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