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_journal_iter.h"
9 #include "btree_update.h"
10 #include "btree_update_interior.h"
12 #include "btree_iter.h"
13 #include "btree_locking.h"
19 #include "journal_reclaim.h"
25 #include <linux/random.h>
27 static int bch2_btree_insert_node(struct btree_update *, struct btree_trans *,
28 btree_path_idx_t, struct btree *,
29 struct keylist *, unsigned);
30 static void bch2_btree_update_add_new_node(struct btree_update *, struct btree *);
32 static btree_path_idx_t get_unlocked_mut_path(struct btree_trans *trans,
33 enum btree_id btree_id,
37 btree_path_idx_t path_idx = bch2_path_get(trans, btree_id, pos, level + 1, level,
38 BTREE_ITER_NOPRESERVE|
39 BTREE_ITER_INTENT, _RET_IP_);
40 path_idx = bch2_btree_path_make_mut(trans, path_idx, true, _RET_IP_);
42 struct btree_path *path = trans->paths + path_idx;
43 bch2_btree_path_downgrade(trans, path);
44 __bch2_btree_path_unlock(trans, path);
51 * Verify that child nodes correctly span parent node's range:
53 static void btree_node_interior_verify(struct bch_fs *c, struct btree *b)
55 #ifdef CONFIG_BCACHEFS_DEBUG
56 struct bpos next_node = b->data->min_key;
57 struct btree_node_iter iter;
59 struct bkey_s_c_btree_ptr_v2 bp;
61 struct printbuf buf1 = PRINTBUF, buf2 = PRINTBUF;
65 if (!test_bit(JOURNAL_REPLAY_DONE, &c->journal.flags))
68 bch2_btree_node_iter_init_from_start(&iter, b);
71 k = bch2_btree_node_iter_peek_unpack(&iter, b, &unpacked);
72 if (k.k->type != KEY_TYPE_btree_ptr_v2)
74 bp = bkey_s_c_to_btree_ptr_v2(k);
76 if (!bpos_eq(next_node, bp.v->min_key)) {
77 bch2_dump_btree_node(c, b);
78 bch2_bpos_to_text(&buf1, next_node);
79 bch2_bpos_to_text(&buf2, bp.v->min_key);
80 panic("expected next min_key %s got %s\n", buf1.buf, buf2.buf);
83 bch2_btree_node_iter_advance(&iter, b);
85 if (bch2_btree_node_iter_end(&iter)) {
86 if (!bpos_eq(k.k->p, b->key.k.p)) {
87 bch2_dump_btree_node(c, b);
88 bch2_bpos_to_text(&buf1, b->key.k.p);
89 bch2_bpos_to_text(&buf2, k.k->p);
90 panic("expected end %s got %s\n", buf1.buf, buf2.buf);
95 next_node = bpos_successor(k.k->p);
100 /* Calculate ideal packed bkey format for new btree nodes: */
102 static void __bch2_btree_calc_format(struct bkey_format_state *s, struct btree *b)
104 struct bkey_packed *k;
109 bset_tree_for_each_key(b, t, k)
110 if (!bkey_deleted(k)) {
111 uk = bkey_unpack_key(b, k);
112 bch2_bkey_format_add_key(s, &uk);
116 static struct bkey_format bch2_btree_calc_format(struct btree *b)
118 struct bkey_format_state s;
120 bch2_bkey_format_init(&s);
121 bch2_bkey_format_add_pos(&s, b->data->min_key);
122 bch2_bkey_format_add_pos(&s, b->data->max_key);
123 __bch2_btree_calc_format(&s, b);
125 return bch2_bkey_format_done(&s);
128 static size_t btree_node_u64s_with_format(struct btree_nr_keys nr,
129 struct bkey_format *old_f,
130 struct bkey_format *new_f)
132 /* stupid integer promotion rules */
134 (((int) new_f->key_u64s - old_f->key_u64s) *
135 (int) nr.packed_keys) +
136 (((int) new_f->key_u64s - BKEY_U64s) *
137 (int) nr.unpacked_keys);
139 BUG_ON(delta + nr.live_u64s < 0);
141 return nr.live_u64s + delta;
145 * bch2_btree_node_format_fits - check if we could rewrite node with a new format
147 * @c: filesystem handle
148 * @b: btree node to rewrite
149 * @nr: number of keys for new node (i.e. b->nr)
150 * @new_f: bkey format to translate keys to
152 * Returns: true if all re-packed keys will be able to fit in a new node.
154 * Assumes all keys will successfully pack with the new format.
156 static bool bch2_btree_node_format_fits(struct bch_fs *c, struct btree *b,
157 struct btree_nr_keys nr,
158 struct bkey_format *new_f)
160 size_t u64s = btree_node_u64s_with_format(nr, &b->format, new_f);
162 return __vstruct_bytes(struct btree_node, u64s) < btree_buf_bytes(b);
165 /* Btree node freeing/allocation: */
167 static void __btree_node_free(struct btree_trans *trans, struct btree *b)
169 struct bch_fs *c = trans->c;
171 trace_and_count(c, btree_node_free, trans, b);
173 BUG_ON(btree_node_write_blocked(b));
174 BUG_ON(btree_node_dirty(b));
175 BUG_ON(btree_node_need_write(b));
176 BUG_ON(b == btree_node_root(c, b));
178 BUG_ON(!list_empty(&b->write_blocked));
179 BUG_ON(b->will_make_reachable);
181 clear_btree_node_noevict(b);
183 mutex_lock(&c->btree_cache.lock);
184 list_move(&b->list, &c->btree_cache.freeable);
185 mutex_unlock(&c->btree_cache.lock);
188 static void bch2_btree_node_free_inmem(struct btree_trans *trans,
189 struct btree_path *path,
192 struct bch_fs *c = trans->c;
193 unsigned i, level = b->c.level;
195 bch2_btree_node_lock_write_nofail(trans, path, &b->c);
196 bch2_btree_node_hash_remove(&c->btree_cache, b);
197 __btree_node_free(trans, b);
198 six_unlock_write(&b->c.lock);
199 mark_btree_node_locked_noreset(path, level, BTREE_NODE_INTENT_LOCKED);
201 trans_for_each_path(trans, path, i)
202 if (path->l[level].b == b) {
203 btree_node_unlock(trans, path, level);
204 path->l[level].b = ERR_PTR(-BCH_ERR_no_btree_node_init);
208 static void bch2_btree_node_free_never_used(struct btree_update *as,
209 struct btree_trans *trans,
212 struct bch_fs *c = as->c;
213 struct prealloc_nodes *p = &as->prealloc_nodes[b->c.lock.readers != NULL];
214 struct btree_path *path;
215 unsigned i, level = b->c.level;
217 BUG_ON(!list_empty(&b->write_blocked));
218 BUG_ON(b->will_make_reachable != (1UL|(unsigned long) as));
220 b->will_make_reachable = 0;
221 closure_put(&as->cl);
223 clear_btree_node_will_make_reachable(b);
224 clear_btree_node_accessed(b);
225 clear_btree_node_dirty_acct(c, b);
226 clear_btree_node_need_write(b);
228 mutex_lock(&c->btree_cache.lock);
229 list_del_init(&b->list);
230 bch2_btree_node_hash_remove(&c->btree_cache, b);
231 mutex_unlock(&c->btree_cache.lock);
233 BUG_ON(p->nr >= ARRAY_SIZE(p->b));
236 six_unlock_intent(&b->c.lock);
238 trans_for_each_path(trans, path, i)
239 if (path->l[level].b == b) {
240 btree_node_unlock(trans, path, level);
241 path->l[level].b = ERR_PTR(-BCH_ERR_no_btree_node_init);
245 static struct btree *__bch2_btree_node_alloc(struct btree_trans *trans,
246 struct disk_reservation *res,
251 struct bch_fs *c = trans->c;
252 struct write_point *wp;
254 BKEY_PADDED_ONSTACK(k, BKEY_BTREE_PTR_VAL_U64s_MAX) tmp;
255 struct open_buckets obs = { .nr = 0 };
256 struct bch_devs_list devs_have = (struct bch_devs_list) { 0 };
257 enum bch_watermark watermark = flags & BCH_WATERMARK_MASK;
258 unsigned nr_reserve = watermark > BCH_WATERMARK_reclaim
263 mutex_lock(&c->btree_reserve_cache_lock);
264 if (c->btree_reserve_cache_nr > nr_reserve) {
265 struct btree_alloc *a =
266 &c->btree_reserve_cache[--c->btree_reserve_cache_nr];
269 bkey_copy(&tmp.k, &a->k);
270 mutex_unlock(&c->btree_reserve_cache_lock);
273 mutex_unlock(&c->btree_reserve_cache_lock);
276 ret = bch2_alloc_sectors_start_trans(trans,
277 c->opts.metadata_target ?:
278 c->opts.foreground_target,
280 writepoint_ptr(&c->btree_write_point),
283 c->opts.metadata_replicas_required,
284 watermark, 0, cl, &wp);
288 if (wp->sectors_free < btree_sectors(c)) {
289 struct open_bucket *ob;
292 open_bucket_for_each(c, &wp->ptrs, ob, i)
293 if (ob->sectors_free < btree_sectors(c))
294 ob->sectors_free = 0;
296 bch2_alloc_sectors_done(c, wp);
300 bkey_btree_ptr_v2_init(&tmp.k);
301 bch2_alloc_sectors_append_ptrs(c, wp, &tmp.k, btree_sectors(c), false);
303 bch2_open_bucket_get(c, wp, &obs);
304 bch2_alloc_sectors_done(c, wp);
306 b = bch2_btree_node_mem_alloc(trans, interior_node);
307 six_unlock_write(&b->c.lock);
308 six_unlock_intent(&b->c.lock);
310 /* we hold cannibalize_lock: */
314 bkey_copy(&b->key, &tmp.k);
320 static struct btree *bch2_btree_node_alloc(struct btree_update *as,
321 struct btree_trans *trans,
