bcachefs: Sort & deduplicate updates in bch2_trans_update()

Previously, when doing multiple update in the same transaction commit
that overwrote each other, we relied on doing the updates in the same
order as the bch2_trans_update() calls in order to get the correct
result. But that wasn't correct for triggers; bch2_trans_mark_update()
when marking overwrites would do the wrong thing because it hadn't seen
the update that was being overwritten.
Signed-off-by: Kent Overstreet <kent.overstreet@gmail.com>
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

fs/bcachefs/btree_iter.c

@@ -1793,10 +1793,9 @@ int bch2_trans_iter_free(struct btree_trans *trans,
 static int bch2_trans_realloc_iters(struct btree_trans *trans,
 				    unsigned new_size)
 {
-	void *new_iters, *new_updates, *new_sorted;
+	void *new_iters, *new_updates;
 	size_t iters_bytes;
 	size_t updates_bytes;
-	size_t sorted_bytes;
 
 	new_size = roundup_pow_of_two(new_size);
 
@@ -1811,11 +1810,8 @@ static int bch2_trans_realloc_iters(struct btree_trans *trans,
 
 	iters_bytes	= sizeof(struct btree_iter) * new_size;
 	updates_bytes	= sizeof(struct btree_insert_entry) * new_size;
-	sorted_bytes	= sizeof(u8) * new_size;
 
-	new_iters = kmalloc(iters_bytes +
-			    updates_bytes +
-			    sorted_bytes, GFP_NOFS);
+	new_iters = kmalloc(iters_bytes + updates_bytes, GFP_NOFS);
 	if (new_iters)
 		goto success;
 
@@ -1825,7 +1821,6 @@ static int bch2_trans_realloc_iters(struct btree_trans *trans,
 	trans->used_mempool = true;
 success:
 	new_updates	= new_iters + iters_bytes;
-	new_sorted	= new_updates + updates_bytes;
 
 	memcpy(new_iters, trans->iters,
 	       sizeof(struct btree_iter) * trans->nr_iters);
@@ -1842,7 +1837,6 @@ static int bch2_trans_realloc_iters(struct btree_trans *trans,
 
 	trans->iters		= new_iters;
 	trans->updates		= new_updates;
-	trans->updates_sorted	= new_sorted;
 	trans->size		= new_size;
 
 	if (trans->iters_live) {
@@ -1891,6 +1885,7 @@ static struct btree_iter *btree_trans_iter_alloc(struct btree_trans *trans)
 got_slot:
 	BUG_ON(trans->iters_linked & (1ULL << idx));
 	trans->iters_linked |= 1ULL << idx;
+	trans->iters[idx].flags = 0;
 	return &trans->iters[idx];
 }
 
@@ -1906,6 +1901,9 @@ static inline void btree_iter_copy(struct btree_iter *dst,
 		if (btree_node_locked(dst, i))
 			six_lock_increment(&dst->l[i].b->c.lock,
 					   __btree_lock_want(dst, i));
+
+	dst->flags &= ~BTREE_ITER_KEEP_UNTIL_COMMIT;
+	dst->flags &= ~BTREE_ITER_SET_POS_AFTER_COMMIT;
 }
 
 static inline struct bpos bpos_diff(struct bpos l, struct bpos r)
@@ -1956,7 +1954,6 @@ static struct btree_iter *__btree_trans_get_iter(struct btree_trans *trans,
 		iter = best;
 	}
 
-	iter->flags &= ~BTREE_ITER_KEEP_UNTIL_COMMIT;
 	iter->flags &= ~(BTREE_ITER_SLOTS|BTREE_ITER_INTENT|BTREE_ITER_PREFETCH);
 	iter->flags |= flags & (BTREE_ITER_SLOTS|BTREE_ITER_INTENT|BTREE_ITER_PREFETCH);
 
