xref: /openbmc/linux/fs/ext4/extents_status.c (revision e23feb16)
1 /*
2  *  fs/ext4/extents_status.c
3  *
4  * Written by Yongqiang Yang <xiaoqiangnk@gmail.com>
5  * Modified by
6  *	Allison Henderson <achender@linux.vnet.ibm.com>
7  *	Hugh Dickins <hughd@google.com>
8  *	Zheng Liu <wenqing.lz@taobao.com>
9  *
10  * Ext4 extents status tree core functions.
11  */
12 #include <linux/rbtree.h>
13 #include <linux/list_sort.h>
14 #include "ext4.h"
15 #include "extents_status.h"
16 
17 #include <trace/events/ext4.h>
18 
19 /*
20  * According to previous discussion in Ext4 Developer Workshop, we
21  * will introduce a new structure called io tree to track all extent
22  * status in order to solve some problems that we have met
23  * (e.g. Reservation space warning), and provide extent-level locking.
24  * Delay extent tree is the first step to achieve this goal.  It is
25  * original built by Yongqiang Yang.  At that time it is called delay
26  * extent tree, whose goal is only track delayed extents in memory to
27  * simplify the implementation of fiemap and bigalloc, and introduce
28  * lseek SEEK_DATA/SEEK_HOLE support.  That is why it is still called
29  * delay extent tree at the first commit.  But for better understand
30  * what it does, it has been rename to extent status tree.
31  *
32  * Step1:
33  * Currently the first step has been done.  All delayed extents are
34  * tracked in the tree.  It maintains the delayed extent when a delayed
35  * allocation is issued, and the delayed extent is written out or
36  * invalidated.  Therefore the implementation of fiemap and bigalloc
37  * are simplified, and SEEK_DATA/SEEK_HOLE are introduced.
38  *
39  * The following comment describes the implemenmtation of extent
40  * status tree and future works.
41  *
42  * Step2:
43  * In this step all extent status are tracked by extent status tree.
44  * Thus, we can first try to lookup a block mapping in this tree before
45  * finding it in extent tree.  Hence, single extent cache can be removed
46  * because extent status tree can do a better job.  Extents in status
47  * tree are loaded on-demand.  Therefore, the extent status tree may not
48  * contain all of the extents in a file.  Meanwhile we define a shrinker
49  * to reclaim memory from extent status tree because fragmented extent
50  * tree will make status tree cost too much memory.  written/unwritten/-
51  * hole extents in the tree will be reclaimed by this shrinker when we
52  * are under high memory pressure.  Delayed extents will not be
53  * reclimed because fiemap, bigalloc, and seek_data/hole need it.
54  */
55 
56 /*
57  * Extent status tree implementation for ext4.
58  *
59  *
60  * ==========================================================================
61  * Extent status tree tracks all extent status.
62  *
63  * 1. Why we need to implement extent status tree?
64  *
65  * Without extent status tree, ext4 identifies a delayed extent by looking
66  * up page cache, this has several deficiencies - complicated, buggy,
67  * and inefficient code.
68  *
69  * FIEMAP, SEEK_HOLE/DATA, bigalloc, and writeout all need to know if a
70  * block or a range of blocks are belonged to a delayed extent.
71  *
72  * Let us have a look at how they do without extent status tree.
73  *   --	FIEMAP
74  *	FIEMAP looks up page cache to identify delayed allocations from holes.
75  *
76  *   --	SEEK_HOLE/DATA
77  *	SEEK_HOLE/DATA has the same problem as FIEMAP.
78  *
79  *   --	bigalloc
80  *	bigalloc looks up page cache to figure out if a block is
81  *	already under delayed allocation or not to determine whether
82  *	quota reserving is needed for the cluster.
83  *
84  *   --	writeout
85  *	Writeout looks up whole page cache to see if a buffer is
86  *	mapped, If there are not very many delayed buffers, then it is
87  *	time comsuming.
88  *
89  * With extent status tree implementation, FIEMAP, SEEK_HOLE/DATA,
90  * bigalloc and writeout can figure out if a block or a range of
91  * blocks is under delayed allocation(belonged to a delayed extent) or
92  * not by searching the extent tree.
93  *
94  *
95  * ==========================================================================
96  * 2. Ext4 extent status tree impelmentation
97  *
98  *   --	extent
99  *	A extent is a range of blocks which are contiguous logically and
100  *	physically.  Unlike extent in extent tree, this extent in ext4 is
101  *	a in-memory struct, there is no corresponding on-disk data.  There
102  *	is no limit on length of extent, so an extent can contain as many
103  *	blocks as they are contiguous logically and physically.
104  *
105  *   --	extent status tree
106  *	Every inode has an extent status tree and all allocation blocks
107  *	are added to the tree with different status.  The extent in the
108  *	tree are ordered by logical block no.
109  *
110  *   --	operations on a extent status tree
111  *	There are three important operations on a delayed extent tree: find
112  *	next extent, adding a extent(a range of blocks) and removing a extent.
113  *
114  *   --	race on a extent status tree
115  *	Extent status tree is protected by inode->i_es_lock.
116  *
117  *   --	memory consumption
118  *      Fragmented extent tree will make extent status tree cost too much
119  *      memory.  Hence, we will reclaim written/unwritten/hole extents from
120  *      the tree under a heavy memory pressure.
121  *
122  *
123  * ==========================================================================
124  * 3. Performance analysis
125  *
126  *   --	overhead
127  *	1. There is a cache extent for write access, so if writes are
128  *	not very random, adding space operaions are in O(1) time.
