fs/btrfs/inode.c

   1 /*
   2  * Copyright (C) 2007 Oracle.  All rights reserved.
   3  *
   4  * This program is free software; you can redistribute it and/or
   5  * modify it under the terms of the GNU General Public
   6  * License v2 as published by the Free Software Foundation.
   7  *
   8  * This program is distributed in the hope that it will be useful,
   9  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  11  * General Public License for more details.
  12  *
  13  * You should have received a copy of the GNU General Public
  14  * License along with this program; if not, write to the
  15  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  16  * Boston, MA 021110-1307, USA.
  17  */
  18
  19 #include <linux/kernel.h>
  20 #include <linux/bio.h>
  21 #include <linux/buffer_head.h>
  22 #include <linux/file.h>
  23 #include <linux/fs.h>
  24 #include <linux/pagemap.h>
  25 #include <linux/highmem.h>
  26 #include <linux/time.h>
  27 #include <linux/init.h>
  28 #include <linux/string.h>
  29 #include <linux/backing-dev.h>
  30 #include <linux/mpage.h>
  31 #include <linux/swap.h>
  32 #include <linux/writeback.h>
  33 #include <linux/statfs.h>
  34 #include <linux/compat.h>
  35 #include <linux/bit_spinlock.h>
  36 #include <linux/xattr.h>
  37 #include <linux/posix_acl.h>
  38 #include <linux/falloc.h>
  39 #include <linux/slab.h>
  40 #include <linux/ratelimit.h>
  41 #include <linux/mount.h>
  42 #include <linux/btrfs.h>
  43 #include <linux/blkdev.h>
  44 #include <linux/posix_acl_xattr.h>
  45 #include <linux/uio.h>
  46 #include "ctree.h"
  47 #include "disk-io.h"
  48 #include "transaction.h"
  49 #include "btrfs_inode.h"
  50 #include "print-tree.h"
  51 #include "ordered-data.h"
  52 #include "xattr.h"
  53 #include "tree-log.h"
  54 #include "volumes.h"
  55 #include "compression.h"
  56 #include "locking.h"
  57 #include "free-space-cache.h"
  58 #include "inode-map.h"
  59 #include "backref.h"
  60 #include "hash.h"
  61 #include "props.h"
  62 #include "qgroup.h"
  63
  64 struct btrfs_iget_args {
  65         struct btrfs_key *location;
  66         struct btrfs_root *root;
  67 };
  68
  69 struct btrfs_dio_data {
  70         u64 outstanding_extents;
  71         u64 reserve;
  72         u64 unsubmitted_oe_range_start;
  73         u64 unsubmitted_oe_range_end;
  74 };
  75
  76 static const struct inode_operations btrfs_dir_inode_operations;
  77 static const struct inode_operations btrfs_symlink_inode_operations;
  78 static const struct inode_operations btrfs_dir_ro_inode_operations;
  79 static const struct inode_operations btrfs_special_inode_operations;
  80 static const struct inode_operations btrfs_file_inode_operations;
  81 static const struct address_space_operations btrfs_aops;
  82 static const struct address_space_operations btrfs_symlink_aops;
  83 static const struct file_operations btrfs_dir_file_operations;
  84 static const struct extent_io_ops btrfs_extent_io_ops;
  85
  86 static struct kmem_cache *btrfs_inode_cachep;
  87 struct kmem_cache *btrfs_trans_handle_cachep;
  88 struct kmem_cache *btrfs_transaction_cachep;
  89 struct kmem_cache *btrfs_path_cachep;
  90 struct kmem_cache *btrfs_free_space_cachep;
  91
  92 #define S_SHIFT 12
  93 static const unsigned char btrfs_type_by_mode[S_IFMT >> S_SHIFT] = {
  94         [S_IFREG >> S_SHIFT]    = BTRFS_FT_REG_FILE,
  95         [S_IFDIR >> S_SHIFT]    = BTRFS_FT_DIR,
  96         [S_IFCHR >> S_SHIFT]    = BTRFS_FT_CHRDEV,
  97         [S_IFBLK >> S_SHIFT]    = BTRFS_FT_BLKDEV,
  98         [S_IFIFO >> S_SHIFT]    = BTRFS_FT_FIFO,
  99         [S_IFSOCK >> S_SHIFT]   = BTRFS_FT_SOCK,
 100         [S_IFLNK >> S_SHIFT]    = BTRFS_FT_SYMLINK,
 101 };
 102
 103 static int btrfs_setsize(struct inode *inode, struct iattr *attr);
 104 static int btrfs_truncate(struct inode *inode);
 105 static int btrfs_finish_ordered_io(struct btrfs_ordered_extent *ordered_extent);
 106 static noinline int cow_file_range(struct inode *inode,
 107                                    struct page *locked_page,
 108                                    u64 start, u64 end, int *page_started,
 109                                    unsigned long *nr_written, int unlock);
 110 static struct extent_map *create_pinned_em(struct inode *inode, u64 start,
 111                                            u64 len, u64 orig_start,
 112                                            u64 block_start, u64 block_len,
 113                                            u64 orig_block_len, u64 ram_bytes,
 114                                            int type);
 115
 116 static int btrfs_dirty_inode(struct inode *inode);
 117
 118 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
 119 void btrfs_test_inode_set_ops(struct inode *inode)
 120 {
 121         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
 122 }
 123 #endif
 124
 125 static int btrfs_init_inode_security(struct btrfs_trans_handle *trans,
 126                                      struct inode *inode,  struct inode *dir,
 127                                      const struct qstr *qstr)
 128 {
 129         int err;
 130
 131         err = btrfs_init_acl(trans, inode, dir);
 132         if (!err)
 133                 err = btrfs_xattr_security_init(trans, inode, dir, qstr);
 134         return err;
 135 }
 136
 137 /*
 138  * this does all the hard work for inserting an inline extent into
 139  * the btree.  The caller should have done a btrfs_drop_extents so that
 140  * no overlapping inline items exist in the btree
 141  */
 142 static int insert_inline_extent(struct btrfs_trans_handle *trans,
 143                                 struct btrfs_path *path, int extent_inserted,
 144                                 struct btrfs_root *root, struct inode *inode,
 145                                 u64 start, size_t size, size_t compressed_size,
 146                                 int compress_type,
 147                                 struct page **compressed_pages)
 148 {
 149         struct extent_buffer *leaf;
 150         struct page *page = NULL;
 151         char *kaddr;
 152         unsigned long ptr;
 153         struct btrfs_file_extent_item *ei;
 154         int err = 0;
 155         int ret;
 156         size_t cur_size = size;
 157         unsigned long offset;
 158
 159         if (compressed_size && compressed_pages)
 160                 cur_size = compressed_size;
 161
 162         inode_add_bytes(inode, size);
 163
 164         if (!extent_inserted) {
 165                 struct btrfs_key key;
 166                 size_t datasize;
 167
 168                 key.objectid = btrfs_ino(inode);
 169                 key.offset = start;
 170                 key.type = BTRFS_EXTENT_DATA_KEY;
 171
 172                 datasize = btrfs_file_extent_calc_inline_size(cur_size);
 173                 path->leave_spinning = 1;
 174                 ret = btrfs_insert_empty_item(trans, root, path, &key,
 175                                               datasize);
 176                 if (ret) {
 177                         err = ret;
 178                         goto fail;
 179                 }
 180         }
 181         leaf = path->nodes[0];
 182         ei = btrfs_item_ptr(leaf, path->slots[0],
 183                             struct btrfs_file_extent_item);
 184         btrfs_set_file_extent_generation(leaf, ei, trans->transid);
 185         btrfs_set_file_extent_type(leaf, ei, BTRFS_FILE_EXTENT_INLINE);
 186         btrfs_set_file_extent_encryption(leaf, ei, 0);
 187         btrfs_set_file_extent_other_encoding(leaf, ei, 0);
 188         btrfs_set_file_extent_ram_bytes(leaf, ei, size);
 189         ptr = btrfs_file_extent_inline_start(ei);
 190
 191         if (compress_type != BTRFS_COMPRESS_NONE) {
 192                 struct page *cpage;
 193                 int i = 0;
 194                 while (compressed_size > 0) {
 195                         cpage = compressed_pages[i];
 196                         cur_size = min_t(unsigned long, compressed_size,
 197                                        PAGE_SIZE);
 198
 199                         kaddr = kmap_atomic(cpage);
 200                         write_extent_buffer(leaf, kaddr, ptr, cur_size);
 201                         kunmap_atomic(kaddr);
 202
 203                         i++;
 204                         ptr += cur_size;
 205                         compressed_size -= cur_size;
 206                 }
 207                 btrfs_set_file_extent_compression(leaf, ei,
 208                                                   compress_type);
 209         } else {
 210                 page = find_get_page(inode->i_mapping,
 211                                      start >> PAGE_SHIFT);
 212                 btrfs_set_file_extent_compression(leaf, ei, 0);
 213                 kaddr = kmap_atomic(page);
 214                 offset = start & (PAGE_SIZE - 1);
 215                 write_extent_buffer(leaf, kaddr + offset, ptr, size);
 216                 kunmap_atomic(kaddr);
 217                 put_page(page);
 218         }
 219         btrfs_mark_buffer_dirty(leaf);
 220         btrfs_release_path(path);
 221
 222         /*
 223          * we're an inline extent, so nobody can
 224          * extend the file past i_size without locking
 225          * a page we already have locked.
 226          *
 227          * We must do any isize and inode updates
 228          * before we unlock the pages.  Otherwise we
 229          * could end up racing with unlink.
 230          */
 231         BTRFS_I(inode)->disk_i_size = inode->i_size;
 232         ret = btrfs_update_inode(trans, root, inode);
 233
 234         return ret;
 235 fail:
 236         return err;
 237 }
 238
 239
 240 /*
 241  * conditionally insert an inline extent into the file.  This
 242  * does the checks required to make sure the data is small enough
 243  * to fit as an inline extent.
 244  */
 245 static noinline int cow_file_range_inline(struct btrfs_root *root,
 246                                           struct inode *inode, u64 start,
 247                                           u64 end, size_t compressed_size,
 248                                           int compress_type,
 249                                           struct page **compressed_pages)
 250 {
 251         struct btrfs_trans_handle *trans;
 252         u64 isize = i_size_read(inode);
 253         u64 actual_end = min(end + 1, isize);
 254         u64 inline_len = actual_end - start;
 255         u64 aligned_end = ALIGN(end, root->sectorsize);
 256         u64 data_len = inline_len;
 257         int ret;
 258         struct btrfs_path *path;
 259         int extent_inserted = 0;
 260         u32 extent_item_size;
 261
 262         if (compressed_size)
 263                 data_len = compressed_size;
 264
 265         if (start > 0 ||
 266             actual_end > root->sectorsize ||
 267             data_len > BTRFS_MAX_INLINE_DATA_SIZE(root) ||
 268             (!compressed_size &&
 269             (actual_end & (root->sectorsize - 1)) == 0) ||
 270             end + 1 < isize ||
 271             data_len > root->fs_info->max_inline) {
 272                 return 1;
 273         }
 274
 275         path = btrfs_alloc_path();
 276         if (!path)
 277                 return -ENOMEM;
 278
 279         trans = btrfs_join_transaction(root);
 280         if (IS_ERR(trans)) {
 281                 btrfs_free_path(path);
 282                 return PTR_ERR(trans);
 283         }
 284         trans->block_rsv = &root->fs_info->delalloc_block_rsv;
 285
 286         if (compressed_size && compressed_pages)
 287                 extent_item_size = btrfs_file_extent_calc_inline_size(
 288                    compressed_size);
 289         else
 290                 extent_item_size = btrfs_file_extent_calc_inline_size(
 291                     inline_len);
 292
 293         ret = __btrfs_drop_extents(trans, root, inode, path,
 294                                    start, aligned_end, NULL,
 295                                    1, 1, extent_item_size, &extent_inserted);
 296         if (ret) {
 297                 btrfs_abort_transaction(trans, root, ret);
 298                 goto out;
 299         }
 300
 301         if (isize > actual_end)
 302                 inline_len = min_t(u64, isize, actual_end);
 303         ret = insert_inline_extent(trans, path, extent_inserted,
 304                                    root, inode, start,
 305                                    inline_len, compressed_size,
 306                                    compress_type, compressed_pages);
 307         if (ret && ret != -ENOSPC) {
 308                 btrfs_abort_transaction(trans, root, ret);
 309                 goto out;
 310         } else if (ret == -ENOSPC) {
 311                 ret = 1;
 312                 goto out;
 313         }
 314
 315         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &BTRFS_I(inode)->runtime_flags);
 316         btrfs_delalloc_release_metadata(inode, end + 1 - start);
 317         btrfs_drop_extent_cache(inode, start, aligned_end - 1, 0);
 318 out:
 319         /*
 320          * Don't forget to free the reserved space, as for inlined extent
 321          * it won't count as data extent, free them directly here.
 322          * And at reserve time, it's always aligned to page size, so
 323          * just free one page here.
 324          */
 325         btrfs_qgroup_free_data(inode, 0, PAGE_SIZE);
 326         btrfs_free_path(path);
 327         btrfs_end_transaction(trans, root);
 328         return ret;
 329 }
 330
 331 struct async_extent {
 332         u64 start;
 333         u64 ram_size;
 334         u64 compressed_size;
 335         struct page **pages;
 336         unsigned long nr_pages;
 337         int compress_type;
 338         struct list_head list;
 339 };
 340
 341 struct async_cow {
 342         struct inode *inode;
 343         struct btrfs_root *root;
 344         struct page *locked_page;
 345         u64 start;
 346         u64 end;
 347         struct list_head extents;
 348         struct btrfs_work work;
 349 };
 350
 351 static noinline int add_async_extent(struct async_cow *cow,
 352                                      u64 start, u64 ram_size,
 353                                      u64 compressed_size,
 354                                      struct page **pages,
 355                                      unsigned long nr_pages,
 356                                      int compress_type)
 357 {
 358         struct async_extent *async_extent;
 359
 360         async_extent = kmalloc(sizeof(*async_extent), GFP_NOFS);
 361         BUG_ON(!async_extent); /* -ENOMEM */
 362         async_extent->start = start;
 363         async_extent->ram_size = ram_size;
 364         async_extent->compressed_size = compressed_size;
 365         async_extent->pages = pages;
 366         async_extent->nr_pages = nr_pages;
 367         async_extent->compress_type = compress_type;
 368         list_add_tail(&async_extent->list, &cow->extents);
 369         return 0;
 370 }
 371
 372 static inline int inode_need_compress(struct inode *inode)
 373 {
 374         struct btrfs_root *root = BTRFS_I(inode)->root;
 375
 376         /* force compress */
 377         if (btrfs_test_opt(root, FORCE_COMPRESS))
 378                 return 1;
 379         /* bad compression ratios */
 380         if (BTRFS_I(inode)->flags & BTRFS_INODE_NOCOMPRESS)
 381                 return 0;
 382         if (btrfs_test_opt(root, COMPRESS) ||
 383             BTRFS_I(inode)->flags & BTRFS_INODE_COMPRESS ||
 384             BTRFS_I(inode)->force_compress)
 385                 return 1;
 386         return 0;
 387 }
 388
 389 /*
 390  * we create compressed extents in two phases.  The first
 391  * phase compresses a range of pages that have already been
 392  * locked (both pages and state bits are locked).
 393  *
 394  * This is done inside an ordered work queue, and the compression
 395  * is spread across many cpus.  The actual IO submission is step
 396  * two, and the ordered work queue takes care of making sure that
 397  * happens in the same order things were put onto the queue by
 398  * writepages and friends.
 399  *
 400  * If this code finds it can't get good compression, it puts an
 401  * entry onto the work queue to write the uncompressed bytes.  This
 402  * makes sure that both compressed inodes and uncompressed inodes
 403  * are written in the same order that the flusher thread sent them
 404  * down.
 405  */
 406 static noinline void compress_file_range(struct inode *inode,
 407                                         struct page *locked_page,
 408                                         u64 start, u64 end,
 409                                         struct async_cow *async_cow,
 410                                         int *num_added)
 411 {
 412         struct btrfs_root *root = BTRFS_I(inode)->root;
 413         u64 num_bytes;
 414         u64 blocksize = root->sectorsize;
 415         u64 actual_end;
 416         u64 isize = i_size_read(inode);
 417         int ret = 0;
 418         struct page **pages = NULL;
 419         unsigned long nr_pages;
 420         unsigned long nr_pages_ret = 0;
 421         unsigned long total_compressed = 0;
 422         unsigned long total_in = 0;
 423         unsigned long max_compressed = SZ_128K;
 424         unsigned long max_uncompressed = SZ_128K;
 425         int i;
 426         int will_compress;
 427         int compress_type = root->fs_info->compress_type;
 428         int redirty = 0;
 429
 430         /* if this is a small write inside eof, kick off a defrag */
 431         if ((end - start + 1) < SZ_16K &&
 432             (start > 0 || end + 1 < BTRFS_I(inode)->disk_i_size))
 433                 btrfs_add_inode_defrag(NULL, inode);
 434
 435         actual_end = min_t(u64, isize, end + 1);
 436 again:
 437         will_compress = 0;
 438         nr_pages = (end >> PAGE_SHIFT) - (start >> PAGE_SHIFT) + 1;
 439         nr_pages = min_t(unsigned long, nr_pages, SZ_128K / PAGE_SIZE);
 440
 441         /*
 442          * we don't want to send crud past the end of i_size through
 443          * compression, that's just a waste of CPU time.  So, if the
 444          * end of the file is before the start of our current
 445          * requested range of bytes, we bail out to the uncompressed
 446          * cleanup code that can deal with all of this.
 447          *
 448          * It isn't really the fastest way to fix things, but this is a
 449          * very uncommon corner.
 450          */
 451         if (actual_end <= start)
 452                 goto cleanup_and_bail_uncompressed;
 453
 454         total_compressed = actual_end - start;
 455
 456         /*
 457          * skip compression for a small file range(<=blocksize) that
 458          * isn't an inline extent, since it doesn't save disk space at all.
 459          */
 460         if (total_compressed <= blocksize &&
 461            (start > 0 || end + 1 < BTRFS_I(inode)->disk_i_size))
 462                 goto cleanup_and_bail_uncompressed;
 463
 464         /* we want to make sure that amount of ram required to uncompress
 465          * an extent is reasonable, so we limit the total size in ram
 466          * of a compressed extent to 128k.  This is a crucial number
 467          * because it also controls how easily we can spread reads across
 468          * cpus for decompression.
 469          *
 470          * We also want to make sure the amount of IO required to do
 471          * a random read is reasonably small, so we limit the size of
 472          * a compressed extent to 128k.
 473          */
 474         total_compressed = min(total_compressed, max_uncompressed);
 475         num_bytes = ALIGN(end - start + 1, blocksize);
 476         num_bytes = max(blocksize,  num_bytes);
 477         total_in = 0;
 478         ret = 0;
 479
 480         /*
 481          * we do compression for mount -o compress and when the
 482          * inode has not been flagged as nocompress.  This flag can
 483          * change at any time if we discover bad compression ratios.
 484          */
 485         if (inode_need_compress(inode)) {
 486                 WARN_ON(pages);
 487                 pages = kcalloc(nr_pages, sizeof(struct page *), GFP_NOFS);
 488                 if (!pages) {
 489                         /* just bail out to the uncompressed code */
 490                         goto cont;
 491                 }
 492
 493                 if (BTRFS_I(inode)->force_compress)
 494                         compress_type = BTRFS_I(inode)->force_compress;
 495
 496                 /*
 497                  * we need to call clear_page_dirty_for_io on each
 498                  * page in the range.  Otherwise applications with the file
 499                  * mmap'd can wander in and change the page contents while
 500                  * we are compressing them.
 501                  *
 502                  * If the compression fails for any reason, we set the pages
 503                  * dirty again later on.
 504                  */
 505                 extent_range_clear_dirty_for_io(inode, start, end);
 506                 redirty = 1;
 507                 ret = btrfs_compress_pages(compress_type,
 508                                            inode->i_mapping, start,
 509                                            total_compressed, pages,
 510                                            nr_pages, &nr_pages_ret,
 511                                            &total_in,
 512                                            &total_compressed,
 513                                            max_compressed);
 514
 515                 if (!ret) {
 516                         unsigned long offset = total_compressed &
 517                                 (PAGE_SIZE - 1);
 518                         struct page *page = pages[nr_pages_ret - 1];
 519                         char *kaddr;
 520
 521                         /* zero the tail end of the last page, we might be
 522                          * sending it down to disk
 523                          */
 524                         if (offset) {
 525                                 kaddr = kmap_atomic(page);
 526                                 memset(kaddr + offset, 0,
 527                                        PAGE_SIZE - offset);
 528                                 kunmap_atomic(kaddr);
 529                         }
 530                         will_compress = 1;
 531                 }
 532         }
 533 cont:
 534         if (start == 0) {
 535                 /* lets try to make an inline extent */
 536                 if (ret || total_in < (actual_end - start)) {
 537                         /* we didn't compress the entire range, try
 538                          * to make an uncompressed inline extent.
 539                          */
 540                         ret = cow_file_range_inline(root, inode, start, end,
 541                                                     0, 0, NULL);
 542                 } else {
 543                         /* try making a compressed inline extent */
 544                         ret = cow_file_range_inline(root, inode, start, end,
 545                                                     total_compressed,
 546                                                     compress_type, pages);
 547                 }
 548                 if (ret <= 0) {
 549                         unsigned long clear_flags = EXTENT_DELALLOC |
 550                                 EXTENT_DEFRAG;
 551                         unsigned long page_error_op;
 552
 553                         clear_flags |= (ret < 0) ? EXTENT_DO_ACCOUNTING : 0;
 554                         page_error_op = ret < 0 ? PAGE_SET_ERROR : 0;
 555
 556                         /*
 557                          * inline extent creation worked or returned error,
 558                          * we don't need to create any more async work items.
 559                          * Unlock and free up our temp pages.
 560                          */
 561                         extent_clear_unlock_delalloc(inode, start, end, NULL,
 562                                                      clear_flags, PAGE_UNLOCK |
 563                                                      PAGE_CLEAR_DIRTY |
 564                                                      PAGE_SET_WRITEBACK |
 565                                                      page_error_op |
 566                                                      PAGE_END_WRITEBACK);
 567                         goto free_pages_out;
 568                 }
 569         }
 570
 571         if (will_compress) {
 572                 /*
 573                  * we aren't doing an inline extent round the compressed size
 574                  * up to a block size boundary so the allocator does sane
 575                  * things
 576                  */
 577                 total_compressed = ALIGN(total_compressed, blocksize);
 578
 579                 /*
 580                  * one last check to make sure the compression is really a
 581                  * win, compare the page count read with the blocks on disk
 582                  */
 583                 total_in = ALIGN(total_in, PAGE_SIZE);
 584                 if (total_compressed >= total_in) {
 585                         will_compress = 0;
 586                 } else {
 587                         num_bytes = total_in;
 588                 }
 589         }
 590         if (!will_compress && pages) {
 591                 /*
 592                  * the compression code ran but failed to make things smaller,
 593                  * free any pages it allocated and our page pointer array
 594                  */
 595                 for (i = 0; i < nr_pages_ret; i++) {
 596                         WARN_ON(pages[i]->mapping);
 597                         put_page(pages[i]);
 598                 }
 599                 kfree(pages);
 600                 pages = NULL;
 601                 total_compressed = 0;
 602                 nr_pages_ret = 0;
 603
 604                 /* flag the file so we don't compress in the future */
 605                 if (!btrfs_test_opt(root, FORCE_COMPRESS) &&
 606                     !(BTRFS_I(inode)->force_compress)) {
 607                         BTRFS_I(inode)->flags |= BTRFS_INODE_NOCOMPRESS;
 608                 }
 609         }
 610         if (will_compress) {
 611                 *num_added += 1;
 612
 613                 /* the async work queues will take care of doing actual
 614                  * allocation on disk for these compressed pages,
 615                  * and will submit them to the elevator.
 616                  */
 617                 add_async_extent(async_cow, start, num_bytes,
 618                                  total_compressed, pages, nr_pages_ret,
 619                                  compress_type);
 620
 621                 if (start + num_bytes < end) {
 622                         start += num_bytes;
 623                         pages = NULL;
 624                         cond_resched();
 625                         goto again;
 626                 }
 627         } else {
 628 cleanup_and_bail_uncompressed:
 629                 /*
 630                  * No compression, but we still need to write the pages in
 631                  * the file we've been given so far.  redirty the locked
 632                  * page if it corresponds to our extent and set things up
 633                  * for the async work queue to run cow_file_range to do
 634                  * the normal delalloc dance
 635                  */
 636                 if (page_offset(locked_page) >= start &&
 637                     page_offset(locked_page) <= end) {
 638                         __set_page_dirty_nobuffers(locked_page);
 639                         /* unlocked later on in the async handlers */
 640                 }
 641                 if (redirty)
 642                         extent_range_redirty_for_io(inode, start, end);
 643                 add_async_extent(async_cow, start, end - start + 1,
 644                                  0, NULL, 0, BTRFS_COMPRESS_NONE);
 645                 *num_added += 1;
 646         }
 647
 648         return;
 649
 650 free_pages_out:
 651         for (i = 0; i < nr_pages_ret; i++) {
 652                 WARN_ON(pages[i]->mapping);
 653                 put_page(pages[i]);
 654         }
 655         kfree(pages);
 656 }
 657
 658 static void free_async_extent_pages(struct async_extent *async_extent)
 659 {
 660         int i;
 661
 662         if (!async_extent->pages)
 663                 return;
 664
 665         for (i = 0; i < async_extent->nr_pages; i++) {
 666                 WARN_ON(async_extent->pages[i]->mapping);
 667                 put_page(async_extent->pages[i]);
 668         }
 669         kfree(async_extent->pages);
 670         async_extent->nr_pages = 0;
 671         async_extent->pages = NULL;
 672 }
 673
 674 /*
 675  * phase two of compressed writeback.  This is the ordered portion
 676  * of the code, which only gets called in the order the work was
 677  * queued.  We walk all the async extents created by compress_file_range
 678  * and send them down to the disk.
 679  */
 680 static noinline void submit_compressed_extents(struct inode *inode,
 681                                               struct async_cow *async_cow)
 682 {
 683         struct async_extent *async_extent;
 684         u64 alloc_hint = 0;
 685         struct btrfs_key ins;
 686         struct extent_map *em;
 687         struct btrfs_root *root = BTRFS_I(inode)->root;
 688         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
 689         struct extent_io_tree *io_tree;
 690         int ret = 0;
 691
 692 again:
 693         while (!list_empty(&async_cow->extents)) {
 694                 async_extent = list_entry(async_cow->extents.next,
 695                                           struct async_extent, list);
 696                 list_del(&async_extent->list);
 697
 698                 io_tree = &BTRFS_I(inode)->io_tree;
 699
 700 retry:
 701                 /* did the compression code fall back to uncompressed IO? */
 702                 if (!async_extent->pages) {
 703                         int page_started = 0;
 704                         unsigned long nr_written = 0;
 705
 706                         lock_extent(io_tree, async_extent->start,
 707                                          async_extent->start +
 708                                          async_extent->ram_size - 1);
 709
 710                         /* allocate blocks */
 711                         ret = cow_file_range(inode, async_cow->locked_page,
 712                                              async_extent->start,
 713                                              async_extent->start +
 714                                              async_extent->ram_size - 1,
 715                                              &page_started, &nr_written, 0);
 716
 717                         /* JDM XXX */
 718
 719                         /*
 720                          * if page_started, cow_file_range inserted an
 721                          * inline extent and took care of all the unlocking
 722                          * and IO for us.  Otherwise, we need to submit
 723                          * all those pages down to the drive.
 724                          */
 725                         if (!page_started && !ret)
 726                                 extent_write_locked_range(io_tree,
 727                                                   inode, async_extent->start,
 728                                                   async_extent->start +
 729                                                   async_extent->ram_size - 1,
 730                                                   btrfs_get_extent,
 731                                                   WB_SYNC_ALL);
 732                         else if (ret)
 733                                 unlock_page(async_cow->locked_page);
 734                         kfree(async_extent);
 735                         cond_resched();
 736                         continue;
 737                 }
 738
 739                 lock_extent(io_tree, async_extent->start,
 740                             async_extent->start + async_extent->ram_size - 1);
 741
 742                 ret = btrfs_reserve_extent(root,
 743                                            async_extent->compressed_size,
 744                                            async_extent->compressed_size,
 745                                            0, alloc_hint, &ins, 1, 1);
 746                 if (ret) {
 747                         free_async_extent_pages(async_extent);
 748
 749                         if (ret == -ENOSPC) {
 750                                 unlock_extent(io_tree, async_extent->start,
 751                                               async_extent->start +
 752                                               async_extent->ram_size - 1);
 753
 754                                 /*
 755                                  * we need to redirty the pages if we decide to
 756                                  * fallback to uncompressed IO, otherwise we
 757                                  * will not submit these pages down to lower
 758                                  * layers.
 759                                  */
 760                                 extent_range_redirty_for_io(inode,
 761                                                 async_extent->start,
 762                                                 async_extent->start +
 763                                                 async_extent->ram_size - 1);
 764
 765                                 goto retry;
 766                         }
 767                         goto out_free;
 768                 }
 769                 /*
 770                  * here we're doing allocation and writeback of the
 771                  * compressed pages
 772                  */
 773                 btrfs_drop_extent_cache(inode, async_extent->start,
 774                                         async_extent->start +
 775                                         async_extent->ram_size - 1, 0);
 776
 777                 em = alloc_extent_map();
 778                 if (!em) {
 779                         ret = -ENOMEM;
 780                         goto out_free_reserve;
 781                 }
 782                 em->start = async_extent->start;
 783                 em->len = async_extent->ram_size;
 784                 em->orig_start = em->start;
 785                 em->mod_start = em->start;
 786                 em->mod_len = em->len;
 787
 788                 em->block_start = ins.objectid;
 789                 em->block_len = ins.offset;
 790                 em->orig_block_len = ins.offset;
 791                 em->ram_bytes = async_extent->ram_size;
 792                 em->bdev = root->fs_info->fs_devices->latest_bdev;
 793                 em->compress_type = async_extent->compress_type;
 794                 set_bit(EXTENT_FLAG_PINNED, &em->flags);
 795                 set_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
 796                 em->generation = -1;
 797
 798                 while (1) {
 799                         write_lock(&em_tree->lock);
 800                         ret = add_extent_mapping(em_tree, em, 1);
 801                         write_unlock(&em_tree->lock);
 802                         if (ret != -EEXIST) {
 803                                 free_extent_map(em);
 804                                 break;
 805                         }
 806                         btrfs_drop_extent_cache(inode, async_extent->start,
 807                                                 async_extent->start +
 808                                                 async_extent->ram_size - 1, 0);
 809                 }
 810
 811                 if (ret)
 812                         goto out_free_reserve;
 813
 814                 ret = btrfs_add_ordered_extent_compress(inode,
 815                                                 async_extent->start,
 816                                                 ins.objectid,
 817                                                 async_extent->ram_size,
 818                                                 ins.offset,
 819                                                 BTRFS_ORDERED_COMPRESSED,
 820                                                 async_extent->compress_type);
 821                 if (ret) {
 822                         btrfs_drop_extent_cache(inode, async_extent->start,
 823                                                 async_extent->start +
 824                                                 async_extent->ram_size - 1, 0);
 825                         goto out_free_reserve;
 826                 }
 827                 btrfs_dec_block_group_reservations(root->fs_info, ins.objectid);
 828
 829                 /*
 830                  * clear dirty, set writeback and unlock the pages.
 831                  */
 832                 extent_clear_unlock_delalloc(inode, async_extent->start,
 833                                 async_extent->start +
 834                                 async_extent->ram_size - 1,
 835                                 NULL, EXTENT_LOCKED | EXTENT_DELALLOC,
 836                                 PAGE_UNLOCK | PAGE_CLEAR_DIRTY |
 837                                 PAGE_SET_WRITEBACK);
 838                 ret = btrfs_submit_compressed_write(inode,
 839                                     async_extent->start,
 840                                     async_extent->ram_size,
 841                                     ins.objectid,
 842                                     ins.offset, async_extent->pages,
 843                                     async_extent->nr_pages);
 844                 if (ret) {
 845                         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
 846                         struct page *p = async_extent->pages[0];
 847                         const u64 start = async_extent->start;
 848                         const u64 end = start + async_extent->ram_size - 1;
 849
 850                         p->mapping = inode->i_mapping;
 851                         tree->ops->writepage_end_io_hook(p, start, end,
 852                                                          NULL, 0);
 853                         p->mapping = NULL;
 854                         extent_clear_unlock_delalloc(inode, start, end, NULL, 0,
 855                                                      PAGE_END_WRITEBACK |
 856                                                      PAGE_SET_ERROR);
 857                         free_async_extent_pages(async_extent);
 858                 }
 859                 alloc_hint = ins.objectid + ins.offset;
 860                 kfree(async_extent);
 861                 cond_resched();
 862         }
 863         return;
 864 out_free_reserve:
 865         btrfs_dec_block_group_reservations(root->fs_info, ins.objectid);
 866         btrfs_free_reserved_extent(root, ins.objectid, ins.offset, 1);
 867 out_free:
 868         extent_clear_unlock_delalloc(inode, async_extent->start,
 869                                      async_extent->start +
 870                                      async_extent->ram_size - 1,
 871                                      NULL, EXTENT_LOCKED | EXTENT_DELALLOC |
 872                                      EXTENT_DEFRAG | EXTENT_DO_ACCOUNTING,
 873                                      PAGE_UNLOCK | PAGE_CLEAR_DIRTY |
 874                                      PAGE_SET_WRITEBACK | PAGE_END_WRITEBACK |
 875                                      PAGE_SET_ERROR);
 876         free_async_extent_pages(async_extent);
 877         kfree(async_extent);
 878         goto again;
 879 }
 880
 881 static u64 get_extent_allocation_hint(struct inode *inode, u64 start,
 882                                       u64 num_bytes)
 883 {
 884         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
 885         struct extent_map *em;
 886         u64 alloc_hint = 0;
 887
 888         read_lock(&em_tree->lock);
 889         em = search_extent_mapping(em_tree, start, num_bytes);
 890         if (em) {
 891                 /*
 892                  * if block start isn't an actual block number then find the
 893                  * first block in this inode and use that as a hint.  If that
 894                  * block is also bogus then just don't worry about it.
 895                  */
 896                 if (em->block_start >= EXTENT_MAP_LAST_BYTE) {
 897                         free_extent_map(em);
 898                         em = search_extent_mapping(em_tree, 0, 0);
 899                         if (em && em->block_start < EXTENT_MAP_LAST_BYTE)
 900                                 alloc_hint = em->block_start;
 901                         if (em)
 902                                 free_extent_map(em);
 903                 } else {
 904                         alloc_hint = em->block_start;
 905                         free_extent_map(em);
 906                 }
 907         }
 908         read_unlock(&em_tree->lock);
 909
 910         return alloc_hint;
 911 }
 912
 913 /*
 914  * when extent_io.c finds a delayed allocation range in the file,
 915  * the call backs end up in this code.  The basic idea is to
 916  * allocate extents on disk for the range, and create ordered data structs
 917  * in ram to track those extents.
 918  *
 919  * locked_page is the page that writepage had locked already.  We use
 920  * it to make sure we don't do extra locks or unlocks.
 921  *
 922  * *page_started is set to one if we unlock locked_page and do everything
 923  * required to start IO on it.  It may be clean and already done with
 924  * IO when we return.
 925  */
 926 static noinline int cow_file_range(struct inode *inode,
 927                                    struct page *locked_page,
 928                                    u64 start, u64 end, int *page_started,
 929                                    unsigned long *nr_written,
 930                                    int unlock)
 931 {
 932         struct btrfs_root *root = BTRFS_I(inode)->root;
 933         u64 alloc_hint = 0;
 934         u64 num_bytes;
 935         unsigned long ram_size;
 936         u64 disk_num_bytes;
 937         u64 cur_alloc_size;
 938         u64 blocksize = root->sectorsize;
 939         struct btrfs_key ins;
 940         struct extent_map *em;
 941         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
 942         int ret = 0;
 943
 944         if (btrfs_is_free_space_inode(inode)) {
 945                 WARN_ON_ONCE(1);
 946                 ret = -EINVAL;
 947                 goto out_unlock;
 948         }
 949
 950         num_bytes = ALIGN(end - start + 1, blocksize);
 951         num_bytes = max(blocksize,  num_bytes);
 952         disk_num_bytes = num_bytes;
 953
 954         /* if this is a small write inside eof, kick off defrag */
 955         if (num_bytes < SZ_64K &&
 956             (start > 0 || end + 1 < BTRFS_I(inode)->disk_i_size))
 957                 btrfs_add_inode_defrag(NULL, inode);
 958
 959         if (start == 0) {
 960                 /* lets try to make an inline extent */
 961                 ret = cow_file_range_inline(root, inode, start, end, 0, 0,
 962                                             NULL);
 963                 if (ret == 0) {
 964                         extent_clear_unlock_delalloc(inode, start, end, NULL,
 965                                      EXTENT_LOCKED | EXTENT_DELALLOC |
 966                                      EXTENT_DEFRAG, PAGE_UNLOCK |
 967                                      PAGE_CLEAR_DIRTY | PAGE_SET_WRITEBACK |
 968                                      PAGE_END_WRITEBACK);
 969
 970                         *nr_written = *nr_written +
 971                              (end - start + PAGE_SIZE) / PAGE_SIZE;
 972                         *page_started = 1;
 973                         goto out;
 974                 } else if (ret < 0) {
 975                         goto out_unlock;
 976                 }
 977         }
 978
 979         BUG_ON(disk_num_bytes >
 980                btrfs_super_total_bytes(root->fs_info->super_copy));
 981
 982         alloc_hint = get_extent_allocation_hint(inode, start, num_bytes);
 983         btrfs_drop_extent_cache(inode, start, start + num_bytes - 1, 0);
 984
 985         while (disk_num_bytes > 0) {
 986                 unsigned long op;
 987
 988                 cur_alloc_size = disk_num_bytes;
 989                 ret = btrfs_reserve_extent(root, cur_alloc_size,
 990                                            root->sectorsize, 0, alloc_hint,
 991                                            &ins, 1, 1);
 992                 if (ret < 0)
 993                         goto out_unlock;
 994
 995                 em = alloc_extent_map();
 996                 if (!em) {
 997                         ret = -ENOMEM;
 998                         goto out_reserve;
 999                 }
1000                 em->start = start;
1001                 em->orig_start = em->start;
1002                 ram_size = ins.offset;
1003                 em->len = ins.offset;
1004                 em->mod_start = em->start;
1005                 em->mod_len = em->len;
1006
1007                 em->block_start = ins.objectid;
1008                 em->block_len = ins.offset;
1009                 em->orig_block_len = ins.offset;
1010                 em->ram_bytes = ram_size;
1011                 em->bdev = root->fs_info->fs_devices->latest_bdev;
1012                 set_bit(EXTENT_FLAG_PINNED, &em->flags);
1013                 em->generation = -1;
1014
1015                 while (1) {
1016                         write_lock(&em_tree->lock);
1017                         ret = add_extent_mapping(em_tree, em, 1);
1018                         write_unlock(&em_tree->lock);
1019                         if (ret != -EEXIST) {
1020                                 free_extent_map(em);
1021                                 break;
1022                         }
1023                         btrfs_drop_extent_cache(inode, start,
1024                                                 start + ram_size - 1, 0);
1025                 }
1026                 if (ret)
1027                         goto out_reserve;
1028
1029                 cur_alloc_size = ins.offset;
1030                 ret = btrfs_add_ordered_extent(inode, start, ins.objectid,
1031                                                ram_size, cur_alloc_size, 0);
1032                 if (ret)
1033                         goto out_drop_extent_cache;
1034
1035                 if (root->root_key.objectid ==
1036                     BTRFS_DATA_RELOC_TREE_OBJECTID) {
1037                         ret = btrfs_reloc_clone_csums(inode, start,
1038                                                       cur_alloc_size);
1039                         if (ret)
1040                                 goto out_drop_extent_cache;
1041                 }
1042
1043                 btrfs_dec_block_group_reservations(root->fs_info, ins.objectid);
1044
1045                 if (disk_num_bytes < cur_alloc_size)
1046                         break;
1047
1048                 /* we're not doing compressed IO, don't unlock the first
1049                  * page (which the caller expects to stay locked), don't
1050                  * clear any dirty bits and don't set any writeback bits
1051                  *
1052                  * Do set the Private2 bit so we know this page was properly
1053                  * setup for writepage
1054                  */
1055                 op = unlock ? PAGE_UNLOCK : 0;
1056                 op |= PAGE_SET_PRIVATE2;
1057
1058                 extent_clear_unlock_delalloc(inode, start,
1059                                              start + ram_size - 1, locked_page,
1060                                              EXTENT_LOCKED | EXTENT_DELALLOC,
1061                                              op);
1062                 disk_num_bytes -= cur_alloc_size;
1063                 num_bytes -= cur_alloc_size;
1064                 alloc_hint = ins.objectid + ins.offset;
1065                 start += cur_alloc_size;
1066         }
1067 out:
1068         return ret;
1069
1070 out_drop_extent_cache:
1071         btrfs_drop_extent_cache(inode, start, start + ram_size - 1, 0);
1072 out_reserve:
1073         btrfs_dec_block_group_reservations(root->fs_info, ins.objectid);
1074         btrfs_free_reserved_extent(root, ins.objectid, ins.offset, 1);
1075 out_unlock:
1076         extent_clear_unlock_delalloc(inode, start, end, locked_page,
1077                                      EXTENT_LOCKED | EXTENT_DO_ACCOUNTING |
1078                                      EXTENT_DELALLOC | EXTENT_DEFRAG,
1079                                      PAGE_UNLOCK | PAGE_CLEAR_DIRTY |
1080                                      PAGE_SET_WRITEBACK | PAGE_END_WRITEBACK);
1081         goto out;
1082 }
1083
1084 /*
1085  * work queue call back to started compression on a file and pages
1086  */
1087 static noinline void async_cow_start(struct btrfs_work *work)
1088 {
1089         struct async_cow *async_cow;
1090         int num_added = 0;
1091         async_cow = container_of(work, struct async_cow, work);
1092
1093         compress_file_range(async_cow->inode, async_cow->locked_page,
1094                             async_cow->start, async_cow->end, async_cow,
1095                             &num_added);
1096         if (num_added == 0) {
1097                 btrfs_add_delayed_iput(async_cow->inode);
1098                 async_cow->inode = NULL;
1099         }
1100 }
1101
1102 /*
1103  * work queue call back to submit previously compressed pages
1104  */
1105 static noinline void async_cow_submit(struct btrfs_work *work)
1106 {
1107         struct async_cow *async_cow;
1108         struct btrfs_root *root;
1109         unsigned long nr_pages;
1110
1111         async_cow = container_of(work, struct async_cow, work);
1112
1113         root = async_cow->root;
1114         nr_pages = (async_cow->end - async_cow->start + PAGE_SIZE) >>
1115                 PAGE_SHIFT;
1116
1117         /*
1118          * atomic_sub_return implies a barrier for waitqueue_active
1119          */
1120         if (atomic_sub_return(nr_pages, &root->fs_info->async_delalloc_pages) <
1121             5 * SZ_1M &&
1122             waitqueue_active(&root->fs_info->async_submit_wait))
1123                 wake_up(&root->fs_info->async_submit_wait);
1124
1125         if (async_cow->inode)
1126                 submit_compressed_extents(async_cow->inode, async_cow);
1127 }
1128
1129 static noinline void async_cow_free(struct btrfs_work *work)
1130 {
1131         struct async_cow *async_cow;
1132         async_cow = container_of(work, struct async_cow, work);
1133         if (async_cow->inode)
1134                 btrfs_add_delayed_iput(async_cow->inode);
1135         kfree(async_cow);
1136 }
1137
1138 static int cow_file_range_async(struct inode *inode, struct page *locked_page,
1139                                 u64 start, u64 end, int *page_started,
1140                                 unsigned long *nr_written)
1141 {
1142         struct async_cow *async_cow;
1143         struct btrfs_root *root = BTRFS_I(inode)->root;
1144         unsigned long nr_pages;
1145         u64 cur_end;
1146         int limit = 10 * SZ_1M;
1147
1148         clear_extent_bit(&BTRFS_I(inode)->io_tree, start, end, EXTENT_LOCKED,
1149                          1, 0, NULL, GFP_NOFS);
1150         while (start < end) {
1151                 async_cow = kmalloc(sizeof(*async_cow), GFP_NOFS);
1152                 BUG_ON(!async_cow); /* -ENOMEM */
1153                 async_cow->inode = igrab(inode);
1154                 async_cow->root = root;
1155                 async_cow->locked_page = locked_page;
1156                 async_cow->start = start;
1157
1158                 if (BTRFS_I(inode)->flags & BTRFS_INODE_NOCOMPRESS &&
1159                     !btrfs_test_opt(root, FORCE_COMPRESS))
1160                         cur_end = end;
1161                 else
1162                         cur_end = min(end, start + SZ_512K - 1);
1163
1164                 async_cow->end = cur_end;
1165                 INIT_LIST_HEAD(&async_cow->extents);
1166
1167                 btrfs_init_work(&async_cow->work,
1168                                 btrfs_delalloc_helper,
1169                                 async_cow_start, async_cow_submit,
1170                                 async_cow_free);
1171
1172                 nr_pages = (cur_end - start + PAGE_SIZE) >>
1173                         PAGE_SHIFT;
1174                 atomic_add(nr_pages, &root->fs_info->async_delalloc_pages);
1175
1176                 btrfs_queue_work(root->fs_info->delalloc_workers,
1177                                  &async_cow->work);
1178
1179                 if (atomic_read(&root->fs_info->async_delalloc_pages) > limit) {
1180                         wait_event(root->fs_info->async_submit_wait,
1181                            (atomic_read(&root->fs_info->async_delalloc_pages) <
1182                             limit));
1183                 }
1184
1185                 while (atomic_read(&root->fs_info->async_submit_draining) &&
1186                       atomic_read(&root->fs_info->async_delalloc_pages)) {
1187                         wait_event(root->fs_info->async_submit_wait,
1188                           (atomic_read(&root->fs_info->async_delalloc_pages) ==
1189                            0));
1190                 }
1191
1192                 *nr_written += nr_pages;
1193                 start = cur_end + 1;
1194         }
1195         *page_started = 1;
1196         return 0;
1197 }
1198
1199 static noinline int csum_exist_in_range(struct btrfs_root *root,
1200                                         u64 bytenr, u64 num_bytes)
1201 {
1202         int ret;
1203         struct btrfs_ordered_sum *sums;
1204         LIST_HEAD(list);
1205
1206         ret = btrfs_lookup_csums_range(root->fs_info->csum_root, bytenr,
1207                                        bytenr + num_bytes - 1, &list, 0);
1208         if (ret == 0 && list_empty(&list))
1209                 return 0;
1210
1211         while (!list_empty(&list)) {
1212                 sums = list_entry(list.next, struct btrfs_ordered_sum, list);
1213                 list_del(&sums->list);
1214                 kfree(sums);
1215         }
1216         return 1;
1217 }
1218
1219 /*
1220  * when nowcow writeback call back.  This checks for snapshots or COW copies
1221  * of the extents that exist in the file, and COWs the file as required.
1222  *
1223  * If no cow copies or snapshots exist, we write directly to the existing
1224  * blocks on disk
1225  */
1226 static noinline int run_delalloc_nocow(struct inode *inode,
1227                                        struct page *locked_page,
1228                               u64 start, u64 end, int *page_started, int force,
1229                               unsigned long *nr_written)
1230 {
1231         struct btrfs_root *root = BTRFS_I(inode)->root;
1232         struct btrfs_trans_handle *trans;
1233         struct extent_buffer *leaf;
1234         struct btrfs_path *path;
1235         struct btrfs_file_extent_item *fi;
1236         struct btrfs_key found_key;
1237         u64 cow_start;
1238         u64 cur_offset;
1239         u64 extent_end;
1240         u64 extent_offset;
1241         u64 disk_bytenr;
1242         u64 num_bytes;
1243         u64 disk_num_bytes;
1244         u64 ram_bytes;
1245         int extent_type;
1246         int ret, err;
1247         int type;
1248         int nocow;
1249         int check_prev = 1;
1250         bool nolock;
1251         u64 ino = btrfs_ino(inode);
1252
1253         path = btrfs_alloc_path();
1254         if (!path) {
1255                 extent_clear_unlock_delalloc(inode, start, end, locked_page,
1256                                              EXTENT_LOCKED | EXTENT_DELALLOC |
1257                                              EXTENT_DO_ACCOUNTING |
1258                                              EXTENT_DEFRAG, PAGE_UNLOCK |
1259                                              PAGE_CLEAR_DIRTY |
1260                                              PAGE_SET_WRITEBACK |
1261                                              PAGE_END_WRITEBACK);
1262                 return -ENOMEM;
1263         }
1264
1265         nolock = btrfs_is_free_space_inode(inode);
1266
1267         if (nolock)
1268                 trans = btrfs_join_transaction_nolock(root);
1269         else
1270                 trans = btrfs_join_transaction(root);
1271
1272         if (IS_ERR(trans)) {
1273                 extent_clear_unlock_delalloc(inode, start, end, locked_page,
1274                                              EXTENT_LOCKED | EXTENT_DELALLOC |
1275                                              EXTENT_DO_ACCOUNTING |
1276                                              EXTENT_DEFRAG, PAGE_UNLOCK |
1277                                              PAGE_CLEAR_DIRTY |
1278                                              PAGE_SET_WRITEBACK |
1279                                              PAGE_END_WRITEBACK);
1280                 btrfs_free_path(path);
1281                 return PTR_ERR(trans);
1282         }
1283
1284         trans->block_rsv = &root->fs_info->delalloc_block_rsv;
1285
1286         cow_start = (u64)-1;
1287         cur_offset = start;
1288         while (1) {
1289                 ret = btrfs_lookup_file_extent(trans, root, path, ino,
1290                                                cur_offset, 0);
1291                 if (ret < 0)
1292                         goto error;
1293                 if (ret > 0 && path->slots[0] > 0 && check_prev) {
1294                         leaf = path->nodes[0];
1295                         btrfs_item_key_to_cpu(leaf, &found_key,
1296                                               path->slots[0] - 1);
1297                         if (found_key.objectid == ino &&
1298                             found_key.type == BTRFS_EXTENT_DATA_KEY)
1299                                 path->slots[0]--;
1300                 }
1301                 check_prev = 0;
1302 next_slot:
1303                 leaf = path->nodes[0];
1304                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
1305                         ret = btrfs_next_leaf(root, path);
1306                         if (ret < 0)
1307                                 goto error;
1308                         if (ret > 0)
1309                                 break;
1310                         leaf = path->nodes[0];
1311                 }
1312
1313                 nocow = 0;
1314                 disk_bytenr = 0;
1315                 num_bytes = 0;
1316                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
1317
1318                 if (found_key.objectid > ino)
1319                         break;
1320                 if (WARN_ON_ONCE(found_key.objectid < ino) ||
1321                     found_key.type < BTRFS_EXTENT_DATA_KEY) {
1322                         path->slots[0]++;
1323                         goto next_slot;
1324                 }
1325                 if (found_key.type > BTRFS_EXTENT_DATA_KEY ||
1326                     found_key.offset > end)
1327                         break;
1328
1329                 if (found_key.offset > cur_offset) {
1330                         extent_end = found_key.offset;
1331                         extent_type = 0;
1332                         goto out_check;
1333                 }
1334
1335                 fi = btrfs_item_ptr(leaf, path->slots[0],
1336                                     struct btrfs_file_extent_item);
1337                 extent_type = btrfs_file_extent_type(leaf, fi);
1338
1339                 ram_bytes = btrfs_file_extent_ram_bytes(leaf, fi);
1340                 if (extent_type == BTRFS_FILE_EXTENT_REG ||
1341                     extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
1342                         disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1343                         extent_offset = btrfs_file_extent_offset(leaf, fi);
1344                         extent_end = found_key.offset +
1345                                 btrfs_file_extent_num_bytes(leaf, fi);
1346                         disk_num_bytes =
1347                                 btrfs_file_extent_disk_num_bytes(leaf, fi);
1348                         if (extent_end <= start) {
1349                                 path->slots[0]++;
1350                                 goto next_slot;
1351                         }
1352                         if (disk_bytenr == 0)
1353                                 goto out_check;
1354                         if (btrfs_file_extent_compression(leaf, fi) ||
1355                             btrfs_file_extent_encryption(leaf, fi) ||
1356                             btrfs_file_extent_other_encoding(leaf, fi))
1357                                 goto out_check;
1358                         if (extent_type == BTRFS_FILE_EXTENT_REG && !force)
1359                                 goto out_check;
1360                         if (btrfs_extent_readonly(root, disk_bytenr))
1361                                 goto out_check;
1362                         if (btrfs_cross_ref_exist(trans, root, ino,
1363                                                   found_key.offset -
1364                                                   extent_offset, disk_bytenr))
1365                                 goto out_check;
1366                         disk_bytenr += extent_offset;
1367                         disk_bytenr += cur_offset - found_key.offset;
1368                         num_bytes = min(end + 1, extent_end) - cur_offset;
1369                         /*
1370                          * if there are pending snapshots for this root,
1371                          * we fall into common COW way.
