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 = crc64_be(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 &b->c->cache->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 #define cmp_int(l, r) ((l > r) - (l < r))
564 #ifdef CONFIG_PROVE_LOCKING
565 static int btree_lock_cmp_fn(const struct lockdep_map *_a,
566 const struct lockdep_map *_b)
568 const struct btree *a = container_of(_a, struct btree, lock.dep_map);
569 const struct btree *b = container_of(_b, struct btree, lock.dep_map);
571 return -cmp_int(a->level, b->level) ?: bkey_cmp(&a->key, &b->key);
574 static void btree_lock_print_fn(const struct lockdep_map *map)
576 const struct btree *b = container_of(map, struct btree, lock.dep_map);
578 printk(KERN_CONT " l=%u %llu:%llu", b->level,
579 KEY_INODE(&b->key), KEY_OFFSET(&b->key));
583 static struct btree *mca_bucket_alloc(struct cache_set *c,
584 struct bkey *k, gfp_t gfp)
587 * kzalloc() is necessary here for initialization,
588 * see code comments in bch_btree_keys_init().
590 struct btree *b = kzalloc(sizeof(struct btree), gfp);
595 init_rwsem(&b->lock);
596 lock_set_cmp_fn(&b->lock, btree_lock_cmp_fn, btree_lock_print_fn);
597 mutex_init(&b->write_lock);
598 lockdep_set_novalidate_class(&b->write_lock);
599 INIT_LIST_HEAD(&b->list);
600 INIT_DELAYED_WORK(&b->work, btree_node_write_work);
602 sema_init(&b->io_mutex, 1);
604 mca_data_alloc(b, k, gfp);
608 static int mca_reap(struct btree *b, unsigned int min_order, bool flush)
612 closure_init_stack(&cl);
613 lockdep_assert_held(&b->c->bucket_lock);
615 if (!down_write_trylock(&b->lock))
618 BUG_ON(btree_node_dirty(b) && !b->keys.set[0].data);
620 if (b->keys.page_order < min_order)
624 if (btree_node_dirty(b))
627 if (down_trylock(&b->io_mutex))
634 * BTREE_NODE_dirty might be cleared in btree_flush_btree() by
635 * __bch_btree_node_write(). To avoid an extra flush, acquire
636 * b->write_lock before checking BTREE_NODE_dirty bit.
638 mutex_lock(&b->write_lock);
640 * If this btree node is selected in btree_flush_write() by journal
641 * code, delay and retry until the node is flushed by journal code
642 * and BTREE_NODE_journal_flush bit cleared by btree_flush_write().
644 if (btree_node_journal_flush(b)) {
645 pr_debug("bnode %p is flushing by journal, retry\n", b);
646 mutex_unlock(&b->write_lock);
651 if (btree_node_dirty(b))
652 __bch_btree_node_write(b, &cl);
653 mutex_unlock(&b->write_lock);
657 /* wait for any in flight btree write */
667 static unsigned long bch_mca_scan(struct shrinker *shrink,
668 struct shrink_control *sc)
670 struct cache_set *c = container_of(shrink, struct cache_set, shrink);
672 unsigned long i, nr = sc->nr_to_scan;
673 unsigned long freed = 0;
674 unsigned int btree_cache_used;
676 if (c->shrinker_disabled)
679 if (c->btree_cache_alloc_lock)
682 /* Return -1 if we can't do anything right now */
683 if (sc->gfp_mask & __GFP_IO)
684 mutex_lock(&c->bucket_lock);
685 else if (!mutex_trylock(&c->bucket_lock))
689 * It's _really_ critical that we don't free too many btree nodes - we
690 * have to always leave ourselves a reserve. The reserve is how we
691 * guarantee that allocating memory for a new btree node can always
692 * succeed, so that inserting keys into the btree can always succeed and
693 * IO can always make forward progress:
695 nr /= c->btree_pages;
698 nr = min_t(unsigned long, nr, mca_can_free(c));
701 btree_cache_used = c->btree_cache_used;
702 list_for_each_entry_safe_reverse(b, t, &c->btree_cache_freeable, list) {
706 if (!mca_reap(b, 0, false)) {
715 list_for_each_entry_safe_reverse(b, t, &c->btree_cache, list) {
716 if (nr <= 0 || i >= btree_cache_used)
719 if (!mca_reap(b, 0, false)) {
730 mutex_unlock(&c->bucket_lock);
731 return freed * c->btree_pages;
734 static unsigned long bch_mca_count(struct shrinker *shrink,
735 struct shrink_control *sc)
737 struct cache_set *c = container_of(shrink, struct cache_set, shrink);
739 if (c->shrinker_disabled)
742 if (c->btree_cache_alloc_lock)
745 return mca_can_free(c) * c->btree_pages;
748 void bch_btree_cache_free(struct cache_set *c)
753 closure_init_stack(&cl);
755 if (c->shrink.list.next)
756 unregister_shrinker(&c->shrink);
758 mutex_lock(&c->bucket_lock);
760 #ifdef CONFIG_BCACHE_DEBUG
762 list_move(&c->verify_data->list, &c->btree_cache);
764 free_pages((unsigned long) c->verify_ondisk, ilog2(meta_bucket_pages(&c->cache->sb)));
767 list_splice(&c->btree_cache_freeable,
770 while (!list_empty(&c->btree_cache)) {
771 b = list_first_entry(&c->btree_cache, struct btree, list);
774 * This function is called by cache_set_free(), no I/O
775 * request on cache now, it is unnecessary to acquire
776 * b->write_lock before clearing BTREE_NODE_dirty anymore.
778 if (btree_node_dirty(b)) {
779 btree_complete_write(b, btree_current_write(b));
780 clear_bit(BTREE_NODE_dirty, &b->flags);
785 while (!list_empty(&c->btree_cache_freed)) {
786 b = list_first_entry(&c->btree_cache_freed,
789 cancel_delayed_work_sync(&b->work);
793 mutex_unlock(&c->bucket_lock);
796 int bch_btree_cache_alloc(struct cache_set *c)
800 for (i = 0; i < mca_reserve(c); i++)
801 if (!mca_bucket_alloc(c, &ZERO_KEY, GFP_KERNEL))
804 list_splice_init(&c->btree_cache,
805 &c->btree_cache_freeable);
807 #ifdef CONFIG_BCACHE_DEBUG
808 mutex_init(&c->verify_lock);
810 c->verify_ondisk = (void *)
811 __get_free_pages(GFP_KERNEL|__GFP_COMP,
812 ilog2(meta_bucket_pages(&c->cache->sb)));
813 if (!c->verify_ondisk) {
815 * Don't worry about the mca_rereserve buckets
816 * allocated in previous for-loop, they will be
817 * handled properly in bch_cache_set_unregister().
