fs/btrfs/file.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/fs.h>
  20 #include <linux/pagemap.h>
  21 #include <linux/highmem.h>
  22 #include <linux/time.h>
  23 #include <linux/init.h>
  24 #include <linux/string.h>
  25 #include <linux/backing-dev.h>
  26 #include <linux/mpage.h>
  27 #include <linux/falloc.h>
  28 #include <linux/swap.h>
  29 #include <linux/writeback.h>
  30 #include <linux/statfs.h>
  31 #include <linux/compat.h>
  32 #include <linux/slab.h>
  33 #include "ctree.h"
  34 #include "disk-io.h"
  35 #include "transaction.h"
  36 #include "btrfs_inode.h"
  37 #include "ioctl.h"
  38 #include "print-tree.h"
  39 #include "tree-log.h"
  40 #include "locking.h"
  41 #include "compat.h"
  42 #include "volumes.h"
  43
  44 static struct kmem_cache *btrfs_inode_defrag_cachep;
  45 /*
  46  * when auto defrag is enabled we
  47  * queue up these defrag structs to remember which
  48  * inodes need defragging passes
  49  */
  50 struct inode_defrag {
  51         struct rb_node rb_node;
  52         /* objectid */
  53         u64 ino;
  54         /*
  55          * transid where the defrag was added, we search for
  56          * extents newer than this
  57          */
  58         u64 transid;
  59
  60         /* root objectid */
  61         u64 root;
  62
  63         /* last offset we were able to defrag */
  64         u64 last_offset;
  65
  66         /* if we've wrapped around back to zero once already */
  67         int cycled;
  68 };
  69
  70 static int __compare_inode_defrag(struct inode_defrag *defrag1,
  71                                   struct inode_defrag *defrag2)
  72 {
  73         if (defrag1->root > defrag2->root)
  74                 return 1;
  75         else if (defrag1->root < defrag2->root)
  76                 return -1;
  77         else if (defrag1->ino > defrag2->ino)
  78                 return 1;
  79         else if (defrag1->ino < defrag2->ino)
  80                 return -1;
  81         else
  82                 return 0;
  83 }
  84
  85 /* pop a record for an inode into the defrag tree.  The lock
  86  * must be held already
  87  *
  88  * If you're inserting a record for an older transid than an
  89  * existing record, the transid already in the tree is lowered
  90  *
  91  * If an existing record is found the defrag item you
  92  * pass in is freed
  93  */
  94 static int __btrfs_add_inode_defrag(struct inode *inode,
  95                                     struct inode_defrag *defrag)
  96 {
  97         struct btrfs_root *root = BTRFS_I(inode)->root;
  98         struct inode_defrag *entry;
  99         struct rb_node **p;
 100         struct rb_node *parent = NULL;
 101         int ret;
 102
 103         p = &root->fs_info->defrag_inodes.rb_node;
 104         while (*p) {
 105                 parent = *p;
 106                 entry = rb_entry(parent, struct inode_defrag, rb_node);
 107
 108                 ret = __compare_inode_defrag(defrag, entry);
 109                 if (ret < 0)
 110                         p = &parent->rb_left;
 111                 else if (ret > 0)
 112                         p = &parent->rb_right;
 113                 else {
 114                         /* if we're reinserting an entry for
 115                          * an old defrag run, make sure to
 116                          * lower the transid of our existing record
 117                          */
 118                         if (defrag->transid < entry->transid)
 119                                 entry->transid = defrag->transid;
 120                         if (defrag->last_offset > entry->last_offset)
 121                                 entry->last_offset = defrag->last_offset;
 122                         return -EEXIST;
 123                 }
 124         }
 125         set_bit(BTRFS_INODE_IN_DEFRAG, &BTRFS_I(inode)->runtime_flags);
 126         rb_link_node(&defrag->rb_node, parent, p);
 127         rb_insert_color(&defrag->rb_node, &root->fs_info->defrag_inodes);
 128         return 0;
 129 }
 130
 131 static inline int __need_auto_defrag(struct btrfs_root *root)
 132 {
 133         if (!btrfs_test_opt(root, AUTO_DEFRAG))
 134                 return 0;
 135
 136         if (btrfs_fs_closing(root->fs_info))
 137                 return 0;
 138
 139         return 1;
 140 }
 141
 142 /*
 143  * insert a defrag record for this inode if auto defrag is
 144  * enabled
 145  */
 146 int btrfs_add_inode_defrag(struct btrfs_trans_handle *trans,
 147                            struct inode *inode)
 148 {
 149         struct btrfs_root *root = BTRFS_I(inode)->root;
 150         struct inode_defrag *defrag;
 151         u64 transid;
 152         int ret;
 153
 154         if (!__need_auto_defrag(root))
 155                 return 0;
 156
 157         if (test_bit(BTRFS_INODE_IN_DEFRAG, &BTRFS_I(inode)->runtime_flags))
 158                 return 0;
 159
 160         if (trans)
 161                 transid = trans->transid;
 162         else
 163                 transid = BTRFS_I(inode)->root->last_trans;
 164
 165         defrag = kmem_cache_zalloc(btrfs_inode_defrag_cachep, GFP_NOFS);
 166         if (!defrag)
 167                 return -ENOMEM;
 168
 169         defrag->ino = btrfs_ino(inode);
 170         defrag->transid = transid;
 171         defrag->root = root->root_key.objectid;
 172
 173         spin_lock(&root->fs_info->defrag_inodes_lock);
 174         if (!test_bit(BTRFS_INODE_IN_DEFRAG, &BTRFS_I(inode)->runtime_flags)) {
 175                 /*
 176                  * If we set IN_DEFRAG flag and evict the inode from memory,
 177                  * and then re-read this inode, this new inode doesn't have
 178                  * IN_DEFRAG flag. At the case, we may find the existed defrag.
 179                  */
 180                 ret = __btrfs_add_inode_defrag(inode, defrag);
 181                 if (ret)
 182                         kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
 183         } else {
 184                 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
 185         }
 186         spin_unlock(&root->fs_info->defrag_inodes_lock);
 187         return 0;
 188 }
 189
 190 /*
 191  * Requeue the defrag object. If there is a defrag object that points to
 192  * the same inode in the tree, we will merge them together (by
 193  * __btrfs_add_inode_defrag()) and free the one that we want to requeue.
 194  */
 195 void btrfs_requeue_inode_defrag(struct inode *inode,
 196                                 struct inode_defrag *defrag)
 197 {
 198         struct btrfs_root *root = BTRFS_I(inode)->root;
 199         int ret;
 200
 201         if (!__need_auto_defrag(root))
 202                 goto out;
 203
 204         /*
 205          * Here we don't check the IN_DEFRAG flag, because we need merge
 206          * them together.
 207          */
 208         spin_lock(&root->fs_info->defrag_inodes_lock);
 209         ret = __btrfs_add_inode_defrag(inode, defrag);
 210         spin_unlock(&root->fs_info->defrag_inodes_lock);
 211         if (ret)
 212                 goto out;
 213         return;
 214 out:
 215         kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
 216 }
 217
 218 /*
 219  * pick the defragable inode that we want, if it doesn't exist, we will get
 220  * the next one.
 221  */
 222 static struct inode_defrag *
 223 btrfs_pick_defrag_inode(struct btrfs_fs_info *fs_info, u64 root, u64 ino)
 224 {
 225         struct inode_defrag *entry = NULL;
 226         struct inode_defrag tmp;
 227         struct rb_node *p;
 228         struct rb_node *parent = NULL;
 229         int ret;
 230
 231         tmp.ino = ino;
 232         tmp.root = root;
 233
 234         spin_lock(&fs_info->defrag_inodes_lock);
 235         p = fs_info->defrag_inodes.rb_node;
 236         while (p) {
 237                 parent = p;
 238                 entry = rb_entry(parent, struct inode_defrag, rb_node);
 239
 240                 ret = __compare_inode_defrag(&tmp, entry);
 241                 if (ret < 0)
 242                         p = parent->rb_left;
 243                 else if (ret > 0)
 244                         p = parent->rb_right;
 245                 else
 246                         goto out;
 247         }
 248
 249         if (parent && __compare_inode_defrag(&tmp, entry) > 0) {
 250                 parent = rb_next(parent);
 251                 if (parent)
 252                         entry = rb_entry(parent, struct inode_defrag, rb_node);
 253                 else
 254                         entry = NULL;
 255         }
 256 out:
 257         if (entry)
 258                 rb_erase(parent, &fs_info->defrag_inodes);
 259         spin_unlock(&fs_info->defrag_inodes_lock);
 260         return entry;
 261 }
 262
 263 void btrfs_cleanup_defrag_inodes(struct btrfs_fs_info *fs_info)
 264 {
 265         struct inode_defrag *defrag;
 266         struct rb_node *node;
 267
 268         spin_lock(&fs_info->defrag_inodes_lock);
 269         node = rb_first(&fs_info->defrag_inodes);
 270         while (node) {
 271                 rb_erase(node, &fs_info->defrag_inodes);
 272                 defrag = rb_entry(node, struct inode_defrag, rb_node);
 273                 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
 274
 275                 if (need_resched()) {
 276                         spin_unlock(&fs_info->defrag_inodes_lock);
 277                         cond_resched();
 278                         spin_lock(&fs_info->defrag_inodes_lock);
 279                 }
 280
 281                 node = rb_first(&fs_info->defrag_inodes);
 282         }
 283         spin_unlock(&fs_info->defrag_inodes_lock);
 284 }
 285
 286 #define BTRFS_DEFRAG_BATCH      1024
 287
 288 static int __btrfs_run_defrag_inode(struct btrfs_fs_info *fs_info,
 289                                     struct inode_defrag *defrag)
 290 {
 291         struct btrfs_root *inode_root;
 292         struct inode *inode;
 293         struct btrfs_key key;
 294         struct btrfs_ioctl_defrag_range_args range;
 295         int num_defrag;
 296
 297         /* get the inode */
 298         key.objectid = defrag->root;
 299         btrfs_set_key_type(&key, BTRFS_ROOT_ITEM_KEY);
 300         key.offset = (u64)-1;
 301         inode_root = btrfs_read_fs_root_no_name(fs_info, &key);
 302         if (IS_ERR(inode_root)) {
 303                 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
 304                 return PTR_ERR(inode_root);
 305         }
 306
 307         key.objectid = defrag->ino;
 308         btrfs_set_key_type(&key, BTRFS_INODE_ITEM_KEY);
 309         key.offset = 0;
 310         inode = btrfs_iget(fs_info->sb, &key, inode_root, NULL);
 311         if (IS_ERR(inode)) {
 312                 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
 313                 return PTR_ERR(inode);
 314         }
 315
 316         /* do a chunk of defrag */
 317         clear_bit(BTRFS_INODE_IN_DEFRAG, &BTRFS_I(inode)->runtime_flags);
 318         memset(&range, 0, sizeof(range));
 319         range.len = (u64)-1;
 320         range.start = defrag->last_offset;
 321
 322         sb_start_write(fs_info->sb);
 323         num_defrag = btrfs_defrag_file(inode, NULL, &range, defrag->transid,
 324                                        BTRFS_DEFRAG_BATCH);
 325         sb_end_write(fs_info->sb);
 326         /*
 327          * if we filled the whole defrag batch, there
 328          * must be more work to do.  Queue this defrag
 329          * again
 330          */
 331         if (num_defrag == BTRFS_DEFRAG_BATCH) {
 332                 defrag->last_offset = range.start;
 333                 btrfs_requeue_inode_defrag(inode, defrag);
 334         } else if (defrag->last_offset && !defrag->cycled) {
 335                 /*
 336                  * we didn't fill our defrag batch, but
 337                  * we didn't start at zero.  Make sure we loop
 338                  * around to the start of the file.
