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 | 74 | int nzones; |
895a11b8 | 75 | int scheduled; |
7414a03f AJ |
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, | |
1e7970c0 | 104 | struct btrfs_key *top, u64 generation); |
7414a03f AJ |
105 | |
106 | /* recurses */ | |
107 | /* in case of err, eb might be NULL */ | |
02873e43 ZL |
108 | static void __readahead_hook(struct btrfs_fs_info *fs_info, |
109 | struct reada_extent *re, struct extent_buffer *eb, | |
bcdc51b2 | 110 | int err) |
7414a03f | 111 | { |
7414a03f AJ |
112 | int nritems; |
113 | int i; | |
114 | u64 bytenr; | |
115 | u64 generation; | |
7414a03f | 116 | struct list_head list; |
7414a03f | 117 | |
7414a03f AJ |
118 | spin_lock(&re->lock); |
119 | /* | |
120 | * just take the full list from the extent. afterwards we | |
121 | * don't need the lock anymore | |
122 | */ | |
123 | list_replace_init(&re->extctl, &list); | |
895a11b8 | 124 | re->scheduled = 0; |
7414a03f AJ |
125 | spin_unlock(&re->lock); |
126 | ||
57f16e08 ZL |
127 | /* |
128 | * this is the error case, the extent buffer has not been | |
129 | * read correctly. We won't access anything from it and | |
130 | * just cleanup our data structures. Effectively this will | |
131 | * cut the branch below this node from read ahead. | |
132 | */ | |
133 | if (err) | |
134 | goto cleanup; | |
7414a03f | 135 | |
57f16e08 ZL |
136 | /* |
137 | * FIXME: currently we just set nritems to 0 if this is a leaf, | |
138 | * effectively ignoring the content. In a next step we could | |
139 | * trigger more readahead depending from the content, e.g. | |
140 | * fetch the checksums for the extents in the leaf. | |
141 | */ | |
04998b33 | 142 | if (!btrfs_header_level(eb)) |
57f16e08 ZL |
143 | goto cleanup; |
144 | ||
145 | nritems = btrfs_header_nritems(eb); | |
146 | generation = btrfs_header_generation(eb); | |
7414a03f AJ |
147 | for (i = 0; i < nritems; i++) { |
148 | struct reada_extctl *rec; | |
149 | u64 n_gen; | |
150 | struct btrfs_key key; | |
151 | struct btrfs_key next_key; | |
152 | ||
153 | btrfs_node_key_to_cpu(eb, &key, i); | |
154 | if (i + 1 < nritems) | |
155 | btrfs_node_key_to_cpu(eb, &next_key, i + 1); | |
156 | else | |
157 | next_key = re->top; | |
158 | bytenr = btrfs_node_blockptr(eb, i); | |
159 | n_gen = btrfs_node_ptr_generation(eb, i); | |
160 | ||
161 | list_for_each_entry(rec, &list, list) { | |
162 | struct reada_control *rc = rec->rc; | |
163 | ||
164 | /* | |
165 | * if the generation doesn't match, just ignore this | |
166 | * extctl. This will probably cut off a branch from | |
167 | * prefetch. Alternatively one could start a new (sub-) | |
168 | * prefetch for this branch, starting again from root. | |
169 | * FIXME: move the generation check out of this loop | |
170 | */ | |
171 | #ifdef DEBUG | |
172 | if (rec->generation != generation) { | |
02873e43 ZL |
173 | btrfs_debug(fs_info, |
174 | "generation mismatch for (%llu,%d,%llu) %llu != %llu", | |
175 | key.objectid, key.type, key.offset, | |
176 | rec->generation, generation); | |
7414a03f AJ |
177 | } |
178 | #endif | |
179 | if (rec->generation == generation && | |
180 | btrfs_comp_cpu_keys(&key, &rc->key_end) < 0 && | |
181 | btrfs_comp_cpu_keys(&next_key, &rc->key_start) > 0) | |
1e7970c0 | 182 | reada_add_block(rc, bytenr, &next_key, n_gen); |
7414a03f AJ |
183 | } |
184 | } | |
57f16e08 ZL |
185 | |
186 | cleanup: | |
7414a03f AJ |
187 | /* |
188 | * free extctl records | |
189 | */ | |
190 | while (!list_empty(&list)) { | |
191 | struct reada_control *rc; | |
192 | struct reada_extctl *rec; | |
193 | ||
194 | rec = list_first_entry(&list, struct reada_extctl, list); | |
195 | list_del(&rec->list); | |
196 | rc = rec->rc; | |
197 | kfree(rec); | |
198 | ||
199 | kref_get(&rc->refcnt); | |
200 | if (atomic_dec_and_test(&rc->elems)) { | |
201 | kref_put(&rc->refcnt, reada_control_release); | |
202 | wake_up(&rc->wait); | |
203 | } | |
204 | kref_put(&rc->refcnt, reada_control_release); | |
