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/compat.h>
  31 #include <linux/slab.h>
  32 #include <linux/btrfs.h>
  33 #include <linux/uio.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 "volumes.h"
  42 #include "qgroup.h"
  43 #include "compression.h"
  44
  45 static struct kmem_cache *btrfs_inode_defrag_cachep;
  46 /*
  47  * when auto defrag is enabled we
  48  * queue up these defrag structs to remember which
  49  * inodes need defragging passes
  50  */
  51 struct inode_defrag {
  52         struct rb_node rb_node;
  53         /* objectid */
  54         u64 ino;
  55         /*
  56          * transid where the defrag was added, we search for
  57          * extents newer than this
  58          */
  59         u64 transid;
  60
  61         /* root objectid */
  62         u64 root;
  63
  64         /* last offset we were able to defrag */
  65         u64 last_offset;
  66
  67         /* if we've wrapped around back to zero once already */
  68         int cycled;
  69 };
  70
  71 static int __compare_inode_defrag(struct inode_defrag *defrag1,
  72                                   struct inode_defrag *defrag2)
  73 {
  74         if (defrag1->root > defrag2->root)
  75                 return 1;
  76         else if (defrag1->root < defrag2->root)
  77                 return -1;
  78         else if (defrag1->ino > defrag2->ino)
  79                 return 1;
  80         else if (defrag1->ino < defrag2->ino)
  81                 return -1;
  82         else
  83                 return 0;
  84 }
  85
  86 /* pop a record for an inode into the defrag tree.  The lock
  87  * must be held already
  88  *
  89  * If you're inserting a record for an older transid than an
  90  * existing record, the transid already in the tree is lowered
  91  *
  92  * If an existing record is found the defrag item you
  93  * pass in is freed
  94  */
  95 static int __btrfs_add_inode_defrag(struct btrfs_inode *inode,
  96                                     struct inode_defrag *defrag)
  97 {
  98         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
  99         struct inode_defrag *entry;
 100         struct rb_node **p;
 101         struct rb_node *parent = NULL;
 102         int ret;
 103
 104         p = &fs_info->defrag_inodes.rb_node;
 105         while (*p) {
 106                 parent = *p;
 107                 entry = rb_entry(parent, struct inode_defrag, rb_node);
 108
 109                 ret = __compare_inode_defrag(defrag, entry);
 110                 if (ret < 0)
 111                         p = &parent->rb_left;
 112                 else if (ret > 0)
 113                         p = &parent->rb_right;
 114                 else {
 115                         /* if we're reinserting an entry for
 116                          * an old defrag run, make sure to
 117                          * lower the transid of our existing record
 118                          */
 119                         if (defrag->transid < entry->transid)
 120                                 entry->transid = defrag->transid;
 121                         if (defrag->last_offset > entry->last_offset)
 122                                 entry->last_offset = defrag->last_offset;
 123                         return -EEXIST;
 124                 }
 125         }
 126         set_bit(BTRFS_INODE_IN_DEFRAG, &inode->runtime_flags);
 127         rb_link_node(&defrag->rb_node, parent, p);
 128         rb_insert_color(&defrag->rb_node, &fs_info->defrag_inodes);
 129         return 0;
 130 }
 131
 132 static inline int __need_auto_defrag(struct btrfs_fs_info *fs_info)
 133 {
 134         if (!btrfs_test_opt(fs_info, AUTO_DEFRAG))
 135                 return 0;
 136
 137         if (btrfs_fs_closing(fs_info))
 138                 return 0;
 139
 140         return 1;
 141 }
 142
 143 /*
 144  * insert a defrag record for this inode if auto defrag is
 145  * enabled
 146  */
 147 int btrfs_add_inode_defrag(struct btrfs_trans_handle *trans,
 148                            struct btrfs_inode *inode)
 149 {
 150         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
 151         struct btrfs_root *root = inode->root;
 152         struct inode_defrag *defrag;
 153         u64 transid;
 154         int ret;
 155
 156         if (!__need_auto_defrag(fs_info))
 157                 return 0;
 158
 159         if (test_bit(BTRFS_INODE_IN_DEFRAG, &inode->runtime_flags))
 160                 return 0;
 161
 162         if (trans)
 163                 transid = trans->transid;
 164         else
 165                 transid = inode->root->last_trans;
 166
 167         defrag = kmem_cache_zalloc(btrfs_inode_defrag_cachep, GFP_NOFS);
 168         if (!defrag)
 169                 return -ENOMEM;
 170
 171         defrag->ino = btrfs_ino(inode);
 172         defrag->transid = transid;
 173         defrag->root = root->root_key.objectid;
 174
 175         spin_lock(&fs_info->defrag_inodes_lock);
 176         if (!test_bit(BTRFS_INODE_IN_DEFRAG, &inode->runtime_flags)) {
 177                 /*
 178                  * If we set IN_DEFRAG flag and evict the inode from memory,
 179                  * and then re-read this inode, this new inode doesn't have
 180                  * IN_DEFRAG flag. At the case, we may find the existed defrag.
 181                  */
 182                 ret = __btrfs_add_inode_defrag(inode, defrag);
 183                 if (ret)
 184                         kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
 185         } else {
 186                 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
 187         }
 188         spin_unlock(&fs_info->defrag_inodes_lock);
 189         return 0;
 190 }
 191
 192 /*
 193  * Requeue the defrag object. If there is a defrag object that points to
 194  * the same inode in the tree, we will merge them together (by
 195  * __btrfs_add_inode_defrag()) and free the one that we want to requeue.
 196  */
 197 static void btrfs_requeue_inode_defrag(struct btrfs_inode *inode,
 198                                        struct inode_defrag *defrag)
 199 {
 200         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
 201         int ret;
 202
 203         if (!__need_auto_defrag(fs_info))
 204                 goto out;
 205
 206         /*
 207          * Here we don't check the IN_DEFRAG flag, because we need merge
 208          * them together.
 209          */
 210         spin_lock(&fs_info->defrag_inodes_lock);
 211         ret = __btrfs_add_inode_defrag(inode, defrag);
 212         spin_unlock(&fs_info->defrag_inodes_lock);
 213         if (ret)
 214                 goto out;
 215         return;
 216 out:
 217         kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
 218 }
 219
 220 /*
 221  * pick the defragable inode that we want, if it doesn't exist, we will get
 222  * the next one.
 223  */
 224 static struct inode_defrag *
 225 btrfs_pick_defrag_inode(struct btrfs_fs_info *fs_info, u64 root, u64 ino)
 226 {
 227         struct inode_defrag *entry = NULL;
 228         struct inode_defrag tmp;
 229         struct rb_node *p;
 230         struct rb_node *parent = NULL;
 231         int ret;
 232
 233         tmp.ino = ino;
 234         tmp.root = root;
 235
 236         spin_lock(&fs_info->defrag_inodes_lock);
 237         p = fs_info->defrag_inodes.rb_node;
 238         while (p) {
 239                 parent = p;
 240                 entry = rb_entry(parent, struct inode_defrag, rb_node);
 241
 242                 ret = __compare_inode_defrag(&tmp, entry);
 243                 if (ret < 0)
 244                         p = parent->rb_left;
 245                 else if (ret > 0)
 246                         p = parent->rb_right;
 247                 else
 248                         goto out;
 249         }
 250
 251         if (parent && __compare_inode_defrag(&tmp, entry) > 0) {
 252                 parent = rb_next(parent);
 253                 if (parent)
 254                         entry = rb_entry(parent, struct inode_defrag, rb_node);
 255                 else
 256                         entry = NULL;
 257         }
 258 out:
 259         if (entry)
 260                 rb_erase(parent, &fs_info->defrag_inodes);
 261         spin_unlock(&fs_info->defrag_inodes_lock);
 262         return entry;
 263 }
 264
 265 void btrfs_cleanup_defrag_inodes(struct btrfs_fs_info *fs_info)
 266 {
 267         struct inode_defrag *defrag;
 268         struct rb_node *node;
 269
 270         spin_lock(&fs_info->defrag_inodes_lock);
 271         node = rb_first(&fs_info->defrag_inodes);
 272         while (node) {
 273                 rb_erase(node, &fs_info->defrag_inodes);
 274                 defrag = rb_entry(node, struct inode_defrag, rb_node);
 275                 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
 276
 277                 cond_resched_lock(&fs_info->defrag_inodes_lock);
 278
 279                 node = rb_first(&fs_info->defrag_inodes);
 280         }
 281         spin_unlock(&fs_info->defrag_inodes_lock);
 282 }
 283
 284 #define BTRFS_DEFRAG_BATCH      1024
 285
 286 static int __btrfs_run_defrag_inode(struct btrfs_fs_info *fs_info,
 287                                     struct inode_defrag *defrag)
 288 {
 289         struct btrfs_root *inode_root;
 290         struct inode *inode;
 291         struct btrfs_key key;
 292         struct btrfs_ioctl_defrag_range_args range;
 293         int num_defrag;
 294         int index;
 295         int ret;
 296
 297         /* get the inode */
 298         key.objectid = defrag->root;
 299         key.type = BTRFS_ROOT_ITEM_KEY;
 300         key.offset = (u64)-1;
 301
 302         index = srcu_read_lock(&fs_info->subvol_srcu);
 303
 304         inode_root = btrfs_read_fs_root_no_name(fs_info, &key);
 305         if (IS_ERR(inode_root)) {
 306                 ret = PTR_ERR(inode_root);
 307                 goto cleanup;
 308         }
 309
 310         key.objectid = defrag->ino;
 311         key.type = BTRFS_INODE_ITEM_KEY;
 312         key.offset = 0;
 313         inode = btrfs_iget(fs_info->sb, &key, inode_root, NULL);
 314         if (IS_ERR(inode)) {
 315                 ret = PTR_ERR(inode);
 316                 goto cleanup;
 317         }
 318         srcu_read_unlock(&fs_info->subvol_srcu, index);
 319
 320         /* do a chunk of defrag */
 321         clear_bit(BTRFS_INODE_IN_DEFRAG, &BTRFS_I(inode)->runtime_flags);
 322         memset(&range, 0, sizeof(range));
 323         range.len = (u64)-1;
 324         range.start = defrag->last_offset;
 325
 326         sb_start_write(fs_info->sb);
 327         num_defrag = btrfs_defrag_file(inode, NULL, &range, defrag->transid,
 328                                        BTRFS_DEFRAG_BATCH);
 329         sb_end_write(fs_info->sb);
 330         /*
 331          * if we filled the whole defrag batch, there
 332          * must be more work to do.  Queue this defrag
 333          * again
 334          */
 335         if (num_defrag == BTRFS_DEFRAG_BATCH) {
 336                 defrag->last_offset = range.start;
 337                 btrfs_requeue_inode_defrag(BTRFS_I(inode), defrag);
 338         } else if (defrag->last_offset && !defrag->cycled) {
 339                 /*
 340                  * we didn't fill our defrag batch, but
 341                  * we didn't start at zero.  Make sure we loop
 342                  * around to the start of the file.
 343                  */
 344                 defrag->last_offset = 0;
 345                 defrag->cycled = 1;
 346                 btrfs_requeue_inode_defrag(BTRFS_I(inode), defrag);
 347         } else {
 348                 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
 349         }
 350
 351         iput(inode);
 352         return 0;
 353 cleanup:
 354         srcu_read_unlock(&fs_info->subvol_srcu, index);
 355         kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
 356         return ret;
 357 }
 358
 359 /*
 360  * run through the list of inodes in the FS that need
 361  * defragging
 362  */
 363 int btrfs_run_defrag_inodes(struct btrfs_fs_info *fs_info)
 364 {
 365         struct inode_defrag *defrag;
 366         u64 first_ino = 0;
 367         u64 root_objectid = 0;
 368
 369         atomic_inc(&fs_info->defrag_running);
 370         while (1) {
 371                 /* Pause the auto defragger. */
 372                 if (test_bit(BTRFS_FS_STATE_REMOUNTING,
 373                              &fs_info->fs_state))
 374                         break;
 375
 376                 if (!__need_auto_defrag(fs_info))
 377                         break;
 378
 379                 /* find an inode to defrag */
 380                 defrag = btrfs_pick_defrag_inode(fs_info, root_objectid,
 381                                                  first_ino);
 382                 if (!defrag) {
 383                         if (root_objectid || first_ino) {
 384                                 root_objectid = 0;
 385                                 first_ino = 0;
 386                                 continue;
 387                         } else {
 388                                 break;
 389                         }
 390                 }
 391
 392                 first_ino = defrag->ino + 1;
 393                 root_objectid = defrag->root;
 394
 395                 __btrfs_run_defrag_inode(fs_info, defrag);
 396         }
 397         atomic_dec(&fs_info->defrag_running);
 398
 399         /*
 400          * during unmount, we use the transaction_wait queue to
 401          * wait for the defragger to stop
 402          */
 403         wake_up(&fs_info->transaction_wait);
 404         return 0;
 405 }
 406
 407 /* simple helper to fault in pages and copy.  This should go away
 408  * and be replaced with calls into generic code.
 409  */
 410 static noinline int btrfs_copy_from_user(loff_t pos, size_t write_bytes,
 411                                          struct page **prepared_pages,
 412                                          struct iov_iter *i)
 413 {
 414         size_t copied = 0;
 415         size_t total_copied = 0;
 416         int pg = 0;
 417         int offset = pos & (PAGE_SIZE - 1);
 418
 419         while (write_bytes > 0) {
 420                 size_t count = min_t(size_t,
 421                                      PAGE_SIZE - offset, write_bytes);
 422                 struct page *page = prepared_pages[pg];
 423                 /*
 424                  * Copy data from userspace to the current page
 425                  */
 426                 copied = iov_iter_copy_from_user_atomic(page, i, offset, count);
 427
 428                 /* Flush processor's dcache for this page */
 429                 flush_dcache_page(page);
 430
 431                 /*
 432                  * if we get a partial write, we can end up with
 433                  * partially up to date pages.  These add
 434                  * a lot of complexity, so make sure they don't
 435                  * happen by forcing this copy to be retried.
 436                  *
 437                  * The rest of the btrfs_file_write code will fall
 438                  * back to page at a time copies after we return 0.
