Commit | Line | Data |
---|---|---|
cafe5635 KO |
1 | /* |
2 | * bcache journalling code, for btree insertions | |
3 | * | |
4 | * Copyright 2012 Google, Inc. | |
5 | */ | |
6 | ||
7 | #include "bcache.h" | |
8 | #include "btree.h" | |
9 | #include "debug.h" | |
10 | #include "request.h" | |
11 | ||
c37511b8 KO |
12 | #include <trace/events/bcache.h> |
13 | ||
cafe5635 KO |
14 | /* |
15 | * Journal replay/recovery: | |
16 | * | |
17 | * This code is all driven from run_cache_set(); we first read the journal | |
18 | * entries, do some other stuff, then we mark all the keys in the journal | |
19 | * entries (same as garbage collection would), then we replay them - reinserting | |
20 | * them into the cache in precisely the same order as they appear in the | |
21 | * journal. | |
22 | * | |
23 | * We only journal keys that go in leaf nodes, which simplifies things quite a | |
24 | * bit. | |
25 | */ | |
26 | ||
27 | static void journal_read_endio(struct bio *bio, int error) | |
28 | { | |
29 | struct closure *cl = bio->bi_private; | |
30 | closure_put(cl); | |
31 | } | |
32 | ||
33 | static int journal_read_bucket(struct cache *ca, struct list_head *list, | |
34 | struct btree_op *op, unsigned bucket_index) | |
35 | { | |
36 | struct journal_device *ja = &ca->journal; | |
37 | struct bio *bio = &ja->bio; | |
38 | ||
39 | struct journal_replay *i; | |
40 | struct jset *j, *data = ca->set->journal.w[0].data; | |
41 | unsigned len, left, offset = 0; | |
42 | int ret = 0; | |
43 | sector_t bucket = bucket_to_sector(ca->set, ca->sb.d[bucket_index]); | |
44 | ||
45 | pr_debug("reading %llu", (uint64_t) bucket); | |
46 | ||
47 | while (offset < ca->sb.bucket_size) { | |
48 | reread: left = ca->sb.bucket_size - offset; | |
49 | len = min_t(unsigned, left, PAGE_SECTORS * 8); | |
50 | ||
51 | bio_reset(bio); | |
52 | bio->bi_sector = bucket + offset; | |
53 | bio->bi_bdev = ca->bdev; | |
54 | bio->bi_rw = READ; | |
55 | bio->bi_size = len << 9; | |
56 | ||
57 | bio->bi_end_io = journal_read_endio; | |
58 | bio->bi_private = &op->cl; | |
169ef1cf | 59 | bch_bio_map(bio, data); |
cafe5635 KO |
60 | |
61 | closure_bio_submit(bio, &op->cl, ca); | |
62 | closure_sync(&op->cl); | |
63 | ||
64 | /* This function could be simpler now since we no longer write | |
65 | * journal entries that overlap bucket boundaries; this means | |
66 | * the start of a bucket will always have a valid journal entry | |
67 | * if it has any journal entries at all. | |
68 | */ | |
69 | ||
70 | j = data; | |
71 | while (len) { | |
72 | struct list_head *where; | |
73 | size_t blocks, bytes = set_bytes(j); | |
74 | ||
75 | if (j->magic != jset_magic(ca->set)) | |
76 | return ret; | |
77 | ||
78 | if (bytes > left << 9) | |
79 | return ret; | |
80 | ||
81 | if (bytes > len << 9) | |
82 | goto reread; | |
83 | ||
84 | if (j->csum != csum_set(j)) | |
85 | return ret; | |
86 | ||
87 | blocks = set_blocks(j, ca->set); | |
88 | ||
89 | while (!list_empty(list)) { | |
90 | i = list_first_entry(list, | |
91 | struct journal_replay, list); | |
92 | if (i->j.seq >= j->last_seq) | |
93 | break; | |
