]>
Commit | Line | Data |
---|---|---|
b2441318 | 1 | // SPDX-License-Identifier: GPL-2.0 |
cafe5635 KO |
2 | /* |
3 | * Primary bucket allocation code | |
4 | * | |
5 | * Copyright 2012 Google, Inc. | |
6 | * | |
7 | * Allocation in bcache is done in terms of buckets: | |
8 | * | |
9 | * Each bucket has associated an 8 bit gen; this gen corresponds to the gen in | |
10 | * btree pointers - they must match for the pointer to be considered valid. | |
11 | * | |
12 | * Thus (assuming a bucket has no dirty data or metadata in it) we can reuse a | |
13 | * bucket simply by incrementing its gen. | |
14 | * | |
15 | * The gens (along with the priorities; it's really the gens are important but | |
16 | * the code is named as if it's the priorities) are written in an arbitrary list | |
17 | * of buckets on disk, with a pointer to them in the journal header. | |
18 | * | |
19 | * When we invalidate a bucket, we have to write its new gen to disk and wait | |
20 | * for that write to complete before we use it - otherwise after a crash we | |
21 | * could have pointers that appeared to be good but pointed to data that had | |
22 | * been overwritten. | |
23 | * | |
24 | * Since the gens and priorities are all stored contiguously on disk, we can | |
25 | * batch this up: We fill up the free_inc list with freshly invalidated buckets, | |
26 | * call prio_write(), and when prio_write() finishes we pull buckets off the | |
27 | * free_inc list and optionally discard them. | |
28 | * | |
29 | * free_inc isn't the only freelist - if it was, we'd often to sleep while | |
30 | * priorities and gens were being written before we could allocate. c->free is a | |
31 | * smaller freelist, and buckets on that list are always ready to be used. | |
32 | * | |
33 | * If we've got discards enabled, that happens when a bucket moves from the | |
34 | * free_inc list to the free list. | |
35 | * | |
36 | * There is another freelist, because sometimes we have buckets that we know | |
37 | * have nothing pointing into them - these we can reuse without waiting for | |
38 | * priorities to be rewritten. These come from freed btree nodes and buckets | |
39 | * that garbage collection discovered no longer had valid keys pointing into | |
40 | * them (because they were overwritten). That's the unused list - buckets on the | |
41 | * unused list move to the free list, optionally being discarded in the process. | |
42 | * | |
43 | * It's also important to ensure that gens don't wrap around - with respect to | |
44 | * either the oldest gen in the btree or the gen on disk. This is quite | |
45 | * difficult to do in practice, but we explicitly guard against it anyways - if | |
46 | * a bucket is in danger of wrapping around we simply skip invalidating it that | |
47 | * time around, and we garbage collect or rewrite the priorities sooner than we | |
48 | * would have otherwise. | |
49 | * | |
50 | * bch_bucket_alloc() allocates a single bucket from a specific cache. | |
51 | * | |
52 | * bch_bucket_alloc_set() allocates one or more buckets from different caches | |
53 | * out of a cache set. | |
54 | * | |
55 | * free_some_buckets() drives all the processes described above. It's called | |
56 | * from bch_bucket_alloc() and a few other places that need to make sure free | |
