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/bcache.txt.
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>
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
94 #define PTR_DIRTY_BIT (((uint64_t) 1 << 36))
96 #define PTR_HASH(c, k) \
97 (((k)->ptr[0] >> c->bucket_bits) | PTR_GEN(k, 0))
99 #define insert_lock(s, b) ((b)->level <= (s)->lock)
102 * These macros are for recursing down the btree - they handle the details of
103 * locking and looking up nodes in the cache for you. They're best treated as
104 * mere syntax when reading code that uses them.
106 * op->lock determines whether we take a read or a write lock at a given depth.
107 * If you've got a read lock and find that you need a write lock (i.e. you're
108 * going to have to split), set op->lock and return -EINTR; btree_root() will
109 * call you again and you'll have the correct lock.
113 * btree - recurse down the btree on a specified key
114 * @fn: function to call, which will be passed the child node
115 * @key: key to recurse on
116 * @b: parent btree node
117 * @op: pointer to struct btree_op
119 #define btree(fn, key, b, op, ...) \
121 int _r, l = (b)->level - 1; \
122 bool _w = l <= (op)->lock; \
123 struct btree *_child = bch_btree_node_get((b)->c, op, key, l, \
125 if (!IS_ERR(_child)) { \
126 _r = bch_btree_ ## fn(_child, op, ##__VA_ARGS__); \
127 rw_unlock(_w, _child); \
129 _r = PTR_ERR(_child); \
134 * btree_root - call a function on the root of the btree
135 * @fn: function to call, which will be passed the child node
137 * @op: pointer to struct btree_op
139 #define btree_root(fn, c, op, ...) \
143 struct btree *_b = (c)->root; \
144 bool _w = insert_lock(op, _b); \
145 rw_lock(_w, _b, _b->level); \
146 if (_b == (c)->root && \
147 _w == insert_lock(op, _b)) { \
148 _r = bch_btree_ ## fn(_b, op, ##__VA_ARGS__); \
151 bch_cannibalize_unlock(c); \
154 } while (_r == -EINTR); \
156 finish_wait(&(c)->btree_cache_wait, &(op)->wait); \
160 static inline struct bset *write_block(struct btree *b)
162 return ((void *) btree_bset_first(b)) + b->written * block_bytes(b->c);
165 static void bch_btree_init_next(struct btree *b)
167 /* If not a leaf node, always sort */
168 if (b->level && b->keys.nsets)
169 bch_btree_sort(&b->keys, &b->c->sort);
171 bch_btree_sort_lazy(&b->keys, &b->c->sort);
173 if (b->written < btree_blocks(b))
174 bch_bset_init_next(&b->keys, write_block(b),
175 bset_magic(&b->c->sb));
179 /* Btree key manipulation */
181 void bkey_put(struct cache_set *c, struct bkey *k)
185 for (i = 0; i < KEY_PTRS(k); i++)
186 if (ptr_available(c, k, i))
187 atomic_dec_bug(&PTR_BUCKET(c, k, i)->pin);
192 static uint64_t btree_csum_set(struct btree *b, struct bset *i)
194 uint64_t crc = b->key.ptr[0];
195 void *data = (void *) i + 8, *end = bset_bkey_last(i);
197 crc = bch_crc64_update(crc, data, end - data);
198 return crc ^ 0xffffffffffffffffULL;
201 void bch_btree_node_read_done(struct btree *b)
203 const char *err = "bad btree header";
204 struct bset *i = btree_bset_first(b);
205 struct btree_iter *iter;
207 iter = mempool_alloc(b->c->fill_iter, GFP_NOIO);
208 iter->size = b->c->sb.bucket_size / b->c->sb.block_size;
211 #ifdef CONFIG_BCACHE_DEBUG
219 b->written < btree_blocks(b) && i->seq == b->keys.set[0].data->seq;
220 i = write_block(b)) {
221 err = "unsupported bset version";
222 if (i->version > BCACHE_BSET_VERSION)
225 err = "bad btree header";
226 if (b->written + set_blocks(i, block_bytes(b->c)) >
231 if (i->magic != bset_magic(&b->c->sb))
234 err = "bad checksum";
235 switch (i->version) {
237 if (i->csum != csum_set(i))
240 case BCACHE_BSET_VERSION:
241 if (i->csum != btree_csum_set(b, i))
247 if (i != b->keys.set[0].data && !i->keys)
250 bch_btree_iter_push(iter, i->start, bset_bkey_last(i));
252 b->written += set_blocks(i, block_bytes(b->c));
255 err = "corrupted btree";
256 for (i = write_block(b);
257 bset_sector_offset(&b->keys, i) < KEY_SIZE(&b->key);
258 i = ((void *) i) + block_bytes(b->c))
259 if (i->seq == b->keys.set[0].data->seq)
262 bch_btree_sort_and_fix_extents(&b->keys, iter, &b->c->sort);
264 i = b->keys.set[0].data;
265 err = "short btree key";
266 if (b->keys.set[0].size &&
267 bkey_cmp(&b->key, &b->keys.set[0].end) < 0)
270 if (b->written < btree_blocks(b))
271 bch_bset_init_next(&b->keys, write_block(b),
272 bset_magic(&b->c->sb));
274 mempool_free(iter, b->c->fill_iter);
277 set_btree_node_io_error(b);
278 bch_cache_set_error(b->c, "%s at bucket %zu, block %u, %u keys",
279 err, PTR_BUCKET_NR(b->c, &b->key, 0),
280 bset_block_offset(b, i), i->keys);
284 static void btree_node_read_endio(struct bio *bio)
286 struct closure *cl = bio->bi_private;
290 static void bch_btree_node_read(struct btree *b)
292 uint64_t start_time = local_clock();
296 trace_bcache_btree_read(b);
298 closure_init_stack(&cl);
300 bio = bch_bbio_alloc(b->c);
301 bio->bi_iter.bi_size = KEY_SIZE(&b->key) << 9;
302 bio->bi_end_io = btree_node_read_endio;
303 bio->bi_private = &cl;
304 bio->bi_opf = REQ_OP_READ | REQ_META;
306 bch_bio_map(bio, b->keys.set[0].data);
308 bch_submit_bbio(bio, b->c, &b->key, 0);
312 set_btree_node_io_error(b);
314 bch_bbio_free(bio, b->c);
316 if (btree_node_io_error(b))
319 bch_btree_node_read_done(b);
320 bch_time_stats_update(&b->c->btree_read_time, start_time);
324 bch_cache_set_error(b->c, "io error reading bucket %zu",
325 PTR_BUCKET_NR(b->c, &b->key, 0));
328 static void btree_complete_write(struct btree *b, struct btree_write *w)
330 if (w->prio_blocked &&
331 !atomic_sub_return(w->prio_blocked, &b->c->prio_blocked))
332 wake_up_allocators(b->c);
335 atomic_dec_bug(w->journal);
336 __closure_wake_up(&b->c->journal.wait);
343 static void btree_node_write_unlock(struct closure *cl)
345 struct btree *b = container_of(cl, struct btree, io);
350 static void __btree_node_write_done(struct closure *cl)
352 struct btree *b = container_of(cl, struct btree, io);
353 struct btree_write *w = btree_prev_write(b);
355 bch_bbio_free(b->bio, b->c);
357 btree_complete_write(b, w);
359 if (btree_node_dirty(b))
360 schedule_delayed_work(&b->work, 30 * HZ);
362 closure_return_with_destructor(cl, btree_node_write_unlock);
365 static void btree_node_write_done(struct closure *cl)
367 struct btree *b = container_of(cl, struct btree, io);
369 bio_free_pages(b->bio);
370 __btree_node_write_done(cl);
373 static void btree_node_write_endio(struct bio *bio)
375 struct closure *cl = bio->bi_private;
376 struct btree *b = container_of(cl, struct btree, io);
379 set_btree_node_io_error(b);
381 bch_bbio_count_io_errors(b->c, bio, bio->bi_status, "writing btree");
385 static void do_btree_node_write(struct btree *b)
387 struct closure *cl = &b->io;
388 struct bset *i = btree_bset_last(b);
391 i->version = BCACHE_BSET_VERSION;
392 i->csum = btree_csum_set(b, i);
395 b->bio = bch_bbio_alloc(b->c);
397 b->bio->bi_end_io = btree_node_write_endio;
398 b->bio->bi_private = cl;
399 b->bio->bi_iter.bi_size = roundup(set_bytes(i), block_bytes(b->c));
400 b->bio->bi_opf = REQ_OP_WRITE | REQ_META | REQ_FUA;
401 bch_bio_map(b->bio, i);
404 * If we're appending to a leaf node, we don't technically need FUA -
405 * this write just needs to be persisted before the next journal write,
406 * which will be marked FLUSH|FUA.