1372                          */
1373                         if (!nolock) {
1374                                 err = btrfs_start_write_no_snapshoting(root);
1375                                 if (!err)
1376                                         goto out_check;
1377                         }
1378                         /*
1379                          * force cow if csum exists in the range.
1380                          * this ensure that csum for a given extent are
1381                          * either valid or do not exist.
1382                          */
1383                         if (csum_exist_in_range(root, disk_bytenr, num_bytes))
1384                                 goto out_check;
1385                         if (!btrfs_inc_nocow_writers(root->fs_info,
1386                                                      disk_bytenr))
1387                                 goto out_check;
1388                         nocow = 1;
1389                 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
1390                         extent_end = found_key.offset +
1391                                 btrfs_file_extent_inline_len(leaf,
1392                                                      path->slots[0], fi);
1393                         extent_end = ALIGN(extent_end, root->sectorsize);
1394                 } else {
1395                         BUG_ON(1);
1396                 }
1397 out_check:
1398                 if (extent_end <= start) {
1399                         path->slots[0]++;
1400                         if (!nolock && nocow)
1401                                 btrfs_end_write_no_snapshoting(root);
1402                         if (nocow)
1403                                 btrfs_dec_nocow_writers(root->fs_info,
1404                                                         disk_bytenr);
1405                         goto next_slot;
1406                 }
1407                 if (!nocow) {
1408                         if (cow_start == (u64)-1)
1409                                 cow_start = cur_offset;
1410                         cur_offset = extent_end;
1411                         if (cur_offset > end)
1412                                 break;
1413                         path->slots[0]++;
1414                         goto next_slot;
1415                 }
1416
1417                 btrfs_release_path(path);
1418                 if (cow_start != (u64)-1) {
1419                         ret = cow_file_range(inode, locked_page,
1420                                              cow_start, found_key.offset - 1,
1421                                              page_started, nr_written, 1);
1422                         if (ret) {
1423                                 if (!nolock && nocow)
1424                                         btrfs_end_write_no_snapshoting(root);
1425                                 if (nocow)
1426                                         btrfs_dec_nocow_writers(root->fs_info,
1427                                                                 disk_bytenr);
1428                                 goto error;
1429                         }
1430                         cow_start = (u64)-1;
1431                 }
1432
1433                 if (extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
1434                         struct extent_map *em;
1435                         struct extent_map_tree *em_tree;
1436                         em_tree = &BTRFS_I(inode)->extent_tree;
1437                         em = alloc_extent_map();
1438                         BUG_ON(!em); /* -ENOMEM */
1439                         em->start = cur_offset;
1440                         em->orig_start = found_key.offset - extent_offset;
1441                         em->len = num_bytes;
1442                         em->block_len = num_bytes;
1443                         em->block_start = disk_bytenr;
1444                         em->orig_block_len = disk_num_bytes;
1445                         em->ram_bytes = ram_bytes;
1446                         em->bdev = root->fs_info->fs_devices->latest_bdev;
1447                         em->mod_start = em->start;
1448                         em->mod_len = em->len;
1449                         set_bit(EXTENT_FLAG_PINNED, &em->flags);
1450                         set_bit(EXTENT_FLAG_FILLING, &em->flags);
1451                         em->generation = -1;
1452                         while (1) {
1453                                 write_lock(&em_tree->lock);
1454                                 ret = add_extent_mapping(em_tree, em, 1);
1455                                 write_unlock(&em_tree->lock);
1456                                 if (ret != -EEXIST) {
1457                                         free_extent_map(em);
1458                                         break;
1459                                 }
1460                                 btrfs_drop_extent_cache(inode, em->start,
1461                                                 em->start + em->len - 1, 0);
1462                         }
1463                         type = BTRFS_ORDERED_PREALLOC;
1464                 } else {
1465                         type = BTRFS_ORDERED_NOCOW;
1466                 }
1467
1468                 ret = btrfs_add_ordered_extent(inode, cur_offset, disk_bytenr,
1469                                                num_bytes, num_bytes, type);
1470                 if (nocow)
1471                         btrfs_dec_nocow_writers(root->fs_info, disk_bytenr);
1472                 BUG_ON(ret); /* -ENOMEM */
1473
1474                 if (root->root_key.objectid ==
1475                     BTRFS_DATA_RELOC_TREE_OBJECTID) {
1476                         ret = btrfs_reloc_clone_csums(inode, cur_offset,
1477                                                       num_bytes);
1478                         if (ret) {
1479                                 if (!nolock && nocow)
1480                                         btrfs_end_write_no_snapshoting(root);
1481                                 goto error;
1482                         }
1483                 }
1484
1485                 extent_clear_unlock_delalloc(inode, cur_offset,
1486                                              cur_offset + num_bytes - 1,
1487                                              locked_page, EXTENT_LOCKED |
1488                                              EXTENT_DELALLOC, PAGE_UNLOCK |
1489                                              PAGE_SET_PRIVATE2);
1490                 if (!nolock && nocow)
1491                         btrfs_end_write_no_snapshoting(root);
1492                 cur_offset = extent_end;
1493                 if (cur_offset > end)
1494                         break;
1495         }
1496         btrfs_release_path(path);
1497
1498         if (cur_offset <= end && cow_start == (u64)-1) {
1499                 cow_start = cur_offset;
1500                 cur_offset = end;
1501         }
1502
1503         if (cow_start != (u64)-1) {
1504                 ret = cow_file_range(inode, locked_page, cow_start, end,
1505                                      page_started, nr_written, 1);
1506                 if (ret)
1507                         goto error;
1508         }
1509
1510 error:
1511         err = btrfs_end_transaction(trans, root);
1512         if (!ret)
1513                 ret = err;
1514
1515         if (ret && cur_offset < end)
1516                 extent_clear_unlock_delalloc(inode, cur_offset, end,
1517                                              locked_page, EXTENT_LOCKED |
1518                                              EXTENT_DELALLOC | EXTENT_DEFRAG |
1519                                              EXTENT_DO_ACCOUNTING, PAGE_UNLOCK |
1520                                              PAGE_CLEAR_DIRTY |
1521                                              PAGE_SET_WRITEBACK |
1522                                              PAGE_END_WRITEBACK);
1523         btrfs_free_path(path);
1524         return ret;
1525 }
1526
1527 static inline int need_force_cow(struct inode *inode, u64 start, u64 end)
1528 {
1529
1530         if (!(BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW) &&
1531             !(BTRFS_I(inode)->flags & BTRFS_INODE_PREALLOC))
1532                 return 0;
1533
1534         /*
1535          * @defrag_bytes is a hint value, no spinlock held here,
1536          * if is not zero, it means the file is defragging.
1537          * Force cow if given extent needs to be defragged.
1538          */
1539         if (BTRFS_I(inode)->defrag_bytes &&
1540             test_range_bit(&BTRFS_I(inode)->io_tree, start, end,
1541                            EXTENT_DEFRAG, 0, NULL))
1542                 return 1;
1543
1544         return 0;
1545 }
1546
1547 /*
1548  * extent_io.c call back to do delayed allocation processing
1549  */
1550 static int run_delalloc_range(struct inode *inode, struct page *locked_page,
1551                               u64 start, u64 end, int *page_started,
1552                               unsigned long *nr_written)
1553 {
1554         int ret;
1555         int force_cow = need_force_cow(inode, start, end);
1556
1557         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW && !force_cow) {
1558                 ret = run_delalloc_nocow(inode, locked_page, start, end,
1559                                          page_started, 1, nr_written);
1560         } else if (BTRFS_I(inode)->flags & BTRFS_INODE_PREALLOC && !force_cow) {
1561                 ret = run_delalloc_nocow(inode, locked_page, start, end,
1562                                          page_started, 0, nr_written);
1563         } else if (!inode_need_compress(inode)) {
1564                 ret = cow_file_range(inode, locked_page, start, end,
1565                                       page_started, nr_written, 1);
1566         } else {
1567                 set_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
1568                         &BTRFS_I(inode)->runtime_flags);
1569                 ret = cow_file_range_async(inode, locked_page, start, end,
1570                                            page_started, nr_written);
1571         }
1572         return ret;
1573 }
1574
1575 static void btrfs_split_extent_hook(struct inode *inode,
1576                                     struct extent_state *orig, u64 split)
1577 {
1578         u64 size;
1579
1580         /* not delalloc, ignore it */
1581         if (!(orig->state & EXTENT_DELALLOC))
1582                 return;
1583
1584         size = orig->end - orig->start + 1;
1585         if (size > BTRFS_MAX_EXTENT_SIZE) {
1586                 u64 num_extents;
1587                 u64 new_size;
1588
1589                 /*
1590                  * See the explanation in btrfs_merge_extent_hook, the same
1591                  * applies here, just in reverse.
1592                  */
1593                 new_size = orig->end - split + 1;
1594                 num_extents = div64_u64(new_size + BTRFS_MAX_EXTENT_SIZE - 1,
1595                                         BTRFS_MAX_EXTENT_SIZE);
1596                 new_size = split - orig->start;
1597                 num_extents += div64_u64(new_size + BTRFS_MAX_EXTENT_SIZE - 1,
1598                                         BTRFS_MAX_EXTENT_SIZE);
1599                 if (div64_u64(size + BTRFS_MAX_EXTENT_SIZE - 1,
1600                               BTRFS_MAX_EXTENT_SIZE) >= num_extents)
1601                         return;
1602         }
1603
1604         spin_lock(&BTRFS_I(inode)->lock);
1605         BTRFS_I(inode)->outstanding_extents++;
1606         spin_unlock(&BTRFS_I(inode)->lock);
1607 }
1608
1609 /*
1610  * extent_io.c merge_extent_hook, used to track merged delayed allocation
1611  * extents so we can keep track of new extents that are just merged onto old
1612  * extents, such as when we are doing sequential writes, so we can properly
1613  * account for the metadata space we'll need.
1614  */
1615 static void btrfs_merge_extent_hook(struct inode *inode,
1616                                     struct extent_state *new,
1617                                     struct extent_state *other)
1618 {
1619         u64 new_size, old_size;
1620         u64 num_extents;
1621
1622         /* not delalloc, ignore it */
1623         if (!(other->state & EXTENT_DELALLOC))
1624                 return;
1625
1626         if (new->start > other->start)
1627                 new_size = new->end - other->start + 1;
1628         else
1629                 new_size = other->end - new->start + 1;
1630
1631         /* we're not bigger than the max, unreserve the space and go */
1632         if (new_size <= BTRFS_MAX_EXTENT_SIZE) {
1633                 spin_lock(&BTRFS_I(inode)->lock);
1634                 BTRFS_I(inode)->outstanding_extents--;
1635                 spin_unlock(&BTRFS_I(inode)->lock);
1636                 return;
1637         }
1638
1639         /*
1640          * We have to add up either side to figure out how many extents were
1641          * accounted for before we merged into one big extent.  If the number of
1642          * extents we accounted for is <= the amount we need for the new range
1643          * then we can return, otherwise drop.  Think of it like this
1644          *
1645          * [ 4k][MAX_SIZE]
1646          *
1647          * So we've grown the extent by a MAX_SIZE extent, this would mean we
1648          * need 2 outstanding extents, on one side we have 1 and the other side
1649          * we have 1 so they are == and we can return.  But in this case
1650          *
1651          * [MAX_SIZE+4k][MAX_SIZE+4k]
1652          *
1653          * Each range on their own accounts for 2 extents, but merged together
1654          * they are only 3 extents worth of accounting, so we need to drop in
1655          * this case.
1656          */
1657         old_size = other->end - other->start + 1;
1658         num_extents = div64_u64(old_size + BTRFS_MAX_EXTENT_SIZE - 1,
1659                                 BTRFS_MAX_EXTENT_SIZE);
1660         old_size = new->end - new->start + 1;
1661         num_extents += div64_u64(old_size + BTRFS_MAX_EXTENT_SIZE - 1,
1662                                  BTRFS_MAX_EXTENT_SIZE);
1663
1664         if (div64_u64(new_size + BTRFS_MAX_EXTENT_SIZE - 1,
1665                       BTRFS_MAX_EXTENT_SIZE) >= num_extents)
1666                 return;
1667
1668         spin_lock(&BTRFS_I(inode)->lock);
1669         BTRFS_I(inode)->outstanding_extents--;
1670         spin_unlock(&BTRFS_I(inode)->lock);
1671 }
1672
1673 static void btrfs_add_delalloc_inodes(struct btrfs_root *root,
1674                                       struct inode *inode)
1675 {
1676         spin_lock(&root->delalloc_lock);
1677         if (list_empty(&BTRFS_I(inode)->delalloc_inodes)) {
1678                 list_add_tail(&BTRFS_I(inode)->delalloc_inodes,
1679                               &root->delalloc_inodes);
1680                 set_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1681                         &BTRFS_I(inode)->runtime_flags);
1682                 root->nr_delalloc_inodes++;
1683                 if (root->nr_delalloc_inodes == 1) {
1684                         spin_lock(&root->fs_info->delalloc_root_lock);
1685                         BUG_ON(!list_empty(&root->delalloc_root));
1686                         list_add_tail(&root->delalloc_root,
1687                                       &root->fs_info->delalloc_roots);
1688                         spin_unlock(&root->fs_info->delalloc_root_lock);
1689                 }
1690         }
1691         spin_unlock(&root->delalloc_lock);
1692 }
1693
1694 static void btrfs_del_delalloc_inode(struct btrfs_root *root,
1695                                      struct inode *inode)
1696 {
1697         spin_lock(&root->delalloc_lock);
1698         if (!list_empty(&BTRFS_I(inode)->delalloc_inodes)) {
1699                 list_del_init(&BTRFS_I(inode)->delalloc_inodes);
1700                 clear_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1701                           &BTRFS_I(inode)->runtime_flags);
1702                 root->nr_delalloc_inodes--;
1703                 if (!root->nr_delalloc_inodes) {
1704                         spin_lock(&root->fs_info->delalloc_root_lock);
1705                         BUG_ON(list_empty(&root->delalloc_root));
1706                         list_del_init(&root->delalloc_root);
1707                         spin_unlock(&root->fs_info->delalloc_root_lock);
1708                 }
1709         }
1710         spin_unlock(&root->delalloc_lock);
1711 }
1712
1713 /*
1714  * extent_io.c set_bit_hook, used to track delayed allocation
1715  * bytes in this file, and to maintain the list of inodes that
1716  * have pending delalloc work to be done.
1717  */
1718 static void btrfs_set_bit_hook(struct inode *inode,
1719                                struct extent_state *state, unsigned *bits)
1720 {
1721
1722         if ((*bits & EXTENT_DEFRAG) && !(*bits & EXTENT_DELALLOC))
1723                 WARN_ON(1);
1724         /*
1725          * set_bit and clear bit hooks normally require _irqsave/restore
1726          * but in this case, we are only testing for the DELALLOC
1727          * bit, which is only set or cleared with irqs on
1728          */
1729         if (!(state->state & EXTENT_DELALLOC) && (*bits & EXTENT_DELALLOC)) {
1730                 struct btrfs_root *root = BTRFS_I(inode)->root;
1731                 u64 len = state->end + 1 - state->start;
1732                 bool do_list = !btrfs_is_free_space_inode(inode);
1733
1734                 if (*bits & EXTENT_FIRST_DELALLOC) {
1735                         *bits &= ~EXTENT_FIRST_DELALLOC;
1736                 } else {
1737                         spin_lock(&BTRFS_I(inode)->lock);
1738                         BTRFS_I(inode)->outstanding_extents++;
1739                         spin_unlock(&BTRFS_I(inode)->lock);
1740                 }
1741
1742                 /* For sanity tests */
1743                 if (btrfs_test_is_dummy_root(root))
1744                         return;
1745
1746                 __percpu_counter_add(&root->fs_info->delalloc_bytes, len,
1747                                      root->fs_info->delalloc_batch);
1748                 spin_lock(&BTRFS_I(inode)->lock);
1749                 BTRFS_I(inode)->delalloc_bytes += len;
1750                 if (*bits & EXTENT_DEFRAG)
1751                         BTRFS_I(inode)->defrag_bytes += len;
1752                 if (do_list && !test_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1753                                          &BTRFS_I(inode)->runtime_flags))
1754                         btrfs_add_delalloc_inodes(root, inode);
1755                 spin_unlock(&BTRFS_I(inode)->lock);
1756         }
1757 }
1758
1759 /*
1760  * extent_io.c clear_bit_hook, see set_bit_hook for why
1761  */
1762 static void btrfs_clear_bit_hook(struct inode *inode,
1763                                  struct extent_state *state,
1764                                  unsigned *bits)
1765 {
1766         u64 len = state->end + 1 - state->start;
1767         u64 num_extents = div64_u64(len + BTRFS_MAX_EXTENT_SIZE -1,
1768                                     BTRFS_MAX_EXTENT_SIZE);
1769
1770         spin_lock(&BTRFS_I(inode)->lock);
1771         if ((state->state & EXTENT_DEFRAG) && (*bits & EXTENT_DEFRAG))
1772                 BTRFS_I(inode)->defrag_bytes -= len;
1773         spin_unlock(&BTRFS_I(inode)->lock);
1774
1775         /*
1776          * set_bit and clear bit hooks normally require _irqsave/restore
1777          * but in this case, we are only testing for the DELALLOC
1778          * bit, which is only set or cleared with irqs on
1779          */
1780         if ((state->state & EXTENT_DELALLOC) && (*bits & EXTENT_DELALLOC)) {
1781                 struct btrfs_root *root = BTRFS_I(inode)->root;
1782                 bool do_list = !btrfs_is_free_space_inode(inode);
1783
1784                 if (*bits & EXTENT_FIRST_DELALLOC) {
1785                         *bits &= ~EXTENT_FIRST_DELALLOC;
1786                 } else if (!(*bits & EXTENT_DO_ACCOUNTING)) {
1787                         spin_lock(&BTRFS_I(inode)->lock);
1788                         BTRFS_I(inode)->outstanding_extents -= num_extents;
1789                         spin_unlock(&BTRFS_I(inode)->lock);
1790                 }
1791
1792                 /*
1793                  * We don't reserve metadata space for space cache inodes so we
1794                  * don't need to call dellalloc_release_metadata if there is an
1795                  * error.
1796                  */
1797                 if (*bits & EXTENT_DO_ACCOUNTING &&
1798                     root != root->fs_info->tree_root)
1799                         btrfs_delalloc_release_metadata(inode, len);
1800
1801                 /* For sanity tests. */
1802                 if (btrfs_test_is_dummy_root(root))
1803                         return;
1804
1805                 if (root->root_key.objectid != BTRFS_DATA_RELOC_TREE_OBJECTID
1806                     && do_list && !(state->state & EXTENT_NORESERVE))
1807                         btrfs_free_reserved_data_space_noquota(inode,
1808                                         state->start, len);
1809
1810                 __percpu_counter_add(&root->fs_info->delalloc_bytes, -len,
1811                                      root->fs_info->delalloc_batch);
1812                 spin_lock(&BTRFS_I(inode)->lock);
1813                 BTRFS_I(inode)->delalloc_bytes -= len;
1814                 if (do_list && BTRFS_I(inode)->delalloc_bytes == 0 &&
1815                     test_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1816                              &BTRFS_I(inode)->runtime_flags))
1817                         btrfs_del_delalloc_inode(root, inode);
1818                 spin_unlock(&BTRFS_I(inode)->lock);
1819         }
1820 }
1821
1822 /*
1823  * extent_io.c merge_bio_hook, this must check the chunk tree to make sure
1824  * we don't create bios that span stripes or chunks
1825  */
1826 int btrfs_merge_bio_hook(int rw, struct page *page, unsigned long offset,
1827                          size_t size, struct bio *bio,
1828                          unsigned long bio_flags)
1829 {
1830         struct btrfs_root *root = BTRFS_I(page->mapping->host)->root;
1831         u64 logical = (u64)bio->bi_iter.bi_sector << 9;
1832         u64 length = 0;
1833         u64 map_length;
1834         int ret;
1835
1836         if (bio_flags & EXTENT_BIO_COMPRESSED)
1837                 return 0;
1838
1839         length = bio->bi_iter.bi_size;
1840         map_length = length;
1841         ret = btrfs_map_block(root->fs_info, rw, logical,
1842                               &map_length, NULL, 0);
1843         /* Will always return 0 with map_multi == NULL */
1844         BUG_ON(ret < 0);
1845         if (map_length < length + size)
1846                 return 1;
1847         return 0;
1848 }
1849
1850 /*
1851  * in order to insert checksums into the metadata in large chunks,
1852  * we wait until bio submission time.   All the pages in the bio are
1853  * checksummed and sums are attached onto the ordered extent record.
1854  *
1855  * At IO completion time the cums attached on the ordered extent record
1856  * are inserted into the btree
1857  */
1858 static int __btrfs_submit_bio_start(struct inode *inode, int rw,
1859                                     struct bio *bio, int mirror_num,
1860                                     unsigned long bio_flags,
1861                                     u64 bio_offset)
1862 {
1863         struct btrfs_root *root = BTRFS_I(inode)->root;
1864         int ret = 0;
1865
1866         ret = btrfs_csum_one_bio(root, inode, bio, 0, 0);
1867         BUG_ON(ret); /* -ENOMEM */
1868         return 0;
1869 }
1870
1871 /*
1872  * in order to insert checksums into the metadata in large chunks,
1873  * we wait until bio submission time.   All the pages in the bio are
1874  * checksummed and sums are attached onto the ordered extent record.
1875  *
1876  * At IO completion time the cums attached on the ordered extent record
1877  * are inserted into the btree
1878  */
1879 static int __btrfs_submit_bio_done(struct inode *inode, int rw, struct bio *bio,
1880                           int mirror_num, unsigned long bio_flags,
1881                           u64 bio_offset)
1882 {
1883         struct btrfs_root *root = BTRFS_I(inode)->root;
1884         int ret;
1885
1886         ret = btrfs_map_bio(root, rw, bio, mirror_num, 1);
1887         if (ret) {
1888                 bio->bi_error = ret;
1889                 bio_endio(bio);
1890         }
1891         return ret;
1892 }
1893
1894 /*
1895  * extent_io.c submission hook. This does the right thing for csum calculation
1896  * on write, or reading the csums from the tree before a read
1897  */
1898 static int btrfs_submit_bio_hook(struct inode *inode, int rw, struct bio *bio,
1899                           int mirror_num, unsigned long bio_flags,
1900                           u64 bio_offset)
1901 {
1902         struct btrfs_root *root = BTRFS_I(inode)->root;
1903         enum btrfs_wq_endio_type metadata = BTRFS_WQ_ENDIO_DATA;
1904         int ret = 0;
1905         int skip_sum;
1906         int async = !atomic_read(&BTRFS_I(inode)->sync_writers);
1907
1908         skip_sum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM;
1909
1910         if (btrfs_is_free_space_inode(inode))
1911                 metadata = BTRFS_WQ_ENDIO_FREE_SPACE;
1912
1913         if (!(rw & REQ_WRITE)) {
1914                 ret = btrfs_bio_wq_end_io(root->fs_info, bio, metadata);
1915                 if (ret)
1916                         goto out;
1917
1918                 if (bio_flags & EXTENT_BIO_COMPRESSED) {
1919                         ret = btrfs_submit_compressed_read(inode, bio,
1920                                                            mirror_num,
1921                                                            bio_flags);
1922                         goto out;
1923                 } else if (!skip_sum) {
1924                         ret = btrfs_lookup_bio_sums(root, inode, bio, NULL);
1925                         if (ret)
1926                                 goto out;
1927                 }
1928                 goto mapit;
1929         } else if (async && !skip_sum) {
1930                 /* csum items have already been cloned */
1931                 if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
1932                         goto mapit;
1933                 /* we're doing a write, do the async checksumming */
1934                 ret = btrfs_wq_submit_bio(BTRFS_I(inode)->root->fs_info,
1935                                    inode, rw, bio, mirror_num,
1936                                    bio_flags, bio_offset,
1937                                    __btrfs_submit_bio_start,
1938                                    __btrfs_submit_bio_done);
1939                 goto out;
1940         } else if (!skip_sum) {
1941                 ret = btrfs_csum_one_bio(root, inode, bio, 0, 0);
1942                 if (ret)
1943                         goto out;
1944         }
1945
1946 mapit:
1947         ret = btrfs_map_bio(root, rw, bio, mirror_num, 0);
1948
1949 out:
1950         if (ret < 0) {
1951                 bio->bi_error = ret;
1952                 bio_endio(bio);
1953         }
1954         return ret;
1955 }
1956
1957 /*
1958  * given a list of ordered sums record them in the inode.  This happens
1959  * at IO completion time based on sums calculated at bio submission time.
1960  */
1961 static noinline int add_pending_csums(struct btrfs_trans_handle *trans,
1962                              struct inode *inode, u64 file_offset,
1963                              struct list_head *list)
1964 {
1965         struct btrfs_ordered_sum *sum;
1966
1967         list_for_each_entry(sum, list, list) {
1968                 trans->adding_csums = 1;
1969                 btrfs_csum_file_blocks(trans,
1970                        BTRFS_I(inode)->root->fs_info->csum_root, sum);
1971                 trans->adding_csums = 0;
1972         }
1973         return 0;
1974 }
1975
1976 int btrfs_set_extent_delalloc(struct inode *inode, u64 start, u64 end,
1977                               struct extent_state **cached_state)
1978 {
1979         WARN_ON((end & (PAGE_SIZE - 1)) == 0);
1980         return set_extent_delalloc(&BTRFS_I(inode)->io_tree, start, end,
1981                                    cached_state);
1982 }
1983
1984 /* see btrfs_writepage_start_hook for details on why this is required */
1985 struct btrfs_writepage_fixup {
1986         struct page *page;
1987         struct btrfs_work work;
1988 };
1989
1990 static void btrfs_writepage_fixup_worker(struct btrfs_work *work)
1991 {
1992         struct btrfs_writepage_fixup *fixup;
1993         struct btrfs_ordered_extent *ordered;
1994         struct extent_state *cached_state = NULL;
1995         struct page *page;
1996         struct inode *inode;
1997         u64 page_start;
1998         u64 page_end;
1999         int ret;
2000
2001         fixup = container_of(work, struct btrfs_writepage_fixup, work);
2002         page = fixup->page;
2003 again:
2004         lock_page(page);
2005         if (!page->mapping || !PageDirty(page) || !PageChecked(page)) {
2006                 ClearPageChecked(page);
2007                 goto out_page;
2008         }
2009
2010         inode = page->mapping->host;
2011         page_start = page_offset(page);
2012         page_end = page_offset(page) + PAGE_SIZE - 1;
2013
2014         lock_extent_bits(&BTRFS_I(inode)->io_tree, page_start, page_end,
2015                          &cached_state);
2016
2017         /* already ordered? We're done */
2018         if (PagePrivate2(page))
2019                 goto out;
2020
2021         ordered = btrfs_lookup_ordered_range(inode, page_start,
2022                                         PAGE_SIZE);
2023         if (ordered) {
2024                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, page_start,
2025                                      page_end, &cached_state, GFP_NOFS);
2026                 unlock_page(page);
2027                 btrfs_start_ordered_extent(inode, ordered, 1);
2028                 btrfs_put_ordered_extent(ordered);
2029                 goto again;
2030         }
2031
2032         ret = btrfs_delalloc_reserve_space(inode, page_start,
2033                                            PAGE_SIZE);
2034         if (ret) {
2035                 mapping_set_error(page->mapping, ret);
2036                 end_extent_writepage(page, ret, page_start, page_end);
2037                 ClearPageChecked(page);
2038                 goto out;
2039          }
2040
2041         btrfs_set_extent_delalloc(inode, page_start, page_end, &cached_state);
2042         ClearPageChecked(page);
2043         set_page_dirty(page);
2044 out:
2045         unlock_extent_cached(&BTRFS_I(inode)->io_tree, page_start, page_end,
2046                              &cached_state, GFP_NOFS);
2047 out_page:
2048         unlock_page(page);
2049         put_page(page);
2050         kfree(fixup);
2051 }
2052
2053 /*
2054  * There are a few paths in the higher layers of the kernel that directly
2055  * set the page dirty bit without asking the filesystem if it is a
2056  * good idea.  This causes problems because we want to make sure COW
2057  * properly happens and the data=ordered rules are followed.
2058  *
2059  * In our case any range that doesn't have the ORDERED bit set
2060  * hasn't been properly setup for IO.  We kick off an async process
2061  * to fix it up.  The async helper will wait for ordered extents, set
2062  * the delalloc bit and make it safe to write the page.
2063  */
2064 static int btrfs_writepage_start_hook(struct page *page, u64 start, u64 end)
2065 {
2066         struct inode *inode = page->mapping->host;
2067         struct btrfs_writepage_fixup *fixup;
2068         struct btrfs_root *root = BTRFS_I(inode)->root;
2069
2070         /* this page is properly in the ordered list */
2071         if (TestClearPagePrivate2(page))
2072                 return 0;
2073
2074         if (PageChecked(page))
2075                 return -EAGAIN;
2076
2077         fixup = kzalloc(sizeof(*fixup), GFP_NOFS);
2078         if (!fixup)
2079                 return -EAGAIN;
2080
2081         SetPageChecked(page);
2082         get_page(page);
2083         btrfs_init_work(&fixup->work, btrfs_fixup_helper,
2084                         btrfs_writepage_fixup_worker, NULL, NULL);
2085         fixup->page = page;
2086         btrfs_queue_work(root->fs_info->fixup_workers, &fixup->work);
2087         return -EBUSY;
2088 }
2089
2090 static int insert_reserved_file_extent(struct btrfs_trans_handle *trans,
2091                                        struct inode *inode, u64 file_pos,
2092                                        u64 disk_bytenr, u64 disk_num_bytes,
2093                                        u64 num_bytes, u64 ram_bytes,
2094                                        u8 compression, u8 encryption,
2095                                        u16 other_encoding, int extent_type)
2096 {
2097         struct btrfs_root *root = BTRFS_I(inode)->root;
2098         struct btrfs_file_extent_item *fi;
2099         struct btrfs_path *path;
2100         struct extent_buffer *leaf;
2101         struct btrfs_key ins;
2102         int extent_inserted = 0;
2103         int ret;
2104
2105         path = btrfs_alloc_path();
2106         if (!path)
2107                 return -ENOMEM;
2108
2109         /*
2110          * we may be replacing one extent in the tree with another.
2111          * The new extent is pinned in the extent map, and we don't want
2112          * to drop it from the cache until it is completely in the btree.
2113          *
2114          * So, tell btrfs_drop_extents to leave this extent in the cache.
2115          * the caller is expected to unpin it and allow it to be merged
2116          * with the others.
2117          */
2118         ret = __btrfs_drop_extents(trans, root, inode, path, file_pos,
2119                                    file_pos + num_bytes, NULL, 0,
2120                                    1, sizeof(*fi), &extent_inserted);
2121         if (ret)
2122                 goto out;
2123
2124         if (!extent_inserted) {
2125                 ins.objectid = btrfs_ino(inode);
2126                 ins.offset = file_pos;
2127                 ins.type = BTRFS_EXTENT_DATA_KEY;
2128
2129                 path->leave_spinning = 1;
2130                 ret = btrfs_insert_empty_item(trans, root, path, &ins,
2131                                               sizeof(*fi));
2132                 if (ret)
2133                         goto out;
2134         }
2135         leaf = path->nodes[0];
2136         fi = btrfs_item_ptr(leaf, path->slots[0],
2137                             struct btrfs_file_extent_item);
2138         btrfs_set_file_extent_generation(leaf, fi, trans->transid);
2139         btrfs_set_file_extent_type(leaf, fi, extent_type);
2140         btrfs_set_file_extent_disk_bytenr(leaf, fi, disk_bytenr);
2141         btrfs_set_file_extent_disk_num_bytes(leaf, fi, disk_num_bytes);
2142         btrfs_set_file_extent_offset(leaf, fi, 0);
2143         btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2144         btrfs_set_file_extent_ram_bytes(leaf, fi, ram_bytes);
2145         btrfs_set_file_extent_compression(leaf, fi, compression);
2146         btrfs_set_file_extent_encryption(leaf, fi, encryption);
2147         btrfs_set_file_extent_other_encoding(leaf, fi, other_encoding);
2148
2149         btrfs_mark_buffer_dirty(leaf);
2150         btrfs_release_path(path);
2151
2152         inode_add_bytes(inode, num_bytes);
2153
2154         ins.objectid = disk_bytenr;
2155         ins.offset = disk_num_bytes;
2156         ins.type = BTRFS_EXTENT_ITEM_KEY;
2157         ret = btrfs_alloc_reserved_file_extent(trans, root,
2158                                         root->root_key.objectid,
2159                                         btrfs_ino(inode), file_pos,
2160                                         ram_bytes, &ins);
2161         /*
2162          * Release the reserved range from inode dirty range map, as it is
2163          * already moved into delayed_ref_head
2164          */
2165         btrfs_qgroup_release_data(inode, file_pos, ram_bytes);
2166 out:
2167         btrfs_free_path(path);
2168
2169         return ret;
2170 }
2171
2172 /* snapshot-aware defrag */
2173 struct sa_defrag_extent_backref {
2174         struct rb_node node;
2175         struct old_sa_defrag_extent *old;
2176         u64 root_id;
2177         u64 inum;
2178         u64 file_pos;
2179         u64 extent_offset;
2180         u64 num_bytes;
2181         u64 generation;
2182 };
2183
2184 struct old_sa_defrag_extent {
2185         struct list_head list;
2186         struct new_sa_defrag_extent *new;
2187
2188         u64 extent_offset;
2189         u64 bytenr;
2190         u64 offset;
2191         u64 len;
2192         int count;
2193 };
2194
2195 struct new_sa_defrag_extent {
2196         struct rb_root root;
2197         struct list_head head;
2198         struct btrfs_path *path;
2199         struct inode *inode;
2200         u64 file_pos;
2201         u64 len;
2202         u64 bytenr;
2203         u64 disk_len;
2204         u8 compress_type;
2205 };
2206
2207 static int backref_comp(struct sa_defrag_extent_backref *b1,
2208                         struct sa_defrag_extent_backref *b2)
2209 {
2210         if (b1->root_id < b2->root_id)
2211                 return -1;
2212         else if (b1->root_id > b2->root_id)
2213                 return 1;
2214
2215         if (b1->inum < b2->inum)
2216                 return -1;
2217         else if (b1->inum > b2->inum)
2218                 return 1;
2219
2220         if (b1->file_pos < b2->file_pos)
2221                 return -1;
2222         else if (b1->file_pos > b2->file_pos)
2223                 return 1;
2224
2225         /*
2226          * [------------------------------] ===> (a range of space)
2227          *     |<--->|   |<---->| =============> (fs/file tree A)
2228          * |<---------------------------->| ===> (fs/file tree B)
2229          *
2230          * A range of space can refer to two file extents in one tree while
2231          * refer to only one file extent in another tree.
2232          *
2233          * So we may process a disk offset more than one time(two extents in A)
2234          * and locate at the same extent(one extent in B), then insert two same
2235          * backrefs(both refer to the extent in B).
2236          */
2237         return 0;
2238 }
2239
2240 static void backref_insert(struct rb_root *root,
2241                            struct sa_defrag_extent_backref *backref)
2242 {
2243         struct rb_node **p = &root->rb_node;
2244         struct rb_node *parent = NULL;
2245         struct sa_defrag_extent_backref *entry;
2246         int ret;
2247
2248         while (*p) {
2249                 parent = *p;
2250                 entry = rb_entry(parent, struct sa_defrag_extent_backref, node);
2251
2252                 ret = backref_comp(backref, entry);
2253                 if (ret < 0)
2254                         p = &(*p)->rb_left;
2255                 else
2256                         p = &(*p)->rb_right;
2257         }
2258
2259         rb_link_node(&backref->node, parent, p);
2260         rb_insert_color(&backref->node, root);
2261 }
2262
2263 /*
2264  * Note the backref might has changed, and in this case we just return 0.
2265  */
2266 static noinline int record_one_backref(u64 inum, u64 offset, u64 root_id,
2267                                        void *ctx)
2268 {
2269         struct btrfs_file_extent_item *extent;
2270         struct btrfs_fs_info *fs_info;
2271         struct old_sa_defrag_extent *old = ctx;
2272         struct new_sa_defrag_extent *new = old->new;
2273         struct btrfs_path *path = new->path;
2274         struct btrfs_key key;
2275         struct btrfs_root *root;
2276         struct sa_defrag_extent_backref *backref;
2277         struct extent_buffer *leaf;
2278         struct inode *inode = new->inode;
2279         int slot;
2280         int ret;
2281         u64 extent_offset;
2282         u64 num_bytes;
2283
2284         if (BTRFS_I(inode)->root->root_key.objectid == root_id &&
2285             inum == btrfs_ino(inode))
2286                 return 0;
2287
2288         key.objectid = root_id;
2289         key.type = BTRFS_ROOT_ITEM_KEY;
2290         key.offset = (u64)-1;
2291
2292         fs_info = BTRFS_I(inode)->root->fs_info;
2293         root = btrfs_read_fs_root_no_name(fs_info, &key);
2294         if (IS_ERR(root)) {
2295                 if (PTR_ERR(root) == -ENOENT)
2296                         return 0;
2297                 WARN_ON(1);
2298                 pr_debug("inum=%llu, offset=%llu, root_id=%llu\n",
2299                          inum, offset, root_id);
2300                 return PTR_ERR(root);
2301         }
2302
2303         key.objectid = inum;
2304         key.type = BTRFS_EXTENT_DATA_KEY;
2305         if (offset > (u64)-1 << 32)
2306                 key.offset = 0;
2307         else
2308                 key.offset = offset;
2309
2310         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2311         if (WARN_ON(ret < 0))
2312                 return ret;
2313         ret = 0;
2314
2315         while (1) {
2316                 cond_resched();
2317
2318                 leaf = path->nodes[0];
2319                 slot = path->slots[0];
2320
2321                 if (slot >= btrfs_header_nritems(leaf)) {
2322                         ret = btrfs_next_leaf(root, path);
2323                         if (ret < 0) {
2324                                 goto out;
2325                         } else if (ret > 0) {
2326                                 ret = 0;
2327                                 goto out;
2328                         }
2329                         continue;
2330                 }
2331
2332                 path->slots[0]++;
2333
2334                 btrfs_item_key_to_cpu(leaf, &key, slot);
2335
2336                 if (key.objectid > inum)
2337                         goto out;
2338
2339                 if (key.objectid < inum || key.type != BTRFS_EXTENT_DATA_KEY)
2340                         continue;
2341
2342                 extent = btrfs_item_ptr(leaf, slot,
2343                                         struct btrfs_file_extent_item);
2344
2345                 if (btrfs_file_extent_disk_bytenr(leaf, extent) != old->bytenr)
2346                         continue;
2347
2348                 /*
2349                  * 'offset' refers to the exact key.offset,
2350                  * NOT the 'offset' field in btrfs_extent_data_ref, ie.
2351                  * (key.offset - extent_offset).
2352                  */
2353                 if (key.offset != offset)
2354                         continue;
2355
2356                 extent_offset = btrfs_file_extent_offset(leaf, extent);
2357                 num_bytes = btrfs_file_extent_num_bytes(leaf, extent);
2358
2359                 if (extent_offset >= old->extent_offset + old->offset +
2360                     old->len || extent_offset + num_bytes <=
2361                     old->extent_offset + old->offset)
2362                         continue;
2363                 break;
2364         }
2365
2366         backref = kmalloc(sizeof(*backref), GFP_NOFS);
2367         if (!backref) {
2368                 ret = -ENOENT;
2369                 goto out;
2370         }
2371
2372         backref->root_id = root_id;
2373         backref->inum = inum;
2374         backref->file_pos = offset;
2375         backref->num_bytes = num_bytes;
2376         backref->extent_offset = extent_offset;
2377         backref->generation = btrfs_file_extent_generation(leaf, extent);
2378         backref->old = old;
2379         backref_insert(&new->root, backref);
2380         old->count++;
2381 out:
2382         btrfs_release_path(path);
2383         WARN_ON(ret);
2384         return ret;
2385 }
2386
2387 static noinline bool record_extent_backrefs(struct btrfs_path *path,
2388                                    struct new_sa_defrag_extent *new)
2389 {
2390         struct btrfs_fs_info *fs_info = BTRFS_I(new->inode)->root->fs_info;
2391         struct old_sa_defrag_extent *old, *tmp;
2392         int ret;
2393
2394         new->path = path;
2395
2396         list_for_each_entry_safe(old, tmp, &new->head, list) {
2397                 ret = iterate_inodes_from_logical(old->bytenr +
2398                                                   old->extent_offset, fs_info,
2399                                                   path, record_one_backref,
2400                                                   old);
2401                 if (ret < 0 && ret != -ENOENT)
2402                         return false;
2403
2404                 /* no backref to be processed for this extent */
2405                 if (!old->count) {
2406                         list_del(&old->list);
2407                         kfree(old);
2408                 }
2409         }
2410
2411         if (list_empty(&new->head))
2412                 return false;
2413
2414         return true;
2415 }
2416
2417 static int relink_is_mergable(struct extent_buffer *leaf,
2418                               struct btrfs_file_extent_item *fi,
2419                               struct new_sa_defrag_extent *new)
2420 {
2421         if (btrfs_file_extent_disk_bytenr(leaf, fi) != new->bytenr)
2422                 return 0;
2423
2424         if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG)
2425                 return 0;
2426
2427         if (btrfs_file_extent_compression(leaf, fi) != new->compress_type)
2428                 return 0;
2429
2430         if (btrfs_file_extent_encryption(leaf, fi) ||
2431             btrfs_file_extent_other_encoding(leaf, fi))
2432                 return 0;
2433
2434         return 1;
2435 }
2436
2437 /*
2438  * Note the backref might has changed, and in this case we just return 0.