 339                  */
 340                 defrag->last_offset = 0;
 341                 defrag->cycled = 1;
 342                 btrfs_requeue_inode_defrag(inode, defrag);
 343         } else {
 344                 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
 345         }
 346
 347         iput(inode);
 348         return 0;
 349 }
 350
 351 /*
 352  * run through the list of inodes in the FS that need
 353  * defragging
 354  */
 355 int btrfs_run_defrag_inodes(struct btrfs_fs_info *fs_info)
 356 {
 357         struct inode_defrag *defrag;
 358         u64 first_ino = 0;
 359         u64 root_objectid = 0;
 360
 361         atomic_inc(&fs_info->defrag_running);
 362         while(1) {
 363                 if (!__need_auto_defrag(fs_info->tree_root))
 364                         break;
 365
 366                 /* find an inode to defrag */
 367                 defrag = btrfs_pick_defrag_inode(fs_info, root_objectid,
 368                                                  first_ino);
 369                 if (!defrag) {
 370                         if (root_objectid || first_ino) {
 371                                 root_objectid = 0;
 372                                 first_ino = 0;
 373                                 continue;
 374                         } else {
 375                                 break;
 376                         }
 377                 }
 378
 379                 first_ino = defrag->ino + 1;
 380                 root_objectid = defrag->root;
 381
 382                 __btrfs_run_defrag_inode(fs_info, defrag);
 383         }
 384         atomic_dec(&fs_info->defrag_running);
 385
 386         /*
 387          * during unmount, we use the transaction_wait queue to
 388          * wait for the defragger to stop
 389          */
 390         wake_up(&fs_info->transaction_wait);
 391         return 0;
 392 }
 393
 394 /* simple helper to fault in pages and copy.  This should go away
 395  * and be replaced with calls into generic code.
 396  */
 397 static noinline int btrfs_copy_from_user(loff_t pos, int num_pages,
 398                                          size_t write_bytes,
 399                                          struct page **prepared_pages,
 400                                          struct iov_iter *i)
 401 {
 402         size_t copied = 0;
 403         size_t total_copied = 0;
 404         int pg = 0;
 405         int offset = pos & (PAGE_CACHE_SIZE - 1);
 406
 407         while (write_bytes > 0) {
 408                 size_t count = min_t(size_t,
 409                                      PAGE_CACHE_SIZE - offset, write_bytes);
 410                 struct page *page = prepared_pages[pg];
 411                 /*
 412                  * Copy data from userspace to the current page
 413                  *
 414                  * Disable pagefault to avoid recursive lock since
 415                  * the pages are already locked
 416                  */
 417                 pagefault_disable();
 418                 copied = iov_iter_copy_from_user_atomic(page, i, offset, count);
 419                 pagefault_enable();
 420
 421                 /* Flush processor's dcache for this page */
 422                 flush_dcache_page(page);
 423
 424                 /*
 425                  * if we get a partial write, we can end up with
 426                  * partially up to date pages.  These add
 427                  * a lot of complexity, so make sure they don't
 428                  * happen by forcing this copy to be retried.
 429                  *
 430                  * The rest of the btrfs_file_write code will fall
 431                  * back to page at a time copies after we return 0.
 432                  */
 433                 if (!PageUptodate(page) && copied < count)
 434                         copied = 0;
 435
 436                 iov_iter_advance(i, copied);
 437                 write_bytes -= copied;
 438                 total_copied += copied;
 439
 440                 /* Return to btrfs_file_aio_write to fault page */
 441                 if (unlikely(copied == 0))
 442                         break;
 443
 444                 if (unlikely(copied < PAGE_CACHE_SIZE - offset)) {
 445                         offset += copied;
 446                 } else {
 447                         pg++;
 448                         offset = 0;
 449                 }
 450         }
 451         return total_copied;
 452 }
 453
 454 /*
 455  * unlocks pages after btrfs_file_write is done with them
 456  */
 457 void btrfs_drop_pages(struct page **pages, size_t num_pages)
 458 {
 459         size_t i;
 460         for (i = 0; i < num_pages; i++) {
 461                 /* page checked is some magic around finding pages that
 462                  * have been modified without going through btrfs_set_page_dirty
 463                  * clear it here
 464                  */
 465                 ClearPageChecked(pages[i]);
 466                 unlock_page(pages[i]);
 467                 mark_page_accessed(pages[i]);
 468                 page_cache_release(pages[i]);
 469         }
 470 }
 471
 472 /*
 473  * after copy_from_user, pages need to be dirtied and we need to make
 474  * sure holes are created between the current EOF and the start of
 475  * any next extents (if required).
 476  *
 477  * this also makes the decision about creating an inline extent vs
 478  * doing real data extents, marking pages dirty and delalloc as required.
 479  */
 480 int btrfs_dirty_pages(struct btrfs_root *root, struct inode *inode,
 481                       struct page **pages, size_t num_pages,
 482                       loff_t pos, size_t write_bytes,
 483                       struct extent_state **cached)
 484 {
 485         int err = 0;
 486         int i;
 487         u64 num_bytes;
 488         u64 start_pos;
 489         u64 end_of_last_block;
 490         u64 end_pos = pos + write_bytes;
 491         loff_t isize = i_size_read(inode);
 492
 493         start_pos = pos & ~((u64)root->sectorsize - 1);
 494         num_bytes = (write_bytes + pos - start_pos +
 495                     root->sectorsize - 1) & ~((u64)root->sectorsize - 1);
 496
 497         end_of_last_block = start_pos + num_bytes - 1;
 498         err = btrfs_set_extent_delalloc(inode, start_pos, end_of_last_block,
 499                                         cached);
 500         if (err)
 501                 return err;
 502
 503         for (i = 0; i < num_pages; i++) {
 504                 struct page *p = pages[i];
 505                 SetPageUptodate(p);
 506                 ClearPageChecked(p);
 507                 set_page_dirty(p);
 508         }
 509
 510         /*
 511          * we've only changed i_size in ram, and we haven't updated
 512          * the disk i_size.  There is no need to log the inode
 513          * at this time.
 514          */
 515         if (end_pos > isize)
 516                 i_size_write(inode, end_pos);
 517         return 0;
 518 }
 519
 520 /*
 521  * this drops all the extents in the cache that intersect the range
 522  * [start, end].  Existing extents are split as required.
 523  */
 524 void btrfs_drop_extent_cache(struct inode *inode, u64 start, u64 end,
 525                              int skip_pinned)
 526 {
 527         struct extent_map *em;
 528         struct extent_map *split = NULL;
 529         struct extent_map *split2 = NULL;
 530         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
 531         u64 len = end - start + 1;
 532         u64 gen;
 533         int ret;
 534         int testend = 1;
 535         unsigned long flags;
 536         int compressed = 0;
 537
 538         WARN_ON(end < start);
 539         if (end == (u64)-1) {
 540                 len = (u64)-1;
 541                 testend = 0;
 542         }
 543         while (1) {
 544                 int no_splits = 0;
 545
 546                 if (!split)
 547                         split = alloc_extent_map();
 548                 if (!split2)
 549                         split2 = alloc_extent_map();
 550                 if (!split || !split2)
 551                         no_splits = 1;
 552
 553                 write_lock(&em_tree->lock);
 554                 em = lookup_extent_mapping(em_tree, start, len);
 555                 if (!em) {
 556                         write_unlock(&em_tree->lock);
 557                         break;
 558                 }
 559                 flags = em->flags;
 560                 gen = em->generation;
 561                 if (skip_pinned && test_bit(EXTENT_FLAG_PINNED, &em->flags)) {
 562                         if (testend && em->start + em->len >= start + len) {
 563                                 free_extent_map(em);
 564                                 write_unlock(&em_tree->lock);
 565                                 break;
 566                         }
 567                         start = em->start + em->len;
 568                         if (testend)
 569                                 len = start + len - (em->start + em->len);
 570                         free_extent_map(em);
 571                         write_unlock(&em_tree->lock);
 572                         continue;
 573                 }
 574                 compressed = test_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
 575                 clear_bit(EXTENT_FLAG_PINNED, &em->flags);
 576                 remove_extent_mapping(em_tree, em);
 577                 if (no_splits)
 578                         goto next;
 579
 580                 if (em->block_start < EXTENT_MAP_LAST_BYTE &&
 581                     em->start < start) {
 582                         split->start = em->start;
 583                         split->len = start - em->start;
 584                         split->orig_start = em->orig_start;
 585                         split->block_start = em->block_start;
 586
 587                         if (compressed)
 588                                 split->block_len = em->block_len;
 589                         else
 590                                 split->block_len = split->len;
 591                         split->generation = gen;
 592                         split->bdev = em->bdev;
 593                         split->flags = flags;
 594                         split->compress_type = em->compress_type;
 595                         ret = add_extent_mapping(em_tree, split);
 596                         BUG_ON(ret); /* Logic error */
 597                         list_move(&split->list, &em_tree->modified_extents);
 598                         free_extent_map(split);
 599                         split = split2;
 600                         split2 = NULL;
 601                 }
 602                 if (em->block_start < EXTENT_MAP_LAST_BYTE &&
 603                     testend && em->start + em->len > start + len) {
 604                         u64 diff = start + len - em->start;
 605
 606                         split->start = start + len;
 607                         split->len = em->start + em->len - (start + len);
 608                         split->bdev = em->bdev;
 609                         split->flags = flags;
 610                         split->compress_type = em->compress_type;
 611                         split->generation = gen;
 612
 613                         if (compressed) {
 614                                 split->block_len = em->block_len;
 615                                 split->block_start = em->block_start;
 616                                 split->orig_start = em->orig_start;
 617                         } else {
 618                                 split->block_len = split->len;
 619                                 split->block_start = em->block_start + diff;
 620                                 split->orig_start = split->start;
 621                         }
 622
 623                         ret = add_extent_mapping(em_tree, split);
 624                         BUG_ON(ret); /* Logic error */
 625                         list_move(&split->list, &em_tree->modified_extents);
 626                         free_extent_map(split);
 627                         split = NULL;
 628                 }
 629 next:
 630                 write_unlock(&em_tree->lock);
 631
 632                 /* once for us */
 633                 free_extent_map(em);
 634                 /* once for the tree*/
 635                 free_extent_map(em);
 636         }
 637         if (split)
 638                 free_extent_map(split);
 639         if (split2)
 640                 free_extent_map(split2);
 641 }
 642
 643 /*
 644  * this is very complex, but the basic idea is to drop all extents
 645  * in the range start - end.  hint_block is filled in with a block number
 646  * that would be a good hint to the block allocator for this file.
