Pileus Git - ~andy/linux/blob - 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 <linux/btrfs.h>
  34 #include "ctree.h"
  35 #include "disk-io.h"
  36 #include "transaction.h"
  37 #include "btrfs_inode.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         int index;
 297         int ret;
 298
 299         /* get the inode */
 300         key.objectid = defrag->root;
 301         btrfs_set_key_type(&key, BTRFS_ROOT_ITEM_KEY);
 302         key.offset = (u64)-1;
 303
 304         index = srcu_read_lock(&fs_info->subvol_srcu);
 305
 306         inode_root = btrfs_read_fs_root_no_name(fs_info, &key);
 307         if (IS_ERR(inode_root)) {
 308                 ret = PTR_ERR(inode_root);
 309                 goto cleanup;
 310         }
 311         if (btrfs_root_refs(&inode_root->root_item) == 0) {
 312                 ret = -ENOENT;
 313                 goto cleanup;
 314         }
 315
 316         key.objectid = defrag->ino;
 317         btrfs_set_key_type(&key, BTRFS_INODE_ITEM_KEY);
 318         key.offset = 0;
 319         inode = btrfs_iget(fs_info->sb, &key, inode_root, NULL);
 320         if (IS_ERR(inode)) {
 321                 ret = PTR_ERR(inode);
 322                 goto cleanup;
 323         }
 324         srcu_read_unlock(&fs_info->subvol_srcu, index);
 325
 326         /* do a chunk of defrag */
 327         clear_bit(BTRFS_INODE_IN_DEFRAG, &BTRFS_I(inode)->runtime_flags);
 328         memset(&range, 0, sizeof(range));
 329         range.len = (u64)-1;
 330         range.start = defrag->last_offset;
 331
 332         sb_start_write(fs_info->sb);
 333         num_defrag = btrfs_defrag_file(inode, NULL, &range, defrag->transid,
 334                                        BTRFS_DEFRAG_BATCH);
 335         sb_end_write(fs_info->sb);
 336         /*
 337          * if we filled the whole defrag batch, there
 338          * must be more work to do.  Queue this defrag
 339          * again
 340          */
 341         if (num_defrag == BTRFS_DEFRAG_BATCH) {
 342                 defrag->last_offset = range.start;
 343                 btrfs_requeue_inode_defrag(inode, defrag);
 344         } else if (defrag->last_offset && !defrag->cycled) {
 345                 /*
 346                  * we didn't fill our defrag batch, but
 347                  * we didn't start at zero.  Make sure we loop
 348                  * around to the start of the file.
 349                  */
 350                 defrag->last_offset = 0;
 351                 defrag->cycled = 1;
 352                 btrfs_requeue_inode_defrag(inode, defrag);
 353         } else {
 354                 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
 355         }
 356
 357         iput(inode);
 358         return 0;
 359 cleanup:
 360         srcu_read_unlock(&fs_info->subvol_srcu, index);
 361         kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
 362         return ret;
 363 }
 364
 365 /*
 366  * run through the list of inodes in the FS that need
 367  * defragging
 368  */
 369 int btrfs_run_defrag_inodes(struct btrfs_fs_info *fs_info)
 370 {
 371         struct inode_defrag *defrag;
 372         u64 first_ino = 0;
 373         u64 root_objectid = 0;
 374
 375         atomic_inc(&fs_info->defrag_running);
 376         while(1) {
 377                 /* Pause the auto defragger. */
 378                 if (test_bit(BTRFS_FS_STATE_REMOUNTING,
 379                              &fs_info->fs_state))
 380                         break;
 381
 382                 if (!__need_auto_defrag(fs_info->tree_root))
 383                         break;
 384
 385                 /* find an inode to defrag */
 386                 defrag = btrfs_pick_defrag_inode(fs_info, root_objectid,
 387                                                  first_ino);
 388                 if (!defrag) {
 389                         if (root_objectid || first_ino) {
 390                                 root_objectid = 0;
 391                                 first_ino = 0;
 392                                 continue;
 393                         } else {
 394                                 break;
 395                         }
 396                 }
 397
 398                 first_ino = defrag->ino + 1;
 399                 root_objectid = defrag->root;
 400
 401                 __btrfs_run_defrag_inode(fs_info, defrag);
 402         }
 403         atomic_dec(&fs_info->defrag_running);
 404
 405         /*
 406          * during unmount, we use the transaction_wait queue to
 407          * wait for the defragger to stop
 408          */
 409         wake_up(&fs_info->transaction_wait);
 410         return 0;
 411 }
 412
 413 /* simple helper to fault in pages and copy.  This should go away
 414  * and be replaced with calls into generic code.
 415  */
 416 static noinline int btrfs_copy_from_user(loff_t pos, int num_pages,
 417                                          size_t write_bytes,
 418                                          struct page **prepared_pages,
 419                                          struct iov_iter *i)
 420 {
 421         size_t copied = 0;
 422         size_t total_copied = 0;
 423         int pg = 0;
 424         int offset = pos & (PAGE_CACHE_SIZE - 1);
 425
 426         while (write_bytes > 0) {
 427                 size_t count = min_t(size_t,
 428                                      PAGE_CACHE_SIZE - offset, write_bytes);
 429                 struct page *page = prepared_pages[pg];
 430                 /*
 431                  * Copy data from userspace to the current page
 432                  *
 433                  * Disable pagefault to avoid recursive lock since
 434                  * the pages are already locked
 435                  */
 436                 pagefault_disable();
 437                 copied = iov_iter_copy_from_user_atomic(page, i, offset, count);
 438                 pagefault_enable();
 439
 440                 /* Flush processor's dcache for this page */
 441                 flush_dcache_page(page);
 442
 443                 /*
 444                  * if we get a partial write, we can end up with
 445                  * partially up to date pages.  These add
 446                  * a lot of complexity, so make sure they don't
 447                  * happen by forcing this copy to be retried.
 448                  *
 449                  * The rest of the btrfs_file_write code will fall
 450                  * back to page at a time copies after we return 0.
 451                  */
 452                 if (!PageUptodate(page) && copied < count)
 453                         copied = 0;
 454
 455                 iov_iter_advance(i, copied);
 456                 write_bytes -= copied;
 457                 total_copied += copied;
 458
 459                 /* Return to btrfs_file_aio_write to fault page */
 460                 if (unlikely(copied == 0))
 461                         break;
 462
 463                 if (unlikely(copied < PAGE_CACHE_SIZE - offset)) {
 464                         offset += copied;
 465                 } else {
 466                         pg++;
 467                         offset = 0;
 468                 }
 469         }
 470         return total_copied;
 471 }
 472
 473 /*
 474  * unlocks pages after btrfs_file_write is done with them
 475  */
 476 void btrfs_drop_pages(struct page **pages, size_t num_pages)
 477 {
 478         size_t i;
 479         for (i = 0; i < num_pages; i++) {
 480                 /* page checked is some magic around finding pages that
 481                  * have been modified without going through btrfs_set_page_dirty
 482                  * clear it here
 483                  */
 484                 ClearPageChecked(pages[i]);
 485                 unlock_page(pages[i]);
 486                 mark_page_accessed(pages[i]);
 487                 page_cache_release(pages[i]);
 488         }
 489 }
 490
 491 /*
 492  * after copy_from_user, pages need to be dirtied and we need to make
 493  * sure holes are created between the current EOF and the start of
 494  * any next extents (if required).
 495  *
 496  * this also makes the decision about creating an inline extent vs
 497  * doing real data extents, marking pages dirty and delalloc as required.
 498  */
 499 int btrfs_dirty_pages(struct btrfs_root *root, struct inode *inode,
 500                       struct page **pages, size_t num_pages,
 501                       loff_t pos, size_t write_bytes,
 502                       struct extent_state **cached)
 503 {
 504         int err = 0;
 505         int i;
 506         u64 num_bytes;
 507         u64 start_pos;
 508         u64 end_of_last_block;
 509         u64 end_pos = pos + write_bytes;
 510         loff_t isize = i_size_read(inode);
 511
 512         start_pos = pos & ~((u64)root->sectorsize - 1);
 513         num_bytes = ALIGN(write_bytes + pos - start_pos, root->sectorsize);
 514
 515         end_of_last_block = start_pos + num_bytes - 1;
 516         err = btrfs_set_extent_delalloc(inode, start_pos, end_of_last_block,
 517                                         cached);
 518         if (err)
 519                 return err;
 520
 521         for (i = 0; i < num_pages; i++) {
 522                 struct page *p = pages[i];
 523                 SetPageUptodate(p);
 524                 ClearPageChecked(p);
 525                 set_page_dirty(p);
 526         }
 527
 528         /*
 529          * we've only changed i_size in ram, and we haven't updated
 530          * the disk i_size.  There is no need to log the inode
 531          * at this time.
 532          */
 533         if (end_pos > isize)
 534                 i_size_write(inode, end_pos);
 535         return 0;
 536 }
 537
 538 /*
 539  * this drops all the extents in the cache that intersect the range
 540  * [start, end].  Existing extents are split as required.
