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/compat.h>
  34 #include <linux/bit_spinlock.h>
  35 #include <linux/xattr.h>
  36 #include <linux/posix_acl.h>
  37 #include <linux/falloc.h>
  38 #include <linux/slab.h>
  39 #include <linux/ratelimit.h>
  40 #include <linux/mount.h>
  41 #include <linux/btrfs.h>
  42 #include <linux/blkdev.h>
  43 #include <linux/posix_acl_xattr.h>
  44 #include <linux/uio.h>
  45 #include "ctree.h"
  46 #include "disk-io.h"
  47 #include "transaction.h"
  48 #include "btrfs_inode.h"
  49 #include "print-tree.h"
  50 #include "ordered-data.h"
  51 #include "xattr.h"
  52 #include "tree-log.h"
  53 #include "volumes.h"
  54 #include "compression.h"
  55 #include "locking.h"
  56 #include "free-space-cache.h"
  57 #include "inode-map.h"
  58 #include "backref.h"
  59 #include "hash.h"
  60 #include "props.h"
  61 #include "qgroup.h"
  62 #include "dedupe.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         int overwrite;
  75 };
  76
  77 static const struct inode_operations btrfs_dir_inode_operations;
  78 static const struct inode_operations btrfs_symlink_inode_operations;
  79 static const struct inode_operations btrfs_dir_ro_inode_operations;
  80 static const struct inode_operations btrfs_special_inode_operations;
  81 static const struct inode_operations btrfs_file_inode_operations;
  82 static const struct address_space_operations btrfs_aops;
  83 static const struct address_space_operations btrfs_symlink_aops;
  84 static const struct file_operations btrfs_dir_file_operations;
  85 static const struct extent_io_ops btrfs_extent_io_ops;
  86
  87 static struct kmem_cache *btrfs_inode_cachep;
  88 struct kmem_cache *btrfs_trans_handle_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, u64 delalloc_end,
 109                                    int *page_started, unsigned long *nr_written,
 110                                    int unlock, struct btrfs_dedupe_hash *hash);
 111 static struct extent_map *create_io_em(struct inode *inode, u64 start, u64 len,
 112                                        u64 orig_start, u64 block_start,
 113                                        u64 block_len, u64 orig_block_len,
 114                                        u64 ram_bytes, int compress_type,
 115                                        int type);
 116
 117 static void __endio_write_update_ordered(struct inode *inode,
 118                                          const u64 offset, const u64 bytes,
 119                                          const bool uptodate);
 120
 121 /*
 122  * Cleanup all submitted ordered extents in specified range to handle errors
 123  * from the fill_dellaloc() callback.
 124  *
 125  * NOTE: caller must ensure that when an error happens, it can not call
 126  * extent_clear_unlock_delalloc() to clear both the bits EXTENT_DO_ACCOUNTING
 127  * and EXTENT_DELALLOC simultaneously, because that causes the reserved metadata
 128  * to be released, which we want to happen only when finishing the ordered
 129  * extent (btrfs_finish_ordered_io()). Also note that the caller of the
 130  * fill_delalloc() callback already does proper cleanup for the first page of
 131  * the range, that is, it invokes the callback writepage_end_io_hook() for the
 132  * range of the first page.
 133  */
 134 static inline void btrfs_cleanup_ordered_extents(struct inode *inode,
 135                                                  const u64 offset,
 136                                                  const u64 bytes)
 137 {
 138         return __endio_write_update_ordered(inode, offset + PAGE_SIZE,
 139                                             bytes - PAGE_SIZE, false);
 140 }
 141
 142 static int btrfs_dirty_inode(struct inode *inode);
 143
 144 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
 145 void btrfs_test_inode_set_ops(struct inode *inode)
 146 {
 147         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
 148 }
 149 #endif
 150
 151 static int btrfs_init_inode_security(struct btrfs_trans_handle *trans,
 152                                      struct inode *inode,  struct inode *dir,
 153                                      const struct qstr *qstr)
 154 {
 155         int err;
 156
 157         err = btrfs_init_acl(trans, inode, dir);
 158         if (!err)
 159                 err = btrfs_xattr_security_init(trans, inode, dir, qstr);
 160         return err;
 161 }
 162
 163 /*
 164  * this does all the hard work for inserting an inline extent into
 165  * the btree.  The caller should have done a btrfs_drop_extents so that
 166  * no overlapping inline items exist in the btree
 167  */
 168 static int insert_inline_extent(struct btrfs_trans_handle *trans,
 169                                 struct btrfs_path *path, int extent_inserted,
 170                                 struct btrfs_root *root, struct inode *inode,
 171                                 u64 start, size_t size, size_t compressed_size,
 172                                 int compress_type,
 173                                 struct page **compressed_pages)
 174 {
 175         struct extent_buffer *leaf;
 176         struct page *page = NULL;
 177         char *kaddr;
 178         unsigned long ptr;
 179         struct btrfs_file_extent_item *ei;
 180         int ret;
 181         size_t cur_size = size;
 182         unsigned long offset;
 183
 184         if (compressed_size && compressed_pages)
 185                 cur_size = compressed_size;
 186
 187         inode_add_bytes(inode, size);
 188
 189         if (!extent_inserted) {
 190                 struct btrfs_key key;
 191                 size_t datasize;
 192
 193                 key.objectid = btrfs_ino(BTRFS_I(inode));
 194                 key.offset = start;
 195                 key.type = BTRFS_EXTENT_DATA_KEY;
 196
 197                 datasize = btrfs_file_extent_calc_inline_size(cur_size);
 198                 path->leave_spinning = 1;
 199                 ret = btrfs_insert_empty_item(trans, root, path, &key,
 200                                               datasize);
 201                 if (ret)
 202                         goto fail;
 203         }
 204         leaf = path->nodes[0];
 205         ei = btrfs_item_ptr(leaf, path->slots[0],
 206                             struct btrfs_file_extent_item);
 207         btrfs_set_file_extent_generation(leaf, ei, trans->transid);
 208         btrfs_set_file_extent_type(leaf, ei, BTRFS_FILE_EXTENT_INLINE);
 209         btrfs_set_file_extent_encryption(leaf, ei, 0);
 210         btrfs_set_file_extent_other_encoding(leaf, ei, 0);
 211         btrfs_set_file_extent_ram_bytes(leaf, ei, size);
 212         ptr = btrfs_file_extent_inline_start(ei);
 213
 214         if (compress_type != BTRFS_COMPRESS_NONE) {
 215                 struct page *cpage;
 216                 int i = 0;
 217                 while (compressed_size > 0) {
 218                         cpage = compressed_pages[i];
 219                         cur_size = min_t(unsigned long, compressed_size,
 220                                        PAGE_SIZE);
 221
 222                         kaddr = kmap_atomic(cpage);
 223                         write_extent_buffer(leaf, kaddr, ptr, cur_size);
 224                         kunmap_atomic(kaddr);
 225
 226                         i++;
 227                         ptr += cur_size;
 228                         compressed_size -= cur_size;
 229                 }
 230                 btrfs_set_file_extent_compression(leaf, ei,
 231                                                   compress_type);
 232         } else {
 233                 page = find_get_page(inode->i_mapping,
 234                                      start >> PAGE_SHIFT);
 235                 btrfs_set_file_extent_compression(leaf, ei, 0);
 236                 kaddr = kmap_atomic(page);
 237                 offset = start & (PAGE_SIZE - 1);
 238                 write_extent_buffer(leaf, kaddr + offset, ptr, size);
 239                 kunmap_atomic(kaddr);
 240                 put_page(page);
 241         }
 242         btrfs_mark_buffer_dirty(leaf);
 243         btrfs_release_path(path);
 244
 245         /*
 246          * we're an inline extent, so nobody can
 247          * extend the file past i_size without locking
 248          * a page we already have locked.
 249          *
 250          * We must do any isize and inode updates
 251          * before we unlock the pages.  Otherwise we
 252          * could end up racing with unlink.
 253          */
 254         BTRFS_I(inode)->disk_i_size = inode->i_size;
 255         ret = btrfs_update_inode(trans, root, inode);
 256
 257 fail:
 258         return ret;
 259 }
 260
 261
 262 /*
 263  * conditionally insert an inline extent into the file.  This
 264  * does the checks required to make sure the data is small enough
 265  * to fit as an inline extent.
 266  */
 267 static noinline int cow_file_range_inline(struct btrfs_root *root,
 268                                           struct inode *inode, u64 start,
 269                                           u64 end, size_t compressed_size,
 270                                           int compress_type,
 271                                           struct page **compressed_pages)
 272 {
 273         struct btrfs_fs_info *fs_info = root->fs_info;
 274         struct btrfs_trans_handle *trans;
 275         u64 isize = i_size_read(inode);
 276         u64 actual_end = min(end + 1, isize);
 277         u64 inline_len = actual_end - start;
 278         u64 aligned_end = ALIGN(end, fs_info->sectorsize);
 279         u64 data_len = inline_len;
 280         int ret;
 281         struct btrfs_path *path;
 282         int extent_inserted = 0;
 283         u32 extent_item_size;
 284
 285         if (compressed_size)
 286                 data_len = compressed_size;
 287
 288         if (start > 0 ||
 289             actual_end > fs_info->sectorsize ||
 290             data_len > BTRFS_MAX_INLINE_DATA_SIZE(fs_info) ||
 291             (!compressed_size &&
 292             (actual_end & (fs_info->sectorsize - 1)) == 0) ||
 293             end + 1 < isize ||
 294             data_len > fs_info->max_inline) {
 295                 return 1;
 296         }
 297
 298         path = btrfs_alloc_path();
 299         if (!path)
 300                 return -ENOMEM;
 301
 302         trans = btrfs_join_transaction(root);
 303         if (IS_ERR(trans)) {
 304                 btrfs_free_path(path);
 305                 return PTR_ERR(trans);
 306         }
 307         trans->block_rsv = &fs_info->delalloc_block_rsv;
 308
 309         if (compressed_size && compressed_pages)
 310                 extent_item_size = btrfs_file_extent_calc_inline_size(
 311                    compressed_size);
 312         else
 313                 extent_item_size = btrfs_file_extent_calc_inline_size(
 314                     inline_len);
 315
 316         ret = __btrfs_drop_extents(trans, root, inode, path,
 317                                    start, aligned_end, NULL,
 318                                    1, 1, extent_item_size, &extent_inserted);
 319         if (ret) {
 320                 btrfs_abort_transaction(trans, ret);
 321                 goto out;
 322         }
 323
 324         if (isize > actual_end)
 325                 inline_len = min_t(u64, isize, actual_end);
 326         ret = insert_inline_extent(trans, path, extent_inserted,
 327                                    root, inode, start,
 328                                    inline_len, compressed_size,
 329                                    compress_type, compressed_pages);
 330         if (ret && ret != -ENOSPC) {
 331                 btrfs_abort_transaction(trans, ret);
 332                 goto out;
 333         } else if (ret == -ENOSPC) {
 334                 ret = 1;
 335                 goto out;
 336         }
 337
 338         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &BTRFS_I(inode)->runtime_flags);
 339         btrfs_delalloc_release_metadata(BTRFS_I(inode), end + 1 - start);
 340         btrfs_drop_extent_cache(BTRFS_I(inode), start, aligned_end - 1, 0);
 341 out:
 342         /*
 343          * Don't forget to free the reserved space, as for inlined extent
 344          * it won't count as data extent, free them directly here.
 345          * And at reserve time, it's always aligned to page size, so
 346          * just free one page here.
 347          */
 348         btrfs_qgroup_free_data(inode, NULL, 0, PAGE_SIZE);
 349         btrfs_free_path(path);
 350         btrfs_end_transaction(trans);
 351         return ret;
 352 }
 353
 354 struct async_extent {
 355         u64 start;
 356         u64 ram_size;
 357         u64 compressed_size;
 358         struct page **pages;
 359         unsigned long nr_pages;
 360         int compress_type;
 361         struct list_head list;
 362 };
 363
 364 struct async_cow {
 365         struct inode *inode;
 366         struct btrfs_root *root;
 367         struct page *locked_page;
 368         u64 start;
 369         u64 end;
 370         struct list_head extents;
 371         struct btrfs_work work;
 372 };
 373
 374 static noinline int add_async_extent(struct async_cow *cow,
 375                                      u64 start, u64 ram_size,
 376                                      u64 compressed_size,
 377                                      struct page **pages,
 378                                      unsigned long nr_pages,
 379                                      int compress_type)
 380 {
 381         struct async_extent *async_extent;
 382
 383         async_extent = kmalloc(sizeof(*async_extent), GFP_NOFS);
 384         BUG_ON(!async_extent); /* -ENOMEM */
 385         async_extent->start = start;
 386         async_extent->ram_size = ram_size;
 387         async_extent->compressed_size = compressed_size;
 388         async_extent->pages = pages;
 389         async_extent->nr_pages = nr_pages;
 390         async_extent->compress_type = compress_type;
 391         list_add_tail(&async_extent->list, &cow->extents);
 392         return 0;
 393 }
 394
 395 static inline int inode_need_compress(struct inode *inode, u64 start, u64 end)
 396 {
 397         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
 398
 399         /* force compress */
 400         if (btrfs_test_opt(fs_info, FORCE_COMPRESS))
 401                 return 1;
 402         /* defrag ioctl */
 403         if (BTRFS_I(inode)->defrag_compress)
 404                 return 1;
 405         /* bad compression ratios */
 406         if (BTRFS_I(inode)->flags & BTRFS_INODE_NOCOMPRESS)
 407                 return 0;
 408         if (btrfs_test_opt(fs_info, COMPRESS) ||
 409             BTRFS_I(inode)->flags & BTRFS_INODE_COMPRESS ||
 410             BTRFS_I(inode)->prop_compress)
 411                 return btrfs_compress_heuristic(inode, start, end);
 412         return 0;
 413 }
 414
 415 static inline void inode_should_defrag(struct btrfs_inode *inode,
 416                 u64 start, u64 end, u64 num_bytes, u64 small_write)
 417 {
 418         /* If this is a small write inside eof, kick off a defrag */
 419         if (num_bytes < small_write &&
 420             (start > 0 || end + 1 < inode->disk_i_size))
 421                 btrfs_add_inode_defrag(NULL, inode);
 422 }
 423
 424 /*
 425  * we create compressed extents in two phases.  The first
 426  * phase compresses a range of pages that have already been
 427  * locked (both pages and state bits are locked).
 428  *
 429  * This is done inside an ordered work queue, and the compression
 430  * is spread across many cpus.  The actual IO submission is step
 431  * two, and the ordered work queue takes care of making sure that
 432  * happens in the same order things were put onto the queue by
 433  * writepages and friends.
 434  *
 435  * If this code finds it can't get good compression, it puts an
 436  * entry onto the work queue to write the uncompressed bytes.  This
 437  * makes sure that both compressed inodes and uncompressed inodes
 438  * are written in the same order that the flusher thread sent them
 439  * down.
 440  */
 441 static noinline void compress_file_range(struct inode *inode,
 442                                         struct page *locked_page,
 443                                         u64 start, u64 end,
 444                                         struct async_cow *async_cow,
 445                                         int *num_added)
 446 {
 447         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
 448         struct btrfs_root *root = BTRFS_I(inode)->root;
 449         u64 num_bytes;
 450         u64 blocksize = fs_info->sectorsize;
 451         u64 actual_end;
 452         u64 isize = i_size_read(inode);
 453         int ret = 0;
 454         struct page **pages = NULL;
 455         unsigned long nr_pages;
 456         unsigned long total_compressed = 0;
 457         unsigned long total_in = 0;
 458         int i;
 459         int will_compress;
 460         int compress_type = fs_info->compress_type;
 461         int redirty = 0;
 462
 463         inode_should_defrag(BTRFS_I(inode), start, end, end - start + 1,
 464                         SZ_16K);
 465
 466         actual_end = min_t(u64, isize, end + 1);
 467 again:
 468         will_compress = 0;
 469         nr_pages = (end >> PAGE_SHIFT) - (start >> PAGE_SHIFT) + 1;
 470         BUILD_BUG_ON((BTRFS_MAX_COMPRESSED % PAGE_SIZE) != 0);
 471         nr_pages = min_t(unsigned long, nr_pages,
 472                         BTRFS_MAX_COMPRESSED / PAGE_SIZE);
 473
 474         /*
 475          * we don't want to send crud past the end of i_size through
 476          * compression, that's just a waste of CPU time.  So, if the
 477          * end of the file is before the start of our current
 478          * requested range of bytes, we bail out to the uncompressed
 479          * cleanup code that can deal with all of this.
 480          *
 481          * It isn't really the fastest way to fix things, but this is a
 482          * very uncommon corner.
 483          */
 484         if (actual_end <= start)
 485                 goto cleanup_and_bail_uncompressed;
 486
 487         total_compressed = actual_end - start;
 488
 489         /*
 490          * skip compression for a small file range(<=blocksize) that
 491          * isn't an inline extent, since it doesn't save disk space at all.
 492          */
 493         if (total_compressed <= blocksize &&
 494            (start > 0 || end + 1 < BTRFS_I(inode)->disk_i_size))
 495                 goto cleanup_and_bail_uncompressed;
 496
 497         total_compressed = min_t(unsigned long, total_compressed,
 498                         BTRFS_MAX_UNCOMPRESSED);
 499         num_bytes = ALIGN(end - start + 1, blocksize);
 500         num_bytes = max(blocksize,  num_bytes);
 501         total_in = 0;
 502         ret = 0;
 503
 504         /*
 505          * we do compression for mount -o compress and when the
 506          * inode has not been flagged as nocompress.  This flag can
 507          * change at any time if we discover bad compression ratios.
 508          */
 509         if (inode_need_compress(inode, start, end)) {
 510                 WARN_ON(pages);
 511                 pages = kcalloc(nr_pages, sizeof(struct page *), GFP_NOFS);
 512                 if (!pages) {
 513                         /* just bail out to the uncompressed code */
 514                         goto cont;
 515                 }
 516
 517                 if (BTRFS_I(inode)->defrag_compress)
 518                         compress_type = BTRFS_I(inode)->defrag_compress;
 519                 else if (BTRFS_I(inode)->prop_compress)
 520                         compress_type = BTRFS_I(inode)->prop_compress;
 521
 522                 /*
 523                  * we need to call clear_page_dirty_for_io on each
 524                  * page in the range.  Otherwise applications with the file
 525                  * mmap'd can wander in and change the page contents while
 526                  * we are compressing them.
 527                  *
 528                  * If the compression fails for any reason, we set the pages
 529                  * dirty again later on.
 530                  */
 531                 extent_range_clear_dirty_for_io(inode, start, end);
 532                 redirty = 1;
 533                 ret = btrfs_compress_pages(compress_type,
 534                                            inode->i_mapping, start,
 535                                            pages,
 536                                            &nr_pages,
 537                                            &total_in,
 538                                            &total_compressed);
 539
 540                 if (!ret) {
 541                         unsigned long offset = total_compressed &
 542                                 (PAGE_SIZE - 1);
 543                         struct page *page = pages[nr_pages - 1];
 544                         char *kaddr;
 545
 546                         /* zero the tail end of the last page, we might be
 547                          * sending it down to disk
 548                          */
 549                         if (offset) {
 550                                 kaddr = kmap_atomic(page);
 551                                 memset(kaddr + offset, 0,
 552                                        PAGE_SIZE - offset);
 553                                 kunmap_atomic(kaddr);
 554                         }
 555                         will_compress = 1;
 556                 }
 557         }
 558 cont:
 559         if (start == 0) {
 560                 /* lets try to make an inline extent */
 561                 if (ret || total_in < (actual_end - start)) {
 562                         /* we didn't compress the entire range, try
 563                          * to make an uncompressed inline extent.
 564                          */
 565                         ret = cow_file_range_inline(root, inode, start, end,
 566                                             0, BTRFS_COMPRESS_NONE, NULL);
 567                 } else {
 568                         /* try making a compressed inline extent */
 569                         ret = cow_file_range_inline(root, inode, start, end,
 570                                                     total_compressed,
 571                                                     compress_type, pages);
 572                 }
 573                 if (ret <= 0) {
 574                         unsigned long clear_flags = EXTENT_DELALLOC |
 575                                 EXTENT_DELALLOC_NEW | EXTENT_DEFRAG;
 576                         unsigned long page_error_op;
 577
 578                         clear_flags |= (ret < 0) ? EXTENT_DO_ACCOUNTING : 0;
 579                         page_error_op = ret < 0 ? PAGE_SET_ERROR : 0;
 580
 581                         /*
 582                          * inline extent creation worked or returned error,
 583                          * we don't need to create any more async work items.
 584                          * Unlock and free up our temp pages.
 585                          */
 586                         extent_clear_unlock_delalloc(inode, start, end, end,
 587                                                      NULL, clear_flags,
 588                                                      PAGE_UNLOCK |
 589                                                      PAGE_CLEAR_DIRTY |
 590                                                      PAGE_SET_WRITEBACK |
 591                                                      page_error_op |
 592                                                      PAGE_END_WRITEBACK);
 593                         if (ret == 0)
 594                                 btrfs_free_reserved_data_space_noquota(inode,
 595                                                                start,
 596                                                                end - start + 1);
 597                         goto free_pages_out;
 598                 }
 599         }
 600
 601         if (will_compress) {
 602                 /*
 603                  * we aren't doing an inline extent round the compressed size
 604                  * up to a block size boundary so the allocator does sane
 605                  * things
 606                  */
 607                 total_compressed = ALIGN(total_compressed, blocksize);
 608
 609                 /*
 610                  * one last check to make sure the compression is really a
 611                  * win, compare the page count read with the blocks on disk,
 612                  * compression must free at least one sector size
 613                  */
 614                 total_in = ALIGN(total_in, PAGE_SIZE);
 615                 if (total_compressed + blocksize <= total_in) {
 616                         num_bytes = total_in;
 617                         *num_added += 1;
 618
 619                         /*
 620                          * The async work queues will take care of doing actual
 621                          * allocation on disk for these compressed pages, and
 622                          * will submit them to the elevator.
 623                          */
 624                         add_async_extent(async_cow, start, num_bytes,
 625                                         total_compressed, pages, nr_pages,
 626                                         compress_type);
 627
 628                         if (start + num_bytes < end) {
 629                                 start += num_bytes;
 630                                 pages = NULL;
 631                                 cond_resched();
 632                                 goto again;
 633                         }
 634                         return;
 635                 }
 636         }
 637         if (pages) {
 638                 /*
 639                  * the compression code ran but failed to make things smaller,
 640                  * free any pages it allocated and our page pointer array
 641                  */
 642                 for (i = 0; i < nr_pages; i++) {
 643                         WARN_ON(pages[i]->mapping);
 644                         put_page(pages[i]);
 645                 }
 646                 kfree(pages);
 647                 pages = NULL;
 648                 total_compressed = 0;
 649                 nr_pages = 0;
 650
 651                 /* flag the file so we don't compress in the future */
 652                 if (!btrfs_test_opt(fs_info, FORCE_COMPRESS) &&
 653                     !(BTRFS_I(inode)->prop_compress)) {
 654                         BTRFS_I(inode)->flags |= BTRFS_INODE_NOCOMPRESS;
 655                 }
 656         }
 657 cleanup_and_bail_uncompressed:
 658         /*
 659          * No compression, but we still need to write the pages in the file
 660          * we've been given so far.  redirty the locked page if it corresponds
 661          * to our extent and set things up for the async work queue to run
 662          * cow_file_range to do the normal delalloc dance.
 663          */
 664         if (page_offset(locked_page) >= start &&
 665             page_offset(locked_page) <= end)
 666                 __set_page_dirty_nobuffers(locked_page);
 667                 /* unlocked later on in the async handlers */
 668
 669         if (redirty)
 670                 extent_range_redirty_for_io(inode, start, end);
 671         add_async_extent(async_cow, start, end - start + 1, 0, NULL, 0,
 672                          BTRFS_COMPRESS_NONE);
 673         *num_added += 1;
 674
 675         return;
 676
 677 free_pages_out:
 678         for (i = 0; i < nr_pages; i++) {
 679                 WARN_ON(pages[i]->mapping);
 680                 put_page(pages[i]);
 681         }
 682         kfree(pages);
 683 }
 684
 685 static void free_async_extent_pages(struct async_extent *async_extent)
 686 {
 687         int i;
 688
 689         if (!async_extent->pages)
 690                 return;
 691
 692         for (i = 0; i < async_extent->nr_pages; i++) {
 693                 WARN_ON(async_extent->pages[i]->mapping);
 694                 put_page(async_extent->pages[i]);
 695         }
 696         kfree(async_extent->pages);
 697         async_extent->nr_pages = 0;
 698         async_extent->pages = NULL;
 699 }
 700
 701 /*
 702  * phase two of compressed writeback.  This is the ordered portion
 703  * of the code, which only gets called in the order the work was
 704  * queued.  We walk all the async extents created by compress_file_range
 705  * and send them down to the disk.
 706  */
 707 static noinline void submit_compressed_extents(struct inode *inode,
 708                                               struct async_cow *async_cow)
 709 {
 710         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
 711         struct async_extent *async_extent;
 712         u64 alloc_hint = 0;
 713         struct btrfs_key ins;
 714         struct extent_map *em;
 715         struct btrfs_root *root = BTRFS_I(inode)->root;
 716         struct extent_io_tree *io_tree;
 717         int ret = 0;
 718
 719 again:
 720         while (!list_empty(&async_cow->extents)) {
 721                 async_extent = list_entry(async_cow->extents.next,
 722                                           struct async_extent, list);
 723                 list_del(&async_extent->list);
 724
 725                 io_tree = &BTRFS_I(inode)->io_tree;
 726
 727 retry:
 728                 /* did the compression code fall back to uncompressed IO? */
 729                 if (!async_extent->pages) {
 730                         int page_started = 0;
 731                         unsigned long nr_written = 0;
 732
 733                         lock_extent(io_tree, async_extent->start,
 734                                          async_extent->start +
 735                                          async_extent->ram_size - 1);
 736
 737                         /* allocate blocks */
 738                         ret = cow_file_range(inode, async_cow->locked_page,
 739                                              async_extent->start,
 740                                              async_extent->start +
 741                                              async_extent->ram_size - 1,
 742                                              async_extent->start +
 743                                              async_extent->ram_size - 1,
 744                                              &page_started, &nr_written, 0,
 745                                              NULL);
 746
 747                         /* JDM XXX */
 748
 749                         /*
 750                          * if page_started, cow_file_range inserted an
 751                          * inline extent and took care of all the unlocking
 752                          * and IO for us.  Otherwise, we need to submit
 753                          * all those pages down to the drive.
 754                          */
 755                         if (!page_started && !ret)
 756                                 extent_write_locked_range(io_tree,
 757                                                   inode, async_extent->start,
 758                                                   async_extent->start +
 759                                                   async_extent->ram_size - 1,
 760                                                   btrfs_get_extent,
 761                                                   WB_SYNC_ALL);
 762                         else if (ret)
 763                                 unlock_page(async_cow->locked_page);
 764                         kfree(async_extent);
 765                         cond_resched();
 766                         continue;
 767                 }
 768
 769                 lock_extent(io_tree, async_extent->start,
 770                             async_extent->start + async_extent->ram_size - 1);
 771
 772                 ret = btrfs_reserve_extent(root, async_extent->ram_size,
 773                                            async_extent->compressed_size,
 774                                            async_extent->compressed_size,
 775                                            0, alloc_hint, &ins, 1, 1);
 776                 if (ret) {
 777                         free_async_extent_pages(async_extent);
 778
 779                         if (ret == -ENOSPC) {
 780                                 unlock_extent(io_tree, async_extent->start,
 781                                               async_extent->start +
 782                                               async_extent->ram_size - 1);
 783
 784                                 /*
 785                                  * we need to redirty the pages if we decide to
 786                                  * fallback to uncompressed IO, otherwise we
 787                                  * will not submit these pages down to lower
 788                                  * layers.
 789                                  */
 790                                 extent_range_redirty_for_io(inode,
 791                                                 async_extent->start,
 792                                                 async_extent->start +
 793                                                 async_extent->ram_size - 1);
 794
 795                                 goto retry;
 796                         }
 797                         goto out_free;
 798                 }
 799                 /*
 800                  * here we're doing allocation and writeback of the
 801                  * compressed pages
 802                  */
 803                 em = create_io_em(inode, async_extent->start,
 804                                   async_extent->ram_size, /* len */
 805                                   async_extent->start, /* orig_start */
 806                                   ins.objectid, /* block_start */
 807                                   ins.offset, /* block_len */
 808                                   ins.offset, /* orig_block_len */
 809                                   async_extent->ram_size, /* ram_bytes */
 810                                   async_extent->compress_type,
 811                                   BTRFS_ORDERED_COMPRESSED);
 812                 if (IS_ERR(em))
 813                         /* ret value is not necessary due to void function */
 814                         goto out_free_reserve;
 815                 free_extent_map(em);
 816
 817                 ret = btrfs_add_ordered_extent_compress(inode,
 818                                                 async_extent->start,
 819                                                 ins.objectid,
 820                                                 async_extent->ram_size,
 821                                                 ins.offset,
 822                                                 BTRFS_ORDERED_COMPRESSED,
 823                                                 async_extent->compress_type);
 824                 if (ret) {
 825                         btrfs_drop_extent_cache(BTRFS_I(inode),
 826                                                 async_extent->start,
 827                                                 async_extent->start +
 828                                                 async_extent->ram_size - 1, 0);
 829                         goto out_free_reserve;
 830                 }
 831                 btrfs_dec_block_group_reservations(fs_info, ins.objectid);
 832
 833                 /*
 834                  * clear dirty, set writeback and unlock the pages.
 835                  */
 836                 extent_clear_unlock_delalloc(inode, async_extent->start,
 837                                 async_extent->start +
 838                                 async_extent->ram_size - 1,
 839                                 async_extent->start +
 840                                 async_extent->ram_size - 1,
 841                                 NULL, EXTENT_LOCKED | EXTENT_DELALLOC,
 842                                 PAGE_UNLOCK | PAGE_CLEAR_DIRTY |
 843                                 PAGE_SET_WRITEBACK);
 844                 if (btrfs_submit_compressed_write(inode,
 845                                     async_extent->start,
 846                                     async_extent->ram_size,
 847                                     ins.objectid,
 848                                     ins.offset, async_extent->pages,
 849                                     async_extent->nr_pages)) {
 850                         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
 851                         struct page *p = async_extent->pages[0];
 852                         const u64 start = async_extent->start;
 853                         const u64 end = start + async_extent->ram_size - 1;
 854
 855                         p->mapping = inode->i_mapping;
 856                         tree->ops->writepage_end_io_hook(p, start, end,
 857                                                          NULL, 0);
 858                         p->mapping = NULL;
 859                         extent_clear_unlock_delalloc(inode, start, end, end,
 860                                                      NULL, 0,
 861                                                      PAGE_END_WRITEBACK |
 862                                                      PAGE_SET_ERROR);
 863                         free_async_extent_pages(async_extent);
 864                 }
 865                 alloc_hint = ins.objectid + ins.offset;
 866                 kfree(async_extent);
 867                 cond_resched();
 868         }
 869         return;
 870 out_free_reserve:
 871         btrfs_dec_block_group_reservations(fs_info, ins.objectid);
 872         btrfs_free_reserved_extent(fs_info, ins.objectid, ins.offset, 1);
 873 out_free:
 874         extent_clear_unlock_delalloc(inode, async_extent->start,
 875                                      async_extent->start +
 876                                      async_extent->ram_size - 1,
 877                                      async_extent->start +
 878                                      async_extent->ram_size - 1,
 879                                      NULL, EXTENT_LOCKED | EXTENT_DELALLOC |
 880                                      EXTENT_DELALLOC_NEW |
 881                                      EXTENT_DEFRAG | EXTENT_DO_ACCOUNTING,
 882                                      PAGE_UNLOCK | PAGE_CLEAR_DIRTY |
 883                                      PAGE_SET_WRITEBACK | PAGE_END_WRITEBACK |
 884                                      PAGE_SET_ERROR);
 885         free_async_extent_pages(async_extent);
 886         kfree(async_extent);
 887         goto again;
 888 }
 889
 890 static u64 get_extent_allocation_hint(struct inode *inode, u64 start,
 891                                       u64 num_bytes)
 892 {
 893         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
 894         struct extent_map *em;
 895         u64 alloc_hint = 0;
 896
 897         read_lock(&em_tree->lock);
 898         em = search_extent_mapping(em_tree, start, num_bytes);
 899         if (em) {
 900                 /*
 901                  * if block start isn't an actual block number then find the
 902                  * first block in this inode and use that as a hint.  If that
 903                  * block is also bogus then just don't worry about it.
 904                  */
 905                 if (em->block_start >= EXTENT_MAP_LAST_BYTE) {
 906                         free_extent_map(em);
 907                         em = search_extent_mapping(em_tree, 0, 0);
 908                         if (em && em->block_start < EXTENT_MAP_LAST_BYTE)
 909                                 alloc_hint = em->block_start;
 910                         if (em)
 911                                 free_extent_map(em);
 912                 } else {
 913                         alloc_hint = em->block_start;
 914                         free_extent_map(em);
 915                 }
 916         }
 917         read_unlock(&em_tree->lock);
 918
 919         return alloc_hint;
 920 }
 921
 922 /*
 923  * when extent_io.c finds a delayed allocation range in the file,
 924  * the call backs end up in this code.  The basic idea is to
 925  * allocate extents on disk for the range, and create ordered data structs
 926  * in ram to track those extents.
 927  *
 928  * locked_page is the page that writepage had locked already.  We use
 929  * it to make sure we don't do extra locks or unlocks.
 930  *
 931  * *page_started is set to one if we unlock locked_page and do everything
 932  * required to start IO on it.  It may be clean and already done with
 933  * IO when we return.
 934  */
 935 static noinline int cow_file_range(struct inode *inode,
 936                                    struct page *locked_page,
 937                                    u64 start, u64 end, u64 delalloc_end,
 938                                    int *page_started, unsigned long *nr_written,
 939                                    int unlock, struct btrfs_dedupe_hash *hash)
 940 {
 941         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
 942         struct btrfs_root *root = BTRFS_I(inode)->root;
 943         u64 alloc_hint = 0;
 944         u64 num_bytes;
 945         unsigned long ram_size;
 946         u64 disk_num_bytes;
 947         u64 cur_alloc_size = 0;
 948         u64 blocksize = fs_info->sectorsize;
 949         struct btrfs_key ins;
 950         struct extent_map *em;
 951         unsigned clear_bits;
 952         unsigned long page_ops;
 953         bool extent_reserved = false;
 954         int ret = 0;
 955
 956         if (btrfs_is_free_space_inode(BTRFS_I(inode))) {
 957                 WARN_ON_ONCE(1);
 958                 ret = -EINVAL;
 959                 goto out_unlock;
 960         }
 961
 962         num_bytes = ALIGN(end - start + 1, blocksize);
 963         num_bytes = max(blocksize,  num_bytes);
 964         disk_num_bytes = num_bytes;
 965
 966         inode_should_defrag(BTRFS_I(inode), start, end, num_bytes, SZ_64K);
 967
 968         if (start == 0) {
 969                 /* lets try to make an inline extent */
 970                 ret = cow_file_range_inline(root, inode, start, end, 0,
 971                                         BTRFS_COMPRESS_NONE, NULL);
 972                 if (ret == 0) {
 973                         extent_clear_unlock_delalloc(inode, start, end,
 974                                      delalloc_end, NULL,
 975                                      EXTENT_LOCKED | EXTENT_DELALLOC |
 976                                      EXTENT_DELALLOC_NEW |
 977                                      EXTENT_DEFRAG, PAGE_UNLOCK |
 978                                      PAGE_CLEAR_DIRTY | PAGE_SET_WRITEBACK |
 979                                      PAGE_END_WRITEBACK);
 980                         btrfs_free_reserved_data_space_noquota(inode, start,
 981                                                 end - start + 1);
 982                         *nr_written = *nr_written +
 983                              (end - start + PAGE_SIZE) / PAGE_SIZE;
 984                         *page_started = 1;
 985                         goto out;
 986                 } else if (ret < 0) {
 987                         goto out_unlock;
 988                 }
 989         }
 990
 991         BUG_ON(disk_num_bytes >
 992                btrfs_super_total_bytes(fs_info->super_copy));
 993
 994         alloc_hint = get_extent_allocation_hint(inode, start, num_bytes);
 995         btrfs_drop_extent_cache(BTRFS_I(inode), start,
 996                         start + num_bytes - 1, 0);
 997
 998         while (disk_num_bytes > 0) {
 999                 cur_alloc_size = disk_num_bytes;
1000                 ret = btrfs_reserve_extent(root, cur_alloc_size, cur_alloc_size,
1001                                            fs_info->sectorsize, 0, alloc_hint,
1002                                            &ins, 1, 1);
1003                 if (ret < 0)
1004                         goto out_unlock;
1005                 cur_alloc_size = ins.offset;
1006                 extent_reserved = true;
1007
1008                 ram_size = ins.offset;
1009                 em = create_io_em(inode, start, ins.offset, /* len */
1010                                   start, /* orig_start */
1011                                   ins.objectid, /* block_start */
1012                                   ins.offset, /* block_len */
1013                                   ins.offset, /* orig_block_len */
1014                                   ram_size, /* ram_bytes */
1015                                   BTRFS_COMPRESS_NONE, /* compress_type */
1016                                   BTRFS_ORDERED_REGULAR /* type */);
1017                 if (IS_ERR(em))
1018                         goto out_reserve;
1019                 free_extent_map(em);
1020
1021                 ret = btrfs_add_ordered_extent(inode, start, ins.objectid,
1022                                                ram_size, cur_alloc_size, 0);
1023                 if (ret)
1024                         goto out_drop_extent_cache;
1025
1026                 if (root->root_key.objectid ==
1027                     BTRFS_DATA_RELOC_TREE_OBJECTID) {
1028                         ret = btrfs_reloc_clone_csums(inode, start,
1029                                                       cur_alloc_size);
1030                         /*
1031                          * Only drop cache here, and process as normal.
1032                          *
1033                          * We must not allow extent_clear_unlock_delalloc()
1034                          * at out_unlock label to free meta of this ordered
1035                          * extent, as its meta should be freed by
1036                          * btrfs_finish_ordered_io().
1037                          *
1038                          * So we must continue until @start is increased to
1039                          * skip current ordered extent.
1040                          */
1041                         if (ret)
1042                                 btrfs_drop_extent_cache(BTRFS_I(inode), start,
1043                                                 start + ram_size - 1, 0);
1044                 }
1045
1046                 btrfs_dec_block_group_reservations(fs_info, ins.objectid);
1047
1048                 /* we're not doing compressed IO, don't unlock the first
1049                  * page (which the caller expects to stay locked), don't
1050                  * clear any dirty bits and don't set any writeback bits
1051                  *
1052                  * Do set the Private2 bit so we know this page was properly
1053                  * setup for writepage
1054                  */
1055                 page_ops = unlock ? PAGE_UNLOCK : 0;
1056                 page_ops |= PAGE_SET_PRIVATE2;
1057
1058                 extent_clear_unlock_delalloc(inode, start,
1059                                              start + ram_size - 1,
1060                                              delalloc_end, locked_page,
1061                                              EXTENT_LOCKED | EXTENT_DELALLOC,
1062                                              page_ops);
1063                 if (disk_num_bytes < cur_alloc_size)
1064                         disk_num_bytes = 0;
1065                 else
1066                         disk_num_bytes -= cur_alloc_size;
1067                 num_bytes -= cur_alloc_size;
1068                 alloc_hint = ins.objectid + ins.offset;
1069                 start += cur_alloc_size;
1070                 extent_reserved = false;
1071
1072                 /*
1073                  * btrfs_reloc_clone_csums() error, since start is increased
1074                  * extent_clear_unlock_delalloc() at out_unlock label won't
1075                  * free metadata of current ordered extent, we're OK to exit.
1076                  */
1077                 if (ret)
1078                         goto out_unlock;
1079         }
1080 out:
1081         return ret;
1082
1083 out_drop_extent_cache:
1084         btrfs_drop_extent_cache(BTRFS_I(inode), start, start + ram_size - 1, 0);
1085 out_reserve:
1086         btrfs_dec_block_group_reservations(fs_info, ins.objectid);
1087         btrfs_free_reserved_extent(fs_info, ins.objectid, ins.offset, 1);
1088 out_unlock:
1089         clear_bits = EXTENT_LOCKED | EXTENT_DELALLOC | EXTENT_DELALLOC_NEW |
1090                 EXTENT_DEFRAG | EXTENT_CLEAR_META_RESV;
1091         page_ops = PAGE_UNLOCK | PAGE_CLEAR_DIRTY | PAGE_SET_WRITEBACK |
1092                 PAGE_END_WRITEBACK;
1093         /*
1094          * If we reserved an extent for our delalloc range (or a subrange) and
1095          * failed to create the respective ordered extent, then it means that
1096          * when we reserved the extent we decremented the extent's size from
1097          * the data space_info's bytes_may_use counter and incremented the
1098          * space_info's bytes_reserved counter by the same amount. We must make
1099          * sure extent_clear_unlock_delalloc() does not try to decrement again
1100          * the data space_info's bytes_may_use counter, therefore we do not pass
1101          * it the flag EXTENT_CLEAR_DATA_RESV.
1102          */
1103         if (extent_reserved) {
1104                 extent_clear_unlock_delalloc(inode, start,
1105                                              start + cur_alloc_size,
1106                                              start + cur_alloc_size,
1107                                              locked_page,
1108                                              clear_bits,
1109                                              page_ops);
1110                 start += cur_alloc_size;
1111                 if (start >= end)
1112                         goto out;
1113         }
1114         extent_clear_unlock_delalloc(inode, start, end, delalloc_end,
1115                                      locked_page,
1116                                      clear_bits | EXTENT_CLEAR_DATA_RESV,
1117                                      page_ops);
1118         goto out;
1119 }
1120
1121 /*
1122  * work queue call back to started compression on a file and pages
1123  */
1124 static noinline void async_cow_start(struct btrfs_work *work)
1125 {
1126         struct async_cow *async_cow;
1127         int num_added = 0;
1128         async_cow = container_of(work, struct async_cow, work);
1129
1130         compress_file_range(async_cow->inode, async_cow->locked_page,
1131                             async_cow->start, async_cow->end, async_cow,
1132                             &num_added);
1133         if (num_added == 0) {
1134                 btrfs_add_delayed_iput(async_cow->inode);
1135                 async_cow->inode = NULL;
1136         }
1137 }
1138
1139 /*
1140  * work queue call back to submit previously compressed pages
1141  */
1142 static noinline void async_cow_submit(struct btrfs_work *work)
1143 {
1144         struct btrfs_fs_info *fs_info;
1145         struct async_cow *async_cow;
1146         struct btrfs_root *root;
1147         unsigned long nr_pages;
1148
1149         async_cow = container_of(work, struct async_cow, work);
1150
1151         root = async_cow->root;
1152         fs_info = root->fs_info;
1153         nr_pages = (async_cow->end - async_cow->start + PAGE_SIZE) >>
1154                 PAGE_SHIFT;
1155
1156         /*
1157          * atomic_sub_return implies a barrier for waitqueue_active
1158          */
1159         if (atomic_sub_return(nr_pages, &fs_info->async_delalloc_pages) <
1160             5 * SZ_1M &&
1161             waitqueue_active(&fs_info->async_submit_wait))
1162                 wake_up(&fs_info->async_submit_wait);
1163
1164         if (async_cow->inode)
1165                 submit_compressed_extents(async_cow->inode, async_cow);
1166 }
1167
1168 static noinline void async_cow_free(struct btrfs_work *work)
1169 {
1170         struct async_cow *async_cow;
1171         async_cow = container_of(work, struct async_cow, work);
1172         if (async_cow->inode)
1173                 btrfs_add_delayed_iput(async_cow->inode);
1174         kfree(async_cow);
1175 }
1176
1177 static int cow_file_range_async(struct inode *inode, struct page *locked_page,
1178                                 u64 start, u64 end, int *page_started,
1179                                 unsigned long *nr_written)
1180 {
1181         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1182         struct async_cow *async_cow;
1183         struct btrfs_root *root = BTRFS_I(inode)->root;
1184         unsigned long nr_pages;
1185         u64 cur_end;
1186
1187         clear_extent_bit(&BTRFS_I(inode)->io_tree, start, end, EXTENT_LOCKED,
1188                          1, 0, NULL, GFP_NOFS);
1189         while (start < end) {
1190                 async_cow = kmalloc(sizeof(*async_cow), GFP_NOFS);
1191                 BUG_ON(!async_cow); /* -ENOMEM */
1192                 async_cow->inode = igrab(inode);
1193                 async_cow->root = root;
1194                 async_cow->locked_page = locked_page;
1195                 async_cow->start = start;
1196
1197                 if (BTRFS_I(inode)->flags & BTRFS_INODE_NOCOMPRESS &&
1198                     !btrfs_test_opt(fs_info, FORCE_COMPRESS))
1199                         cur_end = end;
1200                 else
1201                         cur_end = min(end, start + SZ_512K - 1);
1202
1203                 async_cow->end = cur_end;
1204                 INIT_LIST_HEAD(&async_cow->extents);
1205
1206                 btrfs_init_work(&async_cow->work,
1207                                 btrfs_delalloc_helper,
1208                                 async_cow_start, async_cow_submit,
1209                                 async_cow_free);
1210
1211                 nr_pages = (cur_end - start + PAGE_SIZE) >>
1212                         PAGE_SHIFT;
1213                 atomic_add(nr_pages, &fs_info->async_delalloc_pages);
1214
1215                 btrfs_queue_work(fs_info->delalloc_workers, &async_cow->work);
1216
1217                 while (atomic_read(&fs_info->async_submit_draining) &&
1218                        atomic_read(&fs_info->async_delalloc_pages)) {
1219                         wait_event(fs_info->async_submit_wait,
1220                                    (atomic_read(&fs_info->async_delalloc_pages) ==
1221                                     0));
1222                 }
1223
1224                 *nr_written += nr_pages;
1225                 start = cur_end + 1;
1226         }
1227         *page_started = 1;
1228         return 0;
1229 }
1230
1231 static noinline int csum_exist_in_range(struct btrfs_fs_info *fs_info,
1232                                         u64 bytenr, u64 num_bytes)
1233 {
1234         int ret;
1235         struct btrfs_ordered_sum *sums;
1236         LIST_HEAD(list);
1237
1238         ret = btrfs_lookup_csums_range(fs_info->csum_root, bytenr,
1239                                        bytenr + num_bytes - 1, &list, 0);
1240         if (ret == 0 && list_empty(&list))
1241                 return 0;
1242
1243         while (!list_empty(&list)) {
1244                 sums = list_entry(list.next, struct btrfs_ordered_sum, list);
1245                 list_del(&sums->list);
1246                 kfree(sums);
1247         }
1248         return 1;
1249 }
1250
1251 /*
1252  * when nowcow writeback call back.  This checks for snapshots or COW copies
1253  * of the extents that exist in the file, and COWs the file as required.
1254  *
1255  * If no cow copies or snapshots exist, we write directly to the existing
1256  * blocks on disk
1257  */
1258 static noinline int run_delalloc_nocow(struct inode *inode,
1259                                        struct page *locked_page,
1260                               u64 start, u64 end, int *page_started, int force,
1261                               unsigned long *nr_written)
1262 {
1263         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1264         struct btrfs_root *root = BTRFS_I(inode)->root;
1265         struct extent_buffer *leaf;
1266         struct btrfs_path *path;
1267         struct btrfs_file_extent_item *fi;
1268         struct btrfs_key found_key;
1269         struct extent_map *em;
1270         u64 cow_start;
1271         u64 cur_offset;
1272         u64 extent_end;
1273         u64 extent_offset;
1274         u64 disk_bytenr;
1275         u64 num_bytes;
1276         u64 disk_num_bytes;
1277         u64 ram_bytes;
1278         int extent_type;
1279         int ret, err;
1280         int type;
1281         int nocow;
1282         int check_prev = 1;
1283         bool nolock;
1284         u64 ino = btrfs_ino(BTRFS_I(inode));
1285
1286         path = btrfs_alloc_path();
1287         if (!path) {
1288                 extent_clear_unlock_delalloc(inode, start, end, end,
1289                                              locked_page,
1290                                              EXTENT_LOCKED | EXTENT_DELALLOC |
1291                                              EXTENT_DO_ACCOUNTING |
1292                                              EXTENT_DEFRAG, PAGE_UNLOCK |
1293                                              PAGE_CLEAR_DIRTY |
1294                                              PAGE_SET_WRITEBACK |
1295                                              PAGE_END_WRITEBACK);
1296                 return -ENOMEM;
1297         }
1298
1299         nolock = btrfs_is_free_space_inode(BTRFS_I(inode));
1300
1301         cow_start = (u64)-1;
1302         cur_offset = start;
1303         while (1) {
1304                 ret = btrfs_lookup_file_extent(NULL, root, path, ino,
1305                                                cur_offset, 0);
1306                 if (ret < 0)
1307                         goto error;
1308                 if (ret > 0 && path->slots[0] > 0 && check_prev) {
1309                         leaf = path->nodes[0];
1310                         btrfs_item_key_to_cpu(leaf, &found_key,
1311                                               path->slots[0] - 1);
1312                         if (found_key.objectid == ino &&
1313                             found_key.type == BTRFS_EXTENT_DATA_KEY)
1314                                 path->slots[0]--;
1315                 }
1316                 check_prev = 0;
1317 next_slot:
1318                 leaf = path->nodes[0];
1319                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
1320                         ret = btrfs_next_leaf(root, path);
1321                         if (ret < 0)
1322                                 goto error;
1323                         if (ret > 0)
1324                                 break;
1325                         leaf = path->nodes[0];
1326                 }
1327
1328                 nocow = 0;
1329                 disk_bytenr = 0;
1330                 num_bytes = 0;
1331                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
1332
1333                 if (found_key.objectid > ino)
1334                         break;
1335                 if (WARN_ON_ONCE(found_key.objectid < ino) ||
1336                     found_key.type < BTRFS_EXTENT_DATA_KEY) {
1337                         path->slots[0]++;
1338                         goto next_slot;
1339                 }
1340                 if (found_key.type > BTRFS_EXTENT_DATA_KEY ||
1341                     found_key.offset > end)
1342                         break;
1343
1344                 if (found_key.offset > cur_offset) {
1345                         extent_end = found_key.offset;
1346                         extent_type = 0;
1347                         goto out_check;
1348                 }
1349
1350                 fi = btrfs_item_ptr(leaf, path->slots[0],
1351                                     struct btrfs_file_extent_item);
1352                 extent_type = btrfs_file_extent_type(leaf, fi);
1353
1354                 ram_bytes = btrfs_file_extent_ram_bytes(leaf, fi);
1355                 if (extent_type == BTRFS_FILE_EXTENT_REG ||
1356                     extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
1357                         disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1358                         extent_offset = btrfs_file_extent_offset(leaf, fi);
1359                         extent_end = found_key.offset +
1360                                 btrfs_file_extent_num_bytes(leaf, fi);
1361                         disk_num_bytes =
1362                                 btrfs_file_extent_disk_num_bytes(leaf, fi);
1363                         if (extent_end <= start) {
1364                                 path->slots[0]++;
1365                                 goto next_slot;
1366                         }
1367                         if (disk_bytenr == 0)
1368                                 goto out_check;
1369                         if (btrfs_file_extent_compression(leaf, fi) ||
1370                             btrfs_file_extent_encryption(leaf, fi) ||
1371                             btrfs_file_extent_other_encoding(leaf, fi))
1372                                 goto out_check;
1373                         if (extent_type == BTRFS_FILE_EXTENT_REG && !force)
1374                                 goto out_check;
1375                         if (btrfs_extent_readonly(fs_info, disk_bytenr))
1376                                 goto out_check;
1377                         if (btrfs_cross_ref_exist(root, ino,
1378                                                   found_key.offset -
1379                                                   extent_offset, disk_bytenr))
1380                                 goto out_check;
1381                         disk_bytenr += extent_offset;
1382                         disk_bytenr += cur_offset - found_key.offset;
1383                         num_bytes = min(end + 1, extent_end) - cur_offset;
1384                         /*
1385                          * if there are pending snapshots for this root,
1386                          * we fall into common COW way.
