extents_status.c

   1// SPDX-License-Identifier: GPL-2.0
   2/*
   3 *  fs/ext4/extents_status.c
   4 *
   5 * Written by Yongqiang Yang <xiaoqiangnk@gmail.com>
   6 * Modified by
   7 *	Allison Henderson <achender@linux.vnet.ibm.com>
   8 *	Hugh Dickins <hughd@google.com>
   9 *	Zheng Liu <wenqing.lz@taobao.com>
  10 *
  11 * Ext4 extents status tree core functions.
  12 */
  13#include <linux/list_sort.h>
  14#include <linux/proc_fs.h>
  15#include <linux/seq_file.h>
  16#include "ext4.h"
  17
  18#include <trace/events/ext4.h>
  19
  20/*
  21 * According to previous discussion in Ext4 Developer Workshop, we
  22 * will introduce a new structure called io tree to track all extent
  23 * status in order to solve some problems that we have met
  24 * (e.g. Reservation space warning), and provide extent-level locking.
  25 * Delay extent tree is the first step to achieve this goal.  It is
  26 * original built by Yongqiang Yang.  At that time it is called delay
  27 * extent tree, whose goal is only track delayed extents in memory to
  28 * simplify the implementation of fiemap and bigalloc, and introduce
  29 * lseek SEEK_DATA/SEEK_HOLE support.  That is why it is still called
  30 * delay extent tree at the first commit.  But for better understand
  31 * what it does, it has been rename to extent status tree.
  32 *
  33 * Step1:
  34 * Currently the first step has been done.  All delayed extents are
  35 * tracked in the tree.  It maintains the delayed extent when a delayed
  36 * allocation is issued, and the delayed extent is written out or
  37 * invalidated.  Therefore the implementation of fiemap and bigalloc
  38 * are simplified, and SEEK_DATA/SEEK_HOLE are introduced.
  39 *
  40 * The following comment describes the implemenmtation of extent
  41 * status tree and future works.
  42 *
  43 * Step2:
  44 * In this step all extent status are tracked by extent status tree.
  45 * Thus, we can first try to lookup a block mapping in this tree before
  46 * finding it in extent tree.  Hence, single extent cache can be removed
  47 * because extent status tree can do a better job.  Extents in status
  48 * tree are loaded on-demand.  Therefore, the extent status tree may not
  49 * contain all of the extents in a file.  Meanwhile we define a shrinker
  50 * to reclaim memory from extent status tree because fragmented extent
  51 * tree will make status tree cost too much memory.  written/unwritten/-
  52 * hole extents in the tree will be reclaimed by this shrinker when we
  53 * are under high memory pressure.  Delayed extents will not be
  54 * reclimed because fiemap, bigalloc, and seek_data/hole need it.
  55 */
  56
  57/*
  58 * Extent status tree implementation for ext4.
  59 *
  60 *
  61 * ==========================================================================
  62 * Extent status tree tracks all extent status.
  63 *
  64 * 1. Why we need to implement extent status tree?
  65 *
  66 * Without extent status tree, ext4 identifies a delayed extent by looking
  67 * up page cache, this has several deficiencies - complicated, buggy,
  68 * and inefficient code.
  69 *
  70 * FIEMAP, SEEK_HOLE/DATA, bigalloc, and writeout all need to know if a
  71 * block or a range of blocks are belonged to a delayed extent.
  72 *
  73 * Let us have a look at how they do without extent status tree.
  74 *   --	FIEMAP
  75 *	FIEMAP looks up page cache to identify delayed allocations from holes.
  76 *
  77 *   --	SEEK_HOLE/DATA
  78 *	SEEK_HOLE/DATA has the same problem as FIEMAP.
  79 *
  80 *   --	bigalloc
  81 *	bigalloc looks up page cache to figure out if a block is
  82 *	already under delayed allocation or not to determine whether
  83 *	quota reserving is needed for the cluster.
  84 *
  85 *   --	writeout
  86 *	Writeout looks up whole page cache to see if a buffer is
  87 *	mapped, If there are not very many delayed buffers, then it is
  88 *	time consuming.
  89 *
  90 * With extent status tree implementation, FIEMAP, SEEK_HOLE/DATA,
  91 * bigalloc and writeout can figure out if a block or a range of
  92 * blocks is under delayed allocation(belonged to a delayed extent) or
  93 * not by searching the extent tree.
  94 *
  95 *
  96 * ==========================================================================
  97 * 2. Ext4 extent status tree impelmentation
  98 *
  99 *   --	extent
 100 *	A extent is a range of blocks which are contiguous logically and
 101 *	physically.  Unlike extent in extent tree, this extent in ext4 is
 102 *	a in-memory struct, there is no corresponding on-disk data.  There
 103 *	is no limit on length of extent, so an extent can contain as many
 104 *	blocks as they are contiguous logically and physically.
 105 *
 106 *   --	extent status tree
 107 *	Every inode has an extent status tree and all allocation blocks
 108 *	are added to the tree with different status.  The extent in the
 109 *	tree are ordered by logical block no.
 110 *
 111 *   --	operations on a extent status tree
 112 *	There are three important operations on a delayed extent tree: find
 113 *	next extent, adding a extent(a range of blocks) and removing a extent.
 114 *
 115 *   --	race on a extent status tree
 116 *	Extent status tree is protected by inode->i_es_lock.
 117 *
 118 *   --	memory consumption
 119 *      Fragmented extent tree will make extent status tree cost too much
 120 *      memory.  Hence, we will reclaim written/unwritten/hole extents from
 121 *      the tree under a heavy memory pressure.
 122 *
 123 * ==========================================================================
 124 * 3. Assurance of Ext4 extent status tree consistency
 125 *
 126 * When mapping blocks, Ext4 queries the extent status tree first and should
 127 * always trusts that the extent status tree is consistent and up to date.
 128 * Therefore, it is important to adheres to the following rules when createing,
 129 * modifying and removing extents.
 130 *
 131 *  1. Besides fastcommit replay, when Ext4 creates or queries block mappings,
 132 *     the extent information should always be processed through the extent
 133 *     status tree instead of being organized manually through the on-disk
 134 *     extent tree.
 135 *
 136 *  2. When updating the extent tree, Ext4 should acquire the i_data_sem
 137 *     exclusively and update the extent status tree atomically. If the extents
 138 *     to be modified are large enough to exceed the range that a single
 139 *     i_data_sem can process (as ext4_datasem_ensure_credits() may drop
 140 *     i_data_sem to restart a transaction), it must (e.g. as ext4_punch_hole()
 141 *     does):
 142 *
 143 *     a) Hold the i_rwsem and invalidate_lock exclusively. This ensures
 144 *        exclusion against page faults, as well as reads and writes that may
 145 *        concurrently modify the extent status tree.
 146 *     b) Evict all page cache in the affected range and recommend rebuilding
 147 *        or dropping the extent status tree after modifying the on-disk
 148 *        extent tree. This ensures exclusion against concurrent writebacks
 149 *        that do not hold those locks but only holds a folio lock.
 150 *
 151 *  3. Based on the rules above, when querying block mappings, Ext4 should at
 152 *     least hold the i_rwsem or invalidate_lock or folio lock(s) for the
 153 *     specified querying range.
 154 *
 155 * ==========================================================================
 156 * 4. Performance analysis
 157 *
 158 *   --	overhead
 159 *	1. There is a cache extent for write access, so if writes are
 160 *	not very random, adding space operaions are in O(1) time.
 161 *
 162 *   --	gain
 163 *	2. Code is much simpler, more readable, more maintainable and
 164 *	more efficient.
 165 *
 166 *
 167 * ==========================================================================
 168 * 5. TODO list
 169 *
 170 *   -- Refactor delayed space reservation
 171 *
 172 *   -- Extent-level locking
 173 */
 174
 175static struct kmem_cache *ext4_es_cachep;
 176static struct kmem_cache *ext4_pending_cachep;
 177
 178static int __es_insert_extent(struct inode *inode, struct extent_status *newes,
 179			      struct extent_status *prealloc);
 180static int __es_remove_extent(struct inode *inode, ext4_lblk_t lblk,
 181			      ext4_lblk_t end, int *reserved,
 182			      struct extent_status *prealloc);
 183static int es_reclaim_extents(struct ext4_inode_info *ei, int *nr_to_scan);
 184static int __es_shrink(struct ext4_sb_info *sbi, int nr_to_scan,
 185		       struct ext4_inode_info *locked_ei);
 186static int __revise_pending(struct inode *inode, ext4_lblk_t lblk,
 187			    ext4_lblk_t len,
 188			    struct pending_reservation **prealloc);
 189
 190int __init ext4_init_es(void)
 191{
 192	ext4_es_cachep = KMEM_CACHE(extent_status, SLAB_RECLAIM_ACCOUNT);
 193	if (ext4_es_cachep == NULL)
 194		return -ENOMEM;
 195	return 0;
 196}
 197
 198void ext4_exit_es(void)
 199{
 200	kmem_cache_destroy(ext4_es_cachep);
 201}
 202
 203void ext4_es_init_tree(struct ext4_es_tree *tree)
 204{
 205	tree->root = RB_ROOT;
 206	tree->cache_es = NULL;
 207}
 208
 209#ifdef ES_DEBUG__
 210static void ext4_es_print_tree(struct inode *inode)
 211{
 212	struct ext4_es_tree *tree;
 213	struct rb_node *node;
 214
 215	printk(KERN_DEBUG "status extents for inode %lu:", inode->i_ino);
 216	tree = &EXT4_I(inode)->i_es_tree;
 217	node = rb_first(&tree->root);
 218	while (node) {
 219		struct extent_status *es;
 220		es = rb_entry(node, struct extent_status, rb_node);
 221		printk(KERN_DEBUG " [%u/%u) %llu %x",
 222		       es->es_lblk, es->es_len,
 223		       ext4_es_pblock(es), ext4_es_status(es));
 224		node = rb_next(node);
 225	}
 226	printk(KERN_DEBUG "\n");
 227}
 228#else
 229#define ext4_es_print_tree(inode)
 230#endif
 231
 232static inline ext4_lblk_t ext4_es_end(struct extent_status *es)
 233{
 234	BUG_ON(es->es_lblk + es->es_len < es->es_lblk);
 235	return es->es_lblk + es->es_len - 1;
 236}
 237
 238static inline void ext4_es_inc_seq(struct inode *inode)
 239{
 240	struct ext4_inode_info *ei = EXT4_I(inode);
 241
 242	WRITE_ONCE(ei->i_es_seq, ei->i_es_seq + 1);
 243}
 244
 245/*
 246 * search through the tree for an delayed extent with a given offset.  If
 247 * it can't be found, try to find next extent.
