2 * Copyright (C) 2010 Kent Overstreet <kent.overstreet@gmail.com>
4 * Uses a block device as cache for other block devices; optimized for SSDs.
5 * All allocation is done in buckets, which should match the erase block size
8 * Buckets containing cached data are kept on a heap sorted by priority;
9 * bucket priority is increased on cache hit, and periodically all the buckets
10 * on the heap have their priority scaled down. This currently is just used as
11 * an LRU but in the future should allow for more intelligent heuristics.
13 * Buckets have an 8 bit counter; freeing is accomplished by incrementing the
14 * counter. Garbage collection is used to remove stale pointers.
16 * Indexing is done via a btree; nodes are not necessarily fully sorted, rather
17 * as keys are inserted we only sort the pages that have not yet been written.
18 * When garbage collection is run, we resort the entire node.
20 * All configuration is done via sysfs; see Documentation/bcache.txt.
28 #include <linux/slab.h>
29 #include <linux/bitops.h>
30 #include <linux/hash.h>
31 #include <linux/prefetch.h>
32 #include <linux/random.h>
33 #include <linux/rcupdate.h>
34 #include <trace/events/bcache.h>
38 * register_bcache: Return errors out to userspace correctly
40 * Writeback: don't undirty key until after a cache flush
42 * Create an iterator for key pointers
44 * On btree write error, mark bucket such that it won't be freed from the cache
47 * Check for bad keys in replay
49 * Refcount journal entries in journal_replay
52 * Finish incremental gc
53 * Gc should free old UUIDs, data for invalid UUIDs
55 * Provide a way to list backing device UUIDs we have data cached for, and
56 * probably how long it's been since we've seen them, and a way to invalidate
57 * dirty data for devices that will never be attached again
59 * Keep 1 min/5 min/15 min statistics of how busy a block device has been, so
60 * that based on that and how much dirty data we have we can keep writeback
63 * Add a tracepoint or somesuch to watch for writeback starvation
65 * When btree depth > 1 and splitting an interior node, we have to make sure
66 * alloc_bucket() cannot fail. This should be true but is not completely
69 * Make sure all allocations get charged to the root cgroup
73 * If data write is less than hard sector size of ssd, round up offset in open
74 * bucket to the next whole sector
76 * Also lookup by cgroup in get_open_bucket()
78 * Superblock needs to be fleshed out for multiple cache devices
80 * Add a sysfs tunable for the number of writeback IOs in flight
82 * Add a sysfs tunable for the number of open data buckets
84 * IO tracking: Can we track when one process is doing io on behalf of another?
85 * IO tracking: Don't use just an average, weigh more recent stuff higher
87 * Test module load/unload
90 static const char * const op_types[] = {
94 static const char *op_type(struct btree_op *op)
96 return op_types[op->type];
99 #define MAX_NEED_GC 64
100 #define MAX_SAVE_PRIO 72
102 #define PTR_DIRTY_BIT (((uint64_t) 1 << 36))
104 #define PTR_HASH(c, k) \
105 (((k)->ptr[0] >> c->bucket_bits) | PTR_GEN(k, 0))
107 struct workqueue_struct *bch_gc_wq;
108 static struct workqueue_struct *btree_io_wq;
110 void bch_btree_op_init_stack(struct btree_op *op)
112 memset(op, 0, sizeof(struct btree_op));
113 closure_init_stack(&op->cl);
115 bch_keylist_init(&op->keys);
118 /* Btree key manipulation */
120 static void bkey_put(struct cache_set *c, struct bkey *k, int level)
122 if ((level && KEY_OFFSET(k)) || !level)
128 static uint64_t btree_csum_set(struct btree *b, struct bset *i)
130 uint64_t crc = b->key.ptr[0];
131 void *data = (void *) i + 8, *end = end(i);
133 crc = bch_crc64_update(crc, data, end - data);
134 return crc ^ 0xffffffffffffffffULL;
137 static void btree_bio_endio(struct bio *bio, int error,
138 struct batch_complete *batch)
140 struct closure *cl = bio->bi_private;
141 struct btree *b = container_of(cl, struct btree, io.cl);
144 set_btree_node_io_error(b);
146 bch_bbio_count_io_errors(b->c, bio, error, (bio->bi_rw & WRITE)
147 ? "writing btree" : "reading btree");
151 static void btree_bio_init(struct btree *b)
154 b->bio = bch_bbio_alloc(b->c);
156 b->bio->bi_end_io = btree_bio_endio;
157 b->bio->bi_private = &b->io.cl;
160 void bch_btree_read_done(struct closure *cl)
162 struct btree *b = container_of(cl, struct btree, io.cl);
163 struct bset *i = b->sets[0].data;
164 struct btree_iter *iter = b->c->fill_iter;
165 const char *err = "bad btree header";
166 BUG_ON(b->nsets || b->written);
168 bch_bbio_free(b->bio, b->c);
171 mutex_lock(&b->c->fill_lock);
174 if (btree_node_io_error(b) ||
179 b->written < btree_blocks(b) && i->seq == b->sets[0].data->seq;
180 i = write_block(b)) {
181 err = "unsupported bset version";
182 if (i->version > BCACHE_BSET_VERSION)
185 err = "bad btree header";
186 if (b->written + set_blocks(i, b->c) > btree_blocks(b))
190 if (i->magic != bset_magic(b->c))
193 err = "bad checksum";
194 switch (i->version) {
196 if (i->csum != csum_set(i))
199 case BCACHE_BSET_VERSION:
200 if (i->csum != btree_csum_set(b, i))
206 if (i != b->sets[0].data && !i->keys)
209 bch_btree_iter_push(iter, i->start, end(i));
211 b->written += set_blocks(i, b->c);
214 err = "corrupted btree";
215 for (i = write_block(b);
216 index(i, b) < btree_blocks(b);
217 i = ((void *) i) + block_bytes(b->c))
218 if (i->seq == b->sets[0].data->seq)
221 bch_btree_sort_and_fix_extents(b, iter);
224 err = "short btree key";
225 if (b->sets[0].size &&
226 bkey_cmp(&b->key, &b->sets[0].end) < 0)
229 if (b->written < btree_blocks(b))
230 bch_bset_init_next(b);
233 mutex_unlock(&b->c->fill_lock);
235 spin_lock(&b->c->btree_read_time_lock);
236 bch_time_stats_update(&b->c->btree_read_time, b->io_start_time);
237 spin_unlock(&b->c->btree_read_time_lock);
239 smp_wmb(); /* read_done is our write lock */
240 set_btree_node_read_done(b);
244 set_btree_node_io_error(b);
245 bch_cache_set_error(b->c, "%s at bucket %zu, block %zu, %u keys",
246 err, PTR_BUCKET_NR(b->c, &b->key, 0),
247 index(i, b), i->keys);
251 void bch_btree_read(struct btree *b)
253 BUG_ON(b->nsets || b->written);
255 if (!closure_trylock(&b->io.cl, &b->c->cl))
258 b->io_start_time = local_clock();
261 b->bio->bi_rw = REQ_META|READ_SYNC;
262 b->bio->bi_size = KEY_SIZE(&b->key) << 9;
264 bch_bio_map(b->bio, b->sets[0].data);
266 pr_debug("%s", pbtree(b));
267 trace_bcache_btree_read(b->bio);
268 bch_submit_bbio(b->bio, b->c, &b->key, 0);
270 continue_at(&b->io.cl, bch_btree_read_done, system_wq);
