1 // SPDX-License-Identifier: GPL-2.0
4 #include "btree_update.h"
5 #include "btree_update_interior.h"
8 #include "btree_iter.h"
9 #include "btree_key_cache.h"
10 #include "btree_locking.h"
14 #include "extent_update.h"
16 #include "journal_reclaim.h"
20 #include <linux/prefetch.h>
21 #include <linux/sort.h>
22 #include <trace/events/bcachefs.h>
24 static inline bool same_leaf_as_prev(struct btree_trans *trans,
25 struct btree_insert_entry *i)
27 return i != trans->updates2 &&
28 iter_l(i[0].iter)->b == iter_l(i[-1].iter)->b;
31 inline void bch2_btree_node_lock_for_insert(struct bch_fs *c, struct btree *b,
32 struct btree_iter *iter)
34 bch2_btree_node_lock_write(b, iter);
36 if (btree_iter_type(iter) == BTREE_ITER_CACHED)
39 if (unlikely(btree_node_just_written(b)) &&
40 bch2_btree_post_write_cleanup(c, b))
41 bch2_btree_iter_reinit_node(iter, b);
44 * If the last bset has been written, or if it's gotten too big - start
45 * a new bset to insert into:
47 if (want_new_bset(c, b))
48 bch2_btree_init_next(c, b, iter);
51 /* Inserting into a given leaf node (last stage of insert): */
53 /* Handle overwrites and do insert, for non extents: */
54 bool bch2_btree_bset_insert_key(struct btree_iter *iter,
56 struct btree_node_iter *node_iter,
57 struct bkey_i *insert)
59 struct bkey_packed *k;
60 unsigned clobber_u64s = 0, new_u64s = 0;
62 EBUG_ON(btree_node_just_written(b));
63 EBUG_ON(bset_written(b, btree_bset_last(b)));
64 EBUG_ON(bkey_deleted(&insert->k) && bkey_val_u64s(&insert->k));
65 EBUG_ON(bkey_cmp(b->data->min_key, POS_MIN) &&
66 bkey_cmp(bkey_start_pos(&insert->k),
67 bkey_predecessor(b->data->min_key)) < 0);
68 EBUG_ON(bkey_cmp(insert->k.p, b->data->min_key) < 0);
69 EBUG_ON(bkey_cmp(insert->k.p, b->data->max_key) > 0);
70 EBUG_ON(insert->k.u64s >
71 bch_btree_keys_u64s_remaining(iter->trans->c, b));
72 EBUG_ON(iter->flags & BTREE_ITER_IS_EXTENTS);
74 k = bch2_btree_node_iter_peek_all(node_iter, b);
75 if (k && bkey_cmp_packed(b, k, &insert->k))
78 /* @k is the key being overwritten/deleted, if any: */
79 EBUG_ON(k && bkey_whiteout(k));
81 /* Deleting, but not found? nothing to do: */
82 if (bkey_whiteout(&insert->k) && !k)
85 if (bkey_whiteout(&insert->k)) {
87 btree_account_key_drop(b, k);
88 k->type = KEY_TYPE_deleted;
90 if (k->needs_whiteout)
91 push_whiteout(iter->trans->c, b, insert->k.p);
92 k->needs_whiteout = false;
94 if (k >= btree_bset_last(b)->start) {
95 clobber_u64s = k->u64s;
96 bch2_bset_delete(b, k, clobber_u64s);
99 bch2_btree_iter_fix_key_modified(iter, b, k);
107 btree_account_key_drop(b, k);
108 k->type = KEY_TYPE_deleted;
110 insert->k.needs_whiteout = k->needs_whiteout;
111 k->needs_whiteout = false;
113 if (k >= btree_bset_last(b)->start) {
114 clobber_u64s = k->u64s;
117 bch2_btree_iter_fix_key_modified(iter, b, k);
121 k = bch2_btree_node_iter_bset_pos(node_iter, b, bset_tree_last(b));
123 bch2_bset_insert(b, node_iter, k, insert, clobber_u64s);
126 if (clobber_u64s != new_u64s)
127 bch2_btree_node_iter_fix(iter, b, node_iter, k,
128 clobber_u64s, new_u64s);
132 static void __btree_node_flush(struct journal *j, struct journal_entry_pin *pin,
135 struct bch_fs *c = container_of(j, struct bch_fs, journal);
136 struct btree_write *w = container_of(pin, struct btree_write, journal);
137 struct btree *b = container_of(w, struct btree, writes[i]);
139 btree_node_lock_type(c, b, SIX_LOCK_read);
140 bch2_btree_node_write_cond(c, b,
