Linux Audio

Check our new training course

Loading...
v5.4
   1// SPDX-License-Identifier: GPL-2.0
   2/*
   3 * Copyright (C) 2007 Oracle.  All rights reserved.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
   4 */
   5
   6#include <linux/fs.h>
   7#include <linux/pagemap.h>
 
   8#include <linux/time.h>
   9#include <linux/init.h>
  10#include <linux/string.h>
  11#include <linux/backing-dev.h>
 
  12#include <linux/falloc.h>
 
  13#include <linux/writeback.h>
 
  14#include <linux/compat.h>
  15#include <linux/slab.h>
  16#include <linux/btrfs.h>
  17#include <linux/uio.h>
  18#include <linux/iversion.h>
  19#include "ctree.h"
  20#include "disk-io.h"
  21#include "transaction.h"
  22#include "btrfs_inode.h"
  23#include "print-tree.h"
  24#include "tree-log.h"
  25#include "locking.h"
  26#include "volumes.h"
  27#include "qgroup.h"
  28#include "compression.h"
  29#include "delalloc-space.h"
  30
  31static struct kmem_cache *btrfs_inode_defrag_cachep;
  32/*
  33 * when auto defrag is enabled we
  34 * queue up these defrag structs to remember which
  35 * inodes need defragging passes
  36 */
  37struct inode_defrag {
  38	struct rb_node rb_node;
  39	/* objectid */
  40	u64 ino;
  41	/*
  42	 * transid where the defrag was added, we search for
  43	 * extents newer than this
  44	 */
  45	u64 transid;
  46
  47	/* root objectid */
  48	u64 root;
  49
  50	/* last offset we were able to defrag */
  51	u64 last_offset;
  52
  53	/* if we've wrapped around back to zero once already */
  54	int cycled;
  55};
  56
  57static int __compare_inode_defrag(struct inode_defrag *defrag1,
  58				  struct inode_defrag *defrag2)
  59{
  60	if (defrag1->root > defrag2->root)
  61		return 1;
  62	else if (defrag1->root < defrag2->root)
  63		return -1;
  64	else if (defrag1->ino > defrag2->ino)
  65		return 1;
  66	else if (defrag1->ino < defrag2->ino)
  67		return -1;
  68	else
  69		return 0;
  70}
  71
  72/* pop a record for an inode into the defrag tree.  The lock
  73 * must be held already
  74 *
  75 * If you're inserting a record for an older transid than an
  76 * existing record, the transid already in the tree is lowered
  77 *
  78 * If an existing record is found the defrag item you
  79 * pass in is freed
  80 */
  81static int __btrfs_add_inode_defrag(struct btrfs_inode *inode,
  82				    struct inode_defrag *defrag)
  83{
  84	struct btrfs_fs_info *fs_info = inode->root->fs_info;
  85	struct inode_defrag *entry;
  86	struct rb_node **p;
  87	struct rb_node *parent = NULL;
  88	int ret;
  89
  90	p = &fs_info->defrag_inodes.rb_node;
  91	while (*p) {
  92		parent = *p;
  93		entry = rb_entry(parent, struct inode_defrag, rb_node);
  94
  95		ret = __compare_inode_defrag(defrag, entry);
  96		if (ret < 0)
  97			p = &parent->rb_left;
  98		else if (ret > 0)
  99			p = &parent->rb_right;
 100		else {
 101			/* if we're reinserting an entry for
 102			 * an old defrag run, make sure to
 103			 * lower the transid of our existing record
 104			 */
 105			if (defrag->transid < entry->transid)
 106				entry->transid = defrag->transid;
 107			if (defrag->last_offset > entry->last_offset)
 108				entry->last_offset = defrag->last_offset;
 109			return -EEXIST;
 110		}
 111	}
 112	set_bit(BTRFS_INODE_IN_DEFRAG, &inode->runtime_flags);
 113	rb_link_node(&defrag->rb_node, parent, p);
 114	rb_insert_color(&defrag->rb_node, &fs_info->defrag_inodes);
 115	return 0;
 116}
 117
 118static inline int __need_auto_defrag(struct btrfs_fs_info *fs_info)
 119{
 120	if (!btrfs_test_opt(fs_info, AUTO_DEFRAG))
 121		return 0;
 122
 123	if (btrfs_fs_closing(fs_info))
 124		return 0;
 125
 126	return 1;
 127}
 128
 129/*
 130 * insert a defrag record for this inode if auto defrag is
 131 * enabled
 132 */
 133int btrfs_add_inode_defrag(struct btrfs_trans_handle *trans,
 134			   struct btrfs_inode *inode)
 135{
 136	struct btrfs_root *root = inode->root;
 137	struct btrfs_fs_info *fs_info = root->fs_info;
 138	struct inode_defrag *defrag;
 139	u64 transid;
 140	int ret;
 141
 142	if (!__need_auto_defrag(fs_info))
 143		return 0;
 144
 145	if (test_bit(BTRFS_INODE_IN_DEFRAG, &inode->runtime_flags))
 146		return 0;
 147
 148	if (trans)
 149		transid = trans->transid;
 150	else
 151		transid = inode->root->last_trans;
 152
 153	defrag = kmem_cache_zalloc(btrfs_inode_defrag_cachep, GFP_NOFS);
 154	if (!defrag)
 155		return -ENOMEM;
 156
 157	defrag->ino = btrfs_ino(inode);
 158	defrag->transid = transid;
 159	defrag->root = root->root_key.objectid;
 160
 161	spin_lock(&fs_info->defrag_inodes_lock);
 162	if (!test_bit(BTRFS_INODE_IN_DEFRAG, &inode->runtime_flags)) {
 163		/*
 164		 * If we set IN_DEFRAG flag and evict the inode from memory,
 165		 * and then re-read this inode, this new inode doesn't have
 166		 * IN_DEFRAG flag. At the case, we may find the existed defrag.
 167		 */
 168		ret = __btrfs_add_inode_defrag(inode, defrag);
 169		if (ret)
 170			kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
 171	} else {
 172		kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
 173	}
 174	spin_unlock(&fs_info->defrag_inodes_lock);
 175	return 0;
 176}
 177
 178/*
 179 * Requeue the defrag object. If there is a defrag object that points to
 180 * the same inode in the tree, we will merge them together (by
 181 * __btrfs_add_inode_defrag()) and free the one that we want to requeue.
 182 */
 183static void btrfs_requeue_inode_defrag(struct btrfs_inode *inode,
 184				       struct inode_defrag *defrag)
 185{
 186	struct btrfs_fs_info *fs_info = inode->root->fs_info;
 187	int ret;
 188
 189	if (!__need_auto_defrag(fs_info))
 190		goto out;
 191
 192	/*
 193	 * Here we don't check the IN_DEFRAG flag, because we need merge
 194	 * them together.
 195	 */
 196	spin_lock(&fs_info->defrag_inodes_lock);
 197	ret = __btrfs_add_inode_defrag(inode, defrag);
 198	spin_unlock(&fs_info->defrag_inodes_lock);
 199	if (ret)
 200		goto out;
 201	return;
 202out:
 203	kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
 204}
 205
 206/*
 207 * pick the defragable inode that we want, if it doesn't exist, we will get
 208 * the next one.
 209 */
 210static struct inode_defrag *
 211btrfs_pick_defrag_inode(struct btrfs_fs_info *fs_info, u64 root, u64 ino)
 212{
 213	struct inode_defrag *entry = NULL;
 214	struct inode_defrag tmp;
 215	struct rb_node *p;
 216	struct rb_node *parent = NULL;
 217	int ret;
 218
 219	tmp.ino = ino;
 220	tmp.root = root;
 221
 222	spin_lock(&fs_info->defrag_inodes_lock);
 223	p = fs_info->defrag_inodes.rb_node;
 224	while (p) {
 225		parent = p;
 226		entry = rb_entry(parent, struct inode_defrag, rb_node);
 227
 228		ret = __compare_inode_defrag(&tmp, entry);
 229		if (ret < 0)
 230			p = parent->rb_left;
 231		else if (ret > 0)
 232			p = parent->rb_right;
 233		else
 234			goto out;
 235	}
 236
 237	if (parent && __compare_inode_defrag(&tmp, entry) > 0) {
 238		parent = rb_next(parent);
 239		if (parent)
 240			entry = rb_entry(parent, struct inode_defrag, rb_node);
 241		else
 242			entry = NULL;
 243	}
 244out:
 245	if (entry)
 246		rb_erase(parent, &fs_info->defrag_inodes);
 247	spin_unlock(&fs_info->defrag_inodes_lock);
 248	return entry;
 249}
 250
 251void btrfs_cleanup_defrag_inodes(struct btrfs_fs_info *fs_info)
 252{
 253	struct inode_defrag *defrag;
 254	struct rb_node *node;
 255
 256	spin_lock(&fs_info->defrag_inodes_lock);
 257	node = rb_first(&fs_info->defrag_inodes);
 258	while (node) {
 259		rb_erase(node, &fs_info->defrag_inodes);
 260		defrag = rb_entry(node, struct inode_defrag, rb_node);
 261		kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
 262
 263		cond_resched_lock(&fs_info->defrag_inodes_lock);
 264
 265		node = rb_first(&fs_info->defrag_inodes);
 266	}
 267	spin_unlock(&fs_info->defrag_inodes_lock);
 268}
 269
 270#define BTRFS_DEFRAG_BATCH	1024
 271
 272static int __btrfs_run_defrag_inode(struct btrfs_fs_info *fs_info,
 273				    struct inode_defrag *defrag)
 274{
 275	struct btrfs_root *inode_root;
 276	struct inode *inode;
 277	struct btrfs_key key;
 278	struct btrfs_ioctl_defrag_range_args range;
 279	int num_defrag;
 280	int index;
 281	int ret;
 282
 283	/* get the inode */
 284	key.objectid = defrag->root;
 285	key.type = BTRFS_ROOT_ITEM_KEY;
 286	key.offset = (u64)-1;
 287
 288	index = srcu_read_lock(&fs_info->subvol_srcu);
 289
 290	inode_root = btrfs_read_fs_root_no_name(fs_info, &key);
 291	if (IS_ERR(inode_root)) {
 292		ret = PTR_ERR(inode_root);
 293		goto cleanup;
 294	}
 295
 296	key.objectid = defrag->ino;
 297	key.type = BTRFS_INODE_ITEM_KEY;
 298	key.offset = 0;
 299	inode = btrfs_iget(fs_info->sb, &key, inode_root, NULL);
 300	if (IS_ERR(inode)) {
 301		ret = PTR_ERR(inode);
 302		goto cleanup;
 303	}
 304	srcu_read_unlock(&fs_info->subvol_srcu, index);
 305
 306	/* do a chunk of defrag */
 307	clear_bit(BTRFS_INODE_IN_DEFRAG, &BTRFS_I(inode)->runtime_flags);
 308	memset(&range, 0, sizeof(range));
 309	range.len = (u64)-1;
 310	range.start = defrag->last_offset;
 311
 312	sb_start_write(fs_info->sb);
 313	num_defrag = btrfs_defrag_file(inode, NULL, &range, defrag->transid,
 314				       BTRFS_DEFRAG_BATCH);
 315	sb_end_write(fs_info->sb);
 316	/*
 317	 * if we filled the whole defrag batch, there
 318	 * must be more work to do.  Queue this defrag
 319	 * again
 320	 */
 321	if (num_defrag == BTRFS_DEFRAG_BATCH) {
 322		defrag->last_offset = range.start;
 323		btrfs_requeue_inode_defrag(BTRFS_I(inode), defrag);
 324	} else if (defrag->last_offset && !defrag->cycled) {
 325		/*
 326		 * we didn't fill our defrag batch, but
 327		 * we didn't start at zero.  Make sure we loop
 328		 * around to the start of the file.
 329		 */
 330		defrag->last_offset = 0;
 331		defrag->cycled = 1;
 332		btrfs_requeue_inode_defrag(BTRFS_I(inode), defrag);
 333	} else {
 334		kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
 335	}
 336
 337	iput(inode);
 338	return 0;
 339cleanup:
 340	srcu_read_unlock(&fs_info->subvol_srcu, index);
 341	kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
 342	return ret;
 343}
 344
 345/*
 346 * run through the list of inodes in the FS that need
 347 * defragging
 348 */
 349int btrfs_run_defrag_inodes(struct btrfs_fs_info *fs_info)
 350{
 351	struct inode_defrag *defrag;
 352	u64 first_ino = 0;
 353	u64 root_objectid = 0;
 354
 355	atomic_inc(&fs_info->defrag_running);
 356	while (1) {
 357		/* Pause the auto defragger. */
 358		if (test_bit(BTRFS_FS_STATE_REMOUNTING,
 359			     &fs_info->fs_state))
 360			break;
 361
 362		if (!__need_auto_defrag(fs_info))
 363			break;
 364
 365		/* find an inode to defrag */
 366		defrag = btrfs_pick_defrag_inode(fs_info, root_objectid,
 367						 first_ino);
 368		if (!defrag) {
 369			if (root_objectid || first_ino) {
 370				root_objectid = 0;
 371				first_ino = 0;
 372				continue;
 373			} else {
 374				break;
 375			}
 376		}
 377
 378		first_ino = defrag->ino + 1;
 379		root_objectid = defrag->root;
 380
 381		__btrfs_run_defrag_inode(fs_info, defrag);
 382	}
 383	atomic_dec(&fs_info->defrag_running);
 384
 385	/*
 386	 * during unmount, we use the transaction_wait queue to
 387	 * wait for the defragger to stop
 388	 */
 389	wake_up(&fs_info->transaction_wait);
 390	return 0;
 391}
 392
 393/* simple helper to fault in pages and copy.  This should go away
 394 * and be replaced with calls into generic code.
 395 */
 396static noinline int btrfs_copy_from_user(loff_t pos, size_t write_bytes,
 397					 struct page **prepared_pages,
 398					 struct iov_iter *i)
 399{
 400	size_t copied = 0;
 401	size_t total_copied = 0;
 402	int pg = 0;
 403	int offset = offset_in_page(pos);
 404
 405	while (write_bytes > 0) {
 406		size_t count = min_t(size_t,
 407				     PAGE_SIZE - offset, write_bytes);
 408		struct page *page = prepared_pages[pg];
 409		/*
 410		 * Copy data from userspace to the current page
 411		 */
 412		copied = iov_iter_copy_from_user_atomic(page, i, offset, count);
 413
 414		/* Flush processor's dcache for this page */
 415		flush_dcache_page(page);
 416
 417		/*
 418		 * if we get a partial write, we can end up with
 419		 * partially up to date pages.  These add
 420		 * a lot of complexity, so make sure they don't
 421		 * happen by forcing this copy to be retried.
 422		 *
 423		 * The rest of the btrfs_file_write code will fall
 424		 * back to page at a time copies after we return 0.
 425		 */
 426		if (!PageUptodate(page) && copied < count)
 427			copied = 0;
 428
 429		iov_iter_advance(i, copied);
 430		write_bytes -= copied;
 431		total_copied += copied;
 432
 433		/* Return to btrfs_file_write_iter to fault page */
 434		if (unlikely(copied == 0))
 435			break;
 436
 437		if (copied < PAGE_SIZE - offset) {
 438			offset += copied;
 439		} else {
 440			pg++;
 441			offset = 0;
 442		}
 443	}
 444	return total_copied;
 445}
 446
 447/*
 448 * unlocks pages after btrfs_file_write is done with them
 449 */
 450static void btrfs_drop_pages(struct page **pages, size_t num_pages)
 451{
 452	size_t i;
 453	for (i = 0; i < num_pages; i++) {
 454		/* page checked is some magic around finding pages that
 455		 * have been modified without going through btrfs_set_page_dirty
 456		 * clear it here. There should be no need to mark the pages
 457		 * accessed as prepare_pages should have marked them accessed
 458		 * in prepare_pages via find_or_create_page()
 459		 */
 460		ClearPageChecked(pages[i]);
 461		unlock_page(pages[i]);
 462		put_page(pages[i]);
 463	}
 464}
 465
 466static int btrfs_find_new_delalloc_bytes(struct btrfs_inode *inode,
 467					 const u64 start,
 468					 const u64 len,
 469					 struct extent_state **cached_state)
 470{
 471	u64 search_start = start;
 472	const u64 end = start + len - 1;
 473
 474	while (search_start < end) {
 475		const u64 search_len = end - search_start + 1;
 476		struct extent_map *em;
 477		u64 em_len;
 478		int ret = 0;
 479
 480		em = btrfs_get_extent(inode, NULL, 0, search_start,
 481				      search_len, 0);
 482		if (IS_ERR(em))
 483			return PTR_ERR(em);
 484
 485		if (em->block_start != EXTENT_MAP_HOLE)
 486			goto next;
 487
 488		em_len = em->len;
 489		if (em->start < search_start)
 490			em_len -= search_start - em->start;
 491		if (em_len > search_len)
 492			em_len = search_len;
 493
 494		ret = set_extent_bit(&inode->io_tree, search_start,
 495				     search_start + em_len - 1,
 496				     EXTENT_DELALLOC_NEW,
 497				     NULL, cached_state, GFP_NOFS);
 498next:
 499		search_start = extent_map_end(em);
 500		free_extent_map(em);
 501		if (ret)
 502			return ret;
 503	}
 504	return 0;
 505}
 506
 507/*
 508 * after copy_from_user, pages need to be dirtied and we need to make
 509 * sure holes are created between the current EOF and the start of
 510 * any next extents (if required).
 511 *
 512 * this also makes the decision about creating an inline extent vs
 513 * doing real data extents, marking pages dirty and delalloc as required.
 514 */
 515int btrfs_dirty_pages(struct inode *inode, struct page **pages,
 516		      size_t num_pages, loff_t pos, size_t write_bytes,
 517		      struct extent_state **cached)
 
 518{
 519	struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
 520	int err = 0;
 521	int i;
 522	u64 num_bytes;
 523	u64 start_pos;
 524	u64 end_of_last_block;
 525	u64 end_pos = pos + write_bytes;
 526	loff_t isize = i_size_read(inode);
 527	unsigned int extra_bits = 0;
 528
 529	start_pos = pos & ~((u64) fs_info->sectorsize - 1);
 530	num_bytes = round_up(write_bytes + pos - start_pos,
 531			     fs_info->sectorsize);
 532
 533	end_of_last_block = start_pos + num_bytes - 1;
 534
 535	/*
 536	 * The pages may have already been dirty, clear out old accounting so
 537	 * we can set things up properly
 538	 */
 539	clear_extent_bit(&BTRFS_I(inode)->io_tree, start_pos, end_of_last_block,
 540			 EXTENT_DELALLOC | EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG,
 541			 0, 0, cached);
 542
 543	if (!btrfs_is_free_space_inode(BTRFS_I(inode))) {
 544		if (start_pos >= isize &&
 545		    !(BTRFS_I(inode)->flags & BTRFS_INODE_PREALLOC)) {
 546			/*
 547			 * There can't be any extents following eof in this case
 548			 * so just set the delalloc new bit for the range
 549			 * directly.
 550			 */
 551			extra_bits |= EXTENT_DELALLOC_NEW;
 552		} else {
 553			err = btrfs_find_new_delalloc_bytes(BTRFS_I(inode),
 554							    start_pos,
 555							    num_bytes, cached);
 556			if (err)
 557				return err;
 558		}
 559	}
 560
 561	err = btrfs_set_extent_delalloc(inode, start_pos, end_of_last_block,
 562					extra_bits, cached);
 563	if (err)
 564		return err;
 565
 566	for (i = 0; i < num_pages; i++) {
 567		struct page *p = pages[i];
 568		SetPageUptodate(p);
 569		ClearPageChecked(p);
 570		set_page_dirty(p);
 571	}
 572
 573	/*
 574	 * we've only changed i_size in ram, and we haven't updated
 575	 * the disk i_size.  There is no need to log the inode
 576	 * at this time.
 577	 */
 578	if (end_pos > isize)
 579		i_size_write(inode, end_pos);
 580	return 0;
 581}
 582
 583/*
 584 * this drops all the extents in the cache that intersect the range
 585 * [start, end].  Existing extents are split as required.
 586 */
 587void btrfs_drop_extent_cache(struct btrfs_inode *inode, u64 start, u64 end,
 588			     int skip_pinned)
 589{
 590	struct extent_map *em;
 591	struct extent_map *split = NULL;
 592	struct extent_map *split2 = NULL;
 593	struct extent_map_tree *em_tree = &inode->extent_tree;
 594	u64 len = end - start + 1;
 595	u64 gen;
 596	int ret;
 597	int testend = 1;
 598	unsigned long flags;
 599	int compressed = 0;
 600	bool modified;
 601
 602	WARN_ON(end < start);
 603	if (end == (u64)-1) {
 604		len = (u64)-1;
 605		testend = 0;
 606	}
 607	while (1) {
 608		int no_splits = 0;
 609
 610		modified = false;
 611		if (!split)
 612			split = alloc_extent_map();
 613		if (!split2)
 614			split2 = alloc_extent_map();
 615		if (!split || !split2)
 616			no_splits = 1;
 617
 618		write_lock(&em_tree->lock);
 619		em = lookup_extent_mapping(em_tree, start, len);
 620		if (!em) {
 621			write_unlock(&em_tree->lock);
 622			break;
 623		}
 624		flags = em->flags;
 625		gen = em->generation;
 626		if (skip_pinned && test_bit(EXTENT_FLAG_PINNED, &em->flags)) {
 627			if (testend && em->start + em->len >= start + len) {
 628				free_extent_map(em);
 629				write_unlock(&em_tree->lock);
 630				break;
 631			}
 632			start = em->start + em->len;
 633			if (testend)
 634				len = start + len - (em->start + em->len);
 635			free_extent_map(em);
 636			write_unlock(&em_tree->lock);
 637			continue;
 638		}
 639		compressed = test_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
 640		clear_bit(EXTENT_FLAG_PINNED, &em->flags);
 641		clear_bit(EXTENT_FLAG_LOGGING, &flags);
 642		modified = !list_empty(&em->list);
 643		if (no_splits)
 644			goto next;
 645
 646		if (em->start < start) {
 647			split->start = em->start;
 648			split->len = start - em->start;
 649
 650			if (em->block_start < EXTENT_MAP_LAST_BYTE) {
 651				split->orig_start = em->orig_start;
 652				split->block_start = em->block_start;
 653
 654				if (compressed)
 655					split->block_len = em->block_len;
 656				else
 657					split->block_len = split->len;
 658				split->orig_block_len = max(split->block_len,
 659						em->orig_block_len);
 660				split->ram_bytes = em->ram_bytes;
 661			} else {
 662				split->orig_start = split->start;
 663				split->block_len = 0;
 664				split->block_start = em->block_start;
 665				split->orig_block_len = 0;
 666				split->ram_bytes = split->len;
 667			}
 668
 669			split->generation = gen;
 670			split->bdev = em->bdev;
 671			split->flags = flags;
 672			split->compress_type = em->compress_type;
 673			replace_extent_mapping(em_tree, em, split, modified);
 674			free_extent_map(split);
 675			split = split2;
 676			split2 = NULL;
 677		}
 678		if (testend && em->start + em->len > start + len) {
 679			u64 diff = start + len - em->start;
 680
 681			split->start = start + len;
 682			split->len = em->start + em->len - (start + len);
 683			split->bdev = em->bdev;
 684			split->flags = flags;
 685			split->compress_type = em->compress_type;
 686			split->generation = gen;
 687
 688			if (em->block_start < EXTENT_MAP_LAST_BYTE) {
 689				split->orig_block_len = max(em->block_len,
 690						    em->orig_block_len);
 691
 692				split->ram_bytes = em->ram_bytes;
 693				if (compressed) {
 694					split->block_len = em->block_len;
 695					split->block_start = em->block_start;
 696					split->orig_start = em->orig_start;
 697				} else {
 698					split->block_len = split->len;
 699					split->block_start = em->block_start
 700						+ diff;
 701					split->orig_start = em->orig_start;
 702				}
 703			} else {
 704				split->ram_bytes = split->len;
 705				split->orig_start = split->start;
 706				split->block_len = 0;
 707				split->block_start = em->block_start;
 708				split->orig_block_len = 0;
 709			}
 710
 711			if (extent_map_in_tree(em)) {
 712				replace_extent_mapping(em_tree, em, split,
 713						       modified);
 714			} else {
 715				ret = add_extent_mapping(em_tree, split,
 716							 modified);
 717				ASSERT(ret == 0); /* Logic error */
 718			}
 719			free_extent_map(split);
 720			split = NULL;
 721		}
 722next:
 723		if (extent_map_in_tree(em))
 724			remove_extent_mapping(em_tree, em);
 725		write_unlock(&em_tree->lock);
 726
 727		/* once for us */
 728		free_extent_map(em);
 729		/* once for the tree*/
 730		free_extent_map(em);
 731	}
 732	if (split)
 733		free_extent_map(split);
 734	if (split2)
 735		free_extent_map(split2);
 736}
 737
 738/*
 739 * this is very complex, but the basic idea is to drop all extents
 740 * in the range start - end.  hint_block is filled in with a block number
 741 * that would be a good hint to the block allocator for this file.
 742 *
 743 * If an extent intersects the range but is not entirely inside the range
 744 * it is either truncated or split.  Anything entirely inside the range
 745 * is deleted from the tree.
 746 */
 747int __btrfs_drop_extents(struct btrfs_trans_handle *trans,
 748			 struct btrfs_root *root, struct inode *inode,
 749			 struct btrfs_path *path, u64 start, u64 end,
 750			 u64 *drop_end, int drop_cache,
 751			 int replace_extent,
 752			 u32 extent_item_size,
 753			 int *key_inserted)
 754{
 755	struct btrfs_fs_info *fs_info = root->fs_info;
 756	struct extent_buffer *leaf;
 757	struct btrfs_file_extent_item *fi;
 758	struct btrfs_ref ref = { 0 };
 759	struct btrfs_key key;
 760	struct btrfs_key new_key;
 761	u64 ino = btrfs_ino(BTRFS_I(inode));
 762	u64 search_start = start;
 763	u64 disk_bytenr = 0;
 764	u64 num_bytes = 0;
 765	u64 extent_offset = 0;
 766	u64 extent_end = 0;
 767	u64 last_end = start;
 768	int del_nr = 0;
 769	int del_slot = 0;
 770	int extent_type;
 771	int recow;
 772	int ret;
 773	int modify_tree = -1;
 774	int update_refs;
 775	int found = 0;
 776	int leafs_visited = 0;
 777
 778	if (drop_cache)
 779		btrfs_drop_extent_cache(BTRFS_I(inode), start, end - 1, 0);
 780
 781	if (start >= BTRFS_I(inode)->disk_i_size && !replace_extent)
 782		modify_tree = 0;
 783
 784	update_refs = (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
 785		       root == fs_info->tree_root);
 786	while (1) {
 787		recow = 0;
 788		ret = btrfs_lookup_file_extent(trans, root, path, ino,
 789					       search_start, modify_tree);
 790		if (ret < 0)
 791			break;
 792		if (ret > 0 && path->slots[0] > 0 && search_start == start) {
 793			leaf = path->nodes[0];
 794			btrfs_item_key_to_cpu(leaf, &key, path->slots[0] - 1);
 795			if (key.objectid == ino &&
 796			    key.type == BTRFS_EXTENT_DATA_KEY)
 797				path->slots[0]--;
 798		}
 799		ret = 0;
 800		leafs_visited++;
 801next_slot:
 802		leaf = path->nodes[0];
 803		if (path->slots[0] >= btrfs_header_nritems(leaf)) {
 804			BUG_ON(del_nr > 0);
 805			ret = btrfs_next_leaf(root, path);
 806			if (ret < 0)
 807				break;
 808			if (ret > 0) {
 809				ret = 0;
 810				break;
 811			}
 812			leafs_visited++;
 813			leaf = path->nodes[0];
 814			recow = 1;
 815		}
 816
 817		btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 818
 819		if (key.objectid > ino)
 820			break;
 821		if (WARN_ON_ONCE(key.objectid < ino) ||
 822		    key.type < BTRFS_EXTENT_DATA_KEY) {
 823			ASSERT(del_nr == 0);
 824			path->slots[0]++;
 825			goto next_slot;
 826		}
 827		if (key.type > BTRFS_EXTENT_DATA_KEY || key.offset >= end)
 828			break;
 829
 830		fi = btrfs_item_ptr(leaf, path->slots[0],
 831				    struct btrfs_file_extent_item);
 832		extent_type = btrfs_file_extent_type(leaf, fi);
 833
 834		if (extent_type == BTRFS_FILE_EXTENT_REG ||
 835		    extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
 836			disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
 837			num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
 838			extent_offset = btrfs_file_extent_offset(leaf, fi);
 839			extent_end = key.offset +
 840				btrfs_file_extent_num_bytes(leaf, fi);
 841		} else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
 842			extent_end = key.offset +
 843				btrfs_file_extent_ram_bytes(leaf, fi);
 