324 struct bch_fs *c = as->c;
326 struct prealloc_nodes *p = &as->prealloc_nodes[!!level];
329 BUG_ON(level >= BTREE_MAX_DEPTH);
334 btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_intent);
335 btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_write);
337 set_btree_node_accessed(b);
338 set_btree_node_dirty_acct(c, b);
339 set_btree_node_need_write(b);
341 bch2_bset_init_first(b, &b->data->keys);
343 b->c.btree_id = as->btree_id;
344 b->version_ondisk = c->sb.version;
346 memset(&b->nr, 0, sizeof(b->nr));
347 b->data->magic = cpu_to_le64(bset_magic(c));
348 memset(&b->data->_ptr, 0, sizeof(b->data->_ptr));
350 SET_BTREE_NODE_ID(b->data, as->btree_id);
351 SET_BTREE_NODE_LEVEL(b->data, level);
353 if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
354 struct bkey_i_btree_ptr_v2 *bp = bkey_i_to_btree_ptr_v2(&b->key);
357 bp->v.seq = b->data->keys.seq;
358 bp->v.sectors_written = 0;
361 SET_BTREE_NODE_NEW_EXTENT_OVERWRITE(b->data, true);
363 bch2_btree_build_aux_trees(b);
365 ret = bch2_btree_node_hash_insert(&c->btree_cache, b, level, as->btree_id);
368 trace_and_count(c, btree_node_alloc, trans, b);
369 bch2_increment_clock(c, btree_sectors(c), WRITE);
373 static void btree_set_min(struct btree *b, struct bpos pos)
375 if (b->key.k.type == KEY_TYPE_btree_ptr_v2)
376 bkey_i_to_btree_ptr_v2(&b->key)->v.min_key = pos;
377 b->data->min_key = pos;
380 static void btree_set_max(struct btree *b, struct bpos pos)
383 b->data->max_key = pos;
386 static struct btree *bch2_btree_node_alloc_replacement(struct btree_update *as,
387 struct btree_trans *trans,
390 struct btree *n = bch2_btree_node_alloc(as, trans, b->c.level);
391 struct bkey_format format = bch2_btree_calc_format(b);
394 * The keys might expand with the new format - if they wouldn't fit in
395 * the btree node anymore, use the old format for now:
397 if (!bch2_btree_node_format_fits(as->c, b, b->nr, &format))
400 SET_BTREE_NODE_SEQ(n->data, BTREE_NODE_SEQ(b->data) + 1);
402 btree_set_min(n, b->data->min_key);
403 btree_set_max(n, b->data->max_key);
405 n->data->format = format;
406 btree_node_set_format(n, format);
408 bch2_btree_sort_into(as->c, n, b);
410 btree_node_reset_sib_u64s(n);
414 static struct btree *__btree_root_alloc(struct btree_update *as,
415 struct btree_trans *trans, unsigned level)
417 struct btree *b = bch2_btree_node_alloc(as, trans, level);
419 btree_set_min(b, POS_MIN);
420 btree_set_max(b, SPOS_MAX);
421 b->data->format = bch2_btree_calc_format(b);
423 btree_node_set_format(b, b->data->format);
424 bch2_btree_build_aux_trees(b);
429 static void bch2_btree_reserve_put(struct btree_update *as, struct btree_trans *trans)
431 struct bch_fs *c = as->c;
432 struct prealloc_nodes *p;
434 for (p = as->prealloc_nodes;
435 p < as->prealloc_nodes + ARRAY_SIZE(as->prealloc_nodes);
438 struct btree *b = p->b[--p->nr];
440 mutex_lock(&c->btree_reserve_cache_lock);
442 if (c->btree_reserve_cache_nr <
443 ARRAY_SIZE(c->btree_reserve_cache)) {
444 struct btree_alloc *a =
445 &c->btree_reserve_cache[c->btree_reserve_cache_nr++];
449 bkey_copy(&a->k, &b->key);
451 bch2_open_buckets_put(c, &b->ob);
454 mutex_unlock(&c->btree_reserve_cache_lock);
456 btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_intent);
457 btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_write);
458 __btree_node_free(trans, b);
459 six_unlock_write(&b->c.lock);
460 six_unlock_intent(&b->c.lock);
465 static int bch2_btree_reserve_get(struct btree_trans *trans,
466 struct btree_update *as,
467 unsigned nr_nodes[2],
475 BUG_ON(nr_nodes[0] + nr_nodes[1] > BTREE_RESERVE_MAX);
478 * Protects reaping from the btree node cache and using the btree node
479 * open bucket reserve:
481 ret = bch2_btree_cache_cannibalize_lock(trans, cl);
485 for (interior = 0; interior < 2; interior++) {
486 struct prealloc_nodes *p = as->prealloc_nodes + interior;
488 while (p->nr < nr_nodes[interior]) {
489 b = __bch2_btree_node_alloc(trans, &as->disk_res, cl,
500 bch2_btree_cache_cannibalize_unlock(trans);
504 /* Asynchronous interior node update machinery */
506 static void bch2_btree_update_free(struct btree_update *as, struct btree_trans *trans)
508 struct bch_fs *c = as->c;
510 if (as->took_gc_lock)
511 up_read(&c->gc_lock);
512 as->took_gc_lock = false;
514 bch2_journal_pin_drop(&c->journal, &as->journal);
515 bch2_journal_pin_flush(&c->journal, &as->journal);
516 bch2_disk_reservation_put(c, &as->disk_res);
517 bch2_btree_reserve_put(as, trans);
519 bch2_time_stats_update(&c->times[BCH_TIME_btree_interior_update_total],
522 mutex_lock(&c->btree_interior_update_lock);
523 list_del(&as->unwritten_list);
526 closure_debug_destroy(&as->cl);
527 mempool_free(as, &c->btree_interior_update_pool);
530 * Have to do the wakeup with btree_interior_update_lock still held,
531 * since being on btree_interior_update_list is our ref on @c:
533 closure_wake_up(&c->btree_interior_update_wait);
535 mutex_unlock(&c->btree_interior_update_lock);
538 static void btree_update_add_key(struct btree_update *as,
539 struct keylist *keys, struct btree *b)
541 struct bkey_i *k = &b->key;
543 BUG_ON(bch2_keylist_u64s(keys) + k->k.u64s >
544 ARRAY_SIZE(as->_old_keys));
546 bkey_copy(keys->top, k);
547 bkey_i_to_btree_ptr_v2(keys->top)->v.mem_ptr = b->c.level + 1;
549 bch2_keylist_push(keys);
553 * The transactional part of an interior btree node update, where we journal the
554 * update we did to the interior node and update alloc info:
556 static int btree_update_nodes_written_trans(struct btree_trans *trans,
557 struct btree_update *as)
559 struct jset_entry *e = bch2_trans_jset_entry_alloc(trans, as->journal_u64s);
560 int ret = PTR_ERR_OR_ZERO(e);
564 memcpy(e, as->journal_entries, as->journal_u64s * sizeof(u64));
566 trans->journal_pin = &as->journal;
568 for_each_keylist_key(&as->old_keys, k) {
569 unsigned level = bkey_i_to_btree_ptr_v2(k)->v.mem_ptr;
571 ret = bch2_key_trigger_old(trans, as->btree_id, level, bkey_i_to_s_c(k),
572 BTREE_TRIGGER_TRANSACTIONAL);
577 for_each_keylist_key(&as->new_keys, k) {
578 unsigned level = bkey_i_to_btree_ptr_v2(k)->v.mem_ptr;
580 ret = bch2_key_trigger_new(trans, as->btree_id, level, bkey_i_to_s(k),
581 BTREE_TRIGGER_TRANSACTIONAL);
589 static void btree_update_nodes_written(struct btree_update *as)
591 struct bch_fs *c = as->c;
593 struct btree_trans *trans = bch2_trans_get(c);
599 * If we're already in an error state, it might be because a btree node
600 * was never written, and we might be trying to free that same btree
601 * node here, but it won't have been marked as allocated and we'll see
602 * spurious disk usage inconsistencies in the transactional part below
603 * if we don't skip it:
605 ret = bch2_journal_error(&c->journal);
610 * Wait for any in flight writes to finish before we free the old nodes
613 for (i = 0; i < as->nr_old_nodes; i++) {
616 b = as->old_nodes[i];
618 btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_read);
619 seq = b->data ? b->data->keys.seq : 0;
620 six_unlock_read(&b->c.lock);
622 if (seq == as->old_nodes_seq[i])
623 wait_on_bit_io(&b->flags, BTREE_NODE_write_in_flight_inner,
624 TASK_UNINTERRUPTIBLE);
628 * We did an update to a parent node where the pointers we added pointed
629 * to child nodes that weren't written yet: now, the child nodes have
630 * been written so we can write out the update to the interior node.
634 * We can't call into journal reclaim here: we'd block on the journal
635 * reclaim lock, but we may need to release the open buckets we have
636 * pinned in order for other btree updates to make forward progress, and
637 * journal reclaim does btree updates when flushing bkey_cached entries,
638 * which may require allocations as well.