1387                          */
1388                         if (!nolock) {
1389                                 err = btrfs_start_write_no_snapshotting(root);
1390                                 if (!err)
1391                                         goto out_check;
1392                         }
1393                         /*
1394                          * force cow if csum exists in the range.
1395                          * this ensure that csum for a given extent are
1396                          * either valid or do not exist.
1397                          */
1398                         if (csum_exist_in_range(fs_info, disk_bytenr,
1399                                                 num_bytes)) {
1400                                 if (!nolock)
1401                                         btrfs_end_write_no_snapshotting(root);
1402                                 goto out_check;
1403                         }
1404                         if (!btrfs_inc_nocow_writers(fs_info, disk_bytenr)) {
1405                                 if (!nolock)
1406                                         btrfs_end_write_no_snapshotting(root);
1407                                 goto out_check;
1408                         }
1409                         nocow = 1;
1410                 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
1411                         extent_end = found_key.offset +
1412                                 btrfs_file_extent_inline_len(leaf,
1413                                                      path->slots[0], fi);
1414                         extent_end = ALIGN(extent_end,
1415                                            fs_info->sectorsize);
1416                 } else {
1417                         BUG_ON(1);
1418                 }
1419 out_check:
1420                 if (extent_end <= start) {
1421                         path->slots[0]++;
1422                         if (!nolock && nocow)
1423                                 btrfs_end_write_no_snapshotting(root);
1424                         if (nocow)
1425                                 btrfs_dec_nocow_writers(fs_info, disk_bytenr);
1426                         goto next_slot;
1427                 }
1428                 if (!nocow) {
1429                         if (cow_start == (u64)-1)
1430                                 cow_start = cur_offset;
1431                         cur_offset = extent_end;
1432                         if (cur_offset > end)
1433                                 break;
1434                         path->slots[0]++;
1435                         goto next_slot;
1436                 }
1437
1438                 btrfs_release_path(path);
1439                 if (cow_start != (u64)-1) {
1440                         ret = cow_file_range(inode, locked_page,
1441                                              cow_start, found_key.offset - 1,
1442                                              end, page_started, nr_written, 1,
1443                                              NULL);
1444                         if (ret) {
1445                                 if (!nolock && nocow)
1446                                         btrfs_end_write_no_snapshotting(root);
1447                                 if (nocow)
1448                                         btrfs_dec_nocow_writers(fs_info,
1449                                                                 disk_bytenr);
1450                                 goto error;
1451                         }
1452                         cow_start = (u64)-1;
1453                 }
1454
1455                 if (extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
1456                         u64 orig_start = found_key.offset - extent_offset;
1457
1458                         em = create_io_em(inode, cur_offset, num_bytes,
1459                                           orig_start,
1460                                           disk_bytenr, /* block_start */
1461                                           num_bytes, /* block_len */
1462                                           disk_num_bytes, /* orig_block_len */
1463                                           ram_bytes, BTRFS_COMPRESS_NONE,
1464                                           BTRFS_ORDERED_PREALLOC);
1465                         if (IS_ERR(em)) {
1466                                 if (!nolock && nocow)
1467                                         btrfs_end_write_no_snapshotting(root);
1468                                 if (nocow)
1469                                         btrfs_dec_nocow_writers(fs_info,
1470                                                                 disk_bytenr);
1471                                 ret = PTR_ERR(em);
1472                                 goto error;
1473                         }
1474                         free_extent_map(em);
1475                 }
1476
1477                 if (extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
1478                         type = BTRFS_ORDERED_PREALLOC;
1479                 } else {
1480                         type = BTRFS_ORDERED_NOCOW;
1481                 }
1482
1483                 ret = btrfs_add_ordered_extent(inode, cur_offset, disk_bytenr,
1484                                                num_bytes, num_bytes, type);
1485                 if (nocow)
1486                         btrfs_dec_nocow_writers(fs_info, disk_bytenr);
1487                 BUG_ON(ret); /* -ENOMEM */
1488
1489                 if (root->root_key.objectid ==
1490                     BTRFS_DATA_RELOC_TREE_OBJECTID)
1491                         /*
1492                          * Error handled later, as we must prevent
1493                          * extent_clear_unlock_delalloc() in error handler
1494                          * from freeing metadata of created ordered extent.
1495                          */
1496                         ret = btrfs_reloc_clone_csums(inode, cur_offset,
1497                                                       num_bytes);
1498
1499                 extent_clear_unlock_delalloc(inode, cur_offset,
1500                                              cur_offset + num_bytes - 1, end,
1501                                              locked_page, EXTENT_LOCKED |
1502                                              EXTENT_DELALLOC |
1503                                              EXTENT_CLEAR_DATA_RESV,
1504                                              PAGE_UNLOCK | PAGE_SET_PRIVATE2);
1505
1506                 if (!nolock && nocow)
1507                         btrfs_end_write_no_snapshotting(root);
1508                 cur_offset = extent_end;
1509
1510                 /*
1511                  * btrfs_reloc_clone_csums() error, now we're OK to call error
1512                  * handler, as metadata for created ordered extent will only
1513                  * be freed by btrfs_finish_ordered_io().
1514                  */
1515                 if (ret)
1516                         goto error;
1517                 if (cur_offset > end)
1518                         break;
1519         }
1520         btrfs_release_path(path);
1521
1522         if (cur_offset <= end && cow_start == (u64)-1) {
1523                 cow_start = cur_offset;
1524                 cur_offset = end;
1525         }
1526
1527         if (cow_start != (u64)-1) {
1528                 ret = cow_file_range(inode, locked_page, cow_start, end, end,
1529                                      page_started, nr_written, 1, NULL);
1530                 if (ret)
1531                         goto error;
1532         }
1533
1534 error:
1535         if (ret && cur_offset < end)
1536                 extent_clear_unlock_delalloc(inode, cur_offset, end, end,
1537                                              locked_page, EXTENT_LOCKED |
1538                                              EXTENT_DELALLOC | EXTENT_DEFRAG |
1539                                              EXTENT_DO_ACCOUNTING, PAGE_UNLOCK |
1540                                              PAGE_CLEAR_DIRTY |
1541                                              PAGE_SET_WRITEBACK |
1542                                              PAGE_END_WRITEBACK);
1543         btrfs_free_path(path);
1544         return ret;
1545 }
1546
1547 static inline int need_force_cow(struct inode *inode, u64 start, u64 end)
1548 {
1549
1550         if (!(BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW) &&
1551             !(BTRFS_I(inode)->flags & BTRFS_INODE_PREALLOC))
1552                 return 0;
1553
1554         /*
1555          * @defrag_bytes is a hint value, no spinlock held here,
1556          * if is not zero, it means the file is defragging.
1557          * Force cow if given extent needs to be defragged.
1558          */
1559         if (BTRFS_I(inode)->defrag_bytes &&
1560             test_range_bit(&BTRFS_I(inode)->io_tree, start, end,
1561                            EXTENT_DEFRAG, 0, NULL))
1562                 return 1;
1563
1564         return 0;
1565 }
1566
1567 /*
1568  * extent_io.c call back to do delayed allocation processing
1569  */
1570 static int run_delalloc_range(void *private_data, struct page *locked_page,
1571                               u64 start, u64 end, int *page_started,
1572                               unsigned long *nr_written)
1573 {
1574         struct inode *inode = private_data;
1575         int ret;
1576         int force_cow = need_force_cow(inode, start, end);
1577
1578         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW && !force_cow) {
1579                 ret = run_delalloc_nocow(inode, locked_page, start, end,
1580                                          page_started, 1, nr_written);
1581         } else if (BTRFS_I(inode)->flags & BTRFS_INODE_PREALLOC && !force_cow) {
1582                 ret = run_delalloc_nocow(inode, locked_page, start, end,
1583                                          page_started, 0, nr_written);
1584         } else if (!inode_need_compress(inode, start, end)) {
1585                 ret = cow_file_range(inode, locked_page, start, end, end,
1586                                       page_started, nr_written, 1, NULL);
1587         } else {
1588                 set_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
1589                         &BTRFS_I(inode)->runtime_flags);
1590                 ret = cow_file_range_async(inode, locked_page, start, end,
1591                                            page_started, nr_written);
1592         }
1593         if (ret)
1594                 btrfs_cleanup_ordered_extents(inode, start, end - start + 1);
1595         return ret;
1596 }
1597
1598 static void btrfs_split_extent_hook(void *private_data,
1599                                     struct extent_state *orig, u64 split)
1600 {
1601         struct inode *inode = private_data;
1602         u64 size;
1603
1604         /* not delalloc, ignore it */
1605         if (!(orig->state & EXTENT_DELALLOC))
1606                 return;
1607
1608         size = orig->end - orig->start + 1;
1609         if (size > BTRFS_MAX_EXTENT_SIZE) {
1610                 u32 num_extents;
1611                 u64 new_size;
1612
1613                 /*
1614                  * See the explanation in btrfs_merge_extent_hook, the same
1615                  * applies here, just in reverse.
1616                  */
1617                 new_size = orig->end - split + 1;
1618                 num_extents = count_max_extents(new_size);
1619                 new_size = split - orig->start;
1620                 num_extents += count_max_extents(new_size);
1621                 if (count_max_extents(size) >= num_extents)
1622                         return;
1623         }
1624
1625         spin_lock(&BTRFS_I(inode)->lock);
1626         BTRFS_I(inode)->outstanding_extents++;
1627         spin_unlock(&BTRFS_I(inode)->lock);
1628 }
1629
1630 /*
1631  * extent_io.c merge_extent_hook, used to track merged delayed allocation
1632  * extents so we can keep track of new extents that are just merged onto old
1633  * extents, such as when we are doing sequential writes, so we can properly
1634  * account for the metadata space we'll need.
1635  */
1636 static void btrfs_merge_extent_hook(void *private_data,
1637                                     struct extent_state *new,
1638                                     struct extent_state *other)
1639 {
1640         struct inode *inode = private_data;
1641         u64 new_size, old_size;
1642         u32 num_extents;
1643
1644         /* not delalloc, ignore it */
1645         if (!(other->state & EXTENT_DELALLOC))
1646                 return;
1647
1648         if (new->start > other->start)
1649                 new_size = new->end - other->start + 1;
1650         else
1651                 new_size = other->end - new->start + 1;
1652
1653         /* we're not bigger than the max, unreserve the space and go */
1654         if (new_size <= BTRFS_MAX_EXTENT_SIZE) {
1655                 spin_lock(&BTRFS_I(inode)->lock);
1656                 BTRFS_I(inode)->outstanding_extents--;
1657                 spin_unlock(&BTRFS_I(inode)->lock);
1658                 return;
1659         }
1660
1661         /*
1662          * We have to add up either side to figure out how many extents were
1663          * accounted for before we merged into one big extent.  If the number of
1664          * extents we accounted for is <= the amount we need for the new range
1665          * then we can return, otherwise drop.  Think of it like this
1666          *
1667          * [ 4k][MAX_SIZE]
1668          *
1669          * So we've grown the extent by a MAX_SIZE extent, this would mean we
1670          * need 2 outstanding extents, on one side we have 1 and the other side
1671          * we have 1 so they are == and we can return.  But in this case
1672          *
1673          * [MAX_SIZE+4k][MAX_SIZE+4k]
1674          *
1675          * Each range on their own accounts for 2 extents, but merged together
1676          * they are only 3 extents worth of accounting, so we need to drop in
1677          * this case.
1678          */
1679         old_size = other->end - other->start + 1;
1680         num_extents = count_max_extents(old_size);
1681         old_size = new->end - new->start + 1;
1682         num_extents += count_max_extents(old_size);
1683         if (count_max_extents(new_size) >= num_extents)
1684                 return;
1685
1686         spin_lock(&BTRFS_I(inode)->lock);
1687         BTRFS_I(inode)->outstanding_extents--;
1688         spin_unlock(&BTRFS_I(inode)->lock);
1689 }
1690
1691 static void btrfs_add_delalloc_inodes(struct btrfs_root *root,
1692                                       struct inode *inode)
1693 {
1694         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1695
1696         spin_lock(&root->delalloc_lock);
1697         if (list_empty(&BTRFS_I(inode)->delalloc_inodes)) {
1698                 list_add_tail(&BTRFS_I(inode)->delalloc_inodes,
1699                               &root->delalloc_inodes);
1700                 set_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1701                         &BTRFS_I(inode)->runtime_flags);
1702                 root->nr_delalloc_inodes++;
1703                 if (root->nr_delalloc_inodes == 1) {
1704                         spin_lock(&fs_info->delalloc_root_lock);
1705                         BUG_ON(!list_empty(&root->delalloc_root));
1706                         list_add_tail(&root->delalloc_root,
1707                                       &fs_info->delalloc_roots);
1708                         spin_unlock(&fs_info->delalloc_root_lock);
1709                 }
1710         }
1711         spin_unlock(&root->delalloc_lock);
1712 }
1713
1714 static void btrfs_del_delalloc_inode(struct btrfs_root *root,
1715                                      struct btrfs_inode *inode)
1716 {
1717         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
1718
1719         spin_lock(&root->delalloc_lock);
1720         if (!list_empty(&inode->delalloc_inodes)) {
1721                 list_del_init(&inode->delalloc_inodes);
1722                 clear_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1723                           &inode->runtime_flags);
1724                 root->nr_delalloc_inodes--;
1725                 if (!root->nr_delalloc_inodes) {
1726                         spin_lock(&fs_info->delalloc_root_lock);
1727                         BUG_ON(list_empty(&root->delalloc_root));
1728                         list_del_init(&root->delalloc_root);
1729                         spin_unlock(&fs_info->delalloc_root_lock);
1730                 }
1731         }
1732         spin_unlock(&root->delalloc_lock);
1733 }
1734
1735 /*
1736  * extent_io.c set_bit_hook, used to track delayed allocation
1737  * bytes in this file, and to maintain the list of inodes that
1738  * have pending delalloc work to be done.
1739  */
1740 static void btrfs_set_bit_hook(void *private_data,
1741                                struct extent_state *state, unsigned *bits)
1742 {
1743         struct inode *inode = private_data;
1744
1745         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1746
1747         if ((*bits & EXTENT_DEFRAG) && !(*bits & EXTENT_DELALLOC))
1748                 WARN_ON(1);
1749         /*
1750          * set_bit and clear bit hooks normally require _irqsave/restore
1751          * but in this case, we are only testing for the DELALLOC
1752          * bit, which is only set or cleared with irqs on
1753          */
1754         if (!(state->state & EXTENT_DELALLOC) && (*bits & EXTENT_DELALLOC)) {
1755                 struct btrfs_root *root = BTRFS_I(inode)->root;
1756                 u64 len = state->end + 1 - state->start;
1757                 bool do_list = !btrfs_is_free_space_inode(BTRFS_I(inode));
1758
1759                 if (*bits & EXTENT_FIRST_DELALLOC) {
1760                         *bits &= ~EXTENT_FIRST_DELALLOC;
1761                 } else {
1762                         spin_lock(&BTRFS_I(inode)->lock);
1763                         BTRFS_I(inode)->outstanding_extents++;
1764                         spin_unlock(&BTRFS_I(inode)->lock);
1765                 }
1766
1767                 /* For sanity tests */
1768                 if (btrfs_is_testing(fs_info))
1769                         return;
1770
1771                 percpu_counter_add_batch(&fs_info->delalloc_bytes, len,
1772                                          fs_info->delalloc_batch);
1773                 spin_lock(&BTRFS_I(inode)->lock);
1774                 BTRFS_I(inode)->delalloc_bytes += len;
1775                 if (*bits & EXTENT_DEFRAG)
1776                         BTRFS_I(inode)->defrag_bytes += len;
1777                 if (do_list && !test_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1778                                          &BTRFS_I(inode)->runtime_flags))
1779                         btrfs_add_delalloc_inodes(root, inode);
1780                 spin_unlock(&BTRFS_I(inode)->lock);
1781         }
1782
1783         if (!(state->state & EXTENT_DELALLOC_NEW) &&
1784             (*bits & EXTENT_DELALLOC_NEW)) {
1785                 spin_lock(&BTRFS_I(inode)->lock);
1786                 BTRFS_I(inode)->new_delalloc_bytes += state->end + 1 -
1787                         state->start;
1788                 spin_unlock(&BTRFS_I(inode)->lock);
1789         }
1790 }
1791
1792 /*
1793  * extent_io.c clear_bit_hook, see set_bit_hook for why
1794  */
1795 static void btrfs_clear_bit_hook(void *private_data,
1796                                  struct extent_state *state,
1797                                  unsigned *bits)
1798 {
1799         struct btrfs_inode *inode = BTRFS_I((struct inode *)private_data);
1800         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
1801         u64 len = state->end + 1 - state->start;
1802         u32 num_extents = count_max_extents(len);
1803
1804         if ((state->state & EXTENT_DEFRAG) && (*bits & EXTENT_DEFRAG)) {
1805                 spin_lock(&inode->lock);
1806                 inode->defrag_bytes -= len;
1807                 spin_unlock(&inode->lock);
1808         }
1809
1810         /*
1811          * set_bit and clear bit hooks normally require _irqsave/restore
1812          * but in this case, we are only testing for the DELALLOC
1813          * bit, which is only set or cleared with irqs on
1814          */
1815         if ((state->state & EXTENT_DELALLOC) && (*bits & EXTENT_DELALLOC)) {
1816                 struct btrfs_root *root = inode->root;
1817                 bool do_list = !btrfs_is_free_space_inode(inode);
1818
1819                 if (*bits & EXTENT_FIRST_DELALLOC) {
1820                         *bits &= ~EXTENT_FIRST_DELALLOC;
1821                 } else if (!(*bits & EXTENT_CLEAR_META_RESV)) {
1822                         spin_lock(&inode->lock);
1823                         inode->outstanding_extents -= num_extents;
1824                         spin_unlock(&inode->lock);
1825                 }
1826
1827                 /*
1828                  * We don't reserve metadata space for space cache inodes so we
1829                  * don't need to call dellalloc_release_metadata if there is an
1830                  * error.
1831                  */
1832                 if (*bits & EXTENT_CLEAR_META_RESV &&
1833                     root != fs_info->tree_root)
1834                         btrfs_delalloc_release_metadata(inode, len);
1835
1836                 /* For sanity tests. */
1837                 if (btrfs_is_testing(fs_info))
1838                         return;
1839
1840                 if (root->root_key.objectid != BTRFS_DATA_RELOC_TREE_OBJECTID &&
1841                     do_list && !(state->state & EXTENT_NORESERVE) &&
1842                     (*bits & EXTENT_CLEAR_DATA_RESV))
1843                         btrfs_free_reserved_data_space_noquota(
1844                                         &inode->vfs_inode,
1845                                         state->start, len);
1846
1847                 percpu_counter_add_batch(&fs_info->delalloc_bytes, -len,
1848                                          fs_info->delalloc_batch);
1849                 spin_lock(&inode->lock);
1850                 inode->delalloc_bytes -= len;
1851                 if (do_list && inode->delalloc_bytes == 0 &&
1852                     test_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1853                                         &inode->runtime_flags))
1854                         btrfs_del_delalloc_inode(root, inode);
1855                 spin_unlock(&inode->lock);
1856         }
1857
1858         if ((state->state & EXTENT_DELALLOC_NEW) &&
1859             (*bits & EXTENT_DELALLOC_NEW)) {
1860                 spin_lock(&inode->lock);
1861                 ASSERT(inode->new_delalloc_bytes >= len);
1862                 inode->new_delalloc_bytes -= len;
1863                 spin_unlock(&inode->lock);
1864         }
1865 }
1866
1867 /*
1868  * extent_io.c merge_bio_hook, this must check the chunk tree to make sure
1869  * we don't create bios that span stripes or chunks
1870  *
1871  * return 1 if page cannot be merged to bio
1872  * return 0 if page can be merged to bio
1873  * return error otherwise
1874  */
1875 int btrfs_merge_bio_hook(struct page *page, unsigned long offset,
1876                          size_t size, struct bio *bio,
1877                          unsigned long bio_flags)
1878 {
1879         struct inode *inode = page->mapping->host;
1880         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1881         u64 logical = (u64)bio->bi_iter.bi_sector << 9;
1882         u64 length = 0;
1883         u64 map_length;
1884         int ret;
1885
1886         if (bio_flags & EXTENT_BIO_COMPRESSED)
1887                 return 0;
1888
1889         length = bio->bi_iter.bi_size;
1890         map_length = length;
1891         ret = btrfs_map_block(fs_info, btrfs_op(bio), logical, &map_length,
1892                               NULL, 0);
1893         if (ret < 0)
1894                 return ret;
1895         if (map_length < length + size)
1896                 return 1;
1897         return 0;
1898 }
1899
1900 /*
1901  * in order to insert checksums into the metadata in large chunks,
1902  * we wait until bio submission time.   All the pages in the bio are
1903  * checksummed and sums are attached onto the ordered extent record.
1904  *
1905  * At IO completion time the cums attached on the ordered extent record
1906  * are inserted into the btree
1907  */
1908 static blk_status_t __btrfs_submit_bio_start(void *private_data, struct bio *bio,
1909                                     int mirror_num, unsigned long bio_flags,
1910                                     u64 bio_offset)
1911 {
1912         struct inode *inode = private_data;
1913         blk_status_t ret = 0;
1914
1915         ret = btrfs_csum_one_bio(inode, bio, 0, 0);
1916         BUG_ON(ret); /* -ENOMEM */
1917         return 0;
1918 }
1919
1920 /*
1921  * in order to insert checksums into the metadata in large chunks,
1922  * we wait until bio submission time.   All the pages in the bio are
1923  * checksummed and sums are attached onto the ordered extent record.
1924  *
1925  * At IO completion time the cums attached on the ordered extent record
1926  * are inserted into the btree
1927  */
1928 static blk_status_t __btrfs_submit_bio_done(void *private_data, struct bio *bio,
1929                           int mirror_num, unsigned long bio_flags,
1930                           u64 bio_offset)
1931 {
1932         struct inode *inode = private_data;
1933         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1934         blk_status_t ret;
1935
1936         ret = btrfs_map_bio(fs_info, bio, mirror_num, 1);
1937         if (ret) {
1938                 bio->bi_status = ret;
1939                 bio_endio(bio);
1940         }
1941         return ret;
1942 }
1943
1944 /*
1945  * extent_io.c submission hook. This does the right thing for csum calculation
1946  * on write, or reading the csums from the tree before a read
1947  */
1948 static blk_status_t btrfs_submit_bio_hook(void *private_data, struct bio *bio,
1949                                  int mirror_num, unsigned long bio_flags,
1950                                  u64 bio_offset)
1951 {
1952         struct inode *inode = private_data;
1953         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1954         struct btrfs_root *root = BTRFS_I(inode)->root;
1955         enum btrfs_wq_endio_type metadata = BTRFS_WQ_ENDIO_DATA;
1956         blk_status_t ret = 0;
1957         int skip_sum;
1958         int async = !atomic_read(&BTRFS_I(inode)->sync_writers);
1959
1960         skip_sum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM;
1961
1962         if (btrfs_is_free_space_inode(BTRFS_I(inode)))
1963                 metadata = BTRFS_WQ_ENDIO_FREE_SPACE;
1964
1965         if (bio_op(bio) != REQ_OP_WRITE) {
1966                 ret = btrfs_bio_wq_end_io(fs_info, bio, metadata);
1967                 if (ret)
1968                         goto out;
1969
1970                 if (bio_flags & EXTENT_BIO_COMPRESSED) {
1971                         ret = btrfs_submit_compressed_read(inode, bio,
1972                                                            mirror_num,
1973                                                            bio_flags);
1974                         goto out;
1975                 } else if (!skip_sum) {
1976                         ret = btrfs_lookup_bio_sums(inode, bio, NULL);
1977                         if (ret)
1978                                 goto out;
1979                 }
1980                 goto mapit;
1981         } else if (async && !skip_sum) {
1982                 /* csum items have already been cloned */
1983                 if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
1984                         goto mapit;
1985                 /* we're doing a write, do the async checksumming */
1986                 ret = btrfs_wq_submit_bio(fs_info, bio, mirror_num, bio_flags,
1987                                           bio_offset, inode,
1988                                           __btrfs_submit_bio_start,
1989                                           __btrfs_submit_bio_done);
1990                 goto out;
1991         } else if (!skip_sum) {
1992                 ret = btrfs_csum_one_bio(inode, bio, 0, 0);
1993                 if (ret)
1994                         goto out;
1995         }
1996
1997 mapit:
1998         ret = btrfs_map_bio(fs_info, bio, mirror_num, 0);
1999
2000 out:
2001         if (ret) {
2002                 bio->bi_status = ret;
2003                 bio_endio(bio);
2004         }
2005         return ret;
2006 }
2007
2008 /*
2009  * given a list of ordered sums record them in the inode.  This happens
2010  * at IO completion time based on sums calculated at bio submission time.
2011  */
2012 static noinline int add_pending_csums(struct btrfs_trans_handle *trans,
2013                              struct inode *inode, struct list_head *list)
2014 {
2015         struct btrfs_ordered_sum *sum;
2016
2017         list_for_each_entry(sum, list, list) {
2018                 trans->adding_csums = 1;
2019                 btrfs_csum_file_blocks(trans,
2020                        BTRFS_I(inode)->root->fs_info->csum_root, sum);
2021                 trans->adding_csums = 0;
2022         }
2023         return 0;
2024 }
2025
2026 int btrfs_set_extent_delalloc(struct inode *inode, u64 start, u64 end,
2027                               struct extent_state **cached_state, int dedupe)
2028 {
2029         WARN_ON((end & (PAGE_SIZE - 1)) == 0);
2030         return set_extent_delalloc(&BTRFS_I(inode)->io_tree, start, end,
2031                                    cached_state);
2032 }
2033
2034 /* see btrfs_writepage_start_hook for details on why this is required */
2035 struct btrfs_writepage_fixup {
2036         struct page *page;
2037         struct btrfs_work work;
2038 };
2039
2040 static void btrfs_writepage_fixup_worker(struct btrfs_work *work)
2041 {
2042         struct btrfs_writepage_fixup *fixup;
2043         struct btrfs_ordered_extent *ordered;
2044         struct extent_state *cached_state = NULL;
2045         struct extent_changeset *data_reserved = NULL;
2046         struct page *page;
2047         struct inode *inode;
2048         u64 page_start;
2049         u64 page_end;
2050         int ret;
2051
2052         fixup = container_of(work, struct btrfs_writepage_fixup, work);
2053         page = fixup->page;
2054 again:
2055         lock_page(page);
2056         if (!page->mapping || !PageDirty(page) || !PageChecked(page)) {
2057                 ClearPageChecked(page);
2058                 goto out_page;
2059         }
2060
2061         inode = page->mapping->host;
2062         page_start = page_offset(page);
2063         page_end = page_offset(page) + PAGE_SIZE - 1;
2064
2065         lock_extent_bits(&BTRFS_I(inode)->io_tree, page_start, page_end,
2066                          &cached_state);
2067
2068         /* already ordered? We're done */
2069         if (PagePrivate2(page))
2070                 goto out;
2071
2072         ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), page_start,
2073                                         PAGE_SIZE);
2074         if (ordered) {
2075                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, page_start,
2076                                      page_end, &cached_state, GFP_NOFS);
2077                 unlock_page(page);
2078                 btrfs_start_ordered_extent(inode, ordered, 1);
2079                 btrfs_put_ordered_extent(ordered);
2080                 goto again;
2081         }
2082
2083         ret = btrfs_delalloc_reserve_space(inode, &data_reserved, page_start,
2084                                            PAGE_SIZE);
2085         if (ret) {
2086                 mapping_set_error(page->mapping, ret);
2087                 end_extent_writepage(page, ret, page_start, page_end);
2088                 ClearPageChecked(page);
2089                 goto out;
2090          }
2091
2092         btrfs_set_extent_delalloc(inode, page_start, page_end, &cached_state,
2093                                   0);
2094         ClearPageChecked(page);
2095         set_page_dirty(page);
2096 out:
2097         unlock_extent_cached(&BTRFS_I(inode)->io_tree, page_start, page_end,
2098                              &cached_state, GFP_NOFS);
2099 out_page:
2100         unlock_page(page);
2101         put_page(page);
2102         kfree(fixup);
2103         extent_changeset_free(data_reserved);
2104 }
2105
2106 /*
2107  * There are a few paths in the higher layers of the kernel that directly
2108  * set the page dirty bit without asking the filesystem if it is a
2109  * good idea.  This causes problems because we want to make sure COW
2110  * properly happens and the data=ordered rules are followed.
2111  *
2112  * In our case any range that doesn't have the ORDERED bit set
2113  * hasn't been properly setup for IO.  We kick off an async process
2114  * to fix it up.  The async helper will wait for ordered extents, set
2115  * the delalloc bit and make it safe to write the page.
2116  */
2117 static int btrfs_writepage_start_hook(struct page *page, u64 start, u64 end)
2118 {
2119         struct inode *inode = page->mapping->host;
2120         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2121         struct btrfs_writepage_fixup *fixup;
2122
2123         /* this page is properly in the ordered list */
2124         if (TestClearPagePrivate2(page))
2125                 return 0;
2126
2127         if (PageChecked(page))
2128                 return -EAGAIN;
2129
2130         fixup = kzalloc(sizeof(*fixup), GFP_NOFS);
2131         if (!fixup)
2132                 return -EAGAIN;
2133
2134         SetPageChecked(page);
2135         get_page(page);
2136         btrfs_init_work(&fixup->work, btrfs_fixup_helper,
2137                         btrfs_writepage_fixup_worker, NULL, NULL);
2138         fixup->page = page;
2139         btrfs_queue_work(fs_info->fixup_workers, &fixup->work);
2140         return -EBUSY;
2141 }
2142
2143 static int insert_reserved_file_extent(struct btrfs_trans_handle *trans,
2144                                        struct inode *inode, u64 file_pos,
2145                                        u64 disk_bytenr, u64 disk_num_bytes,
2146                                        u64 num_bytes, u64 ram_bytes,
2147                                        u8 compression, u8 encryption,
2148                                        u16 other_encoding, int extent_type)
2149 {
2150         struct btrfs_root *root = BTRFS_I(inode)->root;
2151         struct btrfs_file_extent_item *fi;
2152         struct btrfs_path *path;
2153         struct extent_buffer *leaf;
2154         struct btrfs_key ins;
2155         u64 qg_released;
2156         int extent_inserted = 0;
2157         int ret;
2158
2159         path = btrfs_alloc_path();
2160         if (!path)
2161                 return -ENOMEM;
2162
2163         /*
2164          * we may be replacing one extent in the tree with another.
2165          * The new extent is pinned in the extent map, and we don't want
2166          * to drop it from the cache until it is completely in the btree.
2167          *
2168          * So, tell btrfs_drop_extents to leave this extent in the cache.
2169          * the caller is expected to unpin it and allow it to be merged
2170          * with the others.
2171          */
2172         ret = __btrfs_drop_extents(trans, root, inode, path, file_pos,
2173                                    file_pos + num_bytes, NULL, 0,
2174                                    1, sizeof(*fi), &extent_inserted);
2175         if (ret)
2176                 goto out;
2177
2178         if (!extent_inserted) {
2179                 ins.objectid = btrfs_ino(BTRFS_I(inode));
2180                 ins.offset = file_pos;
2181                 ins.type = BTRFS_EXTENT_DATA_KEY;
2182
2183                 path->leave_spinning = 1;
2184                 ret = btrfs_insert_empty_item(trans, root, path, &ins,
2185                                               sizeof(*fi));
2186                 if (ret)
2187                         goto out;
2188         }
2189         leaf = path->nodes[0];
2190         fi = btrfs_item_ptr(leaf, path->slots[0],
2191                             struct btrfs_file_extent_item);
2192         btrfs_set_file_extent_generation(leaf, fi, trans->transid);
2193         btrfs_set_file_extent_type(leaf, fi, extent_type);
2194         btrfs_set_file_extent_disk_bytenr(leaf, fi, disk_bytenr);
2195         btrfs_set_file_extent_disk_num_bytes(leaf, fi, disk_num_bytes);
2196         btrfs_set_file_extent_offset(leaf, fi, 0);
2197         btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2198         btrfs_set_file_extent_ram_bytes(leaf, fi, ram_bytes);
2199         btrfs_set_file_extent_compression(leaf, fi, compression);
2200         btrfs_set_file_extent_encryption(leaf, fi, encryption);
2201         btrfs_set_file_extent_other_encoding(leaf, fi, other_encoding);
2202
2203         btrfs_mark_buffer_dirty(leaf);
2204         btrfs_release_path(path);
2205
2206         inode_add_bytes(inode, num_bytes);
2207
2208         ins.objectid = disk_bytenr;
2209         ins.offset = disk_num_bytes;
2210         ins.type = BTRFS_EXTENT_ITEM_KEY;
2211
2212         /*
2213          * Release the reserved range from inode dirty range map, as it is
2214          * already moved into delayed_ref_head
2215          */
2216         ret = btrfs_qgroup_release_data(inode, file_pos, ram_bytes);
2217         if (ret < 0)
2218                 goto out;
2219         qg_released = ret;
2220         ret = btrfs_alloc_reserved_file_extent(trans, root->root_key.objectid,
2221                         btrfs_ino(BTRFS_I(inode)), file_pos, qg_released, &ins);
2222 out:
2223         btrfs_free_path(path);
2224
2225         return ret;
2226 }
2227
2228 /* snapshot-aware defrag */
2229 struct sa_defrag_extent_backref {
2230         struct rb_node node;
2231         struct old_sa_defrag_extent *old;
2232         u64 root_id;
2233         u64 inum;
2234         u64 file_pos;
2235         u64 extent_offset;
2236         u64 num_bytes;
2237         u64 generation;
2238 };
2239
2240 struct old_sa_defrag_extent {
2241         struct list_head list;
2242         struct new_sa_defrag_extent *new;
2243
2244         u64 extent_offset;
2245         u64 bytenr;
2246         u64 offset;
2247         u64 len;
2248         int count;
2249 };
2250
2251 struct new_sa_defrag_extent {
2252         struct rb_root root;
2253         struct list_head head;
2254         struct btrfs_path *path;
2255         struct inode *inode;
2256         u64 file_pos;
2257         u64 len;
2258         u64 bytenr;
2259         u64 disk_len;
2260         u8 compress_type;
2261 };
2262
2263 static int backref_comp(struct sa_defrag_extent_backref *b1,
2264                         struct sa_defrag_extent_backref *b2)
2265 {
2266         if (b1->root_id < b2->root_id)
2267                 return -1;
2268         else if (b1->root_id > b2->root_id)
2269                 return 1;
2270
2271         if (b1->inum < b2->inum)
2272                 return -1;
2273         else if (b1->inum > b2->inum)
2274                 return 1;
2275
2276         if (b1->file_pos < b2->file_pos)
2277                 return -1;
2278         else if (b1->file_pos > b2->file_pos)
2279                 return 1;
2280
2281         /*
2282          * [------------------------------] ===> (a range of space)
2283          *     |<--->|   |<---->| =============> (fs/file tree A)
2284          * |<---------------------------->| ===> (fs/file tree B)
2285          *
2286          * A range of space can refer to two file extents in one tree while
2287          * refer to only one file extent in another tree.
2288          *
2289          * So we may process a disk offset more than one time(two extents in A)
2290          * and locate at the same extent(one extent in B), then insert two same
2291          * backrefs(both refer to the extent in B).
2292          */
2293         return 0;
2294 }
2295
2296 static void backref_insert(struct rb_root *root,
2297                            struct sa_defrag_extent_backref *backref)
2298 {
2299         struct rb_node **p = &root->rb_node;
2300         struct rb_node *parent = NULL;
2301         struct sa_defrag_extent_backref *entry;
2302         int ret;
2303
2304         while (*p) {
2305                 parent = *p;
2306                 entry = rb_entry(parent, struct sa_defrag_extent_backref, node);
2307
2308                 ret = backref_comp(backref, entry);
2309                 if (ret < 0)
2310                         p = &(*p)->rb_left;
2311                 else
2312                         p = &(*p)->rb_right;
2313         }
2314
2315         rb_link_node(&backref->node, parent, p);
2316         rb_insert_color(&backref->node, root);
2317 }
2318
2319 /*
2320  * Note the backref might has changed, and in this case we just return 0.
2321  */
2322 static noinline int record_one_backref(u64 inum, u64 offset, u64 root_id,
2323                                        void *ctx)
2324 {
2325         struct btrfs_file_extent_item *extent;
2326         struct old_sa_defrag_extent *old = ctx;
2327         struct new_sa_defrag_extent *new = old->new;
2328         struct btrfs_path *path = new->path;
2329         struct btrfs_key key;
2330         struct btrfs_root *root;
2331         struct sa_defrag_extent_backref *backref;
2332         struct extent_buffer *leaf;
2333         struct inode *inode = new->inode;
2334         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2335         int slot;
2336         int ret;
2337         u64 extent_offset;
2338         u64 num_bytes;
2339
2340         if (BTRFS_I(inode)->root->root_key.objectid == root_id &&
2341             inum == btrfs_ino(BTRFS_I(inode)))
2342                 return 0;
2343
2344         key.objectid = root_id;
2345         key.type = BTRFS_ROOT_ITEM_KEY;
2346         key.offset = (u64)-1;
2347
2348         root = btrfs_read_fs_root_no_name(fs_info, &key);
2349         if (IS_ERR(root)) {
2350                 if (PTR_ERR(root) == -ENOENT)
2351                         return 0;
2352                 WARN_ON(1);
2353                 btrfs_debug(fs_info, "inum=%llu, offset=%llu, root_id=%llu",
2354                          inum, offset, root_id);
2355                 return PTR_ERR(root);
2356         }
2357
2358         key.objectid = inum;
2359         key.type = BTRFS_EXTENT_DATA_KEY;
2360         if (offset > (u64)-1 << 32)
2361                 key.offset = 0;
2362         else
2363                 key.offset = offset;
2364
2365         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2366         if (WARN_ON(ret < 0))
2367                 return ret;
2368         ret = 0;
2369
2370         while (1) {
2371                 cond_resched();
2372
2373                 leaf = path->nodes[0];
2374                 slot = path->slots[0];
2375
2376                 if (slot >= btrfs_header_nritems(leaf)) {
2377                         ret = btrfs_next_leaf(root, path);
2378                         if (ret < 0) {
2379                                 goto out;
2380                         } else if (ret > 0) {
2381                                 ret = 0;
2382                                 goto out;
2383                         }
2384                         continue;
2385                 }
2386
2387                 path->slots[0]++;
2388
2389                 btrfs_item_key_to_cpu(leaf, &key, slot);
2390
2391                 if (key.objectid > inum)
2392                         goto out;
2393
2394                 if (key.objectid < inum || key.type != BTRFS_EXTENT_DATA_KEY)
2395                         continue;
2396
2397                 extent = btrfs_item_ptr(leaf, slot,
2398                                         struct btrfs_file_extent_item);
2399
2400                 if (btrfs_file_extent_disk_bytenr(leaf, extent) != old->bytenr)
2401                         continue;
2402
2403                 /*
2404                  * 'offset' refers to the exact key.offset,
2405                  * NOT the 'offset' field in btrfs_extent_data_ref, ie.
2406                  * (key.offset - extent_offset).
2407                  */
2408                 if (key.offset != offset)
2409                         continue;
2410
2411                 extent_offset = btrfs_file_extent_offset(leaf, extent);
2412                 num_bytes = btrfs_file_extent_num_bytes(leaf, extent);
2413
2414                 if (extent_offset >= old->extent_offset + old->offset +
2415                     old->len || extent_offset + num_bytes <=
2416                     old->extent_offset + old->offset)
2417                         continue;
2418                 break;
2419         }
2420
2421         backref = kmalloc(sizeof(*backref), GFP_NOFS);
2422         if (!backref) {
2423                 ret = -ENOENT;
2424                 goto out;
2425         }
2426
2427         backref->root_id = root_id;
2428         backref->inum = inum;
2429         backref->file_pos = offset;
2430         backref->num_bytes = num_bytes;
2431         backref->extent_offset = extent_offset;
2432         backref->generation = btrfs_file_extent_generation(leaf, extent);
2433         backref->old = old;
2434         backref_insert(&new->root, backref);
2435         old->count++;
2436 out:
2437         btrfs_release_path(path);
2438         WARN_ON(ret);
2439         return ret;
2440 }
2441
2442 static noinline bool record_extent_backrefs(struct btrfs_path *path,
2443                                    struct new_sa_defrag_extent *new)
2444 {
2445         struct btrfs_fs_info *fs_info = btrfs_sb(new->inode->i_sb);
2446         struct old_sa_defrag_extent *old, *tmp;
2447         int ret;
2448
2449         new->path = path;
2450
2451         list_for_each_entry_safe(old, tmp, &new->head, list) {
2452                 ret = iterate_inodes_from_logical(old->bytenr +
2453                                                   old->extent_offset, fs_info,
2454                                                   path, record_one_backref,
2455                                                   old);
2456                 if (ret < 0 && ret != -ENOENT)
2457                         return false;
2458
2459                 /* no backref to be processed for this extent */
2460                 if (!old->count) {
2461                         list_del(&old->list);
2462                         kfree(old);
2463                 }
2464         }
2465
2466         if (list_empty(&new->head))
2467                 return false;
2468
2469         return true;
2470 }
2471
2472 static int relink_is_mergable(struct extent_buffer *leaf,
2473                               struct btrfs_file_extent_item *fi,
2474                               struct new_sa_defrag_extent *new)
2475 {
2476         if (btrfs_file_extent_disk_bytenr(leaf, fi) != new->bytenr)
2477                 return 0;
2478
2479         if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG)
2480                 return 0;
2481
2482         if (btrfs_file_extent_compression(leaf, fi) != new->compress_type)
2483                 return 0;
2484
2485         if (btrfs_file_extent_encryption(leaf, fi) ||
2486             btrfs_file_extent_other_encoding(leaf, fi))
2487                 return 0;
2488
2489         return 1;
2490 }
2491
2492 /*
2493  * Note the backref might has changed, and in this case we just return 0.
2494  */
2495 static noinline int relink_extent_backref(struct btrfs_path *path,
2496                                  struct sa_defrag_extent_backref *prev,
2497                                  struct sa_defrag_extent_backref *backref)
2498 {
2499         struct btrfs_file_extent_item *extent;
2500         struct btrfs_file_extent_item *item;
2501         struct btrfs_ordered_extent *ordered;
2502         struct btrfs_trans_handle *trans;
2503         struct btrfs_root *root;
2504         struct btrfs_key key;
2505         struct extent_buffer *leaf;
2506         struct old_sa_defrag_extent *old = backref->old;
2507         struct new_sa_defrag_extent *new = old->new;
2508         struct btrfs_fs_info *fs_info = btrfs_sb(new->inode->i_sb);
2509         struct inode *inode;
2510         struct extent_state *cached = NULL;
2511         int ret = 0;
2512         u64 start;
2513         u64 len;
2514         u64 lock_start;
2515         u64 lock_end;
2516         bool merge = false;
2517         int index;
2518
2519         if (prev && prev->root_id == backref->root_id &&
2520             prev->inum == backref->inum &&
2521             prev->file_pos + prev->num_bytes == backref->file_pos)
2522                 merge = true;
2523
2524         /* step 1: get root */
2525         key.objectid = backref->root_id;
2526         key.type = BTRFS_ROOT_ITEM_KEY;
2527         key.offset = (u64)-1;
2528
2529         index = srcu_read_lock(&fs_info->subvol_srcu);
2530
2531         root = btrfs_read_fs_root_no_name(fs_info, &key);
2532         if (IS_ERR(root)) {
2533                 srcu_read_unlock(&fs_info->subvol_srcu, index);
2534                 if (PTR_ERR(root) == -ENOENT)
2535                         return 0;
2536                 return PTR_ERR(root);
2537         }
2538
2539         if (btrfs_root_readonly(root)) {
2540                 srcu_read_unlock(&fs_info->subvol_srcu, index);
2541                 return 0;
2542         }
2543
2544         /* step 2: get inode */
2545         key.objectid = backref->inum;
2546         key.type = BTRFS_INODE_ITEM_KEY;
2547         key.offset = 0;
2548
2549         inode = btrfs_iget(fs_info->sb, &key, root, NULL);
2550         if (IS_ERR(inode)) {
2551                 srcu_read_unlock(&fs_info->subvol_srcu, index);
2552                 return 0;
2553         }
2554
2555         srcu_read_unlock(&fs_info->subvol_srcu, index);
2556
2557         /* step 3: relink backref */
2558         lock_start = backref->file_pos;
2559         lock_end = backref->file_pos + backref->num_bytes - 1;
2560         lock_extent_bits(&BTRFS_I(inode)->io_tree, lock_start, lock_end,
2561                          &cached);
2562
2563         ordered = btrfs_lookup_first_ordered_extent(inode, lock_end);
2564         if (ordered) {
2565                 btrfs_put_ordered_extent(ordered);
2566                 goto out_unlock;
2567         }
2568
2569         trans = btrfs_join_transaction(root);
2570         if (IS_ERR(trans)) {
2571                 ret = PTR_ERR(trans);
2572                 goto out_unlock;
2573         }
2574
2575         key.objectid = backref->inum;
2576         key.type = BTRFS_EXTENT_DATA_KEY;
2577         key.offset = backref->file_pos;
2578
2579         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2580         if (ret < 0) {
2581                 goto out_free_path;
2582         } else if (ret > 0) {
2583                 ret = 0;
2584                 goto out_free_path;
2585         }
2586
2587         extent = btrfs_item_ptr(path->nodes[0], path->slots[0],
2588                                 struct btrfs_file_extent_item);
2589
2590         if (btrfs_file_extent_generation(path->nodes[0], extent) !=
2591             backref->generation)
2592                 goto out_free_path;
2593
2594         btrfs_release_path(path);
2595
2596         start = backref->file_pos;
2597         if (backref->extent_offset < old->extent_offset + old->offset)
2598                 start += old->extent_offset + old->offset -
2599                          backref->extent_offset;
2600
2601         len = min(backref->extent_offset + backref->num_bytes,
2602                   old->extent_offset + old->offset + old->len);
2603         len -= max(backref->extent_offset, old->extent_offset + old->offset);
2604
2605         ret = btrfs_drop_extents(trans, root, inode, start,
2606                                  start + len, 1);
2607         if (ret)
2608                 goto out_free_path;
2609 again:
2610         key.objectid = btrfs_ino(BTRFS_I(inode));
2611         key.type = BTRFS_EXTENT_DATA_KEY;
2612         key.offset = start;
2613
2614         path->leave_spinning = 1;
2615         if (merge) {
2616                 struct btrfs_file_extent_item *fi;
2617                 u64 extent_len;
2618                 struct btrfs_key found_key;
2619
2620                 ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
2621                 if (ret < 0)
2622                         goto out_free_path;
2623
2624                 path->slots[0]--;
2625                 leaf = path->nodes[0];
2626                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
2627
2628                 fi = btrfs_item_ptr(leaf, path->slots[0],
2629                                     struct btrfs_file_extent_item);
2630                 extent_len = btrfs_file_extent_num_bytes(leaf, fi);
2631
2632                 if (extent_len + found_key.offset == start &&
2633                     relink_is_mergable(leaf, fi, new)) {
2634                         btrfs_set_file_extent_num_bytes(leaf, fi,
2635                                                         extent_len + len);
2636                         btrfs_mark_buffer_dirty(leaf);
2637                         inode_add_bytes(inode, len);
2638
2639                         ret = 1;
2640                         goto out_free_path;
2641                 } else {
2642                         merge = false;
2643                         btrfs_release_path(path);
2644                         goto again;
2645                 }
2646         }
2647
2648         ret = btrfs_insert_empty_item(trans, root, path, &key,
2649                                         sizeof(*extent));
2650         if (ret) {
2651                 btrfs_abort_transaction(trans, ret);
2652                 goto out_free_path;
2653         }
2654
2655         leaf = path->nodes[0];
2656         item = btrfs_item_ptr(leaf, path->slots[0],
2657                                 struct btrfs_file_extent_item);
2658         btrfs_set_file_extent_disk_bytenr(leaf, item, new->bytenr);
2659         btrfs_set_file_extent_disk_num_bytes(leaf, item, new->disk_len);
2660         btrfs_set_file_extent_offset(leaf, item, start - new->file_pos);
2661         btrfs_set_file_extent_num_bytes(leaf, item, len);
2662         btrfs_set_file_extent_ram_bytes(leaf, item, new->len);
2663         btrfs_set_file_extent_generation(leaf, item, trans->transid);
2664         btrfs_set_file_extent_type(leaf, item, BTRFS_FILE_EXTENT_REG);
2665         btrfs_set_file_extent_compression(leaf, item, new->compress_type);
2666         btrfs_set_file_extent_encryption(leaf, item, 0);
2667         btrfs_set_file_extent_other_encoding(leaf, item, 0);
2668
2669         btrfs_mark_buffer_dirty(leaf);
2670         inode_add_bytes(inode, len);
2671         btrfs_release_path(path);
2672
2673         ret = btrfs_inc_extent_ref(trans, fs_info, new->bytenr,
2674                         new->disk_len, 0,
2675                         backref->root_id, backref->inum,
2676                         new->file_pos); /* start - extent_offset */
2677         if (ret) {
2678                 btrfs_abort_transaction(trans, ret);
2679                 goto out_free_path;
2680         }
2681
2682         ret = 1;
2683 out_free_path:
2684         btrfs_release_path(path);
2685         path->leave_spinning = 0;
2686         btrfs_end_transaction(trans);
2687 out_unlock:
2688         unlock_extent_cached(&BTRFS_I(inode)->io_tree, lock_start, lock_end,
2689                              &cached, GFP_NOFS);
2690         iput(inode);
2691         return ret;
2692 }
2693
2694 static void free_sa_defrag_extent(struct new_sa_defrag_extent *new)
2695 {
2696         struct old_sa_defrag_extent *old, *tmp;
2697
2698         if (!new)
2699                 return;
2700
2701         list_for_each_entry_safe(old, tmp, &new->head, list) {
2702                 kfree(old);
2703         }
2704         kfree(new);
2705 }
2706
2707 static void relink_file_extents(struct new_sa_defrag_extent *new)
2708 {
2709         struct btrfs_fs_info *fs_info = btrfs_sb(new->inode->i_sb);
2710         struct btrfs_path *path;
2711         struct sa_defrag_extent_backref *backref;
2712         struct sa_defrag_extent_backref *prev = NULL;
2713         struct inode *inode;
2714         struct btrfs_root *root;
2715         struct rb_node *node;
2716         int ret;
2717
2718         inode = new->inode;
2719         root = BTRFS_I(inode)->root;
2720
2721         path = btrfs_alloc_path();
2722         if (!path)
2723                 return;
2724
2725         if (!record_extent_backrefs(path, new)) {
2726                 btrfs_free_path(path);
2727                 goto out;
2728         }
2729         btrfs_release_path(path);
2730
2731         while (1) {
2732                 node = rb_first(&new->root);
2733                 if (!node)
2734                         break;
2735                 rb_erase(node, &new->root);
2736
2737                 backref = rb_entry(node, struct sa_defrag_extent_backref, node);
2738
2739                 ret = relink_extent_backref(path, prev, backref);
2740                 WARN_ON(ret < 0);
2741
2742                 kfree(prev);
2743
2744                 if (ret == 1)
2745                         prev = backref;
2746                 else
2747                         prev = NULL;
2748                 cond_resched();
2749         }
2750         kfree(prev);
2751
2752         btrfs_free_path(path);
2753 out:
2754         free_sa_defrag_extent(new);
2755
2756         atomic_dec(&fs_info->defrag_running);
2757         wake_up(&fs_info->transaction_wait);
2758 }
2759
2760 static struct new_sa_defrag_extent *
2761 record_old_file_extents(struct inode *inode,
2762                         struct btrfs_ordered_extent *ordered)
2763 {
2764         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2765         struct btrfs_root *root = BTRFS_I(inode)->root;
2766         struct btrfs_path *path;
2767         struct btrfs_key key;
2768         struct old_sa_defrag_extent *old;
2769         struct new_sa_defrag_extent *new;
2770         int ret;
2771
2772         new = kmalloc(sizeof(*new), GFP_NOFS);
2773         if (!new)
2774                 return NULL;
2775
2776         new->inode = inode;
2777         new->file_pos = ordered->file_offset;
2778         new->len = ordered->len;
2779         new->bytenr = ordered->start;
2780         new->disk_len = ordered->disk_len;
2781         new->compress_type = ordered->compress_type;
2782         new->root = RB_ROOT;
2783         INIT_LIST_HEAD(&new->head);
2784
2785         path = btrfs_alloc_path();
2786         if (!path)
2787                 goto out_kfree;
2788
2789         key.objectid = btrfs_ino(BTRFS_I(inode));
2790         key.type = BTRFS_EXTENT_DATA_KEY;
2791         key.offset = new->file_pos;
2792
2793         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2794         if (ret < 0)
2795                 goto out_free_path;
2796         if (ret > 0 && path->slots[0] > 0)
2797                 path->slots[0]--;
2798
2799         /* find out all the old extents for the file range */
2800         while (1) {
2801                 struct btrfs_file_extent_item *extent;
2802                 struct extent_buffer *l;
2803                 int slot;
2804                 u64 num_bytes;
2805                 u64 offset;
2806                 u64 end;
2807                 u64 disk_bytenr;
2808                 u64 extent_offset;
2809
2810                 l = path->nodes[0];
2811                 slot = path->slots[0];
2812
2813                 if (slot >= btrfs_header_nritems(l)) {
2814                         ret = btrfs_next_leaf(root, path);
2815                         if (ret < 0)
2816                                 goto out_free_path;
2817                         else if (ret > 0)
2818                                 break;
2819                         continue;
2820                 }
2821
2822                 btrfs_item_key_to_cpu(l, &key, slot);
2823
2824                 if (key.objectid != btrfs_ino(BTRFS_I(inode)))
2825                         break;
2826                 if (key.type != BTRFS_EXTENT_DATA_KEY)
2827                         break;
2828                 if (key.offset >= new->file_pos + new->len)
2829                         break;
2830
2831                 extent = btrfs_item_ptr(l, slot, struct btrfs_file_extent_item);
2832
2833                 num_bytes = btrfs_file_extent_num_bytes(l, extent);
2834                 if (key.offset + num_bytes < new->file_pos)
2835                         goto next;
2836
2837                 disk_bytenr = btrfs_file_extent_disk_bytenr(l, extent);
2838                 if (!disk_bytenr)
2839                         goto next;
2840
2841                 extent_offset = btrfs_file_extent_offset(l, extent);
2842
2843                 old = kmalloc(sizeof(*old), GFP_NOFS);
2844                 if (!old)
2845                         goto out_free_path;
2846
2847                 offset = max(new->file_pos, key.offset);
2848                 end = min(new->file_pos + new->len, key.offset + num_bytes);
2849
2850                 old->bytenr = disk_bytenr;
2851                 old->extent_offset = extent_offset;
2852                 old->offset = offset - key.offset;
2853                 old->len = end - offset;
2854                 old->new = new;
2855                 old->count = 0;
2856                 list_add_tail(&old->list, &new->head);
2857 next:
2858                 path->slots[0]++;
2859                 cond_resched();
2860         }
2861
2862         btrfs_free_path(path);
2863         atomic_inc(&fs_info->defrag_running);
2864
2865         return new;
2866
2867 out_free_path:
2868         btrfs_free_path(path);
2869 out_kfree:
2870         free_sa_defrag_extent(new);
2871         return NULL;
2872 }
2873
2874 static void btrfs_release_delalloc_bytes(struct btrfs_fs_info *fs_info,
2875                                          u64 start, u64 len)
2876 {
2877         struct btrfs_block_group_cache *cache;
2878
2879         cache = btrfs_lookup_block_group(fs_info, start);
2880         ASSERT(cache);
2881
2882         spin_lock(&cache->lock);
2883         cache->delalloc_bytes -= len;
2884         spin_unlock(&cache->lock);
2885
2886         btrfs_put_block_group(cache);
2887 }
2888
2889 /* as ordered data IO finishes, this gets called so we can finish
2890  * an ordered extent if the range of bytes in the file it covers are
2891  * fully written.
