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 for (p = as->prealloc_nodes;
385 p < as->prealloc_nodes + ARRAY_SIZE(as->prealloc_nodes);
388 struct btree *b = p->b[--p->nr];
390 mutex_lock(&c->btree_reserve_cache_lock);
392 if (c->btree_reserve_cache_nr <
393 ARRAY_SIZE(c->btree_reserve_cache)) {
394 struct btree_alloc *a =
395 &c->btree_reserve_cache[c->btree_reserve_cache_nr++];
399 bkey_copy(&a->k, &b->key);
401 bch2_open_buckets_put(c, &b->ob);
404 mutex_unlock(&c->btree_reserve_cache_lock);
406 six_lock_intent(&b->c.lock, NULL, NULL);
407 six_lock_write(&b->c.lock, NULL, NULL);
408 __btree_node_free(c, b);
409 six_unlock_write(&b->c.lock);
410 six_unlock_intent(&b->c.lock);
415 static int bch2_btree_reserve_get(struct btree_update *as,
416 unsigned nr_nodes[2],
419 struct bch_fs *c = as->c;
425 closure_init_stack(&cl);
428 BUG_ON(nr_nodes[0] + nr_nodes[1] > BTREE_RESERVE_MAX);
431 * Protects reaping from the btree node cache and using the btree node
432 * open bucket reserve:
434 * BTREE_INSERT_NOWAIT only applies to btree node allocation, not
435 * blocking on this lock:
437 ret = bch2_btree_cache_cannibalize_lock(c, &cl);
441 for (interior = 0; interior < 2; interior++) {
442 struct prealloc_nodes *p = as->prealloc_nodes + interior;
444 while (p->nr < nr_nodes[interior]) {
445 b = __bch2_btree_node_alloc(c, &as->disk_res,
446 flags & BTREE_INSERT_NOWAIT
458 bch2_btree_cache_cannibalize_unlock(c);
462 bch2_btree_cache_cannibalize_unlock(c);
468 trace_btree_reserve_get_fail(c, nr_nodes[0] + nr_nodes[1], &cl);
472 /* Asynchronous interior node update machinery */
474 static void bch2_btree_update_free(struct btree_update *as)
476 struct bch_fs *c = as->c;
478 if (as->took_gc_lock)
479 up_read(&c->gc_lock);
480 as->took_gc_lock = false;
482 bch2_journal_preres_put(&c->journal, &as->journal_preres);
484 bch2_journal_pin_drop(&c->journal, &as->journal);
485 bch2_journal_pin_flush(&c->journal, &as->journal);
486 bch2_disk_reservation_put(c, &as->disk_res);
487 bch2_btree_reserve_put(as);
489 bch2_time_stats_update(&c->times[BCH_TIME_btree_interior_update_total],
492 mutex_lock(&c->btree_interior_update_lock);
493 list_del(&as->unwritten_list);
496 closure_debug_destroy(&as->cl);
497 mempool_free(as, &c->btree_interior_update_pool);
500 * Have to do the wakeup with btree_interior_update_lock still held,
501 * since being on btree_interior_update_list is our ref on @c:
503 closure_wake_up(&c->btree_interior_update_wait);
505 mutex_unlock(&c->btree_interior_update_lock);
508 static void btree_update_add_key(struct btree_update *as,
509 struct keylist *keys, struct btree *b)
511 struct bkey_i *k = &b->key;
513 BUG_ON(bch2_keylist_u64s(keys) + k->k.u64s >
514 ARRAY_SIZE(as->_old_keys));
516 bkey_copy(keys->top, k);
517 bkey_i_to_btree_ptr_v2(keys->top)->v.mem_ptr = b->c.level + 1;
519 bch2_keylist_push(keys);
523 * The transactional part of an interior btree node update, where we journal the
524 * update we did to the interior node and update alloc info:
526 static int btree_update_nodes_written_trans(struct btree_trans *trans,
527 struct btree_update *as)
532 ret = darray_make_room(&trans->extra_journal_entries, as->journal_u64s);
536 memcpy(&darray_top(trans->extra_journal_entries),
538 as->journal_u64s * sizeof(u64));
539 trans->extra_journal_entries.nr += as->journal_u64s;
541 trans->journal_pin = &as->journal;
543 for_each_keylist_key(&as->old_keys, k) {
544 unsigned level = bkey_i_to_btree_ptr_v2(k)->v.mem_ptr;
546 ret = bch2_trans_mark_old(trans, as->btree_id, level, bkey_i_to_s_c(k), 0);
551 for_each_keylist_key(&as->new_keys, k) {
552 unsigned level = bkey_i_to_btree_ptr_v2(k)->v.mem_ptr;
554 ret = bch2_trans_mark_new(trans, as->btree_id, level, k, 0);
562 static void btree_update_nodes_written(struct btree_update *as)
564 struct bch_fs *c = as->c;
565 struct btree *b = as->b;
566 struct btree_trans trans;
572 * If we're already in an error state, it might be because a btree node
573 * was never written, and we might be trying to free that same btree
574 * node here, but it won't have been marked as allocated and we'll see
575 * spurious disk usage inconsistencies in the transactional part below
576 * if we don't skip it:
578 ret = bch2_journal_error(&c->journal);
583 * Wait for any in flight writes to finish before we free the old nodes
586 for (i = 0; i < as->nr_old_nodes; i++) {
587 struct btree *old = as->old_nodes[i];
590 six_lock_read(&old->c.lock, NULL, NULL);
591 seq = old->data ? old->data->keys.seq : 0;
592 six_unlock_read(&old->c.lock);
594 if (seq == as->old_nodes_seq[i])
595 wait_on_bit_io(&old->flags, BTREE_NODE_write_in_flight_inner,
596 TASK_UNINTERRUPTIBLE);
600 * We did an update to a parent node where the pointers we added pointed
601 * to child nodes that weren't written yet: now, the child nodes have
602 * been written so we can write out the update to the interior node.
606 * We can't call into journal reclaim here: we'd block on the journal
607 * reclaim lock, but we may need to release the open buckets we have
608 * pinned in order for other btree updates to make forward progress, and
609 * journal reclaim does btree updates when flushing bkey_cached entries,
610 * which may require allocations as well.
