1 // SPDX-License-Identifier: GPL-2.0
6 #include "btree_cache.h"
8 #include "btree_iter.h"
9 #include "btree_locking.h"
16 #include <linux/prefetch.h>
17 #include <linux/sched/mm.h>
19 const char * const bch2_btree_node_flags[] = {
26 void bch2_recalc_btree_reserve(struct bch_fs *c)
28 unsigned i, reserve = 16;
30 if (!c->btree_roots_known[0].b)
33 for (i = 0; i < btree_id_nr_alive(c); i++) {
34 struct btree_root *r = bch2_btree_id_root(c, i);
37 reserve += min_t(unsigned, 1, r->b->c.level) * 8;
40 c->btree_cache.reserve = reserve;
43 static inline unsigned btree_cache_can_free(struct btree_cache *bc)
45 return max_t(int, 0, bc->used - bc->reserve);
48 static void btree_node_to_freedlist(struct btree_cache *bc, struct btree *b)
50 if (b->c.lock.readers)
51 list_move(&b->list, &bc->freed_pcpu);
53 list_move(&b->list, &bc->freed_nonpcpu);
56 static void btree_node_data_free(struct bch_fs *c, struct btree *b)
58 struct btree_cache *bc = &c->btree_cache;
60 EBUG_ON(btree_node_write_in_flight(b));
62 clear_btree_node_just_written(b);
69 munmap(b->aux_data, btree_aux_data_bytes(b));
75 btree_node_to_freedlist(bc, b);
78 static int bch2_btree_cache_cmp_fn(struct rhashtable_compare_arg *arg,
81 const struct btree *b = obj;
82 const u64 *v = arg->key;
84 return b->hash_val == *v ? 0 : 1;
87 static const struct rhashtable_params bch_btree_cache_params = {
88 .head_offset = offsetof(struct btree, hash),
89 .key_offset = offsetof(struct btree, hash_val),
90 .key_len = sizeof(u64),
91 .obj_cmpfn = bch2_btree_cache_cmp_fn,
94 static int btree_node_data_alloc(struct bch_fs *c, struct btree *b, gfp_t gfp)
96 BUG_ON(b->data || b->aux_data);
98 b->data = kvmalloc(btree_buf_bytes(b), gfp);
100 return -BCH_ERR_ENOMEM_btree_node_mem_alloc;
102 b->aux_data = kvmalloc(btree_aux_data_bytes(b), gfp);
104 b->aux_data = mmap(NULL, btree_aux_data_bytes(b),
105 PROT_READ|PROT_WRITE|PROT_EXEC,
106 MAP_PRIVATE|MAP_ANONYMOUS, 0, 0);
107 if (b->aux_data == MAP_FAILED)
113 return -BCH_ERR_ENOMEM_btree_node_mem_alloc;
119 static struct btree *__btree_node_mem_alloc(struct bch_fs *c, gfp_t gfp)
123 b = kzalloc(sizeof(struct btree), gfp);
127 bkey_btree_ptr_init(&b->key);
128 INIT_LIST_HEAD(&b->list);
129 INIT_LIST_HEAD(&b->write_blocked);
130 b->byte_order = ilog2(c->opts.btree_node_size);
134 struct btree *__bch2_btree_node_mem_alloc(struct bch_fs *c)
136 struct btree_cache *bc = &c->btree_cache;
139 b = __btree_node_mem_alloc(c, GFP_KERNEL);
143 if (btree_node_data_alloc(c, b, GFP_KERNEL)) {
148 bch2_btree_lock_init(&b->c, 0);
151 list_add(&b->list, &bc->freeable);
155 /* Btree in memory cache - hash table */
157 void bch2_btree_node_hash_remove(struct btree_cache *bc, struct btree *b)
159 int ret = rhashtable_remove_fast(&bc->table, &b->hash, bch_btree_cache_params);
163 /* Cause future lookups for this node to fail: */
167 int __bch2_btree_node_hash_insert(struct btree_cache *bc, struct btree *b)
170 b->hash_val = btree_ptr_hash_val(&b->key);
172 return rhashtable_lookup_insert_fast(&bc->table, &b->hash,
173 bch_btree_cache_params);
176 int bch2_btree_node_hash_insert(struct btree_cache *bc, struct btree *b,
177 unsigned level, enum btree_id id)
184 mutex_lock(&bc->lock);
185 ret = __bch2_btree_node_hash_insert(bc, b);
187 list_add_tail(&b->list, &bc->live);
188 mutex_unlock(&bc->lock);
194 static inline struct btree *btree_cache_find(struct btree_cache *bc,
195 const struct bkey_i *k)
197 u64 v = btree_ptr_hash_val(k);
199 return rhashtable_lookup_fast(&bc->table, &v, bch_btree_cache_params);