 439                  */
 440                 if (!PageUptodate(page) && copied < count)
 441                         copied = 0;
 442
 443                 iov_iter_advance(i, copied);
 444                 write_bytes -= copied;
 445                 total_copied += copied;
 446
 447                 /* Return to btrfs_file_write_iter to fault page */
 448                 if (unlikely(copied == 0))
 449                         break;
 450
 451                 if (copied < PAGE_SIZE - offset) {
 452                         offset += copied;
 453                 } else {
 454                         pg++;
 455                         offset = 0;
 456                 }
 457         }
 458         return total_copied;
 459 }
 460
 461 /*
 462  * unlocks pages after btrfs_file_write is done with them
 463  */
 464 static void btrfs_drop_pages(struct page **pages, size_t num_pages)
 465 {
 466         size_t i;
 467         for (i = 0; i < num_pages; i++) {
 468                 /* page checked is some magic around finding pages that
 469                  * have been modified without going through btrfs_set_page_dirty
 470                  * clear it here. There should be no need to mark the pages
 471                  * accessed as prepare_pages should have marked them accessed
 472                  * in prepare_pages via find_or_create_page()
 473                  */
 474                 ClearPageChecked(pages[i]);
 475                 unlock_page(pages[i]);
 476                 put_page(pages[i]);
 477         }
 478 }
 479
 480 /*
 481  * after copy_from_user, pages need to be dirtied and we need to make
 482  * sure holes are created between the current EOF and the start of
 483  * any next extents (if required).
 484  *
 485  * this also makes the decision about creating an inline extent vs
 486  * doing real data extents, marking pages dirty and delalloc as required.
 487  */
 488 int btrfs_dirty_pages(struct inode *inode, struct page **pages,
 489                       size_t num_pages, loff_t pos, size_t write_bytes,
 490                       struct extent_state **cached)
 491 {
 492         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
 493         int err = 0;
 494         int i;
 495         u64 num_bytes;
 496         u64 start_pos;
 497         u64 end_of_last_block;
 498         u64 end_pos = pos + write_bytes;
 499         loff_t isize = i_size_read(inode);
 500
 501         start_pos = pos & ~((u64) fs_info->sectorsize - 1);
 502         num_bytes = round_up(write_bytes + pos - start_pos,
 503                              fs_info->sectorsize);
 504
 505         end_of_last_block = start_pos + num_bytes - 1;
 506         err = btrfs_set_extent_delalloc(inode, start_pos, end_of_last_block,
 507                                         cached, 0);
 508         if (err)
 509                 return err;
 510
 511         for (i = 0; i < num_pages; i++) {
 512                 struct page *p = pages[i];
 513                 SetPageUptodate(p);
 514                 ClearPageChecked(p);
 515                 set_page_dirty(p);
 516         }
 517
 518         /*
 519          * we've only changed i_size in ram, and we haven't updated
 520          * the disk i_size.  There is no need to log the inode
 521          * at this time.
 522          */
 523         if (end_pos > isize)
 524                 i_size_write(inode, end_pos);
 525         return 0;
 526 }
 527
 528 /*
 529  * this drops all the extents in the cache that intersect the range
 530  * [start, end].  Existing extents are split as required.
 531  */
 532 void btrfs_drop_extent_cache(struct btrfs_inode *inode, u64 start, u64 end,
 533                              int skip_pinned)
 534 {
 535         struct extent_map *em;
 536         struct extent_map *split = NULL;
 537         struct extent_map *split2 = NULL;
 538         struct extent_map_tree *em_tree = &inode->extent_tree;
 539         u64 len = end - start + 1;
 540         u64 gen;
 541         int ret;
 542         int testend = 1;
 543         unsigned long flags;
 544         int compressed = 0;
 545         bool modified;
 546
 547         WARN_ON(end < start);
 548         if (end == (u64)-1) {
 549                 len = (u64)-1;
 550                 testend = 0;
 551         }
 552         while (1) {
 553                 int no_splits = 0;
 554
 555                 modified = false;
 556                 if (!split)
 557                         split = alloc_extent_map();
 558                 if (!split2)
 559                         split2 = alloc_extent_map();
 560                 if (!split || !split2)
 561                         no_splits = 1;
 562
 563                 write_lock(&em_tree->lock);
 564                 em = lookup_extent_mapping(em_tree, start, len);
 565                 if (!em) {
 566                         write_unlock(&em_tree->lock);
 567                         break;
 568                 }
 569                 flags = em->flags;
 570                 gen = em->generation;
 571                 if (skip_pinned && test_bit(EXTENT_FLAG_PINNED, &em->flags)) {
 572                         if (testend && em->start + em->len >= start + len) {
 573                                 free_extent_map(em);
 574                                 write_unlock(&em_tree->lock);
 575                                 break;
 576                         }
 577                         start = em->start + em->len;
 578                         if (testend)
 579                                 len = start + len - (em->start + em->len);
 580                         free_extent_map(em);
 581                         write_unlock(&em_tree->lock);
 582                         continue;
 583                 }
 584                 compressed = test_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
 585                 clear_bit(EXTENT_FLAG_PINNED, &em->flags);
 586                 clear_bit(EXTENT_FLAG_LOGGING, &flags);
 587                 modified = !list_empty(&em->list);
 588                 if (no_splits)
 589                         goto next;
 590
 591                 if (em->start < start) {
 592                         split->start = em->start;
 593                         split->len = start - em->start;
 594
 595                         if (em->block_start < EXTENT_MAP_LAST_BYTE) {
 596                                 split->orig_start = em->orig_start;
 597                                 split->block_start = em->block_start;
 598
 599                                 if (compressed)
 600                                         split->block_len = em->block_len;
 601                                 else
 602                                         split->block_len = split->len;
 603                                 split->orig_block_len = max(split->block_len,
 604                                                 em->orig_block_len);
 605                                 split->ram_bytes = em->ram_bytes;
 606                         } else {
 607                                 split->orig_start = split->start;
 608                                 split->block_len = 0;
 609                                 split->block_start = em->block_start;
 610                                 split->orig_block_len = 0;
 611                                 split->ram_bytes = split->len;
 612                         }
 613
 614                         split->generation = gen;
 615                         split->bdev = em->bdev;
 616                         split->flags = flags;
 617                         split->compress_type = em->compress_type;
 618                         replace_extent_mapping(em_tree, em, split, modified);
 619                         free_extent_map(split);
 620                         split = split2;
 621                         split2 = NULL;
 622                 }
 623                 if (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
 633                         if (em->block_start < EXTENT_MAP_LAST_BYTE) {
 634                                 split->orig_block_len = max(em->block_len,
 635                                                     em->orig_block_len);
 636
 637                                 split->ram_bytes = em->ram_bytes;
 638                                 if (compressed) {
 639                                         split->block_len = em->block_len;
 640                                         split->block_start = em->block_start;
 641                                         split->orig_start = em->orig_start;
 642                                 } else {
 643                                         split->block_len = split->len;
 644                                         split->block_start = em->block_start
 645                                                 + diff;
 646                                         split->orig_start = em->orig_start;
 647                                 }
 648                         } else {
 649                                 split->ram_bytes = split->len;
 650                                 split->orig_start = split->start;
 651                                 split->block_len = 0;
 652                                 split->block_start = em->block_start;
 653                                 split->orig_block_len = 0;
 654                         }
 655
 656                         if (extent_map_in_tree(em)) {
 657                                 replace_extent_mapping(em_tree, em, split,
 658                                                        modified);
 659                         } else {
 660                                 ret = add_extent_mapping(em_tree, split,
 661                                                          modified);
 662                                 ASSERT(ret == 0); /* Logic error */
 663                         }
 664                         free_extent_map(split);
 665                         split = NULL;
 666                 }
 667 next:
 668                 if (extent_map_in_tree(em))
 669                         remove_extent_mapping(em_tree, em);
 670                 write_unlock(&em_tree->lock);
 671
 672                 /* once for us */
 673                 free_extent_map(em);
 674                 /* once for the tree*/
 675                 free_extent_map(em);
 676         }
 677         if (split)
 678                 free_extent_map(split);
 679         if (split2)
 680                 free_extent_map(split2);
 681 }
 682
 683 /*
 684  * this is very complex, but the basic idea is to drop all extents
 685  * in the range start - end.  hint_block is filled in with a block number
 686  * that would be a good hint to the block allocator for this file.
 687  *
 688  * If an extent intersects the range but is not entirely inside the range
 689  * it is either truncated or split.  Anything entirely inside the range
 690  * is deleted from the tree.
 691  */
 692 int __btrfs_drop_extents(struct btrfs_trans_handle *trans,
 693                          struct btrfs_root *root, struct inode *inode,
 694                          struct btrfs_path *path, u64 start, u64 end,
 695                          u64 *drop_end, int drop_cache,
 696                          int replace_extent,
 697                          u32 extent_item_size,
 698                          int *key_inserted)
 699 {
 700         struct btrfs_fs_info *fs_info = root->fs_info;
 701         struct extent_buffer *leaf;
 702         struct btrfs_file_extent_item *fi;
 703         struct btrfs_key key;
 704         struct btrfs_key new_key;
 705         u64 ino = btrfs_ino(BTRFS_I(inode));
 706         u64 search_start = start;
 707         u64 disk_bytenr = 0;
 708         u64 num_bytes = 0;
 709         u64 extent_offset = 0;
 710         u64 extent_end = 0;
 711         u64 last_end = start;
 712         int del_nr = 0;
 713         int del_slot = 0;
 714         int extent_type;
 715         int recow;
 716         int ret;
 717         int modify_tree = -1;
 718         int update_refs;
 719         int found = 0;
 720         int leafs_visited = 0;
 721
 722         if (drop_cache)
 723                 btrfs_drop_extent_cache(BTRFS_I(inode), start, end - 1, 0);
 724
 725         if (start >= BTRFS_I(inode)->disk_i_size && !replace_extent)
 726                 modify_tree = 0;
 727
 728         update_refs = (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
 729                        root == fs_info->tree_root);
 730         while (1) {
 731                 recow = 0;
 732                 ret = btrfs_lookup_file_extent(trans, root, path, ino,
 733                                                search_start, modify_tree);
 734                 if (ret < 0)
 735                         break;
 736                 if (ret > 0 && path->slots[0] > 0 && search_start == start) {
 737                         leaf = path->nodes[0];
 738                         btrfs_item_key_to_cpu(leaf, &key, path->slots[0] - 1);
 739                         if (key.objectid == ino &&
 740                             key.type == BTRFS_EXTENT_DATA_KEY)
 741                                 path->slots[0]--;
 742                 }
 743                 ret = 0;
 744                 leafs_visited++;
 745 next_slot:
 746                 leaf = path->nodes[0];
 747                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
 748                         BUG_ON(del_nr > 0);
 749                         ret = btrfs_next_leaf(root, path);
 750                         if (ret < 0)
 751                                 break;
 752                         if (ret > 0) {
 753                                 ret = 0;
 754                                 break;
 755                         }
 756                         leafs_visited++;
 757                         leaf = path->nodes[0];
 758                         recow = 1;
 759                 }
 760
 761                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 762
 763                 if (key.objectid > ino)
 764                         break;
 765                 if (WARN_ON_ONCE(key.objectid < ino) ||
 766                     key.type < BTRFS_EXTENT_DATA_KEY) {
 767                         ASSERT(del_nr == 0);
 768                         path->slots[0]++;
 769                         goto next_slot;
 770                 }
 771                 if (key.type > BTRFS_EXTENT_DATA_KEY || key.offset >= end)
 772                         break;
 773
 774                 fi = btrfs_item_ptr(leaf, path->slots[0],
 775                                     struct btrfs_file_extent_item);
 776                 extent_type = btrfs_file_extent_type(leaf, fi);
 777
 778                 if (extent_type == BTRFS_FILE_EXTENT_REG ||
 779                     extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
 780                         disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
 781                         num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
 782                         extent_offset = btrfs_file_extent_offset(leaf, fi);
 783                         extent_end = key.offset +
 784                                 btrfs_file_extent_num_bytes(leaf, fi);
 785                 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
 786                         extent_end = key.offset +
 787                                 btrfs_file_extent_ram_bytes(leaf, fi);
 788                 } else {
 789                         /* can't happen */
 790                         BUG();
 791                 }
 792
 793                 /*
 794                  * Don't skip extent items representing 0 byte lengths. They
 795                  * used to be created (bug) if while punching holes we hit
 796                  * -ENOSPC condition. So if we find one here, just ensure we
 797                  * delete it, otherwise we would insert a new file extent item
 798                  * with the same key (offset) as that 0 bytes length file
 799                  * extent item in the call to setup_items_for_insert() later
 800                  * in this function.
 801                  */
 802                 if (extent_end == key.offset && extent_end >= search_start) {
 803                         last_end = extent_end;
 804                         goto delete_extent_item;
 805                 }
 806
 807                 if (extent_end <= search_start) {
 808                         path->slots[0]++;
 809                         goto next_slot;
 810                 }
 811
 812                 found = 1;
 813                 search_start = max(key.offset, start);
 814                 if (recow || !modify_tree) {
 815                         modify_tree = -1;
 816                         btrfs_release_path(path);
 817                         continue;
 818                 }
 819
 820                 /*
 821                  *     | - range to drop - |
 822                  *  | -------- extent -------- |
 823                  */
 824                 if (start > key.offset && end < extent_end) {
 825                         BUG_ON(del_nr > 0);
 826                         if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
 827                                 ret = -EOPNOTSUPP;
 828                                 break;
 829                         }
 830
 831                         memcpy(&new_key, &key, sizeof(new_key));
 832                         new_key.offset = start;
 833                         ret = btrfs_duplicate_item(trans, root, path,
 834                                                    &new_key);
 835                         if (ret == -EAGAIN) {
 836                                 btrfs_release_path(path);
 837                                 continue;
 838                         }
 839                         if (ret < 0)
 840                                 break;
 841
 842                         leaf = path->nodes[0];
 843                         fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
 844                                             struct btrfs_file_extent_item);
 845                         btrfs_set_file_extent_num_bytes(leaf, fi,
 846                                                         start - key.offset);
 847
 848                         fi = btrfs_item_ptr(leaf, path->slots[0],
 849                                             struct btrfs_file_extent_item);
 850
 851                         extent_offset += start - key.offset;
 852                         btrfs_set_file_extent_offset(leaf, fi, extent_offset);
 853                         btrfs_set_file_extent_num_bytes(leaf, fi,
 854                                                         extent_end - start);
 855                         btrfs_mark_buffer_dirty(leaf);
 856
 857                         if (update_refs && disk_bytenr > 0) {
 858                                 ret = btrfs_inc_extent_ref(trans, fs_info,
 859                                                 disk_bytenr, num_bytes, 0,
 860                                                 root->root_key.objectid,
 861                                                 new_key.objectid,
 862                                                 start - extent_offset);
 863                                 BUG_ON(ret); /* -ENOMEM */
 864                         }
 865                         key.offset = start;
 866                 }
 867                 /*
 868                  * From here on out we will have actually dropped something, so
 869                  * last_end can be updated.