2892  */
2893 static int btrfs_finish_ordered_io(struct btrfs_ordered_extent *ordered_extent)
2894 {
2895         struct inode *inode = ordered_extent->inode;
2896         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2897         struct btrfs_root *root = BTRFS_I(inode)->root;
2898         struct btrfs_trans_handle *trans = NULL;
2899         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
2900         struct extent_state *cached_state = NULL;
2901         struct new_sa_defrag_extent *new = NULL;
2902         int compress_type = 0;
2903         int ret = 0;
2904         u64 logical_len = ordered_extent->len;
2905         bool nolock;
2906         bool truncated = false;
2907         bool range_locked = false;
2908         bool clear_new_delalloc_bytes = false;
2909
2910         if (!test_bit(BTRFS_ORDERED_NOCOW, &ordered_extent->flags) &&
2911             !test_bit(BTRFS_ORDERED_PREALLOC, &ordered_extent->flags) &&
2912             !test_bit(BTRFS_ORDERED_DIRECT, &ordered_extent->flags))
2913                 clear_new_delalloc_bytes = true;
2914
2915         nolock = btrfs_is_free_space_inode(BTRFS_I(inode));
2916
2917         if (test_bit(BTRFS_ORDERED_IOERR, &ordered_extent->flags)) {
2918                 ret = -EIO;
2919                 goto out;
2920         }
2921
2922         btrfs_free_io_failure_record(BTRFS_I(inode),
2923                         ordered_extent->file_offset,
2924                         ordered_extent->file_offset +
2925                         ordered_extent->len - 1);
2926
2927         if (test_bit(BTRFS_ORDERED_TRUNCATED, &ordered_extent->flags)) {
2928                 truncated = true;
2929                 logical_len = ordered_extent->truncated_len;
2930                 /* Truncated the entire extent, don't bother adding */
2931                 if (!logical_len)
2932                         goto out;
2933         }
2934
2935         if (test_bit(BTRFS_ORDERED_NOCOW, &ordered_extent->flags)) {
2936                 BUG_ON(!list_empty(&ordered_extent->list)); /* Logic error */
2937
2938                 /*
2939                  * For mwrite(mmap + memset to write) case, we still reserve
2940                  * space for NOCOW range.
2941                  * As NOCOW won't cause a new delayed ref, just free the space
2942                  */
2943                 btrfs_qgroup_free_data(inode, NULL, ordered_extent->file_offset,
2944                                        ordered_extent->len);
2945                 btrfs_ordered_update_i_size(inode, 0, ordered_extent);
2946                 if (nolock)
2947                         trans = btrfs_join_transaction_nolock(root);
2948                 else
2949                         trans = btrfs_join_transaction(root);
2950                 if (IS_ERR(trans)) {
2951                         ret = PTR_ERR(trans);
2952                         trans = NULL;
2953                         goto out;
2954                 }
2955                 trans->block_rsv = &fs_info->delalloc_block_rsv;
2956                 ret = btrfs_update_inode_fallback(trans, root, inode);
2957                 if (ret) /* -ENOMEM or corruption */
2958                         btrfs_abort_transaction(trans, ret);
2959                 goto out;
2960         }
2961
2962         range_locked = true;
2963         lock_extent_bits(io_tree, ordered_extent->file_offset,
2964                          ordered_extent->file_offset + ordered_extent->len - 1,
2965                          &cached_state);
2966
2967         ret = test_range_bit(io_tree, ordered_extent->file_offset,
2968                         ordered_extent->file_offset + ordered_extent->len - 1,
2969                         EXTENT_DEFRAG, 0, cached_state);
2970         if (ret) {
2971                 u64 last_snapshot = btrfs_root_last_snapshot(&root->root_item);
2972                 if (0 && last_snapshot >= BTRFS_I(inode)->generation)
2973                         /* the inode is shared */
2974                         new = record_old_file_extents(inode, ordered_extent);
2975
2976                 clear_extent_bit(io_tree, ordered_extent->file_offset,
2977                         ordered_extent->file_offset + ordered_extent->len - 1,
2978                         EXTENT_DEFRAG, 0, 0, &cached_state, GFP_NOFS);
2979         }
2980
2981         if (nolock)
2982                 trans = btrfs_join_transaction_nolock(root);
2983         else
2984                 trans = btrfs_join_transaction(root);
2985         if (IS_ERR(trans)) {
2986                 ret = PTR_ERR(trans);
2987                 trans = NULL;
2988                 goto out;
2989         }
2990
2991         trans->block_rsv = &fs_info->delalloc_block_rsv;
2992
2993         if (test_bit(BTRFS_ORDERED_COMPRESSED, &ordered_extent->flags))
2994                 compress_type = ordered_extent->compress_type;
2995         if (test_bit(BTRFS_ORDERED_PREALLOC, &ordered_extent->flags)) {
2996                 BUG_ON(compress_type);
2997                 ret = btrfs_mark_extent_written(trans, BTRFS_I(inode),
2998                                                 ordered_extent->file_offset,
2999                                                 ordered_extent->file_offset +
3000                                                 logical_len);
3001         } else {
3002                 BUG_ON(root == fs_info->tree_root);
3003                 ret = insert_reserved_file_extent(trans, inode,
3004                                                 ordered_extent->file_offset,
3005                                                 ordered_extent->start,
3006                                                 ordered_extent->disk_len,
3007                                                 logical_len, logical_len,
3008                                                 compress_type, 0, 0,
3009                                                 BTRFS_FILE_EXTENT_REG);
3010                 if (!ret)
3011                         btrfs_release_delalloc_bytes(fs_info,
3012                                                      ordered_extent->start,
3013                                                      ordered_extent->disk_len);
3014         }
3015         unpin_extent_cache(&BTRFS_I(inode)->extent_tree,
3016                            ordered_extent->file_offset, ordered_extent->len,
3017                            trans->transid);
3018         if (ret < 0) {
3019                 btrfs_abort_transaction(trans, ret);
3020                 goto out;
3021         }
3022
3023         add_pending_csums(trans, inode, &ordered_extent->list);
3024
3025         btrfs_ordered_update_i_size(inode, 0, ordered_extent);
3026         ret = btrfs_update_inode_fallback(trans, root, inode);
3027         if (ret) { /* -ENOMEM or corruption */
3028                 btrfs_abort_transaction(trans, ret);
3029                 goto out;
3030         }
3031         ret = 0;
3032 out:
3033         if (range_locked || clear_new_delalloc_bytes) {
3034                 unsigned int clear_bits = 0;
3035
3036                 if (range_locked)
3037                         clear_bits |= EXTENT_LOCKED;
3038                 if (clear_new_delalloc_bytes)
3039                         clear_bits |= EXTENT_DELALLOC_NEW;
3040                 clear_extent_bit(&BTRFS_I(inode)->io_tree,
3041                                  ordered_extent->file_offset,
3042                                  ordered_extent->file_offset +
3043                                  ordered_extent->len - 1,
3044                                  clear_bits,
3045                                  (clear_bits & EXTENT_LOCKED) ? 1 : 0,
3046                                  0, &cached_state, GFP_NOFS);
3047         }
3048
3049         if (root != fs_info->tree_root)
3050                 btrfs_delalloc_release_metadata(BTRFS_I(inode),
3051                                 ordered_extent->len);
3052         if (trans)
3053                 btrfs_end_transaction(trans);
3054
3055         if (ret || truncated) {
3056                 u64 start, end;
3057
3058                 if (truncated)
3059                         start = ordered_extent->file_offset + logical_len;
3060                 else
3061                         start = ordered_extent->file_offset;
3062                 end = ordered_extent->file_offset + ordered_extent->len - 1;
3063                 clear_extent_uptodate(io_tree, start, end, NULL, GFP_NOFS);
3064
3065                 /* Drop the cache for the part of the extent we didn't write. */
3066                 btrfs_drop_extent_cache(BTRFS_I(inode), start, end, 0);
3067
3068                 /*
3069                  * If the ordered extent had an IOERR or something else went
3070                  * wrong we need to return the space for this ordered extent
3071                  * back to the allocator.  We only free the extent in the
3072                  * truncated case if we didn't write out the extent at all.
3073                  */
3074                 if ((ret || !logical_len) &&
3075                     !test_bit(BTRFS_ORDERED_NOCOW, &ordered_extent->flags) &&
3076                     !test_bit(BTRFS_ORDERED_PREALLOC, &ordered_extent->flags))
3077                         btrfs_free_reserved_extent(fs_info,
3078                                                    ordered_extent->start,
3079                                                    ordered_extent->disk_len, 1);
3080         }
3081
3082
3083         /*
3084          * This needs to be done to make sure anybody waiting knows we are done
3085          * updating everything for this ordered extent.
3086          */
3087         btrfs_remove_ordered_extent(inode, ordered_extent);
3088
3089         /* for snapshot-aware defrag */
3090         if (new) {
3091                 if (ret) {
3092                         free_sa_defrag_extent(new);
3093                         atomic_dec(&fs_info->defrag_running);
3094                 } else {
3095                         relink_file_extents(new);
3096                 }
3097         }
3098
3099         /* once for us */
3100         btrfs_put_ordered_extent(ordered_extent);
3101         /* once for the tree */
3102         btrfs_put_ordered_extent(ordered_extent);
3103
3104         return ret;
3105 }
3106
3107 static void finish_ordered_fn(struct btrfs_work *work)
3108 {
3109         struct btrfs_ordered_extent *ordered_extent;
3110         ordered_extent = container_of(work, struct btrfs_ordered_extent, work);
3111         btrfs_finish_ordered_io(ordered_extent);
3112 }
3113
3114 static void btrfs_writepage_end_io_hook(struct page *page, u64 start, u64 end,
3115                                 struct extent_state *state, int uptodate)
3116 {
3117         struct inode *inode = page->mapping->host;
3118         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
3119         struct btrfs_ordered_extent *ordered_extent = NULL;
3120         struct btrfs_workqueue *wq;
3121         btrfs_work_func_t func;
3122
3123         trace_btrfs_writepage_end_io_hook(page, start, end, uptodate);
3124
3125         ClearPagePrivate2(page);
3126         if (!btrfs_dec_test_ordered_pending(inode, &ordered_extent, start,
3127                                             end - start + 1, uptodate))
3128                 return;
3129
3130         if (btrfs_is_free_space_inode(BTRFS_I(inode))) {
3131                 wq = fs_info->endio_freespace_worker;
3132                 func = btrfs_freespace_write_helper;
3133         } else {
3134                 wq = fs_info->endio_write_workers;
3135                 func = btrfs_endio_write_helper;
3136         }
3137
3138         btrfs_init_work(&ordered_extent->work, func, finish_ordered_fn, NULL,
3139                         NULL);
3140         btrfs_queue_work(wq, &ordered_extent->work);
3141 }
3142
3143 static int __readpage_endio_check(struct inode *inode,
3144                                   struct btrfs_io_bio *io_bio,
3145                                   int icsum, struct page *page,
3146                                   int pgoff, u64 start, size_t len)
3147 {
3148         char *kaddr;
3149         u32 csum_expected;
3150         u32 csum = ~(u32)0;
3151
3152         csum_expected = *(((u32 *)io_bio->csum) + icsum);
3153
3154         kaddr = kmap_atomic(page);
3155         csum = btrfs_csum_data(kaddr + pgoff, csum,  len);
3156         btrfs_csum_final(csum, (u8 *)&csum);
3157         if (csum != csum_expected)
3158                 goto zeroit;
3159
3160         kunmap_atomic(kaddr);
3161         return 0;
3162 zeroit:
3163         btrfs_print_data_csum_error(BTRFS_I(inode), start, csum, csum_expected,
3164                                     io_bio->mirror_num);
3165         memset(kaddr + pgoff, 1, len);
3166         flush_dcache_page(page);
3167         kunmap_atomic(kaddr);
3168         return -EIO;
3169 }
3170
3171 /*
3172  * when reads are done, we need to check csums to verify the data is correct
3173  * if there's a match, we allow the bio to finish.  If not, the code in
3174  * extent_io.c will try to find good copies for us.
3175  */
3176 static int btrfs_readpage_end_io_hook(struct btrfs_io_bio *io_bio,
3177                                       u64 phy_offset, struct page *page,
3178                                       u64 start, u64 end, int mirror)
3179 {
3180         size_t offset = start - page_offset(page);
3181         struct inode *inode = page->mapping->host;
3182         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
3183         struct btrfs_root *root = BTRFS_I(inode)->root;
3184
3185         if (PageChecked(page)) {
3186                 ClearPageChecked(page);
3187                 return 0;
3188         }
3189
3190         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)
3191                 return 0;
3192
3193         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID &&
3194             test_range_bit(io_tree, start, end, EXTENT_NODATASUM, 1, NULL)) {
3195                 clear_extent_bits(io_tree, start, end, EXTENT_NODATASUM);
3196                 return 0;
3197         }
3198
3199         phy_offset >>= inode->i_sb->s_blocksize_bits;
3200         return __readpage_endio_check(inode, io_bio, phy_offset, page, offset,
3201                                       start, (size_t)(end - start + 1));
3202 }
3203
3204 void btrfs_add_delayed_iput(struct inode *inode)
3205 {
3206         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
3207         struct btrfs_inode *binode = BTRFS_I(inode);
3208
3209         if (atomic_add_unless(&inode->i_count, -1, 1))
3210                 return;
3211
3212         spin_lock(&fs_info->delayed_iput_lock);
3213         if (binode->delayed_iput_count == 0) {
3214                 ASSERT(list_empty(&binode->delayed_iput));
3215                 list_add_tail(&binode->delayed_iput, &fs_info->delayed_iputs);
3216         } else {
3217                 binode->delayed_iput_count++;
3218         }
3219         spin_unlock(&fs_info->delayed_iput_lock);
3220 }
3221
3222 void btrfs_run_delayed_iputs(struct btrfs_fs_info *fs_info)
3223 {
3224
3225         spin_lock(&fs_info->delayed_iput_lock);
3226         while (!list_empty(&fs_info->delayed_iputs)) {
3227                 struct btrfs_inode *inode;
3228
3229                 inode = list_first_entry(&fs_info->delayed_iputs,
3230                                 struct btrfs_inode, delayed_iput);
3231                 if (inode->delayed_iput_count) {
3232                         inode->delayed_iput_count--;
3233                         list_move_tail(&inode->delayed_iput,
3234                                         &fs_info->delayed_iputs);
3235                 } else {
3236                         list_del_init(&inode->delayed_iput);
3237                 }
3238                 spin_unlock(&fs_info->delayed_iput_lock);
3239                 iput(&inode->vfs_inode);
3240                 spin_lock(&fs_info->delayed_iput_lock);
3241         }
3242         spin_unlock(&fs_info->delayed_iput_lock);
3243 }
3244
3245 /*
3246  * This is called in transaction commit time. If there are no orphan
3247  * files in the subvolume, it removes orphan item and frees block_rsv
3248  * structure.
3249  */
3250 void btrfs_orphan_commit_root(struct btrfs_trans_handle *trans,
3251                               struct btrfs_root *root)
3252 {
3253         struct btrfs_fs_info *fs_info = root->fs_info;
3254         struct btrfs_block_rsv *block_rsv;
3255         int ret;
3256
3257         if (atomic_read(&root->orphan_inodes) ||
3258             root->orphan_cleanup_state != ORPHAN_CLEANUP_DONE)
3259                 return;
3260
3261         spin_lock(&root->orphan_lock);
3262         if (atomic_read(&root->orphan_inodes)) {
3263                 spin_unlock(&root->orphan_lock);
3264                 return;
3265         }
3266
3267         if (root->orphan_cleanup_state != ORPHAN_CLEANUP_DONE) {
3268                 spin_unlock(&root->orphan_lock);
3269                 return;
3270         }
3271
3272         block_rsv = root->orphan_block_rsv;
3273         root->orphan_block_rsv = NULL;
3274         spin_unlock(&root->orphan_lock);
3275
3276         if (test_bit(BTRFS_ROOT_ORPHAN_ITEM_INSERTED, &root->state) &&
3277             btrfs_root_refs(&root->root_item) > 0) {
3278                 ret = btrfs_del_orphan_item(trans, fs_info->tree_root,
3279                                             root->root_key.objectid);
3280                 if (ret)
3281                         btrfs_abort_transaction(trans, ret);
3282                 else
3283                         clear_bit(BTRFS_ROOT_ORPHAN_ITEM_INSERTED,
3284                                   &root->state);
3285         }
3286
3287         if (block_rsv) {
3288                 WARN_ON(block_rsv->size > 0);
3289                 btrfs_free_block_rsv(fs_info, block_rsv);
3290         }
3291 }
3292
3293 /*
3294  * This creates an orphan entry for the given inode in case something goes
3295  * wrong in the middle of an unlink/truncate.
3296  *
3297  * NOTE: caller of this function should reserve 5 units of metadata for
3298  *       this function.
3299  */
3300 int btrfs_orphan_add(struct btrfs_trans_handle *trans,
3301                 struct btrfs_inode *inode)
3302 {
3303         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
3304         struct btrfs_root *root = inode->root;
3305         struct btrfs_block_rsv *block_rsv = NULL;
3306         int reserve = 0;
3307         int insert = 0;
3308         int ret;
3309
3310         if (!root->orphan_block_rsv) {
3311                 block_rsv = btrfs_alloc_block_rsv(fs_info,
3312                                                   BTRFS_BLOCK_RSV_TEMP);
3313                 if (!block_rsv)
3314                         return -ENOMEM;
3315         }
3316
3317         spin_lock(&root->orphan_lock);
3318         if (!root->orphan_block_rsv) {
3319                 root->orphan_block_rsv = block_rsv;
3320         } else if (block_rsv) {
3321                 btrfs_free_block_rsv(fs_info, block_rsv);
3322                 block_rsv = NULL;
3323         }
3324
3325         if (!test_and_set_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
3326                               &inode->runtime_flags)) {
3327 #if 0
3328                 /*
3329                  * For proper ENOSPC handling, we should do orphan
3330                  * cleanup when mounting. But this introduces backward
3331                  * compatibility issue.
3332                  */
3333                 if (!xchg(&root->orphan_item_inserted, 1))
3334                         insert = 2;
3335                 else
3336                         insert = 1;
3337 #endif
3338                 insert = 1;
3339                 atomic_inc(&root->orphan_inodes);
3340         }
3341
3342         if (!test_and_set_bit(BTRFS_INODE_ORPHAN_META_RESERVED,
3343                               &inode->runtime_flags))
3344                 reserve = 1;
3345         spin_unlock(&root->orphan_lock);
3346
3347         /* grab metadata reservation from transaction handle */
3348         if (reserve) {
3349                 ret = btrfs_orphan_reserve_metadata(trans, inode);
3350                 ASSERT(!ret);
3351                 if (ret) {
3352                         atomic_dec(&root->orphan_inodes);
3353                         clear_bit(BTRFS_INODE_ORPHAN_META_RESERVED,
3354                                   &inode->runtime_flags);
3355                         if (insert)
3356                                 clear_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
3357                                           &inode->runtime_flags);
3358                         return ret;
3359                 }
3360         }
3361
3362         /* insert an orphan item to track this unlinked/truncated file */
3363         if (insert >= 1) {
3364                 ret = btrfs_insert_orphan_item(trans, root, btrfs_ino(inode));
3365                 if (ret) {
3366                         atomic_dec(&root->orphan_inodes);
3367                         if (reserve) {
3368                                 clear_bit(BTRFS_INODE_ORPHAN_META_RESERVED,
3369                                           &inode->runtime_flags);
3370                                 btrfs_orphan_release_metadata(inode);
3371                         }
3372                         if (ret != -EEXIST) {
3373                                 clear_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
3374                                           &inode->runtime_flags);
3375                                 btrfs_abort_transaction(trans, ret);
3376                                 return ret;
3377                         }
3378                 }
3379                 ret = 0;
3380         }
3381
3382         /* insert an orphan item to track subvolume contains orphan files */
3383         if (insert >= 2) {
3384                 ret = btrfs_insert_orphan_item(trans, fs_info->tree_root,
3385                                                root->root_key.objectid);
3386                 if (ret && ret != -EEXIST) {
3387                         btrfs_abort_transaction(trans, ret);
3388                         return ret;
3389                 }
3390         }
3391         return 0;
3392 }
3393
3394 /*
3395  * We have done the truncate/delete so we can go ahead and remove the orphan
3396  * item for this particular inode.
3397  */
3398 static int btrfs_orphan_del(struct btrfs_trans_handle *trans,
3399                             struct btrfs_inode *inode)
3400 {
3401         struct btrfs_root *root = inode->root;
3402         int delete_item = 0;
3403         int release_rsv = 0;
3404         int ret = 0;
3405
3406         spin_lock(&root->orphan_lock);
3407         if (test_and_clear_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
3408                                &inode->runtime_flags))
3409                 delete_item = 1;
3410
3411         if (test_and_clear_bit(BTRFS_INODE_ORPHAN_META_RESERVED,
3412                                &inode->runtime_flags))
3413                 release_rsv = 1;
3414         spin_unlock(&root->orphan_lock);
3415
3416         if (delete_item) {
3417                 atomic_dec(&root->orphan_inodes);
3418                 if (trans)
3419                         ret = btrfs_del_orphan_item(trans, root,
3420                                                     btrfs_ino(inode));
3421         }
3422
3423         if (release_rsv)
3424                 btrfs_orphan_release_metadata(inode);
3425
3426         return ret;
3427 }
3428
3429 /*
3430  * this cleans up any orphans that may be left on the list from the last use
3431  * of this root.
3432  */
3433 int btrfs_orphan_cleanup(struct btrfs_root *root)
3434 {
3435         struct btrfs_fs_info *fs_info = root->fs_info;
3436         struct btrfs_path *path;
3437         struct extent_buffer *leaf;
3438         struct btrfs_key key, found_key;
3439         struct btrfs_trans_handle *trans;
3440         struct inode *inode;
3441         u64 last_objectid = 0;
3442         int ret = 0, nr_unlink = 0, nr_truncate = 0;
3443
3444         if (cmpxchg(&root->orphan_cleanup_state, 0, ORPHAN_CLEANUP_STARTED))
3445                 return 0;
3446
3447         path = btrfs_alloc_path();
3448         if (!path) {
3449                 ret = -ENOMEM;
3450                 goto out;
3451         }
3452         path->reada = READA_BACK;
3453
3454         key.objectid = BTRFS_ORPHAN_OBJECTID;
3455         key.type = BTRFS_ORPHAN_ITEM_KEY;
3456         key.offset = (u64)-1;
3457
3458         while (1) {
3459                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3460                 if (ret < 0)
3461                         goto out;
3462
3463                 /*
3464                  * if ret == 0 means we found what we were searching for, which
3465                  * is weird, but possible, so only screw with path if we didn't
3466                  * find the key and see if we have stuff that matches
3467                  */
3468                 if (ret > 0) {
3469                         ret = 0;
3470                         if (path->slots[0] == 0)
3471                                 break;
3472                         path->slots[0]--;
3473                 }
3474
3475                 /* pull out the item */
3476                 leaf = path->nodes[0];
3477                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
3478
3479                 /* make sure the item matches what we want */
3480                 if (found_key.objectid != BTRFS_ORPHAN_OBJECTID)
3481                         break;
3482                 if (found_key.type != BTRFS_ORPHAN_ITEM_KEY)
3483                         break;
3484
3485                 /* release the path since we're done with it */
3486                 btrfs_release_path(path);
3487
3488                 /*
3489                  * this is where we are basically btrfs_lookup, without the
3490                  * crossing root thing.  we store the inode number in the
3491                  * offset of the orphan item.
3492                  */
3493
3494                 if (found_key.offset == last_objectid) {
3495                         btrfs_err(fs_info,
3496                                   "Error removing orphan entry, stopping orphan cleanup");
3497                         ret = -EINVAL;
3498                         goto out;
3499                 }
3500
3501                 last_objectid = found_key.offset;
3502
3503                 found_key.objectid = found_key.offset;
3504                 found_key.type = BTRFS_INODE_ITEM_KEY;
3505                 found_key.offset = 0;
3506                 inode = btrfs_iget(fs_info->sb, &found_key, root, NULL);
3507                 ret = PTR_ERR_OR_ZERO(inode);
3508                 if (ret && ret != -ENOENT)
3509                         goto out;
3510
3511                 if (ret == -ENOENT && root == fs_info->tree_root) {
3512                         struct btrfs_root *dead_root;
3513                         struct btrfs_fs_info *fs_info = root->fs_info;
3514                         int is_dead_root = 0;
3515
3516                         /*
3517                          * this is an orphan in the tree root. Currently these
3518                          * could come from 2 sources:
3519                          *  a) a snapshot deletion in progress
3520                          *  b) a free space cache inode
3521                          * We need to distinguish those two, as the snapshot
3522                          * orphan must not get deleted.
3523                          * find_dead_roots already ran before us, so if this
3524                          * is a snapshot deletion, we should find the root
3525                          * in the dead_roots list
3526                          */
3527                         spin_lock(&fs_info->trans_lock);
3528                         list_for_each_entry(dead_root, &fs_info->dead_roots,
3529                                             root_list) {
3530                                 if (dead_root->root_key.objectid ==
3531                                     found_key.objectid) {
3532                                         is_dead_root = 1;
3533                                         break;
3534                                 }
3535                         }
3536                         spin_unlock(&fs_info->trans_lock);
3537                         if (is_dead_root) {
3538                                 /* prevent this orphan from being found again */
3539                                 key.offset = found_key.objectid - 1;
3540                                 continue;
3541                         }
3542                 }
3543                 /*
3544                  * Inode is already gone but the orphan item is still there,
3545                  * kill the orphan item.
3546                  */
3547                 if (ret == -ENOENT) {
3548                         trans = btrfs_start_transaction(root, 1);
3549                         if (IS_ERR(trans)) {
3550                                 ret = PTR_ERR(trans);
3551                                 goto out;
3552                         }
3553                         btrfs_debug(fs_info, "auto deleting %Lu",
3554                                     found_key.objectid);
3555                         ret = btrfs_del_orphan_item(trans, root,
3556                                                     found_key.objectid);
3557                         btrfs_end_transaction(trans);
3558                         if (ret)
3559                                 goto out;
3560                         continue;
3561                 }
3562
3563                 /*
3564                  * add this inode to the orphan list so btrfs_orphan_del does
3565                  * the proper thing when we hit it
3566                  */
3567                 set_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
3568                         &BTRFS_I(inode)->runtime_flags);
3569                 atomic_inc(&root->orphan_inodes);
3570
3571                 /* if we have links, this was a truncate, lets do that */
3572                 if (inode->i_nlink) {
3573                         if (WARN_ON(!S_ISREG(inode->i_mode))) {
3574                                 iput(inode);
3575                                 continue;
3576                         }
3577                         nr_truncate++;
3578
3579                         /* 1 for the orphan item deletion. */
3580                         trans = btrfs_start_transaction(root, 1);
3581                         if (IS_ERR(trans)) {
3582                                 iput(inode);
3583                                 ret = PTR_ERR(trans);
3584                                 goto out;
3585                         }
3586                         ret = btrfs_orphan_add(trans, BTRFS_I(inode));
3587                         btrfs_end_transaction(trans);
3588                         if (ret) {
3589                                 iput(inode);
3590                                 goto out;
3591                         }
3592
3593                         ret = btrfs_truncate(inode);
3594                         if (ret)
3595                                 btrfs_orphan_del(NULL, BTRFS_I(inode));
3596                 } else {
3597                         nr_unlink++;
3598                 }
3599
3600                 /* this will do delete_inode and everything for us */
3601                 iput(inode);
3602                 if (ret)
3603                         goto out;
3604         }
3605         /* release the path since we're done with it */
3606         btrfs_release_path(path);
3607
3608         root->orphan_cleanup_state = ORPHAN_CLEANUP_DONE;
3609
3610         if (root->orphan_block_rsv)
3611                 btrfs_block_rsv_release(fs_info, root->orphan_block_rsv,
3612                                         (u64)-1);
3613
3614         if (root->orphan_block_rsv ||
3615             test_bit(BTRFS_ROOT_ORPHAN_ITEM_INSERTED, &root->state)) {
3616                 trans = btrfs_join_transaction(root);
3617                 if (!IS_ERR(trans))
3618                         btrfs_end_transaction(trans);
3619         }
3620
3621         if (nr_unlink)
3622                 btrfs_debug(fs_info, "unlinked %d orphans", nr_unlink);
3623         if (nr_truncate)
3624                 btrfs_debug(fs_info, "truncated %d orphans", nr_truncate);
3625
3626 out:
3627         if (ret)
3628                 btrfs_err(fs_info, "could not do orphan cleanup %d", ret);
3629         btrfs_free_path(path);
3630         return ret;
3631 }
3632
3633 /*
3634  * very simple check to peek ahead in the leaf looking for xattrs.  If we
3635  * don't find any xattrs, we know there can't be any acls.
3636  *
3637  * slot is the slot the inode is in, objectid is the objectid of the inode
3638  */
3639 static noinline int acls_after_inode_item(struct extent_buffer *leaf,
3640                                           int slot, u64 objectid,
3641                                           int *first_xattr_slot)
3642 {
3643         u32 nritems = btrfs_header_nritems(leaf);
3644         struct btrfs_key found_key;
3645         static u64 xattr_access = 0;
3646         static u64 xattr_default = 0;
3647         int scanned = 0;
3648
3649         if (!xattr_access) {
3650                 xattr_access = btrfs_name_hash(XATTR_NAME_POSIX_ACL_ACCESS,
3651                                         strlen(XATTR_NAME_POSIX_ACL_ACCESS));
3652                 xattr_default = btrfs_name_hash(XATTR_NAME_POSIX_ACL_DEFAULT,
3653                                         strlen(XATTR_NAME_POSIX_ACL_DEFAULT));
3654         }
3655
3656         slot++;
3657         *first_xattr_slot = -1;
3658         while (slot < nritems) {
3659                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
3660
3661                 /* we found a different objectid, there must not be acls */
3662                 if (found_key.objectid != objectid)
3663                         return 0;
3664
3665                 /* we found an xattr, assume we've got an acl */
3666                 if (found_key.type == BTRFS_XATTR_ITEM_KEY) {
3667                         if (*first_xattr_slot == -1)
3668                                 *first_xattr_slot = slot;
3669                         if (found_key.offset == xattr_access ||
3670                             found_key.offset == xattr_default)
3671                                 return 1;
3672                 }
3673
3674                 /*
3675                  * we found a key greater than an xattr key, there can't
3676                  * be any acls later on
3677                  */
3678                 if (found_key.type > BTRFS_XATTR_ITEM_KEY)
3679                         return 0;
3680
3681                 slot++;
3682                 scanned++;
3683
3684                 /*
3685                  * it goes inode, inode backrefs, xattrs, extents,
3686                  * so if there are a ton of hard links to an inode there can
3687                  * be a lot of backrefs.  Don't waste time searching too hard,
3688                  * this is just an optimization
3689                  */
3690                 if (scanned >= 8)
3691                         break;
3692         }
3693         /* we hit the end of the leaf before we found an xattr or
3694          * something larger than an xattr.  We have to assume the inode
3695          * has acls
3696          */
3697         if (*first_xattr_slot == -1)
3698                 *first_xattr_slot = slot;
3699         return 1;
3700 }
3701
3702 /*
3703  * read an inode from the btree into the in-memory inode
3704  */
3705 static int btrfs_read_locked_inode(struct inode *inode)
3706 {
3707         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
3708         struct btrfs_path *path;
3709         struct extent_buffer *leaf;
3710         struct btrfs_inode_item *inode_item;
3711         struct btrfs_root *root = BTRFS_I(inode)->root;
3712         struct btrfs_key location;
3713         unsigned long ptr;
3714         int maybe_acls;
3715         u32 rdev;
3716         int ret;
3717         bool filled = false;
3718         int first_xattr_slot;
3719
3720         ret = btrfs_fill_inode(inode, &rdev);
3721         if (!ret)
3722                 filled = true;
3723
3724         path = btrfs_alloc_path();
3725         if (!path) {
3726                 ret = -ENOMEM;
3727                 goto make_bad;
3728         }
3729
3730         memcpy(&location, &BTRFS_I(inode)->location, sizeof(location));
3731
3732         ret = btrfs_lookup_inode(NULL, root, path, &location, 0);
3733         if (ret) {
3734                 if (ret > 0)
3735                         ret = -ENOENT;
3736                 goto make_bad;
3737         }
3738
3739         leaf = path->nodes[0];
3740
3741         if (filled)
3742                 goto cache_index;
3743
3744         inode_item = btrfs_item_ptr(leaf, path->slots[0],
3745                                     struct btrfs_inode_item);
3746         inode->i_mode = btrfs_inode_mode(leaf, inode_item);
3747         set_nlink(inode, btrfs_inode_nlink(leaf, inode_item));
3748         i_uid_write(inode, btrfs_inode_uid(leaf, inode_item));
3749         i_gid_write(inode, btrfs_inode_gid(leaf, inode_item));
3750         btrfs_i_size_write(BTRFS_I(inode), btrfs_inode_size(leaf, inode_item));
3751
3752         inode->i_atime.tv_sec = btrfs_timespec_sec(leaf, &inode_item->atime);
3753         inode->i_atime.tv_nsec = btrfs_timespec_nsec(leaf, &inode_item->atime);
3754
3755         inode->i_mtime.tv_sec = btrfs_timespec_sec(leaf, &inode_item->mtime);
3756         inode->i_mtime.tv_nsec = btrfs_timespec_nsec(leaf, &inode_item->mtime);
3757
3758         inode->i_ctime.tv_sec = btrfs_timespec_sec(leaf, &inode_item->ctime);
3759         inode->i_ctime.tv_nsec = btrfs_timespec_nsec(leaf, &inode_item->ctime);
3760
3761         BTRFS_I(inode)->i_otime.tv_sec =
3762                 btrfs_timespec_sec(leaf, &inode_item->otime);
3763         BTRFS_I(inode)->i_otime.tv_nsec =
3764                 btrfs_timespec_nsec(leaf, &inode_item->otime);
3765
3766         inode_set_bytes(inode, btrfs_inode_nbytes(leaf, inode_item));
3767         BTRFS_I(inode)->generation = btrfs_inode_generation(leaf, inode_item);
3768         BTRFS_I(inode)->last_trans = btrfs_inode_transid(leaf, inode_item);
3769
3770         inode->i_version = btrfs_inode_sequence(leaf, inode_item);
3771         inode->i_generation = BTRFS_I(inode)->generation;
3772         inode->i_rdev = 0;
3773         rdev = btrfs_inode_rdev(leaf, inode_item);
3774
3775         BTRFS_I(inode)->index_cnt = (u64)-1;
3776         BTRFS_I(inode)->flags = btrfs_inode_flags(leaf, inode_item);
3777
3778 cache_index:
3779         /*
3780          * If we were modified in the current generation and evicted from memory
3781          * and then re-read we need to do a full sync since we don't have any
3782          * idea about which extents were modified before we were evicted from
3783          * cache.
3784          *
3785          * This is required for both inode re-read from disk and delayed inode
3786          * in delayed_nodes_tree.
3787          */
3788         if (BTRFS_I(inode)->last_trans == fs_info->generation)
3789                 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
3790                         &BTRFS_I(inode)->runtime_flags);
3791
3792         /*
3793          * We don't persist the id of the transaction where an unlink operation
3794          * against the inode was last made. So here we assume the inode might
3795          * have been evicted, and therefore the exact value of last_unlink_trans
3796          * lost, and set it to last_trans to avoid metadata inconsistencies
3797          * between the inode and its parent if the inode is fsync'ed and the log
3798          * replayed. For example, in the scenario:
3799          *
3800          * touch mydir/foo
3801          * ln mydir/foo mydir/bar
3802          * sync
3803          * unlink mydir/bar
3804          * echo 2 > /proc/sys/vm/drop_caches   # evicts inode
3805          * xfs_io -c fsync mydir/foo
3806          * <power failure>
3807          * mount fs, triggers fsync log replay
3808          *
3809          * We must make sure that when we fsync our inode foo we also log its
3810          * parent inode, otherwise after log replay the parent still has the
3811          * dentry with the "bar" name but our inode foo has a link count of 1
3812          * and doesn't have an inode ref with the name "bar" anymore.
3813          *
3814          * Setting last_unlink_trans to last_trans is a pessimistic approach,
3815          * but it guarantees correctness at the expense of occasional full
3816          * transaction commits on fsync if our inode is a directory, or if our
3817          * inode is not a directory, logging its parent unnecessarily.
3818          */
3819         BTRFS_I(inode)->last_unlink_trans = BTRFS_I(inode)->last_trans;
3820
3821         path->slots[0]++;
3822         if (inode->i_nlink != 1 ||
3823             path->slots[0] >= btrfs_header_nritems(leaf))
3824                 goto cache_acl;
3825
3826         btrfs_item_key_to_cpu(leaf, &location, path->slots[0]);
3827         if (location.objectid != btrfs_ino(BTRFS_I(inode)))
3828                 goto cache_acl;
3829
3830         ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
3831         if (location.type == BTRFS_INODE_REF_KEY) {
3832                 struct btrfs_inode_ref *ref;
3833
3834                 ref = (struct btrfs_inode_ref *)ptr;
3835                 BTRFS_I(inode)->dir_index = btrfs_inode_ref_index(leaf, ref);
3836         } else if (location.type == BTRFS_INODE_EXTREF_KEY) {
3837                 struct btrfs_inode_extref *extref;
3838
3839                 extref = (struct btrfs_inode_extref *)ptr;
3840                 BTRFS_I(inode)->dir_index = btrfs_inode_extref_index(leaf,
3841                                                                      extref);
3842         }
3843 cache_acl:
3844         /*
3845          * try to precache a NULL acl entry for files that don't have
3846          * any xattrs or acls
3847          */
3848         maybe_acls = acls_after_inode_item(leaf, path->slots[0],
3849                         btrfs_ino(BTRFS_I(inode)), &first_xattr_slot);
3850         if (first_xattr_slot != -1) {
3851                 path->slots[0] = first_xattr_slot;
3852                 ret = btrfs_load_inode_props(inode, path);
3853                 if (ret)
3854                         btrfs_err(fs_info,
3855                                   "error loading props for ino %llu (root %llu): %d",
3856                                   btrfs_ino(BTRFS_I(inode)),
3857                                   root->root_key.objectid, ret);
3858         }
3859         btrfs_free_path(path);
3860
3861         if (!maybe_acls)
3862                 cache_no_acl(inode);
3863
3864         switch (inode->i_mode & S_IFMT) {
3865         case S_IFREG:
3866                 inode->i_mapping->a_ops = &btrfs_aops;
3867                 BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
3868                 inode->i_fop = &btrfs_file_operations;
3869                 inode->i_op = &btrfs_file_inode_operations;
3870                 break;
3871         case S_IFDIR:
3872                 inode->i_fop = &btrfs_dir_file_operations;
3873                 inode->i_op = &btrfs_dir_inode_operations;
3874                 break;
3875         case S_IFLNK:
3876                 inode->i_op = &btrfs_symlink_inode_operations;
3877                 inode_nohighmem(inode);
3878                 inode->i_mapping->a_ops = &btrfs_symlink_aops;
3879                 break;
3880         default:
3881                 inode->i_op = &btrfs_special_inode_operations;
3882                 init_special_inode(inode, inode->i_mode, rdev);
3883                 break;
3884         }
3885
3886         btrfs_update_iflags(inode);
3887         return 0;
3888
3889 make_bad:
3890         btrfs_free_path(path);
3891         make_bad_inode(inode);
3892         return ret;
3893 }
3894
3895 /*
3896  * given a leaf and an inode, copy the inode fields into the leaf
3897  */
3898 static void fill_inode_item(struct btrfs_trans_handle *trans,
3899                             struct extent_buffer *leaf,
3900                             struct btrfs_inode_item *item,
3901                             struct inode *inode)
3902 {
3903         struct btrfs_map_token token;
3904
3905         btrfs_init_map_token(&token);
3906
3907         btrfs_set_token_inode_uid(leaf, item, i_uid_read(inode), &token);
3908         btrfs_set_token_inode_gid(leaf, item, i_gid_read(inode), &token);
3909         btrfs_set_token_inode_size(leaf, item, BTRFS_I(inode)->disk_i_size,
3910                                    &token);
3911         btrfs_set_token_inode_mode(leaf, item, inode->i_mode, &token);
3912         btrfs_set_token_inode_nlink(leaf, item, inode->i_nlink, &token);
3913
3914         btrfs_set_token_timespec_sec(leaf, &item->atime,
3915                                      inode->i_atime.tv_sec, &token);
3916         btrfs_set_token_timespec_nsec(leaf, &item->atime,
3917                                       inode->i_atime.tv_nsec, &token);
3918
3919         btrfs_set_token_timespec_sec(leaf, &item->mtime,
3920                                      inode->i_mtime.tv_sec, &token);
3921         btrfs_set_token_timespec_nsec(leaf, &item->mtime,
3922                                       inode->i_mtime.tv_nsec, &token);
3923
3924         btrfs_set_token_timespec_sec(leaf, &item->ctime,
3925                                      inode->i_ctime.tv_sec, &token);
3926         btrfs_set_token_timespec_nsec(leaf, &item->ctime,
3927                                       inode->i_ctime.tv_nsec, &token);
3928
3929         btrfs_set_token_timespec_sec(leaf, &item->otime,
3930                                      BTRFS_I(inode)->i_otime.tv_sec, &token);
3931         btrfs_set_token_timespec_nsec(leaf, &item->otime,
3932                                       BTRFS_I(inode)->i_otime.tv_nsec, &token);
3933
3934         btrfs_set_token_inode_nbytes(leaf, item, inode_get_bytes(inode),
3935                                      &token);
3936         btrfs_set_token_inode_generation(leaf, item, BTRFS_I(inode)->generation,
3937                                          &token);
3938         btrfs_set_token_inode_sequence(leaf, item, inode->i_version, &token);
3939         btrfs_set_token_inode_transid(leaf, item, trans->transid, &token);
3940         btrfs_set_token_inode_rdev(leaf, item, inode->i_rdev, &token);
3941         btrfs_set_token_inode_flags(leaf, item, BTRFS_I(inode)->flags, &token);
3942         btrfs_set_token_inode_block_group(leaf, item, 0, &token);
3943 }
3944
3945 /*
3946  * copy everything in the in-memory inode into the btree.
3947  */
3948 static noinline int btrfs_update_inode_item(struct btrfs_trans_handle *trans,
3949                                 struct btrfs_root *root, struct inode *inode)
3950 {
3951         struct btrfs_inode_item *inode_item;
3952         struct btrfs_path *path;
3953         struct extent_buffer *leaf;
3954         int ret;
3955
3956         path = btrfs_alloc_path();
3957         if (!path)
3958                 return -ENOMEM;
3959
3960         path->leave_spinning = 1;
3961         ret = btrfs_lookup_inode(trans, root, path, &BTRFS_I(inode)->location,
3962                                  1);
3963         if (ret) {
3964                 if (ret > 0)
3965                         ret = -ENOENT;
3966                 goto failed;
3967         }
3968
3969         leaf = path->nodes[0];
3970         inode_item = btrfs_item_ptr(leaf, path->slots[0],
3971                                     struct btrfs_inode_item);
3972
3973         fill_inode_item(trans, leaf, inode_item, inode);
3974         btrfs_mark_buffer_dirty(leaf);
3975         btrfs_set_inode_last_trans(trans, inode);
3976         ret = 0;
3977 failed:
3978         btrfs_free_path(path);
3979         return ret;
3980 }
3981
3982 /*
3983  * copy everything in the in-memory inode into the btree.
3984  */
3985 noinline int btrfs_update_inode(struct btrfs_trans_handle *trans,
3986                                 struct btrfs_root *root, struct inode *inode)
3987 {
3988         struct btrfs_fs_info *fs_info = root->fs_info;
3989         int ret;
3990
3991         /*
3992          * If the inode is a free space inode, we can deadlock during commit
3993          * if we put it into the delayed code.
3994          *
3995          * The data relocation inode should also be directly updated
3996          * without delay
3997          */
3998         if (!btrfs_is_free_space_inode(BTRFS_I(inode))
3999             && root->root_key.objectid != BTRFS_DATA_RELOC_TREE_OBJECTID
4000             && !test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags)) {
4001                 btrfs_update_root_times(trans, root);
4002
4003                 ret = btrfs_delayed_update_inode(trans, root, inode);
4004                 if (!ret)
4005                         btrfs_set_inode_last_trans(trans, inode);
4006                 return ret;
4007         }
4008
4009         return btrfs_update_inode_item(trans, root, inode);
4010 }
4011
4012 noinline int btrfs_update_inode_fallback(struct btrfs_trans_handle *trans,
4013                                          struct btrfs_root *root,
4014                                          struct inode *inode)
4015 {
4016         int ret;
4017
4018         ret = btrfs_update_inode(trans, root, inode);
4019         if (ret == -ENOSPC)
4020                 return btrfs_update_inode_item(trans, root, inode);
4021         return ret;
4022 }
4023
4024 /*
4025  * unlink helper that gets used here in inode.c and in the tree logging
4026  * recovery code.  It remove a link in a directory with a given name, and
4027  * also drops the back refs in the inode to the directory
4028  */
4029 static int __btrfs_unlink_inode(struct btrfs_trans_handle *trans,
4030                                 struct btrfs_root *root,
4031                                 struct btrfs_inode *dir,
4032                                 struct btrfs_inode *inode,
4033                                 const char *name, int name_len)
4034 {
4035         struct btrfs_fs_info *fs_info = root->fs_info;
4036         struct btrfs_path *path;
4037         int ret = 0;
4038         struct extent_buffer *leaf;
4039         struct btrfs_dir_item *di;
4040         struct btrfs_key key;
4041         u64 index;
4042         u64 ino = btrfs_ino(inode);
4043         u64 dir_ino = btrfs_ino(dir);
4044
4045         path = btrfs_alloc_path();
4046         if (!path) {
4047                 ret = -ENOMEM;
4048                 goto out;
4049         }
4050
4051         path->leave_spinning = 1;
4052         di = btrfs_lookup_dir_item(trans, root, path, dir_ino,
4053                                     name, name_len, -1);
4054         if (IS_ERR(di)) {
4055                 ret = PTR_ERR(di);
4056                 goto err;
4057         }
4058         if (!di) {
4059                 ret = -ENOENT;
4060                 goto err;
4061         }
4062         leaf = path->nodes[0];
4063         btrfs_dir_item_key_to_cpu(leaf, di, &key);
4064         ret = btrfs_delete_one_dir_name(trans, root, path, di);
4065         if (ret)
4066                 goto err;
4067         btrfs_release_path(path);
4068
4069         /*
4070          * If we don't have dir index, we have to get it by looking up
4071          * the inode ref, since we get the inode ref, remove it directly,
4072          * it is unnecessary to do delayed deletion.
4073          *
4074          * But if we have dir index, needn't search inode ref to get it.
4075          * Since the inode ref is close to the inode item, it is better
4076          * that we delay to delete it, and just do this deletion when
4077          * we update the inode item.