129  *
130  *   --	gain
131  *	2. Code is much simpler, more readable, more maintainable and
132  *	more efficient.
133  *
134  *
135  * ==========================================================================
136  * 4. TODO list
137  *
138  *   -- Refactor delayed space reservation
139  *
140  *   -- Extent-level locking
141  */
142 
143 static struct kmem_cache *ext4_es_cachep;
144 
145 static int __es_insert_extent(struct inode *inode, struct extent_status *newes);
146 static int __es_remove_extent(struct inode *inode, ext4_lblk_t lblk,
147 			      ext4_lblk_t end);
148 static int __es_try_to_reclaim_extents(struct ext4_inode_info *ei,
149 				       int nr_to_scan);
150 static int __ext4_es_shrink(struct ext4_sb_info *sbi, int nr_to_scan,
151 			    struct ext4_inode_info *locked_ei);
152 
153 int __init ext4_init_es(void)
154 {
155 	ext4_es_cachep = kmem_cache_create("ext4_extent_status",
156 					   sizeof(struct extent_status),
157 					   0, (SLAB_RECLAIM_ACCOUNT), NULL);
158 	if (ext4_es_cachep == NULL)
159 		return -ENOMEM;
160 	return 0;
161 }
162 
163 void ext4_exit_es(void)
164 {
165 	if (ext4_es_cachep)
166 		kmem_cache_destroy(ext4_es_cachep);
167 }
168 
169 void ext4_es_init_tree(struct ext4_es_tree *tree)
170 {
171 	tree->root = RB_ROOT;
172 	tree->cache_es = NULL;
173 }
174 
175 #ifdef ES_DEBUG__
176 static void ext4_es_print_tree(struct inode *inode)
177 {
178 	struct ext4_es_tree *tree;
179 	struct rb_node *node;
180 
181 	printk(KERN_DEBUG "status extents for inode %lu:", inode->i_ino);
182 	tree = &EXT4_I(inode)->i_es_tree;
183 	node = rb_first(&tree->root);
184 	while (node) {
185 		struct extent_status *es;
186 		es = rb_entry(node, struct extent_status, rb_node);
187 		printk(KERN_DEBUG " [%u/%u) %llu %llx",
188 		       es->es_lblk, es->es_len,
189 		       ext4_es_pblock(es), ext4_es_status(es));
190 		node = rb_next(node);
191 	}
192 	printk(KERN_DEBUG "\n");
193 }
194 #else
195 #define ext4_es_print_tree(inode)
196 #endif
197 
198 static inline ext4_lblk_t ext4_es_end(struct extent_status *es)
199 {
200 	BUG_ON(es->es_lblk + es->es_len < es->es_lblk);
201 	return es->es_lblk + es->es_len - 1;
202 }
203 
204 /*
205  * search through the tree for an delayed extent with a given offset.  If
206  * it can't be found, try to find next extent.
207  */
208 static struct extent_status *__es_tree_search(struct rb_root *root,
209 					      ext4_lblk_t lblk)
210 {
211 	struct rb_node *node = root->rb_node;
212 	struct extent_status *es = NULL;
213 
214 	while (node) {
215 		es = rb_entry(node, struct extent_status, rb_node);
216 		if (lblk < es->es_lblk)
217 			node = node->rb_left;
218 		else if (lblk > ext4_es_end(es))
219 			node = node->rb_right;
220 		else
221 			return es;
222 	}
223 
224 	if (es && lblk < es->es_lblk)
225 		return es;
226 
227 	if (es && lblk > ext4_es_end(es)) {
228 		node = rb_next(&es->rb_node);
229 		return node ? rb_entry(node, struct extent_status, rb_node) :
230 			      NULL;
231 	}
232 
233 	return NULL;
234 }
235 
236 /*
237  * ext4_es_find_delayed_extent_range: find the 1st delayed extent covering
238  * @es->lblk if it exists, otherwise, the next extent after @es->lblk.
239  *
240  * @inode: the inode which owns delayed extents
241  * @lblk: the offset where we start to search
242  * @end: the offset where we stop to search
243  * @es: delayed extent that we found
244  */
245 void ext4_es_find_delayed_extent_range(struct inode *inode,
246 				 ext4_lblk_t lblk, ext4_lblk_t end,
247 				 struct extent_status *es)
248 {
249 	struct ext4_es_tree *tree = NULL;
250 	struct extent_status *es1 = NULL;
251 	struct rb_node *node;
252 
253 	BUG_ON(es == NULL);
254 	BUG_ON(end < lblk);
255 	trace_ext4_es_find_delayed_extent_range_enter(inode, lblk);
256 
257 	read_lock(&EXT4_I(inode)->i_es_lock);
258 	tree = &EXT4_I(inode)->i_es_tree;
259 
260 	/* find extent in cache firstly */
261 	es->es_lblk = es->es_len = es->es_pblk = 0;
262 	if (tree->cache_es) {
263 		es1 = tree->cache_es;
264 		if (in_range(lblk, es1->es_lblk, es1->es_len)) {
265 			es_debug("%u cached by [%u/%u) %llu %x\n",
266 				 lblk, es1->es_lblk, es1->es_len,
267 				 ext4_es_pblock(es1), ext4_es_status(es1));
268 			goto out;
269 		}
270 	}
271 
272 	es1 = __es_tree_search(&tree->root, lblk);
273 
274 out:
275 	if (es1 && !ext4_es_is_delayed(es1)) {
276 		while ((node = rb_next(&es1->rb_node)) != NULL) {
277 			es1 = rb_entry(node, struct extent_status, rb_node);
278 			if (es1->es_lblk > end) {
279 				es1 = NULL;