273 static void btree_complete_write(struct btree *b, struct btree_write *w)
275 if (w->prio_blocked &&
276 !atomic_sub_return(w->prio_blocked, &b->c->prio_blocked))
277 wake_up(&b->c->alloc_wait);
280 atomic_dec_bug(w->journal);
281 __closure_wake_up(&b->c->journal.wait);
285 closure_put(w->owner);
292 static void __btree_write_done(struct closure *cl)
294 struct btree *b = container_of(cl, struct btree, io.cl);
295 struct btree_write *w = btree_prev_write(b);
297 bch_bbio_free(b->bio, b->c);
299 btree_complete_write(b, w);
301 if (btree_node_dirty(b))
302 queue_delayed_work(btree_io_wq, &b->work,
303 msecs_to_jiffies(30000));
308 static void btree_write_done(struct closure *cl)
310 struct btree *b = container_of(cl, struct btree, io.cl);
314 __bio_for_each_segment(bv, b->bio, n, 0)
315 __free_page(bv->bv_page);
317 __btree_write_done(cl);
320 static void do_btree_write(struct btree *b)
322 struct closure *cl = &b->io.cl;
323 struct bset *i = b->sets[b->nsets].data;
326 i->version = BCACHE_BSET_VERSION;
327 i->csum = btree_csum_set(b, i);
330 b->bio->bi_rw = REQ_META|WRITE_SYNC;
331 b->bio->bi_size = set_blocks(i, b->c) * block_bytes(b->c);
332 bch_bio_map(b->bio, i);
334 bkey_copy(&k.key, &b->key);
335 SET_PTR_OFFSET(&k.key, 0, PTR_OFFSET(&k.key, 0) + bset_offset(b, i));
337 if (!bch_bio_alloc_pages(b->bio, GFP_NOIO)) {
340 void *base = (void *) ((unsigned long) i & ~(PAGE_SIZE - 1));
342 bio_for_each_segment(bv, b->bio, j)
343 memcpy(page_address(bv->bv_page),
344 base + j * PAGE_SIZE, PAGE_SIZE);
346 trace_bcache_btree_write(b->bio);
347 bch_submit_bbio(b->bio, b->c, &k.key, 0);
349 continue_at(cl, btree_write_done, NULL);
352 bch_bio_map(b->bio, i);
354 trace_bcache_btree_write(b->bio);
355 bch_submit_bbio(b->bio, b->c, &k.key, 0);
358 __btree_write_done(cl);
362 static void __btree_write(struct btree *b)
364 struct bset *i = b->sets[b->nsets].data;
366 BUG_ON(current->bio_list);
368 closure_lock(&b->io, &b->c->cl);
369 cancel_delayed_work(&b->work);
371 clear_bit(BTREE_NODE_dirty, &b->flags);
372 change_bit(BTREE_NODE_write_idx, &b->flags);
374 bch_check_key_order(b, i);
375 BUG_ON(b->written && !i->keys);
379 pr_debug("%s block %i keys %i", pbtree(b), b->written, i->keys);
381 b->written += set_blocks(i, b->c);
382 atomic_long_add(set_blocks(i, b->c) * b->c->sb.block_size,
383 &PTR_CACHE(b->c, &b->key, 0)->btree_sectors_written);
385 bch_btree_sort_lazy(b);
387 if (b->written < btree_blocks(b))
388 bch_bset_init_next(b);
391 static void btree_write_work(struct work_struct *w)
393 struct btree *b = container_of(to_delayed_work(w), struct btree, work);
395 down_write(&b->lock);
397 if (btree_node_dirty(b))
402 void bch_btree_write(struct btree *b, bool now, struct btree_op *op)
404 struct bset *i = b->sets[b->nsets].data;
405 struct btree_write *w = btree_current_write(b);
408 (b->written >= btree_blocks(b) ||
409 i->seq != b->sets[0].data->seq ||
412 if (!btree_node_dirty(b)) {
413 set_btree_node_dirty(b);
414 queue_delayed_work(btree_io_wq, &b->work,
415 msecs_to_jiffies(30000));
418 w->prio_blocked += b->prio_blocked;
421 if (op && op->journal && !b->level) {
423 journal_pin_cmp(b->c, w, op)) {
424 atomic_dec_bug(w->journal);
429 w->journal = op->journal;
430 atomic_inc(w->journal);
434 if (current->bio_list)
437 /* Force write if set is too big */
440 set_bytes(i) > PAGE_SIZE - 48) {
442 /* Must wait on multiple writes */
445 closure_get(&op->cl);
454 * Btree in memory cache - allocation/freeing
455 * mca -> memory cache
458 static void mca_reinit(struct btree *b)
466 for (i = 0; i < MAX_BSETS; i++)
469 * Second loop starts at 1 because b->sets[0]->data is the memory we
472 for (i = 1; i < MAX_BSETS; i++)
473 b->sets[i].data = NULL;
476 #define mca_reserve(c) (((c->root && c->root->level) \
477 ? c->root->level : 1) * 8 + 16)
478 #define mca_can_free(c) \
479 max_t(int, 0, c->bucket_cache_used - mca_reserve(c))
481 static void mca_data_free(struct btree *b)
483 struct bset_tree *t = b->sets;
484 BUG_ON(!closure_is_unlocked(&b->io.cl));
486 if (bset_prev_bytes(b) < PAGE_SIZE)
489 free_pages((unsigned long) t->prev,
490 get_order(bset_prev_bytes(b)));
492 if (bset_tree_bytes(b) < PAGE_SIZE)
495 free_pages((unsigned long) t->tree,
496 get_order(bset_tree_bytes(b)));
498 free_pages((unsigned long) t->data, b->page_order);
503 list_move(&b->list, &b->c->btree_cache_freed);
504 b->c->bucket_cache_used--;
507 static void mca_bucket_free(struct btree *b)
509 BUG_ON(btree_node_dirty(b));
512 hlist_del_init_rcu(&b->hash);
513 list_move(&b->list, &b->c->btree_cache_freeable);
516 static unsigned btree_order(struct bkey *k)
518 return ilog2(KEY_SIZE(k) / PAGE_SECTORS ?: 1);
521 static void mca_data_alloc(struct btree *b, struct bkey *k, gfp_t gfp)
523 struct bset_tree *t = b->sets;
526 b->page_order = max_t(unsigned,
527 ilog2(b->c->btree_pages),
530 t->data = (void *) __get_free_pages(gfp, b->page_order);
534 t->tree = bset_tree_bytes(b) < PAGE_SIZE
535 ? kmalloc(bset_tree_bytes(b), gfp)
536 : (void *) __get_free_pages(gfp, get_order(bset_tree_bytes(b)));
540 t->prev = bset_prev_bytes(b) < PAGE_SIZE
541 ? kmalloc(bset_prev_bytes(b), gfp)
542 : (void *) __get_free_pages(gfp, get_order(bset_prev_bytes(b)));
546 list_move(&b->list, &b->c->btree_cache);
547 b->c->bucket_cache_used++;
553 static struct btree *mca_bucket_alloc(struct cache_set *c,
554 struct bkey *k, gfp_t gfp)
556 struct btree *b = kzalloc(sizeof(struct btree), gfp);
560 init_rwsem(&b->lock);
561 lockdep_set_novalidate_class(&b->lock);
562 INIT_LIST_HEAD(&b->list);
563 INIT_DELAYED_WORK(&b->work, btree_write_work);
565 closure_init_unlocked(&b->io);
567 mca_data_alloc(b, k, gfp);
571 static int mca_reap(struct btree *b, struct closure *cl, unsigned min_order)
573 lockdep_assert_held(&b->c->bucket_lock);
575 if (!down_write_trylock(&b->lock))
578 if (b->page_order < min_order) {
583 BUG_ON(btree_node_dirty(b) && !b->sets[0].data);
585 if (cl && btree_node_dirty(b))
586 bch_btree_write(b, true, NULL);
589 closure_wait_event_async(&b->io.wait, cl,
590 atomic_read(&b->io.cl.remaining) == -1);
592 if (btree_node_dirty(b) ||
593 !closure_is_unlocked(&b->io.cl) ||
594 work_pending(&b->work.work)) {
602 static int bch_mca_shrink(struct shrinker *shrink, struct shrink_control *sc)
604 struct cache_set *c = container_of(shrink, struct cache_set, shrink);
606 unsigned long i, nr = sc->nr_to_scan;
608 if (c->shrinker_disabled)
615 * If nr == 0, we're supposed to return the number of items we have
616 * cached. Not allowed to return -1.