141 (btree_current_write(b) == w && w->journal.seq == seq));
142 six_unlock_read(&b->c.lock);
145 static void btree_node_flush0(struct journal *j, struct journal_entry_pin *pin, u64 seq)
147 return __btree_node_flush(j, pin, 0, seq);
150 static void btree_node_flush1(struct journal *j, struct journal_entry_pin *pin, u64 seq)
152 return __btree_node_flush(j, pin, 1, seq);
155 inline void bch2_btree_add_journal_pin(struct bch_fs *c,
156 struct btree *b, u64 seq)
158 struct btree_write *w = btree_current_write(b);
160 bch2_journal_pin_add(&c->journal, seq, &w->journal,
161 btree_node_write_idx(b) == 0
163 : btree_node_flush1);
167 * btree_insert_key - insert a key one key into a leaf node
169 static bool btree_insert_key_leaf(struct btree_trans *trans,
170 struct btree_iter *iter,
171 struct bkey_i *insert)
173 struct bch_fs *c = trans->c;
174 struct btree *b = iter_l(iter)->b;
175 struct bset_tree *t = bset_tree_last(b);
176 struct bset *i = bset(b, t);
177 int old_u64s = bset_u64s(t);
178 int old_live_u64s = b->nr.live_u64s;
179 int live_u64s_added, u64s_added;
181 EBUG_ON(!iter->level &&
182 !test_bit(BCH_FS_BTREE_INTERIOR_REPLAY_DONE, &c->flags));
184 if (unlikely(!bch2_btree_bset_insert_key(iter, b,
185 &iter_l(iter)->iter, insert)))
188 i->journal_seq = cpu_to_le64(max(trans->journal_res.seq,
189 le64_to_cpu(i->journal_seq)));
191 bch2_btree_add_journal_pin(c, b, trans->journal_res.seq);
193 if (unlikely(!btree_node_dirty(b)))
194 set_btree_node_dirty(b);
196 live_u64s_added = (int) b->nr.live_u64s - old_live_u64s;
197 u64s_added = (int) bset_u64s(t) - old_u64s;
199 if (b->sib_u64s[0] != U16_MAX && live_u64s_added < 0)
200 b->sib_u64s[0] = max(0, (int) b->sib_u64s[0] + live_u64s_added);
201 if (b->sib_u64s[1] != U16_MAX && live_u64s_added < 0)
202 b->sib_u64s[1] = max(0, (int) b->sib_u64s[1] + live_u64s_added);
204 if (u64s_added > live_u64s_added &&
205 bch2_maybe_compact_whiteouts(c, b))
206 bch2_btree_iter_reinit_node(iter, b);
208 trace_btree_insert_key(c, b, insert);
212 /* Cached btree updates: */
214 /* Normal update interface: */
216 static inline void btree_insert_entry_checks(struct btree_trans *trans,
217 struct btree_iter *iter,
218 struct bkey_i *insert)
220 struct bch_fs *c = trans->c;
222 BUG_ON(bkey_cmp(insert->k.p, iter->pos));
223 BUG_ON(debug_check_bkeys(c) &&
224 bch2_bkey_invalid(c, bkey_i_to_s_c(insert),
225 __btree_node_type(iter->level, iter->btree_id)));
229 bch2_trans_journal_preres_get_cold(struct btree_trans *trans, unsigned u64s)
231 struct bch_fs *c = trans->c;
234 bch2_trans_unlock(trans);
236 ret = bch2_journal_preres_get(&c->journal,
237 &trans->journal_preres, u64s, 0);
241 if (!bch2_trans_relock(trans)) {
242 trace_trans_restart_journal_preres_get(trans->ip);
249 static inline int bch2_trans_journal_res_get(struct btree_trans *trans,
252 struct bch_fs *c = trans->c;
255 if (trans->flags & BTREE_INSERT_JOURNAL_RESERVED)
256 flags |= JOURNAL_RES_GET_RESERVED;
258 ret = bch2_journal_res_get(&c->journal, &trans->journal_res,
259 trans->journal_u64s, flags);
261 return ret == -EAGAIN ? BTREE_INSERT_NEED_JOURNAL_RES : ret;
264 static enum btree_insert_ret
265 btree_key_can_insert(struct btree_trans *trans,
266 struct btree_iter *iter,
269 struct bch_fs *c = trans->c;
270 struct btree *b = iter_l(iter)->b;
272 if (!bch2_btree_node_insert_fits(c, b, u64s))
273 return BTREE_INSERT_BTREE_NODE_FULL;
275 return BTREE_INSERT_OK;
