]>
Commit | Line | Data |
---|---|---|
7414a03f AJ |
1 | /* |
2 | * Copyright (C) 2011 STRATO. All rights reserved. | |
3 | * | |
4 | * This program is free software; you can redistribute it and/or | |
5 | * modify it under the terms of the GNU General Public | |
6 | * License v2 as published by the Free Software Foundation. | |
7 | * | |
8 | * This program is distributed in the hope that it will be useful, | |
9 | * but WITHOUT ANY WARRANTY; without even the implied warranty of | |
10 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU | |
11 | * General Public License for more details. | |
12 | * | |
13 | * You should have received a copy of the GNU General Public | |
14 | * License along with this program; if not, write to the | |
15 | * Free Software Foundation, Inc., 59 Temple Place - Suite 330, | |
16 | * Boston, MA 021110-1307, USA. | |
17 | */ | |
18 | ||
19 | #include <linux/sched.h> | |
20 | #include <linux/pagemap.h> | |
21 | #include <linux/writeback.h> | |
22 | #include <linux/blkdev.h> | |
23 | #include <linux/rbtree.h> | |
24 | #include <linux/slab.h> | |
25 | #include <linux/workqueue.h> | |
26 | #include "ctree.h" | |
27 | #include "volumes.h" | |
28 | #include "disk-io.h" | |
29 | #include "transaction.h" | |
8dabb742 | 30 | #include "dev-replace.h" |
7414a03f AJ |
31 | |
32 | #undef DEBUG | |
33 | ||
34 | /* | |
35 | * This is the implementation for the generic read ahead framework. | |
36 | * | |
37 | * To trigger a readahead, btrfs_reada_add must be called. It will start | |
38 | * a read ahead for the given range [start, end) on tree root. The returned | |
39 | * handle can either be used to wait on the readahead to finish | |
40 | * (btrfs_reada_wait), or to send it to the background (btrfs_reada_detach). | |
41 | * | |
42 | * The read ahead works as follows: | |
43 | * On btrfs_reada_add, the root of the tree is inserted into a radix_tree. | |
44 | * reada_start_machine will then search for extents to prefetch and trigger | |
45 | * some reads. When a read finishes for a node, all contained node/leaf | |
46 | * pointers that lie in the given range will also be enqueued. The reads will | |
47 | * be triggered in sequential order, thus giving a big win over a naive | |
48 | * enumeration. It will also make use of multi-device layouts. Each disk | |
49 | * will have its on read pointer and all disks will by utilized in parallel. | |
50 | * Also will no two disks read both sides of a mirror simultaneously, as this | |
51 | * would waste seeking capacity. Instead both disks will read different parts | |
52 | * of the filesystem. | |
53 | * Any number of readaheads can be started in parallel. The read order will be | |
54 | * determined globally, i.e. 2 parallel readaheads will normally finish faster | |
55 | * than the 2 started one after another. | |
56 | */ | |
57 | ||
7414a03f AJ |
58 | #define MAX_IN_FLIGHT 6 |
59 | ||
60 | struct reada_extctl { | |
61 | struct list_head list; | |
62 | struct reada_control *rc; | |
63 | u64 generation; | |
64 | }; | |
65 | ||
66 | struct reada_extent { | |
67 | u64 logical; | |
68 | struct btrfs_key top; | |
7414a03f AJ |
69 | int err; |
70 | struct list_head extctl; | |
99621b44 | 71 | int refcnt; |
7414a03f | 72 | spinlock_t lock; |
94598ba8 | 73 | struct reada_zone *zones[BTRFS_MAX_MIRRORS]; |
7414a03f AJ |
74 | int nzones; |
75 | struct btrfs_device *scheduled_for; | |
76 | }; | |
77 | ||
78 | struct reada_zone { | |
79 | u64 start; | |
80 | u64 end; | |
81 | u64 elems; | |
82 | struct list_head list; | |
83 | spinlock_t lock; | |
84 | int locked; | |
85 | struct btrfs_device *device; | |
94598ba8 SB |
86 | struct btrfs_device *devs[BTRFS_MAX_MIRRORS]; /* full list, incl |
87 | * self */ | |
7414a03f AJ |
88 | int ndevs; |
89 | struct kref refcnt; | |
90 | }; | |
91 | ||
92 | struct reada_machine_work { | |
d458b054 | 93 | struct btrfs_work work; |
7414a03f AJ |
94 | struct btrfs_fs_info *fs_info; |
95 | }; | |
96 | ||
97 | static void reada_extent_put(struct btrfs_fs_info *, struct reada_extent *); | |
98 | static void reada_control_release(struct kref *kref); | |
99 | static void reada_zone_release(struct kref *kref); | |
100 | static void reada_start_machine(struct btrfs_fs_info *fs_info); | |
101 | static void __reada_start_machine(struct btrfs_fs_info *fs_info); | |
102 | ||
