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