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/aio.h>
  36 #include <linux/bit_spinlock.h>
  37 #include <linux/xattr.h>
  38 #include <linux/posix_acl.h>
  39 #include <linux/falloc.h>
  40 #include <linux/slab.h>
  41 #include <linux/ratelimit.h>
  42 #include <linux/mount.h>
  43 #include <linux/btrfs.h>
  44 #include <linux/blkdev.h>
  45 #include <linux/posix_acl_xattr.h>
  46 #include "ctree.h"
  47 #include "disk-io.h"
  48 #include "transaction.h"
  49 #include "btrfs_inode.h"
  50 #include "print-tree.h"
  51 #include "ordered-data.h"
  52 #include "xattr.h"
  53 #include "tree-log.h"
  54 #include "volumes.h"
  55 #include "compression.h"
  56 #include "locking.h"
  57 #include "free-space-cache.h"
  58 #include "inode-map.h"
  59 #include "backref.h"
  60 #include "hash.h"
  61
  62 struct btrfs_iget_args {
  63         u64 ino;
  64         struct btrfs_root *root;
  65 };
  66
  67 static const struct inode_operations btrfs_dir_inode_operations;
  68 static const struct inode_operations btrfs_symlink_inode_operations;
  69 static const struct inode_operations btrfs_dir_ro_inode_operations;
  70 static const struct inode_operations btrfs_special_inode_operations;
  71 static const struct inode_operations btrfs_file_inode_operations;
  72 static const struct address_space_operations btrfs_aops;
  73 static const struct address_space_operations btrfs_symlink_aops;
  74 static const struct file_operations btrfs_dir_file_operations;
  75 static struct extent_io_ops btrfs_extent_io_ops;
  76
  77 static struct kmem_cache *btrfs_inode_cachep;
  78 static struct kmem_cache *btrfs_delalloc_work_cachep;
  79 struct kmem_cache *btrfs_trans_handle_cachep;
  80 struct kmem_cache *btrfs_transaction_cachep;
  81 struct kmem_cache *btrfs_path_cachep;
  82 struct kmem_cache *btrfs_free_space_cachep;
  83
  84 #define S_SHIFT 12
  85 static unsigned char btrfs_type_by_mode[S_IFMT >> S_SHIFT] = {
  86         [S_IFREG >> S_SHIFT]    = BTRFS_FT_REG_FILE,
  87         [S_IFDIR >> S_SHIFT]    = BTRFS_FT_DIR,
  88         [S_IFCHR >> S_SHIFT]    = BTRFS_FT_CHRDEV,
  89         [S_IFBLK >> S_SHIFT]    = BTRFS_FT_BLKDEV,
  90         [S_IFIFO >> S_SHIFT]    = BTRFS_FT_FIFO,
  91         [S_IFSOCK >> S_SHIFT]   = BTRFS_FT_SOCK,
  92         [S_IFLNK >> S_SHIFT]    = BTRFS_FT_SYMLINK,
  93 };
  94
  95 static int btrfs_setsize(struct inode *inode, struct iattr *attr);
  96 static int btrfs_truncate(struct inode *inode);
  97 static int btrfs_finish_ordered_io(struct btrfs_ordered_extent *ordered_extent);
  98 static noinline int cow_file_range(struct inode *inode,
  99                                    struct page *locked_page,
 100                                    u64 start, u64 end, int *page_started,
 101                                    unsigned long *nr_written, int unlock);
 102 static struct extent_map *create_pinned_em(struct inode *inode, u64 start,
 103                                            u64 len, u64 orig_start,
 104                                            u64 block_start, u64 block_len,
 105                                            u64 orig_block_len, u64 ram_bytes,
 106                                            int type);
 107
 108 static int btrfs_dirty_inode(struct inode *inode);
 109
 110 static int btrfs_init_inode_security(struct btrfs_trans_handle *trans,
 111                                      struct inode *inode,  struct inode *dir,
 112                                      const struct qstr *qstr)
 113 {
 114         int err;
 115
 116         err = btrfs_init_acl(trans, inode, dir);
 117         if (!err)
 118                 err = btrfs_xattr_security_init(trans, inode, dir, qstr);
 119         return err;
 120 }
 121
 122 /*
 123  * this does all the hard work for inserting an inline extent into
 124  * the btree.  The caller should have done a btrfs_drop_extents so that
 125  * no overlapping inline items exist in the btree
 126  */
 127 static noinline int insert_inline_extent(struct btrfs_trans_handle *trans,
 128                                 struct btrfs_root *root, struct inode *inode,
 129                                 u64 start, size_t size, size_t compressed_size,
 130                                 int compress_type,
 131                                 struct page **compressed_pages)
 132 {
 133         struct btrfs_key key;
 134         struct btrfs_path *path;
 135         struct extent_buffer *leaf;
 136         struct page *page = NULL;
 137         char *kaddr;
 138         unsigned long ptr;
 139         struct btrfs_file_extent_item *ei;
 140         int err = 0;
 141         int ret;
 142         size_t cur_size = size;
 143         size_t datasize;
 144         unsigned long offset;
 145
 146         if (compressed_size && compressed_pages)
 147                 cur_size = compressed_size;
 148
 149         path = btrfs_alloc_path();
 150         if (!path)
 151                 return -ENOMEM;
 152
 153         path->leave_spinning = 1;
 154
 155         key.objectid = btrfs_ino(inode);
 156         key.offset = start;
 157         btrfs_set_key_type(&key, BTRFS_EXTENT_DATA_KEY);
 158         datasize = btrfs_file_extent_calc_inline_size(cur_size);
 159
 160         inode_add_bytes(inode, size);
 161         ret = btrfs_insert_empty_item(trans, root, path, &key,
 162                                       datasize);
 163         if (ret) {
 164                 err = ret;
 165                 goto fail;
 166         }
 167         leaf = path->nodes[0];
 168         ei = btrfs_item_ptr(leaf, path->slots[0],
 169                             struct btrfs_file_extent_item);
 170         btrfs_set_file_extent_generation(leaf, ei, trans->transid);
 171         btrfs_set_file_extent_type(leaf, ei, BTRFS_FILE_EXTENT_INLINE);
 172         btrfs_set_file_extent_encryption(leaf, ei, 0);
 173         btrfs_set_file_extent_other_encoding(leaf, ei, 0);
 174         btrfs_set_file_extent_ram_bytes(leaf, ei, size);
 175         ptr = btrfs_file_extent_inline_start(ei);
 176
 177         if (compress_type != BTRFS_COMPRESS_NONE) {
 178                 struct page *cpage;
 179                 int i = 0;
 180                 while (compressed_size > 0) {
 181                         cpage = compressed_pages[i];
 182                         cur_size = min_t(unsigned long, compressed_size,
 183                                        PAGE_CACHE_SIZE);
 184
 185                         kaddr = kmap_atomic(cpage);
 186                         write_extent_buffer(leaf, kaddr, ptr, cur_size);
 187                         kunmap_atomic(kaddr);
 188
 189                         i++;
 190                         ptr += cur_size;
 191                         compressed_size -= cur_size;
 192                 }
 193                 btrfs_set_file_extent_compression(leaf, ei,
 194                                                   compress_type);
 195         } else {
 196                 page = find_get_page(inode->i_mapping,
 197                                      start >> PAGE_CACHE_SHIFT);
 198                 btrfs_set_file_extent_compression(leaf, ei, 0);
 199                 kaddr = kmap_atomic(page);
 200                 offset = start & (PAGE_CACHE_SIZE - 1);
 201                 write_extent_buffer(leaf, kaddr + offset, ptr, size);
 202                 kunmap_atomic(kaddr);
 203                 page_cache_release(page);
 204         }
 205         btrfs_mark_buffer_dirty(leaf);
 206         btrfs_free_path(path);
 207
 208         /*
 209          * we're an inline extent, so nobody can
 210          * extend the file past i_size without locking
 211          * a page we already have locked.
 212          *
 213          * We must do any isize and inode updates
 214          * before we unlock the pages.  Otherwise we
 215          * could end up racing with unlink.
 216          */
 217         BTRFS_I(inode)->disk_i_size = inode->i_size;
 218         ret = btrfs_update_inode(trans, root, inode);
 219
 220         return ret;
 221 fail:
 222         btrfs_free_path(path);
 223         return err;
 224 }
 225
 226
 227 /*
 228  * conditionally insert an inline extent into the file.  This
 229  * does the checks required to make sure the data is small enough
 230  * to fit as an inline extent.
 231  */
 232 static noinline int cow_file_range_inline(struct btrfs_root *root,
 233                                           struct inode *inode, u64 start,
 234                                           u64 end, size_t compressed_size,
 235                                           int compress_type,
 236                                           struct page **compressed_pages)
 237 {
 238         struct btrfs_trans_handle *trans;
 239         u64 isize = i_size_read(inode);
 240         u64 actual_end = min(end + 1, isize);
 241         u64 inline_len = actual_end - start;
 242         u64 aligned_end = ALIGN(end, root->sectorsize);
 243         u64 data_len = inline_len;
 244         int ret;
 245
 246         if (compressed_size)
 247                 data_len = compressed_size;
 248
 249         if (start > 0 ||
 250             actual_end >= PAGE_CACHE_SIZE ||
 251             data_len >= BTRFS_MAX_INLINE_DATA_SIZE(root) ||
 252             (!compressed_size &&
 253             (actual_end & (root->sectorsize - 1)) == 0) ||
 254             end + 1 < isize ||
 255             data_len > root->fs_info->max_inline) {
 256                 return 1;
 257         }
 258
 259         trans = btrfs_join_transaction(root);
 260         if (IS_ERR(trans))
 261                 return PTR_ERR(trans);
 262         trans->block_rsv = &root->fs_info->delalloc_block_rsv;
 263
 264         ret = btrfs_drop_extents(trans, root, inode, start, aligned_end, 1);
 265         if (ret) {
 266                 btrfs_abort_transaction(trans, root, ret);
 267                 goto out;
 268         }
 269
 270         if (isize > actual_end)
 271                 inline_len = min_t(u64, isize, actual_end);
 272         ret = insert_inline_extent(trans, root, inode, start,
 273                                    inline_len, compressed_size,
 274                                    compress_type, compressed_pages);
 275         if (ret && ret != -ENOSPC) {
 276                 btrfs_abort_transaction(trans, root, ret);
 277                 goto out;
 278         } else if (ret == -ENOSPC) {
 279                 ret = 1;
 280                 goto out;
 281         }
 282
 283         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &BTRFS_I(inode)->runtime_flags);
 284         btrfs_delalloc_release_metadata(inode, end + 1 - start);
 285         btrfs_drop_extent_cache(inode, start, aligned_end - 1, 0);
 286 out:
 287         btrfs_end_transaction(trans, root);
 288         return ret;
 289 }
 290
 291 struct async_extent {
 292         u64 start;
 293         u64 ram_size;
 294         u64 compressed_size;
 295         struct page **pages;
 296         unsigned long nr_pages;
 297         int compress_type;
 298         struct list_head list;
 299 };
 300
 301 struct async_cow {
 302         struct inode *inode;
 303         struct btrfs_root *root;
 304         struct page *locked_page;
 305         u64 start;
 306         u64 end;
 307         struct list_head extents;
 308         struct btrfs_work work;
 309 };
 310
 311 static noinline int add_async_extent(struct async_cow *cow,
 312                                      u64 start, u64 ram_size,
 313                                      u64 compressed_size,
 314                                      struct page **pages,
 315                                      unsigned long nr_pages,
 316                                      int compress_type)
 317 {
 318         struct async_extent *async_extent;
 319
 320         async_extent = kmalloc(sizeof(*async_extent), GFP_NOFS);
 321         BUG_ON(!async_extent); /* -ENOMEM */
 322         async_extent->start = start;
 323         async_extent->ram_size = ram_size;
 324         async_extent->compressed_size = compressed_size;
 325         async_extent->pages = pages;
 326         async_extent->nr_pages = nr_pages;
 327         async_extent->compress_type = compress_type;
 328         list_add_tail(&async_extent->list, &cow->extents);
 329         return 0;
 330 }
 331
 332 /*
 333  * we create compressed extents in two phases.  The first
 334  * phase compresses a range of pages that have already been
 335  * locked (both pages and state bits are locked).
 336  *
 337  * This is done inside an ordered work queue, and the compression
 338  * is spread across many cpus.  The actual IO submission is step
 339  * two, and the ordered work queue takes care of making sure that
 340  * happens in the same order things were put onto the queue by
 341  * writepages and friends.
 342  *
 343  * If this code finds it can't get good compression, it puts an
 344  * entry onto the work queue to write the uncompressed bytes.  This
 345  * makes sure that both compressed inodes and uncompressed inodes
 346  * are written in the same order that the flusher thread sent them
 347  * down.
 348  */
 349 static noinline int compress_file_range(struct inode *inode,
 350                                         struct page *locked_page,
 351                                         u64 start, u64 end,
 352                                         struct async_cow *async_cow,
 353                                         int *num_added)
 354 {
 355         struct btrfs_root *root = BTRFS_I(inode)->root;
 356         u64 num_bytes;
 357         u64 blocksize = root->sectorsize;
 358         u64 actual_end;
 359         u64 isize = i_size_read(inode);
 360         int ret = 0;
 361         struct page **pages = NULL;
 362         unsigned long nr_pages;
 363         unsigned long nr_pages_ret = 0;
 364         unsigned long total_compressed = 0;
 365         unsigned long total_in = 0;
 366         unsigned long max_compressed = 128 * 1024;
 367         unsigned long max_uncompressed = 128 * 1024;
 368         int i;
 369         int will_compress;
 370         int compress_type = root->fs_info->compress_type;
 371         int redirty = 0;
 372
 373         /* if this is a small write inside eof, kick off a defrag */
 374         if ((end - start + 1) < 16 * 1024 &&
 375             (start > 0 || end + 1 < BTRFS_I(inode)->disk_i_size))
 376                 btrfs_add_inode_defrag(NULL, inode);
 377
 378         actual_end = min_t(u64, isize, end + 1);
 379 again:
 380         will_compress = 0;
 381         nr_pages = (end >> PAGE_CACHE_SHIFT) - (start >> PAGE_CACHE_SHIFT) + 1;
 382         nr_pages = min(nr_pages, (128 * 1024UL) / PAGE_CACHE_SIZE);
 383
 384         /*
 385          * we don't want to send crud past the end of i_size through
 386          * compression, that's just a waste of CPU time.  So, if the
 387          * end of the file is before the start of our current
 388          * requested range of bytes, we bail out to the uncompressed
 389          * cleanup code that can deal with all of this.
 390          *
 391          * It isn't really the fastest way to fix things, but this is a
 392          * very uncommon corner.
 393          */
 394         if (actual_end <= start)
 395                 goto cleanup_and_bail_uncompressed;
 396
 397         total_compressed = actual_end - start;
 398
 399         /* we want to make sure that amount of ram required to uncompress
 400          * an extent is reasonable, so we limit the total size in ram
 401          * of a compressed extent to 128k.  This is a crucial number
 402          * because it also controls how easily we can spread reads across
 403          * cpus for decompression.
 404          *
 405          * We also want to make sure the amount of IO required to do
 406          * a random read is reasonably small, so we limit the size of
 407          * a compressed extent to 128k.
 408          */
 409         total_compressed = min(total_compressed, max_uncompressed);
 410         num_bytes = ALIGN(end - start + 1, blocksize);
 411         num_bytes = max(blocksize,  num_bytes);
 412         total_in = 0;
 413         ret = 0;
 414
 415         /*
 416          * we do compression for mount -o compress and when the
 417          * inode has not been flagged as nocompress.  This flag can
 418          * change at any time if we discover bad compression ratios.
 419          */
 420         if (!(BTRFS_I(inode)->flags & BTRFS_INODE_NOCOMPRESS) &&
 421             (btrfs_test_opt(root, COMPRESS) ||
 422              (BTRFS_I(inode)->force_compress) ||
 423              (BTRFS_I(inode)->flags & BTRFS_INODE_COMPRESS))) {
 424                 WARN_ON(pages);
 425                 pages = kzalloc(sizeof(struct page *) * nr_pages, GFP_NOFS);
 426                 if (!pages) {
 427                         /* just bail out to the uncompressed code */
 428                         goto cont;
 429                 }
 430
 431                 if (BTRFS_I(inode)->force_compress)
 432                         compress_type = BTRFS_I(inode)->force_compress;
 433
 434                 /*
 435                  * we need to call clear_page_dirty_for_io on each
 436                  * page in the range.  Otherwise applications with the file
 437                  * mmap'd can wander in and change the page contents while
 438                  * we are compressing them.
 439                  *
 440                  * If the compression fails for any reason, we set the pages
 441                  * dirty again later on.
 442                  */
 443                 extent_range_clear_dirty_for_io(inode, start, end);
 444                 redirty = 1;
 445                 ret = btrfs_compress_pages(compress_type,
 446                                            inode->i_mapping, start,
 447                                            total_compressed, pages,
 448                                            nr_pages, &nr_pages_ret,
 449                                            &total_in,
 450                                            &total_compressed,
 451                                            max_compressed);
 452
 453                 if (!ret) {
 454                         unsigned long offset = total_compressed &
 455                                 (PAGE_CACHE_SIZE - 1);
 456                         struct page *page = pages[nr_pages_ret - 1];
 457                         char *kaddr;
 458
 459                         /* zero the tail end of the last page, we might be
 460                          * sending it down to disk
 461                          */
 462                         if (offset) {
 463                                 kaddr = kmap_atomic(page);
 464                                 memset(kaddr + offset, 0,
 465                                        PAGE_CACHE_SIZE - offset);
 466                                 kunmap_atomic(kaddr);
 467                         }
 468                         will_compress = 1;
 469                 }
 470         }
 471 cont:
 472         if (start == 0) {
 473                 /* lets try to make an inline extent */
 474                 if (ret || total_in < (actual_end - start)) {
 475                         /* we didn't compress the entire range, try
 476                          * to make an uncompressed inline extent.
 477                          */
 478                         ret = cow_file_range_inline(root, inode, start, end,
 479                                                     0, 0, NULL);
 480                 } else {
 481                         /* try making a compressed inline extent */
 482                         ret = cow_file_range_inline(root, inode, start, end,
 483                                                     total_compressed,
 484                                                     compress_type, pages);
 485                 }
 486                 if (ret <= 0) {
 487                         unsigned long clear_flags = EXTENT_DELALLOC |
 488                                 EXTENT_DEFRAG;
 489                         clear_flags |= (ret < 0) ? EXTENT_DO_ACCOUNTING : 0;
 490
 491                         /*
 492                          * inline extent creation worked or returned error,
 493                          * we don't need to create any more async work items.
 494                          * Unlock and free up our temp pages.
 495                          */
 496                         extent_clear_unlock_delalloc(inode, start, end, NULL,
 497                                                      clear_flags, PAGE_UNLOCK |
 498                                                      PAGE_CLEAR_DIRTY |
 499                                                      PAGE_SET_WRITEBACK |
 500                                                      PAGE_END_WRITEBACK);
 501                         goto free_pages_out;
 502                 }
 503         }
 504
 505         if (will_compress) {
 506                 /*
 507                  * we aren't doing an inline extent round the compressed size
 508                  * up to a block size boundary so the allocator does sane
 509                  * things
 510                  */
 511                 total_compressed = ALIGN(total_compressed, blocksize);
 512
 513                 /*
 514                  * one last check to make sure the compression is really a
 515                  * win, compare the page count read with the blocks on disk
 516                  */
 517                 total_in = ALIGN(total_in, PAGE_CACHE_SIZE);
 518                 if (total_compressed >= total_in) {
 519                         will_compress = 0;
 520                 } else {
 521                         num_bytes = total_in;
 522                 }
 523         }
 524         if (!will_compress && pages) {
 525                 /*
 526                  * the compression code ran but failed to make things smaller,
 527                  * free any pages it allocated and our page pointer array
 528                  */
 529                 for (i = 0; i < nr_pages_ret; i++) {
 530                         WARN_ON(pages[i]->mapping);
 531                         page_cache_release(pages[i]);
 532                 }
 533                 kfree(pages);
 534                 pages = NULL;
 535                 total_compressed = 0;
 536                 nr_pages_ret = 0;
 537
 538                 /* flag the file so we don't compress in the future */
 539                 if (!btrfs_test_opt(root, FORCE_COMPRESS) &&
 540                     !(BTRFS_I(inode)->force_compress)) {
 541                         BTRFS_I(inode)->flags |= BTRFS_INODE_NOCOMPRESS;
 542                 }
 543         }
 544         if (will_compress) {
 545                 *num_added += 1;
 546
 547                 /* the async work queues will take care of doing actual
 548                  * allocation on disk for these compressed pages,
 549                  * and will submit them to the elevator.
 550                  */
 551                 add_async_extent(async_cow, start, num_bytes,
 552                                  total_compressed, pages, nr_pages_ret,
 553                                  compress_type);
 554
 555                 if (start + num_bytes < end) {
 556                         start += num_bytes;
 557                         pages = NULL;
 558                         cond_resched();
 559                         goto again;
 560                 }
 561         } else {
 562 cleanup_and_bail_uncompressed:
 563                 /*
 564                  * No compression, but we still need to write the pages in
 565                  * the file we've been given so far.  redirty the locked
 566                  * page if it corresponds to our extent and set things up
 567                  * for the async work queue to run cow_file_range to do
 568                  * the normal delalloc dance
 569                  */
 570                 if (page_offset(locked_page) >= start &&
 571                     page_offset(locked_page) <= end) {
 572                         __set_page_dirty_nobuffers(locked_page);
 573                         /* unlocked later on in the async handlers */
 574                 }
 575                 if (redirty)
 576                         extent_range_redirty_for_io(inode, start, end);
 577                 add_async_extent(async_cow, start, end - start + 1,
 578                                  0, NULL, 0, BTRFS_COMPRESS_NONE);
 579                 *num_added += 1;
 580         }
 581
 582 out:
 583         return ret;
 584
 585 free_pages_out:
 586         for (i = 0; i < nr_pages_ret; i++) {
 587                 WARN_ON(pages[i]->mapping);
 588                 page_cache_release(pages[i]);
 589         }
 590         kfree(pages);
 591
 592         goto out;
 593 }
 594
 595 /*
 596  * phase two of compressed writeback.  This is the ordered portion
 597  * of the code, which only gets called in the order the work was
 598  * queued.  We walk all the async extents created by compress_file_range
 599  * and send them down to the disk.
 600  */
 601 static noinline int submit_compressed_extents(struct inode *inode,
 602                                               struct async_cow *async_cow)
 603 {
 604         struct async_extent *async_extent;
 605         u64 alloc_hint = 0;
 606         struct btrfs_key ins;
 607         struct extent_map *em;
 608         struct btrfs_root *root = BTRFS_I(inode)->root;
 609         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
 610         struct extent_io_tree *io_tree;
 611         int ret = 0;
 612
 613         if (list_empty(&async_cow->extents))
 614                 return 0;
 615
 616 again:
 617         while (!list_empty(&async_cow->extents)) {
 618                 async_extent = list_entry(async_cow->extents.next,
 619                                           struct async_extent, list);
 620                 list_del(&async_extent->list);
 621
 622                 io_tree = &BTRFS_I(inode)->io_tree;
 623
 624 retry:
 625                 /* did the compression code fall back to uncompressed IO? */
 626                 if (!async_extent->pages) {
 627                         int page_started = 0;
 628                         unsigned long nr_written = 0;
 629
 630                         lock_extent(io_tree, async_extent->start,
 631                                          async_extent->start +
 632                                          async_extent->ram_size - 1);
 633
 634                         /* allocate blocks */
 635                         ret = cow_file_range(inode, async_cow->locked_page,
 636                                              async_extent->start,
 637                                              async_extent->start +
 638                                              async_extent->ram_size - 1,
 639                                              &page_started, &nr_written, 0);
 640
 641                         /* JDM XXX */
 642
 643                         /*
 644                          * if page_started, cow_file_range inserted an
 645                          * inline extent and took care of all the unlocking
 646                          * and IO for us.  Otherwise, we need to submit
 647                          * all those pages down to the drive.
 648                          */
 649                         if (!page_started && !ret)
 650                                 extent_write_locked_range(io_tree,
 651                                                   inode, async_extent->start,
 652                                                   async_extent->start +
 653                                                   async_extent->ram_size - 1,
 654                                                   btrfs_get_extent,
 655                                                   WB_SYNC_ALL);
 656                         else if (ret)
 657                                 unlock_page(async_cow->locked_page);
 658                         kfree(async_extent);
 659                         cond_resched();
 660                         continue;
 661                 }
 662
 663                 lock_extent(io_tree, async_extent->start,
 664                             async_extent->start + async_extent->ram_size - 1);
 665
 666                 ret = btrfs_reserve_extent(root,
 667                                            async_extent->compressed_size,
 668                                            async_extent->compressed_size,
 669                                            0, alloc_hint, &ins, 1);
 670                 if (ret) {
 671                         int i;
 672
 673                         for (i = 0; i < async_extent->nr_pages; i++) {
 674                                 WARN_ON(async_extent->pages[i]->mapping);
 675                                 page_cache_release(async_extent->pages[i]);
 676                         }
 677                         kfree(async_extent->pages);
 678                         async_extent->nr_pages = 0;
 679                         async_extent->pages = NULL;
 680
 681                         if (ret == -ENOSPC) {
 682                                 unlock_extent(io_tree, async_extent->start,
 683                                               async_extent->start +
 684                                               async_extent->ram_size - 1);
 685                                 goto retry;
 686                         }
 687                         goto out_free;
 688                 }
 689
 690                 /*
 691                  * here we're doing allocation and writeback of the
 692                  * compressed pages
 693                  */
 694                 btrfs_drop_extent_cache(inode, async_extent->start,
 695                                         async_extent->start +
 696                                         async_extent->ram_size - 1, 0);
 697
 698                 em = alloc_extent_map();
 699                 if (!em) {
 700                         ret = -ENOMEM;
 701                         goto out_free_reserve;
 702                 }
 703                 em->start = async_extent->start;
 704                 em->len = async_extent->ram_size;
 705                 em->orig_start = em->start;
 706                 em->mod_start = em->start;
 707                 em->mod_len = em->len;
 708
 709                 em->block_start = ins.objectid;
 710                 em->block_len = ins.offset;
 711                 em->orig_block_len = ins.offset;
 712                 em->ram_bytes = async_extent->ram_size;
 713                 em->bdev = root->fs_info->fs_devices->latest_bdev;
 714                 em->compress_type = async_extent->compress_type;
 715                 set_bit(EXTENT_FLAG_PINNED, &em->flags);
 716                 set_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
 717                 em->generation = -1;
 718
 719                 while (1) {
 720                         write_lock(&em_tree->lock);
 721                         ret = add_extent_mapping(em_tree, em, 1);
 722                         write_unlock(&em_tree->lock);
 723                         if (ret != -EEXIST) {
 724                                 free_extent_map(em);
 725                                 break;
 726                         }
 727                         btrfs_drop_extent_cache(inode, async_extent->start,
 728                                                 async_extent->start +
 729                                                 async_extent->ram_size - 1, 0);
 730                 }
 731
 732                 if (ret)
 733                         goto out_free_reserve;
 734
 735                 ret = btrfs_add_ordered_extent_compress(inode,
 736                                                 async_extent->start,
 737                                                 ins.objectid,
 738                                                 async_extent->ram_size,
 739                                                 ins.offset,
 740                                                 BTRFS_ORDERED_COMPRESSED,
 741                                                 async_extent->compress_type);
 742                 if (ret)
 743                         goto out_free_reserve;
 744
 745                 /*
 746                  * clear dirty, set writeback and unlock the pages.
 747                  */
 748                 extent_clear_unlock_delalloc(inode, async_extent->start,
 749                                 async_extent->start +
 750                                 async_extent->ram_size - 1,
 751                                 NULL, EXTENT_LOCKED | EXTENT_DELALLOC,
 752                                 PAGE_UNLOCK | PAGE_CLEAR_DIRTY |
 753                                 PAGE_SET_WRITEBACK);
 754                 ret = btrfs_submit_compressed_write(inode,
 755                                     async_extent->start,
 756                                     async_extent->ram_size,
 757                                     ins.objectid,
 758                                     ins.offset, async_extent->pages,
 759                                     async_extent->nr_pages);
 760                 alloc_hint = ins.objectid + ins.offset;
 761                 kfree(async_extent);
 762                 if (ret)
 763                         goto out;
 764                 cond_resched();
 765         }
 766         ret = 0;
 767 out:
 768         return ret;
 769 out_free_reserve:
 770         btrfs_free_reserved_extent(root, ins.objectid, ins.offset);
 771 out_free:
 772         extent_clear_unlock_delalloc(inode, async_extent->start,
 773                                      async_extent->start +
 774                                      async_extent->ram_size - 1,
 775                                      NULL, EXTENT_LOCKED | EXTENT_DELALLOC |
 776                                      EXTENT_DEFRAG | EXTENT_DO_ACCOUNTING,
 777                                      PAGE_UNLOCK | PAGE_CLEAR_DIRTY |
 778                                      PAGE_SET_WRITEBACK | PAGE_END_WRITEBACK);
 779         kfree(async_extent);
 780         goto again;
 781 }
 782
 783 static u64 get_extent_allocation_hint(struct inode *inode, u64 start,
 784                                       u64 num_bytes)
 785 {
 786         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
 787         struct extent_map *em;
 788         u64 alloc_hint = 0;
 789
 790         read_lock(&em_tree->lock);
 791         em = search_extent_mapping(em_tree, start, num_bytes);
 792         if (em) {
 793                 /*
 794                  * if block start isn't an actual block number then find the
 795                  * first block in this inode and use that as a hint.  If that
 796                  * block is also bogus then just don't worry about it.
 797                  */
 798                 if (em->block_start >= EXTENT_MAP_LAST_BYTE) {
 799                         free_extent_map(em);
 800                         em = search_extent_mapping(em_tree, 0, 0);
 801                         if (em && em->block_start < EXTENT_MAP_LAST_BYTE)
 802                                 alloc_hint = em->block_start;
 803                         if (em)
 804                                 free_extent_map(em);
 805                 } else {
 806                         alloc_hint = em->block_start;
 807                         free_extent_map(em);
 808                 }
 809         }
 810         read_unlock(&em_tree->lock);
 811
 812         return alloc_hint;
 813 }
 814
 815 /*
 816  * when extent_io.c finds a delayed allocation range in the file,
 817  * the call backs end up in this code.  The basic idea is to
 818  * allocate extents on disk for the range, and create ordered data structs
 819  * in ram to track those extents.
 820  *
 821  * locked_page is the page that writepage had locked already.  We use
 822  * it to make sure we don't do extra locks or unlocks.
 823  *
 824  * *page_started is set to one if we unlock locked_page and do everything
 825  * required to start IO on it.  It may be clean and already done with
 826  * IO when we return.
 827  */
 828 static noinline int cow_file_range(struct inode *inode,
 829                                    struct page *locked_page,
 830                                    u64 start, u64 end, int *page_started,
 831                                    unsigned long *nr_written,
 832                                    int unlock)
 833 {
 834         struct btrfs_root *root = BTRFS_I(inode)->root;
 835         u64 alloc_hint = 0;
 836         u64 num_bytes;
 837         unsigned long ram_size;
 838         u64 disk_num_bytes;
 839         u64 cur_alloc_size;
 840         u64 blocksize = root->sectorsize;
 841         struct btrfs_key ins;
 842         struct extent_map *em;
 843         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
 844         int ret = 0;
 845
 846         if (btrfs_is_free_space_inode(inode)) {
 847                 WARN_ON_ONCE(1);
 848                 return -EINVAL;
 849         }
 850
 851         num_bytes = ALIGN(end - start + 1, blocksize);
 852         num_bytes = max(blocksize,  num_bytes);
 853         disk_num_bytes = num_bytes;
 854
 855         /* if this is a small write inside eof, kick off defrag */
 856         if (num_bytes < 64 * 1024 &&
 857             (start > 0 || end + 1 < BTRFS_I(inode)->disk_i_size))
 858                 btrfs_add_inode_defrag(NULL, inode);
 859
 860         if (start == 0) {
 861                 /* lets try to make an inline extent */
 862                 ret = cow_file_range_inline(root, inode, start, end, 0, 0,
 863                                             NULL);
 864                 if (ret == 0) {
 865                         extent_clear_unlock_delalloc(inode, start, end, NULL,
 866                                      EXTENT_LOCKED | EXTENT_DELALLOC |
 867                                      EXTENT_DEFRAG, PAGE_UNLOCK |
 868                                      PAGE_CLEAR_DIRTY | PAGE_SET_WRITEBACK |
 869                                      PAGE_END_WRITEBACK);
 870
 871                         *nr_written = *nr_written +
 872                              (end - start + PAGE_CACHE_SIZE) / PAGE_CACHE_SIZE;
 873                         *page_started = 1;
 874                         goto out;
 875                 } else if (ret < 0) {
 876                         goto out_unlock;
 877                 }
 878         }
 879
 880         BUG_ON(disk_num_bytes >
 881                btrfs_super_total_bytes(root->fs_info->super_copy));
 882
 883         alloc_hint = get_extent_allocation_hint(inode, start, num_bytes);
 884         btrfs_drop_extent_cache(inode, start, start + num_bytes - 1, 0);
 885
 886         while (disk_num_bytes > 0) {
 887                 unsigned long op;
 888
 889                 cur_alloc_size = disk_num_bytes;
 890                 ret = btrfs_reserve_extent(root, cur_alloc_size,
 891                                            root->sectorsize, 0, alloc_hint,
 892                                            &ins, 1);
 893                 if (ret < 0)
 894                         goto out_unlock;
 895
 896                 em = alloc_extent_map();
 897                 if (!em) {
 898                         ret = -ENOMEM;
 899                         goto out_reserve;
 900                 }
 901                 em->start = start;
 902                 em->orig_start = em->start;
 903                 ram_size = ins.offset;
 904                 em->len = ins.offset;
 905                 em->mod_start = em->start;
 906                 em->mod_len = em->len;
 907
 908                 em->block_start = ins.objectid;
 909                 em->block_len = ins.offset;
 910                 em->orig_block_len = ins.offset;
 911                 em->ram_bytes = ram_size;
 912                 em->bdev = root->fs_info->fs_devices->latest_bdev;
 913                 set_bit(EXTENT_FLAG_PINNED, &em->flags);
 914                 em->generation = -1;
 915
 916                 while (1) {
 917                         write_lock(&em_tree->lock);
 918                         ret = add_extent_mapping(em_tree, em, 1);
 919                         write_unlock(&em_tree->lock);
 920                         if (ret != -EEXIST) {
 921                                 free_extent_map(em);
 922                                 break;
 923                         }
 924                         btrfs_drop_extent_cache(inode, start,
 925                                                 start + ram_size - 1, 0);
 926                 }
 927                 if (ret)
 928                         goto out_reserve;
 929
 930                 cur_alloc_size = ins.offset;
 931                 ret = btrfs_add_ordered_extent(inode, start, ins.objectid,
 932                                                ram_size, cur_alloc_size, 0);
 933                 if (ret)
 934                         goto out_reserve;
 935
 936                 if (root->root_key.objectid ==
 937                     BTRFS_DATA_RELOC_TREE_OBJECTID) {
 938                         ret = btrfs_reloc_clone_csums(inode, start,
 939                                                       cur_alloc_size);
 940                         if (ret)
 941                                 goto out_reserve;
 942                 }
 943
 944                 if (disk_num_bytes < cur_alloc_size)
 945                         break;
 946
 947                 /* we're not doing compressed IO, don't unlock the first
 948                  * page (which the caller expects to stay locked), don't
 949                  * clear any dirty bits and don't set any writeback bits
 950                  *
 951                  * Do set the Private2 bit so we know this page was properly
 952                  * setup for writepage
 953                  */
 954                 op = unlock ? PAGE_UNLOCK : 0;
 955                 op |= PAGE_SET_PRIVATE2;
 956
 957                 extent_clear_unlock_delalloc(inode, start,
 958                                              start + ram_size - 1, locked_page,
 959                                              EXTENT_LOCKED | EXTENT_DELALLOC,
 960                                              op);
 961                 disk_num_bytes -= cur_alloc_size;
 962                 num_bytes -= cur_alloc_size;
 963                 alloc_hint = ins.objectid + ins.offset;
 964                 start += cur_alloc_size;
 965         }
 966 out:
 967         return ret;
 968
 969 out_reserve:
 970         btrfs_free_reserved_extent(root, ins.objectid, ins.offset);
 971 out_unlock:
 972         extent_clear_unlock_delalloc(inode, start, end, locked_page,
 973                                      EXTENT_LOCKED | EXTENT_DO_ACCOUNTING |
 974                                      EXTENT_DELALLOC | EXTENT_DEFRAG,
 975                                      PAGE_UNLOCK | PAGE_CLEAR_DIRTY |
 976                                      PAGE_SET_WRITEBACK | PAGE_END_WRITEBACK);
 977         goto out;
 978 }
 979
 980 /*
 981  * work queue call back to started compression on a file and pages
 982  */
 983 static noinline void async_cow_start(struct btrfs_work *work)
 984 {
 985         struct async_cow *async_cow;
 986         int num_added = 0;
 987         async_cow = container_of(work, struct async_cow, work);
 988
 989         compress_file_range(async_cow->inode, async_cow->locked_page,
 990                             async_cow->start, async_cow->end, async_cow,
 991                             &num_added);
 992         if (num_added == 0) {
 993                 btrfs_add_delayed_iput(async_cow->inode);
 994                 async_cow->inode = NULL;
 995         }
 996 }
 997
 998 /*
 999  * work queue call back to submit previously compressed pages
1000  */
1001 static noinline void async_cow_submit(struct btrfs_work *work)
1002 {
1003         struct async_cow *async_cow;
1004         struct btrfs_root *root;
1005         unsigned long nr_pages;
1006
1007         async_cow = container_of(work, struct async_cow, work);
1008
1009         root = async_cow->root;
1010         nr_pages = (async_cow->end - async_cow->start + PAGE_CACHE_SIZE) >>
1011                 PAGE_CACHE_SHIFT;
1012
1013         if (atomic_sub_return(nr_pages, &root->fs_info->async_delalloc_pages) <
1014             5 * 1024 * 1024 &&
1015             waitqueue_active(&root->fs_info->async_submit_wait))
1016                 wake_up(&root->fs_info->async_submit_wait);
1017
1018         if (async_cow->inode)
1019                 submit_compressed_extents(async_cow->inode, async_cow);
1020 }
1021
1022 static noinline void async_cow_free(struct btrfs_work *work)
1023 {
1024         struct async_cow *async_cow;
1025         async_cow = container_of(work, struct async_cow, work);
1026         if (async_cow->inode)
1027                 btrfs_add_delayed_iput(async_cow->inode);
1028         kfree(async_cow);
1029 }
1030
1031 static int cow_file_range_async(struct inode *inode, struct page *locked_page,
1032                                 u64 start, u64 end, int *page_started,
1033                                 unsigned long *nr_written)
1034 {
1035         struct async_cow *async_cow;
1036         struct btrfs_root *root = BTRFS_I(inode)->root;
1037         unsigned long nr_pages;
1038         u64 cur_end;
1039         int limit = 10 * 1024 * 1024;
1040
1041         clear_extent_bit(&BTRFS_I(inode)->io_tree, start, end, EXTENT_LOCKED,
1042                          1, 0, NULL, GFP_NOFS);
1043         while (start < end) {
1044                 async_cow = kmalloc(sizeof(*async_cow), GFP_NOFS);
1045                 BUG_ON(!async_cow); /* -ENOMEM */
1046                 async_cow->inode = igrab(inode);
1047                 async_cow->root = root;
1048                 async_cow->locked_page = locked_page;
1049                 async_cow->start = start;
1050
1051                 if (BTRFS_I(inode)->flags & BTRFS_INODE_NOCOMPRESS)
1052                         cur_end = end;
1053                 else
1054                         cur_end = min(end, start + 512 * 1024 - 1);
1055
1056                 async_cow->end = cur_end;
1057                 INIT_LIST_HEAD(&async_cow->extents);
1058
1059                 async_cow->work.func = async_cow_start;
1060                 async_cow->work.ordered_func = async_cow_submit;
1061                 async_cow->work.ordered_free = async_cow_free;
1062                 async_cow->work.flags = 0;
1063
1064                 nr_pages = (cur_end - start + PAGE_CACHE_SIZE) >>
1065                         PAGE_CACHE_SHIFT;
1066                 atomic_add(nr_pages, &root->fs_info->async_delalloc_pages);
1067
1068                 btrfs_queue_worker(&root->fs_info->delalloc_workers,
1069                                    &async_cow->work);
1070
1071                 if (atomic_read(&root->fs_info->async_delalloc_pages) > limit) {
1072                         wait_event(root->fs_info->async_submit_wait,
1073                            (atomic_read(&root->fs_info->async_delalloc_pages) <
1074                             limit));
1075                 }
1076
1077                 while (atomic_read(&root->fs_info->async_submit_draining) &&
1078                       atomic_read(&root->fs_info->async_delalloc_pages)) {
1079                         wait_event(root->fs_info->async_submit_wait,
1080                           (atomic_read(&root->fs_info->async_delalloc_pages) ==
1081                            0));
1082                 }
1083
1084                 *nr_written += nr_pages;
1085                 start = cur_end + 1;
1086         }
1087         *page_started = 1;
1088         return 0;
1089 }
1090
1091 static noinline int csum_exist_in_range(struct btrfs_root *root,
1092                                         u64 bytenr, u64 num_bytes)
1093 {
1094         int ret;
1095         struct btrfs_ordered_sum *sums;
1096         LIST_HEAD(list);
1097
1098         ret = btrfs_lookup_csums_range(root->fs_info->csum_root, bytenr,
1099                                        bytenr + num_bytes - 1, &list, 0);
1100         if (ret == 0 && list_empty(&list))
1101                 return 0;
1102
1103         while (!list_empty(&list)) {
1104                 sums = list_entry(list.next, struct btrfs_ordered_sum, list);
1105                 list_del(&sums->list);
1106                 kfree(sums);
1107         }
1108         return 1;
1109 }
1110
1111 /*
1112  * when nowcow writeback call back.  This checks for snapshots or COW copies
1113  * of the extents that exist in the file, and COWs the file as required.