103 | static int reada_add_block(struct reada_control *rc, u64 logical, | |
104 | struct btrfs_key *top, int level, u64 generation); | |
105 | ||
106 | /* recurses */ | |
107 | /* in case of err, eb might be NULL */ | |
108 | static int __readahead_hook(struct btrfs_root *root, struct extent_buffer *eb, | |
109 | u64 start, int err) | |
110 | { | |
111 | int level = 0; | |
112 | int nritems; | |
113 | int i; | |
114 | u64 bytenr; | |
115 | u64 generation; | |
116 | struct reada_extent *re; | |
117 | struct btrfs_fs_info *fs_info = root->fs_info; | |
118 | struct list_head list; | |
119 | unsigned long index = start >> PAGE_CACHE_SHIFT; | |
120 | struct btrfs_device *for_dev; | |
121 | ||
122 | if (eb) | |
123 | level = btrfs_header_level(eb); | |
124 | ||
125 | /* find extent */ | |
126 | spin_lock(&fs_info->reada_lock); | |
127 | re = radix_tree_lookup(&fs_info->reada_tree, index); | |
128 | if (re) | |
99621b44 | 129 | re->refcnt++; |
7414a03f AJ |
130 | spin_unlock(&fs_info->reada_lock); |
131 | ||
132 | if (!re) | |
133 | return -1; | |
134 | ||
135 | spin_lock(&re->lock); | |
136 | /* | |
137 | * just take the full list from the extent. afterwards we | |
138 | * don't need the lock anymore | |
139 | */ | |
140 | list_replace_init(&re->extctl, &list); | |
141 | for_dev = re->scheduled_for; | |
142 | re->scheduled_for = NULL; | |
143 | spin_unlock(&re->lock); | |
144 | ||
145 | if (err == 0) { | |
146 | nritems = level ? btrfs_header_nritems(eb) : 0; | |
147 | generation = btrfs_header_generation(eb); | |
148 | /* | |
149 | * FIXME: currently we just set nritems to 0 if this is a leaf, | |
150 | * effectively ignoring the content. In a next step we could | |
151 | * trigger more readahead depending from the content, e.g. | |
152 | * fetch the checksums for the extents in the leaf. | |
153 | */ | |
154 | } else { | |
155 | /* | |
156 | * this is the error case, the extent buffer has not been | |
157 | * read correctly. We won't access anything from it and | |
158 | * just cleanup our data structures. Effectively this will | |
159 | * cut the branch below this node from read ahead. | |
160 | */ | |
161 | nritems = 0; | |
162 | generation = 0; | |
163 | } | |
164 | ||
165 | for (i = 0; i < nritems; i++) { | |
166 | struct reada_extctl *rec; | |
167 | u64 n_gen; | |
168 | struct btrfs_key key; | |
169 | struct btrfs_key next_key; | |
170 | ||
171 | btrfs_node_key_to_cpu(eb, &key, i); | |
172 | if (i + 1 < nritems) | |
173 | btrfs_node_key_to_cpu(eb, &next_key, i + 1); | |
174 | else | |
175 | next_key = re->top; | |
176 | bytenr = btrfs_node_blockptr(eb, i); | |
177 | n_gen = btrfs_node_ptr_generation(eb, i); | |
178 | ||
179 | list_for_each_entry(rec, &list, list) { | |
180 | struct reada_control *rc = rec->rc; | |
181 | ||
182 | /* | |
183 | * if the generation doesn't match, just ignore this | |
184 | * extctl. This will probably cut off a branch from | |
185 | * prefetch. Alternatively one could start a new (sub-) | |
186 | * prefetch for this branch, starting again from root. | |
187 | * FIXME: move the generation check out of this loop | |
188 | */ | |
189 | #ifdef DEBUG | |
190 | if (rec->generation != generation) { | |
efe120a0 FH |
191 | btrfs_debug(root->fs_info, |
192 | "generation mismatch for (%llu,%d,%llu) %llu != %llu", | |
7414a03f AJ |
193 | key.objectid, key.type, key.offset, |
194 | rec->generation, generation); | |
195 | } | |
196 | #endif | |
197 | if (rec->generation == generation && | |
198 | btrfs_comp_cpu_keys(&key, &rc->key_end) < 0 && | |
199 | btrfs_comp_cpu_keys(&next_key, &rc->key_start) > 0) | |
200 | reada_add_block(rc, bytenr, &next_key, | |
201 | level - 1, n_gen); | |
202 | } | |
203 | } | |
204 | /* | |
205 | * free extctl records | |
206 | */ | |
207 | while (!list_empty(&list)) { | |
208 | struct reada_control *rc; | |
209 | struct reada_extctl *rec; | |
210 | ||
211 | rec = list_first_entry(&list, struct reada_extctl, list); | |
212 | list_del(&rec->list); | |
213 | rc = rec->rc; | |
214 | kfree(rec); | |
215 | ||
216 | kref_get(&rc->refcnt); | |
217 | if (atomic_dec_and_test(&rc->elems)) { | |
218 | kref_put(&rc->refcnt, reada_control_release); | |
219 | wake_up(&rc->wait); | |
220 | } | |
221 | kref_put(&rc->refcnt, reada_control_release); | |
222 | ||
223 | reada_extent_put(fs_info, re); /* one ref for each entry */ | |