203 * this version is for btree nodes that have already been freed (we're not
204 * reaping a real btree node)
206 static int __btree_node_reclaim(struct bch_fs *c, struct btree *b, bool flush)
208 struct btree_cache *bc = &c->btree_cache;
211 lockdep_assert_held(&bc->lock);
213 struct bbpos pos = BBPOS(b->c.btree_id, b->key.k.p);
215 u64 mask = b->c.level
216 ? bc->pinned_nodes_interior_mask
217 : bc->pinned_nodes_leaf_mask;
219 if ((mask & BIT_ULL(b->c.btree_id)) &&
220 bbpos_cmp(bc->pinned_nodes_start, pos) < 0 &&
221 bbpos_cmp(bc->pinned_nodes_end, pos) >= 0)
222 return -BCH_ERR_ENOMEM_btree_node_reclaim;
225 if (b->flags & ((1U << BTREE_NODE_dirty)|
226 (1U << BTREE_NODE_read_in_flight)|
227 (1U << BTREE_NODE_write_in_flight))) {
229 return -BCH_ERR_ENOMEM_btree_node_reclaim;
231 /* XXX: waiting on IO with btree cache lock held */
232 bch2_btree_node_wait_on_read(b);
233 bch2_btree_node_wait_on_write(b);
236 if (!six_trylock_intent(&b->c.lock))
237 return -BCH_ERR_ENOMEM_btree_node_reclaim;
239 if (!six_trylock_write(&b->c.lock))
240 goto out_unlock_intent;
242 /* recheck under lock */
243 if (b->flags & ((1U << BTREE_NODE_read_in_flight)|
244 (1U << BTREE_NODE_write_in_flight))) {
247 six_unlock_write(&b->c.lock);
248 six_unlock_intent(&b->c.lock);
252 if (btree_node_noevict(b) ||
253 btree_node_write_blocked(b) ||
254 btree_node_will_make_reachable(b))
257 if (btree_node_dirty(b)) {
261 * Using the underscore version because we don't want to compact
262 * bsets after the write, since this node is about to be evicted
263 * - unless btree verify mode is enabled, since it runs out of
264 * the post write cleanup:
266 if (bch2_verify_btree_ondisk)
267 bch2_btree_node_write(c, b, SIX_LOCK_intent,
268 BTREE_WRITE_cache_reclaim);
270 __bch2_btree_node_write(c, b,
271 BTREE_WRITE_cache_reclaim);
273 six_unlock_write(&b->c.lock);
274 six_unlock_intent(&b->c.lock);
278 if (b->hash_val && !ret)
279 trace_and_count(c, btree_cache_reap, c, b);
282 six_unlock_write(&b->c.lock);
284 six_unlock_intent(&b->c.lock);
285 ret = -BCH_ERR_ENOMEM_btree_node_reclaim;
289 static int btree_node_reclaim(struct bch_fs *c, struct btree *b)
291 return __btree_node_reclaim(c, b, false);
294 static int btree_node_write_and_reclaim(struct bch_fs *c, struct btree *b)
296 return __btree_node_reclaim(c, b, true);
299 static unsigned long bch2_btree_cache_scan(struct shrinker *shrink,
300 struct shrink_control *sc)
302 struct bch_fs *c = shrink->private_data;
303 struct btree_cache *bc = &c->btree_cache;
305 unsigned long nr = sc->nr_to_scan;
306 unsigned long can_free = 0;
307 unsigned long freed = 0;
308 unsigned long touched = 0;
310 unsigned long ret = SHRINK_STOP;
311 bool trigger_writes = atomic_read(&bc->dirty) + nr >=
314 if (bch2_btree_shrinker_disabled)
317 mutex_lock(&bc->lock);
318 flags = memalloc_nofs_save();
321 * It's _really_ critical that we don't free too many btree nodes - we
322 * have to always leave ourselves a reserve. The reserve is how we
323 * guarantee that allocating memory for a new btree node can always
324 * succeed, so that inserting keys into the btree can always succeed and
325 * IO can always make forward progress:
327 can_free = btree_cache_can_free(bc);
328 nr = min_t(unsigned long, nr, can_free);
331 list_for_each_entry_safe(b, t, &bc->freeable, list) {
333 * Leave a few nodes on the freeable list, so that a btree split
334 * won't have to hit the system allocator:
344 if (!btree_node_reclaim(c, b)) {
345 btree_node_data_free(c, b);
346 six_unlock_write(&b->c.lock);