612 bch2_trans_init(&trans, c, 0, 512);
613 ret = __bch2_trans_do(&trans, &as->disk_res, &journal_seq,
615 BTREE_INSERT_NOCHECK_RW|
616 BTREE_INSERT_JOURNAL_RECLAIM|
617 JOURNAL_WATERMARK_reserved,
618 btree_update_nodes_written_trans(&trans, as));
619 bch2_trans_exit(&trans);
621 bch2_fs_fatal_err_on(ret && !bch2_journal_error(&c->journal), c,
622 "error %i in btree_update_nodes_written()", ret);
626 * @b is the node we did the final insert into:
628 * On failure to get a journal reservation, we still have to
629 * unblock the write and allow most of the write path to happen
630 * so that shutdown works, but the i->journal_seq mechanism
631 * won't work to prevent the btree write from being visible (we
632 * didn't get a journal sequence number) - instead
633 * __bch2_btree_node_write() doesn't do the actual write if
634 * we're in journal error state:
637 six_lock_intent(&b->c.lock, NULL, NULL);
638 six_lock_write(&b->c.lock, NULL, NULL);
639 mutex_lock(&c->btree_interior_update_lock);
641 list_del(&as->write_blocked_list);
642 if (list_empty(&b->write_blocked))
643 clear_btree_node_write_blocked(b);
646 * Node might have been freed, recheck under
647 * btree_interior_update_lock:
650 struct bset *i = btree_bset_last(b);
653 BUG_ON(!btree_node_dirty(b));
656 i->journal_seq = cpu_to_le64(
658 le64_to_cpu(i->journal_seq)));
660 bch2_btree_add_journal_pin(c, b, journal_seq);
663 * If we didn't get a journal sequence number we
664 * can't write this btree node, because recovery
665 * won't know to ignore this write:
667 set_btree_node_never_write(b);
671 mutex_unlock(&c->btree_interior_update_lock);
672 six_unlock_write(&b->c.lock);
674 btree_node_write_if_need(c, b, SIX_LOCK_intent);
675 six_unlock_intent(&b->c.lock);
678 bch2_journal_pin_drop(&c->journal, &as->journal);
680 bch2_journal_preres_put(&c->journal, &as->journal_preres);
682 mutex_lock(&c->btree_interior_update_lock);
683 for (i = 0; i < as->nr_new_nodes; i++) {
684 b = as->new_nodes[i];
686 BUG_ON(b->will_make_reachable != (unsigned long) as);
687 b->will_make_reachable = 0;
688 clear_btree_node_will_make_reachable(b);
690 mutex_unlock(&c->btree_interior_update_lock);
692 for (i = 0; i < as->nr_new_nodes; i++) {
693 b = as->new_nodes[i];
695 six_lock_read(&b->c.lock, NULL, NULL);
696 btree_node_write_if_need(c, b, SIX_LOCK_read);
697 six_unlock_read(&b->c.lock);
700 for (i = 0; i < as->nr_open_buckets; i++)
701 bch2_open_bucket_put(c, c->open_buckets + as->open_buckets[i]);
703 bch2_btree_update_free(as);
706 static void btree_interior_update_work(struct work_struct *work)
709 container_of(work, struct bch_fs, btree_interior_update_work);
710 struct btree_update *as;
713 mutex_lock(&c->btree_interior_update_lock);
714 as = list_first_entry_or_null(&c->btree_interior_updates_unwritten,
715 struct btree_update, unwritten_list);
716 if (as && !as->nodes_written)
718 mutex_unlock(&c->btree_interior_update_lock);
723 btree_update_nodes_written(as);
727 static void btree_update_set_nodes_written(struct closure *cl)
729 struct btree_update *as = container_of(cl, struct btree_update, cl);
730 struct bch_fs *c = as->c;
732 mutex_lock(&c->btree_interior_update_lock);
733 as->nodes_written = true;
734 mutex_unlock(&c->btree_interior_update_lock);
736 queue_work(c->btree_interior_update_worker, &c->btree_interior_update_work);
740 * We're updating @b with pointers to nodes that haven't finished writing yet:
741 * block @b from being written until @as completes
743 static void btree_update_updated_node(struct btree_update *as, struct btree *b)
745 struct bch_fs *c = as->c;
747 mutex_lock(&c->btree_interior_update_lock);
748 list_add_tail(&as->unwritten_list, &c->btree_interior_updates_unwritten);
750 BUG_ON(as->mode != BTREE_INTERIOR_NO_UPDATE);
751 BUG_ON(!btree_node_dirty(b));
753 as->mode = BTREE_INTERIOR_UPDATING_NODE;
756 set_btree_node_write_blocked(b);
757 list_add(&as->write_blocked_list, &b->write_blocked);
759 mutex_unlock(&c->btree_interior_update_lock);
762 static void btree_update_reparent(struct btree_update *as,
763 struct btree_update *child)
765 struct bch_fs *c = as->c;
767 lockdep_assert_held(&c->btree_interior_update_lock);
770 child->mode = BTREE_INTERIOR_UPDATING_AS;
772 bch2_journal_pin_copy(&c->journal, &as->journal, &child->journal, NULL);
775 static void btree_update_updated_root(struct btree_update *as, struct btree *b)
777 struct bkey_i *insert = &b->key;
778 struct bch_fs *c = as->c;
780 BUG_ON(as->mode != BTREE_INTERIOR_NO_UPDATE);
782 BUG_ON(as->journal_u64s + jset_u64s(insert->k.u64s) >
783 ARRAY_SIZE(as->journal_entries));
786 journal_entry_set((void *) &as->journal_entries[as->journal_u64s],
787 BCH_JSET_ENTRY_btree_root,
788 b->c.btree_id, b->c.level,
789 insert, insert->k.u64s);
791 mutex_lock(&c->btree_interior_update_lock);
792 list_add_tail(&as->unwritten_list, &c->btree_interior_updates_unwritten);
794 as->mode = BTREE_INTERIOR_UPDATING_ROOT;
795 mutex_unlock(&c->btree_interior_update_lock);
799 * bch2_btree_update_add_new_node:
801 * This causes @as to wait on @b to be written, before it gets to
802 * bch2_btree_update_nodes_written
804 * Additionally, it sets b->will_make_reachable to prevent any additional writes
805 * to @b from happening besides the first until @b is reachable on disk
807 * And it adds @b to the list of @as's new nodes, so that we can update sector
808 * counts in bch2_btree_update_nodes_written:
810 static void bch2_btree_update_add_new_node(struct btree_update *as, struct btree *b)
812 struct bch_fs *c = as->c;
814 closure_get(&as->cl);
816 mutex_lock(&c->btree_interior_update_lock);
817 BUG_ON(as->nr_new_nodes >= ARRAY_SIZE(as->new_nodes));
818 BUG_ON(b->will_make_reachable);
820 as->new_nodes[as->nr_new_nodes++] = b;
821 b->will_make_reachable = 1UL|(unsigned long) as;
822 set_btree_node_will_make_reachable(b);
824 mutex_unlock(&c->btree_interior_update_lock);
826 btree_update_add_key(as, &as->new_keys, b);
830 * returns true if @b was a new node
832 static void btree_update_drop_new_node(struct bch_fs *c, struct btree *b)
834 struct btree_update *as;
838 mutex_lock(&c->btree_interior_update_lock);
840 * When b->will_make_reachable != 0, it owns a ref on as->cl that's
841 * dropped when it gets written by bch2_btree_complete_write - the
842 * xchg() is for synchronization with bch2_btree_complete_write:
844 v = xchg(&b->will_make_reachable, 0);
845 clear_btree_node_will_make_reachable(b);
846 as = (struct btree_update *) (v & ~1UL);
849 mutex_unlock(&c->btree_interior_update_lock);
853 for (i = 0; i < as->nr_new_nodes; i++)
854 if (as->new_nodes[i] == b)
859 array_remove_item(as->new_nodes, as->nr_new_nodes, i);
860 mutex_unlock(&c->btree_interior_update_lock);
863 closure_put(&as->cl);
866 static void bch2_btree_update_get_open_buckets(struct btree_update *as, struct btree *b)
869 as->open_buckets[as->nr_open_buckets++] =
874 * @b is being split/rewritten: it may have pointers to not-yet-written btree
875 * nodes and thus outstanding btree_updates - redirect @b's
876 * btree_updates to point to this btree_update:
878 static void bch2_btree_interior_update_will_free_node(struct btree_update *as,
881 struct bch_fs *c = as->c;
882 struct btree_update *p, *n;
883 struct btree_write *w;
885 set_btree_node_dying(b);
887 if (btree_node_fake(b))
890 mutex_lock(&c->btree_interior_update_lock);
893 * Does this node have any btree_update operations preventing
894 * it from being written?