408 * Similarly if we're writing a new btree root - the pointer is going to
409 * be in the next journal entry.
411 * But if we're writing a new btree node (that isn't a root) or
412 * appending to a non leaf btree node, we need either FUA or a flush
413 * when we write the parent with the new pointer. FUA is cheaper than a
414 * flush, and writes appending to leaf nodes aren't blocking anything so
415 * just make all btree node writes FUA to keep things sane.
418 bkey_copy(&k.key, &b->key);
419 SET_PTR_OFFSET(&k.key, 0, PTR_OFFSET(&k.key, 0) +
420 bset_sector_offset(&b->keys, i));
422 if (!bio_alloc_pages(b->bio, __GFP_NOWARN|GFP_NOWAIT)) {
425 void *base = (void *) ((unsigned long) i & ~(PAGE_SIZE - 1));
427 bio_for_each_segment_all(bv, b->bio, j)
428 memcpy(page_address(bv->bv_page),
429 base + j * PAGE_SIZE, PAGE_SIZE);
431 bch_submit_bbio(b->bio, b->c, &k.key, 0);
433 continue_at(cl, btree_node_write_done, NULL);
436 bch_bio_map(b->bio, i);
438 bch_submit_bbio(b->bio, b->c, &k.key, 0);
441 continue_at_nobarrier(cl, __btree_node_write_done, NULL);
445 void __bch_btree_node_write(struct btree *b, struct closure *parent)
447 struct bset *i = btree_bset_last(b);
449 lockdep_assert_held(&b->write_lock);
451 trace_bcache_btree_write(b);
453 BUG_ON(current->bio_list);
454 BUG_ON(b->written >= btree_blocks(b));
455 BUG_ON(b->written && !i->keys);
456 BUG_ON(btree_bset_first(b)->seq != i->seq);
457 bch_check_keys(&b->keys, "writing");
459 cancel_delayed_work(&b->work);
461 /* If caller isn't waiting for write, parent refcount is cache set */
463 closure_init(&b->io, parent ?: &b->c->cl);
465 clear_bit(BTREE_NODE_dirty, &b->flags);
466 change_bit(BTREE_NODE_write_idx, &b->flags);
468 do_btree_node_write(b);
470 atomic_long_add(set_blocks(i, block_bytes(b->c)) * b->c->sb.block_size,
471 &PTR_CACHE(b->c, &b->key, 0)->btree_sectors_written);
473 b->written += set_blocks(i, block_bytes(b->c));
476 void bch_btree_node_write(struct btree *b, struct closure *parent)
478 unsigned nsets = b->keys.nsets;
480 lockdep_assert_held(&b->lock);
482 __bch_btree_node_write(b, parent);
485 * do verify if there was more than one set initially (i.e. we did a
486 * sort) and we sorted down to a single set:
488 if (nsets && !b->keys.nsets)
491 bch_btree_init_next(b);
494 static void bch_btree_node_write_sync(struct btree *b)
498 closure_init_stack(&cl);
500 mutex_lock(&b->write_lock);
501 bch_btree_node_write(b, &cl);
502 mutex_unlock(&b->write_lock);
507 static void btree_node_write_work(struct work_struct *w)
509 struct btree *b = container_of(to_delayed_work(w), struct btree, work);
511 mutex_lock(&b->write_lock);
512 if (btree_node_dirty(b))
513 __bch_btree_node_write(b, NULL);
514 mutex_unlock(&b->write_lock);
517 static void bch_btree_leaf_dirty(struct btree *b, atomic_t *journal_ref)
519 struct bset *i = btree_bset_last(b);
520 struct btree_write *w = btree_current_write(b);
522 lockdep_assert_held(&b->write_lock);
527 if (!btree_node_dirty(b))
528 schedule_delayed_work(&b->work, 30 * HZ);
530 set_btree_node_dirty(b);
534 journal_pin_cmp(b->c, w->journal, journal_ref)) {
535 atomic_dec_bug(w->journal);
540 w->journal = journal_ref;
541 atomic_inc(w->journal);
545 /* Force write if set is too big */
546 if (set_bytes(i) > PAGE_SIZE - 48 &&
548 bch_btree_node_write(b, NULL);
552 * Btree in memory cache - allocation/freeing
553 * mca -> memory cache
556 #define mca_reserve(c) (((c->root && c->root->level) \
557 ? c->root->level : 1) * 8 + 16)
558 #define mca_can_free(c) \
559 max_t(int, 0, c->btree_cache_used - mca_reserve(c))
561 static void mca_data_free(struct btree *b)
563 BUG_ON(b->io_mutex.count != 1);
565 bch_btree_keys_free(&b->keys);
567 b->c->btree_cache_used--;
568 list_move(&b->list, &b->c->btree_cache_freed);
571 static void mca_bucket_free(struct btree *b)
573 BUG_ON(btree_node_dirty(b));
576 hlist_del_init_rcu(&b->hash);
577 list_move(&b->list, &b->c->btree_cache_freeable);
580 static unsigned btree_order(struct bkey *k)
582 return ilog2(KEY_SIZE(k) / PAGE_SECTORS ?: 1);
585 static void mca_data_alloc(struct btree *b, struct bkey *k, gfp_t gfp)
587 if (!bch_btree_keys_alloc(&b->keys,
589 ilog2(b->c->btree_pages),
592 b->c->btree_cache_used++;
593 list_move(&b->list, &b->c->btree_cache);
595 list_move(&b->list, &b->c->btree_cache_freed);
599 static struct btree *mca_bucket_alloc(struct cache_set *c,
600 struct bkey *k, gfp_t gfp)
602 struct btree *b = kzalloc(sizeof(struct btree), gfp);
606 init_rwsem(&b->lock);
607 lockdep_set_novalidate_class(&b->lock);
608 mutex_init(&b->write_lock);
609 lockdep_set_novalidate_class(&b->write_lock);
610 INIT_LIST_HEAD(&b->list);
611 INIT_DELAYED_WORK(&b->work, btree_node_write_work);
613 sema_init(&b->io_mutex, 1);
615 mca_data_alloc(b, k, gfp);
619 static int mca_reap(struct btree *b, unsigned min_order, bool flush)
623 closure_init_stack(&cl);
624 lockdep_assert_held(&b->c->bucket_lock);
626 if (!down_write_trylock(&b->lock))
629 BUG_ON(btree_node_dirty(b) && !b->keys.set[0].data);
631 if (b->keys.page_order < min_order)
635 if (btree_node_dirty(b))
638 if (down_trylock(&b->io_mutex))
643 mutex_lock(&b->write_lock);
644 if (btree_node_dirty(b))
645 __bch_btree_node_write(b, &cl);
646 mutex_unlock(&b->write_lock);
650 /* wait for any in flight btree write */
660 static unsigned long bch_mca_scan(struct shrinker *shrink,
661 struct shrink_control *sc)
663 struct cache_set *c = container_of(shrink, struct cache_set, shrink);
665 unsigned long i, nr = sc->nr_to_scan;
666 unsigned long freed = 0;
668 if (c->shrinker_disabled)
671 if (c->btree_cache_alloc_lock)
674 /* Return -1 if we can't do anything right now */
675 if (sc->gfp_mask & __GFP_IO)
676 mutex_lock(&c->bucket_lock);
677 else if (!mutex_trylock(&c->bucket_lock))
681 * It's _really_ critical that we don't free too many btree nodes - we
