1 // SPDX-License-Identifier: GPL-2.0
3 * bcache journalling code, for btree insertions
5 * Copyright 2012 Google, Inc.
13 #include <trace/events/bcache.h>
16 * Journal replay/recovery:
18 * This code is all driven from run_cache_set(); we first read the journal
19 * entries, do some other stuff, then we mark all the keys in the journal
20 * entries (same as garbage collection would), then we replay them - reinserting
21 * them into the cache in precisely the same order as they appear in the
24 * We only journal keys that go in leaf nodes, which simplifies things quite a
28 static void journal_read_endio(struct bio *bio)
30 struct closure *cl = bio->bi_private;
35 static int journal_read_bucket(struct cache *ca, struct list_head *list,
36 unsigned int bucket_index)
38 struct journal_device *ja = &ca->journal;
39 struct bio *bio = &ja->bio;
41 struct journal_replay *i;
42 struct jset *j, *data = ca->set->journal.w[0].data;
44 unsigned int len, left, offset = 0;
46 sector_t bucket = bucket_to_sector(ca->set, ca->sb.d[bucket_index]);
48 closure_init_stack(&cl);
50 pr_debug("reading %u\n", bucket_index);
52 while (offset < ca->sb.bucket_size) {
53 reread: left = ca->sb.bucket_size - offset;
54 len = min_t(unsigned int, left, PAGE_SECTORS << JSET_BITS);
57 bio->bi_iter.bi_sector = bucket + offset;
58 bio_set_dev(bio, ca->bdev);
59 bio->bi_iter.bi_size = len << 9;
61 bio->bi_end_io = journal_read_endio;
62 bio->bi_private = &cl;
63 bio_set_op_attrs(bio, REQ_OP_READ, 0);
64 bch_bio_map(bio, data);
66 closure_bio_submit(ca->set, bio, &cl);
69 /* This function could be simpler now since we no longer write
70 * journal entries that overlap bucket boundaries; this means
71 * the start of a bucket will always have a valid journal entry
72 * if it has any journal entries at all.
77 struct list_head *where;
78 size_t blocks, bytes = set_bytes(j);
80 if (j->magic != jset_magic(&ca->sb)) {
81 pr_debug("%u: bad magic\n", bucket_index);
85 if (bytes > left << 9 ||
86 bytes > PAGE_SIZE << JSET_BITS) {
87 pr_info("%u: too big, %zu bytes, offset %u\n",
88 bucket_index, bytes, offset);
95 if (j->csum != csum_set(j)) {
96 pr_info("%u: bad csum, %zu bytes, offset %u\n",
97 bucket_index, bytes, offset);
101 blocks = set_blocks(j, block_bytes(ca));
104 * Nodes in 'list' are in linear increasing order of
105 * i->j.seq, the node on head has the smallest (oldest)
106 * journal seq, the node on tail has the biggest
107 * (latest) journal seq.
111 * Check from the oldest jset for last_seq. If
112 * i->j.seq < j->last_seq, it means the oldest jset
113 * in list is expired and useless, remove it from
114 * this list. Otherwise, j is a condidate jset for
115 * further following checks.
117 while (!list_empty(list)) {
118 i = list_first_entry(list,
119 struct journal_replay, list);
120 if (i->j.seq >= j->last_seq)
126 /* iterate list in reverse order (from latest jset) */
127 list_for_each_entry_reverse(i, list, list) {
128 if (j->seq == i->j.seq)
132 * if j->seq is less than any i->j.last_seq
133 * in list, j is an expired and useless jset.
