2 * bcache journalling code, for btree insertions
4 * Copyright 2012 Google, Inc.
12 #include <trace/events/bcache.h>
15 * Journal replay/recovery:
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
23 * We only journal keys that go in leaf nodes, which simplifies things quite a
27 static void journal_read_endio(struct bio *bio, int error)
29 struct closure *cl = bio->bi_private;
33 static int journal_read_bucket(struct cache *ca, struct list_head *list,
34 struct btree_op *op, unsigned bucket_index)
36 struct journal_device *ja = &ca->journal;
37 struct bio *bio = &ja->bio;
39 struct journal_replay *i;
40 struct jset *j, *data = ca->set->journal.w[0].data;
41 unsigned len, left, offset = 0;
43 sector_t bucket = bucket_to_sector(ca->set, ca->sb.d[bucket_index]);
45 pr_debug("reading %llu", (uint64_t) bucket);
47 while (offset < ca->sb.bucket_size) {
48 reread: left = ca->sb.bucket_size - offset;
49 len = min_t(unsigned, left, PAGE_SECTORS * 8);
52 bio->bi_sector = bucket + offset;
53 bio->bi_bdev = ca->bdev;
55 bio->bi_size = len << 9;
57 bio->bi_end_io = journal_read_endio;
58 bio->bi_private = &op->cl;
59 bch_bio_map(bio, data);
61 closure_bio_submit(bio, &op->cl, ca);
62 closure_sync(&op->cl);
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.
72 struct list_head *where;
73 size_t blocks, bytes = set_bytes(j);
75 if (j->magic != jset_magic(ca->set))
78 if (bytes > left << 9)
84 if (j->csum != csum_set(j))
87 blocks = set_blocks(j, ca->set);
89 while (!list_empty(list)) {
90 i = list_first_entry(list,
91 struct journal_replay, list);
92 if (i->j.seq >= j->last_seq)
98 list_for_each_entry_reverse(i, list, list) {
99 if (j->seq == i->j.seq)
102 if (j->seq < i->j.last_seq)
105 if (j->seq > i->j.seq) {
113 i = kmalloc(offsetof(struct journal_replay, j) +
117 memcpy(&i->j, j, bytes);
118 list_add(&i->list, where);
121 ja->seq[bucket_index] = j->seq;
123 offset += blocks * ca->sb.block_size;
124 len -= blocks * ca->sb.block_size;
125 j = ((void *) j) + blocks * block_bytes(ca);
132 int bch_journal_read(struct cache_set *c, struct list_head *list,
135 #define read_bucket(b) \
137 int ret = journal_read_bucket(ca, list, op, b); \
138 __set_bit(b, bitmap); \
147 for_each_cache(ca, c, iter) {
148 struct journal_device *ja = &ca->journal;
149 unsigned long bitmap[SB_JOURNAL_BUCKETS / BITS_PER_LONG];
153 bitmap_zero(bitmap, SB_JOURNAL_BUCKETS);
154 pr_debug("%u journal buckets", ca->sb.njournal_buckets);
157 * Read journal buckets ordered by golden ratio hash to quickly
158 * find a sequence of buckets with valid journal entries
160 for (i = 0; i < ca->sb.njournal_buckets; i++) {
161 l = (i * 2654435769U) % ca->sb.njournal_buckets;
163 if (test_bit(l, bitmap))
171 * If that fails, check all the buckets we haven't checked
174 pr_debug("falling back to linear search");
176 for (l = find_first_zero_bit(bitmap, ca->sb.njournal_buckets);
177 l < ca->sb.njournal_buckets;
178 l = find_next_zero_bit(bitmap, ca->sb.njournal_buckets, l + 1))
182 if (list_empty(list))
186 m = r = find_next_bit(bitmap, ca->sb.njournal_buckets, l + 1);
187 pr_debug("starting binary search, l %u r %u", l, r);
190 seq = list_entry(list->prev, struct journal_replay,
196 if (seq != list_entry(list->prev, struct journal_replay,
204 * Read buckets in reverse order until we stop finding more
207 pr_debug("finishing up: m %u njournal_buckets %u",
208 m, ca->sb.njournal_buckets);
213 l = ca->sb.njournal_buckets - 1;
218 if (test_bit(l, bitmap))
227 for (i = 0; i < ca->sb.njournal_buckets; i++)
228 if (ja->seq[i] > seq) {
230 ja->cur_idx = ja->discard_idx =
236 if (!list_empty(list))
237 c->journal.seq = list_entry(list->prev,
238 struct journal_replay,
245 void bch_journal_mark(struct cache_set *c, struct list_head *list)
249 struct journal_replay *i;
250 struct journal *j = &c->journal;
251 uint64_t last = j->seq;
254 * journal.pin should never fill up - we never write a journal
255 * entry when it would fill up. But if for some reason it does, we
256 * iterate over the list in reverse order so that we can just skip that
257 * refcount instead of bugging.
