1 // SPDX-License-Identifier: GPL-2.0
4 #include "alloc_foreground.h"
5 #include "bkey_methods.h"
6 #include "btree_cache.h"
8 #include "btree_update.h"
9 #include "btree_update_interior.h"
11 #include "btree_iter.h"
12 #include "btree_locking.h"
16 #include "journal_reclaim.h"
21 #include <linux/random.h>
22 #include <trace/events/bcachefs.h>
27 * Verify that child nodes correctly span parent node's range:
29 static void btree_node_interior_verify(struct bch_fs *c, struct btree *b)
31 #ifdef CONFIG_BCACHEFS_DEBUG
32 struct bpos next_node = b->data->min_key;
33 struct btree_node_iter iter;
35 struct bkey_s_c_btree_ptr_v2 bp;
40 if (!test_bit(BCH_FS_BTREE_INTERIOR_REPLAY_DONE, &c->flags))
43 bch2_btree_node_iter_init_from_start(&iter, b);
46 k = bch2_btree_node_iter_peek_unpack(&iter, b, &unpacked);
47 if (k.k->type != KEY_TYPE_btree_ptr_v2)
49 bp = bkey_s_c_to_btree_ptr_v2(k);
51 BUG_ON(bkey_cmp(next_node, bp.v->min_key));
53 bch2_btree_node_iter_advance(&iter, b);
55 if (bch2_btree_node_iter_end(&iter)) {
56 BUG_ON(bkey_cmp(k.k->p, b->key.k.p));
60 next_node = bkey_successor(k.k->p);
65 /* Calculate ideal packed bkey format for new btree nodes: */
67 void __bch2_btree_calc_format(struct bkey_format_state *s, struct btree *b)
69 struct bkey_packed *k;
73 bch2_bkey_format_add_pos(s, b->data->min_key);
76 bset_tree_for_each_key(b, t, k)
77 if (!bkey_whiteout(k)) {
78 uk = bkey_unpack_key(b, k);
79 bch2_bkey_format_add_key(s, &uk);
83 static struct bkey_format bch2_btree_calc_format(struct btree *b)
85 struct bkey_format_state s;
87 bch2_bkey_format_init(&s);
88 __bch2_btree_calc_format(&s, b);
90 return bch2_bkey_format_done(&s);
93 static size_t btree_node_u64s_with_format(struct btree *b,
94 struct bkey_format *new_f)
96 struct bkey_format *old_f = &b->format;
98 /* stupid integer promotion rules */
100 (((int) new_f->key_u64s - old_f->key_u64s) *
101 (int) b->nr.packed_keys) +
102 (((int) new_f->key_u64s - BKEY_U64s) *
103 (int) b->nr.unpacked_keys);
105 BUG_ON(delta + b->nr.live_u64s < 0);
107 return b->nr.live_u64s + delta;
111 * btree_node_format_fits - check if we could rewrite node with a new format
113 * This assumes all keys can pack with the new format -- it just checks if
114 * the re-packed keys would fit inside the node itself.
116 bool bch2_btree_node_format_fits(struct bch_fs *c, struct btree *b,
117 struct bkey_format *new_f)
119 size_t u64s = btree_node_u64s_with_format(b, new_f);
121 return __vstruct_bytes(struct btree_node, u64s) < btree_bytes(c);
124 /* Btree node freeing/allocation: */
126 static void __btree_node_free(struct bch_fs *c, struct btree *b)
128 trace_btree_node_free(c, b);
130 BUG_ON(btree_node_dirty(b));
131 BUG_ON(btree_node_need_write(b));
132 BUG_ON(b == btree_node_root(c, b));
134 BUG_ON(!list_empty(&b->write_blocked));
135 BUG_ON(b->will_make_reachable);
137 clear_btree_node_noevict(b);
139 bch2_btree_node_hash_remove(&c->btree_cache, b);
141 mutex_lock(&c->btree_cache.lock);
142 list_move(&b->list, &c->btree_cache.freeable);
143 mutex_unlock(&c->btree_cache.lock);
146 void bch2_btree_node_free_never_inserted(struct bch_fs *c, struct btree *b)
148 struct open_buckets ob = b->ob;
152 clear_btree_node_dirty(b);
154 btree_node_lock_type(c, b, SIX_LOCK_write);
155 __btree_node_free(c, b);
156 six_unlock_write(&b->c.lock);
158 bch2_open_buckets_put(c, &ob);
161 void bch2_btree_node_free_inmem(struct bch_fs *c, struct btree *b,
162 struct btree_iter *iter)
164 struct btree_iter *linked;
166 trans_for_each_iter(iter->trans, linked)
167 BUG_ON(linked->l[b->c.level].b == b);
169 six_lock_write(&b->c.lock, NULL, NULL);
170 __btree_node_free(c, b);
171 six_unlock_write(&b->c.lock);
172 six_unlock_intent(&b->c.lock);
175 static struct btree *__bch2_btree_node_alloc(struct bch_fs *c,
176 struct disk_reservation *res,
180 struct write_point *wp;
183 struct open_buckets ob = { .nr = 0 };
184 struct bch_devs_list devs_have = (struct bch_devs_list) { 0 };
186 enum alloc_reserve alloc_reserve;
188 if (flags & BTREE_INSERT_USE_ALLOC_RESERVE) {
190 alloc_reserve = RESERVE_ALLOC;
191 } else if (flags & BTREE_INSERT_USE_RESERVE) {
192 nr_reserve = BTREE_NODE_RESERVE / 2;
193 alloc_reserve = RESERVE_BTREE;
195 nr_reserve = BTREE_NODE_RESERVE;
196 alloc_reserve = RESERVE_NONE;
199 mutex_lock(&c->btree_reserve_cache_lock);
200 if (c->btree_reserve_cache_nr > nr_reserve) {
201 struct btree_alloc *a =
202 &c->btree_reserve_cache[--c->btree_reserve_cache_nr];
205 bkey_copy(&tmp.k, &a->k);
206 mutex_unlock(&c->btree_reserve_cache_lock);
209 mutex_unlock(&c->btree_reserve_cache_lock);
212 wp = bch2_alloc_sectors_start(c, c->opts.foreground_target, 0,
213 writepoint_ptr(&c->btree_write_point),
216 c->opts.metadata_replicas_required,
217 alloc_reserve, 0, cl);
221 if (wp->sectors_free < c->opts.btree_node_size) {
222 struct open_bucket *ob;
225 open_bucket_for_each(c, &wp->ptrs, ob, i)
226 if (ob->sectors_free < c->opts.btree_node_size)
227 ob->sectors_free = 0;
229 bch2_alloc_sectors_done(c, wp);
233 if (c->sb.features & (1ULL << BCH_FEATURE_btree_ptr_v2))
234 bkey_btree_ptr_v2_init(&tmp.k);
236 bkey_btree_ptr_init(&tmp.k);
238 bch2_alloc_sectors_append_ptrs(c, wp, &tmp.k, c->opts.btree_node_size);
240 bch2_open_bucket_get(c, wp, &ob);
241 bch2_alloc_sectors_done(c, wp);
243 b = bch2_btree_node_mem_alloc(c);
245 /* we hold cannibalize_lock: */
249 bkey_copy(&b->key, &tmp.k);
255 static struct btree *bch2_btree_node_alloc(struct btree_update *as, unsigned level)
257 struct bch_fs *c = as->c;
261 BUG_ON(level >= BTREE_MAX_DEPTH);
262 BUG_ON(!as->nr_prealloc_nodes);
264 b = as->prealloc_nodes[--as->nr_prealloc_nodes];
266 set_btree_node_accessed(b);
267 set_btree_node_dirty(b);
268 set_btree_node_need_write(b);
270 bch2_bset_init_first(b, &b->data->keys);
272 b->c.btree_id = as->btree_id;
274 memset(&b->nr, 0, sizeof(b->nr));
275 b->data->magic = cpu_to_le64(bset_magic(c));
277 SET_BTREE_NODE_ID(b->data, as->btree_id);
278 SET_BTREE_NODE_LEVEL(b->data, level);
279 b->data->ptr = bch2_bkey_ptrs_c(bkey_i_to_s_c(&b->key)).start->ptr;
281 if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