822 c->verify_data = mca_bucket_alloc(c, &ZERO_KEY, GFP_KERNEL);
824 if (c->verify_data &&
825 c->verify_data->keys.set->data)
826 list_del_init(&c->verify_data->list);
828 c->verify_data = NULL;
831 c->shrink.count_objects = bch_mca_count;
832 c->shrink.scan_objects = bch_mca_scan;
834 c->shrink.batch = c->btree_pages * 2;
836 if (register_shrinker(&c->shrink, "md-bcache:%pU", c->set_uuid))
837 pr_warn("bcache: %s: could not register shrinker\n",
843 /* Btree in memory cache - hash table */
845 static struct hlist_head *mca_hash(struct cache_set *c, struct bkey *k)
847 return &c->bucket_hash[hash_32(PTR_HASH(c, k), BUCKET_HASH_BITS)];
850 static struct btree *mca_find(struct cache_set *c, struct bkey *k)
855 hlist_for_each_entry_rcu(b, mca_hash(c, k), hash)
856 if (PTR_HASH(c, &b->key) == PTR_HASH(c, k))
864 static int mca_cannibalize_lock(struct cache_set *c, struct btree_op *op)
866 spin_lock(&c->btree_cannibalize_lock);
867 if (likely(c->btree_cache_alloc_lock == NULL)) {
868 c->btree_cache_alloc_lock = current;
869 } else if (c->btree_cache_alloc_lock != current) {
871 prepare_to_wait(&c->btree_cache_wait, &op->wait,
872 TASK_UNINTERRUPTIBLE);
873 spin_unlock(&c->btree_cannibalize_lock);
876 spin_unlock(&c->btree_cannibalize_lock);
881 static struct btree *mca_cannibalize(struct cache_set *c, struct btree_op *op,
886 trace_bcache_btree_cache_cannibalize(c);
888 if (mca_cannibalize_lock(c, op))
889 return ERR_PTR(-EINTR);
891 list_for_each_entry_reverse(b, &c->btree_cache, list)
892 if (!mca_reap(b, btree_order(k), false))
895 list_for_each_entry_reverse(b, &c->btree_cache, list)
896 if (!mca_reap(b, btree_order(k), true))
899 WARN(1, "btree cache cannibalize failed\n");
900 return ERR_PTR(-ENOMEM);
904 * We can only have one thread cannibalizing other cached btree nodes at a time,
905 * or we'll deadlock. We use an open coded mutex to ensure that, which a
906 * cannibalize_bucket() will take. This means every time we unlock the root of
907 * the btree, we need to release this lock if we have it held.
909 void bch_cannibalize_unlock(struct cache_set *c)
911 spin_lock(&c->btree_cannibalize_lock);
912 if (c->btree_cache_alloc_lock == current) {
913 c->btree_cache_alloc_lock = NULL;
914 wake_up(&c->btree_cache_wait);
916 spin_unlock(&c->btree_cannibalize_lock);
919 static struct btree *mca_alloc(struct cache_set *c, struct btree_op *op,
920 struct bkey *k, int level)
924 BUG_ON(current->bio_list);
926 lockdep_assert_held(&c->bucket_lock);
931 /* btree_free() doesn't free memory; it sticks the node on the end of
932 * the list. Check if there's any freed nodes there:
934 list_for_each_entry(b, &c->btree_cache_freeable, list)
935 if (!mca_reap(b, btree_order(k), false))
938 /* We never free struct btree itself, just the memory that holds the on
939 * disk node. Check the freed list before allocating a new one:
941 list_for_each_entry(b, &c->btree_cache_freed, list)
942 if (!mca_reap(b, 0, false)) {
943 mca_data_alloc(b, k, __GFP_NOWARN|GFP_NOIO);
944 if (!b->keys.set[0].data)
950 b = mca_bucket_alloc(c, k, __GFP_NOWARN|GFP_NOIO);
954 BUG_ON(!down_write_trylock(&b->lock));
955 if (!b->keys.set->data)
958 BUG_ON(b->io_mutex.count != 1);
960 bkey_copy(&b->key, k);
961 list_move(&b->list, &c->btree_cache);
962 hlist_del_init_rcu(&b->hash);
963 hlist_add_head_rcu(&b->hash, mca_hash(c, k));
965 lock_set_subclass(&b->lock.dep_map, level + 1, _THIS_IP_);
966 b->parent = (void *) ~0UL;
972 bch_btree_keys_init(&b->keys, &bch_extent_keys_ops,
973 &b->c->expensive_debug_checks);
975 bch_btree_keys_init(&b->keys, &bch_btree_keys_ops,
976 &b->c->expensive_debug_checks);
983 b = mca_cannibalize(c, op, k);
991 * bch_btree_node_get - find a btree node in the cache and lock it, reading it
992 * in from disk if necessary.
994 * If IO is necessary and running under submit_bio_noacct, returns -EAGAIN.
996 * The btree node will have either a read or a write lock held, depending on
997 * level and op->lock.
999 struct btree *bch_btree_node_get(struct cache_set *c, struct btree_op *op,
1000 struct bkey *k, int level, bool write,
1001 struct btree *parent)
1011 if (current->bio_list)
1012 return ERR_PTR(-EAGAIN);
1014 mutex_lock(&c->bucket_lock);
1015 b = mca_alloc(c, op, k, level);
1016 mutex_unlock(&c->bucket_lock);
1023 bch_btree_node_read(b);
1026 downgrade_write(&b->lock);
1028 rw_lock(write, b, level);
1029 if (PTR_HASH(c, &b->key) != PTR_HASH(c, k)) {
1030 rw_unlock(write, b);
1033 BUG_ON(b->level != level);
1036 if (btree_node_io_error(b)) {
1037 rw_unlock(write, b);
1038 return ERR_PTR(-EIO);
1041 BUG_ON(!b->written);
1045 for (; i <= b->keys.nsets && b->keys.set[i].size; i++) {
1046 prefetch(b->keys.set[i].tree);
1047 prefetch(b->keys.set[i].data);
1050 for (; i <= b->keys.nsets; i++)
1051 prefetch(b->keys.set[i].data);
1056 static void btree_node_prefetch(struct btree *parent, struct bkey *k)
1060 mutex_lock(&parent->c->bucket_lock);
1061 b = mca_alloc(parent->c, NULL, k, parent->level - 1);
1062 mutex_unlock(&parent->c->bucket_lock);
1064 if (!IS_ERR_OR_NULL(b)) {
1066 bch_btree_node_read(b);
1073 static void btree_node_free(struct btree *b)
1075 trace_bcache_btree_node_free(b);
1077 BUG_ON(b == b->c->root);
1080 mutex_lock(&b->write_lock);
1082 * If the btree node is selected and flushing in btree_flush_write(),
1083 * delay and retry until the BTREE_NODE_journal_flush bit cleared,
1084 * then it is safe to free the btree node here. Otherwise this btree
1085 * node will be in race condition.
