3 #include "btree_cache.h"
5 #include "btree_iter.h"
6 #include "btree_locking.h"
10 #include <linux/prefetch.h>
11 #include <trace/events/bcachefs.h>
13 #define DEF_BTREE_ID(kwd, val, name) name,
15 const char * const bch2_btree_ids[] = {
16 DEFINE_BCH_BTREE_IDS()
22 void bch2_recalc_btree_reserve(struct bch_fs *c)
24 unsigned i, reserve = 16;
26 if (!c->btree_roots[0].b)
29 for (i = 0; i < BTREE_ID_NR; i++)
30 if (c->btree_roots[i].b)
31 reserve += min_t(unsigned, 1,
32 c->btree_roots[i].b->level) * 8;
34 c->btree_cache_reserve = reserve;
37 #define mca_can_free(c) \
38 max_t(int, 0, c->btree_cache_used - c->btree_cache_reserve)
40 static void __mca_data_free(struct bch_fs *c, struct btree *b)
42 EBUG_ON(btree_node_write_in_flight(b));
44 free_pages((unsigned long) b->data, btree_page_order(c));
46 bch2_btree_keys_free(b);
49 static void mca_data_free(struct bch_fs *c, struct btree *b)
51 __mca_data_free(c, b);
52 c->btree_cache_used--;
53 list_move(&b->list, &c->btree_cache_freed);
56 #define PTR_HASH(_k) (bkey_i_to_extent_c(_k)->v._data[0])
58 static const struct rhashtable_params bch_btree_cache_params = {
59 .head_offset = offsetof(struct btree, hash),
60 .key_offset = offsetof(struct btree, key.v),
61 .key_len = sizeof(struct bch_extent_ptr),
64 static void mca_data_alloc(struct bch_fs *c, struct btree *b, gfp_t gfp)
66 unsigned order = ilog2(btree_pages(c));
68 b->data = (void *) __get_free_pages(gfp, order);
72 if (bch2_btree_keys_alloc(b, order, gfp))
75 c->btree_cache_used++;
76 list_move(&b->list, &c->btree_cache_freeable);
79 free_pages((unsigned long) b->data, order);
81 list_move(&b->list, &c->btree_cache_freed);
84 static struct btree *mca_bucket_alloc(struct bch_fs *c, gfp_t gfp)
86 struct btree *b = kzalloc(sizeof(struct btree), gfp);
90 six_lock_init(&b->lock);
91 INIT_LIST_HEAD(&b->list);
92 INIT_LIST_HEAD(&b->write_blocked);
94 mca_data_alloc(c, b, gfp);
95 return b->data ? b : NULL;
98 /* Btree in memory cache - hash table */
100 void bch2_btree_node_hash_remove(struct bch_fs *c, struct btree *b)
102 BUG_ON(btree_node_dirty(b));
106 rhashtable_remove_fast(&c->btree_cache_table, &b->hash,
107 bch_btree_cache_params);
109 /* Cause future lookups for this node to fail: */
110 bkey_i_to_extent(&b->key)->v._data[0] = 0;
113 int bch2_btree_node_hash_insert(struct bch_fs *c, struct btree *b,
114 unsigned level, enum btree_id id)
120 ret = rhashtable_lookup_insert_fast(&c->btree_cache_table, &b->hash,
121 bch_btree_cache_params);
125 mutex_lock(&c->btree_cache_lock);
126 list_add(&b->list, &c->btree_cache);
127 mutex_unlock(&c->btree_cache_lock);
133 static inline struct btree *mca_find(struct bch_fs *c,
134 const struct bkey_i *k)
136 return rhashtable_lookup_fast(&c->btree_cache_table, &PTR_HASH(k),
137 bch_btree_cache_params);
141 * this version is for btree nodes that have already been freed (we're not
142 * reaping a real btree node)
144 static int mca_reap_notrace(struct bch_fs *c, struct btree *b, bool flush)
146 lockdep_assert_held(&c->btree_cache_lock);
148 if (!six_trylock_intent(&b->lock))
151 if (!six_trylock_write(&b->lock))
152 goto out_unlock_intent;