2439  */
2440 static noinline int relink_extent_backref(struct btrfs_path *path,
2441                                  struct sa_defrag_extent_backref *prev,
2442                                  struct sa_defrag_extent_backref *backref)
2443 {
2444         struct btrfs_file_extent_item *extent;
2445         struct btrfs_file_extent_item *item;
2446         struct btrfs_ordered_extent *ordered;
2447         struct btrfs_trans_handle *trans;
2448         struct btrfs_fs_info *fs_info;
2449         struct btrfs_root *root;
2450         struct btrfs_key key;
2451         struct extent_buffer *leaf;
2452         struct old_sa_defrag_extent *old = backref->old;
2453         struct new_sa_defrag_extent *new = old->new;
2454         struct inode *src_inode = new->inode;
2455         struct inode *inode;
2456         struct extent_state *cached = NULL;
2457         int ret = 0;
2458         u64 start;
2459         u64 len;
2460         u64 lock_start;
2461         u64 lock_end;
2462         bool merge = false;
2463         int index;
2464
2465         if (prev && prev->root_id == backref->root_id &&
2466             prev->inum == backref->inum &&
2467             prev->file_pos + prev->num_bytes == backref->file_pos)
2468                 merge = true;
2469
2470         /* step 1: get root */
2471         key.objectid = backref->root_id;
2472         key.type = BTRFS_ROOT_ITEM_KEY;
2473         key.offset = (u64)-1;
2474
2475         fs_info = BTRFS_I(src_inode)->root->fs_info;
2476         index = srcu_read_lock(&fs_info->subvol_srcu);
2477
2478         root = btrfs_read_fs_root_no_name(fs_info, &key);
2479         if (IS_ERR(root)) {
2480                 srcu_read_unlock(&fs_info->subvol_srcu, index);
2481                 if (PTR_ERR(root) == -ENOENT)
2482                         return 0;
2483                 return PTR_ERR(root);
2484         }
2485
2486         if (btrfs_root_readonly(root)) {
2487                 srcu_read_unlock(&fs_info->subvol_srcu, index);
2488                 return 0;
2489         }
2490
2491         /* step 2: get inode */
2492         key.objectid = backref->inum;
2493         key.type = BTRFS_INODE_ITEM_KEY;
2494         key.offset = 0;
2495
2496         inode = btrfs_iget(fs_info->sb, &key, root, NULL);
2497         if (IS_ERR(inode)) {
2498                 srcu_read_unlock(&fs_info->subvol_srcu, index);
2499                 return 0;
2500         }
2501
2502         srcu_read_unlock(&fs_info->subvol_srcu, index);
2503
2504         /* step 3: relink backref */
2505         lock_start = backref->file_pos;
2506         lock_end = backref->file_pos + backref->num_bytes - 1;
2507         lock_extent_bits(&BTRFS_I(inode)->io_tree, lock_start, lock_end,
2508                          &cached);
2509
2510         ordered = btrfs_lookup_first_ordered_extent(inode, lock_end);
2511         if (ordered) {
2512                 btrfs_put_ordered_extent(ordered);
2513                 goto out_unlock;
2514         }
2515
2516         trans = btrfs_join_transaction(root);
2517         if (IS_ERR(trans)) {
2518                 ret = PTR_ERR(trans);
2519                 goto out_unlock;
2520         }
2521
2522         key.objectid = backref->inum;
2523         key.type = BTRFS_EXTENT_DATA_KEY;
2524         key.offset = backref->file_pos;
2525
2526         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2527         if (ret < 0) {
2528                 goto out_free_path;
2529         } else if (ret > 0) {
2530                 ret = 0;
2531                 goto out_free_path;
2532         }
2533
2534         extent = btrfs_item_ptr(path->nodes[0], path->slots[0],
2535                                 struct btrfs_file_extent_item);
2536
2537         if (btrfs_file_extent_generation(path->nodes[0], extent) !=
2538             backref->generation)
2539                 goto out_free_path;
2540
2541         btrfs_release_path(path);
2542
2543         start = backref->file_pos;
2544         if (backref->extent_offset < old->extent_offset + old->offset)
2545                 start += old->extent_offset + old->offset -
2546                          backref->extent_offset;
2547
2548         len = min(backref->extent_offset + backref->num_bytes,
2549                   old->extent_offset + old->offset + old->len);
2550         len -= max(backref->extent_offset, old->extent_offset + old->offset);
2551
2552         ret = btrfs_drop_extents(trans, root, inode, start,
2553                                  start + len, 1);
2554         if (ret)
2555                 goto out_free_path;
2556 again:
2557         key.objectid = btrfs_ino(inode);
2558         key.type = BTRFS_EXTENT_DATA_KEY;
2559         key.offset = start;
2560
2561         path->leave_spinning = 1;
2562         if (merge) {
2563                 struct btrfs_file_extent_item *fi;
2564                 u64 extent_len;
2565                 struct btrfs_key found_key;
2566
2567                 ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
2568                 if (ret < 0)
2569                         goto out_free_path;
2570
2571                 path->slots[0]--;
2572                 leaf = path->nodes[0];
2573                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
2574
2575                 fi = btrfs_item_ptr(leaf, path->slots[0],
2576                                     struct btrfs_file_extent_item);
2577                 extent_len = btrfs_file_extent_num_bytes(leaf, fi);
2578
2579                 if (extent_len + found_key.offset == start &&
2580                     relink_is_mergable(leaf, fi, new)) {
2581                         btrfs_set_file_extent_num_bytes(leaf, fi,
2582                                                         extent_len + len);
2583                         btrfs_mark_buffer_dirty(leaf);
2584                         inode_add_bytes(inode, len);
2585
2586                         ret = 1;
2587                         goto out_free_path;
2588                 } else {
2589                         merge = false;
2590                         btrfs_release_path(path);
2591                         goto again;
2592                 }
2593         }
2594
2595         ret = btrfs_insert_empty_item(trans, root, path, &key,
2596                                         sizeof(*extent));
2597         if (ret) {
2598                 btrfs_abort_transaction(trans, root, ret);
2599                 goto out_free_path;
2600         }
2601
2602         leaf = path->nodes[0];
2603         item = btrfs_item_ptr(leaf, path->slots[0],
2604                                 struct btrfs_file_extent_item);
2605         btrfs_set_file_extent_disk_bytenr(leaf, item, new->bytenr);
2606         btrfs_set_file_extent_disk_num_bytes(leaf, item, new->disk_len);
2607         btrfs_set_file_extent_offset(leaf, item, start - new->file_pos);
2608         btrfs_set_file_extent_num_bytes(leaf, item, len);
2609         btrfs_set_file_extent_ram_bytes(leaf, item, new->len);
2610         btrfs_set_file_extent_generation(leaf, item, trans->transid);
2611         btrfs_set_file_extent_type(leaf, item, BTRFS_FILE_EXTENT_REG);
2612         btrfs_set_file_extent_compression(leaf, item, new->compress_type);
2613         btrfs_set_file_extent_encryption(leaf, item, 0);
2614         btrfs_set_file_extent_other_encoding(leaf, item, 0);
2615
2616         btrfs_mark_buffer_dirty(leaf);
2617         inode_add_bytes(inode, len);
2618         btrfs_release_path(path);
2619
2620         ret = btrfs_inc_extent_ref(trans, root, new->bytenr,
2621                         new->disk_len, 0,
2622                         backref->root_id, backref->inum,
2623                         new->file_pos); /* start - extent_offset */
2624         if (ret) {
2625                 btrfs_abort_transaction(trans, root, ret);
2626                 goto out_free_path;
2627         }
2628
2629         ret = 1;
2630 out_free_path:
2631         btrfs_release_path(path);
2632         path->leave_spinning = 0;
2633         btrfs_end_transaction(trans, root);
2634 out_unlock:
2635         unlock_extent_cached(&BTRFS_I(inode)->io_tree, lock_start, lock_end,
2636                              &cached, GFP_NOFS);
2637         iput(inode);
2638         return ret;
2639 }
2640
2641 static void free_sa_defrag_extent(struct new_sa_defrag_extent *new)
2642 {
2643         struct old_sa_defrag_extent *old, *tmp;
2644
2645         if (!new)
2646                 return;
2647
2648         list_for_each_entry_safe(old, tmp, &new->head, list) {
2649                 kfree(old);
2650         }
2651         kfree(new);
2652 }
2653
2654 static void relink_file_extents(struct new_sa_defrag_extent *new)
2655 {
2656         struct btrfs_path *path;
2657         struct sa_defrag_extent_backref *backref;
2658         struct sa_defrag_extent_backref *prev = NULL;
2659         struct inode *inode;
2660         struct btrfs_root *root;
2661         struct rb_node *node;
2662         int ret;
2663
2664         inode = new->inode;
2665         root = BTRFS_I(inode)->root;
2666
2667         path = btrfs_alloc_path();
2668         if (!path)
2669                 return;
2670
2671         if (!record_extent_backrefs(path, new)) {
2672                 btrfs_free_path(path);
2673                 goto out;
2674         }
2675         btrfs_release_path(path);
2676
2677         while (1) {
2678                 node = rb_first(&new->root);
2679                 if (!node)
2680                         break;
2681                 rb_erase(node, &new->root);
2682
2683                 backref = rb_entry(node, struct sa_defrag_extent_backref, node);
2684
2685                 ret = relink_extent_backref(path, prev, backref);
2686                 WARN_ON(ret < 0);
2687
2688                 kfree(prev);
2689
2690                 if (ret == 1)
2691                         prev = backref;
2692                 else
2693                         prev = NULL;
2694                 cond_resched();
2695         }
2696         kfree(prev);
2697
2698         btrfs_free_path(path);
2699 out:
2700         free_sa_defrag_extent(new);
2701
2702         atomic_dec(&root->fs_info->defrag_running);
2703         wake_up(&root->fs_info->transaction_wait);
2704 }
2705
2706 static struct new_sa_defrag_extent *
2707 record_old_file_extents(struct inode *inode,
2708                         struct btrfs_ordered_extent *ordered)
2709 {
2710         struct btrfs_root *root = BTRFS_I(inode)->root;
2711         struct btrfs_path *path;
2712         struct btrfs_key key;
2713         struct old_sa_defrag_extent *old;
2714         struct new_sa_defrag_extent *new;
2715         int ret;
2716
2717         new = kmalloc(sizeof(*new), GFP_NOFS);
2718         if (!new)
2719                 return NULL;
2720
2721         new->inode = inode;
2722         new->file_pos = ordered->file_offset;
2723         new->len = ordered->len;
2724         new->bytenr = ordered->start;
2725         new->disk_len = ordered->disk_len;
2726         new->compress_type = ordered->compress_type;
2727         new->root = RB_ROOT;
2728         INIT_LIST_HEAD(&new->head);
2729
2730         path = btrfs_alloc_path();
2731         if (!path)
2732                 goto out_kfree;
2733
2734         key.objectid = btrfs_ino(inode);
2735         key.type = BTRFS_EXTENT_DATA_KEY;
2736         key.offset = new->file_pos;
2737
2738         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2739         if (ret < 0)
2740                 goto out_free_path;
2741         if (ret > 0 && path->slots[0] > 0)
2742                 path->slots[0]--;
2743
2744         /* find out all the old extents for the file range */
2745         while (1) {
2746                 struct btrfs_file_extent_item *extent;
2747                 struct extent_buffer *l;
2748                 int slot;
2749                 u64 num_bytes;
2750                 u64 offset;
2751                 u64 end;
2752                 u64 disk_bytenr;
2753                 u64 extent_offset;
2754
2755                 l = path->nodes[0];
2756                 slot = path->slots[0];
2757
2758                 if (slot >= btrfs_header_nritems(l)) {
2759                         ret = btrfs_next_leaf(root, path);
2760                         if (ret < 0)
2761                                 goto out_free_path;
2762                         else if (ret > 0)
2763                                 break;
2764                         continue;
2765                 }
2766
2767                 btrfs_item_key_to_cpu(l, &key, slot);
2768
2769                 if (key.objectid != btrfs_ino(inode))
2770                         break;
2771                 if (key.type != BTRFS_EXTENT_DATA_KEY)
2772                         break;
2773                 if (key.offset >= new->file_pos + new->len)
2774                         break;
2775
2776                 extent = btrfs_item_ptr(l, slot, struct btrfs_file_extent_item);
2777
2778                 num_bytes = btrfs_file_extent_num_bytes(l, extent);
2779                 if (key.offset + num_bytes < new->file_pos)
2780                         goto next;
2781
2782                 disk_bytenr = btrfs_file_extent_disk_bytenr(l, extent);
2783                 if (!disk_bytenr)
2784                         goto next;
2785
2786                 extent_offset = btrfs_file_extent_offset(l, extent);
2787
2788                 old = kmalloc(sizeof(*old), GFP_NOFS);
2789                 if (!old)
2790                         goto out_free_path;
2791
2792                 offset = max(new->file_pos, key.offset);
2793                 end = min(new->file_pos + new->len, key.offset + num_bytes);
2794
2795                 old->bytenr = disk_bytenr;
2796                 old->extent_offset = extent_offset;
2797                 old->offset = offset - key.offset;
2798                 old->len = end - offset;
2799                 old->new = new;
2800                 old->count = 0;
2801                 list_add_tail(&old->list, &new->head);
2802 next:
2803                 path->slots[0]++;
2804                 cond_resched();
2805         }
2806
2807         btrfs_free_path(path);
2808         atomic_inc(&root->fs_info->defrag_running);
2809
2810         return new;
2811
2812 out_free_path:
2813         btrfs_free_path(path);
2814 out_kfree:
2815         free_sa_defrag_extent(new);
2816         return NULL;
2817 }
2818
2819 static void btrfs_release_delalloc_bytes(struct btrfs_root *root,
2820                                          u64 start, u64 len)
2821 {
2822         struct btrfs_block_group_cache *cache;
2823
2824         cache = btrfs_lookup_block_group(root->fs_info, start);
2825         ASSERT(cache);
2826
2827         spin_lock(&cache->lock);
2828         cache->delalloc_bytes -= len;
2829         spin_unlock(&cache->lock);
2830
2831         btrfs_put_block_group(cache);
2832 }
2833
2834 /* as ordered data IO finishes, this gets called so we can finish
2835  * an ordered extent if the range of bytes in the file it covers are
2836  * fully written.
2837  */
2838 static int btrfs_finish_ordered_io(struct btrfs_ordered_extent *ordered_extent)
2839 {
2840         struct inode *inode = ordered_extent->inode;
2841         struct btrfs_root *root = BTRFS_I(inode)->root;
2842         struct btrfs_trans_handle *trans = NULL;
2843         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
2844         struct extent_state *cached_state = NULL;
2845         struct new_sa_defrag_extent *new = NULL;
2846         int compress_type = 0;
2847         int ret = 0;
2848         u64 logical_len = ordered_extent->len;
2849         bool nolock;
2850         bool truncated = false;
2851
2852         nolock = btrfs_is_free_space_inode(inode);
2853
2854         if (test_bit(BTRFS_ORDERED_IOERR, &ordered_extent->flags)) {
2855                 ret = -EIO;
2856                 goto out;
2857         }
2858
2859         btrfs_free_io_failure_record(inode, ordered_extent->file_offset,
2860                                      ordered_extent->file_offset +
2861                                      ordered_extent->len - 1);
2862
2863         if (test_bit(BTRFS_ORDERED_TRUNCATED, &ordered_extent->flags)) {
2864                 truncated = true;
2865                 logical_len = ordered_extent->truncated_len;
2866                 /* Truncated the entire extent, don't bother adding */
2867                 if (!logical_len)
2868                         goto out;
2869         }
2870
2871         if (test_bit(BTRFS_ORDERED_NOCOW, &ordered_extent->flags)) {
2872                 BUG_ON(!list_empty(&ordered_extent->list)); /* Logic error */
2873
2874                 /*
2875                  * For mwrite(mmap + memset to write) case, we still reserve
2876                  * space for NOCOW range.
2877                  * As NOCOW won't cause a new delayed ref, just free the space
2878                  */
2879                 btrfs_qgroup_free_data(inode, ordered_extent->file_offset,
2880                                        ordered_extent->len);
2881                 btrfs_ordered_update_i_size(inode, 0, ordered_extent);
2882                 if (nolock)
2883                         trans = btrfs_join_transaction_nolock(root);
2884                 else
2885                         trans = btrfs_join_transaction(root);
2886                 if (IS_ERR(trans)) {
2887                         ret = PTR_ERR(trans);
2888                         trans = NULL;
2889                         goto out;
2890                 }
2891                 trans->block_rsv = &root->fs_info->delalloc_block_rsv;
2892                 ret = btrfs_update_inode_fallback(trans, root, inode);
2893                 if (ret) /* -ENOMEM or corruption */
2894                         btrfs_abort_transaction(trans, root, ret);
2895                 goto out;
2896         }
2897
2898         lock_extent_bits(io_tree, ordered_extent->file_offset,
2899                          ordered_extent->file_offset + ordered_extent->len - 1,
2900                          &cached_state);
2901
2902         ret = test_range_bit(io_tree, ordered_extent->file_offset,
2903                         ordered_extent->file_offset + ordered_extent->len - 1,
2904                         EXTENT_DEFRAG, 1, cached_state);
2905         if (ret) {
2906                 u64 last_snapshot = btrfs_root_last_snapshot(&root->root_item);
2907                 if (0 && last_snapshot >= BTRFS_I(inode)->generation)
2908                         /* the inode is shared */
2909                         new = record_old_file_extents(inode, ordered_extent);
2910
2911                 clear_extent_bit(io_tree, ordered_extent->file_offset,
2912                         ordered_extent->file_offset + ordered_extent->len - 1,
2913                         EXTENT_DEFRAG, 0, 0, &cached_state, GFP_NOFS);
2914         }
2915
2916         if (nolock)
2917                 trans = btrfs_join_transaction_nolock(root);
2918         else
2919                 trans = btrfs_join_transaction(root);
2920         if (IS_ERR(trans)) {
2921                 ret = PTR_ERR(trans);
2922                 trans = NULL;
2923                 goto out_unlock;
2924         }
2925
2926         trans->block_rsv = &root->fs_info->delalloc_block_rsv;
2927
2928         if (test_bit(BTRFS_ORDERED_COMPRESSED, &ordered_extent->flags))
2929                 compress_type = ordered_extent->compress_type;
2930         if (test_bit(BTRFS_ORDERED_PREALLOC, &ordered_extent->flags)) {
2931                 BUG_ON(compress_type);
2932                 ret = btrfs_mark_extent_written(trans, inode,
2933                                                 ordered_extent->file_offset,
2934                                                 ordered_extent->file_offset +
2935                                                 logical_len);
2936         } else {
2937                 BUG_ON(root == root->fs_info->tree_root);
2938                 ret = insert_reserved_file_extent(trans, inode,
2939                                                 ordered_extent->file_offset,
2940                                                 ordered_extent->start,
2941                                                 ordered_extent->disk_len,
2942                                                 logical_len, logical_len,
2943                                                 compress_type, 0, 0,
2944                                                 BTRFS_FILE_EXTENT_REG);
2945                 if (!ret)
2946                         btrfs_release_delalloc_bytes(root,
2947                                                      ordered_extent->start,
2948                                                      ordered_extent->disk_len);
2949         }
2950         unpin_extent_cache(&BTRFS_I(inode)->extent_tree,
2951                            ordered_extent->file_offset, ordered_extent->len,
2952                            trans->transid);
2953         if (ret < 0) {
2954                 btrfs_abort_transaction(trans, root, ret);
2955                 goto out_unlock;
2956         }
2957
2958         add_pending_csums(trans, inode, ordered_extent->file_offset,
2959                           &ordered_extent->list);
2960
2961         btrfs_ordered_update_i_size(inode, 0, ordered_extent);
2962         ret = btrfs_update_inode_fallback(trans, root, inode);
2963         if (ret) { /* -ENOMEM or corruption */
2964                 btrfs_abort_transaction(trans, root, ret);
2965                 goto out_unlock;
2966         }
2967         ret = 0;
2968 out_unlock:
2969         unlock_extent_cached(io_tree, ordered_extent->file_offset,
2970                              ordered_extent->file_offset +
2971                              ordered_extent->len - 1, &cached_state, GFP_NOFS);
2972 out:
2973         if (root != root->fs_info->tree_root)
2974                 btrfs_delalloc_release_metadata(inode, ordered_extent->len);
2975         if (trans)
2976                 btrfs_end_transaction(trans, root);
2977
2978         if (ret || truncated) {
2979                 u64 start, end;
2980
2981                 if (truncated)
2982                         start = ordered_extent->file_offset + logical_len;
2983                 else
2984                         start = ordered_extent->file_offset;
2985                 end = ordered_extent->file_offset + ordered_extent->len - 1;
2986                 clear_extent_uptodate(io_tree, start, end, NULL, GFP_NOFS);
2987
2988                 /* Drop the cache for the part of the extent we didn't write. */
2989                 btrfs_drop_extent_cache(inode, start, end, 0);
2990
2991                 /*
2992                  * If the ordered extent had an IOERR or something else went
2993                  * wrong we need to return the space for this ordered extent
2994                  * back to the allocator.  We only free the extent in the
2995                  * truncated case if we didn't write out the extent at all.
2996                  */
2997                 if ((ret || !logical_len) &&
2998                     !test_bit(BTRFS_ORDERED_NOCOW, &ordered_extent->flags) &&
2999                     !test_bit(BTRFS_ORDERED_PREALLOC, &ordered_extent->flags))
3000                         btrfs_free_reserved_extent(root, ordered_extent->start,
3001                                                    ordered_extent->disk_len, 1);
3002         }
3003
3004
3005         /*
3006          * This needs to be done to make sure anybody waiting knows we are done
3007          * updating everything for this ordered extent.
3008          */
3009         btrfs_remove_ordered_extent(inode, ordered_extent);
3010
3011         /* for snapshot-aware defrag */
3012         if (new) {
3013                 if (ret) {
3014                         free_sa_defrag_extent(new);
3015                         atomic_dec(&root->fs_info->defrag_running);
3016                 } else {
3017                         relink_file_extents(new);
3018                 }
3019         }
3020
3021         /* once for us */
3022         btrfs_put_ordered_extent(ordered_extent);
3023         /* once for the tree */
3024         btrfs_put_ordered_extent(ordered_extent);
3025
3026         return ret;
3027 }
3028
3029 static void finish_ordered_fn(struct btrfs_work *work)
3030 {
3031         struct btrfs_ordered_extent *ordered_extent;
3032         ordered_extent = container_of(work, struct btrfs_ordered_extent, work);
3033         btrfs_finish_ordered_io(ordered_extent);
3034 }
3035
3036 static int btrfs_writepage_end_io_hook(struct page *page, u64 start, u64 end,
3037                                 struct extent_state *state, int uptodate)
3038 {
3039         struct inode *inode = page->mapping->host;
3040         struct btrfs_root *root = BTRFS_I(inode)->root;
3041         struct btrfs_ordered_extent *ordered_extent = NULL;
3042         struct btrfs_workqueue *wq;
3043         btrfs_work_func_t func;
3044
3045         trace_btrfs_writepage_end_io_hook(page, start, end, uptodate);
3046
3047         ClearPagePrivate2(page);
3048         if (!btrfs_dec_test_ordered_pending(inode, &ordered_extent, start,
3049                                             end - start + 1, uptodate))
3050                 return 0;
3051
3052         if (btrfs_is_free_space_inode(inode)) {
3053                 wq = root->fs_info->endio_freespace_worker;
3054                 func = btrfs_freespace_write_helper;
3055         } else {
3056                 wq = root->fs_info->endio_write_workers;
3057                 func = btrfs_endio_write_helper;
3058         }
3059
3060         btrfs_init_work(&ordered_extent->work, func, finish_ordered_fn, NULL,
3061                         NULL);
3062         btrfs_queue_work(wq, &ordered_extent->work);
3063
3064         return 0;
3065 }
3066
3067 static int __readpage_endio_check(struct inode *inode,
3068                                   struct btrfs_io_bio *io_bio,
3069                                   int icsum, struct page *page,
3070                                   int pgoff, u64 start, size_t len)
3071 {
3072         char *kaddr;
3073         u32 csum_expected;
3074         u32 csum = ~(u32)0;
3075
3076         csum_expected = *(((u32 *)io_bio->csum) + icsum);
3077
3078         kaddr = kmap_atomic(page);
3079         csum = btrfs_csum_data(kaddr + pgoff, csum,  len);
3080         btrfs_csum_final(csum, (char *)&csum);
3081         if (csum != csum_expected)
3082                 goto zeroit;
3083
3084         kunmap_atomic(kaddr);
3085         return 0;
3086 zeroit:
3087         btrfs_warn_rl(BTRFS_I(inode)->root->fs_info,
3088                 "csum failed ino %llu off %llu csum %u expected csum %u",
3089                            btrfs_ino(inode), start, csum, csum_expected);
3090         memset(kaddr + pgoff, 1, len);
3091         flush_dcache_page(page);
3092         kunmap_atomic(kaddr);
3093         if (csum_expected == 0)
3094                 return 0;
3095         return -EIO;
3096 }
3097
3098 /*
3099  * when reads are done, we need to check csums to verify the data is correct
3100  * if there's a match, we allow the bio to finish.  If not, the code in
3101  * extent_io.c will try to find good copies for us.
3102  */
3103 static int btrfs_readpage_end_io_hook(struct btrfs_io_bio *io_bio,
3104                                       u64 phy_offset, struct page *page,
3105                                       u64 start, u64 end, int mirror)
3106 {
3107         size_t offset = start - page_offset(page);
3108         struct inode *inode = page->mapping->host;
3109         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
3110         struct btrfs_root *root = BTRFS_I(inode)->root;
3111
3112         if (PageChecked(page)) {
3113                 ClearPageChecked(page);
3114                 return 0;
3115         }
3116
3117         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)
3118                 return 0;
3119
3120         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID &&
3121             test_range_bit(io_tree, start, end, EXTENT_NODATASUM, 1, NULL)) {
3122                 clear_extent_bits(io_tree, start, end, EXTENT_NODATASUM);
3123                 return 0;
3124         }
3125
3126         phy_offset >>= inode->i_sb->s_blocksize_bits;
3127         return __readpage_endio_check(inode, io_bio, phy_offset, page, offset,
3128                                       start, (size_t)(end - start + 1));
3129 }
3130
3131 void btrfs_add_delayed_iput(struct inode *inode)
3132 {
3133         struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
3134         struct btrfs_inode *binode = BTRFS_I(inode);
3135
3136         if (atomic_add_unless(&inode->i_count, -1, 1))
3137                 return;
3138
3139         spin_lock(&fs_info->delayed_iput_lock);
3140         if (binode->delayed_iput_count == 0) {
3141                 ASSERT(list_empty(&binode->delayed_iput));
3142                 list_add_tail(&binode->delayed_iput, &fs_info->delayed_iputs);
3143         } else {
3144                 binode->delayed_iput_count++;
3145         }
3146         spin_unlock(&fs_info->delayed_iput_lock);
3147 }
3148
3149 void btrfs_run_delayed_iputs(struct btrfs_root *root)
3150 {
3151         struct btrfs_fs_info *fs_info = root->fs_info;
3152
3153         spin_lock(&fs_info->delayed_iput_lock);
3154         while (!list_empty(&fs_info->delayed_iputs)) {
3155                 struct btrfs_inode *inode;
3156
3157                 inode = list_first_entry(&fs_info->delayed_iputs,
3158                                 struct btrfs_inode, delayed_iput);
3159                 if (inode->delayed_iput_count) {
3160                         inode->delayed_iput_count--;
3161                         list_move_tail(&inode->delayed_iput,
3162                                         &fs_info->delayed_iputs);
3163                 } else {
3164                         list_del_init(&inode->delayed_iput);
3165                 }
3166                 spin_unlock(&fs_info->delayed_iput_lock);
3167                 iput(&inode->vfs_inode);
3168                 spin_lock(&fs_info->delayed_iput_lock);
3169         }
3170         spin_unlock(&fs_info->delayed_iput_lock);
3171 }
3172
3173 /*
3174  * This is called in transaction commit time. If there are no orphan
3175  * files in the subvolume, it removes orphan item and frees block_rsv
3176  * structure.
3177  */
3178 void btrfs_orphan_commit_root(struct btrfs_trans_handle *trans,
3179                               struct btrfs_root *root)
3180 {
3181         struct btrfs_block_rsv *block_rsv;
3182         int ret;
3183
3184         if (atomic_read(&root->orphan_inodes) ||
3185             root->orphan_cleanup_state != ORPHAN_CLEANUP_DONE)
3186                 return;
3187
3188         spin_lock(&root->orphan_lock);
3189         if (atomic_read(&root->orphan_inodes)) {
3190                 spin_unlock(&root->orphan_lock);
3191                 return;
3192         }
3193
3194         if (root->orphan_cleanup_state != ORPHAN_CLEANUP_DONE) {
3195                 spin_unlock(&root->orphan_lock);
3196                 return;
3197         }
3198
3199         block_rsv = root->orphan_block_rsv;
3200         root->orphan_block_rsv = NULL;
3201         spin_unlock(&root->orphan_lock);
3202
3203         if (test_bit(BTRFS_ROOT_ORPHAN_ITEM_INSERTED, &root->state) &&
3204             btrfs_root_refs(&root->root_item) > 0) {
3205                 ret = btrfs_del_orphan_item(trans, root->fs_info->tree_root,
3206                                             root->root_key.objectid);
3207                 if (ret)
3208                         btrfs_abort_transaction(trans, root, ret);
3209                 else
3210                         clear_bit(BTRFS_ROOT_ORPHAN_ITEM_INSERTED,
3211                                   &root->state);
3212         }
3213
3214         if (block_rsv) {
3215                 WARN_ON(block_rsv->size > 0);
3216                 btrfs_free_block_rsv(root, block_rsv);
3217         }
3218 }
3219
3220 /*
3221  * This creates an orphan entry for the given inode in case something goes
3222  * wrong in the middle of an unlink/truncate.
3223  *
3224  * NOTE: caller of this function should reserve 5 units of metadata for
3225  *       this function.
3226  */
3227 int btrfs_orphan_add(struct btrfs_trans_handle *trans, struct inode *inode)
3228 {
3229         struct btrfs_root *root = BTRFS_I(inode)->root;
3230         struct btrfs_block_rsv *block_rsv = NULL;
3231         int reserve = 0;
3232         int insert = 0;
3233         int ret;
3234
3235         if (!root->orphan_block_rsv) {
3236                 block_rsv = btrfs_alloc_block_rsv(root, BTRFS_BLOCK_RSV_TEMP);
3237                 if (!block_rsv)
3238                         return -ENOMEM;
3239         }
3240
3241         spin_lock(&root->orphan_lock);
3242         if (!root->orphan_block_rsv) {
3243                 root->orphan_block_rsv = block_rsv;
3244         } else if (block_rsv) {
3245                 btrfs_free_block_rsv(root, block_rsv);
3246                 block_rsv = NULL;
3247         }
3248
3249         if (!test_and_set_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
3250                               &BTRFS_I(inode)->runtime_flags)) {
3251 #if 0
3252                 /*
3253                  * For proper ENOSPC handling, we should do orphan
3254                  * cleanup when mounting. But this introduces backward
3255                  * compatibility issue.
3256                  */
3257                 if (!xchg(&root->orphan_item_inserted, 1))
3258                         insert = 2;
3259                 else
3260                         insert = 1;
3261 #endif
3262                 insert = 1;
3263                 atomic_inc(&root->orphan_inodes);
3264         }
3265
3266         if (!test_and_set_bit(BTRFS_INODE_ORPHAN_META_RESERVED,
3267                               &BTRFS_I(inode)->runtime_flags))
3268                 reserve = 1;
3269         spin_unlock(&root->orphan_lock);
3270
3271         /* grab metadata reservation from transaction handle */
3272         if (reserve) {
3273                 ret = btrfs_orphan_reserve_metadata(trans, inode);
3274                 ASSERT(!ret);
3275                 if (ret) {
3276                         atomic_dec(&root->orphan_inodes);
3277                         clear_bit(BTRFS_INODE_ORPHAN_META_RESERVED,
3278                                   &BTRFS_I(inode)->runtime_flags);
3279                         if (insert)
3280                                 clear_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
3281                                           &BTRFS_I(inode)->runtime_flags);
3282                         return ret;
3283                 }
3284         }
3285
3286         /* insert an orphan item to track this unlinked/truncated file */
3287         if (insert >= 1) {
3288                 ret = btrfs_insert_orphan_item(trans, root, btrfs_ino(inode));
3289                 if (ret) {
3290                         atomic_dec(&root->orphan_inodes);
3291                         if (reserve) {
3292                                 clear_bit(BTRFS_INODE_ORPHAN_META_RESERVED,
3293                                           &BTRFS_I(inode)->runtime_flags);
3294                                 btrfs_orphan_release_metadata(inode);
3295                         }
3296                         if (ret != -EEXIST) {
3297                                 clear_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
3298                                           &BTRFS_I(inode)->runtime_flags);
3299                                 btrfs_abort_transaction(trans, root, ret);
3300                                 return ret;
3301                         }
3302                 }
3303                 ret = 0;
3304         }
3305
3306         /* insert an orphan item to track subvolume contains orphan files */
3307         if (insert >= 2) {
3308                 ret = btrfs_insert_orphan_item(trans, root->fs_info->tree_root,
3309                                                root->root_key.objectid);
3310                 if (ret && ret != -EEXIST) {
3311                         btrfs_abort_transaction(trans, root, ret);
3312                         return ret;
3313                 }
3314         }
3315         return 0;
3316 }
3317
3318 /*
3319  * We have done the truncate/delete so we can go ahead and remove the orphan
3320  * item for this particular inode.
3321  */
3322 static int btrfs_orphan_del(struct btrfs_trans_handle *trans,
3323                             struct inode *inode)
3324 {
3325         struct btrfs_root *root = BTRFS_I(inode)->root;
3326         int delete_item = 0;
3327         int release_rsv = 0;
3328         int ret = 0;
3329
3330         spin_lock(&root->orphan_lock);
3331         if (test_and_clear_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
3332                                &BTRFS_I(inode)->runtime_flags))
3333                 delete_item = 1;
3334
3335         if (test_and_clear_bit(BTRFS_INODE_ORPHAN_META_RESERVED,
3336                                &BTRFS_I(inode)->runtime_flags))
3337                 release_rsv = 1;
3338         spin_unlock(&root->orphan_lock);
3339
3340         if (delete_item) {
3341                 atomic_dec(&root->orphan_inodes);
3342                 if (trans)
3343                         ret = btrfs_del_orphan_item(trans, root,
3344                                                     btrfs_ino(inode));
3345         }
3346
3347         if (release_rsv)
3348                 btrfs_orphan_release_metadata(inode);
3349
3350         return ret;
3351 }
3352
3353 /*
3354  * this cleans up any orphans that may be left on the list from the last use
3355  * of this root.
3356  */
3357 int btrfs_orphan_cleanup(struct btrfs_root *root)
3358 {
3359         struct btrfs_path *path;
3360         struct extent_buffer *leaf;
3361         struct btrfs_key key, found_key;
3362         struct btrfs_trans_handle *trans;
3363         struct inode *inode;
3364         u64 last_objectid = 0;
3365         int ret = 0, nr_unlink = 0, nr_truncate = 0;
3366
3367         if (cmpxchg(&root->orphan_cleanup_state, 0, ORPHAN_CLEANUP_STARTED))
3368                 return 0;
3369
3370         path = btrfs_alloc_path();
3371         if (!path) {
3372                 ret = -ENOMEM;
3373                 goto out;
3374         }
3375         path->reada = READA_BACK;
3376
3377         key.objectid = BTRFS_ORPHAN_OBJECTID;
3378         key.type = BTRFS_ORPHAN_ITEM_KEY;
3379         key.offset = (u64)-1;
3380
3381         while (1) {
3382                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3383                 if (ret < 0)
3384                         goto out;
3385
3386                 /*
3387                  * if ret == 0 means we found what we were searching for, which
3388                  * is weird, but possible, so only screw with path if we didn't
3389                  * find the key and see if we have stuff that matches
3390                  */
3391                 if (ret > 0) {
3392                         ret = 0;
3393                         if (path->slots[0] == 0)
3394                                 break;
3395                         path->slots[0]--;
3396                 }
3397
3398                 /* pull out the item */
3399                 leaf = path->nodes[0];
3400                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
3401
3402                 /* make sure the item matches what we want */
3403                 if (found_key.objectid != BTRFS_ORPHAN_OBJECTID)
3404                         break;
3405                 if (found_key.type != BTRFS_ORPHAN_ITEM_KEY)
3406                         break;
3407
3408                 /* release the path since we're done with it */
3409                 btrfs_release_path(path);
3410
3411                 /*
3412                  * this is where we are basically btrfs_lookup, without the
3413                  * crossing root thing.  we store the inode number in the
3414                  * offset of the orphan item.
3415                  */
3416
3417                 if (found_key.offset == last_objectid) {
3418                         btrfs_err(root->fs_info,
3419                                 "Error removing orphan entry, stopping orphan cleanup");
3420                         ret = -EINVAL;
3421                         goto out;
3422                 }
3423
3424                 last_objectid = found_key.offset;
3425
3426                 found_key.objectid = found_key.offset;
3427                 found_key.type = BTRFS_INODE_ITEM_KEY;
3428                 found_key.offset = 0;
3429                 inode = btrfs_iget(root->fs_info->sb, &found_key, root, NULL);
3430                 ret = PTR_ERR_OR_ZERO(inode);
3431                 if (ret && ret != -ENOENT)
3432                         goto out;
3433
3434                 if (ret == -ENOENT && root == root->fs_info->tree_root) {
3435                         struct btrfs_root *dead_root;
3436                         struct btrfs_fs_info *fs_info = root->fs_info;
3437                         int is_dead_root = 0;
3438
3439                         /*
3440                          * this is an orphan in the tree root. Currently these
3441                          * could come from 2 sources:
3442                          *  a) a snapshot deletion in progress
3443                          *  b) a free space cache inode
3444                          * We need to distinguish those two, as the snapshot
3445                          * orphan must not get deleted.
3446                          * find_dead_roots already ran before us, so if this
3447                          * is a snapshot deletion, we should find the root
3448                          * in the dead_roots list
3449                          */
3450                         spin_lock(&fs_info->trans_lock);
3451                         list_for_each_entry(dead_root, &fs_info->dead_roots,
3452                                             root_list) {
3453                                 if (dead_root->root_key.objectid ==
3454                                     found_key.objectid) {
3455                                         is_dead_root = 1;
3456                                         break;
3457                                 }
3458                         }
3459                         spin_unlock(&fs_info->trans_lock);
3460                         if (is_dead_root) {
3461                                 /* prevent this orphan from being found again */
3462                                 key.offset = found_key.objectid - 1;
3463                                 continue;
3464                         }
3465                 }
3466                 /*
3467                  * Inode is already gone but the orphan item is still there,
3468                  * kill the orphan item.
3469                  */
3470                 if (ret == -ENOENT) {
3471                         trans = btrfs_start_transaction(root, 1);
3472                         if (IS_ERR(trans)) {
3473                                 ret = PTR_ERR(trans);
3474                                 goto out;
3475                         }
3476                         btrfs_debug(root->fs_info, "auto deleting %Lu",
3477                                 found_key.objectid);
3478                         ret = btrfs_del_orphan_item(trans, root,
3479                                                     found_key.objectid);
3480                         btrfs_end_transaction(trans, root);
3481                         if (ret)
3482                                 goto out;
3483                         continue;
3484                 }
3485
3486                 /*
3487                  * add this inode to the orphan list so btrfs_orphan_del does
3488                  * the proper thing when we hit it
3489                  */
3490                 set_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
3491                         &BTRFS_I(inode)->runtime_flags);
3492                 atomic_inc(&root->orphan_inodes);
3493
3494                 /* if we have links, this was a truncate, lets do that */
3495                 if (inode->i_nlink) {
3496                         if (WARN_ON(!S_ISREG(inode->i_mode))) {
3497                                 iput(inode);
3498                                 continue;
3499                         }
3500                         nr_truncate++;
3501
3502                         /* 1 for the orphan item deletion. */
3503                         trans = btrfs_start_transaction(root, 1);
3504                         if (IS_ERR(trans)) {
3505                                 iput(inode);
3506                                 ret = PTR_ERR(trans);
3507                                 goto out;
3508                         }
3509                         ret = btrfs_orphan_add(trans, inode);
3510                         btrfs_end_transaction(trans, root);
3511                         if (ret) {
3512                                 iput(inode);
3513                                 goto out;
3514                         }
3515
3516                         ret = btrfs_truncate(inode);
3517                         if (ret)
3518                                 btrfs_orphan_del(NULL, inode);
3519                 } else {
3520                         nr_unlink++;
3521                 }
3522
3523                 /* this will do delete_inode and everything for us */
3524                 iput(inode);
3525                 if (ret)
3526                         goto out;
3527         }
3528         /* release the path since we're done with it */
3529         btrfs_release_path(path);
3530
3531         root->orphan_cleanup_state = ORPHAN_CLEANUP_DONE;
3532
3533         if (root->orphan_block_rsv)
3534                 btrfs_block_rsv_release(root, root->orphan_block_rsv,
3535                                         (u64)-1);
3536
3537         if (root->orphan_block_rsv ||
3538             test_bit(BTRFS_ROOT_ORPHAN_ITEM_INSERTED, &root->state)) {
3539                 trans = btrfs_join_transaction(root);
3540                 if (!IS_ERR(trans))
3541                         btrfs_end_transaction(trans, root);
3542         }
3543
3544         if (nr_unlink)
3545                 btrfs_debug(root->fs_info, "unlinked %d orphans", nr_unlink);
3546         if (nr_truncate)
3547                 btrfs_debug(root->fs_info, "truncated %d orphans", nr_truncate);
3548
3549 out:
3550         if (ret)
3551                 btrfs_err(root->fs_info,
3552                         "could not do orphan cleanup %d", ret);
3553         btrfs_free_path(path);
3554         return ret;
3555 }
3556
3557 /*
3558  * very simple check to peek ahead in the leaf looking for xattrs.  If we
3559  * don't find any xattrs, we know there can't be any acls.
3560  *
3561  * slot is the slot the inode is in, objectid is the objectid of the inode
3562  */
3563 static noinline int acls_after_inode_item(struct extent_buffer *leaf,
3564                                           int slot, u64 objectid,
3565                                           int *first_xattr_slot)
3566 {
3567         u32 nritems = btrfs_header_nritems(leaf);
3568         struct btrfs_key found_key;
3569         static u64 xattr_access = 0;
3570         static u64 xattr_default = 0;
3571         int scanned = 0;
3572
3573         if (!xattr_access) {
3574                 xattr_access = btrfs_name_hash(XATTR_NAME_POSIX_ACL_ACCESS,
3575                                         strlen(XATTR_NAME_POSIX_ACL_ACCESS));
3576                 xattr_default = btrfs_name_hash(XATTR_NAME_POSIX_ACL_DEFAULT,
3577                                         strlen(XATTR_NAME_POSIX_ACL_DEFAULT));
3578         }
3579
3580         slot++;
3581         *first_xattr_slot = -1;
3582         while (slot < nritems) {
3583                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
3584
3585                 /* we found a different objectid, there must not be acls */
3586                 if (found_key.objectid != objectid)
3587                         return 0;
3588
3589                 /* we found an xattr, assume we've got an acl */
3590                 if (found_key.type == BTRFS_XATTR_ITEM_KEY) {
3591                         if (*first_xattr_slot == -1)
3592                                 *first_xattr_slot = slot;
3593                         if (found_key.offset == xattr_access ||
3594                             found_key.offset == xattr_default)
3595                                 return 1;
3596                 }
3597
3598                 /*
3599                  * we found a key greater than an xattr key, there can't
3600                  * be any acls later on
3601                  */
3602                 if (found_key.type > BTRFS_XATTR_ITEM_KEY)
3603                         return 0;
3604
3605                 slot++;
3606                 scanned++;
3607
3608                 /*
3609                  * it goes inode, inode backrefs, xattrs, extents,
3610                  * so if there are a ton of hard links to an inode there can
3611                  * be a lot of backrefs.  Don't waste time searching too hard,
3612                  * this is just an optimization
3613                  */
3614                 if (scanned >= 8)
3615                         break;
3616         }
3617         /* we hit the end of the leaf before we found an xattr or
3618          * something larger than an xattr.  We have to assume the inode
3619          * has acls
3620          */
3621         if (*first_xattr_slot == -1)
3622                 *first_xattr_slot = slot;
3623         return 1;
3624 }
3625
3626 /*
3627  * read an inode from the btree into the in-memory inode
3628  */
3629 static int btrfs_read_locked_inode(struct inode *inode)
3630 {
3631         struct btrfs_path *path;
3632         struct extent_buffer *leaf;
3633         struct btrfs_inode_item *inode_item;
3634         struct btrfs_root *root = BTRFS_I(inode)->root;
3635         struct btrfs_key location;
3636         unsigned long ptr;
3637         int maybe_acls;
3638         u32 rdev;
3639         int ret;
3640         bool filled = false;
3641         int first_xattr_slot;
3642
3643         ret = btrfs_fill_inode(inode, &rdev);
3644         if (!ret)
3645                 filled = true;
3646
3647         path = btrfs_alloc_path();
3648         if (!path) {
3649                 ret = -ENOMEM;
3650                 goto make_bad;
3651         }
3652
3653         memcpy(&location, &BTRFS_I(inode)->location, sizeof(location));
3654
3655         ret = btrfs_lookup_inode(NULL, root, path, &location, 0);
3656         if (ret) {
3657                 if (ret > 0)
3658                         ret = -ENOENT;
3659                 goto make_bad;
3660         }
3661
3662         leaf = path->nodes[0];
3663
3664         if (filled)
3665                 goto cache_index;
3666
3667         inode_item = btrfs_item_ptr(leaf, path->slots[0],
3668                                     struct btrfs_inode_item);
3669         inode->i_mode = btrfs_inode_mode(leaf, inode_item);
3670         set_nlink(inode, btrfs_inode_nlink(leaf, inode_item));
3671         i_uid_write(inode, btrfs_inode_uid(leaf, inode_item));
3672         i_gid_write(inode, btrfs_inode_gid(leaf, inode_item));
3673         btrfs_i_size_write(inode, btrfs_inode_size(leaf, inode_item));
3674
3675         inode->i_atime.tv_sec = btrfs_timespec_sec(leaf, &inode_item->atime);
3676         inode->i_atime.tv_nsec = btrfs_timespec_nsec(leaf, &inode_item->atime);
3677
3678         inode->i_mtime.tv_sec = btrfs_timespec_sec(leaf, &inode_item->mtime);
3679         inode->i_mtime.tv_nsec = btrfs_timespec_nsec(leaf, &inode_item->mtime);
3680
3681         inode->i_ctime.tv_sec = btrfs_timespec_sec(leaf, &inode_item->ctime);
3682         inode->i_ctime.tv_nsec = btrfs_timespec_nsec(leaf, &inode_item->ctime);
3683
3684         BTRFS_I(inode)->i_otime.tv_sec =
3685                 btrfs_timespec_sec(leaf, &inode_item->otime);
3686         BTRFS_I(inode)->i_otime.tv_nsec =
3687                 btrfs_timespec_nsec(leaf, &inode_item->otime);
3688
3689         inode_set_bytes(inode, btrfs_inode_nbytes(leaf, inode_item));
3690         BTRFS_I(inode)->generation = btrfs_inode_generation(leaf, inode_item);
3691         BTRFS_I(inode)->last_trans = btrfs_inode_transid(leaf, inode_item);
3692
3693         inode->i_version = btrfs_inode_sequence(leaf, inode_item);
3694         inode->i_generation = BTRFS_I(inode)->generation;
3695         inode->i_rdev = 0;
3696         rdev = btrfs_inode_rdev(leaf, inode_item);
3697
3698         BTRFS_I(inode)->index_cnt = (u64)-1;
3699         BTRFS_I(inode)->flags = btrfs_inode_flags(leaf, inode_item);
3700
3701 cache_index:
3702         /*
3703          * If we were modified in the current generation and evicted from memory
3704          * and then re-read we need to do a full sync since we don't have any
3705          * idea about which extents were modified before we were evicted from
3706          * cache.
3707          *
3708          * This is required for both inode re-read from disk and delayed inode
3709          * in delayed_nodes_tree.
3710          */
3711         if (BTRFS_I(inode)->last_trans == root->fs_info->generation)
3712                 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
3713                         &BTRFS_I(inode)->runtime_flags);
3714
3715         /*
3716          * We don't persist the id of the transaction where an unlink operation
3717          * against the inode was last made. So here we assume the inode might
3718          * have been evicted, and therefore the exact value of last_unlink_trans
3719          * lost, and set it to last_trans to avoid metadata inconsistencies
3720          * between the inode and its parent if the inode is fsync'ed and the log
3721          * replayed. For example, in the scenario:
3722          *
3723          * touch mydir/foo
3724          * ln mydir/foo mydir/bar
3725          * sync
3726          * unlink mydir/bar
3727          * echo 2 > /proc/sys/vm/drop_caches   # evicts inode
3728          * xfs_io -c fsync mydir/foo
3729          * <power failure>
3730          * mount fs, triggers fsync log replay
3731          *
3732          * We must make sure that when we fsync our inode foo we also log its
3733          * parent inode, otherwise after log replay the parent still has the
3734          * dentry with the "bar" name but our inode foo has a link count of 1
3735          * and doesn't have an inode ref with the name "bar" anymore.
3736          *
3737          * Setting last_unlink_trans to last_trans is a pessimistic approach,
3738          * but it guarantees correctness at the expense of occasional full
3739          * transaction commits on fsync if our inode is a directory, or if our
3740          * inode is not a directory, logging its parent unnecessarily.
3741          */
3742         BTRFS_I(inode)->last_unlink_trans = BTRFS_I(inode)->last_trans;
3743
3744         path->slots[0]++;
3745         if (inode->i_nlink != 1 ||
3746             path->slots[0] >= btrfs_header_nritems(leaf))
3747                 goto cache_acl;
3748
3749         btrfs_item_key_to_cpu(leaf, &location, path->slots[0]);
3750         if (location.objectid != btrfs_ino(inode))
3751                 goto cache_acl;
3752
3753         ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
3754         if (location.type == BTRFS_INODE_REF_KEY) {
3755                 struct btrfs_inode_ref *ref;
3756
3757                 ref = (struct btrfs_inode_ref *)ptr;
3758                 BTRFS_I(inode)->dir_index = btrfs_inode_ref_index(leaf, ref);
3759         } else if (location.type == BTRFS_INODE_EXTREF_KEY) {
3760                 struct btrfs_inode_extref *extref;
3761
3762                 extref = (struct btrfs_inode_extref *)ptr;
3763                 BTRFS_I(inode)->dir_index = btrfs_inode_extref_index(leaf,
3764                                                                      extref);
3765         }
3766 cache_acl:
3767         /*
3768          * try to precache a NULL acl entry for files that don't have
3769          * any xattrs or acls
3770          */
3771         maybe_acls = acls_after_inode_item(leaf, path->slots[0],
3772                                            btrfs_ino(inode), &first_xattr_slot);
3773         if (first_xattr_slot != -1) {
3774                 path->slots[0] = first_xattr_slot;
3775                 ret = btrfs_load_inode_props(inode, path);
3776                 if (ret)
3777                         btrfs_err(root->fs_info,
3778                                   "error loading props for ino %llu (root %llu): %d",
3779                                   btrfs_ino(inode),
3780                                   root->root_key.objectid, ret);
3781         }
3782         btrfs_free_path(path);
3783
3784         if (!maybe_acls)
3785                 cache_no_acl(inode);
3786
3787         switch (inode->i_mode & S_IFMT) {
3788         case S_IFREG:
3789                 inode->i_mapping->a_ops = &btrfs_aops;
3790                 BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
3791                 inode->i_fop = &btrfs_file_operations;
3792                 inode->i_op = &btrfs_file_inode_operations;
3793                 break;
3794         case S_IFDIR:
3795                 inode->i_fop = &btrfs_dir_file_operations;
3796                 if (root == root->fs_info->tree_root)
3797                         inode->i_op = &btrfs_dir_ro_inode_operations;
3798                 else
3799                         inode->i_op = &btrfs_dir_inode_operations;
3800                 break;
3801         case S_IFLNK:
3802                 inode->i_op = &btrfs_symlink_inode_operations;
3803                 inode_nohighmem(inode);
3804                 inode->i_mapping->a_ops = &btrfs_symlink_aops;
3805                 break;
3806         default:
3807                 inode->i_op = &btrfs_special_inode_operations;
3808                 init_special_inode(inode, inode->i_mode, rdev);
3809                 break;
3810         }
3811
3812         btrfs_update_iflags(inode);
3813         return 0;
3814
3815 make_bad:
3816         btrfs_free_path(path);
3817         make_bad_inode(inode);
3818         return ret;
3819 }
3820
3821 /*
3822  * given a leaf and an inode, copy the inode fields into the leaf
3823  */
3824 static void fill_inode_item(struct btrfs_trans_handle *trans,
3825                             struct extent_buffer *leaf,
3826                             struct btrfs_inode_item *item,
3827                             struct inode *inode)
3828 {
3829         struct btrfs_map_token token;
3830
3831         btrfs_init_map_token(&token);
3832
3833         btrfs_set_token_inode_uid(leaf, item, i_uid_read(inode), &token);
3834         btrfs_set_token_inode_gid(leaf, item, i_gid_read(inode), &token);
3835         btrfs_set_token_inode_size(leaf, item, BTRFS_I(inode)->disk_i_size,
3836                                    &token);
3837         btrfs_set_token_inode_mode(leaf, item, inode->i_mode, &token);
3838         btrfs_set_token_inode_nlink(leaf, item, inode->i_nlink, &token);
3839
3840         btrfs_set_token_timespec_sec(leaf, &item->atime,
3841                                      inode->i_atime.tv_sec, &token);
3842         btrfs_set_token_timespec_nsec(leaf, &item->atime,
3843                                       inode->i_atime.tv_nsec, &token);
3844
3845         btrfs_set_token_timespec_sec(leaf, &item->mtime,
3846                                      inode->i_mtime.tv_sec, &token);
3847         btrfs_set_token_timespec_nsec(leaf, &item->mtime,
3848                                       inode->i_mtime.tv_nsec, &token);
3849
3850         btrfs_set_token_timespec_sec(leaf, &item->ctime,
3851                                      inode->i_ctime.tv_sec, &token);
3852         btrfs_set_token_timespec_nsec(leaf, &item->ctime,
3853                                       inode->i_ctime.tv_nsec, &token);
3854
3855         btrfs_set_token_timespec_sec(leaf, &item->otime,
3856                                      BTRFS_I(inode)->i_otime.tv_sec, &token);
3857         btrfs_set_token_timespec_nsec(leaf, &item->otime,
3858                                       BTRFS_I(inode)->i_otime.tv_nsec, &token);
3859
3860         btrfs_set_token_inode_nbytes(leaf, item, inode_get_bytes(inode),
3861                                      &token);
3862         btrfs_set_token_inode_generation(leaf, item, BTRFS_I(inode)->generation,
3863                                          &token);
3864         btrfs_set_token_inode_sequence(leaf, item, inode->i_version, &token);
3865         btrfs_set_token_inode_transid(leaf, item, trans->transid, &token);
3866         btrfs_set_token_inode_rdev(leaf, item, inode->i_rdev, &token);
3867         btrfs_set_token_inode_flags(leaf, item, BTRFS_I(inode)->flags, &token);
3868         btrfs_set_token_inode_block_group(leaf, item, 0, &token);
3869 }
3870
3871 /*
3872  * copy everything in the in-memory inode into the btree.
3873  */
3874 static noinline int btrfs_update_inode_item(struct btrfs_trans_handle *trans,
3875                                 struct btrfs_root *root, struct inode *inode)
3876 {
3877         struct btrfs_inode_item *inode_item;
3878         struct btrfs_path *path;
3879         struct extent_buffer *leaf;
3880         int ret;
3881
3882         path = btrfs_alloc_path();
3883         if (!path)
3884                 return -ENOMEM;
3885
3886         path->leave_spinning = 1;
3887         ret = btrfs_lookup_inode(trans, root, path, &BTRFS_I(inode)->location,
3888                                  1);
3889         if (ret) {
3890                 if (ret > 0)
3891                         ret = -ENOENT;
3892                 goto failed;
3893         }
3894
3895         leaf = path->nodes[0];
3896         inode_item = btrfs_item_ptr(leaf, path->slots[0],
3897                                     struct btrfs_inode_item);
3898
3899         fill_inode_item(trans, leaf, inode_item, inode);
3900         btrfs_mark_buffer_dirty(leaf);
3901         btrfs_set_inode_last_trans(trans, inode);
3902         ret = 0;
3903 failed:
3904         btrfs_free_path(path);
3905         return ret;
3906 }
3907
3908 /*
3909  * copy everything in the in-memory inode into the btree.