94 | list_del(&i->list); | |
95 | kfree(i); | |
96 | } | |
97 | ||
98 | list_for_each_entry_reverse(i, list, list) { | |
99 | if (j->seq == i->j.seq) | |
100 | goto next_set; | |
101 | ||
102 | if (j->seq < i->j.last_seq) | |
103 | goto next_set; | |
104 | ||
105 | if (j->seq > i->j.seq) { | |
106 | where = &i->list; | |
107 | goto add; | |
108 | } | |
109 | } | |
110 | ||
111 | where = list; | |
112 | add: | |
113 | i = kmalloc(offsetof(struct journal_replay, j) + | |
114 | bytes, GFP_KERNEL); | |
115 | if (!i) | |
116 | return -ENOMEM; | |
117 | memcpy(&i->j, j, bytes); | |
118 | list_add(&i->list, where); | |
119 | ret = 1; | |
120 | ||
121 | ja->seq[bucket_index] = j->seq; | |
122 | next_set: | |
123 | offset += blocks * ca->sb.block_size; | |
124 | len -= blocks * ca->sb.block_size; | |
125 | j = ((void *) j) + blocks * block_bytes(ca); | |
126 | } | |
127 | } | |
128 | ||
129 | return ret; | |
130 | } | |
131 | ||
132 | int bch_journal_read(struct cache_set *c, struct list_head *list, | |
133 | struct btree_op *op) | |
134 | { | |
135 | #define read_bucket(b) \ | |
136 | ({ \ | |
137 | int ret = journal_read_bucket(ca, list, op, b); \ | |
138 | __set_bit(b, bitmap); \ | |
139 | if (ret < 0) \ | |
140 | return ret; \ | |
141 | ret; \ | |
142 | }) | |
143 | ||
144 | struct cache *ca; | |
145 | unsigned iter; | |
146 | ||
147 | for_each_cache(ca, c, iter) { | |
148 | struct journal_device *ja = &ca->journal; | |
149 | unsigned long bitmap[SB_JOURNAL_BUCKETS / BITS_PER_LONG]; | |
150 | unsigned i, l, r, m; | |
151 | uint64_t seq; | |
152 | ||
153 | bitmap_zero(bitmap, SB_JOURNAL_BUCKETS); | |
154 | pr_debug("%u journal buckets", ca->sb.njournal_buckets); | |
155 | ||
156 | /* Read journal buckets ordered by golden ratio hash to quickly | |
157 | * find a sequence of buckets with valid journal entries | |
158 | */ | |
159 | for (i = 0; i < ca->sb.njournal_buckets; i++) { | |
160 | l = (i * 2654435769U) % ca->sb.njournal_buckets; | |
161 | ||
162 | if (test_bit(l, bitmap)) | |
163 | break; | |
164 | ||
165 | if (read_bucket(l)) | |
166 | goto bsearch; | |
167 | } | |
168 | ||
169 | /* If that fails, check all the buckets we haven't checked | |
170 | * already | |
171 | */ | |
172 | pr_debug("falling back to linear search"); | |
173 | ||
174 | for (l = 0; l < ca->sb.njournal_buckets; l++) { | |
175 | if (test_bit(l, bitmap)) | |
176 | continue; | |
177 | ||
178 | if (read_bucket(l)) | |
179 | goto bsearch; | |
180 | } | |
181 | bsearch: | |
182 | /* Binary search */ | |
183 | m = r = find_next_bit(bitmap, ca->sb.njournal_buckets, l + 1); | |
184 | pr_debug("starting binary search, l %u r %u", l, r); | |
185 | ||
186 | while (l + 1 < r) { | |
faa56736 KO |
187 | seq = list_entry(list->prev, struct journal_replay, |
188 | list)->j.seq; | |
189 | ||
cafe5635 | 190 | m = (l + r) >> 1; |
faa56736 | 191 | read_bucket(m); |
cafe5635 | 192 | |
faa56736 KO |
193 | if (seq != list_entry(list->prev, struct journal_replay, |
194 | list)->j.seq) | |
cafe5635 KO |
195 | l = m; |
196 | else | |
197 | r = m; | |
198 | } | |
199 | ||
200 | /* Read buckets in reverse order until we stop finding more | |