 844		} else {
 845			/* can't happen */
 846			BUG();
 847		}
 848
 849		/*
 850		 * Don't skip extent items representing 0 byte lengths. They
 851		 * used to be created (bug) if while punching holes we hit
 852		 * -ENOSPC condition. So if we find one here, just ensure we
 853		 * delete it, otherwise we would insert a new file extent item
 854		 * with the same key (offset) as that 0 bytes length file
 855		 * extent item in the call to setup_items_for_insert() later
 856		 * in this function.
 857		 */
 858		if (extent_end == key.offset && extent_end >= search_start) {
 859			last_end = extent_end;
 860			goto delete_extent_item;
 861		}
 862
 863		if (extent_end <= search_start) {
 864			path->slots[0]++;
 865			goto next_slot;
 866		}
 867
 868		found = 1;
 869		search_start = max(key.offset, start);
 870		if (recow || !modify_tree) {
 871			modify_tree = -1;
 872			btrfs_release_path(path);
 873			continue;
 874		}
 875
 876		/*
 877		 *     | - range to drop - |
 878		 *  | -------- extent -------- |
 879		 */
 880		if (start > key.offset && end < extent_end) {
 881			BUG_ON(del_nr > 0);
 882			if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
 883				ret = -EOPNOTSUPP;
 884				break;
 885			}
 886
 887			memcpy(&new_key, &key, sizeof(new_key));
 888			new_key.offset = start;
 889			ret = btrfs_duplicate_item(trans, root, path,
 890						   &new_key);
 891			if (ret == -EAGAIN) {
 892				btrfs_release_path(path);
 893				continue;
 894			}
 895			if (ret < 0)
 896				break;
 897
 898			leaf = path->nodes[0];
 899			fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
 900					    struct btrfs_file_extent_item);
 901			btrfs_set_file_extent_num_bytes(leaf, fi,
 902							start - key.offset);
 903
 904			fi = btrfs_item_ptr(leaf, path->slots[0],
 905					    struct btrfs_file_extent_item);
 906
 907			extent_offset += start - key.offset;
 908			btrfs_set_file_extent_offset(leaf, fi, extent_offset);
 909			btrfs_set_file_extent_num_bytes(leaf, fi,
 910							extent_end - start);
 911			btrfs_mark_buffer_dirty(leaf);
 912
 913			if (update_refs && disk_bytenr > 0) {
 914				btrfs_init_generic_ref(&ref,
 915						BTRFS_ADD_DELAYED_REF,
 916						disk_bytenr, num_bytes, 0);
 917				btrfs_init_data_ref(&ref,
 918						root->root_key.objectid,
 919						new_key.objectid,
 920						start - extent_offset);
 921				ret = btrfs_inc_extent_ref(trans, &ref);
 922				BUG_ON(ret); /* -ENOMEM */
 923			}
 924			key.offset = start;
 925		}
 926		/*
 927		 * From here on out we will have actually dropped something, so
 928		 * last_end can be updated.
 929		 */
 930		last_end = extent_end;
 931
 932		/*
 933		 *  | ---- range to drop ----- |
 934		 *      | -------- extent -------- |
 935		 */
 936		if (start <= key.offset && end < extent_end) {
 937			if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
 938				ret = -EOPNOTSUPP;
 939				break;
 940			}
 941
 942			memcpy(&new_key, &key, sizeof(new_key));
 943			new_key.offset = end;
 944			btrfs_set_item_key_safe(fs_info, path, &new_key);
 945
 946			extent_offset += end - key.offset;
 947			btrfs_set_file_extent_offset(leaf, fi, extent_offset);
 948			btrfs_set_file_extent_num_bytes(leaf, fi,
 949							extent_end - end);
 950			btrfs_mark_buffer_dirty(leaf);
 951			if (update_refs && disk_bytenr > 0)
 952				inode_sub_bytes(inode, end - key.offset);
 953			break;
 954		}
 955
 956		search_start = extent_end;
 957		/*
 958		 *       | ---- range to drop ----- |
 959		 *  | -------- extent -------- |
 960		 */
 961		if (start > key.offset && end >= extent_end) {
 962			BUG_ON(del_nr > 0);
 963			if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
 964				ret = -EOPNOTSUPP;
 965				break;
 966			}
 967
 968			btrfs_set_file_extent_num_bytes(leaf, fi,
 969							start - key.offset);
 970			btrfs_mark_buffer_dirty(leaf);
 971			if (update_refs && disk_bytenr > 0)
 972				inode_sub_bytes(inode, extent_end - start);
 973			if (end == extent_end)
 974				break;
 975
 976			path->slots[0]++;
 977			goto next_slot;
 978		}
 979
 980		/*
 981		 *  | ---- range to drop ----- |
 982		 *    | ------ extent ------ |
 983		 */
 984		if (start <= key.offset && end >= extent_end) {
 985delete_extent_item:
 986			if (del_nr == 0) {
 987				del_slot = path->slots[0];
 988				del_nr = 1;
 989			} else {
 990				BUG_ON(del_slot + del_nr != path->slots[0]);
 991				del_nr++;
 992			}
 993
 994			if (update_refs &&
 995			    extent_type == BTRFS_FILE_EXTENT_INLINE) {
 996				inode_sub_bytes(inode,
 997						extent_end - key.offset);
 998				extent_end = ALIGN(extent_end,
 999						   fs_info->sectorsize);
1000			} else if (update_refs && disk_bytenr > 0) {
1001				btrfs_init_generic_ref(&ref,
1002						BTRFS_DROP_DELAYED_REF,
1003						disk_bytenr, num_bytes, 0);
1004				btrfs_init_data_ref(&ref,
1005						root->root_key.objectid,
1006						key.objectid,
1007						key.offset - extent_offset);
1008				ret = btrfs_free_extent(trans, &ref);
1009				BUG_ON(ret); /* -ENOMEM */
1010				inode_sub_bytes(inode,
1011						extent_end - key.offset);
1012			}
1013
1014			if (end == extent_end)
1015				break;
1016
1017			if (path->slots[0] + 1 < btrfs_header_nritems(leaf)) {
1018				path->slots[0]++;
1019				goto next_slot;
1020			}
1021
1022			ret = btrfs_del_items(trans, root, path, del_slot,
1023					      del_nr);
1024			if (ret) {
1025				btrfs_abort_transaction(trans, ret);
1026				break;
1027			}
1028
1029			del_nr = 0;
1030			del_slot = 0;
1031
1032			btrfs_release_path(path);
1033			continue;
1034		}
1035
1036		BUG();
1037	}
1038
1039	if (!ret && del_nr > 0) {
1040		/*
1041		 * Set path->slots[0] to first slot, so that after the delete
1042		 * if items are move off from our leaf to its immediate left or
1043		 * right neighbor leafs, we end up with a correct and adjusted
1044		 * path->slots[0] for our insertion (if replace_extent != 0).
1045		 */
1046		path->slots[0] = del_slot;
1047		ret = btrfs_del_items(trans, root, path, del_slot, del_nr);
1048		if (ret)
1049			btrfs_abort_transaction(trans, ret);
1050	}
1051
1052	leaf = path->nodes[0];
1053	/*
1054	 * If btrfs_del_items() was called, it might have deleted a leaf, in
1055	 * which case it unlocked our path, so check path->locks[0] matches a
1056	 * write lock.
1057	 */
1058	if (!ret && replace_extent && leafs_visited == 1 &&
1059	    (path->locks[0] == BTRFS_WRITE_LOCK_BLOCKING ||
1060	     path->locks[0] == BTRFS_WRITE_LOCK) &&
1061	    btrfs_leaf_free_space(leaf) >=
1062	    sizeof(struct btrfs_item) + extent_item_size) {
1063
1064		key.objectid = ino;
1065		key.type = BTRFS_EXTENT_DATA_KEY;
1066		key.offset = start;
1067		if (!del_nr && path->slots[0] < btrfs_header_nritems(leaf)) {
1068			struct btrfs_key slot_key;
1069
1070			btrfs_item_key_to_cpu(leaf, &slot_key, path->slots[0]);
1071			if (btrfs_comp_cpu_keys(&key, &slot_key) > 0)
1072				path->slots[0]++;
1073		}
1074		setup_items_for_insert(root, path, &key,
1075				       &extent_item_size,
1076				       extent_item_size,
1077				       sizeof(struct btrfs_item) +
1078				       extent_item_size, 1);
1079		*key_inserted = 1;
1080	}
1081
1082	if (!replace_extent || !(*key_inserted))
1083		btrfs_release_path(path);
1084	if (drop_end)
1085		*drop_end = found ? min(end, last_end) : end;
1086	return ret;
1087}
1088
1089int btrfs_drop_extents(struct btrfs_trans_handle *trans,
1090		       struct btrfs_root *root, struct inode *inode, u64 start,
1091		       u64 end, int drop_cache)
1092{
1093	struct btrfs_path *path;
1094	int ret;
1095
1096	path = btrfs_alloc_path();
1097	if (!path)
1098		return -ENOMEM;
1099	ret = __btrfs_drop_extents(trans, root, inode, path, start, end, NULL,
1100				   drop_cache, 0, 0, NULL);
1101	btrfs_free_path(path);
1102	return ret;
1103}
1104
1105static int extent_mergeable(struct extent_buffer *leaf, int slot,
1106			    u64 objectid, u64 bytenr, u64 orig_offset,
1107			    u64 *start, u64 *end)
1108{
1109	struct btrfs_file_extent_item *fi;
1110	struct btrfs_key key;
1111	u64 extent_end;
1112
1113	if (slot < 0 || slot >= btrfs_header_nritems(leaf))
1114		return 0;
1115
1116	btrfs_item_key_to_cpu(leaf, &key, slot);
1117	if (key.objectid != objectid || key.type != BTRFS_EXTENT_DATA_KEY)
1118		return 0;
1119
1120	fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
1121	if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG ||
1122	    btrfs_file_extent_disk_bytenr(leaf, fi) != bytenr ||
1123	    btrfs_file_extent_offset(leaf, fi) != key.offset - orig_offset ||
1124	    btrfs_file_extent_compression(leaf, fi) ||
1125	    btrfs_file_extent_encryption(leaf, fi) ||
1126	    btrfs_file_extent_other_encoding(leaf, fi))
1127		return 0;
1128
1129	extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
1130	if ((*start && *start != key.offset) || (*end && *end != extent_end))
1131		return 0;
1132
1133	*start = key.offset;
1134	*end = extent_end;
1135	return 1;
1136}
1137
1138/*
1139 * Mark extent in the range start - end as written.
1140 *
1141 * This changes extent type from 'pre-allocated' to 'regular'. If only
1142 * part of extent is marked as written, the extent will be split into
1143 * two or three.
1144 */
1145int btrfs_mark_extent_written(struct btrfs_trans_handle *trans,
1146			      struct btrfs_inode *inode, u64 start, u64 end)
1147{
1148	struct btrfs_fs_info *fs_info = trans->fs_info;
1149	struct btrfs_root *root = inode->root;
1150	struct extent_buffer *leaf;
1151	struct btrfs_path *path;
1152	struct btrfs_file_extent_item *fi;
1153	struct btrfs_ref ref = { 0 };
1154	struct btrfs_key key;
1155	struct btrfs_key new_key;
1156	u64 bytenr;
1157	u64 num_bytes;
1158	u64 extent_end;
1159	u64 orig_offset;
1160	u64 other_start;
1161	u64 other_end;
1162	u64 split;
1163	int del_nr = 0;
1164	int del_slot = 0;
1165	int recow;
1166	int ret;
1167	u64 ino = btrfs_ino(inode);
1168
1169	path = btrfs_alloc_path();
1170	if (!path)
1171		return -ENOMEM;
1172again:
1173	recow = 0;
1174	split = start;
1175	key.objectid = ino;
1176	key.type = BTRFS_EXTENT_DATA_KEY;
1177	key.offset = split;
1178
1179	ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1180	if (ret < 0)
1181		goto out;
1182	if (ret > 0 && path->slots[0] > 0)
1183		path->slots[0]--;
1184
1185	leaf = path->nodes[0];
1186	btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1187	if (key.objectid != ino ||
1188	    key.type != BTRFS_EXTENT_DATA_KEY) {
1189		ret = -EINVAL;
1190		btrfs_abort_transaction(trans, ret);
1191		goto out;
1192	}
1193	fi = btrfs_item_ptr(leaf, path->slots[0],
1194			    struct btrfs_file_extent_item);
1195	if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_PREALLOC) {
1196		ret = -EINVAL;
1197		btrfs_abort_transaction(trans, ret);
1198		goto out;
1199	}
1200	extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
1201	if (key.offset > start || extent_end < end) {
1202		ret = -EINVAL;
1203		btrfs_abort_transaction(trans, ret);
1204		goto out;
1205	}
1206
1207	bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1208	num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
1209	orig_offset = key.offset - btrfs_file_extent_offset(leaf, fi);
1210	memcpy(&new_key, &key, sizeof(new_key));
1211
1212	if (start == key.offset && end < extent_end) {
1213		other_start = 0;
1214		other_end = start;
1215		if (extent_mergeable(leaf, path->slots[0] - 1,
1216				     ino, bytenr, orig_offset,
1217				     &other_start, &other_end)) {
1218			new_key.offset = end;
1219			btrfs_set_item_key_safe(fs_info, path, &new_key);
1220			fi = btrfs_item_ptr(leaf, path->slots[0],
1221					    struct btrfs_file_extent_item);
1222			btrfs_set_file_extent_generation(leaf, fi,
1223							 trans->transid);
1224			btrfs_set_file_extent_num_bytes(leaf, fi,
1225							extent_end - end);
1226			btrfs_set_file_extent_offset(leaf, fi,
1227						     end - orig_offset);
1228			fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
1229					    struct btrfs_file_extent_item);
1230			btrfs_set_file_extent_generation(leaf, fi,
1231							 trans->transid);
1232			btrfs_set_file_extent_num_bytes(leaf, fi,
1233							end - other_start);
1234			btrfs_mark_buffer_dirty(leaf);
1235			goto out;
1236		}
1237	}
1238
1239	if (start > key.offset && end == extent_end) {
1240		other_start = end;
1241		other_end = 0;
1242		if (extent_mergeable(leaf, path->slots[0] + 1,
1243				     ino, bytenr, orig_offset,
1244				     &other_start, &other_end)) {
1245			fi = btrfs_item_ptr(leaf, path->slots[0],
1246					    struct btrfs_file_extent_item);
1247			btrfs_set_file_extent_num_bytes(leaf, fi,
1248							start - key.offset);
1249			btrfs_set_file_extent_generation(leaf, fi,
1250							 trans->transid);
1251			path->slots[0]++;
1252			new_key.offset = start;
1253			btrfs_set_item_key_safe(fs_info, path, &new_key);
1254
1255			fi = btrfs_item_ptr(leaf, path->slots[0],
1256					    struct btrfs_file_extent_item);
1257			btrfs_set_file_extent_generation(leaf, fi,
1258							 trans->transid);
1259			btrfs_set_file_extent_num_bytes(leaf, fi,
1260							other_end - start);
1261			btrfs_set_file_extent_offset(leaf, fi,
1262						     start - orig_offset);
1263			btrfs_mark_buffer_dirty(leaf);
1264			goto out;
1265		}
1266	}
1267
1268	while (start > key.offset || end < extent_end) {
1269		if (key.offset == start)
1270			split = end;
1271
1272		new_key.offset = split;
1273		ret = btrfs_duplicate_item(trans, root, path, &new_key);
1274		if (ret == -EAGAIN) {
1275			btrfs_release_path(path);
1276			goto again;
1277		}
1278		if (ret < 0) {
1279			btrfs_abort_transaction(trans, ret);
1280			goto out;
1281		}
1282
1283		leaf = path->nodes[0];
1284		fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
1285				    struct btrfs_file_extent_item);
1286		btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1287		btrfs_set_file_extent_num_bytes(leaf, fi,
1288						split - key.offset);
1289
1290		fi = btrfs_item_ptr(leaf, path->slots[0],
1291				    struct btrfs_file_extent_item);
1292
1293		btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1294		btrfs_set_file_extent_offset(leaf, fi, split - orig_offset);
1295		btrfs_set_file_extent_num_bytes(leaf, fi,
1296						extent_end - split);
1297		btrfs_mark_buffer_dirty(leaf);
1298
1299		btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF, bytenr,
1300				       num_bytes, 0);
1301		btrfs_init_data_ref(&ref, root->root_key.objectid, ino,
1302				    orig_offset);
1303		ret = btrfs_inc_extent_ref(trans, &ref);
1304		if (ret) {
1305			btrfs_abort_transaction(trans, ret);
1306			goto out;
1307		}
1308
1309		if (split == start) {
1310			key.offset = start;
1311		} else {
1312			if (start != key.offset) {
1313				ret = -EINVAL;
1314				btrfs_abort_transaction(trans, ret);
1315				goto out;
1316			}
1317			path->slots[0]--;
1318			extent_end = end;
1319		}
1320		recow = 1;
1321	}
1322
1323	other_start = end;
1324	other_end = 0;
1325	btrfs_init_generic_ref(&ref, BTRFS_DROP_DELAYED_REF, bytenr,
1326			       num_bytes, 0);
1327	btrfs_init_data_ref(&ref, root->root_key.objectid, ino, orig_offset);
1328	if (extent_mergeable(leaf, path->slots[0] + 1,
1329			     ino, bytenr, orig_offset,
1330			     &other_start, &other_end)) {
1331		if (recow) {
1332			btrfs_release_path(path);
1333			goto again;
1334		}
1335		extent_end = other_end;
1336		del_slot = path->slots[0] + 1;
1337		del_nr++;
1338		ret = btrfs_free_extent(trans, &ref);
1339		if (ret) {
1340			btrfs_abort_transaction(trans, ret);
1341			goto out;
1342		}
1343	}
1344	other_start = 0;
1345	other_end = start;
1346	if (extent_mergeable(leaf, path->slots[0] - 1,
1347			     ino, bytenr, orig_offset,
1348			     &other_start, &other_end)) {
1349		if (recow) {
1350			btrfs_release_path(path);
1351			goto again;
1352		}
1353		key.offset = other_start;
1354		del_slot = path->slots[0];
1355		del_nr++;
1356		ret = btrfs_free_extent(trans, &ref);
1357		if (ret) {
1358			btrfs_abort_transaction(trans, ret);
1359			goto out;
1360		}
1361	}
1362	if (del_nr == 0) {
1363		fi = btrfs_item_ptr(leaf, path->slots[0],
1364			   struct btrfs_file_extent_item);
1365		btrfs_set_file_extent_type(leaf, fi,
1366					   BTRFS_FILE_EXTENT_REG);
1367		btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1368		btrfs_mark_buffer_dirty(leaf);
1369	} else {
1370		fi = btrfs_item_ptr(leaf, del_slot - 1,
1371			   struct btrfs_file_extent_item);
1372		btrfs_set_file_extent_type(leaf, fi,
1373					   BTRFS_FILE_EXTENT_REG);
1374		btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1375		btrfs_set_file_extent_num_bytes(leaf, fi,
1376						extent_end - key.offset);
1377		btrfs_mark_buffer_dirty(leaf);
1378
1379		ret = btrfs_del_items(trans, root, path, del_slot, del_nr);
1380		if (ret < 0) {
1381			btrfs_abort_transaction(trans, ret);
1382			goto out;
1383		}
1384	}
1385out:
1386	btrfs_free_path(path);
1387	return 0;
1388}
1389
1390/*
1391 * on error we return an unlocked page and the error value
1392 * on success we return a locked page and 0
1393 */
1394static int prepare_uptodate_page(struct inode *inode,
1395				 struct page *page, u64 pos,
1396				 bool force_uptodate)
1397{
1398	int ret = 0;
1399
1400	if (((pos & (PAGE_SIZE - 1)) || force_uptodate) &&
1401	    !PageUptodate(page)) {
1402		ret = btrfs_readpage(NULL, page);
1403		if (ret)
1404			return ret;
1405		lock_page(page);
1406		if (!PageUptodate(page)) {
1407			unlock_page(page);
1408			return -EIO;
1409		}
1410		if (page->mapping != inode->i_mapping) {
1411			unlock_page(page);
1412			return -EAGAIN;
1413		}
1414	}
1415	return 0;
1416}
1417
1418/*
1419 * this just gets pages into the page cache and locks them down.
1420 */
1421static noinline int prepare_pages(struct inode *inode, struct page **pages,
1422				  size_t num_pages, loff_t pos,
1423				  size_t write_bytes, bool force_uptodate)
1424{
1425	int i;
1426	unsigned long index = pos >> PAGE_SHIFT;
1427	gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping);
1428	int err = 0;
1429	int faili;
1430
1431	for (i = 0; i < num_pages; i++) {
1432again:
1433		pages[i] = find_or_create_page(inode->i_mapping, index + i,
1434					       mask | __GFP_WRITE);
1435		if (!pages[i]) {
1436			faili = i - 1;
1437			err = -ENOMEM;
1438			goto fail;
1439		}
1440
1441		if (i == 0)
1442			err = prepare_uptodate_page(inode, pages[i], pos,
1443						    force_uptodate);
1444		if (!err && i == num_pages - 1)
1445			err = prepare_uptodate_page(inode, pages[i],
1446						    pos + write_bytes, false);
1447		if (err) {
1448			put_page(pages[i]);
1449			if (err == -EAGAIN) {
1450				err = 0;
1451				goto again;
1452			}
1453			faili = i - 1;
1454			goto fail;
1455		}
1456		wait_on_page_writeback(pages[i]);
1457	}
1458
1459	return 0;
1460fail:
1461	while (faili >= 0) {
1462		unlock_page(pages[faili]);
1463		put_page(pages[faili]);
1464		faili--;
1465	}
1466	return err;
1467
1468}
1469
1470/*
1471 * This function locks the extent and properly waits for data=ordered extents
1472 * to finish before allowing the pages to be modified if need.
1473 *
1474 * The return value:
1475 * 1 - the extent is locked
1476 * 0 - the extent is not locked, and everything is OK
1477 * -EAGAIN - need re-prepare the pages
1478 * the other < 0 number - Something wrong happens
1479 */
1480static noinline int
1481lock_and_cleanup_extent_if_need(struct btrfs_inode *inode, struct page **pages,
1482				size_t num_pages, loff_t pos,
1483				size_t write_bytes,
1484				u64 *lockstart, u64 *lockend,
1485				struct extent_state **cached_state)
1486{
1487	struct btrfs_fs_info *fs_info = inode->root->fs_info;
1488	u64 start_pos;
1489	u64 last_pos;
1490	int i;
1491	int ret = 0;
1492
1493	start_pos = round_down(pos, fs_info->sectorsize);
1494	last_pos = start_pos
1495		+ round_up(pos + write_bytes - start_pos,
1496			   fs_info->sectorsize) - 1;
1497
1498	if (start_pos < inode->vfs_inode.i_size) {
1499		struct btrfs_ordered_extent *ordered;
1500
1501		lock_extent_bits(&inode->io_tree, start_pos, last_pos,
1502				cached_state);
1503		ordered = btrfs_lookup_ordered_range(inode, start_pos,
1504						     last_pos - start_pos + 1);
1505		if (ordered &&
1506		    ordered->file_offset + ordered->len > start_pos &&
1507		    ordered->file_offset <= last_pos) {
1508			unlock_extent_cached(&inode->io_tree, start_pos,
1509					last_pos, cached_state);
 
1510			for (i = 0; i < num_pages; i++) {
1511				unlock_page(pages[i]);
1512				put_page(pages[i]);
1513			}
1514			btrfs_start_ordered_extent(&inode->vfs_inode,
1515					ordered, 1);
1516			btrfs_put_ordered_extent(ordered);
1517			return -EAGAIN;
1518		}
1519		if (ordered)
1520			btrfs_put_ordered_extent(ordered);
1521
 
 
 
 
1522		*lockstart = start_pos;
1523		*lockend = last_pos;
1524		ret = 1;
1525	}
1526
1527	/*
1528	 * It's possible the pages are dirty right now, but we don't want
1529	 * to clean them yet because copy_from_user may catch a page fault
1530	 * and we might have to fall back to one page at a time.  If that
1531	 * happens, we'll unlock these pages and we'd have a window where
1532	 * reclaim could sneak in and drop the once-dirty page on the floor
1533	 * without writing it.
1534	 *
1535	 * We have the pages locked and the extent range locked, so there's
1536	 * no way someone can start IO on any dirty pages in this range.
1537	 *
1538	 * We'll call btrfs_dirty_pages() later on, and that will flip around
1539	 * delalloc bits and dirty the pages as required.
1540	 */
1541	for (i = 0; i < num_pages; i++) {
 
 
1542		set_page_extent_mapped(pages[i]);
1543		WARN_ON(!PageLocked(pages[i]));
1544	}
1545
1546	return ret;
1547}
1548
1549static noinline int check_can_nocow(struct btrfs_inode *inode, loff_t pos,
1550				    size_t *write_bytes)
1551{
1552	struct btrfs_fs_info *fs_info = inode->root->fs_info;
1553	struct btrfs_root *root = inode->root;
1554	u64 lockstart, lockend;
1555	u64 num_bytes;
1556	int ret;
1557
1558	ret = btrfs_start_write_no_snapshotting(root);
1559	if (!ret)
1560		return -EAGAIN;
1561
1562	lockstart = round_down(pos, fs_info->sectorsize);
1563	lockend = round_up(pos + *write_bytes,
1564			   fs_info->sectorsize) - 1;
1565
1566	btrfs_lock_and_flush_ordered_range(&inode->io_tree, inode, lockstart,
1567					   lockend, NULL);
 
 
 
 
 
 
 
 
 
1568
1569	num_bytes = lockend - lockstart + 1;
1570	ret = can_nocow_extent(&inode->vfs_inode, lockstart, &num_bytes,
1571			NULL, NULL, NULL);
1572	if (ret <= 0) {
1573		ret = 0;
1574		btrfs_end_write_no_snapshotting(root);
1575	} else {
1576		*write_bytes = min_t(size_t, *write_bytes ,
1577				     num_bytes - pos + lockstart);
1578	}
1579
1580	unlock_extent(&inode->io_tree, lockstart, lockend);
1581
1582	return ret;
1583}
1584
1585static noinline ssize_t btrfs_buffered_write(struct kiocb *iocb,
1586					       struct iov_iter *i)
 
1587{
1588	struct file *file = iocb->ki_filp;
1589	loff_t pos = iocb->ki_pos;
1590	struct inode *inode = file_inode(file);
1591	struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1592	struct btrfs_root *root = BTRFS_I(inode)->root;
1593	struct page **pages = NULL;
1594	struct extent_changeset *data_reserved = NULL;
1595	u64 release_bytes = 0;
1596	u64 lockstart;
1597	u64 lockend;
1598	size_t num_written = 0;
1599	int nrptrs;
1600	int ret = 0;
1601	bool only_release_metadata = false;
1602	bool force_page_uptodate = false;
 