260 list_for_each_entry_reverse(i, list, list) {
261 BUG_ON(last < i->j.seq);
264 while (last-- != i->j.seq)
265 if (fifo_free(&j->pin) > 1) {
266 fifo_push_front(&j->pin, p);
267 atomic_set(&fifo_front(&j->pin), 0);
270 if (fifo_free(&j->pin) > 1) {
271 fifo_push_front(&j->pin, p);
272 i->pin = &fifo_front(&j->pin);
273 atomic_set(i->pin, 1);
281 for (j = 0; j < KEY_PTRS(k); j++) {
282 struct bucket *g = PTR_BUCKET(c, k, j);
285 if (g->prio == BTREE_PRIO &&
287 g->prio = INITIAL_PRIO;
290 __bch_btree_mark_key(c, 0, k);
295 int bch_journal_replay(struct cache_set *s, struct list_head *list,
298 int ret = 0, keys = 0, entries = 0;
300 struct journal_replay *i =
301 list_entry(list->prev, struct journal_replay, list);
303 uint64_t start = i->j.last_seq, end = i->j.seq, n = start;
304 struct keylist keylist;
306 bch_keylist_init(&keylist);
308 list_for_each_entry(i, list, list) {
309 BUG_ON(i->pin && atomic_read(i->pin) != 1);
311 cache_set_err_on(n != i->j.seq, s,
312 "bcache: journal entries %llu-%llu missing! (replaying %llu-%llu)",
313 n, i->j.seq - 1, start, end);
318 trace_bcache_journal_replay_key(k);
320 bkey_copy(keylist.top, k);
321 bch_keylist_push(&keylist);
323 op->journal = i->pin;
325 ret = bch_btree_insert(op, s, &keylist);
329 BUG_ON(!bch_keylist_empty(&keylist));
341 pr_info("journal replay done, %i keys in %i entries, seq %llu",
344 while (!list_empty(list)) {
345 i = list_first_entry(list, struct journal_replay, list);
350 closure_sync(&op->cl);
356 static void btree_flush_write(struct cache_set *c)
359 * Try to find the btree node with that references the oldest journal
360 * entry, best is our current candidate and is locked if non NULL:
362 struct btree *b, *best;
367 for_each_cached_btree(b, c, i)
368 if (btree_current_write(b)->journal) {
371 else if (journal_pin_cmp(c,
372 btree_current_write(best),
373 btree_current_write(b))) {
380 rw_lock(true, b, b->level);
382 if (!btree_current_write(b)->journal) {
388 bch_btree_node_write(b, NULL);
393 #define last_seq(j) ((j)->seq - fifo_used(&(j)->pin) + 1)
395 static void journal_discard_endio(struct bio *bio, int error)
397 struct journal_device *ja =
398 container_of(bio, struct journal_device, discard_bio);
399 struct cache *ca = container_of(ja, struct cache, journal);
401 atomic_set(&ja->discard_in_flight, DISCARD_DONE);
403 closure_wake_up(&ca->set->journal.wait);
404 closure_put(&ca->set->cl);
407 static void journal_discard_work(struct work_struct *work)
409 struct journal_device *ja =
410 container_of(work, struct journal_device, discard_work);
412 submit_bio(0, &ja->discard_bio);
415 static void do_journal_discard(struct cache *ca)
417 struct journal_device *ja = &ca->journal;
418 struct bio *bio = &ja->discard_bio;
421 ja->discard_idx = ja->last_idx;
425 switch (atomic_read(&ja->discard_in_flight)) {
426 case DISCARD_IN_FLIGHT:
430 ja->discard_idx = (ja->discard_idx + 1) %
431 ca->sb.njournal_buckets;
433 atomic_set(&ja->discard_in_flight, DISCARD_READY);
437 if (ja->discard_idx == ja->last_idx)
440 atomic_set(&ja->discard_in_flight, DISCARD_IN_FLIGHT);
443 bio->bi_sector = bucket_to_sector(ca->set,
444 ca->sb.d[ja->discard_idx]);
445 bio->bi_bdev = ca->bdev;
446 bio->bi_rw = REQ_WRITE|REQ_DISCARD;
447 bio->bi_max_vecs = 1;
448 bio->bi_io_vec = bio->bi_inline_vecs;
449 bio->bi_size = bucket_bytes(ca);
450 bio->bi_end_io = journal_discard_endio;
452 closure_get(&ca->set->cl);
453 INIT_WORK(&ja->discard_work, journal_discard_work);
454 schedule_work(&ja->discard_work);
458 static void journal_reclaim(struct cache_set *c)
460 struct bkey *k = &c->journal.key;
463 unsigned iter, n = 0;
