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/kthread.h>
32 #include <linux/prefetch.h>
33 #include <linux/random.h>
34 #include <linux/rcupdate.h>
35 #include <trace/events/bcache.h>
39 * register_bcache: Return errors out to userspace correctly
41 * Writeback: don't undirty key until after a cache flush
43 * Create an iterator for key pointers
45 * On btree write error, mark bucket such that it won't be freed from the cache
48 * Check for bad keys in replay
50 * Refcount journal entries in journal_replay
53 * Finish incremental gc
54 * Gc should free old UUIDs, data for invalid UUIDs
56 * Provide a way to list backing device UUIDs we have data cached for, and
57 * probably how long it's been since we've seen them, and a way to invalidate
58 * dirty data for devices that will never be attached again
60 * Keep 1 min/5 min/15 min statistics of how busy a block device has been, so
61 * that based on that and how much dirty data we have we can keep writeback
64 * Add a tracepoint or somesuch to watch for writeback starvation
66 * When btree depth > 1 and splitting an interior node, we have to make sure
67 * alloc_bucket() cannot fail. This should be true but is not completely
72 * If data write is less than hard sector size of ssd, round up offset in open
73 * bucket to the next whole sector
75 * Superblock needs to be fleshed out for multiple cache devices
77 * Add a sysfs tunable for the number of writeback IOs in flight
79 * Add a sysfs tunable for the number of open data buckets
81 * IO tracking: Can we track when one process is doing io on behalf of another?
82 * IO tracking: Don't use just an average, weigh more recent stuff higher
84 * Test module load/unload
87 #define MAX_NEED_GC 64
88 #define MAX_SAVE_PRIO 72
90 #define PTR_DIRTY_BIT (((uint64_t) 1 << 36))
92 #define PTR_HASH(c, k) \
93 (((k)->ptr[0] >> c->bucket_bits) | PTR_GEN(k, 0))
95 #define insert_lock(s, b) ((b)->level <= (s)->lock)
98 * These macros are for recursing down the btree - they handle the details of
99 * locking and looking up nodes in the cache for you. They're best treated as
100 * mere syntax when reading code that uses them.
102 * op->lock determines whether we take a read or a write lock at a given depth.
103 * If you've got a read lock and find that you need a write lock (i.e. you're
104 * going to have to split), set op->lock and return -EINTR; btree_root() will
105 * call you again and you'll have the correct lock.
109 * btree - recurse down the btree on a specified key
110 * @fn: function to call, which will be passed the child node
111 * @key: key to recurse on
112 * @b: parent btree node
113 * @op: pointer to struct btree_op
115 #define btree(fn, key, b, op, ...) \
117 int _r, l = (b)->level - 1; \
118 bool _w = l <= (op)->lock; \
119 struct btree *_child = bch_btree_node_get((b)->c, op, key, l, \
121 if (!IS_ERR(_child)) { \
122 _r = bch_btree_ ## fn(_child, op, ##__VA_ARGS__); \
123 rw_unlock(_w, _child); \
125 _r = PTR_ERR(_child); \
130 * btree_root - call a function on the root of the btree
131 * @fn: function to call, which will be passed the child node
133 * @op: pointer to struct btree_op
135 #define btree_root(fn, c, op, ...) \
139 struct btree *_b = (c)->root; \
140 bool _w = insert_lock(op, _b); \
141 rw_lock(_w, _b, _b->level); \
142 if (_b == (c)->root && \
143 _w == insert_lock(op, _b)) { \
144 _r = bch_btree_ ## fn(_b, op, ##__VA_ARGS__); \
147 bch_cannibalize_unlock(c); \
150 } while (_r == -EINTR); \
152 finish_wait(&(c)->btree_cache_wait, &(op)->wait); \
156 static inline struct bset *write_block(struct btree *b)
158 return ((void *) btree_bset_first(b)) + b->written * block_bytes(b->c);
161 static void bch_btree_init_next(struct btree *b)
163 /* If not a leaf node, always sort */
164 if (b->level && b->keys.nsets)
165 bch_btree_sort(&b->keys, &b->c->sort);
167 bch_btree_sort_lazy(&b->keys, &b->c->sort);
169 if (b->written < btree_blocks(b))
170 bch_bset_init_next(&b->keys, write_block(b),
171 bset_magic(&b->c->sb));
175 /* Btree key manipulation */
177 void bkey_put(struct cache_set *c, struct bkey *k)
181 for (i = 0; i < KEY_PTRS(k); i++)
182 if (ptr_available(c, k, i))
183 atomic_dec_bug(&PTR_BUCKET(c, k, i)->pin);
188 static uint64_t btree_csum_set(struct btree *b, struct bset *i)
190 uint64_t crc = b->key.ptr[0];
191 void *data = (void *) i + 8, *end = bset_bkey_last(i);
193 crc = bch_crc64_update(crc, data, end - data);
194 return crc ^ 0xffffffffffffffffULL;
197 void bch_btree_node_read_done(struct btree *b)
199 const char *err = "bad btree header";
200 struct bset *i = btree_bset_first(b);
201 struct btree_iter *iter;
203 iter = mempool_alloc(b->c->fill_iter, GFP_NOIO);
204 iter->size = b->c->sb.bucket_size / b->c->sb.block_size;
207 #ifdef CONFIG_BCACHE_DEBUG
215 b->written < btree_blocks(b) && i->seq == b->keys.set[0].data->seq;
216 i = write_block(b)) {
217 err = "unsupported bset version";
218 if (i->version > BCACHE_BSET_VERSION)
221 err = "bad btree header";
222 if (b->written + set_blocks(i, block_bytes(b->c)) >
227 if (i->magic != bset_magic(&b->c->sb))
230 err = "bad checksum";
231 switch (i->version) {
233 if (i->csum != csum_set(i))
236 case BCACHE_BSET_VERSION:
237 if (i->csum != btree_csum_set(b, i))
243 if (i != b->keys.set[0].data && !i->keys)
246 bch_btree_iter_push(iter, i->start, bset_bkey_last(i));
248 b->written += set_blocks(i, block_bytes(b->c));
251 err = "corrupted btree";
252 for (i = write_block(b);
253 bset_sector_offset(&b->keys, i) < KEY_SIZE(&b->key);
254 i = ((void *) i) + block_bytes(b->c))
255 if (i->seq == b->keys.set[0].data->seq)
258 bch_btree_sort_and_fix_extents(&b->keys, iter, &b->c->sort);
260 i = b->keys.set[0].data;
261 err = "short btree key";
262 if (b->keys.set[0].size &&
263 bkey_cmp(&b->key, &b->keys.set[0].end) < 0)
266 if (b->written < btree_blocks(b))
267 bch_bset_init_next(&b->keys, write_block(b),
268 bset_magic(&b->c->sb));
270 mempool_free(iter, b->c->fill_iter);
273 set_btree_node_io_error(b);
274 bch_cache_set_error(b->c, "%s at bucket %zu, block %u, %u keys",
275 err, PTR_BUCKET_NR(b->c, &b->key, 0),
276 bset_block_offset(b, i), i->keys);
280 static void btree_node_read_endio(struct bio *bio)
282 struct closure *cl = bio->bi_private;
286 static void bch_btree_node_read(struct btree *b)
288 uint64_t start_time = local_clock();
292 trace_bcache_btree_read(b);
294 closure_init_stack(&cl);
296 bio = bch_bbio_alloc(b->c);
297 bio->bi_iter.bi_size = KEY_SIZE(&b->key) << 9;
298 bio->bi_end_io = btree_node_read_endio;
299 bio->bi_private = &cl;
300 bio_set_op_attrs(bio, REQ_OP_READ, REQ_META|READ_SYNC);
302 bch_bio_map(bio, b->keys.set[0].data);
304 bch_submit_bbio(bio, b->c, &b->key, 0);
308 set_btree_node_io_error(b);
310 bch_bbio_free(bio, b->c);
312 if (btree_node_io_error(b))
315 bch_btree_node_read_done(b);
316 bch_time_stats_update(&b->c->btree_read_time, start_time);
320 bch_cache_set_error(b->c, "io error reading bucket %zu",
321 PTR_BUCKET_NR(b->c, &b->key, 0));
324 static void btree_complete_write(struct btree *b, struct btree_write *w)
326 if (w->prio_blocked &&
327 !atomic_sub_return(w->prio_blocked, &b->c->prio_blocked))
328 wake_up_allocators(b->c);
331 atomic_dec_bug(w->journal);
332 __closure_wake_up(&b->c->journal.wait);
339 static void btree_node_write_unlock(struct closure *cl)
341 struct btree *b = container_of(cl, struct btree, io);
346 static void __btree_node_write_done(struct closure *cl)
348 struct btree *b = container_of(cl, struct btree, io);
349 struct btree_write *w = btree_prev_write(b);
351 bch_bbio_free(b->bio, b->c);
353 btree_complete_write(b, w);
355 if (btree_node_dirty(b))
356 schedule_delayed_work(&b->work, 30 * HZ);
358 closure_return_with_destructor(cl, btree_node_write_unlock);
361 static void btree_node_write_done(struct closure *cl)
363 struct btree *b = container_of(cl, struct btree, io);
365 bio_free_pages(b->bio);
366 __btree_node_write_done(cl);
369 static void btree_node_write_endio(struct bio *bio)
371 struct closure *cl = bio->bi_private;
372 struct btree *b = container_of(cl, struct btree, io);
375 set_btree_node_io_error(b);
377 bch_bbio_count_io_errors(b->c, bio, bio->bi_error, "writing btree");
381 static void do_btree_node_write(struct btree *b)
383 struct closure *cl = &b->io;
384 struct bset *i = btree_bset_last(b);
387 i->version = BCACHE_BSET_VERSION;
388 i->csum = btree_csum_set(b, i);
391 b->bio = bch_bbio_alloc(b->c);
393 b->bio->bi_end_io = btree_node_write_endio;
394 b->bio->bi_private = cl;
395 b->bio->bi_iter.bi_size = roundup(set_bytes(i), block_bytes(b->c));
396 bio_set_op_attrs(b->bio, REQ_OP_WRITE, REQ_META|WRITE_SYNC|REQ_FUA);
397 bch_bio_map(b->bio, i);
400 * If we're appending to a leaf node, we don't technically need FUA -
401 * this write just needs to be persisted before the next journal write,
402 * which will be marked FLUSH|FUA.
