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 > root->sectorsize ||
 267             data_len > BTRFS_MAX_INLINE_DATA_SIZE(root) ||
 268             (!compressed_size &&
 269             (actual_end & (root->sectorsize - 1)) == 0) ||
 270             end + 1 < isize ||
 271             data_len > root->fs_info->max_inline) {
 272                 return 1;
 273         }
 274
 275         path = btrfs_alloc_path();
 276         if (!path)
 277                 return -ENOMEM;
 278
 279         trans = btrfs_join_transaction(root);
 280         if (IS_ERR(trans)) {
 281                 btrfs_free_path(path);
 282                 return PTR_ERR(trans);
 283         }
 284         trans->block_rsv = &root->fs_info->delalloc_block_rsv;
 285
 286         if (compressed_size && compressed_pages)
 287                 extent_item_size = btrfs_file_extent_calc_inline_size(
 288                    compressed_size);
 289         else
 290                 extent_item_size = btrfs_file_extent_calc_inline_size(
 291                     inline_len);
 292
 293         ret = __btrfs_drop_extents(trans, root, inode, path,
 294                                    start, aligned_end, NULL,
 295                                    1, 1, extent_item_size, &extent_inserted);
 296         if (ret) {
 297                 btrfs_abort_transaction(trans, root, ret);
 298                 goto out;
 299         }
 300
 301         if (isize > actual_end)
 302                 inline_len = min_t(u64, isize, actual_end);
 303         ret = insert_inline_extent(trans, path, extent_inserted,
 304                                    root, inode, start,
 305                                    inline_len, compressed_size,
 306                                    compress_type, compressed_pages);
 307         if (ret && ret != -ENOSPC) {
 308                 btrfs_abort_transaction(trans, root, ret);
 309                 goto out;
 310         } else if (ret == -ENOSPC) {
 311                 ret = 1;
 312                 goto out;
 313         }
 314
 315         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &BTRFS_I(inode)->runtime_flags);
 316         btrfs_delalloc_release_metadata(inode, end + 1 - start);
 317         btrfs_drop_extent_cache(inode, start, aligned_end - 1, 0);
 318 out:
 319         /*
 320          * Don't forget to free the reserved space, as for inlined extent
 321          * it won't count as data extent, free them directly here.
 322          * And at reserve time, it's always aligned to page size, so
 323          * just free one page here.
 324          */
 325         btrfs_qgroup_free_data(inode, 0, PAGE_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_range(inode, page_start,
2006                                         PAGE_CACHE_SIZE);
2007         if (ordered) {
2008                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, page_start,
2009                                      page_end, &cached_state, GFP_NOFS);
2010                 unlock_page(page);
2011                 btrfs_start_ordered_extent(inode, ordered, 1);
2012                 btrfs_put_ordered_extent(ordered);
2013                 goto again;
2014         }
2015
2016         ret = btrfs_delalloc_reserve_space(inode, page_start,
2017                                            PAGE_CACHE_SIZE);
2018         if (ret) {
2019                 mapping_set_error(page->mapping, ret);
2020                 end_extent_writepage(page, ret, page_start, page_end);
2021                 ClearPageChecked(page);
2022                 goto out;
2023          }
2024
2025         btrfs_set_extent_delalloc(inode, page_start, page_end, &cached_state);
2026         ClearPageChecked(page);
2027         set_page_dirty(page);
2028 out:
2029         unlock_extent_cached(&BTRFS_I(inode)->io_tree, page_start, page_end,
2030                              &cached_state, GFP_NOFS);
2031 out_page:
2032         unlock_page(page);
2033         page_cache_release(page);
2034         kfree(fixup);
2035 }
2036
2037 /*
2038  * There are a few paths in the higher layers of the kernel that directly
2039  * set the page dirty bit without asking the filesystem if it is a
2040  * good idea.  This causes problems because we want to make sure COW
2041  * properly happens and the data=ordered rules are followed.
2042  *
2043  * In our case any range that doesn't have the ORDERED bit set
2044  * hasn't been properly setup for IO.  We kick off an async process
2045  * to fix it up.  The async helper will wait for ordered extents, set
2046  * the delalloc bit and make it safe to write the page.
2047  */
2048 static int btrfs_writepage_start_hook(struct page *page, u64 start, u64 end)
2049 {
2050         struct inode *inode = page->mapping->host;
2051         struct btrfs_writepage_fixup *fixup;
2052         struct btrfs_root *root = BTRFS_I(inode)->root;
2053
2054         /* this page is properly in the ordered list */
2055         if (TestClearPagePrivate2(page))
2056                 return 0;
2057
2058         if (PageChecked(page))
2059                 return -EAGAIN;
2060
2061         fixup = kzalloc(sizeof(*fixup), GFP_NOFS);
2062         if (!fixup)
2063                 return -EAGAIN;
2064
2065         SetPageChecked(page);
2066         page_cache_get(page);
2067         btrfs_init_work(&fixup->work, btrfs_fixup_helper,
2068                         btrfs_writepage_fixup_worker, NULL, NULL);
2069         fixup->page = page;
2070         btrfs_queue_work(root->fs_info->fixup_workers, &fixup->work);
2071         return -EBUSY;
2072 }
2073
2074 static int insert_reserved_file_extent(struct btrfs_trans_handle *trans,
2075                                        struct inode *inode, u64 file_pos,
2076                                        u64 disk_bytenr, u64 disk_num_bytes,
2077                                        u64 num_bytes, u64 ram_bytes,
2078                                        u8 compression, u8 encryption,
2079                                        u16 other_encoding, int extent_type)
2080 {
2081         struct btrfs_root *root = BTRFS_I(inode)->root;
2082         struct btrfs_file_extent_item *fi;
2083         struct btrfs_path *path;
2084         struct extent_buffer *leaf;
2085         struct btrfs_key ins;
2086         int extent_inserted = 0;
2087         int ret;
2088
2089         path = btrfs_alloc_path();
2090         if (!path)
2091                 return -ENOMEM;
2092
2093         /*
2094          * we may be replacing one extent in the tree with another.
2095          * The new extent is pinned in the extent map, and we don't want
2096          * to drop it from the cache until it is completely in the btree.
2097          *
2098          * So, tell btrfs_drop_extents to leave this extent in the cache.
2099          * the caller is expected to unpin it and allow it to be merged
2100          * with the others.
2101          */
2102         ret = __btrfs_drop_extents(trans, root, inode, path, file_pos,
2103                                    file_pos + num_bytes, NULL, 0,
2104                                    1, sizeof(*fi), &extent_inserted);
2105         if (ret)
2106                 goto out;
2107
2108         if (!extent_inserted) {
2109                 ins.objectid = btrfs_ino(inode);
2110                 ins.offset = file_pos;
2111                 ins.type = BTRFS_EXTENT_DATA_KEY;
2112
2113                 path->leave_spinning = 1;
2114                 ret = btrfs_insert_empty_item(trans, root, path, &ins,
2115                                               sizeof(*fi));
2116                 if (ret)
2117                         goto out;
2118         }
2119         leaf = path->nodes[0];
2120         fi = btrfs_item_ptr(leaf, path->slots[0],
2121                             struct btrfs_file_extent_item);
2122         btrfs_set_file_extent_generation(leaf, fi, trans->transid);
2123         btrfs_set_file_extent_type(leaf, fi, extent_type);
2124         btrfs_set_file_extent_disk_bytenr(leaf, fi, disk_bytenr);
2125         btrfs_set_file_extent_disk_num_bytes(leaf, fi, disk_num_bytes);
2126         btrfs_set_file_extent_offset(leaf, fi, 0);
2127         btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2128         btrfs_set_file_extent_ram_bytes(leaf, fi, ram_bytes);
2129         btrfs_set_file_extent_compression(leaf, fi, compression);
2130         btrfs_set_file_extent_encryption(leaf, fi, encryption);
2131         btrfs_set_file_extent_other_encoding(leaf, fi, other_encoding);
2132
2133         btrfs_mark_buffer_dirty(leaf);
2134         btrfs_release_path(path);
2135
2136         inode_add_bytes(inode, num_bytes);
2137
2138         ins.objectid = disk_bytenr;
2139         ins.offset = disk_num_bytes;
2140         ins.type = BTRFS_EXTENT_ITEM_KEY;
2141         ret = btrfs_alloc_reserved_file_extent(trans, root,
2142                                         root->root_key.objectid,
2143                                         btrfs_ino(inode), file_pos,
2144                                         ram_bytes, &ins);
2145         /*
2146          * Release the reserved range from inode dirty range map, as it is
2147          * already moved into delayed_ref_head
2148          */
2149         btrfs_qgroup_release_data(inode, file_pos, ram_bytes);
2150 out:
2151         btrfs_free_path(path);
2152
2153         return ret;
2154 }
2155
2156 /* snapshot-aware defrag */
2157 struct sa_defrag_extent_backref {
2158         struct rb_node node;
2159         struct old_sa_defrag_extent *old;
2160         u64 root_id;
2161         u64 inum;
2162         u64 file_pos;
2163         u64 extent_offset;
2164         u64 num_bytes;
2165         u64 generation;
2166 };
2167
2168 struct old_sa_defrag_extent {
2169         struct list_head list;
2170         struct new_sa_defrag_extent *new;
2171
2172         u64 extent_offset;
2173         u64 bytenr;
2174         u64 offset;
2175         u64 len;
2176         int count;
2177 };
2178
2179 struct new_sa_defrag_extent {
2180         struct rb_root root;
2181         struct list_head head;
2182         struct btrfs_path *path;
2183         struct inode *inode;
2184         u64 file_pos;
2185         u64 len;
2186         u64 bytenr;
2187         u64 disk_len;
2188         u8 compress_type;
2189 };
2190
2191 static int backref_comp(struct sa_defrag_extent_backref *b1,
2192                         struct sa_defrag_extent_backref *b2)
2193 {
2194         if (b1->root_id < b2->root_id)
2195                 return -1;
2196         else if (b1->root_id > b2->root_id)
2197                 return 1;
2198
2199         if (b1->inum < b2->inum)
2200                 return -1;
2201         else if (b1->inum > b2->inum)
2202                 return 1;
2203
2204         if (b1->file_pos < b2->file_pos)
2205                 return -1;
2206         else if (b1->file_pos > b2->file_pos)
2207                 return 1;
2208
2209         /*
2210          * [------------------------------] ===> (a range of space)
2211          *     |<--->|   |<---->| =============> (fs/file tree A)
2212          * |<---------------------------->| ===> (fs/file tree B)
2213          *
2214          * A range of space can refer to two file extents in one tree while
2215          * refer to only one file extent in another tree.
2216          *
2217          * So we may process a disk offset more than one time(two extents in A)
2218          * and locate at the same extent(one extent in B), then insert two same
2219          * backrefs(both refer to the extent in B).
2220          */
2221         return 0;
2222 }
2223
2224 static void backref_insert(struct rb_root *root,
2225                            struct sa_defrag_extent_backref *backref)
2226 {
2227         struct rb_node **p = &root->rb_node;
2228         struct rb_node *parent = NULL;
2229         struct sa_defrag_extent_backref *entry;
2230         int ret;
2231
2232         while (*p) {
2233                 parent = *p;
2234                 entry = rb_entry(parent, struct sa_defrag_extent_backref, node);
2235
2236                 ret = backref_comp(backref, entry);
2237                 if (ret < 0)
2238                         p = &(*p)->rb_left;
2239                 else
2240                         p = &(*p)->rb_right;
2241         }
2242
2243         rb_link_node(&backref->node, parent, p);
2244         rb_insert_color(&backref->node, root);
2245 }
2246
2247 /*
2248  * Note the backref might has changed, and in this case we just return 0.
2249  */
2250 static noinline int record_one_backref(u64 inum, u64 offset, u64 root_id,
2251                                        void *ctx)
2252 {
2253         struct btrfs_file_extent_item *extent;
2254         struct btrfs_fs_info *fs_info;
2255         struct old_sa_defrag_extent *old = ctx;
2256         struct new_sa_defrag_extent *new = old->new;
2257         struct btrfs_path *path = new->path;
2258         struct btrfs_key key;
2259         struct btrfs_root *root;
2260         struct sa_defrag_extent_backref *backref;
2261         struct extent_buffer *leaf;
2262         struct inode *inode = new->inode;
2263         int slot;
2264         int ret;
2265         u64 extent_offset;
2266         u64 num_bytes;
2267
2268         if (BTRFS_I(inode)->root->root_key.objectid == root_id &&
2269             inum == btrfs_ino(inode))
2270                 return 0;
2271
2272         key.objectid = root_id;
2273         key.type = BTRFS_ROOT_ITEM_KEY;
2274         key.offset = (u64)-1;
2275
2276         fs_info = BTRFS_I(inode)->root->fs_info;
2277         root = btrfs_read_fs_root_no_name(fs_info, &key);
2278         if (IS_ERR(root)) {
2279                 if (PTR_ERR(root) == -ENOENT)
2280                         return 0;
2281                 WARN_ON(1);
2282                 pr_debug("inum=%llu, offset=%llu, root_id=%llu\n",
2283                          inum, offset, root_id);
2284                 return PTR_ERR(root);
2285         }
2286
2287         key.objectid = inum;
2288         key.type = BTRFS_EXTENT_DATA_KEY;
2289         if (offset > (u64)-1 << 32)
2290                 key.offset = 0;
2291         else
2292                 key.offset = offset;
2293
2294         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2295         if (WARN_ON(ret < 0))
2296                 return ret;
2297         ret = 0;
2298
2299         while (1) {
2300                 cond_resched();
2301
2302                 leaf = path->nodes[0];
2303                 slot = path->slots[0];
2304
2305                 if (slot >= btrfs_header_nritems(leaf)) {
2306                         ret = btrfs_next_leaf(root, path);
2307                         if (ret < 0) {
2308                                 goto out;
2309                         } else if (ret > 0) {
2310                                 ret = 0;
2311                                 goto out;
2312                         }
2313                         continue;
2314                 }
2315
2316                 path->slots[0]++;
2317
2318                 btrfs_item_key_to_cpu(leaf, &key, slot);
2319
2320                 if (key.objectid > inum)
2321                         goto out;
2322
2323                 if (key.objectid < inum || key.type != BTRFS_EXTENT_DATA_KEY)
2324                         continue;
2325
2326                 extent = btrfs_item_ptr(leaf, slot,
2327                                         struct btrfs_file_extent_item);
2328
2329                 if (btrfs_file_extent_disk_bytenr(leaf, extent) != old->bytenr)
2330                         continue;
2331
2332                 /*
2333                  * 'offset' refers to the exact key.offset,
2334                  * NOT the 'offset' field in btrfs_extent_data_ref, ie.
2335                  * (key.offset - extent_offset).
2336                  */
2337                 if (key.offset != offset)
2338                         continue;
2339
2340                 extent_offset = btrfs_file_extent_offset(leaf, extent);
2341                 num_bytes = btrfs_file_extent_num_bytes(leaf, extent);
2342
2343                 if (extent_offset >= old->extent_offset + old->offset +
2344                     old->len || extent_offset + num_bytes <=
2345                     old->extent_offset + old->offset)
2346                         continue;
2347                 break;
2348         }
2349
2350         backref = kmalloc(sizeof(*backref), GFP_NOFS);
2351         if (!backref) {
2352                 ret = -ENOENT;
2353                 goto out;
2354         }
2355
2356         backref->root_id = root_id;
2357         backref->inum = inum;
2358         backref->file_pos = offset;
2359         backref->num_bytes = num_bytes;
2360         backref->extent_offset = extent_offset;
2361         backref->generation = btrfs_file_extent_generation(leaf, extent);
2362         backref->old = old;
2363         backref_insert(&new->root, backref);
2364         old->count++;
2365 out:
2366         btrfs_release_path(path);
2367         WARN_ON(ret);
2368         return ret;
2369 }
2370
2371 static noinline bool record_extent_backrefs(struct btrfs_path *path,
2372                                    struct new_sa_defrag_extent *new)
2373 {
2374         struct btrfs_fs_info *fs_info = BTRFS_I(new->inode)->root->fs_info;
2375         struct old_sa_defrag_extent *old, *tmp;
2376         int ret;
2377
2378         new->path = path;
2379
2380         list_for_each_entry_safe(old, tmp, &new->head, list) {
2381                 ret = iterate_inodes_from_logical(old->bytenr +
2382                                                   old->extent_offset, fs_info,
2383                                                   path, record_one_backref,
2384                                                   old);
2385                 if (ret < 0 && ret != -ENOENT)
2386                         return false;
2387
2388                 /* no backref to be processed for this extent */
2389                 if (!old->count) {
2390                         list_del(&old->list);
2391                         kfree(old);
2392                 }
2393         }
2394
2395         if (list_empty(&new->head))
2396                 return false;
2397
2398         return true;
2399 }
2400
2401 static int relink_is_mergable(struct extent_buffer *leaf,
2402                               struct btrfs_file_extent_item *fi,
2403                               struct new_sa_defrag_extent *new)
2404 {
2405         if (btrfs_file_extent_disk_bytenr(leaf, fi) != new->bytenr)
2406                 return 0;
2407
2408         if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG)
2409                 return 0;
2410
2411         if (btrfs_file_extent_compression(leaf, fi) != new->compress_type)
2412                 return 0;
2413
2414         if (btrfs_file_extent_encryption(leaf, fi) ||
2415             btrfs_file_extent_other_encoding(leaf, fi))
2416                 return 0;
2417
2418         return 1;
2419 }
2420
2421 /*
2422  * Note the backref might has changed, and in this case we just return 0.
2423  */
2424 static noinline int relink_extent_backref(struct btrfs_path *path,
2425                                  struct sa_defrag_extent_backref *prev,
2426                                  struct sa_defrag_extent_backref *backref)
2427 {
2428         struct btrfs_file_extent_item *extent;
2429         struct btrfs_file_extent_item *item;
2430         struct btrfs_ordered_extent *ordered;
2431         struct btrfs_trans_handle *trans;
2432         struct btrfs_fs_info *fs_info;
2433         struct btrfs_root *root;
2434         struct btrfs_key key;
2435         struct extent_buffer *leaf;
2436         struct old_sa_defrag_extent *old = backref->old;
2437         struct new_sa_defrag_extent *new = old->new;
2438         struct inode *src_inode = new->inode;
2439         struct inode *inode;
2440         struct extent_state *cached = NULL;
2441         int ret = 0;
2442         u64 start;
2443         u64 len;
2444         u64 lock_start;
2445         u64 lock_end;
2446         bool merge = false;
2447         int index;
2448
2449         if (prev && prev->root_id == backref->root_id &&
2450             prev->inum == backref->inum &&
2451             prev->file_pos + prev->num_bytes == backref->file_pos)
2452                 merge = true;
2453
2454         /* step 1: get root */
2455         key.objectid = backref->root_id;
2456         key.type = BTRFS_ROOT_ITEM_KEY;
2457         key.offset = (u64)-1;
2458
2459         fs_info = BTRFS_I(src_inode)->root->fs_info;
2460         index = srcu_read_lock(&fs_info->subvol_srcu);
2461
2462         root = btrfs_read_fs_root_no_name(fs_info, &key);
2463         if (IS_ERR(root)) {
2464                 srcu_read_unlock(&fs_info->subvol_srcu, index);
2465                 if (PTR_ERR(root) == -ENOENT)
2466                         return 0;
2467                 return PTR_ERR(root);
2468         }
2469
2470         if (btrfs_root_readonly(root)) {
2471                 srcu_read_unlock(&fs_info->subvol_srcu, index);
2472                 return 0;
2473         }
2474
2475         /* step 2: get inode */
2476         key.objectid = backref->inum;
2477         key.type = BTRFS_INODE_ITEM_KEY;
2478         key.offset = 0;
2479
2480         inode = btrfs_iget(fs_info->sb, &key, root, NULL);
2481         if (IS_ERR(inode)) {
2482                 srcu_read_unlock(&fs_info->subvol_srcu, index);
2483                 return 0;
2484         }
2485
2486         srcu_read_unlock(&fs_info->subvol_srcu, index);
2487
2488         /* step 3: relink backref */
2489         lock_start = backref->file_pos;
2490         lock_end = backref->file_pos + backref->num_bytes - 1;
2491         lock_extent_bits(&BTRFS_I(inode)->io_tree, lock_start, lock_end,
2492                          &cached);
2493
2494         ordered = btrfs_lookup_first_ordered_extent(inode, lock_end);
2495         if (ordered) {
2496                 btrfs_put_ordered_extent(ordered);
2497                 goto out_unlock;
2498         }
2499
2500         trans = btrfs_join_transaction(root);
2501         if (IS_ERR(trans)) {
2502                 ret = PTR_ERR(trans);
2503                 goto out_unlock;
2504         }
2505
2506         key.objectid = backref->inum;
2507         key.type = BTRFS_EXTENT_DATA_KEY;
2508         key.offset = backref->file_pos;
2509
2510         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2511         if (ret < 0) {
2512                 goto out_free_path;
2513         } else if (ret > 0) {
2514                 ret = 0;
2515                 goto out_free_path;
2516         }
2517
2518         extent = btrfs_item_ptr(path->nodes[0], path->slots[0],
2519                                 struct btrfs_file_extent_item);
2520
2521         if (btrfs_file_extent_generation(path->nodes[0], extent) !=
2522             backref->generation)
2523                 goto out_free_path;
2524
2525         btrfs_release_path(path);
2526
2527         start = backref->file_pos;
2528         if (backref->extent_offset < old->extent_offset + old->offset)
2529                 start += old->extent_offset + old->offset -
2530                          backref->extent_offset;
2531
2532         len = min(backref->extent_offset + backref->num_bytes,
2533                   old->extent_offset + old->offset + old->len);
2534         len -= max(backref->extent_offset, old->extent_offset + old->offset);
2535
2536         ret = btrfs_drop_extents(trans, root, inode, start,
2537                                  start + len, 1);
2538         if (ret)
2539                 goto out_free_path;
2540 again:
2541         key.objectid = btrfs_ino(inode);
2542         key.type = BTRFS_EXTENT_DATA_KEY;
2543         key.offset = start;
2544
2545         path->leave_spinning = 1;
2546         if (merge) {
2547                 struct btrfs_file_extent_item *fi;
2548                 u64 extent_len;
2549                 struct btrfs_key found_key;
2550
2551                 ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
2552                 if (ret < 0)
2553                         goto out_free_path;
2554
2555                 path->slots[0]--;
2556                 leaf = path->nodes[0];
2557                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
2558
2559                 fi = btrfs_item_ptr(leaf, path->slots[0],
2560                                     struct btrfs_file_extent_item);
2561                 extent_len = btrfs_file_extent_num_bytes(leaf, fi);
2562
2563                 if (extent_len + found_key.offset == start &&
2564                     relink_is_mergable(leaf, fi, new)) {
2565                         btrfs_set_file_extent_num_bytes(leaf, fi,
2566                                                         extent_len + len);
2567                         btrfs_mark_buffer_dirty(leaf);
2568                         inode_add_bytes(inode, len);
2569
2570                         ret = 1;
2571                         goto out_free_path;
2572                 } else {
2573                         merge = false;
2574                         btrfs_release_path(path);
2575                         goto again;
2576                 }
2577         }
2578
2579         ret = btrfs_insert_empty_item(trans, root, path, &key,
2580                                         sizeof(*extent));
2581         if (ret) {
2582                 btrfs_abort_transaction(trans, root, ret);
2583                 goto out_free_path;
2584         }
2585
2586         leaf = path->nodes[0];
2587         item = btrfs_item_ptr(leaf, path->slots[0],
2588                                 struct btrfs_file_extent_item);
2589         btrfs_set_file_extent_disk_bytenr(leaf, item, new->bytenr);
2590         btrfs_set_file_extent_disk_num_bytes(leaf, item, new->disk_len);
2591         btrfs_set_file_extent_offset(leaf, item, start - new->file_pos);
2592         btrfs_set_file_extent_num_bytes(leaf, item, len);
2593         btrfs_set_file_extent_ram_bytes(leaf, item, new->len);
2594         btrfs_set_file_extent_generation(leaf, item, trans->transid);
2595         btrfs_set_file_extent_type(leaf, item, BTRFS_FILE_EXTENT_REG);
2596         btrfs_set_file_extent_compression(leaf, item, new->compress_type);
2597         btrfs_set_file_extent_encryption(leaf, item, 0);
2598         btrfs_set_file_extent_other_encoding(leaf, item, 0);
2599
2600         btrfs_mark_buffer_dirty(leaf);
2601         inode_add_bytes(inode, len);
2602         btrfs_release_path(path);
2603
2604         ret = btrfs_inc_extent_ref(trans, root, new->bytenr,
2605                         new->disk_len, 0,
2606                         backref->root_id, backref->inum,
2607                         new->file_pos); /* start - extent_offset */
2608         if (ret) {
2609                 btrfs_abort_transaction(trans, root, ret);
2610                 goto out_free_path;
2611         }
2612
2613         ret = 1;
2614 out_free_path:
2615         btrfs_release_path(path);
2616         path->leave_spinning = 0;
2617         btrfs_end_transaction(trans, root);
2618 out_unlock:
2619         unlock_extent_cached(&BTRFS_I(inode)->io_tree, lock_start, lock_end,
2620                              &cached, GFP_NOFS);
2621         iput(inode);
2622         return ret;
2623 }
2624
2625 static void free_sa_defrag_extent(struct new_sa_defrag_extent *new)
2626 {
2627         struct old_sa_defrag_extent *old, *tmp;
2628
2629         if (!new)
2630                 return;
2631
2632         list_for_each_entry_safe(old, tmp, &new->head, list) {
2633                 kfree(old);
2634         }
2635         kfree(new);
2636 }
2637
2638 static void relink_file_extents(struct new_sa_defrag_extent *new)
2639 {
2640         struct btrfs_path *path;
2641         struct sa_defrag_extent_backref *backref;
2642         struct sa_defrag_extent_backref *prev = NULL;
2643         struct inode *inode;
2644         struct btrfs_root *root;
2645         struct rb_node *node;
2646         int ret;
2647
2648         inode = new->inode;
2649         root = BTRFS_I(inode)->root;
2650
2651         path = btrfs_alloc_path();
2652         if (!path)
2653                 return;
2654
2655         if (!record_extent_backrefs(path, new)) {
2656                 btrfs_free_path(path);
2657                 goto out;
2658         }
2659         btrfs_release_path(path);
2660
2661         while (1) {
2662                 node = rb_first(&new->root);
2663                 if (!node)
2664                         break;
2665                 rb_erase(node, &new->root);
2666
2667                 backref = rb_entry(node, struct sa_defrag_extent_backref, node);
2668
2669                 ret = relink_extent_backref(path, prev, backref);
2670                 WARN_ON(ret < 0);
2671
2672                 kfree(prev);
2673
2674                 if (ret == 1)
2675                         prev = backref;
2676                 else
2677                         prev = NULL;
2678                 cond_resched();
2679         }
2680         kfree(prev);
2681
2682         btrfs_free_path(path);
2683 out:
2684         free_sa_defrag_extent(new);
2685
2686         atomic_dec(&root->fs_info->defrag_running);
2687         wake_up(&root->fs_info->transaction_wait);
2688 }
2689
2690 static struct new_sa_defrag_extent *
2691 record_old_file_extents(struct inode *inode,
2692                         struct btrfs_ordered_extent *ordered)
2693 {
2694         struct btrfs_root *root = BTRFS_I(inode)->root;
2695         struct btrfs_path *path;
2696         struct btrfs_key key;
2697         struct old_sa_defrag_extent *old;
2698         struct new_sa_defrag_extent *new;
2699         int ret;
2700
2701         new = kmalloc(sizeof(*new), GFP_NOFS);
2702         if (!new)
2703                 return NULL;
2704
2705         new->inode = inode;
2706         new->file_pos = ordered->file_offset;
2707         new->len = ordered->len;
2708         new->bytenr = ordered->start;
2709         new->disk_len = ordered->disk_len;
2710         new->compress_type = ordered->compress_type;
2711         new->root = RB_ROOT;
2712         INIT_LIST_HEAD(&new->head);
2713
2714         path = btrfs_alloc_path();
2715         if (!path)
2716                 goto out_kfree;
2717
2718         key.objectid = btrfs_ino(inode);
2719         key.type = BTRFS_EXTENT_DATA_KEY;
2720         key.offset = new->file_pos;
2721
2722         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2723         if (ret < 0)
2724                 goto out_free_path;
2725         if (ret > 0 && path->slots[0] > 0)
2726                 path->slots[0]--;
2727
2728         /* find out all the old extents for the file range */
2729         while (1) {
2730                 struct btrfs_file_extent_item *extent;
2731                 struct extent_buffer *l;
2732                 int slot;
2733                 u64 num_bytes;
2734                 u64 offset;
2735                 u64 end;
2736                 u64 disk_bytenr;
2737                 u64 extent_offset;
2738
2739                 l = path->nodes[0];
2740                 slot = path->slots[0];
2741
2742                 if (slot >= btrfs_header_nritems(l)) {
2743                         ret = btrfs_next_leaf(root, path);
2744                         if (ret < 0)
2745                                 goto out_free_path;
2746                         else if (ret > 0)
2747                                 break;
2748                         continue;
2749                 }
2750
2751                 btrfs_item_key_to_cpu(l, &key, slot);
2752
2753                 if (key.objectid != btrfs_ino(inode))
2754                         break;
2755                 if (key.type != BTRFS_EXTENT_DATA_KEY)
2756                         break;
2757                 if (key.offset >= new->file_pos + new->len)
2758                         break;
2759
2760                 extent = btrfs_item_ptr(l, slot, struct btrfs_file_extent_item);
2761
2762                 num_bytes = btrfs_file_extent_num_bytes(l, extent);
2763                 if (key.offset + num_bytes < new->file_pos)
2764                         goto next;
2765
2766                 disk_bytenr = btrfs_file_extent_disk_bytenr(l, extent);
2767                 if (!disk_bytenr)
2768                         goto next;
2769
2770                 extent_offset = btrfs_file_extent_offset(l, extent);
2771
2772                 old = kmalloc(sizeof(*old), GFP_NOFS);
2773                 if (!old)
2774                         goto out_free_path;
2775
2776                 offset = max(new->file_pos, key.offset);
2777                 end = min(new->file_pos + new->len, key.offset + num_bytes);
2778
2779                 old->bytenr = disk_bytenr;
2780                 old->extent_offset = extent_offset;
2781                 old->offset = offset - key.offset;
2782                 old->len = end - offset;
2783                 old->new = new;
2784                 old->count = 0;
2785                 list_add_tail(&old->list, &new->head);
2786 next:
2787                 path->slots[0]++;
2788                 cond_resched();
2789         }
2790
2791         btrfs_free_path(path);
2792         atomic_inc(&root->fs_info->defrag_running);
2793
2794         return new;
2795
2796 out_free_path:
2797         btrfs_free_path(path);
2798 out_kfree:
2799         free_sa_defrag_extent(new);
2800         return NULL;
2801 }
2802
2803 static void btrfs_release_delalloc_bytes(struct btrfs_root *root,
2804                                          u64 start, u64 len)
2805 {
2806         struct btrfs_block_group_cache *cache;
2807
2808         cache = btrfs_lookup_block_group(root->fs_info, start);
2809         ASSERT(cache);
2810
2811         spin_lock(&cache->lock);
2812         cache->delalloc_bytes -= len;
2813         spin_unlock(&cache->lock);
2814
2815         btrfs_put_block_group(cache);
2816 }
2817
2818 /* as ordered data IO finishes, this gets called so we can finish
2819  * an ordered extent if the range of bytes in the file it covers are
2820  * fully written.
2821  */
2822 static int btrfs_finish_ordered_io(struct btrfs_ordered_extent *ordered_extent)
2823 {
2824         struct inode *inode = ordered_extent->inode;
2825         struct btrfs_root *root = BTRFS_I(inode)->root;
2826         struct btrfs_trans_handle *trans = NULL;
2827         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
2828         struct extent_state *cached_state = NULL;
2829         struct new_sa_defrag_extent *new = NULL;
2830         int compress_type = 0;
2831         int ret = 0;
2832         u64 logical_len = ordered_extent->len;
2833         bool nolock;
2834         bool truncated = false;
2835
2836         nolock = btrfs_is_free_space_inode(inode);
2837
2838         if (test_bit(BTRFS_ORDERED_IOERR, &ordered_extent->flags)) {
2839                 ret = -EIO;
2840                 goto out;
2841         }
2842
2843         btrfs_free_io_failure_record(inode, ordered_extent->file_offset,
2844                                      ordered_extent->file_offset +
2845                                      ordered_extent->len - 1);
2846
2847         if (test_bit(BTRFS_ORDERED_TRUNCATED, &ordered_extent->flags)) {
2848                 truncated = true;
2849                 logical_len = ordered_extent->truncated_len;
2850                 /* Truncated the entire extent, don't bother adding */
2851                 if (!logical_len)
2852                         goto out;
2853         }
2854
2855         if (test_bit(BTRFS_ORDERED_NOCOW, &ordered_extent->flags)) {
2856                 BUG_ON(!list_empty(&ordered_extent->list)); /* Logic error */
2857
2858                 /*
2859                  * For mwrite(mmap + memset to write) case, we still reserve
2860                  * space for NOCOW range.
2861                  * As NOCOW won't cause a new delayed ref, just free the space
2862                  */
2863                 btrfs_qgroup_free_data(inode, ordered_extent->file_offset,
2864                                        ordered_extent->len);
2865                 btrfs_ordered_update_i_size(inode, 0, ordered_extent);
2866                 if (nolock)
2867                         trans = btrfs_join_transaction_nolock(root);
2868                 else
2869                         trans = btrfs_join_transaction(root);
2870                 if (IS_ERR(trans)) {
2871                         ret = PTR_ERR(trans);
2872                         trans = NULL;
2873                         goto out;
2874                 }
2875                 trans->block_rsv = &root->fs_info->delalloc_block_rsv;
2876                 ret = btrfs_update_inode_fallback(trans, root, inode);
2877                 if (ret) /* -ENOMEM or corruption */
2878                         btrfs_abort_transaction(trans, root, ret);
2879                 goto out;
2880         }
2881
2882         lock_extent_bits(io_tree, ordered_extent->file_offset,
2883                          ordered_extent->file_offset + ordered_extent->len - 1,
2884                          &cached_state);
2885
2886         ret = test_range_bit(io_tree, ordered_extent->file_offset,
2887                         ordered_extent->file_offset + ordered_extent->len - 1,
2888                         EXTENT_DEFRAG, 1, cached_state);
2889         if (ret) {
2890                 u64 last_snapshot = btrfs_root_last_snapshot(&root->root_item);
2891                 if (0 && last_snapshot >= BTRFS_I(inode)->generation)
2892                         /* the inode is shared */
2893                         new = record_old_file_extents(inode, ordered_extent);
2894
2895                 clear_extent_bit(io_tree, ordered_extent->file_offset,
2896                         ordered_extent->file_offset + ordered_extent->len - 1,
2897                         EXTENT_DEFRAG, 0, 0, &cached_state, GFP_NOFS);
2898         }
2899
2900         if (nolock)
2901                 trans = btrfs_join_transaction_nolock(root);
2902         else
2903                 trans = btrfs_join_transaction(root);
2904         if (IS_ERR(trans)) {
2905                 ret = PTR_ERR(trans);
2906                 trans = NULL;
2907                 goto out_unlock;
2908         }
2909
2910         trans->block_rsv = &root->fs_info->delalloc_block_rsv;
2911
2912         if (test_bit(BTRFS_ORDERED_COMPRESSED, &ordered_extent->flags))
2913                 compress_type = ordered_extent->compress_type;
2914         if (test_bit(BTRFS_ORDERED_PREALLOC, &ordered_extent->flags)) {
2915                 BUG_ON(compress_type);
2916                 ret = btrfs_mark_extent_written(trans, inode,
2917                                                 ordered_extent->file_offset,
2918                                                 ordered_extent->file_offset +
2919                                                 logical_len);
2920         } else {
2921                 BUG_ON(root == root->fs_info->tree_root);
2922                 ret = insert_reserved_file_extent(trans, inode,
2923                                                 ordered_extent->file_offset,
2924                                                 ordered_extent->start,
2925                                                 ordered_extent->disk_len,
2926                                                 logical_len, logical_len,
2927                                                 compress_type, 0, 0,
2928                                                 BTRFS_FILE_EXTENT_REG);
2929                 if (!ret)
2930                         btrfs_release_delalloc_bytes(root,
2931                                                      ordered_extent->start,
2932                                                      ordered_extent->disk_len);
2933         }
2934         unpin_extent_cache(&BTRFS_I(inode)->extent_tree,
2935                            ordered_extent->file_offset, ordered_extent->len,
2936                            trans->transid);
2937         if (ret < 0) {
2938                 btrfs_abort_transaction(trans, root, ret);
2939                 goto out_unlock;
2940         }
2941
2942         add_pending_csums(trans, inode, ordered_extent->file_offset,
2943                           &ordered_extent->list);
2944
2945         btrfs_ordered_update_i_size(inode, 0, ordered_extent);
2946         ret = btrfs_update_inode_fallback(trans, root, inode);
2947         if (ret) { /* -ENOMEM or corruption */
2948                 btrfs_abort_transaction(trans, root, ret);
2949                 goto out_unlock;
2950         }
2951         ret = 0;
2952 out_unlock:
2953         unlock_extent_cached(io_tree, ordered_extent->file_offset,
2954                              ordered_extent->file_offset +
2955                              ordered_extent->len - 1, &cached_state, GFP_NOFS);
2956 out:
2957         if (root != root->fs_info->tree_root)
2958                 btrfs_delalloc_release_metadata(inode, ordered_extent->len);
2959         if (trans)
2960                 btrfs_end_transaction(trans, root);
2961
2962         if (ret || truncated) {
2963                 u64 start, end;
2964
2965                 if (truncated)
2966                         start = ordered_extent->file_offset + logical_len;
2967                 else
2968                         start = ordered_extent->file_offset;
2969                 end = ordered_extent->file_offset + ordered_extent->len - 1;
2970                 clear_extent_uptodate(io_tree, start, end, NULL, GFP_NOFS);
2971
2972                 /* Drop the cache for the part of the extent we didn't write. */
2973                 btrfs_drop_extent_cache(inode, start, end, 0);
2974
2975                 /*
2976                  * If the ordered extent had an IOERR or something else went
2977                  * wrong we need to return the space for this ordered extent
2978                  * back to the allocator.  We only free the extent in the
2979                  * truncated case if we didn't write out the extent at all.