224 | } | |
225 | reada_extent_put(fs_info, re); /* our ref */ | |
226 | if (for_dev) | |
227 | atomic_dec(&for_dev->reada_in_flight); | |
228 | ||
229 | return 0; | |
230 | } | |
231 | ||
232 | /* | |
233 | * start is passed separately in case eb in NULL, which may be the case with | |
234 | * failed I/O | |
235 | */ | |
236 | int btree_readahead_hook(struct btrfs_root *root, struct extent_buffer *eb, | |
237 | u64 start, int err) | |
238 | { | |
239 | int ret; | |
240 | ||
241 | ret = __readahead_hook(root, eb, start, err); | |
242 | ||
243 | reada_start_machine(root->fs_info); | |
244 | ||
245 | return ret; | |
246 | } | |
247 | ||
248 | static struct reada_zone *reada_find_zone(struct btrfs_fs_info *fs_info, | |
249 | struct btrfs_device *dev, u64 logical, | |
21ca543e | 250 | struct btrfs_bio *bbio) |
7414a03f AJ |
251 | { |
252 | int ret; | |
7414a03f AJ |
253 | struct reada_zone *zone; |
254 | struct btrfs_block_group_cache *cache = NULL; | |
255 | u64 start; | |
256 | u64 end; | |
257 | int i; | |
258 | ||
7414a03f AJ |
259 | zone = NULL; |
260 | spin_lock(&fs_info->reada_lock); | |
261 | ret = radix_tree_gang_lookup(&dev->reada_zones, (void **)&zone, | |
262 | logical >> PAGE_CACHE_SHIFT, 1); | |
c37f49c7 | 263 | if (ret == 1 && logical >= zone->start && logical <= zone->end) { |
7414a03f | 264 | kref_get(&zone->refcnt); |
7414a03f | 265 | spin_unlock(&fs_info->reada_lock); |
c37f49c7 | 266 | return zone; |
7414a03f AJ |
267 | } |
268 | ||
c37f49c7 ZL |
269 | spin_unlock(&fs_info->reada_lock); |
270 | ||
7414a03f AJ |
271 | cache = btrfs_lookup_block_group(fs_info, logical); |
272 | if (!cache) | |
273 | return NULL; | |
274 | ||
275 | start = cache->key.objectid; | |
276 | end = start + cache->key.offset - 1; | |
277 | btrfs_put_block_group(cache); | |
278 | ||
279 | zone = kzalloc(sizeof(*zone), GFP_NOFS); | |
280 | if (!zone) | |
281 | return NULL; | |
282 | ||
283 | zone->start = start; | |
284 | zone->end = end; | |
285 | INIT_LIST_HEAD(&zone->list); | |
286 | spin_lock_init(&zone->lock); | |
287 | zone->locked = 0; | |
288 | kref_init(&zone->refcnt); | |
289 | zone->elems = 0; | |
290 | zone->device = dev; /* our device always sits at index 0 */ | |
21ca543e | 291 | for (i = 0; i < bbio->num_stripes; ++i) { |
7414a03f | 292 | /* bounds have already been checked */ |
21ca543e | 293 | zone->devs[i] = bbio->stripes[i].dev; |
7414a03f | 294 | } |
21ca543e | 295 | zone->ndevs = bbio->num_stripes; |
7414a03f AJ |
296 | |
297 | spin_lock(&fs_info->reada_lock); | |
298 | ret = radix_tree_insert(&dev->reada_zones, | |
a175423c | 299 | (unsigned long)(zone->end >> PAGE_CACHE_SHIFT), |
7414a03f | 300 | zone); |
7414a03f | 301 | |
8c9c2bf7 | 302 | if (ret == -EEXIST) { |
7414a03f | 303 | kfree(zone); |
8c9c2bf7 AJ |
304 | ret = radix_tree_gang_lookup(&dev->reada_zones, (void **)&zone, |
305 | logical >> PAGE_CACHE_SHIFT, 1); | |
306 | if (ret == 1) | |
307 | kref_get(&zone->refcnt); | |
7414a03f | 308 | } |
8c9c2bf7 | 309 | spin_unlock(&fs_info->reada_lock); |
7414a03f AJ |
310 | |
311 | return zone; | |
312 | } | |
313 | ||
314 | static struct reada_extent *reada_find_extent(struct btrfs_root *root, | |
315 | u64 logical, | |
316 | struct btrfs_key *top, int level) | |
317 | { | |
318 | int ret; | |
7414a03f | 319 | struct reada_extent *re = NULL; |
8c9c2bf7 | 320 | struct reada_extent *re_exist = NULL; |
7414a03f | 321 | struct btrfs_fs_info *fs_info = root->fs_info; |
21ca543e | 322 | struct btrfs_bio *bbio = NULL; |
7414a03f | 323 | struct btrfs_device *dev; |
207a232c | 324 | struct btrfs_device *prev_dev; |
7414a03f AJ |
325 | u32 blocksize; |
326 | u64 length; | |
7cb2c420 | 327 | int real_stripes; |
7414a03f AJ |
328 | int nzones = 0; |
329 | int i; | |
330 | unsigned long index = logical >> PAGE_CACHE_SHIFT; | |
8dabb742 | 331 | int dev_replace_is_ongoing; |
7414a03f | 332 | |
7414a03f AJ |
333 | spin_lock(&fs_info->reada_lock); |
334 | re = radix_tree_lookup(&fs_info->reada_tree, index); | |
335 | if (re) | |
99621b44 | 336 | re->refcnt++; |
7414a03f AJ |
337 | spin_unlock(&fs_info->reada_lock); |
338 | ||
8c9c2bf7 | 339 | if (re) |
7414a03f AJ |
340 | return re; |
341 | ||
342 | re = kzalloc(sizeof(*re), GFP_NOFS); | |
343 | if (!re) | |
344 | return NULL; | |
345 | ||
707e8a07 | 346 | blocksize = root->nodesize; |
7414a03f | 347 | re->logical = logical; |
