projects / deliverable / linux.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
blame | history | raw | HEAD
Merge tag 'regulator-fix-v4.0-rc4' of git://git.kernel.org/pub/scm/linux/kernel/git...
[deliverable/linux.git] / fs / ext4 / extents_status.c
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 <linux/proc_fs.h>
15 #include <linux/seq_file.h>
16 #include "ext4.h"
17 #include "extents_status.h"
18
19 #include <trace/events/ext4.h>
20
21 /*
22 * According to previous discussion in Ext4 Developer Workshop, we
23 * will introduce a new structure called io tree to track all extent
24 * status in order to solve some problems that we have met
25 * (e.g. Reservation space warning), and provide extent-level locking.
26 * Delay extent tree is the first step to achieve this goal. It is
27 * original built by Yongqiang Yang. At that time it is called delay
28 * extent tree, whose goal is only track delayed extents in memory to
29 * simplify the implementation of fiemap and bigalloc, and introduce
30 * lseek SEEK_DATA/SEEK_HOLE support. That is why it is still called
31 * delay extent tree at the first commit. But for better understand
32 * what it does, it has been rename to extent status tree.
33 *
34 * Step1:
35 * Currently the first step has been done. All delayed extents are
36 * tracked in the tree. It maintains the delayed extent when a delayed
37 * allocation is issued, and the delayed extent is written out or
38 * invalidated. Therefore the implementation of fiemap and bigalloc
39 * are simplified, and SEEK_DATA/SEEK_HOLE are introduced.
40 *
41 * The following comment describes the implemenmtation of extent
42 * status tree and future works.
43 *
44 * Step2:
45 * In this step all extent status are tracked by extent status tree.
46 * Thus, we can first try to lookup a block mapping in this tree before
47 * finding it in extent tree. Hence, single extent cache can be removed
48 * because extent status tree can do a better job. Extents in status
49 * tree are loaded on-demand. Therefore, the extent status tree may not
50 * contain all of the extents in a file. Meanwhile we define a shrinker
51 * to reclaim memory from extent status tree because fragmented extent
52 * tree will make status tree cost too much memory. written/unwritten/-
53 * hole extents in the tree will be reclaimed by this shrinker when we
54 * are under high memory pressure. Delayed extents will not be
55 * reclimed because fiemap, bigalloc, and seek_data/hole need it.
56 */
57
58 /*
59 * Extent status tree implementation for ext4.
60 *
61 *
62 * ==========================================================================
63 * Extent status tree tracks all extent status.
64 *
65 * 1. Why we need to implement extent status tree?
66 *
67 * Without extent status tree, ext4 identifies a delayed extent by looking
68 * up page cache, this has several deficiencies - complicated, buggy,
69 * and inefficient code.
70 *
71 * FIEMAP, SEEK_HOLE/DATA, bigalloc, and writeout all need to know if a
72 * block or a range of blocks are belonged to a delayed extent.
73 *
74 * Let us have a look at how they do without extent status tree.
75 * -- FIEMAP
76 * FIEMAP looks up page cache to identify delayed allocations from holes.
77 *
78 * -- SEEK_HOLE/DATA
79 * SEEK_HOLE/DATA has the same problem as FIEMAP.
80 *
81 * -- bigalloc
82 * bigalloc looks up page cache to figure out if a block is
83 * already under delayed allocation or not to determine whether
84 * quota reserving is needed for the cluster.
85 *
86 * -- writeout
87 * Writeout looks up whole page cache to see if a buffer is
88 * mapped, If there are not very many delayed buffers, then it is
89 * time comsuming.
90 *
91 * With extent status tree implementation, FIEMAP, SEEK_HOLE/DATA,
92 * bigalloc and writeout can figure out if a block or a range of
93 * blocks is under delayed allocation(belonged to a delayed extent) or
94 * not by searching the extent tree.
95 *
96 *
97 * ==========================================================================
98 * 2. Ext4 extent status tree impelmentation
99 *
100 * -- extent
101 * A extent is a range of blocks which are contiguous logically and
102 * physically. Unlike extent in extent tree, this extent in ext4 is
103 * a in-memory struct, there is no corresponding on-disk data. There
104 * is no limit on length of extent, so an extent can contain as many
105 * blocks as they are contiguous logically and physically.
106 *
107 * -- extent status tree
108 * Every inode has an extent status tree and all allocation blocks
109 * are added to the tree with different status. The extent in the
110 * tree are ordered by logical block no.
111 *
112 * -- operations on a extent status tree
113 * There are three important operations on a delayed extent tree: find
114 * next extent, adding a extent(a range of blocks) and removing a extent.
115 *
116 * -- race on a extent status tree
117 * Extent status tree is protected by inode->i_es_lock.
118 *
119 * -- memory consumption
120 * Fragmented extent tree will make extent status tree cost too much
121 * memory. Hence, we will reclaim written/unwritten/hole extents from
122 * the tree under a heavy memory pressure.
123 *
124 *
125 * ==========================================================================
126 * 3. Performance analysis
127 *
128 * -- overhead
129 * 1. There is a cache extent for write access, so if writes are
130 * not very random, adding space operaions are in O(1) time.
131 *
132 * -- gain
133 * 2. Code is much simpler, more readable, more maintainable and
134 * more efficient.
