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/freezer.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 <trace/events/bcache.h>
40 * register_bcache: Return errors out to userspace correctly
42 * Writeback: don't undirty key until after a cache flush
44 * Create an iterator for key pointers
46 * On btree write error, mark bucket such that it won't be freed from the cache
49 * Check for bad keys in replay
51 * Refcount journal entries in journal_replay
54 * Finish incremental gc
55 * Gc should free old UUIDs, data for invalid UUIDs
57 * Provide a way to list backing device UUIDs we have data cached for, and
58 * probably how long it's been since we've seen them, and a way to invalidate
59 * dirty data for devices that will never be attached again
61 * Keep 1 min/5 min/15 min statistics of how busy a block device has been, so
62 * that based on that and how much dirty data we have we can keep writeback
65 * Add a tracepoint or somesuch to watch for writeback starvation
67 * When btree depth > 1 and splitting an interior node, we have to make sure
68 * alloc_bucket() cannot fail. This should be true but is not completely
73 * If data write is less than hard sector size of ssd, round up offset in open
74 * bucket to the next whole sector
76 * Superblock needs to be fleshed out for multiple cache devices
78 * Add a sysfs tunable for the number of writeback IOs in flight
80 * Add a sysfs tunable for the number of open data buckets
82 * IO tracking: Can we track when one process is doing io on behalf of another?
83 * IO tracking: Don't use just an average, weigh more recent stuff higher
85 * Test module load/unload
88 #define MAX_NEED_GC 64
89 #define MAX_SAVE_PRIO 72
91 #define PTR_DIRTY_BIT (((uint64_t) 1 << 36))
93 #define PTR_HASH(c, k) \
94 (((k)->ptr[0] >> c->bucket_bits) | PTR_GEN(k, 0))
96 #define insert_lock(s, b) ((b)->level <= (s)->lock)
99 * These macros are for recursing down the btree - they handle the details of
100 * locking and looking up nodes in the cache for you. They're best treated as
101 * mere syntax when reading code that uses them.
103 * op->lock determines whether we take a read or a write lock at a given depth.
104 * If you've got a read lock and find that you need a write lock (i.e. you're
105 * going to have to split), set op->lock and return -EINTR; btree_root() will
106 * call you again and you'll have the correct lock.
110 * btree - recurse down the btree on a specified key
111 * @fn: function to call, which will be passed the child node
112 * @key: key to recurse on
113 * @b: parent btree node
114 * @op: pointer to struct btree_op
116 #define btree(fn, key, b, op, ...) \
118 int _r, l = (b)->level - 1; \
119 bool _w = l <= (op)->lock; \
120 struct btree *_child = bch_btree_node_get((b)->c, op, key, l, _w);\
121 if (!IS_ERR(_child)) { \
122 _child->parent = (b); \
123 _r = bch_btree_ ## fn(_child, op, ##__VA_ARGS__); \
124 rw_unlock(_w, _child); \
126 _r = PTR_ERR(_child); \
131 * btree_root - call a function on the root of the btree
132 * @fn: function to call, which will be passed the child node
134 * @op: pointer to struct btree_op
136 #define btree_root(fn, c, op, ...) \
140 struct btree *_b = (c)->root; \
141 bool _w = insert_lock(op, _b); \
142 rw_lock(_w, _b, _b->level); \
143 if (_b == (c)->root && \
144 _w == insert_lock(op, _b)) { \
146 _r = bch_btree_ ## fn(_b, op, ##__VA_ARGS__); \
149 bch_cannibalize_unlock(c); \
152 } while (_r == -EINTR); \
154 finish_wait(&(c)->btree_cache_wait, &(op)->wait); \
158 static inline struct bset *write_block(struct btree *b)
160 return ((void *) btree_bset_first(b)) + b->written * block_bytes(b->c);
163 static void bch_btree_init_next(struct btree *b)
165 /* If not a leaf node, always sort */
166 if (b->level && b->keys.nsets)
167 bch_btree_sort(&b->keys, &b->c->sort);
169 bch_btree_sort_lazy(&b->keys, &b->c->sort);
171 if (b->written < btree_blocks(b))
172 bch_bset_init_next(&b->keys, write_block(b),
173 bset_magic(&b->c->sb));
177 /* Btree key manipulation */
179 void bkey_put(struct cache_set *c, struct bkey *k)
183 for (i = 0; i < KEY_PTRS(k); i++)
184 if (ptr_available(c, k, i))
185 atomic_dec_bug(&PTR_BUCKET(c, k, i)->pin);
190 static uint64_t btree_csum_set(struct btree *b, struct bset *i)
192 uint64_t crc = b->key.ptr[0];
193 void *data = (void *) i + 8, *end = bset_bkey_last(i);
195 crc = bch_crc64_update(crc, data, end - data);
196 return crc ^ 0xffffffffffffffffULL;
199 void bch_btree_node_read_done(struct btree *b)
201 const char *err = "bad btree header";
202 struct bset *i = btree_bset_first(b);
203 struct btree_iter *iter;
205 iter = mempool_alloc(b->c->fill_iter, GFP_NOWAIT);
206 iter->size = b->c->sb.bucket_size / b->c->sb.block_size;
209 #ifdef CONFIG_BCACHE_DEBUG
217 b->written < btree_blocks(b) && i->seq == b->keys.set[0].data->seq;
218 i = write_block(b)) {
219 err = "unsupported bset version";
220 if (i->version > BCACHE_BSET_VERSION)
223 err = "bad btree header";
224 if (b->written + set_blocks(i, block_bytes(b->c)) >
229 if (i->magic != bset_magic(&b->c->sb))
232 err = "bad checksum";
233 switch (i->version) {
235 if (i->csum != csum_set(i))
238 case BCACHE_BSET_VERSION:
239 if (i->csum != btree_csum_set(b, i))
245 if (i != b->keys.set[0].data && !i->keys)
248 bch_btree_iter_push(iter, i->start, bset_bkey_last(i));
250 b->written += set_blocks(i, block_bytes(b->c));
253 err = "corrupted btree";
254 for (i = write_block(b);
255 bset_sector_offset(&b->keys, i) < KEY_SIZE(&b->key);
256 i = ((void *) i) + block_bytes(b->c))
257 if (i->seq == b->keys.set[0].data->seq)
260 bch_btree_sort_and_fix_extents(&b->keys, iter, &b->c->sort);
262 i = b->keys.set[0].data;
263 err = "short btree key";
264 if (b->keys.set[0].size &&
265 bkey_cmp(&b->key, &b->keys.set[0].end) < 0)
268 if (b->written < btree_blocks(b))
269 bch_bset_init_next(&b->keys, write_block(b),
270 bset_magic(&b->c->sb));
272 mempool_free(iter, b->c->fill_iter);
275 set_btree_node_io_error(b);
276 bch_cache_set_error(b->c, "%s at bucket %zu, block %u, %u keys",
277 err, PTR_BUCKET_NR(b->c, &b->key, 0),
278 bset_block_offset(b, i), i->keys);
282 static void btree_node_read_endio(struct bio *bio, int error)
284 struct closure *cl = bio->bi_private;
288 static void bch_btree_node_read(struct btree *b)
290 uint64_t start_time = local_clock();
294 trace_bcache_btree_read(b);
296 closure_init_stack(&cl);
298 bio = bch_bbio_alloc(b->c);
299 bio->bi_rw = REQ_META|READ_SYNC;
300 bio->bi_iter.bi_size = KEY_SIZE(&b->key) << 9;
301 bio->bi_end_io = btree_node_read_endio;
302 bio->bi_private = &cl;
304 bch_bio_map(bio, b->keys.set[0].data);
306 bch_submit_bbio(bio, b->c, &b->key, 0);
309 if (!test_bit(BIO_UPTODATE, &bio->bi_flags))
310 set_btree_node_io_error(b);
312 bch_bbio_free(bio, b->c);
314 if (btree_node_io_error(b))
317 bch_btree_node_read_done(b);
318 bch_time_stats_update(&b->c->btree_read_time, start_time);
322 bch_cache_set_error(b->c, "io error reading bucket %zu",
323 PTR_BUCKET_NR(b->c, &b->key, 0));
326 static void btree_complete_write(struct btree *b, struct btree_write *w)
328 if (w->prio_blocked &&
329 !atomic_sub_return(w->prio_blocked, &b->c->prio_blocked))
330 wake_up_allocators(b->c);
333 atomic_dec_bug(w->journal);
334 __closure_wake_up(&b->c->journal.wait);
341 static void btree_node_write_unlock(struct closure *cl)
343 struct btree *b = container_of(cl, struct btree, io);
348 static void __btree_node_write_done(struct closure *cl)
350 struct btree *b = container_of(cl, struct btree, io);
351 struct btree_write *w = btree_prev_write(b);
353 bch_bbio_free(b->bio, b->c);
355 btree_complete_write(b, w);
357 if (btree_node_dirty(b))
358 schedule_delayed_work(&b->work, 30 * HZ);
360 closure_return_with_destructor(cl, btree_node_write_unlock);
363 static void btree_node_write_done(struct closure *cl)
365 struct btree *b = container_of(cl, struct btree, io);
369 bio_for_each_segment_all(bv, b->bio, n)
370 __free_page(bv->bv_page);
372 __btree_node_write_done(cl);
375 static void btree_node_write_endio(struct bio *bio, int error)
377 struct closure *cl = bio->bi_private;
378 struct btree *b = container_of(cl, struct btree, io);
381 set_btree_node_io_error(b);
383 bch_bbio_count_io_errors(b->c, bio, error, "writing btree");
387 static void do_btree_node_write(struct btree *b)
389 struct closure *cl = &b->io;
390 struct bset *i = btree_bset_last(b);
393 i->version = BCACHE_BSET_VERSION;
394 i->csum = btree_csum_set(b, i);
397 b->bio = bch_bbio_alloc(b->c);
399 b->bio->bi_end_io = btree_node_write_endio;
400 b->bio->bi_private = cl;
401 b->bio->bi_rw = REQ_META|WRITE_SYNC|REQ_FUA;
402 b->bio->bi_iter.bi_size = roundup(set_bytes(i), block_bytes(b->c));
403 bch_bio_map(b->bio, i);
406 * If we're appending to a leaf node, we don't technically need FUA -
407 * this write just needs to be persisted before the next journal write,
408 * which will be marked FLUSH|FUA.