896 * If so, redirect them to point to this btree_update: we can
897 * write out our new nodes, but we won't make them visible until those
898 * operations complete
900 list_for_each_entry_safe(p, n, &b->write_blocked, write_blocked_list) {
901 list_del_init(&p->write_blocked_list);
902 btree_update_reparent(as, p);
905 * for flush_held_btree_writes() waiting on updates to flush or
906 * nodes to be writeable:
908 closure_wake_up(&c->btree_interior_update_wait);
911 clear_btree_node_dirty_acct(c, b);
912 clear_btree_node_need_write(b);
915 * Does this node have unwritten data that has a pin on the journal?
917 * If so, transfer that pin to the btree_update operation -
918 * note that if we're freeing multiple nodes, we only need to keep the
919 * oldest pin of any of the nodes we're freeing. We'll release the pin
920 * when the new nodes are persistent and reachable on disk:
922 w = btree_current_write(b);
923 bch2_journal_pin_copy(&c->journal, &as->journal, &w->journal, NULL);
924 bch2_journal_pin_drop(&c->journal, &w->journal);
926 w = btree_prev_write(b);
927 bch2_journal_pin_copy(&c->journal, &as->journal, &w->journal, NULL);
928 bch2_journal_pin_drop(&c->journal, &w->journal);
930 mutex_unlock(&c->btree_interior_update_lock);
933 * Is this a node that isn't reachable on disk yet?
935 * Nodes that aren't reachable yet have writes blocked until they're
936 * reachable - now that we've cancelled any pending writes and moved
937 * things waiting on that write to wait on this update, we can drop this
938 * node from the list of nodes that the other update is making
939 * reachable, prior to freeing it:
941 btree_update_drop_new_node(c, b);
943 btree_update_add_key(as, &as->old_keys, b);
945 as->old_nodes[as->nr_old_nodes] = b;
946 as->old_nodes_seq[as->nr_old_nodes] = b->data->keys.seq;
950 static void bch2_btree_update_done(struct btree_update *as)
952 struct bch_fs *c = as->c;
953 u64 start_time = as->start_time;
955 BUG_ON(as->mode == BTREE_INTERIOR_NO_UPDATE);
957 if (as->took_gc_lock)
958 up_read(&as->c->gc_lock);
959 as->took_gc_lock = false;
961 bch2_btree_reserve_put(as);
963 continue_at(&as->cl, btree_update_set_nodes_written,
964 as->c->btree_interior_update_worker);
966 bch2_time_stats_update(&c->times[BCH_TIME_btree_interior_update_foreground],
970 static struct btree_update *
971 bch2_btree_update_start(struct btree_trans *trans, struct btree_path *path,
972 unsigned level, bool split, unsigned flags)
974 struct bch_fs *c = trans->c;
975 struct btree_update *as;
976 u64 start_time = local_clock();
977 int disk_res_flags = (flags & BTREE_INSERT_NOFAIL)
978 ? BCH_DISK_RESERVATION_NOFAIL : 0;
979 unsigned nr_nodes[2] = { 0, 0 };
980 unsigned update_level = level;
981 int journal_flags = flags & JOURNAL_WATERMARK_MASK;
984 BUG_ON(!path->should_be_locked);
986 if (flags & BTREE_INSERT_JOURNAL_RECLAIM)
987 journal_flags |= JOURNAL_RES_GET_NONBLOCK;
990 nr_nodes[!!update_level] += 1 + split;
993 if (!btree_path_node(path, update_level))
997 * XXX: figure out how far we might need to split,
998 * instead of locking/reserving all the way to the root:
1000 split = update_level + 1 < BTREE_MAX_DEPTH;
1003 /* Might have to allocate a new root: */
1004 if (update_level < BTREE_MAX_DEPTH)
1007 if (!bch2_btree_path_upgrade(trans, path, U8_MAX)) {
1008 trace_trans_restart_iter_upgrade(trans->fn, _RET_IP_,
1009 path->btree_id, &path->pos);
1010 ret = btree_trans_restart(trans);
1011 return ERR_PTR(ret);
1014 if (flags & BTREE_INSERT_GC_LOCK_HELD)
1015 lockdep_assert_held(&c->gc_lock);
1016 else if (!down_read_trylock(&c->gc_lock)) {
1017 bch2_trans_unlock(trans);
1018 down_read(&c->gc_lock);
1019 if (!bch2_trans_relock(trans)) {
1020 up_read(&c->gc_lock);
1021 return ERR_PTR(-EINTR);
1025 as = mempool_alloc(&c->btree_interior_update_pool, GFP_NOIO);
1026 memset(as, 0, sizeof(*as));
1027 closure_init(&as->cl, NULL);
1029 as->start_time = start_time;
1030 as->mode = BTREE_INTERIOR_NO_UPDATE;
1031 as->took_gc_lock = !(flags & BTREE_INSERT_GC_LOCK_HELD);
1032 as->btree_id = path->btree_id;
1033 INIT_LIST_HEAD(&as->list);
1034 INIT_LIST_HEAD(&as->unwritten_list);
1035 INIT_LIST_HEAD(&as->write_blocked_list);
1036 bch2_keylist_init(&as->old_keys, as->_old_keys);
1037 bch2_keylist_init(&as->new_keys, as->_new_keys);
1038 bch2_keylist_init(&as->parent_keys, as->inline_keys);
1040 mutex_lock(&c->btree_interior_update_lock);
1041 list_add_tail(&as->list, &c->btree_interior_update_list);
1042 mutex_unlock(&c->btree_interior_update_lock);
1045 * We don't want to allocate if we're in an error state, that can cause
1046 * deadlock on emergency shutdown due to open buckets getting stuck in
1047 * the btree_reserve_cache after allocator shutdown has cleared it out.
1048 * This check needs to come after adding us to the btree_interior_update
1049 * list but before calling bch2_btree_reserve_get, to synchronize with
1050 * __bch2_fs_read_only().
1052 ret = bch2_journal_error(&c->journal);
1056 bch2_trans_unlock(trans);
1058 ret = bch2_journal_preres_get(&c->journal, &as->journal_preres,
1059 BTREE_UPDATE_JOURNAL_RES,
1062 bch2_btree_update_free(as);
1063 trace_trans_restart_journal_preres_get(trans->fn, _RET_IP_);
1064 btree_trans_restart(trans);
1065 return ERR_PTR(ret);
1068 ret = bch2_disk_reservation_get(c, &as->disk_res,
1069 (nr_nodes[0] + nr_nodes[1]) * btree_sectors(c),
1070 c->opts.metadata_replicas,
1075 ret = bch2_btree_reserve_get(as, nr_nodes, flags);
1079 if (!bch2_trans_relock(trans)) {
1086 bch2_btree_update_free(as);
1087 return ERR_PTR(ret);
1090 /* Btree root updates: */
1092 static void bch2_btree_set_root_inmem(struct bch_fs *c, struct btree *b)
1094 /* Root nodes cannot be reaped */
1095 mutex_lock(&c->btree_cache.lock);
1096 list_del_init(&b->list);
1097 mutex_unlock(&c->btree_cache.lock);
1099 mutex_lock(&c->btree_root_lock);
1100 BUG_ON(btree_node_root(c, b) &&
1101 (b->c.level < btree_node_root(c, b)->c.level ||
1102 !btree_node_dying(btree_node_root(c, b))));
1104 btree_node_root(c, b) = b;
1105 mutex_unlock(&c->btree_root_lock);
1107 bch2_recalc_btree_reserve(c);
1111 * bch_btree_set_root - update the root in memory and on disk
1113 * To ensure forward progress, the current task must not be holding any
1114 * btree node write locks. However, you must hold an intent lock on the
1117 * Note: This allocates a journal entry but doesn't add any keys to
1118 * it. All the btree roots are part of every journal write, so there
1119 * is nothing new to be done. This just guarantees that there is a
1122 static void bch2_btree_set_root(struct btree_update *as,
1123 struct btree_trans *trans,
1124 struct btree_path *path,
1127 struct bch_fs *c = as->c;
1130 trace_btree_set_root(c, b);
1131 BUG_ON(!b->written);
1133 old = btree_node_root(c, b);
1136 * Ensure no one is using the old root while we switch to the
1139 bch2_btree_node_lock_write(trans, path, old);
1141 bch2_btree_set_root_inmem(c, b);
1143 btree_update_updated_root(as, b);
1146 * Unlock old root after new root is visible:
1148 * The new root isn't persistent, but that's ok: we still have
1149 * an intent lock on the new root, and any updates that would
1150 * depend on the new root would have to update the new root.