3910  */
3911 noinline int btrfs_update_inode(struct btrfs_trans_handle *trans,
3912                                 struct btrfs_root *root, struct inode *inode)
3913 {
3914         int ret;
3915
3916         /*
3917          * If the inode is a free space inode, we can deadlock during commit
3918          * if we put it into the delayed code.
3919          *
3920          * The data relocation inode should also be directly updated
3921          * without delay
3922          */
3923         if (!btrfs_is_free_space_inode(inode)
3924             && root->root_key.objectid != BTRFS_DATA_RELOC_TREE_OBJECTID
3925             && !root->fs_info->log_root_recovering) {
3926                 btrfs_update_root_times(trans, root);
3927
3928                 ret = btrfs_delayed_update_inode(trans, root, inode);
3929                 if (!ret)
3930                         btrfs_set_inode_last_trans(trans, inode);
3931                 return ret;
3932         }
3933
3934         return btrfs_update_inode_item(trans, root, inode);
3935 }
3936
3937 noinline int btrfs_update_inode_fallback(struct btrfs_trans_handle *trans,
3938                                          struct btrfs_root *root,
3939                                          struct inode *inode)
3940 {
3941         int ret;
3942
3943         ret = btrfs_update_inode(trans, root, inode);
3944         if (ret == -ENOSPC)
3945                 return btrfs_update_inode_item(trans, root, inode);
3946         return ret;
3947 }
3948
3949 /*
3950  * unlink helper that gets used here in inode.c and in the tree logging
3951  * recovery code.  It remove a link in a directory with a given name, and
3952  * also drops the back refs in the inode to the directory
3953  */
3954 static int __btrfs_unlink_inode(struct btrfs_trans_handle *trans,
3955                                 struct btrfs_root *root,
3956                                 struct inode *dir, struct inode *inode,
3957                                 const char *name, int name_len)
3958 {
3959         struct btrfs_path *path;
3960         int ret = 0;
3961         struct extent_buffer *leaf;
3962         struct btrfs_dir_item *di;
3963         struct btrfs_key key;
3964         u64 index;
3965         u64 ino = btrfs_ino(inode);
3966         u64 dir_ino = btrfs_ino(dir);
3967
3968         path = btrfs_alloc_path();
3969         if (!path) {
3970                 ret = -ENOMEM;
3971                 goto out;
3972         }
3973
3974         path->leave_spinning = 1;
3975         di = btrfs_lookup_dir_item(trans, root, path, dir_ino,
3976                                     name, name_len, -1);
3977         if (IS_ERR(di)) {
3978                 ret = PTR_ERR(di);
3979                 goto err;
3980         }
3981         if (!di) {
3982                 ret = -ENOENT;
3983                 goto err;
3984         }
3985         leaf = path->nodes[0];
3986         btrfs_dir_item_key_to_cpu(leaf, di, &key);
3987         ret = btrfs_delete_one_dir_name(trans, root, path, di);
3988         if (ret)
3989                 goto err;
3990         btrfs_release_path(path);
3991
3992         /*
3993          * If we don't have dir index, we have to get it by looking up
3994          * the inode ref, since we get the inode ref, remove it directly,
3995          * it is unnecessary to do delayed deletion.
3996          *
3997          * But if we have dir index, needn't search inode ref to get it.
3998          * Since the inode ref is close to the inode item, it is better
3999          * that we delay to delete it, and just do this deletion when
4000          * we update the inode item.
4001          */
4002         if (BTRFS_I(inode)->dir_index) {
4003                 ret = btrfs_delayed_delete_inode_ref(inode);
4004                 if (!ret) {
4005                         index = BTRFS_I(inode)->dir_index;
4006                         goto skip_backref;
4007                 }
4008         }
4009
4010         ret = btrfs_del_inode_ref(trans, root, name, name_len, ino,
4011                                   dir_ino, &index);
4012         if (ret) {
4013                 btrfs_info(root->fs_info,
4014                         "failed to delete reference to %.*s, inode %llu parent %llu",
4015                         name_len, name, ino, dir_ino);
4016                 btrfs_abort_transaction(trans, root, ret);
4017                 goto err;
4018         }
4019 skip_backref:
4020         ret = btrfs_delete_delayed_dir_index(trans, root, dir, index);
4021         if (ret) {
4022                 btrfs_abort_transaction(trans, root, ret);
4023                 goto err;
4024         }
4025
4026         ret = btrfs_del_inode_ref_in_log(trans, root, name, name_len,
4027                                          inode, dir_ino);
4028         if (ret != 0 && ret != -ENOENT) {
4029                 btrfs_abort_transaction(trans, root, ret);
4030                 goto err;
4031         }
4032
4033         ret = btrfs_del_dir_entries_in_log(trans, root, name, name_len,
4034                                            dir, index);
4035         if (ret == -ENOENT)
4036                 ret = 0;
4037         else if (ret)
4038                 btrfs_abort_transaction(trans, root, ret);
4039 err:
4040         btrfs_free_path(path);
4041         if (ret)
4042                 goto out;
4043
4044         btrfs_i_size_write(dir, dir->i_size - name_len * 2);
4045         inode_inc_iversion(inode);
4046         inode_inc_iversion(dir);
4047         inode->i_ctime = dir->i_mtime =
4048                 dir->i_ctime = current_fs_time(inode->i_sb);
4049         ret = btrfs_update_inode(trans, root, dir);
4050 out:
4051         return ret;
4052 }
4053
4054 int btrfs_unlink_inode(struct btrfs_trans_handle *trans,
4055                        struct btrfs_root *root,
4056                        struct inode *dir, struct inode *inode,
4057                        const char *name, int name_len)
4058 {
4059         int ret;
4060         ret = __btrfs_unlink_inode(trans, root, dir, inode, name, name_len);
4061         if (!ret) {
4062                 drop_nlink(inode);
4063                 ret = btrfs_update_inode(trans, root, inode);
4064         }
4065         return ret;
4066 }
4067
4068 /*
4069  * helper to start transaction for unlink and rmdir.
4070  *
4071  * unlink and rmdir are special in btrfs, they do not always free space, so
4072  * if we cannot make our reservations the normal way try and see if there is
4073  * plenty of slack room in the global reserve to migrate, otherwise we cannot
4074  * allow the unlink to occur.
4075  */
4076 static struct btrfs_trans_handle *__unlink_start_trans(struct inode *dir)
4077 {
4078         struct btrfs_root *root = BTRFS_I(dir)->root;
4079
4080         /*
4081          * 1 for the possible orphan item
4082          * 1 for the dir item
4083          * 1 for the dir index
4084          * 1 for the inode ref
4085          * 1 for the inode
4086          */
4087         return btrfs_start_transaction_fallback_global_rsv(root, 5, 5);
4088 }
4089
4090 static int btrfs_unlink(struct inode *dir, struct dentry *dentry)
4091 {
4092         struct btrfs_root *root = BTRFS_I(dir)->root;
4093         struct btrfs_trans_handle *trans;
4094         struct inode *inode = d_inode(dentry);
4095         int ret;
4096
4097         trans = __unlink_start_trans(dir);
4098         if (IS_ERR(trans))
4099                 return PTR_ERR(trans);
4100
4101         btrfs_record_unlink_dir(trans, dir, d_inode(dentry), 0);
4102
4103         ret = btrfs_unlink_inode(trans, root, dir, d_inode(dentry),
4104                                  dentry->d_name.name, dentry->d_name.len);
4105         if (ret)
4106                 goto out;
4107
4108         if (inode->i_nlink == 0) {
4109                 ret = btrfs_orphan_add(trans, inode);
4110                 if (ret)
4111                         goto out;
4112         }
4113
4114 out:
4115         btrfs_end_transaction(trans, root);
4116         btrfs_btree_balance_dirty(root);
4117         return ret;
4118 }
4119
4120 int btrfs_unlink_subvol(struct btrfs_trans_handle *trans,
4121                         struct btrfs_root *root,
4122                         struct inode *dir, u64 objectid,
4123                         const char *name, int name_len)
4124 {
4125         struct btrfs_path *path;
4126         struct extent_buffer *leaf;
4127         struct btrfs_dir_item *di;
4128         struct btrfs_key key;
4129         u64 index;
4130         int ret;
4131         u64 dir_ino = btrfs_ino(dir);
4132
4133         path = btrfs_alloc_path();
4134         if (!path)
4135                 return -ENOMEM;
4136
4137         di = btrfs_lookup_dir_item(trans, root, path, dir_ino,
4138                                    name, name_len, -1);
4139         if (IS_ERR_OR_NULL(di)) {
4140                 if (!di)
4141                         ret = -ENOENT;
4142                 else
4143                         ret = PTR_ERR(di);
4144                 goto out;
4145         }
4146
4147         leaf = path->nodes[0];
4148         btrfs_dir_item_key_to_cpu(leaf, di, &key);
4149         WARN_ON(key.type != BTRFS_ROOT_ITEM_KEY || key.objectid != objectid);
4150         ret = btrfs_delete_one_dir_name(trans, root, path, di);
4151         if (ret) {
4152                 btrfs_abort_transaction(trans, root, ret);
4153                 goto out;
4154         }
4155         btrfs_release_path(path);
4156
4157         ret = btrfs_del_root_ref(trans, root->fs_info->tree_root,
4158                                  objectid, root->root_key.objectid,
4159                                  dir_ino, &index, name, name_len);
4160         if (ret < 0) {
4161                 if (ret != -ENOENT) {
4162                         btrfs_abort_transaction(trans, root, ret);
4163                         goto out;
4164                 }
4165                 di = btrfs_search_dir_index_item(root, path, dir_ino,
4166                                                  name, name_len);
4167                 if (IS_ERR_OR_NULL(di)) {
4168                         if (!di)
4169                                 ret = -ENOENT;
4170                         else
4171                                 ret = PTR_ERR(di);
4172                         btrfs_abort_transaction(trans, root, ret);
4173                         goto out;
4174                 }
4175
4176                 leaf = path->nodes[0];
4177                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
4178                 btrfs_release_path(path);
4179                 index = key.offset;
4180         }
4181         btrfs_release_path(path);
4182
4183         ret = btrfs_delete_delayed_dir_index(trans, root, dir, index);
4184         if (ret) {
4185                 btrfs_abort_transaction(trans, root, ret);
4186                 goto out;
4187         }
4188
4189         btrfs_i_size_write(dir, dir->i_size - name_len * 2);
4190         inode_inc_iversion(dir);
4191         dir->i_mtime = dir->i_ctime = current_fs_time(dir->i_sb);
4192         ret = btrfs_update_inode_fallback(trans, root, dir);
4193         if (ret)
4194                 btrfs_abort_transaction(trans, root, ret);
4195 out:
4196         btrfs_free_path(path);
4197         return ret;
4198 }
4199
4200 static int btrfs_rmdir(struct inode *dir, struct dentry *dentry)
4201 {
4202         struct inode *inode = d_inode(dentry);
4203         int err = 0;
4204         struct btrfs_root *root = BTRFS_I(dir)->root;
4205         struct btrfs_trans_handle *trans;
4206         u64 last_unlink_trans;
4207
4208         if (inode->i_size > BTRFS_EMPTY_DIR_SIZE)
4209                 return -ENOTEMPTY;
4210         if (btrfs_ino(inode) == BTRFS_FIRST_FREE_OBJECTID)
4211                 return -EPERM;
4212
4213         trans = __unlink_start_trans(dir);
4214         if (IS_ERR(trans))
4215                 return PTR_ERR(trans);
4216
4217         if (unlikely(btrfs_ino(inode) == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)) {
4218                 err = btrfs_unlink_subvol(trans, root, dir,
4219                                           BTRFS_I(inode)->location.objectid,
4220                                           dentry->d_name.name,
4221                                           dentry->d_name.len);
4222                 goto out;
4223         }
4224
4225         err = btrfs_orphan_add(trans, inode);
4226         if (err)
4227                 goto out;
4228
4229         last_unlink_trans = BTRFS_I(inode)->last_unlink_trans;
4230
4231         /* now the directory is empty */
4232         err = btrfs_unlink_inode(trans, root, dir, d_inode(dentry),
4233                                  dentry->d_name.name, dentry->d_name.len);
4234         if (!err) {
4235                 btrfs_i_size_write(inode, 0);
4236                 /*
4237                  * Propagate the last_unlink_trans value of the deleted dir to
4238                  * its parent directory. This is to prevent an unrecoverable
4239                  * log tree in the case we do something like this:
4240                  * 1) create dir foo
4241                  * 2) create snapshot under dir foo
4242                  * 3) delete the snapshot
4243                  * 4) rmdir foo
4244                  * 5) mkdir foo
4245                  * 6) fsync foo or some file inside foo
4246                  */
4247                 if (last_unlink_trans >= trans->transid)
4248                         BTRFS_I(dir)->last_unlink_trans = last_unlink_trans;
4249         }
4250 out:
4251         btrfs_end_transaction(trans, root);
4252         btrfs_btree_balance_dirty(root);
4253
4254         return err;
4255 }
4256
4257 static int truncate_space_check(struct btrfs_trans_handle *trans,
4258                                 struct btrfs_root *root,
4259                                 u64 bytes_deleted)
4260 {
4261         int ret;
4262
4263         /*
4264          * This is only used to apply pressure to the enospc system, we don't
4265          * intend to use this reservation at all.
4266          */
4267         bytes_deleted = btrfs_csum_bytes_to_leaves(root, bytes_deleted);
4268         bytes_deleted *= root->nodesize;
4269         ret = btrfs_block_rsv_add(root, &root->fs_info->trans_block_rsv,
4270                                   bytes_deleted, BTRFS_RESERVE_NO_FLUSH);
4271         if (!ret) {
4272                 trace_btrfs_space_reservation(root->fs_info, "transaction",
4273                                               trans->transid,
4274                                               bytes_deleted, 1);
4275                 trans->bytes_reserved += bytes_deleted;
4276         }
4277         return ret;
4278
4279 }
4280
4281 static int truncate_inline_extent(struct inode *inode,
4282                                   struct btrfs_path *path,
4283                                   struct btrfs_key *found_key,
4284                                   const u64 item_end,
4285                                   const u64 new_size)
4286 {
4287         struct extent_buffer *leaf = path->nodes[0];
4288         int slot = path->slots[0];
4289         struct btrfs_file_extent_item *fi;
4290         u32 size = (u32)(new_size - found_key->offset);
4291         struct btrfs_root *root = BTRFS_I(inode)->root;
4292
4293         fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
4294
4295         if (btrfs_file_extent_compression(leaf, fi) != BTRFS_COMPRESS_NONE) {
4296                 loff_t offset = new_size;
4297                 loff_t page_end = ALIGN(offset, PAGE_SIZE);
4298
4299                 /*
4300                  * Zero out the remaining of the last page of our inline extent,
4301                  * instead of directly truncating our inline extent here - that
4302                  * would be much more complex (decompressing all the data, then
4303                  * compressing the truncated data, which might be bigger than
4304                  * the size of the inline extent, resize the extent, etc).
4305                  * We release the path because to get the page we might need to
4306                  * read the extent item from disk (data not in the page cache).
4307                  */
4308                 btrfs_release_path(path);
4309                 return btrfs_truncate_block(inode, offset, page_end - offset,
4310                                         0);
4311         }
4312
4313         btrfs_set_file_extent_ram_bytes(leaf, fi, size);
4314         size = btrfs_file_extent_calc_inline_size(size);
4315         btrfs_truncate_item(root, path, size, 1);
4316
4317         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state))
4318                 inode_sub_bytes(inode, item_end + 1 - new_size);
4319
4320         return 0;
4321 }
4322
4323 /*
4324  * this can truncate away extent items, csum items and directory items.
4325  * It starts at a high offset and removes keys until it can't find
4326  * any higher than new_size
4327  *
4328  * csum items that cross the new i_size are truncated to the new size
4329  * as well.
4330  *
4331  * min_type is the minimum key type to truncate down to.  If set to 0, this
4332  * will kill all the items on this inode, including the INODE_ITEM_KEY.
4333  */
4334 int btrfs_truncate_inode_items(struct btrfs_trans_handle *trans,
4335                                struct btrfs_root *root,
4336                                struct inode *inode,
4337                                u64 new_size, u32 min_type)
4338 {
4339         struct btrfs_path *path;
4340         struct extent_buffer *leaf;
4341         struct btrfs_file_extent_item *fi;
4342         struct btrfs_key key;
4343         struct btrfs_key found_key;
4344         u64 extent_start = 0;
4345         u64 extent_num_bytes = 0;
4346         u64 extent_offset = 0;
4347         u64 item_end = 0;
4348         u64 last_size = new_size;
4349         u32 found_type = (u8)-1;
4350         int found_extent;
4351         int del_item;
4352         int pending_del_nr = 0;
4353         int pending_del_slot = 0;
4354         int extent_type = -1;
4355         int ret;
4356         int err = 0;
4357         u64 ino = btrfs_ino(inode);
4358         u64 bytes_deleted = 0;
4359         bool be_nice = 0;
4360         bool should_throttle = 0;
4361         bool should_end = 0;
4362
4363         BUG_ON(new_size > 0 && min_type != BTRFS_EXTENT_DATA_KEY);
4364
4365         /*
4366          * for non-free space inodes and ref cows, we want to back off from
4367          * time to time
4368          */
4369         if (!btrfs_is_free_space_inode(inode) &&
4370             test_bit(BTRFS_ROOT_REF_COWS, &root->state))
4371                 be_nice = 1;
4372
4373         path = btrfs_alloc_path();
4374         if (!path)
4375                 return -ENOMEM;
4376         path->reada = READA_BACK;
4377
4378         /*
4379          * We want to drop from the next block forward in case this new size is
4380          * not block aligned since we will be keeping the last block of the
4381          * extent just the way it is.
4382          */
4383         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
4384             root == root->fs_info->tree_root)
4385                 btrfs_drop_extent_cache(inode, ALIGN(new_size,
4386                                         root->sectorsize), (u64)-1, 0);
4387
4388         /*
4389          * This function is also used to drop the items in the log tree before
4390          * we relog the inode, so if root != BTRFS_I(inode)->root, it means
4391          * it is used to drop the loged items. So we shouldn't kill the delayed
4392          * items.
4393          */
4394         if (min_type == 0 && root == BTRFS_I(inode)->root)
4395                 btrfs_kill_delayed_inode_items(inode);
4396
4397         key.objectid = ino;
4398         key.offset = (u64)-1;
4399         key.type = (u8)-1;
4400
4401 search_again:
4402         /*
4403          * with a 16K leaf size and 128MB extents, you can actually queue
4404          * up a huge file in a single leaf.  Most of the time that
4405          * bytes_deleted is > 0, it will be huge by the time we get here
4406          */
4407         if (be_nice && bytes_deleted > SZ_32M) {
4408                 if (btrfs_should_end_transaction(trans, root)) {
4409                         err = -EAGAIN;
4410                         goto error;
4411                 }
4412         }
4413
4414
4415         path->leave_spinning = 1;
4416         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
4417         if (ret < 0) {
4418                 err = ret;
4419                 goto out;
4420         }
4421
4422         if (ret > 0) {
4423                 /* there are no items in the tree for us to truncate, we're
4424                  * done
4425                  */
4426                 if (path->slots[0] == 0)
4427                         goto out;
4428                 path->slots[0]--;
4429         }
4430
4431         while (1) {
4432                 fi = NULL;
4433                 leaf = path->nodes[0];
4434                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
4435                 found_type = found_key.type;
4436
4437                 if (found_key.objectid != ino)
4438                         break;
4439
4440                 if (found_type < min_type)
4441                         break;
4442
4443                 item_end = found_key.offset;
4444                 if (found_type == BTRFS_EXTENT_DATA_KEY) {
4445                         fi = btrfs_item_ptr(leaf, path->slots[0],
4446                                             struct btrfs_file_extent_item);
4447                         extent_type = btrfs_file_extent_type(leaf, fi);
4448                         if (extent_type != BTRFS_FILE_EXTENT_INLINE) {
4449                                 item_end +=
4450                                     btrfs_file_extent_num_bytes(leaf, fi);
4451                         } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
4452                                 item_end += btrfs_file_extent_inline_len(leaf,
4453                                                          path->slots[0], fi);
4454                         }
4455                         item_end--;
4456                 }
4457                 if (found_type > min_type) {
4458                         del_item = 1;
4459                 } else {
4460                         if (item_end < new_size)
4461                                 break;
4462                         if (found_key.offset >= new_size)
4463                                 del_item = 1;
4464                         else
4465                                 del_item = 0;
4466                 }
4467                 found_extent = 0;
4468                 /* FIXME, shrink the extent if the ref count is only 1 */
4469                 if (found_type != BTRFS_EXTENT_DATA_KEY)
4470                         goto delete;
4471
4472                 if (del_item)
4473                         last_size = found_key.offset;
4474                 else
4475                         last_size = new_size;
4476
4477                 if (extent_type != BTRFS_FILE_EXTENT_INLINE) {
4478                         u64 num_dec;
4479                         extent_start = btrfs_file_extent_disk_bytenr(leaf, fi);
4480                         if (!del_item) {
4481                                 u64 orig_num_bytes =
4482                                         btrfs_file_extent_num_bytes(leaf, fi);
4483                                 extent_num_bytes = ALIGN(new_size -
4484                                                 found_key.offset,
4485                                                 root->sectorsize);
4486                                 btrfs_set_file_extent_num_bytes(leaf, fi,
4487                                                          extent_num_bytes);
4488                                 num_dec = (orig_num_bytes -
4489                                            extent_num_bytes);
4490                                 if (test_bit(BTRFS_ROOT_REF_COWS,
4491                                              &root->state) &&
4492                                     extent_start != 0)
4493                                         inode_sub_bytes(inode, num_dec);
4494                                 btrfs_mark_buffer_dirty(leaf);
4495                         } else {
4496                                 extent_num_bytes =
4497                                         btrfs_file_extent_disk_num_bytes(leaf,
4498                                                                          fi);
4499                                 extent_offset = found_key.offset -
4500                                         btrfs_file_extent_offset(leaf, fi);
4501
4502                                 /* FIXME blocksize != 4096 */
4503                                 num_dec = btrfs_file_extent_num_bytes(leaf, fi);
4504                                 if (extent_start != 0) {
4505                                         found_extent = 1;
4506                                         if (test_bit(BTRFS_ROOT_REF_COWS,
4507                                                      &root->state))
4508                                                 inode_sub_bytes(inode, num_dec);
4509                                 }
4510                         }
4511                 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
4512                         /*
4513                          * we can't truncate inline items that have had
4514                          * special encodings
4515                          */
4516                         if (!del_item &&
4517                             btrfs_file_extent_encryption(leaf, fi) == 0 &&
4518                             btrfs_file_extent_other_encoding(leaf, fi) == 0) {
4519
4520                                 /*
4521                                  * Need to release path in order to truncate a
4522                                  * compressed extent. So delete any accumulated
4523                                  * extent items so far.
4524                                  */
4525                                 if (btrfs_file_extent_compression(leaf, fi) !=
4526                                     BTRFS_COMPRESS_NONE && pending_del_nr) {
4527                                         err = btrfs_del_items(trans, root, path,
4528                                                               pending_del_slot,
4529                                                               pending_del_nr);
4530                                         if (err) {
4531                                                 btrfs_abort_transaction(trans,
4532                                                                         root,
4533                                                                         err);
4534                                                 goto error;
4535                                         }
4536                                         pending_del_nr = 0;
4537                                 }
4538
4539                                 err = truncate_inline_extent(inode, path,
4540                                                              &found_key,
4541                                                              item_end,
4542                                                              new_size);
4543                                 if (err) {
4544                                         btrfs_abort_transaction(trans,
4545                                                                 root, err);
4546                                         goto error;
4547                                 }
4548                         } else if (test_bit(BTRFS_ROOT_REF_COWS,
4549                                             &root->state)) {
4550                                 inode_sub_bytes(inode, item_end + 1 - new_size);
4551                         }
4552                 }
4553 delete:
4554                 if (del_item) {
4555                         if (!pending_del_nr) {
4556                                 /* no pending yet, add ourselves */
4557                                 pending_del_slot = path->slots[0];
4558                                 pending_del_nr = 1;
4559                         } else if (pending_del_nr &&
4560                                    path->slots[0] + 1 == pending_del_slot) {
4561                                 /* hop on the pending chunk */
4562                                 pending_del_nr++;
4563                                 pending_del_slot = path->slots[0];
4564                         } else {
4565                                 BUG();
4566                         }
4567                 } else {
4568                         break;
4569                 }
4570                 should_throttle = 0;
4571
4572                 if (found_extent &&
4573                     (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
4574                      root == root->fs_info->tree_root)) {
4575                         btrfs_set_path_blocking(path);
4576                         bytes_deleted += extent_num_bytes;
4577                         ret = btrfs_free_extent(trans, root, extent_start,
4578                                                 extent_num_bytes, 0,
4579                                                 btrfs_header_owner(leaf),
4580                                                 ino, extent_offset);
4581                         BUG_ON(ret);
4582                         if (btrfs_should_throttle_delayed_refs(trans, root))
4583                                 btrfs_async_run_delayed_refs(root,
4584                                                              trans->transid,
4585                                         trans->delayed_ref_updates * 2, 0);
4586                         if (be_nice) {
4587                                 if (truncate_space_check(trans, root,
4588                                                          extent_num_bytes)) {
4589                                         should_end = 1;
4590                                 }
4591                                 if (btrfs_should_throttle_delayed_refs(trans,
4592                                                                        root)) {
4593                                         should_throttle = 1;
4594                                 }
4595                         }
4596                 }
4597
4598                 if (found_type == BTRFS_INODE_ITEM_KEY)
4599                         break;
4600
4601                 if (path->slots[0] == 0 ||
4602                     path->slots[0] != pending_del_slot ||
4603                     should_throttle || should_end) {
4604                         if (pending_del_nr) {
4605                                 ret = btrfs_del_items(trans, root, path,
4606                                                 pending_del_slot,
4607                                                 pending_del_nr);
4608                                 if (ret) {
4609                                         btrfs_abort_transaction(trans,
4610                                                                 root, ret);
4611                                         goto error;
4612                                 }
4613                                 pending_del_nr = 0;
4614                         }
4615                         btrfs_release_path(path);
4616                         if (should_throttle) {
4617                                 unsigned long updates = trans->delayed_ref_updates;
4618                                 if (updates) {
4619                                         trans->delayed_ref_updates = 0;
4620                                         ret = btrfs_run_delayed_refs(trans, root, updates * 2);
4621                                         if (ret && !err)
4622                                                 err = ret;
4623                                 }
4624                         }
4625                         /*
4626                          * if we failed to refill our space rsv, bail out
4627                          * and let the transaction restart
4628                          */
4629                         if (should_end) {
4630                                 err = -EAGAIN;
4631                                 goto error;
4632                         }
4633                         goto search_again;
4634                 } else {
4635                         path->slots[0]--;
4636                 }
4637         }
4638 out:
4639         if (pending_del_nr) {
4640                 ret = btrfs_del_items(trans, root, path, pending_del_slot,
4641                                       pending_del_nr);
4642                 if (ret)
4643                         btrfs_abort_transaction(trans, root, ret);
4644         }
4645 error:
4646         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID)
4647                 btrfs_ordered_update_i_size(inode, last_size, NULL);
4648
4649         btrfs_free_path(path);
4650
4651         if (be_nice && bytes_deleted > SZ_32M) {
4652                 unsigned long updates = trans->delayed_ref_updates;
4653                 if (updates) {
4654                         trans->delayed_ref_updates = 0;
4655                         ret = btrfs_run_delayed_refs(trans, root, updates * 2);
4656                         if (ret && !err)
4657                                 err = ret;
4658                 }
4659         }
4660         return err;
4661 }
4662
4663 /*
4664  * btrfs_truncate_block - read, zero a chunk and write a block
4665  * @inode - inode that we're zeroing
4666  * @from - the offset to start zeroing
4667  * @len - the length to zero, 0 to zero the entire range respective to the
4668  *      offset
4669  * @front - zero up to the offset instead of from the offset on
4670  *
4671  * This will find the block for the "from" offset and cow the block and zero the
4672  * part we want to zero.  This is used with truncate and hole punching.
4673  */
4674 int btrfs_truncate_block(struct inode *inode, loff_t from, loff_t len,
4675                         int front)
4676 {
4677         struct address_space *mapping = inode->i_mapping;
4678         struct btrfs_root *root = BTRFS_I(inode)->root;
4679         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
4680         struct btrfs_ordered_extent *ordered;
4681         struct extent_state *cached_state = NULL;
4682         char *kaddr;
4683         u32 blocksize = root->sectorsize;
4684         pgoff_t index = from >> PAGE_SHIFT;
4685         unsigned offset = from & (blocksize - 1);
4686         struct page *page;
4687         gfp_t mask = btrfs_alloc_write_mask(mapping);
4688         int ret = 0;
4689         u64 block_start;
4690         u64 block_end;
4691
4692         if ((offset & (blocksize - 1)) == 0 &&
4693             (!len || ((len & (blocksize - 1)) == 0)))
4694                 goto out;
4695
4696         ret = btrfs_delalloc_reserve_space(inode,
4697                         round_down(from, blocksize), blocksize);
4698         if (ret)
4699                 goto out;
4700
4701 again:
4702         page = find_or_create_page(mapping, index, mask);
4703         if (!page) {
4704                 btrfs_delalloc_release_space(inode,
4705                                 round_down(from, blocksize),
4706                                 blocksize);
4707                 ret = -ENOMEM;
4708                 goto out;
4709         }
4710
4711         block_start = round_down(from, blocksize);
4712         block_end = block_start + blocksize - 1;
4713
4714         if (!PageUptodate(page)) {
4715                 ret = btrfs_readpage(NULL, page);
4716                 lock_page(page);
4717                 if (page->mapping != mapping) {
4718                         unlock_page(page);
4719                         put_page(page);
4720                         goto again;
4721                 }
4722                 if (!PageUptodate(page)) {
4723                         ret = -EIO;
4724                         goto out_unlock;
4725                 }
4726         }
4727         wait_on_page_writeback(page);
4728
4729         lock_extent_bits(io_tree, block_start, block_end, &cached_state);
4730         set_page_extent_mapped(page);
4731
4732         ordered = btrfs_lookup_ordered_extent(inode, block_start);
4733         if (ordered) {
4734                 unlock_extent_cached(io_tree, block_start, block_end,
4735                                      &cached_state, GFP_NOFS);
4736                 unlock_page(page);
4737                 put_page(page);
4738                 btrfs_start_ordered_extent(inode, ordered, 1);
4739                 btrfs_put_ordered_extent(ordered);
4740                 goto again;
4741         }
4742
4743         clear_extent_bit(&BTRFS_I(inode)->io_tree, block_start, block_end,
4744                           EXTENT_DIRTY | EXTENT_DELALLOC |
4745                           EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG,
4746                           0, 0, &cached_state, GFP_NOFS);
4747
4748         ret = btrfs_set_extent_delalloc(inode, block_start, block_end,
4749                                         &cached_state);
4750         if (ret) {
4751                 unlock_extent_cached(io_tree, block_start, block_end,
4752                                      &cached_state, GFP_NOFS);
4753                 goto out_unlock;
4754         }
4755
4756         if (offset != blocksize) {
4757                 if (!len)
4758                         len = blocksize - offset;
4759                 kaddr = kmap(page);
4760                 if (front)
4761                         memset(kaddr + (block_start - page_offset(page)),
4762                                 0, offset);
4763                 else
4764                         memset(kaddr + (block_start - page_offset(page)) +  offset,
4765                                 0, len);
4766                 flush_dcache_page(page);
4767                 kunmap(page);
4768         }
4769         ClearPageChecked(page);
4770         set_page_dirty(page);
4771         unlock_extent_cached(io_tree, block_start, block_end, &cached_state,
4772                              GFP_NOFS);
4773
4774 out_unlock:
4775         if (ret)
4776                 btrfs_delalloc_release_space(inode, block_start,
4777                                              blocksize);
4778         unlock_page(page);
4779         put_page(page);
4780 out:
4781         return ret;
4782 }
4783
4784 static int maybe_insert_hole(struct btrfs_root *root, struct inode *inode,
4785                              u64 offset, u64 len)
4786 {
4787         struct btrfs_trans_handle *trans;
4788         int ret;
4789
4790         /*
4791          * Still need to make sure the inode looks like it's been updated so
4792          * that any holes get logged if we fsync.
4793          */
4794         if (btrfs_fs_incompat(root->fs_info, NO_HOLES)) {
4795                 BTRFS_I(inode)->last_trans = root->fs_info->generation;
4796                 BTRFS_I(inode)->last_sub_trans = root->log_transid;
4797                 BTRFS_I(inode)->last_log_commit = root->last_log_commit;
4798                 return 0;
4799         }
4800
4801         /*
4802          * 1 - for the one we're dropping
4803          * 1 - for the one we're adding
4804          * 1 - for updating the inode.
4805          */
4806         trans = btrfs_start_transaction(root, 3);
4807         if (IS_ERR(trans))
4808                 return PTR_ERR(trans);
4809
4810         ret = btrfs_drop_extents(trans, root, inode, offset, offset + len, 1);
4811         if (ret) {
4812                 btrfs_abort_transaction(trans, root, ret);
4813                 btrfs_end_transaction(trans, root);
4814                 return ret;
4815         }
4816
4817         ret = btrfs_insert_file_extent(trans, root, btrfs_ino(inode), offset,
4818                                        0, 0, len, 0, len, 0, 0, 0);
4819         if (ret)
4820                 btrfs_abort_transaction(trans, root, ret);
4821         else
4822                 btrfs_update_inode(trans, root, inode);
4823         btrfs_end_transaction(trans, root);
4824         return ret;
4825 }
4826
4827 /*
4828  * This function puts in dummy file extents for the area we're creating a hole
4829  * for.  So if we are truncating this file to a larger size we need to insert
4830  * these file extents so that btrfs_get_extent will return a EXTENT_MAP_HOLE for
4831  * the range between oldsize and size
4832  */
4833 int btrfs_cont_expand(struct inode *inode, loff_t oldsize, loff_t size)
4834 {
4835         struct btrfs_root *root = BTRFS_I(inode)->root;
4836         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
4837         struct extent_map *em = NULL;
4838         struct extent_state *cached_state = NULL;
4839         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
4840         u64 hole_start = ALIGN(oldsize, root->sectorsize);
4841         u64 block_end = ALIGN(size, root->sectorsize);
4842         u64 last_byte;
4843         u64 cur_offset;
4844         u64 hole_size;
4845         int err = 0;
4846
4847         /*
4848          * If our size started in the middle of a block we need to zero out the
4849          * rest of the block before we expand the i_size, otherwise we could
4850          * expose stale data.
4851          */
4852         err = btrfs_truncate_block(inode, oldsize, 0, 0);
4853         if (err)
4854                 return err;
4855
4856         if (size <= hole_start)
4857                 return 0;
4858
4859         while (1) {
4860                 struct btrfs_ordered_extent *ordered;
4861
4862                 lock_extent_bits(io_tree, hole_start, block_end - 1,
4863                                  &cached_state);
4864                 ordered = btrfs_lookup_ordered_range(inode, hole_start,
4865                                                      block_end - hole_start);
4866                 if (!ordered)
4867                         break;
4868                 unlock_extent_cached(io_tree, hole_start, block_end - 1,
4869                                      &cached_state, GFP_NOFS);
4870                 btrfs_start_ordered_extent(inode, ordered, 1);
4871                 btrfs_put_ordered_extent(ordered);
4872         }
4873
4874         cur_offset = hole_start;
4875         while (1) {
4876                 em = btrfs_get_extent(inode, NULL, 0, cur_offset,
4877                                 block_end - cur_offset, 0);
4878                 if (IS_ERR(em)) {
4879                         err = PTR_ERR(em);
4880                         em = NULL;
4881                         break;
4882                 }
4883                 last_byte = min(extent_map_end(em), block_end);
4884                 last_byte = ALIGN(last_byte , root->sectorsize);
4885                 if (!test_bit(EXTENT_FLAG_PREALLOC, &em->flags)) {
4886                         struct extent_map *hole_em;
4887                         hole_size = last_byte - cur_offset;
4888
4889                         err = maybe_insert_hole(root, inode, cur_offset,
4890                                                 hole_size);
4891                         if (err)
4892                                 break;
4893                         btrfs_drop_extent_cache(inode, cur_offset,
4894                                                 cur_offset + hole_size - 1, 0);
4895                         hole_em = alloc_extent_map();
4896                         if (!hole_em) {
4897                                 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
4898                                         &BTRFS_I(inode)->runtime_flags);
4899                                 goto next;
4900                         }
4901                         hole_em->start = cur_offset;
4902                         hole_em->len = hole_size;
4903                         hole_em->orig_start = cur_offset;
4904
4905                         hole_em->block_start = EXTENT_MAP_HOLE;
4906                         hole_em->block_len = 0;
4907                         hole_em->orig_block_len = 0;
4908                         hole_em->ram_bytes = hole_size;
4909                         hole_em->bdev = root->fs_info->fs_devices->latest_bdev;
4910                         hole_em->compress_type = BTRFS_COMPRESS_NONE;
4911                         hole_em->generation = root->fs_info->generation;
4912
4913                         while (1) {
4914                                 write_lock(&em_tree->lock);
4915                                 err = add_extent_mapping(em_tree, hole_em, 1);
4916                                 write_unlock(&em_tree->lock);
4917                                 if (err != -EEXIST)
4918                                         break;
4919                                 btrfs_drop_extent_cache(inode, cur_offset,
4920                                                         cur_offset +
4921                                                         hole_size - 1, 0);
4922                         }
4923                         free_extent_map(hole_em);
4924                 }
4925 next:
4926                 free_extent_map(em);
4927                 em = NULL;
4928                 cur_offset = last_byte;
4929                 if (cur_offset >= block_end)
4930                         break;
4931         }
4932         free_extent_map(em);
4933         unlock_extent_cached(io_tree, hole_start, block_end - 1, &cached_state,
4934                              GFP_NOFS);
4935         return err;
4936 }
4937
4938 static int btrfs_setsize(struct inode *inode, struct iattr *attr)
4939 {
4940         struct btrfs_root *root = BTRFS_I(inode)->root;
4941         struct btrfs_trans_handle *trans;
4942         loff_t oldsize = i_size_read(inode);
4943         loff_t newsize = attr->ia_size;
4944         int mask = attr->ia_valid;
4945         int ret;
4946
4947         /*
4948          * The regular truncate() case without ATTR_CTIME and ATTR_MTIME is a
4949          * special case where we need to update the times despite not having
4950          * these flags set.  For all other operations the VFS set these flags
4951          * explicitly if it wants a timestamp update.
4952          */
4953         if (newsize != oldsize) {
4954                 inode_inc_iversion(inode);
4955                 if (!(mask & (ATTR_CTIME | ATTR_MTIME)))
4956                         inode->i_ctime = inode->i_mtime =
4957                                 current_fs_time(inode->i_sb);
4958         }
4959
4960         if (newsize > oldsize) {
4961                 /*
4962                  * Don't do an expanding truncate while snapshoting is ongoing.
4963                  * This is to ensure the snapshot captures a fully consistent
4964                  * state of this file - if the snapshot captures this expanding
4965                  * truncation, it must capture all writes that happened before
4966                  * this truncation.
4967                  */
4968                 btrfs_wait_for_snapshot_creation(root);
4969                 ret = btrfs_cont_expand(inode, oldsize, newsize);
4970                 if (ret) {
4971                         btrfs_end_write_no_snapshoting(root);
4972                         return ret;
4973                 }
4974
4975                 trans = btrfs_start_transaction(root, 1);
4976                 if (IS_ERR(trans)) {
4977                         btrfs_end_write_no_snapshoting(root);
4978                         return PTR_ERR(trans);
4979                 }
4980
4981                 i_size_write(inode, newsize);
4982                 btrfs_ordered_update_i_size(inode, i_size_read(inode), NULL);
4983                 pagecache_isize_extended(inode, oldsize, newsize);
4984                 ret = btrfs_update_inode(trans, root, inode);
4985                 btrfs_end_write_no_snapshoting(root);
4986                 btrfs_end_transaction(trans, root);
4987         } else {
4988
4989                 /*
4990                  * We're truncating a file that used to have good data down to
4991                  * zero. Make sure it gets into the ordered flush list so that
4992                  * any new writes get down to disk quickly.
4993                  */
4994                 if (newsize == 0)
4995                         set_bit(BTRFS_INODE_ORDERED_DATA_CLOSE,
4996                                 &BTRFS_I(inode)->runtime_flags);
4997
4998                 /*
4999                  * 1 for the orphan item we're going to add
5000                  * 1 for the orphan item deletion.
5001                  */
5002                 trans = btrfs_start_transaction(root, 2);
5003                 if (IS_ERR(trans))
5004                         return PTR_ERR(trans);
5005
5006                 /*
5007                  * We need to do this in case we fail at _any_ point during the
5008                  * actual truncate.  Once we do the truncate_setsize we could
5009                  * invalidate pages which forces any outstanding ordered io to
5010                  * be instantly completed which will give us extents that need
5011                  * to be truncated.  If we fail to get an orphan inode down we
5012                  * could have left over extents that were never meant to live,
5013                  * so we need to guarantee from this point on that everything
5014                  * will be consistent.
5015                  */
5016                 ret = btrfs_orphan_add(trans, inode);
5017                 btrfs_end_transaction(trans, root);
5018                 if (ret)
5019                         return ret;
5020
5021                 /* we don't support swapfiles, so vmtruncate shouldn't fail */
5022                 truncate_setsize(inode, newsize);
5023
5024                 /* Disable nonlocked read DIO to avoid the end less truncate */
5025                 btrfs_inode_block_unlocked_dio(inode);
5026                 inode_dio_wait(inode);
5027                 btrfs_inode_resume_unlocked_dio(inode);
5028
5029                 ret = btrfs_truncate(inode);
5030                 if (ret && inode->i_nlink) {
5031                         int err;
5032
5033                         /*
5034                          * failed to truncate, disk_i_size is only adjusted down
5035                          * as we remove extents, so it should represent the true
5036                          * size of the inode, so reset the in memory size and
5037                          * delete our orphan entry.
5038                          */
5039                         trans = btrfs_join_transaction(root);
5040                         if (IS_ERR(trans)) {
5041                                 btrfs_orphan_del(NULL, inode);
5042                                 return ret;
5043                         }
5044                         i_size_write(inode, BTRFS_I(inode)->disk_i_size);
5045                         err = btrfs_orphan_del(trans, inode);
5046                         if (err)
5047                                 btrfs_abort_transaction(trans, root, err);
5048                         btrfs_end_transaction(trans, root);
5049                 }
5050         }
5051
5052         return ret;
5053 }
5054
5055 static int btrfs_setattr(struct dentry *dentry, struct iattr *attr)
5056 {
5057         struct inode *inode = d_inode(dentry);
5058         struct btrfs_root *root = BTRFS_I(inode)->root;
5059         int err;
5060
5061         if (btrfs_root_readonly(root))
5062                 return -EROFS;
5063
5064         err = inode_change_ok(inode, attr);
5065         if (err)
5066                 return err;
5067
5068         if (S_ISREG(inode->i_mode) && (attr->ia_valid & ATTR_SIZE)) {
5069                 err = btrfs_setsize(inode, attr);
5070                 if (err)
5071                         return err;
5072         }
5073
5074         if (attr->ia_valid) {
5075                 setattr_copy(inode, attr);
5076                 inode_inc_iversion(inode);
5077                 err = btrfs_dirty_inode(inode);
5078
5079                 if (!err && attr->ia_valid & ATTR_MODE)
5080                         err = posix_acl_chmod(inode, inode->i_mode);
5081         }
5082
5083         return err;
5084 }
5085
5086 /*
5087  * While truncating the inode pages during eviction, we get the VFS calling
5088  * btrfs_invalidatepage() against each page of the inode. This is slow because
5089  * the calls to btrfs_invalidatepage() result in a huge amount of calls to
5090  * lock_extent_bits() and clear_extent_bit(), which keep merging and splitting
5091  * extent_state structures over and over, wasting lots of time.
5092  *
5093  * Therefore if the inode is being evicted, let btrfs_invalidatepage() skip all
5094  * those expensive operations on a per page basis and do only the ordered io
5095  * finishing, while we release here the extent_map and extent_state structures,
5096  * without the excessive merging and splitting.
5097  */
5098 static void evict_inode_truncate_pages(struct inode *inode)
5099 {
5100         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
5101         struct extent_map_tree *map_tree = &BTRFS_I(inode)->extent_tree;
5102         struct rb_node *node;
5103
5104         ASSERT(inode->i_state & I_FREEING);
5105         truncate_inode_pages_final(&inode->i_data);
5106
5107         write_lock(&map_tree->lock);
5108         while (!RB_EMPTY_ROOT(&map_tree->map)) {
5109                 struct extent_map *em;
5110
5111                 node = rb_first(&map_tree->map);
5112                 em = rb_entry(node, struct extent_map, rb_node);
5113                 clear_bit(EXTENT_FLAG_PINNED, &em->flags);
5114                 clear_bit(EXTENT_FLAG_LOGGING, &em->flags);
5115                 remove_extent_mapping(map_tree, em);
5116                 free_extent_map(em);
5117                 if (need_resched()) {
5118                         write_unlock(&map_tree->lock);
5119                         cond_resched();
5120                         write_lock(&map_tree->lock);
5121                 }
5122         }
5123         write_unlock(&map_tree->lock);
5124
5125         /*
5126          * Keep looping until we have no more ranges in the io tree.
5127          * We can have ongoing bios started by readpages (called from readahead)
5128          * that have their endio callback (extent_io.c:end_bio_extent_readpage)
5129          * still in progress (unlocked the pages in the bio but did not yet
5130          * unlocked the ranges in the io tree). Therefore this means some
5131          * ranges can still be locked and eviction started because before
5132          * submitting those bios, which are executed by a separate task (work
5133          * queue kthread), inode references (inode->i_count) were not taken
5134          * (which would be dropped in the end io callback of each bio).
5135          * Therefore here we effectively end up waiting for those bios and
5136          * anyone else holding locked ranges without having bumped the inode's
5137          * reference count - if we don't do it, when they access the inode's
5138          * io_tree to unlock a range it may be too late, leading to an
5139          * use-after-free issue.
5140          */
5141         spin_lock(&io_tree->lock);
5142         while (!RB_EMPTY_ROOT(&io_tree->state)) {
5143                 struct extent_state *state;
5144                 struct extent_state *cached_state = NULL;
5145                 u64 start;
5146                 u64 end;
5147
5148                 node = rb_first(&io_tree->state);
5149                 state = rb_entry(node, struct extent_state, rb_node);
5150                 start = state->start;
5151                 end = state->end;
5152                 spin_unlock(&io_tree->lock);
5153
5154                 lock_extent_bits(io_tree, start, end, &cached_state);
5155
5156                 /*
5157                  * If still has DELALLOC flag, the extent didn't reach disk,
5158                  * and its reserved space won't be freed by delayed_ref.
5159                  * So we need to free its reserved space here.
5160                  * (Refer to comment in btrfs_invalidatepage, case 2)
5161                  *
5162                  * Note, end is the bytenr of last byte, so we need + 1 here.
5163                  */
5164                 if (state->state & EXTENT_DELALLOC)
5165                         btrfs_qgroup_free_data(inode, start, end - start + 1);
5166
5167                 clear_extent_bit(io_tree, start, end,
5168                                  EXTENT_LOCKED | EXTENT_DIRTY |
5169                                  EXTENT_DELALLOC | EXTENT_DO_ACCOUNTING |
5170                                  EXTENT_DEFRAG, 1, 1,
5171                                  &cached_state, GFP_NOFS);
5172
5173                 cond_resched();
5174                 spin_lock(&io_tree->lock);
5175         }
5176         spin_unlock(&io_tree->lock);
5177 }
5178
5179 void btrfs_evict_inode(struct inode *inode)
5180 {
5181         struct btrfs_trans_handle *trans;
5182         struct btrfs_root *root = BTRFS_I(inode)->root;
5183         struct btrfs_block_rsv *rsv, *global_rsv;
5184         int steal_from_global = 0;
5185         u64 min_size = btrfs_calc_trunc_metadata_size(root, 1);
5186         int ret;
5187
5188         trace_btrfs_inode_evict(inode);
5189
5190         evict_inode_truncate_pages(inode);
5191
5192         if (inode->i_nlink &&
5193             ((btrfs_root_refs(&root->root_item) != 0 &&
5194               root->root_key.objectid != BTRFS_ROOT_TREE_OBJECTID) ||
5195              btrfs_is_free_space_inode(inode)))
5196                 goto no_delete;
5197
5198         if (is_bad_inode(inode)) {
5199                 btrfs_orphan_del(NULL, inode);
5200                 goto no_delete;
5201         }
5202         /* do we really want it for ->i_nlink > 0 and zero btrfs_root_refs? */
5203         if (!special_file(inode->i_mode))
5204                 btrfs_wait_ordered_range(inode, 0, (u64)-1);
5205
5206         btrfs_free_io_failure_record(inode, 0, (u64)-1);
5207
5208         if (root->fs_info->log_root_recovering) {
5209                 BUG_ON(test_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
5210                                  &BTRFS_I(inode)->runtime_flags));
5211                 goto no_delete;
5212         }
5213
5214         if (inode->i_nlink > 0) {
5215                 BUG_ON(btrfs_root_refs(&root->root_item) != 0 &&
5216                        root->root_key.objectid != BTRFS_ROOT_TREE_OBJECTID);
5217                 goto no_delete;
5218         }
5219
5220         ret = btrfs_commit_inode_delayed_inode(inode);
5221         if (ret) {
5222                 btrfs_orphan_del(NULL, inode);
5223                 goto no_delete;
5224         }
5225
5226         rsv = btrfs_alloc_block_rsv(root, BTRFS_BLOCK_RSV_TEMP);
5227         if (!rsv) {
5228                 btrfs_orphan_del(NULL, inode);
5229                 goto no_delete;
5230         }
5231         rsv->size = min_size;
5232         rsv->failfast = 1;
5233         global_rsv = &root->fs_info->global_block_rsv;
5234
5235         btrfs_i_size_write(inode, 0);
5236
5237         /*
5238          * This is a bit simpler than btrfs_truncate since we've already
5239          * reserved our space for our orphan item in the unlink, so we just
5240          * need to reserve some slack space in case we add bytes and update
5241          * inode item when doing the truncate.
5242          */
5243         while (1) {
5244                 ret = btrfs_block_rsv_refill(root, rsv, min_size,
5245                                              BTRFS_RESERVE_FLUSH_LIMIT);
5246
5247                 /*
5248                  * Try and steal from the global reserve since we will
5249                  * likely not use this space anyway, we want to try as
5250                  * hard as possible to get this to work.
