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"
17 #include "journal_reclaim.h"
23 #include <linux/random.h>
24 #include <trace/events/bcachefs.h>
26 static void bch2_btree_insert_node(struct btree_update *, struct btree_trans *,
27 struct btree_path *, struct btree *,
28 struct keylist *, unsigned);
29 static void bch2_btree_update_add_new_node(struct btree_update *, struct btree *);
34 * Verify that child nodes correctly span parent node's range:
36 static void btree_node_interior_verify(struct bch_fs *c, struct btree *b)
38 #ifdef CONFIG_BCACHEFS_DEBUG
39 struct bpos next_node = b->data->min_key;
40 struct btree_node_iter iter;
42 struct bkey_s_c_btree_ptr_v2 bp;
44 struct printbuf buf1 = PRINTBUF, buf2 = PRINTBUF;
48 if (!test_bit(JOURNAL_REPLAY_DONE, &c->journal.flags))
51 bch2_btree_node_iter_init_from_start(&iter, b);
54 k = bch2_btree_node_iter_peek_unpack(&iter, b, &unpacked);
55 if (k.k->type != KEY_TYPE_btree_ptr_v2)
57 bp = bkey_s_c_to_btree_ptr_v2(k);
59 if (bpos_cmp(next_node, bp.v->min_key)) {
60 bch2_dump_btree_node(c, b);
61 bch2_bpos_to_text(&buf1, next_node);
62 bch2_bpos_to_text(&buf2, bp.v->min_key);
63 panic("expected next min_key %s got %s\n", buf1.buf, buf2.buf);
66 bch2_btree_node_iter_advance(&iter, b);
68 if (bch2_btree_node_iter_end(&iter)) {
69 if (bpos_cmp(k.k->p, b->key.k.p)) {
70 bch2_dump_btree_node(c, b);
71 bch2_bpos_to_text(&buf1, b->key.k.p);
72 bch2_bpos_to_text(&buf2, k.k->p);
73 panic("expected end %s got %s\n", buf1.buf, buf2.buf);
78 next_node = bpos_successor(k.k->p);
83 /* Calculate ideal packed bkey format for new btree nodes: */
85 void __bch2_btree_calc_format(struct bkey_format_state *s, struct btree *b)
87 struct bkey_packed *k;
92 bset_tree_for_each_key(b, t, k)
93 if (!bkey_deleted(k)) {
94 uk = bkey_unpack_key(b, k);
95 bch2_bkey_format_add_key(s, &uk);
99 static struct bkey_format bch2_btree_calc_format(struct btree *b)
101 struct bkey_format_state s;
103 bch2_bkey_format_init(&s);
104 bch2_bkey_format_add_pos(&s, b->data->min_key);
105 bch2_bkey_format_add_pos(&s, b->data->max_key);
106 __bch2_btree_calc_format(&s, b);
108 return bch2_bkey_format_done(&s);
111 static size_t btree_node_u64s_with_format(struct btree *b,
112 struct bkey_format *new_f)
114 struct bkey_format *old_f = &b->format;
116 /* stupid integer promotion rules */
118 (((int) new_f->key_u64s - old_f->key_u64s) *
119 (int) b->nr.packed_keys) +
120 (((int) new_f->key_u64s - BKEY_U64s) *
121 (int) b->nr.unpacked_keys);
123 BUG_ON(delta + b->nr.live_u64s < 0);
125 return b->nr.live_u64s + delta;
129 * btree_node_format_fits - check if we could rewrite node with a new format
131 * This assumes all keys can pack with the new format -- it just checks if
132 * the re-packed keys would fit inside the node itself.
134 bool bch2_btree_node_format_fits(struct bch_fs *c, struct btree *b,
135 struct bkey_format *new_f)
137 size_t u64s = btree_node_u64s_with_format(b, new_f);
139 return __vstruct_bytes(struct btree_node, u64s) < btree_bytes(c);
142 /* Btree node freeing/allocation: */
144 static void __btree_node_free(struct bch_fs *c, struct btree *b)
146 trace_btree_node_free(c, b);
148 BUG_ON(btree_node_dirty(b));
149 BUG_ON(btree_node_need_write(b));
150 BUG_ON(b == btree_node_root(c, b));
152 BUG_ON(!list_empty(&b->write_blocked));
153 BUG_ON(b->will_make_reachable);
155 clear_btree_node_noevict(b);
157 mutex_lock(&c->btree_cache.lock);
158 list_move(&b->list, &c->btree_cache.freeable);
159 mutex_unlock(&c->btree_cache.lock);
162 static void bch2_btree_node_free_inmem(struct btree_trans *trans,
165 struct bch_fs *c = trans->c;
166 struct btree_path *path;
168 trans_for_each_path(trans, path)
169 BUG_ON(path->l[b->c.level].b == b &&
170 path->l[b->c.level].lock_seq == b->c.lock.state.seq);
172 six_lock_write(&b->c.lock, NULL, NULL);
174 bch2_btree_node_hash_remove(&c->btree_cache, b);
175 __btree_node_free(c, b);
177 six_unlock_write(&b->c.lock);
178 six_unlock_intent(&b->c.lock);
181 static struct btree *__bch2_btree_node_alloc(struct bch_fs *c,
182 struct disk_reservation *res,
187 struct write_point *wp;
189 __BKEY_PADDED(k, BKEY_BTREE_PTR_VAL_U64s_MAX) tmp;
190 struct open_buckets ob = { .nr = 0 };
191 struct bch_devs_list devs_have = (struct bch_devs_list) { 0 };
193 enum alloc_reserve alloc_reserve;
195 if (flags & BTREE_INSERT_USE_RESERVE) {
197 alloc_reserve = RESERVE_BTREE_MOVINGGC;
199 nr_reserve = BTREE_NODE_RESERVE;
200 alloc_reserve = RESERVE_BTREE;
203 mutex_lock(&c->btree_reserve_cache_lock);
204 if (c->btree_reserve_cache_nr > nr_reserve) {
205 struct btree_alloc *a =
206 &c->btree_reserve_cache[--c->btree_reserve_cache_nr];
209 bkey_copy(&tmp.k, &a->k);
210 mutex_unlock(&c->btree_reserve_cache_lock);
213 mutex_unlock(&c->btree_reserve_cache_lock);
216 wp = bch2_alloc_sectors_start(c,
217 c->opts.metadata_target ?:
218 c->opts.foreground_target,
220 writepoint_ptr(&c->btree_write_point),
223 c->opts.metadata_replicas_required,
224 alloc_reserve, 0, cl);
228 if (wp->sectors_free < btree_sectors(c)) {
229 struct open_bucket *ob;
232 open_bucket_for_each(c, &wp->ptrs, ob, i)
233 if (ob->sectors_free < btree_sectors(c))
234 ob->sectors_free = 0;
236 bch2_alloc_sectors_done(c, wp);
240 bkey_btree_ptr_v2_init(&tmp.k);
241 bch2_alloc_sectors_append_ptrs(c, wp, &tmp.k, btree_sectors(c), false);
243 bch2_open_bucket_get(c, wp, &ob);
244 bch2_alloc_sectors_done(c, wp);
246 b = bch2_btree_node_mem_alloc(c, interior_node);
247 six_unlock_write(&b->c.lock);
248 six_unlock_intent(&b->c.lock);
250 /* we hold cannibalize_lock: */
254 bkey_copy(&b->key, &tmp.k);
260 static struct btree *bch2_btree_node_alloc(struct btree_update *as, unsigned level)
262 struct bch_fs *c = as->c;
264 struct prealloc_nodes *p = &as->prealloc_nodes[!!level];
267 BUG_ON(level >= BTREE_MAX_DEPTH);
272 six_lock_intent(&b->c.lock, NULL, NULL);
273 six_lock_write(&b->c.lock, NULL, NULL);
275 set_btree_node_accessed(b);
276 set_btree_node_dirty_acct(c, b);
277 set_btree_node_need_write(b);
279 bch2_bset_init_first(b, &b->data->keys);
281 b->c.btree_id = as->btree_id;
282 b->version_ondisk = c->sb.version;
284 memset(&b->nr, 0, sizeof(b->nr));
285 b->data->magic = cpu_to_le64(bset_magic(c));
286 memset(&b->data->_ptr, 0, sizeof(b->data->_ptr));
288 SET_BTREE_NODE_ID(b->data, as->btree_id);
289 SET_BTREE_NODE_LEVEL(b->data, level);
291 if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
292 struct bkey_i_btree_ptr_v2 *bp = bkey_i_to_btree_ptr_v2(&b->key);
295 bp->v.seq = b->data->keys.seq;
296 bp->v.sectors_written = 0;
299 SET_BTREE_NODE_NEW_EXTENT_OVERWRITE(b->data, true);
301 bch2_btree_build_aux_trees(b);
303 ret = bch2_btree_node_hash_insert(&c->btree_cache, b, level, as->btree_id);