135 *
136 *
137 * ==========================================================================
138 * 4. TODO list
139 *
140 * -- Refactor delayed space reservation
141 *
142 * -- Extent-level locking
143 */
144
145 static struct kmem_cache *ext4_es_cachep;
146
147 static int __es_insert_extent(struct inode *inode, struct extent_status *newes);
148 static int __es_remove_extent(struct inode *inode, ext4_lblk_t lblk,
149 ext4_lblk_t end);
150 static int es_reclaim_extents(struct ext4_inode_info *ei, int *nr_to_scan);
151 static int __es_shrink(struct ext4_sb_info *sbi, int nr_to_scan,
152 struct ext4_inode_info *locked_ei);
153
154 int __init ext4_init_es(void)
155 {
156 ext4_es_cachep = kmem_cache_create("ext4_extent_status",
157 sizeof(struct extent_status),
158 0, (SLAB_RECLAIM_ACCOUNT), NULL);
159 if (ext4_es_cachep == NULL)
160 return -ENOMEM;
161 return 0;
162 }
163
164 void ext4_exit_es(void)
165 {
166 if (ext4_es_cachep)
167 kmem_cache_destroy(ext4_es_cachep);
168 }
169
170 void ext4_es_init_tree(struct ext4_es_tree *tree)
171 {
172 tree->root = RB_ROOT;
173 tree->cache_es = NULL;
174 }
175
176 #ifdef ES_DEBUG__
177 static void ext4_es_print_tree(struct inode *inode)
178 {
179 struct ext4_es_tree *tree;
180 struct rb_node *node;
181
182 printk(KERN_DEBUG "status extents for inode %lu:", inode->i_ino);
183 tree = &EXT4_I(inode)->i_es_tree;
184 node = rb_first(&tree->root);
185 while (node) {
186 struct extent_status *es;
187 es = rb_entry(node, struct extent_status, rb_node);
188 printk(KERN_DEBUG " [%u/%u) %llu %x",
189 es->es_lblk, es->es_len,
190 ext4_es_pblock(es), ext4_es_status(es));
191 node = rb_next(node);
192 }
193 printk(KERN_DEBUG "\n");
194 }
195 #else
196 #define ext4_es_print_tree(inode)
197 #endif
198
199 static inline ext4_lblk_t ext4_es_end(struct extent_status *es)
200 {
201 BUG_ON(es->es_lblk + es->es_len < es->es_lblk);
202 return es->es_lblk + es->es_len - 1;
203 }
204
205 /*
206 * search through the tree for an delayed extent with a given offset. If
207 * it can't be found, try to find next extent.
208 */
209 static struct extent_status *__es_tree_search(struct rb_root *root,
210 ext4_lblk_t lblk)
211 {
212 struct rb_node *node = root->rb_node;
213 struct extent_status *es = NULL;
214
215 while (node) {
216 es = rb_entry(node, struct extent_status, rb_node);
217 if (lblk < es->es_lblk)
218 node = node->rb_left;
219 else if (lblk > ext4_es_end(es))
220 node = node->rb_right;
221 else
222 return es;
223 }
224
225 if (es && lblk < es->es_lblk)
226 return es;
227
228 if (es && lblk > ext4_es_end(es)) {
229 node = rb_next(&es->rb_node);
230 return node ? rb_entry(node, struct extent_status, rb_node) :
231 NULL;
232 }
233
234 return NULL;
235 }
236
237 /*
238 * ext4_es_find_delayed_extent_range: find the 1st delayed extent covering
239 * @es->lblk if it exists, otherwise, the next extent after @es->lblk.
240 *
241 * @inode: the inode which owns delayed extents
242 * @lblk: the offset where we start to search
243 * @end: the offset where we stop to search
244 * @es: delayed extent that we found
245 */
246 void ext4_es_find_delayed_extent_range(struct inode *inode,
247 ext4_lblk_t lblk, ext4_lblk_t end,
248 struct extent_status *es)
249 {
250 struct ext4_es_tree *tree = NULL;
251 struct extent_status *es1 = NULL;
252 struct rb_node *node;
253
254 BUG_ON(es == NULL);
255 BUG_ON(end < lblk);
256 trace_ext4_es_find_delayed_extent_range_enter(inode, lblk);
257
258 read_lock(&EXT4_I(inode)->i_es_lock);
259 tree = &EXT4_I(inode)->i_es_tree;
260
261 /* find extent in cache firstly */
262 es->es_lblk = es->es_len = es->es_pblk = 0;
263 if (tree->cache_es) {
264 es1 = tree->cache_es;
265 if (in_range(lblk, es1->es_lblk, es1->es_len)) {
266 es_debug("%u cached by [%u/%u) %llu %x\n",
267 lblk, es1->es_lblk, es1->es_len,
268 ext4_es_pblock(es1), ext4_es_status(es1));
269 goto out;
270 }
271 }
272
273 es1 = __es_tree_search(&tree->root, lblk);
274
275 out:
276 if (es1 && !ext4_es_is_delayed(es1)) {
277 while ((node = rb_next(&es1->rb_node)) != NULL) {
278 es1 = rb_entry(node, struct extent_status, rb_node);
279 if (es1->es_lblk > end) {
280 es1 = NULL;
281 break;
282 }
283 if (ext4_es_is_delayed(es1))
284 break;
285 }
286 }
287
288 if (es1 && ext4_es_is_delayed(es1)) {
289 tree->cache_es = es1;
290 es->es_lblk = es1->es_lblk;
291 es->es_len = es1->es_len;
292 es->es_pblk = es1->es_pblk;
293 }
294
295 read_unlock(&EXT4_I(inode)->i_es_lock);
296
297 trace_ext4_es_find_delayed_extent_range_exit(inode, es);
298 }
299
300 static void ext4_es_list_add(struct inode *inode)
301 {
302 struct ext4_inode_info *ei = EXT4_I(inode);
303 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
304
305 if (!list_empty(&ei->i_es_list))
306 return;
307
308 spin_lock(&sbi->s_es_lock);
309 if (list_empty(&ei->i_es_list)) {
310 list_add_tail(&ei->i_es_list, &sbi->s_es_list);
311 sbi->s_es_nr_inode++;
312 }
313 spin_unlock(&sbi->s_es_lock);
314 }
315
316 static void ext4_es_list_del(struct inode *inode)
317 {
318 struct ext4_inode_info *ei = EXT4_I(inode);
319 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
320
321 spin_lock(&sbi->s_es_lock);
322 if (!list_empty(&ei->i_es_list)) {
323 list_del_init(&ei->i_es_list);
324 sbi->s_es_nr_inode--;
325 WARN_ON_ONCE(sbi->s_es_nr_inode < 0);
