Loading...
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/list_sort.h>
13#include <linux/proc_fs.h>
14#include <linux/seq_file.h>
15#include "ext4.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
143static struct kmem_cache *ext4_es_cachep;
144
145static int __es_insert_extent(struct inode *inode, struct extent_status *newes);
146static int __es_remove_extent(struct inode *inode, ext4_lblk_t lblk,
147 ext4_lblk_t end);
148static int es_reclaim_extents(struct ext4_inode_info *ei, int *nr_to_scan);
149static int __es_shrink(struct ext4_sb_info *sbi, int nr_to_scan,
150 struct ext4_inode_info *locked_ei);
151
152int __init ext4_init_es(void)
153{
154 ext4_es_cachep = kmem_cache_create("ext4_extent_status",
155 sizeof(struct extent_status),
156 0, (SLAB_RECLAIM_ACCOUNT), NULL);
157 if (ext4_es_cachep == NULL)
158 return -ENOMEM;
159 return 0;
160}
161
162void ext4_exit_es(void)
163{
164 if (ext4_es_cachep)
165 kmem_cache_destroy(ext4_es_cachep);
166}
167
168void ext4_es_init_tree(struct ext4_es_tree *tree)
169{
170 tree->root = RB_ROOT;
171 tree->cache_es = NULL;
172}
173
174#ifdef ES_DEBUG__
175static void ext4_es_print_tree(struct inode *inode)
176{
177 struct ext4_es_tree *tree;
178 struct rb_node *node;
179
180 printk(KERN_DEBUG "status extents for inode %lu:", inode->i_ino);
181 tree = &EXT4_I(inode)->i_es_tree;
182 node = rb_first(&tree->root);
183 while (node) {
184 struct extent_status *es;
185 es = rb_entry(node, struct extent_status, rb_node);
186 printk(KERN_DEBUG " [%u/%u) %llu %x",
187 es->es_lblk, es->es_len,
188 ext4_es_pblock(es), ext4_es_status(es));
189 node = rb_next(node);
190 }
191 printk(KERN_DEBUG "\n");
192}
193#else
194#define ext4_es_print_tree(inode)
195#endif
196
197static inline ext4_lblk_t ext4_es_end(struct extent_status *es)
198{
199 BUG_ON(es->es_lblk + es->es_len < es->es_lblk);
200 return es->es_lblk + es->es_len - 1;
201}
202
203/*
204 * search through the tree for an delayed extent with a given offset. If
205 * it can't be found, try to find next extent.
206 */
207static struct extent_status *__es_tree_search(struct rb_root *root,
208 ext4_lblk_t lblk)
209{
210 struct rb_node *node = root->rb_node;
211 struct extent_status *es = NULL;
212
213 while (node) {
214 es = rb_entry(node, struct extent_status, rb_node);
215 if (lblk < es->es_lblk)
216 node = node->rb_left;
217 else if (lblk > ext4_es_end(es))
218 node = node->rb_right;
219 else
220 return es;
221 }
222
223 if (es && lblk < es->es_lblk)
224 return es;
225
226 if (es && lblk > ext4_es_end(es)) {
227 node = rb_next(&es->rb_node);
228 return node ? rb_entry(node, struct extent_status, rb_node) :
229 NULL;
230 }
231
232 return NULL;
233}
234
235/*
236 * ext4_es_find_delayed_extent_range: find the 1st delayed extent covering
237 * @es->lblk if it exists, otherwise, the next extent after @es->lblk.
238 *
239 * @inode: the inode which owns delayed extents
240 * @lblk: the offset where we start to search
241 * @end: the offset where we stop to search
242 * @es: delayed extent that we found
243 */
244void ext4_es_find_delayed_extent_range(struct inode *inode,
245 ext4_lblk_t lblk, ext4_lblk_t end,
246 struct extent_status *es)
247{
248 struct ext4_es_tree *tree = NULL;
249 struct extent_status *es1 = NULL;
250 struct rb_node *node;
251
252 BUG_ON(es == NULL);
253 BUG_ON(end < lblk);
254 trace_ext4_es_find_delayed_extent_range_enter(inode, lblk);
255
256 read_lock(&EXT4_I(inode)->i_es_lock);
257 tree = &EXT4_I(inode)->i_es_tree;
258
259 /* find extent in cache firstly */
260 es->es_lblk = es->es_len = es->es_pblk = 0;
261 if (tree->cache_es) {
262 es1 = tree->cache_es;
263 if (in_range(lblk, es1->es_lblk, es1->es_len)) {
264 es_debug("%u cached by [%u/%u) %llu %x\n",
265 lblk, es1->es_lblk, es1->es_len,
266 ext4_es_pblock(es1), ext4_es_status(es1));
267 goto out;
268 }
269 }
270
271 es1 = __es_tree_search(&tree->root, lblk);
272
273out:
274 if (es1 && !ext4_es_is_delayed(es1)) {
275 while ((node = rb_next(&es1->rb_node)) != NULL) {
276 es1 = rb_entry(node, struct extent_status, rb_node);
277 if (es1->es_lblk > end) {
278 es1 = NULL;
279 break;
280 }
281 if (ext4_es_is_delayed(es1))
282 break;
283 }
284 }
285
286 if (es1 && ext4_es_is_delayed(es1)) {
287 tree->cache_es = es1;
288 es->es_lblk = es1->es_lblk;
289 es->es_len = es1->es_len;
290 es->es_pblk = es1->es_pblk;
291 }
292
293 read_unlock(&EXT4_I(inode)->i_es_lock);
294
295 trace_ext4_es_find_delayed_extent_range_exit(inode, es);
296}
297
298static void ext4_es_list_add(struct inode *inode)
299{
300 struct ext4_inode_info *ei = EXT4_I(inode);
301 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
302
303 if (!list_empty(&ei->i_es_list))
304 return;
305
306 spin_lock(&sbi->s_es_lock);
307 if (list_empty(&ei->i_es_list)) {
308 list_add_tail(&ei->i_es_list, &sbi->s_es_list);
309 sbi->s_es_nr_inode++;
310 }
311 spin_unlock(&sbi->s_es_lock);
312}
313
314static void ext4_es_list_del(struct inode *inode)
315{
316 struct ext4_inode_info *ei = EXT4_I(inode);
317 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
318
319 spin_lock(&sbi->s_es_lock);
320 if (!list_empty(&ei->i_es_list)) {
321 list_del_init(&ei->i_es_list);
322 sbi->s_es_nr_inode--;
323 WARN_ON_ONCE(sbi->s_es_nr_inode < 0);
324 }
325 spin_unlock(&sbi->s_es_lock);
326}
327
328static struct extent_status *
329ext4_es_alloc_extent(struct inode *inode, ext4_lblk_t lblk, ext4_lblk_t len,
330 ext4_fsblk_t pblk)
331{
332 struct extent_status *es;
333 es = kmem_cache_alloc(ext4_es_cachep, GFP_ATOMIC);
334 if (es == NULL)
335 return NULL;
336 es->es_lblk = lblk;
337 es->es_len = len;
338 es->es_pblk = pblk;
339
340 /*
341 * We don't count delayed extent because we never try to reclaim them
342 */
343 if (!ext4_es_is_delayed(es)) {
344 if (!EXT4_I(inode)->i_es_shk_nr++)
345 ext4_es_list_add(inode);
346 percpu_counter_inc(&EXT4_SB(inode->i_sb)->
347 s_es_stats.es_stats_shk_cnt);
348 }
349
350 EXT4_I(inode)->i_es_all_nr++;
351 percpu_counter_inc(&EXT4_SB(inode->i_sb)->s_es_stats.es_stats_all_cnt);
352
353 return es;
354}
355
356static void ext4_es_free_extent(struct inode *inode, struct extent_status *es)