404 * Similarly if we're writing a new btree root - the pointer is going to
405 * be in the next journal entry.
407 * But if we're writing a new btree node (that isn't a root) or
408 * appending to a non leaf btree node, we need either FUA or a flush
409 * when we write the parent with the new pointer. FUA is cheaper than a
410 * flush, and writes appending to leaf nodes aren't blocking anything so
411 * just make all btree node writes FUA to keep things sane.
414 bkey_copy(&k.key, &b->key);
415 SET_PTR_OFFSET(&k.key, 0, PTR_OFFSET(&k.key, 0) +
416 bset_sector_offset(&b->keys, i));
418 if (!bio_alloc_pages(b->bio, __GFP_NOWARN|GFP_NOWAIT)) {
421 void *base = (void *) ((unsigned long) i & ~(PAGE_SIZE - 1));
423 bio_for_each_segment_all(bv, b->bio, j)
424 memcpy(page_address(bv->bv_page),
425 base + j * PAGE_SIZE, PAGE_SIZE);
427 bch_submit_bbio(b->bio, b->c, &k.key, 0);
429 continue_at(cl, btree_node_write_done, NULL);
432 bch_bio_map(b->bio, i);
434 bch_submit_bbio(b->bio, b->c, &k.key, 0);
437 continue_at_nobarrier(cl, __btree_node_write_done, NULL);
441 void __bch_btree_node_write(struct btree *b, struct closure *parent)
443 struct bset *i = btree_bset_last(b);
445 lockdep_assert_held(&b->write_lock);
447 trace_bcache_btree_write(b);
449 BUG_ON(current->bio_list);
450 BUG_ON(b->written >= btree_blocks(b));
451 BUG_ON(b->written && !i->keys);
452 BUG_ON(btree_bset_first(b)->seq != i->seq);
453 bch_check_keys(&b->keys, "writing");
455 cancel_delayed_work(&b->work);
457 /* If caller isn't waiting for write, parent refcount is cache set */
459 closure_init(&b->io, parent ?: &b->c->cl);
461 clear_bit(BTREE_NODE_dirty, &b->flags);
462 change_bit(BTREE_NODE_write_idx, &b->flags);
464 do_btree_node_write(b);
466 atomic_long_add(set_blocks(i, block_bytes(b->c)) * b->c->sb.block_size,
467 &PTR_CACHE(b->c, &b->key, 0)->btree_sectors_written);
469 b->written += set_blocks(i, block_bytes(b->c));
472 void bch_btree_node_write(struct btree *b, struct closure *parent)
474 unsigned nsets = b->keys.nsets;
476 lockdep_assert_held(&b->lock);
478 __bch_btree_node_write(b, parent);
481 * do verify if there was more than one set initially (i.e. we did a
482 * sort) and we sorted down to a single set:
484 if (nsets && !b->keys.nsets)
487 bch_btree_init_next(b);
490 static void bch_btree_node_write_sync(struct btree *b)
494 closure_init_stack(&cl);
496 mutex_lock(&b->write_lock);
497 bch_btree_node_write(b, &cl);
498 mutex_unlock(&b->write_lock);
503 static void btree_node_write_work(struct work_struct *w)
505 struct btree *b = container_of(to_delayed_work(w), struct btree, work);
507 mutex_lock(&b->write_lock);
508 if (btree_node_dirty(b))
509 __bch_btree_node_write(b, NULL);
510 mutex_unlock(&b->write_lock);
513 static void bch_btree_leaf_dirty(struct btree *b, atomic_t *journal_ref)
515 struct bset *i = btree_bset_last(b);
516 struct btree_write *w = btree_current_write(b);
518 lockdep_assert_held(&b->write_lock);
523 if (!btree_node_dirty(b))
524 schedule_delayed_work(&b->work, 30 * HZ);
526 set_btree_node_dirty(b);
530 journal_pin_cmp(b->c, w->journal, journal_ref)) {
531 atomic_dec_bug(w->journal);
536 w->journal = journal_ref;
537 atomic_inc(w->journal);
541 /* Force write if set is too big */
542 if (set_bytes(i) > PAGE_SIZE - 48 &&
544 bch_btree_node_write(b, NULL);
548 * Btree in memory cache - allocation/freeing
549 * mca -> memory cache
552 #define mca_reserve(c) (((c->root && c->root->level) \
553 ? c->root->level : 1) * 8 + 16)
554 #define mca_can_free(c) \
555 max_t(int, 0, c->btree_cache_used - mca_reserve(c))
557 static void mca_data_free(struct btree *b)
559 BUG_ON(b->io_mutex.count != 1);
561 bch_btree_keys_free(&b->keys);
563 b->c->btree_cache_used--;
564 list_move(&b->list, &b->c->btree_cache_freed);
567 static void mca_bucket_free(struct btree *b)
569 BUG_ON(btree_node_dirty(b));
572 hlist_del_init_rcu(&b->hash);
573 list_move(&b->list, &b->c->btree_cache_freeable);
576 static unsigned btree_order(struct bkey *k)
578 return ilog2(KEY_SIZE(k) / PAGE_SECTORS ?: 1);
581 static void mca_data_alloc(struct btree *b, struct bkey *k, gfp_t gfp)
583 if (!bch_btree_keys_alloc(&b->keys,
585 ilog2(b->c->btree_pages),
588 b->c->btree_cache_used++;
589 list_move(&b->list, &b->c->btree_cache);
591 list_move(&b->list, &b->c->btree_cache_freed);
595 static struct btree *mca_bucket_alloc(struct cache_set *c,
596 struct bkey *k, gfp_t gfp)
598 struct btree *b = kzalloc(sizeof(struct btree), gfp);
602 init_rwsem(&b->lock);
603 lockdep_set_novalidate_class(&b->lock);
604 mutex_init(&b->write_lock);
605 lockdep_set_novalidate_class(&b->write_lock);
606 INIT_LIST_HEAD(&b->list);
607 INIT_DELAYED_WORK(&b->work, btree_node_write_work);
609 sema_init(&b->io_mutex, 1);
611 mca_data_alloc(b, k, gfp);
615 static int mca_reap(struct btree *b, unsigned min_order, bool flush)
619 closure_init_stack(&cl);
620 lockdep_assert_held(&b->c->bucket_lock);
622 if (!down_write_trylock(&b->lock))
625 BUG_ON(btree_node_dirty(b) && !b->keys.set[0].data);
627 if (b->keys.page_order < min_order)
631 if (btree_node_dirty(b))
634 if (down_trylock(&b->io_mutex))
639 mutex_lock(&b->write_lock);
640 if (btree_node_dirty(b))
641 __bch_btree_node_write(b, &cl);
642 mutex_unlock(&b->write_lock);
646 /* wait for any in flight btree write */
656 static unsigned long bch_mca_scan(struct shrinker *shrink,
657 struct shrink_control *sc)
659 struct cache_set *c = container_of(shrink, struct cache_set, shrink);
661 unsigned long i, nr = sc->nr_to_scan;
662 unsigned long freed = 0;
664 if (c->shrinker_disabled)
