1 // SPDX-License-Identifier: GPL-2.0
5 #include "btree_cache.h"
7 #include "btree_iter.h"
8 #include "btree_locking.h"
12 #include <linux/prefetch.h>
13 #include <linux/sched/mm.h>
14 #include <trace/events/bcachefs.h>
16 struct lock_class_key bch2_btree_node_lock_key;
18 const char * const bch2_btree_node_flags[] = {
25 void bch2_recalc_btree_reserve(struct bch_fs *c)
27 unsigned i, reserve = 16;
29 if (!c->btree_roots[0].b)
32 for (i = 0; i < BTREE_ID_NR; i++)
33 if (c->btree_roots[i].b)
34 reserve += min_t(unsigned, 1,
35 c->btree_roots[i].b->c.level) * 8;
37 c->btree_cache.reserve = reserve;
40 static inline unsigned btree_cache_can_free(struct btree_cache *bc)
42 return max_t(int, 0, bc->used - bc->reserve);
45 static void btree_node_to_freedlist(struct btree_cache *bc, struct btree *b)
47 if (b->c.lock.readers)
48 list_move(&b->list, &bc->freed_pcpu);
50 list_move(&b->list, &bc->freed_nonpcpu);
53 static void btree_node_data_free(struct bch_fs *c, struct btree *b)
55 struct btree_cache *bc = &c->btree_cache;
57 EBUG_ON(btree_node_write_in_flight(b));
59 kvpfree(b->data, btree_bytes(c));
64 munmap(b->aux_data, btree_aux_data_bytes(b));
70 btree_node_to_freedlist(bc, b);
73 static int bch2_btree_cache_cmp_fn(struct rhashtable_compare_arg *arg,
76 const struct btree *b = obj;
77 const u64 *v = arg->key;
79 return b->hash_val == *v ? 0 : 1;
82 static const struct rhashtable_params bch_btree_cache_params = {
83 .head_offset = offsetof(struct btree, hash),
84 .key_offset = offsetof(struct btree, hash_val),
85 .key_len = sizeof(u64),
86 .obj_cmpfn = bch2_btree_cache_cmp_fn,
89 static int btree_node_data_alloc(struct bch_fs *c, struct btree *b, gfp_t gfp)
91 BUG_ON(b->data || b->aux_data);
93 b->data = kvpmalloc(btree_bytes(c), gfp);
97 b->aux_data = vmalloc_exec(btree_aux_data_bytes(b), gfp);
99 b->aux_data = mmap(NULL, btree_aux_data_bytes(b),
100 PROT_READ|PROT_WRITE|PROT_EXEC,
101 MAP_PRIVATE|MAP_ANONYMOUS, 0, 0);
102 if (b->aux_data == MAP_FAILED)
106 kvpfree(b->data, btree_bytes(c));
114 static struct btree *__btree_node_mem_alloc(struct bch_fs *c)
116 struct btree *b = kzalloc(sizeof(struct btree), GFP_KERNEL);
120 bkey_btree_ptr_init(&b->key);
121 __six_lock_init(&b->c.lock, "b->c.lock", &bch2_btree_node_lock_key);
122 INIT_LIST_HEAD(&b->list);
123 INIT_LIST_HEAD(&b->write_blocked);
124 b->byte_order = ilog2(btree_bytes(c));
128 struct btree *__bch2_btree_node_mem_alloc(struct bch_fs *c)
130 struct btree_cache *bc = &c->btree_cache;
131 struct btree *b = __btree_node_mem_alloc(c);
135 if (btree_node_data_alloc(c, b, GFP_KERNEL)) {
141 list_add(&b->list, &bc->freeable);
145 /* Btree in memory cache - hash table */
147 void bch2_btree_node_hash_remove(struct btree_cache *bc, struct btree *b)
149 int ret = rhashtable_remove_fast(&bc->table, &b->hash, bch_btree_cache_params);
152 /* Cause future lookups for this node to fail: */
155 six_lock_wakeup_all(&b->c.lock);
158 int __bch2_btree_node_hash_insert(struct btree_cache *bc, struct btree *b)
161 b->hash_val = btree_ptr_hash_val(&b->key);
163 return rhashtable_lookup_insert_fast(&bc->table, &b->hash,
164 bch_btree_cache_params);
167 int bch2_btree_node_hash_insert(struct btree_cache *bc, struct btree *b,