 541  */
 542 void btrfs_drop_extent_cache(struct inode *inode, u64 start, u64 end,
 543                              int skip_pinned)
 544 {
 545         struct extent_map *em;
 546         struct extent_map *split = NULL;
 547         struct extent_map *split2 = NULL;
 548         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
 549         u64 len = end - start + 1;
 550         u64 gen;
 551         int ret;
 552         int testend = 1;
 553         unsigned long flags;
 554         int compressed = 0;
 555
 556         WARN_ON(end < start);
 557         if (end == (u64)-1) {
 558                 len = (u64)-1;
 559                 testend = 0;
 560         }
 561         while (1) {
 562                 int no_splits = 0;
 563
 564                 if (!split)
 565                         split = alloc_extent_map();
 566                 if (!split2)
 567                         split2 = alloc_extent_map();
 568                 if (!split || !split2)
 569                         no_splits = 1;
 570
 571                 write_lock(&em_tree->lock);
 572                 em = lookup_extent_mapping(em_tree, start, len);
 573                 if (!em) {
 574                         write_unlock(&em_tree->lock);
 575                         break;
 576                 }
 577                 flags = em->flags;
 578                 gen = em->generation;
 579                 if (skip_pinned && test_bit(EXTENT_FLAG_PINNED, &em->flags)) {
 580                         if (testend && em->start + em->len >= start + len) {
 581                                 free_extent_map(em);
 582                                 write_unlock(&em_tree->lock);
 583                                 break;
 584                         }
 585                         start = em->start + em->len;
 586                         if (testend)
 587                                 len = start + len - (em->start + em->len);
 588                         free_extent_map(em);
 589                         write_unlock(&em_tree->lock);
 590                         continue;
 591                 }
 592                 compressed = test_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
 593                 clear_bit(EXTENT_FLAG_PINNED, &em->flags);
 594                 remove_extent_mapping(em_tree, em);
 595                 if (no_splits)
 596                         goto next;
 597
 598                 if (em->block_start < EXTENT_MAP_LAST_BYTE &&
 599                     em->start < start) {
 600                         split->start = em->start;
 601                         split->len = start - em->start;
 602                         split->orig_start = em->orig_start;
 603                         split->block_start = em->block_start;
 604
 605                         if (compressed)
 606                                 split->block_len = em->block_len;
 607                         else
 608                                 split->block_len = split->len;
 609                         split->orig_block_len = max(split->block_len,
 610                                                     em->orig_block_len);
 611                         split->generation = gen;
 612                         split->bdev = em->bdev;
 613                         split->flags = flags;
 614                         split->compress_type = em->compress_type;
 615                         ret = add_extent_mapping(em_tree, split);
 616                         BUG_ON(ret); /* Logic error */
 617                         list_move(&split->list, &em_tree->modified_extents);
 618                         free_extent_map(split);
 619                         split = split2;
 620                         split2 = NULL;
 621                 }
 622                 if (em->block_start < EXTENT_MAP_LAST_BYTE &&
 623                     testend && em->start + em->len > start + len) {
 624                         u64 diff = start + len - em->start;
 625
 626                         split->start = start + len;
 627                         split->len = em->start + em->len - (start + len);
 628                         split->bdev = em->bdev;
 629                         split->flags = flags;
 630                         split->compress_type = em->compress_type;
 631                         split->generation = gen;
 632                         split->orig_block_len = max(em->block_len,
 633                                                     em->orig_block_len);
 634
 635                         if (compressed) {
 636                                 split->block_len = em->block_len;
 637                                 split->block_start = em->block_start;
 638                                 split->orig_start = em->orig_start;
 639                         } else {
 640                                 split->block_len = split->len;
 641                                 split->block_start = em->block_start + diff;
 642                                 split->orig_start = em->orig_start;
 643                         }
 644
 645                         ret = add_extent_mapping(em_tree, split);
 646                         BUG_ON(ret); /* Logic error */
 647                         list_move(&split->list, &em_tree->modified_extents);
 648                         free_extent_map(split);
 649                         split = NULL;
 650                 }
 651 next:
 652                 write_unlock(&em_tree->lock);
 653
 654                 /* once for us */
 655                 free_extent_map(em);
 656                 /* once for the tree*/
 657                 free_extent_map(em);
 658         }
 659         if (split)
 660                 free_extent_map(split);
 661         if (split2)
 662                 free_extent_map(split2);
 663 }
 664
 665 /*
 666  * this is very complex, but the basic idea is to drop all extents
 667  * in the range start - end.  hint_block is filled in with a block number
 668  * that would be a good hint to the block allocator for this file.
 669  *
 670  * If an extent intersects the range but is not entirely inside the range
 671  * it is either truncated or split.  Anything entirely inside the range
 672  * is deleted from the tree.
 673  */
 674 int __btrfs_drop_extents(struct btrfs_trans_handle *trans,
 675                          struct btrfs_root *root, struct inode *inode,
 676                          struct btrfs_path *path, u64 start, u64 end,
 677                          u64 *drop_end, int drop_cache)
 678 {
 679         struct extent_buffer *leaf;
 680         struct btrfs_file_extent_item *fi;
 681         struct btrfs_key key;
 682         struct btrfs_key new_key;
 683         u64 ino = btrfs_ino(inode);
 684         u64 search_start = start;
 685         u64 disk_bytenr = 0;
 686         u64 num_bytes = 0;
 687         u64 extent_offset = 0;
 688         u64 extent_end = 0;
 689         int del_nr = 0;
 690         int del_slot = 0;
 691         int extent_type;
 692         int recow;
 693         int ret;
 694         int modify_tree = -1;
 695         int update_refs = (root->ref_cows || root == root->fs_info->tree_root);
 696         int found = 0;
 697
 698         if (drop_cache)
 699                 btrfs_drop_extent_cache(inode, start, end - 1, 0);
 700
 701         if (start >= BTRFS_I(inode)->disk_i_size)
 702                 modify_tree = 0;
 703
 704         while (1) {
 705                 recow = 0;
 706                 ret = btrfs_lookup_file_extent(trans, root, path, ino,
 707                                                search_start, modify_tree);
 708                 if (ret < 0)
 709                         break;
 710                 if (ret > 0 && path->slots[0] > 0 && search_start == start) {
 711                         leaf = path->nodes[0];
 712                         btrfs_item_key_to_cpu(leaf, &key, path->slots[0] - 1);
 713                         if (key.objectid == ino &&
 714                             key.type == BTRFS_EXTENT_DATA_KEY)
 715                                 path->slots[0]--;
 716                 }
 717                 ret = 0;
 718 next_slot:
 719                 leaf = path->nodes[0];
 720                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
 721                         BUG_ON(del_nr > 0);
 722                         ret = btrfs_next_leaf(root, path);
 723                         if (ret < 0)
 724                                 break;
 725                         if (ret > 0) {
 726                                 ret = 0;
 727                                 break;
 728                         }
 729                         leaf = path->nodes[0];
 730                         recow = 1;
 731                 }
 732
 733                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 734                 if (key.objectid > ino ||
 735                     key.type > BTRFS_EXTENT_DATA_KEY || key.offset >= end)
 736                         break;
 737
 738                 fi = btrfs_item_ptr(leaf, path->slots[0],
 739                                     struct btrfs_file_extent_item);
 740                 extent_type = btrfs_file_extent_type(leaf, fi);
 741
 742                 if (extent_type == BTRFS_FILE_EXTENT_REG ||
 743                     extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
 744                         disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
 745                         num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
 746                         extent_offset = btrfs_file_extent_offset(leaf, fi);
 747                         extent_end = key.offset +
 748                                 btrfs_file_extent_num_bytes(leaf, fi);
 749                 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
 750                         extent_end = key.offset +
 751                                 btrfs_file_extent_inline_len(leaf, fi);
 752                 } else {
 753                         WARN_ON(1);
 754                         extent_end = search_start;
 755                 }
 756
 757                 if (extent_end <= search_start) {
 758                         path->slots[0]++;
 759                         goto next_slot;
 760                 }
 761
 762                 found = 1;
 763                 search_start = max(key.offset, start);
 764                 if (recow || !modify_tree) {
 765                         modify_tree = -1;
 766                         btrfs_release_path(path);
 767                         continue;
 768                 }
 769
 770                 /*
 771                  *     | - range to drop - |
 772                  *  | -------- extent -------- |
 773                  */
 774                 if (start > key.offset && end < extent_end) {
 775                         BUG_ON(del_nr > 0);
 776                         BUG_ON(extent_type == BTRFS_FILE_EXTENT_INLINE);
 777
 778                         memcpy(&new_key, &key, sizeof(new_key));
 779                         new_key.offset = start;
 780                         ret = btrfs_duplicate_item(trans, root, path,
 781                                                    &new_key);
 782                         if (ret == -EAGAIN) {
 783                                 btrfs_release_path(path);
 784                                 continue;
 785                         }
 786                         if (ret < 0)
 787                                 break;
 788
 789                         leaf = path->nodes[0];
 790                         fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
 791                                             struct btrfs_file_extent_item);
 792                         btrfs_set_file_extent_num_bytes(leaf, fi,
 793                                                         start - key.offset);
 794
 795                         fi = btrfs_item_ptr(leaf, path->slots[0],
 796                                             struct btrfs_file_extent_item);
 797
 798                         extent_offset += start - key.offset;
 799                         btrfs_set_file_extent_offset(leaf, fi, extent_offset);
 800                         btrfs_set_file_extent_num_bytes(leaf, fi,
 801                                                         extent_end - start);
 802                         btrfs_mark_buffer_dirty(leaf);
 803
 804                         if (update_refs && disk_bytenr > 0) {
 805                                 ret = btrfs_inc_extent_ref(trans, root,
 806                                                 disk_bytenr, num_bytes, 0,
 807                                                 root->root_key.objectid,
 808                                                 new_key.objectid,
 809                                                 start - extent_offset, 0);
 810                                 BUG_ON(ret); /* -ENOMEM */
 811                         }
 812                         key.offset = start;
 813                 }
 814                 /*
 815                  *  | ---- range to drop ----- |
 816                  *      | -------- extent -------- |
 817                  */
 818                 if (start <= key.offset && end < extent_end) {
 819                         BUG_ON(extent_type == BTRFS_FILE_EXTENT_INLINE);
 820
 821                         memcpy(&new_key, &key, sizeof(new_key));
 822                         new_key.offset = end;
 823                         btrfs_set_item_key_safe(trans, root, path, &new_key);
 824
