fs/btrfs/ordered-data.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/slab.h>
  20 #include <linux/blkdev.h>
  21 #include <linux/writeback.h>
  22 #include <linux/pagevec.h>
  23 #include "ctree.h"
  24 #include "transaction.h"
  25 #include "btrfs_inode.h"
  26 #include "extent_io.h"
  27
  28 static struct kmem_cache *btrfs_ordered_extent_cache;
  29
  30 static u64 entry_end(struct btrfs_ordered_extent *entry)
  31 {
  32         if (entry->file_offset + entry->len < entry->file_offset)
  33                 return (u64)-1;
  34         return entry->file_offset + entry->len;
  35 }
  36
  37 /* returns NULL if the insertion worked, or it returns the node it did find
  38  * in the tree
  39  */
  40 static struct rb_node *tree_insert(struct rb_root *root, u64 file_offset,
  41                                    struct rb_node *node)
  42 {
  43         struct rb_node **p = &root->rb_node;
  44         struct rb_node *parent = NULL;
  45         struct btrfs_ordered_extent *entry;
  46
  47         while (*p) {
  48                 parent = *p;
  49                 entry = rb_entry(parent, struct btrfs_ordered_extent, rb_node);
  50
  51                 if (file_offset < entry->file_offset)
  52                         p = &(*p)->rb_left;
  53                 else if (file_offset >= entry_end(entry))
  54                         p = &(*p)->rb_right;
  55                 else
  56                         return parent;
  57         }
  58
  59         rb_link_node(node, parent, p);
  60         rb_insert_color(node, root);
  61         return NULL;
  62 }
  63
  64 static void ordered_data_tree_panic(struct inode *inode, int errno,
  65                                                u64 offset)
  66 {
  67         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
  68         btrfs_panic(fs_info, errno, "Inconsistency in ordered tree at offset "
  69                     "%llu\n", (unsigned long long)offset);
  70 }
  71
  72 /*
  73  * look for a given offset in the tree, and if it can't be found return the
  74  * first lesser offset
  75  */
  76 static struct rb_node *__tree_search(struct rb_root *root, u64 file_offset,
  77                                      struct rb_node **prev_ret)
  78 {
  79         struct rb_node *n = root->rb_node;
  80         struct rb_node *prev = NULL;
  81         struct rb_node *test;
  82         struct btrfs_ordered_extent *entry;
  83         struct btrfs_ordered_extent *prev_entry = NULL;
  84
  85         while (n) {
  86                 entry = rb_entry(n, struct btrfs_ordered_extent, rb_node);
  87                 prev = n;
  88                 prev_entry = entry;
  89
  90                 if (file_offset < entry->file_offset)
  91                         n = n->rb_left;
  92                 else if (file_offset >= entry_end(entry))
  93                         n = n->rb_right;
  94                 else
  95                         return n;
  96         }
  97         if (!prev_ret)
  98                 return NULL;
  99
 100         while (prev && file_offset >= entry_end(prev_entry)) {
 101                 test = rb_next(prev);
 102                 if (!test)
 103                         break;
 104                 prev_entry = rb_entry(test, struct btrfs_ordered_extent,
 105                                       rb_node);
 106                 if (file_offset < entry_end(prev_entry))
 107                         break;
 108
 109                 prev = test;
 110         }
 111         if (prev)
 112                 prev_entry = rb_entry(prev, struct btrfs_ordered_extent,
 113                                       rb_node);
 114         while (prev && file_offset < entry_end(prev_entry)) {
 115                 test = rb_prev(prev);
 116                 if (!test)
 117                         break;
 118                 prev_entry = rb_entry(test, struct btrfs_ordered_extent,
 119                                       rb_node);
 120                 prev = test;
 121         }
 122         *prev_ret = prev;
 123         return NULL;
 124 }
 125
 126 /*
 127  * helper to check if a given offset is inside a given entry
 128  */
 129 static int offset_in_entry(struct btrfs_ordered_extent *entry, u64 file_offset)
 130 {
 131         if (file_offset < entry->file_offset ||
 132             entry->file_offset + entry->len <= file_offset)
 133                 return 0;
 134         return 1;
 135 }
 136
 137 static int range_overlaps(struct btrfs_ordered_extent *entry, u64 file_offset,
 138                           u64 len)
 139 {
 140         if (file_offset + len <= entry->file_offset ||
 141             entry->file_offset + entry->len <= file_offset)
 142                 return 0;
 143         return 1;
 144 }
 145
 146 /*
 147  * look find the first ordered struct that has this offset, otherwise
 148  * the first one less than this offset
 149  */
 150 static inline struct rb_node *tree_search(struct btrfs_ordered_inode_tree *tree,
 151                                           u64 file_offset)
 152 {
 153         struct rb_root *root = &tree->tree;
 154         struct rb_node *prev = NULL;
 155         struct rb_node *ret;
 156         struct btrfs_ordered_extent *entry;
 157
 158         if (tree->last) {
 159                 entry = rb_entry(tree->last, struct btrfs_ordered_extent,
 160                                  rb_node);
 161                 if (offset_in_entry(entry, file_offset))
 162                         return tree->last;
 163         }
 164         ret = __tree_search(root, file_offset, &prev);
 165         if (!ret)
 166                 ret = prev;
 167         if (ret)
 168                 tree->last = ret;
 169         return ret;
 170 }
 171
 172 /* allocate and add a new ordered_extent into the per-inode tree.
