Commit | Line | Data |
---|---|---|
654598be ZL |
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 "ext4.h" | |
14 | #include "extents_status.h" | |
15 | #include "ext4_extents.h" | |
16 | ||
992e9fdd ZL |
17 | #include <trace/events/ext4.h> |
18 | ||
654598be ZL |
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 | |
06b0c886 | 26 | * extent tree, whose goal is only track delayed extents in memory to |
654598be ZL |
27 | * simplify the implementation of fiemap and bigalloc, and introduce |
28 | * lseek SEEK_DATA/SEEK_HOLE support. That is why it is still called | |
06b0c886 ZL |
29 | * delay extent tree at the first commit. But for better understand |
30 | * what it does, it has been rename to extent status tree. | |
654598be | 31 | * |
06b0c886 ZL |
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 | |
654598be ZL |
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. | |
06b0c886 ZL |
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. | |
654598be ZL |
54 | */ |
55 | ||
56 | /* | |
06b0c886 | 57 | * Extent status tree implementation for ext4. |
654598be ZL |
58 | * |
59 | * | |
60 | * ========================================================================== | |
06b0c886 | 61 | * Extent status tree tracks all extent status. |
654598be | 62 | * |
06b0c886 | 63 | * 1. Why we need to implement extent status tree? |
654598be | 64 | * |
06b0c886 | 65 | * Without extent status tree, ext4 identifies a delayed extent by looking |
654598be ZL |
66 | * up page cache, this has several deficiencies - complicated, buggy, |
67 | * and inefficient code. | |
68 | * | |
06b0c886 ZL |
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. | |
654598be | 71 | * |
06b0c886 | 72 | * Let us have a look at how they do without extent status tree. |
654598be ZL |
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 | * | |
654598be ZL |
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 | * | |
06b0c886 | 89 | * With extent status tree implementation, FIEMAP, SEEK_HOLE/DATA, |
654598be ZL |
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 | |
06b0c886 | 92 | * not by searching the extent tree. |
654598be ZL |
93 | * |
94 | * | |
95 | * ========================================================================== | |
06b0c886 ZL |
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. | |
654598be | 104 | * |
06b0c886 ZL |
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. | |
654598be | 109 | * |
06b0c886 ZL |
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. | |
654598be | 113 | * |
06b0c886 ZL |
114 | * -- race on a extent status tree |
115 | * Extent status tree is protected by inode->i_es_lock. | |
654598be | 116 | * |
06b0c886 ZL |
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. | |
654598be ZL |
121 | * |
122 | * | |
123 | * ========================================================================== | |
06b0c886 ZL |
124 | * 3. Performance analysis |
125 | * | |
654598be ZL |
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 | |
654598be | 137 | * |
06b0c886 | 138 | * -- Refactor delayed space reservation |
654598be ZL |
139 | * |
140 | * -- Extent-level locking | |
141 | */ | |
142 | ||
143 | static struct kmem_cache *ext4_es_cachep; | |
144 | ||
06b0c886 ZL |
145 | static int __es_insert_extent(struct ext4_es_tree *tree, |