619 return mca_can_free(c) * c->btree_pages;
621 /* Return -1 if we can't do anything right now */
622 if (sc->gfp_mask & __GFP_WAIT)
623 mutex_lock(&c->bucket_lock);
624 else if (!mutex_trylock(&c->bucket_lock))
627 nr /= c->btree_pages;
628 nr = min_t(unsigned long, nr, mca_can_free(c));
631 list_for_each_entry_safe(b, t, &c->btree_cache_freeable, list) {
636 !mca_reap(b, NULL, 0)) {
644 * Can happen right when we first start up, before we've read in any
647 if (list_empty(&c->btree_cache))
650 for (i = 0; nr && i < c->bucket_cache_used; i++) {
651 b = list_first_entry(&c->btree_cache, struct btree, list);
652 list_rotate_left(&c->btree_cache);
655 !mca_reap(b, NULL, 0)) {
664 nr = mca_can_free(c) * c->btree_pages;
665 mutex_unlock(&c->bucket_lock);
669 void bch_btree_cache_free(struct cache_set *c)
673 closure_init_stack(&cl);
675 if (c->shrink.list.next)
676 unregister_shrinker(&c->shrink);
678 mutex_lock(&c->bucket_lock);
680 #ifdef CONFIG_BCACHE_DEBUG
682 list_move(&c->verify_data->list, &c->btree_cache);
685 list_splice(&c->btree_cache_freeable,
688 while (!list_empty(&c->btree_cache)) {
689 b = list_first_entry(&c->btree_cache, struct btree, list);
691 if (btree_node_dirty(b))
692 btree_complete_write(b, btree_current_write(b));
693 clear_bit(BTREE_NODE_dirty, &b->flags);
698 while (!list_empty(&c->btree_cache_freed)) {
699 b = list_first_entry(&c->btree_cache_freed,
702 cancel_delayed_work_sync(&b->work);
706 mutex_unlock(&c->bucket_lock);
709 int bch_btree_cache_alloc(struct cache_set *c)
713 /* XXX: doesn't check for errors */
715 closure_init_unlocked(&c->gc);
717 for (i = 0; i < mca_reserve(c); i++)
718 mca_bucket_alloc(c, &ZERO_KEY, GFP_KERNEL);
720 list_splice_init(&c->btree_cache,
721 &c->btree_cache_freeable);
723 #ifdef CONFIG_BCACHE_DEBUG
724 mutex_init(&c->verify_lock);
726 c->verify_data = mca_bucket_alloc(c, &ZERO_KEY, GFP_KERNEL);
728 if (c->verify_data &&
729 c->verify_data->sets[0].data)
730 list_del_init(&c->verify_data->list);
732 c->verify_data = NULL;
735 c->shrink.shrink = bch_mca_shrink;
737 c->shrink.batch = c->btree_pages * 2;
738 register_shrinker(&c->shrink);
743 /* Btree in memory cache - hash table */
745 static struct hlist_head *mca_hash(struct cache_set *c, struct bkey *k)
747 return &c->bucket_hash[hash_32(PTR_HASH(c, k), BUCKET_HASH_BITS)];
750 static struct btree *mca_find(struct cache_set *c, struct bkey *k)
755 hlist_for_each_entry_rcu(b, mca_hash(c, k), hash)
756 if (PTR_HASH(c, &b->key) == PTR_HASH(c, k))
764 static struct btree *mca_cannibalize(struct cache_set *c, struct bkey *k,
765 int level, struct closure *cl)
771 return ERR_PTR(-ENOMEM);
774 * Trying to free up some memory - i.e. reuse some btree nodes - may
775 * require initiating IO to flush the dirty part of the node. If we're
776 * running under generic_make_request(), that IO will never finish and
777 * we would deadlock. Returning -EAGAIN causes the cache lookup code to
778 * punt to workqueue and retry.
780 if (current->bio_list)
781 return ERR_PTR(-EAGAIN);
783 if (c->try_harder && c->try_harder != cl) {
784 closure_wait_event_async(&c->try_wait, cl, !c->try_harder);
785 return ERR_PTR(-EAGAIN);
788 /* XXX: tracepoint */
790 c->try_harder_start = local_clock();
792 list_for_each_entry_reverse(i, &c->btree_cache, list) {
793 int r = mca_reap(i, cl, btree_order(k));
800 if (ret == -EAGAIN &&
801 closure_blocking(cl)) {
802 mutex_unlock(&c->bucket_lock);
804 mutex_lock(&c->bucket_lock);
812 * We can only have one thread cannibalizing other cached btree nodes at a time,
813 * or we'll deadlock. We use an open coded mutex to ensure that, which a
814 * cannibalize_bucket() will take. This means every time we unlock the root of
815 * the btree, we need to release this lock if we have it held.
817 void bch_cannibalize_unlock(struct cache_set *c, struct closure *cl)
819 if (c->try_harder == cl) {
820 bch_time_stats_update(&c->try_harder_time, c->try_harder_start);
821 c->try_harder = NULL;
822 __closure_wake_up(&c->try_wait);
826 static struct btree *mca_alloc(struct cache_set *c, struct bkey *k,
827 int level, struct closure *cl)
831 lockdep_assert_held(&c->bucket_lock);
836 /* btree_free() doesn't free memory; it sticks the node on the end of
837 * the list. Check if there's any freed nodes there:
839 list_for_each_entry(b, &c->btree_cache_freeable, list)
840 if (!mca_reap(b, NULL, btree_order(k)))
843 /* We never free struct btree itself, just the memory that holds the on
844 * disk node. Check the freed list before allocating a new one:
846 list_for_each_entry(b, &c->btree_cache_freed, list)
847 if (!mca_reap(b, NULL, 0)) {
848 mca_data_alloc(b, k, __GFP_NOWARN|GFP_NOIO);
849 if (!b->sets[0].data)
855 b = mca_bucket_alloc(c, k, __GFP_NOWARN|GFP_NOIO);
859 BUG_ON(!down_write_trylock(&b->lock));
863 BUG_ON(!closure_is_unlocked(&b->io.cl));
865 bkey_copy(&b->key, k);
866 list_move(&b->list, &c->btree_cache);
867 hlist_del_init_rcu(&b->hash);
868 hlist_add_head_rcu(&b->hash, mca_hash(c, k));
870 lock_set_subclass(&b->lock.dep_map, level + 1, _THIS_IP_);
880 b = mca_cannibalize(c, k, level, cl);
888 * bch_btree_node_get - find a btree node in the cache and lock it, reading it
889 * in from disk if necessary.
891 * If IO is necessary, it uses the closure embedded in struct btree_op to wait;
892 * if that closure is in non blocking mode, will return -EAGAIN.
894 * The btree node will have either a read or a write lock held, depending on
895 * level and op->lock.