 248 */
 249static struct extent_status *__es_tree_search(struct rb_root *root,
 250					      ext4_lblk_t lblk)
 251{
 252	struct rb_node *node = root->rb_node;
 253	struct extent_status *es = NULL;
 254
 255	while (node) {
 256		es = rb_entry(node, struct extent_status, rb_node);
 257		if (lblk < es->es_lblk)
 258			node = node->rb_left;
 259		else if (lblk > ext4_es_end(es))
 260			node = node->rb_right;
 261		else
 262			return es;
 263	}
 264
 265	if (es && lblk < es->es_lblk)
 266		return es;
 267
 268	if (es && lblk > ext4_es_end(es)) {
 269		node = rb_next(&es->rb_node);
 270		return node ? rb_entry(node, struct extent_status, rb_node) :
 271			      NULL;
 272	}
 273
 274	return NULL;
 275}
 276
 277/*
 278 * ext4_es_find_extent_range - find extent with specified status within block
 279 *                             range or next extent following block range in
 280 *                             extents status tree
 281 *
 282 * @inode - file containing the range
 283 * @matching_fn - pointer to function that matches extents with desired status
 284 * @lblk - logical block defining start of range
 285 * @end - logical block defining end of range
 286 * @es - extent found, if any
 287 *
 288 * Find the first extent within the block range specified by @lblk and @end
 289 * in the extents status tree that satisfies @matching_fn.  If a match
 290 * is found, it's returned in @es.  If not, and a matching extent is found
 291 * beyond the block range, it's returned in @es.  If no match is found, an
 292 * extent is returned in @es whose es_lblk, es_len, and es_pblk components
 293 * are 0.
 294 */
 295static void __es_find_extent_range(struct inode *inode,
 296				   int (*matching_fn)(struct extent_status *es),
 297				   ext4_lblk_t lblk, ext4_lblk_t end,
 298				   struct extent_status *es)
 299{
 300	struct ext4_es_tree *tree = NULL;
 301	struct extent_status *es1 = NULL;
 302	struct rb_node *node;
 303
 304	WARN_ON(es == NULL);
 305	WARN_ON(end < lblk);
 306
 307	tree = &EXT4_I(inode)->i_es_tree;
 308
 309	/* see if the extent has been cached */
 310	es->es_lblk = es->es_len = es->es_pblk = 0;
 311	es1 = READ_ONCE(tree->cache_es);
 312	if (es1 && in_range(lblk, es1->es_lblk, es1->es_len)) {
 313		es_debug("%u cached by [%u/%u) %llu %x\n",
 314			 lblk, es1->es_lblk, es1->es_len,
 315			 ext4_es_pblock(es1), ext4_es_status(es1));
 316		goto out;
 317	}
 318
 319	es1 = __es_tree_search(&tree->root, lblk);
 320
 321out:
 322	if (es1 && !matching_fn(es1)) {
 323		while ((node = rb_next(&es1->rb_node)) != NULL) {
 324			es1 = rb_entry(node, struct extent_status, rb_node);
 325			if (es1->es_lblk > end) {
 326				es1 = NULL;
 327				break;
 328			}
 329			if (matching_fn(es1))
 330				break;
 331		}
 332	}
 333
 334	if (es1 && matching_fn(es1)) {
 335		WRITE_ONCE(tree->cache_es, es1);
 336		es->es_lblk = es1->es_lblk;
 337		es->es_len = es1->es_len;
 338		es->es_pblk = es1->es_pblk;
 339	}
 340
 341}
 342
 343/*
 344 * Locking for __es_find_extent_range() for external use
 345 */
 346void ext4_es_find_extent_range(struct inode *inode,
 347			       int (*matching_fn)(struct extent_status *es),
 348			       ext4_lblk_t lblk, ext4_lblk_t end,
 349			       struct extent_status *es)
 350{
 351	es->es_lblk = es->es_len = es->es_pblk = 0;
 352
 353	if (EXT4_SB(inode->i_sb)->s_mount_state & EXT4_FC_REPLAY)
 354		return;
 355
 356	trace_ext4_es_find_extent_range_enter(inode, lblk);
 357
 358	read_lock(&EXT4_I(inode)->i_es_lock);
 359	__es_find_extent_range(inode, matching_fn, lblk, end, es);
 360	read_unlock(&EXT4_I(inode)->i_es_lock);
 361
 362	trace_ext4_es_find_extent_range_exit(inode, es);
 363}
 364
 365/*
 366 * __es_scan_range - search block range for block with specified status
 367 *                   in extents status tree
 368 *
 369 * @inode - file containing the range
 370 * @matching_fn - pointer to function that matches extents with desired status
 371 * @lblk - logical block defining start of range
 372 * @end - logical block defining end of range
 373 *
 374 * Returns true if at least one block in the specified block range satisfies
 375 * the criterion specified by @matching_fn, and false if not.  If at least
 376 * one extent has the specified status, then there is at least one block
 377 * in the cluster with that status.  Should only be called by code that has
 378 * taken i_es_lock.
 379 */
 380static bool __es_scan_range(struct inode *inode,
 381			    int (*matching_fn)(struct extent_status *es),
 382			    ext4_lblk_t start, ext4_lblk_t end)
 383{
 384	struct extent_status es;
 385
 386	__es_find_extent_range(inode, matching_fn, start, end, &es);
 387	if (es.es_len == 0)
 388		return false;   /* no matching extent in the tree */
 389	else if (es.es_lblk <= start &&
 390		 start < es.es_lblk + es.es_len)
 391		return true;
 392	else if (start <= es.es_lblk && es.es_lblk <= end)
 393		return true;
 394	else
 395		return false;
 396}
 397/*
 398 * Locking for __es_scan_range() for external use
 399 */
 400bool ext4_es_scan_range(struct inode *inode,
 401			int (*matching_fn)(struct extent_status *es),
 402			ext4_lblk_t lblk, ext4_lblk_t end)
 403{
 404	bool ret;
 405
 406	if (EXT4_SB(inode->i_sb)->s_mount_state & EXT4_FC_REPLAY)
 407		return false;
 408
 409	read_lock(&EXT4_I(inode)->i_es_lock);
 410	ret = __es_scan_range(inode, matching_fn, lblk, end);
 411	read_unlock(&EXT4_I(inode)->i_es_lock);
 412
 413	return ret;
 414}
 415
 416/*
 417 * __es_scan_clu - search cluster for block with specified status in
 418 *                 extents status tree
 419 *
 420 * @inode - file containing the cluster
 421 * @matching_fn - pointer to function that matches extents with desired status
 422 * @lblk - logical block in cluster to be searched
 423 *
 424 * Returns true if at least one extent in the cluster containing @lblk
 425 * satisfies the criterion specified by @matching_fn, and false if not.  If at
 426 * least one extent has the specified status, then there is at least one block
 427 * in the cluster with that status.  Should only be called by code that has
 428 * taken i_es_lock.
 429 */
 430static bool __es_scan_clu(struct inode *inode,
 431			  int (*matching_fn)(struct extent_status *es),
 432			  ext4_lblk_t lblk)
 433{
 434	struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
 435	ext4_lblk_t lblk_start, lblk_end;
 436
 437	lblk_start = EXT4_LBLK_CMASK(sbi, lblk);
 438	lblk_end = lblk_start + sbi->s_cluster_ratio - 1;
 439
 440	return __es_scan_range(inode, matching_fn, lblk_start, lblk_end);
 441}
 442
 443/*
 444 * Locking for __es_scan_clu() for external use
 445 */
 446bool ext4_es_scan_clu(struct inode *inode,
 447		      int (*matching_fn)(struct extent_status *es),
 448		      ext4_lblk_t lblk)
 449{
 450	bool ret;
 451
 452	if (EXT4_SB(inode->i_sb)->s_mount_state & EXT4_FC_REPLAY)
 453		return false;
 454
 455	read_lock(&EXT4_I(inode)->i_es_lock);
 456	ret = __es_scan_clu(inode, matching_fn, lblk);
 457	read_unlock(&EXT4_I(inode)->i_es_lock);
 458
 459	return ret;
 460}
 461
 462static void ext4_es_list_add(struct inode *inode)
 463{
 464	struct ext4_inode_info *ei = EXT4_I(inode);
 465	struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
 466
 467	if (!list_empty(&ei->i_es_list))
 468		return;
 469
 470	spin_lock(&sbi->s_es_lock);
 471	if (list_empty(&ei->i_es_list)) {
 472		list_add_tail(&ei->i_es_list, &sbi->s_es_list);
 473		sbi->s_es_nr_inode++;
 474	}
 475	spin_unlock(&sbi->s_es_lock);
 476}
 477
 478static void ext4_es_list_del(struct inode *inode)
 479{
 480	struct ext4_inode_info *ei = EXT4_I(inode);
 481	struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
 482
 483	spin_lock(&sbi->s_es_lock);
 484	if (!list_empty(&ei->i_es_list)) {
 485		list_del_init(&ei->i_es_list);
 486		sbi->s_es_nr_inode--;
 487		WARN_ON_ONCE(sbi->s_es_nr_inode < 0);
 488	}
 489	spin_unlock(&sbi->s_es_lock);
 490}
 491
 492static inline struct pending_reservation *__alloc_pending(bool nofail)
 493{
 494	if (!nofail)
 495		return kmem_cache_alloc(ext4_pending_cachep, GFP_ATOMIC);
 496
 497	return kmem_cache_zalloc(ext4_pending_cachep, GFP_KERNEL | __GFP_NOFAIL);
 498}
 499
 500static inline void __free_pending(struct pending_reservation *pr)
 501{
 502	kmem_cache_free(ext4_pending_cachep, pr);
 503}
 504
 505/*
 506 * Returns true if we cannot fail to allocate memory for this extent_status
 507 * entry and cannot reclaim it until its status changes.