357{
358 EXT4_I(inode)->i_es_all_nr--;
359 percpu_counter_dec(&EXT4_SB(inode->i_sb)->s_es_stats.es_stats_all_cnt);
360
361 /* Decrease the shrink counter when this es is not delayed */
362 if (!ext4_es_is_delayed(es)) {
363 BUG_ON(EXT4_I(inode)->i_es_shk_nr == 0);
364 if (!--EXT4_I(inode)->i_es_shk_nr)
365 ext4_es_list_del(inode);
366 percpu_counter_dec(&EXT4_SB(inode->i_sb)->
367 s_es_stats.es_stats_shk_cnt);
368 }
369
370 kmem_cache_free(ext4_es_cachep, es);
371}
372
373/*
374 * Check whether or not two extents can be merged
375 * Condition:
376 * - logical block number is contiguous
377 * - physical block number is contiguous
378 * - status is equal
379 */
380static int ext4_es_can_be_merged(struct extent_status *es1,
381 struct extent_status *es2)
382{
383 if (ext4_es_type(es1) != ext4_es_type(es2))
384 return 0;
385
386 if (((__u64) es1->es_len) + es2->es_len > EXT_MAX_BLOCKS) {
387 pr_warn("ES assertion failed when merging extents. "
388 "The sum of lengths of es1 (%d) and es2 (%d) "
389 "is bigger than allowed file size (%d)\n",
390 es1->es_len, es2->es_len, EXT_MAX_BLOCKS);
391 WARN_ON(1);
392 return 0;
393 }
394
395 if (((__u64) es1->es_lblk) + es1->es_len != es2->es_lblk)
396 return 0;
397
398 if ((ext4_es_is_written(es1) || ext4_es_is_unwritten(es1)) &&
399 (ext4_es_pblock(es1) + es1->es_len == ext4_es_pblock(es2)))
400 return 1;
401
402 if (ext4_es_is_hole(es1))
403 return 1;
404
405 /* we need to check delayed extent is without unwritten status */
406 if (ext4_es_is_delayed(es1) && !ext4_es_is_unwritten(es1))
407 return 1;
408
409 return 0;
410}
411
412static struct extent_status *
413ext4_es_try_to_merge_left(struct inode *inode, struct extent_status *es)
414{
415 struct ext4_es_tree *tree = &EXT4_I(inode)->i_es_tree;
416 struct extent_status *es1;
417 struct rb_node *node;
418
419 node = rb_prev(&es->rb_node);
420 if (!node)
421 return es;
422
423 es1 = rb_entry(node, struct extent_status, rb_node);
424 if (ext4_es_can_be_merged(es1, es)) {
425 es1->es_len += es->es_len;
426 if (ext4_es_is_referenced(es))
427 ext4_es_set_referenced(es1);
428 rb_erase(&es->rb_node, &tree->root);
429 ext4_es_free_extent(inode, es);
430 es = es1;
431 }
432
433 return es;
434}
435
436static struct extent_status *
437ext4_es_try_to_merge_right(struct inode *inode, struct extent_status *es)
438{
439 struct ext4_es_tree *tree = &EXT4_I(inode)->i_es_tree;
440 struct extent_status *es1;
441 struct rb_node *node;
442
443 node = rb_next(&es->rb_node);
444 if (!node)
445 return es;
446
447 es1 = rb_entry(node, struct extent_status, rb_node);
448 if (ext4_es_can_be_merged(es, es1)) {
449 es->es_len += es1->es_len;
450 if (ext4_es_is_referenced(es1))
451 ext4_es_set_referenced(es);
452 rb_erase(node, &tree->root);
453 ext4_es_free_extent(inode, es1);
454 }
455
456 return es;
457}
458
459#ifdef ES_AGGRESSIVE_TEST
460#include "ext4_extents.h" /* Needed when ES_AGGRESSIVE_TEST is defined */
461
462static void ext4_es_insert_extent_ext_check(struct inode *inode,
463 struct extent_status *es)
464{
465 struct ext4_ext_path *path = NULL;
466 struct ext4_extent *ex;
467 ext4_lblk_t ee_block;
468 ext4_fsblk_t ee_start;
469 unsigned short ee_len;
470 int depth, ee_status, es_status;
471
472 path = ext4_find_extent(inode, es->es_lblk, NULL, EXT4_EX_NOCACHE);
473 if (IS_ERR(path))
474 return;
475
476 depth = ext_depth(inode);
477 ex = path[depth].p_ext;
478
479 if (ex) {
480
481 ee_block = le32_to_cpu(ex->ee_block);
482 ee_start = ext4_ext_pblock(ex);
483 ee_len = ext4_ext_get_actual_len(ex);
484
485 ee_status = ext4_ext_is_unwritten(ex) ? 1 : 0;
486 es_status = ext4_es_is_unwritten(es) ? 1 : 0;
487
488 /*
489 * Make sure ex and es are not overlap when we try to insert
490 * a delayed/hole extent.
491 */
492 if (!ext4_es_is_written(es) && !ext4_es_is_unwritten(es)) {
493 if (in_range(es->es_lblk, ee_block, ee_len)) {
494 pr_warn("ES insert assertion failed for "
495 "inode: %lu we can find an extent "
496 "at block [%d/%d/%llu/%c], but we "
497 "want to add a delayed/hole extent "
498 "[%d/%d/%llu/%x]\n",
499 inode->i_ino, ee_block, ee_len,
500 ee_start, ee_status ? 'u' : 'w',
501 es->es_lblk, es->es_len,
502 ext4_es_pblock(es), ext4_es_status(es));
503 }
504 goto out;
505 }
506
507 /*
508 * We don't check ee_block == es->es_lblk, etc. because es
509 * might be a part of whole extent, vice versa.
510 */
511 if (es->es_lblk < ee_block ||
512 ext4_es_pblock(es) != ee_start + es->es_lblk - ee_block) {
513 pr_warn("ES insert assertion failed for inode: %lu "
514 "ex_status [%d/%d/%llu/%c] != "
515 "es_status [%d/%d/%llu/%c]\n", inode->i_ino,
516 ee_block, ee_len, ee_start,
517 ee_status ? 'u' : 'w', es->es_lblk, es->es_len,
518 ext4_es_pblock(es), es_status ? 'u' : 'w');
519 goto out;
520 }
521
522 if (ee_status ^ es_status) {
523 pr_warn("ES insert assertion failed for inode: %lu "
524 "ex_status [%d/%d/%llu/%c] != "
525 "es_status [%d/%d/%llu/%c]\n", inode->i_ino,
526 ee_block, ee_len, ee_start,
527 ee_status ? 'u' : 'w', es->es_lblk, es->es_len,
528 ext4_es_pblock(es), es_status ? 'u' : 'w');
529 }
530 } else {
531 /*
532 * We can't find an extent on disk. So we need to make sure
533 * that we don't want to add an written/unwritten extent.
534 */
535 if (!ext4_es_is_delayed(es) && !ext4_es_is_hole(es)) {
536 pr_warn("ES insert assertion failed for inode: %lu "
537 "can't find an extent at block %d but we want "
538 "to add a written/unwritten extent "
539 "[%d/%d/%llu/%x]\n", inode->i_ino,
540 es->es_lblk, es->es_lblk, es->es_len,
541 ext4_es_pblock(es), ext4_es_status(es));
542 }
543 }
544out:
545 ext4_ext_drop_refs(path);
546 kfree(path);
547}
548
549static void ext4_es_insert_extent_ind_check(struct inode *inode,
550 struct extent_status *es)
551{
552 struct ext4_map_blocks map;
553 int retval;
554
555 /*
556 * Here we call ext4_ind_map_blocks to lookup a block mapping because
557 * 'Indirect' structure is defined in indirect.c. So we couldn't
558 * access direct/indirect tree from outside. It is too dirty to define
559 * this function in indirect.c file.