282 struct bkey_i_btree_ptr_v2 *bp = bkey_i_to_btree_ptr_v2(&b->key);
285 bp->v.seq = b->data->keys.seq;
286 bp->v.sectors_written = 0;
287 bp->v.sectors = cpu_to_le16(c->opts.btree_node_size);
290 if (c->sb.features & (1ULL << BCH_FEATURE_new_extent_overwrite))
291 SET_BTREE_NODE_NEW_EXTENT_OVERWRITE(b->data, true);
293 if (btree_node_is_extents(b) &&
294 !BTREE_NODE_NEW_EXTENT_OVERWRITE(b->data)) {
295 set_btree_node_old_extent_overwrite(b);
296 set_btree_node_need_rewrite(b);
299 bch2_btree_build_aux_trees(b);
301 ret = bch2_btree_node_hash_insert(&c->btree_cache, b, level, as->btree_id);
304 trace_btree_node_alloc(c, b);
308 static void btree_set_min(struct btree *b, struct bpos pos)
310 if (b->key.k.type == KEY_TYPE_btree_ptr_v2)
311 bkey_i_to_btree_ptr_v2(&b->key)->v.min_key = pos;
312 b->data->min_key = pos;
315 static void btree_set_max(struct btree *b, struct bpos pos)
318 b->data->max_key = pos;
321 struct btree *__bch2_btree_node_alloc_replacement(struct btree_update *as,
323 struct bkey_format format)
327 n = bch2_btree_node_alloc(as, b->c.level);
329 SET_BTREE_NODE_SEQ(n->data, BTREE_NODE_SEQ(b->data) + 1);
331 btree_set_min(n, b->data->min_key);
332 btree_set_max(n, b->data->max_key);
334 n->data->format = format;
335 btree_node_set_format(n, format);
337 bch2_btree_sort_into(as->c, n, b);
339 btree_node_reset_sib_u64s(n);
341 n->key.k.p = b->key.k.p;
345 static struct btree *bch2_btree_node_alloc_replacement(struct btree_update *as,
348 struct bkey_format new_f = bch2_btree_calc_format(b);
351 * The keys might expand with the new format - if they wouldn't fit in
352 * the btree node anymore, use the old format for now:
354 if (!bch2_btree_node_format_fits(as->c, b, &new_f))
357 return __bch2_btree_node_alloc_replacement(as, b, new_f);
360 static struct btree *__btree_root_alloc(struct btree_update *as, unsigned level)
362 struct btree *b = bch2_btree_node_alloc(as, level);
364 btree_set_min(b, POS_MIN);
365 btree_set_max(b, POS_MAX);
366 b->data->format = bch2_btree_calc_format(b);
368 btree_node_set_format(b, b->data->format);
369 bch2_btree_build_aux_trees(b);
371 bch2_btree_update_add_new_node(as, b);
372 six_unlock_write(&b->c.lock);
377 static void bch2_btree_reserve_put(struct btree_update *as)
379 struct bch_fs *c = as->c;
381 mutex_lock(&c->btree_reserve_cache_lock);
383 while (as->nr_prealloc_nodes) {
384 struct btree *b = as->prealloc_nodes[--as->nr_prealloc_nodes];
386 six_unlock_write(&b->c.lock);
388 if (c->btree_reserve_cache_nr <
389 ARRAY_SIZE(c->btree_reserve_cache)) {
390 struct btree_alloc *a =
391 &c->btree_reserve_cache[c->btree_reserve_cache_nr++];
395 bkey_copy(&a->k, &b->key);
397 bch2_open_buckets_put(c, &b->ob);
400 btree_node_lock_type(c, b, SIX_LOCK_write);
401 __btree_node_free(c, b);
402 six_unlock_write(&b->c.lock);
404 six_unlock_intent(&b->c.lock);
407 mutex_unlock(&c->btree_reserve_cache_lock);
410 static int bch2_btree_reserve_get(struct btree_update *as, unsigned nr_nodes,
411 unsigned flags, struct closure *cl)
413 struct bch_fs *c = as->c;
417 BUG_ON(nr_nodes > BTREE_RESERVE_MAX);
420 * Protects reaping from the btree node cache and using the btree node
421 * open bucket reserve:
423 ret = bch2_btree_cache_cannibalize_lock(c, cl);
427 while (as->nr_prealloc_nodes < nr_nodes) {
428 b = __bch2_btree_node_alloc(c, &as->disk_res,
429 flags & BTREE_INSERT_NOWAIT
436 ret = bch2_mark_bkey_replicas(c, bkey_i_to_s_c(&b->key));
440 as->prealloc_nodes[as->nr_prealloc_nodes++] = b;
443 bch2_btree_cache_cannibalize_unlock(c);
446 bch2_btree_cache_cannibalize_unlock(c);
447 trace_btree_reserve_get_fail(c, nr_nodes, cl);
451 /* Asynchronous interior node update machinery */
453 static void bch2_btree_update_free(struct btree_update *as)
455 struct bch_fs *c = as->c;
457 bch2_journal_preres_put(&c->journal, &as->journal_preres);
459 bch2_journal_pin_drop(&c->journal, &as->journal);
460 bch2_journal_pin_flush(&c->journal, &as->journal);
461 bch2_disk_reservation_put(c, &as->disk_res);
462 bch2_btree_reserve_put(as);
464 mutex_lock(&c->btree_interior_update_lock);
465 list_del(&as->unwritten_list);
467 mutex_unlock(&c->btree_interior_update_lock);
469 closure_debug_destroy(&as->cl);
470 mempool_free(as, &c->btree_interior_update_pool);
472 closure_wake_up(&c->btree_interior_update_wait);
475 static void btree_update_will_delete_key(struct btree_update *as,
478 BUG_ON(bch2_keylist_u64s(&as->old_keys) + k->k.u64s >
479 ARRAY_SIZE(as->_old_keys));
480 bch2_keylist_add(&as->old_keys, k);
483 static void btree_update_will_add_key(struct btree_update *as,
486 BUG_ON(bch2_keylist_u64s(&as->new_keys) + k->k.u64s >
487 ARRAY_SIZE(as->_new_keys));
488 bch2_keylist_add(&as->new_keys, k);
492 * The transactional part of an interior btree node update, where we journal the
493 * update we did to the interior node and update alloc info:
495 static int btree_update_nodes_written_trans(struct btree_trans *trans,
496 struct btree_update *as)
501 trans->extra_journal_entries = (void *) &as->journal_entries[0];
502 trans->extra_journal_entry_u64s = as->journal_u64s;
503 trans->journal_pin = &as->journal;
505 for_each_keylist_key(&as->new_keys, k) {
506 ret = bch2_trans_mark_key(trans, bkey_i_to_s_c(k),
507 0, 0, BTREE_TRIGGER_INSERT);
512 for_each_keylist_key(&as->old_keys, k) {
513 ret = bch2_trans_mark_key(trans, bkey_i_to_s_c(k),
514 0, 0, BTREE_TRIGGER_OVERWRITE);
522 static void btree_update_nodes_written(struct btree_update *as)
524 struct bch_fs *c = as->c;
525 struct btree *b = as->b;
531 * We did an update to a parent node where the pointers we added pointed
532 * to child nodes that weren't written yet: now, the child nodes have
533 * been written so we can write out the update to the interior node.
537 * We can't call into journal reclaim here: we'd block on the journal
538 * reclaim lock, but we may need to release the open buckets we have
539 * pinned in order for other btree updates to make forward progress, and
540 * journal reclaim does btree updates when flushing bkey_cached entries,
541 * which may require allocations as well.