 508 */
 509static inline bool ext4_es_must_keep(struct extent_status *es)
 510{
 511	/* fiemap, bigalloc, and seek_data/hole need to use it. */
 512	if (ext4_es_is_delayed(es))
 513		return true;
 514
 515	return false;
 516}
 517
 518static inline struct extent_status *__es_alloc_extent(bool nofail)
 519{
 520	if (!nofail)
 521		return kmem_cache_alloc(ext4_es_cachep, GFP_ATOMIC);
 522
 523	return kmem_cache_zalloc(ext4_es_cachep, GFP_KERNEL | __GFP_NOFAIL);
 524}
 525
 526static void ext4_es_init_extent(struct inode *inode, struct extent_status *es,
 527		ext4_lblk_t lblk, ext4_lblk_t len, ext4_fsblk_t pblk)
 528{
 529	es->es_lblk = lblk;
 530	es->es_len = len;
 531	es->es_pblk = pblk;
 532
 533	/* We never try to reclaim a must kept extent, so we don't count it. */
 534	if (!ext4_es_must_keep(es)) {
 535		if (!EXT4_I(inode)->i_es_shk_nr++)
 536			ext4_es_list_add(inode);
 537		percpu_counter_inc(&EXT4_SB(inode->i_sb)->
 538					s_es_stats.es_stats_shk_cnt);
 539	}
 540
 541	EXT4_I(inode)->i_es_all_nr++;
 542	percpu_counter_inc(&EXT4_SB(inode->i_sb)->s_es_stats.es_stats_all_cnt);
 543}
 544
 545static inline void __es_free_extent(struct extent_status *es)
 546{
 547	kmem_cache_free(ext4_es_cachep, es);
 548}
 549
 550static void ext4_es_free_extent(struct inode *inode, struct extent_status *es)
 551{
 552	EXT4_I(inode)->i_es_all_nr--;
 553	percpu_counter_dec(&EXT4_SB(inode->i_sb)->s_es_stats.es_stats_all_cnt);
 554
 555	/* Decrease the shrink counter when we can reclaim the extent. */
 556	if (!ext4_es_must_keep(es)) {
 557		BUG_ON(EXT4_I(inode)->i_es_shk_nr == 0);
 558		if (!--EXT4_I(inode)->i_es_shk_nr)
 559			ext4_es_list_del(inode);
 560		percpu_counter_dec(&EXT4_SB(inode->i_sb)->
 561					s_es_stats.es_stats_shk_cnt);
 562	}
 563
 564	__es_free_extent(es);
 565}
 566
 567/*
 568 * Check whether or not two extents can be merged
 569 * Condition:
 570 *  - logical block number is contiguous
 571 *  - physical block number is contiguous
 572 *  - status is equal
 573 */
 574static int ext4_es_can_be_merged(struct extent_status *es1,
 575				 struct extent_status *es2)
 576{
 577	if (ext4_es_type(es1) != ext4_es_type(es2))
 578		return 0;
 579
 580	if (((__u64) es1->es_len) + es2->es_len > EXT_MAX_BLOCKS) {
 581		pr_warn("ES assertion failed when merging extents. "
 582			"The sum of lengths of es1 (%d) and es2 (%d) "
 583			"is bigger than allowed file size (%d)\n",
 584			es1->es_len, es2->es_len, EXT_MAX_BLOCKS);
 585		WARN_ON(1);
 586		return 0;
 587	}
 588
 589	if (((__u64) es1->es_lblk) + es1->es_len != es2->es_lblk)
 590		return 0;
 591
 592	if ((ext4_es_is_written(es1) || ext4_es_is_unwritten(es1)) &&
 593	    (ext4_es_pblock(es1) + es1->es_len == ext4_es_pblock(es2)))
 594		return 1;
 595
 596	if (ext4_es_is_hole(es1))
 597		return 1;
 598
 599	/* we need to check delayed extent */
 600	if (ext4_es_is_delayed(es1))
 601		return 1;
 602
 603	return 0;
 604}
 605
 606static struct extent_status *
 607ext4_es_try_to_merge_left(struct inode *inode, struct extent_status *es)
 608{
 609	struct ext4_es_tree *tree = &EXT4_I(inode)->i_es_tree;
 610	struct extent_status *es1;
 611	struct rb_node *node;
 612
 613	node = rb_prev(&es->rb_node);
 614	if (!node)
 615		return es;
 616
 617	es1 = rb_entry(node, struct extent_status, rb_node);
 618	if (ext4_es_can_be_merged(es1, es)) {
 619		es1->es_len += es->es_len;
 620		if (ext4_es_is_referenced(es))
 621			ext4_es_set_referenced(es1);
 622		rb_erase(&es->rb_node, &tree->root);
 623		ext4_es_free_extent(inode, es);
 624		es = es1;
 625	}
 626
 627	return es;
 628}
 629
 630static struct extent_status *
 631ext4_es_try_to_merge_right(struct inode *inode, struct extent_status *es)
 632{
 633	struct ext4_es_tree *tree = &EXT4_I(inode)->i_es_tree;
 634	struct extent_status *es1;
 635	struct rb_node *node;
 636
 637	node = rb_next(&es->rb_node);
 638	if (!node)
 639		return es;
 640
 641	es1 = rb_entry(node, struct extent_status, rb_node);
 642	if (ext4_es_can_be_merged(es, es1)) {
 643		es->es_len += es1->es_len;
 644		if (ext4_es_is_referenced(es1))
 645			ext4_es_set_referenced(es);
 646		rb_erase(node, &tree->root);
 647		ext4_es_free_extent(inode, es1);
 648	}
 649
 650	return es;
 651}
 652
 653#ifdef ES_AGGRESSIVE_TEST
 654#include "ext4_extents.h"	/* Needed when ES_AGGRESSIVE_TEST is defined */
 655
 656static void ext4_es_insert_extent_ext_check(struct inode *inode,
 657					    struct extent_status *es)
 658{
 659	struct ext4_ext_path *path = NULL;
 660	struct ext4_extent *ex;
 661	ext4_lblk_t ee_block;
 662	ext4_fsblk_t ee_start;
 663	unsigned short ee_len;
 664	int depth, ee_status, es_status;
 665
 666	path = ext4_find_extent(inode, es->es_lblk, NULL, EXT4_EX_NOCACHE);
 667	if (IS_ERR(path))
 668		return;
 669
 670	depth = ext_depth(inode);
 671	ex = path[depth].p_ext;
 672
 673	if (ex) {
 674
 675		ee_block = le32_to_cpu(ex->ee_block);
 676		ee_start = ext4_ext_pblock(ex);
 677		ee_len = ext4_ext_get_actual_len(ex);
 678
 679		ee_status = ext4_ext_is_unwritten(ex) ? 1 : 0;
 680		es_status = ext4_es_is_unwritten(es) ? 1 : 0;
 681
 682		/*
 683		 * Make sure ex and es are not overlap when we try to insert
 684		 * a delayed/hole extent.
 685		 */
 686		if (!ext4_es_is_written(es) && !ext4_es_is_unwritten(es)) {
 687			if (in_range(es->es_lblk, ee_block, ee_len)) {
 688				pr_warn("ES insert assertion failed for "
 689					"inode: %lu we can find an extent "
 690					"at block [%d/%d/%llu/%c], but we "
 691					"want to add a delayed/hole extent "
 692					"[%d/%d/%llu/%x]\n",
 693					inode->i_ino, ee_block, ee_len,
 694					ee_start, ee_status ? 'u' : 'w',
 695					es->es_lblk, es->es_len,
 696					ext4_es_pblock(es), ext4_es_status(es));
 697			}
 698			goto out;
 699		}
 700
 701		/*
 702		 * We don't check ee_block == es->es_lblk, etc. because es
 703		 * might be a part of whole extent, vice versa.
 704		 */
 705		if (es->es_lblk < ee_block ||
 706		    ext4_es_pblock(es) != ee_start + es->es_lblk - ee_block) {
 707			pr_warn("ES insert assertion failed for inode: %lu "
 708				"ex_status [%d/%d/%llu/%c] != "
 709				"es_status [%d/%d/%llu/%c]\n", inode->i_ino,
 710				ee_block, ee_len, ee_start,
 711				ee_status ? 'u' : 'w', es->es_lblk, es->es_len,
 712				ext4_es_pblock(es), es_status ? 'u' : 'w');
 713			goto out;
 714		}
 715
 716		if (ee_status ^ es_status) {
 717			pr_warn("ES insert assertion failed for inode: %lu "
 718				"ex_status [%d/%d/%llu/%c] != "
 719				"es_status [%d/%d/%llu/%c]\n", inode->i_ino,
 720				ee_block, ee_len, ee_start,
 721				ee_status ? 'u' : 'w', es->es_lblk, es->es_len,
 722				ext4_es_pblock(es), es_status ? 'u' : 'w');
 723		}
 724	} else {
 725		/*
 726		 * We can't find an extent on disk.  So we need to make sure
 727		 * that we don't want to add an written/unwritten extent.
 728		 */
 729		if (!ext4_es_is_delayed(es) && !ext4_es_is_hole(es)) {
 730			pr_warn("ES insert assertion failed for inode: %lu "
 731				"can't find an extent at block %d but we want "
 732				"to add a written/unwritten extent "
 733				"[%d/%d/%llu/%x]\n", inode->i_ino,
 734				es->es_lblk, es->es_lblk, es->es_len,
 735				ext4_es_pblock(es), ext4_es_status(es));
 736		}
 737	}
 738out:
 739	ext4_free_ext_path(path);
 740}
 741
 742static void ext4_es_insert_extent_ind_check(struct inode *inode,
 743					    struct extent_status *es)
 744{
 745	struct ext4_map_blocks map;
 746	int retval;
 747
 748	/*
 749	 * Here we call ext4_ind_map_blocks to lookup a block mapping because
 750	 * 'Indirect' structure is defined in indirect.c.  So we couldn't
 751	 * access direct/indirect tree from outside.  It is too dirty to define
 752	 * this function in indirect.c file.
 753	 */
 754
 755	map.m_lblk = es->es_lblk;
 756	map.m_len = es->es_len;
 757
 758	retval = ext4_ind_map_blocks(NULL, inode, &map, 0);
 759	if (retval > 0) {
 760		if (ext4_es_is_delayed(es) || ext4_es_is_hole(es)) {
 761			/*
 762			 * We want to add a delayed/hole extent but this
 763			 * block has been allocated.
 764			 */
 765			pr_warn("ES insert assertion failed for inode: %lu "
 766				"We can find blocks but we want to add a "
 767				"delayed/hole extent [%d/%d/%llu/%x]\n",
 768				inode->i_ino, es->es_lblk, es->es_len,
 769				ext4_es_pblock(es), ext4_es_status(es));
 770			return;
 771		} else if (ext4_es_is_written(es)) {
 772			if (retval != es->es_len) {
 773				pr_warn("ES insert assertion failed for "
 774					"inode: %lu retval %d != es_len %d\n",
 775					inode->i_ino, retval, es->es_len);
 776				return;
 777			}
 778			if (map.m_pblk != ext4_es_pblock(es)) {
 779				pr_warn("ES insert assertion failed for "
 780					"inode: %lu m_pblk %llu != "
 781					"es_pblk %llu\n",
 782					inode->i_ino, map.m_pblk,
 783					ext4_es_pblock(es));
 784				return;
 785			}
 786		} else {
 787			/*
 788			 * We don't need to check unwritten extent because
 789			 * indirect-based file doesn't have it.
 790			 */
 791			BUG();
 792		}
 793	} else if (retval == 0) {
 794		if (ext4_es_is_written(es)) {
 795			pr_warn("ES insert assertion failed for inode: %lu "
 796				"We can't find the block but we want to add "
 797				"a written extent [%d/%d/%llu/%x]\n",
 798				inode->i_ino, es->es_lblk, es->es_len,
 799				ext4_es_pblock(es), ext4_es_status(es));
 800			return;
 801		}
 802	}
 803}
 804
 805static inline void ext4_es_insert_extent_check(struct inode *inode,
 806					       struct extent_status *es)
 807{
 808	/*
 809	 * We don't need to worry about the race condition because
 810	 * caller takes i_data_sem locking.
 811	 */
 812	BUG_ON(!rwsem_is_locked(&EXT4_I(inode)->i_data_sem));
 813	if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
 814		ext4_es_insert_extent_ext_check(inode, es);
 815	else
 816		ext4_es_insert_extent_ind_check(inode, es);
 817}
 818#else
 819static inline void ext4_es_insert_extent_check(struct inode *inode,
 820					       struct extent_status *es)
 821{
 822}
 823#endif
 824
 825static int __es_insert_extent(struct inode *inode, struct extent_status *newes,
 826			      struct extent_status *prealloc)
 827{
 828	struct ext4_es_tree *tree = &EXT4_I(inode)->i_es_tree;
 829	struct rb_node **p = &tree->root.rb_node;
 830	struct rb_node *parent = NULL;
 831	struct extent_status *es;
 832
 833	while (*p) {
 834		parent = *p;
 835		es = rb_entry(parent, struct extent_status, rb_node);
 836
 837		if (newes->es_lblk < es->es_lblk) {
 838			if (ext4_es_can_be_merged(newes, es)) {
 839				/*
 840				 * Here we can modify es_lblk directly
 841				 * because it isn't overlapped.