410 * Similarly if we're writing a new btree root - the pointer is going to
411 * be in the next journal entry.
413 * But if we're writing a new btree node (that isn't a root) or
414 * appending to a non leaf btree node, we need either FUA or a flush
415 * when we write the parent with the new pointer. FUA is cheaper than a
416 * flush, and writes appending to leaf nodes aren't blocking anything so
417 * just make all btree node writes FUA to keep things sane.
420 bkey_copy(&k.key, &b->key);
421 SET_PTR_OFFSET(&k.key, 0, PTR_OFFSET(&k.key, 0) +
422 bset_sector_offset(&b->keys, i));
424 if (!bio_alloc_pages(b->bio, GFP_NOIO)) {
427 void *base = (void *) ((unsigned long) i & ~(PAGE_SIZE - 1));
429 bio_for_each_segment_all(bv, b->bio, j)
430 memcpy(page_address(bv->bv_page),
431 base + j * PAGE_SIZE, PAGE_SIZE);
433 bch_submit_bbio(b->bio, b->c, &k.key, 0);
435 continue_at(cl, btree_node_write_done, NULL);
438 bch_bio_map(b->bio, i);
440 bch_submit_bbio(b->bio, b->c, &k.key, 0);
443 continue_at_nobarrier(cl, __btree_node_write_done, NULL);
447 void __bch_btree_node_write(struct btree *b, struct closure *parent)
449 struct bset *i = btree_bset_last(b);
451 lockdep_assert_held(&b->write_lock);
453 trace_bcache_btree_write(b);
455 BUG_ON(current->bio_list);
456 BUG_ON(b->written >= btree_blocks(b));
457 BUG_ON(b->written && !i->keys);
458 BUG_ON(btree_bset_first(b)->seq != i->seq);
459 bch_check_keys(&b->keys, "writing");
461 cancel_delayed_work(&b->work);
463 /* If caller isn't waiting for write, parent refcount is cache set */
465 closure_init(&b->io, parent ?: &b->c->cl);
467 clear_bit(BTREE_NODE_dirty, &b->flags);
468 change_bit(BTREE_NODE_write_idx, &b->flags);
470 do_btree_node_write(b);
472 atomic_long_add(set_blocks(i, block_bytes(b->c)) * b->c->sb.block_size,
473 &PTR_CACHE(b->c, &b->key, 0)->btree_sectors_written);
475 b->written += set_blocks(i, block_bytes(b->c));
478 void bch_btree_node_write(struct btree *b, struct closure *parent)
480 unsigned nsets = b->keys.nsets;
482 lockdep_assert_held(&b->lock);
484 __bch_btree_node_write(b, parent);
487 * do verify if there was more than one set initially (i.e. we did a
488 * sort) and we sorted down to a single set:
490 if (nsets && !b->keys.nsets)
493 bch_btree_init_next(b);
496 static void bch_btree_node_write_sync(struct btree *b)
500 closure_init_stack(&cl);
502 mutex_lock(&b->write_lock);
503 bch_btree_node_write(b, &cl);
504 mutex_unlock(&b->write_lock);
509 static void btree_node_write_work(struct work_struct *w)
511 struct btree *b = container_of(to_delayed_work(w), struct btree, work);
513 mutex_lock(&b->write_lock);
514 if (btree_node_dirty(b))
515 __bch_btree_node_write(b, NULL);
516 mutex_unlock(&b->write_lock);
519 static void bch_btree_leaf_dirty(struct btree *b, atomic_t *journal_ref)
521 struct bset *i = btree_bset_last(b);
522 struct btree_write *w = btree_current_write(b);
524 lockdep_assert_held(&b->write_lock);
529 if (!btree_node_dirty(b))
530 schedule_delayed_work(&b->work, 30 * HZ);
532 set_btree_node_dirty(b);
536 journal_pin_cmp(b->c, w->journal, journal_ref)) {
537 atomic_dec_bug(w->journal);
542 w->journal = journal_ref;
543 atomic_inc(w->journal);
547 /* Force write if set is too big */
548 if (set_bytes(i) > PAGE_SIZE - 48 &&
550 bch_btree_node_write(b, NULL);
554 * Btree in memory cache - allocation/freeing
555 * mca -> memory cache
558 #define mca_reserve(c) (((c->root && c->root->level) \
559 ? c->root->level : 1) * 8 + 16)
560 #define mca_can_free(c) \
561 max_t(int, 0, c->btree_cache_used - mca_reserve(c))
563 static void mca_data_free(struct btree *b)
565 BUG_ON(b->io_mutex.count != 1);
567 bch_btree_keys_free(&b->keys);
569 b->c->btree_cache_used--;
570 list_move(&b->list, &b->c->btree_cache_freed);
573 static void mca_bucket_free(struct btree *b)
575 BUG_ON(btree_node_dirty(b));
578 hlist_del_init_rcu(&b->hash);
579 list_move(&b->list, &b->c->btree_cache_freeable);
582 static unsigned btree_order(struct bkey *k)
584 return ilog2(KEY_SIZE(k) / PAGE_SECTORS ?: 1);
587 static void mca_data_alloc(struct btree *b, struct bkey *k, gfp_t gfp)
589 if (!bch_btree_keys_alloc(&b->keys,
591 ilog2(b->c->btree_pages),
594 b->c->btree_cache_used++;
595 list_move(&b->list, &b->c->btree_cache);
597 list_move(&b->list, &b->c->btree_cache_freed);
601 static struct btree *mca_bucket_alloc(struct cache_set *c,
602 struct bkey *k, gfp_t gfp)
604 struct btree *b = kzalloc(sizeof(struct btree), gfp);
608 init_rwsem(&b->lock);
609 lockdep_set_novalidate_class(&b->lock);
610 mutex_init(&b->write_lock);
611 lockdep_set_novalidate_class(&b->write_lock);
612 INIT_LIST_HEAD(&b->list);
613 INIT_DELAYED_WORK(&b->work, btree_node_write_work);
615 sema_init(&b->io_mutex, 1);
617 mca_data_alloc(b, k, gfp);
621 static int mca_reap(struct btree *b, unsigned min_order, bool flush)
625 closure_init_stack(&cl);
626 lockdep_assert_held(&b->c->bucket_lock);
628 if (!down_write_trylock(&b->lock))
631 BUG_ON(btree_node_dirty(b) && !b->keys.set[0].data);
633 if (b->keys.page_order < min_order)
637 if (btree_node_dirty(b))
640 if (down_trylock(&b->io_mutex))
645 mutex_lock(&b->write_lock);
646 if (btree_node_dirty(b))
647 __bch_btree_node_write(b, &cl);
648 mutex_unlock(&b->write_lock);
652 /* wait for any in flight btree write */
662 static unsigned long bch_mca_scan(struct shrinker *shrink,
663 struct shrink_control *sc)
665 struct cache_set *c = container_of(shrink, struct cache_set, shrink);
667 unsigned long i, nr = sc->nr_to_scan;
668 unsigned long freed = 0;
670 if (c->shrinker_disabled)
673 if (c->btree_cache_alloc_lock)
676 /* Return -1 if we can't do anything right now */
677 if (sc->gfp_mask & __GFP_IO)
678 mutex_lock(&c->bucket_lock);