306 trace_btree_node_alloc(c, b);
310 static void btree_set_min(struct btree *b, struct bpos pos)
312 if (b->key.k.type == KEY_TYPE_btree_ptr_v2)
313 bkey_i_to_btree_ptr_v2(&b->key)->v.min_key = pos;
314 b->data->min_key = pos;
317 static void btree_set_max(struct btree *b, struct bpos pos)
320 b->data->max_key = pos;
323 struct btree *__bch2_btree_node_alloc_replacement(struct btree_update *as,
325 struct bkey_format format)
329 n = bch2_btree_node_alloc(as, b->c.level);
331 SET_BTREE_NODE_SEQ(n->data, BTREE_NODE_SEQ(b->data) + 1);
333 btree_set_min(n, b->data->min_key);
334 btree_set_max(n, b->data->max_key);
336 n->data->format = format;
337 btree_node_set_format(n, format);
339 bch2_btree_sort_into(as->c, n, b);
341 btree_node_reset_sib_u64s(n);
343 n->key.k.p = b->key.k.p;
347 static struct btree *bch2_btree_node_alloc_replacement(struct btree_update *as,
350 struct bkey_format new_f = bch2_btree_calc_format(b);
353 * The keys might expand with the new format - if they wouldn't fit in
354 * the btree node anymore, use the old format for now:
356 if (!bch2_btree_node_format_fits(as->c, b, &new_f))
359 return __bch2_btree_node_alloc_replacement(as, b, new_f);
362 static struct btree *__btree_root_alloc(struct btree_update *as, unsigned level)
364 struct btree *b = bch2_btree_node_alloc(as, level);
366 btree_set_min(b, POS_MIN);
367 btree_set_max(b, SPOS_MAX);
368 b->data->format = bch2_btree_calc_format(b);
370 btree_node_set_format(b, b->data->format);
371 bch2_btree_build_aux_trees(b);
373 bch2_btree_update_add_new_node(as, b);
374 six_unlock_write(&b->c.lock);
379 static void bch2_btree_reserve_put(struct btree_update *as)
381 struct bch_fs *c = as->c;
382 struct prealloc_nodes *p;
384 mutex_lock(&c->btree_reserve_cache_lock);
386 for (p = as->prealloc_nodes;
387 p < as->prealloc_nodes + ARRAY_SIZE(as->prealloc_nodes);
390 struct btree *b = p->b[--p->nr];
392 six_lock_intent(&b->c.lock, NULL, NULL);
393 six_lock_write(&b->c.lock, NULL, NULL);
395 if (c->btree_reserve_cache_nr <
396 ARRAY_SIZE(c->btree_reserve_cache)) {
397 struct btree_alloc *a =
398 &c->btree_reserve_cache[c->btree_reserve_cache_nr++];
402 bkey_copy(&a->k, &b->key);
404 bch2_open_buckets_put(c, &b->ob);
407 __btree_node_free(c, b);
408 six_unlock_write(&b->c.lock);
409 six_unlock_intent(&b->c.lock);
413 mutex_unlock(&c->btree_reserve_cache_lock);
416 static int bch2_btree_reserve_get(struct btree_update *as,
417 unsigned nr_nodes[2],
420 struct bch_fs *c = as->c;
426 closure_init_stack(&cl);
429 BUG_ON(nr_nodes[0] + nr_nodes[1] > BTREE_RESERVE_MAX);
432 * Protects reaping from the btree node cache and using the btree node
433 * open bucket reserve:
435 * BTREE_INSERT_NOWAIT only applies to btree node allocation, not
436 * blocking on this lock:
438 ret = bch2_btree_cache_cannibalize_lock(c, &cl);
442 for (interior = 0; interior < 2; interior++) {
443 struct prealloc_nodes *p = as->prealloc_nodes + interior;
445 while (p->nr < nr_nodes[interior]) {
446 b = __bch2_btree_node_alloc(c, &as->disk_res,
447 flags & BTREE_INSERT_NOWAIT
459 bch2_btree_cache_cannibalize_unlock(c);
463 bch2_btree_cache_cannibalize_unlock(c);
469 trace_btree_reserve_get_fail(c, nr_nodes[0] + nr_nodes[1], &cl);
473 /* Asynchronous interior node update machinery */
475 static void bch2_btree_update_free(struct btree_update *as)
477 struct bch_fs *c = as->c;
479 if (as->took_gc_lock)
480 up_read(&c->gc_lock);
481 as->took_gc_lock = false;
483 bch2_journal_preres_put(&c->journal, &as->journal_preres);
485 bch2_journal_pin_drop(&c->journal, &as->journal);
486 bch2_journal_pin_flush(&c->journal, &as->journal);
487 bch2_disk_reservation_put(c, &as->disk_res);
488 bch2_btree_reserve_put(as);
490 bch2_time_stats_update(&c->times[BCH_TIME_btree_interior_update_total],
493 mutex_lock(&c->btree_interior_update_lock);
494 list_del(&as->unwritten_list);
497 closure_debug_destroy(&as->cl);
498 mempool_free(as, &c->btree_interior_update_pool);
501 * Have to do the wakeup with btree_interior_update_lock still held,
502 * since being on btree_interior_update_list is our ref on @c:
504 closure_wake_up(&c->btree_interior_update_wait);
506 mutex_unlock(&c->btree_interior_update_lock);
509 static void btree_update_will_delete_key(struct btree_update *as,
512 BUG_ON(bch2_keylist_u64s(&as->old_keys) + k->k.u64s >
513 ARRAY_SIZE(as->_old_keys));
514 bch2_keylist_add(&as->old_keys, k);
517 static void btree_update_will_add_key(struct btree_update *as,
520 BUG_ON(bch2_keylist_u64s(&as->new_keys) + k->k.u64s >
521 ARRAY_SIZE(as->_new_keys));
522 bch2_keylist_add(&as->new_keys, k);
526 * The transactional part of an interior btree node update, where we journal the
527 * update we did to the interior node and update alloc info:
529 static int btree_update_nodes_written_trans(struct btree_trans *trans,
530 struct btree_update *as)
535 trans->extra_journal_entries = (void *) &as->journal_entries[0];
536 trans->extra_journal_entry_u64s = as->journal_u64s;
537 trans->journal_pin = &as->journal;
539 for_each_keylist_key(&as->new_keys, k) {
540 ret = bch2_trans_mark_new(trans, k, 0);
545 for_each_keylist_key(&as->old_keys, k) {
546 ret = bch2_trans_mark_old(trans, bkey_i_to_s_c(k), 0);
554 static void btree_update_nodes_written(struct btree_update *as)
556 struct bch_fs *c = as->c;
557 struct btree *b = as->b;
558 struct btree_trans trans;
564 * If we're already in an error state, it might be because a btree node
565 * was never written, and we might be trying to free that same btree
566 * node here, but it won't have been marked as allocated and we'll see
567 * spurious disk usage inconsistencies in the transactional part below
568 * if we don't skip it:
570 ret = bch2_journal_error(&c->journal);
575 * Wait for any in flight writes to finish before we free the old nodes
578 for (i = 0; i < as->nr_old_nodes; i++) {
579 struct btree *old = as->old_nodes[i];
582 six_lock_read(&old->c.lock, NULL, NULL);
583 seq = old->data ? old->data->keys.seq : 0;
584 six_unlock_read(&old->c.lock);
586 if (seq == as->old_nodes_seq[i])
587 wait_on_bit_io(&old->flags, BTREE_NODE_write_in_flight_inner,
588 TASK_UNINTERRUPTIBLE);
592 * We did an update to a parent node where the pointers we added pointed
593 * to child nodes that weren't written yet: now, the child nodes have
594 * been written so we can write out the update to the interior node.
598 * We can't call into journal reclaim here: we'd block on the journal
599 * reclaim lock, but we may need to release the open buckets we have
600 * pinned in order for other btree updates to make forward progress, and
601 * journal reclaim does btree updates when flushing bkey_cached entries,
602 * which may require allocations as well.