205 | ||
206 | reada_extent_put(fs_info, re); /* one ref for each entry */ | |
207 | } | |
7414a03f | 208 | |
6e39dbe8 | 209 | return; |
7414a03f AJ |
210 | } |
211 | ||
02873e43 | 212 | int btree_readahead_hook(struct btrfs_fs_info *fs_info, |
fc2e901f | 213 | struct extent_buffer *eb, int err) |
7414a03f | 214 | { |
6e39dbe8 ZL |
215 | int ret = 0; |
216 | struct reada_extent *re; | |
7414a03f | 217 | |
6e39dbe8 ZL |
218 | /* find extent */ |
219 | spin_lock(&fs_info->reada_lock); | |
220 | re = radix_tree_lookup(&fs_info->reada_tree, | |
fc2e901f | 221 | eb->start >> PAGE_SHIFT); |
6e39dbe8 ZL |
222 | if (re) |
223 | re->refcnt++; | |
224 | spin_unlock(&fs_info->reada_lock); | |
225 | if (!re) { | |
226 | ret = -1; | |
227 | goto start_machine; | |
228 | } | |
7414a03f | 229 | |
bcdc51b2 | 230 | __readahead_hook(fs_info, re, eb, err); |
6e39dbe8 | 231 | reada_extent_put(fs_info, re); /* our ref */ |
7414a03f | 232 | |
6e39dbe8 ZL |
233 | start_machine: |
234 | reada_start_machine(fs_info); | |
7414a03f AJ |
235 | return ret; |
236 | } | |
237 | ||
238 | static struct reada_zone *reada_find_zone(struct btrfs_fs_info *fs_info, | |
239 | struct btrfs_device *dev, u64 logical, | |
21ca543e | 240 | struct btrfs_bio *bbio) |
7414a03f AJ |
241 | { |
242 | int ret; | |
7414a03f AJ |
243 | struct reada_zone *zone; |
244 | struct btrfs_block_group_cache *cache = NULL; | |
245 | u64 start; | |
246 | u64 end; | |
247 | int i; | |
248 | ||
7414a03f AJ |
249 | zone = NULL; |
250 | spin_lock(&fs_info->reada_lock); | |
251 | ret = radix_tree_gang_lookup(&dev->reada_zones, (void **)&zone, | |
09cbfeaf | 252 | logical >> PAGE_SHIFT, 1); |
c37f49c7 | 253 | if (ret == 1 && logical >= zone->start && logical <= zone->end) { |
7414a03f | 254 | kref_get(&zone->refcnt); |
7414a03f | 255 | spin_unlock(&fs_info->reada_lock); |
c37f49c7 | 256 | return zone; |
7414a03f AJ |
257 | } |
258 | ||
c37f49c7 ZL |
259 | spin_unlock(&fs_info->reada_lock); |
260 | ||
7414a03f AJ |
261 | cache = btrfs_lookup_block_group(fs_info, logical); |
262 | if (!cache) | |
263 | return NULL; | |
264 | ||
265 | start = cache->key.objectid; | |
266 | end = start + cache->key.offset - 1; | |
267 | btrfs_put_block_group(cache); | |
268 | ||
ed0244fa | 269 | zone = kzalloc(sizeof(*zone), GFP_KERNEL); |
7414a03f AJ |
270 | if (!zone) |
271 | return NULL; | |
272 | ||
cc8385b5 DS |
273 | ret = radix_tree_preload(GFP_KERNEL); |
274 | if (ret) { | |
275 | kfree(zone); | |
276 | return NULL; | |
277 | } | |
278 | ||
7414a03f AJ |
279 | zone->start = start; |
280 | zone->end = end; | |
281 | INIT_LIST_HEAD(&zone->list); | |
282 | spin_lock_init(&zone->lock); | |
283 | zone->locked = 0; | |
284 | kref_init(&zone->refcnt); | |
285 | zone->elems = 0; | |
286 | zone->device = dev; /* our device always sits at index 0 */ | |
21ca543e | 287 | for (i = 0; i < bbio->num_stripes; ++i) { |
7414a03f | 288 | /* bounds have already been checked */ |
21ca543e | 289 | zone->devs[i] = bbio->stripes[i].dev; |
7414a03f | 290 | } |
21ca543e | 291 | zone->ndevs = bbio->num_stripes; |
7414a03f AJ |
292 | |
293 | spin_lock(&fs_info->reada_lock); | |
294 | ret = radix_tree_insert(&dev->reada_zones, | |
09cbfeaf | 295 | (unsigned long)(zone->end >> PAGE_SHIFT), |
7414a03f | 296 | zone); |
7414a03f | 297 | |
8c9c2bf7 | 298 | if (ret == -EEXIST) { |
7414a03f | 299 | kfree(zone); |
8c9c2bf7 | 300 | ret = radix_tree_gang_lookup(&dev->reada_zones, (void **)&zone, |
09cbfeaf | 301 | logical >> PAGE_SHIFT, 1); |
8e9aa51f | 302 | if (ret == 1 && logical >= zone->start && logical <= zone->end) |
8c9c2bf7 | 303 | kref_get(&zone->refcnt); |
8e9aa51f ZL |
304 | else |
305 | zone = NULL; | |
7414a03f | 306 | } |
8c9c2bf7 | 307 | spin_unlock(&fs_info->reada_lock); |
cc8385b5 | 308 | radix_tree_preload_end(); |
7414a03f AJ |
309 | |
310 | return zone; | |
311 | } | |
312 | ||
2ff7e61e | 313 | static struct reada_extent *reada_find_extent(struct btrfs_fs_info *fs_info, |
7414a03f | 314 | u64 logical, |