 842				 */
 843				es->es_lblk = newes->es_lblk;
 844				es->es_len += newes->es_len;
 845				if (ext4_es_is_written(es) ||
 846				    ext4_es_is_unwritten(es))
 847					ext4_es_store_pblock(es,
 848							     newes->es_pblk);
 849				es = ext4_es_try_to_merge_left(inode, es);
 850				goto out;
 851			}
 852			p = &(*p)->rb_left;
 853		} else if (newes->es_lblk > ext4_es_end(es)) {
 854			if (ext4_es_can_be_merged(es, newes)) {
 855				es->es_len += newes->es_len;
 856				es = ext4_es_try_to_merge_right(inode, es);
 857				goto out;
 858			}
 859			p = &(*p)->rb_right;
 860		} else {
 861			BUG();
 862			return -EINVAL;
 863		}
 864	}
 865
 866	if (prealloc)
 867		es = prealloc;
 868	else
 869		es = __es_alloc_extent(false);
 870	if (!es)
 871		return -ENOMEM;
 872	ext4_es_init_extent(inode, es, newes->es_lblk, newes->es_len,
 873			    newes->es_pblk);
 874
 875	rb_link_node(&es->rb_node, parent, p);
 876	rb_insert_color(&es->rb_node, &tree->root);
 877
 878out:
 879	tree->cache_es = es;
 880	return 0;
 881}
 882
 883/*
 884 * ext4_es_insert_extent() adds information to an inode's extent
 885 * status tree.
 886 */
 887void ext4_es_insert_extent(struct inode *inode, ext4_lblk_t lblk,
 888			   ext4_lblk_t len, ext4_fsblk_t pblk,
 889			   unsigned int status, bool delalloc_reserve_used)
 890{
 891	struct extent_status newes;
 892	ext4_lblk_t end = lblk + len - 1;
 893	int err1 = 0, err2 = 0, err3 = 0;
 894	int resv_used = 0, pending = 0;
 895	struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
 896	struct extent_status *es1 = NULL;
 897	struct extent_status *es2 = NULL;
 898	struct pending_reservation *pr = NULL;
 899	bool revise_pending = false;
 900
 901	if (EXT4_SB(inode->i_sb)->s_mount_state & EXT4_FC_REPLAY)
 902		return;
 903
 904	es_debug("add [%u/%u) %llu %x %d to extent status tree of inode %lu\n",
 905		 lblk, len, pblk, status, delalloc_reserve_used, inode->i_ino);
 906
 907	if (!len)
 908		return;
 909
 910	BUG_ON(end < lblk);
 911	WARN_ON_ONCE(status & EXTENT_STATUS_DELAYED);
 912
 913	newes.es_lblk = lblk;
 914	newes.es_len = len;
 915	ext4_es_store_pblock_status(&newes, pblk, status);
 916
 917	ext4_es_insert_extent_check(inode, &newes);
 918
 919	revise_pending = sbi->s_cluster_ratio > 1 &&
 920			 test_opt(inode->i_sb, DELALLOC) &&
 921			 (status & (EXTENT_STATUS_WRITTEN |
 922				    EXTENT_STATUS_UNWRITTEN));
 923retry:
 924	if (err1 && !es1)
 925		es1 = __es_alloc_extent(true);
 926	if ((err1 || err2) && !es2)
 927		es2 = __es_alloc_extent(true);
 928	if ((err1 || err2 || err3 < 0) && revise_pending && !pr)
 929		pr = __alloc_pending(true);
 930	write_lock(&EXT4_I(inode)->i_es_lock);
 931
 932	err1 = __es_remove_extent(inode, lblk, end, &resv_used, es1);
 933	if (err1 != 0)
 934		goto error;
 935	/* Free preallocated extent if it didn't get used. */
 936	if (es1) {
 937		if (!es1->es_len)
 938			__es_free_extent(es1);
 939		es1 = NULL;
 940	}
 941
 942	err2 = __es_insert_extent(inode, &newes, es2);
 943	if (err2 == -ENOMEM && !ext4_es_must_keep(&newes))
 944		err2 = 0;
 945	if (err2 != 0)
 946		goto error;
 947	/* Free preallocated extent if it didn't get used. */
 948	if (es2) {
 949		if (!es2->es_len)
 950			__es_free_extent(es2);
 951		es2 = NULL;
 952	}
 953
 954	if (revise_pending) {
 955		err3 = __revise_pending(inode, lblk, len, &pr);
 956		if (err3 < 0)
 957			goto error;
 958		if (pr) {
 959			__free_pending(pr);
 960			pr = NULL;
 961		}
 962		pending = err3;
 963	}
 964	/*
 965	 * TODO: For cache on-disk extents, there is no need to increment
 966	 * the sequence counter, this requires future optimization.
 967	 */
 968	ext4_es_inc_seq(inode);
 969error:
 970	write_unlock(&EXT4_I(inode)->i_es_lock);
 971	/*
 972	 * Reduce the reserved cluster count to reflect successful deferred
 973	 * allocation of delayed allocated clusters or direct allocation of
 974	 * clusters discovered to be delayed allocated.  Once allocated, a
 975	 * cluster is not included in the reserved count.
 976	 *
 977	 * When direct allocating (from fallocate, filemap, DIO, or clusters
 978	 * allocated when delalloc has been disabled by ext4_nonda_switch())
 979	 * an extent either 1) contains delayed blocks but start with
 980	 * non-delayed allocated blocks (e.g. hole) or 2) contains non-delayed
 981	 * allocated blocks which belong to delayed allocated clusters when
 982	 * bigalloc feature is enabled, quota has already been claimed by
 983	 * ext4_mb_new_blocks(), so release the quota reservations made for
 984	 * any previously delayed allocated clusters instead of claim them
 985	 * again.
 986	 */
 987	resv_used += pending;
 988	if (resv_used)
 989		ext4_da_update_reserve_space(inode, resv_used,
 990					     delalloc_reserve_used);
 991
 992	if (err1 || err2 || err3 < 0)
 993		goto retry;
 994
 995	trace_ext4_es_insert_extent(inode, &newes);
 996	ext4_es_print_tree(inode);
 997	return;
 998}
 999
1000/*
1001 * ext4_es_cache_extent() inserts information into the extent status
1002 * tree if and only if there isn't information about the range in
1003 * question already.
1004 */
1005void ext4_es_cache_extent(struct inode *inode, ext4_lblk_t lblk,
1006			  ext4_lblk_t len, ext4_fsblk_t pblk,
1007			  unsigned int status)
1008{
1009	struct extent_status *es;
1010	struct extent_status newes;
1011	ext4_lblk_t end = lblk + len - 1;
1012
1013	if (EXT4_SB(inode->i_sb)->s_mount_state & EXT4_FC_REPLAY)
1014		return;
1015
1016	newes.es_lblk = lblk;
1017	newes.es_len = len;
1018	ext4_es_store_pblock_status(&newes, pblk, status);
1019	trace_ext4_es_cache_extent(inode, &newes);
1020
1021	if (!len)
1022		return;
1023
1024	BUG_ON(end < lblk);
1025
1026	write_lock(&EXT4_I(inode)->i_es_lock);
1027
1028	es = __es_tree_search(&EXT4_I(inode)->i_es_tree.root, lblk);
1029	if (!es || es->es_lblk > end)
1030		__es_insert_extent(inode, &newes, NULL);
1031	write_unlock(&EXT4_I(inode)->i_es_lock);
1032}
1033
1034/*
1035 * ext4_es_lookup_extent() looks up an extent in extent status tree.
1036 *
1037 * ext4_es_lookup_extent is called by ext4_map_blocks/ext4_da_map_blocks.
1038 *
1039 * Return: 1 on found, 0 on not
1040 */
1041int ext4_es_lookup_extent(struct inode *inode, ext4_lblk_t lblk,
1042			  ext4_lblk_t *next_lblk, struct extent_status *es,
1043			  u64 *pseq)
1044{
1045	struct ext4_es_tree *tree;
1046	struct ext4_es_stats *stats;
1047	struct extent_status *es1 = NULL;
1048	struct rb_node *node;
1049	int found = 0;
1050
1051	if (EXT4_SB(inode->i_sb)->s_mount_state & EXT4_FC_REPLAY)
1052		return 0;
1053
1054	trace_ext4_es_lookup_extent_enter(inode, lblk);
1055	es_debug("lookup extent in block %u\n", lblk);
1056
1057	tree = &EXT4_I(inode)->i_es_tree;
1058	read_lock(&EXT4_I(inode)->i_es_lock);
1059
1060	/* find extent in cache firstly */
1061	es->es_lblk = es->es_len = es->es_pblk = 0;
1062	es1 = READ_ONCE(tree->cache_es);
1063	if (es1 && in_range(lblk, es1->es_lblk, es1->es_len)) {
1064		es_debug("%u cached by [%u/%u)\n",
1065			 lblk, es1->es_lblk, es1->es_len);
1066		found = 1;
1067		goto out;
1068	}
1069
1070	node = tree->root.rb_node;
1071	while (node) {
1072		es1 = rb_entry(node, struct extent_status, rb_node);
1073		if (lblk < es1->es_lblk)
1074			node = node->rb_left;
1075		else if (lblk > ext4_es_end(es1))
1076			node = node->rb_right;
1077		else {
1078			found = 1;
1079			break;
1080		}
1081	}
1082
1083out:
1084	stats = &EXT4_SB(inode->i_sb)->s_es_stats;
1085	if (found) {
1086		BUG_ON(!es1);
1087		es->es_lblk = es1->es_lblk;
1088		es->es_len = es1->es_len;
1089		es->es_pblk = es1->es_pblk;
1090		if (!ext4_es_is_referenced(es1))
1091			ext4_es_set_referenced(es1);
1092		percpu_counter_inc(&stats->es_stats_cache_hits);
1093		if (next_lblk) {
1094			node = rb_next(&es1->rb_node);
1095			if (node) {
1096				es1 = rb_entry(node, struct extent_status,
1097					       rb_node);
1098				*next_lblk = es1->es_lblk;
1099			} else
1100				*next_lblk = 0;
1101		}
1102		if (pseq)
1103			*pseq = EXT4_I(inode)->i_es_seq;
1104	} else {
1105		percpu_counter_inc(&stats->es_stats_cache_misses);
1106	}
1107
1108	read_unlock(&EXT4_I(inode)->i_es_lock);
1109
1110	trace_ext4_es_lookup_extent_exit(inode, es, found);
1111	return found;
1112}
1113
1114struct rsvd_count {
1115	int ndelayed;
1116	bool first_do_lblk_found;
1117	ext4_lblk_t first_do_lblk;
1118	ext4_lblk_t last_do_lblk;
1119	struct extent_status *left_es;
1120	bool partial;
1121	ext4_lblk_t lclu;
1122};
1123
1124/*
1125 * init_rsvd - initialize reserved count data before removing block range
1126 *	       in file from extent status tree
1127 *
1128 * @inode - file containing range
1129 * @lblk - first block in range
1130 * @es - pointer to first extent in range
1131 * @rc - pointer to reserved count data
1132 *
1133 * Assumes es is not NULL
1134 */
1135static void init_rsvd(struct inode *inode, ext4_lblk_t lblk,
1136		      struct extent_status *es, struct rsvd_count *rc)
1137{
1138	struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1139	struct rb_node *node;
1140
1141	rc->ndelayed = 0;
1142
1143	/*
1144	 * for bigalloc, note the first delayed block in the range has not