@@ -1968,6 +1965,7 @@ static struct btree_iter *__btree_trans_get_iter(struct btree_trans *trans,
 	BUG_ON(iter->btree_id != btree_id);
 	BUG_ON((iter->flags ^ flags) & BTREE_ITER_TYPE);
 	BUG_ON(iter->flags & BTREE_ITER_KEEP_UNTIL_COMMIT);
+	BUG_ON(iter->flags & BTREE_ITER_SET_POS_AFTER_COMMIT);
 	BUG_ON(trans->iters_live & (1ULL << iter->idx));
 
 	trans->iters_live	|= 1ULL << iter->idx;
@@ -2030,7 +2028,6 @@ struct btree_iter *bch2_trans_copy_iter(struct btree_trans *trans,
 	 * it's cheap to copy it again:
 	 */
 	trans->iters_touched &= ~(1ULL << iter->idx);
-	iter->flags &= ~BTREE_ITER_KEEP_UNTIL_COMMIT;
 
 	return iter;
 }
@@ -2090,7 +2087,8 @@ void bch2_trans_reset(struct btree_trans *trans, unsigned flags)
 	struct btree_iter *iter;
 
 	trans_for_each_iter(trans, iter)
-		iter->flags &= ~BTREE_ITER_KEEP_UNTIL_COMMIT;
+		iter->flags &= ~(BTREE_ITER_KEEP_UNTIL_COMMIT|
+				 BTREE_ITER_SET_POS_AFTER_COMMIT);
 
 	bch2_trans_unlink_iters(trans);
 
@@ -2099,6 +2097,7 @@ void bch2_trans_reset(struct btree_trans *trans, unsigned flags)
 
 	trans->iters_touched &= trans->iters_live;
 
+	trans->need_reset		= 0;
 	trans->nr_updates		= 0;
 
 	if (flags & TRANS_RESET_MEM)
@@ -2127,7 +2126,6 @@ void bch2_trans_init(struct btree_trans *trans, struct bch_fs *c,
 	trans->size		= ARRAY_SIZE(trans->iters_onstack);
 	trans->iters		= trans->iters_onstack;
 	trans->updates		= trans->updates_onstack;
-	trans->updates_sorted	= trans->updates_sorted_onstack;
 	trans->fs_usage_deltas	= NULL;
 
 	if (expected_nr_iters > trans->size)

fs/bcachefs/btree_types.h

@@ -197,6 +197,7 @@ enum btree_iter_type {
  */
 #define BTREE_ITER_IS_EXTENTS		(1 << 6)
 #define BTREE_ITER_ERROR		(1 << 7)
+#define BTREE_ITER_SET_POS_AFTER_COMMIT	(1 << 8)
 
 enum btree_iter_uptodate {
 	BTREE_ITER_UPTODATE		= 0,
@@ -213,12 +214,13 @@ enum btree_iter_uptodate {
  * @nodes_intent_locked	- bitmask indicating which locks are intent locks
  */
 struct btree_iter {
-	u8			idx;
-
 	struct btree_trans	*trans;
 	struct bpos		pos;
+	struct bpos		pos_after_commit;
+
+	u16			flags;
+	u8			idx;
 
-	u8			flags;
 	enum btree_iter_uptodate uptodate:4;
 	enum btree_id		btree_id:4;
 	unsigned		level:4,
@@ -246,6 +248,7 @@ static inline enum btree_iter_type btree_iter_type(struct btree_iter *iter)
 
 struct btree_insert_entry {
 	unsigned		trigger_flags;
+	unsigned		trans_triggers_run:1;
 	struct bkey_i		*k;
 	struct btree_iter	*iter;
 };
@@ -266,6 +269,7 @@ struct btree_trans {
 	unsigned		used_mempool:1;
 	unsigned		error:1;
 	unsigned		nounlock:1;
+	unsigned		need_reset:1;
 
 	unsigned		mem_top;
 	unsigned		mem_bytes;
@@ -273,7 +277,6 @@ struct btree_trans {
 
 	struct btree_iter	*iters;
 	struct btree_insert_entry *updates;
-	u8			*updates_sorted;
 
 	/* update path: */
 	struct journal_res	journal_res;
@@ -287,7 +290,6 @@ struct btree_trans {
 
 	struct btree_iter	iters_onstack[2];
 	struct btree_insert_entry updates_onstack[2];
-	u8			updates_sorted_onstack[2];
 };
 