1114  *
1115  * If no cow copies or snapshots exist, we write directly to the existing
1116  * blocks on disk
1117  */
1118 static noinline int run_delalloc_nocow(struct inode *inode,
1119                                        struct page *locked_page,
1120                               u64 start, u64 end, int *page_started, int force,
1121                               unsigned long *nr_written)
1122 {
1123         struct btrfs_root *root = BTRFS_I(inode)->root;
1124         struct btrfs_trans_handle *trans;
1125         struct extent_buffer *leaf;
1126         struct btrfs_path *path;
1127         struct btrfs_file_extent_item *fi;
1128         struct btrfs_key found_key;
1129         u64 cow_start;
1130         u64 cur_offset;
1131         u64 extent_end;
1132         u64 extent_offset;
1133         u64 disk_bytenr;
1134         u64 num_bytes;
1135         u64 disk_num_bytes;
1136         u64 ram_bytes;
1137         int extent_type;
1138         int ret, err;
1139         int type;
1140         int nocow;
1141         int check_prev = 1;
1142         bool nolock;
1143         u64 ino = btrfs_ino(inode);
1144
1145         path = btrfs_alloc_path();
1146         if (!path) {
1147                 extent_clear_unlock_delalloc(inode, start, end, locked_page,
1148                                              EXTENT_LOCKED | EXTENT_DELALLOC |
1149                                              EXTENT_DO_ACCOUNTING |
1150                                              EXTENT_DEFRAG, PAGE_UNLOCK |
1151                                              PAGE_CLEAR_DIRTY |
1152                                              PAGE_SET_WRITEBACK |
1153                                              PAGE_END_WRITEBACK);
1154                 return -ENOMEM;
1155         }
1156
1157         nolock = btrfs_is_free_space_inode(inode);
1158
1159         if (nolock)
1160                 trans = btrfs_join_transaction_nolock(root);
1161         else
1162                 trans = btrfs_join_transaction(root);
1163
1164         if (IS_ERR(trans)) {
1165                 extent_clear_unlock_delalloc(inode, start, end, locked_page,
1166                                              EXTENT_LOCKED | EXTENT_DELALLOC |
1167                                              EXTENT_DO_ACCOUNTING |
1168                                              EXTENT_DEFRAG, PAGE_UNLOCK |
1169                                              PAGE_CLEAR_DIRTY |
1170                                              PAGE_SET_WRITEBACK |
1171                                              PAGE_END_WRITEBACK);
1172                 btrfs_free_path(path);
1173                 return PTR_ERR(trans);
1174         }
1175
1176         trans->block_rsv = &root->fs_info->delalloc_block_rsv;
1177
1178         cow_start = (u64)-1;
1179         cur_offset = start;
1180         while (1) {
1181                 ret = btrfs_lookup_file_extent(trans, root, path, ino,
1182                                                cur_offset, 0);
1183                 if (ret < 0)
1184                         goto error;
1185                 if (ret > 0 && path->slots[0] > 0 && check_prev) {
1186                         leaf = path->nodes[0];
1187                         btrfs_item_key_to_cpu(leaf, &found_key,
1188                                               path->slots[0] - 1);
1189                         if (found_key.objectid == ino &&
1190                             found_key.type == BTRFS_EXTENT_DATA_KEY)
1191                                 path->slots[0]--;
1192                 }
1193                 check_prev = 0;
1194 next_slot:
1195                 leaf = path->nodes[0];
1196                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
1197                         ret = btrfs_next_leaf(root, path);
1198                         if (ret < 0)
1199                                 goto error;
1200                         if (ret > 0)
1201                                 break;
1202                         leaf = path->nodes[0];
1203                 }
1204
1205                 nocow = 0;
1206                 disk_bytenr = 0;
1207                 num_bytes = 0;
1208                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
1209
1210                 if (found_key.objectid > ino ||
1211                     found_key.type > BTRFS_EXTENT_DATA_KEY ||
1212                     found_key.offset > end)
1213                         break;
1214
1215                 if (found_key.offset > cur_offset) {
1216                         extent_end = found_key.offset;
1217                         extent_type = 0;
1218                         goto out_check;
1219                 }
1220
1221                 fi = btrfs_item_ptr(leaf, path->slots[0],
1222                                     struct btrfs_file_extent_item);
1223                 extent_type = btrfs_file_extent_type(leaf, fi);
1224
1225                 ram_bytes = btrfs_file_extent_ram_bytes(leaf, fi);
1226                 if (extent_type == BTRFS_FILE_EXTENT_REG ||
1227                     extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
1228                         disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1229                         extent_offset = btrfs_file_extent_offset(leaf, fi);
1230                         extent_end = found_key.offset +
1231                                 btrfs_file_extent_num_bytes(leaf, fi);
1232                         disk_num_bytes =
1233                                 btrfs_file_extent_disk_num_bytes(leaf, fi);
1234                         if (extent_end <= start) {
1235                                 path->slots[0]++;
1236                                 goto next_slot;
1237                         }
1238                         if (disk_bytenr == 0)
1239                                 goto out_check;
1240                         if (btrfs_file_extent_compression(leaf, fi) ||
1241                             btrfs_file_extent_encryption(leaf, fi) ||
1242                             btrfs_file_extent_other_encoding(leaf, fi))
1243                                 goto out_check;
1244                         if (extent_type == BTRFS_FILE_EXTENT_REG && !force)
1245                                 goto out_check;
1246                         if (btrfs_extent_readonly(root, disk_bytenr))
1247                                 goto out_check;
1248                         if (btrfs_cross_ref_exist(trans, root, ino,
1249                                                   found_key.offset -
1250                                                   extent_offset, disk_bytenr))
1251                                 goto out_check;
1252                         disk_bytenr += extent_offset;
1253                         disk_bytenr += cur_offset - found_key.offset;
1254                         num_bytes = min(end + 1, extent_end) - cur_offset;
1255                         /*
1256                          * force cow if csum exists in the range.
1257                          * this ensure that csum for a given extent are
1258                          * either valid or do not exist.
1259                          */
1260                         if (csum_exist_in_range(root, disk_bytenr, num_bytes))
1261                                 goto out_check;
1262                         nocow = 1;
1263                 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
1264                         extent_end = found_key.offset +
1265                                 btrfs_file_extent_inline_len(leaf, fi);
1266                         extent_end = ALIGN(extent_end, root->sectorsize);
1267                 } else {
1268                         BUG_ON(1);
1269                 }
1270 out_check:
1271                 if (extent_end <= start) {
1272                         path->slots[0]++;
1273                         goto next_slot;
1274                 }
1275                 if (!nocow) {
1276                         if (cow_start == (u64)-1)
1277                                 cow_start = cur_offset;
1278                         cur_offset = extent_end;
1279                         if (cur_offset > end)
1280                                 break;
1281                         path->slots[0]++;
1282                         goto next_slot;
1283                 }
1284
1285                 btrfs_release_path(path);
1286                 if (cow_start != (u64)-1) {
1287                         ret = cow_file_range(inode, locked_page,
1288                                              cow_start, found_key.offset - 1,
1289                                              page_started, nr_written, 1);
1290                         if (ret)
1291                                 goto error;
1292                         cow_start = (u64)-1;
1293                 }
1294
1295                 if (extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
1296                         struct extent_map *em;
1297                         struct extent_map_tree *em_tree;
1298                         em_tree = &BTRFS_I(inode)->extent_tree;
1299                         em = alloc_extent_map();
1300                         BUG_ON(!em); /* -ENOMEM */
1301                         em->start = cur_offset;
1302                         em->orig_start = found_key.offset - extent_offset;
1303                         em->len = num_bytes;
1304                         em->block_len = num_bytes;
1305                         em->block_start = disk_bytenr;
1306                         em->orig_block_len = disk_num_bytes;
1307                         em->ram_bytes = ram_bytes;
1308                         em->bdev = root->fs_info->fs_devices->latest_bdev;
1309                         em->mod_start = em->start;
1310                         em->mod_len = em->len;
1311                         set_bit(EXTENT_FLAG_PINNED, &em->flags);
1312                         set_bit(EXTENT_FLAG_FILLING, &em->flags);
1313                         em->generation = -1;
1314                         while (1) {
1315                                 write_lock(&em_tree->lock);
1316                                 ret = add_extent_mapping(em_tree, em, 1);
1317                                 write_unlock(&em_tree->lock);
1318                                 if (ret != -EEXIST) {
1319                                         free_extent_map(em);
1320                                         break;
1321                                 }
1322                                 btrfs_drop_extent_cache(inode, em->start,
1323                                                 em->start + em->len - 1, 0);
1324                         }
1325                         type = BTRFS_ORDERED_PREALLOC;
1326                 } else {
1327                         type = BTRFS_ORDERED_NOCOW;
1328                 }
1329
1330                 ret = btrfs_add_ordered_extent(inode, cur_offset, disk_bytenr,
1331                                                num_bytes, num_bytes, type);
1332                 BUG_ON(ret); /* -ENOMEM */
1333
1334                 if (root->root_key.objectid ==
1335                     BTRFS_DATA_RELOC_TREE_OBJECTID) {
1336                         ret = btrfs_reloc_clone_csums(inode, cur_offset,
1337                                                       num_bytes);
1338                         if (ret)
1339                                 goto error;
1340                 }
1341
1342                 extent_clear_unlock_delalloc(inode, cur_offset,
1343                                              cur_offset + num_bytes - 1,
1344                                              locked_page, EXTENT_LOCKED |
1345                                              EXTENT_DELALLOC, PAGE_UNLOCK |
1346                                              PAGE_SET_PRIVATE2);
1347                 cur_offset = extent_end;
1348                 if (cur_offset > end)
1349                         break;
1350         }
1351         btrfs_release_path(path);
1352
1353         if (cur_offset <= end && cow_start == (u64)-1) {
1354                 cow_start = cur_offset;
1355                 cur_offset = end;
1356         }
1357
1358         if (cow_start != (u64)-1) {
1359                 ret = cow_file_range(inode, locked_page, cow_start, end,
1360                                      page_started, nr_written, 1);
1361                 if (ret)
1362                         goto error;
1363         }
1364
1365 error:
1366         err = btrfs_end_transaction(trans, root);
1367         if (!ret)
1368                 ret = err;
1369
1370         if (ret && cur_offset < end)
1371                 extent_clear_unlock_delalloc(inode, cur_offset, end,
1372                                              locked_page, EXTENT_LOCKED |
1373                                              EXTENT_DELALLOC | EXTENT_DEFRAG |
1374                                              EXTENT_DO_ACCOUNTING, PAGE_UNLOCK |
1375                                              PAGE_CLEAR_DIRTY |
1376                                              PAGE_SET_WRITEBACK |
1377                                              PAGE_END_WRITEBACK);
1378         btrfs_free_path(path);
1379         return ret;
1380 }
1381
1382 /*
1383  * extent_io.c call back to do delayed allocation processing
1384  */
1385 static int run_delalloc_range(struct inode *inode, struct page *locked_page,
1386                               u64 start, u64 end, int *page_started,
1387                               unsigned long *nr_written)
1388 {
1389         int ret;
1390         struct btrfs_root *root = BTRFS_I(inode)->root;
1391
1392         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW) {
1393                 ret = run_delalloc_nocow(inode, locked_page, start, end,
1394                                          page_started, 1, nr_written);
1395         } else if (BTRFS_I(inode)->flags & BTRFS_INODE_PREALLOC) {
1396                 ret = run_delalloc_nocow(inode, locked_page, start, end,
1397                                          page_started, 0, nr_written);
1398         } else if (!btrfs_test_opt(root, COMPRESS) &&
1399                    !(BTRFS_I(inode)->force_compress) &&
1400                    !(BTRFS_I(inode)->flags & BTRFS_INODE_COMPRESS)) {
1401                 ret = cow_file_range(inode, locked_page, start, end,
1402                                       page_started, nr_written, 1);
1403         } else {
1404                 set_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
1405                         &BTRFS_I(inode)->runtime_flags);
1406                 ret = cow_file_range_async(inode, locked_page, start, end,
1407                                            page_started, nr_written);
1408         }
1409         return ret;
1410 }
1411
1412 static void btrfs_split_extent_hook(struct inode *inode,
1413                                     struct extent_state *orig, u64 split)
1414 {
1415         /* not delalloc, ignore it */
1416         if (!(orig->state & EXTENT_DELALLOC))
1417                 return;
1418
1419         spin_lock(&BTRFS_I(inode)->lock);
1420         BTRFS_I(inode)->outstanding_extents++;
1421         spin_unlock(&BTRFS_I(inode)->lock);
1422 }
1423
1424 /*
1425  * extent_io.c merge_extent_hook, used to track merged delayed allocation
1426  * extents so we can keep track of new extents that are just merged onto old
1427  * extents, such as when we are doing sequential writes, so we can properly
1428  * account for the metadata space we'll need.
1429  */
1430 static void btrfs_merge_extent_hook(struct inode *inode,
1431                                     struct extent_state *new,
1432                                     struct extent_state *other)
1433 {
1434         /* not delalloc, ignore it */
1435         if (!(other->state & EXTENT_DELALLOC))
1436                 return;
1437
1438         spin_lock(&BTRFS_I(inode)->lock);
1439         BTRFS_I(inode)->outstanding_extents--;
1440         spin_unlock(&BTRFS_I(inode)->lock);
1441 }
1442
1443 static void btrfs_add_delalloc_inodes(struct btrfs_root *root,
1444                                       struct inode *inode)
1445 {
1446         spin_lock(&root->delalloc_lock);
1447         if (list_empty(&BTRFS_I(inode)->delalloc_inodes)) {
1448                 list_add_tail(&BTRFS_I(inode)->delalloc_inodes,
1449                               &root->delalloc_inodes);
1450                 set_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1451                         &BTRFS_I(inode)->runtime_flags);
1452                 root->nr_delalloc_inodes++;
1453                 if (root->nr_delalloc_inodes == 1) {
1454                         spin_lock(&root->fs_info->delalloc_root_lock);
1455                         BUG_ON(!list_empty(&root->delalloc_root));
1456                         list_add_tail(&root->delalloc_root,
1457                                       &root->fs_info->delalloc_roots);
1458                         spin_unlock(&root->fs_info->delalloc_root_lock);
1459                 }
1460         }
1461         spin_unlock(&root->delalloc_lock);
1462 }
1463
1464 static void btrfs_del_delalloc_inode(struct btrfs_root *root,
1465                                      struct inode *inode)
1466 {
1467         spin_lock(&root->delalloc_lock);
1468         if (!list_empty(&BTRFS_I(inode)->delalloc_inodes)) {
1469                 list_del_init(&BTRFS_I(inode)->delalloc_inodes);
1470                 clear_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1471                           &BTRFS_I(inode)->runtime_flags);
1472                 root->nr_delalloc_inodes--;
1473                 if (!root->nr_delalloc_inodes) {
1474                         spin_lock(&root->fs_info->delalloc_root_lock);
1475                         BUG_ON(list_empty(&root->delalloc_root));
1476                         list_del_init(&root->delalloc_root);
1477                         spin_unlock(&root->fs_info->delalloc_root_lock);
1478                 }
1479         }
1480         spin_unlock(&root->delalloc_lock);
1481 }
1482
1483 /*
1484  * extent_io.c set_bit_hook, used to track delayed allocation
1485  * bytes in this file, and to maintain the list of inodes that
1486  * have pending delalloc work to be done.
1487  */
1488 static void btrfs_set_bit_hook(struct inode *inode,
1489                                struct extent_state *state, unsigned long *bits)
1490 {
1491
1492         /*
1493          * set_bit and clear bit hooks normally require _irqsave/restore
1494          * but in this case, we are only testing for the DELALLOC
1495          * bit, which is only set or cleared with irqs on
1496          */
1497         if (!(state->state & EXTENT_DELALLOC) && (*bits & EXTENT_DELALLOC)) {
1498                 struct btrfs_root *root = BTRFS_I(inode)->root;
1499                 u64 len = state->end + 1 - state->start;
1500                 bool do_list = !btrfs_is_free_space_inode(inode);
1501
1502                 if (*bits & EXTENT_FIRST_DELALLOC) {
1503                         *bits &= ~EXTENT_FIRST_DELALLOC;
1504                 } else {
1505                         spin_lock(&BTRFS_I(inode)->lock);
1506                         BTRFS_I(inode)->outstanding_extents++;
1507                         spin_unlock(&BTRFS_I(inode)->lock);
1508                 }
1509
1510                 __percpu_counter_add(&root->fs_info->delalloc_bytes, len,
1511                                      root->fs_info->delalloc_batch);
1512                 spin_lock(&BTRFS_I(inode)->lock);
1513                 BTRFS_I(inode)->delalloc_bytes += len;
1514                 if (do_list && !test_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1515                                          &BTRFS_I(inode)->runtime_flags))
1516                         btrfs_add_delalloc_inodes(root, inode);
1517                 spin_unlock(&BTRFS_I(inode)->lock);
1518         }
1519 }
1520
1521 /*
1522  * extent_io.c clear_bit_hook, see set_bit_hook for why
1523  */
1524 static void btrfs_clear_bit_hook(struct inode *inode,
1525                                  struct extent_state *state,
1526                                  unsigned long *bits)
1527 {
1528         /*
1529          * set_bit and clear bit hooks normally require _irqsave/restore
1530          * but in this case, we are only testing for the DELALLOC
1531          * bit, which is only set or cleared with irqs on
1532          */
1533         if ((state->state & EXTENT_DELALLOC) && (*bits & EXTENT_DELALLOC)) {
1534                 struct btrfs_root *root = BTRFS_I(inode)->root;
1535                 u64 len = state->end + 1 - state->start;
1536                 bool do_list = !btrfs_is_free_space_inode(inode);
1537
1538                 if (*bits & EXTENT_FIRST_DELALLOC) {
1539                         *bits &= ~EXTENT_FIRST_DELALLOC;
1540                 } else if (!(*bits & EXTENT_DO_ACCOUNTING)) {
1541                         spin_lock(&BTRFS_I(inode)->lock);
1542                         BTRFS_I(inode)->outstanding_extents--;
1543                         spin_unlock(&BTRFS_I(inode)->lock);
1544                 }
1545
1546                 /*
1547                  * We don't reserve metadata space for space cache inodes so we
1548                  * don't need to call dellalloc_release_metadata if there is an
1549                  * error.
1550                  */
1551                 if (*bits & EXTENT_DO_ACCOUNTING &&
1552                     root != root->fs_info->tree_root)
1553                         btrfs_delalloc_release_metadata(inode, len);
1554
1555                 if (root->root_key.objectid != BTRFS_DATA_RELOC_TREE_OBJECTID
1556                     && do_list && !(state->state & EXTENT_NORESERVE))
1557                         btrfs_free_reserved_data_space(inode, len);
1558
1559                 __percpu_counter_add(&root->fs_info->delalloc_bytes, -len,
1560                                      root->fs_info->delalloc_batch);
1561                 spin_lock(&BTRFS_I(inode)->lock);
1562                 BTRFS_I(inode)->delalloc_bytes -= len;
1563                 if (do_list && BTRFS_I(inode)->delalloc_bytes == 0 &&
1564                     test_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1565                              &BTRFS_I(inode)->runtime_flags))
1566                         btrfs_del_delalloc_inode(root, inode);
1567                 spin_unlock(&BTRFS_I(inode)->lock);
1568         }
1569 }
1570
1571 /*
1572  * extent_io.c merge_bio_hook, this must check the chunk tree to make sure
1573  * we don't create bios that span stripes or chunks
1574  */
1575 int btrfs_merge_bio_hook(int rw, struct page *page, unsigned long offset,
1576                          size_t size, struct bio *bio,
1577                          unsigned long bio_flags)
1578 {
1579         struct btrfs_root *root = BTRFS_I(page->mapping->host)->root;
1580         u64 logical = (u64)bio->bi_sector << 9;
1581         u64 length = 0;
1582         u64 map_length;
1583         int ret;
1584
1585         if (bio_flags & EXTENT_BIO_COMPRESSED)
1586                 return 0;
1587
1588         length = bio->bi_size;
1589         map_length = length;
1590         ret = btrfs_map_block(root->fs_info, rw, logical,
1591                               &map_length, NULL, 0);
1592         /* Will always return 0 with map_multi == NULL */
1593         BUG_ON(ret < 0);
1594         if (map_length < length + size)
1595                 return 1;
1596         return 0;
1597 }
1598
1599 /*
1600  * in order to insert checksums into the metadata in large chunks,
1601  * we wait until bio submission time.   All the pages in the bio are
1602  * checksummed and sums are attached onto the ordered extent record.
1603  *
1604  * At IO completion time the cums attached on the ordered extent record
1605  * are inserted into the btree
1606  */
1607 static int __btrfs_submit_bio_start(struct inode *inode, int rw,
1608                                     struct bio *bio, int mirror_num,
1609                                     unsigned long bio_flags,
1610                                     u64 bio_offset)
1611 {
1612         struct btrfs_root *root = BTRFS_I(inode)->root;
1613         int ret = 0;
1614
1615         ret = btrfs_csum_one_bio(root, inode, bio, 0, 0);
1616         BUG_ON(ret); /* -ENOMEM */
1617         return 0;
1618 }
1619
1620 /*
1621  * in order to insert checksums into the metadata in large chunks,
1622  * we wait until bio submission time.   All the pages in the bio are
1623  * checksummed and sums are attached onto the ordered extent record.
1624  *
1625  * At IO completion time the cums attached on the ordered extent record
1626  * are inserted into the btree
1627  */
1628 static int __btrfs_submit_bio_done(struct inode *inode, int rw, struct bio *bio,
1629                           int mirror_num, unsigned long bio_flags,
1630                           u64 bio_offset)
1631 {
1632         struct btrfs_root *root = BTRFS_I(inode)->root;
1633         int ret;
1634
1635         ret = btrfs_map_bio(root, rw, bio, mirror_num, 1);
1636         if (ret)
1637                 bio_endio(bio, ret);
1638         return ret;
1639 }
1640
1641 /*
1642  * extent_io.c submission hook. This does the right thing for csum calculation
1643  * on write, or reading the csums from the tree before a read
1644  */
1645 static int btrfs_submit_bio_hook(struct inode *inode, int rw, struct bio *bio,
1646                           int mirror_num, unsigned long bio_flags,
1647                           u64 bio_offset)
1648 {
1649         struct btrfs_root *root = BTRFS_I(inode)->root;
1650         int ret = 0;
1651         int skip_sum;
1652         int metadata = 0;
1653         int async = !atomic_read(&BTRFS_I(inode)->sync_writers);
1654
1655         skip_sum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM;
1656
1657         if (btrfs_is_free_space_inode(inode))
1658                 metadata = 2;
1659
1660         if (!(rw & REQ_WRITE)) {
1661                 ret = btrfs_bio_wq_end_io(root->fs_info, bio, metadata);
1662                 if (ret)
1663                         goto out;
1664
1665                 if (bio_flags & EXTENT_BIO_COMPRESSED) {
1666                         ret = btrfs_submit_compressed_read(inode, bio,
1667                                                            mirror_num,
1668                                                            bio_flags);
1669                         goto out;
1670                 } else if (!skip_sum) {
1671                         ret = btrfs_lookup_bio_sums(root, inode, bio, NULL);
1672                         if (ret)
1673                                 goto out;
1674                 }
1675                 goto mapit;
1676         } else if (async && !skip_sum) {
1677                 /* csum items have already been cloned */
1678                 if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
1679                         goto mapit;
1680                 /* we're doing a write, do the async checksumming */
1681                 ret = btrfs_wq_submit_bio(BTRFS_I(inode)->root->fs_info,
1682                                    inode, rw, bio, mirror_num,
1683                                    bio_flags, bio_offset,
1684                                    __btrfs_submit_bio_start,
1685                                    __btrfs_submit_bio_done);
1686                 goto out;
1687         } else if (!skip_sum) {
1688                 ret = btrfs_csum_one_bio(root, inode, bio, 0, 0);
1689                 if (ret)
1690                         goto out;
1691         }
1692
1693 mapit:
1694         ret = btrfs_map_bio(root, rw, bio, mirror_num, 0);
1695
1696 out:
1697         if (ret < 0)
1698                 bio_endio(bio, ret);
1699         return ret;
1700 }
1701
1702 /*
1703  * given a list of ordered sums record them in the inode.  This happens
1704  * at IO completion time based on sums calculated at bio submission time.
1705  */
1706 static noinline int add_pending_csums(struct btrfs_trans_handle *trans,
1707                              struct inode *inode, u64 file_offset,
1708                              struct list_head *list)
1709 {
1710         struct btrfs_ordered_sum *sum;
1711
1712         list_for_each_entry(sum, list, list) {
1713                 trans->adding_csums = 1;
1714                 btrfs_csum_file_blocks(trans,
1715                        BTRFS_I(inode)->root->fs_info->csum_root, sum);
1716                 trans->adding_csums = 0;
1717         }
1718         return 0;
1719 }
1720
1721 int btrfs_set_extent_delalloc(struct inode *inode, u64 start, u64 end,
1722                               struct extent_state **cached_state)
1723 {
1724         WARN_ON((end & (PAGE_CACHE_SIZE - 1)) == 0);
1725         return set_extent_delalloc(&BTRFS_I(inode)->io_tree, start, end,
1726                                    cached_state, GFP_NOFS);
1727 }
1728
1729 /* see btrfs_writepage_start_hook for details on why this is required */
1730 struct btrfs_writepage_fixup {
1731         struct page *page;
1732         struct btrfs_work work;
1733 };
1734
1735 static void btrfs_writepage_fixup_worker(struct btrfs_work *work)
1736 {
1737         struct btrfs_writepage_fixup *fixup;
1738         struct btrfs_ordered_extent *ordered;
1739         struct extent_state *cached_state = NULL;
1740         struct page *page;
1741         struct inode *inode;
1742         u64 page_start;
1743         u64 page_end;
1744         int ret;
1745
1746         fixup = container_of(work, struct btrfs_writepage_fixup, work);
1747         page = fixup->page;
1748 again:
1749         lock_page(page);
1750         if (!page->mapping || !PageDirty(page) || !PageChecked(page)) {
1751                 ClearPageChecked(page);
1752                 goto out_page;
1753         }
1754
1755         inode = page->mapping->host;
1756         page_start = page_offset(page);
1757         page_end = page_offset(page) + PAGE_CACHE_SIZE - 1;
1758
1759         lock_extent_bits(&BTRFS_I(inode)->io_tree, page_start, page_end, 0,
1760                          &cached_state);
1761
1762         /* already ordered? We're done */
1763         if (PagePrivate2(page))
1764                 goto out;
1765
1766         ordered = btrfs_lookup_ordered_extent(inode, page_start);
1767         if (ordered) {
1768                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, page_start,
1769                                      page_end, &cached_state, GFP_NOFS);
1770                 unlock_page(page);
1771                 btrfs_start_ordered_extent(inode, ordered, 1);
1772                 btrfs_put_ordered_extent(ordered);
1773                 goto again;
1774         }
1775
1776         ret = btrfs_delalloc_reserve_space(inode, PAGE_CACHE_SIZE);
1777         if (ret) {
1778                 mapping_set_error(page->mapping, ret);
1779                 end_extent_writepage(page, ret, page_start, page_end);
1780                 ClearPageChecked(page);
1781                 goto out;
1782          }
1783
1784         btrfs_set_extent_delalloc(inode, page_start, page_end, &cached_state);
1785         ClearPageChecked(page);
1786         set_page_dirty(page);
1787 out:
1788         unlock_extent_cached(&BTRFS_I(inode)->io_tree, page_start, page_end,
1789                              &cached_state, GFP_NOFS);
1790 out_page:
1791         unlock_page(page);
1792         page_cache_release(page);
1793         kfree(fixup);
1794 }
1795
1796 /*
1797  * There are a few paths in the higher layers of the kernel that directly
1798  * set the page dirty bit without asking the filesystem if it is a
1799  * good idea.  This causes problems because we want to make sure COW
1800  * properly happens and the data=ordered rules are followed.
1801  *
1802  * In our case any range that doesn't have the ORDERED bit set
1803  * hasn't been properly setup for IO.  We kick off an async process
1804  * to fix it up.  The async helper will wait for ordered extents, set
1805  * the delalloc bit and make it safe to write the page.
1806  */
1807 static int btrfs_writepage_start_hook(struct page *page, u64 start, u64 end)
1808 {
1809         struct inode *inode = page->mapping->host;
1810         struct btrfs_writepage_fixup *fixup;
1811         struct btrfs_root *root = BTRFS_I(inode)->root;
1812
1813         /* this page is properly in the ordered list */
1814         if (TestClearPagePrivate2(page))
1815                 return 0;
1816
1817         if (PageChecked(page))
1818                 return -EAGAIN;
1819
1820         fixup = kzalloc(sizeof(*fixup), GFP_NOFS);
1821         if (!fixup)
1822                 return -EAGAIN;
1823
1824         SetPageChecked(page);
1825         page_cache_get(page);
1826         fixup->work.func = btrfs_writepage_fixup_worker;
1827         fixup->page = page;
1828         btrfs_queue_worker(&root->fs_info->fixup_workers, &fixup->work);
1829         return -EBUSY;
1830 }
1831
1832 static int insert_reserved_file_extent(struct btrfs_trans_handle *trans,
1833                                        struct inode *inode, u64 file_pos,
1834                                        u64 disk_bytenr, u64 disk_num_bytes,
1835                                        u64 num_bytes, u64 ram_bytes,
1836                                        u8 compression, u8 encryption,
1837                                        u16 other_encoding, int extent_type)
1838 {
1839         struct btrfs_root *root = BTRFS_I(inode)->root;
1840         struct btrfs_file_extent_item *fi;
1841         struct btrfs_path *path;
1842         struct extent_buffer *leaf;
1843         struct btrfs_key ins;
1844         int ret;
1845
1846         path = btrfs_alloc_path();
1847         if (!path)
1848                 return -ENOMEM;
1849
1850         path->leave_spinning = 1;
1851
1852         /*
1853          * we may be replacing one extent in the tree with another.
1854          * The new extent is pinned in the extent map, and we don't want
1855          * to drop it from the cache until it is completely in the btree.
1856          *
1857          * So, tell btrfs_drop_extents to leave this extent in the cache.
1858          * the caller is expected to unpin it and allow it to be merged
1859          * with the others.
1860          */
1861         ret = btrfs_drop_extents(trans, root, inode, file_pos,
1862                                  file_pos + num_bytes, 0);
1863         if (ret)
1864                 goto out;
1865
1866         ins.objectid = btrfs_ino(inode);
1867         ins.offset = file_pos;
1868         ins.type = BTRFS_EXTENT_DATA_KEY;
1869         ret = btrfs_insert_empty_item(trans, root, path, &ins, sizeof(*fi));
1870         if (ret)
1871                 goto out;
1872         leaf = path->nodes[0];
1873         fi = btrfs_item_ptr(leaf, path->slots[0],
1874                             struct btrfs_file_extent_item);
1875         btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1876         btrfs_set_file_extent_type(leaf, fi, extent_type);
1877         btrfs_set_file_extent_disk_bytenr(leaf, fi, disk_bytenr);
1878         btrfs_set_file_extent_disk_num_bytes(leaf, fi, disk_num_bytes);
1879         btrfs_set_file_extent_offset(leaf, fi, 0);
1880         btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
1881         btrfs_set_file_extent_ram_bytes(leaf, fi, ram_bytes);
1882         btrfs_set_file_extent_compression(leaf, fi, compression);
1883         btrfs_set_file_extent_encryption(leaf, fi, encryption);
1884         btrfs_set_file_extent_other_encoding(leaf, fi, other_encoding);
1885
1886         btrfs_mark_buffer_dirty(leaf);
1887         btrfs_release_path(path);
1888
1889         inode_add_bytes(inode, num_bytes);
1890
1891         ins.objectid = disk_bytenr;
1892         ins.offset = disk_num_bytes;
1893         ins.type = BTRFS_EXTENT_ITEM_KEY;
1894         ret = btrfs_alloc_reserved_file_extent(trans, root,
1895                                         root->root_key.objectid,
1896                                         btrfs_ino(inode), file_pos, &ins);
1897 out:
1898         btrfs_free_path(path);
1899
1900         return ret;
1901 }
1902
1903 /* snapshot-aware defrag */
1904 struct sa_defrag_extent_backref {
1905         struct rb_node node;
1906         struct old_sa_defrag_extent *old;
1907         u64 root_id;
1908         u64 inum;
1909         u64 file_pos;
1910         u64 extent_offset;
1911         u64 num_bytes;
1912         u64 generation;
1913 };
1914
1915 struct old_sa_defrag_extent {
1916         struct list_head list;
1917         struct new_sa_defrag_extent *new;
1918
1919         u64 extent_offset;
1920         u64 bytenr;
1921         u64 offset;
1922         u64 len;
1923         int count;
1924 };
1925
1926 struct new_sa_defrag_extent {
1927         struct rb_root root;
1928         struct list_head head;
1929         struct btrfs_path *path;
1930         struct inode *inode;
1931         u64 file_pos;
1932         u64 len;
1933         u64 bytenr;
1934         u64 disk_len;
1935         u8 compress_type;
1936 };
1937
1938 static int backref_comp(struct sa_defrag_extent_backref *b1,
1939                         struct sa_defrag_extent_backref *b2)
1940 {
1941         if (b1->root_id < b2->root_id)
1942                 return -1;
1943         else if (b1->root_id > b2->root_id)
1944                 return 1;
1945
1946         if (b1->inum < b2->inum)
1947                 return -1;
1948         else if (b1->inum > b2->inum)
1949                 return 1;
1950
1951         if (b1->file_pos < b2->file_pos)
1952                 return -1;
1953         else if (b1->file_pos > b2->file_pos)
1954                 return 1;
1955
1956         /*
1957          * [------------------------------] ===> (a range of space)
1958          *     |<--->|   |<---->| =============> (fs/file tree A)
1959          * |<---------------------------->| ===> (fs/file tree B)
1960          *
1961          * A range of space can refer to two file extents in one tree while
1962          * refer to only one file extent in another tree.
1963          *
1964          * So we may process a disk offset more than one time(two extents in A)
1965          * and locate at the same extent(one extent in B), then insert two same
1966          * backrefs(both refer to the extent in B).
1967          */
1968         return 0;
1969 }
1970
1971 static void backref_insert(struct rb_root *root,
1972                            struct sa_defrag_extent_backref *backref)
1973 {
1974         struct rb_node **p = &root->rb_node;
1975         struct rb_node *parent = NULL;
1976         struct sa_defrag_extent_backref *entry;
1977         int ret;
1978
1979         while (*p) {
1980                 parent = *p;
1981                 entry = rb_entry(parent, struct sa_defrag_extent_backref, node);
1982
1983                 ret = backref_comp(backref, entry);
1984                 if (ret < 0)
1985                         p = &(*p)->rb_left;
1986                 else
1987                         p = &(*p)->rb_right;
1988         }
1989
1990         rb_link_node(&backref->node, parent, p);
1991         rb_insert_color(&backref->node, root);
1992 }
1993
1994 /*
1995  * Note the backref might has changed, and in this case we just return 0.
1996  */
1997 static noinline int record_one_backref(u64 inum, u64 offset, u64 root_id,
1998                                        void *ctx)
1999 {
2000         struct btrfs_file_extent_item *extent;
2001         struct btrfs_fs_info *fs_info;
2002         struct old_sa_defrag_extent *old = ctx;
2003         struct new_sa_defrag_extent *new = old->new;
2004         struct btrfs_path *path = new->path;
2005         struct btrfs_key key;
2006         struct btrfs_root *root;
2007         struct sa_defrag_extent_backref *backref;
2008         struct extent_buffer *leaf;
2009         struct inode *inode = new->inode;
2010         int slot;
2011         int ret;
2012         u64 extent_offset;
2013         u64 num_bytes;
2014
2015         if (BTRFS_I(inode)->root->root_key.objectid == root_id &&
2016             inum == btrfs_ino(inode))
2017                 return 0;
2018
2019         key.objectid = root_id;
2020         key.type = BTRFS_ROOT_ITEM_KEY;
2021         key.offset = (u64)-1;
2022
2023         fs_info = BTRFS_I(inode)->root->fs_info;
2024         root = btrfs_read_fs_root_no_name(fs_info, &key);
2025         if (IS_ERR(root)) {
2026                 if (PTR_ERR(root) == -ENOENT)
2027                         return 0;
2028                 WARN_ON(1);
2029                 pr_debug("inum=%llu, offset=%llu, root_id=%llu\n",
2030                          inum, offset, root_id);
2031                 return PTR_ERR(root);
2032         }
2033
2034         key.objectid = inum;
2035         key.type = BTRFS_EXTENT_DATA_KEY;
2036         if (offset > (u64)-1 << 32)
2037                 key.offset = 0;
2038         else
2039                 key.offset = offset;
2040
2041         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2042         if (WARN_ON(ret < 0))
2043                 return ret;
2044         ret = 0;
2045
2046         while (1) {
2047                 cond_resched();
2048
2049                 leaf = path->nodes[0];
2050                 slot = path->slots[0];
2051
2052                 if (slot >= btrfs_header_nritems(leaf)) {
2053                         ret = btrfs_next_leaf(root, path);
2054                         if (ret < 0) {
2055                                 goto out;
2056                         } else if (ret > 0) {
2057                                 ret = 0;
2058                                 goto out;
2059                         }
2060                         continue;
2061                 }
2062
2063                 path->slots[0]++;
2064
2065                 btrfs_item_key_to_cpu(leaf, &key, slot);
2066
2067                 if (key.objectid > inum)
2068                         goto out;
2069
2070                 if (key.objectid < inum || key.type != BTRFS_EXTENT_DATA_KEY)
2071                         continue;
2072
2073                 extent = btrfs_item_ptr(leaf, slot,
2074                                         struct btrfs_file_extent_item);
2075
2076                 if (btrfs_file_extent_disk_bytenr(leaf, extent) != old->bytenr)
2077                         continue;
2078
2079                 /*
2080                  * 'offset' refers to the exact key.offset,
2081                  * NOT the 'offset' field in btrfs_extent_data_ref, ie.
2082                  * (key.offset - extent_offset).
2083                  */
2084                 if (key.offset != offset)
2085                         continue;
2086
2087                 extent_offset = btrfs_file_extent_offset(leaf, extent);
2088                 num_bytes = btrfs_file_extent_num_bytes(leaf, extent);
2089
2090                 if (extent_offset >= old->extent_offset + old->offset +
2091                     old->len || extent_offset + num_bytes <=
2092                     old->extent_offset + old->offset)
2093                         continue;
2094                 break;
2095         }
2096
2097         backref = kmalloc(sizeof(*backref), GFP_NOFS);
2098         if (!backref) {
2099                 ret = -ENOENT;
2100                 goto out;
2101         }
2102
2103         backref->root_id = root_id;
2104         backref->inum = inum;
2105         backref->file_pos = offset;
2106         backref->num_bytes = num_bytes;
2107         backref->extent_offset = extent_offset;
2108         backref->generation = btrfs_file_extent_generation(leaf, extent);
2109         backref->old = old;
2110         backref_insert(&new->root, backref);
2111         old->count++;
2112 out:
2113         btrfs_release_path(path);
2114         WARN_ON(ret);
2115         return ret;
2116 }
2117
2118 static noinline bool record_extent_backrefs(struct btrfs_path *path,
2119                                    struct new_sa_defrag_extent *new)
2120 {
2121         struct btrfs_fs_info *fs_info = BTRFS_I(new->inode)->root->fs_info;
2122         struct old_sa_defrag_extent *old, *tmp;
2123         int ret;
2124
2125         new->path = path;
2126
2127         list_for_each_entry_safe(old, tmp, &new->head, list) {
2128                 ret = iterate_inodes_from_logical(old->bytenr +
2129                                                   old->extent_offset, fs_info,
2130                                                   path, record_one_backref,
2131                                                   old);
2132                 if (ret < 0 && ret != -ENOENT)
2133                         return false;
2134
2135                 /* no backref to be processed for this extent */
2136                 if (!old->count) {
2137                         list_del(&old->list);
2138                         kfree(old);
2139                 }
2140         }
2141
2142         if (list_empty(&new->head))
2143                 return false;
2144
2145         return true;
2146 }
2147
2148 static int relink_is_mergable(struct extent_buffer *leaf,
2149                               struct btrfs_file_extent_item *fi,
2150                               struct new_sa_defrag_extent *new)
2151 {
2152         if (btrfs_file_extent_disk_bytenr(leaf, fi) != new->bytenr)
2153                 return 0;
2154
2155         if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG)
2156                 return 0;
2157
2158         if (btrfs_file_extent_compression(leaf, fi) != new->compress_type)
2159                 return 0;
2160
2161         if (btrfs_file_extent_encryption(leaf, fi) ||
2162             btrfs_file_extent_other_encoding(leaf, fi))
2163                 return 0;
2164
2165         return 1;
2166 }
2167
2168 /*
2169  * Note the backref might has changed, and in this case we just return 0.