640 ret = commit_do(trans, &as->disk_res, &journal_seq,
641 BCH_WATERMARK_reclaim|
642 BCH_TRANS_COMMIT_no_enospc|
643 BCH_TRANS_COMMIT_no_check_rw|
644 BCH_TRANS_COMMIT_journal_reclaim,
645 btree_update_nodes_written_trans(trans, as));
646 bch2_trans_unlock(trans);
648 bch2_fs_fatal_err_on(ret && !bch2_journal_error(&c->journal), c,
649 "%s(): error %s", __func__, bch2_err_str(ret));
654 btree_path_idx_t path_idx = get_unlocked_mut_path(trans,
655 as->btree_id, b->c.level, b->key.k.p);
656 struct btree_path *path = trans->paths + path_idx;
658 * @b is the node we did the final insert into:
660 * On failure to get a journal reservation, we still have to
661 * unblock the write and allow most of the write path to happen
662 * so that shutdown works, but the i->journal_seq mechanism
663 * won't work to prevent the btree write from being visible (we
664 * didn't get a journal sequence number) - instead
665 * __bch2_btree_node_write() doesn't do the actual write if
666 * we're in journal error state:
670 * Ensure transaction is unlocked before using
671 * btree_node_lock_nopath() (the use of which is always suspect,
672 * we need to work on removing this in the future)
674 * It should be, but get_unlocked_mut_path() -> bch2_path_get()
675 * calls bch2_path_upgrade(), before we call path_make_mut(), so
676 * we may rarely end up with a locked path besides the one we
679 bch2_trans_unlock(trans);
680 btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_intent);
681 mark_btree_node_locked(trans, path, b->c.level, BTREE_NODE_INTENT_LOCKED);
682 path->l[b->c.level].lock_seq = six_lock_seq(&b->c.lock);
683 path->l[b->c.level].b = b;
685 bch2_btree_node_lock_write_nofail(trans, path, &b->c);
687 mutex_lock(&c->btree_interior_update_lock);
689 list_del(&as->write_blocked_list);
690 if (list_empty(&b->write_blocked))
691 clear_btree_node_write_blocked(b);
694 * Node might have been freed, recheck under
695 * btree_interior_update_lock:
699 BUG_ON(!btree_node_dirty(b));
702 struct bset *last = btree_bset_last(b);
704 last->journal_seq = cpu_to_le64(
706 le64_to_cpu(last->journal_seq)));
708 bch2_btree_add_journal_pin(c, b, journal_seq);
711 * If we didn't get a journal sequence number we
712 * can't write this btree node, because recovery
713 * won't know to ignore this write:
715 set_btree_node_never_write(b);
719 mutex_unlock(&c->btree_interior_update_lock);
721 mark_btree_node_locked_noreset(path, b->c.level, BTREE_NODE_INTENT_LOCKED);
722 six_unlock_write(&b->c.lock);
724 btree_node_write_if_need(c, b, SIX_LOCK_intent);
725 btree_node_unlock(trans, path, b->c.level);
726 bch2_path_put(trans, path_idx, true);
729 bch2_journal_pin_drop(&c->journal, &as->journal);
731 mutex_lock(&c->btree_interior_update_lock);
732 for (i = 0; i < as->nr_new_nodes; i++) {
733 b = as->new_nodes[i];
735 BUG_ON(b->will_make_reachable != (unsigned long) as);
736 b->will_make_reachable = 0;
737 clear_btree_node_will_make_reachable(b);
739 mutex_unlock(&c->btree_interior_update_lock);
741 for (i = 0; i < as->nr_new_nodes; i++) {
742 b = as->new_nodes[i];
744 btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_read);
745 btree_node_write_if_need(c, b, SIX_LOCK_read);
746 six_unlock_read(&b->c.lock);
749 for (i = 0; i < as->nr_open_buckets; i++)
750 bch2_open_bucket_put(c, c->open_buckets + as->open_buckets[i]);
752 bch2_btree_update_free(as, trans);
753 bch2_trans_put(trans);
756 static void btree_interior_update_work(struct work_struct *work)
759 container_of(work, struct bch_fs, btree_interior_update_work);
760 struct btree_update *as;
763 mutex_lock(&c->btree_interior_update_lock);
764 as = list_first_entry_or_null(&c->btree_interior_updates_unwritten,
765 struct btree_update, unwritten_list);
766 if (as && !as->nodes_written)
768 mutex_unlock(&c->btree_interior_update_lock);
773 btree_update_nodes_written(as);
777 static CLOSURE_CALLBACK(btree_update_set_nodes_written)
779 closure_type(as, struct btree_update, cl);
780 struct bch_fs *c = as->c;
782 mutex_lock(&c->btree_interior_update_lock);
783 as->nodes_written = true;
784 mutex_unlock(&c->btree_interior_update_lock);
786 queue_work(c->btree_interior_update_worker, &c->btree_interior_update_work);
790 * We're updating @b with pointers to nodes that haven't finished writing yet:
791 * block @b from being written until @as completes
793 static void btree_update_updated_node(struct btree_update *as, struct btree *b)
795 struct bch_fs *c = as->c;
797 mutex_lock(&c->btree_interior_update_lock);
798 list_add_tail(&as->unwritten_list, &c->btree_interior_updates_unwritten);
800 BUG_ON(as->mode != BTREE_INTERIOR_NO_UPDATE);
801 BUG_ON(!btree_node_dirty(b));
804 as->mode = BTREE_INTERIOR_UPDATING_NODE;
807 set_btree_node_write_blocked(b);
808 list_add(&as->write_blocked_list, &b->write_blocked);
810 mutex_unlock(&c->btree_interior_update_lock);
813 static int bch2_update_reparent_journal_pin_flush(struct journal *j,
814 struct journal_entry_pin *_pin, u64 seq)
819 static void btree_update_reparent(struct btree_update *as,
820 struct btree_update *child)
822 struct bch_fs *c = as->c;
824 lockdep_assert_held(&c->btree_interior_update_lock);
827 child->mode = BTREE_INTERIOR_UPDATING_AS;
829 bch2_journal_pin_copy(&c->journal, &as->journal, &child->journal,
830 bch2_update_reparent_journal_pin_flush);
833 static void btree_update_updated_root(struct btree_update *as, struct btree *b)
835 struct bkey_i *insert = &b->key;
836 struct bch_fs *c = as->c;
838 BUG_ON(as->mode != BTREE_INTERIOR_NO_UPDATE);
840 BUG_ON(as->journal_u64s + jset_u64s(insert->k.u64s) >
841 ARRAY_SIZE(as->journal_entries));
844 journal_entry_set((void *) &as->journal_entries[as->journal_u64s],
845 BCH_JSET_ENTRY_btree_root,
846 b->c.btree_id, b->c.level,
847 insert, insert->k.u64s);
849 mutex_lock(&c->btree_interior_update_lock);
850 list_add_tail(&as->unwritten_list, &c->btree_interior_updates_unwritten);
852 as->mode = BTREE_INTERIOR_UPDATING_ROOT;
853 mutex_unlock(&c->btree_interior_update_lock);
857 * bch2_btree_update_add_new_node:
859 * This causes @as to wait on @b to be written, before it gets to
860 * bch2_btree_update_nodes_written
862 * Additionally, it sets b->will_make_reachable to prevent any additional writes
863 * to @b from happening besides the first until @b is reachable on disk
865 * And it adds @b to the list of @as's new nodes, so that we can update sector
866 * counts in bch2_btree_update_nodes_written:
868 static void bch2_btree_update_add_new_node(struct btree_update *as, struct btree *b)
870 struct bch_fs *c = as->c;
872 closure_get(&as->cl);
874 mutex_lock(&c->btree_interior_update_lock);
875 BUG_ON(as->nr_new_nodes >= ARRAY_SIZE(as->new_nodes));
876 BUG_ON(b->will_make_reachable);
878 as->new_nodes[as->nr_new_nodes++] = b;
879 b->will_make_reachable = 1UL|(unsigned long) as;
880 set_btree_node_will_make_reachable(b);
882 mutex_unlock(&c->btree_interior_update_lock);
884 btree_update_add_key(as, &as->new_keys, b);
886 if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
887 unsigned bytes = vstruct_end(&b->data->keys) - (void *) b->data;
888 unsigned sectors = round_up(bytes, block_bytes(c)) >> 9;
890 bkey_i_to_btree_ptr_v2(&b->key)->v.sectors_written =
891 cpu_to_le16(sectors);
896 * returns true if @b was a new node
898 static void btree_update_drop_new_node(struct bch_fs *c, struct btree *b)
900 struct btree_update *as;
904 mutex_lock(&c->btree_interior_update_lock);
906 * When b->will_make_reachable != 0, it owns a ref on as->cl that's
907 * dropped when it gets written by bch2_btree_complete_write - the
908 * xchg() is for synchronization with bch2_btree_complete_write:
910 v = xchg(&b->will_make_reachable, 0);
911 clear_btree_node_will_make_reachable(b);
912 as = (struct btree_update *) (v & ~1UL);
915 mutex_unlock(&c->btree_interior_update_lock);
919 for (i = 0; i < as->nr_new_nodes; i++)
920 if (as->new_nodes[i] == b)
925 array_remove_item(as->new_nodes, as->nr_new_nodes, i);
926 mutex_unlock(&c->btree_interior_update_lock);
929 closure_put(&as->cl);
932 static void bch2_btree_update_get_open_buckets(struct btree_update *as, struct btree *b)
935 as->open_buckets[as->nr_open_buckets++] =
939 static int bch2_btree_update_will_free_node_journal_pin_flush(struct journal *j,
940 struct journal_entry_pin *_pin, u64 seq)
946 * @b is being split/rewritten: it may have pointers to not-yet-written btree
947 * nodes and thus outstanding btree_updates - redirect @b's
948 * btree_updates to point to this btree_update:
950 static void bch2_btree_interior_update_will_free_node(struct btree_update *as,
953 struct bch_fs *c = as->c;
954 struct btree_update *p, *n;
955 struct btree_write *w;
957 set_btree_node_dying(b);
959 if (btree_node_fake(b))
962 mutex_lock(&c->btree_interior_update_lock);
965 * Does this node have any btree_update operations preventing
966 * it from being written?
968 * If so, redirect them to point to this btree_update: we can
969 * write out our new nodes, but we won't make them visible until those
970 * operations complete
972 list_for_each_entry_safe(p, n, &b->write_blocked, write_blocked_list) {
973 list_del_init(&p->write_blocked_list);
974 btree_update_reparent(as, p);
977 * for flush_held_btree_writes() waiting on updates to flush or
978 * nodes to be writeable:
980 closure_wake_up(&c->btree_interior_update_wait);
983 clear_btree_node_dirty_acct(c, b);
984 clear_btree_node_need_write(b);
985 clear_btree_node_write_blocked(b);
988 * Does this node have unwritten data that has a pin on the journal?
990 * If so, transfer that pin to the btree_update operation -
991 * note that if we're freeing multiple nodes, we only need to keep the
992 * oldest pin of any of the nodes we're freeing. We'll release the pin
993 * when the new nodes are persistent and reachable on disk:
995 w = btree_current_write(b);
996 bch2_journal_pin_copy(&c->journal, &as->journal, &w->journal,
997 bch2_btree_update_will_free_node_journal_pin_flush);
998 bch2_journal_pin_drop(&c->journal, &w->journal);
1000 w = btree_prev_write(b);
1001 bch2_journal_pin_copy(&c->journal, &as->journal, &w->journal,
1002 bch2_btree_update_will_free_node_journal_pin_flush);
1003 bch2_journal_pin_drop(&c->journal, &w->journal);
1005 mutex_unlock(&c->btree_interior_update_lock);
1008 * Is this a node that isn't reachable on disk yet?