347 six_unlock_intent(&b->c.lock);
352 list_for_each_entry_safe(b, t, &bc->live, list) {
355 if (btree_node_accessed(b)) {
356 clear_btree_node_accessed(b);
357 } else if (!btree_node_reclaim(c, b)) {
359 btree_node_data_free(c, b);
361 bch2_btree_node_hash_remove(bc, b);
362 six_unlock_write(&b->c.lock);
363 six_unlock_intent(&b->c.lock);
367 } else if (trigger_writes &&
368 btree_node_dirty(b) &&
369 !btree_node_will_make_reachable(b) &&
370 !btree_node_write_blocked(b) &&
371 six_trylock_read(&b->c.lock)) {
372 list_move(&bc->live, &b->list);
373 mutex_unlock(&bc->lock);
374 __bch2_btree_node_write(c, b, BTREE_WRITE_cache_reclaim);
375 six_unlock_read(&b->c.lock);
378 mutex_lock(&bc->lock);
386 if (&t->list != &bc->live)
387 list_move_tail(&bc->live, &t->list);
389 mutex_unlock(&bc->lock);
392 memalloc_nofs_restore(flags);
393 trace_and_count(c, btree_cache_scan, sc->nr_to_scan, can_free, ret);
397 static unsigned long bch2_btree_cache_count(struct shrinker *shrink,
398 struct shrink_control *sc)
400 struct bch_fs *c = shrink->private_data;
401 struct btree_cache *bc = &c->btree_cache;
403 if (bch2_btree_shrinker_disabled)
406 return btree_cache_can_free(bc);
409 void bch2_fs_btree_cache_exit(struct bch_fs *c)
411 struct btree_cache *bc = &c->btree_cache;
415 shrinker_free(bc->shrink);
417 /* vfree() can allocate memory: */
418 flags = memalloc_nofs_save();
419 mutex_lock(&bc->lock);
422 list_move(&c->verify_data->list, &bc->live);
424 kvfree(c->verify_ondisk);
426 for (i = 0; i < btree_id_nr_alive(c); i++) {
427 struct btree_root *r = bch2_btree_id_root(c, i);
430 list_add(&r->b->list, &bc->live);
433 list_splice(&bc->freeable, &bc->live);
435 while (!list_empty(&bc->live)) {
436 b = list_first_entry(&bc->live, struct btree, list);
438 BUG_ON(btree_node_read_in_flight(b) ||
439 btree_node_write_in_flight(b));
441 btree_node_data_free(c, b);
444 BUG_ON(!bch2_journal_error(&c->journal) &&
445 atomic_read(&c->btree_cache.dirty));
447 list_splice(&bc->freed_pcpu, &bc->freed_nonpcpu);
449 while (!list_empty(&bc->freed_nonpcpu)) {
450 b = list_first_entry(&bc->freed_nonpcpu, struct btree, list);
452 six_lock_exit(&b->c.lock);
456 mutex_unlock(&bc->lock);
457 memalloc_nofs_restore(flags);
459 if (bc->table_init_done)
460 rhashtable_destroy(&bc->table);
463 int bch2_fs_btree_cache_init(struct bch_fs *c)
465 struct btree_cache *bc = &c->btree_cache;
466 struct shrinker *shrink;
470 ret = rhashtable_init(&bc->table, &bch_btree_cache_params);
474 bc->table_init_done = true;
476 bch2_recalc_btree_reserve(c);
478 for (i = 0; i < bc->reserve; i++)
479 if (!__bch2_btree_node_mem_alloc(c))
482 list_splice_init(&bc->live, &bc->freeable);
484 mutex_init(&c->verify_lock);
486 shrink = shrinker_alloc(0, "%s-btree_cache", c->name);
490 shrink->count_objects = bch2_btree_cache_count;
491 shrink->scan_objects = bch2_btree_cache_scan;
493 shrink->private_data = c;
494 shrinker_register(shrink);
498 return -BCH_ERR_ENOMEM_fs_btree_cache_init;
501 void bch2_fs_btree_cache_init_early(struct btree_cache *bc)
503 mutex_init(&bc->lock);
504 INIT_LIST_HEAD(&bc->live);
505 INIT_LIST_HEAD(&bc->freeable);
506 INIT_LIST_HEAD(&bc->freed_pcpu);
507 INIT_LIST_HEAD(&bc->freed_nonpcpu);
511 * We can only have one thread cannibalizing other cached btree nodes at a time,
512 * or we'll deadlock. We use an open coded mutex to ensure that, which a
513 * cannibalize_bucket() will take. This means every time we unlock the root of
514 * the btree, we need to release this lock if we have it held.