667 if (c->btree_cache_alloc_lock)
670 /* Return -1 if we can't do anything right now */
671 if (sc->gfp_mask & __GFP_IO)
672 mutex_lock(&c->bucket_lock);
673 else if (!mutex_trylock(&c->bucket_lock))
677 * It's _really_ critical that we don't free too many btree nodes - we
678 * have to always leave ourselves a reserve. The reserve is how we
679 * guarantee that allocating memory for a new btree node can always
680 * succeed, so that inserting keys into the btree can always succeed and
681 * IO can always make forward progress:
683 nr /= c->btree_pages;
686 nr = min_t(unsigned long, nr, mca_can_free(c));
689 list_for_each_entry_safe(b, t, &c->btree_cache_freeable, list) {
694 !mca_reap(b, 0, false)) {
701 for (i = 0; (nr--) && i < c->btree_cache_used; i++) {
702 if (list_empty(&c->btree_cache))
705 b = list_first_entry(&c->btree_cache, struct btree, list);
706 list_rotate_left(&c->btree_cache);
709 !mca_reap(b, 0, false)) {
718 mutex_unlock(&c->bucket_lock);
722 static unsigned long bch_mca_count(struct shrinker *shrink,
723 struct shrink_control *sc)
725 struct cache_set *c = container_of(shrink, struct cache_set, shrink);
727 if (c->shrinker_disabled)
730 if (c->btree_cache_alloc_lock)
733 return mca_can_free(c) * c->btree_pages;
736 void bch_btree_cache_free(struct cache_set *c)
740 closure_init_stack(&cl);
742 if (c->shrink.list.next)
743 unregister_shrinker(&c->shrink);
745 mutex_lock(&c->bucket_lock);
747 #ifdef CONFIG_BCACHE_DEBUG
749 list_move(&c->verify_data->list, &c->btree_cache);
751 free_pages((unsigned long) c->verify_ondisk, ilog2(bucket_pages(c)));
754 list_splice(&c->btree_cache_freeable,
757 while (!list_empty(&c->btree_cache)) {
758 b = list_first_entry(&c->btree_cache, struct btree, list);
760 if (btree_node_dirty(b))
761 btree_complete_write(b, btree_current_write(b));
762 clear_bit(BTREE_NODE_dirty, &b->flags);
767 while (!list_empty(&c->btree_cache_freed)) {
768 b = list_first_entry(&c->btree_cache_freed,
771 cancel_delayed_work_sync(&b->work);
775 mutex_unlock(&c->bucket_lock);
778 int bch_btree_cache_alloc(struct cache_set *c)
782 for (i = 0; i < mca_reserve(c); i++)
783 if (!mca_bucket_alloc(c, &ZERO_KEY, GFP_KERNEL))
786 list_splice_init(&c->btree_cache,
787 &c->btree_cache_freeable);
789 #ifdef CONFIG_BCACHE_DEBUG
790 mutex_init(&c->verify_lock);
792 c->verify_ondisk = (void *)
793 __get_free_pages(GFP_KERNEL|__GFP_COMP, ilog2(bucket_pages(c)));
795 c->verify_data = mca_bucket_alloc(c, &ZERO_KEY, GFP_KERNEL);
797 if (c->verify_data &&
798 c->verify_data->keys.set->data)
799 list_del_init(&c->verify_data->list);
801 c->verify_data = NULL;
804 c->shrink.count_objects = bch_mca_count;
805 c->shrink.scan_objects = bch_mca_scan;
807 c->shrink.batch = c->btree_pages * 2;
809 if (register_shrinker(&c->shrink))
810 pr_warn("bcache: %s: could not register shrinker",
816 /* Btree in memory cache - hash table */
818 static struct hlist_head *mca_hash(struct cache_set *c, struct bkey *k)
820 return &c->bucket_hash[hash_32(PTR_HASH(c, k), BUCKET_HASH_BITS)];
823 static struct btree *mca_find(struct cache_set *c, struct bkey *k)
828 hlist_for_each_entry_rcu(b, mca_hash(c, k), hash)
829 if (PTR_HASH(c, &b->key) == PTR_HASH(c, k))
837 static int mca_cannibalize_lock(struct cache_set *c, struct btree_op *op)
839 spin_lock(&c->btree_cannibalize_lock);
840 if (likely(c->btree_cache_alloc_lock == NULL)) {
841 c->btree_cache_alloc_lock = current;
842 } else if (c->btree_cache_alloc_lock != current) {
844 prepare_to_wait(&c->btree_cache_wait, &op->wait,
845 TASK_UNINTERRUPTIBLE);
846 spin_unlock(&c->btree_cannibalize_lock);
849 spin_unlock(&c->btree_cannibalize_lock);
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 spin_lock(&c->btree_cannibalize_lock);
885 if (c->btree_cache_alloc_lock == current) {
886 c->btree_cache_alloc_lock = NULL;
887 wake_up(&c->btree_cache_wait);
889 spin_unlock(&c->btree_cannibalize_lock);
892 static struct btree *mca_alloc(struct cache_set *c, struct btree_op *op,
893 struct bkey *k, int level)
897 BUG_ON(current->bio_list);
899 lockdep_assert_held(&c->bucket_lock);
904 /* btree_free() doesn't free memory; it sticks the node on the end of
905 * the list. Check if there's any freed nodes there:
907 list_for_each_entry(b, &c->btree_cache_freeable, list)
908 if (!mca_reap(b, btree_order(k), false))
911 /* We never free struct btree itself, just the memory that holds the on
912 * disk node. Check the freed list before allocating a new one:
914 list_for_each_entry(b, &c->btree_cache_freed, list)
915 if (!mca_reap(b, 0, false)) {
916 mca_data_alloc(b, k, __GFP_NOWARN|GFP_NOIO);
917 if (!b->keys.set[0].data)
923 b = mca_bucket_alloc(c, k, __GFP_NOWARN|GFP_NOIO);
927 BUG_ON(!down_write_trylock(&b->lock));
928 if (!b->keys.set->data)
931 BUG_ON(b->io_mutex.count != 1);
933 bkey_copy(&b->key, k);
934 list_move(&b->list, &c->btree_cache);
935 hlist_del_init_rcu(&b->hash);
936 hlist_add_head_rcu(&b->hash, mca_hash(c, k));
938 lock_set_subclass(&b->lock.dep_map, level + 1, _THIS_IP_);
939 b->parent = (void *) ~0UL;
945 bch_btree_keys_init(&b->keys, &bch_extent_keys_ops,
946 &b->c->expensive_debug_checks);
948 bch_btree_keys_init(&b->keys, &bch_btree_keys_ops,
949 &b->c->expensive_debug_checks);
956 b = mca_cannibalize(c, op, k);
964 * bch_btree_node_get - find a btree node in the cache and lock it, reading it
965 * in from disk if necessary.
967 * If IO is necessary and running under generic_make_request, returns -EAGAIN.
969 * The btree node will have either a read or a write lock held, depending on
970 * level and op->lock.