1010 * Nodes that aren't reachable yet have writes blocked until they're
1011 * reachable - now that we've cancelled any pending writes and moved
1012 * things waiting on that write to wait on this update, we can drop this
1013 * node from the list of nodes that the other update is making
1014 * reachable, prior to freeing it:
1016 btree_update_drop_new_node(c, b);
1018 btree_update_add_key(as, &as->old_keys, b);
1020 as->old_nodes[as->nr_old_nodes] = b;
1021 as->old_nodes_seq[as->nr_old_nodes] = b->data->keys.seq;
1025 static void bch2_btree_update_done(struct btree_update *as, struct btree_trans *trans)
1027 struct bch_fs *c = as->c;
1028 u64 start_time = as->start_time;
1030 BUG_ON(as->mode == BTREE_INTERIOR_NO_UPDATE);
1032 if (as->took_gc_lock)
1033 up_read(&as->c->gc_lock);
1034 as->took_gc_lock = false;
1036 bch2_btree_reserve_put(as, trans);
1038 continue_at(&as->cl, btree_update_set_nodes_written,
1039 as->c->btree_interior_update_worker);
1041 bch2_time_stats_update(&c->times[BCH_TIME_btree_interior_update_foreground],
1045 static struct btree_update *
1046 bch2_btree_update_start(struct btree_trans *trans, struct btree_path *path,
1047 unsigned level, bool split, unsigned flags)
1049 struct bch_fs *c = trans->c;
1050 struct btree_update *as;
1051 u64 start_time = local_clock();
1052 int disk_res_flags = (flags & BCH_TRANS_COMMIT_no_enospc)
1053 ? BCH_DISK_RESERVATION_NOFAIL : 0;
1054 unsigned nr_nodes[2] = { 0, 0 };
1055 unsigned update_level = level;
1056 enum bch_watermark watermark = flags & BCH_WATERMARK_MASK;
1058 u32 restart_count = trans->restart_count;
1060 BUG_ON(!path->should_be_locked);
1062 if (watermark == BCH_WATERMARK_copygc)
1063 watermark = BCH_WATERMARK_btree_copygc;
1064 if (watermark < BCH_WATERMARK_btree)
1065 watermark = BCH_WATERMARK_btree;
1067 flags &= ~BCH_WATERMARK_MASK;
1070 if (!(flags & BCH_TRANS_COMMIT_journal_reclaim) &&
1071 watermark < c->journal.watermark) {
1072 struct journal_res res = { 0 };
1074 ret = drop_locks_do(trans,
1075 bch2_journal_res_get(&c->journal, &res, 1,
1076 watermark|JOURNAL_RES_GET_CHECK));
1078 return ERR_PTR(ret);
1082 nr_nodes[!!update_level] += 1 + split;
1085 ret = bch2_btree_path_upgrade(trans, path, update_level + 1);
1087 return ERR_PTR(ret);
1089 if (!btree_path_node(path, update_level)) {
1090 /* Allocating new root? */
1091 nr_nodes[1] += split;
1092 update_level = BTREE_MAX_DEPTH;
1097 * Always check for space for two keys, even if we won't have to
1098 * split at prior level - it might have been a merge instead:
1100 if (bch2_btree_node_insert_fits(path->l[update_level].b,
1101 BKEY_BTREE_PTR_U64s_MAX * 2))
1104 split = path->l[update_level].b->nr.live_u64s > BTREE_SPLIT_THRESHOLD(c);
1107 if (!down_read_trylock(&c->gc_lock)) {
1108 ret = drop_locks_do(trans, (down_read(&c->gc_lock), 0));
1110 up_read(&c->gc_lock);
1111 return ERR_PTR(ret);
1115 as = mempool_alloc(&c->btree_interior_update_pool, GFP_NOFS);
1116 memset(as, 0, sizeof(*as));
1117 closure_init(&as->cl, NULL);
1119 as->start_time = start_time;
1120 as->mode = BTREE_INTERIOR_NO_UPDATE;
1121 as->took_gc_lock = true;
1122 as->btree_id = path->btree_id;
1123 as->update_level = update_level;
1124 INIT_LIST_HEAD(&as->list);
1125 INIT_LIST_HEAD(&as->unwritten_list);
1126 INIT_LIST_HEAD(&as->write_blocked_list);
1127 bch2_keylist_init(&as->old_keys, as->_old_keys);
1128 bch2_keylist_init(&as->new_keys, as->_new_keys);
1129 bch2_keylist_init(&as->parent_keys, as->inline_keys);
1131 mutex_lock(&c->btree_interior_update_lock);
1132 list_add_tail(&as->list, &c->btree_interior_update_list);
1133 mutex_unlock(&c->btree_interior_update_lock);
1136 * We don't want to allocate if we're in an error state, that can cause
1137 * deadlock on emergency shutdown due to open buckets getting stuck in
1138 * the btree_reserve_cache after allocator shutdown has cleared it out.
1139 * This check needs to come after adding us to the btree_interior_update
1140 * list but before calling bch2_btree_reserve_get, to synchronize with
1141 * __bch2_fs_read_only().
1143 ret = bch2_journal_error(&c->journal);
1147 ret = bch2_disk_reservation_get(c, &as->disk_res,
1148 (nr_nodes[0] + nr_nodes[1]) * btree_sectors(c),
1149 c->opts.metadata_replicas,
1154 ret = bch2_btree_reserve_get(trans, as, nr_nodes, flags, NULL);
1155 if (bch2_err_matches(ret, ENOSPC) ||
1156 bch2_err_matches(ret, ENOMEM)) {
1160 * XXX: this should probably be a separate BTREE_INSERT_NONBLOCK
1163 if (bch2_err_matches(ret, ENOSPC) &&
1164 (flags & BCH_TRANS_COMMIT_journal_reclaim) &&
1165 watermark != BCH_WATERMARK_reclaim) {
1166 ret = -BCH_ERR_journal_reclaim_would_deadlock;
1170 closure_init_stack(&cl);
1173 ret = bch2_btree_reserve_get(trans, as, nr_nodes, flags, &cl);
1175 bch2_trans_unlock(trans);
1177 } while (bch2_err_matches(ret, BCH_ERR_operation_blocked));
1181 trace_and_count(c, btree_reserve_get_fail, trans->fn,
1182 _RET_IP_, nr_nodes[0] + nr_nodes[1], ret);
1186 ret = bch2_trans_relock(trans);
1190 bch2_trans_verify_not_restarted(trans, restart_count);
1193 bch2_btree_update_free(as, trans);
1194 if (!bch2_err_matches(ret, ENOSPC) &&
1195 !bch2_err_matches(ret, EROFS))
1196 bch_err_fn_ratelimited(c, ret);
1197 return ERR_PTR(ret);
1200 /* Btree root updates: */
1202 static void bch2_btree_set_root_inmem(struct bch_fs *c, struct btree *b)
1204 /* Root nodes cannot be reaped */
1205 mutex_lock(&c->btree_cache.lock);
1206 list_del_init(&b->list);
1207 mutex_unlock(&c->btree_cache.lock);
1209 mutex_lock(&c->btree_root_lock);
1210 BUG_ON(btree_node_root(c, b) &&
1211 (b->c.level < btree_node_root(c, b)->c.level ||
1212 !btree_node_dying(btree_node_root(c, b))));
1214 bch2_btree_id_root(c, b->c.btree_id)->b = b;
1215 mutex_unlock(&c->btree_root_lock);
1217 bch2_recalc_btree_reserve(c);
1220 static void bch2_btree_set_root(struct btree_update *as,
1221 struct btree_trans *trans,
1222 struct btree_path *path,
1225 struct bch_fs *c = as->c;
1228 trace_and_count(c, btree_node_set_root, trans, b);
1230 old = btree_node_root(c, b);
1233 * Ensure no one is using the old root while we switch to the
1236 bch2_btree_node_lock_write_nofail(trans, path, &old->c);
1238 bch2_btree_set_root_inmem(c, b);
1240 btree_update_updated_root(as, b);
1243 * Unlock old root after new root is visible:
1245 * The new root isn't persistent, but that's ok: we still have
1246 * an intent lock on the new root, and any updates that would
1247 * depend on the new root would have to update the new root.