135 if (j->seq < i->j.last_seq)
139 * 'where' points to first jset in list which
142 if (j->seq > i->j.seq) {
150 i = kmalloc(offsetof(struct journal_replay, j) +
154 memcpy(&i->j, j, bytes);
155 /* Add to the location after 'where' points to */
156 list_add(&i->list, where);
159 if (j->seq > ja->seq[bucket_index])
160 ja->seq[bucket_index] = j->seq;
162 offset += blocks * ca->sb.block_size;
163 len -= blocks * ca->sb.block_size;
164 j = ((void *) j) + blocks * block_bytes(ca);
171 int bch_journal_read(struct cache_set *c, struct list_head *list)
173 #define read_bucket(b) \
175 ret = journal_read_bucket(ca, list, b); \
176 __set_bit(b, bitmap); \
182 struct cache *ca = c->cache;
184 struct journal_device *ja = &ca->journal;
185 DECLARE_BITMAP(bitmap, SB_JOURNAL_BUCKETS);
186 unsigned int i, l, r, m;
189 bitmap_zero(bitmap, SB_JOURNAL_BUCKETS);
190 pr_debug("%u journal buckets\n", ca->sb.njournal_buckets);
193 * Read journal buckets ordered by golden ratio hash to quickly
194 * find a sequence of buckets with valid journal entries
196 for (i = 0; i < ca->sb.njournal_buckets; i++) {
198 * We must try the index l with ZERO first for
199 * correctness due to the scenario that the journal
200 * bucket is circular buffer which might have wrapped
202 l = (i * 2654435769U) % ca->sb.njournal_buckets;
204 if (test_bit(l, bitmap))
212 * If that fails, check all the buckets we haven't checked
215 pr_debug("falling back to linear search\n");
217 for_each_clear_bit(l, bitmap, ca->sb.njournal_buckets)
221 /* no journal entries on this device? */
222 if (l == ca->sb.njournal_buckets)
225 BUG_ON(list_empty(list));
229 r = find_next_bit(bitmap, ca->sb.njournal_buckets, l + 1);
230 pr_debug("starting binary search, l %u r %u\n", l, r);
233 seq = list_entry(list->prev, struct journal_replay,
239 if (seq != list_entry(list->prev, struct journal_replay,
247 * Read buckets in reverse order until we stop finding more
250 pr_debug("finishing up: m %u njournal_buckets %u\n",
251 m, ca->sb.njournal_buckets);
256 l = ca->sb.njournal_buckets - 1;
261 if (test_bit(l, bitmap))
270 for (i = 0; i < ca->sb.njournal_buckets; i++)
271 if (ja->seq[i] > seq) {
274 * When journal_reclaim() goes to allocate for
275 * the first time, it'll use the bucket after
279 ja->last_idx = ja->discard_idx = (i + 1) %
280 ca->sb.njournal_buckets;
285 if (!list_empty(list))
286 c->journal.seq = list_entry(list->prev,
287 struct journal_replay,
294 void bch_journal_mark(struct cache_set *c, struct list_head *list)
298 struct journal_replay *i;
299 struct journal *j = &c->journal;
300 uint64_t last = j->seq;
303 * journal.pin should never fill up - we never write a journal
304 * entry when it would fill up. But if for some reason it does, we
305 * iterate over the list in reverse order so that we can just skip that
306 * refcount instead of bugging.
309 list_for_each_entry_reverse(i, list, list) {
310 BUG_ON(last < i->j.seq);
313 while (last-- != i->j.seq)
314 if (fifo_free(&j->pin) > 1) {
315 fifo_push_front(&j->pin, p);
316 atomic_set(&fifo_front(&j->pin), 0);
319 if (fifo_free(&j->pin) > 1) {
320 fifo_push_front(&j->pin, p);
321 i->pin = &fifo_front(&j->pin);
322 atomic_set(i->pin, 1);
326 k < bset_bkey_last(&i->j);
328 if (!__bch_extent_invalid(c, k)) {
331 for (j = 0; j < KEY_PTRS(k); j++)
332 if (ptr_available(c, k, j))
333 atomic_inc(&PTR_BUCKET(c, k, j)->pin);
335 bch_initial_mark_key(c, 0, k);
340 static bool is_discard_enabled(struct cache_set *s)
342 struct cache *ca = s->cache;
350 int bch_journal_replay(struct cache_set *s, struct list_head *list)
352 int ret = 0, keys = 0, entries = 0;
354 struct journal_replay *i =
355 list_entry(list->prev, struct journal_replay, list);