604 bch2_trans_init(&trans, c, 0, 512);
605 ret = __bch2_trans_do(&trans, &as->disk_res, &journal_seq,
607 BTREE_INSERT_NOCHECK_RW|
608 BTREE_INSERT_JOURNAL_RECLAIM|
609 BTREE_INSERT_JOURNAL_RESERVED,
610 btree_update_nodes_written_trans(&trans, as));
611 bch2_trans_exit(&trans);
613 bch2_fs_fatal_err_on(ret && !bch2_journal_error(&c->journal), c,
614 "error %i in btree_update_nodes_written()", ret);
618 * @b is the node we did the final insert into:
620 * On failure to get a journal reservation, we still have to
621 * unblock the write and allow most of the write path to happen
622 * so that shutdown works, but the i->journal_seq mechanism
623 * won't work to prevent the btree write from being visible (we
624 * didn't get a journal sequence number) - instead
625 * __bch2_btree_node_write() doesn't do the actual write if
626 * we're in journal error state:
629 six_lock_intent(&b->c.lock, NULL, NULL);
630 six_lock_write(&b->c.lock, NULL, NULL);
631 mutex_lock(&c->btree_interior_update_lock);
633 list_del(&as->write_blocked_list);
634 if (list_empty(&b->write_blocked))
635 clear_btree_node_write_blocked(b);
638 * Node might have been freed, recheck under
639 * btree_interior_update_lock:
642 struct bset *i = btree_bset_last(b);
645 BUG_ON(!btree_node_dirty(b));
648 i->journal_seq = cpu_to_le64(
650 le64_to_cpu(i->journal_seq)));
652 bch2_btree_add_journal_pin(c, b, journal_seq);
655 * If we didn't get a journal sequence number we
656 * can't write this btree node, because recovery
657 * won't know to ignore this write:
659 set_btree_node_never_write(b);
663 mutex_unlock(&c->btree_interior_update_lock);
664 six_unlock_write(&b->c.lock);
666 btree_node_write_if_need(c, b, SIX_LOCK_intent);
667 six_unlock_intent(&b->c.lock);
670 bch2_journal_pin_drop(&c->journal, &as->journal);
672 bch2_journal_preres_put(&c->journal, &as->journal_preres);
674 mutex_lock(&c->btree_interior_update_lock);
675 for (i = 0; i < as->nr_new_nodes; i++) {
676 b = as->new_nodes[i];
678 BUG_ON(b->will_make_reachable != (unsigned long) as);
679 b->will_make_reachable = 0;
680 clear_btree_node_will_make_reachable(b);
682 mutex_unlock(&c->btree_interior_update_lock);
684 for (i = 0; i < as->nr_new_nodes; i++) {
685 b = as->new_nodes[i];
687 six_lock_read(&b->c.lock, NULL, NULL);
688 btree_node_write_if_need(c, b, SIX_LOCK_read);
689 six_unlock_read(&b->c.lock);
692 for (i = 0; i < as->nr_open_buckets; i++)
693 bch2_open_bucket_put(c, c->open_buckets + as->open_buckets[i]);
695 bch2_btree_update_free(as);
698 static void btree_interior_update_work(struct work_struct *work)
701 container_of(work, struct bch_fs, btree_interior_update_work);
702 struct btree_update *as;
705 mutex_lock(&c->btree_interior_update_lock);
706 as = list_first_entry_or_null(&c->btree_interior_updates_unwritten,
707 struct btree_update, unwritten_list);
708 if (as && !as->nodes_written)
710 mutex_unlock(&c->btree_interior_update_lock);
715 btree_update_nodes_written(as);
719 static void btree_update_set_nodes_written(struct closure *cl)
721 struct btree_update *as = container_of(cl, struct btree_update, cl);
722 struct bch_fs *c = as->c;
724 mutex_lock(&c->btree_interior_update_lock);
725 as->nodes_written = true;
726 mutex_unlock(&c->btree_interior_update_lock);
728 queue_work(c->btree_interior_update_worker, &c->btree_interior_update_work);
732 * We're updating @b with pointers to nodes that haven't finished writing yet:
733 * block @b from being written until @as completes
735 static void btree_update_updated_node(struct btree_update *as, struct btree *b)
737 struct bch_fs *c = as->c;
739 mutex_lock(&c->btree_interior_update_lock);
740 list_add_tail(&as->unwritten_list, &c->btree_interior_updates_unwritten);
742 BUG_ON(as->mode != BTREE_INTERIOR_NO_UPDATE);
743 BUG_ON(!btree_node_dirty(b));
745 as->mode = BTREE_INTERIOR_UPDATING_NODE;
748 set_btree_node_write_blocked(b);
749 list_add(&as->write_blocked_list, &b->write_blocked);
751 mutex_unlock(&c->btree_interior_update_lock);
754 static void btree_update_reparent(struct btree_update *as,
755 struct btree_update *child)
757 struct bch_fs *c = as->c;
759 lockdep_assert_held(&c->btree_interior_update_lock);
762 child->mode = BTREE_INTERIOR_UPDATING_AS;
764 bch2_journal_pin_copy(&c->journal, &as->journal, &child->journal, NULL);
767 static void btree_update_updated_root(struct btree_update *as, struct btree *b)
769 struct bkey_i *insert = &b->key;
770 struct bch_fs *c = as->c;
772 BUG_ON(as->mode != BTREE_INTERIOR_NO_UPDATE);
774 BUG_ON(as->journal_u64s + jset_u64s(insert->k.u64s) >
775 ARRAY_SIZE(as->journal_entries));
778 journal_entry_set((void *) &as->journal_entries[as->journal_u64s],
779 BCH_JSET_ENTRY_btree_root,
780 b->c.btree_id, b->c.level,
781 insert, insert->k.u64s);
783 mutex_lock(&c->btree_interior_update_lock);
784 list_add_tail(&as->unwritten_list, &c->btree_interior_updates_unwritten);
786 as->mode = BTREE_INTERIOR_UPDATING_ROOT;
787 mutex_unlock(&c->btree_interior_update_lock);
791 * bch2_btree_update_add_new_node:
793 * This causes @as to wait on @b to be written, before it gets to
794 * bch2_btree_update_nodes_written
796 * Additionally, it sets b->will_make_reachable to prevent any additional writes
797 * to @b from happening besides the first until @b is reachable on disk
799 * And it adds @b to the list of @as's new nodes, so that we can update sector
800 * counts in bch2_btree_update_nodes_written:
802 static void bch2_btree_update_add_new_node(struct btree_update *as, struct btree *b)
804 struct bch_fs *c = as->c;
806 closure_get(&as->cl);
808 mutex_lock(&c->btree_interior_update_lock);
809 BUG_ON(as->nr_new_nodes >= ARRAY_SIZE(as->new_nodes));
810 BUG_ON(b->will_make_reachable);
812 as->new_nodes[as->nr_new_nodes++] = b;
813 b->will_make_reachable = 1UL|(unsigned long) as;
814 set_btree_node_will_make_reachable(b);
816 mutex_unlock(&c->btree_interior_update_lock);
818 btree_update_will_add_key(as, &b->key);
822 * returns true if @b was a new node
824 static void btree_update_drop_new_node(struct bch_fs *c, struct btree *b)
826 struct btree_update *as;
830 mutex_lock(&c->btree_interior_update_lock);
832 * When b->will_make_reachable != 0, it owns a ref on as->cl that's
833 * dropped when it gets written by bch2_btree_complete_write - the
834 * xchg() is for synchronization with bch2_btree_complete_write:
836 v = xchg(&b->will_make_reachable, 0);
837 clear_btree_node_will_make_reachable(b);
838 as = (struct btree_update *) (v & ~1UL);
841 mutex_unlock(&c->btree_interior_update_lock);
845 for (i = 0; i < as->nr_new_nodes; i++)
846 if (as->new_nodes[i] == b)
851 array_remove_item(as->new_nodes, as->nr_new_nodes, i);
852 mutex_unlock(&c->btree_interior_update_lock);
855 closure_put(&as->cl);
858 static void bch2_btree_update_get_open_buckets(struct btree_update *as, struct btree *b)
861 as->open_buckets[as->nr_open_buckets++] =
866 * @b is being split/rewritten: it may have pointers to not-yet-written btree
867 * nodes and thus outstanding btree_updates - redirect @b's
868 * btree_updates to point to this btree_update:
870 static void bch2_btree_interior_update_will_free_node(struct btree_update *as,
873 struct bch_fs *c = as->c;
874 struct btree_update *p, *n;
875 struct btree_write *w;
877 set_btree_node_dying(b);
879 if (btree_node_fake(b))
882 mutex_lock(&c->btree_interior_update_lock);
885 * Does this node have any btree_update operations preventing
886 * it from being written?
888 * If so, redirect them to point to this btree_update: we can
889 * write out our new nodes, but we won't make them visible until those
890 * operations complete
892 list_for_each_entry_safe(p, n, &b->write_blocked, write_blocked_list) {
893 list_del_init(&p->write_blocked_list);
894 btree_update_reparent(as, p);
897 * for flush_held_btree_writes() waiting on updates to flush or
898 * nodes to be writeable:
900 closure_wake_up(&c->btree_interior_update_wait);
903 clear_btree_node_dirty_acct(c, b);
904 clear_btree_node_need_write(b);
907 * Does this node have unwritten data that has a pin on the journal?
