1 // SPDX-License-Identifier: GPL-2.0
3 * Copyright (C) 2010 Kent Overstreet <kent.overstreet@gmail.com>
5 * Uses a block device as cache for other block devices; optimized for SSDs.
6 * All allocation is done in buckets, which should match the erase block size
9 * Buckets containing cached data are kept on a heap sorted by priority;
10 * bucket priority is increased on cache hit, and periodically all the buckets
11 * on the heap have their priority scaled down. This currently is just used as
12 * an LRU but in the future should allow for more intelligent heuristics.
14 * Buckets have an 8 bit counter; freeing is accomplished by incrementing the
15 * counter. Garbage collection is used to remove stale pointers.
17 * Indexing is done via a btree; nodes are not necessarily fully sorted, rather
18 * as keys are inserted we only sort the pages that have not yet been written.
19 * When garbage collection is run, we resort the entire node.
21 * All configuration is done via sysfs; see Documentation/admin-guide/bcache.rst.
29 #include <linux/slab.h>
30 #include <linux/bitops.h>
31 #include <linux/hash.h>
32 #include <linux/kthread.h>
33 #include <linux/prefetch.h>
34 #include <linux/random.h>
35 #include <linux/rcupdate.h>
36 #include <linux/sched/clock.h>
37 #include <linux/rculist.h>
38 #include <linux/delay.h>
39 #include <trace/events/bcache.h>
43 * register_bcache: Return errors out to userspace correctly
45 * Writeback: don't undirty key until after a cache flush
47 * Create an iterator for key pointers
49 * On btree write error, mark bucket such that it won't be freed from the cache
52 * Check for bad keys in replay
54 * Refcount journal entries in journal_replay
57 * Finish incremental gc
58 * Gc should free old UUIDs, data for invalid UUIDs
60 * Provide a way to list backing device UUIDs we have data cached for, and
61 * probably how long it's been since we've seen them, and a way to invalidate
62 * dirty data for devices that will never be attached again
64 * Keep 1 min/5 min/15 min statistics of how busy a block device has been, so
65 * that based on that and how much dirty data we have we can keep writeback
68 * Add a tracepoint or somesuch to watch for writeback starvation
70 * When btree depth > 1 and splitting an interior node, we have to make sure
71 * alloc_bucket() cannot fail. This should be true but is not completely
76 * If data write is less than hard sector size of ssd, round up offset in open
77 * bucket to the next whole sector
79 * Superblock needs to be fleshed out for multiple cache devices
81 * Add a sysfs tunable for the number of writeback IOs in flight
83 * Add a sysfs tunable for the number of open data buckets
85 * IO tracking: Can we track when one process is doing io on behalf of another?
86 * IO tracking: Don't use just an average, weigh more recent stuff higher
88 * Test module load/unload
91 #define MAX_NEED_GC 64
92 #define MAX_SAVE_PRIO 72
93 #define MAX_GC_TIMES 100
94 #define MIN_GC_NODES 100
95 #define GC_SLEEP_MS 100
97 #define PTR_DIRTY_BIT (((uint64_t) 1 << 36))
99 #define PTR_HASH(c, k) \
100 (((k)->ptr[0] >> c->bucket_bits) | PTR_GEN(k, 0))
102 static struct workqueue_struct *btree_io_wq;
104 #define insert_lock(s, b) ((b)->level <= (s)->lock)
107 static inline struct bset *write_block(struct btree *b)
109 return ((void *) btree_bset_first(b)) + b->written * block_bytes(b->c->cache);
112 static void bch_btree_init_next(struct btree *b)
114 /* If not a leaf node, always sort */
115 if (b->level && b->keys.nsets)
116 bch_btree_sort(&b->keys, &b->c->sort);
118 bch_btree_sort_lazy(&b->keys, &b->c->sort);
120 if (b->written < btree_blocks(b))
121 bch_bset_init_next(&b->keys, write_block(b),
122 bset_magic(&b->c->cache->sb));
126 /* Btree key manipulation */
128 void bkey_put(struct cache_set *c, struct bkey *k)
132 for (i = 0; i < KEY_PTRS(k); i++)
133 if (ptr_available(c, k, i))
134 atomic_dec_bug(&PTR_BUCKET(c, k, i)->pin);
139 static uint64_t btree_csum_set(struct btree *b, struct bset *i)
141 uint64_t crc = b->key.ptr[0];
142 void *data = (void *) i + 8, *end = bset_bkey_last(i);
144 crc = bch_crc64_update(crc, data, end - data);
145 return crc ^ 0xffffffffffffffffULL;
148 void bch_btree_node_read_done(struct btree *b)
150 const char *err = "bad btree header";
151 struct bset *i = btree_bset_first(b);
152 struct btree_iter *iter;
155 * c->fill_iter can allocate an iterator with more memory space
156 * than static MAX_BSETS.
157 * See the comment arount cache_set->fill_iter.
159 iter = mempool_alloc(&b->c->fill_iter, GFP_NOIO);
160 iter->size = b->c->cache->sb.bucket_size / b->c->cache->sb.block_size;
163 #ifdef CONFIG_BCACHE_DEBUG
171 b->written < btree_blocks(b) && i->seq == b->keys.set[0].data->seq;
172 i = write_block(b)) {
173 err = "unsupported bset version";
174 if (i->version > BCACHE_BSET_VERSION)
177 err = "bad btree header";
178 if (b->written + set_blocks(i, block_bytes(b->c->cache)) >
183 if (i->magic != bset_magic(&b->c->cache->sb))
186 err = "bad checksum";
187 switch (i->version) {
189 if (i->csum != csum_set(i))
192 case BCACHE_BSET_VERSION:
193 if (i->csum != btree_csum_set(b, i))
199 if (i != b->keys.set[0].data && !i->keys)
202 bch_btree_iter_push(iter, i->start, bset_bkey_last(i));
204 b->written += set_blocks(i, block_bytes(b->c->cache));
207 err = "corrupted btree";
208 for (i = write_block(b);
209 bset_sector_offset(&b->keys, i) < KEY_SIZE(&b->key);
210 i = ((void *) i) + block_bytes(b->c->cache))
211 if (i->seq == b->keys.set[0].data->seq)
214 bch_btree_sort_and_fix_extents(&b->keys, iter, &b->c->sort);
216 i = b->keys.set[0].data;
217 err = "short btree key";
218 if (b->keys.set[0].size &&
219 bkey_cmp(&b->key, &b->keys.set[0].end) < 0)
222 if (b->written < btree_blocks(b))
223 bch_bset_init_next(&b->keys, write_block(b),
224 bset_magic(&b->c->cache->sb));
226 mempool_free(iter, &b->c->fill_iter);
229 set_btree_node_io_error(b);
230 bch_cache_set_error(b->c, "%s at bucket %zu, block %u, %u keys",
231 err, PTR_BUCKET_NR(b->c, &b->key, 0),
232 bset_block_offset(b, i), i->keys);
236 static void btree_node_read_endio(struct bio *bio)
238 struct closure *cl = bio->bi_private;
243 static void bch_btree_node_read(struct btree *b)
245 uint64_t start_time = local_clock();
249 trace_bcache_btree_read(b);
251 closure_init_stack(&cl);
253 bio = bch_bbio_alloc(b->c);
254 bio->bi_iter.bi_size = KEY_SIZE(&b->key) << 9;
255 bio->bi_end_io = btree_node_read_endio;
256 bio->bi_private = &cl;
257 bio->bi_opf = REQ_OP_READ | REQ_META;
259 bch_bio_map(bio, b->keys.set[0].data);
261 bch_submit_bbio(bio, b->c, &b->key, 0);
265 set_btree_node_io_error(b);
267 bch_bbio_free(bio, b->c);
269 if (btree_node_io_error(b))
272 bch_btree_node_read_done(b);
273 bch_time_stats_update(&b->c->btree_read_time, start_time);
277 bch_cache_set_error(b->c, "io error reading bucket %zu",
278 PTR_BUCKET_NR(b->c, &b->key, 0));
281 static void btree_complete_write(struct btree *b, struct btree_write *w)
283 if (w->prio_blocked &&
284 !atomic_sub_return(w->prio_blocked, &b->c->prio_blocked))
285 wake_up_allocators(b->c);
288 atomic_dec_bug(w->journal);
289 __closure_wake_up(&b->c->journal.wait);
296 static void btree_node_write_unlock(struct closure *cl)
298 struct btree *b = container_of(cl, struct btree, io);
303 static void __btree_node_write_done(struct closure *cl)
305 struct btree *b = container_of(cl, struct btree, io);
306 struct btree_write *w = btree_prev_write(b);
308 bch_bbio_free(b->bio, b->c);
310 btree_complete_write(b, w);
312 if (btree_node_dirty(b))
313 queue_delayed_work(btree_io_wq, &b->work, 30 * HZ);
315 closure_return_with_destructor(cl, btree_node_write_unlock);
318 static void btree_node_write_done(struct closure *cl)
320 struct btree *b = container_of(cl, struct btree, io);
322 bio_free_pages(b->bio);
323 __btree_node_write_done(cl);
326 static void btree_node_write_endio(struct bio *bio)
328 struct closure *cl = bio->bi_private;
329 struct btree *b = container_of(cl, struct btree, io);
332 set_btree_node_io_error(b);
334 bch_bbio_count_io_errors(b->c, bio, bio->bi_status, "writing btree");
338 static void do_btree_node_write(struct btree *b)
340 struct closure *cl = &b->io;
341 struct bset *i = btree_bset_last(b);
344 i->version = BCACHE_BSET_VERSION;
345 i->csum = btree_csum_set(b, i);
348 b->bio = bch_bbio_alloc(b->c);
350 b->bio->bi_end_io = btree_node_write_endio;
351 b->bio->bi_private = cl;
352 b->bio->bi_iter.bi_size = roundup(set_bytes(i), block_bytes(b->c->cache));
353 b->bio->bi_opf = REQ_OP_WRITE | REQ_META | REQ_FUA;
354 bch_bio_map(b->bio, i);
357 * If we're appending to a leaf node, we don't technically need FUA -
358 * this write just needs to be persisted before the next journal write,
359 * which will be marked FLUSH|FUA.