168 unsigned level, enum btree_id id)
175 mutex_lock(&bc->lock);
176 ret = __bch2_btree_node_hash_insert(bc, b);
178 list_add(&b->list, &bc->live);
179 mutex_unlock(&bc->lock);
185 static inline struct btree *btree_cache_find(struct btree_cache *bc,
186 const struct bkey_i *k)
188 u64 v = btree_ptr_hash_val(k);
190 return rhashtable_lookup_fast(&bc->table, &v, bch_btree_cache_params);
194 * this version is for btree nodes that have already been freed (we're not
195 * reaping a real btree node)
197 static int __btree_node_reclaim(struct bch_fs *c, struct btree *b, bool flush)
199 struct btree_cache *bc = &c->btree_cache;
202 lockdep_assert_held(&bc->lock);
204 if (b->flags & ((1U << BTREE_NODE_dirty)|
205 (1U << BTREE_NODE_read_in_flight)|
206 (1U << BTREE_NODE_write_in_flight))) {
210 /* XXX: waiting on IO with btree cache lock held */
211 bch2_btree_node_wait_on_read(b);
212 bch2_btree_node_wait_on_write(b);
215 if (!six_trylock_intent(&b->c.lock))
218 if (!six_trylock_write(&b->c.lock))
219 goto out_unlock_intent;
221 /* recheck under lock */
222 if (b->flags & ((1U << BTREE_NODE_read_in_flight)|
223 (1U << BTREE_NODE_write_in_flight))) {
226 six_unlock_write(&b->c.lock);
227 six_unlock_intent(&b->c.lock);
231 if (btree_node_noevict(b) ||
232 btree_node_write_blocked(b) ||
233 btree_node_will_make_reachable(b))
236 if (btree_node_dirty(b)) {
240 * Using the underscore version because we don't want to compact
241 * bsets after the write, since this node is about to be evicted
242 * - unless btree verify mode is enabled, since it runs out of
243 * the post write cleanup:
245 if (bch2_verify_btree_ondisk)
246 bch2_btree_node_write(c, b, SIX_LOCK_intent, 0);
248 __bch2_btree_node_write(c, b, 0);
250 six_unlock_write(&b->c.lock);
251 six_unlock_intent(&b->c.lock);
255 if (b->hash_val && !ret)
256 trace_btree_node_reap(c, b);
259 six_unlock_write(&b->c.lock);
261 six_unlock_intent(&b->c.lock);
266 static int btree_node_reclaim(struct bch_fs *c, struct btree *b)
268 return __btree_node_reclaim(c, b, false);
271 static int btree_node_write_and_reclaim(struct bch_fs *c, struct btree *b)
273 return __btree_node_reclaim(c, b, true);
276 static unsigned long bch2_btree_cache_scan(struct shrinker *shrink,
277 struct shrink_control *sc)
279 struct bch_fs *c = container_of(shrink, struct bch_fs,
281 struct btree_cache *bc = &c->btree_cache;
283 unsigned long nr = sc->nr_to_scan;
284 unsigned long can_free;
285 unsigned long touched = 0;
286 unsigned long freed = 0;
288 unsigned long ret = SHRINK_STOP;
290 if (bch2_btree_shrinker_disabled)
293 /* Return -1 if we can't do anything right now */
294 if (sc->gfp_mask & __GFP_FS)
295 mutex_lock(&bc->lock);
296 else if (!mutex_trylock(&bc->lock))
299 flags = memalloc_nofs_save();
302 * It's _really_ critical that we don't free too many btree nodes - we
303 * have to always leave ourselves a reserve. The reserve is how we
304 * guarantee that allocating memory for a new btree node can always
305 * succeed, so that inserting keys into the btree can always succeed and
306 * IO can always make forward progress:
308 nr /= btree_pages(c);
309 can_free = btree_cache_can_free(bc);