326 }
327 spin_unlock(&sbi->s_es_lock);
328 }
329
330 static struct extent_status *
331 ext4_es_alloc_extent(struct inode *inode, ext4_lblk_t lblk, ext4_lblk_t len,
332 ext4_fsblk_t pblk)
333 {
334 struct extent_status *es;
335 es = kmem_cache_alloc(ext4_es_cachep, GFP_ATOMIC);
336 if (es == NULL)
337 return NULL;
338 es->es_lblk = lblk;
339 es->es_len = len;
340 es->es_pblk = pblk;
341
342 /*
343 * We don't count delayed extent because we never try to reclaim them
344 */
345 if (!ext4_es_is_delayed(es)) {
346 if (!EXT4_I(inode)->i_es_shk_nr++)
347 ext4_es_list_add(inode);
348 percpu_counter_inc(&EXT4_SB(inode->i_sb)->
349 s_es_stats.es_stats_shk_cnt);
350 }
351
352 EXT4_I(inode)->i_es_all_nr++;
353 percpu_counter_inc(&EXT4_SB(inode->i_sb)->s_es_stats.es_stats_all_cnt);
354
355 return es;
356 }
357
358 static void ext4_es_free_extent(struct inode *inode, struct extent_status *es)
359 {
360 EXT4_I(inode)->i_es_all_nr--;
361 percpu_counter_dec(&EXT4_SB(inode->i_sb)->s_es_stats.es_stats_all_cnt);
362
363 /* Decrease the shrink counter when this es is not delayed */
364 if (!ext4_es_is_delayed(es)) {
365 BUG_ON(EXT4_I(inode)->i_es_shk_nr == 0);
366 if (!--EXT4_I(inode)->i_es_shk_nr)
367 ext4_es_list_del(inode);
368 percpu_counter_dec(&EXT4_SB(inode->i_sb)->
369 s_es_stats.es_stats_shk_cnt);
370 }
371
372 kmem_cache_free(ext4_es_cachep, es);
373 }
374
375 /*
376 * Check whether or not two extents can be merged
377 * Condition:
378 * - logical block number is contiguous
379 * - physical block number is contiguous
380 * - status is equal
381 */
382 static int ext4_es_can_be_merged(struct extent_status *es1,
383 struct extent_status *es2)
384 {
385 if (ext4_es_type(es1) != ext4_es_type(es2))
386 return 0;
387
388 if (((__u64) es1->es_len) + es2->es_len > EXT_MAX_BLOCKS) {
389 pr_warn("ES assertion failed when merging extents. "
390 "The sum of lengths of es1 (%d) and es2 (%d) "
391 "is bigger than allowed file size (%d)\n",
392 es1->es_len, es2->es_len, EXT_MAX_BLOCKS);
393 WARN_ON(1);
394 return 0;
395 }
396
397 if (((__u64) es1->es_lblk) + es1->es_len != es2->es_lblk)
398 return 0;
399
400 if ((ext4_es_is_written(es1) || ext4_es_is_unwritten(es1)) &&
401 (ext4_es_pblock(es1) + es1->es_len == ext4_es_pblock(es2)))
402 return 1;
403
404 if (ext4_es_is_hole(es1))
405 return 1;
406
407 /* we need to check delayed extent is without unwritten status */
408 if (ext4_es_is_delayed(es1) && !ext4_es_is_unwritten(es1))
409 return 1;
410
411 return 0;
412 }
413
414 static struct extent_status *
415 ext4_es_try_to_merge_left(struct inode *inode, struct extent_status *es)
416 {
417 struct ext4_es_tree *tree = &EXT4_I(inode)->i_es_tree;
418 struct extent_status *es1;
419 struct rb_node *node;
420
421 node = rb_prev(&es->rb_node);
422 if (!node)
423 return es;
424
425 es1 = rb_entry(node, struct extent_status, rb_node);
426 if (ext4_es_can_be_merged(es1, es)) {
427 es1->es_len += es->es_len;
428 if (ext4_es_is_referenced(es))
429 ext4_es_set_referenced(es1);
430 rb_erase(&es->rb_node, &tree->root);
431 ext4_es_free_extent(inode, es);
432 es = es1;
433 }
434
435 return es;
436 }
437
438 static struct extent_status *
439 ext4_es_try_to_merge_right(struct inode *inode, struct extent_status *es)
440 {
441 struct ext4_es_tree *tree = &EXT4_I(inode)->i_es_tree;
442 struct extent_status *es1;
443 struct rb_node *node;
444
445 node = rb_next(&es->rb_node);
446 if (!node)
447 return es;
448
449 es1 = rb_entry(node, struct extent_status, rb_node);
450 if (ext4_es_can_be_merged(es, es1)) {
451 es->es_len += es1->es_len;
452 if (ext4_es_is_referenced(es1))
453 ext4_es_set_referenced(es);
454 rb_erase(node, &tree->root);
455 ext4_es_free_extent(inode, es1);
456 }
457
458 return es;
459 }
460
461 #ifdef ES_AGGRESSIVE_TEST
462 #include "ext4_extents.h" /* Needed when ES_AGGRESSIVE_TEST is defined */
463
464 static void ext4_es_insert_extent_ext_check(struct inode *inode,
465 struct extent_status *es)
466 {
467 struct ext4_ext_path *path = NULL;
468 struct ext4_extent *ex;
469 ext4_lblk_t ee_block;
470 ext4_fsblk_t ee_start;
471 unsigned short ee_len;
472 int depth, ee_status, es_status;
473
474 path = ext4_find_extent(inode, es->es_lblk, NULL, EXT4_EX_NOCACHE);
475 if (IS_ERR(path))
476 return;
477
478 depth = ext_depth(inode);
479 ex = path[depth].p_ext;
480
481 if (ex) {
482
483 ee_block = le32_to_cpu(ex->ee_block);
484 ee_start = ext4_ext_pblock(ex);
485 ee_len = ext4_ext_get_actual_len(ex);
486
487 ee_status = ext4_ext_is_unwritten(ex) ? 1 : 0;
488 es_status = ext4_es_is_unwritten(es) ? 1 : 0;
489
490 /*
491 * Make sure ex and es are not overlap when we try to insert
492 * a delayed/hole extent.
493 */
494 if (!ext4_es_is_written(es) && !ext4_es_is_unwritten(es)) {
495 if (in_range(es->es_lblk, ee_block, ee_len)) {
496 pr_warn("ES insert assertion failed for "
497 "inode: %lu we can find an extent "
498 "at block [%d/%d/%llu/%c], but we "
499 "want to add a delayed/hole extent "
500 "[%d/%d/%llu/%x]\n",
501 inode->i_ino, ee_block, ee_len,
502 ee_start, ee_status ? 'u' : 'w',
503 es->es_lblk, es->es_len,
504 ext4_es_pblock(es), ext4_es_status(es));
505 }
506 goto out;
507 }
508
509 /*
510 * We don't check ee_block == es->es_lblk, etc. because es
511 * might be a part of whole extent, vice versa.
