1 // SPDX-License-Identifier: GPL-2.0
5 #include "btree_cache.h"
7 #include "btree_iter.h"
8 #include "btree_locking.h"
13 #include <linux/prefetch.h>
14 #include <linux/sched/mm.h>
15 #include <trace/events/bcachefs.h>
17 const char * const bch2_btree_node_flags[] = {
24 void bch2_recalc_btree_reserve(struct bch_fs *c)
26 unsigned i, reserve = 16;
28 if (!c->btree_roots[0].b)
31 for (i = 0; i < BTREE_ID_NR; i++)
32 if (c->btree_roots[i].b)
33 reserve += min_t(unsigned, 1,
34 c->btree_roots[i].b->c.level) * 8;
36 c->btree_cache.reserve = reserve;
39 static inline unsigned btree_cache_can_free(struct btree_cache *bc)
41 return max_t(int, 0, bc->used - bc->reserve);
44 static void btree_node_to_freedlist(struct btree_cache *bc, struct btree *b)
46 if (b->c.lock.readers)
47 list_move(&b->list, &bc->freed_pcpu);
49 list_move(&b->list, &bc->freed_nonpcpu);
52 static void btree_node_data_free(struct bch_fs *c, struct btree *b)
54 struct btree_cache *bc = &c->btree_cache;
56 EBUG_ON(btree_node_write_in_flight(b));
58 kvpfree(b->data, btree_bytes(c));
63 munmap(b->aux_data, btree_aux_data_bytes(b));
69 btree_node_to_freedlist(bc, b);
72 static int bch2_btree_cache_cmp_fn(struct rhashtable_compare_arg *arg,
75 const struct btree *b = obj;
76 const u64 *v = arg->key;
78 return b->hash_val == *v ? 0 : 1;
81 static const struct rhashtable_params bch_btree_cache_params = {
82 .head_offset = offsetof(struct btree, hash),
83 .key_offset = offsetof(struct btree, hash_val),
84 .key_len = sizeof(u64),
85 .obj_cmpfn = bch2_btree_cache_cmp_fn,
88 static int btree_node_data_alloc(struct bch_fs *c, struct btree *b, gfp_t gfp)
90 BUG_ON(b->data || b->aux_data);
92 b->data = kvpmalloc(btree_bytes(c), gfp);
96 b->aux_data = vmalloc_exec(btree_aux_data_bytes(b), gfp);
98 b->aux_data = mmap(NULL, btree_aux_data_bytes(b),
99 PROT_READ|PROT_WRITE|PROT_EXEC,
100 MAP_PRIVATE|MAP_ANONYMOUS, 0, 0);
101 if (b->aux_data == MAP_FAILED)
105 kvpfree(b->data, btree_bytes(c));
113 static struct btree *__btree_node_mem_alloc(struct bch_fs *c, gfp_t gfp)
115 struct btree *b = kzalloc(sizeof(struct btree), gfp);
119 bkey_btree_ptr_init(&b->key);
120 __six_lock_init(&b->c.lock, "b->c.lock", &bch2_btree_node_lock_key);
121 #ifdef CONFIG_DEBUG_LOCK_ALLOC
122 lockdep_set_no_check_recursion(&b->c.lock.dep_map);
124 INIT_LIST_HEAD(&b->list);
125 INIT_LIST_HEAD(&b->write_blocked);
126 b->byte_order = ilog2(btree_bytes(c));
130 struct btree *__bch2_btree_node_mem_alloc(struct bch_fs *c)
132 struct btree_cache *bc = &c->btree_cache;
133 struct btree *b = __btree_node_mem_alloc(c, GFP_KERNEL);
137 if (btree_node_data_alloc(c, b, GFP_KERNEL)) {
143 list_add(&b->list, &bc->freeable);
147 /* Btree in memory cache - hash table */
149 void bch2_btree_node_hash_remove(struct btree_cache *bc, struct btree *b)
151 int ret = rhashtable_remove_fast(&bc->table, &b->hash, bch_btree_cache_params);
154 /* Cause future lookups for this node to fail: */
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_and_count(c, btree_cache_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 = 0;
285 unsigned long freed = 0;
286 unsigned long touched = 0;
288 unsigned long ret = SHRINK_STOP;
289 bool trigger_writes = atomic_read(&bc->dirty) + nr >=
292 if (bch2_btree_shrinker_disabled)
295 mutex_lock(&bc->lock);
296 flags = memalloc_nofs_save();
299 * It's _really_ critical that we don't free too many btree nodes - we
300 * have to always leave ourselves a reserve. The reserve is how we
301 * guarantee that allocating memory for a new btree node can always