57 | * buckets are ready. | |
58 | * | |
59 | * invalidate_buckets_(lru|fifo)() find buckets that are available to be | |
60 | * invalidated, and then invalidate them and stick them on the free_inc list - | |
61 | * in either lru or fifo order. | |
62 | */ | |
63 | ||
64 | #include "bcache.h" | |
65 | #include "btree.h" | |
66 | ||
49b1212d | 67 | #include <linux/blkdev.h> |
119ba0f8 | 68 | #include <linux/kthread.h> |
cafe5635 | 69 | #include <linux/random.h> |
c37511b8 | 70 | #include <trace/events/bcache.h> |
cafe5635 | 71 | |
89b1fc54 TJ |
72 | #define MAX_OPEN_BUCKETS 128 |
73 | ||
cafe5635 KO |
74 | /* Bucket heap / gen */ |
75 | ||
76 | uint8_t bch_inc_gen(struct cache *ca, struct bucket *b) | |
77 | { | |
78 | uint8_t ret = ++b->gen; | |
79 | ||
80 | ca->set->need_gc = max(ca->set->need_gc, bucket_gc_gen(b)); | |
81 | WARN_ON_ONCE(ca->set->need_gc > BUCKET_GC_GEN_MAX); | |
82 | ||
cafe5635 KO |
83 | return ret; |
84 | } | |
85 | ||
86 | void bch_rescale_priorities(struct cache_set *c, int sectors) | |
87 | { | |
88 | struct cache *ca; | |
89 | struct bucket *b; | |
90 | unsigned next = c->nbuckets * c->sb.bucket_size / 1024; | |
91 | unsigned i; | |
92 | int r; | |
93 | ||
94 | atomic_sub(sectors, &c->rescale); | |
95 | ||
96 | do { | |
97 | r = atomic_read(&c->rescale); | |
98 | ||
99 | if (r >= 0) | |
100 | return; | |
101 | } while (atomic_cmpxchg(&c->rescale, r, r + next) != r); | |
102 | ||
103 | mutex_lock(&c->bucket_lock); | |
104 | ||
105 | c->min_prio = USHRT_MAX; | |
106 | ||
107 | for_each_cache(ca, c, i) | |
108 | for_each_bucket(b, ca) | |
109 | if (b->prio && | |
110 | b->prio != BTREE_PRIO && | |
111 | !atomic_read(&b->pin)) { | |
112 | b->prio--; | |
113 | c->min_prio = min(c->min_prio, b->prio); | |
114 | } | |
115 | ||
116 | mutex_unlock(&c->bucket_lock); | |
117 | } | |
118 | ||
2531d9ee KO |
119 | /* |
120 | * Background allocation thread: scans for buckets to be invalidated, | |
121 | * invalidates them, rewrites prios/gens (marking them as invalidated on disk), | |
122 | * then optionally issues discard commands to the newly free buckets, then puts | |
123 | * them on the various freelists. | |
124 | */ | |
cafe5635 KO |
125 | |
126 | static inline bool can_inc_bucket_gen(struct bucket *b) | |
127 | { | |
2531d9ee | 128 | return bucket_gc_gen(b) < BUCKET_GC_GEN_MAX; |
cafe5635 KO |
129 | } |
130 | ||
2531d9ee | 131 | bool bch_can_invalidate_bucket(struct cache *ca, struct bucket *b) |
cafe5635 | 132 | { |
2531d9ee | 133 | BUG_ON(!ca->set->gc_mark_valid); |
cafe5635 | 134 | |
4fe6a816 KO |
135 | return (!GC_MARK(b) || |
136 | GC_MARK(b) == GC_MARK_RECLAIMABLE) && | |
cafe5635 KO |
137 | !atomic_read(&b->pin) && |
138 | can_inc_bucket_gen(b); | |
139 | } | |
140 | ||
2531d9ee | 141 | void __bch_invalidate_one_bucket(struct cache *ca, struct bucket *b) |
cafe5635 | 142 | { |
2531d9ee KO |
143 | lockdep_assert_held(&ca->set->bucket_lock); |
144 | BUG_ON(GC_MARK(b) && GC_MARK(b) != GC_MARK_RECLAIMABLE); | |
7159b1ad KO |
145 | |
146 | if (GC_SECTORS_USED(b)) | |
2531d9ee | 147 | trace_bcache_invalidate(ca, b - ca->buckets); |
7159b1ad | 148 | |
cafe5635 KO |
149 | bch_inc_gen(ca, b); |
150 | b->prio = INITIAL_PRIO; | |
151 | atomic_inc(&b->pin); | |
2531d9ee KO |