 647  *
 648  * If an extent intersects the range but is not entirely inside the range
 649  * it is either truncated or split.  Anything entirely inside the range
 650  * is deleted from the tree.
 651  */
 652 int __btrfs_drop_extents(struct btrfs_trans_handle *trans,
 653                          struct btrfs_root *root, struct inode *inode,
 654                          struct btrfs_path *path, u64 start, u64 end,
 655                          u64 *drop_end, int drop_cache)
 656 {
 657         struct extent_buffer *leaf;
 658         struct btrfs_file_extent_item *fi;
 659         struct btrfs_key key;
 660         struct btrfs_key new_key;
 661         u64 ino = btrfs_ino(inode);
 662         u64 search_start = start;
 663         u64 disk_bytenr = 0;
 664         u64 num_bytes = 0;
 665         u64 extent_offset = 0;
 666         u64 extent_end = 0;
 667         int del_nr = 0;
 668         int del_slot = 0;
 669         int extent_type;
 670         int recow;
 671         int ret;
 672         int modify_tree = -1;
 673         int update_refs = (root->ref_cows || root == root->fs_info->tree_root);
 674         int found = 0;
 675
 676         if (drop_cache)
 677                 btrfs_drop_extent_cache(inode, start, end - 1, 0);
 678
 679         if (start >= BTRFS_I(inode)->disk_i_size)
 680                 modify_tree = 0;
 681
 682         while (1) {
 683                 recow = 0;
 684                 ret = btrfs_lookup_file_extent(trans, root, path, ino,
 685                                                search_start, modify_tree);
 686                 if (ret < 0)
 687                         break;
 688                 if (ret > 0 && path->slots[0] > 0 && search_start == start) {
 689                         leaf = path->nodes[0];
 690                         btrfs_item_key_to_cpu(leaf, &key, path->slots[0] - 1);
 691                         if (key.objectid == ino &&
 692                             key.type == BTRFS_EXTENT_DATA_KEY)
 693                                 path->slots[0]--;
 694                 }
 695                 ret = 0;
 696 next_slot:
 697                 leaf = path->nodes[0];
 698                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
 699                         BUG_ON(del_nr > 0);
 700                         ret = btrfs_next_leaf(root, path);
 701                         if (ret < 0)
 702                                 break;
 703                         if (ret > 0) {
 704                                 ret = 0;
 705                                 break;
 706                         }
 707                         leaf = path->nodes[0];
 708                         recow = 1;
 709                 }
 710
 711                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 712                 if (key.objectid > ino ||
 713                     key.type > BTRFS_EXTENT_DATA_KEY || key.offset >= end)
 714                         break;
 715
 716                 fi = btrfs_item_ptr(leaf, path->slots[0],
 717                                     struct btrfs_file_extent_item);
 718                 extent_type = btrfs_file_extent_type(leaf, fi);
 719
 720                 if (extent_type == BTRFS_FILE_EXTENT_REG ||
 721                     extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
 722                         disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
 723                         num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
 724                         extent_offset = btrfs_file_extent_offset(leaf, fi);
 725                         extent_end = key.offset +
 726                                 btrfs_file_extent_num_bytes(leaf, fi);
 727                 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
 728                         extent_end = key.offset +
 729                                 btrfs_file_extent_inline_len(leaf, fi);
 730                 } else {
 731                         WARN_ON(1);
 732                         extent_end = search_start;
 733                 }
 734
 735                 if (extent_end <= search_start) {
 736                         path->slots[0]++;
 737                         goto next_slot;
 738                 }
 739
 740                 found = 1;
 741                 search_start = max(key.offset, start);
 742                 if (recow || !modify_tree) {
 743                         modify_tree = -1;
 744                         btrfs_release_path(path);
 745                         continue;
 746                 }
 747
 748                 /*
 749                  *     | - range to drop - |
 750                  *  | -------- extent -------- |
 751                  */
 752                 if (start > key.offset && end < extent_end) {
 753                         BUG_ON(del_nr > 0);
 754                         BUG_ON(extent_type == BTRFS_FILE_EXTENT_INLINE);
 755
 756                         memcpy(&new_key, &key, sizeof(new_key));
 757                         new_key.offset = start;
 758                         ret = btrfs_duplicate_item(trans, root, path,
 759                                                    &new_key);
 760                         if (ret == -EAGAIN) {
 761                                 btrfs_release_path(path);
 762                                 continue;
 763                         }
 764                         if (ret < 0)
 765                                 break;
 766
 767                         leaf = path->nodes[0];
 768                         fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
 769                                             struct btrfs_file_extent_item);
 770                         btrfs_set_file_extent_num_bytes(leaf, fi,
 771                                                         start - key.offset);
 772
 773                         fi = btrfs_item_ptr(leaf, path->slots[0],
 774                                             struct btrfs_file_extent_item);
 775
 776                         extent_offset += start - key.offset;
 777                         btrfs_set_file_extent_offset(leaf, fi, extent_offset);
 778                         btrfs_set_file_extent_num_bytes(leaf, fi,
 779                                                         extent_end - start);
 780                         btrfs_mark_buffer_dirty(leaf);
 781
 782                         if (update_refs && disk_bytenr > 0) {
 783                                 ret = btrfs_inc_extent_ref(trans, root,
 784                                                 disk_bytenr, num_bytes, 0,
 785                                                 root->root_key.objectid,
 786                                                 new_key.objectid,
 787                                                 start - extent_offset, 0);
 788                                 BUG_ON(ret); /* -ENOMEM */
 789                         }
 790                         key.offset = start;
 791                 }
 792                 /*
 793                  *  | ---- range to drop ----- |
 794                  *      | -------- extent -------- |
 795                  */
 796                 if (start <= key.offset && end < extent_end) {
 797                         BUG_ON(extent_type == BTRFS_FILE_EXTENT_INLINE);
 798
 799                         memcpy(&new_key, &key, sizeof(new_key));
 800                         new_key.offset = end;
 801                         btrfs_set_item_key_safe(trans, root, path, &new_key);
 802
 803                         extent_offset += end - key.offset;
 804                         btrfs_set_file_extent_offset(leaf, fi, extent_offset);
 805                         btrfs_set_file_extent_num_bytes(leaf, fi,
 806                                                         extent_end - end);
 807                         btrfs_mark_buffer_dirty(leaf);
 808                         if (update_refs && disk_bytenr > 0)
 809                                 inode_sub_bytes(inode, end - key.offset);
 810                         break;
 811                 }
 812
 813                 search_start = extent_end;
 814                 /*
 815                  *       | ---- range to drop ----- |
 816                  *  | -------- extent -------- |
 817                  */
 818                 if (start > key.offset && end >= extent_end) {
 819                         BUG_ON(del_nr > 0);
 820                         BUG_ON(extent_type == BTRFS_FILE_EXTENT_INLINE);
 821
 822                         btrfs_set_file_extent_num_bytes(leaf, fi,
 823                                                         start - key.offset);
 824                         btrfs_mark_buffer_dirty(leaf);
 825                         if (update_refs && disk_bytenr > 0)
 826                                 inode_sub_bytes(inode, extent_end - start);
 827                         if (end == extent_end)
 828                                 break;
 829
 830                         path->slots[0]++;
 831                         goto next_slot;
 832                 }
 833
 834                 /*
 835                  *  | ---- range to drop ----- |
 836                  *    | ------ extent ------ |
 837                  */
 838                 if (start <= key.offset && end >= extent_end) {
 839                         if (del_nr == 0) {
 840                                 del_slot = path->slots[0];
 841                                 del_nr = 1;
 842                         } else {
 843                                 BUG_ON(del_slot + del_nr != path->slots[0]);
 844                                 del_nr++;
 845                         }
 846
 847                         if (update_refs &&
 848                             extent_type == BTRFS_FILE_EXTENT_INLINE) {
 849                                 inode_sub_bytes(inode,
 850                                                 extent_end - key.offset);
 851                                 extent_end = ALIGN(extent_end,
 852                                                    root->sectorsize);
 853                         } else if (update_refs && disk_bytenr > 0) {
 854                                 ret = btrfs_free_extent(trans, root,
 855                                                 disk_bytenr, num_bytes, 0,
 856                                                 root->root_key.objectid,
 857                                                 key.objectid, key.offset -
 858                                                 extent_offset, 0);
 859                                 BUG_ON(ret); /* -ENOMEM */
 860                                 inode_sub_bytes(inode,
 861                                                 extent_end - key.offset);
 862                         }
 863
 864                         if (end == extent_end)
 865                                 break;
 866
 867                         if (path->slots[0] + 1 < btrfs_header_nritems(leaf)) {
 868                                 path->slots[0]++;
 869                                 goto next_slot;
 870                         }
 871
 872                         ret = btrfs_del_items(trans, root, path, del_slot,
 873                                               del_nr);
 874                         if (ret) {
 875                                 btrfs_abort_transaction(trans, root, ret);
 876                                 break;
 877                         }
 878
 879                         del_nr = 0;
 880                         del_slot = 0;
 881
 882                         btrfs_release_path(path);
 883                         continue;
 884                 }
 885
 886                 BUG_ON(1);
 887         }
 888
 889         if (!ret && del_nr > 0) {
 890                 ret = btrfs_del_items(trans, root, path, del_slot, del_nr);
 891                 if (ret)
 892                         btrfs_abort_transaction(trans, root, ret);
 893         }
 894
 895         if (drop_end)
 896                 *drop_end = found ? min(end, extent_end) : end;
 897         btrfs_release_path(path);
 898         return ret;
 899 }
 900
 901 int btrfs_drop_extents(struct btrfs_trans_handle *trans,
 902                        struct btrfs_root *root, struct inode *inode, u64 start,
 903                        u64 end, int drop_cache)
 904 {
 905         struct btrfs_path *path;
 906         int ret;
 907
 908         path = btrfs_alloc_path();
 909         if (!path)
 910                 return -ENOMEM;
 911         ret = __btrfs_drop_extents(trans, root, inode, path, start, end, NULL,
 912                                    drop_cache);
 913         btrfs_free_path(path);
 914         return ret;
 915 }
 916
 917 static int extent_mergeable(struct extent_buffer *leaf, int slot,
 918                             u64 objectid, u64 bytenr, u64 orig_offset,
 919                             u64 *start, u64 *end)
 920 {
 921         struct btrfs_file_extent_item *fi;
 922         struct btrfs_key key;
 923         u64 extent_end;
 924
 925         if (slot < 0 || slot >= btrfs_header_nritems(leaf))
 926                 return 0;
 927
 928         btrfs_item_key_to_cpu(leaf, &key, slot);
 929         if (key.objectid != objectid || key.type != BTRFS_EXTENT_DATA_KEY)
 930                 return 0;
 931
 932         fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
 933         if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG ||
 934             btrfs_file_extent_disk_bytenr(leaf, fi) != bytenr ||
 935             btrfs_file_extent_offset(leaf, fi) != key.offset - orig_offset ||
 936             btrfs_file_extent_compression(leaf, fi) ||
 937             btrfs_file_extent_encryption(leaf, fi) ||
 938             btrfs_file_extent_other_encoding(leaf, fi))
 939                 return 0;
 940
 941         extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
 942         if ((*start && *start != key.offset) || (*end && *end != extent_end))
 943                 return 0;
 944
 945         *start = key.offset;
 946         *end = extent_end;
 947         return 1;
 948 }
 949
 950 /*
 951  * Mark extent in the range start - end as written.