361 * Similarly if we're writing a new btree root - the pointer is going to
362 * be in the next journal entry.
364 * But if we're writing a new btree node (that isn't a root) or
365 * appending to a non leaf btree node, we need either FUA or a flush
366 * when we write the parent with the new pointer. FUA is cheaper than a
367 * flush, and writes appending to leaf nodes aren't blocking anything so
368 * just make all btree node writes FUA to keep things sane.
371 bkey_copy(&k.key, &b->key);
372 SET_PTR_OFFSET(&k.key, 0, PTR_OFFSET(&k.key, 0) +
373 bset_sector_offset(&b->keys, i));
375 if (!bch_bio_alloc_pages(b->bio, __GFP_NOWARN|GFP_NOWAIT)) {
377 void *addr = (void *) ((unsigned long) i & ~(PAGE_SIZE - 1));
378 struct bvec_iter_all iter_all;
380 bio_for_each_segment_all(bv, b->bio, iter_all) {
381 memcpy(page_address(bv->bv_page), addr, PAGE_SIZE);
385 bch_submit_bbio(b->bio, b->c, &k.key, 0);
387 continue_at(cl, btree_node_write_done, NULL);
390 * No problem for multipage bvec since the bio is
394 bch_bio_map(b->bio, i);
396 bch_submit_bbio(b->bio, b->c, &k.key, 0);
399 continue_at_nobarrier(cl, __btree_node_write_done, NULL);
403 void __bch_btree_node_write(struct btree *b, struct closure *parent)
405 struct bset *i = btree_bset_last(b);
407 lockdep_assert_held(&b->write_lock);
409 trace_bcache_btree_write(b);
411 BUG_ON(current->bio_list);
412 BUG_ON(b->written >= btree_blocks(b));
413 BUG_ON(b->written && !i->keys);
414 BUG_ON(btree_bset_first(b)->seq != i->seq);
415 bch_check_keys(&b->keys, "writing");
417 cancel_delayed_work(&b->work);
419 /* If caller isn't waiting for write, parent refcount is cache set */
421 closure_init(&b->io, parent ?: &b->c->cl);
423 clear_bit(BTREE_NODE_dirty, &b->flags);
424 change_bit(BTREE_NODE_write_idx, &b->flags);
426 do_btree_node_write(b);
428 atomic_long_add(set_blocks(i, block_bytes(b->c->cache)) * b->c->cache->sb.block_size,
429 &PTR_CACHE(b->c, &b->key, 0)->btree_sectors_written);
431 b->written += set_blocks(i, block_bytes(b->c->cache));
434 void bch_btree_node_write(struct btree *b, struct closure *parent)
436 unsigned int nsets = b->keys.nsets;
438 lockdep_assert_held(&b->lock);
440 __bch_btree_node_write(b, parent);
443 * do verify if there was more than one set initially (i.e. we did a
444 * sort) and we sorted down to a single set:
446 if (nsets && !b->keys.nsets)
449 bch_btree_init_next(b);
452 static void bch_btree_node_write_sync(struct btree *b)
456 closure_init_stack(&cl);
458 mutex_lock(&b->write_lock);
459 bch_btree_node_write(b, &cl);
460 mutex_unlock(&b->write_lock);
465 static void btree_node_write_work(struct work_struct *w)
467 struct btree *b = container_of(to_delayed_work(w), struct btree, work);
469 mutex_lock(&b->write_lock);
470 if (btree_node_dirty(b))
471 __bch_btree_node_write(b, NULL);
472 mutex_unlock(&b->write_lock);
475 static void bch_btree_leaf_dirty(struct btree *b, atomic_t *journal_ref)
477 struct bset *i = btree_bset_last(b);
478 struct btree_write *w = btree_current_write(b);
480 lockdep_assert_held(&b->write_lock);
485 if (!btree_node_dirty(b))
486 queue_delayed_work(btree_io_wq, &b->work, 30 * HZ);
488 set_btree_node_dirty(b);
491 * w->journal is always the oldest journal pin of all bkeys
492 * in the leaf node, to make sure the oldest jset seq won't
493 * be increased before this btree node is flushed.
497 journal_pin_cmp(b->c, w->journal, journal_ref)) {
498 atomic_dec_bug(w->journal);
503 w->journal = journal_ref;
504 atomic_inc(w->journal);
508 /* Force write if set is too big */
509 if (set_bytes(i) > PAGE_SIZE - 48 &&
511 bch_btree_node_write(b, NULL);
515 * Btree in memory cache - allocation/freeing
516 * mca -> memory cache
519 #define mca_reserve(c) (((!IS_ERR_OR_NULL(c->root) && c->root->level) \
520 ? c->root->level : 1) * 8 + 16)
521 #define mca_can_free(c) \
522 max_t(int, 0, c->btree_cache_used - mca_reserve(c))
524 static void mca_data_free(struct btree *b)
526 BUG_ON(b->io_mutex.count != 1);
528 bch_btree_keys_free(&b->keys);
530 b->c->btree_cache_used--;
531 list_move(&b->list, &b->c->btree_cache_freed);
534 static void mca_bucket_free(struct btree *b)
536 BUG_ON(btree_node_dirty(b));
539 hlist_del_init_rcu(&b->hash);
540 list_move(&b->list, &b->c->btree_cache_freeable);
543 static unsigned int btree_order(struct bkey *k)
545 return ilog2(KEY_SIZE(k) / PAGE_SECTORS ?: 1);
548 static void mca_data_alloc(struct btree *b, struct bkey *k, gfp_t gfp)
550 if (!bch_btree_keys_alloc(&b->keys,
552 ilog2(b->c->btree_pages),
555 b->c->btree_cache_used++;
556 list_move(&b->list, &b->c->btree_cache);
558 list_move(&b->list, &b->c->btree_cache_freed);
562 static struct btree *mca_bucket_alloc(struct cache_set *c,
563 struct bkey *k, gfp_t gfp)
566 * kzalloc() is necessary here for initialization,
567 * see code comments in bch_btree_keys_init().
569 struct btree *b = kzalloc(sizeof(struct btree), gfp);
574 init_rwsem(&b->lock);
575 lockdep_set_novalidate_class(&b->lock);
576 mutex_init(&b->write_lock);
577 lockdep_set_novalidate_class(&b->write_lock);
578 INIT_LIST_HEAD(&b->list);
579 INIT_DELAYED_WORK(&b->work, btree_node_write_work);
581 sema_init(&b->io_mutex, 1);
583 mca_data_alloc(b, k, gfp);
587 static int mca_reap(struct btree *b, unsigned int min_order, bool flush)
591 closure_init_stack(&cl);
592 lockdep_assert_held(&b->c->bucket_lock);
594 if (!down_write_trylock(&b->lock))
597 BUG_ON(btree_node_dirty(b) && !b->keys.set[0].data);
599 if (b->keys.page_order < min_order)
603 if (btree_node_dirty(b))
606 if (down_trylock(&b->io_mutex))
613 * BTREE_NODE_dirty might be cleared in btree_flush_btree() by
614 * __bch_btree_node_write(). To avoid an extra flush, acquire
615 * b->write_lock before checking BTREE_NODE_dirty bit.
617 mutex_lock(&b->write_lock);
619 * If this btree node is selected in btree_flush_write() by journal
620 * code, delay and retry until the node is flushed by journal code
621 * and BTREE_NODE_journal_flush bit cleared by btree_flush_write().