278 static enum btree_insert_ret
279 btree_key_can_insert_cached(struct btree_trans *trans,
280 struct btree_iter *iter,
283 struct bkey_cached *ck = (void *) iter->l[0].b;
285 struct bkey_i *new_k;
289 if (u64s <= ck->u64s)
290 return BTREE_INSERT_OK;
292 new_u64s = roundup_pow_of_two(u64s);
293 new_k = krealloc(ck->k, new_u64s * sizeof(u64), GFP_NOFS);
299 return BTREE_INSERT_OK;
302 static inline void do_btree_insert_one(struct btree_trans *trans,
303 struct btree_iter *iter,
304 struct bkey_i *insert)
306 struct bch_fs *c = trans->c;
307 struct journal *j = &c->journal;
310 EBUG_ON(trans->journal_res.ref !=
311 !(trans->flags & BTREE_INSERT_JOURNAL_REPLAY));
313 insert->k.needs_whiteout = false;
315 did_work = (btree_iter_type(iter) != BTREE_ITER_CACHED)
316 ? btree_insert_key_leaf(trans, iter, insert)
317 : bch2_btree_insert_key_cached(trans, iter, insert);
321 if (likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY))) {
322 bch2_journal_add_keys(j, &trans->journal_res,
323 iter->btree_id, insert);
325 bch2_journal_set_has_inode(j, &trans->journal_res,
328 if (trans->journal_seq)
329 *trans->journal_seq = trans->journal_res.seq;
333 static inline bool iter_has_trans_triggers(struct btree_iter *iter)
335 return BTREE_NODE_TYPE_HAS_TRANS_TRIGGERS & (1U << iter->btree_id);
338 static inline bool iter_has_nontrans_triggers(struct btree_iter *iter)
340 return (((BTREE_NODE_TYPE_HAS_TRIGGERS &
341 ~BTREE_NODE_TYPE_HAS_TRANS_TRIGGERS)) |
342 (1U << BTREE_ID_EC)) &
343 (1U << iter->btree_id);
346 static noinline void bch2_btree_iter_unlock_noinline(struct btree_iter *iter)
348 __bch2_btree_iter_unlock(iter);
351 static noinline void bch2_trans_mark_gc(struct btree_trans *trans)
353 struct bch_fs *c = trans->c;
354 struct btree_insert_entry *i;
356 trans_for_each_update(trans, i) {
358 * XXX: synchronization of cached update triggers with gc
360 BUG_ON(btree_iter_type(i->iter) == BTREE_ITER_CACHED);
362 if (gc_visited(c, gc_pos_btree_node(i->iter->l[0].b)))
363 bch2_mark_update(trans, i->iter, i->k, NULL,
364 i->trigger_flags|BTREE_TRIGGER_GC);
369 bch2_trans_commit_write_locked(struct btree_trans *trans,
370 struct btree_insert_entry **stopped_at)
372 struct bch_fs *c = trans->c;
373 struct bch_fs_usage *fs_usage = NULL;
374 struct btree_insert_entry *i;
376 bool marking = false;
380 trace_trans_restart_fault_inject(trans->ip);
385 * Check if the insert will fit in the leaf node with the write lock
386 * held, otherwise another thread could write the node changing the
387 * amount of space available:
390 prefetch(&trans->c->journal.flags);
392 trans_for_each_update2(trans, i) {
393 /* Multiple inserts might go to same leaf: */
394 if (!same_leaf_as_prev(trans, i))
397 u64s += i->k->k.u64s;
398 ret = btree_iter_type(i->iter) != BTREE_ITER_CACHED
399 ? btree_key_can_insert(trans, i->iter, u64s)
400 : btree_key_can_insert_cached(trans, i->iter, u64s);
406 if (btree_node_type_needs_gc(i->iter->btree_id))
411 percpu_down_read(&c->mark_lock);
412 fs_usage = bch2_fs_usage_scratch_get(c);
416 * Don't get journal reservation until after we know insert will
419 if (likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY))) {
420 ret = bch2_trans_journal_res_get(trans,
421 JOURNAL_RES_GET_NONBLOCK);
425 trans->journal_res.seq = c->journal.replay_journal_seq;
428 if (unlikely(trans->extra_journal_entry_u64s)) {
429 memcpy_u64s_small(journal_res_entry(&c->journal, &trans->journal_res),