146 | struct extent_status *newes); | |
147 | static int __es_remove_extent(struct ext4_es_tree *tree, ext4_lblk_t lblk, | |
148 | ext4_lblk_t end); | |
149 | ||
654598be ZL |
150 | int __init ext4_init_es(void) |
151 | { | |
152 | ext4_es_cachep = KMEM_CACHE(extent_status, SLAB_RECLAIM_ACCOUNT); | |
153 | if (ext4_es_cachep == NULL) | |
154 | return -ENOMEM; | |
155 | return 0; | |
156 | } | |
157 | ||
158 | void ext4_exit_es(void) | |
159 | { | |
160 | if (ext4_es_cachep) | |
161 | kmem_cache_destroy(ext4_es_cachep); | |
162 | } | |
163 | ||
164 | void ext4_es_init_tree(struct ext4_es_tree *tree) | |
165 | { | |
166 | tree->root = RB_ROOT; | |
167 | tree->cache_es = NULL; | |
168 | } | |
169 | ||
170 | #ifdef ES_DEBUG__ | |
171 | static void ext4_es_print_tree(struct inode *inode) | |
172 | { | |
173 | struct ext4_es_tree *tree; | |
174 | struct rb_node *node; | |
175 | ||
176 | printk(KERN_DEBUG "status extents for inode %lu:", inode->i_ino); | |
177 | tree = &EXT4_I(inode)->i_es_tree; | |
178 | node = rb_first(&tree->root); | |
179 | while (node) { | |
180 | struct extent_status *es; | |
181 | es = rb_entry(node, struct extent_status, rb_node); | |
06b0c886 | 182 | printk(KERN_DEBUG " [%u/%u)", es->es_lblk, es->es_len); |
654598be ZL |
183 | node = rb_next(node); |
184 | } | |
185 | printk(KERN_DEBUG "\n"); | |
186 | } | |
187 | #else | |
188 | #define ext4_es_print_tree(inode) | |
189 | #endif | |
190 | ||
06b0c886 | 191 | static inline ext4_lblk_t ext4_es_end(struct extent_status *es) |
654598be | 192 | { |
06b0c886 ZL |
193 | BUG_ON(es->es_lblk + es->es_len < es->es_lblk); |
194 | return es->es_lblk + es->es_len - 1; | |
654598be ZL |
195 | } |
196 | ||
197 | /* | |
198 | * search through the tree for an delayed extent with a given offset. If | |
199 | * it can't be found, try to find next extent. | |
200 | */ | |
201 | static struct extent_status *__es_tree_search(struct rb_root *root, | |
06b0c886 | 202 | ext4_lblk_t lblk) |
654598be ZL |
203 | { |
204 | struct rb_node *node = root->rb_node; | |
205 | struct extent_status *es = NULL; | |
206 | ||
207 | while (node) { | |
208 | es = rb_entry(node, struct extent_status, rb_node); | |
06b0c886 | 209 | if (lblk < es->es_lblk) |
654598be | 210 | node = node->rb_left; |
06b0c886 | 211 | else if (lblk > ext4_es_end(es)) |
654598be ZL |
212 | node = node->rb_right; |
213 | else | |
214 | return es; | |
215 | } | |
216 | ||
06b0c886 | 217 | if (es && lblk < es->es_lblk) |
654598be ZL |
218 | return es; |
219 | ||
06b0c886 | 220 | if (es && lblk > ext4_es_end(es)) { |
654598be ZL |
221 | node = rb_next(&es->rb_node); |
222 | return node ? rb_entry(node, struct extent_status, rb_node) : | |
223 | NULL; | |
224 | } | |
225 | ||
226 | return NULL; | |
227 | } | |
228 | ||
229 | /* | |
06b0c886 ZL |
230 | * ext4_es_find_extent: find the 1st delayed extent covering @es->lblk |
231 | * if it exists, otherwise, the next extent after @es->lblk. | |
654598be ZL |
232 | * |
233 | * @inode: the inode which owns delayed extents | |
234 | * @es: delayed extent that we found | |
235 | * | |
236 | * Returns the first block of the next extent after es, otherwise | |
237 | * EXT_MAX_BLOCKS if no delay extent is found. | |
238 | * Delayed extent is returned via @es. | |
239 | */ | |