516 void bch2_btree_cache_cannibalize_unlock(struct btree_trans *trans)
518 struct bch_fs *c = trans->c;
519 struct btree_cache *bc = &c->btree_cache;
521 if (bc->alloc_lock == current) {
522 trace_and_count(c, btree_cache_cannibalize_unlock, trans);
523 bc->alloc_lock = NULL;
524 closure_wake_up(&bc->alloc_wait);
528 int bch2_btree_cache_cannibalize_lock(struct btree_trans *trans, struct closure *cl)
530 struct bch_fs *c = trans->c;
531 struct btree_cache *bc = &c->btree_cache;
532 struct task_struct *old;
534 old = cmpxchg(&bc->alloc_lock, NULL, current);
535 if (old == NULL || old == current)
539 trace_and_count(c, btree_cache_cannibalize_lock_fail, trans);
540 return -BCH_ERR_ENOMEM_btree_cache_cannibalize_lock;
543 closure_wait(&bc->alloc_wait, cl);
545 /* Try again, after adding ourselves to waitlist */
546 old = cmpxchg(&bc->alloc_lock, NULL, current);
547 if (old == NULL || old == current) {
549 closure_wake_up(&bc->alloc_wait);
553 trace_and_count(c, btree_cache_cannibalize_lock_fail, trans);
554 return -BCH_ERR_btree_cache_cannibalize_lock_blocked;
557 trace_and_count(c, btree_cache_cannibalize_lock, trans);
561 static struct btree *btree_node_cannibalize(struct bch_fs *c)
563 struct btree_cache *bc = &c->btree_cache;
566 list_for_each_entry_reverse(b, &bc->live, list)
567 if (!btree_node_reclaim(c, b))
571 list_for_each_entry_reverse(b, &bc->live, list)
572 if (!btree_node_write_and_reclaim(c, b))
576 * Rare case: all nodes were intent-locked.
579 WARN_ONCE(1, "btree cache cannibalize failed\n");
584 struct btree *bch2_btree_node_mem_alloc(struct btree_trans *trans, bool pcpu_read_locks)
586 struct bch_fs *c = trans->c;
587 struct btree_cache *bc = &c->btree_cache;
588 struct list_head *freed = pcpu_read_locks
590 : &bc->freed_nonpcpu;
591 struct btree *b, *b2;
592 u64 start_time = local_clock();
595 flags = memalloc_nofs_save();
596 mutex_lock(&bc->lock);
599 * We never free struct btree itself, just the memory that holds the on
600 * disk node. Check the freed list before allocating a new one:
602 list_for_each_entry(b, freed, list)
603 if (!btree_node_reclaim(c, b)) {
604 list_del_init(&b->list);
608 b = __btree_node_mem_alloc(c, GFP_NOWAIT|__GFP_NOWARN);
610 mutex_unlock(&bc->lock);
611 bch2_trans_unlock(trans);
612 b = __btree_node_mem_alloc(c, GFP_KERNEL);
615 mutex_lock(&bc->lock);
618 bch2_btree_lock_init(&b->c, pcpu_read_locks ? SIX_LOCK_INIT_PCPU : 0);
620 BUG_ON(!six_trylock_intent(&b->c.lock));
621 BUG_ON(!six_trylock_write(&b->c.lock));
625 * btree_free() doesn't free memory; it sticks the node on the end of
626 * the list. Check if there's any freed nodes there:
628 list_for_each_entry(b2, &bc->freeable, list)
629 if (!btree_node_reclaim(c, b2)) {
630 swap(b->data, b2->data);
631 swap(b->aux_data, b2->aux_data);
632 btree_node_to_freedlist(bc, b2);
633 six_unlock_write(&b2->c.lock);
634 six_unlock_intent(&b2->c.lock);
638 mutex_unlock(&bc->lock);
640 if (btree_node_data_alloc(c, b, GFP_NOWAIT|__GFP_NOWARN)) {
641 bch2_trans_unlock(trans);
642 if (btree_node_data_alloc(c, b, GFP_KERNEL|__GFP_NOWARN))
646 mutex_lock(&bc->lock);
649 mutex_unlock(&bc->lock);
651 BUG_ON(btree_node_hashed(b));
652 BUG_ON(btree_node_dirty(b));
653 BUG_ON(btree_node_write_in_flight(b));
660 b->whiteout_u64s = 0;
661 bch2_btree_keys_init(b);
662 set_btree_node_accessed(b);