560 */
561
562 map.m_lblk = es->es_lblk;
563 map.m_len = es->es_len;
564
565 retval = ext4_ind_map_blocks(NULL, inode, &map, 0);
566 if (retval > 0) {
567 if (ext4_es_is_delayed(es) || ext4_es_is_hole(es)) {
568 /*
569 * We want to add a delayed/hole extent but this
570 * block has been allocated.
571 */
572 pr_warn("ES insert assertion failed for inode: %lu "
573 "We can find blocks but we want to add a "
574 "delayed/hole extent [%d/%d/%llu/%x]\n",
575 inode->i_ino, es->es_lblk, es->es_len,
576 ext4_es_pblock(es), ext4_es_status(es));
577 return;
578 } else if (ext4_es_is_written(es)) {
579 if (retval != es->es_len) {
580 pr_warn("ES insert assertion failed for "
581 "inode: %lu retval %d != es_len %d\n",
582 inode->i_ino, retval, es->es_len);
583 return;
584 }
585 if (map.m_pblk != ext4_es_pblock(es)) {
586 pr_warn("ES insert assertion failed for "
587 "inode: %lu m_pblk %llu != "
588 "es_pblk %llu\n",
589 inode->i_ino, map.m_pblk,
590 ext4_es_pblock(es));
591 return;
592 }
593 } else {
594 /*
595 * We don't need to check unwritten extent because
596 * indirect-based file doesn't have it.
597 */
598 BUG_ON(1);
599 }
600 } else if (retval == 0) {
601 if (ext4_es_is_written(es)) {
602 pr_warn("ES insert assertion failed for inode: %lu "
603 "We can't find the block but we want to add "
604 "a written extent [%d/%d/%llu/%x]\n",
605 inode->i_ino, es->es_lblk, es->es_len,
606 ext4_es_pblock(es), ext4_es_status(es));
607 return;
608 }
609 }
610}
611
612static inline void ext4_es_insert_extent_check(struct inode *inode,
613 struct extent_status *es)
614{
615 /*
616 * We don't need to worry about the race condition because
617 * caller takes i_data_sem locking.
618 */
619 BUG_ON(!rwsem_is_locked(&EXT4_I(inode)->i_data_sem));
620 if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
621 ext4_es_insert_extent_ext_check(inode, es);
622 else
623 ext4_es_insert_extent_ind_check(inode, es);
624}
625#else
626static inline void ext4_es_insert_extent_check(struct inode *inode,
627 struct extent_status *es)
628{
629}
630#endif
631
632static int __es_insert_extent(struct inode *inode, struct extent_status *newes)
633{
634 struct ext4_es_tree *tree = &EXT4_I(inode)->i_es_tree;
635 struct rb_node **p = &tree->root.rb_node;
636 struct rb_node *parent = NULL;
637 struct extent_status *es;
638
639 while (*p) {
640 parent = *p;
641 es = rb_entry(parent, struct extent_status, rb_node);
642
643 if (newes->es_lblk < es->es_lblk) {
644 if (ext4_es_can_be_merged(newes, es)) {
645 /*
646 * Here we can modify es_lblk directly
647 * because it isn't overlapped.
648 */
649 es->es_lblk = newes->es_lblk;
650 es->es_len += newes->es_len;
651 if (ext4_es_is_written(es) ||
652 ext4_es_is_unwritten(es))
653 ext4_es_store_pblock(es,
654 newes->es_pblk);
655 es = ext4_es_try_to_merge_left(inode, es);
656 goto out;
657 }
658 p = &(*p)->rb_left;
659 } else if (newes->es_lblk > ext4_es_end(es)) {
660 if (ext4_es_can_be_merged(es, newes)) {
661 es->es_len += newes->es_len;
662 es = ext4_es_try_to_merge_right(inode, es);
663 goto out;
664 }
665 p = &(*p)->rb_right;
666 } else {
667 BUG_ON(1);
668 return -EINVAL;
669 }
670 }
671
672 es = ext4_es_alloc_extent(inode, newes->es_lblk, newes->es_len,
673 newes->es_pblk);
674 if (!es)
675 return -ENOMEM;
676 rb_link_node(&es->rb_node, parent, p);
677 rb_insert_color(&es->rb_node, &tree->root);
678
679out:
680 tree->cache_es = es;
681 return 0;
682}
683
684/*
685 * ext4_es_insert_extent() adds information to an inode's extent
686 * status tree.
687 *
688 * Return 0 on success, error code on failure.
689 */
690int ext4_es_insert_extent(struct inode *inode, ext4_lblk_t lblk,
691 ext4_lblk_t len, ext4_fsblk_t pblk,
692 unsigned int status)
693{
694 struct extent_status newes;
695 ext4_lblk_t end = lblk + len - 1;
696 int err = 0;
697
698 es_debug("add [%u/%u) %llu %x to extent status tree of inode %lu\n",
699 lblk, len, pblk, status, inode->i_ino);
700
701 if (!len)
702 return 0;
703
704 BUG_ON(end < lblk);
705
706 if ((status & EXTENT_STATUS_DELAYED) &&
707 (status & EXTENT_STATUS_WRITTEN)) {
708 ext4_warning(inode->i_sb, "Inserting extent [%u/%u] as "
709 " delayed and written which can potentially "
710 " cause data loss.", lblk, len);
711 WARN_ON(1);
712 }
713
714 newes.es_lblk = lblk;
715 newes.es_len = len;
716 ext4_es_store_pblock_status(&newes, pblk, status);
717 trace_ext4_es_insert_extent(inode, &newes);
718
719 ext4_es_insert_extent_check(inode, &newes);
720
721 write_lock(&EXT4_I(inode)->i_es_lock);
722 err = __es_remove_extent(inode, lblk, end);
723 if (err != 0)
724 goto error;
725retry:
726 err = __es_insert_extent(inode, &newes);
727 if (err == -ENOMEM && __es_shrink(EXT4_SB(inode->i_sb),
728 128, EXT4_I(inode)))
729 goto retry;
730 if (err == -ENOMEM && !ext4_es_is_delayed(&newes))
731 err = 0;
732
733error:
734 write_unlock(&EXT4_I(inode)->i_es_lock);
735
736 ext4_es_print_tree(inode);
737
738 return err;
739}
740
741/*
742 * ext4_es_cache_extent() inserts information into the extent status
743 * tree if and only if there isn't information about the range in
744 * question already.
745 */
746void ext4_es_cache_extent(struct inode *inode, ext4_lblk_t lblk,
747 ext4_lblk_t len, ext4_fsblk_t pblk,
748 unsigned int status)
749{
750 struct extent_status *es;
751 struct extent_status newes;
752 ext4_lblk_t end = lblk + len - 1;
753
754 newes.es_lblk = lblk;
755 newes.es_len = len;
756 ext4_es_store_pblock_status(&newes, pblk, status);
757 trace_ext4_es_cache_extent(inode, &newes);
758
759 if (!len)
760 return;
761
762 BUG_ON(end < lblk);
763
764 write_lock(&EXT4_I(inode)->i_es_lock);
765
766 es = __es_tree_search(&EXT4_I(inode)->i_es_tree.root, lblk);
767 if (!es || es->es_lblk > end)
768 __es_insert_extent(inode, &newes);
769 write_unlock(&EXT4_I(inode)->i_es_lock);
770}
771
772/*
773 * ext4_es_lookup_extent() looks up an extent in extent status tree.