240 | ext4_lblk_t ext4_es_find_extent(struct inode *inode, struct extent_status *es) | |
241 | { | |
242 | struct ext4_es_tree *tree = NULL; | |
243 | struct extent_status *es1 = NULL; | |
244 | struct rb_node *node; | |
245 | ext4_lblk_t ret = EXT_MAX_BLOCKS; | |
246 | ||
06b0c886 | 247 | trace_ext4_es_find_extent_enter(inode, es->es_lblk); |
992e9fdd | 248 | |
654598be ZL |
249 | read_lock(&EXT4_I(inode)->i_es_lock); |
250 | tree = &EXT4_I(inode)->i_es_tree; | |
251 | ||
252 | /* find delay extent in cache firstly */ | |
253 | if (tree->cache_es) { | |
254 | es1 = tree->cache_es; | |
06b0c886 | 255 | if (in_range(es->es_lblk, es1->es_lblk, es1->es_len)) { |
654598be | 256 | es_debug("%u cached by [%u/%u)\n", |
06b0c886 | 257 | es->es_lblk, es1->es_lblk, es1->es_len); |
654598be ZL |
258 | goto out; |
259 | } | |
260 | } | |
261 | ||
06b0c886 ZL |
262 | es->es_len = 0; |
263 | es1 = __es_tree_search(&tree->root, es->es_lblk); | |
654598be ZL |
264 | |
265 | out: | |
266 | if (es1) { | |
267 | tree->cache_es = es1; | |
06b0c886 ZL |
268 | es->es_lblk = es1->es_lblk; |
269 | es->es_len = es1->es_len; | |
654598be ZL |
270 | node = rb_next(&es1->rb_node); |
271 | if (node) { | |
272 | es1 = rb_entry(node, struct extent_status, rb_node); | |
06b0c886 | 273 | ret = es1->es_lblk; |
654598be ZL |
274 | } |
275 | } | |
276 | ||
277 | read_unlock(&EXT4_I(inode)->i_es_lock); | |
992e9fdd ZL |
278 | |
279 | trace_ext4_es_find_extent_exit(inode, es, ret); | |
654598be ZL |
280 | return ret; |
281 | } | |
282 | ||
283 | static struct extent_status * | |
06b0c886 | 284 | ext4_es_alloc_extent(ext4_lblk_t lblk, ext4_lblk_t len) |
654598be ZL |
285 | { |
286 | struct extent_status *es; | |
287 | es = kmem_cache_alloc(ext4_es_cachep, GFP_ATOMIC); | |
288 | if (es == NULL) | |
289 | return NULL; | |
06b0c886 ZL |
290 | es->es_lblk = lblk; |
291 | es->es_len = len; | |
654598be ZL |
292 | return es; |
293 | } | |
294 | ||
295 | static void ext4_es_free_extent(struct extent_status *es) | |
296 | { | |
297 | kmem_cache_free(ext4_es_cachep, es); | |
298 | } | |
299 | ||
06b0c886 ZL |
300 | /* |
301 | * Check whether or not two extents can be merged | |
302 | * Condition: | |
303 | * - logical block number is contiguous | |
304 | */ | |
305 | static int ext4_es_can_be_merged(struct extent_status *es1, | |
306 | struct extent_status *es2) | |
307 | { | |
308 | if (es1->es_lblk + es1->es_len != es2->es_lblk) | |
309 | return 0; | |
310 | ||
311 | return 1; | |
312 | } | |
313 | ||
654598be ZL |
314 | static struct extent_status * |
315 | ext4_es_try_to_merge_left(struct ext4_es_tree *tree, struct extent_status *es) | |
316 | { | |
317 | struct extent_status *es1; | |
318 | struct rb_node *node; | |
319 | ||
320 | node = rb_prev(&es->rb_node); | |
321 | if (!node) | |
322 | return es; | |
323 | ||
324 | es1 = rb_entry(node, struct extent_status, rb_node); | |
06b0c886 ZL |
325 | if (ext4_es_can_be_merged(es1, es)) { |
326 | es1->es_len += es->es_len; | |
654598be ZL |
327 | rb_erase(&es->rb_node, &tree->root); |
328 | ext4_es_free_extent(es); | |
329 | es = es1; | |
330 | } | |
331 | ||
332 | return es; | |
333 | } | |
334 | ||
335 | static struct extent_status * | |
336 | ext4_es_try_to_merge_right(struct ext4_es_tree *tree, struct extent_status *es) | |
337 | { | |
338 | struct extent_status *es1; | |