310 nr = min_t(unsigned long, nr, can_free);
313 list_for_each_entry_safe(b, t, &bc->freeable, list) {
315 * Leave a few nodes on the freeable list, so that a btree split
316 * won't have to hit the system allocator:
326 if (!btree_node_reclaim(c, b)) {
327 btree_node_data_free(c, b);
328 six_unlock_write(&b->c.lock);
329 six_unlock_intent(&b->c.lock);
334 list_for_each_entry_safe(b, t, &bc->live, list) {
336 if (btree_node_accessed(b)) {
337 clear_btree_node_accessed(b);
341 if (!btree_node_reclaim(c, b)) {
342 /* can't call bch2_btree_node_hash_remove under lock */
344 if (&t->list != &bc->live)
345 list_move_tail(&bc->live, &t->list);
347 btree_node_data_free(c, b);
348 mutex_unlock(&bc->lock);
350 bch2_btree_node_hash_remove(bc, b);
351 six_unlock_write(&b->c.lock);
352 six_unlock_intent(&b->c.lock);
357 if (sc->gfp_mask & __GFP_FS)
358 mutex_lock(&bc->lock);
359 else if (!mutex_trylock(&bc->lock))
370 if (&t->list != &bc->live)
371 list_move_tail(&bc->live, &t->list);
376 mutex_unlock(&bc->lock);
378 ret = (unsigned long) freed * btree_pages(c);
379 memalloc_nofs_restore(flags);
381 trace_btree_cache_scan(sc->nr_to_scan,
382 sc->nr_to_scan / btree_pages(c),
383 btree_cache_can_free(bc),
388 static unsigned long bch2_btree_cache_count(struct shrinker *shrink,
389 struct shrink_control *sc)
391 struct bch_fs *c = container_of(shrink, struct bch_fs,
393 struct btree_cache *bc = &c->btree_cache;
395 if (bch2_btree_shrinker_disabled)
398 return btree_cache_can_free(bc) * btree_pages(c);
401 void bch2_fs_btree_cache_exit(struct bch_fs *c)
403 struct btree_cache *bc = &c->btree_cache;
407 if (bc->shrink.list.next)
408 unregister_shrinker(&bc->shrink);
410 /* vfree() can allocate memory: */
411 flags = memalloc_nofs_save();
412 mutex_lock(&bc->lock);
415 list_move(&c->verify_data->list, &bc->live);
417 kvpfree(c->verify_ondisk, btree_bytes(c));
419 for (i = 0; i < BTREE_ID_NR; i++)
420 if (c->btree_roots[i].b)
421 list_add(&c->btree_roots[i].b->list, &bc->live);
423 list_splice(&bc->freeable, &bc->live);
425 while (!list_empty(&bc->live)) {
426 b = list_first_entry(&bc->live, struct btree, list);
428 BUG_ON(btree_node_read_in_flight(b) ||
429 btree_node_write_in_flight(b));
431 if (btree_node_dirty(b))
432 bch2_btree_complete_write(c, b, btree_current_write(b));
433 clear_btree_node_dirty_acct(c, b);
435 btree_node_data_free(c, b);
438 BUG_ON(atomic_read(&c->btree_cache.dirty));
440 list_splice(&bc->freed_pcpu, &bc->freed_nonpcpu);
442 while (!list_empty(&bc->freed_nonpcpu)) {
443 b = list_first_entry(&bc->freed_nonpcpu, struct btree, list);
445 six_lock_pcpu_free(&b->c.lock);
449 mutex_unlock(&bc->lock);
450 memalloc_nofs_restore(flags);
452 if (bc->table_init_done)
453 rhashtable_destroy(&bc->table);
456 int bch2_fs_btree_cache_init(struct bch_fs *c)
458 struct btree_cache *bc = &c->btree_cache;
462 pr_verbose_init(c->opts, "");
464 ret = rhashtable_init(&bc->table, &bch_btree_cache_params);
468 bc->table_init_done = true;
470 bch2_recalc_btree_reserve(c);
472 for (i = 0; i < bc->reserve; i++)
473 if (!__bch2_btree_node_mem_alloc(c)) {
478 list_splice_init(&bc->live, &bc->freeable);
480 mutex_init(&c->verify_lock);
482 bc->shrink.count_objects = bch2_btree_cache_count;