543 ret = bch2_trans_do(c, &as->disk_res, &journal_seq,
545 BTREE_INSERT_USE_RESERVE|
546 BTREE_INSERT_USE_ALLOC_RESERVE|
547 BTREE_INSERT_NOCHECK_RW|
548 BTREE_INSERT_JOURNAL_RECLAIM|
549 BTREE_INSERT_JOURNAL_RESERVED,
550 btree_update_nodes_written_trans(&trans, as));
551 BUG_ON(ret && !bch2_journal_error(&c->journal));
555 * @b is the node we did the final insert into:
557 * On failure to get a journal reservation, we still have to
558 * unblock the write and allow most of the write path to happen
559 * so that shutdown works, but the i->journal_seq mechanism
560 * won't work to prevent the btree write from being visible (we
561 * didn't get a journal sequence number) - instead
562 * __bch2_btree_node_write() doesn't do the actual write if
563 * we're in journal error state:
566 btree_node_lock_type(c, b, SIX_LOCK_intent);
567 btree_node_lock_type(c, b, SIX_LOCK_write);
568 mutex_lock(&c->btree_interior_update_lock);
570 list_del(&as->write_blocked_list);
572 if (!ret && as->b == b) {
573 struct bset *i = btree_bset_last(b);
576 BUG_ON(!btree_node_dirty(b));
578 i->journal_seq = cpu_to_le64(
580 le64_to_cpu(i->journal_seq)));
582 bch2_btree_add_journal_pin(c, b, journal_seq);
585 mutex_unlock(&c->btree_interior_update_lock);
586 six_unlock_write(&b->c.lock);
588 btree_node_write_if_need(c, b, SIX_LOCK_intent);
589 six_unlock_intent(&b->c.lock);
592 bch2_journal_pin_drop(&c->journal, &as->journal);
594 bch2_journal_preres_put(&c->journal, &as->journal_preres);
596 mutex_lock(&c->btree_interior_update_lock);
597 for (i = 0; i < as->nr_new_nodes; i++) {
598 b = as->new_nodes[i];
600 BUG_ON(b->will_make_reachable != (unsigned long) as);
601 b->will_make_reachable = 0;
603 mutex_unlock(&c->btree_interior_update_lock);
605 for (i = 0; i < as->nr_new_nodes; i++) {
606 b = as->new_nodes[i];
608 btree_node_lock_type(c, b, SIX_LOCK_read);
609 btree_node_write_if_need(c, b, SIX_LOCK_read);
610 six_unlock_read(&b->c.lock);
613 for (i = 0; i < as->nr_open_buckets; i++)
614 bch2_open_bucket_put(c, c->open_buckets + as->open_buckets[i]);
616 bch2_btree_update_free(as);
619 static void btree_interior_update_work(struct work_struct *work)
622 container_of(work, struct bch_fs, btree_interior_update_work);
623 struct btree_update *as;
626 mutex_lock(&c->btree_interior_update_lock);
627 as = list_first_entry_or_null(&c->btree_interior_updates_unwritten,
628 struct btree_update, unwritten_list);
629 if (as && !as->nodes_written)
631 mutex_unlock(&c->btree_interior_update_lock);
636 btree_update_nodes_written(as);
640 static void btree_update_set_nodes_written(struct closure *cl)
642 struct btree_update *as = container_of(cl, struct btree_update, cl);
643 struct bch_fs *c = as->c;
645 mutex_lock(&c->btree_interior_update_lock);
646 as->nodes_written = true;
647 mutex_unlock(&c->btree_interior_update_lock);
649 queue_work(c->btree_interior_update_worker, &c->btree_interior_update_work);
653 * We're updating @b with pointers to nodes that haven't finished writing yet:
654 * block @b from being written until @as completes
656 static void btree_update_updated_node(struct btree_update *as, struct btree *b)
658 struct bch_fs *c = as->c;
660 mutex_lock(&c->btree_interior_update_lock);
661 list_add_tail(&as->unwritten_list, &c->btree_interior_updates_unwritten);
663 BUG_ON(as->mode != BTREE_INTERIOR_NO_UPDATE);
664 BUG_ON(!btree_node_dirty(b));
666 as->mode = BTREE_INTERIOR_UPDATING_NODE;
668 list_add(&as->write_blocked_list, &b->write_blocked);
670 mutex_unlock(&c->btree_interior_update_lock);
673 static void btree_update_reparent(struct btree_update *as,
674 struct btree_update *child)
676 struct bch_fs *c = as->c;
678 lockdep_assert_held(&c->btree_interior_update_lock);
681 child->mode = BTREE_INTERIOR_UPDATING_AS;
684 * When we write a new btree root, we have to drop our journal pin
685 * _before_ the new nodes are technically reachable; see
686 * btree_update_nodes_written().
688 * This goes for journal pins that are recursively blocked on us - so,
689 * just transfer the journal pin to the new interior update so
690 * btree_update_nodes_written() can drop it.
692 bch2_journal_pin_copy(&c->journal, &as->journal, &child->journal, NULL);
693 bch2_journal_pin_drop(&c->journal, &child->journal);
696 static void btree_update_updated_root(struct btree_update *as, struct btree *b)
698 struct bkey_i *insert = &b->key;
699 struct bch_fs *c = as->c;
701 BUG_ON(as->mode != BTREE_INTERIOR_NO_UPDATE);
703 BUG_ON(as->journal_u64s + jset_u64s(insert->k.u64s) >
704 ARRAY_SIZE(as->journal_entries));
707 journal_entry_set((void *) &as->journal_entries[as->journal_u64s],
708 BCH_JSET_ENTRY_btree_root,
709 b->c.btree_id, b->c.level,
710 insert, insert->k.u64s);
712 mutex_lock(&c->btree_interior_update_lock);
713 list_add_tail(&as->unwritten_list, &c->btree_interior_updates_unwritten);
715 as->mode = BTREE_INTERIOR_UPDATING_ROOT;
716 mutex_unlock(&c->btree_interior_update_lock);
720 * bch2_btree_update_add_new_node:
722 * This causes @as to wait on @b to be written, before it gets to
723 * bch2_btree_update_nodes_written
725 * Additionally, it sets b->will_make_reachable to prevent any additional writes
726 * to @b from happening besides the first until @b is reachable on disk
728 * And it adds @b to the list of @as's new nodes, so that we can update sector
729 * counts in bch2_btree_update_nodes_written:
731 void bch2_btree_update_add_new_node(struct btree_update *as, struct btree *b)
733 struct bch_fs *c = as->c;
735 closure_get(&as->cl);
737 mutex_lock(&c->btree_interior_update_lock);
738 BUG_ON(as->nr_new_nodes >= ARRAY_SIZE(as->new_nodes));
739 BUG_ON(b->will_make_reachable);
741 as->new_nodes[as->nr_new_nodes++] = b;
742 b->will_make_reachable = 1UL|(unsigned long) as;
744 mutex_unlock(&c->btree_interior_update_lock);
746 btree_update_will_add_key(as, &b->key);
750 * returns true if @b was a new node
752 static void btree_update_drop_new_node(struct bch_fs *c, struct btree *b)
754 struct btree_update *as;
758 mutex_lock(&c->btree_interior_update_lock);
760 * When b->will_make_reachable != 0, it owns a ref on as->cl that's
761 * dropped when it gets written by bch2_btree_complete_write - the
762 * xchg() is for synchronization with bch2_btree_complete_write:
764 v = xchg(&b->will_make_reachable, 0);
765 as = (struct btree_update *) (v & ~1UL);
768 mutex_unlock(&c->btree_interior_update_lock);
772 for (i = 0; i < as->nr_new_nodes; i++)
773 if (as->new_nodes[i] == b)
778 array_remove_item(as->new_nodes, as->nr_new_nodes, i);
779 mutex_unlock(&c->btree_interior_update_lock);
782 closure_put(&as->cl);
785 void bch2_btree_update_get_open_buckets(struct btree_update *as, struct btree *b)
788 as->open_buckets[as->nr_open_buckets++] =
793 * @b is being split/rewritten: it may have pointers to not-yet-written btree
794 * nodes and thus outstanding btree_updates - redirect @b's
795 * btree_updates to point to this btree_update:
797 void bch2_btree_interior_update_will_free_node(struct btree_update *as,
800 struct bch_fs *c = as->c;
801 struct btree_update *p, *n;
802 struct btree_write *w;
804 set_btree_node_dying(b);
806 if (btree_node_fake(b))
809 mutex_lock(&c->btree_interior_update_lock);
812 * Does this node have any btree_update operations preventing
813 * it from being written?