152 | } |
153 | ||
154 | static void bch_invalidate_one_bucket(struct cache *ca, struct bucket *b) | |
155 | { | |
156 | __bch_invalidate_one_bucket(ca, b); | |
157 | ||
158 | fifo_push(&ca->free_inc, b - ca->buckets); | |
cafe5635 KO |
159 | } |
160 | ||
e0a985a4 KO |
161 | /* |
162 | * Determines what order we're going to reuse buckets, smallest bucket_prio() | |
163 | * first: we also take into account the number of sectors of live data in that | |
164 | * bucket, and in order for that multiply to make sense we have to scale bucket | |
165 | * | |
166 | * Thus, we scale the bucket priorities so that the bucket with the smallest | |
167 | * prio is worth 1/8th of what INITIAL_PRIO is worth. | |
168 | */ | |
169 | ||
170 | #define bucket_prio(b) \ | |
171 | ({ \ | |
172 | unsigned min_prio = (INITIAL_PRIO - ca->set->min_prio) / 8; \ | |
173 | \ | |
174 | (b->prio - ca->set->min_prio + min_prio) * GC_SECTORS_USED(b); \ | |
175 | }) | |
cafe5635 | 176 | |
b1a67b0f KO |
177 | #define bucket_max_cmp(l, r) (bucket_prio(l) < bucket_prio(r)) |
178 | #define bucket_min_cmp(l, r) (bucket_prio(l) > bucket_prio(r)) | |
cafe5635 | 179 | |
b1a67b0f KO |
180 | static void invalidate_buckets_lru(struct cache *ca) |
181 | { | |
cafe5635 KO |
182 | struct bucket *b; |
183 | ssize_t i; | |
184 | ||
185 | ca->heap.used = 0; | |
186 | ||
187 | for_each_bucket(b, ca) { | |
2531d9ee | 188 | if (!bch_can_invalidate_bucket(ca, b)) |
86b26b82 KO |
189 | continue; |
190 | ||
191 | if (!heap_full(&ca->heap)) | |
192 | heap_add(&ca->heap, b, bucket_max_cmp); | |
193 | else if (bucket_max_cmp(b, heap_peek(&ca->heap))) { | |
194 | ca->heap.data[0] = b; | |
195 | heap_sift(&ca->heap, 0, bucket_max_cmp); | |
cafe5635 KO |
196 | } |
197 | } | |
198 | ||
cafe5635 KO |
199 | for (i = ca->heap.used / 2 - 1; i >= 0; --i) |
200 | heap_sift(&ca->heap, i, bucket_min_cmp); | |
201 | ||
202 | while (!fifo_full(&ca->free_inc)) { | |
203 | if (!heap_pop(&ca->heap, b, bucket_min_cmp)) { | |
86b26b82 KO |
204 | /* |
205 | * We don't want to be calling invalidate_buckets() | |
cafe5635 KO |
206 | * multiple times when it can't do anything |
207 | */ | |
208 | ca->invalidate_needs_gc = 1; | |
72a44517 | 209 | wake_up_gc(ca->set); |
cafe5635 KO |
210 | return; |
211 | } | |
212 | ||
2531d9ee | 213 | bch_invalidate_one_bucket(ca, b); |
cafe5635 KO |
214 | } |
215 | } | |
216 | ||
217 | static void invalidate_buckets_fifo(struct cache *ca) | |
218 | { | |
219 | struct bucket *b; | |
220 | size_t checked = 0; | |
221 | ||
222 | while (!fifo_full(&ca->free_inc)) { | |
223 | if (ca->fifo_last_bucket < ca->sb.first_bucket || | |
224 | ca->fifo_last_bucket >= ca->sb.nbuckets) | |
225 | ca->fifo_last_bucket = ca->sb.first_bucket; | |
226 | ||
227 | b = ca->buckets + ca->fifo_last_bucket++; | |
228 | ||
2531d9ee KO |
229 | if (bch_can_invalidate_bucket(ca, b)) |
230 | bch_invalidate_one_bucket(ca, b); | |
cafe5635 KO |
231 | |
232 | if (++checked >= ca->sb.nbuckets) { | |
233 | ca->invalidate_needs_gc = 1; | |
72a44517 | 234 | wake_up_gc(ca->set); |
cafe5635 KO |
235 | return; |
236 | } | |
237 | } | |
238 | } | |
239 | ||
240 | static void invalidate_buckets_random(struct cache *ca) | |
241 | { | |
242 | struct bucket *b; | |
243 | size_t checked = 0; | |