682 * have to always leave ourselves a reserve. The reserve is how we
683 * guarantee that allocating memory for a new btree node can always
684 * succeed, so that inserting keys into the btree can always succeed and
685 * IO can always make forward progress:
687 nr /= c->btree_pages;
688 nr = min_t(unsigned long, nr, mca_can_free(c));
691 list_for_each_entry_safe(b, t, &c->btree_cache_freeable, list) {
696 !mca_reap(b, 0, false)) {
703 for (i = 0; (nr--) && i < c->btree_cache_used; i++) {
704 if (list_empty(&c->btree_cache))
707 b = list_first_entry(&c->btree_cache, struct btree, list);
708 list_rotate_left(&c->btree_cache);
711 !mca_reap(b, 0, false)) {
720 mutex_unlock(&c->bucket_lock);
724 static unsigned long bch_mca_count(struct shrinker *shrink,
725 struct shrink_control *sc)
727 struct cache_set *c = container_of(shrink, struct cache_set, shrink);
729 if (c->shrinker_disabled)
732 if (c->btree_cache_alloc_lock)
735 return mca_can_free(c) * c->btree_pages;
738 void bch_btree_cache_free(struct cache_set *c)
742 closure_init_stack(&cl);
744 if (c->shrink.list.next)
745 unregister_shrinker(&c->shrink);
747 mutex_lock(&c->bucket_lock);
749 #ifdef CONFIG_BCACHE_DEBUG
751 list_move(&c->verify_data->list, &c->btree_cache);
753 free_pages((unsigned long) c->verify_ondisk, ilog2(bucket_pages(c)));
756 list_splice(&c->btree_cache_freeable,
759 while (!list_empty(&c->btree_cache)) {
760 b = list_first_entry(&c->btree_cache, struct btree, list);
762 if (btree_node_dirty(b))
763 btree_complete_write(b, btree_current_write(b));
764 clear_bit(BTREE_NODE_dirty, &b->flags);
769 while (!list_empty(&c->btree_cache_freed)) {
770 b = list_first_entry(&c->btree_cache_freed,
773 cancel_delayed_work_sync(&b->work);
777 mutex_unlock(&c->bucket_lock);
780 int bch_btree_cache_alloc(struct cache_set *c)
784 for (i = 0; i < mca_reserve(c); i++)
785 if (!mca_bucket_alloc(c, &ZERO_KEY, GFP_KERNEL))
788 list_splice_init(&c->btree_cache,
789 &c->btree_cache_freeable);
791 #ifdef CONFIG_BCACHE_DEBUG
792 mutex_init(&c->verify_lock);
794 c->verify_ondisk = (void *)
795 __get_free_pages(GFP_KERNEL, ilog2(bucket_pages(c)));
797 c->verify_data = mca_bucket_alloc(c, &ZERO_KEY, GFP_KERNEL);
799 if (c->verify_data &&
800 c->verify_data->keys.set->data)
801 list_del_init(&c->verify_data->list);
803 c->verify_data = NULL;
806 c->shrink.count_objects = bch_mca_count;
807 c->shrink.scan_objects = bch_mca_scan;
809 c->shrink.batch = c->btree_pages * 2;
811 if (register_shrinker(&c->shrink))
812 pr_warn("bcache: %s: could not register shrinker",
818 /* Btree in memory cache - hash table */
820 static struct hlist_head *mca_hash(struct cache_set *c, struct bkey *k)
822 return &c->bucket_hash[hash_32(PTR_HASH(c, k), BUCKET_HASH_BITS)];
825 static struct btree *mca_find(struct cache_set *c, struct bkey *k)
830 hlist_for_each_entry_rcu(b, mca_hash(c, k), hash)
831 if (PTR_HASH(c, &b->key) == PTR_HASH(c, k))
839 static int mca_cannibalize_lock(struct cache_set *c, struct btree_op *op)
841 struct task_struct *old;
843 old = cmpxchg(&c->btree_cache_alloc_lock, NULL, current);
844 if (old && old != current) {
846 prepare_to_wait(&c->btree_cache_wait, &op->wait,
847 TASK_UNINTERRUPTIBLE);
854 static struct btree *mca_cannibalize(struct cache_set *c, struct btree_op *op,
859 trace_bcache_btree_cache_cannibalize(c);
861 if (mca_cannibalize_lock(c, op))
862 return ERR_PTR(-EINTR);
864 list_for_each_entry_reverse(b, &c->btree_cache, list)
865 if (!mca_reap(b, btree_order(k), false))
868 list_for_each_entry_reverse(b, &c->btree_cache, list)
869 if (!mca_reap(b, btree_order(k), true))
872 WARN(1, "btree cache cannibalize failed\n");
873 return ERR_PTR(-ENOMEM);
877 * We can only have one thread cannibalizing other cached btree nodes at a time,
878 * or we'll deadlock. We use an open coded mutex to ensure that, which a
879 * cannibalize_bucket() will take. This means every time we unlock the root of
880 * the btree, we need to release this lock if we have it held.
882 static void bch_cannibalize_unlock(struct cache_set *c)
884 if (c->btree_cache_alloc_lock == current) {
885 c->btree_cache_alloc_lock = NULL;
886 wake_up(&c->btree_cache_wait);
890 static struct btree *mca_alloc(struct cache_set *c, struct btree_op *op,
891 struct bkey *k, int level)
895 BUG_ON(current->bio_list);
897 lockdep_assert_held(&c->bucket_lock);
902 /* btree_free() doesn't free memory; it sticks the node on the end of
903 * the list. Check if there's any freed nodes there:
905 list_for_each_entry(b, &c->btree_cache_freeable, list)
906 if (!mca_reap(b, btree_order(k), false))
909 /* We never free struct btree itself, just the memory that holds the on
910 * disk node. Check the freed list before allocating a new one:
912 list_for_each_entry(b, &c->btree_cache_freed, list)
913 if (!mca_reap(b, 0, false)) {
914 mca_data_alloc(b, k, __GFP_NOWARN|GFP_NOIO);
915 if (!b->keys.set[0].data)
921 b = mca_bucket_alloc(c, k, __GFP_NOWARN|GFP_NOIO);
925 BUG_ON(!down_write_trylock(&b->lock));
926 if (!b->keys.set->data)
929 BUG_ON(b->io_mutex.count != 1);
931 bkey_copy(&b->key, k);
932 list_move(&b->list, &c->btree_cache);
933 hlist_del_init_rcu(&b->hash);
934 hlist_add_head_rcu(&b->hash, mca_hash(c, k));
936 lock_set_subclass(&b->lock.dep_map, level + 1, _THIS_IP_);
937 b->parent = (void *) ~0UL;
943 bch_btree_keys_init(&b->keys, &bch_extent_keys_ops,
944 &b->c->expensive_debug_checks);
946 bch_btree_keys_init(&b->keys, &bch_btree_keys_ops,
947 &b->c->expensive_debug_checks);
954 b = mca_cannibalize(c, op, k);
962 * bch_btree_node_get - find a btree node in the cache and lock it, reading it
963 * in from disk if necessary.
965 * If IO is necessary and running under generic_make_request, returns -EAGAIN.
967 * The btree node will have either a read or a write lock held, depending on
968 * level and op->lock.