483 bc->shrink.scan_objects = bch2_btree_cache_scan;
484 bc->shrink.seeks = 4;
485 bc->shrink.batch = btree_pages(c) * 2;
486 ret = register_shrinker(&bc->shrink);
488 pr_verbose_init(c->opts, "ret %i", ret);
492 void bch2_fs_btree_cache_init_early(struct btree_cache *bc)
494 mutex_init(&bc->lock);
495 INIT_LIST_HEAD(&bc->live);
496 INIT_LIST_HEAD(&bc->freeable);
497 INIT_LIST_HEAD(&bc->freed_pcpu);
498 INIT_LIST_HEAD(&bc->freed_nonpcpu);
502 * We can only have one thread cannibalizing other cached btree nodes at a time,
503 * or we'll deadlock. We use an open coded mutex to ensure that, which a
504 * cannibalize_bucket() will take. This means every time we unlock the root of
505 * the btree, we need to release this lock if we have it held.
507 void bch2_btree_cache_cannibalize_unlock(struct bch_fs *c)
509 struct btree_cache *bc = &c->btree_cache;
511 if (bc->alloc_lock == current) {
512 trace_btree_node_cannibalize_unlock(c);
513 bc->alloc_lock = NULL;
514 closure_wake_up(&bc->alloc_wait);
518 int bch2_btree_cache_cannibalize_lock(struct bch_fs *c, struct closure *cl)
520 struct btree_cache *bc = &c->btree_cache;
521 struct task_struct *old;
523 old = cmpxchg(&bc->alloc_lock, NULL, current);
524 if (old == NULL || old == current)
528 trace_btree_node_cannibalize_lock_fail(c);
532 closure_wait(&bc->alloc_wait, cl);
534 /* Try again, after adding ourselves to waitlist */
535 old = cmpxchg(&bc->alloc_lock, NULL, current);
536 if (old == NULL || old == current) {
538 closure_wake_up(&bc->alloc_wait);
542 trace_btree_node_cannibalize_lock_fail(c);
546 trace_btree_node_cannibalize_lock(c);
550 static struct btree *btree_node_cannibalize(struct bch_fs *c)
552 struct btree_cache *bc = &c->btree_cache;
555 list_for_each_entry_reverse(b, &bc->live, list)
556 if (!btree_node_reclaim(c, b))
560 list_for_each_entry_reverse(b, &bc->live, list)
561 if (!btree_node_write_and_reclaim(c, b))
565 * Rare case: all nodes were intent-locked.
568 WARN_ONCE(1, "btree cache cannibalize failed\n");
573 struct btree *bch2_btree_node_mem_alloc(struct bch_fs *c, bool pcpu_read_locks)
575 struct btree_cache *bc = &c->btree_cache;
576 struct list_head *freed = pcpu_read_locks
578 : &bc->freed_nonpcpu;
579 struct btree *b, *b2;
580 u64 start_time = local_clock();
583 flags = memalloc_nofs_save();
584 mutex_lock(&bc->lock);
587 * We never free struct btree itself, just the memory that holds the on
588 * disk node. Check the freed list before allocating a new one:
590 list_for_each_entry(b, freed, list)
591 if (!btree_node_reclaim(c, b)) {
592 list_del_init(&b->list);
596 b = __btree_node_mem_alloc(c);
601 six_lock_pcpu_alloc(&b->c.lock);
603 BUG_ON(!six_trylock_intent(&b->c.lock));
604 BUG_ON(!six_trylock_write(&b->c.lock));
608 * btree_free() doesn't free memory; it sticks the node on the end of
609 * the list. Check if there's any freed nodes there:
611 list_for_each_entry(b2, &bc->freeable, list)
612 if (!btree_node_reclaim(c, b2)) {
613 swap(b->data, b2->data);
614 swap(b->aux_data, b2->aux_data);
615 btree_node_to_freedlist(bc, b2);
616 six_unlock_write(&b2->c.lock);
617 six_unlock_intent(&b2->c.lock);
621 mutex_unlock(&bc->lock);