1087 if (btree_node_journal_flush(b)) {
1088 mutex_unlock(&b->write_lock);
1089 pr_debug("bnode %p journal_flush set, retry\n", b);
1094 if (btree_node_dirty(b)) {
1095 btree_complete_write(b, btree_current_write(b));
1096 clear_bit(BTREE_NODE_dirty, &b->flags);
1099 mutex_unlock(&b->write_lock);
1101 cancel_delayed_work(&b->work);
1103 mutex_lock(&b->c->bucket_lock);
1104 bch_bucket_free(b->c, &b->key);
1106 mutex_unlock(&b->c->bucket_lock);
1109 struct btree *__bch_btree_node_alloc(struct cache_set *c, struct btree_op *op,
1110 int level, bool wait,
1111 struct btree *parent)
1116 mutex_lock(&c->bucket_lock);
1118 /* return ERR_PTR(-EAGAIN) when it fails */
1119 b = ERR_PTR(-EAGAIN);
1120 if (__bch_bucket_alloc_set(c, RESERVE_BTREE, &k.key, wait))
1123 bkey_put(c, &k.key);
1124 SET_KEY_SIZE(&k.key, c->btree_pages * PAGE_SECTORS);
1126 b = mca_alloc(c, op, &k.key, level);
1132 "Tried to allocate bucket that was in btree cache");
1137 bch_bset_init_next(&b->keys, b->keys.set->data, bset_magic(&b->c->cache->sb));
1139 mutex_unlock(&c->bucket_lock);
1141 trace_bcache_btree_node_alloc(b);
1144 bch_bucket_free(c, &k.key);
1146 mutex_unlock(&c->bucket_lock);
1148 trace_bcache_btree_node_alloc_fail(c);
1152 static struct btree *bch_btree_node_alloc(struct cache_set *c,
1153 struct btree_op *op, int level,
1154 struct btree *parent)
1156 return __bch_btree_node_alloc(c, op, level, op != NULL, parent);
1159 static struct btree *btree_node_alloc_replacement(struct btree *b,
1160 struct btree_op *op)
1162 struct btree *n = bch_btree_node_alloc(b->c, op, b->level, b->parent);
1165 mutex_lock(&n->write_lock);
1166 bch_btree_sort_into(&b->keys, &n->keys, &b->c->sort);
1167 bkey_copy_key(&n->key, &b->key);
1168 mutex_unlock(&n->write_lock);
1174 static void make_btree_freeing_key(struct btree *b, struct bkey *k)
1178 mutex_lock(&b->c->bucket_lock);
1180 atomic_inc(&b->c->prio_blocked);
1182 bkey_copy(k, &b->key);
1183 bkey_copy_key(k, &ZERO_KEY);
1185 for (i = 0; i < KEY_PTRS(k); i++)
1187 bch_inc_gen(b->c->cache,
1188 PTR_BUCKET(b->c, &b->key, i)));
1190 mutex_unlock(&b->c->bucket_lock);
1193 static int btree_check_reserve(struct btree *b, struct btree_op *op)
1195 struct cache_set *c = b->c;
1196 struct cache *ca = c->cache;
1197 unsigned int reserve = (c->root->level - b->level) * 2 + 1;
1199 mutex_lock(&c->bucket_lock);
1201 if (fifo_used(&ca->free[RESERVE_BTREE]) < reserve) {
1203 prepare_to_wait(&c->btree_cache_wait, &op->wait,
1204 TASK_UNINTERRUPTIBLE);
1205 mutex_unlock(&c->bucket_lock);
1209 mutex_unlock(&c->bucket_lock);
1211 return mca_cannibalize_lock(b->c, op);
1214 /* Garbage collection */
1216 static uint8_t __bch_btree_mark_key(struct cache_set *c, int level,
1224 * ptr_invalid() can't return true for the keys that mark btree nodes as
1225 * freed, but since ptr_bad() returns true we'll never actually use them
1226 * for anything and thus we don't want mark their pointers here
1228 if (!bkey_cmp(k, &ZERO_KEY))
1231 for (i = 0; i < KEY_PTRS(k); i++) {
1232 if (!ptr_available(c, k, i))
1235 g = PTR_BUCKET(c, k, i);
1237 if (gen_after(g->last_gc, PTR_GEN(k, i)))
1238 g->last_gc = PTR_GEN(k, i);
1240 if (ptr_stale(c, k, i)) {
1241 stale = max(stale, ptr_stale(c, k, i));
1245 cache_bug_on(GC_MARK(g) &&
1246 (GC_MARK(g) == GC_MARK_METADATA) != (level != 0),
1247 c, "inconsistent ptrs: mark = %llu, level = %i",
1251 SET_GC_MARK(g, GC_MARK_METADATA);
1252 else if (KEY_DIRTY(k))
1253 SET_GC_MARK(g, GC_MARK_DIRTY);
1254 else if (!GC_MARK(g))
1255 SET_GC_MARK(g, GC_MARK_RECLAIMABLE);
1257 /* guard against overflow */
1258 SET_GC_SECTORS_USED(g, min_t(unsigned int,
1259 GC_SECTORS_USED(g) + KEY_SIZE(k),
1260 MAX_GC_SECTORS_USED));
1262 BUG_ON(!GC_SECTORS_USED(g));
1268 #define btree_mark_key(b, k) __bch_btree_mark_key(b->c, b->level, k)
1270 void bch_initial_mark_key(struct cache_set *c, int level, struct bkey *k)
1274 for (i = 0; i < KEY_PTRS(k); i++)
1275 if (ptr_available(c, k, i) &&
1276 !ptr_stale(c, k, i)) {
1277 struct bucket *b = PTR_BUCKET(c, k, i);
1279 b->gen = PTR_GEN(k, i);
1281 if (level && bkey_cmp(k, &ZERO_KEY))
1282 b->prio = BTREE_PRIO;
1283 else if (!level && b->prio == BTREE_PRIO)
1284 b->prio = INITIAL_PRIO;
1287 __bch_btree_mark_key(c, level, k);
1290 void bch_update_bucket_in_use(struct cache_set *c, struct gc_stat *stats)
1292 stats->in_use = (c->nbuckets - c->avail_nbuckets) * 100 / c->nbuckets;
1295 static bool btree_gc_mark_node(struct btree *b, struct gc_stat *gc)
1298 unsigned int keys = 0, good_keys = 0;
1300 struct btree_iter iter;
1301 struct bset_tree *t;
1305 for_each_key_filter(&b->keys, k, &iter, bch_ptr_invalid) {
1306 stale = max(stale, btree_mark_key(b, k));
1309 if (bch_ptr_bad(&b->keys, k))
1312 gc->key_bytes += bkey_u64s(k);
1316 gc->data += KEY_SIZE(k);
1319 for (t = b->keys.set; t <= &b->keys.set[b->keys.nsets]; t++)
1320 btree_bug_on(t->size &&
1321 bset_written(&b->keys, t) &&
1322 bkey_cmp(&b->key, &t->end) < 0,
1323 b, "found short btree key in gc");
1325 if (b->c->gc_always_rewrite)
1331 if ((keys - good_keys) * 2 > keys)
1337 #define GC_MERGE_NODES 4U
1339 struct gc_merge_info {
1344 static int bch_btree_insert_node(struct btree *b, struct btree_op *op,
1345 struct keylist *insert_keys,