909 * If so, transfer that pin to the btree_update operation -
910 * note that if we're freeing multiple nodes, we only need to keep the
911 * oldest pin of any of the nodes we're freeing. We'll release the pin
912 * when the new nodes are persistent and reachable on disk:
914 w = btree_current_write(b);
915 bch2_journal_pin_copy(&c->journal, &as->journal, &w->journal, NULL);
916 bch2_journal_pin_drop(&c->journal, &w->journal);
918 w = btree_prev_write(b);
919 bch2_journal_pin_copy(&c->journal, &as->journal, &w->journal, NULL);
920 bch2_journal_pin_drop(&c->journal, &w->journal);
922 mutex_unlock(&c->btree_interior_update_lock);
925 * Is this a node that isn't reachable on disk yet?
927 * Nodes that aren't reachable yet have writes blocked until they're
928 * reachable - now that we've cancelled any pending writes and moved
929 * things waiting on that write to wait on this update, we can drop this
930 * node from the list of nodes that the other update is making
931 * reachable, prior to freeing it:
933 btree_update_drop_new_node(c, b);
935 btree_update_will_delete_key(as, &b->key);
937 as->old_nodes[as->nr_old_nodes] = b;
938 as->old_nodes_seq[as->nr_old_nodes] = b->data->keys.seq;
942 static void bch2_btree_update_done(struct btree_update *as)
944 struct bch_fs *c = as->c;
945 u64 start_time = as->start_time;
947 BUG_ON(as->mode == BTREE_INTERIOR_NO_UPDATE);
949 if (as->took_gc_lock)
950 up_read(&as->c->gc_lock);
951 as->took_gc_lock = false;
953 bch2_btree_reserve_put(as);
955 continue_at(&as->cl, btree_update_set_nodes_written,
956 as->c->btree_interior_update_worker);
958 bch2_time_stats_update(&c->times[BCH_TIME_btree_interior_update_foreground],
962 static struct btree_update *
963 bch2_btree_update_start(struct btree_trans *trans, struct btree_path *path,
964 unsigned level, bool split, unsigned flags)
966 struct bch_fs *c = trans->c;
967 struct btree_update *as;
968 u64 start_time = local_clock();
969 int disk_res_flags = (flags & BTREE_INSERT_NOFAIL)
970 ? BCH_DISK_RESERVATION_NOFAIL : 0;
971 unsigned nr_nodes[2] = { 0, 0 };
972 unsigned update_level = level;
973 int journal_flags = 0;
976 BUG_ON(!path->should_be_locked);
978 if (flags & BTREE_INSERT_JOURNAL_RESERVED)
979 journal_flags |= JOURNAL_RES_GET_RESERVED;
980 if (flags & BTREE_INSERT_JOURNAL_RECLAIM)
981 journal_flags |= JOURNAL_RES_GET_NONBLOCK;
984 nr_nodes[!!update_level] += 1 + split;
987 if (!btree_path_node(path, update_level))
991 * XXX: figure out how far we might need to split,
992 * instead of locking/reserving all the way to the root:
994 split = update_level + 1 < BTREE_MAX_DEPTH;
997 /* Might have to allocate a new root: */
998 if (update_level < BTREE_MAX_DEPTH)
1001 if (!bch2_btree_path_upgrade(trans, path, U8_MAX)) {
1002 trace_trans_restart_iter_upgrade(trans->fn, _RET_IP_,
1003 path->btree_id, &path->pos);
1004 ret = btree_trans_restart(trans);
1005 return ERR_PTR(ret);
1008 if (flags & BTREE_INSERT_GC_LOCK_HELD)
1009 lockdep_assert_held(&c->gc_lock);
1010 else if (!down_read_trylock(&c->gc_lock)) {
1011 bch2_trans_unlock(trans);
1012 down_read(&c->gc_lock);
1013 if (!bch2_trans_relock(trans)) {
1014 up_read(&c->gc_lock);
1015 return ERR_PTR(-EINTR);
1019 as = mempool_alloc(&c->btree_interior_update_pool, GFP_NOIO);
1020 memset(as, 0, sizeof(*as));
1021 closure_init(&as->cl, NULL);
1023 as->start_time = start_time;
1024 as->mode = BTREE_INTERIOR_NO_UPDATE;
1025 as->took_gc_lock = !(flags & BTREE_INSERT_GC_LOCK_HELD);
1026 as->btree_id = path->btree_id;
1027 INIT_LIST_HEAD(&as->list);
1028 INIT_LIST_HEAD(&as->unwritten_list);
1029 INIT_LIST_HEAD(&as->write_blocked_list);
1030 bch2_keylist_init(&as->old_keys, as->_old_keys);
1031 bch2_keylist_init(&as->new_keys, as->_new_keys);
1032 bch2_keylist_init(&as->parent_keys, as->inline_keys);
1034 mutex_lock(&c->btree_interior_update_lock);
1035 list_add_tail(&as->list, &c->btree_interior_update_list);
1036 mutex_unlock(&c->btree_interior_update_lock);
1039 * We don't want to allocate if we're in an error state, that can cause
1040 * deadlock on emergency shutdown due to open buckets getting stuck in
1041 * the btree_reserve_cache after allocator shutdown has cleared it out.
1042 * This check needs to come after adding us to the btree_interior_update
1043 * list but before calling bch2_btree_reserve_get, to synchronize with
1044 * __bch2_fs_read_only().
1046 ret = bch2_journal_error(&c->journal);
1050 bch2_trans_unlock(trans);
1052 ret = bch2_journal_preres_get(&c->journal, &as->journal_preres,
1053 BTREE_UPDATE_JOURNAL_RES,
1056 bch2_btree_update_free(as);
1057 trace_trans_restart_journal_preres_get(trans->fn, _RET_IP_);
1058 btree_trans_restart(trans);
1059 return ERR_PTR(ret);
1062 ret = bch2_disk_reservation_get(c, &as->disk_res,
1063 (nr_nodes[0] + nr_nodes[1]) * btree_sectors(c),
1064 c->opts.metadata_replicas,
1069 ret = bch2_btree_reserve_get(as, nr_nodes, flags);
1073 if (!bch2_trans_relock(trans)) {
1080 bch2_btree_update_free(as);
1081 return ERR_PTR(ret);
1084 /* Btree root updates: */
1086 static void bch2_btree_set_root_inmem(struct bch_fs *c, struct btree *b)
1088 /* Root nodes cannot be reaped */
1089 mutex_lock(&c->btree_cache.lock);
1090 list_del_init(&b->list);
1091 mutex_unlock(&c->btree_cache.lock);
1094 six_lock_pcpu_alloc(&b->c.lock);
1096 six_lock_pcpu_free(&b->c.lock);
1098 mutex_lock(&c->btree_root_lock);
1099 BUG_ON(btree_node_root(c, b) &&
1100 (b->c.level < btree_node_root(c, b)->c.level ||
1101 !btree_node_dying(btree_node_root(c, b))));
1103 btree_node_root(c, b) = b;
1104 mutex_unlock(&c->btree_root_lock);
1106 bch2_recalc_btree_reserve(c);
1110 * bch_btree_set_root - update the root in memory and on disk
1112 * To ensure forward progress, the current task must not be holding any
1113 * btree node write locks. However, you must hold an intent lock on the
1116 * Note: This allocates a journal entry but doesn't add any keys to
1117 * it. All the btree roots are part of every journal write, so there
1118 * is nothing new to be done. This just guarantees that there is a
1121 static void bch2_btree_set_root(struct btree_update *as,
1122 struct btree_trans *trans,
1123 struct btree_path *path,
1126 struct bch_fs *c = as->c;
1129 trace_btree_set_root(c, b);
1130 BUG_ON(!b->written);
1132 old = btree_node_root(c, b);
1135 * Ensure no one is using the old root while we switch to the
1138 bch2_btree_node_lock_write(trans, path, old);
1140 bch2_btree_set_root_inmem(c, b);
1142 btree_update_updated_root(as, b);
1145 * Unlock old root after new root is visible:
1147 * The new root isn't persistent, but that's ok: we still have
1148 * an intent lock on the new root, and any updates that would
1149 * depend on the new root would have to update the new root.