2170  */
2171 static noinline int relink_extent_backref(struct btrfs_path *path,
2172                                  struct sa_defrag_extent_backref *prev,
2173                                  struct sa_defrag_extent_backref *backref)
2174 {
2175         struct btrfs_file_extent_item *extent;
2176         struct btrfs_file_extent_item *item;
2177         struct btrfs_ordered_extent *ordered;
2178         struct btrfs_trans_handle *trans;
2179         struct btrfs_fs_info *fs_info;
2180         struct btrfs_root *root;
2181         struct btrfs_key key;
2182         struct extent_buffer *leaf;
2183         struct old_sa_defrag_extent *old = backref->old;
2184         struct new_sa_defrag_extent *new = old->new;
2185         struct inode *src_inode = new->inode;
2186         struct inode *inode;
2187         struct extent_state *cached = NULL;
2188         int ret = 0;
2189         u64 start;
2190         u64 len;
2191         u64 lock_start;
2192         u64 lock_end;
2193         bool merge = false;
2194         int index;
2195
2196         if (prev && prev->root_id == backref->root_id &&
2197             prev->inum == backref->inum &&
2198             prev->file_pos + prev->num_bytes == backref->file_pos)
2199                 merge = true;
2200
2201         /* step 1: get root */
2202         key.objectid = backref->root_id;
2203         key.type = BTRFS_ROOT_ITEM_KEY;
2204         key.offset = (u64)-1;
2205
2206         fs_info = BTRFS_I(src_inode)->root->fs_info;
2207         index = srcu_read_lock(&fs_info->subvol_srcu);
2208
2209         root = btrfs_read_fs_root_no_name(fs_info, &key);
2210         if (IS_ERR(root)) {
2211                 srcu_read_unlock(&fs_info->subvol_srcu, index);
2212                 if (PTR_ERR(root) == -ENOENT)
2213                         return 0;
2214                 return PTR_ERR(root);
2215         }
2216
2217         /* step 2: get inode */
2218         key.objectid = backref->inum;
2219         key.type = BTRFS_INODE_ITEM_KEY;
2220         key.offset = 0;
2221
2222         inode = btrfs_iget(fs_info->sb, &key, root, NULL);
2223         if (IS_ERR(inode)) {
2224                 srcu_read_unlock(&fs_info->subvol_srcu, index);
2225                 return 0;
2226         }
2227
2228         srcu_read_unlock(&fs_info->subvol_srcu, index);
2229
2230         /* step 3: relink backref */
2231         lock_start = backref->file_pos;
2232         lock_end = backref->file_pos + backref->num_bytes - 1;
2233         lock_extent_bits(&BTRFS_I(inode)->io_tree, lock_start, lock_end,
2234                          0, &cached);
2235
2236         ordered = btrfs_lookup_first_ordered_extent(inode, lock_end);
2237         if (ordered) {
2238                 btrfs_put_ordered_extent(ordered);
2239                 goto out_unlock;
2240         }
2241
2242         trans = btrfs_join_transaction(root);
2243         if (IS_ERR(trans)) {
2244                 ret = PTR_ERR(trans);
2245                 goto out_unlock;
2246         }
2247
2248         key.objectid = backref->inum;
2249         key.type = BTRFS_EXTENT_DATA_KEY;
2250         key.offset = backref->file_pos;
2251
2252         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2253         if (ret < 0) {
2254                 goto out_free_path;
2255         } else if (ret > 0) {
2256                 ret = 0;
2257                 goto out_free_path;
2258         }
2259
2260         extent = btrfs_item_ptr(path->nodes[0], path->slots[0],
2261                                 struct btrfs_file_extent_item);
2262
2263         if (btrfs_file_extent_generation(path->nodes[0], extent) !=
2264             backref->generation)
2265                 goto out_free_path;
2266
2267         btrfs_release_path(path);
2268
2269         start = backref->file_pos;
2270         if (backref->extent_offset < old->extent_offset + old->offset)
2271                 start += old->extent_offset + old->offset -
2272                          backref->extent_offset;
2273
2274         len = min(backref->extent_offset + backref->num_bytes,
2275                   old->extent_offset + old->offset + old->len);
2276         len -= max(backref->extent_offset, old->extent_offset + old->offset);
2277
2278         ret = btrfs_drop_extents(trans, root, inode, start,
2279                                  start + len, 1);
2280         if (ret)
2281                 goto out_free_path;
2282 again:
2283         key.objectid = btrfs_ino(inode);
2284         key.type = BTRFS_EXTENT_DATA_KEY;
2285         key.offset = start;
2286
2287         path->leave_spinning = 1;
2288         if (merge) {
2289                 struct btrfs_file_extent_item *fi;
2290                 u64 extent_len;
2291                 struct btrfs_key found_key;
2292
2293                 ret = btrfs_search_slot(trans, root, &key, path, 1, 1);
2294                 if (ret < 0)
2295                         goto out_free_path;
2296
2297                 path->slots[0]--;
2298                 leaf = path->nodes[0];
2299                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
2300
2301                 fi = btrfs_item_ptr(leaf, path->slots[0],
2302                                     struct btrfs_file_extent_item);
2303                 extent_len = btrfs_file_extent_num_bytes(leaf, fi);
2304
2305                 if (extent_len + found_key.offset == start &&
2306                     relink_is_mergable(leaf, fi, new)) {
2307                         btrfs_set_file_extent_num_bytes(leaf, fi,
2308                                                         extent_len + len);
2309                         btrfs_mark_buffer_dirty(leaf);
2310                         inode_add_bytes(inode, len);
2311
2312                         ret = 1;
2313                         goto out_free_path;
2314                 } else {
2315                         merge = false;
2316                         btrfs_release_path(path);
2317                         goto again;
2318                 }
2319         }
2320
2321         ret = btrfs_insert_empty_item(trans, root, path, &key,
2322                                         sizeof(*extent));
2323         if (ret) {
2324                 btrfs_abort_transaction(trans, root, ret);
2325                 goto out_free_path;
2326         }
2327
2328         leaf = path->nodes[0];
2329         item = btrfs_item_ptr(leaf, path->slots[0],
2330                                 struct btrfs_file_extent_item);
2331         btrfs_set_file_extent_disk_bytenr(leaf, item, new->bytenr);
2332         btrfs_set_file_extent_disk_num_bytes(leaf, item, new->disk_len);
2333         btrfs_set_file_extent_offset(leaf, item, start - new->file_pos);
2334         btrfs_set_file_extent_num_bytes(leaf, item, len);
2335         btrfs_set_file_extent_ram_bytes(leaf, item, new->len);
2336         btrfs_set_file_extent_generation(leaf, item, trans->transid);
2337         btrfs_set_file_extent_type(leaf, item, BTRFS_FILE_EXTENT_REG);
2338         btrfs_set_file_extent_compression(leaf, item, new->compress_type);
2339         btrfs_set_file_extent_encryption(leaf, item, 0);
2340         btrfs_set_file_extent_other_encoding(leaf, item, 0);
2341
2342         btrfs_mark_buffer_dirty(leaf);
2343         inode_add_bytes(inode, len);
2344         btrfs_release_path(path);
2345
2346         ret = btrfs_inc_extent_ref(trans, root, new->bytenr,
2347                         new->disk_len, 0,
2348                         backref->root_id, backref->inum,
2349                         new->file_pos, 0);      /* start - extent_offset */
2350         if (ret) {
2351                 btrfs_abort_transaction(trans, root, ret);
2352                 goto out_free_path;
2353         }
2354
2355         ret = 1;
2356 out_free_path:
2357         btrfs_release_path(path);
2358         path->leave_spinning = 0;
2359         btrfs_end_transaction(trans, root);
2360 out_unlock:
2361         unlock_extent_cached(&BTRFS_I(inode)->io_tree, lock_start, lock_end,
2362                              &cached, GFP_NOFS);
2363         iput(inode);
2364         return ret;
2365 }
2366
2367 static void free_sa_defrag_extent(struct new_sa_defrag_extent *new)
2368 {
2369         struct old_sa_defrag_extent *old, *tmp;
2370
2371         if (!new)
2372                 return;
2373
2374         list_for_each_entry_safe(old, tmp, &new->head, list) {
2375                 list_del(&old->list);
2376                 kfree(old);
2377         }
2378         kfree(new);
2379 }
2380
2381 static void relink_file_extents(struct new_sa_defrag_extent *new)
2382 {
2383         struct btrfs_path *path;
2384         struct sa_defrag_extent_backref *backref;
2385         struct sa_defrag_extent_backref *prev = NULL;
2386         struct inode *inode;
2387         struct btrfs_root *root;
2388         struct rb_node *node;
2389         int ret;
2390
2391         inode = new->inode;
2392         root = BTRFS_I(inode)->root;
2393
2394         path = btrfs_alloc_path();
2395         if (!path)
2396                 return;
2397
2398         if (!record_extent_backrefs(path, new)) {
2399                 btrfs_free_path(path);
2400                 goto out;
2401         }
2402         btrfs_release_path(path);
2403
2404         while (1) {
2405                 node = rb_first(&new->root);
2406                 if (!node)
2407                         break;
2408                 rb_erase(node, &new->root);
2409
2410                 backref = rb_entry(node, struct sa_defrag_extent_backref, node);
2411
2412                 ret = relink_extent_backref(path, prev, backref);
2413                 WARN_ON(ret < 0);
2414
2415                 kfree(prev);
2416
2417                 if (ret == 1)
2418                         prev = backref;
2419                 else
2420                         prev = NULL;
2421                 cond_resched();
2422         }
2423         kfree(prev);
2424
2425         btrfs_free_path(path);
2426 out:
2427         free_sa_defrag_extent(new);
2428
2429         atomic_dec(&root->fs_info->defrag_running);
2430         wake_up(&root->fs_info->transaction_wait);
2431 }
2432
2433 static struct new_sa_defrag_extent *
2434 record_old_file_extents(struct inode *inode,
2435                         struct btrfs_ordered_extent *ordered)
2436 {
2437         struct btrfs_root *root = BTRFS_I(inode)->root;
2438         struct btrfs_path *path;
2439         struct btrfs_key key;
2440         struct old_sa_defrag_extent *old;
2441         struct new_sa_defrag_extent *new;
2442         int ret;
2443
2444         new = kmalloc(sizeof(*new), GFP_NOFS);
2445         if (!new)
2446                 return NULL;
2447
2448         new->inode = inode;
2449         new->file_pos = ordered->file_offset;
2450         new->len = ordered->len;
2451         new->bytenr = ordered->start;
2452         new->disk_len = ordered->disk_len;
2453         new->compress_type = ordered->compress_type;
2454         new->root = RB_ROOT;
2455         INIT_LIST_HEAD(&new->head);
2456
2457         path = btrfs_alloc_path();
2458         if (!path)
2459                 goto out_kfree;
2460
2461         key.objectid = btrfs_ino(inode);
2462         key.type = BTRFS_EXTENT_DATA_KEY;
2463         key.offset = new->file_pos;
2464
2465         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2466         if (ret < 0)
2467                 goto out_free_path;
2468         if (ret > 0 && path->slots[0] > 0)
2469                 path->slots[0]--;
2470
2471         /* find out all the old extents for the file range */
2472         while (1) {
2473                 struct btrfs_file_extent_item *extent;
2474                 struct extent_buffer *l;
2475                 int slot;
2476                 u64 num_bytes;
2477                 u64 offset;
2478                 u64 end;
2479                 u64 disk_bytenr;
2480                 u64 extent_offset;
2481
2482                 l = path->nodes[0];
2483                 slot = path->slots[0];
2484
2485                 if (slot >= btrfs_header_nritems(l)) {
2486                         ret = btrfs_next_leaf(root, path);
2487                         if (ret < 0)
2488                                 goto out_free_path;
2489                         else if (ret > 0)
2490                                 break;
2491                         continue;
2492                 }
2493
2494                 btrfs_item_key_to_cpu(l, &key, slot);
2495
2496                 if (key.objectid != btrfs_ino(inode))
2497                         break;
2498                 if (key.type != BTRFS_EXTENT_DATA_KEY)
2499                         break;
2500                 if (key.offset >= new->file_pos + new->len)
2501                         break;
2502
2503                 extent = btrfs_item_ptr(l, slot, struct btrfs_file_extent_item);
2504
2505                 num_bytes = btrfs_file_extent_num_bytes(l, extent);
2506                 if (key.offset + num_bytes < new->file_pos)
2507                         goto next;
2508
2509                 disk_bytenr = btrfs_file_extent_disk_bytenr(l, extent);
2510                 if (!disk_bytenr)
2511                         goto next;
2512
2513                 extent_offset = btrfs_file_extent_offset(l, extent);
2514
2515                 old = kmalloc(sizeof(*old), GFP_NOFS);
2516                 if (!old)
2517                         goto out_free_path;
2518
2519                 offset = max(new->file_pos, key.offset);
2520                 end = min(new->file_pos + new->len, key.offset + num_bytes);
2521
2522                 old->bytenr = disk_bytenr;
2523                 old->extent_offset = extent_offset;
2524                 old->offset = offset - key.offset;
2525                 old->len = end - offset;
2526                 old->new = new;
2527                 old->count = 0;
2528                 list_add_tail(&old->list, &new->head);
2529 next:
2530                 path->slots[0]++;
2531                 cond_resched();
2532         }
2533
2534         btrfs_free_path(path);
2535         atomic_inc(&root->fs_info->defrag_running);
2536
2537         return new;
2538
2539 out_free_path:
2540         btrfs_free_path(path);
2541 out_kfree:
2542         free_sa_defrag_extent(new);
2543         return NULL;
2544 }
2545
2546 /*
2547  * helper function for btrfs_finish_ordered_io, this
2548  * just reads in some of the csum leaves to prime them into ram
2549  * before we start the transaction.  It limits the amount of btree
2550  * reads required while inside the transaction.
2551  */
2552 /* as ordered data IO finishes, this gets called so we can finish
2553  * an ordered extent if the range of bytes in the file it covers are
2554  * fully written.
2555  */
2556 static int btrfs_finish_ordered_io(struct btrfs_ordered_extent *ordered_extent)
2557 {
2558         struct inode *inode = ordered_extent->inode;
2559         struct btrfs_root *root = BTRFS_I(inode)->root;
2560         struct btrfs_trans_handle *trans = NULL;
2561         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
2562         struct extent_state *cached_state = NULL;
2563         struct new_sa_defrag_extent *new = NULL;
2564         int compress_type = 0;
2565         int ret = 0;
2566         u64 logical_len = ordered_extent->len;
2567         bool nolock;
2568         bool truncated = false;
2569
2570         nolock = btrfs_is_free_space_inode(inode);
2571
2572         if (test_bit(BTRFS_ORDERED_IOERR, &ordered_extent->flags)) {
2573                 ret = -EIO;
2574                 goto out;
2575         }
2576
2577         if (test_bit(BTRFS_ORDERED_TRUNCATED, &ordered_extent->flags)) {
2578                 truncated = true;
2579                 logical_len = ordered_extent->truncated_len;
2580                 /* Truncated the entire extent, don't bother adding */
2581                 if (!logical_len)
2582                         goto out;
2583         }
2584
2585         if (test_bit(BTRFS_ORDERED_NOCOW, &ordered_extent->flags)) {
2586                 BUG_ON(!list_empty(&ordered_extent->list)); /* Logic error */
2587                 btrfs_ordered_update_i_size(inode, 0, ordered_extent);
2588                 if (nolock)
2589                         trans = btrfs_join_transaction_nolock(root);
2590                 else
2591                         trans = btrfs_join_transaction(root);
2592                 if (IS_ERR(trans)) {
2593                         ret = PTR_ERR(trans);
2594                         trans = NULL;
2595                         goto out;
2596                 }
2597                 trans->block_rsv = &root->fs_info->delalloc_block_rsv;
2598                 ret = btrfs_update_inode_fallback(trans, root, inode);
2599                 if (ret) /* -ENOMEM or corruption */
2600                         btrfs_abort_transaction(trans, root, ret);
2601                 goto out;
2602         }
2603
2604         lock_extent_bits(io_tree, ordered_extent->file_offset,
2605                          ordered_extent->file_offset + ordered_extent->len - 1,
2606                          0, &cached_state);
2607
2608         ret = test_range_bit(io_tree, ordered_extent->file_offset,
2609                         ordered_extent->file_offset + ordered_extent->len - 1,
2610                         EXTENT_DEFRAG, 1, cached_state);
2611         if (ret) {
2612                 u64 last_snapshot = btrfs_root_last_snapshot(&root->root_item);
2613                 if (last_snapshot >= BTRFS_I(inode)->generation)
2614                         /* the inode is shared */
2615                         new = record_old_file_extents(inode, ordered_extent);
2616
2617                 clear_extent_bit(io_tree, ordered_extent->file_offset,
2618                         ordered_extent->file_offset + ordered_extent->len - 1,
2619                         EXTENT_DEFRAG, 0, 0, &cached_state, GFP_NOFS);
2620         }
2621
2622         if (nolock)
2623                 trans = btrfs_join_transaction_nolock(root);
2624         else
2625                 trans = btrfs_join_transaction(root);
2626         if (IS_ERR(trans)) {
2627                 ret = PTR_ERR(trans);
2628                 trans = NULL;
2629                 goto out_unlock;
2630         }
2631         trans->block_rsv = &root->fs_info->delalloc_block_rsv;
2632
2633         if (test_bit(BTRFS_ORDERED_COMPRESSED, &ordered_extent->flags))
2634                 compress_type = ordered_extent->compress_type;
2635         if (test_bit(BTRFS_ORDERED_PREALLOC, &ordered_extent->flags)) {
2636                 BUG_ON(compress_type);
2637                 ret = btrfs_mark_extent_written(trans, inode,
2638                                                 ordered_extent->file_offset,
2639                                                 ordered_extent->file_offset +
2640                                                 logical_len);
2641         } else {
2642                 BUG_ON(root == root->fs_info->tree_root);
2643                 ret = insert_reserved_file_extent(trans, inode,
2644                                                 ordered_extent->file_offset,
2645                                                 ordered_extent->start,
2646                                                 ordered_extent->disk_len,
2647                                                 logical_len, logical_len,
2648                                                 compress_type, 0, 0,
2649                                                 BTRFS_FILE_EXTENT_REG);
2650         }
2651         unpin_extent_cache(&BTRFS_I(inode)->extent_tree,
2652                            ordered_extent->file_offset, ordered_extent->len,
2653                            trans->transid);
2654         if (ret < 0) {
2655                 btrfs_abort_transaction(trans, root, ret);
2656                 goto out_unlock;
2657         }
2658
2659         add_pending_csums(trans, inode, ordered_extent->file_offset,
2660                           &ordered_extent->list);
2661
2662         btrfs_ordered_update_i_size(inode, 0, ordered_extent);
2663         ret = btrfs_update_inode_fallback(trans, root, inode);
2664         if (ret) { /* -ENOMEM or corruption */
2665                 btrfs_abort_transaction(trans, root, ret);
2666                 goto out_unlock;
2667         }
2668         ret = 0;
2669 out_unlock:
2670         unlock_extent_cached(io_tree, ordered_extent->file_offset,
2671                              ordered_extent->file_offset +
2672                              ordered_extent->len - 1, &cached_state, GFP_NOFS);
2673 out:
2674         if (root != root->fs_info->tree_root)
2675                 btrfs_delalloc_release_metadata(inode, ordered_extent->len);
2676         if (trans)
2677                 btrfs_end_transaction(trans, root);
2678
2679         if (ret || truncated) {
2680                 u64 start, end;
2681
2682                 if (truncated)
2683                         start = ordered_extent->file_offset + logical_len;
2684                 else
2685                         start = ordered_extent->file_offset;
2686                 end = ordered_extent->file_offset + ordered_extent->len - 1;
2687                 clear_extent_uptodate(io_tree, start, end, NULL, GFP_NOFS);
2688
2689                 /* Drop the cache for the part of the extent we didn't write. */
2690                 btrfs_drop_extent_cache(inode, start, end, 0);
2691
2692                 /*
2693                  * If the ordered extent had an IOERR or something else went
2694                  * wrong we need to return the space for this ordered extent
2695                  * back to the allocator.  We only free the extent in the
2696                  * truncated case if we didn't write out the extent at all.
2697                  */
2698                 if ((ret || !logical_len) &&
2699                     !test_bit(BTRFS_ORDERED_NOCOW, &ordered_extent->flags) &&
2700                     !test_bit(BTRFS_ORDERED_PREALLOC, &ordered_extent->flags))
2701                         btrfs_free_reserved_extent(root, ordered_extent->start,
2702                                                    ordered_extent->disk_len);
2703         }
2704
2705
2706         /*
2707          * This needs to be done to make sure anybody waiting knows we are done
2708          * updating everything for this ordered extent.
2709          */
2710         btrfs_remove_ordered_extent(inode, ordered_extent);
2711
2712         /* for snapshot-aware defrag */
2713         if (new) {
2714                 if (ret) {
2715                         free_sa_defrag_extent(new);
2716                         atomic_dec(&root->fs_info->defrag_running);
2717                 } else {
2718                         relink_file_extents(new);
2719                 }
2720         }
2721
2722         /* once for us */
2723         btrfs_put_ordered_extent(ordered_extent);
2724         /* once for the tree */
2725         btrfs_put_ordered_extent(ordered_extent);
2726
2727         return ret;
2728 }
2729
2730 static void finish_ordered_fn(struct btrfs_work *work)
2731 {
2732         struct btrfs_ordered_extent *ordered_extent;
2733         ordered_extent = container_of(work, struct btrfs_ordered_extent, work);
2734         btrfs_finish_ordered_io(ordered_extent);
2735 }
2736
2737 static int btrfs_writepage_end_io_hook(struct page *page, u64 start, u64 end,
2738                                 struct extent_state *state, int uptodate)
2739 {
2740         struct inode *inode = page->mapping->host;
2741         struct btrfs_root *root = BTRFS_I(inode)->root;
2742         struct btrfs_ordered_extent *ordered_extent = NULL;
2743         struct btrfs_workers *workers;
2744
2745         trace_btrfs_writepage_end_io_hook(page, start, end, uptodate);
2746
2747         ClearPagePrivate2(page);
2748         if (!btrfs_dec_test_ordered_pending(inode, &ordered_extent, start,
2749                                             end - start + 1, uptodate))
2750                 return 0;
2751
2752         ordered_extent->work.func = finish_ordered_fn;
2753         ordered_extent->work.flags = 0;
2754
2755         if (btrfs_is_free_space_inode(inode))
2756                 workers = &root->fs_info->endio_freespace_worker;
2757         else
2758                 workers = &root->fs_info->endio_write_workers;
2759         btrfs_queue_worker(workers, &ordered_extent->work);
2760
2761         return 0;
2762 }
2763
2764 /*
2765  * when reads are done, we need to check csums to verify the data is correct
2766  * if there's a match, we allow the bio to finish.  If not, the code in
2767  * extent_io.c will try to find good copies for us.
2768  */
2769 static int btrfs_readpage_end_io_hook(struct btrfs_io_bio *io_bio,
2770                                       u64 phy_offset, struct page *page,
2771                                       u64 start, u64 end, int mirror)
2772 {
2773         size_t offset = start - page_offset(page);
2774         struct inode *inode = page->mapping->host;
2775         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
2776         char *kaddr;
2777         struct btrfs_root *root = BTRFS_I(inode)->root;
2778         u32 csum_expected;
2779         u32 csum = ~(u32)0;
2780         static DEFINE_RATELIMIT_STATE(_rs, DEFAULT_RATELIMIT_INTERVAL,
2781                                       DEFAULT_RATELIMIT_BURST);
2782
2783         if (PageChecked(page)) {
2784                 ClearPageChecked(page);
2785                 goto good;
2786         }
2787
2788         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)
2789                 goto good;
2790
2791         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID &&
2792             test_range_bit(io_tree, start, end, EXTENT_NODATASUM, 1, NULL)) {
2793                 clear_extent_bits(io_tree, start, end, EXTENT_NODATASUM,
2794                                   GFP_NOFS);
2795                 return 0;
2796         }
2797
2798         phy_offset >>= inode->i_sb->s_blocksize_bits;
2799         csum_expected = *(((u32 *)io_bio->csum) + phy_offset);
2800
2801         kaddr = kmap_atomic(page);
2802         csum = btrfs_csum_data(kaddr + offset, csum,  end - start + 1);
2803         btrfs_csum_final(csum, (char *)&csum);
2804         if (csum != csum_expected)
2805                 goto zeroit;
2806
2807         kunmap_atomic(kaddr);
2808 good:
2809         return 0;
2810
2811 zeroit:
2812         if (__ratelimit(&_rs))
2813                 btrfs_info(root->fs_info, "csum failed ino %llu off %llu csum %u expected csum %u",
2814                         btrfs_ino(page->mapping->host), start, csum, csum_expected);
2815         memset(kaddr + offset, 1, end - start + 1);
2816         flush_dcache_page(page);
2817         kunmap_atomic(kaddr);
2818         if (csum_expected == 0)
2819                 return 0;
2820         return -EIO;
2821 }
2822
2823 struct delayed_iput {
2824         struct list_head list;
2825         struct inode *inode;
2826 };
2827
2828 /* JDM: If this is fs-wide, why can't we add a pointer to
2829  * btrfs_inode instead and avoid the allocation? */
2830 void btrfs_add_delayed_iput(struct inode *inode)
2831 {
2832         struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
2833         struct delayed_iput *delayed;
2834
2835         if (atomic_add_unless(&inode->i_count, -1, 1))
2836                 return;
2837
2838         delayed = kmalloc(sizeof(*delayed), GFP_NOFS | __GFP_NOFAIL);
2839         delayed->inode = inode;
2840
2841         spin_lock(&fs_info->delayed_iput_lock);
2842         list_add_tail(&delayed->list, &fs_info->delayed_iputs);
2843         spin_unlock(&fs_info->delayed_iput_lock);
2844 }
2845
2846 void btrfs_run_delayed_iputs(struct btrfs_root *root)
2847 {
2848         LIST_HEAD(list);
2849         struct btrfs_fs_info *fs_info = root->fs_info;
2850         struct delayed_iput *delayed;
2851         int empty;
2852
2853         spin_lock(&fs_info->delayed_iput_lock);
2854         empty = list_empty(&fs_info->delayed_iputs);
2855         spin_unlock(&fs_info->delayed_iput_lock);
2856         if (empty)
2857                 return;
2858
2859         spin_lock(&fs_info->delayed_iput_lock);
2860         list_splice_init(&fs_info->delayed_iputs, &list);
2861         spin_unlock(&fs_info->delayed_iput_lock);
2862
2863         while (!list_empty(&list)) {
2864                 delayed = list_entry(list.next, struct delayed_iput, list);
2865                 list_del(&delayed->list);
2866                 iput(delayed->inode);
2867                 kfree(delayed);
2868         }
2869 }
2870
2871 /*
2872  * This is called in transaction commit time. If there are no orphan
2873  * files in the subvolume, it removes orphan item and frees block_rsv
2874  * structure.
2875  */
2876 void btrfs_orphan_commit_root(struct btrfs_trans_handle *trans,
2877                               struct btrfs_root *root)
2878 {
2879         struct btrfs_block_rsv *block_rsv;
2880         int ret;
2881
2882         if (atomic_read(&root->orphan_inodes) ||
2883             root->orphan_cleanup_state != ORPHAN_CLEANUP_DONE)
2884                 return;
2885
2886         spin_lock(&root->orphan_lock);
2887         if (atomic_read(&root->orphan_inodes)) {
2888                 spin_unlock(&root->orphan_lock);
2889                 return;
2890         }
2891
2892         if (root->orphan_cleanup_state != ORPHAN_CLEANUP_DONE) {
2893                 spin_unlock(&root->orphan_lock);
2894                 return;
2895         }
2896
2897         block_rsv = root->orphan_block_rsv;
2898         root->orphan_block_rsv = NULL;
2899         spin_unlock(&root->orphan_lock);
2900
2901         if (root->orphan_item_inserted &&
2902             btrfs_root_refs(&root->root_item) > 0) {
2903                 ret = btrfs_del_orphan_item(trans, root->fs_info->tree_root,
2904                                             root->root_key.objectid);
2905                 if (ret)
2906                         btrfs_abort_transaction(trans, root, ret);
2907                 else
2908                         root->orphan_item_inserted = 0;
2909         }
2910
2911         if (block_rsv) {
2912                 WARN_ON(block_rsv->size > 0);
2913                 btrfs_free_block_rsv(root, block_rsv);
2914         }
2915 }
2916
2917 /*
2918  * This creates an orphan entry for the given inode in case something goes
2919  * wrong in the middle of an unlink/truncate.
2920  *
2921  * NOTE: caller of this function should reserve 5 units of metadata for
2922  *       this function.
2923  */
2924 int btrfs_orphan_add(struct btrfs_trans_handle *trans, struct inode *inode)
2925 {
2926         struct btrfs_root *root = BTRFS_I(inode)->root;
2927         struct btrfs_block_rsv *block_rsv = NULL;
2928         int reserve = 0;
2929         int insert = 0;
2930         int ret;
2931
2932         if (!root->orphan_block_rsv) {
2933                 block_rsv = btrfs_alloc_block_rsv(root, BTRFS_BLOCK_RSV_TEMP);
2934                 if (!block_rsv)
2935                         return -ENOMEM;
2936         }
2937
2938         spin_lock(&root->orphan_lock);
2939         if (!root->orphan_block_rsv) {
2940                 root->orphan_block_rsv = block_rsv;
2941         } else if (block_rsv) {
2942                 btrfs_free_block_rsv(root, block_rsv);
2943                 block_rsv = NULL;
2944         }
2945
2946         if (!test_and_set_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
2947                               &BTRFS_I(inode)->runtime_flags)) {
2948 #if 0
2949                 /*
2950                  * For proper ENOSPC handling, we should do orphan
2951                  * cleanup when mounting. But this introduces backward
2952                  * compatibility issue.
2953                  */
2954                 if (!xchg(&root->orphan_item_inserted, 1))
2955                         insert = 2;
2956                 else
2957                         insert = 1;
2958 #endif
2959                 insert = 1;
2960                 atomic_inc(&root->orphan_inodes);
2961         }
2962
2963         if (!test_and_set_bit(BTRFS_INODE_ORPHAN_META_RESERVED,
2964                               &BTRFS_I(inode)->runtime_flags))
2965                 reserve = 1;
2966         spin_unlock(&root->orphan_lock);
2967
2968         /* grab metadata reservation from transaction handle */
2969         if (reserve) {
2970                 ret = btrfs_orphan_reserve_metadata(trans, inode);
2971                 BUG_ON(ret); /* -ENOSPC in reservation; Logic error? JDM */
2972         }
2973
2974         /* insert an orphan item to track this unlinked/truncated file */
2975         if (insert >= 1) {
2976                 ret = btrfs_insert_orphan_item(trans, root, btrfs_ino(inode));
2977                 if (ret) {
2978                         atomic_dec(&root->orphan_inodes);
2979                         if (reserve) {
2980                                 clear_bit(BTRFS_INODE_ORPHAN_META_RESERVED,
2981                                           &BTRFS_I(inode)->runtime_flags);
2982                                 btrfs_orphan_release_metadata(inode);
2983                         }
2984                         if (ret != -EEXIST) {
2985                                 clear_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
2986                                           &BTRFS_I(inode)->runtime_flags);
2987                                 btrfs_abort_transaction(trans, root, ret);
2988                                 return ret;
2989                         }
2990                 }
2991                 ret = 0;
2992         }
2993
2994         /* insert an orphan item to track subvolume contains orphan files */
2995         if (insert >= 2) {
2996                 ret = btrfs_insert_orphan_item(trans, root->fs_info->tree_root,
2997                                                root->root_key.objectid);
2998                 if (ret && ret != -EEXIST) {
2999                         btrfs_abort_transaction(trans, root, ret);
3000                         return ret;
3001                 }
3002         }
3003         return 0;
3004 }
3005
3006 /*
3007  * We have done the truncate/delete so we can go ahead and remove the orphan
3008  * item for this particular inode.
3009  */
3010 static int btrfs_orphan_del(struct btrfs_trans_handle *trans,
3011                             struct inode *inode)
3012 {
3013         struct btrfs_root *root = BTRFS_I(inode)->root;
3014         int delete_item = 0;
3015         int release_rsv = 0;
3016         int ret = 0;
3017
3018         spin_lock(&root->orphan_lock);
3019         if (test_and_clear_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
3020                                &BTRFS_I(inode)->runtime_flags))
3021                 delete_item = 1;
3022
3023         if (test_and_clear_bit(BTRFS_INODE_ORPHAN_META_RESERVED,
3024                                &BTRFS_I(inode)->runtime_flags))
3025                 release_rsv = 1;
3026         spin_unlock(&root->orphan_lock);
3027
3028         if (delete_item) {
3029                 atomic_dec(&root->orphan_inodes);
3030                 if (trans)
3031                         ret = btrfs_del_orphan_item(trans, root,
3032                                                     btrfs_ino(inode));
3033         }
3034
3035         if (release_rsv)
3036                 btrfs_orphan_release_metadata(inode);
3037
3038         return ret;
3039 }
3040
3041 /*
3042  * this cleans up any orphans that may be left on the list from the last use
3043  * of this root.
3044  */
3045 int btrfs_orphan_cleanup(struct btrfs_root *root)
3046 {
3047         struct btrfs_path *path;
3048         struct extent_buffer *leaf;
3049         struct btrfs_key key, found_key;
3050         struct btrfs_trans_handle *trans;
3051         struct inode *inode;
3052         u64 last_objectid = 0;
3053         int ret = 0, nr_unlink = 0, nr_truncate = 0;
3054
3055         if (cmpxchg(&root->orphan_cleanup_state, 0, ORPHAN_CLEANUP_STARTED))
3056                 return 0;
3057
3058         path = btrfs_alloc_path();
3059         if (!path) {
3060                 ret = -ENOMEM;
3061                 goto out;
3062         }
3063         path->reada = -1;
3064
3065         key.objectid = BTRFS_ORPHAN_OBJECTID;
3066         btrfs_set_key_type(&key, BTRFS_ORPHAN_ITEM_KEY);
3067         key.offset = (u64)-1;
3068
3069         while (1) {
3070                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3071                 if (ret < 0)
3072                         goto out;
3073
3074                 /*
3075                  * if ret == 0 means we found what we were searching for, which
3076                  * is weird, but possible, so only screw with path if we didn't
3077                  * find the key and see if we have stuff that matches
3078                  */
3079                 if (ret > 0) {
3080                         ret = 0;
3081                         if (path->slots[0] == 0)
3082                                 break;
3083                         path->slots[0]--;
3084                 }
3085
3086                 /* pull out the item */
3087                 leaf = path->nodes[0];
3088                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
3089
3090                 /* make sure the item matches what we want */
3091                 if (found_key.objectid != BTRFS_ORPHAN_OBJECTID)
3092                         break;
3093                 if (btrfs_key_type(&found_key) != BTRFS_ORPHAN_ITEM_KEY)
3094                         break;
3095
3096                 /* release the path since we're done with it */
3097                 btrfs_release_path(path);
3098
3099                 /*
3100                  * this is where we are basically btrfs_lookup, without the
3101                  * crossing root thing.  we store the inode number in the
3102                  * offset of the orphan item.
3103                  */
3104
3105                 if (found_key.offset == last_objectid) {
3106                         btrfs_err(root->fs_info,
3107                                 "Error removing orphan entry, stopping orphan cleanup");
3108                         ret = -EINVAL;
3109                         goto out;
3110                 }
3111
3112                 last_objectid = found_key.offset;
3113
3114                 found_key.objectid = found_key.offset;
3115                 found_key.type = BTRFS_INODE_ITEM_KEY;
3116                 found_key.offset = 0;
3117                 inode = btrfs_iget(root->fs_info->sb, &found_key, root, NULL);
3118                 ret = PTR_ERR_OR_ZERO(inode);
3119                 if (ret && ret != -ESTALE)
3120                         goto out;
3121
3122                 if (ret == -ESTALE && root == root->fs_info->tree_root) {
3123                         struct btrfs_root *dead_root;
3124                         struct btrfs_fs_info *fs_info = root->fs_info;
3125                         int is_dead_root = 0;
3126
3127                         /*
3128                          * this is an orphan in the tree root. Currently these
3129                          * could come from 2 sources:
3130                          *  a) a snapshot deletion in progress
3131                          *  b) a free space cache inode
3132                          * We need to distinguish those two, as the snapshot
3133                          * orphan must not get deleted.
3134                          * find_dead_roots already ran before us, so if this
3135                          * is a snapshot deletion, we should find the root
3136                          * in the dead_roots list
3137                          */
3138                         spin_lock(&fs_info->trans_lock);
3139                         list_for_each_entry(dead_root, &fs_info->dead_roots,
3140                                             root_list) {
3141                                 if (dead_root->root_key.objectid ==
3142                                     found_key.objectid) {
3143                                         is_dead_root = 1;
3144                                         break;
3145                                 }
3146                         }
3147                         spin_unlock(&fs_info->trans_lock);
3148                         if (is_dead_root) {
3149                                 /* prevent this orphan from being found again */
3150                                 key.offset = found_key.objectid - 1;
3151                                 continue;
3152                         }
3153                 }
3154                 /*
3155                  * Inode is already gone but the orphan item is still there,
3156                  * kill the orphan item.
3157                  */
3158                 if (ret == -ESTALE) {
3159                         trans = btrfs_start_transaction(root, 1);
3160                         if (IS_ERR(trans)) {
3161                                 ret = PTR_ERR(trans);
3162                                 goto out;
3163                         }
3164                         btrfs_debug(root->fs_info, "auto deleting %Lu",
3165                                 found_key.objectid);
3166                         ret = btrfs_del_orphan_item(trans, root,
3167                                                     found_key.objectid);
3168                         btrfs_end_transaction(trans, root);
3169                         if (ret)
3170                                 goto out;
3171                         continue;
3172                 }
3173
3174                 /*
3175                  * add this inode to the orphan list so btrfs_orphan_del does
3176                  * the proper thing when we hit it
3177                  */
3178                 set_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
3179                         &BTRFS_I(inode)->runtime_flags);
3180                 atomic_inc(&root->orphan_inodes);
3181
3182                 /* if we have links, this was a truncate, lets do that */
3183                 if (inode->i_nlink) {
3184                         if (WARN_ON(!S_ISREG(inode->i_mode))) {
3185                                 iput(inode);
3186                                 continue;
3187                         }
3188                         nr_truncate++;
3189
3190                         /* 1 for the orphan item deletion. */
3191                         trans = btrfs_start_transaction(root, 1);
3192                         if (IS_ERR(trans)) {
3193                                 iput(inode);
3194                                 ret = PTR_ERR(trans);
3195                                 goto out;
3196                         }
3197                         ret = btrfs_orphan_add(trans, inode);
3198                         btrfs_end_transaction(trans, root);
3199                         if (ret) {
3200                                 iput(inode);
3201                                 goto out;
3202                         }
3203
3204                         ret = btrfs_truncate(inode);
3205                         if (ret)
3206                                 btrfs_orphan_del(NULL, inode);
3207                 } else {
3208                         nr_unlink++;
3209                 }
3210
3211                 /* this will do delete_inode and everything for us */
3212                 iput(inode);
3213                 if (ret)
3214                         goto out;
3215         }
3216         /* release the path since we're done with it */
3217         btrfs_release_path(path);
3218
3219         root->orphan_cleanup_state = ORPHAN_CLEANUP_DONE;
3220
3221         if (root->orphan_block_rsv)
3222                 btrfs_block_rsv_release(root, root->orphan_block_rsv,
3223                                         (u64)-1);
3224
3225         if (root->orphan_block_rsv || root->orphan_item_inserted) {
3226                 trans = btrfs_join_transaction(root);
3227                 if (!IS_ERR(trans))
3228                         btrfs_end_transaction(trans, root);
3229         }
3230
3231         if (nr_unlink)
3232                 btrfs_debug(root->fs_info, "unlinked %d orphans", nr_unlink);
3233         if (nr_truncate)
3234                 btrfs_debug(root->fs_info, "truncated %d orphans", nr_truncate);
3235
3236 out:
3237         if (ret)
3238                 btrfs_crit(root->fs_info,
3239                         "could not do orphan cleanup %d", ret);
3240         btrfs_free_path(path);
3241         return ret;
3242 }
3243
3244 /*
3245  * very simple check to peek ahead in the leaf looking for xattrs.  If we
3246  * don't find any xattrs, we know there can't be any acls.