774 *
775 * ext4_es_lookup_extent is called by ext4_map_blocks/ext4_da_map_blocks.
776 *
777 * Return: 1 on found, 0 on not
778 */
779int ext4_es_lookup_extent(struct inode *inode, ext4_lblk_t lblk,
780 struct extent_status *es)
781{
782 struct ext4_es_tree *tree;
783 struct ext4_es_stats *stats;
784 struct extent_status *es1 = NULL;
785 struct rb_node *node;
786 int found = 0;
787
788 trace_ext4_es_lookup_extent_enter(inode, lblk);
789 es_debug("lookup extent in block %u\n", lblk);
790
791 tree = &EXT4_I(inode)->i_es_tree;
792 read_lock(&EXT4_I(inode)->i_es_lock);
793
794 /* find extent in cache firstly */
795 es->es_lblk = es->es_len = es->es_pblk = 0;
796 if (tree->cache_es) {
797 es1 = tree->cache_es;
798 if (in_range(lblk, es1->es_lblk, es1->es_len)) {
799 es_debug("%u cached by [%u/%u)\n",
800 lblk, es1->es_lblk, es1->es_len);
801 found = 1;
802 goto out;
803 }
804 }
805
806 node = tree->root.rb_node;
807 while (node) {
808 es1 = rb_entry(node, struct extent_status, rb_node);
809 if (lblk < es1->es_lblk)
810 node = node->rb_left;
811 else if (lblk > ext4_es_end(es1))
812 node = node->rb_right;
813 else {
814 found = 1;
815 break;
816 }
817 }
818
819out:
820 stats = &EXT4_SB(inode->i_sb)->s_es_stats;
821 if (found) {
822 BUG_ON(!es1);
823 es->es_lblk = es1->es_lblk;
824 es->es_len = es1->es_len;
825 es->es_pblk = es1->es_pblk;
826 if (!ext4_es_is_referenced(es1))
827 ext4_es_set_referenced(es1);
828 stats->es_stats_cache_hits++;
829 } else {
830 stats->es_stats_cache_misses++;
831 }
832
833 read_unlock(&EXT4_I(inode)->i_es_lock);
834
835 trace_ext4_es_lookup_extent_exit(inode, es, found);
836 return found;
837}
838
839static int __es_remove_extent(struct inode *inode, ext4_lblk_t lblk,
840 ext4_lblk_t end)
841{
842 struct ext4_es_tree *tree = &EXT4_I(inode)->i_es_tree;
843 struct rb_node *node;
844 struct extent_status *es;
845 struct extent_status orig_es;
846 ext4_lblk_t len1, len2;
847 ext4_fsblk_t block;
848 int err;
849
850retry:
851 err = 0;
852 es = __es_tree_search(&tree->root, lblk);
853 if (!es)
854 goto out;
855 if (es->es_lblk > end)
856 goto out;
857
858 /* Simply invalidate cache_es. */
859 tree->cache_es = NULL;
860
861 orig_es.es_lblk = es->es_lblk;
862 orig_es.es_len = es->es_len;
863 orig_es.es_pblk = es->es_pblk;
864
865 len1 = lblk > es->es_lblk ? lblk - es->es_lblk : 0;
866 len2 = ext4_es_end(es) > end ? ext4_es_end(es) - end : 0;
867 if (len1 > 0)
868 es->es_len = len1;
869 if (len2 > 0) {
870 if (len1 > 0) {
871 struct extent_status newes;
872
873 newes.es_lblk = end + 1;
874 newes.es_len = len2;
875 block = 0x7FDEADBEEFULL;
876 if (ext4_es_is_written(&orig_es) ||
877 ext4_es_is_unwritten(&orig_es))
878 block = ext4_es_pblock(&orig_es) +
879 orig_es.es_len - len2;
880 ext4_es_store_pblock_status(&newes, block,
881 ext4_es_status(&orig_es));
882 err = __es_insert_extent(inode, &newes);
883 if (err) {
884 es->es_lblk = orig_es.es_lblk;
885 es->es_len = orig_es.es_len;
886 if ((err == -ENOMEM) &&
887 __es_shrink(EXT4_SB(inode->i_sb),
888 128, EXT4_I(inode)))
889 goto retry;
890 goto out;
891 }
892 } else {
893 es->es_lblk = end + 1;
894 es->es_len = len2;
895 if (ext4_es_is_written(es) ||
896 ext4_es_is_unwritten(es)) {
897 block = orig_es.es_pblk + orig_es.es_len - len2;
898 ext4_es_store_pblock(es, block);
899 }
900 }
901 goto out;
902 }
903
904 if (len1 > 0) {
905 node = rb_next(&es->rb_node);
906 if (node)
907 es = rb_entry(node, struct extent_status, rb_node);
908 else
909 es = NULL;
910 }
911
912 while (es && ext4_es_end(es) <= end) {
913 node = rb_next(&es->rb_node);
914 rb_erase(&es->rb_node, &tree->root);
915 ext4_es_free_extent(inode, es);
916 if (!node) {
917 es = NULL;
918 break;
919 }
920 es = rb_entry(node, struct extent_status, rb_node);
921 }
922
923 if (es && es->es_lblk < end + 1) {
924 ext4_lblk_t orig_len = es->es_len;
925
926 len1 = ext4_es_end(es) - end;
927 es->es_lblk = end + 1;
928 es->es_len = len1;
929 if (ext4_es_is_written(es) || ext4_es_is_unwritten(es)) {
930 block = es->es_pblk + orig_len - len1;
931 ext4_es_store_pblock(es, block);
932 }
933 }
934
935out:
936 return err;
937}
938
939/*
940 * ext4_es_remove_extent() removes a space from a extent status tree.
941 *
942 * Return 0 on success, error code on failure.
943 */
944int ext4_es_remove_extent(struct inode *inode, ext4_lblk_t lblk,
945 ext4_lblk_t len)
946{
947 ext4_lblk_t end;
948 int err = 0;
949
950 trace_ext4_es_remove_extent(inode, lblk, len);
951 es_debug("remove [%u/%u) from extent status tree of inode %lu\n",
952 lblk, len, inode->i_ino);
953
954 if (!len)
955 return err;
956
957 end = lblk + len - 1;
958 BUG_ON(end < lblk);
959
960 /*
961 * ext4_clear_inode() depends on us taking i_es_lock unconditionally
962 * so that we are sure __es_shrink() is done with the inode before it
963 * is reclaimed.
964 */
965 write_lock(&EXT4_I(inode)->i_es_lock);
966 err = __es_remove_extent(inode, lblk, end);
967 write_unlock(&EXT4_I(inode)->i_es_lock);
968 ext4_es_print_tree(inode);
969 return err;
970}
971
972static int __es_shrink(struct ext4_sb_info *sbi, int nr_to_scan,
973 struct ext4_inode_info *locked_ei)
974{
975 struct ext4_inode_info *ei;
976 struct ext4_es_stats *es_stats;
977 ktime_t start_time;
978 u64 scan_time;
979 int nr_to_walk;
980 int nr_shrunk = 0;
981 int retried = 0, nr_skipped = 0;
982
983 es_stats = &sbi->s_es_stats;
984 start_time = ktime_get();
985
986retry:
987 spin_lock(&sbi->s_es_lock);
988 nr_to_walk = sbi->s_es_nr_inode;
989 while (nr_to_walk-- > 0) {
990 if (list_empty(&sbi->s_es_list)) {
991 spin_unlock(&sbi->s_es_lock);
992 goto out;
993 }
994 ei = list_first_entry(&sbi->s_es_list, struct ext4_inode_info,
995 i_es_list);
996 /* Move the inode to the tail */
997 list_move_tail(&ei->i_es_list, &sbi->s_es_list);
998
999 /*
1000 * Normally we try hard to avoid shrinking precached inodes,
1001 * but we will as a last resort.