201 | * journal entries | |
202 | */ | |
203 | pr_debug("finishing up"); | |
204 | l = m; | |
205 | ||
206 | while (1) { | |
207 | if (!l--) | |
208 | l = ca->sb.njournal_buckets - 1; | |
209 | ||
210 | if (l == m) | |
211 | break; | |
212 | ||
213 | if (test_bit(l, bitmap)) | |
214 | continue; | |
215 | ||
216 | if (!read_bucket(l)) | |
217 | break; | |
218 | } | |
219 | ||
220 | seq = 0; | |
221 | ||
222 | for (i = 0; i < ca->sb.njournal_buckets; i++) | |
223 | if (ja->seq[i] > seq) { | |
224 | seq = ja->seq[i]; | |
225 | ja->cur_idx = ja->discard_idx = | |
226 | ja->last_idx = i; | |
227 | ||
228 | } | |
229 | } | |
230 | ||
231 | c->journal.seq = list_entry(list->prev, | |
232 | struct journal_replay, | |
233 | list)->j.seq; | |
234 | ||
235 | return 0; | |
236 | #undef read_bucket | |
237 | } | |
238 | ||
239 | void bch_journal_mark(struct cache_set *c, struct list_head *list) | |
240 | { | |
241 | atomic_t p = { 0 }; | |
242 | struct bkey *k; | |
243 | struct journal_replay *i; | |
244 | struct journal *j = &c->journal; | |
245 | uint64_t last = j->seq; | |
246 | ||
247 | /* | |
248 | * journal.pin should never fill up - we never write a journal | |
249 | * entry when it would fill up. But if for some reason it does, we | |
250 | * iterate over the list in reverse order so that we can just skip that | |
251 | * refcount instead of bugging. | |
252 | */ | |
253 | ||
254 | list_for_each_entry_reverse(i, list, list) { | |
255 | BUG_ON(last < i->j.seq); | |
256 | i->pin = NULL; | |
257 | ||
258 | while (last-- != i->j.seq) | |
259 | if (fifo_free(&j->pin) > 1) { | |
260 | fifo_push_front(&j->pin, p); | |
261 | atomic_set(&fifo_front(&j->pin), 0); | |
262 | } | |
263 | ||
264 | if (fifo_free(&j->pin) > 1) { | |
265 | fifo_push_front(&j->pin, p); | |
266 | i->pin = &fifo_front(&j->pin); | |
267 | atomic_set(i->pin, 1); | |
268 | } | |
269 | ||
270 | for (k = i->j.start; | |
271 | k < end(&i->j); | |
272 | k = bkey_next(k)) { | |
273 | unsigned j; | |
274 | ||
275 | for (j = 0; j < KEY_PTRS(k); j++) { | |
276 | struct bucket *g = PTR_BUCKET(c, k, j); | |
277 | atomic_inc(&g->pin); | |
278 | ||
279 | if (g->prio == BTREE_PRIO && | |
280 | !ptr_stale(c, k, j)) | |
281 | g->prio = INITIAL_PRIO; | |
282 | } | |
283 | ||
284 | __bch_btree_mark_key(c, 0, k); | |
285 | } | |
286 | } | |
287 | } | |
288 | ||
289 | int bch_journal_replay(struct cache_set *s, struct list_head *list, | |
290 | struct btree_op *op) | |
291 | { | |
292 | int ret = 0, keys = 0, entries = 0; | |
293 | struct bkey *k; | |
294 | struct journal_replay *i = | |
295 | list_entry(list->prev, struct journal_replay, list); | |
296 | ||
297 | uint64_t start = i->j.last_seq, end = i->j.seq, n = start; | |
298 | ||
299 | list_for_each_entry(i, list, list) { | |
300 | BUG_ON(i->pin && atomic_read(i->pin) != 1); | |
301 | ||
302 | if (n != i->j.seq) | |
b1a67b0f KO |
303 | pr_err( |
304 | "journal entries %llu-%llu missing! (replaying %llu-%llu)\n", | |
305 | n, i->j.seq - 1, start, end); | |
cafe5635 KO |
306 | |
307 | for (k = i->j.start; | |
308 | k < end(&i->j); | |
309 | k = bkey_next(k)) { | |