512 */
513 if (es->es_lblk < ee_block ||
514 ext4_es_pblock(es) != ee_start + es->es_lblk - ee_block) {
515 pr_warn("ES insert assertion failed for inode: %lu "
516 "ex_status [%d/%d/%llu/%c] != "
517 "es_status [%d/%d/%llu/%c]\n", inode->i_ino,
518 ee_block, ee_len, ee_start,
519 ee_status ? 'u' : 'w', es->es_lblk, es->es_len,
520 ext4_es_pblock(es), es_status ? 'u' : 'w');
521 goto out;
522 }
523
524 if (ee_status ^ es_status) {
525 pr_warn("ES insert assertion failed for inode: %lu "
526 "ex_status [%d/%d/%llu/%c] != "
527 "es_status [%d/%d/%llu/%c]\n", inode->i_ino,
528 ee_block, ee_len, ee_start,
529 ee_status ? 'u' : 'w', es->es_lblk, es->es_len,
530 ext4_es_pblock(es), es_status ? 'u' : 'w');
531 }
532 } else {
533 /*
534 * We can't find an extent on disk. So we need to make sure
535 * that we don't want to add an written/unwritten extent.
536 */
537 if (!ext4_es_is_delayed(es) && !ext4_es_is_hole(es)) {
538 pr_warn("ES insert assertion failed for inode: %lu "
539 "can't find an extent at block %d but we want "
540 "to add a written/unwritten extent "
541 "[%d/%d/%llu/%x]\n", inode->i_ino,
542 es->es_lblk, es->es_lblk, es->es_len,
543 ext4_es_pblock(es), ext4_es_status(es));
544 }
545 }
546 out:
547 ext4_ext_drop_refs(path);
548 kfree(path);
549 }
550
551 static void ext4_es_insert_extent_ind_check(struct inode *inode,
552 struct extent_status *es)
553 {
554 struct ext4_map_blocks map;
555 int retval;
556
557 /*
558 * Here we call ext4_ind_map_blocks to lookup a block mapping because
559 * 'Indirect' structure is defined in indirect.c. So we couldn't
560 * access direct/indirect tree from outside. It is too dirty to define
561 * this function in indirect.c file.
562 */
563
564 map.m_lblk = es->es_lblk;
565 map.m_len = es->es_len;
566
567 retval = ext4_ind_map_blocks(NULL, inode, &map, 0);
568 if (retval > 0) {
569 if (ext4_es_is_delayed(es) || ext4_es_is_hole(es)) {
570 /*
571 * We want to add a delayed/hole extent but this
572 * block has been allocated.
573 */
574 pr_warn("ES insert assertion failed for inode: %lu "
575 "We can find blocks but we want to add a "
576 "delayed/hole extent [%d/%d/%llu/%x]\n",
577 inode->i_ino, es->es_lblk, es->es_len,
578 ext4_es_pblock(es), ext4_es_status(es));
579 return;
580 } else if (ext4_es_is_written(es)) {
581 if (retval != es->es_len) {
582 pr_warn("ES insert assertion failed for "
583 "inode: %lu retval %d != es_len %d\n",
584 inode->i_ino, retval, es->es_len);
585 return;
586 }
587 if (map.m_pblk != ext4_es_pblock(es)) {
588 pr_warn("ES insert assertion failed for "
589 "inode: %lu m_pblk %llu != "
590 "es_pblk %llu\n",
591 inode->i_ino, map.m_pblk,
592 ext4_es_pblock(es));
593 return;
594 }
595 } else {
596 /*
597 * We don't need to check unwritten extent because
598 * indirect-based file doesn't have it.
599 */
600 BUG_ON(1);
601 }
602 } else if (retval == 0) {
603 if (ext4_es_is_written(es)) {
604 pr_warn("ES insert assertion failed for inode: %lu "
605 "We can't find the block but we want to add "
606 "a written extent [%d/%d/%llu/%x]\n",
607 inode->i_ino, es->es_lblk, es->es_len,
608 ext4_es_pblock(es), ext4_es_status(es));
609 return;
610 }
611 }
612 }
613
614 static inline void ext4_es_insert_extent_check(struct inode *inode,
615 struct extent_status *es)
616 {
617 /*
618 * We don't need to worry about the race condition because
619 * caller takes i_data_sem locking.
620 */
621 BUG_ON(!rwsem_is_locked(&EXT4_I(inode)->i_data_sem));
622 if (ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))
623 ext4_es_insert_extent_ext_check(inode, es);
624 else
625 ext4_es_insert_extent_ind_check(inode, es);
626 }
627 #else
628 static inline void ext4_es_insert_extent_check(struct inode *inode,
629 struct extent_status *es)
630 {
631 }
632 #endif
633
634 static int __es_insert_extent(struct inode *inode, struct extent_status *newes)
635 {
636 struct ext4_es_tree *tree = &EXT4_I(inode)->i_es_tree;
637 struct rb_node **p = &tree->root.rb_node;
638 struct rb_node *parent = NULL;
639 struct extent_status *es;
640
641 while (*p) {
642 parent = *p;
643 es = rb_entry(parent, struct extent_status, rb_node);
644
645 if (newes->es_lblk < es->es_lblk) {
646 if (ext4_es_can_be_merged(newes, es)) {
647 /*
648 * Here we can modify es_lblk directly
649 * because it isn't overlapped.