357 uint64_t start = i->j.last_seq, end = i->j.seq, n = start;
358 struct keylist keylist;
360 list_for_each_entry(i, list, list) {
361 BUG_ON(i->pin && atomic_read(i->pin) != 1);
364 if (n == start && is_discard_enabled(s))
365 pr_info("journal entries %llu-%llu may be discarded! (replaying %llu-%llu)\n",
366 n, i->j.seq - 1, start, end);
368 pr_err("journal entries %llu-%llu missing! (replaying %llu-%llu)\n",
369 n, i->j.seq - 1, start, end);
376 k < bset_bkey_last(&i->j);
378 trace_bcache_journal_replay_key(k);
380 bch_keylist_init_single(&keylist, k);
382 ret = bch_btree_insert(s, &keylist, i->pin, NULL);
386 BUG_ON(!bch_keylist_empty(&keylist));
398 pr_info("journal replay done, %i keys in %i entries, seq %llu\n",
401 while (!list_empty(list)) {
402 i = list_first_entry(list, struct journal_replay, list);
412 static void btree_flush_write(struct cache_set *c)
414 struct btree *b, *t, *btree_nodes[BTREE_FLUSH_NR];
417 atomic_t *fifo_front_p, *now_fifo_front_p;
420 if (c->journal.btree_flushing)
423 spin_lock(&c->journal.flush_write_lock);
424 if (c->journal.btree_flushing) {
425 spin_unlock(&c->journal.flush_write_lock);
428 c->journal.btree_flushing = true;
429 spin_unlock(&c->journal.flush_write_lock);
431 /* get the oldest journal entry and check its refcount */
432 spin_lock(&c->journal.lock);
433 fifo_front_p = &fifo_front(&c->journal.pin);
434 ref_nr = atomic_read(fifo_front_p);
437 * do nothing if no btree node references
438 * the oldest journal entry
440 spin_unlock(&c->journal.lock);
443 spin_unlock(&c->journal.lock);
445 mask = c->journal.pin.mask;
447 atomic_long_inc(&c->flush_write);
448 memset(btree_nodes, 0, sizeof(btree_nodes));
450 mutex_lock(&c->bucket_lock);
451 list_for_each_entry_safe_reverse(b, t, &c->btree_cache, list) {
453 * It is safe to get now_fifo_front_p without holding
454 * c->journal.lock here, because we don't need to know
455 * the exactly accurate value, just check whether the
456 * front pointer of c->journal.pin is changed.
458 now_fifo_front_p = &fifo_front(&c->journal.pin);
460 * If the oldest journal entry is reclaimed and front
461 * pointer of c->journal.pin changes, it is unnecessary
462 * to scan c->btree_cache anymore, just quit the loop and
463 * flush out what we have already.
465 if (now_fifo_front_p != fifo_front_p)
468 * quit this loop if all matching btree nodes are
469 * scanned and record in btree_nodes[] already.
471 ref_nr = atomic_read(fifo_front_p);
475 if (btree_node_journal_flush(b))
476 pr_err("BUG: flush_write bit should not be set here!\n");
478 mutex_lock(&b->write_lock);
480 if (!btree_node_dirty(b)) {
481 mutex_unlock(&b->write_lock);
485 if (!btree_current_write(b)->journal) {
486 mutex_unlock(&b->write_lock);
491 * Only select the btree node which exactly references
492 * the oldest journal entry.
494 * If the journal entry pointed by fifo_front_p is
495 * reclaimed in parallel, don't worry:
496 * - the list_for_each_xxx loop will quit when checking
497 * next now_fifo_front_p.
498 * - If there are matched nodes recorded in btree_nodes[],
499 * they are clean now (this is why and how the oldest
500 * journal entry can be reclaimed). These selected nodes
501 * will be ignored and skipped in the folowing for-loop.
503 if (((btree_current_write(b)->journal - fifo_front_p) &
505 mutex_unlock(&b->write_lock);
509 set_btree_node_journal_flush(b);
511 mutex_unlock(&b->write_lock);
513 btree_nodes[nr++] = b;
515 * To avoid holding c->bucket_lock too long time,
516 * only scan for BTREE_FLUSH_NR matched btree nodes
517 * at most. If there are more btree nodes reference
518 * the oldest journal entry, try to flush them next
519 * time when btree_flush_write() is called.