244 | ||
245 | while (!fifo_full(&ca->free_inc)) { | |
246 | size_t n; | |
247 | get_random_bytes(&n, sizeof(n)); | |
248 | ||
249 | n %= (size_t) (ca->sb.nbuckets - ca->sb.first_bucket); | |
250 | n += ca->sb.first_bucket; | |
251 | ||
252 | b = ca->buckets + n; | |
253 | ||
2531d9ee KO |
254 | if (bch_can_invalidate_bucket(ca, b)) |
255 | bch_invalidate_one_bucket(ca, b); | |
cafe5635 KO |
256 | |
257 | if (++checked >= ca->sb.nbuckets / 2) { | |
258 | ca->invalidate_needs_gc = 1; | |
72a44517 | 259 | wake_up_gc(ca->set); |
cafe5635 KO |
260 | return; |
261 | } | |
262 | } | |
263 | } | |
264 | ||
265 | static void invalidate_buckets(struct cache *ca) | |
266 | { | |
2531d9ee | 267 | BUG_ON(ca->invalidate_needs_gc); |
cafe5635 KO |
268 | |
269 | switch (CACHE_REPLACEMENT(&ca->sb)) { | |
270 | case CACHE_REPLACEMENT_LRU: | |
271 | invalidate_buckets_lru(ca); | |
272 | break; | |
273 | case CACHE_REPLACEMENT_FIFO: | |
274 | invalidate_buckets_fifo(ca); | |
275 | break; | |
276 | case CACHE_REPLACEMENT_RANDOM: | |
277 | invalidate_buckets_random(ca); | |
278 | break; | |
279 | } | |
280 | } | |
281 | ||
282 | #define allocator_wait(ca, cond) \ | |
283 | do { \ | |
86b26b82 | 284 | while (1) { \ |
119ba0f8 | 285 | set_current_state(TASK_INTERRUPTIBLE); \ |
86b26b82 KO |
286 | if (cond) \ |
287 | break; \ | |
cafe5635 KO |
288 | \ |
289 | mutex_unlock(&(ca)->set->bucket_lock); \ | |
79826c35 | 290 | if (kthread_should_stop()) \ |
119ba0f8 | 291 | return 0; \ |
cafe5635 KO |
292 | \ |
293 | schedule(); \ | |
cafe5635 KO |
294 | mutex_lock(&(ca)->set->bucket_lock); \ |
295 | } \ | |
119ba0f8 | 296 | __set_current_state(TASK_RUNNING); \ |
cafe5635 KO |
297 | } while (0) |
298 | ||
78365411 KO |
299 | static int bch_allocator_push(struct cache *ca, long bucket) |
300 | { | |
301 | unsigned i; | |
302 | ||
303 | /* Prios/gens are actually the most important reserve */ | |
304 | if (fifo_push(&ca->free[RESERVE_PRIO], bucket)) | |
305 | return true; | |
306 | ||
307 | for (i = 0; i < RESERVE_NR; i++) | |
308 | if (fifo_push(&ca->free[i], bucket)) | |
309 | return true; | |
310 | ||
311 | return false; | |
312 | } | |
313 | ||
119ba0f8 | 314 | static int bch_allocator_thread(void *arg) |
cafe5635 | 315 | { |
119ba0f8 | 316 | struct cache *ca = arg; |
cafe5635 KO |
317 | |
318 | mutex_lock(&ca->set->bucket_lock); | |
319 | ||
320 | while (1) { | |
86b26b82 KO |
321 | /* |
322 | * First, we pull buckets off of the unused and free_inc lists, | |
323 | * possibly issue discards to them, then we add the bucket to | |
324 | * the free list: | |
325 | */ | |
2531d9ee | 326 | while (!fifo_empty(&ca->free_inc)) { |
cafe5635 KO |
327 | long bucket; |
328 | ||
2531d9ee | 329 | fifo_pop(&ca->free_inc, bucket); |
cafe5635 | 330 | |
cafe5635 | 331 | if (ca->discard) { |
49b1212d KO |
332 | mutex_unlock(&ca->set->bucket_lock); |
333 | blkdev_issue_discard(ca->bdev, | |
334 | bucket_to_sector(ca->set, bucket), | |
8b326d3a | 335 | ca->sb.bucket_size, GFP_KERNEL, 0); |
49b1212d | 336 | mutex_lock(&ca->set->bucket_lock); |
cafe5635 | 337 | } |
49b1212d | 338 | |
78365411 | 339 | allocator_wait(ca, bch_allocator_push(ca, bucket)); |
0a63b66d | 340 | wake_up(&ca->set->btree_cache_wait); |
35fcd848 | 341 | wake_up(&ca->set->bucket_wait); |
