1 // SPDX-License-Identifier: GPL-2.0
4 #include "btree_update.h"
5 #include "btree_iter.h"
6 #include "btree_journal_iter.h"
7 #include "btree_locking.h"
17 static inline int btree_insert_entry_cmp(const struct btree_insert_entry *l,
18 const struct btree_insert_entry *r)
20 return cmp_int(l->btree_id, r->btree_id) ?:
21 cmp_int(l->cached, r->cached) ?:
22 -cmp_int(l->level, r->level) ?:
23 bpos_cmp(l->k->k.p, r->k->k.p);
26 static int __must_check
27 bch2_trans_update_by_path(struct btree_trans *, btree_path_idx_t,
28 struct bkey_i *, enum btree_update_flags,
31 static noinline int extent_front_merge(struct btree_trans *trans,
32 struct btree_iter *iter,
34 struct bkey_i **insert,
35 enum btree_update_flags flags)
37 struct bch_fs *c = trans->c;
38 struct bkey_i *update;
41 update = bch2_bkey_make_mut_noupdate(trans, k);
42 ret = PTR_ERR_OR_ZERO(update);
46 if (!bch2_bkey_merge(c, bkey_i_to_s(update), bkey_i_to_s_c(*insert)))
49 ret = bch2_key_has_snapshot_overwrites(trans, iter->btree_id, k.k->p) ?:
50 bch2_key_has_snapshot_overwrites(trans, iter->btree_id, (*insert)->k.p);
56 ret = bch2_btree_delete_at(trans, iter, flags);
64 static noinline int extent_back_merge(struct btree_trans *trans,
65 struct btree_iter *iter,
66 struct bkey_i *insert,
69 struct bch_fs *c = trans->c;
72 ret = bch2_key_has_snapshot_overwrites(trans, iter->btree_id, insert->k.p) ?:
73 bch2_key_has_snapshot_overwrites(trans, iter->btree_id, k.k->p);
79 bch2_bkey_merge(c, bkey_i_to_s(insert), k);
84 * When deleting, check if we need to emit a whiteout (because we're overwriting
85 * something in an ancestor snapshot)
87 static int need_whiteout_for_snapshot(struct btree_trans *trans,
88 enum btree_id btree_id, struct bpos pos)
90 struct btree_iter iter;
92 u32 snapshot = pos.snapshot;
95 if (!bch2_snapshot_parent(trans->c, pos.snapshot))
100 for_each_btree_key_norestart(trans, iter, btree_id, pos,
101 BTREE_ITER_ALL_SNAPSHOTS|
102 BTREE_ITER_NOPRESERVE, k, ret) {
103 if (!bkey_eq(k.k->p, pos))
106 if (bch2_snapshot_is_ancestor(trans->c, snapshot,
108 ret = !bkey_whiteout(k.k);
112 bch2_trans_iter_exit(trans, &iter);
117 int __bch2_insert_snapshot_whiteouts(struct btree_trans *trans,
122 struct bch_fs *c = trans->c;
123 struct btree_iter old_iter, new_iter = { NULL };
124 struct bkey_s_c old_k, new_k;
126 struct bkey_i *update;
129 if (!bch2_snapshot_has_children(c, old_pos.snapshot))
134 bch2_trans_iter_init(trans, &old_iter, id, old_pos,
135 BTREE_ITER_NOT_EXTENTS|
136 BTREE_ITER_ALL_SNAPSHOTS);
137 while ((old_k = bch2_btree_iter_prev(&old_iter)).k &&
138 !(ret = bkey_err(old_k)) &&
139 bkey_eq(old_pos, old_k.k->p)) {
140 struct bpos whiteout_pos =
141 SPOS(new_pos.inode, new_pos.offset, old_k.k->p.snapshot);;
143 if (!bch2_snapshot_is_ancestor(c, old_k.k->p.snapshot, old_pos.snapshot) ||
144 snapshot_list_has_ancestor(c, &s, old_k.k->p.snapshot))
147 new_k = bch2_bkey_get_iter(trans, &new_iter, id, whiteout_pos,
148 BTREE_ITER_NOT_EXTENTS|
150 ret = bkey_err(new_k);
154 if (new_k.k->type == KEY_TYPE_deleted) {
155 update = bch2_trans_kmalloc(trans, sizeof(struct bkey_i));
156 ret = PTR_ERR_OR_ZERO(update);
160 bkey_init(&update->k);
161 update->k.p = whiteout_pos;
162 update->k.type = KEY_TYPE_whiteout;
164 ret = bch2_trans_update(trans, &new_iter, update,
165 BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE);