302 * succeed, so that inserting keys into the btree can always succeed and
303 * IO can always make forward progress:
305 can_free = btree_cache_can_free(bc);
306 nr = min_t(unsigned long, nr, can_free);
309 list_for_each_entry_safe(b, t, &bc->freeable, list) {
311 * Leave a few nodes on the freeable list, so that a btree split
312 * won't have to hit the system allocator:
322 if (!btree_node_reclaim(c, b)) {
323 btree_node_data_free(c, b);
324 six_unlock_write(&b->c.lock);
325 six_unlock_intent(&b->c.lock);
330 list_for_each_entry_safe(b, t, &bc->live, list) {
333 if (btree_node_accessed(b)) {
334 clear_btree_node_accessed(b);
335 } else if (!btree_node_reclaim(c, b)) {
337 btree_node_data_free(c, b);
339 bch2_btree_node_hash_remove(bc, b);
340 six_unlock_write(&b->c.lock);
341 six_unlock_intent(&b->c.lock);
345 } else if (trigger_writes &&
346 btree_node_dirty(b) &&
347 !btree_node_will_make_reachable(b) &&
348 !btree_node_write_blocked(b) &&
349 six_trylock_read(&b->c.lock)) {
350 list_move(&bc->live, &b->list);
351 mutex_unlock(&bc->lock);
352 __bch2_btree_node_write(c, b, 0);
353 six_unlock_read(&b->c.lock);
356 mutex_lock(&bc->lock);
364 if (&t->list != &bc->live)
365 list_move_tail(&bc->live, &t->list);
367 mutex_unlock(&bc->lock);
370 memalloc_nofs_restore(flags);
371 trace_and_count(c, btree_cache_scan, sc->nr_to_scan, can_free, ret);
375 static unsigned long bch2_btree_cache_count(struct shrinker *shrink,
376 struct shrink_control *sc)
378 struct bch_fs *c = container_of(shrink, struct bch_fs,
380 struct btree_cache *bc = &c->btree_cache;
382 if (bch2_btree_shrinker_disabled)
385 return btree_cache_can_free(bc);
388 static void bch2_btree_cache_shrinker_to_text(struct printbuf *out, struct shrinker *shrink)
390 struct bch_fs *c = container_of(shrink, struct bch_fs,
393 bch2_btree_cache_to_text(out, c);
396 void bch2_fs_btree_cache_exit(struct bch_fs *c)
398 struct btree_cache *bc = &c->btree_cache;
402 if (bc->shrink.list.next)
403 unregister_shrinker(&bc->shrink);
405 /* vfree() can allocate memory: */
406 flags = memalloc_nofs_save();
407 mutex_lock(&bc->lock);
410 list_move(&c->verify_data->list, &bc->live);
412 kvpfree(c->verify_ondisk, btree_bytes(c));
414 for (i = 0; i < BTREE_ID_NR; i++)
415 if (c->btree_roots[i].b)
416 list_add(&c->btree_roots[i].b->list, &bc->live);
418 list_splice(&bc->freeable, &bc->live);
420 while (!list_empty(&bc->live)) {
421 b = list_first_entry(&bc->live, struct btree, list);
423 BUG_ON(btree_node_read_in_flight(b) ||
424 btree_node_write_in_flight(b));
426 if (btree_node_dirty(b))
427 bch2_btree_complete_write(c, b, btree_current_write(b));
428 clear_btree_node_dirty_acct(c, b);
430 btree_node_data_free(c, b);
433 BUG_ON(atomic_read(&c->btree_cache.dirty));
435 list_splice(&bc->freed_pcpu, &bc->freed_nonpcpu);
437 while (!list_empty(&bc->freed_nonpcpu)) {
438 b = list_first_entry(&bc->freed_nonpcpu, struct btree, list);
440 six_lock_pcpu_free(&b->c.lock);
444 mutex_unlock(&bc->lock);
445 memalloc_nofs_restore(flags);
447 if (bc->table_init_done)
448 rhashtable_destroy(&bc->table);
451 int bch2_fs_btree_cache_init(struct bch_fs *c)
453 struct btree_cache *bc = &c->btree_cache;
457 pr_verbose_init(c->opts, "");
459 ret = rhashtable_init(&bc->table, &bch_btree_cache_params);
463 bc->table_init_done = true;
465 bch2_recalc_btree_reserve(c);
467 for (i = 0; i < bc->reserve; i++)
468 if (!__bch2_btree_node_mem_alloc(c)) {
473 list_splice_init(&bc->live, &bc->freeable);
475 mutex_init(&c->verify_lock);
477 bc->shrink.count_objects = bch2_btree_cache_count;