 870                  */
 871                 last_end = extent_end;
 872
 873                 /*
 874                  *  | ---- range to drop ----- |
 875                  *      | -------- extent -------- |
 876                  */
 877                 if (start <= key.offset && end < extent_end) {
 878                         if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
 879                                 ret = -EOPNOTSUPP;
 880                                 break;
 881                         }
 882
 883                         memcpy(&new_key, &key, sizeof(new_key));
 884                         new_key.offset = end;
 885                         btrfs_set_item_key_safe(fs_info, path, &new_key);
 886
 887                         extent_offset += end - key.offset;
 888                         btrfs_set_file_extent_offset(leaf, fi, extent_offset);
 889                         btrfs_set_file_extent_num_bytes(leaf, fi,
 890                                                         extent_end - end);
 891                         btrfs_mark_buffer_dirty(leaf);
 892                         if (update_refs && disk_bytenr > 0)
 893                                 inode_sub_bytes(inode, end - key.offset);
 894                         break;
 895                 }
 896
 897                 search_start = extent_end;
 898                 /*
 899                  *       | ---- range to drop ----- |
 900                  *  | -------- extent -------- |
 901                  */
 902                 if (start > key.offset && end >= extent_end) {
 903                         BUG_ON(del_nr > 0);
 904                         if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
 905                                 ret = -EOPNOTSUPP;
 906                                 break;
 907                         }
 908
 909                         btrfs_set_file_extent_num_bytes(leaf, fi,
 910                                                         start - key.offset);
 911                         btrfs_mark_buffer_dirty(leaf);
 912                         if (update_refs && disk_bytenr > 0)
 913                                 inode_sub_bytes(inode, extent_end - start);
 914                         if (end == extent_end)
 915                                 break;
 916
 917                         path->slots[0]++;
 918                         goto next_slot;
 919                 }
 920
 921                 /*
 922                  *  | ---- range to drop ----- |
 923                  *    | ------ extent ------ |
 924                  */
 925                 if (start <= key.offset && end >= extent_end) {
 926 delete_extent_item:
 927                         if (del_nr == 0) {
 928                                 del_slot = path->slots[0];
 929                                 del_nr = 1;
 930                         } else {
 931                                 BUG_ON(del_slot + del_nr != path->slots[0]);
 932                                 del_nr++;
 933                         }
 934
 935                         if (update_refs &&
 936                             extent_type == BTRFS_FILE_EXTENT_INLINE) {
 937                                 inode_sub_bytes(inode,
 938                                                 extent_end - key.offset);
 939                                 extent_end = ALIGN(extent_end,
 940                                                    fs_info->sectorsize);
 941                         } else if (update_refs && disk_bytenr > 0) {
 942                                 ret = btrfs_free_extent(trans, fs_info,
 943                                                 disk_bytenr, num_bytes, 0,
 944                                                 root->root_key.objectid,
 945                                                 key.objectid, key.offset -
 946                                                 extent_offset);
 947                                 BUG_ON(ret); /* -ENOMEM */
 948                                 inode_sub_bytes(inode,
 949                                                 extent_end - key.offset);
 950                         }
 951
 952                         if (end == extent_end)
 953                                 break;
 954
 955                         if (path->slots[0] + 1 < btrfs_header_nritems(leaf)) {
 956                                 path->slots[0]++;
 957                                 goto next_slot;
 958                         }
 959
 960                         ret = btrfs_del_items(trans, root, path, del_slot,
 961                                               del_nr);
 962                         if (ret) {
 963                                 btrfs_abort_transaction(trans, ret);
 964                                 break;
 965                         }
 966
 967                         del_nr = 0;
 968                         del_slot = 0;
 969
 970                         btrfs_release_path(path);
 971                         continue;
 972                 }
 973
 974                 BUG_ON(1);
 975         }
 976
 977         if (!ret && del_nr > 0) {
 978                 /*
 979                  * Set path->slots[0] to first slot, so that after the delete
 980                  * if items are move off from our leaf to its immediate left or
 981                  * right neighbor leafs, we end up with a correct and adjusted
 982                  * path->slots[0] for our insertion (if replace_extent != 0).
 983                  */
 984                 path->slots[0] = del_slot;
 985                 ret = btrfs_del_items(trans, root, path, del_slot, del_nr);
 986                 if (ret)
 987                         btrfs_abort_transaction(trans, ret);
 988         }
 989
 990         leaf = path->nodes[0];
 991         /*
 992          * If btrfs_del_items() was called, it might have deleted a leaf, in
 993          * which case it unlocked our path, so check path->locks[0] matches a
 994          * write lock.
 995          */
 996         if (!ret && replace_extent && leafs_visited == 1 &&
 997             (path->locks[0] == BTRFS_WRITE_LOCK_BLOCKING ||
 998              path->locks[0] == BTRFS_WRITE_LOCK) &&
 999             btrfs_leaf_free_space(fs_info, leaf) >=
1000             sizeof(struct btrfs_item) + extent_item_size) {
1001
1002                 key.objectid = ino;
1003                 key.type = BTRFS_EXTENT_DATA_KEY;
1004                 key.offset = start;
1005                 if (!del_nr && path->slots[0] < btrfs_header_nritems(leaf)) {
1006                         struct btrfs_key slot_key;
1007
1008                         btrfs_item_key_to_cpu(leaf, &slot_key, path->slots[0]);
1009                         if (btrfs_comp_cpu_keys(&key, &slot_key) > 0)
1010                                 path->slots[0]++;
1011                 }
1012                 setup_items_for_insert(root, path, &key,
1013                                        &extent_item_size,
1014                                        extent_item_size,
1015                                        sizeof(struct btrfs_item) +
1016                                        extent_item_size, 1);
1017                 *key_inserted = 1;
1018         }
1019
1020         if (!replace_extent || !(*key_inserted))
1021                 btrfs_release_path(path);
1022         if (drop_end)
1023                 *drop_end = found ? min(end, last_end) : end;
1024         return ret;
1025 }
1026
1027 int btrfs_drop_extents(struct btrfs_trans_handle *trans,
1028                        struct btrfs_root *root, struct inode *inode, u64 start,
1029                        u64 end, int drop_cache)
1030 {
1031         struct btrfs_path *path;
1032         int ret;
1033
1034         path = btrfs_alloc_path();
1035         if (!path)
1036                 return -ENOMEM;
1037         ret = __btrfs_drop_extents(trans, root, inode, path, start, end, NULL,
1038                                    drop_cache, 0, 0, NULL);
1039         btrfs_free_path(path);
1040         return ret;
1041 }
1042
1043 static int extent_mergeable(struct extent_buffer *leaf, int slot,
1044                             u64 objectid, u64 bytenr, u64 orig_offset,
1045                             u64 *start, u64 *end)
1046 {
1047         struct btrfs_file_extent_item *fi;
1048         struct btrfs_key key;
1049         u64 extent_end;
1050
1051         if (slot < 0 || slot >= btrfs_header_nritems(leaf))
1052                 return 0;
1053
1054         btrfs_item_key_to_cpu(leaf, &key, slot);
1055         if (key.objectid != objectid || key.type != BTRFS_EXTENT_DATA_KEY)
1056                 return 0;
1057
1058         fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
1059         if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG ||
1060             btrfs_file_extent_disk_bytenr(leaf, fi) != bytenr ||
1061             btrfs_file_extent_offset(leaf, fi) != key.offset - orig_offset ||
1062             btrfs_file_extent_compression(leaf, fi) ||
1063             btrfs_file_extent_encryption(leaf, fi) ||
1064             btrfs_file_extent_other_encoding(leaf, fi))
1065                 return 0;
1066
1067         extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
1068         if ((*start && *start != key.offset) || (*end && *end != extent_end))
1069                 return 0;
1070
1071         *start = key.offset;
1072         *end = extent_end;
1073         return 1;
1074 }
1075
1076 /*
1077  * Mark extent in the range start - end as written.
1078  *
1079  * This changes extent type from 'pre-allocated' to 'regular'. If only
1080  * part of extent is marked as written, the extent will be split into
1081  * two or three.
1082  */
1083 int btrfs_mark_extent_written(struct btrfs_trans_handle *trans,
1084                               struct btrfs_inode *inode, u64 start, u64 end)
1085 {
1086         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
1087         struct btrfs_root *root = inode->root;
1088         struct extent_buffer *leaf;
1089         struct btrfs_path *path;
1090         struct btrfs_file_extent_item *fi;
1091         struct btrfs_key key;
1092         struct btrfs_key new_key;
1093         u64 bytenr;
1094         u64 num_bytes;
1095         u64 extent_end;
1096         u64 orig_offset;
1097         u64 other_start;
1098         u64 other_end;
1099         u64 split;
1100         int del_nr = 0;
1101         int del_slot = 0;
1102         int recow;
1103         int ret = 0;
1104         u64 ino = btrfs_ino(inode);
1105
1106         path = btrfs_alloc_path();
1107         if (!path)
1108                 return -ENOMEM;
1109 again:
1110         recow = 0;
1111         split = start;
1112         key.objectid = ino;
1113         key.type = BTRFS_EXTENT_DATA_KEY;
1114         key.offset = split;
1115
1116         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1117         if (ret < 0)
1118                 goto out;
1119         if (ret > 0 && path->slots[0] > 0)
1120                 path->slots[0]--;
1121
1122         leaf = path->nodes[0];
1123         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1124         if (key.objectid != ino ||
1125             key.type != BTRFS_EXTENT_DATA_KEY) {
1126                 ret = -EINVAL;
1127                 btrfs_abort_transaction(trans, ret);
1128                 goto out;
1129         }
1130         fi = btrfs_item_ptr(leaf, path->slots[0],
1131                             struct btrfs_file_extent_item);
1132         if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_PREALLOC) {
1133                 ret = -EINVAL;
1134                 btrfs_abort_transaction(trans, ret);
1135                 goto out;
1136         }
1137         extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
1138         if (key.offset > start || extent_end < end) {
1139                 ret = -EINVAL;
1140                 btrfs_abort_transaction(trans, ret);
1141                 goto out;
1142         }
1143
1144         bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1145         num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
1146         orig_offset = key.offset - btrfs_file_extent_offset(leaf, fi);
1147         memcpy(&new_key, &key, sizeof(new_key));
1148
1149         if (start == key.offset && end < extent_end) {
1150                 other_start = 0;
1151                 other_end = start;
1152                 if (extent_mergeable(leaf, path->slots[0] - 1,
1153                                      ino, bytenr, orig_offset,
1154                                      &other_start, &other_end)) {
1155                         new_key.offset = end;
1156                         btrfs_set_item_key_safe(fs_info, path, &new_key);
1157                         fi = btrfs_item_ptr(leaf, path->slots[0],
1158                                             struct btrfs_file_extent_item);
1159                         btrfs_set_file_extent_generation(leaf, fi,
1160                                                          trans->transid);
1161                         btrfs_set_file_extent_num_bytes(leaf, fi,
1162                                                         extent_end - end);
1163                         btrfs_set_file_extent_offset(leaf, fi,
1164                                                      end - orig_offset);
1165                         fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
1166                                             struct btrfs_file_extent_item);
1167                         btrfs_set_file_extent_generation(leaf, fi,
1168                                                          trans->transid);
1169                         btrfs_set_file_extent_num_bytes(leaf, fi,
1170                                                         end - other_start);
1171                         btrfs_mark_buffer_dirty(leaf);
1172                         goto out;
1173                 }
1174         }
1175
1176         if (start > key.offset && end == extent_end) {
1177                 other_start = end;
1178                 other_end = 0;
1179                 if (extent_mergeable(leaf, path->slots[0] + 1,
1180                                      ino, bytenr, orig_offset,
1181                                      &other_start, &other_end)) {
1182                         fi = btrfs_item_ptr(leaf, path->slots[0],
1183                                             struct btrfs_file_extent_item);
1184                         btrfs_set_file_extent_num_bytes(leaf, fi,
1185                                                         start - key.offset);
1186                         btrfs_set_file_extent_generation(leaf, fi,
1187                                                          trans->transid);
1188                         path->slots[0]++;
1189                         new_key.offset = start;
1190                         btrfs_set_item_key_safe(fs_info, path, &new_key);
1191
1192                         fi = btrfs_item_ptr(leaf, path->slots[0],
1193                                             struct btrfs_file_extent_item);
1194                         btrfs_set_file_extent_generation(leaf, fi,
1195                                                          trans->transid);
1196                         btrfs_set_file_extent_num_bytes(leaf, fi,
1197                                                         other_end - start);
1198                         btrfs_set_file_extent_offset(leaf, fi,
1199                                                      start - orig_offset);
1200                         btrfs_mark_buffer_dirty(leaf);
1201                         goto out;
1202                 }
1203         }
1204
1205         while (start > key.offset || end < extent_end) {
1206                 if (key.offset == start)
1207                         split = end;
1208
1209                 new_key.offset = split;
1210                 ret = btrfs_duplicate_item(trans, root, path, &new_key);
1211                 if (ret == -EAGAIN) {
1212                         btrfs_release_path(path);
1213                         goto again;
1214                 }
1215                 if (ret < 0) {
1216                         btrfs_abort_transaction(trans, ret);
1217                         goto out;
1218                 }
1219
1220                 leaf = path->nodes[0];
1221                 fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
1222                                     struct btrfs_file_extent_item);
1223                 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1224                 btrfs_set_file_extent_num_bytes(leaf, fi,
1225                                                 split - key.offset);
1226
1227                 fi = btrfs_item_ptr(leaf, path->slots[0],
1228                                     struct btrfs_file_extent_item);
1229
1230                 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1231                 btrfs_set_file_extent_offset(leaf, fi, split - orig_offset);
1232                 btrfs_set_file_extent_num_bytes(leaf, fi,
1233                                                 extent_end - split);
1234                 btrfs_mark_buffer_dirty(leaf);
1235
1236                 ret = btrfs_inc_extent_ref(trans, fs_info, bytenr, num_bytes,
1237                                            0, root->root_key.objectid,
1238                                            ino, orig_offset);
1239                 if (ret) {
1240                         btrfs_abort_transaction(trans, ret);
1241                         goto out;
1242                 }
1243
1244                 if (split == start) {
1245                         key.offset = start;
1246                 } else {
1247                         if (start != key.offset) {
1248                                 ret = -EINVAL;
1249                                 btrfs_abort_transaction(trans, ret);
1250                                 goto out;
1251                         }
1252                         path->slots[0]--;
1253                         extent_end = end;
1254                 }
1255                 recow = 1;
1256         }
1257
1258         other_start = end;
1259         other_end = 0;
1260         if (extent_mergeable(leaf, path->slots[0] + 1,
1261                              ino, bytenr, orig_offset,
1262                              &other_start, &other_end)) {
1263                 if (recow) {
1264                         btrfs_release_path(path);
1265                         goto again;
1266                 }
1267                 extent_end = other_end;
1268                 del_slot = path->slots[0] + 1;
1269                 del_nr++;
1270                 ret = btrfs_free_extent(trans, fs_info, bytenr, num_bytes,
1271                                         0, root->root_key.objectid,
1272                                         ino, orig_offset);
1273                 if (ret) {
1274                         btrfs_abort_transaction(trans, ret);
1275                         goto out;
1276                 }
1277         }
1278         other_start = 0;
1279         other_end = start;
1280         if (extent_mergeable(leaf, path->slots[0] - 1,
1281                              ino, bytenr, orig_offset,
1282                              &other_start, &other_end)) {
1283                 if (recow) {
1284                         btrfs_release_path(path);
1285                         goto again;
1286                 }
1287                 key.offset = other_start;
1288                 del_slot = path->slots[0];
1289                 del_nr++;
1290                 ret = btrfs_free_extent(trans, fs_info, bytenr, num_bytes,
1291                                         0, root->root_key.objectid,
1292                                         ino, orig_offset);
1293                 if (ret) {
1294                         btrfs_abort_transaction(trans, ret);
1295                         goto out;
1296                 }
1297         }
1298         if (del_nr == 0) {
1299                 fi = btrfs_item_ptr(leaf, path->slots[0],
1300                            struct btrfs_file_extent_item);
1301                 btrfs_set_file_extent_type(leaf, fi,
1302                                            BTRFS_FILE_EXTENT_REG);
1303                 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1304                 btrfs_mark_buffer_dirty(leaf);
1305         } else {
1306                 fi = btrfs_item_ptr(leaf, del_slot - 1,
1307                            struct btrfs_file_extent_item);
1308                 btrfs_set_file_extent_type(leaf, fi,
1309                                            BTRFS_FILE_EXTENT_REG);
1310                 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1311                 btrfs_set_file_extent_num_bytes(leaf, fi,
1312                                                 extent_end - key.offset);
1313                 btrfs_mark_buffer_dirty(leaf);
1314
1315                 ret = btrfs_del_items(trans, root, path, del_slot, del_nr);
1316                 if (ret < 0) {
1317                         btrfs_abort_transaction(trans, ret);
1318                         goto out;
1319                 }
1320         }
1321 out:
1322         btrfs_free_path(path);
1323         return ret;
1324 }
1325
1326 /*
1327  * on error we return an unlocked page and the error value
1328  * on success we return a locked page and 0
1329  */
1330 static int prepare_uptodate_page(struct inode *inode,
1331                                  struct page *page, u64 pos,
1332                                  bool force_uptodate)
1333 {
1334         int ret = 0;
1335
1336         if (((pos & (PAGE_SIZE - 1)) || force_uptodate) &&
1337             !PageUptodate(page)) {
1338                 ret = btrfs_readpage(NULL, page);
1339                 if (ret)
1340                         return ret;
1341                 lock_page(page);
1342                 if (!PageUptodate(page)) {
1343                         unlock_page(page);
1344                         return -EIO;
1345                 }
1346                 if (page->mapping != inode->i_mapping) {
1347                         unlock_page(page);
1348                         return -EAGAIN;
1349                 }
1350         }
1351         return 0;
1352 }
1353
1354 /*
1355  * this just gets pages into the page cache and locks them down.