cafe5635 KO |
342 | } |
343 | ||
86b26b82 KO |
344 | /* |
345 | * We've run out of free buckets, we need to find some buckets | |
346 | * we can invalidate. First, invalidate them in memory and add | |
347 | * them to the free_inc list: | |
348 | */ | |
cafe5635 | 349 | |
2531d9ee | 350 | retry_invalidate: |
86b26b82 | 351 | allocator_wait(ca, ca->set->gc_mark_valid && |
2531d9ee | 352 | !ca->invalidate_needs_gc); |
86b26b82 | 353 | invalidate_buckets(ca); |
cafe5635 | 354 | |
86b26b82 KO |
355 | /* |
356 | * Now, we write their new gens to disk so we can start writing | |
357 | * new stuff to them: | |
358 | */ | |
359 | allocator_wait(ca, !atomic_read(&ca->set->prio_blocked)); | |
2531d9ee KO |
360 | if (CACHE_SYNC(&ca->set->sb)) { |
361 | /* | |
362 | * This could deadlock if an allocation with a btree | |
363 | * node locked ever blocked - having the btree node | |
364 | * locked would block garbage collection, but here we're | |
365 | * waiting on garbage collection before we invalidate | |
366 | * and free anything. | |
367 | * | |
368 | * But this should be safe since the btree code always | |
369 | * uses btree_check_reserve() before allocating now, and | |
370 | * if it fails it blocks without btree nodes locked. | |
371 | */ | |
372 | if (!fifo_full(&ca->free_inc)) | |
373 | goto retry_invalidate; | |
374 | ||
cafe5635 | 375 | bch_prio_write(ca); |
2531d9ee | 376 | } |
cafe5635 KO |
377 | } |
378 | } | |
379 | ||
2531d9ee KO |
380 | /* Allocation */ |
381 | ||
78365411 | 382 | long bch_bucket_alloc(struct cache *ca, unsigned reserve, bool wait) |
cafe5635 | 383 | { |
35fcd848 KO |
384 | DEFINE_WAIT(w); |
385 | struct bucket *b; | |
386 | long r; | |
387 | ||
388 | /* fastpath */ | |
78365411 KO |
389 | if (fifo_pop(&ca->free[RESERVE_NONE], r) || |
390 | fifo_pop(&ca->free[reserve], r)) | |
35fcd848 | 391 | goto out; |
35fcd848 | 392 | |
7159b1ad KO |
393 | if (!wait) { |
394 | trace_bcache_alloc_fail(ca, reserve); | |
35fcd848 | 395 | return -1; |
7159b1ad | 396 | } |
35fcd848 | 397 | |
78365411 | 398 | do { |
35fcd848 KO |
399 | prepare_to_wait(&ca->set->bucket_wait, &w, |
400 | TASK_UNINTERRUPTIBLE); | |
401 | ||
402 | mutex_unlock(&ca->set->bucket_lock); | |
403 | schedule(); | |
404 | mutex_lock(&ca->set->bucket_lock); | |
78365411 KO |
405 | } while (!fifo_pop(&ca->free[RESERVE_NONE], r) && |
406 | !fifo_pop(&ca->free[reserve], r)); | |
35fcd848 KO |
407 | |
408 | finish_wait(&ca->set->bucket_wait, &w); | |
409 | out: | |
91af8300 CL |
410 | if (ca->alloc_thread) |
411 | wake_up_process(ca->alloc_thread); | |
cafe5635 | 412 | |
7159b1ad KO |
413 | trace_bcache_alloc(ca, reserve); |
414 | ||
280481d0 | 415 | if (expensive_debug_checks(ca->set)) { |
cafe5635 KO |
416 | size_t iter; |
417 | long i; | |
78365411 | 418 | unsigned j; |
cafe5635 KO |
419 | |
420 | for (iter = 0; iter < prio_buckets(ca) * 2; iter++) | |
421 | BUG_ON(ca->prio_buckets[iter] == (uint64_t) r); | |
422 | ||
78365411 KO |
423 | for (j = 0; j < RESERVE_NR; j++) |
424 | fifo_for_each(i, &ca->free[j], iter) | |
425 | BUG_ON(i == r); | |
cafe5635 KO |
426 | fifo_for_each(i, &ca->free_inc, iter) |
427 | BUG_ON(i == r); | |
cafe5635 | 428 | } |
280481d0 | 429 | |
35fcd848 | 430 | b = ca->buckets + r; |
cafe5635 | 431 | |