970 struct btree *bch_btree_node_get(struct cache_set *c, struct btree_op *op,
971 struct bkey *k, int level, bool write,
972 struct btree *parent)
982 if (current->bio_list)
983 return ERR_PTR(-EAGAIN);
985 mutex_lock(&c->bucket_lock);
986 b = mca_alloc(c, op, k, level);
987 mutex_unlock(&c->bucket_lock);
994 bch_btree_node_read(b);
997 downgrade_write(&b->lock);
999 rw_lock(write, b, level);
1000 if (PTR_HASH(c, &b->key) != PTR_HASH(c, k)) {
1001 rw_unlock(write, b);
1004 BUG_ON(b->level != level);
1010 for (; i <= b->keys.nsets && b->keys.set[i].size; i++) {
1011 prefetch(b->keys.set[i].tree);
1012 prefetch(b->keys.set[i].data);
1015 for (; i <= b->keys.nsets; i++)
1016 prefetch(b->keys.set[i].data);
1018 if (btree_node_io_error(b)) {
1019 rw_unlock(write, b);
1020 return ERR_PTR(-EIO);
1023 BUG_ON(!b->written);
1028 static void btree_node_prefetch(struct btree *parent, struct bkey *k)
1032 mutex_lock(&parent->c->bucket_lock);
1033 b = mca_alloc(parent->c, NULL, k, parent->level - 1);
1034 mutex_unlock(&parent->c->bucket_lock);
1036 if (!IS_ERR_OR_NULL(b)) {
1038 bch_btree_node_read(b);
1045 static void btree_node_free(struct btree *b)
1047 trace_bcache_btree_node_free(b);
1049 BUG_ON(b == b->c->root);
1051 mutex_lock(&b->write_lock);
1053 if (btree_node_dirty(b))
1054 btree_complete_write(b, btree_current_write(b));
1055 clear_bit(BTREE_NODE_dirty, &b->flags);
1057 mutex_unlock(&b->write_lock);
1059 cancel_delayed_work(&b->work);
1061 mutex_lock(&b->c->bucket_lock);
1062 bch_bucket_free(b->c, &b->key);
1064 mutex_unlock(&b->c->bucket_lock);
1067 struct btree *__bch_btree_node_alloc(struct cache_set *c, struct btree_op *op,
1068 int level, bool wait,
1069 struct btree *parent)
1072 struct btree *b = ERR_PTR(-EAGAIN);
1074 mutex_lock(&c->bucket_lock);
1076 if (__bch_bucket_alloc_set(c, RESERVE_BTREE, &k.key, 1, wait))
1079 bkey_put(c, &k.key);
1080 SET_KEY_SIZE(&k.key, c->btree_pages * PAGE_SECTORS);
1082 b = mca_alloc(c, op, &k.key, level);
1088 "Tried to allocate bucket that was in btree cache");
1094 bch_bset_init_next(&b->keys, b->keys.set->data, bset_magic(&b->c->sb));
1096 mutex_unlock(&c->bucket_lock);
1098 trace_bcache_btree_node_alloc(b);
1101 bch_bucket_free(c, &k.key);
1103 mutex_unlock(&c->bucket_lock);
1105 trace_bcache_btree_node_alloc_fail(c);
1109 static struct btree *bch_btree_node_alloc(struct cache_set *c,
1110 struct btree_op *op, int level,
1111 struct btree *parent)
1113 return __bch_btree_node_alloc(c, op, level, op != NULL, parent);
1116 static struct btree *btree_node_alloc_replacement(struct btree *b,
1117 struct btree_op *op)
1119 struct btree *n = bch_btree_node_alloc(b->c, op, b->level, b->parent);
1120 if (!IS_ERR_OR_NULL(n)) {
1121 mutex_lock(&n->write_lock);
1122 bch_btree_sort_into(&b->keys, &n->keys, &b->c->sort);
1123 bkey_copy_key(&n->key, &b->key);
1124 mutex_unlock(&n->write_lock);
1130 static void make_btree_freeing_key(struct btree *b, struct bkey *k)
1134 mutex_lock(&b->c->bucket_lock);
1136 atomic_inc(&b->c->prio_blocked);
1138 bkey_copy(k, &b->key);
1139 bkey_copy_key(k, &ZERO_KEY);
1141 for (i = 0; i < KEY_PTRS(k); i++)
1143 bch_inc_gen(PTR_CACHE(b->c, &b->key, i),
1144 PTR_BUCKET(b->c, &b->key, i)));
1146 mutex_unlock(&b->c->bucket_lock);
1149 static int btree_check_reserve(struct btree *b, struct btree_op *op)
1151 struct cache_set *c = b->c;
1153 unsigned i, reserve = (c->root->level - b->level) * 2 + 1;
1155 mutex_lock(&c->bucket_lock);
1157 for_each_cache(ca, c, i)
1158 if (fifo_used(&ca->free[RESERVE_BTREE]) < reserve) {
1160 prepare_to_wait(&c->btree_cache_wait, &op->wait,
1161 TASK_UNINTERRUPTIBLE);
1162 mutex_unlock(&c->bucket_lock);
1166 mutex_unlock(&c->bucket_lock);
1168 return mca_cannibalize_lock(b->c, op);
1171 /* Garbage collection */
1173 static uint8_t __bch_btree_mark_key(struct cache_set *c, int level,
1181 * ptr_invalid() can't return true for the keys that mark btree nodes as
1182 * freed, but since ptr_bad() returns true we'll never actually use them
1183 * for anything and thus we don't want mark their pointers here
1185 if (!bkey_cmp(k, &ZERO_KEY))
1188 for (i = 0; i < KEY_PTRS(k); i++) {
1189 if (!ptr_available(c, k, i))
1192 g = PTR_BUCKET(c, k, i);
1194 if (gen_after(g->last_gc, PTR_GEN(k, i)))
1195 g->last_gc = PTR_GEN(k, i);
1197 if (ptr_stale(c, k, i)) {
1198 stale = max(stale, ptr_stale(c, k, i));
1202 cache_bug_on(GC_MARK(g) &&
1203 (GC_MARK(g) == GC_MARK_METADATA) != (level != 0),
1204 c, "inconsistent ptrs: mark = %llu, level = %i",
1208 SET_GC_MARK(g, GC_MARK_METADATA);
1209 else if (KEY_DIRTY(k))
1210 SET_GC_MARK(g, GC_MARK_DIRTY);
1211 else if (!GC_MARK(g))
1212 SET_GC_MARK(g, GC_MARK_RECLAIMABLE);
1214 /* guard against overflow */
1215 SET_GC_SECTORS_USED(g, min_t(unsigned,
1216 GC_SECTORS_USED(g) + KEY_SIZE(k),
1217 MAX_GC_SECTORS_USED));
1219 BUG_ON(!GC_SECTORS_USED(g));
1225 #define btree_mark_key(b, k) __bch_btree_mark_key(b->c, b->level, k)
1227 void bch_initial_mark_key(struct cache_set *c, int level, struct bkey *k)
1231 for (i = 0; i < KEY_PTRS(k); i++)
1232 if (ptr_available(c, k, i) &&
1233 !ptr_stale(c, k, i)) {
1234 struct bucket *b = PTR_BUCKET(c, k, i);
1236 b->gen = PTR_GEN(k, i);
1238 if (level && bkey_cmp(k, &ZERO_KEY))
1239 b->prio = BTREE_PRIO;
1240 else if (!level && b->prio == BTREE_PRIO)
1241 b->prio = INITIAL_PRIO;
1244 __bch_btree_mark_key(c, level, k);
1247 void bch_update_bucket_in_use(struct cache_set *c, struct gc_stat *stats)
1249 stats->in_use = (c->nbuckets - c->avail_nbuckets) * 100 / c->nbuckets;
1252 static bool btree_gc_mark_node(struct btree *b, struct gc_stat *gc)
1255 unsigned keys = 0, good_keys = 0;
1257 struct btree_iter iter;
1258 struct bset_tree *t;
1262 for_each_key_filter(&b->keys, k, &iter, bch_ptr_invalid) {