 825                         extent_offset += end - key.offset;
 826                         btrfs_set_file_extent_offset(leaf, fi, extent_offset);
 827                         btrfs_set_file_extent_num_bytes(leaf, fi,
 828                                                         extent_end - end);
 829                         btrfs_mark_buffer_dirty(leaf);
 830                         if (update_refs && disk_bytenr > 0)
 831                                 inode_sub_bytes(inode, end - key.offset);
 832                         break;
 833                 }
 834
 835                 search_start = extent_end;
 836                 /*
 837                  *       | ---- range to drop ----- |
 838                  *  | -------- extent -------- |
 839                  */
 840                 if (start > key.offset && end >= extent_end) {
 841                         BUG_ON(del_nr > 0);
 842                         BUG_ON(extent_type == BTRFS_FILE_EXTENT_INLINE);
 843
 844                         btrfs_set_file_extent_num_bytes(leaf, fi,
 845                                                         start - key.offset);
 846                         btrfs_mark_buffer_dirty(leaf);
 847                         if (update_refs && disk_bytenr > 0)
 848                                 inode_sub_bytes(inode, extent_end - start);
 849                         if (end == extent_end)
 850                                 break;
 851
 852                         path->slots[0]++;
 853                         goto next_slot;
 854                 }
 855
 856                 /*
 857                  *  | ---- range to drop ----- |
 858                  *    | ------ extent ------ |
 859                  */
 860                 if (start <= key.offset && end >= extent_end) {
 861                         if (del_nr == 0) {
 862                                 del_slot = path->slots[0];
 863                                 del_nr = 1;
 864                         } else {
 865                                 BUG_ON(del_slot + del_nr != path->slots[0]);
 866                                 del_nr++;
 867                         }
 868
 869                         if (update_refs &&
 870                             extent_type == BTRFS_FILE_EXTENT_INLINE) {
 871                                 inode_sub_bytes(inode,
 872                                                 extent_end - key.offset);
 873                                 extent_end = ALIGN(extent_end,
 874                                                    root->sectorsize);
 875                         } else if (update_refs && disk_bytenr > 0) {
 876                                 ret = btrfs_free_extent(trans, root,
 877                                                 disk_bytenr, num_bytes, 0,
 878                                                 root->root_key.objectid,
 879                                                 key.objectid, key.offset -
 880                                                 extent_offset, 0);
 881                                 BUG_ON(ret); /* -ENOMEM */
 882                                 inode_sub_bytes(inode,
 883                                                 extent_end - key.offset);
 884                         }
 885
 886                         if (end == extent_end)
 887                                 break;
 888
 889                         if (path->slots[0] + 1 < btrfs_header_nritems(leaf)) {
 890                                 path->slots[0]++;
 891                                 goto next_slot;
 892                         }
 893
 894                         ret = btrfs_del_items(trans, root, path, del_slot,
 895                                               del_nr);
 896                         if (ret) {
 897                                 btrfs_abort_transaction(trans, root, ret);
 898                                 break;
 899                         }
 900
 901                         del_nr = 0;
 902                         del_slot = 0;
 903
 904                         btrfs_release_path(path);
 905                         continue;
 906                 }
 907
 908                 BUG_ON(1);
 909         }
 910
 911         if (!ret && del_nr > 0) {
 912                 ret = btrfs_del_items(trans, root, path, del_slot, del_nr);
 913                 if (ret)
 914                         btrfs_abort_transaction(trans, root, ret);
 915         }
 916
 917         if (drop_end)
 918                 *drop_end = found ? min(end, extent_end) : end;
 919         btrfs_release_path(path);
 920         return ret;
 921 }
 922
 923 int btrfs_drop_extents(struct btrfs_trans_handle *trans,
 924                        struct btrfs_root *root, struct inode *inode, u64 start,
 925                        u64 end, int drop_cache)
 926 {
 927         struct btrfs_path *path;
 928         int ret;
 929
 930         path = btrfs_alloc_path();
 931         if (!path)
 932                 return -ENOMEM;
 933         ret = __btrfs_drop_extents(trans, root, inode, path, start, end, NULL,
 934                                    drop_cache);
 935         btrfs_free_path(path);
 936         return ret;
 937 }
 938
 939 static int extent_mergeable(struct extent_buffer *leaf, int slot,
 940                             u64 objectid, u64 bytenr, u64 orig_offset,
 941                             u64 *start, u64 *end)
 942 {
 943         struct btrfs_file_extent_item *fi;
 944         struct btrfs_key key;
 945         u64 extent_end;
 946
 947         if (slot < 0 || slot >= btrfs_header_nritems(leaf))
 948                 return 0;
 949
 950         btrfs_item_key_to_cpu(leaf, &key, slot);
 951         if (key.objectid != objectid || key.type != BTRFS_EXTENT_DATA_KEY)
 952                 return 0;
 953
 954         fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
 955         if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG ||
 956             btrfs_file_extent_disk_bytenr(leaf, fi) != bytenr ||
 957             btrfs_file_extent_offset(leaf, fi) != key.offset - orig_offset ||
 958             btrfs_file_extent_compression(leaf, fi) ||
 959             btrfs_file_extent_encryption(leaf, fi) ||
 960             btrfs_file_extent_other_encoding(leaf, fi))
 961                 return 0;
 962
 963         extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
 964         if ((*start && *start != key.offset) || (*end && *end != extent_end))
 965                 return 0;
 966
 967         *start = key.offset;
 968         *end = extent_end;
 969         return 1;
 970 }
 971
 972 /*
 973  * Mark extent in the range start - end as written.
 974  *
 975  * This changes extent type from 'pre-allocated' to 'regular'. If only
 976  * part of extent is marked as written, the extent will be split into
 977  * two or three.
 978  */
 979 int btrfs_mark_extent_written(struct btrfs_trans_handle *trans,
 980                               struct inode *inode, u64 start, u64 end)
 981 {
 982         struct btrfs_root *root = BTRFS_I(inode)->root;
 983         struct extent_buffer *leaf;
 984         struct btrfs_path *path;
 985         struct btrfs_file_extent_item *fi;
 986         struct btrfs_key key;
 987         struct btrfs_key new_key;
 988         u64 bytenr;
 989         u64 num_bytes;
 990         u64 extent_end;
 991         u64 orig_offset;
 992         u64 other_start;
 993         u64 other_end;
 994         u64 split;
 995         int del_nr = 0;
 996         int del_slot = 0;
 997         int recow;
 998         int ret;
 999         u64 ino = btrfs_ino(inode);
1000
1001         path = btrfs_alloc_path();
1002         if (!path)
1003                 return -ENOMEM;
1004 again:
1005         recow = 0;
1006         split = start;
1007         key.objectid = ino;
1008         key.type = BTRFS_EXTENT_DATA_KEY;
1009         key.offset = split;
1010
1011         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1012         if (ret < 0)
1013                 goto out;
1014         if (ret > 0 && path->slots[0] > 0)
1015                 path->slots[0]--;
1016
1017         leaf = path->nodes[0];
1018         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1019         BUG_ON(key.objectid != ino || key.type != BTRFS_EXTENT_DATA_KEY);
1020         fi = btrfs_item_ptr(leaf, path->slots[0],
1021                             struct btrfs_file_extent_item);
1022         BUG_ON(btrfs_file_extent_type(leaf, fi) !=
1023                BTRFS_FILE_EXTENT_PREALLOC);
1024         extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
1025         BUG_ON(key.offset > start || extent_end < end);
1026
1027         bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1028         num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
1029         orig_offset = key.offset - btrfs_file_extent_offset(leaf, fi);
1030         memcpy(&new_key, &key, sizeof(new_key));
1031
1032         if (start == key.offset && end < extent_end) {
1033                 other_start = 0;
1034                 other_end = start;
1035                 if (extent_mergeable(leaf, path->slots[0] - 1,
1036                                      ino, bytenr, orig_offset,
1037                                      &other_start, &other_end)) {
1038                         new_key.offset = end;
1039                         btrfs_set_item_key_safe(trans, root, path, &new_key);
1040                         fi = btrfs_item_ptr(leaf, path->slots[0],
1041                                             struct btrfs_file_extent_item);
1042                         btrfs_set_file_extent_generation(leaf, fi,
1043                                                          trans->transid);
1044                         btrfs_set_file_extent_num_bytes(leaf, fi,
1045                                                         extent_end - end);
1046                         btrfs_set_file_extent_offset(leaf, fi,
1047                                                      end - orig_offset);
1048                         fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
1049                                             struct btrfs_file_extent_item);
1050                         btrfs_set_file_extent_generation(leaf, fi,
1051                                                          trans->transid);
1052                         btrfs_set_file_extent_num_bytes(leaf, fi,
1053                                                         end - other_start);
1054                         btrfs_mark_buffer_dirty(leaf);
1055                         goto out;
1056                 }
1057         }
1058
1059         if (start > key.offset && end == extent_end) {
1060                 other_start = end;
1061                 other_end = 0;
1062                 if (extent_mergeable(leaf, path->slots[0] + 1,
1063                                      ino, bytenr, orig_offset,
1064                                      &other_start, &other_end)) {
1065                         fi = btrfs_item_ptr(leaf, path->slots[0],
1066                                             struct btrfs_file_extent_item);
1067                         btrfs_set_file_extent_num_bytes(leaf, fi,
1068                                                         start - key.offset);
1069                         btrfs_set_file_extent_generation(leaf, fi,
1070                                                          trans->transid);
1071                         path->slots[0]++;
1072                         new_key.offset = start;
1073                         btrfs_set_item_key_safe(trans, root, path, &new_key);
1074
1075                         fi = btrfs_item_ptr(leaf, path->slots[0],
1076                                             struct btrfs_file_extent_item);
1077                         btrfs_set_file_extent_generation(leaf, fi,
1078                                                          trans->transid);
1079                         btrfs_set_file_extent_num_bytes(leaf, fi,
1080                                                         other_end - start);
1081                         btrfs_set_file_extent_offset(leaf, fi,
1082                                                      start - orig_offset);
1083                         btrfs_mark_buffer_dirty(leaf);
1084                         goto out;
1085                 }
1086         }
1087
1088         while (start > key.offset || end < extent_end) {
1089                 if (key.offset == start)
1090                         split = end;
1091
1092                 new_key.offset = split;
1093                 ret = btrfs_duplicate_item(trans, root, path, &new_key);
1094                 if (ret == -EAGAIN) {
1095                         btrfs_release_path(path);
1096                         goto again;
1097                 }
1098                 if (ret < 0) {
1099                         btrfs_abort_transaction(trans, root, ret);
1100                         goto out;
1101                 }
1102
1103                 leaf = path->nodes[0];
1104                 fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
1105                                     struct btrfs_file_extent_item);
1106                 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1107                 btrfs_set_file_extent_num_bytes(leaf, fi,
1108                                                 split - key.offset);
1109
1110                 fi = btrfs_item_ptr(leaf, path->slots[0],
1111                                     struct btrfs_file_extent_item);