1356  */
1357 static noinline int prepare_pages(struct inode *inode, struct page **pages,
1358                                   size_t num_pages, loff_t pos,
1359                                   size_t write_bytes, bool force_uptodate)
1360 {
1361         int i;
1362         unsigned long index = pos >> PAGE_SHIFT;
1363         gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping);
1364         int err = 0;
1365         int faili;
1366
1367         for (i = 0; i < num_pages; i++) {
1368 again:
1369                 pages[i] = find_or_create_page(inode->i_mapping, index + i,
1370                                                mask | __GFP_WRITE);
1371                 if (!pages[i]) {
1372                         faili = i - 1;
1373                         err = -ENOMEM;
1374                         goto fail;
1375                 }
1376
1377                 if (i == 0)
1378                         err = prepare_uptodate_page(inode, pages[i], pos,
1379                                                     force_uptodate);
1380                 if (!err && i == num_pages - 1)
1381                         err = prepare_uptodate_page(inode, pages[i],
1382                                                     pos + write_bytes, false);
1383                 if (err) {
1384                         put_page(pages[i]);
1385                         if (err == -EAGAIN) {
1386                                 err = 0;
1387                                 goto again;
1388                         }
1389                         faili = i - 1;
1390                         goto fail;
1391                 }
1392                 wait_on_page_writeback(pages[i]);
1393         }
1394
1395         return 0;
1396 fail:
1397         while (faili >= 0) {
1398                 unlock_page(pages[faili]);
1399                 put_page(pages[faili]);
1400                 faili--;
1401         }
1402         return err;
1403
1404 }
1405
1406 static int btrfs_find_new_delalloc_bytes(struct btrfs_inode *inode,
1407                                          const u64 start,
1408                                          const u64 len,
1409                                          struct extent_state **cached_state)
1410 {
1411         u64 search_start = start;
1412         const u64 end = start + len - 1;
1413
1414         while (search_start < end) {
1415                 const u64 search_len = end - search_start + 1;
1416                 struct extent_map *em;
1417                 u64 em_len;
1418                 int ret = 0;
1419
1420                 em = btrfs_get_extent(inode, NULL, 0, search_start,
1421                                       search_len, 0);
1422                 if (IS_ERR(em))
1423                         return PTR_ERR(em);
1424
1425                 if (em->block_start != EXTENT_MAP_HOLE)
1426                         goto next;
1427
1428                 em_len = em->len;
1429                 if (em->start < search_start)
1430                         em_len -= search_start - em->start;
1431                 if (em_len > search_len)
1432                         em_len = search_len;
1433
1434                 ret = set_extent_bit(&inode->io_tree, search_start,
1435                                      search_start + em_len - 1,
1436                                      EXTENT_DELALLOC_NEW,
1437                                      NULL, cached_state, GFP_NOFS);
1438 next:
1439                 search_start = extent_map_end(em);
1440                 free_extent_map(em);
1441                 if (ret)
1442                         return ret;
1443         }
1444         return 0;
1445 }
1446
1447 /*
1448  * This function locks the extent and properly waits for data=ordered extents
1449  * to finish before allowing the pages to be modified if need.
1450  *
1451  * The return value:
1452  * 1 - the extent is locked
1453  * 0 - the extent is not locked, and everything is OK
1454  * -EAGAIN - need re-prepare the pages
1455  * the other < 0 number - Something wrong happens
1456  */
1457 static noinline int
1458 lock_and_cleanup_extent_if_need(struct btrfs_inode *inode, struct page **pages,
1459                                 size_t num_pages, loff_t pos,
1460                                 size_t write_bytes,
1461                                 u64 *lockstart, u64 *lockend,
1462                                 struct extent_state **cached_state)
1463 {
1464         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
1465         u64 start_pos;
1466         u64 last_pos;
1467         int i;
1468         int ret = 0;
1469
1470         start_pos = round_down(pos, fs_info->sectorsize);
1471         last_pos = start_pos
1472                 + round_up(pos + write_bytes - start_pos,
1473                            fs_info->sectorsize) - 1;
1474
1475         if (start_pos < inode->vfs_inode.i_size ||
1476             (inode->flags & BTRFS_INODE_PREALLOC)) {
1477                 struct btrfs_ordered_extent *ordered;
1478                 unsigned int clear_bits;
1479
1480                 lock_extent_bits(&inode->io_tree, start_pos, last_pos,
1481                                 cached_state);
1482                 ordered = btrfs_lookup_ordered_range(inode, start_pos,
1483                                                      last_pos - start_pos + 1);
1484                 if (ordered &&
1485                     ordered->file_offset + ordered->len > start_pos &&
1486                     ordered->file_offset <= last_pos) {
1487                         unlock_extent_cached(&inode->io_tree, start_pos,
1488                                         last_pos, cached_state, GFP_NOFS);
1489                         for (i = 0; i < num_pages; i++) {
1490                                 unlock_page(pages[i]);
1491                                 put_page(pages[i]);
1492                         }
1493                         btrfs_start_ordered_extent(&inode->vfs_inode,
1494                                         ordered, 1);
1495                         btrfs_put_ordered_extent(ordered);
1496                         return -EAGAIN;
1497                 }
1498                 if (ordered)
1499                         btrfs_put_ordered_extent(ordered);
1500                 ret = btrfs_find_new_delalloc_bytes(inode, start_pos,
1501                                                     last_pos - start_pos + 1,
1502                                                     cached_state);
1503                 clear_bits = EXTENT_DIRTY | EXTENT_DELALLOC |
1504                         EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG;
1505                 if (ret)
1506                         clear_bits |= EXTENT_DELALLOC_NEW | EXTENT_LOCKED;
1507                 clear_extent_bit(&inode->io_tree, start_pos,
1508                                  last_pos, clear_bits,
1509                                  (clear_bits & EXTENT_LOCKED) ? 1 : 0,
1510                                  0, cached_state, GFP_NOFS);
1511                 if (ret)
1512                         return ret;
1513                 *lockstart = start_pos;
1514                 *lockend = last_pos;
1515                 ret = 1;
1516         }
1517
1518         for (i = 0; i < num_pages; i++) {
1519                 if (clear_page_dirty_for_io(pages[i]))
1520                         account_page_redirty(pages[i]);
1521                 set_page_extent_mapped(pages[i]);
1522                 WARN_ON(!PageLocked(pages[i]));
1523         }
1524
1525         return ret;
1526 }
1527
1528 static noinline int check_can_nocow(struct btrfs_inode *inode, loff_t pos,
1529                                     size_t *write_bytes)
1530 {
1531         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
1532         struct btrfs_root *root = inode->root;
1533         struct btrfs_ordered_extent *ordered;
1534         u64 lockstart, lockend;
1535         u64 num_bytes;
1536         int ret;
1537
1538         ret = btrfs_start_write_no_snapshotting(root);
1539         if (!ret)
1540                 return -ENOSPC;
1541
1542         lockstart = round_down(pos, fs_info->sectorsize);
1543         lockend = round_up(pos + *write_bytes,
1544                            fs_info->sectorsize) - 1;
1545
1546         while (1) {
1547                 lock_extent(&inode->io_tree, lockstart, lockend);
1548                 ordered = btrfs_lookup_ordered_range(inode, lockstart,
1549                                                      lockend - lockstart + 1);
1550                 if (!ordered) {
1551                         break;
1552                 }
1553                 unlock_extent(&inode->io_tree, lockstart, lockend);
1554                 btrfs_start_ordered_extent(&inode->vfs_inode, ordered, 1);
1555                 btrfs_put_ordered_extent(ordered);
1556         }
1557
1558         num_bytes = lockend - lockstart + 1;
1559         ret = can_nocow_extent(&inode->vfs_inode, lockstart, &num_bytes,
1560                         NULL, NULL, NULL);
1561         if (ret <= 0) {
1562                 ret = 0;
1563                 btrfs_end_write_no_snapshotting(root);
1564         } else {
1565                 *write_bytes = min_t(size_t, *write_bytes ,
1566                                      num_bytes - pos + lockstart);
1567         }
1568
1569         unlock_extent(&inode->io_tree, lockstart, lockend);
1570
1571         return ret;
1572 }
1573
1574 static noinline ssize_t __btrfs_buffered_write(struct file *file,
1575                                                struct iov_iter *i,
1576                                                loff_t pos)
1577 {
1578         struct inode *inode = file_inode(file);
1579         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1580         struct btrfs_root *root = BTRFS_I(inode)->root;
1581         struct page **pages = NULL;
1582         struct extent_state *cached_state = NULL;
1583         struct extent_changeset *data_reserved = NULL;
1584         u64 release_bytes = 0;
1585         u64 lockstart;
1586         u64 lockend;
1587         size_t num_written = 0;
1588         int nrptrs;
1589         int ret = 0;
1590         bool only_release_metadata = false;
1591         bool force_page_uptodate = false;
1592         bool need_unlock;
1593
1594         nrptrs = min(DIV_ROUND_UP(iov_iter_count(i), PAGE_SIZE),
1595                         PAGE_SIZE / (sizeof(struct page *)));
1596         nrptrs = min(nrptrs, current->nr_dirtied_pause - current->nr_dirtied);
1597         nrptrs = max(nrptrs, 8);
1598         pages = kmalloc_array(nrptrs, sizeof(struct page *), GFP_KERNEL);
1599         if (!pages)
1600                 return -ENOMEM;
1601
1602         while (iov_iter_count(i) > 0) {
1603                 size_t offset = pos & (PAGE_SIZE - 1);
1604                 size_t sector_offset;
1605                 size_t write_bytes = min(iov_iter_count(i),
1606                                          nrptrs * (size_t)PAGE_SIZE -
1607                                          offset);
1608                 size_t num_pages = DIV_ROUND_UP(write_bytes + offset,
1609                                                 PAGE_SIZE);
1610                 size_t reserve_bytes;
1611                 size_t dirty_pages;
1612                 size_t copied;
1613                 size_t dirty_sectors;
1614                 size_t num_sectors;
1615
1616                 WARN_ON(num_pages > nrptrs);
1617
1618                 /*
1619                  * Fault pages before locking them in prepare_pages
1620                  * to avoid recursive lock
1621                  */
1622                 if (unlikely(iov_iter_fault_in_readable(i, write_bytes))) {
1623                         ret = -EFAULT;
1624                         break;
1625                 }
1626
1627                 only_release_metadata = false;
1628                 sector_offset = pos & (fs_info->sectorsize - 1);
1629                 reserve_bytes = round_up(write_bytes + sector_offset,
1630                                 fs_info->sectorsize);
1631
1632                 extent_changeset_release(data_reserved);
1633                 ret = btrfs_check_data_free_space(inode, &data_reserved, pos,
1634                                                   write_bytes);
1635                 if (ret < 0) {
1636                         if ((BTRFS_I(inode)->flags & (BTRFS_INODE_NODATACOW |
1637                                                       BTRFS_INODE_PREALLOC)) &&
1638                             check_can_nocow(BTRFS_I(inode), pos,
1639                                         &write_bytes) > 0) {
1640                                 /*
1641                                  * For nodata cow case, no need to reserve
1642                                  * data space.