7414a03f AJ |
348 | re->top = *top; |
349 | INIT_LIST_HEAD(&re->extctl); | |
350 | spin_lock_init(&re->lock); | |
99621b44 | 351 | re->refcnt = 1; |
7414a03f AJ |
352 | |
353 | /* | |
354 | * map block | |
355 | */ | |
356 | length = blocksize; | |
29a8d9a0 SB |
357 | ret = btrfs_map_block(fs_info, REQ_GET_READ_MIRRORS, logical, &length, |
358 | &bbio, 0); | |
21ca543e | 359 | if (ret || !bbio || length < blocksize) |
7414a03f AJ |
360 | goto error; |
361 | ||
94598ba8 | 362 | if (bbio->num_stripes > BTRFS_MAX_MIRRORS) { |
efe120a0 FH |
363 | btrfs_err(root->fs_info, |
364 | "readahead: more than %d copies not supported", | |
365 | BTRFS_MAX_MIRRORS); | |
7414a03f AJ |
366 | goto error; |
367 | } | |
368 | ||
7cb2c420 OS |
369 | real_stripes = bbio->num_stripes - bbio->num_tgtdevs; |
370 | for (nzones = 0; nzones < real_stripes; ++nzones) { | |
7414a03f AJ |
371 | struct reada_zone *zone; |
372 | ||
21ca543e ID |
373 | dev = bbio->stripes[nzones].dev; |
374 | zone = reada_find_zone(fs_info, dev, logical, bbio); | |
7414a03f AJ |
375 | if (!zone) |
376 | break; | |
377 | ||
378 | re->zones[nzones] = zone; | |
379 | spin_lock(&zone->lock); | |
380 | if (!zone->elems) | |
381 | kref_get(&zone->refcnt); | |
382 | ++zone->elems; | |
383 | spin_unlock(&zone->lock); | |
384 | spin_lock(&fs_info->reada_lock); | |
385 | kref_put(&zone->refcnt, reada_zone_release); | |
386 | spin_unlock(&fs_info->reada_lock); | |
387 | } | |
388 | re->nzones = nzones; | |
389 | if (nzones == 0) { | |
390 | /* not a single zone found, error and out */ | |
391 | goto error; | |
392 | } | |
393 | ||
394 | /* insert extent in reada_tree + all per-device trees, all or nothing */ | |
8dabb742 | 395 | btrfs_dev_replace_lock(&fs_info->dev_replace); |
7414a03f AJ |
396 | spin_lock(&fs_info->reada_lock); |
397 | ret = radix_tree_insert(&fs_info->reada_tree, index, re); | |
8c9c2bf7 AJ |
398 | if (ret == -EEXIST) { |
399 | re_exist = radix_tree_lookup(&fs_info->reada_tree, index); | |
400 | BUG_ON(!re_exist); | |
99621b44 | 401 | re_exist->refcnt++; |
8c9c2bf7 | 402 | spin_unlock(&fs_info->reada_lock); |
8dabb742 | 403 | btrfs_dev_replace_unlock(&fs_info->dev_replace); |
8c9c2bf7 AJ |
404 | goto error; |
405 | } | |
7414a03f AJ |
406 | if (ret) { |
407 | spin_unlock(&fs_info->reada_lock); | |
8dabb742 | 408 | btrfs_dev_replace_unlock(&fs_info->dev_replace); |
7414a03f AJ |
409 | goto error; |
410 | } | |
207a232c | 411 | prev_dev = NULL; |
8dabb742 SB |
412 | dev_replace_is_ongoing = btrfs_dev_replace_is_ongoing( |
413 | &fs_info->dev_replace); | |
7414a03f | 414 | for (i = 0; i < nzones; ++i) { |
21ca543e | 415 | dev = bbio->stripes[i].dev; |
207a232c AJ |
416 | if (dev == prev_dev) { |
417 | /* | |
418 | * in case of DUP, just add the first zone. As both | |
419 | * are on the same device, there's nothing to gain | |
420 | * from adding both. | |
421 | * Also, it wouldn't work, as the tree is per device | |
422 | * and adding would fail with EEXIST | |
423 | */ | |
424 | continue; | |
425 | } | |
ff023aac | 426 | if (!dev->bdev) { |
5fbc7c59 WS |
427 | /* |
428 | * cannot read ahead on missing device, but for RAID5/6, | |
429 | * REQ_GET_READ_MIRRORS return 1. So don't skip missing | |
430 | * device for such case. | |
431 | */ | |
432 | if (nzones > 1) | |
433 | continue; | |
ff023aac | 434 | } |
8dabb742 SB |
435 | if (dev_replace_is_ongoing && |
436 | dev == fs_info->dev_replace.tgtdev) { | |
437 | /* | |
438 | * as this device is selected for reading only as | |
439 | * a last resort, skip it for read ahead. | |
440 | */ | |
441 | continue; | |
442 | } | |
207a232c | 443 | prev_dev = dev; |
7414a03f AJ |
444 | ret = radix_tree_insert(&dev->reada_extents, index, re); |
445 | if (ret) { | |
446 | while (--i >= 0) { | |
21ca543e | 447 | dev = bbio->stripes[i].dev; |
7414a03f | 448 | BUG_ON(dev == NULL); |
ff023aac | 449 | /* ignore whether the entry was inserted */ |
7414a03f AJ |
450 | radix_tree_delete(&dev->reada_extents, index); |
451 | } | |
452 | BUG_ON(fs_info == NULL); | |
453 | radix_tree_delete(&fs_info->reada_tree, index); | |
454 | spin_unlock(&fs_info->reada_lock); | |
8dabb742 | 455 | btrfs_dev_replace_unlock(&fs_info->dev_replace); |
7414a03f AJ |
456 | goto error; |
457 | } | |
458 | } | |
459 | spin_unlock(&fs_info->reada_lock); | |