1e7970c0 | 315 | struct btrfs_key *top) |
7414a03f AJ |
316 | { |
317 | int ret; | |
7414a03f | 318 | struct reada_extent *re = NULL; |
8c9c2bf7 | 319 | struct reada_extent *re_exist = NULL; |
21ca543e | 320 | struct btrfs_bio *bbio = NULL; |
7414a03f | 321 | struct btrfs_device *dev; |
207a232c | 322 | struct btrfs_device *prev_dev; |
7414a03f AJ |
323 | u32 blocksize; |
324 | u64 length; | |
7cb2c420 | 325 | int real_stripes; |
7414a03f | 326 | int nzones = 0; |
09cbfeaf | 327 | unsigned long index = logical >> PAGE_SHIFT; |
8dabb742 | 328 | int dev_replace_is_ongoing; |
31945021 | 329 | int have_zone = 0; |
7414a03f | 330 | |
7414a03f AJ |
331 | spin_lock(&fs_info->reada_lock); |
332 | re = radix_tree_lookup(&fs_info->reada_tree, index); | |
333 | if (re) | |
99621b44 | 334 | re->refcnt++; |
7414a03f AJ |
335 | spin_unlock(&fs_info->reada_lock); |
336 | ||
8c9c2bf7 | 337 | if (re) |
7414a03f AJ |
338 | return re; |
339 | ||
ed0244fa | 340 | re = kzalloc(sizeof(*re), GFP_KERNEL); |
7414a03f AJ |
341 | if (!re) |
342 | return NULL; | |
343 | ||
0b246afa | 344 | blocksize = fs_info->nodesize; |
7414a03f | 345 | re->logical = logical; |
7414a03f AJ |
346 | re->top = *top; |
347 | INIT_LIST_HEAD(&re->extctl); | |
348 | spin_lock_init(&re->lock); | |
99621b44 | 349 | re->refcnt = 1; |
7414a03f AJ |
350 | |
351 | /* | |
352 | * map block | |
353 | */ | |
354 | length = blocksize; | |
cf8cddd3 CH |
355 | ret = btrfs_map_block(fs_info, BTRFS_MAP_GET_READ_MIRRORS, logical, |
356 | &length, &bbio, 0); | |
21ca543e | 357 | if (ret || !bbio || length < blocksize) |
7414a03f AJ |
358 | goto error; |
359 | ||
94598ba8 | 360 | if (bbio->num_stripes > BTRFS_MAX_MIRRORS) { |
0b246afa | 361 | btrfs_err(fs_info, |
efe120a0 FH |
362 | "readahead: more than %d copies not supported", |
363 | BTRFS_MAX_MIRRORS); | |
7414a03f AJ |
364 | goto error; |
365 | } | |
366 | ||
7cb2c420 OS |
367 | real_stripes = bbio->num_stripes - bbio->num_tgtdevs; |
368 | for (nzones = 0; nzones < real_stripes; ++nzones) { | |
7414a03f AJ |
369 | struct reada_zone *zone; |
370 | ||
21ca543e | 371 | dev = bbio->stripes[nzones].dev; |
7aff8cf4 ZL |
372 | |
373 | /* cannot read ahead on missing device. */ | |
374 | if (!dev->bdev) | |
375 | continue; | |
376 | ||
21ca543e | 377 | zone = reada_find_zone(fs_info, dev, logical, bbio); |
7414a03f | 378 | if (!zone) |
6a159d2a | 379 | continue; |
7414a03f | 380 | |
6a159d2a | 381 | re->zones[re->nzones++] = zone; |
7414a03f AJ |
382 | spin_lock(&zone->lock); |
383 | if (!zone->elems) | |
384 | kref_get(&zone->refcnt); | |
385 | ++zone->elems; | |
386 | spin_unlock(&zone->lock); | |
387 | spin_lock(&fs_info->reada_lock); | |
388 | kref_put(&zone->refcnt, reada_zone_release); | |
389 | spin_unlock(&fs_info->reada_lock); | |
390 | } | |
6a159d2a | 391 | if (re->nzones == 0) { |
7414a03f AJ |
392 | /* not a single zone found, error and out */ |
393 | goto error; | |
394 | } | |
395 | ||
7ef70b4d DS |
396 | ret = radix_tree_preload(GFP_KERNEL); |
397 | if (ret) | |
398 | goto error; | |
399 | ||
7414a03f | 400 | /* insert extent in reada_tree + all per-device trees, all or nothing */ |
73beece9 | 401 | btrfs_dev_replace_lock(&fs_info->dev_replace, 0); |
7414a03f AJ |
402 | spin_lock(&fs_info->reada_lock); |
403 | ret = radix_tree_insert(&fs_info->reada_tree, index, re); | |
8c9c2bf7 AJ |
404 | if (ret == -EEXIST) { |
405 | re_exist = radix_tree_lookup(&fs_info->reada_tree, index); | |
99621b44 | 406 | re_exist->refcnt++; |
8c9c2bf7 | 407 | spin_unlock(&fs_info->reada_lock); |
73beece9 | 408 | btrfs_dev_replace_unlock(&fs_info->dev_replace, 0); |
7ef70b4d | 409 | radix_tree_preload_end(); |
8c9c2bf7 AJ |
410 | goto error; |
411 | } | |
7414a03f AJ |
412 | if (ret) { |
413 | spin_unlock(&fs_info->reada_lock); | |
73beece9 | 414 | btrfs_dev_replace_unlock(&fs_info->dev_replace, 0); |
7ef70b4d | 415 | radix_tree_preload_end(); |
7414a03f AJ |
416 | goto error; |
417 | } | |
7ef70b4d | 418 | radix_tree_preload_end(); |