 #define BTREE_FLAG(flag)						\

fs/bcachefs/btree_update.h

@@ -72,6 +72,8 @@ int bch2_btree_node_rewrite(struct bch_fs *c, struct btree_iter *,
 int bch2_btree_node_update_key(struct bch_fs *, struct btree_iter *,
 			       struct btree *, struct bkey_i_btree_ptr *);
 
+int bch2_trans_update(struct btree_trans *, struct btree_iter *,
+		      struct bkey_i *, enum btree_trigger_flags);
 int __bch2_trans_commit(struct btree_trans *);
 
 /**
@@ -96,19 +98,6 @@ static inline int bch2_trans_commit(struct btree_trans *trans,
 	return __bch2_trans_commit(trans);
 }
 
-static inline void bch2_trans_update(struct btree_trans *trans,
-				     struct btree_iter *iter, struct bkey_i *k,
-				     enum btree_trigger_flags flags)
-{
-	EBUG_ON(trans->nr_updates >= trans->nr_iters);
-
-	iter->flags |= BTREE_ITER_KEEP_UNTIL_COMMIT;
-
-	trans->updates[trans->nr_updates++] = (struct btree_insert_entry) {
-		.trigger_flags = flags, .iter = iter, .k = k
-	};
-}
-
 #define __bch2_trans_do(_trans, _disk_res, _journal_seq,		\
 			_flags,	_reset_flags, _do)			\
 ({									\

fs/bcachefs/btree_update_leaf.c

@@ -21,18 +21,12 @@
 #include <linux/sort.h>
 
 static inline bool same_leaf_as_prev(struct btree_trans *trans,
-				     unsigned idx)
+				     struct btree_insert_entry *i)
 {
-	return idx &&
-		trans->updates[trans->updates_sorted[idx]].iter->l[0].b ==
-		trans->updates[trans->updates_sorted[idx - 1]].iter->l[0].b;
+	return i != trans->updates &&
+		i[0].iter->l[0].b == i[-1].iter->l[0].b;
 }
 
-#define trans_for_each_update_sorted(_trans, _i, _iter)			\
-	for (_iter = 0;							\
-	     _iter < _trans->nr_updates &&				\
-	     (_i = _trans->updates + _trans->updates_sorted[_iter], 1);	\
-	     _iter++)
 
 inline void bch2_btree_node_lock_for_insert(struct bch_fs *c, struct btree *b,
 					    struct btree_iter *iter)
@@ -51,28 +45,6 @@ inline void bch2_btree_node_lock_for_insert(struct bch_fs *c, struct btree *b,
 		bch2_btree_init_next(c, b, iter);
 }
 
-static inline void btree_trans_sort_updates(struct btree_trans *trans)
-{
-	struct btree_insert_entry *l, *r;
-	unsigned nr = 0, pos;
-
-	trans_for_each_update(trans, l) {
-		for (pos = 0; pos < nr; pos++) {
-			r = trans->updates + trans->updates_sorted[pos];
-
-			if (btree_iter_cmp(l->iter, r->iter) <= 0)
-				break;
-		}
-
-		memmove(&trans->updates_sorted[pos + 1],
-			&trans->updates_sorted[pos],
-			(nr - pos) * sizeof(trans->updates_sorted[0]));
-
-		trans->updates_sorted[pos] = l - trans->updates;
-		nr++;
-	}
-}
-
 /* Inserting into a given leaf node (last stage of insert): */
 
 /* Handle overwrites and do insert, for non extents: */
@@ -409,7 +381,7 @@ bch2_trans_commit_write_locked(struct btree_trans *trans,
 	struct bch_fs *c = trans->c;
 	struct bch_fs_usage_online *fs_usage = NULL;
 	struct btree_insert_entry *i;
-	unsigned iter, u64s = 0;
+	unsigned u64s = 0;
 	bool marking = false;
 	int ret;
 
@@ -426,9 +398,9 @@ bch2_trans_commit_write_locked(struct btree_trans *trans,
 
 	prefetch(&trans->c->journal.flags);
 