280 				break;
281 			}
282 			if (ext4_es_is_delayed(es1))
283 				break;
284 		}
285 	}
286 
287 	if (es1 && ext4_es_is_delayed(es1)) {
288 		tree->cache_es = es1;
289 		es->es_lblk = es1->es_lblk;
290 		es->es_len = es1->es_len;
291 		es->es_pblk = es1->es_pblk;
292 	}
293 
294 	read_unlock(&EXT4_I(inode)->i_es_lock);
295 
296 	trace_ext4_es_find_delayed_extent_range_exit(inode, es);
297 }
298 
299 static struct extent_status *
300 ext4_es_alloc_extent(struct inode *inode, ext4_lblk_t lblk, ext4_lblk_t len,
301 		     ext4_fsblk_t pblk)
302 {
303 	struct extent_status *es;
304 	es = kmem_cache_alloc(ext4_es_cachep, GFP_ATOMIC);
305 	if (es == NULL)
306 		return NULL;
307 	es->es_lblk = lblk;
308 	es->es_len = len;
309 	es->es_pblk = pblk;
310 
311 	/*
312 	 * We don't count delayed extent because we never try to reclaim them
313 	 */
314 	if (!ext4_es_is_delayed(es)) {
315 		EXT4_I(inode)->i_es_lru_nr++;
316 		percpu_counter_inc(&EXT4_SB(inode->i_sb)->s_extent_cache_cnt);
317 	}
318 
319 	return es;
320 }
321 
322 static void ext4_es_free_extent(struct inode *inode, struct extent_status *es)
323 {
324 	/* Decrease the lru counter when this es is not delayed */
325 	if (!ext4_es_is_delayed(es)) {
326 		BUG_ON(EXT4_I(inode)->i_es_lru_nr == 0);
327 		EXT4_I(inode)->i_es_lru_nr--;
328 		percpu_counter_dec(&EXT4_SB(inode->i_sb)->s_extent_cache_cnt);
329 	}
330 
331 	kmem_cache_free(ext4_es_cachep, es);
332 }
333 
334 /*
335  * Check whether or not two extents can be merged
336  * Condition:
337  *  - logical block number is contiguous
338  *  - physical block number is contiguous
339  *  - status is equal
340  */
341 static int ext4_es_can_be_merged(struct extent_status *es1,
342 				 struct extent_status *es2)
343 {
344 	if (ext4_es_status(es1) != ext4_es_status(es2))
345 		return 0;
346 
347 	if (((__u64) es1->es_len) + es2->es_len > 0xFFFFFFFFULL)
348 		return 0;
349 
350 	if (((__u64) es1->es_lblk) + es1->es_len != es2->es_lblk)
351 		return 0;
352 
353 	if ((ext4_es_is_written(es1) || ext4_es_is_unwritten(es1)) &&
354 	    (ext4_es_pblock(es1) + es1->es_len == ext4_es_pblock(es2)))
355 		return 1;
356 
357 	if (ext4_es_is_hole(es1))
358 		return 1;
359 
360 	/* we need to check delayed extent is without unwritten status */
361 	if (ext4_es_is_delayed(es1) && !ext4_es_is_unwritten(es1))
362 		return 1;
363 
364 	return 0;
365 }
366 
367 static struct extent_status *
368 ext4_es_try_to_merge_left(struct inode *inode, struct extent_status *es)
369 {
370 	struct ext4_es_tree *tree = &EXT4_I(inode)->i_es_tree;
371 	struct extent_status *es1;
372 	struct rb_node *node;
373 
374 	node = rb_prev(&es->rb_node);
375 	if (!node)
376 		return es;
377 
378 	es1 = rb_entry(node, struct extent_status, rb_node);
379 	if (ext4_es_can_be_merged(es1, es)) {
380 		es1->es_len += es->es_len;
381 		rb_erase(&es->rb_node, &tree->root);
382 		ext4_es_free_extent(inode, es);
383 		es = es1;
384 	}
385 
386 	return es;
387 }
388 
389 static struct extent_status *
390 ext4_es_try_to_merge_right(struct inode *inode, struct extent_status *es)
391 {
392 	struct ext4_es_tree *tree = &EXT4_I(inode)->i_es_tree;
393 	struct extent_status *es1;
394 	struct rb_node *node;
395 
396 	node = rb_next(&es->rb_node);
397 	if (!node)
398 		return es;
399 
400 	es1 = rb_entry(node, struct extent_status, rb_node);
401 	if (ext4_es_can_be_merged(es, es1)) {
402 		es->es_len += es1->es_len;
403 		rb_erase(node, &tree->root);
404 		ext4_es_free_extent(inode, es1);
405 	}
406 
407 	return es;
408 }
409 
410 #ifdef ES_AGGRESSIVE_TEST
411 #include "ext4_extents.h"	/* Needed when ES_AGGRESSIVE_TEST is defined */
412 
413 static void ext4_es_insert_extent_ext_check(struct inode *inode,
414 					    struct extent_status *es)
415 {
416 	struct ext4_ext_path *path = NULL;
417 	struct ext4_extent *ex;
418 	ext4_lblk_t ee_block;
419 	ext4_fsblk_t ee_start;
420 	unsigned short ee_len;
421 	int depth, ee_status, es_status;
422 
423 	path = ext4_ext_find_extent(inode, es->es_lblk, NULL, EXT4_EX_NOCACHE);
424 	if (IS_ERR(path))
425 		return;
426 
427 	depth = ext_depth(inode);
428 	ex = path[depth].p_ext;
429 
430 	if (ex) {
431 
432 		ee_block = le32_to_cpu(ex->ee_block);
433 		ee_start = ext4_ext_pblock(ex);
434 		ee_len = ext4_ext_get_actual_len(ex);
435 
436 		ee_status = ext4_ext_is_uninitialized(ex) ? 1 : 0;
437 		es_status = ext4_es_is_unwritten(es) ? 1 : 0;
438 
439 		/*
440 		 * Make sure ex and es are not overlap when we try to insert
441 		 * a delayed/hole extent.