478 bc->shrink.scan_objects = bch2_btree_cache_scan;
479 bc->shrink.to_text = bch2_btree_cache_shrinker_to_text;
480 bc->shrink.seeks = 4;
481 ret = register_shrinker(&bc->shrink, "%s/btree_cache", c->name);
483 pr_verbose_init(c->opts, "ret %i", ret);
487 void bch2_fs_btree_cache_init_early(struct btree_cache *bc)
489 mutex_init(&bc->lock);
490 INIT_LIST_HEAD(&bc->live);
491 INIT_LIST_HEAD(&bc->freeable);
492 INIT_LIST_HEAD(&bc->freed_pcpu);
493 INIT_LIST_HEAD(&bc->freed_nonpcpu);
497 * We can only have one thread cannibalizing other cached btree nodes at a time,
498 * or we'll deadlock. We use an open coded mutex to ensure that, which a
499 * cannibalize_bucket() will take. This means every time we unlock the root of
500 * the btree, we need to release this lock if we have it held.
502 void bch2_btree_cache_cannibalize_unlock(struct bch_fs *c)
504 struct btree_cache *bc = &c->btree_cache;
506 if (bc->alloc_lock == current) {
507 trace_and_count(c, btree_cache_cannibalize_unlock, c);
508 bc->alloc_lock = NULL;
509 closure_wake_up(&bc->alloc_wait);
513 int bch2_btree_cache_cannibalize_lock(struct bch_fs *c, struct closure *cl)
515 struct btree_cache *bc = &c->btree_cache;
516 struct task_struct *old;
518 old = cmpxchg(&bc->alloc_lock, NULL, current);
519 if (old == NULL || old == current)
523 trace_and_count(c, btree_cache_cannibalize_lock_fail, c);
527 closure_wait(&bc->alloc_wait, cl);
529 /* Try again, after adding ourselves to waitlist */
530 old = cmpxchg(&bc->alloc_lock, NULL, current);
531 if (old == NULL || old == current) {
533 closure_wake_up(&bc->alloc_wait);
537 trace_and_count(c, btree_cache_cannibalize_lock_fail, c);
541 trace_and_count(c, btree_cache_cannibalize_lock, c);
545 static struct btree *btree_node_cannibalize(struct bch_fs *c)
547 struct btree_cache *bc = &c->btree_cache;
550 list_for_each_entry_reverse(b, &bc->live, list)
551 if (!btree_node_reclaim(c, b))
555 list_for_each_entry_reverse(b, &bc->live, list)
556 if (!btree_node_write_and_reclaim(c, b))
560 * Rare case: all nodes were intent-locked.
563 WARN_ONCE(1, "btree cache cannibalize failed\n");
568 struct btree *bch2_btree_node_mem_alloc(struct bch_fs *c, bool pcpu_read_locks)
570 struct btree_cache *bc = &c->btree_cache;
571 struct list_head *freed = pcpu_read_locks
573 : &bc->freed_nonpcpu;
574 struct btree *b, *b2;
575 u64 start_time = local_clock();
578 flags = memalloc_nofs_save();
579 mutex_lock(&bc->lock);
582 * We never free struct btree itself, just the memory that holds the on
583 * disk node. Check the freed list before allocating a new one:
585 list_for_each_entry(b, freed, list)
586 if (!btree_node_reclaim(c, b)) {
587 list_del_init(&b->list);
591 b = __btree_node_mem_alloc(c, __GFP_NOWARN);
593 mutex_unlock(&bc->lock);
594 b = __btree_node_mem_alloc(c, GFP_KERNEL);
597 mutex_lock(&bc->lock);
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_and_count(c, btree_cache_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_and_count(c, trans_restart_relock_parent_for_fill, trans, _THIS_IP_, path);
700 return ERR_PTR(btree_trans_restart(trans, BCH_ERR_transaction_restart_fill_relock));
703 b = bch2_btree_node_mem_alloc(c, level != 0);
705 if (trans && b == ERR_PTR(-ENOMEM)) {
706 trans->memory_allocation_failure = true;
707 trace_and_count(c, trans_restart_memory_allocation_failure, trans, _THIS_IP_, path);
708 return ERR_PTR(btree_trans_restart(trans, BCH_ERR_transaction_restart_fill_mem_alloc_fail));
714 bkey_copy(&b->key, k);
715 if (bch2_btree_node_hash_insert(bc, b, level, btree_id)) {