623 if (btree_node_journal_flush(b)) {
624 pr_debug("bnode %p is flushing by journal, retry\n", b);
625 mutex_unlock(&b->write_lock);
630 if (btree_node_dirty(b))
631 __bch_btree_node_write(b, &cl);
632 mutex_unlock(&b->write_lock);
636 /* wait for any in flight btree write */
646 static unsigned long bch_mca_scan(struct shrinker *shrink,
647 struct shrink_control *sc)
649 struct cache_set *c = container_of(shrink, struct cache_set, shrink);
651 unsigned long i, nr = sc->nr_to_scan;
652 unsigned long freed = 0;
653 unsigned int btree_cache_used;
655 if (c->shrinker_disabled)
658 if (c->btree_cache_alloc_lock)
661 /* Return -1 if we can't do anything right now */
662 if (sc->gfp_mask & __GFP_IO)
663 mutex_lock(&c->bucket_lock);
664 else if (!mutex_trylock(&c->bucket_lock))
668 * It's _really_ critical that we don't free too many btree nodes - we
669 * have to always leave ourselves a reserve. The reserve is how we
670 * guarantee that allocating memory for a new btree node can always
671 * succeed, so that inserting keys into the btree can always succeed and
672 * IO can always make forward progress:
674 nr /= c->btree_pages;
677 nr = min_t(unsigned long, nr, mca_can_free(c));
680 btree_cache_used = c->btree_cache_used;
681 list_for_each_entry_safe_reverse(b, t, &c->btree_cache_freeable, list) {
685 if (!mca_reap(b, 0, false)) {
694 list_for_each_entry_safe_reverse(b, t, &c->btree_cache, list) {
695 if (nr <= 0 || i >= btree_cache_used)
698 if (!mca_reap(b, 0, false)) {
709 mutex_unlock(&c->bucket_lock);
710 return freed * c->btree_pages;
713 static unsigned long bch_mca_count(struct shrinker *shrink,
714 struct shrink_control *sc)
716 struct cache_set *c = container_of(shrink, struct cache_set, shrink);
718 if (c->shrinker_disabled)
721 if (c->btree_cache_alloc_lock)
724 return mca_can_free(c) * c->btree_pages;
727 void bch_btree_cache_free(struct cache_set *c)
732 closure_init_stack(&cl);
734 if (c->shrink.list.next)
735 unregister_shrinker(&c->shrink);
737 mutex_lock(&c->bucket_lock);
739 #ifdef CONFIG_BCACHE_DEBUG
741 list_move(&c->verify_data->list, &c->btree_cache);
743 free_pages((unsigned long) c->verify_ondisk, ilog2(meta_bucket_pages(&c->cache->sb)));
746 list_splice(&c->btree_cache_freeable,
749 while (!list_empty(&c->btree_cache)) {
750 b = list_first_entry(&c->btree_cache, struct btree, list);
753 * This function is called by cache_set_free(), no I/O
754 * request on cache now, it is unnecessary to acquire
755 * b->write_lock before clearing BTREE_NODE_dirty anymore.
757 if (btree_node_dirty(b)) {
758 btree_complete_write(b, btree_current_write(b));
759 clear_bit(BTREE_NODE_dirty, &b->flags);
764 while (!list_empty(&c->btree_cache_freed)) {
765 b = list_first_entry(&c->btree_cache_freed,
768 cancel_delayed_work_sync(&b->work);
772 mutex_unlock(&c->bucket_lock);
775 int bch_btree_cache_alloc(struct cache_set *c)
779 for (i = 0; i < mca_reserve(c); i++)
780 if (!mca_bucket_alloc(c, &ZERO_KEY, GFP_KERNEL))
783 list_splice_init(&c->btree_cache,
784 &c->btree_cache_freeable);
786 #ifdef CONFIG_BCACHE_DEBUG
787 mutex_init(&c->verify_lock);
789 c->verify_ondisk = (void *)
790 __get_free_pages(GFP_KERNEL|__GFP_COMP,
791 ilog2(meta_bucket_pages(&c->cache->sb)));
792 if (!c->verify_ondisk) {
794 * Don't worry about the mca_rereserve buckets
795 * allocated in previous for-loop, they will be
796 * handled properly in bch_cache_set_unregister().
801 c->verify_data = mca_bucket_alloc(c, &ZERO_KEY, GFP_KERNEL);
803 if (c->verify_data &&
804 c->verify_data->keys.set->data)
805 list_del_init(&c->verify_data->list);
807 c->verify_data = NULL;
810 c->shrink.count_objects = bch_mca_count;
811 c->shrink.scan_objects = bch_mca_scan;
813 c->shrink.batch = c->btree_pages * 2;
815 if (register_shrinker(&c->shrink))
816 pr_warn("bcache: %s: could not register shrinker\n",
822 /* Btree in memory cache - hash table */
824 static struct hlist_head *mca_hash(struct cache_set *c, struct bkey *k)
826 return &c->bucket_hash[hash_32(PTR_HASH(c, k), BUCKET_HASH_BITS)];
829 static struct btree *mca_find(struct cache_set *c, struct bkey *k)
834 hlist_for_each_entry_rcu(b, mca_hash(c, k), hash)
835 if (PTR_HASH(c, &b->key) == PTR_HASH(c, k))
843 static int mca_cannibalize_lock(struct cache_set *c, struct btree_op *op)
845 spin_lock(&c->btree_cannibalize_lock);
846 if (likely(c->btree_cache_alloc_lock == NULL)) {
847 c->btree_cache_alloc_lock = current;
848 } else if (c->btree_cache_alloc_lock != current) {
850 prepare_to_wait(&c->btree_cache_wait, &op->wait,
851 TASK_UNINTERRUPTIBLE);
852 spin_unlock(&c->btree_cannibalize_lock);
855 spin_unlock(&c->btree_cannibalize_lock);
860 static struct btree *mca_cannibalize(struct cache_set *c, struct btree_op *op,
865 trace_bcache_btree_cache_cannibalize(c);
867 if (mca_cannibalize_lock(c, op))
868 return ERR_PTR(-EINTR);
870 list_for_each_entry_reverse(b, &c->btree_cache, list)
871 if (!mca_reap(b, btree_order(k), false))
874 list_for_each_entry_reverse(b, &c->btree_cache, list)
875 if (!mca_reap(b, btree_order(k), true))
878 WARN(1, "btree cache cannibalize failed\n");
879 return ERR_PTR(-ENOMEM);
883 * We can only have one thread cannibalizing other cached btree nodes at a time,
884 * or we'll deadlock. We use an open coded mutex to ensure that, which a
885 * cannibalize_bucket() will take. This means every time we unlock the root of
886 * the btree, we need to release this lock if we have it held.
888 void bch_cannibalize_unlock(struct cache_set *c)
890 spin_lock(&c->btree_cannibalize_lock);
891 if (c->btree_cache_alloc_lock == current) {
892 c->btree_cache_alloc_lock = NULL;
893 wake_up(&c->btree_cache_wait);
895 spin_unlock(&c->btree_cannibalize_lock);
898 static struct btree *mca_alloc(struct cache_set *c, struct btree_op *op,
899 struct bkey *k, int level)
903 BUG_ON(current->bio_list);
905 lockdep_assert_held(&c->bucket_lock);
910 /* btree_free() doesn't free memory; it sticks the node on the end of
911 * the list. Check if there's any freed nodes there:
913 list_for_each_entry(b, &c->btree_cache_freeable, list)
914 if (!mca_reap(b, btree_order(k), false))
917 /* We never free struct btree itself, just the memory that holds the on
918 * disk node. Check the freed list before allocating a new one:
920 list_for_each_entry(b, &c->btree_cache_freed, list)
921 if (!mca_reap(b, 0, false)) {
922 mca_data_alloc(b, k, __GFP_NOWARN|GFP_NOIO);
923 if (!b->keys.set[0].data)
929 b = mca_bucket_alloc(c, k, __GFP_NOWARN|GFP_NOIO);
933 BUG_ON(!down_write_trylock(&b->lock));
934 if (!b->keys.set->data)
937 BUG_ON(b->io_mutex.count != 1);
939 bkey_copy(&b->key, k);
940 list_move(&b->list, &c->btree_cache);
941 hlist_del_init_rcu(&b->hash);
942 hlist_add_head_rcu(&b->hash, mca_hash(c, k));
944 lock_set_subclass(&b->lock.dep_map, level + 1, _THIS_IP_);
945 b->parent = (void *) ~0UL;
951 bch_btree_keys_init(&b->keys, &bch_extent_keys_ops,
952 &b->c->expensive_debug_checks);
954 bch_btree_keys_init(&b->keys, &bch_btree_keys_ops,
955 &b->c->expensive_debug_checks);
962 b = mca_cannibalize(c, op, k);
970 * bch_btree_node_get - find a btree node in the cache and lock it, reading it
971 * in from disk if necessary.
973 * If IO is necessary and running under submit_bio_noacct, returns -EAGAIN.
975 * The btree node will have either a read or a write lock held, depending on
976 * level and op->lock.
978 * Note: Only error code or btree pointer will be returned, it is unncessary
979 * for callers to check NULL pointer.
981 struct btree *bch_btree_node_get(struct cache_set *c, struct btree_op *op,
982 struct bkey *k, int level, bool write,
983 struct btree *parent)
993 if (current->bio_list)
994 return ERR_PTR(-EAGAIN);
996 mutex_lock(&c->bucket_lock);
997 b = mca_alloc(c, op, k, level);
998 mutex_unlock(&c->bucket_lock);
1005 bch_btree_node_read(b);
1008 downgrade_write(&b->lock);
1010 rw_lock(write, b, level);
1011 if (PTR_HASH(c, &b->key) != PTR_HASH(c, k)) {
1012 rw_unlock(write, b);
1015 BUG_ON(b->level != level);
1018 if (btree_node_io_error(b)) {
1019 rw_unlock(write, b);
1020 return ERR_PTR(-EIO);
1023 BUG_ON(!b->written);
1027 for (; i <= b->keys.nsets && b->keys.set[i].size; i++) {
1028 prefetch(b->keys.set[i].tree);
1029 prefetch(b->keys.set[i].data);
1032 for (; i <= b->keys.nsets; i++)
1033 prefetch(b->keys.set[i].data);
1038 static void btree_node_prefetch(struct btree *parent, struct bkey *k)
1042 mutex_lock(&parent->c->bucket_lock);
1043 b = mca_alloc(parent->c, NULL, k, parent->level - 1);
1044 mutex_unlock(&parent->c->bucket_lock);
1046 if (!IS_ERR_OR_NULL(b)) {
1048 bch_btree_node_read(b);
1055 static void btree_node_free(struct btree *b)
1057 trace_bcache_btree_node_free(b);
1059 BUG_ON(b == b->c->root);
1062 mutex_lock(&b->write_lock);
1064 * If the btree node is selected and flushing in btree_flush_write(),
1065 * delay and retry until the BTREE_NODE_journal_flush bit cleared,
1066 * then it is safe to free the btree node here. Otherwise this btree
1067 * node will be in race condition.