167 bch2_trans_iter_exit(trans, &new_iter);
169 ret = snapshot_list_add(c, &s, old_k.k->p.snapshot);
173 bch2_trans_iter_exit(trans, &new_iter);
174 bch2_trans_iter_exit(trans, &old_iter);
180 int bch2_trans_update_extent_overwrite(struct btree_trans *trans,
181 struct btree_iter *iter,
182 enum btree_update_flags flags,
186 enum btree_id btree_id = iter->btree_id;
187 struct bkey_i *update;
188 struct bpos new_start = bkey_start_pos(new.k);
189 unsigned front_split = bkey_lt(bkey_start_pos(old.k), new_start);
190 unsigned back_split = bkey_gt(old.k->p, new.k->p);
191 unsigned middle_split = (front_split || back_split) &&
192 old.k->p.snapshot != new.k->p.snapshot;
193 unsigned nr_splits = front_split + back_split + middle_split;
194 int ret = 0, compressed_sectors;
197 * If we're going to be splitting a compressed extent, note it
198 * so that __bch2_trans_commit() can increase our disk
202 (compressed_sectors = bch2_bkey_sectors_compressed(old)))
203 trans->extra_disk_res += compressed_sectors * (nr_splits - 1);
206 update = bch2_bkey_make_mut_noupdate(trans, old);
207 if ((ret = PTR_ERR_OR_ZERO(update)))
210 bch2_cut_back(new_start, update);
212 ret = bch2_insert_snapshot_whiteouts(trans, btree_id,
213 old.k->p, update->k.p) ?:
214 bch2_btree_insert_nonextent(trans, btree_id, update,
215 BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|flags);
220 /* If we're overwriting in a different snapshot - middle split: */
222 update = bch2_bkey_make_mut_noupdate(trans, old);
223 if ((ret = PTR_ERR_OR_ZERO(update)))
226 bch2_cut_front(new_start, update);
227 bch2_cut_back(new.k->p, update);
229 ret = bch2_insert_snapshot_whiteouts(trans, btree_id,
230 old.k->p, update->k.p) ?:
231 bch2_btree_insert_nonextent(trans, btree_id, update,
232 BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|flags);
237 if (bkey_le(old.k->p, new.k->p)) {
238 update = bch2_trans_kmalloc(trans, sizeof(*update));
239 if ((ret = PTR_ERR_OR_ZERO(update)))
242 bkey_init(&update->k);
243 update->k.p = old.k->p;
244 update->k.p.snapshot = new.k->p.snapshot;
246 if (new.k->p.snapshot != old.k->p.snapshot) {
247 update->k.type = KEY_TYPE_whiteout;
248 } else if (btree_type_has_snapshots(btree_id)) {
249 ret = need_whiteout_for_snapshot(trans, btree_id, update->k.p);
253 update->k.type = KEY_TYPE_whiteout;
256 ret = bch2_btree_insert_nonextent(trans, btree_id, update,
257 BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|flags);
263 update = bch2_bkey_make_mut_noupdate(trans, old);
264 if ((ret = PTR_ERR_OR_ZERO(update)))
267 bch2_cut_front(new.k->p, update);
269 ret = bch2_trans_update_by_path(trans, iter->path, update,
270 BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|
279 static int bch2_trans_update_extent(struct btree_trans *trans,
280 struct btree_iter *orig_iter,
281 struct bkey_i *insert,
282 enum btree_update_flags flags)
284 struct btree_iter iter;
286 enum btree_id btree_id = orig_iter->btree_id;
289 bch2_trans_iter_init(trans, &iter, btree_id, bkey_start_pos(&insert->k),
291 BTREE_ITER_WITH_UPDATES|
292 BTREE_ITER_NOT_EXTENTS);
293 k = bch2_btree_iter_peek_upto(&iter, POS(insert->k.p.inode, U64_MAX));
294 if ((ret = bkey_err(k)))
299 if (bkey_eq(k.k->p, bkey_start_pos(&insert->k))) {
300 if (bch2_bkey_maybe_mergable(k.k, &insert->k)) {
301 ret = extent_front_merge(trans, &iter, k, &insert, flags);
309 while (bkey_gt(insert->k.p, bkey_start_pos(k.k))) {