1346 atomic_t *journal_ref,
1347 struct bkey *replace_key);
1349 static int btree_gc_coalesce(struct btree *b, struct btree_op *op,
1350 struct gc_stat *gc, struct gc_merge_info *r)
1352 unsigned int i, nodes = 0, keys = 0, blocks;
1353 struct btree *new_nodes[GC_MERGE_NODES];
1354 struct keylist keylist;
1358 bch_keylist_init(&keylist);
1360 if (btree_check_reserve(b, NULL))
1363 memset(new_nodes, 0, sizeof(new_nodes));
1364 closure_init_stack(&cl);
1366 while (nodes < GC_MERGE_NODES && !IS_ERR(r[nodes].b))
1367 keys += r[nodes++].keys;
1369 blocks = btree_default_blocks(b->c) * 2 / 3;
1372 __set_blocks(b->keys.set[0].data, keys,
1373 block_bytes(b->c->cache)) > blocks * (nodes - 1))
1376 for (i = 0; i < nodes; i++) {
1377 new_nodes[i] = btree_node_alloc_replacement(r[i].b, NULL);
1378 if (IS_ERR(new_nodes[i]))
1379 goto out_nocoalesce;
1383 * We have to check the reserve here, after we've allocated our new
1384 * nodes, to make sure the insert below will succeed - we also check
1385 * before as an optimization to potentially avoid a bunch of expensive
1388 if (btree_check_reserve(b, NULL))
1389 goto out_nocoalesce;
1391 for (i = 0; i < nodes; i++)
1392 mutex_lock(&new_nodes[i]->write_lock);
1394 for (i = nodes - 1; i > 0; --i) {
1395 struct bset *n1 = btree_bset_first(new_nodes[i]);
1396 struct bset *n2 = btree_bset_first(new_nodes[i - 1]);
1397 struct bkey *k, *last = NULL;
1403 k < bset_bkey_last(n2);
1405 if (__set_blocks(n1, n1->keys + keys +
1407 block_bytes(b->c->cache)) > blocks)
1411 keys += bkey_u64s(k);
1415 * Last node we're not getting rid of - we're getting
1416 * rid of the node at r[0]. Have to try and fit all of
1417 * the remaining keys into this node; we can't ensure
1418 * they will always fit due to rounding and variable
1419 * length keys (shouldn't be possible in practice,
1422 if (__set_blocks(n1, n1->keys + n2->keys,
1423 block_bytes(b->c->cache)) >
1424 btree_blocks(new_nodes[i]))
1425 goto out_unlock_nocoalesce;
1428 /* Take the key of the node we're getting rid of */
1432 BUG_ON(__set_blocks(n1, n1->keys + keys, block_bytes(b->c->cache)) >
1433 btree_blocks(new_nodes[i]));
1436 bkey_copy_key(&new_nodes[i]->key, last);
1438 memcpy(bset_bkey_last(n1),
1440 (void *) bset_bkey_idx(n2, keys) - (void *) n2->start);
1443 r[i].keys = n1->keys;
1446 bset_bkey_idx(n2, keys),
1447 (void *) bset_bkey_last(n2) -
1448 (void *) bset_bkey_idx(n2, keys));
1452 if (__bch_keylist_realloc(&keylist,
1453 bkey_u64s(&new_nodes[i]->key)))
1454 goto out_unlock_nocoalesce;
1456 bch_btree_node_write(new_nodes[i], &cl);
1457 bch_keylist_add(&keylist, &new_nodes[i]->key);
1460 for (i = 0; i < nodes; i++)
1461 mutex_unlock(&new_nodes[i]->write_lock);
1465 /* We emptied out this node */
1466 BUG_ON(btree_bset_first(new_nodes[0])->keys);
1467 btree_node_free(new_nodes[0]);
1468 rw_unlock(true, new_nodes[0]);
1469 new_nodes[0] = NULL;
1471 for (i = 0; i < nodes; i++) {
1472 if (__bch_keylist_realloc(&keylist, bkey_u64s(&r[i].b->key)))
1473 goto out_nocoalesce;
1475 make_btree_freeing_key(r[i].b, keylist.top);
1476 bch_keylist_push(&keylist);
1479 bch_btree_insert_node(b, op, &keylist, NULL, NULL);
1480 BUG_ON(!bch_keylist_empty(&keylist));
1482 for (i = 0; i < nodes; i++) {
1483 btree_node_free(r[i].b);
1484 rw_unlock(true, r[i].b);
1486 r[i].b = new_nodes[i];
1489 memmove(r, r + 1, sizeof(r[0]) * (nodes - 1));
1490 r[nodes - 1].b = ERR_PTR(-EINTR);
1492 trace_bcache_btree_gc_coalesce(nodes);
1495 bch_keylist_free(&keylist);
1497 /* Invalidated our iterator */
1500 out_unlock_nocoalesce:
1501 for (i = 0; i < nodes; i++)
1502 mutex_unlock(&new_nodes[i]->write_lock);
1507 while ((k = bch_keylist_pop(&keylist)))
1508 if (!bkey_cmp(k, &ZERO_KEY))
1509 atomic_dec(&b->c->prio_blocked);
1510 bch_keylist_free(&keylist);
1512 for (i = 0; i < nodes; i++)
1513 if (!IS_ERR(new_nodes[i])) {
1514 btree_node_free(new_nodes[i]);
1515 rw_unlock(true, new_nodes[i]);
1520 static int btree_gc_rewrite_node(struct btree *b, struct btree_op *op,
1521 struct btree *replace)
1523 struct keylist keys;
1526 if (btree_check_reserve(b, NULL))
1529 n = btree_node_alloc_replacement(replace, NULL);
1531 /* recheck reserve after allocating replacement node */
1532 if (btree_check_reserve(b, NULL)) {
1538 bch_btree_node_write_sync(n);
1540 bch_keylist_init(&keys);
1541 bch_keylist_add(&keys, &n->key);
1543 make_btree_freeing_key(replace, keys.top);
1544 bch_keylist_push(&keys);
1546 bch_btree_insert_node(b, op, &keys, NULL, NULL);
1547 BUG_ON(!bch_keylist_empty(&keys));
1549 btree_node_free(replace);
1552 /* Invalidated our iterator */
1556 static unsigned int btree_gc_count_keys(struct btree *b)
1559 struct btree_iter iter;
1560 unsigned int ret = 0;
1562 for_each_key_filter(&b->keys, k, &iter, bch_ptr_bad)
1563 ret += bkey_u64s(k);
1568 static size_t btree_gc_min_nodes(struct cache_set *c)
1573 * Since incremental GC would stop 100ms when front
1574 * side I/O comes, so when there are many btree nodes,
1575 * if GC only processes constant (100) nodes each time,
1576 * GC would last a long time, and the front side I/Os
1577 * would run out of the buckets (since no new bucket
1578 * can be allocated during GC), and be blocked again.