c37511b8 KO |
310 | trace_bcache_journal_replay_key(k); |
311 | ||
cafe5635 KO |
312 | bkey_copy(op->keys.top, k); |
313 | bch_keylist_push(&op->keys); | |
314 | ||
315 | op->journal = i->pin; | |
316 | atomic_inc(op->journal); | |
317 | ||
318 | ret = bch_btree_insert(op, s); | |
319 | if (ret) | |
320 | goto err; | |
321 | ||
322 | BUG_ON(!bch_keylist_empty(&op->keys)); | |
323 | keys++; | |
324 | ||
325 | cond_resched(); | |
326 | } | |
327 | ||
328 | if (i->pin) | |
329 | atomic_dec(i->pin); | |
330 | n = i->j.seq + 1; | |
331 | entries++; | |
332 | } | |
333 | ||
334 | pr_info("journal replay done, %i keys in %i entries, seq %llu", | |
335 | keys, entries, end); | |
336 | ||
337 | while (!list_empty(list)) { | |
338 | i = list_first_entry(list, struct journal_replay, list); | |
339 | list_del(&i->list); | |
340 | kfree(i); | |
341 | } | |
342 | err: | |
343 | closure_sync(&op->cl); | |
344 | return ret; | |
345 | } | |
346 | ||
347 | /* Journalling */ | |
348 | ||
349 | static void btree_flush_write(struct cache_set *c) | |
350 | { | |
351 | /* | |
352 | * Try to find the btree node with that references the oldest journal | |
353 | * entry, best is our current candidate and is locked if non NULL: | |
354 | */ | |
355 | struct btree *b, *best = NULL; | |
356 | unsigned iter; | |
357 | ||
358 | for_each_cached_btree(b, c, iter) { | |
359 | if (!down_write_trylock(&b->lock)) | |
360 | continue; | |
361 | ||
362 | if (!btree_node_dirty(b) || | |
363 | !btree_current_write(b)->journal) { | |
364 | rw_unlock(true, b); | |
365 | continue; | |
366 | } | |
367 | ||
368 | if (!best) | |
369 | best = b; | |
370 | else if (journal_pin_cmp(c, | |
371 | btree_current_write(best), | |
372 | btree_current_write(b))) { | |
373 | rw_unlock(true, best); | |
374 | best = b; | |
375 | } else | |
376 | rw_unlock(true, b); | |
377 | } | |
378 | ||
379 | if (best) | |
380 | goto out; | |
381 | ||
382 | /* We can't find the best btree node, just pick the first */ | |
383 | list_for_each_entry(b, &c->btree_cache, list) | |
384 | if (!b->level && btree_node_dirty(b)) { | |
385 | best = b; | |
386 | rw_lock(true, best, best->level); | |
387 | goto found; | |
388 | } | |
389 | ||
390 | out: | |
391 | if (!best) | |
392 | return; | |
393 | found: | |
394 | if (btree_node_dirty(best)) | |
57943511 | 395 | bch_btree_node_write(best, NULL); |
cafe5635 KO |
396 | rw_unlock(true, best); |
397 | } | |
398 | ||
399 | #define last_seq(j) ((j)->seq - fifo_used(&(j)->pin) + 1) | |
400 | ||
401 | static void journal_discard_endio(struct bio *bio, int error) | |
402 | { | |
403 | struct journal_device *ja = | |
404 | container_of(bio, struct journal_device, discard_bio); | |
405 | struct cache *ca = container_of(ja, struct cache, journal); | |
406 | ||
407 | atomic_set(&ja->discard_in_flight, DISCARD_DONE); | |
408 | ||
409 | closure_wake_up(&ca->set->journal.wait); | |
410 | closure_put(&ca->set->cl); | |
411 | } | |
412 | ||
413 | static void journal_discard_work(struct work_struct *work) | |
414 | { | |
415 | struct journal_device *ja = | |
416 | container_of(work, struct journal_device, discard_work); | |