664 time_stats_update(&c->times[BCH_TIME_btree_node_mem_alloc],
667 memalloc_nofs_restore(flags);
670 mutex_lock(&bc->lock);
672 /* Try to cannibalize another cached btree node: */
673 if (bc->alloc_lock == current) {
674 b2 = btree_node_cannibalize(c);
675 clear_btree_node_just_written(b2);
676 bch2_btree_node_hash_remove(bc, b2);
679 swap(b->data, b2->data);
680 swap(b->aux_data, b2->aux_data);
681 btree_node_to_freedlist(bc, b2);
682 six_unlock_write(&b2->c.lock);
683 six_unlock_intent(&b2->c.lock);
686 list_del_init(&b->list);
689 mutex_unlock(&bc->lock);
691 trace_and_count(c, btree_cache_cannibalize, trans);
695 mutex_unlock(&bc->lock);
696 memalloc_nofs_restore(flags);
697 return ERR_PTR(-BCH_ERR_ENOMEM_btree_node_mem_alloc);
700 /* Slowpath, don't want it inlined into btree_iter_traverse() */
701 static noinline struct btree *bch2_btree_node_fill(struct btree_trans *trans,
702 struct btree_path *path,
703 const struct bkey_i *k,
704 enum btree_id btree_id,
706 enum six_lock_type lock_type,
709 struct bch_fs *c = trans->c;
710 struct btree_cache *bc = &c->btree_cache;
714 BUG_ON(level + 1 >= BTREE_MAX_DEPTH);
716 * Parent node must be locked, else we could read in a btree node that's
719 if (path && !bch2_btree_node_relock(trans, path, level + 1)) {
720 trace_and_count(c, trans_restart_relock_parent_for_fill, trans, _THIS_IP_, path);
721 return ERR_PTR(btree_trans_restart(trans, BCH_ERR_transaction_restart_fill_relock));
724 b = bch2_btree_node_mem_alloc(trans, level != 0);
726 if (bch2_err_matches(PTR_ERR_OR_ZERO(b), ENOMEM)) {
730 trans->memory_allocation_failure = true;
731 trace_and_count(c, trans_restart_memory_allocation_failure, trans, _THIS_IP_, path);
732 return ERR_PTR(btree_trans_restart(trans, BCH_ERR_transaction_restart_fill_mem_alloc_fail));
738 bkey_copy(&b->key, k);
739 if (bch2_btree_node_hash_insert(bc, b, level, btree_id)) {
740 /* raced with another fill: */
742 /* mark as unhashed... */
745 mutex_lock(&bc->lock);
746 list_add(&b->list, &bc->freeable);
747 mutex_unlock(&bc->lock);
749 six_unlock_write(&b->c.lock);
750 six_unlock_intent(&b->c.lock);
754 set_btree_node_read_in_flight(b);
756 six_unlock_write(&b->c.lock);
757 seq = six_lock_seq(&b->c.lock);
758 six_unlock_intent(&b->c.lock);
760 /* Unlock before doing IO: */
762 bch2_trans_unlock_noassert(trans);
764 bch2_btree_node_read(trans, b, sync);
770 int ret = bch2_trans_relock(trans) ?:
771 bch2_btree_path_relock_intent(trans, path);
773 BUG_ON(!trans->restarted);
778 if (!six_relock_type(&b->c.lock, lock_type, seq)) {
781 trace_and_count(c, trans_restart_relock_after_fill, trans, _THIS_IP_, path);
782 return ERR_PTR(btree_trans_restart(trans, BCH_ERR_transaction_restart_relock_after_fill));
788 static noinline void btree_bad_header(struct bch_fs *c, struct btree *b)
790 struct printbuf buf = PRINTBUF;
792 if (c->curr_recovery_pass <= BCH_RECOVERY_PASS_check_allocations)
796 "btree node header doesn't match ptr\n"
797 "btree %s level %u\n"
799 bch2_btree_id_str(b->c.btree_id), b->c.level);
800 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(&b->key));
802 prt_printf(&buf, "\nheader: btree %s level %llu\n"
804 bch2_btree_id_str(BTREE_NODE_ID(b->data)),
805 BTREE_NODE_LEVEL(b->data));
806 bch2_bpos_to_text(&buf, b->data->min_key);
808 prt_printf(&buf, "\nmax ");
809 bch2_bpos_to_text(&buf, b->data->max_key);