154 if (btree_node_write_error(b) ||
155 btree_node_noevict(b))
158 if (!list_empty(&b->write_blocked))
162 (btree_node_dirty(b) ||
163 btree_node_write_in_flight(b)))
167 * Using the underscore version because we don't want to compact bsets
168 * after the write, since this node is about to be evicted - unless
169 * btree verify mode is enabled, since it runs out of the post write
172 if (btree_node_dirty(b)) {
173 if (verify_btree_ondisk(c))
174 bch2_btree_node_write(c, b, NULL, SIX_LOCK_intent, -1);
176 __bch2_btree_node_write(c, b, NULL, SIX_LOCK_read, -1);
179 /* wait for any in flight btree write */
180 wait_on_bit_io(&b->flags, BTREE_NODE_write_in_flight,
181 TASK_UNINTERRUPTIBLE);
185 six_unlock_write(&b->lock);
187 six_unlock_intent(&b->lock);
191 static int mca_reap(struct bch_fs *c, struct btree *b, bool flush)
193 int ret = mca_reap_notrace(c, b, flush);
195 trace_btree_node_reap(c, b, ret);
199 static unsigned long bch2_mca_scan(struct shrinker *shrink,
200 struct shrink_control *sc)
202 struct bch_fs *c = container_of(shrink, struct bch_fs,
205 unsigned long nr = sc->nr_to_scan;
206 unsigned long can_free;
207 unsigned long touched = 0;
208 unsigned long freed = 0;
211 if (btree_shrinker_disabled(c))
214 if (c->btree_cache_alloc_lock)
217 /* Return -1 if we can't do anything right now */
218 if (sc->gfp_mask & __GFP_IO)
219 mutex_lock(&c->btree_cache_lock);
220 else if (!mutex_trylock(&c->btree_cache_lock))
224 * It's _really_ critical that we don't free too many btree nodes - we
225 * have to always leave ourselves a reserve. The reserve is how we
226 * guarantee that allocating memory for a new btree node can always
227 * succeed, so that inserting keys into the btree can always succeed and
228 * IO can always make forward progress:
230 nr /= btree_pages(c);
231 can_free = mca_can_free(c);
232 nr = min_t(unsigned long, nr, can_free);
235 list_for_each_entry_safe(b, t, &c->btree_cache_freeable, list) {
242 !mca_reap_notrace(c, b, false)) {
244 six_unlock_write(&b->lock);
245 six_unlock_intent(&b->lock);
250 list_for_each_entry_safe(b, t, &c->btree_cache, list) {
255 if (&t->list != &c->btree_cache)
256 list_move_tail(&c->btree_cache, &t->list);
260 if (!btree_node_accessed(b) &&
261 !mca_reap(c, b, false)) {
262 /* can't call bch2_btree_node_hash_remove under btree_cache_lock */
264 if (&t->list != &c->btree_cache)
265 list_move_tail(&c->btree_cache, &t->list);
268 mutex_unlock(&c->btree_cache_lock);
270 bch2_btree_node_hash_remove(c, b);
271 six_unlock_write(&b->lock);
272 six_unlock_intent(&b->lock);
277 if (sc->gfp_mask & __GFP_IO)
278 mutex_lock(&c->btree_cache_lock);
279 else if (!mutex_trylock(&c->btree_cache_lock))
283 clear_btree_node_accessed(b);
286 mutex_unlock(&c->btree_cache_lock);
288 return (unsigned long) freed * btree_pages(c);
291 static unsigned long bch2_mca_count(struct shrinker *shrink,
292 struct shrink_control *sc)
294 struct bch_fs *c = container_of(shrink, struct bch_fs,
297 if (btree_shrinker_disabled(c))
300 if (c->btree_cache_alloc_lock)
303 return mca_can_free(c) * btree_pages(c);
306 void bch2_fs_btree_exit(struct bch_fs *c)
311 if (c->btree_cache_shrink.list.next)