679 else if (!mutex_trylock(&c->bucket_lock))
683 * It's _really_ critical that we don't free too many btree nodes - we
684 * have to always leave ourselves a reserve. The reserve is how we
685 * guarantee that allocating memory for a new btree node can always
686 * succeed, so that inserting keys into the btree can always succeed and
687 * IO can always make forward progress:
689 nr /= c->btree_pages;
690 nr = min_t(unsigned long, nr, mca_can_free(c));
693 list_for_each_entry_safe(b, t, &c->btree_cache_freeable, list) {
698 !mca_reap(b, 0, false)) {
705 for (i = 0; (nr--) && i < c->btree_cache_used; i++) {
706 if (list_empty(&c->btree_cache))
709 b = list_first_entry(&c->btree_cache, struct btree, list);
710 list_rotate_left(&c->btree_cache);
713 !mca_reap(b, 0, false)) {
722 mutex_unlock(&c->bucket_lock);
726 static unsigned long bch_mca_count(struct shrinker *shrink,
727 struct shrink_control *sc)
729 struct cache_set *c = container_of(shrink, struct cache_set, shrink);
731 if (c->shrinker_disabled)
734 if (c->btree_cache_alloc_lock)
737 return mca_can_free(c) * c->btree_pages;
740 void bch_btree_cache_free(struct cache_set *c)
744 closure_init_stack(&cl);
746 if (c->shrink.list.next)
747 unregister_shrinker(&c->shrink);
749 mutex_lock(&c->bucket_lock);
751 #ifdef CONFIG_BCACHE_DEBUG
753 list_move(&c->verify_data->list, &c->btree_cache);
755 free_pages((unsigned long) c->verify_ondisk, ilog2(bucket_pages(c)));
758 list_splice(&c->btree_cache_freeable,
761 while (!list_empty(&c->btree_cache)) {
762 b = list_first_entry(&c->btree_cache, struct btree, list);
764 if (btree_node_dirty(b))
765 btree_complete_write(b, btree_current_write(b));
766 clear_bit(BTREE_NODE_dirty, &b->flags);
771 while (!list_empty(&c->btree_cache_freed)) {
772 b = list_first_entry(&c->btree_cache_freed,
775 cancel_delayed_work_sync(&b->work);
779 mutex_unlock(&c->bucket_lock);
782 int bch_btree_cache_alloc(struct cache_set *c)
786 for (i = 0; i < mca_reserve(c); i++)
787 if (!mca_bucket_alloc(c, &ZERO_KEY, GFP_KERNEL))
790 list_splice_init(&c->btree_cache,
791 &c->btree_cache_freeable);
793 #ifdef CONFIG_BCACHE_DEBUG
794 mutex_init(&c->verify_lock);
796 c->verify_ondisk = (void *)
797 __get_free_pages(GFP_KERNEL, ilog2(bucket_pages(c)));
799 c->verify_data = mca_bucket_alloc(c, &ZERO_KEY, GFP_KERNEL);
801 if (c->verify_data &&
802 c->verify_data->keys.set->data)
803 list_del_init(&c->verify_data->list);
805 c->verify_data = NULL;
808 c->shrink.count_objects = bch_mca_count;
809 c->shrink.scan_objects = bch_mca_scan;
811 c->shrink.batch = c->btree_pages * 2;
812 register_shrinker(&c->shrink);
817 /* Btree in memory cache - hash table */
819 static struct hlist_head *mca_hash(struct cache_set *c, struct bkey *k)
821 return &c->bucket_hash[hash_32(PTR_HASH(c, k), BUCKET_HASH_BITS)];
824 static struct btree *mca_find(struct cache_set *c, struct bkey *k)
829 hlist_for_each_entry_rcu(b, mca_hash(c, k), hash)
830 if (PTR_HASH(c, &b->key) == PTR_HASH(c, k))
838 static int mca_cannibalize_lock(struct cache_set *c, struct btree_op *op)
840 struct task_struct *old;
842 old = cmpxchg(&c->btree_cache_alloc_lock, NULL, current);
843 if (old && old != current) {
845 prepare_to_wait(&c->btree_cache_wait, &op->wait,
846 TASK_UNINTERRUPTIBLE);
853 static struct btree *mca_cannibalize(struct cache_set *c, struct btree_op *op,
858 trace_bcache_btree_cache_cannibalize(c);
860 if (mca_cannibalize_lock(c, op))
861 return ERR_PTR(-EINTR);
863 list_for_each_entry_reverse(b, &c->btree_cache, list)
864 if (!mca_reap(b, btree_order(k), false))
867 list_for_each_entry_reverse(b, &c->btree_cache, list)
868 if (!mca_reap(b, btree_order(k), true))
871 WARN(1, "btree cache cannibalize failed\n");
872 return ERR_PTR(-ENOMEM);
876 * We can only have one thread cannibalizing other cached btree nodes at a time,
877 * or we'll deadlock. We use an open coded mutex to ensure that, which a
878 * cannibalize_bucket() will take. This means every time we unlock the root of
879 * the btree, we need to release this lock if we have it held.
881 static void bch_cannibalize_unlock(struct cache_set *c)
883 if (c->btree_cache_alloc_lock == current) {
884 c->btree_cache_alloc_lock = NULL;
885 wake_up(&c->btree_cache_wait);
889 static struct btree *mca_alloc(struct cache_set *c, struct btree_op *op,
890 struct bkey *k, int level)
894 BUG_ON(current->bio_list);
896 lockdep_assert_held(&c->bucket_lock);
901 /* btree_free() doesn't free memory; it sticks the node on the end of
902 * the list. Check if there's any freed nodes there:
904 list_for_each_entry(b, &c->btree_cache_freeable, list)
905 if (!mca_reap(b, btree_order(k), false))
908 /* We never free struct btree itself, just the memory that holds the on
909 * disk node. Check the freed list before allocating a new one:
911 list_for_each_entry(b, &c->btree_cache_freed, list)
912 if (!mca_reap(b, 0, false)) {
913 mca_data_alloc(b, k, __GFP_NOWARN|GFP_NOIO);
914 if (!b->keys.set[0].data)
920 b = mca_bucket_alloc(c, k, __GFP_NOWARN|GFP_NOIO);
924 BUG_ON(!down_write_trylock(&b->lock));
925 if (!b->keys.set->data)
928 BUG_ON(b->io_mutex.count != 1);
930 bkey_copy(&b->key, k);
931 list_move(&b->list, &c->btree_cache);
932 hlist_del_init_rcu(&b->hash);
933 hlist_add_head_rcu(&b->hash, mca_hash(c, k));
935 lock_set_subclass(&b->lock.dep_map, level + 1, _THIS_IP_);
936 b->parent = (void *) ~0UL;
942 bch_btree_keys_init(&b->keys, &bch_extent_keys_ops,
943 &b->c->expensive_debug_checks);
945 bch_btree_keys_init(&b->keys, &bch_btree_keys_ops,
946 &b->c->expensive_debug_checks);
953 b = mca_cannibalize(c, op, k);
961 * bch_btree_node_get - find a btree node in the cache and lock it, reading it
962 * in from disk if necessary.