430 trans->extra_journal_entries,
431 trans->extra_journal_entry_u64s);
433 trans->journal_res.offset += trans->extra_journal_entry_u64s;
434 trans->journal_res.u64s -= trans->extra_journal_entry_u64s;
438 * Not allowed to fail after we've gotten our journal reservation - we
442 if (!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY)) {
443 if (journal_seq_verify(c))
444 trans_for_each_update2(trans, i)
445 i->k->k.version.lo = trans->journal_res.seq;
446 else if (inject_invalid_keys(c))
447 trans_for_each_update2(trans, i)
448 i->k->k.version = MAX_VERSION;
451 /* Must be called under mark_lock: */
452 if (marking && trans->fs_usage_deltas &&
453 bch2_replicas_delta_list_apply(c, fs_usage,
454 trans->fs_usage_deltas)) {
455 ret = BTREE_INSERT_NEED_MARK_REPLICAS;
459 trans_for_each_update(trans, i)
460 if (iter_has_nontrans_triggers(i->iter))
461 bch2_mark_update(trans, i->iter, i->k,
462 fs_usage, i->trigger_flags);
465 bch2_trans_fs_usage_apply(trans, fs_usage);
467 if (unlikely(c->gc_pos.phase))
468 bch2_trans_mark_gc(trans);
470 trans_for_each_update2(trans, i)
471 do_btree_insert_one(trans, i->iter, i->k);
474 bch2_fs_usage_scratch_put(c, fs_usage);
475 percpu_up_read(&c->mark_lock);
482 * Get journal reservation, take write locks, and attempt to do btree update(s):
484 static inline int do_bch2_trans_commit(struct btree_trans *trans,
485 struct btree_insert_entry **stopped_at)
487 struct btree_insert_entry *i;
488 struct btree_iter *iter;
491 trans_for_each_update2(trans, i)
492 BUG_ON(!btree_node_intent_locked(i->iter, i->iter->level));
494 ret = bch2_journal_preres_get(&trans->c->journal,
495 &trans->journal_preres, trans->journal_preres_u64s,
496 JOURNAL_RES_GET_NONBLOCK|
497 ((trans->flags & BTREE_INSERT_JOURNAL_RECLAIM)
498 ? JOURNAL_RES_GET_RECLAIM : 0));
499 if (unlikely(ret == -EAGAIN))
500 ret = bch2_trans_journal_preres_get_cold(trans,
501 trans->journal_preres_u64s);
506 * Can't be holding any read locks when we go to take write locks:
508 * note - this must be done after bch2_trans_journal_preres_get_cold()
509 * or anything else that might call bch2_trans_relock(), since that
510 * would just retake the read locks:
512 trans_for_each_iter(trans, iter) {
513 if (iter->nodes_locked != iter->nodes_intent_locked) {
514 EBUG_ON(iter->flags & BTREE_ITER_KEEP_UNTIL_COMMIT);
515 EBUG_ON(trans->iters_live & (1ULL << iter->idx));
516 bch2_btree_iter_unlock_noinline(iter);
520 if (IS_ENABLED(CONFIG_BCACHEFS_DEBUG))
521 trans_for_each_update2(trans, i)
522 btree_insert_entry_checks(trans, i->iter, i->k);
523 bch2_btree_trans_verify_locks(trans);
525 trans_for_each_update2(trans, i)
526 if (!same_leaf_as_prev(trans, i))
527 bch2_btree_node_lock_for_insert(trans->c,
528 iter_l(i->iter)->b, i->iter);
530 ret = bch2_trans_commit_write_locked(trans, stopped_at);
532 trans_for_each_update2(trans, i)
533 if (!same_leaf_as_prev(trans, i))
534 bch2_btree_node_unlock_write_inlined(iter_l(i->iter)->b,
537 if (!ret && trans->journal_pin)
538 bch2_journal_pin_add(&trans->c->journal, trans->journal_res.seq,
539 trans->journal_pin, NULL);
542 * Drop journal reservation after dropping write locks, since dropping
543 * the journal reservation may kick off a journal write:
545 bch2_journal_res_put(&trans->c->journal, &trans->journal_res);
550 if (trans->flags & BTREE_INSERT_NOUNLOCK)
551 trans->nounlock = true;
553 trans_for_each_update2(trans, i)