1152 bch2_btree_node_unlock_write(trans, path, old);
1155 /* Interior node updates: */
1157 static void bch2_insert_fixup_btree_ptr(struct btree_update *as,
1158 struct btree_trans *trans,
1159 struct btree_path *path,
1161 struct btree_node_iter *node_iter,
1162 struct bkey_i *insert)
1164 struct bch_fs *c = as->c;
1165 struct bkey_packed *k;
1166 struct printbuf buf = PRINTBUF;
1168 BUG_ON(insert->k.type == KEY_TYPE_btree_ptr_v2 &&
1169 !btree_ptr_sectors_written(insert));
1171 if (unlikely(!test_bit(JOURNAL_REPLAY_DONE, &c->journal.flags)))
1172 bch2_journal_key_overwritten(c, b->c.btree_id, b->c.level, insert->k.p);
1174 if (bch2_bkey_invalid(c, bkey_i_to_s_c(insert),
1175 btree_node_type(b), WRITE, &buf) ?:
1176 bch2_bkey_in_btree_node(b, bkey_i_to_s_c(insert), &buf)) {
1177 printbuf_reset(&buf);
1178 prt_printf(&buf, "inserting invalid bkey\n ");
1179 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(insert));
1180 prt_printf(&buf, "\n ");
1181 bch2_bkey_invalid(c, bkey_i_to_s_c(insert),
1182 btree_node_type(b), WRITE, &buf);
1183 bch2_bkey_in_btree_node(b, bkey_i_to_s_c(insert), &buf);
1185 bch2_fs_inconsistent(c, "%s", buf.buf);
1189 BUG_ON(as->journal_u64s + jset_u64s(insert->k.u64s) >
1190 ARRAY_SIZE(as->journal_entries));
1193 journal_entry_set((void *) &as->journal_entries[as->journal_u64s],
1194 BCH_JSET_ENTRY_btree_keys,
1195 b->c.btree_id, b->c.level,
1196 insert, insert->k.u64s);
1198 while ((k = bch2_btree_node_iter_peek_all(node_iter, b)) &&
1199 bkey_iter_pos_cmp(b, k, &insert->k.p) < 0)
1200 bch2_btree_node_iter_advance(node_iter, b);
1202 bch2_btree_bset_insert_key(trans, path, b, node_iter, insert);
1203 set_btree_node_dirty_acct(c, b);
1204 set_btree_node_need_write(b);
1206 printbuf_exit(&buf);
1210 __bch2_btree_insert_keys_interior(struct btree_update *as,
1211 struct btree_trans *trans,
1212 struct btree_path *path,
1214 struct btree_node_iter node_iter,
1215 struct keylist *keys)
1217 struct bkey_i *insert = bch2_keylist_front(keys);
1218 struct bkey_packed *k;
1220 BUG_ON(btree_node_type(b) != BKEY_TYPE_btree);
1222 while ((k = bch2_btree_node_iter_prev_all(&node_iter, b)) &&
1223 (bkey_cmp_left_packed(b, k, &insert->k.p) >= 0))
1226 while (!bch2_keylist_empty(keys)) {
1227 bch2_insert_fixup_btree_ptr(as, trans, path, b,
1228 &node_iter, bch2_keylist_front(keys));
1229 bch2_keylist_pop_front(keys);
1234 * Move keys from n1 (original replacement node, now lower node) to n2 (higher
1237 static struct btree *__btree_split_node(struct btree_update *as,
1240 struct bkey_format_state s;
1241 size_t nr_packed = 0, nr_unpacked = 0;
1243 struct bset *set1, *set2;
1244 struct bkey_packed *k, *set2_start, *set2_end, *out, *prev = NULL;
1247 n2 = bch2_btree_node_alloc(as, n1->c.level);
1249 n2->data->max_key = n1->data->max_key;
1250 n2->data->format = n1->format;
1251 SET_BTREE_NODE_SEQ(n2->data, BTREE_NODE_SEQ(n1->data));
1252 n2->key.k.p = n1->key.k.p;
1254 bch2_btree_update_add_new_node(as, n2);
1256 set1 = btree_bset_first(n1);
1257 set2 = btree_bset_first(n2);
1260 * Has to be a linear search because we don't have an auxiliary
1265 struct bkey_packed *n = bkey_next(k);
1267 if (n == vstruct_last(set1))
1269 if (k->_data - set1->_data >= (le16_to_cpu(set1->u64s) * 3) / 5)
1283 set2_end = vstruct_last(set1);
1285 set1->u64s = cpu_to_le16((u64 *) set2_start - set1->_data);
1286 set_btree_bset_end(n1, n1->set);
1288 n1->nr.live_u64s = le16_to_cpu(set1->u64s);
1289 n1->nr.bset_u64s[0] = le16_to_cpu(set1->u64s);
1290 n1->nr.packed_keys = nr_packed;
1291 n1->nr.unpacked_keys = nr_unpacked;
1293 n1_pos = bkey_unpack_pos(n1, prev);
1294 if (as->c->sb.version < bcachefs_metadata_version_snapshot)
1295 n1_pos.snapshot = U32_MAX;
1297 btree_set_max(n1, n1_pos);
1298 btree_set_min(n2, bpos_successor(n1->key.k.p));
1300 bch2_bkey_format_init(&s);
1301 bch2_bkey_format_add_pos(&s, n2->data->min_key);
1302 bch2_bkey_format_add_pos(&s, n2->data->max_key);
1304 for (k = set2_start; k != set2_end; k = bkey_next(k)) {
1305 struct bkey uk = bkey_unpack_key(n1, k);
1306 bch2_bkey_format_add_key(&s, &uk);
1309 n2->data->format = bch2_bkey_format_done(&s);
1310 btree_node_set_format(n2, n2->data->format);
1313 memset(&n2->nr, 0, sizeof(n2->nr));
1315 for (k = set2_start; k != set2_end; k = bkey_next(k)) {
1316 BUG_ON(!bch2_bkey_transform(&n2->format, out, bkey_packed(k)
1317 ? &n1->format : &bch2_bkey_format_current, k));
1318 out->format = KEY_FORMAT_LOCAL_BTREE;
1319 btree_keys_account_key_add(&n2->nr, 0, out);
1320 out = bkey_next(out);
1323 set2->u64s = cpu_to_le16((u64 *) out - set2->_data);
1324 set_btree_bset_end(n2, n2->set);
1326 BUG_ON(!set1->u64s);
1327 BUG_ON(!set2->u64s);
1329 btree_node_reset_sib_u64s(n1);
1330 btree_node_reset_sib_u64s(n2);
1332 bch2_verify_btree_nr_keys(n1);
1333 bch2_verify_btree_nr_keys(n2);
1336 btree_node_interior_verify(as->c, n1);
1337 btree_node_interior_verify(as->c, n2);
1344 * For updates to interior nodes, we've got to do the insert before we split
1345 * because the stuff we're inserting has to be inserted atomically. Post split,
1346 * the keys might have to go in different nodes and the split would no longer be
1349 * Worse, if the insert is from btree node coalescing, if we do the insert after
1350 * we do the split (and pick the pivot) - the pivot we pick might be between