207a232c | 419 | prev_dev = NULL; |
8dabb742 SB |
420 | dev_replace_is_ongoing = btrfs_dev_replace_is_ongoing( |
421 | &fs_info->dev_replace); | |
6a159d2a ZL |
422 | for (nzones = 0; nzones < re->nzones; ++nzones) { |
423 | dev = re->zones[nzones]->device; | |
424 | ||
207a232c AJ |
425 | if (dev == prev_dev) { |
426 | /* | |
427 | * in case of DUP, just add the first zone. As both | |
428 | * are on the same device, there's nothing to gain | |
429 | * from adding both. | |
430 | * Also, it wouldn't work, as the tree is per device | |
431 | * and adding would fail with EEXIST | |
432 | */ | |
433 | continue; | |
434 | } | |
7aff8cf4 ZL |
435 | if (!dev->bdev) |
436 | continue; | |
437 | ||
8dabb742 SB |
438 | if (dev_replace_is_ongoing && |
439 | dev == fs_info->dev_replace.tgtdev) { | |
440 | /* | |
441 | * as this device is selected for reading only as | |
442 | * a last resort, skip it for read ahead. | |
443 | */ | |
444 | continue; | |
445 | } | |
207a232c | 446 | prev_dev = dev; |
7414a03f AJ |
447 | ret = radix_tree_insert(&dev->reada_extents, index, re); |
448 | if (ret) { | |
6a159d2a ZL |
449 | while (--nzones >= 0) { |
450 | dev = re->zones[nzones]->device; | |
7414a03f | 451 | BUG_ON(dev == NULL); |
ff023aac | 452 | /* ignore whether the entry was inserted */ |
7414a03f AJ |
453 | radix_tree_delete(&dev->reada_extents, index); |
454 | } | |
7414a03f AJ |
455 | radix_tree_delete(&fs_info->reada_tree, index); |
456 | spin_unlock(&fs_info->reada_lock); | |
73beece9 | 457 | btrfs_dev_replace_unlock(&fs_info->dev_replace, 0); |
7414a03f AJ |
458 | goto error; |
459 | } | |
31945021 | 460 | have_zone = 1; |
7414a03f AJ |
461 | } |
462 | spin_unlock(&fs_info->reada_lock); | |
73beece9 | 463 | btrfs_dev_replace_unlock(&fs_info->dev_replace, 0); |
7414a03f | 464 | |
31945021 ZL |
465 | if (!have_zone) |
466 | goto error; | |
467 | ||
6e9606d2 | 468 | btrfs_put_bbio(bbio); |
7414a03f AJ |
469 | return re; |
470 | ||
471 | error: | |
6a159d2a | 472 | for (nzones = 0; nzones < re->nzones; ++nzones) { |
7414a03f AJ |
473 | struct reada_zone *zone; |
474 | ||
7414a03f AJ |
475 | zone = re->zones[nzones]; |
476 | kref_get(&zone->refcnt); | |
477 | spin_lock(&zone->lock); | |
478 | --zone->elems; | |
479 | if (zone->elems == 0) { | |
480 | /* | |
481 | * no fs_info->reada_lock needed, as this can't be | |
482 | * the last ref | |
483 | */ | |
484 | kref_put(&zone->refcnt, reada_zone_release); | |
485 | } | |
486 | spin_unlock(&zone->lock); | |
487 | ||
488 | spin_lock(&fs_info->reada_lock); | |
489 | kref_put(&zone->refcnt, reada_zone_release); | |
490 | spin_unlock(&fs_info->reada_lock); | |
491 | } | |
6e9606d2 | 492 | btrfs_put_bbio(bbio); |
7414a03f | 493 | kfree(re); |
8c9c2bf7 | 494 | return re_exist; |
7414a03f AJ |
495 | } |
496 | ||
7414a03f AJ |
497 | static void reada_extent_put(struct btrfs_fs_info *fs_info, |
498 | struct reada_extent *re) | |
499 | { | |
500 | int i; | |
09cbfeaf | 501 | unsigned long index = re->logical >> PAGE_SHIFT; |
7414a03f AJ |
502 | |
503 | spin_lock(&fs_info->reada_lock); | |
99621b44 | 504 | if (--re->refcnt) { |
7414a03f AJ |
505 | spin_unlock(&fs_info->reada_lock); |
506 | return; | |
507 | } | |
508 | ||
509 | radix_tree_delete(&fs_info->reada_tree, index); | |
510 | for (i = 0; i < re->nzones; ++i) { | |
511 | struct reada_zone *zone = re->zones[i]; | |
512 | ||
513 | radix_tree_delete(&zone->device->reada_extents, index); | |
514 | } | |
515 | ||
516 | spin_unlock(&fs_info->reada_lock); | |
517 | ||
518 | for (i = 0; i < re->nzones; ++i) { | |
519 | struct reada_zone *zone = re->zones[i]; | |
520 | ||
521 | kref_get(&zone->refcnt); | |
522 | spin_lock(&zone->lock); | |
523 | --zone->elems; | |
524 | if (zone->elems == 0) { | |
525 | /* no fs_info->reada_lock needed, as this can't be | |
526 | * the last ref */ | |
527 | kref_put(&zone->refcnt, reada_zone_release); | |
528 | } | |
529 | spin_unlock(&zone->lock); | |
530 | ||