4078          */
4079         if (inode->dir_index) {
4080                 ret = btrfs_delayed_delete_inode_ref(inode);
4081                 if (!ret) {
4082                         index = inode->dir_index;
4083                         goto skip_backref;
4084                 }
4085         }
4086
4087         ret = btrfs_del_inode_ref(trans, root, name, name_len, ino,
4088                                   dir_ino, &index);
4089         if (ret) {
4090                 btrfs_info(fs_info,
4091                         "failed to delete reference to %.*s, inode %llu parent %llu",
4092                         name_len, name, ino, dir_ino);
4093                 btrfs_abort_transaction(trans, ret);
4094                 goto err;
4095         }
4096 skip_backref:
4097         ret = btrfs_delete_delayed_dir_index(trans, fs_info, dir, index);
4098         if (ret) {
4099                 btrfs_abort_transaction(trans, ret);
4100                 goto err;
4101         }
4102
4103         ret = btrfs_del_inode_ref_in_log(trans, root, name, name_len, inode,
4104                         dir_ino);
4105         if (ret != 0 && ret != -ENOENT) {
4106                 btrfs_abort_transaction(trans, ret);
4107                 goto err;
4108         }
4109
4110         ret = btrfs_del_dir_entries_in_log(trans, root, name, name_len, dir,
4111                         index);
4112         if (ret == -ENOENT)
4113                 ret = 0;
4114         else if (ret)
4115                 btrfs_abort_transaction(trans, ret);
4116 err:
4117         btrfs_free_path(path);
4118         if (ret)
4119                 goto out;
4120
4121         btrfs_i_size_write(dir, dir->vfs_inode.i_size - name_len * 2);
4122         inode_inc_iversion(&inode->vfs_inode);
4123         inode_inc_iversion(&dir->vfs_inode);
4124         inode->vfs_inode.i_ctime = dir->vfs_inode.i_mtime =
4125                 dir->vfs_inode.i_ctime = current_time(&inode->vfs_inode);
4126         ret = btrfs_update_inode(trans, root, &dir->vfs_inode);
4127 out:
4128         return ret;
4129 }
4130
4131 int btrfs_unlink_inode(struct btrfs_trans_handle *trans,
4132                        struct btrfs_root *root,
4133                        struct btrfs_inode *dir, struct btrfs_inode *inode,
4134                        const char *name, int name_len)
4135 {
4136         int ret;
4137         ret = __btrfs_unlink_inode(trans, root, dir, inode, name, name_len);
4138         if (!ret) {
4139                 drop_nlink(&inode->vfs_inode);
4140                 ret = btrfs_update_inode(trans, root, &inode->vfs_inode);
4141         }
4142         return ret;
4143 }
4144
4145 /*
4146  * helper to start transaction for unlink and rmdir.
4147  *
4148  * unlink and rmdir are special in btrfs, they do not always free space, so
4149  * if we cannot make our reservations the normal way try and see if there is
4150  * plenty of slack room in the global reserve to migrate, otherwise we cannot
4151  * allow the unlink to occur.
4152  */
4153 static struct btrfs_trans_handle *__unlink_start_trans(struct inode *dir)
4154 {
4155         struct btrfs_root *root = BTRFS_I(dir)->root;
4156
4157         /*
4158          * 1 for the possible orphan item
4159          * 1 for the dir item
4160          * 1 for the dir index
4161          * 1 for the inode ref
4162          * 1 for the inode
4163          */
4164         return btrfs_start_transaction_fallback_global_rsv(root, 5, 5);
4165 }
4166
4167 static int btrfs_unlink(struct inode *dir, struct dentry *dentry)
4168 {
4169         struct btrfs_root *root = BTRFS_I(dir)->root;
4170         struct btrfs_trans_handle *trans;
4171         struct inode *inode = d_inode(dentry);
4172         int ret;
4173
4174         trans = __unlink_start_trans(dir);
4175         if (IS_ERR(trans))
4176                 return PTR_ERR(trans);
4177
4178         btrfs_record_unlink_dir(trans, BTRFS_I(dir), BTRFS_I(d_inode(dentry)),
4179                         0);
4180
4181         ret = btrfs_unlink_inode(trans, root, BTRFS_I(dir),
4182                         BTRFS_I(d_inode(dentry)), dentry->d_name.name,
4183                         dentry->d_name.len);
4184         if (ret)
4185                 goto out;
4186
4187         if (inode->i_nlink == 0) {
4188                 ret = btrfs_orphan_add(trans, BTRFS_I(inode));
4189                 if (ret)
4190                         goto out;
4191         }
4192
4193 out:
4194         btrfs_end_transaction(trans);
4195         btrfs_btree_balance_dirty(root->fs_info);
4196         return ret;
4197 }
4198
4199 int btrfs_unlink_subvol(struct btrfs_trans_handle *trans,
4200                         struct btrfs_root *root,
4201                         struct inode *dir, u64 objectid,
4202                         const char *name, int name_len)
4203 {
4204         struct btrfs_fs_info *fs_info = root->fs_info;
4205         struct btrfs_path *path;
4206         struct extent_buffer *leaf;
4207         struct btrfs_dir_item *di;
4208         struct btrfs_key key;
4209         u64 index;
4210         int ret;
4211         u64 dir_ino = btrfs_ino(BTRFS_I(dir));
4212
4213         path = btrfs_alloc_path();
4214         if (!path)
4215                 return -ENOMEM;
4216
4217         di = btrfs_lookup_dir_item(trans, root, path, dir_ino,
4218                                    name, name_len, -1);
4219         if (IS_ERR_OR_NULL(di)) {
4220                 if (!di)
4221                         ret = -ENOENT;
4222                 else
4223                         ret = PTR_ERR(di);
4224                 goto out;
4225         }
4226
4227         leaf = path->nodes[0];
4228         btrfs_dir_item_key_to_cpu(leaf, di, &key);
4229         WARN_ON(key.type != BTRFS_ROOT_ITEM_KEY || key.objectid != objectid);
4230         ret = btrfs_delete_one_dir_name(trans, root, path, di);
4231         if (ret) {
4232                 btrfs_abort_transaction(trans, ret);
4233                 goto out;
4234         }
4235         btrfs_release_path(path);
4236
4237         ret = btrfs_del_root_ref(trans, fs_info, objectid,
4238                                  root->root_key.objectid, dir_ino,
4239                                  &index, name, name_len);
4240         if (ret < 0) {
4241                 if (ret != -ENOENT) {
4242                         btrfs_abort_transaction(trans, ret);
4243                         goto out;
4244                 }
4245                 di = btrfs_search_dir_index_item(root, path, dir_ino,
4246                                                  name, name_len);
4247                 if (IS_ERR_OR_NULL(di)) {
4248                         if (!di)
4249                                 ret = -ENOENT;
4250                         else
4251                                 ret = PTR_ERR(di);
4252                         btrfs_abort_transaction(trans, ret);
4253                         goto out;
4254                 }
4255
4256                 leaf = path->nodes[0];
4257                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
4258                 btrfs_release_path(path);
4259                 index = key.offset;
4260         }
4261         btrfs_release_path(path);
4262
4263         ret = btrfs_delete_delayed_dir_index(trans, fs_info, BTRFS_I(dir), index);
4264         if (ret) {
4265                 btrfs_abort_transaction(trans, ret);
4266                 goto out;
4267         }
4268
4269         btrfs_i_size_write(BTRFS_I(dir), dir->i_size - name_len * 2);
4270         inode_inc_iversion(dir);
4271         dir->i_mtime = dir->i_ctime = current_time(dir);
4272         ret = btrfs_update_inode_fallback(trans, root, dir);
4273         if (ret)
4274                 btrfs_abort_transaction(trans, ret);
4275 out:
4276         btrfs_free_path(path);
4277         return ret;
4278 }
4279
4280 static int btrfs_rmdir(struct inode *dir, struct dentry *dentry)
4281 {
4282         struct inode *inode = d_inode(dentry);
4283         int err = 0;
4284         struct btrfs_root *root = BTRFS_I(dir)->root;
4285         struct btrfs_trans_handle *trans;
4286         u64 last_unlink_trans;
4287
4288         if (inode->i_size > BTRFS_EMPTY_DIR_SIZE)
4289                 return -ENOTEMPTY;
4290         if (btrfs_ino(BTRFS_I(inode)) == BTRFS_FIRST_FREE_OBJECTID)
4291                 return -EPERM;
4292
4293         trans = __unlink_start_trans(dir);
4294         if (IS_ERR(trans))
4295                 return PTR_ERR(trans);
4296
4297         if (unlikely(btrfs_ino(BTRFS_I(inode)) == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)) {
4298                 err = btrfs_unlink_subvol(trans, root, dir,
4299                                           BTRFS_I(inode)->location.objectid,
4300                                           dentry->d_name.name,
4301                                           dentry->d_name.len);
4302                 goto out;
4303         }
4304
4305         err = btrfs_orphan_add(trans, BTRFS_I(inode));
4306         if (err)
4307                 goto out;
4308
4309         last_unlink_trans = BTRFS_I(inode)->last_unlink_trans;
4310
4311         /* now the directory is empty */
4312         err = btrfs_unlink_inode(trans, root, BTRFS_I(dir),
4313                         BTRFS_I(d_inode(dentry)), dentry->d_name.name,
4314                         dentry->d_name.len);
4315         if (!err) {
4316                 btrfs_i_size_write(BTRFS_I(inode), 0);
4317                 /*
4318                  * Propagate the last_unlink_trans value of the deleted dir to
4319                  * its parent directory. This is to prevent an unrecoverable
4320                  * log tree in the case we do something like this:
4321                  * 1) create dir foo
4322                  * 2) create snapshot under dir foo
4323                  * 3) delete the snapshot
4324                  * 4) rmdir foo
4325                  * 5) mkdir foo
4326                  * 6) fsync foo or some file inside foo
4327                  */
4328                 if (last_unlink_trans >= trans->transid)
4329                         BTRFS_I(dir)->last_unlink_trans = last_unlink_trans;
4330         }
4331 out:
4332         btrfs_end_transaction(trans);
4333         btrfs_btree_balance_dirty(root->fs_info);
4334
4335         return err;
4336 }
4337
4338 static int truncate_space_check(struct btrfs_trans_handle *trans,
4339                                 struct btrfs_root *root,
4340                                 u64 bytes_deleted)
4341 {
4342         struct btrfs_fs_info *fs_info = root->fs_info;
4343         int ret;
4344
4345         /*
4346          * This is only used to apply pressure to the enospc system, we don't
4347          * intend to use this reservation at all.
4348          */
4349         bytes_deleted = btrfs_csum_bytes_to_leaves(fs_info, bytes_deleted);
4350         bytes_deleted *= fs_info->nodesize;
4351         ret = btrfs_block_rsv_add(root, &fs_info->trans_block_rsv,
4352                                   bytes_deleted, BTRFS_RESERVE_NO_FLUSH);
4353         if (!ret) {
4354                 trace_btrfs_space_reservation(fs_info, "transaction",
4355                                               trans->transid,
4356                                               bytes_deleted, 1);
4357                 trans->bytes_reserved += bytes_deleted;
4358         }
4359         return ret;
4360
4361 }
4362
4363 static int truncate_inline_extent(struct inode *inode,
4364                                   struct btrfs_path *path,
4365                                   struct btrfs_key *found_key,
4366                                   const u64 item_end,
4367                                   const u64 new_size)
4368 {
4369         struct extent_buffer *leaf = path->nodes[0];
4370         int slot = path->slots[0];
4371         struct btrfs_file_extent_item *fi;
4372         u32 size = (u32)(new_size - found_key->offset);
4373         struct btrfs_root *root = BTRFS_I(inode)->root;
4374
4375         fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
4376
4377         if (btrfs_file_extent_compression(leaf, fi) != BTRFS_COMPRESS_NONE) {
4378                 loff_t offset = new_size;
4379                 loff_t page_end = ALIGN(offset, PAGE_SIZE);
4380
4381                 /*
4382                  * Zero out the remaining of the last page of our inline extent,
4383                  * instead of directly truncating our inline extent here - that
4384                  * would be much more complex (decompressing all the data, then
4385                  * compressing the truncated data, which might be bigger than
4386                  * the size of the inline extent, resize the extent, etc).
4387                  * We release the path because to get the page we might need to
4388                  * read the extent item from disk (data not in the page cache).
4389                  */
4390                 btrfs_release_path(path);
4391                 return btrfs_truncate_block(inode, offset, page_end - offset,
4392                                         0);
4393         }
4394
4395         btrfs_set_file_extent_ram_bytes(leaf, fi, size);
4396         size = btrfs_file_extent_calc_inline_size(size);
4397         btrfs_truncate_item(root->fs_info, path, size, 1);
4398
4399         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state))
4400                 inode_sub_bytes(inode, item_end + 1 - new_size);
4401
4402         return 0;
4403 }
4404
4405 /*
4406  * this can truncate away extent items, csum items and directory items.
4407  * It starts at a high offset and removes keys until it can't find
4408  * any higher than new_size
4409  *
4410  * csum items that cross the new i_size are truncated to the new size
4411  * as well.
4412  *
4413  * min_type is the minimum key type to truncate down to.  If set to 0, this
4414  * will kill all the items on this inode, including the INODE_ITEM_KEY.
4415  */
4416 int btrfs_truncate_inode_items(struct btrfs_trans_handle *trans,
4417                                struct btrfs_root *root,
4418                                struct inode *inode,
4419                                u64 new_size, u32 min_type)
4420 {
4421         struct btrfs_fs_info *fs_info = root->fs_info;
4422         struct btrfs_path *path;
4423         struct extent_buffer *leaf;
4424         struct btrfs_file_extent_item *fi;
4425         struct btrfs_key key;
4426         struct btrfs_key found_key;
4427         u64 extent_start = 0;
4428         u64 extent_num_bytes = 0;
4429         u64 extent_offset = 0;
4430         u64 item_end = 0;
4431         u64 last_size = new_size;
4432         u32 found_type = (u8)-1;
4433         int found_extent;
4434         int del_item;
4435         int pending_del_nr = 0;
4436         int pending_del_slot = 0;
4437         int extent_type = -1;
4438         int ret;
4439         int err = 0;
4440         u64 ino = btrfs_ino(BTRFS_I(inode));
4441         u64 bytes_deleted = 0;
4442         bool be_nice = 0;
4443         bool should_throttle = 0;
4444         bool should_end = 0;
4445
4446         BUG_ON(new_size > 0 && min_type != BTRFS_EXTENT_DATA_KEY);
4447
4448         /*
4449          * for non-free space inodes and ref cows, we want to back off from
4450          * time to time
4451          */
4452         if (!btrfs_is_free_space_inode(BTRFS_I(inode)) &&
4453             test_bit(BTRFS_ROOT_REF_COWS, &root->state))
4454                 be_nice = 1;
4455
4456         path = btrfs_alloc_path();
4457         if (!path)
4458                 return -ENOMEM;
4459         path->reada = READA_BACK;
4460
4461         /*
4462          * We want to drop from the next block forward in case this new size is
4463          * not block aligned since we will be keeping the last block of the
4464          * extent just the way it is.
4465          */
4466         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
4467             root == fs_info->tree_root)
4468                 btrfs_drop_extent_cache(BTRFS_I(inode), ALIGN(new_size,
4469                                         fs_info->sectorsize),
4470                                         (u64)-1, 0);
4471
4472         /*
4473          * This function is also used to drop the items in the log tree before
4474          * we relog the inode, so if root != BTRFS_I(inode)->root, it means
4475          * it is used to drop the loged items. So we shouldn't kill the delayed
4476          * items.
4477          */
4478         if (min_type == 0 && root == BTRFS_I(inode)->root)
4479                 btrfs_kill_delayed_inode_items(BTRFS_I(inode));
4480
4481         key.objectid = ino;
4482         key.offset = (u64)-1;
4483         key.type = (u8)-1;
4484
4485 search_again:
4486         /*
4487          * with a 16K leaf size and 128MB extents, you can actually queue
4488          * up a huge file in a single leaf.  Most of the time that
4489          * bytes_deleted is > 0, it will be huge by the time we get here
4490          */
4491         if (be_nice && bytes_deleted > SZ_32M) {
4492                 if (btrfs_should_end_transaction(trans)) {
4493                         err = -EAGAIN;
4494                         goto error;
4495                 }
4496         }
4497
4498
4499         path->leave_spinning = 1;
4500         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
4501         if (ret < 0) {
4502                 err = ret;
4503                 goto out;
4504         }
4505
4506         if (ret > 0) {
4507                 /* there are no items in the tree for us to truncate, we're
4508                  * done
4509                  */
4510                 if (path->slots[0] == 0)
4511                         goto out;
4512                 path->slots[0]--;
4513         }
4514
4515         while (1) {
4516                 fi = NULL;
4517                 leaf = path->nodes[0];
4518                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
4519                 found_type = found_key.type;
4520
4521                 if (found_key.objectid != ino)
4522                         break;
4523
4524                 if (found_type < min_type)
4525                         break;
4526
4527                 item_end = found_key.offset;
4528                 if (found_type == BTRFS_EXTENT_DATA_KEY) {
4529                         fi = btrfs_item_ptr(leaf, path->slots[0],
4530                                             struct btrfs_file_extent_item);
4531                         extent_type = btrfs_file_extent_type(leaf, fi);
4532                         if (extent_type != BTRFS_FILE_EXTENT_INLINE) {
4533                                 item_end +=
4534                                     btrfs_file_extent_num_bytes(leaf, fi);
4535
4536                                 trace_btrfs_truncate_show_fi_regular(
4537                                         BTRFS_I(inode), leaf, fi,
4538                                         found_key.offset);
4539                         } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
4540                                 item_end += btrfs_file_extent_inline_len(leaf,
4541                                                          path->slots[0], fi);
4542
4543                                 trace_btrfs_truncate_show_fi_inline(
4544                                         BTRFS_I(inode), leaf, fi, path->slots[0],
4545                                         found_key.offset);
4546                         }
4547                         item_end--;
4548                 }
4549                 if (found_type > min_type) {
4550                         del_item = 1;
4551                 } else {
4552                         if (item_end < new_size)
4553                                 break;
4554                         if (found_key.offset >= new_size)
4555                                 del_item = 1;
4556                         else
4557                                 del_item = 0;
4558                 }
4559                 found_extent = 0;
4560                 /* FIXME, shrink the extent if the ref count is only 1 */
4561                 if (found_type != BTRFS_EXTENT_DATA_KEY)
4562                         goto delete;
4563
4564                 if (del_item)
4565                         last_size = found_key.offset;
4566                 else
4567                         last_size = new_size;
4568
4569                 if (extent_type != BTRFS_FILE_EXTENT_INLINE) {
4570                         u64 num_dec;
4571                         extent_start = btrfs_file_extent_disk_bytenr(leaf, fi);
4572                         if (!del_item) {
4573                                 u64 orig_num_bytes =
4574                                         btrfs_file_extent_num_bytes(leaf, fi);
4575                                 extent_num_bytes = ALIGN(new_size -
4576                                                 found_key.offset,
4577                                                 fs_info->sectorsize);
4578                                 btrfs_set_file_extent_num_bytes(leaf, fi,
4579                                                          extent_num_bytes);
4580                                 num_dec = (orig_num_bytes -
4581                                            extent_num_bytes);
4582                                 if (test_bit(BTRFS_ROOT_REF_COWS,
4583                                              &root->state) &&
4584                                     extent_start != 0)
4585                                         inode_sub_bytes(inode, num_dec);
4586                                 btrfs_mark_buffer_dirty(leaf);
4587                         } else {
4588                                 extent_num_bytes =
4589                                         btrfs_file_extent_disk_num_bytes(leaf,
4590                                                                          fi);
4591                                 extent_offset = found_key.offset -
4592                                         btrfs_file_extent_offset(leaf, fi);
4593
4594                                 /* FIXME blocksize != 4096 */
4595                                 num_dec = btrfs_file_extent_num_bytes(leaf, fi);
4596                                 if (extent_start != 0) {
4597                                         found_extent = 1;
4598                                         if (test_bit(BTRFS_ROOT_REF_COWS,
4599                                                      &root->state))
4600                                                 inode_sub_bytes(inode, num_dec);
4601                                 }
4602                         }
4603                 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
4604                         /*
4605                          * we can't truncate inline items that have had
4606                          * special encodings
4607                          */
4608                         if (!del_item &&
4609                             btrfs_file_extent_encryption(leaf, fi) == 0 &&
4610                             btrfs_file_extent_other_encoding(leaf, fi) == 0) {
4611
4612                                 /*
4613                                  * Need to release path in order to truncate a
4614                                  * compressed extent. So delete any accumulated
4615                                  * extent items so far.
4616                                  */
4617                                 if (btrfs_file_extent_compression(leaf, fi) !=
4618                                     BTRFS_COMPRESS_NONE && pending_del_nr) {
4619                                         err = btrfs_del_items(trans, root, path,
4620                                                               pending_del_slot,
4621                                                               pending_del_nr);
4622                                         if (err) {
4623                                                 btrfs_abort_transaction(trans,
4624                                                                         err);
4625                                                 goto error;
4626                                         }
4627                                         pending_del_nr = 0;
4628                                 }
4629
4630                                 err = truncate_inline_extent(inode, path,
4631                                                              &found_key,
4632                                                              item_end,
4633                                                              new_size);
4634                                 if (err) {
4635                                         btrfs_abort_transaction(trans, err);
4636                                         goto error;
4637                                 }
4638                         } else if (test_bit(BTRFS_ROOT_REF_COWS,
4639                                             &root->state)) {
4640                                 inode_sub_bytes(inode, item_end + 1 - new_size);
4641                         }
4642                 }
4643 delete:
4644                 if (del_item) {
4645                         if (!pending_del_nr) {
4646                                 /* no pending yet, add ourselves */
4647                                 pending_del_slot = path->slots[0];
4648                                 pending_del_nr = 1;
4649                         } else if (pending_del_nr &&
4650                                    path->slots[0] + 1 == pending_del_slot) {
4651                                 /* hop on the pending chunk */
4652                                 pending_del_nr++;
4653                                 pending_del_slot = path->slots[0];
4654                         } else {
4655                                 BUG();
4656                         }
4657                 } else {
4658                         break;
4659                 }
4660                 should_throttle = 0;
4661
4662                 if (found_extent &&
4663                     (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
4664                      root == fs_info->tree_root)) {
4665                         btrfs_set_path_blocking(path);
4666                         bytes_deleted += extent_num_bytes;
4667                         ret = btrfs_free_extent(trans, fs_info, extent_start,
4668                                                 extent_num_bytes, 0,
4669                                                 btrfs_header_owner(leaf),
4670                                                 ino, extent_offset);
4671                         BUG_ON(ret);
4672                         if (btrfs_should_throttle_delayed_refs(trans, fs_info))
4673                                 btrfs_async_run_delayed_refs(fs_info,
4674                                         trans->delayed_ref_updates * 2,
4675                                         trans->transid, 0);
4676                         if (be_nice) {
4677                                 if (truncate_space_check(trans, root,
4678                                                          extent_num_bytes)) {
4679                                         should_end = 1;
4680                                 }
4681                                 if (btrfs_should_throttle_delayed_refs(trans,
4682                                                                        fs_info))
4683                                         should_throttle = 1;
4684                         }
4685                 }
4686
4687                 if (found_type == BTRFS_INODE_ITEM_KEY)
4688                         break;
4689
4690                 if (path->slots[0] == 0 ||
4691                     path->slots[0] != pending_del_slot ||
4692                     should_throttle || should_end) {
4693                         if (pending_del_nr) {
4694                                 ret = btrfs_del_items(trans, root, path,
4695                                                 pending_del_slot,
4696                                                 pending_del_nr);
4697                                 if (ret) {
4698                                         btrfs_abort_transaction(trans, ret);
4699                                         goto error;
4700                                 }
4701                                 pending_del_nr = 0;
4702                         }
4703                         btrfs_release_path(path);
4704                         if (should_throttle) {
4705                                 unsigned long updates = trans->delayed_ref_updates;
4706                                 if (updates) {
4707                                         trans->delayed_ref_updates = 0;
4708                                         ret = btrfs_run_delayed_refs(trans,
4709                                                                    fs_info,
4710                                                                    updates * 2);
4711                                         if (ret && !err)
4712                                                 err = ret;
4713                                 }
4714                         }
4715                         /*
4716                          * if we failed to refill our space rsv, bail out
4717                          * and let the transaction restart
4718                          */
4719                         if (should_end) {
4720                                 err = -EAGAIN;
4721                                 goto error;
4722                         }
4723                         goto search_again;
4724                 } else {
4725                         path->slots[0]--;
4726                 }
4727         }
4728 out:
4729         if (pending_del_nr) {
4730                 ret = btrfs_del_items(trans, root, path, pending_del_slot,
4731                                       pending_del_nr);
4732                 if (ret)
4733                         btrfs_abort_transaction(trans, ret);
4734         }
4735 error:
4736         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
4737                 ASSERT(last_size >= new_size);
4738                 if (!err && last_size > new_size)
4739                         last_size = new_size;
4740                 btrfs_ordered_update_i_size(inode, last_size, NULL);
4741         }
4742
4743         btrfs_free_path(path);
4744
4745         if (be_nice && bytes_deleted > SZ_32M) {
4746                 unsigned long updates = trans->delayed_ref_updates;
4747                 if (updates) {
4748                         trans->delayed_ref_updates = 0;
4749                         ret = btrfs_run_delayed_refs(trans, fs_info,
4750                                                      updates * 2);
4751                         if (ret && !err)
4752                                 err = ret;
4753                 }
4754         }
4755         return err;
4756 }
4757
4758 /*
4759  * btrfs_truncate_block - read, zero a chunk and write a block
4760  * @inode - inode that we're zeroing
4761  * @from - the offset to start zeroing
4762  * @len - the length to zero, 0 to zero the entire range respective to the
4763  *      offset
4764  * @front - zero up to the offset instead of from the offset on
4765  *
4766  * This will find the block for the "from" offset and cow the block and zero the
4767  * part we want to zero.  This is used with truncate and hole punching.
4768  */
4769 int btrfs_truncate_block(struct inode *inode, loff_t from, loff_t len,
4770                         int front)
4771 {
4772         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
4773         struct address_space *mapping = inode->i_mapping;
4774         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
4775         struct btrfs_ordered_extent *ordered;
4776         struct extent_state *cached_state = NULL;
4777         struct extent_changeset *data_reserved = NULL;
4778         char *kaddr;
4779         u32 blocksize = fs_info->sectorsize;
4780         pgoff_t index = from >> PAGE_SHIFT;
4781         unsigned offset = from & (blocksize - 1);
4782         struct page *page;
4783         gfp_t mask = btrfs_alloc_write_mask(mapping);
4784         int ret = 0;
4785         u64 block_start;
4786         u64 block_end;
4787
4788         if ((offset & (blocksize - 1)) == 0 &&
4789             (!len || ((len & (blocksize - 1)) == 0)))
4790                 goto out;
4791
4792         ret = btrfs_delalloc_reserve_space(inode, &data_reserved,
4793                         round_down(from, blocksize), blocksize);
4794         if (ret)
4795                 goto out;
4796
4797 again:
4798         page = find_or_create_page(mapping, index, mask);
4799         if (!page) {
4800                 btrfs_delalloc_release_space(inode, data_reserved,
4801                                 round_down(from, blocksize),
4802                                 blocksize);
4803                 ret = -ENOMEM;
4804                 goto out;
4805         }
4806
4807         block_start = round_down(from, blocksize);
4808         block_end = block_start + blocksize - 1;
4809
4810         if (!PageUptodate(page)) {
4811                 ret = btrfs_readpage(NULL, page);
4812                 lock_page(page);
4813                 if (page->mapping != mapping) {
4814                         unlock_page(page);
4815                         put_page(page);
4816                         goto again;
4817                 }
4818                 if (!PageUptodate(page)) {
4819                         ret = -EIO;
4820                         goto out_unlock;
4821                 }
4822         }
4823         wait_on_page_writeback(page);
4824
4825         lock_extent_bits(io_tree, block_start, block_end, &cached_state);
4826         set_page_extent_mapped(page);
4827
4828         ordered = btrfs_lookup_ordered_extent(inode, block_start);
4829         if (ordered) {
4830                 unlock_extent_cached(io_tree, block_start, block_end,
4831                                      &cached_state, GFP_NOFS);
4832                 unlock_page(page);
4833                 put_page(page);
4834                 btrfs_start_ordered_extent(inode, ordered, 1);
4835                 btrfs_put_ordered_extent(ordered);
4836                 goto again;
4837         }
4838
4839         clear_extent_bit(&BTRFS_I(inode)->io_tree, block_start, block_end,
4840                           EXTENT_DIRTY | EXTENT_DELALLOC |
4841                           EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG,
4842                           0, 0, &cached_state, GFP_NOFS);
4843
4844         ret = btrfs_set_extent_delalloc(inode, block_start, block_end,
4845                                         &cached_state, 0);
4846         if (ret) {
4847                 unlock_extent_cached(io_tree, block_start, block_end,
4848                                      &cached_state, GFP_NOFS);
4849                 goto out_unlock;
4850         }
4851
4852         if (offset != blocksize) {
4853                 if (!len)
4854                         len = blocksize - offset;
4855                 kaddr = kmap(page);
4856                 if (front)
4857                         memset(kaddr + (block_start - page_offset(page)),
4858                                 0, offset);
4859                 else
4860                         memset(kaddr + (block_start - page_offset(page)) +  offset,
4861                                 0, len);
4862                 flush_dcache_page(page);
4863                 kunmap(page);
4864         }
4865         ClearPageChecked(page);
4866         set_page_dirty(page);
4867         unlock_extent_cached(io_tree, block_start, block_end, &cached_state,
4868                              GFP_NOFS);
4869
4870 out_unlock:
4871         if (ret)
4872                 btrfs_delalloc_release_space(inode, data_reserved, block_start,
4873                                              blocksize);
4874         unlock_page(page);
4875         put_page(page);
4876 out:
4877         extent_changeset_free(data_reserved);
4878         return ret;
4879 }
4880
4881 static int maybe_insert_hole(struct btrfs_root *root, struct inode *inode,
4882                              u64 offset, u64 len)
4883 {
4884         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
4885         struct btrfs_trans_handle *trans;
4886         int ret;
4887
4888         /*
4889          * Still need to make sure the inode looks like it's been updated so
4890          * that any holes get logged if we fsync.
4891          */
4892         if (btrfs_fs_incompat(fs_info, NO_HOLES)) {
4893                 BTRFS_I(inode)->last_trans = fs_info->generation;
4894                 BTRFS_I(inode)->last_sub_trans = root->log_transid;
4895                 BTRFS_I(inode)->last_log_commit = root->last_log_commit;
4896                 return 0;
4897         }
4898
4899         /*
4900          * 1 - for the one we're dropping
4901          * 1 - for the one we're adding
4902          * 1 - for updating the inode.
4903          */
4904         trans = btrfs_start_transaction(root, 3);
4905         if (IS_ERR(trans))
4906                 return PTR_ERR(trans);
4907
4908         ret = btrfs_drop_extents(trans, root, inode, offset, offset + len, 1);
4909         if (ret) {
4910                 btrfs_abort_transaction(trans, ret);
4911                 btrfs_end_transaction(trans);
4912                 return ret;
4913         }
4914
4915         ret = btrfs_insert_file_extent(trans, root, btrfs_ino(BTRFS_I(inode)),
4916                         offset, 0, 0, len, 0, len, 0, 0, 0);
4917         if (ret)
4918                 btrfs_abort_transaction(trans, ret);
4919         else
4920                 btrfs_update_inode(trans, root, inode);
4921         btrfs_end_transaction(trans);
4922         return ret;
4923 }
4924
4925 /*
4926  * This function puts in dummy file extents for the area we're creating a hole
4927  * for.  So if we are truncating this file to a larger size we need to insert
4928  * these file extents so that btrfs_get_extent will return a EXTENT_MAP_HOLE for
4929  * the range between oldsize and size
4930  */
4931 int btrfs_cont_expand(struct inode *inode, loff_t oldsize, loff_t size)
4932 {
4933         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
4934         struct btrfs_root *root = BTRFS_I(inode)->root;
4935         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
4936         struct extent_map *em = NULL;
4937         struct extent_state *cached_state = NULL;
4938         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
4939         u64 hole_start = ALIGN(oldsize, fs_info->sectorsize);
4940         u64 block_end = ALIGN(size, fs_info->sectorsize);
4941         u64 last_byte;
4942         u64 cur_offset;
4943         u64 hole_size;
4944         int err = 0;
4945
4946         /*
4947          * If our size started in the middle of a block we need to zero out the
4948          * rest of the block before we expand the i_size, otherwise we could
4949          * expose stale data.
4950          */
4951         err = btrfs_truncate_block(inode, oldsize, 0, 0);
4952         if (err)
4953                 return err;
4954
4955         if (size <= hole_start)
4956                 return 0;
4957
4958         while (1) {
4959                 struct btrfs_ordered_extent *ordered;
4960
4961                 lock_extent_bits(io_tree, hole_start, block_end - 1,
4962                                  &cached_state);
4963                 ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), hole_start,
4964                                                      block_end - hole_start);
4965                 if (!ordered)
4966                         break;
4967                 unlock_extent_cached(io_tree, hole_start, block_end - 1,
4968                                      &cached_state, GFP_NOFS);
4969                 btrfs_start_ordered_extent(inode, ordered, 1);
4970                 btrfs_put_ordered_extent(ordered);
4971         }
4972
4973         cur_offset = hole_start;
4974         while (1) {
4975                 em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, cur_offset,
4976                                 block_end - cur_offset, 0);
4977                 if (IS_ERR(em)) {
4978                         err = PTR_ERR(em);
4979                         em = NULL;
4980                         break;
4981                 }
4982                 last_byte = min(extent_map_end(em), block_end);
4983                 last_byte = ALIGN(last_byte, fs_info->sectorsize);
4984                 if (!test_bit(EXTENT_FLAG_PREALLOC, &em->flags)) {
4985                         struct extent_map *hole_em;
4986                         hole_size = last_byte - cur_offset;
4987
4988                         err = maybe_insert_hole(root, inode, cur_offset,
4989                                                 hole_size);
4990                         if (err)
4991                                 break;
4992                         btrfs_drop_extent_cache(BTRFS_I(inode), cur_offset,
4993                                                 cur_offset + hole_size - 1, 0);
4994                         hole_em = alloc_extent_map();
4995                         if (!hole_em) {
4996                                 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
4997                                         &BTRFS_I(inode)->runtime_flags);
4998                                 goto next;
4999                         }
5000                         hole_em->start = cur_offset;
5001                         hole_em->len = hole_size;
5002                         hole_em->orig_start = cur_offset;
5003
5004                         hole_em->block_start = EXTENT_MAP_HOLE;
5005                         hole_em->block_len = 0;
5006                         hole_em->orig_block_len = 0;
5007                         hole_em->ram_bytes = hole_size;
5008                         hole_em->bdev = fs_info->fs_devices->latest_bdev;
5009                         hole_em->compress_type = BTRFS_COMPRESS_NONE;
5010                         hole_em->generation = fs_info->generation;
5011
5012                         while (1) {
5013                                 write_lock(&em_tree->lock);
5014                                 err = add_extent_mapping(em_tree, hole_em, 1);
5015                                 write_unlock(&em_tree->lock);
5016                                 if (err != -EEXIST)
5017                                         break;
5018                                 btrfs_drop_extent_cache(BTRFS_I(inode),
5019                                                         cur_offset,
5020                                                         cur_offset +
5021                                                         hole_size - 1, 0);
5022                         }
5023                         free_extent_map(hole_em);
5024                 }
5025 next:
5026                 free_extent_map(em);
5027                 em = NULL;
5028                 cur_offset = last_byte;
5029                 if (cur_offset >= block_end)
5030                         break;
5031         }
5032         free_extent_map(em);
5033         unlock_extent_cached(io_tree, hole_start, block_end - 1, &cached_state,
5034                              GFP_NOFS);
5035         return err;
5036 }
5037
5038 static int btrfs_setsize(struct inode *inode, struct iattr *attr)
5039 {
5040         struct btrfs_root *root = BTRFS_I(inode)->root;
5041         struct btrfs_trans_handle *trans;
5042         loff_t oldsize = i_size_read(inode);
5043         loff_t newsize = attr->ia_size;
5044         int mask = attr->ia_valid;
5045         int ret;
5046
5047         /*
5048          * The regular truncate() case without ATTR_CTIME and ATTR_MTIME is a
5049          * special case where we need to update the times despite not having
5050          * these flags set.  For all other operations the VFS set these flags
5051          * explicitly if it wants a timestamp update.
5052          */
5053         if (newsize != oldsize) {
5054                 inode_inc_iversion(inode);
5055                 if (!(mask & (ATTR_CTIME | ATTR_MTIME)))
5056                         inode->i_ctime = inode->i_mtime =
5057                                 current_time(inode);
5058         }
5059
5060         if (newsize > oldsize) {
5061                 /*
5062                  * Don't do an expanding truncate while snapshotting is ongoing.
5063                  * This is to ensure the snapshot captures a fully consistent
5064                  * state of this file - if the snapshot captures this expanding
5065                  * truncation, it must capture all writes that happened before
5066                  * this truncation.
5067                  */
5068                 btrfs_wait_for_snapshot_creation(root);
5069                 ret = btrfs_cont_expand(inode, oldsize, newsize);
5070                 if (ret) {
5071                         btrfs_end_write_no_snapshotting(root);
5072                         return ret;
5073                 }
5074
5075                 trans = btrfs_start_transaction(root, 1);
5076                 if (IS_ERR(trans)) {
5077                         btrfs_end_write_no_snapshotting(root);
5078                         return PTR_ERR(trans);
5079                 }
5080
5081                 i_size_write(inode, newsize);
5082                 btrfs_ordered_update_i_size(inode, i_size_read(inode), NULL);
5083                 pagecache_isize_extended(inode, oldsize, newsize);
5084                 ret = btrfs_update_inode(trans, root, inode);
5085                 btrfs_end_write_no_snapshotting(root);
5086                 btrfs_end_transaction(trans);
5087         } else {
5088
5089                 /*
5090                  * We're truncating a file that used to have good data down to
5091                  * zero. Make sure it gets into the ordered flush list so that
5092                  * any new writes get down to disk quickly.
5093                  */
5094                 if (newsize == 0)
5095                         set_bit(BTRFS_INODE_ORDERED_DATA_CLOSE,
5096                                 &BTRFS_I(inode)->runtime_flags);
5097
5098                 /*
5099                  * 1 for the orphan item we're going to add
5100                  * 1 for the orphan item deletion.
5101                  */
5102                 trans = btrfs_start_transaction(root, 2);
5103                 if (IS_ERR(trans))
5104                         return PTR_ERR(trans);
5105
5106                 /*
5107                  * We need to do this in case we fail at _any_ point during the
5108                  * actual truncate.  Once we do the truncate_setsize we could
5109                  * invalidate pages which forces any outstanding ordered io to
5110                  * be instantly completed which will give us extents that need
5111                  * to be truncated.  If we fail to get an orphan inode down we
5112                  * could have left over extents that were never meant to live,
5113                  * so we need to guarantee from this point on that everything
5114                  * will be consistent.
5115                  */
5116                 ret = btrfs_orphan_add(trans, BTRFS_I(inode));
5117                 btrfs_end_transaction(trans);
5118                 if (ret)
5119                         return ret;
5120
5121                 /* we don't support swapfiles, so vmtruncate shouldn't fail */
5122                 truncate_setsize(inode, newsize);
5123
5124                 /* Disable nonlocked read DIO to avoid the end less truncate */
5125                 btrfs_inode_block_unlocked_dio(BTRFS_I(inode));
5126                 inode_dio_wait(inode);
5127                 btrfs_inode_resume_unlocked_dio(BTRFS_I(inode));
5128
5129                 ret = btrfs_truncate(inode);
5130                 if (ret && inode->i_nlink) {
5131                         int err;
5132
5133                         /* To get a stable disk_i_size */
5134                         err = btrfs_wait_ordered_range(inode, 0, (u64)-1);
5135                         if (err) {
5136                                 btrfs_orphan_del(NULL, BTRFS_I(inode));
5137                                 return err;
5138                         }
5139
5140                         /*
5141                          * failed to truncate, disk_i_size is only adjusted down
5142                          * as we remove extents, so it should represent the true
5143                          * size of the inode, so reset the in memory size and
5144                          * delete our orphan entry.
5145                          */
5146                         trans = btrfs_join_transaction(root);
5147                         if (IS_ERR(trans)) {
5148                                 btrfs_orphan_del(NULL, BTRFS_I(inode));
5149                                 return ret;
5150                         }
5151                         i_size_write(inode, BTRFS_I(inode)->disk_i_size);
5152                         err = btrfs_orphan_del(trans, BTRFS_I(inode));
5153                         if (err)
5154                                 btrfs_abort_transaction(trans, err);
5155                         btrfs_end_transaction(trans);
5156                 }
5157         }
5158
5159         return ret;
5160 }
5161
5162 static int btrfs_setattr(struct dentry *dentry, struct iattr *attr)
5163 {
5164         struct inode *inode = d_inode(dentry);
5165         struct btrfs_root *root = BTRFS_I(inode)->root;
5166         int err;
5167
5168         if (btrfs_root_readonly(root))
5169                 return -EROFS;
5170
5171         err = setattr_prepare(dentry, attr);
5172         if (err)
5173                 return err;
5174
5175         if (S_ISREG(inode->i_mode) && (attr->ia_valid & ATTR_SIZE)) {
5176                 err = btrfs_setsize(inode, attr);
5177                 if (err)
5178                         return err;
5179         }
5180
5181         if (attr->ia_valid) {
5182                 setattr_copy(inode, attr);
5183                 inode_inc_iversion(inode);
5184                 err = btrfs_dirty_inode(inode);
5185
5186                 if (!err && attr->ia_valid & ATTR_MODE)
5187                         err = posix_acl_chmod(inode, inode->i_mode);
5188         }
5189
5190         return err;
5191 }
5192
5193 /*
5194  * While truncating the inode pages during eviction, we get the VFS calling
5195  * btrfs_invalidatepage() against each page of the inode. This is slow because
5196  * the calls to btrfs_invalidatepage() result in a huge amount of calls to
5197  * lock_extent_bits() and clear_extent_bit(), which keep merging and splitting
5198  * extent_state structures over and over, wasting lots of time.
5199  *
5200  * Therefore if the inode is being evicted, let btrfs_invalidatepage() skip all
5201  * those expensive operations on a per page basis and do only the ordered io
5202  * finishing, while we release here the extent_map and extent_state structures,
5203  * without the excessive merging and splitting.
5204  */
5205 static void evict_inode_truncate_pages(struct inode *inode)
5206 {
5207         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
5208         struct extent_map_tree *map_tree = &BTRFS_I(inode)->extent_tree;
5209         struct rb_node *node;
5210
5211         ASSERT(inode->i_state & I_FREEING);
5212         truncate_inode_pages_final(&inode->i_data);
5213
5214         write_lock(&map_tree->lock);
5215         while (!RB_EMPTY_ROOT(&map_tree->map)) {
5216                 struct extent_map *em;
5217
5218                 node = rb_first(&map_tree->map);
5219                 em = rb_entry(node, struct extent_map, rb_node);
5220                 clear_bit(EXTENT_FLAG_PINNED, &em->flags);
5221                 clear_bit(EXTENT_FLAG_LOGGING, &em->flags);
5222                 remove_extent_mapping(map_tree, em);
5223                 free_extent_map(em);
5224                 if (need_resched()) {
5225                         write_unlock(&map_tree->lock);
5226                         cond_resched();
5227                         write_lock(&map_tree->lock);
5228                 }
5229         }
5230         write_unlock(&map_tree->lock);
5231
5232         /*
5233          * Keep looping until we have no more ranges in the io tree.
5234          * We can have ongoing bios started by readpages (called from readahead)
5235          * that have their endio callback (extent_io.c:end_bio_extent_readpage)
5236          * still in progress (unlocked the pages in the bio but did not yet
5237          * unlocked the ranges in the io tree). Therefore this means some
5238          * ranges can still be locked and eviction started because before
5239          * submitting those bios, which are executed by a separate task (work
5240          * queue kthread), inode references (inode->i_count) were not taken
5241          * (which would be dropped in the end io callback of each bio).
5242          * Therefore here we effectively end up waiting for those bios and
5243          * anyone else holding locked ranges without having bumped the inode's
5244          * reference count - if we don't do it, when they access the inode's
5245          * io_tree to unlock a range it may be too late, leading to an
5246          * use-after-free issue.
5247          */
5248         spin_lock(&io_tree->lock);
5249         while (!RB_EMPTY_ROOT(&io_tree->state)) {
5250                 struct extent_state *state;
5251                 struct extent_state *cached_state = NULL;
5252                 u64 start;
5253                 u64 end;
5254
5255                 node = rb_first(&io_tree->state);
5256                 state = rb_entry(node, struct extent_state, rb_node);
5257                 start = state->start;
5258                 end = state->end;
5259                 spin_unlock(&io_tree->lock);
5260
5261                 lock_extent_bits(io_tree, start, end, &cached_state);
5262
5263                 /*
5264                  * If still has DELALLOC flag, the extent didn't reach disk,
5265                  * and its reserved space won't be freed by delayed_ref.
5266                  * So we need to free its reserved space here.
5267                  * (Refer to comment in btrfs_invalidatepage, case 2)
5268                  *
5269                  * Note, end is the bytenr of last byte, so we need + 1 here.
5270                  */
5271                 if (state->state & EXTENT_DELALLOC)
5272                         btrfs_qgroup_free_data(inode, NULL, start, end - start + 1);
5273
5274                 clear_extent_bit(io_tree, start, end,
5275                                  EXTENT_LOCKED | EXTENT_DIRTY |
5276                                  EXTENT_DELALLOC | EXTENT_DO_ACCOUNTING |
5277                                  EXTENT_DEFRAG, 1, 1,
5278                                  &cached_state, GFP_NOFS);
5279
5280                 cond_resched();
5281                 spin_lock(&io_tree->lock);
5282         }
5283         spin_unlock(&io_tree->lock);
5284 }
5285
5286 void btrfs_evict_inode(struct inode *inode)
5287 {
5288         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
5289         struct btrfs_trans_handle *trans;
5290         struct btrfs_root *root = BTRFS_I(inode)->root;
5291         struct btrfs_block_rsv *rsv, *global_rsv;
5292         int steal_from_global = 0;
5293         u64 min_size;
5294         int ret;
5295
5296         trace_btrfs_inode_evict(inode);
5297
5298         if (!root) {
5299                 kmem_cache_free(btrfs_inode_cachep, BTRFS_I(inode));
5300                 return;
5301         }
5302
5303         min_size = btrfs_calc_trunc_metadata_size(fs_info, 1);
5304
5305         evict_inode_truncate_pages(inode);
5306
5307         if (inode->i_nlink &&
5308             ((btrfs_root_refs(&root->root_item) != 0 &&
5309               root->root_key.objectid != BTRFS_ROOT_TREE_OBJECTID) ||
5310              btrfs_is_free_space_inode(BTRFS_I(inode))))
5311                 goto no_delete;
5312
5313         if (is_bad_inode(inode)) {
5314                 btrfs_orphan_del(NULL, BTRFS_I(inode));
5315                 goto no_delete;
5316         }
5317         /* do we really want it for ->i_nlink > 0 and zero btrfs_root_refs? */
5318         if (!special_file(inode->i_mode))
5319                 btrfs_wait_ordered_range(inode, 0, (u64)-1);
5320
5321         btrfs_free_io_failure_record(BTRFS_I(inode), 0, (u64)-1);
5322
5323         if (test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags)) {
5324                 BUG_ON(test_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
5325                                  &BTRFS_I(inode)->runtime_flags));
5326                 goto no_delete;
5327         }
5328
5329         if (inode->i_nlink > 0) {
5330                 BUG_ON(btrfs_root_refs(&root->root_item) != 0 &&
5331                        root->root_key.objectid != BTRFS_ROOT_TREE_OBJECTID);
5332                 goto no_delete;
5333         }
5334
5335         ret = btrfs_commit_inode_delayed_inode(BTRFS_I(inode));
5336         if (ret) {
5337                 btrfs_orphan_del(NULL, BTRFS_I(inode));
5338                 goto no_delete;
5339         }
5340
5341         rsv = btrfs_alloc_block_rsv(fs_info, BTRFS_BLOCK_RSV_TEMP);
5342         if (!rsv) {
5343                 btrfs_orphan_del(NULL, BTRFS_I(inode));
5344                 goto no_delete;
5345         }
5346         rsv->size = min_size;
5347         rsv->failfast = 1;
5348         global_rsv = &fs_info->global_block_rsv;
5349
5350         btrfs_i_size_write(BTRFS_I(inode), 0);
5351
5352         /*
5353          * This is a bit simpler than btrfs_truncate since we've already
5354          * reserved our space for our orphan item in the unlink, so we just
5355          * need to reserve some slack space in case we add bytes and update
5356          * inode item when doing the truncate.
5357          */
5358         while (1) {
5359                 ret = btrfs_block_rsv_refill(root, rsv, min_size,
5360                                              BTRFS_RESERVE_FLUSH_LIMIT);
5361
5362                 /*
5363                  * Try and steal from the global reserve since we will
5364                  * likely not use this space anyway, we want to try as
5365                  * hard as possible to get this to work.
5366                  */
5367                 if (ret)
5368                         steal_from_global++;
5369                 else
5370                         steal_from_global = 0;
5371                 ret = 0;
5372
5373                 /*
5374                  * steal_from_global == 0: we reserved stuff, hooray!
5375                  * steal_from_global == 1: we didn't reserve stuff, boo!
5376                  * steal_from_global == 2: we've committed, still not a lot of
5377                  * room but maybe we'll have room in the global reserve this
5378                  * time.
5379                  * steal_from_global == 3: abandon all hope!
5380                  */
5381                 if (steal_from_global > 2) {
5382                         btrfs_warn(fs_info,
5383                                    "Could not get space for a delete, will truncate on mount %d",
5384                                    ret);
5385                         btrfs_orphan_del(NULL, BTRFS_I(inode));
5386                         btrfs_free_block_rsv(fs_info, rsv);
5387                         goto no_delete;
5388                 }
5389
5390                 trans = btrfs_join_transaction(root);
5391                 if (IS_ERR(trans)) {
5392                         btrfs_orphan_del(NULL, BTRFS_I(inode));
5393                         btrfs_free_block_rsv(fs_info, rsv);
5394                         goto no_delete;
5395                 }
5396
5397                 /*
5398                  * We can't just steal from the global reserve, we need to make
5399                  * sure there is room to do it, if not we need to commit and try
5400                  * again.
5401                  */
5402                 if (steal_from_global) {
5403                         if (!btrfs_check_space_for_delayed_refs(trans, fs_info))
5404                                 ret = btrfs_block_rsv_migrate(global_rsv, rsv,
5405                                                               min_size, 0);
5406                         else
5407                                 ret = -ENOSPC;
5408                 }
5409
5410                 /*
5411                  * Couldn't steal from the global reserve, we have too much
5412                  * pending stuff built up, commit the transaction and try it
5413                  * again.
5414                  */
5415                 if (ret) {
5416                         ret = btrfs_commit_transaction(trans);
5417                         if (ret) {
5418                                 btrfs_orphan_del(NULL, BTRFS_I(inode));
5419                                 btrfs_free_block_rsv(fs_info, rsv);
5420                                 goto no_delete;
5421                         }
5422                         continue;
5423                 } else {
5424                         steal_from_global = 0;
5425                 }
5426
5427                 trans->block_rsv = rsv;
5428
5429                 ret = btrfs_truncate_inode_items(trans, root, inode, 0, 0);
5430                 if (ret != -ENOSPC && ret != -EAGAIN)
5431                         break;
5432
5433                 trans->block_rsv = &fs_info->trans_block_rsv;
5434                 btrfs_end_transaction(trans);
5435                 trans = NULL;
5436                 btrfs_btree_balance_dirty(fs_info);
5437         }
5438
5439         btrfs_free_block_rsv(fs_info, rsv);
5440
5441         /*
5442          * Errors here aren't a big deal, it just means we leave orphan items
5443          * in the tree.  They will be cleaned up on the next mount.