1603
1604	nrptrs = min(DIV_ROUND_UP(iov_iter_count(i), PAGE_SIZE),
1605			PAGE_SIZE / (sizeof(struct page *)));
1606	nrptrs = min(nrptrs, current->nr_dirtied_pause - current->nr_dirtied);
1607	nrptrs = max(nrptrs, 8);
1608	pages = kmalloc_array(nrptrs, sizeof(struct page *), GFP_KERNEL);
1609	if (!pages)
1610		return -ENOMEM;
1611
1612	while (iov_iter_count(i) > 0) {
1613		struct extent_state *cached_state = NULL;
1614		size_t offset = offset_in_page(pos);
1615		size_t sector_offset;
1616		size_t write_bytes = min(iov_iter_count(i),
1617					 nrptrs * (size_t)PAGE_SIZE -
1618					 offset);
1619		size_t num_pages = DIV_ROUND_UP(write_bytes + offset,
1620						PAGE_SIZE);
1621		size_t reserve_bytes;
1622		size_t dirty_pages;
1623		size_t copied;
1624		size_t dirty_sectors;
1625		size_t num_sectors;
1626		int extents_locked;
1627
1628		WARN_ON(num_pages > nrptrs);
1629
1630		/*
1631		 * Fault pages before locking them in prepare_pages
1632		 * to avoid recursive lock
1633		 */
1634		if (unlikely(iov_iter_fault_in_readable(i, write_bytes))) {
1635			ret = -EFAULT;
1636			break;
1637		}
1638
1639		sector_offset = pos & (fs_info->sectorsize - 1);
1640		reserve_bytes = round_up(write_bytes + sector_offset,
1641				fs_info->sectorsize);
1642
1643		extent_changeset_release(data_reserved);
1644		ret = btrfs_check_data_free_space(inode, &data_reserved, pos,
1645						  write_bytes);
1646		if (ret < 0) {
1647			if ((BTRFS_I(inode)->flags & (BTRFS_INODE_NODATACOW |
1648						      BTRFS_INODE_PREALLOC)) &&
1649			    check_can_nocow(BTRFS_I(inode), pos,
1650					&write_bytes) > 0) {
1651				/*
1652				 * For nodata cow case, no need to reserve
1653				 * data space.
1654				 */
1655				only_release_metadata = true;
1656				/*
1657				 * our prealloc extent may be smaller than
1658				 * write_bytes, so scale down.
1659				 */
1660				num_pages = DIV_ROUND_UP(write_bytes + offset,
1661							 PAGE_SIZE);
1662				reserve_bytes = round_up(write_bytes +
1663							 sector_offset,
1664							 fs_info->sectorsize);
1665			} else {
1666				break;
1667			}
1668		}
1669
1670		WARN_ON(reserve_bytes == 0);
1671		ret = btrfs_delalloc_reserve_metadata(BTRFS_I(inode),
1672				reserve_bytes);
 
 
 
1673		if (ret) {
1674			if (!only_release_metadata)
1675				btrfs_free_reserved_data_space(inode,
1676						data_reserved, pos,
1677						write_bytes);
1678			else
1679				btrfs_end_write_no_snapshotting(root);
1680			break;
1681		}
1682
1683		release_bytes = reserve_bytes;
 
1684again:
1685		/*
1686		 * This is going to setup the pages array with the number of
1687		 * pages we want, so we don't really need to worry about the
1688		 * contents of pages from loop to loop
1689		 */
1690		ret = prepare_pages(inode, pages, num_pages,
1691				    pos, write_bytes,
1692				    force_page_uptodate);
1693		if (ret) {
1694			btrfs_delalloc_release_extents(BTRFS_I(inode),
1695						       reserve_bytes);
1696			break;
1697		}
1698
1699		extents_locked = lock_and_cleanup_extent_if_need(
1700				BTRFS_I(inode), pages,
1701				num_pages, pos, write_bytes, &lockstart,
1702				&lockend, &cached_state);
1703		if (extents_locked < 0) {
1704			if (extents_locked == -EAGAIN)
1705				goto again;
1706			btrfs_delalloc_release_extents(BTRFS_I(inode),
1707						       reserve_bytes);
1708			ret = extents_locked;
1709			break;
 
 
 
1710		}
1711
1712		copied = btrfs_copy_from_user(pos, write_bytes, pages, i);
1713
1714		num_sectors = BTRFS_BYTES_TO_BLKS(fs_info, reserve_bytes);
1715		dirty_sectors = round_up(copied + sector_offset,
1716					fs_info->sectorsize);
1717		dirty_sectors = BTRFS_BYTES_TO_BLKS(fs_info, dirty_sectors);
1718
1719		/*
1720		 * if we have trouble faulting in the pages, fall
1721		 * back to one page at a time
1722		 */
1723		if (copied < write_bytes)
1724			nrptrs = 1;
1725
1726		if (copied == 0) {
1727			force_page_uptodate = true;
1728			dirty_sectors = 0;
1729			dirty_pages = 0;
1730		} else {
1731			force_page_uptodate = false;
1732			dirty_pages = DIV_ROUND_UP(copied + offset,
1733						   PAGE_SIZE);
1734		}
1735
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1736		if (num_sectors > dirty_sectors) {
1737			/* release everything except the sectors we dirtied */
1738			release_bytes -= dirty_sectors <<
1739						fs_info->sb->s_blocksize_bits;
 
 
 
 
1740			if (only_release_metadata) {
1741				btrfs_delalloc_release_metadata(BTRFS_I(inode),
1742							release_bytes, true);
1743			} else {
1744				u64 __pos;
1745
1746				__pos = round_down(pos,
1747						   fs_info->sectorsize) +
1748					(dirty_pages << PAGE_SHIFT);
1749				btrfs_delalloc_release_space(inode,
1750						data_reserved, __pos,
1751						release_bytes, true);
1752			}
1753		}
1754
1755		release_bytes = round_up(copied + sector_offset,
1756					fs_info->sectorsize);
1757
1758		if (copied > 0)
1759			ret = btrfs_dirty_pages(inode, pages, dirty_pages,
1760						pos, copied, &cached_state);
1761
1762		/*
1763		 * If we have not locked the extent range, because the range's
1764		 * start offset is >= i_size, we might still have a non-NULL
1765		 * cached extent state, acquired while marking the extent range
1766		 * as delalloc through btrfs_dirty_pages(). Therefore free any
1767		 * possible cached extent state to avoid a memory leak.
1768		 */
1769		if (extents_locked)
1770			unlock_extent_cached(&BTRFS_I(inode)->io_tree,
1771					     lockstart, lockend, &cached_state);
1772		else
1773			free_extent_state(cached_state);
1774
1775		btrfs_delalloc_release_extents(BTRFS_I(inode), reserve_bytes);
1776		if (ret) {
1777			btrfs_drop_pages(pages, num_pages);
1778			break;
1779		}
1780
1781		release_bytes = 0;
1782		if (only_release_metadata)
1783			btrfs_end_write_no_snapshotting(root);
1784
1785		if (only_release_metadata && copied > 0) {
1786			lockstart = round_down(pos,
1787					       fs_info->sectorsize);
1788			lockend = round_up(pos + copied,
1789					   fs_info->sectorsize) - 1;
1790
1791			set_extent_bit(&BTRFS_I(inode)->io_tree, lockstart,
1792				       lockend, EXTENT_NORESERVE, NULL,
1793				       NULL, GFP_NOFS);
1794			only_release_metadata = false;
1795		}
1796
1797		btrfs_drop_pages(pages, num_pages);
1798
1799		cond_resched();
1800
1801		balance_dirty_pages_ratelimited(inode->i_mapping);
1802		if (dirty_pages < (fs_info->nodesize >> PAGE_SHIFT) + 1)
1803			btrfs_btree_balance_dirty(fs_info);
1804
1805		pos += copied;
1806		num_written += copied;
1807	}
1808
1809	kfree(pages);
1810
1811	if (release_bytes) {
1812		if (only_release_metadata) {
1813			btrfs_end_write_no_snapshotting(root);
1814			btrfs_delalloc_release_metadata(BTRFS_I(inode),
1815					release_bytes, true);
1816		} else {
1817			btrfs_delalloc_release_space(inode, data_reserved,
1818					round_down(pos, fs_info->sectorsize),
1819					release_bytes, true);
1820		}
1821	}
1822
1823	extent_changeset_free(data_reserved);
1824	return num_written ? num_written : ret;
1825}
1826
1827static ssize_t __btrfs_direct_write(struct kiocb *iocb, struct iov_iter *from)
 
 
1828{
1829	struct file *file = iocb->ki_filp;
1830	struct inode *inode = file_inode(file);
1831	loff_t pos;
1832	ssize_t written;
1833	ssize_t written_buffered;
1834	loff_t endbyte;
1835	int err;
1836
1837	written = generic_file_direct_write(iocb, from);
1838
1839	if (written < 0 || !iov_iter_count(from))
1840		return written;
1841
1842	pos = iocb->ki_pos;
1843	written_buffered = btrfs_buffered_write(iocb, from);
1844	if (written_buffered < 0) {
1845		err = written_buffered;
1846		goto out;
1847	}
1848	/*
1849	 * Ensure all data is persisted. We want the next direct IO read to be
1850	 * able to read what was just written.
1851	 */
1852	endbyte = pos + written_buffered - 1;
1853	err = btrfs_fdatawrite_range(inode, pos, endbyte);
1854	if (err)
1855		goto out;
1856	err = filemap_fdatawait_range(inode->i_mapping, pos, endbyte);
1857	if (err)
1858		goto out;
1859	written += written_buffered;
1860	iocb->ki_pos = pos + written_buffered;
1861	invalidate_mapping_pages(file->f_mapping, pos >> PAGE_SHIFT,
1862				 endbyte >> PAGE_SHIFT);
1863out:
1864	return written ? written : err;
1865}
1866
1867static void update_time_for_write(struct inode *inode)
1868{
1869	struct timespec64 now;
1870
1871	if (IS_NOCMTIME(inode))
1872		return;
1873
1874	now = current_time(inode);
1875	if (!timespec64_equal(&inode->i_mtime, &now))
1876		inode->i_mtime = now;
1877
1878	if (!timespec64_equal(&inode->i_ctime, &now))
1879		inode->i_ctime = now;
1880
1881	if (IS_I_VERSION(inode))
1882		inode_inc_iversion(inode);
1883}
1884
1885static ssize_t btrfs_file_write_iter(struct kiocb *iocb,
1886				    struct iov_iter *from)
1887{
1888	struct file *file = iocb->ki_filp;
1889	struct inode *inode = file_inode(file);
1890	struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1891	struct btrfs_root *root = BTRFS_I(inode)->root;
1892	u64 start_pos;
1893	u64 end_pos;
1894	ssize_t num_written = 0;
1895	const bool sync = iocb->ki_flags & IOCB_DSYNC;
1896	ssize_t err;
1897	loff_t pos;
1898	size_t count;
1899	loff_t oldsize;
1900	int clean_page = 0;
1901
1902	if (!(iocb->ki_flags & IOCB_DIRECT) &&
1903	    (iocb->ki_flags & IOCB_NOWAIT))
1904		return -EOPNOTSUPP;
1905
1906	if (!inode_trylock(inode)) {
1907		if (iocb->ki_flags & IOCB_NOWAIT)
1908			return -EAGAIN;
1909		inode_lock(inode);
1910	}
1911
1912	err = generic_write_checks(iocb, from);
1913	if (err <= 0) {
1914		inode_unlock(inode);
1915		return err;
1916	}
1917
1918	pos = iocb->ki_pos;
1919	count = iov_iter_count(from);
1920	if (iocb->ki_flags & IOCB_NOWAIT) {
1921		/*
1922		 * We will allocate space in case nodatacow is not set,
1923		 * so bail
1924		 */
1925		if (!(BTRFS_I(inode)->flags & (BTRFS_INODE_NODATACOW |
1926					      BTRFS_INODE_PREALLOC)) ||
1927		    check_can_nocow(BTRFS_I(inode), pos, &count) <= 0) {
1928			inode_unlock(inode);
1929			return -EAGAIN;
1930		}
1931	}
1932
1933	current->backing_dev_info = inode_to_bdi(inode);
1934	err = file_remove_privs(file);
1935	if (err) {
1936		inode_unlock(inode);
1937		goto out;
1938	}
1939
1940	/*
1941	 * If BTRFS flips readonly due to some impossible error
1942	 * (fs_info->fs_state now has BTRFS_SUPER_FLAG_ERROR),
1943	 * although we have opened a file as writable, we have
1944	 * to stop this write operation to ensure FS consistency.
1945	 */
1946	if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state)) {
1947		inode_unlock(inode);
1948		err = -EROFS;
1949		goto out;
1950	}
1951
1952	/*
1953	 * We reserve space for updating the inode when we reserve space for the
1954	 * extent we are going to write, so we will enospc out there.  We don't
1955	 * need to start yet another transaction to update the inode as we will
1956	 * update the inode when we finish writing whatever data we write.
1957	 */
1958	update_time_for_write(inode);
1959
1960	start_pos = round_down(pos, fs_info->sectorsize);
 
 
1961	oldsize = i_size_read(inode);
1962	if (start_pos > oldsize) {
1963		/* Expand hole size to cover write data, preventing empty gap */
1964		end_pos = round_up(pos + count,
1965				   fs_info->sectorsize);
1966		err = btrfs_cont_expand(inode, oldsize, end_pos);
1967		if (err) {
1968			inode_unlock(inode);
1969			goto out;
1970		}
1971		if (start_pos > round_up(oldsize, fs_info->sectorsize))
1972			clean_page = 1;
1973	}
1974
1975	if (sync)
1976		atomic_inc(&BTRFS_I(inode)->sync_writers);
1977
1978	if (iocb->ki_flags & IOCB_DIRECT) {
1979		num_written = __btrfs_direct_write(iocb, from);
1980	} else {
1981		num_written = btrfs_buffered_write(iocb, from);
1982		if (num_written > 0)
1983			iocb->ki_pos = pos + num_written;
1984		if (clean_page)
1985			pagecache_isize_extended(inode, oldsize,
1986						i_size_read(inode));
1987	}
1988
1989	inode_unlock(inode);
1990
1991	/*
1992	 * We also have to set last_sub_trans to the current log transid,
1993	 * otherwise subsequent syncs to a file that's been synced in this
1994	 * transaction will appear to have already occurred.
1995	 */
1996	spin_lock(&BTRFS_I(inode)->lock);
1997	BTRFS_I(inode)->last_sub_trans = root->log_transid;
1998	spin_unlock(&BTRFS_I(inode)->lock);
1999	if (num_written > 0)
2000		num_written = generic_write_sync(iocb, num_written);
 
 
 
2001
2002	if (sync)
2003		atomic_dec(&BTRFS_I(inode)->sync_writers);
2004out:
2005	current->backing_dev_info = NULL;
2006	return num_written ? num_written : err;
2007}
2008
2009int btrfs_release_file(struct inode *inode, struct file *filp)
2010{
2011	struct btrfs_file_private *private = filp->private_data;
2012
2013	if (private && private->filldir_buf)
2014		kfree(private->filldir_buf);
2015	kfree(private);
2016	filp->private_data = NULL;
2017
2018	/*
2019	 * ordered_data_close is set by setattr when we are about to truncate
2020	 * a file from a non-zero size to a zero size.  This tries to
2021	 * flush down new bytes that may have been written if the
2022	 * application were using truncate to replace a file in place.
2023	 */
2024	if (test_and_clear_bit(BTRFS_INODE_ORDERED_DATA_CLOSE,
2025			       &BTRFS_I(inode)->runtime_flags))
2026			filemap_flush(inode->i_mapping);
2027	return 0;
2028}
2029
2030static int start_ordered_ops(struct inode *inode, loff_t start, loff_t end)
2031{
2032	int ret;
2033	struct blk_plug plug;
2034
2035	/*
2036	 * This is only called in fsync, which would do synchronous writes, so
2037	 * a plug can merge adjacent IOs as much as possible.  Esp. in case of
2038	 * multiple disks using raid profile, a large IO can be split to
2039	 * several segments of stripe length (currently 64K).
2040	 */
2041	blk_start_plug(&plug);
2042	atomic_inc(&BTRFS_I(inode)->sync_writers);
2043	ret = btrfs_fdatawrite_range(inode, start, end);
2044	atomic_dec(&BTRFS_I(inode)->sync_writers);
2045	blk_finish_plug(&plug);
2046
2047	return ret;
2048}
2049
2050/*
2051 * fsync call for both files and directories.  This logs the inode into
2052 * the tree log instead of forcing full commits whenever possible.
2053 *
2054 * It needs to call filemap_fdatawait so that all ordered extent updates are
2055 * in the metadata btree are up to date for copying to the log.
2056 *
2057 * It drops the inode mutex before doing the tree log commit.  This is an
2058 * important optimization for directories because holding the mutex prevents
2059 * new operations on the dir while we write to disk.
2060 */
2061int btrfs_sync_file(struct file *file, loff_t start, loff_t end, int datasync)
2062{
2063	struct dentry *dentry = file_dentry(file);
2064	struct inode *inode = d_inode(dentry);
2065	struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2066	struct btrfs_root *root = BTRFS_I(inode)->root;
2067	struct btrfs_trans_handle *trans;
2068	struct btrfs_log_ctx ctx;
2069	int ret = 0, err;
 
 
2070
 
 
 
 
 
2071	trace_btrfs_sync_file(file, datasync);
2072
2073	btrfs_init_log_ctx(&ctx, inode);
2074
2075	/*
2076	 * We write the dirty pages in the range and wait until they complete
2077	 * out of the ->i_mutex. If so, we can flush the dirty pages by
2078	 * multi-task, and make the performance up.  See
2079	 * btrfs_wait_ordered_range for an explanation of the ASYNC check.
2080	 */
2081	ret = start_ordered_ops(inode, start, end);
2082	if (ret)
2083		goto out;
2084
2085	inode_lock(inode);
2086
2087	/*
2088	 * We take the dio_sem here because the tree log stuff can race with
2089	 * lockless dio writes and get an extent map logged for an extent we
2090	 * never waited on.  We need it this high up for lockdep reasons.
2091	 */
2092	down_write(&BTRFS_I(inode)->dio_sem);
2093
2094	atomic_inc(&root->log_batch);
2095
 
2096	/*
2097	 * If the inode needs a full sync, make sure we use a full range to
2098	 * avoid log tree corruption, due to hole detection racing with ordered
2099	 * extent completion for adjacent ranges, and assertion failures during
2100	 * hole detection. Do this while holding the inode lock, to avoid races
2101	 * with other tasks.
2102	 */
2103	if (test_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2104		     &BTRFS_I(inode)->runtime_flags)) {
2105		start = 0;
2106		end = LLONG_MAX;
2107	}
2108
2109	/*
2110	 * Before we acquired the inode's lock, someone may have dirtied more
2111	 * pages in the target range. We need to make sure that writeback for
2112	 * any such pages does not start while we are logging the inode, because
2113	 * if it does, any of the following might happen when we are not doing a
2114	 * full inode sync:
2115	 *
2116	 * 1) We log an extent after its writeback finishes but before its
2117	 *    checksums are added to the csum tree, leading to -EIO errors
2118	 *    when attempting to read the extent after a log replay.
2119	 *
2120	 * 2) We can end up logging an extent before its writeback finishes.
2121	 *    Therefore after the log replay we will have a file extent item
2122	 *    pointing to an unwritten extent (and no data checksums as well).
2123	 *
2124	 * So trigger writeback for any eventual new dirty pages and then we
2125	 * wait for all ordered extents to complete below.
2126	 */
2127	ret = start_ordered_ops(inode, start, end);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2128	if (ret) {
2129		inode_unlock(inode);
2130		goto out;
2131	}
 
2132
2133	/*
2134	 * We have to do this here to avoid the priority inversion of waiting on
2135	 * IO of a lower priority task while holding a transaction open.
 
2136	 *
2137	 * Also, the range length can be represented by u64, we have to do the
2138	 * typecasts to avoid signed overflow if it's [0, LLONG_MAX].
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2139	 */
2140	ret = btrfs_wait_ordered_range(inode, start, (u64)end - (u64)start + 1);
2141	if (ret) {
2142		up_write(&BTRFS_I(inode)->dio_sem);
2143		inode_unlock(inode);
2144		goto out;
2145	}
2146	atomic_inc(&root->log_batch);
2147
2148	smp_mb();
2149	if (btrfs_inode_in_log(BTRFS_I(inode), fs_info->generation) ||
2150	    BTRFS_I(inode)->last_trans <= fs_info->last_trans_committed) {
 
 
 
 
2151		/*
2152		 * We've had everything committed since the last time we were
2153		 * modified so clear this flag in case it was set for whatever
2154		 * reason, it's no longer relevant.
2155		 */
2156		clear_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2157			  &BTRFS_I(inode)->runtime_flags);
2158		/*
2159		 * An ordered extent might have started before and completed
2160		 * already with io errors, in which case the inode was not
2161		 * updated and we end up here. So check the inode's mapping
2162		 * for any errors that might have happened since we last
2163		 * checked called fsync.
2164		 */
2165		ret = filemap_check_wb_err(inode->i_mapping, file->f_wb_err);
2166		up_write(&BTRFS_I(inode)->dio_sem);
2167		inode_unlock(inode);
2168		goto out;
2169	}
2170
2171	/*
 
 
 
 
 
 
2172	 * We use start here because we will need to wait on the IO to complete
2173	 * in btrfs_sync_log, which could require joining a transaction (for
2174	 * example checking cross references in the nocow path).  If we use join
2175	 * here we could get into a situation where we're waiting on IO to
2176	 * happen that is blocked on a transaction trying to commit.  With start
2177	 * we inc the extwriter counter, so we wait for all extwriters to exit
2178	 * before we start blocking joiners.  This comment is to keep somebody
2179	 * from thinking they are super smart and changing this to
2180	 * btrfs_join_transaction *cough*Josef*cough*.
2181	 */
2182	trans = btrfs_start_transaction(root, 0);
2183	if (IS_ERR(trans)) {
2184		ret = PTR_ERR(trans);
2185		up_write(&BTRFS_I(inode)->dio_sem);
2186		inode_unlock(inode);
2187		goto out;
2188	}
 
 
 
2189
2190	ret = btrfs_log_dentry_safe(trans, dentry, start, end, &ctx);
2191	if (ret < 0) {
2192		/* Fallthrough and commit/free transaction. */
2193		ret = 1;
2194	}
2195
2196	/* we've logged all the items and now have a consistent
2197	 * version of the file in the log.  It is possible that
2198	 * someone will come in and modify the file, but that's
2199	 * fine because the log is consistent on disk, and we
2200	 * have references to all of the file's extents
2201	 *
2202	 * It is possible that someone will come in and log the
2203	 * file again, but that will end up using the synchronization
2204	 * inside btrfs_sync_log to keep things safe.
2205	 */
2206	up_write(&BTRFS_I(inode)->dio_sem);
2207	inode_unlock(inode);
2208
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2209	if (ret != BTRFS_NO_LOG_SYNC) {
2210		if (!ret) {
2211			ret = btrfs_sync_log(trans, root, &ctx);
2212			if (!ret) {
2213				ret = btrfs_end_transaction(trans);
 
 
 
 
 
 
 
2214				goto out;
2215			}
2216		}
2217		ret = btrfs_commit_transaction(trans);
2218	} else {
2219		ret = btrfs_end_transaction(trans);
2220	}
2221out:
2222	ASSERT(list_empty(&ctx.list));
2223	err = file_check_and_advance_wb_err(file);
2224	if (!ret)
2225		ret = err;
2226	return ret > 0 ? -EIO : ret;
2227}
2228
2229static const struct vm_operations_struct btrfs_file_vm_ops = {
2230	.fault		= filemap_fault,
2231	.map_pages	= filemap_map_pages,
2232	.page_mkwrite	= btrfs_page_mkwrite,
2233};
2234
2235static int btrfs_file_mmap(struct file	*filp, struct vm_area_struct *vma)
2236{
2237	struct address_space *mapping = filp->f_mapping;
2238
2239	if (!mapping->a_ops->readpage)
2240		return -ENOEXEC;
2241
2242	file_accessed(filp);
2243	vma->vm_ops = &btrfs_file_vm_ops;
2244
2245	return 0;
2246}
2247
2248static int hole_mergeable(struct btrfs_inode *inode, struct extent_buffer *leaf,
2249			  int slot, u64 start, u64 end)
2250{
2251	struct btrfs_file_extent_item *fi;
2252	struct btrfs_key key;
2253
2254	if (slot < 0 || slot >= btrfs_header_nritems(leaf))
2255		return 0;
2256
2257	btrfs_item_key_to_cpu(leaf, &key, slot);
2258	if (key.objectid != btrfs_ino(inode) ||
2259	    key.type != BTRFS_EXTENT_DATA_KEY)
2260		return 0;
2261
2262	fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
2263
2264	if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG)
2265		return 0;
2266
2267	if (btrfs_file_extent_disk_bytenr(leaf, fi))
2268		return 0;
2269
2270	if (key.offset == end)
2271		return 1;
2272	if (key.offset + btrfs_file_extent_num_bytes(leaf, fi) == start)
2273		return 1;
2274	return 0;
2275}
2276
2277static int fill_holes(struct btrfs_trans_handle *trans,
2278		struct btrfs_inode *inode,
2279		struct btrfs_path *path, u64 offset, u64 end)
2280{
2281	struct btrfs_fs_info *fs_info = trans->fs_info;
2282	struct btrfs_root *root = inode->root;
2283	struct extent_buffer *leaf;
2284	struct btrfs_file_extent_item *fi;
2285	struct extent_map *hole_em;
2286	struct extent_map_tree *em_tree = &inode->extent_tree;
2287	struct btrfs_key key;
2288	int ret;
2289
2290	if (btrfs_fs_incompat(fs_info, NO_HOLES))
2291		goto out;
2292
2293	key.objectid = btrfs_ino(inode);
2294	key.type = BTRFS_EXTENT_DATA_KEY;
2295	key.offset = offset;
2296
2297	ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
2298	if (ret <= 0) {
2299		/*
2300		 * We should have dropped this offset, so if we find it then
2301		 * something has gone horribly wrong.
2302		 */
2303		if (ret == 0)
2304			ret = -EINVAL;
2305		return ret;
2306	}
2307
2308	leaf = path->nodes[0];
2309	if (hole_mergeable(inode, leaf, path->slots[0] - 1, offset, end)) {
2310		u64 num_bytes;
2311
2312		path->slots[0]--;
2313		fi = btrfs_item_ptr(leaf, path->slots[0],
2314				    struct btrfs_file_extent_item);
2315		num_bytes = btrfs_file_extent_num_bytes(leaf, fi) +
2316			end - offset;
2317		btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2318		btrfs_set_file_extent_ram_bytes(leaf, fi, num_bytes);
2319		btrfs_set_file_extent_offset(leaf, fi, 0);
2320		btrfs_mark_buffer_dirty(leaf);
2321		goto out;
2322	}
2323
2324	if (hole_mergeable(inode, leaf, path->slots[0], offset, end)) {
2325		u64 num_bytes;
2326
2327		key.offset = offset;
2328		btrfs_set_item_key_safe(fs_info, path, &key);
2329		fi = btrfs_item_ptr(leaf, path->slots[0],
2330				    struct btrfs_file_extent_item);
2331		num_bytes = btrfs_file_extent_num_bytes(leaf, fi) + end -
2332			offset;
2333		btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2334		btrfs_set_file_extent_ram_bytes(leaf, fi, num_bytes);
2335		btrfs_set_file_extent_offset(leaf, fi, 0);
2336		btrfs_mark_buffer_dirty(leaf);
2337		goto out;
2338	}
2339	btrfs_release_path(path);
2340
2341	ret = btrfs_insert_file_extent(trans, root, btrfs_ino(inode),
2342			offset, 0, 0, end - offset, 0, end - offset, 0, 0, 0);
 
2343	if (ret)
2344		return ret;
2345
2346out:
2347	btrfs_release_path(path);
2348
2349	hole_em = alloc_extent_map();
2350	if (!hole_em) {
2351		btrfs_drop_extent_cache(inode, offset, end - 1, 0);
2352		set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &inode->runtime_flags);
 