972 struct btree *bch_btree_node_get(struct cache_set *c, struct btree_op *op,
973 struct bkey *k, int level, bool write,
974 struct btree *parent)
984 if (current->bio_list)
985 return ERR_PTR(-EAGAIN);
987 mutex_lock(&c->bucket_lock);
988 b = mca_alloc(c, op, k, level);
989 mutex_unlock(&c->bucket_lock);
996 bch_btree_node_read(b);
999 downgrade_write(&b->lock);
1001 rw_lock(write, b, level);
1002 if (PTR_HASH(c, &b->key) != PTR_HASH(c, k)) {
1003 rw_unlock(write, b);
1006 BUG_ON(b->level != level);
1012 for (; i <= b->keys.nsets && b->keys.set[i].size; i++) {
1013 prefetch(b->keys.set[i].tree);
1014 prefetch(b->keys.set[i].data);
1017 for (; i <= b->keys.nsets; i++)
1018 prefetch(b->keys.set[i].data);
1020 if (btree_node_io_error(b)) {
1021 rw_unlock(write, b);
1022 return ERR_PTR(-EIO);
1025 BUG_ON(!b->written);
1030 static void btree_node_prefetch(struct btree *parent, struct bkey *k)
1034 mutex_lock(&parent->c->bucket_lock);
1035 b = mca_alloc(parent->c, NULL, k, parent->level - 1);
1036 mutex_unlock(&parent->c->bucket_lock);
1038 if (!IS_ERR_OR_NULL(b)) {
1040 bch_btree_node_read(b);
1047 static void btree_node_free(struct btree *b)
1049 trace_bcache_btree_node_free(b);
1051 BUG_ON(b == b->c->root);
1053 mutex_lock(&b->write_lock);
1055 if (btree_node_dirty(b))
1056 btree_complete_write(b, btree_current_write(b));
1057 clear_bit(BTREE_NODE_dirty, &b->flags);
1059 mutex_unlock(&b->write_lock);
1061 cancel_delayed_work(&b->work);
1063 mutex_lock(&b->c->bucket_lock);
1064 bch_bucket_free(b->c, &b->key);
1066 mutex_unlock(&b->c->bucket_lock);
1069 struct btree *__bch_btree_node_alloc(struct cache_set *c, struct btree_op *op,
1070 int level, bool wait,
1071 struct btree *parent)
1074 struct btree *b = ERR_PTR(-EAGAIN);
1076 mutex_lock(&c->bucket_lock);
1078 if (__bch_bucket_alloc_set(c, RESERVE_BTREE, &k.key, 1, wait))
1081 bkey_put(c, &k.key);
1082 SET_KEY_SIZE(&k.key, c->btree_pages * PAGE_SECTORS);
1084 b = mca_alloc(c, op, &k.key, level);
1090 "Tried to allocate bucket that was in btree cache");
1096 bch_bset_init_next(&b->keys, b->keys.set->data, bset_magic(&b->c->sb));
1098 mutex_unlock(&c->bucket_lock);
1100 trace_bcache_btree_node_alloc(b);
1103 bch_bucket_free(c, &k.key);
1105 mutex_unlock(&c->bucket_lock);
1107 trace_bcache_btree_node_alloc_fail(c);
1111 static struct btree *bch_btree_node_alloc(struct cache_set *c,
1112 struct btree_op *op, int level,
1113 struct btree *parent)
1115 return __bch_btree_node_alloc(c, op, level, op != NULL, parent);
1118 static struct btree *btree_node_alloc_replacement(struct btree *b,
1119 struct btree_op *op)
1121 struct btree *n = bch_btree_node_alloc(b->c, op, b->level, b->parent);
1122 if (!IS_ERR_OR_NULL(n)) {
1123 mutex_lock(&n->write_lock);
1124 bch_btree_sort_into(&b->keys, &n->keys, &b->c->sort);
1125 bkey_copy_key(&n->key, &b->key);
1126 mutex_unlock(&n->write_lock);
1132 static void make_btree_freeing_key(struct btree *b, struct bkey *k)
1136 mutex_lock(&b->c->bucket_lock);
1138 atomic_inc(&b->c->prio_blocked);
1140 bkey_copy(k, &b->key);
1141 bkey_copy_key(k, &ZERO_KEY);
1143 for (i = 0; i < KEY_PTRS(k); i++)
1145 bch_inc_gen(PTR_CACHE(b->c, &b->key, i),
1146 PTR_BUCKET(b->c, &b->key, i)));
1148 mutex_unlock(&b->c->bucket_lock);
1151 static int btree_check_reserve(struct btree *b, struct btree_op *op)
1153 struct cache_set *c = b->c;
1155 unsigned i, reserve = (c->root->level - b->level) * 2 + 1;
1157 mutex_lock(&c->bucket_lock);
1159 for_each_cache(ca, c, i)
1160 if (fifo_used(&ca->free[RESERVE_BTREE]) < reserve) {
1162 prepare_to_wait(&c->btree_cache_wait, &op->wait,
1163 TASK_UNINTERRUPTIBLE);
1164 mutex_unlock(&c->bucket_lock);
1168 mutex_unlock(&c->bucket_lock);
1170 return mca_cannibalize_lock(b->c, op);
1173 /* Garbage collection */
1175 static uint8_t __bch_btree_mark_key(struct cache_set *c, int level,
1183 * ptr_invalid() can't return true for the keys that mark btree nodes as
1184 * freed, but since ptr_bad() returns true we'll never actually use them
1185 * for anything and thus we don't want mark their pointers here
1187 if (!bkey_cmp(k, &ZERO_KEY))
1190 for (i = 0; i < KEY_PTRS(k); i++) {
1191 if (!ptr_available(c, k, i))
1194 g = PTR_BUCKET(c, k, i);
1196 if (gen_after(g->last_gc, PTR_GEN(k, i)))
1197 g->last_gc = PTR_GEN(k, i);
1199 if (ptr_stale(c, k, i)) {
1200 stale = max(stale, ptr_stale(c, k, i));
1204 cache_bug_on(GC_MARK(g) &&
1205 (GC_MARK(g) == GC_MARK_METADATA) != (level != 0),
1206 c, "inconsistent ptrs: mark = %llu, level = %i",
1210 SET_GC_MARK(g, GC_MARK_METADATA);
1211 else if (KEY_DIRTY(k))
1212 SET_GC_MARK(g, GC_MARK_DIRTY);
1213 else if (!GC_MARK(g))
1214 SET_GC_MARK(g, GC_MARK_RECLAIMABLE);
1216 /* guard against overflow */
1217 SET_GC_SECTORS_USED(g, min_t(unsigned,
1218 GC_SECTORS_USED(g) + KEY_SIZE(k),
1219 MAX_GC_SECTORS_USED));
1221 BUG_ON(!GC_SECTORS_USED(g));
1227 #define btree_mark_key(b, k) __bch_btree_mark_key(b->c, b->level, k)
1229 void bch_initial_mark_key(struct cache_set *c, int level, struct bkey *k)
1233 for (i = 0; i < KEY_PTRS(k); i++)
1234 if (ptr_available(c, k, i) &&
1235 !ptr_stale(c, k, i)) {
1236 struct bucket *b = PTR_BUCKET(c, k, i);
1238 b->gen = PTR_GEN(k, i);
1240 if (level && bkey_cmp(k, &ZERO_KEY))
1241 b->prio = BTREE_PRIO;
1242 else if (!level && b->prio == BTREE_PRIO)
1243 b->prio = INITIAL_PRIO;
1246 __bch_btree_mark_key(c, level, k);
1249 static bool btree_gc_mark_node(struct btree *b, struct gc_stat *gc)
1252 unsigned keys = 0, good_keys = 0;
1254 struct btree_iter iter;
1255 struct bset_tree *t;
1259 for_each_key_filter(&b->keys, k, &iter, bch_ptr_invalid) {
1260 stale = max(stale, btree_mark_key(b, k));
1263 if (bch_ptr_bad(&b->keys, k))
1266 gc->key_bytes += bkey_u64s(k);
1270 gc->data += KEY_SIZE(k);