521 if (nr == BTREE_FLUSH_NR)
524 mutex_unlock(&c->bucket_lock);
526 for (i = 0; i < nr; i++) {
529 pr_err("BUG: btree_nodes[%d] is NULL\n", i);
533 /* safe to check without holding b->write_lock */
534 if (!btree_node_journal_flush(b)) {
535 pr_err("BUG: bnode %p: journal_flush bit cleaned\n", b);
539 mutex_lock(&b->write_lock);
540 if (!btree_current_write(b)->journal) {
541 clear_bit(BTREE_NODE_journal_flush, &b->flags);
542 mutex_unlock(&b->write_lock);
543 pr_debug("bnode %p: written by others\n", b);
547 if (!btree_node_dirty(b)) {
548 clear_bit(BTREE_NODE_journal_flush, &b->flags);
549 mutex_unlock(&b->write_lock);
550 pr_debug("bnode %p: dirty bit cleaned by others\n", b);
554 __bch_btree_node_write(b, NULL);
555 clear_bit(BTREE_NODE_journal_flush, &b->flags);
556 mutex_unlock(&b->write_lock);
560 spin_lock(&c->journal.flush_write_lock);
561 c->journal.btree_flushing = false;
562 spin_unlock(&c->journal.flush_write_lock);
565 #define last_seq(j) ((j)->seq - fifo_used(&(j)->pin) + 1)
567 static void journal_discard_endio(struct bio *bio)
569 struct journal_device *ja =
570 container_of(bio, struct journal_device, discard_bio);
571 struct cache *ca = container_of(ja, struct cache, journal);
573 atomic_set(&ja->discard_in_flight, DISCARD_DONE);
575 closure_wake_up(&ca->set->journal.wait);
576 closure_put(&ca->set->cl);
579 static void journal_discard_work(struct work_struct *work)
581 struct journal_device *ja =
582 container_of(work, struct journal_device, discard_work);
584 submit_bio(&ja->discard_bio);
587 static void do_journal_discard(struct cache *ca)
589 struct journal_device *ja = &ca->journal;
590 struct bio *bio = &ja->discard_bio;
593 ja->discard_idx = ja->last_idx;
597 switch (atomic_read(&ja->discard_in_flight)) {
598 case DISCARD_IN_FLIGHT:
602 ja->discard_idx = (ja->discard_idx + 1) %
603 ca->sb.njournal_buckets;
605 atomic_set(&ja->discard_in_flight, DISCARD_READY);
609 if (ja->discard_idx == ja->last_idx)
612 atomic_set(&ja->discard_in_flight, DISCARD_IN_FLIGHT);
614 bio_init(bio, bio->bi_inline_vecs, 1);
615 bio_set_op_attrs(bio, REQ_OP_DISCARD, 0);
616 bio->bi_iter.bi_sector = bucket_to_sector(ca->set,
617 ca->sb.d[ja->discard_idx]);
618 bio_set_dev(bio, ca->bdev);
619 bio->bi_iter.bi_size = bucket_bytes(ca);
620 bio->bi_end_io = journal_discard_endio;
622 closure_get(&ca->set->cl);
623 INIT_WORK(&ja->discard_work, journal_discard_work);
624 queue_work(bch_journal_wq, &ja->discard_work);
628 static void journal_reclaim(struct cache_set *c)
630 struct bkey *k = &c->journal.key;
631 struct cache *ca = c->cache;
634 struct journal_device *ja = &ca->journal;
635 atomic_t p __maybe_unused;
637 atomic_long_inc(&c->reclaim);
639 while (!atomic_read(&fifo_front(&c->journal.pin)))
640 fifo_pop(&c->journal.pin, p);
642 last_seq = last_seq(&c->journal);
644 /* Update last_idx */
646 while (ja->last_idx != ja->cur_idx &&
647 ja->seq[ja->last_idx] < last_seq)
648 ja->last_idx = (ja->last_idx + 1) %
649 ca->sb.njournal_buckets;
651 do_journal_discard(ca);
653 if (c->journal.blocks_free)
656 next = (ja->cur_idx + 1) % ca->sb.njournal_buckets;
657 /* No space available on this device */
658 if (next == ja->discard_idx)
662 k->ptr[0] = MAKE_PTR(0,
663 bucket_to_sector(c, ca->sb.d[ja->cur_idx]),
665 atomic_long_inc(&c->reclaimed_journal_buckets);
669 c->journal.blocks_free = ca->sb.bucket_size >> c->block_bits;
672 if (!journal_full(&c->journal))