964 * If IO is necessary and running under generic_make_request, returns -EAGAIN.
966 * The btree node will have either a read or a write lock held, depending on
967 * level and op->lock.
969 struct btree *bch_btree_node_get(struct cache_set *c, struct btree_op *op,
970 struct bkey *k, int level, bool write)
980 if (current->bio_list)
981 return ERR_PTR(-EAGAIN);
983 mutex_lock(&c->bucket_lock);
984 b = mca_alloc(c, op, k, level);
985 mutex_unlock(&c->bucket_lock);
992 bch_btree_node_read(b);
995 downgrade_write(&b->lock);
997 rw_lock(write, b, level);
998 if (PTR_HASH(c, &b->key) != PTR_HASH(c, k)) {
1002 BUG_ON(b->level != level);
1007 for (; i <= b->keys.nsets && b->keys.set[i].size; i++) {
1008 prefetch(b->keys.set[i].tree);
1009 prefetch(b->keys.set[i].data);
1012 for (; i <= b->keys.nsets; i++)
1013 prefetch(b->keys.set[i].data);
1015 if (btree_node_io_error(b)) {
1016 rw_unlock(write, b);
1017 return ERR_PTR(-EIO);
1020 BUG_ON(!b->written);
1025 static void btree_node_prefetch(struct cache_set *c, struct bkey *k, int level)
1029 mutex_lock(&c->bucket_lock);
1030 b = mca_alloc(c, NULL, k, level);
1031 mutex_unlock(&c->bucket_lock);
1033 if (!IS_ERR_OR_NULL(b)) {
1034 bch_btree_node_read(b);
1041 static void btree_node_free(struct btree *b)
1043 trace_bcache_btree_node_free(b);
1045 BUG_ON(b == b->c->root);
1047 mutex_lock(&b->write_lock);
1049 if (btree_node_dirty(b))
1050 btree_complete_write(b, btree_current_write(b));
1051 clear_bit(BTREE_NODE_dirty, &b->flags);
1053 mutex_unlock(&b->write_lock);
1055 cancel_delayed_work(&b->work);
1057 mutex_lock(&b->c->bucket_lock);
1058 bch_bucket_free(b->c, &b->key);
1060 mutex_unlock(&b->c->bucket_lock);
1063 struct btree *bch_btree_node_alloc(struct cache_set *c, struct btree_op *op,
1067 struct btree *b = ERR_PTR(-EAGAIN);
1069 mutex_lock(&c->bucket_lock);
1071 if (__bch_bucket_alloc_set(c, RESERVE_BTREE, &k.key, 1, op != NULL))
1074 bkey_put(c, &k.key);
1075 SET_KEY_SIZE(&k.key, c->btree_pages * PAGE_SECTORS);
1077 b = mca_alloc(c, op, &k.key, level);
1083 "Tried to allocate bucket that was in btree cache");
1088 bch_bset_init_next(&b->keys, b->keys.set->data, bset_magic(&b->c->sb));
1090 mutex_unlock(&c->bucket_lock);
1092 trace_bcache_btree_node_alloc(b);
1095 bch_bucket_free(c, &k.key);
1097 mutex_unlock(&c->bucket_lock);
1099 trace_bcache_btree_node_alloc_fail(b);
1103 static struct btree *btree_node_alloc_replacement(struct btree *b,
1104 struct btree_op *op)
1106 struct btree *n = bch_btree_node_alloc(b->c, op, b->level);
1107 if (!IS_ERR_OR_NULL(n)) {
1108 mutex_lock(&n->write_lock);
1109 bch_btree_sort_into(&b->keys, &n->keys, &b->c->sort);
1110 bkey_copy_key(&n->key, &b->key);
1111 mutex_unlock(&n->write_lock);
1117 static void make_btree_freeing_key(struct btree *b, struct bkey *k)
1121 mutex_lock(&b->c->bucket_lock);
1123 atomic_inc(&b->c->prio_blocked);
1125 bkey_copy(k, &b->key);
1126 bkey_copy_key(k, &ZERO_KEY);
1128 for (i = 0; i < KEY_PTRS(k); i++)
1130 bch_inc_gen(PTR_CACHE(b->c, &b->key, i),
1131 PTR_BUCKET(b->c, &b->key, i)));
1133 mutex_unlock(&b->c->bucket_lock);
1136 static int btree_check_reserve(struct btree *b, struct btree_op *op)
1138 struct cache_set *c = b->c;
1140 unsigned i, reserve = (c->root->level - b->level) * 2 + 1;
1142 mutex_lock(&c->bucket_lock);
1144 for_each_cache(ca, c, i)
1145 if (fifo_used(&ca->free[RESERVE_BTREE]) < reserve) {
1147 prepare_to_wait(&c->btree_cache_wait, &op->wait,
1148 TASK_UNINTERRUPTIBLE);
1149 mutex_unlock(&c->bucket_lock);
1153 mutex_unlock(&c->bucket_lock);
1155 return mca_cannibalize_lock(b->c, op);
1158 /* Garbage collection */
1160 static uint8_t __bch_btree_mark_key(struct cache_set *c, int level,
1168 * ptr_invalid() can't return true for the keys that mark btree nodes as
1169 * freed, but since ptr_bad() returns true we'll never actually use them
1170 * for anything and thus we don't want mark their pointers here
1172 if (!bkey_cmp(k, &ZERO_KEY))
1175 for (i = 0; i < KEY_PTRS(k); i++) {
1176 if (!ptr_available(c, k, i))
1179 g = PTR_BUCKET(c, k, i);
1181 if (gen_after(g->gc_gen, PTR_GEN(k, i)))
1182 g->gc_gen = PTR_GEN(k, i);
1184 if (ptr_stale(c, k, i)) {
1185 stale = max(stale, ptr_stale(c, k, i));
1189 cache_bug_on(GC_MARK(g) &&
1190 (GC_MARK(g) == GC_MARK_METADATA) != (level != 0),
1191 c, "inconsistent ptrs: mark = %llu, level = %i",
1195 SET_GC_MARK(g, GC_MARK_METADATA);
1196 else if (KEY_DIRTY(k))
1197 SET_GC_MARK(g, GC_MARK_DIRTY);
1198 else if (!GC_MARK(g))
1199 SET_GC_MARK(g, GC_MARK_RECLAIMABLE);
1201 /* guard against overflow */
1202 SET_GC_SECTORS_USED(g, min_t(unsigned,
1203 GC_SECTORS_USED(g) + KEY_SIZE(k),
1204 MAX_GC_SECTORS_USED));
1206 BUG_ON(!GC_SECTORS_USED(g));
1212 #define btree_mark_key(b, k) __bch_btree_mark_key(b->c, b->level, k)
1214 void bch_initial_mark_key(struct cache_set *c, int level, struct bkey *k)
1218 for (i = 0; i < KEY_PTRS(k); i++)
1219 if (ptr_available(c, k, i) &&
1220 !ptr_stale(c, k, i)) {
1221 struct bucket *b = PTR_BUCKET(c, k, i);
1223 b->gen = PTR_GEN(k, i);
1225 if (level && bkey_cmp(k, &ZERO_KEY))
1226 b->prio = BTREE_PRIO;
1227 else if (!level && b->prio == BTREE_PRIO)
1228 b->prio = INITIAL_PRIO;
1231 __bch_btree_mark_key(c, level, k);
1234 static bool btree_gc_mark_node(struct btree *b, struct gc_stat *gc)
1237 unsigned keys = 0, good_keys = 0;
1239 struct btree_iter iter;
1240 struct bset_tree *t;
1244 for_each_key_filter(&b->keys, k, &iter, bch_ptr_invalid) {
1245 stale = max(stale, btree_mark_key(b, k));
1248 if (bch_ptr_bad(&b->keys, k))
1251 gc->key_bytes += bkey_u64s(k);
1255 gc->data += KEY_SIZE(k);
1258 for (t = b->keys.set; t <= &b->keys.set[b->keys.nsets]; t++)
1259 btree_bug_on(t->size &&
1260 bset_written(&b->keys, t) &&
1261 bkey_cmp(&b->key, &t->end) < 0,
1262 b, "found short btree key in gc");
1264 if (b->c->gc_always_rewrite)