442 		 */
443 		if (!ext4_es_is_written(es) && !ext4_es_is_unwritten(es)) {
444 			if (in_range(es->es_lblk, ee_block, ee_len)) {
445 				pr_warn("ES insert assertion failed for "
446 					"inode: %lu we can find an extent "
447 					"at block [%d/%d/%llu/%c], but we "
448 					"want to add an delayed/hole extent "
449 					"[%d/%d/%llu/%llx]\n",
450 					inode->i_ino, ee_block, ee_len,
451 					ee_start, ee_status ? 'u' : 'w',
452 					es->es_lblk, es->es_len,
453 					ext4_es_pblock(es), ext4_es_status(es));
454 			}
455 			goto out;
456 		}
457 
458 		/*
459 		 * We don't check ee_block == es->es_lblk, etc. because es
460 		 * might be a part of whole extent, vice versa.
461 		 */
462 		if (es->es_lblk < ee_block ||
463 		    ext4_es_pblock(es) != ee_start + es->es_lblk - ee_block) {
464 			pr_warn("ES insert assertion failed for inode: %lu "
465 				"ex_status [%d/%d/%llu/%c] != "
466 				"es_status [%d/%d/%llu/%c]\n", inode->i_ino,
467 				ee_block, ee_len, ee_start,
468 				ee_status ? 'u' : 'w', es->es_lblk, es->es_len,
469 				ext4_es_pblock(es), es_status ? 'u' : 'w');
470 			goto out;
471 		}
472 
473 		if (ee_status ^ es_status) {
474 			pr_warn("ES insert assertion failed for inode: %lu "
475 				"ex_status [%d/%d/%llu/%c] != "
476 				"es_status [%d/%d/%llu/%c]\n", inode->i_ino,
477 				ee_block, ee_len, ee_start,
478 				ee_status ? 'u' : 'w', es->es_lblk, es->es_len,
479 				ext4_es_pblock(es), es_status ? 'u' : 'w');
480 		}
481 	} else {
482 		/*
483 		 * We can't find an extent on disk.  So we need to make sure
484 		 * that we don't want to add an written/unwritten extent.
485 		 */
486 		if (!ext4_es_is_delayed(es) && !ext4_es_is_hole(es)) {
487 			pr_warn("ES insert assertion failed for inode: %lu "
488 				"can't find an extent at block %d but we want "
489 				"to add an written/unwritten extent "
490 				"[%d/%d/%llu/%llx]\n", inode->i_ino,
491 				es->es_lblk, es->es_lblk, es->es_len,
492 				ext4_es_pblock(es), ext4_es_status(es));
493 		}
494 	}
495 out:
496 	if (path) {
497 		ext4_ext_drop_refs(path);
498 		kfree(path);
499 	}
500 }
501 
502 static void ext4_es_insert_extent_ind_check(struct inode *inode,
503 					    struct extent_status *es)
504 {
505 	struct ext4_map_blocks map;
506 	int retval;
507 
508 	/*
509 	 * Here we call ext4_ind_map_blocks to lookup a block mapping because
510 	 * 'Indirect' structure is defined in indirect.c.  So we couldn't
511 	 * access direct/indirect tree from outside.  It is too dirty to define
512 	 * this function in indirect.c file.
513 	 */
514 
515 	map.m_lblk = es->es_lblk;
516 	map.m_len = es->es_len;
517 
518 	retval = ext4_ind_map_blocks(NULL, inode, &map, 0);
519 	if (retval > 0) {
520 		if (ext4_es_is_delayed(es) || ext4_es_is_hole(es)) {
521 			/*
522 			 * We want to add a delayed/hole extent but this
523 			 * block has been allocated.
524 			 */
525 			pr_warn("ES insert assertion failed for inode: %lu "
526 				"We can find blocks but we want to add a "
527 				"delayed/hole extent [%d/%d/%llu/%llx]\n",
528 				inode->i_ino, es->es_lblk, es->es_len,
529 				ext4_es_pblock(es), ext4_es_status(es));
530 			return;
531 		} else if (ext4_es_is_written(es)) {
532 			if (retval != es->es_len) {
533 				pr_warn("ES insert assertion failed for "
534 					"inode: %lu retval %d != es_len %d\n",
535 					inode->i_ino, retval, es->es_len);
536 				return;
537 			}
538 			if (map.m_pblk != ext4_es_pblock(es)) {
539 				pr_warn("ES insert assertion failed for "
540 					"inode: %lu m_pblk %llu != "
541 					"es_pblk %llu\n",
542 					inode->i_ino, map.m_pblk,
543 					ext4_es_pblock(es));
544 				return;
545 			}
546 		} else {
547 			/*
548 			 * We don't need to check unwritten extent because
549 			 * indirect-based file doesn't have it.