-	trans_for_each_update_sorted(trans, i, iter) {
+	trans_for_each_update(trans, i) {
 		/* Multiple inserts might go to same leaf: */
-		if (!same_leaf_as_prev(trans, iter))
+		if (!same_leaf_as_prev(trans, i))
 			u64s = 0;
 
 		u64s += i->k->k.u64s;
@@ -510,7 +482,6 @@ static inline int do_bch2_trans_commit(struct btree_trans *trans,
 {
 	struct btree_insert_entry *i;
 	struct btree_iter *iter;
-	unsigned idx;
 	int ret;
 
 	trans_for_each_update(trans, i)
@@ -545,21 +516,15 @@ static inline int do_bch2_trans_commit(struct btree_trans *trans,
 			btree_insert_entry_checks(trans, i->iter, i->k);
 	bch2_btree_trans_verify_locks(trans);
 
-	/*
-	 * No more updates can be added - sort updates so we can take write
-	 * locks in the correct order:
-	 */
-	btree_trans_sort_updates(trans);
-
-	trans_for_each_update_sorted(trans, i, idx)
-		if (!same_leaf_as_prev(trans, idx))
+	trans_for_each_update(trans, i)
+		if (!same_leaf_as_prev(trans, i))
 			bch2_btree_node_lock_for_insert(trans->c,
 						i->iter->l[0].b, i->iter);
 
 	ret = bch2_trans_commit_write_locked(trans, stopped_at);
 
-	trans_for_each_update_sorted(trans, i, idx)
-		if (!same_leaf_as_prev(trans, idx))
+	trans_for_each_update(trans, i)
+		if (!same_leaf_as_prev(trans, i))
 			bch2_btree_node_unlock_write_inlined(i->iter->l[0].b,
 							     i->iter);
 
@@ -575,8 +540,8 @@ static inline int do_bch2_trans_commit(struct btree_trans *trans,
 	if (trans->flags & BTREE_INSERT_NOUNLOCK)
 		trans->nounlock = true;
 
-	trans_for_each_update_sorted(trans, i, idx)
-		if (!same_leaf_as_prev(trans, idx))
+	trans_for_each_update(trans, i)
+		if (!same_leaf_as_prev(trans, i))
 			bch2_foreground_maybe_merge(trans->c, i->iter,
 						    0, trans->flags);
 
@@ -708,9 +673,13 @@ bch2_trans_commit_get_rw_cold(struct btree_trans *trans)
 int __bch2_trans_commit(struct btree_trans *trans)
 {
 	struct btree_insert_entry *i = NULL;
+	struct btree_iter *iter;
+	bool trans_trigger_run;
 	unsigned u64s;
 	int ret = 0;
 
+	BUG_ON(trans->need_reset);
+
 	if (!trans->nr_updates)
 		goto out_noupdates;
 
@@ -730,9 +699,29 @@ int __bch2_trans_commit(struct btree_trans *trans)
 	}
 
 	/*
-	 * note: running triggers will append more updates to the list of
-	 * updates as we're walking it:
+	 * Running triggers will append more updates to the list of updates as
+	 * we're walking it:
 	 */
+	do {
+		trans_trigger_run = false;
+
+		trans_for_each_update(trans, i) {
+			if (iter_has_trans_triggers(i->iter) &&
+			    !i->trans_triggers_run) {
+				i->trans_triggers_run = true;
+				trans_trigger_run = true;
+
+				ret = bch2_trans_mark_update(trans, i->iter, i->k,
+							     i->trigger_flags);
+				if (unlikely(ret)) {
+					if (ret == -EINTR)
+						trace_trans_restart_mark(trans->ip);
+					goto out;
+				}
+			}
+		}
+	} while (trans_trigger_run);
+
 	trans_for_each_update(trans, i) {
 		/* we know trans->nounlock won't be set here: */
 		if (unlikely(!(i->iter->locks_want < 1
@@ -743,16 +732,6 @@ int __bch2_trans_commit(struct btree_trans *trans)
 			goto out;
 		}
 