 173  * file_offset is the logical offset in the file
 174  *
 175  * start is the disk block number of an extent already reserved in the
 176  * extent allocation tree
 177  *
 178  * len is the length of the extent
 179  *
 180  * The tree is given a single reference on the ordered extent that was
 181  * inserted.
 182  */
 183 static int __btrfs_add_ordered_extent(struct inode *inode, u64 file_offset,
 184                                       u64 start, u64 len, u64 disk_len,
 185                                       int type, int dio, int compress_type)
 186 {
 187         struct btrfs_ordered_inode_tree *tree;
 188         struct rb_node *node;
 189         struct btrfs_ordered_extent *entry;
 190
 191         tree = &BTRFS_I(inode)->ordered_tree;
 192         entry = kmem_cache_zalloc(btrfs_ordered_extent_cache, GFP_NOFS);
 193         if (!entry)
 194                 return -ENOMEM;
 195
 196         entry->file_offset = file_offset;
 197         entry->start = start;
 198         entry->len = len;
 199         entry->disk_len = disk_len;
 200         entry->bytes_left = len;
 201         entry->inode = igrab(inode);
 202         entry->compress_type = compress_type;
 203         if (type != BTRFS_ORDERED_IO_DONE && type != BTRFS_ORDERED_COMPLETE)
 204                 set_bit(type, &entry->flags);
 205
 206         if (dio)
 207                 set_bit(BTRFS_ORDERED_DIRECT, &entry->flags);
 208
 209         /* one ref for the tree */
 210         atomic_set(&entry->refs, 1);
 211         init_waitqueue_head(&entry->wait);
 212         INIT_LIST_HEAD(&entry->list);
 213         INIT_LIST_HEAD(&entry->root_extent_list);
 214
 215         trace_btrfs_ordered_extent_add(inode, entry);
 216
 217         spin_lock_irq(&tree->lock);
 218         node = tree_insert(&tree->tree, file_offset,
 219                            &entry->rb_node);
 220         if (node)
 221                 ordered_data_tree_panic(inode, -EEXIST, file_offset);
 222         spin_unlock_irq(&tree->lock);
 223
 224         spin_lock(&BTRFS_I(inode)->root->fs_info->ordered_extent_lock);
 225         list_add_tail(&entry->root_extent_list,
 226                       &BTRFS_I(inode)->root->fs_info->ordered_extents);
 227         spin_unlock(&BTRFS_I(inode)->root->fs_info->ordered_extent_lock);
 228
 229         return 0;
 230 }
 231
 232 int btrfs_add_ordered_extent(struct inode *inode, u64 file_offset,
 233                              u64 start, u64 len, u64 disk_len, int type)
 234 {
 235         return __btrfs_add_ordered_extent(inode, file_offset, start, len,
 236                                           disk_len, type, 0,
 237                                           BTRFS_COMPRESS_NONE);
 238 }
 239
 240 int btrfs_add_ordered_extent_dio(struct inode *inode, u64 file_offset,
 241                                  u64 start, u64 len, u64 disk_len, int type)
 242 {
 243         return __btrfs_add_ordered_extent(inode, file_offset, start, len,
 244                                           disk_len, type, 1,
 245                                           BTRFS_COMPRESS_NONE);
 246 }
 247
 248 int btrfs_add_ordered_extent_compress(struct inode *inode, u64 file_offset,
 249                                       u64 start, u64 len, u64 disk_len,
 250                                       int type, int compress_type)
 251 {
 252         return __btrfs_add_ordered_extent(inode, file_offset, start, len,
 253                                           disk_len, type, 0,
 254                                           compress_type);
 255 }
 256
 257 /*
 258  * Add a struct btrfs_ordered_sum into the list of checksums to be inserted
 259  * when an ordered extent is finished.  If the list covers more than one
 260  * ordered extent, it is split across multiples.