1263 stale = max(stale, btree_mark_key(b, k));
1266 if (bch_ptr_bad(&b->keys, k))
1269 gc->key_bytes += bkey_u64s(k);
1273 gc->data += KEY_SIZE(k);
1276 for (t = b->keys.set; t <= &b->keys.set[b->keys.nsets]; t++)
1277 btree_bug_on(t->size &&
1278 bset_written(&b->keys, t) &&
1279 bkey_cmp(&b->key, &t->end) < 0,
1280 b, "found short btree key in gc");
1282 if (b->c->gc_always_rewrite)
1288 if ((keys - good_keys) * 2 > keys)
1294 #define GC_MERGE_NODES 4U
1296 struct gc_merge_info {
1301 static int bch_btree_insert_node(struct btree *, struct btree_op *,
1302 struct keylist *, atomic_t *, struct bkey *);
1304 static int btree_gc_coalesce(struct btree *b, struct btree_op *op,
1305 struct gc_stat *gc, struct gc_merge_info *r)
1307 unsigned i, nodes = 0, keys = 0, blocks;
1308 struct btree *new_nodes[GC_MERGE_NODES];
1309 struct keylist keylist;
1313 bch_keylist_init(&keylist);
1315 if (btree_check_reserve(b, NULL))
1318 memset(new_nodes, 0, sizeof(new_nodes));
1319 closure_init_stack(&cl);
1321 while (nodes < GC_MERGE_NODES && !IS_ERR_OR_NULL(r[nodes].b))
1322 keys += r[nodes++].keys;
1324 blocks = btree_default_blocks(b->c) * 2 / 3;
1327 __set_blocks(b->keys.set[0].data, keys,
1328 block_bytes(b->c)) > blocks * (nodes - 1))
1331 for (i = 0; i < nodes; i++) {
1332 new_nodes[i] = btree_node_alloc_replacement(r[i].b, NULL);
1333 if (IS_ERR_OR_NULL(new_nodes[i]))
1334 goto out_nocoalesce;
1338 * We have to check the reserve here, after we've allocated our new
1339 * nodes, to make sure the insert below will succeed - we also check
1340 * before as an optimization to potentially avoid a bunch of expensive
1343 if (btree_check_reserve(b, NULL))
1344 goto out_nocoalesce;
1346 for (i = 0; i < nodes; i++)
1347 mutex_lock(&new_nodes[i]->write_lock);
1349 for (i = nodes - 1; i > 0; --i) {
1350 struct bset *n1 = btree_bset_first(new_nodes[i]);
1351 struct bset *n2 = btree_bset_first(new_nodes[i - 1]);
1352 struct bkey *k, *last = NULL;
1358 k < bset_bkey_last(n2);
1360 if (__set_blocks(n1, n1->keys + keys +
1362 block_bytes(b->c)) > blocks)
1366 keys += bkey_u64s(k);
1370 * Last node we're not getting rid of - we're getting
1371 * rid of the node at r[0]. Have to try and fit all of
1372 * the remaining keys into this node; we can't ensure
1373 * they will always fit due to rounding and variable
1374 * length keys (shouldn't be possible in practice,
1377 if (__set_blocks(n1, n1->keys + n2->keys,
1378 block_bytes(b->c)) >
1379 btree_blocks(new_nodes[i]))
1380 goto out_nocoalesce;
1383 /* Take the key of the node we're getting rid of */
1387 BUG_ON(__set_blocks(n1, n1->keys + keys, block_bytes(b->c)) >
1388 btree_blocks(new_nodes[i]));
1391 bkey_copy_key(&new_nodes[i]->key, last);
1393 memcpy(bset_bkey_last(n1),
1395 (void *) bset_bkey_idx(n2, keys) - (void *) n2->start);
1398 r[i].keys = n1->keys;
1401 bset_bkey_idx(n2, keys),
1402 (void *) bset_bkey_last(n2) -
1403 (void *) bset_bkey_idx(n2, keys));
1407 if (__bch_keylist_realloc(&keylist,
1408 bkey_u64s(&new_nodes[i]->key)))
1409 goto out_nocoalesce;
1411 bch_btree_node_write(new_nodes[i], &cl);
1412 bch_keylist_add(&keylist, &new_nodes[i]->key);
1415 for (i = 0; i < nodes; i++)
1416 mutex_unlock(&new_nodes[i]->write_lock);
1420 /* We emptied out this node */
1421 BUG_ON(btree_bset_first(new_nodes[0])->keys);
1422 btree_node_free(new_nodes[0]);
1423 rw_unlock(true, new_nodes[0]);
1424 new_nodes[0] = NULL;
1426 for (i = 0; i < nodes; i++) {
1427 if (__bch_keylist_realloc(&keylist, bkey_u64s(&r[i].b->key)))
1428 goto out_nocoalesce;
1430 make_btree_freeing_key(r[i].b, keylist.top);
1431 bch_keylist_push(&keylist);
1434 bch_btree_insert_node(b, op, &keylist, NULL, NULL);
1435 BUG_ON(!bch_keylist_empty(&keylist));
1437 for (i = 0; i < nodes; i++) {
1438 btree_node_free(r[i].b);
1439 rw_unlock(true, r[i].b);
1441 r[i].b = new_nodes[i];
1444 memmove(r, r + 1, sizeof(r[0]) * (nodes - 1));
1445 r[nodes - 1].b = ERR_PTR(-EINTR);
1447 trace_bcache_btree_gc_coalesce(nodes);
1450 bch_keylist_free(&keylist);
1452 /* Invalidated our iterator */
1457 bch_keylist_free(&keylist);
1459 while ((k = bch_keylist_pop(&keylist)))
1460 if (!bkey_cmp(k, &ZERO_KEY))
1461 atomic_dec(&b->c->prio_blocked);
1463 for (i = 0; i < nodes; i++)
1464 if (!IS_ERR_OR_NULL(new_nodes[i])) {
1465 btree_node_free(new_nodes[i]);
1466 rw_unlock(true, new_nodes[i]);
1471 static int btree_gc_rewrite_node(struct btree *b, struct btree_op *op,
1472 struct btree *replace)
1474 struct keylist keys;
1477 if (btree_check_reserve(b, NULL))
1480 n = btree_node_alloc_replacement(replace, NULL);
1482 /* recheck reserve after allocating replacement node */
1483 if (btree_check_reserve(b, NULL)) {
1489 bch_btree_node_write_sync(n);
1491 bch_keylist_init(&keys);
1492 bch_keylist_add(&keys, &n->key);
1494 make_btree_freeing_key(replace, keys.top);
1495 bch_keylist_push(&keys);
1497 bch_btree_insert_node(b, op, &keys, NULL, NULL);
1498 BUG_ON(!bch_keylist_empty(&keys));
1500 btree_node_free(replace);
1503 /* Invalidated our iterator */
1507 static unsigned btree_gc_count_keys(struct btree *b)
1510 struct btree_iter iter;
1513 for_each_key_filter(&b->keys, k, &iter, bch_ptr_bad)
1514 ret += bkey_u64s(k);
1519 static int btree_gc_recurse(struct btree *b, struct btree_op *op,
1520 struct closure *writes, struct gc_stat *gc)
1523 bool should_rewrite;
1525 struct btree_iter iter;
1526 struct gc_merge_info r[GC_MERGE_NODES];
1527 struct gc_merge_info *i, *last = r + ARRAY_SIZE(r) - 1;
1529 bch_btree_iter_init(&b->keys, &iter, &b->c->gc_done);
1531 for (i = r; i < r + ARRAY_SIZE(r); i++)
1532 i->b = ERR_PTR(-EINTR);
1535 k = bch_btree_iter_next_filter(&iter, &b->keys, bch_ptr_bad);
1537 r->b = bch_btree_node_get(b->c, op, k, b->level - 1,
1540 ret = PTR_ERR(r->b);