1270 if ((keys - good_keys) * 2 > keys)
1276 #define GC_MERGE_NODES 4U
1278 struct gc_merge_info {
1283 static int bch_btree_insert_node(struct btree *, struct btree_op *,
1284 struct keylist *, atomic_t *, struct bkey *);
1286 static int btree_gc_coalesce(struct btree *b, struct btree_op *op,
1287 struct gc_stat *gc, struct gc_merge_info *r)
1289 unsigned i, nodes = 0, keys = 0, blocks;
1290 struct btree *new_nodes[GC_MERGE_NODES];
1291 struct keylist keylist;
1295 bch_keylist_init(&keylist);
1297 if (btree_check_reserve(b, NULL))
1300 memset(new_nodes, 0, sizeof(new_nodes));
1301 closure_init_stack(&cl);
1303 while (nodes < GC_MERGE_NODES && !IS_ERR_OR_NULL(r[nodes].b))
1304 keys += r[nodes++].keys;
1306 blocks = btree_default_blocks(b->c) * 2 / 3;
1309 __set_blocks(b->keys.set[0].data, keys,
1310 block_bytes(b->c)) > blocks * (nodes - 1))
1313 for (i = 0; i < nodes; i++) {
1314 new_nodes[i] = btree_node_alloc_replacement(r[i].b, NULL);
1315 if (IS_ERR_OR_NULL(new_nodes[i]))
1316 goto out_nocoalesce;
1320 * We have to check the reserve here, after we've allocated our new
1321 * nodes, to make sure the insert below will succeed - we also check
1322 * before as an optimization to potentially avoid a bunch of expensive
1325 if (btree_check_reserve(b, NULL))
1326 goto out_nocoalesce;
1328 for (i = 0; i < nodes; i++)
1329 mutex_lock(&new_nodes[i]->write_lock);
1331 for (i = nodes - 1; i > 0; --i) {
1332 struct bset *n1 = btree_bset_first(new_nodes[i]);
1333 struct bset *n2 = btree_bset_first(new_nodes[i - 1]);
1334 struct bkey *k, *last = NULL;
1340 k < bset_bkey_last(n2);
1342 if (__set_blocks(n1, n1->keys + keys +
1344 block_bytes(b->c)) > blocks)
1348 keys += bkey_u64s(k);
1352 * Last node we're not getting rid of - we're getting
1353 * rid of the node at r[0]. Have to try and fit all of
1354 * the remaining keys into this node; we can't ensure
1355 * they will always fit due to rounding and variable
1356 * length keys (shouldn't be possible in practice,
1359 if (__set_blocks(n1, n1->keys + n2->keys,
1360 block_bytes(b->c)) >
1361 btree_blocks(new_nodes[i]))
1362 goto out_nocoalesce;
1365 /* Take the key of the node we're getting rid of */
1369 BUG_ON(__set_blocks(n1, n1->keys + keys, block_bytes(b->c)) >
1370 btree_blocks(new_nodes[i]));
1373 bkey_copy_key(&new_nodes[i]->key, last);
1375 memcpy(bset_bkey_last(n1),
1377 (void *) bset_bkey_idx(n2, keys) - (void *) n2->start);
1380 r[i].keys = n1->keys;
1383 bset_bkey_idx(n2, keys),
1384 (void *) bset_bkey_last(n2) -
1385 (void *) bset_bkey_idx(n2, keys));
1389 if (__bch_keylist_realloc(&keylist,
1390 bkey_u64s(&new_nodes[i]->key)))
1391 goto out_nocoalesce;
1393 bch_btree_node_write(new_nodes[i], &cl);
1394 bch_keylist_add(&keylist, &new_nodes[i]->key);
1397 for (i = 0; i < nodes; i++)
1398 mutex_unlock(&new_nodes[i]->write_lock);
1402 /* We emptied out this node */
1403 BUG_ON(btree_bset_first(new_nodes[0])->keys);
1404 btree_node_free(new_nodes[0]);
1405 rw_unlock(true, new_nodes[0]);
1407 for (i = 0; i < nodes; i++) {
1408 if (__bch_keylist_realloc(&keylist, bkey_u64s(&r[i].b->key)))
1409 goto out_nocoalesce;
1411 make_btree_freeing_key(r[i].b, keylist.top);
1412 bch_keylist_push(&keylist);
1415 bch_btree_insert_node(b, op, &keylist, NULL, NULL);
1416 BUG_ON(!bch_keylist_empty(&keylist));
1418 for (i = 0; i < nodes; i++) {
1419 btree_node_free(r[i].b);
1420 rw_unlock(true, r[i].b);
1422 r[i].b = new_nodes[i];
1425 memmove(r, r + 1, sizeof(r[0]) * (nodes - 1));
1426 r[nodes - 1].b = ERR_PTR(-EINTR);
1428 trace_bcache_btree_gc_coalesce(nodes);
1431 bch_keylist_free(&keylist);
1433 /* Invalidated our iterator */
1438 bch_keylist_free(&keylist);
1440 while ((k = bch_keylist_pop(&keylist)))
1441 if (!bkey_cmp(k, &ZERO_KEY))
1442 atomic_dec(&b->c->prio_blocked);
1444 for (i = 0; i < nodes; i++)
1445 if (!IS_ERR_OR_NULL(new_nodes[i])) {
1446 btree_node_free(new_nodes[i]);
1447 rw_unlock(true, new_nodes[i]);
1452 static int btree_gc_rewrite_node(struct btree *b, struct btree_op *op,
1453 struct btree *replace)
1455 struct keylist keys;
1458 if (btree_check_reserve(b, NULL))
1461 n = btree_node_alloc_replacement(replace, NULL);
1463 /* recheck reserve after allocating replacement node */
1464 if (btree_check_reserve(b, NULL)) {
1470 bch_btree_node_write_sync(n);
1472 bch_keylist_init(&keys);
1473 bch_keylist_add(&keys, &n->key);
1475 make_btree_freeing_key(replace, keys.top);
1476 bch_keylist_push(&keys);
1478 bch_btree_insert_node(b, op, &keys, NULL, NULL);
1479 BUG_ON(!bch_keylist_empty(&keys));
1481 btree_node_free(replace);
1484 /* Invalidated our iterator */
1488 static unsigned btree_gc_count_keys(struct btree *b)
1491 struct btree_iter iter;
1494 for_each_key_filter(&b->keys, k, &iter, bch_ptr_bad)
1495 ret += bkey_u64s(k);
1500 static int btree_gc_recurse(struct btree *b, struct btree_op *op,
1501 struct closure *writes, struct gc_stat *gc)
1504 bool should_rewrite;
1506 struct btree_iter iter;
1507 struct gc_merge_info r[GC_MERGE_NODES];
1508 struct gc_merge_info *i, *last = r + ARRAY_SIZE(r) - 1;
1510 bch_btree_iter_init(&b->keys, &iter, &b->c->gc_done);
1512 for (i = r; i < r + ARRAY_SIZE(r); i++)
1513 i->b = ERR_PTR(-EINTR);
1516 k = bch_btree_iter_next_filter(&iter, &b->keys, bch_ptr_bad);
1518 r->b = bch_btree_node_get(b->c, op, k, b->level - 1,
1521 ret = PTR_ERR(r->b);
1525 r->keys = btree_gc_count_keys(r->b);
1527 ret = btree_gc_coalesce(b, op, gc, r);
1535 if (!IS_ERR(last->b)) {
1536 should_rewrite = btree_gc_mark_node(last->b, gc);
1537 if (should_rewrite) {
1538 ret = btree_gc_rewrite_node(b, op, last->b);
1543 if (last->b->level) {
1544 ret = btree_gc_recurse(last->b, op, writes, gc);
1549 bkey_copy_key(&b->c->gc_done, &last->b->key);
1552 * Must flush leaf nodes before gc ends, since replace
1553 * operations aren't journalled