2980                  */
2981                 if ((ret || !logical_len) &&
2982                     !test_bit(BTRFS_ORDERED_NOCOW, &ordered_extent->flags) &&
2983                     !test_bit(BTRFS_ORDERED_PREALLOC, &ordered_extent->flags))
2984                         btrfs_free_reserved_extent(root, ordered_extent->start,
2985                                                    ordered_extent->disk_len, 1);
2986         }
2987
2988
2989         /*
2990          * This needs to be done to make sure anybody waiting knows we are done
2991          * updating everything for this ordered extent.
2992          */
2993         btrfs_remove_ordered_extent(inode, ordered_extent);
2994
2995         /* for snapshot-aware defrag */
2996         if (new) {
2997                 if (ret) {
2998                         free_sa_defrag_extent(new);
2999                         atomic_dec(&root->fs_info->defrag_running);
3000                 } else {
3001                         relink_file_extents(new);
3002                 }
3003         }
3004
3005         /* once for us */
3006         btrfs_put_ordered_extent(ordered_extent);
3007         /* once for the tree */
3008         btrfs_put_ordered_extent(ordered_extent);
3009
3010         return ret;
3011 }
3012
3013 static void finish_ordered_fn(struct btrfs_work *work)
3014 {
3015         struct btrfs_ordered_extent *ordered_extent;
3016         ordered_extent = container_of(work, struct btrfs_ordered_extent, work);
3017         btrfs_finish_ordered_io(ordered_extent);
3018 }
3019
3020 static int btrfs_writepage_end_io_hook(struct page *page, u64 start, u64 end,
3021                                 struct extent_state *state, int uptodate)
3022 {
3023         struct inode *inode = page->mapping->host;
3024         struct btrfs_root *root = BTRFS_I(inode)->root;
3025         struct btrfs_ordered_extent *ordered_extent = NULL;
3026         struct btrfs_workqueue *wq;
3027         btrfs_work_func_t func;
3028
3029         trace_btrfs_writepage_end_io_hook(page, start, end, uptodate);
3030
3031         ClearPagePrivate2(page);
3032         if (!btrfs_dec_test_ordered_pending(inode, &ordered_extent, start,
3033                                             end - start + 1, uptodate))
3034                 return 0;
3035
3036         if (btrfs_is_free_space_inode(inode)) {
3037                 wq = root->fs_info->endio_freespace_worker;
3038                 func = btrfs_freespace_write_helper;
3039         } else {
3040                 wq = root->fs_info->endio_write_workers;
3041                 func = btrfs_endio_write_helper;
3042         }
3043
3044         btrfs_init_work(&ordered_extent->work, func, finish_ordered_fn, NULL,
3045                         NULL);
3046         btrfs_queue_work(wq, &ordered_extent->work);
3047
3048         return 0;
3049 }
3050
3051 static int __readpage_endio_check(struct inode *inode,
3052                                   struct btrfs_io_bio *io_bio,
3053                                   int icsum, struct page *page,
3054                                   int pgoff, u64 start, size_t len)
3055 {
3056         char *kaddr;
3057         u32 csum_expected;
3058         u32 csum = ~(u32)0;
3059
3060         csum_expected = *(((u32 *)io_bio->csum) + icsum);
3061
3062         kaddr = kmap_atomic(page);
3063         csum = btrfs_csum_data(kaddr + pgoff, csum,  len);
3064         btrfs_csum_final(csum, (char *)&csum);
3065         if (csum != csum_expected)
3066                 goto zeroit;
3067
3068         kunmap_atomic(kaddr);
3069         return 0;
3070 zeroit:
3071         btrfs_warn_rl(BTRFS_I(inode)->root->fs_info,
3072                 "csum failed ino %llu off %llu csum %u expected csum %u",
3073                            btrfs_ino(inode), start, csum, csum_expected);
3074         memset(kaddr + pgoff, 1, len);
3075         flush_dcache_page(page);
3076         kunmap_atomic(kaddr);
3077         if (csum_expected == 0)
3078                 return 0;
3079         return -EIO;
3080 }
3081
3082 /*
3083  * when reads are done, we need to check csums to verify the data is correct
3084  * if there's a match, we allow the bio to finish.  If not, the code in
3085  * extent_io.c will try to find good copies for us.
3086  */
3087 static int btrfs_readpage_end_io_hook(struct btrfs_io_bio *io_bio,
3088                                       u64 phy_offset, struct page *page,
3089                                       u64 start, u64 end, int mirror)
3090 {
3091         size_t offset = start - page_offset(page);
3092         struct inode *inode = page->mapping->host;
3093         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
3094         struct btrfs_root *root = BTRFS_I(inode)->root;
3095
3096         if (PageChecked(page)) {
3097                 ClearPageChecked(page);
3098                 return 0;
3099         }
3100
3101         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)
3102                 return 0;
3103
3104         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID &&
3105             test_range_bit(io_tree, start, end, EXTENT_NODATASUM, 1, NULL)) {
3106                 clear_extent_bits(io_tree, start, end, EXTENT_NODATASUM,
3107                                   GFP_NOFS);
3108                 return 0;
3109         }
3110
3111         phy_offset >>= inode->i_sb->s_blocksize_bits;
3112         return __readpage_endio_check(inode, io_bio, phy_offset, page, offset,
3113                                       start, (size_t)(end - start + 1));
3114 }
3115
3116 void btrfs_add_delayed_iput(struct inode *inode)
3117 {
3118         struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
3119         struct btrfs_inode *binode = BTRFS_I(inode);
3120
3121         if (atomic_add_unless(&inode->i_count, -1, 1))
3122                 return;
3123
3124         spin_lock(&fs_info->delayed_iput_lock);
3125         if (binode->delayed_iput_count == 0) {
3126                 ASSERT(list_empty(&binode->delayed_iput));
3127                 list_add_tail(&binode->delayed_iput, &fs_info->delayed_iputs);
3128         } else {
3129                 binode->delayed_iput_count++;
3130         }
3131         spin_unlock(&fs_info->delayed_iput_lock);
3132 }
3133
3134 void btrfs_run_delayed_iputs(struct btrfs_root *root)
3135 {
3136         struct btrfs_fs_info *fs_info = root->fs_info;
3137
3138         spin_lock(&fs_info->delayed_iput_lock);
3139         while (!list_empty(&fs_info->delayed_iputs)) {
3140                 struct btrfs_inode *inode;
3141
3142                 inode = list_first_entry(&fs_info->delayed_iputs,
3143                                 struct btrfs_inode, delayed_iput);
3144                 if (inode->delayed_iput_count) {
3145                         inode->delayed_iput_count--;
3146                         list_move_tail(&inode->delayed_iput,
3147                                         &fs_info->delayed_iputs);
3148                 } else {
3149                         list_del_init(&inode->delayed_iput);
3150                 }
3151                 spin_unlock(&fs_info->delayed_iput_lock);
3152                 iput(&inode->vfs_inode);
3153                 spin_lock(&fs_info->delayed_iput_lock);
3154         }
3155         spin_unlock(&fs_info->delayed_iput_lock);
3156 }
3157
3158 /*
3159  * This is called in transaction commit time. If there are no orphan
3160  * files in the subvolume, it removes orphan item and frees block_rsv
3161  * structure.
3162  */
3163 void btrfs_orphan_commit_root(struct btrfs_trans_handle *trans,
3164                               struct btrfs_root *root)
3165 {
3166         struct btrfs_block_rsv *block_rsv;
3167         int ret;
3168
3169         if (atomic_read(&root->orphan_inodes) ||
3170             root->orphan_cleanup_state != ORPHAN_CLEANUP_DONE)
3171                 return;
3172
3173         spin_lock(&root->orphan_lock);
3174         if (atomic_read(&root->orphan_inodes)) {
3175                 spin_unlock(&root->orphan_lock);
3176                 return;
3177         }
3178
3179         if (root->orphan_cleanup_state != ORPHAN_CLEANUP_DONE) {
3180                 spin_unlock(&root->orphan_lock);
3181                 return;
3182         }
3183
3184         block_rsv = root->orphan_block_rsv;
3185         root->orphan_block_rsv = NULL;
3186         spin_unlock(&root->orphan_lock);
3187
3188         if (test_bit(BTRFS_ROOT_ORPHAN_ITEM_INSERTED, &root->state) &&
3189             btrfs_root_refs(&root->root_item) > 0) {
3190                 ret = btrfs_del_orphan_item(trans, root->fs_info->tree_root,
3191                                             root->root_key.objectid);
3192                 if (ret)
3193                         btrfs_abort_transaction(trans, root, ret);
3194                 else
3195                         clear_bit(BTRFS_ROOT_ORPHAN_ITEM_INSERTED,
3196                                   &root->state);
3197         }
3198
3199         if (block_rsv) {
3200                 WARN_ON(block_rsv->size > 0);
3201                 btrfs_free_block_rsv(root, block_rsv);
3202         }
3203 }
3204
3205 /*
3206  * This creates an orphan entry for the given inode in case something goes
3207  * wrong in the middle of an unlink/truncate.
3208  *
3209  * NOTE: caller of this function should reserve 5 units of metadata for
3210  *       this function.
3211  */
3212 int btrfs_orphan_add(struct btrfs_trans_handle *trans, struct inode *inode)
3213 {
3214         struct btrfs_root *root = BTRFS_I(inode)->root;
3215         struct btrfs_block_rsv *block_rsv = NULL;
3216         int reserve = 0;
3217         int insert = 0;
3218         int ret;
3219
3220         if (!root->orphan_block_rsv) {
3221                 block_rsv = btrfs_alloc_block_rsv(root, BTRFS_BLOCK_RSV_TEMP);
3222                 if (!block_rsv)
3223                         return -ENOMEM;
3224         }
3225
3226         spin_lock(&root->orphan_lock);
3227         if (!root->orphan_block_rsv) {
3228                 root->orphan_block_rsv = block_rsv;
3229         } else if (block_rsv) {
3230                 btrfs_free_block_rsv(root, block_rsv);
3231                 block_rsv = NULL;
3232         }
3233
3234         if (!test_and_set_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
3235                               &BTRFS_I(inode)->runtime_flags)) {
3236 #if 0
3237                 /*
3238                  * For proper ENOSPC handling, we should do orphan
3239                  * cleanup when mounting. But this introduces backward
3240                  * compatibility issue.
3241                  */
3242                 if (!xchg(&root->orphan_item_inserted, 1))
3243                         insert = 2;
3244                 else
3245                         insert = 1;
3246 #endif
3247                 insert = 1;
3248                 atomic_inc(&root->orphan_inodes);
3249         }
3250
3251         if (!test_and_set_bit(BTRFS_INODE_ORPHAN_META_RESERVED,
3252                               &BTRFS_I(inode)->runtime_flags))
3253                 reserve = 1;
3254         spin_unlock(&root->orphan_lock);
3255
3256         /* grab metadata reservation from transaction handle */
3257         if (reserve) {
3258                 ret = btrfs_orphan_reserve_metadata(trans, inode);
3259                 BUG_ON(ret); /* -ENOSPC in reservation; Logic error? JDM */
3260         }
3261
3262         /* insert an orphan item to track this unlinked/truncated file */
3263         if (insert >= 1) {
3264                 ret = btrfs_insert_orphan_item(trans, root, btrfs_ino(inode));
3265                 if (ret) {
3266                         atomic_dec(&root->orphan_inodes);
3267                         if (reserve) {
3268                                 clear_bit(BTRFS_INODE_ORPHAN_META_RESERVED,
3269                                           &BTRFS_I(inode)->runtime_flags);
3270                                 btrfs_orphan_release_metadata(inode);
3271                         }
3272                         if (ret != -EEXIST) {
3273                                 clear_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
3274                                           &BTRFS_I(inode)->runtime_flags);
3275                                 btrfs_abort_transaction(trans, root, ret);
3276                                 return ret;
3277                         }
3278                 }
3279                 ret = 0;
3280         }
3281
3282         /* insert an orphan item to track subvolume contains orphan files */
3283         if (insert >= 2) {
3284                 ret = btrfs_insert_orphan_item(trans, root->fs_info->tree_root,
3285                                                root->root_key.objectid);
3286                 if (ret && ret != -EEXIST) {
3287                         btrfs_abort_transaction(trans, root, ret);
3288                         return ret;
3289                 }
3290         }
3291         return 0;
3292 }
3293
3294 /*
3295  * We have done the truncate/delete so we can go ahead and remove the orphan
3296  * item for this particular inode.
3297  */
3298 static int btrfs_orphan_del(struct btrfs_trans_handle *trans,
3299                             struct inode *inode)
3300 {
3301         struct btrfs_root *root = BTRFS_I(inode)->root;
3302         int delete_item = 0;
3303         int release_rsv = 0;
3304         int ret = 0;
3305
3306         spin_lock(&root->orphan_lock);
3307         if (test_and_clear_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
3308                                &BTRFS_I(inode)->runtime_flags))
3309                 delete_item = 1;
3310
3311         if (test_and_clear_bit(BTRFS_INODE_ORPHAN_META_RESERVED,
3312                                &BTRFS_I(inode)->runtime_flags))
3313                 release_rsv = 1;
3314         spin_unlock(&root->orphan_lock);
3315
3316         if (delete_item) {
3317                 atomic_dec(&root->orphan_inodes);
3318                 if (trans)
3319                         ret = btrfs_del_orphan_item(trans, root,
3320                                                     btrfs_ino(inode));
3321         }
3322
3323         if (release_rsv)
3324                 btrfs_orphan_release_metadata(inode);
3325
3326         return ret;
3327 }
3328
3329 /*
3330  * this cleans up any orphans that may be left on the list from the last use
3331  * of this root.
3332  */
3333 int btrfs_orphan_cleanup(struct btrfs_root *root)
3334 {
3335         struct btrfs_path *path;
3336         struct extent_buffer *leaf;
3337         struct btrfs_key key, found_key;
3338         struct btrfs_trans_handle *trans;
3339         struct inode *inode;
3340         u64 last_objectid = 0;
3341         int ret = 0, nr_unlink = 0, nr_truncate = 0;
3342
3343         if (cmpxchg(&root->orphan_cleanup_state, 0, ORPHAN_CLEANUP_STARTED))
3344                 return 0;
3345
3346         path = btrfs_alloc_path();
3347         if (!path) {
3348                 ret = -ENOMEM;
3349                 goto out;
3350         }
3351         path->reada = READA_BACK;
3352
3353         key.objectid = BTRFS_ORPHAN_OBJECTID;
3354         key.type = BTRFS_ORPHAN_ITEM_KEY;
3355         key.offset = (u64)-1;
3356
3357         while (1) {
3358                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3359                 if (ret < 0)
3360                         goto out;
3361
3362                 /*
3363                  * if ret == 0 means we found what we were searching for, which
3364                  * is weird, but possible, so only screw with path if we didn't
3365                  * find the key and see if we have stuff that matches
3366                  */
3367                 if (ret > 0) {
3368                         ret = 0;
3369                         if (path->slots[0] == 0)
3370                                 break;
3371                         path->slots[0]--;
3372                 }
3373
3374                 /* pull out the item */
3375                 leaf = path->nodes[0];
3376                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
3377
3378                 /* make sure the item matches what we want */
3379                 if (found_key.objectid != BTRFS_ORPHAN_OBJECTID)
3380                         break;
3381                 if (found_key.type != BTRFS_ORPHAN_ITEM_KEY)
3382                         break;
3383
3384                 /* release the path since we're done with it */
3385                 btrfs_release_path(path);
3386
3387                 /*
3388                  * this is where we are basically btrfs_lookup, without the
3389                  * crossing root thing.  we store the inode number in the
3390                  * offset of the orphan item.
3391                  */
3392
3393                 if (found_key.offset == last_objectid) {
3394                         btrfs_err(root->fs_info,
3395                                 "Error removing orphan entry, stopping orphan cleanup");
3396                         ret = -EINVAL;
3397                         goto out;
3398                 }
3399
3400                 last_objectid = found_key.offset;
3401
3402                 found_key.objectid = found_key.offset;
3403                 found_key.type = BTRFS_INODE_ITEM_KEY;
3404                 found_key.offset = 0;
3405                 inode = btrfs_iget(root->fs_info->sb, &found_key, root, NULL);
3406                 ret = PTR_ERR_OR_ZERO(inode);
3407                 if (ret && ret != -ESTALE)
3408                         goto out;
3409
3410                 if (ret == -ESTALE && root == root->fs_info->tree_root) {
3411                         struct btrfs_root *dead_root;
3412                         struct btrfs_fs_info *fs_info = root->fs_info;
3413                         int is_dead_root = 0;
3414
3415                         /*
3416                          * this is an orphan in the tree root. Currently these
3417                          * could come from 2 sources:
3418                          *  a) a snapshot deletion in progress
3419                          *  b) a free space cache inode
3420                          * We need to distinguish those two, as the snapshot
3421                          * orphan must not get deleted.
3422                          * find_dead_roots already ran before us, so if this
3423                          * is a snapshot deletion, we should find the root
3424                          * in the dead_roots list
3425                          */
3426                         spin_lock(&fs_info->trans_lock);
3427                         list_for_each_entry(dead_root, &fs_info->dead_roots,
3428                                             root_list) {
3429                                 if (dead_root->root_key.objectid ==
3430                                     found_key.objectid) {
3431                                         is_dead_root = 1;
3432                                         break;
3433                                 }
3434                         }
3435                         spin_unlock(&fs_info->trans_lock);
3436                         if (is_dead_root) {
3437                                 /* prevent this orphan from being found again */
3438                                 key.offset = found_key.objectid - 1;
3439                                 continue;
3440                         }
3441                 }
3442                 /*
3443                  * Inode is already gone but the orphan item is still there,
3444                  * kill the orphan item.
3445                  */
3446                 if (ret == -ESTALE) {
3447                         trans = btrfs_start_transaction(root, 1);
3448                         if (IS_ERR(trans)) {
3449                                 ret = PTR_ERR(trans);
3450                                 goto out;
3451                         }
3452                         btrfs_debug(root->fs_info, "auto deleting %Lu",
3453                                 found_key.objectid);
3454                         ret = btrfs_del_orphan_item(trans, root,
3455                                                     found_key.objectid);
3456                         btrfs_end_transaction(trans, root);
3457                         if (ret)
3458                                 goto out;
3459                         continue;
3460                 }
3461
3462                 /*
3463                  * add this inode to the orphan list so btrfs_orphan_del does
3464                  * the proper thing when we hit it
3465                  */
3466                 set_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
3467                         &BTRFS_I(inode)->runtime_flags);
3468                 atomic_inc(&root->orphan_inodes);
3469
3470                 /* if we have links, this was a truncate, lets do that */
3471                 if (inode->i_nlink) {
3472                         if (WARN_ON(!S_ISREG(inode->i_mode))) {
3473                                 iput(inode);
3474                                 continue;
3475                         }
3476                         nr_truncate++;
3477
3478                         /* 1 for the orphan item deletion. */
3479                         trans = btrfs_start_transaction(root, 1);
3480                         if (IS_ERR(trans)) {
3481                                 iput(inode);
3482                                 ret = PTR_ERR(trans);
3483                                 goto out;
3484                         }
3485                         ret = btrfs_orphan_add(trans, inode);
3486                         btrfs_end_transaction(trans, root);
3487                         if (ret) {
3488                                 iput(inode);
3489                                 goto out;
3490                         }
3491
3492                         ret = btrfs_truncate(inode);
3493                         if (ret)
3494                                 btrfs_orphan_del(NULL, inode);
3495                 } else {
3496                         nr_unlink++;
3497                 }
3498
3499                 /* this will do delete_inode and everything for us */
3500                 iput(inode);
3501                 if (ret)
3502                         goto out;
3503         }
3504         /* release the path since we're done with it */
3505         btrfs_release_path(path);
3506
3507         root->orphan_cleanup_state = ORPHAN_CLEANUP_DONE;
3508
3509         if (root->orphan_block_rsv)
3510                 btrfs_block_rsv_release(root, root->orphan_block_rsv,
3511                                         (u64)-1);
3512
3513         if (root->orphan_block_rsv ||
3514             test_bit(BTRFS_ROOT_ORPHAN_ITEM_INSERTED, &root->state)) {
3515                 trans = btrfs_join_transaction(root);
3516                 if (!IS_ERR(trans))
3517                         btrfs_end_transaction(trans, root);
3518         }
3519
3520         if (nr_unlink)
3521                 btrfs_debug(root->fs_info, "unlinked %d orphans", nr_unlink);
3522         if (nr_truncate)
3523                 btrfs_debug(root->fs_info, "truncated %d orphans", nr_truncate);
3524
3525 out:
3526         if (ret)
3527                 btrfs_err(root->fs_info,
3528                         "could not do orphan cleanup %d", ret);
3529         btrfs_free_path(path);
3530         return ret;
3531 }
3532
3533 /*
3534  * very simple check to peek ahead in the leaf looking for xattrs.  If we
3535  * don't find any xattrs, we know there can't be any acls.
3536  *
3537  * slot is the slot the inode is in, objectid is the objectid of the inode
3538  */
3539 static noinline int acls_after_inode_item(struct extent_buffer *leaf,
3540                                           int slot, u64 objectid,
3541                                           int *first_xattr_slot)
3542 {
3543         u32 nritems = btrfs_header_nritems(leaf);
3544         struct btrfs_key found_key;
3545         static u64 xattr_access = 0;
3546         static u64 xattr_default = 0;
3547         int scanned = 0;
3548
3549         if (!xattr_access) {
3550                 xattr_access = btrfs_name_hash(POSIX_ACL_XATTR_ACCESS,
3551                                         strlen(POSIX_ACL_XATTR_ACCESS));
3552                 xattr_default = btrfs_name_hash(POSIX_ACL_XATTR_DEFAULT,
3553                                         strlen(POSIX_ACL_XATTR_DEFAULT));
3554         }
3555
3556         slot++;
3557         *first_xattr_slot = -1;
3558         while (slot < nritems) {
3559                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
3560
3561                 /* we found a different objectid, there must not be acls */
3562                 if (found_key.objectid != objectid)
3563                         return 0;
3564
3565                 /* we found an xattr, assume we've got an acl */
3566                 if (found_key.type == BTRFS_XATTR_ITEM_KEY) {
3567                         if (*first_xattr_slot == -1)
3568                                 *first_xattr_slot = slot;
3569                         if (found_key.offset == xattr_access ||
3570                             found_key.offset == xattr_default)
3571                                 return 1;
3572                 }
3573
3574                 /*
3575                  * we found a key greater than an xattr key, there can't
3576                  * be any acls later on
3577                  */
3578                 if (found_key.type > BTRFS_XATTR_ITEM_KEY)
3579                         return 0;
3580
3581                 slot++;
3582                 scanned++;
3583
3584                 /*
3585                  * it goes inode, inode backrefs, xattrs, extents,
3586                  * so if there are a ton of hard links to an inode there can
3587                  * be a lot of backrefs.  Don't waste time searching too hard,
3588                  * this is just an optimization
3589                  */
3590                 if (scanned >= 8)
3591                         break;
3592         }
3593         /* we hit the end of the leaf before we found an xattr or
3594          * something larger than an xattr.  We have to assume the inode
3595          * has acls
3596          */
3597         if (*first_xattr_slot == -1)
3598                 *first_xattr_slot = slot;
3599         return 1;
3600 }
3601
3602 /*
3603  * read an inode from the btree into the in-memory inode
3604  */
3605 static void btrfs_read_locked_inode(struct inode *inode)
3606 {
3607         struct btrfs_path *path;
3608         struct extent_buffer *leaf;
3609         struct btrfs_inode_item *inode_item;
3610         struct btrfs_root *root = BTRFS_I(inode)->root;
3611         struct btrfs_key location;
3612         unsigned long ptr;
3613         int maybe_acls;
3614         u32 rdev;
3615         int ret;
3616         bool filled = false;
3617         int first_xattr_slot;
3618
3619         ret = btrfs_fill_inode(inode, &rdev);
3620         if (!ret)
3621                 filled = true;
3622
3623         path = btrfs_alloc_path();
3624         if (!path)
3625                 goto make_bad;
3626
3627         memcpy(&location, &BTRFS_I(inode)->location, sizeof(location));
3628
3629         ret = btrfs_lookup_inode(NULL, root, path, &location, 0);
3630         if (ret)
3631                 goto make_bad;
3632
3633         leaf = path->nodes[0];
3634
3635         if (filled)
3636                 goto cache_index;
3637
3638         inode_item = btrfs_item_ptr(leaf, path->slots[0],
3639                                     struct btrfs_inode_item);
3640         inode->i_mode = btrfs_inode_mode(leaf, inode_item);
3641         set_nlink(inode, btrfs_inode_nlink(leaf, inode_item));
3642         i_uid_write(inode, btrfs_inode_uid(leaf, inode_item));
3643         i_gid_write(inode, btrfs_inode_gid(leaf, inode_item));
3644         btrfs_i_size_write(inode, btrfs_inode_size(leaf, inode_item));
3645
3646         inode->i_atime.tv_sec = btrfs_timespec_sec(leaf, &inode_item->atime);
3647         inode->i_atime.tv_nsec = btrfs_timespec_nsec(leaf, &inode_item->atime);
3648
3649         inode->i_mtime.tv_sec = btrfs_timespec_sec(leaf, &inode_item->mtime);
3650         inode->i_mtime.tv_nsec = btrfs_timespec_nsec(leaf, &inode_item->mtime);
3651
3652         inode->i_ctime.tv_sec = btrfs_timespec_sec(leaf, &inode_item->ctime);
3653         inode->i_ctime.tv_nsec = btrfs_timespec_nsec(leaf, &inode_item->ctime);
3654
3655         BTRFS_I(inode)->i_otime.tv_sec =
3656                 btrfs_timespec_sec(leaf, &inode_item->otime);
3657         BTRFS_I(inode)->i_otime.tv_nsec =
3658                 btrfs_timespec_nsec(leaf, &inode_item->otime);
3659
3660         inode_set_bytes(inode, btrfs_inode_nbytes(leaf, inode_item));
3661         BTRFS_I(inode)->generation = btrfs_inode_generation(leaf, inode_item);
3662         BTRFS_I(inode)->last_trans = btrfs_inode_transid(leaf, inode_item);
3663
3664         inode->i_version = btrfs_inode_sequence(leaf, inode_item);
3665         inode->i_generation = BTRFS_I(inode)->generation;
3666         inode->i_rdev = 0;
3667         rdev = btrfs_inode_rdev(leaf, inode_item);
3668
3669         BTRFS_I(inode)->index_cnt = (u64)-1;
3670         BTRFS_I(inode)->flags = btrfs_inode_flags(leaf, inode_item);
3671
3672 cache_index:
3673         /*
3674          * If we were modified in the current generation and evicted from memory
3675          * and then re-read we need to do a full sync since we don't have any
3676          * idea about which extents were modified before we were evicted from
3677          * cache.
3678          *
3679          * This is required for both inode re-read from disk and delayed inode
3680          * in delayed_nodes_tree.
3681          */
3682         if (BTRFS_I(inode)->last_trans == root->fs_info->generation)
3683                 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
3684                         &BTRFS_I(inode)->runtime_flags);
3685
3686         /*
3687          * We don't persist the id of the transaction where an unlink operation
3688          * against the inode was last made. So here we assume the inode might
3689          * have been evicted, and therefore the exact value of last_unlink_trans
3690          * lost, and set it to last_trans to avoid metadata inconsistencies
3691          * between the inode and its parent if the inode is fsync'ed and the log
3692          * replayed. For example, in the scenario:
3693          *
3694          * touch mydir/foo
3695          * ln mydir/foo mydir/bar
3696          * sync
3697          * unlink mydir/bar
3698          * echo 2 > /proc/sys/vm/drop_caches   # evicts inode
3699          * xfs_io -c fsync mydir/foo
3700          * <power failure>
3701          * mount fs, triggers fsync log replay
3702          *
3703          * We must make sure that when we fsync our inode foo we also log its
3704          * parent inode, otherwise after log replay the parent still has the
3705          * dentry with the "bar" name but our inode foo has a link count of 1
3706          * and doesn't have an inode ref with the name "bar" anymore.
3707          *
3708          * Setting last_unlink_trans to last_trans is a pessimistic approach,
3709          * but it guarantees correctness at the expense of ocassional full
3710          * transaction commits on fsync if our inode is a directory, or if our
3711          * inode is not a directory, logging its parent unnecessarily.
3712          */
3713         BTRFS_I(inode)->last_unlink_trans = BTRFS_I(inode)->last_trans;
3714
3715         path->slots[0]++;
3716         if (inode->i_nlink != 1 ||
3717             path->slots[0] >= btrfs_header_nritems(leaf))
3718                 goto cache_acl;
3719
3720         btrfs_item_key_to_cpu(leaf, &location, path->slots[0]);
3721         if (location.objectid != btrfs_ino(inode))
3722                 goto cache_acl;
3723
3724         ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
3725         if (location.type == BTRFS_INODE_REF_KEY) {
3726                 struct btrfs_inode_ref *ref;
3727
3728                 ref = (struct btrfs_inode_ref *)ptr;
3729                 BTRFS_I(inode)->dir_index = btrfs_inode_ref_index(leaf, ref);
3730         } else if (location.type == BTRFS_INODE_EXTREF_KEY) {
3731                 struct btrfs_inode_extref *extref;
3732
3733                 extref = (struct btrfs_inode_extref *)ptr;
3734                 BTRFS_I(inode)->dir_index = btrfs_inode_extref_index(leaf,
3735                                                                      extref);
3736         }
3737 cache_acl:
3738         /*
3739          * try to precache a NULL acl entry for files that don't have
3740          * any xattrs or acls
3741          */
3742         maybe_acls = acls_after_inode_item(leaf, path->slots[0],
3743                                            btrfs_ino(inode), &first_xattr_slot);
3744         if (first_xattr_slot != -1) {
3745                 path->slots[0] = first_xattr_slot;
3746                 ret = btrfs_load_inode_props(inode, path);
3747                 if (ret)
3748                         btrfs_err(root->fs_info,
3749                                   "error loading props for ino %llu (root %llu): %d",
3750                                   btrfs_ino(inode),
3751                                   root->root_key.objectid, ret);
3752         }
3753         btrfs_free_path(path);
3754
3755         if (!maybe_acls)
3756                 cache_no_acl(inode);
3757
3758         switch (inode->i_mode & S_IFMT) {
3759         case S_IFREG:
3760                 inode->i_mapping->a_ops = &btrfs_aops;
3761                 BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
3762                 inode->i_fop = &btrfs_file_operations;
3763                 inode->i_op = &btrfs_file_inode_operations;
3764                 break;
3765         case S_IFDIR:
3766                 inode->i_fop = &btrfs_dir_file_operations;
3767                 if (root == root->fs_info->tree_root)
3768                         inode->i_op = &btrfs_dir_ro_inode_operations;
3769                 else
3770                         inode->i_op = &btrfs_dir_inode_operations;
3771                 break;
3772         case S_IFLNK:
3773                 inode->i_op = &btrfs_symlink_inode_operations;
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         bytes_deleted = btrfs_csum_bytes_to_leaves(root, bytes_deleted);
4215         ret = btrfs_block_rsv_add(root, &root->fs_info->trans_block_rsv,
4216                                   bytes_deleted, BTRFS_RESERVE_NO_FLUSH);
4217         if (!ret)
4218                 trans->bytes_reserved += bytes_deleted;
4219         return ret;
4220
4221 }
4222
4223 static int truncate_inline_extent(struct inode *inode,
4224                                   struct btrfs_path *path,
4225                                   struct btrfs_key *found_key,
4226                                   const u64 item_end,
4227                                   const u64 new_size)
4228 {
4229         struct extent_buffer *leaf = path->nodes[0];
4230         int slot = path->slots[0];
4231         struct btrfs_file_extent_item *fi;
4232         u32 size = (u32)(new_size - found_key->offset);
4233         struct btrfs_root *root = BTRFS_I(inode)->root;
4234
4235         fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
4236
4237         if (btrfs_file_extent_compression(leaf, fi) != BTRFS_COMPRESS_NONE) {
4238                 loff_t offset = new_size;
4239                 loff_t page_end = ALIGN(offset, PAGE_CACHE_SIZE);
4240
4241                 /*
4242                  * Zero out the remaining of the last page of our inline extent,
4243                  * instead of directly truncating our inline extent here - that
4244                  * would be much more complex (decompressing all the data, then
4245                  * compressing the truncated data, which might be bigger than
4246                  * the size of the inline extent, resize the extent, etc).
4247                  * We release the path because to get the page we might need to
4248                  * read the extent item from disk (data not in the page cache).
4249                  */
4250                 btrfs_release_path(path);
4251                 return btrfs_truncate_block(inode, offset, page_end - offset,
4252                                         0);
4253         }
4254
4255         btrfs_set_file_extent_ram_bytes(leaf, fi, size);
4256         size = btrfs_file_extent_calc_inline_size(size);
4257         btrfs_truncate_item(root, path, size, 1);
4258
4259         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state))
4260                 inode_sub_bytes(inode, item_end + 1 - new_size);
4261
4262         return 0;
4263 }
4264
4265 /*
4266  * this can truncate away extent items, csum items and directory items.
4267  * It starts at a high offset and removes keys until it can't find
4268  * any higher than new_size
4269  *
4270  * csum items that cross the new i_size are truncated to the new size
4271  * as well.
4272  *
4273  * min_type is the minimum key type to truncate down to.  If set to 0, this
4274  * will kill all the items on this inode, including the INODE_ITEM_KEY.
4275  */
4276 int btrfs_truncate_inode_items(struct btrfs_trans_handle *trans,
4277                                struct btrfs_root *root,
4278                                struct inode *inode,
4279                                u64 new_size, u32 min_type)
4280 {
4281         struct btrfs_path *path;
4282         struct extent_buffer *leaf;
4283         struct btrfs_file_extent_item *fi;
4284         struct btrfs_key key;
4285         struct btrfs_key found_key;
4286         u64 extent_start = 0;
4287         u64 extent_num_bytes = 0;
4288         u64 extent_offset = 0;
4289         u64 item_end = 0;
4290         u64 last_size = new_size;
4291         u32 found_type = (u8)-1;
4292         int found_extent;
4293         int del_item;
4294         int pending_del_nr = 0;
4295         int pending_del_slot = 0;
4296         int extent_type = -1;
4297         int ret;
4298         int err = 0;
4299         u64 ino = btrfs_ino(inode);
4300         u64 bytes_deleted = 0;
4301         bool be_nice = 0;
4302         bool should_throttle = 0;
4303         bool should_end = 0;
4304
4305         BUG_ON(new_size > 0 && min_type != BTRFS_EXTENT_DATA_KEY);
4306
4307         /*
4308          * for non-free space inodes and ref cows, we want to back off from
4309          * time to time
4310          */
4311         if (!btrfs_is_free_space_inode(inode) &&
4312             test_bit(BTRFS_ROOT_REF_COWS, &root->state))
4313                 be_nice = 1;
4314
4315         path = btrfs_alloc_path();
4316         if (!path)
4317                 return -ENOMEM;
4318         path->reada = READA_BACK;
4319
4320         /*
4321          * We want to drop from the next block forward in case this new size is
4322          * not block aligned since we will be keeping the last block of the
4323          * extent just the way it is.
4324          */
4325         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
4326             root == root->fs_info->tree_root)
4327                 btrfs_drop_extent_cache(inode, ALIGN(new_size,
4328                                         root->sectorsize), (u64)-1, 0);
4329
4330         /*
4331          * This function is also used to drop the items in the log tree before
4332          * we relog the inode, so if root != BTRFS_I(inode)->root, it means
4333          * it is used to drop the loged items. So we shouldn't kill the delayed
4334          * items.