417 | ||
418 | submit_bio(0, &ja->discard_bio); | |
419 | } | |
420 | ||
421 | static void do_journal_discard(struct cache *ca) | |
422 | { | |
423 | struct journal_device *ja = &ca->journal; | |
424 | struct bio *bio = &ja->discard_bio; | |
425 | ||
426 | if (!ca->discard) { | |
427 | ja->discard_idx = ja->last_idx; | |
428 | return; | |
429 | } | |
430 | ||
431 | switch (atomic_read(&ja->discard_in_flight) == DISCARD_IN_FLIGHT) { | |
432 | case DISCARD_IN_FLIGHT: | |
433 | return; | |
434 | ||
435 | case DISCARD_DONE: | |
436 | ja->discard_idx = (ja->discard_idx + 1) % | |
437 | ca->sb.njournal_buckets; | |
438 | ||
439 | atomic_set(&ja->discard_in_flight, DISCARD_READY); | |
440 | /* fallthrough */ | |
441 | ||
442 | case DISCARD_READY: | |
443 | if (ja->discard_idx == ja->last_idx) | |
444 | return; | |
445 | ||
446 | atomic_set(&ja->discard_in_flight, DISCARD_IN_FLIGHT); | |
447 | ||
448 | bio_init(bio); | |
449 | bio->bi_sector = bucket_to_sector(ca->set, | |
b1a67b0f | 450 | ca->sb.d[ja->discard_idx]); |
cafe5635 KO |
451 | bio->bi_bdev = ca->bdev; |
452 | bio->bi_rw = REQ_WRITE|REQ_DISCARD; | |
453 | bio->bi_max_vecs = 1; | |
454 | bio->bi_io_vec = bio->bi_inline_vecs; | |
455 | bio->bi_size = bucket_bytes(ca); | |
456 | bio->bi_end_io = journal_discard_endio; | |
457 | ||
458 | closure_get(&ca->set->cl); | |
459 | INIT_WORK(&ja->discard_work, journal_discard_work); | |
460 | schedule_work(&ja->discard_work); | |
461 | } | |
462 | } | |
463 | ||
464 | static void journal_reclaim(struct cache_set *c) | |
465 | { | |
466 | struct bkey *k = &c->journal.key; | |
467 | struct cache *ca; | |
468 | uint64_t last_seq; | |
469 | unsigned iter, n = 0; | |
470 | atomic_t p; | |
471 | ||
472 | while (!atomic_read(&fifo_front(&c->journal.pin))) | |
473 | fifo_pop(&c->journal.pin, p); | |
474 | ||
475 | last_seq = last_seq(&c->journal); | |
476 | ||
477 | /* Update last_idx */ | |
478 | ||
479 | for_each_cache(ca, c, iter) { | |
480 | struct journal_device *ja = &ca->journal; | |
481 | ||
482 | while (ja->last_idx != ja->cur_idx && | |
483 | ja->seq[ja->last_idx] < last_seq) | |
484 | ja->last_idx = (ja->last_idx + 1) % | |
485 | ca->sb.njournal_buckets; | |
486 | } | |
487 | ||
488 | for_each_cache(ca, c, iter) | |
489 | do_journal_discard(ca); | |
490 | ||
491 | if (c->journal.blocks_free) | |
492 | return; | |
493 | ||
494 | /* | |
495 | * Allocate: | |
496 | * XXX: Sort by free journal space | |
497 | */ | |
498 | ||
499 | for_each_cache(ca, c, iter) { | |
500 | struct journal_device *ja = &ca->journal; | |
501 | unsigned next = (ja->cur_idx + 1) % ca->sb.njournal_buckets; | |
502 | ||
503 | /* No space available on this device */ | |
504 | if (next == ja->discard_idx) | |
505 | continue; | |
506 | ||
507 | ja->cur_idx = next; | |
508 | k->ptr[n++] = PTR(0, | |
509 | bucket_to_sector(c, ca->sb.d[ja->cur_idx]), | |
510 | ca->sb.nr_this_dev); | |
511 | } | |
512 | ||
513 | bkey_init(k); | |
514 | SET_KEY_PTRS(k, n); | |
515 | ||
516 | if (n) | |
517 | c->journal.blocks_free = c->sb.bucket_size >> c->block_bits; | |
518 | ||
519 | if (!journal_full(&c->journal)) | |