1112
1113                 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1114                 btrfs_set_file_extent_offset(leaf, fi, split - orig_offset);
1115                 btrfs_set_file_extent_num_bytes(leaf, fi,
1116                                                 extent_end - split);
1117                 btrfs_mark_buffer_dirty(leaf);
1118
1119                 ret = btrfs_inc_extent_ref(trans, root, bytenr, num_bytes, 0,
1120                                            root->root_key.objectid,
1121                                            ino, orig_offset, 0);
1122                 BUG_ON(ret); /* -ENOMEM */
1123
1124                 if (split == start) {
1125                         key.offset = start;
1126                 } else {
1127                         BUG_ON(start != key.offset);
1128                         path->slots[0]--;
1129                         extent_end = end;
1130                 }
1131                 recow = 1;
1132         }
1133
1134         other_start = end;
1135         other_end = 0;
1136         if (extent_mergeable(leaf, path->slots[0] + 1,
1137                              ino, bytenr, orig_offset,
1138                              &other_start, &other_end)) {
1139                 if (recow) {
1140                         btrfs_release_path(path);
1141                         goto again;
1142                 }
1143                 extent_end = other_end;
1144                 del_slot = path->slots[0] + 1;
1145                 del_nr++;
1146                 ret = btrfs_free_extent(trans, root, bytenr, num_bytes,
1147                                         0, root->root_key.objectid,
1148                                         ino, orig_offset, 0);
1149                 BUG_ON(ret); /* -ENOMEM */
1150         }
1151         other_start = 0;
1152         other_end = start;
1153         if (extent_mergeable(leaf, path->slots[0] - 1,
1154                              ino, bytenr, orig_offset,
1155                              &other_start, &other_end)) {
1156                 if (recow) {
1157                         btrfs_release_path(path);
1158                         goto again;
1159                 }
1160                 key.offset = other_start;
1161                 del_slot = path->slots[0];
1162                 del_nr++;
1163                 ret = btrfs_free_extent(trans, root, bytenr, num_bytes,
1164                                         0, root->root_key.objectid,
1165                                         ino, orig_offset, 0);
1166                 BUG_ON(ret); /* -ENOMEM */
1167         }
1168         if (del_nr == 0) {
1169                 fi = btrfs_item_ptr(leaf, path->slots[0],
1170                            struct btrfs_file_extent_item);
1171                 btrfs_set_file_extent_type(leaf, fi,
1172                                            BTRFS_FILE_EXTENT_REG);
1173                 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1174                 btrfs_mark_buffer_dirty(leaf);
1175         } else {
1176                 fi = btrfs_item_ptr(leaf, del_slot - 1,
1177                            struct btrfs_file_extent_item);
1178                 btrfs_set_file_extent_type(leaf, fi,
1179                                            BTRFS_FILE_EXTENT_REG);
1180                 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1181                 btrfs_set_file_extent_num_bytes(leaf, fi,
1182                                                 extent_end - key.offset);
1183                 btrfs_mark_buffer_dirty(leaf);
1184
1185                 ret = btrfs_del_items(trans, root, path, del_slot, del_nr);
1186                 if (ret < 0) {
1187                         btrfs_abort_transaction(trans, root, ret);
1188                         goto out;
1189                 }
1190         }
1191 out:
1192         btrfs_free_path(path);
1193         return 0;
1194 }
1195
1196 /*
1197  * on error we return an unlocked page and the error value
1198  * on success we return a locked page and 0
1199  */
1200 static int prepare_uptodate_page(struct page *page, u64 pos,
1201                                  bool force_uptodate)
1202 {
1203         int ret = 0;
1204
1205         if (((pos & (PAGE_CACHE_SIZE - 1)) || force_uptodate) &&
1206             !PageUptodate(page)) {
1207                 ret = btrfs_readpage(NULL, page);
1208                 if (ret)
1209                         return ret;
1210                 lock_page(page);
1211                 if (!PageUptodate(page)) {
1212                         unlock_page(page);
1213                         return -EIO;
1214                 }
1215         }
1216         return 0;
1217 }
1218
1219 /*
1220  * this gets pages into the page cache and locks them down, it also properly
1221  * waits for data=ordered extents to finish before allowing the pages to be
1222  * modified.
1223  */
1224 static noinline int prepare_pages(struct btrfs_root *root, struct file *file,
1225                          struct page **pages, size_t num_pages,
1226                          loff_t pos, unsigned long first_index,
1227                          size_t write_bytes, bool force_uptodate)
1228 {
1229         struct extent_state *cached_state = NULL;
1230         int i;
1231         unsigned long index = pos >> PAGE_CACHE_SHIFT;
1232         struct inode *inode = file_inode(file);
1233         gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping);
1234         int err = 0;
1235         int faili = 0;
1236         u64 start_pos;
1237         u64 last_pos;
1238
1239         start_pos = pos & ~((u64)root->sectorsize - 1);
1240         last_pos = ((u64)index + num_pages) << PAGE_CACHE_SHIFT;
1241
1242 again:
1243         for (i = 0; i < num_pages; i++) {
1244                 pages[i] = find_or_create_page(inode->i_mapping, index + i,
1245                                                mask | __GFP_WRITE);
1246                 if (!pages[i]) {
1247                         faili = i - 1;
1248                         err = -ENOMEM;
1249                         goto fail;
1250                 }
1251
1252                 if (i == 0)
1253                         err = prepare_uptodate_page(pages[i], pos,
1254                                                     force_uptodate);
1255                 if (i == num_pages - 1)
1256                         err = prepare_uptodate_page(pages[i],
1257                                                     pos + write_bytes, false);
1258                 if (err) {
1259                         page_cache_release(pages[i]);
1260                         faili = i - 1;
1261                         goto fail;
1262                 }
1263                 wait_on_page_writeback(pages[i]);
1264         }
1265         err = 0;
1266         if (start_pos < inode->i_size) {
1267                 struct btrfs_ordered_extent *ordered;
1268                 lock_extent_bits(&BTRFS_I(inode)->io_tree,
1269                                  start_pos, last_pos - 1, 0, &cached_state);
1270                 ordered = btrfs_lookup_first_ordered_extent(inode,
1271                                                             last_pos - 1);
1272                 if (ordered &&
1273                     ordered->file_offset + ordered->len > start_pos &&
1274                     ordered->file_offset < last_pos) {
1275                         btrfs_put_ordered_extent(ordered);
1276                         unlock_extent_cached(&BTRFS_I(inode)->io_tree,
1277                                              start_pos, last_pos - 1,
1278                                              &cached_state, GFP_NOFS);
1279                         for (i = 0; i < num_pages; i++) {
1280                                 unlock_page(pages[i]);
1281                                 page_cache_release(pages[i]);
1282                         }
1283                         btrfs_wait_ordered_range(inode, start_pos,
1284                                                  last_pos - start_pos);
1285                         goto again;
1286                 }
1287                 if (ordered)
1288                         btrfs_put_ordered_extent(ordered);
1289
1290                 clear_extent_bit(&BTRFS_I(inode)->io_tree, start_pos,
1291                                   last_pos - 1, EXTENT_DIRTY | EXTENT_DELALLOC |
1292                                   EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG,
1293                                   0, 0, &cached_state, GFP_NOFS);
1294                 unlock_extent_cached(&BTRFS_I(inode)->io_tree,
1295                                      start_pos, last_pos - 1, &cached_state,
1296                                      GFP_NOFS);
1297         }
1298         for (i = 0; i < num_pages; i++) {
1299                 if (clear_page_dirty_for_io(pages[i]))
1300                         account_page_redirty(pages[i]);
1301                 set_page_extent_mapped(pages[i]);
1302                 WARN_ON(!PageLocked(pages[i]));
1303         }
1304         return 0;
1305 fail:
1306         while (faili >= 0) {
1307                 unlock_page(pages[faili]);
1308                 page_cache_release(pages[faili]);
1309                 faili--;
1310         }
1311         return err;
1312
1313 }
1314
1315 static noinline ssize_t __btrfs_buffered_write(struct file *file,
1316                                                struct iov_iter *i,
1317                                                loff_t pos)
1318 {
1319         struct inode *inode = file_inode(file);
1320         struct btrfs_root *root = BTRFS_I(inode)->root;
1321         struct page **pages = NULL;
1322         unsigned long first_index;
1323         size_t num_written = 0;
1324         int nrptrs;
1325         int ret = 0;
1326         bool force_page_uptodate = false;
1327
1328         nrptrs = min((iov_iter_count(i) + PAGE_CACHE_SIZE - 1) /
1329                      PAGE_CACHE_SIZE, PAGE_CACHE_SIZE /
1330                      (sizeof(struct page *)));
1331         nrptrs = min(nrptrs, current->nr_dirtied_pause - current->nr_dirtied);
1332         nrptrs = max(nrptrs, 8);
1333         pages = kmalloc(nrptrs * sizeof(struct page *), GFP_KERNEL);
1334         if (!pages)
1335                 return -ENOMEM;
1336
1337         first_index = pos >> PAGE_CACHE_SHIFT;
1338
1339         while (iov_iter_count(i) > 0) {
1340                 size_t offset = pos & (PAGE_CACHE_SIZE - 1);
1341                 size_t write_bytes = min(iov_iter_count(i),
1342                                          nrptrs * (size_t)PAGE_CACHE_SIZE -
1343                                          offset);
1344                 size_t num_pages = (write_bytes + offset +
1345                                     PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1346                 size_t dirty_pages;
1347                 size_t copied;
1348
1349                 WARN_ON(num_pages > nrptrs);
1350
1351                 /*
1352                  * Fault pages before locking them in prepare_pages
1353                  * to avoid recursive lock
1354                  */
1355                 if (unlikely(iov_iter_fault_in_readable(i, write_bytes))) {
1356                         ret = -EFAULT;
1357                         break;
1358                 }
1359
1360                 ret = btrfs_delalloc_reserve_space(inode,
1361                                         num_pages << PAGE_CACHE_SHIFT);
1362                 if (ret)
1363                         break;
1364
1365                 /*
1366                  * This is going to setup the pages array with the number of
1367                  * pages we want, so we don't really need to worry about the
1368                  * contents of pages from loop to loop
1369                  */
1370                 ret = prepare_pages(root, file, pages, num_pages,
1371                                     pos, first_index, write_bytes,
1372                                     force_page_uptodate);
1373                 if (ret) {
1374                         btrfs_delalloc_release_space(inode,
1375                                         num_pages << PAGE_CACHE_SHIFT);
1376                         break;
1377                 }
1378
1379                 copied = btrfs_copy_from_user(pos, num_pages,
1380                                            write_bytes, pages, i);
1381
1382                 /*
1383                  * if we have trouble faulting in the pages, fall
1384                  * back to one page at a time
1385                  */
1386                 if (copied < write_bytes)
1387                         nrptrs = 1;
1388
1389                 if (copied == 0) {
1390                         force_page_uptodate = true;
1391                         dirty_pages = 0;
1392                 } else {
1393                         force_page_uptodate = false;
1394                         dirty_pages = (copied + offset +
1395                                        PAGE_CACHE_SIZE - 1) >>
1396                                        PAGE_CACHE_SHIFT;
1397                 }
1398
1399                 /*
1400                  * If we had a short copy we need to release the excess delaloc
1401                  * bytes we reserved.  We need to increment outstanding_extents
1402                  * because btrfs_delalloc_release_space will decrement it, but
1403                  * we still have an outstanding extent for the chunk we actually
1404                  * managed to copy.