1069 if (btree_node_journal_flush(b)) {
1070 mutex_unlock(&b->write_lock);
1071 pr_debug("bnode %p journal_flush set, retry\n", b);
1076 if (btree_node_dirty(b)) {
1077 btree_complete_write(b, btree_current_write(b));
1078 clear_bit(BTREE_NODE_dirty, &b->flags);
1081 mutex_unlock(&b->write_lock);
1083 cancel_delayed_work(&b->work);
1085 mutex_lock(&b->c->bucket_lock);
1086 bch_bucket_free(b->c, &b->key);
1088 mutex_unlock(&b->c->bucket_lock);
1092 * Only error code or btree pointer will be returned, it is unncessary for
1093 * callers to check NULL pointer.
1095 struct btree *__bch_btree_node_alloc(struct cache_set *c, struct btree_op *op,
1096 int level, bool wait,
1097 struct btree *parent)
1102 mutex_lock(&c->bucket_lock);
1104 /* return ERR_PTR(-EAGAIN) when it fails */
1105 b = ERR_PTR(-EAGAIN);
1106 if (__bch_bucket_alloc_set(c, RESERVE_BTREE, &k.key, wait))
1109 bkey_put(c, &k.key);
1110 SET_KEY_SIZE(&k.key, c->btree_pages * PAGE_SECTORS);
1112 b = mca_alloc(c, op, &k.key, level);
1118 "Tried to allocate bucket that was in btree cache");
1123 bch_bset_init_next(&b->keys, b->keys.set->data, bset_magic(&b->c->cache->sb));
1125 mutex_unlock(&c->bucket_lock);
1127 trace_bcache_btree_node_alloc(b);
1130 bch_bucket_free(c, &k.key);
1132 mutex_unlock(&c->bucket_lock);
1134 trace_bcache_btree_node_alloc_fail(c);
1138 static struct btree *bch_btree_node_alloc(struct cache_set *c,
1139 struct btree_op *op, int level,
1140 struct btree *parent)
1142 return __bch_btree_node_alloc(c, op, level, op != NULL, parent);
1145 static struct btree *btree_node_alloc_replacement(struct btree *b,
1146 struct btree_op *op)
1148 struct btree *n = bch_btree_node_alloc(b->c, op, b->level, b->parent);
1151 mutex_lock(&n->write_lock);
1152 bch_btree_sort_into(&b->keys, &n->keys, &b->c->sort);
1153 bkey_copy_key(&n->key, &b->key);
1154 mutex_unlock(&n->write_lock);
1160 static void make_btree_freeing_key(struct btree *b, struct bkey *k)
1164 mutex_lock(&b->c->bucket_lock);
1166 atomic_inc(&b->c->prio_blocked);
1168 bkey_copy(k, &b->key);
1169 bkey_copy_key(k, &ZERO_KEY);
1171 for (i = 0; i < KEY_PTRS(k); i++)
1173 bch_inc_gen(PTR_CACHE(b->c, &b->key, i),
1174 PTR_BUCKET(b->c, &b->key, i)));
1176 mutex_unlock(&b->c->bucket_lock);
1179 static int btree_check_reserve(struct btree *b, struct btree_op *op)
1181 struct cache_set *c = b->c;
1182 struct cache *ca = c->cache;
1183 unsigned int reserve = (c->root->level - b->level) * 2 + 1;
1185 mutex_lock(&c->bucket_lock);
1187 if (fifo_used(&ca->free[RESERVE_BTREE]) < reserve) {
1189 prepare_to_wait(&c->btree_cache_wait, &op->wait,
1190 TASK_UNINTERRUPTIBLE);
1191 mutex_unlock(&c->bucket_lock);
1195 mutex_unlock(&c->bucket_lock);
1197 return mca_cannibalize_lock(b->c, op);
1200 /* Garbage collection */
1202 static uint8_t __bch_btree_mark_key(struct cache_set *c, int level,
1210 * ptr_invalid() can't return true for the keys that mark btree nodes as
1211 * freed, but since ptr_bad() returns true we'll never actually use them
1212 * for anything and thus we don't want mark their pointers here
1214 if (!bkey_cmp(k, &ZERO_KEY))
1217 for (i = 0; i < KEY_PTRS(k); i++) {
1218 if (!ptr_available(c, k, i))
1221 g = PTR_BUCKET(c, k, i);
1223 if (gen_after(g->last_gc, PTR_GEN(k, i)))
1224 g->last_gc = PTR_GEN(k, i);
1226 if (ptr_stale(c, k, i)) {
1227 stale = max(stale, ptr_stale(c, k, i));
1231 cache_bug_on(GC_MARK(g) &&
1232 (GC_MARK(g) == GC_MARK_METADATA) != (level != 0),
1233 c, "inconsistent ptrs: mark = %llu, level = %i",
1237 SET_GC_MARK(g, GC_MARK_METADATA);
1238 else if (KEY_DIRTY(k))
1239 SET_GC_MARK(g, GC_MARK_DIRTY);
1240 else if (!GC_MARK(g))
1241 SET_GC_MARK(g, GC_MARK_RECLAIMABLE);
1243 /* guard against overflow */
1244 SET_GC_SECTORS_USED(g, min_t(unsigned int,
1245 GC_SECTORS_USED(g) + KEY_SIZE(k),
1246 MAX_GC_SECTORS_USED));
1248 BUG_ON(!GC_SECTORS_USED(g));
1254 #define btree_mark_key(b, k) __bch_btree_mark_key(b->c, b->level, k)
1256 void bch_initial_mark_key(struct cache_set *c, int level, struct bkey *k)
1260 for (i = 0; i < KEY_PTRS(k); i++)
1261 if (ptr_available(c, k, i) &&
1262 !ptr_stale(c, k, i)) {
1263 struct bucket *b = PTR_BUCKET(c, k, i);
1265 b->gen = PTR_GEN(k, i);
1267 if (level && bkey_cmp(k, &ZERO_KEY))
1268 b->prio = BTREE_PRIO;
1269 else if (!level && b->prio == BTREE_PRIO)
1270 b->prio = INITIAL_PRIO;
1273 __bch_btree_mark_key(c, level, k);
1276 void bch_update_bucket_in_use(struct cache_set *c, struct gc_stat *stats)
1278 stats->in_use = (c->nbuckets - c->avail_nbuckets) * 100 / c->nbuckets;
1281 static bool btree_gc_mark_node(struct btree *b, struct gc_stat *gc)
1284 unsigned int keys = 0, good_keys = 0;
1286 struct btree_iter iter;
1287 struct bset_tree *t;
1291 for_each_key_filter(&b->keys, k, &iter, bch_ptr_invalid) {
1292 stale = max(stale, btree_mark_key(b, k));
1295 if (bch_ptr_bad(&b->keys, k))
1298 gc->key_bytes += bkey_u64s(k);
1302 gc->data += KEY_SIZE(k);
1305 for (t = b->keys.set; t <= &b->keys.set[b->keys.nsets]; t++)
1306 btree_bug_on(t->size &&
1307 bset_written(&b->keys, t) &&
1308 bkey_cmp(&b->key, &t->end) < 0,
1309 b, "found short btree key in gc");
1311 if (b->c->gc_always_rewrite)
1317 if ((keys - good_keys) * 2 > keys)
1323 #define GC_MERGE_NODES 4U
1325 struct gc_merge_info {
1330 static int bch_btree_insert_node(struct btree *b, struct btree_op *op,
1331 struct keylist *insert_keys,
1332 atomic_t *journal_ref,
1333 struct bkey *replace_key);
1335 static int btree_gc_coalesce(struct btree *b, struct btree_op *op,
1336 struct gc_stat *gc, struct gc_merge_info *r)
1338 unsigned int i, nodes = 0, keys = 0, blocks;
1339 struct btree *new_nodes[GC_MERGE_NODES];
1340 struct keylist keylist;
1344 bch_keylist_init(&keylist);
1346 if (btree_check_reserve(b, NULL))
1349 memset(new_nodes, 0, sizeof(new_nodes));
1350 closure_init_stack(&cl);
1352 while (nodes < GC_MERGE_NODES && !IS_ERR_OR_NULL(r[nodes].b))
1353 keys += r[nodes++].keys;
1355 blocks = btree_default_blocks(b->c) * 2 / 3;
1358 __set_blocks(b->keys.set[0].data, keys,
1359 block_bytes(b->c->cache)) > blocks * (nodes - 1))
1362 for (i = 0; i < nodes; i++) {
1363 new_nodes[i] = btree_node_alloc_replacement(r[i].b, NULL);
1364 if (IS_ERR(new_nodes[i]))
1365 goto out_nocoalesce;
1369 * We have to check the reserve here, after we've allocated our new
1370 * nodes, to make sure the insert below will succeed - we also check
1371 * before as an optimization to potentially avoid a bunch of expensive
1374 if (btree_check_reserve(b, NULL))
1375 goto out_nocoalesce;
1377 for (i = 0; i < nodes; i++)
1378 mutex_lock(&new_nodes[i]->write_lock);
1380 for (i = nodes - 1; i > 0; --i) {
1381 struct bset *n1 = btree_bset_first(new_nodes[i]);
1382 struct bset *n2 = btree_bset_first(new_nodes[i - 1]);
1383 struct bkey *k, *last = NULL;
1389 k < bset_bkey_last(n2);
1391 if (__set_blocks(n1, n1->keys + keys +
1393 block_bytes(b->c->cache)) > blocks)