 952  *
 953  * This changes extent type from 'pre-allocated' to 'regular'. If only
 954  * part of extent is marked as written, the extent will be split into
 955  * two or three.
 956  */
 957 int btrfs_mark_extent_written(struct btrfs_trans_handle *trans,
 958                               struct inode *inode, u64 start, u64 end)
 959 {
 960         struct btrfs_root *root = BTRFS_I(inode)->root;
 961         struct extent_buffer *leaf;
 962         struct btrfs_path *path;
 963         struct btrfs_file_extent_item *fi;
 964         struct btrfs_key key;
 965         struct btrfs_key new_key;
 966         u64 bytenr;
 967         u64 num_bytes;
 968         u64 extent_end;
 969         u64 orig_offset;
 970         u64 other_start;
 971         u64 other_end;
 972         u64 split;
 973         int del_nr = 0;
 974         int del_slot = 0;
 975         int recow;
 976         int ret;
 977         u64 ino = btrfs_ino(inode);
 978
 979         path = btrfs_alloc_path();
 980         if (!path)
 981                 return -ENOMEM;
 982 again:
 983         recow = 0;
 984         split = start;
 985         key.objectid = ino;
 986         key.type = BTRFS_EXTENT_DATA_KEY;
 987         key.offset = split;
 988
 989         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
 990         if (ret < 0)
 991                 goto out;
 992         if (ret > 0 && path->slots[0] > 0)
 993                 path->slots[0]--;
 994
 995         leaf = path->nodes[0];
 996         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 997         BUG_ON(key.objectid != ino || key.type != BTRFS_EXTENT_DATA_KEY);
 998         fi = btrfs_item_ptr(leaf, path->slots[0],
 999                             struct btrfs_file_extent_item);
1000         BUG_ON(btrfs_file_extent_type(leaf, fi) !=
1001                BTRFS_FILE_EXTENT_PREALLOC);
1002         extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
1003         BUG_ON(key.offset > start || extent_end < end);
1004
1005         bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1006         num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
1007         orig_offset = key.offset - btrfs_file_extent_offset(leaf, fi);
1008         memcpy(&new_key, &key, sizeof(new_key));
1009
1010         if (start == key.offset && end < extent_end) {
1011                 other_start = 0;
1012                 other_end = start;
1013                 if (extent_mergeable(leaf, path->slots[0] - 1,
1014                                      ino, bytenr, orig_offset,
1015                                      &other_start, &other_end)) {
1016                         new_key.offset = end;
1017                         btrfs_set_item_key_safe(trans, root, path, &new_key);
1018                         fi = btrfs_item_ptr(leaf, path->slots[0],
1019                                             struct btrfs_file_extent_item);
1020                         btrfs_set_file_extent_generation(leaf, fi,
1021                                                          trans->transid);
1022                         btrfs_set_file_extent_num_bytes(leaf, fi,
1023                                                         extent_end - end);
1024                         btrfs_set_file_extent_offset(leaf, fi,
1025                                                      end - orig_offset);
1026                         fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
1027                                             struct btrfs_file_extent_item);
1028                         btrfs_set_file_extent_generation(leaf, fi,
1029                                                          trans->transid);
1030                         btrfs_set_file_extent_num_bytes(leaf, fi,
1031                                                         end - other_start);
1032                         btrfs_mark_buffer_dirty(leaf);
1033                         goto out;
1034                 }
1035         }
1036
1037         if (start > key.offset && end == extent_end) {
1038                 other_start = end;
1039                 other_end = 0;
1040                 if (extent_mergeable(leaf, path->slots[0] + 1,
1041                                      ino, bytenr, orig_offset,
1042                                      &other_start, &other_end)) {
1043                         fi = btrfs_item_ptr(leaf, path->slots[0],
1044                                             struct btrfs_file_extent_item);
1045                         btrfs_set_file_extent_num_bytes(leaf, fi,
1046                                                         start - key.offset);
1047                         btrfs_set_file_extent_generation(leaf, fi,
1048                                                          trans->transid);
1049                         path->slots[0]++;
1050                         new_key.offset = start;
1051                         btrfs_set_item_key_safe(trans, root, path, &new_key);
1052
1053                         fi = btrfs_item_ptr(leaf, path->slots[0],
1054                                             struct btrfs_file_extent_item);
1055                         btrfs_set_file_extent_generation(leaf, fi,
1056                                                          trans->transid);
1057                         btrfs_set_file_extent_num_bytes(leaf, fi,
1058                                                         other_end - start);
1059                         btrfs_set_file_extent_offset(leaf, fi,
1060                                                      start - orig_offset);
1061                         btrfs_mark_buffer_dirty(leaf);
1062                         goto out;
1063                 }
1064         }
1065
1066         while (start > key.offset || end < extent_end) {
1067                 if (key.offset == start)
1068                         split = end;
1069
1070                 new_key.offset = split;
1071                 ret = btrfs_duplicate_item(trans, root, path, &new_key);
1072                 if (ret == -EAGAIN) {
1073                         btrfs_release_path(path);
1074                         goto again;
1075                 }
1076                 if (ret < 0) {
1077                         btrfs_abort_transaction(trans, root, ret);
1078                         goto out;
1079                 }
1080
1081                 leaf = path->nodes[0];
1082                 fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
1083                                     struct btrfs_file_extent_item);
1084                 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1085                 btrfs_set_file_extent_num_bytes(leaf, fi,
1086                                                 split - key.offset);
1087
1088                 fi = btrfs_item_ptr(leaf, path->slots[0],
1089                                     struct btrfs_file_extent_item);
1090
1091                 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1092                 btrfs_set_file_extent_offset(leaf, fi, split - orig_offset);
1093                 btrfs_set_file_extent_num_bytes(leaf, fi,
1094                                                 extent_end - split);
1095                 btrfs_mark_buffer_dirty(leaf);
1096
1097                 ret = btrfs_inc_extent_ref(trans, root, bytenr, num_bytes, 0,
1098                                            root->root_key.objectid,
1099                                            ino, orig_offset, 0);
1100                 BUG_ON(ret); /* -ENOMEM */
1101
1102                 if (split == start) {
1103                         key.offset = start;
1104                 } else {
1105                         BUG_ON(start != key.offset);
1106                         path->slots[0]--;
1107                         extent_end = end;
1108                 }
1109                 recow = 1;
1110         }
1111
1112         other_start = end;
1113         other_end = 0;
1114         if (extent_mergeable(leaf, path->slots[0] + 1,
1115                              ino, bytenr, orig_offset,
1116                              &other_start, &other_end)) {
1117                 if (recow) {
1118                         btrfs_release_path(path);
1119                         goto again;
1120                 }
1121                 extent_end = other_end;
1122                 del_slot = path->slots[0] + 1;
1123                 del_nr++;
1124                 ret = btrfs_free_extent(trans, root, bytenr, num_bytes,
1125                                         0, root->root_key.objectid,
1126                                         ino, orig_offset, 0);
1127                 BUG_ON(ret); /* -ENOMEM */
1128         }
1129         other_start = 0;
1130         other_end = start;
1131         if (extent_mergeable(leaf, path->slots[0] - 1,
1132                              ino, bytenr, orig_offset,
1133                              &other_start, &other_end)) {
1134                 if (recow) {
1135                         btrfs_release_path(path);
1136                         goto again;
1137                 }
1138                 key.offset = other_start;
1139                 del_slot = path->slots[0];
1140                 del_nr++;
1141                 ret = btrfs_free_extent(trans, root, bytenr, num_bytes,
1142                                         0, root->root_key.objectid,
1143                                         ino, orig_offset, 0);
1144                 BUG_ON(ret); /* -ENOMEM */
1145         }
1146         if (del_nr == 0) {
1147                 fi = btrfs_item_ptr(leaf, path->slots[0],
1148                            struct btrfs_file_extent_item);
1149                 btrfs_set_file_extent_type(leaf, fi,
1150                                            BTRFS_FILE_EXTENT_REG);
1151                 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1152                 btrfs_mark_buffer_dirty(leaf);
1153         } else {
1154                 fi = btrfs_item_ptr(leaf, del_slot - 1,
1155                            struct btrfs_file_extent_item);
1156                 btrfs_set_file_extent_type(leaf, fi,
1157                                            BTRFS_FILE_EXTENT_REG);
1158                 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1159                 btrfs_set_file_extent_num_bytes(leaf, fi,
1160                                                 extent_end - key.offset);
1161                 btrfs_mark_buffer_dirty(leaf);
1162
1163                 ret = btrfs_del_items(trans, root, path, del_slot, del_nr);
1164                 if (ret < 0) {
1165                         btrfs_abort_transaction(trans, root, ret);
1166                         goto out;
1167                 }
1168         }
1169 out:
1170         btrfs_free_path(path);
1171         return 0;
1172 }
1173
1174 /*
1175  * on error we return an unlocked page and the error value
1176  * on success we return a locked page and 0
1177  */
1178 static int prepare_uptodate_page(struct page *page, u64 pos,
1179                                  bool force_uptodate)
1180 {
1181         int ret = 0;
1182
1183         if (((pos & (PAGE_CACHE_SIZE - 1)) || force_uptodate) &&
1184             !PageUptodate(page)) {
1185                 ret = btrfs_readpage(NULL, page);
1186                 if (ret)
1187                         return ret;
1188                 lock_page(page);
1189                 if (!PageUptodate(page)) {
1190                         unlock_page(page);
1191                         return -EIO;
1192                 }
1193         }
1194         return 0;
1195 }
1196
1197 /*
1198  * this gets pages into the page cache and locks them down, it also properly
1199  * waits for data=ordered extents to finish before allowing the pages to be
1200  * modified.