650 */
651 es->es_lblk = newes->es_lblk;
652 es->es_len += newes->es_len;
653 if (ext4_es_is_written(es) ||
654 ext4_es_is_unwritten(es))
655 ext4_es_store_pblock(es,
656 newes->es_pblk);
657 es = ext4_es_try_to_merge_left(inode, es);
658 goto out;
659 }
660 p = &(*p)->rb_left;
661 } else if (newes->es_lblk > ext4_es_end(es)) {
662 if (ext4_es_can_be_merged(es, newes)) {
663 es->es_len += newes->es_len;
664 es = ext4_es_try_to_merge_right(inode, es);
665 goto out;
666 }
667 p = &(*p)->rb_right;
668 } else {
669 BUG_ON(1);
670 return -EINVAL;
671 }
672 }
673
674 es = ext4_es_alloc_extent(inode, newes->es_lblk, newes->es_len,
675 newes->es_pblk);
676 if (!es)
677 return -ENOMEM;
678 rb_link_node(&es->rb_node, parent, p);
679 rb_insert_color(&es->rb_node, &tree->root);
680
681 out:
682 tree->cache_es = es;
683 return 0;
684 }
685
686 /*
687 * ext4_es_insert_extent() adds information to an inode's extent
688 * status tree.
689 *
690 * Return 0 on success, error code on failure.
691 */
692 int ext4_es_insert_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 newes;
697 ext4_lblk_t end = lblk + len - 1;
698 int err = 0;
699
700 es_debug("add [%u/%u) %llu %x to extent status tree of inode %lu\n",
701 lblk, len, pblk, status, inode->i_ino);
702
703 if (!len)
704 return 0;
705
706 BUG_ON(end < lblk);
707
708 newes.es_lblk = lblk;
709 newes.es_len = len;
710 ext4_es_store_pblock_status(&newes, pblk, status);
711 trace_ext4_es_insert_extent(inode, &newes);
712
713 ext4_es_insert_extent_check(inode, &newes);
714
715 write_lock(&EXT4_I(inode)->i_es_lock);
716 err = __es_remove_extent(inode, lblk, end);
717 if (err != 0)
718 goto error;
719 retry:
720 err = __es_insert_extent(inode, &newes);
721 if (err == -ENOMEM && __es_shrink(EXT4_SB(inode->i_sb),
722 128, EXT4_I(inode)))
723 goto retry;
724 if (err == -ENOMEM && !ext4_es_is_delayed(&newes))
725 err = 0;
726
727 error:
728 write_unlock(&EXT4_I(inode)->i_es_lock);
729
730 ext4_es_print_tree(inode);
731
732 return err;
733 }
734
735 /*
736 * ext4_es_cache_extent() inserts information into the extent status
737 * tree if and only if there isn't information about the range in
738 * question already.
739 */
740 void ext4_es_cache_extent(struct inode *inode, ext4_lblk_t lblk,
741 ext4_lblk_t len, ext4_fsblk_t pblk,
742 unsigned int status)
743 {
744 struct extent_status *es;
745 struct extent_status newes;
746 ext4_lblk_t end = lblk + len - 1;
747
748 newes.es_lblk = lblk;
749 newes.es_len = len;
750 ext4_es_store_pblock_status(&newes, pblk, status);
751 trace_ext4_es_cache_extent(inode, &newes);
752
753 if (!len)
754 return;
755
756 BUG_ON(end < lblk);
757
758 write_lock(&EXT4_I(inode)->i_es_lock);
759
760 es = __es_tree_search(&EXT4_I(inode)->i_es_tree.root, lblk);
761 if (!es || es->es_lblk > end)
762 __es_insert_extent(inode, &newes);
763 write_unlock(&EXT4_I(inode)->i_es_lock);
764 }
765
766 /*
767 * ext4_es_lookup_extent() looks up an extent in extent status tree.
768 *
769 * ext4_es_lookup_extent is called by ext4_map_blocks/ext4_da_map_blocks.
770 *
771 * Return: 1 on found, 0 on not
772 */
773 int ext4_es_lookup_extent(struct inode *inode, ext4_lblk_t lblk,
774 struct extent_status *es)
775 {
776 struct ext4_es_tree *tree;
777 struct ext4_es_stats *stats;
778 struct extent_status *es1 = NULL;
779 struct rb_node *node;
780 int found = 0;
781
782 trace_ext4_es_lookup_extent_enter(inode, lblk);
783 es_debug("lookup extent in block %u\n", lblk);
784
785 tree = &EXT4_I(inode)->i_es_tree;
786 read_lock(&EXT4_I(inode)->i_es_lock);
787
788 /* find extent in cache firstly */
789 es->es_lblk = es->es_len = es->es_pblk = 0;
790 if (tree->cache_es) {
791 es1 = tree->cache_es;
792 if (in_range(lblk, es1->es_lblk, es1->es_len)) {
793 es_debug("%u cached by [%u/%u)\n",
794 lblk, es1->es_lblk, es1->es_len);
795 found = 1;
796 goto out;
797 }
798 }
799
800 node = tree->root.rb_node;
801 while (node) {
802 es1 = rb_entry(node, struct extent_status, rb_node);
803 if (lblk < es1->es_lblk)
804 node = node->rb_left;
805 else if (lblk > ext4_es_end(es1))
806 node = node->rb_right;
807 else {
808 found = 1;
809 break;
810 }
811 }
812
813 out:
814 stats = &EXT4_SB(inode->i_sb)->s_es_stats;
815 if (found) {
816 BUG_ON(!es1);
817 es->es_lblk = es1->es_lblk;
818 es->es_len = es1->es_len;
819 es->es_pblk = es1->es_pblk;
820 if (!ext4_es_is_referenced(es))
821 ext4_es_set_referenced(es);
822 stats->es_stats_cache_hits++;
823 } else {
824 stats->es_stats_cache_misses++;
825 }
826
827 read_unlock(&EXT4_I(inode)->i_es_lock);
828
829 trace_ext4_es_lookup_extent_exit(inode, es, found);
830 return found;
831 }
832
833 static int __es_remove_extent(struct inode *inode, ext4_lblk_t lblk,
834 ext4_lblk_t end)
835 {
836 struct ext4_es_tree *tree = &EXT4_I(inode)->i_es_tree;
837 struct rb_node *node;