5251                  */
5252                 if (ret)
5253                         steal_from_global++;
5254                 else
5255                         steal_from_global = 0;
5256                 ret = 0;
5257
5258                 /*
5259                  * steal_from_global == 0: we reserved stuff, hooray!
5260                  * steal_from_global == 1: we didn't reserve stuff, boo!
5261                  * steal_from_global == 2: we've committed, still not a lot of
5262                  * room but maybe we'll have room in the global reserve this
5263                  * time.
5264                  * steal_from_global == 3: abandon all hope!
5265                  */
5266                 if (steal_from_global > 2) {
5267                         btrfs_warn(root->fs_info,
5268                                 "Could not get space for a delete, will truncate on mount %d",
5269                                 ret);
5270                         btrfs_orphan_del(NULL, inode);
5271                         btrfs_free_block_rsv(root, rsv);
5272                         goto no_delete;
5273                 }
5274
5275                 trans = btrfs_join_transaction(root);
5276                 if (IS_ERR(trans)) {
5277                         btrfs_orphan_del(NULL, inode);
5278                         btrfs_free_block_rsv(root, rsv);
5279                         goto no_delete;
5280                 }
5281
5282                 /*
5283                  * We can't just steal from the global reserve, we need to make
5284                  * sure there is room to do it, if not we need to commit and try
5285                  * again.
5286                  */
5287                 if (steal_from_global) {
5288                         if (!btrfs_check_space_for_delayed_refs(trans, root))
5289                                 ret = btrfs_block_rsv_migrate(global_rsv, rsv,
5290                                                               min_size, 0);
5291                         else
5292                                 ret = -ENOSPC;
5293                 }
5294
5295                 /*
5296                  * Couldn't steal from the global reserve, we have too much
5297                  * pending stuff built up, commit the transaction and try it
5298                  * again.
5299                  */
5300                 if (ret) {
5301                         ret = btrfs_commit_transaction(trans, root);
5302                         if (ret) {
5303                                 btrfs_orphan_del(NULL, inode);
5304                                 btrfs_free_block_rsv(root, rsv);
5305                                 goto no_delete;
5306                         }
5307                         continue;
5308                 } else {
5309                         steal_from_global = 0;
5310                 }
5311
5312                 trans->block_rsv = rsv;
5313
5314                 ret = btrfs_truncate_inode_items(trans, root, inode, 0, 0);
5315                 if (ret != -ENOSPC && ret != -EAGAIN)
5316                         break;
5317
5318                 trans->block_rsv = &root->fs_info->trans_block_rsv;
5319                 btrfs_end_transaction(trans, root);
5320                 trans = NULL;
5321                 btrfs_btree_balance_dirty(root);
5322         }
5323
5324         btrfs_free_block_rsv(root, rsv);
5325
5326         /*
5327          * Errors here aren't a big deal, it just means we leave orphan items
5328          * in the tree.  They will be cleaned up on the next mount.
5329          */
5330         if (ret == 0) {
5331                 trans->block_rsv = root->orphan_block_rsv;
5332                 btrfs_orphan_del(trans, inode);
5333         } else {
5334                 btrfs_orphan_del(NULL, inode);
5335         }
5336
5337         trans->block_rsv = &root->fs_info->trans_block_rsv;
5338         if (!(root == root->fs_info->tree_root ||
5339               root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID))
5340                 btrfs_return_ino(root, btrfs_ino(inode));
5341
5342         btrfs_end_transaction(trans, root);
5343         btrfs_btree_balance_dirty(root);
5344 no_delete:
5345         btrfs_remove_delayed_node(inode);
5346         clear_inode(inode);
5347 }
5348
5349 /*
5350  * this returns the key found in the dir entry in the location pointer.
5351  * If no dir entries were found, location->objectid is 0.
5352  */
5353 static int btrfs_inode_by_name(struct inode *dir, struct dentry *dentry,
5354                                struct btrfs_key *location)
5355 {
5356         const char *name = dentry->d_name.name;
5357         int namelen = dentry->d_name.len;
5358         struct btrfs_dir_item *di;
5359         struct btrfs_path *path;
5360         struct btrfs_root *root = BTRFS_I(dir)->root;
5361         int ret = 0;
5362
5363         path = btrfs_alloc_path();
5364         if (!path)
5365                 return -ENOMEM;
5366
5367         di = btrfs_lookup_dir_item(NULL, root, path, btrfs_ino(dir), name,
5368                                     namelen, 0);
5369         if (IS_ERR(di))
5370                 ret = PTR_ERR(di);
5371
5372         if (IS_ERR_OR_NULL(di))
5373                 goto out_err;
5374
5375         btrfs_dir_item_key_to_cpu(path->nodes[0], di, location);
5376 out:
5377         btrfs_free_path(path);
5378         return ret;
5379 out_err:
5380         location->objectid = 0;
5381         goto out;
5382 }
5383
5384 /*
5385  * when we hit a tree root in a directory, the btrfs part of the inode
5386  * needs to be changed to reflect the root directory of the tree root.  This
5387  * is kind of like crossing a mount point.
5388  */
5389 static int fixup_tree_root_location(struct btrfs_root *root,
5390                                     struct inode *dir,
5391                                     struct dentry *dentry,
5392                                     struct btrfs_key *location,
5393                                     struct btrfs_root **sub_root)
5394 {
5395         struct btrfs_path *path;
5396         struct btrfs_root *new_root;
5397         struct btrfs_root_ref *ref;
5398         struct extent_buffer *leaf;
5399         struct btrfs_key key;
5400         int ret;
5401         int err = 0;
5402
5403         path = btrfs_alloc_path();
5404         if (!path) {
5405                 err = -ENOMEM;
5406                 goto out;
5407         }
5408
5409         err = -ENOENT;
5410         key.objectid = BTRFS_I(dir)->root->root_key.objectid;
5411         key.type = BTRFS_ROOT_REF_KEY;
5412         key.offset = location->objectid;
5413
5414         ret = btrfs_search_slot(NULL, root->fs_info->tree_root, &key, path,
5415                                 0, 0);
5416         if (ret) {
5417                 if (ret < 0)
5418                         err = ret;
5419                 goto out;
5420         }
5421
5422         leaf = path->nodes[0];
5423         ref = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_root_ref);
5424         if (btrfs_root_ref_dirid(leaf, ref) != btrfs_ino(dir) ||
5425             btrfs_root_ref_name_len(leaf, ref) != dentry->d_name.len)
5426                 goto out;
5427
5428         ret = memcmp_extent_buffer(leaf, dentry->d_name.name,
5429                                    (unsigned long)(ref + 1),
5430                                    dentry->d_name.len);
5431         if (ret)
5432                 goto out;
5433
5434         btrfs_release_path(path);
5435
5436         new_root = btrfs_read_fs_root_no_name(root->fs_info, location);
5437         if (IS_ERR(new_root)) {
5438                 err = PTR_ERR(new_root);
5439                 goto out;
5440         }
5441
5442         *sub_root = new_root;
5443         location->objectid = btrfs_root_dirid(&new_root->root_item);
5444         location->type = BTRFS_INODE_ITEM_KEY;
5445         location->offset = 0;
5446         err = 0;
5447 out:
5448         btrfs_free_path(path);
5449         return err;
5450 }
5451
5452 static void inode_tree_add(struct inode *inode)
5453 {
5454         struct btrfs_root *root = BTRFS_I(inode)->root;
5455         struct btrfs_inode *entry;
5456         struct rb_node **p;
5457         struct rb_node *parent;
5458         struct rb_node *new = &BTRFS_I(inode)->rb_node;
5459         u64 ino = btrfs_ino(inode);
5460
5461         if (inode_unhashed(inode))
5462                 return;
5463         parent = NULL;
5464         spin_lock(&root->inode_lock);
5465         p = &root->inode_tree.rb_node;
5466         while (*p) {
5467                 parent = *p;
5468                 entry = rb_entry(parent, struct btrfs_inode, rb_node);
5469
5470                 if (ino < btrfs_ino(&entry->vfs_inode))
5471                         p = &parent->rb_left;
5472                 else if (ino > btrfs_ino(&entry->vfs_inode))
5473                         p = &parent->rb_right;
5474                 else {
5475                         WARN_ON(!(entry->vfs_inode.i_state &
5476                                   (I_WILL_FREE | I_FREEING)));
5477                         rb_replace_node(parent, new, &root->inode_tree);
5478                         RB_CLEAR_NODE(parent);
5479                         spin_unlock(&root->inode_lock);
5480                         return;
5481                 }
5482         }
5483         rb_link_node(new, parent, p);
5484         rb_insert_color(new, &root->inode_tree);
5485         spin_unlock(&root->inode_lock);
5486 }
5487
5488 static void inode_tree_del(struct inode *inode)
5489 {
5490         struct btrfs_root *root = BTRFS_I(inode)->root;
5491         int empty = 0;
5492
5493         spin_lock(&root->inode_lock);
5494         if (!RB_EMPTY_NODE(&BTRFS_I(inode)->rb_node)) {
5495                 rb_erase(&BTRFS_I(inode)->rb_node, &root->inode_tree);
5496                 RB_CLEAR_NODE(&BTRFS_I(inode)->rb_node);
5497                 empty = RB_EMPTY_ROOT(&root->inode_tree);
5498         }
5499         spin_unlock(&root->inode_lock);
5500
5501         if (empty && btrfs_root_refs(&root->root_item) == 0) {
5502                 synchronize_srcu(&root->fs_info->subvol_srcu);
5503                 spin_lock(&root->inode_lock);
5504                 empty = RB_EMPTY_ROOT(&root->inode_tree);
5505                 spin_unlock(&root->inode_lock);
5506                 if (empty)
5507                         btrfs_add_dead_root(root);
5508         }
5509 }
5510
5511 void btrfs_invalidate_inodes(struct btrfs_root *root)
5512 {
5513         struct rb_node *node;
5514         struct rb_node *prev;
5515         struct btrfs_inode *entry;
5516         struct inode *inode;
5517         u64 objectid = 0;
5518
5519         if (!test_bit(BTRFS_FS_STATE_ERROR, &root->fs_info->fs_state))
5520                 WARN_ON(btrfs_root_refs(&root->root_item) != 0);
5521
5522         spin_lock(&root->inode_lock);
5523 again:
5524         node = root->inode_tree.rb_node;
5525         prev = NULL;
5526         while (node) {
5527                 prev = node;
5528                 entry = rb_entry(node, struct btrfs_inode, rb_node);
5529
5530                 if (objectid < btrfs_ino(&entry->vfs_inode))
5531                         node = node->rb_left;
5532                 else if (objectid > btrfs_ino(&entry->vfs_inode))
5533                         node = node->rb_right;
5534                 else
5535                         break;
5536         }
5537         if (!node) {
5538                 while (prev) {
5539                         entry = rb_entry(prev, struct btrfs_inode, rb_node);
5540                         if (objectid <= btrfs_ino(&entry->vfs_inode)) {
5541                                 node = prev;
5542                                 break;
5543                         }
5544                         prev = rb_next(prev);
5545                 }
5546         }
5547         while (node) {
5548                 entry = rb_entry(node, struct btrfs_inode, rb_node);
5549                 objectid = btrfs_ino(&entry->vfs_inode) + 1;
5550                 inode = igrab(&entry->vfs_inode);
5551                 if (inode) {
5552                         spin_unlock(&root->inode_lock);
5553                         if (atomic_read(&inode->i_count) > 1)
5554                                 d_prune_aliases(inode);
5555                         /*
5556                          * btrfs_drop_inode will have it removed from
5557                          * the inode cache when its usage count
5558                          * hits zero.
5559                          */
5560                         iput(inode);
5561                         cond_resched();
5562                         spin_lock(&root->inode_lock);
5563                         goto again;
5564                 }
5565
5566                 if (cond_resched_lock(&root->inode_lock))
5567                         goto again;
5568
5569                 node = rb_next(node);
5570         }
5571         spin_unlock(&root->inode_lock);
5572 }
5573
5574 static int btrfs_init_locked_inode(struct inode *inode, void *p)
5575 {
5576         struct btrfs_iget_args *args = p;
5577         inode->i_ino = args->location->objectid;
5578         memcpy(&BTRFS_I(inode)->location, args->location,
5579                sizeof(*args->location));
5580         BTRFS_I(inode)->root = args->root;
5581         return 0;
5582 }
5583
5584 static int btrfs_find_actor(struct inode *inode, void *opaque)
5585 {
5586         struct btrfs_iget_args *args = opaque;
5587         return args->location->objectid == BTRFS_I(inode)->location.objectid &&
5588                 args->root == BTRFS_I(inode)->root;
5589 }
5590
5591 static struct inode *btrfs_iget_locked(struct super_block *s,
5592                                        struct btrfs_key *location,
5593                                        struct btrfs_root *root)
5594 {
5595         struct inode *inode;
5596         struct btrfs_iget_args args;
5597         unsigned long hashval = btrfs_inode_hash(location->objectid, root);
5598
5599         args.location = location;
5600         args.root = root;
5601
5602         inode = iget5_locked(s, hashval, btrfs_find_actor,
5603                              btrfs_init_locked_inode,
5604                              (void *)&args);
5605         return inode;
5606 }
5607
5608 /* Get an inode object given its location and corresponding root.
5609  * Returns in *is_new if the inode was read from disk
5610  */
5611 struct inode *btrfs_iget(struct super_block *s, struct btrfs_key *location,
5612                          struct btrfs_root *root, int *new)
5613 {
5614         struct inode *inode;
5615
5616         inode = btrfs_iget_locked(s, location, root);
5617         if (!inode)
5618                 return ERR_PTR(-ENOMEM);
5619
5620         if (inode->i_state & I_NEW) {
5621                 int ret;
5622
5623                 ret = btrfs_read_locked_inode(inode);
5624                 if (!is_bad_inode(inode)) {
5625                         inode_tree_add(inode);
5626                         unlock_new_inode(inode);
5627                         if (new)
5628                                 *new = 1;
5629                 } else {
5630                         unlock_new_inode(inode);
5631                         iput(inode);
5632                         ASSERT(ret < 0);
5633                         inode = ERR_PTR(ret < 0 ? ret : -ESTALE);
5634                 }
5635         }
5636
5637         return inode;
5638 }
5639
5640 static struct inode *new_simple_dir(struct super_block *s,
5641                                     struct btrfs_key *key,
5642                                     struct btrfs_root *root)
5643 {
5644         struct inode *inode = new_inode(s);
5645
5646         if (!inode)
5647                 return ERR_PTR(-ENOMEM);
5648
5649         BTRFS_I(inode)->root = root;
5650         memcpy(&BTRFS_I(inode)->location, key, sizeof(*key));
5651         set_bit(BTRFS_INODE_DUMMY, &BTRFS_I(inode)->runtime_flags);
5652
5653         inode->i_ino = BTRFS_EMPTY_SUBVOL_DIR_OBJECTID;
5654         inode->i_op = &btrfs_dir_ro_inode_operations;
5655         inode->i_fop = &simple_dir_operations;
5656         inode->i_mode = S_IFDIR | S_IRUGO | S_IWUSR | S_IXUGO;
5657         inode->i_mtime = current_fs_time(inode->i_sb);
5658         inode->i_atime = inode->i_mtime;
5659         inode->i_ctime = inode->i_mtime;
5660         BTRFS_I(inode)->i_otime = inode->i_mtime;
5661
5662         return inode;
5663 }
5664
5665 struct inode *btrfs_lookup_dentry(struct inode *dir, struct dentry *dentry)
5666 {
5667         struct inode *inode;
5668         struct btrfs_root *root = BTRFS_I(dir)->root;
5669         struct btrfs_root *sub_root = root;
5670         struct btrfs_key location;
5671         int index;
5672         int ret = 0;
5673
5674         if (dentry->d_name.len > BTRFS_NAME_LEN)
5675                 return ERR_PTR(-ENAMETOOLONG);
5676
5677         ret = btrfs_inode_by_name(dir, dentry, &location);
5678         if (ret < 0)
5679                 return ERR_PTR(ret);
5680
5681         if (location.objectid == 0)
5682                 return ERR_PTR(-ENOENT);
5683
5684         if (location.type == BTRFS_INODE_ITEM_KEY) {
5685                 inode = btrfs_iget(dir->i_sb, &location, root, NULL);
5686                 return inode;
5687         }
5688
5689         BUG_ON(location.type != BTRFS_ROOT_ITEM_KEY);
5690
5691         index = srcu_read_lock(&root->fs_info->subvol_srcu);
5692         ret = fixup_tree_root_location(root, dir, dentry,
5693                                        &location, &sub_root);
5694         if (ret < 0) {
5695                 if (ret != -ENOENT)
5696                         inode = ERR_PTR(ret);
5697                 else
5698                         inode = new_simple_dir(dir->i_sb, &location, sub_root);
5699         } else {
5700                 inode = btrfs_iget(dir->i_sb, &location, sub_root, NULL);
5701         }
5702         srcu_read_unlock(&root->fs_info->subvol_srcu, index);
5703
5704         if (!IS_ERR(inode) && root != sub_root) {
5705                 down_read(&root->fs_info->cleanup_work_sem);
5706                 if (!(inode->i_sb->s_flags & MS_RDONLY))
5707                         ret = btrfs_orphan_cleanup(sub_root);
5708                 up_read(&root->fs_info->cleanup_work_sem);
5709                 if (ret) {
5710                         iput(inode);
5711                         inode = ERR_PTR(ret);
5712                 }
5713         }
5714
5715         return inode;
5716 }
5717
5718 static int btrfs_dentry_delete(const struct dentry *dentry)
5719 {
5720         struct btrfs_root *root;
5721         struct inode *inode = d_inode(dentry);
5722
5723         if (!inode && !IS_ROOT(dentry))
5724                 inode = d_inode(dentry->d_parent);
5725
5726         if (inode) {
5727                 root = BTRFS_I(inode)->root;
5728                 if (btrfs_root_refs(&root->root_item) == 0)
5729                         return 1;
5730
5731                 if (btrfs_ino(inode) == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)
5732                         return 1;
5733         }
5734         return 0;
5735 }
5736
5737 static void btrfs_dentry_release(struct dentry *dentry)
5738 {
5739         kfree(dentry->d_fsdata);
5740 }
5741
5742 static struct dentry *btrfs_lookup(struct inode *dir, struct dentry *dentry,
5743                                    unsigned int flags)
5744 {
5745         struct inode *inode;
5746
5747         inode = btrfs_lookup_dentry(dir, dentry);
5748         if (IS_ERR(inode)) {
5749                 if (PTR_ERR(inode) == -ENOENT)
5750                         inode = NULL;
5751                 else
5752                         return ERR_CAST(inode);
5753         }
5754
5755         return d_splice_alias(inode, dentry);
5756 }
5757
5758 unsigned char btrfs_filetype_table[] = {
5759         DT_UNKNOWN, DT_REG, DT_DIR, DT_CHR, DT_BLK, DT_FIFO, DT_SOCK, DT_LNK
5760 };
5761
5762 static int btrfs_real_readdir(struct file *file, struct dir_context *ctx)
5763 {
5764         struct inode *inode = file_inode(file);
5765         struct btrfs_root *root = BTRFS_I(inode)->root;
5766         struct btrfs_item *item;
5767         struct btrfs_dir_item *di;
5768         struct btrfs_key key;
5769         struct btrfs_key found_key;
5770         struct btrfs_path *path;
5771         struct list_head ins_list;
5772         struct list_head del_list;
5773         int ret;
5774         struct extent_buffer *leaf;
5775         int slot;
5776         unsigned char d_type;
5777         int over = 0;
5778         u32 di_cur;
5779         u32 di_total;
5780         u32 di_len;
5781         int key_type = BTRFS_DIR_INDEX_KEY;
5782         char tmp_name[32];
5783         char *name_ptr;
5784         int name_len;
5785         int is_curr = 0;        /* ctx->pos points to the current index? */
5786         bool emitted;
5787         bool put = false;
5788
5789         /* FIXME, use a real flag for deciding about the key type */
5790         if (root->fs_info->tree_root == root)
5791                 key_type = BTRFS_DIR_ITEM_KEY;
5792
5793         if (!dir_emit_dots(file, ctx))
5794                 return 0;
5795
5796         path = btrfs_alloc_path();
5797         if (!path)
5798                 return -ENOMEM;
5799
5800         path->reada = READA_FORWARD;
5801
5802         if (key_type == BTRFS_DIR_INDEX_KEY) {
5803                 INIT_LIST_HEAD(&ins_list);
5804                 INIT_LIST_HEAD(&del_list);
5805                 put = btrfs_readdir_get_delayed_items(inode, &ins_list,
5806                                                       &del_list);
5807         }
5808
5809         key.type = key_type;
5810         key.offset = ctx->pos;
5811         key.objectid = btrfs_ino(inode);
5812
5813         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
5814         if (ret < 0)
5815                 goto err;
5816
5817         emitted = false;
5818         while (1) {
5819                 leaf = path->nodes[0];
5820                 slot = path->slots[0];
5821                 if (slot >= btrfs_header_nritems(leaf)) {
5822                         ret = btrfs_next_leaf(root, path);
5823                         if (ret < 0)
5824                                 goto err;
5825                         else if (ret > 0)
5826                                 break;
5827                         continue;
5828                 }
5829
5830                 item = btrfs_item_nr(slot);
5831                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
5832
5833                 if (found_key.objectid != key.objectid)
5834                         break;
5835                 if (found_key.type != key_type)
5836                         break;
5837                 if (found_key.offset < ctx->pos)
5838                         goto next;
5839                 if (key_type == BTRFS_DIR_INDEX_KEY &&
5840                     btrfs_should_delete_dir_index(&del_list,
5841                                                   found_key.offset))
5842                         goto next;
5843
5844                 ctx->pos = found_key.offset;
5845                 is_curr = 1;
5846
5847                 di = btrfs_item_ptr(leaf, slot, struct btrfs_dir_item);
5848                 di_cur = 0;
5849                 di_total = btrfs_item_size(leaf, item);
5850
5851                 while (di_cur < di_total) {
5852                         struct btrfs_key location;
5853
5854                         if (verify_dir_item(root, leaf, di))
5855                                 break;
5856
5857                         name_len = btrfs_dir_name_len(leaf, di);
5858                         if (name_len <= sizeof(tmp_name)) {
5859                                 name_ptr = tmp_name;
5860                         } else {
5861                                 name_ptr = kmalloc(name_len, GFP_KERNEL);
5862                                 if (!name_ptr) {
5863                                         ret = -ENOMEM;
5864                                         goto err;
5865                                 }
5866                         }
5867                         read_extent_buffer(leaf, name_ptr,
5868                                            (unsigned long)(di + 1), name_len);
5869
5870                         d_type = btrfs_filetype_table[btrfs_dir_type(leaf, di)];
5871                         btrfs_dir_item_key_to_cpu(leaf, di, &location);
5872
5873
5874                         /* is this a reference to our own snapshot? If so
5875                          * skip it.
5876                          *
5877                          * In contrast to old kernels, we insert the snapshot's
5878                          * dir item and dir index after it has been created, so
5879                          * we won't find a reference to our own snapshot. We
5880                          * still keep the following code for backward
5881                          * compatibility.
5882                          */
5883                         if (location.type == BTRFS_ROOT_ITEM_KEY &&
5884                             location.objectid == root->root_key.objectid) {
5885                                 over = 0;
5886                                 goto skip;
5887                         }
5888                         over = !dir_emit(ctx, name_ptr, name_len,
5889                                        location.objectid, d_type);
5890
5891 skip:
5892                         if (name_ptr != tmp_name)
5893                                 kfree(name_ptr);
5894
5895                         if (over)
5896                                 goto nopos;
5897                         emitted = true;
5898                         di_len = btrfs_dir_name_len(leaf, di) +
5899                                  btrfs_dir_data_len(leaf, di) + sizeof(*di);
5900                         di_cur += di_len;
5901                         di = (struct btrfs_dir_item *)((char *)di + di_len);
5902                 }
5903 next:
5904                 path->slots[0]++;
5905         }
5906
5907         if (key_type == BTRFS_DIR_INDEX_KEY) {
5908                 if (is_curr)
5909                         ctx->pos++;
5910                 ret = btrfs_readdir_delayed_dir_index(ctx, &ins_list, &emitted);
5911                 if (ret)
5912                         goto nopos;
5913         }
5914
5915         /*
5916          * If we haven't emitted any dir entry, we must not touch ctx->pos as
5917          * it was was set to the termination value in previous call. We assume
5918          * that "." and ".." were emitted if we reach this point and set the
5919          * termination value as well for an empty directory.
5920          */
5921         if (ctx->pos > 2 && !emitted)
5922                 goto nopos;
5923
5924         /* Reached end of directory/root. Bump pos past the last item. */
5925         ctx->pos++;
5926
5927         /*
5928          * Stop new entries from being returned after we return the last
5929          * entry.
5930          *
5931          * New directory entries are assigned a strictly increasing
5932          * offset.  This means that new entries created during readdir
5933          * are *guaranteed* to be seen in the future by that readdir.
5934          * This has broken buggy programs which operate on names as
5935          * they're returned by readdir.  Until we re-use freed offsets
5936          * we have this hack to stop new entries from being returned
5937          * under the assumption that they'll never reach this huge
5938          * offset.
5939          *
5940          * This is being careful not to overflow 32bit loff_t unless the
5941          * last entry requires it because doing so has broken 32bit apps
5942          * in the past.
5943          */
5944         if (key_type == BTRFS_DIR_INDEX_KEY) {
5945                 if (ctx->pos >= INT_MAX)
5946                         ctx->pos = LLONG_MAX;
5947                 else
5948                         ctx->pos = INT_MAX;
5949         }
5950 nopos:
5951         ret = 0;
5952 err:
5953         if (put)
5954                 btrfs_readdir_put_delayed_items(inode, &ins_list, &del_list);
5955         btrfs_free_path(path);
5956         return ret;
5957 }
5958
5959 int btrfs_write_inode(struct inode *inode, struct writeback_control *wbc)
5960 {
5961         struct btrfs_root *root = BTRFS_I(inode)->root;
5962         struct btrfs_trans_handle *trans;
5963         int ret = 0;
5964         bool nolock = false;
5965
5966         if (test_bit(BTRFS_INODE_DUMMY, &BTRFS_I(inode)->runtime_flags))
5967                 return 0;
5968
5969         if (btrfs_fs_closing(root->fs_info) && btrfs_is_free_space_inode(inode))
5970                 nolock = true;
5971
5972         if (wbc->sync_mode == WB_SYNC_ALL) {
5973                 if (nolock)
5974                         trans = btrfs_join_transaction_nolock(root);
5975                 else
5976                         trans = btrfs_join_transaction(root);
5977                 if (IS_ERR(trans))
5978                         return PTR_ERR(trans);
5979                 ret = btrfs_commit_transaction(trans, root);
5980         }
5981         return ret;
5982 }
5983
5984 /*
5985  * This is somewhat expensive, updating the tree every time the
5986  * inode changes.  But, it is most likely to find the inode in cache.
5987  * FIXME, needs more benchmarking...there are no reasons other than performance
5988  * to keep or drop this code.
5989  */
5990 static int btrfs_dirty_inode(struct inode *inode)
5991 {
5992         struct btrfs_root *root = BTRFS_I(inode)->root;
5993         struct btrfs_trans_handle *trans;
5994         int ret;
5995
5996         if (test_bit(BTRFS_INODE_DUMMY, &BTRFS_I(inode)->runtime_flags))
5997                 return 0;
5998
5999         trans = btrfs_join_transaction(root);
6000         if (IS_ERR(trans))
6001                 return PTR_ERR(trans);
6002
6003         ret = btrfs_update_inode(trans, root, inode);
6004         if (ret && ret == -ENOSPC) {
6005                 /* whoops, lets try again with the full transaction */
6006                 btrfs_end_transaction(trans, root);
6007                 trans = btrfs_start_transaction(root, 1);
6008                 if (IS_ERR(trans))
6009                         return PTR_ERR(trans);
6010
6011                 ret = btrfs_update_inode(trans, root, inode);
6012         }
6013         btrfs_end_transaction(trans, root);
6014         if (BTRFS_I(inode)->delayed_node)
6015                 btrfs_balance_delayed_items(root);
6016
6017         return ret;
6018 }
6019
6020 /*
6021  * This is a copy of file_update_time.  We need this so we can return error on
6022  * ENOSPC for updating the inode in the case of file write and mmap writes.
6023  */
6024 static int btrfs_update_time(struct inode *inode, struct timespec *now,
6025                              int flags)
6026 {
6027         struct btrfs_root *root = BTRFS_I(inode)->root;
6028
6029         if (btrfs_root_readonly(root))
6030                 return -EROFS;
6031
6032         if (flags & S_VERSION)
6033                 inode_inc_iversion(inode);
6034         if (flags & S_CTIME)
6035                 inode->i_ctime = *now;
6036         if (flags & S_MTIME)
6037                 inode->i_mtime = *now;
6038         if (flags & S_ATIME)
6039                 inode->i_atime = *now;
6040         return btrfs_dirty_inode(inode);
6041 }
6042
6043 /*
6044  * find the highest existing sequence number in a directory
6045  * and then set the in-memory index_cnt variable to reflect
6046  * free sequence numbers
6047  */
6048 static int btrfs_set_inode_index_count(struct inode *inode)
6049 {
6050         struct btrfs_root *root = BTRFS_I(inode)->root;
6051         struct btrfs_key key, found_key;
6052         struct btrfs_path *path;
6053         struct extent_buffer *leaf;
6054         int ret;
6055
6056         key.objectid = btrfs_ino(inode);
6057         key.type = BTRFS_DIR_INDEX_KEY;
6058         key.offset = (u64)-1;
6059
6060         path = btrfs_alloc_path();
6061         if (!path)
6062                 return -ENOMEM;
6063
6064         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
6065         if (ret < 0)
6066                 goto out;
6067         /* FIXME: we should be able to handle this */
6068         if (ret == 0)
6069                 goto out;
6070         ret = 0;
6071
6072         /*
6073          * MAGIC NUMBER EXPLANATION:
6074          * since we search a directory based on f_pos we have to start at 2
6075          * since '.' and '..' have f_pos of 0 and 1 respectively, so everybody
6076          * else has to start at 2
6077          */
6078         if (path->slots[0] == 0) {
6079                 BTRFS_I(inode)->index_cnt = 2;
6080                 goto out;
6081         }
6082
6083         path->slots[0]--;
6084
6085         leaf = path->nodes[0];
6086         btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
6087
6088         if (found_key.objectid != btrfs_ino(inode) ||
6089             found_key.type != BTRFS_DIR_INDEX_KEY) {
6090                 BTRFS_I(inode)->index_cnt = 2;
6091                 goto out;
6092         }
6093
6094         BTRFS_I(inode)->index_cnt = found_key.offset + 1;
6095 out:
6096         btrfs_free_path(path);
6097         return ret;
6098 }
6099
6100 /*
6101  * helper to find a free sequence number in a given directory.  This current
6102  * code is very simple, later versions will do smarter things in the btree
6103  */
6104 int btrfs_set_inode_index(struct inode *dir, u64 *index)
6105 {
6106         int ret = 0;
6107
6108         if (BTRFS_I(dir)->index_cnt == (u64)-1) {
6109                 ret = btrfs_inode_delayed_dir_index_count(dir);
6110                 if (ret) {
6111                         ret = btrfs_set_inode_index_count(dir);
6112                         if (ret)
6113                                 return ret;
6114                 }
6115         }
6116
6117         *index = BTRFS_I(dir)->index_cnt;
6118         BTRFS_I(dir)->index_cnt++;
6119
6120         return ret;
6121 }
6122
6123 static int btrfs_insert_inode_locked(struct inode *inode)
6124 {
6125         struct btrfs_iget_args args;
6126         args.location = &BTRFS_I(inode)->location;
6127         args.root = BTRFS_I(inode)->root;
6128
6129         return insert_inode_locked4(inode,
6130                    btrfs_inode_hash(inode->i_ino, BTRFS_I(inode)->root),
6131                    btrfs_find_actor, &args);
6132 }
6133
6134 static struct inode *btrfs_new_inode(struct btrfs_trans_handle *trans,
6135                                      struct btrfs_root *root,
6136                                      struct inode *dir,
6137                                      const char *name, int name_len,
6138                                      u64 ref_objectid, u64 objectid,
6139                                      umode_t mode, u64 *index)
6140 {
6141         struct inode *inode;
6142         struct btrfs_inode_item *inode_item;
6143         struct btrfs_key *location;
6144         struct btrfs_path *path;
6145         struct btrfs_inode_ref *ref;
6146         struct btrfs_key key[2];
6147         u32 sizes[2];
6148         int nitems = name ? 2 : 1;
6149         unsigned long ptr;
6150         int ret;
6151
6152         path = btrfs_alloc_path();
6153         if (!path)
6154                 return ERR_PTR(-ENOMEM);
6155
6156         inode = new_inode(root->fs_info->sb);
6157         if (!inode) {
6158                 btrfs_free_path(path);
6159                 return ERR_PTR(-ENOMEM);
6160         }
6161
6162         /*
6163          * O_TMPFILE, set link count to 0, so that after this point,
6164          * we fill in an inode item with the correct link count.
6165          */
6166         if (!name)
6167                 set_nlink(inode, 0);
6168
6169         /*
6170          * we have to initialize this early, so we can reclaim the inode
6171          * number if we fail afterwards in this function.
6172          */
6173         inode->i_ino = objectid;
6174
6175         if (dir && name) {
6176                 trace_btrfs_inode_request(dir);
6177
6178                 ret = btrfs_set_inode_index(dir, index);
6179                 if (ret) {
6180                         btrfs_free_path(path);
6181                         iput(inode);
6182                         return ERR_PTR(ret);
6183                 }
6184         } else if (dir) {
6185                 *index = 0;
6186         }
6187         /*
6188          * index_cnt is ignored for everything but a dir,
6189          * btrfs_get_inode_index_count has an explanation for the magic
6190          * number
6191          */
6192         BTRFS_I(inode)->index_cnt = 2;
6193         BTRFS_I(inode)->dir_index = *index;
6194         BTRFS_I(inode)->root = root;
6195         BTRFS_I(inode)->generation = trans->transid;
6196         inode->i_generation = BTRFS_I(inode)->generation;
6197
6198         /*
6199          * We could have gotten an inode number from somebody who was fsynced
6200          * and then removed in this same transaction, so let's just set full
6201          * sync since it will be a full sync anyway and this will blow away the
6202          * old info in the log.
6203          */
6204         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &BTRFS_I(inode)->runtime_flags);
6205
6206         key[0].objectid = objectid;
6207         key[0].type = BTRFS_INODE_ITEM_KEY;
6208         key[0].offset = 0;
6209
6210         sizes[0] = sizeof(struct btrfs_inode_item);
6211
6212         if (name) {
6213                 /*
6214                  * Start new inodes with an inode_ref. This is slightly more
6215                  * efficient for small numbers of hard links since they will
6216                  * be packed into one item. Extended refs will kick in if we
6217                  * add more hard links than can fit in the ref item.
6218                  */
6219                 key[1].objectid = objectid;
6220                 key[1].type = BTRFS_INODE_REF_KEY;
6221                 key[1].offset = ref_objectid;
6222
6223                 sizes[1] = name_len + sizeof(*ref);
6224         }
6225
6226         location = &BTRFS_I(inode)->location;
6227         location->objectid = objectid;
6228         location->offset = 0;
6229         location->type = BTRFS_INODE_ITEM_KEY;
6230
6231         ret = btrfs_insert_inode_locked(inode);
6232         if (ret < 0)
6233                 goto fail;
6234
6235         path->leave_spinning = 1;
6236         ret = btrfs_insert_empty_items(trans, root, path, key, sizes, nitems);
6237         if (ret != 0)
6238                 goto fail_unlock;
6239
6240         inode_init_owner(inode, dir, mode);
6241         inode_set_bytes(inode, 0);
6242
6243         inode->i_mtime = current_fs_time(inode->i_sb);
6244         inode->i_atime = inode->i_mtime;
6245         inode->i_ctime = inode->i_mtime;
6246         BTRFS_I(inode)->i_otime = inode->i_mtime;
6247
6248         inode_item = btrfs_item_ptr(path->nodes[0], path->slots[0],
6249                                   struct btrfs_inode_item);
6250         memset_extent_buffer(path->nodes[0], 0, (unsigned long)inode_item,
6251                              sizeof(*inode_item));
6252         fill_inode_item(trans, path->nodes[0], inode_item, inode);
6253
6254         if (name) {
6255                 ref = btrfs_item_ptr(path->nodes[0], path->slots[0] + 1,
6256                                      struct btrfs_inode_ref);
6257                 btrfs_set_inode_ref_name_len(path->nodes[0], ref, name_len);
6258                 btrfs_set_inode_ref_index(path->nodes[0], ref, *index);
6259                 ptr = (unsigned long)(ref + 1);
6260                 write_extent_buffer(path->nodes[0], name, ptr, name_len);
6261         }
6262
6263         btrfs_mark_buffer_dirty(path->nodes[0]);
6264         btrfs_free_path(path);
6265
6266         btrfs_inherit_iflags(inode, dir);
6267
6268         if (S_ISREG(mode)) {
6269                 if (btrfs_test_opt(root, NODATASUM))
6270                         BTRFS_I(inode)->flags |= BTRFS_INODE_NODATASUM;
6271                 if (btrfs_test_opt(root, NODATACOW))
6272                         BTRFS_I(inode)->flags |= BTRFS_INODE_NODATACOW |
6273                                 BTRFS_INODE_NODATASUM;
6274         }
6275
6276         inode_tree_add(inode);
6277
6278         trace_btrfs_inode_new(inode);
6279         btrfs_set_inode_last_trans(trans, inode);
6280
6281         btrfs_update_root_times(trans, root);
6282
6283         ret = btrfs_inode_inherit_props(trans, inode, dir);
6284         if (ret)
6285                 btrfs_err(root->fs_info,
6286                           "error inheriting props for ino %llu (root %llu): %d",
6287                           btrfs_ino(inode), root->root_key.objectid, ret);
6288
6289         return inode;
6290
6291 fail_unlock:
6292         unlock_new_inode(inode);
6293 fail:
6294         if (dir && name)
6295                 BTRFS_I(dir)->index_cnt--;
6296         btrfs_free_path(path);
6297         iput(inode);
6298         return ERR_PTR(ret);
6299 }
6300
6301 static inline u8 btrfs_inode_type(struct inode *inode)
6302 {
6303         return btrfs_type_by_mode[(inode->i_mode & S_IFMT) >> S_SHIFT];
6304 }
6305
6306 /*
6307  * utility function to add 'inode' into 'parent_inode' with
6308  * a give name and a given sequence number.
6309  * if 'add_backref' is true, also insert a backref from the
6310  * inode to the parent directory.
6311  */
6312 int btrfs_add_link(struct btrfs_trans_handle *trans,
6313                    struct inode *parent_inode, struct inode *inode,
6314                    const char *name, int name_len, int add_backref, u64 index)
6315 {
6316         int ret = 0;
6317         struct btrfs_key key;
6318         struct btrfs_root *root = BTRFS_I(parent_inode)->root;
6319         u64 ino = btrfs_ino(inode);
6320         u64 parent_ino = btrfs_ino(parent_inode);
6321
6322         if (unlikely(ino == BTRFS_FIRST_FREE_OBJECTID)) {
6323                 memcpy(&key, &BTRFS_I(inode)->root->root_key, sizeof(key));
6324         } else {
6325                 key.objectid = ino;
6326                 key.type = BTRFS_INODE_ITEM_KEY;
6327                 key.offset = 0;
6328         }
6329
6330         if (unlikely(ino == BTRFS_FIRST_FREE_OBJECTID)) {
6331                 ret = btrfs_add_root_ref(trans, root->fs_info->tree_root,
6332                                          key.objectid, root->root_key.objectid,
6333                                          parent_ino, index, name, name_len);
6334         } else if (add_backref) {
6335                 ret = btrfs_insert_inode_ref(trans, root, name, name_len, ino,
6336                                              parent_ino, index);
6337         }
6338
6339         /* Nothing to clean up yet */
6340         if (ret)
6341                 return ret;
6342
6343         ret = btrfs_insert_dir_item(trans, root, name, name_len,
6344                                     parent_inode, &key,
6345                                     btrfs_inode_type(inode), index);
6346         if (ret == -EEXIST || ret == -EOVERFLOW)
6347                 goto fail_dir_item;
6348         else if (ret) {
6349                 btrfs_abort_transaction(trans, root, ret);
6350                 return ret;
6351         }
6352
6353         btrfs_i_size_write(parent_inode, parent_inode->i_size +
6354                            name_len * 2);
6355         inode_inc_iversion(parent_inode);
6356         parent_inode->i_mtime = parent_inode->i_ctime =
6357                 current_fs_time(parent_inode->i_sb);
6358         ret = btrfs_update_inode(trans, root, parent_inode);
6359         if (ret)
6360                 btrfs_abort_transaction(trans, root, ret);
6361         return ret;
6362
6363 fail_dir_item:
6364         if (unlikely(ino == BTRFS_FIRST_FREE_OBJECTID)) {
6365                 u64 local_index;
6366                 int err;
6367                 err = btrfs_del_root_ref(trans, root->fs_info->tree_root,
6368                                  key.objectid, root->root_key.objectid,
6369                                  parent_ino, &local_index, name, name_len);
6370
6371         } else if (add_backref) {
6372                 u64 local_index;
6373                 int err;
6374
6375                 err = btrfs_del_inode_ref(trans, root, name, name_len,
6376                                           ino, parent_ino, &local_index);
6377         }
6378         return ret;
6379 }
6380
6381 static int btrfs_add_nondir(struct btrfs_trans_handle *trans,
6382                             struct inode *dir, struct dentry *dentry,
6383                             struct inode *inode, int backref, u64 index)
6384 {
6385         int err = btrfs_add_link(trans, dir, inode,
6386                                  dentry->d_name.name, dentry->d_name.len,
6387                                  backref, index);
6388         if (err > 0)
6389                 err = -EEXIST;
6390         return err;
6391 }
6392
6393 static int btrfs_mknod(struct inode *dir, struct dentry *dentry,
6394                         umode_t mode, dev_t rdev)
6395 {
6396         struct btrfs_trans_handle *trans;
6397         struct btrfs_root *root = BTRFS_I(dir)->root;
6398         struct inode *inode = NULL;
6399         int err;
6400         int drop_inode = 0;
6401         u64 objectid;
6402         u64 index = 0;
6403
6404         /*
6405          * 2 for inode item and ref
6406          * 2 for dir items
6407          * 1 for xattr if selinux is on
6408          */
6409         trans = btrfs_start_transaction(root, 5);
6410         if (IS_ERR(trans))
6411                 return PTR_ERR(trans);
6412
6413         err = btrfs_find_free_ino(root, &objectid);
6414         if (err)
6415                 goto out_unlock;
6416
6417         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
6418                                 dentry->d_name.len, btrfs_ino(dir), objectid,
6419                                 mode, &index);
6420         if (IS_ERR(inode)) {
6421                 err = PTR_ERR(inode);
6422                 goto out_unlock;
6423         }
6424
6425         /*
6426         * If the active LSM wants to access the inode during
6427         * d_instantiate it needs these. Smack checks to see
6428         * if the filesystem supports xattrs by looking at the
6429         * ops vector.
6430         */
6431         inode->i_op = &btrfs_special_inode_operations;
6432         init_special_inode(inode, inode->i_mode, rdev);
6433
6434         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
6435         if (err)
6436                 goto out_unlock_inode;
6437
6438         err = btrfs_add_nondir(trans, dir, dentry, inode, 0, index);
6439         if (err) {
6440                 goto out_unlock_inode;
6441         } else {
6442                 btrfs_update_inode(trans, root, inode);
6443                 unlock_new_inode(inode);
6444                 d_instantiate(dentry, inode);
6445         }
6446
6447 out_unlock:
6448         btrfs_end_transaction(trans, root);
6449         btrfs_balance_delayed_items(root);
6450         btrfs_btree_balance_dirty(root);
6451         if (drop_inode) {
6452                 inode_dec_link_count(inode);
6453                 iput(inode);
6454         }
6455         return err;
6456
6457 out_unlock_inode:
6458         drop_inode = 1;
6459         unlock_new_inode(inode);
6460         goto out_unlock;
6461
6462 }
6463
6464 static int btrfs_create(struct inode *dir, struct dentry *dentry,
6465                         umode_t mode, bool excl)
6466 {
6467         struct btrfs_trans_handle *trans;
6468         struct btrfs_root *root = BTRFS_I(dir)->root;
6469         struct inode *inode = NULL;
6470         int drop_inode_on_err = 0;
6471         int err;
6472         u64 objectid;
6473         u64 index = 0;
6474
6475         /*
6476          * 2 for inode item and ref
6477          * 2 for dir items
6478          * 1 for xattr if selinux is on
6479          */
6480         trans = btrfs_start_transaction(root, 5);
6481         if (IS_ERR(trans))
6482                 return PTR_ERR(trans);
6483
6484         err = btrfs_find_free_ino(root, &objectid);
6485         if (err)
6486                 goto out_unlock;
6487
6488         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
6489                                 dentry->d_name.len, btrfs_ino(dir), objectid,
6490                                 mode, &index);
6491         if (IS_ERR(inode)) {
6492                 err = PTR_ERR(inode);
6493                 goto out_unlock;
6494         }
6495         drop_inode_on_err = 1;
6496         /*
6497         * If the active LSM wants to access the inode during
6498         * d_instantiate it needs these. Smack checks to see
6499         * if the filesystem supports xattrs by looking at the
6500         * ops vector.
6501         */
6502         inode->i_fop = &btrfs_file_operations;
6503         inode->i_op = &btrfs_file_inode_operations;
6504         inode->i_mapping->a_ops = &btrfs_aops;
6505
6506         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
6507         if (err)
6508                 goto out_unlock_inode;
6509
6510         err = btrfs_update_inode(trans, root, inode);
6511         if (err)
6512                 goto out_unlock_inode;
6513
6514         err = btrfs_add_nondir(trans, dir, dentry, inode, 0, index);
6515         if (err)
6516                 goto out_unlock_inode;
6517
6518         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
6519         unlock_new_inode(inode);
6520         d_instantiate(dentry, inode);
6521
6522 out_unlock:
6523         btrfs_end_transaction(trans, root);
6524         if (err && drop_inode_on_err) {
6525                 inode_dec_link_count(inode);
6526                 iput(inode);
6527         }
6528         btrfs_balance_delayed_items(root);
6529         btrfs_btree_balance_dirty(root);
6530         return err;
6531
6532 out_unlock_inode:
6533         unlock_new_inode(inode);
6534         goto out_unlock;
6535
6536 }
6537
6538 static int btrfs_link(struct dentry *old_dentry, struct inode *dir,
6539                       struct dentry *dentry)
6540 {
6541         struct btrfs_trans_handle *trans = NULL;
6542         struct btrfs_root *root = BTRFS_I(dir)->root;
6543         struct inode *inode = d_inode(old_dentry);
6544         u64 index;
6545         int err;
6546         int drop_inode = 0;
6547
6548         /* do not allow sys_link's with other subvols of the same device */
6549         if (root->objectid != BTRFS_I(inode)->root->objectid)
6550                 return -EXDEV;
6551
6552         if (inode->i_nlink >= BTRFS_LINK_MAX)
6553                 return -EMLINK;
6554
6555         err = btrfs_set_inode_index(dir, &index);
6556         if (err)
6557                 goto fail;
6558
6559         /*
6560          * 2 items for inode and inode ref
6561          * 2 items for dir items
6562          * 1 item for parent inode
6563          */
6564         trans = btrfs_start_transaction(root, 5);
6565         if (IS_ERR(trans)) {
6566                 err = PTR_ERR(trans);
6567                 trans = NULL;
6568                 goto fail;
6569         }
6570
6571         /* There are several dir indexes for this inode, clear the cache. */
6572         BTRFS_I(inode)->dir_index = 0ULL;
6573         inc_nlink(inode);
6574         inode_inc_iversion(inode);
6575         inode->i_ctime = current_fs_time(inode->i_sb);
6576         ihold(inode);
6577         set_bit(BTRFS_INODE_COPY_EVERYTHING, &BTRFS_I(inode)->runtime_flags);
6578
6579         err = btrfs_add_nondir(trans, dir, dentry, inode, 1, index);
6580
6581         if (err) {
6582                 drop_inode = 1;
6583         } else {
6584                 struct dentry *parent = dentry->d_parent;
6585                 err = btrfs_update_inode(trans, root, inode);
6586                 if (err)
6587                         goto fail;
6588                 if (inode->i_nlink == 1) {
6589                         /*
6590                          * If new hard link count is 1, it's a file created
6591                          * with open(2) O_TMPFILE flag.
6592                          */
6593                         err = btrfs_orphan_del(trans, inode);
6594                         if (err)
6595                                 goto fail;
6596                 }
6597                 d_instantiate(dentry, inode);
6598                 btrfs_log_new_name(trans, inode, NULL, parent);
6599         }
6600
6601         btrfs_balance_delayed_items(root);
6602 fail:
6603         if (trans)
6604                 btrfs_end_transaction(trans, root);
6605         if (drop_inode) {
6606                 inode_dec_link_count(inode);
6607                 iput(inode);
6608         }
6609         btrfs_btree_balance_dirty(root);
6610         return err;
6611 }
6612
6613 static int btrfs_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode)
6614 {
6615         struct inode *inode = NULL;
6616         struct btrfs_trans_handle *trans;
6617         struct btrfs_root *root = BTRFS_I(dir)->root;
6618         int err = 0;
6619         int drop_on_err = 0;
6620         u64 objectid = 0;
6621         u64 index = 0;
6622
6623         /*
6624          * 2 items for inode and ref
6625          * 2 items for dir items
6626          * 1 for xattr if selinux is on
6627          */
6628         trans = btrfs_start_transaction(root, 5);
6629         if (IS_ERR(trans))
6630                 return PTR_ERR(trans);
6631
6632         err = btrfs_find_free_ino(root, &objectid);
6633         if (err)
6634                 goto out_fail;
6635
6636         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
6637                                 dentry->d_name.len, btrfs_ino(dir), objectid,
6638                                 S_IFDIR | mode, &index);
6639         if (IS_ERR(inode)) {
6640                 err = PTR_ERR(inode);
6641                 goto out_fail;
6642         }
6643
6644         drop_on_err = 1;
6645         /* these must be set before we unlock the inode */
6646         inode->i_op = &btrfs_dir_inode_operations;
6647         inode->i_fop = &btrfs_dir_file_operations;
6648
6649         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
6650         if (err)
6651                 goto out_fail_inode;
6652
6653         btrfs_i_size_write(inode, 0);
6654         err = btrfs_update_inode(trans, root, inode);
6655         if (err)
6656                 goto out_fail_inode;
6657
6658         err = btrfs_add_link(trans, dir, inode, dentry->d_name.name,
6659                              dentry->d_name.len, 0, index);
6660         if (err)
6661                 goto out_fail_inode;
6662
6663         d_instantiate(dentry, inode);
6664         /*
6665          * mkdir is special.  We're unlocking after we call d_instantiate
6666          * to avoid a race with nfsd calling d_instantiate.