1405                  */
1406                 if (num_pages > dirty_pages) {
1407                         if (copied > 0) {
1408                                 spin_lock(&BTRFS_I(inode)->lock);
1409                                 BTRFS_I(inode)->outstanding_extents++;
1410                                 spin_unlock(&BTRFS_I(inode)->lock);
1411                         }
1412                         btrfs_delalloc_release_space(inode,
1413                                         (num_pages - dirty_pages) <<
1414                                         PAGE_CACHE_SHIFT);
1415                 }
1416
1417                 if (copied > 0) {
1418                         ret = btrfs_dirty_pages(root, inode, pages,
1419                                                 dirty_pages, pos, copied,
1420                                                 NULL);
1421                         if (ret) {
1422                                 btrfs_delalloc_release_space(inode,
1423                                         dirty_pages << PAGE_CACHE_SHIFT);
1424                                 btrfs_drop_pages(pages, num_pages);
1425                                 break;
1426                         }
1427                 }
1428
1429                 btrfs_drop_pages(pages, num_pages);
1430
1431                 cond_resched();
1432
1433                 balance_dirty_pages_ratelimited(inode->i_mapping);
1434                 if (dirty_pages < (root->leafsize >> PAGE_CACHE_SHIFT) + 1)
1435                         btrfs_btree_balance_dirty(root);
1436
1437                 pos += copied;
1438                 num_written += copied;
1439         }
1440
1441         kfree(pages);
1442
1443         return num_written ? num_written : ret;
1444 }
1445
1446 static ssize_t __btrfs_direct_write(struct kiocb *iocb,
1447                                     const struct iovec *iov,
1448                                     unsigned long nr_segs, loff_t pos,
1449                                     loff_t *ppos, size_t count, size_t ocount)
1450 {
1451         struct file *file = iocb->ki_filp;
1452         struct iov_iter i;
1453         ssize_t written;
1454         ssize_t written_buffered;
1455         loff_t endbyte;
1456         int err;
1457
1458         written = generic_file_direct_write(iocb, iov, &nr_segs, pos, ppos,
1459                                             count, ocount);
1460
1461         if (written < 0 || written == count)
1462                 return written;
1463
1464         pos += written;
1465         count -= written;
1466         iov_iter_init(&i, iov, nr_segs, count, written);
1467         written_buffered = __btrfs_buffered_write(file, &i, pos);
1468         if (written_buffered < 0) {
1469                 err = written_buffered;
1470                 goto out;
1471         }
1472         endbyte = pos + written_buffered - 1;
1473         err = filemap_write_and_wait_range(file->f_mapping, pos, endbyte);
1474         if (err)
1475                 goto out;
1476         written += written_buffered;
1477         *ppos = pos + written_buffered;
1478         invalidate_mapping_pages(file->f_mapping, pos >> PAGE_CACHE_SHIFT,
1479                                  endbyte >> PAGE_CACHE_SHIFT);
1480 out:
1481         return written ? written : err;
1482 }
1483
1484 static void update_time_for_write(struct inode *inode)
1485 {
1486         struct timespec now;
1487
1488         if (IS_NOCMTIME(inode))
1489                 return;
1490
1491         now = current_fs_time(inode->i_sb);
1492         if (!timespec_equal(&inode->i_mtime, &now))
1493                 inode->i_mtime = now;
1494
1495         if (!timespec_equal(&inode->i_ctime, &now))
1496                 inode->i_ctime = now;
1497
1498         if (IS_I_VERSION(inode))
1499                 inode_inc_iversion(inode);
1500 }
1501
1502 static ssize_t btrfs_file_aio_write(struct kiocb *iocb,
1503                                     const struct iovec *iov,
1504                                     unsigned long nr_segs, loff_t pos)
1505 {
1506         struct file *file = iocb->ki_filp;
1507         struct inode *inode = file_inode(file);
1508         struct btrfs_root *root = BTRFS_I(inode)->root;
1509         loff_t *ppos = &iocb->ki_pos;
1510         u64 start_pos;
1511         ssize_t num_written = 0;
1512         ssize_t err = 0;
1513         size_t count, ocount;
1514         bool sync = (file->f_flags & O_DSYNC) || IS_SYNC(file->f_mapping->host);
1515
1516         sb_start_write(inode->i_sb);
1517
1518         mutex_lock(&inode->i_mutex);
1519
1520         err = generic_segment_checks(iov, &nr_segs, &ocount, VERIFY_READ);
1521         if (err) {
1522                 mutex_unlock(&inode->i_mutex);
1523                 goto out;
1524         }
1525         count = ocount;
1526
1527         current->backing_dev_info = inode->i_mapping->backing_dev_info;
1528         err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode));
1529         if (err) {
1530                 mutex_unlock(&inode->i_mutex);
1531                 goto out;
1532         }
1533
1534         if (count == 0) {
1535                 mutex_unlock(&inode->i_mutex);
1536                 goto out;
1537         }
1538
1539         err = file_remove_suid(file);
1540         if (err) {
1541                 mutex_unlock(&inode->i_mutex);
1542                 goto out;
1543         }
1544
1545         /*
1546          * If BTRFS flips readonly due to some impossible error
1547          * (fs_info->fs_state now has BTRFS_SUPER_FLAG_ERROR),
1548          * although we have opened a file as writable, we have
1549          * to stop this write operation to ensure FS consistency.
1550          */
1551         if (test_bit(BTRFS_FS_STATE_ERROR, &root->fs_info->fs_state)) {
1552                 mutex_unlock(&inode->i_mutex);
1553                 err = -EROFS;
1554                 goto out;
1555         }
1556
1557         /*
1558          * We reserve space for updating the inode when we reserve space for the
1559          * extent we are going to write, so we will enospc out there.  We don't
1560          * need to start yet another transaction to update the inode as we will
1561          * update the inode when we finish writing whatever data we write.
1562          */
1563         update_time_for_write(inode);
1564
1565         start_pos = round_down(pos, root->sectorsize);
1566         if (start_pos > i_size_read(inode)) {
1567                 err = btrfs_cont_expand(inode, i_size_read(inode), start_pos);
1568                 if (err) {
1569                         mutex_unlock(&inode->i_mutex);
1570                         goto out;
1571                 }
1572         }
1573
1574         if (sync)
1575                 atomic_inc(&BTRFS_I(inode)->sync_writers);
1576
1577         if (unlikely(file->f_flags & O_DIRECT)) {
1578                 num_written = __btrfs_direct_write(iocb, iov, nr_segs,
1579                                                    pos, ppos, count, ocount);
1580         } else {
1581                 struct iov_iter i;
1582
1583                 iov_iter_init(&i, iov, nr_segs, count, num_written);
1584
1585                 num_written = __btrfs_buffered_write(file, &i, pos);
1586                 if (num_written > 0)
1587                         *ppos = pos + num_written;
1588         }
1589
1590         mutex_unlock(&inode->i_mutex);
1591
1592         /*
1593          * we want to make sure fsync finds this change
1594          * but we haven't joined a transaction running right now.
1595          *
1596          * Later on, someone is sure to update the inode and get the
1597          * real transid recorded.
1598          *
1599          * We set last_trans now to the fs_info generation + 1,
1600          * this will either be one more than the running transaction
1601          * or the generation used for the next transaction if there isn't
1602          * one running right now.
1603          *
1604          * We also have to set last_sub_trans to the current log transid,
1605          * otherwise subsequent syncs to a file that's been synced in this
1606          * transaction will appear to have already occured.
1607          */
1608         BTRFS_I(inode)->last_trans = root->fs_info->generation + 1;
1609         BTRFS_I(inode)->last_sub_trans = root->log_transid;
1610         if (num_written > 0 || num_written == -EIOCBQUEUED) {
1611                 err = generic_write_sync(file, pos, num_written);
1612                 if (err < 0 && num_written > 0)
1613                         num_written = err;
1614         }
1615
1616         if (sync)
1617                 atomic_dec(&BTRFS_I(inode)->sync_writers);
1618 out:
1619         sb_end_write(inode->i_sb);
1620         current->backing_dev_info = NULL;
1621         return num_written ? num_written : err;
1622 }
1623
1624 int btrfs_release_file(struct inode *inode, struct file *filp)
1625 {
1626         /*
1627          * ordered_data_close is set by settattr when we are about to truncate
1628          * a file from a non-zero size to a zero size.  This tries to
1629          * flush down new bytes that may have been written if the
1630          * application were using truncate to replace a file in place.
1631          */
1632         if (test_and_clear_bit(BTRFS_INODE_ORDERED_DATA_CLOSE,
1633                                &BTRFS_I(inode)->runtime_flags)) {
1634                 struct btrfs_trans_handle *trans;
1635                 struct btrfs_root *root = BTRFS_I(inode)->root;
1636
1637                 /*
1638                  * We need to block on a committing transaction to keep us from
1639                  * throwing a ordered operation on to the list and causing
1640                  * something like sync to deadlock trying to flush out this
1641                  * inode.