310 bool done = bkey_lt(insert->k.p, k.k->p);
312 ret = bch2_trans_update_extent_overwrite(trans, &iter, flags, k, bkey_i_to_s_c(insert));
319 bch2_btree_iter_advance(&iter);
320 k = bch2_btree_iter_peek_upto(&iter, POS(insert->k.p.inode, U64_MAX));
321 if ((ret = bkey_err(k)))
327 if (bch2_bkey_maybe_mergable(&insert->k, k.k)) {
328 ret = extent_back_merge(trans, &iter, insert, k);
333 if (!bkey_deleted(&insert->k))
334 ret = bch2_btree_insert_nonextent(trans, btree_id, insert, flags);
336 bch2_trans_iter_exit(trans, &iter);
341 static noinline int flush_new_cached_update(struct btree_trans *trans,
342 struct btree_insert_entry *i,
343 enum btree_update_flags flags,
349 btree_path_idx_t path_idx =
350 bch2_path_get(trans, i->btree_id, i->old_k.p, 1, 0,
351 BTREE_ITER_INTENT, _THIS_IP_);
352 ret = bch2_btree_path_traverse(trans, path_idx, 0);
356 struct btree_path *btree_path = trans->paths + path_idx;
359 * The old key in the insert entry might actually refer to an existing
360 * key in the btree that has been deleted from cache and not yet
361 * flushed. Check for this and skip the flush so we don't run triggers
362 * against a stale key.
364 bch2_btree_path_peek_slot_exact(btree_path, &k);
365 if (!bkey_deleted(&k))
368 i->key_cache_already_flushed = true;
369 i->flags |= BTREE_TRIGGER_NORUN;
371 btree_path_set_should_be_locked(btree_path);
372 ret = bch2_trans_update_by_path(trans, path_idx, i->k, flags, ip);
374 bch2_path_put(trans, path_idx, true);
378 static int __must_check
379 bch2_trans_update_by_path(struct btree_trans *trans, btree_path_idx_t path_idx,
380 struct bkey_i *k, enum btree_update_flags flags,
383 struct bch_fs *c = trans->c;
384 struct btree_insert_entry *i, n;
387 struct btree_path *path = trans->paths + path_idx;
388 EBUG_ON(!path->should_be_locked);
389 EBUG_ON(trans->nr_updates >= trans->nr_paths);
390 EBUG_ON(!bpos_eq(k->k.p, path->pos));
392 n = (struct btree_insert_entry) {
394 .bkey_type = __btree_node_type(path->level, path->btree_id),
395 .btree_id = path->btree_id,
396 .level = path->level,
397 .cached = path->cached,
403 #ifdef CONFIG_BCACHEFS_DEBUG
404 trans_for_each_update(trans, i)
405 BUG_ON(i != trans->updates &&
406 btree_insert_entry_cmp(i - 1, i) >= 0);
410 * Pending updates are kept sorted: first, find position of new update,
411 * then delete/trim any updates the new update overwrites:
413 for (i = trans->updates; i < trans->updates + trans->nr_updates; i++) {
414 cmp = btree_insert_entry_cmp(&n, i);
419 if (!cmp && i < trans->updates + trans->nr_updates) {
420 EBUG_ON(i->insert_trigger_run || i->overwrite_trigger_run);
422 bch2_path_put(trans, i->path, true);
424 i->cached = n.cached;
427 i->ip_allocated = n.ip_allocated;
429 array_insert_item(trans->updates, trans->nr_updates,
430 i - trans->updates, n);
432 i->old_v = bch2_btree_path_peek_slot_exact(path, &i->old_k).v;
433 i->old_btree_u64s = !bkey_deleted(&i->old_k) ? i->old_k.u64s : 0;
435 if (unlikely(trans->journal_replay_not_finished)) {
437 bch2_journal_keys_peek_slot(c, n.btree_id, n.level, k->k.p);
446 __btree_path_get(trans->paths + i->path, true);
449 * If a key is present in the key cache, it must also exist in the
450 * btree - this is necessary for cache coherency. When iterating over
451 * a btree that's cached in the key cache, the btree iter code checks