811 bch2_fs_inconsistent(c, "%s", buf.buf);
815 static inline void btree_check_header(struct bch_fs *c, struct btree *b)
817 if (b->c.btree_id != BTREE_NODE_ID(b->data) ||
818 b->c.level != BTREE_NODE_LEVEL(b->data) ||
819 !bpos_eq(b->data->max_key, b->key.k.p) ||
820 (b->key.k.type == KEY_TYPE_btree_ptr_v2 &&
821 !bpos_eq(b->data->min_key,
822 bkey_i_to_btree_ptr_v2(&b->key)->v.min_key)))
823 btree_bad_header(c, b);
826 static struct btree *__bch2_btree_node_get(struct btree_trans *trans, struct btree_path *path,
827 const struct bkey_i *k, unsigned level,
828 enum six_lock_type lock_type,
829 unsigned long trace_ip)
831 struct bch_fs *c = trans->c;
832 struct btree_cache *bc = &c->btree_cache;
835 bool need_relock = false;
838 EBUG_ON(level >= BTREE_MAX_DEPTH);
840 b = btree_cache_find(bc, k);
843 * We must have the parent locked to call bch2_btree_node_fill(),
844 * else we could read in a btree node from disk that's been
847 b = bch2_btree_node_fill(trans, path, k, path->btree_id,
848 level, lock_type, true);
851 /* We raced and found the btree node in the cache */
858 if (btree_node_read_locked(path, level + 1))
859 btree_node_unlock(trans, path, level + 1);
861 ret = btree_node_lock(trans, path, &b->c, level, lock_type, trace_ip);
862 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
867 if (unlikely(b->hash_val != btree_ptr_hash_val(k) ||
868 b->c.level != level ||
870 six_unlock_type(&b->c.lock, lock_type);
871 if (bch2_btree_node_relock(trans, path, level + 1))
874 trace_and_count(c, trans_restart_btree_node_reused, trans, trace_ip, path);
875 return ERR_PTR(btree_trans_restart(trans, BCH_ERR_transaction_restart_lock_node_reused));
878 /* avoid atomic set bit if it's not needed: */
879 if (!btree_node_accessed(b))
880 set_btree_node_accessed(b);
883 if (unlikely(btree_node_read_in_flight(b))) {
884 u32 seq = six_lock_seq(&b->c.lock);
886 six_unlock_type(&b->c.lock, lock_type);
887 bch2_trans_unlock(trans);
890 bch2_btree_node_wait_on_read(b);
893 * should_be_locked is not set on this path yet, so we need to
894 * relock it specifically:
896 if (!six_relock_type(&b->c.lock, lock_type, seq))
900 if (unlikely(need_relock)) {
901 ret = bch2_trans_relock(trans) ?:
902 bch2_btree_path_relock_intent(trans, path);
904 six_unlock_type(&b->c.lock, lock_type);
909 prefetch(b->aux_data);
911 for_each_bset(b, t) {
912 void *p = (u64 *) b->aux_data + t->aux_data_offset;
914 prefetch(p + L1_CACHE_BYTES * 0);
915 prefetch(p + L1_CACHE_BYTES * 1);
916 prefetch(p + L1_CACHE_BYTES * 2);
919 if (unlikely(btree_node_read_error(b))) {
920 six_unlock_type(&b->c.lock, lock_type);
921 return ERR_PTR(-BCH_ERR_btree_node_read_error);
924 EBUG_ON(b->c.btree_id != path->btree_id);
925 EBUG_ON(BTREE_NODE_LEVEL(b->data) != level);
926 btree_check_header(c, b);
932 * bch2_btree_node_get - find a btree node in the cache and lock it, reading it
933 * in from disk if necessary.
935 * @trans: btree transaction object
936 * @path: btree_path being traversed
937 * @k: pointer to btree node (generally KEY_TYPE_btree_ptr_v2)
938 * @level: level of btree node being looked up (0 == leaf node)
939 * @lock_type: SIX_LOCK_read or SIX_LOCK_intent
940 * @trace_ip: ip of caller of btree iterator code (i.e. caller of bch2_btree_iter_peek())
942 * The btree node will have either a read or a write lock held, depending on
943 * the @write parameter.