 261  */
 262 void btrfs_add_ordered_sum(struct inode *inode,
 263                            struct btrfs_ordered_extent *entry,
 264                            struct btrfs_ordered_sum *sum)
 265 {
 266         struct btrfs_ordered_inode_tree *tree;
 267
 268         tree = &BTRFS_I(inode)->ordered_tree;
 269         spin_lock_irq(&tree->lock);
 270         list_add_tail(&sum->list, &entry->list);
 271         spin_unlock_irq(&tree->lock);
 272 }
 273
 274 /*
 275  * this is used to account for finished IO across a given range
 276  * of the file.  The IO may span ordered extents.  If
 277  * a given ordered_extent is completely done, 1 is returned, otherwise
 278  * 0.
 279  *
 280  * test_and_set_bit on a flag in the struct btrfs_ordered_extent is used
 281  * to make sure this function only returns 1 once for a given ordered extent.
 282  *
 283  * file_offset is updated to one byte past the range that is recorded as
 284  * complete.  This allows you to walk forward in the file.
 285  */
 286 int btrfs_dec_test_first_ordered_pending(struct inode *inode,
 287                                    struct btrfs_ordered_extent **cached,
 288                                    u64 *file_offset, u64 io_size, int uptodate)
 289 {
 290         struct btrfs_ordered_inode_tree *tree;
 291         struct rb_node *node;
 292         struct btrfs_ordered_extent *entry = NULL;
 293         int ret;
 294         unsigned long flags;
 295         u64 dec_end;
 296         u64 dec_start;
 297         u64 to_dec;
 298
 299         tree = &BTRFS_I(inode)->ordered_tree;
 300         spin_lock_irqsave(&tree->lock, flags);
 301         node = tree_search(tree, *file_offset);
 302         if (!node) {
 303                 ret = 1;
 304                 goto out;
 305         }
 306
 307         entry = rb_entry(node, struct btrfs_ordered_extent, rb_node);
 308         if (!offset_in_entry(entry, *file_offset)) {
 309                 ret = 1;
 310                 goto out;
 311         }
 312
 313         dec_start = max(*file_offset, entry->file_offset);
 314         dec_end = min(*file_offset + io_size, entry->file_offset +
 315                       entry->len);
 316         *file_offset = dec_end;
 317         if (dec_start > dec_end) {
 318                 printk(KERN_CRIT "bad ordering dec_start %llu end %llu\n",
 319                        (unsigned long long)dec_start,
 320                        (unsigned long long)dec_end);
 321         }
 322         to_dec = dec_end - dec_start;
 323         if (to_dec > entry->bytes_left) {
 324                 printk(KERN_CRIT "bad ordered accounting left %llu size %llu\n",
 325                        (unsigned long long)entry->bytes_left,
 326                        (unsigned long long)to_dec);
 327         }
 328         entry->bytes_left -= to_dec;
 329         if (!uptodate)
 330                 set_bit(BTRFS_ORDERED_IOERR, &entry->flags);
 331
 332         if (entry->bytes_left == 0)
 333                 ret = test_and_set_bit(BTRFS_ORDERED_IO_DONE, &entry->flags);
 334         else
 335                 ret = 1;
 336 out:
 337         if (!ret && cached && entry) {
 338                 *cached = entry;
 339                 atomic_inc(&entry->refs);
 340         }
 341         spin_unlock_irqrestore(&tree->lock, flags);
 342         return ret == 0;
 343 }
 344
 345 /*
 346  * this is used to account for finished IO across a given range
 347  * of the file.  The IO should not span ordered extents.  If
 348  * a given ordered_extent is completely done, 1 is returned, otherwise
 349  * 0.
 350  *
 351  * test_and_set_bit on a flag in the struct btrfs_ordered_extent is used
 352  * to make sure this function only returns 1 once for a given ordered extent.