-		if (iter_has_trans_triggers(i->iter)) {
-			ret = bch2_trans_mark_update(trans, i->iter, i->k,
-						     i->trigger_flags);
-			if (unlikely(ret)) {
-				if (ret == -EINTR)
-					trace_trans_restart_mark(trans->ip);
-				goto out;
-			}
-		}
-
 		u64s = jset_u64s(i->k->k.u64s);
 		if (0)
 			trans->journal_preres_u64s += u64s;
@@ -768,6 +747,15 @@ int __bch2_trans_commit(struct btree_trans *trans)
 
 	if (ret)
 		goto err;
+
+	trans_for_each_iter(trans, iter)
+		if ((trans->iters_live & (1ULL << iter->idx)) &&
+		    (iter->flags & BTREE_ITER_SET_POS_AFTER_COMMIT)) {
+			if (trans->flags & BTREE_INSERT_NOUNLOCK)
+				bch2_btree_iter_set_pos_same_leaf(iter, iter->pos_after_commit);
+			else
+				bch2_btree_iter_set_pos(iter, iter->pos_after_commit);
+		}
 out:
 	bch2_journal_preres_put(&trans->c->journal, &trans->journal_preres);
 
@@ -785,6 +773,76 @@ int __bch2_trans_commit(struct btree_trans *trans)
 	goto retry;
 }
 
+int bch2_trans_update(struct btree_trans *trans, struct btree_iter *iter,
+		      struct bkey_i *k, enum btree_trigger_flags flags)
+{
+	struct btree_insert_entry *i, n = (struct btree_insert_entry) {
+		.trigger_flags = flags, .iter = iter, .k = k
+	};
+
+	EBUG_ON(bkey_cmp(iter->pos, bkey_start_pos(&k->k)));
+
+	iter->flags |= BTREE_ITER_KEEP_UNTIL_COMMIT;
+
+	if (iter->flags & BTREE_ITER_IS_EXTENTS) {
+		iter->pos_after_commit = k->k.p;
+		iter->flags |= BTREE_ITER_SET_POS_AFTER_COMMIT;
+	}
+
+	/*
+	 * Pending updates are kept sorted: first, find position of new update:
+	 */
+	trans_for_each_update(trans, i)
+		if (btree_iter_cmp(iter, i->iter) <= 0)
+			break;
+
+	/*
+	 * Now delete/trim any updates the new update overwrites:
+	 */
+	if (i > trans->updates &&
+	    i[-1].iter->btree_id == iter->btree_id &&
+	    bkey_cmp(iter->pos, i[-1].k->k.p) < 0)
+		bch2_cut_back(n.iter->pos, i[-1].k);
+
+	while (i < trans->updates + trans->nr_updates &&
+	       iter->btree_id == i->iter->btree_id &&
+	       bkey_cmp(n.k->k.p, i->k->k.p) >= 0)
+		array_remove_item(trans->updates, trans->nr_updates,
+				  i - trans->updates);
+
+	if (i < trans->updates + trans->nr_updates &&
+	    iter->btree_id == i->iter->btree_id &&
+	    bkey_cmp(n.k->k.p, i->iter->pos) > 0) {
+		/*
+		 * When we have an extent that overwrites the start of another
+		 * update, trimming that extent will mean the iterator's
+		 * position has to change since the iterator position has to
+		 * match the extent's start pos - but we don't want to change
+		 * the iterator pos if some other code is using it, so we may
+		 * need to clone it:
+		 */
+		if (trans->iters_live & (1ULL << i->iter->idx)) {
+			i->iter = bch2_trans_copy_iter(trans, i->iter);
+			if (IS_ERR(i->iter)) {
+				trans->need_reset = true;
+				return PTR_ERR(i->iter);
+			}
+
+			i->iter->flags |= BTREE_ITER_KEEP_UNTIL_COMMIT;
+			bch2_trans_iter_put(trans, i->iter);
+		}
+
+		bch2_cut_front(n.k->k.p, i->k);
+		bch2_btree_iter_set_pos(i->iter, n.k->k.p);
+	}
+
+	EBUG_ON(trans->nr_updates >= trans->nr_iters);
+
+	array_insert_item(trans->updates, trans->nr_updates,
+			  i - trans->updates, n);
+	return 0;
+}
+
 static int __bch2_btree_insert(struct btree_trans *trans,
 			       enum btree_id id, struct bkey_i *k)
 {

fs/bcachefs/buckets.c

@@ -1433,30 +1433,6 @@ static int trans_get_key(struct btree_trans *trans,
 	return ret;
 }
 