312 unregister_shrinker(&c->btree_cache_shrink);
314 mutex_lock(&c->btree_cache_lock);
316 #ifdef CONFIG_BCACHEFS_DEBUG
318 list_move(&c->verify_data->list, &c->btree_cache);
320 free_pages((unsigned long) c->verify_ondisk, ilog2(btree_pages(c)));
323 for (i = 0; i < BTREE_ID_NR; i++)
324 if (c->btree_roots[i].b)
325 list_add(&c->btree_roots[i].b->list, &c->btree_cache);
327 list_splice(&c->btree_cache_freeable,
330 while (!list_empty(&c->btree_cache)) {
331 b = list_first_entry(&c->btree_cache, struct btree, list);
333 if (btree_node_dirty(b))
334 bch2_btree_complete_write(c, b, btree_current_write(b));
335 clear_btree_node_dirty(b);
340 while (!list_empty(&c->btree_cache_freed)) {
341 b = list_first_entry(&c->btree_cache_freed,
347 mutex_unlock(&c->btree_cache_lock);
349 if (c->btree_cache_table_init_done)
350 rhashtable_destroy(&c->btree_cache_table);
353 int bch2_fs_btree_init(struct bch_fs *c)
358 ret = rhashtable_init(&c->btree_cache_table, &bch_btree_cache_params);
362 c->btree_cache_table_init_done = true;
364 bch2_recalc_btree_reserve(c);
366 for (i = 0; i < c->btree_cache_reserve; i++)
367 if (!mca_bucket_alloc(c, GFP_KERNEL))
370 list_splice_init(&c->btree_cache,
371 &c->btree_cache_freeable);
373 #ifdef CONFIG_BCACHEFS_DEBUG
374 mutex_init(&c->verify_lock);
376 c->verify_ondisk = (void *)
377 __get_free_pages(GFP_KERNEL, ilog2(btree_pages(c)));
378 if (!c->verify_ondisk)
381 c->verify_data = mca_bucket_alloc(c, GFP_KERNEL);
385 list_del_init(&c->verify_data->list);
388 c->btree_cache_shrink.count_objects = bch2_mca_count;
389 c->btree_cache_shrink.scan_objects = bch2_mca_scan;
390 c->btree_cache_shrink.seeks = 4;
391 c->btree_cache_shrink.batch = btree_pages(c) * 2;
392 register_shrinker(&c->btree_cache_shrink);
398 * We can only have one thread cannibalizing other cached btree nodes at a time,
399 * or we'll deadlock. We use an open coded mutex to ensure that, which a
400 * cannibalize_bucket() will take. This means every time we unlock the root of
401 * the btree, we need to release this lock if we have it held.
403 void bch2_btree_node_cannibalize_unlock(struct bch_fs *c)
405 if (c->btree_cache_alloc_lock == current) {
406 trace_btree_node_cannibalize_unlock(c);
407 c->btree_cache_alloc_lock = NULL;
408 closure_wake_up(&c->mca_wait);
412 int bch2_btree_node_cannibalize_lock(struct bch_fs *c, struct closure *cl)
414 struct task_struct *old;
416 old = cmpxchg(&c->btree_cache_alloc_lock, NULL, current);
417 if (old == NULL || old == current)
421 trace_btree_node_cannibalize_lock_fail(c);
425 closure_wait(&c->mca_wait, cl);
427 /* Try again, after adding ourselves to waitlist */
428 old = cmpxchg(&c->btree_cache_alloc_lock, NULL, current);
429 if (old == NULL || old == current) {
431 closure_wake_up(&c->mca_wait);
435 trace_btree_node_cannibalize_lock_fail(c);
439 trace_btree_node_cannibalize_lock(c);
443 static struct btree *mca_cannibalize(struct bch_fs *c)
447 list_for_each_entry_reverse(b, &c->btree_cache, list)
448 if (!mca_reap(c, b, false))
452 list_for_each_entry_reverse(b, &c->btree_cache, list)
453 if (!mca_reap(c, b, true))
457 * Rare case: all nodes were intent-locked.