1002 */
1003 if (!retried && ext4_test_inode_state(&ei->vfs_inode,
1004 EXT4_STATE_EXT_PRECACHED)) {
1005 nr_skipped++;
1006 continue;
1007 }
1008
1009 if (ei == locked_ei || !write_trylock(&ei->i_es_lock)) {
1010 nr_skipped++;
1011 continue;
1012 }
1013 /*
1014 * Now we hold i_es_lock which protects us from inode reclaim
1015 * freeing inode under us
1016 */
1017 spin_unlock(&sbi->s_es_lock);
1018
1019 nr_shrunk += es_reclaim_extents(ei, &nr_to_scan);
1020 write_unlock(&ei->i_es_lock);
1021
1022 if (nr_to_scan <= 0)
1023 goto out;
1024 spin_lock(&sbi->s_es_lock);
1025 }
1026 spin_unlock(&sbi->s_es_lock);
1027
1028 /*
1029 * If we skipped any inodes, and we weren't able to make any
1030 * forward progress, try again to scan precached inodes.
1031 */
1032 if ((nr_shrunk == 0) && nr_skipped && !retried) {
1033 retried++;
1034 goto retry;
1035 }
1036
1037 if (locked_ei && nr_shrunk == 0)
1038 nr_shrunk = es_reclaim_extents(locked_ei, &nr_to_scan);
1039
1040out:
1041 scan_time = ktime_to_ns(ktime_sub(ktime_get(), start_time));
1042 if (likely(es_stats->es_stats_scan_time))
1043 es_stats->es_stats_scan_time = (scan_time +
1044 es_stats->es_stats_scan_time*3) / 4;
1045 else
1046 es_stats->es_stats_scan_time = scan_time;
1047 if (scan_time > es_stats->es_stats_max_scan_time)
1048 es_stats->es_stats_max_scan_time = scan_time;
1049 if (likely(es_stats->es_stats_shrunk))
1050 es_stats->es_stats_shrunk = (nr_shrunk +
1051 es_stats->es_stats_shrunk*3) / 4;
1052 else
1053 es_stats->es_stats_shrunk = nr_shrunk;
1054
1055 trace_ext4_es_shrink(sbi->s_sb, nr_shrunk, scan_time,
1056 nr_skipped, retried);
1057 return nr_shrunk;
1058}
1059
1060static unsigned long ext4_es_count(struct shrinker *shrink,
1061 struct shrink_control *sc)
1062{
1063 unsigned long nr;
1064 struct ext4_sb_info *sbi;
1065
1066 sbi = container_of(shrink, struct ext4_sb_info, s_es_shrinker);
1067 nr = percpu_counter_read_positive(&sbi->s_es_stats.es_stats_shk_cnt);
1068 trace_ext4_es_shrink_count(sbi->s_sb, sc->nr_to_scan, nr);
1069 return nr;
1070}
1071
1072static unsigned long ext4_es_scan(struct shrinker *shrink,
1073 struct shrink_control *sc)
1074{
1075 struct ext4_sb_info *sbi = container_of(shrink,
1076 struct ext4_sb_info, s_es_shrinker);
1077 int nr_to_scan = sc->nr_to_scan;
1078 int ret, nr_shrunk;
1079
1080 ret = percpu_counter_read_positive(&sbi->s_es_stats.es_stats_shk_cnt);
1081 trace_ext4_es_shrink_scan_enter(sbi->s_sb, nr_to_scan, ret);
1082
1083 if (!nr_to_scan)
1084 return ret;
1085
1086 nr_shrunk = __es_shrink(sbi, nr_to_scan, NULL);
1087
1088 trace_ext4_es_shrink_scan_exit(sbi->s_sb, nr_shrunk, ret);
1089 return nr_shrunk;
1090}
1091
1092int ext4_seq_es_shrinker_info_show(struct seq_file *seq, void *v)
1093{
1094 struct ext4_sb_info *sbi = EXT4_SB((struct super_block *) seq->private);
1095 struct ext4_es_stats *es_stats = &sbi->s_es_stats;
1096 struct ext4_inode_info *ei, *max = NULL;
1097 unsigned int inode_cnt = 0;
1098
1099 if (v != SEQ_START_TOKEN)
1100 return 0;
1101
1102 /* here we just find an inode that has the max nr. of objects */
1103 spin_lock(&sbi->s_es_lock);
1104 list_for_each_entry(ei, &sbi->s_es_list, i_es_list) {
1105 inode_cnt++;
1106 if (max && max->i_es_all_nr < ei->i_es_all_nr)
1107 max = ei;
1108 else if (!max)
1109 max = ei;
1110 }
1111 spin_unlock(&sbi->s_es_lock);
1112
1113 seq_printf(seq, "stats:\n %lld objects\n %lld reclaimable objects\n",
1114 percpu_counter_sum_positive(&es_stats->es_stats_all_cnt),
1115 percpu_counter_sum_positive(&es_stats->es_stats_shk_cnt));
1116 seq_printf(seq, " %lu/%lu cache hits/misses\n",
1117 es_stats->es_stats_cache_hits,
1118 es_stats->es_stats_cache_misses);
1119 if (inode_cnt)
1120 seq_printf(seq, " %d inodes on list\n", inode_cnt);
1121
1122 seq_printf(seq, "average:\n %llu us scan time\n",
1123 div_u64(es_stats->es_stats_scan_time, 1000));
1124 seq_printf(seq, " %lu shrunk objects\n", es_stats->es_stats_shrunk);
1125 if (inode_cnt)
1126 seq_printf(seq,
1127 "maximum:\n %lu inode (%u objects, %u reclaimable)\n"
1128 " %llu us max scan time\n",
1129 max->vfs_inode.i_ino, max->i_es_all_nr, max->i_es_shk_nr,
1130 div_u64(es_stats->es_stats_max_scan_time, 1000));
1131
1132 return 0;
1133}
1134
1135int ext4_es_register_shrinker(struct ext4_sb_info *sbi)
1136{
1137 int err;
1138
1139 /* Make sure we have enough bits for physical block number */
1140 BUILD_BUG_ON(ES_SHIFT < 48);
1141 INIT_LIST_HEAD(&sbi->s_es_list);
1142 sbi->s_es_nr_inode = 0;
1143 spin_lock_init(&sbi->s_es_lock);
1144 sbi->s_es_stats.es_stats_shrunk = 0;
1145 sbi->s_es_stats.es_stats_cache_hits = 0;
1146 sbi->s_es_stats.es_stats_cache_misses = 0;
1147 sbi->s_es_stats.es_stats_scan_time = 0;
1148 sbi->s_es_stats.es_stats_max_scan_time = 0;
1149 err = percpu_counter_init(&sbi->s_es_stats.es_stats_all_cnt, 0, GFP_KERNEL);
1150 if (err)
1151 return err;
1152 err = percpu_counter_init(&sbi->s_es_stats.es_stats_shk_cnt, 0, GFP_KERNEL);
1153 if (err)
1154 goto err1;
1155
1156 sbi->s_es_shrinker.scan_objects = ext4_es_scan;
1157 sbi->s_es_shrinker.count_objects = ext4_es_count;
1158 sbi->s_es_shrinker.seeks = DEFAULT_SEEKS;
1159 err = register_shrinker(&sbi->s_es_shrinker);
1160 if (err)
1161 goto err2;
1162
1163 return 0;
1164
1165err2:
1166 percpu_counter_destroy(&sbi->s_es_stats.es_stats_shk_cnt);
1167err1:
1168 percpu_counter_destroy(&sbi->s_es_stats.es_stats_all_cnt);
1169 return err;
1170}
1171
1172void ext4_es_unregister_shrinker(struct ext4_sb_info *sbi)
1173{
1174 percpu_counter_destroy(&sbi->s_es_stats.es_stats_all_cnt);
1175 percpu_counter_destroy(&sbi->s_es_stats.es_stats_shk_cnt);
1176 unregister_shrinker(&sbi->s_es_shrinker);
1177}
1178
1179/*
1180 * Shrink extents in given inode from ei->i_es_shrink_lblk till end. Scan at
1181 * most *nr_to_scan extents, update *nr_to_scan accordingly.
