fs/btrfs/inode.c

   1 /*
   2  * Copyright (C) 2007 Oracle.  All rights reserved.
   3  *
   4  * This program is free software; you can redistribute it and/or
   5  * modify it under the terms of the GNU General Public
   6  * License v2 as published by the Free Software Foundation.
   7  *
   8  * This program is distributed in the hope that it will be useful,
   9  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  11  * General Public License for more details.
  12  *
  13  * You should have received a copy of the GNU General Public
  14  * License along with this program; if not, write to the
  15  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  16  * Boston, MA 021110-1307, USA.
  17  */
  18
  19 #include <linux/kernel.h>
  20 #include <linux/bio.h>
  21 #include <linux/buffer_head.h>
  22 #include <linux/file.h>
  23 #include <linux/fs.h>
  24 #include <linux/pagemap.h>
  25 #include <linux/highmem.h>
  26 #include <linux/time.h>
  27 #include <linux/init.h>
  28 #include <linux/string.h>
  29 #include <linux/backing-dev.h>
  30 #include <linux/mpage.h>
  31 #include <linux/swap.h>
  32 #include <linux/writeback.h>
  33 #include <linux/statfs.h>
  34 #include <linux/compat.h>
  35 #include <linux/bit_spinlock.h>
  36 #include <linux/xattr.h>
  37 #include <linux/posix_acl.h>
  38 #include <linux/falloc.h>
  39 #include <linux/slab.h>
  40 #include <linux/ratelimit.h>
  41 #include "compat.h"
  42 #include "ctree.h"
  43 #include "disk-io.h"
  44 #include "transaction.h"
  45 #include "btrfs_inode.h"
  46 #include "ioctl.h"
  47 #include "print-tree.h"
  48 #include "volumes.h"
  49 #include "ordered-data.h"
  50 #include "xattr.h"
  51 #include "tree-log.h"
  52 #include "compression.h"
  53 #include "locking.h"
  54 #include "free-space-cache.h"
  55 #include "inode-map.h"
  56
  57 struct btrfs_iget_args {
  58         u64 ino;
  59         struct btrfs_root *root;
  60 };
  61
  62 static const struct inode_operations btrfs_dir_inode_operations;
  63 static const struct inode_operations btrfs_symlink_inode_operations;
  64 static const struct inode_operations btrfs_dir_ro_inode_operations;
  65 static const struct inode_operations btrfs_special_inode_operations;
  66 static const struct inode_operations btrfs_file_inode_operations;
  67 static const struct address_space_operations btrfs_aops;
  68 static const struct address_space_operations btrfs_symlink_aops;
  69 static const struct file_operations btrfs_dir_file_operations;
  70 static struct extent_io_ops btrfs_extent_io_ops;
  71
  72 static struct kmem_cache *btrfs_inode_cachep;
  73 struct kmem_cache *btrfs_trans_handle_cachep;
  74 struct kmem_cache *btrfs_transaction_cachep;
  75 struct kmem_cache *btrfs_path_cachep;
  76 struct kmem_cache *btrfs_free_space_cachep;
  77
  78 #define S_SHIFT 12
  79 static unsigned char btrfs_type_by_mode[S_IFMT >> S_SHIFT] = {
  80         [S_IFREG >> S_SHIFT]    = BTRFS_FT_REG_FILE,
  81         [S_IFDIR >> S_SHIFT]    = BTRFS_FT_DIR,
  82         [S_IFCHR >> S_SHIFT]    = BTRFS_FT_CHRDEV,
  83         [S_IFBLK >> S_SHIFT]    = BTRFS_FT_BLKDEV,
  84         [S_IFIFO >> S_SHIFT]    = BTRFS_FT_FIFO,
  85         [S_IFSOCK >> S_SHIFT]   = BTRFS_FT_SOCK,
  86         [S_IFLNK >> S_SHIFT]    = BTRFS_FT_SYMLINK,
  87 };
  88
  89 static int btrfs_setsize(struct inode *inode, loff_t newsize);
  90 static int btrfs_truncate(struct inode *inode);
  91 static int btrfs_finish_ordered_io(struct inode *inode, u64 start, u64 end);
  92 static noinline int cow_file_range(struct inode *inode,
  93                                    struct page *locked_page,
  94                                    u64 start, u64 end, int *page_started,
  95                                    unsigned long *nr_written, int unlock);
  96
  97 static int btrfs_init_inode_security(struct btrfs_trans_handle *trans,
  98                                      struct inode *inode,  struct inode *dir,
  99                                      const struct qstr *qstr)
 100 {
 101         int err;
 102
 103         err = btrfs_init_acl(trans, inode, dir);
 104         if (!err)
 105                 err = btrfs_xattr_security_init(trans, inode, dir, qstr);
 106         return err;
 107 }
 108
 109 /*
 110  * this does all the hard work for inserting an inline extent into
 111  * the btree.  The caller should have done a btrfs_drop_extents so that
 112  * no overlapping inline items exist in the btree
 113  */
 114 static noinline int insert_inline_extent(struct btrfs_trans_handle *trans,
 115                                 struct btrfs_root *root, struct inode *inode,
 116                                 u64 start, size_t size, size_t compressed_size,
 117                                 int compress_type,
 118                                 struct page **compressed_pages)
 119 {
 120         struct btrfs_key key;
 121         struct btrfs_path *path;
 122         struct extent_buffer *leaf;
 123         struct page *page = NULL;
 124         char *kaddr;
 125         unsigned long ptr;
 126         struct btrfs_file_extent_item *ei;
 127         int err = 0;
 128         int ret;
 129         size_t cur_size = size;
 130         size_t datasize;
 131         unsigned long offset;
 132
 133         if (compressed_size && compressed_pages)
 134                 cur_size = compressed_size;
 135
 136         path = btrfs_alloc_path();
 137         if (!path)
 138                 return -ENOMEM;
 139
 140         path->leave_spinning = 1;
 141
 142         key.objectid = btrfs_ino(inode);
 143         key.offset = start;
 144         btrfs_set_key_type(&key, BTRFS_EXTENT_DATA_KEY);
 145         datasize = btrfs_file_extent_calc_inline_size(cur_size);
 146
 147         inode_add_bytes(inode, size);
 148         ret = btrfs_insert_empty_item(trans, root, path, &key,
 149                                       datasize);
 150         BUG_ON(ret);
 151         if (ret) {
 152                 err = ret;
 153                 goto fail;
 154         }
 155         leaf = path->nodes[0];
 156         ei = btrfs_item_ptr(leaf, path->slots[0],
 157                             struct btrfs_file_extent_item);
 158         btrfs_set_file_extent_generation(leaf, ei, trans->transid);
 159         btrfs_set_file_extent_type(leaf, ei, BTRFS_FILE_EXTENT_INLINE);
 160         btrfs_set_file_extent_encryption(leaf, ei, 0);
 161         btrfs_set_file_extent_other_encoding(leaf, ei, 0);
 162         btrfs_set_file_extent_ram_bytes(leaf, ei, size);
 163         ptr = btrfs_file_extent_inline_start(ei);
 164
 165         if (compress_type != BTRFS_COMPRESS_NONE) {
 166                 struct page *cpage;
 167                 int i = 0;
 168                 while (compressed_size > 0) {
 169                         cpage = compressed_pages[i];
 170                         cur_size = min_t(unsigned long, compressed_size,
 171                                        PAGE_CACHE_SIZE);
 172
 173                         kaddr = kmap_atomic(cpage, KM_USER0);
 174                         write_extent_buffer(leaf, kaddr, ptr, cur_size);
 175                         kunmap_atomic(kaddr, KM_USER0);
 176
 177                         i++;
 178                         ptr += cur_size;
 179                         compressed_size -= cur_size;
 180                 }
 181                 btrfs_set_file_extent_compression(leaf, ei,
 182                                                   compress_type);
 183         } else {
 184                 page = find_get_page(inode->i_mapping,
 185                                      start >> PAGE_CACHE_SHIFT);
 186                 btrfs_set_file_extent_compression(leaf, ei, 0);
 187                 kaddr = kmap_atomic(page, KM_USER0);
 188                 offset = start & (PAGE_CACHE_SIZE - 1);
 189                 write_extent_buffer(leaf, kaddr + offset, ptr, size);
 190                 kunmap_atomic(kaddr, KM_USER0);
 191                 page_cache_release(page);
 192         }
 193         btrfs_mark_buffer_dirty(leaf);
 194         btrfs_free_path(path);
 195
 196         /*
 197          * we're an inline extent, so nobody can
 198          * extend the file past i_size without locking
 199          * a page we already have locked.
 200          *
 201          * We must do any isize and inode updates
 202          * before we unlock the pages.  Otherwise we
 203          * could end up racing with unlink.
 204          */
 205         BTRFS_I(inode)->disk_i_size = inode->i_size;
 206         btrfs_update_inode(trans, root, inode);
 207
 208         return 0;
 209 fail:
 210         btrfs_free_path(path);
 211         return err;
 212 }
 213
 214
 215 /*
 216  * conditionally insert an inline extent into the file.  This
 217  * does the checks required to make sure the data is small enough
 218  * to fit as an inline extent.
 219  */
 220 static noinline int cow_file_range_inline(struct btrfs_trans_handle *trans,
 221                                  struct btrfs_root *root,
 222                                  struct inode *inode, u64 start, u64 end,
 223                                  size_t compressed_size, int compress_type,
 224                                  struct page **compressed_pages)
 225 {
 226         u64 isize = i_size_read(inode);
 227         u64 actual_end = min(end + 1, isize);
 228         u64 inline_len = actual_end - start;
 229         u64 aligned_end = (end + root->sectorsize - 1) &
 230                         ~((u64)root->sectorsize - 1);
 231         u64 hint_byte;
 232         u64 data_len = inline_len;
 233         int ret;
 234
 235         if (compressed_size)
 236                 data_len = compressed_size;
 237
 238         if (start > 0 ||
 239             actual_end >= PAGE_CACHE_SIZE ||
 240             data_len >= BTRFS_MAX_INLINE_DATA_SIZE(root) ||
 241             (!compressed_size &&
 242             (actual_end & (root->sectorsize - 1)) == 0) ||
 243             end + 1 < isize ||
 244             data_len > root->fs_info->max_inline) {
 245                 return 1;
 246         }
 247
 248         ret = btrfs_drop_extents(trans, inode, start, aligned_end,
 249                                  &hint_byte, 1);
 250         BUG_ON(ret);
 251
 252         if (isize > actual_end)
 253                 inline_len = min_t(u64, isize, actual_end);
 254         ret = insert_inline_extent(trans, root, inode, start,
 255                                    inline_len, compressed_size,
 256                                    compress_type, compressed_pages);
 257         BUG_ON(ret);
 258         btrfs_delalloc_release_metadata(inode, end + 1 - start);
 259         btrfs_drop_extent_cache(inode, start, aligned_end - 1, 0);
 260         return 0;
 261 }
 262
 263 struct async_extent {
 264         u64 start;
 265         u64 ram_size;
 266         u64 compressed_size;
 267         struct page **pages;
 268         unsigned long nr_pages;
 269         int compress_type;
 270         struct list_head list;
 271 };
 272
 273 struct async_cow {
 274         struct inode *inode;
 275         struct btrfs_root *root;
 276         struct page *locked_page;
 277         u64 start;
 278         u64 end;
 279         struct list_head extents;
 280         struct btrfs_work work;
 281 };
 282
 283 static noinline int add_async_extent(struct async_cow *cow,
 284                                      u64 start, u64 ram_size,
 285                                      u64 compressed_size,
 286                                      struct page **pages,
 287                                      unsigned long nr_pages,
 288                                      int compress_type)
 289 {
 290         struct async_extent *async_extent;
 291
 292         async_extent = kmalloc(sizeof(*async_extent), GFP_NOFS);
 293         BUG_ON(!async_extent);
 294         async_extent->start = start;
 295         async_extent->ram_size = ram_size;
 296         async_extent->compressed_size = compressed_size;
 297         async_extent->pages = pages;
 298         async_extent->nr_pages = nr_pages;
 299         async_extent->compress_type = compress_type;
 300         list_add_tail(&async_extent->list, &cow->extents);
 301         return 0;
 302 }
 303
 304 /*
 305  * we create compressed extents in two phases.  The first
 306  * phase compresses a range of pages that have already been
 307  * locked (both pages and state bits are locked).
 308  *
 309  * This is done inside an ordered work queue, and the compression
 310  * is spread across many cpus.  The actual IO submission is step
 311  * two, and the ordered work queue takes care of making sure that
 312  * happens in the same order things were put onto the queue by
 313  * writepages and friends.
 314  *
 315  * If this code finds it can't get good compression, it puts an
 316  * entry onto the work queue to write the uncompressed bytes.  This
 317  * makes sure that both compressed inodes and uncompressed inodes
 318  * are written in the same order that pdflush sent them down.
 319  */
 320 static noinline int compress_file_range(struct inode *inode,
 321                                         struct page *locked_page,
 322                                         u64 start, u64 end,
 323                                         struct async_cow *async_cow,
 324                                         int *num_added)
 325 {
 326         struct btrfs_root *root = BTRFS_I(inode)->root;
 327         struct btrfs_trans_handle *trans;
 328         u64 num_bytes;
 329         u64 blocksize = root->sectorsize;
 330         u64 actual_end;
 331         u64 isize = i_size_read(inode);
 332         int ret = 0;
 333         struct page **pages = NULL;
 334         unsigned long nr_pages;
 335         unsigned long nr_pages_ret = 0;
 336         unsigned long total_compressed = 0;
 337         unsigned long total_in = 0;
 338         unsigned long max_compressed = 128 * 1024;
 339         unsigned long max_uncompressed = 128 * 1024;
 340         int i;
 341         int will_compress;
 342         int compress_type = root->fs_info->compress_type;
 343
 344         /* if this is a small write inside eof, kick off a defragbot */
 345         if (end <= BTRFS_I(inode)->disk_i_size && (end - start + 1) < 16 * 1024)
 346                 btrfs_add_inode_defrag(NULL, inode);
 347
 348         actual_end = min_t(u64, isize, end + 1);
 349 again:
 350         will_compress = 0;
 351         nr_pages = (end >> PAGE_CACHE_SHIFT) - (start >> PAGE_CACHE_SHIFT) + 1;
 352         nr_pages = min(nr_pages, (128 * 1024UL) / PAGE_CACHE_SIZE);
 353
 354         /*
 355          * we don't want to send crud past the end of i_size through
 356          * compression, that's just a waste of CPU time.  So, if the
 357          * end of the file is before the start of our current
 358          * requested range of bytes, we bail out to the uncompressed
 359          * cleanup code that can deal with all of this.
 360          *
 361          * It isn't really the fastest way to fix things, but this is a
 362          * very uncommon corner.
 363          */
 364         if (actual_end <= start)
 365                 goto cleanup_and_bail_uncompressed;
 366
 367         total_compressed = actual_end - start;
 368
 369         /* we want to make sure that amount of ram required to uncompress
 370          * an extent is reasonable, so we limit the total size in ram
 371          * of a compressed extent to 128k.  This is a crucial number
 372          * because it also controls how easily we can spread reads across
 373          * cpus for decompression.
 374          *
 375          * We also want to make sure the amount of IO required to do
 376          * a random read is reasonably small, so we limit the size of
 377          * a compressed extent to 128k.