1642                  */
1643                 trans = btrfs_start_transaction(root, 0);
1644                 if (IS_ERR(trans))
1645                         return PTR_ERR(trans);
1646                 btrfs_add_ordered_operation(trans, BTRFS_I(inode)->root, inode);
1647                 btrfs_end_transaction(trans, root);
1648                 if (inode->i_size > BTRFS_ORDERED_OPERATIONS_FLUSH_LIMIT)
1649                         filemap_flush(inode->i_mapping);
1650         }
1651         if (filp->private_data)
1652                 btrfs_ioctl_trans_end(filp);
1653         return 0;
1654 }
1655
1656 /*
1657  * fsync call for both files and directories.  This logs the inode into
1658  * the tree log instead of forcing full commits whenever possible.
1659  *
1660  * It needs to call filemap_fdatawait so that all ordered extent updates are
1661  * in the metadata btree are up to date for copying to the log.
1662  *
1663  * It drops the inode mutex before doing the tree log commit.  This is an
1664  * important optimization for directories because holding the mutex prevents
1665  * new operations on the dir while we write to disk.
1666  */
1667 int btrfs_sync_file(struct file *file, loff_t start, loff_t end, int datasync)
1668 {
1669         struct dentry *dentry = file->f_path.dentry;
1670         struct inode *inode = dentry->d_inode;
1671         struct btrfs_root *root = BTRFS_I(inode)->root;
1672         int ret = 0;
1673         struct btrfs_trans_handle *trans;
1674         bool full_sync = 0;
1675
1676         trace_btrfs_sync_file(file, datasync);
1677
1678         /*
1679          * We write the dirty pages in the range and wait until they complete
1680          * out of the ->i_mutex. If so, we can flush the dirty pages by
1681          * multi-task, and make the performance up.  See
1682          * btrfs_wait_ordered_range for an explanation of the ASYNC check.
1683          */
1684         atomic_inc(&BTRFS_I(inode)->sync_writers);
1685         ret = filemap_fdatawrite_range(inode->i_mapping, start, end);
1686         if (!ret && test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
1687                              &BTRFS_I(inode)->runtime_flags))
1688                 ret = filemap_fdatawrite_range(inode->i_mapping, start, end);
1689         atomic_dec(&BTRFS_I(inode)->sync_writers);
1690         if (ret)
1691                 return ret;
1692
1693         mutex_lock(&inode->i_mutex);
1694
1695         /*
1696          * We flush the dirty pages again to avoid some dirty pages in the
1697          * range being left.
1698          */
1699         atomic_inc(&root->log_batch);
1700         full_sync = test_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
1701                              &BTRFS_I(inode)->runtime_flags);
1702         if (full_sync)
1703                 btrfs_wait_ordered_range(inode, start, end - start + 1);
1704         atomic_inc(&root->log_batch);
1705
1706         /*
1707          * check the transaction that last modified this inode
1708          * and see if its already been committed
1709          */
1710         if (!BTRFS_I(inode)->last_trans) {
1711                 mutex_unlock(&inode->i_mutex);
1712                 goto out;
1713         }
1714
1715         /*
1716          * if the last transaction that changed this file was before
1717          * the current transaction, we can bail out now without any
1718          * syncing
1719          */
1720         smp_mb();
1721         if (btrfs_inode_in_log(inode, root->fs_info->generation) ||
1722             BTRFS_I(inode)->last_trans <=
1723             root->fs_info->last_trans_committed) {
1724                 BTRFS_I(inode)->last_trans = 0;
1725
1726                 /*
1727                  * We'v had everything committed since the last time we were
1728                  * modified so clear this flag in case it was set for whatever
1729                  * reason, it's no longer relevant.
1730                  */
1731                 clear_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
1732                           &BTRFS_I(inode)->runtime_flags);
1733                 mutex_unlock(&inode->i_mutex);
1734                 goto out;
1735         }
1736
1737         /*
1738          * ok we haven't committed the transaction yet, lets do a commit
1739          */
1740         if (file->private_data)
1741                 btrfs_ioctl_trans_end(file);
1742
1743         trans = btrfs_start_transaction(root, 0);
1744         if (IS_ERR(trans)) {
1745                 ret = PTR_ERR(trans);
1746                 mutex_unlock(&inode->i_mutex);
1747                 goto out;
1748         }
1749
1750         ret = btrfs_log_dentry_safe(trans, root, dentry);
1751         if (ret < 0) {
1752                 mutex_unlock(&inode->i_mutex);
1753                 goto out;
1754         }
1755
1756         /* we've logged all the items and now have a consistent
1757          * version of the file in the log.  It is possible that
1758          * someone will come in and modify the file, but that's
1759          * fine because the log is consistent on disk, and we
1760          * have references to all of the file's extents
1761          *
1762          * It is possible that someone will come in and log the
1763          * file again, but that will end up using the synchronization
1764          * inside btrfs_sync_log to keep things safe.
1765          */
1766         mutex_unlock(&inode->i_mutex);
1767
1768         if (ret != BTRFS_NO_LOG_SYNC) {
1769                 if (ret > 0) {
1770                         /*
1771                          * If we didn't already wait for ordered extents we need
1772                          * to do that now.
1773                          */
1774                         if (!full_sync)
1775                                 btrfs_wait_ordered_range(inode, start,
1776                                                          end - start + 1);
1777                         ret = btrfs_commit_transaction(trans, root);
1778                 } else {
1779                         ret = btrfs_sync_log(trans, root);
1780                         if (ret == 0) {
1781                                 ret = btrfs_end_transaction(trans, root);
1782                         } else {
1783                                 if (!full_sync)
1784                                         btrfs_wait_ordered_range(inode, start,
1785                                                                  end -
1786                                                                  start + 1);
1787                                 ret = btrfs_commit_transaction(trans, root);
1788                         }
1789                 }
1790         } else {
1791                 ret = btrfs_end_transaction(trans, root);
1792         }
1793 out:
1794         return ret > 0 ? -EIO : ret;
1795 }
1796
1797 static const struct vm_operations_struct btrfs_file_vm_ops = {
1798         .fault          = filemap_fault,
1799         .page_mkwrite   = btrfs_page_mkwrite,
1800         .remap_pages    = generic_file_remap_pages,
1801 };
1802
1803 static int btrfs_file_mmap(struct file  *filp, struct vm_area_struct *vma)
1804 {
1805         struct address_space *mapping = filp->f_mapping;
1806
1807         if (!mapping->a_ops->readpage)
1808                 return -ENOEXEC;
1809
1810         file_accessed(filp);
1811         vma->vm_ops = &btrfs_file_vm_ops;
1812
1813         return 0;
1814 }
1815
1816 static int hole_mergeable(struct inode *inode, struct extent_buffer *leaf,
1817                           int slot, u64 start, u64 end)
1818 {
1819         struct btrfs_file_extent_item *fi;
1820         struct btrfs_key key;
1821
1822         if (slot < 0 || slot >= btrfs_header_nritems(leaf))
1823                 return 0;
1824
1825         btrfs_item_key_to_cpu(leaf, &key, slot);
1826         if (key.objectid != btrfs_ino(inode) ||
1827             key.type != BTRFS_EXTENT_DATA_KEY)
1828                 return 0;
1829
1830         fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
1831
1832         if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG)
1833                 return 0;
1834
1835         if (btrfs_file_extent_disk_bytenr(leaf, fi))
1836                 return 0;
1837
1838         if (key.offset == end)
1839                 return 1;
1840         if (key.offset + btrfs_file_extent_num_bytes(leaf, fi) == start)
1841                 return 1;
1842         return 0;
1843 }
1844
1845 static int fill_holes(struct btrfs_trans_handle *trans, struct inode *inode,
1846                       struct btrfs_path *path, u64 offset, u64 end)
1847 {
1848         struct btrfs_root *root = BTRFS_I(inode)->root;
1849         struct extent_buffer *leaf;
1850         struct btrfs_file_extent_item *fi;
1851         struct extent_map *hole_em;
1852         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
1853         struct btrfs_key key;
1854         int ret;
1855
1856         key.objectid = btrfs_ino(inode);
1857         key.type = BTRFS_EXTENT_DATA_KEY;
1858         key.offset = offset;
1859
1860
1861         ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
1862         if (ret < 0)
1863                 return ret;
1864         BUG_ON(!ret);
1865
1866         leaf = path->nodes[0];
1867         if (hole_mergeable(inode, leaf, path->slots[0]-1, offset, end)) {
1868                 u64 num_bytes;
1869
1870                 path->slots[0]--;
1871                 fi = btrfs_item_ptr(leaf, path->slots[0],
1872                                     struct btrfs_file_extent_item);
1873                 num_bytes = btrfs_file_extent_num_bytes(leaf, fi) +
1874                         end - offset;
1875                 btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
1876                 btrfs_set_file_extent_ram_bytes(leaf, fi, num_bytes);
1877                 btrfs_set_file_extent_offset(leaf, fi, 0);
1878                 btrfs_mark_buffer_dirty(leaf);
1879                 goto out;
1880         }
1881
1882         if (hole_mergeable(inode, leaf, path->slots[0]+1, offset, end)) {
1883                 u64 num_bytes;
1884
1885                 path->slots[0]++;
1886                 key.offset = offset;
1887                 btrfs_set_item_key_safe(trans, root, path, &key);
1888                 fi = btrfs_item_ptr(leaf, path->slots[0],
1889                                     struct btrfs_file_extent_item);
1890                 num_bytes = btrfs_file_extent_num_bytes(leaf, fi) + end -
1891                         offset;
1892                 btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
1893                 btrfs_set_file_extent_ram_bytes(leaf, fi, num_bytes);
1894                 btrfs_set_file_extent_offset(leaf, fi, 0);
1895                 btrfs_mark_buffer_dirty(leaf);
1896                 goto out;
1897         }
1898         btrfs_release_path(path);
1899
1900         ret = btrfs_insert_file_extent(trans, root, btrfs_ino(inode), offset,
1901                                        0, 0, end - offset, 0, end - offset,
1902                                        0, 0, 0);
1903         if (ret)
1904                 return ret;
1905
1906 out:
1907         btrfs_release_path(path);
1908
1909         hole_em = alloc_extent_map();
1910         if (!hole_em) {
1911                 btrfs_drop_extent_cache(inode, offset, end - 1, 0);
1912                 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
1913                         &BTRFS_I(inode)->runtime_flags);
1914         } else {
1915                 hole_em->start = offset;
1916                 hole_em->len = end - offset;
1917                 hole_em->orig_start = offset;
1918
1919                 hole_em->block_start = EXTENT_MAP_HOLE;
1920                 hole_em->block_len = 0;
1921                 hole_em->orig_block_len = 0;
1922                 hole_em->bdev = root->fs_info->fs_devices->latest_bdev;
1923                 hole_em->compress_type = BTRFS_COMPRESS_NONE;
1924                 hole_em->generation = trans->transid;
1925
1926                 do {
1927                         btrfs_drop_extent_cache(inode, offset, end - 1, 0);
1928                         write_lock(&em_tree->lock);
1929                         ret = add_extent_mapping(em_tree, hole_em);
1930                         if (!ret)
1931                                 list_move(&hole_em->list,
1932                                           &em_tree->modified_extents);
1933                         write_unlock(&em_tree->lock);
1934                 } while (ret == -EEXIST);
1935                 free_extent_map(hole_em);
1936                 if (ret)
1937                         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
1938                                 &BTRFS_I(inode)->runtime_flags);
1939         }
1940
1941         return 0;
1942 }
1943
1944 static int btrfs_punch_hole(struct inode *inode, loff_t offset, loff_t len)
1945 {
1946         struct btrfs_root *root = BTRFS_I(inode)->root;
1947         struct extent_state *cached_state = NULL;
1948         struct btrfs_path *path;
1949         struct btrfs_block_rsv *rsv;
1950         struct btrfs_trans_handle *trans;
1951         u64 lockstart = round_up(offset, BTRFS_I(inode)->root->sectorsize);
1952         u64 lockend = round_down(offset + len,
1953                                  BTRFS_I(inode)->root->sectorsize) - 1;
1954         u64 cur_offset = lockstart;
1955         u64 min_size = btrfs_calc_trunc_metadata_size(root, 1);
1956         u64 drop_end;
1957         int ret = 0;
1958         int err = 0;
1959         bool same_page = ((offset >> PAGE_CACHE_SHIFT) ==
1960                           ((offset + len - 1) >> PAGE_CACHE_SHIFT));
1961
1962         btrfs_wait_ordered_range(inode, offset, len);
1963
1964         mutex_lock(&inode->i_mutex);
1965         /*
1966          * We needn't truncate any page which is beyond the end of the file
1967          * because we are sure there is no data there.