1201  */
1202 static noinline int prepare_pages(struct btrfs_root *root, struct file *file,
1203                          struct page **pages, size_t num_pages,
1204                          loff_t pos, unsigned long first_index,
1205                          size_t write_bytes, bool force_uptodate)
1206 {
1207         struct extent_state *cached_state = NULL;
1208         int i;
1209         unsigned long index = pos >> PAGE_CACHE_SHIFT;
1210         struct inode *inode = fdentry(file)->d_inode;
1211         gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping);
1212         int err = 0;
1213         int faili = 0;
1214         u64 start_pos;
1215         u64 last_pos;
1216
1217         start_pos = pos & ~((u64)root->sectorsize - 1);
1218         last_pos = ((u64)index + num_pages) << PAGE_CACHE_SHIFT;
1219
1220 again:
1221         for (i = 0; i < num_pages; i++) {
1222                 pages[i] = find_or_create_page(inode->i_mapping, index + i,
1223                                                mask | __GFP_WRITE);
1224                 if (!pages[i]) {
1225                         faili = i - 1;
1226                         err = -ENOMEM;
1227                         goto fail;
1228                 }
1229
1230                 if (i == 0)
1231                         err = prepare_uptodate_page(pages[i], pos,
1232                                                     force_uptodate);
1233                 if (i == num_pages - 1)
1234                         err = prepare_uptodate_page(pages[i],
1235                                                     pos + write_bytes, false);
1236                 if (err) {
1237                         page_cache_release(pages[i]);
1238                         faili = i - 1;
1239                         goto fail;
1240                 }
1241                 wait_on_page_writeback(pages[i]);
1242         }
1243         err = 0;
1244         if (start_pos < inode->i_size) {
1245                 struct btrfs_ordered_extent *ordered;
1246                 lock_extent_bits(&BTRFS_I(inode)->io_tree,
1247                                  start_pos, last_pos - 1, 0, &cached_state);
1248                 ordered = btrfs_lookup_first_ordered_extent(inode,
1249                                                             last_pos - 1);
1250                 if (ordered &&
1251                     ordered->file_offset + ordered->len > start_pos &&
1252                     ordered->file_offset < last_pos) {
1253                         btrfs_put_ordered_extent(ordered);
1254                         unlock_extent_cached(&BTRFS_I(inode)->io_tree,
1255                                              start_pos, last_pos - 1,
1256                                              &cached_state, GFP_NOFS);
1257                         for (i = 0; i < num_pages; i++) {
1258                                 unlock_page(pages[i]);
1259                                 page_cache_release(pages[i]);
1260                         }
1261                         btrfs_wait_ordered_range(inode, start_pos,
1262                                                  last_pos - start_pos);
1263                         goto again;
1264                 }
1265                 if (ordered)
1266                         btrfs_put_ordered_extent(ordered);
1267
1268                 clear_extent_bit(&BTRFS_I(inode)->io_tree, start_pos,
1269                                   last_pos - 1, EXTENT_DIRTY | EXTENT_DELALLOC |
1270                                   EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG,
1271                                   0, 0, &cached_state, GFP_NOFS);
1272                 unlock_extent_cached(&BTRFS_I(inode)->io_tree,
1273                                      start_pos, last_pos - 1, &cached_state,
1274                                      GFP_NOFS);
1275         }
1276         for (i = 0; i < num_pages; i++) {
1277                 if (clear_page_dirty_for_io(pages[i]))
1278                         account_page_redirty(pages[i]);
1279                 set_page_extent_mapped(pages[i]);
1280                 WARN_ON(!PageLocked(pages[i]));
1281         }
1282         return 0;
1283 fail:
1284         while (faili >= 0) {
1285                 unlock_page(pages[faili]);
1286                 page_cache_release(pages[faili]);
1287                 faili--;
1288         }
1289         return err;
1290
1291 }
1292
1293 static noinline ssize_t __btrfs_buffered_write(struct file *file,
1294                                                struct iov_iter *i,
1295                                                loff_t pos)
1296 {
1297         struct inode *inode = fdentry(file)->d_inode;
1298         struct btrfs_root *root = BTRFS_I(inode)->root;
1299         struct page **pages = NULL;
1300         unsigned long first_index;
1301         size_t num_written = 0;
1302         int nrptrs;
1303         int ret = 0;
1304         bool force_page_uptodate = false;
1305
1306         nrptrs = min((iov_iter_count(i) + PAGE_CACHE_SIZE - 1) /
1307                      PAGE_CACHE_SIZE, PAGE_CACHE_SIZE /
1308                      (sizeof(struct page *)));
1309         nrptrs = min(nrptrs, current->nr_dirtied_pause - current->nr_dirtied);
1310         nrptrs = max(nrptrs, 8);
1311         pages = kmalloc(nrptrs * sizeof(struct page *), GFP_KERNEL);
1312         if (!pages)
1313                 return -ENOMEM;
1314
1315         first_index = pos >> PAGE_CACHE_SHIFT;
1316
1317         while (iov_iter_count(i) > 0) {
1318                 size_t offset = pos & (PAGE_CACHE_SIZE - 1);
1319                 size_t write_bytes = min(iov_iter_count(i),
1320                                          nrptrs * (size_t)PAGE_CACHE_SIZE -
1321                                          offset);
1322                 size_t num_pages = (write_bytes + offset +
1323                                     PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1324                 size_t dirty_pages;
1325                 size_t copied;
1326
1327                 WARN_ON(num_pages > nrptrs);
1328
1329                 /*
1330                  * Fault pages before locking them in prepare_pages
1331                  * to avoid recursive lock
1332                  */
1333                 if (unlikely(iov_iter_fault_in_readable(i, write_bytes))) {
1334                         ret = -EFAULT;
1335                         break;
1336                 }
1337
1338                 ret = btrfs_delalloc_reserve_space(inode,
1339                                         num_pages << PAGE_CACHE_SHIFT);
1340                 if (ret)
1341                         break;
1342
1343                 /*
1344                  * This is going to setup the pages array with the number of
1345                  * pages we want, so we don't really need to worry about the
1346                  * contents of pages from loop to loop
1347                  */
1348                 ret = prepare_pages(root, file, pages, num_pages,
1349                                     pos, first_index, write_bytes,
1350                                     force_page_uptodate);
1351                 if (ret) {
1352                         btrfs_delalloc_release_space(inode,
1353                                         num_pages << PAGE_CACHE_SHIFT);
1354                         break;
1355                 }
1356
1357                 copied = btrfs_copy_from_user(pos, num_pages,
1358                                            write_bytes, pages, i);
1359
1360                 /*
1361                  * if we have trouble faulting in the pages, fall
1362                  * back to one page at a time
1363                  */
1364                 if (copied < write_bytes)
1365                         nrptrs = 1;
1366
1367                 if (copied == 0) {
1368                         force_page_uptodate = true;
1369                         dirty_pages = 0;
1370                 } else {
1371                         force_page_uptodate = false;
1372                         dirty_pages = (copied + offset +
1373                                        PAGE_CACHE_SIZE - 1) >>
1374                                        PAGE_CACHE_SHIFT;
1375                 }
1376
1377                 /*
1378                  * If we had a short copy we need to release the excess delaloc
1379                  * bytes we reserved.  We need to increment outstanding_extents
1380                  * because btrfs_delalloc_release_space will decrement it, but
1381                  * we still have an outstanding extent for the chunk we actually
1382                  * managed to copy.
1383                  */
1384                 if (num_pages > dirty_pages) {
1385                         if (copied > 0) {
1386                                 spin_lock(&BTRFS_I(inode)->lock);
1387                                 BTRFS_I(inode)->outstanding_extents++;
1388                                 spin_unlock(&BTRFS_I(inode)->lock);
1389                         }
1390                         btrfs_delalloc_release_space(inode,
1391                                         (num_pages - dirty_pages) <<
1392                                         PAGE_CACHE_SHIFT);
1393                 }
1394
1395                 if (copied > 0) {
1396                         ret = btrfs_dirty_pages(root, inode, pages,
1397                                                 dirty_pages, pos, copied,
1398                                                 NULL);
1399                         if (ret) {
1400                                 btrfs_delalloc_release_space(inode,
1401                                         dirty_pages << PAGE_CACHE_SHIFT);
1402                                 btrfs_drop_pages(pages, num_pages);
1403                                 break;
1404                         }
1405                 }
1406
1407                 btrfs_drop_pages(pages, num_pages);
1408
1409                 cond_resched();
1410
1411                 balance_dirty_pages_ratelimited_nr(inode->i_mapping,
1412                                                    dirty_pages);
1413                 if (dirty_pages < (root->leafsize >> PAGE_CACHE_SHIFT) + 1)
1414                         btrfs_btree_balance_dirty(root);
1415
1416                 pos += copied;
1417                 num_written += copied;
1418         }
1419
1420         kfree(pages);
1421
1422         return num_written ? num_written : ret;
1423 }
1424
1425 static ssize_t __btrfs_direct_write(struct kiocb *iocb,
1426                                     const struct iovec *iov,
1427                                     unsigned long nr_segs, loff_t pos,
1428                                     loff_t *ppos, size_t count, size_t ocount)
1429 {
1430         struct file *file = iocb->ki_filp;
1431         struct iov_iter i;
1432         ssize_t written;
1433         ssize_t written_buffered;
1434         loff_t endbyte;
1435         int err;
1436
1437         written = generic_file_direct_write(iocb, iov, &nr_segs, pos, ppos,
1438                                             count, ocount);
1439
1440         if (written < 0 || written == count)
1441                 return written;
1442
1443         pos += written;
1444         count -= written;
1445         iov_iter_init(&i, iov, nr_segs, count, written);
1446         written_buffered = __btrfs_buffered_write(file, &i, pos);
1447         if (written_buffered < 0) {
1448                 err = written_buffered;
1449                 goto out;
1450         }
1451         endbyte = pos + written_buffered - 1;
1452         err = filemap_write_and_wait_range(file->f_mapping, pos, endbyte);
1453         if (err)
1454                 goto out;
1455         written += written_buffered;
1456         *ppos = pos + written_buffered;
1457         invalidate_mapping_pages(file->f_mapping, pos >> PAGE_CACHE_SHIFT,
1458                                  endbyte >> PAGE_CACHE_SHIFT);
1459 out:
1460         return written ? written : err;
1461 }
1462
1463 static ssize_t btrfs_file_aio_write(struct kiocb *iocb,
1464                                     const struct iovec *iov,
1465                                     unsigned long nr_segs, loff_t pos)
1466 {
1467         struct file *file = iocb->ki_filp;
1468         struct inode *inode = fdentry(file)->d_inode;
1469         struct btrfs_root *root = BTRFS_I(inode)->root;
1470         loff_t *ppos = &iocb->ki_pos;
1471         u64 start_pos;
1472         ssize_t num_written = 0;
1473         ssize_t err = 0;
1474         size_t count, ocount;
1475
1476         sb_start_write(inode->i_sb);
1477
1478         mutex_lock(&inode->i_mutex);
1479
1480         err = generic_segment_checks(iov, &nr_segs, &ocount, VERIFY_READ);
1481         if (err) {
1482                 mutex_unlock(&inode->i_mutex);
1483                 goto out;
1484         }
1485         count = ocount;
1486
1487         current->backing_dev_info = inode->i_mapping->backing_dev_info;
1488         err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode));
1489         if (err) {
1490                 mutex_unlock(&inode->i_mutex);
1491                 goto out;
1492         }
1493
1494         if (count == 0) {
1495                 mutex_unlock(&inode->i_mutex);
1496                 goto out;
1497         }
1498
1499         err = file_remove_suid(file);
1500         if (err) {
1501                 mutex_unlock(&inode->i_mutex);
1502                 goto out;
1503         }
1504
1505         /*
1506          * If BTRFS flips readonly due to some impossible error
1507          * (fs_info->fs_state now has BTRFS_SUPER_FLAG_ERROR),
1508          * although we have opened a file as writable, we have
1509          * to stop this write operation to ensure FS consistency.