945 * Returns: btree node or ERR_PTR()
947 struct btree *bch2_btree_node_get(struct btree_trans *trans, struct btree_path *path,
948 const struct bkey_i *k, unsigned level,
949 enum six_lock_type lock_type,
950 unsigned long trace_ip)
952 struct bch_fs *c = trans->c;
957 EBUG_ON(level >= BTREE_MAX_DEPTH);
959 b = btree_node_mem_ptr(k);
962 * Check b->hash_val _before_ calling btree_node_lock() - this might not
963 * be the node we want anymore, and trying to lock the wrong node could
964 * cause an unneccessary transaction restart:
966 if (unlikely(!c->opts.btree_node_mem_ptr_optimization ||
968 b->hash_val != btree_ptr_hash_val(k)))
969 return __bch2_btree_node_get(trans, path, k, level, lock_type, trace_ip);
971 if (btree_node_read_locked(path, level + 1))
972 btree_node_unlock(trans, path, level + 1);
974 ret = btree_node_lock(trans, path, &b->c, level, lock_type, trace_ip);
975 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
980 if (unlikely(b->hash_val != btree_ptr_hash_val(k) ||
981 b->c.level != level ||
983 six_unlock_type(&b->c.lock, lock_type);
984 if (bch2_btree_node_relock(trans, path, level + 1))
985 return __bch2_btree_node_get(trans, path, k, level, lock_type, trace_ip);
987 trace_and_count(c, trans_restart_btree_node_reused, trans, trace_ip, path);
988 return ERR_PTR(btree_trans_restart(trans, BCH_ERR_transaction_restart_lock_node_reused));
991 if (unlikely(btree_node_read_in_flight(b))) {
992 six_unlock_type(&b->c.lock, lock_type);
993 return __bch2_btree_node_get(trans, path, k, level, lock_type, trace_ip);
996 prefetch(b->aux_data);
998 for_each_bset(b, t) {
999 void *p = (u64 *) b->aux_data + t->aux_data_offset;
1001 prefetch(p + L1_CACHE_BYTES * 0);
1002 prefetch(p + L1_CACHE_BYTES * 1);
1003 prefetch(p + L1_CACHE_BYTES * 2);
1006 /* avoid atomic set bit if it's not needed: */
1007 if (!btree_node_accessed(b))
1008 set_btree_node_accessed(b);
1010 if (unlikely(btree_node_read_error(b))) {
1011 six_unlock_type(&b->c.lock, lock_type);
1012 return ERR_PTR(-BCH_ERR_btree_node_read_error);
1015 EBUG_ON(b->c.btree_id != path->btree_id);
1016 EBUG_ON(BTREE_NODE_LEVEL(b->data) != level);
1017 btree_check_header(c, b);
1022 struct btree *bch2_btree_node_get_noiter(struct btree_trans *trans,
1023 const struct bkey_i *k,
1024 enum btree_id btree_id,
1028 struct bch_fs *c = trans->c;
1029 struct btree_cache *bc = &c->btree_cache;
1031 struct bset_tree *t;
1034 EBUG_ON(level >= BTREE_MAX_DEPTH);
1036 if (c->opts.btree_node_mem_ptr_optimization) {
1037 b = btree_node_mem_ptr(k);
1042 b = btree_cache_find(bc, k);
1047 b = bch2_btree_node_fill(trans, NULL, k, btree_id,
1048 level, SIX_LOCK_read, true);
1050 /* We raced and found the btree node in the cache */
1055 !bch2_btree_cache_cannibalize_lock(trans, NULL))
1062 ret = btree_node_lock_nopath(trans, &b->c, SIX_LOCK_read, _THIS_IP_);
1063 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
1064 return ERR_PTR(ret);
1068 if (unlikely(b->hash_val != btree_ptr_hash_val(k) ||
1069 b->c.btree_id != btree_id ||
1070 b->c.level != level)) {
1071 six_unlock_read(&b->c.lock);
1076 /* XXX: waiting on IO with btree locks held: */
1077 __bch2_btree_node_wait_on_read(b);
1079 prefetch(b->aux_data);
1081 for_each_bset(b, t) {
1082 void *p = (u64 *) b->aux_data + t->aux_data_offset;
1084 prefetch(p + L1_CACHE_BYTES * 0);
1085 prefetch(p + L1_CACHE_BYTES * 1);
1086 prefetch(p + L1_CACHE_BYTES * 2);
1089 /* avoid atomic set bit if it's not needed: */