1145	 * been found, record the extent containing the block to the left of
1146	 * the region to be removed, if any, and note that there's no partial
1147	 * cluster to track
1148	 */
1149	if (sbi->s_cluster_ratio > 1) {
1150		rc->first_do_lblk_found = false;
1151		if (lblk > es->es_lblk) {
1152			rc->left_es = es;
1153		} else {
1154			node = rb_prev(&es->rb_node);
1155			rc->left_es = node ? rb_entry(node,
1156						      struct extent_status,
1157						      rb_node) : NULL;
1158		}
1159		rc->partial = false;
1160	}
1161}
1162
1163/*
1164 * count_rsvd - count the clusters containing delayed blocks in a range
1165 *	        within an extent and add to the running tally in rsvd_count
1166 *
1167 * @inode - file containing extent
1168 * @lblk - first block in range
1169 * @len - length of range in blocks
1170 * @es - pointer to extent containing clusters to be counted
1171 * @rc - pointer to reserved count data
1172 *
1173 * Tracks partial clusters found at the beginning and end of extents so
1174 * they aren't overcounted when they span adjacent extents
1175 */
1176static void count_rsvd(struct inode *inode, ext4_lblk_t lblk, long len,
1177		       struct extent_status *es, struct rsvd_count *rc)
1178{
1179	struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1180	ext4_lblk_t i, end, nclu;
1181
1182	if (!ext4_es_is_delayed(es))
1183		return;
1184
1185	WARN_ON(len <= 0);
1186
1187	if (sbi->s_cluster_ratio == 1) {
1188		rc->ndelayed += (int) len;
1189		return;
1190	}
1191
1192	/* bigalloc */
1193
1194	i = (lblk < es->es_lblk) ? es->es_lblk : lblk;
1195	end = lblk + (ext4_lblk_t) len - 1;
1196	end = (end > ext4_es_end(es)) ? ext4_es_end(es) : end;
1197
1198	/* record the first block of the first delayed extent seen */
1199	if (!rc->first_do_lblk_found) {
1200		rc->first_do_lblk = i;
1201		rc->first_do_lblk_found = true;
1202	}
1203
1204	/* update the last lblk in the region seen so far */
1205	rc->last_do_lblk = end;
1206
1207	/*
1208	 * if we're tracking a partial cluster and the current extent
1209	 * doesn't start with it, count it and stop tracking
1210	 */
1211	if (rc->partial && (rc->lclu != EXT4_B2C(sbi, i))) {
1212		rc->ndelayed++;
1213		rc->partial = false;
1214	}
1215
1216	/*
1217	 * if the first cluster doesn't start on a cluster boundary but
1218	 * ends on one, count it
1219	 */
1220	if (EXT4_LBLK_COFF(sbi, i) != 0) {
1221		if (end >= EXT4_LBLK_CFILL(sbi, i)) {
1222			rc->ndelayed++;
1223			rc->partial = false;
1224			i = EXT4_LBLK_CFILL(sbi, i) + 1;
1225		}
1226	}
1227
1228	/*
1229	 * if the current cluster starts on a cluster boundary, count the
1230	 * number of whole delayed clusters in the extent
1231	 */
1232	if ((i + sbi->s_cluster_ratio - 1) <= end) {
1233		nclu = (end - i + 1) >> sbi->s_cluster_bits;
1234		rc->ndelayed += nclu;
1235		i += nclu << sbi->s_cluster_bits;
1236	}
1237
1238	/*
1239	 * start tracking a partial cluster if there's a partial at the end
1240	 * of the current extent and we're not already tracking one
1241	 */
1242	if (!rc->partial && i <= end) {
1243		rc->partial = true;
1244		rc->lclu = EXT4_B2C(sbi, i);
1245	}
1246}
1247
1248/*
1249 * __pr_tree_search - search for a pending cluster reservation
1250 *
1251 * @root - root of pending reservation tree
1252 * @lclu - logical cluster to search for
1253 *
1254 * Returns the pending reservation for the cluster identified by @lclu
1255 * if found.  If not, returns a reservation for the next cluster if any,
1256 * and if not, returns NULL.
1257 */
1258static struct pending_reservation *__pr_tree_search(struct rb_root *root,
1259						    ext4_lblk_t lclu)
1260{
1261	struct rb_node *node = root->rb_node;
1262	struct pending_reservation *pr = NULL;
1263
1264	while (node) {
1265		pr = rb_entry(node, struct pending_reservation, rb_node);
1266		if (lclu < pr->lclu)
1267			node = node->rb_left;
1268		else if (lclu > pr->lclu)
1269			node = node->rb_right;
1270		else
1271			return pr;
1272	}
1273	if (pr && lclu < pr->lclu)
1274		return pr;
1275	if (pr && lclu > pr->lclu) {
1276		node = rb_next(&pr->rb_node);
1277		return node ? rb_entry(node, struct pending_reservation,
1278				       rb_node) : NULL;
1279	}
1280	return NULL;
1281}
1282
1283/*
1284 * get_rsvd - calculates and returns the number of cluster reservations to be
1285 *	      released when removing a block range from the extent status tree
1286 *	      and releases any pending reservations within the range
1287 *
1288 * @inode - file containing block range
1289 * @end - last block in range
1290 * @right_es - pointer to extent containing next block beyond end or NULL
1291 * @rc - pointer to reserved count data
1292 *
1293 * The number of reservations to be released is equal to the number of
1294 * clusters containing delayed blocks within the range, minus the number of
1295 * clusters still containing delayed blocks at the ends of the range, and
1296 * minus the number of pending reservations within the range.
1297 */
1298static unsigned int get_rsvd(struct inode *inode, ext4_lblk_t end,
1299			     struct extent_status *right_es,
1300			     struct rsvd_count *rc)
1301{
1302	struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1303	struct pending_reservation *pr;
1304	struct ext4_pending_tree *tree = &EXT4_I(inode)->i_pending_tree;
1305	struct rb_node *node;
1306	ext4_lblk_t first_lclu, last_lclu;
1307	bool left_delayed, right_delayed, count_pending;
1308	struct extent_status *es;
1309
1310	if (sbi->s_cluster_ratio > 1) {
1311		/* count any remaining partial cluster */
1312		if (rc->partial)
1313			rc->ndelayed++;
1314
1315		if (rc->ndelayed == 0)
1316			return 0;
1317
1318		first_lclu = EXT4_B2C(sbi, rc->first_do_lblk);
1319		last_lclu = EXT4_B2C(sbi, rc->last_do_lblk);
1320
1321		/*
1322		 * decrease the delayed count by the number of clusters at the
1323		 * ends of the range that still contain delayed blocks -
1324		 * these clusters still need to be reserved
1325		 */
1326		left_delayed = right_delayed = false;
1327
1328		es = rc->left_es;
1329		while (es && ext4_es_end(es) >=
1330		       EXT4_LBLK_CMASK(sbi, rc->first_do_lblk)) {
1331			if (ext4_es_is_delayed(es)) {
1332				rc->ndelayed--;
1333				left_delayed = true;
1334				break;
1335			}
1336			node = rb_prev(&es->rb_node);
1337			if (!node)
1338				break;
1339			es = rb_entry(node, struct extent_status, rb_node);
1340		}
1341		if (right_es && (!left_delayed || first_lclu != last_lclu)) {
1342			if (end < ext4_es_end(right_es)) {
1343				es = right_es;
1344			} else {
1345				node = rb_next(&right_es->rb_node);
1346				es = node ? rb_entry(node, struct extent_status,
1347						     rb_node) : NULL;
1348			}
1349			while (es && es->es_lblk <=
1350			       EXT4_LBLK_CFILL(sbi, rc->last_do_lblk)) {
1351				if (ext4_es_is_delayed(es)) {
1352					rc->ndelayed--;
1353					right_delayed = true;
1354					break;
1355				}
1356				node = rb_next(&es->rb_node);
1357				if (!node)
1358					break;
1359				es = rb_entry(node, struct extent_status,
1360					      rb_node);
1361			}
1362		}
1363
1364		/*
1365		 * Determine the block range that should be searched for
1366		 * pending reservations, if any.  Clusters on the ends of the
1367		 * original removed range containing delayed blocks are
1368		 * excluded.  They've already been accounted for and it's not
1369		 * possible to determine if an associated pending reservation
1370		 * should be released with the information available in the
1371		 * extents status tree.
1372		 */
1373		if (first_lclu == last_lclu) {
1374			if (left_delayed | right_delayed)
1375				count_pending = false;
1376			else
1377				count_pending = true;
1378		} else {
1379			if (left_delayed)
1380				first_lclu++;
1381			if (right_delayed)
1382				last_lclu--;
1383			if (first_lclu <= last_lclu)
1384				count_pending = true;
1385			else
1386				count_pending = false;
1387		}
1388
1389		/*
1390		 * a pending reservation found between first_lclu and last_lclu
1391		 * represents an allocated cluster that contained at least one
1392		 * delayed block, so the delayed total must be reduced by one
1393		 * for each pending reservation found and released
1394		 */
1395		if (count_pending) {
1396			pr = __pr_tree_search(&tree->root, first_lclu);
1397			while (pr && pr->lclu <= last_lclu) {
1398				rc->ndelayed--;
1399				node = rb_next(&pr->rb_node);
1400				rb_erase(&pr->rb_node, &tree->root);
1401				__free_pending(pr);
1402				if (!node)
1403					break;
1404				pr = rb_entry(node, struct pending_reservation,
1405					      rb_node);
1406			}
1407		}
1408	}
1409	return rc->ndelayed;
1410}
1411
1412
1413/*
1414 * __es_remove_extent - removes block range from extent status tree
1415 *
1416 * @inode - file containing range
1417 * @lblk - first block in range
1418 * @end - last block in range
1419 * @reserved - number of cluster reservations released
1420 * @prealloc - pre-allocated es to avoid memory allocation failures
1421 *
1422 * If @reserved is not NULL and delayed allocation is enabled, counts
1423 * block/cluster reservations freed by removing range and if bigalloc
1424 * enabled cancels pending reservations as needed. Returns 0 on success,
1425 * error code on failure.