623 if (btree_node_data_alloc(c, b, __GFP_NOWARN|GFP_KERNEL))
626 mutex_lock(&bc->lock);
629 mutex_unlock(&bc->lock);
631 BUG_ON(btree_node_hashed(b));
632 BUG_ON(btree_node_dirty(b));
633 BUG_ON(btree_node_write_in_flight(b));
640 b->whiteout_u64s = 0;
641 bch2_btree_keys_init(b);
642 set_btree_node_accessed(b);
644 bch2_time_stats_update(&c->times[BCH_TIME_btree_node_mem_alloc],
647 memalloc_nofs_restore(flags);
650 mutex_lock(&bc->lock);
652 /* Try to cannibalize another cached btree node: */
653 if (bc->alloc_lock == current) {
654 b2 = btree_node_cannibalize(c);
655 bch2_btree_node_hash_remove(bc, b2);
658 swap(b->data, b2->data);
659 swap(b->aux_data, b2->aux_data);
660 btree_node_to_freedlist(bc, b2);
661 six_unlock_write(&b2->c.lock);
662 six_unlock_intent(&b2->c.lock);
665 list_del_init(&b->list);
668 mutex_unlock(&bc->lock);
670 trace_btree_node_cannibalize(c);
674 mutex_unlock(&bc->lock);
675 memalloc_nofs_restore(flags);
676 return ERR_PTR(-ENOMEM);
679 /* Slowpath, don't want it inlined into btree_iter_traverse() */
680 static noinline struct btree *bch2_btree_node_fill(struct bch_fs *c,
681 struct btree_trans *trans,
682 struct btree_path *path,
683 const struct bkey_i *k,
684 enum btree_id btree_id,
686 enum six_lock_type lock_type,
689 struct btree_cache *bc = &c->btree_cache;
693 BUG_ON(level + 1 >= BTREE_MAX_DEPTH);
695 * Parent node must be locked, else we could read in a btree node that's
698 if (trans && !bch2_btree_node_relock(trans, path, level + 1)) {
699 trace_trans_restart_relock_parent_for_fill(trans->fn,
700 _THIS_IP_, btree_id, &path->pos);
701 btree_trans_restart(trans);
702 return ERR_PTR(-EINTR);
705 b = bch2_btree_node_mem_alloc(c, level != 0);
707 if (trans && b == ERR_PTR(-ENOMEM)) {
708 trans->memory_allocation_failure = true;
709 trace_trans_restart_memory_allocation_failure(trans->fn,
710 _THIS_IP_, btree_id, &path->pos);
711 btree_trans_restart(trans);
712 return ERR_PTR(-EINTR);
718 bkey_copy(&b->key, k);
719 if (bch2_btree_node_hash_insert(bc, b, level, btree_id)) {
720 /* raced with another fill: */
722 /* mark as unhashed... */
725 mutex_lock(&bc->lock);
726 list_add(&b->list, &bc->freeable);
727 mutex_unlock(&bc->lock);
729 six_unlock_write(&b->c.lock);
730 six_unlock_intent(&b->c.lock);
734 set_btree_node_read_in_flight(b);
736 six_unlock_write(&b->c.lock);
737 seq = b->c.lock.state.seq;
738 six_unlock_intent(&b->c.lock);
740 /* Unlock before doing IO: */
742 bch2_trans_unlock(trans);
744 bch2_btree_node_read(c, b, sync);
750 (!bch2_trans_relock(trans) ||
751 !bch2_btree_path_relock_intent(trans, path))) {
752 BUG_ON(!trans->restarted);
753 return ERR_PTR(-EINTR);
756 if (!six_relock_type(&b->c.lock, lock_type, seq)) {
757 trace_trans_restart_relock_after_fill(trans->fn, _THIS_IP_,
758 btree_id, &path->pos);
759 btree_trans_restart(trans);
760 return ERR_PTR(-EINTR);
766 static int lock_node_check_fn(struct six_lock *lock, void *p)
768 struct btree *b = container_of(lock, struct btree, c.lock);
769 const struct bkey_i *k = p;
771 return b->hash_val == btree_ptr_hash_val(k) ? 0 : -1;
774 static noinline void btree_bad_header(struct bch_fs *c, struct btree *b)
776 struct printbuf buf1 = PRINTBUF;