554 if (btree_iter_type(i->iter) != BTREE_ITER_CACHED &&
555 !same_leaf_as_prev(trans, i))
556 bch2_foreground_maybe_merge(trans->c, i->iter,
559 trans->nounlock = false;
561 bch2_trans_downgrade(trans);
567 int bch2_trans_commit_error(struct btree_trans *trans,
568 struct btree_insert_entry *i,
571 struct bch_fs *c = trans->c;
572 unsigned flags = trans->flags;
575 * BTREE_INSERT_NOUNLOCK means don't unlock _after_ successful btree
576 * update; if we haven't done anything yet it doesn't apply
578 flags &= ~BTREE_INSERT_NOUNLOCK;
581 case BTREE_INSERT_BTREE_NODE_FULL:
582 ret = bch2_btree_split_leaf(c, i->iter, flags);
585 * if the split succeeded without dropping locks the insert will
586 * still be atomic (what the caller peeked() and is overwriting
587 * won't have changed)
592 * split -> btree node merging (of parent node) might still drop
593 * locks when we're not passing it BTREE_INSERT_NOUNLOCK
595 * we don't want to pass BTREE_INSERT_NOUNLOCK to split as that
596 * will inhibit merging - but we don't have a reliable way yet
597 * (do we?) of checking if we dropped locks in this path
604 * don't care if we got ENOSPC because we told split it
609 (flags & BTREE_INSERT_NOUNLOCK)) {
610 trace_trans_restart_btree_node_split(trans->ip);
614 case BTREE_INSERT_ENOSPC:
617 case BTREE_INSERT_NEED_MARK_REPLICAS:
618 bch2_trans_unlock(trans);
620 trans_for_each_update(trans, i) {
621 ret = bch2_mark_bkey_replicas(c, bkey_i_to_s_c(i->k));
626 if (bch2_trans_relock(trans))
629 trace_trans_restart_mark_replicas(trans->ip);
632 case BTREE_INSERT_NEED_JOURNAL_RES:
633 bch2_trans_unlock(trans);
635 ret = bch2_trans_journal_res_get(trans, JOURNAL_RES_GET_CHECK);
639 if (bch2_trans_relock(trans))
642 trace_trans_restart_journal_res_get(trans->ip);
651 int ret2 = bch2_btree_iter_traverse_all(trans);
654 trace_trans_restart_traverse(trans->ip);
658 trace_trans_restart_atomic(trans->ip);
665 bch2_trans_commit_get_rw_cold(struct btree_trans *trans)
667 struct bch_fs *c = trans->c;
670 if (likely(!(trans->flags & BTREE_INSERT_LAZY_RW)))
673 bch2_trans_unlock(trans);
675 ret = bch2_fs_read_write_early(c);
679 percpu_ref_get(&c->writes);
683 static void bch2_trans_update2(struct btree_trans *trans,
684 struct btree_iter *iter,
685 struct bkey_i *insert)
687 struct btree_insert_entry *i, n = (struct btree_insert_entry) {
688 .iter = iter, .k = insert
691 btree_insert_entry_checks(trans, n.iter, n.k);
693 BUG_ON(iter->uptodate > BTREE_ITER_NEED_PEEK);
695 EBUG_ON(trans->nr_updates2 >= trans->nr_iters);
697 iter->flags |= BTREE_ITER_KEEP_UNTIL_COMMIT;
699 trans_for_each_update2(trans, i) {
700 if (btree_iter_cmp(n.iter, i->iter) == 0) {
705 if (btree_iter_cmp(n.iter, i->iter) <= 0)
709 array_insert_item(trans->updates2, trans->nr_updates2,
710 i - trans->updates2, n);
713 static int extent_update_to_keys(struct btree_trans *trans,
714 struct btree_iter *orig_iter,
715 struct bkey_i *insert)
717 struct btree_iter *iter;
720 ret = bch2_extent_can_insert(trans, orig_iter, insert);
724 if (bkey_deleted(&insert->k))
727 iter = bch2_trans_copy_iter(trans, orig_iter);
729 return PTR_ERR(iter);
731 iter->flags |= BTREE_ITER_INTENT;
732 __bch2_btree_iter_set_pos(iter, insert->k.p, false);
733 bch2_trans_update2(trans, iter, insert);
734 bch2_trans_iter_put(trans, iter);
738 static int extent_handle_overwrites(struct btree_trans *trans,
739 enum btree_id btree_id,
740 struct bpos start, struct bpos end)