1182 *
1183 * Return 0 if we hit end of tree / interval, 1 if we exhausted nr_to_scan.
1184 * Increment *nr_shrunk by the number of reclaimed extents. Also update
1185 * ei->i_es_shrink_lblk to where we should continue scanning.
1186 */
1187static int es_do_reclaim_extents(struct ext4_inode_info *ei, ext4_lblk_t end,
1188 int *nr_to_scan, int *nr_shrunk)
1189{
1190 struct inode *inode = &ei->vfs_inode;
1191 struct ext4_es_tree *tree = &ei->i_es_tree;
1192 struct extent_status *es;
1193 struct rb_node *node;
1194
1195 es = __es_tree_search(&tree->root, ei->i_es_shrink_lblk);
1196 if (!es)
1197 goto out_wrap;
1198 node = &es->rb_node;
1199 while (*nr_to_scan > 0) {
1200 if (es->es_lblk > end) {
1201 ei->i_es_shrink_lblk = end + 1;
1202 return 0;
1203 }
1204
1205 (*nr_to_scan)--;
1206 node = rb_next(&es->rb_node);
1207 /*
1208 * We can't reclaim delayed extent from status tree because
1209 * fiemap, bigallic, and seek_data/hole need to use it.
1210 */
1211 if (ext4_es_is_delayed(es))
1212 goto next;
1213 if (ext4_es_is_referenced(es)) {
1214 ext4_es_clear_referenced(es);
1215 goto next;
1216 }
1217
1218 rb_erase(&es->rb_node, &tree->root);
1219 ext4_es_free_extent(inode, es);
1220 (*nr_shrunk)++;
1221next:
1222 if (!node)
1223 goto out_wrap;
1224 es = rb_entry(node, struct extent_status, rb_node);
1225 }
1226 ei->i_es_shrink_lblk = es->es_lblk;
1227 return 1;
1228out_wrap:
1229 ei->i_es_shrink_lblk = 0;
1230 return 0;
1231}
1232
1233static int es_reclaim_extents(struct ext4_inode_info *ei, int *nr_to_scan)
1234{
1235 struct inode *inode = &ei->vfs_inode;
1236 int nr_shrunk = 0;
1237 ext4_lblk_t start = ei->i_es_shrink_lblk;
1238 static DEFINE_RATELIMIT_STATE(_rs, DEFAULT_RATELIMIT_INTERVAL,
1239 DEFAULT_RATELIMIT_BURST);
1240
1241 if (ei->i_es_shk_nr == 0)
1242 return 0;
1243
1244 if (ext4_test_inode_state(inode, EXT4_STATE_EXT_PRECACHED) &&
1245 __ratelimit(&_rs))
1246 ext4_warning(inode->i_sb, "forced shrink of precached extents");
1247
1248 if (!es_do_reclaim_extents(ei, EXT_MAX_BLOCKS, nr_to_scan, &nr_shrunk) &&
1249 start != 0)
1250 es_do_reclaim_extents(ei, start - 1, nr_to_scan, &nr_shrunk);
1251
1252 ei->i_es_tree.cache_es = NULL;
1253 return nr_shrunk;
1254}
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
143static struct kmem_cache *ext4_es_cachep;
144
145static int __es_insert_extent(struct inode *inode, struct extent_status *newes);
146static int __es_remove_extent(struct inode *inode, ext4_lblk_t lblk,
147 ext4_lblk_t end);
148static int __es_try_to_reclaim_extents(struct ext4_inode_info *ei,
149 int nr_to_scan);
150static int __ext4_es_shrink(struct ext4_sb_info *sbi, int nr_to_scan,
151 struct ext4_inode_info *locked_ei);
152
153int __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
163void ext4_exit_es(void)
164{
165 if (ext4_es_cachep)
166 kmem_cache_destroy(ext4_es_cachep);
167}
168
169void 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__
176static 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 %x",
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
198static 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 */
208static 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 */
245void 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
274out:
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
299static struct extent_status *
300ext4_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
322static 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 */
341static 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
367static struct extent_status *
368ext4_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
389static struct extent_status *
390ext4_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
413static 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 a delayed/hole extent "
449 "[%d/%d/%llu/%x]\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 a written/unwritten extent "
490 "[%d/%d/%llu/%x]\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 }
495out:
496 if (path) {
497 ext4_ext_drop_refs(path);
498 kfree(path);
499 }
500}
501
502static 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/%x]\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 "a written extent [%d/%d/%llu/%x]\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
565static 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
579static inline void ext4_es_insert_extent_check(struct inode *inode,
580 struct extent_status *es)
581{
582}
583#endif
584
585static 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
632out:
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 */
643int 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_status(&newes, pblk, status);
662 trace_ext4_es_insert_extent(inode, &newes);
663
664 ext4_es_insert_extent_check(inode, &newes);
665
666 write_lock(&EXT4_I(inode)->i_es_lock);
667 err = __es_remove_extent(inode, lblk, end);
668 if (err != 0)
669 goto error;
670retry:
671 err = __es_insert_extent(inode, &newes);
672 if (err == -ENOMEM && __ext4_es_shrink(EXT4_SB(inode->i_sb), 1,
673 EXT4_I(inode)))
674 goto retry;
675 if (err == -ENOMEM && !ext4_es_is_delayed(&newes))
676 err = 0;
677
678error:
679 write_unlock(&EXT4_I(inode)->i_es_lock);
680
681 ext4_es_print_tree(inode);
682
683 return err;
684}
685
686/*
687 * ext4_es_cache_extent() inserts information into the extent status
688 * tree if and only if there isn't information about the range in
689 * question already.
690 */
691void ext4_es_cache_extent(struct inode *inode, ext4_lblk_t lblk,
692 ext4_lblk_t len, ext4_fsblk_t pblk,
693 unsigned int status)
694{
695 struct extent_status *es;
696 struct extent_status newes;
697 ext4_lblk_t end = lblk + len - 1;
698
699 newes.es_lblk = lblk;
700 newes.es_len = len;
701 ext4_es_store_pblock_status(&newes, pblk, status);
702 trace_ext4_es_cache_extent(inode, &newes);
703
704 if (!len)
705 return;
706
707 BUG_ON(end < lblk);
708
709 write_lock(&EXT4_I(inode)->i_es_lock);
710
711 es = __es_tree_search(&EXT4_I(inode)->i_es_tree.root, lblk);
712 if (!es || es->es_lblk > end)
713 __es_insert_extent(inode, &newes);
714 write_unlock(&EXT4_I(inode)->i_es_lock);
715}
716
717/*
718 * ext4_es_lookup_extent() looks up an extent in extent status tree.