897 struct btree *bch_btree_node_get(struct cache_set *c, struct bkey *k,
898 int level, struct btree_op *op)
901 bool write = level <= op->lock;
909 mutex_lock(&c->bucket_lock);
910 b = mca_alloc(c, k, level, &op->cl);
911 mutex_unlock(&c->bucket_lock);
921 downgrade_write(&b->lock);
923 rw_lock(write, b, level);
924 if (PTR_HASH(c, &b->key) != PTR_HASH(c, k)) {
928 BUG_ON(b->level != level);
933 for (; i <= b->nsets && b->sets[i].size; i++) {
934 prefetch(b->sets[i].tree);
935 prefetch(b->sets[i].data);
938 for (; i <= b->nsets; i++)
939 prefetch(b->sets[i].data);
941 if (!closure_wait_event(&b->io.wait, &op->cl,
942 btree_node_read_done(b))) {
944 b = ERR_PTR(-EAGAIN);
945 } else if (btree_node_io_error(b)) {
954 static void btree_node_prefetch(struct cache_set *c, struct bkey *k, int level)
958 mutex_lock(&c->bucket_lock);
959 b = mca_alloc(c, k, level, NULL);
960 mutex_unlock(&c->bucket_lock);
962 if (!IS_ERR_OR_NULL(b)) {
970 static void btree_node_free(struct btree *b, struct btree_op *op)
975 * The BUG_ON() in btree_node_get() implies that we must have a write
976 * lock on parent to free or even invalidate a node
978 BUG_ON(op->lock <= b->level);
979 BUG_ON(b == b->c->root);
980 pr_debug("bucket %s", pbtree(b));
982 if (btree_node_dirty(b))
983 btree_complete_write(b, btree_current_write(b));
984 clear_bit(BTREE_NODE_dirty, &b->flags);
986 if (b->prio_blocked &&
987 !atomic_sub_return(b->prio_blocked, &b->c->prio_blocked))
988 wake_up(&b->c->alloc_wait);
992 cancel_delayed_work(&b->work);
994 mutex_lock(&b->c->bucket_lock);
996 for (i = 0; i < KEY_PTRS(&b->key); i++) {
997 BUG_ON(atomic_read(&PTR_BUCKET(b->c, &b->key, i)->pin));
999 bch_inc_gen(PTR_CACHE(b->c, &b->key, i),
1000 PTR_BUCKET(b->c, &b->key, i));
1003 bch_bucket_free(b->c, &b->key);
1005 mutex_unlock(&b->c->bucket_lock);
1008 struct btree *bch_btree_node_alloc(struct cache_set *c, int level,
1012 struct btree *b = ERR_PTR(-EAGAIN);
1014 mutex_lock(&c->bucket_lock);
1016 if (__bch_bucket_alloc_set(c, WATERMARK_METADATA, &k.key, 1, cl))
1019 SET_KEY_SIZE(&k.key, c->btree_pages * PAGE_SECTORS);
1021 b = mca_alloc(c, &k.key, level, cl);
1027 "Tried to allocate bucket that was in btree cache");
1028 __bkey_put(c, &k.key);
1032 set_btree_node_read_done(b);
1034 bch_bset_init_next(b);
1036 mutex_unlock(&c->bucket_lock);
1039 bch_bucket_free(c, &k.key);
1040 __bkey_put(c, &k.key);
1042 mutex_unlock(&c->bucket_lock);
1046 static struct btree *btree_node_alloc_replacement(struct btree *b,
1049 struct btree *n = bch_btree_node_alloc(b->c, b->level, cl);
1050 if (!IS_ERR_OR_NULL(n))
1051 bch_btree_sort_into(b, n);
1056 /* Garbage collection */
1058 uint8_t __bch_btree_mark_key(struct cache_set *c, int level, struct bkey *k)
1065 * ptr_invalid() can't return true for the keys that mark btree nodes as
1066 * freed, but since ptr_bad() returns true we'll never actually use them
1067 * for anything and thus we don't want mark their pointers here
1069 if (!bkey_cmp(k, &ZERO_KEY))
1072 for (i = 0; i < KEY_PTRS(k); i++) {
1073 if (!ptr_available(c, k, i))
1076 g = PTR_BUCKET(c, k, i);
1078 if (gen_after(g->gc_gen, PTR_GEN(k, i)))
1079 g->gc_gen = PTR_GEN(k, i);
1081 if (ptr_stale(c, k, i)) {
1082 stale = max(stale, ptr_stale(c, k, i));
1086 cache_bug_on(GC_MARK(g) &&
1087 (GC_MARK(g) == GC_MARK_METADATA) != (level != 0),
1088 c, "inconsistent ptrs: mark = %llu, level = %i",
1092 SET_GC_MARK(g, GC_MARK_METADATA);
1093 else if (KEY_DIRTY(k))
1094 SET_GC_MARK(g, GC_MARK_DIRTY);
1096 /* guard against overflow */
1097 SET_GC_SECTORS_USED(g, min_t(unsigned,
1098 GC_SECTORS_USED(g) + KEY_SIZE(k),
1101 BUG_ON(!GC_SECTORS_USED(g));
1107 #define btree_mark_key(b, k) __bch_btree_mark_key(b->c, b->level, k)
1109 static int btree_gc_mark_node(struct btree *b, unsigned *keys,
1113 unsigned last_dev = -1;
1114 struct bcache_device *d = NULL;
1116 struct btree_iter iter;
1117 struct bset_tree *t;
1121 for_each_key_filter(b, k, &iter, bch_ptr_invalid) {
1122 if (last_dev != KEY_INODE(k)) {
1123 last_dev = KEY_INODE(k);
1125 d = KEY_INODE(k) < b->c->nr_uuids
1126 ? b->c->devices[last_dev]
1130 stale = max(stale, btree_mark_key(b, k));
1132 if (bch_ptr_bad(b, k))
1135 *keys += bkey_u64s(k);
1137 gc->key_bytes += bkey_u64s(k);
1140 gc->data += KEY_SIZE(k);
1142 gc->dirty += KEY_SIZE(k);
1144 d->sectors_dirty_gc += KEY_SIZE(k);
1148 for (t = b->sets; t <= &b->sets[b->nsets]; t++)
1149 btree_bug_on(t->size &&
1150 bset_written(b, t) &&
1151 bkey_cmp(&b->key, &t->end) < 0,
1152 b, "found short btree key in gc");
1157 static struct btree *btree_gc_alloc(struct btree *b, struct bkey *k,
1158 struct btree_op *op)
1161 * We block priorities from being written for the duration of garbage
1162 * collection, so we can't sleep in btree_alloc() ->
1163 * bch_bucket_alloc_set(), or we'd risk deadlock - so we don't pass it
1166 struct btree *n = btree_node_alloc_replacement(b, NULL);
1168 if (!IS_ERR_OR_NULL(n)) {
1171 memcpy(k->ptr, b->key.ptr,
1172 sizeof(uint64_t) * KEY_PTRS(&b->key));
1174 __bkey_put(b->c, &b->key);
1175 atomic_inc(&b->c->prio_blocked);
1178 btree_node_free(n, op);
1186 * Leaving this at 2 until we've got incremental garbage collection done; it
1187 * could be higher (and has been tested with 4) except that garbage collection
1188 * could take much longer, adversely affecting latency.
1190 #define GC_MERGE_NODES 2U
1192 struct gc_merge_info {
1198 static void btree_gc_coalesce(struct btree *b, struct btree_op *op,
1199 struct gc_stat *gc, struct gc_merge_info *r)
1201 unsigned nodes = 0, keys = 0, blocks;
1204 while (nodes < GC_MERGE_NODES && r[nodes].b)
1205 keys += r[nodes++].keys;
1207 blocks = btree_default_blocks(b->c) * 2 / 3;
1210 __set_blocks(b->sets[0].data, keys, b->c) > blocks * (nodes - 1))
1213 for (i = nodes - 1; i >= 0; --i) {
1214 if (r[i].b->written)
1215 r[i].b = btree_gc_alloc(r[i].b, r[i].k, op);
1217 if (r[i].b->written)