 378          */
 379         total_compressed = min(total_compressed, max_uncompressed);
 380         num_bytes = (end - start + blocksize) & ~(blocksize - 1);
 381         num_bytes = max(blocksize,  num_bytes);
 382         total_in = 0;
 383         ret = 0;
 384
 385         /*
 386          * we do compression for mount -o compress and when the
 387          * inode has not been flagged as nocompress.  This flag can
 388          * change at any time if we discover bad compression ratios.
 389          */
 390         if (!(BTRFS_I(inode)->flags & BTRFS_INODE_NOCOMPRESS) &&
 391             (btrfs_test_opt(root, COMPRESS) ||
 392              (BTRFS_I(inode)->force_compress) ||
 393              (BTRFS_I(inode)->flags & BTRFS_INODE_COMPRESS))) {
 394                 WARN_ON(pages);
 395                 pages = kzalloc(sizeof(struct page *) * nr_pages, GFP_NOFS);
 396                 BUG_ON(!pages);
 397
 398                 if (BTRFS_I(inode)->force_compress)
 399                         compress_type = BTRFS_I(inode)->force_compress;
 400
 401                 ret = btrfs_compress_pages(compress_type,
 402                                            inode->i_mapping, start,
 403                                            total_compressed, pages,
 404                                            nr_pages, &nr_pages_ret,
 405                                            &total_in,
 406                                            &total_compressed,
 407                                            max_compressed);
 408
 409                 if (!ret) {
 410                         unsigned long offset = total_compressed &
 411                                 (PAGE_CACHE_SIZE - 1);
 412                         struct page *page = pages[nr_pages_ret - 1];
 413                         char *kaddr;
 414
 415                         /* zero the tail end of the last page, we might be
 416                          * sending it down to disk
 417                          */
 418                         if (offset) {
 419                                 kaddr = kmap_atomic(page, KM_USER0);
 420                                 memset(kaddr + offset, 0,
 421                                        PAGE_CACHE_SIZE - offset);
 422                                 kunmap_atomic(kaddr, KM_USER0);
 423                         }
 424                         will_compress = 1;
 425                 }
 426         }
 427         if (start == 0) {
 428                 trans = btrfs_join_transaction(root);
 429                 BUG_ON(IS_ERR(trans));
 430                 trans->block_rsv = &root->fs_info->delalloc_block_rsv;
 431
 432                 /* lets try to make an inline extent */
 433                 if (ret || total_in < (actual_end - start)) {
 434                         /* we didn't compress the entire range, try
 435                          * to make an uncompressed inline extent.
 436                          */
 437                         ret = cow_file_range_inline(trans, root, inode,
 438                                                     start, end, 0, 0, NULL);
 439                 } else {
 440                         /* try making a compressed inline extent */
 441                         ret = cow_file_range_inline(trans, root, inode,
 442                                                     start, end,
 443                                                     total_compressed,
 444                                                     compress_type, pages);
 445                 }
 446                 if (ret == 0) {
 447                         /*
 448                          * inline extent creation worked, we don't need
 449                          * to create any more async work items.  Unlock
 450                          * and free up our temp pages.
 451                          */
 452                         extent_clear_unlock_delalloc(inode,
 453                              &BTRFS_I(inode)->io_tree,
 454                              start, end, NULL,
 455                              EXTENT_CLEAR_UNLOCK_PAGE | EXTENT_CLEAR_DIRTY |
 456                              EXTENT_CLEAR_DELALLOC |
 457                              EXTENT_SET_WRITEBACK | EXTENT_END_WRITEBACK);
 458
 459                         btrfs_end_transaction(trans, root);
 460                         goto free_pages_out;
 461                 }
 462                 btrfs_end_transaction(trans, root);
 463         }
 464
 465         if (will_compress) {
 466                 /*
 467                  * we aren't doing an inline extent round the compressed size
 468                  * up to a block size boundary so the allocator does sane
 469                  * things
 470                  */
 471                 total_compressed = (total_compressed + blocksize - 1) &
 472                         ~(blocksize - 1);
 473
 474                 /*
 475                  * one last check to make sure the compression is really a
 476                  * win, compare the page count read with the blocks on disk
 477                  */
 478                 total_in = (total_in + PAGE_CACHE_SIZE - 1) &
 479                         ~(PAGE_CACHE_SIZE - 1);
 480                 if (total_compressed >= total_in) {
 481                         will_compress = 0;
 482                 } else {
 483                         num_bytes = total_in;
 484                 }
 485         }
 486         if (!will_compress && pages) {
 487                 /*
 488                  * the compression code ran but failed to make things smaller,
 489                  * free any pages it allocated and our page pointer array
 490                  */
 491                 for (i = 0; i < nr_pages_ret; i++) {
 492                         WARN_ON(pages[i]->mapping);
 493                         page_cache_release(pages[i]);
 494                 }
 495                 kfree(pages);
 496                 pages = NULL;
 497                 total_compressed = 0;
 498                 nr_pages_ret = 0;
 499
 500                 /* flag the file so we don't compress in the future */
 501                 if (!btrfs_test_opt(root, FORCE_COMPRESS) &&
 502                     !(BTRFS_I(inode)->force_compress)) {
 503                         BTRFS_I(inode)->flags |= BTRFS_INODE_NOCOMPRESS;
 504                 }
 505         }
 506         if (will_compress) {
 507                 *num_added += 1;
 508
 509                 /* the async work queues will take care of doing actual
 510                  * allocation on disk for these compressed pages,
 511                  * and will submit them to the elevator.
 512                  */
 513                 add_async_extent(async_cow, start, num_bytes,
 514                                  total_compressed, pages, nr_pages_ret,
 515                                  compress_type);
 516
 517                 if (start + num_bytes < end) {
 518                         start += num_bytes;
 519                         pages = NULL;
 520                         cond_resched();
 521                         goto again;
 522                 }
 523         } else {
 524 cleanup_and_bail_uncompressed:
 525                 /*
 526                  * No compression, but we still need to write the pages in
 527                  * the file we've been given so far.  redirty the locked
 528                  * page if it corresponds to our extent and set things up
 529                  * for the async work queue to run cow_file_range to do
 530                  * the normal delalloc dance
 531                  */
 532                 if (page_offset(locked_page) >= start &&
 533                     page_offset(locked_page) <= end) {
 534                         __set_page_dirty_nobuffers(locked_page);
 535                         /* unlocked later on in the async handlers */
 536                 }
 537                 add_async_extent(async_cow, start, end - start + 1,
 538                                  0, NULL, 0, BTRFS_COMPRESS_NONE);
 539                 *num_added += 1;
 540         }
 541
 542 out:
 543         return 0;
 544
 545 free_pages_out:
 546         for (i = 0; i < nr_pages_ret; i++) {
 547                 WARN_ON(pages[i]->mapping);
 548                 page_cache_release(pages[i]);
 549         }
 550         kfree(pages);
 551
 552         goto out;
 553 }
 554
 555 /*
 556  * phase two of compressed writeback.  This is the ordered portion
 557  * of the code, which only gets called in the order the work was
 558  * queued.  We walk all the async extents created by compress_file_range
 559  * and send them down to the disk.
 560  */
 561 static noinline int submit_compressed_extents(struct inode *inode,
 562                                               struct async_cow *async_cow)
 563 {
 564         struct async_extent *async_extent;
 565         u64 alloc_hint = 0;
 566         struct btrfs_trans_handle *trans;
 567         struct btrfs_key ins;
 568         struct extent_map *em;
 569         struct btrfs_root *root = BTRFS_I(inode)->root;
 570         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
 571         struct extent_io_tree *io_tree;
 572         int ret = 0;
 573
 574         if (list_empty(&async_cow->extents))
 575                 return 0;
 576
 577
 578         while (!list_empty(&async_cow->extents)) {
 579                 async_extent = list_entry(async_cow->extents.next,
 580                                           struct async_extent, list);
 581                 list_del(&async_extent->list);
 582
 583                 io_tree = &BTRFS_I(inode)->io_tree;
 584
 585 retry:
 586                 /* did the compression code fall back to uncompressed IO? */
 587                 if (!async_extent->pages) {
 588                         int page_started = 0;
 589                         unsigned long nr_written = 0;
 590
 591                         lock_extent(io_tree, async_extent->start,
 592                                          async_extent->start +
 593                                          async_extent->ram_size - 1, GFP_NOFS);
 594
 595                         /* allocate blocks */
 596                         ret = cow_file_range(inode, async_cow->locked_page,
 597                                              async_extent->start,
 598                                              async_extent->start +
 599                                              async_extent->ram_size - 1,
 600                                              &page_started, &nr_written, 0);
 601
 602                         /*
 603                          * if page_started, cow_file_range inserted an
 604                          * inline extent and took care of all the unlocking
 605                          * and IO for us.  Otherwise, we need to submit
 606                          * all those pages down to the drive.
 607                          */
 608                         if (!page_started && !ret)
 609                                 extent_write_locked_range(io_tree,
 610                                                   inode, async_extent->start,
 611                                                   async_extent->start +
 612                                                   async_extent->ram_size - 1,
 613                                                   btrfs_get_extent,
 614                                                   WB_SYNC_ALL);
 615                         kfree(async_extent);
 616                         cond_resched();
 617                         continue;
 618                 }
 619
 620                 lock_extent(io_tree, async_extent->start,
 621                             async_extent->start + async_extent->ram_size - 1,
 622                             GFP_NOFS);
 623
 624                 trans = btrfs_join_transaction(root);
 625                 BUG_ON(IS_ERR(trans));
 626                 trans->block_rsv = &root->fs_info->delalloc_block_rsv;
 627                 ret = btrfs_reserve_extent(trans, root,
 628                                            async_extent->compressed_size,
 629                                            async_extent->compressed_size,
 630                                            0, alloc_hint,
 631                                            (u64)-1, &ins, 1);
 632                 btrfs_end_transaction(trans, root);
 633
 634                 if (ret) {
 635                         int i;
 636                         for (i = 0; i < async_extent->nr_pages; i++) {
 637                                 WARN_ON(async_extent->pages[i]->mapping);
 638                                 page_cache_release(async_extent->pages[i]);
 639                         }
 640                         kfree(async_extent->pages);
 641                         async_extent->nr_pages = 0;
 642                         async_extent->pages = NULL;
 643                         unlock_extent(io_tree, async_extent->start,
 644                                       async_extent->start +
 645                                       async_extent->ram_size - 1, GFP_NOFS);
 646                         goto retry;
 647                 }
 648
 649                 /*
 650                  * here we're doing allocation and writeback of the
 651                  * compressed pages
 652                  */
 653                 btrfs_drop_extent_cache(inode, async_extent->start,
 654                                         async_extent->start +
 655                                         async_extent->ram_size - 1, 0);
 656
 657                 em = alloc_extent_map();
 658                 BUG_ON(!em);
 659                 em->start = async_extent->start;
 660                 em->len = async_extent->ram_size;
 661                 em->orig_start = em->start;
 662
 663                 em->block_start = ins.objectid;
 664                 em->block_len = ins.offset;
 665                 em->bdev = root->fs_info->fs_devices->latest_bdev;
 666                 em->compress_type = async_extent->compress_type;
 667                 set_bit(EXTENT_FLAG_PINNED, &em->flags);
 668                 set_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
 669
 670                 while (1) {
 671                         write_lock(&em_tree->lock);
 672                         ret = add_extent_mapping(em_tree, em);
 673                         write_unlock(&em_tree->lock);
 674                         if (ret != -EEXIST) {
 675                                 free_extent_map(em);
 676                                 break;
 677                         }
 678                         btrfs_drop_extent_cache(inode, async_extent->start,
 679                                                 async_extent->start +
 680                                                 async_extent->ram_size - 1, 0);
 681                 }
 682
 683                 ret = btrfs_add_ordered_extent_compress(inode,
 684                                                 async_extent->start,
 685                                                 ins.objectid,
 686                                                 async_extent->ram_size,
 687                                                 ins.offset,
 688                                                 BTRFS_ORDERED_COMPRESSED,
 689                                                 async_extent->compress_type);
 690                 BUG_ON(ret);
 691
 692                 /*
 693                  * clear dirty, set writeback and unlock the pages.
 694                  */
 695                 extent_clear_unlock_delalloc(inode,
 696                                 &BTRFS_I(inode)->io_tree,
 697                                 async_extent->start,
 698                                 async_extent->start +
 699                                 async_extent->ram_size - 1,
 700                                 NULL, EXTENT_CLEAR_UNLOCK_PAGE |
 701                                 EXTENT_CLEAR_UNLOCK |
 702                                 EXTENT_CLEAR_DELALLOC |
 703                                 EXTENT_CLEAR_DIRTY | EXTENT_SET_WRITEBACK);
 704
 705                 ret = btrfs_submit_compressed_write(inode,
 706                                     async_extent->start,
 707                                     async_extent->ram_size,
 708                                     ins.objectid,
 709                                     ins.offset, async_extent->pages,
 710                                     async_extent->nr_pages);
 711
 712                 BUG_ON(ret);
 713                 alloc_hint = ins.objectid + ins.offset;
 714                 kfree(async_extent);
 715                 cond_resched();
 716         }
 717
 718         return 0;
 719 }
 720
 721 static u64 get_extent_allocation_hint(struct inode *inode, u64 start,
 722                                       u64 num_bytes)
 723 {
 724         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
 725         struct extent_map *em;
 726         u64 alloc_hint = 0;
 727
 728         read_lock(&em_tree->lock);
 729         em = search_extent_mapping(em_tree, start, num_bytes);
 730         if (em) {
 731                 /*
 732                  * if block start isn't an actual block number then find the
 733                  * first block in this inode and use that as a hint.  If that
 734                  * block is also bogus then just don't worry about it.
 735                  */
 736                 if (em->block_start >= EXTENT_MAP_LAST_BYTE) {
 737                         free_extent_map(em);
 738                         em = search_extent_mapping(em_tree, 0, 0);
 739                         if (em && em->block_start < EXTENT_MAP_LAST_BYTE)
 740                                 alloc_hint = em->block_start;
 741                         if (em)
 742                                 free_extent_map(em);
 743                 } else {
 744                         alloc_hint = em->block_start;
 745                         free_extent_map(em);
 746                 }
 747         }
 748         read_unlock(&em_tree->lock);
 749
 750         return alloc_hint;
 751 }
 752
 753 /*
 754  * when extent_io.c finds a delayed allocation range in the file,
 755  * the call backs end up in this code.  The basic idea is to
 756  * allocate extents on disk for the range, and create ordered data structs
 757  * in ram to track those extents.
 758  *
 759  * locked_page is the page that writepage had locked already.  We use
 760  * it to make sure we don't do extra locks or unlocks.
 761  *
 762  * *page_started is set to one if we unlock locked_page and do everything
 763  * required to start IO on it.  It may be clean and already done with
 764  * IO when we return.