815 * If so, redirect them to point to this btree_update: we can
816 * write out our new nodes, but we won't make them visible until those
817 * operations complete
819 list_for_each_entry_safe(p, n, &b->write_blocked, write_blocked_list) {
820 list_del_init(&p->write_blocked_list);
821 btree_update_reparent(as, p);
824 * for flush_held_btree_writes() waiting on updates to flush or
825 * nodes to be writeable:
827 closure_wake_up(&c->btree_interior_update_wait);
830 clear_btree_node_dirty(b);
831 clear_btree_node_need_write(b);
834 * Does this node have unwritten data that has a pin on the journal?
836 * If so, transfer that pin to the btree_update operation -
837 * note that if we're freeing multiple nodes, we only need to keep the
838 * oldest pin of any of the nodes we're freeing. We'll release the pin
839 * when the new nodes are persistent and reachable on disk:
841 w = btree_current_write(b);
842 bch2_journal_pin_copy(&c->journal, &as->journal, &w->journal, NULL);
843 bch2_journal_pin_drop(&c->journal, &w->journal);
845 w = btree_prev_write(b);
846 bch2_journal_pin_copy(&c->journal, &as->journal, &w->journal, NULL);
847 bch2_journal_pin_drop(&c->journal, &w->journal);
849 mutex_unlock(&c->btree_interior_update_lock);
852 * Is this a node that isn't reachable on disk yet?
854 * Nodes that aren't reachable yet have writes blocked until they're
855 * reachable - now that we've cancelled any pending writes and moved
856 * things waiting on that write to wait on this update, we can drop this
857 * node from the list of nodes that the other update is making
858 * reachable, prior to freeing it:
860 btree_update_drop_new_node(c, b);
862 btree_update_will_delete_key(as, &b->key);
865 void bch2_btree_update_done(struct btree_update *as)
867 BUG_ON(as->mode == BTREE_INTERIOR_NO_UPDATE);
869 bch2_btree_reserve_put(as);
871 continue_at(&as->cl, btree_update_set_nodes_written, system_freezable_wq);
874 struct btree_update *
875 bch2_btree_update_start(struct btree_trans *trans, enum btree_id id,
876 unsigned nr_nodes, unsigned flags,
879 struct bch_fs *c = trans->c;
880 struct btree_update *as;
881 int disk_res_flags = (flags & BTREE_INSERT_NOFAIL)
882 ? BCH_DISK_RESERVATION_NOFAIL : 0;
883 int journal_flags = (flags & BTREE_INSERT_JOURNAL_RESERVED)
884 ? JOURNAL_RES_GET_RECLAIM : 0;
888 * This check isn't necessary for correctness - it's just to potentially
889 * prevent us from doing a lot of work that'll end up being wasted:
891 ret = bch2_journal_error(&c->journal);
895 as = mempool_alloc(&c->btree_interior_update_pool, GFP_NOIO);
896 memset(as, 0, sizeof(*as));
897 closure_init(&as->cl, NULL);
899 as->mode = BTREE_INTERIOR_NO_UPDATE;
901 INIT_LIST_HEAD(&as->list);
902 INIT_LIST_HEAD(&as->unwritten_list);
903 INIT_LIST_HEAD(&as->write_blocked_list);
904 bch2_keylist_init(&as->old_keys, as->_old_keys);
905 bch2_keylist_init(&as->new_keys, as->_new_keys);
906 bch2_keylist_init(&as->parent_keys, as->inline_keys);
908 ret = bch2_journal_preres_get(&c->journal, &as->journal_preres,
909 BTREE_UPDATE_JOURNAL_RES,
910 journal_flags|JOURNAL_RES_GET_NONBLOCK);
911 if (ret == -EAGAIN) {
912 if (flags & BTREE_INSERT_NOUNLOCK)
913 return ERR_PTR(-EINTR);
915 bch2_trans_unlock(trans);
917 ret = bch2_journal_preres_get(&c->journal, &as->journal_preres,
918 BTREE_UPDATE_JOURNAL_RES,
923 if (!bch2_trans_relock(trans)) {
929 ret = bch2_disk_reservation_get(c, &as->disk_res,
930 nr_nodes * c->opts.btree_node_size,
931 c->opts.metadata_replicas,
936 ret = bch2_btree_reserve_get(as, nr_nodes, flags, cl);
940 mutex_lock(&c->btree_interior_update_lock);
941 list_add_tail(&as->list, &c->btree_interior_update_list);
942 mutex_unlock(&c->btree_interior_update_lock);
946 bch2_btree_update_free(as);
950 /* Btree root updates: */
952 static void bch2_btree_set_root_inmem(struct bch_fs *c, struct btree *b)
954 /* Root nodes cannot be reaped */
955 mutex_lock(&c->btree_cache.lock);
956 list_del_init(&b->list);
957 mutex_unlock(&c->btree_cache.lock);
959 mutex_lock(&c->btree_root_lock);
960 BUG_ON(btree_node_root(c, b) &&
961 (b->c.level < btree_node_root(c, b)->c.level ||
962 !btree_node_dying(btree_node_root(c, b))));
964 btree_node_root(c, b) = b;
965 mutex_unlock(&c->btree_root_lock);
967 bch2_recalc_btree_reserve(c);
971 * bch_btree_set_root - update the root in memory and on disk
973 * To ensure forward progress, the current task must not be holding any
974 * btree node write locks. However, you must hold an intent lock on the
977 * Note: This allocates a journal entry but doesn't add any keys to
978 * it. All the btree roots are part of every journal write, so there
979 * is nothing new to be done. This just guarantees that there is a
982 static void bch2_btree_set_root(struct btree_update *as, struct btree *b,
983 struct btree_iter *iter)
985 struct bch_fs *c = as->c;
988 trace_btree_set_root(c, b);
989 BUG_ON(!b->written &&
990 !test_bit(BCH_FS_HOLD_BTREE_WRITES, &c->flags));
992 old = btree_node_root(c, b);
995 * Ensure no one is using the old root while we switch to the
998 bch2_btree_node_lock_write(old, iter);
1000 bch2_btree_set_root_inmem(c, b);
1002 btree_update_updated_root(as, b);
1005 * Unlock old root after new root is visible:
1007 * The new root isn't persistent, but that's ok: we still have
1008 * an intent lock on the new root, and any updates that would
1009 * depend on the new root would have to update the new root.