1273 for (t = b->keys.set; t <= &b->keys.set[b->keys.nsets]; t++)
1274 btree_bug_on(t->size &&
1275 bset_written(&b->keys, t) &&
1276 bkey_cmp(&b->key, &t->end) < 0,
1277 b, "found short btree key in gc");
1279 if (b->c->gc_always_rewrite)
1285 if ((keys - good_keys) * 2 > keys)
1291 #define GC_MERGE_NODES 4U
1293 struct gc_merge_info {
1298 static int bch_btree_insert_node(struct btree *, struct btree_op *,
1299 struct keylist *, atomic_t *, struct bkey *);
1301 static int btree_gc_coalesce(struct btree *b, struct btree_op *op,
1302 struct gc_stat *gc, struct gc_merge_info *r)
1304 unsigned i, nodes = 0, keys = 0, blocks;
1305 struct btree *new_nodes[GC_MERGE_NODES];
1306 struct keylist keylist;
1310 bch_keylist_init(&keylist);
1312 if (btree_check_reserve(b, NULL))
1315 memset(new_nodes, 0, sizeof(new_nodes));
1316 closure_init_stack(&cl);
1318 while (nodes < GC_MERGE_NODES && !IS_ERR_OR_NULL(r[nodes].b))
1319 keys += r[nodes++].keys;
1321 blocks = btree_default_blocks(b->c) * 2 / 3;
1324 __set_blocks(b->keys.set[0].data, keys,
1325 block_bytes(b->c)) > blocks * (nodes - 1))
1328 for (i = 0; i < nodes; i++) {
1329 new_nodes[i] = btree_node_alloc_replacement(r[i].b, NULL);
1330 if (IS_ERR_OR_NULL(new_nodes[i]))
1331 goto out_nocoalesce;
1335 * We have to check the reserve here, after we've allocated our new
1336 * nodes, to make sure the insert below will succeed - we also check
1337 * before as an optimization to potentially avoid a bunch of expensive
1340 if (btree_check_reserve(b, NULL))
1341 goto out_nocoalesce;
1343 for (i = 0; i < nodes; i++)
1344 mutex_lock(&new_nodes[i]->write_lock);
1346 for (i = nodes - 1; i > 0; --i) {
1347 struct bset *n1 = btree_bset_first(new_nodes[i]);
1348 struct bset *n2 = btree_bset_first(new_nodes[i - 1]);
1349 struct bkey *k, *last = NULL;
1355 k < bset_bkey_last(n2);
1357 if (__set_blocks(n1, n1->keys + keys +
1359 block_bytes(b->c)) > blocks)
1363 keys += bkey_u64s(k);
1367 * Last node we're not getting rid of - we're getting
1368 * rid of the node at r[0]. Have to try and fit all of
1369 * the remaining keys into this node; we can't ensure
1370 * they will always fit due to rounding and variable
1371 * length keys (shouldn't be possible in practice,
1374 if (__set_blocks(n1, n1->keys + n2->keys,
1375 block_bytes(b->c)) >
1376 btree_blocks(new_nodes[i]))
1377 goto out_unlock_nocoalesce;
1380 /* Take the key of the node we're getting rid of */
1384 BUG_ON(__set_blocks(n1, n1->keys + keys, block_bytes(b->c)) >
1385 btree_blocks(new_nodes[i]));
1388 bkey_copy_key(&new_nodes[i]->key, last);
1390 memcpy(bset_bkey_last(n1),
1392 (void *) bset_bkey_idx(n2, keys) - (void *) n2->start);
1395 r[i].keys = n1->keys;
1398 bset_bkey_idx(n2, keys),
1399 (void *) bset_bkey_last(n2) -
1400 (void *) bset_bkey_idx(n2, keys));
1404 if (__bch_keylist_realloc(&keylist,
1405 bkey_u64s(&new_nodes[i]->key)))
1406 goto out_unlock_nocoalesce;
1408 bch_btree_node_write(new_nodes[i], &cl);
1409 bch_keylist_add(&keylist, &new_nodes[i]->key);
1412 for (i = 0; i < nodes; i++)
1413 mutex_unlock(&new_nodes[i]->write_lock);
1417 /* We emptied out this node */
1418 BUG_ON(btree_bset_first(new_nodes[0])->keys);
1419 btree_node_free(new_nodes[0]);
1420 rw_unlock(true, new_nodes[0]);
1421 new_nodes[0] = NULL;
1423 for (i = 0; i < nodes; i++) {
1424 if (__bch_keylist_realloc(&keylist, bkey_u64s(&r[i].b->key)))
1425 goto out_nocoalesce;
1427 make_btree_freeing_key(r[i].b, keylist.top);
1428 bch_keylist_push(&keylist);
1431 bch_btree_insert_node(b, op, &keylist, NULL, NULL);
1432 BUG_ON(!bch_keylist_empty(&keylist));
1434 for (i = 0; i < nodes; i++) {
1435 btree_node_free(r[i].b);
1436 rw_unlock(true, r[i].b);
1438 r[i].b = new_nodes[i];
1441 memmove(r, r + 1, sizeof(r[0]) * (nodes - 1));
1442 r[nodes - 1].b = ERR_PTR(-EINTR);
1444 trace_bcache_btree_gc_coalesce(nodes);
1447 bch_keylist_free(&keylist);
1449 /* Invalidated our iterator */
1452 out_unlock_nocoalesce:
1453 for (i = 0; i < nodes; i++)
1454 mutex_unlock(&new_nodes[i]->write_lock);
1458 bch_keylist_free(&keylist);
1460 while ((k = bch_keylist_pop(&keylist)))
1461 if (!bkey_cmp(k, &ZERO_KEY))
1462 atomic_dec(&b->c->prio_blocked);
1464 for (i = 0; i < nodes; i++)
1465 if (!IS_ERR_OR_NULL(new_nodes[i])) {
1466 btree_node_free(new_nodes[i]);
1467 rw_unlock(true, new_nodes[i]);
1472 static int btree_gc_rewrite_node(struct btree *b, struct btree_op *op,
1473 struct btree *replace)
1475 struct keylist keys;
1478 if (btree_check_reserve(b, NULL))
1481 n = btree_node_alloc_replacement(replace, NULL);
1483 /* recheck reserve after allocating replacement node */
1484 if (btree_check_reserve(b, NULL)) {
1490 bch_btree_node_write_sync(n);
1492 bch_keylist_init(&keys);
1493 bch_keylist_add(&keys, &n->key);
1495 make_btree_freeing_key(replace, keys.top);
1496 bch_keylist_push(&keys);
1498 bch_btree_insert_node(b, op, &keys, NULL, NULL);
1499 BUG_ON(!bch_keylist_empty(&keys));
1501 btree_node_free(replace);
1504 /* Invalidated our iterator */
1508 static unsigned btree_gc_count_keys(struct btree *b)
1511 struct btree_iter iter;
1514 for_each_key_filter(&b->keys, k, &iter, bch_ptr_bad)
1515 ret += bkey_u64s(k);
1520 static int btree_gc_recurse(struct btree *b, struct btree_op *op,
1521 struct closure *writes, struct gc_stat *gc)
1524 bool should_rewrite;
1526 struct btree_iter iter;
1527 struct gc_merge_info r[GC_MERGE_NODES];
1528 struct gc_merge_info *i, *last = r + ARRAY_SIZE(r) - 1;
1530 bch_btree_iter_init(&b->keys, &iter, &b->c->gc_done);
1532 for (i = r; i < r + ARRAY_SIZE(r); i++)
1533 i->b = ERR_PTR(-EINTR);
1536 k = bch_btree_iter_next_filter(&iter, &b->keys, bch_ptr_bad);
1538 r->b = bch_btree_node_get(b->c, op, k, b->level - 1,
1541 ret = PTR_ERR(r->b);
1545 r->keys = btree_gc_count_keys(r->b);