742 struct btree_iter *iter = NULL, *update_iter;
743 struct bkey_i *update;
747 iter = bch2_trans_get_iter(trans, btree_id, start, BTREE_ITER_INTENT);
748 ret = PTR_ERR_OR_ZERO(iter);
752 k = bch2_btree_iter_peek_with_updates(iter);
754 while (k.k && !(ret = bkey_err(k))) {
755 if (bkey_cmp(end, bkey_start_pos(k.k)) <= 0)
758 if (bkey_cmp(bkey_start_pos(k.k), start) < 0) {
759 update_iter = bch2_trans_copy_iter(trans, iter);
760 if ((ret = PTR_ERR_OR_ZERO(update_iter)))
763 update = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
764 if ((ret = PTR_ERR_OR_ZERO(update)))
767 bkey_reassemble(update, k);
768 bch2_cut_back(start, update);
770 __bch2_btree_iter_set_pos(update_iter, update->k.p, false);
771 bch2_trans_update2(trans, update_iter, update);
772 bch2_trans_iter_put(trans, update_iter);
775 if (bkey_cmp(k.k->p, end) > 0) {
776 update_iter = bch2_trans_copy_iter(trans, iter);
777 if ((ret = PTR_ERR_OR_ZERO(update_iter)))
780 update = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
781 if ((ret = PTR_ERR_OR_ZERO(update)))
784 bkey_reassemble(update, k);
785 bch2_cut_front(end, update);
787 __bch2_btree_iter_set_pos(update_iter, update->k.p, false);
788 bch2_trans_update2(trans, update_iter, update);
789 bch2_trans_iter_put(trans, update_iter);
791 update_iter = bch2_trans_copy_iter(trans, iter);
792 if ((ret = PTR_ERR_OR_ZERO(update_iter)))
795 update = bch2_trans_kmalloc(trans, sizeof(struct bkey));
796 if ((ret = PTR_ERR_OR_ZERO(update)))
800 set_bkey_val_u64s(&update->k, 0);
801 update->k.type = KEY_TYPE_deleted;
804 __bch2_btree_iter_set_pos(update_iter, update->k.p, false);
805 bch2_trans_update2(trans, update_iter, update);
806 bch2_trans_iter_put(trans, update_iter);
809 k = bch2_btree_iter_next_with_updates(iter);
812 if (!IS_ERR_OR_NULL(iter))
813 bch2_trans_iter_put(trans, iter);
817 int __bch2_trans_commit(struct btree_trans *trans)
819 struct btree_insert_entry *i = NULL;
820 struct btree_iter *iter;
821 bool trans_trigger_run;
825 BUG_ON(trans->need_reset);
827 if (!trans->nr_updates)
830 if (trans->flags & BTREE_INSERT_GC_LOCK_HELD)
831 lockdep_assert_held(&trans->c->gc_lock);
833 memset(&trans->journal_preres, 0, sizeof(trans->journal_preres));
835 trans->journal_u64s = trans->extra_journal_entry_u64s;
836 trans->journal_preres_u64s = 0;
838 if (!(trans->flags & BTREE_INSERT_NOCHECK_RW) &&
839 unlikely(!percpu_ref_tryget(&trans->c->writes))) {
840 ret = bch2_trans_commit_get_rw_cold(trans);
845 #ifdef CONFIG_BCACHEFS_DEBUG
846 trans_for_each_update(trans, i)
847 if (btree_iter_type(i->iter) != BTREE_ITER_CACHED &&
848 !(i->trigger_flags & BTREE_TRIGGER_NORUN))
849 bch2_btree_key_cache_verify_clean(trans,
850 i->iter->btree_id, i->iter->pos);
854 * Running triggers will append more updates to the list of updates as
858 trans_trigger_run = false;
860 trans_for_each_update(trans, i) {
861 if (unlikely(i->iter->uptodate > BTREE_ITER_NEED_PEEK &&
862 (ret = bch2_btree_iter_traverse(i->iter)))) {
863 trace_trans_restart_traverse(trans->ip);
868 * We're not using bch2_btree_iter_upgrade here because
869 * we know trans->nounlock can't be set:
871 if (unlikely(i->iter->locks_want < 1 &&
872 !__bch2_btree_iter_upgrade(i->iter, 1))) {
873 trace_trans_restart_upgrade(trans->ip);
878 if (iter_has_trans_triggers(i->iter) &&
879 !i->trans_triggers_run) {
880 i->trans_triggers_run = true;
881 trans_trigger_run = true;
883 ret = bch2_trans_mark_update(trans, i->iter, i->k,