-static void *trans_update_key(struct btree_trans *trans,
-			      struct btree_iter *iter,
-			      unsigned u64s)
-{
-	struct btree_insert_entry *i;
-	struct bkey_i *new_k;
-
-	new_k = bch2_trans_kmalloc(trans, u64s * sizeof(u64));
-	if (IS_ERR(new_k))
-		return new_k;
-
-	bkey_init(&new_k->k);
-	new_k->k.p = iter->pos;
-
-	trans_for_each_update(trans, i)
-		if (i->iter == iter) {
-			i->k = new_k;
-			return new_k;
-		}
-
-	bch2_trans_update(trans, iter, new_k, 0);
-	return new_k;
-}
-
 static int bch2_trans_mark_pointer(struct btree_trans *trans,
 			struct extent_ptr_decoded p,
 			s64 sectors, enum bch_data_type data_type)
@@ -1540,7 +1516,7 @@ static int bch2_trans_mark_pointer(struct btree_trans *trans,
 	u.data_type = u.dirty_sectors || u.cached_sectors
 		? data_type : 0;
 
-	a = trans_update_key(trans, iter, BKEY_ALLOC_U64s_MAX);
+	a = bch2_trans_kmalloc(trans, BKEY_ALLOC_U64s_MAX * 8);
 	ret = PTR_ERR_OR_ZERO(a);
 	if (ret)
 		goto out;
@@ -1548,6 +1524,7 @@ static int bch2_trans_mark_pointer(struct btree_trans *trans,
 	bkey_alloc_init(&a->k_i);
 	a->k.p = iter->pos;
 	bch2_alloc_pack(a, u);
+	bch2_trans_update(trans, iter, &a->k_i, 0);
 out:
 	bch2_trans_iter_put(trans, iter);
 	return ret;
@@ -1562,9 +1539,8 @@ static int bch2_trans_mark_stripe_ptr(struct btree_trans *trans,
 {
 	struct bch_fs *c = trans->c;
 	struct btree_iter *iter;
-	struct bkey_i *new_k;
 	struct bkey_s_c k;
-	struct bkey_s_stripe s;
+	struct bkey_i_stripe *s;
 	int ret = 0;
 
 	ret = trans_get_key(trans, BTREE_ID_EC, POS(0, p.idx), &iter, &k);
@@ -1579,21 +1555,21 @@ static int bch2_trans_mark_stripe_ptr(struct btree_trans *trans,
 		goto out;
 	}
 
-	new_k = trans_update_key(trans, iter, k.k->u64s);
-	ret = PTR_ERR_OR_ZERO(new_k);
+	s = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
+	ret = PTR_ERR_OR_ZERO(s);
 	if (ret)
 		goto out;
 
-	bkey_reassemble(new_k, k);
-	s = bkey_i_to_s_stripe(new_k);
+	bkey_reassemble(&s->k_i, k);
 
-	stripe_blockcount_set(s.v, p.block,
-		stripe_blockcount_get(s.v, p.block) +
+	stripe_blockcount_set(&s->v, p.block,
+		stripe_blockcount_get(&s->v, p.block) +
 		sectors);
 
-	*nr_data	= s.v->nr_blocks - s.v->nr_redundant;
-	*nr_parity	= s.v->nr_redundant;
-	bch2_bkey_to_replicas(&r->e, s.s_c);
+	*nr_data	= s->v.nr_blocks - s->v.nr_redundant;
+	*nr_parity	= s->v.nr_redundant;
+	bch2_bkey_to_replicas(&r->e, bkey_i_to_s_c(&s->k_i));
+	bch2_trans_update(trans, iter, &s->k_i, 0);
 out:
 	bch2_trans_iter_put(trans, iter);
 	return ret;
@@ -1674,7 +1650,6 @@ static int __bch2_trans_mark_reflink_p(struct btree_trans *trans,
 {
 	struct bch_fs *c = trans->c;
 	struct btree_iter *iter;
-	struct bkey_i *new_k;
 	struct bkey_s_c k;
 	struct bkey_i_reflink_v *r_v;
 	s64 ret;
@@ -1700,13 +1675,12 @@ static int __bch2_trans_mark_reflink_p(struct btree_trans *trans,
 	bch2_btree_iter_set_pos(iter, bkey_start_pos(k.k));
 	BUG_ON(iter->uptodate > BTREE_ITER_NEED_PEEK);
 