2353	} else {
2354		hole_em->start = offset;
2355		hole_em->len = end - offset;
2356		hole_em->ram_bytes = hole_em->len;
2357		hole_em->orig_start = offset;
2358
2359		hole_em->block_start = EXTENT_MAP_HOLE;
2360		hole_em->block_len = 0;
2361		hole_em->orig_block_len = 0;
2362		hole_em->bdev = fs_info->fs_devices->latest_bdev;
2363		hole_em->compress_type = BTRFS_COMPRESS_NONE;
2364		hole_em->generation = trans->transid;
2365
2366		do {
2367			btrfs_drop_extent_cache(inode, offset, end - 1, 0);
2368			write_lock(&em_tree->lock);
2369			ret = add_extent_mapping(em_tree, hole_em, 1);
2370			write_unlock(&em_tree->lock);
2371		} while (ret == -EEXIST);
2372		free_extent_map(hole_em);
2373		if (ret)
2374			set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2375					&inode->runtime_flags);
2376	}
2377
2378	return 0;
2379}
2380
2381/*
2382 * Find a hole extent on given inode and change start/len to the end of hole
2383 * extent.(hole/vacuum extent whose em->start <= start &&
2384 *	   em->start + em->len > start)
2385 * When a hole extent is found, return 1 and modify start/len.
2386 */
2387static int find_first_non_hole(struct inode *inode, u64 *start, u64 *len)
2388{
2389	struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2390	struct extent_map *em;
2391	int ret = 0;
2392
2393	em = btrfs_get_extent(BTRFS_I(inode), NULL, 0,
2394			      round_down(*start, fs_info->sectorsize),
2395			      round_up(*len, fs_info->sectorsize), 0);
2396	if (IS_ERR(em))
2397		return PTR_ERR(em);
 
 
 
2398
2399	/* Hole or vacuum extent(only exists in no-hole mode) */
2400	if (em->block_start == EXTENT_MAP_HOLE) {
2401		ret = 1;
2402		*len = em->start + em->len > *start + *len ?
2403		       0 : *start + *len - em->start - em->len;
2404		*start = em->start + em->len;
2405	}
2406	free_extent_map(em);
2407	return ret;
2408}
2409
2410static int btrfs_punch_hole_lock_range(struct inode *inode,
2411				       const u64 lockstart,
2412				       const u64 lockend,
2413				       struct extent_state **cached_state)
2414{
2415	while (1) {
2416		struct btrfs_ordered_extent *ordered;
2417		int ret;
2418
2419		truncate_pagecache_range(inode, lockstart, lockend);
2420
2421		lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2422				 cached_state);
2423		ordered = btrfs_lookup_first_ordered_extent(inode, lockend);
2424
2425		/*
2426		 * We need to make sure we have no ordered extents in this range
2427		 * and nobody raced in and read a page in this range, if we did
2428		 * we need to try again.
2429		 */
2430		if ((!ordered ||
2431		    (ordered->file_offset + ordered->len <= lockstart ||
2432		     ordered->file_offset > lockend)) &&
2433		     !filemap_range_has_page(inode->i_mapping,
2434					     lockstart, lockend)) {
2435			if (ordered)
2436				btrfs_put_ordered_extent(ordered);
2437			break;
2438		}
2439		if (ordered)
2440			btrfs_put_ordered_extent(ordered);
2441		unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart,
2442				     lockend, cached_state);
2443		ret = btrfs_wait_ordered_range(inode, lockstart,
2444					       lockend - lockstart + 1);
2445		if (ret)
2446			return ret;
2447	}
2448	return 0;
2449}
2450
2451static int btrfs_insert_clone_extent(struct btrfs_trans_handle *trans,
2452				     struct inode *inode,
2453				     struct btrfs_path *path,
2454				     struct btrfs_clone_extent_info *clone_info,
2455				     const u64 clone_len)
2456{
2457	struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2458	struct btrfs_root *root = BTRFS_I(inode)->root;
2459	struct btrfs_file_extent_item *extent;
2460	struct extent_buffer *leaf;
2461	struct btrfs_key key;
2462	int slot;
2463	struct btrfs_ref ref = { 0 };
2464	u64 ref_offset;
2465	int ret;
2466
2467	if (clone_len == 0)
2468		return 0;
2469
2470	if (clone_info->disk_offset == 0 &&
2471	    btrfs_fs_incompat(fs_info, NO_HOLES))
2472		return 0;
2473
2474	key.objectid = btrfs_ino(BTRFS_I(inode));
2475	key.type = BTRFS_EXTENT_DATA_KEY;
2476	key.offset = clone_info->file_offset;
2477	ret = btrfs_insert_empty_item(trans, root, path, &key,
2478				      clone_info->item_size);
2479	if (ret)
2480		return ret;
2481	leaf = path->nodes[0];
2482	slot = path->slots[0];
2483	write_extent_buffer(leaf, clone_info->extent_buf,
2484			    btrfs_item_ptr_offset(leaf, slot),
2485			    clone_info->item_size);
2486	extent = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
2487	btrfs_set_file_extent_offset(leaf, extent, clone_info->data_offset);
2488	btrfs_set_file_extent_num_bytes(leaf, extent, clone_len);
2489	btrfs_mark_buffer_dirty(leaf);
2490	btrfs_release_path(path);
2491
2492	/* If it's a hole, nothing more needs to be done. */
2493	if (clone_info->disk_offset == 0)
2494		return 0;
2495
2496	inode_add_bytes(inode, clone_len);
2497	btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF,
2498			       clone_info->disk_offset,
2499			       clone_info->disk_len, 0);
2500	ref_offset = clone_info->file_offset - clone_info->data_offset;
2501	btrfs_init_data_ref(&ref, root->root_key.objectid,
2502			    btrfs_ino(BTRFS_I(inode)), ref_offset);
2503	ret = btrfs_inc_extent_ref(trans, &ref);
2504
2505	return ret;
2506}
2507
2508/*
2509 * The respective range must have been previously locked, as well as the inode.
2510 * The end offset is inclusive (last byte of the range).
2511 * @clone_info is NULL for fallocate's hole punching and non-NULL for extent
2512 * cloning.
2513 * When cloning, we don't want to end up in a state where we dropped extents
2514 * without inserting a new one, so we must abort the transaction to avoid a
2515 * corruption.
2516 */
2517int btrfs_punch_hole_range(struct inode *inode, struct btrfs_path *path,
2518			   const u64 start, const u64 end,
2519			   struct btrfs_clone_extent_info *clone_info,
2520			   struct btrfs_trans_handle **trans_out)
2521{
2522	struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2523	u64 min_size = btrfs_calc_insert_metadata_size(fs_info, 1);
2524	u64 ino_size = round_up(inode->i_size, fs_info->sectorsize);
2525	struct btrfs_root *root = BTRFS_I(inode)->root;
2526	struct btrfs_trans_handle *trans = NULL;
2527	struct btrfs_block_rsv *rsv;
2528	unsigned int rsv_count;
2529	u64 cur_offset;
2530	u64 drop_end;
2531	u64 len = end - start;
2532	int ret = 0;
2533
2534	if (end <= start)
2535		return -EINVAL;
2536
2537	rsv = btrfs_alloc_block_rsv(fs_info, BTRFS_BLOCK_RSV_TEMP);
2538	if (!rsv) {
2539		ret = -ENOMEM;
2540		goto out;
2541	}
2542	rsv->size = btrfs_calc_insert_metadata_size(fs_info, 1);
2543	rsv->failfast = 1;
2544
2545	/*
2546	 * 1 - update the inode
2547	 * 1 - removing the extents in the range
2548	 * 1 - adding the hole extent if no_holes isn't set or if we are cloning
2549	 *     an extent
2550	 */
2551	if (!btrfs_fs_incompat(fs_info, NO_HOLES) || clone_info)
2552		rsv_count = 3;
2553	else
2554		rsv_count = 2;
2555
2556	trans = btrfs_start_transaction(root, rsv_count);
2557	if (IS_ERR(trans)) {
2558		ret = PTR_ERR(trans);
2559		trans = NULL;
2560		goto out_free;
2561	}
2562
2563	ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv, rsv,
2564				      min_size, false);
2565	BUG_ON(ret);
2566	trans->block_rsv = rsv;
2567
2568	cur_offset = start;
2569	while (cur_offset < end) {
2570		ret = __btrfs_drop_extents(trans, root, inode, path,
2571					   cur_offset, end + 1, &drop_end,
2572					   1, 0, 0, NULL);
2573		if (ret != -ENOSPC) {
2574			/*
2575			 * When cloning we want to avoid transaction aborts when
2576			 * nothing was done and we are attempting to clone parts
2577			 * of inline extents, in such cases -EOPNOTSUPP is
2578			 * returned by __btrfs_drop_extents() without having
2579			 * changed anything in the file.
2580			 */
2581			if (clone_info && ret && ret != -EOPNOTSUPP)
2582				btrfs_abort_transaction(trans, ret);
2583			break;
2584		}
2585
2586		trans->block_rsv = &fs_info->trans_block_rsv;
2587
2588		if (!clone_info && cur_offset < drop_end &&
2589		    cur_offset < ino_size) {
2590			ret = fill_holes(trans, BTRFS_I(inode), path,
2591					cur_offset, drop_end);
2592			if (ret) {
2593				/*
2594				 * If we failed then we didn't insert our hole
2595				 * entries for the area we dropped, so now the
2596				 * fs is corrupted, so we must abort the
2597				 * transaction.
2598				 */
2599				btrfs_abort_transaction(trans, ret);
2600				break;
2601			}
2602		}
2603
2604		if (clone_info) {
2605			u64 clone_len = drop_end - cur_offset;
2606
2607			ret = btrfs_insert_clone_extent(trans, inode, path,
2608							clone_info, clone_len);
2609			if (ret) {
2610				btrfs_abort_transaction(trans, ret);
2611				break;
2612			}
2613			clone_info->data_len -= clone_len;
2614			clone_info->data_offset += clone_len;
2615			clone_info->file_offset += clone_len;
2616		}
2617
2618		cur_offset = drop_end;
2619
2620		ret = btrfs_update_inode(trans, root, inode);
2621		if (ret)
2622			break;
2623
2624		btrfs_end_transaction(trans);
2625		btrfs_btree_balance_dirty(fs_info);
2626
2627		trans = btrfs_start_transaction(root, rsv_count);
2628		if (IS_ERR(trans)) {
2629			ret = PTR_ERR(trans);
2630			trans = NULL;
2631			break;
2632		}
2633
2634		ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv,
2635					      rsv, min_size, false);
2636		BUG_ON(ret);	/* shouldn't happen */
2637		trans->block_rsv = rsv;
2638
2639		if (!clone_info) {
2640			ret = find_first_non_hole(inode, &cur_offset, &len);
2641			if (unlikely(ret < 0))
2642				break;
2643			if (ret && !len) {
2644				ret = 0;
2645				break;
2646			}
2647		}
2648	}
2649
2650	/*
2651	 * If we were cloning, force the next fsync to be a full one since we
2652	 * we replaced (or just dropped in the case of cloning holes when
2653	 * NO_HOLES is enabled) extents and extent maps.
2654	 * This is for the sake of simplicity, and cloning into files larger
2655	 * than 16Mb would force the full fsync any way (when
2656	 * try_release_extent_mapping() is invoked during page cache truncation.
2657	 */
2658	if (clone_info)
2659		set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2660			&BTRFS_I(inode)->runtime_flags);
2661
2662	if (ret)
2663		goto out_trans;
2664
2665	trans->block_rsv = &fs_info->trans_block_rsv;
2666	/*
2667	 * If we are using the NO_HOLES feature we might have had already an
2668	 * hole that overlaps a part of the region [lockstart, lockend] and
2669	 * ends at (or beyond) lockend. Since we have no file extent items to
2670	 * represent holes, drop_end can be less than lockend and so we must
2671	 * make sure we have an extent map representing the existing hole (the
2672	 * call to __btrfs_drop_extents() might have dropped the existing extent
2673	 * map representing the existing hole), otherwise the fast fsync path
2674	 * will not record the existence of the hole region
2675	 * [existing_hole_start, lockend].
2676	 */
2677	if (drop_end <= end)
2678		drop_end = end + 1;
2679	/*
2680	 * Don't insert file hole extent item if it's for a range beyond eof
2681	 * (because it's useless) or if it represents a 0 bytes range (when
2682	 * cur_offset == drop_end).
2683	 */
2684	if (!clone_info && cur_offset < ino_size && cur_offset < drop_end) {
2685		ret = fill_holes(trans, BTRFS_I(inode), path,
2686				cur_offset, drop_end);
2687		if (ret) {
2688			/* Same comment as above. */
2689			btrfs_abort_transaction(trans, ret);
2690			goto out_trans;
2691		}
2692	}
2693	if (clone_info) {
2694		ret = btrfs_insert_clone_extent(trans, inode, path, clone_info,
2695						clone_info->data_len);
2696		if (ret) {
2697			btrfs_abort_transaction(trans, ret);
2698			goto out_trans;
2699		}
2700	}
2701
2702out_trans:
2703	if (!trans)
2704		goto out_free;
2705
2706	trans->block_rsv = &fs_info->trans_block_rsv;
2707	if (ret)
2708		btrfs_end_transaction(trans);
2709	else
2710		*trans_out = trans;
2711out_free:
2712	btrfs_free_block_rsv(fs_info, rsv);
2713out:
2714	return ret;
2715}
2716
2717static int btrfs_punch_hole(struct inode *inode, loff_t offset, loff_t len)
2718{
2719	struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2720	struct btrfs_root *root = BTRFS_I(inode)->root;
2721	struct extent_state *cached_state = NULL;
2722	struct btrfs_path *path;
2723	struct btrfs_trans_handle *trans = NULL;
 
2724	u64 lockstart;
2725	u64 lockend;
2726	u64 tail_start;
2727	u64 tail_len;
2728	u64 orig_start = offset;
 
 
 
2729	int ret = 0;
 
 
2730	bool same_block;
 
2731	u64 ino_size;
2732	bool truncated_block = false;
2733	bool updated_inode = false;
2734
2735	ret = btrfs_wait_ordered_range(inode, offset, len);
2736	if (ret)
2737		return ret;
2738
2739	inode_lock(inode);
2740	ino_size = round_up(inode->i_size, fs_info->sectorsize);
2741	ret = find_first_non_hole(inode, &offset, &len);
2742	if (ret < 0)
2743		goto out_only_mutex;
2744	if (ret && !len) {
2745		/* Already in a large hole */
2746		ret = 0;
2747		goto out_only_mutex;
2748	}
2749
2750	lockstart = round_up(offset, btrfs_inode_sectorsize(inode));
2751	lockend = round_down(offset + len,
2752			     btrfs_inode_sectorsize(inode)) - 1;
2753	same_block = (BTRFS_BYTES_TO_BLKS(fs_info, offset))
2754		== (BTRFS_BYTES_TO_BLKS(fs_info, offset + len - 1));
2755	/*
2756	 * We needn't truncate any block which is beyond the end of the file
2757	 * because we are sure there is no data there.
2758	 */
2759	/*
2760	 * Only do this if we are in the same block and we aren't doing the
2761	 * entire block.
2762	 */
2763	if (same_block && len < fs_info->sectorsize) {
2764		if (offset < ino_size) {
2765			truncated_block = true;
2766			ret = btrfs_truncate_block(inode, offset, len, 0);
2767		} else {
2768			ret = 0;
2769		}
2770		goto out_only_mutex;
2771	}
2772
2773	/* zero back part of the first block */
2774	if (offset < ino_size) {
2775		truncated_block = true;
2776		ret = btrfs_truncate_block(inode, offset, 0, 0);
2777		if (ret) {
2778			inode_unlock(inode);
2779			return ret;
2780		}
2781	}
2782
2783	/* Check the aligned pages after the first unaligned page,
2784	 * if offset != orig_start, which means the first unaligned page
2785	 * including several following pages are already in holes,
2786	 * the extra check can be skipped */
2787	if (offset == orig_start) {
2788		/* after truncate page, check hole again */
2789		len = offset + len - lockstart;
2790		offset = lockstart;
2791		ret = find_first_non_hole(inode, &offset, &len);
2792		if (ret < 0)
2793			goto out_only_mutex;
2794		if (ret && !len) {
2795			ret = 0;
2796			goto out_only_mutex;
2797		}
2798		lockstart = offset;
2799	}
2800
2801	/* Check the tail unaligned part is in a hole */
2802	tail_start = lockend + 1;
2803	tail_len = offset + len - tail_start;
2804	if (tail_len) {
2805		ret = find_first_non_hole(inode, &tail_start, &tail_len);
2806		if (unlikely(ret < 0))
2807			goto out_only_mutex;
2808		if (!ret) {
2809			/* zero the front end of the last page */
2810			if (tail_start + tail_len < ino_size) {
2811				truncated_block = true;
2812				ret = btrfs_truncate_block(inode,
2813							tail_start + tail_len,
2814							0, 1);
2815				if (ret)
2816					goto out_only_mutex;
2817			}
2818		}
2819	}
2820
2821	if (lockend < lockstart) {
2822		ret = 0;
2823		goto out_only_mutex;
2824	}
2825
2826	ret = btrfs_punch_hole_lock_range(inode, lockstart, lockend,
2827					  &cached_state);
2828	if (ret)
2829		goto out_only_mutex;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2830
2831	path = btrfs_alloc_path();
2832	if (!path) {
2833		ret = -ENOMEM;
2834		goto out;
2835	}
2836
2837	ret = btrfs_punch_hole_range(inode, path, lockstart, lockend, NULL,
2838				     &trans);
2839	btrfs_free_path(path);
2840	if (ret)
2841		goto out;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2842
2843	ASSERT(trans != NULL);
2844	inode_inc_iversion(inode);
2845	inode->i_mtime = inode->i_ctime = current_time(inode);
 
 
2846	ret = btrfs_update_inode(trans, root, inode);
2847	updated_inode = true;
2848	btrfs_end_transaction(trans);
2849	btrfs_btree_balance_dirty(fs_info);
 
 
 
2850out:
2851	unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2852			     &cached_state);
2853out_only_mutex:
2854	if (!updated_inode && truncated_block && !ret) {
2855		/*
2856		 * If we only end up zeroing part of a page, we still need to
2857		 * update the inode item, so that all the time fields are
2858		 * updated as well as the necessary btrfs inode in memory fields
2859		 * for detecting, at fsync time, if the inode isn't yet in the
2860		 * log tree or it's there but not up to date.
2861		 */
2862		struct timespec64 now = current_time(inode);
2863
2864		inode_inc_iversion(inode);
2865		inode->i_mtime = now;
2866		inode->i_ctime = now;
2867		trans = btrfs_start_transaction(root, 1);
2868		if (IS_ERR(trans)) {
2869			ret = PTR_ERR(trans);
2870		} else {
2871			int ret2;
2872
2873			ret = btrfs_update_inode(trans, root, inode);
2874			ret2 = btrfs_end_transaction(trans);
2875			if (!ret)
2876				ret = ret2;
2877		}
2878	}
2879	inode_unlock(inode);
2880	return ret;
 
 
2881}
2882
2883/* Helper structure to record which range is already reserved */
2884struct falloc_range {
2885	struct list_head list;
2886	u64 start;
2887	u64 len;
2888};
2889
2890/*
2891 * Helper function to add falloc range
2892 *
2893 * Caller should have locked the larger range of extent containing
2894 * [start, len)
2895 */
2896static int add_falloc_range(struct list_head *head, u64 start, u64 len)
2897{
2898	struct falloc_range *prev = NULL;
2899	struct falloc_range *range = NULL;
2900
2901	if (list_empty(head))
2902		goto insert;
2903
2904	/*
2905	 * As fallocate iterate by bytenr order, we only need to check
2906	 * the last range.
2907	 */
2908	prev = list_entry(head->prev, struct falloc_range, list);
2909	if (prev->start + prev->len == start) {
2910		prev->len += len;
2911		return 0;
2912	}
2913insert:
2914	range = kmalloc(sizeof(*range), GFP_KERNEL);
2915	if (!range)
2916		return -ENOMEM;
2917	range->start = start;
2918	range->len = len;
2919	list_add_tail(&range->list, head);
2920	return 0;
2921}
2922
2923static int btrfs_fallocate_update_isize(struct inode *inode,
2924					const u64 end,
2925					const int mode)
2926{
2927	struct btrfs_trans_handle *trans;
2928	struct btrfs_root *root = BTRFS_I(inode)->root;
2929	int ret;
2930	int ret2;
2931
2932	if (mode & FALLOC_FL_KEEP_SIZE || end <= i_size_read(inode))
2933		return 0;
2934
2935	trans = btrfs_start_transaction(root, 1);
2936	if (IS_ERR(trans))
2937		return PTR_ERR(trans);
2938
2939	inode->i_ctime = current_time(inode);
2940	i_size_write(inode, end);
2941	btrfs_ordered_update_i_size(inode, end, NULL);
2942	ret = btrfs_update_inode(trans, root, inode);
2943	ret2 = btrfs_end_transaction(trans);
2944
2945	return ret ? ret : ret2;
2946}
2947
2948enum {
2949	RANGE_BOUNDARY_WRITTEN_EXTENT,
2950	RANGE_BOUNDARY_PREALLOC_EXTENT,
2951	RANGE_BOUNDARY_HOLE,
2952};
2953
2954static int btrfs_zero_range_check_range_boundary(struct inode *inode,
2955						 u64 offset)
2956{
2957	const u64 sectorsize = btrfs_inode_sectorsize(inode);
2958	struct extent_map *em;
2959	int ret;
2960
2961	offset = round_down(offset, sectorsize);
2962	em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, offset, sectorsize, 0);
2963	if (IS_ERR(em))
2964		return PTR_ERR(em);
2965
2966	if (em->block_start == EXTENT_MAP_HOLE)
2967		ret = RANGE_BOUNDARY_HOLE;
2968	else if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
2969		ret = RANGE_BOUNDARY_PREALLOC_EXTENT;
2970	else
2971		ret = RANGE_BOUNDARY_WRITTEN_EXTENT;
2972
2973	free_extent_map(em);
2974	return ret;
2975}
2976
2977static int btrfs_zero_range(struct inode *inode,
2978			    loff_t offset,
2979			    loff_t len,
2980			    const int mode)
2981{
2982	struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
2983	struct extent_map *em;
2984	struct extent_changeset *data_reserved = NULL;
2985	int ret;
2986	u64 alloc_hint = 0;
2987	const u64 sectorsize = btrfs_inode_sectorsize(inode);
2988	u64 alloc_start = round_down(offset, sectorsize);
2989	u64 alloc_end = round_up(offset + len, sectorsize);
2990	u64 bytes_to_reserve = 0;
2991	bool space_reserved = false;
2992
2993	inode_dio_wait(inode);
2994
2995	em = btrfs_get_extent(BTRFS_I(inode), NULL, 0,
2996			      alloc_start, alloc_end - alloc_start, 0);
2997	if (IS_ERR(em)) {
2998		ret = PTR_ERR(em);
2999		goto out;
3000	}
3001
3002	/*
3003	 * Avoid hole punching and extent allocation for some cases. More cases
3004	 * could be considered, but these are unlikely common and we keep things
3005	 * as simple as possible for now. Also, intentionally, if the target
3006	 * range contains one or more prealloc extents together with regular
3007	 * extents and holes, we drop all the existing extents and allocate a
3008	 * new prealloc extent, so that we get a larger contiguous disk extent.
3009	 */
3010	if (em->start <= alloc_start &&
3011	    test_bit(EXTENT_FLAG_PREALLOC, &em->flags)) {
3012		const u64 em_end = em->start + em->len;
3013
3014		if (em_end >= offset + len) {
3015			/*
3016			 * The whole range is already a prealloc extent,
3017			 * do nothing except updating the inode's i_size if
3018			 * needed.
3019			 */
3020			free_extent_map(em);
3021			ret = btrfs_fallocate_update_isize(inode, offset + len,
3022							   mode);
3023			goto out;
3024		}
3025		/*
3026		 * Part of the range is already a prealloc extent, so operate
3027		 * only on the remaining part of the range.
3028		 */
3029		alloc_start = em_end;
3030		ASSERT(IS_ALIGNED(alloc_start, sectorsize));
3031		len = offset + len - alloc_start;
3032		offset = alloc_start;
3033		alloc_hint = em->block_start + em->len;
3034	}
3035	free_extent_map(em);
3036
3037	if (BTRFS_BYTES_TO_BLKS(fs_info, offset) ==
3038	    BTRFS_BYTES_TO_BLKS(fs_info, offset + len - 1)) {
3039		em = btrfs_get_extent(BTRFS_I(inode), NULL, 0,
3040				      alloc_start, sectorsize, 0);
3041		if (IS_ERR(em)) {
3042			ret = PTR_ERR(em);
3043			goto out;
3044		}
3045
3046		if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags)) {
3047			free_extent_map(em);
3048			ret = btrfs_fallocate_update_isize(inode, offset + len,
3049							   mode);
3050			goto out;
3051		}
3052		if (len < sectorsize && em->block_start != EXTENT_MAP_HOLE) {
3053			free_extent_map(em);
3054			ret = btrfs_truncate_block(inode, offset, len, 0);
3055			if (!ret)
3056				ret = btrfs_fallocate_update_isize(inode,
3057								   offset + len,
3058								   mode);
3059			return ret;
3060		}
3061		free_extent_map(em);
3062		alloc_start = round_down(offset, sectorsize);
3063		alloc_end = alloc_start + sectorsize;
3064		goto reserve_space;
3065	}
3066
3067	alloc_start = round_up(offset, sectorsize);
3068	alloc_end = round_down(offset + len, sectorsize);
3069
3070	/*
3071	 * For unaligned ranges, check the pages at the boundaries, they might
3072	 * map to an extent, in which case we need to partially zero them, or
3073	 * they might map to a hole, in which case we need our allocation range
3074	 * to cover them.
3075	 */
3076	if (!IS_ALIGNED(offset, sectorsize)) {
3077		ret = btrfs_zero_range_check_range_boundary(inode, offset);
3078		if (ret < 0)
3079			goto out;
3080		if (ret == RANGE_BOUNDARY_HOLE) {
3081			alloc_start = round_down(offset, sectorsize);
3082			ret = 0;
3083		} else if (ret == RANGE_BOUNDARY_WRITTEN_EXTENT) {
3084			ret = btrfs_truncate_block(inode, offset, 0, 0);
3085			if (ret)
3086				goto out;
3087		} else {
3088			ret = 0;
3089		}
3090	}
3091
3092	if (!IS_ALIGNED(offset + len, sectorsize)) {
3093		ret = btrfs_zero_range_check_range_boundary(inode,
3094							    offset + len);
3095		if (ret < 0)
3096			goto out;
3097		if (ret == RANGE_BOUNDARY_HOLE) {
3098			alloc_end = round_up(offset + len, sectorsize);
3099			ret = 0;
3100		} else if (ret == RANGE_BOUNDARY_WRITTEN_EXTENT) {
3101			ret = btrfs_truncate_block(inode, offset + len, 0, 1);
3102			if (ret)
3103				goto out;
3104		} else {
3105			ret = 0;
3106		}
3107	}
3108
3109reserve_space:
3110	if (alloc_start < alloc_end) {
3111		struct extent_state *cached_state = NULL;
3112		const u64 lockstart = alloc_start;
3113		const u64 lockend = alloc_end - 1;
3114
3115		bytes_to_reserve = alloc_end - alloc_start;
3116		ret = btrfs_alloc_data_chunk_ondemand(BTRFS_I(inode),
3117						      bytes_to_reserve);
3118		if (ret < 0)
3119			goto out;
3120		space_reserved = true;
3121		ret = btrfs_qgroup_reserve_data(inode, &data_reserved,
3122						alloc_start, bytes_to_reserve);
3123		if (ret)
3124			goto out;
3125		ret = btrfs_punch_hole_lock_range(inode, lockstart, lockend,
3126						  &cached_state);
3127		if (ret)
3128			goto out;
3129		ret = btrfs_prealloc_file_range(inode, mode, alloc_start,
3130						alloc_end - alloc_start,
3131						i_blocksize(inode),
3132						offset + len, &alloc_hint);
3133		unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart,
3134				     lockend, &cached_state);
3135		/* btrfs_prealloc_file_range releases reserved space on error */
3136		if (ret) {
3137			space_reserved = false;
3138			goto out;
3139		}
3140	}
3141	ret = btrfs_fallocate_update_isize(inode, offset + len, mode);
3142 out:
3143	if (ret && space_reserved)
3144		btrfs_free_reserved_data_space(inode, data_reserved,
3145					       alloc_start, bytes_to_reserve);
3146	extent_changeset_free(data_reserved);
3147
3148	return ret;
3149}
3150
3151static long btrfs_fallocate(struct file *file, int mode,
3152			    loff_t offset, loff_t len)
3153{
3154	struct inode *inode = file_inode(file);
3155	struct extent_state *cached_state = NULL;
3156	struct extent_changeset *data_reserved = NULL;
3157	struct falloc_range *range;
3158	struct falloc_range *tmp;
3159	struct list_head reserve_list;
3160	u64 cur_offset;
3161	u64 last_byte;
3162	u64 alloc_start;
3163	u64 alloc_end;
3164	u64 alloc_hint = 0;
3165	u64 locked_end;
3166	u64 actual_end = 0;
3167	struct extent_map *em;
3168	int blocksize = btrfs_inode_sectorsize(inode);
3169	int ret;
3170
3171	alloc_start = round_down(offset, blocksize);
3172	alloc_end = round_up(offset + len, blocksize);
3173	cur_offset = alloc_start;
3174
3175	/* Make sure we aren't being give some crap mode */
3176	if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE |
3177		     FALLOC_FL_ZERO_RANGE))
3178		return -EOPNOTSUPP;
3179
3180	if (mode & FALLOC_FL_PUNCH_HOLE)
3181		return btrfs_punch_hole(inode, offset, len);
3182
3183	/*
3184	 * Only trigger disk allocation, don't trigger qgroup reserve
3185	 *
3186	 * For qgroup space, it will be checked later.
3187	 */
3188	if (!(mode & FALLOC_FL_ZERO_RANGE)) {
3189		ret = btrfs_alloc_data_chunk_ondemand(BTRFS_I(inode),
3190						      alloc_end - alloc_start);
3191		if (ret < 0)
3192			return ret;
3193	}
3194
3195	inode_lock(inode);
3196
3197	if (!(mode & FALLOC_FL_KEEP_SIZE) && offset + len > inode->i_size) {
3198		ret = inode_newsize_ok(inode, offset + len);
3199		if (ret)
3200			goto out;
3201	}
3202
3203	/*
3204	 * TODO: Move these two operations after we have checked
3205	 * accurate reserved space, or fallocate can still fail but
3206	 * with page truncated or size expanded.
3207	 *
3208	 * But that's a minor problem and won't do much harm BTW.
3209	 */
3210	if (alloc_start > inode->i_size) {
3211		ret = btrfs_cont_expand(inode, i_size_read(inode),
3212					alloc_start);
3213		if (ret)
3214			goto out;
3215	} else if (offset + len > inode->i_size) {
3216		/*
3217		 * If we are fallocating from the end of the file onward we
3218		 * need to zero out the end of the block if i_size lands in the
3219		 * middle of a block.
3220		 */
3221		ret = btrfs_truncate_block(inode, inode->i_size, 0, 0);
3222		if (ret)
3223			goto out;
3224	}
3225
3226	/*
3227	 * wait for ordered IO before we have any locks.  We'll loop again
3228	 * below with the locks held.
3229	 */
3230	ret = btrfs_wait_ordered_range(inode, alloc_start,
3231				       alloc_end - alloc_start);
3232	if (ret)
3233		goto out;
3234
3235	if (mode & FALLOC_FL_ZERO_RANGE) {
3236		ret = btrfs_zero_range(inode, offset, len, mode);
3237		inode_unlock(inode);
3238		return ret;
3239	}
3240
3241	locked_end = alloc_end - 1;
3242	while (1) {
3243		struct btrfs_ordered_extent *ordered;
3244
3245		/* the extent lock is ordered inside the running
3246		 * transaction
3247		 */
3248		lock_extent_bits(&BTRFS_I(inode)->io_tree, alloc_start,
3249				 locked_end, &cached_state);
3250		ordered = btrfs_lookup_first_ordered_extent(inode, locked_end);
3251
3252		if (ordered &&
3253		    ordered->file_offset + ordered->len > alloc_start &&
3254		    ordered->file_offset < alloc_end) {
3255			btrfs_put_ordered_extent(ordered);
3256			unlock_extent_cached(&BTRFS_I(inode)->io_tree,
3257					     alloc_start, locked_end,
3258					     &cached_state);
3259			/*
3260			 * we can't wait on the range with the transaction
3261			 * running or with the extent lock held
3262			 */
3263			ret = btrfs_wait_ordered_range(inode, alloc_start,
3264						       alloc_end - alloc_start);
3265			if (ret)
3266				goto out;
3267		} else {
3268			if (ordered)
3269				btrfs_put_ordered_extent(ordered);
3270			break;
3271		}
3272	}
3273
3274	/* First, check if we exceed the qgroup limit */
3275	INIT_LIST_HEAD(&reserve_list);
3276	while (cur_offset < alloc_end) {
3277		em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, cur_offset,
 