8dabb742 | 460 | btrfs_dev_replace_unlock(&fs_info->dev_replace); |
7414a03f | 461 | |
6e9606d2 | 462 | btrfs_put_bbio(bbio); |
7414a03f AJ |
463 | return re; |
464 | ||
465 | error: | |
466 | while (nzones) { | |
467 | struct reada_zone *zone; | |
468 | ||
469 | --nzones; | |
470 | zone = re->zones[nzones]; | |
471 | kref_get(&zone->refcnt); | |
472 | spin_lock(&zone->lock); | |
473 | --zone->elems; | |
474 | if (zone->elems == 0) { | |
475 | /* | |
476 | * no fs_info->reada_lock needed, as this can't be | |
477 | * the last ref | |
478 | */ | |
479 | kref_put(&zone->refcnt, reada_zone_release); | |
480 | } | |
481 | spin_unlock(&zone->lock); | |
482 | ||
483 | spin_lock(&fs_info->reada_lock); | |
484 | kref_put(&zone->refcnt, reada_zone_release); | |
485 | spin_unlock(&fs_info->reada_lock); | |
486 | } | |
6e9606d2 | 487 | btrfs_put_bbio(bbio); |
7414a03f | 488 | kfree(re); |
8c9c2bf7 | 489 | return re_exist; |
7414a03f AJ |
490 | } |
491 | ||
7414a03f AJ |
492 | static void reada_extent_put(struct btrfs_fs_info *fs_info, |
493 | struct reada_extent *re) | |
494 | { | |
495 | int i; | |
496 | unsigned long index = re->logical >> PAGE_CACHE_SHIFT; | |
497 | ||
498 | spin_lock(&fs_info->reada_lock); | |
99621b44 | 499 | if (--re->refcnt) { |
7414a03f AJ |
500 | spin_unlock(&fs_info->reada_lock); |
501 | return; | |
502 | } | |
503 | ||
504 | radix_tree_delete(&fs_info->reada_tree, index); | |
505 | for (i = 0; i < re->nzones; ++i) { | |
506 | struct reada_zone *zone = re->zones[i]; | |
507 | ||
508 | radix_tree_delete(&zone->device->reada_extents, index); | |
509 | } | |
510 | ||
511 | spin_unlock(&fs_info->reada_lock); | |
512 | ||
513 | for (i = 0; i < re->nzones; ++i) { | |
514 | struct reada_zone *zone = re->zones[i]; | |
515 | ||
516 | kref_get(&zone->refcnt); | |
517 | spin_lock(&zone->lock); | |
518 | --zone->elems; | |
519 | if (zone->elems == 0) { | |
520 | /* no fs_info->reada_lock needed, as this can't be | |
521 | * the last ref */ | |
522 | kref_put(&zone->refcnt, reada_zone_release); | |
523 | } | |
524 | spin_unlock(&zone->lock); | |
525 | ||
526 | spin_lock(&fs_info->reada_lock); | |
527 | kref_put(&zone->refcnt, reada_zone_release); | |
528 | spin_unlock(&fs_info->reada_lock); | |
529 | } | |
530 | if (re->scheduled_for) | |
531 | atomic_dec(&re->scheduled_for->reada_in_flight); | |
532 | ||
533 | kfree(re); | |
534 | } | |
535 | ||
536 | static void reada_zone_release(struct kref *kref) | |
537 | { | |
538 | struct reada_zone *zone = container_of(kref, struct reada_zone, refcnt); | |
539 | ||
540 | radix_tree_delete(&zone->device->reada_zones, | |
541 | zone->end >> PAGE_CACHE_SHIFT); | |
542 | ||
543 | kfree(zone); | |
544 | } | |
545 | ||
546 | static void reada_control_release(struct kref *kref) | |
547 | { | |
548 | struct reada_control *rc = container_of(kref, struct reada_control, | |
549 | refcnt); | |
550 | ||
551 | kfree(rc); | |
552 | } | |
553 | ||
554 | static int reada_add_block(struct reada_control *rc, u64 logical, | |
555 | struct btrfs_key *top, int level, u64 generation) | |
556 | { | |
557 | struct btrfs_root *root = rc->root; | |
558 | struct reada_extent *re; | |
559 | struct reada_extctl *rec; | |
560 | ||
561 | re = reada_find_extent(root, logical, top, level); /* takes one ref */ | |
562 | if (!re) | |
563 | return -1; | |
564 | ||
565 | rec = kzalloc(sizeof(*rec), GFP_NOFS); | |
566 | if (!rec) { | |
567 | reada_extent_put(root->fs_info, re); | |
ddd664f4 | 568 | return -ENOMEM; |
7414a03f AJ |
569 | } |
570 | ||
571 | rec->rc = rc; | |
572 | rec->generation = generation; | |
573 | atomic_inc(&rc->elems); | |
574 | ||
575 | spin_lock(&re->lock); | |
576 | list_add_tail(&rec->list, &re->extctl); | |
577 | spin_unlock(&re->lock); | |
578 | ||
579 | /* leave the ref on the extent */ | |
580 | ||
581 | return 0; | |
582 | } | |
583 | ||
584 | /* | |
585 | * called with fs_info->reada_lock held | |
586 | */ | |
587 | static void reada_peer_zones_set_lock(struct reada_zone *zone, int lock) | |
588 | { | |
589 | int i; | |
590 | unsigned long index = zone->end >> PAGE_CACHE_SHIFT; | |
591 | ||
592 | for (i = 0; i < zone->ndevs; ++i) { | |
593 | struct reada_zone *peer; | |
594 | peer = radix_tree_lookup(&zone->devs[i]->reada_zones, index); | |
595 | if (peer && peer->device != zone->device) | |
596 | peer->locked = lock; | |
597 | } | |