-	new_k = trans_update_key(trans, iter, k.k->u64s);
-	ret = PTR_ERR_OR_ZERO(new_k);
+	r_v = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
+	ret = PTR_ERR_OR_ZERO(r_v);
 	if (ret)
 		goto err;
 
-	bkey_reassemble(new_k, k);
-	r_v = bkey_i_to_reflink_v(new_k);
+	bkey_reassemble(&r_v->k_i, k);
 
 	le64_add_cpu(&r_v->v.refcount,
 		     !(flags & BTREE_TRIGGER_OVERWRITE) ? 1 : -1);
@@ -1715,6 +1689,8 @@ static int __bch2_trans_mark_reflink_p(struct btree_trans *trans,
 		r_v->k.type = KEY_TYPE_deleted;
 		set_bkey_val_u64s(&r_v->k, 0);
 	}
+
+	bch2_trans_update(trans, iter, &r_v->k_i, 0);
 out:
 	ret = k.k->p.offset - idx;
 err:

fs/bcachefs/fs-io.c

@@ -2383,7 +2383,7 @@ static long bchfs_fcollapse_finsert(struct bch_inode_info *inode,
 	struct address_space *mapping = inode->v.i_mapping;
 	struct bkey_on_stack copy;
 	struct btree_trans trans;
-	struct btree_iter *src, *dst, *del = NULL;
+	struct btree_iter *src, *dst;
 	loff_t shift, new_size;
 	u64 src_start;
 	int ret;
@@ -2513,29 +2513,6 @@ static long bchfs_fcollapse_finsert(struct bch_inode_info *inode,
 
 		next_pos = insert ? bkey_start_pos(&delete.k) : delete.k.p;
 
-		/*
-		 * If the new and old keys overlap (because we're moving an
-		 * extent that's bigger than the amount we're collapsing by),
-		 * we need to trim the delete key here so they don't overlap
-		 * because overlaps on insertions aren't handled before
-		 * triggers are run, so the overwrite will get double counted
-		 * by the triggers machinery:
-		 */
-		if (insert &&
-		    bkey_cmp(bkey_start_pos(&copy.k->k), delete.k.p) < 0) {
-			bch2_cut_back(bkey_start_pos(&copy.k->k), &delete);
-		} else if (!insert &&
-			   bkey_cmp(copy.k->k.p,
-				    bkey_start_pos(&delete.k)) > 0) {
-			bch2_cut_front(copy.k->k.p, &delete);
-
-			del = bch2_trans_copy_iter(&trans, src);
-			BUG_ON(IS_ERR_OR_NULL(del));
-
-			bch2_btree_iter_set_pos(del,
-				bkey_start_pos(&delete.k));
-		}
-
 		if (copy.k->k.size == k.k->size) {
 			/*
 			 * If we're moving the entire extent, we can skip
@@ -2553,18 +2530,13 @@ static long bchfs_fcollapse_finsert(struct bch_inode_info *inode,
 			BUG_ON(ret);
 		}
 
-		bch2_trans_update(&trans, dst, copy.k, trigger_flags);
-		bch2_trans_update(&trans, del ?: src, &delete, trigger_flags);
-
-		ret = bch2_trans_commit(&trans, &disk_res,
-					&inode->ei_journal_seq,
-					BTREE_INSERT_NOFAIL);
+		ret =   bch2_trans_update(&trans, src, &delete, trigger_flags) ?:
+			bch2_trans_update(&trans, dst, copy.k, trigger_flags) ?:
+			bch2_trans_commit(&trans, &disk_res,
+					  &inode->ei_journal_seq,
+					  BTREE_INSERT_NOFAIL);
 		bch2_disk_reservation_put(c, &disk_res);
 bkey_err:
-		if (del)
-			bch2_trans_iter_put(&trans, del);
-		del = NULL;
-
 		if (!ret)
 			bch2_btree_iter_set_pos(src, next_pos);