716 /* raced with another fill: */
718 /* mark as unhashed... */
721 mutex_lock(&bc->lock);
722 list_add(&b->list, &bc->freeable);
723 mutex_unlock(&bc->lock);
725 six_unlock_write(&b->c.lock);
726 six_unlock_intent(&b->c.lock);
730 set_btree_node_read_in_flight(b);
732 six_unlock_write(&b->c.lock);
733 seq = b->c.lock.state.seq;
734 six_unlock_intent(&b->c.lock);
736 /* Unlock before doing IO: */
738 bch2_trans_unlock(trans);
740 bch2_btree_node_read(c, b, sync);
746 int ret = bch2_trans_relock(trans) ?:
747 bch2_btree_path_relock_intent(trans, path);
749 BUG_ON(!trans->restarted);
754 if (!six_relock_type(&b->c.lock, lock_type, seq)) {
756 trace_and_count(c, trans_restart_relock_after_fill, trans, _THIS_IP_, path);
757 return ERR_PTR(btree_trans_restart(trans, BCH_ERR_transaction_restart_relock_after_fill));
763 static noinline void btree_bad_header(struct bch_fs *c, struct btree *b)
765 struct printbuf buf = PRINTBUF;
767 if (!test_bit(BCH_FS_INITIAL_GC_DONE, &c->flags))
771 "btree node header doesn't match ptr\n"
772 "btree %s level %u\n"
774 bch2_btree_ids[b->c.btree_id], b->c.level);
775 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(&b->key));
777 prt_printf(&buf, "\nheader: btree %s level %llu\n"
779 bch2_btree_ids[BTREE_NODE_ID(b->data)],
780 BTREE_NODE_LEVEL(b->data));
781 bch2_bpos_to_text(&buf, b->data->min_key);
783 prt_printf(&buf, "\nmax ");
784 bch2_bpos_to_text(&buf, b->data->max_key);
786 bch2_fs_inconsistent(c, "%s", buf.buf);
790 static inline void btree_check_header(struct bch_fs *c, struct btree *b)
792 if (b->c.btree_id != BTREE_NODE_ID(b->data) ||
793 b->c.level != BTREE_NODE_LEVEL(b->data) ||
794 bpos_cmp(b->data->max_key, b->key.k.p) ||
795 (b->key.k.type == KEY_TYPE_btree_ptr_v2 &&
796 bpos_cmp(b->data->min_key,
797 bkey_i_to_btree_ptr_v2(&b->key)->v.min_key)))
798 btree_bad_header(c, b);
802 * bch_btree_node_get - find a btree node in the cache and lock it, reading it
803 * in from disk if necessary.
805 * If IO is necessary and running under generic_make_request, returns -EAGAIN.
807 * The btree node will have either a read or a write lock held, depending on
808 * the @write parameter.
810 struct btree *bch2_btree_node_get(struct btree_trans *trans, struct btree_path *path,
811 const struct bkey_i *k, unsigned level,
812 enum six_lock_type lock_type,
813 unsigned long trace_ip)
815 struct bch_fs *c = trans->c;
816 struct btree_cache *bc = &c->btree_cache;
821 EBUG_ON(level >= BTREE_MAX_DEPTH);
823 b = btree_node_mem_ptr(k);
826 * Check b->hash_val _before_ calling btree_node_lock() - this might not
827 * be the node we want anymore, and trying to lock the wrong node could
828 * cause an unneccessary transaction restart:
830 if (likely(c->opts.btree_node_mem_ptr_optimization &&
832 b->hash_val == btree_ptr_hash_val(k)))
835 b = btree_cache_find(bc, k);
838 * We must have the parent locked to call bch2_btree_node_fill(),
839 * else we could read in a btree node from disk that's been
842 b = bch2_btree_node_fill(c, trans, path, k, path->btree_id,
843 level, lock_type, true);
845 /* We raced and found the btree node in the cache */
854 * There's a potential deadlock with splits and insertions into
855 * interior nodes we have to avoid:
857 * The other thread might be holding an intent lock on the node
858 * we want, and they want to update its parent node so they're
859 * going to upgrade their intent lock on the parent node to a
862 * But if we're holding a read lock on the parent, and we're
863 * trying to get the intent lock they're holding, we deadlock.