673 __closure_wake_up(&c->journal.wait);
676 void bch_journal_next(struct journal *j)
680 j->cur = (j->cur == j->w)
685 * The fifo_push() needs to happen at the same time as j->seq is
686 * incremented for last_seq() to be calculated correctly
688 BUG_ON(!fifo_push(&j->pin, p));
689 atomic_set(&fifo_back(&j->pin), 1);
691 j->cur->data->seq = ++j->seq;
692 j->cur->dirty = false;
693 j->cur->need_write = false;
694 j->cur->data->keys = 0;
696 if (fifo_full(&j->pin))
697 pr_debug("journal_pin full (%zu)\n", fifo_used(&j->pin));
700 static void journal_write_endio(struct bio *bio)
702 struct journal_write *w = bio->bi_private;
704 cache_set_err_on(bio->bi_status, w->c, "journal io error");
705 closure_put(&w->c->journal.io);
708 static void journal_write(struct closure *cl);
710 static void journal_write_done(struct closure *cl)
712 struct journal *j = container_of(cl, struct journal, io);
713 struct journal_write *w = (j->cur == j->w)
717 __closure_wake_up(&w->wait);
718 continue_at_nobarrier(cl, journal_write, bch_journal_wq);
721 static void journal_write_unlock(struct closure *cl)
722 __releases(&c->journal.lock)
724 struct cache_set *c = container_of(cl, struct cache_set, journal.io);
726 c->journal.io_in_flight = 0;
727 spin_unlock(&c->journal.lock);
730 static void journal_write_unlocked(struct closure *cl)
731 __releases(c->journal.lock)
733 struct cache_set *c = container_of(cl, struct cache_set, journal.io);
734 struct cache *ca = c->cache;
735 struct journal_write *w = c->journal.cur;
736 struct bkey *k = &c->journal.key;
737 unsigned int i, sectors = set_blocks(w->data, block_bytes(ca)) *
741 struct bio_list list;
743 bio_list_init(&list);
745 if (!w->need_write) {
746 closure_return_with_destructor(cl, journal_write_unlock);
748 } else if (journal_full(&c->journal)) {
750 spin_unlock(&c->journal.lock);
752 btree_flush_write(c);
753 continue_at(cl, journal_write, bch_journal_wq);
757 c->journal.blocks_free -= set_blocks(w->data, block_bytes(ca));
759 w->data->btree_level = c->root->level;
761 bkey_copy(&w->data->btree_root, &c->root->key);
762 bkey_copy(&w->data->uuid_bucket, &c->uuid_bucket);
764 w->data->prio_bucket[ca->sb.nr_this_dev] = ca->prio_buckets[0];
765 w->data->magic = jset_magic(&ca->sb);
766 w->data->version = BCACHE_JSET_VERSION;
767 w->data->last_seq = last_seq(&c->journal);
768 w->data->csum = csum_set(w->data);
770 for (i = 0; i < KEY_PTRS(k); i++) {
771 ca = PTR_CACHE(c, k, i);
772 bio = &ca->journal.bio;
774 atomic_long_add(sectors, &ca->meta_sectors_written);
777 bio->bi_iter.bi_sector = PTR_OFFSET(k, i);
778 bio_set_dev(bio, ca->bdev);
779 bio->bi_iter.bi_size = sectors << 9;
781 bio->bi_end_io = journal_write_endio;
783 bio_set_op_attrs(bio, REQ_OP_WRITE,
784 REQ_SYNC|REQ_META|REQ_PREFLUSH|REQ_FUA);
785 bch_bio_map(bio, w->data);
787 trace_bcache_journal_write(bio, w->data->keys);
788 bio_list_add(&list, bio);
790 SET_PTR_OFFSET(k, i, PTR_OFFSET(k, i) + sectors);
792 ca->journal.seq[ca->journal.cur_idx] = w->data->seq;
795 /* If KEY_PTRS(k) == 0, this jset gets lost in air */
798 atomic_dec_bug(&fifo_back(&c->journal.pin));
799 bch_journal_next(&c->journal);
802 spin_unlock(&c->journal.lock);
804 while ((bio = bio_list_pop(&list)))
805 closure_bio_submit(c, bio, cl);
807 continue_at(cl, journal_write_done, NULL);
810 static void journal_write(struct closure *cl)
812 struct cache_set *c = container_of(cl, struct cache_set, journal.io);