531 | spin_lock(&fs_info->reada_lock); | |
532 | kref_put(&zone->refcnt, reada_zone_release); | |
533 | spin_unlock(&fs_info->reada_lock); | |
534 | } | |
7414a03f AJ |
535 | |
536 | kfree(re); | |
537 | } | |
538 | ||
539 | static void reada_zone_release(struct kref *kref) | |
540 | { | |
541 | struct reada_zone *zone = container_of(kref, struct reada_zone, refcnt); | |
542 | ||
543 | radix_tree_delete(&zone->device->reada_zones, | |
09cbfeaf | 544 | zone->end >> PAGE_SHIFT); |
7414a03f AJ |
545 | |
546 | kfree(zone); | |
547 | } | |
548 | ||
549 | static void reada_control_release(struct kref *kref) | |
550 | { | |
551 | struct reada_control *rc = container_of(kref, struct reada_control, | |
552 | refcnt); | |
553 | ||
554 | kfree(rc); | |
555 | } | |
556 | ||
557 | static int reada_add_block(struct reada_control *rc, u64 logical, | |
1e7970c0 | 558 | struct btrfs_key *top, u64 generation) |
7414a03f | 559 | { |
c28f158e | 560 | struct btrfs_fs_info *fs_info = rc->fs_info; |
7414a03f AJ |
561 | struct reada_extent *re; |
562 | struct reada_extctl *rec; | |
563 | ||
c28f158e | 564 | /* takes one ref */ |
2ff7e61e | 565 | re = reada_find_extent(fs_info, logical, top); |
7414a03f AJ |
566 | if (!re) |
567 | return -1; | |
568 | ||
ed0244fa | 569 | rec = kzalloc(sizeof(*rec), GFP_KERNEL); |
7414a03f | 570 | if (!rec) { |
c28f158e | 571 | reada_extent_put(fs_info, re); |
ddd664f4 | 572 | return -ENOMEM; |
7414a03f AJ |
573 | } |
574 | ||
575 | rec->rc = rc; | |
576 | rec->generation = generation; | |
577 | atomic_inc(&rc->elems); | |
578 | ||
579 | spin_lock(&re->lock); | |
580 | list_add_tail(&rec->list, &re->extctl); | |
581 | spin_unlock(&re->lock); | |
582 | ||
583 | /* leave the ref on the extent */ | |
584 | ||
585 | return 0; | |
586 | } | |
587 | ||
588 | /* | |
589 | * called with fs_info->reada_lock held | |
590 | */ | |
591 | static void reada_peer_zones_set_lock(struct reada_zone *zone, int lock) | |
592 | { | |
593 | int i; | |
09cbfeaf | 594 | unsigned long index = zone->end >> PAGE_SHIFT; |
7414a03f AJ |
595 | |
596 | for (i = 0; i < zone->ndevs; ++i) { | |
597 | struct reada_zone *peer; | |
598 | peer = radix_tree_lookup(&zone->devs[i]->reada_zones, index); | |
599 | if (peer && peer->device != zone->device) | |
600 | peer->locked = lock; | |
601 | } | |
602 | } | |
603 | ||
604 | /* | |
605 | * called with fs_info->reada_lock held | |
606 | */ | |
607 | static int reada_pick_zone(struct btrfs_device *dev) | |
608 | { | |
609 | struct reada_zone *top_zone = NULL; | |
610 | struct reada_zone *top_locked_zone = NULL; | |
611 | u64 top_elems = 0; | |
612 | u64 top_locked_elems = 0; | |
613 | unsigned long index = 0; | |
614 | int ret; | |
615 | ||
616 | if (dev->reada_curr_zone) { | |
617 | reada_peer_zones_set_lock(dev->reada_curr_zone, 0); | |
618 | kref_put(&dev->reada_curr_zone->refcnt, reada_zone_release); | |
619 | dev->reada_curr_zone = NULL; | |
620 | } | |
621 | /* pick the zone with the most elements */ | |
622 | while (1) { | |
623 | struct reada_zone *zone; | |
624 | ||
625 | ret = radix_tree_gang_lookup(&dev->reada_zones, | |
626 | (void **)&zone, index, 1); | |
627 | if (ret == 0) | |
628 | break; | |
09cbfeaf | 629 | index = (zone->end >> PAGE_SHIFT) + 1; |
7414a03f AJ |
630 | if (zone->locked) { |
631 | if (zone->elems > top_locked_elems) { | |
632 | top_locked_elems = zone->elems; | |
633 | top_locked_zone = zone; | |
634 | } | |
635 | } else { | |
636 | if (zone->elems > top_elems) { | |
637 | top_elems = zone->elems; | |
638 | top_zone = zone; | |
639 | } | |
640 | } | |
641 | } | |
642 | if (top_zone) | |
643 | dev->reada_curr_zone = top_zone; | |
644 | else if (top_locked_zone) | |
645 | dev->reada_curr_zone = top_locked_zone; | |
646 | else | |
647 | return 0; | |
648 | ||
649 | dev->reada_next = dev->reada_curr_zone->start; | |
650 | kref_get(&dev->reada_curr_zone->refcnt); | |
651 | reada_peer_zones_set_lock(dev->reada_curr_zone, 1); | |
652 | ||
653 | return 1; | |
654 | } | |
655 | ||