5444          */
5445         if (ret == 0) {
5446                 trans->block_rsv = root->orphan_block_rsv;
5447                 btrfs_orphan_del(trans, BTRFS_I(inode));
5448         } else {
5449                 btrfs_orphan_del(NULL, BTRFS_I(inode));
5450         }
5451
5452         trans->block_rsv = &fs_info->trans_block_rsv;
5453         if (!(root == fs_info->tree_root ||
5454               root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID))
5455                 btrfs_return_ino(root, btrfs_ino(BTRFS_I(inode)));
5456
5457         btrfs_end_transaction(trans);
5458         btrfs_btree_balance_dirty(fs_info);
5459 no_delete:
5460         btrfs_remove_delayed_node(BTRFS_I(inode));
5461         clear_inode(inode);
5462 }
5463
5464 /*
5465  * this returns the key found in the dir entry in the location pointer.
5466  * If no dir entries were found, location->objectid is 0.
5467  */
5468 static int btrfs_inode_by_name(struct inode *dir, struct dentry *dentry,
5469                                struct btrfs_key *location)
5470 {
5471         const char *name = dentry->d_name.name;
5472         int namelen = dentry->d_name.len;
5473         struct btrfs_dir_item *di;
5474         struct btrfs_path *path;
5475         struct btrfs_root *root = BTRFS_I(dir)->root;
5476         int ret = 0;
5477
5478         path = btrfs_alloc_path();
5479         if (!path)
5480                 return -ENOMEM;
5481
5482         di = btrfs_lookup_dir_item(NULL, root, path, btrfs_ino(BTRFS_I(dir)),
5483                         name, namelen, 0);
5484         if (IS_ERR(di))
5485                 ret = PTR_ERR(di);
5486
5487         if (IS_ERR_OR_NULL(di))
5488                 goto out_err;
5489
5490         btrfs_dir_item_key_to_cpu(path->nodes[0], di, location);
5491 out:
5492         btrfs_free_path(path);
5493         return ret;
5494 out_err:
5495         location->objectid = 0;
5496         goto out;
5497 }
5498
5499 /*
5500  * when we hit a tree root in a directory, the btrfs part of the inode
5501  * needs to be changed to reflect the root directory of the tree root.  This
5502  * is kind of like crossing a mount point.
5503  */
5504 static int fixup_tree_root_location(struct btrfs_fs_info *fs_info,
5505                                     struct inode *dir,
5506                                     struct dentry *dentry,
5507                                     struct btrfs_key *location,
5508                                     struct btrfs_root **sub_root)
5509 {
5510         struct btrfs_path *path;
5511         struct btrfs_root *new_root;
5512         struct btrfs_root_ref *ref;
5513         struct extent_buffer *leaf;
5514         struct btrfs_key key;
5515         int ret;
5516         int err = 0;
5517
5518         path = btrfs_alloc_path();
5519         if (!path) {
5520                 err = -ENOMEM;
5521                 goto out;
5522         }
5523
5524         err = -ENOENT;
5525         key.objectid = BTRFS_I(dir)->root->root_key.objectid;
5526         key.type = BTRFS_ROOT_REF_KEY;
5527         key.offset = location->objectid;
5528
5529         ret = btrfs_search_slot(NULL, fs_info->tree_root, &key, path, 0, 0);
5530         if (ret) {
5531                 if (ret < 0)
5532                         err = ret;
5533                 goto out;
5534         }
5535
5536         leaf = path->nodes[0];
5537         ref = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_root_ref);
5538         if (btrfs_root_ref_dirid(leaf, ref) != btrfs_ino(BTRFS_I(dir)) ||
5539             btrfs_root_ref_name_len(leaf, ref) != dentry->d_name.len)
5540                 goto out;
5541
5542         ret = memcmp_extent_buffer(leaf, dentry->d_name.name,
5543                                    (unsigned long)(ref + 1),
5544                                    dentry->d_name.len);
5545         if (ret)
5546                 goto out;
5547
5548         btrfs_release_path(path);
5549
5550         new_root = btrfs_read_fs_root_no_name(fs_info, location);
5551         if (IS_ERR(new_root)) {
5552                 err = PTR_ERR(new_root);
5553                 goto out;
5554         }
5555
5556         *sub_root = new_root;
5557         location->objectid = btrfs_root_dirid(&new_root->root_item);
5558         location->type = BTRFS_INODE_ITEM_KEY;
5559         location->offset = 0;
5560         err = 0;
5561 out:
5562         btrfs_free_path(path);
5563         return err;
5564 }
5565
5566 static void inode_tree_add(struct inode *inode)
5567 {
5568         struct btrfs_root *root = BTRFS_I(inode)->root;
5569         struct btrfs_inode *entry;
5570         struct rb_node **p;
5571         struct rb_node *parent;
5572         struct rb_node *new = &BTRFS_I(inode)->rb_node;
5573         u64 ino = btrfs_ino(BTRFS_I(inode));
5574
5575         if (inode_unhashed(inode))
5576                 return;
5577         parent = NULL;
5578         spin_lock(&root->inode_lock);
5579         p = &root->inode_tree.rb_node;
5580         while (*p) {
5581                 parent = *p;
5582                 entry = rb_entry(parent, struct btrfs_inode, rb_node);
5583
5584                 if (ino < btrfs_ino(BTRFS_I(&entry->vfs_inode)))
5585                         p = &parent->rb_left;
5586                 else if (ino > btrfs_ino(BTRFS_I(&entry->vfs_inode)))
5587                         p = &parent->rb_right;
5588                 else {
5589                         WARN_ON(!(entry->vfs_inode.i_state &
5590                                   (I_WILL_FREE | I_FREEING)));
5591                         rb_replace_node(parent, new, &root->inode_tree);
5592                         RB_CLEAR_NODE(parent);
5593                         spin_unlock(&root->inode_lock);
5594                         return;
5595                 }
5596         }
5597         rb_link_node(new, parent, p);
5598         rb_insert_color(new, &root->inode_tree);
5599         spin_unlock(&root->inode_lock);
5600 }
5601
5602 static void inode_tree_del(struct inode *inode)
5603 {
5604         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
5605         struct btrfs_root *root = BTRFS_I(inode)->root;
5606         int empty = 0;
5607
5608         spin_lock(&root->inode_lock);
5609         if (!RB_EMPTY_NODE(&BTRFS_I(inode)->rb_node)) {
5610                 rb_erase(&BTRFS_I(inode)->rb_node, &root->inode_tree);
5611                 RB_CLEAR_NODE(&BTRFS_I(inode)->rb_node);
5612                 empty = RB_EMPTY_ROOT(&root->inode_tree);
5613         }
5614         spin_unlock(&root->inode_lock);
5615
5616         if (empty && btrfs_root_refs(&root->root_item) == 0) {
5617                 synchronize_srcu(&fs_info->subvol_srcu);
5618                 spin_lock(&root->inode_lock);
5619                 empty = RB_EMPTY_ROOT(&root->inode_tree);
5620                 spin_unlock(&root->inode_lock);
5621                 if (empty)
5622                         btrfs_add_dead_root(root);
5623         }
5624 }
5625
5626 void btrfs_invalidate_inodes(struct btrfs_root *root)
5627 {
5628         struct btrfs_fs_info *fs_info = root->fs_info;
5629         struct rb_node *node;
5630         struct rb_node *prev;
5631         struct btrfs_inode *entry;
5632         struct inode *inode;
5633         u64 objectid = 0;
5634
5635         if (!test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state))
5636                 WARN_ON(btrfs_root_refs(&root->root_item) != 0);
5637
5638         spin_lock(&root->inode_lock);
5639 again:
5640         node = root->inode_tree.rb_node;
5641         prev = NULL;
5642         while (node) {
5643                 prev = node;
5644                 entry = rb_entry(node, struct btrfs_inode, rb_node);
5645
5646                 if (objectid < btrfs_ino(BTRFS_I(&entry->vfs_inode)))
5647                         node = node->rb_left;
5648                 else if (objectid > btrfs_ino(BTRFS_I(&entry->vfs_inode)))
5649                         node = node->rb_right;
5650                 else
5651                         break;
5652         }
5653         if (!node) {
5654                 while (prev) {
5655                         entry = rb_entry(prev, struct btrfs_inode, rb_node);
5656                         if (objectid <= btrfs_ino(BTRFS_I(&entry->vfs_inode))) {
5657                                 node = prev;
5658                                 break;
5659                         }
5660                         prev = rb_next(prev);
5661                 }
5662         }
5663         while (node) {
5664                 entry = rb_entry(node, struct btrfs_inode, rb_node);
5665                 objectid = btrfs_ino(BTRFS_I(&entry->vfs_inode)) + 1;
5666                 inode = igrab(&entry->vfs_inode);
5667                 if (inode) {
5668                         spin_unlock(&root->inode_lock);
5669                         if (atomic_read(&inode->i_count) > 1)
5670                                 d_prune_aliases(inode);
5671                         /*
5672                          * btrfs_drop_inode will have it removed from
5673                          * the inode cache when its usage count
5674                          * hits zero.
5675                          */
5676                         iput(inode);
5677                         cond_resched();
5678                         spin_lock(&root->inode_lock);
5679                         goto again;
5680                 }
5681
5682                 if (cond_resched_lock(&root->inode_lock))
5683                         goto again;
5684
5685                 node = rb_next(node);
5686         }
5687         spin_unlock(&root->inode_lock);
5688 }
5689
5690 static int btrfs_init_locked_inode(struct inode *inode, void *p)
5691 {
5692         struct btrfs_iget_args *args = p;
5693         inode->i_ino = args->location->objectid;
5694         memcpy(&BTRFS_I(inode)->location, args->location,
5695                sizeof(*args->location));
5696         BTRFS_I(inode)->root = args->root;
5697         return 0;
5698 }
5699
5700 static int btrfs_find_actor(struct inode *inode, void *opaque)
5701 {
5702         struct btrfs_iget_args *args = opaque;
5703         return args->location->objectid == BTRFS_I(inode)->location.objectid &&
5704                 args->root == BTRFS_I(inode)->root;
5705 }
5706
5707 static struct inode *btrfs_iget_locked(struct super_block *s,
5708                                        struct btrfs_key *location,
5709                                        struct btrfs_root *root)
5710 {
5711         struct inode *inode;
5712         struct btrfs_iget_args args;
5713         unsigned long hashval = btrfs_inode_hash(location->objectid, root);
5714
5715         args.location = location;
5716         args.root = root;
5717
5718         inode = iget5_locked(s, hashval, btrfs_find_actor,
5719                              btrfs_init_locked_inode,
5720                              (void *)&args);
5721         return inode;
5722 }
5723
5724 /* Get an inode object given its location and corresponding root.
5725  * Returns in *is_new if the inode was read from disk
5726  */
5727 struct inode *btrfs_iget(struct super_block *s, struct btrfs_key *location,
5728                          struct btrfs_root *root, int *new)
5729 {
5730         struct inode *inode;
5731
5732         inode = btrfs_iget_locked(s, location, root);
5733         if (!inode)
5734                 return ERR_PTR(-ENOMEM);
5735
5736         if (inode->i_state & I_NEW) {
5737                 int ret;
5738
5739                 ret = btrfs_read_locked_inode(inode);
5740                 if (!is_bad_inode(inode)) {
5741                         inode_tree_add(inode);
5742                         unlock_new_inode(inode);
5743                         if (new)
5744                                 *new = 1;
5745                 } else {
5746                         unlock_new_inode(inode);
5747                         iput(inode);
5748                         ASSERT(ret < 0);
5749                         inode = ERR_PTR(ret < 0 ? ret : -ESTALE);
5750                 }
5751         }
5752
5753         return inode;
5754 }
5755
5756 static struct inode *new_simple_dir(struct super_block *s,
5757                                     struct btrfs_key *key,
5758                                     struct btrfs_root *root)
5759 {
5760         struct inode *inode = new_inode(s);
5761
5762         if (!inode)
5763                 return ERR_PTR(-ENOMEM);
5764
5765         BTRFS_I(inode)->root = root;
5766         memcpy(&BTRFS_I(inode)->location, key, sizeof(*key));
5767         set_bit(BTRFS_INODE_DUMMY, &BTRFS_I(inode)->runtime_flags);
5768
5769         inode->i_ino = BTRFS_EMPTY_SUBVOL_DIR_OBJECTID;
5770         inode->i_op = &btrfs_dir_ro_inode_operations;
5771         inode->i_opflags &= ~IOP_XATTR;
5772         inode->i_fop = &simple_dir_operations;
5773         inode->i_mode = S_IFDIR | S_IRUGO | S_IWUSR | S_IXUGO;
5774         inode->i_mtime = current_time(inode);
5775         inode->i_atime = inode->i_mtime;
5776         inode->i_ctime = inode->i_mtime;
5777         BTRFS_I(inode)->i_otime = inode->i_mtime;
5778
5779         return inode;
5780 }
5781
5782 struct inode *btrfs_lookup_dentry(struct inode *dir, struct dentry *dentry)
5783 {
5784         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
5785         struct inode *inode;
5786         struct btrfs_root *root = BTRFS_I(dir)->root;
5787         struct btrfs_root *sub_root = root;
5788         struct btrfs_key location;
5789         int index;
5790         int ret = 0;
5791
5792         if (dentry->d_name.len > BTRFS_NAME_LEN)
5793                 return ERR_PTR(-ENAMETOOLONG);
5794
5795         ret = btrfs_inode_by_name(dir, dentry, &location);
5796         if (ret < 0)
5797                 return ERR_PTR(ret);
5798
5799         if (location.objectid == 0)
5800                 return ERR_PTR(-ENOENT);
5801
5802         if (location.type == BTRFS_INODE_ITEM_KEY) {
5803                 inode = btrfs_iget(dir->i_sb, &location, root, NULL);
5804                 return inode;
5805         }
5806
5807         BUG_ON(location.type != BTRFS_ROOT_ITEM_KEY);
5808
5809         index = srcu_read_lock(&fs_info->subvol_srcu);
5810         ret = fixup_tree_root_location(fs_info, dir, dentry,
5811                                        &location, &sub_root);
5812         if (ret < 0) {
5813                 if (ret != -ENOENT)
5814                         inode = ERR_PTR(ret);
5815                 else
5816                         inode = new_simple_dir(dir->i_sb, &location, sub_root);
5817         } else {
5818                 inode = btrfs_iget(dir->i_sb, &location, sub_root, NULL);
5819         }
5820         srcu_read_unlock(&fs_info->subvol_srcu, index);
5821
5822         if (!IS_ERR(inode) && root != sub_root) {
5823                 down_read(&fs_info->cleanup_work_sem);
5824                 if (!(inode->i_sb->s_flags & MS_RDONLY))
5825                         ret = btrfs_orphan_cleanup(sub_root);
5826                 up_read(&fs_info->cleanup_work_sem);
5827                 if (ret) {
5828                         iput(inode);
5829                         inode = ERR_PTR(ret);
5830                 }
5831         }
5832
5833         return inode;
5834 }
5835
5836 static int btrfs_dentry_delete(const struct dentry *dentry)
5837 {
5838         struct btrfs_root *root;
5839         struct inode *inode = d_inode(dentry);
5840
5841         if (!inode && !IS_ROOT(dentry))
5842                 inode = d_inode(dentry->d_parent);
5843
5844         if (inode) {
5845                 root = BTRFS_I(inode)->root;
5846                 if (btrfs_root_refs(&root->root_item) == 0)
5847                         return 1;
5848
5849                 if (btrfs_ino(BTRFS_I(inode)) == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)
5850                         return 1;
5851         }
5852         return 0;
5853 }
5854
5855 static void btrfs_dentry_release(struct dentry *dentry)
5856 {
5857         kfree(dentry->d_fsdata);
5858 }
5859
5860 static struct dentry *btrfs_lookup(struct inode *dir, struct dentry *dentry,
5861                                    unsigned int flags)
5862 {
5863         struct inode *inode;
5864
5865         inode = btrfs_lookup_dentry(dir, dentry);
5866         if (IS_ERR(inode)) {
5867                 if (PTR_ERR(inode) == -ENOENT)
5868                         inode = NULL;
5869                 else
5870                         return ERR_CAST(inode);
5871         }
5872
5873         return d_splice_alias(inode, dentry);
5874 }
5875
5876 unsigned char btrfs_filetype_table[] = {
5877         DT_UNKNOWN, DT_REG, DT_DIR, DT_CHR, DT_BLK, DT_FIFO, DT_SOCK, DT_LNK
5878 };
5879
5880 /*
5881  * All this infrastructure exists because dir_emit can fault, and we are holding
5882  * the tree lock when doing readdir.  For now just allocate a buffer and copy
5883  * our information into that, and then dir_emit from the buffer.  This is
5884  * similar to what NFS does, only we don't keep the buffer around in pagecache
5885  * because I'm afraid I'll mess that up.  Long term we need to make filldir do
5886  * copy_to_user_inatomic so we don't have to worry about page faulting under the
5887  * tree lock.
5888  */
5889 static int btrfs_opendir(struct inode *inode, struct file *file)
5890 {
5891         struct btrfs_file_private *private;
5892
5893         private = kzalloc(sizeof(struct btrfs_file_private), GFP_KERNEL);
5894         if (!private)
5895                 return -ENOMEM;
5896         private->filldir_buf = kzalloc(PAGE_SIZE, GFP_KERNEL);
5897         if (!private->filldir_buf) {
5898                 kfree(private);
5899                 return -ENOMEM;
5900         }
5901         file->private_data = private;
5902         return 0;
5903 }
5904
5905 struct dir_entry {
5906         u64 ino;
5907         u64 offset;
5908         unsigned type;
5909         int name_len;
5910 };
5911
5912 static int btrfs_filldir(void *addr, int entries, struct dir_context *ctx)
5913 {
5914         while (entries--) {
5915                 struct dir_entry *entry = addr;
5916                 char *name = (char *)(entry + 1);
5917
5918                 ctx->pos = entry->offset;
5919                 if (!dir_emit(ctx, name, entry->name_len, entry->ino,
5920                               entry->type))
5921                         return 1;
5922                 addr += sizeof(struct dir_entry) + entry->name_len;
5923                 ctx->pos++;
5924         }
5925         return 0;
5926 }
5927
5928 static int btrfs_real_readdir(struct file *file, struct dir_context *ctx)
5929 {
5930         struct inode *inode = file_inode(file);
5931         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
5932         struct btrfs_root *root = BTRFS_I(inode)->root;
5933         struct btrfs_file_private *private = file->private_data;
5934         struct btrfs_dir_item *di;
5935         struct btrfs_key key;
5936         struct btrfs_key found_key;
5937         struct btrfs_path *path;
5938         void *addr;
5939         struct list_head ins_list;
5940         struct list_head del_list;
5941         int ret;
5942         struct extent_buffer *leaf;
5943         int slot;
5944         char *name_ptr;
5945         int name_len;
5946         int entries = 0;
5947         int total_len = 0;
5948         bool put = false;
5949         struct btrfs_key location;
5950
5951         if (!dir_emit_dots(file, ctx))
5952                 return 0;
5953
5954         path = btrfs_alloc_path();
5955         if (!path)
5956                 return -ENOMEM;
5957
5958         addr = private->filldir_buf;
5959         path->reada = READA_FORWARD;
5960
5961         INIT_LIST_HEAD(&ins_list);
5962         INIT_LIST_HEAD(&del_list);
5963         put = btrfs_readdir_get_delayed_items(inode, &ins_list, &del_list);
5964
5965 again:
5966         key.type = BTRFS_DIR_INDEX_KEY;
5967         key.offset = ctx->pos;
5968         key.objectid = btrfs_ino(BTRFS_I(inode));
5969
5970         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
5971         if (ret < 0)
5972                 goto err;
5973
5974         while (1) {
5975                 struct dir_entry *entry;
5976
5977                 leaf = path->nodes[0];
5978                 slot = path->slots[0];
5979                 if (slot >= btrfs_header_nritems(leaf)) {
5980                         ret = btrfs_next_leaf(root, path);
5981                         if (ret < 0)
5982                                 goto err;
5983                         else if (ret > 0)
5984                                 break;
5985                         continue;
5986                 }
5987
5988                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
5989
5990                 if (found_key.objectid != key.objectid)
5991                         break;
5992                 if (found_key.type != BTRFS_DIR_INDEX_KEY)
5993                         break;
5994                 if (found_key.offset < ctx->pos)
5995                         goto next;
5996                 if (btrfs_should_delete_dir_index(&del_list, found_key.offset))
5997                         goto next;
5998                 di = btrfs_item_ptr(leaf, slot, struct btrfs_dir_item);
5999                 if (verify_dir_item(fs_info, leaf, slot, di))
6000                         goto next;
6001
6002                 name_len = btrfs_dir_name_len(leaf, di);
6003                 if ((total_len + sizeof(struct dir_entry) + name_len) >=
6004                     PAGE_SIZE) {
6005                         btrfs_release_path(path);
6006                         ret = btrfs_filldir(private->filldir_buf, entries, ctx);
6007                         if (ret)
6008                                 goto nopos;
6009                         addr = private->filldir_buf;
6010                         entries = 0;
6011                         total_len = 0;
6012                         goto again;
6013                 }
6014
6015                 entry = addr;
6016                 entry->name_len = name_len;
6017                 name_ptr = (char *)(entry + 1);
6018                 read_extent_buffer(leaf, name_ptr, (unsigned long)(di + 1),
6019                                    name_len);
6020                 entry->type = btrfs_filetype_table[btrfs_dir_type(leaf, di)];
6021                 btrfs_dir_item_key_to_cpu(leaf, di, &location);
6022                 entry->ino = location.objectid;
6023                 entry->offset = found_key.offset;
6024                 entries++;
6025                 addr += sizeof(struct dir_entry) + name_len;
6026                 total_len += sizeof(struct dir_entry) + name_len;
6027 next:
6028                 path->slots[0]++;
6029         }
6030         btrfs_release_path(path);
6031
6032         ret = btrfs_filldir(private->filldir_buf, entries, ctx);
6033         if (ret)
6034                 goto nopos;
6035
6036         ret = btrfs_readdir_delayed_dir_index(ctx, &ins_list);
6037         if (ret)
6038                 goto nopos;
6039
6040         /*
6041          * Stop new entries from being returned after we return the last
6042          * entry.
6043          *
6044          * New directory entries are assigned a strictly increasing
6045          * offset.  This means that new entries created during readdir
6046          * are *guaranteed* to be seen in the future by that readdir.
6047          * This has broken buggy programs which operate on names as
6048          * they're returned by readdir.  Until we re-use freed offsets
6049          * we have this hack to stop new entries from being returned
6050          * under the assumption that they'll never reach this huge
6051          * offset.
6052          *
6053          * This is being careful not to overflow 32bit loff_t unless the
6054          * last entry requires it because doing so has broken 32bit apps
6055          * in the past.
6056          */
6057         if (ctx->pos >= INT_MAX)
6058                 ctx->pos = LLONG_MAX;
6059         else
6060                 ctx->pos = INT_MAX;
6061 nopos:
6062         ret = 0;
6063 err:
6064         if (put)
6065                 btrfs_readdir_put_delayed_items(inode, &ins_list, &del_list);
6066         btrfs_free_path(path);
6067         return ret;
6068 }
6069
6070 int btrfs_write_inode(struct inode *inode, struct writeback_control *wbc)
6071 {
6072         struct btrfs_root *root = BTRFS_I(inode)->root;
6073         struct btrfs_trans_handle *trans;
6074         int ret = 0;
6075         bool nolock = false;
6076
6077         if (test_bit(BTRFS_INODE_DUMMY, &BTRFS_I(inode)->runtime_flags))
6078                 return 0;
6079
6080         if (btrfs_fs_closing(root->fs_info) &&
6081                         btrfs_is_free_space_inode(BTRFS_I(inode)))
6082                 nolock = true;
6083
6084         if (wbc->sync_mode == WB_SYNC_ALL) {
6085                 if (nolock)
6086                         trans = btrfs_join_transaction_nolock(root);
6087                 else
6088                         trans = btrfs_join_transaction(root);
6089                 if (IS_ERR(trans))
6090                         return PTR_ERR(trans);
6091                 ret = btrfs_commit_transaction(trans);
6092         }
6093         return ret;
6094 }
6095
6096 /*
6097  * This is somewhat expensive, updating the tree every time the
6098  * inode changes.  But, it is most likely to find the inode in cache.
6099  * FIXME, needs more benchmarking...there are no reasons other than performance
6100  * to keep or drop this code.
6101  */
6102 static int btrfs_dirty_inode(struct inode *inode)
6103 {
6104         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
6105         struct btrfs_root *root = BTRFS_I(inode)->root;
6106         struct btrfs_trans_handle *trans;
6107         int ret;
6108
6109         if (test_bit(BTRFS_INODE_DUMMY, &BTRFS_I(inode)->runtime_flags))
6110                 return 0;
6111
6112         trans = btrfs_join_transaction(root);
6113         if (IS_ERR(trans))
6114                 return PTR_ERR(trans);
6115
6116         ret = btrfs_update_inode(trans, root, inode);
6117         if (ret && ret == -ENOSPC) {
6118                 /* whoops, lets try again with the full transaction */
6119                 btrfs_end_transaction(trans);
6120                 trans = btrfs_start_transaction(root, 1);
6121                 if (IS_ERR(trans))
6122                         return PTR_ERR(trans);
6123
6124                 ret = btrfs_update_inode(trans, root, inode);
6125         }
6126         btrfs_end_transaction(trans);
6127         if (BTRFS_I(inode)->delayed_node)
6128                 btrfs_balance_delayed_items(fs_info);
6129
6130         return ret;
6131 }
6132
6133 /*
6134  * This is a copy of file_update_time.  We need this so we can return error on
6135  * ENOSPC for updating the inode in the case of file write and mmap writes.
6136  */
6137 static int btrfs_update_time(struct inode *inode, struct timespec *now,
6138                              int flags)
6139 {
6140         struct btrfs_root *root = BTRFS_I(inode)->root;
6141
6142         if (btrfs_root_readonly(root))
6143                 return -EROFS;
6144
6145         if (flags & S_VERSION)
6146                 inode_inc_iversion(inode);
6147         if (flags & S_CTIME)
6148                 inode->i_ctime = *now;
6149         if (flags & S_MTIME)
6150                 inode->i_mtime = *now;
6151         if (flags & S_ATIME)
6152                 inode->i_atime = *now;
6153         return btrfs_dirty_inode(inode);
6154 }
6155
6156 /*
6157  * find the highest existing sequence number in a directory
6158  * and then set the in-memory index_cnt variable to reflect
6159  * free sequence numbers
6160  */
6161 static int btrfs_set_inode_index_count(struct btrfs_inode *inode)
6162 {
6163         struct btrfs_root *root = inode->root;
6164         struct btrfs_key key, found_key;
6165         struct btrfs_path *path;
6166         struct extent_buffer *leaf;
6167         int ret;
6168
6169         key.objectid = btrfs_ino(inode);
6170         key.type = BTRFS_DIR_INDEX_KEY;
6171         key.offset = (u64)-1;
6172
6173         path = btrfs_alloc_path();
6174         if (!path)
6175                 return -ENOMEM;
6176
6177         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
6178         if (ret < 0)
6179                 goto out;
6180         /* FIXME: we should be able to handle this */
6181         if (ret == 0)
6182                 goto out;
6183         ret = 0;
6184
6185         /*
6186          * MAGIC NUMBER EXPLANATION:
6187          * since we search a directory based on f_pos we have to start at 2
6188          * since '.' and '..' have f_pos of 0 and 1 respectively, so everybody
6189          * else has to start at 2
6190          */
6191         if (path->slots[0] == 0) {
6192                 inode->index_cnt = 2;
6193                 goto out;
6194         }
6195
6196         path->slots[0]--;
6197
6198         leaf = path->nodes[0];
6199         btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
6200
6201         if (found_key.objectid != btrfs_ino(inode) ||
6202             found_key.type != BTRFS_DIR_INDEX_KEY) {
6203                 inode->index_cnt = 2;
6204                 goto out;
6205         }
6206
6207         inode->index_cnt = found_key.offset + 1;
6208 out:
6209         btrfs_free_path(path);
6210         return ret;
6211 }
6212
6213 /*
6214  * helper to find a free sequence number in a given directory.  This current
6215  * code is very simple, later versions will do smarter things in the btree
6216  */
6217 int btrfs_set_inode_index(struct btrfs_inode *dir, u64 *index)
6218 {
6219         int ret = 0;
6220
6221         if (dir->index_cnt == (u64)-1) {
6222                 ret = btrfs_inode_delayed_dir_index_count(dir);
6223                 if (ret) {
6224                         ret = btrfs_set_inode_index_count(dir);
6225                         if (ret)
6226                                 return ret;
6227                 }
6228         }
6229
6230         *index = dir->index_cnt;
6231         dir->index_cnt++;
6232
6233         return ret;
6234 }
6235
6236 static int btrfs_insert_inode_locked(struct inode *inode)
6237 {
6238         struct btrfs_iget_args args;
6239         args.location = &BTRFS_I(inode)->location;
6240         args.root = BTRFS_I(inode)->root;
6241
6242         return insert_inode_locked4(inode,
6243                    btrfs_inode_hash(inode->i_ino, BTRFS_I(inode)->root),
6244                    btrfs_find_actor, &args);
6245 }
6246
6247 /*
6248  * Inherit flags from the parent inode.
6249  *
6250  * Currently only the compression flags and the cow flags are inherited.
6251  */
6252 static void btrfs_inherit_iflags(struct inode *inode, struct inode *dir)
6253 {
6254         unsigned int flags;
6255
6256         if (!dir)
6257                 return;
6258
6259         flags = BTRFS_I(dir)->flags;
6260
6261         if (flags & BTRFS_INODE_NOCOMPRESS) {
6262                 BTRFS_I(inode)->flags &= ~BTRFS_INODE_COMPRESS;
6263                 BTRFS_I(inode)->flags |= BTRFS_INODE_NOCOMPRESS;
6264         } else if (flags & BTRFS_INODE_COMPRESS) {
6265                 BTRFS_I(inode)->flags &= ~BTRFS_INODE_NOCOMPRESS;
6266                 BTRFS_I(inode)->flags |= BTRFS_INODE_COMPRESS;
6267         }
6268
6269         if (flags & BTRFS_INODE_NODATACOW) {
6270                 BTRFS_I(inode)->flags |= BTRFS_INODE_NODATACOW;
6271                 if (S_ISREG(inode->i_mode))
6272                         BTRFS_I(inode)->flags |= BTRFS_INODE_NODATASUM;
6273         }
6274
6275         btrfs_update_iflags(inode);
6276 }
6277
6278 static struct inode *btrfs_new_inode(struct btrfs_trans_handle *trans,
6279                                      struct btrfs_root *root,
6280                                      struct inode *dir,
6281                                      const char *name, int name_len,
6282                                      u64 ref_objectid, u64 objectid,
6283                                      umode_t mode, u64 *index)
6284 {
6285         struct btrfs_fs_info *fs_info = root->fs_info;
6286         struct inode *inode;
6287         struct btrfs_inode_item *inode_item;
6288         struct btrfs_key *location;
6289         struct btrfs_path *path;
6290         struct btrfs_inode_ref *ref;
6291         struct btrfs_key key[2];
6292         u32 sizes[2];
6293         int nitems = name ? 2 : 1;
6294         unsigned long ptr;
6295         int ret;
6296
6297         path = btrfs_alloc_path();
6298         if (!path)
6299                 return ERR_PTR(-ENOMEM);
6300
6301         inode = new_inode(fs_info->sb);
6302         if (!inode) {
6303                 btrfs_free_path(path);
6304                 return ERR_PTR(-ENOMEM);
6305         }
6306
6307         /*
6308          * O_TMPFILE, set link count to 0, so that after this point,
6309          * we fill in an inode item with the correct link count.
6310          */
6311         if (!name)
6312                 set_nlink(inode, 0);
6313
6314         /*
6315          * we have to initialize this early, so we can reclaim the inode
6316          * number if we fail afterwards in this function.
6317          */
6318         inode->i_ino = objectid;
6319
6320         if (dir && name) {
6321                 trace_btrfs_inode_request(dir);
6322
6323                 ret = btrfs_set_inode_index(BTRFS_I(dir), index);
6324                 if (ret) {
6325                         btrfs_free_path(path);
6326                         iput(inode);
6327                         return ERR_PTR(ret);
6328                 }
6329         } else if (dir) {
6330                 *index = 0;
6331         }
6332         /*
6333          * index_cnt is ignored for everything but a dir,
6334          * btrfs_get_inode_index_count has an explanation for the magic
6335          * number
6336          */
6337         BTRFS_I(inode)->index_cnt = 2;
6338         BTRFS_I(inode)->dir_index = *index;
6339         BTRFS_I(inode)->root = root;
6340         BTRFS_I(inode)->generation = trans->transid;
6341         inode->i_generation = BTRFS_I(inode)->generation;
6342
6343         /*
6344          * We could have gotten an inode number from somebody who was fsynced
6345          * and then removed in this same transaction, so let's just set full
6346          * sync since it will be a full sync anyway and this will blow away the
6347          * old info in the log.
6348          */
6349         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &BTRFS_I(inode)->runtime_flags);
6350
6351         key[0].objectid = objectid;
6352         key[0].type = BTRFS_INODE_ITEM_KEY;
6353         key[0].offset = 0;
6354
6355         sizes[0] = sizeof(struct btrfs_inode_item);
6356
6357         if (name) {
6358                 /*
6359                  * Start new inodes with an inode_ref. This is slightly more
6360                  * efficient for small numbers of hard links since they will
6361                  * be packed into one item. Extended refs will kick in if we
6362                  * add more hard links than can fit in the ref item.
6363                  */
6364                 key[1].objectid = objectid;
6365                 key[1].type = BTRFS_INODE_REF_KEY;
6366                 key[1].offset = ref_objectid;
6367
6368                 sizes[1] = name_len + sizeof(*ref);
6369         }
6370
6371         location = &BTRFS_I(inode)->location;
6372         location->objectid = objectid;
6373         location->offset = 0;
6374         location->type = BTRFS_INODE_ITEM_KEY;
6375
6376         ret = btrfs_insert_inode_locked(inode);
6377         if (ret < 0)
6378                 goto fail;
6379
6380         path->leave_spinning = 1;
6381         ret = btrfs_insert_empty_items(trans, root, path, key, sizes, nitems);
6382         if (ret != 0)
6383                 goto fail_unlock;
6384
6385         inode_init_owner(inode, dir, mode);
6386         inode_set_bytes(inode, 0);
6387
6388         inode->i_mtime = current_time(inode);
6389         inode->i_atime = inode->i_mtime;
6390         inode->i_ctime = inode->i_mtime;
6391         BTRFS_I(inode)->i_otime = inode->i_mtime;
6392
6393         inode_item = btrfs_item_ptr(path->nodes[0], path->slots[0],
6394                                   struct btrfs_inode_item);
6395         memzero_extent_buffer(path->nodes[0], (unsigned long)inode_item,
6396                              sizeof(*inode_item));
6397         fill_inode_item(trans, path->nodes[0], inode_item, inode);
6398
6399         if (name) {
6400                 ref = btrfs_item_ptr(path->nodes[0], path->slots[0] + 1,
6401                                      struct btrfs_inode_ref);
6402                 btrfs_set_inode_ref_name_len(path->nodes[0], ref, name_len);
6403                 btrfs_set_inode_ref_index(path->nodes[0], ref, *index);
6404                 ptr = (unsigned long)(ref + 1);
6405                 write_extent_buffer(path->nodes[0], name, ptr, name_len);
6406         }
6407
6408         btrfs_mark_buffer_dirty(path->nodes[0]);
6409         btrfs_free_path(path);
6410
6411         btrfs_inherit_iflags(inode, dir);
6412
6413         if (S_ISREG(mode)) {
6414                 if (btrfs_test_opt(fs_info, NODATASUM))
6415                         BTRFS_I(inode)->flags |= BTRFS_INODE_NODATASUM;
6416                 if (btrfs_test_opt(fs_info, NODATACOW))
6417                         BTRFS_I(inode)->flags |= BTRFS_INODE_NODATACOW |
6418                                 BTRFS_INODE_NODATASUM;
6419         }
6420
6421         inode_tree_add(inode);
6422
6423         trace_btrfs_inode_new(inode);
6424         btrfs_set_inode_last_trans(trans, inode);
6425
6426         btrfs_update_root_times(trans, root);
6427
6428         ret = btrfs_inode_inherit_props(trans, inode, dir);
6429         if (ret)
6430                 btrfs_err(fs_info,
6431                           "error inheriting props for ino %llu (root %llu): %d",
6432                         btrfs_ino(BTRFS_I(inode)), root->root_key.objectid, ret);
6433
6434         return inode;
6435
6436 fail_unlock:
6437         unlock_new_inode(inode);
6438 fail:
6439         if (dir && name)
6440                 BTRFS_I(dir)->index_cnt--;
6441         btrfs_free_path(path);
6442         iput(inode);
6443         return ERR_PTR(ret);
6444 }
6445
6446 static inline u8 btrfs_inode_type(struct inode *inode)
6447 {
6448         return btrfs_type_by_mode[(inode->i_mode & S_IFMT) >> S_SHIFT];
6449 }
6450
6451 /*
6452  * utility function to add 'inode' into 'parent_inode' with
6453  * a give name and a given sequence number.
6454  * if 'add_backref' is true, also insert a backref from the
6455  * inode to the parent directory.
6456  */
6457 int btrfs_add_link(struct btrfs_trans_handle *trans,
6458                    struct btrfs_inode *parent_inode, struct btrfs_inode *inode,
6459                    const char *name, int name_len, int add_backref, u64 index)
6460 {
6461         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
6462         int ret = 0;
6463         struct btrfs_key key;
6464         struct btrfs_root *root = parent_inode->root;
6465         u64 ino = btrfs_ino(inode);
6466         u64 parent_ino = btrfs_ino(parent_inode);
6467
6468         if (unlikely(ino == BTRFS_FIRST_FREE_OBJECTID)) {
6469                 memcpy(&key, &inode->root->root_key, sizeof(key));
6470         } else {
6471                 key.objectid = ino;
6472                 key.type = BTRFS_INODE_ITEM_KEY;
6473                 key.offset = 0;
6474         }
6475
6476         if (unlikely(ino == BTRFS_FIRST_FREE_OBJECTID)) {
6477                 ret = btrfs_add_root_ref(trans, fs_info, key.objectid,
6478                                          root->root_key.objectid, parent_ino,
6479                                          index, name, name_len);
6480         } else if (add_backref) {
6481                 ret = btrfs_insert_inode_ref(trans, root, name, name_len, ino,
6482                                              parent_ino, index);
6483         }
6484
6485         /* Nothing to clean up yet */
6486         if (ret)
6487                 return ret;
6488
6489         ret = btrfs_insert_dir_item(trans, root, name, name_len,
6490                                     parent_inode, &key,
6491                                     btrfs_inode_type(&inode->vfs_inode), index);
6492         if (ret == -EEXIST || ret == -EOVERFLOW)
6493                 goto fail_dir_item;
6494         else if (ret) {
6495                 btrfs_abort_transaction(trans, ret);
6496                 return ret;
6497         }
6498
6499         btrfs_i_size_write(parent_inode, parent_inode->vfs_inode.i_size +
6500                            name_len * 2);
6501         inode_inc_iversion(&parent_inode->vfs_inode);
6502         parent_inode->vfs_inode.i_mtime = parent_inode->vfs_inode.i_ctime =
6503                 current_time(&parent_inode->vfs_inode);
6504         ret = btrfs_update_inode(trans, root, &parent_inode->vfs_inode);
6505         if (ret)
6506                 btrfs_abort_transaction(trans, ret);
6507         return ret;
6508
6509 fail_dir_item:
6510         if (unlikely(ino == BTRFS_FIRST_FREE_OBJECTID)) {
6511                 u64 local_index;
6512                 int err;
6513                 err = btrfs_del_root_ref(trans, fs_info, key.objectid,
6514                                          root->root_key.objectid, parent_ino,
6515                                          &local_index, name, name_len);
6516
6517         } else if (add_backref) {
6518                 u64 local_index;
6519                 int err;
6520
6521                 err = btrfs_del_inode_ref(trans, root, name, name_len,
6522                                           ino, parent_ino, &local_index);
6523         }
6524         return ret;
6525 }
6526
6527 static int btrfs_add_nondir(struct btrfs_trans_handle *trans,
6528                             struct btrfs_inode *dir, struct dentry *dentry,
6529                             struct btrfs_inode *inode, int backref, u64 index)
6530 {
6531         int err = btrfs_add_link(trans, dir, inode,
6532                                  dentry->d_name.name, dentry->d_name.len,
6533                                  backref, index);
6534         if (err > 0)
6535                 err = -EEXIST;
6536         return err;
6537 }
6538
6539 static int btrfs_mknod(struct inode *dir, struct dentry *dentry,
6540                         umode_t mode, dev_t rdev)
6541 {
6542         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
6543         struct btrfs_trans_handle *trans;
6544         struct btrfs_root *root = BTRFS_I(dir)->root;
6545         struct inode *inode = NULL;
6546         int err;
6547         int drop_inode = 0;
6548         u64 objectid;
6549         u64 index = 0;
6550
6551         /*
6552          * 2 for inode item and ref
6553          * 2 for dir items
6554          * 1 for xattr if selinux is on
6555          */
6556         trans = btrfs_start_transaction(root, 5);
6557         if (IS_ERR(trans))
6558                 return PTR_ERR(trans);
6559
6560         err = btrfs_find_free_ino(root, &objectid);
6561         if (err)
6562                 goto out_unlock;
6563
6564         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
6565                         dentry->d_name.len, btrfs_ino(BTRFS_I(dir)), objectid,
6566                         mode, &index);
6567         if (IS_ERR(inode)) {
6568                 err = PTR_ERR(inode);
6569                 goto out_unlock;
6570         }
6571
6572         /*
6573         * If the active LSM wants to access the inode during
6574         * d_instantiate it needs these. Smack checks to see
6575         * if the filesystem supports xattrs by looking at the
6576         * ops vector.
6577         */
6578         inode->i_op = &btrfs_special_inode_operations;
6579         init_special_inode(inode, inode->i_mode, rdev);
6580
6581         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
6582         if (err)
6583                 goto out_unlock_inode;
6584
6585         err = btrfs_add_nondir(trans, BTRFS_I(dir), dentry, BTRFS_I(inode),
6586                         0, index);
6587         if (err) {
6588                 goto out_unlock_inode;
6589         } else {
6590                 btrfs_update_inode(trans, root, inode);
6591                 unlock_new_inode(inode);
6592                 d_instantiate(dentry, inode);
6593         }
6594
6595 out_unlock:
6596         btrfs_end_transaction(trans);
6597         btrfs_balance_delayed_items(fs_info);
6598         btrfs_btree_balance_dirty(fs_info);
6599         if (drop_inode) {
6600                 inode_dec_link_count(inode);
6601                 iput(inode);
6602         }
6603         return err;
6604
6605 out_unlock_inode:
6606         drop_inode = 1;
6607         unlock_new_inode(inode);
6608         goto out_unlock;
6609
6610 }
6611
6612 static int btrfs_create(struct inode *dir, struct dentry *dentry,
6613                         umode_t mode, bool excl)
6614 {
6615         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
6616         struct btrfs_trans_handle *trans;
6617         struct btrfs_root *root = BTRFS_I(dir)->root;
6618         struct inode *inode = NULL;
6619         int drop_inode_on_err = 0;
6620         int err;
6621         u64 objectid;
6622         u64 index = 0;
6623
6624         /*
6625          * 2 for inode item and ref
6626          * 2 for dir items
6627          * 1 for xattr if selinux is on
6628          */
6629         trans = btrfs_start_transaction(root, 5);
6630         if (IS_ERR(trans))
6631                 return PTR_ERR(trans);
6632
6633         err = btrfs_find_free_ino(root, &objectid);
6634         if (err)
6635                 goto out_unlock;
6636
6637         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
6638                         dentry->d_name.len, btrfs_ino(BTRFS_I(dir)), objectid,
6639                         mode, &index);
6640         if (IS_ERR(inode)) {
6641                 err = PTR_ERR(inode);
6642                 goto out_unlock;
6643         }
6644         drop_inode_on_err = 1;
6645         /*
6646         * If the active LSM wants to access the inode during
6647         * d_instantiate it needs these. Smack checks to see
6648         * if the filesystem supports xattrs by looking at the
6649         * ops vector.
6650         */
6651         inode->i_fop = &btrfs_file_operations;
6652         inode->i_op = &btrfs_file_inode_operations;
6653         inode->i_mapping->a_ops = &btrfs_aops;
6654
6655         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
6656         if (err)
6657                 goto out_unlock_inode;
6658
6659         err = btrfs_update_inode(trans, root, inode);
6660         if (err)
6661                 goto out_unlock_inode;
6662
6663         err = btrfs_add_nondir(trans, BTRFS_I(dir), dentry, BTRFS_I(inode),
6664                         0, index);
6665         if (err)
6666                 goto out_unlock_inode;
6667
6668         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
6669         unlock_new_inode(inode);
6670         d_instantiate(dentry, inode);
6671
6672 out_unlock:
6673         btrfs_end_transaction(trans);
6674         if (err && drop_inode_on_err) {
6675                 inode_dec_link_count(inode);
6676                 iput(inode);
6677         }
6678         btrfs_balance_delayed_items(fs_info);
6679         btrfs_btree_balance_dirty(fs_info);
6680         return err;
6681
6682 out_unlock_inode:
6683         unlock_new_inode(inode);
6684         goto out_unlock;
6685
6686 }
6687
6688 static int btrfs_link(struct dentry *old_dentry, struct inode *dir,
6689                       struct dentry *dentry)
6690 {
6691         struct btrfs_trans_handle *trans = NULL;
6692         struct btrfs_root *root = BTRFS_I(dir)->root;
6693         struct inode *inode = d_inode(old_dentry);
6694         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
6695         u64 index;
6696         int err;
6697         int drop_inode = 0;
6698
6699         /* do not allow sys_link's with other subvols of the same device */
6700         if (root->objectid != BTRFS_I(inode)->root->objectid)
6701                 return -EXDEV;
6702
6703         if (inode->i_nlink >= BTRFS_LINK_MAX)
6704                 return -EMLINK;
6705
6706         err = btrfs_set_inode_index(BTRFS_I(dir), &index);
6707         if (err)
6708                 goto fail;
6709
6710         /*
6711          * 2 items for inode and inode ref
6712          * 2 items for dir items
6713          * 1 item for parent inode
6714          */
6715         trans = btrfs_start_transaction(root, 5);
6716         if (IS_ERR(trans)) {
6717                 err = PTR_ERR(trans);
6718                 trans = NULL;
6719                 goto fail;
6720         }
6721
6722         /* There are several dir indexes for this inode, clear the cache. */
6723         BTRFS_I(inode)->dir_index = 0ULL;
6724         inc_nlink(inode);
6725         inode_inc_iversion(inode);
6726         inode->i_ctime = current_time(inode);
6727         ihold(inode);
6728         set_bit(BTRFS_INODE_COPY_EVERYTHING, &BTRFS_I(inode)->runtime_flags);
6729
6730         err = btrfs_add_nondir(trans, BTRFS_I(dir), dentry, BTRFS_I(inode),
6731                         1, index);
6732
6733         if (err) {
6734                 drop_inode = 1;
6735         } else {
6736                 struct dentry *parent = dentry->d_parent;
6737                 err = btrfs_update_inode(trans, root, inode);
6738                 if (err)
6739                         goto fail;
6740                 if (inode->i_nlink == 1) {
6741                         /*
6742                          * If new hard link count is 1, it's a file created
6743                          * with open(2) O_TMPFILE flag.
6744                          */
6745                         err = btrfs_orphan_del(trans, BTRFS_I(inode));
6746                         if (err)
6747                                 goto fail;
6748                 }
6749                 d_instantiate(dentry, inode);
6750                 btrfs_log_new_name(trans, BTRFS_I(inode), NULL, parent);
6751         }
6752
6753         btrfs_balance_delayed_items(fs_info);
6754 fail:
6755         if (trans)
6756                 btrfs_end_transaction(trans);
6757         if (drop_inode) {
6758                 inode_dec_link_count(inode);
6759                 iput(inode);
6760         }
6761         btrfs_btree_balance_dirty(fs_info);
6762         return err;
6763 }
6764
6765 static int btrfs_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode)
6766 {
6767         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
6768         struct inode *inode = NULL;
6769         struct btrfs_trans_handle *trans;
6770         struct btrfs_root *root = BTRFS_I(dir)->root;
6771         int err = 0;
6772         int drop_on_err = 0;
6773         u64 objectid = 0;
6774         u64 index = 0;
6775
6776         /*
6777          * 2 items for inode and ref
6778          * 2 items for dir items
6779          * 1 for xattr if selinux is on
6780          */
6781         trans = btrfs_start_transaction(root, 5);
6782         if (IS_ERR(trans))
6783                 return PTR_ERR(trans);
6784
6785         err = btrfs_find_free_ino(root, &objectid);
6786         if (err)
6787                 goto out_fail;
6788
6789         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
6790                         dentry->d_name.len, btrfs_ino(BTRFS_I(dir)), objectid,
6791                         S_IFDIR | mode, &index);
6792         if (IS_ERR(inode)) {
6793                 err = PTR_ERR(inode);
6794                 goto out_fail;
6795         }
6796
6797         drop_on_err = 1;
6798         /* these must be set before we unlock the inode */
6799         inode->i_op = &btrfs_dir_inode_operations;
6800         inode->i_fop = &btrfs_dir_file_operations;
6801
6802         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
6803         if (err)
6804                 goto out_fail_inode;
6805
6806         btrfs_i_size_write(BTRFS_I(inode), 0);
6807         err = btrfs_update_inode(trans, root, inode);
6808         if (err)
6809                 goto out_fail_inode;
6810
6811         err = btrfs_add_link(trans, BTRFS_I(dir), BTRFS_I(inode),
6812                         dentry->d_name.name,
6813                         dentry->d_name.len, 0, index);
6814         if (err)
6815                 goto out_fail_inode;
6816
6817         d_instantiate(dentry, inode);
6818         /*
6819          * mkdir is special.  We're unlocking after we call d_instantiate
6820          * to avoid a race with nfsd calling d_instantiate.
6821          */
6822         unlock_new_inode(inode);
6823         drop_on_err = 0;
6824
6825 out_fail:
6826         btrfs_end_transaction(trans);
6827         if (drop_on_err) {
6828                 inode_dec_link_count(inode);
6829                 iput(inode);
6830         }
6831         btrfs_balance_delayed_items(fs_info);
6832         btrfs_btree_balance_dirty(fs_info);
6833         return err;
6834
6835 out_fail_inode:
6836         unlock_new_inode(inode);
6837         goto out_fail;
6838 }
6839
6840 /* Find next extent map of a given extent map, caller needs to ensure locks */
6841 static struct extent_map *next_extent_map(struct extent_map *em)
6842 {
6843         struct rb_node *next;
6844
6845         next = rb_next(&em->rb_node);
6846         if (!next)
6847                 return NULL;
6848         return container_of(next, struct extent_map, rb_node);
6849 }
6850
6851 static struct extent_map *prev_extent_map(struct extent_map *em)
6852 {
6853         struct rb_node *prev;
6854
6855         prev = rb_prev(&em->rb_node);
6856         if (!prev)
6857                 return NULL;
6858         return container_of(prev, struct extent_map, rb_node);
6859 }
6860
6861 /* helper for btfs_get_extent.  Given an existing extent in the tree,
6862  * the existing extent is the nearest extent to map_start,
6863  * and an extent that you want to insert, deal with overlap and insert
6864  * the best fitted new extent into the tree.