838 struct extent_status *es;
839 struct extent_status orig_es;
840 ext4_lblk_t len1, len2;
841 ext4_fsblk_t block;
842 int err;
843
844 retry:
845 err = 0;
846 es = __es_tree_search(&tree->root, lblk);
847 if (!es)
848 goto out;
849 if (es->es_lblk > end)
850 goto out;
851
852 /* Simply invalidate cache_es. */
853 tree->cache_es = NULL;
854
855 orig_es.es_lblk = es->es_lblk;
856 orig_es.es_len = es->es_len;
857 orig_es.es_pblk = es->es_pblk;
858
859 len1 = lblk > es->es_lblk ? lblk - es->es_lblk : 0;
860 len2 = ext4_es_end(es) > end ? ext4_es_end(es) - end : 0;
861 if (len1 > 0)
862 es->es_len = len1;
863 if (len2 > 0) {
864 if (len1 > 0) {
865 struct extent_status newes;
866
867 newes.es_lblk = end + 1;
868 newes.es_len = len2;
869 block = 0x7FDEADBEEFULL;
870 if (ext4_es_is_written(&orig_es) ||
871 ext4_es_is_unwritten(&orig_es))
872 block = ext4_es_pblock(&orig_es) +
873 orig_es.es_len - len2;
874 ext4_es_store_pblock_status(&newes, block,
875 ext4_es_status(&orig_es));
876 err = __es_insert_extent(inode, &newes);
877 if (err) {
878 es->es_lblk = orig_es.es_lblk;
879 es->es_len = orig_es.es_len;
880 if ((err == -ENOMEM) &&
881 __es_shrink(EXT4_SB(inode->i_sb),
882 128, EXT4_I(inode)))
883 goto retry;
884 goto out;
885 }
886 } else {
887 es->es_lblk = end + 1;
888 es->es_len = len2;
889 if (ext4_es_is_written(es) ||
890 ext4_es_is_unwritten(es)) {
891 block = orig_es.es_pblk + orig_es.es_len - len2;
892 ext4_es_store_pblock(es, block);
893 }
894 }
895 goto out;
896 }
897
898 if (len1 > 0) {
899 node = rb_next(&es->rb_node);
900 if (node)
901 es = rb_entry(node, struct extent_status, rb_node);
902 else
903 es = NULL;
904 }
905
906 while (es && ext4_es_end(es) <= end) {
907 node = rb_next(&es->rb_node);
908 rb_erase(&es->rb_node, &tree->root);
909 ext4_es_free_extent(inode, es);
910 if (!node) {
911 es = NULL;
912 break;
913 }
914 es = rb_entry(node, struct extent_status, rb_node);
915 }
916
917 if (es && es->es_lblk < end + 1) {
918 ext4_lblk_t orig_len = es->es_len;
919
920 len1 = ext4_es_end(es) - end;
921 es->es_lblk = end + 1;
922 es->es_len = len1;
923 if (ext4_es_is_written(es) || ext4_es_is_unwritten(es)) {
924 block = es->es_pblk + orig_len - len1;
925 ext4_es_store_pblock(es, block);
926 }
927 }
928
929 out:
930 return err;
931 }
932
933 /*
934 * ext4_es_remove_extent() removes a space from a extent status tree.
935 *
936 * Return 0 on success, error code on failure.
937 */
938 int ext4_es_remove_extent(struct inode *inode, ext4_lblk_t lblk,
939 ext4_lblk_t len)
940 {
941 ext4_lblk_t end;
942 int err = 0;
943
944 trace_ext4_es_remove_extent(inode, lblk, len);
945 es_debug("remove [%u/%u) from extent status tree of inode %lu\n",
946 lblk, len, inode->i_ino);
947
948 if (!len)
949 return err;
950
951 end = lblk + len - 1;
952 BUG_ON(end < lblk);
953
954 /*
955 * ext4_clear_inode() depends on us taking i_es_lock unconditionally
956 * so that we are sure __es_shrink() is done with the inode before it
957 * is reclaimed.
958 */
959 write_lock(&EXT4_I(inode)->i_es_lock);
960 err = __es_remove_extent(inode, lblk, end);
961 write_unlock(&EXT4_I(inode)->i_es_lock);
962 ext4_es_print_tree(inode);
963 return err;
964 }
965
966 static int __es_shrink(struct ext4_sb_info *sbi, int nr_to_scan,
967 struct ext4_inode_info *locked_ei)
968 {
969 struct ext4_inode_info *ei;
970 struct ext4_es_stats *es_stats;
971 ktime_t start_time;
972 u64 scan_time;
973 int nr_to_walk;
974 int nr_shrunk = 0;
975 int retried = 0, nr_skipped = 0;
976
977 es_stats = &sbi->s_es_stats;
978 start_time = ktime_get();
979
980 retry:
981 spin_lock(&sbi->s_es_lock);
982 nr_to_walk = sbi->s_es_nr_inode;
983 while (nr_to_walk-- > 0) {
984 if (list_empty(&sbi->s_es_list)) {
985 spin_unlock(&sbi->s_es_lock);
986 goto out;
987 }
988 ei = list_first_entry(&sbi->s_es_list, struct ext4_inode_info,
989 i_es_list);
990 /* Move the inode to the tail */
991 list_move_tail(&ei->i_es_list, &sbi->s_es_list);
992
993 /*
994 * Normally we try hard to avoid shrinking precached inodes,
995 * but we will as a last resort.
996 */
997 if (!retried && ext4_test_inode_state(&ei->vfs_inode,
998 EXT4_STATE_EXT_PRECACHED)) {
999 nr_skipped++;
1000 continue;
1001 }
1002
1003 if (ei == locked_ei || !write_trylock(&ei->i_es_lock)) {
1004 nr_skipped++;
1005 continue;
1006 }
1007 /*
1008 * Now we hold i_es_lock which protects us from inode reclaim
1009 * freeing inode under us
1010 */
1011 spin_unlock(&sbi->s_es_lock);
1012
1013 nr_shrunk += es_reclaim_extents(ei, &nr_to_scan);
1014 write_unlock(&ei->i_es_lock);
1015
1016 if (nr_to_scan <= 0)
1017 goto out;
1018 spin_lock(&sbi->s_es_lock);
1019 }
1020 spin_unlock(&sbi->s_es_lock);
1021
1022 /*
1023 * If we skipped any inodes, and we weren't able to make any
1024 * forward progress, try again to scan precached inodes.