3278				      alloc_end - cur_offset, 0);
3279		if (IS_ERR(em)) {
3280			ret = PTR_ERR(em);
 
 
 
3281			break;
3282		}
3283		last_byte = min(extent_map_end(em), alloc_end);
3284		actual_end = min_t(u64, extent_map_end(em), offset + len);
3285		last_byte = ALIGN(last_byte, blocksize);
3286		if (em->block_start == EXTENT_MAP_HOLE ||
3287		    (cur_offset >= inode->i_size &&
3288		     !test_bit(EXTENT_FLAG_PREALLOC, &em->flags))) {
3289			ret = add_falloc_range(&reserve_list, cur_offset,
3290					       last_byte - cur_offset);
3291			if (ret < 0) {
3292				free_extent_map(em);
3293				break;
3294			}
3295			ret = btrfs_qgroup_reserve_data(inode, &data_reserved,
3296					cur_offset, last_byte - cur_offset);
3297			if (ret < 0) {
3298				cur_offset = last_byte;
3299				free_extent_map(em);
3300				break;
3301			}
3302		} else {
3303			/*
3304			 * Do not need to reserve unwritten extent for this
3305			 * range, free reserved data space first, otherwise
3306			 * it'll result in false ENOSPC error.
3307			 */
3308			btrfs_free_reserved_data_space(inode, data_reserved,
3309					cur_offset, last_byte - cur_offset);
3310		}
3311		free_extent_map(em);
3312		cur_offset = last_byte;
 
 
3313	}
3314
3315	/*
3316	 * If ret is still 0, means we're OK to fallocate.
3317	 * Or just cleanup the list and exit.
3318	 */
3319	list_for_each_entry_safe(range, tmp, &reserve_list, list) {
3320		if (!ret)
3321			ret = btrfs_prealloc_file_range(inode, mode,
3322					range->start,
3323					range->len, i_blocksize(inode),
3324					offset + len, &alloc_hint);
3325		else
3326			btrfs_free_reserved_data_space(inode,
3327					data_reserved, range->start,
3328					range->len);
3329		list_del(&range->list);
3330		kfree(range);
3331	}
3332	if (ret < 0)
3333		goto out_unlock;
3334
3335	/*
3336	 * We didn't need to allocate any more space, but we still extended the
3337	 * size of the file so we need to update i_size and the inode item.
3338	 */
3339	ret = btrfs_fallocate_update_isize(inode, actual_end, mode);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3340out_unlock:
3341	unlock_extent_cached(&BTRFS_I(inode)->io_tree, alloc_start, locked_end,
3342			     &cached_state);
3343out:
 
 
 
 
 
 
 
 
3344	inode_unlock(inode);
3345	/* Let go of our reservation. */
3346	if (ret != 0 && !(mode & FALLOC_FL_ZERO_RANGE))
3347		btrfs_free_reserved_data_space(inode, data_reserved,
3348				cur_offset, alloc_end - cur_offset);
3349	extent_changeset_free(data_reserved);
3350	return ret;
3351}
3352
3353static int find_desired_extent(struct inode *inode, loff_t *offset, int whence)
3354{
3355	struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
3356	struct extent_map *em = NULL;
3357	struct extent_state *cached_state = NULL;
3358	u64 lockstart;
3359	u64 lockend;
3360	u64 start;
3361	u64 len;
3362	int ret = 0;
3363
3364	if (inode->i_size == 0)
3365		return -ENXIO;
3366
3367	/*
3368	 * *offset can be negative, in this case we start finding DATA/HOLE from
3369	 * the very start of the file.
3370	 */
3371	start = max_t(loff_t, 0, *offset);
3372
3373	lockstart = round_down(start, fs_info->sectorsize);
3374	lockend = round_up(i_size_read(inode),
3375			   fs_info->sectorsize);
3376	if (lockend <= lockstart)
3377		lockend = lockstart + fs_info->sectorsize;
3378	lockend--;
3379	len = lockend - lockstart + 1;
3380
3381	lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend,
3382			 &cached_state);
3383
3384	while (start < inode->i_size) {
3385		em = btrfs_get_extent_fiemap(BTRFS_I(inode), start, len);
3386		if (IS_ERR(em)) {
3387			ret = PTR_ERR(em);
3388			em = NULL;
3389			break;
3390		}
3391
3392		if (whence == SEEK_HOLE &&
3393		    (em->block_start == EXTENT_MAP_HOLE ||
3394		     test_bit(EXTENT_FLAG_PREALLOC, &em->flags)))
3395			break;
3396		else if (whence == SEEK_DATA &&
3397			   (em->block_start != EXTENT_MAP_HOLE &&
3398			    !test_bit(EXTENT_FLAG_PREALLOC, &em->flags)))
3399			break;
3400
3401		start = em->start + em->len;
3402		free_extent_map(em);
3403		em = NULL;
3404		cond_resched();
3405	}
3406	free_extent_map(em);
3407	if (!ret) {
3408		if (whence == SEEK_DATA && start >= inode->i_size)
3409			ret = -ENXIO;
3410		else
3411			*offset = min_t(loff_t, start, inode->i_size);
3412	}
3413	unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
3414			     &cached_state);
3415	return ret;
3416}
3417
3418static loff_t btrfs_file_llseek(struct file *file, loff_t offset, int whence)
3419{
3420	struct inode *inode = file->f_mapping->host;
3421	int ret;
3422
3423	inode_lock(inode);
3424	switch (whence) {
3425	case SEEK_END:
3426	case SEEK_CUR:
3427		offset = generic_file_llseek(file, offset, whence);
3428		goto out;
3429	case SEEK_DATA:
3430	case SEEK_HOLE:
3431		if (offset >= i_size_read(inode)) {
3432			inode_unlock(inode);
3433			return -ENXIO;
3434		}
3435
3436		ret = find_desired_extent(inode, &offset, whence);
3437		if (ret) {
3438			inode_unlock(inode);
3439			return ret;
3440		}
3441	}
3442
3443	offset = vfs_setpos(file, offset, inode->i_sb->s_maxbytes);
3444out:
3445	inode_unlock(inode);
3446	return offset;
3447}
3448
3449static int btrfs_file_open(struct inode *inode, struct file *filp)
3450{
3451	filp->f_mode |= FMODE_NOWAIT;
3452	return generic_file_open(inode, filp);
3453}
3454
3455const struct file_operations btrfs_file_operations = {
3456	.llseek		= btrfs_file_llseek,
3457	.read_iter      = generic_file_read_iter,
3458	.splice_read	= generic_file_splice_read,
3459	.write_iter	= btrfs_file_write_iter,
3460	.mmap		= btrfs_file_mmap,
3461	.open		= btrfs_file_open,
3462	.release	= btrfs_release_file,
3463	.fsync		= btrfs_sync_file,
3464	.fallocate	= btrfs_fallocate,
3465	.unlocked_ioctl	= btrfs_ioctl,
3466#ifdef CONFIG_COMPAT
3467	.compat_ioctl	= btrfs_compat_ioctl,
3468#endif
3469	.remap_file_range = btrfs_remap_file_range,
 
 
3470};
3471
3472void __cold btrfs_auto_defrag_exit(void)
3473{
3474	kmem_cache_destroy(btrfs_inode_defrag_cachep);
3475}
3476
3477int __init btrfs_auto_defrag_init(void)
3478{
3479	btrfs_inode_defrag_cachep = kmem_cache_create("btrfs_inode_defrag",
3480					sizeof(struct inode_defrag), 0,
3481					SLAB_MEM_SPREAD,
3482					NULL);
3483	if (!btrfs_inode_defrag_cachep)
3484		return -ENOMEM;
3485
3486	return 0;
3487}
3488
3489int btrfs_fdatawrite_range(struct inode *inode, loff_t start, loff_t end)
3490{
3491	int ret;
3492
3493	/*
3494	 * So with compression we will find and lock a dirty page and clear the
3495	 * first one as dirty, setup an async extent, and immediately return
3496	 * with the entire range locked but with nobody actually marked with
3497	 * writeback.  So we can't just filemap_write_and_wait_range() and
3498	 * expect it to work since it will just kick off a thread to do the
3499	 * actual work.  So we need to call filemap_fdatawrite_range _again_
3500	 * since it will wait on the page lock, which won't be unlocked until
3501	 * after the pages have been marked as writeback and so we're good to go
3502	 * from there.  We have to do this otherwise we'll miss the ordered
3503	 * extents and that results in badness.  Please Josef, do not think you
3504	 * know better and pull this out at some point in the future, it is
3505	 * right and you are wrong.
3506	 */
3507	ret = filemap_fdatawrite_range(inode->i_mapping, start, end);
3508	if (!ret && test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
3509			     &BTRFS_I(inode)->runtime_flags))
3510		ret = filemap_fdatawrite_range(inode->i_mapping, start, end);
3511
3512	return ret;
3513}
v4.6
 
   1/*
   2 * Copyright (C) 2007 Oracle.  All rights reserved.
   3 *
   4 * This program is free software; you can redistribute it and/or
   5 * modify it under the terms of the GNU General Public
   6 * License v2 as published by the Free Software Foundation.
   7 *
   8 * This program is distributed in the hope that it will be useful,
   9 * but WITHOUT ANY WARRANTY; without even the implied warranty of
  10 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  11 * General Public License for more details.
  12 *
  13 * You should have received a copy of the GNU General Public
  14 * License along with this program; if not, write to the
  15 * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  16 * Boston, MA 021110-1307, USA.
  17 */
  18
  19#include <linux/fs.h>
  20#include <linux/pagemap.h>
  21#include <linux/highmem.h>
  22#include <linux/time.h>
  23#include <linux/init.h>
  24#include <linux/string.h>
  25#include <linux/backing-dev.h>
  26#include <linux/mpage.h>
  27#include <linux/falloc.h>
  28#include <linux/swap.h>
  29#include <linux/writeback.h>
  30#include <linux/statfs.h>
  31#include <linux/compat.h>
  32#include <linux/slab.h>
  33#include <linux/btrfs.h>
  34#include <linux/uio.h>
 
  35#include "ctree.h"
  36#include "disk-io.h"
  37#include "transaction.h"
  38#include "btrfs_inode.h"
  39#include "print-tree.h"
  40#include "tree-log.h"
  41#include "locking.h"
  42#include "volumes.h"
  43#include "qgroup.h"
  44#include "compression.h"
 
  45
  46static struct kmem_cache *btrfs_inode_defrag_cachep;
  47/*
  48 * when auto defrag is enabled we
  49 * queue up these defrag structs to remember which
  50 * inodes need defragging passes
  51 */
  52struct inode_defrag {
  53	struct rb_node rb_node;
  54	/* objectid */
  55	u64 ino;
  56	/*
  57	 * transid where the defrag was added, we search for
  58	 * extents newer than this
  59	 */
  60	u64 transid;
  61
  62	/* root objectid */
  63	u64 root;
  64
  65	/* last offset we were able to defrag */
  66	u64 last_offset;
  67
  68	/* if we've wrapped around back to zero once already */
  69	int cycled;
  70};
  71
  72static int __compare_inode_defrag(struct inode_defrag *defrag1,
  73				  struct inode_defrag *defrag2)
  74{
  75	if (defrag1->root > defrag2->root)
  76		return 1;
  77	else if (defrag1->root < defrag2->root)
  78		return -1;
  79	else if (defrag1->ino > defrag2->ino)
  80		return 1;
  81	else if (defrag1->ino < defrag2->ino)
  82		return -1;
  83	else
  84		return 0;
  85}
  86
  87/* pop a record for an inode into the defrag tree.  The lock
  88 * must be held already
  89 *
  90 * If you're inserting a record for an older transid than an
  91 * existing record, the transid already in the tree is lowered
  92 *
  93 * If an existing record is found the defrag item you
  94 * pass in is freed
  95 */
  96static int __btrfs_add_inode_defrag(struct inode *inode,
  97				    struct inode_defrag *defrag)
  98{
  99	struct btrfs_root *root = BTRFS_I(inode)->root;
 100	struct inode_defrag *entry;
 101	struct rb_node **p;
 102	struct rb_node *parent = NULL;
 103	int ret;
 104
 105	p = &root->fs_info->defrag_inodes.rb_node;
 106	while (*p) {
 107		parent = *p;
 108		entry = rb_entry(parent, struct inode_defrag, rb_node);
 109
 110		ret = __compare_inode_defrag(defrag, entry);
 111		if (ret < 0)
 112			p = &parent->rb_left;
 113		else if (ret > 0)
 114			p = &parent->rb_right;
 115		else {
 116			/* if we're reinserting an entry for
 117			 * an old defrag run, make sure to
 118			 * lower the transid of our existing record
 119			 */
 120			if (defrag->transid < entry->transid)
 121				entry->transid = defrag->transid;
 122			if (defrag->last_offset > entry->last_offset)
 123				entry->last_offset = defrag->last_offset;
 124			return -EEXIST;
 125		}
 126	}
 127	set_bit(BTRFS_INODE_IN_DEFRAG, &BTRFS_I(inode)->runtime_flags);
 128	rb_link_node(&defrag->rb_node, parent, p);
 129	rb_insert_color(&defrag->rb_node, &root->fs_info->defrag_inodes);
 130	return 0;
 131}
 132
 133static inline int __need_auto_defrag(struct btrfs_root *root)
 134{
 135	if (!btrfs_test_opt(root, AUTO_DEFRAG))
 136		return 0;
 137
 138	if (btrfs_fs_closing(root->fs_info))
 139		return 0;
 140
 141	return 1;
 142}
 143
 144/*
 145 * insert a defrag record for this inode if auto defrag is
 146 * enabled
 147 */
 148int btrfs_add_inode_defrag(struct btrfs_trans_handle *trans,
 149			   struct inode *inode)
 150{
 151	struct btrfs_root *root = BTRFS_I(inode)->root;
 
 152	struct inode_defrag *defrag;
 153	u64 transid;
 154	int ret;
 155
 156	if (!__need_auto_defrag(root))
 157		return 0;
 158
 159	if (test_bit(BTRFS_INODE_IN_DEFRAG, &BTRFS_I(inode)->runtime_flags))
 160		return 0;
 161
 162	if (trans)
 163		transid = trans->transid;
 164	else
 165		transid = BTRFS_I(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(&root->fs_info->defrag_inodes_lock);
 176	if (!test_bit(BTRFS_INODE_IN_DEFRAG, &BTRFS_I(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(&root->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 */
 197static void btrfs_requeue_inode_defrag(struct inode *inode,
 198				       struct inode_defrag *defrag)
 199{
 200	struct btrfs_root *root = BTRFS_I(inode)->root;
 201	int ret;
 202
 203	if (!__need_auto_defrag(root))
 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(&root->fs_info->defrag_inodes_lock);
 211	ret = __btrfs_add_inode_defrag(inode, defrag);
 212	spin_unlock(&root->fs_info->defrag_inodes_lock);
 213	if (ret)
 214		goto out;
 215	return;
 216out:
 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 */
 224static struct inode_defrag *
 225btrfs_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	}
 258out:
 259	if (entry)
 260		rb_erase(parent, &fs_info->defrag_inodes);
 261	spin_unlock(&fs_info->defrag_inodes_lock);
 262	return entry;
 263}
 264
 265void 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
 286static 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(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(inode, defrag);
 347	} else {
 348		kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
 349	}
 350
 351	iput(inode);
 352	return 0;
 353cleanup:
 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 */
 363int 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->tree_root))
 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 */
 410static 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 */
 464static 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 */
 488int btrfs_dirty_pages(struct btrfs_root *root, struct inode *inode,
 489			     struct page **pages, size_t num_pages,
 490			     loff_t pos, size_t write_bytes,
 491			     struct extent_state **cached)
 492{
 
 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)root->sectorsize - 1);
 502	num_bytes = round_up(write_bytes + pos - start_pos, root->sectorsize);
 
 503
 504	end_of_last_block = start_pos + num_bytes - 1;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 505	err = btrfs_set_extent_delalloc(inode, start_pos, end_of_last_block,
 506					cached);
 507	if (err)
 508		return err;
 509
 510	for (i = 0; i < num_pages; i++) {
 511		struct page *p = pages[i];
 512		SetPageUptodate(p);
 513		ClearPageChecked(p);
 514		set_page_dirty(p);
 515	}
 516
 517	/*
 518	 * we've only changed i_size in ram, and we haven't updated
 519	 * the disk i_size.  There is no need to log the inode
 520	 * at this time.
 521	 */
 522	if (end_pos > isize)
 523		i_size_write(inode, end_pos);
 524	return 0;
 525}
 526
 527/*
 528 * this drops all the extents in the cache that intersect the range
 529 * [start, end].  Existing extents are split as required.
 530 */
 531void btrfs_drop_extent_cache(struct inode *inode, u64 start, u64 end,
 532			     int skip_pinned)
 533{
 534	struct extent_map *em;
 535	struct extent_map *split = NULL;
 536	struct extent_map *split2 = NULL;
 537	struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
 538	u64 len = end - start + 1;
 539	u64 gen;
 540	int ret;
 541	int testend = 1;
 542	unsigned long flags;
 543	int compressed = 0;
 544	bool modified;
 545
 546	WARN_ON(end < start);
 547	if (end == (u64)-1) {
 548		len = (u64)-1;
 549		testend = 0;
 550	}
 551	while (1) {
 552		int no_splits = 0;
 553
 554		modified = false;
 555		if (!split)
 556			split = alloc_extent_map();
 557		if (!split2)
 558			split2 = alloc_extent_map();
 559		if (!split || !split2)
 560			no_splits = 1;
 561
 562		write_lock(&em_tree->lock);
 563		em = lookup_extent_mapping(em_tree, start, len);
 564		if (!em) {
 565			write_unlock(&em_tree->lock);
 566			break;
 567		}
 568		flags = em->flags;
 569		gen = em->generation;
 570		if (skip_pinned && test_bit(EXTENT_FLAG_PINNED, &em->flags)) {
 571			if (testend && em->start + em->len >= start + len) {
 572				free_extent_map(em);
 573				write_unlock(&em_tree->lock);
 574				break;
 575			}
 576			start = em->start + em->len;
 577			if (testend)
 578				len = start + len - (em->start + em->len);
 579			free_extent_map(em);
 580			write_unlock(&em_tree->lock);
 581			continue;
 582		}
 583		compressed = test_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
 584		clear_bit(EXTENT_FLAG_PINNED, &em->flags);
 585		clear_bit(EXTENT_FLAG_LOGGING, &flags);
 586		modified = !list_empty(&em->list);
 587		if (no_splits)
 588			goto next;
 589
 590		if (em->start < start) {
 591			split->start = em->start;
 592			split->len = start - em->start;
 593
 594			if (em->block_start < EXTENT_MAP_LAST_BYTE) {
 595				split->orig_start = em->orig_start;
 596				split->block_start = em->block_start;
 597
 598				if (compressed)
 599					split->block_len = em->block_len;
 600				else
 601					split->block_len = split->len;
 602				split->orig_block_len = max(split->block_len,
 603						em->orig_block_len);
 604				split->ram_bytes = em->ram_bytes;
 605			} else {
 606				split->orig_start = split->start;
 607				split->block_len = 0;
 608				split->block_start = em->block_start;
 609				split->orig_block_len = 0;
 610				split->ram_bytes = split->len;
 611			}
 612
 613			split->generation = gen;
 614			split->bdev = em->bdev;
 615			split->flags = flags;
 616			split->compress_type = em->compress_type;
 617			replace_extent_mapping(em_tree, em, split, modified);
 618			free_extent_map(split);
 619			split = split2;
 620			split2 = NULL;
 621		}
 622		if (testend && em->start + em->len > start + len) {
 623			u64 diff = start + len - em->start;
 624
 625			split->start = start + len;
 626			split->len = em->start + em->len - (start + len);
 627			split->bdev = em->bdev;
 628			split->flags = flags;
 629			split->compress_type = em->compress_type;
 630			split->generation = gen;
 631
 632			if (em->block_start < EXTENT_MAP_LAST_BYTE) {
 633				split->orig_block_len = max(em->block_len,
 634						    em->orig_block_len);
 635
 636				split->ram_bytes = em->ram_bytes;
 637				if (compressed) {
 638					split->block_len = em->block_len;
 639					split->block_start = em->block_start;
 640					split->orig_start = em->orig_start;
 641				} else {
 642					split->block_len = split->len;
 643					split->block_start = em->block_start
 644						+ diff;
 645					split->orig_start = em->orig_start;
 646				}
 647			} else {
 648				split->ram_bytes = split->len;
 649				split->orig_start = split->start;
 650				split->block_len = 0;
 651				split->block_start = em->block_start;
 652				split->orig_block_len = 0;
 653			}
 654
 655			if (extent_map_in_tree(em)) {
 656				replace_extent_mapping(em_tree, em, split,
 657						       modified);
 658			} else {
 659				ret = add_extent_mapping(em_tree, split,
 660							 modified);
 661				ASSERT(ret == 0); /* Logic error */
 662			}
 663			free_extent_map(split);
 664			split = NULL;
 665		}
 666next:
 667		if (extent_map_in_tree(em))
 668			remove_extent_mapping(em_tree, em);
 669		write_unlock(&em_tree->lock);
 670
 671		/* once for us */
 672		free_extent_map(em);
 673		/* once for the tree*/
 674		free_extent_map(em);
 675	}
 676	if (split)
 677		free_extent_map(split);
 678	if (split2)
 679		free_extent_map(split2);
 680}
 681
 682/*
 683 * this is very complex, but the basic idea is to drop all extents
 684 * in the range start - end.  hint_block is filled in with a block number
 685 * that would be a good hint to the block allocator for this file.
 686 *
 687 * If an extent intersects the range but is not entirely inside the range
 688 * it is either truncated or split.  Anything entirely inside the range
 689 * is deleted from the tree.
 690 */
 691int __btrfs_drop_extents(struct btrfs_trans_handle *trans,
 692			 struct btrfs_root *root, struct inode *inode,
 693			 struct btrfs_path *path, u64 start, u64 end,
 694			 u64 *drop_end, int drop_cache,
 695			 int replace_extent,
 696			 u32 extent_item_size,
 697			 int *key_inserted)
 698{
 
 699	struct extent_buffer *leaf;
 700	struct btrfs_file_extent_item *fi;
 
 701	struct btrfs_key key;
 702	struct btrfs_key new_key;
 703	u64 ino = btrfs_ino(inode);
 704	u64 search_start = start;
 705	u64 disk_bytenr = 0;
 706	u64 num_bytes = 0;
 707	u64 extent_offset = 0;
 708	u64 extent_end = 0;
 