1397 keys += bkey_u64s(k);
1401 * Last node we're not getting rid of - we're getting
1402 * rid of the node at r[0]. Have to try and fit all of
1403 * the remaining keys into this node; we can't ensure
1404 * they will always fit due to rounding and variable
1405 * length keys (shouldn't be possible in practice,
1408 if (__set_blocks(n1, n1->keys + n2->keys,
1409 block_bytes(b->c->cache)) >
1410 btree_blocks(new_nodes[i]))
1411 goto out_unlock_nocoalesce;
1414 /* Take the key of the node we're getting rid of */
1418 BUG_ON(__set_blocks(n1, n1->keys + keys, block_bytes(b->c->cache)) >
1419 btree_blocks(new_nodes[i]));
1422 bkey_copy_key(&new_nodes[i]->key, last);
1424 memcpy(bset_bkey_last(n1),
1426 (void *) bset_bkey_idx(n2, keys) - (void *) n2->start);
1429 r[i].keys = n1->keys;
1432 bset_bkey_idx(n2, keys),
1433 (void *) bset_bkey_last(n2) -
1434 (void *) bset_bkey_idx(n2, keys));
1438 if (__bch_keylist_realloc(&keylist,
1439 bkey_u64s(&new_nodes[i]->key)))
1440 goto out_unlock_nocoalesce;
1442 bch_btree_node_write(new_nodes[i], &cl);
1443 bch_keylist_add(&keylist, &new_nodes[i]->key);
1446 for (i = 0; i < nodes; i++)
1447 mutex_unlock(&new_nodes[i]->write_lock);
1451 /* We emptied out this node */
1452 BUG_ON(btree_bset_first(new_nodes[0])->keys);
1453 btree_node_free(new_nodes[0]);
1454 rw_unlock(true, new_nodes[0]);
1455 new_nodes[0] = NULL;
1457 for (i = 0; i < nodes; i++) {
1458 if (__bch_keylist_realloc(&keylist, bkey_u64s(&r[i].b->key)))
1459 goto out_nocoalesce;
1461 make_btree_freeing_key(r[i].b, keylist.top);
1462 bch_keylist_push(&keylist);
1465 bch_btree_insert_node(b, op, &keylist, NULL, NULL);
1466 BUG_ON(!bch_keylist_empty(&keylist));
1468 for (i = 0; i < nodes; i++) {
1469 btree_node_free(r[i].b);
1470 rw_unlock(true, r[i].b);
1472 r[i].b = new_nodes[i];
1475 memmove(r, r + 1, sizeof(r[0]) * (nodes - 1));
1476 r[nodes - 1].b = ERR_PTR(-EINTR);
1478 trace_bcache_btree_gc_coalesce(nodes);
1481 bch_keylist_free(&keylist);
1483 /* Invalidated our iterator */
1486 out_unlock_nocoalesce:
1487 for (i = 0; i < nodes; i++)
1488 mutex_unlock(&new_nodes[i]->write_lock);
1493 while ((k = bch_keylist_pop(&keylist)))
1494 if (!bkey_cmp(k, &ZERO_KEY))
1495 atomic_dec(&b->c->prio_blocked);
1496 bch_keylist_free(&keylist);
1498 for (i = 0; i < nodes; i++)
1499 if (!IS_ERR_OR_NULL(new_nodes[i])) {
1500 btree_node_free(new_nodes[i]);
1501 rw_unlock(true, new_nodes[i]);
1506 static int btree_gc_rewrite_node(struct btree *b, struct btree_op *op,
1507 struct btree *replace)
1509 struct keylist keys;
1512 if (btree_check_reserve(b, NULL))
1515 n = btree_node_alloc_replacement(replace, NULL);
1519 /* recheck reserve after allocating replacement node */
1520 if (btree_check_reserve(b, NULL)) {
1526 bch_btree_node_write_sync(n);
1528 bch_keylist_init(&keys);
1529 bch_keylist_add(&keys, &n->key);
1531 make_btree_freeing_key(replace, keys.top);
1532 bch_keylist_push(&keys);
1534 bch_btree_insert_node(b, op, &keys, NULL, NULL);
1535 BUG_ON(!bch_keylist_empty(&keys));
1537 btree_node_free(replace);
1540 /* Invalidated our iterator */
1544 static unsigned int btree_gc_count_keys(struct btree *b)
1547 struct btree_iter iter;
1548 unsigned int ret = 0;
1550 for_each_key_filter(&b->keys, k, &iter, bch_ptr_bad)
1551 ret += bkey_u64s(k);
1556 static size_t btree_gc_min_nodes(struct cache_set *c)
1561 * Since incremental GC would stop 100ms when front
1562 * side I/O comes, so when there are many btree nodes,
1563 * if GC only processes constant (100) nodes each time,
1564 * GC would last a long time, and the front side I/Os
1565 * would run out of the buckets (since no new bucket
1566 * can be allocated during GC), and be blocked again.
1567 * So GC should not process constant nodes, but varied
1568 * nodes according to the number of btree nodes, which
1569 * realized by dividing GC into constant(100) times,
1570 * so when there are many btree nodes, GC can process
1571 * more nodes each time, otherwise, GC will process less
1572 * nodes each time (but no less than MIN_GC_NODES)
1574 min_nodes = c->gc_stats.nodes / MAX_GC_TIMES;
1575 if (min_nodes < MIN_GC_NODES)
1576 min_nodes = MIN_GC_NODES;
1582 static int btree_gc_recurse(struct btree *b, struct btree_op *op,
1583 struct closure *writes, struct gc_stat *gc)
1586 bool should_rewrite;
1588 struct btree_iter iter;
1589 struct gc_merge_info r[GC_MERGE_NODES];
1590 struct gc_merge_info *i, *last = r + ARRAY_SIZE(r) - 1;
1592 bch_btree_iter_init(&b->keys, &iter, &b->c->gc_done);
1594 for (i = r; i < r + ARRAY_SIZE(r); i++)
1595 i->b = ERR_PTR(-EINTR);
1598 k = bch_btree_iter_next_filter(&iter, &b->keys, bch_ptr_bad);
1600 r->b = bch_btree_node_get(b->c, op, k, b->level - 1,
1603 ret = PTR_ERR(r->b);
1607 r->keys = btree_gc_count_keys(r->b);
1609 ret = btree_gc_coalesce(b, op, gc, r);
1617 if (!IS_ERR(last->b)) {
1618 should_rewrite = btree_gc_mark_node(last->b, gc);
1619 if (should_rewrite) {
1620 ret = btree_gc_rewrite_node(b, op, last->b);
1625 if (last->b->level) {
1626 ret = btree_gc_recurse(last->b, op, writes, gc);
1631 bkey_copy_key(&b->c->gc_done, &last->b->key);
1634 * Must flush leaf nodes before gc ends, since replace
1635 * operations aren't journalled
1637 mutex_lock(&last->b->write_lock);
1638 if (btree_node_dirty(last->b))
1639 bch_btree_node_write(last->b, writes);
1640 mutex_unlock(&last->b->write_lock);
1641 rw_unlock(true, last->b);
1644 memmove(r + 1, r, sizeof(r[0]) * (GC_MERGE_NODES - 1));
1647 if (atomic_read(&b->c->search_inflight) &&
1648 gc->nodes >= gc->nodes_pre + btree_gc_min_nodes(b->c)) {
1649 gc->nodes_pre = gc->nodes;
1654 if (need_resched()) {
1660 for (i = r; i < r + ARRAY_SIZE(r); i++)
1661 if (!IS_ERR_OR_NULL(i->b)) {
1662 mutex_lock(&i->b->write_lock);
1663 if (btree_node_dirty(i->b))
1664 bch_btree_node_write(i->b, writes);
1665 mutex_unlock(&i->b->write_lock);
1666 rw_unlock(true, i->b);
1672 static int bch_btree_gc_root(struct btree *b, struct btree_op *op,
1673 struct closure *writes, struct gc_stat *gc)
1675 struct btree *n = NULL;
1677 bool should_rewrite;
1679 should_rewrite = btree_gc_mark_node(b, gc);
1680 if (should_rewrite) {
1681 n = btree_node_alloc_replacement(b, NULL);
1684 bch_btree_node_write_sync(n);
1686 bch_btree_set_root(n);
1694 __bch_btree_mark_key(b->c, b->level + 1, &b->key);
1697 ret = btree_gc_recurse(b, op, writes, gc);
1702 bkey_copy_key(&b->c->gc_done, &b->key);
1707 static void btree_gc_start(struct cache_set *c)
1712 if (!c->gc_mark_valid)
1715 mutex_lock(&c->bucket_lock);
1717 c->gc_mark_valid = 0;
1718 c->gc_done = ZERO_KEY;
1721 for_each_bucket(b, ca) {
1722 b->last_gc = b->gen;
1723 if (!atomic_read(&b->pin)) {
1725 SET_GC_SECTORS_USED(b, 0);
1729 mutex_unlock(&c->bucket_lock);
1732 static void bch_btree_gc_finish(struct cache_set *c)
1739 mutex_lock(&c->bucket_lock);
1742 c->gc_mark_valid = 1;
1745 for (i = 0; i < KEY_PTRS(&c->uuid_bucket); i++)