4335          */
4336         if (min_type == 0 && root == BTRFS_I(inode)->root)
4337                 btrfs_kill_delayed_inode_items(inode);
4338
4339         key.objectid = ino;
4340         key.offset = (u64)-1;
4341         key.type = (u8)-1;
4342
4343 search_again:
4344         /*
4345          * with a 16K leaf size and 128MB extents, you can actually queue
4346          * up a huge file in a single leaf.  Most of the time that
4347          * bytes_deleted is > 0, it will be huge by the time we get here
4348          */
4349         if (be_nice && bytes_deleted > SZ_32M) {
4350                 if (btrfs_should_end_transaction(trans, root)) {
4351                         err = -EAGAIN;
4352                         goto error;
4353                 }
4354         }
4355
4356
4357         path->leave_spinning = 1;
4358         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
4359         if (ret < 0) {
4360                 err = ret;
4361                 goto out;
4362         }
4363
4364         if (ret > 0) {
4365                 /* there are no items in the tree for us to truncate, we're
4366                  * done
4367                  */
4368                 if (path->slots[0] == 0)
4369                         goto out;
4370                 path->slots[0]--;
4371         }
4372
4373         while (1) {
4374                 fi = NULL;
4375                 leaf = path->nodes[0];
4376                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
4377                 found_type = found_key.type;
4378
4379                 if (found_key.objectid != ino)
4380                         break;
4381
4382                 if (found_type < min_type)
4383                         break;
4384
4385                 item_end = found_key.offset;
4386                 if (found_type == BTRFS_EXTENT_DATA_KEY) {
4387                         fi = btrfs_item_ptr(leaf, path->slots[0],
4388                                             struct btrfs_file_extent_item);
4389                         extent_type = btrfs_file_extent_type(leaf, fi);
4390                         if (extent_type != BTRFS_FILE_EXTENT_INLINE) {
4391                                 item_end +=
4392                                     btrfs_file_extent_num_bytes(leaf, fi);
4393                         } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
4394                                 item_end += btrfs_file_extent_inline_len(leaf,
4395                                                          path->slots[0], fi);
4396                         }
4397                         item_end--;
4398                 }
4399                 if (found_type > min_type) {
4400                         del_item = 1;
4401                 } else {
4402                         if (item_end < new_size)
4403                                 break;
4404                         if (found_key.offset >= new_size)
4405                                 del_item = 1;
4406                         else
4407                                 del_item = 0;
4408                 }
4409                 found_extent = 0;
4410                 /* FIXME, shrink the extent if the ref count is only 1 */
4411                 if (found_type != BTRFS_EXTENT_DATA_KEY)
4412                         goto delete;
4413
4414                 if (del_item)
4415                         last_size = found_key.offset;
4416                 else
4417                         last_size = new_size;
4418
4419                 if (extent_type != BTRFS_FILE_EXTENT_INLINE) {
4420                         u64 num_dec;
4421                         extent_start = btrfs_file_extent_disk_bytenr(leaf, fi);
4422                         if (!del_item) {
4423                                 u64 orig_num_bytes =
4424                                         btrfs_file_extent_num_bytes(leaf, fi);
4425                                 extent_num_bytes = ALIGN(new_size -
4426                                                 found_key.offset,
4427                                                 root->sectorsize);
4428                                 btrfs_set_file_extent_num_bytes(leaf, fi,
4429                                                          extent_num_bytes);
4430                                 num_dec = (orig_num_bytes -
4431                                            extent_num_bytes);
4432                                 if (test_bit(BTRFS_ROOT_REF_COWS,
4433                                              &root->state) &&
4434                                     extent_start != 0)
4435                                         inode_sub_bytes(inode, num_dec);
4436                                 btrfs_mark_buffer_dirty(leaf);
4437                         } else {
4438                                 extent_num_bytes =
4439                                         btrfs_file_extent_disk_num_bytes(leaf,
4440                                                                          fi);
4441                                 extent_offset = found_key.offset -
4442                                         btrfs_file_extent_offset(leaf, fi);
4443
4444                                 /* FIXME blocksize != 4096 */
4445                                 num_dec = btrfs_file_extent_num_bytes(leaf, fi);
4446                                 if (extent_start != 0) {
4447                                         found_extent = 1;
4448                                         if (test_bit(BTRFS_ROOT_REF_COWS,
4449                                                      &root->state))
4450                                                 inode_sub_bytes(inode, num_dec);
4451                                 }
4452                         }
4453                 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
4454                         /*
4455                          * we can't truncate inline items that have had
4456                          * special encodings
4457                          */
4458                         if (!del_item &&
4459                             btrfs_file_extent_encryption(leaf, fi) == 0 &&
4460                             btrfs_file_extent_other_encoding(leaf, fi) == 0) {
4461
4462                                 /*
4463                                  * Need to release path in order to truncate a
4464                                  * compressed extent. So delete any accumulated
4465                                  * extent items so far.
4466                                  */
4467                                 if (btrfs_file_extent_compression(leaf, fi) !=
4468                                     BTRFS_COMPRESS_NONE && pending_del_nr) {
4469                                         err = btrfs_del_items(trans, root, path,
4470                                                               pending_del_slot,
4471                                                               pending_del_nr);
4472                                         if (err) {
4473                                                 btrfs_abort_transaction(trans,
4474                                                                         root,
4475                                                                         err);
4476                                                 goto error;
4477                                         }
4478                                         pending_del_nr = 0;
4479                                 }
4480
4481                                 err = truncate_inline_extent(inode, path,
4482                                                              &found_key,
4483                                                              item_end,
4484                                                              new_size);
4485                                 if (err) {
4486                                         btrfs_abort_transaction(trans,
4487                                                                 root, err);
4488                                         goto error;
4489                                 }
4490                         } else if (test_bit(BTRFS_ROOT_REF_COWS,
4491                                             &root->state)) {
4492                                 inode_sub_bytes(inode, item_end + 1 - new_size);
4493                         }
4494                 }
4495 delete:
4496                 if (del_item) {
4497                         if (!pending_del_nr) {
4498                                 /* no pending yet, add ourselves */
4499                                 pending_del_slot = path->slots[0];
4500                                 pending_del_nr = 1;
4501                         } else if (pending_del_nr &&
4502                                    path->slots[0] + 1 == pending_del_slot) {
4503                                 /* hop on the pending chunk */
4504                                 pending_del_nr++;
4505                                 pending_del_slot = path->slots[0];
4506                         } else {
4507                                 BUG();
4508                         }
4509                 } else {
4510                         break;
4511                 }
4512                 should_throttle = 0;
4513
4514                 if (found_extent &&
4515                     (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
4516                      root == root->fs_info->tree_root)) {
4517                         btrfs_set_path_blocking(path);
4518                         bytes_deleted += extent_num_bytes;
4519                         ret = btrfs_free_extent(trans, root, extent_start,
4520                                                 extent_num_bytes, 0,
4521                                                 btrfs_header_owner(leaf),
4522                                                 ino, extent_offset);
4523                         BUG_ON(ret);
4524                         if (btrfs_should_throttle_delayed_refs(trans, root))
4525                                 btrfs_async_run_delayed_refs(root,
4526                                         trans->delayed_ref_updates * 2, 0);
4527                         if (be_nice) {
4528                                 if (truncate_space_check(trans, root,
4529                                                          extent_num_bytes)) {
4530                                         should_end = 1;
4531                                 }
4532                                 if (btrfs_should_throttle_delayed_refs(trans,
4533                                                                        root)) {
4534                                         should_throttle = 1;
4535                                 }
4536                         }
4537                 }
4538
4539                 if (found_type == BTRFS_INODE_ITEM_KEY)
4540                         break;
4541
4542                 if (path->slots[0] == 0 ||
4543                     path->slots[0] != pending_del_slot ||
4544                     should_throttle || should_end) {
4545                         if (pending_del_nr) {
4546                                 ret = btrfs_del_items(trans, root, path,
4547                                                 pending_del_slot,
4548                                                 pending_del_nr);
4549                                 if (ret) {
4550                                         btrfs_abort_transaction(trans,
4551                                                                 root, ret);
4552                                         goto error;
4553                                 }
4554                                 pending_del_nr = 0;
4555                         }
4556                         btrfs_release_path(path);
4557                         if (should_throttle) {
4558                                 unsigned long updates = trans->delayed_ref_updates;
4559                                 if (updates) {
4560                                         trans->delayed_ref_updates = 0;
4561                                         ret = btrfs_run_delayed_refs(trans, root, updates * 2);
4562                                         if (ret && !err)
4563                                                 err = ret;
4564                                 }
4565                         }
4566                         /*
4567                          * if we failed to refill our space rsv, bail out
4568                          * and let the transaction restart
4569                          */
4570                         if (should_end) {
4571                                 err = -EAGAIN;
4572                                 goto error;
4573                         }
4574                         goto search_again;
4575                 } else {
4576                         path->slots[0]--;
4577                 }
4578         }
4579 out:
4580         if (pending_del_nr) {
4581                 ret = btrfs_del_items(trans, root, path, pending_del_slot,
4582                                       pending_del_nr);
4583                 if (ret)
4584                         btrfs_abort_transaction(trans, root, ret);
4585         }
4586 error:
4587         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID)
4588                 btrfs_ordered_update_i_size(inode, last_size, NULL);
4589
4590         btrfs_free_path(path);
4591
4592         if (be_nice && bytes_deleted > SZ_32M) {
4593                 unsigned long updates = trans->delayed_ref_updates;
4594                 if (updates) {
4595                         trans->delayed_ref_updates = 0;
4596                         ret = btrfs_run_delayed_refs(trans, root, updates * 2);
4597                         if (ret && !err)
4598                                 err = ret;
4599                 }
4600         }
4601         return err;
4602 }
4603
4604 /*
4605  * btrfs_truncate_block - read, zero a chunk and write a block
4606  * @inode - inode that we're zeroing
4607  * @from - the offset to start zeroing
4608  * @len - the length to zero, 0 to zero the entire range respective to the
4609  *      offset
4610  * @front - zero up to the offset instead of from the offset on
4611  *
4612  * This will find the block for the "from" offset and cow the block and zero the
4613  * part we want to zero.  This is used with truncate and hole punching.
4614  */
4615 int btrfs_truncate_block(struct inode *inode, loff_t from, loff_t len,
4616                         int front)
4617 {
4618         struct address_space *mapping = inode->i_mapping;
4619         struct btrfs_root *root = BTRFS_I(inode)->root;
4620         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
4621         struct btrfs_ordered_extent *ordered;
4622         struct extent_state *cached_state = NULL;
4623         char *kaddr;
4624         u32 blocksize = root->sectorsize;
4625         pgoff_t index = from >> PAGE_CACHE_SHIFT;
4626         unsigned offset = from & (blocksize - 1);
4627         struct page *page;
4628         gfp_t mask = btrfs_alloc_write_mask(mapping);
4629         int ret = 0;
4630         u64 block_start;
4631         u64 block_end;
4632
4633         if ((offset & (blocksize - 1)) == 0 &&
4634             (!len || ((len & (blocksize - 1)) == 0)))
4635                 goto out;
4636
4637         ret = btrfs_delalloc_reserve_space(inode,
4638                         round_down(from, blocksize), blocksize);
4639         if (ret)
4640                 goto out;
4641
4642 again:
4643         page = find_or_create_page(mapping, index, mask);
4644         if (!page) {
4645                 btrfs_delalloc_release_space(inode,
4646                                 round_down(from, blocksize),
4647                                 blocksize);
4648                 ret = -ENOMEM;
4649                 goto out;
4650         }
4651
4652         block_start = round_down(from, blocksize);
4653         block_end = block_start + blocksize - 1;
4654
4655         if (!PageUptodate(page)) {
4656                 ret = btrfs_readpage(NULL, page);
4657                 lock_page(page);
4658                 if (page->mapping != mapping) {
4659                         unlock_page(page);
4660                         page_cache_release(page);
4661                         goto again;
4662                 }
4663                 if (!PageUptodate(page)) {
4664                         ret = -EIO;
4665                         goto out_unlock;
4666                 }
4667         }
4668         wait_on_page_writeback(page);
4669
4670         lock_extent_bits(io_tree, block_start, block_end, &cached_state);
4671         set_page_extent_mapped(page);
4672
4673         ordered = btrfs_lookup_ordered_extent(inode, block_start);
4674         if (ordered) {
4675                 unlock_extent_cached(io_tree, block_start, block_end,
4676                                      &cached_state, GFP_NOFS);
4677                 unlock_page(page);
4678                 page_cache_release(page);
4679                 btrfs_start_ordered_extent(inode, ordered, 1);
4680                 btrfs_put_ordered_extent(ordered);
4681                 goto again;
4682         }
4683
4684         clear_extent_bit(&BTRFS_I(inode)->io_tree, block_start, block_end,
4685                           EXTENT_DIRTY | EXTENT_DELALLOC |
4686                           EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG,
4687                           0, 0, &cached_state, GFP_NOFS);
4688
4689         ret = btrfs_set_extent_delalloc(inode, block_start, block_end,
4690                                         &cached_state);
4691         if (ret) {
4692                 unlock_extent_cached(io_tree, block_start, block_end,
4693                                      &cached_state, GFP_NOFS);
4694                 goto out_unlock;
4695         }
4696
4697         if (offset != blocksize) {
4698                 if (!len)
4699                         len = blocksize - offset;
4700                 kaddr = kmap(page);
4701                 if (front)
4702                         memset(kaddr + (block_start - page_offset(page)),
4703                                 0, offset);
4704                 else
4705                         memset(kaddr + (block_start - page_offset(page)) +  offset,
4706                                 0, len);
4707                 flush_dcache_page(page);
4708                 kunmap(page);
4709         }
4710         ClearPageChecked(page);
4711         set_page_dirty(page);
4712         unlock_extent_cached(io_tree, block_start, block_end, &cached_state,
4713                              GFP_NOFS);
4714
4715 out_unlock:
4716         if (ret)
4717                 btrfs_delalloc_release_space(inode, block_start,
4718                                              blocksize);
4719         unlock_page(page);
4720         page_cache_release(page);
4721 out:
4722         return ret;
4723 }
4724
4725 static int maybe_insert_hole(struct btrfs_root *root, struct inode *inode,
4726                              u64 offset, u64 len)
4727 {
4728         struct btrfs_trans_handle *trans;
4729         int ret;
4730
4731         /*
4732          * Still need to make sure the inode looks like it's been updated so
4733          * that any holes get logged if we fsync.
4734          */
4735         if (btrfs_fs_incompat(root->fs_info, NO_HOLES)) {
4736                 BTRFS_I(inode)->last_trans = root->fs_info->generation;
4737                 BTRFS_I(inode)->last_sub_trans = root->log_transid;
4738                 BTRFS_I(inode)->last_log_commit = root->last_log_commit;
4739                 return 0;
4740         }
4741
4742         /*
4743          * 1 - for the one we're dropping
4744          * 1 - for the one we're adding
4745          * 1 - for updating the inode.
4746          */
4747         trans = btrfs_start_transaction(root, 3);
4748         if (IS_ERR(trans))
4749                 return PTR_ERR(trans);
4750
4751         ret = btrfs_drop_extents(trans, root, inode, offset, offset + len, 1);
4752         if (ret) {
4753                 btrfs_abort_transaction(trans, root, ret);
4754                 btrfs_end_transaction(trans, root);
4755                 return ret;
4756         }
4757
4758         ret = btrfs_insert_file_extent(trans, root, btrfs_ino(inode), offset,
4759                                        0, 0, len, 0, len, 0, 0, 0);
4760         if (ret)
4761                 btrfs_abort_transaction(trans, root, ret);
4762         else
4763                 btrfs_update_inode(trans, root, inode);
4764         btrfs_end_transaction(trans, root);
4765         return ret;
4766 }
4767
4768 /*
4769  * This function puts in dummy file extents for the area we're creating a hole
4770  * for.  So if we are truncating this file to a larger size we need to insert
4771  * these file extents so that btrfs_get_extent will return a EXTENT_MAP_HOLE for
4772  * the range between oldsize and size
4773  */
4774 int btrfs_cont_expand(struct inode *inode, loff_t oldsize, loff_t size)
4775 {
4776         struct btrfs_root *root = BTRFS_I(inode)->root;
4777         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
4778         struct extent_map *em = NULL;
4779         struct extent_state *cached_state = NULL;
4780         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
4781         u64 hole_start = ALIGN(oldsize, root->sectorsize);
4782         u64 block_end = ALIGN(size, root->sectorsize);
4783         u64 last_byte;
4784         u64 cur_offset;
4785         u64 hole_size;
4786         int err = 0;
4787
4788         /*
4789          * If our size started in the middle of a block we need to zero out the
4790          * rest of the block before we expand the i_size, otherwise we could
4791          * expose stale data.
4792          */
4793         err = btrfs_truncate_block(inode, oldsize, 0, 0);
4794         if (err)
4795                 return err;
4796
4797         if (size <= hole_start)
4798                 return 0;
4799
4800         while (1) {
4801                 struct btrfs_ordered_extent *ordered;
4802
4803                 lock_extent_bits(io_tree, hole_start, block_end - 1,
4804                                  &cached_state);
4805                 ordered = btrfs_lookup_ordered_range(inode, hole_start,
4806                                                      block_end - hole_start);
4807                 if (!ordered)
4808                         break;
4809                 unlock_extent_cached(io_tree, hole_start, block_end - 1,
4810                                      &cached_state, GFP_NOFS);
4811                 btrfs_start_ordered_extent(inode, ordered, 1);
4812                 btrfs_put_ordered_extent(ordered);
4813         }
4814
4815         cur_offset = hole_start;
4816         while (1) {
4817                 em = btrfs_get_extent(inode, NULL, 0, cur_offset,
4818                                 block_end - cur_offset, 0);
4819                 if (IS_ERR(em)) {
4820                         err = PTR_ERR(em);
4821                         em = NULL;
4822                         break;
4823                 }
4824                 last_byte = min(extent_map_end(em), block_end);
4825                 last_byte = ALIGN(last_byte , root->sectorsize);
4826                 if (!test_bit(EXTENT_FLAG_PREALLOC, &em->flags)) {
4827                         struct extent_map *hole_em;
4828                         hole_size = last_byte - cur_offset;
4829
4830                         err = maybe_insert_hole(root, inode, cur_offset,
4831                                                 hole_size);
4832                         if (err)
4833                                 break;
4834                         btrfs_drop_extent_cache(inode, cur_offset,
4835                                                 cur_offset + hole_size - 1, 0);
4836                         hole_em = alloc_extent_map();
4837                         if (!hole_em) {
4838                                 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
4839                                         &BTRFS_I(inode)->runtime_flags);
4840                                 goto next;
4841                         }
4842                         hole_em->start = cur_offset;
4843                         hole_em->len = hole_size;
4844                         hole_em->orig_start = cur_offset;
4845
4846                         hole_em->block_start = EXTENT_MAP_HOLE;
4847                         hole_em->block_len = 0;
4848                         hole_em->orig_block_len = 0;
4849                         hole_em->ram_bytes = hole_size;
4850                         hole_em->bdev = root->fs_info->fs_devices->latest_bdev;
4851                         hole_em->compress_type = BTRFS_COMPRESS_NONE;
4852                         hole_em->generation = root->fs_info->generation;
4853
4854                         while (1) {
4855                                 write_lock(&em_tree->lock);
4856                                 err = add_extent_mapping(em_tree, hole_em, 1);
4857                                 write_unlock(&em_tree->lock);
4858                                 if (err != -EEXIST)
4859                                         break;
4860                                 btrfs_drop_extent_cache(inode, cur_offset,
4861                                                         cur_offset +
4862                                                         hole_size - 1, 0);
4863                         }
4864                         free_extent_map(hole_em);
4865                 }
4866 next:
4867                 free_extent_map(em);
4868                 em = NULL;
4869                 cur_offset = last_byte;
4870                 if (cur_offset >= block_end)
4871                         break;
4872         }
4873         free_extent_map(em);
4874         unlock_extent_cached(io_tree, hole_start, block_end - 1, &cached_state,
4875                              GFP_NOFS);
4876         return err;
4877 }
4878
4879 static int btrfs_setsize(struct inode *inode, struct iattr *attr)
4880 {
4881         struct btrfs_root *root = BTRFS_I(inode)->root;
4882         struct btrfs_trans_handle *trans;
4883         loff_t oldsize = i_size_read(inode);
4884         loff_t newsize = attr->ia_size;
4885         int mask = attr->ia_valid;
4886         int ret;
4887
4888         /*
4889          * The regular truncate() case without ATTR_CTIME and ATTR_MTIME is a
4890          * special case where we need to update the times despite not having
4891          * these flags set.  For all other operations the VFS set these flags
4892          * explicitly if it wants a timestamp update.
4893          */
4894         if (newsize != oldsize) {
4895                 inode_inc_iversion(inode);
4896                 if (!(mask & (ATTR_CTIME | ATTR_MTIME)))
4897                         inode->i_ctime = inode->i_mtime =
4898                                 current_fs_time(inode->i_sb);
4899         }
4900
4901         if (newsize > oldsize) {
4902                 truncate_pagecache(inode, newsize);
4903                 /*
4904                  * Don't do an expanding truncate while snapshoting is ongoing.
4905                  * This is to ensure the snapshot captures a fully consistent
4906                  * state of this file - if the snapshot captures this expanding
4907                  * truncation, it must capture all writes that happened before
4908                  * this truncation.
4909                  */
4910                 btrfs_wait_for_snapshot_creation(root);
4911                 ret = btrfs_cont_expand(inode, oldsize, newsize);
4912                 if (ret) {
4913                         btrfs_end_write_no_snapshoting(root);
4914                         return ret;
4915                 }
4916
4917                 trans = btrfs_start_transaction(root, 1);
4918                 if (IS_ERR(trans)) {
4919                         btrfs_end_write_no_snapshoting(root);
4920                         return PTR_ERR(trans);
4921                 }
4922
4923                 i_size_write(inode, newsize);
4924                 btrfs_ordered_update_i_size(inode, i_size_read(inode), NULL);
4925                 ret = btrfs_update_inode(trans, root, inode);
4926                 btrfs_end_write_no_snapshoting(root);
4927                 btrfs_end_transaction(trans, root);
4928         } else {
4929
4930                 /*
4931                  * We're truncating a file that used to have good data down to
4932                  * zero. Make sure it gets into the ordered flush list so that
4933                  * any new writes get down to disk quickly.
4934                  */
4935                 if (newsize == 0)
4936                         set_bit(BTRFS_INODE_ORDERED_DATA_CLOSE,
4937                                 &BTRFS_I(inode)->runtime_flags);
4938
4939                 /*
4940                  * 1 for the orphan item we're going to add
4941                  * 1 for the orphan item deletion.
4942                  */
4943                 trans = btrfs_start_transaction(root, 2);
4944                 if (IS_ERR(trans))
4945                         return PTR_ERR(trans);
4946
4947                 /*
4948                  * We need to do this in case we fail at _any_ point during the
4949                  * actual truncate.  Once we do the truncate_setsize we could
4950                  * invalidate pages which forces any outstanding ordered io to
4951                  * be instantly completed which will give us extents that need
4952                  * to be truncated.  If we fail to get an orphan inode down we
4953                  * could have left over extents that were never meant to live,
4954                  * so we need to garuntee from this point on that everything
4955                  * will be consistent.
4956                  */
4957                 ret = btrfs_orphan_add(trans, inode);
4958                 btrfs_end_transaction(trans, root);
4959                 if (ret)
4960                         return ret;
4961
4962                 /* we don't support swapfiles, so vmtruncate shouldn't fail */
4963                 truncate_setsize(inode, newsize);
4964
4965                 /* Disable nonlocked read DIO to avoid the end less truncate */
4966                 btrfs_inode_block_unlocked_dio(inode);
4967                 inode_dio_wait(inode);
4968                 btrfs_inode_resume_unlocked_dio(inode);
4969
4970                 ret = btrfs_truncate(inode);
4971                 if (ret && inode->i_nlink) {
4972                         int err;
4973
4974                         /*
4975                          * failed to truncate, disk_i_size is only adjusted down
4976                          * as we remove extents, so it should represent the true
4977                          * size of the inode, so reset the in memory size and
4978                          * delete our orphan entry.
4979                          */
4980                         trans = btrfs_join_transaction(root);
4981                         if (IS_ERR(trans)) {
4982                                 btrfs_orphan_del(NULL, inode);
4983                                 return ret;
4984                         }
4985                         i_size_write(inode, BTRFS_I(inode)->disk_i_size);
4986                         err = btrfs_orphan_del(trans, inode);
4987                         if (err)
4988                                 btrfs_abort_transaction(trans, root, err);
4989                         btrfs_end_transaction(trans, root);
4990                 }
4991         }
4992
4993         return ret;
4994 }
4995
4996 static int btrfs_setattr(struct dentry *dentry, struct iattr *attr)
4997 {
4998         struct inode *inode = d_inode(dentry);
4999         struct btrfs_root *root = BTRFS_I(inode)->root;
5000         int err;
5001
5002         if (btrfs_root_readonly(root))
5003                 return -EROFS;
5004
5005         err = inode_change_ok(inode, attr);
5006         if (err)
5007                 return err;
5008
5009         if (S_ISREG(inode->i_mode) && (attr->ia_valid & ATTR_SIZE)) {
5010                 err = btrfs_setsize(inode, attr);
5011                 if (err)
5012                         return err;
5013         }
5014
5015         if (attr->ia_valid) {
5016                 setattr_copy(inode, attr);
5017                 inode_inc_iversion(inode);
5018                 err = btrfs_dirty_inode(inode);
5019
5020                 if (!err && attr->ia_valid & ATTR_MODE)
5021                         err = posix_acl_chmod(inode, inode->i_mode);
5022         }
5023
5024         return err;
5025 }
5026
5027 /*
5028  * While truncating the inode pages during eviction, we get the VFS calling
5029  * btrfs_invalidatepage() against each page of the inode. This is slow because
5030  * the calls to btrfs_invalidatepage() result in a huge amount of calls to
5031  * lock_extent_bits() and clear_extent_bit(), which keep merging and splitting
5032  * extent_state structures over and over, wasting lots of time.
5033  *
5034  * Therefore if the inode is being evicted, let btrfs_invalidatepage() skip all
5035  * those expensive operations on a per page basis and do only the ordered io
5036  * finishing, while we release here the extent_map and extent_state structures,
5037  * without the excessive merging and splitting.
5038  */
5039 static void evict_inode_truncate_pages(struct inode *inode)
5040 {
5041         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
5042         struct extent_map_tree *map_tree = &BTRFS_I(inode)->extent_tree;
5043         struct rb_node *node;
5044
5045         ASSERT(inode->i_state & I_FREEING);
5046         truncate_inode_pages_final(&inode->i_data);
5047
5048         write_lock(&map_tree->lock);
5049         while (!RB_EMPTY_ROOT(&map_tree->map)) {
5050                 struct extent_map *em;
5051
5052                 node = rb_first(&map_tree->map);
5053                 em = rb_entry(node, struct extent_map, rb_node);
5054                 clear_bit(EXTENT_FLAG_PINNED, &em->flags);
5055                 clear_bit(EXTENT_FLAG_LOGGING, &em->flags);
5056                 remove_extent_mapping(map_tree, em);
5057                 free_extent_map(em);
5058                 if (need_resched()) {
5059                         write_unlock(&map_tree->lock);
5060                         cond_resched();
5061                         write_lock(&map_tree->lock);
5062                 }
5063         }
5064         write_unlock(&map_tree->lock);
5065
5066         /*
5067          * Keep looping until we have no more ranges in the io tree.
5068          * We can have ongoing bios started by readpages (called from readahead)
5069          * that have their endio callback (extent_io.c:end_bio_extent_readpage)
5070          * still in progress (unlocked the pages in the bio but did not yet
5071          * unlocked the ranges in the io tree). Therefore this means some
5072          * ranges can still be locked and eviction started because before
5073          * submitting those bios, which are executed by a separate task (work
5074          * queue kthread), inode references (inode->i_count) were not taken
5075          * (which would be dropped in the end io callback of each bio).
5076          * Therefore here we effectively end up waiting for those bios and
5077          * anyone else holding locked ranges without having bumped the inode's
5078          * reference count - if we don't do it, when they access the inode's
5079          * io_tree to unlock a range it may be too late, leading to an
5080          * use-after-free issue.
5081          */
5082         spin_lock(&io_tree->lock);
5083         while (!RB_EMPTY_ROOT(&io_tree->state)) {
5084                 struct extent_state *state;
5085                 struct extent_state *cached_state = NULL;
5086                 u64 start;
5087                 u64 end;
5088
5089                 node = rb_first(&io_tree->state);
5090                 state = rb_entry(node, struct extent_state, rb_node);
5091                 start = state->start;
5092                 end = state->end;
5093                 spin_unlock(&io_tree->lock);
5094
5095                 lock_extent_bits(io_tree, start, end, &cached_state);
5096
5097                 /*
5098                  * If still has DELALLOC flag, the extent didn't reach disk,
5099                  * and its reserved space won't be freed by delayed_ref.
5100                  * So we need to free its reserved space here.
5101                  * (Refer to comment in btrfs_invalidatepage, case 2)
5102                  *
5103                  * Note, end is the bytenr of last byte, so we need + 1 here.
5104                  */
5105                 if (state->state & EXTENT_DELALLOC)
5106                         btrfs_qgroup_free_data(inode, start, end - start + 1);
5107
5108                 clear_extent_bit(io_tree, start, end,
5109                                  EXTENT_LOCKED | EXTENT_DIRTY |
5110                                  EXTENT_DELALLOC | EXTENT_DO_ACCOUNTING |
5111                                  EXTENT_DEFRAG, 1, 1,
5112                                  &cached_state, GFP_NOFS);
5113
5114                 cond_resched();
5115                 spin_lock(&io_tree->lock);
5116         }
5117         spin_unlock(&io_tree->lock);
5118 }
5119
5120 void btrfs_evict_inode(struct inode *inode)
5121 {
5122         struct btrfs_trans_handle *trans;
5123         struct btrfs_root *root = BTRFS_I(inode)->root;
5124         struct btrfs_block_rsv *rsv, *global_rsv;
5125         int steal_from_global = 0;
5126         u64 min_size = btrfs_calc_trunc_metadata_size(root, 1);
5127         int ret;
5128
5129         trace_btrfs_inode_evict(inode);
5130
5131         evict_inode_truncate_pages(inode);
5132
5133         if (inode->i_nlink &&
5134             ((btrfs_root_refs(&root->root_item) != 0 &&
5135               root->root_key.objectid != BTRFS_ROOT_TREE_OBJECTID) ||
5136              btrfs_is_free_space_inode(inode)))
5137                 goto no_delete;
5138
5139         if (is_bad_inode(inode)) {
5140                 btrfs_orphan_del(NULL, inode);
5141                 goto no_delete;
5142         }
5143         /* do we really want it for ->i_nlink > 0 and zero btrfs_root_refs? */
5144         if (!special_file(inode->i_mode))
5145                 btrfs_wait_ordered_range(inode, 0, (u64)-1);
5146
5147         btrfs_free_io_failure_record(inode, 0, (u64)-1);
5148
5149         if (root->fs_info->log_root_recovering) {
5150                 BUG_ON(test_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
5151                                  &BTRFS_I(inode)->runtime_flags));
5152                 goto no_delete;
5153         }
5154
5155         if (inode->i_nlink > 0) {
5156                 BUG_ON(btrfs_root_refs(&root->root_item) != 0 &&
5157                        root->root_key.objectid != BTRFS_ROOT_TREE_OBJECTID);
5158                 goto no_delete;
5159         }
5160
5161         ret = btrfs_commit_inode_delayed_inode(inode);
5162         if (ret) {
5163                 btrfs_orphan_del(NULL, inode);
5164                 goto no_delete;
5165         }
5166
5167         rsv = btrfs_alloc_block_rsv(root, BTRFS_BLOCK_RSV_TEMP);
5168         if (!rsv) {
5169                 btrfs_orphan_del(NULL, inode);
5170                 goto no_delete;
5171         }
5172         rsv->size = min_size;
5173         rsv->failfast = 1;
5174         global_rsv = &root->fs_info->global_block_rsv;
5175
5176         btrfs_i_size_write(inode, 0);
5177
5178         /*
5179          * This is a bit simpler than btrfs_truncate since we've already
5180          * reserved our space for our orphan item in the unlink, so we just
5181          * need to reserve some slack space in case we add bytes and update
5182          * inode item when doing the truncate.
5183          */
5184         while (1) {
5185                 ret = btrfs_block_rsv_refill(root, rsv, min_size,
5186                                              BTRFS_RESERVE_FLUSH_LIMIT);
5187
5188                 /*
5189                  * Try and steal from the global reserve since we will
5190                  * likely not use this space anyway, we want to try as
5191                  * hard as possible to get this to work.
5192                  */
5193                 if (ret)
5194                         steal_from_global++;
5195                 else
5196                         steal_from_global = 0;
5197                 ret = 0;
5198
5199                 /*
5200                  * steal_from_global == 0: we reserved stuff, hooray!
5201                  * steal_from_global == 1: we didn't reserve stuff, boo!
5202                  * steal_from_global == 2: we've committed, still not a lot of
5203                  * room but maybe we'll have room in the global reserve this
5204                  * time.
5205                  * steal_from_global == 3: abandon all hope!
5206                  */
5207                 if (steal_from_global > 2) {
5208                         btrfs_warn(root->fs_info,
5209                                 "Could not get space for a delete, will truncate on mount %d",
5210                                 ret);
5211                         btrfs_orphan_del(NULL, inode);
5212                         btrfs_free_block_rsv(root, rsv);
5213                         goto no_delete;
5214                 }
5215
5216                 trans = btrfs_join_transaction(root);
5217                 if (IS_ERR(trans)) {
5218                         btrfs_orphan_del(NULL, inode);
5219                         btrfs_free_block_rsv(root, rsv);
5220                         goto no_delete;
5221                 }
5222
5223                 /*
5224                  * We can't just steal from the global reserve, we need tomake
5225                  * sure there is room to do it, if not we need to commit and try
5226                  * again.
5227                  */
5228                 if (steal_from_global) {
5229                         if (!btrfs_check_space_for_delayed_refs(trans, root))
5230                                 ret = btrfs_block_rsv_migrate(global_rsv, rsv,
5231                                                               min_size);
5232                         else
5233                                 ret = -ENOSPC;
5234                 }
5235
5236                 /*
5237                  * Couldn't steal from the global reserve, we have too much
5238                  * pending stuff built up, commit the transaction and try it
5239                  * again.
5240                  */
5241                 if (ret) {
5242                         ret = btrfs_commit_transaction(trans, root);
5243                         if (ret) {
5244                                 btrfs_orphan_del(NULL, inode);
5245                                 btrfs_free_block_rsv(root, rsv);
5246                                 goto no_delete;
5247                         }
5248                         continue;
5249                 } else {
5250                         steal_from_global = 0;
5251                 }
5252
5253                 trans->block_rsv = rsv;
5254
5255                 ret = btrfs_truncate_inode_items(trans, root, inode, 0, 0);
5256                 if (ret != -ENOSPC && ret != -EAGAIN)
5257                         break;
5258
5259                 trans->block_rsv = &root->fs_info->trans_block_rsv;
5260                 btrfs_end_transaction(trans, root);
5261                 trans = NULL;
5262                 btrfs_btree_balance_dirty(root);
5263         }
5264
5265         btrfs_free_block_rsv(root, rsv);
5266
5267         /*
5268          * Errors here aren't a big deal, it just means we leave orphan items
5269          * in the tree.  They will be cleaned up on the next mount.
5270          */
5271         if (ret == 0) {
5272                 trans->block_rsv = root->orphan_block_rsv;
5273                 btrfs_orphan_del(trans, inode);
5274         } else {
5275                 btrfs_orphan_del(NULL, inode);
5276         }
5277
5278         trans->block_rsv = &root->fs_info->trans_block_rsv;
5279         if (!(root == root->fs_info->tree_root ||
5280               root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID))
5281                 btrfs_return_ino(root, btrfs_ino(inode));
5282
5283         btrfs_end_transaction(trans, root);
5284         btrfs_btree_balance_dirty(root);
5285 no_delete:
5286         btrfs_remove_delayed_node(inode);
5287         clear_inode(inode);
5288 }
5289
5290 /*
5291  * this returns the key found in the dir entry in the location pointer.
5292  * If no dir entries were found, location->objectid is 0.