 709	int del_nr = 0;
 710	int del_slot = 0;
 711	int extent_type;
 712	int recow;
 713	int ret;
 714	int modify_tree = -1;
 715	int update_refs;
 716	int found = 0;
 717	int leafs_visited = 0;
 718
 719	if (drop_cache)
 720		btrfs_drop_extent_cache(inode, start, end - 1, 0);
 721
 722	if (start >= BTRFS_I(inode)->disk_i_size && !replace_extent)
 723		modify_tree = 0;
 724
 725	update_refs = (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
 726		       root == root->fs_info->tree_root);
 727	while (1) {
 728		recow = 0;
 729		ret = btrfs_lookup_file_extent(trans, root, path, ino,
 730					       search_start, modify_tree);
 731		if (ret < 0)
 732			break;
 733		if (ret > 0 && path->slots[0] > 0 && search_start == start) {
 734			leaf = path->nodes[0];
 735			btrfs_item_key_to_cpu(leaf, &key, path->slots[0] - 1);
 736			if (key.objectid == ino &&
 737			    key.type == BTRFS_EXTENT_DATA_KEY)
 738				path->slots[0]--;
 739		}
 740		ret = 0;
 741		leafs_visited++;
 742next_slot:
 743		leaf = path->nodes[0];
 744		if (path->slots[0] >= btrfs_header_nritems(leaf)) {
 745			BUG_ON(del_nr > 0);
 746			ret = btrfs_next_leaf(root, path);
 747			if (ret < 0)
 748				break;
 749			if (ret > 0) {
 750				ret = 0;
 751				break;
 752			}
 753			leafs_visited++;
 754			leaf = path->nodes[0];
 755			recow = 1;
 756		}
 757
 758		btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 759
 760		if (key.objectid > ino)
 761			break;
 762		if (WARN_ON_ONCE(key.objectid < ino) ||
 763		    key.type < BTRFS_EXTENT_DATA_KEY) {
 764			ASSERT(del_nr == 0);
 765			path->slots[0]++;
 766			goto next_slot;
 767		}
 768		if (key.type > BTRFS_EXTENT_DATA_KEY || key.offset >= end)
 769			break;
 770
 771		fi = btrfs_item_ptr(leaf, path->slots[0],
 772				    struct btrfs_file_extent_item);
 773		extent_type = btrfs_file_extent_type(leaf, fi);
 774
 775		if (extent_type == BTRFS_FILE_EXTENT_REG ||
 776		    extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
 777			disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
 778			num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
 779			extent_offset = btrfs_file_extent_offset(leaf, fi);
 780			extent_end = key.offset +
 781				btrfs_file_extent_num_bytes(leaf, fi);
 782		} else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
 783			extent_end = key.offset +
 784				btrfs_file_extent_inline_len(leaf,
 785						     path->slots[0], fi);
 786		} else {
 787			/* can't happen */
 788			BUG();
 789		}
 790
 791		/*
 792		 * Don't skip extent items representing 0 byte lengths. They
 793		 * used to be created (bug) if while punching holes we hit
 794		 * -ENOSPC condition. So if we find one here, just ensure we
 795		 * delete it, otherwise we would insert a new file extent item
 796		 * with the same key (offset) as that 0 bytes length file
 797		 * extent item in the call to setup_items_for_insert() later
 798		 * in this function.
 799		 */
 800		if (extent_end == key.offset && extent_end >= search_start)
 
 801			goto delete_extent_item;
 
 802
 803		if (extent_end <= search_start) {
 804			path->slots[0]++;
 805			goto next_slot;
 806		}
 807
 808		found = 1;
 809		search_start = max(key.offset, start);
 810		if (recow || !modify_tree) {
 811			modify_tree = -1;
 812			btrfs_release_path(path);
 813			continue;
 814		}
 815
 816		/*
 817		 *     | - range to drop - |
 818		 *  | -------- extent -------- |
 819		 */
 820		if (start > key.offset && end < extent_end) {
 821			BUG_ON(del_nr > 0);
 822			if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
 823				ret = -EOPNOTSUPP;
 824				break;
 825			}
 826
 827			memcpy(&new_key, &key, sizeof(new_key));
 828			new_key.offset = start;
 829			ret = btrfs_duplicate_item(trans, root, path,
 830						   &new_key);
 831			if (ret == -EAGAIN) {
 832				btrfs_release_path(path);
 833				continue;
 834			}
 835			if (ret < 0)
 836				break;
 837
 838			leaf = path->nodes[0];
 839			fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
 840					    struct btrfs_file_extent_item);
 841			btrfs_set_file_extent_num_bytes(leaf, fi,
 842							start - key.offset);
 843
 844			fi = btrfs_item_ptr(leaf, path->slots[0],
 845					    struct btrfs_file_extent_item);
 846
 847			extent_offset += start - key.offset;
 848			btrfs_set_file_extent_offset(leaf, fi, extent_offset);
 849			btrfs_set_file_extent_num_bytes(leaf, fi,
 850							extent_end - start);
 851			btrfs_mark_buffer_dirty(leaf);
 852
 853			if (update_refs && disk_bytenr > 0) {
 854				ret = btrfs_inc_extent_ref(trans, root,
 855						disk_bytenr, num_bytes, 0,
 
 
 856						root->root_key.objectid,
 857						new_key.objectid,
 858						start - extent_offset);
 
 859				BUG_ON(ret); /* -ENOMEM */
 860			}
 861			key.offset = start;
 862		}
 863		/*
 
 
 
 
 
 
 864		 *  | ---- range to drop ----- |
 865		 *      | -------- extent -------- |
 866		 */
 867		if (start <= key.offset && end < extent_end) {
 868			if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
 869				ret = -EOPNOTSUPP;
 870				break;
 871			}
 872
 873			memcpy(&new_key, &key, sizeof(new_key));
 874			new_key.offset = end;
 875			btrfs_set_item_key_safe(root->fs_info, path, &new_key);
 876
 877			extent_offset += end - key.offset;
 878			btrfs_set_file_extent_offset(leaf, fi, extent_offset);
 879			btrfs_set_file_extent_num_bytes(leaf, fi,
 880							extent_end - end);
 881			btrfs_mark_buffer_dirty(leaf);
 882			if (update_refs && disk_bytenr > 0)
 883				inode_sub_bytes(inode, end - key.offset);
 884			break;
 885		}
 886
 887		search_start = extent_end;
 888		/*
 889		 *       | ---- range to drop ----- |
 890		 *  | -------- extent -------- |
 891		 */
 892		if (start > key.offset && end >= extent_end) {
 893			BUG_ON(del_nr > 0);
 894			if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
 895				ret = -EOPNOTSUPP;
 896				break;
 897			}
 898
 899			btrfs_set_file_extent_num_bytes(leaf, fi,
 900							start - key.offset);
 901			btrfs_mark_buffer_dirty(leaf);
 902			if (update_refs && disk_bytenr > 0)
 903				inode_sub_bytes(inode, extent_end - start);
 904			if (end == extent_end)
 905				break;
 906
 907			path->slots[0]++;
 908			goto next_slot;
 909		}
 910
 911		/*
 912		 *  | ---- range to drop ----- |
 913		 *    | ------ extent ------ |
 914		 */
 915		if (start <= key.offset && end >= extent_end) {
 916delete_extent_item:
 917			if (del_nr == 0) {
 918				del_slot = path->slots[0];
 919				del_nr = 1;
 920			} else {
 921				BUG_ON(del_slot + del_nr != path->slots[0]);
 922				del_nr++;
 923			}
 924
 925			if (update_refs &&
 926			    extent_type == BTRFS_FILE_EXTENT_INLINE) {
 927				inode_sub_bytes(inode,
 928						extent_end - key.offset);
 929				extent_end = ALIGN(extent_end,
 930						   root->sectorsize);
 931			} else if (update_refs && disk_bytenr > 0) {
 932				ret = btrfs_free_extent(trans, root,
 933						disk_bytenr, num_bytes, 0,
 
 
 934						root->root_key.objectid,
 935						key.objectid, key.offset -
 936						extent_offset);
 
 937				BUG_ON(ret); /* -ENOMEM */
 938				inode_sub_bytes(inode,
 939						extent_end - key.offset);
 940			}
 941
 942			if (end == extent_end)
 943				break;
 944
 945			if (path->slots[0] + 1 < btrfs_header_nritems(leaf)) {
 946				path->slots[0]++;
 947				goto next_slot;
 948			}
 949
 950			ret = btrfs_del_items(trans, root, path, del_slot,
 951					      del_nr);
 952			if (ret) {
 953				btrfs_abort_transaction(trans, root, ret);
 954				break;
 955			}
 956
 957			del_nr = 0;
 958			del_slot = 0;
 959
 960			btrfs_release_path(path);
 961			continue;
 962		}
 963
 964		BUG_ON(1);
 965	}
 966
 967	if (!ret && del_nr > 0) {
 968		/*
 969		 * Set path->slots[0] to first slot, so that after the delete
 970		 * if items are move off from our leaf to its immediate left or
 971		 * right neighbor leafs, we end up with a correct and adjusted
 972		 * path->slots[0] for our insertion (if replace_extent != 0).
 973		 */
 974		path->slots[0] = del_slot;
 975		ret = btrfs_del_items(trans, root, path, del_slot, del_nr);
 976		if (ret)
 977			btrfs_abort_transaction(trans, root, ret);
 978	}
 979
 980	leaf = path->nodes[0];
 981	/*
 982	 * If btrfs_del_items() was called, it might have deleted a leaf, in
 983	 * which case it unlocked our path, so check path->locks[0] matches a
 984	 * write lock.
 985	 */
 986	if (!ret && replace_extent && leafs_visited == 1 &&
 987	    (path->locks[0] == BTRFS_WRITE_LOCK_BLOCKING ||
 988	     path->locks[0] == BTRFS_WRITE_LOCK) &&
 989	    btrfs_leaf_free_space(root, leaf) >=
 990	    sizeof(struct btrfs_item) + extent_item_size) {
 991
 992		key.objectid = ino;
 993		key.type = BTRFS_EXTENT_DATA_KEY;
 994		key.offset = start;
 995		if (!del_nr && path->slots[0] < btrfs_header_nritems(leaf)) {
 996			struct btrfs_key slot_key;
 997
 998			btrfs_item_key_to_cpu(leaf, &slot_key, path->slots[0]);
 999			if (btrfs_comp_cpu_keys(&key, &slot_key) > 0)
1000				path->slots[0]++;
1001		}
1002		setup_items_for_insert(root, path, &key,
1003				       &extent_item_size,
1004				       extent_item_size,
1005				       sizeof(struct btrfs_item) +
1006				       extent_item_size, 1);
1007		*key_inserted = 1;
1008	}
1009
1010	if (!replace_extent || !(*key_inserted))
1011		btrfs_release_path(path);
1012	if (drop_end)
1013		*drop_end = found ? min(end, extent_end) : end;
1014	return ret;
1015}
1016
1017int btrfs_drop_extents(struct btrfs_trans_handle *trans,
1018		       struct btrfs_root *root, struct inode *inode, u64 start,
1019		       u64 end, int drop_cache)
1020{
1021	struct btrfs_path *path;
1022	int ret;
1023
1024	path = btrfs_alloc_path();
1025	if (!path)
1026		return -ENOMEM;
1027	ret = __btrfs_drop_extents(trans, root, inode, path, start, end, NULL,
1028				   drop_cache, 0, 0, NULL);
1029	btrfs_free_path(path);
1030	return ret;
1031}
1032
1033static int extent_mergeable(struct extent_buffer *leaf, int slot,
1034			    u64 objectid, u64 bytenr, u64 orig_offset,
1035			    u64 *start, u64 *end)
1036{
1037	struct btrfs_file_extent_item *fi;
1038	struct btrfs_key key;
1039	u64 extent_end;
1040
1041	if (slot < 0 || slot >= btrfs_header_nritems(leaf))
1042		return 0;
1043
1044	btrfs_item_key_to_cpu(leaf, &key, slot);
1045	if (key.objectid != objectid || key.type != BTRFS_EXTENT_DATA_KEY)
1046		return 0;
1047
1048	fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
1049	if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG ||
1050	    btrfs_file_extent_disk_bytenr(leaf, fi) != bytenr ||
1051	    btrfs_file_extent_offset(leaf, fi) != key.offset - orig_offset ||
1052	    btrfs_file_extent_compression(leaf, fi) ||
1053	    btrfs_file_extent_encryption(leaf, fi) ||
1054	    btrfs_file_extent_other_encoding(leaf, fi))
1055		return 0;
1056
1057	extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
1058	if ((*start && *start != key.offset) || (*end && *end != extent_end))
1059		return 0;
1060
1061	*start = key.offset;
1062	*end = extent_end;
1063	return 1;
1064}
1065
1066/*
1067 * Mark extent in the range start - end as written.
1068 *
1069 * This changes extent type from 'pre-allocated' to 'regular'. If only
1070 * part of extent is marked as written, the extent will be split into
1071 * two or three.
1072 */
1073int btrfs_mark_extent_written(struct btrfs_trans_handle *trans,
1074			      struct inode *inode, u64 start, u64 end)
1075{
1076	struct btrfs_root *root = BTRFS_I(inode)->root;
 
1077	struct extent_buffer *leaf;
1078	struct btrfs_path *path;
1079	struct btrfs_file_extent_item *fi;
 
1080	struct btrfs_key key;
1081	struct btrfs_key new_key;
1082	u64 bytenr;
1083	u64 num_bytes;
1084	u64 extent_end;
1085	u64 orig_offset;
1086	u64 other_start;
1087	u64 other_end;
1088	u64 split;
1089	int del_nr = 0;
1090	int del_slot = 0;
1091	int recow;
1092	int ret;
1093	u64 ino = btrfs_ino(inode);
1094
1095	path = btrfs_alloc_path();
1096	if (!path)
1097		return -ENOMEM;
1098again:
1099	recow = 0;
1100	split = start;
1101	key.objectid = ino;
1102	key.type = BTRFS_EXTENT_DATA_KEY;
1103	key.offset = split;
1104
1105	ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1106	if (ret < 0)
1107		goto out;
1108	if (ret > 0 && path->slots[0] > 0)
1109		path->slots[0]--;
1110
1111	leaf = path->nodes[0];
1112	btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1113	BUG_ON(key.objectid != ino || key.type != BTRFS_EXTENT_DATA_KEY);
 
 
 
 
 
1114	fi = btrfs_item_ptr(leaf, path->slots[0],
1115			    struct btrfs_file_extent_item);
1116	BUG_ON(btrfs_file_extent_type(leaf, fi) !=
1117	       BTRFS_FILE_EXTENT_PREALLOC);
 
 
 
1118	extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
1119	BUG_ON(key.offset > start || extent_end < end);
 
 
 
 
1120
1121	bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1122	num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
1123	orig_offset = key.offset - btrfs_file_extent_offset(leaf, fi);
1124	memcpy(&new_key, &key, sizeof(new_key));
1125
1126	if (start == key.offset && end < extent_end) {
1127		other_start = 0;
1128		other_end = start;
1129		if (extent_mergeable(leaf, path->slots[0] - 1,
1130				     ino, bytenr, orig_offset,
1131				     &other_start, &other_end)) {
1132			new_key.offset = end;
1133			btrfs_set_item_key_safe(root->fs_info, path, &new_key);
1134			fi = btrfs_item_ptr(leaf, path->slots[0],
1135					    struct btrfs_file_extent_item);
1136			btrfs_set_file_extent_generation(leaf, fi,
1137							 trans->transid);
1138			btrfs_set_file_extent_num_bytes(leaf, fi,
1139							extent_end - end);
1140			btrfs_set_file_extent_offset(leaf, fi,
1141						     end - orig_offset);
1142			fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
1143					    struct btrfs_file_extent_item);
1144			btrfs_set_file_extent_generation(leaf, fi,
1145							 trans->transid);
1146			btrfs_set_file_extent_num_bytes(leaf, fi,
1147							end - other_start);
1148			btrfs_mark_buffer_dirty(leaf);
1149			goto out;
1150		}
1151	}
1152
1153	if (start > key.offset && end == extent_end) {
1154		other_start = end;
1155		other_end = 0;
1156		if (extent_mergeable(leaf, path->slots[0] + 1,
1157				     ino, bytenr, orig_offset,
1158				     &other_start, &other_end)) {
1159			fi = btrfs_item_ptr(leaf, path->slots[0],
1160					    struct btrfs_file_extent_item);
1161			btrfs_set_file_extent_num_bytes(leaf, fi,
1162							start - key.offset);
1163			btrfs_set_file_extent_generation(leaf, fi,
1164							 trans->transid);
1165			path->slots[0]++;
1166			new_key.offset = start;
1167			btrfs_set_item_key_safe(root->fs_info, path, &new_key);
1168
1169			fi = btrfs_item_ptr(leaf, path->slots[0],
1170					    struct btrfs_file_extent_item);
1171			btrfs_set_file_extent_generation(leaf, fi,
1172							 trans->transid);
1173			btrfs_set_file_extent_num_bytes(leaf, fi,
1174							other_end - start);
1175			btrfs_set_file_extent_offset(leaf, fi,
1176						     start - orig_offset);
1177			btrfs_mark_buffer_dirty(leaf);
1178			goto out;
1179		}
1180	}
1181
1182	while (start > key.offset || end < extent_end) {
1183		if (key.offset == start)
1184			split = end;
1185
1186		new_key.offset = split;
1187		ret = btrfs_duplicate_item(trans, root, path, &new_key);
1188		if (ret == -EAGAIN) {
1189			btrfs_release_path(path);
1190			goto again;
1191		}
1192		if (ret < 0) {
1193			btrfs_abort_transaction(trans, root, ret);
1194			goto out;
1195		}
1196
1197		leaf = path->nodes[0];
1198		fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
1199				    struct btrfs_file_extent_item);
1200		btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1201		btrfs_set_file_extent_num_bytes(leaf, fi,
1202						split - key.offset);
1203
1204		fi = btrfs_item_ptr(leaf, path->slots[0],
1205				    struct btrfs_file_extent_item);
1206
1207		btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1208		btrfs_set_file_extent_offset(leaf, fi, split - orig_offset);
1209		btrfs_set_file_extent_num_bytes(leaf, fi,
1210						extent_end - split);
1211		btrfs_mark_buffer_dirty(leaf);
1212
1213		ret = btrfs_inc_extent_ref(trans, root, bytenr, num_bytes, 0,
1214					   root->root_key.objectid,
1215					   ino, orig_offset);
1216		BUG_ON(ret); /* -ENOMEM */
 
 
 
 
 
1217
1218		if (split == start) {
1219			key.offset = start;
1220		} else {
1221			BUG_ON(start != key.offset);
 
 
 
 
1222			path->slots[0]--;
1223			extent_end = end;
1224		}
1225		recow = 1;
1226	}
1227
1228	other_start = end;
1229	other_end = 0;
 
 
 
1230	if (extent_mergeable(leaf, path->slots[0] + 1,
1231			     ino, bytenr, orig_offset,
1232			     &other_start, &other_end)) {
1233		if (recow) {
1234			btrfs_release_path(path);
1235			goto again;
1236		}
1237		extent_end = other_end;
1238		del_slot = path->slots[0] + 1;
1239		del_nr++;
1240		ret = btrfs_free_extent(trans, root, bytenr, num_bytes,
1241					0, root->root_key.objectid,
1242					ino, orig_offset);
1243		BUG_ON(ret); /* -ENOMEM */
 
1244	}
1245	other_start = 0;
1246	other_end = start;
1247	if (extent_mergeable(leaf, path->slots[0] - 1,
1248			     ino, bytenr, orig_offset,
1249			     &other_start, &other_end)) {
1250		if (recow) {
1251			btrfs_release_path(path);
1252			goto again;
1253		}
1254		key.offset = other_start;
1255		del_slot = path->slots[0];
1256		del_nr++;
1257		ret = btrfs_free_extent(trans, root, bytenr, num_bytes,
1258					0, root->root_key.objectid,
1259					ino, orig_offset);
1260		BUG_ON(ret); /* -ENOMEM */
 
1261	}
1262	if (del_nr == 0) {
1263		fi = btrfs_item_ptr(leaf, path->slots[0],
1264			   struct btrfs_file_extent_item);
1265		btrfs_set_file_extent_type(leaf, fi,
1266					   BTRFS_FILE_EXTENT_REG);
1267		btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1268		btrfs_mark_buffer_dirty(leaf);
1269	} else {
1270		fi = btrfs_item_ptr(leaf, del_slot - 1,
1271			   struct btrfs_file_extent_item);
1272		btrfs_set_file_extent_type(leaf, fi,
1273					   BTRFS_FILE_EXTENT_REG);
1274		btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1275		btrfs_set_file_extent_num_bytes(leaf, fi,
1276						extent_end - key.offset);
1277		btrfs_mark_buffer_dirty(leaf);
1278
1279		ret = btrfs_del_items(trans, root, path, del_slot, del_nr);
1280		if (ret < 0) {
1281			btrfs_abort_transaction(trans, root, ret);
1282			goto out;
1283		}
1284	}
1285out:
1286	btrfs_free_path(path);
1287	return 0;
1288}
1289
1290/*
1291 * on error we return an unlocked page and the error value
1292 * on success we return a locked page and 0
1293 */
1294static int prepare_uptodate_page(struct inode *inode,
1295				 struct page *page, u64 pos,
1296				 bool force_uptodate)
1297{
1298	int ret = 0;
1299
1300	if (((pos & (PAGE_SIZE - 1)) || force_uptodate) &&
1301	    !PageUptodate(page)) {
1302		ret = btrfs_readpage(NULL, page);
1303		if (ret)
1304			return ret;
1305		lock_page(page);
1306		if (!PageUptodate(page)) {
1307			unlock_page(page);
1308			return -EIO;
1309		}
1310		if (page->mapping != inode->i_mapping) {
1311			unlock_page(page);
1312			return -EAGAIN;
1313		}
1314	}
1315	return 0;
1316}
1317
1318/*
1319 * this just gets pages into the page cache and locks them down.
1320 */
1321static noinline int prepare_pages(struct inode *inode, struct page **pages,
1322				  size_t num_pages, loff_t pos,
1323				  size_t write_bytes, bool force_uptodate)
1324{
1325	int i;
1326	unsigned long index = pos >> PAGE_SHIFT;
1327	gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping);
1328	int err = 0;
1329	int faili;
1330
1331	for (i = 0; i < num_pages; i++) {
1332again:
1333		pages[i] = find_or_create_page(inode->i_mapping, index + i,
1334					       mask | __GFP_WRITE);
1335		if (!pages[i]) {
1336			faili = i - 1;
1337			err = -ENOMEM;
1338			goto fail;
1339		}
1340
1341		if (i == 0)
1342			err = prepare_uptodate_page(inode, pages[i], pos,
1343						    force_uptodate);
1344		if (!err && i == num_pages - 1)
1345			err = prepare_uptodate_page(inode, pages[i],
1346						    pos + write_bytes, false);
1347		if (err) {
1348			put_page(pages[i]);
1349			if (err == -EAGAIN) {
1350				err = 0;
1351				goto again;
1352			}
1353			faili = i - 1;
1354			goto fail;
1355		}
1356		wait_on_page_writeback(pages[i]);
1357	}
1358
1359	return 0;
1360fail:
1361	while (faili >= 0) {
1362		unlock_page(pages[faili]);
1363		put_page(pages[faili]);
1364		faili--;
1365	}
1366	return err;
1367
1368}
1369
1370/*
1371 * This function locks the extent and properly waits for data=ordered extents
1372 * to finish before allowing the pages to be modified if need.
1373 *
1374 * The return value:
1375 * 1 - the extent is locked
1376 * 0 - the extent is not locked, and everything is OK
1377 * -EAGAIN - need re-prepare the pages
1378 * the other < 0 number - Something wrong happens
1379 */
1380static noinline int
1381lock_and_cleanup_extent_if_need(struct inode *inode, struct page **pages,
1382				size_t num_pages, loff_t pos,
1383				size_t write_bytes,
1384				u64 *lockstart, u64 *lockend,
1385				struct extent_state **cached_state)
1386{
1387	struct btrfs_root *root = BTRFS_I(inode)->root;
1388	u64 start_pos;
1389	u64 last_pos;
1390	int i;
1391	int ret = 0;
1392
1393	start_pos = round_down(pos, root->sectorsize);
1394	last_pos = start_pos
1395		+ round_up(pos + write_bytes - start_pos, root->sectorsize) - 1;
 
1396
1397	if (start_pos < inode->i_size) {
1398		struct btrfs_ordered_extent *ordered;
1399		lock_extent_bits(&BTRFS_I(inode)->io_tree,
1400				 start_pos, last_pos, cached_state);
 
1401		ordered = btrfs_lookup_ordered_range(inode, start_pos,
1402						     last_pos - start_pos + 1);
1403		if (ordered &&
1404		    ordered->file_offset + ordered->len > start_pos &&
1405		    ordered->file_offset <= last_pos) {
1406			unlock_extent_cached(&BTRFS_I(inode)->io_tree,
1407					     start_pos, last_pos,
1408					     cached_state, GFP_NOFS);
1409			for (i = 0; i < num_pages; i++) {
1410				unlock_page(pages[i]);
1411				put_page(pages[i]);
1412			}
1413			btrfs_start_ordered_extent(inode, ordered, 1);
 
1414			btrfs_put_ordered_extent(ordered);
1415			return -EAGAIN;
1416		}
1417		if (ordered)
1418			btrfs_put_ordered_extent(ordered);
1419
1420		clear_extent_bit(&BTRFS_I(inode)->io_tree, start_pos,
1421				  last_pos, EXTENT_DIRTY | EXTENT_DELALLOC |
1422				  EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG,
1423				  0, 0, cached_state, GFP_NOFS);
1424		*lockstart = start_pos;
1425		*lockend = last_pos;
1426		ret = 1;
1427	}
1428
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1429	for (i = 0; i < num_pages; i++) {
1430		if (clear_page_dirty_for_io(pages[i]))
1431			account_page_redirty(pages[i]);
1432		set_page_extent_mapped(pages[i]);
1433		WARN_ON(!PageLocked(pages[i]));
1434	}
1435
1436	return ret;
1437}
1438
1439static noinline int check_can_nocow(struct inode *inode, loff_t pos,
1440				    size_t *write_bytes)
1441{
1442	struct btrfs_root *root = BTRFS_I(inode)->root;
1443	struct btrfs_ordered_extent *ordered;
1444	u64 lockstart, lockend;
1445	u64 num_bytes;
1446	int ret;
1447
1448	ret = btrfs_start_write_no_snapshoting(root);
1449	if (!ret)
1450		return -ENOSPC;
1451
1452	lockstart = round_down(pos, root->sectorsize);
1453	lockend = round_up(pos + *write_bytes, root->sectorsize) - 1;
 
1454
1455	while (1) {
1456		lock_extent(&BTRFS_I(inode)->io_tree, lockstart, lockend);
1457		ordered = btrfs_lookup_ordered_range(inode, lockstart,
1458						     lockend - lockstart + 1);
1459		if (!ordered) {
1460			break;
1461		}
1462		unlock_extent(&BTRFS_I(inode)->io_tree, lockstart, lockend);
1463		btrfs_start_ordered_extent(inode, ordered, 1);
1464		btrfs_put_ordered_extent(ordered);
1465	}
1466
1467	num_bytes = lockend - lockstart + 1;
1468	ret = can_nocow_extent(inode, lockstart, &num_bytes, NULL, NULL, NULL);
 
1469	if (ret <= 0) {
1470		ret = 0;
1471		btrfs_end_write_no_snapshoting(root);
1472	} else {
1473		*write_bytes = min_t(size_t, *write_bytes ,
1474				     num_bytes - pos + lockstart);
1475	}
1476
1477	unlock_extent(&BTRFS_I(inode)->io_tree, lockstart, lockend);
1478
1479	return ret;
1480}
1481
1482static noinline ssize_t __btrfs_buffered_write(struct file *file,
1483					       struct iov_iter *i,
1484					       loff_t pos)
1485{
 
 
1486	struct inode *inode = file_inode(file);
 
1487	struct btrfs_root *root = BTRFS_I(inode)->root;
1488	struct page **pages = NULL;
1489	struct extent_state *cached_state = NULL;
1490	u64 release_bytes = 0;
1491	u64 lockstart;
1492	u64 lockend;
1493	size_t num_written = 0;
1494	int nrptrs;
1495	int ret = 0;
1496	bool only_release_metadata = false;
1497	bool force_page_uptodate = false;
1498	bool need_unlock;
1499
1500	nrptrs = min(DIV_ROUND_UP(iov_iter_count(i), PAGE_SIZE),
1501			PAGE_SIZE / (sizeof(struct page *)));
1502	nrptrs = min(nrptrs, current->nr_dirtied_pause - current->nr_dirtied);
1503	nrptrs = max(nrptrs, 8);
1504	pages = kmalloc_array(nrptrs, sizeof(struct page *), GFP_KERNEL);
1505	if (!pages)
1506		return -ENOMEM;
1507
1508	while (iov_iter_count(i) > 0) {
1509		size_t offset = pos & (PAGE_SIZE - 1);
 