1151 bch2_btree_node_unlock_write(trans, path, old);
1154 /* Interior node updates: */
1156 static void bch2_insert_fixup_btree_ptr(struct btree_update *as,
1157 struct btree_trans *trans,
1158 struct btree_path *path,
1160 struct btree_node_iter *node_iter,
1161 struct bkey_i *insert)
1163 struct bch_fs *c = as->c;
1164 struct bkey_packed *k;
1165 const char *invalid;
1167 BUG_ON(insert->k.type == KEY_TYPE_btree_ptr_v2 &&
1168 !btree_ptr_sectors_written(insert));
1170 if (unlikely(!test_bit(JOURNAL_REPLAY_DONE, &c->journal.flags)))
1171 bch2_journal_key_overwritten(c, b->c.btree_id, b->c.level, insert->k.p);
1173 invalid = bch2_bkey_invalid(c, bkey_i_to_s_c(insert), btree_node_type(b)) ?:
1174 bch2_bkey_in_btree_node(b, bkey_i_to_s_c(insert));
1176 struct printbuf buf = PRINTBUF;
1178 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(insert));
1179 bch2_fs_inconsistent(c, "inserting invalid bkey %s: %s", buf.buf, invalid);
1180 printbuf_exit(&buf);
1184 BUG_ON(as->journal_u64s + jset_u64s(insert->k.u64s) >
1185 ARRAY_SIZE(as->journal_entries));
1188 journal_entry_set((void *) &as->journal_entries[as->journal_u64s],
1189 BCH_JSET_ENTRY_btree_keys,
1190 b->c.btree_id, b->c.level,
1191 insert, insert->k.u64s);
1193 while ((k = bch2_btree_node_iter_peek_all(node_iter, b)) &&
1194 bkey_iter_pos_cmp(b, k, &insert->k.p) < 0)
1195 bch2_btree_node_iter_advance(node_iter, b);
1197 bch2_btree_bset_insert_key(trans, path, b, node_iter, insert);
1198 set_btree_node_dirty_acct(c, b);
1199 set_btree_node_need_write(b);
1203 __bch2_btree_insert_keys_interior(struct btree_update *as,
1204 struct btree_trans *trans,
1205 struct btree_path *path,
1207 struct btree_node_iter node_iter,
1208 struct keylist *keys)
1210 struct bkey_i *insert = bch2_keylist_front(keys);
1211 struct bkey_packed *k;
1213 BUG_ON(btree_node_type(b) != BKEY_TYPE_btree);
1215 while ((k = bch2_btree_node_iter_prev_all(&node_iter, b)) &&
1216 (bkey_cmp_left_packed(b, k, &insert->k.p) >= 0))
1219 while (!bch2_keylist_empty(keys)) {
1220 bch2_insert_fixup_btree_ptr(as, trans, path, b,
1221 &node_iter, bch2_keylist_front(keys));
1222 bch2_keylist_pop_front(keys);
1227 * Move keys from n1 (original replacement node, now lower node) to n2 (higher
1230 static struct btree *__btree_split_node(struct btree_update *as,
1233 struct bkey_format_state s;
1234 size_t nr_packed = 0, nr_unpacked = 0;
1236 struct bset *set1, *set2;
1237 struct bkey_packed *k, *set2_start, *set2_end, *out, *prev = NULL;
1240 n2 = bch2_btree_node_alloc(as, n1->c.level);
1241 bch2_btree_update_add_new_node(as, n2);
1243 n2->data->max_key = n1->data->max_key;
1244 n2->data->format = n1->format;
1245 SET_BTREE_NODE_SEQ(n2->data, BTREE_NODE_SEQ(n1->data));
1246 n2->key.k.p = n1->key.k.p;
1248 set1 = btree_bset_first(n1);
1249 set2 = btree_bset_first(n2);
1252 * Has to be a linear search because we don't have an auxiliary
1257 struct bkey_packed *n = bkey_next(k);
1259 if (n == vstruct_last(set1))
1261 if (k->_data - set1->_data >= (le16_to_cpu(set1->u64s) * 3) / 5)
1275 set2_end = vstruct_last(set1);
1277 set1->u64s = cpu_to_le16((u64 *) set2_start - set1->_data);
1278 set_btree_bset_end(n1, n1->set);
1280 n1->nr.live_u64s = le16_to_cpu(set1->u64s);
1281 n1->nr.bset_u64s[0] = le16_to_cpu(set1->u64s);
1282 n1->nr.packed_keys = nr_packed;
1283 n1->nr.unpacked_keys = nr_unpacked;
1285 n1_pos = bkey_unpack_pos(n1, prev);
1286 if (as->c->sb.version < bcachefs_metadata_version_snapshot)
1287 n1_pos.snapshot = U32_MAX;
1289 btree_set_max(n1, n1_pos);
1290 btree_set_min(n2, bpos_successor(n1->key.k.p));
1292 bch2_bkey_format_init(&s);
1293 bch2_bkey_format_add_pos(&s, n2->data->min_key);
1294 bch2_bkey_format_add_pos(&s, n2->data->max_key);
1296 for (k = set2_start; k != set2_end; k = bkey_next(k)) {
1297 struct bkey uk = bkey_unpack_key(n1, k);
1298 bch2_bkey_format_add_key(&s, &uk);
1301 n2->data->format = bch2_bkey_format_done(&s);
1302 btree_node_set_format(n2, n2->data->format);
1305 memset(&n2->nr, 0, sizeof(n2->nr));
1307 for (k = set2_start; k != set2_end; k = bkey_next(k)) {
1308 BUG_ON(!bch2_bkey_transform(&n2->format, out, bkey_packed(k)
1309 ? &n1->format : &bch2_bkey_format_current, k));
1310 out->format = KEY_FORMAT_LOCAL_BTREE;
1311 btree_keys_account_key_add(&n2->nr, 0, out);
1312 out = bkey_next(out);
1315 set2->u64s = cpu_to_le16((u64 *) out - set2->_data);
1316 set_btree_bset_end(n2, n2->set);
1318 BUG_ON(!set1->u64s);
1319 BUG_ON(!set2->u64s);
1321 btree_node_reset_sib_u64s(n1);
1322 btree_node_reset_sib_u64s(n2);
1324 bch2_verify_btree_nr_keys(n1);
1325 bch2_verify_btree_nr_keys(n2);
1328 btree_node_interior_verify(as->c, n1);
1329 btree_node_interior_verify(as->c, n2);
1336 * For updates to interior nodes, we've got to do the insert before we split
1337 * because the stuff we're inserting has to be inserted atomically. Post split,
1338 * the keys might have to go in different nodes and the split would no longer be
1341 * Worse, if the insert is from btree node coalescing, if we do the insert after
1342 * we do the split (and pick the pivot) - the pivot we pick might be between
1343 * nodes that were coalesced, and thus in the middle of a child node post
1346 static void btree_split_insert_keys(struct btree_update *as,
1347 struct btree_trans *trans,
1348 struct btree_path *path,
1350 struct keylist *keys)
1352 struct btree_node_iter node_iter;
1353 struct bkey_i *k = bch2_keylist_front(keys);
1354 struct bkey_packed *src, *dst, *n;
1357 bch2_btree_node_iter_init(&node_iter, b, &k->k.p);
1359 __bch2_btree_insert_keys_interior(as, trans, path, b, node_iter, keys);
1362 * We can't tolerate whiteouts here - with whiteouts there can be
1363 * duplicate keys, and it would be rather bad if we picked a duplicate
1366 i = btree_bset_first(b);
1367 src = dst = i->start;
1368 while (src != vstruct_last(i)) {
1370 if (!bkey_deleted(src)) {
1371 memmove_u64s_down(dst, src, src->u64s);
1372 dst = bkey_next(dst);
1377 /* Also clear out the unwritten whiteouts area: */
1378 b->whiteout_u64s = 0;
1380 i->u64s = cpu_to_le16((u64 *) dst - i->_data);
1381 set_btree_bset_end(b, b->set);
1383 BUG_ON(b->nsets != 1 ||
1384 b->nr.live_u64s != le16_to_cpu(btree_bset_first(b)->u64s));
1386 btree_node_interior_verify(as->c, b);
1389 static void btree_split(struct btree_update *as, struct btree_trans *trans,
1390 struct btree_path *path, struct btree *b,
1391 struct keylist *keys, unsigned flags)
1393 struct bch_fs *c = as->c;
1394 struct btree *parent = btree_node_parent(path, b);
1395 struct btree *n1, *n2 = NULL, *n3 = NULL;
1396 u64 start_time = local_clock();
1398 BUG_ON(!parent && (b != btree_node_root(c, b)));
1399 BUG_ON(!btree_node_intent_locked(path, btree_node_root(c, b)->c.level));