 765  */
 766 static noinline int cow_file_range(struct inode *inode,
 767                                    struct page *locked_page,
 768                                    u64 start, u64 end, int *page_started,
 769                                    unsigned long *nr_written,
 770                                    int unlock)
 771 {
 772         struct btrfs_root *root = BTRFS_I(inode)->root;
 773         struct btrfs_trans_handle *trans;
 774         u64 alloc_hint = 0;
 775         u64 num_bytes;
 776         unsigned long ram_size;
 777         u64 disk_num_bytes;
 778         u64 cur_alloc_size;
 779         u64 blocksize = root->sectorsize;
 780         struct btrfs_key ins;
 781         struct extent_map *em;
 782         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
 783         int ret = 0;
 784
 785         BUG_ON(btrfs_is_free_space_inode(root, inode));
 786         trans = btrfs_join_transaction(root);
 787         BUG_ON(IS_ERR(trans));
 788         trans->block_rsv = &root->fs_info->delalloc_block_rsv;
 789
 790         num_bytes = (end - start + blocksize) & ~(blocksize - 1);
 791         num_bytes = max(blocksize,  num_bytes);
 792         disk_num_bytes = num_bytes;
 793         ret = 0;
 794
 795         /* if this is a small write inside eof, kick off defrag */
 796         if (end <= BTRFS_I(inode)->disk_i_size && num_bytes < 64 * 1024)
 797                 btrfs_add_inode_defrag(trans, inode);
 798
 799         if (start == 0) {
 800                 /* lets try to make an inline extent */
 801                 ret = cow_file_range_inline(trans, root, inode,
 802                                             start, end, 0, 0, NULL);
 803                 if (ret == 0) {
 804                         extent_clear_unlock_delalloc(inode,
 805                                      &BTRFS_I(inode)->io_tree,
 806                                      start, end, NULL,
 807                                      EXTENT_CLEAR_UNLOCK_PAGE |
 808                                      EXTENT_CLEAR_UNLOCK |
 809                                      EXTENT_CLEAR_DELALLOC |
 810                                      EXTENT_CLEAR_DIRTY |
 811                                      EXTENT_SET_WRITEBACK |
 812                                      EXTENT_END_WRITEBACK);
 813
 814                         *nr_written = *nr_written +
 815                              (end - start + PAGE_CACHE_SIZE) / PAGE_CACHE_SIZE;
 816                         *page_started = 1;
 817                         ret = 0;
 818                         goto out;
 819                 }
 820         }
 821
 822         BUG_ON(disk_num_bytes >
 823                btrfs_super_total_bytes(&root->fs_info->super_copy));
 824
 825         alloc_hint = get_extent_allocation_hint(inode, start, num_bytes);
 826         btrfs_drop_extent_cache(inode, start, start + num_bytes - 1, 0);
 827
 828         while (disk_num_bytes > 0) {
 829                 unsigned long op;
 830
 831                 cur_alloc_size = disk_num_bytes;
 832                 ret = btrfs_reserve_extent(trans, root, cur_alloc_size,
 833                                            root->sectorsize, 0, alloc_hint,
 834                                            (u64)-1, &ins, 1);
 835                 BUG_ON(ret);
 836
 837                 em = alloc_extent_map();
 838                 BUG_ON(!em);
 839                 em->start = start;
 840                 em->orig_start = em->start;
 841                 ram_size = ins.offset;
 842                 em->len = ins.offset;
 843
 844                 em->block_start = ins.objectid;
 845                 em->block_len = ins.offset;
 846                 em->bdev = root->fs_info->fs_devices->latest_bdev;
 847                 set_bit(EXTENT_FLAG_PINNED, &em->flags);
 848
 849                 while (1) {
 850                         write_lock(&em_tree->lock);
 851                         ret = add_extent_mapping(em_tree, em);
 852                         write_unlock(&em_tree->lock);
 853                         if (ret != -EEXIST) {
 854                                 free_extent_map(em);
 855                                 break;
 856                         }
 857                         btrfs_drop_extent_cache(inode, start,
 858                                                 start + ram_size - 1, 0);
 859                 }
 860
 861                 cur_alloc_size = ins.offset;
 862                 ret = btrfs_add_ordered_extent(inode, start, ins.objectid,
 863                                                ram_size, cur_alloc_size, 0);
 864                 BUG_ON(ret);
 865
 866                 if (root->root_key.objectid ==
 867                     BTRFS_DATA_RELOC_TREE_OBJECTID) {
 868                         ret = btrfs_reloc_clone_csums(inode, start,
 869                                                       cur_alloc_size);
 870                         BUG_ON(ret);
 871                 }
 872
 873                 if (disk_num_bytes < cur_alloc_size)
 874                         break;
 875
 876                 /* we're not doing compressed IO, don't unlock the first
 877                  * page (which the caller expects to stay locked), don't
 878                  * clear any dirty bits and don't set any writeback bits
 879                  *
 880                  * Do set the Private2 bit so we know this page was properly
 881                  * setup for writepage
 882                  */
 883                 op = unlock ? EXTENT_CLEAR_UNLOCK_PAGE : 0;
 884                 op |= EXTENT_CLEAR_UNLOCK | EXTENT_CLEAR_DELALLOC |
 885                         EXTENT_SET_PRIVATE2;
 886
 887                 extent_clear_unlock_delalloc(inode, &BTRFS_I(inode)->io_tree,
 888                                              start, start + ram_size - 1,
 889                                              locked_page, op);
 890                 disk_num_bytes -= cur_alloc_size;
 891                 num_bytes -= cur_alloc_size;
 892                 alloc_hint = ins.objectid + ins.offset;
 893                 start += cur_alloc_size;
 894         }
 895 out:
 896         ret = 0;
 897         btrfs_end_transaction(trans, root);
 898
 899         return ret;
 900 }
 901
 902 /*
 903  * work queue call back to started compression on a file and pages
 904  */
 905 static noinline void async_cow_start(struct btrfs_work *work)
 906 {
 907         struct async_cow *async_cow;
 908         int num_added = 0;
 909         async_cow = container_of(work, struct async_cow, work);
 910
 911         compress_file_range(async_cow->inode, async_cow->locked_page,
 912                             async_cow->start, async_cow->end, async_cow,
 913                             &num_added);
 914         if (num_added == 0)
 915                 async_cow->inode = NULL;
 916 }
 917
 918 /*
 919  * work queue call back to submit previously compressed pages
 920  */
 921 static noinline void async_cow_submit(struct btrfs_work *work)
 922 {
 923         struct async_cow *async_cow;
 924         struct btrfs_root *root;
 925         unsigned long nr_pages;
 926
 927         async_cow = container_of(work, struct async_cow, work);
 928
 929         root = async_cow->root;
 930         nr_pages = (async_cow->end - async_cow->start + PAGE_CACHE_SIZE) >>
 931                 PAGE_CACHE_SHIFT;
 932
 933         atomic_sub(nr_pages, &root->fs_info->async_delalloc_pages);
 934
 935         if (atomic_read(&root->fs_info->async_delalloc_pages) <
 936             5 * 1042 * 1024 &&
 937             waitqueue_active(&root->fs_info->async_submit_wait))
 938                 wake_up(&root->fs_info->async_submit_wait);
 939
 940         if (async_cow->inode)
 941                 submit_compressed_extents(async_cow->inode, async_cow);
 942 }
 943
 944 static noinline void async_cow_free(struct btrfs_work *work)
 945 {
 946         struct async_cow *async_cow;
 947         async_cow = container_of(work, struct async_cow, work);
 948         kfree(async_cow);
 949 }
 950
 951 static int cow_file_range_async(struct inode *inode, struct page *locked_page,
 952                                 u64 start, u64 end, int *page_started,
 953                                 unsigned long *nr_written)
 954 {
 955         struct async_cow *async_cow;
 956         struct btrfs_root *root = BTRFS_I(inode)->root;
 957         unsigned long nr_pages;
 958         u64 cur_end;
 959         int limit = 10 * 1024 * 1042;
 960
 961         clear_extent_bit(&BTRFS_I(inode)->io_tree, start, end, EXTENT_LOCKED,
 962                          1, 0, NULL, GFP_NOFS);
 963         while (start < end) {
 964                 async_cow = kmalloc(sizeof(*async_cow), GFP_NOFS);
 965                 BUG_ON(!async_cow);
 966                 async_cow->inode = inode;
 967                 async_cow->root = root;
 968                 async_cow->locked_page = locked_page;
 969                 async_cow->start = start;
 970
 971                 if (BTRFS_I(inode)->flags & BTRFS_INODE_NOCOMPRESS)
 972                         cur_end = end;
 973                 else
 974                         cur_end = min(end, start + 512 * 1024 - 1);
 975
 976                 async_cow->end = cur_end;
 977                 INIT_LIST_HEAD(&async_cow->extents);
 978
 979                 async_cow->work.func = async_cow_start;
 980                 async_cow->work.ordered_func = async_cow_submit;
 981                 async_cow->work.ordered_free = async_cow_free;
 982                 async_cow->work.flags = 0;
 983
 984                 nr_pages = (cur_end - start + PAGE_CACHE_SIZE) >>
 985                         PAGE_CACHE_SHIFT;
 986                 atomic_add(nr_pages, &root->fs_info->async_delalloc_pages);
 987
 988                 btrfs_queue_worker(&root->fs_info->delalloc_workers,
 989                                    &async_cow->work);
 990
 991                 if (atomic_read(&root->fs_info->async_delalloc_pages) > limit) {
 992                         wait_event(root->fs_info->async_submit_wait,
 993                            (atomic_read(&root->fs_info->async_delalloc_pages) <
 994                             limit));
 995                 }
 996
 997                 while (atomic_read(&root->fs_info->async_submit_draining) &&
 998                       atomic_read(&root->fs_info->async_delalloc_pages)) {
 999                         wait_event(root->fs_info->async_submit_wait,
1000                           (atomic_read(&root->fs_info->async_delalloc_pages) ==
1001                            0));
1002                 }
1003
1004                 *nr_written += nr_pages;
1005                 start = cur_end + 1;
1006         }
1007         *page_started = 1;
1008         return 0;
1009 }
1010
1011 static noinline int csum_exist_in_range(struct btrfs_root *root,
1012                                         u64 bytenr, u64 num_bytes)
1013 {
1014         int ret;
1015         struct btrfs_ordered_sum *sums;
1016         LIST_HEAD(list);
1017
1018         ret = btrfs_lookup_csums_range(root->fs_info->csum_root, bytenr,
1019                                        bytenr + num_bytes - 1, &list, 0);
1020         if (ret == 0 && list_empty(&list))
1021                 return 0;
1022
1023         while (!list_empty(&list)) {
1024                 sums = list_entry(list.next, struct btrfs_ordered_sum, list);
1025                 list_del(&sums->list);
1026                 kfree(sums);
1027         }
1028         return 1;
1029 }
1030
1031 /*
1032  * when nowcow writeback call back.  This checks for snapshots or COW copies
1033  * of the extents that exist in the file, and COWs the file as required.
1034  *
1035  * If no cow copies or snapshots exist, we write directly to the existing
1036  * blocks on disk
1037  */
1038 static noinline int run_delalloc_nocow(struct inode *inode,
1039                                        struct page *locked_page,
1040                               u64 start, u64 end, int *page_started, int force,
1041                               unsigned long *nr_written)
1042 {
1043         struct btrfs_root *root = BTRFS_I(inode)->root;
1044         struct btrfs_trans_handle *trans;
1045         struct extent_buffer *leaf;
1046         struct btrfs_path *path;
1047         struct btrfs_file_extent_item *fi;
1048         struct btrfs_key found_key;
1049         u64 cow_start;
1050         u64 cur_offset;
1051         u64 extent_end;
1052         u64 extent_offset;
1053         u64 disk_bytenr;
1054         u64 num_bytes;
1055         int extent_type;
1056         int ret;
1057         int type;
1058         int nocow;
1059         int check_prev = 1;
1060         bool nolock;
1061         u64 ino = btrfs_ino(inode);
1062
1063         path = btrfs_alloc_path();
1064         if (!path)
1065                 return -ENOMEM;
1066
1067         nolock = btrfs_is_free_space_inode(root, inode);
1068
1069         if (nolock)
1070                 trans = btrfs_join_transaction_nolock(root);
1071         else
1072                 trans = btrfs_join_transaction(root);
1073
1074         BUG_ON(IS_ERR(trans));
1075         trans->block_rsv = &root->fs_info->delalloc_block_rsv;
1076
1077         cow_start = (u64)-1;
1078         cur_offset = start;
1079         while (1) {
1080                 ret = btrfs_lookup_file_extent(trans, root, path, ino,
1081                                                cur_offset, 0);
1082                 BUG_ON(ret < 0);
1083                 if (ret > 0 && path->slots[0] > 0 && check_prev) {
1084                         leaf = path->nodes[0];
1085                         btrfs_item_key_to_cpu(leaf, &found_key,
1086                                               path->slots[0] - 1);
1087                         if (found_key.objectid == ino &&
1088                             found_key.type == BTRFS_EXTENT_DATA_KEY)
1089                                 path->slots[0]--;
1090                 }
1091                 check_prev = 0;
1092 next_slot:
1093                 leaf = path->nodes[0];
1094                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
1095                         ret = btrfs_next_leaf(root, path);
1096                         if (ret < 0)
1097                                 BUG_ON(1);
1098                         if (ret > 0)
1099                                 break;
1100                         leaf = path->nodes[0];
1101                 }
1102
1103                 nocow = 0;
1104                 disk_bytenr = 0;
1105                 num_bytes = 0;
1106                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
1107
1108                 if (found_key.objectid > ino ||
1109                     found_key.type > BTRFS_EXTENT_DATA_KEY ||
1110                     found_key.offset > end)
1111                         break;
1112
1113                 if (found_key.offset > cur_offset) {
1114                         extent_end = found_key.offset;
1115                         extent_type = 0;
1116                         goto out_check;
1117                 }
1118
1119                 fi = btrfs_item_ptr(leaf, path->slots[0],
1120                                     struct btrfs_file_extent_item);
1121                 extent_type = btrfs_file_extent_type(leaf, fi);
1122
1123                 if (extent_type == BTRFS_FILE_EXTENT_REG ||
1124                     extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
1125                         disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1126                         extent_offset = btrfs_file_extent_offset(leaf, fi);
1127                         extent_end = found_key.offset +
1128                                 btrfs_file_extent_num_bytes(leaf, fi);
1129                         if (extent_end <= start) {
1130                                 path->slots[0]++;
1131                                 goto next_slot;
1132                         }
1133                         if (disk_bytenr == 0)
1134                                 goto out_check;
1135                         if (btrfs_file_extent_compression(leaf, fi) ||
1136                             btrfs_file_extent_encryption(leaf, fi) ||
1137                             btrfs_file_extent_other_encoding(leaf, fi))
1138                                 goto out_check;
1139                         if (extent_type == BTRFS_FILE_EXTENT_REG && !force)
1140                                 goto out_check;
1141                         if (btrfs_extent_readonly(root, disk_bytenr))
1142                                 goto out_check;
1143                         if (btrfs_cross_ref_exist(trans, root, ino,
1144                                                   found_key.offset -
1145                                                   extent_offset, disk_bytenr))
1146                                 goto out_check;
1147                         disk_bytenr += extent_offset;
1148                         disk_bytenr += cur_offset - found_key.offset;
1149                         num_bytes = min(end + 1, extent_end) - cur_offset;
1150                         /*
1151                          * force cow if csum exists in the range.
1152                          * this ensure that csum for a given extent are
1153                          * either valid or do not exist.