550 			 */
551 			BUG_ON(1);
552 		}
553 	} else if (retval == 0) {
554 		if (ext4_es_is_written(es)) {
555 			pr_warn("ES insert assertion failed for inode: %lu "
556 				"We can't find the block but we want to add "
557 				"an written extent [%d/%d/%llu/%llx]\n",
558 				inode->i_ino, es->es_lblk, es->es_len,
559 				ext4_es_pblock(es), ext4_es_status(es));
560 			return;
561 		}
562 	}
563 }
564 
565 static inline void ext4_es_insert_extent_check(struct inode *inode,
566 					       struct extent_status *es)
567 {
568 	/*
569 	 * We don't need to worry about the race condition because
570 	 * caller takes i_data_sem locking.
571 	 */
572 	BUG_ON(!rwsem_is_locked(&EXT4_I(inode)->i_data_sem));
573 	if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
574 		ext4_es_insert_extent_ext_check(inode, es);
575 	else
576 		ext4_es_insert_extent_ind_check(inode, es);
577 }
578 #else
579 static inline void ext4_es_insert_extent_check(struct inode *inode,
580 					       struct extent_status *es)
581 {
582 }
583 #endif
584 
585 static int __es_insert_extent(struct inode *inode, struct extent_status *newes)
586 {
587 	struct ext4_es_tree *tree = &EXT4_I(inode)->i_es_tree;
588 	struct rb_node **p = &tree->root.rb_node;
589 	struct rb_node *parent = NULL;
590 	struct extent_status *es;
591 
592 	while (*p) {
593 		parent = *p;
594 		es = rb_entry(parent, struct extent_status, rb_node);
595 
596 		if (newes->es_lblk < es->es_lblk) {
597 			if (ext4_es_can_be_merged(newes, es)) {
598 				/*
599 				 * Here we can modify es_lblk directly
600 				 * because it isn't overlapped.
601 				 */
602 				es->es_lblk = newes->es_lblk;
603 				es->es_len += newes->es_len;
604 				if (ext4_es_is_written(es) ||
605 				    ext4_es_is_unwritten(es))
606 					ext4_es_store_pblock(es,
607 							     newes->es_pblk);
608 				es = ext4_es_try_to_merge_left(inode, es);
609 				goto out;
610 			}
611 			p = &(*p)->rb_left;
612 		} else if (newes->es_lblk > ext4_es_end(es)) {
613 			if (ext4_es_can_be_merged(es, newes)) {
614 				es->es_len += newes->es_len;
615 				es = ext4_es_try_to_merge_right(inode, es);
616 				goto out;
617 			}
618 			p = &(*p)->rb_right;
619 		} else {
620 			BUG_ON(1);
621 			return -EINVAL;
622 		}
623 	}
624 
625 	es = ext4_es_alloc_extent(inode, newes->es_lblk, newes->es_len,
626 				  newes->es_pblk);
627 	if (!es)
628 		return -ENOMEM;
629 	rb_link_node(&es->rb_node, parent, p);
630 	rb_insert_color(&es->rb_node, &tree->root);
631 
632 out:
633 	tree->cache_es = es;
634 	return 0;
635 }
636 
637 /*
638  * ext4_es_insert_extent() adds information to an inode's extent
639  * status tree.
640  *
641  * Return 0 on success, error code on failure.
642  */
643 int ext4_es_insert_extent(struct inode *inode, ext4_lblk_t lblk,
644 			  ext4_lblk_t len, ext4_fsblk_t pblk,
645 			  unsigned int status)
646 {
647 	struct extent_status newes;
648 	ext4_lblk_t end = lblk + len - 1;
649 	int err = 0;
650 
651 	es_debug("add [%u/%u) %llu %x to extent status tree of inode %lu\n",
652 		 lblk, len, pblk, status, inode->i_ino);
653 
654 	if (!len)
655 		return 0;
656 
657 	BUG_ON(end < lblk);
658 
659 	newes.es_lblk = lblk;
660 	newes.es_len = len;
661 	ext4_es_store_pblock(&newes, pblk);
662 	ext4_es_store_status(&newes, status);
663 	trace_ext4_es_insert_extent(inode, &newes);
664 
665 	ext4_es_insert_extent_check(inode, &newes);
666 
667 	write_lock(&EXT4_I(inode)->i_es_lock);
668 	err = __es_remove_extent(inode, lblk, end);
669 	if (err != 0)
670 		goto error;
671 retry:
672 	err = __es_insert_extent(inode, &newes);
673 	if (err == -ENOMEM && __ext4_es_shrink(EXT4_SB(inode->i_sb), 1,
674 					       EXT4_I(inode)))
675 		goto retry;
676 	if (err == -ENOMEM && !ext4_es_is_delayed(&newes))
677 		err = 0;
678 
679 error:
680 	write_unlock(&EXT4_I(inode)->i_es_lock);
681 
682 	ext4_es_print_tree(inode);
683 
684 	return err;
685 }
686 
687 /*
688  * ext4_es_cache_extent() inserts information into the extent status
689  * tree if and only if there isn't information about the range in
690  * question already.
691  */
692 void ext4_es_cache_extent(struct inode *inode, ext4_lblk_t lblk,
693 			  ext4_lblk_t len, ext4_fsblk_t pblk,
694 			  unsigned int status)
695 {
696 	struct extent_status *es;
697 	struct extent_status newes;
698 	ext4_lblk_t end = lblk + len - 1;
699 
700 	newes.es_lblk = lblk;
701 	newes.es_len = len;
702 	ext4_es_store_pblock(&newes, pblk);
703 	ext4_es_store_status(&newes, status);
704 	trace_ext4_es_cache_extent(inode, &newes);
705 
706 	if (!len)
707 		return;
708 
709 	BUG_ON(end < lblk);
710 
711 	write_lock(&EXT4_I(inode)->i_es_lock);
712 
713 	es = __es_tree_search(&EXT4_I(inode)->i_es_tree.root, lblk);
714 	if (!es || es->es_lblk > end)
715 		__es_insert_extent(inode, &newes);
716 	write_unlock(&EXT4_I(inode)->i_es_lock);
717 }
718 
719 /*
720  * ext4_es_lookup_extent() looks up an extent in extent status tree.