1249 bch2_btree_node_unlock_write(trans, path, old);
1252 /* Interior node updates: */
1254 static void bch2_insert_fixup_btree_ptr(struct btree_update *as,
1255 struct btree_trans *trans,
1256 struct btree_path *path,
1258 struct btree_node_iter *node_iter,
1259 struct bkey_i *insert)
1261 struct bch_fs *c = as->c;
1262 struct bkey_packed *k;
1263 struct printbuf buf = PRINTBUF;
1264 unsigned long old, new, v;
1266 BUG_ON(insert->k.type == KEY_TYPE_btree_ptr_v2 &&
1267 !btree_ptr_sectors_written(insert));
1269 if (unlikely(!test_bit(JOURNAL_REPLAY_DONE, &c->journal.flags)))
1270 bch2_journal_key_overwritten(c, b->c.btree_id, b->c.level, insert->k.p);
1272 if (bch2_bkey_invalid(c, bkey_i_to_s_c(insert),
1273 btree_node_type(b), WRITE, &buf) ?:
1274 bch2_bkey_in_btree_node(c, b, bkey_i_to_s_c(insert), &buf)) {
1275 printbuf_reset(&buf);
1276 prt_printf(&buf, "inserting invalid bkey\n ");
1277 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(insert));
1278 prt_printf(&buf, "\n ");
1279 bch2_bkey_invalid(c, bkey_i_to_s_c(insert),
1280 btree_node_type(b), WRITE, &buf);
1281 bch2_bkey_in_btree_node(c, b, bkey_i_to_s_c(insert), &buf);
1283 bch2_fs_inconsistent(c, "%s", buf.buf);
1287 BUG_ON(as->journal_u64s + jset_u64s(insert->k.u64s) >
1288 ARRAY_SIZE(as->journal_entries));
1291 journal_entry_set((void *) &as->journal_entries[as->journal_u64s],
1292 BCH_JSET_ENTRY_btree_keys,
1293 b->c.btree_id, b->c.level,
1294 insert, insert->k.u64s);
1296 while ((k = bch2_btree_node_iter_peek_all(node_iter, b)) &&
1297 bkey_iter_pos_cmp(b, k, &insert->k.p) < 0)
1298 bch2_btree_node_iter_advance(node_iter, b);
1300 bch2_btree_bset_insert_key(trans, path, b, node_iter, insert);
1301 set_btree_node_dirty_acct(c, b);
1303 v = READ_ONCE(b->flags);
1307 new &= ~BTREE_WRITE_TYPE_MASK;
1308 new |= BTREE_WRITE_interior;
1309 new |= 1 << BTREE_NODE_need_write;
1310 } while ((v = cmpxchg(&b->flags, old, new)) != old);
1312 printbuf_exit(&buf);
1316 __bch2_btree_insert_keys_interior(struct btree_update *as,
1317 struct btree_trans *trans,
1318 struct btree_path *path,
1320 struct btree_node_iter node_iter,
1321 struct keylist *keys)
1323 struct bkey_i *insert = bch2_keylist_front(keys);
1324 struct bkey_packed *k;
1326 BUG_ON(btree_node_type(b) != BKEY_TYPE_btree);
1328 while ((k = bch2_btree_node_iter_prev_all(&node_iter, b)) &&
1329 (bkey_cmp_left_packed(b, k, &insert->k.p) >= 0))
1332 while (!bch2_keylist_empty(keys)) {
1333 insert = bch2_keylist_front(keys);
1335 if (bpos_gt(insert->k.p, b->key.k.p))
1338 bch2_insert_fixup_btree_ptr(as, trans, path, b, &node_iter, insert);
1339 bch2_keylist_pop_front(keys);
1344 * Move keys from n1 (original replacement node, now lower node) to n2 (higher
1347 static void __btree_split_node(struct btree_update *as,
1348 struct btree_trans *trans,
1352 struct bkey_packed *k;
1353 struct bpos n1_pos = POS_MIN;
1354 struct btree_node_iter iter;
1355 struct bset *bsets[2];
1356 struct bkey_format_state format[2];
1357 struct bkey_packed *out[2];
1359 unsigned u64s, n1_u64s = (b->nr.live_u64s * 3) / 5;
1360 struct { unsigned nr_keys, val_u64s; } nr_keys[2];
1363 memset(&nr_keys, 0, sizeof(nr_keys));
1365 for (i = 0; i < 2; i++) {
1366 BUG_ON(n[i]->nsets != 1);
1368 bsets[i] = btree_bset_first(n[i]);
1369 out[i] = bsets[i]->start;
1371 SET_BTREE_NODE_SEQ(n[i]->data, BTREE_NODE_SEQ(b->data) + 1);
1372 bch2_bkey_format_init(&format[i]);
1376 for_each_btree_node_key(b, k, &iter) {
1377 if (bkey_deleted(k))
1380 i = u64s >= n1_u64s;
1382 uk = bkey_unpack_key(b, k);
1385 bch2_bkey_format_add_key(&format[i], &uk);
1387 nr_keys[i].nr_keys++;
1388 nr_keys[i].val_u64s += bkeyp_val_u64s(&b->format, k);
1391 btree_set_min(n[0], b->data->min_key);
1392 btree_set_max(n[0], n1_pos);
1393 btree_set_min(n[1], bpos_successor(n1_pos));
1394 btree_set_max(n[1], b->data->max_key);
1396 for (i = 0; i < 2; i++) {
1397 bch2_bkey_format_add_pos(&format[i], n[i]->data->min_key);
1398 bch2_bkey_format_add_pos(&format[i], n[i]->data->max_key);
1400 n[i]->data->format = bch2_bkey_format_done(&format[i]);
1402 unsigned u64s = nr_keys[i].nr_keys * n[i]->data->format.key_u64s +
1403 nr_keys[i].val_u64s;
1404 if (__vstruct_bytes(struct btree_node, u64s) > btree_buf_bytes(b))
1405 n[i]->data->format = b->format;
1407 btree_node_set_format(n[i], n[i]->data->format);
1411 for_each_btree_node_key(b, k, &iter) {
1412 if (bkey_deleted(k))
1415 i = u64s >= n1_u64s;
1418 if (bch2_bkey_transform(&n[i]->format, out[i], bkey_packed(k)
1419 ? &b->format: &bch2_bkey_format_current, k))
1420 out[i]->format = KEY_FORMAT_LOCAL_BTREE;
1422 bch2_bkey_unpack(b, (void *) out[i], k);
1424 out[i]->needs_whiteout = false;
1426 btree_keys_account_key_add(&n[i]->nr, 0, out[i]);
1427 out[i] = bkey_p_next(out[i]);
1430 for (i = 0; i < 2; i++) {
1431 bsets[i]->u64s = cpu_to_le16((u64 *) out[i] - bsets[i]->_data);
1433 BUG_ON(!bsets[i]->u64s);
1435 set_btree_bset_end(n[i], n[i]->set);
1437 btree_node_reset_sib_u64s(n[i]);
1439 bch2_verify_btree_nr_keys(n[i]);
1442 btree_node_interior_verify(as->c, n[i]);
1447 * For updates to interior nodes, we've got to do the insert before we split
1448 * because the stuff we're inserting has to be inserted atomically. Post split,
1449 * the keys might have to go in different nodes and the split would no longer be
1452 * Worse, if the insert is from btree node coalescing, if we do the insert after
1453 * we do the split (and pick the pivot) - the pivot we pick might be between
1454 * nodes that were coalesced, and thus in the middle of a child node post
1457 static void btree_split_insert_keys(struct btree_update *as,
1458 struct btree_trans *trans,
1459 btree_path_idx_t path_idx,
1461 struct keylist *keys)
1463 struct btree_path *path = trans->paths + path_idx;
1465 if (!bch2_keylist_empty(keys) &&
1466 bpos_le(bch2_keylist_front(keys)->k.p, b->data->max_key)) {
1467 struct btree_node_iter node_iter;
1469 bch2_btree_node_iter_init(&node_iter, b, &bch2_keylist_front(keys)->k.p);
1471 __bch2_btree_insert_keys_interior(as, trans, path, b, node_iter, keys);
1473 btree_node_interior_verify(as->c, b);
1477 static int btree_split(struct btree_update *as, struct btree_trans *trans,
1478 btree_path_idx_t path, struct btree *b,
1479 struct keylist *keys, unsigned flags)
1481 struct bch_fs *c = as->c;
1482 struct btree *parent = btree_node_parent(trans->paths + path, b);
1483 struct btree *n1, *n2 = NULL, *n3 = NULL;
1484 btree_path_idx_t path1 = 0, path2 = 0;
1485 u64 start_time = local_clock();
1488 BUG_ON(!parent && (b != btree_node_root(c, b)));
1489 BUG_ON(parent && !btree_node_intent_locked(trans->paths + path, b->c.level + 1));
1491 bch2_btree_interior_update_will_free_node(as, b);
1493 if (b->nr.live_u64s > BTREE_SPLIT_THRESHOLD(c)) {
1496 trace_and_count(c, btree_node_split, trans, b);
1498 n[0] = n1 = bch2_btree_node_alloc(as, trans, b->c.level);
1499 n[1] = n2 = bch2_btree_node_alloc(as, trans, b->c.level);
1501 __btree_split_node(as, trans, b, n);
1504 btree_split_insert_keys(as, trans, path, n1, keys);
1505 btree_split_insert_keys(as, trans, path, n2, keys);
1506 BUG_ON(!bch2_keylist_empty(keys));
1509 bch2_btree_build_aux_trees(n2);
1510 bch2_btree_build_aux_trees(n1);
1512 bch2_btree_update_add_new_node(as, n1);
1513 bch2_btree_update_add_new_node(as, n2);
1514 six_unlock_write(&n2->c.lock);
1515 six_unlock_write(&n1->c.lock);
1517 path1 = get_unlocked_mut_path(trans, as->btree_id, n1->c.level, n1->key.k.p);
1518 six_lock_increment(&n1->c.lock, SIX_LOCK_intent);
1519 mark_btree_node_locked(trans, trans->paths + path1, n1->c.level, BTREE_NODE_INTENT_LOCKED);
1520 bch2_btree_path_level_init(trans, trans->paths + path1, n1);
1522 path2 = get_unlocked_mut_path(trans, as->btree_id, n2->c.level, n2->key.k.p);
1523 six_lock_increment(&n2->c.lock, SIX_LOCK_intent);
1524 mark_btree_node_locked(trans, trans->paths + path2, n2->c.level, BTREE_NODE_INTENT_LOCKED);
1525 bch2_btree_path_level_init(trans, trans->paths + path2, n2);
1528 * Note that on recursive parent_keys == keys, so we
1529 * can't start adding new keys to parent_keys before emptying it
1530 * out (which we did with btree_split_insert_keys() above)
1532 bch2_keylist_add(&as->parent_keys, &n1->key);