452 * the key cache - but the key has to exist in the btree for that to
455 if (path->cached && bkey_deleted(&i->old_k))
456 return flush_new_cached_update(trans, i, flags, ip);
461 static noinline int bch2_trans_update_get_key_cache(struct btree_trans *trans,
462 struct btree_iter *iter,
463 struct btree_path *path)
465 struct btree_path *key_cache_path = btree_iter_key_cache_path(trans, iter);
467 if (!key_cache_path ||
468 !key_cache_path->should_be_locked ||
469 !bpos_eq(key_cache_path->pos, iter->pos)) {
470 struct bkey_cached *ck;
473 if (!iter->key_cache_path)
474 iter->key_cache_path =
475 bch2_path_get(trans, path->btree_id, path->pos, 1, 0,
477 BTREE_ITER_CACHED, _THIS_IP_);
479 iter->key_cache_path =
480 bch2_btree_path_set_pos(trans, iter->key_cache_path, path->pos,
481 iter->flags & BTREE_ITER_INTENT,
484 ret = bch2_btree_path_traverse(trans, iter->key_cache_path, BTREE_ITER_CACHED);
488 ck = (void *) trans->paths[iter->key_cache_path].l[0].b;
490 if (test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
491 trace_and_count(trans->c, trans_restart_key_cache_raced, trans, _RET_IP_);
492 return btree_trans_restart(trans, BCH_ERR_transaction_restart_key_cache_raced);
495 btree_path_set_should_be_locked(trans->paths + iter->key_cache_path);
501 int __must_check bch2_trans_update(struct btree_trans *trans, struct btree_iter *iter,
502 struct bkey_i *k, enum btree_update_flags flags)
504 btree_path_idx_t path_idx = iter->update_path ?: iter->path;
507 if (iter->flags & BTREE_ITER_IS_EXTENTS)
508 return bch2_trans_update_extent(trans, iter, k, flags);
510 if (bkey_deleted(&k->k) &&
511 !(flags & BTREE_UPDATE_KEY_CACHE_RECLAIM) &&
512 (iter->flags & BTREE_ITER_FILTER_SNAPSHOTS)) {
513 ret = need_whiteout_for_snapshot(trans, iter->btree_id, k->k.p);
514 if (unlikely(ret < 0))
518 k->k.type = KEY_TYPE_whiteout;
522 * Ensure that updates to cached btrees go to the key cache:
524 struct btree_path *path = trans->paths + path_idx;
525 if (!(flags & BTREE_UPDATE_KEY_CACHE_RECLAIM) &&
528 btree_id_cached(trans->c, path->btree_id)) {
529 ret = bch2_trans_update_get_key_cache(trans, iter, path);
533 path_idx = iter->key_cache_path;
536 return bch2_trans_update_by_path(trans, path_idx, k, flags, _RET_IP_);
539 int bch2_btree_insert_clone_trans(struct btree_trans *trans,
543 struct bkey_i *n = bch2_trans_kmalloc(trans, bkey_bytes(&k->k));
544 int ret = PTR_ERR_OR_ZERO(n);
549 return bch2_btree_insert_trans(trans, btree, n, 0);
552 struct jset_entry *__bch2_trans_jset_entry_alloc(struct btree_trans *trans, unsigned u64s)
554 unsigned new_top = trans->journal_entries_u64s + u64s;
555 unsigned old_size = trans->journal_entries_size;
557 if (new_top > trans->journal_entries_size) {
558 trans->journal_entries_size = roundup_pow_of_two(new_top);
560 btree_trans_stats(trans)->journal_entries_size = trans->journal_entries_size;
563 struct jset_entry *n =
564 bch2_trans_kmalloc_nomemzero(trans,
565 trans->journal_entries_size * sizeof(u64));
569 if (trans->journal_entries)
570 memcpy(n, trans->journal_entries, old_size * sizeof(u64));
571 trans->journal_entries = n;
573 struct jset_entry *e = btree_trans_journal_entries_top(trans);
574 trans->journal_entries_u64s = new_top;
578 int bch2_bkey_get_empty_slot(struct btree_trans *trans, struct btree_iter *iter,
579 enum btree_id btree, struct bpos end)
584 bch2_trans_iter_init(trans, iter, btree, POS_MAX, BTREE_ITER_INTENT);