1401 bch2_btree_interior_update_will_free_node(as, b);
1403 n1 = bch2_btree_node_alloc_replacement(as, b);
1404 bch2_btree_update_add_new_node(as, n1);
1407 btree_split_insert_keys(as, trans, path, n1, keys);
1409 if (bset_u64s(&n1->set[0]) > BTREE_SPLIT_THRESHOLD(c)) {
1410 trace_btree_split(c, b);
1412 n2 = __btree_split_node(as, n1);
1414 bch2_btree_build_aux_trees(n2);
1415 bch2_btree_build_aux_trees(n1);
1416 six_unlock_write(&n2->c.lock);
1417 six_unlock_write(&n1->c.lock);
1419 bch2_btree_node_write(c, n1, SIX_LOCK_intent, 0);
1420 bch2_btree_node_write(c, n2, SIX_LOCK_intent, 0);
1423 * Note that on recursive parent_keys == keys, so we
1424 * can't start adding new keys to parent_keys before emptying it
1425 * out (which we did with btree_split_insert_keys() above)
1427 bch2_keylist_add(&as->parent_keys, &n1->key);
1428 bch2_keylist_add(&as->parent_keys, &n2->key);
1431 /* Depth increases, make a new root */
1432 n3 = __btree_root_alloc(as, b->c.level + 1);
1434 n3->sib_u64s[0] = U16_MAX;
1435 n3->sib_u64s[1] = U16_MAX;
1437 btree_split_insert_keys(as, trans, path, n3, &as->parent_keys);
1439 bch2_btree_node_write(c, n3, SIX_LOCK_intent, 0);
1442 trace_btree_compact(c, b);
1444 bch2_btree_build_aux_trees(n1);
1445 six_unlock_write(&n1->c.lock);
1447 bch2_btree_node_write(c, n1, SIX_LOCK_intent, 0);
1450 bch2_keylist_add(&as->parent_keys, &n1->key);
1453 /* New nodes all written, now make them visible: */
1456 /* Split a non root node */
1457 bch2_btree_insert_node(as, trans, path, parent, &as->parent_keys, flags);
1459 bch2_btree_set_root(as, trans, path, n3);
1461 /* Root filled up but didn't need to be split */
1462 bch2_btree_set_root(as, trans, path, n1);
1465 bch2_btree_update_get_open_buckets(as, n1);
1467 bch2_btree_update_get_open_buckets(as, n2);
1469 bch2_btree_update_get_open_buckets(as, n3);
1471 /* Successful split, update the path to point to the new nodes: */
1473 six_lock_increment(&b->c.lock, SIX_LOCK_intent);
1475 bch2_trans_node_add(trans, n3);
1477 bch2_trans_node_add(trans, n2);
1478 bch2_trans_node_add(trans, n1);
1481 * The old node must be freed (in memory) _before_ unlocking the new
1482 * nodes - else another thread could re-acquire a read lock on the old
1483 * node after another thread has locked and updated the new node, thus
1484 * seeing stale data:
1486 bch2_btree_node_free_inmem(trans, b);
1489 six_unlock_intent(&n3->c.lock);
1491 six_unlock_intent(&n2->c.lock);
1492 six_unlock_intent(&n1->c.lock);
1494 bch2_trans_verify_locks(trans);
1496 bch2_time_stats_update(&c->times[n2
1497 ? BCH_TIME_btree_node_split
1498 : BCH_TIME_btree_node_compact],
1503 bch2_btree_insert_keys_interior(struct btree_update *as,
1504 struct btree_trans *trans,
1505 struct btree_path *path,
1507 struct keylist *keys)
1509 struct btree_path *linked;
1511 __bch2_btree_insert_keys_interior(as, trans, path, b,
1512 path->l[b->c.level].iter, keys);
1514 btree_update_updated_node(as, b);
1516 trans_for_each_path_with_node(trans, b, linked)
1517 bch2_btree_node_iter_peek(&linked->l[b->c.level].iter, b);
1519 bch2_trans_verify_paths(trans);
1523 * bch_btree_insert_node - insert bkeys into a given btree node
1525 * @iter: btree iterator
1526 * @keys: list of keys to insert
1527 * @hook: insert callback
1528 * @persistent: if not null, @persistent will wait on journal write
1530 * Inserts as many keys as it can into a given btree node, splitting it if full.
1531 * If a split occurred, this function will return early. This can only happen
1532 * for leaf nodes -- inserts into interior nodes have to be atomic.
1534 static void bch2_btree_insert_node(struct btree_update *as, struct btree_trans *trans,
1535 struct btree_path *path, struct btree *b,
1536 struct keylist *keys, unsigned flags)
1538 struct bch_fs *c = as->c;
1539 int old_u64s = le16_to_cpu(btree_bset_last(b)->u64s);
1540 int old_live_u64s = b->nr.live_u64s;
1541 int live_u64s_added, u64s_added;
1543 lockdep_assert_held(&c->gc_lock);
1544 BUG_ON(!btree_node_intent_locked(path, btree_node_root(c, b)->c.level));
1545 BUG_ON(!b->c.level);
1546 BUG_ON(!as || as->b);
1547 bch2_verify_keylist_sorted(keys);
1549 bch2_btree_node_lock_for_insert(trans, path, b);
1551 if (!bch2_btree_node_insert_fits(c, b, bch2_keylist_u64s(keys))) {
1552 bch2_btree_node_unlock_write(trans, path, b);
1556 btree_node_interior_verify(c, b);
1558 bch2_btree_insert_keys_interior(as, trans, path, b, keys);
1560 live_u64s_added = (int) b->nr.live_u64s - old_live_u64s;
1561 u64s_added = (int) le16_to_cpu(btree_bset_last(b)->u64s) - old_u64s;
1563 if (b->sib_u64s[0] != U16_MAX && live_u64s_added < 0)
1564 b->sib_u64s[0] = max(0, (int) b->sib_u64s[0] + live_u64s_added);
1565 if (b->sib_u64s[1] != U16_MAX && live_u64s_added < 0)
1566 b->sib_u64s[1] = max(0, (int) b->sib_u64s[1] + live_u64s_added);
1568 if (u64s_added > live_u64s_added &&
1569 bch2_maybe_compact_whiteouts(c, b))
1570 bch2_trans_node_reinit_iter(trans, b);
1572 bch2_btree_node_unlock_write(trans, path, b);
1574 btree_node_interior_verify(c, b);
1577 btree_split(as, trans, path, b, keys, flags);
1580 int bch2_btree_split_leaf(struct btree_trans *trans,
1581 struct btree_path *path,
1584 struct btree *b = path_l(path)->b;
1585 struct btree_update *as;
1589 as = bch2_btree_update_start(trans, path, path->level,
1594 btree_split(as, trans, path, b, NULL, flags);
1595 bch2_btree_update_done(as);
1597 for (l = path->level + 1; btree_path_node(path, l) && !ret; l++)
1598 ret = bch2_foreground_maybe_merge(trans, path, l, flags);
1603 int __bch2_foreground_maybe_merge(struct btree_trans *trans,
1604 struct btree_path *path,
1607 enum btree_node_sibling sib)
1609 struct bch_fs *c = trans->c;
1610 struct btree_path *sib_path = NULL;
1611 struct btree_update *as;
1612 struct bkey_format_state new_s;
1613 struct bkey_format new_f;
1614 struct bkey_i delete;
1615 struct btree *b, *m, *n, *prev, *next, *parent;
1616 struct bpos sib_pos;
1618 u64 start_time = local_clock();
1621 BUG_ON(!path->should_be_locked);
1622 BUG_ON(!btree_node_locked(path, level));
1624 b = path->l[level].b;
1626 if ((sib == btree_prev_sib && !bpos_cmp(b->data->min_key, POS_MIN)) ||
1627 (sib == btree_next_sib && !bpos_cmp(b->data->max_key, SPOS_MAX))) {
1628 b->sib_u64s[sib] = U16_MAX;
1632 sib_pos = sib == btree_prev_sib
1633 ? bpos_predecessor(b->data->min_key)
1634 : bpos_successor(b->data->max_key);
1636 sib_path = bch2_path_get(trans, path->btree_id, sib_pos,
1637 U8_MAX, level, BTREE_ITER_INTENT, _THIS_IP_);
1638 ret = bch2_btree_path_traverse(trans, sib_path, false);
1642 sib_path->should_be_locked = true;
1644 m = sib_path->l[level].b;
1646 if (btree_node_parent(path, b) !=