719 *
720 * ext4_es_lookup_extent is called by ext4_map_blocks/ext4_da_map_blocks.
721 *
722 * Return: 1 on found, 0 on not
723 */
724int ext4_es_lookup_extent(struct inode *inode, ext4_lblk_t lblk,
725 struct extent_status *es)
726{
727 struct ext4_es_tree *tree;
728 struct extent_status *es1 = NULL;
729 struct rb_node *node;
730 int found = 0;
731
732 trace_ext4_es_lookup_extent_enter(inode, lblk);
733 es_debug("lookup extent in block %u\n", lblk);
734
735 tree = &EXT4_I(inode)->i_es_tree;
736 read_lock(&EXT4_I(inode)->i_es_lock);
737
738 /* find extent in cache firstly */
739 es->es_lblk = es->es_len = es->es_pblk = 0;
740 if (tree->cache_es) {
741 es1 = tree->cache_es;
742 if (in_range(lblk, es1->es_lblk, es1->es_len)) {
743 es_debug("%u cached by [%u/%u)\n",
744 lblk, es1->es_lblk, es1->es_len);
745 found = 1;
746 goto out;
747 }
748 }
749
750 node = tree->root.rb_node;
751 while (node) {
752 es1 = rb_entry(node, struct extent_status, rb_node);
753 if (lblk < es1->es_lblk)
754 node = node->rb_left;
755 else if (lblk > ext4_es_end(es1))
756 node = node->rb_right;
757 else {
758 found = 1;
759 break;
760 }
761 }
762
763out:
764 if (found) {
765 BUG_ON(!es1);
766 es->es_lblk = es1->es_lblk;
767 es->es_len = es1->es_len;
768 es->es_pblk = es1->es_pblk;
769 }
770
771 read_unlock(&EXT4_I(inode)->i_es_lock);
772
773 trace_ext4_es_lookup_extent_exit(inode, es, found);
774 return found;
775}
776
777static int __es_remove_extent(struct inode *inode, ext4_lblk_t lblk,
778 ext4_lblk_t end)
779{
780 struct ext4_es_tree *tree = &EXT4_I(inode)->i_es_tree;
781 struct rb_node *node;
782 struct extent_status *es;
783 struct extent_status orig_es;
784 ext4_lblk_t len1, len2;
785 ext4_fsblk_t block;
786 int err;
787
788retry:
789 err = 0;
790 es = __es_tree_search(&tree->root, lblk);
791 if (!es)
792 goto out;
793 if (es->es_lblk > end)
794 goto out;
795
796 /* Simply invalidate cache_es. */
797 tree->cache_es = NULL;
798
799 orig_es.es_lblk = es->es_lblk;
800 orig_es.es_len = es->es_len;
801 orig_es.es_pblk = es->es_pblk;
802
803 len1 = lblk > es->es_lblk ? lblk - es->es_lblk : 0;
804 len2 = ext4_es_end(es) > end ? ext4_es_end(es) - end : 0;
805 if (len1 > 0)
806 es->es_len = len1;
807 if (len2 > 0) {
808 if (len1 > 0) {
809 struct extent_status newes;
810
811 newes.es_lblk = end + 1;
812 newes.es_len = len2;
813 block = 0x7FDEADBEEFULL;
814 if (ext4_es_is_written(&orig_es) ||
815 ext4_es_is_unwritten(&orig_es))
816 block = ext4_es_pblock(&orig_es) +
817 orig_es.es_len - len2;
818 ext4_es_store_pblock_status(&newes, block,
819 ext4_es_status(&orig_es));
820 err = __es_insert_extent(inode, &newes);
821 if (err) {
822 es->es_lblk = orig_es.es_lblk;
823 es->es_len = orig_es.es_len;
824 if ((err == -ENOMEM) &&
825 __ext4_es_shrink(EXT4_SB(inode->i_sb), 1,
826 EXT4_I(inode)))
827 goto retry;
828 goto out;
829 }
830 } else {
831 es->es_lblk = end + 1;
832 es->es_len = len2;
833 if (ext4_es_is_written(es) ||
834 ext4_es_is_unwritten(es)) {
835 block = orig_es.es_pblk + orig_es.es_len - len2;
836 ext4_es_store_pblock(es, block);
837 }
838 }
839 goto out;
840 }
841
842 if (len1 > 0) {
843 node = rb_next(&es->rb_node);
844 if (node)
845 es = rb_entry(node, struct extent_status, rb_node);
846 else
847 es = NULL;
848 }
849
850 while (es && ext4_es_end(es) <= end) {
851 node = rb_next(&es->rb_node);
852 rb_erase(&es->rb_node, &tree->root);
853 ext4_es_free_extent(inode, es);
854 if (!node) {
855 es = NULL;
856 break;
857 }
858 es = rb_entry(node, struct extent_status, rb_node);
859 }
860
861 if (es && es->es_lblk < end + 1) {
862 ext4_lblk_t orig_len = es->es_len;
863
864 len1 = ext4_es_end(es) - end;
865 es->es_lblk = end + 1;
866 es->es_len = len1;
867 if (ext4_es_is_written(es) || ext4_es_is_unwritten(es)) {
868 block = es->es_pblk + orig_len - len1;
869 ext4_es_store_pblock(es, block);
870 }
871 }
872
873out:
874 return err;
875}
876
877/*
878 * ext4_es_remove_extent() removes a space from a extent status tree.
879 *
880 * Return 0 on success, error code on failure.
881 */
882int ext4_es_remove_extent(struct inode *inode, ext4_lblk_t lblk,
883 ext4_lblk_t len)
884{
885 ext4_lblk_t end;
886 int err = 0;
887
888 trace_ext4_es_remove_extent(inode, lblk, len);
889 es_debug("remove [%u/%u) from extent status tree of inode %lu\n",
890 lblk, len, inode->i_ino);
891
892 if (!len)
893 return err;
894
895 end = lblk + len - 1;
896 BUG_ON(end < lblk);
897
898 write_lock(&EXT4_I(inode)->i_es_lock);
899 err = __es_remove_extent(inode, lblk, end);
900 write_unlock(&EXT4_I(inode)->i_es_lock);
901 ext4_es_print_tree(inode);
902 return err;
903}
904
905static int ext4_inode_touch_time_cmp(void *priv, struct list_head *a,
906 struct list_head *b)
907{
908 struct ext4_inode_info *eia, *eib;
909 eia = list_entry(a, struct ext4_inode_info, i_es_lru);
910 eib = list_entry(b, struct ext4_inode_info, i_es_lru);
911
912 if (ext4_test_inode_state(&eia->vfs_inode, EXT4_STATE_EXT_PRECACHED) &&
913 !ext4_test_inode_state(&eib->vfs_inode, EXT4_STATE_EXT_PRECACHED))
914 return 1;
915 if (!ext4_test_inode_state(&eia->vfs_inode, EXT4_STATE_EXT_PRECACHED) &&
916 ext4_test_inode_state(&eib->vfs_inode, EXT4_STATE_EXT_PRECACHED))
917 return -1;
918 if (eia->i_touch_when == eib->i_touch_when)
919 return 0;
920 if (time_after(eia->i_touch_when, eib->i_touch_when))
921 return 1;
922 else
923 return -1;
924}
925
926static int __ext4_es_shrink(struct ext4_sb_info *sbi, int nr_to_scan,
927 struct ext4_inode_info *locked_ei)
928{
929 struct ext4_inode_info *ei;
930 struct list_head *cur, *tmp;
931 LIST_HEAD(skipped);
932 int nr_shrunk = 0;
933 int retried = 0, skip_precached = 1, nr_skipped = 0;
934
935 spin_lock(&sbi->s_es_lru_lock);
936
937retry:
938 list_for_each_safe(cur, tmp, &sbi->s_es_lru) {
939 int shrunk;
940
941 /*
942 * If we have already reclaimed all extents from extent
943 * status tree, just stop the loop immediately.