1154                          */
1155                         if (csum_exist_in_range(root, disk_bytenr, num_bytes))
1156                                 goto out_check;
1157                         nocow = 1;
1158                 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
1159                         extent_end = found_key.offset +
1160                                 btrfs_file_extent_inline_len(leaf, fi);
1161                         extent_end = ALIGN(extent_end, root->sectorsize);
1162                 } else {
1163                         BUG_ON(1);
1164                 }
1165 out_check:
1166                 if (extent_end <= start) {
1167                         path->slots[0]++;
1168                         goto next_slot;
1169                 }
1170                 if (!nocow) {
1171                         if (cow_start == (u64)-1)
1172                                 cow_start = cur_offset;
1173                         cur_offset = extent_end;
1174                         if (cur_offset > end)
1175                                 break;
1176                         path->slots[0]++;
1177                         goto next_slot;
1178                 }
1179
1180                 btrfs_release_path(path);
1181                 if (cow_start != (u64)-1) {
1182                         ret = cow_file_range(inode, locked_page, cow_start,
1183                                         found_key.offset - 1, page_started,
1184                                         nr_written, 1);
1185                         BUG_ON(ret);
1186                         cow_start = (u64)-1;
1187                 }
1188
1189                 if (extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
1190                         struct extent_map *em;
1191                         struct extent_map_tree *em_tree;
1192                         em_tree = &BTRFS_I(inode)->extent_tree;
1193                         em = alloc_extent_map();
1194                         BUG_ON(!em);
1195                         em->start = cur_offset;
1196                         em->orig_start = em->start;
1197                         em->len = num_bytes;
1198                         em->block_len = num_bytes;
1199                         em->block_start = disk_bytenr;
1200                         em->bdev = root->fs_info->fs_devices->latest_bdev;
1201                         set_bit(EXTENT_FLAG_PINNED, &em->flags);
1202                         while (1) {
1203                                 write_lock(&em_tree->lock);
1204                                 ret = add_extent_mapping(em_tree, em);
1205                                 write_unlock(&em_tree->lock);
1206                                 if (ret != -EEXIST) {
1207                                         free_extent_map(em);
1208                                         break;
1209                                 }
1210                                 btrfs_drop_extent_cache(inode, em->start,
1211                                                 em->start + em->len - 1, 0);
1212                         }
1213                         type = BTRFS_ORDERED_PREALLOC;
1214                 } else {
1215                         type = BTRFS_ORDERED_NOCOW;
1216                 }
1217
1218                 ret = btrfs_add_ordered_extent(inode, cur_offset, disk_bytenr,
1219                                                num_bytes, num_bytes, type);
1220                 BUG_ON(ret);
1221
1222                 if (root->root_key.objectid ==
1223                     BTRFS_DATA_RELOC_TREE_OBJECTID) {
1224                         ret = btrfs_reloc_clone_csums(inode, cur_offset,
1225                                                       num_bytes);
1226                         BUG_ON(ret);
1227                 }
1228
1229                 extent_clear_unlock_delalloc(inode, &BTRFS_I(inode)->io_tree,
1230                                 cur_offset, cur_offset + num_bytes - 1,
1231                                 locked_page, EXTENT_CLEAR_UNLOCK_PAGE |
1232                                 EXTENT_CLEAR_UNLOCK | EXTENT_CLEAR_DELALLOC |
1233                                 EXTENT_SET_PRIVATE2);
1234                 cur_offset = extent_end;
1235                 if (cur_offset > end)
1236                         break;
1237         }
1238         btrfs_release_path(path);
1239
1240         if (cur_offset <= end && cow_start == (u64)-1)
1241                 cow_start = cur_offset;
1242         if (cow_start != (u64)-1) {
1243                 ret = cow_file_range(inode, locked_page, cow_start, end,
1244                                      page_started, nr_written, 1);
1245                 BUG_ON(ret);
1246         }
1247
1248         if (nolock) {
1249                 ret = btrfs_end_transaction_nolock(trans, root);
1250                 BUG_ON(ret);
1251         } else {
1252                 ret = btrfs_end_transaction(trans, root);
1253                 BUG_ON(ret);
1254         }
1255         btrfs_free_path(path);
1256         return 0;
1257 }
1258
1259 /*
1260  * extent_io.c call back to do delayed allocation processing
1261  */
1262 static int run_delalloc_range(struct inode *inode, struct page *locked_page,
1263                               u64 start, u64 end, int *page_started,
1264                               unsigned long *nr_written)
1265 {
1266         int ret;
1267         struct btrfs_root *root = BTRFS_I(inode)->root;
1268
1269         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW)
1270                 ret = run_delalloc_nocow(inode, locked_page, start, end,
1271                                          page_started, 1, nr_written);
1272         else if (BTRFS_I(inode)->flags & BTRFS_INODE_PREALLOC)
1273                 ret = run_delalloc_nocow(inode, locked_page, start, end,
1274                                          page_started, 0, nr_written);
1275         else if (!btrfs_test_opt(root, COMPRESS) &&
1276                  !(BTRFS_I(inode)->force_compress) &&
1277                  !(BTRFS_I(inode)->flags & BTRFS_INODE_COMPRESS))
1278                 ret = cow_file_range(inode, locked_page, start, end,
1279                                       page_started, nr_written, 1);
1280         else
1281                 ret = cow_file_range_async(inode, locked_page, start, end,
1282                                            page_started, nr_written);
1283         return ret;
1284 }
1285
1286 static void btrfs_split_extent_hook(struct inode *inode,
1287                                     struct extent_state *orig, u64 split)
1288 {
1289         /* not delalloc, ignore it */
1290         if (!(orig->state & EXTENT_DELALLOC))
1291                 return;
1292
1293         spin_lock(&BTRFS_I(inode)->lock);
1294         BTRFS_I(inode)->outstanding_extents++;
1295         spin_unlock(&BTRFS_I(inode)->lock);
1296 }
1297
1298 /*
1299  * extent_io.c merge_extent_hook, used to track merged delayed allocation
1300  * extents so we can keep track of new extents that are just merged onto old
1301  * extents, such as when we are doing sequential writes, so we can properly
1302  * account for the metadata space we'll need.
1303  */
1304 static void btrfs_merge_extent_hook(struct inode *inode,
1305                                     struct extent_state *new,
1306                                     struct extent_state *other)
1307 {
1308         /* not delalloc, ignore it */
1309         if (!(other->state & EXTENT_DELALLOC))
1310                 return;
1311
1312         spin_lock(&BTRFS_I(inode)->lock);
1313         BTRFS_I(inode)->outstanding_extents--;
1314         spin_unlock(&BTRFS_I(inode)->lock);
1315 }
1316
1317 /*
1318  * extent_io.c set_bit_hook, used to track delayed allocation
1319  * bytes in this file, and to maintain the list of inodes that
1320  * have pending delalloc work to be done.
1321  */
1322 static void btrfs_set_bit_hook(struct inode *inode,
1323                                struct extent_state *state, int *bits)
1324 {
1325
1326         /*
1327          * set_bit and clear bit hooks normally require _irqsave/restore
1328          * but in this case, we are only testing for the DELALLOC
1329          * bit, which is only set or cleared with irqs on
1330          */
1331         if (!(state->state & EXTENT_DELALLOC) && (*bits & EXTENT_DELALLOC)) {
1332                 struct btrfs_root *root = BTRFS_I(inode)->root;
1333                 u64 len = state->end + 1 - state->start;
1334                 bool do_list = !btrfs_is_free_space_inode(root, inode);
1335
1336                 if (*bits & EXTENT_FIRST_DELALLOC) {
1337                         *bits &= ~EXTENT_FIRST_DELALLOC;
1338                 } else {
1339                         spin_lock(&BTRFS_I(inode)->lock);
1340                         BTRFS_I(inode)->outstanding_extents++;
1341                         spin_unlock(&BTRFS_I(inode)->lock);
1342                 }
1343
1344                 spin_lock(&root->fs_info->delalloc_lock);
1345                 BTRFS_I(inode)->delalloc_bytes += len;
1346                 root->fs_info->delalloc_bytes += len;
1347                 if (do_list && list_empty(&BTRFS_I(inode)->delalloc_inodes)) {
1348                         list_add_tail(&BTRFS_I(inode)->delalloc_inodes,
1349                                       &root->fs_info->delalloc_inodes);
1350                 }
1351                 spin_unlock(&root->fs_info->delalloc_lock);
1352         }
1353 }
1354
1355 /*
1356  * extent_io.c clear_bit_hook, see set_bit_hook for why
1357  */
1358 static void btrfs_clear_bit_hook(struct inode *inode,
1359                                  struct extent_state *state, int *bits)
1360 {
1361         /*
1362          * set_bit and clear bit hooks normally require _irqsave/restore
1363          * but in this case, we are only testing for the DELALLOC
1364          * bit, which is only set or cleared with irqs on
1365          */
1366         if ((state->state & EXTENT_DELALLOC) && (*bits & EXTENT_DELALLOC)) {
1367                 struct btrfs_root *root = BTRFS_I(inode)->root;
1368                 u64 len = state->end + 1 - state->start;
1369                 bool do_list = !btrfs_is_free_space_inode(root, inode);
1370
1371                 if (*bits & EXTENT_FIRST_DELALLOC) {
1372                         *bits &= ~EXTENT_FIRST_DELALLOC;
1373                 } else if (!(*bits & EXTENT_DO_ACCOUNTING)) {
1374                         spin_lock(&BTRFS_I(inode)->lock);
1375                         BTRFS_I(inode)->outstanding_extents--;
1376                         spin_unlock(&BTRFS_I(inode)->lock);
1377                 }
1378
1379                 if (*bits & EXTENT_DO_ACCOUNTING)
1380                         btrfs_delalloc_release_metadata(inode, len);
1381
1382                 if (root->root_key.objectid != BTRFS_DATA_RELOC_TREE_OBJECTID
1383                     && do_list)
1384                         btrfs_free_reserved_data_space(inode, len);
1385
1386                 spin_lock(&root->fs_info->delalloc_lock);
1387                 root->fs_info->delalloc_bytes -= len;
1388                 BTRFS_I(inode)->delalloc_bytes -= len;
1389
1390                 if (do_list && BTRFS_I(inode)->delalloc_bytes == 0 &&
1391                     !list_empty(&BTRFS_I(inode)->delalloc_inodes)) {
1392                         list_del_init(&BTRFS_I(inode)->delalloc_inodes);
1393                 }
1394                 spin_unlock(&root->fs_info->delalloc_lock);
1395         }
1396 }
1397
1398 /*
1399  * extent_io.c merge_bio_hook, this must check the chunk tree to make sure
1400  * we don't create bios that span stripes or chunks
1401  */
1402 int btrfs_merge_bio_hook(struct page *page, unsigned long offset,
1403                          size_t size, struct bio *bio,
1404                          unsigned long bio_flags)
1405 {
1406         struct btrfs_root *root = BTRFS_I(page->mapping->host)->root;
1407         struct btrfs_mapping_tree *map_tree;
1408         u64 logical = (u64)bio->bi_sector << 9;
1409         u64 length = 0;
1410         u64 map_length;
1411         int ret;
1412
1413         if (bio_flags & EXTENT_BIO_COMPRESSED)
1414                 return 0;
1415
1416         length = bio->bi_size;
1417         map_tree = &root->fs_info->mapping_tree;
1418         map_length = length;
1419         ret = btrfs_map_block(map_tree, READ, logical,
1420                               &map_length, NULL, 0);
1421
1422         if (map_length < length + size)
1423                 return 1;
1424         return ret;
1425 }
1426
1427 /*
1428  * in order to insert checksums into the metadata in large chunks,
1429  * we wait until bio submission time.   All the pages in the bio are
1430  * checksummed and sums are attached onto the ordered extent record.
1431  *
1432  * At IO completion time the cums attached on the ordered extent record
1433  * are inserted into the btree
1434  */
1435 static int __btrfs_submit_bio_start(struct inode *inode, int rw,
1436                                     struct bio *bio, int mirror_num,
1437                                     unsigned long bio_flags,
1438                                     u64 bio_offset)
1439 {
1440         struct btrfs_root *root = BTRFS_I(inode)->root;
1441         int ret = 0;
1442
1443         ret = btrfs_csum_one_bio(root, inode, bio, 0, 0);
1444         BUG_ON(ret);
1445         return 0;
1446 }
1447
1448 /*
1449  * in order to insert checksums into the metadata in large chunks,
1450  * we wait until bio submission time.   All the pages in the bio are
1451  * checksummed and sums are attached onto the ordered extent record.
1452  *
1453  * At IO completion time the cums attached on the ordered extent record
1454  * are inserted into the btree
1455  */
1456 static int __btrfs_submit_bio_done(struct inode *inode, int rw, struct bio *bio,
1457                           int mirror_num, unsigned long bio_flags,
1458                           u64 bio_offset)
1459 {
1460         struct btrfs_root *root = BTRFS_I(inode)->root;
1461         return btrfs_map_bio(root, rw, bio, mirror_num, 1);
1462 }
1463
1464 /*
1465  * extent_io.c submission hook. This does the right thing for csum calculation
1466  * on write, or reading the csums from the tree before a read
1467  */
1468 static int btrfs_submit_bio_hook(struct inode *inode, int rw, struct bio *bio,
1469                           int mirror_num, unsigned long bio_flags,
1470                           u64 bio_offset)
1471 {
1472         struct btrfs_root *root = BTRFS_I(inode)->root;
1473         int ret = 0;
1474         int skip_sum;
1475
1476         skip_sum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM;
1477
1478         if (btrfs_is_free_space_inode(root, inode))
1479                 ret = btrfs_bio_wq_end_io(root->fs_info, bio, 2);
1480         else
1481                 ret = btrfs_bio_wq_end_io(root->fs_info, bio, 0);
1482         BUG_ON(ret);
1483
1484         if (!(rw & REQ_WRITE)) {
1485                 if (bio_flags & EXTENT_BIO_COMPRESSED) {
1486                         return btrfs_submit_compressed_read(inode, bio,
1487                                                     mirror_num, bio_flags);
1488                 } else if (!skip_sum) {
1489                         ret = btrfs_lookup_bio_sums(root, inode, bio, NULL);
1490                         if (ret)
1491                                 return ret;
1492                 }
1493                 goto mapit;
1494         } else if (!skip_sum) {
1495                 /* csum items have already been cloned */
1496                 if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
1497                         goto mapit;
1498                 /* we're doing a write, do the async checksumming */
1499                 return btrfs_wq_submit_bio(BTRFS_I(inode)->root->fs_info,
1500                                    inode, rw, bio, mirror_num,
1501                                    bio_flags, bio_offset,
1502                                    __btrfs_submit_bio_start,
1503                                    __btrfs_submit_bio_done);
1504         }
1505
1506 mapit:
1507         return btrfs_map_bio(root, rw, bio, mirror_num, 0);
1508 }
1509
1510 /*
1511  * given a list of ordered sums record them in the inode.  This happens
1512  * at IO completion time based on sums calculated at bio submission time.