1647 btree_node_parent(sib_path, m)) {
1648 b->sib_u64s[sib] = U16_MAX;
1652 if (sib == btree_prev_sib) {
1660 if (bkey_cmp(bpos_successor(prev->data->max_key), next->data->min_key)) {
1661 struct printbuf buf1 = PRINTBUF, buf2 = PRINTBUF;
1663 bch2_bpos_to_text(&buf1, prev->data->max_key);
1664 bch2_bpos_to_text(&buf2, next->data->min_key);
1666 "btree topology error in btree merge:\n"
1667 " prev ends at %s\n"
1668 " next starts at %s",
1669 buf1.buf, buf2.buf);
1670 printbuf_exit(&buf1);
1671 printbuf_exit(&buf2);
1672 bch2_topology_error(c);
1677 bch2_bkey_format_init(&new_s);
1678 bch2_bkey_format_add_pos(&new_s, prev->data->min_key);
1679 __bch2_btree_calc_format(&new_s, prev);
1680 __bch2_btree_calc_format(&new_s, next);
1681 bch2_bkey_format_add_pos(&new_s, next->data->max_key);
1682 new_f = bch2_bkey_format_done(&new_s);
1684 sib_u64s = btree_node_u64s_with_format(b, &new_f) +
1685 btree_node_u64s_with_format(m, &new_f);
1687 if (sib_u64s > BTREE_FOREGROUND_MERGE_HYSTERESIS(c)) {
1688 sib_u64s -= BTREE_FOREGROUND_MERGE_HYSTERESIS(c);
1690 sib_u64s += BTREE_FOREGROUND_MERGE_HYSTERESIS(c);
1693 sib_u64s = min(sib_u64s, btree_max_u64s(c));
1694 sib_u64s = min(sib_u64s, (size_t) U16_MAX - 1);
1695 b->sib_u64s[sib] = sib_u64s;
1697 if (b->sib_u64s[sib] > c->btree_foreground_merge_threshold)
1700 parent = btree_node_parent(path, b);
1701 as = bch2_btree_update_start(trans, path, level, false,
1702 BTREE_INSERT_NOFAIL|
1703 BTREE_INSERT_USE_RESERVE|
1705 ret = PTR_ERR_OR_ZERO(as);
1709 trace_btree_merge(c, b);
1711 bch2_btree_interior_update_will_free_node(as, b);
1712 bch2_btree_interior_update_will_free_node(as, m);
1714 n = bch2_btree_node_alloc(as, b->c.level);
1715 bch2_btree_update_add_new_node(as, n);
1717 SET_BTREE_NODE_SEQ(n->data,
1718 max(BTREE_NODE_SEQ(b->data),
1719 BTREE_NODE_SEQ(m->data)) + 1);
1721 btree_set_min(n, prev->data->min_key);
1722 btree_set_max(n, next->data->max_key);
1723 n->data->format = new_f;
1725 btree_node_set_format(n, new_f);
1727 bch2_btree_sort_into(c, n, prev);
1728 bch2_btree_sort_into(c, n, next);
1730 bch2_btree_build_aux_trees(n);
1731 six_unlock_write(&n->c.lock);
1733 bch2_btree_node_write(c, n, SIX_LOCK_intent, 0);
1735 bkey_init(&delete.k);
1736 delete.k.p = prev->key.k.p;
1737 bch2_keylist_add(&as->parent_keys, &delete);
1738 bch2_keylist_add(&as->parent_keys, &n->key);
1740 bch2_trans_verify_paths(trans);
1742 bch2_btree_insert_node(as, trans, path, parent, &as->parent_keys, flags);
1744 bch2_trans_verify_paths(trans);
1746 bch2_btree_update_get_open_buckets(as, n);
1748 six_lock_increment(&b->c.lock, SIX_LOCK_intent);
1749 six_lock_increment(&m->c.lock, SIX_LOCK_intent);
1751 bch2_trans_node_add(trans, n);
1753 bch2_trans_verify_paths(trans);
1755 bch2_btree_node_free_inmem(trans, b);
1756 bch2_btree_node_free_inmem(trans, m);
1758 six_unlock_intent(&n->c.lock);
1760 bch2_btree_update_done(as);
1762 bch2_time_stats_update(&c->times[BCH_TIME_btree_node_merge], start_time);
1765 bch2_path_put(trans, sib_path, true);
1766 bch2_trans_verify_locks(trans);
1771 * bch_btree_node_rewrite - Rewrite/move a btree node
1773 int bch2_btree_node_rewrite(struct btree_trans *trans,
1774 struct btree_iter *iter,
1778 struct bch_fs *c = trans->c;
1779 struct btree *n, *parent;
1780 struct btree_update *as;
1783 flags |= BTREE_INSERT_NOFAIL;
1785 parent = btree_node_parent(iter->path, b);
1786 as = bch2_btree_update_start(trans, iter->path, b->c.level,
1788 ret = PTR_ERR_OR_ZERO(as);
1790 trace_btree_gc_rewrite_node_fail(c, b);
1794 bch2_btree_interior_update_will_free_node(as, b);
1796 n = bch2_btree_node_alloc_replacement(as, b);
1797 bch2_btree_update_add_new_node(as, n);
1799 bch2_btree_build_aux_trees(n);
1800 six_unlock_write(&n->c.lock);
1802 trace_btree_gc_rewrite_node(c, b);
1804 bch2_btree_node_write(c, n, SIX_LOCK_intent, 0);
1807 bch2_keylist_add(&as->parent_keys, &n->key);
1808 bch2_btree_insert_node(as, trans, iter->path, parent,
1809 &as->parent_keys, flags);
1811 bch2_btree_set_root(as, trans, iter->path, n);
1814 bch2_btree_update_get_open_buckets(as, n);
1816 six_lock_increment(&b->c.lock, SIX_LOCK_intent);
1817 bch2_trans_node_add(trans, n);
1818 bch2_btree_node_free_inmem(trans, b);
1819 six_unlock_intent(&n->c.lock);
1821 bch2_btree_update_done(as);
1823 bch2_btree_path_downgrade(iter->path);
1827 struct async_btree_rewrite {
1829 struct work_struct work;
1830 enum btree_id btree_id;
1836 static int async_btree_node_rewrite_trans(struct btree_trans *trans,
1837 struct async_btree_rewrite *a)
1839 struct btree_iter iter;
1843 bch2_trans_node_iter_init(trans, &iter, a->btree_id, a->pos,
1844 BTREE_MAX_DEPTH, a->level, 0);
1845 b = bch2_btree_iter_peek_node(&iter);
1846 ret = PTR_ERR_OR_ZERO(b);
1850 if (!b || b->data->keys.seq != a->seq)
1853 ret = bch2_btree_node_rewrite(trans, &iter, b, 0);
1855 bch2_trans_iter_exit(trans, &iter);
1860 void async_btree_node_rewrite_work(struct work_struct *work)
1862 struct async_btree_rewrite *a =
1863 container_of(work, struct async_btree_rewrite, work);
1864 struct bch_fs *c = a->c;
1866 bch2_trans_do(c, NULL, NULL, 0,
1867 async_btree_node_rewrite_trans(&trans, a));
1868 percpu_ref_put(&c->writes);
1872 void bch2_btree_node_rewrite_async(struct bch_fs *c, struct btree *b)
1874 struct async_btree_rewrite *a;
1876 if (!percpu_ref_tryget(&c->writes))
1879 a = kmalloc(sizeof(*a), GFP_NOFS);
1881 percpu_ref_put(&c->writes);
1886 a->btree_id = b->c.btree_id;
1887 a->level = b->c.level;
1888 a->pos = b->key.k.p;
1889 a->seq = b->data->keys.seq;
1891 INIT_WORK(&a->work, async_btree_node_rewrite_work);
1892 queue_work(c->btree_interior_update_worker, &a->work);
1895 static int __bch2_btree_node_update_key(struct btree_trans *trans,
1896 struct btree_iter *iter,
1897 struct btree *b, struct btree *new_hash,
1898 struct bkey_i *new_key,
1901 struct bch_fs *c = trans->c;
1902 struct btree_iter iter2 = { NULL };
1903 struct btree *parent;
1904 u64 journal_entries[BKEY_BTREE_PTR_U64s_MAX];
1907 if (!skip_triggers) {
1908 ret = bch2_trans_mark_new(trans, new_key, 0);
1912 ret = bch2_trans_mark_old(trans, bkey_i_to_s_c(&b->key), 0);
1918 bkey_copy(&new_hash->key, new_key);
1919 ret = bch2_btree_node_hash_insert(&c->btree_cache,
1920 new_hash, b->c.level, b->c.btree_id);
1924 parent = btree_node_parent(iter->path, b);
1926 bch2_trans_copy_iter(&iter2, iter);
1928 iter2.path = bch2_btree_path_make_mut(trans, iter2.path,
1929 iter2.flags & BTREE_ITER_INTENT,
1932 BUG_ON(iter2.path->level != b->c.level);
1933 BUG_ON(bpos_cmp(iter2.path->pos, new_key->k.p));
1935 btree_node_unlock(iter2.path, iter2.path->level);
1936 path_l(iter2.path)->b = BTREE_ITER_NO_NODE_UP;
1937 iter2.path->level++;
1939 bch2_btree_path_check_sort(trans, iter2.path, 0);