5293  */
5294 static int btrfs_inode_by_name(struct inode *dir, struct dentry *dentry,
5295                                struct btrfs_key *location)
5296 {
5297         const char *name = dentry->d_name.name;
5298         int namelen = dentry->d_name.len;
5299         struct btrfs_dir_item *di;
5300         struct btrfs_path *path;
5301         struct btrfs_root *root = BTRFS_I(dir)->root;
5302         int ret = 0;
5303
5304         path = btrfs_alloc_path();
5305         if (!path)
5306                 return -ENOMEM;
5307
5308         di = btrfs_lookup_dir_item(NULL, root, path, btrfs_ino(dir), name,
5309                                     namelen, 0);
5310         if (IS_ERR(di))
5311                 ret = PTR_ERR(di);
5312
5313         if (IS_ERR_OR_NULL(di))
5314                 goto out_err;
5315
5316         btrfs_dir_item_key_to_cpu(path->nodes[0], di, location);
5317 out:
5318         btrfs_free_path(path);
5319         return ret;
5320 out_err:
5321         location->objectid = 0;
5322         goto out;
5323 }
5324
5325 /*
5326  * when we hit a tree root in a directory, the btrfs part of the inode
5327  * needs to be changed to reflect the root directory of the tree root.  This
5328  * is kind of like crossing a mount point.
5329  */
5330 static int fixup_tree_root_location(struct btrfs_root *root,
5331                                     struct inode *dir,
5332                                     struct dentry *dentry,
5333                                     struct btrfs_key *location,
5334                                     struct btrfs_root **sub_root)
5335 {
5336         struct btrfs_path *path;
5337         struct btrfs_root *new_root;
5338         struct btrfs_root_ref *ref;
5339         struct extent_buffer *leaf;
5340         struct btrfs_key key;
5341         int ret;
5342         int err = 0;
5343
5344         path = btrfs_alloc_path();
5345         if (!path) {
5346                 err = -ENOMEM;
5347                 goto out;
5348         }
5349
5350         err = -ENOENT;
5351         key.objectid = BTRFS_I(dir)->root->root_key.objectid;
5352         key.type = BTRFS_ROOT_REF_KEY;
5353         key.offset = location->objectid;
5354
5355         ret = btrfs_search_slot(NULL, root->fs_info->tree_root, &key, path,
5356                                 0, 0);
5357         if (ret) {
5358                 if (ret < 0)
5359                         err = ret;
5360                 goto out;
5361         }
5362
5363         leaf = path->nodes[0];
5364         ref = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_root_ref);
5365         if (btrfs_root_ref_dirid(leaf, ref) != btrfs_ino(dir) ||
5366             btrfs_root_ref_name_len(leaf, ref) != dentry->d_name.len)
5367                 goto out;
5368
5369         ret = memcmp_extent_buffer(leaf, dentry->d_name.name,
5370                                    (unsigned long)(ref + 1),
5371                                    dentry->d_name.len);
5372         if (ret)
5373                 goto out;
5374
5375         btrfs_release_path(path);
5376
5377         new_root = btrfs_read_fs_root_no_name(root->fs_info, location);
5378         if (IS_ERR(new_root)) {
5379                 err = PTR_ERR(new_root);
5380                 goto out;
5381         }
5382
5383         *sub_root = new_root;
5384         location->objectid = btrfs_root_dirid(&new_root->root_item);
5385         location->type = BTRFS_INODE_ITEM_KEY;
5386         location->offset = 0;
5387         err = 0;
5388 out:
5389         btrfs_free_path(path);
5390         return err;
5391 }
5392
5393 static void inode_tree_add(struct inode *inode)
5394 {
5395         struct btrfs_root *root = BTRFS_I(inode)->root;
5396         struct btrfs_inode *entry;
5397         struct rb_node **p;
5398         struct rb_node *parent;
5399         struct rb_node *new = &BTRFS_I(inode)->rb_node;
5400         u64 ino = btrfs_ino(inode);
5401
5402         if (inode_unhashed(inode))
5403                 return;
5404         parent = NULL;
5405         spin_lock(&root->inode_lock);
5406         p = &root->inode_tree.rb_node;
5407         while (*p) {
5408                 parent = *p;
5409                 entry = rb_entry(parent, struct btrfs_inode, rb_node);
5410
5411                 if (ino < btrfs_ino(&entry->vfs_inode))
5412                         p = &parent->rb_left;
5413                 else if (ino > btrfs_ino(&entry->vfs_inode))
5414                         p = &parent->rb_right;
5415                 else {
5416                         WARN_ON(!(entry->vfs_inode.i_state &
5417                                   (I_WILL_FREE | I_FREEING)));
5418                         rb_replace_node(parent, new, &root->inode_tree);
5419                         RB_CLEAR_NODE(parent);
5420                         spin_unlock(&root->inode_lock);
5421                         return;
5422                 }
5423         }
5424         rb_link_node(new, parent, p);
5425         rb_insert_color(new, &root->inode_tree);
5426         spin_unlock(&root->inode_lock);
5427 }
5428
5429 static void inode_tree_del(struct inode *inode)
5430 {
5431         struct btrfs_root *root = BTRFS_I(inode)->root;
5432         int empty = 0;
5433
5434         spin_lock(&root->inode_lock);
5435         if (!RB_EMPTY_NODE(&BTRFS_I(inode)->rb_node)) {
5436                 rb_erase(&BTRFS_I(inode)->rb_node, &root->inode_tree);
5437                 RB_CLEAR_NODE(&BTRFS_I(inode)->rb_node);
5438                 empty = RB_EMPTY_ROOT(&root->inode_tree);
5439         }
5440         spin_unlock(&root->inode_lock);
5441
5442         if (empty && btrfs_root_refs(&root->root_item) == 0) {
5443                 synchronize_srcu(&root->fs_info->subvol_srcu);
5444                 spin_lock(&root->inode_lock);
5445                 empty = RB_EMPTY_ROOT(&root->inode_tree);
5446                 spin_unlock(&root->inode_lock);
5447                 if (empty)
5448                         btrfs_add_dead_root(root);
5449         }
5450 }
5451
5452 void btrfs_invalidate_inodes(struct btrfs_root *root)
5453 {
5454         struct rb_node *node;
5455         struct rb_node *prev;
5456         struct btrfs_inode *entry;
5457         struct inode *inode;
5458         u64 objectid = 0;
5459
5460         if (!test_bit(BTRFS_FS_STATE_ERROR, &root->fs_info->fs_state))
5461                 WARN_ON(btrfs_root_refs(&root->root_item) != 0);
5462
5463         spin_lock(&root->inode_lock);
5464 again:
5465         node = root->inode_tree.rb_node;
5466         prev = NULL;
5467         while (node) {
5468                 prev = node;
5469                 entry = rb_entry(node, struct btrfs_inode, rb_node);
5470
5471                 if (objectid < btrfs_ino(&entry->vfs_inode))
5472                         node = node->rb_left;
5473                 else if (objectid > btrfs_ino(&entry->vfs_inode))
5474                         node = node->rb_right;
5475                 else
5476                         break;
5477         }
5478         if (!node) {
5479                 while (prev) {
5480                         entry = rb_entry(prev, struct btrfs_inode, rb_node);
5481                         if (objectid <= btrfs_ino(&entry->vfs_inode)) {
5482                                 node = prev;
5483                                 break;
5484                         }
5485                         prev = rb_next(prev);
5486                 }
5487         }
5488         while (node) {
5489                 entry = rb_entry(node, struct btrfs_inode, rb_node);
5490                 objectid = btrfs_ino(&entry->vfs_inode) + 1;
5491                 inode = igrab(&entry->vfs_inode);
5492                 if (inode) {
5493                         spin_unlock(&root->inode_lock);
5494                         if (atomic_read(&inode->i_count) > 1)
5495                                 d_prune_aliases(inode);
5496                         /*
5497                          * btrfs_drop_inode will have it removed from
5498                          * the inode cache when its usage count
5499                          * hits zero.
5500                          */
5501                         iput(inode);
5502                         cond_resched();
5503                         spin_lock(&root->inode_lock);
5504                         goto again;
5505                 }
5506
5507                 if (cond_resched_lock(&root->inode_lock))
5508                         goto again;
5509
5510                 node = rb_next(node);
5511         }
5512         spin_unlock(&root->inode_lock);
5513 }
5514
5515 static int btrfs_init_locked_inode(struct inode *inode, void *p)
5516 {
5517         struct btrfs_iget_args *args = p;
5518         inode->i_ino = args->location->objectid;
5519         memcpy(&BTRFS_I(inode)->location, args->location,
5520                sizeof(*args->location));
5521         BTRFS_I(inode)->root = args->root;
5522         return 0;
5523 }
5524
5525 static int btrfs_find_actor(struct inode *inode, void *opaque)
5526 {
5527         struct btrfs_iget_args *args = opaque;
5528         return args->location->objectid == BTRFS_I(inode)->location.objectid &&
5529                 args->root == BTRFS_I(inode)->root;
5530 }
5531
5532 static struct inode *btrfs_iget_locked(struct super_block *s,
5533                                        struct btrfs_key *location,
5534                                        struct btrfs_root *root)
5535 {
5536         struct inode *inode;
5537         struct btrfs_iget_args args;
5538         unsigned long hashval = btrfs_inode_hash(location->objectid, root);
5539
5540         args.location = location;
5541         args.root = root;
5542
5543         inode = iget5_locked(s, hashval, btrfs_find_actor,
5544                              btrfs_init_locked_inode,
5545                              (void *)&args);
5546         return inode;
5547 }
5548
5549 /* Get an inode object given its location and corresponding root.
5550  * Returns in *is_new if the inode was read from disk
5551  */
5552 struct inode *btrfs_iget(struct super_block *s, struct btrfs_key *location,
5553                          struct btrfs_root *root, int *new)
5554 {
5555         struct inode *inode;
5556
5557         inode = btrfs_iget_locked(s, location, root);
5558         if (!inode)
5559                 return ERR_PTR(-ENOMEM);
5560
5561         if (inode->i_state & I_NEW) {
5562                 btrfs_read_locked_inode(inode);
5563                 if (!is_bad_inode(inode)) {
5564                         inode_tree_add(inode);
5565                         unlock_new_inode(inode);
5566                         if (new)
5567                                 *new = 1;
5568                 } else {
5569                         unlock_new_inode(inode);
5570                         iput(inode);
5571                         inode = ERR_PTR(-ESTALE);
5572                 }
5573         }
5574
5575         return inode;
5576 }
5577
5578 static struct inode *new_simple_dir(struct super_block *s,
5579                                     struct btrfs_key *key,
5580                                     struct btrfs_root *root)
5581 {
5582         struct inode *inode = new_inode(s);
5583
5584         if (!inode)
5585                 return ERR_PTR(-ENOMEM);
5586
5587         BTRFS_I(inode)->root = root;
5588         memcpy(&BTRFS_I(inode)->location, key, sizeof(*key));
5589         set_bit(BTRFS_INODE_DUMMY, &BTRFS_I(inode)->runtime_flags);
5590
5591         inode->i_ino = BTRFS_EMPTY_SUBVOL_DIR_OBJECTID;
5592         inode->i_op = &btrfs_dir_ro_inode_operations;
5593         inode->i_fop = &simple_dir_operations;
5594         inode->i_mode = S_IFDIR | S_IRUGO | S_IWUSR | S_IXUGO;
5595         inode->i_mtime = CURRENT_TIME;
5596         inode->i_atime = inode->i_mtime;
5597         inode->i_ctime = inode->i_mtime;
5598         BTRFS_I(inode)->i_otime = inode->i_mtime;
5599
5600         return inode;
5601 }
5602
5603 struct inode *btrfs_lookup_dentry(struct inode *dir, struct dentry *dentry)
5604 {
5605         struct inode *inode;
5606         struct btrfs_root *root = BTRFS_I(dir)->root;
5607         struct btrfs_root *sub_root = root;
5608         struct btrfs_key location;
5609         int index;
5610         int ret = 0;
5611
5612         if (dentry->d_name.len > BTRFS_NAME_LEN)
5613                 return ERR_PTR(-ENAMETOOLONG);
5614
5615         ret = btrfs_inode_by_name(dir, dentry, &location);
5616         if (ret < 0)
5617                 return ERR_PTR(ret);
5618
5619         if (location.objectid == 0)
5620                 return ERR_PTR(-ENOENT);
5621
5622         if (location.type == BTRFS_INODE_ITEM_KEY) {
5623                 inode = btrfs_iget(dir->i_sb, &location, root, NULL);
5624                 return inode;
5625         }
5626
5627         BUG_ON(location.type != BTRFS_ROOT_ITEM_KEY);
5628
5629         index = srcu_read_lock(&root->fs_info->subvol_srcu);
5630         ret = fixup_tree_root_location(root, dir, dentry,
5631                                        &location, &sub_root);
5632         if (ret < 0) {
5633                 if (ret != -ENOENT)
5634                         inode = ERR_PTR(ret);
5635                 else
5636                         inode = new_simple_dir(dir->i_sb, &location, sub_root);
5637         } else {
5638                 inode = btrfs_iget(dir->i_sb, &location, sub_root, NULL);
5639         }
5640         srcu_read_unlock(&root->fs_info->subvol_srcu, index);
5641
5642         if (!IS_ERR(inode) && root != sub_root) {
5643                 down_read(&root->fs_info->cleanup_work_sem);
5644                 if (!(inode->i_sb->s_flags & MS_RDONLY))
5645                         ret = btrfs_orphan_cleanup(sub_root);
5646                 up_read(&root->fs_info->cleanup_work_sem);
5647                 if (ret) {
5648                         iput(inode);
5649                         inode = ERR_PTR(ret);
5650                 }
5651         }
5652
5653         return inode;
5654 }
5655
5656 static int btrfs_dentry_delete(const struct dentry *dentry)
5657 {
5658         struct btrfs_root *root;
5659         struct inode *inode = d_inode(dentry);
5660
5661         if (!inode && !IS_ROOT(dentry))
5662                 inode = d_inode(dentry->d_parent);
5663
5664         if (inode) {
5665                 root = BTRFS_I(inode)->root;
5666                 if (btrfs_root_refs(&root->root_item) == 0)
5667                         return 1;
5668
5669                 if (btrfs_ino(inode) == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)
5670                         return 1;
5671         }
5672         return 0;
5673 }
5674
5675 static void btrfs_dentry_release(struct dentry *dentry)
5676 {
5677         kfree(dentry->d_fsdata);
5678 }
5679
5680 static struct dentry *btrfs_lookup(struct inode *dir, struct dentry *dentry,
5681                                    unsigned int flags)
5682 {
5683         struct inode *inode;
5684
5685         inode = btrfs_lookup_dentry(dir, dentry);
5686         if (IS_ERR(inode)) {
5687                 if (PTR_ERR(inode) == -ENOENT)
5688                         inode = NULL;
5689                 else
5690                         return ERR_CAST(inode);
5691         }
5692
5693         return d_splice_alias(inode, dentry);
5694 }
5695
5696 unsigned char btrfs_filetype_table[] = {
5697         DT_UNKNOWN, DT_REG, DT_DIR, DT_CHR, DT_BLK, DT_FIFO, DT_SOCK, DT_LNK
5698 };
5699
5700 static int btrfs_real_readdir(struct file *file, struct dir_context *ctx)
5701 {
5702         struct inode *inode = file_inode(file);
5703         struct btrfs_root *root = BTRFS_I(inode)->root;
5704         struct btrfs_item *item;
5705         struct btrfs_dir_item *di;
5706         struct btrfs_key key;
5707         struct btrfs_key found_key;
5708         struct btrfs_path *path;
5709         struct list_head ins_list;
5710         struct list_head del_list;
5711         int ret;
5712         struct extent_buffer *leaf;
5713         int slot;
5714         unsigned char d_type;
5715         int over = 0;
5716         u32 di_cur;
5717         u32 di_total;
5718         u32 di_len;
5719         int key_type = BTRFS_DIR_INDEX_KEY;
5720         char tmp_name[32];
5721         char *name_ptr;
5722         int name_len;
5723         int is_curr = 0;        /* ctx->pos points to the current index? */
5724
5725         /* FIXME, use a real flag for deciding about the key type */
5726         if (root->fs_info->tree_root == root)
5727                 key_type = BTRFS_DIR_ITEM_KEY;
5728
5729         if (!dir_emit_dots(file, ctx))
5730                 return 0;
5731
5732         path = btrfs_alloc_path();
5733         if (!path)
5734                 return -ENOMEM;
5735
5736         path->reada = READA_FORWARD;
5737
5738         if (key_type == BTRFS_DIR_INDEX_KEY) {
5739                 INIT_LIST_HEAD(&ins_list);
5740                 INIT_LIST_HEAD(&del_list);
5741                 btrfs_get_delayed_items(inode, &ins_list, &del_list);
5742         }
5743
5744         key.type = key_type;
5745         key.offset = ctx->pos;
5746         key.objectid = btrfs_ino(inode);
5747
5748         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
5749         if (ret < 0)
5750                 goto err;
5751
5752         while (1) {
5753                 leaf = path->nodes[0];
5754                 slot = path->slots[0];
5755                 if (slot >= btrfs_header_nritems(leaf)) {
5756                         ret = btrfs_next_leaf(root, path);
5757                         if (ret < 0)
5758                                 goto err;
5759                         else if (ret > 0)
5760                                 break;
5761                         continue;
5762                 }
5763
5764                 item = btrfs_item_nr(slot);
5765                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
5766
5767                 if (found_key.objectid != key.objectid)
5768                         break;
5769                 if (found_key.type != key_type)
5770                         break;
5771                 if (found_key.offset < ctx->pos)
5772                         goto next;
5773                 if (key_type == BTRFS_DIR_INDEX_KEY &&
5774                     btrfs_should_delete_dir_index(&del_list,
5775                                                   found_key.offset))
5776                         goto next;
5777
5778                 ctx->pos = found_key.offset;
5779                 is_curr = 1;
5780
5781                 di = btrfs_item_ptr(leaf, slot, struct btrfs_dir_item);
5782                 di_cur = 0;
5783                 di_total = btrfs_item_size(leaf, item);
5784
5785                 while (di_cur < di_total) {
5786                         struct btrfs_key location;
5787
5788                         if (verify_dir_item(root, leaf, di))
5789                                 break;
5790
5791                         name_len = btrfs_dir_name_len(leaf, di);
5792                         if (name_len <= sizeof(tmp_name)) {
5793                                 name_ptr = tmp_name;
5794                         } else {
5795                                 name_ptr = kmalloc(name_len, GFP_NOFS);
5796                                 if (!name_ptr) {
5797                                         ret = -ENOMEM;
5798                                         goto err;
5799                                 }
5800                         }
5801                         read_extent_buffer(leaf, name_ptr,
5802                                            (unsigned long)(di + 1), name_len);
5803
5804                         d_type = btrfs_filetype_table[btrfs_dir_type(leaf, di)];
5805                         btrfs_dir_item_key_to_cpu(leaf, di, &location);
5806
5807
5808                         /* is this a reference to our own snapshot? If so
5809                          * skip it.
5810                          *
5811                          * In contrast to old kernels, we insert the snapshot's
5812                          * dir item and dir index after it has been created, so
5813                          * we won't find a reference to our own snapshot. We
5814                          * still keep the following code for backward
5815                          * compatibility.
5816                          */
5817                         if (location.type == BTRFS_ROOT_ITEM_KEY &&
5818                             location.objectid == root->root_key.objectid) {
5819                                 over = 0;
5820                                 goto skip;
5821                         }
5822                         over = !dir_emit(ctx, name_ptr, name_len,
5823                                        location.objectid, d_type);
5824
5825 skip:
5826                         if (name_ptr != tmp_name)
5827                                 kfree(name_ptr);
5828
5829                         if (over)
5830                                 goto nopos;
5831                         di_len = btrfs_dir_name_len(leaf, di) +
5832                                  btrfs_dir_data_len(leaf, di) + sizeof(*di);
5833                         di_cur += di_len;
5834                         di = (struct btrfs_dir_item *)((char *)di + di_len);
5835                 }
5836 next:
5837                 path->slots[0]++;
5838         }
5839
5840         if (key_type == BTRFS_DIR_INDEX_KEY) {
5841                 if (is_curr)
5842                         ctx->pos++;
5843                 ret = btrfs_readdir_delayed_dir_index(ctx, &ins_list);
5844                 if (ret)
5845                         goto nopos;
5846         }
5847
5848         /* Reached end of directory/root. Bump pos past the last item. */
5849         ctx->pos++;
5850
5851         /*
5852          * Stop new entries from being returned after we return the last
5853          * entry.
5854          *
5855          * New directory entries are assigned a strictly increasing
5856          * offset.  This means that new entries created during readdir
5857          * are *guaranteed* to be seen in the future by that readdir.
5858          * This has broken buggy programs which operate on names as
5859          * they're returned by readdir.  Until we re-use freed offsets
5860          * we have this hack to stop new entries from being returned
5861          * under the assumption that they'll never reach this huge
5862          * offset.
5863          *
5864          * This is being careful not to overflow 32bit loff_t unless the
5865          * last entry requires it because doing so has broken 32bit apps
5866          * in the past.
5867          */
5868         if (key_type == BTRFS_DIR_INDEX_KEY) {
5869                 if (ctx->pos >= INT_MAX)
5870                         ctx->pos = LLONG_MAX;
5871                 else
5872                         ctx->pos = INT_MAX;
5873         }
5874 nopos:
5875         ret = 0;
5876 err:
5877         if (key_type == BTRFS_DIR_INDEX_KEY)
5878                 btrfs_put_delayed_items(&ins_list, &del_list);
5879         btrfs_free_path(path);
5880         return ret;
5881 }
5882
5883 int btrfs_write_inode(struct inode *inode, struct writeback_control *wbc)
5884 {
5885         struct btrfs_root *root = BTRFS_I(inode)->root;
5886         struct btrfs_trans_handle *trans;
5887         int ret = 0;
5888         bool nolock = false;
5889
5890         if (test_bit(BTRFS_INODE_DUMMY, &BTRFS_I(inode)->runtime_flags))
5891                 return 0;
5892
5893         if (btrfs_fs_closing(root->fs_info) && btrfs_is_free_space_inode(inode))
5894                 nolock = true;
5895
5896         if (wbc->sync_mode == WB_SYNC_ALL) {
5897                 if (nolock)
5898                         trans = btrfs_join_transaction_nolock(root);
5899                 else
5900                         trans = btrfs_join_transaction(root);
5901                 if (IS_ERR(trans))
5902                         return PTR_ERR(trans);
5903                 ret = btrfs_commit_transaction(trans, root);
5904         }
5905         return ret;
5906 }
5907
5908 /*
5909  * This is somewhat expensive, updating the tree every time the
5910  * inode changes.  But, it is most likely to find the inode in cache.
5911  * FIXME, needs more benchmarking...there are no reasons other than performance
5912  * to keep or drop this code.
5913  */
5914 static int btrfs_dirty_inode(struct inode *inode)
5915 {
5916         struct btrfs_root *root = BTRFS_I(inode)->root;
5917         struct btrfs_trans_handle *trans;
5918         int ret;
5919
5920         if (test_bit(BTRFS_INODE_DUMMY, &BTRFS_I(inode)->runtime_flags))
5921                 return 0;
5922
5923         trans = btrfs_join_transaction(root);
5924         if (IS_ERR(trans))
5925                 return PTR_ERR(trans);
5926
5927         ret = btrfs_update_inode(trans, root, inode);
5928         if (ret && ret == -ENOSPC) {
5929                 /* whoops, lets try again with the full transaction */
5930                 btrfs_end_transaction(trans, root);
5931                 trans = btrfs_start_transaction(root, 1);
5932                 if (IS_ERR(trans))
5933                         return PTR_ERR(trans);
5934
5935                 ret = btrfs_update_inode(trans, root, inode);
5936         }
5937         btrfs_end_transaction(trans, root);
5938         if (BTRFS_I(inode)->delayed_node)
5939                 btrfs_balance_delayed_items(root);
5940
5941         return ret;
5942 }
5943
5944 /*
5945  * This is a copy of file_update_time.  We need this so we can return error on
5946  * ENOSPC for updating the inode in the case of file write and mmap writes.
5947  */
5948 static int btrfs_update_time(struct inode *inode, struct timespec *now,
5949                              int flags)
5950 {
5951         struct btrfs_root *root = BTRFS_I(inode)->root;
5952
5953         if (btrfs_root_readonly(root))
5954                 return -EROFS;
5955
5956         if (flags & S_VERSION)
5957                 inode_inc_iversion(inode);
5958         if (flags & S_CTIME)
5959                 inode->i_ctime = *now;
5960         if (flags & S_MTIME)
5961                 inode->i_mtime = *now;
5962         if (flags & S_ATIME)
5963                 inode->i_atime = *now;
5964         return btrfs_dirty_inode(inode);
5965 }
5966
5967 /*
5968  * find the highest existing sequence number in a directory
5969  * and then set the in-memory index_cnt variable to reflect
5970  * free sequence numbers
5971  */
5972 static int btrfs_set_inode_index_count(struct inode *inode)
5973 {
5974         struct btrfs_root *root = BTRFS_I(inode)->root;
5975         struct btrfs_key key, found_key;
5976         struct btrfs_path *path;
5977         struct extent_buffer *leaf;
5978         int ret;
5979
5980         key.objectid = btrfs_ino(inode);
5981         key.type = BTRFS_DIR_INDEX_KEY;
5982         key.offset = (u64)-1;
5983
5984         path = btrfs_alloc_path();
5985         if (!path)
5986                 return -ENOMEM;
5987
5988         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
5989         if (ret < 0)
5990                 goto out;
5991         /* FIXME: we should be able to handle this */
5992         if (ret == 0)
5993                 goto out;
5994         ret = 0;
5995
5996         /*
5997          * MAGIC NUMBER EXPLANATION:
5998          * since we search a directory based on f_pos we have to start at 2
5999          * since '.' and '..' have f_pos of 0 and 1 respectively, so everybody
6000          * else has to start at 2
6001          */
6002         if (path->slots[0] == 0) {
6003                 BTRFS_I(inode)->index_cnt = 2;
6004                 goto out;
6005         }
6006
6007         path->slots[0]--;
6008
6009         leaf = path->nodes[0];
6010         btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
6011
6012         if (found_key.objectid != btrfs_ino(inode) ||
6013             found_key.type != BTRFS_DIR_INDEX_KEY) {
6014                 BTRFS_I(inode)->index_cnt = 2;
6015                 goto out;
6016         }
6017
6018         BTRFS_I(inode)->index_cnt = found_key.offset + 1;
6019 out:
6020         btrfs_free_path(path);
6021         return ret;
6022 }
6023
6024 /*
6025  * helper to find a free sequence number in a given directory.  This current
6026  * code is very simple, later versions will do smarter things in the btree
6027  */
6028 int btrfs_set_inode_index(struct inode *dir, u64 *index)
6029 {
6030         int ret = 0;
6031
6032         if (BTRFS_I(dir)->index_cnt == (u64)-1) {
6033                 ret = btrfs_inode_delayed_dir_index_count(dir);
6034                 if (ret) {
6035                         ret = btrfs_set_inode_index_count(dir);
6036                         if (ret)
6037                                 return ret;
6038                 }
6039         }
6040
6041         *index = BTRFS_I(dir)->index_cnt;
6042         BTRFS_I(dir)->index_cnt++;
6043
6044         return ret;
6045 }
6046
6047 static int btrfs_insert_inode_locked(struct inode *inode)
6048 {
6049         struct btrfs_iget_args args;
6050         args.location = &BTRFS_I(inode)->location;
6051         args.root = BTRFS_I(inode)->root;
6052
6053         return insert_inode_locked4(inode,
6054                    btrfs_inode_hash(inode->i_ino, BTRFS_I(inode)->root),
6055                    btrfs_find_actor, &args);
6056 }
6057
6058 static struct inode *btrfs_new_inode(struct btrfs_trans_handle *trans,
6059                                      struct btrfs_root *root,
6060                                      struct inode *dir,
6061                                      const char *name, int name_len,
6062                                      u64 ref_objectid, u64 objectid,
6063                                      umode_t mode, u64 *index)
6064 {
6065         struct inode *inode;
6066         struct btrfs_inode_item *inode_item;
6067         struct btrfs_key *location;
6068         struct btrfs_path *path;
6069         struct btrfs_inode_ref *ref;
6070         struct btrfs_key key[2];
6071         u32 sizes[2];
6072         int nitems = name ? 2 : 1;
6073         unsigned long ptr;
6074         int ret;
6075
6076         path = btrfs_alloc_path();
6077         if (!path)
6078                 return ERR_PTR(-ENOMEM);
6079
6080         inode = new_inode(root->fs_info->sb);
6081         if (!inode) {
6082                 btrfs_free_path(path);
6083                 return ERR_PTR(-ENOMEM);
6084         }
6085
6086         /*
6087          * O_TMPFILE, set link count to 0, so that after this point,
6088          * we fill in an inode item with the correct link count.
6089          */
6090         if (!name)
6091                 set_nlink(inode, 0);
6092
6093         /*
6094          * we have to initialize this early, so we can reclaim the inode
6095          * number if we fail afterwards in this function.
6096          */
6097         inode->i_ino = objectid;
6098
6099         if (dir && name) {
6100                 trace_btrfs_inode_request(dir);
6101
6102                 ret = btrfs_set_inode_index(dir, index);
6103                 if (ret) {
6104                         btrfs_free_path(path);
6105                         iput(inode);
6106                         return ERR_PTR(ret);
6107                 }
6108         } else if (dir) {
6109                 *index = 0;
6110         }
6111         /*
6112          * index_cnt is ignored for everything but a dir,
6113          * btrfs_get_inode_index_count has an explanation for the magic
6114          * number
6115          */
6116         BTRFS_I(inode)->index_cnt = 2;
6117         BTRFS_I(inode)->dir_index = *index;
6118         BTRFS_I(inode)->root = root;
6119         BTRFS_I(inode)->generation = trans->transid;
6120         inode->i_generation = BTRFS_I(inode)->generation;
6121
6122         /*
6123          * We could have gotten an inode number from somebody who was fsynced
6124          * and then removed in this same transaction, so let's just set full
6125          * sync since it will be a full sync anyway and this will blow away the
6126          * old info in the log.
6127          */
6128         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &BTRFS_I(inode)->runtime_flags);
6129
6130         key[0].objectid = objectid;
6131         key[0].type = BTRFS_INODE_ITEM_KEY;
6132         key[0].offset = 0;
6133
6134         sizes[0] = sizeof(struct btrfs_inode_item);
6135
6136         if (name) {
6137                 /*
6138                  * Start new inodes with an inode_ref. This is slightly more
6139                  * efficient for small numbers of hard links since they will
6140                  * be packed into one item. Extended refs will kick in if we
6141                  * add more hard links than can fit in the ref item.
6142                  */
6143                 key[1].objectid = objectid;
6144                 key[1].type = BTRFS_INODE_REF_KEY;
6145                 key[1].offset = ref_objectid;
6146
6147                 sizes[1] = name_len + sizeof(*ref);
6148         }
6149
6150         location = &BTRFS_I(inode)->location;
6151         location->objectid = objectid;
6152         location->offset = 0;
6153         location->type = BTRFS_INODE_ITEM_KEY;
6154
6155         ret = btrfs_insert_inode_locked(inode);
6156         if (ret < 0)
6157                 goto fail;
6158
6159         path->leave_spinning = 1;
6160         ret = btrfs_insert_empty_items(trans, root, path, key, sizes, nitems);
6161         if (ret != 0)
6162                 goto fail_unlock;
6163
6164         inode_init_owner(inode, dir, mode);
6165         inode_set_bytes(inode, 0);
6166
6167         inode->i_mtime = CURRENT_TIME;
6168         inode->i_atime = inode->i_mtime;
6169         inode->i_ctime = inode->i_mtime;
6170         BTRFS_I(inode)->i_otime = inode->i_mtime;
6171
6172         inode_item = btrfs_item_ptr(path->nodes[0], path->slots[0],
6173                                   struct btrfs_inode_item);
6174         memset_extent_buffer(path->nodes[0], 0, (unsigned long)inode_item,
6175                              sizeof(*inode_item));
6176         fill_inode_item(trans, path->nodes[0], inode_item, inode);
6177
6178         if (name) {
6179                 ref = btrfs_item_ptr(path->nodes[0], path->slots[0] + 1,
6180                                      struct btrfs_inode_ref);
6181                 btrfs_set_inode_ref_name_len(path->nodes[0], ref, name_len);
6182                 btrfs_set_inode_ref_index(path->nodes[0], ref, *index);
6183                 ptr = (unsigned long)(ref + 1);
6184                 write_extent_buffer(path->nodes[0], name, ptr, name_len);
6185         }
6186
6187         btrfs_mark_buffer_dirty(path->nodes[0]);
6188         btrfs_free_path(path);
6189
6190         btrfs_inherit_iflags(inode, dir);
6191
6192         if (S_ISREG(mode)) {
6193                 if (btrfs_test_opt(root, NODATASUM))
6194                         BTRFS_I(inode)->flags |= BTRFS_INODE_NODATASUM;
6195                 if (btrfs_test_opt(root, NODATACOW))
6196                         BTRFS_I(inode)->flags |= BTRFS_INODE_NODATACOW |
6197                                 BTRFS_INODE_NODATASUM;
6198         }
6199
6200         inode_tree_add(inode);
6201
6202         trace_btrfs_inode_new(inode);
6203         btrfs_set_inode_last_trans(trans, inode);
6204
6205         btrfs_update_root_times(trans, root);
6206
6207         ret = btrfs_inode_inherit_props(trans, inode, dir);
6208         if (ret)
6209                 btrfs_err(root->fs_info,
6210                           "error inheriting props for ino %llu (root %llu): %d",
6211                           btrfs_ino(inode), root->root_key.objectid, ret);
6212
6213         return inode;
6214
6215 fail_unlock:
6216         unlock_new_inode(inode);
6217 fail:
6218         if (dir && name)
6219                 BTRFS_I(dir)->index_cnt--;
6220         btrfs_free_path(path);
6221         iput(inode);
6222         return ERR_PTR(ret);
6223 }
6224
6225 static inline u8 btrfs_inode_type(struct inode *inode)
6226 {
6227         return btrfs_type_by_mode[(inode->i_mode & S_IFMT) >> S_SHIFT];
6228 }
6229
6230 /*
6231  * utility function to add 'inode' into 'parent_inode' with
6232  * a give name and a given sequence number.
6233  * if 'add_backref' is true, also insert a backref from the
6234  * inode to the parent directory.
6235  */
6236 int btrfs_add_link(struct btrfs_trans_handle *trans,
6237                    struct inode *parent_inode, struct inode *inode,
6238                    const char *name, int name_len, int add_backref, u64 index)
6239 {
6240         int ret = 0;
6241         struct btrfs_key key;
6242         struct btrfs_root *root = BTRFS_I(parent_inode)->root;
6243         u64 ino = btrfs_ino(inode);
6244         u64 parent_ino = btrfs_ino(parent_inode);
6245
6246         if (unlikely(ino == BTRFS_FIRST_FREE_OBJECTID)) {
6247                 memcpy(&key, &BTRFS_I(inode)->root->root_key, sizeof(key));
6248         } else {
6249                 key.objectid = ino;
6250                 key.type = BTRFS_INODE_ITEM_KEY;
6251                 key.offset = 0;
6252         }
6253
6254         if (unlikely(ino == BTRFS_FIRST_FREE_OBJECTID)) {
6255                 ret = btrfs_add_root_ref(trans, root->fs_info->tree_root,
6256                                          key.objectid, root->root_key.objectid,
6257                                          parent_ino, index, name, name_len);
6258         } else if (add_backref) {
6259                 ret = btrfs_insert_inode_ref(trans, root, name, name_len, ino,
6260                                              parent_ino, index);
6261         }
6262
6263         /* Nothing to clean up yet */
6264         if (ret)
6265                 return ret;
6266
6267         ret = btrfs_insert_dir_item(trans, root, name, name_len,
6268                                     parent_inode, &key,
6269                                     btrfs_inode_type(inode), index);
6270         if (ret == -EEXIST || ret == -EOVERFLOW)
6271                 goto fail_dir_item;
6272         else if (ret) {
6273                 btrfs_abort_transaction(trans, root, ret);
6274                 return ret;
6275         }
6276
6277         btrfs_i_size_write(parent_inode, parent_inode->i_size +
6278                            name_len * 2);
6279         inode_inc_iversion(parent_inode);
6280         parent_inode->i_mtime = parent_inode->i_ctime = CURRENT_TIME;
6281         ret = btrfs_update_inode(trans, root, parent_inode);
6282         if (ret)
6283                 btrfs_abort_transaction(trans, root, ret);
6284         return ret;
6285
6286 fail_dir_item:
6287         if (unlikely(ino == BTRFS_FIRST_FREE_OBJECTID)) {
6288                 u64 local_index;
6289                 int err;
6290                 err = btrfs_del_root_ref(trans, root->fs_info->tree_root,
6291                                  key.objectid, root->root_key.objectid,
6292                                  parent_ino, &local_index, name, name_len);
6293
6294         } else if (add_backref) {
6295                 u64 local_index;
6296                 int err;
6297
6298                 err = btrfs_del_inode_ref(trans, root, name, name_len,
6299                                           ino, parent_ino, &local_index);
6300         }
6301         return ret;
6302 }
6303
6304 static int btrfs_add_nondir(struct btrfs_trans_handle *trans,
6305                             struct inode *dir, struct dentry *dentry,
6306                             struct inode *inode, int backref, u64 index)
6307 {
6308         int err = btrfs_add_link(trans, dir, inode,
6309                                  dentry->d_name.name, dentry->d_name.len,
6310                                  backref, index);
6311         if (err > 0)
6312                 err = -EEXIST;
6313         return err;
6314 }
6315
6316 static int btrfs_mknod(struct inode *dir, struct dentry *dentry,
6317                         umode_t mode, dev_t rdev)
6318 {
6319         struct btrfs_trans_handle *trans;
6320         struct btrfs_root *root = BTRFS_I(dir)->root;
6321         struct inode *inode = NULL;
6322         int err;
6323         int drop_inode = 0;
6324         u64 objectid;
6325         u64 index = 0;
6326
6327         /*
6328          * 2 for inode item and ref
6329          * 2 for dir items
6330          * 1 for xattr if selinux is on
6331          */
6332         trans = btrfs_start_transaction(root, 5);
6333         if (IS_ERR(trans))
6334                 return PTR_ERR(trans);
6335
6336         err = btrfs_find_free_ino(root, &objectid);
6337         if (err)
6338                 goto out_unlock;
6339
6340         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
6341                                 dentry->d_name.len, btrfs_ino(dir), objectid,
6342                                 mode, &index);
6343         if (IS_ERR(inode)) {
6344                 err = PTR_ERR(inode);
6345                 goto out_unlock;
6346         }
6347
6348         /*
6349         * If the active LSM wants to access the inode during
6350         * d_instantiate it needs these. Smack checks to see
6351         * if the filesystem supports xattrs by looking at the
6352         * ops vector.