1746 SET_GC_MARK(PTR_BUCKET(c, &c->uuid_bucket, i),
1749 /* don't reclaim buckets to which writeback keys point */
1751 for (i = 0; i < c->devices_max_used; i++) {
1752 struct bcache_device *d = c->devices[i];
1753 struct cached_dev *dc;
1754 struct keybuf_key *w, *n;
1756 if (!d || UUID_FLASH_ONLY(&c->uuids[i]))
1758 dc = container_of(d, struct cached_dev, disk);
1760 spin_lock(&dc->writeback_keys.lock);
1761 rbtree_postorder_for_each_entry_safe(w, n,
1762 &dc->writeback_keys.keys, node)
1763 for (j = 0; j < KEY_PTRS(&w->key); j++)
1764 SET_GC_MARK(PTR_BUCKET(c, &w->key, j),
1766 spin_unlock(&dc->writeback_keys.lock);
1770 c->avail_nbuckets = 0;
1773 ca->invalidate_needs_gc = 0;
1775 for (k = ca->sb.d; k < ca->sb.d + ca->sb.keys; k++)
1776 SET_GC_MARK(ca->buckets + *k, GC_MARK_METADATA);
1778 for (k = ca->prio_buckets;
1779 k < ca->prio_buckets + prio_buckets(ca) * 2; k++)
1780 SET_GC_MARK(ca->buckets + *k, GC_MARK_METADATA);
1782 for_each_bucket(b, ca) {
1783 c->need_gc = max(c->need_gc, bucket_gc_gen(b));
1785 if (atomic_read(&b->pin))
1788 BUG_ON(!GC_MARK(b) && GC_SECTORS_USED(b));
1790 if (!GC_MARK(b) || GC_MARK(b) == GC_MARK_RECLAIMABLE)
1791 c->avail_nbuckets++;
1794 mutex_unlock(&c->bucket_lock);
1797 static void bch_btree_gc(struct cache_set *c)
1800 struct gc_stat stats;
1801 struct closure writes;
1803 uint64_t start_time = local_clock();
1805 trace_bcache_gc_start(c);
1807 memset(&stats, 0, sizeof(struct gc_stat));
1808 closure_init_stack(&writes);
1809 bch_btree_op_init(&op, SHRT_MAX);
1813 /* if CACHE_SET_IO_DISABLE set, gc thread should stop too */
1815 ret = bcache_btree_root(gc_root, c, &op, &writes, &stats);
1816 closure_sync(&writes);
1820 schedule_timeout_interruptible(msecs_to_jiffies
1823 pr_warn("gc failed!\n");
1824 } while (ret && !test_bit(CACHE_SET_IO_DISABLE, &c->flags));
1826 bch_btree_gc_finish(c);
1827 wake_up_allocators(c);
1829 bch_time_stats_update(&c->btree_gc_time, start_time);
1831 stats.key_bytes *= sizeof(uint64_t);
1833 bch_update_bucket_in_use(c, &stats);
1834 memcpy(&c->gc_stats, &stats, sizeof(struct gc_stat));
1836 trace_bcache_gc_end(c);
1841 static bool gc_should_run(struct cache_set *c)
1843 struct cache *ca = c->cache;
1845 if (ca->invalidate_needs_gc)
1848 if (atomic_read(&c->sectors_to_gc) < 0)
1854 static int bch_gc_thread(void *arg)
1856 struct cache_set *c = arg;
1859 wait_event_interruptible(c->gc_wait,
1860 kthread_should_stop() ||
1861 test_bit(CACHE_SET_IO_DISABLE, &c->flags) ||
1864 if (kthread_should_stop() ||
1865 test_bit(CACHE_SET_IO_DISABLE, &c->flags))
1872 wait_for_kthread_stop();
1876 int bch_gc_thread_start(struct cache_set *c)
1878 c->gc_thread = kthread_run(bch_gc_thread, c, "bcache_gc");
1879 return PTR_ERR_OR_ZERO(c->gc_thread);
1882 /* Initial partial gc */
1884 static int bch_btree_check_recurse(struct btree *b, struct btree_op *op)
1887 struct bkey *k, *p = NULL;
1888 struct btree_iter iter;
1890 for_each_key_filter(&b->keys, k, &iter, bch_ptr_invalid)
1891 bch_initial_mark_key(b->c, b->level, k);
1893 bch_initial_mark_key(b->c, b->level + 1, &b->key);
1896 bch_btree_iter_init(&b->keys, &iter, NULL);
1899 k = bch_btree_iter_next_filter(&iter, &b->keys,
1902 btree_node_prefetch(b, k);
1904 * initiallize c->gc_stats.nodes
1905 * for incremental GC
1907 b->c->gc_stats.nodes++;
1911 ret = bcache_btree(check_recurse, p, b, op);
1914 } while (p && !ret);
1921 static int bch_btree_check_thread(void *arg)
1924 struct btree_check_info *info = arg;
1925 struct btree_check_state *check_state = info->state;
1926 struct cache_set *c = check_state->c;
1927 struct btree_iter iter;
1929 int cur_idx, prev_idx, skip_nr;
1932 cur_idx = prev_idx = 0;
1935 /* root node keys are checked before thread created */
1936 bch_btree_iter_init(&c->root->keys, &iter, NULL);
1937 k = bch_btree_iter_next_filter(&iter, &c->root->keys, bch_ptr_bad);
1943 * Fetch a root node key index, skip the keys which
1944 * should be fetched by other threads, then check the
1945 * sub-tree indexed by the fetched key.
1947 spin_lock(&check_state->idx_lock);
1948 cur_idx = check_state->key_idx;
1949 check_state->key_idx++;
1950 spin_unlock(&check_state->idx_lock);
1952 skip_nr = cur_idx - prev_idx;
1955 k = bch_btree_iter_next_filter(&iter,
1962 * No more keys to check in root node,
1963 * current checking threads are enough,
1964 * stop creating more.
1966 atomic_set(&check_state->enough, 1);
1967 /* Update check_state->enough earlier */
1968 smp_mb__after_atomic();
1978 btree_node_prefetch(c->root, p);
1979 c->gc_stats.nodes++;
1980 bch_btree_op_init(&op, 0);
1981 ret = bcache_btree(check_recurse, p, c->root, &op);
1983 * The op may be added to cache_set's btree_cache_wait
1984 * in mca_cannibalize(), must ensure it is removed from
1985 * the list and release btree_cache_alloc_lock before
1987 * Otherwise, the btree_cache_wait will be damaged.
1989 bch_cannibalize_unlock(c);
1990 finish_wait(&c->btree_cache_wait, &(&op)->wait);
2001 /* update check_state->started among all CPUs */
2002 smp_mb__before_atomic();
2003 if (atomic_dec_and_test(&check_state->started))
2004 wake_up(&check_state->wait);
2011 static int bch_btree_chkthread_nr(void)
2013 int n = num_online_cpus()/2;
2017 else if (n > BCH_BTR_CHKTHREAD_MAX)
2018 n = BCH_BTR_CHKTHREAD_MAX;
2023 int bch_btree_check(struct cache_set *c)
2027 struct bkey *k = NULL;
2028 struct btree_iter iter;
2029 struct btree_check_state check_state;
2031 /* check and mark root node keys */
2032 for_each_key_filter(&c->root->keys, k, &iter, bch_ptr_invalid)
2033 bch_initial_mark_key(c, c->root->level, k);
2035 bch_initial_mark_key(c, c->root->level + 1, &c->root->key);
2037 if (c->root->level == 0)
2040 memset(&check_state, 0, sizeof(struct btree_check_state));
2042 check_state.total_threads = bch_btree_chkthread_nr();
2043 check_state.key_idx = 0;
2044 spin_lock_init(&check_state.idx_lock);
2045 atomic_set(&check_state.started, 0);
2046 atomic_set(&check_state.enough, 0);
2047 init_waitqueue_head(&check_state.wait);
2049 rw_lock(0, c->root, c->root->level);
2051 * Run multiple threads to check btree nodes in parallel,
2052 * if check_state.enough is non-zero, it means current
2053 * running check threads are enough, unncessary to create
2056 for (i = 0; i < check_state.total_threads; i++) {
2057 /* fetch latest check_state.enough earlier */
2058 smp_mb__before_atomic();
2059 if (atomic_read(&check_state.enough))
2062 check_state.infos[i].result = 0;
2063 check_state.infos[i].state = &check_state;
2065 check_state.infos[i].thread =
2066 kthread_run(bch_btree_check_thread,
2067 &check_state.infos[i],
2068 "bch_btrchk[%d]", i);
2069 if (IS_ERR(check_state.infos[i].thread)) {
2070 pr_err("fails to run thread bch_btrchk[%d]\n", i);
2071 for (--i; i >= 0; i--)
2072 kthread_stop(check_state.infos[i].thread);
2076 atomic_inc(&check_state.started);
2080 * Must wait for all threads to stop.