1643                                  */
1644                                 only_release_metadata = true;
1645                                 /*
1646                                  * our prealloc extent may be smaller than
1647                                  * write_bytes, so scale down.
1648                                  */
1649                                 num_pages = DIV_ROUND_UP(write_bytes + offset,
1650                                                          PAGE_SIZE);
1651                                 reserve_bytes = round_up(write_bytes +
1652                                                          sector_offset,
1653                                                          fs_info->sectorsize);
1654                         } else {
1655                                 break;
1656                         }
1657                 }
1658
1659                 ret = btrfs_delalloc_reserve_metadata(BTRFS_I(inode),
1660                                 reserve_bytes);
1661                 if (ret) {
1662                         if (!only_release_metadata)
1663                                 btrfs_free_reserved_data_space(inode,
1664                                                 data_reserved, pos,
1665                                                 write_bytes);
1666                         else
1667                                 btrfs_end_write_no_snapshotting(root);
1668                         break;
1669                 }
1670
1671                 release_bytes = reserve_bytes;
1672                 need_unlock = false;
1673 again:
1674                 /*
1675                  * This is going to setup the pages array with the number of
1676                  * pages we want, so we don't really need to worry about the
1677                  * contents of pages from loop to loop
1678                  */
1679                 ret = prepare_pages(inode, pages, num_pages,
1680                                     pos, write_bytes,
1681                                     force_page_uptodate);
1682                 if (ret)
1683                         break;
1684
1685                 ret = lock_and_cleanup_extent_if_need(BTRFS_I(inode), pages,
1686                                 num_pages, pos, write_bytes, &lockstart,
1687                                 &lockend, &cached_state);
1688                 if (ret < 0) {
1689                         if (ret == -EAGAIN)
1690                                 goto again;
1691                         break;
1692                 } else if (ret > 0) {
1693                         need_unlock = true;
1694                         ret = 0;
1695                 }
1696
1697                 copied = btrfs_copy_from_user(pos, write_bytes, pages, i);
1698
1699                 num_sectors = BTRFS_BYTES_TO_BLKS(fs_info, reserve_bytes);
1700                 dirty_sectors = round_up(copied + sector_offset,
1701                                         fs_info->sectorsize);
1702                 dirty_sectors = BTRFS_BYTES_TO_BLKS(fs_info, dirty_sectors);
1703
1704                 /*
1705                  * if we have trouble faulting in the pages, fall
1706                  * back to one page at a time
1707                  */
1708                 if (copied < write_bytes)
1709                         nrptrs = 1;
1710
1711                 if (copied == 0) {
1712                         force_page_uptodate = true;
1713                         dirty_sectors = 0;
1714                         dirty_pages = 0;
1715                 } else {
1716                         force_page_uptodate = false;
1717                         dirty_pages = DIV_ROUND_UP(copied + offset,
1718                                                    PAGE_SIZE);
1719                 }
1720
1721                 /*
1722                  * If we had a short copy we need to release the excess delaloc
1723                  * bytes we reserved.  We need to increment outstanding_extents
1724                  * because btrfs_delalloc_release_space and
1725                  * btrfs_delalloc_release_metadata will decrement it, but
1726                  * we still have an outstanding extent for the chunk we actually
1727                  * managed to copy.
1728                  */
1729                 if (num_sectors > dirty_sectors) {
1730                         /* release everything except the sectors we dirtied */
1731                         release_bytes -= dirty_sectors <<
1732                                                 fs_info->sb->s_blocksize_bits;
1733                         if (copied > 0) {
1734                                 spin_lock(&BTRFS_I(inode)->lock);
1735                                 BTRFS_I(inode)->outstanding_extents++;
1736                                 spin_unlock(&BTRFS_I(inode)->lock);
1737                         }
1738                         if (only_release_metadata) {
1739                                 btrfs_delalloc_release_metadata(BTRFS_I(inode),
1740                                                                 release_bytes);
1741                         } else {
1742                                 u64 __pos;
1743
1744                                 __pos = round_down(pos,
1745                                                    fs_info->sectorsize) +
1746                                         (dirty_pages << PAGE_SHIFT);
1747                                 btrfs_delalloc_release_space(inode,
1748                                                 data_reserved, __pos,
1749                                                 release_bytes);
1750                         }
1751                 }
1752
1753                 release_bytes = round_up(copied + sector_offset,
1754                                         fs_info->sectorsize);
1755
1756                 if (copied > 0)
1757                         ret = btrfs_dirty_pages(inode, pages, dirty_pages,
1758                                                 pos, copied, NULL);
1759                 if (need_unlock)
1760                         unlock_extent_cached(&BTRFS_I(inode)->io_tree,
1761                                              lockstart, lockend, &cached_state,
1762                                              GFP_NOFS);
1763                 if (ret) {
1764                         btrfs_drop_pages(pages, num_pages);
1765                         break;
1766                 }
1767
1768                 release_bytes = 0;
1769                 if (only_release_metadata)
1770                         btrfs_end_write_no_snapshotting(root);
1771
1772                 if (only_release_metadata && copied > 0) {
1773                         lockstart = round_down(pos,
1774                                                fs_info->sectorsize);
1775                         lockend = round_up(pos + copied,
1776                                            fs_info->sectorsize) - 1;
1777
1778                         set_extent_bit(&BTRFS_I(inode)->io_tree, lockstart,
1779                                        lockend, EXTENT_NORESERVE, NULL,
1780                                        NULL, GFP_NOFS);
1781                 }
1782
1783                 btrfs_drop_pages(pages, num_pages);
1784
1785                 cond_resched();
1786
1787                 balance_dirty_pages_ratelimited(inode->i_mapping);
1788                 if (dirty_pages < (fs_info->nodesize >> PAGE_SHIFT) + 1)
1789                         btrfs_btree_balance_dirty(fs_info);
1790
1791                 pos += copied;
1792                 num_written += copied;
1793         }
1794
1795         kfree(pages);
1796
1797         if (release_bytes) {
1798                 if (only_release_metadata) {
1799                         btrfs_end_write_no_snapshotting(root);
1800                         btrfs_delalloc_release_metadata(BTRFS_I(inode),
1801                                         release_bytes);
1802                 } else {
1803                         btrfs_delalloc_release_space(inode, data_reserved,
1804                                         round_down(pos, fs_info->sectorsize),
1805                                         release_bytes);
1806                 }
1807         }
1808
1809         extent_changeset_free(data_reserved);
1810         return num_written ? num_written : ret;
1811 }
1812
1813 static ssize_t __btrfs_direct_write(struct kiocb *iocb, struct iov_iter *from)
1814 {
1815         struct file *file = iocb->ki_filp;
1816         struct inode *inode = file_inode(file);
1817         loff_t pos = iocb->ki_pos;
1818         ssize_t written;
1819         ssize_t written_buffered;
1820         loff_t endbyte;
1821         int err;
1822
1823         written = generic_file_direct_write(iocb, from);
1824
1825         if (written < 0 || !iov_iter_count(from))
1826                 return written;
1827
1828         pos += written;
1829         written_buffered = __btrfs_buffered_write(file, from, pos);
1830         if (written_buffered < 0) {
1831                 err = written_buffered;
1832                 goto out;
1833         }
1834         /*
1835          * Ensure all data is persisted. We want the next direct IO read to be
1836          * able to read what was just written.
1837          */
1838         endbyte = pos + written_buffered - 1;
1839         err = btrfs_fdatawrite_range(inode, pos, endbyte);
1840         if (err)
1841                 goto out;
1842         err = filemap_fdatawait_range(inode->i_mapping, pos, endbyte);
1843         if (err)
1844                 goto out;
1845         written += written_buffered;
1846         iocb->ki_pos = pos + written_buffered;
1847         invalidate_mapping_pages(file->f_mapping, pos >> PAGE_SHIFT,
1848                                  endbyte >> PAGE_SHIFT);
1849 out:
1850         return written ? written : err;
1851 }
1852
1853 static void update_time_for_write(struct inode *inode)
1854 {
1855         struct timespec now;
1856
1857         if (IS_NOCMTIME(inode))
1858                 return;
1859
1860         now = current_time(inode);
1861         if (!timespec_equal(&inode->i_mtime, &now))
1862                 inode->i_mtime = now;
1863
1864         if (!timespec_equal(&inode->i_ctime, &now))
1865                 inode->i_ctime = now;
1866
1867         if (IS_I_VERSION(inode))
1868                 inode_inc_iversion(inode);
1869 }
1870
1871 static ssize_t btrfs_file_write_iter(struct kiocb *iocb,
1872                                     struct iov_iter *from)
1873 {
1874         struct file *file = iocb->ki_filp;
1875         struct inode *inode = file_inode(file);
1876         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1877         struct btrfs_root *root = BTRFS_I(inode)->root;
1878         u64 start_pos;
1879         u64 end_pos;
1880         ssize_t num_written = 0;
1881         bool sync = (file->f_flags & O_DSYNC) || IS_SYNC(file->f_mapping->host);
1882         ssize_t err;
1883         loff_t pos;
1884         size_t count;
1885         loff_t oldsize;
1886         int clean_page = 0;
1887
1888         if (!(iocb->ki_flags & IOCB_DIRECT) &&
1889             (iocb->ki_flags & IOCB_NOWAIT))
1890                 return -EOPNOTSUPP;
1891
1892         if (iocb->ki_flags & IOCB_NOWAIT) {
1893                 if (!inode_trylock(inode))
1894                         return -EAGAIN;
1895         } else {
1896                 inode_lock(inode);
1897         }
1898
1899         err = generic_write_checks(iocb, from);
1900         if (err <= 0) {
1901                 inode_unlock(inode);
1902                 return err;
1903         }
1904
1905         pos = iocb->ki_pos;
1906         count = iov_iter_count(from);
1907         if (iocb->ki_flags & IOCB_NOWAIT) {
1908                 /*
1909                  * We will allocate space in case nodatacow is not set,
1910                  * so bail
1911                  */
1912                 if (!(BTRFS_I(inode)->flags & (BTRFS_INODE_NODATACOW |
1913                                               BTRFS_INODE_PREALLOC)) ||
1914                     check_can_nocow(BTRFS_I(inode), pos, &count) <= 0) {
1915                         inode_unlock(inode);
1916                         return -EAGAIN;
1917                 }
1918         }
1919
1920         current->backing_dev_info = inode_to_bdi(inode);
1921         err = file_remove_privs(file);
1922         if (err) {
1923                 inode_unlock(inode);
1924                 goto out;
1925         }
1926
1927         /*
1928          * If BTRFS flips readonly due to some impossible error
1929          * (fs_info->fs_state now has BTRFS_SUPER_FLAG_ERROR),
1930          * although we have opened a file as writable, we have
1931          * to stop this write operation to ensure FS consistency.
1932          */
1933         if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state)) {
1934                 inode_unlock(inode);
1935                 err = -EROFS;
1936                 goto out;
1937         }
1938
1939         /*
1940          * We reserve space for updating the inode when we reserve space for the
1941          * extent we are going to write, so we will enospc out there.  We don't
1942          * need to start yet another transaction to update the inode as we will
1943          * update the inode when we finish writing whatever data we write.
1944          */
1945         update_time_for_write(inode);
1946
1947         start_pos = round_down(pos, fs_info->sectorsize);
1948         oldsize = i_size_read(inode);
1949         if (start_pos > oldsize) {
1950                 /* Expand hole size to cover write data, preventing empty gap */
1951                 end_pos = round_up(pos + count,
1952                                    fs_info->sectorsize);
1953                 err = btrfs_cont_expand(inode, oldsize, end_pos);
1954                 if (err) {
1955                         inode_unlock(inode);
1956                         goto out;
1957                 }
1958                 if (start_pos > round_up(oldsize, fs_info->sectorsize))
1959                         clean_page = 1;
1960         }
1961
1962         if (sync)
1963                 atomic_inc(&BTRFS_I(inode)->sync_writers);
1964
1965         if (iocb->ki_flags & IOCB_DIRECT) {
1966                 num_written = __btrfs_direct_write(iocb, from);
1967         } else {
1968                 num_written = __btrfs_buffered_write(file, from, pos);
1969                 if (num_written > 0)
1970                         iocb->ki_pos = pos + num_written;
1971                 if (clean_page)
1972                         pagecache_isize_extended(inode, oldsize,
1973                                                 i_size_read(inode));
1974         }
1975
1976         inode_unlock(inode);
1977
1978         /*
1979          * We also have to set last_sub_trans to the current log transid,
1980          * otherwise subsequent syncs to a file that's been synced in this
1981          * transaction will appear to have already occurred.
1982          */
1983         spin_lock(&BTRFS_I(inode)->lock);
1984         BTRFS_I(inode)->last_sub_trans = root->log_transid;
1985         spin_unlock(&BTRFS_I(inode)->lock);
1986         if (num_written > 0)
1987                 num_written = generic_write_sync(iocb, num_written);
1988
1989         if (sync)
1990                 atomic_dec(&BTRFS_I(inode)->sync_writers);
1991 out:
1992         current->backing_dev_info = NULL;
1993         return num_written ? num_written : err;
1994 }
1995
1996 int btrfs_release_file(struct inode *inode, struct file *filp)
1997 {
1998         struct btrfs_file_private *private = filp->private_data;
1999
2000         if (private && private->trans)
2001                 btrfs_ioctl_trans_end(filp);
2002         if (private && private->filldir_buf)
2003                 kfree(private->filldir_buf);
2004         kfree(private);
2005         filp->private_data = NULL;
2006
2007         /*
2008          * ordered_data_close is set by settattr when we are about to truncate
2009          * a file from a non-zero size to a zero size.  This tries to
2010          * flush down new bytes that may have been written if the
2011          * application were using truncate to replace a file in place.
2012          */
2013         if (test_and_clear_bit(BTRFS_INODE_ORDERED_DATA_CLOSE,
2014                                &BTRFS_I(inode)->runtime_flags))
2015                         filemap_flush(inode->i_mapping);
2016         return 0;
2017 }
2018
2019 static int start_ordered_ops(struct inode *inode, loff_t start, loff_t end)
2020 {
2021         int ret;
2022         struct blk_plug plug;
2023
2024         /*
2025          * This is only called in fsync, which would do synchronous writes, so
2026          * a plug can merge adjacent IOs as much as possible.  Esp. in case of
2027          * multiple disks using raid profile, a large IO can be split to
2028          * several segments of stripe length (currently 64K).