1025 */
1026 if ((nr_shrunk == 0) && nr_skipped && !retried) {
1027 retried++;
1028 goto retry;
1029 }
1030
1031 if (locked_ei && nr_shrunk == 0)
1032 nr_shrunk = es_reclaim_extents(locked_ei, &nr_to_scan);
1033
1034 out:
1035 scan_time = ktime_to_ns(ktime_sub(ktime_get(), start_time));
1036 if (likely(es_stats->es_stats_scan_time))
1037 es_stats->es_stats_scan_time = (scan_time +
1038 es_stats->es_stats_scan_time*3) / 4;
1039 else
1040 es_stats->es_stats_scan_time = scan_time;
1041 if (scan_time > es_stats->es_stats_max_scan_time)
1042 es_stats->es_stats_max_scan_time = scan_time;
1043 if (likely(es_stats->es_stats_shrunk))
1044 es_stats->es_stats_shrunk = (nr_shrunk +
1045 es_stats->es_stats_shrunk*3) / 4;
1046 else
1047 es_stats->es_stats_shrunk = nr_shrunk;
1048
1049 trace_ext4_es_shrink(sbi->s_sb, nr_shrunk, scan_time,
1050 nr_skipped, retried);
1051 return nr_shrunk;
1052 }
1053
1054 static unsigned long ext4_es_count(struct shrinker *shrink,
1055 struct shrink_control *sc)
1056 {
1057 unsigned long nr;
1058 struct ext4_sb_info *sbi;
1059
1060 sbi = container_of(shrink, struct ext4_sb_info, s_es_shrinker);
1061 nr = percpu_counter_read_positive(&sbi->s_es_stats.es_stats_shk_cnt);
1062 trace_ext4_es_shrink_count(sbi->s_sb, sc->nr_to_scan, nr);
1063 return nr;
1064 }
1065
1066 static unsigned long ext4_es_scan(struct shrinker *shrink,
1067 struct shrink_control *sc)
1068 {
1069 struct ext4_sb_info *sbi = container_of(shrink,
1070 struct ext4_sb_info, s_es_shrinker);
1071 int nr_to_scan = sc->nr_to_scan;
1072 int ret, nr_shrunk;
1073
1074 ret = percpu_counter_read_positive(&sbi->s_es_stats.es_stats_shk_cnt);
1075 trace_ext4_es_shrink_scan_enter(sbi->s_sb, nr_to_scan, ret);
1076
1077 if (!nr_to_scan)
1078 return ret;
1079
1080 nr_shrunk = __es_shrink(sbi, nr_to_scan, NULL);
1081
1082 trace_ext4_es_shrink_scan_exit(sbi->s_sb, nr_shrunk, ret);
1083 return nr_shrunk;
1084 }
1085
1086 static void *ext4_es_seq_shrinker_info_start(struct seq_file *seq, loff_t *pos)
1087 {
1088 return *pos ? NULL : SEQ_START_TOKEN;
1089 }
1090
1091 static void *
1092 ext4_es_seq_shrinker_info_next(struct seq_file *seq, void *v, loff_t *pos)
1093 {
1094 return NULL;
1095 }
1096
1097 static int ext4_es_seq_shrinker_info_show(struct seq_file *seq, void *v)
1098 {
1099 struct ext4_sb_info *sbi = seq->private;
1100 struct ext4_es_stats *es_stats = &sbi->s_es_stats;
1101 struct ext4_inode_info *ei, *max = NULL;
1102 unsigned int inode_cnt = 0;
1103
1104 if (v != SEQ_START_TOKEN)
1105 return 0;
1106
1107 /* here we just find an inode that has the max nr. of objects */
1108 spin_lock(&sbi->s_es_lock);
1109 list_for_each_entry(ei, &sbi->s_es_list, i_es_list) {
1110 inode_cnt++;
1111 if (max && max->i_es_all_nr < ei->i_es_all_nr)
1112 max = ei;
1113 else if (!max)
1114 max = ei;
1115 }
1116 spin_unlock(&sbi->s_es_lock);
1117
1118 seq_printf(seq, "stats:\n %lld objects\n %lld reclaimable objects\n",
1119 percpu_counter_sum_positive(&es_stats->es_stats_all_cnt),
1120 percpu_counter_sum_positive(&es_stats->es_stats_shk_cnt));
1121 seq_printf(seq, " %lu/%lu cache hits/misses\n",
1122 es_stats->es_stats_cache_hits,
1123 es_stats->es_stats_cache_misses);
1124 if (inode_cnt)
1125 seq_printf(seq, " %d inodes on list\n", inode_cnt);
1126
1127 seq_printf(seq, "average:\n %llu us scan time\n",
1128 div_u64(es_stats->es_stats_scan_time, 1000));
1129 seq_printf(seq, " %lu shrunk objects\n", es_stats->es_stats_shrunk);
1130 if (inode_cnt)
1131 seq_printf(seq,
1132 "maximum:\n %lu inode (%u objects, %u reclaimable)\n"
1133 " %llu us max scan time\n",
1134 max->vfs_inode.i_ino, max->i_es_all_nr, max->i_es_shk_nr,
1135 div_u64(es_stats->es_stats_max_scan_time, 1000));
1136
1137 return 0;
1138 }
1139
1140 static void ext4_es_seq_shrinker_info_stop(struct seq_file *seq, void *v)
1141 {
1142 }
1143
1144 static const struct seq_operations ext4_es_seq_shrinker_info_ops = {
1145 .start = ext4_es_seq_shrinker_info_start,
1146 .next = ext4_es_seq_shrinker_info_next,
1147 .stop = ext4_es_seq_shrinker_info_stop,
1148 .show = ext4_es_seq_shrinker_info_show,
1149 };
1150
1151 static int
1152 ext4_es_seq_shrinker_info_open(struct inode *inode, struct file *file)
1153 {
1154 int ret;
1155
1156 ret = seq_open(file, &ext4_es_seq_shrinker_info_ops);
1157 if (!ret) {
1158 struct seq_file *m = file->private_data;
1159 m->private = PDE_DATA(inode);
1160 }
1161
1162 return ret;
1163 }
1164
1165 static int
1166 ext4_es_seq_shrinker_info_release(struct inode *inode, struct file *file)