865 * So to avoid this we drop the read locks on parent nodes when
866 * we're starting to take intent locks - and handle the race.
868 * The race is that they might be about to free the node we
869 * want, and dropping our read lock on the parent node lets them
870 * update the parent marking the node we want as freed, and then
873 * To guard against this, btree nodes are evicted from the cache
874 * when they're freed - and b->hash_val is zeroed out, which we
875 * check for after we lock the node.
877 * Then, bch2_btree_node_relock() on the parent will fail - because
878 * the parent was modified, when the pointer to the node we want
879 * was removed - and we'll bail out:
881 if (btree_node_read_locked(path, level + 1))
882 btree_node_unlock(trans, path, level + 1);
884 ret = btree_node_lock(trans, path, &b->c, level, lock_type, trace_ip);
885 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
890 if (unlikely(b->hash_val != btree_ptr_hash_val(k) ||
891 b->c.level != level ||
893 six_unlock_type(&b->c.lock, lock_type);
894 if (bch2_btree_node_relock(trans, path, level + 1))
897 trace_and_count(c, trans_restart_btree_node_reused, trans, trace_ip, path);
898 return ERR_PTR(btree_trans_restart(trans, BCH_ERR_transaction_restart_lock_node_reused));
902 if (unlikely(btree_node_read_in_flight(b))) {
903 u32 seq = b->c.lock.state.seq;
905 six_unlock_type(&b->c.lock, lock_type);
906 bch2_trans_unlock(trans);
908 bch2_btree_node_wait_on_read(b);
911 * should_be_locked is not set on this path yet, so we need to
912 * relock it specifically:
915 int ret = bch2_trans_relock(trans) ?:
916 bch2_btree_path_relock_intent(trans, path);
918 BUG_ON(!trans->restarted);
923 if (!six_relock_type(&b->c.lock, lock_type, seq))
927 prefetch(b->aux_data);
929 for_each_bset(b, t) {
930 void *p = (u64 *) b->aux_data + t->aux_data_offset;
932 prefetch(p + L1_CACHE_BYTES * 0);
933 prefetch(p + L1_CACHE_BYTES * 1);
934 prefetch(p + L1_CACHE_BYTES * 2);
937 /* avoid atomic set bit if it's not needed: */
938 if (!btree_node_accessed(b))
939 set_btree_node_accessed(b);
941 if (unlikely(btree_node_read_error(b))) {
942 six_unlock_type(&b->c.lock, lock_type);
943 return ERR_PTR(-EIO);
946 EBUG_ON(b->c.btree_id != path->btree_id);
947 EBUG_ON(BTREE_NODE_LEVEL(b->data) != level);
948 btree_check_header(c, b);
953 struct btree *bch2_btree_node_get_noiter(struct btree_trans *trans,
954 const struct bkey_i *k,
955 enum btree_id btree_id,
959 struct bch_fs *c = trans->c;
960 struct btree_cache *bc = &c->btree_cache;
965 EBUG_ON(level >= BTREE_MAX_DEPTH);
967 if (c->opts.btree_node_mem_ptr_optimization) {
968 b = btree_node_mem_ptr(k);
973 b = btree_cache_find(bc, k);
978 b = bch2_btree_node_fill(c, NULL, NULL, k, btree_id,
979 level, SIX_LOCK_read, true);
981 /* We raced and found the btree node in the cache */
986 !bch2_btree_cache_cannibalize_lock(c, NULL))
993 ret = btree_node_lock_nopath(trans, &b->c, SIX_LOCK_read);
994 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
999 if (unlikely(b->hash_val != btree_ptr_hash_val(k) ||
1000 b->c.btree_id != btree_id ||
1001 b->c.level != level)) {
1002 six_unlock_read(&b->c.lock);
1007 /* XXX: waiting on IO with btree locks held: */
1008 __bch2_btree_node_wait_on_read(b);
1010 prefetch(b->aux_data);
1012 for_each_bset(b, t) {