1579 * So GC should not process constant nodes, but varied
1580 * nodes according to the number of btree nodes, which
1581 * realized by dividing GC into constant(100) times,
1582 * so when there are many btree nodes, GC can process
1583 * more nodes each time, otherwise, GC will process less
1584 * nodes each time (but no less than MIN_GC_NODES)
1586 min_nodes = c->gc_stats.nodes / MAX_GC_TIMES;
1587 if (min_nodes < MIN_GC_NODES)
1588 min_nodes = MIN_GC_NODES;
1594 static int btree_gc_recurse(struct btree *b, struct btree_op *op,
1595 struct closure *writes, struct gc_stat *gc)
1598 bool should_rewrite;
1600 struct btree_iter iter;
1601 struct gc_merge_info r[GC_MERGE_NODES];
1602 struct gc_merge_info *i, *last = r + ARRAY_SIZE(r) - 1;
1604 bch_btree_iter_init(&b->keys, &iter, &b->c->gc_done);
1606 for (i = r; i < r + ARRAY_SIZE(r); i++)
1607 i->b = ERR_PTR(-EINTR);
1610 k = bch_btree_iter_next_filter(&iter, &b->keys, bch_ptr_bad);
1612 r->b = bch_btree_node_get(b->c, op, k, b->level - 1,
1615 ret = PTR_ERR(r->b);
1619 r->keys = btree_gc_count_keys(r->b);
1621 ret = btree_gc_coalesce(b, op, gc, r);
1629 if (!IS_ERR(last->b)) {
1630 should_rewrite = btree_gc_mark_node(last->b, gc);
1631 if (should_rewrite) {
1632 ret = btree_gc_rewrite_node(b, op, last->b);
1637 if (last->b->level) {
1638 ret = btree_gc_recurse(last->b, op, writes, gc);
1643 bkey_copy_key(&b->c->gc_done, &last->b->key);
1646 * Must flush leaf nodes before gc ends, since replace
1647 * operations aren't journalled
1649 mutex_lock(&last->b->write_lock);
1650 if (btree_node_dirty(last->b))
1651 bch_btree_node_write(last->b, writes);
1652 mutex_unlock(&last->b->write_lock);
1653 rw_unlock(true, last->b);
1656 memmove(r + 1, r, sizeof(r[0]) * (GC_MERGE_NODES - 1));
1659 if (atomic_read(&b->c->search_inflight) &&
1660 gc->nodes >= gc->nodes_pre + btree_gc_min_nodes(b->c)) {
1661 gc->nodes_pre = gc->nodes;
1666 if (need_resched()) {
1672 for (i = r; i < r + ARRAY_SIZE(r); i++)
1673 if (!IS_ERR_OR_NULL(i->b)) {
1674 mutex_lock(&i->b->write_lock);
1675 if (btree_node_dirty(i->b))
1676 bch_btree_node_write(i->b, writes);
1677 mutex_unlock(&i->b->write_lock);
1678 rw_unlock(true, i->b);
1684 static int bch_btree_gc_root(struct btree *b, struct btree_op *op,
1685 struct closure *writes, struct gc_stat *gc)
1687 struct btree *n = NULL;
1689 bool should_rewrite;
1691 should_rewrite = btree_gc_mark_node(b, gc);
1692 if (should_rewrite) {
1693 n = btree_node_alloc_replacement(b, NULL);
1696 bch_btree_node_write_sync(n);
1698 bch_btree_set_root(n);
1706 __bch_btree_mark_key(b->c, b->level + 1, &b->key);
1709 ret = btree_gc_recurse(b, op, writes, gc);
1714 bkey_copy_key(&b->c->gc_done, &b->key);
1719 static void btree_gc_start(struct cache_set *c)
1724 if (!c->gc_mark_valid)
1727 mutex_lock(&c->bucket_lock);
1729 c->gc_mark_valid = 0;
1730 c->gc_done = ZERO_KEY;
1733 for_each_bucket(b, ca) {
1734 b->last_gc = b->gen;
1735 if (!atomic_read(&b->pin)) {
1737 SET_GC_SECTORS_USED(b, 0);
1741 mutex_unlock(&c->bucket_lock);
1744 static void bch_btree_gc_finish(struct cache_set *c)
1751 mutex_lock(&c->bucket_lock);
1754 c->gc_mark_valid = 1;
1757 for (i = 0; i < KEY_PTRS(&c->uuid_bucket); i++)
1758 SET_GC_MARK(PTR_BUCKET(c, &c->uuid_bucket, i),
1761 /* don't reclaim buckets to which writeback keys point */
1763 for (i = 0; i < c->devices_max_used; i++) {
1764 struct bcache_device *d = c->devices[i];
1765 struct cached_dev *dc;
1766 struct keybuf_key *w, *n;
1768 if (!d || UUID_FLASH_ONLY(&c->uuids[i]))
1770 dc = container_of(d, struct cached_dev, disk);
1772 spin_lock(&dc->writeback_keys.lock);
1773 rbtree_postorder_for_each_entry_safe(w, n,
1774 &dc->writeback_keys.keys, node)
1775 for (j = 0; j < KEY_PTRS(&w->key); j++)
1776 SET_GC_MARK(PTR_BUCKET(c, &w->key, j),
1778 spin_unlock(&dc->writeback_keys.lock);
1782 c->avail_nbuckets = 0;
1785 ca->invalidate_needs_gc = 0;
1787 for (k = ca->sb.d; k < ca->sb.d + ca->sb.keys; k++)
1788 SET_GC_MARK(ca->buckets + *k, GC_MARK_METADATA);
1790 for (k = ca->prio_buckets;
1791 k < ca->prio_buckets + prio_buckets(ca) * 2; k++)
1792 SET_GC_MARK(ca->buckets + *k, GC_MARK_METADATA);
1794 for_each_bucket(b, ca) {
1795 c->need_gc = max(c->need_gc, bucket_gc_gen(b));
1797 if (atomic_read(&b->pin))
1800 BUG_ON(!GC_MARK(b) && GC_SECTORS_USED(b));
1802 if (!GC_MARK(b) || GC_MARK(b) == GC_MARK_RECLAIMABLE)
1803 c->avail_nbuckets++;
1806 mutex_unlock(&c->bucket_lock);
1809 static void bch_btree_gc(struct cache_set *c)
1812 struct gc_stat stats;
1813 struct closure writes;
1815 uint64_t start_time = local_clock();
1817 trace_bcache_gc_start(c);
1819 memset(&stats, 0, sizeof(struct gc_stat));
1820 closure_init_stack(&writes);
1821 bch_btree_op_init(&op, SHRT_MAX);
1825 /* if CACHE_SET_IO_DISABLE set, gc thread should stop too */
1827 ret = bcache_btree_root(gc_root, c, &op, &writes, &stats);
1828 closure_sync(&writes);
1832 schedule_timeout_interruptible(msecs_to_jiffies
1835 pr_warn("gc failed!\n");
1836 } while (ret && !test_bit(CACHE_SET_IO_DISABLE, &c->flags));
1838 bch_btree_gc_finish(c);
1839 wake_up_allocators(c);
1841 bch_time_stats_update(&c->btree_gc_time, start_time);
1843 stats.key_bytes *= sizeof(uint64_t);
1845 bch_update_bucket_in_use(c, &stats);
1846 memcpy(&c->gc_stats, &stats, sizeof(struct gc_stat));
1848 trace_bcache_gc_end(c);
1853 static bool gc_should_run(struct cache_set *c)
1855 struct cache *ca = c->cache;
1857 if (ca->invalidate_needs_gc)
1860 if (atomic_read(&c->sectors_to_gc) < 0)
1866 static int bch_gc_thread(void *arg)
1868 struct cache_set *c = arg;
1871 wait_event_interruptible(c->gc_wait,
1872 kthread_should_stop() ||
1873 test_bit(CACHE_SET_IO_DISABLE, &c->flags) ||
1876 if (kthread_should_stop() ||
1877 test_bit(CACHE_SET_IO_DISABLE, &c->flags))
1884 wait_for_kthread_stop();
1888 int bch_gc_thread_start(struct cache_set *c)
1890 c->gc_thread = kthread_run(bch_gc_thread, c, "bcache_gc");
1891 return PTR_ERR_OR_ZERO(c->gc_thread);
1894 /* Initial partial gc */
1896 static int bch_btree_check_recurse(struct btree *b, struct btree_op *op)
1899 struct bkey *k, *p = NULL;
1900 struct btree_iter iter;
1902 for_each_key_filter(&b->keys, k, &iter, bch_ptr_invalid)
1903 bch_initial_mark_key(b->c, b->level, k);
1905 bch_initial_mark_key(b->c, b->level + 1, &b->key);
1908 bch_btree_iter_init(&b->keys, &iter, NULL);
1911 k = bch_btree_iter_next_filter(&iter, &b->keys,
1914 btree_node_prefetch(b, k);
1916 * initiallize c->gc_stats.nodes
1917 * for incremental GC
1919 b->c->gc_stats.nodes++;
1923 ret = bcache_btree(check_recurse, p, b, op);
1926 } while (p && !ret);
1933 static int bch_btree_check_thread(void *arg)
1936 struct btree_check_info *info = arg;
1937 struct btree_check_state *check_state = info->state;
1938 struct cache_set *c = check_state->c;
1939 struct btree_iter iter;
1941 int cur_idx, prev_idx, skip_nr;
1944 cur_idx = prev_idx = 0;
1947 /* root node keys are checked before thread created */
1948 bch_btree_iter_init(&c->root->keys, &iter, NULL);
1949 k = bch_btree_iter_next_filter(&iter, &c->root->keys, bch_ptr_bad);
1955 * Fetch a root node key index, skip the keys which
1956 * should be fetched by other threads, then check the
1957 * sub-tree indexed by the fetched key.