1011 bch2_btree_node_unlock_write(old, iter);
1014 /* Interior node updates: */
1016 static void bch2_insert_fixup_btree_ptr(struct btree_update *as, struct btree *b,
1017 struct btree_iter *iter,
1018 struct bkey_i *insert,
1019 struct btree_node_iter *node_iter)
1021 struct bkey_packed *k;
1023 BUG_ON(as->journal_u64s + jset_u64s(insert->k.u64s) >
1024 ARRAY_SIZE(as->journal_entries));
1027 journal_entry_set((void *) &as->journal_entries[as->journal_u64s],
1028 BCH_JSET_ENTRY_btree_keys,
1029 b->c.btree_id, b->c.level,
1030 insert, insert->k.u64s);
1032 while ((k = bch2_btree_node_iter_peek_all(node_iter, b)) &&
1033 bkey_iter_pos_cmp(b, k, &insert->k.p) < 0)
1034 bch2_btree_node_iter_advance(node_iter, b);
1036 bch2_btree_bset_insert_key(iter, b, node_iter, insert);
1037 set_btree_node_dirty(b);
1038 set_btree_node_need_write(b);
1042 * Move keys from n1 (original replacement node, now lower node) to n2 (higher
1045 static struct btree *__btree_split_node(struct btree_update *as,
1047 struct btree_iter *iter)
1049 size_t nr_packed = 0, nr_unpacked = 0;
1051 struct bset *set1, *set2;
1052 struct bkey_packed *k, *prev = NULL;
1054 n2 = bch2_btree_node_alloc(as, n1->c.level);
1055 bch2_btree_update_add_new_node(as, n2);
1057 n2->data->max_key = n1->data->max_key;
1058 n2->data->format = n1->format;
1059 SET_BTREE_NODE_SEQ(n2->data, BTREE_NODE_SEQ(n1->data));
1060 n2->key.k.p = n1->key.k.p;
1062 btree_node_set_format(n2, n2->data->format);
1064 set1 = btree_bset_first(n1);
1065 set2 = btree_bset_first(n2);
1068 * Has to be a linear search because we don't have an auxiliary
1073 struct bkey_packed *n = bkey_next_skip_noops(k, vstruct_last(set1));
1075 if (n == vstruct_last(set1))
1077 if (k->_data - set1->_data >= (le16_to_cpu(set1->u64s) * 3) / 5)
1091 btree_set_max(n1, bkey_unpack_pos(n1, prev));
1092 btree_set_min(n2, bkey_successor(n1->key.k.p));
1094 set2->u64s = cpu_to_le16((u64 *) vstruct_end(set1) - (u64 *) k);
1095 set1->u64s = cpu_to_le16(le16_to_cpu(set1->u64s) - le16_to_cpu(set2->u64s));
1097 set_btree_bset_end(n1, n1->set);
1098 set_btree_bset_end(n2, n2->set);
1100 n2->nr.live_u64s = le16_to_cpu(set2->u64s);
1101 n2->nr.bset_u64s[0] = le16_to_cpu(set2->u64s);
1102 n2->nr.packed_keys = n1->nr.packed_keys - nr_packed;
1103 n2->nr.unpacked_keys = n1->nr.unpacked_keys - nr_unpacked;
1105 n1->nr.live_u64s = le16_to_cpu(set1->u64s);
1106 n1->nr.bset_u64s[0] = le16_to_cpu(set1->u64s);
1107 n1->nr.packed_keys = nr_packed;
1108 n1->nr.unpacked_keys = nr_unpacked;
1110 BUG_ON(!set1->u64s);
1111 BUG_ON(!set2->u64s);
1113 memcpy_u64s(set2->start,
1115 le16_to_cpu(set2->u64s));
1117 btree_node_reset_sib_u64s(n1);
1118 btree_node_reset_sib_u64s(n2);
1120 bch2_verify_btree_nr_keys(n1);
1121 bch2_verify_btree_nr_keys(n2);
1124 btree_node_interior_verify(as->c, n1);
1125 btree_node_interior_verify(as->c, n2);
1132 * For updates to interior nodes, we've got to do the insert before we split
1133 * because the stuff we're inserting has to be inserted atomically. Post split,
1134 * the keys might have to go in different nodes and the split would no longer be
1137 * Worse, if the insert is from btree node coalescing, if we do the insert after
1138 * we do the split (and pick the pivot) - the pivot we pick might be between
1139 * nodes that were coalesced, and thus in the middle of a child node post
1142 static void btree_split_insert_keys(struct btree_update *as, struct btree *b,
1143 struct btree_iter *iter,
1144 struct keylist *keys)
1146 struct btree_node_iter node_iter;
1147 struct bkey_i *k = bch2_keylist_front(keys);
1148 struct bkey_packed *src, *dst, *n;
1151 BUG_ON(btree_node_type(b) != BKEY_TYPE_BTREE);
1153 bch2_btree_node_iter_init(&node_iter, b, &k->k.p);
1155 while (!bch2_keylist_empty(keys)) {
1156 k = bch2_keylist_front(keys);
1158 bch2_insert_fixup_btree_ptr(as, b, iter, k, &node_iter);
1159 bch2_keylist_pop_front(keys);
1163 * We can't tolerate whiteouts here - with whiteouts there can be
1164 * duplicate keys, and it would be rather bad if we picked a duplicate
1167 i = btree_bset_first(b);
1168 src = dst = i->start;
1169 while (src != vstruct_last(i)) {
1170 n = bkey_next_skip_noops(src, vstruct_last(i));
1171 if (!bkey_deleted(src)) {
1172 memmove_u64s_down(dst, src, src->u64s);
1173 dst = bkey_next(dst);
1178 i->u64s = cpu_to_le16((u64 *) dst - i->_data);
1179 set_btree_bset_end(b, b->set);
1181 BUG_ON(b->nsets != 1 ||
1182 b->nr.live_u64s != le16_to_cpu(btree_bset_first(b)->u64s));
1184 btree_node_interior_verify(as->c, b);
1187 static void btree_split(struct btree_update *as, struct btree *b,
1188 struct btree_iter *iter, struct keylist *keys,
1191 struct bch_fs *c = as->c;
1192 struct btree *parent = btree_node_parent(iter, b);
1193 struct btree *n1, *n2 = NULL, *n3 = NULL;
1194 u64 start_time = local_clock();
1196 BUG_ON(!parent && (b != btree_node_root(c, b)));
1197 BUG_ON(!btree_node_intent_locked(iter, btree_node_root(c, b)->c.level));
1199 bch2_btree_interior_update_will_free_node(as, b);
1201 n1 = bch2_btree_node_alloc_replacement(as, b);
1202 bch2_btree_update_add_new_node(as, n1);
1205 btree_split_insert_keys(as, n1, iter, keys);
1207 if (bset_u64s(&n1->set[0]) > BTREE_SPLIT_THRESHOLD(c)) {
1208 trace_btree_split(c, b);
1210 n2 = __btree_split_node(as, n1, iter);
1212 bch2_btree_build_aux_trees(n2);
1213 bch2_btree_build_aux_trees(n1);
1214 six_unlock_write(&n2->c.lock);
1215 six_unlock_write(&n1->c.lock);
1217 bch2_btree_node_write(c, n2, SIX_LOCK_intent);
1220 * Note that on recursive parent_keys == keys, so we
1221 * can't start adding new keys to parent_keys before emptying it
1222 * out (which we did with btree_split_insert_keys() above)
1224 bch2_keylist_add(&as->parent_keys, &n1->key);
1225 bch2_keylist_add(&as->parent_keys, &n2->key);
1228 /* Depth increases, make a new root */
1229 n3 = __btree_root_alloc(as, b->c.level + 1);
1231 n3->sib_u64s[0] = U16_MAX;
1232 n3->sib_u64s[1] = U16_MAX;
1234 btree_split_insert_keys(as, n3, iter, &as->parent_keys);
1236 bch2_btree_node_write(c, n3, SIX_LOCK_intent);
1239 trace_btree_compact(c, b);
1241 bch2_btree_build_aux_trees(n1);
1242 six_unlock_write(&n1->c.lock);
1245 bch2_keylist_add(&as->parent_keys, &n1->key);
1248 bch2_btree_node_write(c, n1, SIX_LOCK_intent);
1250 /* New nodes all written, now make them visible: */
1253 /* Split a non root node */
1254 bch2_btree_insert_node(as, parent, iter, &as->parent_keys, flags);
1256 bch2_btree_set_root(as, n3, iter);
1258 /* Root filled up but didn't need to be split */
1259 bch2_btree_set_root(as, n1, iter);
1262 bch2_btree_update_get_open_buckets(as, n1);
1264 bch2_btree_update_get_open_buckets(as, n2);
1266 bch2_btree_update_get_open_buckets(as, n3);
1268 /* Successful split, update the iterator to point to the new nodes: */
1270 six_lock_increment(&b->c.lock, SIX_LOCK_intent);
1271 bch2_btree_iter_node_drop(iter, b);
1273 bch2_btree_iter_node_replace(iter, n3);
1275 bch2_btree_iter_node_replace(iter, n2);
1276 bch2_btree_iter_node_replace(iter, n1);
1279 * The old node must be freed (in memory) _before_ unlocking the new
1280 * nodes - else another thread could re-acquire a read lock on the old
1281 * node after another thread has locked and updated the new node, thus
1282 * seeing stale data:
1284 bch2_btree_node_free_inmem(c, b, iter);
1287 six_unlock_intent(&n3->c.lock);
1289 six_unlock_intent(&n2->c.lock);
1290 six_unlock_intent(&n1->c.lock);
1292 bch2_btree_trans_verify_locks(iter->trans);
1294 bch2_time_stats_update(&c->times[BCH_TIME_btree_node_split],
1299 bch2_btree_insert_keys_interior(struct btree_update *as, struct btree *b,
1300 struct btree_iter *iter, struct keylist *keys)
1302 struct btree_iter *linked;
1303 struct btree_node_iter node_iter;
1304 struct bkey_i *insert = bch2_keylist_front(keys);
1305 struct bkey_packed *k;
1307 /* Don't screw up @iter's position: */
1308 node_iter = iter->l[b->c.level].iter;
1311 * btree_split(), btree_gc_coalesce() will insert keys before
1312 * the iterator's current position - they know the keys go in
1313 * the node the iterator points to:
1315 while ((k = bch2_btree_node_iter_prev_all(&node_iter, b)) &&
1316 (bkey_cmp_packed(b, k, &insert->k) >= 0))
1319 for_each_keylist_key(keys, insert)
1320 bch2_insert_fixup_btree_ptr(as, b, iter, insert, &node_iter);
1322 btree_update_updated_node(as, b);
1324 trans_for_each_iter_with_node(iter->trans, b, linked)
1325 bch2_btree_node_iter_peek(&linked->l[b->c.level].iter, b);
1327 bch2_btree_trans_verify_iters(iter->trans, b);
1331 * bch_btree_insert_node - insert bkeys into a given btree node
1333 * @iter: btree iterator
1334 * @keys: list of keys to insert
1335 * @hook: insert callback
1336 * @persistent: if not null, @persistent will wait on journal write
1338 * Inserts as many keys as it can into a given btree node, splitting it if full.