3247  *
3248  * slot is the slot the inode is in, objectid is the objectid of the inode
3249  */
3250 static noinline int acls_after_inode_item(struct extent_buffer *leaf,
3251                                           int slot, u64 objectid)
3252 {
3253         u32 nritems = btrfs_header_nritems(leaf);
3254         struct btrfs_key found_key;
3255         static u64 xattr_access = 0;
3256         static u64 xattr_default = 0;
3257         int scanned = 0;
3258
3259         if (!xattr_access) {
3260                 xattr_access = btrfs_name_hash(POSIX_ACL_XATTR_ACCESS,
3261                                         strlen(POSIX_ACL_XATTR_ACCESS));
3262                 xattr_default = btrfs_name_hash(POSIX_ACL_XATTR_DEFAULT,
3263                                         strlen(POSIX_ACL_XATTR_DEFAULT));
3264         }
3265
3266         slot++;
3267         while (slot < nritems) {
3268                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
3269
3270                 /* we found a different objectid, there must not be acls */
3271                 if (found_key.objectid != objectid)
3272                         return 0;
3273
3274                 /* we found an xattr, assume we've got an acl */
3275                 if (found_key.type == BTRFS_XATTR_ITEM_KEY) {
3276                         if (found_key.offset == xattr_access ||
3277                             found_key.offset == xattr_default)
3278                                 return 1;
3279                 }
3280
3281                 /*
3282                  * we found a key greater than an xattr key, there can't
3283                  * be any acls later on
3284                  */
3285                 if (found_key.type > BTRFS_XATTR_ITEM_KEY)
3286                         return 0;
3287
3288                 slot++;
3289                 scanned++;
3290
3291                 /*
3292                  * it goes inode, inode backrefs, xattrs, extents,
3293                  * so if there are a ton of hard links to an inode there can
3294                  * be a lot of backrefs.  Don't waste time searching too hard,
3295                  * this is just an optimization
3296                  */
3297                 if (scanned >= 8)
3298                         break;
3299         }
3300         /* we hit the end of the leaf before we found an xattr or
3301          * something larger than an xattr.  We have to assume the inode
3302          * has acls
3303          */
3304         return 1;
3305 }
3306
3307 /*
3308  * read an inode from the btree into the in-memory inode
3309  */
3310 static void btrfs_read_locked_inode(struct inode *inode)
3311 {
3312         struct btrfs_path *path;
3313         struct extent_buffer *leaf;
3314         struct btrfs_inode_item *inode_item;
3315         struct btrfs_timespec *tspec;
3316         struct btrfs_root *root = BTRFS_I(inode)->root;
3317         struct btrfs_key location;
3318         int maybe_acls;
3319         u32 rdev;
3320         int ret;
3321         bool filled = false;
3322
3323         ret = btrfs_fill_inode(inode, &rdev);
3324         if (!ret)
3325                 filled = true;
3326
3327         path = btrfs_alloc_path();
3328         if (!path)
3329                 goto make_bad;
3330
3331         path->leave_spinning = 1;
3332         memcpy(&location, &BTRFS_I(inode)->location, sizeof(location));
3333
3334         ret = btrfs_lookup_inode(NULL, root, path, &location, 0);
3335         if (ret)
3336                 goto make_bad;
3337
3338         leaf = path->nodes[0];
3339
3340         if (filled)
3341                 goto cache_acl;
3342
3343         inode_item = btrfs_item_ptr(leaf, path->slots[0],
3344                                     struct btrfs_inode_item);
3345         inode->i_mode = btrfs_inode_mode(leaf, inode_item);
3346         set_nlink(inode, btrfs_inode_nlink(leaf, inode_item));
3347         i_uid_write(inode, btrfs_inode_uid(leaf, inode_item));
3348         i_gid_write(inode, btrfs_inode_gid(leaf, inode_item));
3349         btrfs_i_size_write(inode, btrfs_inode_size(leaf, inode_item));
3350
3351         tspec = btrfs_inode_atime(inode_item);
3352         inode->i_atime.tv_sec = btrfs_timespec_sec(leaf, tspec);
3353         inode->i_atime.tv_nsec = btrfs_timespec_nsec(leaf, tspec);
3354
3355         tspec = btrfs_inode_mtime(inode_item);
3356         inode->i_mtime.tv_sec = btrfs_timespec_sec(leaf, tspec);
3357         inode->i_mtime.tv_nsec = btrfs_timespec_nsec(leaf, tspec);
3358
3359         tspec = btrfs_inode_ctime(inode_item);
3360         inode->i_ctime.tv_sec = btrfs_timespec_sec(leaf, tspec);
3361         inode->i_ctime.tv_nsec = btrfs_timespec_nsec(leaf, tspec);
3362
3363         inode_set_bytes(inode, btrfs_inode_nbytes(leaf, inode_item));
3364         BTRFS_I(inode)->generation = btrfs_inode_generation(leaf, inode_item);
3365         BTRFS_I(inode)->last_trans = btrfs_inode_transid(leaf, inode_item);
3366
3367         /*
3368          * If we were modified in the current generation and evicted from memory
3369          * and then re-read we need to do a full sync since we don't have any
3370          * idea about which extents were modified before we were evicted from
3371          * cache.
3372          */
3373         if (BTRFS_I(inode)->last_trans == root->fs_info->generation)
3374                 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
3375                         &BTRFS_I(inode)->runtime_flags);
3376
3377         inode->i_version = btrfs_inode_sequence(leaf, inode_item);
3378         inode->i_generation = BTRFS_I(inode)->generation;
3379         inode->i_rdev = 0;
3380         rdev = btrfs_inode_rdev(leaf, inode_item);
3381
3382         BTRFS_I(inode)->index_cnt = (u64)-1;
3383         BTRFS_I(inode)->flags = btrfs_inode_flags(leaf, inode_item);
3384 cache_acl:
3385         /*
3386          * try to precache a NULL acl entry for files that don't have
3387          * any xattrs or acls
3388          */
3389         maybe_acls = acls_after_inode_item(leaf, path->slots[0],
3390                                            btrfs_ino(inode));
3391         if (!maybe_acls)
3392                 cache_no_acl(inode);
3393
3394         btrfs_free_path(path);
3395
3396         switch (inode->i_mode & S_IFMT) {
3397         case S_IFREG:
3398                 inode->i_mapping->a_ops = &btrfs_aops;
3399                 inode->i_mapping->backing_dev_info = &root->fs_info->bdi;
3400                 BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
3401                 inode->i_fop = &btrfs_file_operations;
3402                 inode->i_op = &btrfs_file_inode_operations;
3403                 break;
3404         case S_IFDIR:
3405                 inode->i_fop = &btrfs_dir_file_operations;
3406                 if (root == root->fs_info->tree_root)
3407                         inode->i_op = &btrfs_dir_ro_inode_operations;
3408                 else
3409                         inode->i_op = &btrfs_dir_inode_operations;
3410                 break;
3411         case S_IFLNK:
3412                 inode->i_op = &btrfs_symlink_inode_operations;
3413                 inode->i_mapping->a_ops = &btrfs_symlink_aops;
3414                 inode->i_mapping->backing_dev_info = &root->fs_info->bdi;
3415                 break;
3416         default:
3417                 inode->i_op = &btrfs_special_inode_operations;
3418                 init_special_inode(inode, inode->i_mode, rdev);
3419                 break;
3420         }
3421
3422         btrfs_update_iflags(inode);
3423         return;
3424
3425 make_bad:
3426         btrfs_free_path(path);
3427         make_bad_inode(inode);
3428 }
3429
3430 /*
3431  * given a leaf and an inode, copy the inode fields into the leaf
3432  */
3433 static void fill_inode_item(struct btrfs_trans_handle *trans,
3434                             struct extent_buffer *leaf,
3435                             struct btrfs_inode_item *item,
3436                             struct inode *inode)
3437 {
3438         struct btrfs_map_token token;
3439
3440         btrfs_init_map_token(&token);
3441
3442         btrfs_set_token_inode_uid(leaf, item, i_uid_read(inode), &token);
3443         btrfs_set_token_inode_gid(leaf, item, i_gid_read(inode), &token);
3444         btrfs_set_token_inode_size(leaf, item, BTRFS_I(inode)->disk_i_size,
3445                                    &token);
3446         btrfs_set_token_inode_mode(leaf, item, inode->i_mode, &token);
3447         btrfs_set_token_inode_nlink(leaf, item, inode->i_nlink, &token);
3448
3449         btrfs_set_token_timespec_sec(leaf, btrfs_inode_atime(item),
3450                                      inode->i_atime.tv_sec, &token);
3451         btrfs_set_token_timespec_nsec(leaf, btrfs_inode_atime(item),
3452                                       inode->i_atime.tv_nsec, &token);
3453
3454         btrfs_set_token_timespec_sec(leaf, btrfs_inode_mtime(item),
3455                                      inode->i_mtime.tv_sec, &token);
3456         btrfs_set_token_timespec_nsec(leaf, btrfs_inode_mtime(item),
3457                                       inode->i_mtime.tv_nsec, &token);
3458
3459         btrfs_set_token_timespec_sec(leaf, btrfs_inode_ctime(item),
3460                                      inode->i_ctime.tv_sec, &token);
3461         btrfs_set_token_timespec_nsec(leaf, btrfs_inode_ctime(item),
3462                                       inode->i_ctime.tv_nsec, &token);
3463
3464         btrfs_set_token_inode_nbytes(leaf, item, inode_get_bytes(inode),
3465                                      &token);
3466         btrfs_set_token_inode_generation(leaf, item, BTRFS_I(inode)->generation,
3467                                          &token);
3468         btrfs_set_token_inode_sequence(leaf, item, inode->i_version, &token);
3469         btrfs_set_token_inode_transid(leaf, item, trans->transid, &token);
3470         btrfs_set_token_inode_rdev(leaf, item, inode->i_rdev, &token);
3471         btrfs_set_token_inode_flags(leaf, item, BTRFS_I(inode)->flags, &token);
3472         btrfs_set_token_inode_block_group(leaf, item, 0, &token);
3473 }
3474
3475 /*
3476  * copy everything in the in-memory inode into the btree.
3477  */
3478 static noinline int btrfs_update_inode_item(struct btrfs_trans_handle *trans,
3479                                 struct btrfs_root *root, struct inode *inode)
3480 {
3481         struct btrfs_inode_item *inode_item;
3482         struct btrfs_path *path;
3483         struct extent_buffer *leaf;
3484         int ret;
3485
3486         path = btrfs_alloc_path();
3487         if (!path)
3488                 return -ENOMEM;
3489
3490         path->leave_spinning = 1;
3491         ret = btrfs_lookup_inode(trans, root, path, &BTRFS_I(inode)->location,
3492                                  1);
3493         if (ret) {
3494                 if (ret > 0)
3495                         ret = -ENOENT;
3496                 goto failed;
3497         }
3498
3499         btrfs_unlock_up_safe(path, 1);
3500         leaf = path->nodes[0];
3501         inode_item = btrfs_item_ptr(leaf, path->slots[0],
3502                                     struct btrfs_inode_item);
3503
3504         fill_inode_item(trans, leaf, inode_item, inode);
3505         btrfs_mark_buffer_dirty(leaf);
3506         btrfs_set_inode_last_trans(trans, inode);
3507         ret = 0;
3508 failed:
3509         btrfs_free_path(path);
3510         return ret;
3511 }
3512
3513 /*
3514  * copy everything in the in-memory inode into the btree.
3515  */
3516 noinline int btrfs_update_inode(struct btrfs_trans_handle *trans,
3517                                 struct btrfs_root *root, struct inode *inode)
3518 {
3519         int ret;
3520
3521         /*
3522          * If the inode is a free space inode, we can deadlock during commit
3523          * if we put it into the delayed code.
3524          *
3525          * The data relocation inode should also be directly updated
3526          * without delay
3527          */
3528         if (!btrfs_is_free_space_inode(inode)
3529             && root->root_key.objectid != BTRFS_DATA_RELOC_TREE_OBJECTID) {
3530                 btrfs_update_root_times(trans, root);
3531
3532                 ret = btrfs_delayed_update_inode(trans, root, inode);
3533                 if (!ret)
3534                         btrfs_set_inode_last_trans(trans, inode);
3535                 return ret;
3536         }
3537
3538         return btrfs_update_inode_item(trans, root, inode);
3539 }
3540
3541 noinline int btrfs_update_inode_fallback(struct btrfs_trans_handle *trans,
3542                                          struct btrfs_root *root,
3543                                          struct inode *inode)
3544 {
3545         int ret;
3546
3547         ret = btrfs_update_inode(trans, root, inode);
3548         if (ret == -ENOSPC)
3549                 return btrfs_update_inode_item(trans, root, inode);
3550         return ret;
3551 }
3552
3553 /*
3554  * unlink helper that gets used here in inode.c and in the tree logging
3555  * recovery code.  It remove a link in a directory with a given name, and
3556  * also drops the back refs in the inode to the directory
3557  */
3558 static int __btrfs_unlink_inode(struct btrfs_trans_handle *trans,
3559                                 struct btrfs_root *root,
3560                                 struct inode *dir, struct inode *inode,
3561                                 const char *name, int name_len)
3562 {
3563         struct btrfs_path *path;
3564         int ret = 0;
3565         struct extent_buffer *leaf;
3566         struct btrfs_dir_item *di;
3567         struct btrfs_key key;
3568         u64 index;
3569         u64 ino = btrfs_ino(inode);
3570         u64 dir_ino = btrfs_ino(dir);
3571
3572         path = btrfs_alloc_path();
3573         if (!path) {
3574                 ret = -ENOMEM;
3575                 goto out;
3576         }
3577
3578         path->leave_spinning = 1;
3579         di = btrfs_lookup_dir_item(trans, root, path, dir_ino,
3580                                     name, name_len, -1);
3581         if (IS_ERR(di)) {
3582                 ret = PTR_ERR(di);
3583                 goto err;
3584         }
3585         if (!di) {
3586                 ret = -ENOENT;
3587                 goto err;
3588         }
3589         leaf = path->nodes[0];
3590         btrfs_dir_item_key_to_cpu(leaf, di, &key);
3591         ret = btrfs_delete_one_dir_name(trans, root, path, di);
3592         if (ret)
3593                 goto err;
3594         btrfs_release_path(path);
3595
3596         ret = btrfs_del_inode_ref(trans, root, name, name_len, ino,
3597                                   dir_ino, &index);
3598         if (ret) {
3599                 btrfs_info(root->fs_info,
3600                         "failed to delete reference to %.*s, inode %llu parent %llu",
3601                         name_len, name, ino, dir_ino);
3602                 btrfs_abort_transaction(trans, root, ret);
3603                 goto err;
3604         }
3605
3606         ret = btrfs_delete_delayed_dir_index(trans, root, dir, index);
3607         if (ret) {
3608                 btrfs_abort_transaction(trans, root, ret);
3609                 goto err;
3610         }
3611
3612         ret = btrfs_del_inode_ref_in_log(trans, root, name, name_len,
3613                                          inode, dir_ino);
3614         if (ret != 0 && ret != -ENOENT) {
3615                 btrfs_abort_transaction(trans, root, ret);
3616                 goto err;
3617         }
3618
3619         ret = btrfs_del_dir_entries_in_log(trans, root, name, name_len,
3620                                            dir, index);
3621         if (ret == -ENOENT)
3622                 ret = 0;
3623         else if (ret)
3624                 btrfs_abort_transaction(trans, root, ret);
3625 err:
3626         btrfs_free_path(path);
3627         if (ret)
3628                 goto out;
3629
3630         btrfs_i_size_write(dir, dir->i_size - name_len * 2);
3631         inode_inc_iversion(inode);
3632         inode_inc_iversion(dir);
3633         inode->i_ctime = dir->i_mtime = dir->i_ctime = CURRENT_TIME;
3634         ret = btrfs_update_inode(trans, root, dir);
3635 out:
3636         return ret;
3637 }
3638
3639 int btrfs_unlink_inode(struct btrfs_trans_handle *trans,
3640                        struct btrfs_root *root,
3641                        struct inode *dir, struct inode *inode,
3642                        const char *name, int name_len)
3643 {
3644         int ret;
3645         ret = __btrfs_unlink_inode(trans, root, dir, inode, name, name_len);
3646         if (!ret) {
3647                 drop_nlink(inode);
3648                 ret = btrfs_update_inode(trans, root, inode);
3649         }
3650         return ret;
3651 }
3652
3653 /*
3654  * helper to start transaction for unlink and rmdir.
3655  *
3656  * unlink and rmdir are special in btrfs, they do not always free space, so
3657  * if we cannot make our reservations the normal way try and see if there is
3658  * plenty of slack room in the global reserve to migrate, otherwise we cannot
3659  * allow the unlink to occur.
3660  */
3661 static struct btrfs_trans_handle *__unlink_start_trans(struct inode *dir)
3662 {
3663         struct btrfs_trans_handle *trans;
3664         struct btrfs_root *root = BTRFS_I(dir)->root;
3665         int ret;
3666
3667         /*
3668          * 1 for the possible orphan item
3669          * 1 for the dir item
3670          * 1 for the dir index
3671          * 1 for the inode ref
3672          * 1 for the inode
3673          */
3674         trans = btrfs_start_transaction(root, 5);
3675         if (!IS_ERR(trans) || PTR_ERR(trans) != -ENOSPC)
3676                 return trans;
3677
3678         if (PTR_ERR(trans) == -ENOSPC) {
3679                 u64 num_bytes = btrfs_calc_trans_metadata_size(root, 5);
3680
3681                 trans = btrfs_start_transaction(root, 0);
3682                 if (IS_ERR(trans))
3683                         return trans;
3684                 ret = btrfs_cond_migrate_bytes(root->fs_info,
3685                                                &root->fs_info->trans_block_rsv,
3686                                                num_bytes, 5);
3687                 if (ret) {
3688                         btrfs_end_transaction(trans, root);
3689                         return ERR_PTR(ret);
3690                 }
3691                 trans->block_rsv = &root->fs_info->trans_block_rsv;
3692                 trans->bytes_reserved = num_bytes;
3693         }
3694         return trans;
3695 }
3696
3697 static int btrfs_unlink(struct inode *dir, struct dentry *dentry)
3698 {
3699         struct btrfs_root *root = BTRFS_I(dir)->root;
3700         struct btrfs_trans_handle *trans;
3701         struct inode *inode = dentry->d_inode;
3702         int ret;
3703
3704         trans = __unlink_start_trans(dir);
3705         if (IS_ERR(trans))
3706                 return PTR_ERR(trans);
3707
3708         btrfs_record_unlink_dir(trans, dir, dentry->d_inode, 0);
3709
3710         ret = btrfs_unlink_inode(trans, root, dir, dentry->d_inode,
3711                                  dentry->d_name.name, dentry->d_name.len);
3712         if (ret)
3713                 goto out;
3714
3715         if (inode->i_nlink == 0) {
3716                 ret = btrfs_orphan_add(trans, inode);
3717                 if (ret)
3718                         goto out;
3719         }
3720
3721 out:
3722         btrfs_end_transaction(trans, root);
3723         btrfs_btree_balance_dirty(root);
3724         return ret;
3725 }
3726
3727 int btrfs_unlink_subvol(struct btrfs_trans_handle *trans,
3728                         struct btrfs_root *root,
3729                         struct inode *dir, u64 objectid,
3730                         const char *name, int name_len)
3731 {
3732         struct btrfs_path *path;
3733         struct extent_buffer *leaf;
3734         struct btrfs_dir_item *di;
3735         struct btrfs_key key;
3736         u64 index;
3737         int ret;
3738         u64 dir_ino = btrfs_ino(dir);
3739
3740         path = btrfs_alloc_path();
3741         if (!path)
3742                 return -ENOMEM;
3743
3744         di = btrfs_lookup_dir_item(trans, root, path, dir_ino,
3745                                    name, name_len, -1);
3746         if (IS_ERR_OR_NULL(di)) {
3747                 if (!di)
3748                         ret = -ENOENT;
3749                 else
3750                         ret = PTR_ERR(di);
3751                 goto out;
3752         }
3753
3754         leaf = path->nodes[0];
3755         btrfs_dir_item_key_to_cpu(leaf, di, &key);
3756         WARN_ON(key.type != BTRFS_ROOT_ITEM_KEY || key.objectid != objectid);
3757         ret = btrfs_delete_one_dir_name(trans, root, path, di);
3758         if (ret) {
3759                 btrfs_abort_transaction(trans, root, ret);
3760                 goto out;
3761         }
3762         btrfs_release_path(path);
3763
3764         ret = btrfs_del_root_ref(trans, root->fs_info->tree_root,
3765                                  objectid, root->root_key.objectid,
3766                                  dir_ino, &index, name, name_len);
3767         if (ret < 0) {
3768                 if (ret != -ENOENT) {
3769                         btrfs_abort_transaction(trans, root, ret);
3770                         goto out;
3771                 }
3772                 di = btrfs_search_dir_index_item(root, path, dir_ino,
3773                                                  name, name_len);
3774                 if (IS_ERR_OR_NULL(di)) {
3775                         if (!di)
3776                                 ret = -ENOENT;
3777                         else
3778                                 ret = PTR_ERR(di);
3779                         btrfs_abort_transaction(trans, root, ret);
3780                         goto out;
3781                 }
3782
3783                 leaf = path->nodes[0];
3784                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3785                 btrfs_release_path(path);
3786                 index = key.offset;
3787         }
3788         btrfs_release_path(path);
3789
3790         ret = btrfs_delete_delayed_dir_index(trans, root, dir, index);
3791         if (ret) {
3792                 btrfs_abort_transaction(trans, root, ret);
3793                 goto out;
3794         }
3795
3796         btrfs_i_size_write(dir, dir->i_size - name_len * 2);
3797         inode_inc_iversion(dir);
3798         dir->i_mtime = dir->i_ctime = CURRENT_TIME;
3799         ret = btrfs_update_inode_fallback(trans, root, dir);
3800         if (ret)
3801                 btrfs_abort_transaction(trans, root, ret);
3802 out:
3803         btrfs_free_path(path);
3804         return ret;
3805 }
3806
3807 static int btrfs_rmdir(struct inode *dir, struct dentry *dentry)
3808 {
3809         struct inode *inode = dentry->d_inode;
3810         int err = 0;
3811         struct btrfs_root *root = BTRFS_I(dir)->root;
3812         struct btrfs_trans_handle *trans;
3813
3814         if (inode->i_size > BTRFS_EMPTY_DIR_SIZE)
3815                 return -ENOTEMPTY;
3816         if (btrfs_ino(inode) == BTRFS_FIRST_FREE_OBJECTID)
3817                 return -EPERM;
3818
3819         trans = __unlink_start_trans(dir);
3820         if (IS_ERR(trans))
3821                 return PTR_ERR(trans);
3822
3823         if (unlikely(btrfs_ino(inode) == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)) {
3824                 err = btrfs_unlink_subvol(trans, root, dir,
3825                                           BTRFS_I(inode)->location.objectid,
3826                                           dentry->d_name.name,
3827                                           dentry->d_name.len);
3828                 goto out;
3829         }
3830
3831         err = btrfs_orphan_add(trans, inode);
3832         if (err)
3833                 goto out;
3834
3835         /* now the directory is empty */
3836         err = btrfs_unlink_inode(trans, root, dir, dentry->d_inode,
3837                                  dentry->d_name.name, dentry->d_name.len);
3838         if (!err)
3839                 btrfs_i_size_write(inode, 0);
3840 out:
3841         btrfs_end_transaction(trans, root);
3842         btrfs_btree_balance_dirty(root);
3843
3844         return err;
3845 }
3846
3847 /*
3848  * this can truncate away extent items, csum items and directory items.
3849  * It starts at a high offset and removes keys until it can't find
3850  * any higher than new_size
3851  *
3852  * csum items that cross the new i_size are truncated to the new size
3853  * as well.
3854  *
3855  * min_type is the minimum key type to truncate down to.  If set to 0, this
3856  * will kill all the items on this inode, including the INODE_ITEM_KEY.
3857  */
3858 int btrfs_truncate_inode_items(struct btrfs_trans_handle *trans,
3859                                struct btrfs_root *root,
3860                                struct inode *inode,
3861                                u64 new_size, u32 min_type)
3862 {
3863         struct btrfs_path *path;
3864         struct extent_buffer *leaf;
3865         struct btrfs_file_extent_item *fi;
3866         struct btrfs_key key;
3867         struct btrfs_key found_key;
3868         u64 extent_start = 0;
3869         u64 extent_num_bytes = 0;
3870         u64 extent_offset = 0;
3871         u64 item_end = 0;
3872         u64 last_size = (u64)-1;
3873         u32 found_type = (u8)-1;
3874         int found_extent;
3875         int del_item;
3876         int pending_del_nr = 0;
3877         int pending_del_slot = 0;
3878         int extent_type = -1;
3879         int ret;
3880         int err = 0;
3881         u64 ino = btrfs_ino(inode);
3882
3883         BUG_ON(new_size > 0 && min_type != BTRFS_EXTENT_DATA_KEY);
3884
3885         path = btrfs_alloc_path();
3886         if (!path)
3887                 return -ENOMEM;
3888         path->reada = -1;
3889
3890         /*
3891          * We want to drop from the next block forward in case this new size is
3892          * not block aligned since we will be keeping the last block of the
3893          * extent just the way it is.
3894          */
3895         if (root->ref_cows || root == root->fs_info->tree_root)
3896                 btrfs_drop_extent_cache(inode, ALIGN(new_size,
3897                                         root->sectorsize), (u64)-1, 0);
3898
3899         /*
3900          * This function is also used to drop the items in the log tree before
3901          * we relog the inode, so if root != BTRFS_I(inode)->root, it means
3902          * it is used to drop the loged items. So we shouldn't kill the delayed
3903          * items.
3904          */
3905         if (min_type == 0 && root == BTRFS_I(inode)->root)
3906                 btrfs_kill_delayed_inode_items(inode);
3907
3908         key.objectid = ino;
3909         key.offset = (u64)-1;
3910         key.type = (u8)-1;
3911
3912 search_again:
3913         path->leave_spinning = 1;
3914         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
3915         if (ret < 0) {
3916                 err = ret;
3917                 goto out;
3918         }
3919
3920         if (ret > 0) {
3921                 /* there are no items in the tree for us to truncate, we're
3922                  * done
3923                  */
3924                 if (path->slots[0] == 0)
3925                         goto out;
3926                 path->slots[0]--;
3927         }
3928
3929         while (1) {
3930                 fi = NULL;
3931                 leaf = path->nodes[0];
3932                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
3933                 found_type = btrfs_key_type(&found_key);
3934
3935                 if (found_key.objectid != ino)
3936                         break;
3937
3938                 if (found_type < min_type)
3939                         break;
3940
3941                 item_end = found_key.offset;
3942                 if (found_type == BTRFS_EXTENT_DATA_KEY) {
3943                         fi = btrfs_item_ptr(leaf, path->slots[0],
3944                                             struct btrfs_file_extent_item);
3945                         extent_type = btrfs_file_extent_type(leaf, fi);
3946                         if (extent_type != BTRFS_FILE_EXTENT_INLINE) {
3947                                 item_end +=
3948                                     btrfs_file_extent_num_bytes(leaf, fi);
3949                         } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
3950                                 item_end += btrfs_file_extent_inline_len(leaf,
3951                                                                          fi);
3952                         }
3953                         item_end--;
3954                 }
3955                 if (found_type > min_type) {
3956                         del_item = 1;
3957                 } else {
3958                         if (item_end < new_size)
3959                                 break;
3960                         if (found_key.offset >= new_size)
3961                                 del_item = 1;
3962                         else
3963                                 del_item = 0;
3964                 }
3965                 found_extent = 0;
3966                 /* FIXME, shrink the extent if the ref count is only 1 */
3967                 if (found_type != BTRFS_EXTENT_DATA_KEY)
3968                         goto delete;
3969
3970                 if (del_item)
3971                         last_size = found_key.offset;
3972                 else
3973                         last_size = new_size;
3974
3975                 if (extent_type != BTRFS_FILE_EXTENT_INLINE) {
3976                         u64 num_dec;
3977                         extent_start = btrfs_file_extent_disk_bytenr(leaf, fi);
3978                         if (!del_item) {
3979                                 u64 orig_num_bytes =
3980                                         btrfs_file_extent_num_bytes(leaf, fi);
3981                                 extent_num_bytes = ALIGN(new_size -
3982                                                 found_key.offset,
3983                                                 root->sectorsize);
3984                                 btrfs_set_file_extent_num_bytes(leaf, fi,
3985                                                          extent_num_bytes);
3986                                 num_dec = (orig_num_bytes -
3987                                            extent_num_bytes);
3988                                 if (root->ref_cows && extent_start != 0)
3989                                         inode_sub_bytes(inode, num_dec);
3990                                 btrfs_mark_buffer_dirty(leaf);
3991                         } else {
3992                                 extent_num_bytes =
3993                                         btrfs_file_extent_disk_num_bytes(leaf,
3994                                                                          fi);
3995                                 extent_offset = found_key.offset -
3996                                         btrfs_file_extent_offset(leaf, fi);
3997
3998                                 /* FIXME blocksize != 4096 */
3999                                 num_dec = btrfs_file_extent_num_bytes(leaf, fi);
4000                                 if (extent_start != 0) {
4001                                         found_extent = 1;
4002                                         if (root->ref_cows)
4003                                                 inode_sub_bytes(inode, num_dec);
4004                                 }
4005                         }
4006                 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
4007                         /*
4008                          * we can't truncate inline items that have had
4009                          * special encodings
4010                          */
4011                         if (!del_item &&
4012                             btrfs_file_extent_compression(leaf, fi) == 0 &&
4013                             btrfs_file_extent_encryption(leaf, fi) == 0 &&
4014                             btrfs_file_extent_other_encoding(leaf, fi) == 0) {
4015                                 u32 size = new_size - found_key.offset;
4016
4017                                 if (root->ref_cows) {
4018                                         inode_sub_bytes(inode, item_end + 1 -
4019                                                         new_size);
4020                                 }
4021                                 size =
4022                                     btrfs_file_extent_calc_inline_size(size);
4023                                 btrfs_truncate_item(root, path, size, 1);
4024                         } else if (root->ref_cows) {
4025                                 inode_sub_bytes(inode, item_end + 1 -
4026                                                 found_key.offset);
4027                         }
4028                 }
4029 delete:
4030                 if (del_item) {
4031                         if (!pending_del_nr) {
4032                                 /* no pending yet, add ourselves */
4033                                 pending_del_slot = path->slots[0];
4034                                 pending_del_nr = 1;
4035                         } else if (pending_del_nr &&
4036                                    path->slots[0] + 1 == pending_del_slot) {
4037                                 /* hop on the pending chunk */
4038                                 pending_del_nr++;
4039                                 pending_del_slot = path->slots[0];
4040                         } else {
4041                                 BUG();
4042                         }
4043                 } else {
4044                         break;
4045                 }
4046                 if (found_extent && (root->ref_cows ||
4047                                      root == root->fs_info->tree_root)) {
4048                         btrfs_set_path_blocking(path);
4049                         ret = btrfs_free_extent(trans, root, extent_start,
4050                                                 extent_num_bytes, 0,
4051                                                 btrfs_header_owner(leaf),
4052                                                 ino, extent_offset, 0);
4053                         BUG_ON(ret);
4054                 }
4055
4056                 if (found_type == BTRFS_INODE_ITEM_KEY)
4057                         break;
4058
4059                 if (path->slots[0] == 0 ||
4060                     path->slots[0] != pending_del_slot) {
4061                         if (pending_del_nr) {
4062                                 ret = btrfs_del_items(trans, root, path,
4063                                                 pending_del_slot,
4064                                                 pending_del_nr);
4065                                 if (ret) {
4066                                         btrfs_abort_transaction(trans,
4067                                                                 root, ret);
4068                                         goto error;
4069                                 }
4070                                 pending_del_nr = 0;
4071                         }
4072                         btrfs_release_path(path);
4073                         goto search_again;
4074                 } else {
4075                         path->slots[0]--;
4076                 }
4077         }
4078 out:
4079         if (pending_del_nr) {
4080                 ret = btrfs_del_items(trans, root, path, pending_del_slot,
4081                                       pending_del_nr);
4082                 if (ret)
4083                         btrfs_abort_transaction(trans, root, ret);
4084         }
4085 error:
4086         if (last_size != (u64)-1)
4087                 btrfs_ordered_update_i_size(inode, last_size, NULL);
4088         btrfs_free_path(path);
4089         return err;
4090 }
4091
4092 /*
4093  * btrfs_truncate_page - read, zero a chunk and write a page
4094  * @inode - inode that we're zeroing
4095  * @from - the offset to start zeroing
4096  * @len - the length to zero, 0 to zero the entire range respective to the
4097  *      offset
4098  * @front - zero up to the offset instead of from the offset on
4099  *
4100  * This will find the page for the "from" offset and cow the page and zero the
4101  * part we want to zero.  This is used with truncate and hole punching.
4102  */
4103 int btrfs_truncate_page(struct inode *inode, loff_t from, loff_t len,
4104                         int front)
4105 {
4106         struct address_space *mapping = inode->i_mapping;
4107         struct btrfs_root *root = BTRFS_I(inode)->root;
4108         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
4109         struct btrfs_ordered_extent *ordered;
4110         struct extent_state *cached_state = NULL;
4111         char *kaddr;
4112         u32 blocksize = root->sectorsize;
4113         pgoff_t index = from >> PAGE_CACHE_SHIFT;
4114         unsigned offset = from & (PAGE_CACHE_SIZE-1);
4115         struct page *page;
4116         gfp_t mask = btrfs_alloc_write_mask(mapping);
4117         int ret = 0;
4118         u64 page_start;
4119         u64 page_end;
4120
4121         if ((offset & (blocksize - 1)) == 0 &&
4122             (!len || ((len & (blocksize - 1)) == 0)))
4123                 goto out;
4124         ret = btrfs_delalloc_reserve_space(inode, PAGE_CACHE_SIZE);
4125         if (ret)
4126                 goto out;
4127
4128 again:
4129         page = find_or_create_page(mapping, index, mask);
4130         if (!page) {
4131                 btrfs_delalloc_release_space(inode, PAGE_CACHE_SIZE);
4132                 ret = -ENOMEM;
4133                 goto out;
4134         }
4135
4136         page_start = page_offset(page);
4137         page_end = page_start + PAGE_CACHE_SIZE - 1;
4138
4139         if (!PageUptodate(page)) {
4140                 ret = btrfs_readpage(NULL, page);
4141                 lock_page(page);
4142                 if (page->mapping != mapping) {
4143                         unlock_page(page);
4144                         page_cache_release(page);
4145                         goto again;
4146                 }
4147                 if (!PageUptodate(page)) {
4148                         ret = -EIO;
4149                         goto out_unlock;
4150                 }
4151         }
4152         wait_on_page_writeback(page);
4153
4154         lock_extent_bits(io_tree, page_start, page_end, 0, &cached_state);
4155         set_page_extent_mapped(page);
4156
4157         ordered = btrfs_lookup_ordered_extent(inode, page_start);
4158         if (ordered) {
4159                 unlock_extent_cached(io_tree, page_start, page_end,
4160                                      &cached_state, GFP_NOFS);
4161                 unlock_page(page);
4162                 page_cache_release(page);
4163                 btrfs_start_ordered_extent(inode, ordered, 1);
4164                 btrfs_put_ordered_extent(ordered);
4165                 goto again;
4166         }
4167
4168         clear_extent_bit(&BTRFS_I(inode)->io_tree, page_start, page_end,
4169                           EXTENT_DIRTY | EXTENT_DELALLOC |
4170                           EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG,
4171                           0, 0, &cached_state, GFP_NOFS);
4172
4173         ret = btrfs_set_extent_delalloc(inode, page_start, page_end,
4174                                         &cached_state);
4175         if (ret) {
4176                 unlock_extent_cached(io_tree, page_start, page_end,
4177                                      &cached_state, GFP_NOFS);
4178                 goto out_unlock;
4179         }
4180
4181         if (offset != PAGE_CACHE_SIZE) {
4182                 if (!len)
4183                         len = PAGE_CACHE_SIZE - offset;
4184                 kaddr = kmap(page);
4185                 if (front)
4186                         memset(kaddr, 0, offset);
4187                 else
4188                         memset(kaddr + offset, 0, len);
4189                 flush_dcache_page(page);
4190                 kunmap(page);
4191         }
4192         ClearPageChecked(page);
4193         set_page_dirty(page);
4194         unlock_extent_cached(io_tree, page_start, page_end, &cached_state,
4195                              GFP_NOFS);
4196
4197 out_unlock:
4198         if (ret)
4199                 btrfs_delalloc_release_space(inode, PAGE_CACHE_SIZE);
4200         unlock_page(page);
4201         page_cache_release(page);
4202 out:
4203         return ret;
4204 }
4205
4206 static int maybe_insert_hole(struct btrfs_root *root, struct inode *inode,
4207                              u64 offset, u64 len)
4208 {
4209         struct btrfs_trans_handle *trans;
4210         int ret;
4211
4212         /*
4213          * Still need to make sure the inode looks like it's been updated so
4214          * that any holes get logged if we fsync.
4215          */
4216         if (btrfs_fs_incompat(root->fs_info, NO_HOLES)) {
4217                 BTRFS_I(inode)->last_trans = root->fs_info->generation;
4218                 BTRFS_I(inode)->last_sub_trans = root->log_transid;
4219                 BTRFS_I(inode)->last_log_commit = root->last_log_commit;
4220                 return 0;
4221         }
4222
4223         /*
4224          * 1 - for the one we're dropping
4225          * 1 - for the one we're adding
4226          * 1 - for updating the inode.
4227          */
4228         trans = btrfs_start_transaction(root, 3);
4229         if (IS_ERR(trans))
4230                 return PTR_ERR(trans);
4231
4232         ret = btrfs_drop_extents(trans, root, inode, offset, offset + len, 1);
4233         if (ret) {
4234                 btrfs_abort_transaction(trans, root, ret);
4235                 btrfs_end_transaction(trans, root);
4236                 return ret;
4237         }
4238
4239         ret = btrfs_insert_file_extent(trans, root, btrfs_ino(inode), offset,
4240                                        0, 0, len, 0, len, 0, 0, 0);
4241         if (ret)
4242                 btrfs_abort_transaction(trans, root, ret);
4243         else
4244                 btrfs_update_inode(trans, root, inode);
4245         btrfs_end_transaction(trans, root);
4246         return ret;
4247 }
4248
4249 /*
4250  * This function puts in dummy file extents for the area we're creating a hole
4251  * for.  So if we are truncating this file to a larger size we need to insert
4252  * these file extents so that btrfs_get_extent will return a EXTENT_MAP_HOLE for
4253  * the range between oldsize and size
4254  */
4255 int btrfs_cont_expand(struct inode *inode, loff_t oldsize, loff_t size)
4256 {
4257         struct btrfs_root *root = BTRFS_I(inode)->root;
4258         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
4259         struct extent_map *em = NULL;
4260         struct extent_state *cached_state = NULL;
4261         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
4262         u64 hole_start = ALIGN(oldsize, root->sectorsize);
4263         u64 block_end = ALIGN(size, root->sectorsize);
4264         u64 last_byte;
4265         u64 cur_offset;
4266         u64 hole_size;
4267         int err = 0;
4268
4269         /*
4270          * If our size started in the middle of a page we need to zero out the
4271          * rest of the page before we expand the i_size, otherwise we could
4272          * expose stale data.
4273          */
4274         err = btrfs_truncate_page(inode, oldsize, 0, 0);
4275         if (err)
4276                 return err;
4277
4278         if (size <= hole_start)
4279                 return 0;
4280
4281         while (1) {
4282                 struct btrfs_ordered_extent *ordered;
4283
4284                 lock_extent_bits(io_tree, hole_start, block_end - 1, 0,
4285                                  &cached_state);
4286                 ordered = btrfs_lookup_ordered_range(inode, hole_start,
4287                                                      block_end - hole_start);
4288                 if (!ordered)
4289                         break;
4290                 unlock_extent_cached(io_tree, hole_start, block_end - 1,
4291                                      &cached_state, GFP_NOFS);
4292                 btrfs_start_ordered_extent(inode, ordered, 1);
4293                 btrfs_put_ordered_extent(ordered);
4294         }
4295
4296         cur_offset = hole_start;
4297         while (1) {
4298                 em = btrfs_get_extent(inode, NULL, 0, cur_offset,
4299                                 block_end - cur_offset, 0);
4300                 if (IS_ERR(em)) {
4301                         err = PTR_ERR(em);
4302                         em = NULL;
4303                         break;
4304                 }
4305                 last_byte = min(extent_map_end(em), block_end);
4306                 last_byte = ALIGN(last_byte , root->sectorsize);
4307                 if (!test_bit(EXTENT_FLAG_PREALLOC, &em->flags)) {
4308                         struct extent_map *hole_em;
4309                         hole_size = last_byte - cur_offset;
4310
4311                         err = maybe_insert_hole(root, inode, cur_offset,
4312                                                 hole_size);
4313                         if (err)
4314                                 break;
4315                         btrfs_drop_extent_cache(inode, cur_offset,
4316                                                 cur_offset + hole_size - 1, 0);
4317                         hole_em = alloc_extent_map();
4318                         if (!hole_em) {
4319                                 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
4320                                         &BTRFS_I(inode)->runtime_flags);
4321                                 goto next;
4322                         }
4323                         hole_em->start = cur_offset;
4324                         hole_em->len = hole_size;
4325                         hole_em->orig_start = cur_offset;
4326
4327                         hole_em->block_start = EXTENT_MAP_HOLE;
4328                         hole_em->block_len = 0;
4329                         hole_em->orig_block_len = 0;
4330                         hole_em->ram_bytes = hole_size;
4331                         hole_em->bdev = root->fs_info->fs_devices->latest_bdev;
4332                         hole_em->compress_type = BTRFS_COMPRESS_NONE;
4333                         hole_em->generation = root->fs_info->generation;
4334
4335                         while (1) {
4336                                 write_lock(&em_tree->lock);
4337                                 err = add_extent_mapping(em_tree, hole_em, 1);
4338                                 write_unlock(&em_tree->lock);
4339                                 if (err != -EEXIST)
4340                                         break;
4341                                 btrfs_drop_extent_cache(inode, cur_offset,
4342                                                         cur_offset +
4343                                                         hole_size - 1, 0);
4344                         }
4345                         free_extent_map(hole_em);
4346                 }
4347 next:
4348                 free_extent_map(em);
4349                 em = NULL;
4350                 cur_offset = last_byte;
4351                 if (cur_offset >= block_end)
4352                         break;
4353         }
4354         free_extent_map(em);
4355         unlock_extent_cached(io_tree, hole_start, block_end - 1, &cached_state,
4356                              GFP_NOFS);
4357         return err;
4358 }
4359
4360 static int btrfs_setsize(struct inode *inode, struct iattr *attr)
4361 {
4362         struct btrfs_root *root = BTRFS_I(inode)->root;
4363         struct btrfs_trans_handle *trans;
4364         loff_t oldsize = i_size_read(inode);
4365         loff_t newsize = attr->ia_size;
4366         int mask = attr->ia_valid;
4367         int ret;
4368
4369         /*
4370          * The regular truncate() case without ATTR_CTIME and ATTR_MTIME is a
4371          * special case where we need to update the times despite not having
4372          * these flags set.  For all other operations the VFS set these flags
4373          * explicitly if it wants a timestamp update.