1013 void *p = (u64 *) b->aux_data + t->aux_data_offset;
1015 prefetch(p + L1_CACHE_BYTES * 0);
1016 prefetch(p + L1_CACHE_BYTES * 1);
1017 prefetch(p + L1_CACHE_BYTES * 2);
1020 /* avoid atomic set bit if it's not needed: */
1021 if (!btree_node_accessed(b))
1022 set_btree_node_accessed(b);
1024 if (unlikely(btree_node_read_error(b))) {
1025 six_unlock_read(&b->c.lock);
1030 EBUG_ON(b->c.btree_id != btree_id);
1031 EBUG_ON(BTREE_NODE_LEVEL(b->data) != level);
1032 btree_check_header(c, b);
1034 bch2_btree_cache_cannibalize_unlock(c);
1038 int bch2_btree_node_prefetch(struct bch_fs *c,
1039 struct btree_trans *trans,
1040 struct btree_path *path,
1041 const struct bkey_i *k,
1042 enum btree_id btree_id, unsigned level)
1044 struct btree_cache *bc = &c->btree_cache;
1047 BUG_ON(trans && !btree_node_locked(path, level + 1));
1048 BUG_ON(level >= BTREE_MAX_DEPTH);
1050 b = btree_cache_find(bc, k);
1054 b = bch2_btree_node_fill(c, trans, path, k, btree_id,
1055 level, SIX_LOCK_read, false);
1056 return PTR_ERR_OR_ZERO(b);
1059 void bch2_btree_node_evict(struct btree_trans *trans, const struct bkey_i *k)
1061 struct bch_fs *c = trans->c;
1062 struct btree_cache *bc = &c->btree_cache;
1065 b = btree_cache_find(bc, k);
1069 /* not allowed to wait on io with btree locks held: */
1071 /* XXX we're called from btree_gc which will be holding other btree
1074 __bch2_btree_node_wait_on_read(b);
1075 __bch2_btree_node_wait_on_write(b);
1077 btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_intent);
1078 btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_write);
1080 if (btree_node_dirty(b)) {
1081 __bch2_btree_node_write(c, b, 0);
1082 six_unlock_write(&b->c.lock);
1083 six_unlock_intent(&b->c.lock);
1087 BUG_ON(btree_node_dirty(b));
1089 mutex_lock(&bc->lock);
1090 btree_node_data_free(c, b);
1091 bch2_btree_node_hash_remove(bc, b);
1092 mutex_unlock(&bc->lock);
1094 six_unlock_write(&b->c.lock);
1095 six_unlock_intent(&b->c.lock);
1098 void bch2_btree_node_to_text(struct printbuf *out, struct bch_fs *c,
1101 const struct bkey_format *f = &b->format;
1102 struct bset_stats stats;
1104 memset(&stats, 0, sizeof(stats));
1106 bch2_btree_keys_stats(b, &stats);
1108 prt_printf(out, "l %u ", b->c.level);
1109 bch2_bpos_to_text(out, b->data->min_key);
1110 prt_printf(out, " - ");
1111 bch2_bpos_to_text(out, b->data->max_key);
1112 prt_printf(out, ":\n"
1114 bch2_val_to_text(out, c, bkey_i_to_s_c(&b->key));
1116 prt_printf(out, "\n"
1117 " format: u64s %u fields %u %u %u %u %u\n"
1118 " unpack fn len: %u\n"
1119 " bytes used %zu/%zu (%zu%% full)\n"
1120 " sib u64s: %u, %u (merge threshold %u)\n"
1121 " nr packed keys %u\n"
1122 " nr unpacked keys %u\n"
1124 " failed unpacked %zu\n",
1126 f->bits_per_field[0],
1127 f->bits_per_field[1],
1128 f->bits_per_field[2],
1129 f->bits_per_field[3],
1130 f->bits_per_field[4],
1132 b->nr.live_u64s * sizeof(u64),
1133 btree_bytes(c) - sizeof(struct btree_node),
1134 b->nr.live_u64s * 100 / btree_max_u64s(c),
1137 c->btree_foreground_merge_threshold,
1139 b->nr.unpacked_keys,
1144 void bch2_btree_cache_to_text(struct printbuf *out, struct bch_fs *c)
1146 prt_printf(out, "nr nodes:\t\t%u\n", c->btree_cache.used);
1147 prt_printf(out, "nr dirty:\t\t%u\n", atomic_read(&c->btree_cache.dirty));
1148 prt_printf(out, "cannibalize lock:\t%p\n", c->btree_cache.alloc_lock);