6353         */
6354         inode->i_op = &btrfs_special_inode_operations;
6355         init_special_inode(inode, inode->i_mode, rdev);
6356
6357         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
6358         if (err)
6359                 goto out_unlock_inode;
6360
6361         err = btrfs_add_nondir(trans, dir, dentry, inode, 0, index);
6362         if (err) {
6363                 goto out_unlock_inode;
6364         } else {
6365                 btrfs_update_inode(trans, root, inode);
6366                 unlock_new_inode(inode);
6367                 d_instantiate(dentry, inode);
6368         }
6369
6370 out_unlock:
6371         btrfs_end_transaction(trans, root);
6372         btrfs_balance_delayed_items(root);
6373         btrfs_btree_balance_dirty(root);
6374         if (drop_inode) {
6375                 inode_dec_link_count(inode);
6376                 iput(inode);
6377         }
6378         return err;
6379
6380 out_unlock_inode:
6381         drop_inode = 1;
6382         unlock_new_inode(inode);
6383         goto out_unlock;
6384
6385 }
6386
6387 static int btrfs_create(struct inode *dir, struct dentry *dentry,
6388                         umode_t mode, bool excl)
6389 {
6390         struct btrfs_trans_handle *trans;
6391         struct btrfs_root *root = BTRFS_I(dir)->root;
6392         struct inode *inode = NULL;
6393         int drop_inode_on_err = 0;
6394         int err;
6395         u64 objectid;
6396         u64 index = 0;
6397
6398         /*
6399          * 2 for inode item and ref
6400          * 2 for dir items
6401          * 1 for xattr if selinux is on
6402          */
6403         trans = btrfs_start_transaction(root, 5);
6404         if (IS_ERR(trans))
6405                 return PTR_ERR(trans);
6406
6407         err = btrfs_find_free_ino(root, &objectid);
6408         if (err)
6409                 goto out_unlock;
6410
6411         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
6412                                 dentry->d_name.len, btrfs_ino(dir), objectid,
6413                                 mode, &index);
6414         if (IS_ERR(inode)) {
6415                 err = PTR_ERR(inode);
6416                 goto out_unlock;
6417         }
6418         drop_inode_on_err = 1;
6419         /*
6420         * If the active LSM wants to access the inode during
6421         * d_instantiate it needs these. Smack checks to see
6422         * if the filesystem supports xattrs by looking at the
6423         * ops vector.
6424         */
6425         inode->i_fop = &btrfs_file_operations;
6426         inode->i_op = &btrfs_file_inode_operations;
6427         inode->i_mapping->a_ops = &btrfs_aops;
6428
6429         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
6430         if (err)
6431                 goto out_unlock_inode;
6432
6433         err = btrfs_update_inode(trans, root, inode);
6434         if (err)
6435                 goto out_unlock_inode;
6436
6437         err = btrfs_add_nondir(trans, dir, dentry, inode, 0, index);
6438         if (err)
6439                 goto out_unlock_inode;
6440
6441         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
6442         unlock_new_inode(inode);
6443         d_instantiate(dentry, inode);
6444
6445 out_unlock:
6446         btrfs_end_transaction(trans, root);
6447         if (err && drop_inode_on_err) {
6448                 inode_dec_link_count(inode);
6449                 iput(inode);
6450         }
6451         btrfs_balance_delayed_items(root);
6452         btrfs_btree_balance_dirty(root);
6453         return err;
6454
6455 out_unlock_inode:
6456         unlock_new_inode(inode);
6457         goto out_unlock;
6458
6459 }
6460
6461 static int btrfs_link(struct dentry *old_dentry, struct inode *dir,
6462                       struct dentry *dentry)
6463 {
6464         struct btrfs_trans_handle *trans = NULL;
6465         struct btrfs_root *root = BTRFS_I(dir)->root;
6466         struct inode *inode = d_inode(old_dentry);
6467         u64 index;
6468         int err;
6469         int drop_inode = 0;
6470
6471         /* do not allow sys_link's with other subvols of the same device */
6472         if (root->objectid != BTRFS_I(inode)->root->objectid)
6473                 return -EXDEV;
6474
6475         if (inode->i_nlink >= BTRFS_LINK_MAX)
6476                 return -EMLINK;
6477
6478         err = btrfs_set_inode_index(dir, &index);
6479         if (err)
6480                 goto fail;
6481
6482         /*
6483          * 2 items for inode and inode ref
6484          * 2 items for dir items
6485          * 1 item for parent inode
6486          */
6487         trans = btrfs_start_transaction(root, 5);
6488         if (IS_ERR(trans)) {
6489                 err = PTR_ERR(trans);
6490                 trans = NULL;
6491                 goto fail;
6492         }
6493
6494         /* There are several dir indexes for this inode, clear the cache. */
6495         BTRFS_I(inode)->dir_index = 0ULL;
6496         inc_nlink(inode);
6497         inode_inc_iversion(inode);
6498         inode->i_ctime = CURRENT_TIME;
6499         ihold(inode);
6500         set_bit(BTRFS_INODE_COPY_EVERYTHING, &BTRFS_I(inode)->runtime_flags);
6501
6502         err = btrfs_add_nondir(trans, dir, dentry, inode, 1, index);
6503
6504         if (err) {
6505                 drop_inode = 1;
6506         } else {
6507                 struct dentry *parent = dentry->d_parent;
6508                 err = btrfs_update_inode(trans, root, inode);
6509                 if (err)
6510                         goto fail;
6511                 if (inode->i_nlink == 1) {
6512                         /*
6513                          * If new hard link count is 1, it's a file created
6514                          * with open(2) O_TMPFILE flag.
6515                          */
6516                         err = btrfs_orphan_del(trans, inode);
6517                         if (err)
6518                                 goto fail;
6519                 }
6520                 d_instantiate(dentry, inode);
6521                 btrfs_log_new_name(trans, inode, NULL, parent);
6522         }
6523
6524         btrfs_balance_delayed_items(root);
6525 fail:
6526         if (trans)
6527                 btrfs_end_transaction(trans, root);
6528         if (drop_inode) {
6529                 inode_dec_link_count(inode);
6530                 iput(inode);
6531         }
6532         btrfs_btree_balance_dirty(root);
6533         return err;
6534 }
6535
6536 static int btrfs_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode)
6537 {
6538         struct inode *inode = NULL;
6539         struct btrfs_trans_handle *trans;
6540         struct btrfs_root *root = BTRFS_I(dir)->root;
6541         int err = 0;
6542         int drop_on_err = 0;
6543         u64 objectid = 0;
6544         u64 index = 0;
6545
6546         /*
6547          * 2 items for inode and ref
6548          * 2 items for dir items
6549          * 1 for xattr if selinux is on
6550          */
6551         trans = btrfs_start_transaction(root, 5);
6552         if (IS_ERR(trans))
6553                 return PTR_ERR(trans);
6554
6555         err = btrfs_find_free_ino(root, &objectid);
6556         if (err)
6557                 goto out_fail;
6558
6559         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
6560                                 dentry->d_name.len, btrfs_ino(dir), objectid,
6561                                 S_IFDIR | mode, &index);
6562         if (IS_ERR(inode)) {
6563                 err = PTR_ERR(inode);
6564                 goto out_fail;
6565         }
6566
6567         drop_on_err = 1;
6568         /* these must be set before we unlock the inode */
6569         inode->i_op = &btrfs_dir_inode_operations;
6570         inode->i_fop = &btrfs_dir_file_operations;
6571
6572         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
6573         if (err)
6574                 goto out_fail_inode;
6575
6576         btrfs_i_size_write(inode, 0);
6577         err = btrfs_update_inode(trans, root, inode);
6578         if (err)
6579                 goto out_fail_inode;
6580
6581         err = btrfs_add_link(trans, dir, inode, dentry->d_name.name,
6582                              dentry->d_name.len, 0, index);
6583         if (err)
6584                 goto out_fail_inode;
6585
6586         d_instantiate(dentry, inode);
6587         /*
6588          * mkdir is special.  We're unlocking after we call d_instantiate
6589          * to avoid a race with nfsd calling d_instantiate.
6590          */
6591         unlock_new_inode(inode);
6592         drop_on_err = 0;
6593
6594 out_fail:
6595         btrfs_end_transaction(trans, root);
6596         if (drop_on_err) {
6597                 inode_dec_link_count(inode);
6598                 iput(inode);
6599         }
6600         btrfs_balance_delayed_items(root);
6601         btrfs_btree_balance_dirty(root);
6602         return err;
6603
6604 out_fail_inode:
6605         unlock_new_inode(inode);
6606         goto out_fail;
6607 }
6608
6609 /* Find next extent map of a given extent map, caller needs to ensure locks */
6610 static struct extent_map *next_extent_map(struct extent_map *em)
6611 {
6612         struct rb_node *next;
6613
6614         next = rb_next(&em->rb_node);
6615         if (!next)
6616                 return NULL;
6617         return container_of(next, struct extent_map, rb_node);
6618 }
6619
6620 static struct extent_map *prev_extent_map(struct extent_map *em)
6621 {
6622         struct rb_node *prev;
6623
6624         prev = rb_prev(&em->rb_node);
6625         if (!prev)
6626                 return NULL;
6627         return container_of(prev, struct extent_map, rb_node);
6628 }
6629
6630 /* helper for btfs_get_extent.  Given an existing extent in the tree,
6631  * the existing extent is the nearest extent to map_start,
6632  * and an extent that you want to insert, deal with overlap and insert
6633  * the best fitted new extent into the tree.
6634  */
6635 static int merge_extent_mapping(struct extent_map_tree *em_tree,
6636                                 struct extent_map *existing,
6637                                 struct extent_map *em,
6638                                 u64 map_start)
6639 {
6640         struct extent_map *prev;
6641         struct extent_map *next;
6642         u64 start;
6643         u64 end;
6644         u64 start_diff;
6645
6646         BUG_ON(map_start < em->start || map_start >= extent_map_end(em));
6647
6648         if (existing->start > map_start) {
6649                 next = existing;
6650                 prev = prev_extent_map(next);
6651         } else {
6652                 prev = existing;
6653                 next = next_extent_map(prev);
6654         }
6655
6656         start = prev ? extent_map_end(prev) : em->start;
6657         start = max_t(u64, start, em->start);
6658         end = next ? next->start : extent_map_end(em);
6659         end = min_t(u64, end, extent_map_end(em));
6660         start_diff = start - em->start;
6661         em->start = start;
6662         em->len = end - start;
6663         if (em->block_start < EXTENT_MAP_LAST_BYTE &&
6664             !test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
6665                 em->block_start += start_diff;
6666                 em->block_len -= start_diff;
6667         }
6668         return add_extent_mapping(em_tree, em, 0);
6669 }
6670
6671 static noinline int uncompress_inline(struct btrfs_path *path,
6672                                       struct page *page,
6673                                       size_t pg_offset, u64 extent_offset,
6674                                       struct btrfs_file_extent_item *item)
6675 {
6676         int ret;
6677         struct extent_buffer *leaf = path->nodes[0];
6678         char *tmp;
6679         size_t max_size;
6680         unsigned long inline_size;
6681         unsigned long ptr;
6682         int compress_type;
6683
6684         WARN_ON(pg_offset != 0);
6685         compress_type = btrfs_file_extent_compression(leaf, item);
6686         max_size = btrfs_file_extent_ram_bytes(leaf, item);
6687         inline_size = btrfs_file_extent_inline_item_len(leaf,
6688                                         btrfs_item_nr(path->slots[0]));
6689         tmp = kmalloc(inline_size, GFP_NOFS);
6690         if (!tmp)
6691                 return -ENOMEM;
6692         ptr = btrfs_file_extent_inline_start(item);
6693
6694         read_extent_buffer(leaf, tmp, ptr, inline_size);
6695
6696         max_size = min_t(unsigned long, PAGE_CACHE_SIZE, max_size);
6697         ret = btrfs_decompress(compress_type, tmp, page,
6698                                extent_offset, inline_size, max_size);
6699         kfree(tmp);
6700         return ret;
6701 }
6702
6703 /*
6704  * a bit scary, this does extent mapping from logical file offset to the disk.
6705  * the ugly parts come from merging extents from the disk with the in-ram
6706  * representation.  This gets more complex because of the data=ordered code,
6707  * where the in-ram extents might be locked pending data=ordered completion.
6708  *
6709  * This also copies inline extents directly into the page.
6710  */
6711
6712 struct extent_map *btrfs_get_extent(struct inode *inode, struct page *page,
6713                                     size_t pg_offset, u64 start, u64 len,
6714                                     int create)
6715 {
6716         int ret;
6717         int err = 0;
6718         u64 extent_start = 0;
6719         u64 extent_end = 0;
6720         u64 objectid = btrfs_ino(inode);
6721         u32 found_type;
6722         struct btrfs_path *path = NULL;
6723         struct btrfs_root *root = BTRFS_I(inode)->root;
6724         struct btrfs_file_extent_item *item;
6725         struct extent_buffer *leaf;
6726         struct btrfs_key found_key;
6727         struct extent_map *em = NULL;
6728         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
6729         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
6730         struct btrfs_trans_handle *trans = NULL;
6731         const bool new_inline = !page || create;
6732
6733 again:
6734         read_lock(&em_tree->lock);
6735         em = lookup_extent_mapping(em_tree, start, len);
6736         if (em)
6737                 em->bdev = root->fs_info->fs_devices->latest_bdev;
6738         read_unlock(&em_tree->lock);
6739
6740         if (em) {
6741                 if (em->start > start || em->start + em->len <= start)
6742                         free_extent_map(em);
6743                 else if (em->block_start == EXTENT_MAP_INLINE && page)
6744                         free_extent_map(em);
6745                 else
6746                         goto out;
6747         }
6748         em = alloc_extent_map();
6749         if (!em) {
6750                 err = -ENOMEM;
6751                 goto out;
6752         }
6753         em->bdev = root->fs_info->fs_devices->latest_bdev;
6754         em->start = EXTENT_MAP_HOLE;
6755         em->orig_start = EXTENT_MAP_HOLE;
6756         em->len = (u64)-1;
6757         em->block_len = (u64)-1;
6758
6759         if (!path) {
6760                 path = btrfs_alloc_path();
6761                 if (!path) {
6762                         err = -ENOMEM;
6763                         goto out;
6764                 }
6765                 /*
6766                  * Chances are we'll be called again, so go ahead and do
6767                  * readahead
6768                  */
6769                 path->reada = READA_FORWARD;
6770         }
6771
6772         ret = btrfs_lookup_file_extent(trans, root, path,
6773                                        objectid, start, trans != NULL);
6774         if (ret < 0) {
6775                 err = ret;
6776                 goto out;
6777         }
6778
6779         if (ret != 0) {
6780                 if (path->slots[0] == 0)
6781                         goto not_found;
6782                 path->slots[0]--;
6783         }
6784
6785         leaf = path->nodes[0];
6786         item = btrfs_item_ptr(leaf, path->slots[0],
6787                               struct btrfs_file_extent_item);
6788         /* are we inside the extent that was found? */
6789         btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
6790         found_type = found_key.type;
6791         if (found_key.objectid != objectid ||
6792             found_type != BTRFS_EXTENT_DATA_KEY) {
6793                 /*
6794                  * If we backup past the first extent we want to move forward
6795                  * and see if there is an extent in front of us, otherwise we'll
6796                  * say there is a hole for our whole search range which can
6797                  * cause problems.
6798                  */
6799                 extent_end = start;
6800                 goto next;
6801         }
6802
6803         found_type = btrfs_file_extent_type(leaf, item);
6804         extent_start = found_key.offset;
6805         if (found_type == BTRFS_FILE_EXTENT_REG ||
6806             found_type == BTRFS_FILE_EXTENT_PREALLOC) {
6807                 extent_end = extent_start +
6808                        btrfs_file_extent_num_bytes(leaf, item);
6809         } else if (found_type == BTRFS_FILE_EXTENT_INLINE) {
6810                 size_t size;
6811                 size = btrfs_file_extent_inline_len(leaf, path->slots[0], item);
6812                 extent_end = ALIGN(extent_start + size, root->sectorsize);
6813         }
6814 next:
6815         if (start >= extent_end) {
6816                 path->slots[0]++;
6817                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
6818                         ret = btrfs_next_leaf(root, path);
6819                         if (ret < 0) {
6820                                 err = ret;
6821                                 goto out;
6822                         }
6823                         if (ret > 0)
6824                                 goto not_found;
6825                         leaf = path->nodes[0];
6826                 }
6827                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
6828                 if (found_key.objectid != objectid ||
6829                     found_key.type != BTRFS_EXTENT_DATA_KEY)
6830                         goto not_found;
6831                 if (start + len <= found_key.offset)
6832                         goto not_found;
6833                 if (start > found_key.offset)
6834                         goto next;
6835                 em->start = start;
6836                 em->orig_start = start;
6837                 em->len = found_key.offset - start;
6838                 goto not_found_em;
6839         }
6840
6841         btrfs_extent_item_to_extent_map(inode, path, item, new_inline, em);
6842
6843         if (found_type == BTRFS_FILE_EXTENT_REG ||
6844             found_type == BTRFS_FILE_EXTENT_PREALLOC) {
6845                 goto insert;
6846         } else if (found_type == BTRFS_FILE_EXTENT_INLINE) {
6847                 unsigned long ptr;
6848                 char *map;
6849                 size_t size;
6850                 size_t extent_offset;
6851                 size_t copy_size;
6852
6853                 if (new_inline)
6854                         goto out;
6855
6856                 size = btrfs_file_extent_inline_len(leaf, path->slots[0], item);
6857                 extent_offset = page_offset(page) + pg_offset - extent_start;
6858                 copy_size = min_t(u64, PAGE_CACHE_SIZE - pg_offset,
6859                                 size - extent_offset);
6860                 em->start = extent_start + extent_offset;
6861                 em->len = ALIGN(copy_size, root->sectorsize);
6862                 em->orig_block_len = em->len;
6863                 em->orig_start = em->start;
6864                 ptr = btrfs_file_extent_inline_start(item) + extent_offset;
6865                 if (create == 0 && !PageUptodate(page)) {
6866                         if (btrfs_file_extent_compression(leaf, item) !=
6867                             BTRFS_COMPRESS_NONE) {
6868                                 ret = uncompress_inline(path, page, pg_offset,
6869                                                         extent_offset, item);
6870                                 if (ret) {
6871                                         err = ret;
6872                                         goto out;
6873                                 }
6874                         } else {
6875                                 map = kmap(page);
6876                                 read_extent_buffer(leaf, map + pg_offset, ptr,
6877                                                    copy_size);
6878                                 if (pg_offset + copy_size < PAGE_CACHE_SIZE) {
6879                                         memset(map + pg_offset + copy_size, 0,
6880                                                PAGE_CACHE_SIZE - pg_offset -
6881                                                copy_size);
6882                                 }
6883                                 kunmap(page);
6884                         }
6885                         flush_dcache_page(page);
6886                 } else if (create && PageUptodate(page)) {
6887                         BUG();
6888                         if (!trans) {
6889                                 kunmap(page);
6890                                 free_extent_map(em);
6891                                 em = NULL;
6892
6893                                 btrfs_release_path(path);
6894                                 trans = btrfs_join_transaction(root);
6895
6896                                 if (IS_ERR(trans))
6897                                         return ERR_CAST(trans);
6898                                 goto again;
6899                         }
6900                         map = kmap(page);
6901                         write_extent_buffer(leaf, map + pg_offset, ptr,
6902                                             copy_size);
6903                         kunmap(page);
6904                         btrfs_mark_buffer_dirty(leaf);
6905                 }
6906                 set_extent_uptodate(io_tree, em->start,
6907                                     extent_map_end(em) - 1, NULL, GFP_NOFS);
6908                 goto insert;
6909         }
6910 not_found:
6911         em->start = start;
6912         em->orig_start = start;
6913         em->len = len;
6914 not_found_em:
6915         em->block_start = EXTENT_MAP_HOLE;
6916         set_bit(EXTENT_FLAG_VACANCY, &em->flags);
6917 insert:
6918         btrfs_release_path(path);
6919         if (em->start > start || extent_map_end(em) <= start) {
6920                 btrfs_err(root->fs_info, "bad extent! em: [%llu %llu] passed [%llu %llu]",
6921                         em->start, em->len, start, len);
6922                 err = -EIO;
6923                 goto out;
6924         }
6925
6926         err = 0;
6927         write_lock(&em_tree->lock);
6928         ret = add_extent_mapping(em_tree, em, 0);
6929         /* it is possible that someone inserted the extent into the tree
6930          * while we had the lock dropped.  It is also possible that
6931          * an overlapping map exists in the tree
6932          */
6933         if (ret == -EEXIST) {
6934                 struct extent_map *existing;
6935
6936                 ret = 0;
6937
6938                 existing = search_extent_mapping(em_tree, start, len);
6939                 /*
6940                  * existing will always be non-NULL, since there must be
6941                  * extent causing the -EEXIST.
6942                  */
6943                 if (start >= extent_map_end(existing) ||
6944                     start <= existing->start) {
6945                         /*
6946                          * The existing extent map is the one nearest to
6947                          * the [start, start + len) range which overlaps
6948                          */
6949                         err = merge_extent_mapping(em_tree, existing,
6950                                                    em, start);
6951                         free_extent_map(existing);
6952                         if (err) {
6953                                 free_extent_map(em);
6954                                 em = NULL;
6955                         }
6956                 } else {
6957                         free_extent_map(em);
6958                         em = existing;
6959                         err = 0;
6960                 }
6961         }
6962         write_unlock(&em_tree->lock);
6963 out:
6964
6965         trace_btrfs_get_extent(root, em);
6966
6967         btrfs_free_path(path);
6968         if (trans) {
6969                 ret = btrfs_end_transaction(trans, root);
6970                 if (!err)
6971                         err = ret;
6972         }
6973         if (err) {
6974                 free_extent_map(em);
6975                 return ERR_PTR(err);
6976         }
6977         BUG_ON(!em); /* Error is always set */
6978         return em;
6979 }
6980
6981 struct extent_map *btrfs_get_extent_fiemap(struct inode *inode, struct page *page,
6982                                            size_t pg_offset, u64 start, u64 len,
6983                                            int create)
6984 {
6985         struct extent_map *em;
6986         struct extent_map *hole_em = NULL;
6987         u64 range_start = start;
6988         u64 end;
6989         u64 found;
6990         u64 found_end;
6991         int err = 0;
6992
6993         em = btrfs_get_extent(inode, page, pg_offset, start, len, create);
6994         if (IS_ERR(em))
6995                 return em;
6996         if (em) {
6997                 /*
6998                  * if our em maps to
6999                  * -  a hole or
7000                  * -  a pre-alloc extent,
7001                  * there might actually be delalloc bytes behind it.
7002                  */
7003                 if (em->block_start != EXTENT_MAP_HOLE &&
7004                     !test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
7005                         return em;
7006                 else
7007                         hole_em = em;
7008         }
7009
7010         /* check to see if we've wrapped (len == -1 or similar) */
7011         end = start + len;
7012         if (end < start)
7013                 end = (u64)-1;
7014         else
7015                 end -= 1;
7016
7017         em = NULL;
7018
7019         /* ok, we didn't find anything, lets look for delalloc */
7020         found = count_range_bits(&BTRFS_I(inode)->io_tree, &range_start,
7021                                  end, len, EXTENT_DELALLOC, 1);
7022         found_end = range_start + found;
7023         if (found_end < range_start)
7024                 found_end = (u64)-1;
7025
7026         /*
7027          * we didn't find anything useful, return
7028          * the original results from get_extent()
7029          */
7030         if (range_start > end || found_end <= start) {
7031                 em = hole_em;
7032                 hole_em = NULL;
7033                 goto out;
7034         }
7035
7036         /* adjust the range_start to make sure it doesn't
7037          * go backwards from the start they passed in
7038          */
7039         range_start = max(start, range_start);
7040         found = found_end - range_start;
7041
7042         if (found > 0) {
7043                 u64 hole_start = start;
7044                 u64 hole_len = len;
7045
7046                 em = alloc_extent_map();
7047                 if (!em) {
7048                         err = -ENOMEM;
7049                         goto out;
7050                 }
7051                 /*
7052                  * when btrfs_get_extent can't find anything it
7053                  * returns one huge hole
7054                  *
7055                  * make sure what it found really fits our range, and
7056                  * adjust to make sure it is based on the start from
7057                  * the caller
7058                  */
7059                 if (hole_em) {
7060                         u64 calc_end = extent_map_end(hole_em);
7061
7062                         if (calc_end <= start || (hole_em->start > end)) {
7063                                 free_extent_map(hole_em);
7064                                 hole_em = NULL;
7065                         } else {
7066                                 hole_start = max(hole_em->start, start);
7067                                 hole_len = calc_end - hole_start;
7068                         }
7069                 }
7070                 em->bdev = NULL;
7071                 if (hole_em && range_start > hole_start) {
7072                         /* our hole starts before our delalloc, so we
7073                          * have to return just the parts of the hole
7074                          * that go until  the delalloc starts
7075                          */
7076                         em->len = min(hole_len,
7077                                       range_start - hole_start);
7078                         em->start = hole_start;
7079                         em->orig_start = hole_start;
7080                         /*
7081                          * don't adjust block start at all,
7082                          * it is fixed at EXTENT_MAP_HOLE
7083                          */
7084                         em->block_start = hole_em->block_start;
7085                         em->block_len = hole_len;
7086                         if (test_bit(EXTENT_FLAG_PREALLOC, &hole_em->flags))
7087                                 set_bit(EXTENT_FLAG_PREALLOC, &em->flags);
7088                 } else {
7089                         em->start = range_start;
7090                         em->len = found;
7091                         em->orig_start = range_start;
7092                         em->block_start = EXTENT_MAP_DELALLOC;
7093                         em->block_len = found;
7094                 }
7095         } else if (hole_em) {
7096                 return hole_em;
7097         }
7098 out:
7099
7100         free_extent_map(hole_em);
7101         if (err) {
7102                 free_extent_map(em);
7103                 return ERR_PTR(err);
7104         }
7105         return em;
7106 }
7107
7108 static struct extent_map *btrfs_new_extent_direct(struct inode *inode,
7109                                                   u64 start, u64 len)
7110 {
7111         struct btrfs_root *root = BTRFS_I(inode)->root;
7112         struct extent_map *em;
7113         struct btrfs_key ins;
7114         u64 alloc_hint;
7115         int ret;
7116
7117         alloc_hint = get_extent_allocation_hint(inode, start, len);
7118         ret = btrfs_reserve_extent(root, len, root->sectorsize, 0,
7119                                    alloc_hint, &ins, 1, 1);
7120         if (ret)
7121                 return ERR_PTR(ret);
7122
7123         /*
7124          * Create the ordered extent before the extent map. This is to avoid
7125          * races with the fast fsync path that would lead to it logging file
7126          * extent items that point to disk extents that were not yet written to.
7127          * The fast fsync path collects ordered extents into a local list and
7128          * then collects all the new extent maps, so we must create the ordered
7129          * extent first and make sure the fast fsync path collects any new
7130          * ordered extents after collecting new extent maps as well.
7131          * The fsync path simply can not rely on inode_dio_wait() because it
7132          * causes deadlock with AIO.
7133          */
7134         ret = btrfs_add_ordered_extent_dio(inode, start, ins.objectid,
7135                                            ins.offset, ins.offset, 0);
7136         if (ret) {
7137                 btrfs_free_reserved_extent(root, ins.objectid, ins.offset, 1);
7138                 return ERR_PTR(ret);
7139         }
7140
7141         em = create_pinned_em(inode, start, ins.offset, start, ins.objectid,
7142                               ins.offset, ins.offset, ins.offset, 0);
7143         if (IS_ERR(em)) {
7144                 struct btrfs_ordered_extent *oe;
7145
7146                 btrfs_free_reserved_extent(root, ins.objectid, ins.offset, 1);
7147                 oe = btrfs_lookup_ordered_extent(inode, start);
7148                 ASSERT(oe);
7149                 if (WARN_ON(!oe))
7150                         return em;
7151                 set_bit(BTRFS_ORDERED_IOERR, &oe->flags);
7152                 set_bit(BTRFS_ORDERED_IO_DONE, &oe->flags);
7153                 btrfs_remove_ordered_extent(inode, oe);
7154                 /* Once for our lookup and once for the ordered extents tree. */
7155                 btrfs_put_ordered_extent(oe);
7156                 btrfs_put_ordered_extent(oe);
7157         }
7158         return em;
7159 }
7160
7161 /*
7162  * returns 1 when the nocow is safe, < 1 on error, 0 if the
7163  * block must be cow'd
7164  */
7165 noinline int can_nocow_extent(struct inode *inode, u64 offset, u64 *len,
7166                               u64 *orig_start, u64 *orig_block_len,
7167                               u64 *ram_bytes)
7168 {
7169         struct btrfs_trans_handle *trans;
7170         struct btrfs_path *path;
7171         int ret;
7172         struct extent_buffer *leaf;
7173         struct btrfs_root *root = BTRFS_I(inode)->root;
7174         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
7175         struct btrfs_file_extent_item *fi;
7176         struct btrfs_key key;
7177         u64 disk_bytenr;
7178         u64 backref_offset;
7179         u64 extent_end;
7180         u64 num_bytes;
7181         int slot;
7182         int found_type;
7183         bool nocow = (BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW);
7184
7185         path = btrfs_alloc_path();
7186         if (!path)
7187                 return -ENOMEM;
7188
7189         ret = btrfs_lookup_file_extent(NULL, root, path, btrfs_ino(inode),
7190                                        offset, 0);
7191         if (ret < 0)
7192                 goto out;
7193
7194         slot = path->slots[0];
7195         if (ret == 1) {
7196                 if (slot == 0) {
7197                         /* can't find the item, must cow */
7198                         ret = 0;
7199                         goto out;
7200                 }
7201                 slot--;
7202         }
7203         ret = 0;
7204         leaf = path->nodes[0];
7205         btrfs_item_key_to_cpu(leaf, &key, slot);
7206         if (key.objectid != btrfs_ino(inode) ||
7207             key.type != BTRFS_EXTENT_DATA_KEY) {
7208                 /* not our file or wrong item type, must cow */
7209                 goto out;
7210         }
7211
7212         if (key.offset > offset) {
7213                 /* Wrong offset, must cow */
7214                 goto out;
7215         }
7216
7217         fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
7218         found_type = btrfs_file_extent_type(leaf, fi);
7219         if (found_type != BTRFS_FILE_EXTENT_REG &&
7220             found_type != BTRFS_FILE_EXTENT_PREALLOC) {
7221                 /* not a regular extent, must cow */
7222                 goto out;
7223         }
7224
7225         if (!nocow && found_type == BTRFS_FILE_EXTENT_REG)
7226                 goto out;
7227
7228         extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
7229         if (extent_end <= offset)
7230                 goto out;
7231
7232         disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
7233         if (disk_bytenr == 0)
7234                 goto out;
7235
7236         if (btrfs_file_extent_compression(leaf, fi) ||
7237             btrfs_file_extent_encryption(leaf, fi) ||
7238             btrfs_file_extent_other_encoding(leaf, fi))
7239                 goto out;
7240
7241         backref_offset = btrfs_file_extent_offset(leaf, fi);
7242
7243         if (orig_start) {
7244                 *orig_start = key.offset - backref_offset;
7245                 *orig_block_len = btrfs_file_extent_disk_num_bytes(leaf, fi);
7246                 *ram_bytes = btrfs_file_extent_ram_bytes(leaf, fi);
7247         }
7248
7249         if (btrfs_extent_readonly(root, disk_bytenr))
7250                 goto out;
7251
7252         num_bytes = min(offset + *len, extent_end) - offset;
7253         if (!nocow && found_type == BTRFS_FILE_EXTENT_PREALLOC) {
7254                 u64 range_end;
7255
7256                 range_end = round_up(offset + num_bytes, root->sectorsize) - 1;
7257                 ret = test_range_bit(io_tree, offset, range_end,
7258                                      EXTENT_DELALLOC, 0, NULL);
7259                 if (ret) {
7260                         ret = -EAGAIN;
7261                         goto out;
7262                 }
7263         }
7264
7265         btrfs_release_path(path);
7266
7267         /*
7268          * look for other files referencing this extent, if we
7269          * find any we must cow
7270          */
7271         trans = btrfs_join_transaction(root);
7272         if (IS_ERR(trans)) {
7273                 ret = 0;
7274                 goto out;
7275         }
7276
7277         ret = btrfs_cross_ref_exist(trans, root, btrfs_ino(inode),
7278                                     key.offset - backref_offset, disk_bytenr);
7279         btrfs_end_transaction(trans, root);
7280         if (ret) {
7281                 ret = 0;
7282                 goto out;
7283         }
7284
7285         /*
7286          * adjust disk_bytenr and num_bytes to cover just the bytes
7287          * in this extent we are about to write.  If there
7288          * are any csums in that range we have to cow in order
7289          * to keep the csums correct
7290          */
7291         disk_bytenr += backref_offset;
7292         disk_bytenr += offset - key.offset;
7293         if (csum_exist_in_range(root, disk_bytenr, num_bytes))
7294                                 goto out;
7295         /*
7296          * all of the above have passed, it is safe to overwrite this extent
7297          * without cow
7298          */
7299         *len = num_bytes;
7300         ret = 1;
7301 out:
7302         btrfs_free_path(path);
7303         return ret;
7304 }
7305
7306 bool btrfs_page_exists_in_range(struct inode *inode, loff_t start, loff_t end)
7307 {
7308         struct radix_tree_root *root = &inode->i_mapping->page_tree;
7309         int found = false;
7310         void **pagep = NULL;
7311         struct page *page = NULL;
7312         int start_idx;
7313         int end_idx;
7314
7315         start_idx = start >> PAGE_CACHE_SHIFT;
7316
7317         /*
7318          * end is the last byte in the last page.  end == start is legal
7319          */
7320         end_idx = end >> PAGE_CACHE_SHIFT;
7321
7322         rcu_read_lock();
7323
7324         /* Most of the code in this while loop is lifted from
7325          * find_get_page.  It's been modified to begin searching from a
7326          * page and return just the first page found in that range.  If the
7327          * found idx is less than or equal to the end idx then we know that
7328          * a page exists.  If no pages are found or if those pages are
7329          * outside of the range then we're fine (yay!) */
7330         while (page == NULL &&
7331                radix_tree_gang_lookup_slot(root, &pagep, NULL, start_idx, 1)) {
7332                 page = radix_tree_deref_slot(pagep);
7333                 if (unlikely(!page))
7334                         break;
7335
7336                 if (radix_tree_exception(page)) {
7337                         if (radix_tree_deref_retry(page)) {
7338                                 page = NULL;
7339                                 continue;
7340                         }
7341                         /*
7342                          * Otherwise, shmem/tmpfs must be storing a swap entry
7343                          * here as an exceptional entry: so return it without
7344                          * attempting to raise page count.
7345                          */
7346                         page = NULL;
7347                         break; /* TODO: Is this relevant for this use case? */
7348                 }
7349
7350                 if (!page_cache_get_speculative(page)) {
7351                         page = NULL;
7352                         continue;
7353                 }
7354
7355                 /*
7356                  * Has the page moved?