1513  */
1514 static noinline int add_pending_csums(struct btrfs_trans_handle *trans,
1515                              struct inode *inode, u64 file_offset,
1516                              struct list_head *list)
1517 {
1518         struct btrfs_ordered_sum *sum;
1519
1520         list_for_each_entry(sum, list, list) {
1521                 btrfs_csum_file_blocks(trans,
1522                        BTRFS_I(inode)->root->fs_info->csum_root, sum);
1523         }
1524         return 0;
1525 }
1526
1527 int btrfs_set_extent_delalloc(struct inode *inode, u64 start, u64 end,
1528                               struct extent_state **cached_state)
1529 {
1530         if ((end & (PAGE_CACHE_SIZE - 1)) == 0)
1531                 WARN_ON(1);
1532         return set_extent_delalloc(&BTRFS_I(inode)->io_tree, start, end,
1533                                    cached_state, GFP_NOFS);
1534 }
1535
1536 /* see btrfs_writepage_start_hook for details on why this is required */
1537 struct btrfs_writepage_fixup {
1538         struct page *page;
1539         struct btrfs_work work;
1540 };
1541
1542 static void btrfs_writepage_fixup_worker(struct btrfs_work *work)
1543 {
1544         struct btrfs_writepage_fixup *fixup;
1545         struct btrfs_ordered_extent *ordered;
1546         struct extent_state *cached_state = NULL;
1547         struct page *page;
1548         struct inode *inode;
1549         u64 page_start;
1550         u64 page_end;
1551
1552         fixup = container_of(work, struct btrfs_writepage_fixup, work);
1553         page = fixup->page;
1554 again:
1555         lock_page(page);
1556         if (!page->mapping || !PageDirty(page) || !PageChecked(page)) {
1557                 ClearPageChecked(page);
1558                 goto out_page;
1559         }
1560
1561         inode = page->mapping->host;
1562         page_start = page_offset(page);
1563         page_end = page_offset(page) + PAGE_CACHE_SIZE - 1;
1564
1565         lock_extent_bits(&BTRFS_I(inode)->io_tree, page_start, page_end, 0,
1566                          &cached_state, GFP_NOFS);
1567
1568         /* already ordered? We're done */
1569         if (PagePrivate2(page))
1570                 goto out;
1571
1572         ordered = btrfs_lookup_ordered_extent(inode, page_start);
1573         if (ordered) {
1574                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, page_start,
1575                                      page_end, &cached_state, GFP_NOFS);
1576                 unlock_page(page);
1577                 btrfs_start_ordered_extent(inode, ordered, 1);
1578                 goto again;
1579         }
1580
1581         BUG();
1582         btrfs_set_extent_delalloc(inode, page_start, page_end, &cached_state);
1583         ClearPageChecked(page);
1584 out:
1585         unlock_extent_cached(&BTRFS_I(inode)->io_tree, page_start, page_end,
1586                              &cached_state, GFP_NOFS);
1587 out_page:
1588         unlock_page(page);
1589         page_cache_release(page);
1590         kfree(fixup);
1591 }
1592
1593 /*
1594  * There are a few paths in the higher layers of the kernel that directly
1595  * set the page dirty bit without asking the filesystem if it is a
1596  * good idea.  This causes problems because we want to make sure COW
1597  * properly happens and the data=ordered rules are followed.
1598  *
1599  * In our case any range that doesn't have the ORDERED bit set
1600  * hasn't been properly setup for IO.  We kick off an async process
1601  * to fix it up.  The async helper will wait for ordered extents, set
1602  * the delalloc bit and make it safe to write the page.
1603  */
1604 static int btrfs_writepage_start_hook(struct page *page, u64 start, u64 end)
1605 {
1606         struct inode *inode = page->mapping->host;
1607         struct btrfs_writepage_fixup *fixup;
1608         struct btrfs_root *root = BTRFS_I(inode)->root;
1609
1610         /* this page is properly in the ordered list */
1611         if (TestClearPagePrivate2(page))
1612                 return 0;
1613
1614         if (PageChecked(page))
1615                 return -EAGAIN;
1616
1617         fixup = kzalloc(sizeof(*fixup), GFP_NOFS);
1618         if (!fixup)
1619                 return -EAGAIN;
1620
1621         SetPageChecked(page);
1622         page_cache_get(page);
1623         fixup->work.func = btrfs_writepage_fixup_worker;
1624         fixup->page = page;
1625         btrfs_queue_worker(&root->fs_info->fixup_workers, &fixup->work);
1626         return -EAGAIN;
1627 }
1628
1629 static int insert_reserved_file_extent(struct btrfs_trans_handle *trans,
1630                                        struct inode *inode, u64 file_pos,
1631                                        u64 disk_bytenr, u64 disk_num_bytes,
1632                                        u64 num_bytes, u64 ram_bytes,
1633                                        u8 compression, u8 encryption,
1634                                        u16 other_encoding, int extent_type)
1635 {
1636         struct btrfs_root *root = BTRFS_I(inode)->root;
1637         struct btrfs_file_extent_item *fi;
1638         struct btrfs_path *path;
1639         struct extent_buffer *leaf;
1640         struct btrfs_key ins;
1641         u64 hint;
1642         int ret;
1643
1644         path = btrfs_alloc_path();
1645         if (!path)
1646                 return -ENOMEM;
1647
1648         path->leave_spinning = 1;
1649
1650         /*
1651          * we may be replacing one extent in the tree with another.
1652          * The new extent is pinned in the extent map, and we don't want
1653          * to drop it from the cache until it is completely in the btree.
1654          *
1655          * So, tell btrfs_drop_extents to leave this extent in the cache.
1656          * the caller is expected to unpin it and allow it to be merged
1657          * with the others.
1658          */
1659         ret = btrfs_drop_extents(trans, inode, file_pos, file_pos + num_bytes,
1660                                  &hint, 0);
1661         BUG_ON(ret);
1662
1663         ins.objectid = btrfs_ino(inode);
1664         ins.offset = file_pos;
1665         ins.type = BTRFS_EXTENT_DATA_KEY;
1666         ret = btrfs_insert_empty_item(trans, root, path, &ins, sizeof(*fi));
1667         BUG_ON(ret);
1668         leaf = path->nodes[0];
1669         fi = btrfs_item_ptr(leaf, path->slots[0],
1670                             struct btrfs_file_extent_item);
1671         btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1672         btrfs_set_file_extent_type(leaf, fi, extent_type);
1673         btrfs_set_file_extent_disk_bytenr(leaf, fi, disk_bytenr);
1674         btrfs_set_file_extent_disk_num_bytes(leaf, fi, disk_num_bytes);
1675         btrfs_set_file_extent_offset(leaf, fi, 0);
1676         btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
1677         btrfs_set_file_extent_ram_bytes(leaf, fi, ram_bytes);
1678         btrfs_set_file_extent_compression(leaf, fi, compression);
1679         btrfs_set_file_extent_encryption(leaf, fi, encryption);
1680         btrfs_set_file_extent_other_encoding(leaf, fi, other_encoding);
1681
1682         btrfs_unlock_up_safe(path, 1);
1683         btrfs_set_lock_blocking(leaf);
1684
1685         btrfs_mark_buffer_dirty(leaf);
1686
1687         inode_add_bytes(inode, num_bytes);
1688
1689         ins.objectid = disk_bytenr;
1690         ins.offset = disk_num_bytes;
1691         ins.type = BTRFS_EXTENT_ITEM_KEY;
1692         ret = btrfs_alloc_reserved_file_extent(trans, root,
1693                                         root->root_key.objectid,
1694                                         btrfs_ino(inode), file_pos, &ins);
1695         BUG_ON(ret);
1696         btrfs_free_path(path);
1697
1698         return 0;
1699 }
1700
1701 /*
1702  * helper function for btrfs_finish_ordered_io, this
1703  * just reads in some of the csum leaves to prime them into ram
1704  * before we start the transaction.  It limits the amount of btree
1705  * reads required while inside the transaction.
1706  */
1707 /* as ordered data IO finishes, this gets called so we can finish
1708  * an ordered extent if the range of bytes in the file it covers are
1709  * fully written.
1710  */
1711 static int btrfs_finish_ordered_io(struct inode *inode, u64 start, u64 end)
1712 {
1713         struct btrfs_root *root = BTRFS_I(inode)->root;
1714         struct btrfs_trans_handle *trans = NULL;
1715         struct btrfs_ordered_extent *ordered_extent = NULL;
1716         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
1717         struct extent_state *cached_state = NULL;
1718         int compress_type = 0;
1719         int ret;
1720         bool nolock;
1721
1722         ret = btrfs_dec_test_ordered_pending(inode, &ordered_extent, start,
1723                                              end - start + 1);
1724         if (!ret)
1725                 return 0;
1726         BUG_ON(!ordered_extent);
1727
1728         nolock = btrfs_is_free_space_inode(root, inode);
1729
1730         if (test_bit(BTRFS_ORDERED_NOCOW, &ordered_extent->flags)) {
1731                 BUG_ON(!list_empty(&ordered_extent->list));
1732                 ret = btrfs_ordered_update_i_size(inode, 0, ordered_extent);
1733                 if (!ret) {
1734                         if (nolock)
1735                                 trans = btrfs_join_transaction_nolock(root);
1736                         else
1737                                 trans = btrfs_join_transaction(root);
1738                         BUG_ON(IS_ERR(trans));
1739                         trans->block_rsv = &root->fs_info->delalloc_block_rsv;
1740                         ret = btrfs_update_inode(trans, root, inode);
1741                         BUG_ON(ret);
1742                 }
1743                 goto out;
1744         }
1745
1746         lock_extent_bits(io_tree, ordered_extent->file_offset,
1747                          ordered_extent->file_offset + ordered_extent->len - 1,
1748                          0, &cached_state, GFP_NOFS);
1749
1750         if (nolock)
1751                 trans = btrfs_join_transaction_nolock(root);
1752         else
1753                 trans = btrfs_join_transaction(root);
1754         BUG_ON(IS_ERR(trans));
1755         trans->block_rsv = &root->fs_info->delalloc_block_rsv;
1756
1757         if (test_bit(BTRFS_ORDERED_COMPRESSED, &ordered_extent->flags))
1758                 compress_type = ordered_extent->compress_type;
1759         if (test_bit(BTRFS_ORDERED_PREALLOC, &ordered_extent->flags)) {
1760                 BUG_ON(compress_type);
1761                 ret = btrfs_mark_extent_written(trans, inode,
1762                                                 ordered_extent->file_offset,
1763                                                 ordered_extent->file_offset +
1764                                                 ordered_extent->len);
1765                 BUG_ON(ret);
1766         } else {
1767                 BUG_ON(root == root->fs_info->tree_root);
1768                 ret = insert_reserved_file_extent(trans, inode,
1769                                                 ordered_extent->file_offset,
1770                                                 ordered_extent->start,
1771                                                 ordered_extent->disk_len,
1772                                                 ordered_extent->len,
1773                                                 ordered_extent->len,
1774                                                 compress_type, 0, 0,
1775                                                 BTRFS_FILE_EXTENT_REG);
1776                 unpin_extent_cache(&BTRFS_I(inode)->extent_tree,
1777                                    ordered_extent->file_offset,
1778                                    ordered_extent->len);
1779                 BUG_ON(ret);
1780         }
1781         unlock_extent_cached(io_tree, ordered_extent->file_offset,
1782                              ordered_extent->file_offset +
1783                              ordered_extent->len - 1, &cached_state, GFP_NOFS);
1784
1785         add_pending_csums(trans, inode, ordered_extent->file_offset,
1786                           &ordered_extent->list);
1787
1788         ret = btrfs_ordered_update_i_size(inode, 0, ordered_extent);
1789         if (!ret || !test_bit(BTRFS_ORDERED_PREALLOC, &ordered_extent->flags)) {
1790                 ret = btrfs_update_inode(trans, root, inode);
1791                 BUG_ON(ret);
1792         }
1793         ret = 0;
1794 out:
1795         if (nolock) {
1796                 if (trans)
1797                         btrfs_end_transaction_nolock(trans, root);
1798         } else {
1799                 btrfs_delalloc_release_metadata(inode, ordered_extent->len);
1800                 if (trans)
1801                         btrfs_end_transaction(trans, root);
1802         }
1803
1804         /* once for us */
1805         btrfs_put_ordered_extent(ordered_extent);
1806         /* once for the tree */
1807         btrfs_put_ordered_extent(ordered_extent);
1808
1809         return 0;
1810 }
1811
1812 static int btrfs_writepage_end_io_hook(struct page *page, u64 start, u64 end,
1813                                 struct extent_state *state, int uptodate)
1814 {
1815         trace_btrfs_writepage_end_io_hook(page, start, end, uptodate);
1816
1817         ClearPagePrivate2(page);
1818         return btrfs_finish_ordered_io(page->mapping->host, start, end);
1819 }
1820
1821 /*
1822  * When IO fails, either with EIO or csum verification fails, we
1823  * try other mirrors that might have a good copy of the data.  This
1824  * io_failure_record is used to record state as we go through all the
1825  * mirrors.  If another mirror has good data, the page is set up to date
1826  * and things continue.  If a good mirror can't be found, the original
1827  * bio end_io callback is called to indicate things have failed.