6865  */
6866 static int merge_extent_mapping(struct extent_map_tree *em_tree,
6867                                 struct extent_map *existing,
6868                                 struct extent_map *em,
6869                                 u64 map_start)
6870 {
6871         struct extent_map *prev;
6872         struct extent_map *next;
6873         u64 start;
6874         u64 end;
6875         u64 start_diff;
6876
6877         BUG_ON(map_start < em->start || map_start >= extent_map_end(em));
6878
6879         if (existing->start > map_start) {
6880                 next = existing;
6881                 prev = prev_extent_map(next);
6882         } else {
6883                 prev = existing;
6884                 next = next_extent_map(prev);
6885         }
6886
6887         start = prev ? extent_map_end(prev) : em->start;
6888         start = max_t(u64, start, em->start);
6889         end = next ? next->start : extent_map_end(em);
6890         end = min_t(u64, end, extent_map_end(em));
6891         start_diff = start - em->start;
6892         em->start = start;
6893         em->len = end - start;
6894         if (em->block_start < EXTENT_MAP_LAST_BYTE &&
6895             !test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
6896                 em->block_start += start_diff;
6897                 em->block_len -= start_diff;
6898         }
6899         return add_extent_mapping(em_tree, em, 0);
6900 }
6901
6902 static noinline int uncompress_inline(struct btrfs_path *path,
6903                                       struct page *page,
6904                                       size_t pg_offset, u64 extent_offset,
6905                                       struct btrfs_file_extent_item *item)
6906 {
6907         int ret;
6908         struct extent_buffer *leaf = path->nodes[0];
6909         char *tmp;
6910         size_t max_size;
6911         unsigned long inline_size;
6912         unsigned long ptr;
6913         int compress_type;
6914
6915         WARN_ON(pg_offset != 0);
6916         compress_type = btrfs_file_extent_compression(leaf, item);
6917         max_size = btrfs_file_extent_ram_bytes(leaf, item);
6918         inline_size = btrfs_file_extent_inline_item_len(leaf,
6919                                         btrfs_item_nr(path->slots[0]));
6920         tmp = kmalloc(inline_size, GFP_NOFS);
6921         if (!tmp)
6922                 return -ENOMEM;
6923         ptr = btrfs_file_extent_inline_start(item);
6924
6925         read_extent_buffer(leaf, tmp, ptr, inline_size);
6926
6927         max_size = min_t(unsigned long, PAGE_SIZE, max_size);
6928         ret = btrfs_decompress(compress_type, tmp, page,
6929                                extent_offset, inline_size, max_size);
6930
6931         /*
6932          * decompression code contains a memset to fill in any space between the end
6933          * of the uncompressed data and the end of max_size in case the decompressed
6934          * data ends up shorter than ram_bytes.  That doesn't cover the hole between
6935          * the end of an inline extent and the beginning of the next block, so we
6936          * cover that region here.
6937          */
6938
6939         if (max_size + pg_offset < PAGE_SIZE) {
6940                 char *map = kmap(page);
6941                 memset(map + pg_offset + max_size, 0, PAGE_SIZE - max_size - pg_offset);
6942                 kunmap(page);
6943         }
6944         kfree(tmp);
6945         return ret;
6946 }
6947
6948 /*
6949  * a bit scary, this does extent mapping from logical file offset to the disk.
6950  * the ugly parts come from merging extents from the disk with the in-ram
6951  * representation.  This gets more complex because of the data=ordered code,
6952  * where the in-ram extents might be locked pending data=ordered completion.
6953  *
6954  * This also copies inline extents directly into the page.
6955  */
6956 struct extent_map *btrfs_get_extent(struct btrfs_inode *inode,
6957                 struct page *page,
6958             size_t pg_offset, u64 start, u64 len,
6959                 int create)
6960 {
6961         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
6962         int ret;
6963         int err = 0;
6964         u64 extent_start = 0;
6965         u64 extent_end = 0;
6966         u64 objectid = btrfs_ino(inode);
6967         u32 found_type;
6968         struct btrfs_path *path = NULL;
6969         struct btrfs_root *root = inode->root;
6970         struct btrfs_file_extent_item *item;
6971         struct extent_buffer *leaf;
6972         struct btrfs_key found_key;
6973         struct extent_map *em = NULL;
6974         struct extent_map_tree *em_tree = &inode->extent_tree;
6975         struct extent_io_tree *io_tree = &inode->io_tree;
6976         struct btrfs_trans_handle *trans = NULL;
6977         const bool new_inline = !page || create;
6978
6979 again:
6980         read_lock(&em_tree->lock);
6981         em = lookup_extent_mapping(em_tree, start, len);
6982         if (em)
6983                 em->bdev = fs_info->fs_devices->latest_bdev;
6984         read_unlock(&em_tree->lock);
6985
6986         if (em) {
6987                 if (em->start > start || em->start + em->len <= start)
6988                         free_extent_map(em);
6989                 else if (em->block_start == EXTENT_MAP_INLINE && page)
6990                         free_extent_map(em);
6991                 else
6992                         goto out;
6993         }
6994         em = alloc_extent_map();
6995         if (!em) {
6996                 err = -ENOMEM;
6997                 goto out;
6998         }
6999         em->bdev = fs_info->fs_devices->latest_bdev;
7000         em->start = EXTENT_MAP_HOLE;
7001         em->orig_start = EXTENT_MAP_HOLE;
7002         em->len = (u64)-1;
7003         em->block_len = (u64)-1;
7004
7005         if (!path) {
7006                 path = btrfs_alloc_path();
7007                 if (!path) {
7008                         err = -ENOMEM;
7009                         goto out;
7010                 }
7011                 /*
7012                  * Chances are we'll be called again, so go ahead and do
7013                  * readahead
7014                  */
7015                 path->reada = READA_FORWARD;
7016         }
7017
7018         ret = btrfs_lookup_file_extent(trans, root, path,
7019                                        objectid, start, trans != NULL);
7020         if (ret < 0) {
7021                 err = ret;
7022                 goto out;
7023         }
7024
7025         if (ret != 0) {
7026                 if (path->slots[0] == 0)
7027                         goto not_found;
7028                 path->slots[0]--;
7029         }
7030
7031         leaf = path->nodes[0];
7032         item = btrfs_item_ptr(leaf, path->slots[0],
7033                               struct btrfs_file_extent_item);
7034         /* are we inside the extent that was found? */
7035         btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
7036         found_type = found_key.type;
7037         if (found_key.objectid != objectid ||
7038             found_type != BTRFS_EXTENT_DATA_KEY) {
7039                 /*
7040                  * If we backup past the first extent we want to move forward
7041                  * and see if there is an extent in front of us, otherwise we'll
7042                  * say there is a hole for our whole search range which can
7043                  * cause problems.
7044                  */
7045                 extent_end = start;
7046                 goto next;
7047         }
7048
7049         found_type = btrfs_file_extent_type(leaf, item);
7050         extent_start = found_key.offset;
7051         if (found_type == BTRFS_FILE_EXTENT_REG ||
7052             found_type == BTRFS_FILE_EXTENT_PREALLOC) {
7053                 extent_end = extent_start +
7054                        btrfs_file_extent_num_bytes(leaf, item);
7055
7056                 trace_btrfs_get_extent_show_fi_regular(inode, leaf, item,
7057                                                        extent_start);
7058         } else if (found_type == BTRFS_FILE_EXTENT_INLINE) {
7059                 size_t size;
7060                 size = btrfs_file_extent_inline_len(leaf, path->slots[0], item);
7061                 extent_end = ALIGN(extent_start + size,
7062                                    fs_info->sectorsize);
7063
7064                 trace_btrfs_get_extent_show_fi_inline(inode, leaf, item,
7065                                                       path->slots[0],
7066                                                       extent_start);
7067         }
7068 next:
7069         if (start >= extent_end) {
7070                 path->slots[0]++;
7071                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
7072                         ret = btrfs_next_leaf(root, path);
7073                         if (ret < 0) {
7074                                 err = ret;
7075                                 goto out;
7076                         }
7077                         if (ret > 0)
7078                                 goto not_found;
7079                         leaf = path->nodes[0];
7080                 }
7081                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
7082                 if (found_key.objectid != objectid ||
7083                     found_key.type != BTRFS_EXTENT_DATA_KEY)
7084                         goto not_found;
7085                 if (start + len <= found_key.offset)
7086                         goto not_found;
7087                 if (start > found_key.offset)
7088                         goto next;
7089                 em->start = start;
7090                 em->orig_start = start;
7091                 em->len = found_key.offset - start;
7092                 goto not_found_em;
7093         }
7094
7095         btrfs_extent_item_to_extent_map(inode, path, item,
7096                         new_inline, em);
7097
7098         if (found_type == BTRFS_FILE_EXTENT_REG ||
7099             found_type == BTRFS_FILE_EXTENT_PREALLOC) {
7100                 goto insert;
7101         } else if (found_type == BTRFS_FILE_EXTENT_INLINE) {
7102                 unsigned long ptr;
7103                 char *map;
7104                 size_t size;
7105                 size_t extent_offset;
7106                 size_t copy_size;
7107
7108                 if (new_inline)
7109                         goto out;
7110
7111                 size = btrfs_file_extent_inline_len(leaf, path->slots[0], item);
7112                 extent_offset = page_offset(page) + pg_offset - extent_start;
7113                 copy_size = min_t(u64, PAGE_SIZE - pg_offset,
7114                                   size - extent_offset);
7115                 em->start = extent_start + extent_offset;
7116                 em->len = ALIGN(copy_size, fs_info->sectorsize);
7117                 em->orig_block_len = em->len;
7118                 em->orig_start = em->start;
7119                 ptr = btrfs_file_extent_inline_start(item) + extent_offset;
7120                 if (create == 0 && !PageUptodate(page)) {
7121                         if (btrfs_file_extent_compression(leaf, item) !=
7122                             BTRFS_COMPRESS_NONE) {
7123                                 ret = uncompress_inline(path, page, pg_offset,
7124                                                         extent_offset, item);
7125                                 if (ret) {
7126                                         err = ret;
7127                                         goto out;
7128                                 }
7129                         } else {
7130                                 map = kmap(page);
7131                                 read_extent_buffer(leaf, map + pg_offset, ptr,
7132                                                    copy_size);
7133                                 if (pg_offset + copy_size < PAGE_SIZE) {
7134                                         memset(map + pg_offset + copy_size, 0,
7135                                                PAGE_SIZE - pg_offset -
7136                                                copy_size);
7137                                 }
7138                                 kunmap(page);
7139                         }
7140                         flush_dcache_page(page);
7141                 } else if (create && PageUptodate(page)) {
7142                         BUG();
7143                         if (!trans) {
7144                                 kunmap(page);
7145                                 free_extent_map(em);
7146                                 em = NULL;
7147
7148                                 btrfs_release_path(path);
7149                                 trans = btrfs_join_transaction(root);
7150
7151                                 if (IS_ERR(trans))
7152                                         return ERR_CAST(trans);
7153                                 goto again;
7154                         }
7155                         map = kmap(page);
7156                         write_extent_buffer(leaf, map + pg_offset, ptr,
7157                                             copy_size);
7158                         kunmap(page);
7159                         btrfs_mark_buffer_dirty(leaf);
7160                 }
7161                 set_extent_uptodate(io_tree, em->start,
7162                                     extent_map_end(em) - 1, NULL, GFP_NOFS);
7163                 goto insert;
7164         }
7165 not_found:
7166         em->start = start;
7167         em->orig_start = start;
7168         em->len = len;
7169 not_found_em:
7170         em->block_start = EXTENT_MAP_HOLE;
7171         set_bit(EXTENT_FLAG_VACANCY, &em->flags);
7172 insert:
7173         btrfs_release_path(path);
7174         if (em->start > start || extent_map_end(em) <= start) {
7175                 btrfs_err(fs_info,
7176                           "bad extent! em: [%llu %llu] passed [%llu %llu]",
7177                           em->start, em->len, start, len);
7178                 err = -EIO;
7179                 goto out;
7180         }
7181
7182         err = 0;
7183         write_lock(&em_tree->lock);
7184         ret = add_extent_mapping(em_tree, em, 0);
7185         /* it is possible that someone inserted the extent into the tree
7186          * while we had the lock dropped.  It is also possible that
7187          * an overlapping map exists in the tree
7188          */
7189         if (ret == -EEXIST) {
7190                 struct extent_map *existing;
7191
7192                 ret = 0;
7193
7194                 existing = search_extent_mapping(em_tree, start, len);
7195                 /*
7196                  * existing will always be non-NULL, since there must be
7197                  * extent causing the -EEXIST.
7198                  */
7199                 if (existing->start == em->start &&
7200                     extent_map_end(existing) >= extent_map_end(em) &&
7201                     em->block_start == existing->block_start) {
7202                         /*
7203                          * The existing extent map already encompasses the
7204                          * entire extent map we tried to add.
7205                          */
7206                         free_extent_map(em);
7207                         em = existing;
7208                         err = 0;
7209
7210                 } else if (start >= extent_map_end(existing) ||
7211                     start <= existing->start) {
7212                         /*
7213                          * The existing extent map is the one nearest to
7214                          * the [start, start + len) range which overlaps
7215                          */
7216                         err = merge_extent_mapping(em_tree, existing,
7217                                                    em, start);
7218                         free_extent_map(existing);
7219                         if (err) {
7220                                 free_extent_map(em);
7221                                 em = NULL;
7222                         }
7223                 } else {
7224                         free_extent_map(em);
7225                         em = existing;
7226                         err = 0;
7227                 }
7228         }
7229         write_unlock(&em_tree->lock);
7230 out:
7231
7232         trace_btrfs_get_extent(root, inode, em);
7233
7234         btrfs_free_path(path);
7235         if (trans) {
7236                 ret = btrfs_end_transaction(trans);
7237                 if (!err)
7238                         err = ret;
7239         }
7240         if (err) {
7241                 free_extent_map(em);
7242                 return ERR_PTR(err);
7243         }
7244         BUG_ON(!em); /* Error is always set */
7245         return em;
7246 }
7247
7248 struct extent_map *btrfs_get_extent_fiemap(struct btrfs_inode *inode,
7249                 struct page *page,
7250                 size_t pg_offset, u64 start, u64 len,
7251                 int create)
7252 {
7253         struct extent_map *em;
7254         struct extent_map *hole_em = NULL;
7255         u64 range_start = start;
7256         u64 end;
7257         u64 found;
7258         u64 found_end;
7259         int err = 0;
7260
7261         em = btrfs_get_extent(inode, page, pg_offset, start, len, create);
7262         if (IS_ERR(em))
7263                 return em;
7264         /*
7265          * If our em maps to:
7266          * - a hole or
7267          * - a pre-alloc extent,
7268          * there might actually be delalloc bytes behind it.
7269          */
7270         if (em->block_start != EXTENT_MAP_HOLE &&
7271             !test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
7272                 return em;
7273         else
7274                 hole_em = em;
7275
7276         /* check to see if we've wrapped (len == -1 or similar) */
7277         end = start + len;
7278         if (end < start)
7279                 end = (u64)-1;
7280         else
7281                 end -= 1;
7282
7283         em = NULL;
7284
7285         /* ok, we didn't find anything, lets look for delalloc */
7286         found = count_range_bits(&inode->io_tree, &range_start,
7287                                  end, len, EXTENT_DELALLOC, 1);
7288         found_end = range_start + found;
7289         if (found_end < range_start)
7290                 found_end = (u64)-1;
7291
7292         /*
7293          * we didn't find anything useful, return
7294          * the original results from get_extent()
7295          */
7296         if (range_start > end || found_end <= start) {
7297                 em = hole_em;
7298                 hole_em = NULL;
7299                 goto out;
7300         }
7301
7302         /* adjust the range_start to make sure it doesn't
7303          * go backwards from the start they passed in
7304          */
7305         range_start = max(start, range_start);
7306         found = found_end - range_start;
7307
7308         if (found > 0) {
7309                 u64 hole_start = start;
7310                 u64 hole_len = len;
7311
7312                 em = alloc_extent_map();
7313                 if (!em) {
7314                         err = -ENOMEM;
7315                         goto out;
7316                 }
7317                 /*
7318                  * when btrfs_get_extent can't find anything it
7319                  * returns one huge hole
7320                  *
7321                  * make sure what it found really fits our range, and
7322                  * adjust to make sure it is based on the start from
7323                  * the caller
7324                  */
7325                 if (hole_em) {
7326                         u64 calc_end = extent_map_end(hole_em);
7327
7328                         if (calc_end <= start || (hole_em->start > end)) {
7329                                 free_extent_map(hole_em);
7330                                 hole_em = NULL;
7331                         } else {
7332                                 hole_start = max(hole_em->start, start);
7333                                 hole_len = calc_end - hole_start;
7334                         }
7335                 }
7336                 em->bdev = NULL;
7337                 if (hole_em && range_start > hole_start) {
7338                         /* our hole starts before our delalloc, so we
7339                          * have to return just the parts of the hole
7340                          * that go until  the delalloc starts
7341                          */
7342                         em->len = min(hole_len,
7343                                       range_start - hole_start);
7344                         em->start = hole_start;
7345                         em->orig_start = hole_start;
7346                         /*
7347                          * don't adjust block start at all,
7348                          * it is fixed at EXTENT_MAP_HOLE
7349                          */
7350                         em->block_start = hole_em->block_start;
7351                         em->block_len = hole_len;
7352                         if (test_bit(EXTENT_FLAG_PREALLOC, &hole_em->flags))
7353                                 set_bit(EXTENT_FLAG_PREALLOC, &em->flags);
7354                 } else {
7355                         em->start = range_start;
7356                         em->len = found;
7357                         em->orig_start = range_start;
7358                         em->block_start = EXTENT_MAP_DELALLOC;
7359                         em->block_len = found;
7360                 }
7361         } else if (hole_em) {
7362                 return hole_em;
7363         }
7364 out:
7365
7366         free_extent_map(hole_em);
7367         if (err) {
7368                 free_extent_map(em);
7369                 return ERR_PTR(err);
7370         }
7371         return em;
7372 }
7373
7374 static struct extent_map *btrfs_create_dio_extent(struct inode *inode,
7375                                                   const u64 start,
7376                                                   const u64 len,
7377                                                   const u64 orig_start,
7378                                                   const u64 block_start,
7379                                                   const u64 block_len,
7380                                                   const u64 orig_block_len,
7381                                                   const u64 ram_bytes,
7382                                                   const int type)
7383 {
7384         struct extent_map *em = NULL;
7385         int ret;
7386
7387         if (type != BTRFS_ORDERED_NOCOW) {
7388                 em = create_io_em(inode, start, len, orig_start,
7389                                   block_start, block_len, orig_block_len,
7390                                   ram_bytes,
7391                                   BTRFS_COMPRESS_NONE, /* compress_type */
7392                                   type);
7393                 if (IS_ERR(em))
7394                         goto out;
7395         }
7396         ret = btrfs_add_ordered_extent_dio(inode, start, block_start,
7397                                            len, block_len, type);
7398         if (ret) {
7399                 if (em) {
7400                         free_extent_map(em);
7401                         btrfs_drop_extent_cache(BTRFS_I(inode), start,
7402                                                 start + len - 1, 0);
7403                 }
7404                 em = ERR_PTR(ret);
7405         }
7406  out:
7407
7408         return em;
7409 }
7410
7411 static struct extent_map *btrfs_new_extent_direct(struct inode *inode,
7412                                                   u64 start, u64 len)
7413 {
7414         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7415         struct btrfs_root *root = BTRFS_I(inode)->root;
7416         struct extent_map *em;
7417         struct btrfs_key ins;
7418         u64 alloc_hint;
7419         int ret;
7420
7421         alloc_hint = get_extent_allocation_hint(inode, start, len);
7422         ret = btrfs_reserve_extent(root, len, len, fs_info->sectorsize,
7423                                    0, alloc_hint, &ins, 1, 1);
7424         if (ret)
7425                 return ERR_PTR(ret);
7426
7427         em = btrfs_create_dio_extent(inode, start, ins.offset, start,
7428                                      ins.objectid, ins.offset, ins.offset,
7429                                      ins.offset, BTRFS_ORDERED_REGULAR);
7430         btrfs_dec_block_group_reservations(fs_info, ins.objectid);
7431         if (IS_ERR(em))
7432                 btrfs_free_reserved_extent(fs_info, ins.objectid,
7433                                            ins.offset, 1);
7434
7435         return em;
7436 }
7437
7438 /*
7439  * returns 1 when the nocow is safe, < 1 on error, 0 if the
7440  * block must be cow'd
7441  */
7442 noinline int can_nocow_extent(struct inode *inode, u64 offset, u64 *len,
7443                               u64 *orig_start, u64 *orig_block_len,
7444                               u64 *ram_bytes)
7445 {
7446         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7447         struct btrfs_path *path;
7448         int ret;
7449         struct extent_buffer *leaf;
7450         struct btrfs_root *root = BTRFS_I(inode)->root;
7451         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
7452         struct btrfs_file_extent_item *fi;
7453         struct btrfs_key key;
7454         u64 disk_bytenr;
7455         u64 backref_offset;
7456         u64 extent_end;
7457         u64 num_bytes;
7458         int slot;
7459         int found_type;
7460         bool nocow = (BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW);
7461
7462         path = btrfs_alloc_path();
7463         if (!path)
7464                 return -ENOMEM;
7465
7466         ret = btrfs_lookup_file_extent(NULL, root, path,
7467                         btrfs_ino(BTRFS_I(inode)), offset, 0);
7468         if (ret < 0)
7469                 goto out;
7470
7471         slot = path->slots[0];
7472         if (ret == 1) {
7473                 if (slot == 0) {
7474                         /* can't find the item, must cow */
7475                         ret = 0;
7476                         goto out;
7477                 }
7478                 slot--;
7479         }
7480         ret = 0;
7481         leaf = path->nodes[0];
7482         btrfs_item_key_to_cpu(leaf, &key, slot);
7483         if (key.objectid != btrfs_ino(BTRFS_I(inode)) ||
7484             key.type != BTRFS_EXTENT_DATA_KEY) {
7485                 /* not our file or wrong item type, must cow */
7486                 goto out;
7487         }
7488
7489         if (key.offset > offset) {
7490                 /* Wrong offset, must cow */
7491                 goto out;
7492         }
7493
7494         fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
7495         found_type = btrfs_file_extent_type(leaf, fi);
7496         if (found_type != BTRFS_FILE_EXTENT_REG &&
7497             found_type != BTRFS_FILE_EXTENT_PREALLOC) {
7498                 /* not a regular extent, must cow */
7499                 goto out;
7500         }
7501
7502         if (!nocow && found_type == BTRFS_FILE_EXTENT_REG)
7503                 goto out;
7504
7505         extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
7506         if (extent_end <= offset)
7507                 goto out;
7508
7509         disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
7510         if (disk_bytenr == 0)
7511                 goto out;
7512
7513         if (btrfs_file_extent_compression(leaf, fi) ||
7514             btrfs_file_extent_encryption(leaf, fi) ||
7515             btrfs_file_extent_other_encoding(leaf, fi))
7516                 goto out;
7517
7518         backref_offset = btrfs_file_extent_offset(leaf, fi);
7519
7520         if (orig_start) {
7521                 *orig_start = key.offset - backref_offset;
7522                 *orig_block_len = btrfs_file_extent_disk_num_bytes(leaf, fi);
7523                 *ram_bytes = btrfs_file_extent_ram_bytes(leaf, fi);
7524         }
7525
7526         if (btrfs_extent_readonly(fs_info, disk_bytenr))
7527                 goto out;
7528
7529         num_bytes = min(offset + *len, extent_end) - offset;
7530         if (!nocow && found_type == BTRFS_FILE_EXTENT_PREALLOC) {
7531                 u64 range_end;
7532
7533                 range_end = round_up(offset + num_bytes,
7534                                      root->fs_info->sectorsize) - 1;
7535                 ret = test_range_bit(io_tree, offset, range_end,
7536                                      EXTENT_DELALLOC, 0, NULL);
7537                 if (ret) {
7538                         ret = -EAGAIN;
7539                         goto out;
7540                 }
7541         }
7542
7543         btrfs_release_path(path);
7544
7545         /*
7546          * look for other files referencing this extent, if we
7547          * find any we must cow
7548          */
7549
7550         ret = btrfs_cross_ref_exist(root, btrfs_ino(BTRFS_I(inode)),
7551                                     key.offset - backref_offset, disk_bytenr);
7552         if (ret) {
7553                 ret = 0;
7554                 goto out;
7555         }
7556
7557         /*
7558          * adjust disk_bytenr and num_bytes to cover just the bytes
7559          * in this extent we are about to write.  If there
7560          * are any csums in that range we have to cow in order
7561          * to keep the csums correct
7562          */
7563         disk_bytenr += backref_offset;
7564         disk_bytenr += offset - key.offset;
7565         if (csum_exist_in_range(fs_info, disk_bytenr, num_bytes))
7566                 goto out;
7567         /*
7568          * all of the above have passed, it is safe to overwrite this extent
7569          * without cow
7570          */
7571         *len = num_bytes;
7572         ret = 1;
7573 out:
7574         btrfs_free_path(path);
7575         return ret;
7576 }
7577
7578 bool btrfs_page_exists_in_range(struct inode *inode, loff_t start, loff_t end)
7579 {
7580         struct radix_tree_root *root = &inode->i_mapping->page_tree;
7581         bool found = false;
7582         void **pagep = NULL;
7583         struct page *page = NULL;
7584         unsigned long start_idx;
7585         unsigned long end_idx;
7586
7587         start_idx = start >> PAGE_SHIFT;
7588
7589         /*
7590          * end is the last byte in the last page.  end == start is legal
7591          */
7592         end_idx = end >> PAGE_SHIFT;
7593
7594         rcu_read_lock();
7595
7596         /* Most of the code in this while loop is lifted from
7597          * find_get_page.  It's been modified to begin searching from a
7598          * page and return just the first page found in that range.  If the
7599          * found idx is less than or equal to the end idx then we know that
7600          * a page exists.  If no pages are found or if those pages are
7601          * outside of the range then we're fine (yay!) */
7602         while (page == NULL &&
7603                radix_tree_gang_lookup_slot(root, &pagep, NULL, start_idx, 1)) {
7604                 page = radix_tree_deref_slot(pagep);
7605                 if (unlikely(!page))
7606                         break;
7607
7608                 if (radix_tree_exception(page)) {
7609                         if (radix_tree_deref_retry(page)) {
7610                                 page = NULL;
7611                                 continue;
7612                         }
7613                         /*
7614                          * Otherwise, shmem/tmpfs must be storing a swap entry
7615                          * here as an exceptional entry: so return it without
7616                          * attempting to raise page count.
7617                          */
7618                         page = NULL;
7619                         break; /* TODO: Is this relevant for this use case? */
7620                 }
7621
7622                 if (!page_cache_get_speculative(page)) {
7623                         page = NULL;
7624                         continue;
7625                 }
7626
7627                 /*
7628                  * Has the page moved?
7629                  * This is part of the lockless pagecache protocol. See
7630                  * include/linux/pagemap.h for details.
7631                  */
7632                 if (unlikely(page != *pagep)) {
7633                         put_page(page);
7634                         page = NULL;
7635                 }
7636         }
7637
7638         if (page) {
7639                 if (page->index <= end_idx)
7640                         found = true;
7641                 put_page(page);
7642         }
7643
7644         rcu_read_unlock();
7645         return found;
7646 }
7647
7648 static int lock_extent_direct(struct inode *inode, u64 lockstart, u64 lockend,
7649                               struct extent_state **cached_state, int writing)
7650 {
7651         struct btrfs_ordered_extent *ordered;
7652         int ret = 0;
7653
7654         while (1) {
7655                 lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend,
7656                                  cached_state);
7657                 /*
7658                  * We're concerned with the entire range that we're going to be
7659                  * doing DIO to, so we need to make sure there's no ordered
7660                  * extents in this range.
7661                  */
7662                 ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), lockstart,
7663                                                      lockend - lockstart + 1);
7664
7665                 /*
7666                  * We need to make sure there are no buffered pages in this
7667                  * range either, we could have raced between the invalidate in
7668                  * generic_file_direct_write and locking the extent.  The
7669                  * invalidate needs to happen so that reads after a write do not
7670                  * get stale data.
7671                  */
7672                 if (!ordered &&
7673                     (!writing ||
7674                      !btrfs_page_exists_in_range(inode, lockstart, lockend)))
7675                         break;
7676
7677                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
7678                                      cached_state, GFP_NOFS);
7679
7680                 if (ordered) {
7681                         /*
7682                          * If we are doing a DIO read and the ordered extent we
7683                          * found is for a buffered write, we can not wait for it
7684                          * to complete and retry, because if we do so we can
7685                          * deadlock with concurrent buffered writes on page
7686                          * locks. This happens only if our DIO read covers more
7687                          * than one extent map, if at this point has already
7688                          * created an ordered extent for a previous extent map
7689                          * and locked its range in the inode's io tree, and a
7690                          * concurrent write against that previous extent map's
7691                          * range and this range started (we unlock the ranges
7692                          * in the io tree only when the bios complete and
7693                          * buffered writes always lock pages before attempting
7694                          * to lock range in the io tree).
7695                          */
7696                         if (writing ||
7697                             test_bit(BTRFS_ORDERED_DIRECT, &ordered->flags))
7698                                 btrfs_start_ordered_extent(inode, ordered, 1);
7699                         else
7700                                 ret = -ENOTBLK;
7701                         btrfs_put_ordered_extent(ordered);
7702                 } else {
7703                         /*
7704                          * We could trigger writeback for this range (and wait
7705                          * for it to complete) and then invalidate the pages for
7706                          * this range (through invalidate_inode_pages2_range()),
7707                          * but that can lead us to a deadlock with a concurrent
7708                          * call to readpages() (a buffered read or a defrag call
7709                          * triggered a readahead) on a page lock due to an
7710                          * ordered dio extent we created before but did not have
7711                          * yet a corresponding bio submitted (whence it can not
7712                          * complete), which makes readpages() wait for that
7713                          * ordered extent to complete while holding a lock on
7714                          * that page.
7715                          */
7716                         ret = -ENOTBLK;
7717                 }
7718
7719                 if (ret)
7720                         break;
7721
7722                 cond_resched();
7723         }
7724
7725         return ret;
7726 }
7727
7728 /* The callers of this must take lock_extent() */
7729 static struct extent_map *create_io_em(struct inode *inode, u64 start, u64 len,
7730                                        u64 orig_start, u64 block_start,
7731                                        u64 block_len, u64 orig_block_len,
7732                                        u64 ram_bytes, int compress_type,
7733                                        int type)
7734 {
7735         struct extent_map_tree *em_tree;
7736         struct extent_map *em;
7737         struct btrfs_root *root = BTRFS_I(inode)->root;
7738         int ret;
7739
7740         ASSERT(type == BTRFS_ORDERED_PREALLOC ||
7741                type == BTRFS_ORDERED_COMPRESSED ||
7742                type == BTRFS_ORDERED_NOCOW ||
7743                type == BTRFS_ORDERED_REGULAR);
7744
7745         em_tree = &BTRFS_I(inode)->extent_tree;
7746         em = alloc_extent_map();
7747         if (!em)
7748                 return ERR_PTR(-ENOMEM);
7749
7750         em->start = start;
7751         em->orig_start = orig_start;
7752         em->len = len;
7753         em->block_len = block_len;
7754         em->block_start = block_start;
7755         em->bdev = root->fs_info->fs_devices->latest_bdev;
7756         em->orig_block_len = orig_block_len;
7757         em->ram_bytes = ram_bytes;
7758         em->generation = -1;
7759         set_bit(EXTENT_FLAG_PINNED, &em->flags);
7760         if (type == BTRFS_ORDERED_PREALLOC) {
7761                 set_bit(EXTENT_FLAG_FILLING, &em->flags);
7762         } else if (type == BTRFS_ORDERED_COMPRESSED) {
7763                 set_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
7764                 em->compress_type = compress_type;
7765         }
7766
7767         do {
7768                 btrfs_drop_extent_cache(BTRFS_I(inode), em->start,
7769                                 em->start + em->len - 1, 0);
7770                 write_lock(&em_tree->lock);
7771                 ret = add_extent_mapping(em_tree, em, 1);
7772                 write_unlock(&em_tree->lock);
7773                 /*
7774                  * The caller has taken lock_extent(), who could race with us
7775                  * to add em?
7776                  */
7777         } while (ret == -EEXIST);
7778
7779         if (ret) {
7780                 free_extent_map(em);
7781                 return ERR_PTR(ret);
7782         }
7783
7784         /* em got 2 refs now, callers needs to do free_extent_map once. */
7785         return em;
7786 }
7787
7788 static void adjust_dio_outstanding_extents(struct inode *inode,
7789                                            struct btrfs_dio_data *dio_data,
7790                                            const u64 len)
7791 {
7792         unsigned num_extents = count_max_extents(len);
7793
7794         /*
7795          * If we have an outstanding_extents count still set then we're
7796          * within our reservation, otherwise we need to adjust our inode
7797          * counter appropriately.
7798          */
7799         if (dio_data->outstanding_extents >= num_extents) {
7800                 dio_data->outstanding_extents -= num_extents;
7801         } else {
7802                 /*
7803                  * If dio write length has been split due to no large enough
7804                  * contiguous space, we need to compensate our inode counter
7805                  * appropriately.
7806                  */
7807                 u64 num_needed = num_extents - dio_data->outstanding_extents;
7808
7809                 spin_lock(&BTRFS_I(inode)->lock);
7810                 BTRFS_I(inode)->outstanding_extents += num_needed;
7811                 spin_unlock(&BTRFS_I(inode)->lock);
7812         }
7813 }
7814
7815 static int btrfs_get_blocks_direct(struct inode *inode, sector_t iblock,
7816                                    struct buffer_head *bh_result, int create)
7817 {
7818         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7819         struct extent_map *em;
7820         struct extent_state *cached_state = NULL;
7821         struct btrfs_dio_data *dio_data = NULL;
7822         u64 start = iblock << inode->i_blkbits;
7823         u64 lockstart, lockend;
7824         u64 len = bh_result->b_size;
7825         int unlock_bits = EXTENT_LOCKED;
7826         int ret = 0;
7827
7828         if (create)
7829                 unlock_bits |= EXTENT_DIRTY;
7830         else
7831                 len = min_t(u64, len, fs_info->sectorsize);
7832
7833         lockstart = start;
7834         lockend = start + len - 1;
7835
7836         if (current->journal_info) {
7837                 /*
7838                  * Need to pull our outstanding extents and set journal_info to NULL so
7839                  * that anything that needs to check if there's a transaction doesn't get
7840                  * confused.
7841                  */
7842                 dio_data = current->journal_info;
7843                 current->journal_info = NULL;
7844         }
7845
7846         /*
7847          * If this errors out it's because we couldn't invalidate pagecache for
7848          * this range and we need to fallback to buffered.
7849          */
7850         if (lock_extent_direct(inode, lockstart, lockend, &cached_state,
7851                                create)) {
7852                 ret = -ENOTBLK;
7853                 goto err;
7854         }
7855
7856         em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, start, len, 0);
7857         if (IS_ERR(em)) {
7858                 ret = PTR_ERR(em);
7859                 goto unlock_err;
7860         }
7861
7862         /*
7863          * Ok for INLINE and COMPRESSED extents we need to fallback on buffered
7864          * io.  INLINE is special, and we could probably kludge it in here, but
7865          * it's still buffered so for safety lets just fall back to the generic
7866          * buffered path.
7867          *
7868          * For COMPRESSED we _have_ to read the entire extent in so we can
7869          * decompress it, so there will be buffering required no matter what we
7870          * do, so go ahead and fallback to buffered.
7871          *
7872          * We return -ENOTBLK because that's what makes DIO go ahead and go back
7873          * to buffered IO.  Don't blame me, this is the price we pay for using
7874          * the generic code.
7875          */
7876         if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags) ||
7877             em->block_start == EXTENT_MAP_INLINE) {
7878                 free_extent_map(em);
7879                 ret = -ENOTBLK;
7880                 goto unlock_err;
7881         }
7882
7883         /* Just a good old fashioned hole, return */
7884         if (!create && (em->block_start == EXTENT_MAP_HOLE ||
7885                         test_bit(EXTENT_FLAG_PREALLOC, &em->flags))) {
7886                 free_extent_map(em);
7887                 goto unlock_err;
7888         }
7889
7890         /*
7891          * We don't allocate a new extent in the following cases
7892          *
7893          * 1) The inode is marked as NODATACOW.  In this case we'll just use the
7894          * existing extent.
7895          * 2) The extent is marked as PREALLOC.  We're good to go here and can
7896          * just use the extent.
7897          *
7898          */
7899         if (!create) {
7900                 len = min(len, em->len - (start - em->start));
7901                 lockstart = start + len;
7902                 goto unlock;
7903         }
7904
7905         if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags) ||
7906             ((BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW) &&
7907              em->block_start != EXTENT_MAP_HOLE)) {
7908                 int type;
7909                 u64 block_start, orig_start, orig_block_len, ram_bytes;
7910
7911                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
7912                         type = BTRFS_ORDERED_PREALLOC;
7913                 else
7914                         type = BTRFS_ORDERED_NOCOW;
7915                 len = min(len, em->len - (start - em->start));
7916                 block_start = em->block_start + (start - em->start);
7917
7918                 if (can_nocow_extent(inode, start, &len, &orig_start,
7919                                      &orig_block_len, &ram_bytes) == 1 &&
7920                     btrfs_inc_nocow_writers(fs_info, block_start)) {
7921                         struct extent_map *em2;
7922
7923                         em2 = btrfs_create_dio_extent(inode, start, len,
7924                                                       orig_start, block_start,
7925                                                       len, orig_block_len,
7926                                                       ram_bytes, type);
7927                         btrfs_dec_nocow_writers(fs_info, block_start);
7928                         if (type == BTRFS_ORDERED_PREALLOC) {
7929                                 free_extent_map(em);
7930                                 em = em2;
7931                         }
7932                         if (em2 && IS_ERR(em2)) {
7933                                 ret = PTR_ERR(em2);
7934                                 goto unlock_err;
7935                         }
7936                         /*
7937                          * For inode marked NODATACOW or extent marked PREALLOC,
7938                          * use the existing or preallocated extent, so does not
7939                          * need to adjust btrfs_space_info's bytes_may_use.
7940                          */
7941                         btrfs_free_reserved_data_space_noquota(inode,
7942                                         start, len);
7943                         goto unlock;
7944                 }
7945         }
7946
7947         /*
7948          * this will cow the extent, reset the len in case we changed
7949          * it above
7950          */
7951         len = bh_result->b_size;
7952         free_extent_map(em);
7953         em = btrfs_new_extent_direct(inode, start, len);
7954         if (IS_ERR(em)) {
7955                 ret = PTR_ERR(em);
7956                 goto unlock_err;
7957         }
7958         len = min(len, em->len - (start - em->start));
7959 unlock:
7960         bh_result->b_blocknr = (em->block_start + (start - em->start)) >>
7961                 inode->i_blkbits;
7962         bh_result->b_size = len;
7963         bh_result->b_bdev = em->bdev;
7964         set_buffer_mapped(bh_result);
7965         if (create) {
7966                 if (!test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
7967                         set_buffer_new(bh_result);
7968
7969                 /*
7970                  * Need to update the i_size under the extent lock so buffered
7971                  * readers will get the updated i_size when we unlock.
7972                  */
7973                 if (!dio_data->overwrite && start + len > i_size_read(inode))
7974                         i_size_write(inode, start + len);
7975
7976                 adjust_dio_outstanding_extents(inode, dio_data, len);
7977                 WARN_ON(dio_data->reserve < len);
7978                 dio_data->reserve -= len;
7979                 dio_data->unsubmitted_oe_range_end = start + len;
7980                 current->journal_info = dio_data;
7981         }
7982
7983         /*
7984          * In the case of write we need to clear and unlock the entire range,
7985          * in the case of read we need to unlock only the end area that we
7986          * aren't using if there is any left over space.
7987          */
7988         if (lockstart < lockend) {
7989                 clear_extent_bit(&BTRFS_I(inode)->io_tree, lockstart,
7990                                  lockend, unlock_bits, 1, 0,
7991                                  &cached_state, GFP_NOFS);
7992         } else {
7993                 free_extent_state(cached_state);
7994         }
7995
7996         free_extent_map(em);
7997
7998         return 0;
7999
8000 unlock_err:
8001         clear_extent_bit(&BTRFS_I(inode)->io_tree, lockstart, lockend,
8002                          unlock_bits, 1, 0, &cached_state, GFP_NOFS);
8003 err:
8004         if (dio_data)
8005                 current->journal_info = dio_data;
8006         /*
8007          * Compensate the delalloc release we do in btrfs_direct_IO() when we
8008          * write less data then expected, so that we don't underflow our inode's
8009          * outstanding extents counter.
8010          */
8011         if (create && dio_data)
8012                 adjust_dio_outstanding_extents(inode, dio_data, len);
8013
8014         return ret;
8015 }
8016
8017 static inline int submit_dio_repair_bio(struct inode *inode, struct bio *bio,
8018                                         int mirror_num)
8019 {
8020         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
8021         int ret;
8022
8023         BUG_ON(bio_op(bio) == REQ_OP_WRITE);
8024
8025         bio_get(bio);
8026
8027         ret = btrfs_bio_wq_end_io(fs_info, bio, BTRFS_WQ_ENDIO_DIO_REPAIR);
8028         if (ret)
8029                 goto err;
8030
8031         ret = btrfs_map_bio(fs_info, bio, mirror_num, 0);
8032 err:
8033         bio_put(bio);
8034         return ret;
8035 }
8036
8037 static int btrfs_check_dio_repairable(struct inode *inode,
8038                                       struct bio *failed_bio,
8039                                       struct io_failure_record *failrec,
8040                                       int failed_mirror)
8041 {
8042         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
8043         int num_copies;
8044
8045         num_copies = btrfs_num_copies(fs_info, failrec->logical, failrec->len);
8046         if (num_copies == 1) {
8047                 /*
8048                  * we only have a single copy of the data, so don't bother with
8049                  * all the retry and error correction code that follows. no
8050                  * matter what the error is, it is very likely to persist.
8051                  */
8052                 btrfs_debug(fs_info,
8053                         "Check DIO Repairable: cannot repair, num_copies=%d, next_mirror %d, failed_mirror %d",
8054                         num_copies, failrec->this_mirror, failed_mirror);
8055                 return 0;
8056         }
8057
8058         failrec->failed_mirror = failed_mirror;
8059         failrec->this_mirror++;
8060         if (failrec->this_mirror == failed_mirror)
8061                 failrec->this_mirror++;
8062
8063         if (failrec->this_mirror > num_copies) {
8064                 btrfs_debug(fs_info,
8065                         "Check DIO Repairable: (fail) num_copies=%d, next_mirror %d, failed_mirror %d",
8066                         num_copies, failrec->this_mirror, failed_mirror);
8067                 return 0;
8068         }
8069
8070         return 1;
8071 }
8072
8073 static int dio_read_error(struct inode *inode, struct bio *failed_bio,
8074                         struct page *page, unsigned int pgoff,
8075                         u64 start, u64 end, int failed_mirror,
8076                         bio_end_io_t *repair_endio, void *repair_arg)
8077 {
8078         struct io_failure_record *failrec;
8079         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
8080         struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
8081         struct bio *bio;
8082         int isector;
8083         unsigned int read_mode = 0;
8084         int segs;
8085         int ret;
8086
8087         BUG_ON(bio_op(failed_bio) == REQ_OP_WRITE);
8088
8089         ret = btrfs_get_io_failure_record(inode, start, end, &failrec);
8090         if (ret)
8091                 return ret;
8092
8093         ret = btrfs_check_dio_repairable(inode, failed_bio, failrec,
8094                                          failed_mirror);
8095         if (!ret) {
8096                 free_io_failure(failure_tree, io_tree, failrec);
8097                 return -EIO;
8098         }
8099
8100         segs = bio_segments(failed_bio);
8101         if (segs > 1 ||
8102             (failed_bio->bi_io_vec->bv_len > btrfs_inode_sectorsize(inode)))
8103                 read_mode |= REQ_FAILFAST_DEV;
8104
8105         isector = start - btrfs_io_bio(failed_bio)->logical;
8106         isector >>= inode->i_sb->s_blocksize_bits;
8107         bio = btrfs_create_repair_bio(inode, failed_bio, failrec, page,
8108                                 pgoff, isector, repair_endio, repair_arg);
8109         bio_set_op_attrs(bio, REQ_OP_READ, read_mode);
8110
8111         btrfs_debug(BTRFS_I(inode)->root->fs_info,
8112                     "repair DIO read error: submitting new dio read[%#x] to this_mirror=%d, in_validation=%d",
8113                     read_mode, failrec->this_mirror, failrec->in_validation);
8114
8115         ret = submit_dio_repair_bio(inode, bio, failrec->this_mirror);
8116         if (ret) {
8117                 free_io_failure(failure_tree, io_tree, failrec);
8118                 bio_put(bio);
8119         }
8120
8121         return ret;
8122 }
8123
8124 struct btrfs_retry_complete {
8125         struct completion done;
8126         struct inode *inode;
8127         u64 start;
8128         int uptodate;
8129 };
8130
8131 static void btrfs_retry_endio_nocsum(struct bio *bio)
8132 {
8133         struct btrfs_retry_complete *done = bio->bi_private;
8134         struct inode *inode = done->inode;
8135         struct bio_vec *bvec;
8136         struct extent_io_tree *io_tree, *failure_tree;
8137         int i;
8138
8139         if (bio->bi_status)
8140                 goto end;
8141
8142         ASSERT(bio->bi_vcnt == 1);
8143         io_tree = &BTRFS_I(inode)->io_tree;
8144         failure_tree = &BTRFS_I(inode)->io_failure_tree;
8145         ASSERT(bio->bi_io_vec->bv_len == btrfs_inode_sectorsize(inode));
8146
8147         done->uptodate = 1;
8148         ASSERT(!bio_flagged(bio, BIO_CLONED));
8149         bio_for_each_segment_all(bvec, bio, i)
8150                 clean_io_failure(BTRFS_I(inode)->root->fs_info, failure_tree,
8151                                  io_tree, done->start, bvec->bv_page,
8152                                  btrfs_ino(BTRFS_I(inode)), 0);
8153 end:
8154         complete(&done->done);
8155         bio_put(bio);
8156 }
8157
8158 static int __btrfs_correct_data_nocsum(struct inode *inode,
8159                                        struct btrfs_io_bio *io_bio)
8160 {
8161         struct btrfs_fs_info *fs_info;
8162         struct bio_vec bvec;
8163         struct bvec_iter iter;
8164         struct btrfs_retry_complete done;
8165         u64 start;
8166         unsigned int pgoff;
8167         u32 sectorsize;
8168         int nr_sectors;
8169         int ret;
8170         int err = 0;
8171
8172         fs_info = BTRFS_I(inode)->root->fs_info;
8173         sectorsize = fs_info->sectorsize;
8174
8175         start = io_bio->logical;
8176         done.inode = inode;
8177         io_bio->bio.bi_iter = io_bio->iter;
8178
8179         bio_for_each_segment(bvec, &io_bio->bio, iter) {
8180                 nr_sectors = BTRFS_BYTES_TO_BLKS(fs_info, bvec.bv_len);
8181                 pgoff = bvec.bv_offset;
8182
8183 next_block_or_try_again:
8184                 done.uptodate = 0;
8185                 done.start = start;
8186                 init_completion(&done.done);
8187
8188                 ret = dio_read_error(inode, &io_bio->bio, bvec.bv_page,
8189                                 pgoff, start, start + sectorsize - 1,
8190                                 io_bio->mirror_num,
8191                                 btrfs_retry_endio_nocsum, &done);
8192                 if (ret) {
8193                         err = ret;
8194                         goto next;
8195                 }
8196
8197                 wait_for_completion_io(&done.done);
8198
8199                 if (!done.uptodate) {
8200                         /* We might have another mirror, so try again */
8201                         goto next_block_or_try_again;
8202                 }
8203
8204 next:
8205                 start += sectorsize;
8206
8207                 nr_sectors--;
8208                 if (nr_sectors) {
8209                         pgoff += sectorsize;
8210                         ASSERT(pgoff < PAGE_SIZE);
8211                         goto next_block_or_try_again;
8212                 }
8213         }
8214
8215         return err;
8216 }
8217
8218 static void btrfs_retry_endio(struct bio *bio)
8219 {
8220         struct btrfs_retry_complete *done = bio->bi_private;
8221         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
8222         struct extent_io_tree *io_tree, *failure_tree;
8223         struct inode *inode = done->inode;
8224         struct bio_vec *bvec;
8225         int uptodate;
8226         int ret;
8227         int i;
8228
8229         if (bio->bi_status)
8230                 goto end;
8231
8232         uptodate = 1;
8233
8234         ASSERT(bio->bi_vcnt == 1);
8235         ASSERT(bio->bi_io_vec->bv_len == btrfs_inode_sectorsize(done->inode));
8236
8237         io_tree = &BTRFS_I(inode)->io_tree;
8238         failure_tree = &BTRFS_I(inode)->io_failure_tree;
8239
8240         ASSERT(!bio_flagged(bio, BIO_CLONED));
8241         bio_for_each_segment_all(bvec, bio, i) {
8242                 ret = __readpage_endio_check(inode, io_bio, i, bvec->bv_page,
8243                                              bvec->bv_offset, done->start,
8244                                              bvec->bv_len);
8245                 if (!ret)
8246                         clean_io_failure(BTRFS_I(inode)->root->fs_info,
8247                                          failure_tree, io_tree, done->start,
8248                                          bvec->bv_page,
8249                                          btrfs_ino(BTRFS_I(inode)),
8250                                          bvec->bv_offset);
8251                 else
8252                         uptodate = 0;
8253         }
8254
8255         done->uptodate = uptodate;
8256 end:
8257         complete(&done->done);
8258         bio_put(bio);
8259 }
8260
8261 static blk_status_t __btrfs_subio_endio_read(struct inode *inode,
8262                 struct btrfs_io_bio *io_bio, blk_status_t err)
8263 {
8264         struct btrfs_fs_info *fs_info;
8265         struct bio_vec bvec;
8266         struct bvec_iter iter;
8267         struct btrfs_retry_complete done;
8268         u64 start;
8269         u64 offset = 0;
8270         u32 sectorsize;
8271         int nr_sectors;
8272         unsigned int pgoff;
8273         int csum_pos;
8274         bool uptodate = (err == 0);
8275         int ret;
8276
8277         fs_info = BTRFS_I(inode)->root->fs_info;
8278         sectorsize = fs_info->sectorsize;
8279
8280         err = 0;
8281         start = io_bio->logical;
8282         done.inode = inode;
8283         io_bio->bio.bi_iter = io_bio->iter;
8284
8285         bio_for_each_segment(bvec, &io_bio->bio, iter) {
8286                 nr_sectors = BTRFS_BYTES_TO_BLKS(fs_info, bvec.bv_len);
8287
8288                 pgoff = bvec.bv_offset;
8289 next_block:
8290                 if (uptodate) {
8291                         csum_pos = BTRFS_BYTES_TO_BLKS(fs_info, offset);
8292                         ret = __readpage_endio_check(inode, io_bio, csum_pos,
8293                                         bvec.bv_page, pgoff, start, sectorsize);
8294                         if (likely(!ret))
8295                                 goto next;
8296                 }
8297 try_again:
8298                 done.uptodate = 0;
8299                 done.start = start;
8300                 init_completion(&done.done);
8301
8302                 ret = dio_read_error(inode, &io_bio->bio, bvec.bv_page,
8303                                 pgoff, start, start + sectorsize - 1,
8304                                 io_bio->mirror_num,
8305                                 btrfs_retry_endio, &done);
8306                 if (ret) {
8307                         err = errno_to_blk_status(ret);
8308                         goto next;
8309                 }
8310
8311                 wait_for_completion_io(&done.done);
8312
8313                 if (!done.uptodate) {
8314                         /* We might have another mirror, so try again */
8315                         goto try_again;
8316                 }
8317 next:
8318                 offset += sectorsize;
8319                 start += sectorsize;
8320
8321                 ASSERT(nr_sectors);
8322
8323                 nr_sectors--;
8324                 if (nr_sectors) {
8325                         pgoff += sectorsize;
8326                         ASSERT(pgoff < PAGE_SIZE);
8327                         goto next_block;
8328                 }
8329         }
8330
8331         return err;
8332 }
8333
8334 static blk_status_t btrfs_subio_endio_read(struct inode *inode,
8335                 struct btrfs_io_bio *io_bio, blk_status_t err)
8336 {
8337         bool skip_csum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM;
8338
8339         if (skip_csum) {
8340                 if (unlikely(err))
8341                         return __btrfs_correct_data_nocsum(inode, io_bio);
8342                 else
8343                         return 0;
8344         } else {
8345                 return __btrfs_subio_endio_read(inode, io_bio, err);
8346         }
8347 }
8348
8349 static void btrfs_endio_direct_read(struct bio *bio)
8350 {
8351         struct btrfs_dio_private *dip = bio->bi_private;
8352         struct inode *inode = dip->inode;
8353         struct bio *dio_bio;
8354         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
8355         blk_status_t err = bio->bi_status;
8356
8357         if (dip->flags & BTRFS_DIO_ORIG_BIO_SUBMITTED) {
8358                 err = btrfs_subio_endio_read(inode, io_bio, err);
8359                 if (!err)
8360                         bio->bi_status = 0;
8361         }
8362
8363         unlock_extent(&BTRFS_I(inode)->io_tree, dip->logical_offset,
8364                       dip->logical_offset + dip->bytes - 1);
8365         dio_bio = dip->dio_bio;
8366
8367         kfree(dip);
8368
8369         dio_bio->bi_status = bio->bi_status;
8370         dio_end_io(dio_bio);
8371
8372         if (io_bio->end_io)
8373                 io_bio->end_io(io_bio, blk_status_to_errno(err));
8374         bio_put(bio);
8375 }
8376
8377 static void __endio_write_update_ordered(struct inode *inode,
8378                                          const u64 offset, const u64 bytes,
8379                                          const bool uptodate)
8380 {
8381         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
8382         struct btrfs_ordered_extent *ordered = NULL;
8383         struct btrfs_workqueue *wq;
8384         btrfs_work_func_t func;
8385         u64 ordered_offset = offset;
8386         u64 ordered_bytes = bytes;
8387         int ret;
8388
8389         if (btrfs_is_free_space_inode(BTRFS_I(inode))) {
8390                 wq = fs_info->endio_freespace_worker;
8391                 func = btrfs_freespace_write_helper;
8392         } else {
8393                 wq = fs_info->endio_write_workers;
8394                 func = btrfs_endio_write_helper;
8395         }
8396
8397 again:
8398         ret = btrfs_dec_test_first_ordered_pending(inode, &ordered,
8399                                                    &ordered_offset,
8400                                                    ordered_bytes,
8401                                                    uptodate);
8402         if (!ret)
8403                 goto out_test;
8404
8405         btrfs_init_work(&ordered->work, func, finish_ordered_fn, NULL, NULL);
8406         btrfs_queue_work(wq, &ordered->work);
8407 out_test:
8408         /*
8409          * our bio might span multiple ordered extents.  If we haven't
8410          * completed the accounting for the whole dio, go back and try again
8411          */
8412         if (ordered_offset < offset + bytes) {
8413                 ordered_bytes = offset + bytes - ordered_offset;
8414                 ordered = NULL;
8415                 goto again;
8416         }
8417 }
8418
8419 static void btrfs_endio_direct_write(struct bio *bio)
8420 {
8421         struct btrfs_dio_private *dip = bio->bi_private;
8422         struct bio *dio_bio = dip->dio_bio;
8423
8424         __endio_write_update_ordered(dip->inode, dip->logical_offset,
8425                                      dip->bytes, !bio->bi_status);
8426
8427         kfree(dip);
8428
8429         dio_bio->bi_status = bio->bi_status;
8430         dio_end_io(dio_bio);
8431         bio_put(bio);
8432 }
8433
8434 static blk_status_t __btrfs_submit_bio_start_direct_io(void *private_data,
8435                                     struct bio *bio, int mirror_num,
8436                                     unsigned long bio_flags, u64 offset)
8437 {
8438         struct inode *inode = private_data;
8439         blk_status_t ret;
8440         ret = btrfs_csum_one_bio(inode, bio, offset, 1);
8441         BUG_ON(ret); /* -ENOMEM */
8442         return 0;
8443 }
8444
8445 static void btrfs_end_dio_bio(struct bio *bio)
8446 {
8447         struct btrfs_dio_private *dip = bio->bi_private;
8448         blk_status_t err = bio->bi_status;
8449
8450         if (err)
8451                 btrfs_warn(BTRFS_I(dip->inode)->root->fs_info,
8452                            "direct IO failed ino %llu rw %d,%u sector %#Lx len %u err no %d",
8453                            btrfs_ino(BTRFS_I(dip->inode)), bio_op(bio),
8454                            bio->bi_opf,
8455                            (unsigned long long)bio->bi_iter.bi_sector,
8456                            bio->bi_iter.bi_size, err);
8457
8458         if (dip->subio_endio)
8459                 err = dip->subio_endio(dip->inode, btrfs_io_bio(bio), err);
8460
8461         if (err) {
8462                 dip->errors = 1;
8463
8464                 /*
8465                  * before atomic variable goto zero, we must make sure
8466                  * dip->errors is perceived to be set.