1510          */
1511         if (root->fs_info->fs_state & BTRFS_SUPER_FLAG_ERROR) {
1512                 mutex_unlock(&inode->i_mutex);
1513                 err = -EROFS;
1514                 goto out;
1515         }
1516
1517         err = file_update_time(file);
1518         if (err) {
1519                 mutex_unlock(&inode->i_mutex);
1520                 goto out;
1521         }
1522
1523         start_pos = round_down(pos, root->sectorsize);
1524         if (start_pos > i_size_read(inode)) {
1525                 err = btrfs_cont_expand(inode, i_size_read(inode), start_pos);
1526                 if (err) {
1527                         mutex_unlock(&inode->i_mutex);
1528                         goto out;
1529                 }
1530         }
1531
1532         if (unlikely(file->f_flags & O_DIRECT)) {
1533                 num_written = __btrfs_direct_write(iocb, iov, nr_segs,
1534                                                    pos, ppos, count, ocount);
1535         } else {
1536                 struct iov_iter i;
1537
1538                 iov_iter_init(&i, iov, nr_segs, count, num_written);
1539
1540                 num_written = __btrfs_buffered_write(file, &i, pos);
1541                 if (num_written > 0)
1542                         *ppos = pos + num_written;
1543         }
1544
1545         mutex_unlock(&inode->i_mutex);
1546
1547         /*
1548          * we want to make sure fsync finds this change
1549          * but we haven't joined a transaction running right now.
1550          *
1551          * Later on, someone is sure to update the inode and get the
1552          * real transid recorded.
1553          *
1554          * We set last_trans now to the fs_info generation + 1,
1555          * this will either be one more than the running transaction
1556          * or the generation used for the next transaction if there isn't
1557          * one running right now.
1558          */
1559         BTRFS_I(inode)->last_trans = root->fs_info->generation + 1;
1560         if (num_written > 0 || num_written == -EIOCBQUEUED) {
1561                 err = generic_write_sync(file, pos, num_written);
1562                 if (err < 0 && num_written > 0)
1563                         num_written = err;
1564         }
1565 out:
1566         sb_end_write(inode->i_sb);
1567         current->backing_dev_info = NULL;
1568         return num_written ? num_written : err;
1569 }
1570
1571 int btrfs_release_file(struct inode *inode, struct file *filp)
1572 {
1573         /*
1574          * ordered_data_close is set by settattr when we are about to truncate
1575          * a file from a non-zero size to a zero size.  This tries to
1576          * flush down new bytes that may have been written if the
1577          * application were using truncate to replace a file in place.
1578          */
1579         if (test_and_clear_bit(BTRFS_INODE_ORDERED_DATA_CLOSE,
1580                                &BTRFS_I(inode)->runtime_flags)) {
1581                 btrfs_add_ordered_operation(NULL, BTRFS_I(inode)->root, inode);
1582                 if (inode->i_size > BTRFS_ORDERED_OPERATIONS_FLUSH_LIMIT)
1583                         filemap_flush(inode->i_mapping);
1584         }
1585         if (filp->private_data)
1586                 btrfs_ioctl_trans_end(filp);
1587         return 0;
1588 }
1589
1590 /*
1591  * fsync call for both files and directories.  This logs the inode into
1592  * the tree log instead of forcing full commits whenever possible.
1593  *
1594  * It needs to call filemap_fdatawait so that all ordered extent updates are
1595  * in the metadata btree are up to date for copying to the log.
1596  *
1597  * It drops the inode mutex before doing the tree log commit.  This is an
1598  * important optimization for directories because holding the mutex prevents
1599  * new operations on the dir while we write to disk.
1600  */
1601 int btrfs_sync_file(struct file *file, loff_t start, loff_t end, int datasync)
1602 {
1603         struct dentry *dentry = file->f_path.dentry;
1604         struct inode *inode = dentry->d_inode;
1605         struct btrfs_root *root = BTRFS_I(inode)->root;
1606         int ret = 0;
1607         struct btrfs_trans_handle *trans;
1608
1609         trace_btrfs_sync_file(file, datasync);
1610
1611         /*
1612          * We write the dirty pages in the range and wait until they complete
1613          * out of the ->i_mutex. If so, we can flush the dirty pages by
1614          * multi-task, and make the performance up.
1615          */
1616         ret = filemap_write_and_wait_range(inode->i_mapping, start, end);
1617         if (ret)
1618                 return ret;
1619
1620         mutex_lock(&inode->i_mutex);
1621
1622         /*
1623          * We flush the dirty pages again to avoid some dirty pages in the
1624          * range being left.
1625          */
1626         atomic_inc(&root->log_batch);
1627         btrfs_wait_ordered_range(inode, start, end - start + 1);
1628         atomic_inc(&root->log_batch);
1629
1630         /*
1631          * check the transaction that last modified this inode
1632          * and see if its already been committed
1633          */
1634         if (!BTRFS_I(inode)->last_trans) {
1635                 mutex_unlock(&inode->i_mutex);
1636                 goto out;
1637         }
1638
1639         /*
1640          * if the last transaction that changed this file was before
1641          * the current transaction, we can bail out now without any
1642          * syncing
1643          */
1644         smp_mb();
1645         if (btrfs_inode_in_log(inode, root->fs_info->generation) ||
1646             BTRFS_I(inode)->last_trans <=
1647             root->fs_info->last_trans_committed) {
1648                 BTRFS_I(inode)->last_trans = 0;
1649
1650                 /*
1651                  * We'v had everything committed since the last time we were
1652                  * modified so clear this flag in case it was set for whatever
1653                  * reason, it's no longer relevant.
1654                  */
1655                 clear_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
1656                           &BTRFS_I(inode)->runtime_flags);
1657                 mutex_unlock(&inode->i_mutex);
1658                 goto out;
1659         }
1660
1661         /*
1662          * ok we haven't committed the transaction yet, lets do a commit
1663          */
1664         if (file->private_data)
1665                 btrfs_ioctl_trans_end(file);
1666
1667         trans = btrfs_start_transaction(root, 0);
1668         if (IS_ERR(trans)) {
1669                 ret = PTR_ERR(trans);
1670                 mutex_unlock(&inode->i_mutex);
1671                 goto out;
1672         }
1673
1674         ret = btrfs_log_dentry_safe(trans, root, dentry);
1675         if (ret < 0) {
1676                 mutex_unlock(&inode->i_mutex);
1677                 goto out;
1678         }
1679
1680         /* we've logged all the items and now have a consistent
1681          * version of the file in the log.  It is possible that
1682          * someone will come in and modify the file, but that's
1683          * fine because the log is consistent on disk, and we
1684          * have references to all of the file's extents
1685          *
1686          * It is possible that someone will come in and log the
1687          * file again, but that will end up using the synchronization
1688          * inside btrfs_sync_log to keep things safe.
1689          */
1690         mutex_unlock(&inode->i_mutex);
1691
1692         if (ret != BTRFS_NO_LOG_SYNC) {
1693                 if (ret > 0) {
1694                         ret = btrfs_commit_transaction(trans, root);
1695                 } else {
1696                         ret = btrfs_sync_log(trans, root);
1697                         if (ret == 0)
1698                                 ret = btrfs_end_transaction(trans, root);
1699                         else
1700                                 ret = btrfs_commit_transaction(trans, root);
1701                 }
1702         } else {
1703                 ret = btrfs_end_transaction(trans, root);
1704         }
1705 out:
1706         return ret > 0 ? -EIO : ret;
1707 }
1708
1709 static const struct vm_operations_struct btrfs_file_vm_ops = {
1710         .fault          = filemap_fault,
1711         .page_mkwrite   = btrfs_page_mkwrite,
1712         .remap_pages    = generic_file_remap_pages,
1713 };
1714
1715 static int btrfs_file_mmap(struct file  *filp, struct vm_area_struct *vma)
1716 {
1717         struct address_space *mapping = filp->f_mapping;
1718
1719         if (!mapping->a_ops->readpage)
1720                 return -ENOEXEC;
1721
1722         file_accessed(filp);
1723         vma->vm_ops = &btrfs_file_vm_ops;
1724
1725         return 0;
1726 }
1727
1728 static int hole_mergeable(struct inode *inode, struct extent_buffer *leaf,
1729                           int slot, u64 start, u64 end)
1730 {
1731         struct btrfs_file_extent_item *fi;
1732         struct btrfs_key key;
1733
1734         if (slot < 0 || slot >= btrfs_header_nritems(leaf))
1735                 return 0;
1736
1737         btrfs_item_key_to_cpu(leaf, &key, slot);
1738         if (key.objectid != btrfs_ino(inode) ||
1739             key.type != BTRFS_EXTENT_DATA_KEY)
1740                 return 0;
1741
1742         fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
1743
1744         if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG)
1745                 return 0;
1746
1747         if (btrfs_file_extent_disk_bytenr(leaf, fi))
1748                 return 0;
1749
1750         if (key.offset == end)
1751                 return 1;
1752         if (key.offset + btrfs_file_extent_num_bytes(leaf, fi) == start)
1753                 return 1;
1754         return 0;
1755 }
1756
1757 static int fill_holes(struct btrfs_trans_handle *trans, struct inode *inode,
1758                       struct btrfs_path *path, u64 offset, u64 end)
1759 {
1760         struct btrfs_root *root = BTRFS_I(inode)->root;
1761         struct extent_buffer *leaf;
1762         struct btrfs_file_extent_item *fi;
1763         struct extent_map *hole_em;
1764         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
1765         struct btrfs_key key;
1766         int ret;
1767
1768         key.objectid = btrfs_ino(inode);
1769         key.type = BTRFS_EXTENT_DATA_KEY;
1770         key.offset = offset;
1771
1772
1773         ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
1774         if (ret < 0)
1775                 return ret;
1776         BUG_ON(!ret);
1777
1778         leaf = path->nodes[0];
1779         if (hole_mergeable(inode, leaf, path->slots[0]-1, offset, end)) {
1780                 u64 num_bytes;
1781
1782                 path->slots[0]--;
1783                 fi = btrfs_item_ptr(leaf, path->slots[0],
1784                                     struct btrfs_file_extent_item);
1785                 num_bytes = btrfs_file_extent_num_bytes(leaf, fi) +
1786                         end - offset;
1787                 btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
1788                 btrfs_set_file_extent_ram_bytes(leaf, fi, num_bytes);
1789                 btrfs_set_file_extent_offset(leaf, fi, 0);
1790                 btrfs_mark_buffer_dirty(leaf);
1791                 goto out;
1792         }
1793
1794         if (hole_mergeable(inode, leaf, path->slots[0]+1, offset, end)) {