339 | struct rb_node *node; | |
340 | ||
341 | node = rb_next(&es->rb_node); | |
342 | if (!node) | |
343 | return es; | |
344 | ||
345 | es1 = rb_entry(node, struct extent_status, rb_node); | |
06b0c886 ZL |
346 | if (ext4_es_can_be_merged(es, es1)) { |
347 | es->es_len += es1->es_len; | |
654598be ZL |
348 | rb_erase(node, &tree->root); |
349 | ext4_es_free_extent(es1); | |
350 | } | |
351 | ||
352 | return es; | |
353 | } | |
354 | ||
06b0c886 ZL |
355 | static int __es_insert_extent(struct ext4_es_tree *tree, |
356 | struct extent_status *newes) | |
654598be ZL |
357 | { |
358 | struct rb_node **p = &tree->root.rb_node; | |
359 | struct rb_node *parent = NULL; | |
360 | struct extent_status *es; | |
654598be ZL |
361 | |
362 | while (*p) { | |
363 | parent = *p; | |
364 | es = rb_entry(parent, struct extent_status, rb_node); | |
365 | ||
06b0c886 ZL |
366 | if (newes->es_lblk < es->es_lblk) { |
367 | if (ext4_es_can_be_merged(newes, es)) { | |
368 | /* | |
369 | * Here we can modify es_lblk directly | |
370 | * because it isn't overlapped. | |
371 | */ | |
372 | es->es_lblk = newes->es_lblk; | |
373 | es->es_len += newes->es_len; | |
654598be ZL |
374 | es = ext4_es_try_to_merge_left(tree, es); |
375 | goto out; | |
376 | } | |
377 | p = &(*p)->rb_left; | |
06b0c886 ZL |
378 | } else if (newes->es_lblk > ext4_es_end(es)) { |
379 | if (ext4_es_can_be_merged(es, newes)) { | |
380 | es->es_len += newes->es_len; | |
654598be ZL |
381 | es = ext4_es_try_to_merge_right(tree, es); |
382 | goto out; | |
383 | } | |
384 | p = &(*p)->rb_right; | |
385 | } else { | |
06b0c886 ZL |
386 | BUG_ON(1); |
387 | return -EINVAL; | |
654598be ZL |
388 | } |
389 | } | |
390 | ||
06b0c886 | 391 | es = ext4_es_alloc_extent(newes->es_lblk, newes->es_len); |
654598be ZL |
392 | if (!es) |
393 | return -ENOMEM; | |
394 | rb_link_node(&es->rb_node, parent, p); | |
395 | rb_insert_color(&es->rb_node, &tree->root); | |
396 | ||
397 | out: | |
398 | tree->cache_es = es; | |
399 | return 0; | |
400 | } | |
401 | ||
402 | /* | |
06b0c886 | 403 | * ext4_es_insert_extent() adds a space to a extent status tree. |
654598be ZL |
404 | * |
405 | * ext4_es_insert_extent is called by ext4_da_write_begin and | |
406 | * ext4_es_remove_extent. | |
407 | * | |
408 | * Return 0 on success, error code on failure. | |
409 | */ | |
06b0c886 | 410 | int ext4_es_insert_extent(struct inode *inode, ext4_lblk_t lblk, |
654598be ZL |
411 | ext4_lblk_t len) |
412 | { | |
413 | struct ext4_es_tree *tree; | |
06b0c886 ZL |
414 | struct extent_status newes; |
415 | ext4_lblk_t end = lblk + len - 1; | |
654598be ZL |
416 | int err = 0; |
417 | ||
06b0c886 | 418 | trace_ext4_es_insert_extent(inode, lblk, len); |
654598be | 419 | es_debug("add [%u/%u) to extent status tree of inode %lu\n", |
06b0c886 ZL |
420 | lblk, len, inode->i_ino); |
421 | ||
422 | BUG_ON(end < lblk); | |
423 | ||
424 | newes.es_lblk = lblk; | |
425 | newes.es_len = len; | |
654598be ZL |
426 | |
427 | write_lock(&EXT4_I(inode)->i_es_lock); | |
428 | tree = &EXT4_I(inode)->i_es_tree; | |
06b0c886 ZL |
429 | err = __es_remove_extent(tree, lblk, end); |
430 | if (err != 0) | |
431 | goto error; | |
432 | err = __es_insert_extent(tree, &newes); | |
433 | ||
434 | error: | |
654598be ZL |
435 | write_unlock(&EXT4_I(inode)->i_es_lock); |
436 | ||