1351 * nodes that were coalesced, and thus in the middle of a child node post
1354 static void btree_split_insert_keys(struct btree_update *as,
1355 struct btree_trans *trans,
1356 struct btree_path *path,
1358 struct keylist *keys)
1360 struct btree_node_iter node_iter;
1361 struct bkey_i *k = bch2_keylist_front(keys);
1362 struct bkey_packed *src, *dst, *n;
1365 bch2_btree_node_iter_init(&node_iter, b, &k->k.p);
1367 __bch2_btree_insert_keys_interior(as, trans, path, b, node_iter, keys);
1370 * We can't tolerate whiteouts here - with whiteouts there can be
1371 * duplicate keys, and it would be rather bad if we picked a duplicate
1374 i = btree_bset_first(b);
1375 src = dst = i->start;
1376 while (src != vstruct_last(i)) {
1378 if (!bkey_deleted(src)) {
1379 memmove_u64s_down(dst, src, src->u64s);
1380 dst = bkey_next(dst);
1385 /* Also clear out the unwritten whiteouts area: */
1386 b->whiteout_u64s = 0;
1388 i->u64s = cpu_to_le16((u64 *) dst - i->_data);
1389 set_btree_bset_end(b, b->set);
1391 BUG_ON(b->nsets != 1 ||
1392 b->nr.live_u64s != le16_to_cpu(btree_bset_first(b)->u64s));
1394 btree_node_interior_verify(as->c, b);
1397 static void btree_split(struct btree_update *as, struct btree_trans *trans,
1398 struct btree_path *path, struct btree *b,
1399 struct keylist *keys, unsigned flags)
1401 struct bch_fs *c = as->c;
1402 struct btree *parent = btree_node_parent(path, b);
1403 struct btree *n1, *n2 = NULL, *n3 = NULL;
1404 u64 start_time = local_clock();
1406 BUG_ON(!parent && (b != btree_node_root(c, b)));
1407 BUG_ON(!btree_node_intent_locked(path, btree_node_root(c, b)->c.level));
1409 bch2_btree_interior_update_will_free_node(as, b);
1411 n1 = bch2_btree_node_alloc_replacement(as, b);
1414 btree_split_insert_keys(as, trans, path, n1, keys);
1416 if (bset_u64s(&n1->set[0]) > BTREE_SPLIT_THRESHOLD(c)) {
1417 trace_btree_split(c, b);
1419 n2 = __btree_split_node(as, n1);
1421 bch2_btree_build_aux_trees(n2);
1422 bch2_btree_build_aux_trees(n1);
1423 six_unlock_write(&n2->c.lock);
1424 six_unlock_write(&n1->c.lock);
1426 bch2_btree_update_add_new_node(as, n1);
1428 bch2_btree_node_write(c, n1, SIX_LOCK_intent, 0);
1429 bch2_btree_node_write(c, n2, SIX_LOCK_intent, 0);
1432 * Note that on recursive parent_keys == keys, so we
1433 * can't start adding new keys to parent_keys before emptying it
1434 * out (which we did with btree_split_insert_keys() above)
1436 bch2_keylist_add(&as->parent_keys, &n1->key);
1437 bch2_keylist_add(&as->parent_keys, &n2->key);
1440 /* Depth increases, make a new root */
1441 n3 = __btree_root_alloc(as, b->c.level + 1);
1443 n3->sib_u64s[0] = U16_MAX;
1444 n3->sib_u64s[1] = U16_MAX;
1446 btree_split_insert_keys(as, trans, path, n3, &as->parent_keys);
1448 bch2_btree_node_write(c, n3, SIX_LOCK_intent, 0);
1451 trace_btree_compact(c, b);
1453 bch2_btree_build_aux_trees(n1);
1454 six_unlock_write(&n1->c.lock);
1456 bch2_btree_update_add_new_node(as, n1);
1458 bch2_btree_node_write(c, n1, SIX_LOCK_intent, 0);
1461 bch2_keylist_add(&as->parent_keys, &n1->key);
1464 /* New nodes all written, now make them visible: */
1467 /* Split a non root node */
1468 bch2_btree_insert_node(as, trans, path, parent, &as->parent_keys, flags);
1470 bch2_btree_set_root(as, trans, path, n3);
1472 /* Root filled up but didn't need to be split */
1473 bch2_btree_set_root(as, trans, path, n1);
1476 bch2_btree_update_get_open_buckets(as, n1);
1478 bch2_btree_update_get_open_buckets(as, n2);
1480 bch2_btree_update_get_open_buckets(as, n3);
1482 /* Successful split, update the path to point to the new nodes: */
1484 six_lock_increment(&b->c.lock, SIX_LOCK_intent);
1486 bch2_trans_node_add(trans, n3);
1488 bch2_trans_node_add(trans, n2);
1489 bch2_trans_node_add(trans, n1);
1492 * The old node must be freed (in memory) _before_ unlocking the new
1493 * nodes - else another thread could re-acquire a read lock on the old
1494 * node after another thread has locked and updated the new node, thus
1495 * seeing stale data:
1497 bch2_btree_node_free_inmem(trans, b);
1500 six_unlock_intent(&n3->c.lock);
1502 six_unlock_intent(&n2->c.lock);
1503 six_unlock_intent(&n1->c.lock);
1505 bch2_trans_verify_locks(trans);
1507 bch2_time_stats_update(&c->times[n2
1508 ? BCH_TIME_btree_node_split
1509 : BCH_TIME_btree_node_compact],
1514 bch2_btree_insert_keys_interior(struct btree_update *as,
1515 struct btree_trans *trans,
1516 struct btree_path *path,
1518 struct keylist *keys)
1520 struct btree_path *linked;
1522 __bch2_btree_insert_keys_interior(as, trans, path, b,
1523 path->l[b->c.level].iter, keys);
1525 btree_update_updated_node(as, b);
1527 trans_for_each_path_with_node(trans, b, linked)
1528 bch2_btree_node_iter_peek(&linked->l[b->c.level].iter, b);
1530 bch2_trans_verify_paths(trans);
1534 * bch_btree_insert_node - insert bkeys into a given btree node
1536 * @iter: btree iterator
1537 * @keys: list of keys to insert
1538 * @hook: insert callback
1539 * @persistent: if not null, @persistent will wait on journal write
1541 * Inserts as many keys as it can into a given btree node, splitting it if full.
1542 * If a split occurred, this function will return early. This can only happen
1543 * for leaf nodes -- inserts into interior nodes have to be atomic.