1426 */
1427static int __es_remove_extent(struct inode *inode, ext4_lblk_t lblk,
1428			      ext4_lblk_t end, int *reserved,
1429			      struct extent_status *prealloc)
1430{
1431	struct ext4_es_tree *tree = &EXT4_I(inode)->i_es_tree;
1432	struct rb_node *node;
1433	struct extent_status *es;
1434	struct extent_status orig_es;
1435	ext4_lblk_t len1, len2;
1436	ext4_fsblk_t block;
1437	int err = 0;
1438	bool count_reserved = true;
1439	struct rsvd_count rc;
1440
1441	if (reserved == NULL || !test_opt(inode->i_sb, DELALLOC))
1442		count_reserved = false;
1443
1444	es = __es_tree_search(&tree->root, lblk);
1445	if (!es)
1446		goto out;
1447	if (es->es_lblk > end)
1448		goto out;
1449
1450	/* Simply invalidate cache_es. */
1451	tree->cache_es = NULL;
1452	if (count_reserved)
1453		init_rsvd(inode, lblk, es, &rc);
1454
1455	orig_es.es_lblk = es->es_lblk;
1456	orig_es.es_len = es->es_len;
1457	orig_es.es_pblk = es->es_pblk;
1458
1459	len1 = lblk > es->es_lblk ? lblk - es->es_lblk : 0;
1460	len2 = ext4_es_end(es) > end ? ext4_es_end(es) - end : 0;
1461	if (len1 > 0)
1462		es->es_len = len1;
1463	if (len2 > 0) {
1464		if (len1 > 0) {
1465			struct extent_status newes;
1466
1467			newes.es_lblk = end + 1;
1468			newes.es_len = len2;
1469			block = 0x7FDEADBEEFULL;
1470			if (ext4_es_is_written(&orig_es) ||
1471			    ext4_es_is_unwritten(&orig_es))
1472				block = ext4_es_pblock(&orig_es) +
1473					orig_es.es_len - len2;
1474			ext4_es_store_pblock_status(&newes, block,
1475						    ext4_es_status(&orig_es));
1476			err = __es_insert_extent(inode, &newes, prealloc);
1477			if (err) {
1478				if (!ext4_es_must_keep(&newes))
1479					return 0;
1480
1481				es->es_lblk = orig_es.es_lblk;
1482				es->es_len = orig_es.es_len;
1483				goto out;
1484			}
1485		} else {
1486			es->es_lblk = end + 1;
1487			es->es_len = len2;
1488			if (ext4_es_is_written(es) ||
1489			    ext4_es_is_unwritten(es)) {
1490				block = orig_es.es_pblk + orig_es.es_len - len2;
1491				ext4_es_store_pblock(es, block);
1492			}
1493		}
1494		if (count_reserved)
1495			count_rsvd(inode, orig_es.es_lblk + len1,
1496				   orig_es.es_len - len1 - len2, &orig_es, &rc);
1497		goto out_get_reserved;
1498	}
1499
1500	if (len1 > 0) {
1501		if (count_reserved)
1502			count_rsvd(inode, lblk, orig_es.es_len - len1,
1503				   &orig_es, &rc);
1504		node = rb_next(&es->rb_node);
1505		if (node)
1506			es = rb_entry(node, struct extent_status, rb_node);
1507		else
1508			es = NULL;
1509	}
1510
1511	while (es && ext4_es_end(es) <= end) {
1512		if (count_reserved)
1513			count_rsvd(inode, es->es_lblk, es->es_len, es, &rc);
1514		node = rb_next(&es->rb_node);
1515		rb_erase(&es->rb_node, &tree->root);
1516		ext4_es_free_extent(inode, es);
1517		if (!node) {
1518			es = NULL;
1519			break;
1520		}
1521		es = rb_entry(node, struct extent_status, rb_node);
1522	}
1523
1524	if (es && es->es_lblk < end + 1) {
1525		ext4_lblk_t orig_len = es->es_len;
1526
1527		len1 = ext4_es_end(es) - end;
1528		if (count_reserved)
1529			count_rsvd(inode, es->es_lblk, orig_len - len1,
1530				   es, &rc);
1531		es->es_lblk = end + 1;
1532		es->es_len = len1;
1533		if (ext4_es_is_written(es) || ext4_es_is_unwritten(es)) {
1534			block = es->es_pblk + orig_len - len1;
1535			ext4_es_store_pblock(es, block);
1536		}
1537	}
1538
1539out_get_reserved:
1540	if (count_reserved)
1541		*reserved = get_rsvd(inode, end, es, &rc);
1542out:
1543	return err;
1544}
1545
1546/*
1547 * ext4_es_remove_extent - removes block range from extent status tree
1548 *
1549 * @inode - file containing range
1550 * @lblk - first block in range
1551 * @len - number of blocks to remove
1552 *
1553 * Reduces block/cluster reservation count and for bigalloc cancels pending
1554 * reservations as needed.
1555 */
1556void ext4_es_remove_extent(struct inode *inode, ext4_lblk_t lblk,
1557			   ext4_lblk_t len)
1558{
1559	ext4_lblk_t end;
1560	int err = 0;
1561	int reserved = 0;
1562	struct extent_status *es = NULL;
1563
1564	if (EXT4_SB(inode->i_sb)->s_mount_state & EXT4_FC_REPLAY)
1565		return;
1566
1567	es_debug("remove [%u/%u) from extent status tree of inode %lu\n",
1568		 lblk, len, inode->i_ino);
1569
1570	if (!len)
1571		return;
1572
1573	end = lblk + len - 1;
1574	BUG_ON(end < lblk);
1575
1576retry:
1577	if (err && !es)
1578		es = __es_alloc_extent(true);
1579	/*
1580	 * ext4_clear_inode() depends on us taking i_es_lock unconditionally
1581	 * so that we are sure __es_shrink() is done with the inode before it
1582	 * is reclaimed.
1583	 */
1584	write_lock(&EXT4_I(inode)->i_es_lock);
1585	err = __es_remove_extent(inode, lblk, end, &reserved, es);
1586	if (err)
1587		goto error;
1588	/* Free preallocated extent if it didn't get used. */
1589	if (es) {
1590		if (!es->es_len)
1591			__es_free_extent(es);
1592		es = NULL;
1593	}
1594	ext4_es_inc_seq(inode);
1595error:
1596	write_unlock(&EXT4_I(inode)->i_es_lock);
1597	if (err)
1598		goto retry;
1599
1600	trace_ext4_es_remove_extent(inode, lblk, len);
1601	ext4_es_print_tree(inode);
1602	ext4_da_release_space(inode, reserved);
1603}
1604
1605static int __es_shrink(struct ext4_sb_info *sbi, int nr_to_scan,
1606		       struct ext4_inode_info *locked_ei)
1607{
1608	struct ext4_inode_info *ei;
1609	struct ext4_es_stats *es_stats;
1610	ktime_t start_time;
1611	u64 scan_time;
1612	int nr_to_walk;
1613	int nr_shrunk = 0;
1614	int retried = 0, nr_skipped = 0;
1615
1616	es_stats = &sbi->s_es_stats;
1617	start_time = ktime_get();
1618
1619retry:
1620	spin_lock(&sbi->s_es_lock);
1621	nr_to_walk = sbi->s_es_nr_inode;
1622	while (nr_to_walk-- > 0) {
1623		if (list_empty(&sbi->s_es_list)) {
1624			spin_unlock(&sbi->s_es_lock);
1625			goto out;
1626		}
1627		ei = list_first_entry(&sbi->s_es_list, struct ext4_inode_info,
1628				      i_es_list);
1629		/* Move the inode to the tail */
1630		list_move_tail(&ei->i_es_list, &sbi->s_es_list);
1631
1632		/*
1633		 * Normally we try hard to avoid shrinking precached inodes,
1634		 * but we will as a last resort.
1635		 */
1636		if (!retried && ext4_test_inode_state(&ei->vfs_inode,
1637						EXT4_STATE_EXT_PRECACHED)) {
1638			nr_skipped++;
1639			continue;
1640		}
1641
1642		if (ei == locked_ei || !write_trylock(&ei->i_es_lock)) {
1643			nr_skipped++;
1644			continue;
1645		}
1646		/*
1647		 * Now we hold i_es_lock which protects us from inode reclaim
1648		 * freeing inode under us
1649		 */
1650		spin_unlock(&sbi->s_es_lock);
1651
1652		nr_shrunk += es_reclaim_extents(ei, &nr_to_scan);
1653		write_unlock(&ei->i_es_lock);
1654
1655		if (nr_to_scan <= 0)
1656			goto out;
1657		spin_lock(&sbi->s_es_lock);
1658	}
1659	spin_unlock(&sbi->s_es_lock);
1660
1661	/*
1662	 * If we skipped any inodes, and we weren't able to make any
1663	 * forward progress, try again to scan precached inodes.
1664	 */
1665	if ((nr_shrunk == 0) && nr_skipped && !retried) {
1666		retried++;
1667		goto retry;
1668	}
1669
1670	if (locked_ei && nr_shrunk == 0)
1671		nr_shrunk = es_reclaim_extents(locked_ei, &nr_to_scan);
1672
1673out:
1674	scan_time = ktime_to_ns(ktime_sub(ktime_get(), start_time));
1675	if (likely(es_stats->es_stats_scan_time))
1676		es_stats->es_stats_scan_time = (scan_time +
1677				es_stats->es_stats_scan_time*3) / 4;
1678	else
1679		es_stats->es_stats_scan_time = scan_time;
1680	if (scan_time > es_stats->es_stats_max_scan_time)
1681		es_stats->es_stats_max_scan_time = scan_time;
1682	if (likely(es_stats->es_stats_shrunk))
1683		es_stats->es_stats_shrunk = (nr_shrunk +
1684				es_stats->es_stats_shrunk*3) / 4;
1685	else
1686		es_stats->es_stats_shrunk = nr_shrunk;
1687
1688	trace_ext4_es_shrink(sbi->s_sb, nr_shrunk, scan_time,
1689			     nr_skipped, retried);
1690	return nr_shrunk;
1691}
1692
1693static unsigned long ext4_es_count(struct shrinker *shrink,
1694				   struct shrink_control *sc)
1695{
1696	unsigned long nr;
1697	struct ext4_sb_info *sbi;
1698
1699	sbi = shrink->private_data;
1700	nr = percpu_counter_read_positive(&sbi->s_es_stats.es_stats_shk_cnt);
1701	trace_ext4_es_shrink_count(sbi->s_sb, sc->nr_to_scan, nr);
1702	return nr;
1703}
1704
1705static unsigned long ext4_es_scan(struct shrinker *shrink,
1706				  struct shrink_control *sc)
1707{
1708	struct ext4_sb_info *sbi = shrink->private_data;
1709	int nr_to_scan = sc->nr_to_scan;
1710	int ret, nr_shrunk;
1711
1712	ret = percpu_counter_read_positive(&sbi->s_es_stats.es_stats_shk_cnt);
1713	trace_ext4_es_shrink_scan_enter(sbi->s_sb, nr_to_scan, ret);
1714
1715	nr_shrunk = __es_shrink(sbi, nr_to_scan, NULL);
1716
1717	ret = percpu_counter_read_positive(&sbi->s_es_stats.es_stats_shk_cnt);
1718	trace_ext4_es_shrink_scan_exit(sbi->s_sb, nr_shrunk, ret);
1719	return nr_shrunk;
1720}
1721
1722int ext4_seq_es_shrinker_info_show(struct seq_file *seq, void *v)
1723{
1724	struct ext4_sb_info *sbi = EXT4_SB((struct super_block *) seq->private);
1725	struct ext4_es_stats *es_stats = &sbi->s_es_stats;
1726	struct ext4_inode_info *ei, *max = NULL;
1727	unsigned int inode_cnt = 0;