 353  */
 354 int btrfs_dec_test_ordered_pending(struct inode *inode,
 355                                    struct btrfs_ordered_extent **cached,
 356                                    u64 file_offset, u64 io_size, int uptodate)
 357 {
 358         struct btrfs_ordered_inode_tree *tree;
 359         struct rb_node *node;
 360         struct btrfs_ordered_extent *entry = NULL;
 361         unsigned long flags;
 362         int ret;
 363
 364         tree = &BTRFS_I(inode)->ordered_tree;
 365         spin_lock_irqsave(&tree->lock, flags);
 366         if (cached && *cached) {
 367                 entry = *cached;
 368                 goto have_entry;
 369         }
 370
 371         node = tree_search(tree, file_offset);
 372         if (!node) {
 373                 ret = 1;
 374                 goto out;
 375         }
 376
 377         entry = rb_entry(node, struct btrfs_ordered_extent, rb_node);
 378 have_entry:
 379         if (!offset_in_entry(entry, file_offset)) {
 380                 ret = 1;
 381                 goto out;
 382         }
 383
 384         if (io_size > entry->bytes_left) {
 385                 printk(KERN_CRIT "bad ordered accounting left %llu size %llu\n",
 386                        (unsigned long long)entry->bytes_left,
 387                        (unsigned long long)io_size);
 388         }
 389         entry->bytes_left -= io_size;
 390         if (!uptodate)
 391                 set_bit(BTRFS_ORDERED_IOERR, &entry->flags);
 392
 393         if (entry->bytes_left == 0)
 394                 ret = test_and_set_bit(BTRFS_ORDERED_IO_DONE, &entry->flags);
 395         else
 396                 ret = 1;
 397 out:
 398         if (!ret && cached && entry) {
 399                 *cached = entry;
 400                 atomic_inc(&entry->refs);
 401         }
 402         spin_unlock_irqrestore(&tree->lock, flags);
 403         return ret == 0;
 404 }
 405
 406 /*
 407  * used to drop a reference on an ordered extent.  This will free
 408  * the extent if the last reference is dropped
 409  */
 410 void btrfs_put_ordered_extent(struct btrfs_ordered_extent *entry)
 411 {
 412         struct list_head *cur;
 413         struct btrfs_ordered_sum *sum;
 414
 415         trace_btrfs_ordered_extent_put(entry->inode, entry);
 416
 417         if (atomic_dec_and_test(&entry->refs)) {
 418                 if (entry->inode)
 419                         btrfs_add_delayed_iput(entry->inode);
 420                 while (!list_empty(&entry->list)) {
 421                         cur = entry->list.next;
 422                         sum = list_entry(cur, struct btrfs_ordered_sum, list);
 423                         list_del(&sum->list);
 424                         kfree(sum);
 425                 }
 426                 kmem_cache_free(btrfs_ordered_extent_cache, entry);
 427         }
 428 }
 429
 430 /*
 431  * remove an ordered extent from the tree.  No references are dropped
 432  * and waiters are woken up.
 433  */
 434 void btrfs_remove_ordered_extent(struct inode *inode,
 435                                  struct btrfs_ordered_extent *entry)
 436 {
 437         struct btrfs_ordered_inode_tree *tree;
 438         struct btrfs_root *root = BTRFS_I(inode)->root;
 439         struct rb_node *node;
 440
 441         tree = &BTRFS_I(inode)->ordered_tree;
 442         spin_lock_irq(&tree->lock);
 443         node = &entry->rb_node;
 444         rb_erase(node, &tree->tree);
 445         tree->last = NULL;
 446         set_bit(BTRFS_ORDERED_COMPLETE, &entry->flags);
 447         spin_unlock_irq(&tree->lock);
 448
 449         spin_lock(&root->fs_info->ordered_extent_lock);
 450         list_del_init(&entry->root_extent_list);
 451
 452         trace_btrfs_ordered_extent_remove(inode, entry);
 453
 454         /*
 455          * we have no more ordered extents for this inode and
 456          * no dirty pages.  We can safely remove it from the
 457          * list of ordered extents
 458          */
 459         if (RB_EMPTY_ROOT(&tree->tree) &&
 460             !mapping_tagged(inode->i_mapping, PAGECACHE_TAG_DIRTY)) {
 461                 list_del_init(&BTRFS_I(inode)->ordered_operations);
 462         }
 463         spin_unlock(&root->fs_info->ordered_extent_lock);
 464         wake_up(&entry->wait);
 465 }
 466
 467 /*
 468  * wait for all the ordered extents in a root.  This is done when balancing
 469  * space between drives.
 470  */
 471 void btrfs_wait_ordered_extents(struct btrfs_root *root, int delay_iput)
 472 {
 473         struct list_head splice;
 474         struct list_head *cur;
 475         struct btrfs_ordered_extent *ordered;
 476         struct inode *inode;
 477
 478         INIT_LIST_HEAD(&splice);
 479
 480         spin_lock(&root->fs_info->ordered_extent_lock);
 481         list_splice_init(&root->fs_info->ordered_extents, &splice);
 482         while (!list_empty(&splice)) {
 483                 cur = splice.next;
 484                 ordered = list_entry(cur, struct btrfs_ordered_extent,
 485                                      root_extent_list);
 486                 list_del_init(&ordered->root_extent_list);
 487                 atomic_inc(&ordered->refs);
 488
 489                 /*
 490                  * the inode may be getting freed (in sys_unlink path).