1544 r->keys = btree_gc_count_keys(r->b);
1546 ret = btree_gc_coalesce(b, op, gc, r);
1554 if (!IS_ERR(last->b)) {
1555 should_rewrite = btree_gc_mark_node(last->b, gc);
1556 if (should_rewrite) {
1557 ret = btree_gc_rewrite_node(b, op, last->b);
1562 if (last->b->level) {
1563 ret = btree_gc_recurse(last->b, op, writes, gc);
1568 bkey_copy_key(&b->c->gc_done, &last->b->key);
1571 * Must flush leaf nodes before gc ends, since replace
1572 * operations aren't journalled
1574 mutex_lock(&last->b->write_lock);
1575 if (btree_node_dirty(last->b))
1576 bch_btree_node_write(last->b, writes);
1577 mutex_unlock(&last->b->write_lock);
1578 rw_unlock(true, last->b);
1581 memmove(r + 1, r, sizeof(r[0]) * (GC_MERGE_NODES - 1));
1584 if (need_resched()) {
1590 for (i = r; i < r + ARRAY_SIZE(r); i++)
1591 if (!IS_ERR_OR_NULL(i->b)) {
1592 mutex_lock(&i->b->write_lock);
1593 if (btree_node_dirty(i->b))
1594 bch_btree_node_write(i->b, writes);
1595 mutex_unlock(&i->b->write_lock);
1596 rw_unlock(true, i->b);
1602 static int bch_btree_gc_root(struct btree *b, struct btree_op *op,
1603 struct closure *writes, struct gc_stat *gc)
1605 struct btree *n = NULL;
1607 bool should_rewrite;
1609 should_rewrite = btree_gc_mark_node(b, gc);
1610 if (should_rewrite) {
1611 n = btree_node_alloc_replacement(b, NULL);
1613 if (!IS_ERR_OR_NULL(n)) {
1614 bch_btree_node_write_sync(n);
1616 bch_btree_set_root(n);
1624 __bch_btree_mark_key(b->c, b->level + 1, &b->key);
1627 ret = btree_gc_recurse(b, op, writes, gc);
1632 bkey_copy_key(&b->c->gc_done, &b->key);
1637 static void btree_gc_start(struct cache_set *c)
1643 if (!c->gc_mark_valid)
1646 mutex_lock(&c->bucket_lock);
1648 c->gc_mark_valid = 0;
1649 c->gc_done = ZERO_KEY;
1651 for_each_cache(ca, c, i)
1652 for_each_bucket(b, ca) {
1653 b->last_gc = b->gen;
1654 if (!atomic_read(&b->pin)) {
1656 SET_GC_SECTORS_USED(b, 0);
1660 mutex_unlock(&c->bucket_lock);
1663 static void bch_btree_gc_finish(struct cache_set *c)
1669 mutex_lock(&c->bucket_lock);
1672 c->gc_mark_valid = 1;
1675 for (i = 0; i < KEY_PTRS(&c->uuid_bucket); i++)
1676 SET_GC_MARK(PTR_BUCKET(c, &c->uuid_bucket, i),
1679 /* don't reclaim buckets to which writeback keys point */
1681 for (i = 0; i < c->nr_uuids; i++) {
1682 struct bcache_device *d = c->devices[i];
1683 struct cached_dev *dc;
1684 struct keybuf_key *w, *n;
1687 if (!d || UUID_FLASH_ONLY(&c->uuids[i]))
1689 dc = container_of(d, struct cached_dev, disk);
1691 spin_lock(&dc->writeback_keys.lock);
1692 rbtree_postorder_for_each_entry_safe(w, n,
1693 &dc->writeback_keys.keys, node)
1694 for (j = 0; j < KEY_PTRS(&w->key); j++)
1695 SET_GC_MARK(PTR_BUCKET(c, &w->key, j),
1697 spin_unlock(&dc->writeback_keys.lock);
1701 c->avail_nbuckets = 0;
1702 for_each_cache(ca, c, i) {
1705 ca->invalidate_needs_gc = 0;
1707 for (i = ca->sb.d; i < ca->sb.d + ca->sb.keys; i++)
1708 SET_GC_MARK(ca->buckets + *i, GC_MARK_METADATA);
1710 for (i = ca->prio_buckets;
1711 i < ca->prio_buckets + prio_buckets(ca) * 2; i++)
1712 SET_GC_MARK(ca->buckets + *i, GC_MARK_METADATA);
1714 for_each_bucket(b, ca) {
1715 c->need_gc = max(c->need_gc, bucket_gc_gen(b));
1717 if (atomic_read(&b->pin))
1720 BUG_ON(!GC_MARK(b) && GC_SECTORS_USED(b));
1722 if (!GC_MARK(b) || GC_MARK(b) == GC_MARK_RECLAIMABLE)
1723 c->avail_nbuckets++;
1727 mutex_unlock(&c->bucket_lock);
1730 static void bch_btree_gc(struct cache_set *c)
1733 struct gc_stat stats;
1734 struct closure writes;
1736 uint64_t start_time = local_clock();
1738 trace_bcache_gc_start(c);
1740 memset(&stats, 0, sizeof(struct gc_stat));
1741 closure_init_stack(&writes);
1742 bch_btree_op_init(&op, SHRT_MAX);
1747 ret = btree_root(gc_root, c, &op, &writes, &stats);
1748 closure_sync(&writes);
1751 if (ret && ret != -EAGAIN)
1752 pr_warn("gc failed!");
1755 bch_btree_gc_finish(c);
1756 wake_up_allocators(c);
1758 bch_time_stats_update(&c->btree_gc_time, start_time);
1760 stats.key_bytes *= sizeof(uint64_t);
1762 bch_update_bucket_in_use(c, &stats);
1763 memcpy(&c->gc_stats, &stats, sizeof(struct gc_stat));
1765 trace_bcache_gc_end(c);
1770 static bool gc_should_run(struct cache_set *c)
1775 for_each_cache(ca, c, i)
1776 if (ca->invalidate_needs_gc)
1779 if (atomic_read(&c->sectors_to_gc) < 0)
1785 static int bch_gc_thread(void *arg)
1787 struct cache_set *c = arg;
1790 wait_event_interruptible(c->gc_wait,
1791 kthread_should_stop() || gc_should_run(c));
1793 if (kthread_should_stop())
1803 int bch_gc_thread_start(struct cache_set *c)
1805 c->gc_thread = kthread_run(bch_gc_thread, c, "bcache_gc");
1806 if (IS_ERR(c->gc_thread))
1807 return PTR_ERR(c->gc_thread);
1812 /* Initial partial gc */
1814 static int bch_btree_check_recurse(struct btree *b, struct btree_op *op)
1817 struct bkey *k, *p = NULL;
1818 struct btree_iter iter;
1820 for_each_key_filter(&b->keys, k, &iter, bch_ptr_invalid)
1821 bch_initial_mark_key(b->c, b->level, k);
1823 bch_initial_mark_key(b->c, b->level + 1, &b->key);
1826 bch_btree_iter_init(&b->keys, &iter, NULL);
1829 k = bch_btree_iter_next_filter(&iter, &b->keys,
1832 btree_node_prefetch(b, k);
1835 ret = btree(check_recurse, p, b, op);
1838 } while (p && !ret);
1844 int bch_btree_check(struct cache_set *c)
1848 bch_btree_op_init(&op, SHRT_MAX);
1850 return btree_root(check_recurse, c, &op);
1853 void bch_initial_gc_finish(struct cache_set *c)
1859 bch_btree_gc_finish(c);
1861 mutex_lock(&c->bucket_lock);
1864 * We need to put some unused buckets directly on the prio freelist in
1865 * order to get the allocator thread started - it needs freed buckets in
1866 * order to rewrite the prios and gens, and it needs to rewrite prios
1867 * and gens in order to free buckets.
1869 * This is only safe for buckets that have no live data in them, which
1870 * there should always be some of.