1728
1729	if (v != SEQ_START_TOKEN)
1730		return 0;
1731
1732	/* here we just find an inode that has the max nr. of objects */
1733	spin_lock(&sbi->s_es_lock);
1734	list_for_each_entry(ei, &sbi->s_es_list, i_es_list) {
1735		inode_cnt++;
1736		if (max && max->i_es_all_nr < ei->i_es_all_nr)
1737			max = ei;
1738		else if (!max)
1739			max = ei;
1740	}
1741	spin_unlock(&sbi->s_es_lock);
1742
1743	seq_printf(seq, "stats:\n  %lld objects\n  %lld reclaimable objects\n",
1744		   percpu_counter_sum_positive(&es_stats->es_stats_all_cnt),
1745		   percpu_counter_sum_positive(&es_stats->es_stats_shk_cnt));
1746	seq_printf(seq, "  %lld/%lld cache hits/misses\n",
1747		   percpu_counter_sum_positive(&es_stats->es_stats_cache_hits),
1748		   percpu_counter_sum_positive(&es_stats->es_stats_cache_misses));
1749	if (inode_cnt)
1750		seq_printf(seq, "  %d inodes on list\n", inode_cnt);
1751
1752	seq_printf(seq, "average:\n  %llu us scan time\n",
1753	    div_u64(es_stats->es_stats_scan_time, 1000));
1754	seq_printf(seq, "  %lu shrunk objects\n", es_stats->es_stats_shrunk);
1755	if (inode_cnt)
1756		seq_printf(seq,
1757		    "maximum:\n  %lu inode (%u objects, %u reclaimable)\n"
1758		    "  %llu us max scan time\n",
1759		    max->vfs_inode.i_ino, max->i_es_all_nr, max->i_es_shk_nr,
1760		    div_u64(es_stats->es_stats_max_scan_time, 1000));
1761
1762	return 0;
1763}
1764
1765int ext4_es_register_shrinker(struct ext4_sb_info *sbi)
1766{
1767	int err;
1768
1769	/* Make sure we have enough bits for physical block number */
1770	BUILD_BUG_ON(ES_SHIFT < 48);
1771	INIT_LIST_HEAD(&sbi->s_es_list);
1772	sbi->s_es_nr_inode = 0;
1773	spin_lock_init(&sbi->s_es_lock);
1774	sbi->s_es_stats.es_stats_shrunk = 0;
1775	err = percpu_counter_init(&sbi->s_es_stats.es_stats_cache_hits, 0,
1776				  GFP_KERNEL);
1777	if (err)
1778		return err;
1779	err = percpu_counter_init(&sbi->s_es_stats.es_stats_cache_misses, 0,
1780				  GFP_KERNEL);
1781	if (err)
1782		goto err1;
1783	sbi->s_es_stats.es_stats_scan_time = 0;
1784	sbi->s_es_stats.es_stats_max_scan_time = 0;
1785	err = percpu_counter_init(&sbi->s_es_stats.es_stats_all_cnt, 0, GFP_KERNEL);
1786	if (err)
1787		goto err2;
1788	err = percpu_counter_init(&sbi->s_es_stats.es_stats_shk_cnt, 0, GFP_KERNEL);
1789	if (err)
1790		goto err3;
1791
1792	sbi->s_es_shrinker = shrinker_alloc(0, "ext4-es:%s", sbi->s_sb->s_id);
1793	if (!sbi->s_es_shrinker) {
1794		err = -ENOMEM;
1795		goto err4;
1796	}
1797
1798	sbi->s_es_shrinker->scan_objects = ext4_es_scan;
1799	sbi->s_es_shrinker->count_objects = ext4_es_count;
1800	sbi->s_es_shrinker->private_data = sbi;
1801
1802	shrinker_register(sbi->s_es_shrinker);
1803
1804	return 0;
1805err4:
1806	percpu_counter_destroy(&sbi->s_es_stats.es_stats_shk_cnt);
1807err3:
1808	percpu_counter_destroy(&sbi->s_es_stats.es_stats_all_cnt);
1809err2:
1810	percpu_counter_destroy(&sbi->s_es_stats.es_stats_cache_misses);
1811err1:
1812	percpu_counter_destroy(&sbi->s_es_stats.es_stats_cache_hits);
1813	return err;
1814}
1815
1816void ext4_es_unregister_shrinker(struct ext4_sb_info *sbi)
1817{
1818	percpu_counter_destroy(&sbi->s_es_stats.es_stats_cache_hits);
1819	percpu_counter_destroy(&sbi->s_es_stats.es_stats_cache_misses);
1820	percpu_counter_destroy(&sbi->s_es_stats.es_stats_all_cnt);
1821	percpu_counter_destroy(&sbi->s_es_stats.es_stats_shk_cnt);
1822	shrinker_free(sbi->s_es_shrinker);
1823}
1824
1825/*
1826 * Shrink extents in given inode from ei->i_es_shrink_lblk till end. Scan at
1827 * most *nr_to_scan extents, update *nr_to_scan accordingly.
1828 *
1829 * Return 0 if we hit end of tree / interval, 1 if we exhausted nr_to_scan.
1830 * Increment *nr_shrunk by the number of reclaimed extents. Also update
1831 * ei->i_es_shrink_lblk to where we should continue scanning.
1832 */
1833static int es_do_reclaim_extents(struct ext4_inode_info *ei, ext4_lblk_t end,
1834				 int *nr_to_scan, int *nr_shrunk)
1835{
1836	struct inode *inode = &ei->vfs_inode;
1837	struct ext4_es_tree *tree = &ei->i_es_tree;
1838	struct extent_status *es;
1839	struct rb_node *node;
1840
1841	es = __es_tree_search(&tree->root, ei->i_es_shrink_lblk);
1842	if (!es)
1843		goto out_wrap;
1844
1845	while (*nr_to_scan > 0) {
1846		if (es->es_lblk > end) {
1847			ei->i_es_shrink_lblk = end + 1;
1848			return 0;
1849		}
1850
1851		(*nr_to_scan)--;
1852		node = rb_next(&es->rb_node);
1853
1854		if (ext4_es_must_keep(es))
1855			goto next;
1856		if (ext4_es_is_referenced(es)) {
1857			ext4_es_clear_referenced(es);
1858			goto next;
1859		}
1860
1861		rb_erase(&es->rb_node, &tree->root);
1862		ext4_es_free_extent(inode, es);
1863		(*nr_shrunk)++;
1864next:
1865		if (!node)
1866			goto out_wrap;
1867		es = rb_entry(node, struct extent_status, rb_node);
1868	}
1869	ei->i_es_shrink_lblk = es->es_lblk;
1870	return 1;
1871out_wrap:
1872	ei->i_es_shrink_lblk = 0;
1873	return 0;
1874}
1875
1876static int es_reclaim_extents(struct ext4_inode_info *ei, int *nr_to_scan)
1877{
1878	struct inode *inode = &ei->vfs_inode;
1879	int nr_shrunk = 0;
1880	ext4_lblk_t start = ei->i_es_shrink_lblk;
1881	static DEFINE_RATELIMIT_STATE(_rs, DEFAULT_RATELIMIT_INTERVAL,
1882				      DEFAULT_RATELIMIT_BURST);
1883
1884	if (ei->i_es_shk_nr == 0)
1885		return 0;
1886
1887	if (ext4_test_inode_state(inode, EXT4_STATE_EXT_PRECACHED) &&
1888	    __ratelimit(&_rs))
1889		ext4_warning(inode->i_sb, "forced shrink of precached extents");
1890
1891	if (!es_do_reclaim_extents(ei, EXT_MAX_BLOCKS, nr_to_scan, &nr_shrunk) &&
1892	    start != 0)
1893		es_do_reclaim_extents(ei, start - 1, nr_to_scan, &nr_shrunk);
1894
1895	ei->i_es_tree.cache_es = NULL;
1896	return nr_shrunk;
1897}
1898
1899/*
1900 * Called to support EXT4_IOC_CLEAR_ES_CACHE.  We can only remove
1901 * discretionary entries from the extent status cache.  (Some entries
1902 * must be present for proper operations.)
1903 */
1904void ext4_clear_inode_es(struct inode *inode)
1905{
1906	struct ext4_inode_info *ei = EXT4_I(inode);
1907	struct extent_status *es;
1908	struct ext4_es_tree *tree;
1909	struct rb_node *node;
1910
1911	write_lock(&ei->i_es_lock);
1912	tree = &EXT4_I(inode)->i_es_tree;
1913	tree->cache_es = NULL;
1914	node = rb_first(&tree->root);
1915	while (node) {
1916		es = rb_entry(node, struct extent_status, rb_node);
1917		node = rb_next(node);
1918		if (!ext4_es_must_keep(es)) {
1919			rb_erase(&es->rb_node, &tree->root);
1920			ext4_es_free_extent(inode, es);
1921		}
1922	}
1923	ext4_clear_inode_state(inode, EXT4_STATE_EXT_PRECACHED);
1924	write_unlock(&ei->i_es_lock);
1925}
1926
1927#ifdef ES_DEBUG__
1928static void ext4_print_pending_tree(struct inode *inode)
1929{
1930	struct ext4_pending_tree *tree;
1931	struct rb_node *node;
1932	struct pending_reservation *pr;
1933
1934	printk(KERN_DEBUG "pending reservations for inode %lu:", inode->i_ino);
1935	tree = &EXT4_I(inode)->i_pending_tree;
1936	node = rb_first(&tree->root);
1937	while (node) {
1938		pr = rb_entry(node, struct pending_reservation, rb_node);
1939		printk(KERN_DEBUG " %u", pr->lclu);
1940		node = rb_next(node);
1941	}
1942	printk(KERN_DEBUG "\n");
1943}
1944#else
1945#define ext4_print_pending_tree(inode)
1946#endif
1947
1948int __init ext4_init_pending(void)
1949{
1950	ext4_pending_cachep = KMEM_CACHE(pending_reservation, SLAB_RECLAIM_ACCOUNT);
1951	if (ext4_pending_cachep == NULL)
1952		return -ENOMEM;
1953	return 0;
1954}
1955
1956void ext4_exit_pending(void)
1957{
1958	kmem_cache_destroy(ext4_pending_cachep);
1959}
1960
1961void ext4_init_pending_tree(struct ext4_pending_tree *tree)
1962{
1963	tree->root = RB_ROOT;
1964}
1965
1966/*
1967 * __get_pending - retrieve a pointer to a pending reservation
1968 *
1969 * @inode - file containing the pending cluster reservation
1970 * @lclu - logical cluster of interest
1971 *
1972 * Returns a pointer to a pending reservation if it's a member of
1973 * the set, and NULL if not.  Must be called holding i_es_lock.
1974 */
1975static struct pending_reservation *__get_pending(struct inode *inode,
1976						 ext4_lblk_t lclu)
1977{
1978	struct ext4_pending_tree *tree;
1979	struct rb_node *node;
1980	struct pending_reservation *pr = NULL;
1981
1982	tree = &EXT4_I(inode)->i_pending_tree;
1983	node = (&tree->root)->rb_node;
1984
1985	while (node) {
1986		pr = rb_entry(node, struct pending_reservation, rb_node);
1987		if (lclu < pr->lclu)
1988			node = node->rb_left;
1989		else if (lclu > pr->lclu)
1990			node = node->rb_right;
1991		else if (lclu == pr->lclu)
1992			return pr;
1993	}
1994	return NULL;
1995}
1996
1997/*
1998 * __insert_pending - adds a pending cluster reservation to the set of
1999 *                    pending reservations
2000 *
2001 * @inode - file containing the cluster
2002 * @lblk - logical block in the cluster to be added
2003 * @prealloc - preallocated pending entry
2004 *
2005 * Returns 1 on successful insertion and -ENOMEM on failure.  If the
2006 * pending reservation is already in the set, returns successfully.