598 | } | |
599 | ||
600 | /* | |
601 | * called with fs_info->reada_lock held | |
602 | */ | |
603 | static int reada_pick_zone(struct btrfs_device *dev) | |
604 | { | |
605 | struct reada_zone *top_zone = NULL; | |
606 | struct reada_zone *top_locked_zone = NULL; | |
607 | u64 top_elems = 0; | |
608 | u64 top_locked_elems = 0; | |
609 | unsigned long index = 0; | |
610 | int ret; | |
611 | ||
612 | if (dev->reada_curr_zone) { | |
613 | reada_peer_zones_set_lock(dev->reada_curr_zone, 0); | |
614 | kref_put(&dev->reada_curr_zone->refcnt, reada_zone_release); | |
615 | dev->reada_curr_zone = NULL; | |
616 | } | |
617 | /* pick the zone with the most elements */ | |
618 | while (1) { | |
619 | struct reada_zone *zone; | |
620 | ||
621 | ret = radix_tree_gang_lookup(&dev->reada_zones, | |
622 | (void **)&zone, index, 1); | |
623 | if (ret == 0) | |
624 | break; | |
625 | index = (zone->end >> PAGE_CACHE_SHIFT) + 1; | |
626 | if (zone->locked) { | |
627 | if (zone->elems > top_locked_elems) { | |
628 | top_locked_elems = zone->elems; | |
629 | top_locked_zone = zone; | |
630 | } | |
631 | } else { | |
632 | if (zone->elems > top_elems) { | |
633 | top_elems = zone->elems; | |
634 | top_zone = zone; | |
635 | } | |
636 | } | |
637 | } | |
638 | if (top_zone) | |
639 | dev->reada_curr_zone = top_zone; | |
640 | else if (top_locked_zone) | |
641 | dev->reada_curr_zone = top_locked_zone; | |
642 | else | |
643 | return 0; | |
644 | ||
645 | dev->reada_next = dev->reada_curr_zone->start; | |
646 | kref_get(&dev->reada_curr_zone->refcnt); | |
647 | reada_peer_zones_set_lock(dev->reada_curr_zone, 1); | |
648 | ||
649 | return 1; | |
650 | } | |
651 | ||
652 | static int reada_start_machine_dev(struct btrfs_fs_info *fs_info, | |
653 | struct btrfs_device *dev) | |
654 | { | |
655 | struct reada_extent *re = NULL; | |
656 | int mirror_num = 0; | |
657 | struct extent_buffer *eb = NULL; | |
658 | u64 logical; | |
7414a03f AJ |
659 | int ret; |
660 | int i; | |
661 | int need_kick = 0; | |
662 | ||
663 | spin_lock(&fs_info->reada_lock); | |
664 | if (dev->reada_curr_zone == NULL) { | |
665 | ret = reada_pick_zone(dev); | |
666 | if (!ret) { | |
667 | spin_unlock(&fs_info->reada_lock); | |
668 | return 0; | |
669 | } | |
670 | } | |
671 | /* | |
672 | * FIXME currently we issue the reads one extent at a time. If we have | |
673 | * a contiguous block of extents, we could also coagulate them or use | |
674 | * plugging to speed things up | |
675 | */ | |
676 | ret = radix_tree_gang_lookup(&dev->reada_extents, (void **)&re, | |
677 | dev->reada_next >> PAGE_CACHE_SHIFT, 1); | |
50378530 | 678 | if (ret == 0 || re->logical > dev->reada_curr_zone->end) { |
7414a03f AJ |
679 | ret = reada_pick_zone(dev); |
680 | if (!ret) { | |
681 | spin_unlock(&fs_info->reada_lock); | |
682 | return 0; | |
683 | } | |
684 | re = NULL; | |
685 | ret = radix_tree_gang_lookup(&dev->reada_extents, (void **)&re, | |
686 | dev->reada_next >> PAGE_CACHE_SHIFT, 1); | |
687 | } | |
688 | if (ret == 0) { | |
689 | spin_unlock(&fs_info->reada_lock); | |
690 | return 0; | |
691 | } | |
b6ae40ec | 692 | dev->reada_next = re->logical + fs_info->tree_root->nodesize; |
99621b44 | 693 | re->refcnt++; |
7414a03f AJ |
694 | |
695 | spin_unlock(&fs_info->reada_lock); | |
696 | ||
697 | /* | |
698 | * find mirror num | |
699 | */ | |
700 | for (i = 0; i < re->nzones; ++i) { | |
701 | if (re->zones[i]->device == dev) { | |
702 | mirror_num = i + 1; | |
703 | break; | |
704 | } | |
705 | } | |
706 | logical = re->logical; | |
7414a03f AJ |
707 | |
708 | spin_lock(&re->lock); | |
709 | if (re->scheduled_for == NULL) { | |
710 | re->scheduled_for = dev; | |
711 | need_kick = 1; | |
712 | } | |
713 | spin_unlock(&re->lock); | |
714 | ||
715 | reada_extent_put(fs_info, re); | |
716 | ||
717 | if (!need_kick) | |
718 | return 0; | |
719 | ||
720 | atomic_inc(&dev->reada_in_flight); | |
b6ae40ec | 721 | ret = reada_tree_block_flagged(fs_info->extent_root, logical, |
c0dcaa4d | 722 | mirror_num, &eb); |
7414a03f AJ |
723 | if (ret) |
724 | __readahead_hook(fs_info->extent_root, NULL, logical, ret); | |
725 | else if (eb) | |
726 | __readahead_hook(fs_info->extent_root, eb, eb->start, ret); | |
727 | ||
728 | if (eb) | |
729 | free_extent_buffer(eb); | |
730 | ||
731 | return 1; | |
732 | ||
733 | } | |
734 | ||