8467                  */
8468                 smp_mb__before_atomic();
8469         }
8470
8471         /* if there are more bios still pending for this dio, just exit */
8472         if (!atomic_dec_and_test(&dip->pending_bios))
8473                 goto out;
8474
8475         if (dip->errors) {
8476                 bio_io_error(dip->orig_bio);
8477         } else {
8478                 dip->dio_bio->bi_status = 0;
8479                 bio_endio(dip->orig_bio);
8480         }
8481 out:
8482         bio_put(bio);
8483 }
8484
8485 static inline blk_status_t btrfs_lookup_and_bind_dio_csum(struct inode *inode,
8486                                                  struct btrfs_dio_private *dip,
8487                                                  struct bio *bio,
8488                                                  u64 file_offset)
8489 {
8490         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
8491         struct btrfs_io_bio *orig_io_bio = btrfs_io_bio(dip->orig_bio);
8492         blk_status_t ret;
8493
8494         /*
8495          * We load all the csum data we need when we submit
8496          * the first bio to reduce the csum tree search and
8497          * contention.
8498          */
8499         if (dip->logical_offset == file_offset) {
8500                 ret = btrfs_lookup_bio_sums_dio(inode, dip->orig_bio,
8501                                                 file_offset);
8502                 if (ret)
8503                         return ret;
8504         }
8505
8506         if (bio == dip->orig_bio)
8507                 return 0;
8508
8509         file_offset -= dip->logical_offset;
8510         file_offset >>= inode->i_sb->s_blocksize_bits;
8511         io_bio->csum = (u8 *)(((u32 *)orig_io_bio->csum) + file_offset);
8512
8513         return 0;
8514 }
8515
8516 static inline int __btrfs_submit_dio_bio(struct bio *bio, struct inode *inode,
8517                                          u64 file_offset, int async_submit)
8518 {
8519         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
8520         struct btrfs_dio_private *dip = bio->bi_private;
8521         bool write = bio_op(bio) == REQ_OP_WRITE;
8522         blk_status_t ret;
8523
8524         if (async_submit)
8525                 async_submit = !atomic_read(&BTRFS_I(inode)->sync_writers);
8526
8527         bio_get(bio);
8528
8529         if (!write) {
8530                 ret = btrfs_bio_wq_end_io(fs_info, bio, BTRFS_WQ_ENDIO_DATA);
8531                 if (ret)
8532                         goto err;
8533         }
8534
8535         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)
8536                 goto map;
8537
8538         if (write && async_submit) {
8539                 ret = btrfs_wq_submit_bio(fs_info, bio, 0, 0,
8540                                           file_offset, inode,
8541                                           __btrfs_submit_bio_start_direct_io,
8542                                           __btrfs_submit_bio_done);
8543                 goto err;
8544         } else if (write) {
8545                 /*
8546                  * If we aren't doing async submit, calculate the csum of the
8547                  * bio now.
8548                  */
8549                 ret = btrfs_csum_one_bio(inode, bio, file_offset, 1);
8550                 if (ret)
8551                         goto err;
8552         } else {
8553                 ret = btrfs_lookup_and_bind_dio_csum(inode, dip, bio,
8554                                                      file_offset);
8555                 if (ret)
8556                         goto err;
8557         }
8558 map:
8559         ret = btrfs_map_bio(fs_info, bio, 0, async_submit);
8560 err:
8561         bio_put(bio);
8562         return ret;
8563 }
8564
8565 static int btrfs_submit_direct_hook(struct btrfs_dio_private *dip)
8566 {
8567         struct inode *inode = dip->inode;
8568         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
8569         struct bio *bio;
8570         struct bio *orig_bio = dip->orig_bio;
8571         u64 start_sector = orig_bio->bi_iter.bi_sector;
8572         u64 file_offset = dip->logical_offset;
8573         u64 map_length;
8574         int async_submit = 0;
8575         u64 submit_len;
8576         int clone_offset = 0;
8577         int clone_len;
8578         int ret;
8579
8580         map_length = orig_bio->bi_iter.bi_size;
8581         submit_len = map_length;
8582         ret = btrfs_map_block(fs_info, btrfs_op(orig_bio), start_sector << 9,
8583                               &map_length, NULL, 0);
8584         if (ret)
8585                 return -EIO;
8586
8587         if (map_length >= submit_len) {
8588                 bio = orig_bio;
8589                 dip->flags |= BTRFS_DIO_ORIG_BIO_SUBMITTED;
8590                 goto submit;
8591         }
8592
8593         /* async crcs make it difficult to collect full stripe writes. */
8594         if (btrfs_data_alloc_profile(fs_info) & BTRFS_BLOCK_GROUP_RAID56_MASK)
8595                 async_submit = 0;
8596         else
8597                 async_submit = 1;
8598
8599         /* bio split */
8600         ASSERT(map_length <= INT_MAX);
8601         atomic_inc(&dip->pending_bios);
8602         do {
8603                 clone_len = min_t(int, submit_len, map_length);
8604
8605                 /*
8606                  * This will never fail as it's passing GPF_NOFS and
8607                  * the allocation is backed by btrfs_bioset.
8608                  */
8609                 bio = btrfs_bio_clone_partial(orig_bio, clone_offset,
8610                                               clone_len);
8611                 bio->bi_private = dip;
8612                 bio->bi_end_io = btrfs_end_dio_bio;
8613                 btrfs_io_bio(bio)->logical = file_offset;
8614
8615                 ASSERT(submit_len >= clone_len);
8616                 submit_len -= clone_len;
8617                 if (submit_len == 0)
8618                         break;
8619
8620                 /*
8621                  * Increase the count before we submit the bio so we know
8622                  * the end IO handler won't happen before we increase the
8623                  * count. Otherwise, the dip might get freed before we're
8624                  * done setting it up.
8625                  */
8626                 atomic_inc(&dip->pending_bios);
8627
8628                 ret = __btrfs_submit_dio_bio(bio, inode, file_offset,
8629                                              async_submit);
8630                 if (ret) {
8631                         bio_put(bio);
8632                         atomic_dec(&dip->pending_bios);
8633                         goto out_err;
8634                 }
8635
8636                 clone_offset += clone_len;
8637                 start_sector += clone_len >> 9;
8638                 file_offset += clone_len;
8639
8640                 map_length = submit_len;
8641                 ret = btrfs_map_block(fs_info, btrfs_op(orig_bio),
8642                                       start_sector << 9, &map_length, NULL, 0);
8643                 if (ret)
8644                         goto out_err;
8645         } while (submit_len > 0);
8646
8647 submit:
8648         ret = __btrfs_submit_dio_bio(bio, inode, file_offset, async_submit);
8649         if (!ret)
8650                 return 0;
8651
8652         bio_put(bio);
8653 out_err:
8654         dip->errors = 1;
8655         /*
8656          * before atomic variable goto zero, we must
8657          * make sure dip->errors is perceived to be set.
8658          */
8659         smp_mb__before_atomic();
8660         if (atomic_dec_and_test(&dip->pending_bios))
8661                 bio_io_error(dip->orig_bio);
8662
8663         /* bio_end_io() will handle error, so we needn't return it */
8664         return 0;
8665 }
8666
8667 static void btrfs_submit_direct(struct bio *dio_bio, struct inode *inode,
8668                                 loff_t file_offset)
8669 {
8670         struct btrfs_dio_private *dip = NULL;
8671         struct bio *bio = NULL;
8672         struct btrfs_io_bio *io_bio;
8673         bool write = (bio_op(dio_bio) == REQ_OP_WRITE);
8674         int ret = 0;
8675
8676         bio = btrfs_bio_clone(dio_bio);
8677
8678         dip = kzalloc(sizeof(*dip), GFP_NOFS);
8679         if (!dip) {
8680                 ret = -ENOMEM;
8681                 goto free_ordered;
8682         }
8683
8684         dip->private = dio_bio->bi_private;
8685         dip->inode = inode;
8686         dip->logical_offset = file_offset;
8687         dip->bytes = dio_bio->bi_iter.bi_size;
8688         dip->disk_bytenr = (u64)dio_bio->bi_iter.bi_sector << 9;
8689         bio->bi_private = dip;
8690         dip->orig_bio = bio;
8691         dip->dio_bio = dio_bio;
8692         atomic_set(&dip->pending_bios, 0);
8693         io_bio = btrfs_io_bio(bio);
8694         io_bio->logical = file_offset;
8695
8696         if (write) {
8697                 bio->bi_end_io = btrfs_endio_direct_write;
8698         } else {
8699                 bio->bi_end_io = btrfs_endio_direct_read;
8700                 dip->subio_endio = btrfs_subio_endio_read;
8701         }
8702
8703         /*
8704          * Reset the range for unsubmitted ordered extents (to a 0 length range)
8705          * even if we fail to submit a bio, because in such case we do the
8706          * corresponding error handling below and it must not be done a second
8707          * time by btrfs_direct_IO().
8708          */
8709         if (write) {
8710                 struct btrfs_dio_data *dio_data = current->journal_info;
8711
8712                 dio_data->unsubmitted_oe_range_end = dip->logical_offset +
8713                         dip->bytes;
8714                 dio_data->unsubmitted_oe_range_start =
8715                         dio_data->unsubmitted_oe_range_end;
8716         }
8717
8718         ret = btrfs_submit_direct_hook(dip);
8719         if (!ret)
8720                 return;
8721
8722         if (io_bio->end_io)
8723                 io_bio->end_io(io_bio, ret);
8724
8725 free_ordered:
8726         /*
8727          * If we arrived here it means either we failed to submit the dip
8728          * or we either failed to clone the dio_bio or failed to allocate the
8729          * dip. If we cloned the dio_bio and allocated the dip, we can just
8730          * call bio_endio against our io_bio so that we get proper resource
8731          * cleanup if we fail to submit the dip, otherwise, we must do the
8732          * same as btrfs_endio_direct_[write|read] because we can't call these
8733          * callbacks - they require an allocated dip and a clone of dio_bio.
8734          */
8735         if (bio && dip) {
8736                 bio_io_error(bio);
8737                 /*
8738                  * The end io callbacks free our dip, do the final put on bio
8739                  * and all the cleanup and final put for dio_bio (through
8740                  * dio_end_io()).
8741                  */
8742                 dip = NULL;
8743                 bio = NULL;
8744         } else {
8745                 if (write)
8746                         __endio_write_update_ordered(inode,
8747                                                 file_offset,
8748                                                 dio_bio->bi_iter.bi_size,
8749                                                 false);
8750                 else
8751                         unlock_extent(&BTRFS_I(inode)->io_tree, file_offset,
8752                               file_offset + dio_bio->bi_iter.bi_size - 1);
8753
8754                 dio_bio->bi_status = BLK_STS_IOERR;
8755                 /*
8756                  * Releases and cleans up our dio_bio, no need to bio_put()
8757                  * nor bio_endio()/bio_io_error() against dio_bio.
8758                  */
8759                 dio_end_io(dio_bio);
8760         }
8761         if (bio)
8762                 bio_put(bio);
8763         kfree(dip);
8764 }
8765
8766 static ssize_t check_direct_IO(struct btrfs_fs_info *fs_info,
8767                                struct kiocb *iocb,
8768                                const struct iov_iter *iter, loff_t offset)
8769 {
8770         int seg;
8771         int i;
8772         unsigned int blocksize_mask = fs_info->sectorsize - 1;
8773         ssize_t retval = -EINVAL;
8774
8775         if (offset & blocksize_mask)
8776                 goto out;
8777
8778         if (iov_iter_alignment(iter) & blocksize_mask)
8779                 goto out;
8780
8781         /* If this is a write we don't need to check anymore */
8782         if (iov_iter_rw(iter) != READ || !iter_is_iovec(iter))
8783                 return 0;
8784         /*
8785          * Check to make sure we don't have duplicate iov_base's in this
8786          * iovec, if so return EINVAL, otherwise we'll get csum errors
8787          * when reading back.
8788          */
8789         for (seg = 0; seg < iter->nr_segs; seg++) {
8790                 for (i = seg + 1; i < iter->nr_segs; i++) {
8791                         if (iter->iov[seg].iov_base == iter->iov[i].iov_base)
8792                                 goto out;
8793                 }
8794         }
8795         retval = 0;
8796 out:
8797         return retval;
8798 }
8799
8800 static ssize_t btrfs_direct_IO(struct kiocb *iocb, struct iov_iter *iter)
8801 {
8802         struct file *file = iocb->ki_filp;
8803         struct inode *inode = file->f_mapping->host;
8804         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
8805         struct btrfs_dio_data dio_data = { 0 };
8806         struct extent_changeset *data_reserved = NULL;
8807         loff_t offset = iocb->ki_pos;
8808         size_t count = 0;
8809         int flags = 0;
8810         bool wakeup = true;
8811         bool relock = false;
8812         ssize_t ret;
8813
8814         if (check_direct_IO(fs_info, iocb, iter, offset))
8815                 return 0;
8816
8817         inode_dio_begin(inode);
8818
8819         /*
8820          * The generic stuff only does filemap_write_and_wait_range, which
8821          * isn't enough if we've written compressed pages to this area, so
8822          * we need to flush the dirty pages again to make absolutely sure
8823          * that any outstanding dirty pages are on disk.
8824          */
8825         count = iov_iter_count(iter);
8826         if (test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
8827                      &BTRFS_I(inode)->runtime_flags))
8828                 filemap_fdatawrite_range(inode->i_mapping, offset,
8829                                          offset + count - 1);
8830
8831         if (iov_iter_rw(iter) == WRITE) {
8832                 /*
8833                  * If the write DIO is beyond the EOF, we need update
8834                  * the isize, but it is protected by i_mutex. So we can
8835                  * not unlock the i_mutex at this case.
8836                  */
8837                 if (offset + count <= inode->i_size) {
8838                         dio_data.overwrite = 1;
8839                         inode_unlock(inode);
8840                         relock = true;
8841                 } else if (iocb->ki_flags & IOCB_NOWAIT) {
8842                         ret = -EAGAIN;
8843                         goto out;
8844                 }
8845                 ret = btrfs_delalloc_reserve_space(inode, &data_reserved,
8846                                                    offset, count);
8847                 if (ret)
8848                         goto out;
8849                 dio_data.outstanding_extents = count_max_extents(count);
8850
8851                 /*
8852                  * We need to know how many extents we reserved so that we can
8853                  * do the accounting properly if we go over the number we
8854                  * originally calculated.  Abuse current->journal_info for this.
8855                  */
8856                 dio_data.reserve = round_up(count,
8857                                             fs_info->sectorsize);
8858                 dio_data.unsubmitted_oe_range_start = (u64)offset;
8859                 dio_data.unsubmitted_oe_range_end = (u64)offset;
8860                 current->journal_info = &dio_data;
8861                 down_read(&BTRFS_I(inode)->dio_sem);
8862         } else if (test_bit(BTRFS_INODE_READDIO_NEED_LOCK,
8863                                      &BTRFS_I(inode)->runtime_flags)) {
8864                 inode_dio_end(inode);
8865                 flags = DIO_LOCKING | DIO_SKIP_HOLES;
8866                 wakeup = false;
8867         }
8868
8869         ret = __blockdev_direct_IO(iocb, inode,
8870                                    fs_info->fs_devices->latest_bdev,
8871                                    iter, btrfs_get_blocks_direct, NULL,
8872                                    btrfs_submit_direct, flags);
8873         if (iov_iter_rw(iter) == WRITE) {
8874                 up_read(&BTRFS_I(inode)->dio_sem);
8875                 current->journal_info = NULL;
8876                 if (ret < 0 && ret != -EIOCBQUEUED) {
8877                         if (dio_data.reserve)
8878                                 btrfs_delalloc_release_space(inode, data_reserved,
8879                                         offset, dio_data.reserve);
8880                         /*
8881                          * On error we might have left some ordered extents
8882                          * without submitting corresponding bios for them, so
8883                          * cleanup them up to avoid other tasks getting them
8884                          * and waiting for them to complete forever.
8885                          */
8886                         if (dio_data.unsubmitted_oe_range_start <
8887                             dio_data.unsubmitted_oe_range_end)
8888                                 __endio_write_update_ordered(inode,
8889                                         dio_data.unsubmitted_oe_range_start,
8890                                         dio_data.unsubmitted_oe_range_end -
8891                                         dio_data.unsubmitted_oe_range_start,
8892                                         false);
8893                 } else if (ret >= 0 && (size_t)ret < count)
8894                         btrfs_delalloc_release_space(inode, data_reserved,
8895                                         offset, count - (size_t)ret);
8896         }
8897 out:
8898         if (wakeup)
8899                 inode_dio_end(inode);
8900         if (relock)
8901                 inode_lock(inode);
8902
8903         extent_changeset_free(data_reserved);
8904         return ret;
8905 }
8906
8907 #define BTRFS_FIEMAP_FLAGS      (FIEMAP_FLAG_SYNC)
8908
8909 static int btrfs_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
8910                 __u64 start, __u64 len)
8911 {
8912         int     ret;
8913
8914         ret = fiemap_check_flags(fieinfo, BTRFS_FIEMAP_FLAGS);
8915         if (ret)
8916                 return ret;
8917
8918         return extent_fiemap(inode, fieinfo, start, len, btrfs_get_extent_fiemap);
8919 }
8920
8921 int btrfs_readpage(struct file *file, struct page *page)
8922 {
8923         struct extent_io_tree *tree;
8924         tree = &BTRFS_I(page->mapping->host)->io_tree;
8925         return extent_read_full_page(tree, page, btrfs_get_extent, 0);
8926 }
8927
8928 static int btrfs_writepage(struct page *page, struct writeback_control *wbc)
8929 {
8930         struct extent_io_tree *tree;
8931         struct inode *inode = page->mapping->host;
8932         int ret;
8933
8934         if (current->flags & PF_MEMALLOC) {
8935                 redirty_page_for_writepage(wbc, page);
8936                 unlock_page(page);
8937                 return 0;
8938         }
8939
8940         /*
8941          * If we are under memory pressure we will call this directly from the
8942          * VM, we need to make sure we have the inode referenced for the ordered
8943          * extent.  If not just return like we didn't do anything.
8944          */
8945         if (!igrab(inode)) {
8946                 redirty_page_for_writepage(wbc, page);
8947                 return AOP_WRITEPAGE_ACTIVATE;
8948         }
8949         tree = &BTRFS_I(page->mapping->host)->io_tree;
8950         ret = extent_write_full_page(tree, page, btrfs_get_extent, wbc);
8951         btrfs_add_delayed_iput(inode);
8952         return ret;
8953 }
8954
8955 static int btrfs_writepages(struct address_space *mapping,
8956                             struct writeback_control *wbc)
8957 {
8958         struct extent_io_tree *tree;
8959
8960         tree = &BTRFS_I(mapping->host)->io_tree;
8961         return extent_writepages(tree, mapping, btrfs_get_extent, wbc);
8962 }
8963
8964 static int
8965 btrfs_readpages(struct file *file, struct address_space *mapping,
8966                 struct list_head *pages, unsigned nr_pages)
8967 {
8968         struct extent_io_tree *tree;
8969         tree = &BTRFS_I(mapping->host)->io_tree;
8970         return extent_readpages(tree, mapping, pages, nr_pages,
8971                                 btrfs_get_extent);
8972 }
8973 static int __btrfs_releasepage(struct page *page, gfp_t gfp_flags)
8974 {
8975         struct extent_io_tree *tree;
8976         struct extent_map_tree *map;
8977         int ret;
8978
8979         tree = &BTRFS_I(page->mapping->host)->io_tree;
8980         map = &BTRFS_I(page->mapping->host)->extent_tree;
8981         ret = try_release_extent_mapping(map, tree, page, gfp_flags);
8982         if (ret == 1) {
8983                 ClearPagePrivate(page);
8984                 set_page_private(page, 0);
8985                 put_page(page);
8986         }
8987         return ret;
8988 }
8989
8990 static int btrfs_releasepage(struct page *page, gfp_t gfp_flags)
8991 {
8992         if (PageWriteback(page) || PageDirty(page))
8993                 return 0;
8994         return __btrfs_releasepage(page, gfp_flags);
8995 }
8996
8997 static void btrfs_invalidatepage(struct page *page, unsigned int offset,
8998                                  unsigned int length)
8999 {
9000         struct inode *inode = page->mapping->host;
9001         struct extent_io_tree *tree;
9002         struct btrfs_ordered_extent *ordered;
9003         struct extent_state *cached_state = NULL;
9004         u64 page_start = page_offset(page);
9005         u64 page_end = page_start + PAGE_SIZE - 1;
9006         u64 start;
9007         u64 end;
9008         int inode_evicting = inode->i_state & I_FREEING;
9009
9010         /*
9011          * we have the page locked, so new writeback can't start,
9012          * and the dirty bit won't be cleared while we are here.
9013          *
9014          * Wait for IO on this page so that we can safely clear
9015          * the PagePrivate2 bit and do ordered accounting
9016          */
9017         wait_on_page_writeback(page);
9018
9019         tree = &BTRFS_I(inode)->io_tree;
9020         if (offset) {
9021                 btrfs_releasepage(page, GFP_NOFS);
9022                 return;
9023         }
9024
9025         if (!inode_evicting)
9026                 lock_extent_bits(tree, page_start, page_end, &cached_state);
9027 again:
9028         start = page_start;
9029         ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), start,
9030                                         page_end - start + 1);
9031         if (ordered) {
9032                 end = min(page_end, ordered->file_offset + ordered->len - 1);
9033                 /*
9034                  * IO on this page will never be started, so we need
9035                  * to account for any ordered extents now
9036                  */
9037                 if (!inode_evicting)
9038                         clear_extent_bit(tree, start, end,
9039                                          EXTENT_DIRTY | EXTENT_DELALLOC |
9040                                          EXTENT_DELALLOC_NEW |
9041                                          EXTENT_LOCKED | EXTENT_DO_ACCOUNTING |
9042                                          EXTENT_DEFRAG, 1, 0, &cached_state,
9043                                          GFP_NOFS);
9044                 /*
9045                  * whoever cleared the private bit is responsible
9046                  * for the finish_ordered_io
9047                  */
9048                 if (TestClearPagePrivate2(page)) {
9049                         struct btrfs_ordered_inode_tree *tree;
9050                         u64 new_len;
9051
9052                         tree = &BTRFS_I(inode)->ordered_tree;
9053
9054                         spin_lock_irq(&tree->lock);
9055                         set_bit(BTRFS_ORDERED_TRUNCATED, &ordered->flags);
9056                         new_len = start - ordered->file_offset;
9057                         if (new_len < ordered->truncated_len)
9058                                 ordered->truncated_len = new_len;
9059                         spin_unlock_irq(&tree->lock);
9060
9061                         if (btrfs_dec_test_ordered_pending(inode, &ordered,
9062                                                            start,
9063                                                            end - start + 1, 1))
9064                                 btrfs_finish_ordered_io(ordered);
9065                 }
9066                 btrfs_put_ordered_extent(ordered);
9067                 if (!inode_evicting) {
9068                         cached_state = NULL;
9069                         lock_extent_bits(tree, start, end,
9070                                          &cached_state);
9071                 }
9072
9073                 start = end + 1;
9074                 if (start < page_end)
9075                         goto again;
9076         }
9077
9078         /*
9079          * Qgroup reserved space handler
9080          * Page here will be either
9081          * 1) Already written to disk
9082          *    In this case, its reserved space is released from data rsv map
9083          *    and will be freed by delayed_ref handler finally.
9084          *    So even we call qgroup_free_data(), it won't decrease reserved
9085          *    space.
9086          * 2) Not written to disk
9087          *    This means the reserved space should be freed here. However,
9088          *    if a truncate invalidates the page (by clearing PageDirty)
9089          *    and the page is accounted for while allocating extent
9090          *    in btrfs_check_data_free_space() we let delayed_ref to
9091          *    free the entire extent.
9092          */
9093         if (PageDirty(page))
9094                 btrfs_qgroup_free_data(inode, NULL, page_start, PAGE_SIZE);
9095         if (!inode_evicting) {
9096                 clear_extent_bit(tree, page_start, page_end,
9097                                  EXTENT_LOCKED | EXTENT_DIRTY |
9098                                  EXTENT_DELALLOC | EXTENT_DELALLOC_NEW |
9099                                  EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG, 1, 1,
9100                                  &cached_state, GFP_NOFS);
9101
9102                 __btrfs_releasepage(page, GFP_NOFS);
9103         }
9104
9105         ClearPageChecked(page);
9106         if (PagePrivate(page)) {
9107                 ClearPagePrivate(page);
9108                 set_page_private(page, 0);
9109                 put_page(page);
9110         }
9111 }
9112
9113 /*
9114  * btrfs_page_mkwrite() is not allowed to change the file size as it gets
9115  * called from a page fault handler when a page is first dirtied. Hence we must
9116  * be careful to check for EOF conditions here. We set the page up correctly
9117  * for a written page which means we get ENOSPC checking when writing into
9118  * holes and correct delalloc and unwritten extent mapping on filesystems that
9119  * support these features.
9120  *
9121  * We are not allowed to take the i_mutex here so we have to play games to
9122  * protect against truncate races as the page could now be beyond EOF.  Because
9123  * vmtruncate() writes the inode size before removing pages, once we have the
9124  * page lock we can determine safely if the page is beyond EOF. If it is not
9125  * beyond EOF, then the page is guaranteed safe against truncation until we
9126  * unlock the page.
9127  */
9128 int btrfs_page_mkwrite(struct vm_fault *vmf)
9129 {
9130         struct page *page = vmf->page;
9131         struct inode *inode = file_inode(vmf->vma->vm_file);
9132         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
9133         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
9134         struct btrfs_ordered_extent *ordered;
9135         struct extent_state *cached_state = NULL;
9136         struct extent_changeset *data_reserved = NULL;
9137         char *kaddr;
9138         unsigned long zero_start;
9139         loff_t size;
9140         int ret;
9141         int reserved = 0;
9142         u64 reserved_space;
9143         u64 page_start;
9144         u64 page_end;
9145         u64 end;
9146
9147         reserved_space = PAGE_SIZE;
9148
9149         sb_start_pagefault(inode->i_sb);
9150         page_start = page_offset(page);
9151         page_end = page_start + PAGE_SIZE - 1;
9152         end = page_end;
9153
9154         /*
9155          * Reserving delalloc space after obtaining the page lock can lead to
9156          * deadlock. For example, if a dirty page is locked by this function
9157          * and the call to btrfs_delalloc_reserve_space() ends up triggering
9158          * dirty page write out, then the btrfs_writepage() function could
9159          * end up waiting indefinitely to get a lock on the page currently
9160          * being processed by btrfs_page_mkwrite() function.
9161          */
9162         ret = btrfs_delalloc_reserve_space(inode, &data_reserved, page_start,
9163                                            reserved_space);
9164         if (!ret) {
9165                 ret = file_update_time(vmf->vma->vm_file);
9166                 reserved = 1;
9167         }
9168         if (ret) {
9169                 if (ret == -ENOMEM)
9170                         ret = VM_FAULT_OOM;
9171                 else /* -ENOSPC, -EIO, etc */
9172                         ret = VM_FAULT_SIGBUS;
9173                 if (reserved)
9174                         goto out;
9175                 goto out_noreserve;
9176         }
9177
9178         ret = VM_FAULT_NOPAGE; /* make the VM retry the fault */
9179 again:
9180         lock_page(page);
9181         size = i_size_read(inode);
9182
9183         if ((page->mapping != inode->i_mapping) ||
9184             (page_start >= size)) {
9185                 /* page got truncated out from underneath us */
9186                 goto out_unlock;
9187         }
9188         wait_on_page_writeback(page);
9189
9190         lock_extent_bits(io_tree, page_start, page_end, &cached_state);
9191         set_page_extent_mapped(page);
9192
9193         /*
9194          * we can't set the delalloc bits if there are pending ordered
9195          * extents.  Drop our locks and wait for them to finish
9196          */
9197         ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), page_start,
9198                         PAGE_SIZE);
9199         if (ordered) {
9200                 unlock_extent_cached(io_tree, page_start, page_end,
9201                                      &cached_state, GFP_NOFS);
9202                 unlock_page(page);
9203                 btrfs_start_ordered_extent(inode, ordered, 1);
9204                 btrfs_put_ordered_extent(ordered);
9205                 goto again;
9206         }
9207
9208         if (page->index == ((size - 1) >> PAGE_SHIFT)) {
9209                 reserved_space = round_up(size - page_start,
9210                                           fs_info->sectorsize);
9211                 if (reserved_space < PAGE_SIZE) {
9212                         end = page_start + reserved_space - 1;
9213                         spin_lock(&BTRFS_I(inode)->lock);
9214                         BTRFS_I(inode)->outstanding_extents++;
9215                         spin_unlock(&BTRFS_I(inode)->lock);
9216                         btrfs_delalloc_release_space(inode, data_reserved,
9217                                         page_start, PAGE_SIZE - reserved_space);
9218                 }
9219         }
9220
9221         /*
9222          * page_mkwrite gets called when the page is firstly dirtied after it's
9223          * faulted in, but write(2) could also dirty a page and set delalloc
9224          * bits, thus in this case for space account reason, we still need to
9225          * clear any delalloc bits within this page range since we have to
9226          * reserve data&meta space before lock_page() (see above comments).
9227          */
9228         clear_extent_bit(&BTRFS_I(inode)->io_tree, page_start, end,
9229                           EXTENT_DIRTY | EXTENT_DELALLOC |
9230                           EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG,
9231                           0, 0, &cached_state, GFP_NOFS);
9232
9233         ret = btrfs_set_extent_delalloc(inode, page_start, end,
9234                                         &cached_state, 0);
9235         if (ret) {
9236                 unlock_extent_cached(io_tree, page_start, page_end,
9237                                      &cached_state, GFP_NOFS);
9238                 ret = VM_FAULT_SIGBUS;
9239                 goto out_unlock;
9240         }
9241         ret = 0;
9242
9243         /* page is wholly or partially inside EOF */
9244         if (page_start + PAGE_SIZE > size)
9245                 zero_start = size & ~PAGE_MASK;
9246         else
9247                 zero_start = PAGE_SIZE;
9248
9249         if (zero_start != PAGE_SIZE) {
9250                 kaddr = kmap(page);
9251                 memset(kaddr + zero_start, 0, PAGE_SIZE - zero_start);
9252                 flush_dcache_page(page);
9253                 kunmap(page);
9254         }
9255         ClearPageChecked(page);
9256         set_page_dirty(page);
9257         SetPageUptodate(page);
9258
9259         BTRFS_I(inode)->last_trans = fs_info->generation;
9260         BTRFS_I(inode)->last_sub_trans = BTRFS_I(inode)->root->log_transid;
9261         BTRFS_I(inode)->last_log_commit = BTRFS_I(inode)->root->last_log_commit;
9262
9263         unlock_extent_cached(io_tree, page_start, page_end, &cached_state, GFP_NOFS);
9264
9265 out_unlock:
9266         if (!ret) {
9267                 sb_end_pagefault(inode->i_sb);
9268                 extent_changeset_free(data_reserved);
9269                 return VM_FAULT_LOCKED;
9270         }
9271         unlock_page(page);
9272 out:
9273         btrfs_delalloc_release_space(inode, data_reserved, page_start,
9274                                      reserved_space);
9275 out_noreserve:
9276         sb_end_pagefault(inode->i_sb);
9277         extent_changeset_free(data_reserved);
9278         return ret;
9279 }
9280
9281 static int btrfs_truncate(struct inode *inode)
9282 {
9283         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
9284         struct btrfs_root *root = BTRFS_I(inode)->root;
9285         struct btrfs_block_rsv *rsv;
9286         int ret = 0;
9287         int err = 0;
9288         struct btrfs_trans_handle *trans;
9289         u64 mask = fs_info->sectorsize - 1;
9290         u64 min_size = btrfs_calc_trunc_metadata_size(fs_info, 1);
9291
9292         ret = btrfs_wait_ordered_range(inode, inode->i_size & (~mask),
9293                                        (u64)-1);
9294         if (ret)
9295                 return ret;
9296
9297         /*
9298          * Yes ladies and gentlemen, this is indeed ugly.  The fact is we have
9299          * 3 things going on here
9300          *
9301          * 1) We need to reserve space for our orphan item and the space to
9302          * delete our orphan item.  Lord knows we don't want to have a dangling
9303          * orphan item because we didn't reserve space to remove it.
9304          *
9305          * 2) We need to reserve space to update our inode.
9306          *
9307          * 3) We need to have something to cache all the space that is going to
9308          * be free'd up by the truncate operation, but also have some slack
9309          * space reserved in case it uses space during the truncate (thank you
9310          * very much snapshotting).
9311          *
9312          * And we need these to all be separate.  The fact is we can use a lot of
9313          * space doing the truncate, and we have no earthly idea how much space
9314          * we will use, so we need the truncate reservation to be separate so it
9315          * doesn't end up using space reserved for updating the inode or
9316          * removing the orphan item.  We also need to be able to stop the
9317          * transaction and start a new one, which means we need to be able to
9318          * update the inode several times, and we have no idea of knowing how
9319          * many times that will be, so we can't just reserve 1 item for the
9320          * entirety of the operation, so that has to be done separately as well.
9321          * Then there is the orphan item, which does indeed need to be held on
9322          * to for the whole operation, and we need nobody to touch this reserved
9323          * space except the orphan code.
9324          *
9325          * So that leaves us with
9326          *
9327          * 1) root->orphan_block_rsv - for the orphan deletion.
9328          * 2) rsv - for the truncate reservation, which we will steal from the
9329          * transaction reservation.
9330          * 3) fs_info->trans_block_rsv - this will have 1 items worth left for
9331          * updating the inode.
9332          */
9333         rsv = btrfs_alloc_block_rsv(fs_info, BTRFS_BLOCK_RSV_TEMP);
9334         if (!rsv)
9335                 return -ENOMEM;
9336         rsv->size = min_size;
9337         rsv->failfast = 1;
9338
9339         /*
9340          * 1 for the truncate slack space
9341          * 1 for updating the inode.
9342          */
9343         trans = btrfs_start_transaction(root, 2);
9344         if (IS_ERR(trans)) {
9345                 err = PTR_ERR(trans);
9346                 goto out;
9347         }
9348
9349         /* Migrate the slack space for the truncate to our reserve */
9350         ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv, rsv,
9351                                       min_size, 0);
9352         BUG_ON(ret);
9353
9354         /*
9355          * So if we truncate and then write and fsync we normally would just
9356          * write the extents that changed, which is a problem if we need to
9357          * first truncate that entire inode.  So set this flag so we write out
9358          * all of the extents in the inode to the sync log so we're completely
9359          * safe.
9360          */
9361         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &BTRFS_I(inode)->runtime_flags);
9362         trans->block_rsv = rsv;
9363
9364         while (1) {
9365                 ret = btrfs_truncate_inode_items(trans, root, inode,
9366                                                  inode->i_size,
9367                                                  BTRFS_EXTENT_DATA_KEY);
9368                 if (ret != -ENOSPC && ret != -EAGAIN) {
9369                         err = ret;
9370                         break;
9371                 }
9372
9373                 trans->block_rsv = &fs_info->trans_block_rsv;
9374                 ret = btrfs_update_inode(trans, root, inode);
9375                 if (ret) {
9376                         err = ret;
9377                         break;
9378                 }
9379
9380                 btrfs_end_transaction(trans);
9381                 btrfs_btree_balance_dirty(fs_info);
9382
9383                 trans = btrfs_start_transaction(root, 2);
9384                 if (IS_ERR(trans)) {
9385                         ret = err = PTR_ERR(trans);
9386                         trans = NULL;
9387                         break;
9388                 }
9389
9390                 btrfs_block_rsv_release(fs_info, rsv, -1);
9391                 ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv,
9392                                               rsv, min_size, 0);
9393                 BUG_ON(ret);    /* shouldn't happen */
9394                 trans->block_rsv = rsv;
9395         }
9396
9397         if (ret == 0 && inode->i_nlink > 0) {
9398                 trans->block_rsv = root->orphan_block_rsv;
9399                 ret = btrfs_orphan_del(trans, BTRFS_I(inode));
9400                 if (ret)
9401                         err = ret;
9402         }
9403
9404         if (trans) {
9405                 trans->block_rsv = &fs_info->trans_block_rsv;
9406                 ret = btrfs_update_inode(trans, root, inode);
9407                 if (ret && !err)
9408                         err = ret;
9409
9410                 ret = btrfs_end_transaction(trans);
9411                 btrfs_btree_balance_dirty(fs_info);
9412         }
9413 out:
9414         btrfs_free_block_rsv(fs_info, rsv);
9415
9416         if (ret && !err)
9417                 err = ret;
9418
9419         return err;
9420 }
9421
9422 /*
9423  * create a new subvolume directory/inode (helper for the ioctl).
9424  */
9425 int btrfs_create_subvol_root(struct btrfs_trans_handle *trans,
9426                              struct btrfs_root *new_root,
9427                              struct btrfs_root *parent_root,
9428                              u64 new_dirid)
9429 {
9430         struct inode *inode;
9431         int err;
9432         u64 index = 0;
9433
9434         inode = btrfs_new_inode(trans, new_root, NULL, "..", 2,
9435                                 new_dirid, new_dirid,
9436                                 S_IFDIR | (~current_umask() & S_IRWXUGO),
9437                                 &index);
9438         if (IS_ERR(inode))
9439                 return PTR_ERR(inode);
9440         inode->i_op = &btrfs_dir_inode_operations;
9441         inode->i_fop = &btrfs_dir_file_operations;
9442
9443         set_nlink(inode, 1);
9444         btrfs_i_size_write(BTRFS_I(inode), 0);
9445         unlock_new_inode(inode);
9446
9447         err = btrfs_subvol_inherit_props(trans, new_root, parent_root);
9448         if (err)
9449                 btrfs_err(new_root->fs_info,
9450                           "error inheriting subvolume %llu properties: %d",
9451                           new_root->root_key.objectid, err);
9452
9453         err = btrfs_update_inode(trans, new_root, inode);
9454
9455         iput(inode);
9456         return err;
9457 }
9458
9459 struct inode *btrfs_alloc_inode(struct super_block *sb)
9460 {
9461         struct btrfs_inode *ei;
9462         struct inode *inode;
9463
9464         ei = kmem_cache_alloc(btrfs_inode_cachep, GFP_NOFS);
9465         if (!ei)
9466                 return NULL;
9467
9468         ei->root = NULL;
9469         ei->generation = 0;
9470         ei->last_trans = 0;
9471         ei->last_sub_trans = 0;
9472         ei->logged_trans = 0;
9473         ei->delalloc_bytes = 0;
9474         ei->new_delalloc_bytes = 0;
9475         ei->defrag_bytes = 0;
9476         ei->disk_i_size = 0;
9477         ei->flags = 0;
9478         ei->csum_bytes = 0;
9479         ei->index_cnt = (u64)-1;
9480         ei->dir_index = 0;
9481         ei->last_unlink_trans = 0;
9482         ei->last_log_commit = 0;
9483         ei->delayed_iput_count = 0;
9484
9485         spin_lock_init(&ei->lock);
9486         ei->outstanding_extents = 0;
9487         ei->reserved_extents = 0;
9488
9489         ei->runtime_flags = 0;
9490         ei->prop_compress = BTRFS_COMPRESS_NONE;
9491         ei->defrag_compress = BTRFS_COMPRESS_NONE;
9492
9493         ei->delayed_node = NULL;
9494
9495         ei->i_otime.tv_sec = 0;
9496         ei->i_otime.tv_nsec = 0;
9497
9498         inode = &ei->vfs_inode;
9499         extent_map_tree_init(&ei->extent_tree);
9500         extent_io_tree_init(&ei->io_tree, inode);
9501         extent_io_tree_init(&ei->io_failure_tree, inode);
9502         ei->io_tree.track_uptodate = 1;
9503         ei->io_failure_tree.track_uptodate = 1;
9504         atomic_set(&ei->sync_writers, 0);
9505         mutex_init(&ei->log_mutex);
9506         mutex_init(&ei->delalloc_mutex);
9507         btrfs_ordered_inode_tree_init(&ei->ordered_tree);
9508         INIT_LIST_HEAD(&ei->delalloc_inodes);
9509         INIT_LIST_HEAD(&ei->delayed_iput);
9510         RB_CLEAR_NODE(&ei->rb_node);
9511         init_rwsem(&ei->dio_sem);
9512
9513         return inode;
9514 }
9515
9516 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
9517 void btrfs_test_destroy_inode(struct inode *inode)
9518 {
9519         btrfs_drop_extent_cache(BTRFS_I(inode), 0, (u64)-1, 0);
9520         kmem_cache_free(btrfs_inode_cachep, BTRFS_I(inode));
9521 }
9522 #endif
9523
9524 static void btrfs_i_callback(struct rcu_head *head)
9525 {
9526         struct inode *inode = container_of(head, struct inode, i_rcu);
9527         kmem_cache_free(btrfs_inode_cachep, BTRFS_I(inode));
9528 }
9529
9530 void btrfs_destroy_inode(struct inode *inode)
9531 {
9532         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
9533         struct btrfs_ordered_extent *ordered;
9534         struct btrfs_root *root = BTRFS_I(inode)->root;
9535
9536         WARN_ON(!hlist_empty(&inode->i_dentry));
9537         WARN_ON(inode->i_data.nrpages);
9538         WARN_ON(BTRFS_I(inode)->outstanding_extents);
9539         WARN_ON(BTRFS_I(inode)->reserved_extents);
9540         WARN_ON(BTRFS_I(inode)->delalloc_bytes);
9541         WARN_ON(BTRFS_I(inode)->new_delalloc_bytes);
9542         WARN_ON(BTRFS_I(inode)->csum_bytes);
9543         WARN_ON(BTRFS_I(inode)->defrag_bytes);
9544
9545         /*
9546          * This can happen where we create an inode, but somebody else also
9547          * created the same inode and we need to destroy the one we already
9548          * created.
9549          */
9550         if (!root)
9551                 goto free;
9552
9553         if (test_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
9554                      &BTRFS_I(inode)->runtime_flags)) {
9555                 btrfs_info(fs_info, "inode %llu still on the orphan list",
9556                            btrfs_ino(BTRFS_I(inode)));
9557                 atomic_dec(&root->orphan_inodes);
9558         }
9559
9560         while (1) {
9561                 ordered = btrfs_lookup_first_ordered_extent(inode, (u64)-1);
9562                 if (!ordered)
9563                         break;
9564                 else {
9565                         btrfs_err(fs_info,
9566                                   "found ordered extent %llu %llu on inode cleanup",
9567                                   ordered->file_offset, ordered->len);
9568                         btrfs_remove_ordered_extent(inode, ordered);
9569                         btrfs_put_ordered_extent(ordered);
9570                         btrfs_put_ordered_extent(ordered);
9571                 }
9572         }
9573         btrfs_qgroup_check_reserved_leak(inode);
9574         inode_tree_del(inode);
9575         btrfs_drop_extent_cache(BTRFS_I(inode), 0, (u64)-1, 0);
9576 free:
9577         call_rcu(&inode->i_rcu, btrfs_i_callback);
9578 }
9579
9580 int btrfs_drop_inode(struct inode *inode)
9581 {
9582         struct btrfs_root *root = BTRFS_I(inode)->root;
9583
9584         if (root == NULL)
9585                 return 1;
9586
9587         /* the snap/subvol tree is on deleting */
9588         if (btrfs_root_refs(&root->root_item) == 0)
9589                 return 1;
9590         else
9591                 return generic_drop_inode(inode);
9592 }
9593
9594 static void init_once(void *foo)
9595 {
9596         struct btrfs_inode *ei = (struct btrfs_inode *) foo;
9597
9598         inode_init_once(&ei->vfs_inode);
9599 }
9600
9601 void btrfs_destroy_cachep(void)
9602 {
9603         /*
9604          * Make sure all delayed rcu free inodes are flushed before we
9605          * destroy cache.