2007 */
2008static int __insert_pending(struct inode *inode, ext4_lblk_t lblk,
2009			    struct pending_reservation **prealloc)
2010{
2011	struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
2012	struct ext4_pending_tree *tree = &EXT4_I(inode)->i_pending_tree;
2013	struct rb_node **p = &tree->root.rb_node;
2014	struct rb_node *parent = NULL;
2015	struct pending_reservation *pr;
2016	ext4_lblk_t lclu;
2017	int ret = 0;
2018
2019	lclu = EXT4_B2C(sbi, lblk);
2020	/* search to find parent for insertion */
2021	while (*p) {
2022		parent = *p;
2023		pr = rb_entry(parent, struct pending_reservation, rb_node);
2024
2025		if (lclu < pr->lclu) {
2026			p = &(*p)->rb_left;
2027		} else if (lclu > pr->lclu) {
2028			p = &(*p)->rb_right;
2029		} else {
2030			/* pending reservation already inserted */
2031			goto out;
2032		}
2033	}
2034
2035	if (likely(*prealloc == NULL)) {
2036		pr = __alloc_pending(false);
2037		if (!pr) {
2038			ret = -ENOMEM;
2039			goto out;
2040		}
2041	} else {
2042		pr = *prealloc;
2043		*prealloc = NULL;
2044	}
2045	pr->lclu = lclu;
2046
2047	rb_link_node(&pr->rb_node, parent, p);
2048	rb_insert_color(&pr->rb_node, &tree->root);
2049	ret = 1;
2050
2051out:
2052	return ret;
2053}
2054
2055/*
2056 * __remove_pending - removes a pending cluster reservation from the set
2057 *                    of pending reservations
2058 *
2059 * @inode - file containing the cluster
2060 * @lblk - logical block in the pending cluster reservation to be removed
2061 *
2062 * Returns successfully if pending reservation is not a member of the set.
2063 */
2064static void __remove_pending(struct inode *inode, ext4_lblk_t lblk)
2065{
2066	struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
2067	struct pending_reservation *pr;
2068	struct ext4_pending_tree *tree;
2069
2070	pr = __get_pending(inode, EXT4_B2C(sbi, lblk));
2071	if (pr != NULL) {
2072		tree = &EXT4_I(inode)->i_pending_tree;
2073		rb_erase(&pr->rb_node, &tree->root);
2074		__free_pending(pr);
2075	}
2076}
2077
2078/*
2079 * ext4_remove_pending - removes a pending cluster reservation from the set
2080 *                       of pending reservations
2081 *
2082 * @inode - file containing the cluster
2083 * @lblk - logical block in the pending cluster reservation to be removed
2084 *
2085 * Locking for external use of __remove_pending.
2086 */
2087void ext4_remove_pending(struct inode *inode, ext4_lblk_t lblk)
2088{
2089	struct ext4_inode_info *ei = EXT4_I(inode);
2090
2091	write_lock(&ei->i_es_lock);
2092	__remove_pending(inode, lblk);
2093	write_unlock(&ei->i_es_lock);
2094}
2095
2096/*
2097 * ext4_is_pending - determine whether a cluster has a pending reservation
2098 *                   on it
2099 *
2100 * @inode - file containing the cluster
2101 * @lblk - logical block in the cluster
2102 *
2103 * Returns true if there's a pending reservation for the cluster in the
2104 * set of pending reservations, and false if not.
2105 */
2106bool ext4_is_pending(struct inode *inode, ext4_lblk_t lblk)
2107{
2108	struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
2109	struct ext4_inode_info *ei = EXT4_I(inode);
2110	bool ret;
2111
2112	read_lock(&ei->i_es_lock);
2113	ret = (bool)(__get_pending(inode, EXT4_B2C(sbi, lblk)) != NULL);
2114	read_unlock(&ei->i_es_lock);
2115
2116	return ret;
2117}
2118
2119/*
2120 * ext4_es_insert_delayed_extent - adds some delayed blocks to the extents
2121 *                                 status tree, adding a pending reservation
2122 *                                 where needed
2123 *
2124 * @inode - file containing the newly added block
2125 * @lblk - start logical block to be added
2126 * @len - length of blocks to be added
2127 * @lclu_allocated/end_allocated - indicates whether a physical cluster has
2128 *                                 been allocated for the logical cluster
2129 *                                 that contains the start/end block. Note that
2130 *                                 end_allocated should always be set to false
2131 *                                 if the start and the end block are in the
2132 *                                 same cluster
2133 */
2134void ext4_es_insert_delayed_extent(struct inode *inode, ext4_lblk_t lblk,
2135				   ext4_lblk_t len, bool lclu_allocated,
2136				   bool end_allocated)
2137{
2138	struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
2139	struct extent_status newes;
2140	ext4_lblk_t end = lblk + len - 1;
2141	int err1 = 0, err2 = 0, err3 = 0;
2142	struct extent_status *es1 = NULL;
2143	struct extent_status *es2 = NULL;
2144	struct pending_reservation *pr1 = NULL;
2145	struct pending_reservation *pr2 = NULL;
2146
2147	if (EXT4_SB(inode->i_sb)->s_mount_state & EXT4_FC_REPLAY)
2148		return;
2149
2150	es_debug("add [%u/%u) delayed to extent status tree of inode %lu\n",
2151		 lblk, len, inode->i_ino);
2152	if (!len)
2153		return;
2154
2155	WARN_ON_ONCE((EXT4_B2C(sbi, lblk) == EXT4_B2C(sbi, end)) &&
2156		     end_allocated);
2157
2158	newes.es_lblk = lblk;
2159	newes.es_len = len;
2160	ext4_es_store_pblock_status(&newes, ~0, EXTENT_STATUS_DELAYED);
2161
2162	ext4_es_insert_extent_check(inode, &newes);
2163
2164retry:
2165	if (err1 && !es1)
2166		es1 = __es_alloc_extent(true);
2167	if ((err1 || err2) && !es2)
2168		es2 = __es_alloc_extent(true);
2169	if (err1 || err2 || err3 < 0) {
2170		if (lclu_allocated && !pr1)
2171			pr1 = __alloc_pending(true);
2172		if (end_allocated && !pr2)
2173			pr2 = __alloc_pending(true);
2174	}
2175	write_lock(&EXT4_I(inode)->i_es_lock);
2176
2177	err1 = __es_remove_extent(inode, lblk, end, NULL, es1);
2178	if (err1 != 0)
2179		goto error;
2180	/* Free preallocated extent if it didn't get used. */
2181	if (es1) {
2182		if (!es1->es_len)
2183			__es_free_extent(es1);
2184		es1 = NULL;
2185	}
2186
2187	err2 = __es_insert_extent(inode, &newes, es2);
2188	if (err2 != 0)
2189		goto error;
2190	/* Free preallocated extent if it didn't get used. */
2191	if (es2) {
2192		if (!es2->es_len)
2193			__es_free_extent(es2);
2194		es2 = NULL;
2195	}
2196
2197	if (lclu_allocated) {
2198		err3 = __insert_pending(inode, lblk, &pr1);
2199		if (err3 < 0)
2200			goto error;
2201		if (pr1) {
2202			__free_pending(pr1);
2203			pr1 = NULL;
2204		}
2205	}
2206	if (end_allocated) {
2207		err3 = __insert_pending(inode, end, &pr2);
2208		if (err3 < 0)
2209			goto error;
2210		if (pr2) {
2211			__free_pending(pr2);
2212			pr2 = NULL;
2213		}
2214	}
2215	ext4_es_inc_seq(inode);
2216error:
2217	write_unlock(&EXT4_I(inode)->i_es_lock);
2218	if (err1 || err2 || err3 < 0)
2219		goto retry;
2220
2221	trace_ext4_es_insert_delayed_extent(inode, &newes, lclu_allocated,
2222					    end_allocated);
2223	ext4_es_print_tree(inode);
2224	ext4_print_pending_tree(inode);
2225	return;
2226}
2227
2228/*
2229 * __revise_pending - makes, cancels, or leaves unchanged pending cluster
2230 *                    reservations for a specified block range depending
2231 *                    upon the presence or absence of delayed blocks
2232 *                    outside the range within clusters at the ends of the
2233 *                    range
2234 *
2235 * @inode - file containing the range
2236 * @lblk - logical block defining the start of range
2237 * @len  - length of range in blocks
2238 * @prealloc - preallocated pending entry
2239 *
2240 * Used after a newly allocated extent is added to the extents status tree.
2241 * Requires that the extents in the range have either written or unwritten
2242 * status.  Must be called while holding i_es_lock. Returns number of new
2243 * inserts pending cluster on insert pendings, returns 0 on remove pendings,
2244 * return -ENOMEM on failure.
2245 */
2246static int __revise_pending(struct inode *inode, ext4_lblk_t lblk,
2247			    ext4_lblk_t len,
2248			    struct pending_reservation **prealloc)
2249{
2250	struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
2251	ext4_lblk_t end = lblk + len - 1;
2252	ext4_lblk_t first, last;
2253	bool f_del = false, l_del = false;
2254	int pendings = 0;
2255	int ret = 0;
2256
2257	if (len == 0)
2258		return 0;
2259
2260	/*
2261	 * Two cases - block range within single cluster and block range
2262	 * spanning two or more clusters.  Note that a cluster belonging
2263	 * to a range starting and/or ending on a cluster boundary is treated
2264	 * as if it does not contain a delayed extent.  The new range may
2265	 * have allocated space for previously delayed blocks out to the
2266	 * cluster boundary, requiring that any pre-existing pending
2267	 * reservation be canceled.  Because this code only looks at blocks
2268	 * outside the range, it should revise pending reservations
2269	 * correctly even if the extent represented by the range can't be
2270	 * inserted in the extents status tree due to ENOSPC.
2271	 */
2272
2273	if (EXT4_B2C(sbi, lblk) == EXT4_B2C(sbi, end)) {
2274		first = EXT4_LBLK_CMASK(sbi, lblk);
2275		if (first != lblk)
2276			f_del = __es_scan_range(inode, &ext4_es_is_delayed,
2277						first, lblk - 1);
2278		if (f_del) {
2279			ret = __insert_pending(inode, first, prealloc);
2280			if (ret < 0)
2281				goto out;
2282			pendings += ret;
2283		} else {
2284			last = EXT4_LBLK_CMASK(sbi, end) +
2285			       sbi->s_cluster_ratio - 1;
2286			if (last != end)
2287				l_del = __es_scan_range(inode,
2288							&ext4_es_is_delayed,
2289							end + 1, last);
2290			if (l_del) {
2291				ret = __insert_pending(inode, last, prealloc);
2292				if (ret < 0)
2293					goto out;
2294				pendings += ret;
2295			} else
2296				__remove_pending(inode, last);
2297		}
2298	} else {
2299		first = EXT4_LBLK_CMASK(sbi, lblk);
2300		if (first != lblk)
2301			f_del = __es_scan_range(inode, &ext4_es_is_delayed,
2302						first, lblk - 1);
2303		if (f_del) {
2304			ret = __insert_pending(inode, first, prealloc);
2305			if (ret < 0)
2306				goto out;
2307			pendings += ret;
2308		} else
2309			__remove_pending(inode, first);
2310
2311		last = EXT4_LBLK_CMASK(sbi, end) + sbi->s_cluster_ratio - 1;
2312		if (last != end)
2313			l_del = __es_scan_range(inode, &ext4_es_is_delayed,
2314						end + 1, last);
2315		if (l_del) {
2316			ret = __insert_pending(inode, last, prealloc);
2317			if (ret < 0)
2318				goto out;
2319			pendings += ret;
2320		} else
2321			__remove_pending(inode, last);
2322	}
2323out:
2324	return (ret < 0) ? ret : pendings;
2325}