7357                  * This is part of the lockless pagecache protocol. See
7358                  * include/linux/pagemap.h for details.
7359                  */
7360                 if (unlikely(page != *pagep)) {
7361                         page_cache_release(page);
7362                         page = NULL;
7363                 }
7364         }
7365
7366         if (page) {
7367                 if (page->index <= end_idx)
7368                         found = true;
7369                 page_cache_release(page);
7370         }
7371
7372         rcu_read_unlock();
7373         return found;
7374 }
7375
7376 static int lock_extent_direct(struct inode *inode, u64 lockstart, u64 lockend,
7377                               struct extent_state **cached_state, int writing)
7378 {
7379         struct btrfs_ordered_extent *ordered;
7380         int ret = 0;
7381
7382         while (1) {
7383                 lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend,
7384                                  cached_state);
7385                 /*
7386                  * We're concerned with the entire range that we're going to be
7387                  * doing DIO to, so we need to make sure theres no ordered
7388                  * extents in this range.
7389                  */
7390                 ordered = btrfs_lookup_ordered_range(inode, lockstart,
7391                                                      lockend - lockstart + 1);
7392
7393                 /*
7394                  * We need to make sure there are no buffered pages in this
7395                  * range either, we could have raced between the invalidate in
7396                  * generic_file_direct_write and locking the extent.  The
7397                  * invalidate needs to happen so that reads after a write do not
7398                  * get stale data.
7399                  */
7400                 if (!ordered &&
7401                     (!writing ||
7402                      !btrfs_page_exists_in_range(inode, lockstart, lockend)))
7403                         break;
7404
7405                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
7406                                      cached_state, GFP_NOFS);
7407
7408                 if (ordered) {
7409                         btrfs_start_ordered_extent(inode, ordered, 1);
7410                         btrfs_put_ordered_extent(ordered);
7411                 } else {
7412                         /*
7413                          * We could trigger writeback for this range (and wait
7414                          * for it to complete) and then invalidate the pages for
7415                          * this range (through invalidate_inode_pages2_range()),
7416                          * but that can lead us to a deadlock with a concurrent
7417                          * call to readpages() (a buffered read or a defrag call
7418                          * triggered a readahead) on a page lock due to an
7419                          * ordered dio extent we created before but did not have
7420                          * yet a corresponding bio submitted (whence it can not
7421                          * complete), which makes readpages() wait for that
7422                          * ordered extent to complete while holding a lock on
7423                          * that page.
7424                          */
7425                         ret = -ENOTBLK;
7426                         break;
7427                 }
7428
7429                 cond_resched();
7430         }
7431
7432         return ret;
7433 }
7434
7435 static struct extent_map *create_pinned_em(struct inode *inode, u64 start,
7436                                            u64 len, u64 orig_start,
7437                                            u64 block_start, u64 block_len,
7438                                            u64 orig_block_len, u64 ram_bytes,
7439                                            int type)
7440 {
7441         struct extent_map_tree *em_tree;
7442         struct extent_map *em;
7443         struct btrfs_root *root = BTRFS_I(inode)->root;
7444         int ret;
7445
7446         em_tree = &BTRFS_I(inode)->extent_tree;
7447         em = alloc_extent_map();
7448         if (!em)
7449                 return ERR_PTR(-ENOMEM);
7450
7451         em->start = start;
7452         em->orig_start = orig_start;
7453         em->mod_start = start;
7454         em->mod_len = len;
7455         em->len = len;
7456         em->block_len = block_len;
7457         em->block_start = block_start;
7458         em->bdev = root->fs_info->fs_devices->latest_bdev;
7459         em->orig_block_len = orig_block_len;
7460         em->ram_bytes = ram_bytes;
7461         em->generation = -1;
7462         set_bit(EXTENT_FLAG_PINNED, &em->flags);
7463         if (type == BTRFS_ORDERED_PREALLOC)
7464                 set_bit(EXTENT_FLAG_FILLING, &em->flags);
7465
7466         do {
7467                 btrfs_drop_extent_cache(inode, em->start,
7468                                 em->start + em->len - 1, 0);
7469                 write_lock(&em_tree->lock);
7470                 ret = add_extent_mapping(em_tree, em, 1);
7471                 write_unlock(&em_tree->lock);
7472         } while (ret == -EEXIST);
7473
7474         if (ret) {
7475                 free_extent_map(em);
7476                 return ERR_PTR(ret);
7477         }
7478
7479         return em;
7480 }
7481
7482 static void adjust_dio_outstanding_extents(struct inode *inode,
7483                                            struct btrfs_dio_data *dio_data,
7484                                            const u64 len)
7485 {
7486         unsigned num_extents;
7487
7488         num_extents = (unsigned) div64_u64(len + BTRFS_MAX_EXTENT_SIZE - 1,
7489                                            BTRFS_MAX_EXTENT_SIZE);
7490         /*
7491          * If we have an outstanding_extents count still set then we're
7492          * within our reservation, otherwise we need to adjust our inode
7493          * counter appropriately.
7494          */
7495         if (dio_data->outstanding_extents) {
7496                 dio_data->outstanding_extents -= num_extents;
7497         } else {
7498                 spin_lock(&BTRFS_I(inode)->lock);
7499                 BTRFS_I(inode)->outstanding_extents += num_extents;
7500                 spin_unlock(&BTRFS_I(inode)->lock);
7501         }
7502 }
7503
7504 static int btrfs_get_blocks_direct(struct inode *inode, sector_t iblock,
7505                                    struct buffer_head *bh_result, int create)
7506 {
7507         struct extent_map *em;
7508         struct btrfs_root *root = BTRFS_I(inode)->root;
7509         struct extent_state *cached_state = NULL;
7510         struct btrfs_dio_data *dio_data = NULL;
7511         u64 start = iblock << inode->i_blkbits;
7512         u64 lockstart, lockend;
7513         u64 len = bh_result->b_size;
7514         int unlock_bits = EXTENT_LOCKED;
7515         int ret = 0;
7516
7517         if (create)
7518                 unlock_bits |= EXTENT_DIRTY;
7519         else
7520                 len = min_t(u64, len, root->sectorsize);
7521
7522         lockstart = start;
7523         lockend = start + len - 1;
7524
7525         if (current->journal_info) {
7526                 /*
7527                  * Need to pull our outstanding extents and set journal_info to NULL so
7528                  * that anything that needs to check if there's a transction doesn't get
7529                  * confused.
7530                  */
7531                 dio_data = current->journal_info;
7532                 current->journal_info = NULL;
7533         }
7534
7535         /*
7536          * If this errors out it's because we couldn't invalidate pagecache for
7537          * this range and we need to fallback to buffered.
7538          */
7539         if (lock_extent_direct(inode, lockstart, lockend, &cached_state,
7540                                create)) {
7541                 ret = -ENOTBLK;
7542                 goto err;
7543         }
7544
7545         em = btrfs_get_extent(inode, NULL, 0, start, len, 0);
7546         if (IS_ERR(em)) {
7547                 ret = PTR_ERR(em);
7548                 goto unlock_err;
7549         }
7550
7551         /*
7552          * Ok for INLINE and COMPRESSED extents we need to fallback on buffered
7553          * io.  INLINE is special, and we could probably kludge it in here, but
7554          * it's still buffered so for safety lets just fall back to the generic
7555          * buffered path.
7556          *
7557          * For COMPRESSED we _have_ to read the entire extent in so we can
7558          * decompress it, so there will be buffering required no matter what we
7559          * do, so go ahead and fallback to buffered.
7560          *
7561          * We return -ENOTBLK because thats what makes DIO go ahead and go back
7562          * to buffered IO.  Don't blame me, this is the price we pay for using
7563          * the generic code.
7564          */
7565         if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags) ||
7566             em->block_start == EXTENT_MAP_INLINE) {
7567                 free_extent_map(em);
7568                 ret = -ENOTBLK;
7569                 goto unlock_err;
7570         }
7571
7572         /* Just a good old fashioned hole, return */
7573         if (!create && (em->block_start == EXTENT_MAP_HOLE ||
7574                         test_bit(EXTENT_FLAG_PREALLOC, &em->flags))) {
7575                 free_extent_map(em);
7576                 goto unlock_err;
7577         }
7578
7579         /*
7580          * We don't allocate a new extent in the following cases
7581          *
7582          * 1) The inode is marked as NODATACOW.  In this case we'll just use the
7583          * existing extent.
7584          * 2) The extent is marked as PREALLOC.  We're good to go here and can
7585          * just use the extent.
7586          *
7587          */
7588         if (!create) {
7589                 len = min(len, em->len - (start - em->start));
7590                 lockstart = start + len;
7591                 goto unlock;
7592         }
7593
7594         if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags) ||
7595             ((BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW) &&
7596              em->block_start != EXTENT_MAP_HOLE)) {
7597                 int type;
7598                 u64 block_start, orig_start, orig_block_len, ram_bytes;
7599
7600                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
7601                         type = BTRFS_ORDERED_PREALLOC;
7602                 else
7603                         type = BTRFS_ORDERED_NOCOW;
7604                 len = min(len, em->len - (start - em->start));
7605                 block_start = em->block_start + (start - em->start);
7606
7607                 if (can_nocow_extent(inode, start, &len, &orig_start,
7608                                      &orig_block_len, &ram_bytes) == 1) {
7609                         if (type == BTRFS_ORDERED_PREALLOC) {
7610                                 free_extent_map(em);
7611                                 em = create_pinned_em(inode, start, len,
7612                                                        orig_start,
7613                                                        block_start, len,
7614                                                        orig_block_len,
7615                                                        ram_bytes, type);
7616                                 if (IS_ERR(em)) {
7617                                         ret = PTR_ERR(em);
7618                                         goto unlock_err;
7619                                 }
7620                         }
7621
7622                         ret = btrfs_add_ordered_extent_dio(inode, start,
7623                                            block_start, len, len, type);
7624                         if (ret) {
7625                                 free_extent_map(em);
7626                                 goto unlock_err;
7627                         }
7628                         goto unlock;
7629                 }
7630         }
7631
7632         /*
7633          * this will cow the extent, reset the len in case we changed
7634          * it above
7635          */
7636         len = bh_result->b_size;
7637         free_extent_map(em);
7638         em = btrfs_new_extent_direct(inode, start, len);
7639         if (IS_ERR(em)) {
7640                 ret = PTR_ERR(em);
7641                 goto unlock_err;
7642         }
7643         len = min(len, em->len - (start - em->start));
7644 unlock:
7645         bh_result->b_blocknr = (em->block_start + (start - em->start)) >>
7646                 inode->i_blkbits;
7647         bh_result->b_size = len;
7648         bh_result->b_bdev = em->bdev;
7649         set_buffer_mapped(bh_result);
7650         if (create) {
7651                 if (!test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
7652                         set_buffer_new(bh_result);
7653
7654                 /*
7655                  * Need to update the i_size under the extent lock so buffered
7656                  * readers will get the updated i_size when we unlock.
7657                  */
7658                 if (start + len > i_size_read(inode))
7659                         i_size_write(inode, start + len);
7660
7661                 adjust_dio_outstanding_extents(inode, dio_data, len);
7662                 btrfs_free_reserved_data_space(inode, start, len);
7663                 WARN_ON(dio_data->reserve < len);
7664                 dio_data->reserve -= len;
7665                 dio_data->unsubmitted_oe_range_end = start + len;
7666                 current->journal_info = dio_data;
7667         }
7668
7669         /*
7670          * In the case of write we need to clear and unlock the entire range,
7671          * in the case of read we need to unlock only the end area that we
7672          * aren't using if there is any left over space.
7673          */
7674         if (lockstart < lockend) {
7675                 clear_extent_bit(&BTRFS_I(inode)->io_tree, lockstart,
7676                                  lockend, unlock_bits, 1, 0,
7677                                  &cached_state, GFP_NOFS);
7678         } else {
7679                 free_extent_state(cached_state);
7680         }
7681
7682         free_extent_map(em);
7683
7684         return 0;
7685
7686 unlock_err:
7687         clear_extent_bit(&BTRFS_I(inode)->io_tree, lockstart, lockend,
7688                          unlock_bits, 1, 0, &cached_state, GFP_NOFS);
7689 err:
7690         if (dio_data)
7691                 current->journal_info = dio_data;
7692         /*
7693          * Compensate the delalloc release we do in btrfs_direct_IO() when we
7694          * write less data then expected, so that we don't underflow our inode's
7695          * outstanding extents counter.
7696          */
7697         if (create && dio_data)
7698                 adjust_dio_outstanding_extents(inode, dio_data, len);
7699
7700         return ret;
7701 }
7702
7703 static inline int submit_dio_repair_bio(struct inode *inode, struct bio *bio,
7704                                         int rw, int mirror_num)
7705 {
7706         struct btrfs_root *root = BTRFS_I(inode)->root;
7707         int ret;
7708
7709         BUG_ON(rw & REQ_WRITE);
7710
7711         bio_get(bio);
7712
7713         ret = btrfs_bio_wq_end_io(root->fs_info, bio,
7714                                   BTRFS_WQ_ENDIO_DIO_REPAIR);
7715         if (ret)
7716                 goto err;
7717
7718         ret = btrfs_map_bio(root, rw, bio, mirror_num, 0);
7719 err:
7720         bio_put(bio);
7721         return ret;
7722 }
7723
7724 static int btrfs_check_dio_repairable(struct inode *inode,
7725                                       struct bio *failed_bio,
7726                                       struct io_failure_record *failrec,
7727                                       int failed_mirror)
7728 {
7729         int num_copies;
7730
7731         num_copies = btrfs_num_copies(BTRFS_I(inode)->root->fs_info,
7732                                       failrec->logical, failrec->len);
7733         if (num_copies == 1) {
7734                 /*
7735                  * we only have a single copy of the data, so don't bother with
7736                  * all the retry and error correction code that follows. no
7737                  * matter what the error is, it is very likely to persist.
7738                  */
7739                 pr_debug("Check DIO Repairable: cannot repair, num_copies=%d, next_mirror %d, failed_mirror %d\n",
7740                          num_copies, failrec->this_mirror, failed_mirror);
7741                 return 0;
7742         }
7743
7744         failrec->failed_mirror = failed_mirror;
7745         failrec->this_mirror++;
7746         if (failrec->this_mirror == failed_mirror)
7747                 failrec->this_mirror++;
7748
7749         if (failrec->this_mirror > num_copies) {
7750                 pr_debug("Check DIO Repairable: (fail) num_copies=%d, next_mirror %d, failed_mirror %d\n",
7751                          num_copies, failrec->this_mirror, failed_mirror);
7752                 return 0;
7753         }
7754
7755         return 1;
7756 }
7757
7758 static int dio_read_error(struct inode *inode, struct bio *failed_bio,
7759                         struct page *page, unsigned int pgoff,
7760                         u64 start, u64 end, int failed_mirror,
7761                         bio_end_io_t *repair_endio, void *repair_arg)
7762 {
7763         struct io_failure_record *failrec;
7764         struct bio *bio;
7765         int isector;
7766         int read_mode;
7767         int ret;
7768
7769         BUG_ON(failed_bio->bi_rw & REQ_WRITE);
7770
7771         ret = btrfs_get_io_failure_record(inode, start, end, &failrec);
7772         if (ret)
7773                 return ret;
7774
7775         ret = btrfs_check_dio_repairable(inode, failed_bio, failrec,
7776                                          failed_mirror);
7777         if (!ret) {
7778                 free_io_failure(inode, failrec);
7779                 return -EIO;
7780         }
7781
7782         if ((failed_bio->bi_vcnt > 1)
7783                 || (failed_bio->bi_io_vec->bv_len
7784                         > BTRFS_I(inode)->root->sectorsize))
7785                 read_mode = READ_SYNC | REQ_FAILFAST_DEV;
7786         else
7787                 read_mode = READ_SYNC;
7788
7789         isector = start - btrfs_io_bio(failed_bio)->logical;
7790         isector >>= inode->i_sb->s_blocksize_bits;
7791         bio = btrfs_create_repair_bio(inode, failed_bio, failrec, page,
7792                                 pgoff, isector, repair_endio, repair_arg);
7793         if (!bio) {
7794                 free_io_failure(inode, failrec);
7795                 return -EIO;
7796         }
7797
7798         btrfs_debug(BTRFS_I(inode)->root->fs_info,
7799                     "Repair DIO Read Error: submitting new dio read[%#x] to this_mirror=%d, in_validation=%d\n",
7800                     read_mode, failrec->this_mirror, failrec->in_validation);
7801
7802         ret = submit_dio_repair_bio(inode, bio, read_mode,
7803                                     failrec->this_mirror);
7804         if (ret) {
7805                 free_io_failure(inode, failrec);
7806                 bio_put(bio);
7807         }
7808
7809         return ret;
7810 }
7811
7812 struct btrfs_retry_complete {
7813         struct completion done;
7814         struct inode *inode;
7815         u64 start;
7816         int uptodate;
7817 };
7818
7819 static void btrfs_retry_endio_nocsum(struct bio *bio)
7820 {
7821         struct btrfs_retry_complete *done = bio->bi_private;
7822         struct inode *inode;
7823         struct bio_vec *bvec;
7824         int i;
7825
7826         if (bio->bi_error)
7827                 goto end;
7828
7829         ASSERT(bio->bi_vcnt == 1);
7830         inode = bio->bi_io_vec->bv_page->mapping->host;
7831         ASSERT(bio->bi_io_vec->bv_len == BTRFS_I(inode)->root->sectorsize);
7832
7833         done->uptodate = 1;
7834         bio_for_each_segment_all(bvec, bio, i)
7835                 clean_io_failure(done->inode, done->start, bvec->bv_page, 0);
7836 end:
7837         complete(&done->done);
7838         bio_put(bio);
7839 }
7840
7841 static int __btrfs_correct_data_nocsum(struct inode *inode,
7842                                        struct btrfs_io_bio *io_bio)
7843 {
7844         struct btrfs_fs_info *fs_info;
7845         struct bio_vec *bvec;
7846         struct btrfs_retry_complete done;
7847         u64 start;
7848         unsigned int pgoff;
7849         u32 sectorsize;
7850         int nr_sectors;
7851         int i;
7852         int ret;
7853
7854         fs_info = BTRFS_I(inode)->root->fs_info;
7855         sectorsize = BTRFS_I(inode)->root->sectorsize;
7856
7857         start = io_bio->logical;
7858         done.inode = inode;
7859
7860         bio_for_each_segment_all(bvec, &io_bio->bio, i) {
7861                 nr_sectors = BTRFS_BYTES_TO_BLKS(fs_info, bvec->bv_len);
7862                 pgoff = bvec->bv_offset;
7863
7864 next_block_or_try_again:
7865                 done.uptodate = 0;
7866                 done.start = start;
7867                 init_completion(&done.done);
7868
7869                 ret = dio_read_error(inode, &io_bio->bio, bvec->bv_page,
7870                                 pgoff, start, start + sectorsize - 1,
7871                                 io_bio->mirror_num,
7872                                 btrfs_retry_endio_nocsum, &done);
7873                 if (ret)
7874                         return ret;
7875
7876                 wait_for_completion(&done.done);
7877
7878                 if (!done.uptodate) {
7879                         /* We might have another mirror, so try again */
7880                         goto next_block_or_try_again;
7881                 }
7882
7883                 start += sectorsize;
7884
7885                 if (nr_sectors--) {
7886                         pgoff += sectorsize;
7887                         goto next_block_or_try_again;
7888                 }
7889         }
7890
7891         return 0;
7892 }
7893
7894 static void btrfs_retry_endio(struct bio *bio)
7895 {
7896         struct btrfs_retry_complete *done = bio->bi_private;
7897         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
7898         struct inode *inode;
7899         struct bio_vec *bvec;
7900         u64 start;
7901         int uptodate;
7902         int ret;
7903         int i;
7904
7905         if (bio->bi_error)
7906                 goto end;
7907
7908         uptodate = 1;
7909
7910         start = done->start;
7911
7912         ASSERT(bio->bi_vcnt == 1);
7913         inode = bio->bi_io_vec->bv_page->mapping->host;
7914         ASSERT(bio->bi_io_vec->bv_len == BTRFS_I(inode)->root->sectorsize);
7915
7916         bio_for_each_segment_all(bvec, bio, i) {
7917                 ret = __readpage_endio_check(done->inode, io_bio, i,
7918                                         bvec->bv_page, bvec->bv_offset,
7919                                         done->start, bvec->bv_len);
7920                 if (!ret)
7921                         clean_io_failure(done->inode, done->start,
7922                                         bvec->bv_page, bvec->bv_offset);
7923                 else
7924                         uptodate = 0;
7925         }
7926
7927         done->uptodate = uptodate;
7928 end:
7929         complete(&done->done);
7930         bio_put(bio);
7931 }
7932
7933 static int __btrfs_subio_endio_read(struct inode *inode,
7934                                     struct btrfs_io_bio *io_bio, int err)
7935 {
7936         struct btrfs_fs_info *fs_info;
7937         struct bio_vec *bvec;
7938         struct btrfs_retry_complete done;
7939         u64 start;
7940         u64 offset = 0;
7941         u32 sectorsize;
7942         int nr_sectors;
7943         unsigned int pgoff;
7944         int csum_pos;
7945         int i;
7946         int ret;
7947
7948         fs_info = BTRFS_I(inode)->root->fs_info;
7949         sectorsize = BTRFS_I(inode)->root->sectorsize;
7950
7951         err = 0;
7952         start = io_bio->logical;
7953         done.inode = inode;
7954
7955         bio_for_each_segment_all(bvec, &io_bio->bio, i) {
7956                 nr_sectors = BTRFS_BYTES_TO_BLKS(fs_info, bvec->bv_len);
7957
7958                 pgoff = bvec->bv_offset;
7959 next_block:
7960                 csum_pos = BTRFS_BYTES_TO_BLKS(fs_info, offset);
7961                 ret = __readpage_endio_check(inode, io_bio, csum_pos,
7962                                         bvec->bv_page, pgoff, start,
7963                                         sectorsize);
7964                 if (likely(!ret))
7965                         goto next;
7966 try_again:
7967                 done.uptodate = 0;
7968                 done.start = start;
7969                 init_completion(&done.done);
7970
7971                 ret = dio_read_error(inode, &io_bio->bio, bvec->bv_page,
7972                                 pgoff, start, start + sectorsize - 1,
7973                                 io_bio->mirror_num,
7974                                 btrfs_retry_endio, &done);
7975                 if (ret) {
7976                         err = ret;
7977                         goto next;
7978                 }
7979
7980                 wait_for_completion(&done.done);
7981
7982                 if (!done.uptodate) {
7983                         /* We might have another mirror, so try again */
7984                         goto try_again;
7985                 }
7986 next:
7987                 offset += sectorsize;
7988                 start += sectorsize;
7989
7990                 ASSERT(nr_sectors);
7991
7992                 if (--nr_sectors) {
7993                         pgoff += sectorsize;
7994                         goto next_block;
7995                 }
7996         }
7997
7998         return err;
7999 }
8000
8001 static int btrfs_subio_endio_read(struct inode *inode,
8002                                   struct btrfs_io_bio *io_bio, int err)
8003 {
8004         bool skip_csum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM;
8005
8006         if (skip_csum) {
8007                 if (unlikely(err))
8008                         return __btrfs_correct_data_nocsum(inode, io_bio);
8009                 else
8010                         return 0;
8011         } else {
8012                 return __btrfs_subio_endio_read(inode, io_bio, err);
8013         }
8014 }
8015
8016 static void btrfs_endio_direct_read(struct bio *bio)
8017 {
8018         struct btrfs_dio_private *dip = bio->bi_private;
8019         struct inode *inode = dip->inode;
8020         struct bio *dio_bio;
8021         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
8022         int err = bio->bi_error;
8023
8024         if (dip->flags & BTRFS_DIO_ORIG_BIO_SUBMITTED)
8025                 err = btrfs_subio_endio_read(inode, io_bio, err);
8026
8027         unlock_extent(&BTRFS_I(inode)->io_tree, dip->logical_offset,
8028                       dip->logical_offset + dip->bytes - 1);
8029         dio_bio = dip->dio_bio;
8030
8031         kfree(dip);
8032
8033         dio_end_io(dio_bio, bio->bi_error);
8034
8035         if (io_bio->end_io)
8036                 io_bio->end_io(io_bio, err);
8037         bio_put(bio);
8038 }
8039
8040 static void btrfs_endio_direct_write_update_ordered(struct inode *inode,
8041                                                     const u64 offset,
8042                                                     const u64 bytes,
8043                                                     const int uptodate)
8044 {
8045         struct btrfs_root *root = BTRFS_I(inode)->root;
8046         struct btrfs_ordered_extent *ordered = NULL;
8047         u64 ordered_offset = offset;
8048         u64 ordered_bytes = bytes;
8049         int ret;
8050
8051 again:
8052         ret = btrfs_dec_test_first_ordered_pending(inode, &ordered,
8053                                                    &ordered_offset,
8054                                                    ordered_bytes,
8055                                                    uptodate);
8056         if (!ret)
8057                 goto out_test;
8058
8059         btrfs_init_work(&ordered->work, btrfs_endio_write_helper,
8060                         finish_ordered_fn, NULL, NULL);
8061         btrfs_queue_work(root->fs_info->endio_write_workers,
8062                          &ordered->work);
8063 out_test:
8064         /*
8065          * our bio might span multiple ordered extents.  If we haven't
8066          * completed the accounting for the whole dio, go back and try again
8067          */
8068         if (ordered_offset < offset + bytes) {
8069                 ordered_bytes = offset + bytes - ordered_offset;
8070                 ordered = NULL;
8071                 goto again;
8072         }
8073 }
8074
8075 static void btrfs_endio_direct_write(struct bio *bio)
8076 {
8077         struct btrfs_dio_private *dip = bio->bi_private;
8078         struct bio *dio_bio = dip->dio_bio;
8079
8080         btrfs_endio_direct_write_update_ordered(dip->inode,
8081                                                 dip->logical_offset,
8082                                                 dip->bytes,
8083                                                 !bio->bi_error);
8084
8085         kfree(dip);
8086
8087         dio_end_io(dio_bio, bio->bi_error);
8088         bio_put(bio);
8089 }
8090
8091 static int __btrfs_submit_bio_start_direct_io(struct inode *inode, int rw,
8092                                     struct bio *bio, int mirror_num,
8093                                     unsigned long bio_flags, u64 offset)
8094 {
8095         int ret;
8096         struct btrfs_root *root = BTRFS_I(inode)->root;
8097         ret = btrfs_csum_one_bio(root, inode, bio, offset, 1);
8098         BUG_ON(ret); /* -ENOMEM */
8099         return 0;
8100 }
8101
8102 static void btrfs_end_dio_bio(struct bio *bio)
8103 {
8104         struct btrfs_dio_private *dip = bio->bi_private;
8105         int err = bio->bi_error;
8106
8107         if (err)
8108                 btrfs_warn(BTRFS_I(dip->inode)->root->fs_info,
8109                            "direct IO failed ino %llu rw %lu sector %#Lx len %u err no %d",
8110                            btrfs_ino(dip->inode), bio->bi_rw,
8111                            (unsigned long long)bio->bi_iter.bi_sector,
8112                            bio->bi_iter.bi_size, err);
8113
8114         if (dip->subio_endio)
8115                 err = dip->subio_endio(dip->inode, btrfs_io_bio(bio), err);
8116
8117         if (err) {
8118                 dip->errors = 1;
8119
8120                 /*
8121                  * before atomic variable goto zero, we must make sure
8122                  * dip->errors is perceived to be set.
8123                  */
8124                 smp_mb__before_atomic();
8125         }
8126
8127         /* if there are more bios still pending for this dio, just exit */
8128         if (!atomic_dec_and_test(&dip->pending_bios))
8129                 goto out;
8130
8131         if (dip->errors) {
8132                 bio_io_error(dip->orig_bio);
8133         } else {
8134                 dip->dio_bio->bi_error = 0;
8135                 bio_endio(dip->orig_bio);
8136         }
8137 out:
8138         bio_put(bio);
8139 }
8140
8141 static struct bio *btrfs_dio_bio_alloc(struct block_device *bdev,
8142                                        u64 first_sector, gfp_t gfp_flags)
8143 {
8144         struct bio *bio;
8145         bio = btrfs_bio_alloc(bdev, first_sector, BIO_MAX_PAGES, gfp_flags);
8146         if (bio)
8147                 bio_associate_current(bio);
8148         return bio;
8149 }
8150
8151 static inline int btrfs_lookup_and_bind_dio_csum(struct btrfs_root *root,
8152                                                  struct inode *inode,
8153                                                  struct btrfs_dio_private *dip,
8154                                                  struct bio *bio,
8155                                                  u64 file_offset)
8156 {
8157         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
8158         struct btrfs_io_bio *orig_io_bio = btrfs_io_bio(dip->orig_bio);
8159         int ret;
8160
8161         /*
8162          * We load all the csum data we need when we submit
8163          * the first bio to reduce the csum tree search and
8164          * contention.
8165          */
8166         if (dip->logical_offset == file_offset) {
8167                 ret = btrfs_lookup_bio_sums_dio(root, inode, dip->orig_bio,
8168                                                 file_offset);
8169                 if (ret)
8170                         return ret;
8171         }
8172
8173         if (bio == dip->orig_bio)
8174                 return 0;
8175
8176         file_offset -= dip->logical_offset;
8177         file_offset >>= inode->i_sb->s_blocksize_bits;
8178         io_bio->csum = (u8 *)(((u32 *)orig_io_bio->csum) + file_offset);
8179
8180         return 0;
8181 }
8182
8183 static inline int __btrfs_submit_dio_bio(struct bio *bio, struct inode *inode,
8184                                          int rw, u64 file_offset, int skip_sum,
8185                                          int async_submit)
8186 {
8187         struct btrfs_dio_private *dip = bio->bi_private;
8188         int write = rw & REQ_WRITE;
8189         struct btrfs_root *root = BTRFS_I(inode)->root;
8190         int ret;
8191
8192         if (async_submit)
8193                 async_submit = !atomic_read(&BTRFS_I(inode)->sync_writers);
8194
8195         bio_get(bio);
8196
8197         if (!write) {
8198                 ret = btrfs_bio_wq_end_io(root->fs_info, bio,
8199                                 BTRFS_WQ_ENDIO_DATA);
8200                 if (ret)
8201                         goto err;
8202         }
8203
8204         if (skip_sum)
8205                 goto map;
8206
8207         if (write && async_submit) {
8208                 ret = btrfs_wq_submit_bio(root->fs_info,
8209                                    inode, rw, bio, 0, 0,
8210                                    file_offset,
8211                                    __btrfs_submit_bio_start_direct_io,
8212                                    __btrfs_submit_bio_done);
8213                 goto err;
8214         } else if (write) {
8215                 /*
8216                  * If we aren't doing async submit, calculate the csum of the
8217                  * bio now.
8218                  */
8219                 ret = btrfs_csum_one_bio(root, inode, bio, file_offset, 1);
8220                 if (ret)
8221                         goto err;
8222         } else {
8223                 ret = btrfs_lookup_and_bind_dio_csum(root, inode, dip, bio,
8224                                                      file_offset);
8225                 if (ret)
8226                         goto err;
8227         }
8228 map:
8229         ret = btrfs_map_bio(root, rw, bio, 0, async_submit);
8230 err:
8231         bio_put(bio);
8232         return ret;
8233 }
8234
8235 static int btrfs_submit_direct_hook(int rw, struct btrfs_dio_private *dip,
8236                                     int skip_sum)
8237 {
8238         struct inode *inode = dip->inode;
8239         struct btrfs_root *root = BTRFS_I(inode)->root;
8240         struct bio *bio;
8241         struct bio *orig_bio = dip->orig_bio;
8242         struct bio_vec *bvec = orig_bio->bi_io_vec;
8243         u64 start_sector = orig_bio->bi_iter.bi_sector;
8244         u64 file_offset = dip->logical_offset;
8245         u64 submit_len = 0;
8246         u64 map_length;
8247         u32 blocksize = root->sectorsize;
8248         int async_submit = 0;
8249         int nr_sectors;
8250         int ret;
8251         int i;
8252
8253         map_length = orig_bio->bi_iter.bi_size;
8254         ret = btrfs_map_block(root->fs_info, rw, start_sector << 9,
8255                               &map_length, NULL, 0);
8256         if (ret)
8257                 return -EIO;
8258
8259         if (map_length >= orig_bio->bi_iter.bi_size) {
8260                 bio = orig_bio;
8261                 dip->flags |= BTRFS_DIO_ORIG_BIO_SUBMITTED;
8262                 goto submit;
8263         }
8264
8265         /* async crcs make it difficult to collect full stripe writes. */
8266         if (btrfs_get_alloc_profile(root, 1) & BTRFS_BLOCK_GROUP_RAID56_MASK)
8267                 async_submit = 0;
8268         else
8269                 async_submit = 1;
8270
8271         bio = btrfs_dio_bio_alloc(orig_bio->bi_bdev, start_sector, GFP_NOFS);
8272         if (!bio)
8273                 return -ENOMEM;
8274
8275         bio->bi_private = dip;
8276         bio->bi_end_io = btrfs_end_dio_bio;
8277         btrfs_io_bio(bio)->logical = file_offset;
8278         atomic_inc(&dip->pending_bios);
8279
8280         while (bvec <= (orig_bio->bi_io_vec + orig_bio->bi_vcnt - 1)) {
8281                 nr_sectors = BTRFS_BYTES_TO_BLKS(root->fs_info, bvec->bv_len);
8282                 i = 0;
8283 next_block:
8284                 if (unlikely(map_length < submit_len + blocksize ||
8285                     bio_add_page(bio, bvec->bv_page, blocksize,
8286                             bvec->bv_offset + (i * blocksize)) < blocksize)) {
8287                         /*
8288                          * inc the count before we submit the bio so
8289                          * we know the end IO handler won't happen before
8290                          * we inc the count. Otherwise, the dip might get freed
8291                          * before we're done setting it up
8292                          */
8293                         atomic_inc(&dip->pending_bios);
8294                         ret = __btrfs_submit_dio_bio(bio, inode, rw,
8295                                                      file_offset, skip_sum,
8296                                                      async_submit);
8297                         if (ret) {
8298                                 bio_put(bio);
8299                                 atomic_dec(&dip->pending_bios);
8300                                 goto out_err;
8301                         }
8302
8303                         start_sector += submit_len >> 9;
8304                         file_offset += submit_len;
8305
8306                         submit_len = 0;
8307
8308                         bio = btrfs_dio_bio_alloc(orig_bio->bi_bdev,
8309                                                   start_sector, GFP_NOFS);
8310                         if (!bio)
8311                                 goto out_err;
8312                         bio->bi_private = dip;
8313                         bio->bi_end_io = btrfs_end_dio_bio;
8314                         btrfs_io_bio(bio)->logical = file_offset;
8315
8316                         map_length = orig_bio->bi_iter.bi_size;
8317                         ret = btrfs_map_block(root->fs_info, rw,
8318                                               start_sector << 9,
8319                                               &map_length, NULL, 0);
8320                         if (ret) {
8321                                 bio_put(bio);
8322                                 goto out_err;
8323                         }
8324
8325                         goto next_block;
8326                 } else {
8327                         submit_len += blocksize;
8328                         if (--nr_sectors) {
8329                                 i++;
8330                                 goto next_block;
8331                         }
8332                         bvec++;
8333                 }
8334         }
8335
8336 submit:
8337         ret = __btrfs_submit_dio_bio(bio, inode, rw, file_offset, skip_sum,
8338                                      async_submit);
8339         if (!ret)
8340                 return 0;
8341
8342         bio_put(bio);
8343 out_err:
8344         dip->errors = 1;
8345         /*
8346          * before atomic variable goto zero, we must
8347          * make sure dip->errors is perceived to be set.
8348          */
8349         smp_mb__before_atomic();
8350         if (atomic_dec_and_test(&dip->pending_bios))
8351                 bio_io_error(dip->orig_bio);
8352
8353         /* bio_end_io() will handle error, so we needn't return it */
8354         return 0;
8355 }
8356
8357 static void btrfs_submit_direct(int rw, struct bio *dio_bio,
8358                                 struct inode *inode, loff_t file_offset)
8359 {
8360         struct btrfs_dio_private *dip = NULL;
8361         struct bio *io_bio = NULL;
8362         struct btrfs_io_bio *btrfs_bio;
8363         int skip_sum;
8364         int write = rw & REQ_WRITE;
8365         int ret = 0;
8366
8367         skip_sum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM;
8368
8369         io_bio = btrfs_bio_clone(dio_bio, GFP_NOFS);
8370         if (!io_bio) {
8371                 ret = -ENOMEM;
8372                 goto free_ordered;
8373         }
8374
8375         dip = kzalloc(sizeof(*dip), GFP_NOFS);
8376         if (!dip) {
8377                 ret = -ENOMEM;
8378                 goto free_ordered;
8379         }
8380
8381         dip->private = dio_bio->bi_private;
8382         dip->inode = inode;
8383         dip->logical_offset = file_offset;
8384         dip->bytes = dio_bio->bi_iter.bi_size;
8385         dip->disk_bytenr = (u64)dio_bio->bi_iter.bi_sector << 9;
8386         io_bio->bi_private = dip;
8387         dip->orig_bio = io_bio;
8388         dip->dio_bio = dio_bio;
8389         atomic_set(&dip->pending_bios, 0);
8390         btrfs_bio = btrfs_io_bio(io_bio);
8391         btrfs_bio->logical = file_offset;
8392
8393         if (write) {
8394                 io_bio->bi_end_io = btrfs_endio_direct_write;
8395         } else {
8396                 io_bio->bi_end_io = btrfs_endio_direct_read;
8397                 dip->subio_endio = btrfs_subio_endio_read;
8398         }
8399
8400         /*
8401          * Reset the range for unsubmitted ordered extents (to a 0 length range)
8402          * even if we fail to submit a bio, because in such case we do the
8403          * corresponding error handling below and it must not be done a second
8404          * time by btrfs_direct_IO().