1959 spin_lock(&check_state->idx_lock);
1960 cur_idx = check_state->key_idx;
1961 check_state->key_idx++;
1962 spin_unlock(&check_state->idx_lock);
1964 skip_nr = cur_idx - prev_idx;
1967 k = bch_btree_iter_next_filter(&iter,
1974 * No more keys to check in root node,
1975 * current checking threads are enough,
1976 * stop creating more.
1978 atomic_set(&check_state->enough, 1);
1979 /* Update check_state->enough earlier */
1980 smp_mb__after_atomic();
1990 btree_node_prefetch(c->root, p);
1991 c->gc_stats.nodes++;
1992 bch_btree_op_init(&op, 0);
1993 ret = bcache_btree(check_recurse, p, c->root, &op);
1995 * The op may be added to cache_set's btree_cache_wait
1996 * in mca_cannibalize(), must ensure it is removed from
1997 * the list and release btree_cache_alloc_lock before
1999 * Otherwise, the btree_cache_wait will be damaged.
2001 bch_cannibalize_unlock(c);
2002 finish_wait(&c->btree_cache_wait, &(&op)->wait);
2013 /* update check_state->started among all CPUs */
2014 smp_mb__before_atomic();
2015 if (atomic_dec_and_test(&check_state->started))
2016 wake_up(&check_state->wait);
2023 static int bch_btree_chkthread_nr(void)
2025 int n = num_online_cpus()/2;
2029 else if (n > BCH_BTR_CHKTHREAD_MAX)
2030 n = BCH_BTR_CHKTHREAD_MAX;
2035 int bch_btree_check(struct cache_set *c)
2039 struct bkey *k = NULL;
2040 struct btree_iter iter;
2041 struct btree_check_state check_state;
2043 /* check and mark root node keys */
2044 for_each_key_filter(&c->root->keys, k, &iter, bch_ptr_invalid)
2045 bch_initial_mark_key(c, c->root->level, k);
2047 bch_initial_mark_key(c, c->root->level + 1, &c->root->key);
2049 if (c->root->level == 0)
2052 memset(&check_state, 0, sizeof(struct btree_check_state));
2054 check_state.total_threads = bch_btree_chkthread_nr();
2055 check_state.key_idx = 0;
2056 spin_lock_init(&check_state.idx_lock);
2057 atomic_set(&check_state.started, 0);
2058 atomic_set(&check_state.enough, 0);
2059 init_waitqueue_head(&check_state.wait);
2061 rw_lock(0, c->root, c->root->level);
2063 * Run multiple threads to check btree nodes in parallel,
2064 * if check_state.enough is non-zero, it means current
2065 * running check threads are enough, unncessary to create
2068 for (i = 0; i < check_state.total_threads; i++) {
2069 /* fetch latest check_state.enough earlier */
2070 smp_mb__before_atomic();
2071 if (atomic_read(&check_state.enough))
2074 check_state.infos[i].result = 0;
2075 check_state.infos[i].state = &check_state;
2077 check_state.infos[i].thread =
2078 kthread_run(bch_btree_check_thread,
2079 &check_state.infos[i],
2080 "bch_btrchk[%d]", i);
2081 if (IS_ERR(check_state.infos[i].thread)) {
2082 pr_err("fails to run thread bch_btrchk[%d]\n", i);
2083 for (--i; i >= 0; i--)
2084 kthread_stop(check_state.infos[i].thread);
2088 atomic_inc(&check_state.started);
2092 * Must wait for all threads to stop.
2094 wait_event(check_state.wait, atomic_read(&check_state.started) == 0);
2096 for (i = 0; i < check_state.total_threads; i++) {
2097 if (check_state.infos[i].result) {
2098 ret = check_state.infos[i].result;
2104 rw_unlock(0, c->root);
2108 void bch_initial_gc_finish(struct cache_set *c)
2110 struct cache *ca = c->cache;
2113 bch_btree_gc_finish(c);
2115 mutex_lock(&c->bucket_lock);
2118 * We need to put some unused buckets directly on the prio freelist in
2119 * order to get the allocator thread started - it needs freed buckets in
2120 * order to rewrite the prios and gens, and it needs to rewrite prios
2121 * and gens in order to free buckets.
2123 * This is only safe for buckets that have no live data in them, which
2124 * there should always be some of.
2126 for_each_bucket(b, ca) {
2127 if (fifo_full(&ca->free[RESERVE_PRIO]) &&
2128 fifo_full(&ca->free[RESERVE_BTREE]))
2131 if (bch_can_invalidate_bucket(ca, b) &&
2133 __bch_invalidate_one_bucket(ca, b);
2134 if (!fifo_push(&ca->free[RESERVE_PRIO],
2136 fifo_push(&ca->free[RESERVE_BTREE],
2141 mutex_unlock(&c->bucket_lock);
2144 /* Btree insertion */
2146 static bool btree_insert_key(struct btree *b, struct bkey *k,
2147 struct bkey *replace_key)
2149 unsigned int status;
2151 BUG_ON(bkey_cmp(k, &b->key) > 0);
2153 status = bch_btree_insert_key(&b->keys, k, replace_key);
2154 if (status != BTREE_INSERT_STATUS_NO_INSERT) {
2155 bch_check_keys(&b->keys, "%u for %s", status,
2156 replace_key ? "replace" : "insert");
2158 trace_bcache_btree_insert_key(b, k, replace_key != NULL,
2165 static size_t insert_u64s_remaining(struct btree *b)
2167 long ret = bch_btree_keys_u64s_remaining(&b->keys);
2170 * Might land in the middle of an existing extent and have to split it
2172 if (b->keys.ops->is_extents)
2173 ret -= KEY_MAX_U64S;
2175 return max(ret, 0L);
2178 static bool bch_btree_insert_keys(struct btree *b, struct btree_op *op,
2179 struct keylist *insert_keys,
2180 struct bkey *replace_key)
2183 int oldsize = bch_count_data(&b->keys);
2185 while (!bch_keylist_empty(insert_keys)) {
2186 struct bkey *k = insert_keys->keys;
2188 if (bkey_u64s(k) > insert_u64s_remaining(b))
2191 if (bkey_cmp(k, &b->key) <= 0) {
2195 ret |= btree_insert_key(b, k, replace_key);
2196 bch_keylist_pop_front(insert_keys);
2197 } else if (bkey_cmp(&START_KEY(k), &b->key) < 0) {
2198 BKEY_PADDED(key) temp;
2199 bkey_copy(&temp.key, insert_keys->keys);
2201 bch_cut_back(&b->key, &temp.key);
2202 bch_cut_front(&b->key, insert_keys->keys);
2204 ret |= btree_insert_key(b, &temp.key, replace_key);
2212 op->insert_collision = true;
2214 BUG_ON(!bch_keylist_empty(insert_keys) && b->level);
2216 BUG_ON(bch_count_data(&b->keys) < oldsize);
2220 static int btree_split(struct btree *b, struct btree_op *op,