4374          */
4375         if (newsize != oldsize && (!(mask & (ATTR_CTIME | ATTR_MTIME))))
4376                 inode->i_ctime = inode->i_mtime = current_fs_time(inode->i_sb);
4377
4378         if (newsize > oldsize) {
4379                 truncate_pagecache(inode, newsize);
4380                 ret = btrfs_cont_expand(inode, oldsize, newsize);
4381                 if (ret)
4382                         return ret;
4383
4384                 trans = btrfs_start_transaction(root, 1);
4385                 if (IS_ERR(trans))
4386                         return PTR_ERR(trans);
4387
4388                 i_size_write(inode, newsize);
4389                 btrfs_ordered_update_i_size(inode, i_size_read(inode), NULL);
4390                 ret = btrfs_update_inode(trans, root, inode);
4391                 btrfs_end_transaction(trans, root);
4392         } else {
4393
4394                 /*
4395                  * We're truncating a file that used to have good data down to
4396                  * zero. Make sure it gets into the ordered flush list so that
4397                  * any new writes get down to disk quickly.
4398                  */
4399                 if (newsize == 0)
4400                         set_bit(BTRFS_INODE_ORDERED_DATA_CLOSE,
4401                                 &BTRFS_I(inode)->runtime_flags);
4402
4403                 /*
4404                  * 1 for the orphan item we're going to add
4405                  * 1 for the orphan item deletion.
4406                  */
4407                 trans = btrfs_start_transaction(root, 2);
4408                 if (IS_ERR(trans))
4409                         return PTR_ERR(trans);
4410
4411                 /*
4412                  * We need to do this in case we fail at _any_ point during the
4413                  * actual truncate.  Once we do the truncate_setsize we could
4414                  * invalidate pages which forces any outstanding ordered io to
4415                  * be instantly completed which will give us extents that need
4416                  * to be truncated.  If we fail to get an orphan inode down we
4417                  * could have left over extents that were never meant to live,
4418                  * so we need to garuntee from this point on that everything
4419                  * will be consistent.
4420                  */
4421                 ret = btrfs_orphan_add(trans, inode);
4422                 btrfs_end_transaction(trans, root);
4423                 if (ret)
4424                         return ret;
4425
4426                 /* we don't support swapfiles, so vmtruncate shouldn't fail */
4427                 truncate_setsize(inode, newsize);
4428
4429                 /* Disable nonlocked read DIO to avoid the end less truncate */
4430                 btrfs_inode_block_unlocked_dio(inode);
4431                 inode_dio_wait(inode);
4432                 btrfs_inode_resume_unlocked_dio(inode);
4433
4434                 ret = btrfs_truncate(inode);
4435                 if (ret && inode->i_nlink) {
4436                         int err;
4437
4438                         /*
4439                          * failed to truncate, disk_i_size is only adjusted down
4440                          * as we remove extents, so it should represent the true
4441                          * size of the inode, so reset the in memory size and
4442                          * delete our orphan entry.
4443                          */
4444                         trans = btrfs_join_transaction(root);
4445                         if (IS_ERR(trans)) {
4446                                 btrfs_orphan_del(NULL, inode);
4447                                 return ret;
4448                         }
4449                         i_size_write(inode, BTRFS_I(inode)->disk_i_size);
4450                         err = btrfs_orphan_del(trans, inode);
4451                         if (err)
4452                                 btrfs_abort_transaction(trans, root, err);
4453                         btrfs_end_transaction(trans, root);
4454                 }
4455         }
4456
4457         return ret;
4458 }
4459
4460 static int btrfs_setattr(struct dentry *dentry, struct iattr *attr)
4461 {
4462         struct inode *inode = dentry->d_inode;
4463         struct btrfs_root *root = BTRFS_I(inode)->root;
4464         int err;
4465
4466         if (btrfs_root_readonly(root))
4467                 return -EROFS;
4468
4469         err = inode_change_ok(inode, attr);
4470         if (err)
4471                 return err;
4472
4473         if (S_ISREG(inode->i_mode) && (attr->ia_valid & ATTR_SIZE)) {
4474                 err = btrfs_setsize(inode, attr);
4475                 if (err)
4476                         return err;
4477         }
4478
4479         if (attr->ia_valid) {
4480                 setattr_copy(inode, attr);
4481                 inode_inc_iversion(inode);
4482                 err = btrfs_dirty_inode(inode);
4483
4484                 if (!err && attr->ia_valid & ATTR_MODE)
4485                         err = btrfs_acl_chmod(inode);
4486         }
4487
4488         return err;
4489 }
4490
4491 void btrfs_evict_inode(struct inode *inode)
4492 {
4493         struct btrfs_trans_handle *trans;
4494         struct btrfs_root *root = BTRFS_I(inode)->root;
4495         struct btrfs_block_rsv *rsv, *global_rsv;
4496         u64 min_size = btrfs_calc_trunc_metadata_size(root, 1);
4497         int ret;
4498
4499         trace_btrfs_inode_evict(inode);
4500
4501         truncate_inode_pages(&inode->i_data, 0);
4502         if (inode->i_nlink &&
4503             ((btrfs_root_refs(&root->root_item) != 0 &&
4504               root->root_key.objectid != BTRFS_ROOT_TREE_OBJECTID) ||
4505              btrfs_is_free_space_inode(inode)))
4506                 goto no_delete;
4507
4508         if (is_bad_inode(inode)) {
4509                 btrfs_orphan_del(NULL, inode);
4510                 goto no_delete;
4511         }
4512         /* do we really want it for ->i_nlink > 0 and zero btrfs_root_refs? */
4513         btrfs_wait_ordered_range(inode, 0, (u64)-1);
4514
4515         if (root->fs_info->log_root_recovering) {
4516                 BUG_ON(test_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
4517                                  &BTRFS_I(inode)->runtime_flags));
4518                 goto no_delete;
4519         }
4520
4521         if (inode->i_nlink > 0) {
4522                 BUG_ON(btrfs_root_refs(&root->root_item) != 0 &&
4523                        root->root_key.objectid != BTRFS_ROOT_TREE_OBJECTID);
4524                 goto no_delete;
4525         }
4526
4527         ret = btrfs_commit_inode_delayed_inode(inode);
4528         if (ret) {
4529                 btrfs_orphan_del(NULL, inode);
4530                 goto no_delete;
4531         }
4532
4533         rsv = btrfs_alloc_block_rsv(root, BTRFS_BLOCK_RSV_TEMP);
4534         if (!rsv) {
4535                 btrfs_orphan_del(NULL, inode);
4536                 goto no_delete;
4537         }
4538         rsv->size = min_size;
4539         rsv->failfast = 1;
4540         global_rsv = &root->fs_info->global_block_rsv;
4541
4542         btrfs_i_size_write(inode, 0);
4543
4544         /*
4545          * This is a bit simpler than btrfs_truncate since we've already
4546          * reserved our space for our orphan item in the unlink, so we just
4547          * need to reserve some slack space in case we add bytes and update
4548          * inode item when doing the truncate.
4549          */
4550         while (1) {
4551                 ret = btrfs_block_rsv_refill(root, rsv, min_size,
4552                                              BTRFS_RESERVE_FLUSH_LIMIT);
4553
4554                 /*
4555                  * Try and steal from the global reserve since we will
4556                  * likely not use this space anyway, we want to try as
4557                  * hard as possible to get this to work.
4558                  */
4559                 if (ret)
4560                         ret = btrfs_block_rsv_migrate(global_rsv, rsv, min_size);
4561
4562                 if (ret) {
4563                         btrfs_warn(root->fs_info,
4564                                 "Could not get space for a delete, will truncate on mount %d",
4565                                 ret);
4566                         btrfs_orphan_del(NULL, inode);
4567                         btrfs_free_block_rsv(root, rsv);
4568                         goto no_delete;
4569                 }
4570
4571                 trans = btrfs_join_transaction(root);
4572                 if (IS_ERR(trans)) {
4573                         btrfs_orphan_del(NULL, inode);
4574                         btrfs_free_block_rsv(root, rsv);
4575                         goto no_delete;
4576                 }
4577
4578                 trans->block_rsv = rsv;
4579
4580                 ret = btrfs_truncate_inode_items(trans, root, inode, 0, 0);
4581                 if (ret != -ENOSPC)
4582                         break;
4583
4584                 trans->block_rsv = &root->fs_info->trans_block_rsv;
4585                 btrfs_end_transaction(trans, root);
4586                 trans = NULL;
4587                 btrfs_btree_balance_dirty(root);
4588         }
4589
4590         btrfs_free_block_rsv(root, rsv);
4591
4592         /*
4593          * Errors here aren't a big deal, it just means we leave orphan items
4594          * in the tree.  They will be cleaned up on the next mount.
4595          */
4596         if (ret == 0) {
4597                 trans->block_rsv = root->orphan_block_rsv;
4598                 btrfs_orphan_del(trans, inode);
4599         } else {
4600                 btrfs_orphan_del(NULL, inode);
4601         }
4602
4603         trans->block_rsv = &root->fs_info->trans_block_rsv;
4604         if (!(root == root->fs_info->tree_root ||
4605               root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID))
4606                 btrfs_return_ino(root, btrfs_ino(inode));
4607
4608         btrfs_end_transaction(trans, root);
4609         btrfs_btree_balance_dirty(root);
4610 no_delete:
4611         btrfs_remove_delayed_node(inode);
4612         clear_inode(inode);
4613         return;
4614 }
4615
4616 /*
4617  * this returns the key found in the dir entry in the location pointer.
4618  * If no dir entries were found, location->objectid is 0.
4619  */
4620 static int btrfs_inode_by_name(struct inode *dir, struct dentry *dentry,
4621                                struct btrfs_key *location)
4622 {
4623         const char *name = dentry->d_name.name;
4624         int namelen = dentry->d_name.len;
4625         struct btrfs_dir_item *di;
4626         struct btrfs_path *path;
4627         struct btrfs_root *root = BTRFS_I(dir)->root;
4628         int ret = 0;
4629
4630         path = btrfs_alloc_path();
4631         if (!path)
4632                 return -ENOMEM;
4633
4634         di = btrfs_lookup_dir_item(NULL, root, path, btrfs_ino(dir), name,
4635                                     namelen, 0);
4636         if (IS_ERR(di))
4637                 ret = PTR_ERR(di);
4638
4639         if (IS_ERR_OR_NULL(di))
4640                 goto out_err;
4641
4642         btrfs_dir_item_key_to_cpu(path->nodes[0], di, location);
4643 out:
4644         btrfs_free_path(path);
4645         return ret;
4646 out_err:
4647         location->objectid = 0;
4648         goto out;
4649 }
4650
4651 /*
4652  * when we hit a tree root in a directory, the btrfs part of the inode
4653  * needs to be changed to reflect the root directory of the tree root.  This
4654  * is kind of like crossing a mount point.
4655  */
4656 static int fixup_tree_root_location(struct btrfs_root *root,
4657                                     struct inode *dir,
4658                                     struct dentry *dentry,
4659                                     struct btrfs_key *location,
4660                                     struct btrfs_root **sub_root)
4661 {
4662         struct btrfs_path *path;
4663         struct btrfs_root *new_root;
4664         struct btrfs_root_ref *ref;
4665         struct extent_buffer *leaf;
4666         int ret;
4667         int err = 0;
4668
4669         path = btrfs_alloc_path();
4670         if (!path) {
4671                 err = -ENOMEM;
4672                 goto out;
4673         }
4674
4675         err = -ENOENT;
4676         ret = btrfs_find_item(root->fs_info->tree_root, path,
4677                                 BTRFS_I(dir)->root->root_key.objectid,
4678                                 location->objectid, BTRFS_ROOT_REF_KEY, NULL);
4679         if (ret) {
4680                 if (ret < 0)
4681                         err = ret;
4682                 goto out;
4683         }
4684
4685         leaf = path->nodes[0];
4686         ref = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_root_ref);
4687         if (btrfs_root_ref_dirid(leaf, ref) != btrfs_ino(dir) ||
4688             btrfs_root_ref_name_len(leaf, ref) != dentry->d_name.len)
4689                 goto out;
4690
4691         ret = memcmp_extent_buffer(leaf, dentry->d_name.name,
4692                                    (unsigned long)(ref + 1),
4693                                    dentry->d_name.len);
4694         if (ret)
4695                 goto out;
4696
4697         btrfs_release_path(path);
4698
4699         new_root = btrfs_read_fs_root_no_name(root->fs_info, location);
4700         if (IS_ERR(new_root)) {
4701                 err = PTR_ERR(new_root);
4702                 goto out;
4703         }
4704
4705         *sub_root = new_root;
4706         location->objectid = btrfs_root_dirid(&new_root->root_item);
4707         location->type = BTRFS_INODE_ITEM_KEY;
4708         location->offset = 0;
4709         err = 0;
4710 out:
4711         btrfs_free_path(path);
4712         return err;
4713 }
4714
4715 static void inode_tree_add(struct inode *inode)
4716 {
4717         struct btrfs_root *root = BTRFS_I(inode)->root;
4718         struct btrfs_inode *entry;
4719         struct rb_node **p;
4720         struct rb_node *parent;
4721         struct rb_node *new = &BTRFS_I(inode)->rb_node;
4722         u64 ino = btrfs_ino(inode);
4723
4724         if (inode_unhashed(inode))
4725                 return;
4726         parent = NULL;
4727         spin_lock(&root->inode_lock);
4728         p = &root->inode_tree.rb_node;
4729         while (*p) {
4730                 parent = *p;
4731                 entry = rb_entry(parent, struct btrfs_inode, rb_node);
4732
4733                 if (ino < btrfs_ino(&entry->vfs_inode))
4734                         p = &parent->rb_left;
4735                 else if (ino > btrfs_ino(&entry->vfs_inode))
4736                         p = &parent->rb_right;
4737                 else {
4738                         WARN_ON(!(entry->vfs_inode.i_state &
4739                                   (I_WILL_FREE | I_FREEING)));
4740                         rb_replace_node(parent, new, &root->inode_tree);
4741                         RB_CLEAR_NODE(parent);
4742                         spin_unlock(&root->inode_lock);
4743                         return;
4744                 }
4745         }
4746         rb_link_node(new, parent, p);
4747         rb_insert_color(new, &root->inode_tree);
4748         spin_unlock(&root->inode_lock);
4749 }
4750
4751 static void inode_tree_del(struct inode *inode)
4752 {
4753         struct btrfs_root *root = BTRFS_I(inode)->root;
4754         int empty = 0;
4755
4756         spin_lock(&root->inode_lock);
4757         if (!RB_EMPTY_NODE(&BTRFS_I(inode)->rb_node)) {
4758                 rb_erase(&BTRFS_I(inode)->rb_node, &root->inode_tree);
4759                 RB_CLEAR_NODE(&BTRFS_I(inode)->rb_node);
4760                 empty = RB_EMPTY_ROOT(&root->inode_tree);
4761         }
4762         spin_unlock(&root->inode_lock);
4763
4764         if (empty && btrfs_root_refs(&root->root_item) == 0) {
4765                 synchronize_srcu(&root->fs_info->subvol_srcu);
4766                 spin_lock(&root->inode_lock);
4767                 empty = RB_EMPTY_ROOT(&root->inode_tree);
4768                 spin_unlock(&root->inode_lock);
4769                 if (empty)
4770                         btrfs_add_dead_root(root);
4771         }
4772 }
4773
4774 void btrfs_invalidate_inodes(struct btrfs_root *root)
4775 {
4776         struct rb_node *node;
4777         struct rb_node *prev;
4778         struct btrfs_inode *entry;
4779         struct inode *inode;
4780         u64 objectid = 0;
4781
4782         WARN_ON(btrfs_root_refs(&root->root_item) != 0);
4783
4784         spin_lock(&root->inode_lock);
4785 again:
4786         node = root->inode_tree.rb_node;
4787         prev = NULL;
4788         while (node) {
4789                 prev = node;
4790                 entry = rb_entry(node, struct btrfs_inode, rb_node);
4791
4792                 if (objectid < btrfs_ino(&entry->vfs_inode))
4793                         node = node->rb_left;
4794                 else if (objectid > btrfs_ino(&entry->vfs_inode))
4795                         node = node->rb_right;
4796                 else
4797                         break;
4798         }
4799         if (!node) {
4800                 while (prev) {
4801                         entry = rb_entry(prev, struct btrfs_inode, rb_node);
4802                         if (objectid <= btrfs_ino(&entry->vfs_inode)) {
4803                                 node = prev;
4804                                 break;
4805                         }
4806                         prev = rb_next(prev);
4807                 }
4808         }
4809         while (node) {
4810                 entry = rb_entry(node, struct btrfs_inode, rb_node);
4811                 objectid = btrfs_ino(&entry->vfs_inode) + 1;
4812                 inode = igrab(&entry->vfs_inode);
4813                 if (inode) {
4814                         spin_unlock(&root->inode_lock);
4815                         if (atomic_read(&inode->i_count) > 1)
4816                                 d_prune_aliases(inode);
4817                         /*
4818                          * btrfs_drop_inode will have it removed from
4819                          * the inode cache when its usage count
4820                          * hits zero.
4821                          */
4822                         iput(inode);
4823                         cond_resched();
4824                         spin_lock(&root->inode_lock);
4825                         goto again;
4826                 }
4827
4828                 if (cond_resched_lock(&root->inode_lock))
4829                         goto again;
4830
4831                 node = rb_next(node);
4832         }
4833         spin_unlock(&root->inode_lock);
4834 }
4835
4836 static int btrfs_init_locked_inode(struct inode *inode, void *p)
4837 {
4838         struct btrfs_iget_args *args = p;
4839         inode->i_ino = args->ino;
4840         BTRFS_I(inode)->root = args->root;
4841         return 0;
4842 }
4843
4844 static int btrfs_find_actor(struct inode *inode, void *opaque)
4845 {
4846         struct btrfs_iget_args *args = opaque;
4847         return args->ino == btrfs_ino(inode) &&
4848                 args->root == BTRFS_I(inode)->root;
4849 }
4850
4851 static struct inode *btrfs_iget_locked(struct super_block *s,
4852                                        u64 objectid,
4853                                        struct btrfs_root *root)
4854 {
4855         struct inode *inode;
4856         struct btrfs_iget_args args;
4857         unsigned long hashval = btrfs_inode_hash(objectid, root);
4858
4859         args.ino = objectid;
4860         args.root = root;
4861
4862         inode = iget5_locked(s, hashval, btrfs_find_actor,
4863                              btrfs_init_locked_inode,
4864                              (void *)&args);
4865         return inode;
4866 }
4867
4868 /* Get an inode object given its location and corresponding root.
4869  * Returns in *is_new if the inode was read from disk
4870  */
4871 struct inode *btrfs_iget(struct super_block *s, struct btrfs_key *location,
4872                          struct btrfs_root *root, int *new)
4873 {
4874         struct inode *inode;
4875
4876         inode = btrfs_iget_locked(s, location->objectid, root);
4877         if (!inode)
4878                 return ERR_PTR(-ENOMEM);
4879
4880         if (inode->i_state & I_NEW) {
4881                 BTRFS_I(inode)->root = root;
4882                 memcpy(&BTRFS_I(inode)->location, location, sizeof(*location));
4883                 btrfs_read_locked_inode(inode);
4884                 if (!is_bad_inode(inode)) {
4885                         inode_tree_add(inode);
4886                         unlock_new_inode(inode);
4887                         if (new)
4888                                 *new = 1;
4889                 } else {
4890                         unlock_new_inode(inode);
4891                         iput(inode);
4892                         inode = ERR_PTR(-ESTALE);
4893                 }
4894         }
4895
4896         return inode;
4897 }
4898
4899 static struct inode *new_simple_dir(struct super_block *s,
4900                                     struct btrfs_key *key,
4901                                     struct btrfs_root *root)
4902 {
4903         struct inode *inode = new_inode(s);
4904
4905         if (!inode)
4906                 return ERR_PTR(-ENOMEM);
4907
4908         BTRFS_I(inode)->root = root;
4909         memcpy(&BTRFS_I(inode)->location, key, sizeof(*key));
4910         set_bit(BTRFS_INODE_DUMMY, &BTRFS_I(inode)->runtime_flags);
4911
4912         inode->i_ino = BTRFS_EMPTY_SUBVOL_DIR_OBJECTID;
4913         inode->i_op = &btrfs_dir_ro_inode_operations;
4914         inode->i_fop = &simple_dir_operations;
4915         inode->i_mode = S_IFDIR | S_IRUGO | S_IWUSR | S_IXUGO;
4916         inode->i_mtime = inode->i_atime = inode->i_ctime = CURRENT_TIME;
4917
4918         return inode;
4919 }
4920
4921 struct inode *btrfs_lookup_dentry(struct inode *dir, struct dentry *dentry)
4922 {
4923         struct inode *inode;
4924         struct btrfs_root *root = BTRFS_I(dir)->root;
4925         struct btrfs_root *sub_root = root;
4926         struct btrfs_key location;
4927         int index;
4928         int ret = 0;
4929
4930         if (dentry->d_name.len > BTRFS_NAME_LEN)
4931                 return ERR_PTR(-ENAMETOOLONG);
4932
4933         ret = btrfs_inode_by_name(dir, dentry, &location);
4934         if (ret < 0)
4935                 return ERR_PTR(ret);
4936
4937         if (location.objectid == 0)
4938                 return NULL;
4939
4940         if (location.type == BTRFS_INODE_ITEM_KEY) {
4941                 inode = btrfs_iget(dir->i_sb, &location, root, NULL);
4942                 return inode;
4943         }
4944
4945         BUG_ON(location.type != BTRFS_ROOT_ITEM_KEY);
4946
4947         index = srcu_read_lock(&root->fs_info->subvol_srcu);
4948         ret = fixup_tree_root_location(root, dir, dentry,
4949                                        &location, &sub_root);
4950         if (ret < 0) {
4951                 if (ret != -ENOENT)
4952                         inode = ERR_PTR(ret);
4953                 else
4954                         inode = new_simple_dir(dir->i_sb, &location, sub_root);
4955         } else {
4956                 inode = btrfs_iget(dir->i_sb, &location, sub_root, NULL);
4957         }
4958         srcu_read_unlock(&root->fs_info->subvol_srcu, index);
4959
4960         if (!IS_ERR(inode) && root != sub_root) {
4961                 down_read(&root->fs_info->cleanup_work_sem);
4962                 if (!(inode->i_sb->s_flags & MS_RDONLY))
4963                         ret = btrfs_orphan_cleanup(sub_root);
4964                 up_read(&root->fs_info->cleanup_work_sem);
4965                 if (ret) {
4966                         iput(inode);
4967                         inode = ERR_PTR(ret);
4968                 }
4969         }
4970
4971         return inode;
4972 }
4973
4974 static int btrfs_dentry_delete(const struct dentry *dentry)
4975 {
4976         struct btrfs_root *root;
4977         struct inode *inode = dentry->d_inode;
4978
4979         if (!inode && !IS_ROOT(dentry))
4980                 inode = dentry->d_parent->d_inode;
4981
4982         if (inode) {
4983                 root = BTRFS_I(inode)->root;
4984                 if (btrfs_root_refs(&root->root_item) == 0)
4985                         return 1;
4986
4987                 if (btrfs_ino(inode) == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)
4988                         return 1;
4989         }
4990         return 0;
4991 }
4992
4993 static void btrfs_dentry_release(struct dentry *dentry)
4994 {
4995         if (dentry->d_fsdata)
4996                 kfree(dentry->d_fsdata);
4997 }
4998
4999 static struct dentry *btrfs_lookup(struct inode *dir, struct dentry *dentry,
5000                                    unsigned int flags)
5001 {
5002         struct dentry *ret;
5003
5004         ret = d_splice_alias(btrfs_lookup_dentry(dir, dentry), dentry);
5005         return ret;
5006 }
5007
5008 unsigned char btrfs_filetype_table[] = {
5009         DT_UNKNOWN, DT_REG, DT_DIR, DT_CHR, DT_BLK, DT_FIFO, DT_SOCK, DT_LNK
5010 };
5011
5012 static int btrfs_real_readdir(struct file *file, struct dir_context *ctx)
5013 {
5014         struct inode *inode = file_inode(file);
5015         struct btrfs_root *root = BTRFS_I(inode)->root;
5016         struct btrfs_item *item;
5017         struct btrfs_dir_item *di;
5018         struct btrfs_key key;
5019         struct btrfs_key found_key;
5020         struct btrfs_path *path;
5021         struct list_head ins_list;
5022         struct list_head del_list;
5023         int ret;
5024         struct extent_buffer *leaf;
5025         int slot;
5026         unsigned char d_type;
5027         int over = 0;
5028         u32 di_cur;
5029         u32 di_total;
5030         u32 di_len;
5031         int key_type = BTRFS_DIR_INDEX_KEY;
5032         char tmp_name[32];
5033         char *name_ptr;
5034         int name_len;
5035         int is_curr = 0;        /* ctx->pos points to the current index? */
5036
5037         /* FIXME, use a real flag for deciding about the key type */
5038         if (root->fs_info->tree_root == root)
5039                 key_type = BTRFS_DIR_ITEM_KEY;
5040
5041         if (!dir_emit_dots(file, ctx))
5042                 return 0;
5043
5044         path = btrfs_alloc_path();
5045         if (!path)
5046                 return -ENOMEM;
5047
5048         path->reada = 1;
5049
5050         if (key_type == BTRFS_DIR_INDEX_KEY) {
5051                 INIT_LIST_HEAD(&ins_list);
5052                 INIT_LIST_HEAD(&del_list);
5053                 btrfs_get_delayed_items(inode, &ins_list, &del_list);
5054         }
5055
5056         btrfs_set_key_type(&key, key_type);
5057         key.offset = ctx->pos;
5058         key.objectid = btrfs_ino(inode);
5059
5060         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
5061         if (ret < 0)
5062                 goto err;
5063
5064         while (1) {
5065                 leaf = path->nodes[0];
5066                 slot = path->slots[0];
5067                 if (slot >= btrfs_header_nritems(leaf)) {
5068                         ret = btrfs_next_leaf(root, path);
5069                         if (ret < 0)
5070                                 goto err;
5071                         else if (ret > 0)
5072                                 break;
5073                         continue;
5074                 }
5075
5076                 item = btrfs_item_nr(slot);
5077                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
5078
5079                 if (found_key.objectid != key.objectid)
5080                         break;
5081                 if (btrfs_key_type(&found_key) != key_type)
5082                         break;
5083                 if (found_key.offset < ctx->pos)
5084                         goto next;
5085                 if (key_type == BTRFS_DIR_INDEX_KEY &&
5086                     btrfs_should_delete_dir_index(&del_list,
5087                                                   found_key.offset))
5088                         goto next;
5089
5090                 ctx->pos = found_key.offset;
5091                 is_curr = 1;
5092
5093                 di = btrfs_item_ptr(leaf, slot, struct btrfs_dir_item);
5094                 di_cur = 0;
5095                 di_total = btrfs_item_size(leaf, item);
5096
5097                 while (di_cur < di_total) {
5098                         struct btrfs_key location;
5099
5100                         if (verify_dir_item(root, leaf, di))
5101                                 break;
5102
5103                         name_len = btrfs_dir_name_len(leaf, di);
5104                         if (name_len <= sizeof(tmp_name)) {
5105                                 name_ptr = tmp_name;
5106                         } else {
5107                                 name_ptr = kmalloc(name_len, GFP_NOFS);
5108                                 if (!name_ptr) {
5109                                         ret = -ENOMEM;
5110                                         goto err;
5111                                 }
5112                         }
5113                         read_extent_buffer(leaf, name_ptr,
5114                                            (unsigned long)(di + 1), name_len);
5115
5116                         d_type = btrfs_filetype_table[btrfs_dir_type(leaf, di)];
5117                         btrfs_dir_item_key_to_cpu(leaf, di, &location);
5118
5119
5120                         /* is this a reference to our own snapshot? If so
5121                          * skip it.
5122                          *
5123                          * In contrast to old kernels, we insert the snapshot's
5124                          * dir item and dir index after it has been created, so
5125                          * we won't find a reference to our own snapshot. We
5126                          * still keep the following code for backward
5127                          * compatibility.
5128                          */
5129                         if (location.type == BTRFS_ROOT_ITEM_KEY &&
5130                             location.objectid == root->root_key.objectid) {
5131                                 over = 0;
5132                                 goto skip;
5133                         }
5134                         over = !dir_emit(ctx, name_ptr, name_len,
5135                                        location.objectid, d_type);
5136
5137 skip:
5138                         if (name_ptr != tmp_name)
5139                                 kfree(name_ptr);
5140
5141                         if (over)
5142                                 goto nopos;
5143                         di_len = btrfs_dir_name_len(leaf, di) +
5144                                  btrfs_dir_data_len(leaf, di) + sizeof(*di);
5145                         di_cur += di_len;
5146                         di = (struct btrfs_dir_item *)((char *)di + di_len);
5147                 }
5148 next:
5149                 path->slots[0]++;
5150         }
5151
5152         if (key_type == BTRFS_DIR_INDEX_KEY) {
5153                 if (is_curr)
5154                         ctx->pos++;
5155                 ret = btrfs_readdir_delayed_dir_index(ctx, &ins_list);
5156                 if (ret)
5157                         goto nopos;
5158         }
5159
5160         /* Reached end of directory/root. Bump pos past the last item. */
5161         ctx->pos++;
5162
5163         /*
5164          * Stop new entries from being returned after we return the last
5165          * entry.
5166          *
5167          * New directory entries are assigned a strictly increasing
5168          * offset.  This means that new entries created during readdir
5169          * are *guaranteed* to be seen in the future by that readdir.
5170          * This has broken buggy programs which operate on names as
5171          * they're returned by readdir.  Until we re-use freed offsets
5172          * we have this hack to stop new entries from being returned
5173          * under the assumption that they'll never reach this huge
5174          * offset.
5175          *
5176          * This is being careful not to overflow 32bit loff_t unless the
5177          * last entry requires it because doing so has broken 32bit apps
5178          * in the past.
5179          */
5180         if (key_type == BTRFS_DIR_INDEX_KEY) {
5181                 if (ctx->pos >= INT_MAX)
5182                         ctx->pos = LLONG_MAX;
5183                 else
5184                         ctx->pos = INT_MAX;
5185         }
5186 nopos:
5187         ret = 0;
5188 err:
5189         if (key_type == BTRFS_DIR_INDEX_KEY)
5190                 btrfs_put_delayed_items(&ins_list, &del_list);
5191         btrfs_free_path(path);
5192         return ret;
5193 }
5194
5195 int btrfs_write_inode(struct inode *inode, struct writeback_control *wbc)
5196 {
5197         struct btrfs_root *root = BTRFS_I(inode)->root;
5198         struct btrfs_trans_handle *trans;
5199         int ret = 0;
5200         bool nolock = false;
5201
5202         if (test_bit(BTRFS_INODE_DUMMY, &BTRFS_I(inode)->runtime_flags))
5203                 return 0;
5204
5205         if (btrfs_fs_closing(root->fs_info) && btrfs_is_free_space_inode(inode))
5206                 nolock = true;
5207
5208         if (wbc->sync_mode == WB_SYNC_ALL) {
5209                 if (nolock)
5210                         trans = btrfs_join_transaction_nolock(root);
5211                 else
5212                         trans = btrfs_join_transaction(root);
5213                 if (IS_ERR(trans))
5214                         return PTR_ERR(trans);
5215                 ret = btrfs_commit_transaction(trans, root);
5216         }
5217         return ret;
5218 }
5219
5220 /*
5221  * This is somewhat expensive, updating the tree every time the
5222  * inode changes.  But, it is most likely to find the inode in cache.
5223  * FIXME, needs more benchmarking...there are no reasons other than performance
5224  * to keep or drop this code.
5225  */
5226 static int btrfs_dirty_inode(struct inode *inode)
5227 {
5228         struct btrfs_root *root = BTRFS_I(inode)->root;
5229         struct btrfs_trans_handle *trans;
5230         int ret;
5231
5232         if (test_bit(BTRFS_INODE_DUMMY, &BTRFS_I(inode)->runtime_flags))
5233                 return 0;
5234
5235         trans = btrfs_join_transaction(root);
5236         if (IS_ERR(trans))
5237                 return PTR_ERR(trans);
5238
5239         ret = btrfs_update_inode(trans, root, inode);
5240         if (ret && ret == -ENOSPC) {
5241                 /* whoops, lets try again with the full transaction */
5242                 btrfs_end_transaction(trans, root);
5243                 trans = btrfs_start_transaction(root, 1);
5244                 if (IS_ERR(trans))
5245                         return PTR_ERR(trans);
5246
5247                 ret = btrfs_update_inode(trans, root, inode);
5248         }
5249         btrfs_end_transaction(trans, root);
5250         if (BTRFS_I(inode)->delayed_node)
5251                 btrfs_balance_delayed_items(root);
5252
5253         return ret;
5254 }
5255
5256 /*
5257  * This is a copy of file_update_time.  We need this so we can return error on
5258  * ENOSPC for updating the inode in the case of file write and mmap writes.
5259  */
5260 static int btrfs_update_time(struct inode *inode, struct timespec *now,
5261                              int flags)
5262 {
5263         struct btrfs_root *root = BTRFS_I(inode)->root;
5264
5265         if (btrfs_root_readonly(root))
5266                 return -EROFS;
5267
5268         if (flags & S_VERSION)
5269                 inode_inc_iversion(inode);
5270         if (flags & S_CTIME)
5271                 inode->i_ctime = *now;
5272         if (flags & S_MTIME)
5273                 inode->i_mtime = *now;
5274         if (flags & S_ATIME)
5275                 inode->i_atime = *now;
5276         return btrfs_dirty_inode(inode);
5277 }
5278
5279 /*
5280  * find the highest existing sequence number in a directory
5281  * and then set the in-memory index_cnt variable to reflect
5282  * free sequence numbers
5283  */
5284 static int btrfs_set_inode_index_count(struct inode *inode)
5285 {
5286         struct btrfs_root *root = BTRFS_I(inode)->root;
5287         struct btrfs_key key, found_key;
5288         struct btrfs_path *path;
5289         struct extent_buffer *leaf;
5290         int ret;
5291
5292         key.objectid = btrfs_ino(inode);
5293         btrfs_set_key_type(&key, BTRFS_DIR_INDEX_KEY);
5294         key.offset = (u64)-1;
5295
5296         path = btrfs_alloc_path();
5297         if (!path)
5298                 return -ENOMEM;
5299
5300         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
5301         if (ret < 0)
5302                 goto out;
5303         /* FIXME: we should be able to handle this */
5304         if (ret == 0)
5305                 goto out;
5306         ret = 0;
5307
5308         /*
5309          * MAGIC NUMBER EXPLANATION:
5310          * since we search a directory based on f_pos we have to start at 2
5311          * since '.' and '..' have f_pos of 0 and 1 respectively, so everybody
5312          * else has to start at 2
5313          */
5314         if (path->slots[0] == 0) {
5315                 BTRFS_I(inode)->index_cnt = 2;
5316                 goto out;
5317         }
5318
5319         path->slots[0]--;
5320
5321         leaf = path->nodes[0];
5322         btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
5323
5324         if (found_key.objectid != btrfs_ino(inode) ||
5325             btrfs_key_type(&found_key) != BTRFS_DIR_INDEX_KEY) {
5326                 BTRFS_I(inode)->index_cnt = 2;
5327                 goto out;
5328         }
5329
5330         BTRFS_I(inode)->index_cnt = found_key.offset + 1;
5331 out:
5332         btrfs_free_path(path);
5333         return ret;
5334 }
5335
5336 /*
5337  * helper to find a free sequence number in a given directory.  This current
5338  * code is very simple, later versions will do smarter things in the btree
5339  */
5340 int btrfs_set_inode_index(struct inode *dir, u64 *index)
5341 {
5342         int ret = 0;
5343
5344         if (BTRFS_I(dir)->index_cnt == (u64)-1) {
5345                 ret = btrfs_inode_delayed_dir_index_count(dir);
5346                 if (ret) {
5347                         ret = btrfs_set_inode_index_count(dir);
5348                         if (ret)
5349                                 return ret;
5350                 }
5351         }
5352
5353         *index = BTRFS_I(dir)->index_cnt;
5354         BTRFS_I(dir)->index_cnt++;
5355
5356         return ret;
5357 }
5358
5359 static struct inode *btrfs_new_inode(struct btrfs_trans_handle *trans,
5360                                      struct btrfs_root *root,
5361                                      struct inode *dir,
5362                                      const char *name, int name_len,
5363                                      u64 ref_objectid, u64 objectid,
5364                                      umode_t mode, u64 *index)
5365 {
5366         struct inode *inode;
5367         struct btrfs_inode_item *inode_item;
5368         struct btrfs_key *location;
5369         struct btrfs_path *path;
5370         struct btrfs_inode_ref *ref;
5371         struct btrfs_key key[2];
5372         u32 sizes[2];
5373         unsigned long ptr;
5374         int ret;
5375
5376         path = btrfs_alloc_path();
5377         if (!path)
5378                 return ERR_PTR(-ENOMEM);
5379
5380         inode = new_inode(root->fs_info->sb);
5381         if (!inode) {
5382                 btrfs_free_path(path);
5383                 return ERR_PTR(-ENOMEM);
5384         }
5385
5386         /*
5387          * we have to initialize this early, so we can reclaim the inode
5388          * number if we fail afterwards in this function.
5389          */
5390         inode->i_ino = objectid;
5391
5392         if (dir) {
5393                 trace_btrfs_inode_request(dir);
5394
5395                 ret = btrfs_set_inode_index(dir, index);
5396                 if (ret) {
5397                         btrfs_free_path(path);
5398                         iput(inode);
5399                         return ERR_PTR(ret);
5400                 }
5401         }
5402         /*
5403          * index_cnt is ignored for everything but a dir,
5404          * btrfs_get_inode_index_count has an explanation for the magic
5405          * number
5406          */
5407         BTRFS_I(inode)->index_cnt = 2;
5408         BTRFS_I(inode)->root = root;
5409         BTRFS_I(inode)->generation = trans->transid;
5410         inode->i_generation = BTRFS_I(inode)->generation;
5411
5412         /*
5413          * We could have gotten an inode number from somebody who was fsynced
5414          * and then removed in this same transaction, so let's just set full
5415          * sync since it will be a full sync anyway and this will blow away the
5416          * old info in the log.
5417          */
5418         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &BTRFS_I(inode)->runtime_flags);
5419
5420         key[0].objectid = objectid;
5421         btrfs_set_key_type(&key[0], BTRFS_INODE_ITEM_KEY);
5422         key[0].offset = 0;
5423
5424         /*
5425          * Start new inodes with an inode_ref. This is slightly more
5426          * efficient for small numbers of hard links since they will
5427          * be packed into one item. Extended refs will kick in if we
5428          * add more hard links than can fit in the ref item.
5429          */
5430         key[1].objectid = objectid;
5431         btrfs_set_key_type(&key[1], BTRFS_INODE_REF_KEY);
5432         key[1].offset = ref_objectid;
5433
5434         sizes[0] = sizeof(struct btrfs_inode_item);
5435         sizes[1] = name_len + sizeof(*ref);
5436
5437         path->leave_spinning = 1;
5438         ret = btrfs_insert_empty_items(trans, root, path, key, sizes, 2);
5439         if (ret != 0)
5440                 goto fail;
5441
5442         inode_init_owner(inode, dir, mode);
5443         inode_set_bytes(inode, 0);
5444         inode->i_mtime = inode->i_atime = inode->i_ctime = CURRENT_TIME;
5445         inode_item = btrfs_item_ptr(path->nodes[0], path->slots[0],
5446                                   struct btrfs_inode_item);
5447         memset_extent_buffer(path->nodes[0], 0, (unsigned long)inode_item,
5448                              sizeof(*inode_item));
5449         fill_inode_item(trans, path->nodes[0], inode_item, inode);
5450
5451         ref = btrfs_item_ptr(path->nodes[0], path->slots[0] + 1,
5452                              struct btrfs_inode_ref);
5453         btrfs_set_inode_ref_name_len(path->nodes[0], ref, name_len);
5454         btrfs_set_inode_ref_index(path->nodes[0], ref, *index);
5455         ptr = (unsigned long)(ref + 1);
5456         write_extent_buffer(path->nodes[0], name, ptr, name_len);
5457
5458         btrfs_mark_buffer_dirty(path->nodes[0]);
5459         btrfs_free_path(path);
5460
5461         location = &BTRFS_I(inode)->location;
5462         location->objectid = objectid;
5463         location->offset = 0;
5464         btrfs_set_key_type(location, BTRFS_INODE_ITEM_KEY);
5465
5466         btrfs_inherit_iflags(inode, dir);
5467
5468         if (S_ISREG(mode)) {
5469                 if (btrfs_test_opt(root, NODATASUM))
5470                         BTRFS_I(inode)->flags |= BTRFS_INODE_NODATASUM;
5471                 if (btrfs_test_opt(root, NODATACOW))
5472                         BTRFS_I(inode)->flags |= BTRFS_INODE_NODATACOW |
5473                                 BTRFS_INODE_NODATASUM;
5474         }
5475
5476         btrfs_insert_inode_hash(inode);
5477         inode_tree_add(inode);
5478
5479         trace_btrfs_inode_new(inode);
5480         btrfs_set_inode_last_trans(trans, inode);
5481
5482         btrfs_update_root_times(trans, root);
5483
5484         return inode;
5485 fail:
5486         if (dir)
5487                 BTRFS_I(dir)->index_cnt--;
5488         btrfs_free_path(path);
5489         iput(inode);
5490         return ERR_PTR(ret);
5491 }
5492
5493 static inline u8 btrfs_inode_type(struct inode *inode)
5494 {
5495         return btrfs_type_by_mode[(inode->i_mode & S_IFMT) >> S_SHIFT];
5496 }
5497
5498 /*
5499  * utility function to add 'inode' into 'parent_inode' with
5500  * a give name and a given sequence number.