 491                  */
 492                 inode = igrab(ordered->inode);
 493
 494                 spin_unlock(&root->fs_info->ordered_extent_lock);
 495
 496                 if (inode) {
 497                         btrfs_start_ordered_extent(inode, ordered, 1);
 498                         btrfs_put_ordered_extent(ordered);
 499                         if (delay_iput)
 500                                 btrfs_add_delayed_iput(inode);
 501                         else
 502                                 iput(inode);
 503                 } else {
 504                         btrfs_put_ordered_extent(ordered);
 505                 }
 506
 507                 spin_lock(&root->fs_info->ordered_extent_lock);
 508         }
 509         spin_unlock(&root->fs_info->ordered_extent_lock);
 510 }
 511
 512 /*
 513  * this is used during transaction commit to write all the inodes
 514  * added to the ordered operation list.  These files must be fully on
 515  * disk before the transaction commits.
 516  *
 517  * we have two modes here, one is to just start the IO via filemap_flush
 518  * and the other is to wait for all the io.  When we wait, we have an
 519  * extra check to make sure the ordered operation list really is empty
 520  * before we return
 521  */
 522 void btrfs_run_ordered_operations(struct btrfs_root *root, int wait)
 523 {
 524         struct btrfs_inode *btrfs_inode;
 525         struct inode *inode;
 526         struct list_head splice;
 527
 528         INIT_LIST_HEAD(&splice);
 529
 530         mutex_lock(&root->fs_info->ordered_operations_mutex);
 531         spin_lock(&root->fs_info->ordered_extent_lock);
 532 again:
 533         list_splice_init(&root->fs_info->ordered_operations, &splice);
 534
 535         while (!list_empty(&splice)) {
 536                 btrfs_inode = list_entry(splice.next, struct btrfs_inode,
 537                                    ordered_operations);
 538
 539                 inode = &btrfs_inode->vfs_inode;
 540
 541                 list_del_init(&btrfs_inode->ordered_operations);
 542
 543                 /*
 544                  * the inode may be getting freed (in sys_unlink path).
 545                  */
 546                 inode = igrab(inode);
 547
 548                 if (!wait && inode) {
 549                         list_add_tail(&BTRFS_I(inode)->ordered_operations,
 550                               &root->fs_info->ordered_operations);
 551                 }
 552                 spin_unlock(&root->fs_info->ordered_extent_lock);
 553
 554                 if (inode) {
 555                         if (wait)
 556                                 btrfs_wait_ordered_range(inode, 0, (u64)-1);
 557                         else
 558                                 filemap_flush(inode->i_mapping);
 559                         btrfs_add_delayed_iput(inode);
 560                 }
 561
 562                 cond_resched();
 563                 spin_lock(&root->fs_info->ordered_extent_lock);
 564         }
 565         if (wait && !list_empty(&root->fs_info->ordered_operations))
 566                 goto again;
 567
 568         spin_unlock(&root->fs_info->ordered_extent_lock);
 569         mutex_unlock(&root->fs_info->ordered_operations_mutex);
 570 }
 571
 572 /*
 573  * Used to start IO or wait for a given ordered extent to finish.
 574  *
 575  * If wait is one, this effectively waits on page writeback for all the pages
 576  * in the extent, and it waits on the io completion code to insert
 577  * metadata into the btree corresponding to the extent
 578  */
 579 void btrfs_start_ordered_extent(struct inode *inode,
 580                                        struct btrfs_ordered_extent *entry,
 581                                        int wait)
 582 {
 583         u64 start = entry->file_offset;
 584         u64 end = start + entry->len - 1;
 585
 586         trace_btrfs_ordered_extent_start(inode, entry);
 587
 588         /*
 589          * pages in the range can be dirty, clean or writeback.  We
 590          * start IO on any dirty ones so the wait doesn't stall waiting
 591          * for the flusher thread to find them
 592          */
 593         if (!test_bit(BTRFS_ORDERED_DIRECT, &entry->flags))
 594                 filemap_fdatawrite_range(inode->i_mapping, start, end);
 595         if (wait) {
 596                 wait_event(entry->wait, test_bit(BTRFS_ORDERED_COMPLETE,
 597                                                  &entry->flags));
 598         }
 599 }
 600
 601 /*
 602  * Used to wait on ordered extents across a large range of bytes.