721  *
722  * ext4_es_lookup_extent is called by ext4_map_blocks/ext4_da_map_blocks.
723  *
724  * Return: 1 on found, 0 on not
725  */
726 int ext4_es_lookup_extent(struct inode *inode, ext4_lblk_t lblk,
727 			  struct extent_status *es)
728 {
729 	struct ext4_es_tree *tree;
730 	struct extent_status *es1 = NULL;
731 	struct rb_node *node;
732 	int found = 0;
733 
734 	trace_ext4_es_lookup_extent_enter(inode, lblk);
735 	es_debug("lookup extent in block %u\n", lblk);
736 
737 	tree = &EXT4_I(inode)->i_es_tree;
738 	read_lock(&EXT4_I(inode)->i_es_lock);
739 
740 	/* find extent in cache firstly */
741 	es->es_lblk = es->es_len = es->es_pblk = 0;
742 	if (tree->cache_es) {
743 		es1 = tree->cache_es;
744 		if (in_range(lblk, es1->es_lblk, es1->es_len)) {
745 			es_debug("%u cached by [%u/%u)\n",
746 				 lblk, es1->es_lblk, es1->es_len);
747 			found = 1;
748 			goto out;
749 		}
750 	}
751 
752 	node = tree->root.rb_node;
753 	while (node) {
754 		es1 = rb_entry(node, struct extent_status, rb_node);
755 		if (lblk < es1->es_lblk)
756 			node = node->rb_left;
757 		else if (lblk > ext4_es_end(es1))
758 			node = node->rb_right;
759 		else {
760 			found = 1;
761 			break;
762 		}
763 	}
764 
765 out:
766 	if (found) {
767 		BUG_ON(!es1);
768 		es->es_lblk = es1->es_lblk;
769 		es->es_len = es1->es_len;
770 		es->es_pblk = es1->es_pblk;
771 	}
772 
773 	read_unlock(&EXT4_I(inode)->i_es_lock);
774 
775 	trace_ext4_es_lookup_extent_exit(inode, es, found);
776 	return found;
777 }
778 
779 static int __es_remove_extent(struct inode *inode, ext4_lblk_t lblk,
780 			      ext4_lblk_t end)
781 {
782 	struct ext4_es_tree *tree = &EXT4_I(inode)->i_es_tree;
783 	struct rb_node *node;
784 	struct extent_status *es;
785 	struct extent_status orig_es;
786 	ext4_lblk_t len1, len2;
787 	ext4_fsblk_t block;
788 	int err;
789 
790 retry:
791 	err = 0;
792 	es = __es_tree_search(&tree->root, lblk);
793 	if (!es)
794 		goto out;
795 	if (es->es_lblk > end)
796 		goto out;
797 
798 	/* Simply invalidate cache_es. */
799 	tree->cache_es = NULL;
800 
801 	orig_es.es_lblk = es->es_lblk;
802 	orig_es.es_len = es->es_len;
803 	orig_es.es_pblk = es->es_pblk;
804 
805 	len1 = lblk > es->es_lblk ? lblk - es->es_lblk : 0;
806 	len2 = ext4_es_end(es) > end ? ext4_es_end(es) - end : 0;
807 	if (len1 > 0)
808 		es->es_len = len1;
809 	if (len2 > 0) {
810 		if (len1 > 0) {
811 			struct extent_status newes;
812 
813 			newes.es_lblk = end + 1;
814 			newes.es_len = len2;
815 			if (ext4_es_is_written(&orig_es) ||
816 			    ext4_es_is_unwritten(&orig_es)) {
817 				block = ext4_es_pblock(&orig_es) +
818 					orig_es.es_len - len2;
819 				ext4_es_store_pblock(&newes, block);
820 			}
821 			ext4_es_store_status(&newes, ext4_es_status(&orig_es));
822 			err = __es_insert_extent(inode, &newes);
823 			if (err) {
824 				es->es_lblk = orig_es.es_lblk;
825 				es->es_len = orig_es.es_len;
826 				if ((err == -ENOMEM) &&
827 				    __ext4_es_shrink(EXT4_SB(inode->i_sb), 1,
828 						     EXT4_I(inode)))
829 					goto retry;
830 				goto out;
831 			}
832 		} else {
833 			es->es_lblk = end + 1;
834 			es->es_len = len2;
835 			if (ext4_es_is_written(es) ||
836 			    ext4_es_is_unwritten(es)) {
837 				block = orig_es.es_pblk + orig_es.es_len - len2;
838 				ext4_es_store_pblock(es, block);
839 			}
840 		}
841 		goto out;
842 	}
843 
844 	if (len1 > 0) {
845 		node = rb_next(&es->rb_node);
846 		if (node)
847 			es = rb_entry(node, struct extent_status, rb_node);
848 		else
849 			es = NULL;
850 	}
851 
852 	while (es && ext4_es_end(es) <= end) {
853 		node = rb_next(&es->rb_node);
854 		rb_erase(&es->rb_node, &tree->root);
855 		ext4_es_free_extent(inode, es);
856 		if (!node) {
857 			es = NULL;
858 			break;
859 		}
860 		es = rb_entry(node, struct extent_status, rb_node);
861 	}
862 
863 	if (es && es->es_lblk < end + 1) {
864 		ext4_lblk_t orig_len = es->es_len;
865 
866 		len1 = ext4_es_end(es) - end;
867 		es->es_lblk = end + 1;
868 		es->es_len = len1;
869 		if (ext4_es_is_written(es) || ext4_es_is_unwritten(es)) {
870 			block = es->es_pblk + orig_len - len1;
871 			ext4_es_store_pblock(es, block);
872 		}
873 	}
874 
875 out:
876 	return err;
877 }
878 
879 /*
880  * ext4_es_remove_extent() removes a space from a extent status tree.