1339 * If a split occurred, this function will return early. This can only happen
1340 * for leaf nodes -- inserts into interior nodes have to be atomic.
1342 void bch2_btree_insert_node(struct btree_update *as, struct btree *b,
1343 struct btree_iter *iter, struct keylist *keys,
1346 struct bch_fs *c = as->c;
1347 int old_u64s = le16_to_cpu(btree_bset_last(b)->u64s);
1348 int old_live_u64s = b->nr.live_u64s;
1349 int live_u64s_added, u64s_added;
1351 BUG_ON(!btree_node_intent_locked(iter, btree_node_root(c, b)->c.level));
1352 BUG_ON(!b->c.level);
1353 BUG_ON(!as || as->b);
1354 bch2_verify_keylist_sorted(keys);
1356 if (as->must_rewrite)
1359 bch2_btree_node_lock_for_insert(c, b, iter);
1361 if (!bch2_btree_node_insert_fits(c, b, bch2_keylist_u64s(keys))) {
1362 bch2_btree_node_unlock_write(b, iter);
1366 bch2_btree_insert_keys_interior(as, b, iter, keys);
1368 live_u64s_added = (int) b->nr.live_u64s - old_live_u64s;
1369 u64s_added = (int) le16_to_cpu(btree_bset_last(b)->u64s) - old_u64s;
1371 if (b->sib_u64s[0] != U16_MAX && live_u64s_added < 0)
1372 b->sib_u64s[0] = max(0, (int) b->sib_u64s[0] + live_u64s_added);
1373 if (b->sib_u64s[1] != U16_MAX && live_u64s_added < 0)
1374 b->sib_u64s[1] = max(0, (int) b->sib_u64s[1] + live_u64s_added);
1376 if (u64s_added > live_u64s_added &&
1377 bch2_maybe_compact_whiteouts(c, b))
1378 bch2_btree_iter_reinit_node(iter, b);
1380 bch2_btree_node_unlock_write(b, iter);
1382 btree_node_interior_verify(c, b);
1385 * when called from the btree_split path the new nodes aren't added to
1386 * the btree iterator yet, so the merge path's unlock/wait/relock dance
1389 bch2_foreground_maybe_merge(c, iter, b->c.level,
1390 flags|BTREE_INSERT_NOUNLOCK);
1393 btree_split(as, b, iter, keys, flags);
1396 int bch2_btree_split_leaf(struct bch_fs *c, struct btree_iter *iter,
1399 struct btree_trans *trans = iter->trans;
1400 struct btree *b = iter_l(iter)->b;
1401 struct btree_update *as;
1404 struct btree_insert_entry *i;
1407 * We already have a disk reservation and open buckets pinned; this
1408 * allocation must not block:
1410 trans_for_each_update(trans, i)
1411 if (btree_node_type_needs_gc(i->iter->btree_id))
1412 flags |= BTREE_INSERT_USE_RESERVE;
1414 closure_init_stack(&cl);
1416 /* Hack, because gc and splitting nodes doesn't mix yet: */
1417 if (!(flags & BTREE_INSERT_GC_LOCK_HELD) &&
1418 !down_read_trylock(&c->gc_lock)) {
1419 if (flags & BTREE_INSERT_NOUNLOCK) {
1420 trace_transaction_restart_ip(trans->ip, _THIS_IP_);
1424 bch2_trans_unlock(trans);
1425 down_read(&c->gc_lock);
1427 if (!bch2_trans_relock(trans))
1432 * XXX: figure out how far we might need to split,
1433 * instead of locking/reserving all the way to the root:
1435 if (!bch2_btree_iter_upgrade(iter, U8_MAX)) {
1436 trace_trans_restart_iter_upgrade(trans->ip);
1441 as = bch2_btree_update_start(trans, iter->btree_id,
1442 btree_update_reserve_required(c, b), flags,
1443 !(flags & BTREE_INSERT_NOUNLOCK) ? &cl : NULL);
1446 if (ret == -EAGAIN) {
1447 BUG_ON(flags & BTREE_INSERT_NOUNLOCK);
1448 bch2_trans_unlock(trans);
1451 trace_transaction_restart_ip(trans->ip, _THIS_IP_);
1456 btree_split(as, b, iter, NULL, flags);
1457 bch2_btree_update_done(as);
1460 * We haven't successfully inserted yet, so don't downgrade all the way
1461 * back to read locks;
1463 __bch2_btree_iter_downgrade(iter, 1);
1465 if (!(flags & BTREE_INSERT_GC_LOCK_HELD))
1466 up_read(&c->gc_lock);
1471 void __bch2_foreground_maybe_merge(struct bch_fs *c,
1472 struct btree_iter *iter,
1475 enum btree_node_sibling sib)
1477 struct btree_trans *trans = iter->trans;
1478 struct btree_update *as;
1479 struct bkey_format_state new_s;
1480 struct bkey_format new_f;
1481 struct bkey_i delete;
1482 struct btree *b, *m, *n, *prev, *next, *parent;
1487 BUG_ON(!btree_node_locked(iter, level));
1489 closure_init_stack(&cl);
1491 BUG_ON(!btree_node_locked(iter, level));
1493 b = iter->l[level].b;
1495 parent = btree_node_parent(iter, b);
1499 if (b->sib_u64s[sib] > BTREE_FOREGROUND_MERGE_THRESHOLD(c))
1502 /* XXX: can't be holding read locks */
1503 m = bch2_btree_node_get_sibling(c, iter, b, sib);
1509 /* NULL means no sibling: */
1511 b->sib_u64s[sib] = U16_MAX;
1515 if (sib == btree_prev_sib) {
1523 bch2_bkey_format_init(&new_s);
1524 __bch2_btree_calc_format(&new_s, b);
1525 __bch2_btree_calc_format(&new_s, m);
1526 new_f = bch2_bkey_format_done(&new_s);