 603  */
 604 void btrfs_wait_ordered_range(struct inode *inode, u64 start, u64 len)
 605 {
 606         u64 end;
 607         u64 orig_end;
 608         struct btrfs_ordered_extent *ordered;
 609         int found;
 610
 611         if (start + len < start) {
 612                 orig_end = INT_LIMIT(loff_t);
 613         } else {
 614                 orig_end = start + len - 1;
 615                 if (orig_end > INT_LIMIT(loff_t))
 616                         orig_end = INT_LIMIT(loff_t);
 617         }
 618
 619         /* start IO across the range first to instantiate any delalloc
 620          * extents
 621          */
 622         filemap_fdatawrite_range(inode->i_mapping, start, orig_end);
 623
 624         /*
 625          * So with compression we will find and lock a dirty page and clear the
 626          * first one as dirty, setup an async extent, and immediately return
 627          * with the entire range locked but with nobody actually marked with
 628          * writeback.  So we can't just filemap_write_and_wait_range() and
 629          * expect it to work since it will just kick off a thread to do the
 630          * actual work.  So we need to call filemap_fdatawrite_range _again_
 631          * since it will wait on the page lock, which won't be unlocked until
 632          * after the pages have been marked as writeback and so we're good to go
 633          * from there.  We have to do this otherwise we'll miss the ordered
 634          * extents and that results in badness.  Please Josef, do not think you
 635          * know better and pull this out at some point in the future, it is
 636          * right and you are wrong.
 637          */
 638         if (test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
 639                      &BTRFS_I(inode)->runtime_flags))
 640                 filemap_fdatawrite_range(inode->i_mapping, start, orig_end);
 641
 642         filemap_fdatawait_range(inode->i_mapping, start, orig_end);
 643
 644         end = orig_end;
 645         found = 0;
 646         while (1) {
 647                 ordered = btrfs_lookup_first_ordered_extent(inode, end);
 648                 if (!ordered)
 649                         break;
 650                 if (ordered->file_offset > orig_end) {
 651                         btrfs_put_ordered_extent(ordered);
 652                         break;
 653                 }
 654                 if (ordered->file_offset + ordered->len < start) {
 655                         btrfs_put_ordered_extent(ordered);
 656                         break;
 657                 }
 658                 found++;
 659                 btrfs_start_ordered_extent(inode, ordered, 1);
 660                 end = ordered->file_offset;
 661                 btrfs_put_ordered_extent(ordered);
 662                 if (end == 0 || end == start)
 663                         break;
 664                 end--;
 665         }
 666 }
 667
 668 /*
 669  * find an ordered extent corresponding to file_offset.  return NULL if
 670  * nothing is found, otherwise take a reference on the extent and return it
 671  */
 672 struct btrfs_ordered_extent *btrfs_lookup_ordered_extent(struct inode *inode,
 673                                                          u64 file_offset)
 674 {
 675         struct btrfs_ordered_inode_tree *tree;
 676         struct rb_node *node;
 677         struct btrfs_ordered_extent *entry = NULL;
 678
 679         tree = &BTRFS_I(inode)->ordered_tree;
 680         spin_lock_irq(&tree->lock);
 681         node = tree_search(tree, file_offset);
 682         if (!node)
 683                 goto out;
 684
 685         entry = rb_entry(node, struct btrfs_ordered_extent, rb_node);
 686         if (!offset_in_entry(entry, file_offset))
 687                 entry = NULL;
 688         if (entry)
 689                 atomic_inc(&entry->refs);
 690 out:
 691         spin_unlock_irq(&tree->lock);
 692         return entry;
 693 }
 694
 695 /* Since the DIO code tries to lock a wide area we need to look for any ordered
 696  * extents that exist in the range, rather than just the start of the range.
 697  */
 698 struct btrfs_ordered_extent *btrfs_lookup_ordered_range(struct inode *inode,
 699                                                         u64 file_offset,
 700                                                         u64 len)
 701 {
 702         struct btrfs_ordered_inode_tree *tree;
 703         struct rb_node *node;
 704         struct btrfs_ordered_extent *entry = NULL;
 705
 706         tree = &BTRFS_I(inode)->ordered_tree;
 707         spin_lock_irq(&tree->lock);
 708         node = tree_search(tree, file_offset);
 709         if (!node) {
 710                 node = tree_search(tree, file_offset + len);
 711                 if (!node)
 712                         goto out;
 713         }
 714
 715         while (1) {
 716                 entry = rb_entry(node, struct btrfs_ordered_extent, rb_node);
 717                 if (range_overlaps(entry, file_offset, len))
 718                         break;
 719
 720                 if (entry->file_offset >= file_offset + len) {
 721                         entry = NULL;
 722                         break;
 723                 }
 724                 entry = NULL;
 725                 node = rb_next(node);
 726                 if (!node)
 727                         break;
 728         }
 729 out:
 730         if (entry)
 731                 atomic_inc(&entry->refs);
 732         spin_unlock_irq(&tree->lock);
 733         return entry;
 734 }
 735
 736 /*
 737  * lookup and return any extent before 'file_offset'.  NULL is returned
 738  * if none is found
 739  */
 740 struct btrfs_ordered_extent *
 741 btrfs_lookup_first_ordered_extent(struct inode *inode, u64 file_offset)
 742 {
 743         struct btrfs_ordered_inode_tree *tree;
 744         struct rb_node *node;
 745         struct btrfs_ordered_extent *entry = NULL;
 746
 747         tree = &BTRFS_I(inode)->ordered_tree;
 748         spin_lock_irq(&tree->lock);
 749         node = tree_search(tree, file_offset);
 750         if (!node)
 751                 goto out;
 752
 753         entry = rb_entry(node, struct btrfs_ordered_extent, rb_node);
 754         atomic_inc(&entry->refs);
 755 out:
 756         spin_unlock_irq(&tree->lock);
 757         return entry;
 758 }
 759
 760 /*
 761  * After an extent is done, call this to conditionally update the on disk
 762  * i_size.  i_size is updated to cover any fully written part of the file.