1872 for_each_cache(ca, c, i) {
1873 for_each_bucket(b, ca) {
1874 if (fifo_full(&ca->free[RESERVE_PRIO]))
1877 if (bch_can_invalidate_bucket(ca, b) &&
1879 __bch_invalidate_one_bucket(ca, b);
1880 fifo_push(&ca->free[RESERVE_PRIO],
1886 mutex_unlock(&c->bucket_lock);
1889 /* Btree insertion */
1891 static bool btree_insert_key(struct btree *b, struct bkey *k,
1892 struct bkey *replace_key)
1896 BUG_ON(bkey_cmp(k, &b->key) > 0);
1898 status = bch_btree_insert_key(&b->keys, k, replace_key);
1899 if (status != BTREE_INSERT_STATUS_NO_INSERT) {
1900 bch_check_keys(&b->keys, "%u for %s", status,
1901 replace_key ? "replace" : "insert");
1903 trace_bcache_btree_insert_key(b, k, replace_key != NULL,
1910 static size_t insert_u64s_remaining(struct btree *b)
1912 long ret = bch_btree_keys_u64s_remaining(&b->keys);
1915 * Might land in the middle of an existing extent and have to split it
1917 if (b->keys.ops->is_extents)
1918 ret -= KEY_MAX_U64S;
1920 return max(ret, 0L);
1923 static bool bch_btree_insert_keys(struct btree *b, struct btree_op *op,
1924 struct keylist *insert_keys,
1925 struct bkey *replace_key)
1928 int oldsize = bch_count_data(&b->keys);
1930 while (!bch_keylist_empty(insert_keys)) {
1931 struct bkey *k = insert_keys->keys;
1933 if (bkey_u64s(k) > insert_u64s_remaining(b))
1936 if (bkey_cmp(k, &b->key) <= 0) {
1940 ret |= btree_insert_key(b, k, replace_key);
1941 bch_keylist_pop_front(insert_keys);
1942 } else if (bkey_cmp(&START_KEY(k), &b->key) < 0) {
1943 BKEY_PADDED(key) temp;
1944 bkey_copy(&temp.key, insert_keys->keys);
1946 bch_cut_back(&b->key, &temp.key);
1947 bch_cut_front(&b->key, insert_keys->keys);
1949 ret |= btree_insert_key(b, &temp.key, replace_key);
1957 op->insert_collision = true;
1959 BUG_ON(!bch_keylist_empty(insert_keys) && b->level);
1961 BUG_ON(bch_count_data(&b->keys) < oldsize);
1965 static int btree_split(struct btree *b, struct btree_op *op,
1966 struct keylist *insert_keys,
1967 struct bkey *replace_key)
1970 struct btree *n1, *n2 = NULL, *n3 = NULL;
1971 uint64_t start_time = local_clock();
1973 struct keylist parent_keys;
1975 closure_init_stack(&cl);
1976 bch_keylist_init(&parent_keys);
1978 if (btree_check_reserve(b, op)) {
1982 WARN(1, "insufficient reserve for split\n");
1985 n1 = btree_node_alloc_replacement(b, op);
1989 split = set_blocks(btree_bset_first(n1),
1990 block_bytes(n1->c)) > (btree_blocks(b) * 4) / 5;
1995 trace_bcache_btree_node_split(b, btree_bset_first(n1)->keys);
1997 n2 = bch_btree_node_alloc(b->c, op, b->level, b->parent);
2002 n3 = bch_btree_node_alloc(b->c, op, b->level + 1, NULL);
2007 mutex_lock(&n1->write_lock);
2008 mutex_lock(&n2->write_lock);
2010 bch_btree_insert_keys(n1, op, insert_keys, replace_key);
2013 * Has to be a linear search because we don't have an auxiliary
2017 while (keys < (btree_bset_first(n1)->keys * 3) / 5)
2018 keys += bkey_u64s(bset_bkey_idx(btree_bset_first(n1),
2021 bkey_copy_key(&n1->key,
2022 bset_bkey_idx(btree_bset_first(n1), keys));
2023 keys += bkey_u64s(bset_bkey_idx(btree_bset_first(n1), keys));
2025 btree_bset_first(n2)->keys = btree_bset_first(n1)->keys - keys;
2026 btree_bset_first(n1)->keys = keys;
2028 memcpy(btree_bset_first(n2)->start,
2029 bset_bkey_last(btree_bset_first(n1)),
2030 btree_bset_first(n2)->keys * sizeof(uint64_t));
2032 bkey_copy_key(&n2->key, &b->key);
2034 bch_keylist_add(&parent_keys, &n2->key);
2035 bch_btree_node_write(n2, &cl);
2036 mutex_unlock(&n2->write_lock);
2037 rw_unlock(true, n2);
2039 trace_bcache_btree_node_compact(b, btree_bset_first(n1)->keys);
2041 mutex_lock(&n1->write_lock);
2042 bch_btree_insert_keys(n1, op, insert_keys, replace_key);
2045 bch_keylist_add(&parent_keys, &n1->key);
2046 bch_btree_node_write(n1, &cl);
2047 mutex_unlock(&n1->write_lock);
2050 /* Depth increases, make a new root */
2051 mutex_lock(&n3->write_lock);
2052 bkey_copy_key(&n3->key, &MAX_KEY);
2053 bch_btree_insert_keys(n3, op, &parent_keys, NULL);
2054 bch_btree_node_write(n3, &cl);
2055 mutex_unlock(&n3->write_lock);
2058 bch_btree_set_root(n3);
2059 rw_unlock(true, n3);
2060 } else if (!b->parent) {
2061 /* Root filled up but didn't need to be split */
2063 bch_btree_set_root(n1);
2065 /* Split a non root node */
2067 make_btree_freeing_key(b, parent_keys.top);
2068 bch_keylist_push(&parent_keys);
2070 bch_btree_insert_node(b->parent, op, &parent_keys, NULL, NULL);
2071 BUG_ON(!bch_keylist_empty(&parent_keys));
2075 rw_unlock(true, n1);
2077 bch_time_stats_update(&b->c->btree_split_time, start_time);
2081 bkey_put(b->c, &n2->key);
2082 btree_node_free(n2);
2083 rw_unlock(true, n2);
2085 bkey_put(b->c, &n1->key);
2086 btree_node_free(n1);
2087 rw_unlock(true, n1);
2089 WARN(1, "bcache: btree split failed (level %u)", b->level);
2091 if (n3 == ERR_PTR(-EAGAIN) ||
2092 n2 == ERR_PTR(-EAGAIN) ||
2093 n1 == ERR_PTR(-EAGAIN))
2099 static int bch_btree_insert_node(struct btree *b, struct btree_op *op,
2100 struct keylist *insert_keys,
2101 atomic_t *journal_ref,
2102 struct bkey *replace_key)
2106 BUG_ON(b->level && replace_key);
2108 closure_init_stack(&cl);
2110 mutex_lock(&b->write_lock);
2112 if (write_block(b) != btree_bset_last(b) &&
2113 b->keys.last_set_unwritten)
2114 bch_btree_init_next(b); /* just wrote a set */
2116 if (bch_keylist_nkeys(insert_keys) > insert_u64s_remaining(b)) {
2117 mutex_unlock(&b->write_lock);
2121 BUG_ON(write_block(b) != btree_bset_last(b));
2123 if (bch_btree_insert_keys(b, op, insert_keys, replace_key)) {
2125 bch_btree_leaf_dirty(b, journal_ref);
2127 bch_btree_node_write(b, &cl);
2130 mutex_unlock(&b->write_lock);
2132 /* wait for btree node write if necessary, after unlock */
2137 if (current->bio_list) {
2138 op->lock = b->c->root->level + 1;
2140 } else if (op->lock <= b->c->root->level) {
2141 op->lock = b->c->root->level + 1;
2144 /* Invalidated all iterators */
2145 int ret = btree_split(b, op, insert_keys, replace_key);
2147 if (bch_keylist_empty(insert_keys))
2155 int bch_btree_insert_check_key(struct btree *b, struct btree_op *op,
2156 struct bkey *check_key)
2159 uint64_t btree_ptr = b->key.ptr[0];
2160 unsigned long seq = b->seq;
2161 struct keylist insert;
2162 bool upgrade = op->lock == -1;
2164 bch_keylist_init(&insert);
2167 rw_unlock(false, b);
2168 rw_lock(true, b, b->level);
2170 if (b->key.ptr[0] != btree_ptr ||
2171 b->seq != seq + 1) {
2172 op->lock = b->level;
2177 SET_KEY_PTRS(check_key, 1);