1528 sib_u64s = btree_node_u64s_with_format(b, &new_f) +
1529 btree_node_u64s_with_format(m, &new_f);
1531 if (sib_u64s > BTREE_FOREGROUND_MERGE_HYSTERESIS(c)) {
1532 sib_u64s -= BTREE_FOREGROUND_MERGE_HYSTERESIS(c);
1534 sib_u64s += BTREE_FOREGROUND_MERGE_HYSTERESIS(c);
1537 sib_u64s = min(sib_u64s, btree_max_u64s(c));
1538 b->sib_u64s[sib] = sib_u64s;
1540 if (b->sib_u64s[sib] > BTREE_FOREGROUND_MERGE_THRESHOLD(c)) {
1541 six_unlock_intent(&m->c.lock);
1545 /* We're changing btree topology, doesn't mix with gc: */
1546 if (!(flags & BTREE_INSERT_GC_LOCK_HELD) &&
1547 !down_read_trylock(&c->gc_lock))
1548 goto err_cycle_gc_lock;
1550 if (!bch2_btree_iter_upgrade(iter, U8_MAX)) {
1555 as = bch2_btree_update_start(trans, iter->btree_id,
1556 btree_update_reserve_required(c, parent) + 1,
1558 BTREE_INSERT_NOFAIL|
1559 BTREE_INSERT_USE_RESERVE,
1560 !(flags & BTREE_INSERT_NOUNLOCK) ? &cl : NULL);
1566 trace_btree_merge(c, b);
1568 bch2_btree_interior_update_will_free_node(as, b);
1569 bch2_btree_interior_update_will_free_node(as, m);
1571 n = bch2_btree_node_alloc(as, b->c.level);
1572 bch2_btree_update_add_new_node(as, n);
1574 btree_set_min(n, prev->data->min_key);
1575 btree_set_max(n, next->data->max_key);
1576 n->data->format = new_f;
1578 btree_node_set_format(n, new_f);
1580 bch2_btree_sort_into(c, n, prev);
1581 bch2_btree_sort_into(c, n, next);
1583 bch2_btree_build_aux_trees(n);
1584 six_unlock_write(&n->c.lock);
1586 bkey_init(&delete.k);
1587 delete.k.p = prev->key.k.p;
1588 bch2_keylist_add(&as->parent_keys, &delete);
1589 bch2_keylist_add(&as->parent_keys, &n->key);
1591 bch2_btree_node_write(c, n, SIX_LOCK_intent);
1593 bch2_btree_insert_node(as, parent, iter, &as->parent_keys, flags);
1595 bch2_btree_update_get_open_buckets(as, n);
1597 six_lock_increment(&b->c.lock, SIX_LOCK_intent);
1598 bch2_btree_iter_node_drop(iter, b);
1599 bch2_btree_iter_node_drop(iter, m);
1601 bch2_btree_iter_node_replace(iter, n);
1603 bch2_btree_trans_verify_iters(trans, n);
1605 bch2_btree_node_free_inmem(c, b, iter);
1606 bch2_btree_node_free_inmem(c, m, iter);
1608 six_unlock_intent(&n->c.lock);
1610 bch2_btree_update_done(as);
1612 if (!(flags & BTREE_INSERT_GC_LOCK_HELD))
1613 up_read(&c->gc_lock);
1615 bch2_btree_trans_verify_locks(trans);
1618 * Don't downgrade locks here: we're called after successful insert,
1619 * and the caller will downgrade locks after a successful insert
1620 * anyways (in case e.g. a split was required first)
1622 * And we're also called when inserting into interior nodes in the
1623 * split path, and downgrading to read locks in there is potentially
1630 six_unlock_intent(&m->c.lock);
1632 if (flags & BTREE_INSERT_NOUNLOCK)
1635 bch2_trans_unlock(trans);
1637 down_read(&c->gc_lock);
1638 up_read(&c->gc_lock);
1643 six_unlock_intent(&m->c.lock);
1644 if (!(flags & BTREE_INSERT_GC_LOCK_HELD))
1645 up_read(&c->gc_lock);
1647 BUG_ON(ret == -EAGAIN && (flags & BTREE_INSERT_NOUNLOCK));
1649 if ((ret == -EAGAIN || ret == -EINTR) &&
1650 !(flags & BTREE_INSERT_NOUNLOCK)) {
1651 bch2_trans_unlock(trans);
1653 ret = bch2_btree_iter_traverse(iter);
1663 static int __btree_node_rewrite(struct bch_fs *c, struct btree_iter *iter,
1664 struct btree *b, unsigned flags,
1667 struct btree *n, *parent = btree_node_parent(iter, b);
1668 struct btree_update *as;
1670 as = bch2_btree_update_start(iter->trans, iter->btree_id,
1672 ? btree_update_reserve_required(c, parent)
1676 trace_btree_gc_rewrite_node_fail(c, b);
1680 bch2_btree_interior_update_will_free_node(as, b);
1682 n = bch2_btree_node_alloc_replacement(as, b);
1683 bch2_btree_update_add_new_node(as, n);
1685 bch2_btree_build_aux_trees(n);
1686 six_unlock_write(&n->c.lock);
1688 trace_btree_gc_rewrite_node(c, b);
1690 bch2_btree_node_write(c, n, SIX_LOCK_intent);
1693 bch2_keylist_add(&as->parent_keys, &n->key);
1694 bch2_btree_insert_node(as, parent, iter, &as->parent_keys, flags);
1696 bch2_btree_set_root(as, n, iter);
1699 bch2_btree_update_get_open_buckets(as, n);
1701 six_lock_increment(&b->c.lock, SIX_LOCK_intent);
1702 bch2_btree_iter_node_drop(iter, b);
1703 bch2_btree_iter_node_replace(iter, n);
1704 bch2_btree_node_free_inmem(c, b, iter);
1705 six_unlock_intent(&n->c.lock);
1707 bch2_btree_update_done(as);
1712 * bch_btree_node_rewrite - Rewrite/move a btree node
1714 * Returns 0 on success, -EINTR or -EAGAIN on failure (i.e.