2029          */
2030         blk_start_plug(&plug);
2031         atomic_inc(&BTRFS_I(inode)->sync_writers);
2032         ret = btrfs_fdatawrite_range(inode, start, end);
2033         atomic_dec(&BTRFS_I(inode)->sync_writers);
2034         blk_finish_plug(&plug);
2035
2036         return ret;
2037 }
2038
2039 /*
2040  * fsync call for both files and directories.  This logs the inode into
2041  * the tree log instead of forcing full commits whenever possible.
2042  *
2043  * It needs to call filemap_fdatawait so that all ordered extent updates are
2044  * in the metadata btree are up to date for copying to the log.
2045  *
2046  * It drops the inode mutex before doing the tree log commit.  This is an
2047  * important optimization for directories because holding the mutex prevents
2048  * new operations on the dir while we write to disk.
2049  */
2050 int btrfs_sync_file(struct file *file, loff_t start, loff_t end, int datasync)
2051 {
2052         struct dentry *dentry = file_dentry(file);
2053         struct inode *inode = d_inode(dentry);
2054         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2055         struct btrfs_root *root = BTRFS_I(inode)->root;
2056         struct btrfs_trans_handle *trans;
2057         struct btrfs_log_ctx ctx;
2058         int ret = 0, err;
2059         bool full_sync = 0;
2060         u64 len;
2061
2062         /*
2063          * If the inode needs a full sync, make sure we use a full range to
2064          * avoid log tree corruption, due to hole detection racing with ordered
2065          * extent completion for adjacent ranges, and assertion failures during
2066          * hole detection.
2067          */
2068         if (test_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2069                      &BTRFS_I(inode)->runtime_flags)) {
2070                 start = 0;
2071                 end = LLONG_MAX;
2072         }
2073
2074         /*
2075          * The range length can be represented by u64, we have to do the typecasts
2076          * to avoid signed overflow if it's [0, LLONG_MAX] eg. from fsync()
2077          */
2078         len = (u64)end - (u64)start + 1;
2079         trace_btrfs_sync_file(file, datasync);
2080
2081         btrfs_init_log_ctx(&ctx, inode);
2082
2083         /*
2084          * We write the dirty pages in the range and wait until they complete
2085          * out of the ->i_mutex. If so, we can flush the dirty pages by
2086          * multi-task, and make the performance up.  See
2087          * btrfs_wait_ordered_range for an explanation of the ASYNC check.
2088          */
2089         ret = start_ordered_ops(inode, start, end);
2090         if (ret)
2091                 goto out;
2092
2093         inode_lock(inode);
2094
2095         /*
2096          * We take the dio_sem here because the tree log stuff can race with
2097          * lockless dio writes and get an extent map logged for an extent we
2098          * never waited on.  We need it this high up for lockdep reasons.
2099          */
2100         down_write(&BTRFS_I(inode)->dio_sem);
2101
2102         atomic_inc(&root->log_batch);
2103         full_sync = test_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2104                              &BTRFS_I(inode)->runtime_flags);
2105
2106         /*
2107          * We have to do this here to avoid the priority inversion of waiting on
2108          * IO of a lower priority task while holding a transaciton open.
2109          */
2110         ret = btrfs_wait_ordered_range(inode, start, len);
2111         if (ret) {
2112                 up_write(&BTRFS_I(inode)->dio_sem);
2113                 inode_unlock(inode);
2114                 goto out;
2115         }
2116         atomic_inc(&root->log_batch);
2117
2118         /*
2119          * If the last transaction that changed this file was before the current
2120          * transaction and we have the full sync flag set in our inode, we can
2121          * bail out now without any syncing.
2122          *
2123          * Note that we can't bail out if the full sync flag isn't set. This is
2124          * because when the full sync flag is set we start all ordered extents
2125          * and wait for them to fully complete - when they complete they update
2126          * the inode's last_trans field through:
2127          *
2128          *     btrfs_finish_ordered_io() ->
2129          *         btrfs_update_inode_fallback() ->
2130          *             btrfs_update_inode() ->
2131          *                 btrfs_set_inode_last_trans()
2132          *
2133          * So we are sure that last_trans is up to date and can do this check to
2134          * bail out safely. For the fast path, when the full sync flag is not
2135          * set in our inode, we can not do it because we start only our ordered
2136          * extents and don't wait for them to complete (that is when
2137          * btrfs_finish_ordered_io runs), so here at this point their last_trans
2138          * value might be less than or equals to fs_info->last_trans_committed,
2139          * and setting a speculative last_trans for an inode when a buffered
2140          * write is made (such as fs_info->generation + 1 for example) would not
2141          * be reliable since after setting the value and before fsync is called
2142          * any number of transactions can start and commit (transaction kthread
2143          * commits the current transaction periodically), and a transaction
2144          * commit does not start nor waits for ordered extents to complete.
2145          */
2146         smp_mb();
2147         if (btrfs_inode_in_log(BTRFS_I(inode), fs_info->generation) ||
2148             (full_sync && BTRFS_I(inode)->last_trans <=
2149              fs_info->last_trans_committed) ||
2150             (!btrfs_have_ordered_extents_in_range(inode, start, len) &&
2151              BTRFS_I(inode)->last_trans
2152              <= fs_info->last_trans_committed)) {
2153                 /*
2154                  * We've had everything committed since the last time we were
2155                  * modified so clear this flag in case it was set for whatever
2156                  * reason, it's no longer relevant.
2157                  */
2158                 clear_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2159                           &BTRFS_I(inode)->runtime_flags);
2160                 /*
2161                  * An ordered extent might have started before and completed
2162                  * already with io errors, in which case the inode was not
2163                  * updated and we end up here. So check the inode's mapping
2164                  * for any errors that might have happened since we last
2165                  * checked called fsync.
2166                  */
2167                 ret = filemap_check_wb_err(inode->i_mapping, file->f_wb_err);
2168                 up_write(&BTRFS_I(inode)->dio_sem);
2169                 inode_unlock(inode);
2170                 goto out;
2171         }
2172
2173         /*
2174          * ok we haven't committed the transaction yet, lets do a commit
2175          */
2176         if (file->private_data)
2177                 btrfs_ioctl_trans_end(file);
2178
2179         /*
2180          * We use start here because we will need to wait on the IO to complete
2181          * in btrfs_sync_log, which could require joining a transaction (for
2182          * example checking cross references in the nocow path).  If we use join
2183          * here we could get into a situation where we're waiting on IO to
2184          * happen that is blocked on a transaction trying to commit.  With start
2185          * we inc the extwriter counter, so we wait for all extwriters to exit
2186          * before we start blocking join'ers.  This comment is to keep somebody
2187          * from thinking they are super smart and changing this to
2188          * btrfs_join_transaction *cough*Josef*cough*.
2189          */
2190         trans = btrfs_start_transaction(root, 0);
2191         if (IS_ERR(trans)) {
2192                 ret = PTR_ERR(trans);
2193                 up_write(&BTRFS_I(inode)->dio_sem);
2194                 inode_unlock(inode);
2195                 goto out;
2196         }
2197         trans->sync = true;
2198
2199         ret = btrfs_log_dentry_safe(trans, root, dentry, start, end, &ctx);
2200         if (ret < 0) {
2201                 /* Fallthrough and commit/free transaction. */
2202                 ret = 1;
2203         }
2204
2205         /* we've logged all the items and now have a consistent
2206          * version of the file in the log.  It is possible that
2207          * someone will come in and modify the file, but that's
2208          * fine because the log is consistent on disk, and we
2209          * have references to all of the file's extents
2210          *
2211          * It is possible that someone will come in and log the
2212          * file again, but that will end up using the synchronization
2213          * inside btrfs_sync_log to keep things safe.
2214          */
2215         up_write(&BTRFS_I(inode)->dio_sem);
2216         inode_unlock(inode);
2217
2218         /*
2219          * If any of the ordered extents had an error, just return it to user
2220          * space, so that the application knows some writes didn't succeed and
2221          * can take proper action (retry for e.g.). Blindly committing the
2222          * transaction in this case, would fool userspace that everything was
2223          * successful. And we also want to make sure our log doesn't contain
2224          * file extent items pointing to extents that weren't fully written to -
2225          * just like in the non fast fsync path, where we check for the ordered
2226          * operation's error flag before writing to the log tree and return -EIO
2227          * if any of them had this flag set (btrfs_wait_ordered_range) -
2228          * therefore we need to check for errors in the ordered operations,
2229          * which are indicated by ctx.io_err.
2230          */
2231         if (ctx.io_err) {
2232                 btrfs_end_transaction(trans);
2233                 ret = ctx.io_err;
2234                 goto out;
2235         }
2236
2237         if (ret != BTRFS_NO_LOG_SYNC) {
2238                 if (!ret) {
2239                         ret = btrfs_sync_log(trans, root, &ctx);
2240                         if (!ret) {
2241                                 ret = btrfs_end_transaction(trans);
2242                                 goto out;
2243                         }
2244                 }
2245                 ret = btrfs_commit_transaction(trans);
2246         } else {
2247                 ret = btrfs_end_transaction(trans);
2248         }
2249 out:
2250         ASSERT(list_empty(&ctx.list));
2251         err = file_check_and_advance_wb_err(file);
2252         if (!ret)
2253                 ret = err;
2254         return ret > 0 ? -EIO : ret;
2255 }
2256
2257 static const struct vm_operations_struct btrfs_file_vm_ops = {
2258         .fault          = filemap_fault,
2259         .map_pages      = filemap_map_pages,
2260         .page_mkwrite   = btrfs_page_mkwrite,
2261 };
2262
2263 static int btrfs_file_mmap(struct file  *filp, struct vm_area_struct *vma)
2264 {
2265         struct address_space *mapping = filp->f_mapping;
2266
2267         if (!mapping->a_ops->readpage)
2268                 return -ENOEXEC;
2269
2270         file_accessed(filp);
2271         vma->vm_ops = &btrfs_file_vm_ops;
2272
2273         return 0;
2274 }
2275
2276 static int hole_mergeable(struct btrfs_inode *inode, struct extent_buffer *leaf,
2277                           int slot, u64 start, u64 end)
2278 {
2279         struct btrfs_file_extent_item *fi;
2280         struct btrfs_key key;
2281
2282         if (slot < 0 || slot >= btrfs_header_nritems(leaf))
2283                 return 0;
2284
2285         btrfs_item_key_to_cpu(leaf, &key, slot);
2286         if (key.objectid != btrfs_ino(inode) ||
2287             key.type != BTRFS_EXTENT_DATA_KEY)
2288                 return 0;
2289
2290         fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
2291
2292         if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG)
2293                 return 0;
2294
2295         if (btrfs_file_extent_disk_bytenr(leaf, fi))
2296                 return 0;
2297
2298         if (key.offset == end)
2299                 return 1;
2300         if (key.offset + btrfs_file_extent_num_bytes(leaf, fi) == start)
2301                 return 1;
2302         return 0;
2303 }
2304
2305 static int fill_holes(struct btrfs_trans_handle *trans,
2306                 struct btrfs_inode *inode,
2307                 struct btrfs_path *path, u64 offset, u64 end)
2308 {
2309         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
2310         struct btrfs_root *root = inode->root;
2311         struct extent_buffer *leaf;
2312         struct btrfs_file_extent_item *fi;
2313         struct extent_map *hole_em;
2314         struct extent_map_tree *em_tree = &inode->extent_tree;
2315         struct btrfs_key key;
2316         int ret;
2317
2318         if (btrfs_fs_incompat(fs_info, NO_HOLES))
2319                 goto out;
2320
2321         key.objectid = btrfs_ino(inode);
2322         key.type = BTRFS_EXTENT_DATA_KEY;
2323         key.offset = offset;
2324
2325         ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
2326         if (ret <= 0) {
2327                 /*
2328                  * We should have dropped this offset, so if we find it then
2329                  * something has gone horribly wrong.
2330                  */
2331                 if (ret == 0)
2332                         ret = -EINVAL;
2333                 return ret;
2334         }
2335
2336         leaf = path->nodes[0];
2337         if (hole_mergeable(inode, leaf, path->slots[0] - 1, offset, end)) {
2338                 u64 num_bytes;
2339
2340                 path->slots[0]--;
2341                 fi = btrfs_item_ptr(leaf, path->slots[0],
2342                                     struct btrfs_file_extent_item);
2343                 num_bytes = btrfs_file_extent_num_bytes(leaf, fi) +
2344                         end - offset;
2345                 btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2346                 btrfs_set_file_extent_ram_bytes(leaf, fi, num_bytes);
2347                 btrfs_set_file_extent_offset(leaf, fi, 0);
2348                 btrfs_mark_buffer_dirty(leaf);
2349                 goto out;
2350         }
2351
2352         if (hole_mergeable(inode, leaf, path->slots[0], offset, end)) {
2353                 u64 num_bytes;
2354
2355                 key.offset = offset;
2356                 btrfs_set_item_key_safe(fs_info, path, &key);
2357                 fi = btrfs_item_ptr(leaf, path->slots[0],
2358                                     struct btrfs_file_extent_item);
2359                 num_bytes = btrfs_file_extent_num_bytes(leaf, fi) + end -
2360                         offset;
2361                 btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2362                 btrfs_set_file_extent_ram_bytes(leaf, fi, num_bytes);
2363                 btrfs_set_file_extent_offset(leaf, fi, 0);
2364                 btrfs_mark_buffer_dirty(leaf);
2365                 goto out;
2366         }
2367         btrfs_release_path(path);
2368
2369         ret = btrfs_insert_file_extent(trans, root, btrfs_ino(inode),
2370                         offset, 0, 0, end - offset, 0, end - offset, 0, 0, 0);
2371         if (ret)
2372                 return ret;
2373
2374 out:
2375         btrfs_release_path(path);
2376
2377         hole_em = alloc_extent_map();
2378         if (!hole_em) {
2379                 btrfs_drop_extent_cache(inode, offset, end - 1, 0);
2380                 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &inode->runtime_flags);
2381         } else {
2382                 hole_em->start = offset;
2383                 hole_em->len = end - offset;
2384                 hole_em->ram_bytes = hole_em->len;
2385                 hole_em->orig_start = offset;
2386
2387                 hole_em->block_start = EXTENT_MAP_HOLE;
2388                 hole_em->block_len = 0;
2389                 hole_em->orig_block_len = 0;
2390                 hole_em->bdev = fs_info->fs_devices->latest_bdev;
2391                 hole_em->compress_type = BTRFS_COMPRESS_NONE;
2392                 hole_em->generation = trans->transid;
2393
2394                 do {
2395                         btrfs_drop_extent_cache(inode, offset, end - 1, 0);
2396                         write_lock(&em_tree->lock);
2397                         ret = add_extent_mapping(em_tree, hole_em, 1);
2398                         write_unlock(&em_tree->lock);
2399                 } while (ret == -EEXIST);
2400                 free_extent_map(hole_em);
2401                 if (ret)
2402                         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2403                                         &inode->runtime_flags);
2404         }
2405
2406         return 0;
2407 }
2408
2409 /*
2410  * Find a hole extent on given inode and change start/len to the end of hole
2411  * extent.(hole/vacuum extent whose em->start <= start &&
2412  *         em->start + em->len > start)
2413  * When a hole extent is found, return 1 and modify start/len.