1221 for (i = nodes - 1; i > 0; --i) {
1222 struct bset *n1 = r[i].b->sets->data;
1223 struct bset *n2 = r[i - 1].b->sets->data;
1224 struct bkey *k, *last = NULL;
1230 * Last node we're not getting rid of - we're getting
1231 * rid of the node at r[0]. Have to try and fit all of
1232 * the remaining keys into this node; we can't ensure
1233 * they will always fit due to rounding and variable
1234 * length keys (shouldn't be possible in practice,
1237 if (__set_blocks(n1, n1->keys + r->keys,
1238 b->c) > btree_blocks(r[i].b))
1247 if (__set_blocks(n1, n1->keys + keys +
1248 bkey_u64s(k), b->c) > blocks)
1252 keys += bkey_u64s(k);
1255 BUG_ON(__set_blocks(n1, n1->keys + keys,
1256 b->c) > btree_blocks(r[i].b));
1259 bkey_copy_key(&r[i].b->key, last);
1260 bkey_copy_key(r[i].k, last);
1265 (void *) node(n2, keys) - (void *) n2->start);
1271 (void *) end(n2) - (void *) node(n2, keys));
1275 r[i].keys = n1->keys;
1276 r[i - 1].keys = n2->keys;
1279 btree_node_free(r->b, op);
1280 up_write(&r->b->lock);
1282 pr_debug("coalesced %u nodes", nodes);
1287 memmove(&r[0], &r[1], sizeof(struct gc_merge_info) * nodes);
1288 memset(&r[nodes], 0, sizeof(struct gc_merge_info));
1291 static int btree_gc_recurse(struct btree *b, struct btree_op *op,
1292 struct closure *writes, struct gc_stat *gc)
1294 void write(struct btree *r)
1297 bch_btree_write(r, true, op);
1298 else if (btree_node_dirty(r)) {
1299 BUG_ON(btree_current_write(r)->owner);
1300 btree_current_write(r)->owner = writes;
1301 closure_get(writes);
1303 bch_btree_write(r, true, NULL);
1311 struct gc_merge_info r[GC_MERGE_NODES];
1313 memset(r, 0, sizeof(r));
1315 while ((r->k = bch_next_recurse_key(b, &b->c->gc_done))) {
1316 r->b = bch_btree_node_get(b->c, r->k, b->level - 1, op);
1319 ret = PTR_ERR(r->b);
1324 stale = btree_gc_mark_node(r->b, &r->keys, gc);
1327 (r->b->level || stale > 10 ||
1328 b->c->gc_always_rewrite))
1329 r->b = btree_gc_alloc(r->b, r->k, op);
1332 ret = btree_gc_recurse(r->b, op, writes, gc);
1339 bkey_copy_key(&b->c->gc_done, r->k);
1342 btree_gc_coalesce(b, op, gc, r);
1344 if (r[GC_MERGE_NODES - 1].b)
1345 write(r[GC_MERGE_NODES - 1].b);
1347 memmove(&r[1], &r[0],
1348 sizeof(struct gc_merge_info) * (GC_MERGE_NODES - 1));
1350 /* When we've got incremental GC working, we'll want to do
1351 * if (should_resched())
1356 if (need_resched()) {
1363 for (i = 1; i < GC_MERGE_NODES && r[i].b; i++)
1366 /* Might have freed some children, must remove their keys */
1373 static int bch_btree_gc_root(struct btree *b, struct btree_op *op,
1374 struct closure *writes, struct gc_stat *gc)
1376 struct btree *n = NULL;
1378 int ret = 0, stale = btree_gc_mark_node(b, &keys, gc);
1380 if (b->level || stale > 10)
1381 n = btree_node_alloc_replacement(b, NULL);
1383 if (!IS_ERR_OR_NULL(n))
1387 ret = btree_gc_recurse(b, op, writes, gc);
1389 if (!b->written || btree_node_dirty(b)) {
1390 atomic_inc(&b->c->prio_blocked);
1392 bch_btree_write(b, true, n ? op : NULL);
1395 if (!IS_ERR_OR_NULL(n)) {
1396 closure_sync(&op->cl);
1397 bch_btree_set_root(b);
1398 btree_node_free(n, op);
1405 static void btree_gc_start(struct cache_set *c)
1409 struct bcache_device **d;
1412 if (!c->gc_mark_valid)
1415 mutex_lock(&c->bucket_lock);
1417 c->gc_mark_valid = 0;
1418 c->gc_done = ZERO_KEY;
1420 for_each_cache(ca, c, i)
1421 for_each_bucket(b, ca) {
1423 if (!atomic_read(&b->pin))
1424 SET_GC_MARK(b, GC_MARK_RECLAIMABLE);
1427 for (d = c->devices;
1428 d < c->devices + c->nr_uuids;
1431 (*d)->sectors_dirty_gc = 0;
1433 mutex_unlock(&c->bucket_lock);
1436 size_t bch_btree_gc_finish(struct cache_set *c)
1438 size_t available = 0;
1441 struct bcache_device **d;
1444 mutex_lock(&c->bucket_lock);
1447 c->gc_mark_valid = 1;
1451 for (i = 0; i < KEY_PTRS(&c->root->key); i++)
1452 SET_GC_MARK(PTR_BUCKET(c, &c->root->key, i),
1455 for (i = 0; i < KEY_PTRS(&c->uuid_bucket); i++)
1456 SET_GC_MARK(PTR_BUCKET(c, &c->uuid_bucket, i),
1459 for_each_cache(ca, c, i) {
1462 ca->invalidate_needs_gc = 0;
1464 for (i = ca->sb.d; i < ca->sb.d + ca->sb.keys; i++)
1465 SET_GC_MARK(ca->buckets + *i, GC_MARK_METADATA);
1467 for (i = ca->prio_buckets;
1468 i < ca->prio_buckets + prio_buckets(ca) * 2; i++)
1469 SET_GC_MARK(ca->buckets + *i, GC_MARK_METADATA);
1471 for_each_bucket(b, ca) {
1472 b->last_gc = b->gc_gen;
1473 c->need_gc = max(c->need_gc, bucket_gc_gen(b));
1475 if (!atomic_read(&b->pin) &&
1476 GC_MARK(b) == GC_MARK_RECLAIMABLE) {
1478 if (!GC_SECTORS_USED(b))
1479 bch_bucket_add_unused(ca, b);
1484 for (d = c->devices;
1485 d < c->devices + c->nr_uuids;
1488 unsigned long last =
1489 atomic_long_read(&((*d)->sectors_dirty));
1490 long difference = (*d)->sectors_dirty_gc - last;
1492 pr_debug("sectors dirty off by %li", difference);
1494 (*d)->sectors_dirty_last += difference;
1496 atomic_long_set(&((*d)->sectors_dirty),
1497 (*d)->sectors_dirty_gc);
1500 mutex_unlock(&c->bucket_lock);
1504 static void bch_btree_gc(struct closure *cl)
1506 struct cache_set *c = container_of(cl, struct cache_set, gc.cl);
1508 unsigned long available;
1509 struct gc_stat stats;
1510 struct closure writes;
1513 uint64_t start_time = local_clock();
1514 trace_bcache_gc_start(c->sb.set_uuid);
1515 blktrace_msg_all(c, "Starting gc");
1517 memset(&stats, 0, sizeof(struct gc_stat));
1518 closure_init_stack(&writes);
1519 bch_btree_op_init_stack(&op);
1524 ret = btree_root(gc_root, c, &op, &writes, &stats);
1525 closure_sync(&op.cl);
1526 closure_sync(&writes);
1529 blktrace_msg_all(c, "Stopped gc");
1530 pr_warn("gc failed!");
1532 continue_at(cl, bch_btree_gc, bch_gc_wq);
1535 /* Possibly wait for new UUIDs or whatever to hit disk */
1536 bch_journal_meta(c, &op.cl);
1537 closure_sync(&op.cl);
1539 available = bch_btree_gc_finish(c);
1541 bch_time_stats_update(&c->btree_gc_time, start_time);
1543 stats.key_bytes *= sizeof(uint64_t);
1546 stats.in_use = (c->nbuckets - available) * 100 / c->nbuckets;