5501  * if 'add_backref' is true, also insert a backref from the
5502  * inode to the parent directory.
5503  */
5504 int btrfs_add_link(struct btrfs_trans_handle *trans,
5505                    struct inode *parent_inode, struct inode *inode,
5506                    const char *name, int name_len, int add_backref, u64 index)
5507 {
5508         int ret = 0;
5509         struct btrfs_key key;
5510         struct btrfs_root *root = BTRFS_I(parent_inode)->root;
5511         u64 ino = btrfs_ino(inode);
5512         u64 parent_ino = btrfs_ino(parent_inode);
5513
5514         if (unlikely(ino == BTRFS_FIRST_FREE_OBJECTID)) {
5515                 memcpy(&key, &BTRFS_I(inode)->root->root_key, sizeof(key));
5516         } else {
5517                 key.objectid = ino;
5518                 btrfs_set_key_type(&key, BTRFS_INODE_ITEM_KEY);
5519                 key.offset = 0;
5520         }
5521
5522         if (unlikely(ino == BTRFS_FIRST_FREE_OBJECTID)) {
5523                 ret = btrfs_add_root_ref(trans, root->fs_info->tree_root,
5524                                          key.objectid, root->root_key.objectid,
5525                                          parent_ino, index, name, name_len);
5526         } else if (add_backref) {
5527                 ret = btrfs_insert_inode_ref(trans, root, name, name_len, ino,
5528                                              parent_ino, index);
5529         }
5530
5531         /* Nothing to clean up yet */
5532         if (ret)
5533                 return ret;
5534
5535         ret = btrfs_insert_dir_item(trans, root, name, name_len,
5536                                     parent_inode, &key,
5537                                     btrfs_inode_type(inode), index);
5538         if (ret == -EEXIST || ret == -EOVERFLOW)
5539                 goto fail_dir_item;
5540         else if (ret) {
5541                 btrfs_abort_transaction(trans, root, ret);
5542                 return ret;
5543         }
5544
5545         btrfs_i_size_write(parent_inode, parent_inode->i_size +
5546                            name_len * 2);
5547         inode_inc_iversion(parent_inode);
5548         parent_inode->i_mtime = parent_inode->i_ctime = CURRENT_TIME;
5549         ret = btrfs_update_inode(trans, root, parent_inode);
5550         if (ret)
5551                 btrfs_abort_transaction(trans, root, ret);
5552         return ret;
5553
5554 fail_dir_item:
5555         if (unlikely(ino == BTRFS_FIRST_FREE_OBJECTID)) {
5556                 u64 local_index;
5557                 int err;
5558                 err = btrfs_del_root_ref(trans, root->fs_info->tree_root,
5559                                  key.objectid, root->root_key.objectid,
5560                                  parent_ino, &local_index, name, name_len);
5561
5562         } else if (add_backref) {
5563                 u64 local_index;
5564                 int err;
5565
5566                 err = btrfs_del_inode_ref(trans, root, name, name_len,
5567                                           ino, parent_ino, &local_index);
5568         }
5569         return ret;
5570 }
5571
5572 static int btrfs_add_nondir(struct btrfs_trans_handle *trans,
5573                             struct inode *dir, struct dentry *dentry,
5574                             struct inode *inode, int backref, u64 index)
5575 {
5576         int err = btrfs_add_link(trans, dir, inode,
5577                                  dentry->d_name.name, dentry->d_name.len,
5578                                  backref, index);
5579         if (err > 0)
5580                 err = -EEXIST;
5581         return err;
5582 }
5583
5584 static int btrfs_mknod(struct inode *dir, struct dentry *dentry,
5585                         umode_t mode, dev_t rdev)
5586 {
5587         struct btrfs_trans_handle *trans;
5588         struct btrfs_root *root = BTRFS_I(dir)->root;
5589         struct inode *inode = NULL;
5590         int err;
5591         int drop_inode = 0;
5592         u64 objectid;
5593         u64 index = 0;
5594
5595         if (!new_valid_dev(rdev))
5596                 return -EINVAL;
5597
5598         /*
5599          * 2 for inode item and ref
5600          * 2 for dir items
5601          * 1 for xattr if selinux is on
5602          */
5603         trans = btrfs_start_transaction(root, 5);
5604         if (IS_ERR(trans))
5605                 return PTR_ERR(trans);
5606
5607         err = btrfs_find_free_ino(root, &objectid);
5608         if (err)
5609                 goto out_unlock;
5610
5611         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
5612                                 dentry->d_name.len, btrfs_ino(dir), objectid,
5613                                 mode, &index);
5614         if (IS_ERR(inode)) {
5615                 err = PTR_ERR(inode);
5616                 goto out_unlock;
5617         }
5618
5619         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
5620         if (err) {
5621                 drop_inode = 1;
5622                 goto out_unlock;
5623         }
5624
5625         /*
5626         * If the active LSM wants to access the inode during
5627         * d_instantiate it needs these. Smack checks to see
5628         * if the filesystem supports xattrs by looking at the
5629         * ops vector.
5630         */
5631
5632         inode->i_op = &btrfs_special_inode_operations;
5633         err = btrfs_add_nondir(trans, dir, dentry, inode, 0, index);
5634         if (err)
5635                 drop_inode = 1;
5636         else {
5637                 init_special_inode(inode, inode->i_mode, rdev);
5638                 btrfs_update_inode(trans, root, inode);
5639                 d_instantiate(dentry, inode);
5640         }
5641 out_unlock:
5642         btrfs_end_transaction(trans, root);
5643         btrfs_btree_balance_dirty(root);
5644         if (drop_inode) {
5645                 inode_dec_link_count(inode);
5646                 iput(inode);
5647         }
5648         return err;
5649 }
5650
5651 static int btrfs_create(struct inode *dir, struct dentry *dentry,
5652                         umode_t mode, bool excl)
5653 {
5654         struct btrfs_trans_handle *trans;
5655         struct btrfs_root *root = BTRFS_I(dir)->root;
5656         struct inode *inode = NULL;
5657         int drop_inode_on_err = 0;
5658         int err;
5659         u64 objectid;
5660         u64 index = 0;
5661
5662         /*
5663          * 2 for inode item and ref
5664          * 2 for dir items
5665          * 1 for xattr if selinux is on
5666          */
5667         trans = btrfs_start_transaction(root, 5);
5668         if (IS_ERR(trans))
5669                 return PTR_ERR(trans);
5670
5671         err = btrfs_find_free_ino(root, &objectid);
5672         if (err)
5673                 goto out_unlock;
5674
5675         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
5676                                 dentry->d_name.len, btrfs_ino(dir), objectid,
5677                                 mode, &index);
5678         if (IS_ERR(inode)) {
5679                 err = PTR_ERR(inode);
5680                 goto out_unlock;
5681         }
5682         drop_inode_on_err = 1;
5683
5684         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
5685         if (err)
5686                 goto out_unlock;
5687
5688         err = btrfs_update_inode(trans, root, inode);
5689         if (err)
5690                 goto out_unlock;
5691
5692         /*
5693         * If the active LSM wants to access the inode during
5694         * d_instantiate it needs these. Smack checks to see
5695         * if the filesystem supports xattrs by looking at the
5696         * ops vector.
5697         */
5698         inode->i_fop = &btrfs_file_operations;
5699         inode->i_op = &btrfs_file_inode_operations;
5700
5701         err = btrfs_add_nondir(trans, dir, dentry, inode, 0, index);
5702         if (err)
5703                 goto out_unlock;
5704
5705         inode->i_mapping->a_ops = &btrfs_aops;
5706         inode->i_mapping->backing_dev_info = &root->fs_info->bdi;
5707         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
5708         d_instantiate(dentry, inode);
5709
5710 out_unlock:
5711         btrfs_end_transaction(trans, root);
5712         if (err && drop_inode_on_err) {
5713                 inode_dec_link_count(inode);
5714                 iput(inode);
5715         }
5716         btrfs_btree_balance_dirty(root);
5717         return err;
5718 }
5719
5720 static int btrfs_link(struct dentry *old_dentry, struct inode *dir,
5721                       struct dentry *dentry)
5722 {
5723         struct btrfs_trans_handle *trans;
5724         struct btrfs_root *root = BTRFS_I(dir)->root;
5725         struct inode *inode = old_dentry->d_inode;
5726         u64 index;
5727         int err;
5728         int drop_inode = 0;
5729
5730         /* do not allow sys_link's with other subvols of the same device */
5731         if (root->objectid != BTRFS_I(inode)->root->objectid)
5732                 return -EXDEV;
5733
5734         if (inode->i_nlink >= BTRFS_LINK_MAX)
5735                 return -EMLINK;
5736
5737         err = btrfs_set_inode_index(dir, &index);
5738         if (err)
5739                 goto fail;
5740
5741         /*
5742          * 2 items for inode and inode ref
5743          * 2 items for dir items
5744          * 1 item for parent inode
5745          */
5746         trans = btrfs_start_transaction(root, 5);
5747         if (IS_ERR(trans)) {
5748                 err = PTR_ERR(trans);
5749                 goto fail;
5750         }
5751
5752         inc_nlink(inode);
5753         inode_inc_iversion(inode);
5754         inode->i_ctime = CURRENT_TIME;
5755         ihold(inode);
5756         set_bit(BTRFS_INODE_COPY_EVERYTHING, &BTRFS_I(inode)->runtime_flags);
5757
5758         err = btrfs_add_nondir(trans, dir, dentry, inode, 1, index);
5759
5760         if (err) {
5761                 drop_inode = 1;
5762         } else {
5763                 struct dentry *parent = dentry->d_parent;
5764                 err = btrfs_update_inode(trans, root, inode);
5765                 if (err)
5766                         goto fail;
5767                 d_instantiate(dentry, inode);
5768                 btrfs_log_new_name(trans, inode, NULL, parent);
5769         }
5770
5771         btrfs_end_transaction(trans, root);
5772 fail:
5773         if (drop_inode) {
5774                 inode_dec_link_count(inode);
5775                 iput(inode);
5776         }
5777         btrfs_btree_balance_dirty(root);
5778         return err;
5779 }
5780
5781 static int btrfs_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode)
5782 {
5783         struct inode *inode = NULL;
5784         struct btrfs_trans_handle *trans;
5785         struct btrfs_root *root = BTRFS_I(dir)->root;
5786         int err = 0;
5787         int drop_on_err = 0;
5788         u64 objectid = 0;
5789         u64 index = 0;
5790
5791         /*
5792          * 2 items for inode and ref
5793          * 2 items for dir items
5794          * 1 for xattr if selinux is on
5795          */
5796         trans = btrfs_start_transaction(root, 5);
5797         if (IS_ERR(trans))
5798                 return PTR_ERR(trans);
5799
5800         err = btrfs_find_free_ino(root, &objectid);
5801         if (err)
5802                 goto out_fail;
5803
5804         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
5805                                 dentry->d_name.len, btrfs_ino(dir), objectid,
5806                                 S_IFDIR | mode, &index);
5807         if (IS_ERR(inode)) {
5808                 err = PTR_ERR(inode);
5809                 goto out_fail;
5810         }
5811
5812         drop_on_err = 1;
5813
5814         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
5815         if (err)
5816                 goto out_fail;
5817
5818         inode->i_op = &btrfs_dir_inode_operations;
5819         inode->i_fop = &btrfs_dir_file_operations;
5820
5821         btrfs_i_size_write(inode, 0);
5822         err = btrfs_update_inode(trans, root, inode);
5823         if (err)
5824                 goto out_fail;
5825
5826         err = btrfs_add_link(trans, dir, inode, dentry->d_name.name,
5827                              dentry->d_name.len, 0, index);
5828         if (err)
5829                 goto out_fail;
5830
5831         d_instantiate(dentry, inode);
5832         drop_on_err = 0;
5833
5834 out_fail:
5835         btrfs_end_transaction(trans, root);
5836         if (drop_on_err)
5837                 iput(inode);
5838         btrfs_btree_balance_dirty(root);
5839         return err;
5840 }
5841
5842 /* helper for btfs_get_extent.  Given an existing extent in the tree,
5843  * and an extent that you want to insert, deal with overlap and insert
5844  * the new extent into the tree.
5845  */
5846 static int merge_extent_mapping(struct extent_map_tree *em_tree,
5847                                 struct extent_map *existing,
5848                                 struct extent_map *em,
5849                                 u64 map_start, u64 map_len)
5850 {
5851         u64 start_diff;
5852
5853         BUG_ON(map_start < em->start || map_start >= extent_map_end(em));
5854         start_diff = map_start - em->start;
5855         em->start = map_start;
5856         em->len = map_len;
5857         if (em->block_start < EXTENT_MAP_LAST_BYTE &&
5858             !test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
5859                 em->block_start += start_diff;
5860                 em->block_len -= start_diff;
5861         }
5862         return add_extent_mapping(em_tree, em, 0);
5863 }
5864
5865 static noinline int uncompress_inline(struct btrfs_path *path,
5866                                       struct inode *inode, struct page *page,
5867                                       size_t pg_offset, u64 extent_offset,
5868                                       struct btrfs_file_extent_item *item)
5869 {
5870         int ret;
5871         struct extent_buffer *leaf = path->nodes[0];
5872         char *tmp;
5873         size_t max_size;
5874         unsigned long inline_size;
5875         unsigned long ptr;
5876         int compress_type;
5877
5878         WARN_ON(pg_offset != 0);
5879         compress_type = btrfs_file_extent_compression(leaf, item);
5880         max_size = btrfs_file_extent_ram_bytes(leaf, item);
5881         inline_size = btrfs_file_extent_inline_item_len(leaf,
5882                                         btrfs_item_nr(path->slots[0]));
5883         tmp = kmalloc(inline_size, GFP_NOFS);
5884         if (!tmp)
5885                 return -ENOMEM;
5886         ptr = btrfs_file_extent_inline_start(item);
5887
5888         read_extent_buffer(leaf, tmp, ptr, inline_size);
5889
5890         max_size = min_t(unsigned long, PAGE_CACHE_SIZE, max_size);
5891         ret = btrfs_decompress(compress_type, tmp, page,
5892                                extent_offset, inline_size, max_size);
5893         if (ret) {
5894                 char *kaddr = kmap_atomic(page);
5895                 unsigned long copy_size = min_t(u64,
5896                                   PAGE_CACHE_SIZE - pg_offset,
5897                                   max_size - extent_offset);
5898                 memset(kaddr + pg_offset, 0, copy_size);
5899                 kunmap_atomic(kaddr);
5900         }
5901         kfree(tmp);
5902         return 0;
5903 }
5904
5905 /*
5906  * a bit scary, this does extent mapping from logical file offset to the disk.
5907  * the ugly parts come from merging extents from the disk with the in-ram
5908  * representation.  This gets more complex because of the data=ordered code,
5909  * where the in-ram extents might be locked pending data=ordered completion.
5910  *
5911  * This also copies inline extents directly into the page.
5912  */
5913
5914 struct extent_map *btrfs_get_extent(struct inode *inode, struct page *page,
5915                                     size_t pg_offset, u64 start, u64 len,
5916                                     int create)
5917 {
5918         int ret;
5919         int err = 0;
5920         u64 bytenr;
5921         u64 extent_start = 0;
5922         u64 extent_end = 0;
5923         u64 objectid = btrfs_ino(inode);
5924         u32 found_type;
5925         struct btrfs_path *path = NULL;
5926         struct btrfs_root *root = BTRFS_I(inode)->root;
5927         struct btrfs_file_extent_item *item;
5928         struct extent_buffer *leaf;
5929         struct btrfs_key found_key;
5930         struct extent_map *em = NULL;
5931         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
5932         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
5933         struct btrfs_trans_handle *trans = NULL;
5934         int compress_type;
5935
5936 again:
5937         read_lock(&em_tree->lock);
5938         em = lookup_extent_mapping(em_tree, start, len);
5939         if (em)
5940                 em->bdev = root->fs_info->fs_devices->latest_bdev;
5941         read_unlock(&em_tree->lock);
5942
5943         if (em) {
5944                 if (em->start > start || em->start + em->len <= start)
5945                         free_extent_map(em);
5946                 else if (em->block_start == EXTENT_MAP_INLINE && page)
5947                         free_extent_map(em);
5948                 else
5949                         goto out;
5950         }
5951         em = alloc_extent_map();
5952         if (!em) {
5953                 err = -ENOMEM;
5954                 goto out;
5955         }
5956         em->bdev = root->fs_info->fs_devices->latest_bdev;
5957         em->start = EXTENT_MAP_HOLE;
5958         em->orig_start = EXTENT_MAP_HOLE;
5959         em->len = (u64)-1;
5960         em->block_len = (u64)-1;
5961
5962         if (!path) {
5963                 path = btrfs_alloc_path();
5964                 if (!path) {
5965                         err = -ENOMEM;
5966                         goto out;
5967                 }
5968                 /*
5969                  * Chances are we'll be called again, so go ahead and do
5970                  * readahead
5971                  */
5972                 path->reada = 1;
5973         }
5974
5975         ret = btrfs_lookup_file_extent(trans, root, path,
5976                                        objectid, start, trans != NULL);
5977         if (ret < 0) {
5978                 err = ret;
5979                 goto out;
5980         }
5981
5982         if (ret != 0) {
5983                 if (path->slots[0] == 0)
5984                         goto not_found;
5985                 path->slots[0]--;
5986         }
5987
5988         leaf = path->nodes[0];
5989         item = btrfs_item_ptr(leaf, path->slots[0],
5990                               struct btrfs_file_extent_item);
5991         /* are we inside the extent that was found? */
5992         btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
5993         found_type = btrfs_key_type(&found_key);
5994         if (found_key.objectid != objectid ||
5995             found_type != BTRFS_EXTENT_DATA_KEY) {
5996                 /*
5997                  * If we backup past the first extent we want to move forward
5998                  * and see if there is an extent in front of us, otherwise we'll
5999                  * say there is a hole for our whole search range which can
6000                  * cause problems.
6001                  */
6002                 extent_end = start;
6003                 goto next;
6004         }
6005
6006         found_type = btrfs_file_extent_type(leaf, item);
6007         extent_start = found_key.offset;
6008         compress_type = btrfs_file_extent_compression(leaf, item);
6009         if (found_type == BTRFS_FILE_EXTENT_REG ||
6010             found_type == BTRFS_FILE_EXTENT_PREALLOC) {
6011                 extent_end = extent_start +
6012                        btrfs_file_extent_num_bytes(leaf, item);
6013         } else if (found_type == BTRFS_FILE_EXTENT_INLINE) {
6014                 size_t size;
6015                 size = btrfs_file_extent_inline_len(leaf, item);
6016                 extent_end = ALIGN(extent_start + size, root->sectorsize);
6017         }
6018 next:
6019         if (start >= extent_end) {
6020                 path->slots[0]++;
6021                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
6022                         ret = btrfs_next_leaf(root, path);
6023                         if (ret < 0) {
6024                                 err = ret;
6025                                 goto out;
6026                         }
6027                         if (ret > 0)
6028                                 goto not_found;
6029                         leaf = path->nodes[0];
6030                 }
6031                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
6032                 if (found_key.objectid != objectid ||
6033                     found_key.type != BTRFS_EXTENT_DATA_KEY)
6034                         goto not_found;
6035                 if (start + len <= found_key.offset)
6036                         goto not_found;
6037                 em->start = start;
6038                 em->orig_start = start;
6039                 em->len = found_key.offset - start;
6040                 goto not_found_em;
6041         }
6042
6043         em->ram_bytes = btrfs_file_extent_ram_bytes(leaf, item);
6044         if (found_type == BTRFS_FILE_EXTENT_REG ||
6045             found_type == BTRFS_FILE_EXTENT_PREALLOC) {
6046                 em->start = extent_start;
6047                 em->len = extent_end - extent_start;
6048                 em->orig_start = extent_start -
6049                                  btrfs_file_extent_offset(leaf, item);
6050                 em->orig_block_len = btrfs_file_extent_disk_num_bytes(leaf,
6051                                                                       item);
6052                 bytenr = btrfs_file_extent_disk_bytenr(leaf, item);
6053                 if (bytenr == 0) {
6054                         em->block_start = EXTENT_MAP_HOLE;
6055                         goto insert;
6056                 }
6057                 if (compress_type != BTRFS_COMPRESS_NONE) {
6058                         set_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
6059                         em->compress_type = compress_type;
6060                         em->block_start = bytenr;
6061                         em->block_len = em->orig_block_len;
6062                 } else {
6063                         bytenr += btrfs_file_extent_offset(leaf, item);
6064                         em->block_start = bytenr;
6065                         em->block_len = em->len;
6066                         if (found_type == BTRFS_FILE_EXTENT_PREALLOC)
6067                                 set_bit(EXTENT_FLAG_PREALLOC, &em->flags);
6068                 }
6069                 goto insert;
6070         } else if (found_type == BTRFS_FILE_EXTENT_INLINE) {
6071                 unsigned long ptr;
6072                 char *map;
6073                 size_t size;
6074                 size_t extent_offset;
6075                 size_t copy_size;
6076
6077                 em->block_start = EXTENT_MAP_INLINE;
6078                 if (!page || create) {
6079                         em->start = extent_start;
6080                         em->len = extent_end - extent_start;
6081                         goto out;
6082                 }
6083
6084                 size = btrfs_file_extent_inline_len(leaf, item);
6085                 extent_offset = page_offset(page) + pg_offset - extent_start;
6086                 copy_size = min_t(u64, PAGE_CACHE_SIZE - pg_offset,
6087                                 size - extent_offset);
6088                 em->start = extent_start + extent_offset;
6089                 em->len = ALIGN(copy_size, root->sectorsize);
6090                 em->orig_block_len = em->len;
6091                 em->orig_start = em->start;
6092                 if (compress_type) {
6093                         set_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
6094                         em->compress_type = compress_type;
6095                 }
6096                 ptr = btrfs_file_extent_inline_start(item) + extent_offset;
6097                 if (create == 0 && !PageUptodate(page)) {
6098                         if (btrfs_file_extent_compression(leaf, item) !=
6099                             BTRFS_COMPRESS_NONE) {
6100                                 ret = uncompress_inline(path, inode, page,
6101                                                         pg_offset,
6102                                                         extent_offset, item);
6103                                 BUG_ON(ret); /* -ENOMEM */
6104                         } else {
6105                                 map = kmap(page);
6106                                 read_extent_buffer(leaf, map + pg_offset, ptr,
6107                                                    copy_size);
6108                                 if (pg_offset + copy_size < PAGE_CACHE_SIZE) {
6109                                         memset(map + pg_offset + copy_size, 0,
6110                                                PAGE_CACHE_SIZE - pg_offset -
6111                                                copy_size);
6112                                 }
6113                                 kunmap(page);
6114                         }
6115                         flush_dcache_page(page);
6116                 } else if (create && PageUptodate(page)) {
6117                         BUG();
6118                         if (!trans) {
6119                                 kunmap(page);
6120                                 free_extent_map(em);
6121                                 em = NULL;
6122
6123                                 btrfs_release_path(path);
6124                                 trans = btrfs_join_transaction(root);
6125
6126                                 if (IS_ERR(trans))
6127                                         return ERR_CAST(trans);
6128                                 goto again;
6129                         }
6130                         map = kmap(page);
6131                         write_extent_buffer(leaf, map + pg_offset, ptr,
6132                                             copy_size);
6133                         kunmap(page);
6134                         btrfs_mark_buffer_dirty(leaf);
6135                 }
6136                 set_extent_uptodate(io_tree, em->start,
6137                                     extent_map_end(em) - 1, NULL, GFP_NOFS);
6138                 goto insert;
6139         } else {
6140                 WARN(1, KERN_ERR "btrfs unknown found_type %d\n", found_type);
6141         }
6142 not_found:
6143         em->start = start;
6144         em->orig_start = start;
6145         em->len = len;
6146 not_found_em:
6147         em->block_start = EXTENT_MAP_HOLE;
6148         set_bit(EXTENT_FLAG_VACANCY, &em->flags);
6149 insert:
6150         btrfs_release_path(path);
6151         if (em->start > start || extent_map_end(em) <= start) {
6152                 btrfs_err(root->fs_info, "bad extent! em: [%llu %llu] passed [%llu %llu]",
6153                         em->start, em->len, start, len);
6154                 err = -EIO;
6155                 goto out;
6156         }
6157
6158         err = 0;
6159         write_lock(&em_tree->lock);
6160         ret = add_extent_mapping(em_tree, em, 0);
6161         /* it is possible that someone inserted the extent into the tree
6162          * while we had the lock dropped.  It is also possible that
6163          * an overlapping map exists in the tree
6164          */
6165         if (ret == -EEXIST) {
6166                 struct extent_map *existing;
6167
6168                 ret = 0;
6169
6170                 existing = lookup_extent_mapping(em_tree, start, len);
6171                 if (existing && (existing->start > start ||
6172                     existing->start + existing->len <= start)) {
6173                         free_extent_map(existing);
6174                         existing = NULL;
6175                 }
6176                 if (!existing) {
6177                         existing = lookup_extent_mapping(em_tree, em->start,
6178                                                          em->len);
6179                         if (existing) {
6180                                 err = merge_extent_mapping(em_tree, existing,
6181                                                            em, start,
6182                                                            root->sectorsize);
6183                                 free_extent_map(existing);
6184                                 if (err) {
6185                                         free_extent_map(em);
6186                                         em = NULL;
6187                                 }
6188                         } else {
6189                                 err = -EIO;
6190                                 free_extent_map(em);
6191                                 em = NULL;
6192                         }
6193                 } else {
6194                         free_extent_map(em);
6195                         em = existing;
6196                         err = 0;
6197                 }
6198         }
6199         write_unlock(&em_tree->lock);
6200 out:
6201
6202         trace_btrfs_get_extent(root, em);
6203
6204         if (path)
6205                 btrfs_free_path(path);
6206         if (trans) {
6207                 ret = btrfs_end_transaction(trans, root);
6208                 if (!err)
6209                         err = ret;
6210         }
6211         if (err) {
6212                 free_extent_map(em);
6213                 return ERR_PTR(err);
6214         }
6215         BUG_ON(!em); /* Error is always set */
6216         return em;
6217 }
6218
6219 struct extent_map *btrfs_get_extent_fiemap(struct inode *inode, struct page *page,
6220                                            size_t pg_offset, u64 start, u64 len,
6221                                            int create)
6222 {
6223         struct extent_map *em;
6224         struct extent_map *hole_em = NULL;
6225         u64 range_start = start;
6226         u64 end;
6227         u64 found;
6228         u64 found_end;
6229         int err = 0;
6230
6231         em = btrfs_get_extent(inode, page, pg_offset, start, len, create);
6232         if (IS_ERR(em))
6233                 return em;
6234         if (em) {
6235                 /*
6236                  * if our em maps to
6237                  * -  a hole or
6238                  * -  a pre-alloc extent,
6239                  * there might actually be delalloc bytes behind it.
6240                  */
6241                 if (em->block_start != EXTENT_MAP_HOLE &&
6242                     !test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
6243                         return em;
6244                 else
6245                         hole_em = em;
6246         }
6247
6248         /* check to see if we've wrapped (len == -1 or similar) */
6249         end = start + len;
6250         if (end < start)
6251                 end = (u64)-1;
6252         else
6253                 end -= 1;
6254
6255         em = NULL;
6256
6257         /* ok, we didn't find anything, lets look for delalloc */
6258         found = count_range_bits(&BTRFS_I(inode)->io_tree, &range_start,
6259                                  end, len, EXTENT_DELALLOC, 1);
6260         found_end = range_start + found;
6261         if (found_end < range_start)
6262                 found_end = (u64)-1;
6263
6264         /*
6265          * we didn't find anything useful, return
6266          * the original results from get_extent()
6267          */
6268         if (range_start > end || found_end <= start) {
6269                 em = hole_em;
6270                 hole_em = NULL;
6271                 goto out;
6272         }
6273
6274         /* adjust the range_start to make sure it doesn't
6275          * go backwards from the start they passed in
6276          */
6277         range_start = max(start, range_start);
6278         found = found_end - range_start;
6279
6280         if (found > 0) {
6281                 u64 hole_start = start;
6282                 u64 hole_len = len;
6283
6284                 em = alloc_extent_map();
6285                 if (!em) {
6286                         err = -ENOMEM;
6287                         goto out;
6288                 }
6289                 /*
6290                  * when btrfs_get_extent can't find anything it
6291                  * returns one huge hole
6292                  *
6293                  * make sure what it found really fits our range, and
6294                  * adjust to make sure it is based on the start from
6295                  * the caller
6296                  */
6297                 if (hole_em) {
6298                         u64 calc_end = extent_map_end(hole_em);
6299
6300                         if (calc_end <= start || (hole_em->start > end)) {
6301                                 free_extent_map(hole_em);
6302                                 hole_em = NULL;
6303                         } else {
6304                                 hole_start = max(hole_em->start, start);
6305                                 hole_len = calc_end - hole_start;
6306                         }
6307                 }
6308                 em->bdev = NULL;
6309                 if (hole_em && range_start > hole_start) {
6310                         /* our hole starts before our delalloc, so we
6311                          * have to return just the parts of the hole
6312                          * that go until  the delalloc starts
6313                          */
6314                         em->len = min(hole_len,
6315                                       range_start - hole_start);
6316                         em->start = hole_start;
6317                         em->orig_start = hole_start;
6318                         /*
6319                          * don't adjust block start at all,
6320                          * it is fixed at EXTENT_MAP_HOLE
6321                          */
6322                         em->block_start = hole_em->block_start;
6323                         em->block_len = hole_len;
6324                         if (test_bit(EXTENT_FLAG_PREALLOC, &hole_em->flags))
6325                                 set_bit(EXTENT_FLAG_PREALLOC, &em->flags);
6326                 } else {
6327                         em->start = range_start;
6328                         em->len = found;
6329                         em->orig_start = range_start;
6330                         em->block_start = EXTENT_MAP_DELALLOC;
6331                         em->block_len = found;
6332                 }
6333         } else if (hole_em) {
6334                 return hole_em;
6335         }
6336 out:
6337
6338         free_extent_map(hole_em);
6339         if (err) {
6340                 free_extent_map(em);
6341                 return ERR_PTR(err);
6342         }
6343         return em;
6344 }
6345
6346 static struct extent_map *btrfs_new_extent_direct(struct inode *inode,
6347                                                   u64 start, u64 len)
6348 {
6349         struct btrfs_root *root = BTRFS_I(inode)->root;
6350         struct extent_map *em;
6351         struct btrfs_key ins;
6352         u64 alloc_hint;
6353         int ret;
6354
6355         alloc_hint = get_extent_allocation_hint(inode, start, len);
6356         ret = btrfs_reserve_extent(root, len, root->sectorsize, 0,
6357                                    alloc_hint, &ins, 1);
6358         if (ret)
6359                 return ERR_PTR(ret);
6360
6361         em = create_pinned_em(inode, start, ins.offset, start, ins.objectid,
6362                               ins.offset, ins.offset, ins.offset, 0);
6363         if (IS_ERR(em)) {
6364                 btrfs_free_reserved_extent(root, ins.objectid, ins.offset);
6365                 return em;
6366         }
6367
6368         ret = btrfs_add_ordered_extent_dio(inode, start, ins.objectid,
6369                                            ins.offset, ins.offset, 0);
6370         if (ret) {
6371                 btrfs_free_reserved_extent(root, ins.objectid, ins.offset);
6372                 free_extent_map(em);
6373                 return ERR_PTR(ret);
6374         }
6375
6376         return em;
6377 }
6378
6379 /*
6380  * returns 1 when the nocow is safe, < 1 on error, 0 if the
6381  * block must be cow'd
6382  */
6383 noinline int can_nocow_extent(struct inode *inode, u64 offset, u64 *len,
6384                               u64 *orig_start, u64 *orig_block_len,
6385                               u64 *ram_bytes)
6386 {
6387         struct btrfs_trans_handle *trans;
6388         struct btrfs_path *path;
6389         int ret;
6390         struct extent_buffer *leaf;
6391         struct btrfs_root *root = BTRFS_I(inode)->root;
6392         struct btrfs_file_extent_item *fi;
6393         struct btrfs_key key;
6394         u64 disk_bytenr;
6395         u64 backref_offset;
6396         u64 extent_end;
6397         u64 num_bytes;
6398         int slot;
6399         int found_type;
6400         bool nocow = (BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW);
6401         path = btrfs_alloc_path();
6402         if (!path)
6403                 return -ENOMEM;
6404
6405         ret = btrfs_lookup_file_extent(NULL, root, path, btrfs_ino(inode),
6406                                        offset, 0);
6407         if (ret < 0)
6408                 goto out;
6409
6410         slot = path->slots[0];
6411         if (ret == 1) {
6412                 if (slot == 0) {
6413                         /* can't find the item, must cow */
6414                         ret = 0;
6415                         goto out;
6416                 }
6417                 slot--;
6418         }
6419         ret = 0;
6420         leaf = path->nodes[0];
6421         btrfs_item_key_to_cpu(leaf, &key, slot);
6422         if (key.objectid != btrfs_ino(inode) ||
6423             key.type != BTRFS_EXTENT_DATA_KEY) {
6424                 /* not our file or wrong item type, must cow */
6425                 goto out;
6426         }
6427
6428         if (key.offset > offset) {
6429                 /* Wrong offset, must cow */
6430                 goto out;
6431         }
6432
6433         fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
6434         found_type = btrfs_file_extent_type(leaf, fi);
6435         if (found_type != BTRFS_FILE_EXTENT_REG &&
6436             found_type != BTRFS_FILE_EXTENT_PREALLOC) {
6437                 /* not a regular extent, must cow */
6438                 goto out;
6439         }
6440
6441         if (!nocow && found_type == BTRFS_FILE_EXTENT_REG)
6442                 goto out;
6443
6444         disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
6445         if (disk_bytenr == 0)
6446                 goto out;
6447
6448         if (btrfs_file_extent_compression(leaf, fi) ||
6449             btrfs_file_extent_encryption(leaf, fi) ||
6450             btrfs_file_extent_other_encoding(leaf, fi))
6451                 goto out;
6452
6453         backref_offset = btrfs_file_extent_offset(leaf, fi);
6454
6455         if (orig_start) {
6456                 *orig_start = key.offset - backref_offset;
6457                 *orig_block_len = btrfs_file_extent_disk_num_bytes(leaf, fi);
6458                 *ram_bytes = btrfs_file_extent_ram_bytes(leaf, fi);
6459         }
6460
6461         extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
6462
6463         if (btrfs_extent_readonly(root, disk_bytenr))
6464                 goto out;
6465         btrfs_release_path(path);
6466
6467         /*
6468          * look for other files referencing this extent, if we
6469          * find any we must cow
6470          */
6471         trans = btrfs_join_transaction(root);
6472         if (IS_ERR(trans)) {
6473                 ret = 0;
6474                 goto out;
6475         }
6476
6477         ret = btrfs_cross_ref_exist(trans, root, btrfs_ino(inode),
6478                                     key.offset - backref_offset, disk_bytenr);
6479         btrfs_end_transaction(trans, root);
6480         if (ret) {
6481                 ret = 0;
6482                 goto out;
6483         }
6484
6485         /*
6486          * adjust disk_bytenr and num_bytes to cover just the bytes
6487          * in this extent we are about to write.  If there
6488          * are any csums in that range we have to cow in order
6489          * to keep the csums correct
6490          */
6491         disk_bytenr += backref_offset;
6492         disk_bytenr += offset - key.offset;
6493         num_bytes = min(offset + *len, extent_end) - offset;
6494         if (csum_exist_in_range(root, disk_bytenr, num_bytes))
6495                                 goto out;
6496         /*
6497          * all of the above have passed, it is safe to overwrite this extent
6498          * without cow
6499          */
6500         *len = num_bytes;
6501         ret = 1;
6502 out:
6503         btrfs_free_path(path);
6504         return ret;
6505 }
6506
6507 static int lock_extent_direct(struct inode *inode, u64 lockstart, u64 lockend,
6508                               struct extent_state **cached_state, int writing)
6509 {
6510         struct btrfs_ordered_extent *ordered;
6511         int ret = 0;
6512
6513         while (1) {
6514                 lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend,
6515                                  0, cached_state);
6516                 /*
6517                  * We're concerned with the entire range that we're going to be
6518                  * doing DIO to, so we need to make sure theres no ordered
6519                  * extents in this range.
6520                  */
6521                 ordered = btrfs_lookup_ordered_range(inode, lockstart,
6522                                                      lockend - lockstart + 1);
6523
6524                 /*
6525                  * We need to make sure there are no buffered pages in this
6526                  * range either, we could have raced between the invalidate in
6527                  * generic_file_direct_write and locking the extent.  The
6528                  * invalidate needs to happen so that reads after a write do not
6529                  * get stale data.
6530                  */
6531                 if (!ordered && (!writing ||
6532                     !test_range_bit(&BTRFS_I(inode)->io_tree,
6533                                     lockstart, lockend, EXTENT_UPTODATE, 0,
6534                                     *cached_state)))
6535                         break;
6536
6537                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
6538                                      cached_state, GFP_NOFS);
6539
6540                 if (ordered) {
6541                         btrfs_start_ordered_extent(inode, ordered, 1);
6542                         btrfs_put_ordered_extent(ordered);
6543                 } else {
6544                         /* Screw you mmap */
6545                         ret = filemap_write_and_wait_range(inode->i_mapping,
6546                                                            lockstart,
6547                                                            lockend);
6548                         if (ret)
6549                                 break;
6550
6551                         /*
6552                          * If we found a page that couldn't be invalidated just
6553                          * fall back to buffered.
6554                          */
6555                         ret = invalidate_inode_pages2_range(inode->i_mapping,
6556                                         lockstart >> PAGE_CACHE_SHIFT,
6557                                         lockend >> PAGE_CACHE_SHIFT);
6558                         if (ret)
6559                                 break;
6560                 }
6561
6562                 cond_resched();
6563         }
6564
6565         return ret;
6566 }
6567
6568 static struct extent_map *create_pinned_em(struct inode *inode, u64 start,
6569                                            u64 len, u64 orig_start,
6570                                            u64 block_start, u64 block_len,
6571                                            u64 orig_block_len, u64 ram_bytes,
6572                                            int type)
6573 {
6574         struct extent_map_tree *em_tree;
6575         struct extent_map *em;
6576         struct btrfs_root *root = BTRFS_I(inode)->root;
6577         int ret;
6578
6579         em_tree = &BTRFS_I(inode)->extent_tree;
6580         em = alloc_extent_map();
6581         if (!em)
6582                 return ERR_PTR(-ENOMEM);
6583
6584         em->start = start;
6585         em->orig_start = orig_start;
6586         em->mod_start = start;
6587         em->mod_len = len;
6588         em->len = len;
6589         em->block_len = block_len;
6590         em->block_start = block_start;
6591         em->bdev = root->fs_info->fs_devices->latest_bdev;
6592         em->orig_block_len = orig_block_len;
6593         em->ram_bytes = ram_bytes;
6594         em->generation = -1;
6595         set_bit(EXTENT_FLAG_PINNED, &em->flags);
6596         if (type == BTRFS_ORDERED_PREALLOC)
6597                 set_bit(EXTENT_FLAG_FILLING, &em->flags);
6598
6599         do {
6600                 btrfs_drop_extent_cache(inode, em->start,
6601                                 em->start + em->len - 1, 0);
6602                 write_lock(&em_tree->lock);
6603                 ret = add_extent_mapping(em_tree, em, 1);
6604                 write_unlock(&em_tree->lock);
6605         } while (ret == -EEXIST);
6606
6607         if (ret) {
6608                 free_extent_map(em);
6609                 return ERR_PTR(ret);
6610         }
6611
6612         return em;
6613 }
6614
6615
6616 static int btrfs_get_blocks_direct(struct inode *inode, sector_t iblock,
6617                                    struct buffer_head *bh_result, int create)
6618 {
6619         struct extent_map *em;
6620         struct btrfs_root *root = BTRFS_I(inode)->root;
6621         struct extent_state *cached_state = NULL;
6622         u64 start = iblock << inode->i_blkbits;
6623         u64 lockstart, lockend;
6624         u64 len = bh_result->b_size;
6625         int unlock_bits = EXTENT_LOCKED;
6626         int ret = 0;
6627
6628         if (create)
6629                 unlock_bits |= EXTENT_DELALLOC | EXTENT_DIRTY;
6630         else
6631                 len = min_t(u64, len, root->sectorsize);
6632
6633         lockstart = start;
6634         lockend = start + len - 1;
6635
6636         /*
6637          * If this errors out it's because we couldn't invalidate pagecache for
6638          * this range and we need to fallback to buffered.
6639          */
6640         if (lock_extent_direct(inode, lockstart, lockend, &cached_state, create))
6641                 return -ENOTBLK;
6642
6643         em = btrfs_get_extent(inode, NULL, 0, start, len, 0);
6644         if (IS_ERR(em)) {
6645                 ret = PTR_ERR(em);
6646                 goto unlock_err;
6647         }
6648
6649         /*
6650          * Ok for INLINE and COMPRESSED extents we need to fallback on buffered
6651          * io.  INLINE is special, and we could probably kludge it in here, but
6652          * it's still buffered so for safety lets just fall back to the generic
6653          * buffered path.