1545 static void bch2_btree_insert_node(struct btree_update *as, struct btree_trans *trans,
1546 struct btree_path *path, struct btree *b,
1547 struct keylist *keys, unsigned flags)
1549 struct bch_fs *c = as->c;
1550 int old_u64s = le16_to_cpu(btree_bset_last(b)->u64s);
1551 int old_live_u64s = b->nr.live_u64s;
1552 int live_u64s_added, u64s_added;
1554 lockdep_assert_held(&c->gc_lock);
1555 BUG_ON(!btree_node_intent_locked(path, btree_node_root(c, b)->c.level));
1556 BUG_ON(!b->c.level);
1557 BUG_ON(!as || as->b);
1558 bch2_verify_keylist_sorted(keys);
1560 bch2_btree_node_lock_for_insert(trans, path, b);
1562 if (!bch2_btree_node_insert_fits(c, b, bch2_keylist_u64s(keys))) {
1563 bch2_btree_node_unlock_write(trans, path, b);
1567 btree_node_interior_verify(c, b);
1569 bch2_btree_insert_keys_interior(as, trans, path, b, keys);
1571 live_u64s_added = (int) b->nr.live_u64s - old_live_u64s;
1572 u64s_added = (int) le16_to_cpu(btree_bset_last(b)->u64s) - old_u64s;
1574 if (b->sib_u64s[0] != U16_MAX && live_u64s_added < 0)
1575 b->sib_u64s[0] = max(0, (int) b->sib_u64s[0] + live_u64s_added);
1576 if (b->sib_u64s[1] != U16_MAX && live_u64s_added < 0)
1577 b->sib_u64s[1] = max(0, (int) b->sib_u64s[1] + live_u64s_added);
1579 if (u64s_added > live_u64s_added &&
1580 bch2_maybe_compact_whiteouts(c, b))
1581 bch2_trans_node_reinit_iter(trans, b);
1583 bch2_btree_node_unlock_write(trans, path, b);
1585 btree_node_interior_verify(c, b);
1588 btree_split(as, trans, path, b, keys, flags);
1591 int bch2_btree_split_leaf(struct btree_trans *trans,
1592 struct btree_path *path,
1595 struct btree *b = path_l(path)->b;
1596 struct btree_update *as;
1600 as = bch2_btree_update_start(trans, path, path->level,
1605 btree_split(as, trans, path, b, NULL, flags);
1606 bch2_btree_update_done(as);
1608 for (l = path->level + 1; btree_path_node(path, l) && !ret; l++)
1609 ret = bch2_foreground_maybe_merge(trans, path, l, flags);
1614 int __bch2_foreground_maybe_merge(struct btree_trans *trans,
1615 struct btree_path *path,
1618 enum btree_node_sibling sib)
1620 struct bch_fs *c = trans->c;
1621 struct btree_path *sib_path = NULL;
1622 struct btree_update *as;
1623 struct bkey_format_state new_s;
1624 struct bkey_format new_f;
1625 struct bkey_i delete;
1626 struct btree *b, *m, *n, *prev, *next, *parent;
1627 struct bpos sib_pos;
1629 u64 start_time = local_clock();
1632 BUG_ON(!path->should_be_locked);
1633 BUG_ON(!btree_node_locked(path, level));
1635 b = path->l[level].b;
1637 if ((sib == btree_prev_sib && !bpos_cmp(b->data->min_key, POS_MIN)) ||
1638 (sib == btree_next_sib && !bpos_cmp(b->data->max_key, SPOS_MAX))) {
1639 b->sib_u64s[sib] = U16_MAX;
1643 sib_pos = sib == btree_prev_sib
1644 ? bpos_predecessor(b->data->min_key)
1645 : bpos_successor(b->data->max_key);
1647 sib_path = bch2_path_get(trans, path->btree_id, sib_pos,
1648 U8_MAX, level, BTREE_ITER_INTENT, _THIS_IP_);
1649 ret = bch2_btree_path_traverse(trans, sib_path, false);
1653 sib_path->should_be_locked = true;
1655 m = sib_path->l[level].b;
1657 if (btree_node_parent(path, b) !=
1658 btree_node_parent(sib_path, m)) {
1659 b->sib_u64s[sib] = U16_MAX;
1663 if (sib == btree_prev_sib) {
1671 if (bkey_cmp(bpos_successor(prev->data->max_key), next->data->min_key)) {
1672 struct printbuf buf1 = PRINTBUF, buf2 = PRINTBUF;
1674 bch2_bpos_to_text(&buf1, prev->data->max_key);
1675 bch2_bpos_to_text(&buf2, next->data->min_key);
1677 "btree topology error in btree merge:\n"
1678 " prev ends at %s\n"
1679 " next starts at %s",
1680 buf1.buf, buf2.buf);
1681 printbuf_exit(&buf1);
1682 printbuf_exit(&buf2);
1683 bch2_topology_error(c);
1688 bch2_bkey_format_init(&new_s);
1689 bch2_bkey_format_add_pos(&new_s, prev->data->min_key);
1690 __bch2_btree_calc_format(&new_s, prev);
1691 __bch2_btree_calc_format(&new_s, next);
1692 bch2_bkey_format_add_pos(&new_s, next->data->max_key);
1693 new_f = bch2_bkey_format_done(&new_s);
1695 sib_u64s = btree_node_u64s_with_format(b, &new_f) +
1696 btree_node_u64s_with_format(m, &new_f);
1698 if (sib_u64s > BTREE_FOREGROUND_MERGE_HYSTERESIS(c)) {
1699 sib_u64s -= BTREE_FOREGROUND_MERGE_HYSTERESIS(c);
1701 sib_u64s += BTREE_FOREGROUND_MERGE_HYSTERESIS(c);
1704 sib_u64s = min(sib_u64s, btree_max_u64s(c));
1705 sib_u64s = min(sib_u64s, (size_t) U16_MAX - 1);
1706 b->sib_u64s[sib] = sib_u64s;
1708 if (b->sib_u64s[sib] > c->btree_foreground_merge_threshold)
1711 parent = btree_node_parent(path, b);
1712 as = bch2_btree_update_start(trans, path, level, false,
1713 BTREE_INSERT_NOFAIL|
1714 BTREE_INSERT_USE_RESERVE|
1716 ret = PTR_ERR_OR_ZERO(as);
1720 trace_btree_merge(c, b);
1722 bch2_btree_interior_update_will_free_node(as, b);
1723 bch2_btree_interior_update_will_free_node(as, m);
1725 n = bch2_btree_node_alloc(as, b->c.level);
1727 SET_BTREE_NODE_SEQ(n->data,
1728 max(BTREE_NODE_SEQ(b->data),
1729 BTREE_NODE_SEQ(m->data)) + 1);
1731 btree_set_min(n, prev->data->min_key);
1732 btree_set_max(n, next->data->max_key);
1734 bch2_btree_update_add_new_node(as, n);
1736 n->data->format = new_f;
1737 btree_node_set_format(n, new_f);
1739 bch2_btree_sort_into(c, n, prev);
1740 bch2_btree_sort_into(c, n, next);
1742 bch2_btree_build_aux_trees(n);
1743 six_unlock_write(&n->c.lock);
1745 bch2_btree_node_write(c, n, SIX_LOCK_intent, 0);
1747 bkey_init(&delete.k);
1748 delete.k.p = prev->key.k.p;
1749 bch2_keylist_add(&as->parent_keys, &delete);
1750 bch2_keylist_add(&as->parent_keys, &n->key);
1752 bch2_trans_verify_paths(trans);
1754 bch2_btree_insert_node(as, trans, path, parent, &as->parent_keys, flags);
1756 bch2_trans_verify_paths(trans);