585 k = bch2_btree_iter_prev(iter);
590 bch2_btree_iter_advance(iter);
591 k = bch2_btree_iter_peek_slot(iter);
596 BUG_ON(k.k->type != KEY_TYPE_deleted);
598 if (bkey_gt(k.k->p, end)) {
599 ret = -BCH_ERR_ENOSPC_btree_slot;
605 bch2_trans_iter_exit(trans, iter);
609 void bch2_trans_commit_hook(struct btree_trans *trans,
610 struct btree_trans_commit_hook *h)
612 h->next = trans->hooks;
616 int bch2_btree_insert_nonextent(struct btree_trans *trans,
617 enum btree_id btree, struct bkey_i *k,
618 enum btree_update_flags flags)
620 struct btree_iter iter;
623 bch2_trans_iter_init(trans, &iter, btree, k->k.p,
625 BTREE_ITER_NOT_EXTENTS|
627 ret = bch2_btree_iter_traverse(&iter) ?:
628 bch2_trans_update(trans, &iter, k, flags);
629 bch2_trans_iter_exit(trans, &iter);
633 int bch2_btree_insert_trans(struct btree_trans *trans, enum btree_id id,
634 struct bkey_i *k, enum btree_update_flags flags)
636 struct btree_iter iter;
639 bch2_trans_iter_init(trans, &iter, id, bkey_start_pos(&k->k),
642 ret = bch2_btree_iter_traverse(&iter) ?:
643 bch2_trans_update(trans, &iter, k, flags);
644 bch2_trans_iter_exit(trans, &iter);
649 * bch2_btree_insert - insert keys into the extent btree
650 * @c: pointer to struct bch_fs
651 * @id: btree to insert into
653 * @disk_res: must be non-NULL whenever inserting or potentially
654 * splitting data extents
655 * @flags: transaction commit flags
657 * Returns: 0 on success, error code on failure
659 int bch2_btree_insert(struct bch_fs *c, enum btree_id id, struct bkey_i *k,
660 struct disk_reservation *disk_res, int flags)
662 return bch2_trans_do(c, disk_res, NULL, flags,
663 bch2_btree_insert_trans(trans, id, k, 0));
666 int bch2_btree_delete_extent_at(struct btree_trans *trans, struct btree_iter *iter,
667 unsigned len, unsigned update_flags)
671 k = bch2_trans_kmalloc(trans, sizeof(*k));
677 bch2_key_resize(&k->k, len);
678 return bch2_trans_update(trans, iter, k, update_flags);
681 int bch2_btree_delete_at(struct btree_trans *trans,
682 struct btree_iter *iter, unsigned update_flags)
684 return bch2_btree_delete_extent_at(trans, iter, 0, update_flags);
687 int bch2_btree_delete(struct btree_trans *trans,
688 enum btree_id btree, struct bpos pos,
689 unsigned update_flags)
691 struct btree_iter iter;
694 bch2_trans_iter_init(trans, &iter, btree, pos,
697 ret = bch2_btree_iter_traverse(&iter) ?:
698 bch2_btree_delete_at(trans, &iter, update_flags);
699 bch2_trans_iter_exit(trans, &iter);
704 int bch2_btree_delete_range_trans(struct btree_trans *trans, enum btree_id id,
705 struct bpos start, struct bpos end,
706 unsigned update_flags,
709 u32 restart_count = trans->restart_count;
710 struct btree_iter iter;
714 bch2_trans_iter_init(trans, &iter, id, start, BTREE_ITER_INTENT);
715 while ((k = bch2_btree_iter_peek_upto(&iter, end)).k) {
716 struct disk_reservation disk_res =
717 bch2_disk_reservation_init(trans->c, 0);
718 struct bkey_i delete;
724 bkey_init(&delete.k);
727 * This could probably be more efficient for extents:
731 * For extents, iter.pos won't necessarily be the same as
732 * bkey_start_pos(k.k) (for non extents they always will be the
733 * same). It's important that we delete starting from iter.pos
734 * because the range we want to delete could start in the middle
737 * (bch2_btree_iter_peek() does guarantee that iter.pos >=
738 * bkey_start_pos(k.k)).