460 WARN_ONCE(1, "btree cache cannibalize failed\n");
465 struct btree *bch2_btree_node_mem_alloc(struct bch_fs *c)
468 u64 start_time = local_clock();
470 mutex_lock(&c->btree_cache_lock);
473 * btree_free() doesn't free memory; it sticks the node on the end of
474 * the list. Check if there's any freed nodes there:
476 list_for_each_entry(b, &c->btree_cache_freeable, list)
477 if (!mca_reap_notrace(c, b, false))
481 * We never free struct btree itself, just the memory that holds the on
482 * disk node. Check the freed list before allocating a new one:
484 list_for_each_entry(b, &c->btree_cache_freed, list)
485 if (!mca_reap_notrace(c, b, false)) {
486 mca_data_alloc(c, b, __GFP_NOWARN|GFP_NOIO);
490 six_unlock_write(&b->lock);
491 six_unlock_intent(&b->lock);
495 b = mca_bucket_alloc(c, __GFP_NOWARN|GFP_NOIO);
499 BUG_ON(!six_trylock_intent(&b->lock));
500 BUG_ON(!six_trylock_write(&b->lock));
502 BUG_ON(bkey_extent_is_data(&b->key.k) && PTR_HASH(&b->key));
503 BUG_ON(btree_node_write_in_flight(b));
505 list_del_init(&b->list);
506 mutex_unlock(&c->btree_cache_lock);
513 b->whiteout_u64s = 0;
514 b->uncompacted_whiteout_u64s = 0;
515 bch2_btree_keys_init(b, &c->expensive_debug_checks);
517 bch2_time_stats_update(&c->btree_node_mem_alloc_time, start_time);
521 /* Try to cannibalize another cached btree node: */
522 if (c->btree_cache_alloc_lock == current) {
523 b = mca_cannibalize(c);
524 list_del_init(&b->list);
525 mutex_unlock(&c->btree_cache_lock);
527 bch2_btree_node_hash_remove(c, b);
529 trace_btree_node_cannibalize(c);
533 mutex_unlock(&c->btree_cache_lock);
534 return ERR_PTR(-ENOMEM);
537 /* Slowpath, don't want it inlined into btree_iter_traverse() */
538 static noinline struct btree *bch2_btree_node_fill(struct btree_iter *iter,
539 const struct bkey_i *k,
541 enum six_lock_type lock_type)
543 struct bch_fs *c = iter->c;
546 b = bch2_btree_node_mem_alloc(c);
550 bkey_copy(&b->key, k);
551 if (bch2_btree_node_hash_insert(c, b, level, iter->btree_id)) {
552 /* raced with another fill: */
554 /* mark as unhashed... */
555 bkey_i_to_extent(&b->key)->v._data[0] = 0;
557 mutex_lock(&c->btree_cache_lock);
558 list_add(&b->list, &c->btree_cache_freeable);
559 mutex_unlock(&c->btree_cache_lock);
561 six_unlock_write(&b->lock);
562 six_unlock_intent(&b->lock);
567 * If the btree node wasn't cached, we can't drop our lock on
568 * the parent until after it's added to the cache - because
569 * otherwise we could race with a btree_split() freeing the node
570 * we're trying to lock.
572 * But the deadlock described below doesn't exist in this case,
573 * so it's safe to not drop the parent lock until here:
575 if (btree_node_read_locked(iter, level + 1))
576 btree_node_unlock(iter, level + 1);
578 bch2_btree_node_read(c, b);
579 six_unlock_write(&b->lock);
581 if (lock_type == SIX_LOCK_read)
582 six_lock_downgrade(&b->lock);
588 * bch_btree_node_get - find a btree node in the cache and lock it, reading it
589 * in from disk if necessary.
591 * If IO is necessary and running under generic_make_request, returns -EAGAIN.
593 * The btree node will have either a read or a write lock held, depending on
594 * the @write parameter.