2414  */
2415 static int find_first_non_hole(struct inode *inode, u64 *start, u64 *len)
2416 {
2417         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2418         struct extent_map *em;
2419         int ret = 0;
2420
2421         em = btrfs_get_extent(BTRFS_I(inode), NULL, 0,
2422                               round_down(*start, fs_info->sectorsize),
2423                               round_up(*len, fs_info->sectorsize), 0);
2424         if (IS_ERR(em))
2425                 return PTR_ERR(em);
2426
2427         /* Hole or vacuum extent(only exists in no-hole mode) */
2428         if (em->block_start == EXTENT_MAP_HOLE) {
2429                 ret = 1;
2430                 *len = em->start + em->len > *start + *len ?
2431                        0 : *start + *len - em->start - em->len;
2432                 *start = em->start + em->len;
2433         }
2434         free_extent_map(em);
2435         return ret;
2436 }
2437
2438 static int btrfs_punch_hole(struct inode *inode, loff_t offset, loff_t len)
2439 {
2440         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2441         struct btrfs_root *root = BTRFS_I(inode)->root;
2442         struct extent_state *cached_state = NULL;
2443         struct btrfs_path *path;
2444         struct btrfs_block_rsv *rsv;
2445         struct btrfs_trans_handle *trans;
2446         u64 lockstart;
2447         u64 lockend;
2448         u64 tail_start;
2449         u64 tail_len;
2450         u64 orig_start = offset;
2451         u64 cur_offset;
2452         u64 min_size = btrfs_calc_trans_metadata_size(fs_info, 1);
2453         u64 drop_end;
2454         int ret = 0;
2455         int err = 0;
2456         unsigned int rsv_count;
2457         bool same_block;
2458         bool no_holes = btrfs_fs_incompat(fs_info, NO_HOLES);
2459         u64 ino_size;
2460         bool truncated_block = false;
2461         bool updated_inode = false;
2462
2463         ret = btrfs_wait_ordered_range(inode, offset, len);
2464         if (ret)
2465                 return ret;
2466
2467         inode_lock(inode);
2468         ino_size = round_up(inode->i_size, fs_info->sectorsize);
2469         ret = find_first_non_hole(inode, &offset, &len);
2470         if (ret < 0)
2471                 goto out_only_mutex;
2472         if (ret && !len) {
2473                 /* Already in a large hole */
2474                 ret = 0;
2475                 goto out_only_mutex;
2476         }
2477
2478         lockstart = round_up(offset, btrfs_inode_sectorsize(inode));
2479         lockend = round_down(offset + len,
2480                              btrfs_inode_sectorsize(inode)) - 1;
2481         same_block = (BTRFS_BYTES_TO_BLKS(fs_info, offset))
2482                 == (BTRFS_BYTES_TO_BLKS(fs_info, offset + len - 1));
2483         /*
2484          * We needn't truncate any block which is beyond the end of the file
2485          * because we are sure there is no data there.
2486          */
2487         /*
2488          * Only do this if we are in the same block and we aren't doing the
2489          * entire block.
2490          */
2491         if (same_block && len < fs_info->sectorsize) {
2492                 if (offset < ino_size) {
2493                         truncated_block = true;
2494                         ret = btrfs_truncate_block(inode, offset, len, 0);
2495                 } else {
2496                         ret = 0;
2497                 }
2498                 goto out_only_mutex;
2499         }
2500
2501         /* zero back part of the first block */
2502         if (offset < ino_size) {
2503                 truncated_block = true;
2504                 ret = btrfs_truncate_block(inode, offset, 0, 0);
2505                 if (ret) {
2506                         inode_unlock(inode);
2507                         return ret;
2508                 }
2509         }
2510
2511         /* Check the aligned pages after the first unaligned page,
2512          * if offset != orig_start, which means the first unaligned page
2513          * including several following pages are already in holes,
2514          * the extra check can be skipped */
2515         if (offset == orig_start) {
2516                 /* after truncate page, check hole again */
2517                 len = offset + len - lockstart;
2518                 offset = lockstart;
2519                 ret = find_first_non_hole(inode, &offset, &len);
2520                 if (ret < 0)
2521                         goto out_only_mutex;
2522                 if (ret && !len) {
2523                         ret = 0;
2524                         goto out_only_mutex;
2525                 }
2526                 lockstart = offset;
2527         }
2528
2529         /* Check the tail unaligned part is in a hole */
2530         tail_start = lockend + 1;
2531         tail_len = offset + len - tail_start;
2532         if (tail_len) {
2533                 ret = find_first_non_hole(inode, &tail_start, &tail_len);
2534                 if (unlikely(ret < 0))
2535                         goto out_only_mutex;
2536                 if (!ret) {
2537                         /* zero the front end of the last page */
2538                         if (tail_start + tail_len < ino_size) {
2539                                 truncated_block = true;
2540                                 ret = btrfs_truncate_block(inode,
2541                                                         tail_start + tail_len,
2542                                                         0, 1);
2543                                 if (ret)
2544                                         goto out_only_mutex;
2545                         }
2546                 }
2547         }
2548
2549         if (lockend < lockstart) {
2550                 ret = 0;
2551                 goto out_only_mutex;
2552         }
2553
2554         while (1) {
2555                 struct btrfs_ordered_extent *ordered;
2556
2557                 truncate_pagecache_range(inode, lockstart, lockend);
2558
2559                 lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2560                                  &cached_state);
2561                 ordered = btrfs_lookup_first_ordered_extent(inode, lockend);
2562
2563                 /*
2564                  * We need to make sure we have no ordered extents in this range
2565                  * and nobody raced in and read a page in this range, if we did
2566                  * we need to try again.
2567                  */
2568                 if ((!ordered ||
2569                     (ordered->file_offset + ordered->len <= lockstart ||
2570                      ordered->file_offset > lockend)) &&
2571                      !btrfs_page_exists_in_range(inode, lockstart, lockend)) {
2572                         if (ordered)
2573                                 btrfs_put_ordered_extent(ordered);
2574                         break;
2575                 }
2576                 if (ordered)
2577                         btrfs_put_ordered_extent(ordered);
2578                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart,
2579                                      lockend, &cached_state, GFP_NOFS);
2580                 ret = btrfs_wait_ordered_range(inode, lockstart,
2581                                                lockend - lockstart + 1);
2582                 if (ret) {
2583                         inode_unlock(inode);
2584                         return ret;
2585                 }
2586         }
2587
2588         path = btrfs_alloc_path();
2589         if (!path) {
2590                 ret = -ENOMEM;
2591                 goto out;
2592         }
2593
2594         rsv = btrfs_alloc_block_rsv(fs_info, BTRFS_BLOCK_RSV_TEMP);
2595         if (!rsv) {
2596                 ret = -ENOMEM;
2597                 goto out_free;
2598         }
2599         rsv->size = btrfs_calc_trans_metadata_size(fs_info, 1);
2600         rsv->failfast = 1;
2601
2602         /*
2603          * 1 - update the inode
2604          * 1 - removing the extents in the range
2605          * 1 - adding the hole extent if no_holes isn't set
2606          */
2607         rsv_count = no_holes ? 2 : 3;
2608         trans = btrfs_start_transaction(root, rsv_count);
2609         if (IS_ERR(trans)) {
2610                 err = PTR_ERR(trans);
2611                 goto out_free;
2612         }
2613
2614         ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv, rsv,
2615                                       min_size, 0);
2616         BUG_ON(ret);
2617         trans->block_rsv = rsv;
2618
2619         cur_offset = lockstart;
2620         len = lockend - cur_offset;
2621         while (cur_offset < lockend) {
2622                 ret = __btrfs_drop_extents(trans, root, inode, path,
2623                                            cur_offset, lockend + 1,
2624                                            &drop_end, 1, 0, 0, NULL);
2625                 if (ret != -ENOSPC)
2626                         break;
2627
2628                 trans->block_rsv = &fs_info->trans_block_rsv;
2629
2630                 if (cur_offset < drop_end && cur_offset < ino_size) {
2631                         ret = fill_holes(trans, BTRFS_I(inode), path,
2632                                         cur_offset, drop_end);
2633                         if (ret) {
2634                                 /*
2635                                  * If we failed then we didn't insert our hole
2636                                  * entries for the area we dropped, so now the
2637                                  * fs is corrupted, so we must abort the
2638                                  * transaction.
2639                                  */
2640                                 btrfs_abort_transaction(trans, ret);
2641                                 err = ret;
2642                                 break;
2643                         }
2644                 }
2645
2646                 cur_offset = drop_end;
2647
2648                 ret = btrfs_update_inode(trans, root, inode);
2649                 if (ret) {
2650                         err = ret;
2651                         break;
2652                 }
2653
2654                 btrfs_end_transaction(trans);
2655                 btrfs_btree_balance_dirty(fs_info);
2656
2657                 trans = btrfs_start_transaction(root, rsv_count);
2658                 if (IS_ERR(trans)) {
2659                         ret = PTR_ERR(trans);
2660                         trans = NULL;
2661                         break;
2662                 }
2663
2664                 ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv,
2665                                               rsv, min_size, 0);
2666                 BUG_ON(ret);    /* shouldn't happen */
2667                 trans->block_rsv = rsv;
2668
2669                 ret = find_first_non_hole(inode, &cur_offset, &len);
2670                 if (unlikely(ret < 0))
2671                         break;
2672                 if (ret && !len) {
2673                         ret = 0;
2674                         break;
2675                 }
2676         }
2677
2678         if (ret) {
2679                 err = ret;
2680                 goto out_trans;
2681         }
2682
2683         trans->block_rsv = &fs_info->trans_block_rsv;
2684         /*
2685          * If we are using the NO_HOLES feature we might have had already an
2686          * hole that overlaps a part of the region [lockstart, lockend] and
2687          * ends at (or beyond) lockend. Since we have no file extent items to
2688          * represent holes, drop_end can be less than lockend and so we must
2689          * make sure we have an extent map representing the existing hole (the
2690          * call to __btrfs_drop_extents() might have dropped the existing extent
2691          * map representing the existing hole), otherwise the fast fsync path
2692          * will not record the existence of the hole region
2693          * [existing_hole_start, lockend].
2694          */
2695         if (drop_end <= lockend)
2696                 drop_end = lockend + 1;
2697         /*
2698          * Don't insert file hole extent item if it's for a range beyond eof
2699          * (because it's useless) or if it represents a 0 bytes range (when
2700          * cur_offset == drop_end).
2701          */
2702         if (cur_offset < ino_size && cur_offset < drop_end) {
2703                 ret = fill_holes(trans, BTRFS_I(inode), path,
2704                                 cur_offset, drop_end);
2705                 if (ret) {
2706                         /* Same comment as above. */
2707                         btrfs_abort_transaction(trans, ret);
2708                         err = ret;
2709                         goto out_trans;
2710                 }
2711         }
2712
2713 out_trans:
2714         if (!trans)
2715                 goto out_free;
2716
2717         inode_inc_iversion(inode);
2718         inode->i_mtime = inode->i_ctime = current_time(inode);
2719
2720         trans->block_rsv = &fs_info->trans_block_rsv;
2721         ret = btrfs_update_inode(trans, root, inode);
2722         updated_inode = true;
2723         btrfs_end_transaction(trans);
2724         btrfs_btree_balance_dirty(fs_info);
2725 out_free:
2726         btrfs_free_path(path);
2727         btrfs_free_block_rsv(fs_info, rsv);
2728 out:
2729         unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2730                              &cached_state, GFP_NOFS);
2731 out_only_mutex:
2732         if (!updated_inode && truncated_block && !ret && !err) {
2733                 /*
2734                  * If we only end up zeroing part of a page, we still need to
2735                  * update the inode item, so that all the time fields are
2736                  * updated as well as the necessary btrfs inode in memory fields
2737                  * for detecting, at fsync time, if the inode isn't yet in the
2738                  * log tree or it's there but not up to date.
2739                  */
2740                 struct timespec now = current_time(inode);
2741
2742                 inode_inc_iversion(inode);
2743                 inode->i_mtime = now;
2744                 inode->i_ctime = now;
2745                 trans = btrfs_start_transaction(root, 1);
2746                 if (IS_ERR(trans)) {
2747                         err = PTR_ERR(trans);
2748                 } else {
2749                         err = btrfs_update_inode(trans, root, inode);
2750                         ret = btrfs_end_transaction(trans);
2751                 }
2752         }
2753         inode_unlock(inode);
2754         if (ret && !err)
2755                 err = ret;
2756         return err;
2757 }
2758
2759 /* Helper structure to record which range is already reserved */
2760 struct falloc_range {
2761         struct list_head list;
2762         u64 start;
2763         u64 len;
2764 };
2765
2766 /*
2767  * Helper function to add falloc range
2768  *
2769  * Caller should have locked the larger range of extent containing
2770  * [start, len)
2771  */
2772 static int add_falloc_range(struct list_head *head, u64 start, u64 len)
2773 {
2774         struct falloc_range *prev = NULL;
2775         struct falloc_range *range = NULL;
2776
2777         if (list_empty(head))
2778                 goto insert;
2779
2780         /*
2781          * As fallocate iterate by bytenr order, we only need to check
2782          * the last range.