656 | static int reada_start_machine_dev(struct btrfs_fs_info *fs_info, | |
657 | struct btrfs_device *dev) | |
658 | { | |
659 | struct reada_extent *re = NULL; | |
660 | int mirror_num = 0; | |
661 | struct extent_buffer *eb = NULL; | |
662 | u64 logical; | |
7414a03f AJ |
663 | int ret; |
664 | int i; | |
7414a03f AJ |
665 | |
666 | spin_lock(&fs_info->reada_lock); | |
667 | if (dev->reada_curr_zone == NULL) { | |
668 | ret = reada_pick_zone(dev); | |
669 | if (!ret) { | |
670 | spin_unlock(&fs_info->reada_lock); | |
671 | return 0; | |
672 | } | |
673 | } | |
674 | /* | |
675 | * FIXME currently we issue the reads one extent at a time. If we have | |
676 | * a contiguous block of extents, we could also coagulate them or use | |
677 | * plugging to speed things up | |
678 | */ | |
679 | ret = radix_tree_gang_lookup(&dev->reada_extents, (void **)&re, | |
09cbfeaf | 680 | dev->reada_next >> PAGE_SHIFT, 1); |
50378530 | 681 | if (ret == 0 || re->logical > dev->reada_curr_zone->end) { |
7414a03f AJ |
682 | ret = reada_pick_zone(dev); |
683 | if (!ret) { | |
684 | spin_unlock(&fs_info->reada_lock); | |
685 | return 0; | |
686 | } | |
687 | re = NULL; | |
688 | ret = radix_tree_gang_lookup(&dev->reada_extents, (void **)&re, | |
09cbfeaf | 689 | dev->reada_next >> PAGE_SHIFT, 1); |
7414a03f AJ |
690 | } |
691 | if (ret == 0) { | |
692 | spin_unlock(&fs_info->reada_lock); | |
693 | return 0; | |
694 | } | |
da17066c | 695 | dev->reada_next = re->logical + fs_info->nodesize; |
99621b44 | 696 | re->refcnt++; |
7414a03f AJ |
697 | |
698 | spin_unlock(&fs_info->reada_lock); | |
699 | ||
a3f7fde2 | 700 | spin_lock(&re->lock); |
895a11b8 | 701 | if (re->scheduled || list_empty(&re->extctl)) { |
a3f7fde2 ZL |
702 | spin_unlock(&re->lock); |
703 | reada_extent_put(fs_info, re); | |
704 | return 0; | |
705 | } | |
895a11b8 | 706 | re->scheduled = 1; |
a3f7fde2 ZL |
707 | spin_unlock(&re->lock); |
708 | ||
7414a03f AJ |
709 | /* |
710 | * find mirror num | |
711 | */ | |
712 | for (i = 0; i < re->nzones; ++i) { | |
713 | if (re->zones[i]->device == dev) { | |
714 | mirror_num = i + 1; | |
715 | break; | |
716 | } | |
717 | } | |
718 | logical = re->logical; | |
7414a03f | 719 | |
7414a03f | 720 | atomic_inc(&dev->reada_in_flight); |
2ff7e61e | 721 | ret = reada_tree_block_flagged(fs_info, logical, mirror_num, &eb); |
7414a03f | 722 | if (ret) |
bcdc51b2 | 723 | __readahead_hook(fs_info, re, NULL, ret); |
7414a03f | 724 | else if (eb) |
bcdc51b2 | 725 | __readahead_hook(fs_info, re, eb, ret); |
7414a03f AJ |
726 | |
727 | if (eb) | |
728 | free_extent_buffer(eb); | |
729 | ||
895a11b8 | 730 | atomic_dec(&dev->reada_in_flight); |
b257cf50 ZL |
731 | reada_extent_put(fs_info, re); |
732 | ||
7414a03f AJ |
733 | return 1; |
734 | ||
735 | } | |
736 | ||
d458b054 | 737 | static void reada_start_machine_worker(struct btrfs_work *work) |
7414a03f AJ |
738 | { |
739 | struct reada_machine_work *rmw; | |
740 | struct btrfs_fs_info *fs_info; | |
3d136a11 | 741 | int old_ioprio; |
7414a03f AJ |
742 | |
743 | rmw = container_of(work, struct reada_machine_work, work); | |
744 | fs_info = rmw->fs_info; | |
745 | ||
746 | kfree(rmw); | |
747 | ||
3d136a11 SB |
748 | old_ioprio = IOPRIO_PRIO_VALUE(task_nice_ioclass(current), |
749 | task_nice_ioprio(current)); | |
750 | set_task_ioprio(current, BTRFS_IOPRIO_READA); | |
7414a03f | 751 | __reada_start_machine(fs_info); |
3d136a11 | 752 | set_task_ioprio(current, old_ioprio); |
2fefd558 ZL |
753 | |
754 | atomic_dec(&fs_info->reada_works_cnt); | |
7414a03f AJ |
755 | } |
756 | ||
757 | static void __reada_start_machine(struct btrfs_fs_info *fs_info) | |
758 | { | |
759 | struct btrfs_device *device; | |
760 | struct btrfs_fs_devices *fs_devices = fs_info->fs_devices; | |
761 | u64 enqueued; | |
762 | u64 total = 0; | |
763 | int i; | |
764 | ||
765 | do { | |
766 | enqueued = 0; | |
ce7791ff | 767 | mutex_lock(&fs_devices->device_list_mutex); |
7414a03f AJ |