2221 struct keylist *insert_keys,
2222 struct bkey *replace_key)
2225 struct btree *n1, *n2 = NULL, *n3 = NULL;
2226 uint64_t start_time = local_clock();
2228 struct keylist parent_keys;
2230 closure_init_stack(&cl);
2231 bch_keylist_init(&parent_keys);
2233 if (btree_check_reserve(b, op)) {
2237 WARN(1, "insufficient reserve for split\n");
2240 n1 = btree_node_alloc_replacement(b, op);
2244 split = set_blocks(btree_bset_first(n1),
2245 block_bytes(n1->c->cache)) > (btree_blocks(b) * 4) / 5;
2248 unsigned int keys = 0;
2250 trace_bcache_btree_node_split(b, btree_bset_first(n1)->keys);
2252 n2 = bch_btree_node_alloc(b->c, op, b->level, b->parent);
2257 n3 = bch_btree_node_alloc(b->c, op, b->level + 1, NULL);
2262 mutex_lock(&n1->write_lock);
2263 mutex_lock(&n2->write_lock);
2265 bch_btree_insert_keys(n1, op, insert_keys, replace_key);
2268 * Has to be a linear search because we don't have an auxiliary
2272 while (keys < (btree_bset_first(n1)->keys * 3) / 5)
2273 keys += bkey_u64s(bset_bkey_idx(btree_bset_first(n1),
2276 bkey_copy_key(&n1->key,
2277 bset_bkey_idx(btree_bset_first(n1), keys));
2278 keys += bkey_u64s(bset_bkey_idx(btree_bset_first(n1), keys));
2280 btree_bset_first(n2)->keys = btree_bset_first(n1)->keys - keys;
2281 btree_bset_first(n1)->keys = keys;
2283 memcpy(btree_bset_first(n2)->start,
2284 bset_bkey_last(btree_bset_first(n1)),
2285 btree_bset_first(n2)->keys * sizeof(uint64_t));
2287 bkey_copy_key(&n2->key, &b->key);
2289 bch_keylist_add(&parent_keys, &n2->key);
2290 bch_btree_node_write(n2, &cl);
2291 mutex_unlock(&n2->write_lock);
2292 rw_unlock(true, n2);
2294 trace_bcache_btree_node_compact(b, btree_bset_first(n1)->keys);
2296 mutex_lock(&n1->write_lock);
2297 bch_btree_insert_keys(n1, op, insert_keys, replace_key);
2300 bch_keylist_add(&parent_keys, &n1->key);
2301 bch_btree_node_write(n1, &cl);
2302 mutex_unlock(&n1->write_lock);
2305 /* Depth increases, make a new root */
2306 mutex_lock(&n3->write_lock);
2307 bkey_copy_key(&n3->key, &MAX_KEY);
2308 bch_btree_insert_keys(n3, op, &parent_keys, NULL);
2309 bch_btree_node_write(n3, &cl);
2310 mutex_unlock(&n3->write_lock);
2313 bch_btree_set_root(n3);
2314 rw_unlock(true, n3);
2315 } else if (!b->parent) {
2316 /* Root filled up but didn't need to be split */
2318 bch_btree_set_root(n1);
2320 /* Split a non root node */
2322 make_btree_freeing_key(b, parent_keys.top);
2323 bch_keylist_push(&parent_keys);
2325 bch_btree_insert_node(b->parent, op, &parent_keys, NULL, NULL);
2326 BUG_ON(!bch_keylist_empty(&parent_keys));
2330 rw_unlock(true, n1);
2332 bch_time_stats_update(&b->c->btree_split_time, start_time);
2336 bkey_put(b->c, &n2->key);
2337 btree_node_free(n2);
2338 rw_unlock(true, n2);
2340 bkey_put(b->c, &n1->key);
2341 btree_node_free(n1);
2342 rw_unlock(true, n1);
2344 WARN(1, "bcache: btree split failed (level %u)", b->level);
2346 if (n3 == ERR_PTR(-EAGAIN) ||
2347 n2 == ERR_PTR(-EAGAIN) ||
2348 n1 == ERR_PTR(-EAGAIN))
2354 static int bch_btree_insert_node(struct btree *b, struct btree_op *op,
2355 struct keylist *insert_keys,
2356 atomic_t *journal_ref,
2357 struct bkey *replace_key)
2361 BUG_ON(b->level && replace_key);
2363 closure_init_stack(&cl);
2365 mutex_lock(&b->write_lock);
2367 if (write_block(b) != btree_bset_last(b) &&
2368 b->keys.last_set_unwritten)
2369 bch_btree_init_next(b); /* just wrote a set */
2371 if (bch_keylist_nkeys(insert_keys) > insert_u64s_remaining(b)) {
2372 mutex_unlock(&b->write_lock);
2376 BUG_ON(write_block(b) != btree_bset_last(b));
2378 if (bch_btree_insert_keys(b, op, insert_keys, replace_key)) {
2380 bch_btree_leaf_dirty(b, journal_ref);
2382 bch_btree_node_write(b, &cl);
2385 mutex_unlock(&b->write_lock);
2387 /* wait for btree node write if necessary, after unlock */
2392 if (current->bio_list) {
2393 op->lock = b->c->root->level + 1;
2395 } else if (op->lock <= b->c->root->level) {
2396 op->lock = b->c->root->level + 1;
2399 /* Invalidated all iterators */
2400 int ret = btree_split(b, op, insert_keys, replace_key);
2402 if (bch_keylist_empty(insert_keys))
2410 int bch_btree_insert_check_key(struct btree *b, struct btree_op *op,
2411 struct bkey *check_key)
2414 uint64_t btree_ptr = b->key.ptr[0];
2415 unsigned long seq = b->seq;
2416 struct keylist insert;
2417 bool upgrade = op->lock == -1;
2419 bch_keylist_init(&insert);
2422 rw_unlock(false, b);
2423 rw_lock(true, b, b->level);
2425 if (b->key.ptr[0] != btree_ptr ||
2426 b->seq != seq + 1) {
2427 op->lock = b->level;
2432 SET_KEY_PTRS(check_key, 1);
2433 get_random_bytes(&check_key->ptr[0], sizeof(uint64_t));
2435 SET_PTR_DEV(check_key, 0, PTR_CHECK_DEV);
2437 bch_keylist_add(&insert, check_key);
2439 ret = bch_btree_insert_node(b, op, &insert, NULL, NULL);
2441 BUG_ON(!ret && !bch_keylist_empty(&insert));
2444 downgrade_write(&b->lock);
2448 struct btree_insert_op {
2450 struct keylist *keys;
2451 atomic_t *journal_ref;
2452 struct bkey *replace_key;
2455 static int btree_insert_fn(struct btree_op *b_op, struct btree *b)
2457 struct btree_insert_op *op = container_of(b_op,
2458 struct btree_insert_op, op);
2460 int ret = bch_btree_insert_node(b, &op->op, op->keys,
2461 op->journal_ref, op->replace_key);
2462 if (ret && !bch_keylist_empty(op->keys))
2468 int bch_btree_insert(struct cache_set *c, struct keylist *keys,
2469 atomic_t *journal_ref, struct bkey *replace_key)
2471 struct btree_insert_op op;
2474 BUG_ON(current->bio_list);
2475 BUG_ON(bch_keylist_empty(keys));
2477 bch_btree_op_init(&op.op, 0);
2479 op.journal_ref = journal_ref;
2480 op.replace_key = replace_key;
2482 while (!ret && !bch_keylist_empty(keys)) {