1795                 u64 num_bytes;
1796
1797                 path->slots[0]++;
1798                 key.offset = offset;
1799                 btrfs_set_item_key_safe(trans, root, path, &key);
1800                 fi = btrfs_item_ptr(leaf, path->slots[0],
1801                                     struct btrfs_file_extent_item);
1802                 num_bytes = btrfs_file_extent_num_bytes(leaf, fi) + end -
1803                         offset;
1804                 btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
1805                 btrfs_set_file_extent_ram_bytes(leaf, fi, num_bytes);
1806                 btrfs_set_file_extent_offset(leaf, fi, 0);
1807                 btrfs_mark_buffer_dirty(leaf);
1808                 goto out;
1809         }
1810         btrfs_release_path(path);
1811
1812         ret = btrfs_insert_file_extent(trans, root, btrfs_ino(inode), offset,
1813                                        0, 0, end - offset, 0, end - offset,
1814                                        0, 0, 0);
1815         if (ret)
1816                 return ret;
1817
1818 out:
1819         btrfs_release_path(path);
1820
1821         hole_em = alloc_extent_map();
1822         if (!hole_em) {
1823                 btrfs_drop_extent_cache(inode, offset, end - 1, 0);
1824                 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
1825                         &BTRFS_I(inode)->runtime_flags);
1826         } else {
1827                 hole_em->start = offset;
1828                 hole_em->len = end - offset;
1829                 hole_em->orig_start = offset;
1830
1831                 hole_em->block_start = EXTENT_MAP_HOLE;
1832                 hole_em->block_len = 0;
1833                 hole_em->bdev = root->fs_info->fs_devices->latest_bdev;
1834                 hole_em->compress_type = BTRFS_COMPRESS_NONE;
1835                 hole_em->generation = trans->transid;
1836
1837                 do {
1838                         btrfs_drop_extent_cache(inode, offset, end - 1, 0);
1839                         write_lock(&em_tree->lock);
1840                         ret = add_extent_mapping(em_tree, hole_em);
1841                         if (!ret)
1842                                 list_move(&hole_em->list,
1843                                           &em_tree->modified_extents);
1844                         write_unlock(&em_tree->lock);
1845                 } while (ret == -EEXIST);
1846                 free_extent_map(hole_em);
1847                 if (ret)
1848                         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
1849                                 &BTRFS_I(inode)->runtime_flags);
1850         }
1851
1852         return 0;
1853 }
1854
1855 static int btrfs_punch_hole(struct inode *inode, loff_t offset, loff_t len)
1856 {
1857         struct btrfs_root *root = BTRFS_I(inode)->root;
1858         struct extent_state *cached_state = NULL;
1859         struct btrfs_path *path;
1860         struct btrfs_block_rsv *rsv;
1861         struct btrfs_trans_handle *trans;
1862         u64 mask = BTRFS_I(inode)->root->sectorsize - 1;
1863         u64 lockstart = (offset + mask) & ~mask;
1864         u64 lockend = ((offset + len) & ~mask) - 1;
1865         u64 cur_offset = lockstart;
1866         u64 min_size = btrfs_calc_trunc_metadata_size(root, 1);
1867         u64 drop_end;
1868         int ret = 0;
1869         int err = 0;
1870         bool same_page = ((offset >> PAGE_CACHE_SHIFT) ==
1871                           ((offset + len - 1) >> PAGE_CACHE_SHIFT));
1872
1873         btrfs_wait_ordered_range(inode, offset, len);
1874
1875         mutex_lock(&inode->i_mutex);
1876         if (offset >= inode->i_size) {
1877                 mutex_unlock(&inode->i_mutex);
1878                 return 0;
1879         }
1880
1881         /*
1882          * Only do this if we are in the same page and we aren't doing the
1883          * entire page.
1884          */
1885         if (same_page && len < PAGE_CACHE_SIZE) {
1886                 ret = btrfs_truncate_page(inode, offset, len, 0);
1887                 mutex_unlock(&inode->i_mutex);
1888                 return ret;
1889         }
1890
1891         /* zero back part of the first page */
1892         ret = btrfs_truncate_page(inode, offset, 0, 0);
1893         if (ret) {
1894                 mutex_unlock(&inode->i_mutex);
1895                 return ret;
1896         }
1897
1898         /* zero the front end of the last page */
1899         ret = btrfs_truncate_page(inode, offset + len, 0, 1);
1900         if (ret) {
1901                 mutex_unlock(&inode->i_mutex);
1902                 return ret;
1903         }
1904
1905         if (lockend < lockstart) {
1906                 mutex_unlock(&inode->i_mutex);
1907                 return 0;
1908         }
1909
1910         while (1) {
1911                 struct btrfs_ordered_extent *ordered;
1912
1913                 truncate_pagecache_range(inode, lockstart, lockend);
1914
1915                 lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend,
1916                                  0, &cached_state);
1917                 ordered = btrfs_lookup_first_ordered_extent(inode, lockend);
1918
1919                 /*
1920                  * We need to make sure we have no ordered extents in this range
1921                  * and nobody raced in and read a page in this range, if we did
1922                  * we need to try again.
1923                  */
1924                 if ((!ordered ||
1925                     (ordered->file_offset + ordered->len < lockstart ||
1926                      ordered->file_offset > lockend)) &&
1927                      !test_range_bit(&BTRFS_I(inode)->io_tree, lockstart,
1928                                      lockend, EXTENT_UPTODATE, 0,
1929                                      cached_state)) {
1930                         if (ordered)
1931                                 btrfs_put_ordered_extent(ordered);
1932                         break;
1933                 }
1934                 if (ordered)
1935                         btrfs_put_ordered_extent(ordered);
1936                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart,
1937                                      lockend, &cached_state, GFP_NOFS);
1938                 btrfs_wait_ordered_range(inode, lockstart,
1939                                          lockend - lockstart + 1);
1940         }
1941
1942         path = btrfs_alloc_path();
1943         if (!path) {
1944                 ret = -ENOMEM;
1945                 goto out;
1946         }
1947
1948         rsv = btrfs_alloc_block_rsv(root, BTRFS_BLOCK_RSV_TEMP);
1949         if (!rsv) {
1950                 ret = -ENOMEM;
1951                 goto out_free;
1952         }
1953         rsv->size = btrfs_calc_trunc_metadata_size(root, 1);
1954         rsv->failfast = 1;
1955
1956         /*
1957          * 1 - update the inode
1958          * 1 - removing the extents in the range
1959          * 1 - adding the hole extent
1960          */
1961         trans = btrfs_start_transaction(root, 3);
1962         if (IS_ERR(trans)) {
1963                 err = PTR_ERR(trans);
1964                 goto out_free;
1965         }
1966
1967         ret = btrfs_block_rsv_migrate(&root->fs_info->trans_block_rsv, rsv,
1968                                       min_size);
1969         BUG_ON(ret);
1970         trans->block_rsv = rsv;
1971
1972         while (cur_offset < lockend) {
1973                 ret = __btrfs_drop_extents(trans, root, inode, path,
1974                                            cur_offset, lockend + 1,
1975                                            &drop_end, 1);
1976                 if (ret != -ENOSPC)
1977                         break;
1978
1979                 trans->block_rsv = &root->fs_info->trans_block_rsv;
1980
1981                 ret = fill_holes(trans, inode, path, cur_offset, drop_end);
1982                 if (ret) {
1983                         err = ret;
1984                         break;
1985                 }
1986
1987                 cur_offset = drop_end;
1988
1989                 ret = btrfs_update_inode(trans, root, inode);
1990                 if (ret) {
1991                         err = ret;
1992                         break;
1993                 }
1994
1995                 btrfs_end_transaction(trans, root);
1996                 btrfs_btree_balance_dirty(root);
1997
1998                 trans = btrfs_start_transaction(root, 3);
1999                 if (IS_ERR(trans)) {
2000                         ret = PTR_ERR(trans);
2001                         trans = NULL;
2002                         break;
2003                 }
2004
2005                 ret = btrfs_block_rsv_migrate(&root->fs_info->trans_block_rsv,
2006                                               rsv, min_size);
2007                 BUG_ON(ret);    /* shouldn't happen */
2008                 trans->block_rsv = rsv;
2009         }
2010
2011         if (ret) {
2012                 err = ret;
2013                 goto out_trans;
2014         }
2015
2016         trans->block_rsv = &root->fs_info->trans_block_rsv;
2017         ret = fill_holes(trans, inode, path, cur_offset, drop_end);
2018         if (ret) {
2019                 err = ret;
2020                 goto out_trans;
2021         }
2022
2023 out_trans:
2024         if (!trans)
2025                 goto out_free;
2026
2027         inode_inc_iversion(inode);
2028         inode->i_mtime = inode->i_ctime = CURRENT_TIME;
2029
2030         trans->block_rsv = &root->fs_info->trans_block_rsv;
2031         ret = btrfs_update_inode(trans, root, inode);
2032         btrfs_end_transaction(trans, root);
2033         btrfs_btree_balance_dirty(root);
2034 out_free:
2035         btrfs_free_path(path);
2036         btrfs_free_block_rsv(root, rsv);
2037 out:
2038         unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2039                              &cached_state, GFP_NOFS);
2040         mutex_unlock(&inode->i_mutex);
2041         if (ret && !err)
2042                 err = ret;
2043         return err;
2044 }
2045
2046 static long btrfs_fallocate(struct file *file, int mode,
2047                             loff_t offset, loff_t len)
2048 {
2049         struct inode *inode = file->f_path.dentry->d_inode;
2050         struct extent_state *cached_state = NULL;
2051         u64 cur_offset;
2052         u64 last_byte;
2053         u64 alloc_start;
2054         u64 alloc_end;
2055         u64 alloc_hint = 0;
2056         u64 locked_end;
2057         struct extent_map *em;
2058         int blocksize = BTRFS_I(inode)->root->sectorsize;
2059         int ret;
2060
2061         alloc_start = round_down(offset, blocksize);
2062         alloc_end = round_up(offset + len, blocksize);
2063
2064         /* Make sure we aren't being give some crap mode */
2065         if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
2066                 return -EOPNOTSUPP;
2067
2068         if (mode & FALLOC_FL_PUNCH_HOLE)
2069                 return btrfs_punch_hole(inode, offset, len);
2070
2071         /*
2072          * Make sure we have enough space before we do the
2073          * allocation.
2074          */
2075         ret = btrfs_check_data_free_space(inode, alloc_end - alloc_start);
2076         if (ret)
2077                 return ret;
2078
2079         /*
2080          * wait for ordered IO before we have any locks.  We'll loop again
2081          * below with the locks held.