2082 wait_event(check_state.wait, atomic_read(&check_state.started) == 0);
2084 for (i = 0; i < check_state.total_threads; i++) {
2085 if (check_state.infos[i].result) {
2086 ret = check_state.infos[i].result;
2092 rw_unlock(0, c->root);
2096 void bch_initial_gc_finish(struct cache_set *c)
2098 struct cache *ca = c->cache;
2101 bch_btree_gc_finish(c);
2103 mutex_lock(&c->bucket_lock);
2106 * We need to put some unused buckets directly on the prio freelist in
2107 * order to get the allocator thread started - it needs freed buckets in
2108 * order to rewrite the prios and gens, and it needs to rewrite prios
2109 * and gens in order to free buckets.
2111 * This is only safe for buckets that have no live data in them, which
2112 * there should always be some of.
2114 for_each_bucket(b, ca) {
2115 if (fifo_full(&ca->free[RESERVE_PRIO]) &&
2116 fifo_full(&ca->free[RESERVE_BTREE]))
2119 if (bch_can_invalidate_bucket(ca, b) &&
2121 __bch_invalidate_one_bucket(ca, b);
2122 if (!fifo_push(&ca->free[RESERVE_PRIO],
2124 fifo_push(&ca->free[RESERVE_BTREE],
2129 mutex_unlock(&c->bucket_lock);
2132 /* Btree insertion */
2134 static bool btree_insert_key(struct btree *b, struct bkey *k,
2135 struct bkey *replace_key)
2137 unsigned int status;
2139 BUG_ON(bkey_cmp(k, &b->key) > 0);
2141 status = bch_btree_insert_key(&b->keys, k, replace_key);
2142 if (status != BTREE_INSERT_STATUS_NO_INSERT) {
2143 bch_check_keys(&b->keys, "%u for %s", status,
2144 replace_key ? "replace" : "insert");
2146 trace_bcache_btree_insert_key(b, k, replace_key != NULL,
2153 static size_t insert_u64s_remaining(struct btree *b)
2155 long ret = bch_btree_keys_u64s_remaining(&b->keys);
2158 * Might land in the middle of an existing extent and have to split it
2160 if (b->keys.ops->is_extents)
2161 ret -= KEY_MAX_U64S;
2163 return max(ret, 0L);
2166 static bool bch_btree_insert_keys(struct btree *b, struct btree_op *op,
2167 struct keylist *insert_keys,
2168 struct bkey *replace_key)
2171 int oldsize = bch_count_data(&b->keys);
2173 while (!bch_keylist_empty(insert_keys)) {
2174 struct bkey *k = insert_keys->keys;
2176 if (bkey_u64s(k) > insert_u64s_remaining(b))
2179 if (bkey_cmp(k, &b->key) <= 0) {
2183 ret |= btree_insert_key(b, k, replace_key);
2184 bch_keylist_pop_front(insert_keys);
2185 } else if (bkey_cmp(&START_KEY(k), &b->key) < 0) {
2186 BKEY_PADDED(key) temp;
2187 bkey_copy(&temp.key, insert_keys->keys);
2189 bch_cut_back(&b->key, &temp.key);
2190 bch_cut_front(&b->key, insert_keys->keys);
2192 ret |= btree_insert_key(b, &temp.key, replace_key);
2200 op->insert_collision = true;
2202 BUG_ON(!bch_keylist_empty(insert_keys) && b->level);
2204 BUG_ON(bch_count_data(&b->keys) < oldsize);
2208 static int btree_split(struct btree *b, struct btree_op *op,
2209 struct keylist *insert_keys,
2210 struct bkey *replace_key)
2213 struct btree *n1, *n2 = NULL, *n3 = NULL;
2214 uint64_t start_time = local_clock();
2216 struct keylist parent_keys;
2218 closure_init_stack(&cl);
2219 bch_keylist_init(&parent_keys);
2221 if (btree_check_reserve(b, op)) {
2225 WARN(1, "insufficient reserve for split\n");
2228 n1 = btree_node_alloc_replacement(b, op);
2232 split = set_blocks(btree_bset_first(n1),
2233 block_bytes(n1->c->cache)) > (btree_blocks(b) * 4) / 5;
2236 unsigned int keys = 0;
2238 trace_bcache_btree_node_split(b, btree_bset_first(n1)->keys);
2240 n2 = bch_btree_node_alloc(b->c, op, b->level, b->parent);
2245 n3 = bch_btree_node_alloc(b->c, op, b->level + 1, NULL);
2250 mutex_lock(&n1->write_lock);
2251 mutex_lock(&n2->write_lock);
2253 bch_btree_insert_keys(n1, op, insert_keys, replace_key);
2256 * Has to be a linear search because we don't have an auxiliary
2260 while (keys < (btree_bset_first(n1)->keys * 3) / 5)
2261 keys += bkey_u64s(bset_bkey_idx(btree_bset_first(n1),
2264 bkey_copy_key(&n1->key,
2265 bset_bkey_idx(btree_bset_first(n1), keys));
2266 keys += bkey_u64s(bset_bkey_idx(btree_bset_first(n1), keys));
2268 btree_bset_first(n2)->keys = btree_bset_first(n1)->keys - keys;
2269 btree_bset_first(n1)->keys = keys;
2271 memcpy(btree_bset_first(n2)->start,
2272 bset_bkey_last(btree_bset_first(n1)),
2273 btree_bset_first(n2)->keys * sizeof(uint64_t));
2275 bkey_copy_key(&n2->key, &b->key);
2277 bch_keylist_add(&parent_keys, &n2->key);
2278 bch_btree_node_write(n2, &cl);
2279 mutex_unlock(&n2->write_lock);
2280 rw_unlock(true, n2);
2282 trace_bcache_btree_node_compact(b, btree_bset_first(n1)->keys);
2284 mutex_lock(&n1->write_lock);
2285 bch_btree_insert_keys(n1, op, insert_keys, replace_key);
2288 bch_keylist_add(&parent_keys, &n1->key);
2289 bch_btree_node_write(n1, &cl);
2290 mutex_unlock(&n1->write_lock);
2293 /* Depth increases, make a new root */
2294 mutex_lock(&n3->write_lock);
2295 bkey_copy_key(&n3->key, &MAX_KEY);
2296 bch_btree_insert_keys(n3, op, &parent_keys, NULL);
2297 bch_btree_node_write(n3, &cl);
2298 mutex_unlock(&n3->write_lock);
2301 bch_btree_set_root(n3);
2302 rw_unlock(true, n3);
2303 } else if (!b->parent) {
2304 /* Root filled up but didn't need to be split */
2306 bch_btree_set_root(n1);
2308 /* Split a non root node */
2310 make_btree_freeing_key(b, parent_keys.top);
2311 bch_keylist_push(&parent_keys);
2313 bch_btree_insert_node(b->parent, op, &parent_keys, NULL, NULL);
2314 BUG_ON(!bch_keylist_empty(&parent_keys));
2318 rw_unlock(true, n1);
2320 bch_time_stats_update(&b->c->btree_split_time, start_time);
2324 bkey_put(b->c, &n2->key);
2325 btree_node_free(n2);
2326 rw_unlock(true, n2);
2328 bkey_put(b->c, &n1->key);
2329 btree_node_free(n1);
2330 rw_unlock(true, n1);
2332 WARN(1, "bcache: btree split failed (level %u)", b->level);
2334 if (n3 == ERR_PTR(-EAGAIN) ||
2335 n2 == ERR_PTR(-EAGAIN) ||
2336 n1 == ERR_PTR(-EAGAIN))
2342 static int bch_btree_insert_node(struct btree *b, struct btree_op *op,
2343 struct keylist *insert_keys,
2344 atomic_t *journal_ref,
2345 struct bkey *replace_key)
2349 BUG_ON(b->level && replace_key);
2351 closure_init_stack(&cl);
2353 mutex_lock(&b->write_lock);
2355 if (write_block(b) != btree_bset_last(b) &&
2356 b->keys.last_set_unwritten)
2357 bch_btree_init_next(b); /* just wrote a set */
2359 if (bch_keylist_nkeys(insert_keys) > insert_u64s_remaining(b)) {
2360 mutex_unlock(&b->write_lock);
2364 BUG_ON(write_block(b) != btree_bset_last(b));
2366 if (bch_btree_insert_keys(b, op, insert_keys, replace_key)) {
2368 bch_btree_leaf_dirty(b, journal_ref);
2370 bch_btree_node_write(b, &cl);
2373 mutex_unlock(&b->write_lock);
2375 /* wait for btree node write if necessary, after unlock */
2380 if (current->bio_list) {
2381 op->lock = b->c->root->level + 1;
2383 } else if (op->lock <= b->c->root->level) {
2384 op->lock = b->c->root->level + 1;
2387 /* Invalidated all iterators */
2388 int ret = btree_split(b, op, insert_keys, replace_key);
2390 if (bch_keylist_empty(insert_keys))
2398 int bch_btree_insert_check_key(struct btree *b, struct btree_op *op,
2399 struct bkey *check_key)
2402 uint64_t btree_ptr = b->key.ptr[0];