2178 get_random_bytes(&check_key->ptr[0], sizeof(uint64_t));
2180 SET_PTR_DEV(check_key, 0, PTR_CHECK_DEV);
2182 bch_keylist_add(&insert, check_key);
2184 ret = bch_btree_insert_node(b, op, &insert, NULL, NULL);
2186 BUG_ON(!ret && !bch_keylist_empty(&insert));
2189 downgrade_write(&b->lock);
2193 struct btree_insert_op {
2195 struct keylist *keys;
2196 atomic_t *journal_ref;
2197 struct bkey *replace_key;
2200 static int btree_insert_fn(struct btree_op *b_op, struct btree *b)
2202 struct btree_insert_op *op = container_of(b_op,
2203 struct btree_insert_op, op);
2205 int ret = bch_btree_insert_node(b, &op->op, op->keys,
2206 op->journal_ref, op->replace_key);
2207 if (ret && !bch_keylist_empty(op->keys))
2213 int bch_btree_insert(struct cache_set *c, struct keylist *keys,
2214 atomic_t *journal_ref, struct bkey *replace_key)
2216 struct btree_insert_op op;
2219 BUG_ON(current->bio_list);
2220 BUG_ON(bch_keylist_empty(keys));
2222 bch_btree_op_init(&op.op, 0);
2224 op.journal_ref = journal_ref;
2225 op.replace_key = replace_key;
2227 while (!ret && !bch_keylist_empty(keys)) {
2229 ret = bch_btree_map_leaf_nodes(&op.op, c,
2230 &START_KEY(keys->keys),
2237 pr_err("error %i", ret);
2239 while ((k = bch_keylist_pop(keys)))
2241 } else if (op.op.insert_collision)
2247 void bch_btree_set_root(struct btree *b)
2252 closure_init_stack(&cl);
2254 trace_bcache_btree_set_root(b);
2256 BUG_ON(!b->written);
2258 for (i = 0; i < KEY_PTRS(&b->key); i++)
2259 BUG_ON(PTR_BUCKET(b->c, &b->key, i)->prio != BTREE_PRIO);
2261 mutex_lock(&b->c->bucket_lock);
2262 list_del_init(&b->list);
2263 mutex_unlock(&b->c->bucket_lock);
2267 bch_journal_meta(b->c, &cl);
2271 /* Map across nodes or keys */
2273 static int bch_btree_map_nodes_recurse(struct btree *b, struct btree_op *op,
2275 btree_map_nodes_fn *fn, int flags)
2277 int ret = MAP_CONTINUE;
2281 struct btree_iter iter;
2283 bch_btree_iter_init(&b->keys, &iter, from);
2285 while ((k = bch_btree_iter_next_filter(&iter, &b->keys,
2287 ret = btree(map_nodes_recurse, k, b,
2288 op, from, fn, flags);
2291 if (ret != MAP_CONTINUE)
2296 if (!b->level || flags == MAP_ALL_NODES)
2302 int __bch_btree_map_nodes(struct btree_op *op, struct cache_set *c,
2303 struct bkey *from, btree_map_nodes_fn *fn, int flags)
2305 return btree_root(map_nodes_recurse, c, op, from, fn, flags);
2308 static int bch_btree_map_keys_recurse(struct btree *b, struct btree_op *op,
2309 struct bkey *from, btree_map_keys_fn *fn,
2312 int ret = MAP_CONTINUE;
2314 struct btree_iter iter;
2316 bch_btree_iter_init(&b->keys, &iter, from);
2318 while ((k = bch_btree_iter_next_filter(&iter, &b->keys, bch_ptr_bad))) {
2321 : btree(map_keys_recurse, k, b, op, from, fn, flags);
2324 if (ret != MAP_CONTINUE)
2328 if (!b->level && (flags & MAP_END_KEY))
2329 ret = fn(op, b, &KEY(KEY_INODE(&b->key),
2330 KEY_OFFSET(&b->key), 0));
2335 int bch_btree_map_keys(struct btree_op *op, struct cache_set *c,
2336 struct bkey *from, btree_map_keys_fn *fn, int flags)
2338 return btree_root(map_keys_recurse, c, op, from, fn, flags);
2343 static inline int keybuf_cmp(struct keybuf_key *l, struct keybuf_key *r)
2345 /* Overlapping keys compare equal */
2346 if (bkey_cmp(&l->key, &START_KEY(&r->key)) <= 0)
2348 if (bkey_cmp(&START_KEY(&l->key), &r->key) >= 0)
2353 static inline int keybuf_nonoverlapping_cmp(struct keybuf_key *l,
2354 struct keybuf_key *r)
2356 return clamp_t(int64_t, bkey_cmp(&l->key, &r->key), -1, 1);
2364 keybuf_pred_fn *pred;
2367 static int refill_keybuf_fn(struct btree_op *op, struct btree *b,
2370 struct refill *refill = container_of(op, struct refill, op);
2371 struct keybuf *buf = refill->buf;
2372 int ret = MAP_CONTINUE;
2374 if (bkey_cmp(k, refill->end) >= 0) {
2379 if (!KEY_SIZE(k)) /* end key */
2382 if (refill->pred(buf, k)) {
2383 struct keybuf_key *w;
2385 spin_lock(&buf->lock);
2387 w = array_alloc(&buf->freelist);
2389 spin_unlock(&buf->lock);
2394 bkey_copy(&w->key, k);
2396 if (RB_INSERT(&buf->keys, w, node, keybuf_cmp))
2397 array_free(&buf->freelist, w);
2401 if (array_freelist_empty(&buf->freelist))
2404 spin_unlock(&buf->lock);
2407 buf->last_scanned = *k;
2411 void bch_refill_keybuf(struct cache_set *c, struct keybuf *buf,
2412 struct bkey *end, keybuf_pred_fn *pred)
2414 struct bkey start = buf->last_scanned;
2415 struct refill refill;
2419 bch_btree_op_init(&refill.op, -1);
2420 refill.nr_found = 0;
2425 bch_btree_map_keys(&refill.op, c, &buf->last_scanned,
2426 refill_keybuf_fn, MAP_END_KEY);
2428 trace_bcache_keyscan(refill.nr_found,
2429 KEY_INODE(&start), KEY_OFFSET(&start),
2430 KEY_INODE(&buf->last_scanned),
2431 KEY_OFFSET(&buf->last_scanned));
2433 spin_lock(&buf->lock);
2435 if (!RB_EMPTY_ROOT(&buf->keys)) {
2436 struct keybuf_key *w;
2437 w = RB_FIRST(&buf->keys, struct keybuf_key, node);
2438 buf->start = START_KEY(&w->key);
2440 w = RB_LAST(&buf->keys, struct keybuf_key, node);
2443 buf->start = MAX_KEY;
2447 spin_unlock(&buf->lock);
2450 static void __bch_keybuf_del(struct keybuf *buf, struct keybuf_key *w)
2452 rb_erase(&w->node, &buf->keys);
2453 array_free(&buf->freelist, w);
2456 void bch_keybuf_del(struct keybuf *buf, struct keybuf_key *w)
2458 spin_lock(&buf->lock);
2459 __bch_keybuf_del(buf, w);
2460 spin_unlock(&buf->lock);
2463 bool bch_keybuf_check_overlapping(struct keybuf *buf, struct bkey *start,
2467 struct keybuf_key *p, *w, s;
2470 if (bkey_cmp(end, &buf->start) <= 0 ||
2471 bkey_cmp(start, &buf->end) >= 0)
2474 spin_lock(&buf->lock);
2475 w = RB_GREATER(&buf->keys, s, node, keybuf_nonoverlapping_cmp);
2477 while (w && bkey_cmp(&START_KEY(&w->key), end) < 0) {
2479 w = RB_NEXT(w, node);
2484 __bch_keybuf_del(buf, p);
2487 spin_unlock(&buf->lock);
2491 struct keybuf_key *bch_keybuf_next(struct keybuf *buf)
2493 struct keybuf_key *w;
2494 spin_lock(&buf->lock);
2496 w = RB_FIRST(&buf->keys, struct keybuf_key, node);
2498 while (w && w->private)
2499 w = RB_NEXT(w, node);
2502 w->private = ERR_PTR(-EINTR);
2504 spin_unlock(&buf->lock);
2508 struct keybuf_key *bch_keybuf_next_rescan(struct cache_set *c,
2511 keybuf_pred_fn *pred)
2513 struct keybuf_key *ret;
2516 ret = bch_keybuf_next(buf);
2520 if (bkey_cmp(&buf->last_scanned, end) >= 0) {
2521 pr_debug("scan finished");
2525 bch_refill_keybuf(c, buf, end, pred);
2531 void bch_keybuf_init(struct keybuf *buf)
2533 buf->last_scanned = MAX_KEY;
2534 buf->keys = RB_ROOT;
2536 spin_lock_init(&buf->lock);
2537 array_allocator_init(&buf->freelist);