1968          */
1969         /*
1970          * Only do this if we are in the same page and we aren't doing the
1971          * entire page.
1972          */
1973         if (same_page && len < PAGE_CACHE_SIZE) {
1974                 if (offset < round_up(inode->i_size, PAGE_CACHE_SIZE))
1975                         ret = btrfs_truncate_page(inode, offset, len, 0);
1976                 mutex_unlock(&inode->i_mutex);
1977                 return ret;
1978         }
1979
1980         /* zero back part of the first page */
1981         if (offset < round_up(inode->i_size, PAGE_CACHE_SIZE)) {
1982                 ret = btrfs_truncate_page(inode, offset, 0, 0);
1983                 if (ret) {
1984                         mutex_unlock(&inode->i_mutex);
1985                         return ret;
1986                 }
1987         }
1988
1989         /* zero the front end of the last page */
1990         if (offset + len < round_up(inode->i_size, PAGE_CACHE_SIZE)) {
1991                 ret = btrfs_truncate_page(inode, offset + len, 0, 1);
1992                 if (ret) {
1993                         mutex_unlock(&inode->i_mutex);
1994                         return ret;
1995                 }
1996         }
1997
1998         if (lockend < lockstart) {
1999                 mutex_unlock(&inode->i_mutex);
2000                 return 0;
2001         }
2002
2003         while (1) {
2004                 struct btrfs_ordered_extent *ordered;
2005
2006                 truncate_pagecache_range(inode, lockstart, lockend);
2007
2008                 lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2009                                  0, &cached_state);
2010                 ordered = btrfs_lookup_first_ordered_extent(inode, lockend);
2011
2012                 /*
2013                  * We need to make sure we have no ordered extents in this range
2014                  * and nobody raced in and read a page in this range, if we did
2015                  * we need to try again.
2016                  */
2017                 if ((!ordered ||
2018                     (ordered->file_offset + ordered->len < lockstart ||
2019                      ordered->file_offset > lockend)) &&
2020                      !test_range_bit(&BTRFS_I(inode)->io_tree, lockstart,
2021                                      lockend, EXTENT_UPTODATE, 0,
2022                                      cached_state)) {
2023                         if (ordered)
2024                                 btrfs_put_ordered_extent(ordered);
2025                         break;
2026                 }
2027                 if (ordered)
2028                         btrfs_put_ordered_extent(ordered);
2029                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart,
2030                                      lockend, &cached_state, GFP_NOFS);
2031                 btrfs_wait_ordered_range(inode, lockstart,
2032                                          lockend - lockstart + 1);
2033         }
2034
2035         path = btrfs_alloc_path();
2036         if (!path) {
2037                 ret = -ENOMEM;
2038                 goto out;
2039         }
2040
2041         rsv = btrfs_alloc_block_rsv(root, BTRFS_BLOCK_RSV_TEMP);
2042         if (!rsv) {
2043                 ret = -ENOMEM;
2044                 goto out_free;
2045         }
2046         rsv->size = btrfs_calc_trunc_metadata_size(root, 1);
2047         rsv->failfast = 1;
2048
2049         /*
2050          * 1 - update the inode
2051          * 1 - removing the extents in the range
2052          * 1 - adding the hole extent
2053          */
2054         trans = btrfs_start_transaction(root, 3);
2055         if (IS_ERR(trans)) {
2056                 err = PTR_ERR(trans);
2057                 goto out_free;
2058         }
2059
2060         ret = btrfs_block_rsv_migrate(&root->fs_info->trans_block_rsv, rsv,
2061                                       min_size);
2062         BUG_ON(ret);
2063         trans->block_rsv = rsv;
2064
2065         while (cur_offset < lockend) {
2066                 ret = __btrfs_drop_extents(trans, root, inode, path,
2067                                            cur_offset, lockend + 1,
2068                                            &drop_end, 1);
2069                 if (ret != -ENOSPC)
2070                         break;
2071
2072                 trans->block_rsv = &root->fs_info->trans_block_rsv;
2073
2074                 ret = fill_holes(trans, inode, path, cur_offset, drop_end);
2075                 if (ret) {
2076                         err = ret;
2077                         break;
2078                 }
2079
2080                 cur_offset = drop_end;
2081
2082                 ret = btrfs_update_inode(trans, root, inode);
2083                 if (ret) {
2084                         err = ret;
2085                         break;
2086                 }
2087
2088                 btrfs_end_transaction(trans, root);
2089                 btrfs_btree_balance_dirty(root);
2090
2091                 trans = btrfs_start_transaction(root, 3);
2092                 if (IS_ERR(trans)) {
2093                         ret = PTR_ERR(trans);
2094                         trans = NULL;
2095                         break;
2096                 }
2097
2098                 ret = btrfs_block_rsv_migrate(&root->fs_info->trans_block_rsv,
2099                                               rsv, min_size);
2100                 BUG_ON(ret);    /* shouldn't happen */
2101                 trans->block_rsv = rsv;
2102         }
2103
2104         if (ret) {
2105                 err = ret;
2106                 goto out_trans;
2107         }
2108
2109         trans->block_rsv = &root->fs_info->trans_block_rsv;
2110         ret = fill_holes(trans, inode, path, cur_offset, drop_end);
2111         if (ret) {
2112                 err = ret;
2113                 goto out_trans;
2114         }
2115
2116 out_trans:
2117         if (!trans)
2118                 goto out_free;
2119
2120         inode_inc_iversion(inode);
2121         inode->i_mtime = inode->i_ctime = CURRENT_TIME;
2122
2123         trans->block_rsv = &root->fs_info->trans_block_rsv;
2124         ret = btrfs_update_inode(trans, root, inode);
2125         btrfs_end_transaction(trans, root);
2126         btrfs_btree_balance_dirty(root);
2127 out_free:
2128         btrfs_free_path(path);
2129         btrfs_free_block_rsv(root, rsv);
2130 out:
2131         unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2132                              &cached_state, GFP_NOFS);
2133         mutex_unlock(&inode->i_mutex);
2134         if (ret && !err)
2135                 err = ret;
2136         return err;
2137 }
2138
2139 static long btrfs_fallocate(struct file *file, int mode,
2140                             loff_t offset, loff_t len)
2141 {
2142         struct inode *inode = file_inode(file);
2143         struct extent_state *cached_state = NULL;
2144         u64 cur_offset;
2145         u64 last_byte;
2146         u64 alloc_start;
2147         u64 alloc_end;
2148         u64 alloc_hint = 0;
2149         u64 locked_end;
2150         struct extent_map *em;
2151         int blocksize = BTRFS_I(inode)->root->sectorsize;
2152         int ret;
2153
2154         alloc_start = round_down(offset, blocksize);
2155         alloc_end = round_up(offset + len, blocksize);
2156
2157         /* Make sure we aren't being give some crap mode */
2158         if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
2159                 return -EOPNOTSUPP;
2160
2161         if (mode & FALLOC_FL_PUNCH_HOLE)
2162                 return btrfs_punch_hole(inode, offset, len);
2163
2164         /*
2165          * Make sure we have enough space before we do the
2166          * allocation.
2167          */
2168         ret = btrfs_check_data_free_space(inode, alloc_end - alloc_start);
2169         if (ret)
2170                 return ret;
2171
2172         /*
2173          * wait for ordered IO before we have any locks.  We'll loop again
2174          * below with the locks held.