1547 memcpy(&c->gc_stats, &stats, sizeof(struct gc_stat));
1548 blktrace_msg_all(c, "Finished gc");
1550 trace_bcache_gc_end(c->sb.set_uuid);
1551 wake_up(&c->alloc_wait);
1553 continue_at(cl, bch_moving_gc, bch_gc_wq);
1556 void bch_queue_gc(struct cache_set *c)
1558 closure_trylock_call(&c->gc.cl, bch_btree_gc, bch_gc_wq, &c->cl);
1561 /* Initial partial gc */
1563 static int bch_btree_check_recurse(struct btree *b, struct btree_op *op,
1564 unsigned long **seen)
1570 struct btree_iter iter;
1572 for_each_key_filter(b, k, &iter, bch_ptr_invalid) {
1573 for (i = 0; i < KEY_PTRS(k); i++) {
1574 if (!ptr_available(b->c, k, i))
1577 g = PTR_BUCKET(b->c, k, i);
1579 if (!__test_and_set_bit(PTR_BUCKET_NR(b->c, k, i),
1580 seen[PTR_DEV(k, i)]) ||
1581 !ptr_stale(b->c, k, i)) {
1582 g->gen = PTR_GEN(k, i);
1585 g->prio = BTREE_PRIO;
1586 else if (g->prio == BTREE_PRIO)
1587 g->prio = INITIAL_PRIO;
1591 btree_mark_key(b, k);
1595 k = bch_next_recurse_key(b, &ZERO_KEY);
1598 struct bkey *p = bch_next_recurse_key(b, k);
1600 btree_node_prefetch(b->c, p, b->level - 1);
1602 ret = btree(check_recurse, k, b, op, seen);
1613 int bch_btree_check(struct cache_set *c, struct btree_op *op)
1617 unsigned long *seen[MAX_CACHES_PER_SET];
1619 memset(seen, 0, sizeof(seen));
1621 for (i = 0; c->cache[i]; i++) {
1622 size_t n = DIV_ROUND_UP(c->cache[i]->sb.nbuckets, 8);
1623 seen[i] = kmalloc(n, GFP_KERNEL);
1627 /* Disables the seen array until prio_read() uses it too */
1628 memset(seen[i], 0xFF, n);
1631 ret = btree_root(check_recurse, c, op, seen);
1633 for (i = 0; i < MAX_CACHES_PER_SET; i++)
1638 /* Btree insertion */
1640 static void shift_keys(struct btree *b, struct bkey *where, struct bkey *insert)
1642 struct bset *i = b->sets[b->nsets].data;
1644 memmove((uint64_t *) where + bkey_u64s(insert),
1646 (void *) end(i) - (void *) where);
1648 i->keys += bkey_u64s(insert);
1649 bkey_copy(where, insert);
1650 bch_bset_fix_lookup_table(b, where);
1653 static bool fix_overlapping_extents(struct btree *b,
1654 struct bkey *insert,
1655 struct btree_iter *iter,
1656 struct btree_op *op)
1658 void subtract_dirty(struct bkey *k, int sectors)
1660 struct bcache_device *d = b->c->devices[KEY_INODE(k)];
1662 if (KEY_DIRTY(k) && d)
1663 atomic_long_sub(sectors, &d->sectors_dirty);
1666 unsigned old_size, sectors_found = 0;
1669 struct bkey *k = bch_btree_iter_next(iter);
1671 bkey_cmp(&START_KEY(k), insert) >= 0)
1674 if (bkey_cmp(k, &START_KEY(insert)) <= 0)
1677 old_size = KEY_SIZE(k);
1680 * We might overlap with 0 size extents; we can't skip these
1681 * because if they're in the set we're inserting to we have to
1682 * adjust them so they don't overlap with the key we're
1683 * inserting. But we don't want to check them for BTREE_REPLACE
1687 if (op->type == BTREE_REPLACE &&
1690 * k might have been split since we inserted/found the
1691 * key we're replacing
1694 uint64_t offset = KEY_START(k) -
1695 KEY_START(&op->replace);
1697 /* But it must be a subset of the replace key */
1698 if (KEY_START(k) < KEY_START(&op->replace) ||
1699 KEY_OFFSET(k) > KEY_OFFSET(&op->replace))
1702 /* We didn't find a key that we were supposed to */
1703 if (KEY_START(k) > KEY_START(insert) + sectors_found)
1706 if (KEY_PTRS(&op->replace) != KEY_PTRS(k))
1712 BUG_ON(!KEY_PTRS(&op->replace));
1714 for (i = 0; i < KEY_PTRS(&op->replace); i++)
1715 if (k->ptr[i] != op->replace.ptr[i] + offset)
1718 sectors_found = KEY_OFFSET(k) - KEY_START(insert);
1721 if (bkey_cmp(insert, k) < 0 &&
1722 bkey_cmp(&START_KEY(insert), &START_KEY(k)) > 0) {
1724 * We overlapped in the middle of an existing key: that
1725 * means we have to split the old key. But we have to do
1726 * slightly different things depending on whether the
1727 * old key has been written out yet.
1732 subtract_dirty(k, KEY_SIZE(insert));
1734 if (bkey_written(b, k)) {
1736 * We insert a new key to cover the top of the
1737 * old key, and the old key is modified in place
1738 * to represent the bottom split.
1740 * It's completely arbitrary whether the new key
1741 * is the top or the bottom, but it has to match
1742 * up with what btree_sort_fixup() does - it
1743 * doesn't check for this kind of overlap, it
1744 * depends on us inserting a new key for the top
1747 top = bch_bset_search(b, &b->sets[b->nsets],
1749 shift_keys(b, top, k);
1751 BKEY_PADDED(key) temp;
1752 bkey_copy(&temp.key, k);
1753 shift_keys(b, k, &temp.key);
1757 bch_cut_front(insert, top);
1758 bch_cut_back(&START_KEY(insert), k);
1759 bch_bset_fix_invalidated_key(b, k);
1763 if (bkey_cmp(insert, k) < 0) {
1764 bch_cut_front(insert, k);
1766 if (bkey_written(b, k) &&
1767 bkey_cmp(&START_KEY(insert), &START_KEY(k)) <= 0) {
1769 * Completely overwrote, so we don't have to
1770 * invalidate the binary search tree
1772 bch_cut_front(k, k);
1774 __bch_cut_back(&START_KEY(insert), k);
1775 bch_bset_fix_invalidated_key(b, k);
1779 subtract_dirty(k, old_size - KEY_SIZE(k));
1783 if (op->type == BTREE_REPLACE) {
1784 if (!sectors_found) {
1785 op->insert_collision = true;
1787 } else if (sectors_found < KEY_SIZE(insert)) {
1788 SET_KEY_OFFSET(insert, KEY_OFFSET(insert) -
1789 (KEY_SIZE(insert) - sectors_found));
1790 SET_KEY_SIZE(insert, sectors_found);
1797 static bool btree_insert_key(struct btree *b, struct btree_op *op,
1800 struct bset *i = b->sets[b->nsets].data;
1801 struct bkey *m, *prev;
1802 const char *status = "insert";
1804 BUG_ON(bkey_cmp(k, &b->key) > 0);
1805 BUG_ON(b->level && !KEY_PTRS(k));
1806 BUG_ON(!b->level && !KEY_OFFSET(k));
1809 struct btree_iter iter;
1810 struct bkey search = KEY(KEY_INODE(k), KEY_START(k), 0);
1813 * bset_search() returns the first key that is strictly greater
1814 * than the search key - but for back merging, we want to find
1815 * the first key that is greater than or equal to KEY_START(k) -
1816 * unless KEY_START(k) is 0.