887 trace_trans_restart_mark(trans->ip);
892 } while (trans_trigger_run);
894 /* Turn extents updates into keys: */
895 trans_for_each_update(trans, i)
896 if (i->iter->flags & BTREE_ITER_IS_EXTENTS) {
897 struct bpos start = bkey_start_pos(&i->k->k);
899 while (i + 1 < trans->updates + trans->nr_updates &&
900 i[0].iter->btree_id == i[1].iter->btree_id &&
901 !bkey_cmp(i[0].k->k.p, bkey_start_pos(&i[1].k->k)))
904 ret = extent_handle_overwrites(trans, i->iter->btree_id,
910 trans_for_each_update(trans, i) {
911 if (i->iter->flags & BTREE_ITER_IS_EXTENTS) {
912 ret = extent_update_to_keys(trans, i->iter, i->k);
916 bch2_trans_update2(trans, i->iter, i->k);
920 trans_for_each_update2(trans, i) {
921 BUG_ON(i->iter->uptodate > BTREE_ITER_NEED_PEEK);
922 BUG_ON(i->iter->locks_want < 1);
924 u64s = jset_u64s(i->k->k.u64s);
925 if (btree_iter_type(i->iter) == BTREE_ITER_CACHED &&
926 likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY)))
927 trans->journal_preres_u64s += u64s;
928 trans->journal_u64s += u64s;
931 memset(&trans->journal_res, 0, sizeof(trans->journal_res));
933 ret = do_bch2_trans_commit(trans, &i);
935 /* make sure we didn't drop or screw up locks: */
936 bch2_btree_trans_verify_locks(trans);
941 trans_for_each_iter(trans, iter)
942 if ((trans->iters_live & (1ULL << iter->idx)) &&
943 (iter->flags & BTREE_ITER_SET_POS_AFTER_COMMIT)) {
944 if (trans->flags & BTREE_INSERT_NOUNLOCK)
945 bch2_btree_iter_set_pos_same_leaf(iter, iter->pos_after_commit);
947 bch2_btree_iter_set_pos(iter, iter->pos_after_commit);
950 bch2_journal_preres_put(&trans->c->journal, &trans->journal_preres);
952 if (likely(!(trans->flags & BTREE_INSERT_NOCHECK_RW)))
953 percpu_ref_put(&trans->c->writes);
955 bch2_trans_reset(trans, !ret ? TRANS_RESET_NOTRAVERSE : 0);
959 ret = bch2_trans_commit_error(trans, i, ret);
966 int bch2_trans_update(struct btree_trans *trans, struct btree_iter *iter,
967 struct bkey_i *k, enum btree_trigger_flags flags)
969 struct btree_insert_entry *i, n = (struct btree_insert_entry) {
970 .trigger_flags = flags, .iter = iter, .k = k
973 EBUG_ON(bkey_cmp(iter->pos,
974 (iter->flags & BTREE_ITER_IS_EXTENTS)
975 ? bkey_start_pos(&k->k)
978 iter->flags |= BTREE_ITER_KEEP_UNTIL_COMMIT;
980 if (btree_node_type_is_extents(iter->btree_id)) {
981 iter->pos_after_commit = k->k.p;
982 iter->flags |= BTREE_ITER_SET_POS_AFTER_COMMIT;
986 * Pending updates are kept sorted: first, find position of new update:
988 trans_for_each_update(trans, i)
989 if (btree_iter_cmp(iter, i->iter) <= 0)
993 * Now delete/trim any updates the new update overwrites:
995 if (i > trans->updates &&
996 i[-1].iter->btree_id == iter->btree_id &&
997 bkey_cmp(iter->pos, i[-1].k->k.p) < 0)
998 bch2_cut_back(n.iter->pos, i[-1].k);
1000 while (i < trans->updates + trans->nr_updates &&
1001 iter->btree_id == i->iter->btree_id &&
1002 bkey_cmp(n.k->k.p, i->k->k.p) >= 0)
1003 array_remove_item(trans->updates, trans->nr_updates,
1004 i - trans->updates);
1006 if (i < trans->updates + trans->nr_updates &&
1007 iter->btree_id == i->iter->btree_id &&
1008 bkey_cmp(n.k->k.p, i->iter->pos) > 0) {
1010 * When we have an extent that overwrites the start of another
1011 * update, trimming that extent will mean the iterator's
1012 * position has to change since the iterator position has to
1013 * match the extent's start pos - but we don't want to change