1828  */
1829 struct io_failure_record {
1830         struct page *page;
1831         u64 start;
1832         u64 len;
1833         u64 logical;
1834         unsigned long bio_flags;
1835         int last_mirror;
1836 };
1837
1838 static int btrfs_io_failed_hook(struct bio *failed_bio,
1839                          struct page *page, u64 start, u64 end,
1840                          struct extent_state *state)
1841 {
1842         struct io_failure_record *failrec = NULL;
1843         u64 private;
1844         struct extent_map *em;
1845         struct inode *inode = page->mapping->host;
1846         struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
1847         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
1848         struct bio *bio;
1849         int num_copies;
1850         int ret;
1851         int rw;
1852         u64 logical;
1853
1854         ret = get_state_private(failure_tree, start, &private);
1855         if (ret) {
1856                 failrec = kmalloc(sizeof(*failrec), GFP_NOFS);
1857                 if (!failrec)
1858                         return -ENOMEM;
1859                 failrec->start = start;
1860                 failrec->len = end - start + 1;
1861                 failrec->last_mirror = 0;
1862                 failrec->bio_flags = 0;
1863
1864                 read_lock(&em_tree->lock);
1865                 em = lookup_extent_mapping(em_tree, start, failrec->len);
1866                 if (em->start > start || em->start + em->len < start) {
1867                         free_extent_map(em);
1868                         em = NULL;
1869                 }
1870                 read_unlock(&em_tree->lock);
1871
1872                 if (IS_ERR_OR_NULL(em)) {
1873                         kfree(failrec);
1874                         return -EIO;
1875                 }
1876                 logical = start - em->start;
1877                 logical = em->block_start + logical;
1878                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
1879                         logical = em->block_start;
1880                         failrec->bio_flags = EXTENT_BIO_COMPRESSED;
1881                         extent_set_compress_type(&failrec->bio_flags,
1882                                                  em->compress_type);
1883                 }
1884                 failrec->logical = logical;
1885                 free_extent_map(em);
1886                 set_extent_bits(failure_tree, start, end, EXTENT_LOCKED |
1887                                 EXTENT_DIRTY, GFP_NOFS);
1888                 set_state_private(failure_tree, start,
1889                                  (u64)(unsigned long)failrec);
1890         } else {
1891                 failrec = (struct io_failure_record *)(unsigned long)private;
1892         }
1893         num_copies = btrfs_num_copies(
1894                               &BTRFS_I(inode)->root->fs_info->mapping_tree,
1895                               failrec->logical, failrec->len);
1896         failrec->last_mirror++;
1897         if (!state) {
1898                 spin_lock(&BTRFS_I(inode)->io_tree.lock);
1899                 state = find_first_extent_bit_state(&BTRFS_I(inode)->io_tree,
1900                                                     failrec->start,
1901                                                     EXTENT_LOCKED);
1902                 if (state && state->start != failrec->start)
1903                         state = NULL;
1904                 spin_unlock(&BTRFS_I(inode)->io_tree.lock);
1905         }
1906         if (!state || failrec->last_mirror > num_copies) {
1907                 set_state_private(failure_tree, failrec->start, 0);
1908                 clear_extent_bits(failure_tree, failrec->start,
1909                                   failrec->start + failrec->len - 1,
1910                                   EXTENT_LOCKED | EXTENT_DIRTY, GFP_NOFS);
1911                 kfree(failrec);
1912                 return -EIO;
1913         }
1914         bio = bio_alloc(GFP_NOFS, 1);
1915         bio->bi_private = state;
1916         bio->bi_end_io = failed_bio->bi_end_io;
1917         bio->bi_sector = failrec->logical >> 9;
1918         bio->bi_bdev = failed_bio->bi_bdev;
1919         bio->bi_size = 0;
1920
1921         bio_add_page(bio, page, failrec->len, start - page_offset(page));
1922         if (failed_bio->bi_rw & REQ_WRITE)
1923                 rw = WRITE;
1924         else
1925                 rw = READ;
1926
1927         ret = BTRFS_I(inode)->io_tree.ops->submit_bio_hook(inode, rw, bio,
1928                                                       failrec->last_mirror,
1929                                                       failrec->bio_flags, 0);
1930         return ret;
1931 }
1932
1933 /*
1934  * each time an IO finishes, we do a fast check in the IO failure tree
1935  * to see if we need to process or clean up an io_failure_record
1936  */
1937 static int btrfs_clean_io_failures(struct inode *inode, u64 start)
1938 {
1939         u64 private;
1940         u64 private_failure;
1941         struct io_failure_record *failure;
1942         int ret;
1943
1944         private = 0;
1945         if (count_range_bits(&BTRFS_I(inode)->io_failure_tree, &private,
1946                              (u64)-1, 1, EXTENT_DIRTY, 0)) {
1947                 ret = get_state_private(&BTRFS_I(inode)->io_failure_tree,
1948                                         start, &private_failure);
1949                 if (ret == 0) {
1950                         failure = (struct io_failure_record *)(unsigned long)
1951                                    private_failure;
1952                         set_state_private(&BTRFS_I(inode)->io_failure_tree,
1953                                           failure->start, 0);
1954                         clear_extent_bits(&BTRFS_I(inode)->io_failure_tree,
1955                                           failure->start,
1956                                           failure->start + failure->len - 1,
1957                                           EXTENT_DIRTY | EXTENT_LOCKED,
1958                                           GFP_NOFS);
1959                         kfree(failure);
1960                 }
1961         }
1962         return 0;
1963 }
1964
1965 /*
1966  * when reads are done, we need to check csums to verify the data is correct
1967  * if there's a match, we allow the bio to finish.  If not, we go through
1968  * the io_failure_record routines to find good copies
1969  */
1970 static int btrfs_readpage_end_io_hook(struct page *page, u64 start, u64 end,
1971                                struct extent_state *state)
1972 {
1973         size_t offset = start - ((u64)page->index << PAGE_CACHE_SHIFT);
1974         struct inode *inode = page->mapping->host;
1975         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
1976         char *kaddr;
1977         u64 private = ~(u32)0;
1978         int ret;
1979         struct btrfs_root *root = BTRFS_I(inode)->root;
1980         u32 csum = ~(u32)0;
1981
1982         if (PageChecked(page)) {
1983                 ClearPageChecked(page);
1984                 goto good;
1985         }
1986
1987         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)
1988                 goto good;
1989
1990         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID &&
1991             test_range_bit(io_tree, start, end, EXTENT_NODATASUM, 1, NULL)) {
1992                 clear_extent_bits(io_tree, start, end, EXTENT_NODATASUM,
1993                                   GFP_NOFS);
1994                 return 0;
1995         }
1996
1997         if (state && state->start == start) {
1998                 private = state->private;
1999                 ret = 0;
2000         } else {
2001                 ret = get_state_private(io_tree, start, &private);
2002         }
2003         kaddr = kmap_atomic(page, KM_USER0);
2004         if (ret)
2005                 goto zeroit;
2006
2007         csum = btrfs_csum_data(root, kaddr + offset, csum,  end - start + 1);
2008         btrfs_csum_final(csum, (char *)&csum);
2009         if (csum != private)
2010                 goto zeroit;
2011
2012         kunmap_atomic(kaddr, KM_USER0);
2013 good:
2014         /* if the io failure tree for this inode is non-empty,
2015          * check to see if we've recovered from a failed IO
2016          */
2017         btrfs_clean_io_failures(inode, start);
2018         return 0;
2019
2020 zeroit:
2021         printk_ratelimited(KERN_INFO "btrfs csum failed ino %llu off %llu csum %u "
2022                        "private %llu\n",
2023                        (unsigned long long)btrfs_ino(page->mapping->host),
2024                        (unsigned long long)start, csum,
2025                        (unsigned long long)private);
2026         memset(kaddr + offset, 1, end - start + 1);
2027         flush_dcache_page(page);
2028         kunmap_atomic(kaddr, KM_USER0);
2029         if (private == 0)
2030                 return 0;
2031         return -EIO;
2032 }
2033
2034 struct delayed_iput {
2035         struct list_head list;
2036         struct inode *inode;
2037 };
2038
2039 void btrfs_add_delayed_iput(struct inode *inode)
2040 {
2041         struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
2042         struct delayed_iput *delayed;
2043
2044         if (atomic_add_unless(&inode->i_count, -1, 1))
2045                 return;
2046
2047         delayed = kmalloc(sizeof(*delayed), GFP_NOFS | __GFP_NOFAIL);
2048         delayed->inode = inode;
2049
2050         spin_lock(&fs_info->delayed_iput_lock);
2051         list_add_tail(&delayed->list, &fs_info->delayed_iputs);
2052         spin_unlock(&fs_info->delayed_iput_lock);
2053 }
2054
2055 void btrfs_run_delayed_iputs(struct btrfs_root *root)
2056 {
2057         LIST_HEAD(list);
2058         struct btrfs_fs_info *fs_info = root->fs_info;
2059         struct delayed_iput *delayed;
2060         int empty;
2061
2062         spin_lock(&fs_info->delayed_iput_lock);
2063         empty = list_empty(&fs_info->delayed_iputs);
2064         spin_unlock(&fs_info->delayed_iput_lock);
2065         if (empty)
2066                 return;
2067
2068         down_read(&root->fs_info->cleanup_work_sem);
2069         spin_lock(&fs_info->delayed_iput_lock);
2070         list_splice_init(&fs_info->delayed_iputs, &list);
2071         spin_unlock(&fs_info->delayed_iput_lock);
2072
2073         while (!list_empty(&list)) {
2074                 delayed = list_entry(list.next, struct delayed_iput, list);
2075                 list_del(&delayed->list);
2076                 iput(delayed->inode);
2077                 kfree(delayed);
2078         }
2079         up_read(&root->fs_info->cleanup_work_sem);
2080 }
2081
2082 /*
2083  * calculate extra metadata reservation when snapshotting a subvolume
2084  * contains orphan files.
2085  */
2086 void btrfs_orphan_pre_snapshot(struct btrfs_trans_handle *trans,
2087                                 struct btrfs_pending_snapshot *pending,
2088                                 u64 *bytes_to_reserve)
2089 {
2090         struct btrfs_root *root;
2091         struct btrfs_block_rsv *block_rsv;
2092         u64 num_bytes;
2093         int index;
2094
2095         root = pending->root;
2096         if (!root->orphan_block_rsv || list_empty(&root->orphan_list))
2097                 return;
2098
2099         block_rsv = root->orphan_block_rsv;
2100
2101         /* orphan block reservation for the snapshot */
2102         num_bytes = block_rsv->size;
2103
2104         /*
2105          * after the snapshot is created, COWing tree blocks may use more
2106          * space than it frees. So we should make sure there is enough
2107          * reserved space.
2108          */
2109         index = trans->transid & 0x1;
2110         if (block_rsv->reserved + block_rsv->freed[index] < block_rsv->size) {
2111                 num_bytes += block_rsv->size -
2112                              (block_rsv->reserved + block_rsv->freed[index]);
2113         }
2114
2115         *bytes_to_reserve += num_bytes;
2116 }
2117
2118 void btrfs_orphan_post_snapshot(struct btrfs_trans_handle *trans,
2119                                 struct btrfs_pending_snapshot *pending)
2120 {
2121         struct btrfs_root *root = pending->root;
2122         struct btrfs_root *snap = pending->snap;
2123         struct btrfs_block_rsv *block_rsv;
2124         u64 num_bytes;
2125         int index;
2126         int ret;
2127
2128         if (!root->orphan_block_rsv || list_empty(&root->orphan_list))
2129                 return;
2130
2131         /* refill source subvolume's orphan block reservation */
2132         block_rsv = root->orphan_block_rsv;
2133         index = trans->transid & 0x1;
2134         if (block_rsv->reserved + block_rsv->freed[index] < block_rsv->size) {
2135                 num_bytes = block_rsv->size -
2136                             (block_rsv->reserved + block_rsv->freed[index]);
2137                 ret = btrfs_block_rsv_migrate(&pending->block_rsv,
2138                                               root->orphan_block_rsv,
2139                                               num_bytes);
2140                 BUG_ON(ret);
2141         }
2142
2143         /* setup orphan block reservation for the snapshot */
2144         block_rsv = btrfs_alloc_block_rsv(snap);
2145         BUG_ON(!block_rsv);
2146
2147         btrfs_add_durable_block_rsv(root->fs_info, block_rsv);
2148         snap->orphan_block_rsv = block_rsv;
2149
2150         num_bytes = root->orphan_block_rsv->size;
2151         ret = btrfs_block_rsv_migrate(&pending->block_rsv,
2152                                       block_rsv, num_bytes);
2153         BUG_ON(ret);
2154
2155 #if 0
2156         /* insert orphan item for the snapshot */
2157         WARN_ON(!root->orphan_item_inserted);
2158         ret = btrfs_insert_orphan_item(trans, root->fs_info->tree_root,
2159                                        snap->root_key.objectid);
2160         BUG_ON(ret);
2161         snap->orphan_item_inserted = 1;
2162 #endif
2163 }
2164
2165 enum btrfs_orphan_cleanup_state {
2166         ORPHAN_CLEANUP_STARTED  = 1,
2167         ORPHAN_CLEANUP_DONE     = 2,
2168 };
2169
2170 /*
2171  * This is called in transaction commmit time. If there are no orphan
2172  * files in the subvolume, it removes orphan item and frees block_rsv
2173  * structure.
2174  */
2175 void btrfs_orphan_commit_root(struct btrfs_trans_handle *trans,
2176                               struct btrfs_root *root)
2177 {
2178         int ret;
2179
2180         if (!list_empty(&root->orphan_list) ||
2181             root->orphan_cleanup_state != ORPHAN_CLEANUP_DONE)
2182                 return;
2183
2184         if (root->orphan_item_inserted &&
2185             btrfs_root_refs(&root->root_item) > 0) {
2186                 ret = btrfs_del_orphan_item(trans, root->fs_info->tree_root,
2187                                             root->root_key.objectid);
2188                 BUG_ON(ret);
2189                 root->orphan_item_inserted = 0;
2190         }
2191
2192         if (root->orphan_block_rsv) {
2193                 WARN_ON(root->orphan_block_rsv->size > 0);
2194                 btrfs_free_block_rsv(root, root->orphan_block_rsv);
2195                 root->orphan_block_rsv = NULL;
2196         }
2197 }
2198
2199 /*
2200  * This creates an orphan entry for the given inode in case something goes
2201  * wrong in the middle of an unlink/truncate.
2202  *
2203  * NOTE: caller of this function should reserve 5 units of metadata for
2204  *       this function.
2205  */
2206 int btrfs_orphan_add(struct btrfs_trans_handle *trans, struct inode *inode)
2207 {
2208         struct btrfs_root *root = BTRFS_I(inode)->root;
2209         struct btrfs_block_rsv *block_rsv = NULL;
2210         int reserve = 0;
2211         int insert = 0;
2212         int ret;
2213
2214         if (!root->orphan_block_rsv) {
2215                 block_rsv = btrfs_alloc_block_rsv(root);
2216                 if (!block_rsv)
2217                         return -ENOMEM;
2218         }
2219
2220         spin_lock(&root->orphan_lock);
2221         if (!root->orphan_block_rsv) {
2222                 root->orphan_block_rsv = block_rsv;
2223         } else if (block_rsv) {
2224                 btrfs_free_block_rsv(root, block_rsv);
2225                 block_rsv = NULL;
2226         }
2227
2228         if (list_empty(&BTRFS_I(inode)->i_orphan)) {
2229                 list_add(&BTRFS_I(inode)->i_orphan, &root->orphan_list);
2230 #if 0
2231                 /*
2232                  * For proper ENOSPC handling, we should do orphan
2233                  * cleanup when mounting. But this introduces backward
2234                  * compatibility issue.
2235                  */
2236                 if (!xchg(&root->orphan_item_inserted, 1))
2237                         insert = 2;
2238                 else
2239                         insert = 1;
2240 #endif
2241                 insert = 1;
2242         }
2243
2244         if (!BTRFS_I(inode)->orphan_meta_reserved) {
2245                 BTRFS_I(inode)->orphan_meta_reserved = 1;
2246                 reserve = 1;
2247         }
2248         spin_unlock(&root->orphan_lock);
2249
2250         if (block_rsv)
2251                 btrfs_add_durable_block_rsv(root->fs_info, block_rsv);
2252
2253         /* grab metadata reservation from transaction handle */
2254         if (reserve) {
2255                 ret = btrfs_orphan_reserve_metadata(trans, inode);
2256                 BUG_ON(ret);
2257         }
2258
2259         /* insert an orphan item to track this unlinked/truncated file */
2260         if (insert >= 1) {
2261                 ret = btrfs_insert_orphan_item(trans, root, btrfs_ino(inode));
2262                 BUG_ON(ret);
2263         }
2264
2265         /* insert an orphan item to track subvolume contains orphan files */
2266         if (insert >= 2) {
2267                 ret = btrfs_insert_orphan_item(trans, root->fs_info->tree_root,
2268                                                root->root_key.objectid);
2269                 BUG_ON(ret);
2270         }
2271         return 0;
2272 }
2273
2274 /*
2275  * We have done the truncate/delete so we can go ahead and remove the orphan
2276  * item for this particular inode.