9606          */
9607         rcu_barrier();
9608         kmem_cache_destroy(btrfs_inode_cachep);
9609         kmem_cache_destroy(btrfs_trans_handle_cachep);
9610         kmem_cache_destroy(btrfs_path_cachep);
9611         kmem_cache_destroy(btrfs_free_space_cachep);
9612 }
9613
9614 int btrfs_init_cachep(void)
9615 {
9616         btrfs_inode_cachep = kmem_cache_create("btrfs_inode",
9617                         sizeof(struct btrfs_inode), 0,
9618                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD | SLAB_ACCOUNT,
9619                         init_once);
9620         if (!btrfs_inode_cachep)
9621                 goto fail;
9622
9623         btrfs_trans_handle_cachep = kmem_cache_create("btrfs_trans_handle",
9624                         sizeof(struct btrfs_trans_handle), 0,
9625                         SLAB_TEMPORARY | SLAB_MEM_SPREAD, NULL);
9626         if (!btrfs_trans_handle_cachep)
9627                 goto fail;
9628
9629         btrfs_path_cachep = kmem_cache_create("btrfs_path",
9630                         sizeof(struct btrfs_path), 0,
9631                         SLAB_MEM_SPREAD, NULL);
9632         if (!btrfs_path_cachep)
9633                 goto fail;
9634
9635         btrfs_free_space_cachep = kmem_cache_create("btrfs_free_space",
9636                         sizeof(struct btrfs_free_space), 0,
9637                         SLAB_MEM_SPREAD, NULL);
9638         if (!btrfs_free_space_cachep)
9639                 goto fail;
9640
9641         return 0;
9642 fail:
9643         btrfs_destroy_cachep();
9644         return -ENOMEM;
9645 }
9646
9647 static int btrfs_getattr(const struct path *path, struct kstat *stat,
9648                          u32 request_mask, unsigned int flags)
9649 {
9650         u64 delalloc_bytes;
9651         struct inode *inode = d_inode(path->dentry);
9652         u32 blocksize = inode->i_sb->s_blocksize;
9653         u32 bi_flags = BTRFS_I(inode)->flags;
9654
9655         stat->result_mask |= STATX_BTIME;
9656         stat->btime.tv_sec = BTRFS_I(inode)->i_otime.tv_sec;
9657         stat->btime.tv_nsec = BTRFS_I(inode)->i_otime.tv_nsec;
9658         if (bi_flags & BTRFS_INODE_APPEND)
9659                 stat->attributes |= STATX_ATTR_APPEND;
9660         if (bi_flags & BTRFS_INODE_COMPRESS)
9661                 stat->attributes |= STATX_ATTR_COMPRESSED;
9662         if (bi_flags & BTRFS_INODE_IMMUTABLE)
9663                 stat->attributes |= STATX_ATTR_IMMUTABLE;
9664         if (bi_flags & BTRFS_INODE_NODUMP)
9665                 stat->attributes |= STATX_ATTR_NODUMP;
9666
9667         stat->attributes_mask |= (STATX_ATTR_APPEND |
9668                                   STATX_ATTR_COMPRESSED |
9669                                   STATX_ATTR_IMMUTABLE |
9670                                   STATX_ATTR_NODUMP);
9671
9672         generic_fillattr(inode, stat);
9673         stat->dev = BTRFS_I(inode)->root->anon_dev;
9674
9675         spin_lock(&BTRFS_I(inode)->lock);
9676         delalloc_bytes = BTRFS_I(inode)->new_delalloc_bytes;
9677         spin_unlock(&BTRFS_I(inode)->lock);
9678         stat->blocks = (ALIGN(inode_get_bytes(inode), blocksize) +
9679                         ALIGN(delalloc_bytes, blocksize)) >> 9;
9680         return 0;
9681 }
9682
9683 static int btrfs_rename_exchange(struct inode *old_dir,
9684                               struct dentry *old_dentry,
9685                               struct inode *new_dir,
9686                               struct dentry *new_dentry)
9687 {
9688         struct btrfs_fs_info *fs_info = btrfs_sb(old_dir->i_sb);
9689         struct btrfs_trans_handle *trans;
9690         struct btrfs_root *root = BTRFS_I(old_dir)->root;
9691         struct btrfs_root *dest = BTRFS_I(new_dir)->root;
9692         struct inode *new_inode = new_dentry->d_inode;
9693         struct inode *old_inode = old_dentry->d_inode;
9694         struct timespec ctime = current_time(old_inode);
9695         struct dentry *parent;
9696         u64 old_ino = btrfs_ino(BTRFS_I(old_inode));
9697         u64 new_ino = btrfs_ino(BTRFS_I(new_inode));
9698         u64 old_idx = 0;
9699         u64 new_idx = 0;
9700         u64 root_objectid;
9701         int ret;
9702         bool root_log_pinned = false;
9703         bool dest_log_pinned = false;
9704
9705         /* we only allow rename subvolume link between subvolumes */
9706         if (old_ino != BTRFS_FIRST_FREE_OBJECTID && root != dest)
9707                 return -EXDEV;
9708
9709         /* close the race window with snapshot create/destroy ioctl */
9710         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
9711                 down_read(&fs_info->subvol_sem);
9712         if (new_ino == BTRFS_FIRST_FREE_OBJECTID)
9713                 down_read(&fs_info->subvol_sem);
9714
9715         /*
9716          * We want to reserve the absolute worst case amount of items.  So if
9717          * both inodes are subvols and we need to unlink them then that would
9718          * require 4 item modifications, but if they are both normal inodes it
9719          * would require 5 item modifications, so we'll assume their normal
9720          * inodes.  So 5 * 2 is 10, plus 2 for the new links, so 12 total items
9721          * should cover the worst case number of items we'll modify.
9722          */
9723         trans = btrfs_start_transaction(root, 12);
9724         if (IS_ERR(trans)) {
9725                 ret = PTR_ERR(trans);
9726                 goto out_notrans;
9727         }
9728
9729         /*
9730          * We need to find a free sequence number both in the source and
9731          * in the destination directory for the exchange.
9732          */
9733         ret = btrfs_set_inode_index(BTRFS_I(new_dir), &old_idx);
9734         if (ret)
9735                 goto out_fail;
9736         ret = btrfs_set_inode_index(BTRFS_I(old_dir), &new_idx);
9737         if (ret)
9738                 goto out_fail;
9739
9740         BTRFS_I(old_inode)->dir_index = 0ULL;
9741         BTRFS_I(new_inode)->dir_index = 0ULL;
9742
9743         /* Reference for the source. */
9744         if (old_ino == BTRFS_FIRST_FREE_OBJECTID) {
9745                 /* force full log commit if subvolume involved. */
9746                 btrfs_set_log_full_commit(fs_info, trans);
9747         } else {
9748                 btrfs_pin_log_trans(root);
9749                 root_log_pinned = true;
9750                 ret = btrfs_insert_inode_ref(trans, dest,
9751                                              new_dentry->d_name.name,
9752                                              new_dentry->d_name.len,
9753                                              old_ino,
9754                                              btrfs_ino(BTRFS_I(new_dir)),
9755                                              old_idx);
9756                 if (ret)
9757                         goto out_fail;
9758         }
9759
9760         /* And now for the dest. */
9761         if (new_ino == BTRFS_FIRST_FREE_OBJECTID) {
9762                 /* force full log commit if subvolume involved. */
9763                 btrfs_set_log_full_commit(fs_info, trans);
9764         } else {
9765                 btrfs_pin_log_trans(dest);
9766                 dest_log_pinned = true;
9767                 ret = btrfs_insert_inode_ref(trans, root,
9768                                              old_dentry->d_name.name,
9769                                              old_dentry->d_name.len,
9770                                              new_ino,
9771                                              btrfs_ino(BTRFS_I(old_dir)),
9772                                              new_idx);
9773                 if (ret)
9774                         goto out_fail;
9775         }
9776
9777         /* Update inode version and ctime/mtime. */
9778         inode_inc_iversion(old_dir);
9779         inode_inc_iversion(new_dir);
9780         inode_inc_iversion(old_inode);
9781         inode_inc_iversion(new_inode);
9782         old_dir->i_ctime = old_dir->i_mtime = ctime;
9783         new_dir->i_ctime = new_dir->i_mtime = ctime;
9784         old_inode->i_ctime = ctime;
9785         new_inode->i_ctime = ctime;
9786
9787         if (old_dentry->d_parent != new_dentry->d_parent) {
9788                 btrfs_record_unlink_dir(trans, BTRFS_I(old_dir),
9789                                 BTRFS_I(old_inode), 1);
9790                 btrfs_record_unlink_dir(trans, BTRFS_I(new_dir),
9791                                 BTRFS_I(new_inode), 1);
9792         }
9793
9794         /* src is a subvolume */
9795         if (old_ino == BTRFS_FIRST_FREE_OBJECTID) {
9796                 root_objectid = BTRFS_I(old_inode)->root->root_key.objectid;
9797                 ret = btrfs_unlink_subvol(trans, root, old_dir,
9798                                           root_objectid,
9799                                           old_dentry->d_name.name,
9800                                           old_dentry->d_name.len);
9801         } else { /* src is an inode */
9802                 ret = __btrfs_unlink_inode(trans, root, BTRFS_I(old_dir),
9803                                            BTRFS_I(old_dentry->d_inode),
9804                                            old_dentry->d_name.name,
9805                                            old_dentry->d_name.len);
9806                 if (!ret)
9807                         ret = btrfs_update_inode(trans, root, old_inode);
9808         }
9809         if (ret) {
9810                 btrfs_abort_transaction(trans, ret);
9811                 goto out_fail;
9812         }
9813
9814         /* dest is a subvolume */
9815         if (new_ino == BTRFS_FIRST_FREE_OBJECTID) {
9816                 root_objectid = BTRFS_I(new_inode)->root->root_key.objectid;
9817                 ret = btrfs_unlink_subvol(trans, dest, new_dir,
9818                                           root_objectid,
9819                                           new_dentry->d_name.name,
9820                                           new_dentry->d_name.len);
9821         } else { /* dest is an inode */
9822                 ret = __btrfs_unlink_inode(trans, dest, BTRFS_I(new_dir),
9823                                            BTRFS_I(new_dentry->d_inode),
9824                                            new_dentry->d_name.name,
9825                                            new_dentry->d_name.len);
9826                 if (!ret)
9827                         ret = btrfs_update_inode(trans, dest, new_inode);
9828         }
9829         if (ret) {
9830                 btrfs_abort_transaction(trans, ret);
9831                 goto out_fail;
9832         }
9833
9834         ret = btrfs_add_link(trans, BTRFS_I(new_dir), BTRFS_I(old_inode),
9835                              new_dentry->d_name.name,
9836                              new_dentry->d_name.len, 0, old_idx);
9837         if (ret) {
9838                 btrfs_abort_transaction(trans, ret);
9839                 goto out_fail;
9840         }
9841
9842         ret = btrfs_add_link(trans, BTRFS_I(old_dir), BTRFS_I(new_inode),
9843                              old_dentry->d_name.name,
9844                              old_dentry->d_name.len, 0, new_idx);
9845         if (ret) {
9846                 btrfs_abort_transaction(trans, ret);
9847                 goto out_fail;
9848         }
9849
9850         if (old_inode->i_nlink == 1)
9851                 BTRFS_I(old_inode)->dir_index = old_idx;
9852         if (new_inode->i_nlink == 1)
9853                 BTRFS_I(new_inode)->dir_index = new_idx;
9854
9855         if (root_log_pinned) {
9856                 parent = new_dentry->d_parent;
9857                 btrfs_log_new_name(trans, BTRFS_I(old_inode), BTRFS_I(old_dir),
9858                                 parent);
9859                 btrfs_end_log_trans(root);
9860                 root_log_pinned = false;
9861         }
9862         if (dest_log_pinned) {
9863                 parent = old_dentry->d_parent;
9864                 btrfs_log_new_name(trans, BTRFS_I(new_inode), BTRFS_I(new_dir),
9865                                 parent);
9866                 btrfs_end_log_trans(dest);
9867                 dest_log_pinned = false;
9868         }
9869 out_fail:
9870         /*
9871          * If we have pinned a log and an error happened, we unpin tasks
9872          * trying to sync the log and force them to fallback to a transaction
9873          * commit if the log currently contains any of the inodes involved in
9874          * this rename operation (to ensure we do not persist a log with an
9875          * inconsistent state for any of these inodes or leading to any
9876          * inconsistencies when replayed). If the transaction was aborted, the
9877          * abortion reason is propagated to userspace when attempting to commit
9878          * the transaction. If the log does not contain any of these inodes, we
9879          * allow the tasks to sync it.
9880          */
9881         if (ret && (root_log_pinned || dest_log_pinned)) {
9882                 if (btrfs_inode_in_log(BTRFS_I(old_dir), fs_info->generation) ||
9883                     btrfs_inode_in_log(BTRFS_I(new_dir), fs_info->generation) ||
9884                     btrfs_inode_in_log(BTRFS_I(old_inode), fs_info->generation) ||
9885                     (new_inode &&
9886                      btrfs_inode_in_log(BTRFS_I(new_inode), fs_info->generation)))
9887                         btrfs_set_log_full_commit(fs_info, trans);
9888
9889                 if (root_log_pinned) {
9890                         btrfs_end_log_trans(root);
9891                         root_log_pinned = false;
9892                 }
9893                 if (dest_log_pinned) {
9894                         btrfs_end_log_trans(dest);
9895                         dest_log_pinned = false;
9896                 }
9897         }
9898         ret = btrfs_end_transaction(trans);
9899 out_notrans:
9900         if (new_ino == BTRFS_FIRST_FREE_OBJECTID)
9901                 up_read(&fs_info->subvol_sem);
9902         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
9903                 up_read(&fs_info->subvol_sem);
9904
9905         return ret;
9906 }
9907
9908 static int btrfs_whiteout_for_rename(struct btrfs_trans_handle *trans,
9909                                      struct btrfs_root *root,
9910                                      struct inode *dir,
9911                                      struct dentry *dentry)
9912 {
9913         int ret;
9914         struct inode *inode;
9915         u64 objectid;
9916         u64 index;
9917
9918         ret = btrfs_find_free_ino(root, &objectid);
9919         if (ret)
9920                 return ret;
9921
9922         inode = btrfs_new_inode(trans, root, dir,
9923                                 dentry->d_name.name,
9924                                 dentry->d_name.len,
9925                                 btrfs_ino(BTRFS_I(dir)),
9926                                 objectid,
9927                                 S_IFCHR | WHITEOUT_MODE,
9928                                 &index);
9929
9930         if (IS_ERR(inode)) {
9931                 ret = PTR_ERR(inode);
9932                 return ret;
9933         }
9934
9935         inode->i_op = &btrfs_special_inode_operations;
9936         init_special_inode(inode, inode->i_mode,
9937                 WHITEOUT_DEV);
9938
9939         ret = btrfs_init_inode_security(trans, inode, dir,
9940                                 &dentry->d_name);
9941         if (ret)
9942                 goto out;
9943
9944         ret = btrfs_add_nondir(trans, BTRFS_I(dir), dentry,
9945                                 BTRFS_I(inode), 0, index);
9946         if (ret)
9947                 goto out;
9948
9949         ret = btrfs_update_inode(trans, root, inode);
9950 out:
9951         unlock_new_inode(inode);
9952         if (ret)
9953                 inode_dec_link_count(inode);
9954         iput(inode);
9955
9956         return ret;
9957 }
9958
9959 static int btrfs_rename(struct inode *old_dir, struct dentry *old_dentry,
9960                            struct inode *new_dir, struct dentry *new_dentry,
9961                            unsigned int flags)
9962 {
9963         struct btrfs_fs_info *fs_info = btrfs_sb(old_dir->i_sb);
9964         struct btrfs_trans_handle *trans;
9965         unsigned int trans_num_items;
9966         struct btrfs_root *root = BTRFS_I(old_dir)->root;
9967         struct btrfs_root *dest = BTRFS_I(new_dir)->root;
9968         struct inode *new_inode = d_inode(new_dentry);
9969         struct inode *old_inode = d_inode(old_dentry);
9970         u64 index = 0;
9971         u64 root_objectid;
9972         int ret;
9973         u64 old_ino = btrfs_ino(BTRFS_I(old_inode));
9974         bool log_pinned = false;
9975
9976         if (btrfs_ino(BTRFS_I(new_dir)) == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)
9977                 return -EPERM;
9978
9979         /* we only allow rename subvolume link between subvolumes */
9980         if (old_ino != BTRFS_FIRST_FREE_OBJECTID && root != dest)
9981                 return -EXDEV;
9982
9983         if (old_ino == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID ||
9984             (new_inode && btrfs_ino(BTRFS_I(new_inode)) == BTRFS_FIRST_FREE_OBJECTID))
9985                 return -ENOTEMPTY;
9986
9987         if (S_ISDIR(old_inode->i_mode) && new_inode &&
9988             new_inode->i_size > BTRFS_EMPTY_DIR_SIZE)
9989                 return -ENOTEMPTY;
9990
9991
9992         /* check for collisions, even if the  name isn't there */
9993         ret = btrfs_check_dir_item_collision(dest, new_dir->i_ino,
9994                              new_dentry->d_name.name,
9995                              new_dentry->d_name.len);
9996
9997         if (ret) {
9998                 if (ret == -EEXIST) {
9999                         /* we shouldn't get
10000                          * eexist without a new_inode */
10001                         if (WARN_ON(!new_inode)) {
10002                                 return ret;
10003                         }
10004                 } else {
10005                         /* maybe -EOVERFLOW */
10006                         return ret;
10007                 }
10008         }
10009         ret = 0;
10010
10011         /*
10012          * we're using rename to replace one file with another.  Start IO on it
10013          * now so  we don't add too much work to the end of the transaction
10014          */
10015         if (new_inode && S_ISREG(old_inode->i_mode) && new_inode->i_size)
10016                 filemap_flush(old_inode->i_mapping);
10017
10018         /* close the racy window with snapshot create/destroy ioctl */
10019         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
10020                 down_read(&fs_info->subvol_sem);
10021         /*
10022          * We want to reserve the absolute worst case amount of items.  So if
10023          * both inodes are subvols and we need to unlink them then that would
10024          * require 4 item modifications, but if they are both normal inodes it
10025          * would require 5 item modifications, so we'll assume they are normal
10026          * inodes.  So 5 * 2 is 10, plus 1 for the new link, so 11 total items
10027          * should cover the worst case number of items we'll modify.
10028          * If our rename has the whiteout flag, we need more 5 units for the
10029          * new inode (1 inode item, 1 inode ref, 2 dir items and 1 xattr item
10030          * when selinux is enabled).
10031          */
10032         trans_num_items = 11;
10033         if (flags & RENAME_WHITEOUT)
10034                 trans_num_items += 5;
10035         trans = btrfs_start_transaction(root, trans_num_items);
10036         if (IS_ERR(trans)) {
10037                 ret = PTR_ERR(trans);
10038                 goto out_notrans;
10039         }
10040
10041         if (dest != root)
10042                 btrfs_record_root_in_trans(trans, dest);
10043
10044         ret = btrfs_set_inode_index(BTRFS_I(new_dir), &index);
10045         if (ret)
10046                 goto out_fail;
10047
10048         BTRFS_I(old_inode)->dir_index = 0ULL;
10049         if (unlikely(old_ino == BTRFS_FIRST_FREE_OBJECTID)) {
10050                 /* force full log commit if subvolume involved. */
10051                 btrfs_set_log_full_commit(fs_info, trans);
10052         } else {
10053                 btrfs_pin_log_trans(root);
10054                 log_pinned = true;
10055                 ret = btrfs_insert_inode_ref(trans, dest,
10056                                              new_dentry->d_name.name,
10057                                              new_dentry->d_name.len,
10058                                              old_ino,
10059                                              btrfs_ino(BTRFS_I(new_dir)), index);
10060                 if (ret)
10061                         goto out_fail;
10062         }
10063
10064         inode_inc_iversion(old_dir);
10065         inode_inc_iversion(new_dir);
10066         inode_inc_iversion(old_inode);
10067         old_dir->i_ctime = old_dir->i_mtime =
10068         new_dir->i_ctime = new_dir->i_mtime =
10069         old_inode->i_ctime = current_time(old_dir);
10070
10071         if (old_dentry->d_parent != new_dentry->d_parent)
10072                 btrfs_record_unlink_dir(trans, BTRFS_I(old_dir),
10073                                 BTRFS_I(old_inode), 1);
10074
10075         if (unlikely(old_ino == BTRFS_FIRST_FREE_OBJECTID)) {
10076                 root_objectid = BTRFS_I(old_inode)->root->root_key.objectid;
10077                 ret = btrfs_unlink_subvol(trans, root, old_dir, root_objectid,
10078                                         old_dentry->d_name.name,
10079                                         old_dentry->d_name.len);
10080         } else {
10081                 ret = __btrfs_unlink_inode(trans, root, BTRFS_I(old_dir),
10082                                         BTRFS_I(d_inode(old_dentry)),
10083                                         old_dentry->d_name.name,
10084                                         old_dentry->d_name.len);
10085                 if (!ret)
10086                         ret = btrfs_update_inode(trans, root, old_inode);
10087         }
10088         if (ret) {
10089                 btrfs_abort_transaction(trans, ret);
10090                 goto out_fail;
10091         }
10092
10093         if (new_inode) {
10094                 inode_inc_iversion(new_inode);
10095                 new_inode->i_ctime = current_time(new_inode);
10096                 if (unlikely(btrfs_ino(BTRFS_I(new_inode)) ==
10097                              BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)) {
10098                         root_objectid = BTRFS_I(new_inode)->location.objectid;
10099                         ret = btrfs_unlink_subvol(trans, dest, new_dir,
10100                                                 root_objectid,
10101                                                 new_dentry->d_name.name,
10102                                                 new_dentry->d_name.len);
10103                         BUG_ON(new_inode->i_nlink == 0);
10104                 } else {
10105                         ret = btrfs_unlink_inode(trans, dest, BTRFS_I(new_dir),
10106                                                  BTRFS_I(d_inode(new_dentry)),
10107                                                  new_dentry->d_name.name,
10108                                                  new_dentry->d_name.len);
10109                 }
10110                 if (!ret && new_inode->i_nlink == 0)
10111                         ret = btrfs_orphan_add(trans,
10112                                         BTRFS_I(d_inode(new_dentry)));
10113                 if (ret) {
10114                         btrfs_abort_transaction(trans, ret);
10115                         goto out_fail;
10116                 }
10117         }
10118
10119         ret = btrfs_add_link(trans, BTRFS_I(new_dir), BTRFS_I(old_inode),
10120                              new_dentry->d_name.name,
10121                              new_dentry->d_name.len, 0, index);
10122         if (ret) {
10123                 btrfs_abort_transaction(trans, ret);
10124                 goto out_fail;
10125         }
10126
10127         if (old_inode->i_nlink == 1)
10128                 BTRFS_I(old_inode)->dir_index = index;
10129
10130         if (log_pinned) {
10131                 struct dentry *parent = new_dentry->d_parent;
10132
10133                 btrfs_log_new_name(trans, BTRFS_I(old_inode), BTRFS_I(old_dir),
10134                                 parent);
10135                 btrfs_end_log_trans(root);
10136                 log_pinned = false;
10137         }
10138
10139         if (flags & RENAME_WHITEOUT) {
10140                 ret = btrfs_whiteout_for_rename(trans, root, old_dir,
10141                                                 old_dentry);
10142
10143                 if (ret) {
10144                         btrfs_abort_transaction(trans, ret);
10145                         goto out_fail;
10146                 }
10147         }
10148 out_fail:
10149         /*
10150          * If we have pinned the log and an error happened, we unpin tasks
10151          * trying to sync the log and force them to fallback to a transaction
10152          * commit if the log currently contains any of the inodes involved in
10153          * this rename operation (to ensure we do not persist a log with an
10154          * inconsistent state for any of these inodes or leading to any
10155          * inconsistencies when replayed). If the transaction was aborted, the
10156          * abortion reason is propagated to userspace when attempting to commit
10157          * the transaction. If the log does not contain any of these inodes, we
10158          * allow the tasks to sync it.
10159          */
10160         if (ret && log_pinned) {
10161                 if (btrfs_inode_in_log(BTRFS_I(old_dir), fs_info->generation) ||
10162                     btrfs_inode_in_log(BTRFS_I(new_dir), fs_info->generation) ||
10163                     btrfs_inode_in_log(BTRFS_I(old_inode), fs_info->generation) ||
10164                     (new_inode &&
10165                      btrfs_inode_in_log(BTRFS_I(new_inode), fs_info->generation)))
10166                         btrfs_set_log_full_commit(fs_info, trans);
10167
10168                 btrfs_end_log_trans(root);
10169                 log_pinned = false;
10170         }
10171         btrfs_end_transaction(trans);
10172 out_notrans:
10173         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
10174                 up_read(&fs_info->subvol_sem);
10175
10176         return ret;
10177 }
10178
10179 static int btrfs_rename2(struct inode *old_dir, struct dentry *old_dentry,
10180                          struct inode *new_dir, struct dentry *new_dentry,
10181                          unsigned int flags)
10182 {
10183         if (flags & ~(RENAME_NOREPLACE | RENAME_EXCHANGE | RENAME_WHITEOUT))
10184                 return -EINVAL;
10185
10186         if (flags & RENAME_EXCHANGE)
10187                 return btrfs_rename_exchange(old_dir, old_dentry, new_dir,
10188                                           new_dentry);
10189
10190         return btrfs_rename(old_dir, old_dentry, new_dir, new_dentry, flags);
10191 }
10192
10193 static void btrfs_run_delalloc_work(struct btrfs_work *work)
10194 {
10195         struct btrfs_delalloc_work *delalloc_work;
10196         struct inode *inode;
10197
10198         delalloc_work = container_of(work, struct btrfs_delalloc_work,
10199                                      work);
10200         inode = delalloc_work->inode;
10201         filemap_flush(inode->i_mapping);
10202         if (test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
10203                                 &BTRFS_I(inode)->runtime_flags))
10204                 filemap_flush(inode->i_mapping);
10205
10206         if (delalloc_work->delay_iput)
10207                 btrfs_add_delayed_iput(inode);
10208         else
10209                 iput(inode);
10210         complete(&delalloc_work->completion);
10211 }
10212
10213 struct btrfs_delalloc_work *btrfs_alloc_delalloc_work(struct inode *inode,
10214                                                     int delay_iput)
10215 {
10216         struct btrfs_delalloc_work *work;
10217
10218         work = kmalloc(sizeof(*work), GFP_NOFS);
10219         if (!work)
10220                 return NULL;
10221
10222         init_completion(&work->completion);
10223         INIT_LIST_HEAD(&work->list);
10224         work->inode = inode;
10225         work->delay_iput = delay_iput;
10226         WARN_ON_ONCE(!inode);
10227         btrfs_init_work(&work->work, btrfs_flush_delalloc_helper,
10228                         btrfs_run_delalloc_work, NULL, NULL);
10229
10230         return work;
10231 }
10232
10233 void btrfs_wait_and_free_delalloc_work(struct btrfs_delalloc_work *work)
10234 {
10235         wait_for_completion(&work->completion);
10236         kfree(work);
10237 }
10238
10239 /*
10240  * some fairly slow code that needs optimization. This walks the list
10241  * of all the inodes with pending delalloc and forces them to disk.
10242  */
10243 static int __start_delalloc_inodes(struct btrfs_root *root, int delay_iput,
10244                                    int nr)
10245 {
10246         struct btrfs_inode *binode;
10247         struct inode *inode;
10248         struct btrfs_delalloc_work *work, *next;
10249         struct list_head works;
10250         struct list_head splice;
10251         int ret = 0;
10252
10253         INIT_LIST_HEAD(&works);
10254         INIT_LIST_HEAD(&splice);
10255
10256         mutex_lock(&root->delalloc_mutex);
10257         spin_lock(&root->delalloc_lock);
10258         list_splice_init(&root->delalloc_inodes, &splice);
10259         while (!list_empty(&splice)) {
10260                 binode = list_entry(splice.next, struct btrfs_inode,
10261                                     delalloc_inodes);
10262
10263                 list_move_tail(&binode->delalloc_inodes,
10264                                &root->delalloc_inodes);
10265                 inode = igrab(&binode->vfs_inode);
10266                 if (!inode) {
10267                         cond_resched_lock(&root->delalloc_lock);
10268                         continue;
10269                 }
10270                 spin_unlock(&root->delalloc_lock);
10271
10272                 work = btrfs_alloc_delalloc_work(inode, delay_iput);
10273                 if (!work) {
10274                         if (delay_iput)
10275                                 btrfs_add_delayed_iput(inode);
10276                         else
10277                                 iput(inode);
10278                         ret = -ENOMEM;
10279                         goto out;
10280                 }
10281                 list_add_tail(&work->list, &works);
10282                 btrfs_queue_work(root->fs_info->flush_workers,
10283                                  &work->work);
10284                 ret++;
10285                 if (nr != -1 && ret >= nr)
10286                         goto out;
10287                 cond_resched();
10288                 spin_lock(&root->delalloc_lock);
10289         }
10290         spin_unlock(&root->delalloc_lock);
10291
10292 out:
10293         list_for_each_entry_safe(work, next, &works, list) {
10294                 list_del_init(&work->list);
10295                 btrfs_wait_and_free_delalloc_work(work);
10296         }
10297
10298         if (!list_empty_careful(&splice)) {
10299                 spin_lock(&root->delalloc_lock);
10300                 list_splice_tail(&splice, &root->delalloc_inodes);
10301                 spin_unlock(&root->delalloc_lock);
10302         }
10303         mutex_unlock(&root->delalloc_mutex);
10304         return ret;
10305 }
10306
10307 int btrfs_start_delalloc_inodes(struct btrfs_root *root, int delay_iput)
10308 {
10309         struct btrfs_fs_info *fs_info = root->fs_info;
10310         int ret;
10311
10312         if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state))
10313                 return -EROFS;
10314
10315         ret = __start_delalloc_inodes(root, delay_iput, -1);
10316         if (ret > 0)
10317                 ret = 0;
10318         /*
10319          * the filemap_flush will queue IO into the worker threads, but
10320          * we have to make sure the IO is actually started and that
10321          * ordered extents get created before we return
10322          */
10323         atomic_inc(&fs_info->async_submit_draining);
10324         while (atomic_read(&fs_info->nr_async_submits) ||
10325                atomic_read(&fs_info->async_delalloc_pages)) {
10326                 wait_event(fs_info->async_submit_wait,
10327                            (atomic_read(&fs_info->nr_async_submits) == 0 &&
10328                             atomic_read(&fs_info->async_delalloc_pages) == 0));
10329         }
10330         atomic_dec(&fs_info->async_submit_draining);
10331         return ret;
10332 }
10333
10334 int btrfs_start_delalloc_roots(struct btrfs_fs_info *fs_info, int delay_iput,
10335                                int nr)
10336 {
10337         struct btrfs_root *root;
10338         struct list_head splice;
10339         int ret;
10340
10341         if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state))
10342                 return -EROFS;
10343
10344         INIT_LIST_HEAD(&splice);
10345
10346         mutex_lock(&fs_info->delalloc_root_mutex);
10347         spin_lock(&fs_info->delalloc_root_lock);
10348         list_splice_init(&fs_info->delalloc_roots, &splice);
10349         while (!list_empty(&splice) && nr) {
10350                 root = list_first_entry(&splice, struct btrfs_root,
10351                                         delalloc_root);
10352                 root = btrfs_grab_fs_root(root);
10353                 BUG_ON(!root);
10354                 list_move_tail(&root->delalloc_root,
10355                                &fs_info->delalloc_roots);
10356                 spin_unlock(&fs_info->delalloc_root_lock);
10357
10358                 ret = __start_delalloc_inodes(root, delay_iput, nr);
10359                 btrfs_put_fs_root(root);
10360                 if (ret < 0)
10361                         goto out;
10362
10363                 if (nr != -1) {
10364                         nr -= ret;
10365                         WARN_ON(nr < 0);
10366                 }
10367                 spin_lock(&fs_info->delalloc_root_lock);
10368         }
10369         spin_unlock(&fs_info->delalloc_root_lock);
10370
10371         ret = 0;
10372         atomic_inc(&fs_info->async_submit_draining);
10373         while (atomic_read(&fs_info->nr_async_submits) ||
10374               atomic_read(&fs_info->async_delalloc_pages)) {
10375                 wait_event(fs_info->async_submit_wait,
10376                    (atomic_read(&fs_info->nr_async_submits) == 0 &&
10377                     atomic_read(&fs_info->async_delalloc_pages) == 0));
10378         }
10379         atomic_dec(&fs_info->async_submit_draining);
10380 out:
10381         if (!list_empty_careful(&splice)) {
10382                 spin_lock(&fs_info->delalloc_root_lock);
10383                 list_splice_tail(&splice, &fs_info->delalloc_roots);
10384                 spin_unlock(&fs_info->delalloc_root_lock);
10385         }
10386         mutex_unlock(&fs_info->delalloc_root_mutex);
10387         return ret;
10388 }
10389
10390 static int btrfs_symlink(struct inode *dir, struct dentry *dentry,
10391                          const char *symname)
10392 {
10393         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
10394         struct btrfs_trans_handle *trans;
10395         struct btrfs_root *root = BTRFS_I(dir)->root;
10396         struct btrfs_path *path;
10397         struct btrfs_key key;
10398         struct inode *inode = NULL;
10399         int err;
10400         int drop_inode = 0;
10401         u64 objectid;
10402         u64 index = 0;
10403         int name_len;
10404         int datasize;
10405         unsigned long ptr;
10406         struct btrfs_file_extent_item *ei;
10407         struct extent_buffer *leaf;
10408
10409         name_len = strlen(symname);
10410         if (name_len > BTRFS_MAX_INLINE_DATA_SIZE(fs_info))
10411                 return -ENAMETOOLONG;
10412
10413         /*
10414          * 2 items for inode item and ref
10415          * 2 items for dir items
10416          * 1 item for updating parent inode item
10417          * 1 item for the inline extent item
10418          * 1 item for xattr if selinux is on
10419          */
10420         trans = btrfs_start_transaction(root, 7);
10421         if (IS_ERR(trans))
10422                 return PTR_ERR(trans);
10423
10424         err = btrfs_find_free_ino(root, &objectid);
10425         if (err)
10426                 goto out_unlock;
10427
10428         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
10429                                 dentry->d_name.len, btrfs_ino(BTRFS_I(dir)),
10430                                 objectid, S_IFLNK|S_IRWXUGO, &index);
10431         if (IS_ERR(inode)) {
10432                 err = PTR_ERR(inode);
10433                 goto out_unlock;
10434         }
10435
10436         /*
10437         * If the active LSM wants to access the inode during
10438         * d_instantiate it needs these. Smack checks to see
10439         * if the filesystem supports xattrs by looking at the
10440         * ops vector.
10441         */
10442         inode->i_fop = &btrfs_file_operations;
10443         inode->i_op = &btrfs_file_inode_operations;
10444         inode->i_mapping->a_ops = &btrfs_aops;
10445         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
10446
10447         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
10448         if (err)
10449                 goto out_unlock_inode;
10450
10451         path = btrfs_alloc_path();
10452         if (!path) {
10453                 err = -ENOMEM;
10454                 goto out_unlock_inode;
10455         }
10456         key.objectid = btrfs_ino(BTRFS_I(inode));
10457         key.offset = 0;
10458         key.type = BTRFS_EXTENT_DATA_KEY;
10459         datasize = btrfs_file_extent_calc_inline_size(name_len);
10460         err = btrfs_insert_empty_item(trans, root, path, &key,
10461                                       datasize);
10462         if (err) {
10463                 btrfs_free_path(path);
10464                 goto out_unlock_inode;
10465         }
10466         leaf = path->nodes[0];
10467         ei = btrfs_item_ptr(leaf, path->slots[0],
10468                             struct btrfs_file_extent_item);
10469         btrfs_set_file_extent_generation(leaf, ei, trans->transid);
10470         btrfs_set_file_extent_type(leaf, ei,
10471                                    BTRFS_FILE_EXTENT_INLINE);
10472         btrfs_set_file_extent_encryption(leaf, ei, 0);
10473         btrfs_set_file_extent_compression(leaf, ei, 0);
10474         btrfs_set_file_extent_other_encoding(leaf, ei, 0);
10475         btrfs_set_file_extent_ram_bytes(leaf, ei, name_len);
10476
10477         ptr = btrfs_file_extent_inline_start(ei);
10478         write_extent_buffer(leaf, symname, ptr, name_len);
10479         btrfs_mark_buffer_dirty(leaf);
10480         btrfs_free_path(path);
10481
10482         inode->i_op = &btrfs_symlink_inode_operations;
10483         inode_nohighmem(inode);
10484         inode->i_mapping->a_ops = &btrfs_symlink_aops;
10485         inode_set_bytes(inode, name_len);
10486         btrfs_i_size_write(BTRFS_I(inode), name_len);
10487         err = btrfs_update_inode(trans, root, inode);
10488         /*
10489          * Last step, add directory indexes for our symlink inode. This is the
10490          * last step to avoid extra cleanup of these indexes if an error happens
10491          * elsewhere above.
10492          */
10493         if (!err)
10494                 err = btrfs_add_nondir(trans, BTRFS_I(dir), dentry,
10495                                 BTRFS_I(inode), 0, index);
10496         if (err) {
10497                 drop_inode = 1;
10498                 goto out_unlock_inode;
10499         }
10500
10501         unlock_new_inode(inode);
10502         d_instantiate(dentry, inode);
10503
10504 out_unlock:
10505         btrfs_end_transaction(trans);
10506         if (drop_inode) {
10507                 inode_dec_link_count(inode);
10508                 iput(inode);
10509         }
10510         btrfs_btree_balance_dirty(fs_info);
10511         return err;
10512
10513 out_unlock_inode:
10514         drop_inode = 1;
10515         unlock_new_inode(inode);
10516         goto out_unlock;
10517 }
10518
10519 static int __btrfs_prealloc_file_range(struct inode *inode, int mode,
10520                                        u64 start, u64 num_bytes, u64 min_size,
10521                                        loff_t actual_len, u64 *alloc_hint,
10522                                        struct btrfs_trans_handle *trans)
10523 {
10524         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
10525         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
10526         struct extent_map *em;
10527         struct btrfs_root *root = BTRFS_I(inode)->root;
10528         struct btrfs_key ins;
10529         u64 cur_offset = start;
10530         u64 i_size;
10531         u64 cur_bytes;
10532         u64 last_alloc = (u64)-1;
10533         int ret = 0;
10534         bool own_trans = true;
10535         u64 end = start + num_bytes - 1;
10536
10537         if (trans)
10538                 own_trans = false;
10539         while (num_bytes > 0) {
10540                 if (own_trans) {
10541                         trans = btrfs_start_transaction(root, 3);
10542                         if (IS_ERR(trans)) {
10543                                 ret = PTR_ERR(trans);
10544                                 break;
10545                         }
10546                 }
10547
10548                 cur_bytes = min_t(u64, num_bytes, SZ_256M);
10549                 cur_bytes = max(cur_bytes, min_size);
10550                 /*
10551                  * If we are severely fragmented we could end up with really
10552                  * small allocations, so if the allocator is returning small
10553                  * chunks lets make its job easier by only searching for those
10554                  * sized chunks.
10555                  */
10556                 cur_bytes = min(cur_bytes, last_alloc);
10557                 ret = btrfs_reserve_extent(root, cur_bytes, cur_bytes,
10558                                 min_size, 0, *alloc_hint, &ins, 1, 0);
10559                 if (ret) {
10560                         if (own_trans)
10561                                 btrfs_end_transaction(trans);
10562                         break;
10563                 }
10564                 btrfs_dec_block_group_reservations(fs_info, ins.objectid);
10565
10566                 last_alloc = ins.offset;
10567                 ret = insert_reserved_file_extent(trans, inode,
10568                                                   cur_offset, ins.objectid,
10569                                                   ins.offset, ins.offset,
10570                                                   ins.offset, 0, 0, 0,
10571                                                   BTRFS_FILE_EXTENT_PREALLOC);
10572                 if (ret) {
10573                         btrfs_free_reserved_extent(fs_info, ins.objectid,
10574                                                    ins.offset, 0);
10575                         btrfs_abort_transaction(trans, ret);
10576                         if (own_trans)
10577                                 btrfs_end_transaction(trans);
10578                         break;
10579                 }
10580
10581                 btrfs_drop_extent_cache(BTRFS_I(inode), cur_offset,
10582                                         cur_offset + ins.offset -1, 0);
10583
10584                 em = alloc_extent_map();
10585                 if (!em) {
10586                         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
10587                                 &BTRFS_I(inode)->runtime_flags);
10588                         goto next;
10589                 }
10590
10591                 em->start = cur_offset;
10592                 em->orig_start = cur_offset;
10593                 em->len = ins.offset;
10594                 em->block_start = ins.objectid;
10595                 em->block_len = ins.offset;
10596                 em->orig_block_len = ins.offset;
10597                 em->ram_bytes = ins.offset;
10598                 em->bdev = fs_info->fs_devices->latest_bdev;
10599                 set_bit(EXTENT_FLAG_PREALLOC, &em->flags);
10600                 em->generation = trans->transid;
10601
10602                 while (1) {
10603                         write_lock(&em_tree->lock);
10604                         ret = add_extent_mapping(em_tree, em, 1);
10605                         write_unlock(&em_tree->lock);
10606                         if (ret != -EEXIST)
10607                                 break;
10608                         btrfs_drop_extent_cache(BTRFS_I(inode), cur_offset,
10609                                                 cur_offset + ins.offset - 1,
10610                                                 0);
10611                 }
10612                 free_extent_map(em);
10613 next:
10614                 num_bytes -= ins.offset;
10615                 cur_offset += ins.offset;
10616                 *alloc_hint = ins.objectid + ins.offset;
10617
10618                 inode_inc_iversion(inode);
10619                 inode->i_ctime = current_time(inode);
10620                 BTRFS_I(inode)->flags |= BTRFS_INODE_PREALLOC;
10621                 if (!(mode & FALLOC_FL_KEEP_SIZE) &&
10622                     (actual_len > inode->i_size) &&
10623                     (cur_offset > inode->i_size)) {
10624                         if (cur_offset > actual_len)
10625                                 i_size = actual_len;
10626                         else
10627                                 i_size = cur_offset;
10628                         i_size_write(inode, i_size);
10629                         btrfs_ordered_update_i_size(inode, i_size, NULL);
10630                 }
10631
10632                 ret = btrfs_update_inode(trans, root, inode);
10633
10634                 if (ret) {
10635                         btrfs_abort_transaction(trans, ret);
10636                         if (own_trans)
10637                                 btrfs_end_transaction(trans);
10638                         break;
10639                 }
10640
10641                 if (own_trans)
10642                         btrfs_end_transaction(trans);
10643         }
10644         if (cur_offset < end)
10645                 btrfs_free_reserved_data_space(inode, NULL, cur_offset,
10646                         end - cur_offset + 1);
10647         return ret;
10648 }
10649
10650 int btrfs_prealloc_file_range(struct inode *inode, int mode,
10651                               u64 start, u64 num_bytes, u64 min_size,
10652                               loff_t actual_len, u64 *alloc_hint)
10653 {
10654         return __btrfs_prealloc_file_range(inode, mode, start, num_bytes,
10655                                            min_size, actual_len, alloc_hint,
10656                                            NULL);
10657 }
10658
10659 int btrfs_prealloc_file_range_trans(struct inode *inode,
10660                                     struct btrfs_trans_handle *trans, int mode,
10661                                     u64 start, u64 num_bytes, u64 min_size,
10662                                     loff_t actual_len, u64 *alloc_hint)
10663 {
10664         return __btrfs_prealloc_file_range(inode, mode, start, num_bytes,
10665                                            min_size, actual_len, alloc_hint, trans);
10666 }
10667
10668 static int btrfs_set_page_dirty(struct page *page)
10669 {
10670         return __set_page_dirty_nobuffers(page);
10671 }
10672
10673 static int btrfs_permission(struct inode *inode, int mask)
10674 {
10675         struct btrfs_root *root = BTRFS_I(inode)->root;
10676         umode_t mode = inode->i_mode;
10677
10678         if (mask & MAY_WRITE &&
10679             (S_ISREG(mode) || S_ISDIR(mode) || S_ISLNK(mode))) {
10680                 if (btrfs_root_readonly(root))
10681                         return -EROFS;
10682                 if (BTRFS_I(inode)->flags & BTRFS_INODE_READONLY)
10683                         return -EACCES;
10684         }
10685         return generic_permission(inode, mask);
10686 }
10687
10688 static int btrfs_tmpfile(struct inode *dir, struct dentry *dentry, umode_t mode)
10689 {
10690         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
10691         struct btrfs_trans_handle *trans;
10692         struct btrfs_root *root = BTRFS_I(dir)->root;
10693         struct inode *inode = NULL;
10694         u64 objectid;
10695         u64 index;
10696         int ret = 0;
10697
10698         /*
10699          * 5 units required for adding orphan entry
10700          */
10701         trans = btrfs_start_transaction(root, 5);
10702         if (IS_ERR(trans))
10703                 return PTR_ERR(trans);
10704
10705         ret = btrfs_find_free_ino(root, &objectid);
10706         if (ret)
10707                 goto out;
10708
10709         inode = btrfs_new_inode(trans, root, dir, NULL, 0,
10710                         btrfs_ino(BTRFS_I(dir)), objectid, mode, &index);
10711         if (IS_ERR(inode)) {
10712                 ret = PTR_ERR(inode);
10713                 inode = NULL;
10714                 goto out;
10715         }
10716
10717         inode->i_fop = &btrfs_file_operations;
10718         inode->i_op = &btrfs_file_inode_operations;
10719
10720         inode->i_mapping->a_ops = &btrfs_aops;
10721         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
10722
10723         ret = btrfs_init_inode_security(trans, inode, dir, NULL);
10724         if (ret)
10725                 goto out_inode;
10726
10727         ret = btrfs_update_inode(trans, root, inode);
10728         if (ret)
10729                 goto out_inode;
10730         ret = btrfs_orphan_add(trans, BTRFS_I(inode));
10731         if (ret)
10732                 goto out_inode;
10733
10734         /*
10735          * We set number of links to 0 in btrfs_new_inode(), and here we set
10736          * it to 1 because d_tmpfile() will issue a warning if the count is 0,
10737          * through:
10738          *
10739          *    d_tmpfile() -> inode_dec_link_count() -> drop_nlink()
10740          */
10741         set_nlink(inode, 1);
10742         unlock_new_inode(inode);
10743         d_tmpfile(dentry, inode);
10744         mark_inode_dirty(inode);
10745
10746 out:
10747         btrfs_end_transaction(trans);
10748         if (ret)
10749                 iput(inode);
10750         btrfs_balance_delayed_items(fs_info);
10751         btrfs_btree_balance_dirty(fs_info);
10752         return ret;
10753
10754 out_inode:
10755         unlock_new_inode(inode);
10756         goto out;
10757
10758 }
10759
10760 __attribute__((const))
10761 static int btrfs_readpage_io_failed_hook(struct page *page, int failed_mirror)
10762 {
10763         return -EAGAIN;
10764 }
10765
10766 static struct btrfs_fs_info *iotree_fs_info(void *private_data)
10767 {
10768         struct inode *inode = private_data;
10769         return btrfs_sb(inode->i_sb);
10770 }
10771
10772 static void btrfs_check_extent_io_range(void *private_data, const char *caller,
10773                                         u64 start, u64 end)
10774 {
10775         struct inode *inode = private_data;
10776         u64 isize;
10777
10778         isize = i_size_read(inode);
10779         if (end >= PAGE_SIZE && (end % 2) == 0 && end != isize - 1) {
10780                 btrfs_debug_rl(BTRFS_I(inode)->root->fs_info,
10781                     "%s: ino %llu isize %llu odd range [%llu,%llu]",
10782                         caller, btrfs_ino(BTRFS_I(inode)), isize, start, end);
10783         }
10784 }
10785
10786 void btrfs_set_range_writeback(void *private_data, u64 start, u64 end)
10787 {
10788         struct inode *inode = private_data;
10789         unsigned long index = start >> PAGE_SHIFT;
10790         unsigned long end_index = end >> PAGE_SHIFT;
10791         struct page *page;
10792
10793         while (index <= end_index) {
10794                 page = find_get_page(inode->i_mapping, index);
10795                 ASSERT(page); /* Pages should be in the extent_io_tree */
10796                 set_page_writeback(page);
10797                 put_page(page);
10798                 index++;
10799         }
10800 }
10801
10802 static const struct inode_operations btrfs_dir_inode_operations = {
10803         .getattr        = btrfs_getattr,
10804         .lookup         = btrfs_lookup,
10805         .create         = btrfs_create,
10806         .unlink         = btrfs_unlink,
10807         .link           = btrfs_link,
10808         .mkdir          = btrfs_mkdir,
10809         .rmdir          = btrfs_rmdir,
10810         .rename         = btrfs_rename2,
10811         .symlink        = btrfs_symlink,
10812         .setattr        = btrfs_setattr,
10813         .mknod          = btrfs_mknod,
10814         .listxattr      = btrfs_listxattr,
10815         .permission     = btrfs_permission,
10816         .get_acl        = btrfs_get_acl,
10817         .set_acl        = btrfs_set_acl,
10818         .update_time    = btrfs_update_time,
10819         .tmpfile        = btrfs_tmpfile,
10820 };
10821 static const struct inode_operations btrfs_dir_ro_inode_operations = {
10822         .lookup         = btrfs_lookup,
10823         .permission     = btrfs_permission,
10824         .update_time    = btrfs_update_time,
10825 };
10826
10827 static const struct file_operations btrfs_dir_file_operations = {
10828         .llseek         = generic_file_llseek,
10829         .read           = generic_read_dir,
10830         .iterate_shared = btrfs_real_readdir,
10831         .open           = btrfs_opendir,
10832         .unlocked_ioctl = btrfs_ioctl,
10833 #ifdef CONFIG_COMPAT
10834         .compat_ioctl   = btrfs_compat_ioctl,
10835 #endif
10836         .release        = btrfs_release_file,
10837         .fsync          = btrfs_sync_file,
10838 };
10839
10840 static const struct extent_io_ops btrfs_extent_io_ops = {
10841         /* mandatory callbacks */
10842         .submit_bio_hook = btrfs_submit_bio_hook,
10843         .readpage_end_io_hook = btrfs_readpage_end_io_hook,
10844         .merge_bio_hook = btrfs_merge_bio_hook,
10845         .readpage_io_failed_hook = btrfs_readpage_io_failed_hook,
10846         .tree_fs_info = iotree_fs_info,
10847         .set_range_writeback = btrfs_set_range_writeback,
10848
10849         /* optional callbacks */
10850         .fill_delalloc = run_delalloc_range,
10851         .writepage_end_io_hook = btrfs_writepage_end_io_hook,
10852         .writepage_start_hook = btrfs_writepage_start_hook,
10853         .set_bit_hook = btrfs_set_bit_hook,
10854         .clear_bit_hook = btrfs_clear_bit_hook,
10855         .merge_extent_hook = btrfs_merge_extent_hook,
10856         .split_extent_hook = btrfs_split_extent_hook,
10857         .check_extent_io_range = btrfs_check_extent_io_range,
10858 };
10859
10860 /*
10861  * btrfs doesn't support the bmap operation because swapfiles
10862  * use bmap to make a mapping of extents in the file.  They assume
10863  * these extents won't change over the life of the file and they
10864  * use the bmap result to do IO directly to the drive.
10865  *
10866  * the btrfs bmap call would return logical addresses that aren't
10867  * suitable for IO and they also will change frequently as COW
10868  * operations happen.  So, swapfile + btrfs == corruption.
10869  *
10870  * For now we're avoiding this by dropping bmap.
10871  */
10872 static const struct address_space_operations btrfs_aops = {
10873         .readpage       = btrfs_readpage,
10874         .writepage      = btrfs_writepage,
10875         .writepages     = btrfs_writepages,
10876         .readpages      = btrfs_readpages,
10877         .direct_IO      = btrfs_direct_IO,
10878         .invalidatepage = btrfs_invalidatepage,
10879         .releasepage    = btrfs_releasepage,
10880         .set_page_dirty = btrfs_set_page_dirty,
10881         .error_remove_page = generic_error_remove_page,
10882 };
10883
10884 static const struct address_space_operations btrfs_symlink_aops = {
10885         .readpage       = btrfs_readpage,
10886         .writepage      = btrfs_writepage,
10887         .invalidatepage = btrfs_invalidatepage,
10888         .releasepage    = btrfs_releasepage,
10889 };
10890
10891 static const struct inode_operations btrfs_file_inode_operations = {
10892         .getattr        = btrfs_getattr,
10893         .setattr        = btrfs_setattr,
10894         .listxattr      = btrfs_listxattr,
10895         .permission     = btrfs_permission,
10896         .fiemap         = btrfs_fiemap,
10897         .get_acl        = btrfs_get_acl,
10898         .set_acl        = btrfs_set_acl,
10899         .update_time    = btrfs_update_time,
10900 };
10901 static const struct inode_operations btrfs_special_inode_operations = {
10902         .getattr        = btrfs_getattr,
10903         .setattr        = btrfs_setattr,
10904         .permission     = btrfs_permission,
10905         .listxattr      = btrfs_listxattr,
10906         .get_acl        = btrfs_get_acl,
10907         .set_acl        = btrfs_set_acl,
10908         .update_time    = btrfs_update_time,
10909 };
10910 static const struct inode_operations btrfs_symlink_inode_operations = {
10911         .get_link       = page_get_link,
10912         .getattr        = btrfs_getattr,
10913         .setattr        = btrfs_setattr,
10914         .permission     = btrfs_permission,
10915         .listxattr      = btrfs_listxattr,
10916         .update_time    = btrfs_update_time,
10917 };
10918
10919 const struct dentry_operations btrfs_dentry_operations = {
10920         .d_delete       = btrfs_dentry_delete,
10921         .d_release      = btrfs_dentry_release,
10922 };