1818 if (KEY_OFFSET(&search))
1819 SET_KEY_OFFSET(&search, KEY_OFFSET(&search) - 1);
1822 m = bch_btree_iter_init(b, &iter, &search);
1824 if (fix_overlapping_extents(b, k, &iter, op))
1827 while (m != end(i) &&
1828 bkey_cmp(k, &START_KEY(m)) > 0)
1829 prev = m, m = bkey_next(m);
1831 if (key_merging_disabled(b->c))
1834 /* prev is in the tree, if we merge we're done */
1835 status = "back merging";
1837 bch_bkey_try_merge(b, prev, k))
1840 status = "overwrote front";
1842 KEY_PTRS(m) == KEY_PTRS(k) && !KEY_SIZE(m))
1845 status = "front merge";
1847 bch_bkey_try_merge(b, k, m))
1850 m = bch_bset_search(b, &b->sets[b->nsets], k);
1852 insert: shift_keys(b, m, k);
1853 copy: bkey_copy(m, k);
1855 bch_check_keys(b, "%s for %s at %s: %s", status,
1856 op_type(op), pbtree(b), pkey(k));
1857 bch_check_key_order_msg(b, i, "%s for %s at %s: %s", status,
1858 op_type(op), pbtree(b), pkey(k));
1860 if (b->level && !KEY_OFFSET(k))
1863 pr_debug("%s for %s at %s: %s", status,
1864 op_type(op), pbtree(b), pkey(k));
1869 bool bch_btree_insert_keys(struct btree *b, struct btree_op *op)
1873 unsigned oldsize = bch_count_data(b);
1875 while ((k = bch_keylist_pop(&op->keys))) {
1876 bkey_put(b->c, k, b->level);
1877 ret |= btree_insert_key(b, op, k);
1880 BUG_ON(bch_count_data(b) < oldsize);
1884 bool bch_btree_insert_check_key(struct btree *b, struct btree_op *op,
1888 uint64_t btree_ptr = b->key.ptr[0];
1889 unsigned long seq = b->seq;
1892 rw_unlock(false, b);
1893 rw_lock(true, b, b->level);
1895 if (b->key.ptr[0] != btree_ptr ||
1896 b->seq != seq + 1 ||
1900 op->replace = KEY(op->inode, bio_end(bio), bio_sectors(bio));
1902 SET_KEY_PTRS(&op->replace, 1);
1903 get_random_bytes(&op->replace.ptr[0], sizeof(uint64_t));
1905 SET_PTR_DEV(&op->replace, 0, PTR_CHECK_DEV);
1907 bkey_copy(&tmp.k, &op->replace);
1909 BUG_ON(op->type != BTREE_INSERT);
1910 BUG_ON(!btree_insert_key(b, op, &tmp.k));
1911 bch_btree_write(b, false, NULL);
1914 downgrade_write(&b->lock);
1918 static int btree_split(struct btree *b, struct btree_op *op)
1920 bool split, root = b == b->c->root;
1921 struct btree *n1, *n2 = NULL, *n3 = NULL;
1922 uint64_t start_time = local_clock();
1925 set_closure_blocking(&op->cl);
1927 n1 = btree_node_alloc_replacement(b, &op->cl);
1931 split = set_blocks(n1->sets[0].data, n1->c) > (btree_blocks(b) * 4) / 5;
1933 pr_debug("%ssplitting at %s keys %i", split ? "" : "not ",
1934 pbtree(b), n1->sets[0].data->keys);
1939 n2 = bch_btree_node_alloc(b->c, b->level, &op->cl);
1944 n3 = bch_btree_node_alloc(b->c, b->level + 1, &op->cl);
1949 bch_btree_insert_keys(n1, op);
1951 /* Has to be a linear search because we don't have an auxiliary
1955 while (keys < (n1->sets[0].data->keys * 3) / 5)
1956 keys += bkey_u64s(node(n1->sets[0].data, keys));
1958 bkey_copy_key(&n1->key, node(n1->sets[0].data, keys));
1959 keys += bkey_u64s(node(n1->sets[0].data, keys));
1961 n2->sets[0].data->keys = n1->sets[0].data->keys - keys;
1962 n1->sets[0].data->keys = keys;
1964 memcpy(n2->sets[0].data->start,
1965 end(n1->sets[0].data),
1966 n2->sets[0].data->keys * sizeof(uint64_t));
1968 bkey_copy_key(&n2->key, &b->key);
1970 bch_keylist_add(&op->keys, &n2->key);
1971 bch_btree_write(n2, true, op);
1972 rw_unlock(true, n2);
1974 bch_btree_insert_keys(n1, op);
1976 bch_keylist_add(&op->keys, &n1->key);
1977 bch_btree_write(n1, true, op);
1980 bkey_copy_key(&n3->key, &MAX_KEY);
1981 bch_btree_insert_keys(n3, op);
1982 bch_btree_write(n3, true, op);
1984 closure_sync(&op->cl);
1985 bch_btree_set_root(n3);
1986 rw_unlock(true, n3);
1988 op->keys.top = op->keys.bottom;
1989 closure_sync(&op->cl);
1990 bch_btree_set_root(n1);
1994 bkey_copy(op->keys.top, &b->key);
1995 bkey_copy_key(op->keys.top, &ZERO_KEY);
1997 for (i = 0; i < KEY_PTRS(&b->key); i++) {
1998 uint8_t g = PTR_BUCKET(b->c, &b->key, i)->gen + 1;
2000 SET_PTR_GEN(op->keys.top, i, g);
2003 bch_keylist_push(&op->keys);
2004 closure_sync(&op->cl);
2005 atomic_inc(&b->c->prio_blocked);
2008 rw_unlock(true, n1);
2009 btree_node_free(b, op);
2011 bch_time_stats_update(&b->c->btree_split_time, start_time);
2015 __bkey_put(n2->c, &n2->key);
2016 btree_node_free(n2, op);
2017 rw_unlock(true, n2);
2019 __bkey_put(n1->c, &n1->key);
2020 btree_node_free(n1, op);
2021 rw_unlock(true, n1);
2023 if (n3 == ERR_PTR(-EAGAIN) ||
2024 n2 == ERR_PTR(-EAGAIN) ||
2025 n1 == ERR_PTR(-EAGAIN))
2028 pr_warn("couldn't split");
2032 static int bch_btree_insert_recurse(struct btree *b, struct btree_op *op,
2033 struct keylist *stack_keys)
2037 struct bkey *insert = op->keys.bottom;
2038 struct bkey *k = bch_next_recurse_key(b, &START_KEY(insert));
2041 btree_bug(b, "no key to recurse on at level %i/%i",
2042 b->level, b->c->root->level);
2044 op->keys.top = op->keys.bottom;
2048 if (bkey_cmp(insert, k) > 0) {
2051 if (op->type == BTREE_REPLACE) {
2052 __bkey_put(b->c, insert);
2053 op->keys.top = op->keys.bottom;
2054 op->insert_collision = true;
2058 for (i = 0; i < KEY_PTRS(insert); i++)
2059 atomic_inc(&PTR_BUCKET(b->c, insert, i)->pin);
2061 bkey_copy(stack_keys->top, insert);
2063 bch_cut_back(k, insert);
2064 bch_cut_front(k, stack_keys->top);
2066 bch_keylist_push(stack_keys);
2069 ret = btree(insert_recurse, k, b, op, stack_keys);
2074 if (!bch_keylist_empty(&op->keys)) {
2075 if (should_split(b)) {
2076 if (op->lock <= b->c->root->level) {
2078 op->lock = b->c->root->level + 1;
2081 return btree_split(b, op);
2084 BUG_ON(write_block(b) != b->sets[b->nsets].data);
2086 if (bch_btree_insert_keys(b, op))
2087 bch_btree_write(b, false, op);
2093 int bch_btree_insert(struct btree_op *op, struct cache_set *c)
2096 struct keylist stack_keys;
2099 * Don't want to block with the btree locked unless we have to,
2100 * otherwise we get deadlocks with try_harder and between split/gc
2102 clear_closure_blocking(&op->cl);
2104 BUG_ON(bch_keylist_empty(&op->keys));
2105 bch_keylist_copy(&stack_keys, &op->keys);
2106 bch_keylist_init(&op->keys);
2108 while (!bch_keylist_empty(&stack_keys) ||
2109 !bch_keylist_empty(&op->keys)) {
2110 if (bch_keylist_empty(&op->keys)) {
2111 bch_keylist_add(&op->keys,
2112 bch_keylist_pop(&stack_keys));
2116 ret = btree_root(insert_recurse, c, op, &stack_keys);
2118 if (ret == -EAGAIN) {
2120 closure_sync(&op->cl);
2124 pr_err("error %i trying to insert key for %s",
2127 while ((k = bch_keylist_pop(&stack_keys) ?:
2128 bch_keylist_pop(&op->keys)))
2133 bch_keylist_free(&stack_keys);
2136 atomic_dec_bug(op->journal);
2141 void bch_btree_set_root(struct btree *b)
2145 BUG_ON(!b->written);
2147 for (i = 0; i < KEY_PTRS(&b->key); i++)
2148 BUG_ON(PTR_BUCKET(b->c, &b->key, i)->prio != BTREE_PRIO);
2150 mutex_lock(&b->c->bucket_lock);
2151 list_del_init(&b->list);
2152 mutex_unlock(&b->c->bucket_lock);
2155 __bkey_put(b->c, &b->key);
2157 bch_journal_meta(b->c, NULL);
2158 pr_debug("%s for %pf", pbtree(b), __builtin_return_address(0));
2163 static int submit_partial_cache_miss(struct btree *b, struct btree_op *op,
2166 struct search *s = container_of(op, struct search, op);
2167 struct bio *bio = &s->bio.bio;
2172 unsigned sectors = INT_MAX;
2174 if (KEY_INODE(k) == op->inode) {
2175 if (KEY_START(k) <= bio->bi_sector)
2178 sectors = min_t(uint64_t, sectors,
2179 KEY_START(k) - bio->bi_sector);
2182 ret = s->d->cache_miss(b, s, bio, sectors);
2189 * Read from a single key, handling the initial cache miss if the key starts in
2190 * the middle of the bio
2192 static int submit_partial_cache_hit(struct btree *b, struct btree_op *op,
2195 struct search *s = container_of(op, struct search, op);
2196 struct bio *bio = &s->bio.bio;
2200 int ret = submit_partial_cache_miss(b, op, k);
2201 if (ret || op->lookup_done)
2204 /* XXX: figure out best pointer - for multiple cache devices */
2207 PTR_BUCKET(b->c, k, ptr)->prio = INITIAL_PRIO;
2209 while (!op->lookup_done &&
2210 KEY_INODE(k) == op->inode &&
2211 bio->bi_sector < KEY_OFFSET(k)) {
2212 struct bkey *bio_key;
2213 sector_t sector = PTR_OFFSET(k, ptr) +
2214 (bio->bi_sector - KEY_START(k));
2215 unsigned sectors = min_t(uint64_t, INT_MAX,
2216 KEY_OFFSET(k) - bio->bi_sector);
2218 n = bch_bio_split(bio, sectors, GFP_NOIO, s->d->bio_split);
2223 op->lookup_done = true;
2225 bio_key = &container_of(n, struct bbio, bio)->key;
2228 * The bucket we're reading from might be reused while our bio
2229 * is in flight, and we could then end up reading the wrong
2232 * We guard against this by checking (in cache_read_endio()) if
2233 * the pointer is stale again; if so, we treat it as an error
2234 * and reread from the backing device (but we don't pass that
2235 * error up anywhere).