1167 {
1168 return seq_release(inode, file);
1169 }
1170
1171 static const struct file_operations ext4_es_seq_shrinker_info_fops = {
1172 .owner = THIS_MODULE,
1173 .open = ext4_es_seq_shrinker_info_open,
1174 .read = seq_read,
1175 .llseek = seq_lseek,
1176 .release = ext4_es_seq_shrinker_info_release,
1177 };
1178
1179 int ext4_es_register_shrinker(struct ext4_sb_info *sbi)
1180 {
1181 int err;
1182
1183 /* Make sure we have enough bits for physical block number */
1184 BUILD_BUG_ON(ES_SHIFT < 48);
1185 INIT_LIST_HEAD(&sbi->s_es_list);
1186 sbi->s_es_nr_inode = 0;
1187 spin_lock_init(&sbi->s_es_lock);
1188 sbi->s_es_stats.es_stats_shrunk = 0;
1189 sbi->s_es_stats.es_stats_cache_hits = 0;
1190 sbi->s_es_stats.es_stats_cache_misses = 0;
1191 sbi->s_es_stats.es_stats_scan_time = 0;
1192 sbi->s_es_stats.es_stats_max_scan_time = 0;
1193 err = percpu_counter_init(&sbi->s_es_stats.es_stats_all_cnt, 0, GFP_KERNEL);
1194 if (err)
1195 return err;
1196 err = percpu_counter_init(&sbi->s_es_stats.es_stats_shk_cnt, 0, GFP_KERNEL);
1197 if (err)
1198 goto err1;
1199
1200 sbi->s_es_shrinker.scan_objects = ext4_es_scan;
1201 sbi->s_es_shrinker.count_objects = ext4_es_count;
1202 sbi->s_es_shrinker.seeks = DEFAULT_SEEKS;
1203 err = register_shrinker(&sbi->s_es_shrinker);
1204 if (err)
1205 goto err2;
1206
1207 if (sbi->s_proc)
1208 proc_create_data("es_shrinker_info", S_IRUGO, sbi->s_proc,
1209 &ext4_es_seq_shrinker_info_fops, sbi);
1210
1211 return 0;
1212
1213 err2:
1214 percpu_counter_destroy(&sbi->s_es_stats.es_stats_shk_cnt);
1215 err1:
1216 percpu_counter_destroy(&sbi->s_es_stats.es_stats_all_cnt);
1217 return err;
1218 }
1219
1220 void ext4_es_unregister_shrinker(struct ext4_sb_info *sbi)
1221 {
1222 if (sbi->s_proc)
1223 remove_proc_entry("es_shrinker_info", sbi->s_proc);
1224 percpu_counter_destroy(&sbi->s_es_stats.es_stats_all_cnt);
1225 percpu_counter_destroy(&sbi->s_es_stats.es_stats_shk_cnt);
1226 unregister_shrinker(&sbi->s_es_shrinker);
1227 }
1228
1229 /*
1230 * Shrink extents in given inode from ei->i_es_shrink_lblk till end. Scan at
1231 * most *nr_to_scan extents, update *nr_to_scan accordingly.
1232 *
1233 * Return 0 if we hit end of tree / interval, 1 if we exhausted nr_to_scan.
1234 * Increment *nr_shrunk by the number of reclaimed extents. Also update
1235 * ei->i_es_shrink_lblk to where we should continue scanning.
1236 */
1237 static int es_do_reclaim_extents(struct ext4_inode_info *ei, ext4_lblk_t end,
1238 int *nr_to_scan, int *nr_shrunk)
1239 {
1240 struct inode *inode = &ei->vfs_inode;
1241 struct ext4_es_tree *tree = &ei->i_es_tree;
1242 struct extent_status *es;
1243 struct rb_node *node;
1244
1245 es = __es_tree_search(&tree->root, ei->i_es_shrink_lblk);
1246 if (!es)
1247 goto out_wrap;
1248 node = &es->rb_node;
1249 while (*nr_to_scan > 0) {
1250 if (es->es_lblk > end) {
1251 ei->i_es_shrink_lblk = end + 1;
1252 return 0;
1253 }
1254
1255 (*nr_to_scan)--;
1256 node = rb_next(&es->rb_node);
1257 /*
1258 * We can't reclaim delayed extent from status tree because
1259 * fiemap, bigallic, and seek_data/hole need to use it.
1260 */
1261 if (ext4_es_is_delayed(es))
1262 goto next;
1263 if (ext4_es_is_referenced(es)) {
1264 ext4_es_clear_referenced(es);
1265 goto next;
1266 }
1267
1268 rb_erase(&es->rb_node, &tree->root);
1269 ext4_es_free_extent(inode, es);
1270 (*nr_shrunk)++;
1271 next:
1272 if (!node)
1273 goto out_wrap;
1274 es = rb_entry(node, struct extent_status, rb_node);
1275 }
1276 ei->i_es_shrink_lblk = es->es_lblk;
1277 return 1;
1278 out_wrap:
1279 ei->i_es_shrink_lblk = 0;
1280 return 0;
1281 }
1282
1283 static int es_reclaim_extents(struct ext4_inode_info *ei, int *nr_to_scan)
1284 {
1285 struct inode *inode = &ei->vfs_inode;
1286 int nr_shrunk = 0;
1287 ext4_lblk_t start = ei->i_es_shrink_lblk;
1288 static DEFINE_RATELIMIT_STATE(_rs, DEFAULT_RATELIMIT_INTERVAL,
1289 DEFAULT_RATELIMIT_BURST);
1290
1291 if (ei->i_es_shk_nr == 0)
1292 return 0;
1293
1294 if (ext4_test_inode_state(inode, EXT4_STATE_EXT_PRECACHED) &&
1295 __ratelimit(&_rs))
1296 ext4_warning(inode->i_sb, "forced shrink of precached extents");
1297
1298 if (!es_do_reclaim_extents(ei, EXT_MAX_BLOCKS, nr_to_scan, &nr_shrunk) &&
1299 start != 0)
1300 es_do_reclaim_extents(ei, start - 1, nr_to_scan, &nr_shrunk);
1301
1302 ei->i_es_tree.cache_es = NULL;
1303 return nr_shrunk;
1304 }
Linux kernel - Deliverables
This page took 0.0582819999999999 seconds and 5 git commands to generate.