1014 * the iterator pos if some other code is using it, so we may
1017 if (trans->iters_live & (1ULL << i->iter->idx)) {
1018 i->iter = bch2_trans_copy_iter(trans, i->iter);
1019 if (IS_ERR(i->iter)) {
1020 trans->need_reset = true;
1021 return PTR_ERR(i->iter);
1024 i->iter->flags |= BTREE_ITER_KEEP_UNTIL_COMMIT;
1025 bch2_trans_iter_put(trans, i->iter);
1028 bch2_cut_front(n.k->k.p, i->k);
1029 bch2_btree_iter_set_pos(i->iter, n.k->k.p);
1032 EBUG_ON(trans->nr_updates >= trans->nr_iters);
1034 array_insert_item(trans->updates, trans->nr_updates,
1035 i - trans->updates, n);
1039 int __bch2_btree_insert(struct btree_trans *trans,
1040 enum btree_id id, struct bkey_i *k)
1042 struct btree_iter *iter;
1045 iter = bch2_trans_get_iter(trans, id, bkey_start_pos(&k->k),
1048 return PTR_ERR(iter);
1050 ret = bch2_btree_iter_traverse(iter) ?:
1051 bch2_trans_update(trans, iter, k, 0);
1052 bch2_trans_iter_put(trans, iter);
1057 * bch2_btree_insert - insert keys into the extent btree
1058 * @c: pointer to struct bch_fs
1059 * @id: btree to insert into
1060 * @insert_keys: list of keys to insert
1061 * @hook: insert callback
1063 int bch2_btree_insert(struct bch_fs *c, enum btree_id id,
1065 struct disk_reservation *disk_res,
1066 u64 *journal_seq, int flags)
1068 return bch2_trans_do(c, disk_res, journal_seq, flags,
1069 __bch2_btree_insert(&trans, id, k));
1072 int bch2_btree_delete_at_range(struct btree_trans *trans,
1073 struct btree_iter *iter,
1080 while ((k = bch2_btree_iter_peek(iter)).k &&
1081 !(ret = bkey_err(k)) &&
1082 bkey_cmp(iter->pos, end) < 0) {
1083 struct bkey_i delete;
1085 bch2_trans_begin(trans);
1087 bkey_init(&delete.k);
1090 * For extents, iter.pos won't necessarily be the same as
1091 * bkey_start_pos(k.k) (for non extents they always will be the
1092 * same). It's important that we delete starting from iter.pos
1093 * because the range we want to delete could start in the middle
1096 * (bch2_btree_iter_peek() does guarantee that iter.pos >=
1097 * bkey_start_pos(k.k)).
1099 delete.k.p = iter->pos;
1101 if (btree_node_type_is_extents(iter->btree_id)) {
1102 unsigned max_sectors =
1103 KEY_SIZE_MAX & (~0 << trans->c->block_bits);
1105 /* create the biggest key we can */
1106 bch2_key_resize(&delete.k, max_sectors);
1107 bch2_cut_back(end, &delete);
1109 ret = bch2_extent_trim_atomic(&delete, iter);
1114 bch2_trans_update(trans, iter, &delete, 0);
1115 ret = bch2_trans_commit(trans, NULL, journal_seq,
1116 BTREE_INSERT_NOFAIL);
1120 bch2_trans_cond_resched(trans);
1123 if (ret == -EINTR) {
1132 int bch2_btree_delete_at(struct btree_trans *trans,
1133 struct btree_iter *iter, unsigned flags)
1140 bch2_trans_update(trans, iter, &k, 0);
1141 return bch2_trans_commit(trans, NULL, NULL,
1142 BTREE_INSERT_NOFAIL|
1143 BTREE_INSERT_USE_RESERVE|flags);
1147 * bch_btree_delete_range - delete everything within a given range
1149 * Range is a half open interval - [start, end)
1151 int bch2_btree_delete_range(struct bch_fs *c, enum btree_id id,
1152 struct bpos start, struct bpos end,
1155 struct btree_trans trans;
1156 struct btree_iter *iter;
1160 * XXX: whether we need mem/more iters depends on whether this btree id
1163 bch2_trans_init(&trans, c, BTREE_ITER_MAX, 512);
1165 iter = bch2_trans_get_iter(&trans, id, start, BTREE_ITER_INTENT);
1167 ret = bch2_btree_delete_at_range(&trans, iter, end, journal_seq);
1168 ret = bch2_trans_exit(&trans) ?: ret;
1170 BUG_ON(ret == -EINTR);