1555 mutex_lock(&last->b->write_lock);
1556 if (btree_node_dirty(last->b))
1557 bch_btree_node_write(last->b, writes);
1558 mutex_unlock(&last->b->write_lock);
1559 rw_unlock(true, last->b);
1562 memmove(r + 1, r, sizeof(r[0]) * (GC_MERGE_NODES - 1));
1565 if (need_resched()) {
1571 for (i = r; i < r + ARRAY_SIZE(r); i++)
1572 if (!IS_ERR_OR_NULL(i->b)) {
1573 mutex_lock(&i->b->write_lock);
1574 if (btree_node_dirty(i->b))
1575 bch_btree_node_write(i->b, writes);
1576 mutex_unlock(&i->b->write_lock);
1577 rw_unlock(true, i->b);
1583 static int bch_btree_gc_root(struct btree *b, struct btree_op *op,
1584 struct closure *writes, struct gc_stat *gc)
1586 struct btree *n = NULL;
1588 bool should_rewrite;
1590 should_rewrite = btree_gc_mark_node(b, gc);
1591 if (should_rewrite) {
1592 n = btree_node_alloc_replacement(b, NULL);
1594 if (!IS_ERR_OR_NULL(n)) {
1595 bch_btree_node_write_sync(n);
1597 bch_btree_set_root(n);
1605 __bch_btree_mark_key(b->c, b->level + 1, &b->key);
1608 ret = btree_gc_recurse(b, op, writes, gc);
1613 bkey_copy_key(&b->c->gc_done, &b->key);
1618 static void btree_gc_start(struct cache_set *c)
1624 if (!c->gc_mark_valid)
1627 mutex_lock(&c->bucket_lock);
1629 c->gc_mark_valid = 0;
1630 c->gc_done = ZERO_KEY;
1632 for_each_cache(ca, c, i)
1633 for_each_bucket(b, ca) {
1635 if (!atomic_read(&b->pin)) {
1637 SET_GC_SECTORS_USED(b, 0);
1641 mutex_unlock(&c->bucket_lock);
1644 size_t bch_btree_gc_finish(struct cache_set *c)
1646 size_t available = 0;
1651 mutex_lock(&c->bucket_lock);
1654 c->gc_mark_valid = 1;
1657 for (i = 0; i < KEY_PTRS(&c->uuid_bucket); i++)
1658 SET_GC_MARK(PTR_BUCKET(c, &c->uuid_bucket, i),
1661 /* don't reclaim buckets to which writeback keys point */
1663 for (i = 0; i < c->nr_uuids; i++) {
1664 struct bcache_device *d = c->devices[i];
1665 struct cached_dev *dc;
1666 struct keybuf_key *w, *n;
1669 if (!d || UUID_FLASH_ONLY(&c->uuids[i]))
1671 dc = container_of(d, struct cached_dev, disk);
1673 spin_lock(&dc->writeback_keys.lock);
1674 rbtree_postorder_for_each_entry_safe(w, n,
1675 &dc->writeback_keys.keys, node)
1676 for (j = 0; j < KEY_PTRS(&w->key); j++)
1677 SET_GC_MARK(PTR_BUCKET(c, &w->key, j),
1679 spin_unlock(&dc->writeback_keys.lock);
1683 for_each_cache(ca, c, i) {
1686 ca->invalidate_needs_gc = 0;
1688 for (i = ca->sb.d; i < ca->sb.d + ca->sb.keys; i++)
1689 SET_GC_MARK(ca->buckets + *i, GC_MARK_METADATA);
1691 for (i = ca->prio_buckets;
1692 i < ca->prio_buckets + prio_buckets(ca) * 2; i++)
1693 SET_GC_MARK(ca->buckets + *i, GC_MARK_METADATA);
1695 for_each_bucket(b, ca) {
1696 b->last_gc = b->gc_gen;
1697 c->need_gc = max(c->need_gc, bucket_gc_gen(b));
1699 if (atomic_read(&b->pin))
1702 BUG_ON(!GC_MARK(b) && GC_SECTORS_USED(b));
1704 if (!GC_MARK(b) || GC_MARK(b) == GC_MARK_RECLAIMABLE)
1708 bch_bucket_add_unused(ca, b);
1712 mutex_unlock(&c->bucket_lock);
1716 static void bch_btree_gc(struct cache_set *c)
1719 unsigned long available;
1720 struct gc_stat stats;
1721 struct closure writes;
1723 uint64_t start_time = local_clock();
1725 trace_bcache_gc_start(c);
1727 memset(&stats, 0, sizeof(struct gc_stat));
1728 closure_init_stack(&writes);
1729 bch_btree_op_init(&op, SHRT_MAX);
1734 ret = btree_root(gc_root, c, &op, &writes, &stats);
1735 closure_sync(&writes);
1737 if (ret && ret != -EAGAIN)
1738 pr_warn("gc failed!");
1741 available = bch_btree_gc_finish(c);
1742 wake_up_allocators(c);
1744 bch_time_stats_update(&c->btree_gc_time, start_time);
1746 stats.key_bytes *= sizeof(uint64_t);
1748 stats.in_use = (c->nbuckets - available) * 100 / c->nbuckets;
1749 memcpy(&c->gc_stats, &stats, sizeof(struct gc_stat));
1751 trace_bcache_gc_end(c);
1756 static int bch_gc_thread(void *arg)
1758 struct cache_set *c = arg;
1766 set_current_state(TASK_INTERRUPTIBLE);
1767 if (kthread_should_stop())
1770 mutex_lock(&c->bucket_lock);
1772 for_each_cache(ca, c, i)
1773 if (ca->invalidate_needs_gc) {
1774 mutex_unlock(&c->bucket_lock);
1775 set_current_state(TASK_RUNNING);
1779 mutex_unlock(&c->bucket_lock);
1788 int bch_gc_thread_start(struct cache_set *c)
1790 c->gc_thread = kthread_create(bch_gc_thread, c, "bcache_gc");
1791 if (IS_ERR(c->gc_thread))
1792 return PTR_ERR(c->gc_thread);
1794 set_task_state(c->gc_thread, TASK_INTERRUPTIBLE);
1798 /* Initial partial gc */
1800 static int bch_btree_check_recurse(struct btree *b, struct btree_op *op)
1803 struct bkey *k, *p = NULL;
1804 struct btree_iter iter;
1806 for_each_key_filter(&b->keys, k, &iter, bch_ptr_invalid)
1807 bch_initial_mark_key(b->c, b->level, k);
1809 bch_initial_mark_key(b->c, b->level + 1, &b->key);
1812 bch_btree_iter_init(&b->keys, &iter, NULL);
1815 k = bch_btree_iter_next_filter(&iter, &b->keys,
1818 btree_node_prefetch(b->c, k, b->level - 1);
1821 ret = btree(check_recurse, p, b, op);
1824 } while (p && !ret);
1830 int bch_btree_check(struct cache_set *c)
1834 bch_btree_op_init(&op, SHRT_MAX);
1836 return btree_root(check_recurse, c, &op);
1839 /* Btree insertion */
1841 static bool btree_insert_key(struct btree *b, struct bkey *k,
1842 struct bkey *replace_key)
1846 BUG_ON(bkey_cmp(k, &b->key) > 0);
1848 status = bch_btree_insert_key(&b->keys, k, replace_key);
1849 if (status != BTREE_INSERT_STATUS_NO_INSERT) {
1850 bch_check_keys(&b->keys, "%u for %s", status,
1851 replace_key ? "replace" : "insert");
1853 trace_bcache_btree_insert_key(b, k, replace_key != NULL,
1860 static size_t insert_u64s_remaining(struct btree *b)
1862 long ret = bch_btree_keys_u64s_remaining(&b->keys);
1865 * Might land in the middle of an existing extent and have to split it
1867 if (b->keys.ops->is_extents)
1868 ret -= KEY_MAX_U64S;
1870 return max(ret, 0L);
1873 static bool bch_btree_insert_keys(struct btree *b, struct btree_op *op,
1874 struct keylist *insert_keys,
1875 struct bkey *replace_key)