35fcd848 | 432 | BUG_ON(atomic_read(&b->pin) != 1); |
cafe5635 | 433 | |
35fcd848 | 434 | SET_GC_SECTORS_USED(b, ca->sb.bucket_size); |
cafe5635 | 435 | |
78365411 | 436 | if (reserve <= RESERVE_PRIO) { |
35fcd848 | 437 | SET_GC_MARK(b, GC_MARK_METADATA); |
981aa8c0 | 438 | SET_GC_MOVE(b, 0); |
35fcd848 KO |
439 | b->prio = BTREE_PRIO; |
440 | } else { | |
441 | SET_GC_MARK(b, GC_MARK_RECLAIMABLE); | |
981aa8c0 | 442 | SET_GC_MOVE(b, 0); |
35fcd848 | 443 | b->prio = INITIAL_PRIO; |
cafe5635 KO |
444 | } |
445 | ||
d44c2f9e TJ |
446 | if (ca->set->avail_nbuckets > 0) { |
447 | ca->set->avail_nbuckets--; | |
448 | bch_update_bucket_in_use(ca->set, &ca->set->gc_stats); | |
449 | } | |
450 | ||
35fcd848 | 451 | return r; |
cafe5635 KO |
452 | } |
453 | ||
2531d9ee KO |
454 | void __bch_bucket_free(struct cache *ca, struct bucket *b) |
455 | { | |
456 | SET_GC_MARK(b, 0); | |
457 | SET_GC_SECTORS_USED(b, 0); | |
d44c2f9e TJ |
458 | |
459 | if (ca->set->avail_nbuckets < ca->set->nbuckets) { | |
460 | ca->set->avail_nbuckets++; | |
461 | bch_update_bucket_in_use(ca->set, &ca->set->gc_stats); | |
462 | } | |
2531d9ee KO |
463 | } |
464 | ||
cafe5635 KO |
465 | void bch_bucket_free(struct cache_set *c, struct bkey *k) |
466 | { | |
467 | unsigned i; | |
468 | ||
2531d9ee KO |
469 | for (i = 0; i < KEY_PTRS(k); i++) |
470 | __bch_bucket_free(PTR_CACHE(c, k, i), | |
471 | PTR_BUCKET(c, k, i)); | |
cafe5635 KO |
472 | } |
473 | ||
78365411 | 474 | int __bch_bucket_alloc_set(struct cache_set *c, unsigned reserve, |
35fcd848 | 475 | struct bkey *k, int n, bool wait) |
cafe5635 KO |
476 | { |
477 | int i; | |
478 | ||
479 | lockdep_assert_held(&c->bucket_lock); | |
480 | BUG_ON(!n || n > c->caches_loaded || n > 8); | |
481 | ||
482 | bkey_init(k); | |
483 | ||
484 | /* sort by free space/prio of oldest data in caches */ | |
485 | ||
486 | for (i = 0; i < n; i++) { | |
487 | struct cache *ca = c->cache_by_alloc[i]; | |
78365411 | 488 | long b = bch_bucket_alloc(ca, reserve, wait); |
cafe5635 KO |
489 | |
490 | if (b == -1) | |
491 | goto err; | |
492 | ||
cf33c1ee | 493 | k->ptr[i] = MAKE_PTR(ca->buckets[b].gen, |
cafe5635 KO |
494 | bucket_to_sector(c, b), |
495 | ca->sb.nr_this_dev); | |
496 | ||
497 | SET_KEY_PTRS(k, i + 1); | |
498 | } | |
499 | ||
500 | return 0; | |
501 | err: | |
502 | bch_bucket_free(c, k); | |
3a3b6a4e | 503 | bkey_put(c, k); |
cafe5635 KO |
504 | return -1; |
505 | } | |
506 | ||
78365411 | 507 | int bch_bucket_alloc_set(struct cache_set *c, unsigned reserve, |
35fcd848 | 508 | struct bkey *k, int n, bool wait) |
cafe5635 KO |
509 | { |
510 | int ret; | |
511 | mutex_lock(&c->bucket_lock); | |
78365411 | 512 | ret = __bch_bucket_alloc_set(c, reserve, k, n, wait); |
cafe5635 KO |
513 | mutex_unlock(&c->bucket_lock); |
514 | return ret; | |
515 | } | |
516 | ||
2599b53b KO |
517 | /* Sector allocator */ |
518 | ||
519 | struct open_bucket { | |
520 | struct list_head list; | |
521 | unsigned last_write_point; | |
522 | unsigned sectors_free; | |
523 | BKEY_PADDED(key); | |
524 | }; | |
525 | ||
526 | /* | |
527 | * We keep multiple buckets open for writes, and try to segregate different | |
528 | * write streams for better cache utilization: first we look for a bucket where | |