1941 ret = bch2_btree_iter_traverse(&iter2) ?:
1942 bch2_trans_update(trans, &iter2, new_key, BTREE_TRIGGER_NORUN);
1946 BUG_ON(btree_node_root(c, b) != b);
1948 trans->extra_journal_entries = (void *) &journal_entries[0];
1949 trans->extra_journal_entry_u64s =
1950 journal_entry_set((void *) &journal_entries[0],
1951 BCH_JSET_ENTRY_btree_root,
1952 b->c.btree_id, b->c.level,
1953 new_key, new_key->k.u64s);
1956 ret = bch2_trans_commit(trans, NULL, NULL,
1957 BTREE_INSERT_NOFAIL|
1958 BTREE_INSERT_NOCHECK_RW|
1959 BTREE_INSERT_USE_RESERVE|
1960 BTREE_INSERT_JOURNAL_RECLAIM|
1961 BTREE_INSERT_JOURNAL_RESERVED);
1965 bch2_btree_node_lock_write(trans, iter->path, b);
1968 mutex_lock(&c->btree_cache.lock);
1969 bch2_btree_node_hash_remove(&c->btree_cache, new_hash);
1970 bch2_btree_node_hash_remove(&c->btree_cache, b);
1972 bkey_copy(&b->key, new_key);
1973 ret = __bch2_btree_node_hash_insert(&c->btree_cache, b);
1975 mutex_unlock(&c->btree_cache.lock);
1977 bkey_copy(&b->key, new_key);
1980 bch2_btree_node_unlock_write(trans, iter->path, b);
1982 bch2_trans_iter_exit(trans, &iter2);
1986 mutex_lock(&c->btree_cache.lock);
1987 bch2_btree_node_hash_remove(&c->btree_cache, b);
1988 mutex_unlock(&c->btree_cache.lock);
1993 int bch2_btree_node_update_key(struct btree_trans *trans, struct btree_iter *iter,
1994 struct btree *b, struct bkey_i *new_key,
1997 struct bch_fs *c = trans->c;
1998 struct btree *new_hash = NULL;
1999 struct btree_path *path = iter->path;
2003 if (!btree_node_intent_locked(path, b->c.level) &&
2004 !bch2_btree_path_upgrade(trans, path, b->c.level + 1)) {
2005 btree_trans_restart(trans);
2009 closure_init_stack(&cl);
2012 * check btree_ptr_hash_val() after @b is locked by
2013 * btree_iter_traverse():
2015 if (btree_ptr_hash_val(new_key) != b->hash_val) {
2016 ret = bch2_btree_cache_cannibalize_lock(c, &cl);
2018 bch2_trans_unlock(trans);
2020 if (!bch2_trans_relock(trans))
2024 new_hash = bch2_btree_node_mem_alloc(c, false);
2028 ret = __bch2_btree_node_update_key(trans, iter, b, new_hash,
2029 new_key, skip_triggers);
2033 mutex_lock(&c->btree_cache.lock);
2034 list_move(&new_hash->list, &c->btree_cache.freeable);
2035 mutex_unlock(&c->btree_cache.lock);
2037 six_unlock_write(&new_hash->c.lock);
2038 six_unlock_intent(&new_hash->c.lock);
2041 bch2_btree_cache_cannibalize_unlock(c);
2045 int bch2_btree_node_update_key_get_iter(struct btree_trans *trans,
2046 struct btree *b, struct bkey_i *new_key,
2049 struct btree_iter iter;
2052 bch2_trans_node_iter_init(trans, &iter, b->c.btree_id, b->key.k.p,
2053 BTREE_MAX_DEPTH, b->c.level,
2055 ret = bch2_btree_iter_traverse(&iter);
2059 /* has node been freed? */
2060 if (iter.path->l[b->c.level].b != b) {
2061 /* node has been freed: */
2062 BUG_ON(!btree_node_dying(b));
2066 BUG_ON(!btree_node_hashed(b));
2068 ret = bch2_btree_node_update_key(trans, &iter, b, new_key, skip_triggers);
2070 bch2_trans_iter_exit(trans, &iter);
2077 * Only for filesystem bringup, when first reading the btree roots or allocating
2078 * btree roots when initializing a new filesystem:
2080 void bch2_btree_set_root_for_read(struct bch_fs *c, struct btree *b)
2082 BUG_ON(btree_node_root(c, b));
2084 bch2_btree_set_root_inmem(c, b);
2087 void bch2_btree_root_alloc(struct bch_fs *c, enum btree_id id)
2093 closure_init_stack(&cl);
2096 ret = bch2_btree_cache_cannibalize_lock(c, &cl);
2100 b = bch2_btree_node_mem_alloc(c, false);
2101 bch2_btree_cache_cannibalize_unlock(c);
2103 set_btree_node_fake(b);
2104 set_btree_node_need_rewrite(b);
2108 bkey_btree_ptr_init(&b->key);
2109 b->key.k.p = SPOS_MAX;
2110 *((u64 *) bkey_i_to_btree_ptr(&b->key)->v.start) = U64_MAX - id;
2112 bch2_bset_init_first(b, &b->data->keys);
2113 bch2_btree_build_aux_trees(b);
2116 btree_set_min(b, POS_MIN);
2117 btree_set_max(b, SPOS_MAX);
2118 b->data->format = bch2_btree_calc_format(b);
2119 btree_node_set_format(b, b->data->format);
2121 ret = bch2_btree_node_hash_insert(&c->btree_cache, b,
2122 b->c.level, b->c.btree_id);
2125 bch2_btree_set_root_inmem(c, b);
2127 six_unlock_write(&b->c.lock);
2128 six_unlock_intent(&b->c.lock);
2131 void bch2_btree_updates_to_text(struct printbuf *out, struct bch_fs *c)
2133 struct btree_update *as;
2135 mutex_lock(&c->btree_interior_update_lock);
2136 list_for_each_entry(as, &c->btree_interior_update_list, list)
2137 pr_buf(out, "%p m %u w %u r %u j %llu\n",
2141 atomic_read(&as->cl.remaining) & CLOSURE_REMAINING_MASK,
2143 mutex_unlock(&c->btree_interior_update_lock);
2146 size_t bch2_btree_interior_updates_nr_pending(struct bch_fs *c)
2149 struct list_head *i;
2151 mutex_lock(&c->btree_interior_update_lock);
2152 list_for_each(i, &c->btree_interior_update_list)
2154 mutex_unlock(&c->btree_interior_update_lock);
2159 void bch2_journal_entries_to_btree_roots(struct bch_fs *c, struct jset *jset)
2161 struct btree_root *r;
2162 struct jset_entry *entry;
2164 mutex_lock(&c->btree_root_lock);
2166 vstruct_for_each(jset, entry)
2167 if (entry->type == BCH_JSET_ENTRY_btree_root) {
2168 r = &c->btree_roots[entry->btree_id];
2169 r->level = entry->level;
2171 bkey_copy(&r->key, &entry->start[0]);
2174 mutex_unlock(&c->btree_root_lock);
2178 bch2_btree_roots_to_journal_entries(struct bch_fs *c,
2179 struct jset_entry *start,
2180 struct jset_entry *end)
2182 struct jset_entry *entry;
2183 unsigned long have = 0;
2186 for (entry = start; entry < end; entry = vstruct_next(entry))
2187 if (entry->type == BCH_JSET_ENTRY_btree_root)
2188 __set_bit(entry->btree_id, &have);
2190 mutex_lock(&c->btree_root_lock);
2192 for (i = 0; i < BTREE_ID_NR; i++)
2193 if (c->btree_roots[i].alive && !test_bit(i, &have)) {
2194 journal_entry_set(end,
2195 BCH_JSET_ENTRY_btree_root,
2196 i, c->btree_roots[i].level,
2197 &c->btree_roots[i].key,
2198 c->btree_roots[i].key.u64s);
2199 end = vstruct_next(end);
2202 mutex_unlock(&c->btree_root_lock);
2207 void bch2_fs_btree_interior_update_exit(struct bch_fs *c)
2209 if (c->btree_interior_update_worker)
2210 destroy_workqueue(c->btree_interior_update_worker);
2211 mempool_exit(&c->btree_interior_update_pool);
2214 int bch2_fs_btree_interior_update_init(struct bch_fs *c)
2216 mutex_init(&c->btree_reserve_cache_lock);
2217 INIT_LIST_HEAD(&c->btree_interior_update_list);
2218 INIT_LIST_HEAD(&c->btree_interior_updates_unwritten);
2219 mutex_init(&c->btree_interior_update_lock);
2220 INIT_WORK(&c->btree_interior_update_work, btree_interior_update_work);
2222 c->btree_interior_update_worker =
2223 alloc_workqueue("btree_update", WQ_UNBOUND|WQ_MEM_RECLAIM, 1);
2224 if (!c->btree_interior_update_worker)
2227 return mempool_init_kmalloc_pool(&c->btree_interior_update_pool, 1,
2228 sizeof(struct btree_update));