2175          */
2176         btrfs_wait_ordered_range(inode, alloc_start, alloc_end - alloc_start);
2177
2178         mutex_lock(&inode->i_mutex);
2179         ret = inode_newsize_ok(inode, alloc_end);
2180         if (ret)
2181                 goto out;
2182
2183         if (alloc_start > inode->i_size) {
2184                 ret = btrfs_cont_expand(inode, i_size_read(inode),
2185                                         alloc_start);
2186                 if (ret)
2187                         goto out;
2188         }
2189
2190         locked_end = alloc_end - 1;
2191         while (1) {
2192                 struct btrfs_ordered_extent *ordered;
2193
2194                 /* the extent lock is ordered inside the running
2195                  * transaction
2196                  */
2197                 lock_extent_bits(&BTRFS_I(inode)->io_tree, alloc_start,
2198                                  locked_end, 0, &cached_state);
2199                 ordered = btrfs_lookup_first_ordered_extent(inode,
2200                                                             alloc_end - 1);
2201                 if (ordered &&
2202                     ordered->file_offset + ordered->len > alloc_start &&
2203                     ordered->file_offset < alloc_end) {
2204                         btrfs_put_ordered_extent(ordered);
2205                         unlock_extent_cached(&BTRFS_I(inode)->io_tree,
2206                                              alloc_start, locked_end,
2207                                              &cached_state, GFP_NOFS);
2208                         /*
2209                          * we can't wait on the range with the transaction
2210                          * running or with the extent lock held
2211                          */
2212                         btrfs_wait_ordered_range(inode, alloc_start,
2213                                                  alloc_end - alloc_start);
2214                 } else {
2215                         if (ordered)
2216                                 btrfs_put_ordered_extent(ordered);
2217                         break;
2218                 }
2219         }
2220
2221         cur_offset = alloc_start;
2222         while (1) {
2223                 u64 actual_end;
2224
2225                 em = btrfs_get_extent(inode, NULL, 0, cur_offset,
2226                                       alloc_end - cur_offset, 0);
2227                 if (IS_ERR_OR_NULL(em)) {
2228                         if (!em)
2229                                 ret = -ENOMEM;
2230                         else
2231                                 ret = PTR_ERR(em);
2232                         break;
2233                 }
2234                 last_byte = min(extent_map_end(em), alloc_end);
2235                 actual_end = min_t(u64, extent_map_end(em), offset + len);
2236                 last_byte = ALIGN(last_byte, blocksize);
2237
2238                 if (em->block_start == EXTENT_MAP_HOLE ||
2239                     (cur_offset >= inode->i_size &&
2240                      !test_bit(EXTENT_FLAG_PREALLOC, &em->flags))) {
2241                         ret = btrfs_prealloc_file_range(inode, mode, cur_offset,
2242                                                         last_byte - cur_offset,
2243                                                         1 << inode->i_blkbits,
2244                                                         offset + len,
2245                                                         &alloc_hint);
2246
2247                         if (ret < 0) {
2248                                 free_extent_map(em);
2249                                 break;
2250                         }
2251                 } else if (actual_end > inode->i_size &&
2252                            !(mode & FALLOC_FL_KEEP_SIZE)) {
2253                         /*
2254                          * We didn't need to allocate any more space, but we
2255                          * still extended the size of the file so we need to
2256                          * update i_size.
2257                          */
2258                         inode->i_ctime = CURRENT_TIME;
2259                         i_size_write(inode, actual_end);
2260                         btrfs_ordered_update_i_size(inode, actual_end, NULL);
2261                 }
2262                 free_extent_map(em);
2263
2264                 cur_offset = last_byte;
2265                 if (cur_offset >= alloc_end) {
2266                         ret = 0;
2267                         break;
2268                 }
2269         }
2270         unlock_extent_cached(&BTRFS_I(inode)->io_tree, alloc_start, locked_end,
2271                              &cached_state, GFP_NOFS);
2272 out:
2273         mutex_unlock(&inode->i_mutex);
2274         /* Let go of our reservation. */
2275         btrfs_free_reserved_data_space(inode, alloc_end - alloc_start);
2276         return ret;
2277 }
2278
2279 static int find_desired_extent(struct inode *inode, loff_t *offset, int whence)
2280 {
2281         struct btrfs_root *root = BTRFS_I(inode)->root;
2282         struct extent_map *em;
2283         struct extent_state *cached_state = NULL;
2284         u64 lockstart = *offset;
2285         u64 lockend = i_size_read(inode);
2286         u64 start = *offset;
2287         u64 orig_start = *offset;
2288         u64 len = i_size_read(inode);
2289         u64 last_end = 0;
2290         int ret = 0;
2291
2292         lockend = max_t(u64, root->sectorsize, lockend);
2293         if (lockend <= lockstart)
2294                 lockend = lockstart + root->sectorsize;
2295
2296         lockend--;
2297         len = lockend - lockstart + 1;
2298
2299         len = max_t(u64, len, root->sectorsize);
2300         if (inode->i_size == 0)
2301                 return -ENXIO;
2302
2303         lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend, 0,
2304                          &cached_state);
2305
2306         /*
2307          * Delalloc is such a pain.  If we have a hole and we have pending
2308          * delalloc for a portion of the hole we will get back a hole that
2309          * exists for the entire range since it hasn't been actually written
2310          * yet.  So to take care of this case we need to look for an extent just
2311          * before the position we want in case there is outstanding delalloc
2312          * going on here.
2313          */
2314         if (whence == SEEK_HOLE && start != 0) {
2315                 if (start <= root->sectorsize)
2316                         em = btrfs_get_extent_fiemap(inode, NULL, 0, 0,
2317                                                      root->sectorsize, 0);
2318                 else
2319                         em = btrfs_get_extent_fiemap(inode, NULL, 0,
2320                                                      start - root->sectorsize,
2321                                                      root->sectorsize, 0);
2322                 if (IS_ERR(em)) {
2323                         ret = PTR_ERR(em);
2324                         goto out;
2325                 }
2326                 last_end = em->start + em->len;
2327                 if (em->block_start == EXTENT_MAP_DELALLOC)
2328                         last_end = min_t(u64, last_end, inode->i_size);
2329                 free_extent_map(em);
2330         }
2331
2332         while (1) {
2333                 em = btrfs_get_extent_fiemap(inode, NULL, 0, start, len, 0);
2334                 if (IS_ERR(em)) {
2335                         ret = PTR_ERR(em);
2336                         break;
2337                 }
2338
2339                 if (em->block_start == EXTENT_MAP_HOLE) {
2340                         if (test_bit(EXTENT_FLAG_VACANCY, &em->flags)) {
2341                                 if (last_end <= orig_start) {
2342                                         free_extent_map(em);
2343                                         ret = -ENXIO;
2344                                         break;
2345                                 }
2346                         }
2347
2348                         if (whence == SEEK_HOLE) {
2349                                 *offset = start;
2350                                 free_extent_map(em);
2351                                 break;
2352                         }
2353                 } else {
2354                         if (whence == SEEK_DATA) {
2355                                 if (em->block_start == EXTENT_MAP_DELALLOC) {
2356                                         if (start >= inode->i_size) {
2357                                                 free_extent_map(em);
2358                                                 ret = -ENXIO;
2359                                                 break;
2360                                         }
2361                                 }
2362
2363                                 if (!test_bit(EXTENT_FLAG_PREALLOC,
2364                                               &em->flags)) {
2365                                         *offset = start;
2366                                         free_extent_map(em);
2367                                         break;
2368                                 }
2369                         }
2370                 }
2371
2372                 start = em->start + em->len;
2373                 last_end = em->start + em->len;
2374
2375                 if (em->block_start == EXTENT_MAP_DELALLOC)
2376                         last_end = min_t(u64, last_end, inode->i_size);
2377
2378                 if (test_bit(EXTENT_FLAG_VACANCY, &em->flags)) {
2379                         free_extent_map(em);
2380                         ret = -ENXIO;
2381                         break;
2382                 }
2383                 free_extent_map(em);
2384                 cond_resched();
2385         }
2386         if (!ret)
2387                 *offset = min(*offset, inode->i_size);
2388 out:
2389         unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2390                              &cached_state, GFP_NOFS);
2391         return ret;
2392 }
2393
2394 static loff_t btrfs_file_llseek(struct file *file, loff_t offset, int whence)
2395 {
2396         struct inode *inode = file->f_mapping->host;
2397         int ret;
2398
2399         mutex_lock(&inode->i_mutex);
2400         switch (whence) {
2401         case SEEK_END:
2402         case SEEK_CUR:
2403                 offset = generic_file_llseek(file, offset, whence);
2404                 goto out;
2405         case SEEK_DATA:
2406         case SEEK_HOLE:
2407                 if (offset >= i_size_read(inode)) {
2408                         mutex_unlock(&inode->i_mutex);
2409                         return -ENXIO;
2410                 }
2411
2412                 ret = find_desired_extent(inode, &offset, whence);
2413                 if (ret) {
2414                         mutex_unlock(&inode->i_mutex);
2415                         return ret;
2416                 }
2417         }
2418
2419         if (offset < 0 && !(file->f_mode & FMODE_UNSIGNED_OFFSET)) {
2420                 offset = -EINVAL;
2421                 goto out;
2422         }
2423         if (offset > inode->i_sb->s_maxbytes) {
2424                 offset = -EINVAL;
2425                 goto out;
2426         }
2427
2428         /* Special lock needed here? */
2429         if (offset != file->f_pos) {
2430                 file->f_pos = offset;
2431                 file->f_version = 0;
2432         }
2433 out:
2434         mutex_unlock(&inode->i_mutex);
2435         return offset;
2436 }
2437
2438 const struct file_operations btrfs_file_operations = {
2439         .llseek         = btrfs_file_llseek,
2440         .read           = do_sync_read,
2441         .write          = do_sync_write,
2442         .aio_read       = generic_file_aio_read,
2443         .splice_read    = generic_file_splice_read,
2444         .aio_write      = btrfs_file_aio_write,
2445         .mmap           = btrfs_file_mmap,
2446         .open           = generic_file_open,
2447         .release        = btrfs_release_file,
2448         .fsync          = btrfs_sync_file,
2449         .fallocate      = btrfs_fallocate,
2450         .unlocked_ioctl = btrfs_ioctl,
2451 #ifdef CONFIG_COMPAT
2452         .compat_ioctl   = btrfs_ioctl,
2453 #endif
2454 };
2455
2456 void btrfs_auto_defrag_exit(void)
2457 {
2458         if (btrfs_inode_defrag_cachep)
2459                 kmem_cache_destroy(btrfs_inode_defrag_cachep);
2460 }
2461
2462 int btrfs_auto_defrag_init(void)
2463 {
2464         btrfs_inode_defrag_cachep = kmem_cache_create("btrfs_inode_defrag",
2465                                         sizeof(struct inode_defrag), 0,
2466                                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD,
2467                                         NULL);
2468         if (!btrfs_inode_defrag_cachep)
2469                 return -ENOMEM;
2470
2471         return 0;
2472 }