6654          *
6655          * For COMPRESSED we _have_ to read the entire extent in so we can
6656          * decompress it, so there will be buffering required no matter what we
6657          * do, so go ahead and fallback to buffered.
6658          *
6659          * We return -ENOTBLK because thats what makes DIO go ahead and go back
6660          * to buffered IO.  Don't blame me, this is the price we pay for using
6661          * the generic code.
6662          */
6663         if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags) ||
6664             em->block_start == EXTENT_MAP_INLINE) {
6665                 free_extent_map(em);
6666                 ret = -ENOTBLK;
6667                 goto unlock_err;
6668         }
6669
6670         /* Just a good old fashioned hole, return */
6671         if (!create && (em->block_start == EXTENT_MAP_HOLE ||
6672                         test_bit(EXTENT_FLAG_PREALLOC, &em->flags))) {
6673                 free_extent_map(em);
6674                 goto unlock_err;
6675         }
6676
6677         /*
6678          * We don't allocate a new extent in the following cases
6679          *
6680          * 1) The inode is marked as NODATACOW.  In this case we'll just use the
6681          * existing extent.
6682          * 2) The extent is marked as PREALLOC.  We're good to go here and can
6683          * just use the extent.
6684          *
6685          */
6686         if (!create) {
6687                 len = min(len, em->len - (start - em->start));
6688                 lockstart = start + len;
6689                 goto unlock;
6690         }
6691
6692         if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags) ||
6693             ((BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW) &&
6694              em->block_start != EXTENT_MAP_HOLE)) {
6695                 int type;
6696                 int ret;
6697                 u64 block_start, orig_start, orig_block_len, ram_bytes;
6698
6699                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
6700                         type = BTRFS_ORDERED_PREALLOC;
6701                 else
6702                         type = BTRFS_ORDERED_NOCOW;
6703                 len = min(len, em->len - (start - em->start));
6704                 block_start = em->block_start + (start - em->start);
6705
6706                 if (can_nocow_extent(inode, start, &len, &orig_start,
6707                                      &orig_block_len, &ram_bytes) == 1) {
6708                         if (type == BTRFS_ORDERED_PREALLOC) {
6709                                 free_extent_map(em);
6710                                 em = create_pinned_em(inode, start, len,
6711                                                        orig_start,
6712                                                        block_start, len,
6713                                                        orig_block_len,
6714                                                        ram_bytes, type);
6715                                 if (IS_ERR(em))
6716                                         goto unlock_err;
6717                         }
6718
6719                         ret = btrfs_add_ordered_extent_dio(inode, start,
6720                                            block_start, len, len, type);
6721                         if (ret) {
6722                                 free_extent_map(em);
6723                                 goto unlock_err;
6724                         }
6725                         goto unlock;
6726                 }
6727         }
6728
6729         /*
6730          * this will cow the extent, reset the len in case we changed
6731          * it above
6732          */
6733         len = bh_result->b_size;
6734         free_extent_map(em);
6735         em = btrfs_new_extent_direct(inode, start, len);
6736         if (IS_ERR(em)) {
6737                 ret = PTR_ERR(em);
6738                 goto unlock_err;
6739         }
6740         len = min(len, em->len - (start - em->start));
6741 unlock:
6742         bh_result->b_blocknr = (em->block_start + (start - em->start)) >>
6743                 inode->i_blkbits;
6744         bh_result->b_size = len;
6745         bh_result->b_bdev = em->bdev;
6746         set_buffer_mapped(bh_result);
6747         if (create) {
6748                 if (!test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
6749                         set_buffer_new(bh_result);
6750
6751                 /*
6752                  * Need to update the i_size under the extent lock so buffered
6753                  * readers will get the updated i_size when we unlock.
6754                  */
6755                 if (start + len > i_size_read(inode))
6756                         i_size_write(inode, start + len);
6757
6758                 spin_lock(&BTRFS_I(inode)->lock);
6759                 BTRFS_I(inode)->outstanding_extents++;
6760                 spin_unlock(&BTRFS_I(inode)->lock);
6761
6762                 ret = set_extent_bit(&BTRFS_I(inode)->io_tree, lockstart,
6763                                      lockstart + len - 1, EXTENT_DELALLOC, NULL,
6764                                      &cached_state, GFP_NOFS);
6765                 BUG_ON(ret);
6766         }
6767
6768         /*
6769          * In the case of write we need to clear and unlock the entire range,
6770          * in the case of read we need to unlock only the end area that we
6771          * aren't using if there is any left over space.
6772          */
6773         if (lockstart < lockend) {
6774                 clear_extent_bit(&BTRFS_I(inode)->io_tree, lockstart,
6775                                  lockend, unlock_bits, 1, 0,
6776                                  &cached_state, GFP_NOFS);
6777         } else {
6778                 free_extent_state(cached_state);
6779         }
6780
6781         free_extent_map(em);
6782
6783         return 0;
6784
6785 unlock_err:
6786         clear_extent_bit(&BTRFS_I(inode)->io_tree, lockstart, lockend,
6787                          unlock_bits, 1, 0, &cached_state, GFP_NOFS);
6788         return ret;
6789 }
6790
6791 static void btrfs_endio_direct_read(struct bio *bio, int err)
6792 {
6793         struct btrfs_dio_private *dip = bio->bi_private;
6794         struct bio_vec *bvec_end = bio->bi_io_vec + bio->bi_vcnt - 1;
6795         struct bio_vec *bvec = bio->bi_io_vec;
6796         struct inode *inode = dip->inode;
6797         struct btrfs_root *root = BTRFS_I(inode)->root;
6798         struct bio *dio_bio;
6799         u32 *csums = (u32 *)dip->csum;
6800         int index = 0;
6801         u64 start;
6802
6803         start = dip->logical_offset;
6804         do {
6805                 if (!(BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)) {
6806                         struct page *page = bvec->bv_page;
6807                         char *kaddr;
6808                         u32 csum = ~(u32)0;
6809                         unsigned long flags;
6810
6811                         local_irq_save(flags);
6812                         kaddr = kmap_atomic(page);
6813                         csum = btrfs_csum_data(kaddr + bvec->bv_offset,
6814                                                csum, bvec->bv_len);
6815                         btrfs_csum_final(csum, (char *)&csum);
6816                         kunmap_atomic(kaddr);
6817                         local_irq_restore(flags);
6818
6819                         flush_dcache_page(bvec->bv_page);
6820                         if (csum != csums[index]) {
6821                                 btrfs_err(root->fs_info, "csum failed ino %llu off %llu csum %u expected csum %u",
6822                                           btrfs_ino(inode), start, csum,
6823                                           csums[index]);
6824                                 err = -EIO;
6825                         }
6826                 }
6827
6828                 start += bvec->bv_len;
6829                 bvec++;
6830                 index++;
6831         } while (bvec <= bvec_end);
6832
6833         unlock_extent(&BTRFS_I(inode)->io_tree, dip->logical_offset,
6834                       dip->logical_offset + dip->bytes - 1);
6835         dio_bio = dip->dio_bio;
6836
6837         kfree(dip);
6838
6839         /* If we had a csum failure make sure to clear the uptodate flag */
6840         if (err)
6841                 clear_bit(BIO_UPTODATE, &dio_bio->bi_flags);
6842         dio_end_io(dio_bio, err);
6843         bio_put(bio);
6844 }
6845
6846 static void btrfs_endio_direct_write(struct bio *bio, int err)
6847 {
6848         struct btrfs_dio_private *dip = bio->bi_private;
6849         struct inode *inode = dip->inode;
6850         struct btrfs_root *root = BTRFS_I(inode)->root;
6851         struct btrfs_ordered_extent *ordered = NULL;
6852         u64 ordered_offset = dip->logical_offset;
6853         u64 ordered_bytes = dip->bytes;
6854         struct bio *dio_bio;
6855         int ret;
6856
6857         if (err)
6858                 goto out_done;
6859 again:
6860         ret = btrfs_dec_test_first_ordered_pending(inode, &ordered,
6861                                                    &ordered_offset,
6862                                                    ordered_bytes, !err);
6863         if (!ret)
6864                 goto out_test;
6865
6866         ordered->work.func = finish_ordered_fn;
6867         ordered->work.flags = 0;
6868         btrfs_queue_worker(&root->fs_info->endio_write_workers,
6869                            &ordered->work);
6870 out_test:
6871         /*
6872          * our bio might span multiple ordered extents.  If we haven't
6873          * completed the accounting for the whole dio, go back and try again
6874          */
6875         if (ordered_offset < dip->logical_offset + dip->bytes) {
6876                 ordered_bytes = dip->logical_offset + dip->bytes -
6877                         ordered_offset;
6878                 ordered = NULL;
6879                 goto again;
6880         }
6881 out_done:
6882         dio_bio = dip->dio_bio;
6883
6884         kfree(dip);
6885
6886         /* If we had an error make sure to clear the uptodate flag */
6887         if (err)
6888                 clear_bit(BIO_UPTODATE, &dio_bio->bi_flags);
6889         dio_end_io(dio_bio, err);
6890         bio_put(bio);
6891 }
6892
6893 static int __btrfs_submit_bio_start_direct_io(struct inode *inode, int rw,
6894                                     struct bio *bio, int mirror_num,
6895                                     unsigned long bio_flags, u64 offset)
6896 {
6897         int ret;
6898         struct btrfs_root *root = BTRFS_I(inode)->root;
6899         ret = btrfs_csum_one_bio(root, inode, bio, offset, 1);
6900         BUG_ON(ret); /* -ENOMEM */
6901         return 0;
6902 }
6903
6904 static void btrfs_end_dio_bio(struct bio *bio, int err)
6905 {
6906         struct btrfs_dio_private *dip = bio->bi_private;
6907
6908         if (err) {
6909                 printk(KERN_ERR "btrfs direct IO failed ino %llu rw %lu "
6910                       "sector %#Lx len %u err no %d\n",
6911                       btrfs_ino(dip->inode), bio->bi_rw,
6912                       (unsigned long long)bio->bi_sector, bio->bi_size, err);
6913                 dip->errors = 1;
6914
6915                 /*
6916                  * before atomic variable goto zero, we must make sure
6917                  * dip->errors is perceived to be set.
6918                  */
6919                 smp_mb__before_atomic_dec();
6920         }
6921
6922         /* if there are more bios still pending for this dio, just exit */
6923         if (!atomic_dec_and_test(&dip->pending_bios))
6924                 goto out;
6925
6926         if (dip->errors) {
6927                 bio_io_error(dip->orig_bio);
6928         } else {
6929                 set_bit(BIO_UPTODATE, &dip->dio_bio->bi_flags);
6930                 bio_endio(dip->orig_bio, 0);
6931         }
6932 out:
6933         bio_put(bio);
6934 }
6935
6936 static struct bio *btrfs_dio_bio_alloc(struct block_device *bdev,
6937                                        u64 first_sector, gfp_t gfp_flags)
6938 {
6939         int nr_vecs = bio_get_nr_vecs(bdev);
6940         return btrfs_bio_alloc(bdev, first_sector, nr_vecs, gfp_flags);
6941 }
6942
6943 static inline int __btrfs_submit_dio_bio(struct bio *bio, struct inode *inode,
6944                                          int rw, u64 file_offset, int skip_sum,
6945                                          int async_submit)
6946 {
6947         struct btrfs_dio_private *dip = bio->bi_private;
6948         int write = rw & REQ_WRITE;
6949         struct btrfs_root *root = BTRFS_I(inode)->root;
6950         int ret;
6951
6952         if (async_submit)
6953                 async_submit = !atomic_read(&BTRFS_I(inode)->sync_writers);
6954
6955         bio_get(bio);
6956
6957         if (!write) {
6958                 ret = btrfs_bio_wq_end_io(root->fs_info, bio, 0);
6959                 if (ret)
6960                         goto err;
6961         }
6962
6963         if (skip_sum)
6964                 goto map;
6965
6966         if (write && async_submit) {
6967                 ret = btrfs_wq_submit_bio(root->fs_info,
6968                                    inode, rw, bio, 0, 0,
6969                                    file_offset,
6970                                    __btrfs_submit_bio_start_direct_io,
6971                                    __btrfs_submit_bio_done);
6972                 goto err;
6973         } else if (write) {
6974                 /*
6975                  * If we aren't doing async submit, calculate the csum of the
6976                  * bio now.
6977                  */
6978                 ret = btrfs_csum_one_bio(root, inode, bio, file_offset, 1);
6979                 if (ret)
6980                         goto err;
6981         } else if (!skip_sum) {
6982                 ret = btrfs_lookup_bio_sums_dio(root, inode, dip, bio,
6983                                                 file_offset);
6984                 if (ret)
6985                         goto err;
6986         }
6987
6988 map:
6989         ret = btrfs_map_bio(root, rw, bio, 0, async_submit);
6990 err:
6991         bio_put(bio);
6992         return ret;
6993 }
6994
6995 static int btrfs_submit_direct_hook(int rw, struct btrfs_dio_private *dip,
6996                                     int skip_sum)
6997 {
6998         struct inode *inode = dip->inode;
6999         struct btrfs_root *root = BTRFS_I(inode)->root;
7000         struct bio *bio;
7001         struct bio *orig_bio = dip->orig_bio;
7002         struct bio_vec *bvec = orig_bio->bi_io_vec;
7003         u64 start_sector = orig_bio->bi_sector;
7004         u64 file_offset = dip->logical_offset;
7005         u64 submit_len = 0;
7006         u64 map_length;
7007         int nr_pages = 0;
7008         int ret = 0;
7009         int async_submit = 0;
7010
7011         map_length = orig_bio->bi_size;
7012         ret = btrfs_map_block(root->fs_info, rw, start_sector << 9,
7013                               &map_length, NULL, 0);
7014         if (ret) {
7015                 bio_put(orig_bio);
7016                 return -EIO;
7017         }
7018
7019         if (map_length >= orig_bio->bi_size) {
7020                 bio = orig_bio;
7021                 goto submit;
7022         }
7023
7024         /* async crcs make it difficult to collect full stripe writes. */
7025         if (btrfs_get_alloc_profile(root, 1) &
7026             (BTRFS_BLOCK_GROUP_RAID5 | BTRFS_BLOCK_GROUP_RAID6))
7027                 async_submit = 0;
7028         else
7029                 async_submit = 1;
7030
7031         bio = btrfs_dio_bio_alloc(orig_bio->bi_bdev, start_sector, GFP_NOFS);
7032         if (!bio)
7033                 return -ENOMEM;
7034         bio->bi_private = dip;
7035         bio->bi_end_io = btrfs_end_dio_bio;
7036         atomic_inc(&dip->pending_bios);
7037
7038         while (bvec <= (orig_bio->bi_io_vec + orig_bio->bi_vcnt - 1)) {
7039                 if (unlikely(map_length < submit_len + bvec->bv_len ||
7040                     bio_add_page(bio, bvec->bv_page, bvec->bv_len,
7041                                  bvec->bv_offset) < bvec->bv_len)) {
7042                         /*
7043                          * inc the count before we submit the bio so
7044                          * we know the end IO handler won't happen before
7045                          * we inc the count. Otherwise, the dip might get freed
7046                          * before we're done setting it up
7047                          */
7048                         atomic_inc(&dip->pending_bios);
7049                         ret = __btrfs_submit_dio_bio(bio, inode, rw,
7050                                                      file_offset, skip_sum,
7051                                                      async_submit);
7052                         if (ret) {
7053                                 bio_put(bio);
7054                                 atomic_dec(&dip->pending_bios);
7055                                 goto out_err;
7056                         }
7057
7058                         start_sector += submit_len >> 9;
7059                         file_offset += submit_len;
7060
7061                         submit_len = 0;
7062                         nr_pages = 0;
7063
7064                         bio = btrfs_dio_bio_alloc(orig_bio->bi_bdev,
7065                                                   start_sector, GFP_NOFS);
7066                         if (!bio)
7067                                 goto out_err;
7068                         bio->bi_private = dip;
7069                         bio->bi_end_io = btrfs_end_dio_bio;
7070
7071                         map_length = orig_bio->bi_size;
7072                         ret = btrfs_map_block(root->fs_info, rw,
7073                                               start_sector << 9,
7074                                               &map_length, NULL, 0);
7075                         if (ret) {
7076                                 bio_put(bio);
7077                                 goto out_err;
7078                         }
7079                 } else {
7080                         submit_len += bvec->bv_len;
7081                         nr_pages++;
7082                         bvec++;
7083                 }
7084         }
7085
7086 submit:
7087         ret = __btrfs_submit_dio_bio(bio, inode, rw, file_offset, skip_sum,
7088                                      async_submit);
7089         if (!ret)
7090                 return 0;
7091
7092         bio_put(bio);
7093 out_err:
7094         dip->errors = 1;
7095         /*
7096          * before atomic variable goto zero, we must
7097          * make sure dip->errors is perceived to be set.
7098          */
7099         smp_mb__before_atomic_dec();
7100         if (atomic_dec_and_test(&dip->pending_bios))
7101                 bio_io_error(dip->orig_bio);
7102
7103         /* bio_end_io() will handle error, so we needn't return it */
7104         return 0;
7105 }
7106
7107 static void btrfs_submit_direct(int rw, struct bio *dio_bio,
7108                                 struct inode *inode, loff_t file_offset)
7109 {
7110         struct btrfs_root *root = BTRFS_I(inode)->root;
7111         struct btrfs_dio_private *dip;
7112         struct bio *io_bio;
7113         int skip_sum;
7114         int sum_len;
7115         int write = rw & REQ_WRITE;
7116         int ret = 0;
7117         u16 csum_size;
7118
7119         skip_sum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM;
7120
7121         io_bio = btrfs_bio_clone(dio_bio, GFP_NOFS);
7122         if (!io_bio) {
7123                 ret = -ENOMEM;
7124                 goto free_ordered;
7125         }
7126
7127         if (!skip_sum && !write) {
7128                 csum_size = btrfs_super_csum_size(root->fs_info->super_copy);
7129                 sum_len = dio_bio->bi_size >> inode->i_sb->s_blocksize_bits;
7130                 sum_len *= csum_size;
7131         } else {
7132                 sum_len = 0;
7133         }
7134
7135         dip = kmalloc(sizeof(*dip) + sum_len, GFP_NOFS);
7136         if (!dip) {
7137                 ret = -ENOMEM;
7138                 goto free_io_bio;
7139         }
7140
7141         dip->private = dio_bio->bi_private;
7142         dip->inode = inode;
7143         dip->logical_offset = file_offset;
7144         dip->bytes = dio_bio->bi_size;
7145         dip->disk_bytenr = (u64)dio_bio->bi_sector << 9;
7146         io_bio->bi_private = dip;
7147         dip->errors = 0;
7148         dip->orig_bio = io_bio;
7149         dip->dio_bio = dio_bio;
7150         atomic_set(&dip->pending_bios, 0);
7151
7152         if (write)
7153                 io_bio->bi_end_io = btrfs_endio_direct_write;
7154         else
7155                 io_bio->bi_end_io = btrfs_endio_direct_read;
7156
7157         ret = btrfs_submit_direct_hook(rw, dip, skip_sum);
7158         if (!ret)
7159                 return;
7160
7161 free_io_bio:
7162         bio_put(io_bio);
7163
7164 free_ordered:
7165         /*
7166          * If this is a write, we need to clean up the reserved space and kill
7167          * the ordered extent.
7168          */
7169         if (write) {
7170                 struct btrfs_ordered_extent *ordered;
7171                 ordered = btrfs_lookup_ordered_extent(inode, file_offset);
7172                 if (!test_bit(BTRFS_ORDERED_PREALLOC, &ordered->flags) &&
7173                     !test_bit(BTRFS_ORDERED_NOCOW, &ordered->flags))
7174                         btrfs_free_reserved_extent(root, ordered->start,
7175                                                    ordered->disk_len);
7176                 btrfs_put_ordered_extent(ordered);
7177                 btrfs_put_ordered_extent(ordered);
7178         }
7179         bio_endio(dio_bio, ret);
7180 }
7181
7182 static ssize_t check_direct_IO(struct btrfs_root *root, int rw, struct kiocb *iocb,
7183                         const struct iovec *iov, loff_t offset,
7184                         unsigned long nr_segs)
7185 {
7186         int seg;
7187         int i;
7188         size_t size;
7189         unsigned long addr;
7190         unsigned blocksize_mask = root->sectorsize - 1;
7191         ssize_t retval = -EINVAL;
7192         loff_t end = offset;
7193
7194         if (offset & blocksize_mask)
7195                 goto out;
7196
7197         /* Check the memory alignment.  Blocks cannot straddle pages */
7198         for (seg = 0; seg < nr_segs; seg++) {
7199                 addr = (unsigned long)iov[seg].iov_base;
7200                 size = iov[seg].iov_len;
7201                 end += size;
7202                 if ((addr & blocksize_mask) || (size & blocksize_mask))
7203                         goto out;
7204
7205                 /* If this is a write we don't need to check anymore */
7206                 if (rw & WRITE)
7207                         continue;
7208
7209                 /*
7210                  * Check to make sure we don't have duplicate iov_base's in this
7211                  * iovec, if so return EINVAL, otherwise we'll get csum errors
7212                  * when reading back.
7213                  */
7214                 for (i = seg + 1; i < nr_segs; i++) {
7215                         if (iov[seg].iov_base == iov[i].iov_base)
7216                                 goto out;
7217                 }
7218         }
7219         retval = 0;
7220 out:
7221         return retval;
7222 }
7223
7224 static ssize_t btrfs_direct_IO(int rw, struct kiocb *iocb,
7225                         const struct iovec *iov, loff_t offset,
7226                         unsigned long nr_segs)
7227 {
7228         struct file *file = iocb->ki_filp;
7229         struct inode *inode = file->f_mapping->host;
7230         size_t count = 0;
7231         int flags = 0;
7232         bool wakeup = true;
7233         bool relock = false;
7234         ssize_t ret;
7235
7236         if (check_direct_IO(BTRFS_I(inode)->root, rw, iocb, iov,
7237                             offset, nr_segs))
7238                 return 0;
7239
7240         atomic_inc(&inode->i_dio_count);
7241         smp_mb__after_atomic_inc();
7242
7243         /*
7244          * The generic stuff only does filemap_write_and_wait_range, which isn't
7245          * enough if we've written compressed pages to this area, so we need to
7246          * call btrfs_wait_ordered_range to make absolutely sure that any
7247          * outstanding dirty pages are on disk.
7248          */
7249         count = iov_length(iov, nr_segs);
7250         ret = btrfs_wait_ordered_range(inode, offset, count);
7251         if (ret)
7252                 return ret;
7253
7254         if (rw & WRITE) {
7255                 /*
7256                  * If the write DIO is beyond the EOF, we need update
7257                  * the isize, but it is protected by i_mutex. So we can
7258                  * not unlock the i_mutex at this case.
7259                  */
7260                 if (offset + count <= inode->i_size) {
7261                         mutex_unlock(&inode->i_mutex);
7262                         relock = true;
7263                 }
7264                 ret = btrfs_delalloc_reserve_space(inode, count);
7265                 if (ret)
7266                         goto out;
7267         } else if (unlikely(test_bit(BTRFS_INODE_READDIO_NEED_LOCK,
7268                                      &BTRFS_I(inode)->runtime_flags))) {
7269                 inode_dio_done(inode);
7270                 flags = DIO_LOCKING | DIO_SKIP_HOLES;
7271                 wakeup = false;
7272         }
7273
7274         ret = __blockdev_direct_IO(rw, iocb, inode,
7275                         BTRFS_I(inode)->root->fs_info->fs_devices->latest_bdev,
7276                         iov, offset, nr_segs, btrfs_get_blocks_direct, NULL,
7277                         btrfs_submit_direct, flags);
7278         if (rw & WRITE) {
7279                 if (ret < 0 && ret != -EIOCBQUEUED)
7280                         btrfs_delalloc_release_space(inode, count);
7281                 else if (ret >= 0 && (size_t)ret < count)
7282                         btrfs_delalloc_release_space(inode,
7283                                                      count - (size_t)ret);
7284                 else
7285                         btrfs_delalloc_release_metadata(inode, 0);
7286         }
7287 out:
7288         if (wakeup)
7289                 inode_dio_done(inode);
7290         if (relock)
7291                 mutex_lock(&inode->i_mutex);
7292
7293         return ret;
7294 }
7295
7296 #define BTRFS_FIEMAP_FLAGS      (FIEMAP_FLAG_SYNC)
7297
7298 static int btrfs_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
7299                 __u64 start, __u64 len)
7300 {
7301         int     ret;
7302
7303         ret = fiemap_check_flags(fieinfo, BTRFS_FIEMAP_FLAGS);
7304         if (ret)
7305                 return ret;
7306
7307         return extent_fiemap(inode, fieinfo, start, len, btrfs_get_extent_fiemap);
7308 }
7309
7310 int btrfs_readpage(struct file *file, struct page *page)
7311 {
7312         struct extent_io_tree *tree;
7313         tree = &BTRFS_I(page->mapping->host)->io_tree;
7314         return extent_read_full_page(tree, page, btrfs_get_extent, 0);
7315 }
7316
7317 static int btrfs_writepage(struct page *page, struct writeback_control *wbc)
7318 {
7319         struct extent_io_tree *tree;
7320
7321
7322         if (current->flags & PF_MEMALLOC) {
7323                 redirty_page_for_writepage(wbc, page);
7324                 unlock_page(page);
7325                 return 0;
7326         }
7327         tree = &BTRFS_I(page->mapping->host)->io_tree;
7328         return extent_write_full_page(tree, page, btrfs_get_extent, wbc);
7329 }
7330
7331 static int btrfs_writepages(struct address_space *mapping,
7332                             struct writeback_control *wbc)
7333 {
7334         struct extent_io_tree *tree;
7335
7336         tree = &BTRFS_I(mapping->host)->io_tree;
7337         return extent_writepages(tree, mapping, btrfs_get_extent, wbc);
7338 }
7339
7340 static int
7341 btrfs_readpages(struct file *file, struct address_space *mapping,
7342                 struct list_head *pages, unsigned nr_pages)
7343 {
7344         struct extent_io_tree *tree;
7345         tree = &BTRFS_I(mapping->host)->io_tree;
7346         return extent_readpages(tree, mapping, pages, nr_pages,
7347                                 btrfs_get_extent);
7348 }
7349 static int __btrfs_releasepage(struct page *page, gfp_t gfp_flags)
7350 {
7351         struct extent_io_tree *tree;
7352         struct extent_map_tree *map;
7353         int ret;
7354
7355         tree = &BTRFS_I(page->mapping->host)->io_tree;
7356         map = &BTRFS_I(page->mapping->host)->extent_tree;
7357         ret = try_release_extent_mapping(map, tree, page, gfp_flags);
7358         if (ret == 1) {
7359                 ClearPagePrivate(page);
7360                 set_page_private(page, 0);
7361                 page_cache_release(page);
7362         }
7363         return ret;
7364 }
7365
7366 static int btrfs_releasepage(struct page *page, gfp_t gfp_flags)
7367 {
7368         if (PageWriteback(page) || PageDirty(page))
7369                 return 0;
7370         return __btrfs_releasepage(page, gfp_flags & GFP_NOFS);
7371 }
7372
7373 static void btrfs_invalidatepage(struct page *page, unsigned int offset,
7374                                  unsigned int length)
7375 {
7376         struct inode *inode = page->mapping->host;
7377         struct extent_io_tree *tree;
7378         struct btrfs_ordered_extent *ordered;
7379         struct extent_state *cached_state = NULL;
7380         u64 page_start = page_offset(page);
7381         u64 page_end = page_start + PAGE_CACHE_SIZE - 1;
7382
7383         /*
7384          * we have the page locked, so new writeback can't start,
7385          * and the dirty bit won't be cleared while we are here.
7386          *
7387          * Wait for IO on this page so that we can safely clear
7388          * the PagePrivate2 bit and do ordered accounting
7389          */
7390         wait_on_page_writeback(page);
7391
7392         tree = &BTRFS_I(inode)->io_tree;
7393         if (offset) {
7394                 btrfs_releasepage(page, GFP_NOFS);
7395                 return;
7396         }
7397         lock_extent_bits(tree, page_start, page_end, 0, &cached_state);
7398         ordered = btrfs_lookup_ordered_extent(inode, page_offset(page));
7399         if (ordered) {
7400                 /*
7401                  * IO on this page will never be started, so we need
7402                  * to account for any ordered extents now
7403                  */
7404                 clear_extent_bit(tree, page_start, page_end,
7405                                  EXTENT_DIRTY | EXTENT_DELALLOC |
7406                                  EXTENT_LOCKED | EXTENT_DO_ACCOUNTING |
7407                                  EXTENT_DEFRAG, 1, 0, &cached_state, GFP_NOFS);
7408                 /*
7409                  * whoever cleared the private bit is responsible
7410                  * for the finish_ordered_io
7411                  */
7412                 if (TestClearPagePrivate2(page)) {
7413                         struct btrfs_ordered_inode_tree *tree;
7414                         u64 new_len;
7415
7416                         tree = &BTRFS_I(inode)->ordered_tree;
7417
7418                         spin_lock_irq(&tree->lock);
7419                         set_bit(BTRFS_ORDERED_TRUNCATED, &ordered->flags);
7420                         new_len = page_start - ordered->file_offset;
7421                         if (new_len < ordered->truncated_len)
7422                                 ordered->truncated_len = new_len;
7423                         spin_unlock_irq(&tree->lock);
7424
7425                         if (btrfs_dec_test_ordered_pending(inode, &ordered,
7426                                                            page_start,
7427                                                            PAGE_CACHE_SIZE, 1))
7428                                 btrfs_finish_ordered_io(ordered);
7429                 }
7430                 btrfs_put_ordered_extent(ordered);
7431                 cached_state = NULL;
7432                 lock_extent_bits(tree, page_start, page_end, 0, &cached_state);
7433         }
7434         clear_extent_bit(tree, page_start, page_end,
7435                  EXTENT_LOCKED | EXTENT_DIRTY | EXTENT_DELALLOC |
7436                  EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG, 1, 1,
7437                  &cached_state, GFP_NOFS);
7438         __btrfs_releasepage(page, GFP_NOFS);
7439
7440         ClearPageChecked(page);
7441         if (PagePrivate(page)) {
7442                 ClearPagePrivate(page);
7443                 set_page_private(page, 0);
7444                 page_cache_release(page);
7445         }
7446 }
7447
7448 /*
7449  * btrfs_page_mkwrite() is not allowed to change the file size as it gets
7450  * called from a page fault handler when a page is first dirtied. Hence we must
7451  * be careful to check for EOF conditions here. We set the page up correctly
7452  * for a written page which means we get ENOSPC checking when writing into
7453  * holes and correct delalloc and unwritten extent mapping on filesystems that
7454  * support these features.
7455  *
7456  * We are not allowed to take the i_mutex here so we have to play games to
7457  * protect against truncate races as the page could now be beyond EOF.  Because
7458  * vmtruncate() writes the inode size before removing pages, once we have the
7459  * page lock we can determine safely if the page is beyond EOF. If it is not
7460  * beyond EOF, then the page is guaranteed safe against truncation until we
7461  * unlock the page.
7462  */
7463 int btrfs_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
7464 {
7465         struct page *page = vmf->page;
7466         struct inode *inode = file_inode(vma->vm_file);
7467         struct btrfs_root *root = BTRFS_I(inode)->root;
7468         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
7469         struct btrfs_ordered_extent *ordered;
7470         struct extent_state *cached_state = NULL;
7471         char *kaddr;
7472         unsigned long zero_start;
7473         loff_t size;
7474         int ret;
7475         int reserved = 0;
7476         u64 page_start;
7477         u64 page_end;
7478
7479         sb_start_pagefault(inode->i_sb);
7480         ret  = btrfs_delalloc_reserve_space(inode, PAGE_CACHE_SIZE);
7481         if (!ret) {
7482                 ret = file_update_time(vma->vm_file);
7483                 reserved = 1;
7484         }
7485         if (ret) {
7486                 if (ret == -ENOMEM)
7487                         ret = VM_FAULT_OOM;
7488                 else /* -ENOSPC, -EIO, etc */
7489                         ret = VM_FAULT_SIGBUS;
7490                 if (reserved)
7491                         goto out;
7492                 goto out_noreserve;
7493         }
7494
7495         ret = VM_FAULT_NOPAGE; /* make the VM retry the fault */
7496 again:
7497         lock_page(page);
7498         size = i_size_read(inode);
7499         page_start = page_offset(page);
7500         page_end = page_start + PAGE_CACHE_SIZE - 1;
7501
7502         if ((page->mapping != inode->i_mapping) ||
7503             (page_start >= size)) {
7504                 /* page got truncated out from underneath us */
7505                 goto out_unlock;
7506         }
7507         wait_on_page_writeback(page);
7508
7509         lock_extent_bits(io_tree, page_start, page_end, 0, &cached_state);
7510         set_page_extent_mapped(page);
7511
7512         /*
7513          * we can't set the delalloc bits if there are pending ordered
7514          * extents.  Drop our locks and wait for them to finish
7515          */
7516         ordered = btrfs_lookup_ordered_extent(inode, page_start);
7517         if (ordered) {
7518                 unlock_extent_cached(io_tree, page_start, page_end,
7519                                      &cached_state, GFP_NOFS);
7520                 unlock_page(page);
7521                 btrfs_start_ordered_extent(inode, ordered, 1);
7522                 btrfs_put_ordered_extent(ordered);
7523                 goto again;
7524         }
7525
7526         /*
7527          * XXX - page_mkwrite gets called every time the page is dirtied, even
7528          * if it was already dirty, so for space accounting reasons we need to
7529          * clear any delalloc bits for the range we are fixing to save.  There
7530          * is probably a better way to do this, but for now keep consistent with
7531          * prepare_pages in the normal write path.
7532          */
7533         clear_extent_bit(&BTRFS_I(inode)->io_tree, page_start, page_end,
7534                           EXTENT_DIRTY | EXTENT_DELALLOC |
7535                           EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG,
7536                           0, 0, &cached_state, GFP_NOFS);
7537
7538         ret = btrfs_set_extent_delalloc(inode, page_start, page_end,
7539                                         &cached_state);
7540         if (ret) {
7541                 unlock_extent_cached(io_tree, page_start, page_end,
7542                                      &cached_state, GFP_NOFS);
7543                 ret = VM_FAULT_SIGBUS;
7544                 goto out_unlock;
7545         }
7546         ret = 0;
7547
7548         /* page is wholly or partially inside EOF */
7549         if (page_start + PAGE_CACHE_SIZE > size)
7550                 zero_start = size & ~PAGE_CACHE_MASK;
7551         else
7552                 zero_start = PAGE_CACHE_SIZE;
7553
7554         if (zero_start != PAGE_CACHE_SIZE) {
7555                 kaddr = kmap(page);
7556                 memset(kaddr + zero_start, 0, PAGE_CACHE_SIZE - zero_start);
7557                 flush_dcache_page(page);
7558                 kunmap(page);
7559         }
7560         ClearPageChecked(page);
7561         set_page_dirty(page);
7562         SetPageUptodate(page);
7563
7564         BTRFS_I(inode)->last_trans = root->fs_info->generation;
7565         BTRFS_I(inode)->last_sub_trans = BTRFS_I(inode)->root->log_transid;
7566         BTRFS_I(inode)->last_log_commit = BTRFS_I(inode)->root->last_log_commit;
7567
7568         unlock_extent_cached(io_tree, page_start, page_end, &cached_state, GFP_NOFS);
7569
7570 out_unlock:
7571         if (!ret) {
7572                 sb_end_pagefault(inode->i_sb);
7573                 return VM_FAULT_LOCKED;
7574         }
7575         unlock_page(page);
7576 out:
7577         btrfs_delalloc_release_space(inode, PAGE_CACHE_SIZE);
7578 out_noreserve:
7579         sb_end_pagefault(inode->i_sb);
7580         return ret;
7581 }
7582
7583 static int btrfs_truncate(struct inode *inode)
7584 {
7585         struct btrfs_root *root = BTRFS_I(inode)->root;
7586         struct btrfs_block_rsv *rsv;
7587         int ret = 0;
7588         int err = 0;
7589         struct btrfs_trans_handle *trans;
7590         u64 mask = root->sectorsize - 1;
7591         u64 min_size = btrfs_calc_trunc_metadata_size(root, 1);
7592
7593         ret = btrfs_wait_ordered_range(inode, inode->i_size & (~mask),
7594                                        (u64)-1);
7595         if (ret)
7596                 return ret;
7597
7598         /*
7599          * Yes ladies and gentelment, this is indeed ugly.  The fact is we have
7600          * 3 things going on here
7601          *
7602          * 1) We need to reserve space for our orphan item and the space to
7603          * delete our orphan item.  Lord knows we don't want to have a dangling
7604          * orphan item because we didn't reserve space to remove it.
7605          *
7606          * 2) We need to reserve space to update our inode.
7607          *
7608          * 3) We need to have something to cache all the space that is going to
7609          * be free'd up by the truncate operation, but also have some slack
7610          * space reserved in case it uses space during the truncate (thank you
7611          * very much snapshotting).
7612          *
7613          * And we need these to all be seperate.  The fact is we can use alot of
7614          * space doing the truncate, and we have no earthly idea how much space
7615          * we will use, so we need the truncate reservation to be seperate so it
7616          * doesn't end up using space reserved for updating the inode or
7617          * removing the orphan item.  We also need to be able to stop the
7618          * transaction and start a new one, which means we need to be able to
7619          * update the inode several times, and we have no idea of knowing how
7620          * many times that will be, so we can't just reserve 1 item for the
7621          * entirety of the opration, so that has to be done seperately as well.
7622          * Then there is the orphan item, which does indeed need to be held on
7623          * to for the whole operation, and we need nobody to touch this reserved
7624          * space except the orphan code.
7625          *
7626          * So that leaves us with
7627          *
7628          * 1) root->orphan_block_rsv - for the orphan deletion.
7629          * 2) rsv - for the truncate reservation, which we will steal from the
7630          * transaction reservation.
7631          * 3) fs_info->trans_block_rsv - this will have 1 items worth left for
7632          * updating the inode.
7633          */
7634         rsv = btrfs_alloc_block_rsv(root, BTRFS_BLOCK_RSV_TEMP);
7635         if (!rsv)
7636                 return -ENOMEM;
7637         rsv->size = min_size;
7638         rsv->failfast = 1;
7639
7640         /*
7641          * 1 for the truncate slack space
7642          * 1 for updating the inode.
7643          */
7644         trans = btrfs_start_transaction(root, 2);
7645         if (IS_ERR(trans)) {
7646                 err = PTR_ERR(trans);
7647                 goto out;
7648         }
7649
7650         /* Migrate the slack space for the truncate to our reserve */
7651         ret = btrfs_block_rsv_migrate(&root->fs_info->trans_block_rsv, rsv,
7652                                       min_size);
7653         BUG_ON(ret);
7654
7655         /*
7656          * setattr is responsible for setting the ordered_data_close flag,
7657          * but that is only tested during the last file release.  That
7658          * could happen well after the next commit, leaving a great big
7659          * window where new writes may get lost if someone chooses to write
7660          * to this file after truncating to zero
7661          *
7662          * The inode doesn't have any dirty data here, and so if we commit
7663          * this is a noop.  If someone immediately starts writing to the inode
7664          * it is very likely we'll catch some of their writes in this
7665          * transaction, and the commit will find this file on the ordered
7666          * data list with good things to send down.
7667          *
7668          * This is a best effort solution, there is still a window where
7669          * using truncate to replace the contents of the file will
7670          * end up with a zero length file after a crash.
7671          */
7672         if (inode->i_size == 0 && test_bit(BTRFS_INODE_ORDERED_DATA_CLOSE,
7673                                            &BTRFS_I(inode)->runtime_flags))
7674                 btrfs_add_ordered_operation(trans, root, inode);
7675
7676         /*
7677          * So if we truncate and then write and fsync we normally would just
7678          * write the extents that changed, which is a problem if we need to
7679          * first truncate that entire inode.  So set this flag so we write out
7680          * all of the extents in the inode to the sync log so we're completely
7681          * safe.
7682          */
7683         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &BTRFS_I(inode)->runtime_flags);
7684         trans->block_rsv = rsv;
7685
7686         while (1) {
7687                 ret = btrfs_truncate_inode_items(trans, root, inode,
7688                                                  inode->i_size,
7689                                                  BTRFS_EXTENT_DATA_KEY);
7690                 if (ret != -ENOSPC) {
7691                         err = ret;
7692                         break;
7693                 }
7694
7695                 trans->block_rsv = &root->fs_info->trans_block_rsv;
7696                 ret = btrfs_update_inode(trans, root, inode);
7697                 if (ret) {
7698                         err = ret;
7699                         break;
7700                 }
7701
7702                 btrfs_end_transaction(trans, root);
7703                 btrfs_btree_balance_dirty(root);
7704
7705                 trans = btrfs_start_transaction(root, 2);
7706                 if (IS_ERR(trans)) {
7707                         ret = err = PTR_ERR(trans);
7708                         trans = NULL;
7709                         break;
7710                 }
7711
7712                 ret = btrfs_block_rsv_migrate(&root->fs_info->trans_block_rsv,
7713                                               rsv, min_size);
7714                 BUG_ON(ret);    /* shouldn't happen */
7715                 trans->block_rsv = rsv;
7716         }
7717
7718         if (ret == 0 && inode->i_nlink > 0) {
7719                 trans->block_rsv = root->orphan_block_rsv;
7720                 ret = btrfs_orphan_del(trans, inode);
7721                 if (ret)
7722                         err = ret;
7723         }
7724
7725         if (trans) {
7726                 trans->block_rsv = &root->fs_info->trans_block_rsv;
7727                 ret = btrfs_update_inode(trans, root, inode);
7728                 if (ret && !err)
7729                         err = ret;
7730
7731                 ret = btrfs_end_transaction(trans, root);
7732                 btrfs_btree_balance_dirty(root);
7733         }
7734
7735 out:
7736         btrfs_free_block_rsv(root, rsv);
7737
7738         if (ret && !err)
7739                 err = ret;
7740
7741         return err;
7742 }
7743
7744 /*
7745  * create a new subvolume directory/inode (helper for the ioctl).