2783          */
2784         prev = list_entry(head->prev, struct falloc_range, list);
2785         if (prev->start + prev->len == start) {
2786                 prev->len += len;
2787                 return 0;
2788         }
2789 insert:
2790         range = kmalloc(sizeof(*range), GFP_KERNEL);
2791         if (!range)
2792                 return -ENOMEM;
2793         range->start = start;
2794         range->len = len;
2795         list_add_tail(&range->list, head);
2796         return 0;
2797 }
2798
2799 static long btrfs_fallocate(struct file *file, int mode,
2800                             loff_t offset, loff_t len)
2801 {
2802         struct inode *inode = file_inode(file);
2803         struct extent_state *cached_state = NULL;
2804         struct extent_changeset *data_reserved = NULL;
2805         struct falloc_range *range;
2806         struct falloc_range *tmp;
2807         struct list_head reserve_list;
2808         u64 cur_offset;
2809         u64 last_byte;
2810         u64 alloc_start;
2811         u64 alloc_end;
2812         u64 alloc_hint = 0;
2813         u64 locked_end;
2814         u64 actual_end = 0;
2815         struct extent_map *em;
2816         int blocksize = btrfs_inode_sectorsize(inode);
2817         int ret;
2818
2819         alloc_start = round_down(offset, blocksize);
2820         alloc_end = round_up(offset + len, blocksize);
2821         cur_offset = alloc_start;
2822
2823         /* Make sure we aren't being give some crap mode */
2824         if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
2825                 return -EOPNOTSUPP;
2826
2827         if (mode & FALLOC_FL_PUNCH_HOLE)
2828                 return btrfs_punch_hole(inode, offset, len);
2829
2830         /*
2831          * Only trigger disk allocation, don't trigger qgroup reserve
2832          *
2833          * For qgroup space, it will be checked later.
2834          */
2835         ret = btrfs_alloc_data_chunk_ondemand(BTRFS_I(inode),
2836                         alloc_end - alloc_start);
2837         if (ret < 0)
2838                 return ret;
2839
2840         inode_lock(inode);
2841
2842         if (!(mode & FALLOC_FL_KEEP_SIZE) && offset + len > inode->i_size) {
2843                 ret = inode_newsize_ok(inode, offset + len);
2844                 if (ret)
2845                         goto out;
2846         }
2847
2848         /*
2849          * TODO: Move these two operations after we have checked
2850          * accurate reserved space, or fallocate can still fail but
2851          * with page truncated or size expanded.
2852          *
2853          * But that's a minor problem and won't do much harm BTW.
2854          */
2855         if (alloc_start > inode->i_size) {
2856                 ret = btrfs_cont_expand(inode, i_size_read(inode),
2857                                         alloc_start);
2858                 if (ret)
2859                         goto out;
2860         } else if (offset + len > inode->i_size) {
2861                 /*
2862                  * If we are fallocating from the end of the file onward we
2863                  * need to zero out the end of the block if i_size lands in the
2864                  * middle of a block.
2865                  */
2866                 ret = btrfs_truncate_block(inode, inode->i_size, 0, 0);
2867                 if (ret)
2868                         goto out;
2869         }
2870
2871         /*
2872          * wait for ordered IO before we have any locks.  We'll loop again
2873          * below with the locks held.
2874          */
2875         ret = btrfs_wait_ordered_range(inode, alloc_start,
2876                                        alloc_end - alloc_start);
2877         if (ret)
2878                 goto out;
2879
2880         locked_end = alloc_end - 1;
2881         while (1) {
2882                 struct btrfs_ordered_extent *ordered;
2883
2884                 /* the extent lock is ordered inside the running
2885                  * transaction
2886                  */
2887                 lock_extent_bits(&BTRFS_I(inode)->io_tree, alloc_start,
2888                                  locked_end, &cached_state);
2889                 ordered = btrfs_lookup_first_ordered_extent(inode,
2890                                                             alloc_end - 1);
2891                 if (ordered &&
2892                     ordered->file_offset + ordered->len > alloc_start &&
2893                     ordered->file_offset < alloc_end) {
2894                         btrfs_put_ordered_extent(ordered);
2895                         unlock_extent_cached(&BTRFS_I(inode)->io_tree,
2896                                              alloc_start, locked_end,
2897                                              &cached_state, GFP_KERNEL);
2898                         /*
2899                          * we can't wait on the range with the transaction
2900                          * running or with the extent lock held
2901                          */
2902                         ret = btrfs_wait_ordered_range(inode, alloc_start,
2903                                                        alloc_end - alloc_start);
2904                         if (ret)
2905                                 goto out;
2906                 } else {
2907                         if (ordered)
2908                                 btrfs_put_ordered_extent(ordered);
2909                         break;
2910                 }
2911         }
2912
2913         /* First, check if we exceed the qgroup limit */
2914         INIT_LIST_HEAD(&reserve_list);
2915         while (1) {
2916                 em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, cur_offset,
2917                                       alloc_end - cur_offset, 0);
2918                 if (IS_ERR(em)) {
2919                         ret = PTR_ERR(em);
2920                         break;
2921                 }
2922                 last_byte = min(extent_map_end(em), alloc_end);
2923                 actual_end = min_t(u64, extent_map_end(em), offset + len);
2924                 last_byte = ALIGN(last_byte, blocksize);
2925                 if (em->block_start == EXTENT_MAP_HOLE ||
2926                     (cur_offset >= inode->i_size &&
2927                      !test_bit(EXTENT_FLAG_PREALLOC, &em->flags))) {
2928                         ret = add_falloc_range(&reserve_list, cur_offset,
2929                                                last_byte - cur_offset);
2930                         if (ret < 0) {
2931                                 free_extent_map(em);
2932                                 break;
2933                         }
2934                         ret = btrfs_qgroup_reserve_data(inode, &data_reserved,
2935                                         cur_offset, last_byte - cur_offset);
2936                         if (ret < 0) {
2937                                 cur_offset = last_byte;
2938                                 free_extent_map(em);
2939                                 break;
2940                         }
2941                 } else {
2942                         /*
2943                          * Do not need to reserve unwritten extent for this
2944                          * range, free reserved data space first, otherwise
2945                          * it'll result in false ENOSPC error.
2946                          */
2947                         btrfs_free_reserved_data_space(inode, data_reserved,
2948                                         cur_offset, last_byte - cur_offset);
2949                 }
2950                 free_extent_map(em);
2951                 cur_offset = last_byte;
2952                 if (cur_offset >= alloc_end)
2953                         break;
2954         }
2955
2956         /*
2957          * If ret is still 0, means we're OK to fallocate.
2958          * Or just cleanup the list and exit.
2959          */
2960         list_for_each_entry_safe(range, tmp, &reserve_list, list) {
2961                 if (!ret)
2962                         ret = btrfs_prealloc_file_range(inode, mode,
2963                                         range->start,
2964                                         range->len, i_blocksize(inode),
2965                                         offset + len, &alloc_hint);
2966                 else
2967                         btrfs_free_reserved_data_space(inode,
2968                                         data_reserved, range->start,
2969                                         range->len);
2970                 list_del(&range->list);
2971                 kfree(range);
2972         }
2973         if (ret < 0)
2974                 goto out_unlock;
2975
2976         if (actual_end > inode->i_size &&
2977             !(mode & FALLOC_FL_KEEP_SIZE)) {
2978                 struct btrfs_trans_handle *trans;
2979                 struct btrfs_root *root = BTRFS_I(inode)->root;
2980
2981                 /*
2982                  * We didn't need to allocate any more space, but we
2983                  * still extended the size of the file so we need to
2984                  * update i_size and the inode item.
2985                  */
2986                 trans = btrfs_start_transaction(root, 1);
2987                 if (IS_ERR(trans)) {
2988                         ret = PTR_ERR(trans);
2989                 } else {
2990                         inode->i_ctime = current_time(inode);
2991                         i_size_write(inode, actual_end);
2992                         btrfs_ordered_update_i_size(inode, actual_end, NULL);
2993                         ret = btrfs_update_inode(trans, root, inode);
2994                         if (ret)
2995                                 btrfs_end_transaction(trans);
2996                         else
2997                                 ret = btrfs_end_transaction(trans);
2998                 }
2999         }
3000 out_unlock:
3001         unlock_extent_cached(&BTRFS_I(inode)->io_tree, alloc_start, locked_end,
3002                              &cached_state, GFP_KERNEL);
3003 out:
3004         inode_unlock(inode);
3005         /* Let go of our reservation. */
3006         if (ret != 0)
3007                 btrfs_free_reserved_data_space(inode, data_reserved,
3008                                 cur_offset, alloc_end - cur_offset);
3009         extent_changeset_free(data_reserved);
3010         return ret;
3011 }
3012
3013 static int find_desired_extent(struct inode *inode, loff_t *offset, int whence)
3014 {
3015         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
3016         struct extent_map *em = NULL;
3017         struct extent_state *cached_state = NULL;
3018         u64 lockstart;
3019         u64 lockend;
3020         u64 start;
3021         u64 len;
3022         int ret = 0;
3023
3024         if (inode->i_size == 0)
3025                 return -ENXIO;
3026
3027         /*
3028          * *offset can be negative, in this case we start finding DATA/HOLE from
3029          * the very start of the file.
3030          */
3031         start = max_t(loff_t, 0, *offset);
3032
3033         lockstart = round_down(start, fs_info->sectorsize);
3034         lockend = round_up(i_size_read(inode),
3035                            fs_info->sectorsize);
3036         if (lockend <= lockstart)
3037                 lockend = lockstart + fs_info->sectorsize;
3038         lockend--;
3039         len = lockend - lockstart + 1;
3040
3041         lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend,
3042                          &cached_state);
3043
3044         while (start < inode->i_size) {
3045                 em = btrfs_get_extent_fiemap(BTRFS_I(inode), NULL, 0,
3046                                 start, len, 0);
3047                 if (IS_ERR(em)) {
3048                         ret = PTR_ERR(em);
3049                         em = NULL;
3050                         break;
3051                 }
3052
3053                 if (whence == SEEK_HOLE &&
3054                     (em->block_start == EXTENT_MAP_HOLE ||
3055                      test_bit(EXTENT_FLAG_PREALLOC, &em->flags)))
3056                         break;
3057                 else if (whence == SEEK_DATA &&
3058                            (em->block_start != EXTENT_MAP_HOLE &&
3059                             !test_bit(EXTENT_FLAG_PREALLOC, &em->flags)))
3060                         break;
3061
3062                 start = em->start + em->len;
3063                 free_extent_map(em);
3064                 em = NULL;
3065                 cond_resched();
3066         }
3067         free_extent_map(em);
3068         if (!ret) {
3069                 if (whence == SEEK_DATA && start >= inode->i_size)
3070                         ret = -ENXIO;
3071                 else
3072                         *offset = min_t(loff_t, start, inode->i_size);
3073         }
3074         unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
3075                              &cached_state, GFP_NOFS);
3076         return ret;
3077 }
3078
3079 static loff_t btrfs_file_llseek(struct file *file, loff_t offset, int whence)
3080 {
3081         struct inode *inode = file->f_mapping->host;
3082         int ret;
3083
3084         inode_lock(inode);
3085         switch (whence) {
3086         case SEEK_END:
3087         case SEEK_CUR:
3088                 offset = generic_file_llseek(file, offset, whence);
3089                 goto out;
3090         case SEEK_DATA:
3091         case SEEK_HOLE:
3092                 if (offset >= i_size_read(inode)) {
3093                         inode_unlock(inode);
3094                         return -ENXIO;
3095                 }
3096
3097                 ret = find_desired_extent(inode, &offset, whence);
3098                 if (ret) {
3099                         inode_unlock(inode);
3100                         return ret;
3101                 }
3102         }
3103
3104         offset = vfs_setpos(file, offset, inode->i_sb->s_maxbytes);
3105 out:
3106         inode_unlock(inode);
3107         return offset;
3108 }
3109
3110 static int btrfs_file_open(struct inode *inode, struct file *filp)
3111 {
3112         filp->f_mode |= FMODE_NOWAIT;
3113         return generic_file_open(inode, filp);
3114 }
3115
3116 const struct file_operations btrfs_file_operations = {
3117         .llseek         = btrfs_file_llseek,
3118         .read_iter      = generic_file_read_iter,
3119         .splice_read    = generic_file_splice_read,
3120         .write_iter     = btrfs_file_write_iter,
3121         .mmap           = btrfs_file_mmap,
3122         .open           = btrfs_file_open,
3123         .release        = btrfs_release_file,
3124         .fsync          = btrfs_sync_file,
3125         .fallocate      = btrfs_fallocate,
3126         .unlocked_ioctl = btrfs_ioctl,
3127 #ifdef CONFIG_COMPAT
3128         .compat_ioctl   = btrfs_compat_ioctl,
3129 #endif
3130         .clone_file_range = btrfs_clone_file_range,
3131         .dedupe_file_range = btrfs_dedupe_file_range,
3132 };
3133
3134 void btrfs_auto_defrag_exit(void)
3135 {
3136         kmem_cache_destroy(btrfs_inode_defrag_cachep);
3137 }
3138
3139 int btrfs_auto_defrag_init(void)
3140 {
3141         btrfs_inode_defrag_cachep = kmem_cache_create("btrfs_inode_defrag",
3142                                         sizeof(struct inode_defrag), 0,
3143                                         SLAB_MEM_SPREAD,
3144                                         NULL);
3145         if (!btrfs_inode_defrag_cachep)
3146                 return -ENOMEM;
3147
3148         return 0;
3149 }
3150
3151 int btrfs_fdatawrite_range(struct inode *inode, loff_t start, loff_t end)
3152 {
3153         int ret;
3154
3155         /*
3156          * So with compression we will find and lock a dirty page and clear the
3157          * first one as dirty, setup an async extent, and immediately return
3158          * with the entire range locked but with nobody actually marked with
3159          * writeback.  So we can't just filemap_write_and_wait_range() and
3160          * expect it to work since it will just kick off a thread to do the
3161          * actual work.  So we need to call filemap_fdatawrite_range _again_
3162          * since it will wait on the page lock, which won't be unlocked until
3163          * after the pages have been marked as writeback and so we're good to go
3164          * from there.  We have to do this otherwise we'll miss the ordered
3165          * extents and that results in badness.  Please Josef, do not think you
3166          * know better and pull this out at some point in the future, it is
3167          * right and you are wrong.
3168          */
3169         ret = filemap_fdatawrite_range(inode->i_mapping, start, end);
3170         if (!ret && test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
3171                              &BTRFS_I(inode)->runtime_flags))
3172                 ret = filemap_fdatawrite_range(inode->i_mapping, start, end);
3173
3174         return ret;
3175 }