881  *
882  * Return 0 on success, error code on failure.
883  */
884 int ext4_es_remove_extent(struct inode *inode, ext4_lblk_t lblk,
885 			  ext4_lblk_t len)
886 {
887 	ext4_lblk_t end;
888 	int err = 0;
889 
890 	trace_ext4_es_remove_extent(inode, lblk, len);
891 	es_debug("remove [%u/%u) from extent status tree of inode %lu\n",
892 		 lblk, len, inode->i_ino);
893 
894 	if (!len)
895 		return err;
896 
897 	end = lblk + len - 1;
898 	BUG_ON(end < lblk);
899 
900 	write_lock(&EXT4_I(inode)->i_es_lock);
901 	err = __es_remove_extent(inode, lblk, end);
902 	write_unlock(&EXT4_I(inode)->i_es_lock);
903 	ext4_es_print_tree(inode);
904 	return err;
905 }
906 
907 static int ext4_inode_touch_time_cmp(void *priv, struct list_head *a,
908 				     struct list_head *b)
909 {
910 	struct ext4_inode_info *eia, *eib;
911 	eia = list_entry(a, struct ext4_inode_info, i_es_lru);
912 	eib = list_entry(b, struct ext4_inode_info, i_es_lru);
913 
914 	if (ext4_test_inode_state(&eia->vfs_inode, EXT4_STATE_EXT_PRECACHED) &&
915 	    !ext4_test_inode_state(&eib->vfs_inode, EXT4_STATE_EXT_PRECACHED))
916 		return 1;
917 	if (!ext4_test_inode_state(&eia->vfs_inode, EXT4_STATE_EXT_PRECACHED) &&
918 	    ext4_test_inode_state(&eib->vfs_inode, EXT4_STATE_EXT_PRECACHED))
919 		return -1;
920 	if (eia->i_touch_when == eib->i_touch_when)
921 		return 0;
922 	if (time_after(eia->i_touch_when, eib->i_touch_when))
923 		return 1;
924 	else
925 		return -1;
926 }
927 
928 static int __ext4_es_shrink(struct ext4_sb_info *sbi, int nr_to_scan,
929 			    struct ext4_inode_info *locked_ei)
930 {
931 	struct ext4_inode_info *ei;
932 	struct list_head *cur, *tmp;
933 	LIST_HEAD(skipped);
934 	int nr_shrunk = 0;
935 	int retried = 0, skip_precached = 1, nr_skipped = 0;
936 
937 	spin_lock(&sbi->s_es_lru_lock);
938 
939 retry:
940 	list_for_each_safe(cur, tmp, &sbi->s_es_lru) {
941 		int shrunk;
942 
943 		/*
944 		 * If we have already reclaimed all extents from extent
945 		 * status tree, just stop the loop immediately.
946 		 */
947 		if (percpu_counter_read_positive(&sbi->s_extent_cache_cnt) == 0)
948 			break;
949 
950 		ei = list_entry(cur, struct ext4_inode_info, i_es_lru);
951 
952 		/*
953 		 * Skip the inode that is newer than the last_sorted
954 		 * time.  Normally we try hard to avoid shrinking
955 		 * precached inodes, but we will as a last resort.
956 		 */
957 		if ((sbi->s_es_last_sorted < ei->i_touch_when) ||
958 		    (skip_precached && ext4_test_inode_state(&ei->vfs_inode,
959 						EXT4_STATE_EXT_PRECACHED))) {
960 			nr_skipped++;
961 			list_move_tail(cur, &skipped);
962 			continue;
963 		}
964 
965 		if (ei->i_es_lru_nr == 0 || ei == locked_ei)
966 			continue;
967 
968 		write_lock(&ei->i_es_lock);
969 		shrunk = __es_try_to_reclaim_extents(ei, nr_to_scan);
970 		if (ei->i_es_lru_nr == 0)
971 			list_del_init(&ei->i_es_lru);
972 		write_unlock(&ei->i_es_lock);
973 
974 		nr_shrunk += shrunk;
975 		nr_to_scan -= shrunk;
976 		if (nr_to_scan == 0)
977 			break;
978 	}
979 
980 	/* Move the newer inodes into the tail of the LRU list. */
981 	list_splice_tail(&skipped, &sbi->s_es_lru);
982 	INIT_LIST_HEAD(&skipped);
983 
984 	/*
985 	 * If we skipped any inodes, and we weren't able to make any
986 	 * forward progress, sort the list and try again.
987 	 */
988 	if ((nr_shrunk == 0) && nr_skipped && !retried) {
989 		retried++;
990 		list_sort(NULL, &sbi->s_es_lru, ext4_inode_touch_time_cmp);
991 		sbi->s_es_last_sorted = jiffies;
992 		ei = list_first_entry(&sbi->s_es_lru, struct ext4_inode_info,
993 				      i_es_lru);
994 		/*
995 		 * If there are no non-precached inodes left on the
996 		 * list, start releasing precached extents.