7746  */
7747 int btrfs_create_subvol_root(struct btrfs_trans_handle *trans,
7748                              struct btrfs_root *new_root, u64 new_dirid)
7749 {
7750         struct inode *inode;
7751         int err;
7752         u64 index = 0;
7753
7754         inode = btrfs_new_inode(trans, new_root, NULL, "..", 2,
7755                                 new_dirid, new_dirid,
7756                                 S_IFDIR | (~current_umask() & S_IRWXUGO),
7757                                 &index);
7758         if (IS_ERR(inode))
7759                 return PTR_ERR(inode);
7760         inode->i_op = &btrfs_dir_inode_operations;
7761         inode->i_fop = &btrfs_dir_file_operations;
7762
7763         set_nlink(inode, 1);
7764         btrfs_i_size_write(inode, 0);
7765
7766         err = btrfs_update_inode(trans, new_root, inode);
7767
7768         iput(inode);
7769         return err;
7770 }
7771
7772 struct inode *btrfs_alloc_inode(struct super_block *sb)
7773 {
7774         struct btrfs_inode *ei;
7775         struct inode *inode;
7776
7777         ei = kmem_cache_alloc(btrfs_inode_cachep, GFP_NOFS);
7778         if (!ei)
7779                 return NULL;
7780
7781         ei->root = NULL;
7782         ei->generation = 0;
7783         ei->last_trans = 0;
7784         ei->last_sub_trans = 0;
7785         ei->logged_trans = 0;
7786         ei->delalloc_bytes = 0;
7787         ei->disk_i_size = 0;
7788         ei->flags = 0;
7789         ei->csum_bytes = 0;
7790         ei->index_cnt = (u64)-1;
7791         ei->last_unlink_trans = 0;
7792         ei->last_log_commit = 0;
7793
7794         spin_lock_init(&ei->lock);
7795         ei->outstanding_extents = 0;
7796         ei->reserved_extents = 0;
7797
7798         ei->runtime_flags = 0;
7799         ei->force_compress = BTRFS_COMPRESS_NONE;
7800
7801         ei->delayed_node = NULL;
7802
7803         inode = &ei->vfs_inode;
7804         extent_map_tree_init(&ei->extent_tree);
7805         extent_io_tree_init(&ei->io_tree, &inode->i_data);
7806         extent_io_tree_init(&ei->io_failure_tree, &inode->i_data);
7807         ei->io_tree.track_uptodate = 1;
7808         ei->io_failure_tree.track_uptodate = 1;
7809         atomic_set(&ei->sync_writers, 0);
7810         mutex_init(&ei->log_mutex);
7811         mutex_init(&ei->delalloc_mutex);
7812         btrfs_ordered_inode_tree_init(&ei->ordered_tree);
7813         INIT_LIST_HEAD(&ei->delalloc_inodes);
7814         INIT_LIST_HEAD(&ei->ordered_operations);
7815         RB_CLEAR_NODE(&ei->rb_node);
7816
7817         return inode;
7818 }
7819
7820 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
7821 void btrfs_test_destroy_inode(struct inode *inode)
7822 {
7823         btrfs_drop_extent_cache(inode, 0, (u64)-1, 0);
7824         kmem_cache_free(btrfs_inode_cachep, BTRFS_I(inode));
7825 }
7826 #endif
7827
7828 static void btrfs_i_callback(struct rcu_head *head)
7829 {
7830         struct inode *inode = container_of(head, struct inode, i_rcu);
7831         kmem_cache_free(btrfs_inode_cachep, BTRFS_I(inode));
7832 }
7833
7834 void btrfs_destroy_inode(struct inode *inode)
7835 {
7836         struct btrfs_ordered_extent *ordered;
7837         struct btrfs_root *root = BTRFS_I(inode)->root;
7838
7839         WARN_ON(!hlist_empty(&inode->i_dentry));
7840         WARN_ON(inode->i_data.nrpages);
7841         WARN_ON(BTRFS_I(inode)->outstanding_extents);
7842         WARN_ON(BTRFS_I(inode)->reserved_extents);
7843         WARN_ON(BTRFS_I(inode)->delalloc_bytes);
7844         WARN_ON(BTRFS_I(inode)->csum_bytes);
7845
7846         /*
7847          * This can happen where we create an inode, but somebody else also
7848          * created the same inode and we need to destroy the one we already
7849          * created.
7850          */
7851         if (!root)
7852                 goto free;
7853
7854         /*
7855          * Make sure we're properly removed from the ordered operation
7856          * lists.
7857          */
7858         smp_mb();
7859         if (!list_empty(&BTRFS_I(inode)->ordered_operations)) {
7860                 spin_lock(&root->fs_info->ordered_root_lock);
7861                 list_del_init(&BTRFS_I(inode)->ordered_operations);
7862                 spin_unlock(&root->fs_info->ordered_root_lock);
7863         }
7864
7865         if (test_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
7866                      &BTRFS_I(inode)->runtime_flags)) {
7867                 btrfs_info(root->fs_info, "inode %llu still on the orphan list",
7868                         btrfs_ino(inode));
7869                 atomic_dec(&root->orphan_inodes);
7870         }
7871
7872         while (1) {
7873                 ordered = btrfs_lookup_first_ordered_extent(inode, (u64)-1);
7874                 if (!ordered)
7875                         break;
7876                 else {
7877                         btrfs_err(root->fs_info, "found ordered extent %llu %llu on inode cleanup",
7878                                 ordered->file_offset, ordered->len);
7879                         btrfs_remove_ordered_extent(inode, ordered);
7880                         btrfs_put_ordered_extent(ordered);
7881                         btrfs_put_ordered_extent(ordered);
7882                 }
7883         }
7884         inode_tree_del(inode);
7885         btrfs_drop_extent_cache(inode, 0, (u64)-1, 0);
7886 free:
7887         call_rcu(&inode->i_rcu, btrfs_i_callback);
7888 }
7889
7890 int btrfs_drop_inode(struct inode *inode)
7891 {
7892         struct btrfs_root *root = BTRFS_I(inode)->root;
7893
7894         if (root == NULL)
7895                 return 1;
7896
7897         /* the snap/subvol tree is on deleting */
7898         if (btrfs_root_refs(&root->root_item) == 0)
7899                 return 1;
7900         else
7901                 return generic_drop_inode(inode);
7902 }
7903
7904 static void init_once(void *foo)
7905 {
7906         struct btrfs_inode *ei = (struct btrfs_inode *) foo;
7907
7908         inode_init_once(&ei->vfs_inode);
7909 }
7910
7911 void btrfs_destroy_cachep(void)
7912 {
7913         /*
7914          * Make sure all delayed rcu free inodes are flushed before we
7915          * destroy cache.
7916          */
7917         rcu_barrier();
7918         if (btrfs_inode_cachep)
7919                 kmem_cache_destroy(btrfs_inode_cachep);
7920         if (btrfs_trans_handle_cachep)
7921                 kmem_cache_destroy(btrfs_trans_handle_cachep);
7922         if (btrfs_transaction_cachep)
7923                 kmem_cache_destroy(btrfs_transaction_cachep);
7924         if (btrfs_path_cachep)
7925                 kmem_cache_destroy(btrfs_path_cachep);
7926         if (btrfs_free_space_cachep)
7927                 kmem_cache_destroy(btrfs_free_space_cachep);
7928         if (btrfs_delalloc_work_cachep)
7929                 kmem_cache_destroy(btrfs_delalloc_work_cachep);
7930 }
7931
7932 int btrfs_init_cachep(void)
7933 {
7934         btrfs_inode_cachep = kmem_cache_create("btrfs_inode",
7935                         sizeof(struct btrfs_inode), 0,
7936                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, init_once);
7937         if (!btrfs_inode_cachep)
7938                 goto fail;
7939
7940         btrfs_trans_handle_cachep = kmem_cache_create("btrfs_trans_handle",
7941                         sizeof(struct btrfs_trans_handle), 0,
7942                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, NULL);
7943         if (!btrfs_trans_handle_cachep)
7944                 goto fail;
7945
7946         btrfs_transaction_cachep = kmem_cache_create("btrfs_transaction",
7947                         sizeof(struct btrfs_transaction), 0,
7948                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, NULL);
7949         if (!btrfs_transaction_cachep)
7950                 goto fail;
7951
7952         btrfs_path_cachep = kmem_cache_create("btrfs_path",
7953                         sizeof(struct btrfs_path), 0,
7954                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, NULL);
7955         if (!btrfs_path_cachep)
7956                 goto fail;
7957
7958         btrfs_free_space_cachep = kmem_cache_create("btrfs_free_space",
7959                         sizeof(struct btrfs_free_space), 0,
7960                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, NULL);
7961         if (!btrfs_free_space_cachep)
7962                 goto fail;
7963
7964         btrfs_delalloc_work_cachep = kmem_cache_create("btrfs_delalloc_work",
7965                         sizeof(struct btrfs_delalloc_work), 0,
7966                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD,
7967                         NULL);
7968         if (!btrfs_delalloc_work_cachep)
7969                 goto fail;
7970
7971         return 0;
7972 fail:
7973         btrfs_destroy_cachep();
7974         return -ENOMEM;
7975 }
7976
7977 static int btrfs_getattr(struct vfsmount *mnt,
7978                          struct dentry *dentry, struct kstat *stat)
7979 {
7980         u64 delalloc_bytes;
7981         struct inode *inode = dentry->d_inode;
7982         u32 blocksize = inode->i_sb->s_blocksize;
7983
7984         generic_fillattr(inode, stat);
7985         stat->dev = BTRFS_I(inode)->root->anon_dev;
7986         stat->blksize = PAGE_CACHE_SIZE;
7987
7988         spin_lock(&BTRFS_I(inode)->lock);
7989         delalloc_bytes = BTRFS_I(inode)->delalloc_bytes;
7990         spin_unlock(&BTRFS_I(inode)->lock);
7991         stat->blocks = (ALIGN(inode_get_bytes(inode), blocksize) +
7992                         ALIGN(delalloc_bytes, blocksize)) >> 9;
7993         return 0;
7994 }
7995
7996 static int btrfs_rename(struct inode *old_dir, struct dentry *old_dentry,
7997                            struct inode *new_dir, struct dentry *new_dentry)
7998 {
7999         struct btrfs_trans_handle *trans;
8000         struct btrfs_root *root = BTRFS_I(old_dir)->root;
8001         struct btrfs_root *dest = BTRFS_I(new_dir)->root;
8002         struct inode *new_inode = new_dentry->d_inode;
8003         struct inode *old_inode = old_dentry->d_inode;
8004         struct timespec ctime = CURRENT_TIME;
8005         u64 index = 0;
8006         u64 root_objectid;
8007         int ret;
8008         u64 old_ino = btrfs_ino(old_inode);
8009
8010         if (btrfs_ino(new_dir) == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)
8011                 return -EPERM;
8012
8013         /* we only allow rename subvolume link between subvolumes */
8014         if (old_ino != BTRFS_FIRST_FREE_OBJECTID && root != dest)
8015                 return -EXDEV;
8016
8017         if (old_ino == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID ||
8018             (new_inode && btrfs_ino(new_inode) == BTRFS_FIRST_FREE_OBJECTID))
8019                 return -ENOTEMPTY;
8020
8021         if (S_ISDIR(old_inode->i_mode) && new_inode &&
8022             new_inode->i_size > BTRFS_EMPTY_DIR_SIZE)
8023                 return -ENOTEMPTY;
8024
8025
8026         /* check for collisions, even if the  name isn't there */
8027         ret = btrfs_check_dir_item_collision(dest, new_dir->i_ino,
8028                              new_dentry->d_name.name,
8029                              new_dentry->d_name.len);
8030
8031         if (ret) {
8032                 if (ret == -EEXIST) {
8033                         /* we shouldn't get
8034                          * eexist without a new_inode */
8035                         if (WARN_ON(!new_inode)) {
8036                                 return ret;
8037                         }
8038                 } else {
8039                         /* maybe -EOVERFLOW */
8040                         return ret;
8041                 }
8042         }
8043         ret = 0;
8044
8045         /*
8046          * we're using rename to replace one file with another.
8047          * and the replacement file is large.  Start IO on it now so
8048          * we don't add too much work to the end of the transaction
8049          */
8050         if (new_inode && S_ISREG(old_inode->i_mode) && new_inode->i_size &&
8051             old_inode->i_size > BTRFS_ORDERED_OPERATIONS_FLUSH_LIMIT)
8052                 filemap_flush(old_inode->i_mapping);
8053
8054         /* close the racy window with snapshot create/destroy ioctl */
8055         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
8056                 down_read(&root->fs_info->subvol_sem);
8057         /*
8058          * We want to reserve the absolute worst case amount of items.  So if
8059          * both inodes are subvols and we need to unlink them then that would
8060          * require 4 item modifications, but if they are both normal inodes it
8061          * would require 5 item modifications, so we'll assume their normal
8062          * inodes.  So 5 * 2 is 10, plus 1 for the new link, so 11 total items
8063          * should cover the worst case number of items we'll modify.
8064          */
8065         trans = btrfs_start_transaction(root, 11);
8066         if (IS_ERR(trans)) {
8067                 ret = PTR_ERR(trans);
8068                 goto out_notrans;
8069         }
8070
8071         if (dest != root)
8072                 btrfs_record_root_in_trans(trans, dest);
8073
8074         ret = btrfs_set_inode_index(new_dir, &index);
8075         if (ret)
8076                 goto out_fail;
8077
8078         if (unlikely(old_ino == BTRFS_FIRST_FREE_OBJECTID)) {
8079                 /* force full log commit if subvolume involved. */
8080                 root->fs_info->last_trans_log_full_commit = trans->transid;
8081         } else {
8082                 ret = btrfs_insert_inode_ref(trans, dest,
8083                                              new_dentry->d_name.name,
8084                                              new_dentry->d_name.len,
8085                                              old_ino,
8086                                              btrfs_ino(new_dir), index);
8087                 if (ret)
8088                         goto out_fail;
8089                 /*
8090                  * this is an ugly little race, but the rename is required
8091                  * to make sure that if we crash, the inode is either at the
8092                  * old name or the new one.  pinning the log transaction lets
8093                  * us make sure we don't allow a log commit to come in after
8094                  * we unlink the name but before we add the new name back in.
8095                  */
8096                 btrfs_pin_log_trans(root);
8097         }
8098         /*
8099          * make sure the inode gets flushed if it is replacing
8100          * something.
8101          */
8102         if (new_inode && new_inode->i_size && S_ISREG(old_inode->i_mode))
8103                 btrfs_add_ordered_operation(trans, root, old_inode);
8104
8105         inode_inc_iversion(old_dir);
8106         inode_inc_iversion(new_dir);
8107         inode_inc_iversion(old_inode);
8108         old_dir->i_ctime = old_dir->i_mtime = ctime;
8109         new_dir->i_ctime = new_dir->i_mtime = ctime;
8110         old_inode->i_ctime = ctime;
8111
8112         if (old_dentry->d_parent != new_dentry->d_parent)
8113                 btrfs_record_unlink_dir(trans, old_dir, old_inode, 1);
8114
8115         if (unlikely(old_ino == BTRFS_FIRST_FREE_OBJECTID)) {
8116                 root_objectid = BTRFS_I(old_inode)->root->root_key.objectid;
8117                 ret = btrfs_unlink_subvol(trans, root, old_dir, root_objectid,
8118                                         old_dentry->d_name.name,
8119                                         old_dentry->d_name.len);
8120         } else {
8121                 ret = __btrfs_unlink_inode(trans, root, old_dir,
8122                                         old_dentry->d_inode,
8123                                         old_dentry->d_name.name,
8124                                         old_dentry->d_name.len);
8125                 if (!ret)
8126                         ret = btrfs_update_inode(trans, root, old_inode);
8127         }
8128         if (ret) {
8129                 btrfs_abort_transaction(trans, root, ret);
8130                 goto out_fail;
8131         }
8132
8133         if (new_inode) {
8134                 inode_inc_iversion(new_inode);
8135                 new_inode->i_ctime = CURRENT_TIME;
8136                 if (unlikely(btrfs_ino(new_inode) ==
8137                              BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)) {
8138                         root_objectid = BTRFS_I(new_inode)->location.objectid;
8139                         ret = btrfs_unlink_subvol(trans, dest, new_dir,
8140                                                 root_objectid,
8141                                                 new_dentry->d_name.name,
8142                                                 new_dentry->d_name.len);
8143                         BUG_ON(new_inode->i_nlink == 0);
8144                 } else {
8145                         ret = btrfs_unlink_inode(trans, dest, new_dir,
8146                                                  new_dentry->d_inode,
8147                                                  new_dentry->d_name.name,
8148                                                  new_dentry->d_name.len);
8149                 }
8150                 if (!ret && new_inode->i_nlink == 0)
8151                         ret = btrfs_orphan_add(trans, new_dentry->d_inode);
8152                 if (ret) {
8153                         btrfs_abort_transaction(trans, root, ret);
8154                         goto out_fail;
8155                 }
8156         }
8157
8158         ret = btrfs_add_link(trans, new_dir, old_inode,
8159                              new_dentry->d_name.name,
8160                              new_dentry->d_name.len, 0, index);
8161         if (ret) {
8162                 btrfs_abort_transaction(trans, root, ret);
8163                 goto out_fail;
8164         }
8165
8166         if (old_ino != BTRFS_FIRST_FREE_OBJECTID) {
8167                 struct dentry *parent = new_dentry->d_parent;
8168                 btrfs_log_new_name(trans, old_inode, old_dir, parent);
8169                 btrfs_end_log_trans(root);
8170         }
8171 out_fail:
8172         btrfs_end_transaction(trans, root);
8173 out_notrans:
8174         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
8175                 up_read(&root->fs_info->subvol_sem);
8176
8177         return ret;
8178 }
8179
8180 static void btrfs_run_delalloc_work(struct btrfs_work *work)
8181 {
8182         struct btrfs_delalloc_work *delalloc_work;
8183         struct inode *inode;
8184
8185         delalloc_work = container_of(work, struct btrfs_delalloc_work,
8186                                      work);
8187         inode = delalloc_work->inode;
8188         if (delalloc_work->wait) {
8189                 btrfs_wait_ordered_range(inode, 0, (u64)-1);
8190         } else {
8191                 filemap_flush(inode->i_mapping);
8192                 if (test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
8193                              &BTRFS_I(inode)->runtime_flags))
8194                         filemap_flush(inode->i_mapping);
8195         }
8196
8197         if (delalloc_work->delay_iput)
8198                 btrfs_add_delayed_iput(inode);
8199         else
8200                 iput(inode);
8201         complete(&delalloc_work->completion);
8202 }
8203
8204 struct btrfs_delalloc_work *btrfs_alloc_delalloc_work(struct inode *inode,
8205                                                     int wait, int delay_iput)
8206 {
8207         struct btrfs_delalloc_work *work;
8208
8209         work = kmem_cache_zalloc(btrfs_delalloc_work_cachep, GFP_NOFS);
8210         if (!work)
8211                 return NULL;
8212
8213         init_completion(&work->completion);
8214         INIT_LIST_HEAD(&work->list);
8215         work->inode = inode;
8216         work->wait = wait;
8217         work->delay_iput = delay_iput;
8218         work->work.func = btrfs_run_delalloc_work;
8219
8220         return work;
8221 }
8222
8223 void btrfs_wait_and_free_delalloc_work(struct btrfs_delalloc_work *work)
8224 {
8225         wait_for_completion(&work->completion);
8226         kmem_cache_free(btrfs_delalloc_work_cachep, work);
8227 }
8228
8229 /*
8230  * some fairly slow code that needs optimization. This walks the list
8231  * of all the inodes with pending delalloc and forces them to disk.
8232  */
8233 static int __start_delalloc_inodes(struct btrfs_root *root, int delay_iput)
8234 {
8235         struct btrfs_inode *binode;
8236         struct inode *inode;
8237         struct btrfs_delalloc_work *work, *next;
8238         struct list_head works;
8239         struct list_head splice;
8240         int ret = 0;
8241
8242         INIT_LIST_HEAD(&works);
8243         INIT_LIST_HEAD(&splice);
8244
8245         spin_lock(&root->delalloc_lock);
8246         list_splice_init(&root->delalloc_inodes, &splice);
8247         while (!list_empty(&splice)) {
8248                 binode = list_entry(splice.next, struct btrfs_inode,
8249                                     delalloc_inodes);
8250
8251                 list_move_tail(&binode->delalloc_inodes,
8252                                &root->delalloc_inodes);
8253                 inode = igrab(&binode->vfs_inode);
8254                 if (!inode) {
8255                         cond_resched_lock(&root->delalloc_lock);
8256                         continue;
8257                 }
8258                 spin_unlock(&root->delalloc_lock);
8259
8260                 work = btrfs_alloc_delalloc_work(inode, 0, delay_iput);
8261                 if (unlikely(!work)) {
8262                         if (delay_iput)
8263                                 btrfs_add_delayed_iput(inode);
8264                         else
8265                                 iput(inode);
8266                         ret = -ENOMEM;
8267                         goto out;
8268                 }
8269                 list_add_tail(&work->list, &works);
8270                 btrfs_queue_worker(&root->fs_info->flush_workers,
8271                                    &work->work);
8272
8273                 cond_resched();
8274                 spin_lock(&root->delalloc_lock);
8275         }
8276         spin_unlock(&root->delalloc_lock);
8277
8278         list_for_each_entry_safe(work, next, &works, list) {
8279                 list_del_init(&work->list);
8280                 btrfs_wait_and_free_delalloc_work(work);
8281         }
8282         return 0;
8283 out:
8284         list_for_each_entry_safe(work, next, &works, list) {
8285                 list_del_init(&work->list);
8286                 btrfs_wait_and_free_delalloc_work(work);
8287         }
8288
8289         if (!list_empty_careful(&splice)) {
8290                 spin_lock(&root->delalloc_lock);
8291                 list_splice_tail(&splice, &root->delalloc_inodes);
8292                 spin_unlock(&root->delalloc_lock);
8293         }
8294         return ret;
8295 }
8296
8297 int btrfs_start_delalloc_inodes(struct btrfs_root *root, int delay_iput)
8298 {
8299         int ret;
8300
8301         if (root->fs_info->sb->s_flags & MS_RDONLY)
8302                 return -EROFS;
8303
8304         ret = __start_delalloc_inodes(root, delay_iput);
8305         /*
8306          * the filemap_flush will queue IO into the worker threads, but
8307          * we have to make sure the IO is actually started and that
8308          * ordered extents get created before we return
8309          */
8310         atomic_inc(&root->fs_info->async_submit_draining);
8311         while (atomic_read(&root->fs_info->nr_async_submits) ||
8312               atomic_read(&root->fs_info->async_delalloc_pages)) {
8313                 wait_event(root->fs_info->async_submit_wait,
8314                    (atomic_read(&root->fs_info->nr_async_submits) == 0 &&
8315                     atomic_read(&root->fs_info->async_delalloc_pages) == 0));
8316         }
8317         atomic_dec(&root->fs_info->async_submit_draining);
8318         return ret;
8319 }
8320
8321 int btrfs_start_delalloc_roots(struct btrfs_fs_info *fs_info, int delay_iput)
8322 {
8323         struct btrfs_root *root;
8324         struct list_head splice;
8325         int ret;
8326
8327         if (fs_info->sb->s_flags & MS_RDONLY)
8328                 return -EROFS;
8329
8330         INIT_LIST_HEAD(&splice);
8331
8332         spin_lock(&fs_info->delalloc_root_lock);
8333         list_splice_init(&fs_info->delalloc_roots, &splice);
8334         while (!list_empty(&splice)) {
8335                 root = list_first_entry(&splice, struct btrfs_root,
8336                                         delalloc_root);
8337                 root = btrfs_grab_fs_root(root);
8338                 BUG_ON(!root);
8339                 list_move_tail(&root->delalloc_root,
8340                                &fs_info->delalloc_roots);
8341                 spin_unlock(&fs_info->delalloc_root_lock);
8342
8343                 ret = __start_delalloc_inodes(root, delay_iput);
8344                 btrfs_put_fs_root(root);
8345                 if (ret)
8346                         goto out;
8347
8348                 spin_lock(&fs_info->delalloc_root_lock);
8349         }
8350         spin_unlock(&fs_info->delalloc_root_lock);
8351
8352         atomic_inc(&fs_info->async_submit_draining);
8353         while (atomic_read(&fs_info->nr_async_submits) ||
8354               atomic_read(&fs_info->async_delalloc_pages)) {
8355                 wait_event(fs_info->async_submit_wait,
8356                    (atomic_read(&fs_info->nr_async_submits) == 0 &&
8357                     atomic_read(&fs_info->async_delalloc_pages) == 0));
8358         }
8359         atomic_dec(&fs_info->async_submit_draining);
8360         return 0;
8361 out:
8362         if (!list_empty_careful(&splice)) {
8363                 spin_lock(&fs_info->delalloc_root_lock);
8364                 list_splice_tail(&splice, &fs_info->delalloc_roots);
8365                 spin_unlock(&fs_info->delalloc_root_lock);
8366         }
8367         return ret;
8368 }
8369
8370 static int btrfs_symlink(struct inode *dir, struct dentry *dentry,
8371                          const char *symname)
8372 {
8373         struct btrfs_trans_handle *trans;
8374         struct btrfs_root *root = BTRFS_I(dir)->root;
8375         struct btrfs_path *path;
8376         struct btrfs_key key;
8377         struct inode *inode = NULL;
8378         int err;
8379         int drop_inode = 0;
8380         u64 objectid;
8381         u64 index = 0;
8382         int name_len;
8383         int datasize;
8384         unsigned long ptr;
8385         struct btrfs_file_extent_item *ei;
8386         struct extent_buffer *leaf;
8387
8388         name_len = strlen(symname);
8389         if (name_len > BTRFS_MAX_INLINE_DATA_SIZE(root))
8390                 return -ENAMETOOLONG;
8391
8392         /*
8393          * 2 items for inode item and ref
8394          * 2 items for dir items
8395          * 1 item for xattr if selinux is on
8396          */
8397         trans = btrfs_start_transaction(root, 5);
8398         if (IS_ERR(trans))
8399                 return PTR_ERR(trans);
8400
8401         err = btrfs_find_free_ino(root, &objectid);
8402         if (err)
8403                 goto out_unlock;
8404
8405         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
8406                                 dentry->d_name.len, btrfs_ino(dir), objectid,
8407                                 S_IFLNK|S_IRWXUGO, &index);
8408         if (IS_ERR(inode)) {
8409                 err = PTR_ERR(inode);
8410                 goto out_unlock;
8411         }
8412
8413         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
8414         if (err) {
8415                 drop_inode = 1;
8416                 goto out_unlock;
8417         }
8418
8419         /*
8420         * If the active LSM wants to access the inode during
8421         * d_instantiate it needs these. Smack checks to see
8422         * if the filesystem supports xattrs by looking at the
8423         * ops vector.
8424         */
8425         inode->i_fop = &btrfs_file_operations;
8426         inode->i_op = &btrfs_file_inode_operations;
8427
8428         err = btrfs_add_nondir(trans, dir, dentry, inode, 0, index);
8429         if (err)
8430                 drop_inode = 1;
8431         else {
8432                 inode->i_mapping->a_ops = &btrfs_aops;
8433                 inode->i_mapping->backing_dev_info = &root->fs_info->bdi;
8434                 BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
8435         }
8436         if (drop_inode)
8437                 goto out_unlock;
8438
8439         path = btrfs_alloc_path();
8440         if (!path) {
8441                 err = -ENOMEM;
8442                 drop_inode = 1;
8443                 goto out_unlock;
8444         }
8445         key.objectid = btrfs_ino(inode);
8446         key.offset = 0;
8447         btrfs_set_key_type(&key, BTRFS_EXTENT_DATA_KEY);
8448         datasize = btrfs_file_extent_calc_inline_size(name_len);
8449         err = btrfs_insert_empty_item(trans, root, path, &key,
8450                                       datasize);
8451         if (err) {
8452                 drop_inode = 1;
8453                 btrfs_free_path(path);
8454                 goto out_unlock;
8455         }
8456         leaf = path->nodes[0];
8457         ei = btrfs_item_ptr(leaf, path->slots[0],
8458                             struct btrfs_file_extent_item);
8459         btrfs_set_file_extent_generation(leaf, ei, trans->transid);
8460         btrfs_set_file_extent_type(leaf, ei,
8461                                    BTRFS_FILE_EXTENT_INLINE);
8462         btrfs_set_file_extent_encryption(leaf, ei, 0);
8463         btrfs_set_file_extent_compression(leaf, ei, 0);
8464         btrfs_set_file_extent_other_encoding(leaf, ei, 0);
8465         btrfs_set_file_extent_ram_bytes(leaf, ei, name_len);
8466
8467         ptr = btrfs_file_extent_inline_start(ei);
8468         write_extent_buffer(leaf, symname, ptr, name_len);
8469         btrfs_mark_buffer_dirty(leaf);
8470         btrfs_free_path(path);
8471
8472         inode->i_op = &btrfs_symlink_inode_operations;
8473         inode->i_mapping->a_ops = &btrfs_symlink_aops;
8474         inode->i_mapping->backing_dev_info = &root->fs_info->bdi;
8475         inode_set_bytes(inode, name_len);
8476         btrfs_i_size_write(inode, name_len);
8477         err = btrfs_update_inode(trans, root, inode);
8478         if (err)
8479                 drop_inode = 1;
8480
8481 out_unlock:
8482         if (!err)
8483                 d_instantiate(dentry, inode);
8484         btrfs_end_transaction(trans, root);
8485         if (drop_inode) {
8486                 inode_dec_link_count(inode);
8487                 iput(inode);
8488         }
8489         btrfs_btree_balance_dirty(root);
8490         return err;
8491 }
8492
8493 static int __btrfs_prealloc_file_range(struct inode *inode, int mode,
8494                                        u64 start, u64 num_bytes, u64 min_size,
8495                                        loff_t actual_len, u64 *alloc_hint,
8496                                        struct btrfs_trans_handle *trans)
8497 {
8498         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
8499         struct extent_map *em;
8500         struct btrfs_root *root = BTRFS_I(inode)->root;
8501         struct btrfs_key ins;
8502         u64 cur_offset = start;
8503         u64 i_size;
8504         u64 cur_bytes;
8505         int ret = 0;
8506         bool own_trans = true;
8507
8508         if (trans)
8509                 own_trans = false;
8510         while (num_bytes > 0) {
8511                 if (own_trans) {
8512                         trans = btrfs_start_transaction(root, 3);
8513                         if (IS_ERR(trans)) {
8514                                 ret = PTR_ERR(trans);
8515                                 break;
8516                         }
8517                 }
8518
8519                 cur_bytes = min(num_bytes, 256ULL * 1024 * 1024);
8520                 cur_bytes = max(cur_bytes, min_size);
8521                 ret = btrfs_reserve_extent(root, cur_bytes, min_size, 0,
8522                                            *alloc_hint, &ins, 1);
8523                 if (ret) {
8524                         if (own_trans)
8525                                 btrfs_end_transaction(trans, root);
8526                         break;
8527                 }
8528
8529                 ret = insert_reserved_file_extent(trans, inode,
8530                                                   cur_offset, ins.objectid,
8531                                                   ins.offset, ins.offset,
8532                                                   ins.offset, 0, 0, 0,
8533                                                   BTRFS_FILE_EXTENT_PREALLOC);
8534                 if (ret) {
8535                         btrfs_free_reserved_extent(root, ins.objectid,
8536                                                    ins.offset);
8537                         btrfs_abort_transaction(trans, root, ret);
8538                         if (own_trans)
8539                                 btrfs_end_transaction(trans, root);
8540                         break;
8541                 }
8542                 btrfs_drop_extent_cache(inode, cur_offset,
8543                                         cur_offset + ins.offset -1, 0);
8544
8545                 em = alloc_extent_map();
8546                 if (!em) {
8547                         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
8548                                 &BTRFS_I(inode)->runtime_flags);
8549                         goto next;
8550                 }
8551
8552                 em->start = cur_offset;
8553                 em->orig_start = cur_offset;
8554                 em->len = ins.offset;
8555                 em->block_start = ins.objectid;
8556                 em->block_len = ins.offset;
8557                 em->orig_block_len = ins.offset;
8558                 em->ram_bytes = ins.offset;
8559                 em->bdev = root->fs_info->fs_devices->latest_bdev;
8560                 set_bit(EXTENT_FLAG_PREALLOC, &em->flags);
8561                 em->generation = trans->transid;
8562
8563                 while (1) {
8564                         write_lock(&em_tree->lock);
8565                         ret = add_extent_mapping(em_tree, em, 1);
8566                         write_unlock(&em_tree->lock);
8567                         if (ret != -EEXIST)
8568                                 break;
8569                         btrfs_drop_extent_cache(inode, cur_offset,
8570                                                 cur_offset + ins.offset - 1,
8571                                                 0);
8572                 }
8573                 free_extent_map(em);
8574 next:
8575                 num_bytes -= ins.offset;
8576                 cur_offset += ins.offset;
8577                 *alloc_hint = ins.objectid + ins.offset;
8578
8579                 inode_inc_iversion(inode);
8580                 inode->i_ctime = CURRENT_TIME;
8581                 BTRFS_I(inode)->flags |= BTRFS_INODE_PREALLOC;
8582                 if (!(mode & FALLOC_FL_KEEP_SIZE) &&
8583                     (actual_len > inode->i_size) &&
8584                     (cur_offset > inode->i_size)) {
8585                         if (cur_offset > actual_len)
8586                                 i_size = actual_len;
8587                         else
8588                                 i_size = cur_offset;
8589                         i_size_write(inode, i_size);
8590                         btrfs_ordered_update_i_size(inode, i_size, NULL);
8591                 }
8592
8593                 ret = btrfs_update_inode(trans, root, inode);
8594
8595                 if (ret) {
8596                         btrfs_abort_transaction(trans, root, ret);
8597                         if (own_trans)
8598                                 btrfs_end_transaction(trans, root);
8599                         break;
8600                 }
8601
8602                 if (own_trans)
8603                         btrfs_end_transaction(trans, root);
8604         }
8605         return ret;
8606 }
8607
8608 int btrfs_prealloc_file_range(struct inode *inode, int mode,
8609                               u64 start, u64 num_bytes, u64 min_size,
8610                               loff_t actual_len, u64 *alloc_hint)
8611 {
8612         return __btrfs_prealloc_file_range(inode, mode, start, num_bytes,
8613                                            min_size, actual_len, alloc_hint,
8614                                            NULL);
8615 }
8616
8617 int btrfs_prealloc_file_range_trans(struct inode *inode,
8618                                     struct btrfs_trans_handle *trans, int mode,
8619                                     u64 start, u64 num_bytes, u64 min_size,
8620                                     loff_t actual_len, u64 *alloc_hint)
8621 {
8622         return __btrfs_prealloc_file_range(inode, mode, start, num_bytes,
8623                                            min_size, actual_len, alloc_hint, trans);
8624 }
8625
8626 static int btrfs_set_page_dirty(struct page *page)
8627 {
8628         return __set_page_dirty_nobuffers(page);
8629 }
8630
8631 static int btrfs_permission(struct inode *inode, int mask)
8632 {
8633         struct btrfs_root *root = BTRFS_I(inode)->root;
8634         umode_t mode = inode->i_mode;
8635
8636         if (mask & MAY_WRITE &&
8637             (S_ISREG(mode) || S_ISDIR(mode) || S_ISLNK(mode))) {
8638                 if (btrfs_root_readonly(root))
8639                         return -EROFS;
8640                 if (BTRFS_I(inode)->flags & BTRFS_INODE_READONLY)
8641                         return -EACCES;
8642         }
8643         return generic_permission(inode, mask);
8644 }
8645
8646 static const struct inode_operations btrfs_dir_inode_operations = {
8647         .getattr        = btrfs_getattr,
8648         .lookup         = btrfs_lookup,
8649         .create         = btrfs_create,
8650         .unlink         = btrfs_unlink,
8651         .link           = btrfs_link,
8652         .mkdir          = btrfs_mkdir,
8653         .rmdir          = btrfs_rmdir,
8654         .rename         = btrfs_rename,
8655         .symlink        = btrfs_symlink,
8656         .setattr        = btrfs_setattr,
8657         .mknod          = btrfs_mknod,
8658         .setxattr       = btrfs_setxattr,
8659         .getxattr       = btrfs_getxattr,
8660         .listxattr      = btrfs_listxattr,
8661         .removexattr    = btrfs_removexattr,
8662         .permission     = btrfs_permission,
8663         .get_acl        = btrfs_get_acl,
8664         .update_time    = btrfs_update_time,
8665 };
8666 static const struct inode_operations btrfs_dir_ro_inode_operations = {
8667         .lookup         = btrfs_lookup,
8668         .permission     = btrfs_permission,
8669         .get_acl        = btrfs_get_acl,
8670         .update_time    = btrfs_update_time,
8671 };
8672
8673 static const struct file_operations btrfs_dir_file_operations = {
8674         .llseek         = generic_file_llseek,
8675         .read           = generic_read_dir,
8676         .iterate        = btrfs_real_readdir,
8677         .unlocked_ioctl = btrfs_ioctl,
8678 #ifdef CONFIG_COMPAT
8679         .compat_ioctl   = btrfs_ioctl,
8680 #endif
8681         .release        = btrfs_release_file,
8682         .fsync          = btrfs_sync_file,
8683 };
8684
8685 static struct extent_io_ops btrfs_extent_io_ops = {
8686         .fill_delalloc = run_delalloc_range,
8687         .submit_bio_hook = btrfs_submit_bio_hook,
8688         .merge_bio_hook = btrfs_merge_bio_hook,
8689         .readpage_end_io_hook = btrfs_readpage_end_io_hook,
8690         .writepage_end_io_hook = btrfs_writepage_end_io_hook,
8691         .writepage_start_hook = btrfs_writepage_start_hook,
8692         .set_bit_hook = btrfs_set_bit_hook,
8693         .clear_bit_hook = btrfs_clear_bit_hook,
8694         .merge_extent_hook = btrfs_merge_extent_hook,
8695         .split_extent_hook = btrfs_split_extent_hook,
8696 };
8697
8698 /*
8699  * btrfs doesn't support the bmap operation because swapfiles
8700  * use bmap to make a mapping of extents in the file.  They assume
8701  * these extents won't change over the life of the file and they
8702  * use the bmap result to do IO directly to the drive.
8703  *
8704  * the btrfs bmap call would return logical addresses that aren't
8705  * suitable for IO and they also will change frequently as COW
8706  * operations happen.  So, swapfile + btrfs == corruption.
8707  *
8708  * For now we're avoiding this by dropping bmap.
8709  */
8710 static const struct address_space_operations btrfs_aops = {
8711         .readpage       = btrfs_readpage,
8712         .writepage      = btrfs_writepage,
8713         .writepages     = btrfs_writepages,
8714         .readpages      = btrfs_readpages,
8715         .direct_IO      = btrfs_direct_IO,
8716         .invalidatepage = btrfs_invalidatepage,
8717         .releasepage    = btrfs_releasepage,
8718         .set_page_dirty = btrfs_set_page_dirty,
8719         .error_remove_page = generic_error_remove_page,
8720 };
8721
8722 static const struct address_space_operations btrfs_symlink_aops = {
8723         .readpage       = btrfs_readpage,
8724         .writepage      = btrfs_writepage,
8725         .invalidatepage = btrfs_invalidatepage,
8726         .releasepage    = btrfs_releasepage,
8727 };
8728
8729 static const struct inode_operations btrfs_file_inode_operations = {
8730         .getattr        = btrfs_getattr,
8731         .setattr        = btrfs_setattr,
8732         .setxattr       = btrfs_setxattr,
8733         .getxattr       = btrfs_getxattr,
8734         .listxattr      = btrfs_listxattr,
8735         .removexattr    = btrfs_removexattr,
8736         .permission     = btrfs_permission,
8737         .fiemap         = btrfs_fiemap,
8738         .get_acl        = btrfs_get_acl,
8739         .update_time    = btrfs_update_time,
8740 };
8741 static const struct inode_operations btrfs_special_inode_operations = {
8742         .getattr        = btrfs_getattr,
8743         .setattr        = btrfs_setattr,
8744         .permission     = btrfs_permission,
8745         .setxattr       = btrfs_setxattr,
8746         .getxattr       = btrfs_getxattr,
8747         .listxattr      = btrfs_listxattr,
8748         .removexattr    = btrfs_removexattr,
8749         .get_acl        = btrfs_get_acl,
8750         .update_time    = btrfs_update_time,
8751 };
8752 static const struct inode_operations btrfs_symlink_inode_operations = {
8753         .readlink       = generic_readlink,
8754         .follow_link    = page_follow_link_light,
8755         .put_link       = page_put_link,
8756         .getattr        = btrfs_getattr,
8757         .setattr        = btrfs_setattr,
8758         .permission     = btrfs_permission,
8759         .setxattr       = btrfs_setxattr,
8760         .getxattr       = btrfs_getxattr,
8761         .listxattr      = btrfs_listxattr,
8762         .removexattr    = btrfs_removexattr,
8763         .get_acl        = btrfs_get_acl,
8764         .update_time    = btrfs_update_time,
8765 };
8766
8767 const struct dentry_operations btrfs_dentry_operations = {
8768         .d_delete       = btrfs_dentry_delete,
8769         .d_release      = btrfs_dentry_release,
8770 };