1547 ret = btree_gc_coalesce(b, op, gc, r);
1555 if (!IS_ERR(last->b)) {
1556 should_rewrite = btree_gc_mark_node(last->b, gc);
1557 if (should_rewrite) {
1558 ret = btree_gc_rewrite_node(b, op, last->b);
1563 if (last->b->level) {
1564 ret = btree_gc_recurse(last->b, op, writes, gc);
1569 bkey_copy_key(&b->c->gc_done, &last->b->key);
1572 * Must flush leaf nodes before gc ends, since replace
1573 * operations aren't journalled
1575 mutex_lock(&last->b->write_lock);
1576 if (btree_node_dirty(last->b))
1577 bch_btree_node_write(last->b, writes);
1578 mutex_unlock(&last->b->write_lock);
1579 rw_unlock(true, last->b);
1582 memmove(r + 1, r, sizeof(r[0]) * (GC_MERGE_NODES - 1));
1585 if (need_resched()) {
1591 for (i = r; i < r + ARRAY_SIZE(r); i++)
1592 if (!IS_ERR_OR_NULL(i->b)) {
1593 mutex_lock(&i->b->write_lock);
1594 if (btree_node_dirty(i->b))
1595 bch_btree_node_write(i->b, writes);
1596 mutex_unlock(&i->b->write_lock);
1597 rw_unlock(true, i->b);
1603 static int bch_btree_gc_root(struct btree *b, struct btree_op *op,
1604 struct closure *writes, struct gc_stat *gc)
1606 struct btree *n = NULL;
1608 bool should_rewrite;
1610 should_rewrite = btree_gc_mark_node(b, gc);
1611 if (should_rewrite) {
1612 n = btree_node_alloc_replacement(b, NULL);
1614 if (!IS_ERR_OR_NULL(n)) {
1615 bch_btree_node_write_sync(n);
1617 bch_btree_set_root(n);
1625 __bch_btree_mark_key(b->c, b->level + 1, &b->key);
1628 ret = btree_gc_recurse(b, op, writes, gc);
1633 bkey_copy_key(&b->c->gc_done, &b->key);
1638 static void btree_gc_start(struct cache_set *c)
1644 if (!c->gc_mark_valid)
1647 mutex_lock(&c->bucket_lock);
1649 c->gc_mark_valid = 0;
1650 c->gc_done = ZERO_KEY;
1652 for_each_cache(ca, c, i)
1653 for_each_bucket(b, ca) {
1654 b->last_gc = b->gen;
1655 if (!atomic_read(&b->pin)) {
1657 SET_GC_SECTORS_USED(b, 0);
1661 mutex_unlock(&c->bucket_lock);
1664 static size_t bch_btree_gc_finish(struct cache_set *c)
1666 size_t available = 0;
1671 mutex_lock(&c->bucket_lock);
1674 c->gc_mark_valid = 1;
1677 for (i = 0; i < KEY_PTRS(&c->uuid_bucket); i++)
1678 SET_GC_MARK(PTR_BUCKET(c, &c->uuid_bucket, i),
1681 /* don't reclaim buckets to which writeback keys point */
1683 for (i = 0; i < c->nr_uuids; i++) {
1684 struct bcache_device *d = c->devices[i];
1685 struct cached_dev *dc;
1686 struct keybuf_key *w, *n;
1689 if (!d || UUID_FLASH_ONLY(&c->uuids[i]))
1691 dc = container_of(d, struct cached_dev, disk);
1693 spin_lock(&dc->writeback_keys.lock);
1694 rbtree_postorder_for_each_entry_safe(w, n,
1695 &dc->writeback_keys.keys, node)
1696 for (j = 0; j < KEY_PTRS(&w->key); j++)
1697 SET_GC_MARK(PTR_BUCKET(c, &w->key, j),
1699 spin_unlock(&dc->writeback_keys.lock);
1703 for_each_cache(ca, c, i) {
1706 ca->invalidate_needs_gc = 0;
1708 for (i = ca->sb.d; i < ca->sb.d + ca->sb.keys; i++)
1709 SET_GC_MARK(ca->buckets + *i, GC_MARK_METADATA);
1711 for (i = ca->prio_buckets;
1712 i < ca->prio_buckets + prio_buckets(ca) * 2; i++)
1713 SET_GC_MARK(ca->buckets + *i, GC_MARK_METADATA);
1715 for_each_bucket(b, ca) {
1716 c->need_gc = max(c->need_gc, bucket_gc_gen(b));
1718 if (atomic_read(&b->pin))
1721 BUG_ON(!GC_MARK(b) && GC_SECTORS_USED(b));
1723 if (!GC_MARK(b) || GC_MARK(b) == GC_MARK_RECLAIMABLE)
1728 mutex_unlock(&c->bucket_lock);
1732 static void bch_btree_gc(struct cache_set *c)
1735 unsigned long available;
1736 struct gc_stat stats;
1737 struct closure writes;
1739 uint64_t start_time = local_clock();
1741 trace_bcache_gc_start(c);
1743 memset(&stats, 0, sizeof(struct gc_stat));
1744 closure_init_stack(&writes);
1745 bch_btree_op_init(&op, SHRT_MAX);
1750 ret = btree_root(gc_root, c, &op, &writes, &stats);
1751 closure_sync(&writes);
1754 if (ret && ret != -EAGAIN)
1755 pr_warn("gc failed!");
1758 available = bch_btree_gc_finish(c);
1759 wake_up_allocators(c);
1761 bch_time_stats_update(&c->btree_gc_time, start_time);
1763 stats.key_bytes *= sizeof(uint64_t);
1765 stats.in_use = (c->nbuckets - available) * 100 / c->nbuckets;
1766 memcpy(&c->gc_stats, &stats, sizeof(struct gc_stat));
1768 trace_bcache_gc_end(c);
1773 static bool gc_should_run(struct cache_set *c)
1778 for_each_cache(ca, c, i)
1779 if (ca->invalidate_needs_gc)
1782 if (atomic_read(&c->sectors_to_gc) < 0)
1788 static int bch_gc_thread(void *arg)
1790 struct cache_set *c = arg;
1793 wait_event_interruptible(c->gc_wait,
1794 kthread_should_stop() || gc_should_run(c));
1796 if (kthread_should_stop())
1806 int bch_gc_thread_start(struct cache_set *c)
1808 c->gc_thread = kthread_run(bch_gc_thread, c, "bcache_gc");
1809 if (IS_ERR(c->gc_thread))
1810 return PTR_ERR(c->gc_thread);
1815 /* Initial partial gc */
1817 static int bch_btree_check_recurse(struct btree *b, struct btree_op *op)
1820 struct bkey *k, *p = NULL;
1821 struct btree_iter iter;
1823 for_each_key_filter(&b->keys, k, &iter, bch_ptr_invalid)
1824 bch_initial_mark_key(b->c, b->level, k);
1826 bch_initial_mark_key(b->c, b->level + 1, &b->key);
1829 bch_btree_iter_init(&b->keys, &iter, NULL);
1832 k = bch_btree_iter_next_filter(&iter, &b->keys,
1835 btree_node_prefetch(b, k);
1838 ret = btree(check_recurse, p, b, op);
1841 } while (p && !ret);
1847 int bch_btree_check(struct cache_set *c)
1851 bch_btree_op_init(&op, SHRT_MAX);
1853 return btree_root(check_recurse, c, &op);
1856 void bch_initial_gc_finish(struct cache_set *c)
1862 bch_btree_gc_finish(c);
1864 mutex_lock(&c->bucket_lock);
1867 * We need to put some unused buckets directly on the prio freelist in
1868 * order to get the allocator thread started - it needs freed buckets in
1869 * order to rewrite the prios and gens, and it needs to rewrite prios
1870 * and gens in order to free buckets.
1872 * This is only safe for buckets that have no live data in them, which
1873 * there should always be some of.