944 */
945 if (percpu_counter_read_positive(&sbi->s_extent_cache_cnt) == 0)
946 break;
947
948 ei = list_entry(cur, struct ext4_inode_info, i_es_lru);
949
950 /*
951 * Skip the inode that is newer than the last_sorted
952 * time. Normally we try hard to avoid shrinking
953 * precached inodes, but we will as a last resort.
954 */
955 if ((sbi->s_es_last_sorted < ei->i_touch_when) ||
956 (skip_precached && ext4_test_inode_state(&ei->vfs_inode,
957 EXT4_STATE_EXT_PRECACHED))) {
958 nr_skipped++;
959 list_move_tail(cur, &skipped);
960 continue;
961 }
962
963 if (ei->i_es_lru_nr == 0 || ei == locked_ei)
964 continue;
965
966 write_lock(&ei->i_es_lock);
967 shrunk = __es_try_to_reclaim_extents(ei, nr_to_scan);
968 if (ei->i_es_lru_nr == 0)
969 list_del_init(&ei->i_es_lru);
970 write_unlock(&ei->i_es_lock);
971
972 nr_shrunk += shrunk;
973 nr_to_scan -= shrunk;
974 if (nr_to_scan == 0)
975 break;
976 }
977
978 /* Move the newer inodes into the tail of the LRU list. */
979 list_splice_tail(&skipped, &sbi->s_es_lru);
980 INIT_LIST_HEAD(&skipped);
981
982 /*
983 * If we skipped any inodes, and we weren't able to make any
984 * forward progress, sort the list and try again.
985 */
986 if ((nr_shrunk == 0) && nr_skipped && !retried) {
987 retried++;
988 list_sort(NULL, &sbi->s_es_lru, ext4_inode_touch_time_cmp);
989 sbi->s_es_last_sorted = jiffies;
990 ei = list_first_entry(&sbi->s_es_lru, struct ext4_inode_info,
991 i_es_lru);
992 /*
993 * If there are no non-precached inodes left on the
994 * list, start releasing precached extents.
995 */
996 if (ext4_test_inode_state(&ei->vfs_inode,
997 EXT4_STATE_EXT_PRECACHED))
998 skip_precached = 0;
999 goto retry;
1000 }
1001
1002 spin_unlock(&sbi->s_es_lru_lock);
1003
1004 if (locked_ei && nr_shrunk == 0)
1005 nr_shrunk = __es_try_to_reclaim_extents(locked_ei, nr_to_scan);
1006
1007 return nr_shrunk;
1008}
1009
1010static unsigned long ext4_es_count(struct shrinker *shrink,
1011 struct shrink_control *sc)
1012{
1013 unsigned long nr;
1014 struct ext4_sb_info *sbi;
1015
1016 sbi = container_of(shrink, struct ext4_sb_info, s_es_shrinker);
1017 nr = percpu_counter_read_positive(&sbi->s_extent_cache_cnt);
1018 trace_ext4_es_shrink_enter(sbi->s_sb, sc->nr_to_scan, nr);
1019 return nr;
1020}
1021
1022static unsigned long ext4_es_scan(struct shrinker *shrink,
1023 struct shrink_control *sc)
1024{
1025 struct ext4_sb_info *sbi = container_of(shrink,
1026 struct ext4_sb_info, s_es_shrinker);
1027 int nr_to_scan = sc->nr_to_scan;
1028 int ret, nr_shrunk;
1029
1030 ret = percpu_counter_read_positive(&sbi->s_extent_cache_cnt);
1031 trace_ext4_es_shrink_enter(sbi->s_sb, nr_to_scan, ret);
1032
1033 if (!nr_to_scan)
1034 return ret;
1035
1036 nr_shrunk = __ext4_es_shrink(sbi, nr_to_scan, NULL);
1037
1038 trace_ext4_es_shrink_exit(sbi->s_sb, nr_shrunk, ret);
1039 return nr_shrunk;
1040}
1041
1042void ext4_es_register_shrinker(struct ext4_sb_info *sbi)
1043{
1044 INIT_LIST_HEAD(&sbi->s_es_lru);
1045 spin_lock_init(&sbi->s_es_lru_lock);
1046 sbi->s_es_last_sorted = 0;
1047 sbi->s_es_shrinker.scan_objects = ext4_es_scan;
1048 sbi->s_es_shrinker.count_objects = ext4_es_count;
1049 sbi->s_es_shrinker.seeks = DEFAULT_SEEKS;
1050 register_shrinker(&sbi->s_es_shrinker);
1051}
1052
1053void ext4_es_unregister_shrinker(struct ext4_sb_info *sbi)
1054{
1055 unregister_shrinker(&sbi->s_es_shrinker);
1056}
1057
1058void ext4_es_lru_add(struct inode *inode)
1059{
1060 struct ext4_inode_info *ei = EXT4_I(inode);
1061 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1062
1063 ei->i_touch_when = jiffies;
1064
1065 if (!list_empty(&ei->i_es_lru))
1066 return;
1067
1068 spin_lock(&sbi->s_es_lru_lock);
1069 if (list_empty(&ei->i_es_lru))
1070 list_add_tail(&ei->i_es_lru, &sbi->s_es_lru);
1071 spin_unlock(&sbi->s_es_lru_lock);
1072}
1073
1074void ext4_es_lru_del(struct inode *inode)
1075{
1076 struct ext4_inode_info *ei = EXT4_I(inode);
1077 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
1078
1079 spin_lock(&sbi->s_es_lru_lock);
1080 if (!list_empty(&ei->i_es_lru))
1081 list_del_init(&ei->i_es_lru);
1082 spin_unlock(&sbi->s_es_lru_lock);
1083}
1084
1085static int __es_try_to_reclaim_extents(struct ext4_inode_info *ei,
1086 int nr_to_scan)
1087{
1088 struct inode *inode = &ei->vfs_inode;
1089 struct ext4_es_tree *tree = &ei->i_es_tree;
1090 struct rb_node *node;
1091 struct extent_status *es;
1092 unsigned long nr_shrunk = 0;
1093 static DEFINE_RATELIMIT_STATE(_rs, DEFAULT_RATELIMIT_INTERVAL,
1094 DEFAULT_RATELIMIT_BURST);
1095
1096 if (ei->i_es_lru_nr == 0)
1097 return 0;
1098
1099 if (ext4_test_inode_state(inode, EXT4_STATE_EXT_PRECACHED) &&
1100 __ratelimit(&_rs))
1101 ext4_warning(inode->i_sb, "forced shrink of precached extents");
1102
1103 node = rb_first(&tree->root);
1104 while (node != NULL) {
1105 es = rb_entry(node, struct extent_status, rb_node);
1106 node = rb_next(&es->rb_node);
1107 /*
1108 * We can't reclaim delayed extent from status tree because
1109 * fiemap, bigallic, and seek_data/hole need to use it.
1110 */
1111 if (!ext4_es_is_delayed(es)) {
1112 rb_erase(&es->rb_node, &tree->root);
1113 ext4_es_free_extent(inode, es);
1114 nr_shrunk++;
1115 if (--nr_to_scan == 0)
1116 break;
1117 }
1118 }
1119 tree->cache_es = NULL;
1120 return nr_shrunk;
1121}