1510		size_t sector_offset;
1511		size_t write_bytes = min(iov_iter_count(i),
1512					 nrptrs * (size_t)PAGE_SIZE -
1513					 offset);
1514		size_t num_pages = DIV_ROUND_UP(write_bytes + offset,
1515						PAGE_SIZE);
1516		size_t reserve_bytes;
1517		size_t dirty_pages;
1518		size_t copied;
1519		size_t dirty_sectors;
1520		size_t num_sectors;
 
1521
1522		WARN_ON(num_pages > nrptrs);
1523
1524		/*
1525		 * Fault pages before locking them in prepare_pages
1526		 * to avoid recursive lock
1527		 */
1528		if (unlikely(iov_iter_fault_in_readable(i, write_bytes))) {
1529			ret = -EFAULT;
1530			break;
1531		}
1532
1533		sector_offset = pos & (root->sectorsize - 1);
1534		reserve_bytes = round_up(write_bytes + sector_offset,
1535				root->sectorsize);
1536
1537		if ((BTRFS_I(inode)->flags & (BTRFS_INODE_NODATACOW |
1538					      BTRFS_INODE_PREALLOC)) &&
1539		    check_can_nocow(inode, pos, &write_bytes) > 0) {
1540			/*
1541			 * For nodata cow case, no need to reserve
1542			 * data space.
1543			 */
1544			only_release_metadata = true;
1545			/*
1546			 * our prealloc extent may be smaller than
1547			 * write_bytes, so scale down.
1548			 */
1549			num_pages = DIV_ROUND_UP(write_bytes + offset,
1550						 PAGE_SIZE);
1551			reserve_bytes = round_up(write_bytes + sector_offset,
1552					root->sectorsize);
1553			goto reserve_metadata;
 
 
 
 
 
 
 
 
1554		}
1555
1556		ret = btrfs_check_data_free_space(inode, pos, write_bytes);
1557		if (ret < 0)
1558			break;
1559
1560reserve_metadata:
1561		ret = btrfs_delalloc_reserve_metadata(inode, reserve_bytes);
1562		if (ret) {
1563			if (!only_release_metadata)
1564				btrfs_free_reserved_data_space(inode, pos,
1565							       write_bytes);
 
1566			else
1567				btrfs_end_write_no_snapshoting(root);
1568			break;
1569		}
1570
1571		release_bytes = reserve_bytes;
1572		need_unlock = false;
1573again:
1574		/*
1575		 * This is going to setup the pages array with the number of
1576		 * pages we want, so we don't really need to worry about the
1577		 * contents of pages from loop to loop
1578		 */
1579		ret = prepare_pages(inode, pages, num_pages,
1580				    pos, write_bytes,
1581				    force_page_uptodate);
1582		if (ret)
 
 
1583			break;
 
1584
1585		ret = lock_and_cleanup_extent_if_need(inode, pages, num_pages,
1586						pos, write_bytes, &lockstart,
1587						&lockend, &cached_state);
1588		if (ret < 0) {
1589			if (ret == -EAGAIN)
 
1590				goto again;
 
 
 
1591			break;
1592		} else if (ret > 0) {
1593			need_unlock = true;
1594			ret = 0;
1595		}
1596
1597		copied = btrfs_copy_from_user(pos, write_bytes, pages, i);
1598
 
 
 
 
 
1599		/*
1600		 * if we have trouble faulting in the pages, fall
1601		 * back to one page at a time
1602		 */
1603		if (copied < write_bytes)
1604			nrptrs = 1;
1605
1606		if (copied == 0) {
1607			force_page_uptodate = true;
 
1608			dirty_pages = 0;
1609		} else {
1610			force_page_uptodate = false;
1611			dirty_pages = DIV_ROUND_UP(copied + offset,
1612						   PAGE_SIZE);
1613		}
1614
1615		/*
1616		 * If we had a short copy we need to release the excess delaloc
1617		 * bytes we reserved.  We need to increment outstanding_extents
1618		 * because btrfs_delalloc_release_space will decrement it, but
1619		 * we still have an outstanding extent for the chunk we actually
1620		 * managed to copy.
1621		 */
1622		num_sectors = BTRFS_BYTES_TO_BLKS(root->fs_info,
1623						reserve_bytes);
1624		dirty_sectors = round_up(copied + sector_offset,
1625					root->sectorsize);
1626		dirty_sectors = BTRFS_BYTES_TO_BLKS(root->fs_info,
1627						dirty_sectors);
1628
1629		if (num_sectors > dirty_sectors) {
1630			release_bytes = (write_bytes - copied)
1631				& ~((u64)root->sectorsize - 1);
1632			if (copied > 0) {
1633				spin_lock(&BTRFS_I(inode)->lock);
1634				BTRFS_I(inode)->outstanding_extents++;
1635				spin_unlock(&BTRFS_I(inode)->lock);
1636			}
1637			if (only_release_metadata) {
1638				btrfs_delalloc_release_metadata(inode,
1639								release_bytes);
1640			} else {
1641				u64 __pos;
1642
1643				__pos = round_down(pos, root->sectorsize) +
 
1644					(dirty_pages << PAGE_SHIFT);
1645				btrfs_delalloc_release_space(inode, __pos,
1646							     release_bytes);
 
1647			}
1648		}
1649
1650		release_bytes = round_up(copied + sector_offset,
1651					root->sectorsize);
1652
1653		if (copied > 0)
1654			ret = btrfs_dirty_pages(root, inode, pages,
1655						dirty_pages, pos, copied,
1656						NULL);
1657		if (need_unlock)
 
 
 
 
 
 
 
1658			unlock_extent_cached(&BTRFS_I(inode)->io_tree,
1659					     lockstart, lockend, &cached_state,
1660					     GFP_NOFS);
 
 
 
1661		if (ret) {
1662			btrfs_drop_pages(pages, num_pages);
1663			break;
1664		}
1665
1666		release_bytes = 0;
1667		if (only_release_metadata)
1668			btrfs_end_write_no_snapshoting(root);
1669
1670		if (only_release_metadata && copied > 0) {
1671			lockstart = round_down(pos, root->sectorsize);
1672			lockend = round_up(pos + copied, root->sectorsize) - 1;
 
 
1673
1674			set_extent_bit(&BTRFS_I(inode)->io_tree, lockstart,
1675				       lockend, EXTENT_NORESERVE, NULL,
1676				       NULL, GFP_NOFS);
1677			only_release_metadata = false;
1678		}
1679
1680		btrfs_drop_pages(pages, num_pages);
1681
1682		cond_resched();
1683
1684		balance_dirty_pages_ratelimited(inode->i_mapping);
1685		if (dirty_pages < (root->nodesize >> PAGE_SHIFT) + 1)
1686			btrfs_btree_balance_dirty(root);
1687
1688		pos += copied;
1689		num_written += copied;
1690	}
1691
1692	kfree(pages);
1693
1694	if (release_bytes) {
1695		if (only_release_metadata) {
1696			btrfs_end_write_no_snapshoting(root);
1697			btrfs_delalloc_release_metadata(inode, release_bytes);
 
1698		} else {
1699			btrfs_delalloc_release_space(inode, pos, release_bytes);
 
 
1700		}
1701	}
1702
 
1703	return num_written ? num_written : ret;
1704}
1705
1706static ssize_t __btrfs_direct_write(struct kiocb *iocb,
1707				    struct iov_iter *from,
1708				    loff_t pos)
1709{
1710	struct file *file = iocb->ki_filp;
1711	struct inode *inode = file_inode(file);
 
1712	ssize_t written;
1713	ssize_t written_buffered;
1714	loff_t endbyte;
1715	int err;
1716
1717	written = generic_file_direct_write(iocb, from, pos);
1718
1719	if (written < 0 || !iov_iter_count(from))
1720		return written;
1721
1722	pos += written;
1723	written_buffered = __btrfs_buffered_write(file, from, pos);
1724	if (written_buffered < 0) {
1725		err = written_buffered;
1726		goto out;
1727	}
1728	/*
1729	 * Ensure all data is persisted. We want the next direct IO read to be
1730	 * able to read what was just written.
1731	 */
1732	endbyte = pos + written_buffered - 1;
1733	err = btrfs_fdatawrite_range(inode, pos, endbyte);
1734	if (err)
1735		goto out;
1736	err = filemap_fdatawait_range(inode->i_mapping, pos, endbyte);
1737	if (err)
1738		goto out;
1739	written += written_buffered;
1740	iocb->ki_pos = pos + written_buffered;
1741	invalidate_mapping_pages(file->f_mapping, pos >> PAGE_SHIFT,
1742				 endbyte >> PAGE_SHIFT);
1743out:
1744	return written ? written : err;
1745}
1746
1747static void update_time_for_write(struct inode *inode)
1748{
1749	struct timespec now;
1750
1751	if (IS_NOCMTIME(inode))
1752		return;
1753
1754	now = current_fs_time(inode->i_sb);
1755	if (!timespec_equal(&inode->i_mtime, &now))
1756		inode->i_mtime = now;
1757
1758	if (!timespec_equal(&inode->i_ctime, &now))
1759		inode->i_ctime = now;
1760
1761	if (IS_I_VERSION(inode))
1762		inode_inc_iversion(inode);
1763}
1764
1765static ssize_t btrfs_file_write_iter(struct kiocb *iocb,
1766				    struct iov_iter *from)
1767{
1768	struct file *file = iocb->ki_filp;
1769	struct inode *inode = file_inode(file);
 
1770	struct btrfs_root *root = BTRFS_I(inode)->root;
1771	u64 start_pos;
1772	u64 end_pos;
1773	ssize_t num_written = 0;
1774	bool sync = (file->f_flags & O_DSYNC) || IS_SYNC(file->f_mapping->host);
1775	ssize_t err;
1776	loff_t pos;
1777	size_t count;
1778	loff_t oldsize;
1779	int clean_page = 0;
1780
1781	inode_lock(inode);
 
 
 
 
 
 
 
 
 
1782	err = generic_write_checks(iocb, from);
1783	if (err <= 0) {
1784		inode_unlock(inode);
1785		return err;
1786	}
1787
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1788	current->backing_dev_info = inode_to_bdi(inode);
1789	err = file_remove_privs(file);
1790	if (err) {
1791		inode_unlock(inode);
1792		goto out;
1793	}
1794
1795	/*
1796	 * If BTRFS flips readonly due to some impossible error
1797	 * (fs_info->fs_state now has BTRFS_SUPER_FLAG_ERROR),
1798	 * although we have opened a file as writable, we have
1799	 * to stop this write operation to ensure FS consistency.
1800	 */
1801	if (test_bit(BTRFS_FS_STATE_ERROR, &root->fs_info->fs_state)) {
1802		inode_unlock(inode);
1803		err = -EROFS;
1804		goto out;
1805	}
1806
1807	/*
1808	 * We reserve space for updating the inode when we reserve space for the
1809	 * extent we are going to write, so we will enospc out there.  We don't
1810	 * need to start yet another transaction to update the inode as we will
1811	 * update the inode when we finish writing whatever data we write.
1812	 */
1813	update_time_for_write(inode);
1814
1815	pos = iocb->ki_pos;
1816	count = iov_iter_count(from);
1817	start_pos = round_down(pos, root->sectorsize);
1818	oldsize = i_size_read(inode);
1819	if (start_pos > oldsize) {
1820		/* Expand hole size to cover write data, preventing empty gap */
1821		end_pos = round_up(pos + count, root->sectorsize);
 
1822		err = btrfs_cont_expand(inode, oldsize, end_pos);
1823		if (err) {
1824			inode_unlock(inode);
1825			goto out;
1826		}
1827		if (start_pos > round_up(oldsize, root->sectorsize))
1828			clean_page = 1;
1829	}
1830
1831	if (sync)
1832		atomic_inc(&BTRFS_I(inode)->sync_writers);
1833
1834	if (iocb->ki_flags & IOCB_DIRECT) {
1835		num_written = __btrfs_direct_write(iocb, from, pos);
1836	} else {
1837		num_written = __btrfs_buffered_write(file, from, pos);
1838		if (num_written > 0)
1839			iocb->ki_pos = pos + num_written;
1840		if (clean_page)
1841			pagecache_isize_extended(inode, oldsize,
1842						i_size_read(inode));
1843	}
1844
1845	inode_unlock(inode);
1846
1847	/*
1848	 * We also have to set last_sub_trans to the current log transid,
1849	 * otherwise subsequent syncs to a file that's been synced in this
1850	 * transaction will appear to have already occurred.
1851	 */
1852	spin_lock(&BTRFS_I(inode)->lock);
1853	BTRFS_I(inode)->last_sub_trans = root->log_transid;
1854	spin_unlock(&BTRFS_I(inode)->lock);
1855	if (num_written > 0) {
1856		err = generic_write_sync(file, pos, num_written);
1857		if (err < 0)
1858			num_written = err;
1859	}
1860
1861	if (sync)
1862		atomic_dec(&BTRFS_I(inode)->sync_writers);
1863out:
1864	current->backing_dev_info = NULL;
1865	return num_written ? num_written : err;
1866}
1867
1868int btrfs_release_file(struct inode *inode, struct file *filp)
1869{
1870	if (filp->private_data)
1871		btrfs_ioctl_trans_end(filp);
 
 
 
 
 
1872	/*
1873	 * ordered_data_close is set by settattr when we are about to truncate
1874	 * a file from a non-zero size to a zero size.  This tries to
1875	 * flush down new bytes that may have been written if the
1876	 * application were using truncate to replace a file in place.
1877	 */
1878	if (test_and_clear_bit(BTRFS_INODE_ORDERED_DATA_CLOSE,
1879			       &BTRFS_I(inode)->runtime_flags))
1880			filemap_flush(inode->i_mapping);
1881	return 0;
1882}
1883
1884static int start_ordered_ops(struct inode *inode, loff_t start, loff_t end)
1885{
1886	int ret;
 
1887
 
 
 
 
 
 
 
1888	atomic_inc(&BTRFS_I(inode)->sync_writers);
1889	ret = btrfs_fdatawrite_range(inode, start, end);
1890	atomic_dec(&BTRFS_I(inode)->sync_writers);
 
1891
1892	return ret;
1893}
1894
1895/*
1896 * fsync call for both files and directories.  This logs the inode into
1897 * the tree log instead of forcing full commits whenever possible.
1898 *
1899 * It needs to call filemap_fdatawait so that all ordered extent updates are
1900 * in the metadata btree are up to date for copying to the log.
1901 *
1902 * It drops the inode mutex before doing the tree log commit.  This is an
1903 * important optimization for directories because holding the mutex prevents
1904 * new operations on the dir while we write to disk.
1905 */
1906int btrfs_sync_file(struct file *file, loff_t start, loff_t end, int datasync)
1907{
1908	struct dentry *dentry = file_dentry(file);
1909	struct inode *inode = d_inode(dentry);
 
1910	struct btrfs_root *root = BTRFS_I(inode)->root;
1911	struct btrfs_trans_handle *trans;
1912	struct btrfs_log_ctx ctx;
1913	int ret = 0;
1914	bool full_sync = 0;
1915	u64 len;
1916
1917	/*
1918	 * The range length can be represented by u64, we have to do the typecasts
1919	 * to avoid signed overflow if it's [0, LLONG_MAX] eg. from fsync()
1920	 */
1921	len = (u64)end - (u64)start + 1;
1922	trace_btrfs_sync_file(file, datasync);
1923
 
 
1924	/*
1925	 * We write the dirty pages in the range and wait until they complete
1926	 * out of the ->i_mutex. If so, we can flush the dirty pages by
1927	 * multi-task, and make the performance up.  See
1928	 * btrfs_wait_ordered_range for an explanation of the ASYNC check.
1929	 */
1930	ret = start_ordered_ops(inode, start, end);
1931	if (ret)
1932		return ret;
1933
1934	inode_lock(inode);
 
 
 
 
 
 
 
 
1935	atomic_inc(&root->log_batch);
1936	full_sync = test_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
1937			     &BTRFS_I(inode)->runtime_flags);
1938	/*
1939	 * We might have have had more pages made dirty after calling
1940	 * start_ordered_ops and before acquiring the inode's i_mutex.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1941	 */
1942	if (full_sync) {
1943		/*
1944		 * For a full sync, we need to make sure any ordered operations
1945		 * start and finish before we start logging the inode, so that
1946		 * all extents are persisted and the respective file extent
1947		 * items are in the fs/subvol btree.
1948		 */
1949		ret = btrfs_wait_ordered_range(inode, start, len);
1950	} else {
1951		/*
1952		 * Start any new ordered operations before starting to log the
1953		 * inode. We will wait for them to finish in btrfs_sync_log().
1954		 *
1955		 * Right before acquiring the inode's mutex, we might have new
1956		 * writes dirtying pages, which won't immediately start the
1957		 * respective ordered operations - that is done through the
1958		 * fill_delalloc callbacks invoked from the writepage and
1959		 * writepages address space operations. So make sure we start
1960		 * all ordered operations before starting to log our inode. Not
1961		 * doing this means that while logging the inode, writeback
1962		 * could start and invoke writepage/writepages, which would call
1963		 * the fill_delalloc callbacks (cow_file_range,
1964		 * submit_compressed_extents). These callbacks add first an
1965		 * extent map to the modified list of extents and then create
1966		 * the respective ordered operation, which means in
1967		 * tree-log.c:btrfs_log_inode() we might capture all existing
1968		 * ordered operations (with btrfs_get_logged_extents()) before
1969		 * the fill_delalloc callback adds its ordered operation, and by
1970		 * the time we visit the modified list of extent maps (with
1971		 * btrfs_log_changed_extents()), we see and process the extent
1972		 * map they created. We then use the extent map to construct a
1973		 * file extent item for logging without waiting for the
1974		 * respective ordered operation to finish - this file extent
1975		 * item points to a disk location that might not have yet been
1976		 * written to, containing random data - so after a crash a log
1977		 * replay will make our inode have file extent items that point
1978		 * to disk locations containing invalid data, as we returned
1979		 * success to userspace without waiting for the respective
1980		 * ordered operation to finish, because it wasn't captured by
1981		 * btrfs_get_logged_extents().
1982		 */
1983		ret = start_ordered_ops(inode, start, end);
1984	}
1985	if (ret) {
1986		inode_unlock(inode);
1987		goto out;
1988	}
1989	atomic_inc(&root->log_batch);
1990
1991	/*
1992	 * If the last transaction that changed this file was before the current
1993	 * transaction and we have the full sync flag set in our inode, we can
1994	 * bail out now without any syncing.
1995	 *
1996	 * Note that we can't bail out if the full sync flag isn't set. This is
1997	 * because when the full sync flag is set we start all ordered extents
1998	 * and wait for them to fully complete - when they complete they update
1999	 * the inode's last_trans field through:
2000	 *
2001	 *     btrfs_finish_ordered_io() ->
2002	 *         btrfs_update_inode_fallback() ->
2003	 *             btrfs_update_inode() ->
2004	 *                 btrfs_set_inode_last_trans()
2005	 *
2006	 * So we are sure that last_trans is up to date and can do this check to
2007	 * bail out safely. For the fast path, when the full sync flag is not
2008	 * set in our inode, we can not do it because we start only our ordered
2009	 * extents and don't wait for them to complete (that is when
2010	 * btrfs_finish_ordered_io runs), so here at this point their last_trans
2011	 * value might be less than or equals to fs_info->last_trans_committed,
2012	 * and setting a speculative last_trans for an inode when a buffered
2013	 * write is made (such as fs_info->generation + 1 for example) would not
2014	 * be reliable since after setting the value and before fsync is called
2015	 * any number of transactions can start and commit (transaction kthread
2016	 * commits the current transaction periodically), and a transaction
2017	 * commit does not start nor waits for ordered extents to complete.
2018	 */
 
 
 
 
 
 
 
 
2019	smp_mb();
2020	if (btrfs_inode_in_log(inode, root->fs_info->generation) ||
2021	    (full_sync && BTRFS_I(inode)->last_trans <=
2022	     root->fs_info->last_trans_committed) ||
2023	    (!btrfs_have_ordered_extents_in_range(inode, start, len) &&
2024	     BTRFS_I(inode)->last_trans
2025	     <= root->fs_info->last_trans_committed)) {
2026		/*
2027		 * We'v had everything committed since the last time we were
2028		 * modified so clear this flag in case it was set for whatever
2029		 * reason, it's no longer relevant.
2030		 */
2031		clear_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2032			  &BTRFS_I(inode)->runtime_flags);
 
 
 
 
 
 
 
 
 
2033		inode_unlock(inode);
2034		goto out;
2035	}
2036
2037	/*
2038	 * ok we haven't committed the transaction yet, lets do a commit
2039	 */
2040	if (file->private_data)
2041		btrfs_ioctl_trans_end(file);
2042
2043	/*
2044	 * We use start here because we will need to wait on the IO to complete
2045	 * in btrfs_sync_log, which could require joining a transaction (for
2046	 * example checking cross references in the nocow path).  If we use join
2047	 * here we could get into a situation where we're waiting on IO to
2048	 * happen that is blocked on a transaction trying to commit.  With start
2049	 * we inc the extwriter counter, so we wait for all extwriters to exit
2050	 * before we start blocking join'ers.  This comment is to keep somebody
2051	 * from thinking they are super smart and changing this to
2052	 * btrfs_join_transaction *cough*Josef*cough*.
2053	 */
2054	trans = btrfs_start_transaction(root, 0);
2055	if (IS_ERR(trans)) {
2056		ret = PTR_ERR(trans);
 
2057		inode_unlock(inode);
2058		goto out;
2059	}
2060	trans->sync = true;
2061
2062	btrfs_init_log_ctx(&ctx);
2063
2064	ret = btrfs_log_dentry_safe(trans, root, dentry, start, end, &ctx);
2065	if (ret < 0) {
2066		/* Fallthrough and commit/free transaction. */
2067		ret = 1;
2068	}
2069
2070	/* we've logged all the items and now have a consistent
2071	 * version of the file in the log.  It is possible that
2072	 * someone will come in and modify the file, but that's
2073	 * fine because the log is consistent on disk, and we
2074	 * have references to all of the file's extents
2075	 *
2076	 * It is possible that someone will come in and log the
2077	 * file again, but that will end up using the synchronization
2078	 * inside btrfs_sync_log to keep things safe.
2079	 */
 
2080	inode_unlock(inode);
2081
2082	/*
2083	 * If any of the ordered extents had an error, just return it to user
2084	 * space, so that the application knows some writes didn't succeed and
2085	 * can take proper action (retry for e.g.). Blindly committing the
2086	 * transaction in this case, would fool userspace that everything was
2087	 * successful. And we also want to make sure our log doesn't contain
2088	 * file extent items pointing to extents that weren't fully written to -
2089	 * just like in the non fast fsync path, where we check for the ordered
2090	 * operation's error flag before writing to the log tree and return -EIO
2091	 * if any of them had this flag set (btrfs_wait_ordered_range) -
2092	 * therefore we need to check for errors in the ordered operations,
2093	 * which are indicated by ctx.io_err.
2094	 */
2095	if (ctx.io_err) {
2096		btrfs_end_transaction(trans, root);
2097		ret = ctx.io_err;
2098		goto out;
2099	}
2100
2101	if (ret != BTRFS_NO_LOG_SYNC) {
2102		if (!ret) {
2103			ret = btrfs_sync_log(trans, root, &ctx);
2104			if (!ret) {
2105				ret = btrfs_end_transaction(trans, root);
2106				goto out;
2107			}
2108		}
2109		if (!full_sync) {
2110			ret = btrfs_wait_ordered_range(inode, start, len);
2111			if (ret) {
2112				btrfs_end_transaction(trans, root);
2113				goto out;
2114			}
2115		}
2116		ret = btrfs_commit_transaction(trans, root);
2117	} else {
2118		ret = btrfs_end_transaction(trans, root);
2119	}
2120out:
 
 
 
 
2121	return ret > 0 ? -EIO : ret;
2122}
2123
2124static const struct vm_operations_struct btrfs_file_vm_ops = {
2125	.fault		= filemap_fault,
2126	.map_pages	= filemap_map_pages,
2127	.page_mkwrite	= btrfs_page_mkwrite,
2128};
2129
2130static int btrfs_file_mmap(struct file	*filp, struct vm_area_struct *vma)
2131{
2132	struct address_space *mapping = filp->f_mapping;
2133
2134	if (!mapping->a_ops->readpage)
2135		return -ENOEXEC;
2136
2137	file_accessed(filp);
2138	vma->vm_ops = &btrfs_file_vm_ops;
2139
2140	return 0;
2141}
2142
2143static int hole_mergeable(struct inode *inode, struct extent_buffer *leaf,
2144			  int slot, u64 start, u64 end)
2145{
2146	struct btrfs_file_extent_item *fi;
2147	struct btrfs_key key;
2148
2149	if (slot < 0 || slot >= btrfs_header_nritems(leaf))
2150		return 0;
2151
2152	btrfs_item_key_to_cpu(leaf, &key, slot);
2153	if (key.objectid != btrfs_ino(inode) ||
2154	    key.type != BTRFS_EXTENT_DATA_KEY)
2155		return 0;
2156
2157	fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
2158
2159	if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG)
2160		return 0;
2161
2162	if (btrfs_file_extent_disk_bytenr(leaf, fi))
2163		return 0;
2164
2165	if (key.offset == end)
2166		return 1;
2167	if (key.offset + btrfs_file_extent_num_bytes(leaf, fi) == start)
2168		return 1;
2169	return 0;
2170}
2171
2172static int fill_holes(struct btrfs_trans_handle *trans, struct inode *inode,
2173		      struct btrfs_path *path, u64 offset, u64 end)
 
2174{
2175	struct btrfs_root *root = BTRFS_I(inode)->root;
 
2176	struct extent_buffer *leaf;
2177	struct btrfs_file_extent_item *fi;
2178	struct extent_map *hole_em;
2179	struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
2180	struct btrfs_key key;
2181	int ret;
2182
2183	if (btrfs_fs_incompat(root->fs_info, NO_HOLES))
2184		goto out;
2185
2186	key.objectid = btrfs_ino(inode);
2187	key.type = BTRFS_EXTENT_DATA_KEY;
2188	key.offset = offset;
2189
2190	ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
2191	if (ret < 0)
 
 
 
 
 