1715 * btree_check_reserve() has to wait)
1717 int bch2_btree_node_rewrite(struct bch_fs *c, struct btree_iter *iter,
1718 __le64 seq, unsigned flags)
1720 struct btree_trans *trans = iter->trans;
1725 flags |= BTREE_INSERT_NOFAIL;
1727 closure_init_stack(&cl);
1729 bch2_btree_iter_upgrade(iter, U8_MAX);
1731 if (!(flags & BTREE_INSERT_GC_LOCK_HELD)) {
1732 if (!down_read_trylock(&c->gc_lock)) {
1733 bch2_trans_unlock(trans);
1734 down_read(&c->gc_lock);
1739 ret = bch2_btree_iter_traverse(iter);
1743 b = bch2_btree_iter_peek_node(iter);
1744 if (!b || b->data->keys.seq != seq)
1747 ret = __btree_node_rewrite(c, iter, b, flags, &cl);
1748 if (ret != -EAGAIN &&
1752 bch2_trans_unlock(trans);
1756 bch2_btree_iter_downgrade(iter);
1758 if (!(flags & BTREE_INSERT_GC_LOCK_HELD))
1759 up_read(&c->gc_lock);
1765 static void __bch2_btree_node_update_key(struct bch_fs *c,
1766 struct btree_update *as,
1767 struct btree_iter *iter,
1768 struct btree *b, struct btree *new_hash,
1769 struct bkey_i *new_key)
1771 struct btree *parent;
1774 btree_update_will_delete_key(as, &b->key);
1775 btree_update_will_add_key(as, new_key);
1777 parent = btree_node_parent(iter, b);
1780 bkey_copy(&new_hash->key, new_key);
1781 ret = bch2_btree_node_hash_insert(&c->btree_cache,
1782 new_hash, b->c.level, b->c.btree_id);
1786 bch2_keylist_add(&as->parent_keys, new_key);
1787 bch2_btree_insert_node(as, parent, iter, &as->parent_keys, 0);
1790 mutex_lock(&c->btree_cache.lock);
1791 bch2_btree_node_hash_remove(&c->btree_cache, new_hash);
1793 bch2_btree_node_hash_remove(&c->btree_cache, b);
1795 bkey_copy(&b->key, new_key);
1796 ret = __bch2_btree_node_hash_insert(&c->btree_cache, b);
1798 mutex_unlock(&c->btree_cache.lock);
1800 bkey_copy(&b->key, new_key);
1803 BUG_ON(btree_node_root(c, b) != b);
1805 bch2_btree_node_lock_write(b, iter);
1806 bkey_copy(&b->key, new_key);
1808 if (btree_ptr_hash_val(&b->key) != b->hash_val) {
1809 mutex_lock(&c->btree_cache.lock);
1810 bch2_btree_node_hash_remove(&c->btree_cache, b);
1812 ret = __bch2_btree_node_hash_insert(&c->btree_cache, b);
1814 mutex_unlock(&c->btree_cache.lock);
1817 btree_update_updated_root(as, b);
1818 bch2_btree_node_unlock_write(b, iter);
1821 bch2_btree_update_done(as);
1824 int bch2_btree_node_update_key(struct bch_fs *c, struct btree_iter *iter,
1826 struct bkey_i *new_key)
1828 struct btree *parent = btree_node_parent(iter, b);
1829 struct btree_update *as = NULL;
1830 struct btree *new_hash = NULL;
1834 closure_init_stack(&cl);
1836 if (!bch2_btree_iter_upgrade(iter, U8_MAX))
1839 if (!down_read_trylock(&c->gc_lock)) {
1840 bch2_trans_unlock(iter->trans);
1841 down_read(&c->gc_lock);
1843 if (!bch2_trans_relock(iter->trans)) {
1850 * check btree_ptr_hash_val() after @b is locked by
1851 * btree_iter_traverse():
1853 if (btree_ptr_hash_val(new_key) != b->hash_val) {
1854 /* bch2_btree_reserve_get will unlock */
1855 ret = bch2_btree_cache_cannibalize_lock(c, &cl);
1857 bch2_trans_unlock(iter->trans);
1858 up_read(&c->gc_lock);
1860 down_read(&c->gc_lock);
1862 if (!bch2_trans_relock(iter->trans)) {
1868 new_hash = bch2_btree_node_mem_alloc(c);
1871 as = bch2_btree_update_start(iter->trans, iter->btree_id,
1872 parent ? btree_update_reserve_required(c, parent) : 0,
1873 BTREE_INSERT_NOFAIL|
1874 BTREE_INSERT_USE_RESERVE|
1875 BTREE_INSERT_USE_ALLOC_RESERVE,
1883 if (ret == -EINTR) {
1884 bch2_trans_unlock(iter->trans);
1885 up_read(&c->gc_lock);
1887 down_read(&c->gc_lock);
1889 if (bch2_trans_relock(iter->trans))
1896 ret = bch2_mark_bkey_replicas(c, bkey_i_to_s_c(new_key));
1898 goto err_free_update;
1900 __bch2_btree_node_update_key(c, as, iter, b, new_hash, new_key);
1902 bch2_btree_iter_downgrade(iter);
1905 mutex_lock(&c->btree_cache.lock);
1906 list_move(&new_hash->list, &c->btree_cache.freeable);
1907 mutex_unlock(&c->btree_cache.lock);
1909 six_unlock_write(&new_hash->c.lock);
1910 six_unlock_intent(&new_hash->c.lock);
1912 up_read(&c->gc_lock);
1916 bch2_btree_update_free(as);
1923 * Only for filesystem bringup, when first reading the btree roots or allocating
1924 * btree roots when initializing a new filesystem:
1926 void bch2_btree_set_root_for_read(struct bch_fs *c, struct btree *b)
1928 BUG_ON(btree_node_root(c, b));
1930 bch2_btree_set_root_inmem(c, b);
1933 void bch2_btree_root_alloc(struct bch_fs *c, enum btree_id id)
1939 closure_init_stack(&cl);
1942 ret = bch2_btree_cache_cannibalize_lock(c, &cl);
1946 b = bch2_btree_node_mem_alloc(c);
1947 bch2_btree_cache_cannibalize_unlock(c);
1949 set_btree_node_fake(b);
1950 set_btree_node_need_rewrite(b);
1954 bkey_btree_ptr_init(&b->key);
1955 b->key.k.p = POS_MAX;
1956 *((u64 *) bkey_i_to_btree_ptr(&b->key)->v.start) = U64_MAX - id;
1958 bch2_bset_init_first(b, &b->data->keys);
1959 bch2_btree_build_aux_trees(b);
1962 btree_set_min(b, POS_MIN);
1963 btree_set_max(b, POS_MAX);
1964 b->data->format = bch2_btree_calc_format(b);
1965 btree_node_set_format(b, b->data->format);
1967 ret = bch2_btree_node_hash_insert(&c->btree_cache, b,
1968 b->c.level, b->c.btree_id);
1971 bch2_btree_set_root_inmem(c, b);
1973 six_unlock_write(&b->c.lock);
1974 six_unlock_intent(&b->c.lock);
1977 void bch2_btree_updates_to_text(struct printbuf *out, struct bch_fs *c)
1979 struct btree_update *as;
1981 mutex_lock(&c->btree_interior_update_lock);
1982 list_for_each_entry(as, &c->btree_interior_update_list, list)
1983 pr_buf(out, "%p m %u w %u r %u j %llu\n",
1987 atomic_read(&as->cl.remaining) & CLOSURE_REMAINING_MASK,
1989 mutex_unlock(&c->btree_interior_update_lock);
1992 size_t bch2_btree_interior_updates_nr_pending(struct bch_fs *c)
1995 struct list_head *i;
1997 mutex_lock(&c->btree_interior_update_lock);
1998 list_for_each(i, &c->btree_interior_update_list)
2000 mutex_unlock(&c->btree_interior_update_lock);
2005 void bch2_journal_entries_to_btree_roots(struct bch_fs *c, struct jset *jset)
2007 struct btree_root *r;
2008 struct jset_entry *entry;
2010 mutex_lock(&c->btree_root_lock);
2012 vstruct_for_each(jset, entry)
2013 if (entry->type == BCH_JSET_ENTRY_btree_root) {
2014 r = &c->btree_roots[entry->btree_id];
2015 r->level = entry->level;
2017 bkey_copy(&r->key, &entry->start[0]);
2020 mutex_unlock(&c->btree_root_lock);
2024 bch2_btree_roots_to_journal_entries(struct bch_fs *c,
2025 struct jset_entry *start,
2026 struct jset_entry *end)
2028 struct jset_entry *entry;
2029 unsigned long have = 0;
2032 for (entry = start; entry < end; entry = vstruct_next(entry))
2033 if (entry->type == BCH_JSET_ENTRY_btree_root)
2034 __set_bit(entry->btree_id, &have);
2036 mutex_lock(&c->btree_root_lock);
2038 for (i = 0; i < BTREE_ID_NR; i++)
2039 if (c->btree_roots[i].alive && !test_bit(i, &have)) {
2040 journal_entry_set(end,
2041 BCH_JSET_ENTRY_btree_root,
2042 i, c->btree_roots[i].level,
2043 &c->btree_roots[i].key,
2044 c->btree_roots[i].key.u64s);
2045 end = vstruct_next(end);
2048 mutex_unlock(&c->btree_root_lock);
2053 void bch2_fs_btree_interior_update_exit(struct bch_fs *c)
2055 if (c->btree_interior_update_worker)
2056 destroy_workqueue(c->btree_interior_update_worker);
2057 mempool_exit(&c->btree_interior_update_pool);
2060 int bch2_fs_btree_interior_update_init(struct bch_fs *c)
2062 mutex_init(&c->btree_reserve_cache_lock);
2063 INIT_LIST_HEAD(&c->btree_interior_update_list);
2064 INIT_LIST_HEAD(&c->btree_interior_updates_unwritten);
2065 mutex_init(&c->btree_interior_update_lock);
2066 INIT_WORK(&c->btree_interior_update_work, btree_interior_update_work);
2068 c->btree_interior_update_worker =
2069 alloc_workqueue("btree_update", WQ_UNBOUND|WQ_MEM_RECLAIM, 1);
2070 if (!c->btree_interior_update_worker)
2073 return mempool_init_kmalloc_pool(&c->btree_interior_update_pool, 1,
2074 sizeof(struct btree_update));