520 | __closure_wake_up(&c->journal.wait); | |
521 | } | |
522 | ||
523 | void bch_journal_next(struct journal *j) | |
524 | { | |
525 | atomic_t p = { 1 }; | |
526 | ||
527 | j->cur = (j->cur == j->w) | |
528 | ? &j->w[1] | |
529 | : &j->w[0]; | |
530 | ||
531 | /* | |
532 | * The fifo_push() needs to happen at the same time as j->seq is | |
533 | * incremented for last_seq() to be calculated correctly | |
534 | */ | |
535 | BUG_ON(!fifo_push(&j->pin, p)); | |
536 | atomic_set(&fifo_back(&j->pin), 1); | |
537 | ||
538 | j->cur->data->seq = ++j->seq; | |
539 | j->cur->need_write = false; | |
540 | j->cur->data->keys = 0; | |
541 | ||
542 | if (fifo_full(&j->pin)) | |
543 | pr_debug("journal_pin full (%zu)", fifo_used(&j->pin)); | |
544 | } | |
545 | ||
546 | static void journal_write_endio(struct bio *bio, int error) | |
547 | { | |
548 | struct journal_write *w = bio->bi_private; | |
549 | ||
550 | cache_set_err_on(error, w->c, "journal io error"); | |
551 | closure_put(&w->c->journal.io.cl); | |
552 | } | |
553 | ||
554 | static void journal_write(struct closure *); | |
555 | ||
556 | static void journal_write_done(struct closure *cl) | |
557 | { | |
558 | struct journal *j = container_of(cl, struct journal, io.cl); | |
559 | struct cache_set *c = container_of(j, struct cache_set, journal); | |
560 | ||
561 | struct journal_write *w = (j->cur == j->w) | |
562 | ? &j->w[1] | |
563 | : &j->w[0]; | |
564 | ||
565 | __closure_wake_up(&w->wait); | |
566 | ||
567 | if (c->journal_delay_ms) | |
568 | closure_delay(&j->io, msecs_to_jiffies(c->journal_delay_ms)); | |
569 | ||
570 | continue_at(cl, journal_write, system_wq); | |
571 | } | |
572 | ||
573 | static void journal_write_unlocked(struct closure *cl) | |
c19ed23a | 574 | __releases(c->journal.lock) |
cafe5635 KO |
575 | { |
576 | struct cache_set *c = container_of(cl, struct cache_set, journal.io.cl); | |
577 | struct cache *ca; | |
578 | struct journal_write *w = c->journal.cur; | |
579 | struct bkey *k = &c->journal.key; | |
580 | unsigned i, sectors = set_blocks(w->data, c) * c->sb.block_size; | |
581 | ||
582 | struct bio *bio; | |
583 | struct bio_list list; | |
584 | bio_list_init(&list); | |
585 | ||
586 | if (!w->need_write) { | |
587 | /* | |
588 | * XXX: have to unlock closure before we unlock journal lock, | |
589 | * else we race with bch_journal(). But this way we race | |
590 | * against cache set unregister. Doh. | |
591 | */ | |
592 | set_closure_fn(cl, NULL, NULL); | |
593 | closure_sub(cl, CLOSURE_RUNNING + 1); | |
594 | spin_unlock(&c->journal.lock); | |
595 | return; | |
596 | } else if (journal_full(&c->journal)) { | |
597 | journal_reclaim(c); | |
598 | spin_unlock(&c->journal.lock); | |
599 | ||
600 | btree_flush_write(c); | |
601 | continue_at(cl, journal_write, system_wq); | |
602 | } | |
603 | ||
604 | c->journal.blocks_free -= set_blocks(w->data, c); | |
605 | ||
606 | w->data->btree_level = c->root->level; | |
607 | ||
608 | bkey_copy(&w->data->btree_root, &c->root->key); | |
609 | bkey_copy(&w->data->uuid_bucket, &c->uuid_bucket); | |
610 | ||