 
2192		return ret;
2193	BUG_ON(!ret);
2194
2195	leaf = path->nodes[0];
2196	if (hole_mergeable(inode, leaf, path->slots[0]-1, offset, end)) {
2197		u64 num_bytes;
2198
2199		path->slots[0]--;
2200		fi = btrfs_item_ptr(leaf, path->slots[0],
2201				    struct btrfs_file_extent_item);
2202		num_bytes = btrfs_file_extent_num_bytes(leaf, fi) +
2203			end - offset;
2204		btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2205		btrfs_set_file_extent_ram_bytes(leaf, fi, num_bytes);
2206		btrfs_set_file_extent_offset(leaf, fi, 0);
2207		btrfs_mark_buffer_dirty(leaf);
2208		goto out;
2209	}
2210
2211	if (hole_mergeable(inode, leaf, path->slots[0], offset, end)) {
2212		u64 num_bytes;
2213
2214		key.offset = offset;
2215		btrfs_set_item_key_safe(root->fs_info, path, &key);
2216		fi = btrfs_item_ptr(leaf, path->slots[0],
2217				    struct btrfs_file_extent_item);
2218		num_bytes = btrfs_file_extent_num_bytes(leaf, fi) + end -
2219			offset;
2220		btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2221		btrfs_set_file_extent_ram_bytes(leaf, fi, num_bytes);
2222		btrfs_set_file_extent_offset(leaf, fi, 0);
2223		btrfs_mark_buffer_dirty(leaf);
2224		goto out;
2225	}
2226	btrfs_release_path(path);
2227
2228	ret = btrfs_insert_file_extent(trans, root, btrfs_ino(inode), offset,
2229				       0, 0, end - offset, 0, end - offset,
2230				       0, 0, 0);
2231	if (ret)
2232		return ret;
2233
2234out:
2235	btrfs_release_path(path);
2236
2237	hole_em = alloc_extent_map();
2238	if (!hole_em) {
2239		btrfs_drop_extent_cache(inode, offset, end - 1, 0);
2240		set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2241			&BTRFS_I(inode)->runtime_flags);
2242	} else {
2243		hole_em->start = offset;
2244		hole_em->len = end - offset;
2245		hole_em->ram_bytes = hole_em->len;
2246		hole_em->orig_start = offset;
2247
2248		hole_em->block_start = EXTENT_MAP_HOLE;
2249		hole_em->block_len = 0;
2250		hole_em->orig_block_len = 0;
2251		hole_em->bdev = root->fs_info->fs_devices->latest_bdev;
2252		hole_em->compress_type = BTRFS_COMPRESS_NONE;
2253		hole_em->generation = trans->transid;
2254
2255		do {
2256			btrfs_drop_extent_cache(inode, offset, end - 1, 0);
2257			write_lock(&em_tree->lock);
2258			ret = add_extent_mapping(em_tree, hole_em, 1);
2259			write_unlock(&em_tree->lock);
2260		} while (ret == -EEXIST);
2261		free_extent_map(hole_em);
2262		if (ret)
2263			set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2264				&BTRFS_I(inode)->runtime_flags);
2265	}
2266
2267	return 0;
2268}
2269
2270/*
2271 * Find a hole extent on given inode and change start/len to the end of hole
2272 * extent.(hole/vacuum extent whose em->start <= start &&
2273 *	   em->start + em->len > start)
2274 * When a hole extent is found, return 1 and modify start/len.
2275 */
2276static int find_first_non_hole(struct inode *inode, u64 *start, u64 *len)
2277{
 
2278	struct extent_map *em;
2279	int ret = 0;
2280
2281	em = btrfs_get_extent(inode, NULL, 0, *start, *len, 0);
2282	if (IS_ERR_OR_NULL(em)) {
2283		if (!em)
2284			ret = -ENOMEM;
2285		else
2286			ret = PTR_ERR(em);
2287		return ret;
2288	}
2289
2290	/* Hole or vacuum extent(only exists in no-hole mode) */
2291	if (em->block_start == EXTENT_MAP_HOLE) {
2292		ret = 1;
2293		*len = em->start + em->len > *start + *len ?
2294		       0 : *start + *len - em->start - em->len;
2295		*start = em->start + em->len;
2296	}
2297	free_extent_map(em);
2298	return ret;
2299}
2300
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2301static int btrfs_punch_hole(struct inode *inode, loff_t offset, loff_t len)
2302{
 
2303	struct btrfs_root *root = BTRFS_I(inode)->root;
2304	struct extent_state *cached_state = NULL;
2305	struct btrfs_path *path;
2306	struct btrfs_block_rsv *rsv;
2307	struct btrfs_trans_handle *trans;
2308	u64 lockstart;
2309	u64 lockend;
2310	u64 tail_start;
2311	u64 tail_len;
2312	u64 orig_start = offset;
2313	u64 cur_offset;
2314	u64 min_size = btrfs_calc_trunc_metadata_size(root, 1);
2315	u64 drop_end;
2316	int ret = 0;
2317	int err = 0;
2318	unsigned int rsv_count;
2319	bool same_block;
2320	bool no_holes = btrfs_fs_incompat(root->fs_info, NO_HOLES);
2321	u64 ino_size;
2322	bool truncated_block = false;
2323	bool updated_inode = false;
2324
2325	ret = btrfs_wait_ordered_range(inode, offset, len);
2326	if (ret)
2327		return ret;
2328
2329	inode_lock(inode);
2330	ino_size = round_up(inode->i_size, root->sectorsize);
2331	ret = find_first_non_hole(inode, &offset, &len);
2332	if (ret < 0)
2333		goto out_only_mutex;
2334	if (ret && !len) {
2335		/* Already in a large hole */
2336		ret = 0;
2337		goto out_only_mutex;
2338	}
2339
2340	lockstart = round_up(offset, BTRFS_I(inode)->root->sectorsize);
2341	lockend = round_down(offset + len,
2342			     BTRFS_I(inode)->root->sectorsize) - 1;
2343	same_block = (BTRFS_BYTES_TO_BLKS(root->fs_info, offset))
2344		== (BTRFS_BYTES_TO_BLKS(root->fs_info, offset + len - 1));
2345	/*
2346	 * We needn't truncate any block which is beyond the end of the file
2347	 * because we are sure there is no data there.
2348	 */
2349	/*
2350	 * Only do this if we are in the same block and we aren't doing the
2351	 * entire block.
2352	 */
2353	if (same_block && len < root->sectorsize) {
2354		if (offset < ino_size) {
2355			truncated_block = true;
2356			ret = btrfs_truncate_block(inode, offset, len, 0);
2357		} else {
2358			ret = 0;
2359		}
2360		goto out_only_mutex;
2361	}
2362
2363	/* zero back part of the first block */
2364	if (offset < ino_size) {
2365		truncated_block = true;
2366		ret = btrfs_truncate_block(inode, offset, 0, 0);
2367		if (ret) {
2368			inode_unlock(inode);
2369			return ret;
2370		}
2371	}
2372
2373	/* Check the aligned pages after the first unaligned page,
2374	 * if offset != orig_start, which means the first unaligned page
2375	 * including serveral following pages are already in holes,
2376	 * the extra check can be skipped */
2377	if (offset == orig_start) {
2378		/* after truncate page, check hole again */
2379		len = offset + len - lockstart;
2380		offset = lockstart;
2381		ret = find_first_non_hole(inode, &offset, &len);
2382		if (ret < 0)
2383			goto out_only_mutex;
2384		if (ret && !len) {
2385			ret = 0;
2386			goto out_only_mutex;
2387		}
2388		lockstart = offset;
2389	}
2390
2391	/* Check the tail unaligned part is in a hole */
2392	tail_start = lockend + 1;
2393	tail_len = offset + len - tail_start;
2394	if (tail_len) {
2395		ret = find_first_non_hole(inode, &tail_start, &tail_len);
2396		if (unlikely(ret < 0))
2397			goto out_only_mutex;
2398		if (!ret) {
2399			/* zero the front end of the last page */
2400			if (tail_start + tail_len < ino_size) {
2401				truncated_block = true;
2402				ret = btrfs_truncate_block(inode,
2403							tail_start + tail_len,
2404							0, 1);
2405				if (ret)
2406					goto out_only_mutex;
2407			}
2408		}
2409	}
2410
2411	if (lockend < lockstart) {
2412		ret = 0;
2413		goto out_only_mutex;
2414	}
2415
2416	while (1) {
2417		struct btrfs_ordered_extent *ordered;
2418
2419		truncate_pagecache_range(inode, lockstart, lockend);
2420
2421		lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2422				 &cached_state);
2423		ordered = btrfs_lookup_first_ordered_extent(inode, lockend);
2424
2425		/*
2426		 * We need to make sure we have no ordered extents in this range
2427		 * and nobody raced in and read a page in this range, if we did
2428		 * we need to try again.
2429		 */
2430		if ((!ordered ||
2431		    (ordered->file_offset + ordered->len <= lockstart ||
2432		     ordered->file_offset > lockend)) &&
2433		     !btrfs_page_exists_in_range(inode, lockstart, lockend)) {
2434			if (ordered)
2435				btrfs_put_ordered_extent(ordered);
2436			break;
2437		}
2438		if (ordered)
2439			btrfs_put_ordered_extent(ordered);
2440		unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart,
2441				     lockend, &cached_state, GFP_NOFS);
2442		ret = btrfs_wait_ordered_range(inode, lockstart,
2443					       lockend - lockstart + 1);
2444		if (ret) {
2445			inode_unlock(inode);
2446			return ret;
2447		}
2448	}
2449
2450	path = btrfs_alloc_path();
2451	if (!path) {
2452		ret = -ENOMEM;
2453		goto out;
2454	}
2455
2456	rsv = btrfs_alloc_block_rsv(root, BTRFS_BLOCK_RSV_TEMP);
2457	if (!rsv) {
2458		ret = -ENOMEM;
2459		goto out_free;
2460	}
2461	rsv->size = btrfs_calc_trunc_metadata_size(root, 1);
2462	rsv->failfast = 1;
2463
2464	/*
2465	 * 1 - update the inode
2466	 * 1 - removing the extents in the range
2467	 * 1 - adding the hole extent if no_holes isn't set
2468	 */
2469	rsv_count = no_holes ? 2 : 3;
2470	trans = btrfs_start_transaction(root, rsv_count);
2471	if (IS_ERR(trans)) {
2472		err = PTR_ERR(trans);
2473		goto out_free;
2474	}
2475
2476	ret = btrfs_block_rsv_migrate(&root->fs_info->trans_block_rsv, rsv,
2477				      min_size);
2478	BUG_ON(ret);
2479	trans->block_rsv = rsv;
2480
2481	cur_offset = lockstart;
2482	len = lockend - cur_offset;
2483	while (cur_offset < lockend) {
2484		ret = __btrfs_drop_extents(trans, root, inode, path,
2485					   cur_offset, lockend + 1,
2486					   &drop_end, 1, 0, 0, NULL);
2487		if (ret != -ENOSPC)
2488			break;
2489
2490		trans->block_rsv = &root->fs_info->trans_block_rsv;
2491
2492		if (cur_offset < ino_size) {
2493			ret = fill_holes(trans, inode, path, cur_offset,
2494					 drop_end);
2495			if (ret) {
2496				err = ret;
2497				break;
2498			}
2499		}
2500
2501		cur_offset = drop_end;
2502
2503		ret = btrfs_update_inode(trans, root, inode);
2504		if (ret) {
2505			err = ret;
2506			break;
2507		}
2508
2509		btrfs_end_transaction(trans, root);
2510		btrfs_btree_balance_dirty(root);
2511
2512		trans = btrfs_start_transaction(root, rsv_count);
2513		if (IS_ERR(trans)) {
2514			ret = PTR_ERR(trans);
2515			trans = NULL;
2516			break;
2517		}
2518
2519		ret = btrfs_block_rsv_migrate(&root->fs_info->trans_block_rsv,
2520					      rsv, min_size);
2521		BUG_ON(ret);	/* shouldn't happen */
2522		trans->block_rsv = rsv;
2523
2524		ret = find_first_non_hole(inode, &cur_offset, &len);
2525		if (unlikely(ret < 0))
2526			break;
2527		if (ret && !len) {
2528			ret = 0;
2529			break;
2530		}
2531	}
2532
2533	if (ret) {
2534		err = ret;
2535		goto out_trans;
2536	}
2537
2538	trans->block_rsv = &root->fs_info->trans_block_rsv;
2539	/*
2540	 * If we are using the NO_HOLES feature we might have had already an
2541	 * hole that overlaps a part of the region [lockstart, lockend] and
2542	 * ends at (or beyond) lockend. Since we have no file extent items to
2543	 * represent holes, drop_end can be less than lockend and so we must
2544	 * make sure we have an extent map representing the existing hole (the
2545	 * call to __btrfs_drop_extents() might have dropped the existing extent
2546	 * map representing the existing hole), otherwise the fast fsync path
2547	 * will not record the existence of the hole region
2548	 * [existing_hole_start, lockend].
2549	 */
2550	if (drop_end <= lockend)
2551		drop_end = lockend + 1;
2552	/*
2553	 * Don't insert file hole extent item if it's for a range beyond eof
2554	 * (because it's useless) or if it represents a 0 bytes range (when
2555	 * cur_offset == drop_end).
2556	 */
2557	if (cur_offset < ino_size && cur_offset < drop_end) {
2558		ret = fill_holes(trans, inode, path, cur_offset, drop_end);
2559		if (ret) {
2560			err = ret;
2561			goto out_trans;
2562		}
2563	}
2564
2565out_trans:
2566	if (!trans)
2567		goto out_free;
2568
 
2569	inode_inc_iversion(inode);
2570	inode->i_mtime = inode->i_ctime = current_fs_time(inode->i_sb);
2571
2572	trans->block_rsv = &root->fs_info->trans_block_rsv;
2573	ret = btrfs_update_inode(trans, root, inode);
2574	updated_inode = true;
2575	btrfs_end_transaction(trans, root);
2576	btrfs_btree_balance_dirty(root);
2577out_free:
2578	btrfs_free_path(path);
2579	btrfs_free_block_rsv(root, rsv);
2580out:
2581	unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2582			     &cached_state, GFP_NOFS);
2583out_only_mutex:
2584	if (!updated_inode && truncated_block && !ret && !err) {
2585		/*
2586		 * If we only end up zeroing part of a page, we still need to
2587		 * update the inode item, so that all the time fields are
2588		 * updated as well as the necessary btrfs inode in memory fields
2589		 * for detecting, at fsync time, if the inode isn't yet in the
2590		 * log tree or it's there but not up to date.
2591		 */
 
 
 
 
 
2592		trans = btrfs_start_transaction(root, 1);
2593		if (IS_ERR(trans)) {
2594			err = PTR_ERR(trans);
2595		} else {
2596			err = btrfs_update_inode(trans, root, inode);
2597			ret = btrfs_end_transaction(trans, root);
 
 
 
 
2598		}
2599	}
2600	inode_unlock(inode);
2601	if (ret && !err)
2602		err = ret;
2603	return err;
2604}
2605
2606/* Helper structure to record which range is already reserved */
2607struct falloc_range {
2608	struct list_head list;
2609	u64 start;
2610	u64 len;
2611};
2612
2613/*
2614 * Helper function to add falloc range
2615 *
2616 * Caller should have locked the larger range of extent containing
2617 * [start, len)
2618 */
2619static int add_falloc_range(struct list_head *head, u64 start, u64 len)
2620{
2621	struct falloc_range *prev = NULL;
2622	struct falloc_range *range = NULL;
2623
2624	if (list_empty(head))
2625		goto insert;
2626
2627	/*
2628	 * As fallocate iterate by bytenr order, we only need to check
2629	 * the last range.
2630	 */
2631	prev = list_entry(head->prev, struct falloc_range, list);
2632	if (prev->start + prev->len == start) {
2633		prev->len += len;
2634		return 0;
2635	}
2636insert:
2637	range = kmalloc(sizeof(*range), GFP_KERNEL);
2638	if (!range)
2639		return -ENOMEM;
2640	range->start = start;
2641	range->len = len;
2642	list_add_tail(&range->list, head);
2643	return 0;
2644}
2645
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2646static long btrfs_fallocate(struct file *file, int mode,
2647			    loff_t offset, loff_t len)
2648{
2649	struct inode *inode = file_inode(file);
2650	struct extent_state *cached_state = NULL;
 
2651	struct falloc_range *range;
2652	struct falloc_range *tmp;
2653	struct list_head reserve_list;
2654	u64 cur_offset;
2655	u64 last_byte;
2656	u64 alloc_start;
2657	u64 alloc_end;
2658	u64 alloc_hint = 0;
2659	u64 locked_end;
2660	u64 actual_end = 0;
2661	struct extent_map *em;
2662	int blocksize = BTRFS_I(inode)->root->sectorsize;
2663	int ret;
2664
2665	alloc_start = round_down(offset, blocksize);
2666	alloc_end = round_up(offset + len, blocksize);
 
2667
2668	/* Make sure we aren't being give some crap mode */
2669	if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
 
2670		return -EOPNOTSUPP;
2671
2672	if (mode & FALLOC_FL_PUNCH_HOLE)
2673		return btrfs_punch_hole(inode, offset, len);
2674
2675	/*
2676	 * Only trigger disk allocation, don't trigger qgroup reserve
2677	 *
2678	 * For qgroup space, it will be checked later.
2679	 */
2680	ret = btrfs_alloc_data_chunk_ondemand(inode, alloc_end - alloc_start);
2681	if (ret < 0)
2682		return ret;
 
 
 
2683
2684	inode_lock(inode);
2685
2686	if (!(mode & FALLOC_FL_KEEP_SIZE) && offset + len > inode->i_size) {
2687		ret = inode_newsize_ok(inode, offset + len);
2688		if (ret)
2689			goto out;
2690	}
2691
2692	/*
2693	 * TODO: Move these two operations after we have checked
2694	 * accurate reserved space, or fallocate can still fail but
2695	 * with page truncated or size expanded.
2696	 *
2697	 * But that's a minor problem and won't do much harm BTW.
2698	 */
2699	if (alloc_start > inode->i_size) {
2700		ret = btrfs_cont_expand(inode, i_size_read(inode),
2701					alloc_start);
2702		if (ret)
2703			goto out;
2704	} else if (offset + len > inode->i_size) {
2705		/*
2706		 * If we are fallocating from the end of the file onward we
2707		 * need to zero out the end of the block if i_size lands in the
2708		 * middle of a block.
2709		 */
2710		ret = btrfs_truncate_block(inode, inode->i_size, 0, 0);
2711		if (ret)
2712			goto out;
2713	}
2714
2715	/*
2716	 * wait for ordered IO before we have any locks.  We'll loop again
2717	 * below with the locks held.
2718	 */
2719	ret = btrfs_wait_ordered_range(inode, alloc_start,
2720				       alloc_end - alloc_start);
2721	if (ret)
2722		goto out;
2723
 
 
 
 
 
 
2724	locked_end = alloc_end - 1;
2725	while (1) {
2726		struct btrfs_ordered_extent *ordered;
2727
2728		/* the extent lock is ordered inside the running
2729		 * transaction
2730		 */
2731		lock_extent_bits(&BTRFS_I(inode)->io_tree, alloc_start,
2732				 locked_end, &cached_state);
2733		ordered = btrfs_lookup_first_ordered_extent(inode,
2734							    alloc_end - 1);
2735		if (ordered &&
2736		    ordered->file_offset + ordered->len > alloc_start &&
2737		    ordered->file_offset < alloc_end) {
2738			btrfs_put_ordered_extent(ordered);
2739			unlock_extent_cached(&BTRFS_I(inode)->io_tree,
2740					     alloc_start, locked_end,
2741					     &cached_state, GFP_KERNEL);
2742			/*
2743			 * we can't wait on the range with the transaction
2744			 * running or with the extent lock held
2745			 */
2746			ret = btrfs_wait_ordered_range(inode, alloc_start,
2747						       alloc_end - alloc_start);
2748			if (ret)
2749				goto out;
2750		} else {
2751			if (ordered)
2752				btrfs_put_ordered_extent(ordered);
2753			break;
2754		}
2755	}
2756
2757	/* First, check if we exceed the qgroup limit */
2758	INIT_LIST_HEAD(&reserve_list);
2759	cur_offset = alloc_start;
2760	while (1) {
2761		em = btrfs_get_extent(inode, NULL, 0, cur_offset,
2762				      alloc_end - cur_offset, 0);
2763		if (IS_ERR_OR_NULL(em)) {
2764			if (!em)
2765				ret = -ENOMEM;
2766			else
2767				ret = PTR_ERR(em);
2768			break;
2769		}
2770		last_byte = min(extent_map_end(em), alloc_end);
2771		actual_end = min_t(u64, extent_map_end(em), offset + len);
2772		last_byte = ALIGN(last_byte, blocksize);
2773		if (em->block_start == EXTENT_MAP_HOLE ||
2774		    (cur_offset >= inode->i_size &&
2775		     !test_bit(EXTENT_FLAG_PREALLOC, &em->flags))) {
2776			ret = add_falloc_range(&reserve_list, cur_offset,
2777					       last_byte - cur_offset);
2778			if (ret < 0) {
2779				free_extent_map(em);
2780				break;
2781			}
2782			ret = btrfs_qgroup_reserve_data(inode, cur_offset,
2783					last_byte - cur_offset);
2784			if (ret < 0)
 
 
2785				break;
 
 
 
 
 
 
 
 
 
2786		}
2787		free_extent_map(em);
2788		cur_offset = last_byte;
2789		if (cur_offset >= alloc_end)
2790			break;
2791	}
2792
2793	/*
2794	 * If ret is still 0, means we're OK to fallocate.
2795	 * Or just cleanup the list and exit.
2796	 */
2797	list_for_each_entry_safe(range, tmp, &reserve_list, list) {
2798		if (!ret)
2799			ret = btrfs_prealloc_file_range(inode, mode,
2800					range->start,
2801					range->len, 1 << inode->i_blkbits,
2802					offset + len, &alloc_hint);
 
 
 
 
2803		list_del(&range->list);
2804		kfree(range);
2805	}
2806	if (ret < 0)
2807		goto out_unlock;
2808
2809	if (actual_end > inode->i_size &&
2810	    !(mode & FALLOC_FL_KEEP_SIZE)) {
2811		struct btrfs_trans_handle *trans;
2812		struct btrfs_root *root = BTRFS_I(inode)->root;
2813
2814		/*
2815		 * We didn't need to allocate any more space, but we
2816		 * still extended the size of the file so we need to
2817		 * update i_size and the inode item.
2818		 */
2819		trans = btrfs_start_transaction(root, 1);
2820		if (IS_ERR(trans)) {
2821			ret = PTR_ERR(trans);
2822		} else {
2823			inode->i_ctime = current_fs_time(inode->i_sb);
2824			i_size_write(inode, actual_end);
2825			btrfs_ordered_update_i_size(inode, actual_end, NULL);
2826			ret = btrfs_update_inode(trans, root, inode);
2827			if (ret)
2828				btrfs_end_transaction(trans, root);
2829			else
2830				ret = btrfs_end_transaction(trans, root);
2831		}
2832	}
2833out_unlock:
2834	unlock_extent_cached(&BTRFS_I(inode)->io_tree, alloc_start, locked_end,
2835			     &cached_state, GFP_KERNEL);
2836out:
2837	/*
2838	 * As we waited the extent range, the data_rsv_map must be empty
2839	 * in the range, as written data range will be released from it.
2840	 * And for prealloacted extent, it will also be released when
2841	 * its metadata is written.
2842	 * So this is completely used as cleanup.
2843	 */
2844	btrfs_qgroup_free_data(inode, alloc_start, alloc_end - alloc_start);
2845	inode_unlock(inode);
2846	/* Let go of our reservation. */
2847	btrfs_free_reserved_data_space(inode, alloc_start,
2848				       alloc_end - alloc_start);
 
 
2849	return ret;
2850}
2851
2852static int find_desired_extent(struct inode *inode, loff_t *offset, int whence)
2853{
2854	struct btrfs_root *root = BTRFS_I(inode)->root;
2855	struct extent_map *em = NULL;
2856	struct extent_state *cached_state = NULL;
2857	u64 lockstart;
2858	u64 lockend;
2859	u64 start;
2860	u64 len;
2861	int ret = 0;
2862
2863	if (inode->i_size == 0)
2864		return -ENXIO;
2865
2866	/*
2867	 * *offset can be negative, in this case we start finding DATA/HOLE from
2868	 * the very start of the file.
2869	 */
2870	start = max_t(loff_t, 0, *offset);
2871
2872	lockstart = round_down(start, root->sectorsize);
2873	lockend = round_up(i_size_read(inode), root->sectorsize);
 
2874	if (lockend <= lockstart)
2875		lockend = lockstart + root->sectorsize;
2876	lockend--;
2877	len = lockend - lockstart + 1;
2878
2879	lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2880			 &cached_state);
2881
2882	while (start < inode->i_size) {
2883		em = btrfs_get_extent_fiemap(inode, NULL, 0, start, len, 0);
2884		if (IS_ERR(em)) {
2885			ret = PTR_ERR(em);
2886			em = NULL;
2887			break;
2888		}
2889
2890		if (whence == SEEK_HOLE &&
2891		    (em->block_start == EXTENT_MAP_HOLE ||
2892		     test_bit(EXTENT_FLAG_PREALLOC, &em->flags)))
2893			break;
2894		else if (whence == SEEK_DATA &&
2895			   (em->block_start != EXTENT_MAP_HOLE &&
2896			    !test_bit(EXTENT_FLAG_PREALLOC, &em->flags)))
2897			break;
2898
2899		start = em->start + em->len;
2900		free_extent_map(em);
2901		em = NULL;
2902		cond_resched();
2903	}
2904	free_extent_map(em);
2905	if (!ret) {
2906		if (whence == SEEK_DATA && start >= inode->i_size)
2907			ret = -ENXIO;
2908		else
2909			*offset = min_t(loff_t, start, inode->i_size);
2910	}
2911	unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2912			     &cached_state, GFP_NOFS);
2913	return ret;
2914}
2915
2916static loff_t btrfs_file_llseek(struct file *file, loff_t offset, int whence)
2917{
2918	struct inode *inode = file->f_mapping->host;
2919	int ret;
2920
2921	inode_lock(inode);
2922	switch (whence) {
2923	case SEEK_END:
2924	case SEEK_CUR:
2925		offset = generic_file_llseek(file, offset, whence);
2926		goto out;
2927	case SEEK_DATA:
2928	case SEEK_HOLE:
2929		if (offset >= i_size_read(inode)) {
2930			inode_unlock(inode);
2931			return -ENXIO;
2932		}
2933
2934		ret = find_desired_extent(inode, &offset, whence);
2935		if (ret) {
2936			inode_unlock(inode);
2937			return ret;
2938		}
2939	}
2940
2941	offset = vfs_setpos(file, offset, inode->i_sb->s_maxbytes);
2942out:
2943	inode_unlock(inode);
2944	return offset;
2945}
2946
 
 
 
 
 
 
2947const struct file_operations btrfs_file_operations = {
2948	.llseek		= btrfs_file_llseek,
2949	.read_iter      = generic_file_read_iter,
2950	.splice_read	= generic_file_splice_read,
2951	.write_iter	= btrfs_file_write_iter,
2952	.mmap		= btrfs_file_mmap,
2953	.open		= generic_file_open,
2954	.release	= btrfs_release_file,
2955	.fsync		= btrfs_sync_file,
2956	.fallocate	= btrfs_fallocate,
2957	.unlocked_ioctl	= btrfs_ioctl,
2958#ifdef CONFIG_COMPAT
2959	.compat_ioctl	= btrfs_ioctl,
2960#endif
2961	.copy_file_range = btrfs_copy_file_range,
2962	.clone_file_range = btrfs_clone_file_range,
2963	.dedupe_file_range = btrfs_dedupe_file_range,
2964};
2965
2966void btrfs_auto_defrag_exit(void)
2967{
2968	kmem_cache_destroy(btrfs_inode_defrag_cachep);
2969}
2970
2971int btrfs_auto_defrag_init(void)
2972{
2973	btrfs_inode_defrag_cachep = kmem_cache_create("btrfs_inode_defrag",
2974					sizeof(struct inode_defrag), 0,
2975					SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD,
2976					NULL);
2977	if (!btrfs_inode_defrag_cachep)
2978		return -ENOMEM;
2979
2980	return 0;
2981}
2982
2983int btrfs_fdatawrite_range(struct inode *inode, loff_t start, loff_t end)
2984{
2985	int ret;
2986
2987	/*
2988	 * So with compression we will find and lock a dirty page and clear the
2989	 * first one as dirty, setup an async extent, and immediately return
2990	 * with the entire range locked but with nobody actually marked with
2991	 * writeback.  So we can't just filemap_write_and_wait_range() and
2992	 * expect it to work since it will just kick off a thread to do the
2993	 * actual work.  So we need to call filemap_fdatawrite_range _again_
2994	 * since it will wait on the page lock, which won't be unlocked until
2995	 * after the pages have been marked as writeback and so we're good to go
2996	 * from there.  We have to do this otherwise we'll miss the ordered
2997	 * extents and that results in badness.  Please Josef, do not think you
2998	 * know better and pull this out at some point in the future, it is
2999	 * right and you are wrong.
3000	 */
3001	ret = filemap_fdatawrite_range(inode->i_mapping, start, end);
3002	if (!ret && test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
3003			     &BTRFS_I(inode)->runtime_flags))
3004		ret = filemap_fdatawrite_range(inode->i_mapping, start, end);
3005
3006	return ret;
3007}