d458b054 | 735 | static void reada_start_machine_worker(struct btrfs_work *work) |
7414a03f AJ |
736 | { |
737 | struct reada_machine_work *rmw; | |
738 | struct btrfs_fs_info *fs_info; | |
3d136a11 | 739 | int old_ioprio; |
7414a03f AJ |
740 | |
741 | rmw = container_of(work, struct reada_machine_work, work); | |
742 | fs_info = rmw->fs_info; | |
743 | ||
744 | kfree(rmw); | |
745 | ||
3d136a11 SB |
746 | old_ioprio = IOPRIO_PRIO_VALUE(task_nice_ioclass(current), |
747 | task_nice_ioprio(current)); | |
748 | set_task_ioprio(current, BTRFS_IOPRIO_READA); | |
7414a03f | 749 | __reada_start_machine(fs_info); |
3d136a11 | 750 | set_task_ioprio(current, old_ioprio); |
7414a03f AJ |
751 | } |
752 | ||
753 | static void __reada_start_machine(struct btrfs_fs_info *fs_info) | |
754 | { | |
755 | struct btrfs_device *device; | |
756 | struct btrfs_fs_devices *fs_devices = fs_info->fs_devices; | |
757 | u64 enqueued; | |
758 | u64 total = 0; | |
759 | int i; | |
760 | ||
761 | do { | |
762 | enqueued = 0; | |
763 | list_for_each_entry(device, &fs_devices->devices, dev_list) { | |
764 | if (atomic_read(&device->reada_in_flight) < | |
765 | MAX_IN_FLIGHT) | |
766 | enqueued += reada_start_machine_dev(fs_info, | |
767 | device); | |
768 | } | |
769 | total += enqueued; | |
770 | } while (enqueued && total < 10000); | |
771 | ||
772 | if (enqueued == 0) | |
773 | return; | |
774 | ||
775 | /* | |
776 | * If everything is already in the cache, this is effectively single | |
777 | * threaded. To a) not hold the caller for too long and b) to utilize | |
778 | * more cores, we broke the loop above after 10000 iterations and now | |
779 | * enqueue to workers to finish it. This will distribute the load to | |
780 | * the cores. | |
781 | */ | |
782 | for (i = 0; i < 2; ++i) | |
783 | reada_start_machine(fs_info); | |
784 | } | |
785 | ||
786 | static void reada_start_machine(struct btrfs_fs_info *fs_info) | |
787 | { | |
788 | struct reada_machine_work *rmw; | |
789 | ||
790 | rmw = kzalloc(sizeof(*rmw), GFP_NOFS); | |
791 | if (!rmw) { | |
792 | /* FIXME we cannot handle this properly right now */ | |
793 | BUG(); | |
794 | } | |
9e0af237 LB |
795 | btrfs_init_work(&rmw->work, btrfs_readahead_helper, |
796 | reada_start_machine_worker, NULL, NULL); | |
7414a03f AJ |
797 | rmw->fs_info = fs_info; |
798 | ||
736cfa15 | 799 | btrfs_queue_work(fs_info->readahead_workers, &rmw->work); |
7414a03f AJ |
800 | } |
801 | ||
802 | #ifdef DEBUG | |
803 | static void dump_devs(struct btrfs_fs_info *fs_info, int all) | |
804 | { | |
805 | struct btrfs_device *device; | |
806 | struct btrfs_fs_devices *fs_devices = fs_info->fs_devices; | |
807 | unsigned long index; | |
808 | int ret; | |
809 | int i; | |
810 | int j; | |
811 | int cnt; | |
812 | ||
813 | spin_lock(&fs_info->reada_lock); | |
814 | list_for_each_entry(device, &fs_devices->devices, dev_list) { | |
815 | printk(KERN_DEBUG "dev %lld has %d in flight\n", device->devid, | |
816 | atomic_read(&device->reada_in_flight)); | |
817 | index = 0; | |
818 | while (1) { | |
819 | struct reada_zone *zone; | |
820 | ret = radix_tree_gang_lookup(&device->reada_zones, | |
821 | (void **)&zone, index, 1); | |
822 | if (ret == 0) | |
823 | break; | |
824 | printk(KERN_DEBUG " zone %llu-%llu elems %llu locked " | |
825 | "%d devs", zone->start, zone->end, zone->elems, | |
826 | zone->locked); | |
827 | for (j = 0; j < zone->ndevs; ++j) { | |
828 | printk(KERN_CONT " %lld", | |
829 | zone->devs[j]->devid); | |
830 | } | |
831 | if (device->reada_curr_zone == zone) | |
832 | printk(KERN_CONT " curr off %llu", | |
833 | device->reada_next - zone->start); | |
834 | printk(KERN_CONT "\n"); | |
835 | index = (zone->end >> PAGE_CACHE_SHIFT) + 1; | |
836 | } | |
837 | cnt = 0; | |
838 | index = 0; | |
839 | while (all) { | |
840 | struct reada_extent *re = NULL; | |
841 | ||
842 | ret = radix_tree_gang_lookup(&device->reada_extents, | |
843 | (void **)&re, index, 1); | |
844 | if (ret == 0) | |
845 | break; | |
846 | printk(KERN_DEBUG | |
847 | " re: logical %llu size %u empty %d for %lld", | |
b6ae40ec | 848 | re->logical, fs_info->tree_root->nodesize, |
7414a03f AJ |
849 | list_empty(&re->extctl), re->scheduled_for ? |
850 | re->scheduled_for->devid : -1); | |
851 | ||
852 | for (i = 0; i < re->nzones; ++i) { | |
853 | printk(KERN_CONT " zone %llu-%llu devs", | |