1875 for_each_cache(ca, c, i) {
1876 for_each_bucket(b, ca) {
1877 if (fifo_full(&ca->free[RESERVE_PRIO]) &&
1878 fifo_full(&ca->free[RESERVE_BTREE]))
1881 if (bch_can_invalidate_bucket(ca, b) &&
1883 __bch_invalidate_one_bucket(ca, b);
1884 if (!fifo_push(&ca->free[RESERVE_PRIO],
1886 fifo_push(&ca->free[RESERVE_BTREE],
1892 mutex_unlock(&c->bucket_lock);
1895 /* Btree insertion */
1897 static bool btree_insert_key(struct btree *b, struct bkey *k,
1898 struct bkey *replace_key)
1902 BUG_ON(bkey_cmp(k, &b->key) > 0);
1904 status = bch_btree_insert_key(&b->keys, k, replace_key);
1905 if (status != BTREE_INSERT_STATUS_NO_INSERT) {
1906 bch_check_keys(&b->keys, "%u for %s", status,
1907 replace_key ? "replace" : "insert");
1909 trace_bcache_btree_insert_key(b, k, replace_key != NULL,
1916 static size_t insert_u64s_remaining(struct btree *b)
1918 long ret = bch_btree_keys_u64s_remaining(&b->keys);
1921 * Might land in the middle of an existing extent and have to split it
1923 if (b->keys.ops->is_extents)
1924 ret -= KEY_MAX_U64S;
1926 return max(ret, 0L);
1929 static bool bch_btree_insert_keys(struct btree *b, struct btree_op *op,
1930 struct keylist *insert_keys,
1931 struct bkey *replace_key)
1934 int oldsize = bch_count_data(&b->keys);
1936 while (!bch_keylist_empty(insert_keys)) {
1937 struct bkey *k = insert_keys->keys;
1939 if (bkey_u64s(k) > insert_u64s_remaining(b))
1942 if (bkey_cmp(k, &b->key) <= 0) {
1946 ret |= btree_insert_key(b, k, replace_key);
1947 bch_keylist_pop_front(insert_keys);
1948 } else if (bkey_cmp(&START_KEY(k), &b->key) < 0) {
1949 BKEY_PADDED(key) temp;
1950 bkey_copy(&temp.key, insert_keys->keys);
1952 bch_cut_back(&b->key, &temp.key);
1953 bch_cut_front(&b->key, insert_keys->keys);
1955 ret |= btree_insert_key(b, &temp.key, replace_key);
1963 op->insert_collision = true;
1965 BUG_ON(!bch_keylist_empty(insert_keys) && b->level);
1967 BUG_ON(bch_count_data(&b->keys) < oldsize);
1971 static int btree_split(struct btree *b, struct btree_op *op,
1972 struct keylist *insert_keys,
1973 struct bkey *replace_key)
1976 struct btree *n1, *n2 = NULL, *n3 = NULL;
1977 uint64_t start_time = local_clock();
1979 struct keylist parent_keys;
1981 closure_init_stack(&cl);
1982 bch_keylist_init(&parent_keys);
1984 if (btree_check_reserve(b, op)) {
1988 WARN(1, "insufficient reserve for split\n");
1991 n1 = btree_node_alloc_replacement(b, op);
1995 split = set_blocks(btree_bset_first(n1),
1996 block_bytes(n1->c)) > (btree_blocks(b) * 4) / 5;
2001 trace_bcache_btree_node_split(b, btree_bset_first(n1)->keys);
2003 n2 = bch_btree_node_alloc(b->c, op, b->level, b->parent);
2008 n3 = bch_btree_node_alloc(b->c, op, b->level + 1, NULL);
2013 mutex_lock(&n1->write_lock);
2014 mutex_lock(&n2->write_lock);
2016 bch_btree_insert_keys(n1, op, insert_keys, replace_key);
2019 * Has to be a linear search because we don't have an auxiliary
2023 while (keys < (btree_bset_first(n1)->keys * 3) / 5)
2024 keys += bkey_u64s(bset_bkey_idx(btree_bset_first(n1),
2027 bkey_copy_key(&n1->key,
2028 bset_bkey_idx(btree_bset_first(n1), keys));
2029 keys += bkey_u64s(bset_bkey_idx(btree_bset_first(n1), keys));
2031 btree_bset_first(n2)->keys = btree_bset_first(n1)->keys - keys;
2032 btree_bset_first(n1)->keys = keys;
2034 memcpy(btree_bset_first(n2)->start,
2035 bset_bkey_last(btree_bset_first(n1)),
2036 btree_bset_first(n2)->keys * sizeof(uint64_t));
2038 bkey_copy_key(&n2->key, &b->key);
2040 bch_keylist_add(&parent_keys, &n2->key);
2041 bch_btree_node_write(n2, &cl);
2042 mutex_unlock(&n2->write_lock);
2043 rw_unlock(true, n2);
2045 trace_bcache_btree_node_compact(b, btree_bset_first(n1)->keys);
2047 mutex_lock(&n1->write_lock);
2048 bch_btree_insert_keys(n1, op, insert_keys, replace_key);
2051 bch_keylist_add(&parent_keys, &n1->key);
2052 bch_btree_node_write(n1, &cl);
2053 mutex_unlock(&n1->write_lock);
2056 /* Depth increases, make a new root */
2057 mutex_lock(&n3->write_lock);
2058 bkey_copy_key(&n3->key, &MAX_KEY);
2059 bch_btree_insert_keys(n3, op, &parent_keys, NULL);
2060 bch_btree_node_write(n3, &cl);
2061 mutex_unlock(&n3->write_lock);
2064 bch_btree_set_root(n3);
2065 rw_unlock(true, n3);
2066 } else if (!b->parent) {
2067 /* Root filled up but didn't need to be split */
2069 bch_btree_set_root(n1);
2071 /* Split a non root node */
2073 make_btree_freeing_key(b, parent_keys.top);
2074 bch_keylist_push(&parent_keys);
2076 bch_btree_insert_node(b->parent, op, &parent_keys, NULL, NULL);
2077 BUG_ON(!bch_keylist_empty(&parent_keys));
2081 rw_unlock(true, n1);
2083 bch_time_stats_update(&b->c->btree_split_time, start_time);
2087 bkey_put(b->c, &n2->key);
2088 btree_node_free(n2);
2089 rw_unlock(true, n2);
2091 bkey_put(b->c, &n1->key);
2092 btree_node_free(n1);
2093 rw_unlock(true, n1);
2095 WARN(1, "bcache: btree split failed (level %u)", b->level);
2097 if (n3 == ERR_PTR(-EAGAIN) ||
2098 n2 == ERR_PTR(-EAGAIN) ||
2099 n1 == ERR_PTR(-EAGAIN))
2105 static int bch_btree_insert_node(struct btree *b, struct btree_op *op,
2106 struct keylist *insert_keys,
2107 atomic_t *journal_ref,
2108 struct bkey *replace_key)
2112 BUG_ON(b->level && replace_key);
2114 closure_init_stack(&cl);
2116 mutex_lock(&b->write_lock);
2118 if (write_block(b) != btree_bset_last(b) &&
2119 b->keys.last_set_unwritten)
2120 bch_btree_init_next(b); /* just wrote a set */
2122 if (bch_keylist_nkeys(insert_keys) > insert_u64s_remaining(b)) {
2123 mutex_unlock(&b->write_lock);
2127 BUG_ON(write_block(b) != btree_bset_last(b));
2129 if (bch_btree_insert_keys(b, op, insert_keys, replace_key)) {
2131 bch_btree_leaf_dirty(b, journal_ref);
2133 bch_btree_node_write(b, &cl);
2136 mutex_unlock(&b->write_lock);
2138 /* wait for btree node write if necessary, after unlock */
2143 if (current->bio_list) {
2144 op->lock = b->c->root->level + 1;
2146 } else if (op->lock <= b->c->root->level) {
2147 op->lock = b->c->root->level + 1;
2150 /* Invalidated all iterators */
2151 int ret = btree_split(b, op, insert_keys, replace_key);
2153 if (bch_keylist_empty(insert_keys))
2161 int bch_btree_insert_check_key(struct btree *b, struct btree_op *op,
2162 struct bkey *check_key)
2165 uint64_t btree_ptr = b->key.ptr[0];
2166 unsigned long seq = b->seq;
2167 struct keylist insert;
2168 bool upgrade = op->lock == -1;
2170 bch_keylist_init(&insert);
2173 rw_unlock(false, b);
2174 rw_lock(true, b, b->level);
2176 if (b->key.ptr[0] != btree_ptr ||
2177 b->seq != seq + 1) {
2178 op->lock = b->level;