6667          */
6668         unlock_new_inode(inode);
6669         drop_on_err = 0;
6670
6671 out_fail:
6672         btrfs_end_transaction(trans, root);
6673         if (drop_on_err) {
6674                 inode_dec_link_count(inode);
6675                 iput(inode);
6676         }
6677         btrfs_balance_delayed_items(root);
6678         btrfs_btree_balance_dirty(root);
6679         return err;
6680
6681 out_fail_inode:
6682         unlock_new_inode(inode);
6683         goto out_fail;
6684 }
6685
6686 /* Find next extent map of a given extent map, caller needs to ensure locks */
6687 static struct extent_map *next_extent_map(struct extent_map *em)
6688 {
6689         struct rb_node *next;
6690
6691         next = rb_next(&em->rb_node);
6692         if (!next)
6693                 return NULL;
6694         return container_of(next, struct extent_map, rb_node);
6695 }
6696
6697 static struct extent_map *prev_extent_map(struct extent_map *em)
6698 {
6699         struct rb_node *prev;
6700
6701         prev = rb_prev(&em->rb_node);
6702         if (!prev)
6703                 return NULL;
6704         return container_of(prev, struct extent_map, rb_node);
6705 }
6706
6707 /* helper for btfs_get_extent.  Given an existing extent in the tree,
6708  * the existing extent is the nearest extent to map_start,
6709  * and an extent that you want to insert, deal with overlap and insert
6710  * the best fitted new extent into the tree.
6711  */
6712 static int merge_extent_mapping(struct extent_map_tree *em_tree,
6713                                 struct extent_map *existing,
6714                                 struct extent_map *em,
6715                                 u64 map_start)
6716 {
6717         struct extent_map *prev;
6718         struct extent_map *next;
6719         u64 start;
6720         u64 end;
6721         u64 start_diff;
6722
6723         BUG_ON(map_start < em->start || map_start >= extent_map_end(em));
6724
6725         if (existing->start > map_start) {
6726                 next = existing;
6727                 prev = prev_extent_map(next);
6728         } else {
6729                 prev = existing;
6730                 next = next_extent_map(prev);
6731         }
6732
6733         start = prev ? extent_map_end(prev) : em->start;
6734         start = max_t(u64, start, em->start);
6735         end = next ? next->start : extent_map_end(em);
6736         end = min_t(u64, end, extent_map_end(em));
6737         start_diff = start - em->start;
6738         em->start = start;
6739         em->len = end - start;
6740         if (em->block_start < EXTENT_MAP_LAST_BYTE &&
6741             !test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
6742                 em->block_start += start_diff;
6743                 em->block_len -= start_diff;
6744         }
6745         return add_extent_mapping(em_tree, em, 0);
6746 }
6747
6748 static noinline int uncompress_inline(struct btrfs_path *path,
6749                                       struct page *page,
6750                                       size_t pg_offset, u64 extent_offset,
6751                                       struct btrfs_file_extent_item *item)
6752 {
6753         int ret;
6754         struct extent_buffer *leaf = path->nodes[0];
6755         char *tmp;
6756         size_t max_size;
6757         unsigned long inline_size;
6758         unsigned long ptr;
6759         int compress_type;
6760
6761         WARN_ON(pg_offset != 0);
6762         compress_type = btrfs_file_extent_compression(leaf, item);
6763         max_size = btrfs_file_extent_ram_bytes(leaf, item);
6764         inline_size = btrfs_file_extent_inline_item_len(leaf,
6765                                         btrfs_item_nr(path->slots[0]));
6766         tmp = kmalloc(inline_size, GFP_NOFS);
6767         if (!tmp)
6768                 return -ENOMEM;
6769         ptr = btrfs_file_extent_inline_start(item);
6770
6771         read_extent_buffer(leaf, tmp, ptr, inline_size);
6772
6773         max_size = min_t(unsigned long, PAGE_SIZE, max_size);
6774         ret = btrfs_decompress(compress_type, tmp, page,
6775                                extent_offset, inline_size, max_size);
6776         kfree(tmp);
6777         return ret;
6778 }
6779
6780 /*
6781  * a bit scary, this does extent mapping from logical file offset to the disk.
6782  * the ugly parts come from merging extents from the disk with the in-ram
6783  * representation.  This gets more complex because of the data=ordered code,
6784  * where the in-ram extents might be locked pending data=ordered completion.
6785  *
6786  * This also copies inline extents directly into the page.
6787  */
6788
6789 struct extent_map *btrfs_get_extent(struct inode *inode, struct page *page,
6790                                     size_t pg_offset, u64 start, u64 len,
6791                                     int create)
6792 {
6793         int ret;
6794         int err = 0;
6795         u64 extent_start = 0;
6796         u64 extent_end = 0;
6797         u64 objectid = btrfs_ino(inode);
6798         u32 found_type;
6799         struct btrfs_path *path = NULL;
6800         struct btrfs_root *root = BTRFS_I(inode)->root;
6801         struct btrfs_file_extent_item *item;
6802         struct extent_buffer *leaf;
6803         struct btrfs_key found_key;
6804         struct extent_map *em = NULL;
6805         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
6806         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
6807         struct btrfs_trans_handle *trans = NULL;
6808         const bool new_inline = !page || create;
6809
6810 again:
6811         read_lock(&em_tree->lock);
6812         em = lookup_extent_mapping(em_tree, start, len);
6813         if (em)
6814                 em->bdev = root->fs_info->fs_devices->latest_bdev;
6815         read_unlock(&em_tree->lock);
6816
6817         if (em) {
6818                 if (em->start > start || em->start + em->len <= start)
6819                         free_extent_map(em);
6820                 else if (em->block_start == EXTENT_MAP_INLINE && page)
6821                         free_extent_map(em);
6822                 else
6823                         goto out;
6824         }
6825         em = alloc_extent_map();
6826         if (!em) {
6827                 err = -ENOMEM;
6828                 goto out;
6829         }
6830         em->bdev = root->fs_info->fs_devices->latest_bdev;
6831         em->start = EXTENT_MAP_HOLE;
6832         em->orig_start = EXTENT_MAP_HOLE;
6833         em->len = (u64)-1;
6834         em->block_len = (u64)-1;
6835
6836         if (!path) {
6837                 path = btrfs_alloc_path();
6838                 if (!path) {
6839                         err = -ENOMEM;
6840                         goto out;
6841                 }
6842                 /*
6843                  * Chances are we'll be called again, so go ahead and do
6844                  * readahead
6845                  */
6846                 path->reada = READA_FORWARD;
6847         }
6848
6849         ret = btrfs_lookup_file_extent(trans, root, path,
6850                                        objectid, start, trans != NULL);
6851         if (ret < 0) {
6852                 err = ret;
6853                 goto out;
6854         }
6855
6856         if (ret != 0) {
6857                 if (path->slots[0] == 0)
6858                         goto not_found;
6859                 path->slots[0]--;
6860         }
6861
6862         leaf = path->nodes[0];
6863         item = btrfs_item_ptr(leaf, path->slots[0],
6864                               struct btrfs_file_extent_item);
6865         /* are we inside the extent that was found? */
6866         btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
6867         found_type = found_key.type;
6868         if (found_key.objectid != objectid ||
6869             found_type != BTRFS_EXTENT_DATA_KEY) {
6870                 /*
6871                  * If we backup past the first extent we want to move forward
6872                  * and see if there is an extent in front of us, otherwise we'll
6873                  * say there is a hole for our whole search range which can
6874                  * cause problems.
6875                  */
6876                 extent_end = start;
6877                 goto next;
6878         }
6879
6880         found_type = btrfs_file_extent_type(leaf, item);
6881         extent_start = found_key.offset;
6882         if (found_type == BTRFS_FILE_EXTENT_REG ||
6883             found_type == BTRFS_FILE_EXTENT_PREALLOC) {
6884                 extent_end = extent_start +
6885                        btrfs_file_extent_num_bytes(leaf, item);
6886         } else if (found_type == BTRFS_FILE_EXTENT_INLINE) {
6887                 size_t size;
6888                 size = btrfs_file_extent_inline_len(leaf, path->slots[0], item);
6889                 extent_end = ALIGN(extent_start + size, root->sectorsize);
6890         }
6891 next:
6892         if (start >= extent_end) {
6893                 path->slots[0]++;
6894                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
6895                         ret = btrfs_next_leaf(root, path);
6896                         if (ret < 0) {
6897                                 err = ret;
6898                                 goto out;
6899                         }
6900                         if (ret > 0)
6901                                 goto not_found;
6902                         leaf = path->nodes[0];
6903                 }
6904                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
6905                 if (found_key.objectid != objectid ||
6906                     found_key.type != BTRFS_EXTENT_DATA_KEY)
6907                         goto not_found;
6908                 if (start + len <= found_key.offset)
6909                         goto not_found;
6910                 if (start > found_key.offset)
6911                         goto next;
6912                 em->start = start;
6913                 em->orig_start = start;
6914                 em->len = found_key.offset - start;
6915                 goto not_found_em;
6916         }
6917
6918         btrfs_extent_item_to_extent_map(inode, path, item, new_inline, em);
6919
6920         if (found_type == BTRFS_FILE_EXTENT_REG ||
6921             found_type == BTRFS_FILE_EXTENT_PREALLOC) {
6922                 goto insert;
6923         } else if (found_type == BTRFS_FILE_EXTENT_INLINE) {
6924                 unsigned long ptr;
6925                 char *map;
6926                 size_t size;
6927                 size_t extent_offset;
6928                 size_t copy_size;
6929
6930                 if (new_inline)
6931                         goto out;
6932
6933                 size = btrfs_file_extent_inline_len(leaf, path->slots[0], item);
6934                 extent_offset = page_offset(page) + pg_offset - extent_start;
6935                 copy_size = min_t(u64, PAGE_SIZE - pg_offset,
6936                                   size - extent_offset);
6937                 em->start = extent_start + extent_offset;
6938                 em->len = ALIGN(copy_size, root->sectorsize);
6939                 em->orig_block_len = em->len;
6940                 em->orig_start = em->start;
6941                 ptr = btrfs_file_extent_inline_start(item) + extent_offset;
6942                 if (create == 0 && !PageUptodate(page)) {
6943                         if (btrfs_file_extent_compression(leaf, item) !=
6944                             BTRFS_COMPRESS_NONE) {
6945                                 ret = uncompress_inline(path, page, pg_offset,
6946                                                         extent_offset, item);
6947                                 if (ret) {
6948                                         err = ret;
6949                                         goto out;
6950                                 }
6951                         } else {
6952                                 map = kmap(page);
6953                                 read_extent_buffer(leaf, map + pg_offset, ptr,
6954                                                    copy_size);
6955                                 if (pg_offset + copy_size < PAGE_SIZE) {
6956                                         memset(map + pg_offset + copy_size, 0,
6957                                                PAGE_SIZE - pg_offset -
6958                                                copy_size);
6959                                 }
6960                                 kunmap(page);
6961                         }
6962                         flush_dcache_page(page);
6963                 } else if (create && PageUptodate(page)) {
6964                         BUG();
6965                         if (!trans) {
6966                                 kunmap(page);
6967                                 free_extent_map(em);
6968                                 em = NULL;
6969
6970                                 btrfs_release_path(path);
6971                                 trans = btrfs_join_transaction(root);
6972
6973                                 if (IS_ERR(trans))
6974                                         return ERR_CAST(trans);
6975                                 goto again;
6976                         }
6977                         map = kmap(page);
6978                         write_extent_buffer(leaf, map + pg_offset, ptr,
6979                                             copy_size);
6980                         kunmap(page);
6981                         btrfs_mark_buffer_dirty(leaf);
6982                 }
6983                 set_extent_uptodate(io_tree, em->start,
6984                                     extent_map_end(em) - 1, NULL, GFP_NOFS);
6985                 goto insert;
6986         }
6987 not_found:
6988         em->start = start;
6989         em->orig_start = start;
6990         em->len = len;
6991 not_found_em:
6992         em->block_start = EXTENT_MAP_HOLE;
6993         set_bit(EXTENT_FLAG_VACANCY, &em->flags);
6994 insert:
6995         btrfs_release_path(path);
6996         if (em->start > start || extent_map_end(em) <= start) {
6997                 btrfs_err(root->fs_info, "bad extent! em: [%llu %llu] passed [%llu %llu]",
6998                         em->start, em->len, start, len);
6999                 err = -EIO;
7000                 goto out;
7001         }
7002
7003         err = 0;
7004         write_lock(&em_tree->lock);
7005         ret = add_extent_mapping(em_tree, em, 0);
7006         /* it is possible that someone inserted the extent into the tree
7007          * while we had the lock dropped.  It is also possible that
7008          * an overlapping map exists in the tree
7009          */
7010         if (ret == -EEXIST) {
7011                 struct extent_map *existing;
7012
7013                 ret = 0;
7014
7015                 existing = search_extent_mapping(em_tree, start, len);
7016                 /*
7017                  * existing will always be non-NULL, since there must be
7018                  * extent causing the -EEXIST.
7019                  */
7020                 if (existing->start == em->start &&
7021                     extent_map_end(existing) == extent_map_end(em) &&
7022                     em->block_start == existing->block_start) {
7023                         /*
7024                          * these two extents are the same, it happens
7025                          * with inlines especially
7026                          */
7027                         free_extent_map(em);
7028                         em = existing;
7029                         err = 0;
7030
7031                 } else if (start >= extent_map_end(existing) ||
7032                     start <= existing->start) {
7033                         /*
7034                          * The existing extent map is the one nearest to
7035                          * the [start, start + len) range which overlaps
7036                          */
7037                         err = merge_extent_mapping(em_tree, existing,
7038                                                    em, start);
7039                         free_extent_map(existing);
7040                         if (err) {
7041                                 free_extent_map(em);
7042                                 em = NULL;
7043                         }
7044                 } else {
7045                         free_extent_map(em);
7046                         em = existing;
7047                         err = 0;
7048                 }
7049         }
7050         write_unlock(&em_tree->lock);
7051 out:
7052
7053         trace_btrfs_get_extent(root, em);
7054
7055         btrfs_free_path(path);
7056         if (trans) {
7057                 ret = btrfs_end_transaction(trans, root);
7058                 if (!err)
7059                         err = ret;
7060         }
7061         if (err) {
7062                 free_extent_map(em);
7063                 return ERR_PTR(err);
7064         }
7065         BUG_ON(!em); /* Error is always set */
7066         return em;
7067 }
7068
7069 struct extent_map *btrfs_get_extent_fiemap(struct inode *inode, struct page *page,
7070                                            size_t pg_offset, u64 start, u64 len,
7071                                            int create)
7072 {
7073         struct extent_map *em;
7074         struct extent_map *hole_em = NULL;
7075         u64 range_start = start;
7076         u64 end;
7077         u64 found;
7078         u64 found_end;
7079         int err = 0;
7080
7081         em = btrfs_get_extent(inode, page, pg_offset, start, len, create);
7082         if (IS_ERR(em))
7083                 return em;
7084         if (em) {
7085                 /*
7086                  * if our em maps to
7087                  * -  a hole or
7088                  * -  a pre-alloc extent,
7089                  * there might actually be delalloc bytes behind it.
7090                  */
7091                 if (em->block_start != EXTENT_MAP_HOLE &&
7092                     !test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
7093                         return em;
7094                 else
7095                         hole_em = em;
7096         }
7097
7098         /* check to see if we've wrapped (len == -1 or similar) */
7099         end = start + len;
7100         if (end < start)
7101                 end = (u64)-1;
7102         else
7103                 end -= 1;
7104
7105         em = NULL;
7106
7107         /* ok, we didn't find anything, lets look for delalloc */
7108         found = count_range_bits(&BTRFS_I(inode)->io_tree, &range_start,
7109                                  end, len, EXTENT_DELALLOC, 1);
7110         found_end = range_start + found;
7111         if (found_end < range_start)
7112                 found_end = (u64)-1;
7113
7114         /*
7115          * we didn't find anything useful, return
7116          * the original results from get_extent()
7117          */
7118         if (range_start > end || found_end <= start) {
7119                 em = hole_em;
7120                 hole_em = NULL;
7121                 goto out;
7122         }
7123
7124         /* adjust the range_start to make sure it doesn't
7125          * go backwards from the start they passed in
7126          */
7127         range_start = max(start, range_start);
7128         found = found_end - range_start;
7129
7130         if (found > 0) {
7131                 u64 hole_start = start;
7132                 u64 hole_len = len;
7133
7134                 em = alloc_extent_map();
7135                 if (!em) {
7136                         err = -ENOMEM;
7137                         goto out;
7138                 }
7139                 /*
7140                  * when btrfs_get_extent can't find anything it
7141                  * returns one huge hole
7142                  *
7143                  * make sure what it found really fits our range, and
7144                  * adjust to make sure it is based on the start from
7145                  * the caller
7146                  */
7147                 if (hole_em) {
7148                         u64 calc_end = extent_map_end(hole_em);
7149
7150                         if (calc_end <= start || (hole_em->start > end)) {
7151                                 free_extent_map(hole_em);
7152                                 hole_em = NULL;
7153                         } else {
7154                                 hole_start = max(hole_em->start, start);
7155                                 hole_len = calc_end - hole_start;
7156                         }
7157                 }
7158                 em->bdev = NULL;
7159                 if (hole_em && range_start > hole_start) {
7160                         /* our hole starts before our delalloc, so we
7161                          * have to return just the parts of the hole
7162                          * that go until  the delalloc starts
7163                          */
7164                         em->len = min(hole_len,
7165                                       range_start - hole_start);
7166                         em->start = hole_start;
7167                         em->orig_start = hole_start;
7168                         /*
7169                          * don't adjust block start at all,
7170                          * it is fixed at EXTENT_MAP_HOLE
7171                          */
7172                         em->block_start = hole_em->block_start;
7173                         em->block_len = hole_len;
7174                         if (test_bit(EXTENT_FLAG_PREALLOC, &hole_em->flags))
7175                                 set_bit(EXTENT_FLAG_PREALLOC, &em->flags);
7176                 } else {
7177                         em->start = range_start;
7178                         em->len = found;
7179                         em->orig_start = range_start;
7180                         em->block_start = EXTENT_MAP_DELALLOC;
7181                         em->block_len = found;
7182                 }
7183         } else if (hole_em) {
7184                 return hole_em;
7185         }
7186 out:
7187
7188         free_extent_map(hole_em);
7189         if (err) {
7190                 free_extent_map(em);
7191                 return ERR_PTR(err);
7192         }
7193         return em;
7194 }
7195
7196 static struct extent_map *btrfs_create_dio_extent(struct inode *inode,
7197                                                   const u64 start,
7198                                                   const u64 len,
7199                                                   const u64 orig_start,
7200                                                   const u64 block_start,
7201                                                   const u64 block_len,
7202                                                   const u64 orig_block_len,
7203                                                   const u64 ram_bytes,
7204                                                   const int type)
7205 {
7206         struct extent_map *em = NULL;
7207         int ret;
7208
7209         down_read(&BTRFS_I(inode)->dio_sem);
7210         if (type != BTRFS_ORDERED_NOCOW) {
7211                 em = create_pinned_em(inode, start, len, orig_start,
7212                                       block_start, block_len, orig_block_len,
7213                                       ram_bytes, type);
7214                 if (IS_ERR(em))
7215                         goto out;
7216         }
7217         ret = btrfs_add_ordered_extent_dio(inode, start, block_start,
7218                                            len, block_len, type);
7219         if (ret) {
7220                 if (em) {
7221                         free_extent_map(em);
7222                         btrfs_drop_extent_cache(inode, start,
7223                                                 start + len - 1, 0);
7224                 }
7225                 em = ERR_PTR(ret);
7226         }
7227  out:
7228         up_read(&BTRFS_I(inode)->dio_sem);
7229
7230         return em;
7231 }
7232
7233 static struct extent_map *btrfs_new_extent_direct(struct inode *inode,
7234                                                   u64 start, u64 len)
7235 {
7236         struct btrfs_root *root = BTRFS_I(inode)->root;
7237         struct extent_map *em;
7238         struct btrfs_key ins;
7239         u64 alloc_hint;
7240         int ret;
7241
7242         alloc_hint = get_extent_allocation_hint(inode, start, len);
7243         ret = btrfs_reserve_extent(root, len, root->sectorsize, 0,
7244                                    alloc_hint, &ins, 1, 1);
7245         if (ret)
7246                 return ERR_PTR(ret);
7247
7248         em = btrfs_create_dio_extent(inode, start, ins.offset, start,
7249                                      ins.objectid, ins.offset, ins.offset,
7250                                      ins.offset, 0);
7251         btrfs_dec_block_group_reservations(root->fs_info, ins.objectid);
7252         if (IS_ERR(em))
7253                 btrfs_free_reserved_extent(root, ins.objectid, ins.offset, 1);
7254
7255         return em;
7256 }
7257
7258 /*
7259  * returns 1 when the nocow is safe, < 1 on error, 0 if the
7260  * block must be cow'd
7261  */
7262 noinline int can_nocow_extent(struct inode *inode, u64 offset, u64 *len,
7263                               u64 *orig_start, u64 *orig_block_len,
7264                               u64 *ram_bytes)
7265 {
7266         struct btrfs_trans_handle *trans;
7267         struct btrfs_path *path;
7268         int ret;
7269         struct extent_buffer *leaf;
7270         struct btrfs_root *root = BTRFS_I(inode)->root;
7271         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
7272         struct btrfs_file_extent_item *fi;
7273         struct btrfs_key key;
7274         u64 disk_bytenr;
7275         u64 backref_offset;
7276         u64 extent_end;
7277         u64 num_bytes;
7278         int slot;
7279         int found_type;
7280         bool nocow = (BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW);
7281
7282         path = btrfs_alloc_path();
7283         if (!path)
7284                 return -ENOMEM;
7285
7286         ret = btrfs_lookup_file_extent(NULL, root, path, btrfs_ino(inode),
7287                                        offset, 0);
7288         if (ret < 0)
7289                 goto out;
7290
7291         slot = path->slots[0];
7292         if (ret == 1) {
7293                 if (slot == 0) {
7294                         /* can't find the item, must cow */
7295                         ret = 0;
7296                         goto out;
7297                 }
7298                 slot--;
7299         }
7300         ret = 0;
7301         leaf = path->nodes[0];
7302         btrfs_item_key_to_cpu(leaf, &key, slot);
7303         if (key.objectid != btrfs_ino(inode) ||
7304             key.type != BTRFS_EXTENT_DATA_KEY) {
7305                 /* not our file or wrong item type, must cow */
7306                 goto out;
7307         }
7308
7309         if (key.offset > offset) {
7310                 /* Wrong offset, must cow */
7311                 goto out;
7312         }
7313
7314         fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
7315         found_type = btrfs_file_extent_type(leaf, fi);
7316         if (found_type != BTRFS_FILE_EXTENT_REG &&
7317             found_type != BTRFS_FILE_EXTENT_PREALLOC) {
7318                 /* not a regular extent, must cow */
7319                 goto out;
7320         }
7321
7322         if (!nocow && found_type == BTRFS_FILE_EXTENT_REG)
7323                 goto out;
7324
7325         extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
7326         if (extent_end <= offset)
7327                 goto out;
7328
7329         disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
7330         if (disk_bytenr == 0)
7331                 goto out;
7332
7333         if (btrfs_file_extent_compression(leaf, fi) ||
7334             btrfs_file_extent_encryption(leaf, fi) ||
7335             btrfs_file_extent_other_encoding(leaf, fi))
7336                 goto out;
7337
7338         backref_offset = btrfs_file_extent_offset(leaf, fi);
7339
7340         if (orig_start) {
7341                 *orig_start = key.offset - backref_offset;
7342                 *orig_block_len = btrfs_file_extent_disk_num_bytes(leaf, fi);
7343                 *ram_bytes = btrfs_file_extent_ram_bytes(leaf, fi);
7344         }
7345
7346         if (btrfs_extent_readonly(root, disk_bytenr))
7347                 goto out;
7348
7349         num_bytes = min(offset + *len, extent_end) - offset;
7350         if (!nocow && found_type == BTRFS_FILE_EXTENT_PREALLOC) {
7351                 u64 range_end;
7352
7353                 range_end = round_up(offset + num_bytes, root->sectorsize) - 1;
7354                 ret = test_range_bit(io_tree, offset, range_end,
7355                                      EXTENT_DELALLOC, 0, NULL);
7356                 if (ret) {
7357                         ret = -EAGAIN;
7358                         goto out;
7359                 }
7360         }
7361
7362         btrfs_release_path(path);
7363
7364         /*
7365          * look for other files referencing this extent, if we
7366          * find any we must cow
7367          */
7368         trans = btrfs_join_transaction(root);
7369         if (IS_ERR(trans)) {
7370                 ret = 0;
7371                 goto out;
7372         }
7373
7374         ret = btrfs_cross_ref_exist(trans, root, btrfs_ino(inode),
7375                                     key.offset - backref_offset, disk_bytenr);
7376         btrfs_end_transaction(trans, root);
7377         if (ret) {
7378                 ret = 0;
7379                 goto out;
7380         }
7381
7382         /*
7383          * adjust disk_bytenr and num_bytes to cover just the bytes
7384          * in this extent we are about to write.  If there
7385          * are any csums in that range we have to cow in order
7386          * to keep the csums correct
7387          */
7388         disk_bytenr += backref_offset;
7389         disk_bytenr += offset - key.offset;
7390         if (csum_exist_in_range(root, disk_bytenr, num_bytes))
7391                                 goto out;
7392         /*
7393          * all of the above have passed, it is safe to overwrite this extent
7394          * without cow
7395          */
7396         *len = num_bytes;
7397         ret = 1;
7398 out:
7399         btrfs_free_path(path);
7400         return ret;
7401 }
7402
7403 bool btrfs_page_exists_in_range(struct inode *inode, loff_t start, loff_t end)
7404 {
7405         struct radix_tree_root *root = &inode->i_mapping->page_tree;
7406         int found = false;
7407         void **pagep = NULL;
7408         struct page *page = NULL;
7409         int start_idx;
7410         int end_idx;
7411
7412         start_idx = start >> PAGE_SHIFT;
7413
7414         /*
7415          * end is the last byte in the last page.  end == start is legal
7416          */
7417         end_idx = end >> PAGE_SHIFT;
7418
7419         rcu_read_lock();
7420
7421         /* Most of the code in this while loop is lifted from
7422          * find_get_page.  It's been modified to begin searching from a
7423          * page and return just the first page found in that range.  If the
7424          * found idx is less than or equal to the end idx then we know that
7425          * a page exists.  If no pages are found or if those pages are
7426          * outside of the range then we're fine (yay!) */
7427         while (page == NULL &&
7428                radix_tree_gang_lookup_slot(root, &pagep, NULL, start_idx, 1)) {
7429                 page = radix_tree_deref_slot(pagep);
7430                 if (unlikely(!page))
7431                         break;
7432
7433                 if (radix_tree_exception(page)) {
7434                         if (radix_tree_deref_retry(page)) {
7435                                 page = NULL;
7436                                 continue;
7437                         }
7438                         /*
7439                          * Otherwise, shmem/tmpfs must be storing a swap entry
7440                          * here as an exceptional entry: so return it without
7441                          * attempting to raise page count.
7442                          */
7443                         page = NULL;
7444                         break; /* TODO: Is this relevant for this use case? */
7445                 }
7446
7447                 if (!page_cache_get_speculative(page)) {
7448                         page = NULL;
7449                         continue;
7450                 }
7451
7452                 /*
7453                  * Has the page moved?
7454                  * This is part of the lockless pagecache protocol. See
7455                  * include/linux/pagemap.h for details.
7456                  */
7457                 if (unlikely(page != *pagep)) {
7458                         put_page(page);
7459                         page = NULL;
7460                 }
7461         }
7462
7463         if (page) {
7464                 if (page->index <= end_idx)
7465                         found = true;
7466                 put_page(page);
7467         }
7468
7469         rcu_read_unlock();
7470         return found;
7471 }
7472
7473 static int lock_extent_direct(struct inode *inode, u64 lockstart, u64 lockend,
7474                               struct extent_state **cached_state, int writing)
7475 {
7476         struct btrfs_ordered_extent *ordered;
7477         int ret = 0;
7478
7479         while (1) {
7480                 lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend,
7481                                  cached_state);
7482                 /*
7483                  * We're concerned with the entire range that we're going to be
7484                  * doing DIO to, so we need to make sure there's no ordered
7485                  * extents in this range.
7486                  */
7487                 ordered = btrfs_lookup_ordered_range(inode, lockstart,
7488                                                      lockend - lockstart + 1);
7489
7490                 /*
7491                  * We need to make sure there are no buffered pages in this
7492                  * range either, we could have raced between the invalidate in
7493                  * generic_file_direct_write and locking the extent.  The
7494                  * invalidate needs to happen so that reads after a write do not
7495                  * get stale data.
7496                  */
7497                 if (!ordered &&
7498                     (!writing ||
7499                      !btrfs_page_exists_in_range(inode, lockstart, lockend)))
7500                         break;
7501
7502                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
7503                                      cached_state, GFP_NOFS);
7504
7505                 if (ordered) {
7506                         /*
7507                          * If we are doing a DIO read and the ordered extent we
7508                          * found is for a buffered write, we can not wait for it
7509                          * to complete and retry, because if we do so we can
7510                          * deadlock with concurrent buffered writes on page
7511                          * locks. This happens only if our DIO read covers more
7512                          * than one extent map, if at this point has already
7513                          * created an ordered extent for a previous extent map
7514                          * and locked its range in the inode's io tree, and a
7515                          * concurrent write against that previous extent map's
7516                          * range and this range started (we unlock the ranges
7517                          * in the io tree only when the bios complete and
7518                          * buffered writes always lock pages before attempting
7519                          * to lock range in the io tree).
7520                          */
7521                         if (writing ||
7522                             test_bit(BTRFS_ORDERED_DIRECT, &ordered->flags))
7523                                 btrfs_start_ordered_extent(inode, ordered, 1);
7524                         else
7525                                 ret = -ENOTBLK;
7526                         btrfs_put_ordered_extent(ordered);
7527                 } else {
7528                         /*
7529                          * We could trigger writeback for this range (and wait
7530                          * for it to complete) and then invalidate the pages for
7531                          * this range (through invalidate_inode_pages2_range()),
7532                          * but that can lead us to a deadlock with a concurrent
7533                          * call to readpages() (a buffered read or a defrag call
7534                          * triggered a readahead) on a page lock due to an
7535                          * ordered dio extent we created before but did not have
7536                          * yet a corresponding bio submitted (whence it can not
7537                          * complete), which makes readpages() wait for that
7538                          * ordered extent to complete while holding a lock on
7539                          * that page.
7540                          */
7541                         ret = -ENOTBLK;
7542                 }
7543
7544                 if (ret)
7545                         break;
7546
7547                 cond_resched();
7548         }
7549
7550         return ret;
7551 }
7552
7553 static struct extent_map *create_pinned_em(struct inode *inode, u64 start,
7554                                            u64 len, u64 orig_start,
7555                                            u64 block_start, u64 block_len,
7556                                            u64 orig_block_len, u64 ram_bytes,
7557                                            int type)
7558 {
7559         struct extent_map_tree *em_tree;
7560         struct extent_map *em;
7561         struct btrfs_root *root = BTRFS_I(inode)->root;
7562         int ret;
7563
7564         em_tree = &BTRFS_I(inode)->extent_tree;
7565         em = alloc_extent_map();
7566         if (!em)
7567                 return ERR_PTR(-ENOMEM);
7568
7569         em->start = start;
7570         em->orig_start = orig_start;
7571         em->mod_start = start;
7572         em->mod_len = len;
7573         em->len = len;
7574         em->block_len = block_len;
7575         em->block_start = block_start;
7576         em->bdev = root->fs_info->fs_devices->latest_bdev;
7577         em->orig_block_len = orig_block_len;
7578         em->ram_bytes = ram_bytes;
7579         em->generation = -1;
7580         set_bit(EXTENT_FLAG_PINNED, &em->flags);
7581         if (type == BTRFS_ORDERED_PREALLOC)
7582                 set_bit(EXTENT_FLAG_FILLING, &em->flags);
7583
7584         do {
7585                 btrfs_drop_extent_cache(inode, em->start,
7586                                 em->start + em->len - 1, 0);
7587                 write_lock(&em_tree->lock);
7588                 ret = add_extent_mapping(em_tree, em, 1);
7589                 write_unlock(&em_tree->lock);
7590         } while (ret == -EEXIST);
7591
7592         if (ret) {
7593                 free_extent_map(em);
7594                 return ERR_PTR(ret);
7595         }
7596
7597         return em;
7598 }
7599
7600 static void adjust_dio_outstanding_extents(struct inode *inode,
7601                                            struct btrfs_dio_data *dio_data,
7602                                            const u64 len)
7603 {
7604         unsigned num_extents;
7605
7606         num_extents = (unsigned) div64_u64(len + BTRFS_MAX_EXTENT_SIZE - 1,
7607                                            BTRFS_MAX_EXTENT_SIZE);
7608         /*
7609          * If we have an outstanding_extents count still set then we're
7610          * within our reservation, otherwise we need to adjust our inode
7611          * counter appropriately.
7612          */
7613         if (dio_data->outstanding_extents) {
7614                 dio_data->outstanding_extents -= num_extents;
7615         } else {
7616                 spin_lock(&BTRFS_I(inode)->lock);
7617                 BTRFS_I(inode)->outstanding_extents += num_extents;
7618                 spin_unlock(&BTRFS_I(inode)->lock);
7619         }
7620 }
7621
7622 static int btrfs_get_blocks_direct(struct inode *inode, sector_t iblock,
7623                                    struct buffer_head *bh_result, int create)
7624 {
7625         struct extent_map *em;
7626         struct btrfs_root *root = BTRFS_I(inode)->root;
7627         struct extent_state *cached_state = NULL;
7628         struct btrfs_dio_data *dio_data = NULL;
7629         u64 start = iblock << inode->i_blkbits;
7630         u64 lockstart, lockend;
7631         u64 len = bh_result->b_size;
7632         int unlock_bits = EXTENT_LOCKED;
7633         int ret = 0;
7634
7635         if (create)
7636                 unlock_bits |= EXTENT_DIRTY;
7637         else
7638                 len = min_t(u64, len, root->sectorsize);
7639
7640         lockstart = start;
7641         lockend = start + len - 1;
7642
7643         if (current->journal_info) {
7644                 /*
7645                  * Need to pull our outstanding extents and set journal_info to NULL so
7646                  * that anything that needs to check if there's a transaction doesn't get
7647                  * confused.
7648                  */
7649                 dio_data = current->journal_info;
7650                 current->journal_info = NULL;
7651         }
7652
7653         /*
7654          * If this errors out it's because we couldn't invalidate pagecache for
7655          * this range and we need to fallback to buffered.
7656          */
7657         if (lock_extent_direct(inode, lockstart, lockend, &cached_state,
7658                                create)) {
7659                 ret = -ENOTBLK;
7660                 goto err;
7661         }
7662
7663         em = btrfs_get_extent(inode, NULL, 0, start, len, 0);
7664         if (IS_ERR(em)) {
7665                 ret = PTR_ERR(em);
7666                 goto unlock_err;
7667         }
7668
7669         /*
7670          * Ok for INLINE and COMPRESSED extents we need to fallback on buffered
7671          * io.  INLINE is special, and we could probably kludge it in here, but
7672          * it's still buffered so for safety lets just fall back to the generic
7673          * buffered path.
7674          *
7675          * For COMPRESSED we _have_ to read the entire extent in so we can
7676          * decompress it, so there will be buffering required no matter what we
7677          * do, so go ahead and fallback to buffered.
7678          *
7679          * We return -ENOTBLK because that's what makes DIO go ahead and go back
7680          * to buffered IO.  Don't blame me, this is the price we pay for using
7681          * the generic code.
7682          */
7683         if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags) ||
7684             em->block_start == EXTENT_MAP_INLINE) {
7685                 free_extent_map(em);
7686                 ret = -ENOTBLK;
7687                 goto unlock_err;
7688         }
7689
7690         /* Just a good old fashioned hole, return */
7691         if (!create && (em->block_start == EXTENT_MAP_HOLE ||
7692                         test_bit(EXTENT_FLAG_PREALLOC, &em->flags))) {
7693                 free_extent_map(em);
7694                 goto unlock_err;
7695         }
7696
7697         /*
7698          * We don't allocate a new extent in the following cases
7699          *
7700          * 1) The inode is marked as NODATACOW.  In this case we'll just use the
7701          * existing extent.
7702          * 2) The extent is marked as PREALLOC.  We're good to go here and can
7703          * just use the extent.
7704          *
7705          */
7706         if (!create) {
7707                 len = min(len, em->len - (start - em->start));
7708                 lockstart = start + len;
7709                 goto unlock;
7710         }
7711
7712         if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags) ||
7713             ((BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW) &&
7714              em->block_start != EXTENT_MAP_HOLE)) {
7715                 int type;
7716                 u64 block_start, orig_start, orig_block_len, ram_bytes;
7717
7718                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
7719                         type = BTRFS_ORDERED_PREALLOC;
7720                 else
7721                         type = BTRFS_ORDERED_NOCOW;
7722                 len = min(len, em->len - (start - em->start));
7723                 block_start = em->block_start + (start - em->start);
7724
7725                 if (can_nocow_extent(inode, start, &len, &orig_start,
7726                                      &orig_block_len, &ram_bytes) == 1 &&
7727                     btrfs_inc_nocow_writers(root->fs_info, block_start)) {
7728                         struct extent_map *em2;
7729
7730                         em2 = btrfs_create_dio_extent(inode, start, len,
7731                                                       orig_start, block_start,
7732                                                       len, orig_block_len,
7733                                                       ram_bytes, type);
7734                         btrfs_dec_nocow_writers(root->fs_info, block_start);
7735                         if (type == BTRFS_ORDERED_PREALLOC) {
7736                                 free_extent_map(em);
7737                                 em = em2;
7738                         }
7739                         if (em2 && IS_ERR(em2)) {
7740                                 ret = PTR_ERR(em2);
7741                                 goto unlock_err;
7742                         }
7743                         goto unlock;
7744                 }
7745         }
7746
7747         /*
7748          * this will cow the extent, reset the len in case we changed
7749          * it above
7750          */
7751         len = bh_result->b_size;
7752         free_extent_map(em);
7753         em = btrfs_new_extent_direct(inode, start, len);
7754         if (IS_ERR(em)) {
7755                 ret = PTR_ERR(em);
7756                 goto unlock_err;
7757         }
7758         len = min(len, em->len - (start - em->start));
7759 unlock:
7760         bh_result->b_blocknr = (em->block_start + (start - em->start)) >>
7761                 inode->i_blkbits;
7762         bh_result->b_size = len;
7763         bh_result->b_bdev = em->bdev;
7764         set_buffer_mapped(bh_result);
7765         if (create) {
7766                 if (!test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
7767                         set_buffer_new(bh_result);
7768
7769                 /*
7770                  * Need to update the i_size under the extent lock so buffered
7771                  * readers will get the updated i_size when we unlock.
7772                  */
7773                 if (start + len > i_size_read(inode))
7774                         i_size_write(inode, start + len);
7775
7776                 adjust_dio_outstanding_extents(inode, dio_data, len);
7777                 btrfs_free_reserved_data_space(inode, start, len);
7778                 WARN_ON(dio_data->reserve < len);
7779                 dio_data->reserve -= len;
7780                 dio_data->unsubmitted_oe_range_end = start + len;
7781                 current->journal_info = dio_data;
7782         }
7783
7784         /*
7785          * In the case of write we need to clear and unlock the entire range,
7786          * in the case of read we need to unlock only the end area that we
7787          * aren't using if there is any left over space.
7788          */
7789         if (lockstart < lockend) {
7790                 clear_extent_bit(&BTRFS_I(inode)->io_tree, lockstart,
7791                                  lockend, unlock_bits, 1, 0,
7792                                  &cached_state, GFP_NOFS);
7793         } else {
7794                 free_extent_state(cached_state);
7795         }
7796
7797         free_extent_map(em);
7798
7799         return 0;
7800
7801 unlock_err:
7802         clear_extent_bit(&BTRFS_I(inode)->io_tree, lockstart, lockend,
7803                          unlock_bits, 1, 0, &cached_state, GFP_NOFS);
7804 err:
7805         if (dio_data)
7806                 current->journal_info = dio_data;
7807         /*
7808          * Compensate the delalloc release we do in btrfs_direct_IO() when we
7809          * write less data then expected, so that we don't underflow our inode's
7810          * outstanding extents counter.
7811          */
7812         if (create && dio_data)
7813                 adjust_dio_outstanding_extents(inode, dio_data, len);
7814
7815         return ret;
7816 }
7817
7818 static inline int submit_dio_repair_bio(struct inode *inode, struct bio *bio,
7819                                         int rw, int mirror_num)
7820 {
7821         struct btrfs_root *root = BTRFS_I(inode)->root;
7822         int ret;
7823
7824         BUG_ON(rw & REQ_WRITE);
7825
7826         bio_get(bio);
7827
7828         ret = btrfs_bio_wq_end_io(root->fs_info, bio,
7829                                   BTRFS_WQ_ENDIO_DIO_REPAIR);
7830         if (ret)
7831                 goto err;
7832
7833         ret = btrfs_map_bio(root, rw, bio, mirror_num, 0);
7834 err:
7835         bio_put(bio);
7836         return ret;
7837 }
7838
7839 static int btrfs_check_dio_repairable(struct inode *inode,
7840                                       struct bio *failed_bio,
7841                                       struct io_failure_record *failrec,
7842                                       int failed_mirror)
7843 {
7844         int num_copies;
7845
7846         num_copies = btrfs_num_copies(BTRFS_I(inode)->root->fs_info,
7847                                       failrec->logical, failrec->len);
7848         if (num_copies == 1) {
7849                 /*
7850                  * we only have a single copy of the data, so don't bother with
7851                  * all the retry and error correction code that follows. no
7852                  * matter what the error is, it is very likely to persist.
7853                  */
7854                 pr_debug("Check DIO Repairable: cannot repair, num_copies=%d, next_mirror %d, failed_mirror %d\n",
7855                          num_copies, failrec->this_mirror, failed_mirror);
7856                 return 0;
7857         }
7858
7859         failrec->failed_mirror = failed_mirror;
7860         failrec->this_mirror++;
7861         if (failrec->this_mirror == failed_mirror)
7862                 failrec->this_mirror++;
7863
7864         if (failrec->this_mirror > num_copies) {
7865                 pr_debug("Check DIO Repairable: (fail) num_copies=%d, next_mirror %d, failed_mirror %d\n",
7866                          num_copies, failrec->this_mirror, failed_mirror);
7867                 return 0;
7868         }
7869
7870         return 1;
7871 }
7872
7873 static int dio_read_error(struct inode *inode, struct bio *failed_bio,
7874                         struct page *page, unsigned int pgoff,
7875                         u64 start, u64 end, int failed_mirror,
7876                         bio_end_io_t *repair_endio, void *repair_arg)
7877 {
7878         struct io_failure_record *failrec;
7879         struct bio *bio;
7880         int isector;
7881         int read_mode;
7882         int ret;
7883
7884         BUG_ON(failed_bio->bi_rw & REQ_WRITE);
7885
7886         ret = btrfs_get_io_failure_record(inode, start, end, &failrec);
7887         if (ret)
7888                 return ret;
7889
7890         ret = btrfs_check_dio_repairable(inode, failed_bio, failrec,
7891                                          failed_mirror);
7892         if (!ret) {
7893                 free_io_failure(inode, failrec);
7894                 return -EIO;
7895         }
7896
7897         if ((failed_bio->bi_vcnt > 1)
7898                 || (failed_bio->bi_io_vec->bv_len
7899                         > BTRFS_I(inode)->root->sectorsize))
7900                 read_mode = READ_SYNC | REQ_FAILFAST_DEV;
7901         else
7902                 read_mode = READ_SYNC;
7903
7904         isector = start - btrfs_io_bio(failed_bio)->logical;
7905         isector >>= inode->i_sb->s_blocksize_bits;
7906         bio = btrfs_create_repair_bio(inode, failed_bio, failrec, page,
7907                                 pgoff, isector, repair_endio, repair_arg);
7908         if (!bio) {
7909                 free_io_failure(inode, failrec);
7910                 return -EIO;
7911         }
7912
7913         btrfs_debug(BTRFS_I(inode)->root->fs_info,
7914                     "Repair DIO Read Error: submitting new dio read[%#x] to this_mirror=%d, in_validation=%d\n",
7915                     read_mode, failrec->this_mirror, failrec->in_validation);
7916
7917         ret = submit_dio_repair_bio(inode, bio, read_mode,
7918                                     failrec->this_mirror);
7919         if (ret) {
7920                 free_io_failure(inode, failrec);
7921                 bio_put(bio);
7922         }
7923
7924         return ret;
7925 }
7926
7927 struct btrfs_retry_complete {
7928         struct completion done;
7929         struct inode *inode;
7930         u64 start;
7931         int uptodate;
7932 };
7933
7934 static void btrfs_retry_endio_nocsum(struct bio *bio)
7935 {
7936         struct btrfs_retry_complete *done = bio->bi_private;
7937         struct inode *inode;
7938         struct bio_vec *bvec;
7939         int i;
7940
7941         if (bio->bi_error)
7942                 goto end;
7943
7944         ASSERT(bio->bi_vcnt == 1);
7945         inode = bio->bi_io_vec->bv_page->mapping->host;
7946         ASSERT(bio->bi_io_vec->bv_len == BTRFS_I(inode)->root->sectorsize);
7947
7948         done->uptodate = 1;
7949         bio_for_each_segment_all(bvec, bio, i)
7950                 clean_io_failure(done->inode, done->start, bvec->bv_page, 0);
7951 end:
7952         complete(&done->done);
7953         bio_put(bio);
7954 }
7955
7956 static int __btrfs_correct_data_nocsum(struct inode *inode,
7957                                        struct btrfs_io_bio *io_bio)
7958 {
7959         struct btrfs_fs_info *fs_info;
7960         struct bio_vec *bvec;
7961         struct btrfs_retry_complete done;
7962         u64 start;
7963         unsigned int pgoff;
7964         u32 sectorsize;
7965         int nr_sectors;
7966         int i;
7967         int ret;
7968
7969         fs_info = BTRFS_I(inode)->root->fs_info;
7970         sectorsize = BTRFS_I(inode)->root->sectorsize;
7971
7972         start = io_bio->logical;
7973         done.inode = inode;
7974
7975         bio_for_each_segment_all(bvec, &io_bio->bio, i) {
7976                 nr_sectors = BTRFS_BYTES_TO_BLKS(fs_info, bvec->bv_len);
7977                 pgoff = bvec->bv_offset;
7978
7979 next_block_or_try_again:
7980                 done.uptodate = 0;
7981                 done.start = start;
7982                 init_completion(&done.done);
7983
7984                 ret = dio_read_error(inode, &io_bio->bio, bvec->bv_page,
7985                                 pgoff, start, start + sectorsize - 1,
7986                                 io_bio->mirror_num,
7987                                 btrfs_retry_endio_nocsum, &done);
7988                 if (ret)
7989                         return ret;
7990
7991                 wait_for_completion(&done.done);
7992
7993                 if (!done.uptodate) {
7994                         /* We might have another mirror, so try again */
7995                         goto next_block_or_try_again;
7996                 }
7997
7998                 start += sectorsize;
7999
8000                 if (nr_sectors--) {
8001                         pgoff += sectorsize;
8002                         goto next_block_or_try_again;
8003                 }
8004         }
8005
8006         return 0;
8007 }
8008
8009 static void btrfs_retry_endio(struct bio *bio)
8010 {
8011         struct btrfs_retry_complete *done = bio->bi_private;
8012         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
8013         struct inode *inode;
8014         struct bio_vec *bvec;
8015         u64 start;
8016         int uptodate;
8017         int ret;
8018         int i;
8019
8020         if (bio->bi_error)
8021                 goto end;
8022
8023         uptodate = 1;
8024
8025         start = done->start;
8026
8027         ASSERT(bio->bi_vcnt == 1);
8028         inode = bio->bi_io_vec->bv_page->mapping->host;
8029         ASSERT(bio->bi_io_vec->bv_len == BTRFS_I(inode)->root->sectorsize);
8030
8031         bio_for_each_segment_all(bvec, bio, i) {
8032                 ret = __readpage_endio_check(done->inode, io_bio, i,
8033                                         bvec->bv_page, bvec->bv_offset,
8034                                         done->start, bvec->bv_len);
8035                 if (!ret)
8036                         clean_io_failure(done->inode, done->start,
8037                                         bvec->bv_page, bvec->bv_offset);
8038                 else
8039                         uptodate = 0;
8040         }
8041
8042         done->uptodate = uptodate;
8043 end:
8044         complete(&done->done);
8045         bio_put(bio);
8046 }
8047
8048 static int __btrfs_subio_endio_read(struct inode *inode,
8049                                     struct btrfs_io_bio *io_bio, int err)
8050 {
8051         struct btrfs_fs_info *fs_info;
8052         struct bio_vec *bvec;
8053         struct btrfs_retry_complete done;
8054         u64 start;
8055         u64 offset = 0;
8056         u32 sectorsize;
8057         int nr_sectors;
8058         unsigned int pgoff;
8059         int csum_pos;
8060         int i;
8061         int ret;
8062
8063         fs_info = BTRFS_I(inode)->root->fs_info;
8064         sectorsize = BTRFS_I(inode)->root->sectorsize;
8065
8066         err = 0;
8067         start = io_bio->logical;
8068         done.inode = inode;
8069
8070         bio_for_each_segment_all(bvec, &io_bio->bio, i) {
8071                 nr_sectors = BTRFS_BYTES_TO_BLKS(fs_info, bvec->bv_len);
8072
8073                 pgoff = bvec->bv_offset;
8074 next_block:
8075                 csum_pos = BTRFS_BYTES_TO_BLKS(fs_info, offset);
8076                 ret = __readpage_endio_check(inode, io_bio, csum_pos,
8077                                         bvec->bv_page, pgoff, start,
8078                                         sectorsize);
8079                 if (likely(!ret))
8080                         goto next;
8081 try_again:
8082                 done.uptodate = 0;
8083                 done.start = start;
8084                 init_completion(&done.done);
8085
8086                 ret = dio_read_error(inode, &io_bio->bio, bvec->bv_page,
8087                                 pgoff, start, start + sectorsize - 1,
8088                                 io_bio->mirror_num,
8089                                 btrfs_retry_endio, &done);
8090                 if (ret) {
8091                         err = ret;
8092                         goto next;
8093                 }
8094
8095                 wait_for_completion(&done.done);
8096
8097                 if (!done.uptodate) {
8098                         /* We might have another mirror, so try again */
8099                         goto try_again;
8100                 }
8101 next:
8102                 offset += sectorsize;
8103                 start += sectorsize;
8104
8105                 ASSERT(nr_sectors);
8106
8107                 if (--nr_sectors) {
8108                         pgoff += sectorsize;
8109                         goto next_block;
8110                 }
8111         }
8112
8113         return err;
8114 }
8115
8116 static int btrfs_subio_endio_read(struct inode *inode,
8117                                   struct btrfs_io_bio *io_bio, int err)
8118 {
8119         bool skip_csum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM;
8120
8121         if (skip_csum) {
8122                 if (unlikely(err))
8123                         return __btrfs_correct_data_nocsum(inode, io_bio);
8124                 else
8125                         return 0;
8126         } else {
8127                 return __btrfs_subio_endio_read(inode, io_bio, err);
8128         }
8129 }
8130
8131 static void btrfs_endio_direct_read(struct bio *bio)
8132 {
8133         struct btrfs_dio_private *dip = bio->bi_private;
8134         struct inode *inode = dip->inode;
8135         struct bio *dio_bio;
8136         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
8137         int err = bio->bi_error;
8138
8139         if (dip->flags & BTRFS_DIO_ORIG_BIO_SUBMITTED)
8140                 err = btrfs_subio_endio_read(inode, io_bio, err);
8141
8142         unlock_extent(&BTRFS_I(inode)->io_tree, dip->logical_offset,
8143                       dip->logical_offset + dip->bytes - 1);
8144         dio_bio = dip->dio_bio;
8145
8146         kfree(dip);
8147
8148         dio_bio->bi_error = bio->bi_error;
8149         dio_end_io(dio_bio, bio->bi_error);
8150
8151         if (io_bio->end_io)
8152                 io_bio->end_io(io_bio, err);
8153         bio_put(bio);
8154 }
8155
8156 static void btrfs_endio_direct_write_update_ordered(struct inode *inode,
8157                                                     const u64 offset,
8158                                                     const u64 bytes,
8159                                                     const int uptodate)
8160 {
8161         struct btrfs_root *root = BTRFS_I(inode)->root;
8162         struct btrfs_ordered_extent *ordered = NULL;
8163         u64 ordered_offset = offset;
8164         u64 ordered_bytes = bytes;
8165         int ret;
8166
8167 again:
8168         ret = btrfs_dec_test_first_ordered_pending(inode, &ordered,
8169                                                    &ordered_offset,
8170                                                    ordered_bytes,
8171                                                    uptodate);
8172         if (!ret)
8173                 goto out_test;
8174
8175         btrfs_init_work(&ordered->work, btrfs_endio_write_helper,
8176                         finish_ordered_fn, NULL, NULL);
8177         btrfs_queue_work(root->fs_info->endio_write_workers,
8178                          &ordered->work);
8179 out_test:
8180         /*
8181          * our bio might span multiple ordered extents.  If we haven't
8182          * completed the accounting for the whole dio, go back and try again
8183          */
8184         if (ordered_offset < offset + bytes) {
8185                 ordered_bytes = offset + bytes - ordered_offset;
8186                 ordered = NULL;
8187                 goto again;
8188         }
8189 }
8190
8191 static void btrfs_endio_direct_write(struct bio *bio)
8192 {
8193         struct btrfs_dio_private *dip = bio->bi_private;
8194         struct bio *dio_bio = dip->dio_bio;
8195
8196         btrfs_endio_direct_write_update_ordered(dip->inode,
8197                                                 dip->logical_offset,
8198                                                 dip->bytes,
8199                                                 !bio->bi_error);
8200
8201         kfree(dip);
8202
8203         dio_bio->bi_error = bio->bi_error;
8204         dio_end_io(dio_bio, bio->bi_error);
8205         bio_put(bio);
8206 }
8207
8208 static int __btrfs_submit_bio_start_direct_io(struct inode *inode, int rw,
8209                                     struct bio *bio, int mirror_num,
8210                                     unsigned long bio_flags, u64 offset)
8211 {
8212         int ret;
8213         struct btrfs_root *root = BTRFS_I(inode)->root;
8214         ret = btrfs_csum_one_bio(root, inode, bio, offset, 1);
8215         BUG_ON(ret); /* -ENOMEM */
8216         return 0;
8217 }
8218
8219 static void btrfs_end_dio_bio(struct bio *bio)
8220 {
8221         struct btrfs_dio_private *dip = bio->bi_private;
8222         int err = bio->bi_error;
8223
8224         if (err)
8225                 btrfs_warn(BTRFS_I(dip->inode)->root->fs_info,
8226                            "direct IO failed ino %llu rw %lu sector %#Lx len %u err no %d",
8227                            btrfs_ino(dip->inode), bio->bi_rw,
8228                            (unsigned long long)bio->bi_iter.bi_sector,
8229                            bio->bi_iter.bi_size, err);
8230
8231         if (dip->subio_endio)
8232                 err = dip->subio_endio(dip->inode, btrfs_io_bio(bio), err);
8233
8234         if (err) {
8235                 dip->errors = 1;
8236
8237                 /*
8238                  * before atomic variable goto zero, we must make sure
8239                  * dip->errors is perceived to be set.