2277  */
2278 int btrfs_orphan_del(struct btrfs_trans_handle *trans, struct inode *inode)
2279 {
2280         struct btrfs_root *root = BTRFS_I(inode)->root;
2281         int delete_item = 0;
2282         int release_rsv = 0;
2283         int ret = 0;
2284
2285         spin_lock(&root->orphan_lock);
2286         if (!list_empty(&BTRFS_I(inode)->i_orphan)) {
2287                 list_del_init(&BTRFS_I(inode)->i_orphan);
2288                 delete_item = 1;
2289         }
2290
2291         if (BTRFS_I(inode)->orphan_meta_reserved) {
2292                 BTRFS_I(inode)->orphan_meta_reserved = 0;
2293                 release_rsv = 1;
2294         }
2295         spin_unlock(&root->orphan_lock);
2296
2297         if (trans && delete_item) {
2298                 ret = btrfs_del_orphan_item(trans, root, btrfs_ino(inode));
2299                 BUG_ON(ret);
2300         }
2301
2302         if (release_rsv)
2303                 btrfs_orphan_release_metadata(inode);
2304
2305         return 0;
2306 }
2307
2308 /*
2309  * this cleans up any orphans that may be left on the list from the last use
2310  * of this root.
2311  */
2312 int btrfs_orphan_cleanup(struct btrfs_root *root)
2313 {
2314         struct btrfs_path *path;
2315         struct extent_buffer *leaf;
2316         struct btrfs_key key, found_key;
2317         struct btrfs_trans_handle *trans;
2318         struct inode *inode;
2319         int ret = 0, nr_unlink = 0, nr_truncate = 0;
2320
2321         if (cmpxchg(&root->orphan_cleanup_state, 0, ORPHAN_CLEANUP_STARTED))
2322                 return 0;
2323
2324         path = btrfs_alloc_path();
2325         if (!path) {
2326                 ret = -ENOMEM;
2327                 goto out;
2328         }
2329         path->reada = -1;
2330
2331         key.objectid = BTRFS_ORPHAN_OBJECTID;
2332         btrfs_set_key_type(&key, BTRFS_ORPHAN_ITEM_KEY);
2333         key.offset = (u64)-1;
2334
2335         while (1) {
2336                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2337                 if (ret < 0)
2338                         goto out;
2339
2340                 /*
2341                  * if ret == 0 means we found what we were searching for, which
2342                  * is weird, but possible, so only screw with path if we didn't
2343                  * find the key and see if we have stuff that matches
2344                  */
2345                 if (ret > 0) {
2346                         ret = 0;
2347                         if (path->slots[0] == 0)
2348                                 break;
2349                         path->slots[0]--;
2350                 }
2351
2352                 /* pull out the item */
2353                 leaf = path->nodes[0];
2354                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
2355
2356                 /* make sure the item matches what we want */
2357                 if (found_key.objectid != BTRFS_ORPHAN_OBJECTID)
2358                         break;
2359                 if (btrfs_key_type(&found_key) != BTRFS_ORPHAN_ITEM_KEY)
2360                         break;
2361
2362                 /* release the path since we're done with it */
2363                 btrfs_release_path(path);
2364
2365                 /*
2366                  * this is where we are basically btrfs_lookup, without the
2367                  * crossing root thing.  we store the inode number in the
2368                  * offset of the orphan item.
2369                  */
2370                 found_key.objectid = found_key.offset;
2371                 found_key.type = BTRFS_INODE_ITEM_KEY;
2372                 found_key.offset = 0;
2373                 inode = btrfs_iget(root->fs_info->sb, &found_key, root, NULL);
2374                 if (IS_ERR(inode)) {
2375                         ret = PTR_ERR(inode);
2376                         goto out;
2377                 }
2378
2379                 /*
2380                  * add this inode to the orphan list so btrfs_orphan_del does
2381                  * the proper thing when we hit it
2382                  */
2383                 spin_lock(&root->orphan_lock);
2384                 list_add(&BTRFS_I(inode)->i_orphan, &root->orphan_list);
2385                 spin_unlock(&root->orphan_lock);
2386
2387                 /*
2388                  * if this is a bad inode, means we actually succeeded in
2389                  * removing the inode, but not the orphan record, which means
2390                  * we need to manually delete the orphan since iput will just
2391                  * do a destroy_inode
2392                  */
2393                 if (is_bad_inode(inode)) {
2394                         trans = btrfs_start_transaction(root, 0);
2395                         if (IS_ERR(trans)) {
2396                                 ret = PTR_ERR(trans);
2397                                 goto out;
2398                         }
2399                         btrfs_orphan_del(trans, inode);
2400                         btrfs_end_transaction(trans, root);
2401                         iput(inode);
2402                         continue;
2403                 }
2404
2405                 /* if we have links, this was a truncate, lets do that */
2406                 if (inode->i_nlink) {
2407                         if (!S_ISREG(inode->i_mode)) {
2408                                 WARN_ON(1);
2409                                 iput(inode);
2410                                 continue;
2411                         }
2412                         nr_truncate++;
2413                         ret = btrfs_truncate(inode);
2414                 } else {
2415                         nr_unlink++;
2416                 }
2417
2418                 /* this will do delete_inode and everything for us */
2419                 iput(inode);
2420                 if (ret)
2421                         goto out;
2422         }
2423         root->orphan_cleanup_state = ORPHAN_CLEANUP_DONE;
2424
2425         if (root->orphan_block_rsv)
2426                 btrfs_block_rsv_release(root, root->orphan_block_rsv,
2427                                         (u64)-1);
2428
2429         if (root->orphan_block_rsv || root->orphan_item_inserted) {
2430                 trans = btrfs_join_transaction(root);
2431                 if (!IS_ERR(trans))
2432                         btrfs_end_transaction(trans, root);
2433         }
2434
2435         if (nr_unlink)
2436                 printk(KERN_INFO "btrfs: unlinked %d orphans\n", nr_unlink);
2437         if (nr_truncate)
2438                 printk(KERN_INFO "btrfs: truncated %d orphans\n", nr_truncate);
2439
2440 out:
2441         if (ret)
2442                 printk(KERN_CRIT "btrfs: could not do orphan cleanup %d\n", ret);
2443         btrfs_free_path(path);
2444         return ret;
2445 }
2446
2447 /*
2448  * very simple check to peek ahead in the leaf looking for xattrs.  If we
2449  * don't find any xattrs, we know there can't be any acls.
2450  *
2451  * slot is the slot the inode is in, objectid is the objectid of the inode
2452  */
2453 static noinline int acls_after_inode_item(struct extent_buffer *leaf,
2454                                           int slot, u64 objectid)
2455 {
2456         u32 nritems = btrfs_header_nritems(leaf);
2457         struct btrfs_key found_key;
2458         int scanned = 0;
2459
2460         slot++;
2461         while (slot < nritems) {
2462                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
2463
2464                 /* we found a different objectid, there must not be acls */
2465                 if (found_key.objectid != objectid)
2466                         return 0;
2467
2468                 /* we found an xattr, assume we've got an acl */
2469                 if (found_key.type == BTRFS_XATTR_ITEM_KEY)
2470                         return 1;
2471
2472                 /*
2473                  * we found a key greater than an xattr key, there can't
2474                  * be any acls later on
2475                  */
2476                 if (found_key.type > BTRFS_XATTR_ITEM_KEY)
2477                         return 0;
2478
2479                 slot++;
2480                 scanned++;
2481
2482                 /*
2483                  * it goes inode, inode backrefs, xattrs, extents,
2484                  * so if there are a ton of hard links to an inode there can
2485                  * be a lot of backrefs.  Don't waste time searching too hard,
2486                  * this is just an optimization
2487                  */
2488                 if (scanned >= 8)
2489                         break;
2490         }
2491         /* we hit the end of the leaf before we found an xattr or
2492          * something larger than an xattr.  We have to assume the inode
2493          * has acls
2494          */
2495         return 1;
2496 }
2497
2498 /*
2499  * read an inode from the btree into the in-memory inode
2500  */
2501 static void btrfs_read_locked_inode(struct inode *inode)
2502 {
2503         struct btrfs_path *path;
2504         struct extent_buffer *leaf;
2505         struct btrfs_inode_item *inode_item;
2506         struct btrfs_timespec *tspec;
2507         struct btrfs_root *root = BTRFS_I(inode)->root;
2508         struct btrfs_key location;
2509         int maybe_acls;
2510         u32 rdev;
2511         int ret;
2512         bool filled = false;
2513
2514         ret = btrfs_fill_inode(inode, &rdev);
2515         if (!ret)
2516                 filled = true;
2517
2518         path = btrfs_alloc_path();
2519         if (!path)
2520                 goto make_bad;
2521
2522         path->leave_spinning = 1;
2523         memcpy(&location, &BTRFS_I(inode)->location, sizeof(location));
2524
2525         ret = btrfs_lookup_inode(NULL, root, path, &location, 0);
2526         if (ret)
2527                 goto make_bad;
2528
2529         leaf = path->nodes[0];
2530
2531         if (filled)
2532                 goto cache_acl;
2533
2534         inode_item = btrfs_item_ptr(leaf, path->slots[0],
2535                                     struct btrfs_inode_item);
2536         inode->i_mode = btrfs_inode_mode(leaf, inode_item);
2537         inode->i_nlink = btrfs_inode_nlink(leaf, inode_item);
2538         inode->i_uid = btrfs_inode_uid(leaf, inode_item);
2539         inode->i_gid = btrfs_inode_gid(leaf, inode_item);
2540         btrfs_i_size_write(inode, btrfs_inode_size(leaf, inode_item));
2541
2542         tspec = btrfs_inode_atime(inode_item);
2543         inode->i_atime.tv_sec = btrfs_timespec_sec(leaf, tspec);
2544         inode->i_atime.tv_nsec = btrfs_timespec_nsec(leaf, tspec);
2545
2546         tspec = btrfs_inode_mtime(inode_item);
2547         inode->i_mtime.tv_sec = btrfs_timespec_sec(leaf, tspec);
2548         inode->i_mtime.tv_nsec = btrfs_timespec_nsec(leaf, tspec);
2549
2550         tspec = btrfs_inode_ctime(inode_item);
2551         inode->i_ctime.tv_sec = btrfs_timespec_sec(leaf, tspec);
2552         inode->i_ctime.tv_nsec = btrfs_timespec_nsec(leaf, tspec);
2553
2554         inode_set_bytes(inode, btrfs_inode_nbytes(leaf, inode_item));
2555         BTRFS_I(inode)->generation = btrfs_inode_generation(leaf, inode_item);
2556         BTRFS_I(inode)->sequence = btrfs_inode_sequence(leaf, inode_item);
2557         inode->i_generation = BTRFS_I(inode)->generation;
2558         inode->i_rdev = 0;
2559         rdev = btrfs_inode_rdev(leaf, inode_item);
2560
2561         BTRFS_I(inode)->index_cnt = (u64)-1;
2562         BTRFS_I(inode)->flags = btrfs_inode_flags(leaf, inode_item);
2563 cache_acl:
2564         /*
2565          * try to precache a NULL acl entry for files that don't have
2566          * any xattrs or acls
2567          */
2568         maybe_acls = acls_after_inode_item(leaf, path->slots[0],
2569                                            btrfs_ino(inode));
2570         if (!maybe_acls)
2571                 cache_no_acl(inode);
2572
2573         btrfs_free_path(path);
2574
2575         switch (inode->i_mode & S_IFMT) {
2576         case S_IFREG:
2577                 inode->i_mapping->a_ops = &btrfs_aops;
2578                 inode->i_mapping->backing_dev_info = &root->fs_info->bdi;
2579                 BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
2580                 inode->i_fop = &btrfs_file_operations;
2581                 inode->i_op = &btrfs_file_inode_operations;
2582                 break;
2583         case S_IFDIR:
2584                 inode->i_fop = &btrfs_dir_file_operations;
2585                 if (root == root->fs_info->tree_root)
2586                         inode->i_op = &btrfs_dir_ro_inode_operations;
2587                 else
2588                         inode->i_op = &btrfs_dir_inode_operations;
2589                 break;
2590         case S_IFLNK:
2591                 inode->i_op = &btrfs_symlink_inode_operations;
2592                 inode->i_mapping->a_ops = &btrfs_symlink_aops;
2593                 inode->i_mapping->backing_dev_info = &root->fs_info->bdi;
2594                 break;
2595         default:
2596                 inode->i_op = &btrfs_special_inode_operations;
2597                 init_special_inode(inode, inode->i_mode, rdev);
2598                 break;
2599         }
2600
2601         btrfs_update_iflags(inode);
2602         return;
2603
2604 make_bad:
2605         btrfs_free_path(path);
2606         make_bad_inode(inode);
2607 }
2608
2609 /*
2610  * given a leaf and an inode, copy the inode fields into the leaf
2611  */
2612 static void fill_inode_item(struct btrfs_trans_handle *trans,
2613                             struct extent_buffer *leaf,
2614                             struct btrfs_inode_item *item,
2615                             struct inode *inode)
2616 {
2617         btrfs_set_inode_uid(leaf, item, inode->i_uid);
2618         btrfs_set_inode_gid(leaf, item, inode->i_gid);
2619         btrfs_set_inode_size(leaf, item, BTRFS_I(inode)->disk_i_size);
2620         btrfs_set_inode_mode(leaf, item, inode->i_mode);
2621         btrfs_set_inode_nlink(leaf, item, inode->i_nlink);
2622
2623         btrfs_set_timespec_sec(leaf, btrfs_inode_atime(item),
2624                                inode->i_atime.tv_sec);
2625         btrfs_set_timespec_nsec(leaf, btrfs_inode_atime(item),
2626                                 inode->i_atime.tv_nsec);
2627
2628         btrfs_set_timespec_sec(leaf, btrfs_inode_mtime(item),
2629                                inode->i_mtime.tv_sec);
2630         btrfs_set_timespec_nsec(leaf, btrfs_inode_mtime(item),
2631                                 inode->i_mtime.tv_nsec);
2632
2633         btrfs_set_timespec_sec(leaf, btrfs_inode_ctime(item),
2634                                inode->i_ctime.tv_sec);
2635         btrfs_set_timespec_nsec(leaf, btrfs_inode_ctime(item),
2636                                 inode->i_ctime.tv_nsec);
2637
2638         btrfs_set_inode_nbytes(leaf, item, inode_get_bytes(inode));
2639         btrfs_set_inode_generation(leaf, item, BTRFS_I(inode)->generation);
2640         btrfs_set_inode_sequence(leaf, item, BTRFS_I(inode)->sequence);
2641         btrfs_set_inode_transid(leaf, item, trans->transid);
2642         btrfs_set_inode_rdev(leaf, item, inode->i_rdev);
2643         btrfs_set_inode_flags(leaf, item, BTRFS_I(inode)->flags);
2644         btrfs_set_inode_block_group(leaf, item, 0);
2645 }
2646
2647 /*
2648  * copy everything in the in-memory inode into the btree.
2649  */
2650 noinline int btrfs_update_inode(struct btrfs_trans_handle *trans,
2651                                 struct btrfs_root *root, struct inode *inode)
2652 {
2653         struct btrfs_inode_item *inode_item;
2654         struct btrfs_path *path;
2655         struct extent_buffer *leaf;
2656         int ret;
2657
2658         /*
2659          * If the inode is a free space inode, we can deadlock during commit
2660          * if we put it into the delayed code.