1758 bch2_btree_update_get_open_buckets(as, n);
1760 six_lock_increment(&b->c.lock, SIX_LOCK_intent);
1761 six_lock_increment(&m->c.lock, SIX_LOCK_intent);
1763 bch2_trans_node_add(trans, n);
1765 bch2_trans_verify_paths(trans);
1767 bch2_btree_node_free_inmem(trans, b);
1768 bch2_btree_node_free_inmem(trans, m);
1770 six_unlock_intent(&n->c.lock);
1772 bch2_btree_update_done(as);
1774 bch2_time_stats_update(&c->times[BCH_TIME_btree_node_merge], start_time);
1777 bch2_path_put(trans, sib_path, true);
1778 bch2_trans_verify_locks(trans);
1783 * bch_btree_node_rewrite - Rewrite/move a btree node
1785 int bch2_btree_node_rewrite(struct btree_trans *trans,
1786 struct btree_iter *iter,
1790 struct bch_fs *c = trans->c;
1791 struct btree *n, *parent;
1792 struct btree_update *as;
1795 flags |= BTREE_INSERT_NOFAIL;
1797 parent = btree_node_parent(iter->path, b);
1798 as = bch2_btree_update_start(trans, iter->path, b->c.level,
1800 ret = PTR_ERR_OR_ZERO(as);
1804 bch2_btree_interior_update_will_free_node(as, b);
1806 n = bch2_btree_node_alloc_replacement(as, b);
1807 bch2_btree_update_add_new_node(as, n);
1809 bch2_btree_build_aux_trees(n);
1810 six_unlock_write(&n->c.lock);
1812 trace_btree_rewrite(c, b);
1814 bch2_btree_node_write(c, n, SIX_LOCK_intent, 0);
1817 bch2_keylist_add(&as->parent_keys, &n->key);
1818 bch2_btree_insert_node(as, trans, iter->path, parent,
1819 &as->parent_keys, flags);
1821 bch2_btree_set_root(as, trans, iter->path, n);
1824 bch2_btree_update_get_open_buckets(as, n);
1826 six_lock_increment(&b->c.lock, SIX_LOCK_intent);
1827 bch2_trans_node_add(trans, n);
1828 bch2_btree_node_free_inmem(trans, b);
1829 six_unlock_intent(&n->c.lock);
1831 bch2_btree_update_done(as);
1833 bch2_btree_path_downgrade(iter->path);
1837 struct async_btree_rewrite {
1839 struct work_struct work;
1840 enum btree_id btree_id;
1846 static int async_btree_node_rewrite_trans(struct btree_trans *trans,
1847 struct async_btree_rewrite *a)
1849 struct btree_iter iter;
1853 bch2_trans_node_iter_init(trans, &iter, a->btree_id, a->pos,
1854 BTREE_MAX_DEPTH, a->level, 0);
1855 b = bch2_btree_iter_peek_node(&iter);
1856 ret = PTR_ERR_OR_ZERO(b);
1860 if (!b || b->data->keys.seq != a->seq)
1863 ret = bch2_btree_node_rewrite(trans, &iter, b, 0);
1865 bch2_trans_iter_exit(trans, &iter);
1870 void async_btree_node_rewrite_work(struct work_struct *work)
1872 struct async_btree_rewrite *a =
1873 container_of(work, struct async_btree_rewrite, work);
1874 struct bch_fs *c = a->c;
1876 bch2_trans_do(c, NULL, NULL, 0,
1877 async_btree_node_rewrite_trans(&trans, a));
1878 percpu_ref_put(&c->writes);
1882 void bch2_btree_node_rewrite_async(struct bch_fs *c, struct btree *b)
1884 struct async_btree_rewrite *a;
1886 if (!percpu_ref_tryget(&c->writes))
1889 a = kmalloc(sizeof(*a), GFP_NOFS);
1891 percpu_ref_put(&c->writes);
1896 a->btree_id = b->c.btree_id;
1897 a->level = b->c.level;
1898 a->pos = b->key.k.p;
1899 a->seq = b->data->keys.seq;
1901 INIT_WORK(&a->work, async_btree_node_rewrite_work);
1902 queue_work(c->btree_interior_update_worker, &a->work);
1905 static int __bch2_btree_node_update_key(struct btree_trans *trans,
1906 struct btree_iter *iter,
1907 struct btree *b, struct btree *new_hash,
1908 struct bkey_i *new_key,
1911 struct bch_fs *c = trans->c;
1912 struct btree_iter iter2 = { NULL };
1913 struct btree *parent;
1916 if (!skip_triggers) {
1917 ret = bch2_trans_mark_old(trans, b->c.btree_id, b->c.level + 1,
1918 bkey_i_to_s_c(&b->key), 0);
1922 ret = bch2_trans_mark_new(trans, b->c.btree_id, b->c.level + 1,
1929 bkey_copy(&new_hash->key, new_key);
1930 ret = bch2_btree_node_hash_insert(&c->btree_cache,
1931 new_hash, b->c.level, b->c.btree_id);
1935 parent = btree_node_parent(iter->path, b);
1937 bch2_trans_copy_iter(&iter2, iter);
1939 iter2.path = bch2_btree_path_make_mut(trans, iter2.path,
1940 iter2.flags & BTREE_ITER_INTENT,
1943 BUG_ON(iter2.path->level != b->c.level);
1944 BUG_ON(bpos_cmp(iter2.path->pos, new_key->k.p));
1946 btree_node_unlock(iter2.path, iter2.path->level);
1947 path_l(iter2.path)->b = BTREE_ITER_NO_NODE_UP;
1948 iter2.path->level++;
1949 btree_path_set_dirty(iter2.path, BTREE_ITER_NEED_TRAVERSE);
1951 bch2_btree_path_check_sort(trans, iter2.path, 0);
1953 ret = bch2_btree_iter_traverse(&iter2) ?:
1954 bch2_trans_update(trans, &iter2, new_key, BTREE_TRIGGER_NORUN);
1958 BUG_ON(btree_node_root(c, b) != b);
1960 ret = darray_make_room(&trans->extra_journal_entries,
1961 jset_u64s(new_key->k.u64s));
1965 journal_entry_set((void *) &darray_top(trans->extra_journal_entries),
1966 BCH_JSET_ENTRY_btree_root,
1967 b->c.btree_id, b->c.level,
1968 new_key, new_key->k.u64s);
1969 trans->extra_journal_entries.nr += jset_u64s(new_key->k.u64s);
1972 ret = bch2_trans_commit(trans, NULL, NULL,
1973 BTREE_INSERT_NOFAIL|
1974 BTREE_INSERT_NOCHECK_RW|
1975 BTREE_INSERT_USE_RESERVE|
1976 BTREE_INSERT_JOURNAL_RECLAIM|
1977 JOURNAL_WATERMARK_reserved);
1981 bch2_btree_node_lock_write(trans, iter->path, b);
1984 mutex_lock(&c->btree_cache.lock);
1985 bch2_btree_node_hash_remove(&c->btree_cache, new_hash);
1986 bch2_btree_node_hash_remove(&c->btree_cache, b);
1988 bkey_copy(&b->key, new_key);
1989 ret = __bch2_btree_node_hash_insert(&c->btree_cache, b);
1991 mutex_unlock(&c->btree_cache.lock);
1993 bkey_copy(&b->key, new_key);
1996 bch2_btree_node_unlock_write(trans, iter->path, b);
1998 bch2_trans_iter_exit(trans, &iter2);
2002 mutex_lock(&c->btree_cache.lock);
2003 bch2_btree_node_hash_remove(&c->btree_cache, b);
2004 mutex_unlock(&c->btree_cache.lock);
2009 int bch2_btree_node_update_key(struct btree_trans *trans, struct btree_iter *iter,