466 while (!atomic_read(&fifo_front(&c->journal.pin)))
467 fifo_pop(&c->journal.pin, p);
469 last_seq = last_seq(&c->journal);
471 /* Update last_idx */
473 for_each_cache(ca, c, iter) {
474 struct journal_device *ja = &ca->journal;
476 while (ja->last_idx != ja->cur_idx &&
477 ja->seq[ja->last_idx] < last_seq)
478 ja->last_idx = (ja->last_idx + 1) %
479 ca->sb.njournal_buckets;
482 for_each_cache(ca, c, iter)
483 do_journal_discard(ca);
485 if (c->journal.blocks_free)
490 * XXX: Sort by free journal space
493 for_each_cache(ca, c, iter) {
494 struct journal_device *ja = &ca->journal;
495 unsigned next = (ja->cur_idx + 1) % ca->sb.njournal_buckets;
497 /* No space available on this device */
498 if (next == ja->discard_idx)
503 bucket_to_sector(c, ca->sb.d[ja->cur_idx]),
511 c->journal.blocks_free = c->sb.bucket_size >> c->block_bits;
513 if (!journal_full(&c->journal))
514 __closure_wake_up(&c->journal.wait);
517 void bch_journal_next(struct journal *j)
521 j->cur = (j->cur == j->w)
526 * The fifo_push() needs to happen at the same time as j->seq is
527 * incremented for last_seq() to be calculated correctly
529 BUG_ON(!fifo_push(&j->pin, p));
530 atomic_set(&fifo_back(&j->pin), 1);
532 j->cur->data->seq = ++j->seq;
533 j->cur->need_write = false;
534 j->cur->data->keys = 0;
536 if (fifo_full(&j->pin))
537 pr_debug("journal_pin full (%zu)", fifo_used(&j->pin));
540 static void journal_write_endio(struct bio *bio, int error)
542 struct journal_write *w = bio->bi_private;
544 cache_set_err_on(error, w->c, "journal io error");
545 closure_put(&w->c->journal.io);
548 static void journal_write(struct closure *);
550 static void journal_write_done(struct closure *cl)
552 struct journal *j = container_of(cl, struct journal, io);
553 struct journal_write *w = (j->cur == j->w)
557 __closure_wake_up(&w->wait);
558 continue_at_nobarrier(cl, journal_write, system_wq);
561 static void journal_write_unlocked(struct closure *cl)
562 __releases(c->journal.lock)
564 struct cache_set *c = container_of(cl, struct cache_set, journal.io);
566 struct journal_write *w = c->journal.cur;
567 struct bkey *k = &c->journal.key;
568 unsigned i, sectors = set_blocks(w->data, c) * c->sb.block_size;
571 struct bio_list list;
572 bio_list_init(&list);
574 if (!w->need_write) {
576 * XXX: have to unlock closure before we unlock journal lock,
577 * else we race with bch_journal(). But this way we race
578 * against cache set unregister. Doh.
580 set_closure_fn(cl, NULL, NULL);
581 closure_sub(cl, CLOSURE_RUNNING + 1);
582 spin_unlock(&c->journal.lock);
584 } else if (journal_full(&c->journal)) {
586 spin_unlock(&c->journal.lock);
588 btree_flush_write(c);
589 continue_at(cl, journal_write, system_wq);
592 c->journal.blocks_free -= set_blocks(w->data, c);
594 w->data->btree_level = c->root->level;
596 bkey_copy(&w->data->btree_root, &c->root->key);
597 bkey_copy(&w->data->uuid_bucket, &c->uuid_bucket);
599 for_each_cache(ca, c, i)
600 w->data->prio_bucket[ca->sb.nr_this_dev] = ca->prio_buckets[0];
602 w->data->magic = jset_magic(c);
603 w->data->version = BCACHE_JSET_VERSION;
604 w->data->last_seq = last_seq(&c->journal);
605 w->data->csum = csum_set(w->data);
607 for (i = 0; i < KEY_PTRS(k); i++) {
608 ca = PTR_CACHE(c, k, i);
609 bio = &ca->journal.bio;
611 atomic_long_add(sectors, &ca->meta_sectors_written);
614 bio->bi_sector = PTR_OFFSET(k, i);
615 bio->bi_bdev = ca->bdev;
616 bio->bi_rw = REQ_WRITE|REQ_SYNC|REQ_META|REQ_FLUSH|REQ_FUA;
617 bio->bi_size = sectors << 9;
619 bio->bi_end_io = journal_write_endio;
621 bch_bio_map(bio, w->data);
623 trace_bcache_journal_write(bio);
624 bio_list_add(&list, bio);