2082          */
2083         btrfs_wait_ordered_range(inode, alloc_start, alloc_end - alloc_start);
2084
2085         mutex_lock(&inode->i_mutex);
2086         ret = inode_newsize_ok(inode, alloc_end);
2087         if (ret)
2088                 goto out;
2089
2090         if (alloc_start > inode->i_size) {
2091                 ret = btrfs_cont_expand(inode, i_size_read(inode),
2092                                         alloc_start);
2093                 if (ret)
2094                         goto out;
2095         }
2096
2097         locked_end = alloc_end - 1;
2098         while (1) {
2099                 struct btrfs_ordered_extent *ordered;
2100
2101                 /* the extent lock is ordered inside the running
2102                  * transaction
2103                  */
2104                 lock_extent_bits(&BTRFS_I(inode)->io_tree, alloc_start,
2105                                  locked_end, 0, &cached_state);
2106                 ordered = btrfs_lookup_first_ordered_extent(inode,
2107                                                             alloc_end - 1);
2108                 if (ordered &&
2109                     ordered->file_offset + ordered->len > alloc_start &&
2110                     ordered->file_offset < alloc_end) {
2111                         btrfs_put_ordered_extent(ordered);
2112                         unlock_extent_cached(&BTRFS_I(inode)->io_tree,
2113                                              alloc_start, locked_end,
2114                                              &cached_state, GFP_NOFS);
2115                         /*
2116                          * we can't wait on the range with the transaction
2117                          * running or with the extent lock held
2118                          */
2119                         btrfs_wait_ordered_range(inode, alloc_start,
2120                                                  alloc_end - alloc_start);
2121                 } else {
2122                         if (ordered)
2123                                 btrfs_put_ordered_extent(ordered);
2124                         break;
2125                 }
2126         }
2127
2128         cur_offset = alloc_start;
2129         while (1) {
2130                 u64 actual_end;
2131
2132                 em = btrfs_get_extent(inode, NULL, 0, cur_offset,
2133                                       alloc_end - cur_offset, 0);
2134                 if (IS_ERR_OR_NULL(em)) {
2135                         if (!em)
2136                                 ret = -ENOMEM;
2137                         else
2138                                 ret = PTR_ERR(em);
2139                         break;
2140                 }
2141                 last_byte = min(extent_map_end(em), alloc_end);
2142                 actual_end = min_t(u64, extent_map_end(em), offset + len);
2143                 last_byte = ALIGN(last_byte, blocksize);
2144
2145                 if (em->block_start == EXTENT_MAP_HOLE ||
2146                     (cur_offset >= inode->i_size &&
2147                      !test_bit(EXTENT_FLAG_PREALLOC, &em->flags))) {
2148                         ret = btrfs_prealloc_file_range(inode, mode, cur_offset,
2149                                                         last_byte - cur_offset,
2150                                                         1 << inode->i_blkbits,
2151                                                         offset + len,
2152                                                         &alloc_hint);
2153
2154                         if (ret < 0) {
2155                                 free_extent_map(em);
2156                                 break;
2157                         }
2158                 } else if (actual_end > inode->i_size &&
2159                            !(mode & FALLOC_FL_KEEP_SIZE)) {
2160                         /*
2161                          * We didn't need to allocate any more space, but we
2162                          * still extended the size of the file so we need to
2163                          * update i_size.
2164                          */
2165                         inode->i_ctime = CURRENT_TIME;
2166                         i_size_write(inode, actual_end);
2167                         btrfs_ordered_update_i_size(inode, actual_end, NULL);
2168                 }
2169                 free_extent_map(em);
2170
2171                 cur_offset = last_byte;
2172                 if (cur_offset >= alloc_end) {
2173                         ret = 0;
2174                         break;
2175                 }
2176         }
2177         unlock_extent_cached(&BTRFS_I(inode)->io_tree, alloc_start, locked_end,
2178                              &cached_state, GFP_NOFS);
2179 out:
2180         mutex_unlock(&inode->i_mutex);
2181         /* Let go of our reservation. */
2182         btrfs_free_reserved_data_space(inode, alloc_end - alloc_start);
2183         return ret;
2184 }
2185
2186 static int find_desired_extent(struct inode *inode, loff_t *offset, int origin)
2187 {
2188         struct btrfs_root *root = BTRFS_I(inode)->root;
2189         struct extent_map *em;
2190         struct extent_state *cached_state = NULL;
2191         u64 lockstart = *offset;
2192         u64 lockend = i_size_read(inode);
2193         u64 start = *offset;
2194         u64 orig_start = *offset;
2195         u64 len = i_size_read(inode);
2196         u64 last_end = 0;
2197         int ret = 0;
2198
2199         lockend = max_t(u64, root->sectorsize, lockend);
2200         if (lockend <= lockstart)
2201                 lockend = lockstart + root->sectorsize;
2202
2203         len = lockend - lockstart + 1;
2204
2205         len = max_t(u64, len, root->sectorsize);
2206         if (inode->i_size == 0)
2207                 return -ENXIO;
2208
2209         lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend, 0,
2210                          &cached_state);
2211
2212         /*
2213          * Delalloc is such a pain.  If we have a hole and we have pending
2214          * delalloc for a portion of the hole we will get back a hole that
2215          * exists for the entire range since it hasn't been actually written
2216          * yet.  So to take care of this case we need to look for an extent just
2217          * before the position we want in case there is outstanding delalloc
2218          * going on here.
2219          */
2220         if (origin == SEEK_HOLE && start != 0) {
2221                 if (start <= root->sectorsize)
2222                         em = btrfs_get_extent_fiemap(inode, NULL, 0, 0,
2223                                                      root->sectorsize, 0);
2224                 else
2225                         em = btrfs_get_extent_fiemap(inode, NULL, 0,
2226                                                      start - root->sectorsize,
2227                                                      root->sectorsize, 0);
2228                 if (IS_ERR(em)) {
2229                         ret = PTR_ERR(em);
2230                         goto out;
2231                 }
2232                 last_end = em->start + em->len;
2233                 if (em->block_start == EXTENT_MAP_DELALLOC)
2234                         last_end = min_t(u64, last_end, inode->i_size);
2235                 free_extent_map(em);
2236         }
2237
2238         while (1) {
2239                 em = btrfs_get_extent_fiemap(inode, NULL, 0, start, len, 0);
2240                 if (IS_ERR(em)) {
2241                         ret = PTR_ERR(em);
2242                         break;
2243                 }
2244
2245                 if (em->block_start == EXTENT_MAP_HOLE) {
2246                         if (test_bit(EXTENT_FLAG_VACANCY, &em->flags)) {
2247                                 if (last_end <= orig_start) {
2248                                         free_extent_map(em);
2249                                         ret = -ENXIO;
2250                                         break;
2251                                 }
2252                         }
2253
2254                         if (origin == SEEK_HOLE) {
2255                                 *offset = start;
2256                                 free_extent_map(em);
2257                                 break;
2258                         }
2259                 } else {
2260                         if (origin == SEEK_DATA) {
2261                                 if (em->block_start == EXTENT_MAP_DELALLOC) {
2262                                         if (start >= inode->i_size) {
2263                                                 free_extent_map(em);
2264                                                 ret = -ENXIO;
2265                                                 break;
2266                                         }
2267                                 }
2268
2269                                 *offset = start;
2270                                 free_extent_map(em);
2271                                 break;
2272                         }
2273                 }
2274
2275                 start = em->start + em->len;
2276                 last_end = em->start + em->len;
2277
2278                 if (em->block_start == EXTENT_MAP_DELALLOC)
2279                         last_end = min_t(u64, last_end, inode->i_size);
2280
2281                 if (test_bit(EXTENT_FLAG_VACANCY, &em->flags)) {
2282                         free_extent_map(em);
2283                         ret = -ENXIO;
2284                         break;
2285                 }
2286                 free_extent_map(em);
2287                 cond_resched();
2288         }
2289         if (!ret)
2290                 *offset = min(*offset, inode->i_size);
2291 out:
2292         unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2293                              &cached_state, GFP_NOFS);
2294         return ret;
2295 }
2296
2297 static loff_t btrfs_file_llseek(struct file *file, loff_t offset, int origin)
2298 {
2299         struct inode *inode = file->f_mapping->host;
2300         int ret;
2301
2302         mutex_lock(&inode->i_mutex);
2303         switch (origin) {
2304         case SEEK_END:
2305         case SEEK_CUR:
2306                 offset = generic_file_llseek(file, offset, origin);
2307                 goto out;
2308         case SEEK_DATA:
2309         case SEEK_HOLE:
2310                 if (offset >= i_size_read(inode)) {
2311                         mutex_unlock(&inode->i_mutex);
2312                         return -ENXIO;
2313                 }
2314
2315                 ret = find_desired_extent(inode, &offset, origin);
2316                 if (ret) {
2317                         mutex_unlock(&inode->i_mutex);
2318                         return ret;
2319                 }
2320         }
2321
2322         if (offset < 0 && !(file->f_mode & FMODE_UNSIGNED_OFFSET)) {
2323                 offset = -EINVAL;
2324                 goto out;
2325         }
2326         if (offset > inode->i_sb->s_maxbytes) {
2327                 offset = -EINVAL;
2328                 goto out;
2329         }
2330
2331         /* Special lock needed here? */
2332         if (offset != file->f_pos) {
2333                 file->f_pos = offset;
2334                 file->f_version = 0;
2335         }
2336 out:
2337         mutex_unlock(&inode->i_mutex);
2338         return offset;
2339 }
2340
2341 const struct file_operations btrfs_file_operations = {
2342         .llseek         = btrfs_file_llseek,
2343         .read           = do_sync_read,
2344         .write          = do_sync_write,
2345         .aio_read       = generic_file_aio_read,
2346         .splice_read    = generic_file_splice_read,
2347         .aio_write      = btrfs_file_aio_write,
2348         .mmap           = btrfs_file_mmap,
2349         .open           = generic_file_open,
2350         .release        = btrfs_release_file,
2351         .fsync          = btrfs_sync_file,
2352         .fallocate      = btrfs_fallocate,
2353         .unlocked_ioctl = btrfs_ioctl,
2354 #ifdef CONFIG_COMPAT
2355         .compat_ioctl   = btrfs_ioctl,
2356 #endif
2357 };
2358
2359 void btrfs_auto_defrag_exit(void)
2360 {
2361         if (btrfs_inode_defrag_cachep)
2362                 kmem_cache_destroy(btrfs_inode_defrag_cachep);
2363 }
2364
2365 int btrfs_auto_defrag_init(void)
2366 {
2367         btrfs_inode_defrag_cachep = kmem_cache_create("btrfs_inode_defrag",
2368                                         sizeof(struct inode_defrag), 0,
2369                                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD,
2370                                         NULL);
2371         if (!btrfs_inode_defrag_cachep)
2372                 return -ENOMEM;
2373
2374         return 0;
2375 }