768 | list_for_each_entry(device, &fs_devices->devices, dev_list) { |
769 | if (atomic_read(&device->reada_in_flight) < | |
770 | MAX_IN_FLIGHT) | |
771 | enqueued += reada_start_machine_dev(fs_info, | |
772 | device); | |
773 | } | |
ce7791ff | 774 | mutex_unlock(&fs_devices->device_list_mutex); |
7414a03f AJ |
775 | total += enqueued; |
776 | } while (enqueued && total < 10000); | |
777 | ||
778 | if (enqueued == 0) | |
779 | return; | |
780 | ||
781 | /* | |
782 | * If everything is already in the cache, this is effectively single | |
783 | * threaded. To a) not hold the caller for too long and b) to utilize | |
784 | * more cores, we broke the loop above after 10000 iterations and now | |
785 | * enqueue to workers to finish it. This will distribute the load to | |
786 | * the cores. | |
787 | */ | |
2fefd558 | 788 | for (i = 0; i < 2; ++i) { |
7414a03f | 789 | reada_start_machine(fs_info); |
2fefd558 ZL |
790 | if (atomic_read(&fs_info->reada_works_cnt) > |
791 | BTRFS_MAX_MIRRORS * 2) | |
792 | break; | |
793 | } | |
7414a03f AJ |
794 | } |
795 | ||
796 | static void reada_start_machine(struct btrfs_fs_info *fs_info) | |
797 | { | |
798 | struct reada_machine_work *rmw; | |
799 | ||
ed0244fa | 800 | rmw = kzalloc(sizeof(*rmw), GFP_KERNEL); |
7414a03f AJ |
801 | if (!rmw) { |
802 | /* FIXME we cannot handle this properly right now */ | |
803 | BUG(); | |
804 | } | |
9e0af237 LB |
805 | btrfs_init_work(&rmw->work, btrfs_readahead_helper, |
806 | reada_start_machine_worker, NULL, NULL); | |
7414a03f AJ |
807 | rmw->fs_info = fs_info; |
808 | ||
736cfa15 | 809 | btrfs_queue_work(fs_info->readahead_workers, &rmw->work); |
2fefd558 | 810 | atomic_inc(&fs_info->reada_works_cnt); |
7414a03f AJ |
811 | } |
812 | ||
813 | #ifdef DEBUG | |
814 | static void dump_devs(struct btrfs_fs_info *fs_info, int all) | |
815 | { | |
816 | struct btrfs_device *device; | |
817 | struct btrfs_fs_devices *fs_devices = fs_info->fs_devices; | |
818 | unsigned long index; | |
819 | int ret; | |
820 | int i; | |
821 | int j; | |
822 | int cnt; | |
823 | ||
824 | spin_lock(&fs_info->reada_lock); | |
825 | list_for_each_entry(device, &fs_devices->devices, dev_list) { | |
ab8d0fc4 | 826 | btrfs_debug(fs_info, "dev %lld has %d in flight", device->devid, |
7414a03f AJ |
827 | atomic_read(&device->reada_in_flight)); |
828 | index = 0; | |
829 | while (1) { | |
830 | struct reada_zone *zone; | |
831 | ret = radix_tree_gang_lookup(&device->reada_zones, | |
832 | (void **)&zone, index, 1); | |
833 | if (ret == 0) | |
834 | break; | |
62e85577 | 835 | pr_debug(" zone %llu-%llu elems %llu locked %d devs", |
ab8d0fc4 JM |
836 | zone->start, zone->end, zone->elems, |
837 | zone->locked); | |
7414a03f | 838 | for (j = 0; j < zone->ndevs; ++j) { |
62e85577 | 839 | pr_cont(" %lld", |
7414a03f AJ |
840 | zone->devs[j]->devid); |
841 | } | |
842 | if (device->reada_curr_zone == zone) | |
62e85577 | 843 | pr_cont(" curr off %llu", |
7414a03f | 844 | device->reada_next - zone->start); |
62e85577 | 845 | pr_cont("\n"); |
09cbfeaf | 846 | index = (zone->end >> PAGE_SHIFT) + 1; |
7414a03f AJ |
847 | } |
848 | cnt = 0; | |
849 | index = 0; | |
850 | while (all) { | |
851 | struct reada_extent *re = NULL; | |
852 | ||
853 | ret = radix_tree_gang_lookup(&device->reada_extents, | |
854 | (void **)&re, index, 1); | |
855 | if (ret == 0) | |
856 | break; | |
62e85577 | 857 | pr_debug(" re: logical %llu size %u empty %d scheduled %d", |
da17066c | 858 | re->logical, fs_info->nodesize, |
895a11b8 | 859 | list_empty(&re->extctl), re->scheduled); |
7414a03f AJ |
860 | |
861 | for (i = 0; i < re->nzones; ++i) { | |
62e85577 | 862 | pr_cont(" zone %llu-%llu devs", |
7414a03f AJ |
863 | re->zones[i]->start, |
864 | re->zones[i]->end); | |
865 | for (j = 0; j < re->zones[i]->ndevs; ++j) { | |
62e85577 | 866 | pr_cont(" %lld", |
7414a03f AJ |
867 | re->zones[i]->devs[j]->devid); |
868 | } | |
869 | } | |
62e85577 | 870 | pr_cont("\n"); |
09cbfeaf | 871 | index = (re->logical >> PAGE_SHIFT) + 1; |