437 | ext4_es_print_tree(inode); | |
438 | ||
439 | return err; | |
440 | } | |
441 | ||
06b0c886 ZL |
442 | static int __es_remove_extent(struct ext4_es_tree *tree, ext4_lblk_t lblk, |
443 | ext4_lblk_t end) | |
654598be ZL |
444 | { |
445 | struct rb_node *node; | |
654598be ZL |
446 | struct extent_status *es; |
447 | struct extent_status orig_es; | |
06b0c886 | 448 | ext4_lblk_t len1, len2; |
654598be ZL |
449 | int err = 0; |
450 | ||
06b0c886 | 451 | es = __es_tree_search(&tree->root, lblk); |
654598be ZL |
452 | if (!es) |
453 | goto out; | |
06b0c886 | 454 | if (es->es_lblk > end) |
654598be ZL |
455 | goto out; |
456 | ||
457 | /* Simply invalidate cache_es. */ | |
458 | tree->cache_es = NULL; | |
459 | ||
06b0c886 ZL |
460 | orig_es.es_lblk = es->es_lblk; |
461 | orig_es.es_len = es->es_len; | |
462 | len1 = lblk > es->es_lblk ? lblk - es->es_lblk : 0; | |
463 | len2 = ext4_es_end(es) > end ? ext4_es_end(es) - end : 0; | |
654598be | 464 | if (len1 > 0) |
06b0c886 | 465 | es->es_len = len1; |
654598be ZL |
466 | if (len2 > 0) { |
467 | if (len1 > 0) { | |
06b0c886 ZL |
468 | struct extent_status newes; |
469 | ||
470 | newes.es_lblk = end + 1; | |
471 | newes.es_len = len2; | |
472 | err = __es_insert_extent(tree, &newes); | |
654598be | 473 | if (err) { |
06b0c886 ZL |
474 | es->es_lblk = orig_es.es_lblk; |
475 | es->es_len = orig_es.es_len; | |
654598be ZL |
476 | goto out; |
477 | } | |
478 | } else { | |
06b0c886 ZL |
479 | es->es_lblk = end + 1; |
480 | es->es_len = len2; | |
654598be ZL |
481 | } |
482 | goto out; | |
483 | } | |
484 | ||
485 | if (len1 > 0) { | |
486 | node = rb_next(&es->rb_node); | |
487 | if (node) | |
488 | es = rb_entry(node, struct extent_status, rb_node); | |
489 | else | |
490 | es = NULL; | |
491 | } | |
492 | ||
06b0c886 | 493 | while (es && ext4_es_end(es) <= end) { |
654598be ZL |
494 | node = rb_next(&es->rb_node); |
495 | rb_erase(&es->rb_node, &tree->root); | |
496 | ext4_es_free_extent(es); | |
497 | if (!node) { | |
498 | es = NULL; | |
499 | break; | |
500 | } | |
501 | es = rb_entry(node, struct extent_status, rb_node); | |
502 | } | |
503 | ||
06b0c886 ZL |
504 | if (es && es->es_lblk < end + 1) { |
505 | len1 = ext4_es_end(es) - end; | |
506 | es->es_lblk = end + 1; | |
507 | es->es_len = len1; | |
654598be ZL |
508 | } |
509 | ||
510 | out: | |
06b0c886 ZL |
511 | return err; |
512 | } | |
513 | ||
514 | /* | |
515 | * ext4_es_remove_extent() removes a space from a extent status tree. | |
516 | * | |
517 | * Return 0 on success, error code on failure. | |
518 | */ | |
519 | int ext4_es_remove_extent(struct inode *inode, ext4_lblk_t lblk, | |
520 | ext4_lblk_t len) | |
521 | { | |
522 | struct ext4_es_tree *tree; | |
523 | ext4_lblk_t end; | |
524 | int err = 0; | |
525 | ||
526 | trace_ext4_es_remove_extent(inode, lblk, len); | |
527 | es_debug("remove [%u/%u) from extent status tree of inode %lu\n", | |
528 | lblk, len, inode->i_ino); | |
529 | ||
530 | end = lblk + len - 1; | |
531 | BUG_ON(end < lblk); | |
532 | ||
533 | tree = &EXT4_I(inode)->i_es_tree; | |
534 | ||
535 | write_lock(&EXT4_I(inode)->i_es_lock); | |
536 | err = __es_remove_extent(tree, lblk, end); | |
654598be ZL |
537 | write_unlock(&EXT4_I(inode)->i_es_lock); |
538 | ext4_es_print_tree(inode); | |
539 | return err; | |
540 | } |