814 spin_lock(&c->journal.lock);
815 journal_write_unlocked(cl);
818 static void journal_try_write(struct cache_set *c)
819 __releases(c->journal.lock)
821 struct closure *cl = &c->journal.io;
822 struct journal_write *w = c->journal.cur;
824 w->need_write = true;
826 if (!c->journal.io_in_flight) {
827 c->journal.io_in_flight = 1;
828 closure_call(cl, journal_write_unlocked, NULL, &c->cl);
830 spin_unlock(&c->journal.lock);
834 static struct journal_write *journal_wait_for_write(struct cache_set *c,
836 __acquires(&c->journal.lock)
841 struct cache *ca = c->cache;
843 closure_init_stack(&cl);
845 spin_lock(&c->journal.lock);
848 struct journal_write *w = c->journal.cur;
850 sectors = __set_blocks(w->data, w->data->keys + nkeys,
851 block_bytes(ca)) * ca->sb.block_size;
853 if (sectors <= min_t(size_t,
854 c->journal.blocks_free * ca->sb.block_size,
855 PAGE_SECTORS << JSET_BITS))
859 closure_wait(&c->journal.wait, &cl);
861 if (!journal_full(&c->journal)) {
863 trace_bcache_journal_entry_full(c);
866 * XXX: If we were inserting so many keys that they
867 * won't fit in an _empty_ journal write, we'll
868 * deadlock. For now, handle this in
869 * bch_keylist_realloc() - but something to think about.
871 BUG_ON(!w->data->keys);
873 journal_try_write(c); /* unlocks */
876 trace_bcache_journal_full(c);
879 spin_unlock(&c->journal.lock);
881 btree_flush_write(c);
885 spin_lock(&c->journal.lock);
890 static void journal_write_work(struct work_struct *work)
892 struct cache_set *c = container_of(to_delayed_work(work),
895 spin_lock(&c->journal.lock);
896 if (c->journal.cur->dirty)
897 journal_try_write(c);
899 spin_unlock(&c->journal.lock);
903 * Entry point to the journalling code - bio_insert() and btree_invalidate()
904 * pass bch_journal() a list of keys to be journalled, and then
905 * bch_journal() hands those same keys off to btree_insert_async()
908 atomic_t *bch_journal(struct cache_set *c,
909 struct keylist *keys,
910 struct closure *parent)
912 struct journal_write *w;
915 /* No journaling if CACHE_SET_IO_DISABLE set already */
916 if (unlikely(test_bit(CACHE_SET_IO_DISABLE, &c->flags)))
919 if (!CACHE_SYNC(&c->cache->sb))
922 w = journal_wait_for_write(c, bch_keylist_nkeys(keys));
924 memcpy(bset_bkey_last(w->data), keys->keys, bch_keylist_bytes(keys));
925 w->data->keys += bch_keylist_nkeys(keys);
927 ret = &fifo_back(&c->journal.pin);
931 closure_wait(&w->wait, parent);
932 journal_try_write(c);
933 } else if (!w->dirty) {
935 schedule_delayed_work(&c->journal.work,
936 msecs_to_jiffies(c->journal_delay_ms));
937 spin_unlock(&c->journal.lock);
939 spin_unlock(&c->journal.lock);
946 void bch_journal_meta(struct cache_set *c, struct closure *cl)
951 bch_keylist_init(&keys);
953 ref = bch_journal(c, &keys, cl);
958 void bch_journal_free(struct cache_set *c)
960 free_pages((unsigned long) c->journal.w[1].data, JSET_BITS);
961 free_pages((unsigned long) c->journal.w[0].data, JSET_BITS);
962 free_fifo(&c->journal.pin);
965 int bch_journal_alloc(struct cache_set *c)
967 struct journal *j = &c->journal;
969 spin_lock_init(&j->lock);
970 spin_lock_init(&j->flush_write_lock);
971 INIT_DELAYED_WORK(&j->work, journal_write_work);
973 c->journal_delay_ms = 100;
978 if (!(init_fifo(&j->pin, JOURNAL_PIN, GFP_KERNEL)) ||
979 !(j->w[0].data = (void *) __get_free_pages(GFP_KERNEL|__GFP_COMP, JSET_BITS)) ||
980 !(j->w[1].data = (void *) __get_free_pages(GFP_KERNEL|__GFP_COMP, JSET_BITS)))