611 | for_each_cache(ca, c, i) | |
612 | w->data->prio_bucket[ca->sb.nr_this_dev] = ca->prio_buckets[0]; | |
613 | ||
614 | w->data->magic = jset_magic(c); | |
615 | w->data->version = BCACHE_JSET_VERSION; | |
616 | w->data->last_seq = last_seq(&c->journal); | |
617 | w->data->csum = csum_set(w->data); | |
618 | ||
619 | for (i = 0; i < KEY_PTRS(k); i++) { | |
620 | ca = PTR_CACHE(c, k, i); | |
621 | bio = &ca->journal.bio; | |
622 | ||
623 | atomic_long_add(sectors, &ca->meta_sectors_written); | |
624 | ||
625 | bio_reset(bio); | |
626 | bio->bi_sector = PTR_OFFSET(k, i); | |
627 | bio->bi_bdev = ca->bdev; | |
e49c7c37 | 628 | bio->bi_rw = REQ_WRITE|REQ_SYNC|REQ_META|REQ_FLUSH|REQ_FUA; |
cafe5635 KO |
629 | bio->bi_size = sectors << 9; |
630 | ||
631 | bio->bi_end_io = journal_write_endio; | |
632 | bio->bi_private = w; | |
169ef1cf | 633 | bch_bio_map(bio, w->data); |
cafe5635 KO |
634 | |
635 | trace_bcache_journal_write(bio); | |
636 | bio_list_add(&list, bio); | |
637 | ||
638 | SET_PTR_OFFSET(k, i, PTR_OFFSET(k, i) + sectors); | |
639 | ||
640 | ca->journal.seq[ca->journal.cur_idx] = w->data->seq; | |
641 | } | |
642 | ||
643 | atomic_dec_bug(&fifo_back(&c->journal.pin)); | |
644 | bch_journal_next(&c->journal); | |
645 | journal_reclaim(c); | |
646 | ||
647 | spin_unlock(&c->journal.lock); | |
648 | ||
649 | while ((bio = bio_list_pop(&list))) | |
650 | closure_bio_submit(bio, cl, c->cache[0]); | |
651 | ||
652 | continue_at(cl, journal_write_done, NULL); | |
653 | } | |
654 | ||
655 | static void journal_write(struct closure *cl) | |
656 | { | |
657 | struct cache_set *c = container_of(cl, struct cache_set, journal.io.cl); | |
658 | ||
659 | spin_lock(&c->journal.lock); | |
660 | journal_write_unlocked(cl); | |
661 | } | |
662 | ||
663 | static void __journal_try_write(struct cache_set *c, bool noflush) | |
c19ed23a | 664 | __releases(c->journal.lock) |
cafe5635 KO |
665 | { |
666 | struct closure *cl = &c->journal.io.cl; | |
667 | ||
668 | if (!closure_trylock(cl, &c->cl)) | |
669 | spin_unlock(&c->journal.lock); | |
670 | else if (noflush && journal_full(&c->journal)) { | |
671 | spin_unlock(&c->journal.lock); | |
672 | continue_at(cl, journal_write, system_wq); | |
673 | } else | |
674 | journal_write_unlocked(cl); | |
675 | } | |
676 | ||
677 | #define journal_try_write(c) __journal_try_write(c, false) | |
678 | ||
679 | void bch_journal_meta(struct cache_set *c, struct closure *cl) | |
680 | { | |
681 | struct journal_write *w; | |
682 | ||
683 | if (CACHE_SYNC(&c->sb)) { | |
684 | spin_lock(&c->journal.lock); | |
685 | ||
686 | w = c->journal.cur; | |
687 | w->need_write = true; | |
688 | ||
689 | if (cl) | |
690 | BUG_ON(!closure_wait(&w->wait, cl)); | |
691 | ||
692 | __journal_try_write(c, true); | |
693 | } | |
694 | } | |
695 | ||
696 | /* | |
697 | * Entry point to the journalling code - bio_insert() and btree_invalidate() | |
698 | * pass bch_journal() a list of keys to be journalled, and then | |
699 | * bch_journal() hands those same keys off to btree_insert_async() | |
700 | */ | |
701 | ||
702 | void bch_journal(struct closure *cl) | |
703 | { | |