2238 bch_bkey_copy_single_ptr(bio_key, k, ptr);
2239 SET_PTR_OFFSET(bio_key, 0, sector);
2241 n->bi_end_io = bch_cache_read_endio;
2242 n->bi_private = &s->cl;
2244 trace_bcache_cache_hit(n);
2245 __bch_submit_bbio(n, b->c);
2251 int bch_btree_search_recurse(struct btree *b, struct btree_op *op)
2253 struct search *s = container_of(op, struct search, op);
2254 struct bio *bio = &s->bio.bio;
2258 struct btree_iter iter;
2259 bch_btree_iter_init(b, &iter, &KEY(op->inode, bio->bi_sector, 0));
2261 pr_debug("at %s searching for %u:%llu", pbtree(b), op->inode,
2262 (uint64_t) bio->bi_sector);
2265 k = bch_btree_iter_next_filter(&iter, b, bch_ptr_bad);
2268 * b->key would be exactly what we want, except that
2269 * pointers to btree nodes have nonzero size - we
2270 * wouldn't go far enough
2273 ret = submit_partial_cache_miss(b, op,
2274 &KEY(KEY_INODE(&b->key),
2275 KEY_OFFSET(&b->key), 0));
2280 ? btree(search_recurse, k, b, op)
2281 : submit_partial_cache_hit(b, op, k);
2290 static inline int keybuf_cmp(struct keybuf_key *l, struct keybuf_key *r)
2292 /* Overlapping keys compare equal */
2293 if (bkey_cmp(&l->key, &START_KEY(&r->key)) <= 0)
2295 if (bkey_cmp(&START_KEY(&l->key), &r->key) >= 0)
2300 static inline int keybuf_nonoverlapping_cmp(struct keybuf_key *l,
2301 struct keybuf_key *r)
2303 return clamp_t(int64_t, bkey_cmp(&l->key, &r->key), -1, 1);
2306 static int bch_btree_refill_keybuf(struct btree *b, struct btree_op *op,
2307 struct keybuf *buf, struct bkey *end)
2309 struct btree_iter iter;
2310 bch_btree_iter_init(b, &iter, &buf->last_scanned);
2312 while (!array_freelist_empty(&buf->freelist)) {
2313 struct bkey *k = bch_btree_iter_next_filter(&iter, b,
2318 buf->last_scanned = b->key;
2322 buf->last_scanned = *k;
2323 if (bkey_cmp(&buf->last_scanned, end) >= 0)
2326 if (buf->key_predicate(buf, k)) {
2327 struct keybuf_key *w;
2329 pr_debug("%s", pkey(k));
2331 spin_lock(&buf->lock);
2333 w = array_alloc(&buf->freelist);
2336 bkey_copy(&w->key, k);
2338 if (RB_INSERT(&buf->keys, w, node, keybuf_cmp))
2339 array_free(&buf->freelist, w);
2341 spin_unlock(&buf->lock);
2347 btree(refill_keybuf, k, b, op, buf, end);
2349 * Might get an error here, but can't really do anything
2350 * and it'll get logged elsewhere. Just read what we
2354 if (bkey_cmp(&buf->last_scanned, end) >= 0)
2364 void bch_refill_keybuf(struct cache_set *c, struct keybuf *buf,
2367 struct bkey start = buf->last_scanned;
2369 bch_btree_op_init_stack(&op);
2373 btree_root(refill_keybuf, c, &op, buf, end);
2374 closure_sync(&op.cl);
2376 pr_debug("found %s keys from %llu:%llu to %llu:%llu",
2377 RB_EMPTY_ROOT(&buf->keys) ? "no" :
2378 array_freelist_empty(&buf->freelist) ? "some" : "a few",
2379 KEY_INODE(&start), KEY_OFFSET(&start),
2380 KEY_INODE(&buf->last_scanned), KEY_OFFSET(&buf->last_scanned));
2382 spin_lock(&buf->lock);
2384 if (!RB_EMPTY_ROOT(&buf->keys)) {
2385 struct keybuf_key *w;
2386 w = RB_FIRST(&buf->keys, struct keybuf_key, node);
2387 buf->start = START_KEY(&w->key);
2389 w = RB_LAST(&buf->keys, struct keybuf_key, node);
2392 buf->start = MAX_KEY;
2396 spin_unlock(&buf->lock);
2399 static void __bch_keybuf_del(struct keybuf *buf, struct keybuf_key *w)
2401 rb_erase(&w->node, &buf->keys);
2402 array_free(&buf->freelist, w);
2405 void bch_keybuf_del(struct keybuf *buf, struct keybuf_key *w)
2407 spin_lock(&buf->lock);
2408 __bch_keybuf_del(buf, w);
2409 spin_unlock(&buf->lock);
2412 bool bch_keybuf_check_overlapping(struct keybuf *buf, struct bkey *start,
2416 struct keybuf_key *p, *w, s;
2419 if (bkey_cmp(end, &buf->start) <= 0 ||
2420 bkey_cmp(start, &buf->end) >= 0)
2423 spin_lock(&buf->lock);
2424 w = RB_GREATER(&buf->keys, s, node, keybuf_nonoverlapping_cmp);
2426 while (w && bkey_cmp(&START_KEY(&w->key), end) < 0) {
2428 w = RB_NEXT(w, node);
2433 __bch_keybuf_del(buf, p);
2436 spin_unlock(&buf->lock);
2440 struct keybuf_key *bch_keybuf_next(struct keybuf *buf)
2442 struct keybuf_key *w;
2443 spin_lock(&buf->lock);
2445 w = RB_FIRST(&buf->keys, struct keybuf_key, node);
2447 while (w && w->private)
2448 w = RB_NEXT(w, node);
2451 w->private = ERR_PTR(-EINTR);
2453 spin_unlock(&buf->lock);
2457 struct keybuf_key *bch_keybuf_next_rescan(struct cache_set *c,
2461 struct keybuf_key *ret;
2464 ret = bch_keybuf_next(buf);
2468 if (bkey_cmp(&buf->last_scanned, end) >= 0) {
2469 pr_debug("scan finished");
2473 bch_refill_keybuf(c, buf, end);
2479 void bch_keybuf_init(struct keybuf *buf, keybuf_pred_fn *fn)
2481 buf->key_predicate = fn;
2482 buf->last_scanned = MAX_KEY;
2483 buf->keys = RB_ROOT;
2485 spin_lock_init(&buf->lock);
2486 array_allocator_init(&buf->freelist);
2489 void bch_btree_exit(void)
2492 destroy_workqueue(btree_io_wq);
2494 destroy_workqueue(bch_gc_wq);
2497 int __init bch_btree_init(void)
2499 if (!(bch_gc_wq = create_singlethread_workqueue("bch_btree_gc")) ||
2500 !(btree_io_wq = create_singlethread_workqueue("bch_btree_io")))