777 struct printbuf buf2 = PRINTBUF;
778 struct printbuf buf3 = PRINTBUF;
780 if (!test_bit(BCH_FS_INITIAL_GC_DONE, &c->flags))
783 bch2_bkey_val_to_text(&buf1, c, bkey_i_to_s_c(&b->key));
784 bch2_bpos_to_text(&buf2, b->data->min_key);
785 bch2_bpos_to_text(&buf3, b->data->max_key);
787 bch2_fs_inconsistent(c, "btree node header doesn't match ptr\n"
788 "btree %s level %u\n"
790 "header: btree %s level %llu\n"
792 bch2_btree_ids[b->c.btree_id], b->c.level,
794 bch2_btree_ids[BTREE_NODE_ID(b->data)],
795 BTREE_NODE_LEVEL(b->data),
798 printbuf_exit(&buf3);
799 printbuf_exit(&buf2);
800 printbuf_exit(&buf1);
803 static inline void btree_check_header(struct bch_fs *c, struct btree *b)
805 if (b->c.btree_id != BTREE_NODE_ID(b->data) ||
806 b->c.level != BTREE_NODE_LEVEL(b->data) ||
807 bpos_cmp(b->data->max_key, b->key.k.p) ||
808 (b->key.k.type == KEY_TYPE_btree_ptr_v2 &&
809 bpos_cmp(b->data->min_key,
810 bkey_i_to_btree_ptr_v2(&b->key)->v.min_key)))
811 btree_bad_header(c, b);
815 * bch_btree_node_get - find a btree node in the cache and lock it, reading it
816 * in from disk if necessary.
818 * If IO is necessary and running under generic_make_request, returns -EAGAIN.
820 * The btree node will have either a read or a write lock held, depending on
821 * the @write parameter.
823 struct btree *bch2_btree_node_get(struct btree_trans *trans, struct btree_path *path,
824 const struct bkey_i *k, unsigned level,
825 enum six_lock_type lock_type,
826 unsigned long trace_ip)
828 struct bch_fs *c = trans->c;
829 struct btree_cache *bc = &c->btree_cache;
833 EBUG_ON(level >= BTREE_MAX_DEPTH);
835 b = btree_node_mem_ptr(k);
838 * Check b->hash_val _before_ calling btree_node_lock() - this might not
839 * be the node we want anymore, and trying to lock the wrong node could
840 * cause an unneccessary transaction restart:
842 if (likely(c->opts.btree_node_mem_ptr_optimization &&
844 b->hash_val == btree_ptr_hash_val(k)))
847 b = btree_cache_find(bc, k);
850 * We must have the parent locked to call bch2_btree_node_fill(),
851 * else we could read in a btree node from disk that's been
854 b = bch2_btree_node_fill(c, trans, path, k, path->btree_id,
855 level, lock_type, true);
857 /* We raced and found the btree node in the cache */
866 * There's a potential deadlock with splits and insertions into
867 * interior nodes we have to avoid:
869 * The other thread might be holding an intent lock on the node
870 * we want, and they want to update its parent node so they're
871 * going to upgrade their intent lock on the parent node to a
874 * But if we're holding a read lock on the parent, and we're
875 * trying to get the intent lock they're holding, we deadlock.
877 * So to avoid this we drop the read locks on parent nodes when
878 * we're starting to take intent locks - and handle the race.
880 * The race is that they might be about to free the node we
881 * want, and dropping our read lock on the parent node lets them
882 * update the parent marking the node we want as freed, and then
885 * To guard against this, btree nodes are evicted from the cache
886 * when they're freed - and b->hash_val is zeroed out, which we
887 * check for after we lock the node.