626 SET_PTR_OFFSET(k, i, PTR_OFFSET(k, i) + sectors);
628 ca->journal.seq[ca->journal.cur_idx] = w->data->seq;
631 atomic_dec_bug(&fifo_back(&c->journal.pin));
632 bch_journal_next(&c->journal);
635 spin_unlock(&c->journal.lock);
637 while ((bio = bio_list_pop(&list)))
638 closure_bio_submit(bio, cl, c->cache[0]);
640 continue_at(cl, journal_write_done, NULL);
643 static void journal_write(struct closure *cl)
645 struct cache_set *c = container_of(cl, struct cache_set, journal.io);
647 spin_lock(&c->journal.lock);
648 journal_write_unlocked(cl);
651 static void journal_try_write(struct cache_set *c)
652 __releases(c->journal.lock)
654 struct closure *cl = &c->journal.io;
655 struct journal_write *w = c->journal.cur;
657 w->need_write = true;
659 if (closure_trylock(cl, &c->cl))
660 journal_write_unlocked(cl);
662 spin_unlock(&c->journal.lock);
665 static struct journal_write *journal_wait_for_write(struct cache_set *c,
671 closure_init_stack(&cl);
673 spin_lock(&c->journal.lock);
676 struct journal_write *w = c->journal.cur;
678 sectors = __set_blocks(w->data, w->data->keys + nkeys,
679 c) * c->sb.block_size;
681 if (sectors <= min_t(size_t,
682 c->journal.blocks_free * c->sb.block_size,
683 PAGE_SECTORS << JSET_BITS))
686 /* XXX: tracepoint */
687 if (!journal_full(&c->journal)) {
688 trace_bcache_journal_entry_full(c);
691 * XXX: If we were inserting so many keys that they
692 * won't fit in an _empty_ journal write, we'll
693 * deadlock. For now, handle this in
694 * bch_keylist_realloc() - but something to think about.
696 BUG_ON(!w->data->keys);
698 closure_wait(&w->wait, &cl);
699 journal_try_write(c); /* unlocks */
701 trace_bcache_journal_full(c);
703 closure_wait(&c->journal.wait, &cl);
705 spin_unlock(&c->journal.lock);
707 btree_flush_write(c);
711 spin_lock(&c->journal.lock);
715 static void journal_write_work(struct work_struct *work)
717 struct cache_set *c = container_of(to_delayed_work(work),
720 spin_lock(&c->journal.lock);
721 journal_try_write(c);
725 * Entry point to the journalling code - bio_insert() and btree_invalidate()
726 * pass bch_journal() a list of keys to be journalled, and then
727 * bch_journal() hands those same keys off to btree_insert_async()
730 atomic_t *bch_journal(struct cache_set *c,
731 struct keylist *keys,
732 struct closure *parent)
734 struct journal_write *w;
737 if (!CACHE_SYNC(&c->sb))
740 w = journal_wait_for_write(c, bch_keylist_nkeys(keys));
742 memcpy(end(w->data), keys->keys, bch_keylist_bytes(keys));
743 w->data->keys += bch_keylist_nkeys(keys);
745 ret = &fifo_back(&c->journal.pin);
749 closure_wait(&w->wait, parent);
750 journal_try_write(c);
751 } else if (!w->need_write) {
752 schedule_delayed_work(&c->journal.work,
753 msecs_to_jiffies(c->journal_delay_ms));
754 spin_unlock(&c->journal.lock);
756 spin_unlock(&c->journal.lock);
763 void bch_journal_meta(struct cache_set *c, struct closure *cl)
768 bch_keylist_init(&keys);
770 ref = bch_journal(c, &keys, cl);
775 void bch_journal_free(struct cache_set *c)
777 free_pages((unsigned long) c->journal.w[1].data, JSET_BITS);
778 free_pages((unsigned long) c->journal.w[0].data, JSET_BITS);
779 free_fifo(&c->journal.pin);
782 int bch_journal_alloc(struct cache_set *c)
784 struct journal *j = &c->journal;
786 closure_init_unlocked(&j->io);
787 spin_lock_init(&j->lock);
788 INIT_DELAYED_WORK(&j->work, journal_write_work);
790 c->journal_delay_ms = 100;
795 if (!(init_fifo(&j->pin, JOURNAL_PIN, GFP_KERNEL)) ||
796 !(j->w[0].data = (void *) __get_free_pages(GFP_KERNEL, JSET_BITS)) ||
797 !(j->w[1].data = (void *) __get_free_pages(GFP_KERNEL, JSET_BITS)))