1533 bch2_keylist_add(&as->parent_keys, &n2->key);
1536 /* Depth increases, make a new root */
1537 n3 = __btree_root_alloc(as, trans, b->c.level + 1);
1539 bch2_btree_update_add_new_node(as, n3);
1540 six_unlock_write(&n3->c.lock);
1542 trans->paths[path2].locks_want++;
1543 BUG_ON(btree_node_locked(trans->paths + path2, n3->c.level));
1544 six_lock_increment(&n3->c.lock, SIX_LOCK_intent);
1545 mark_btree_node_locked(trans, trans->paths + path2, n3->c.level, BTREE_NODE_INTENT_LOCKED);
1546 bch2_btree_path_level_init(trans, trans->paths + path2, n3);
1548 n3->sib_u64s[0] = U16_MAX;
1549 n3->sib_u64s[1] = U16_MAX;
1551 btree_split_insert_keys(as, trans, path, n3, &as->parent_keys);
1554 trace_and_count(c, btree_node_compact, trans, b);
1556 n1 = bch2_btree_node_alloc_replacement(as, trans, b);
1559 btree_split_insert_keys(as, trans, path, n1, keys);
1560 BUG_ON(!bch2_keylist_empty(keys));
1563 bch2_btree_build_aux_trees(n1);
1564 bch2_btree_update_add_new_node(as, n1);
1565 six_unlock_write(&n1->c.lock);
1567 path1 = get_unlocked_mut_path(trans, as->btree_id, n1->c.level, n1->key.k.p);
1568 six_lock_increment(&n1->c.lock, SIX_LOCK_intent);
1569 mark_btree_node_locked(trans, trans->paths + path1, n1->c.level, BTREE_NODE_INTENT_LOCKED);
1570 bch2_btree_path_level_init(trans, trans->paths + path1, n1);
1573 bch2_keylist_add(&as->parent_keys, &n1->key);
1576 /* New nodes all written, now make them visible: */
1579 /* Split a non root node */
1580 ret = bch2_btree_insert_node(as, trans, path, parent, &as->parent_keys, flags);
1584 bch2_btree_set_root(as, trans, trans->paths + path, n3);
1586 /* Root filled up but didn't need to be split */
1587 bch2_btree_set_root(as, trans, trans->paths + path, n1);
1591 bch2_btree_update_get_open_buckets(as, n3);
1592 bch2_btree_node_write(c, n3, SIX_LOCK_intent, 0);
1595 bch2_btree_update_get_open_buckets(as, n2);
1596 bch2_btree_node_write(c, n2, SIX_LOCK_intent, 0);
1598 bch2_btree_update_get_open_buckets(as, n1);
1599 bch2_btree_node_write(c, n1, SIX_LOCK_intent, 0);
1602 * The old node must be freed (in memory) _before_ unlocking the new
1603 * nodes - else another thread could re-acquire a read lock on the old
1604 * node after another thread has locked and updated the new node, thus
1605 * seeing stale data:
1607 bch2_btree_node_free_inmem(trans, trans->paths + path, b);
1610 bch2_trans_node_add(trans, trans->paths + path, n3);
1612 bch2_trans_node_add(trans, trans->paths + path2, n2);
1613 bch2_trans_node_add(trans, trans->paths + path1, n1);
1616 six_unlock_intent(&n3->c.lock);
1618 six_unlock_intent(&n2->c.lock);
1619 six_unlock_intent(&n1->c.lock);
1622 __bch2_btree_path_unlock(trans, trans->paths + path2);
1623 bch2_path_put(trans, path2, true);
1626 __bch2_btree_path_unlock(trans, trans->paths + path1);
1627 bch2_path_put(trans, path1, true);
1630 bch2_trans_verify_locks(trans);
1632 bch2_time_stats_update(&c->times[n2
1633 ? BCH_TIME_btree_node_split
1634 : BCH_TIME_btree_node_compact],
1639 bch2_btree_node_free_never_used(as, trans, n3);
1641 bch2_btree_node_free_never_used(as, trans, n2);
1642 bch2_btree_node_free_never_used(as, trans, n1);
1647 bch2_btree_insert_keys_interior(struct btree_update *as,
1648 struct btree_trans *trans,
1649 struct btree_path *path,
1651 struct keylist *keys)
1653 struct btree_path *linked;
1656 __bch2_btree_insert_keys_interior(as, trans, path, b,
1657 path->l[b->c.level].iter, keys);
1659 btree_update_updated_node(as, b);
1661 trans_for_each_path_with_node(trans, b, linked, i)
1662 bch2_btree_node_iter_peek(&linked->l[b->c.level].iter, b);
1664 bch2_trans_verify_paths(trans);
1668 * bch2_btree_insert_node - insert bkeys into a given btree node
1670 * @as: btree_update object
1671 * @trans: btree_trans object
1672 * @path_idx: path that points to current node
1673 * @b: node to insert keys into
1674 * @keys: list of keys to insert
1675 * @flags: transaction commit flags
1677 * Returns: 0 on success, typically transaction restart error on failure
1679 * Inserts as many keys as it can into a given btree node, splitting it if full.
1680 * If a split occurred, this function will return early. This can only happen
1681 * for leaf nodes -- inserts into interior nodes have to be atomic.
1683 static int bch2_btree_insert_node(struct btree_update *as, struct btree_trans *trans,
1684 btree_path_idx_t path_idx, struct btree *b,
1685 struct keylist *keys, unsigned flags)
1687 struct bch_fs *c = as->c;
1688 struct btree_path *path = trans->paths + path_idx;
1689 int old_u64s = le16_to_cpu(btree_bset_last(b)->u64s);
1690 int old_live_u64s = b->nr.live_u64s;
1691 int live_u64s_added, u64s_added;
1694 lockdep_assert_held(&c->gc_lock);
1695 BUG_ON(!btree_node_intent_locked(path, b->c.level));
1696 BUG_ON(!b->c.level);
1697 BUG_ON(!as || as->b);
1698 bch2_verify_keylist_sorted(keys);
1700 ret = bch2_btree_node_lock_write(trans, path, &b->c);
1704 bch2_btree_node_prep_for_write(trans, path, b);
1706 if (!bch2_btree_node_insert_fits(b, bch2_keylist_u64s(keys))) {
1707 bch2_btree_node_unlock_write(trans, path, b);
1711 btree_node_interior_verify(c, b);
1713 bch2_btree_insert_keys_interior(as, trans, path, b, keys);
1715 live_u64s_added = (int) b->nr.live_u64s - old_live_u64s;
1716 u64s_added = (int) le16_to_cpu(btree_bset_last(b)->u64s) - old_u64s;
1718 if (b->sib_u64s[0] != U16_MAX && live_u64s_added < 0)
1719 b->sib_u64s[0] = max(0, (int) b->sib_u64s[0] + live_u64s_added);
1720 if (b->sib_u64s[1] != U16_MAX && live_u64s_added < 0)
1721 b->sib_u64s[1] = max(0, (int) b->sib_u64s[1] + live_u64s_added);
1723 if (u64s_added > live_u64s_added &&
1724 bch2_maybe_compact_whiteouts(c, b))
1725 bch2_trans_node_reinit_iter(trans, b);
1727 bch2_btree_node_unlock_write(trans, path, b);
1729 btree_node_interior_verify(c, b);
1733 * We could attempt to avoid the transaction restart, by calling
1734 * bch2_btree_path_upgrade() and allocating more nodes:
1736 if (b->c.level >= as->update_level) {
1737 trace_and_count(c, trans_restart_split_race, trans, _THIS_IP_, b);
1738 return btree_trans_restart(trans, BCH_ERR_transaction_restart_split_race);
1741 return btree_split(as, trans, path_idx, b, keys, flags);
1744 int bch2_btree_split_leaf(struct btree_trans *trans,
1745 btree_path_idx_t path,
1748 /* btree_split & merge may both cause paths array to be reallocated */
1750 struct btree *b = path_l(trans->paths + path)->b;
1751 struct btree_update *as;
1755 as = bch2_btree_update_start(trans, trans->paths + path,
1756 trans->paths[path].level,
1761 ret = btree_split(as, trans, path, b, NULL, flags);
1763 bch2_btree_update_free(as, trans);
1767 bch2_btree_update_done(as, trans);
1769 for (l = trans->paths[path].level + 1;
1770 btree_node_intent_locked(&trans->paths[path], l) && !ret;
1772 ret = bch2_foreground_maybe_merge(trans, path, l, flags);
1777 int __bch2_foreground_maybe_merge(struct btree_trans *trans,
1778 btree_path_idx_t path,
1781 enum btree_node_sibling sib)
1783 struct bch_fs *c = trans->c;
1784 struct btree_update *as;
1785 struct bkey_format_state new_s;
1786 struct bkey_format new_f;
1787 struct bkey_i delete;
1788 struct btree *b, *m, *n, *prev, *next, *parent;
1789 struct bpos sib_pos;
1791 enum btree_id btree = trans->paths[path].btree_id;
1792 btree_path_idx_t sib_path = 0, new_path = 0;
1793 u64 start_time = local_clock();
1796 BUG_ON(!trans->paths[path].should_be_locked);
1797 BUG_ON(!btree_node_locked(&trans->paths[path], level));
1799 b = trans->paths[path].l[level].b;
1801 if ((sib == btree_prev_sib && bpos_eq(b->data->min_key, POS_MIN)) ||
1802 (sib == btree_next_sib && bpos_eq(b->data->max_key, SPOS_MAX))) {
1803 b->sib_u64s[sib] = U16_MAX;
1807 sib_pos = sib == btree_prev_sib
1808 ? bpos_predecessor(b->data->min_key)
1809 : bpos_successor(b->data->max_key);
1811 sib_path = bch2_path_get(trans, btree, sib_pos,
1812 U8_MAX, level, BTREE_ITER_INTENT, _THIS_IP_);
1813 ret = bch2_btree_path_traverse(trans, sib_path, false);
1817 btree_path_set_should_be_locked(trans->paths + sib_path);
1819 m = trans->paths[sib_path].l[level].b;
1821 if (btree_node_parent(trans->paths + path, b) !=
1822 btree_node_parent(trans->paths + sib_path, m)) {
1823 b->sib_u64s[sib] = U16_MAX;