2010 struct btree *b, struct bkey_i *new_key,
2013 struct bch_fs *c = trans->c;
2014 struct btree *new_hash = NULL;
2015 struct btree_path *path = iter->path;
2019 if (!btree_node_intent_locked(path, b->c.level) &&
2020 !bch2_btree_path_upgrade(trans, path, b->c.level + 1)) {
2021 btree_trans_restart(trans);
2025 closure_init_stack(&cl);
2028 * check btree_ptr_hash_val() after @b is locked by
2029 * btree_iter_traverse():
2031 if (btree_ptr_hash_val(new_key) != b->hash_val) {
2032 ret = bch2_btree_cache_cannibalize_lock(c, &cl);
2034 bch2_trans_unlock(trans);
2036 if (!bch2_trans_relock(trans))
2040 new_hash = bch2_btree_node_mem_alloc(c, false);
2044 ret = __bch2_btree_node_update_key(trans, iter, b, new_hash,
2045 new_key, skip_triggers);
2049 mutex_lock(&c->btree_cache.lock);
2050 list_move(&new_hash->list, &c->btree_cache.freeable);
2051 mutex_unlock(&c->btree_cache.lock);
2053 six_unlock_write(&new_hash->c.lock);
2054 six_unlock_intent(&new_hash->c.lock);
2057 bch2_btree_cache_cannibalize_unlock(c);
2061 int bch2_btree_node_update_key_get_iter(struct btree_trans *trans,
2062 struct btree *b, struct bkey_i *new_key,
2065 struct btree_iter iter;
2068 bch2_trans_node_iter_init(trans, &iter, b->c.btree_id, b->key.k.p,
2069 BTREE_MAX_DEPTH, b->c.level,
2071 ret = bch2_btree_iter_traverse(&iter);
2075 /* has node been freed? */
2076 if (iter.path->l[b->c.level].b != b) {
2077 /* node has been freed: */
2078 BUG_ON(!btree_node_dying(b));
2082 BUG_ON(!btree_node_hashed(b));
2084 ret = bch2_btree_node_update_key(trans, &iter, b, new_key, skip_triggers);
2086 bch2_trans_iter_exit(trans, &iter);
2093 * Only for filesystem bringup, when first reading the btree roots or allocating
2094 * btree roots when initializing a new filesystem:
2096 void bch2_btree_set_root_for_read(struct bch_fs *c, struct btree *b)
2098 BUG_ON(btree_node_root(c, b));
2100 bch2_btree_set_root_inmem(c, b);
2103 void bch2_btree_root_alloc(struct bch_fs *c, enum btree_id id)
2109 closure_init_stack(&cl);
2112 ret = bch2_btree_cache_cannibalize_lock(c, &cl);
2116 b = bch2_btree_node_mem_alloc(c, false);
2117 bch2_btree_cache_cannibalize_unlock(c);
2119 set_btree_node_fake(b);
2120 set_btree_node_need_rewrite(b);
2124 bkey_btree_ptr_init(&b->key);
2125 b->key.k.p = SPOS_MAX;
2126 *((u64 *) bkey_i_to_btree_ptr(&b->key)->v.start) = U64_MAX - id;
2128 bch2_bset_init_first(b, &b->data->keys);
2129 bch2_btree_build_aux_trees(b);
2132 btree_set_min(b, POS_MIN);
2133 btree_set_max(b, SPOS_MAX);
2134 b->data->format = bch2_btree_calc_format(b);
2135 btree_node_set_format(b, b->data->format);
2137 ret = bch2_btree_node_hash_insert(&c->btree_cache, b,
2138 b->c.level, b->c.btree_id);
2141 bch2_btree_set_root_inmem(c, b);
2143 six_unlock_write(&b->c.lock);
2144 six_unlock_intent(&b->c.lock);
2147 void bch2_btree_updates_to_text(struct printbuf *out, struct bch_fs *c)
2149 struct btree_update *as;
2151 mutex_lock(&c->btree_interior_update_lock);
2152 list_for_each_entry(as, &c->btree_interior_update_list, list)
2153 prt_printf(out, "%p m %u w %u r %u j %llu\n",
2157 atomic_read(&as->cl.remaining) & CLOSURE_REMAINING_MASK,
2159 mutex_unlock(&c->btree_interior_update_lock);
2162 static bool bch2_btree_interior_updates_pending(struct bch_fs *c)
2166 mutex_lock(&c->btree_interior_update_lock);
2167 ret = !list_empty(&c->btree_interior_update_list);
2168 mutex_unlock(&c->btree_interior_update_lock);
2173 bool bch2_btree_interior_updates_flush(struct bch_fs *c)
2175 bool ret = bch2_btree_interior_updates_pending(c);
2178 closure_wait_event(&c->btree_interior_update_wait,
2179 !bch2_btree_interior_updates_pending(c));
2183 void bch2_journal_entries_to_btree_roots(struct bch_fs *c, struct jset *jset)
2185 struct btree_root *r;
2186 struct jset_entry *entry;
2188 mutex_lock(&c->btree_root_lock);
2190 vstruct_for_each(jset, entry)
2191 if (entry->type == BCH_JSET_ENTRY_btree_root) {
2192 r = &c->btree_roots[entry->btree_id];
2193 r->level = entry->level;
2195 bkey_copy(&r->key, &entry->start[0]);
2198 mutex_unlock(&c->btree_root_lock);
2202 bch2_btree_roots_to_journal_entries(struct bch_fs *c,
2203 struct jset_entry *start,
2204 struct jset_entry *end)
2206 struct jset_entry *entry;
2207 unsigned long have = 0;
2210 for (entry = start; entry < end; entry = vstruct_next(entry))
2211 if (entry->type == BCH_JSET_ENTRY_btree_root)
2212 __set_bit(entry->btree_id, &have);
2214 mutex_lock(&c->btree_root_lock);
2216 for (i = 0; i < BTREE_ID_NR; i++)
2217 if (c->btree_roots[i].alive && !test_bit(i, &have)) {
2218 journal_entry_set(end,
2219 BCH_JSET_ENTRY_btree_root,
2220 i, c->btree_roots[i].level,
2221 &c->btree_roots[i].key,
2222 c->btree_roots[i].key.u64s);
2223 end = vstruct_next(end);
2226 mutex_unlock(&c->btree_root_lock);
2231 void bch2_fs_btree_interior_update_exit(struct bch_fs *c)
2233 if (c->btree_interior_update_worker)
2234 destroy_workqueue(c->btree_interior_update_worker);
2235 mempool_exit(&c->btree_interior_update_pool);
2238 int bch2_fs_btree_interior_update_init(struct bch_fs *c)
2240 mutex_init(&c->btree_reserve_cache_lock);
2241 INIT_LIST_HEAD(&c->btree_interior_update_list);
2242 INIT_LIST_HEAD(&c->btree_interior_updates_unwritten);
2243 mutex_init(&c->btree_interior_update_lock);
2244 INIT_WORK(&c->btree_interior_update_work, btree_interior_update_work);
2246 c->btree_interior_update_worker =
2247 alloc_workqueue("btree_update", WQ_UNBOUND|WQ_MEM_RECLAIM, 1);
2248 if (!c->btree_interior_update_worker)
2251 return mempool_init_kmalloc_pool(&c->btree_interior_update_pool, 1,
2252 sizeof(struct btree_update));