740 delete.k.p = iter.pos;
742 if (iter.flags & BTREE_ITER_IS_EXTENTS)
743 bch2_key_resize(&delete.k,
744 bpos_min(end, k.k->p).offset -
747 ret = bch2_trans_update(trans, &iter, &delete, update_flags) ?:
748 bch2_trans_commit(trans, &disk_res, journal_seq,
749 BCH_TRANS_COMMIT_no_enospc);
750 bch2_disk_reservation_put(trans->c, &disk_res);
753 * the bch2_trans_begin() call is in a weird place because we
754 * need to call it after every transaction commit, to avoid path
755 * overflow, but don't want to call it if the delete operation
756 * is a no-op and we have no work to do:
758 bch2_trans_begin(trans);
760 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
765 bch2_trans_iter_exit(trans, &iter);
767 return ret ?: trans_was_restarted(trans, restart_count);
771 * bch_btree_delete_range - delete everything within a given range
773 * Range is a half open interval - [start, end)
775 int bch2_btree_delete_range(struct bch_fs *c, enum btree_id id,
776 struct bpos start, struct bpos end,
777 unsigned update_flags,
780 int ret = bch2_trans_run(c,
781 bch2_btree_delete_range_trans(trans, id, start, end,
782 update_flags, journal_seq));
783 if (ret == -BCH_ERR_transaction_restart_nested)
788 int bch2_btree_bit_mod(struct btree_trans *trans, enum btree_id btree,
789 struct bpos pos, bool set)
794 k.k.type = set ? KEY_TYPE_set : KEY_TYPE_deleted;
797 return bch2_trans_update_buffered(trans, btree, &k);
800 static int __bch2_trans_log_msg(struct btree_trans *trans, struct printbuf *buf, unsigned u64s)
802 struct jset_entry *e = bch2_trans_jset_entry_alloc(trans, jset_u64s(u64s));
803 int ret = PTR_ERR_OR_ZERO(e);
807 struct jset_entry_log *l = container_of(e, struct jset_entry_log, entry);
808 journal_entry_init(e, BCH_JSET_ENTRY_log, 0, 1, u64s);
809 memcpy(l->d, buf->buf, buf->pos);
815 __bch2_fs_log_msg(struct bch_fs *c, unsigned commit_flags, const char *fmt,
818 struct printbuf buf = PRINTBUF;
819 prt_vprintf(&buf, fmt, args);
821 unsigned u64s = DIV_ROUND_UP(buf.pos, sizeof(u64));
822 prt_chars(&buf, '\0', u64s * sizeof(u64) - buf.pos);
824 int ret = buf.allocation_failure ? -BCH_ERR_ENOMEM_trans_log_msg : 0;
828 if (!test_bit(JOURNAL_STARTED, &c->journal.flags)) {
829 ret = darray_make_room(&c->journal.early_journal_entries, jset_u64s(u64s));
833 struct jset_entry_log *l = (void *) &darray_top(c->journal.early_journal_entries);
834 journal_entry_init(&l->entry, BCH_JSET_ENTRY_log, 0, 1, u64s);
835 memcpy(l->d, buf.buf, buf.pos);
836 c->journal.early_journal_entries.nr += jset_u64s(u64s);
838 ret = bch2_trans_do(c, NULL, NULL,
839 BCH_TRANS_COMMIT_lazy_rw|commit_flags,
840 __bch2_trans_log_msg(trans, &buf, u64s));
848 int bch2_fs_log_msg(struct bch_fs *c, const char *fmt, ...)
854 ret = __bch2_fs_log_msg(c, 0, fmt, args);
860 * Use for logging messages during recovery to enable reserved space and avoid
864 int bch2_journal_log_msg(struct bch_fs *c, const char *fmt, ...)
870 ret = __bch2_fs_log_msg(c, BCH_WATERMARK_reclaim, fmt, args);