8240                  */
8241                 smp_mb__before_atomic();
8242         }
8243
8244         /* if there are more bios still pending for this dio, just exit */
8245         if (!atomic_dec_and_test(&dip->pending_bios))
8246                 goto out;
8247
8248         if (dip->errors) {
8249                 bio_io_error(dip->orig_bio);
8250         } else {
8251                 dip->dio_bio->bi_error = 0;
8252                 bio_endio(dip->orig_bio);
8253         }
8254 out:
8255         bio_put(bio);
8256 }
8257
8258 static struct bio *btrfs_dio_bio_alloc(struct block_device *bdev,
8259                                        u64 first_sector, gfp_t gfp_flags)
8260 {
8261         struct bio *bio;
8262         bio = btrfs_bio_alloc(bdev, first_sector, BIO_MAX_PAGES, gfp_flags);
8263         if (bio)
8264                 bio_associate_current(bio);
8265         return bio;
8266 }
8267
8268 static inline int btrfs_lookup_and_bind_dio_csum(struct btrfs_root *root,
8269                                                  struct inode *inode,
8270                                                  struct btrfs_dio_private *dip,
8271                                                  struct bio *bio,
8272                                                  u64 file_offset)
8273 {
8274         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
8275         struct btrfs_io_bio *orig_io_bio = btrfs_io_bio(dip->orig_bio);
8276         int ret;
8277
8278         /*
8279          * We load all the csum data we need when we submit
8280          * the first bio to reduce the csum tree search and
8281          * contention.
8282          */
8283         if (dip->logical_offset == file_offset) {
8284                 ret = btrfs_lookup_bio_sums_dio(root, inode, dip->orig_bio,
8285                                                 file_offset);
8286                 if (ret)
8287                         return ret;
8288         }
8289
8290         if (bio == dip->orig_bio)
8291                 return 0;
8292
8293         file_offset -= dip->logical_offset;
8294         file_offset >>= inode->i_sb->s_blocksize_bits;
8295         io_bio->csum = (u8 *)(((u32 *)orig_io_bio->csum) + file_offset);
8296
8297         return 0;
8298 }
8299
8300 static inline int __btrfs_submit_dio_bio(struct bio *bio, struct inode *inode,
8301                                          int rw, u64 file_offset, int skip_sum,
8302                                          int async_submit)
8303 {
8304         struct btrfs_dio_private *dip = bio->bi_private;
8305         int write = rw & REQ_WRITE;
8306         struct btrfs_root *root = BTRFS_I(inode)->root;
8307         int ret;
8308
8309         if (async_submit)
8310                 async_submit = !atomic_read(&BTRFS_I(inode)->sync_writers);
8311
8312         bio_get(bio);
8313
8314         if (!write) {
8315                 ret = btrfs_bio_wq_end_io(root->fs_info, bio,
8316                                 BTRFS_WQ_ENDIO_DATA);
8317                 if (ret)
8318                         goto err;
8319         }
8320
8321         if (skip_sum)
8322                 goto map;
8323
8324         if (write && async_submit) {
8325                 ret = btrfs_wq_submit_bio(root->fs_info,
8326                                    inode, rw, bio, 0, 0,
8327                                    file_offset,
8328                                    __btrfs_submit_bio_start_direct_io,
8329                                    __btrfs_submit_bio_done);
8330                 goto err;
8331         } else if (write) {
8332                 /*
8333                  * If we aren't doing async submit, calculate the csum of the
8334                  * bio now.
8335                  */
8336                 ret = btrfs_csum_one_bio(root, inode, bio, file_offset, 1);
8337                 if (ret)
8338                         goto err;
8339         } else {
8340                 ret = btrfs_lookup_and_bind_dio_csum(root, inode, dip, bio,
8341                                                      file_offset);
8342                 if (ret)
8343                         goto err;
8344         }
8345 map:
8346         ret = btrfs_map_bio(root, rw, bio, 0, async_submit);
8347 err:
8348         bio_put(bio);
8349         return ret;
8350 }
8351
8352 static int btrfs_submit_direct_hook(int rw, struct btrfs_dio_private *dip,
8353                                     int skip_sum)
8354 {
8355         struct inode *inode = dip->inode;
8356         struct btrfs_root *root = BTRFS_I(inode)->root;
8357         struct bio *bio;
8358         struct bio *orig_bio = dip->orig_bio;
8359         struct bio_vec *bvec = orig_bio->bi_io_vec;
8360         u64 start_sector = orig_bio->bi_iter.bi_sector;
8361         u64 file_offset = dip->logical_offset;
8362         u64 submit_len = 0;
8363         u64 map_length;
8364         u32 blocksize = root->sectorsize;
8365         int async_submit = 0;
8366         int nr_sectors;
8367         int ret;
8368         int i;
8369
8370         map_length = orig_bio->bi_iter.bi_size;
8371         ret = btrfs_map_block(root->fs_info, rw, start_sector << 9,
8372                               &map_length, NULL, 0);
8373         if (ret)
8374                 return -EIO;
8375
8376         if (map_length >= orig_bio->bi_iter.bi_size) {
8377                 bio = orig_bio;
8378                 dip->flags |= BTRFS_DIO_ORIG_BIO_SUBMITTED;
8379                 goto submit;
8380         }
8381
8382         /* async crcs make it difficult to collect full stripe writes. */
8383         if (btrfs_get_alloc_profile(root, 1) & BTRFS_BLOCK_GROUP_RAID56_MASK)
8384                 async_submit = 0;
8385         else
8386                 async_submit = 1;
8387
8388         bio = btrfs_dio_bio_alloc(orig_bio->bi_bdev, start_sector, GFP_NOFS);
8389         if (!bio)
8390                 return -ENOMEM;
8391
8392         bio->bi_private = dip;
8393         bio->bi_end_io = btrfs_end_dio_bio;
8394         btrfs_io_bio(bio)->logical = file_offset;
8395         atomic_inc(&dip->pending_bios);
8396
8397         while (bvec <= (orig_bio->bi_io_vec + orig_bio->bi_vcnt - 1)) {
8398                 nr_sectors = BTRFS_BYTES_TO_BLKS(root->fs_info, bvec->bv_len);
8399                 i = 0;
8400 next_block:
8401                 if (unlikely(map_length < submit_len + blocksize ||
8402                     bio_add_page(bio, bvec->bv_page, blocksize,
8403                             bvec->bv_offset + (i * blocksize)) < blocksize)) {
8404                         /*
8405                          * inc the count before we submit the bio so
8406                          * we know the end IO handler won't happen before
8407                          * we inc the count. Otherwise, the dip might get freed
8408                          * before we're done setting it up
8409                          */
8410                         atomic_inc(&dip->pending_bios);
8411                         ret = __btrfs_submit_dio_bio(bio, inode, rw,
8412                                                      file_offset, skip_sum,
8413                                                      async_submit);
8414                         if (ret) {
8415                                 bio_put(bio);
8416                                 atomic_dec(&dip->pending_bios);
8417                                 goto out_err;
8418                         }
8419
8420                         start_sector += submit_len >> 9;
8421                         file_offset += submit_len;
8422
8423                         submit_len = 0;
8424
8425                         bio = btrfs_dio_bio_alloc(orig_bio->bi_bdev,
8426                                                   start_sector, GFP_NOFS);
8427                         if (!bio)
8428                                 goto out_err;
8429                         bio->bi_private = dip;
8430                         bio->bi_end_io = btrfs_end_dio_bio;
8431                         btrfs_io_bio(bio)->logical = file_offset;
8432
8433                         map_length = orig_bio->bi_iter.bi_size;
8434                         ret = btrfs_map_block(root->fs_info, rw,
8435                                               start_sector << 9,
8436                                               &map_length, NULL, 0);
8437                         if (ret) {
8438                                 bio_put(bio);
8439                                 goto out_err;
8440                         }
8441
8442                         goto next_block;
8443                 } else {
8444                         submit_len += blocksize;
8445                         if (--nr_sectors) {
8446                                 i++;
8447                                 goto next_block;
8448                         }
8449                         bvec++;
8450                 }
8451         }
8452
8453 submit:
8454         ret = __btrfs_submit_dio_bio(bio, inode, rw, file_offset, skip_sum,
8455                                      async_submit);
8456         if (!ret)
8457                 return 0;
8458
8459         bio_put(bio);
8460 out_err:
8461         dip->errors = 1;
8462         /*
8463          * before atomic variable goto zero, we must
8464          * make sure dip->errors is perceived to be set.
8465          */
8466         smp_mb__before_atomic();
8467         if (atomic_dec_and_test(&dip->pending_bios))
8468                 bio_io_error(dip->orig_bio);
8469
8470         /* bio_end_io() will handle error, so we needn't return it */
8471         return 0;
8472 }
8473
8474 static void btrfs_submit_direct(int rw, struct bio *dio_bio,
8475                                 struct inode *inode, loff_t file_offset)
8476 {
8477         struct btrfs_dio_private *dip = NULL;
8478         struct bio *io_bio = NULL;
8479         struct btrfs_io_bio *btrfs_bio;
8480         int skip_sum;
8481         int write = rw & REQ_WRITE;
8482         int ret = 0;
8483
8484         skip_sum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM;
8485
8486         io_bio = btrfs_bio_clone(dio_bio, GFP_NOFS);
8487         if (!io_bio) {
8488                 ret = -ENOMEM;
8489                 goto free_ordered;
8490         }
8491
8492         dip = kzalloc(sizeof(*dip), GFP_NOFS);
8493         if (!dip) {
8494                 ret = -ENOMEM;
8495                 goto free_ordered;
8496         }
8497
8498         dip->private = dio_bio->bi_private;
8499         dip->inode = inode;
8500         dip->logical_offset = file_offset;
8501         dip->bytes = dio_bio->bi_iter.bi_size;
8502         dip->disk_bytenr = (u64)dio_bio->bi_iter.bi_sector << 9;
8503         io_bio->bi_private = dip;
8504         dip->orig_bio = io_bio;
8505         dip->dio_bio = dio_bio;
8506         atomic_set(&dip->pending_bios, 0);
8507         btrfs_bio = btrfs_io_bio(io_bio);
8508         btrfs_bio->logical = file_offset;
8509
8510         if (write) {
8511                 io_bio->bi_end_io = btrfs_endio_direct_write;
8512         } else {
8513                 io_bio->bi_end_io = btrfs_endio_direct_read;
8514                 dip->subio_endio = btrfs_subio_endio_read;
8515         }
8516
8517         /*
8518          * Reset the range for unsubmitted ordered extents (to a 0 length range)
8519          * even if we fail to submit a bio, because in such case we do the
8520          * corresponding error handling below and it must not be done a second
8521          * time by btrfs_direct_IO().
8522          */
8523         if (write) {
8524                 struct btrfs_dio_data *dio_data = current->journal_info;
8525
8526                 dio_data->unsubmitted_oe_range_end = dip->logical_offset +
8527                         dip->bytes;
8528                 dio_data->unsubmitted_oe_range_start =
8529                         dio_data->unsubmitted_oe_range_end;
8530         }
8531
8532         ret = btrfs_submit_direct_hook(rw, dip, skip_sum);
8533         if (!ret)
8534                 return;
8535
8536         if (btrfs_bio->end_io)
8537                 btrfs_bio->end_io(btrfs_bio, ret);
8538
8539 free_ordered:
8540         /*
8541          * If we arrived here it means either we failed to submit the dip
8542          * or we either failed to clone the dio_bio or failed to allocate the
8543          * dip. If we cloned the dio_bio and allocated the dip, we can just
8544          * call bio_endio against our io_bio so that we get proper resource
8545          * cleanup if we fail to submit the dip, otherwise, we must do the
8546          * same as btrfs_endio_direct_[write|read] because we can't call these
8547          * callbacks - they require an allocated dip and a clone of dio_bio.
8548          */
8549         if (io_bio && dip) {
8550                 io_bio->bi_error = -EIO;
8551                 bio_endio(io_bio);
8552                 /*
8553                  * The end io callbacks free our dip, do the final put on io_bio
8554                  * and all the cleanup and final put for dio_bio (through
8555                  * dio_end_io()).
8556                  */
8557                 dip = NULL;
8558                 io_bio = NULL;
8559         } else {
8560                 if (write)
8561                         btrfs_endio_direct_write_update_ordered(inode,
8562                                                 file_offset,
8563                                                 dio_bio->bi_iter.bi_size,
8564                                                 0);
8565                 else
8566                         unlock_extent(&BTRFS_I(inode)->io_tree, file_offset,
8567                               file_offset + dio_bio->bi_iter.bi_size - 1);
8568
8569                 dio_bio->bi_error = -EIO;
8570                 /*
8571                  * Releases and cleans up our dio_bio, no need to bio_put()
8572                  * nor bio_endio()/bio_io_error() against dio_bio.
8573                  */
8574                 dio_end_io(dio_bio, ret);
8575         }
8576         if (io_bio)
8577                 bio_put(io_bio);
8578         kfree(dip);
8579 }
8580
8581 static ssize_t check_direct_IO(struct btrfs_root *root, struct kiocb *iocb,
8582                         const struct iov_iter *iter, loff_t offset)
8583 {
8584         int seg;
8585         int i;
8586         unsigned blocksize_mask = root->sectorsize - 1;
8587         ssize_t retval = -EINVAL;
8588
8589         if (offset & blocksize_mask)
8590                 goto out;
8591
8592         if (iov_iter_alignment(iter) & blocksize_mask)
8593                 goto out;
8594
8595         /* If this is a write we don't need to check anymore */
8596         if (iov_iter_rw(iter) == WRITE)
8597                 return 0;
8598         /*
8599          * Check to make sure we don't have duplicate iov_base's in this
8600          * iovec, if so return EINVAL, otherwise we'll get csum errors
8601          * when reading back.
8602          */
8603         for (seg = 0; seg < iter->nr_segs; seg++) {
8604                 for (i = seg + 1; i < iter->nr_segs; i++) {
8605                         if (iter->iov[seg].iov_base == iter->iov[i].iov_base)
8606                                 goto out;
8607                 }
8608         }
8609         retval = 0;
8610 out:
8611         return retval;
8612 }
8613
8614 static ssize_t btrfs_direct_IO(struct kiocb *iocb, struct iov_iter *iter)
8615 {
8616         struct file *file = iocb->ki_filp;
8617         struct inode *inode = file->f_mapping->host;
8618         struct btrfs_root *root = BTRFS_I(inode)->root;
8619         struct btrfs_dio_data dio_data = { 0 };
8620         loff_t offset = iocb->ki_pos;
8621         size_t count = 0;
8622         int flags = 0;
8623         bool wakeup = true;
8624         bool relock = false;
8625         ssize_t ret;
8626
8627         if (check_direct_IO(BTRFS_I(inode)->root, iocb, iter, offset))
8628                 return 0;
8629
8630         inode_dio_begin(inode);
8631         smp_mb__after_atomic();
8632
8633         /*
8634          * The generic stuff only does filemap_write_and_wait_range, which
8635          * isn't enough if we've written compressed pages to this area, so
8636          * we need to flush the dirty pages again to make absolutely sure
8637          * that any outstanding dirty pages are on disk.
8638          */
8639         count = iov_iter_count(iter);
8640         if (test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
8641                      &BTRFS_I(inode)->runtime_flags))
8642                 filemap_fdatawrite_range(inode->i_mapping, offset,
8643                                          offset + count - 1);
8644
8645         if (iov_iter_rw(iter) == WRITE) {
8646                 /*
8647                  * If the write DIO is beyond the EOF, we need update
8648                  * the isize, but it is protected by i_mutex. So we can
8649                  * not unlock the i_mutex at this case.
8650                  */
8651                 if (offset + count <= inode->i_size) {
8652                         inode_unlock(inode);
8653                         relock = true;
8654                 }
8655                 ret = btrfs_delalloc_reserve_space(inode, offset, count);
8656                 if (ret)
8657                         goto out;
8658                 dio_data.outstanding_extents = div64_u64(count +
8659                                                 BTRFS_MAX_EXTENT_SIZE - 1,
8660                                                 BTRFS_MAX_EXTENT_SIZE);
8661
8662                 /*
8663                  * We need to know how many extents we reserved so that we can
8664                  * do the accounting properly if we go over the number we
8665                  * originally calculated.  Abuse current->journal_info for this.
8666                  */
8667                 dio_data.reserve = round_up(count, root->sectorsize);
8668                 dio_data.unsubmitted_oe_range_start = (u64)offset;
8669                 dio_data.unsubmitted_oe_range_end = (u64)offset;
8670                 current->journal_info = &dio_data;
8671         } else if (test_bit(BTRFS_INODE_READDIO_NEED_LOCK,
8672                                      &BTRFS_I(inode)->runtime_flags)) {
8673                 inode_dio_end(inode);
8674                 flags = DIO_LOCKING | DIO_SKIP_HOLES;
8675                 wakeup = false;
8676         }
8677
8678         ret = __blockdev_direct_IO(iocb, inode,
8679                                    BTRFS_I(inode)->root->fs_info->fs_devices->latest_bdev,
8680                                    iter, btrfs_get_blocks_direct, NULL,
8681                                    btrfs_submit_direct, flags);
8682         if (iov_iter_rw(iter) == WRITE) {
8683                 current->journal_info = NULL;
8684                 if (ret < 0 && ret != -EIOCBQUEUED) {
8685                         if (dio_data.reserve)
8686                                 btrfs_delalloc_release_space(inode, offset,
8687                                                              dio_data.reserve);
8688                         /*
8689                          * On error we might have left some ordered extents
8690                          * without submitting corresponding bios for them, so
8691                          * cleanup them up to avoid other tasks getting them
8692                          * and waiting for them to complete forever.
8693                          */
8694                         if (dio_data.unsubmitted_oe_range_start <
8695                             dio_data.unsubmitted_oe_range_end)
8696                                 btrfs_endio_direct_write_update_ordered(inode,
8697                                         dio_data.unsubmitted_oe_range_start,
8698                                         dio_data.unsubmitted_oe_range_end -
8699                                         dio_data.unsubmitted_oe_range_start,
8700                                         0);
8701                 } else if (ret >= 0 && (size_t)ret < count)
8702                         btrfs_delalloc_release_space(inode, offset,
8703                                                      count - (size_t)ret);
8704         }
8705 out:
8706         if (wakeup)
8707                 inode_dio_end(inode);
8708         if (relock)
8709                 inode_lock(inode);
8710
8711         return ret;
8712 }
8713
8714 #define BTRFS_FIEMAP_FLAGS      (FIEMAP_FLAG_SYNC)
8715
8716 static int btrfs_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
8717                 __u64 start, __u64 len)
8718 {
8719         int     ret;
8720
8721         ret = fiemap_check_flags(fieinfo, BTRFS_FIEMAP_FLAGS);
8722         if (ret)
8723                 return ret;
8724
8725         return extent_fiemap(inode, fieinfo, start, len, btrfs_get_extent_fiemap);
8726 }
8727
8728 int btrfs_readpage(struct file *file, struct page *page)
8729 {
8730         struct extent_io_tree *tree;
8731         tree = &BTRFS_I(page->mapping->host)->io_tree;
8732         return extent_read_full_page(tree, page, btrfs_get_extent, 0);
8733 }
8734
8735 static int btrfs_writepage(struct page *page, struct writeback_control *wbc)
8736 {
8737         struct extent_io_tree *tree;
8738         struct inode *inode = page->mapping->host;
8739         int ret;
8740
8741         if (current->flags & PF_MEMALLOC) {
8742                 redirty_page_for_writepage(wbc, page);
8743                 unlock_page(page);
8744                 return 0;
8745         }
8746
8747         /*
8748          * If we are under memory pressure we will call this directly from the
8749          * VM, we need to make sure we have the inode referenced for the ordered
8750          * extent.  If not just return like we didn't do anything.
8751          */
8752         if (!igrab(inode)) {
8753                 redirty_page_for_writepage(wbc, page);
8754                 return AOP_WRITEPAGE_ACTIVATE;
8755         }
8756         tree = &BTRFS_I(page->mapping->host)->io_tree;
8757         ret = extent_write_full_page(tree, page, btrfs_get_extent, wbc);
8758         btrfs_add_delayed_iput(inode);
8759         return ret;
8760 }
8761
8762 static int btrfs_writepages(struct address_space *mapping,
8763                             struct writeback_control *wbc)
8764 {
8765         struct extent_io_tree *tree;
8766
8767         tree = &BTRFS_I(mapping->host)->io_tree;
8768         return extent_writepages(tree, mapping, btrfs_get_extent, wbc);
8769 }
8770
8771 static int
8772 btrfs_readpages(struct file *file, struct address_space *mapping,
8773                 struct list_head *pages, unsigned nr_pages)
8774 {
8775         struct extent_io_tree *tree;
8776         tree = &BTRFS_I(mapping->host)->io_tree;
8777         return extent_readpages(tree, mapping, pages, nr_pages,
8778                                 btrfs_get_extent);
8779 }
8780 static int __btrfs_releasepage(struct page *page, gfp_t gfp_flags)
8781 {
8782         struct extent_io_tree *tree;
8783         struct extent_map_tree *map;
8784         int ret;
8785
8786         tree = &BTRFS_I(page->mapping->host)->io_tree;
8787         map = &BTRFS_I(page->mapping->host)->extent_tree;
8788         ret = try_release_extent_mapping(map, tree, page, gfp_flags);
8789         if (ret == 1) {
8790                 ClearPagePrivate(page);
8791                 set_page_private(page, 0);
8792                 put_page(page);
8793         }
8794         return ret;
8795 }
8796
8797 static int btrfs_releasepage(struct page *page, gfp_t gfp_flags)
8798 {
8799         if (PageWriteback(page) || PageDirty(page))
8800                 return 0;
8801         return __btrfs_releasepage(page, gfp_flags & GFP_NOFS);
8802 }
8803
8804 static void btrfs_invalidatepage(struct page *page, unsigned int offset,
8805                                  unsigned int length)
8806 {
8807         struct inode *inode = page->mapping->host;
8808         struct extent_io_tree *tree;
8809         struct btrfs_ordered_extent *ordered;
8810         struct extent_state *cached_state = NULL;
8811         u64 page_start = page_offset(page);
8812         u64 page_end = page_start + PAGE_SIZE - 1;
8813         u64 start;
8814         u64 end;
8815         int inode_evicting = inode->i_state & I_FREEING;
8816
8817         /*
8818          * we have the page locked, so new writeback can't start,
8819          * and the dirty bit won't be cleared while we are here.
8820          *
8821          * Wait for IO on this page so that we can safely clear
8822          * the PagePrivate2 bit and do ordered accounting
8823          */
8824         wait_on_page_writeback(page);
8825
8826         tree = &BTRFS_I(inode)->io_tree;
8827         if (offset) {
8828                 btrfs_releasepage(page, GFP_NOFS);
8829                 return;
8830         }
8831
8832         if (!inode_evicting)
8833                 lock_extent_bits(tree, page_start, page_end, &cached_state);
8834 again:
8835         start = page_start;
8836         ordered = btrfs_lookup_ordered_range(inode, start,
8837                                         page_end - start + 1);
8838         if (ordered) {
8839                 end = min(page_end, ordered->file_offset + ordered->len - 1);
8840                 /*
8841                  * IO on this page will never be started, so we need
8842                  * to account for any ordered extents now
8843                  */
8844                 if (!inode_evicting)
8845                         clear_extent_bit(tree, start, end,
8846                                          EXTENT_DIRTY | EXTENT_DELALLOC |
8847                                          EXTENT_LOCKED | EXTENT_DO_ACCOUNTING |
8848                                          EXTENT_DEFRAG, 1, 0, &cached_state,
8849                                          GFP_NOFS);
8850                 /*
8851                  * whoever cleared the private bit is responsible
8852                  * for the finish_ordered_io
8853                  */
8854                 if (TestClearPagePrivate2(page)) {
8855                         struct btrfs_ordered_inode_tree *tree;
8856                         u64 new_len;
8857
8858                         tree = &BTRFS_I(inode)->ordered_tree;
8859
8860                         spin_lock_irq(&tree->lock);
8861                         set_bit(BTRFS_ORDERED_TRUNCATED, &ordered->flags);
8862                         new_len = start - ordered->file_offset;
8863                         if (new_len < ordered->truncated_len)
8864                                 ordered->truncated_len = new_len;
8865                         spin_unlock_irq(&tree->lock);
8866
8867                         if (btrfs_dec_test_ordered_pending(inode, &ordered,
8868                                                            start,
8869                                                            end - start + 1, 1))
8870                                 btrfs_finish_ordered_io(ordered);
8871                 }
8872                 btrfs_put_ordered_extent(ordered);
8873                 if (!inode_evicting) {
8874                         cached_state = NULL;
8875                         lock_extent_bits(tree, start, end,
8876                                          &cached_state);
8877                 }
8878
8879                 start = end + 1;
8880                 if (start < page_end)
8881                         goto again;
8882         }
8883
8884         /*
8885          * Qgroup reserved space handler
8886          * Page here will be either
8887          * 1) Already written to disk
8888          *    In this case, its reserved space is released from data rsv map
8889          *    and will be freed by delayed_ref handler finally.
8890          *    So even we call qgroup_free_data(), it won't decrease reserved
8891          *    space.
8892          * 2) Not written to disk
8893          *    This means the reserved space should be freed here.
8894          */
8895         btrfs_qgroup_free_data(inode, page_start, PAGE_SIZE);
8896         if (!inode_evicting) {
8897                 clear_extent_bit(tree, page_start, page_end,
8898                                  EXTENT_LOCKED | EXTENT_DIRTY |
8899                                  EXTENT_DELALLOC | EXTENT_DO_ACCOUNTING |
8900                                  EXTENT_DEFRAG, 1, 1,
8901                                  &cached_state, GFP_NOFS);
8902
8903                 __btrfs_releasepage(page, GFP_NOFS);
8904         }
8905
8906         ClearPageChecked(page);
8907         if (PagePrivate(page)) {
8908                 ClearPagePrivate(page);
8909                 set_page_private(page, 0);
8910                 put_page(page);
8911         }
8912 }
8913
8914 /*
8915  * btrfs_page_mkwrite() is not allowed to change the file size as it gets
8916  * called from a page fault handler when a page is first dirtied. Hence we must
8917  * be careful to check for EOF conditions here. We set the page up correctly
8918  * for a written page which means we get ENOSPC checking when writing into
8919  * holes and correct delalloc and unwritten extent mapping on filesystems that
8920  * support these features.
8921  *
8922  * We are not allowed to take the i_mutex here so we have to play games to
8923  * protect against truncate races as the page could now be beyond EOF.  Because
8924  * vmtruncate() writes the inode size before removing pages, once we have the
8925  * page lock we can determine safely if the page is beyond EOF. If it is not
8926  * beyond EOF, then the page is guaranteed safe against truncation until we
8927  * unlock the page.
8928  */
8929 int btrfs_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
8930 {
8931         struct page *page = vmf->page;
8932         struct inode *inode = file_inode(vma->vm_file);
8933         struct btrfs_root *root = BTRFS_I(inode)->root;
8934         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
8935         struct btrfs_ordered_extent *ordered;
8936         struct extent_state *cached_state = NULL;
8937         char *kaddr;
8938         unsigned long zero_start;
8939         loff_t size;
8940         int ret;
8941         int reserved = 0;
8942         u64 reserved_space;
8943         u64 page_start;
8944         u64 page_end;
8945         u64 end;
8946
8947         reserved_space = PAGE_SIZE;
8948
8949         sb_start_pagefault(inode->i_sb);
8950         page_start = page_offset(page);
8951         page_end = page_start + PAGE_SIZE - 1;
8952         end = page_end;
8953
8954         /*
8955          * Reserving delalloc space after obtaining the page lock can lead to
8956          * deadlock. For example, if a dirty page is locked by this function
8957          * and the call to btrfs_delalloc_reserve_space() ends up triggering
8958          * dirty page write out, then the btrfs_writepage() function could
8959          * end up waiting indefinitely to get a lock on the page currently
8960          * being processed by btrfs_page_mkwrite() function.
8961          */
8962         ret = btrfs_delalloc_reserve_space(inode, page_start,
8963                                            reserved_space);
8964         if (!ret) {
8965                 ret = file_update_time(vma->vm_file);
8966                 reserved = 1;
8967         }
8968         if (ret) {
8969                 if (ret == -ENOMEM)
8970                         ret = VM_FAULT_OOM;
8971                 else /* -ENOSPC, -EIO, etc */
8972                         ret = VM_FAULT_SIGBUS;
8973                 if (reserved)
8974                         goto out;
8975                 goto out_noreserve;
8976         }
8977
8978         ret = VM_FAULT_NOPAGE; /* make the VM retry the fault */
8979 again:
8980         lock_page(page);
8981         size = i_size_read(inode);
8982
8983         if ((page->mapping != inode->i_mapping) ||
8984             (page_start >= size)) {
8985                 /* page got truncated out from underneath us */
8986                 goto out_unlock;
8987         }
8988         wait_on_page_writeback(page);
8989
8990         lock_extent_bits(io_tree, page_start, page_end, &cached_state);
8991         set_page_extent_mapped(page);
8992
8993         /*
8994          * we can't set the delalloc bits if there are pending ordered
8995          * extents.  Drop our locks and wait for them to finish
8996          */
8997         ordered = btrfs_lookup_ordered_range(inode, page_start, page_end);
8998         if (ordered) {
8999                 unlock_extent_cached(io_tree, page_start, page_end,
9000                                      &cached_state, GFP_NOFS);
9001                 unlock_page(page);
9002                 btrfs_start_ordered_extent(inode, ordered, 1);
9003                 btrfs_put_ordered_extent(ordered);
9004                 goto again;
9005         }
9006
9007         if (page->index == ((size - 1) >> PAGE_SHIFT)) {
9008                 reserved_space = round_up(size - page_start, root->sectorsize);
9009                 if (reserved_space < PAGE_SIZE) {
9010                         end = page_start + reserved_space - 1;
9011                         spin_lock(&BTRFS_I(inode)->lock);
9012                         BTRFS_I(inode)->outstanding_extents++;
9013                         spin_unlock(&BTRFS_I(inode)->lock);
9014                         btrfs_delalloc_release_space(inode, page_start,
9015                                                 PAGE_SIZE - reserved_space);
9016                 }
9017         }
9018
9019         /*
9020          * XXX - page_mkwrite gets called every time the page is dirtied, even
9021          * if it was already dirty, so for space accounting reasons we need to
9022          * clear any delalloc bits for the range we are fixing to save.  There
9023          * is probably a better way to do this, but for now keep consistent with
9024          * prepare_pages in the normal write path.
9025          */
9026         clear_extent_bit(&BTRFS_I(inode)->io_tree, page_start, end,
9027                           EXTENT_DIRTY | EXTENT_DELALLOC |
9028                           EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG,
9029                           0, 0, &cached_state, GFP_NOFS);
9030
9031         ret = btrfs_set_extent_delalloc(inode, page_start, end,
9032                                         &cached_state);
9033         if (ret) {
9034                 unlock_extent_cached(io_tree, page_start, page_end,
9035                                      &cached_state, GFP_NOFS);
9036                 ret = VM_FAULT_SIGBUS;
9037                 goto out_unlock;
9038         }
9039         ret = 0;
9040
9041         /* page is wholly or partially inside EOF */
9042         if (page_start + PAGE_SIZE > size)
9043                 zero_start = size & ~PAGE_MASK;
9044         else
9045                 zero_start = PAGE_SIZE;
9046
9047         if (zero_start != PAGE_SIZE) {
9048                 kaddr = kmap(page);
9049                 memset(kaddr + zero_start, 0, PAGE_SIZE - zero_start);
9050                 flush_dcache_page(page);
9051                 kunmap(page);
9052         }
9053         ClearPageChecked(page);
9054         set_page_dirty(page);
9055         SetPageUptodate(page);
9056
9057         BTRFS_I(inode)->last_trans = root->fs_info->generation;
9058         BTRFS_I(inode)->last_sub_trans = BTRFS_I(inode)->root->log_transid;
9059         BTRFS_I(inode)->last_log_commit = BTRFS_I(inode)->root->last_log_commit;
9060
9061         unlock_extent_cached(io_tree, page_start, page_end, &cached_state, GFP_NOFS);
9062
9063 out_unlock:
9064         if (!ret) {
9065                 sb_end_pagefault(inode->i_sb);
9066                 return VM_FAULT_LOCKED;
9067         }
9068         unlock_page(page);
9069 out:
9070         btrfs_delalloc_release_space(inode, page_start, reserved_space);
9071 out_noreserve:
9072         sb_end_pagefault(inode->i_sb);
9073         return ret;
9074 }
9075
9076 static int btrfs_truncate(struct inode *inode)
9077 {
9078         struct btrfs_root *root = BTRFS_I(inode)->root;
9079         struct btrfs_block_rsv *rsv;
9080         int ret = 0;
9081         int err = 0;
9082         struct btrfs_trans_handle *trans;
9083         u64 mask = root->sectorsize - 1;
9084         u64 min_size = btrfs_calc_trunc_metadata_size(root, 1);
9085
9086         ret = btrfs_wait_ordered_range(inode, inode->i_size & (~mask),
9087                                        (u64)-1);
9088         if (ret)
9089                 return ret;
9090
9091         /*
9092          * Yes ladies and gentlemen, this is indeed ugly.  The fact is we have
9093          * 3 things going on here
9094          *
9095          * 1) We need to reserve space for our orphan item and the space to
9096          * delete our orphan item.  Lord knows we don't want to have a dangling
9097          * orphan item because we didn't reserve space to remove it.
9098          *
9099          * 2) We need to reserve space to update our inode.
9100          *
9101          * 3) We need to have something to cache all the space that is going to
9102          * be free'd up by the truncate operation, but also have some slack
9103          * space reserved in case it uses space during the truncate (thank you
9104          * very much snapshotting).
9105          *
9106          * And we need these to all be separate.  The fact is we can use a lot of
9107          * space doing the truncate, and we have no earthly idea how much space
9108          * we will use, so we need the truncate reservation to be separate so it
9109          * doesn't end up using space reserved for updating the inode or
9110          * removing the orphan item.  We also need to be able to stop the
9111          * transaction and start a new one, which means we need to be able to
9112          * update the inode several times, and we have no idea of knowing how
9113          * many times that will be, so we can't just reserve 1 item for the
9114          * entirety of the operation, so that has to be done separately as well.
9115          * Then there is the orphan item, which does indeed need to be held on
9116          * to for the whole operation, and we need nobody to touch this reserved
9117          * space except the orphan code.
9118          *
9119          * So that leaves us with
9120          *
9121          * 1) root->orphan_block_rsv - for the orphan deletion.
9122          * 2) rsv - for the truncate reservation, which we will steal from the
9123          * transaction reservation.
9124          * 3) fs_info->trans_block_rsv - this will have 1 items worth left for
9125          * updating the inode.
9126          */
9127         rsv = btrfs_alloc_block_rsv(root, BTRFS_BLOCK_RSV_TEMP);
9128         if (!rsv)
9129                 return -ENOMEM;
9130         rsv->size = min_size;
9131         rsv->failfast = 1;
9132
9133         /*
9134          * 1 for the truncate slack space
9135          * 1 for updating the inode.
9136          */
9137         trans = btrfs_start_transaction(root, 2);
9138         if (IS_ERR(trans)) {
9139                 err = PTR_ERR(trans);
9140                 goto out;
9141         }
9142
9143         /* Migrate the slack space for the truncate to our reserve */
9144         ret = btrfs_block_rsv_migrate(&root->fs_info->trans_block_rsv, rsv,
9145                                       min_size, 0);
9146         BUG_ON(ret);
9147
9148         /*
9149          * So if we truncate and then write and fsync we normally would just
9150          * write the extents that changed, which is a problem if we need to
9151          * first truncate that entire inode.  So set this flag so we write out
9152          * all of the extents in the inode to the sync log so we're completely
9153          * safe.
9154          */
9155         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &BTRFS_I(inode)->runtime_flags);
9156         trans->block_rsv = rsv;
9157
9158         while (1) {
9159                 ret = btrfs_truncate_inode_items(trans, root, inode,
9160                                                  inode->i_size,
9161                                                  BTRFS_EXTENT_DATA_KEY);
9162                 if (ret != -ENOSPC && ret != -EAGAIN) {
9163                         err = ret;
9164                         break;
9165                 }
9166
9167                 trans->block_rsv = &root->fs_info->trans_block_rsv;
9168                 ret = btrfs_update_inode(trans, root, inode);
9169                 if (ret) {
9170                         err = ret;
9171                         break;
9172                 }
9173
9174                 btrfs_end_transaction(trans, root);
9175                 btrfs_btree_balance_dirty(root);
9176
9177                 trans = btrfs_start_transaction(root, 2);
9178                 if (IS_ERR(trans)) {
9179                         ret = err = PTR_ERR(trans);
9180                         trans = NULL;
9181                         break;
9182                 }
9183
9184                 ret = btrfs_block_rsv_migrate(&root->fs_info->trans_block_rsv,
9185                                               rsv, min_size, 0);
9186                 BUG_ON(ret);    /* shouldn't happen */
9187                 trans->block_rsv = rsv;
9188         }
9189
9190         if (ret == 0 && inode->i_nlink > 0) {
9191                 trans->block_rsv = root->orphan_block_rsv;
9192                 ret = btrfs_orphan_del(trans, inode);
9193                 if (ret)
9194                         err = ret;
9195         }
9196
9197         if (trans) {
9198                 trans->block_rsv = &root->fs_info->trans_block_rsv;
9199                 ret = btrfs_update_inode(trans, root, inode);
9200                 if (ret && !err)
9201                         err = ret;
9202
9203                 ret = btrfs_end_transaction(trans, root);
9204                 btrfs_btree_balance_dirty(root);
9205         }
9206 out:
9207         btrfs_free_block_rsv(root, rsv);
9208
9209         if (ret && !err)
9210                 err = ret;
9211
9212         return err;
9213 }
9214
9215 /*
9216  * create a new subvolume directory/inode (helper for the ioctl).
9217  */
9218 int btrfs_create_subvol_root(struct btrfs_trans_handle *trans,
9219                              struct btrfs_root *new_root,
9220                              struct btrfs_root *parent_root,
9221                              u64 new_dirid)
9222 {
9223         struct inode *inode;
9224         int err;
9225         u64 index = 0;
9226
9227         inode = btrfs_new_inode(trans, new_root, NULL, "..", 2,
9228                                 new_dirid, new_dirid,
9229                                 S_IFDIR | (~current_umask() & S_IRWXUGO),
9230                                 &index);
9231         if (IS_ERR(inode))
9232                 return PTR_ERR(inode);
9233         inode->i_op = &btrfs_dir_inode_operations;
9234         inode->i_fop = &btrfs_dir_file_operations;
9235
9236         set_nlink(inode, 1);
9237         btrfs_i_size_write(inode, 0);
9238         unlock_new_inode(inode);
9239
9240         err = btrfs_subvol_inherit_props(trans, new_root, parent_root);
9241         if (err)
9242                 btrfs_err(new_root->fs_info,
9243                           "error inheriting subvolume %llu properties: %d",
9244                           new_root->root_key.objectid, err);
9245
9246         err = btrfs_update_inode(trans, new_root, inode);
9247
9248         iput(inode);
9249         return err;
9250 }
9251
9252 struct inode *btrfs_alloc_inode(struct super_block *sb)
9253 {
9254         struct btrfs_inode *ei;
9255         struct inode *inode;
9256
9257         ei = kmem_cache_alloc(btrfs_inode_cachep, GFP_NOFS);
9258         if (!ei)
9259                 return NULL;
9260
9261         ei->root = NULL;
9262         ei->generation = 0;
9263         ei->last_trans = 0;
9264         ei->last_sub_trans = 0;
9265         ei->logged_trans = 0;
9266         ei->delalloc_bytes = 0;
9267         ei->defrag_bytes = 0;
9268         ei->disk_i_size = 0;
9269         ei->flags = 0;
9270         ei->csum_bytes = 0;
9271         ei->index_cnt = (u64)-1;
9272         ei->dir_index = 0;
9273         ei->last_unlink_trans = 0;
9274         ei->last_log_commit = 0;
9275         ei->delayed_iput_count = 0;
9276
9277         spin_lock_init(&ei->lock);
9278         ei->outstanding_extents = 0;
9279         ei->reserved_extents = 0;
9280
9281         ei->runtime_flags = 0;
9282         ei->force_compress = BTRFS_COMPRESS_NONE;
9283
9284         ei->delayed_node = NULL;
9285
9286         ei->i_otime.tv_sec = 0;
9287         ei->i_otime.tv_nsec = 0;
9288
9289         inode = &ei->vfs_inode;
9290         extent_map_tree_init(&ei->extent_tree);
9291         extent_io_tree_init(&ei->io_tree, &inode->i_data);
9292         extent_io_tree_init(&ei->io_failure_tree, &inode->i_data);
9293         ei->io_tree.track_uptodate = 1;
9294         ei->io_failure_tree.track_uptodate = 1;
9295         atomic_set(&ei->sync_writers, 0);
9296         mutex_init(&ei->log_mutex);
9297         mutex_init(&ei->delalloc_mutex);
9298         btrfs_ordered_inode_tree_init(&ei->ordered_tree);
9299         INIT_LIST_HEAD(&ei->delalloc_inodes);
9300         INIT_LIST_HEAD(&ei->delayed_iput);
9301         RB_CLEAR_NODE(&ei->rb_node);
9302         init_rwsem(&ei->dio_sem);
9303
9304         return inode;
9305 }
9306
9307 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
9308 void btrfs_test_destroy_inode(struct inode *inode)
9309 {
9310         btrfs_drop_extent_cache(inode, 0, (u64)-1, 0);
9311         kmem_cache_free(btrfs_inode_cachep, BTRFS_I(inode));
9312 }
9313 #endif
9314
9315 static void btrfs_i_callback(struct rcu_head *head)
9316 {
9317         struct inode *inode = container_of(head, struct inode, i_rcu);
9318         kmem_cache_free(btrfs_inode_cachep, BTRFS_I(inode));
9319 }
9320
9321 void btrfs_destroy_inode(struct inode *inode)
9322 {
9323         struct btrfs_ordered_extent *ordered;
9324         struct btrfs_root *root = BTRFS_I(inode)->root;
9325
9326         WARN_ON(!hlist_empty(&inode->i_dentry));
9327         WARN_ON(inode->i_data.nrpages);
9328         WARN_ON(BTRFS_I(inode)->outstanding_extents);
9329         WARN_ON(BTRFS_I(inode)->reserved_extents);
9330         WARN_ON(BTRFS_I(inode)->delalloc_bytes);
9331         WARN_ON(BTRFS_I(inode)->csum_bytes);
9332         WARN_ON(BTRFS_I(inode)->defrag_bytes);
9333
9334         /*
9335          * This can happen where we create an inode, but somebody else also
9336          * created the same inode and we need to destroy the one we already
9337          * created.
9338          */
9339         if (!root)
9340                 goto free;
9341
9342         if (test_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
9343                      &BTRFS_I(inode)->runtime_flags)) {
9344                 btrfs_info(root->fs_info, "inode %llu still on the orphan list",
9345                         btrfs_ino(inode));
9346                 atomic_dec(&root->orphan_inodes);
9347         }
9348
9349         while (1) {
9350                 ordered = btrfs_lookup_first_ordered_extent(inode, (u64)-1);
9351                 if (!ordered)
9352                         break;
9353                 else {
9354                         btrfs_err(root->fs_info, "found ordered extent %llu %llu on inode cleanup",
9355                                 ordered->file_offset, ordered->len);
9356                         btrfs_remove_ordered_extent(inode, ordered);
9357                         btrfs_put_ordered_extent(ordered);
9358                         btrfs_put_ordered_extent(ordered);
9359                 }
9360         }
9361         btrfs_qgroup_check_reserved_leak(inode);
9362         inode_tree_del(inode);
9363         btrfs_drop_extent_cache(inode, 0, (u64)-1, 0);
9364 free:
9365         call_rcu(&inode->i_rcu, btrfs_i_callback);
9366 }
9367
9368 int btrfs_drop_inode(struct inode *inode)
9369 {
9370         struct btrfs_root *root = BTRFS_I(inode)->root;
9371
9372         if (root == NULL)
9373                 return 1;
9374
9375         /* the snap/subvol tree is on deleting */
9376         if (btrfs_root_refs(&root->root_item) == 0)
9377                 return 1;
9378         else
9379                 return generic_drop_inode(inode);
9380 }
9381
9382 static void init_once(void *foo)
9383 {
9384         struct btrfs_inode *ei = (struct btrfs_inode *) foo;
9385
9386         inode_init_once(&ei->vfs_inode);
9387 }
9388
9389 void btrfs_destroy_cachep(void)
9390 {
9391         /*
9392          * Make sure all delayed rcu free inodes are flushed before we
9393          * destroy cache.