8405          */
8406         if (write) {
8407                 struct btrfs_dio_data *dio_data = current->journal_info;
8408
8409                 dio_data->unsubmitted_oe_range_end = dip->logical_offset +
8410                         dip->bytes;
8411                 dio_data->unsubmitted_oe_range_start =
8412                         dio_data->unsubmitted_oe_range_end;
8413         }
8414
8415         ret = btrfs_submit_direct_hook(rw, dip, skip_sum);
8416         if (!ret)
8417                 return;
8418
8419         if (btrfs_bio->end_io)
8420                 btrfs_bio->end_io(btrfs_bio, ret);
8421
8422 free_ordered:
8423         /*
8424          * If we arrived here it means either we failed to submit the dip
8425          * or we either failed to clone the dio_bio or failed to allocate the
8426          * dip. If we cloned the dio_bio and allocated the dip, we can just
8427          * call bio_endio against our io_bio so that we get proper resource
8428          * cleanup if we fail to submit the dip, otherwise, we must do the
8429          * same as btrfs_endio_direct_[write|read] because we can't call these
8430          * callbacks - they require an allocated dip and a clone of dio_bio.
8431          */
8432         if (io_bio && dip) {
8433                 io_bio->bi_error = -EIO;
8434                 bio_endio(io_bio);
8435                 /*
8436                  * The end io callbacks free our dip, do the final put on io_bio
8437                  * and all the cleanup and final put for dio_bio (through
8438                  * dio_end_io()).
8439                  */
8440                 dip = NULL;
8441                 io_bio = NULL;
8442         } else {
8443                 if (write)
8444                         btrfs_endio_direct_write_update_ordered(inode,
8445                                                 file_offset,
8446                                                 dio_bio->bi_iter.bi_size,
8447                                                 0);
8448                 else
8449                         unlock_extent(&BTRFS_I(inode)->io_tree, file_offset,
8450                               file_offset + dio_bio->bi_iter.bi_size - 1);
8451
8452                 dio_bio->bi_error = -EIO;
8453                 /*
8454                  * Releases and cleans up our dio_bio, no need to bio_put()
8455                  * nor bio_endio()/bio_io_error() against dio_bio.
8456                  */
8457                 dio_end_io(dio_bio, ret);
8458         }
8459         if (io_bio)
8460                 bio_put(io_bio);
8461         kfree(dip);
8462 }
8463
8464 static ssize_t check_direct_IO(struct btrfs_root *root, struct kiocb *iocb,
8465                         const struct iov_iter *iter, loff_t offset)
8466 {
8467         int seg;
8468         int i;
8469         unsigned blocksize_mask = root->sectorsize - 1;
8470         ssize_t retval = -EINVAL;
8471
8472         if (offset & blocksize_mask)
8473                 goto out;
8474
8475         if (iov_iter_alignment(iter) & blocksize_mask)
8476                 goto out;
8477
8478         /* If this is a write we don't need to check anymore */
8479         if (iov_iter_rw(iter) == WRITE)
8480                 return 0;
8481         /*
8482          * Check to make sure we don't have duplicate iov_base's in this
8483          * iovec, if so return EINVAL, otherwise we'll get csum errors
8484          * when reading back.
8485          */
8486         for (seg = 0; seg < iter->nr_segs; seg++) {
8487                 for (i = seg + 1; i < iter->nr_segs; i++) {
8488                         if (iter->iov[seg].iov_base == iter->iov[i].iov_base)
8489                                 goto out;
8490                 }
8491         }
8492         retval = 0;
8493 out:
8494         return retval;
8495 }
8496
8497 static ssize_t btrfs_direct_IO(struct kiocb *iocb, struct iov_iter *iter,
8498                                loff_t offset)
8499 {
8500         struct file *file = iocb->ki_filp;
8501         struct inode *inode = file->f_mapping->host;
8502         struct btrfs_root *root = BTRFS_I(inode)->root;
8503         struct btrfs_dio_data dio_data = { 0 };
8504         size_t count = 0;
8505         int flags = 0;
8506         bool wakeup = true;
8507         bool relock = false;
8508         ssize_t ret;
8509
8510         if (check_direct_IO(BTRFS_I(inode)->root, iocb, iter, offset))
8511                 return 0;
8512
8513         inode_dio_begin(inode);
8514         smp_mb__after_atomic();
8515
8516         /*
8517          * The generic stuff only does filemap_write_and_wait_range, which
8518          * isn't enough if we've written compressed pages to this area, so
8519          * we need to flush the dirty pages again to make absolutely sure
8520          * that any outstanding dirty pages are on disk.
8521          */
8522         count = iov_iter_count(iter);
8523         if (test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
8524                      &BTRFS_I(inode)->runtime_flags))
8525                 filemap_fdatawrite_range(inode->i_mapping, offset,
8526                                          offset + count - 1);
8527
8528         if (iov_iter_rw(iter) == WRITE) {
8529                 /*
8530                  * If the write DIO is beyond the EOF, we need update
8531                  * the isize, but it is protected by i_mutex. So we can
8532                  * not unlock the i_mutex at this case.
8533                  */
8534                 if (offset + count <= inode->i_size) {
8535                         mutex_unlock(&inode->i_mutex);
8536                         relock = true;
8537                 }
8538                 ret = btrfs_delalloc_reserve_space(inode, offset, count);
8539                 if (ret)
8540                         goto out;
8541                 dio_data.outstanding_extents = div64_u64(count +
8542                                                 BTRFS_MAX_EXTENT_SIZE - 1,
8543                                                 BTRFS_MAX_EXTENT_SIZE);
8544
8545                 /*
8546                  * We need to know how many extents we reserved so that we can
8547                  * do the accounting properly if we go over the number we
8548                  * originally calculated.  Abuse current->journal_info for this.
8549                  */
8550                 dio_data.reserve = round_up(count, root->sectorsize);
8551                 dio_data.unsubmitted_oe_range_start = (u64)offset;
8552                 dio_data.unsubmitted_oe_range_end = (u64)offset;
8553                 current->journal_info = &dio_data;
8554         } else if (test_bit(BTRFS_INODE_READDIO_NEED_LOCK,
8555                                      &BTRFS_I(inode)->runtime_flags)) {
8556                 inode_dio_end(inode);
8557                 flags = DIO_LOCKING | DIO_SKIP_HOLES;
8558                 wakeup = false;
8559         }
8560
8561         ret = __blockdev_direct_IO(iocb, inode,
8562                                    BTRFS_I(inode)->root->fs_info->fs_devices->latest_bdev,
8563                                    iter, offset, btrfs_get_blocks_direct, NULL,
8564                                    btrfs_submit_direct, flags);
8565         if (iov_iter_rw(iter) == WRITE) {
8566                 current->journal_info = NULL;
8567                 if (ret < 0 && ret != -EIOCBQUEUED) {
8568                         if (dio_data.reserve)
8569                                 btrfs_delalloc_release_space(inode, offset,
8570                                                              dio_data.reserve);
8571                         /*
8572                          * On error we might have left some ordered extents
8573                          * without submitting corresponding bios for them, so
8574                          * cleanup them up to avoid other tasks getting them
8575                          * and waiting for them to complete forever.
8576                          */
8577                         if (dio_data.unsubmitted_oe_range_start <
8578                             dio_data.unsubmitted_oe_range_end)
8579                                 btrfs_endio_direct_write_update_ordered(inode,
8580                                         dio_data.unsubmitted_oe_range_start,
8581                                         dio_data.unsubmitted_oe_range_end -
8582                                         dio_data.unsubmitted_oe_range_start,
8583                                         0);
8584                 } else if (ret >= 0 && (size_t)ret < count)
8585                         btrfs_delalloc_release_space(inode, offset,
8586                                                      count - (size_t)ret);
8587         }
8588 out:
8589         if (wakeup)
8590                 inode_dio_end(inode);
8591         if (relock)
8592                 mutex_lock(&inode->i_mutex);
8593
8594         return ret;
8595 }
8596
8597 #define BTRFS_FIEMAP_FLAGS      (FIEMAP_FLAG_SYNC)
8598
8599 static int btrfs_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
8600                 __u64 start, __u64 len)
8601 {
8602         int     ret;
8603
8604         ret = fiemap_check_flags(fieinfo, BTRFS_FIEMAP_FLAGS);
8605         if (ret)
8606                 return ret;
8607
8608         return extent_fiemap(inode, fieinfo, start, len, btrfs_get_extent_fiemap);
8609 }
8610
8611 int btrfs_readpage(struct file *file, struct page *page)
8612 {
8613         struct extent_io_tree *tree;
8614         tree = &BTRFS_I(page->mapping->host)->io_tree;
8615         return extent_read_full_page(tree, page, btrfs_get_extent, 0);
8616 }
8617
8618 static int btrfs_writepage(struct page *page, struct writeback_control *wbc)
8619 {
8620         struct extent_io_tree *tree;
8621         struct inode *inode = page->mapping->host;
8622         int ret;
8623
8624         if (current->flags & PF_MEMALLOC) {
8625                 redirty_page_for_writepage(wbc, page);
8626                 unlock_page(page);
8627                 return 0;
8628         }
8629
8630         /*
8631          * If we are under memory pressure we will call this directly from the
8632          * VM, we need to make sure we have the inode referenced for the ordered
8633          * extent.  If not just return like we didn't do anything.
8634          */
8635         if (!igrab(inode)) {
8636                 redirty_page_for_writepage(wbc, page);
8637                 return AOP_WRITEPAGE_ACTIVATE;
8638         }
8639         tree = &BTRFS_I(page->mapping->host)->io_tree;
8640         ret = extent_write_full_page(tree, page, btrfs_get_extent, wbc);
8641         btrfs_add_delayed_iput(inode);
8642         return ret;
8643 }
8644
8645 static int btrfs_writepages(struct address_space *mapping,
8646                             struct writeback_control *wbc)
8647 {
8648         struct extent_io_tree *tree;
8649
8650         tree = &BTRFS_I(mapping->host)->io_tree;
8651         return extent_writepages(tree, mapping, btrfs_get_extent, wbc);
8652 }
8653
8654 static int
8655 btrfs_readpages(struct file *file, struct address_space *mapping,
8656                 struct list_head *pages, unsigned nr_pages)
8657 {
8658         struct extent_io_tree *tree;
8659         tree = &BTRFS_I(mapping->host)->io_tree;
8660         return extent_readpages(tree, mapping, pages, nr_pages,
8661                                 btrfs_get_extent);
8662 }
8663 static int __btrfs_releasepage(struct page *page, gfp_t gfp_flags)
8664 {
8665         struct extent_io_tree *tree;
8666         struct extent_map_tree *map;
8667         int ret;
8668
8669         tree = &BTRFS_I(page->mapping->host)->io_tree;
8670         map = &BTRFS_I(page->mapping->host)->extent_tree;
8671         ret = try_release_extent_mapping(map, tree, page, gfp_flags);
8672         if (ret == 1) {
8673                 ClearPagePrivate(page);
8674                 set_page_private(page, 0);
8675                 page_cache_release(page);
8676         }
8677         return ret;
8678 }
8679
8680 static int btrfs_releasepage(struct page *page, gfp_t gfp_flags)
8681 {
8682         if (PageWriteback(page) || PageDirty(page))
8683                 return 0;
8684         return __btrfs_releasepage(page, gfp_flags & GFP_NOFS);
8685 }
8686
8687 static void btrfs_invalidatepage(struct page *page, unsigned int offset,
8688                                  unsigned int length)
8689 {
8690         struct inode *inode = page->mapping->host;
8691         struct extent_io_tree *tree;
8692         struct btrfs_ordered_extent *ordered;
8693         struct extent_state *cached_state = NULL;
8694         u64 page_start = page_offset(page);
8695         u64 page_end = page_start + PAGE_CACHE_SIZE - 1;
8696         u64 start;
8697         u64 end;
8698         int inode_evicting = inode->i_state & I_FREEING;
8699
8700         /*
8701          * we have the page locked, so new writeback can't start,
8702          * and the dirty bit won't be cleared while we are here.
8703          *
8704          * Wait for IO on this page so that we can safely clear
8705          * the PagePrivate2 bit and do ordered accounting
8706          */
8707         wait_on_page_writeback(page);
8708
8709         tree = &BTRFS_I(inode)->io_tree;
8710         if (offset) {
8711                 btrfs_releasepage(page, GFP_NOFS);
8712                 return;
8713         }
8714
8715         if (!inode_evicting)
8716                 lock_extent_bits(tree, page_start, page_end, &cached_state);
8717 again:
8718         start = page_start;
8719         ordered = btrfs_lookup_ordered_range(inode, start,
8720                                         page_end - start + 1);
8721         if (ordered) {
8722                 end = min(page_end, ordered->file_offset + ordered->len - 1);
8723                 /*
8724                  * IO on this page will never be started, so we need
8725                  * to account for any ordered extents now
8726                  */
8727                 if (!inode_evicting)
8728                         clear_extent_bit(tree, start, end,
8729                                          EXTENT_DIRTY | EXTENT_DELALLOC |
8730                                          EXTENT_LOCKED | EXTENT_DO_ACCOUNTING |
8731                                          EXTENT_DEFRAG, 1, 0, &cached_state,
8732                                          GFP_NOFS);
8733                 /*
8734                  * whoever cleared the private bit is responsible
8735                  * for the finish_ordered_io
8736                  */
8737                 if (TestClearPagePrivate2(page)) {
8738                         struct btrfs_ordered_inode_tree *tree;
8739                         u64 new_len;
8740
8741                         tree = &BTRFS_I(inode)->ordered_tree;
8742
8743                         spin_lock_irq(&tree->lock);
8744                         set_bit(BTRFS_ORDERED_TRUNCATED, &ordered->flags);
8745                         new_len = start - ordered->file_offset;
8746                         if (new_len < ordered->truncated_len)
8747                                 ordered->truncated_len = new_len;
8748                         spin_unlock_irq(&tree->lock);
8749
8750                         if (btrfs_dec_test_ordered_pending(inode, &ordered,
8751                                                            start,
8752                                                            end - start + 1, 1))
8753                                 btrfs_finish_ordered_io(ordered);
8754                 }
8755                 btrfs_put_ordered_extent(ordered);
8756                 if (!inode_evicting) {
8757                         cached_state = NULL;
8758                         lock_extent_bits(tree, start, end,
8759                                          &cached_state);
8760                 }
8761
8762                 start = end + 1;
8763                 if (start < page_end)
8764                         goto again;
8765         }
8766
8767         /*
8768          * Qgroup reserved space handler
8769          * Page here will be either
8770          * 1) Already written to disk
8771          *    In this case, its reserved space is released from data rsv map
8772          *    and will be freed by delayed_ref handler finally.
8773          *    So even we call qgroup_free_data(), it won't decrease reserved
8774          *    space.
8775          * 2) Not written to disk
8776          *    This means the reserved space should be freed here.
8777          */
8778         btrfs_qgroup_free_data(inode, page_start, PAGE_CACHE_SIZE);
8779         if (!inode_evicting) {
8780                 clear_extent_bit(tree, page_start, page_end,
8781                                  EXTENT_LOCKED | EXTENT_DIRTY |
8782                                  EXTENT_DELALLOC | EXTENT_DO_ACCOUNTING |
8783                                  EXTENT_DEFRAG, 1, 1,
8784                                  &cached_state, GFP_NOFS);
8785
8786                 __btrfs_releasepage(page, GFP_NOFS);
8787         }
8788
8789         ClearPageChecked(page);
8790         if (PagePrivate(page)) {
8791                 ClearPagePrivate(page);
8792                 set_page_private(page, 0);
8793                 page_cache_release(page);
8794         }
8795 }
8796
8797 /*
8798  * btrfs_page_mkwrite() is not allowed to change the file size as it gets
8799  * called from a page fault handler when a page is first dirtied. Hence we must
8800  * be careful to check for EOF conditions here. We set the page up correctly
8801  * for a written page which means we get ENOSPC checking when writing into
8802  * holes and correct delalloc and unwritten extent mapping on filesystems that
8803  * support these features.
8804  *
8805  * We are not allowed to take the i_mutex here so we have to play games to
8806  * protect against truncate races as the page could now be beyond EOF.  Because
8807  * vmtruncate() writes the inode size before removing pages, once we have the
8808  * page lock we can determine safely if the page is beyond EOF. If it is not
8809  * beyond EOF, then the page is guaranteed safe against truncation until we
8810  * unlock the page.
8811  */
8812 int btrfs_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
8813 {
8814         struct page *page = vmf->page;
8815         struct inode *inode = file_inode(vma->vm_file);
8816         struct btrfs_root *root = BTRFS_I(inode)->root;
8817         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
8818         struct btrfs_ordered_extent *ordered;
8819         struct extent_state *cached_state = NULL;
8820         char *kaddr;
8821         unsigned long zero_start;
8822         loff_t size;
8823         int ret;
8824         int reserved = 0;
8825         u64 reserved_space;
8826         u64 page_start;
8827         u64 page_end;
8828         u64 end;
8829
8830         reserved_space = PAGE_CACHE_SIZE;
8831
8832         sb_start_pagefault(inode->i_sb);
8833         page_start = page_offset(page);
8834         page_end = page_start + PAGE_CACHE_SIZE - 1;
8835         end = page_end;
8836
8837         /*
8838          * Reserving delalloc space after obtaining the page lock can lead to
8839          * deadlock. For example, if a dirty page is locked by this function
8840          * and the call to btrfs_delalloc_reserve_space() ends up triggering
8841          * dirty page write out, then the btrfs_writepage() function could
8842          * end up waiting indefinitely to get a lock on the page currently
8843          * being processed by btrfs_page_mkwrite() function.
8844          */
8845         ret = btrfs_delalloc_reserve_space(inode, page_start,
8846                                            reserved_space);
8847         if (!ret) {
8848                 ret = file_update_time(vma->vm_file);
8849                 reserved = 1;
8850         }
8851         if (ret) {
8852                 if (ret == -ENOMEM)
8853                         ret = VM_FAULT_OOM;
8854                 else /* -ENOSPC, -EIO, etc */
8855                         ret = VM_FAULT_SIGBUS;
8856                 if (reserved)
8857                         goto out;
8858                 goto out_noreserve;
8859         }
8860
8861         ret = VM_FAULT_NOPAGE; /* make the VM retry the fault */
8862 again:
8863         lock_page(page);
8864         size = i_size_read(inode);
8865
8866         if ((page->mapping != inode->i_mapping) ||
8867             (page_start >= size)) {
8868                 /* page got truncated out from underneath us */
8869                 goto out_unlock;
8870         }
8871         wait_on_page_writeback(page);
8872
8873         lock_extent_bits(io_tree, page_start, page_end, &cached_state);
8874         set_page_extent_mapped(page);
8875
8876         /*
8877          * we can't set the delalloc bits if there are pending ordered
8878          * extents.  Drop our locks and wait for them to finish
8879          */
8880         ordered = btrfs_lookup_ordered_range(inode, page_start, page_end);
8881         if (ordered) {
8882                 unlock_extent_cached(io_tree, page_start, page_end,
8883                                      &cached_state, GFP_NOFS);
8884                 unlock_page(page);
8885                 btrfs_start_ordered_extent(inode, ordered, 1);
8886                 btrfs_put_ordered_extent(ordered);
8887                 goto again;
8888         }
8889
8890         if (page->index == ((size - 1) >> PAGE_CACHE_SHIFT)) {
8891                 reserved_space = round_up(size - page_start, root->sectorsize);
8892                 if (reserved_space < PAGE_CACHE_SIZE) {
8893                         end = page_start + reserved_space - 1;
8894                         spin_lock(&BTRFS_I(inode)->lock);
8895                         BTRFS_I(inode)->outstanding_extents++;
8896                         spin_unlock(&BTRFS_I(inode)->lock);
8897                         btrfs_delalloc_release_space(inode, page_start,
8898                                                 PAGE_CACHE_SIZE - reserved_space);
8899                 }
8900         }
8901
8902         /*
8903          * XXX - page_mkwrite gets called every time the page is dirtied, even
8904          * if it was already dirty, so for space accounting reasons we need to
8905          * clear any delalloc bits for the range we are fixing to save.  There
8906          * is probably a better way to do this, but for now keep consistent with
8907          * prepare_pages in the normal write path.
8908          */
8909         clear_extent_bit(&BTRFS_I(inode)->io_tree, page_start, end,
8910                           EXTENT_DIRTY | EXTENT_DELALLOC |
8911                           EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG,
8912                           0, 0, &cached_state, GFP_NOFS);
8913
8914         ret = btrfs_set_extent_delalloc(inode, page_start, end,
8915                                         &cached_state);
8916         if (ret) {
8917                 unlock_extent_cached(io_tree, page_start, page_end,
8918                                      &cached_state, GFP_NOFS);
8919                 ret = VM_FAULT_SIGBUS;
8920                 goto out_unlock;
8921         }
8922         ret = 0;
8923
8924         /* page is wholly or partially inside EOF */
8925         if (page_start + PAGE_CACHE_SIZE > size)
8926                 zero_start = size & ~PAGE_CACHE_MASK;
8927         else
8928                 zero_start = PAGE_CACHE_SIZE;
8929
8930         if (zero_start != PAGE_CACHE_SIZE) {
8931                 kaddr = kmap(page);
8932                 memset(kaddr + zero_start, 0, PAGE_CACHE_SIZE - zero_start);
8933                 flush_dcache_page(page);
8934                 kunmap(page);
8935         }
8936         ClearPageChecked(page);
8937         set_page_dirty(page);
8938         SetPageUptodate(page);
8939
8940         BTRFS_I(inode)->last_trans = root->fs_info->generation;
8941         BTRFS_I(inode)->last_sub_trans = BTRFS_I(inode)->root->log_transid;
8942         BTRFS_I(inode)->last_log_commit = BTRFS_I(inode)->root->last_log_commit;
8943
8944         unlock_extent_cached(io_tree, page_start, page_end, &cached_state, GFP_NOFS);
8945
8946 out_unlock:
8947         if (!ret) {
8948                 sb_end_pagefault(inode->i_sb);
8949                 return VM_FAULT_LOCKED;
8950         }
8951         unlock_page(page);
8952 out:
8953         btrfs_delalloc_release_space(inode, page_start, reserved_space);
8954 out_noreserve:
8955         sb_end_pagefault(inode->i_sb);
8956         return ret;
8957 }
8958
8959 static int btrfs_truncate(struct inode *inode)
8960 {
8961         struct btrfs_root *root = BTRFS_I(inode)->root;
8962         struct btrfs_block_rsv *rsv;
8963         int ret = 0;
8964         int err = 0;
8965         struct btrfs_trans_handle *trans;
8966         u64 mask = root->sectorsize - 1;
8967         u64 min_size = btrfs_calc_trunc_metadata_size(root, 1);
8968
8969         ret = btrfs_wait_ordered_range(inode, inode->i_size & (~mask),
8970                                        (u64)-1);
8971         if (ret)
8972                 return ret;
8973
8974         /*
8975          * Yes ladies and gentelment, this is indeed ugly.  The fact is we have
8976          * 3 things going on here
8977          *
8978          * 1) We need to reserve space for our orphan item and the space to
8979          * delete our orphan item.  Lord knows we don't want to have a dangling
8980          * orphan item because we didn't reserve space to remove it.
8981          *
8982          * 2) We need to reserve space to update our inode.
8983          *
8984          * 3) We need to have something to cache all the space that is going to
8985          * be free'd up by the truncate operation, but also have some slack
8986          * space reserved in case it uses space during the truncate (thank you
8987          * very much snapshotting).
8988          *
8989          * And we need these to all be seperate.  The fact is we can use alot of
8990          * space doing the truncate, and we have no earthly idea how much space
8991          * we will use, so we need the truncate reservation to be seperate so it
8992          * doesn't end up using space reserved for updating the inode or
8993          * removing the orphan item.  We also need to be able to stop the
8994          * transaction and start a new one, which means we need to be able to
8995          * update the inode several times, and we have no idea of knowing how
8996          * many times that will be, so we can't just reserve 1 item for the
8997          * entirety of the opration, so that has to be done seperately as well.
8998          * Then there is the orphan item, which does indeed need to be held on
8999          * to for the whole operation, and we need nobody to touch this reserved
9000          * space except the orphan code.
9001          *
9002          * So that leaves us with
9003          *
9004          * 1) root->orphan_block_rsv - for the orphan deletion.
9005          * 2) rsv - for the truncate reservation, which we will steal from the
9006          * transaction reservation.
9007          * 3) fs_info->trans_block_rsv - this will have 1 items worth left for
9008          * updating the inode.
9009          */
9010         rsv = btrfs_alloc_block_rsv(root, BTRFS_BLOCK_RSV_TEMP);
9011         if (!rsv)
9012                 return -ENOMEM;
9013         rsv->size = min_size;
9014         rsv->failfast = 1;
9015
9016         /*
9017          * 1 for the truncate slack space
9018          * 1 for updating the inode.
9019          */
9020         trans = btrfs_start_transaction(root, 2);
9021         if (IS_ERR(trans)) {
9022                 err = PTR_ERR(trans);
9023                 goto out;
9024         }
9025
9026         /* Migrate the slack space for the truncate to our reserve */
9027         ret = btrfs_block_rsv_migrate(&root->fs_info->trans_block_rsv, rsv,
9028                                       min_size);
9029         BUG_ON(ret);
9030
9031         /*
9032          * So if we truncate and then write and fsync we normally would just
9033          * write the extents that changed, which is a problem if we need to
9034          * first truncate that entire inode.  So set this flag so we write out
9035          * all of the extents in the inode to the sync log so we're completely
9036          * safe.
9037          */
9038         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &BTRFS_I(inode)->runtime_flags);
9039         trans->block_rsv = rsv;
9040
9041         while (1) {
9042                 ret = btrfs_truncate_inode_items(trans, root, inode,
9043                                                  inode->i_size,
9044                                                  BTRFS_EXTENT_DATA_KEY);
9045                 if (ret != -ENOSPC && ret != -EAGAIN) {
9046                         err = ret;
9047                         break;
9048                 }
9049
9050                 trans->block_rsv = &root->fs_info->trans_block_rsv;
9051                 ret = btrfs_update_inode(trans, root, inode);
9052                 if (ret) {
9053                         err = ret;
9054                         break;
9055                 }
9056
9057                 btrfs_end_transaction(trans, root);
9058                 btrfs_btree_balance_dirty(root);
9059
9060                 trans = btrfs_start_transaction(root, 2);
9061                 if (IS_ERR(trans)) {
9062                         ret = err = PTR_ERR(trans);
9063                         trans = NULL;
9064                         break;
9065                 }
9066
9067                 ret = btrfs_block_rsv_migrate(&root->fs_info->trans_block_rsv,
9068                                               rsv, min_size);
9069                 BUG_ON(ret);    /* shouldn't happen */
9070                 trans->block_rsv = rsv;
9071         }
9072
9073         if (ret == 0 && inode->i_nlink > 0) {
9074                 trans->block_rsv = root->orphan_block_rsv;
9075                 ret = btrfs_orphan_del(trans, inode);
9076                 if (ret)
9077                         err = ret;
9078         }
9079
9080         if (trans) {
9081                 trans->block_rsv = &root->fs_info->trans_block_rsv;
9082                 ret = btrfs_update_inode(trans, root, inode);
9083                 if (ret && !err)
9084                         err = ret;
9085
9086                 ret = btrfs_end_transaction(trans, root);
9087                 btrfs_btree_balance_dirty(root);
9088         }
9089
9090 out:
9091         btrfs_free_block_rsv(root, rsv);
9092
9093         if (ret && !err)
9094                 err = ret;
9095
9096         return err;
9097 }
9098
9099 /*
9100  * create a new subvolume directory/inode (helper for the ioctl).
9101  */
9102 int btrfs_create_subvol_root(struct btrfs_trans_handle *trans,
9103                              struct btrfs_root *new_root,
9104                              struct btrfs_root *parent_root,
9105                              u64 new_dirid)
9106 {
9107         struct inode *inode;
9108         int err;
9109         u64 index = 0;
9110
9111         inode = btrfs_new_inode(trans, new_root, NULL, "..", 2,
9112                                 new_dirid, new_dirid,
9113                                 S_IFDIR | (~current_umask() & S_IRWXUGO),
9114                                 &index);
9115         if (IS_ERR(inode))
9116                 return PTR_ERR(inode);
9117         inode->i_op = &btrfs_dir_inode_operations;
9118         inode->i_fop = &btrfs_dir_file_operations;
9119
9120         set_nlink(inode, 1);
9121         btrfs_i_size_write(inode, 0);
9122         unlock_new_inode(inode);
9123
9124         err = btrfs_subvol_inherit_props(trans, new_root, parent_root);
9125         if (err)
9126                 btrfs_err(new_root->fs_info,
9127                           "error inheriting subvolume %llu properties: %d",
9128                           new_root->root_key.objectid, err);
9129
9130         err = btrfs_update_inode(trans, new_root, inode);
9131
9132         iput(inode);
9133         return err;
9134 }
9135
9136 struct inode *btrfs_alloc_inode(struct super_block *sb)
9137 {
9138         struct btrfs_inode *ei;
9139         struct inode *inode;
9140
9141         ei = kmem_cache_alloc(btrfs_inode_cachep, GFP_NOFS);
9142         if (!ei)
9143                 return NULL;
9144
9145         ei->root = NULL;
9146         ei->generation = 0;
9147         ei->last_trans = 0;
9148         ei->last_sub_trans = 0;
9149         ei->logged_trans = 0;
9150         ei->delalloc_bytes = 0;
9151         ei->defrag_bytes = 0;
9152         ei->disk_i_size = 0;
9153         ei->flags = 0;
9154         ei->csum_bytes = 0;
9155         ei->index_cnt = (u64)-1;
9156         ei->dir_index = 0;
9157         ei->last_unlink_trans = 0;
9158         ei->last_log_commit = 0;
9159         ei->delayed_iput_count = 0;
9160
9161         spin_lock_init(&ei->lock);
9162         ei->outstanding_extents = 0;
9163         ei->reserved_extents = 0;
9164
9165         ei->runtime_flags = 0;
9166         ei->force_compress = BTRFS_COMPRESS_NONE;
9167
9168         ei->delayed_node = NULL;
9169
9170         ei->i_otime.tv_sec = 0;
9171         ei->i_otime.tv_nsec = 0;
9172
9173         inode = &ei->vfs_inode;
9174         extent_map_tree_init(&ei->extent_tree);
9175         extent_io_tree_init(&ei->io_tree, &inode->i_data);
9176         extent_io_tree_init(&ei->io_failure_tree, &inode->i_data);
9177         ei->io_tree.track_uptodate = 1;
9178         ei->io_failure_tree.track_uptodate = 1;
9179         atomic_set(&ei->sync_writers, 0);
9180         mutex_init(&ei->log_mutex);
9181         mutex_init(&ei->delalloc_mutex);
9182         btrfs_ordered_inode_tree_init(&ei->ordered_tree);
9183         INIT_LIST_HEAD(&ei->delalloc_inodes);
9184         INIT_LIST_HEAD(&ei->delayed_iput);
9185         RB_CLEAR_NODE(&ei->rb_node);
9186
9187         return inode;
9188 }
9189
9190 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
9191 void btrfs_test_destroy_inode(struct inode *inode)
9192 {
9193         btrfs_drop_extent_cache(inode, 0, (u64)-1, 0);
9194         kmem_cache_free(btrfs_inode_cachep, BTRFS_I(inode));
9195 }
9196 #endif
9197
9198 static void btrfs_i_callback(struct rcu_head *head)
9199 {
9200         struct inode *inode = container_of(head, struct inode, i_rcu);
9201         kmem_cache_free(btrfs_inode_cachep, BTRFS_I(inode));
9202 }
9203
9204 void btrfs_destroy_inode(struct inode *inode)
9205 {
9206         struct btrfs_ordered_extent *ordered;
9207         struct btrfs_root *root = BTRFS_I(inode)->root;
9208
9209         WARN_ON(!hlist_empty(&inode->i_dentry));
9210         WARN_ON(inode->i_data.nrpages);
9211         WARN_ON(BTRFS_I(inode)->outstanding_extents);
9212         WARN_ON(BTRFS_I(inode)->reserved_extents);
9213         WARN_ON(BTRFS_I(inode)->delalloc_bytes);
9214         WARN_ON(BTRFS_I(inode)->csum_bytes);
9215         WARN_ON(BTRFS_I(inode)->defrag_bytes);
9216
9217         /*
9218          * This can happen where we create an inode, but somebody else also
9219          * created the same inode and we need to destroy the one we already
9220          * created.
9221          */
9222         if (!root)
9223                 goto free;
9224
9225         if (test_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
9226                      &BTRFS_I(inode)->runtime_flags)) {
9227                 btrfs_info(root->fs_info, "inode %llu still on the orphan list",
9228                         btrfs_ino(inode));
9229                 atomic_dec(&root->orphan_inodes);
9230         }
9231
9232         while (1) {
9233                 ordered = btrfs_lookup_first_ordered_extent(inode, (u64)-1);
9234                 if (!ordered)
9235                         break;
9236                 else {
9237                         btrfs_err(root->fs_info, "found ordered extent %llu %llu on inode cleanup",
9238                                 ordered->file_offset, ordered->len);
9239                         btrfs_remove_ordered_extent(inode, ordered);
9240                         btrfs_put_ordered_extent(ordered);
9241                         btrfs_put_ordered_extent(ordered);
9242                 }
9243         }
9244         btrfs_qgroup_check_reserved_leak(inode);
9245         inode_tree_del(inode);
9246         btrfs_drop_extent_cache(inode, 0, (u64)-1, 0);
9247 free:
9248         call_rcu(&inode->i_rcu, btrfs_i_callback);
9249 }
9250
9251 int btrfs_drop_inode(struct inode *inode)
9252 {
9253         struct btrfs_root *root = BTRFS_I(inode)->root;
9254
9255         if (root == NULL)
9256                 return 1;
9257
9258         /* the snap/subvol tree is on deleting */
9259         if (btrfs_root_refs(&root->root_item) == 0)
9260                 return 1;
9261         else
9262                 return generic_drop_inode(inode);
9263 }
9264
9265 static void init_once(void *foo)
9266 {
9267         struct btrfs_inode *ei = (struct btrfs_inode *) foo;
9268
9269         inode_init_once(&ei->vfs_inode);
9270 }
9271
9272 void btrfs_destroy_cachep(void)
9273 {
9274         /*
9275          * Make sure all delayed rcu free inodes are flushed before we
9276          * destroy cache.