7414a03f AJ |
872 | if (++cnt > 15) |
873 | break; | |
874 | } | |
875 | } | |
876 | ||
877 | index = 0; | |
878 | cnt = 0; | |
879 | while (all) { | |
880 | struct reada_extent *re = NULL; | |
881 | ||
882 | ret = radix_tree_gang_lookup(&fs_info->reada_tree, (void **)&re, | |
883 | index, 1); | |
884 | if (ret == 0) | |
885 | break; | |
895a11b8 | 886 | if (!re->scheduled) { |
09cbfeaf | 887 | index = (re->logical >> PAGE_SHIFT) + 1; |
7414a03f AJ |
888 | continue; |
889 | } | |
62e85577 | 890 | pr_debug("re: logical %llu size %u list empty %d scheduled %d", |
da17066c | 891 | re->logical, fs_info->nodesize, |
895a11b8 | 892 | list_empty(&re->extctl), re->scheduled); |
7414a03f | 893 | for (i = 0; i < re->nzones; ++i) { |
62e85577 | 894 | pr_cont(" zone %llu-%llu devs", |
7414a03f AJ |
895 | re->zones[i]->start, |
896 | re->zones[i]->end); | |
8afd6841 | 897 | for (j = 0; j < re->zones[i]->ndevs; ++j) { |
62e85577 | 898 | pr_cont(" %lld", |
8afd6841 | 899 | re->zones[i]->devs[j]->devid); |
7414a03f AJ |
900 | } |
901 | } | |
62e85577 | 902 | pr_cont("\n"); |
09cbfeaf | 903 | index = (re->logical >> PAGE_SHIFT) + 1; |
7414a03f AJ |
904 | } |
905 | spin_unlock(&fs_info->reada_lock); | |
906 | } | |
907 | #endif | |
908 | ||
909 | /* | |
910 | * interface | |
911 | */ | |
912 | struct reada_control *btrfs_reada_add(struct btrfs_root *root, | |
913 | struct btrfs_key *key_start, struct btrfs_key *key_end) | |
914 | { | |
915 | struct reada_control *rc; | |
916 | u64 start; | |
917 | u64 generation; | |
ddd664f4 | 918 | int ret; |
7414a03f AJ |
919 | struct extent_buffer *node; |
920 | static struct btrfs_key max_key = { | |
921 | .objectid = (u64)-1, | |
922 | .type = (u8)-1, | |
923 | .offset = (u64)-1 | |
924 | }; | |
925 | ||
ed0244fa | 926 | rc = kzalloc(sizeof(*rc), GFP_KERNEL); |
7414a03f AJ |
927 | if (!rc) |
928 | return ERR_PTR(-ENOMEM); | |
929 | ||
c28f158e | 930 | rc->fs_info = root->fs_info; |
7414a03f AJ |
931 | rc->key_start = *key_start; |
932 | rc->key_end = *key_end; | |
933 | atomic_set(&rc->elems, 0); | |
934 | init_waitqueue_head(&rc->wait); | |
935 | kref_init(&rc->refcnt); | |
936 | kref_get(&rc->refcnt); /* one ref for having elements */ | |
937 | ||
938 | node = btrfs_root_node(root); | |
939 | start = node->start; | |
7414a03f AJ |
940 | generation = btrfs_header_generation(node); |
941 | free_extent_buffer(node); | |
942 | ||
1e7970c0 | 943 | ret = reada_add_block(rc, start, &max_key, generation); |
ddd664f4 | 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; | |
c28f158e | 958 | struct btrfs_fs_info *fs_info = rc->fs_info; |
7414a03f AJ |
959 | |
960 | while (atomic_read(&rc->elems)) { | |
4fe7a0e1 ZL |
961 | if (!atomic_read(&fs_info->reada_works_cnt)) |
962 | reada_start_machine(fs_info); | |
7414a03f AJ |
963 | wait_event_timeout(rc->wait, atomic_read(&rc->elems) == 0, |
964 | 5 * HZ); | |
0b246afa | 965 | dump_devs(fs_info, atomic_read(&rc->elems) < 10 ? 1 : 0); |
7414a03f AJ |
966 | } |
967 | ||
0b246afa | 968 | dump_devs(fs_info, atomic_read(&rc->elems) < 10 ? 1 : 0); |
7414a03f AJ |
969 | |
970 | kref_put(&rc->refcnt, reada_control_release); | |
971 | ||
972 | return 0; | |
973 | } | |
974 | #else | |
975 | int btrfs_reada_wait(void *handle) | |
976 | { | |
977 | struct reada_control *rc = handle; | |
c28f158e | 978 | struct btrfs_fs_info *fs_info = rc->fs_info; |
7414a03f AJ |
979 | |
980 | while (atomic_read(&rc->elems)) { | |
4fe7a0e1 ZL |
981 | if (!atomic_read(&fs_info->reada_works_cnt)) |
982 | reada_start_machine(fs_info); | |
983 | wait_event_timeout(rc->wait, atomic_read(&rc->elems) == 0, | |
984 | (HZ + 9) / 10); | |
7414a03f AJ |
985 | } |
986 | ||
987 | kref_put(&rc->refcnt, reada_control_release); | |
988 | ||
989 | return 0; | |
990 | } | |
991 | #endif | |
992 | ||
993 | void btrfs_reada_detach(void *handle) | |
994 | { | |
995 | struct reada_control *rc = handle; | |
996 | ||
997 | kref_put(&rc->refcnt, reada_control_release); | |
998 | } |