889 * Then, bch2_btree_node_relock() on the parent will fail - because
890 * the parent was modified, when the pointer to the node we want
891 * was removed - and we'll bail out:
893 if (btree_node_read_locked(path, level + 1))
894 btree_node_unlock(path, level + 1);
896 if (!btree_node_lock(trans, path, b, k->k.p, level, lock_type,
897 lock_node_check_fn, (void *) k, trace_ip)) {
898 if (!trans->restarted)
900 return ERR_PTR(-EINTR);
903 if (unlikely(b->hash_val != btree_ptr_hash_val(k) ||
904 b->c.level != level ||
906 six_unlock_type(&b->c.lock, lock_type);
907 if (bch2_btree_node_relock(trans, path, level + 1))
910 trace_trans_restart_btree_node_reused(trans->fn,
914 btree_trans_restart(trans);
915 return ERR_PTR(-EINTR);
919 if (unlikely(btree_node_read_in_flight(b))) {
920 u32 seq = b->c.lock.state.seq;
922 six_unlock_type(&b->c.lock, lock_type);
923 bch2_trans_unlock(trans);
925 bch2_btree_node_wait_on_read(b);
928 * should_be_locked is not set on this path yet, so we need to
929 * relock it specifically:
932 (!bch2_trans_relock(trans) ||
933 !bch2_btree_path_relock_intent(trans, path))) {
934 BUG_ON(!trans->restarted);
935 return ERR_PTR(-EINTR);
938 if (!six_relock_type(&b->c.lock, lock_type, seq))
942 prefetch(b->aux_data);
944 for_each_bset(b, t) {
945 void *p = (u64 *) b->aux_data + t->aux_data_offset;
947 prefetch(p + L1_CACHE_BYTES * 0);
948 prefetch(p + L1_CACHE_BYTES * 1);
949 prefetch(p + L1_CACHE_BYTES * 2);
952 /* avoid atomic set bit if it's not needed: */
953 if (!btree_node_accessed(b))
954 set_btree_node_accessed(b);
956 if (unlikely(btree_node_read_error(b))) {
957 six_unlock_type(&b->c.lock, lock_type);
958 return ERR_PTR(-EIO);
961 EBUG_ON(b->c.btree_id != path->btree_id);
962 EBUG_ON(BTREE_NODE_LEVEL(b->data) != level);
963 btree_check_header(c, b);
968 struct btree *bch2_btree_node_get_noiter(struct bch_fs *c,
969 const struct bkey_i *k,
970 enum btree_id btree_id,
974 struct btree_cache *bc = &c->btree_cache;
979 EBUG_ON(level >= BTREE_MAX_DEPTH);
981 if (c->opts.btree_node_mem_ptr_optimization) {
982 b = btree_node_mem_ptr(k);
987 b = btree_cache_find(bc, k);
992 b = bch2_btree_node_fill(c, NULL, NULL, k, btree_id,
993 level, SIX_LOCK_read, true);
995 /* We raced and found the btree node in the cache */
1000 !bch2_btree_cache_cannibalize_lock(c, NULL))
1007 ret = six_lock_read(&b->c.lock, lock_node_check_fn, (void *) k);
1011 if (unlikely(b->hash_val != btree_ptr_hash_val(k) ||
1012 b->c.btree_id != btree_id ||
1013 b->c.level != level)) {
1014 six_unlock_read(&b->c.lock);
1019 /* XXX: waiting on IO with btree locks held: */
1020 __bch2_btree_node_wait_on_read(b);
1022 prefetch(b->aux_data);
1024 for_each_bset(b, t) {
1025 void *p = (u64 *) b->aux_data + t->aux_data_offset;
1027 prefetch(p + L1_CACHE_BYTES * 0);
1028 prefetch(p + L1_CACHE_BYTES * 1);
1029 prefetch(p + L1_CACHE_BYTES * 2);
1032 /* avoid atomic set bit if it's not needed: */