2183 SET_KEY_PTRS(check_key, 1);
2184 get_random_bytes(&check_key->ptr[0], sizeof(uint64_t));
2186 SET_PTR_DEV(check_key, 0, PTR_CHECK_DEV);
2188 bch_keylist_add(&insert, check_key);
2190 ret = bch_btree_insert_node(b, op, &insert, NULL, NULL);
2192 BUG_ON(!ret && !bch_keylist_empty(&insert));
2195 downgrade_write(&b->lock);
2199 struct btree_insert_op {
2201 struct keylist *keys;
2202 atomic_t *journal_ref;
2203 struct bkey *replace_key;
2206 static int btree_insert_fn(struct btree_op *b_op, struct btree *b)
2208 struct btree_insert_op *op = container_of(b_op,
2209 struct btree_insert_op, op);
2211 int ret = bch_btree_insert_node(b, &op->op, op->keys,
2212 op->journal_ref, op->replace_key);
2213 if (ret && !bch_keylist_empty(op->keys))
2219 int bch_btree_insert(struct cache_set *c, struct keylist *keys,
2220 atomic_t *journal_ref, struct bkey *replace_key)
2222 struct btree_insert_op op;
2225 BUG_ON(current->bio_list);
2226 BUG_ON(bch_keylist_empty(keys));
2228 bch_btree_op_init(&op.op, 0);
2230 op.journal_ref = journal_ref;
2231 op.replace_key = replace_key;
2233 while (!ret && !bch_keylist_empty(keys)) {
2235 ret = bch_btree_map_leaf_nodes(&op.op, c,
2236 &START_KEY(keys->keys),
2243 pr_err("error %i", ret);
2245 while ((k = bch_keylist_pop(keys)))
2247 } else if (op.op.insert_collision)
2253 void bch_btree_set_root(struct btree *b)
2258 closure_init_stack(&cl);
2260 trace_bcache_btree_set_root(b);
2262 BUG_ON(!b->written);
2264 for (i = 0; i < KEY_PTRS(&b->key); i++)
2265 BUG_ON(PTR_BUCKET(b->c, &b->key, i)->prio != BTREE_PRIO);
2267 mutex_lock(&b->c->bucket_lock);
2268 list_del_init(&b->list);
2269 mutex_unlock(&b->c->bucket_lock);
2273 bch_journal_meta(b->c, &cl);
2277 /* Map across nodes or keys */
2279 static int bch_btree_map_nodes_recurse(struct btree *b, struct btree_op *op,
2281 btree_map_nodes_fn *fn, int flags)
2283 int ret = MAP_CONTINUE;
2287 struct btree_iter iter;
2289 bch_btree_iter_init(&b->keys, &iter, from);
2291 while ((k = bch_btree_iter_next_filter(&iter, &b->keys,
2293 ret = btree(map_nodes_recurse, k, b,
2294 op, from, fn, flags);
2297 if (ret != MAP_CONTINUE)
2302 if (!b->level || flags == MAP_ALL_NODES)
2308 int __bch_btree_map_nodes(struct btree_op *op, struct cache_set *c,
2309 struct bkey *from, btree_map_nodes_fn *fn, int flags)
2311 return btree_root(map_nodes_recurse, c, op, from, fn, flags);
2314 static int bch_btree_map_keys_recurse(struct btree *b, struct btree_op *op,
2315 struct bkey *from, btree_map_keys_fn *fn,
2318 int ret = MAP_CONTINUE;
2320 struct btree_iter iter;
2322 bch_btree_iter_init(&b->keys, &iter, from);
2324 while ((k = bch_btree_iter_next_filter(&iter, &b->keys, bch_ptr_bad))) {
2327 : btree(map_keys_recurse, k, b, op, from, fn, flags);
2330 if (ret != MAP_CONTINUE)
2334 if (!b->level && (flags & MAP_END_KEY))
2335 ret = fn(op, b, &KEY(KEY_INODE(&b->key),
2336 KEY_OFFSET(&b->key), 0));
2341 int bch_btree_map_keys(struct btree_op *op, struct cache_set *c,
2342 struct bkey *from, btree_map_keys_fn *fn, int flags)
2344 return btree_root(map_keys_recurse, c, op, from, fn, flags);
2349 static inline int keybuf_cmp(struct keybuf_key *l, struct keybuf_key *r)
2351 /* Overlapping keys compare equal */
2352 if (bkey_cmp(&l->key, &START_KEY(&r->key)) <= 0)
2354 if (bkey_cmp(&START_KEY(&l->key), &r->key) >= 0)
2359 static inline int keybuf_nonoverlapping_cmp(struct keybuf_key *l,
2360 struct keybuf_key *r)
2362 return clamp_t(int64_t, bkey_cmp(&l->key, &r->key), -1, 1);
2370 keybuf_pred_fn *pred;
2373 static int refill_keybuf_fn(struct btree_op *op, struct btree *b,
2376 struct refill *refill = container_of(op, struct refill, op);
2377 struct keybuf *buf = refill->buf;
2378 int ret = MAP_CONTINUE;
2380 if (bkey_cmp(k, refill->end) > 0) {
2385 if (!KEY_SIZE(k)) /* end key */
2388 if (refill->pred(buf, k)) {
2389 struct keybuf_key *w;
2391 spin_lock(&buf->lock);
2393 w = array_alloc(&buf->freelist);
2395 spin_unlock(&buf->lock);
2400 bkey_copy(&w->key, k);
2402 if (RB_INSERT(&buf->keys, w, node, keybuf_cmp))
2403 array_free(&buf->freelist, w);
2407 if (array_freelist_empty(&buf->freelist))
2410 spin_unlock(&buf->lock);
2413 buf->last_scanned = *k;
2417 void bch_refill_keybuf(struct cache_set *c, struct keybuf *buf,
2418 struct bkey *end, keybuf_pred_fn *pred)
2420 struct bkey start = buf->last_scanned;
2421 struct refill refill;
2425 bch_btree_op_init(&refill.op, -1);
2426 refill.nr_found = 0;
2431 bch_btree_map_keys(&refill.op, c, &buf->last_scanned,
2432 refill_keybuf_fn, MAP_END_KEY);
2434 trace_bcache_keyscan(refill.nr_found,
2435 KEY_INODE(&start), KEY_OFFSET(&start),
2436 KEY_INODE(&buf->last_scanned),
2437 KEY_OFFSET(&buf->last_scanned));
2439 spin_lock(&buf->lock);
2441 if (!RB_EMPTY_ROOT(&buf->keys)) {
2442 struct keybuf_key *w;
2443 w = RB_FIRST(&buf->keys, struct keybuf_key, node);
2444 buf->start = START_KEY(&w->key);
2446 w = RB_LAST(&buf->keys, struct keybuf_key, node);
2449 buf->start = MAX_KEY;
2453 spin_unlock(&buf->lock);
2456 static void __bch_keybuf_del(struct keybuf *buf, struct keybuf_key *w)
2458 rb_erase(&w->node, &buf->keys);
2459 array_free(&buf->freelist, w);
2462 void bch_keybuf_del(struct keybuf *buf, struct keybuf_key *w)
2464 spin_lock(&buf->lock);
2465 __bch_keybuf_del(buf, w);
2466 spin_unlock(&buf->lock);
2469 bool bch_keybuf_check_overlapping(struct keybuf *buf, struct bkey *start,
2473 struct keybuf_key *p, *w, s;
2476 if (bkey_cmp(end, &buf->start) <= 0 ||
2477 bkey_cmp(start, &buf->end) >= 0)
2480 spin_lock(&buf->lock);
2481 w = RB_GREATER(&buf->keys, s, node, keybuf_nonoverlapping_cmp);
2483 while (w && bkey_cmp(&START_KEY(&w->key), end) < 0) {
2485 w = RB_NEXT(w, node);
2490 __bch_keybuf_del(buf, p);
2493 spin_unlock(&buf->lock);
2497 struct keybuf_key *bch_keybuf_next(struct keybuf *buf)
2499 struct keybuf_key *w;
2500 spin_lock(&buf->lock);
2502 w = RB_FIRST(&buf->keys, struct keybuf_key, node);
2504 while (w && w->private)
2505 w = RB_NEXT(w, node);
2508 w->private = ERR_PTR(-EINTR);
2510 spin_unlock(&buf->lock);
2514 struct keybuf_key *bch_keybuf_next_rescan(struct cache_set *c,
2517 keybuf_pred_fn *pred)
2519 struct keybuf_key *ret;
2522 ret = bch_keybuf_next(buf);
2526 if (bkey_cmp(&buf->last_scanned, end) >= 0) {
2527 pr_debug("scan finished");
2531 bch_refill_keybuf(c, buf, end, pred);
2537 void bch_keybuf_init(struct keybuf *buf)
2539 buf->last_scanned = MAX_KEY;
2540 buf->keys = RB_ROOT;
2542 spin_lock_init(&buf->lock);
2543 array_allocator_init(&buf->freelist);