529 | * the last write to it was sequential with the current write, and failing that | |
530 | * we look for a bucket that was last used by the same task. | |
531 | * | |
532 | * The ideas is if you've got multiple tasks pulling data into the cache at the | |
533 | * same time, you'll get better cache utilization if you try to segregate their | |
534 | * data and preserve locality. | |
535 | * | |
536 | * For example, say you've starting Firefox at the same time you're copying a | |
537 | * bunch of files. Firefox will likely end up being fairly hot and stay in the | |
538 | * cache awhile, but the data you copied might not be; if you wrote all that | |
539 | * data to the same buckets it'd get invalidated at the same time. | |
540 | * | |
541 | * Both of those tasks will be doing fairly random IO so we can't rely on | |
542 | * detecting sequential IO to segregate their data, but going off of the task | |
543 | * should be a sane heuristic. | |
544 | */ | |
545 | static struct open_bucket *pick_data_bucket(struct cache_set *c, | |
546 | const struct bkey *search, | |
547 | unsigned write_point, | |
548 | struct bkey *alloc) | |
549 | { | |
550 | struct open_bucket *ret, *ret_task = NULL; | |
551 | ||
552 | list_for_each_entry_reverse(ret, &c->data_buckets, list) | |
553 | if (!bkey_cmp(&ret->key, search)) | |
554 | goto found; | |
555 | else if (ret->last_write_point == write_point) | |
556 | ret_task = ret; | |
557 | ||
558 | ret = ret_task ?: list_first_entry(&c->data_buckets, | |
559 | struct open_bucket, list); | |
560 | found: | |
561 | if (!ret->sectors_free && KEY_PTRS(alloc)) { | |
562 | ret->sectors_free = c->sb.bucket_size; | |
563 | bkey_copy(&ret->key, alloc); | |
564 | bkey_init(alloc); | |
565 | } | |
566 | ||
567 | if (!ret->sectors_free) | |
568 | ret = NULL; | |
569 | ||
570 | return ret; | |
571 | } | |
572 | ||
573 | /* | |
574 | * Allocates some space in the cache to write to, and k to point to the newly | |
575 | * allocated space, and updates KEY_SIZE(k) and KEY_OFFSET(k) (to point to the | |
576 | * end of the newly allocated space). | |
577 | * | |
578 | * May allocate fewer sectors than @sectors, KEY_SIZE(k) indicates how many | |
579 | * sectors were actually allocated. | |
580 | * | |
581 | * If s->writeback is true, will not fail. | |
582 | */ | |
583 | bool bch_alloc_sectors(struct cache_set *c, struct bkey *k, unsigned sectors, | |
584 | unsigned write_point, unsigned write_prio, bool wait) | |
585 | { | |
586 | struct open_bucket *b; | |
587 | BKEY_PADDED(key) alloc; | |
588 | unsigned i; | |
589 | ||
590 | /* | |
591 | * We might have to allocate a new bucket, which we can't do with a | |
592 | * spinlock held. So if we have to allocate, we drop the lock, allocate | |
593 | * and then retry. KEY_PTRS() indicates whether alloc points to | |
594 | * allocated bucket(s). | |
595 | */ | |
596 | ||
597 | bkey_init(&alloc.key); | |
598 | spin_lock(&c->data_bucket_lock); | |
599 | ||
600 | while (!(b = pick_data_bucket(c, k, write_point, &alloc.key))) { | |
601 | unsigned watermark = write_prio | |
78365411 KO |
602 | ? RESERVE_MOVINGGC |
603 | : RESERVE_NONE; | |
2599b53b KO |
604 | |
605 | spin_unlock(&c->data_bucket_lock); | |
606 | ||
607 | if (bch_bucket_alloc_set(c, watermark, &alloc.key, 1, wait)) | |