1033 if (!btree_node_accessed(b))
1034 set_btree_node_accessed(b);
1036 if (unlikely(btree_node_read_error(b))) {
1037 six_unlock_read(&b->c.lock);
1042 EBUG_ON(b->c.btree_id != btree_id);
1043 EBUG_ON(BTREE_NODE_LEVEL(b->data) != level);
1044 btree_check_header(c, b);
1046 bch2_btree_cache_cannibalize_unlock(c);
1050 int bch2_btree_node_prefetch(struct bch_fs *c,
1051 struct btree_trans *trans,
1052 struct btree_path *path,
1053 const struct bkey_i *k,
1054 enum btree_id btree_id, unsigned level)
1056 struct btree_cache *bc = &c->btree_cache;
1059 BUG_ON(trans && !btree_node_locked(path, level + 1));
1060 BUG_ON(level >= BTREE_MAX_DEPTH);
1062 b = btree_cache_find(bc, k);
1066 b = bch2_btree_node_fill(c, trans, path, k, btree_id,
1067 level, SIX_LOCK_read, false);
1068 return PTR_ERR_OR_ZERO(b);
1071 void bch2_btree_node_evict(struct bch_fs *c, const struct bkey_i *k)
1073 struct btree_cache *bc = &c->btree_cache;
1076 b = btree_cache_find(bc, k);
1080 /* not allowed to wait on io with btree locks held: */
1082 /* XXX we're called from btree_gc which will be holding other btree
1085 __bch2_btree_node_wait_on_read(b);
1086 __bch2_btree_node_wait_on_write(b);
1088 six_lock_intent(&b->c.lock, NULL, NULL);
1089 six_lock_write(&b->c.lock, NULL, NULL);
1091 if (btree_node_dirty(b)) {
1092 __bch2_btree_node_write(c, b, 0);
1093 six_unlock_write(&b->c.lock);
1094 six_unlock_intent(&b->c.lock);
1098 BUG_ON(btree_node_dirty(b));
1100 mutex_lock(&bc->lock);
1101 btree_node_data_free(c, b);
1102 bch2_btree_node_hash_remove(bc, b);
1103 mutex_unlock(&bc->lock);
1105 six_unlock_write(&b->c.lock);
1106 six_unlock_intent(&b->c.lock);
1109 void bch2_btree_node_to_text(struct printbuf *out, struct bch_fs *c,
1112 const struct bkey_format *f = &b->format;
1113 struct bset_stats stats;
1115 memset(&stats, 0, sizeof(stats));
1117 bch2_btree_keys_stats(b, &stats);
1119 pr_buf(out, "l %u ", b->c.level);
1120 bch2_bpos_to_text(out, b->data->min_key);
1122 bch2_bpos_to_text(out, b->data->max_key);
1125 bch2_val_to_text(out, c, bkey_i_to_s_c(&b->key));
1128 " format: u64s %u fields %u %u %u %u %u\n"
1129 " unpack fn len: %u\n"
1130 " bytes used %zu/%zu (%zu%% full)\n"
1131 " sib u64s: %u, %u (merge threshold %u)\n"
1132 " nr packed keys %u\n"
1133 " nr unpacked keys %u\n"
1135 " failed unpacked %zu\n",
1137 f->bits_per_field[0],
1138 f->bits_per_field[1],
1139 f->bits_per_field[2],
1140 f->bits_per_field[3],
1141 f->bits_per_field[4],
1143 b->nr.live_u64s * sizeof(u64),
1144 btree_bytes(c) - sizeof(struct btree_node),
1145 b->nr.live_u64s * 100 / btree_max_u64s(c),
1148 c->btree_foreground_merge_threshold,
1150 b->nr.unpacked_keys,
1155 void bch2_btree_cache_to_text(struct printbuf *out, struct bch_fs *c)
1157 pr_buf(out, "nr nodes:\t\t%u\n", c->btree_cache.used);
1158 pr_buf(out, "nr dirty:\t\t%u\n", atomic_read(&c->btree_cache.dirty));
1159 pr_buf(out, "cannibalize lock:\t%p\n", c->btree_cache.alloc_lock);