2403 unsigned long seq = b->seq;
2404 struct keylist insert;
2405 bool upgrade = op->lock == -1;
2407 bch_keylist_init(&insert);
2410 rw_unlock(false, b);
2411 rw_lock(true, b, b->level);
2413 if (b->key.ptr[0] != btree_ptr ||
2414 b->seq != seq + 1) {
2415 op->lock = b->level;
2420 SET_KEY_PTRS(check_key, 1);
2421 get_random_bytes(&check_key->ptr[0], sizeof(uint64_t));
2423 SET_PTR_DEV(check_key, 0, PTR_CHECK_DEV);
2425 bch_keylist_add(&insert, check_key);
2427 ret = bch_btree_insert_node(b, op, &insert, NULL, NULL);
2429 BUG_ON(!ret && !bch_keylist_empty(&insert));
2432 downgrade_write(&b->lock);
2436 struct btree_insert_op {
2438 struct keylist *keys;
2439 atomic_t *journal_ref;
2440 struct bkey *replace_key;
2443 static int btree_insert_fn(struct btree_op *b_op, struct btree *b)
2445 struct btree_insert_op *op = container_of(b_op,
2446 struct btree_insert_op, op);
2448 int ret = bch_btree_insert_node(b, &op->op, op->keys,
2449 op->journal_ref, op->replace_key);
2450 if (ret && !bch_keylist_empty(op->keys))
2456 int bch_btree_insert(struct cache_set *c, struct keylist *keys,
2457 atomic_t *journal_ref, struct bkey *replace_key)
2459 struct btree_insert_op op;
2462 BUG_ON(current->bio_list);
2463 BUG_ON(bch_keylist_empty(keys));
2465 bch_btree_op_init(&op.op, 0);
2467 op.journal_ref = journal_ref;
2468 op.replace_key = replace_key;
2470 while (!ret && !bch_keylist_empty(keys)) {
2472 ret = bch_btree_map_leaf_nodes(&op.op, c,
2473 &START_KEY(keys->keys),
2480 pr_err("error %i\n", ret);
2482 while ((k = bch_keylist_pop(keys)))
2484 } else if (op.op.insert_collision)
2490 void bch_btree_set_root(struct btree *b)
2495 closure_init_stack(&cl);
2497 trace_bcache_btree_set_root(b);
2499 BUG_ON(!b->written);
2501 for (i = 0; i < KEY_PTRS(&b->key); i++)
2502 BUG_ON(PTR_BUCKET(b->c, &b->key, i)->prio != BTREE_PRIO);
2504 mutex_lock(&b->c->bucket_lock);
2505 list_del_init(&b->list);
2506 mutex_unlock(&b->c->bucket_lock);
2510 bch_journal_meta(b->c, &cl);
2514 /* Map across nodes or keys */
2516 static int bch_btree_map_nodes_recurse(struct btree *b, struct btree_op *op,
2518 btree_map_nodes_fn *fn, int flags)
2520 int ret = MAP_CONTINUE;
2524 struct btree_iter iter;
2526 bch_btree_iter_init(&b->keys, &iter, from);
2528 while ((k = bch_btree_iter_next_filter(&iter, &b->keys,
2530 ret = bcache_btree(map_nodes_recurse, k, b,
2531 op, from, fn, flags);
2534 if (ret != MAP_CONTINUE)
2539 if (!b->level || flags == MAP_ALL_NODES)
2545 int __bch_btree_map_nodes(struct btree_op *op, struct cache_set *c,
2546 struct bkey *from, btree_map_nodes_fn *fn, int flags)
2548 return bcache_btree_root(map_nodes_recurse, c, op, from, fn, flags);
2551 int bch_btree_map_keys_recurse(struct btree *b, struct btree_op *op,
2552 struct bkey *from, btree_map_keys_fn *fn,
2555 int ret = MAP_CONTINUE;
2557 struct btree_iter iter;
2559 bch_btree_iter_init(&b->keys, &iter, from);
2561 while ((k = bch_btree_iter_next_filter(&iter, &b->keys, bch_ptr_bad))) {
2564 : bcache_btree(map_keys_recurse, k,
2565 b, op, from, fn, flags);
2568 if (ret != MAP_CONTINUE)
2572 if (!b->level && (flags & MAP_END_KEY))
2573 ret = fn(op, b, &KEY(KEY_INODE(&b->key),
2574 KEY_OFFSET(&b->key), 0));
2579 int bch_btree_map_keys(struct btree_op *op, struct cache_set *c,
2580 struct bkey *from, btree_map_keys_fn *fn, int flags)
2582 return bcache_btree_root(map_keys_recurse, c, op, from, fn, flags);
2587 static inline int keybuf_cmp(struct keybuf_key *l, struct keybuf_key *r)
2589 /* Overlapping keys compare equal */
2590 if (bkey_cmp(&l->key, &START_KEY(&r->key)) <= 0)
2592 if (bkey_cmp(&START_KEY(&l->key), &r->key) >= 0)
2597 static inline int keybuf_nonoverlapping_cmp(struct keybuf_key *l,
2598 struct keybuf_key *r)
2600 return clamp_t(int64_t, bkey_cmp(&l->key, &r->key), -1, 1);
2605 unsigned int nr_found;
2608 keybuf_pred_fn *pred;
2611 static int refill_keybuf_fn(struct btree_op *op, struct btree *b,
2614 struct refill *refill = container_of(op, struct refill, op);
2615 struct keybuf *buf = refill->buf;
2616 int ret = MAP_CONTINUE;
2618 if (bkey_cmp(k, refill->end) > 0) {
2623 if (!KEY_SIZE(k)) /* end key */
2626 if (refill->pred(buf, k)) {
2627 struct keybuf_key *w;
2629 spin_lock(&buf->lock);
2631 w = array_alloc(&buf->freelist);
2633 spin_unlock(&buf->lock);
2638 bkey_copy(&w->key, k);
2640 if (RB_INSERT(&buf->keys, w, node, keybuf_cmp))
2641 array_free(&buf->freelist, w);
2645 if (array_freelist_empty(&buf->freelist))
2648 spin_unlock(&buf->lock);
2651 buf->last_scanned = *k;
2655 void bch_refill_keybuf(struct cache_set *c, struct keybuf *buf,
2656 struct bkey *end, keybuf_pred_fn *pred)
2658 struct bkey start = buf->last_scanned;
2659 struct refill refill;
2663 bch_btree_op_init(&refill.op, -1);
2664 refill.nr_found = 0;
2669 bch_btree_map_keys(&refill.op, c, &buf->last_scanned,
2670 refill_keybuf_fn, MAP_END_KEY);
2672 trace_bcache_keyscan(refill.nr_found,
2673 KEY_INODE(&start), KEY_OFFSET(&start),
2674 KEY_INODE(&buf->last_scanned),
2675 KEY_OFFSET(&buf->last_scanned));
2677 spin_lock(&buf->lock);
2679 if (!RB_EMPTY_ROOT(&buf->keys)) {
2680 struct keybuf_key *w;
2682 w = RB_FIRST(&buf->keys, struct keybuf_key, node);
2683 buf->start = START_KEY(&w->key);
2685 w = RB_LAST(&buf->keys, struct keybuf_key, node);
2688 buf->start = MAX_KEY;
2692 spin_unlock(&buf->lock);
2695 static void __bch_keybuf_del(struct keybuf *buf, struct keybuf_key *w)
2697 rb_erase(&w->node, &buf->keys);
2698 array_free(&buf->freelist, w);
2701 void bch_keybuf_del(struct keybuf *buf, struct keybuf_key *w)
2703 spin_lock(&buf->lock);
2704 __bch_keybuf_del(buf, w);
2705 spin_unlock(&buf->lock);
2708 bool bch_keybuf_check_overlapping(struct keybuf *buf, struct bkey *start,
2712 struct keybuf_key *p, *w, s;
2716 if (bkey_cmp(end, &buf->start) <= 0 ||
2717 bkey_cmp(start, &buf->end) >= 0)
2720 spin_lock(&buf->lock);
2721 w = RB_GREATER(&buf->keys, s, node, keybuf_nonoverlapping_cmp);
2723 while (w && bkey_cmp(&START_KEY(&w->key), end) < 0) {
2725 w = RB_NEXT(w, node);
2730 __bch_keybuf_del(buf, p);
2733 spin_unlock(&buf->lock);
2737 struct keybuf_key *bch_keybuf_next(struct keybuf *buf)
2739 struct keybuf_key *w;
2741 spin_lock(&buf->lock);
2743 w = RB_FIRST(&buf->keys, struct keybuf_key, node);
2745 while (w && w->private)
2746 w = RB_NEXT(w, node);
2749 w->private = ERR_PTR(-EINTR);
2751 spin_unlock(&buf->lock);
2755 struct keybuf_key *bch_keybuf_next_rescan(struct cache_set *c,
2758 keybuf_pred_fn *pred)
2760 struct keybuf_key *ret;
2763 ret = bch_keybuf_next(buf);
2767 if (bkey_cmp(&buf->last_scanned, end) >= 0) {
2768 pr_debug("scan finished\n");
2772 bch_refill_keybuf(c, buf, end, pred);
2778 void bch_keybuf_init(struct keybuf *buf)
2780 buf->last_scanned = MAX_KEY;
2781 buf->keys = RB_ROOT;
2783 spin_lock_init(&buf->lock);
2784 array_allocator_init(&buf->freelist);
2787 void bch_btree_exit(void)
2790 destroy_workqueue(btree_io_wq);
2793 int __init bch_btree_init(void)
2795 btree_io_wq = alloc_workqueue("bch_btree_io", WQ_MEM_RECLAIM, 0);