1827 if (sib == btree_prev_sib) {
1835 if (!bpos_eq(bpos_successor(prev->data->max_key), next->data->min_key)) {
1836 struct printbuf buf1 = PRINTBUF, buf2 = PRINTBUF;
1838 bch2_bpos_to_text(&buf1, prev->data->max_key);
1839 bch2_bpos_to_text(&buf2, next->data->min_key);
1841 "%s(): btree topology error:\n"
1842 " prev ends at %s\n"
1843 " next starts at %s",
1844 __func__, buf1.buf, buf2.buf);
1845 printbuf_exit(&buf1);
1846 printbuf_exit(&buf2);
1847 bch2_topology_error(c);
1852 bch2_bkey_format_init(&new_s);
1853 bch2_bkey_format_add_pos(&new_s, prev->data->min_key);
1854 __bch2_btree_calc_format(&new_s, prev);
1855 __bch2_btree_calc_format(&new_s, next);
1856 bch2_bkey_format_add_pos(&new_s, next->data->max_key);
1857 new_f = bch2_bkey_format_done(&new_s);
1859 sib_u64s = btree_node_u64s_with_format(b->nr, &b->format, &new_f) +
1860 btree_node_u64s_with_format(m->nr, &m->format, &new_f);
1862 if (sib_u64s > BTREE_FOREGROUND_MERGE_HYSTERESIS(c)) {
1863 sib_u64s -= BTREE_FOREGROUND_MERGE_HYSTERESIS(c);
1865 sib_u64s += BTREE_FOREGROUND_MERGE_HYSTERESIS(c);
1868 sib_u64s = min(sib_u64s, btree_max_u64s(c));
1869 sib_u64s = min(sib_u64s, (size_t) U16_MAX - 1);
1870 b->sib_u64s[sib] = sib_u64s;
1872 if (b->sib_u64s[sib] > c->btree_foreground_merge_threshold)
1875 parent = btree_node_parent(trans->paths + path, b);
1876 as = bch2_btree_update_start(trans, trans->paths + path, level, false,
1877 BCH_TRANS_COMMIT_no_enospc|flags);
1878 ret = PTR_ERR_OR_ZERO(as);
1882 trace_and_count(c, btree_node_merge, trans, b);
1884 bch2_btree_interior_update_will_free_node(as, b);
1885 bch2_btree_interior_update_will_free_node(as, m);
1887 n = bch2_btree_node_alloc(as, trans, b->c.level);
1889 SET_BTREE_NODE_SEQ(n->data,
1890 max(BTREE_NODE_SEQ(b->data),
1891 BTREE_NODE_SEQ(m->data)) + 1);
1893 btree_set_min(n, prev->data->min_key);
1894 btree_set_max(n, next->data->max_key);
1896 n->data->format = new_f;
1897 btree_node_set_format(n, new_f);
1899 bch2_btree_sort_into(c, n, prev);
1900 bch2_btree_sort_into(c, n, next);
1902 bch2_btree_build_aux_trees(n);
1903 bch2_btree_update_add_new_node(as, n);
1904 six_unlock_write(&n->c.lock);
1906 new_path = get_unlocked_mut_path(trans, btree, n->c.level, n->key.k.p);
1907 six_lock_increment(&n->c.lock, SIX_LOCK_intent);
1908 mark_btree_node_locked(trans, trans->paths + new_path, n->c.level, BTREE_NODE_INTENT_LOCKED);
1909 bch2_btree_path_level_init(trans, trans->paths + new_path, n);
1911 bkey_init(&delete.k);
1912 delete.k.p = prev->key.k.p;
1913 bch2_keylist_add(&as->parent_keys, &delete);
1914 bch2_keylist_add(&as->parent_keys, &n->key);
1916 bch2_trans_verify_paths(trans);
1918 ret = bch2_btree_insert_node(as, trans, path, parent, &as->parent_keys, flags);
1920 goto err_free_update;
1922 bch2_trans_verify_paths(trans);
1924 bch2_btree_update_get_open_buckets(as, n);
1925 bch2_btree_node_write(c, n, SIX_LOCK_intent, 0);
1927 bch2_btree_node_free_inmem(trans, trans->paths + path, b);
1928 bch2_btree_node_free_inmem(trans, trans->paths + sib_path, m);
1930 bch2_trans_node_add(trans, trans->paths + path, n);
1932 bch2_trans_verify_paths(trans);
1934 six_unlock_intent(&n->c.lock);
1936 bch2_btree_update_done(as, trans);
1938 bch2_time_stats_update(&c->times[BCH_TIME_btree_node_merge], start_time);
1942 bch2_path_put(trans, new_path, true);
1943 bch2_path_put(trans, sib_path, true);
1944 bch2_trans_verify_locks(trans);
1947 bch2_btree_node_free_never_used(as, trans, n);
1948 bch2_btree_update_free(as, trans);
1952 int bch2_btree_node_rewrite(struct btree_trans *trans,
1953 struct btree_iter *iter,
1957 struct bch_fs *c = trans->c;
1958 struct btree *n, *parent;
1959 struct btree_update *as;
1960 btree_path_idx_t new_path = 0;
1963 flags |= BCH_TRANS_COMMIT_no_enospc;
1965 struct btree_path *path = btree_iter_path(trans, iter);
1966 parent = btree_node_parent(path, b);
1967 as = bch2_btree_update_start(trans, path, b->c.level, false, flags);
1968 ret = PTR_ERR_OR_ZERO(as);
1972 bch2_btree_interior_update_will_free_node(as, b);
1974 n = bch2_btree_node_alloc_replacement(as, trans, b);
1976 bch2_btree_build_aux_trees(n);
1977 bch2_btree_update_add_new_node(as, n);
1978 six_unlock_write(&n->c.lock);
1980 new_path = get_unlocked_mut_path(trans, iter->btree_id, n->c.level, n->key.k.p);
1981 six_lock_increment(&n->c.lock, SIX_LOCK_intent);
1982 mark_btree_node_locked(trans, trans->paths + new_path, n->c.level, BTREE_NODE_INTENT_LOCKED);
1983 bch2_btree_path_level_init(trans, trans->paths + new_path, n);
1985 trace_and_count(c, btree_node_rewrite, trans, b);
1988 bch2_keylist_add(&as->parent_keys, &n->key);
1989 ret = bch2_btree_insert_node(as, trans, iter->path,
1990 parent, &as->parent_keys, flags);
1994 bch2_btree_set_root(as, trans, btree_iter_path(trans, iter), n);
1997 bch2_btree_update_get_open_buckets(as, n);
1998 bch2_btree_node_write(c, n, SIX_LOCK_intent, 0);
2000 bch2_btree_node_free_inmem(trans, btree_iter_path(trans, iter), b);
2002 bch2_trans_node_add(trans, trans->paths + iter->path, n);
2003 six_unlock_intent(&n->c.lock);
2005 bch2_btree_update_done(as, trans);
2008 bch2_path_put(trans, new_path, true);
2009 bch2_trans_downgrade(trans);
2012 bch2_btree_node_free_never_used(as, trans, n);
2013 bch2_btree_update_free(as, trans);
2017 struct async_btree_rewrite {
2019 struct work_struct work;
2020 struct list_head list;
2021 enum btree_id btree_id;
2027 static int async_btree_node_rewrite_trans(struct btree_trans *trans,
2028 struct async_btree_rewrite *a)
2030 struct bch_fs *c = trans->c;
2031 struct btree_iter iter;
2035 bch2_trans_node_iter_init(trans, &iter, a->btree_id, a->pos,
2036 BTREE_MAX_DEPTH, a->level, 0);
2037 b = bch2_btree_iter_peek_node(&iter);
2038 ret = PTR_ERR_OR_ZERO(b);
2042 if (!b || b->data->keys.seq != a->seq) {
2043 struct printbuf buf = PRINTBUF;
2046 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(&b->key));
2048 prt_str(&buf, "(null");
2049 bch_info(c, "%s: node to rewrite not found:, searching for seq %llu, got\n%s",
2050 __func__, a->seq, buf.buf);
2051 printbuf_exit(&buf);
2055 ret = bch2_btree_node_rewrite(trans, &iter, b, 0);
2057 bch2_trans_iter_exit(trans, &iter);
2062 static void async_btree_node_rewrite_work(struct work_struct *work)
2064 struct async_btree_rewrite *a =
2065 container_of(work, struct async_btree_rewrite, work);
2066 struct bch_fs *c = a->c;
2069 ret = bch2_trans_do(c, NULL, NULL, 0,
2070 async_btree_node_rewrite_trans(trans, a));
2072 bch2_write_ref_put(c, BCH_WRITE_REF_node_rewrite);
2076 void bch2_btree_node_rewrite_async(struct bch_fs *c, struct btree *b)
2078 struct async_btree_rewrite *a;
2081 a = kmalloc(sizeof(*a), GFP_NOFS);
2083 bch_err(c, "%s: error allocating memory", __func__);
2088 a->btree_id = b->c.btree_id;
2089 a->level = b->c.level;
2090 a->pos = b->key.k.p;
2091 a->seq = b->data->keys.seq;
2092 INIT_WORK(&a->work, async_btree_node_rewrite_work);
2094 if (unlikely(!test_bit(BCH_FS_may_go_rw, &c->flags))) {
2095 mutex_lock(&c->pending_node_rewrites_lock);
2096 list_add(&a->list, &c->pending_node_rewrites);
2097 mutex_unlock(&c->pending_node_rewrites_lock);
2101 if (!bch2_write_ref_tryget(c, BCH_WRITE_REF_node_rewrite)) {
2102 if (test_bit(BCH_FS_started, &c->flags)) {
2103 bch_err(c, "%s: error getting c->writes ref", __func__);
2108 ret = bch2_fs_read_write_early(c);
2109 bch_err_msg(c, ret, "going read-write");
2115 bch2_write_ref_get(c, BCH_WRITE_REF_node_rewrite);
2118 queue_work(c->btree_interior_update_worker, &a->work);
2121 void bch2_do_pending_node_rewrites(struct bch_fs *c)
2123 struct async_btree_rewrite *a, *n;
2125 mutex_lock(&c->pending_node_rewrites_lock);
2126 list_for_each_entry_safe(a, n, &c->pending_node_rewrites, list) {
2129 bch2_write_ref_get(c, BCH_WRITE_REF_node_rewrite);
2130 queue_work(c->btree_interior_update_worker, &a->work);
2132 mutex_unlock(&c->pending_node_rewrites_lock);
2135 void bch2_free_pending_node_rewrites(struct bch_fs *c)
2137 struct async_btree_rewrite *a, *n;
2139 mutex_lock(&c->pending_node_rewrites_lock);
2140 list_for_each_entry_safe(a, n, &c->pending_node_rewrites, list) {
2145 mutex_unlock(&c->pending_node_rewrites_lock);
2148 static int __bch2_btree_node_update_key(struct btree_trans *trans,