1878 int oldsize = bch_count_data(&b->keys);
1880 while (!bch_keylist_empty(insert_keys)) {
1881 struct bkey *k = insert_keys->keys;
1883 if (bkey_u64s(k) > insert_u64s_remaining(b))
1886 if (bkey_cmp(k, &b->key) <= 0) {
1890 ret |= btree_insert_key(b, k, replace_key);
1891 bch_keylist_pop_front(insert_keys);
1892 } else if (bkey_cmp(&START_KEY(k), &b->key) < 0) {
1893 BKEY_PADDED(key) temp;
1894 bkey_copy(&temp.key, insert_keys->keys);
1896 bch_cut_back(&b->key, &temp.key);
1897 bch_cut_front(&b->key, insert_keys->keys);
1899 ret |= btree_insert_key(b, &temp.key, replace_key);
1907 op->insert_collision = true;
1909 BUG_ON(!bch_keylist_empty(insert_keys) && b->level);
1911 BUG_ON(bch_count_data(&b->keys) < oldsize);
1915 static int btree_split(struct btree *b, struct btree_op *op,
1916 struct keylist *insert_keys,
1917 struct bkey *replace_key)
1920 struct btree *n1, *n2 = NULL, *n3 = NULL;
1921 uint64_t start_time = local_clock();
1923 struct keylist parent_keys;
1925 closure_init_stack(&cl);
1926 bch_keylist_init(&parent_keys);
1928 if (btree_check_reserve(b, op)) {
1932 WARN(1, "insufficient reserve for split\n");
1935 n1 = btree_node_alloc_replacement(b, op);
1939 split = set_blocks(btree_bset_first(n1),
1940 block_bytes(n1->c)) > (btree_blocks(b) * 4) / 5;
1945 trace_bcache_btree_node_split(b, btree_bset_first(n1)->keys);
1947 n2 = bch_btree_node_alloc(b->c, op, b->level);
1952 n3 = bch_btree_node_alloc(b->c, op, b->level + 1);
1957 mutex_lock(&n1->write_lock);
1958 mutex_lock(&n2->write_lock);
1960 bch_btree_insert_keys(n1, op, insert_keys, replace_key);
1963 * Has to be a linear search because we don't have an auxiliary
1967 while (keys < (btree_bset_first(n1)->keys * 3) / 5)
1968 keys += bkey_u64s(bset_bkey_idx(btree_bset_first(n1),
1971 bkey_copy_key(&n1->key,
1972 bset_bkey_idx(btree_bset_first(n1), keys));
1973 keys += bkey_u64s(bset_bkey_idx(btree_bset_first(n1), keys));
1975 btree_bset_first(n2)->keys = btree_bset_first(n1)->keys - keys;
1976 btree_bset_first(n1)->keys = keys;
1978 memcpy(btree_bset_first(n2)->start,
1979 bset_bkey_last(btree_bset_first(n1)),
1980 btree_bset_first(n2)->keys * sizeof(uint64_t));
1982 bkey_copy_key(&n2->key, &b->key);
1984 bch_keylist_add(&parent_keys, &n2->key);
1985 bch_btree_node_write(n2, &cl);
1986 mutex_unlock(&n2->write_lock);
1987 rw_unlock(true, n2);
1989 trace_bcache_btree_node_compact(b, btree_bset_first(n1)->keys);
1991 mutex_lock(&n1->write_lock);
1992 bch_btree_insert_keys(n1, op, insert_keys, replace_key);
1995 bch_keylist_add(&parent_keys, &n1->key);
1996 bch_btree_node_write(n1, &cl);
1997 mutex_unlock(&n1->write_lock);
2000 /* Depth increases, make a new root */
2001 mutex_lock(&n3->write_lock);
2002 bkey_copy_key(&n3->key, &MAX_KEY);
2003 bch_btree_insert_keys(n3, op, &parent_keys, NULL);
2004 bch_btree_node_write(n3, &cl);
2005 mutex_unlock(&n3->write_lock);
2008 bch_btree_set_root(n3);
2009 rw_unlock(true, n3);
2010 } else if (!b->parent) {
2011 /* Root filled up but didn't need to be split */
2013 bch_btree_set_root(n1);
2015 /* Split a non root node */
2017 make_btree_freeing_key(b, parent_keys.top);
2018 bch_keylist_push(&parent_keys);
2020 bch_btree_insert_node(b->parent, op, &parent_keys, NULL, NULL);
2021 BUG_ON(!bch_keylist_empty(&parent_keys));
2025 rw_unlock(true, n1);
2027 bch_time_stats_update(&b->c->btree_split_time, start_time);
2031 bkey_put(b->c, &n2->key);
2032 btree_node_free(n2);
2033 rw_unlock(true, n2);
2035 bkey_put(b->c, &n1->key);
2036 btree_node_free(n1);
2037 rw_unlock(true, n1);
2039 WARN(1, "bcache: btree split failed (level %u)", b->level);
2041 if (n3 == ERR_PTR(-EAGAIN) ||
2042 n2 == ERR_PTR(-EAGAIN) ||
2043 n1 == ERR_PTR(-EAGAIN))
2049 static int bch_btree_insert_node(struct btree *b, struct btree_op *op,
2050 struct keylist *insert_keys,
2051 atomic_t *journal_ref,
2052 struct bkey *replace_key)
2056 BUG_ON(b->level && replace_key);
2058 closure_init_stack(&cl);
2060 mutex_lock(&b->write_lock);
2062 if (write_block(b) != btree_bset_last(b) &&
2063 b->keys.last_set_unwritten)
2064 bch_btree_init_next(b); /* just wrote a set */
2066 if (bch_keylist_nkeys(insert_keys) > insert_u64s_remaining(b)) {
2067 mutex_unlock(&b->write_lock);
2071 BUG_ON(write_block(b) != btree_bset_last(b));
2073 if (bch_btree_insert_keys(b, op, insert_keys, replace_key)) {
2075 bch_btree_leaf_dirty(b, journal_ref);
2077 bch_btree_node_write(b, &cl);
2080 mutex_unlock(&b->write_lock);
2082 /* wait for btree node write if necessary, after unlock */
2087 if (current->bio_list) {
2088 op->lock = b->c->root->level + 1;
2090 } else if (op->lock <= b->c->root->level) {
2091 op->lock = b->c->root->level + 1;
2094 /* Invalidated all iterators */
2095 int ret = btree_split(b, op, insert_keys, replace_key);
2097 if (bch_keylist_empty(insert_keys))
2105 int bch_btree_insert_check_key(struct btree *b, struct btree_op *op,
2106 struct bkey *check_key)
2109 uint64_t btree_ptr = b->key.ptr[0];
2110 unsigned long seq = b->seq;
2111 struct keylist insert;
2112 bool upgrade = op->lock == -1;
2114 bch_keylist_init(&insert);
2117 rw_unlock(false, b);
2118 rw_lock(true, b, b->level);
2120 if (b->key.ptr[0] != btree_ptr ||
2125 SET_KEY_PTRS(check_key, 1);
2126 get_random_bytes(&check_key->ptr[0], sizeof(uint64_t));
2128 SET_PTR_DEV(check_key, 0, PTR_CHECK_DEV);
2130 bch_keylist_add(&insert, check_key);
2132 ret = bch_btree_insert_node(b, op, &insert, NULL, NULL);
2134 BUG_ON(!ret && !bch_keylist_empty(&insert));
2137 downgrade_write(&b->lock);
2141 struct btree_insert_op {
2143 struct keylist *keys;
2144 atomic_t *journal_ref;
2145 struct bkey *replace_key;
2148 static int btree_insert_fn(struct btree_op *b_op, struct btree *b)
2150 struct btree_insert_op *op = container_of(b_op,
2151 struct btree_insert_op, op);
2153 int ret = bch_btree_insert_node(b, &op->op, op->keys,