854 | re->zones[i]->start, | |
855 | re->zones[i]->end); | |
856 | for (j = 0; j < re->zones[i]->ndevs; ++j) { | |
857 | printk(KERN_CONT " %lld", | |
858 | re->zones[i]->devs[j]->devid); | |
859 | } | |
860 | } | |
861 | printk(KERN_CONT "\n"); | |
862 | index = (re->logical >> PAGE_CACHE_SHIFT) + 1; | |
863 | if (++cnt > 15) | |
864 | break; | |
865 | } | |
866 | } | |
867 | ||
868 | index = 0; | |
869 | cnt = 0; | |
870 | while (all) { | |
871 | struct reada_extent *re = NULL; | |
872 | ||
873 | ret = radix_tree_gang_lookup(&fs_info->reada_tree, (void **)&re, | |
874 | index, 1); | |
875 | if (ret == 0) | |
876 | break; | |
877 | if (!re->scheduled_for) { | |
878 | index = (re->logical >> PAGE_CACHE_SHIFT) + 1; | |
879 | continue; | |
880 | } | |
881 | printk(KERN_DEBUG | |
882 | "re: logical %llu size %u list empty %d for %lld", | |
b6ae40ec DS |
883 | re->logical, fs_info->tree_root->nodesize, |
884 | list_empty(&re->extctl), | |
7414a03f AJ |
885 | re->scheduled_for ? re->scheduled_for->devid : -1); |
886 | for (i = 0; i < re->nzones; ++i) { | |
887 | printk(KERN_CONT " zone %llu-%llu devs", | |
888 | re->zones[i]->start, | |
889 | re->zones[i]->end); | |
890 | for (i = 0; i < re->nzones; ++i) { | |
891 | printk(KERN_CONT " zone %llu-%llu devs", | |
892 | re->zones[i]->start, | |
893 | re->zones[i]->end); | |
894 | for (j = 0; j < re->zones[i]->ndevs; ++j) { | |
895 | printk(KERN_CONT " %lld", | |
896 | re->zones[i]->devs[j]->devid); | |
897 | } | |
898 | } | |
899 | } | |
900 | printk(KERN_CONT "\n"); | |
901 | index = (re->logical >> PAGE_CACHE_SHIFT) + 1; | |
902 | } | |
903 | spin_unlock(&fs_info->reada_lock); | |
904 | } | |
905 | #endif | |
906 | ||
907 | /* | |
908 | * interface | |
909 | */ | |
910 | struct reada_control *btrfs_reada_add(struct btrfs_root *root, | |
911 | struct btrfs_key *key_start, struct btrfs_key *key_end) | |
912 | { | |
913 | struct reada_control *rc; | |
914 | u64 start; | |
915 | u64 generation; | |
916 | int level; | |
ddd664f4 | 917 | int ret; |
7414a03f AJ |
918 | struct extent_buffer *node; |
919 | static struct btrfs_key max_key = { | |
920 | .objectid = (u64)-1, | |
921 | .type = (u8)-1, | |
922 | .offset = (u64)-1 | |
923 | }; | |
924 | ||
925 | rc = kzalloc(sizeof(*rc), GFP_NOFS); | |
926 | if (!rc) | |
927 | return ERR_PTR(-ENOMEM); | |
928 | ||
929 | rc->root = root; | |
930 | rc->key_start = *key_start; | |
931 | rc->key_end = *key_end; | |
932 | atomic_set(&rc->elems, 0); | |
933 | init_waitqueue_head(&rc->wait); | |
934 | kref_init(&rc->refcnt); | |
935 | kref_get(&rc->refcnt); /* one ref for having elements */ | |
936 | ||
937 | node = btrfs_root_node(root); | |
938 | start = node->start; | |
939 | level = btrfs_header_level(node); | |
940 | generation = btrfs_header_generation(node); | |
941 | free_extent_buffer(node); | |
942 | ||
ddd664f4 LB |
943 | ret = reada_add_block(rc, start, &max_key, level, generation); |
944 | if (ret) { | |
ff023aac | 945 | kfree(rc); |
ddd664f4 | 946 | return ERR_PTR(ret); |
ff023aac | 947 | } |
7414a03f AJ |
948 | |
949 | reada_start_machine(root->fs_info); | |
950 | ||
951 | return rc; | |
952 | } | |
953 | ||
954 | #ifdef DEBUG | |
955 | int btrfs_reada_wait(void *handle) | |
956 | { | |
957 | struct reada_control *rc = handle; | |
958 | ||
959 | while (atomic_read(&rc->elems)) { | |
960 | wait_event_timeout(rc->wait, atomic_read(&rc->elems) == 0, | |
961 | 5 * HZ); | |
3c59ccd3 V |
962 | dump_devs(rc->root->fs_info, |
963 | atomic_read(&rc->elems) < 10 ? 1 : 0); | |
7414a03f AJ |
964 | } |
965 | ||
3c59ccd3 | 966 | dump_devs(rc->root->fs_info, atomic_read(&rc->elems) < 10 ? 1 : 0); |
7414a03f AJ |
967 | |
968 | kref_put(&rc->refcnt, reada_control_release); | |
969 | ||
970 | return 0; | |
971 | } | |
972 | #else | |
973 | int btrfs_reada_wait(void *handle) | |
974 | { | |
975 | struct reada_control *rc = handle; | |
976 | ||
977 | while (atomic_read(&rc->elems)) { | |
978 | wait_event(rc->wait, atomic_read(&rc->elems) == 0); | |
979 | } | |
980 | ||
981 | kref_put(&rc->refcnt, reada_control_release); | |
982 | ||
983 | return 0; | |
984 | } | |
985 | #endif | |
986 | ||
987 | void btrfs_reada_detach(void *handle) | |
988 | { | |
989 | struct reada_control *rc = handle; | |
990 | ||
991 | kref_put(&rc->refcnt, reada_control_release); | |
992 | } |