608 | return false; | |
609 | ||
610 | spin_lock(&c->data_bucket_lock); | |
611 | } | |
612 | ||
613 | /* | |
614 | * If we had to allocate, we might race and not need to allocate the | |
b1e8139e | 615 | * second time we call pick_data_bucket(). If we allocated a bucket but |
2599b53b KO |
616 | * didn't use it, drop the refcount bch_bucket_alloc_set() took: |
617 | */ | |
618 | if (KEY_PTRS(&alloc.key)) | |
3a3b6a4e | 619 | bkey_put(c, &alloc.key); |
2599b53b KO |
620 | |
621 | for (i = 0; i < KEY_PTRS(&b->key); i++) | |
622 | EBUG_ON(ptr_stale(c, &b->key, i)); | |
623 | ||
624 | /* Set up the pointer to the space we're allocating: */ | |
625 | ||
626 | for (i = 0; i < KEY_PTRS(&b->key); i++) | |
627 | k->ptr[i] = b->key.ptr[i]; | |
628 | ||
629 | sectors = min(sectors, b->sectors_free); | |
630 | ||
631 | SET_KEY_OFFSET(k, KEY_OFFSET(k) + sectors); | |
632 | SET_KEY_SIZE(k, sectors); | |
633 | SET_KEY_PTRS(k, KEY_PTRS(&b->key)); | |
634 | ||
635 | /* | |
636 | * Move b to the end of the lru, and keep track of what this bucket was | |
637 | * last used for: | |
638 | */ | |
639 | list_move_tail(&b->list, &c->data_buckets); | |
640 | bkey_copy_key(&b->key, k); | |
641 | b->last_write_point = write_point; | |
642 | ||
643 | b->sectors_free -= sectors; | |
644 | ||
645 | for (i = 0; i < KEY_PTRS(&b->key); i++) { | |
646 | SET_PTR_OFFSET(&b->key, i, PTR_OFFSET(&b->key, i) + sectors); | |
647 | ||
648 | atomic_long_add(sectors, | |
649 | &PTR_CACHE(c, &b->key, i)->sectors_written); | |
650 | } | |
651 | ||
652 | if (b->sectors_free < c->sb.block_size) | |
653 | b->sectors_free = 0; | |
654 | ||
655 | /* | |
656 | * k takes refcounts on the buckets it points to until it's inserted | |
657 | * into the btree, but if we're done with this bucket we just transfer | |
658 | * get_data_bucket()'s refcount. | |
659 | */ | |
660 | if (b->sectors_free) | |
661 | for (i = 0; i < KEY_PTRS(&b->key); i++) | |
662 | atomic_inc(&PTR_BUCKET(c, &b->key, i)->pin); | |
663 | ||
664 | spin_unlock(&c->data_bucket_lock); | |
665 | return true; | |
666 | } | |
667 | ||
cafe5635 KO |
668 | /* Init */ |
669 | ||
2599b53b KO |
670 | void bch_open_buckets_free(struct cache_set *c) |
671 | { | |
672 | struct open_bucket *b; | |
673 | ||
674 | while (!list_empty(&c->data_buckets)) { | |
675 | b = list_first_entry(&c->data_buckets, | |
676 | struct open_bucket, list); | |
677 | list_del(&b->list); | |
678 | kfree(b); | |
679 | } | |
680 | } | |
681 | ||
682 | int bch_open_buckets_alloc(struct cache_set *c) | |
683 | { | |
684 | int i; | |
685 | ||
686 | spin_lock_init(&c->data_bucket_lock); | |
687 | ||
89b1fc54 | 688 | for (i = 0; i < MAX_OPEN_BUCKETS; i++) { |
2599b53b KO |
689 | struct open_bucket *b = kzalloc(sizeof(*b), GFP_KERNEL); |
690 | if (!b) | |
691 | return -ENOMEM; | |
692 | ||
693 | list_add(&b->list, &c->data_buckets); | |
694 | } | |
695 | ||
696 | return 0; | |
697 | } | |
698 | ||
119ba0f8 KO |
699 | int bch_cache_allocator_start(struct cache *ca) |
700 | { | |
79826c35 KO |
701 | struct task_struct *k = kthread_run(bch_allocator_thread, |
702 | ca, "bcache_allocator"); | |
703 | if (IS_ERR(k)) | |
704 | return PTR_ERR(k); | |
119ba0f8 | 705 | |
79826c35 | 706 | ca->alloc_thread = k; |
119ba0f8 KO |
707 | return 0; |
708 | } |