2154 op->journal_ref, op->replace_key);
2155 if (ret && !bch_keylist_empty(op->keys))
2161 int bch_btree_insert(struct cache_set *c, struct keylist *keys,
2162 atomic_t *journal_ref, struct bkey *replace_key)
2164 struct btree_insert_op op;
2167 BUG_ON(current->bio_list);
2168 BUG_ON(bch_keylist_empty(keys));
2170 bch_btree_op_init(&op.op, 0);
2172 op.journal_ref = journal_ref;
2173 op.replace_key = replace_key;
2175 while (!ret && !bch_keylist_empty(keys)) {
2177 ret = bch_btree_map_leaf_nodes(&op.op, c,
2178 &START_KEY(keys->keys),
2185 pr_err("error %i", ret);
2187 while ((k = bch_keylist_pop(keys)))
2189 } else if (op.op.insert_collision)
2195 void bch_btree_set_root(struct btree *b)
2200 closure_init_stack(&cl);
2202 trace_bcache_btree_set_root(b);
2204 BUG_ON(!b->written);
2206 for (i = 0; i < KEY_PTRS(&b->key); i++)
2207 BUG_ON(PTR_BUCKET(b->c, &b->key, i)->prio != BTREE_PRIO);
2209 mutex_lock(&b->c->bucket_lock);
2210 list_del_init(&b->list);
2211 mutex_unlock(&b->c->bucket_lock);
2215 bch_journal_meta(b->c, &cl);
2219 /* Map across nodes or keys */
2221 static int bch_btree_map_nodes_recurse(struct btree *b, struct btree_op *op,
2223 btree_map_nodes_fn *fn, int flags)
2225 int ret = MAP_CONTINUE;
2229 struct btree_iter iter;
2231 bch_btree_iter_init(&b->keys, &iter, from);
2233 while ((k = bch_btree_iter_next_filter(&iter, &b->keys,
2235 ret = btree(map_nodes_recurse, k, b,
2236 op, from, fn, flags);
2239 if (ret != MAP_CONTINUE)
2244 if (!b->level || flags == MAP_ALL_NODES)
2250 int __bch_btree_map_nodes(struct btree_op *op, struct cache_set *c,
2251 struct bkey *from, btree_map_nodes_fn *fn, int flags)
2253 return btree_root(map_nodes_recurse, c, op, from, fn, flags);
2256 static int bch_btree_map_keys_recurse(struct btree *b, struct btree_op *op,
2257 struct bkey *from, btree_map_keys_fn *fn,
2260 int ret = MAP_CONTINUE;
2262 struct btree_iter iter;
2264 bch_btree_iter_init(&b->keys, &iter, from);
2266 while ((k = bch_btree_iter_next_filter(&iter, &b->keys, bch_ptr_bad))) {
2269 : btree(map_keys_recurse, k, b, op, from, fn, flags);
2272 if (ret != MAP_CONTINUE)
2276 if (!b->level && (flags & MAP_END_KEY))
2277 ret = fn(op, b, &KEY(KEY_INODE(&b->key),
2278 KEY_OFFSET(&b->key), 0));
2283 int bch_btree_map_keys(struct btree_op *op, struct cache_set *c,
2284 struct bkey *from, btree_map_keys_fn *fn, int flags)
2286 return btree_root(map_keys_recurse, c, op, from, fn, flags);
2291 static inline int keybuf_cmp(struct keybuf_key *l, struct keybuf_key *r)
2293 /* Overlapping keys compare equal */
2294 if (bkey_cmp(&l->key, &START_KEY(&r->key)) <= 0)
2296 if (bkey_cmp(&START_KEY(&l->key), &r->key) >= 0)
2301 static inline int keybuf_nonoverlapping_cmp(struct keybuf_key *l,
2302 struct keybuf_key *r)
2304 return clamp_t(int64_t, bkey_cmp(&l->key, &r->key), -1, 1);
2312 keybuf_pred_fn *pred;
2315 static int refill_keybuf_fn(struct btree_op *op, struct btree *b,
2318 struct refill *refill = container_of(op, struct refill, op);
2319 struct keybuf *buf = refill->buf;
2320 int ret = MAP_CONTINUE;
2322 if (bkey_cmp(k, refill->end) >= 0) {
2327 if (!KEY_SIZE(k)) /* end key */
2330 if (refill->pred(buf, k)) {
2331 struct keybuf_key *w;
2333 spin_lock(&buf->lock);
2335 w = array_alloc(&buf->freelist);
2337 spin_unlock(&buf->lock);
2342 bkey_copy(&w->key, k);
2344 if (RB_INSERT(&buf->keys, w, node, keybuf_cmp))
2345 array_free(&buf->freelist, w);
2349 if (array_freelist_empty(&buf->freelist))
2352 spin_unlock(&buf->lock);
2355 buf->last_scanned = *k;
2359 void bch_refill_keybuf(struct cache_set *c, struct keybuf *buf,
2360 struct bkey *end, keybuf_pred_fn *pred)
2362 struct bkey start = buf->last_scanned;
2363 struct refill refill;
2367 bch_btree_op_init(&refill.op, -1);
2368 refill.nr_found = 0;
2373 bch_btree_map_keys(&refill.op, c, &buf->last_scanned,
2374 refill_keybuf_fn, MAP_END_KEY);
2376 trace_bcache_keyscan(refill.nr_found,
2377 KEY_INODE(&start), KEY_OFFSET(&start),
2378 KEY_INODE(&buf->last_scanned),
2379 KEY_OFFSET(&buf->last_scanned));
2381 spin_lock(&buf->lock);
2383 if (!RB_EMPTY_ROOT(&buf->keys)) {
2384 struct keybuf_key *w;
2385 w = RB_FIRST(&buf->keys, struct keybuf_key, node);
2386 buf->start = START_KEY(&w->key);
2388 w = RB_LAST(&buf->keys, struct keybuf_key, node);
2391 buf->start = MAX_KEY;
2395 spin_unlock(&buf->lock);
2398 static void __bch_keybuf_del(struct keybuf *buf, struct keybuf_key *w)
2400 rb_erase(&w->node, &buf->keys);
2401 array_free(&buf->freelist, w);
2404 void bch_keybuf_del(struct keybuf *buf, struct keybuf_key *w)
2406 spin_lock(&buf->lock);
2407 __bch_keybuf_del(buf, w);
2408 spin_unlock(&buf->lock);
2411 bool bch_keybuf_check_overlapping(struct keybuf *buf, struct bkey *start,
2415 struct keybuf_key *p, *w, s;
2418 if (bkey_cmp(end, &buf->start) <= 0 ||
2419 bkey_cmp(start, &buf->end) >= 0)
2422 spin_lock(&buf->lock);
2423 w = RB_GREATER(&buf->keys, s, node, keybuf_nonoverlapping_cmp);
2425 while (w && bkey_cmp(&START_KEY(&w->key), end) < 0) {
2427 w = RB_NEXT(w, node);
2432 __bch_keybuf_del(buf, p);
2435 spin_unlock(&buf->lock);
2439 struct keybuf_key *bch_keybuf_next(struct keybuf *buf)
2441 struct keybuf_key *w;
2442 spin_lock(&buf->lock);
2444 w = RB_FIRST(&buf->keys, struct keybuf_key, node);
2446 while (w && w->private)
2447 w = RB_NEXT(w, node);
2450 w->private = ERR_PTR(-EINTR);
2452 spin_unlock(&buf->lock);
2456 struct keybuf_key *bch_keybuf_next_rescan(struct cache_set *c,
2459 keybuf_pred_fn *pred)
2461 struct keybuf_key *ret;
2464 ret = bch_keybuf_next(buf);
2468 if (bkey_cmp(&buf->last_scanned, end) >= 0) {
2469 pr_debug("scan finished");
2473 bch_refill_keybuf(c, buf, end, pred);
2479 void bch_keybuf_init(struct keybuf *buf)
2481 buf->last_scanned = MAX_KEY;
2482 buf->keys = RB_ROOT;
2484 spin_lock_init(&buf->lock);
2485 array_allocator_init(&buf->freelist);