 763  */
 764 int btrfs_ordered_update_i_size(struct inode *inode, u64 offset,
 765                                 struct btrfs_ordered_extent *ordered)
 766 {
 767         struct btrfs_ordered_inode_tree *tree = &BTRFS_I(inode)->ordered_tree;
 768         u64 disk_i_size;
 769         u64 new_i_size;
 770         u64 i_size = i_size_read(inode);
 771         struct rb_node *node;
 772         struct rb_node *prev = NULL;
 773         struct btrfs_ordered_extent *test;
 774         int ret = 1;
 775
 776         if (ordered)
 777                 offset = entry_end(ordered);
 778         else
 779                 offset = ALIGN(offset, BTRFS_I(inode)->root->sectorsize);
 780
 781         spin_lock_irq(&tree->lock);
 782         disk_i_size = BTRFS_I(inode)->disk_i_size;
 783
 784         /* truncate file */
 785         if (disk_i_size > i_size) {
 786                 BTRFS_I(inode)->disk_i_size = i_size;
 787                 ret = 0;
 788                 goto out;
 789         }
 790
 791         /*
 792          * if the disk i_size is already at the inode->i_size, or
 793          * this ordered extent is inside the disk i_size, we're done
 794          */
 795         if (disk_i_size == i_size || offset <= disk_i_size) {
 796                 goto out;
 797         }
 798
 799         /*
 800          * walk backward from this ordered extent to disk_i_size.
 801          * if we find an ordered extent then we can't update disk i_size
 802          * yet
 803          */
 804         if (ordered) {
 805                 node = rb_prev(&ordered->rb_node);
 806         } else {
 807                 prev = tree_search(tree, offset);
 808                 /*
 809                  * we insert file extents without involving ordered struct,
 810                  * so there should be no ordered struct cover this offset
 811                  */
 812                 if (prev) {
 813                         test = rb_entry(prev, struct btrfs_ordered_extent,
 814                                         rb_node);
 815                         BUG_ON(offset_in_entry(test, offset));
 816                 }
 817                 node = prev;
 818         }
 819         for (; node; node = rb_prev(node)) {
 820                 test = rb_entry(node, struct btrfs_ordered_extent, rb_node);
 821
 822                 /* We treat this entry as if it doesnt exist */
 823                 if (test_bit(BTRFS_ORDERED_UPDATED_ISIZE, &test->flags))
 824                         continue;
 825                 if (test->file_offset + test->len <= disk_i_size)
 826                         break;
 827                 if (test->file_offset >= i_size)
 828                         break;
 829                 if (test->file_offset >= disk_i_size) {
 830                         /*
 831                          * we don't update disk_i_size now, so record this
 832                          * undealt i_size. Or we will not know the real
 833                          * i_size.
 834                          */
 835                         if (test->outstanding_isize < offset)
 836                                 test->outstanding_isize = offset;
 837                         if (ordered &&
 838                             ordered->outstanding_isize >
 839                             test->outstanding_isize)
 840                                 test->outstanding_isize =
 841                                                 ordered->outstanding_isize;
 842                         goto out;
 843                 }
 844         }
 845         new_i_size = min_t(u64, offset, i_size);
 846
 847         /*
 848          * Some ordered extents may completed before the current one, and
 849          * we hold the real i_size in ->outstanding_isize.