596 struct btree *bch2_btree_node_get(struct btree_iter *iter,
597 const struct bkey_i *k, unsigned level,
598 enum six_lock_type lock_type)
603 BUG_ON(level >= BTREE_MAX_DEPTH);
606 b = mca_find(iter->c, k);
611 * We must have the parent locked to call bch2_btree_node_fill(),
612 * else we could read in a btree node from disk that's been
615 b = bch2_btree_node_fill(iter, k, level, lock_type);
617 /* We raced and found the btree node in the cache */
625 * There's a potential deadlock with splits and insertions into
626 * interior nodes we have to avoid:
628 * The other thread might be holding an intent lock on the node
629 * we want, and they want to update its parent node so they're
630 * going to upgrade their intent lock on the parent node to a
633 * But if we're holding a read lock on the parent, and we're
634 * trying to get the intent lock they're holding, we deadlock.
636 * So to avoid this we drop the read locks on parent nodes when
637 * we're starting to take intent locks - and handle the race.
639 * The race is that they might be about to free the node we
640 * want, and dropping our read lock on the parent node lets them
641 * update the parent marking the node we want as freed, and then
644 * To guard against this, btree nodes are evicted from the cache
645 * when they're freed - and PTR_HASH() is zeroed out, which we
646 * check for after we lock the node.
648 * Then, bch2_btree_node_relock() on the parent will fail - because
649 * the parent was modified, when the pointer to the node we want
650 * was removed - and we'll bail out:
652 if (btree_node_read_locked(iter, level + 1))
653 btree_node_unlock(iter, level + 1);
655 if (!btree_node_lock(b, k->k.p, level, iter, lock_type))
656 return ERR_PTR(-EINTR);
658 if (unlikely(PTR_HASH(&b->key) != PTR_HASH(k) ||
661 six_unlock_type(&b->lock, lock_type);
662 if (bch2_btree_node_relock(iter, level + 1))
665 return ERR_PTR(-EINTR);
669 prefetch(b->aux_data);
671 for_each_bset(b, t) {
672 void *p = (u64 *) b->aux_data + t->aux_data_offset;
674 prefetch(p + L1_CACHE_BYTES * 0);
675 prefetch(p + L1_CACHE_BYTES * 1);
676 prefetch(p + L1_CACHE_BYTES * 2);
679 /* avoid atomic set bit if it's not needed: */
680 if (btree_node_accessed(b))
681 set_btree_node_accessed(b);
683 if (unlikely(btree_node_read_error(b))) {
684 six_unlock_type(&b->lock, lock_type);
685 return ERR_PTR(-EIO);
688 EBUG_ON(!b->written);
689 EBUG_ON(b->btree_id != iter->btree_id ||
690 BTREE_NODE_LEVEL(b->data) != level ||
691 bkey_cmp(b->data->max_key, k->k.p));
696 int bch2_print_btree_node(struct bch_fs *c, struct btree *b,
697 char *buf, size_t len)
699 const struct bkey_format *f = &b->format;
700 struct bset_stats stats;
703 memset(&stats, 0, sizeof(stats));
705 bch2_val_to_text(c, BKEY_TYPE_BTREE, ptrs, sizeof(ptrs),
706 bkey_i_to_s_c(&b->key));
707 bch2_btree_keys_stats(b, &stats);
709 return scnprintf(buf, len,
710 "l %u %llu:%llu - %llu:%llu:\n"
712 " format: u64s %u fields %u %u %u %u %u\n"
713 " unpack fn len: %u\n"
714 " bytes used %zu/%zu (%zu%% full)\n"
715 " sib u64s: %u, %u (merge threshold %zu)\n"
716 " nr packed keys %u\n"
717 " nr unpacked keys %u\n"
719 " failed unpacked %zu\n"
721 " failed overflow %zu\n",
723 b->data->min_key.inode,
724 b->data->min_key.offset,
725 b->data->max_key.inode,
726 b->data->max_key.offset,
729 f->bits_per_field[0],
730 f->bits_per_field[1],
731 f->bits_per_field[2],
732 f->bits_per_field[3],
733 f->bits_per_field[4],
735 b->nr.live_u64s * sizeof(u64),
736 btree_bytes(c) - sizeof(struct btree_node),
737 b->nr.live_u64s * 100 / btree_max_u64s(c),
740 BTREE_FOREGROUND_MERGE_THRESHOLD(c),
744 stats.failed_unpacked,
746 stats.failed_overflow);