9394          */
9395         rcu_barrier();
9396         kmem_cache_destroy(btrfs_inode_cachep);
9397         kmem_cache_destroy(btrfs_trans_handle_cachep);
9398         kmem_cache_destroy(btrfs_transaction_cachep);
9399         kmem_cache_destroy(btrfs_path_cachep);
9400         kmem_cache_destroy(btrfs_free_space_cachep);
9401 }
9402
9403 int btrfs_init_cachep(void)
9404 {
9405         btrfs_inode_cachep = kmem_cache_create("btrfs_inode",
9406                         sizeof(struct btrfs_inode), 0,
9407                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD | SLAB_ACCOUNT,
9408                         init_once);
9409         if (!btrfs_inode_cachep)
9410                 goto fail;
9411
9412         btrfs_trans_handle_cachep = kmem_cache_create("btrfs_trans_handle",
9413                         sizeof(struct btrfs_trans_handle), 0,
9414                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, NULL);
9415         if (!btrfs_trans_handle_cachep)
9416                 goto fail;
9417
9418         btrfs_transaction_cachep = kmem_cache_create("btrfs_transaction",
9419                         sizeof(struct btrfs_transaction), 0,
9420                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, NULL);
9421         if (!btrfs_transaction_cachep)
9422                 goto fail;
9423
9424         btrfs_path_cachep = kmem_cache_create("btrfs_path",
9425                         sizeof(struct btrfs_path), 0,
9426                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, NULL);
9427         if (!btrfs_path_cachep)
9428                 goto fail;
9429
9430         btrfs_free_space_cachep = kmem_cache_create("btrfs_free_space",
9431                         sizeof(struct btrfs_free_space), 0,
9432                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, NULL);
9433         if (!btrfs_free_space_cachep)
9434                 goto fail;
9435
9436         return 0;
9437 fail:
9438         btrfs_destroy_cachep();
9439         return -ENOMEM;
9440 }
9441
9442 static int btrfs_getattr(struct vfsmount *mnt,
9443                          struct dentry *dentry, struct kstat *stat)
9444 {
9445         u64 delalloc_bytes;
9446         struct inode *inode = d_inode(dentry);
9447         u32 blocksize = inode->i_sb->s_blocksize;
9448
9449         generic_fillattr(inode, stat);
9450         stat->dev = BTRFS_I(inode)->root->anon_dev;
9451
9452         spin_lock(&BTRFS_I(inode)->lock);
9453         delalloc_bytes = BTRFS_I(inode)->delalloc_bytes;
9454         spin_unlock(&BTRFS_I(inode)->lock);
9455         stat->blocks = (ALIGN(inode_get_bytes(inode), blocksize) +
9456                         ALIGN(delalloc_bytes, blocksize)) >> 9;
9457         return 0;
9458 }
9459
9460 static int btrfs_rename_exchange(struct inode *old_dir,
9461                               struct dentry *old_dentry,
9462                               struct inode *new_dir,
9463                               struct dentry *new_dentry)
9464 {
9465         struct btrfs_trans_handle *trans;
9466         struct btrfs_root *root = BTRFS_I(old_dir)->root;
9467         struct btrfs_root *dest = BTRFS_I(new_dir)->root;
9468         struct inode *new_inode = new_dentry->d_inode;
9469         struct inode *old_inode = old_dentry->d_inode;
9470         struct timespec ctime = CURRENT_TIME;
9471         struct dentry *parent;
9472         u64 old_ino = btrfs_ino(old_inode);
9473         u64 new_ino = btrfs_ino(new_inode);
9474         u64 old_idx = 0;
9475         u64 new_idx = 0;
9476         u64 root_objectid;
9477         int ret;
9478         bool root_log_pinned = false;
9479         bool dest_log_pinned = false;
9480
9481         /* we only allow rename subvolume link between subvolumes */
9482         if (old_ino != BTRFS_FIRST_FREE_OBJECTID && root != dest)
9483                 return -EXDEV;
9484
9485         /* close the race window with snapshot create/destroy ioctl */
9486         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
9487                 down_read(&root->fs_info->subvol_sem);
9488         if (new_ino == BTRFS_FIRST_FREE_OBJECTID)
9489                 down_read(&dest->fs_info->subvol_sem);
9490
9491         /*
9492          * We want to reserve the absolute worst case amount of items.  So if
9493          * both inodes are subvols and we need to unlink them then that would
9494          * require 4 item modifications, but if they are both normal inodes it
9495          * would require 5 item modifications, so we'll assume their normal
9496          * inodes.  So 5 * 2 is 10, plus 2 for the new links, so 12 total items
9497          * should cover the worst case number of items we'll modify.
9498          */
9499         trans = btrfs_start_transaction(root, 12);
9500         if (IS_ERR(trans)) {
9501                 ret = PTR_ERR(trans);
9502                 goto out_notrans;
9503         }
9504
9505         /*
9506          * We need to find a free sequence number both in the source and
9507          * in the destination directory for the exchange.
9508          */
9509         ret = btrfs_set_inode_index(new_dir, &old_idx);
9510         if (ret)
9511                 goto out_fail;
9512         ret = btrfs_set_inode_index(old_dir, &new_idx);
9513         if (ret)
9514                 goto out_fail;
9515
9516         BTRFS_I(old_inode)->dir_index = 0ULL;
9517         BTRFS_I(new_inode)->dir_index = 0ULL;
9518
9519         /* Reference for the source. */
9520         if (old_ino == BTRFS_FIRST_FREE_OBJECTID) {
9521                 /* force full log commit if subvolume involved. */
9522                 btrfs_set_log_full_commit(root->fs_info, trans);
9523         } else {
9524                 btrfs_pin_log_trans(root);
9525                 root_log_pinned = true;
9526                 ret = btrfs_insert_inode_ref(trans, dest,
9527                                              new_dentry->d_name.name,
9528                                              new_dentry->d_name.len,
9529                                              old_ino,
9530                                              btrfs_ino(new_dir), old_idx);
9531                 if (ret)
9532                         goto out_fail;
9533         }
9534
9535         /* And now for the dest. */
9536         if (new_ino == BTRFS_FIRST_FREE_OBJECTID) {
9537                 /* force full log commit if subvolume involved. */
9538                 btrfs_set_log_full_commit(dest->fs_info, trans);
9539         } else {
9540                 btrfs_pin_log_trans(dest);
9541                 dest_log_pinned = true;
9542                 ret = btrfs_insert_inode_ref(trans, root,
9543                                              old_dentry->d_name.name,
9544                                              old_dentry->d_name.len,
9545                                              new_ino,
9546                                              btrfs_ino(old_dir), new_idx);
9547                 if (ret)
9548                         goto out_fail;
9549         }
9550
9551         /* Update inode version and ctime/mtime. */
9552         inode_inc_iversion(old_dir);
9553         inode_inc_iversion(new_dir);
9554         inode_inc_iversion(old_inode);
9555         inode_inc_iversion(new_inode);
9556         old_dir->i_ctime = old_dir->i_mtime = ctime;
9557         new_dir->i_ctime = new_dir->i_mtime = ctime;
9558         old_inode->i_ctime = ctime;
9559         new_inode->i_ctime = ctime;
9560
9561         if (old_dentry->d_parent != new_dentry->d_parent) {
9562                 btrfs_record_unlink_dir(trans, old_dir, old_inode, 1);
9563                 btrfs_record_unlink_dir(trans, new_dir, new_inode, 1);
9564         }
9565
9566         /* src is a subvolume */
9567         if (old_ino == BTRFS_FIRST_FREE_OBJECTID) {
9568                 root_objectid = BTRFS_I(old_inode)->root->root_key.objectid;
9569                 ret = btrfs_unlink_subvol(trans, root, old_dir,
9570                                           root_objectid,
9571                                           old_dentry->d_name.name,
9572                                           old_dentry->d_name.len);
9573         } else { /* src is an inode */
9574                 ret = __btrfs_unlink_inode(trans, root, old_dir,
9575                                            old_dentry->d_inode,
9576                                            old_dentry->d_name.name,
9577                                            old_dentry->d_name.len);
9578                 if (!ret)
9579                         ret = btrfs_update_inode(trans, root, old_inode);
9580         }
9581         if (ret) {
9582                 btrfs_abort_transaction(trans, root, ret);
9583                 goto out_fail;
9584         }
9585
9586         /* dest is a subvolume */
9587         if (new_ino == BTRFS_FIRST_FREE_OBJECTID) {
9588                 root_objectid = BTRFS_I(new_inode)->root->root_key.objectid;
9589                 ret = btrfs_unlink_subvol(trans, dest, new_dir,
9590                                           root_objectid,
9591                                           new_dentry->d_name.name,
9592                                           new_dentry->d_name.len);
9593         } else { /* dest is an inode */
9594                 ret = __btrfs_unlink_inode(trans, dest, new_dir,
9595                                            new_dentry->d_inode,
9596                                            new_dentry->d_name.name,
9597                                            new_dentry->d_name.len);
9598                 if (!ret)
9599                         ret = btrfs_update_inode(trans, dest, new_inode);
9600         }
9601         if (ret) {
9602                 btrfs_abort_transaction(trans, root, ret);
9603                 goto out_fail;
9604         }
9605
9606         ret = btrfs_add_link(trans, new_dir, old_inode,
9607                              new_dentry->d_name.name,
9608                              new_dentry->d_name.len, 0, old_idx);
9609         if (ret) {
9610                 btrfs_abort_transaction(trans, root, ret);
9611                 goto out_fail;
9612         }
9613
9614         ret = btrfs_add_link(trans, old_dir, new_inode,
9615                              old_dentry->d_name.name,
9616                              old_dentry->d_name.len, 0, new_idx);
9617         if (ret) {
9618                 btrfs_abort_transaction(trans, root, ret);
9619                 goto out_fail;
9620         }
9621
9622         if (old_inode->i_nlink == 1)
9623                 BTRFS_I(old_inode)->dir_index = old_idx;
9624         if (new_inode->i_nlink == 1)
9625                 BTRFS_I(new_inode)->dir_index = new_idx;
9626
9627         if (root_log_pinned) {
9628                 parent = new_dentry->d_parent;
9629                 btrfs_log_new_name(trans, old_inode, old_dir, parent);
9630                 btrfs_end_log_trans(root);
9631                 root_log_pinned = false;
9632         }
9633         if (dest_log_pinned) {
9634                 parent = old_dentry->d_parent;
9635                 btrfs_log_new_name(trans, new_inode, new_dir, parent);
9636                 btrfs_end_log_trans(dest);
9637                 dest_log_pinned = false;
9638         }
9639 out_fail:
9640         /*
9641          * If we have pinned a log and an error happened, we unpin tasks
9642          * trying to sync the log and force them to fallback to a transaction
9643          * commit if the log currently contains any of the inodes involved in
9644          * this rename operation (to ensure we do not persist a log with an
9645          * inconsistent state for any of these inodes or leading to any
9646          * inconsistencies when replayed). If the transaction was aborted, the
9647          * abortion reason is propagated to userspace when attempting to commit
9648          * the transaction. If the log does not contain any of these inodes, we
9649          * allow the tasks to sync it.
9650          */
9651         if (ret && (root_log_pinned || dest_log_pinned)) {
9652                 if (btrfs_inode_in_log(old_dir, root->fs_info->generation) ||
9653                     btrfs_inode_in_log(new_dir, root->fs_info->generation) ||
9654                     btrfs_inode_in_log(old_inode, root->fs_info->generation) ||
9655                     (new_inode &&
9656                      btrfs_inode_in_log(new_inode, root->fs_info->generation)))
9657                     btrfs_set_log_full_commit(root->fs_info, trans);
9658
9659                 if (root_log_pinned) {
9660                         btrfs_end_log_trans(root);
9661                         root_log_pinned = false;
9662                 }
9663                 if (dest_log_pinned) {
9664                         btrfs_end_log_trans(dest);
9665                         dest_log_pinned = false;
9666                 }
9667         }
9668         ret = btrfs_end_transaction(trans, root);
9669 out_notrans:
9670         if (new_ino == BTRFS_FIRST_FREE_OBJECTID)
9671                 up_read(&dest->fs_info->subvol_sem);
9672         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
9673                 up_read(&root->fs_info->subvol_sem);
9674
9675         return ret;
9676 }
9677
9678 static int btrfs_whiteout_for_rename(struct btrfs_trans_handle *trans,
9679                                      struct btrfs_root *root,
9680                                      struct inode *dir,
9681                                      struct dentry *dentry)
9682 {
9683         int ret;
9684         struct inode *inode;
9685         u64 objectid;
9686         u64 index;
9687
9688         ret = btrfs_find_free_ino(root, &objectid);
9689         if (ret)
9690                 return ret;
9691
9692         inode = btrfs_new_inode(trans, root, dir,
9693                                 dentry->d_name.name,
9694                                 dentry->d_name.len,
9695                                 btrfs_ino(dir),
9696                                 objectid,
9697                                 S_IFCHR | WHITEOUT_MODE,
9698                                 &index);
9699
9700         if (IS_ERR(inode)) {
9701                 ret = PTR_ERR(inode);
9702                 return ret;
9703         }
9704
9705         inode->i_op = &btrfs_special_inode_operations;
9706         init_special_inode(inode, inode->i_mode,
9707                 WHITEOUT_DEV);
9708
9709         ret = btrfs_init_inode_security(trans, inode, dir,
9710                                 &dentry->d_name);
9711         if (ret)
9712                 goto out;
9713
9714         ret = btrfs_add_nondir(trans, dir, dentry,
9715                                 inode, 0, index);
9716         if (ret)
9717                 goto out;
9718
9719         ret = btrfs_update_inode(trans, root, inode);
9720 out:
9721         unlock_new_inode(inode);
9722         if (ret)
9723                 inode_dec_link_count(inode);
9724         iput(inode);
9725
9726         return ret;
9727 }
9728
9729 static int btrfs_rename(struct inode *old_dir, struct dentry *old_dentry,
9730                            struct inode *new_dir, struct dentry *new_dentry,
9731                            unsigned int flags)
9732 {
9733         struct btrfs_trans_handle *trans;
9734         unsigned int trans_num_items;
9735         struct btrfs_root *root = BTRFS_I(old_dir)->root;
9736         struct btrfs_root *dest = BTRFS_I(new_dir)->root;
9737         struct inode *new_inode = d_inode(new_dentry);
9738         struct inode *old_inode = d_inode(old_dentry);
9739         u64 index = 0;
9740         u64 root_objectid;
9741         int ret;
9742         u64 old_ino = btrfs_ino(old_inode);
9743         bool log_pinned = false;
9744
9745         if (btrfs_ino(new_dir) == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)
9746                 return -EPERM;
9747
9748         /* we only allow rename subvolume link between subvolumes */
9749         if (old_ino != BTRFS_FIRST_FREE_OBJECTID && root != dest)
9750                 return -EXDEV;
9751
9752         if (old_ino == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID ||
9753             (new_inode && btrfs_ino(new_inode) == BTRFS_FIRST_FREE_OBJECTID))
9754                 return -ENOTEMPTY;
9755
9756         if (S_ISDIR(old_inode->i_mode) && new_inode &&
9757             new_inode->i_size > BTRFS_EMPTY_DIR_SIZE)
9758                 return -ENOTEMPTY;
9759
9760
9761         /* check for collisions, even if the  name isn't there */
9762         ret = btrfs_check_dir_item_collision(dest, new_dir->i_ino,
9763                              new_dentry->d_name.name,
9764                              new_dentry->d_name.len);
9765
9766         if (ret) {
9767                 if (ret == -EEXIST) {
9768                         /* we shouldn't get
9769                          * eexist without a new_inode */
9770                         if (WARN_ON(!new_inode)) {
9771                                 return ret;
9772                         }
9773                 } else {
9774                         /* maybe -EOVERFLOW */
9775                         return ret;
9776                 }
9777         }
9778         ret = 0;
9779
9780         /*
9781          * we're using rename to replace one file with another.  Start IO on it
9782          * now so  we don't add too much work to the end of the transaction
9783          */
9784         if (new_inode && S_ISREG(old_inode->i_mode) && new_inode->i_size)
9785                 filemap_flush(old_inode->i_mapping);
9786
9787         /* close the racy window with snapshot create/destroy ioctl */
9788         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
9789                 down_read(&root->fs_info->subvol_sem);
9790         /*
9791          * We want to reserve the absolute worst case amount of items.  So if
9792          * both inodes are subvols and we need to unlink them then that would
9793          * require 4 item modifications, but if they are both normal inodes it
9794          * would require 5 item modifications, so we'll assume they are normal
9795          * inodes.  So 5 * 2 is 10, plus 1 for the new link, so 11 total items
9796          * should cover the worst case number of items we'll modify.
9797          * If our rename has the whiteout flag, we need more 5 units for the
9798          * new inode (1 inode item, 1 inode ref, 2 dir items and 1 xattr item
9799          * when selinux is enabled).
9800          */
9801         trans_num_items = 11;
9802         if (flags & RENAME_WHITEOUT)
9803                 trans_num_items += 5;
9804         trans = btrfs_start_transaction(root, trans_num_items);
9805         if (IS_ERR(trans)) {
9806                 ret = PTR_ERR(trans);
9807                 goto out_notrans;
9808         }
9809
9810         if (dest != root)
9811                 btrfs_record_root_in_trans(trans, dest);
9812
9813         ret = btrfs_set_inode_index(new_dir, &index);
9814         if (ret)
9815                 goto out_fail;
9816
9817         BTRFS_I(old_inode)->dir_index = 0ULL;
9818         if (unlikely(old_ino == BTRFS_FIRST_FREE_OBJECTID)) {
9819                 /* force full log commit if subvolume involved. */
9820                 btrfs_set_log_full_commit(root->fs_info, trans);
9821         } else {
9822                 btrfs_pin_log_trans(root);
9823                 log_pinned = true;
9824                 ret = btrfs_insert_inode_ref(trans, dest,
9825                                              new_dentry->d_name.name,
9826                                              new_dentry->d_name.len,
9827                                              old_ino,
9828                                              btrfs_ino(new_dir), index);
9829                 if (ret)
9830                         goto out_fail;
9831         }
9832
9833         inode_inc_iversion(old_dir);
9834         inode_inc_iversion(new_dir);
9835         inode_inc_iversion(old_inode);
9836         old_dir->i_ctime = old_dir->i_mtime =
9837         new_dir->i_ctime = new_dir->i_mtime =
9838         old_inode->i_ctime = current_fs_time(old_dir->i_sb);
9839
9840         if (old_dentry->d_parent != new_dentry->d_parent)
9841                 btrfs_record_unlink_dir(trans, old_dir, old_inode, 1);
9842
9843         if (unlikely(old_ino == BTRFS_FIRST_FREE_OBJECTID)) {
9844                 root_objectid = BTRFS_I(old_inode)->root->root_key.objectid;
9845                 ret = btrfs_unlink_subvol(trans, root, old_dir, root_objectid,
9846                                         old_dentry->d_name.name,
9847                                         old_dentry->d_name.len);
9848         } else {
9849                 ret = __btrfs_unlink_inode(trans, root, old_dir,
9850                                         d_inode(old_dentry),
9851                                         old_dentry->d_name.name,
9852                                         old_dentry->d_name.len);
9853                 if (!ret)
9854                         ret = btrfs_update_inode(trans, root, old_inode);
9855         }
9856         if (ret) {
9857                 btrfs_abort_transaction(trans, root, ret);
9858                 goto out_fail;
9859         }
9860
9861         if (new_inode) {
9862                 inode_inc_iversion(new_inode);
9863                 new_inode->i_ctime = current_fs_time(new_inode->i_sb);
9864                 if (unlikely(btrfs_ino(new_inode) ==
9865                              BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)) {
9866                         root_objectid = BTRFS_I(new_inode)->location.objectid;
9867                         ret = btrfs_unlink_subvol(trans, dest, new_dir,
9868                                                 root_objectid,
9869                                                 new_dentry->d_name.name,
9870                                                 new_dentry->d_name.len);
9871                         BUG_ON(new_inode->i_nlink == 0);
9872                 } else {
9873                         ret = btrfs_unlink_inode(trans, dest, new_dir,
9874                                                  d_inode(new_dentry),
9875                                                  new_dentry->d_name.name,
9876                                                  new_dentry->d_name.len);
9877                 }
9878                 if (!ret && new_inode->i_nlink == 0)
9879                         ret = btrfs_orphan_add(trans, d_inode(new_dentry));
9880                 if (ret) {
9881                         btrfs_abort_transaction(trans, root, ret);
9882                         goto out_fail;
9883                 }
9884         }
9885
9886         ret = btrfs_add_link(trans, new_dir, old_inode,
9887                              new_dentry->d_name.name,
9888                              new_dentry->d_name.len, 0, index);
9889         if (ret) {
9890                 btrfs_abort_transaction(trans, root, ret);
9891                 goto out_fail;
9892         }
9893
9894         if (old_inode->i_nlink == 1)
9895                 BTRFS_I(old_inode)->dir_index = index;
9896
9897         if (log_pinned) {
9898                 struct dentry *parent = new_dentry->d_parent;
9899
9900                 btrfs_log_new_name(trans, old_inode, old_dir, parent);
9901                 btrfs_end_log_trans(root);
9902                 log_pinned = false;
9903         }
9904
9905         if (flags & RENAME_WHITEOUT) {
9906                 ret = btrfs_whiteout_for_rename(trans, root, old_dir,
9907                                                 old_dentry);
9908
9909                 if (ret) {
9910                         btrfs_abort_transaction(trans, root, ret);
9911                         goto out_fail;
9912                 }
9913         }
9914 out_fail:
9915         /*
9916          * If we have pinned the log and an error happened, we unpin tasks
9917          * trying to sync the log and force them to fallback to a transaction
9918          * commit if the log currently contains any of the inodes involved in
9919          * this rename operation (to ensure we do not persist a log with an
9920          * inconsistent state for any of these inodes or leading to any
9921          * inconsistencies when replayed). If the transaction was aborted, the
9922          * abortion reason is propagated to userspace when attempting to commit
9923          * the transaction. If the log does not contain any of these inodes, we
9924          * allow the tasks to sync it.
9925          */
9926         if (ret && log_pinned) {
9927                 if (btrfs_inode_in_log(old_dir, root->fs_info->generation) ||
9928                     btrfs_inode_in_log(new_dir, root->fs_info->generation) ||
9929                     btrfs_inode_in_log(old_inode, root->fs_info->generation) ||
9930                     (new_inode &&
9931                      btrfs_inode_in_log(new_inode, root->fs_info->generation)))
9932                     btrfs_set_log_full_commit(root->fs_info, trans);
9933
9934                 btrfs_end_log_trans(root);
9935                 log_pinned = false;
9936         }
9937         btrfs_end_transaction(trans, root);
9938 out_notrans:
9939         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
9940                 up_read(&root->fs_info->subvol_sem);
9941
9942         return ret;
9943 }
9944
9945 static int btrfs_rename2(struct inode *old_dir, struct dentry *old_dentry,
9946                          struct inode *new_dir, struct dentry *new_dentry,
9947                          unsigned int flags)
9948 {
9949         if (flags & ~(RENAME_NOREPLACE | RENAME_EXCHANGE | RENAME_WHITEOUT))
9950                 return -EINVAL;
9951
9952         if (flags & RENAME_EXCHANGE)
9953                 return btrfs_rename_exchange(old_dir, old_dentry, new_dir,
9954                                           new_dentry);
9955
9956         return btrfs_rename(old_dir, old_dentry, new_dir, new_dentry, flags);
9957 }
9958
9959 static void btrfs_run_delalloc_work(struct btrfs_work *work)
9960 {
9961         struct btrfs_delalloc_work *delalloc_work;
9962         struct inode *inode;
9963
9964         delalloc_work = container_of(work, struct btrfs_delalloc_work,
9965                                      work);
9966         inode = delalloc_work->inode;
9967         filemap_flush(inode->i_mapping);
9968         if (test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
9969                                 &BTRFS_I(inode)->runtime_flags))
9970                 filemap_flush(inode->i_mapping);
9971
9972         if (delalloc_work->delay_iput)
9973                 btrfs_add_delayed_iput(inode);
9974         else
9975                 iput(inode);
9976         complete(&delalloc_work->completion);
9977 }
9978
9979 struct btrfs_delalloc_work *btrfs_alloc_delalloc_work(struct inode *inode,
9980                                                     int delay_iput)
9981 {
9982         struct btrfs_delalloc_work *work;
9983
9984         work = kmalloc(sizeof(*work), GFP_NOFS);
9985         if (!work)
9986                 return NULL;
9987
9988         init_completion(&work->completion);
9989         INIT_LIST_HEAD(&work->list);
9990         work->inode = inode;
9991         work->delay_iput = delay_iput;
9992         WARN_ON_ONCE(!inode);
9993         btrfs_init_work(&work->work, btrfs_flush_delalloc_helper,
9994                         btrfs_run_delalloc_work, NULL, NULL);
9995
9996         return work;
9997 }
9998
9999 void btrfs_wait_and_free_delalloc_work(struct btrfs_delalloc_work *work)
10000 {
10001         wait_for_completion(&work->completion);
10002         kfree(work);
10003 }
10004
10005 /*
10006  * some fairly slow code that needs optimization. This walks the list
10007  * of all the inodes with pending delalloc and forces them to disk.
10008  */
10009 static int __start_delalloc_inodes(struct btrfs_root *root, int delay_iput,
10010                                    int nr)
10011 {
10012         struct btrfs_inode *binode;
10013         struct inode *inode;
10014         struct btrfs_delalloc_work *work, *next;
10015         struct list_head works;
10016         struct list_head splice;
10017         int ret = 0;
10018
10019         INIT_LIST_HEAD(&works);
10020         INIT_LIST_HEAD(&splice);
10021
10022         mutex_lock(&root->delalloc_mutex);
10023         spin_lock(&root->delalloc_lock);
10024         list_splice_init(&root->delalloc_inodes, &splice);
10025         while (!list_empty(&splice)) {
10026                 binode = list_entry(splice.next, struct btrfs_inode,
10027                                     delalloc_inodes);
10028
10029                 list_move_tail(&binode->delalloc_inodes,
10030                                &root->delalloc_inodes);
10031                 inode = igrab(&binode->vfs_inode);
10032                 if (!inode) {
10033                         cond_resched_lock(&root->delalloc_lock);
10034                         continue;
10035                 }
10036                 spin_unlock(&root->delalloc_lock);
10037
10038                 work = btrfs_alloc_delalloc_work(inode, delay_iput);
10039                 if (!work) {
10040                         if (delay_iput)
10041                                 btrfs_add_delayed_iput(inode);
10042                         else
10043                                 iput(inode);
10044                         ret = -ENOMEM;
10045                         goto out;
10046                 }
10047                 list_add_tail(&work->list, &works);
10048                 btrfs_queue_work(root->fs_info->flush_workers,
10049                                  &work->work);
10050                 ret++;
10051                 if (nr != -1 && ret >= nr)
10052                         goto out;
10053                 cond_resched();
10054                 spin_lock(&root->delalloc_lock);
10055         }
10056         spin_unlock(&root->delalloc_lock);
10057
10058 out:
10059         list_for_each_entry_safe(work, next, &works, list) {
10060                 list_del_init(&work->list);
10061                 btrfs_wait_and_free_delalloc_work(work);
10062         }
10063
10064         if (!list_empty_careful(&splice)) {
10065                 spin_lock(&root->delalloc_lock);
10066                 list_splice_tail(&splice, &root->delalloc_inodes);
10067                 spin_unlock(&root->delalloc_lock);
10068         }
10069         mutex_unlock(&root->delalloc_mutex);
10070         return ret;
10071 }
10072
10073 int btrfs_start_delalloc_inodes(struct btrfs_root *root, int delay_iput)
10074 {
10075         int ret;
10076
10077         if (test_bit(BTRFS_FS_STATE_ERROR, &root->fs_info->fs_state))
10078                 return -EROFS;
10079
10080         ret = __start_delalloc_inodes(root, delay_iput, -1);
10081         if (ret > 0)
10082                 ret = 0;
10083         /*
10084          * the filemap_flush will queue IO into the worker threads, but
10085          * we have to make sure the IO is actually started and that
10086          * ordered extents get created before we return
10087          */
10088         atomic_inc(&root->fs_info->async_submit_draining);
10089         while (atomic_read(&root->fs_info->nr_async_submits) ||
10090               atomic_read(&root->fs_info->async_delalloc_pages)) {
10091                 wait_event(root->fs_info->async_submit_wait,
10092                    (atomic_read(&root->fs_info->nr_async_submits) == 0 &&
10093                     atomic_read(&root->fs_info->async_delalloc_pages) == 0));
10094         }
10095         atomic_dec(&root->fs_info->async_submit_draining);
10096         return ret;
10097 }
10098
10099 int btrfs_start_delalloc_roots(struct btrfs_fs_info *fs_info, int delay_iput,
10100                                int nr)
10101 {
10102         struct btrfs_root *root;
10103         struct list_head splice;
10104         int ret;
10105
10106         if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state))
10107                 return -EROFS;
10108
10109         INIT_LIST_HEAD(&splice);
10110
10111         mutex_lock(&fs_info->delalloc_root_mutex);
10112         spin_lock(&fs_info->delalloc_root_lock);
10113         list_splice_init(&fs_info->delalloc_roots, &splice);
10114         while (!list_empty(&splice) && nr) {
10115                 root = list_first_entry(&splice, struct btrfs_root,
10116                                         delalloc_root);
10117                 root = btrfs_grab_fs_root(root);
10118                 BUG_ON(!root);
10119                 list_move_tail(&root->delalloc_root,
10120                                &fs_info->delalloc_roots);
10121                 spin_unlock(&fs_info->delalloc_root_lock);
10122
10123                 ret = __start_delalloc_inodes(root, delay_iput, nr);
10124                 btrfs_put_fs_root(root);
10125                 if (ret < 0)
10126                         goto out;
10127
10128                 if (nr != -1) {
10129                         nr -= ret;
10130                         WARN_ON(nr < 0);
10131                 }
10132                 spin_lock(&fs_info->delalloc_root_lock);
10133         }
10134         spin_unlock(&fs_info->delalloc_root_lock);
10135
10136         ret = 0;
10137         atomic_inc(&fs_info->async_submit_draining);
10138         while (atomic_read(&fs_info->nr_async_submits) ||
10139               atomic_read(&fs_info->async_delalloc_pages)) {
10140                 wait_event(fs_info->async_submit_wait,
10141                    (atomic_read(&fs_info->nr_async_submits) == 0 &&
10142                     atomic_read(&fs_info->async_delalloc_pages) == 0));
10143         }
10144         atomic_dec(&fs_info->async_submit_draining);
10145 out:
10146         if (!list_empty_careful(&splice)) {
10147                 spin_lock(&fs_info->delalloc_root_lock);
10148                 list_splice_tail(&splice, &fs_info->delalloc_roots);
10149                 spin_unlock(&fs_info->delalloc_root_lock);
10150         }
10151         mutex_unlock(&fs_info->delalloc_root_mutex);
10152         return ret;
10153 }
10154
10155 static int btrfs_symlink(struct inode *dir, struct dentry *dentry,
10156                          const char *symname)
10157 {
10158         struct btrfs_trans_handle *trans;
10159         struct btrfs_root *root = BTRFS_I(dir)->root;
10160         struct btrfs_path *path;
10161         struct btrfs_key key;
10162         struct inode *inode = NULL;
10163         int err;
10164         int drop_inode = 0;
10165         u64 objectid;
10166         u64 index = 0;
10167         int name_len;
10168         int datasize;
10169         unsigned long ptr;
10170         struct btrfs_file_extent_item *ei;
10171         struct extent_buffer *leaf;
10172
10173         name_len = strlen(symname);
10174         if (name_len > BTRFS_MAX_INLINE_DATA_SIZE(root))
10175                 return -ENAMETOOLONG;
10176
10177         /*
10178          * 2 items for inode item and ref
10179          * 2 items for dir items
10180          * 1 item for updating parent inode item
10181          * 1 item for the inline extent item
10182          * 1 item for xattr if selinux is on
10183          */
10184         trans = btrfs_start_transaction(root, 7);
10185         if (IS_ERR(trans))
10186                 return PTR_ERR(trans);
10187
10188         err = btrfs_find_free_ino(root, &objectid);
10189         if (err)
10190                 goto out_unlock;
10191
10192         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
10193                                 dentry->d_name.len, btrfs_ino(dir), objectid,
10194                                 S_IFLNK|S_IRWXUGO, &index);
10195         if (IS_ERR(inode)) {
10196                 err = PTR_ERR(inode);
10197                 goto out_unlock;
10198         }
10199
10200         /*
10201         * If the active LSM wants to access the inode during
10202         * d_instantiate it needs these. Smack checks to see
10203         * if the filesystem supports xattrs by looking at the
10204         * ops vector.
10205         */
10206         inode->i_fop = &btrfs_file_operations;
10207         inode->i_op = &btrfs_file_inode_operations;
10208         inode->i_mapping->a_ops = &btrfs_aops;
10209         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
10210
10211         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
10212         if (err)
10213                 goto out_unlock_inode;
10214
10215         path = btrfs_alloc_path();
10216         if (!path) {
10217                 err = -ENOMEM;
10218                 goto out_unlock_inode;
10219         }
10220         key.objectid = btrfs_ino(inode);
10221         key.offset = 0;
10222         key.type = BTRFS_EXTENT_DATA_KEY;
10223         datasize = btrfs_file_extent_calc_inline_size(name_len);
10224         err = btrfs_insert_empty_item(trans, root, path, &key,
10225                                       datasize);
10226         if (err) {
10227                 btrfs_free_path(path);
10228                 goto out_unlock_inode;
10229         }
10230         leaf = path->nodes[0];
10231         ei = btrfs_item_ptr(leaf, path->slots[0],
10232                             struct btrfs_file_extent_item);
10233         btrfs_set_file_extent_generation(leaf, ei, trans->transid);
10234         btrfs_set_file_extent_type(leaf, ei,
10235                                    BTRFS_FILE_EXTENT_INLINE);
10236         btrfs_set_file_extent_encryption(leaf, ei, 0);
10237         btrfs_set_file_extent_compression(leaf, ei, 0);
10238         btrfs_set_file_extent_other_encoding(leaf, ei, 0);
10239         btrfs_set_file_extent_ram_bytes(leaf, ei, name_len);
10240
10241         ptr = btrfs_file_extent_inline_start(ei);
10242         write_extent_buffer(leaf, symname, ptr, name_len);
10243         btrfs_mark_buffer_dirty(leaf);
10244         btrfs_free_path(path);
10245
10246         inode->i_op = &btrfs_symlink_inode_operations;
10247         inode_nohighmem(inode);
10248         inode->i_mapping->a_ops = &btrfs_symlink_aops;
10249         inode_set_bytes(inode, name_len);
10250         btrfs_i_size_write(inode, name_len);
10251         err = btrfs_update_inode(trans, root, inode);
10252         /*
10253          * Last step, add directory indexes for our symlink inode. This is the
10254          * last step to avoid extra cleanup of these indexes if an error happens
10255          * elsewhere above.
10256          */
10257         if (!err)
10258                 err = btrfs_add_nondir(trans, dir, dentry, inode, 0, index);
10259         if (err) {
10260                 drop_inode = 1;
10261                 goto out_unlock_inode;
10262         }
10263
10264         unlock_new_inode(inode);
10265         d_instantiate(dentry, inode);
10266
10267 out_unlock:
10268         btrfs_end_transaction(trans, root);
10269         if (drop_inode) {
10270                 inode_dec_link_count(inode);
10271                 iput(inode);
10272         }
10273         btrfs_btree_balance_dirty(root);
10274         return err;
10275
10276 out_unlock_inode:
10277         drop_inode = 1;
10278         unlock_new_inode(inode);
10279         goto out_unlock;
10280 }
10281
10282 static int __btrfs_prealloc_file_range(struct inode *inode, int mode,
10283                                        u64 start, u64 num_bytes, u64 min_size,
10284                                        loff_t actual_len, u64 *alloc_hint,
10285                                        struct btrfs_trans_handle *trans)
10286 {
10287         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
10288         struct extent_map *em;
10289         struct btrfs_root *root = BTRFS_I(inode)->root;
10290         struct btrfs_key ins;
10291         u64 cur_offset = start;
10292         u64 i_size;
10293         u64 cur_bytes;
10294         u64 last_alloc = (u64)-1;
10295         int ret = 0;
10296         bool own_trans = true;
10297
10298         if (trans)
10299                 own_trans = false;
10300         while (num_bytes > 0) {
10301                 if (own_trans) {
10302                         trans = btrfs_start_transaction(root, 3);
10303                         if (IS_ERR(trans)) {
10304                                 ret = PTR_ERR(trans);
10305                                 break;
10306                         }
10307                 }
10308
10309                 cur_bytes = min_t(u64, num_bytes, SZ_256M);
10310                 cur_bytes = max(cur_bytes, min_size);
10311                 /*
10312                  * If we are severely fragmented we could end up with really
10313                  * small allocations, so if the allocator is returning small
10314                  * chunks lets make its job easier by only searching for those
10315                  * sized chunks.
10316                  */
10317                 cur_bytes = min(cur_bytes, last_alloc);
10318                 ret = btrfs_reserve_extent(root, cur_bytes, min_size, 0,
10319                                            *alloc_hint, &ins, 1, 0);
10320                 if (ret) {
10321                         if (own_trans)
10322                                 btrfs_end_transaction(trans, root);
10323                         break;
10324                 }
10325                 btrfs_dec_block_group_reservations(root->fs_info, ins.objectid);
10326
10327                 last_alloc = ins.offset;
10328                 ret = insert_reserved_file_extent(trans, inode,
10329                                                   cur_offset, ins.objectid,
10330                                                   ins.offset, ins.offset,
10331                                                   ins.offset, 0, 0, 0,
10332                                                   BTRFS_FILE_EXTENT_PREALLOC);
10333                 if (ret) {
10334                         btrfs_free_reserved_extent(root, ins.objectid,
10335                                                    ins.offset, 0);
10336                         btrfs_abort_transaction(trans, root, ret);
10337                         if (own_trans)
10338                                 btrfs_end_transaction(trans, root);
10339                         break;
10340                 }
10341
10342                 btrfs_drop_extent_cache(inode, cur_offset,
10343                                         cur_offset + ins.offset -1, 0);
10344
10345                 em = alloc_extent_map();
10346                 if (!em) {
10347                         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
10348                                 &BTRFS_I(inode)->runtime_flags);
10349                         goto next;
10350                 }
10351
10352                 em->start = cur_offset;
10353                 em->orig_start = cur_offset;
10354                 em->len = ins.offset;
10355                 em->block_start = ins.objectid;
10356                 em->block_len = ins.offset;
10357                 em->orig_block_len = ins.offset;
10358                 em->ram_bytes = ins.offset;
10359                 em->bdev = root->fs_info->fs_devices->latest_bdev;
10360                 set_bit(EXTENT_FLAG_PREALLOC, &em->flags);
10361                 em->generation = trans->transid;
10362
10363                 while (1) {
10364                         write_lock(&em_tree->lock);
10365                         ret = add_extent_mapping(em_tree, em, 1);
10366                         write_unlock(&em_tree->lock);
10367                         if (ret != -EEXIST)
10368                                 break;
10369                         btrfs_drop_extent_cache(inode, cur_offset,
10370                                                 cur_offset + ins.offset - 1,
10371                                                 0);
10372                 }
10373                 free_extent_map(em);
10374 next:
10375                 num_bytes -= ins.offset;
10376                 cur_offset += ins.offset;
10377                 *alloc_hint = ins.objectid + ins.offset;
10378
10379                 inode_inc_iversion(inode);
10380                 inode->i_ctime = current_fs_time(inode->i_sb);
10381                 BTRFS_I(inode)->flags |= BTRFS_INODE_PREALLOC;
10382                 if (!(mode & FALLOC_FL_KEEP_SIZE) &&
10383                     (actual_len > inode->i_size) &&
10384                     (cur_offset > inode->i_size)) {
10385                         if (cur_offset > actual_len)
10386                                 i_size = actual_len;
10387                         else
10388                                 i_size = cur_offset;
10389                         i_size_write(inode, i_size);
10390                         btrfs_ordered_update_i_size(inode, i_size, NULL);
10391                 }
10392
10393                 ret = btrfs_update_inode(trans, root, inode);
10394
10395                 if (ret) {
10396                         btrfs_abort_transaction(trans, root, ret);
10397                         if (own_trans)
10398                                 btrfs_end_transaction(trans, root);
10399                         break;
10400                 }
10401
10402                 if (own_trans)
10403                         btrfs_end_transaction(trans, root);
10404         }
10405         return ret;
10406 }
10407
10408 int btrfs_prealloc_file_range(struct inode *inode, int mode,
10409                               u64 start, u64 num_bytes, u64 min_size,
10410                               loff_t actual_len, u64 *alloc_hint)
10411 {
10412         return __btrfs_prealloc_file_range(inode, mode, start, num_bytes,
10413                                            min_size, actual_len, alloc_hint,
10414                                            NULL);
10415 }
10416
10417 int btrfs_prealloc_file_range_trans(struct inode *inode,
10418                                     struct btrfs_trans_handle *trans, int mode,
10419                                     u64 start, u64 num_bytes, u64 min_size,
10420                                     loff_t actual_len, u64 *alloc_hint)
10421 {
10422         return __btrfs_prealloc_file_range(inode, mode, start, num_bytes,
10423                                            min_size, actual_len, alloc_hint, trans);
10424 }
10425
10426 static int btrfs_set_page_dirty(struct page *page)
10427 {
10428         return __set_page_dirty_nobuffers(page);
10429 }
10430
10431 static int btrfs_permission(struct inode *inode, int mask)
10432 {
10433         struct btrfs_root *root = BTRFS_I(inode)->root;
10434         umode_t mode = inode->i_mode;
10435
10436         if (mask & MAY_WRITE &&
10437             (S_ISREG(mode) || S_ISDIR(mode) || S_ISLNK(mode))) {
10438                 if (btrfs_root_readonly(root))
10439                         return -EROFS;
10440                 if (BTRFS_I(inode)->flags & BTRFS_INODE_READONLY)
10441                         return -EACCES;
10442         }
10443         return generic_permission(inode, mask);
10444 }
10445
10446 static int btrfs_tmpfile(struct inode *dir, struct dentry *dentry, umode_t mode)
10447 {
10448         struct btrfs_trans_handle *trans;
10449         struct btrfs_root *root = BTRFS_I(dir)->root;
10450         struct inode *inode = NULL;
10451         u64 objectid;
10452         u64 index;
10453         int ret = 0;
10454
10455         /*
10456          * 5 units required for adding orphan entry
10457          */
10458         trans = btrfs_start_transaction(root, 5);
10459         if (IS_ERR(trans))
10460                 return PTR_ERR(trans);
10461
10462         ret = btrfs_find_free_ino(root, &objectid);
10463         if (ret)
10464                 goto out;
10465
10466         inode = btrfs_new_inode(trans, root, dir, NULL, 0,
10467                                 btrfs_ino(dir), objectid, mode, &index);
10468         if (IS_ERR(inode)) {
10469                 ret = PTR_ERR(inode);
10470                 inode = NULL;
10471                 goto out;
10472         }
10473
10474         inode->i_fop = &btrfs_file_operations;
10475         inode->i_op = &btrfs_file_inode_operations;
10476
10477         inode->i_mapping->a_ops = &btrfs_aops;
10478         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
10479
10480         ret = btrfs_init_inode_security(trans, inode, dir, NULL);
10481         if (ret)
10482                 goto out_inode;
10483
10484         ret = btrfs_update_inode(trans, root, inode);
10485         if (ret)
10486                 goto out_inode;
10487         ret = btrfs_orphan_add(trans, inode);
10488         if (ret)
10489                 goto out_inode;
10490
10491         /*
10492          * We set number of links to 0 in btrfs_new_inode(), and here we set
10493          * it to 1 because d_tmpfile() will issue a warning if the count is 0,
10494          * through:
10495          *
10496          *    d_tmpfile() -> inode_dec_link_count() -> drop_nlink()
10497          */
10498         set_nlink(inode, 1);
10499         unlock_new_inode(inode);
10500         d_tmpfile(dentry, inode);
10501         mark_inode_dirty(inode);
10502
10503 out:
10504         btrfs_end_transaction(trans, root);
10505         if (ret)
10506                 iput(inode);
10507         btrfs_balance_delayed_items(root);
10508         btrfs_btree_balance_dirty(root);
10509         return ret;
10510
10511 out_inode:
10512         unlock_new_inode(inode);
10513         goto out;
10514
10515 }
10516
10517 /* Inspired by filemap_check_errors() */
10518 int btrfs_inode_check_errors(struct inode *inode)
10519 {
10520         int ret = 0;
10521
10522         if (test_bit(AS_ENOSPC, &inode->i_mapping->flags) &&
10523             test_and_clear_bit(AS_ENOSPC, &inode->i_mapping->flags))
10524                 ret = -ENOSPC;
10525         if (test_bit(AS_EIO, &inode->i_mapping->flags) &&
10526             test_and_clear_bit(AS_EIO, &inode->i_mapping->flags))
10527                 ret = -EIO;
10528
10529         return ret;
10530 }
10531
10532 static const struct inode_operations btrfs_dir_inode_operations = {
10533         .getattr        = btrfs_getattr,
10534         .lookup         = btrfs_lookup,
10535         .create         = btrfs_create,
10536         .unlink         = btrfs_unlink,
10537         .link           = btrfs_link,
10538         .mkdir          = btrfs_mkdir,
10539         .rmdir          = btrfs_rmdir,
10540         .rename2        = btrfs_rename2,
10541         .symlink        = btrfs_symlink,
10542         .setattr        = btrfs_setattr,
10543         .mknod          = btrfs_mknod,
10544         .setxattr       = generic_setxattr,
10545         .getxattr       = generic_getxattr,
10546         .listxattr      = btrfs_listxattr,
10547         .removexattr    = generic_removexattr,
10548         .permission     = btrfs_permission,
10549         .get_acl        = btrfs_get_acl,
10550         .set_acl        = btrfs_set_acl,
10551         .update_time    = btrfs_update_time,
10552         .tmpfile        = btrfs_tmpfile,
10553 };
10554 static const struct inode_operations btrfs_dir_ro_inode_operations = {
10555         .lookup         = btrfs_lookup,
10556         .permission     = btrfs_permission,
10557         .get_acl        = btrfs_get_acl,
10558         .set_acl        = btrfs_set_acl,
10559         .update_time    = btrfs_update_time,
10560 };
10561
10562 static const struct file_operations btrfs_dir_file_operations = {
10563         .llseek         = generic_file_llseek,
10564         .read           = generic_read_dir,
10565         .iterate_shared = btrfs_real_readdir,
10566         .unlocked_ioctl = btrfs_ioctl,
10567 #ifdef CONFIG_COMPAT
10568         .compat_ioctl   = btrfs_compat_ioctl,
10569 #endif
10570         .release        = btrfs_release_file,
10571         .fsync          = btrfs_sync_file,
10572 };
10573
10574 static const struct extent_io_ops btrfs_extent_io_ops = {
10575         .fill_delalloc = run_delalloc_range,
10576         .submit_bio_hook = btrfs_submit_bio_hook,
10577         .merge_bio_hook = btrfs_merge_bio_hook,
10578         .readpage_end_io_hook = btrfs_readpage_end_io_hook,
10579         .writepage_end_io_hook = btrfs_writepage_end_io_hook,
10580         .writepage_start_hook = btrfs_writepage_start_hook,
10581         .set_bit_hook = btrfs_set_bit_hook,
10582         .clear_bit_hook = btrfs_clear_bit_hook,
10583         .merge_extent_hook = btrfs_merge_extent_hook,
10584         .split_extent_hook = btrfs_split_extent_hook,
10585 };
10586
10587 /*
10588  * btrfs doesn't support the bmap operation because swapfiles
10589  * use bmap to make a mapping of extents in the file.  They assume
10590  * these extents won't change over the life of the file and they
10591  * use the bmap result to do IO directly to the drive.
10592  *
10593  * the btrfs bmap call would return logical addresses that aren't
10594  * suitable for IO and they also will change frequently as COW
10595  * operations happen.  So, swapfile + btrfs == corruption.
10596  *
10597  * For now we're avoiding this by dropping bmap.
10598  */
10599 static const struct address_space_operations btrfs_aops = {
10600         .readpage       = btrfs_readpage,
10601         .writepage      = btrfs_writepage,
10602         .writepages     = btrfs_writepages,
10603         .readpages      = btrfs_readpages,
10604         .direct_IO      = btrfs_direct_IO,
10605         .invalidatepage = btrfs_invalidatepage,
10606         .releasepage    = btrfs_releasepage,
10607         .set_page_dirty = btrfs_set_page_dirty,
10608         .error_remove_page = generic_error_remove_page,
10609 };
10610
10611 static const struct address_space_operations btrfs_symlink_aops = {
10612         .readpage       = btrfs_readpage,
10613         .writepage      = btrfs_writepage,
10614         .invalidatepage = btrfs_invalidatepage,
10615         .releasepage    = btrfs_releasepage,
10616 };
10617
10618 static const struct inode_operations btrfs_file_inode_operations = {
10619         .getattr        = btrfs_getattr,
10620         .setattr        = btrfs_setattr,
10621         .setxattr       = generic_setxattr,
10622         .getxattr       = generic_getxattr,
10623         .listxattr      = btrfs_listxattr,
10624         .removexattr    = generic_removexattr,
10625         .permission     = btrfs_permission,
10626         .fiemap         = btrfs_fiemap,
10627         .get_acl        = btrfs_get_acl,
10628         .set_acl        = btrfs_set_acl,
10629         .update_time    = btrfs_update_time,
10630 };
10631 static const struct inode_operations btrfs_special_inode_operations = {
10632         .getattr        = btrfs_getattr,
10633         .setattr        = btrfs_setattr,
10634         .permission     = btrfs_permission,
10635         .setxattr       = generic_setxattr,
10636         .getxattr       = generic_getxattr,
10637         .listxattr      = btrfs_listxattr,
10638         .removexattr    = generic_removexattr,
10639         .get_acl        = btrfs_get_acl,
10640         .set_acl        = btrfs_set_acl,
10641         .update_time    = btrfs_update_time,
10642 };
10643 static const struct inode_operations btrfs_symlink_inode_operations = {
10644         .readlink       = generic_readlink,
10645         .get_link       = page_get_link,
10646         .getattr        = btrfs_getattr,
10647         .setattr        = btrfs_setattr,
10648         .permission     = btrfs_permission,
10649         .setxattr       = generic_setxattr,
10650         .getxattr       = generic_getxattr,
10651         .listxattr      = btrfs_listxattr,
10652         .removexattr    = generic_removexattr,
10653         .update_time    = btrfs_update_time,
10654 };
10655
10656 const struct dentry_operations btrfs_dentry_operations = {
10657         .d_delete       = btrfs_dentry_delete,
10658         .d_release      = btrfs_dentry_release,
10659 };