704 | struct btree_op *op = container_of(cl, struct btree_op, cl); | |
705 | struct cache_set *c = op->c; | |
706 | struct journal_write *w; | |
707 | size_t b, n = ((uint64_t *) op->keys.top) - op->keys.list; | |
708 | ||
709 | if (op->type != BTREE_INSERT || | |
710 | !CACHE_SYNC(&c->sb)) | |
711 | goto out; | |
712 | ||
713 | /* | |
714 | * If we're looping because we errored, might already be waiting on | |
715 | * another journal write: | |
716 | */ | |
717 | while (atomic_read(&cl->parent->remaining) & CLOSURE_WAITING) | |
718 | closure_sync(cl->parent); | |
719 | ||
720 | spin_lock(&c->journal.lock); | |
721 | ||
722 | if (journal_full(&c->journal)) { | |
c37511b8 KO |
723 | trace_bcache_journal_full(c); |
724 | ||
cafe5635 KO |
725 | closure_wait(&c->journal.wait, cl); |
726 | ||
727 | journal_reclaim(c); | |
728 | spin_unlock(&c->journal.lock); | |
729 | ||
730 | btree_flush_write(c); | |
731 | continue_at(cl, bch_journal, bcache_wq); | |
732 | } | |
733 | ||
734 | w = c->journal.cur; | |
735 | w->need_write = true; | |
736 | b = __set_blocks(w->data, w->data->keys + n, c); | |
737 | ||
738 | if (b * c->sb.block_size > PAGE_SECTORS << JSET_BITS || | |
739 | b > c->journal.blocks_free) { | |
c37511b8 KO |
740 | trace_bcache_journal_entry_full(c); |
741 | ||
742 | /* | |
743 | * XXX: If we were inserting so many keys that they won't fit in | |
cafe5635 KO |
744 | * an _empty_ journal write, we'll deadlock. For now, handle |
745 | * this in bch_keylist_realloc() - but something to think about. | |
746 | */ | |
747 | BUG_ON(!w->data->keys); | |
748 | ||
cafe5635 KO |
749 | BUG_ON(!closure_wait(&w->wait, cl)); |
750 | ||
751 | closure_flush(&c->journal.io); | |
752 | ||
753 | journal_try_write(c); | |
754 | continue_at(cl, bch_journal, bcache_wq); | |
755 | } | |
756 | ||
757 | memcpy(end(w->data), op->keys.list, n * sizeof(uint64_t)); | |
758 | w->data->keys += n; | |
759 | ||
760 | op->journal = &fifo_back(&c->journal.pin); | |
761 | atomic_inc(op->journal); | |
762 | ||
763 | if (op->flush_journal) { | |
764 | closure_flush(&c->journal.io); | |
765 | closure_wait(&w->wait, cl->parent); | |
766 | } | |
767 | ||
768 | journal_try_write(c); | |
769 | out: | |
770 | bch_btree_insert_async(cl); | |
771 | } | |
772 | ||
773 | void bch_journal_free(struct cache_set *c) | |
774 | { | |
775 | free_pages((unsigned long) c->journal.w[1].data, JSET_BITS); | |
776 | free_pages((unsigned long) c->journal.w[0].data, JSET_BITS); | |
777 | free_fifo(&c->journal.pin); | |
778 | } | |
779 | ||
780 | int bch_journal_alloc(struct cache_set *c) | |
781 | { | |
782 | struct journal *j = &c->journal; | |
783 | ||
784 | closure_init_unlocked(&j->io); | |
785 | spin_lock_init(&j->lock); | |
786 | ||
787 | c->journal_delay_ms = 100; | |
788 | ||
789 | j->w[0].c = c; | |
790 | j->w[1].c = c; | |
791 | ||
792 | if (!(init_fifo(&j->pin, JOURNAL_PIN, GFP_KERNEL)) || | |
793 | !(j->w[0].data = (void *) __get_free_pages(GFP_KERNEL, JSET_BITS)) || | |
794 | !(j->w[1].data = (void *) __get_free_pages(GFP_KERNEL, JSET_BITS))) | |
795 | return -ENOMEM; | |
796 | ||
797 | return 0; | |
798 | } |