2484 ret = bch_btree_map_leaf_nodes(&op.op, c,
2485 &START_KEY(keys->keys),
2492 pr_err("error %i\n", ret);
2494 while ((k = bch_keylist_pop(keys)))
2496 } else if (op.op.insert_collision)
2502 void bch_btree_set_root(struct btree *b)
2507 closure_init_stack(&cl);
2509 trace_bcache_btree_set_root(b);
2511 BUG_ON(!b->written);
2513 for (i = 0; i < KEY_PTRS(&b->key); i++)
2514 BUG_ON(PTR_BUCKET(b->c, &b->key, i)->prio != BTREE_PRIO);
2516 mutex_lock(&b->c->bucket_lock);
2517 list_del_init(&b->list);
2518 mutex_unlock(&b->c->bucket_lock);
2522 bch_journal_meta(b->c, &cl);
2526 /* Map across nodes or keys */
2528 static int bch_btree_map_nodes_recurse(struct btree *b, struct btree_op *op,
2530 btree_map_nodes_fn *fn, int flags)
2532 int ret = MAP_CONTINUE;
2536 struct btree_iter iter;
2538 bch_btree_iter_init(&b->keys, &iter, from);
2540 while ((k = bch_btree_iter_next_filter(&iter, &b->keys,
2542 ret = bcache_btree(map_nodes_recurse, k, b,
2543 op, from, fn, flags);
2546 if (ret != MAP_CONTINUE)
2551 if (!b->level || flags == MAP_ALL_NODES)
2557 int __bch_btree_map_nodes(struct btree_op *op, struct cache_set *c,
2558 struct bkey *from, btree_map_nodes_fn *fn, int flags)
2560 return bcache_btree_root(map_nodes_recurse, c, op, from, fn, flags);
2563 int bch_btree_map_keys_recurse(struct btree *b, struct btree_op *op,
2564 struct bkey *from, btree_map_keys_fn *fn,
2567 int ret = MAP_CONTINUE;
2569 struct btree_iter iter;
2571 bch_btree_iter_init(&b->keys, &iter, from);
2573 while ((k = bch_btree_iter_next_filter(&iter, &b->keys, bch_ptr_bad))) {
2576 : bcache_btree(map_keys_recurse, k,
2577 b, op, from, fn, flags);
2580 if (ret != MAP_CONTINUE)
2584 if (!b->level && (flags & MAP_END_KEY))
2585 ret = fn(op, b, &KEY(KEY_INODE(&b->key),
2586 KEY_OFFSET(&b->key), 0));
2591 int bch_btree_map_keys(struct btree_op *op, struct cache_set *c,
2592 struct bkey *from, btree_map_keys_fn *fn, int flags)
2594 return bcache_btree_root(map_keys_recurse, c, op, from, fn, flags);
2599 static inline int keybuf_cmp(struct keybuf_key *l, struct keybuf_key *r)
2601 /* Overlapping keys compare equal */
2602 if (bkey_cmp(&l->key, &START_KEY(&r->key)) <= 0)
2604 if (bkey_cmp(&START_KEY(&l->key), &r->key) >= 0)
2609 static inline int keybuf_nonoverlapping_cmp(struct keybuf_key *l,
2610 struct keybuf_key *r)
2612 return clamp_t(int64_t, bkey_cmp(&l->key, &r->key), -1, 1);
2617 unsigned int nr_found;
2620 keybuf_pred_fn *pred;
2623 static int refill_keybuf_fn(struct btree_op *op, struct btree *b,
2626 struct refill *refill = container_of(op, struct refill, op);
2627 struct keybuf *buf = refill->buf;
2628 int ret = MAP_CONTINUE;
2630 if (bkey_cmp(k, refill->end) > 0) {
2635 if (!KEY_SIZE(k)) /* end key */
2638 if (refill->pred(buf, k)) {
2639 struct keybuf_key *w;
2641 spin_lock(&buf->lock);
2643 w = array_alloc(&buf->freelist);
2645 spin_unlock(&buf->lock);
2650 bkey_copy(&w->key, k);
2652 if (RB_INSERT(&buf->keys, w, node, keybuf_cmp))
2653 array_free(&buf->freelist, w);
2657 if (array_freelist_empty(&buf->freelist))
2660 spin_unlock(&buf->lock);
2663 buf->last_scanned = *k;
2667 void bch_refill_keybuf(struct cache_set *c, struct keybuf *buf,
2668 struct bkey *end, keybuf_pred_fn *pred)
2670 struct bkey start = buf->last_scanned;
2671 struct refill refill;
2675 bch_btree_op_init(&refill.op, -1);
2676 refill.nr_found = 0;
2681 bch_btree_map_keys(&refill.op, c, &buf->last_scanned,
2682 refill_keybuf_fn, MAP_END_KEY);
2684 trace_bcache_keyscan(refill.nr_found,
2685 KEY_INODE(&start), KEY_OFFSET(&start),
2686 KEY_INODE(&buf->last_scanned),
2687 KEY_OFFSET(&buf->last_scanned));
2689 spin_lock(&buf->lock);
2691 if (!RB_EMPTY_ROOT(&buf->keys)) {
2692 struct keybuf_key *w;
2694 w = RB_FIRST(&buf->keys, struct keybuf_key, node);
2695 buf->start = START_KEY(&w->key);
2697 w = RB_LAST(&buf->keys, struct keybuf_key, node);
2700 buf->start = MAX_KEY;
2704 spin_unlock(&buf->lock);
2707 static void __bch_keybuf_del(struct keybuf *buf, struct keybuf_key *w)
2709 rb_erase(&w->node, &buf->keys);
2710 array_free(&buf->freelist, w);
2713 void bch_keybuf_del(struct keybuf *buf, struct keybuf_key *w)
2715 spin_lock(&buf->lock);
2716 __bch_keybuf_del(buf, w);
2717 spin_unlock(&buf->lock);
2720 bool bch_keybuf_check_overlapping(struct keybuf *buf, struct bkey *start,
2724 struct keybuf_key *p, *w, s;
2728 if (bkey_cmp(end, &buf->start) <= 0 ||
2729 bkey_cmp(start, &buf->end) >= 0)
2732 spin_lock(&buf->lock);
2733 w = RB_GREATER(&buf->keys, s, node, keybuf_nonoverlapping_cmp);
2735 while (w && bkey_cmp(&START_KEY(&w->key), end) < 0) {
2737 w = RB_NEXT(w, node);
2742 __bch_keybuf_del(buf, p);
2745 spin_unlock(&buf->lock);
2749 struct keybuf_key *bch_keybuf_next(struct keybuf *buf)
2751 struct keybuf_key *w;
2753 spin_lock(&buf->lock);
2755 w = RB_FIRST(&buf->keys, struct keybuf_key, node);
2757 while (w && w->private)
2758 w = RB_NEXT(w, node);
2761 w->private = ERR_PTR(-EINTR);
2763 spin_unlock(&buf->lock);
2767 struct keybuf_key *bch_keybuf_next_rescan(struct cache_set *c,
2770 keybuf_pred_fn *pred)
2772 struct keybuf_key *ret;
2775 ret = bch_keybuf_next(buf);
2779 if (bkey_cmp(&buf->last_scanned, end) >= 0) {
2780 pr_debug("scan finished\n");
2784 bch_refill_keybuf(c, buf, end, pred);
2790 void bch_keybuf_init(struct keybuf *buf)
2792 buf->last_scanned = MAX_KEY;
2793 buf->keys = RB_ROOT;
2795 spin_lock_init(&buf->lock);
2796 array_allocator_init(&buf->freelist);
2799 void bch_btree_exit(void)
2802 destroy_workqueue(btree_io_wq);
2805 int __init bch_btree_init(void)
2807 btree_io_wq = alloc_workqueue("bch_btree_io", WQ_MEM_RECLAIM, 0);