9277          */
9278         rcu_barrier();
9279         if (btrfs_inode_cachep)
9280                 kmem_cache_destroy(btrfs_inode_cachep);
9281         if (btrfs_trans_handle_cachep)
9282                 kmem_cache_destroy(btrfs_trans_handle_cachep);
9283         if (btrfs_transaction_cachep)
9284                 kmem_cache_destroy(btrfs_transaction_cachep);
9285         if (btrfs_path_cachep)
9286                 kmem_cache_destroy(btrfs_path_cachep);
9287         if (btrfs_free_space_cachep)
9288                 kmem_cache_destroy(btrfs_free_space_cachep);
9289 }
9290
9291 int btrfs_init_cachep(void)
9292 {
9293         btrfs_inode_cachep = kmem_cache_create("btrfs_inode",
9294                         sizeof(struct btrfs_inode), 0,
9295                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, init_once);
9296         if (!btrfs_inode_cachep)
9297                 goto fail;
9298
9299         btrfs_trans_handle_cachep = kmem_cache_create("btrfs_trans_handle",
9300                         sizeof(struct btrfs_trans_handle), 0,
9301                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, NULL);
9302         if (!btrfs_trans_handle_cachep)
9303                 goto fail;
9304
9305         btrfs_transaction_cachep = kmem_cache_create("btrfs_transaction",
9306                         sizeof(struct btrfs_transaction), 0,
9307                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, NULL);
9308         if (!btrfs_transaction_cachep)
9309                 goto fail;
9310
9311         btrfs_path_cachep = kmem_cache_create("btrfs_path",
9312                         sizeof(struct btrfs_path), 0,
9313                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, NULL);
9314         if (!btrfs_path_cachep)
9315                 goto fail;
9316
9317         btrfs_free_space_cachep = kmem_cache_create("btrfs_free_space",
9318                         sizeof(struct btrfs_free_space), 0,
9319                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, NULL);
9320         if (!btrfs_free_space_cachep)
9321                 goto fail;
9322
9323         return 0;
9324 fail:
9325         btrfs_destroy_cachep();
9326         return -ENOMEM;
9327 }
9328
9329 static int btrfs_getattr(struct vfsmount *mnt,
9330                          struct dentry *dentry, struct kstat *stat)
9331 {
9332         u64 delalloc_bytes;
9333         struct inode *inode = d_inode(dentry);
9334         u32 blocksize = inode->i_sb->s_blocksize;
9335
9336         generic_fillattr(inode, stat);
9337         stat->dev = BTRFS_I(inode)->root->anon_dev;
9338         stat->blksize = PAGE_CACHE_SIZE;
9339
9340         spin_lock(&BTRFS_I(inode)->lock);
9341         delalloc_bytes = BTRFS_I(inode)->delalloc_bytes;
9342         spin_unlock(&BTRFS_I(inode)->lock);
9343         stat->blocks = (ALIGN(inode_get_bytes(inode), blocksize) +
9344                         ALIGN(delalloc_bytes, blocksize)) >> 9;
9345         return 0;
9346 }
9347
9348 static int btrfs_rename(struct inode *old_dir, struct dentry *old_dentry,
9349                            struct inode *new_dir, struct dentry *new_dentry)
9350 {
9351         struct btrfs_trans_handle *trans;
9352         struct btrfs_root *root = BTRFS_I(old_dir)->root;
9353         struct btrfs_root *dest = BTRFS_I(new_dir)->root;
9354         struct inode *new_inode = d_inode(new_dentry);
9355         struct inode *old_inode = d_inode(old_dentry);
9356         struct timespec ctime = CURRENT_TIME;
9357         u64 index = 0;
9358         u64 root_objectid;
9359         int ret;
9360         u64 old_ino = btrfs_ino(old_inode);
9361
9362         if (btrfs_ino(new_dir) == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)
9363                 return -EPERM;
9364
9365         /* we only allow rename subvolume link between subvolumes */
9366         if (old_ino != BTRFS_FIRST_FREE_OBJECTID && root != dest)
9367                 return -EXDEV;
9368
9369         if (old_ino == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID ||
9370             (new_inode && btrfs_ino(new_inode) == BTRFS_FIRST_FREE_OBJECTID))
9371                 return -ENOTEMPTY;
9372
9373         if (S_ISDIR(old_inode->i_mode) && new_inode &&
9374             new_inode->i_size > BTRFS_EMPTY_DIR_SIZE)
9375                 return -ENOTEMPTY;
9376
9377
9378         /* check for collisions, even if the  name isn't there */
9379         ret = btrfs_check_dir_item_collision(dest, new_dir->i_ino,
9380                              new_dentry->d_name.name,
9381                              new_dentry->d_name.len);
9382
9383         if (ret) {
9384                 if (ret == -EEXIST) {
9385                         /* we shouldn't get
9386                          * eexist without a new_inode */
9387                         if (WARN_ON(!new_inode)) {
9388                                 return ret;
9389                         }
9390                 } else {
9391                         /* maybe -EOVERFLOW */
9392                         return ret;
9393                 }
9394         }
9395         ret = 0;
9396
9397         /*
9398          * we're using rename to replace one file with another.  Start IO on it
9399          * now so  we don't add too much work to the end of the transaction
9400          */
9401         if (new_inode && S_ISREG(old_inode->i_mode) && new_inode->i_size)
9402                 filemap_flush(old_inode->i_mapping);
9403
9404         /* close the racy window with snapshot create/destroy ioctl */
9405         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
9406                 down_read(&root->fs_info->subvol_sem);
9407         /*
9408          * We want to reserve the absolute worst case amount of items.  So if
9409          * both inodes are subvols and we need to unlink them then that would
9410          * require 4 item modifications, but if they are both normal inodes it
9411          * would require 5 item modifications, so we'll assume their normal
9412          * inodes.  So 5 * 2 is 10, plus 1 for the new link, so 11 total items
9413          * should cover the worst case number of items we'll modify.
9414          */
9415         trans = btrfs_start_transaction(root, 11);
9416         if (IS_ERR(trans)) {
9417                 ret = PTR_ERR(trans);
9418                 goto out_notrans;
9419         }
9420
9421         if (dest != root)
9422                 btrfs_record_root_in_trans(trans, dest);
9423
9424         ret = btrfs_set_inode_index(new_dir, &index);
9425         if (ret)
9426                 goto out_fail;
9427
9428         BTRFS_I(old_inode)->dir_index = 0ULL;
9429         if (unlikely(old_ino == BTRFS_FIRST_FREE_OBJECTID)) {
9430                 /* force full log commit if subvolume involved. */
9431                 btrfs_set_log_full_commit(root->fs_info, trans);
9432         } else {
9433                 ret = btrfs_insert_inode_ref(trans, dest,
9434                                              new_dentry->d_name.name,
9435                                              new_dentry->d_name.len,
9436                                              old_ino,
9437                                              btrfs_ino(new_dir), index);
9438                 if (ret)
9439                         goto out_fail;
9440                 /*
9441                  * this is an ugly little race, but the rename is required
9442                  * to make sure that if we crash, the inode is either at the
9443                  * old name or the new one.  pinning the log transaction lets
9444                  * us make sure we don't allow a log commit to come in after
9445                  * we unlink the name but before we add the new name back in.
9446                  */
9447                 btrfs_pin_log_trans(root);
9448         }
9449
9450         inode_inc_iversion(old_dir);
9451         inode_inc_iversion(new_dir);
9452         inode_inc_iversion(old_inode);
9453         old_dir->i_ctime = old_dir->i_mtime = ctime;
9454         new_dir->i_ctime = new_dir->i_mtime = ctime;
9455         old_inode->i_ctime = ctime;
9456
9457         if (old_dentry->d_parent != new_dentry->d_parent)
9458                 btrfs_record_unlink_dir(trans, old_dir, old_inode, 1);
9459
9460         if (unlikely(old_ino == BTRFS_FIRST_FREE_OBJECTID)) {
9461                 root_objectid = BTRFS_I(old_inode)->root->root_key.objectid;
9462                 ret = btrfs_unlink_subvol(trans, root, old_dir, root_objectid,
9463                                         old_dentry->d_name.name,
9464                                         old_dentry->d_name.len);
9465         } else {
9466                 ret = __btrfs_unlink_inode(trans, root, old_dir,
9467                                         d_inode(old_dentry),
9468                                         old_dentry->d_name.name,
9469                                         old_dentry->d_name.len);
9470                 if (!ret)
9471                         ret = btrfs_update_inode(trans, root, old_inode);
9472         }
9473         if (ret) {
9474                 btrfs_abort_transaction(trans, root, ret);
9475                 goto out_fail;
9476         }
9477
9478         if (new_inode) {
9479                 inode_inc_iversion(new_inode);
9480                 new_inode->i_ctime = CURRENT_TIME;
9481                 if (unlikely(btrfs_ino(new_inode) ==
9482                              BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)) {
9483                         root_objectid = BTRFS_I(new_inode)->location.objectid;
9484                         ret = btrfs_unlink_subvol(trans, dest, new_dir,
9485                                                 root_objectid,
9486                                                 new_dentry->d_name.name,
9487                                                 new_dentry->d_name.len);
9488                         BUG_ON(new_inode->i_nlink == 0);
9489                 } else {
9490                         ret = btrfs_unlink_inode(trans, dest, new_dir,
9491                                                  d_inode(new_dentry),
9492                                                  new_dentry->d_name.name,
9493                                                  new_dentry->d_name.len);
9494                 }
9495                 if (!ret && new_inode->i_nlink == 0)
9496                         ret = btrfs_orphan_add(trans, d_inode(new_dentry));
9497                 if (ret) {
9498                         btrfs_abort_transaction(trans, root, ret);
9499                         goto out_fail;
9500                 }
9501         }
9502
9503         ret = btrfs_add_link(trans, new_dir, old_inode,
9504                              new_dentry->d_name.name,
9505                              new_dentry->d_name.len, 0, index);
9506         if (ret) {
9507                 btrfs_abort_transaction(trans, root, ret);
9508                 goto out_fail;
9509         }
9510
9511         if (old_inode->i_nlink == 1)
9512                 BTRFS_I(old_inode)->dir_index = index;
9513
9514         if (old_ino != BTRFS_FIRST_FREE_OBJECTID) {
9515                 struct dentry *parent = new_dentry->d_parent;
9516                 btrfs_log_new_name(trans, old_inode, old_dir, parent);
9517                 btrfs_end_log_trans(root);
9518         }
9519 out_fail:
9520         btrfs_end_transaction(trans, root);
9521 out_notrans:
9522         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
9523                 up_read(&root->fs_info->subvol_sem);
9524
9525         return ret;
9526 }
9527
9528 static int btrfs_rename2(struct inode *old_dir, struct dentry *old_dentry,
9529                          struct inode *new_dir, struct dentry *new_dentry,
9530                          unsigned int flags)
9531 {
9532         if (flags & ~RENAME_NOREPLACE)
9533                 return -EINVAL;
9534
9535         return btrfs_rename(old_dir, old_dentry, new_dir, new_dentry);
9536 }
9537
9538 static void btrfs_run_delalloc_work(struct btrfs_work *work)
9539 {
9540         struct btrfs_delalloc_work *delalloc_work;
9541         struct inode *inode;
9542
9543         delalloc_work = container_of(work, struct btrfs_delalloc_work,
9544                                      work);
9545         inode = delalloc_work->inode;
9546         filemap_flush(inode->i_mapping);
9547         if (test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
9548                                 &BTRFS_I(inode)->runtime_flags))
9549                 filemap_flush(inode->i_mapping);
9550
9551         if (delalloc_work->delay_iput)
9552                 btrfs_add_delayed_iput(inode);
9553         else
9554                 iput(inode);
9555         complete(&delalloc_work->completion);
9556 }
9557
9558 struct btrfs_delalloc_work *btrfs_alloc_delalloc_work(struct inode *inode,
9559                                                     int delay_iput)
9560 {
9561         struct btrfs_delalloc_work *work;
9562
9563         work = kmalloc(sizeof(*work), GFP_NOFS);
9564         if (!work)
9565                 return NULL;
9566
9567         init_completion(&work->completion);
9568         INIT_LIST_HEAD(&work->list);
9569         work->inode = inode;
9570         work->delay_iput = delay_iput;
9571         WARN_ON_ONCE(!inode);
9572         btrfs_init_work(&work->work, btrfs_flush_delalloc_helper,
9573                         btrfs_run_delalloc_work, NULL, NULL);
9574
9575         return work;
9576 }
9577
9578 void btrfs_wait_and_free_delalloc_work(struct btrfs_delalloc_work *work)
9579 {
9580         wait_for_completion(&work->completion);
9581         kfree(work);
9582 }
9583
9584 /*
9585  * some fairly slow code that needs optimization. This walks the list
9586  * of all the inodes with pending delalloc and forces them to disk.
9587  */
9588 static int __start_delalloc_inodes(struct btrfs_root *root, int delay_iput,
9589                                    int nr)
9590 {
9591         struct btrfs_inode *binode;
9592         struct inode *inode;
9593         struct btrfs_delalloc_work *work, *next;
9594         struct list_head works;
9595         struct list_head splice;
9596         int ret = 0;
9597
9598         INIT_LIST_HEAD(&works);
9599         INIT_LIST_HEAD(&splice);
9600
9601         mutex_lock(&root->delalloc_mutex);
9602         spin_lock(&root->delalloc_lock);
9603         list_splice_init(&root->delalloc_inodes, &splice);
9604         while (!list_empty(&splice)) {
9605                 binode = list_entry(splice.next, struct btrfs_inode,
9606                                     delalloc_inodes);
9607
9608                 list_move_tail(&binode->delalloc_inodes,
9609                                &root->delalloc_inodes);
9610                 inode = igrab(&binode->vfs_inode);
9611                 if (!inode) {
9612                         cond_resched_lock(&root->delalloc_lock);
9613                         continue;
9614                 }
9615                 spin_unlock(&root->delalloc_lock);
9616
9617                 work = btrfs_alloc_delalloc_work(inode, delay_iput);
9618                 if (!work) {
9619                         if (delay_iput)
9620                                 btrfs_add_delayed_iput(inode);
9621                         else
9622                                 iput(inode);
9623                         ret = -ENOMEM;
9624                         goto out;
9625                 }
9626                 list_add_tail(&work->list, &works);
9627                 btrfs_queue_work(root->fs_info->flush_workers,
9628                                  &work->work);
9629                 ret++;
9630                 if (nr != -1 && ret >= nr)
9631                         goto out;
9632                 cond_resched();
9633                 spin_lock(&root->delalloc_lock);
9634         }
9635         spin_unlock(&root->delalloc_lock);
9636
9637 out:
9638         list_for_each_entry_safe(work, next, &works, list) {
9639                 list_del_init(&work->list);
9640                 btrfs_wait_and_free_delalloc_work(work);
9641         }
9642
9643         if (!list_empty_careful(&splice)) {
9644                 spin_lock(&root->delalloc_lock);
9645                 list_splice_tail(&splice, &root->delalloc_inodes);
9646                 spin_unlock(&root->delalloc_lock);
9647         }
9648         mutex_unlock(&root->delalloc_mutex);
9649         return ret;
9650 }
9651
9652 int btrfs_start_delalloc_inodes(struct btrfs_root *root, int delay_iput)
9653 {
9654         int ret;
9655
9656         if (test_bit(BTRFS_FS_STATE_ERROR, &root->fs_info->fs_state))
9657                 return -EROFS;
9658
9659         ret = __start_delalloc_inodes(root, delay_iput, -1);
9660         if (ret > 0)
9661                 ret = 0;
9662         /*
9663          * the filemap_flush will queue IO into the worker threads, but
9664          * we have to make sure the IO is actually started and that
9665          * ordered extents get created before we return
9666          */
9667         atomic_inc(&root->fs_info->async_submit_draining);
9668         while (atomic_read(&root->fs_info->nr_async_submits) ||
9669               atomic_read(&root->fs_info->async_delalloc_pages)) {
9670                 wait_event(root->fs_info->async_submit_wait,
9671                    (atomic_read(&root->fs_info->nr_async_submits) == 0 &&
9672                     atomic_read(&root->fs_info->async_delalloc_pages) == 0));
9673         }
9674         atomic_dec(&root->fs_info->async_submit_draining);
9675         return ret;
9676 }
9677
9678 int btrfs_start_delalloc_roots(struct btrfs_fs_info *fs_info, int delay_iput,
9679                                int nr)
9680 {
9681         struct btrfs_root *root;
9682         struct list_head splice;
9683         int ret;
9684
9685         if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state))
9686                 return -EROFS;
9687
9688         INIT_LIST_HEAD(&splice);
9689
9690         mutex_lock(&fs_info->delalloc_root_mutex);
9691         spin_lock(&fs_info->delalloc_root_lock);
9692         list_splice_init(&fs_info->delalloc_roots, &splice);
9693         while (!list_empty(&splice) && nr) {
9694                 root = list_first_entry(&splice, struct btrfs_root,
9695                                         delalloc_root);
9696                 root = btrfs_grab_fs_root(root);
9697                 BUG_ON(!root);
9698                 list_move_tail(&root->delalloc_root,
9699                                &fs_info->delalloc_roots);
9700                 spin_unlock(&fs_info->delalloc_root_lock);
9701
9702                 ret = __start_delalloc_inodes(root, delay_iput, nr);
9703                 btrfs_put_fs_root(root);
9704                 if (ret < 0)
9705                         goto out;
9706
9707                 if (nr != -1) {
9708                         nr -= ret;
9709                         WARN_ON(nr < 0);
9710                 }
9711                 spin_lock(&fs_info->delalloc_root_lock);
9712         }
9713         spin_unlock(&fs_info->delalloc_root_lock);
9714
9715         ret = 0;
9716         atomic_inc(&fs_info->async_submit_draining);
9717         while (atomic_read(&fs_info->nr_async_submits) ||
9718               atomic_read(&fs_info->async_delalloc_pages)) {
9719                 wait_event(fs_info->async_submit_wait,
9720                    (atomic_read(&fs_info->nr_async_submits) == 0 &&
9721                     atomic_read(&fs_info->async_delalloc_pages) == 0));
9722         }
9723         atomic_dec(&fs_info->async_submit_draining);
9724 out:
9725         if (!list_empty_careful(&splice)) {
9726                 spin_lock(&fs_info->delalloc_root_lock);
9727                 list_splice_tail(&splice, &fs_info->delalloc_roots);
9728                 spin_unlock(&fs_info->delalloc_root_lock);
9729         }
9730         mutex_unlock(&fs_info->delalloc_root_mutex);
9731         return ret;
9732 }
9733
9734 static int btrfs_symlink(struct inode *dir, struct dentry *dentry,
9735                          const char *symname)
9736 {
9737         struct btrfs_trans_handle *trans;
9738         struct btrfs_root *root = BTRFS_I(dir)->root;
9739         struct btrfs_path *path;
9740         struct btrfs_key key;
9741         struct inode *inode = NULL;
9742         int err;
9743         int drop_inode = 0;
9744         u64 objectid;
9745         u64 index = 0;
9746         int name_len;
9747         int datasize;
9748         unsigned long ptr;
9749         struct btrfs_file_extent_item *ei;
9750         struct extent_buffer *leaf;
9751
9752         name_len = strlen(symname);
9753         if (name_len > BTRFS_MAX_INLINE_DATA_SIZE(root))
9754                 return -ENAMETOOLONG;
9755
9756         /*
9757          * 2 items for inode item and ref
9758          * 2 items for dir items
9759          * 1 item for updating parent inode item
9760          * 1 item for the inline extent item
9761          * 1 item for xattr if selinux is on
9762          */
9763         trans = btrfs_start_transaction(root, 7);
9764         if (IS_ERR(trans))
9765                 return PTR_ERR(trans);
9766
9767         err = btrfs_find_free_ino(root, &objectid);
9768         if (err)
9769                 goto out_unlock;
9770
9771         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
9772                                 dentry->d_name.len, btrfs_ino(dir), objectid,
9773                                 S_IFLNK|S_IRWXUGO, &index);
9774         if (IS_ERR(inode)) {
9775                 err = PTR_ERR(inode);
9776                 goto out_unlock;
9777         }
9778
9779         /*
9780         * If the active LSM wants to access the inode during
9781         * d_instantiate it needs these. Smack checks to see
9782         * if the filesystem supports xattrs by looking at the
9783         * ops vector.
9784         */
9785         inode->i_fop = &btrfs_file_operations;
9786         inode->i_op = &btrfs_file_inode_operations;
9787         inode->i_mapping->a_ops = &btrfs_aops;
9788         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
9789
9790         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
9791         if (err)
9792                 goto out_unlock_inode;
9793
9794         path = btrfs_alloc_path();
9795         if (!path) {
9796                 err = -ENOMEM;
9797                 goto out_unlock_inode;
9798         }
9799         key.objectid = btrfs_ino(inode);
9800         key.offset = 0;
9801         key.type = BTRFS_EXTENT_DATA_KEY;
9802         datasize = btrfs_file_extent_calc_inline_size(name_len);
9803         err = btrfs_insert_empty_item(trans, root, path, &key,
9804                                       datasize);
9805         if (err) {
9806                 btrfs_free_path(path);
9807                 goto out_unlock_inode;
9808         }
9809         leaf = path->nodes[0];
9810         ei = btrfs_item_ptr(leaf, path->slots[0],
9811                             struct btrfs_file_extent_item);
9812         btrfs_set_file_extent_generation(leaf, ei, trans->transid);
9813         btrfs_set_file_extent_type(leaf, ei,
9814                                    BTRFS_FILE_EXTENT_INLINE);
9815         btrfs_set_file_extent_encryption(leaf, ei, 0);
9816         btrfs_set_file_extent_compression(leaf, ei, 0);
9817         btrfs_set_file_extent_other_encoding(leaf, ei, 0);
9818         btrfs_set_file_extent_ram_bytes(leaf, ei, name_len);
9819
9820         ptr = btrfs_file_extent_inline_start(ei);
9821         write_extent_buffer(leaf, symname, ptr, name_len);
9822         btrfs_mark_buffer_dirty(leaf);
9823         btrfs_free_path(path);
9824
9825         inode->i_op = &btrfs_symlink_inode_operations;
9826         inode->i_mapping->a_ops = &btrfs_symlink_aops;
9827         inode_set_bytes(inode, name_len);
9828         btrfs_i_size_write(inode, name_len);
9829         err = btrfs_update_inode(trans, root, inode);
9830         /*
9831          * Last step, add directory indexes for our symlink inode. This is the
9832          * last step to avoid extra cleanup of these indexes if an error happens
9833          * elsewhere above.
9834          */
9835         if (!err)
9836                 err = btrfs_add_nondir(trans, dir, dentry, inode, 0, index);
9837         if (err) {
9838                 drop_inode = 1;
9839                 goto out_unlock_inode;
9840         }
9841
9842         unlock_new_inode(inode);
9843         d_instantiate(dentry, inode);
9844
9845 out_unlock:
9846         btrfs_end_transaction(trans, root);
9847         if (drop_inode) {
9848                 inode_dec_link_count(inode);
9849                 iput(inode);
9850         }
9851         btrfs_btree_balance_dirty(root);
9852         return err;
9853
9854 out_unlock_inode:
9855         drop_inode = 1;
9856         unlock_new_inode(inode);
9857         goto out_unlock;
9858 }
9859
9860 static int __btrfs_prealloc_file_range(struct inode *inode, int mode,
9861                                        u64 start, u64 num_bytes, u64 min_size,
9862                                        loff_t actual_len, u64 *alloc_hint,
9863                                        struct btrfs_trans_handle *trans)
9864 {
9865         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
9866         struct extent_map *em;
9867         struct btrfs_root *root = BTRFS_I(inode)->root;
9868         struct btrfs_key ins;
9869         u64 cur_offset = start;
9870         u64 i_size;
9871         u64 cur_bytes;
9872         u64 last_alloc = (u64)-1;
9873         int ret = 0;
9874         bool own_trans = true;
9875
9876         if (trans)
9877                 own_trans = false;
9878         while (num_bytes > 0) {
9879                 if (own_trans) {
9880                         trans = btrfs_start_transaction(root, 3);
9881                         if (IS_ERR(trans)) {
9882                                 ret = PTR_ERR(trans);
9883                                 break;
9884                         }
9885                 }
9886
9887                 cur_bytes = min_t(u64, num_bytes, SZ_256M);
9888                 cur_bytes = max(cur_bytes, min_size);
9889                 /*
9890                  * If we are severely fragmented we could end up with really
9891                  * small allocations, so if the allocator is returning small
9892                  * chunks lets make its job easier by only searching for those
9893                  * sized chunks.
9894                  */
9895                 cur_bytes = min(cur_bytes, last_alloc);
9896                 ret = btrfs_reserve_extent(root, cur_bytes, min_size, 0,
9897                                            *alloc_hint, &ins, 1, 0);
9898                 if (ret) {
9899                         if (own_trans)
9900                                 btrfs_end_transaction(trans, root);
9901                         break;
9902                 }
9903
9904                 last_alloc = ins.offset;
9905                 ret = insert_reserved_file_extent(trans, inode,
9906                                                   cur_offset, ins.objectid,
9907                                                   ins.offset, ins.offset,
9908                                                   ins.offset, 0, 0, 0,
9909                                                   BTRFS_FILE_EXTENT_PREALLOC);
9910                 if (ret) {
9911                         btrfs_free_reserved_extent(root, ins.objectid,
9912                                                    ins.offset, 0);
9913                         btrfs_abort_transaction(trans, root, ret);
9914                         if (own_trans)
9915                                 btrfs_end_transaction(trans, root);
9916                         break;
9917                 }
9918
9919                 btrfs_drop_extent_cache(inode, cur_offset,
9920                                         cur_offset + ins.offset -1, 0);
9921
9922                 em = alloc_extent_map();
9923                 if (!em) {
9924                         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
9925                                 &BTRFS_I(inode)->runtime_flags);
9926                         goto next;
9927                 }
9928
9929                 em->start = cur_offset;
9930                 em->orig_start = cur_offset;
9931                 em->len = ins.offset;
9932                 em->block_start = ins.objectid;
9933                 em->block_len = ins.offset;
9934                 em->orig_block_len = ins.offset;
9935                 em->ram_bytes = ins.offset;
9936                 em->bdev = root->fs_info->fs_devices->latest_bdev;
9937                 set_bit(EXTENT_FLAG_PREALLOC, &em->flags);
9938                 em->generation = trans->transid;
9939
9940                 while (1) {
9941                         write_lock(&em_tree->lock);
9942                         ret = add_extent_mapping(em_tree, em, 1);
9943                         write_unlock(&em_tree->lock);
9944                         if (ret != -EEXIST)
9945                                 break;
9946                         btrfs_drop_extent_cache(inode, cur_offset,
9947                                                 cur_offset + ins.offset - 1,
9948                                                 0);
9949                 }
9950                 free_extent_map(em);
9951 next:
9952                 num_bytes -= ins.offset;
9953                 cur_offset += ins.offset;
9954                 *alloc_hint = ins.objectid + ins.offset;
9955
9956                 inode_inc_iversion(inode);
9957                 inode->i_ctime = CURRENT_TIME;
9958                 BTRFS_I(inode)->flags |= BTRFS_INODE_PREALLOC;
9959                 if (!(mode & FALLOC_FL_KEEP_SIZE) &&
9960                     (actual_len > inode->i_size) &&
9961                     (cur_offset > inode->i_size)) {
9962                         if (cur_offset > actual_len)
9963                                 i_size = actual_len;
9964                         else
9965                                 i_size = cur_offset;
9966                         i_size_write(inode, i_size);
9967                         btrfs_ordered_update_i_size(inode, i_size, NULL);
9968                 }
9969
9970                 ret = btrfs_update_inode(trans, root, inode);
9971
9972                 if (ret) {
9973                         btrfs_abort_transaction(trans, root, ret);
9974                         if (own_trans)
9975                                 btrfs_end_transaction(trans, root);
9976                         break;
9977                 }
9978
9979                 if (own_trans)
9980                         btrfs_end_transaction(trans, root);
9981         }
9982         return ret;
9983 }
9984
9985 int btrfs_prealloc_file_range(struct inode *inode, int mode,
9986                               u64 start, u64 num_bytes, u64 min_size,
9987                               loff_t actual_len, u64 *alloc_hint)
9988 {
9989         return __btrfs_prealloc_file_range(inode, mode, start, num_bytes,
9990                                            min_size, actual_len, alloc_hint,
9991                                            NULL);
9992 }
9993
9994 int btrfs_prealloc_file_range_trans(struct inode *inode,
9995                                     struct btrfs_trans_handle *trans, int mode,
9996                                     u64 start, u64 num_bytes, u64 min_size,
9997                                     loff_t actual_len, u64 *alloc_hint)
9998 {
9999         return __btrfs_prealloc_file_range(inode, mode, start, num_bytes,
10000                                            min_size, actual_len, alloc_hint, trans);
10001 }
10002
10003 static int btrfs_set_page_dirty(struct page *page)
10004 {
10005         return __set_page_dirty_nobuffers(page);
10006 }
10007
10008 static int btrfs_permission(struct inode *inode, int mask)
10009 {
10010         struct btrfs_root *root = BTRFS_I(inode)->root;
10011         umode_t mode = inode->i_mode;
10012
10013         if (mask & MAY_WRITE &&
10014             (S_ISREG(mode) || S_ISDIR(mode) || S_ISLNK(mode))) {
10015                 if (btrfs_root_readonly(root))
10016                         return -EROFS;
10017                 if (BTRFS_I(inode)->flags & BTRFS_INODE_READONLY)
10018                         return -EACCES;
10019         }
10020         return generic_permission(inode, mask);
10021 }
10022
10023 static int btrfs_tmpfile(struct inode *dir, struct dentry *dentry, umode_t mode)
10024 {
10025         struct btrfs_trans_handle *trans;
10026         struct btrfs_root *root = BTRFS_I(dir)->root;
10027         struct inode *inode = NULL;
10028         u64 objectid;
10029         u64 index;
10030         int ret = 0;
10031
10032         /*
10033          * 5 units required for adding orphan entry
10034          */
10035         trans = btrfs_start_transaction(root, 5);
10036         if (IS_ERR(trans))
10037                 return PTR_ERR(trans);
10038
10039         ret = btrfs_find_free_ino(root, &objectid);
10040         if (ret)
10041                 goto out;
10042
10043         inode = btrfs_new_inode(trans, root, dir, NULL, 0,
10044                                 btrfs_ino(dir), objectid, mode, &index);
10045         if (IS_ERR(inode)) {
10046                 ret = PTR_ERR(inode);
10047                 inode = NULL;
10048                 goto out;
10049         }
10050
10051         inode->i_fop = &btrfs_file_operations;
10052         inode->i_op = &btrfs_file_inode_operations;
10053
10054         inode->i_mapping->a_ops = &btrfs_aops;
10055         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
10056
10057         ret = btrfs_init_inode_security(trans, inode, dir, NULL);
10058         if (ret)
10059                 goto out_inode;
10060
10061         ret = btrfs_update_inode(trans, root, inode);
10062         if (ret)
10063                 goto out_inode;
10064         ret = btrfs_orphan_add(trans, inode);
10065         if (ret)
10066                 goto out_inode;
10067
10068         /*
10069          * We set number of links to 0 in btrfs_new_inode(), and here we set
10070          * it to 1 because d_tmpfile() will issue a warning if the count is 0,
10071          * through:
10072          *
10073          *    d_tmpfile() -> inode_dec_link_count() -> drop_nlink()
10074          */
10075         set_nlink(inode, 1);
10076         unlock_new_inode(inode);
10077         d_tmpfile(dentry, inode);
10078         mark_inode_dirty(inode);
10079
10080 out:
10081         btrfs_end_transaction(trans, root);
10082         if (ret)
10083                 iput(inode);
10084         btrfs_balance_delayed_items(root);
10085         btrfs_btree_balance_dirty(root);
10086         return ret;
10087
10088 out_inode:
10089         unlock_new_inode(inode);
10090         goto out;
10091
10092 }
10093
10094 /* Inspired by filemap_check_errors() */
10095 int btrfs_inode_check_errors(struct inode *inode)
10096 {
10097         int ret = 0;
10098
10099         if (test_bit(AS_ENOSPC, &inode->i_mapping->flags) &&
10100             test_and_clear_bit(AS_ENOSPC, &inode->i_mapping->flags))
10101                 ret = -ENOSPC;
10102         if (test_bit(AS_EIO, &inode->i_mapping->flags) &&
10103             test_and_clear_bit(AS_EIO, &inode->i_mapping->flags))
10104                 ret = -EIO;
10105
10106         return ret;
10107 }
10108
10109 static const struct inode_operations btrfs_dir_inode_operations = {
10110         .getattr        = btrfs_getattr,
10111         .lookup         = btrfs_lookup,
10112         .create         = btrfs_create,
10113         .unlink         = btrfs_unlink,
10114         .link           = btrfs_link,
10115         .mkdir          = btrfs_mkdir,
10116         .rmdir          = btrfs_rmdir,
10117         .rename2        = btrfs_rename2,
10118         .symlink        = btrfs_symlink,
10119         .setattr        = btrfs_setattr,
10120         .mknod          = btrfs_mknod,
10121         .setxattr       = btrfs_setxattr,
10122         .getxattr       = btrfs_getxattr,
10123         .listxattr      = btrfs_listxattr,
10124         .removexattr    = btrfs_removexattr,
10125         .permission     = btrfs_permission,
10126         .get_acl        = btrfs_get_acl,
10127         .set_acl        = btrfs_set_acl,
10128         .update_time    = btrfs_update_time,
10129         .tmpfile        = btrfs_tmpfile,
10130 };
10131 static const struct inode_operations btrfs_dir_ro_inode_operations = {
10132         .lookup         = btrfs_lookup,
10133         .permission     = btrfs_permission,
10134         .get_acl        = btrfs_get_acl,
10135         .set_acl        = btrfs_set_acl,
10136         .update_time    = btrfs_update_time,
10137 };
10138
10139 static const struct file_operations btrfs_dir_file_operations = {
10140         .llseek         = generic_file_llseek,
10141         .read           = generic_read_dir,
10142         .iterate        = btrfs_real_readdir,
10143         .unlocked_ioctl = btrfs_ioctl,
10144 #ifdef CONFIG_COMPAT
10145         .compat_ioctl   = btrfs_ioctl,
10146 #endif
10147         .release        = btrfs_release_file,
10148         .fsync          = btrfs_sync_file,
10149 };
10150
10151 static const struct extent_io_ops btrfs_extent_io_ops = {
10152         .fill_delalloc = run_delalloc_range,
10153         .submit_bio_hook = btrfs_submit_bio_hook,
10154         .merge_bio_hook = btrfs_merge_bio_hook,
10155         .readpage_end_io_hook = btrfs_readpage_end_io_hook,
10156         .writepage_end_io_hook = btrfs_writepage_end_io_hook,
10157         .writepage_start_hook = btrfs_writepage_start_hook,
10158         .set_bit_hook = btrfs_set_bit_hook,
10159         .clear_bit_hook = btrfs_clear_bit_hook,
10160         .merge_extent_hook = btrfs_merge_extent_hook,
10161         .split_extent_hook = btrfs_split_extent_hook,
10162 };
10163
10164 /*
10165  * btrfs doesn't support the bmap operation because swapfiles
10166  * use bmap to make a mapping of extents in the file.  They assume
10167  * these extents won't change over the life of the file and they
10168  * use the bmap result to do IO directly to the drive.
10169  *
10170  * the btrfs bmap call would return logical addresses that aren't
10171  * suitable for IO and they also will change frequently as COW
10172  * operations happen.  So, swapfile + btrfs == corruption.
10173  *
10174  * For now we're avoiding this by dropping bmap.
10175  */
10176 static const struct address_space_operations btrfs_aops = {
10177         .readpage       = btrfs_readpage,
10178         .writepage      = btrfs_writepage,
10179         .writepages     = btrfs_writepages,
10180         .readpages      = btrfs_readpages,
10181         .direct_IO      = btrfs_direct_IO,
10182         .invalidatepage = btrfs_invalidatepage,
10183         .releasepage    = btrfs_releasepage,
10184         .set_page_dirty = btrfs_set_page_dirty,
10185         .error_remove_page = generic_error_remove_page,
10186 };
10187
10188 static const struct address_space_operations btrfs_symlink_aops = {
10189         .readpage       = btrfs_readpage,
10190         .writepage      = btrfs_writepage,
10191         .invalidatepage = btrfs_invalidatepage,
10192         .releasepage    = btrfs_releasepage,
10193 };
10194
10195 static const struct inode_operations btrfs_file_inode_operations = {
10196         .getattr        = btrfs_getattr,
10197         .setattr        = btrfs_setattr,
10198         .setxattr       = btrfs_setxattr,
10199         .getxattr       = btrfs_getxattr,
10200         .listxattr      = btrfs_listxattr,
10201         .removexattr    = btrfs_removexattr,
10202         .permission     = btrfs_permission,
10203         .fiemap         = btrfs_fiemap,
10204         .get_acl        = btrfs_get_acl,
10205         .set_acl        = btrfs_set_acl,
10206         .update_time    = btrfs_update_time,
10207 };
10208 static const struct inode_operations btrfs_special_inode_operations = {
10209         .getattr        = btrfs_getattr,
10210         .setattr        = btrfs_setattr,
10211         .permission     = btrfs_permission,
10212         .setxattr       = btrfs_setxattr,
10213         .getxattr       = btrfs_getxattr,
10214         .listxattr      = btrfs_listxattr,
10215         .removexattr    = btrfs_removexattr,
10216         .get_acl        = btrfs_get_acl,
10217         .set_acl        = btrfs_set_acl,
10218         .update_time    = btrfs_update_time,
10219 };
10220 static const struct inode_operations btrfs_symlink_inode_operations = {
10221         .readlink       = generic_readlink,
10222         .follow_link    = page_follow_link_light,
10223         .put_link       = page_put_link,
10224         .getattr        = btrfs_getattr,
10225         .setattr        = btrfs_setattr,
10226         .permission     = btrfs_permission,
10227         .setxattr       = btrfs_setxattr,
10228         .getxattr       = btrfs_getxattr,
10229         .listxattr      = btrfs_listxattr,
10230         .removexattr    = btrfs_removexattr,
10231         .update_time    = btrfs_update_time,
10232 };
10233
10234 const struct dentry_operations btrfs_dentry_operations = {
10235         .d_delete       = btrfs_dentry_delete,
10236         .d_release      = btrfs_dentry_release,
10237 };