997 		 */
998 		if (ext4_test_inode_state(&ei->vfs_inode,
999 					  EXT4_STATE_EXT_PRECACHED))
1000 			skip_precached = 0;
1001 		goto retry;
1002 	}
1003 
1004 	spin_unlock(&sbi->s_es_lru_lock);
1005 
1006 	if (locked_ei && nr_shrunk == 0)
1007 		nr_shrunk = __es_try_to_reclaim_extents(locked_ei, nr_to_scan);
1008 
1009 	return nr_shrunk;
1010 }
1011 
1012 static unsigned long ext4_es_count(struct shrinker *shrink,
1013 				   struct shrink_control *sc)
1014 {
1015 	unsigned long nr;
1016 	struct ext4_sb_info *sbi;
1017 
1018 	sbi = container_of(shrink, struct ext4_sb_info, s_es_shrinker);
1019 	nr = percpu_counter_read_positive(&sbi->s_extent_cache_cnt);
1020 	trace_ext4_es_shrink_enter(sbi->s_sb, sc->nr_to_scan, nr);
1021 	return nr;
1022 }
1023 
1024 static unsigned long ext4_es_scan(struct shrinker *shrink,
1025 				  struct shrink_control *sc)
1026 {
1027 	struct ext4_sb_info *sbi = container_of(shrink,
1028 					struct ext4_sb_info, s_es_shrinker);
1029 	int nr_to_scan = sc->nr_to_scan;
1030 	int ret, nr_shrunk;
1031 
1032 	ret = percpu_counter_read_positive(&sbi->s_extent_cache_cnt);
1033 	trace_ext4_es_shrink_enter(sbi->s_sb, nr_to_scan, ret);
1034 
1035 	if (!nr_to_scan)
1036 		return ret;
1037 
1038 	nr_shrunk = __ext4_es_shrink(sbi, nr_to_scan, NULL);
1039 
1040 	trace_ext4_es_shrink_exit(sbi->s_sb, nr_shrunk, ret);
1041 	return nr_shrunk;
1042 }
1043 
1044 void ext4_es_register_shrinker(struct ext4_sb_info *sbi)
1045 {
1046 	INIT_LIST_HEAD(&sbi->s_es_lru);
1047 	spin_lock_init(&sbi->s_es_lru_lock);
1048 	sbi->s_es_last_sorted = 0;
1049 	sbi->s_es_shrinker.scan_objects = ext4_es_scan;
1050 	sbi->s_es_shrinker.count_objects = ext4_es_count;
1051 	sbi->s_es_shrinker.seeks = DEFAULT_SEEKS;
1052 	register_shrinker(&sbi->s_es_shrinker);
1053 }
1054 
1055 void ext4_es_unregister_shrinker(struct ext4_sb_info *sbi)
1056 {
1057 	unregister_shrinker(&sbi->s_es_shrinker);
1058 }
1059 
1060 void ext4_es_lru_add(struct inode *inode)
1061 {
1062 	struct ext4_inode_info *ei = EXT4_I(inode);
1063 	struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1064 
1065 	ei->i_touch_when = jiffies;
1066 
1067 	if (!list_empty(&ei->i_es_lru))
1068 		return;
1069 
1070 	spin_lock(&sbi->s_es_lru_lock);
1071 	if (list_empty(&ei->i_es_lru))
1072 		list_add_tail(&ei->i_es_lru, &sbi->s_es_lru);
1073 	spin_unlock(&sbi->s_es_lru_lock);
1074 }
1075 
1076 void ext4_es_lru_del(struct inode *inode)
1077 {
1078 	struct ext4_inode_info *ei = EXT4_I(inode);
1079 	struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1080 
1081 	spin_lock(&sbi->s_es_lru_lock);
1082 	if (!list_empty(&ei->i_es_lru))
1083 		list_del_init(&ei->i_es_lru);
1084 	spin_unlock(&sbi->s_es_lru_lock);
1085 }
1086 
1087 static int __es_try_to_reclaim_extents(struct ext4_inode_info *ei,
1088 				       int nr_to_scan)
1089 {
1090 	struct inode *inode = &ei->vfs_inode;
1091 	struct ext4_es_tree *tree = &ei->i_es_tree;
1092 	struct rb_node *node;
1093 	struct extent_status *es;
1094 	unsigned long nr_shrunk = 0;
1095 	static DEFINE_RATELIMIT_STATE(_rs, DEFAULT_RATELIMIT_INTERVAL,
1096 				      DEFAULT_RATELIMIT_BURST);
1097 
1098 	if (ei->i_es_lru_nr == 0)
1099 		return 0;
1100 
1101 	if (ext4_test_inode_state(inode, EXT4_STATE_EXT_PRECACHED) &&
1102 	    __ratelimit(&_rs))
1103 		ext4_warning(inode->i_sb, "forced shrink of precached extents");
1104 
1105 	node = rb_first(&tree->root);
1106 	while (node != NULL) {
1107 		es = rb_entry(node, struct extent_status, rb_node);
1108 		node = rb_next(&es->rb_node);
1109 		/*
1110 		 * We can't reclaim delayed extent from status tree because
1111 		 * fiemap, bigallic, and seek_data/hole need to use it.
1112 		 */
1113 		if (!ext4_es_is_delayed(es)) {
1114 			rb_erase(&es->rb_node, &tree->root);
1115 			ext4_es_free_extent(inode, es);
1116 			nr_shrunk++;
1117 			if (--nr_to_scan == 0)
1118 				break;
1119 		}
1120 	}
1121 	tree->cache_es = NULL;
1122 	return nr_shrunk;
1123 }
1124