2661          *
2662          * The data relocation inode should also be directly updated
2663          * without delay
2664          */
2665         if (!btrfs_is_free_space_inode(root, inode)
2666             && root->root_key.objectid != BTRFS_DATA_RELOC_TREE_OBJECTID) {
2667                 ret = btrfs_delayed_update_inode(trans, root, inode);
2668                 if (!ret)
2669                         btrfs_set_inode_last_trans(trans, inode);
2670                 return ret;
2671         }
2672
2673         path = btrfs_alloc_path();
2674         if (!path)
2675                 return -ENOMEM;
2676
2677         path->leave_spinning = 1;
2678         ret = btrfs_lookup_inode(trans, root, path, &BTRFS_I(inode)->location,
2679                                  1);
2680         if (ret) {
2681                 if (ret > 0)
2682                         ret = -ENOENT;
2683                 goto failed;
2684         }
2685
2686         btrfs_unlock_up_safe(path, 1);
2687         leaf = path->nodes[0];
2688         inode_item = btrfs_item_ptr(leaf, path->slots[0],
2689                                     struct btrfs_inode_item);
2690
2691         fill_inode_item(trans, leaf, inode_item, inode);
2692         btrfs_mark_buffer_dirty(leaf);
2693         btrfs_set_inode_last_trans(trans, inode);
2694         ret = 0;
2695 failed:
2696         btrfs_free_path(path);
2697         return ret;
2698 }
2699
2700 /*
2701  * unlink helper that gets used here in inode.c and in the tree logging
2702  * recovery code.  It remove a link in a directory with a given name, and
2703  * also drops the back refs in the inode to the directory
2704  */
2705 static int __btrfs_unlink_inode(struct btrfs_trans_handle *trans,
2706                                 struct btrfs_root *root,
2707                                 struct inode *dir, struct inode *inode,
2708                                 const char *name, int name_len)
2709 {
2710         struct btrfs_path *path;
2711         int ret = 0;
2712         struct extent_buffer *leaf;
2713         struct btrfs_dir_item *di;
2714         struct btrfs_key key;
2715         u64 index;
2716         u64 ino = btrfs_ino(inode);
2717         u64 dir_ino = btrfs_ino(dir);
2718
2719         path = btrfs_alloc_path();
2720         if (!path) {
2721                 ret = -ENOMEM;
2722                 goto out;
2723         }
2724
2725         path->leave_spinning = 1;
2726         di = btrfs_lookup_dir_item(trans, root, path, dir_ino,
2727                                     name, name_len, -1);
2728         if (IS_ERR(di)) {
2729                 ret = PTR_ERR(di);
2730                 goto err;
2731         }
2732         if (!di) {
2733                 ret = -ENOENT;
2734                 goto err;
2735         }
2736         leaf = path->nodes[0];
2737         btrfs_dir_item_key_to_cpu(leaf, di, &key);
2738         ret = btrfs_delete_one_dir_name(trans, root, path, di);
2739         if (ret)
2740                 goto err;
2741         btrfs_release_path(path);
2742
2743         ret = btrfs_del_inode_ref(trans, root, name, name_len, ino,
2744                                   dir_ino, &index);
2745         if (ret) {
2746                 printk(KERN_INFO "btrfs failed to delete reference to %.*s, "
2747                        "inode %llu parent %llu\n", name_len, name,
2748                        (unsigned long long)ino, (unsigned long long)dir_ino);
2749                 goto err;
2750         }
2751
2752         ret = btrfs_delete_delayed_dir_index(trans, root, dir, index);
2753         if (ret)
2754                 goto err;
2755
2756         ret = btrfs_del_inode_ref_in_log(trans, root, name, name_len,
2757                                          inode, dir_ino);
2758         BUG_ON(ret != 0 && ret != -ENOENT);
2759
2760         ret = btrfs_del_dir_entries_in_log(trans, root, name, name_len,
2761                                            dir, index);
2762         if (ret == -ENOENT)
2763                 ret = 0;
2764 err:
2765         btrfs_free_path(path);
2766         if (ret)
2767                 goto out;
2768
2769         btrfs_i_size_write(dir, dir->i_size - name_len * 2);
2770         inode->i_ctime = dir->i_mtime = dir->i_ctime = CURRENT_TIME;
2771         btrfs_update_inode(trans, root, dir);
2772 out:
2773         return ret;
2774 }
2775
2776 int btrfs_unlink_inode(struct btrfs_trans_handle *trans,
2777                        struct btrfs_root *root,
2778                        struct inode *dir, struct inode *inode,
2779                        const char *name, int name_len)
2780 {
2781         int ret;
2782         ret = __btrfs_unlink_inode(trans, root, dir, inode, name, name_len);
2783         if (!ret) {
2784                 btrfs_drop_nlink(inode);
2785                 ret = btrfs_update_inode(trans, root, inode);
2786         }
2787         return ret;
2788 }
2789
2790
2791 /* helper to check if there is any shared block in the path */
2792 static int check_path_shared(struct btrfs_root *root,
2793                              struct btrfs_path *path)
2794 {
2795         struct extent_buffer *eb;
2796         int level;
2797         u64 refs = 1;
2798
2799         for (level = 0; level < BTRFS_MAX_LEVEL; level++) {
2800                 int ret;
2801
2802                 if (!path->nodes[level])
2803                         break;
2804                 eb = path->nodes[level];
2805                 if (!btrfs_block_can_be_shared(root, eb))
2806                         continue;
2807                 ret = btrfs_lookup_extent_info(NULL, root, eb->start, eb->len,
2808                                                &refs, NULL);
2809                 if (refs > 1)
2810                         return 1;
2811         }
2812         return 0;
2813 }
2814
2815 /*
2816  * helper to start transaction for unlink and rmdir.
2817  *
2818  * unlink and rmdir are special in btrfs, they do not always free space.
2819  * so in enospc case, we should make sure they will free space before
2820  * allowing them to use the global metadata reservation.
2821  */
2822 static struct btrfs_trans_handle *__unlink_start_trans(struct inode *dir,
2823                                                        struct dentry *dentry)
2824 {
2825         struct btrfs_trans_handle *trans;
2826         struct btrfs_root *root = BTRFS_I(dir)->root;
2827         struct btrfs_path *path;
2828         struct btrfs_inode_ref *ref;
2829         struct btrfs_dir_item *di;
2830         struct inode *inode = dentry->d_inode;
2831         u64 index;
2832         int check_link = 1;
2833         int err = -ENOSPC;
2834         int ret;
2835         u64 ino = btrfs_ino(inode);
2836         u64 dir_ino = btrfs_ino(dir);
2837
2838         trans = btrfs_start_transaction(root, 10);
2839         if (!IS_ERR(trans) || PTR_ERR(trans) != -ENOSPC)
2840                 return trans;
2841
2842         if (ino == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)
2843                 return ERR_PTR(-ENOSPC);
2844
2845         /* check if there is someone else holds reference */
2846         if (S_ISDIR(inode->i_mode) && atomic_read(&inode->i_count) > 1)
2847                 return ERR_PTR(-ENOSPC);
2848
2849         if (atomic_read(&inode->i_count) > 2)
2850                 return ERR_PTR(-ENOSPC);
2851
2852         if (xchg(&root->fs_info->enospc_unlink, 1))
2853                 return ERR_PTR(-ENOSPC);
2854
2855         path = btrfs_alloc_path();
2856         if (!path) {
2857                 root->fs_info->enospc_unlink = 0;
2858                 return ERR_PTR(-ENOMEM);
2859         }
2860
2861         trans = btrfs_start_transaction(root, 0);
2862         if (IS_ERR(trans)) {
2863                 btrfs_free_path(path);
2864                 root->fs_info->enospc_unlink = 0;
2865                 return trans;
2866         }
2867
2868         path->skip_locking = 1;
2869         path->search_commit_root = 1;
2870
2871         ret = btrfs_lookup_inode(trans, root, path,
2872                                 &BTRFS_I(dir)->location, 0);
2873         if (ret < 0) {
2874                 err = ret;
2875                 goto out;
2876         }
2877         if (ret == 0) {
2878                 if (check_path_shared(root, path))
2879                         goto out;
2880         } else {
2881                 check_link = 0;
2882         }
2883         btrfs_release_path(path);
2884
2885         ret = btrfs_lookup_inode(trans, root, path,
2886                                 &BTRFS_I(inode)->location, 0);
2887         if (ret < 0) {
2888                 err = ret;
2889                 goto out;
2890         }
2891         if (ret == 0) {
2892                 if (check_path_shared(root, path))
2893                         goto out;
2894         } else {
2895                 check_link = 0;
2896         }
2897         btrfs_release_path(path);
2898
2899         if (ret == 0 && S_ISREG(inode->i_mode)) {
2900                 ret = btrfs_lookup_file_extent(trans, root, path,
2901                                                ino, (u64)-1, 0);
2902                 if (ret < 0) {
2903                         err = ret;
2904                         goto out;
2905                 }
2906                 BUG_ON(ret == 0);
2907                 if (check_path_shared(root, path))
2908                         goto out;
2909                 btrfs_release_path(path);
2910         }
2911
2912         if (!check_link) {
2913                 err = 0;
2914                 goto out;
2915         }
2916
2917         di = btrfs_lookup_dir_item(trans, root, path, dir_ino,
2918                                 dentry->d_name.name, dentry->d_name.len, 0);
2919         if (IS_ERR(di)) {
2920                 err = PTR_ERR(di);
2921                 goto out;
2922         }
2923         if (di) {
2924                 if (check_path_shared(root, path))
2925                         goto out;
2926         } else {
2927                 err = 0;
2928                 goto out;
2929         }
2930         btrfs_release_path(path);
2931
2932         ref = btrfs_lookup_inode_ref(trans, root, path,
2933                                 dentry->d_name.name, dentry->d_name.len,
2934                                 ino, dir_ino, 0);
2935         if (IS_ERR(ref)) {
2936                 err = PTR_ERR(ref);
2937                 goto out;
2938         }
2939         BUG_ON(!ref);
2940         if (check_path_shared(root, path))
2941                 goto out;
2942         index = btrfs_inode_ref_index(path->nodes[0], ref);
2943         btrfs_release_path(path);
2944
2945         /*
2946          * This is a commit root search, if we can lookup inode item and other
2947          * relative items in the commit root, it means the transaction of
2948          * dir/file creation has been committed, and the dir index item that we
2949          * delay to insert has also been inserted into the commit root. So
2950          * we needn't worry about the delayed insertion of the dir index item
2951          * here.
2952          */
2953         di = btrfs_lookup_dir_index_item(trans, root, path, dir_ino, index,
2954                                 dentry->d_name.name, dentry->d_name.len, 0);
2955         if (IS_ERR(di)) {
2956                 err = PTR_ERR(di);
2957                 goto out;
2958         }
2959         BUG_ON(ret == -ENOENT);
2960         if (check_path_shared(root, path))
2961                 goto out;
2962
2963         err = 0;
2964 out:
2965         btrfs_free_path(path);
2966         if (err) {
2967                 btrfs_end_transaction(trans, root);
2968                 root->fs_info->enospc_unlink = 0;
2969                 return ERR_PTR(err);
2970         }
2971
2972         trans->block_rsv = &root->fs_info->global_block_rsv;
2973         return trans;
2974 }
2975
2976 static void __unlink_end_trans(struct btrfs_trans_handle *trans,
2977                                struct btrfs_root *root)
2978 {
2979         if (trans->block_rsv == &root->fs_info->global_block_rsv) {
2980                 BUG_ON(!root->fs_info->enospc_unlink);
2981                 root->fs_info->enospc_unlink = 0;
2982         }
2983         btrfs_end_transaction_throttle(trans, root);
2984 }
2985
2986 static int btrfs_unlink(struct inode *dir, struct dentry *dentry)
2987 {
2988         struct btrfs_root *root = BTRFS_I(dir)->root;
2989         struct btrfs_trans_handle *trans;
2990         struct inode *inode = dentry->d_inode;
2991         int ret;
2992         unsigned long nr = 0;
2993
2994         trans = __unlink_start_trans(dir, dentry);
2995         if (IS_ERR(trans))
2996                 return PTR_ERR(trans);
2997
2998         btrfs_record_unlink_dir(trans, dir, dentry->d_inode, 0);
2999
3000         ret = btrfs_unlink_inode(trans, root, dir, dentry->d_inode,
3001                                  dentry->d_name.name, dentry->d_name.len);
3002         if (ret)
3003                 goto out;
3004
3005         if (inode->i_nlink == 0) {
3006                 ret = btrfs_orphan_add(trans, inode);
3007                 if (ret)
3008                         goto out;
3009         }
3010
3011 out:
3012         nr = trans->blocks_used;
3013         __unlink_end_trans(trans, root);
3014         btrfs_btree_balance_dirty(root, nr);
3015         return ret;
3016 }
3017
3018 int btrfs_unlink_subvol(struct btrfs_trans_handle *trans,
3019                         struct btrfs_root *root,
3020                         struct inode *dir, u64 objectid,
3021                         const char *name, int name_len)
3022 {
3023         struct btrfs_path *path;
3024         struct extent_buffer *leaf;
3025         struct btrfs_dir_item *di;
3026         struct btrfs_key key;
3027         u64 index;
3028         int ret;
3029         u64 dir_ino = btrfs_ino(dir);
3030
3031         path = btrfs_alloc_path();
3032         if (!path)
3033                 return -ENOMEM;
3034
3035         di = btrfs_lookup_dir_item(trans, root, path, dir_ino,
3036                                    name, name_len, -1);
3037         BUG_ON(IS_ERR_OR_NULL(di));
3038
3039         leaf = path->nodes[0];
3040         btrfs_dir_item_key_to_cpu(leaf, di, &key);
3041         WARN_ON(key.type != BTRFS_ROOT_ITEM_KEY || key.objectid != objectid);
3042         ret = btrfs_delete_one_dir_name(trans, root, path, di);
3043         BUG_ON(ret);
3044         btrfs_release_path(path);
3045
3046         ret = btrfs_del_root_ref(trans, root->fs_info->tree_root,
3047                                  objectid, root->root_key.objectid,
3048                                  dir_ino, &index, name, name_len);
3049         if (ret < 0) {
3050                 BUG_ON(ret != -ENOENT);
3051                 di = btrfs_search_dir_index_item(root, path, dir_ino,
3052                                                  name, name_len);
3053                 BUG_ON(IS_ERR_OR_NULL(di));
3054
3055                 leaf = path->nodes[0];
3056                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3057                 btrfs_release_path(path);
3058                 index = key.offset;
3059         }
3060         btrfs_release_path(path);
3061
3062         ret = btrfs_delete_delayed_dir_index(trans, root, dir, index);
3063         BUG_ON(ret);
3064
3065         btrfs_i_size_write(dir, dir->i_size - name_len * 2);
3066         dir->i_mtime = dir->i_ctime = CURRENT_TIME;
3067         ret = btrfs_update_inode(trans, root, dir);
3068         BUG_ON(ret);
3069
3070         btrfs_free_path(path);
3071         return 0;
3072 }
3073
3074 static int btrfs_rmdir(struct inode *dir, struct dentry *dentry)
3075 {
3076         struct inode *inode = dentry->d_inode;
3077         int err = 0;
3078         struct btrfs_root *root = BTRFS_I(dir)->root;
3079         struct btrfs_trans_handle *trans;
3080         unsigned long nr = 0;
3081
3082         if (inode->i_size > BTRFS_EMPTY_DIR_SIZE ||
3083             btrfs_ino(inode) == BTRFS_FIRST_FREE_OBJECTID)
3084                 return -ENOTEMPTY;
3085
3086         trans = __unlink_start_trans(dir, dentry);
3087         if (IS_ERR(trans))
3088                 return PTR_ERR(trans);