1090 if (!btree_node_accessed(b))
1091 set_btree_node_accessed(b);
1093 if (unlikely(btree_node_read_error(b))) {
1094 six_unlock_read(&b->c.lock);
1095 b = ERR_PTR(-BCH_ERR_btree_node_read_error);
1099 EBUG_ON(b->c.btree_id != btree_id);
1100 EBUG_ON(BTREE_NODE_LEVEL(b->data) != level);
1101 btree_check_header(c, b);
1103 bch2_btree_cache_cannibalize_unlock(trans);
1107 int bch2_btree_node_prefetch(struct btree_trans *trans,
1108 struct btree_path *path,
1109 const struct bkey_i *k,
1110 enum btree_id btree_id, unsigned level)
1112 struct bch_fs *c = trans->c;
1113 struct btree_cache *bc = &c->btree_cache;
1116 BUG_ON(path && !btree_node_locked(path, level + 1));
1117 BUG_ON(level >= BTREE_MAX_DEPTH);
1119 b = btree_cache_find(bc, k);
1123 b = bch2_btree_node_fill(trans, path, k, btree_id,
1124 level, SIX_LOCK_read, false);
1125 return PTR_ERR_OR_ZERO(b);
1128 void bch2_btree_node_evict(struct btree_trans *trans, const struct bkey_i *k)
1130 struct bch_fs *c = trans->c;
1131 struct btree_cache *bc = &c->btree_cache;
1134 b = btree_cache_find(bc, k);
1138 /* not allowed to wait on io with btree locks held: */
1140 /* XXX we're called from btree_gc which will be holding other btree
1143 __bch2_btree_node_wait_on_read(b);
1144 __bch2_btree_node_wait_on_write(b);
1146 btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_intent);
1147 btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_write);
1149 if (btree_node_dirty(b)) {
1150 __bch2_btree_node_write(c, b, BTREE_WRITE_cache_reclaim);
1151 six_unlock_write(&b->c.lock);
1152 six_unlock_intent(&b->c.lock);
1156 BUG_ON(btree_node_dirty(b));
1158 mutex_lock(&bc->lock);
1159 btree_node_data_free(c, b);
1160 bch2_btree_node_hash_remove(bc, b);
1161 mutex_unlock(&bc->lock);
1163 six_unlock_write(&b->c.lock);
1164 six_unlock_intent(&b->c.lock);
1167 const char *bch2_btree_id_str(enum btree_id btree)
1169 return btree < BTREE_ID_NR ? __bch2_btree_ids[btree] : "(unknown)";
1172 void bch2_btree_pos_to_text(struct printbuf *out, struct bch_fs *c, const struct btree *b)
1174 prt_printf(out, "%s level %u/%u\n ",
1175 bch2_btree_id_str(b->c.btree_id),
1177 bch2_btree_id_root(c, b->c.btree_id)->level);
1178 bch2_bkey_val_to_text(out, c, bkey_i_to_s_c(&b->key));
1181 void bch2_btree_node_to_text(struct printbuf *out, struct bch_fs *c, const struct btree *b)
1183 struct bset_stats stats;
1185 memset(&stats, 0, sizeof(stats));
1187 bch2_btree_keys_stats(b, &stats);
1189 prt_printf(out, "l %u ", b->c.level);
1190 bch2_bpos_to_text(out, b->data->min_key);
1191 prt_printf(out, " - ");
1192 bch2_bpos_to_text(out, b->data->max_key);
1193 prt_printf(out, ":\n"
1195 bch2_val_to_text(out, c, bkey_i_to_s_c(&b->key));
1200 bch2_bkey_format_to_text(out, &b->format);
1203 " unpack fn len: %u\n"
1204 " bytes used %zu/%zu (%zu%% full)\n"
1205 " sib u64s: %u, %u (merge threshold %u)\n"
1206 " nr packed keys %u\n"
1207 " nr unpacked keys %u\n"
1209 " failed unpacked %zu\n",
1211 b->nr.live_u64s * sizeof(u64),
1212 btree_buf_bytes(b) - sizeof(struct btree_node),
1213 b->nr.live_u64s * 100 / btree_max_u64s(c),
1216 c->btree_foreground_merge_threshold,
1218 b->nr.unpacked_keys,
1223 void bch2_btree_cache_to_text(struct printbuf *out, const struct bch_fs *c)
1225 prt_printf(out, "nr nodes:\t\t%u\n", c->btree_cache.used);
1226 prt_printf(out, "nr dirty:\t\t%u\n", atomic_read(&c->btree_cache.dirty));
1227 prt_printf(out, "cannibalize lock:\t%p\n", c->btree_cache.alloc_lock);