2149 struct btree_iter *iter,
2150 struct btree *b, struct btree *new_hash,
2151 struct bkey_i *new_key,
2152 unsigned commit_flags,
2155 struct bch_fs *c = trans->c;
2156 struct btree_iter iter2 = { NULL };
2157 struct btree *parent;
2160 if (!skip_triggers) {
2161 ret = bch2_key_trigger_old(trans, b->c.btree_id, b->c.level + 1,
2162 bkey_i_to_s_c(&b->key),
2163 BTREE_TRIGGER_TRANSACTIONAL) ?:
2164 bch2_key_trigger_new(trans, b->c.btree_id, b->c.level + 1,
2165 bkey_i_to_s(new_key),
2166 BTREE_TRIGGER_TRANSACTIONAL);
2172 bkey_copy(&new_hash->key, new_key);
2173 ret = bch2_btree_node_hash_insert(&c->btree_cache,
2174 new_hash, b->c.level, b->c.btree_id);
2178 parent = btree_node_parent(btree_iter_path(trans, iter), b);
2180 bch2_trans_copy_iter(&iter2, iter);
2182 iter2.path = bch2_btree_path_make_mut(trans, iter2.path,
2183 iter2.flags & BTREE_ITER_INTENT,
2186 struct btree_path *path2 = btree_iter_path(trans, &iter2);
2187 BUG_ON(path2->level != b->c.level);
2188 BUG_ON(!bpos_eq(path2->pos, new_key->k.p));
2190 btree_path_set_level_up(trans, path2);
2192 trans->paths_sorted = false;
2194 ret = bch2_btree_iter_traverse(&iter2) ?:
2195 bch2_trans_update(trans, &iter2, new_key, BTREE_TRIGGER_NORUN);
2199 BUG_ON(btree_node_root(c, b) != b);
2201 struct jset_entry *e = bch2_trans_jset_entry_alloc(trans,
2202 jset_u64s(new_key->k.u64s));
2203 ret = PTR_ERR_OR_ZERO(e);
2207 journal_entry_set(e,
2208 BCH_JSET_ENTRY_btree_root,
2209 b->c.btree_id, b->c.level,
2210 new_key, new_key->k.u64s);
2213 ret = bch2_trans_commit(trans, NULL, NULL, commit_flags);
2217 bch2_btree_node_lock_write_nofail(trans, btree_iter_path(trans, iter), &b->c);
2220 mutex_lock(&c->btree_cache.lock);
2221 bch2_btree_node_hash_remove(&c->btree_cache, new_hash);
2222 bch2_btree_node_hash_remove(&c->btree_cache, b);
2224 bkey_copy(&b->key, new_key);
2225 ret = __bch2_btree_node_hash_insert(&c->btree_cache, b);
2227 mutex_unlock(&c->btree_cache.lock);
2229 bkey_copy(&b->key, new_key);
2232 bch2_btree_node_unlock_write(trans, btree_iter_path(trans, iter), b);
2234 bch2_trans_iter_exit(trans, &iter2);
2238 mutex_lock(&c->btree_cache.lock);
2239 bch2_btree_node_hash_remove(&c->btree_cache, b);
2240 mutex_unlock(&c->btree_cache.lock);
2245 int bch2_btree_node_update_key(struct btree_trans *trans, struct btree_iter *iter,
2246 struct btree *b, struct bkey_i *new_key,
2247 unsigned commit_flags, bool skip_triggers)
2249 struct bch_fs *c = trans->c;
2250 struct btree *new_hash = NULL;
2251 struct btree_path *path = btree_iter_path(trans, iter);
2255 ret = bch2_btree_path_upgrade(trans, path, b->c.level + 1);
2259 closure_init_stack(&cl);
2262 * check btree_ptr_hash_val() after @b is locked by
2263 * btree_iter_traverse():
2265 if (btree_ptr_hash_val(new_key) != b->hash_val) {
2266 ret = bch2_btree_cache_cannibalize_lock(trans, &cl);
2268 ret = drop_locks_do(trans, (closure_sync(&cl), 0));
2273 new_hash = bch2_btree_node_mem_alloc(trans, false);
2277 ret = __bch2_btree_node_update_key(trans, iter, b, new_hash, new_key,
2278 commit_flags, skip_triggers);
2282 mutex_lock(&c->btree_cache.lock);
2283 list_move(&new_hash->list, &c->btree_cache.freeable);
2284 mutex_unlock(&c->btree_cache.lock);
2286 six_unlock_write(&new_hash->c.lock);
2287 six_unlock_intent(&new_hash->c.lock);
2290 bch2_btree_cache_cannibalize_unlock(trans);
2294 int bch2_btree_node_update_key_get_iter(struct btree_trans *trans,
2295 struct btree *b, struct bkey_i *new_key,
2296 unsigned commit_flags, bool skip_triggers)
2298 struct btree_iter iter;
2301 bch2_trans_node_iter_init(trans, &iter, b->c.btree_id, b->key.k.p,
2302 BTREE_MAX_DEPTH, b->c.level,
2304 ret = bch2_btree_iter_traverse(&iter);
2308 /* has node been freed? */
2309 if (btree_iter_path(trans, &iter)->l[b->c.level].b != b) {
2310 /* node has been freed: */
2311 BUG_ON(!btree_node_dying(b));
2315 BUG_ON(!btree_node_hashed(b));
2317 struct bch_extent_ptr *ptr;
2318 bch2_bkey_drop_ptrs(bkey_i_to_s(new_key), ptr,
2319 !bch2_bkey_has_device(bkey_i_to_s(&b->key), ptr->dev));
2321 ret = bch2_btree_node_update_key(trans, &iter, b, new_key,
2322 commit_flags, skip_triggers);
2324 bch2_trans_iter_exit(trans, &iter);
2331 * Only for filesystem bringup, when first reading the btree roots or allocating
2332 * btree roots when initializing a new filesystem:
2334 void bch2_btree_set_root_for_read(struct bch_fs *c, struct btree *b)
2336 BUG_ON(btree_node_root(c, b));
2338 bch2_btree_set_root_inmem(c, b);
2341 static int __bch2_btree_root_alloc(struct btree_trans *trans, enum btree_id id)
2343 struct bch_fs *c = trans->c;
2348 closure_init_stack(&cl);
2351 ret = bch2_btree_cache_cannibalize_lock(trans, &cl);
2355 b = bch2_btree_node_mem_alloc(trans, false);
2356 bch2_btree_cache_cannibalize_unlock(trans);
2358 set_btree_node_fake(b);
2359 set_btree_node_need_rewrite(b);
2363 bkey_btree_ptr_init(&b->key);
2364 b->key.k.p = SPOS_MAX;
2365 *((u64 *) bkey_i_to_btree_ptr(&b->key)->v.start) = U64_MAX - id;
2367 bch2_bset_init_first(b, &b->data->keys);
2368 bch2_btree_build_aux_trees(b);
2371 btree_set_min(b, POS_MIN);
2372 btree_set_max(b, SPOS_MAX);
2373 b->data->format = bch2_btree_calc_format(b);
2374 btree_node_set_format(b, b->data->format);
2376 ret = bch2_btree_node_hash_insert(&c->btree_cache, b,
2377 b->c.level, b->c.btree_id);
2380 bch2_btree_set_root_inmem(c, b);
2382 six_unlock_write(&b->c.lock);
2383 six_unlock_intent(&b->c.lock);
2387 void bch2_btree_root_alloc(struct bch_fs *c, enum btree_id id)
2389 bch2_trans_run(c, __bch2_btree_root_alloc(trans, id));
2392 void bch2_btree_updates_to_text(struct printbuf *out, struct bch_fs *c)
2394 struct btree_update *as;
2396 mutex_lock(&c->btree_interior_update_lock);
2397 list_for_each_entry(as, &c->btree_interior_update_list, list)
2398 prt_printf(out, "%p m %u w %u r %u j %llu\n",
2402 closure_nr_remaining(&as->cl),
2404 mutex_unlock(&c->btree_interior_update_lock);
2407 static bool bch2_btree_interior_updates_pending(struct bch_fs *c)
2411 mutex_lock(&c->btree_interior_update_lock);
2412 ret = !list_empty(&c->btree_interior_update_list);
2413 mutex_unlock(&c->btree_interior_update_lock);
2418 bool bch2_btree_interior_updates_flush(struct bch_fs *c)
2420 bool ret = bch2_btree_interior_updates_pending(c);
2423 closure_wait_event(&c->btree_interior_update_wait,
2424 !bch2_btree_interior_updates_pending(c));
2428 void bch2_journal_entry_to_btree_root(struct bch_fs *c, struct jset_entry *entry)
2430 struct btree_root *r = bch2_btree_id_root(c, entry->btree_id);
2432 mutex_lock(&c->btree_root_lock);
2434 r->level = entry->level;
2436 bkey_copy(&r->key, (struct bkey_i *) entry->start);
2438 mutex_unlock(&c->btree_root_lock);
2442 bch2_btree_roots_to_journal_entries(struct bch_fs *c,
2443 struct jset_entry *end,
2448 mutex_lock(&c->btree_root_lock);
2450 for (i = 0; i < btree_id_nr_alive(c); i++) {
2451 struct btree_root *r = bch2_btree_id_root(c, i);
2453 if (r->alive && !test_bit(i, &skip)) {
2454 journal_entry_set(end, BCH_JSET_ENTRY_btree_root,
2455 i, r->level, &r->key, r->key.k.u64s);
2456 end = vstruct_next(end);
2460 mutex_unlock(&c->btree_root_lock);
2465 void bch2_fs_btree_interior_update_exit(struct bch_fs *c)
2467 if (c->btree_interior_update_worker)
2468 destroy_workqueue(c->btree_interior_update_worker);
2469 mempool_exit(&c->btree_interior_update_pool);
2472 void bch2_fs_btree_interior_update_init_early(struct bch_fs *c)
2474 mutex_init(&c->btree_reserve_cache_lock);
2475 INIT_LIST_HEAD(&c->btree_interior_update_list);
2476 INIT_LIST_HEAD(&c->btree_interior_updates_unwritten);
2477 mutex_init(&c->btree_interior_update_lock);
2478 INIT_WORK(&c->btree_interior_update_work, btree_interior_update_work);
2480 INIT_LIST_HEAD(&c->pending_node_rewrites);
2481 mutex_init(&c->pending_node_rewrites_lock);
2484 int bch2_fs_btree_interior_update_init(struct bch_fs *c)
2486 c->btree_interior_update_worker =
2487 alloc_workqueue("btree_update", WQ_UNBOUND|WQ_MEM_RECLAIM, 1);
2488 if (!c->btree_interior_update_worker)
2489 return -BCH_ERR_ENOMEM_btree_interior_update_worker_init;
2491 if (mempool_init_kmalloc_pool(&c->btree_interior_update_pool, 1,
2492 sizeof(struct btree_update)))
2493 return -BCH_ERR_ENOMEM_btree_interior_update_pool_init;