 850          */
 851         if (ordered && ordered->outstanding_isize > new_i_size)
 852                 new_i_size = min_t(u64, ordered->outstanding_isize, i_size);
 853         BTRFS_I(inode)->disk_i_size = new_i_size;
 854         ret = 0;
 855 out:
 856         /*
 857          * We need to do this because we can't remove ordered extents until
 858          * after the i_disk_size has been updated and then the inode has been
 859          * updated to reflect the change, so we need to tell anybody who finds
 860          * this ordered extent that we've already done all the real work, we
 861          * just haven't completed all the other work.
 862          */
 863         if (ordered)
 864                 set_bit(BTRFS_ORDERED_UPDATED_ISIZE, &ordered->flags);
 865         spin_unlock_irq(&tree->lock);
 866         return ret;
 867 }
 868
 869 /*
 870  * search the ordered extents for one corresponding to 'offset' and
 871  * try to find a checksum.  This is used because we allow pages to
 872  * be reclaimed before their checksum is actually put into the btree
 873  */
 874 int btrfs_find_ordered_sum(struct inode *inode, u64 offset, u64 disk_bytenr,
 875                            u32 *sum)
 876 {
 877         struct btrfs_ordered_sum *ordered_sum;
 878         struct btrfs_sector_sum *sector_sums;
 879         struct btrfs_ordered_extent *ordered;
 880         struct btrfs_ordered_inode_tree *tree = &BTRFS_I(inode)->ordered_tree;
 881         unsigned long num_sectors;
 882         unsigned long i;
 883         u32 sectorsize = BTRFS_I(inode)->root->sectorsize;
 884         int ret = 1;
 885
 886         ordered = btrfs_lookup_ordered_extent(inode, offset);
 887         if (!ordered)
 888                 return 1;
 889
 890         spin_lock_irq(&tree->lock);
 891         list_for_each_entry_reverse(ordered_sum, &ordered->list, list) {
 892                 if (disk_bytenr >= ordered_sum->bytenr) {
 893                         num_sectors = ordered_sum->len / sectorsize;
 894                         sector_sums = ordered_sum->sums;
 895                         for (i = 0; i < num_sectors; i++) {
 896                                 if (sector_sums[i].bytenr == disk_bytenr) {
 897                                         *sum = sector_sums[i].sum;
 898                                         ret = 0;
 899                                         goto out;
 900                                 }
 901                         }
 902                 }
 903         }
 904 out:
 905         spin_unlock_irq(&tree->lock);
 906         btrfs_put_ordered_extent(ordered);
 907         return ret;
 908 }
 909
 910
 911 /*
 912  * add a given inode to the list of inodes that must be fully on
 913  * disk before a transaction commit finishes.
 914  *
 915  * This basically gives us the ext3 style data=ordered mode, and it is mostly
 916  * used to make sure renamed files are fully on disk.
 917  *
 918  * It is a noop if the inode is already fully on disk.
 919  *
 920  * If trans is not null, we'll do a friendly check for a transaction that
 921  * is already flushing things and force the IO down ourselves.
 922  */
 923 void btrfs_add_ordered_operation(struct btrfs_trans_handle *trans,
 924                                  struct btrfs_root *root, struct inode *inode)
 925 {
 926         u64 last_mod;
 927
 928         last_mod = max(BTRFS_I(inode)->generation, BTRFS_I(inode)->last_trans);
 929
 930         /*
 931          * if this file hasn't been changed since the last transaction
 932          * commit, we can safely return without doing anything
 933          */
 934         if (last_mod < root->fs_info->last_trans_committed)
 935                 return;
 936
 937         /*
 938          * the transaction is already committing.  Just start the IO and
 939          * don't bother with all of this list nonsense
 940          */
 941         if (trans && root->fs_info->running_transaction->blocked) {
 942                 btrfs_wait_ordered_range(inode, 0, (u64)-1);
 943                 return;
 944         }
 945
 946         spin_lock(&root->fs_info->ordered_extent_lock);
 947         if (list_empty(&BTRFS_I(inode)->ordered_operations)) {
 948                 list_add_tail(&BTRFS_I(inode)->ordered_operations,
 949                               &root->fs_info->ordered_operations);
 950         }
 951         spin_unlock(&root->fs_info->ordered_extent_lock);
 952 }
 953
 954 int __init ordered_data_init(void)
 955 {
 956         btrfs_ordered_extent_cache = kmem_cache_create("btrfs_ordered_extent",
 957                                      sizeof(struct btrfs_ordered_extent), 0,
 958                                      SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD,
 959                                      NULL);
 960         if (!btrfs_ordered_extent_cache)
 961                 return -ENOMEM;
 962         return 0;
 963 }
 964
 965 void ordered_data_exit(void)
 966 {
 967         if (btrfs_ordered_extent_cache)
 968                 kmem_cache_destroy(btrfs_ordered_extent_cache);
 969 }