bcachefs: Factor out btree_key_can_insert()

working on getting rid of all the reasons bch2_insert_fixup_extent() can
fail/stop partway, which is needed for other refactorings.

One of the reasons we could have to bail out is if we're splitting a
compressed extent we might need to add to our disk reservation - but we
can check that before actually starting the insert.
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

fs/bcachefs/btree_types.h

@@ -448,7 +448,6 @@ enum btree_insert_ret {
 	/* extent spanned multiple leaf nodes: have to traverse to next node: */
 	BTREE_INSERT_NEED_TRAVERSE,
 	/* write lock held for too long */
-	BTREE_INSERT_NEED_RESCHED,
 	/* leaf node needs to be split */
 	BTREE_INSERT_BTREE_NODE_FULL,
 	BTREE_INSERT_JOURNAL_RES_FULL,

fs/bcachefs/btree_update_leaf.c

@@ -297,6 +297,30 @@ static inline int btree_trans_cmp(struct btree_insert_entry l,
 
 /* Normal update interface: */
 
+static enum btree_insert_ret
+btree_key_can_insert(struct btree_insert *trans,
+		      struct btree_insert_entry *insert,
+		      unsigned *u64s)
+{
+	struct bch_fs *c = trans->c;
+	struct btree *b = insert->iter->l[0].b;
+	static enum btree_insert_ret ret;
+
+	if (unlikely(btree_node_fake(b)))
+		return BTREE_INSERT_BTREE_NODE_FULL;
+
+	ret = !btree_node_is_extents(b)
+		? BTREE_INSERT_OK
+		: bch2_extent_can_insert(trans, insert, u64s);
+	if (ret)
+		return ret;
+
+	if (*u64s > bch_btree_keys_u64s_remaining(c, b))
+		return BTREE_INSERT_BTREE_NODE_FULL;
+
+	return BTREE_INSERT_OK;
+}
+
 /*
  * Get journal reservation, take write locks, and attempt to do btree update(s):
  */
@@ -336,24 +360,34 @@ static inline int do_btree_insert_at(struct btree_insert *trans,
 		goto out;
 	}
 
+	/*
+	 * Check if the insert will fit in the leaf node with the write lock
+	 * held, otherwise another thread could write the node changing the
+	 * amount of space available:
+	 */
 	u64s = 0;
 	trans_for_each_entry(trans, i) {
 		/* Multiple inserts might go to same leaf: */
 		if (!same_leaf_as_prev(trans, i))
 			u64s = 0;
 
-		/*
-		 * bch2_btree_node_insert_fits() must be called under write lock:
-		 * with only an intent lock, another thread can still call
-		 * bch2_btree_node_write(), converting an unwritten bset to a
-		 * written one
-		 */
 		u64s += i->k->k.u64s + i->extra_res;
-		if (!bch2_btree_node_insert_fits(c,
-				i->iter->l[0].b, u64s)) {
+		switch (btree_key_can_insert(trans, i, &u64s)) {
+		case BTREE_INSERT_OK:
+			break;
+		case BTREE_INSERT_BTREE_NODE_FULL:
 			ret = -EINTR;
 			*split = i->iter;
 			goto out;
+		case BTREE_INSERT_ENOSPC:
+			ret = -ENOSPC;
+			goto out;
+		case BTREE_INSERT_NEED_GC_LOCK:
+			ret = -EINTR;
+			*cycle_gc_lock = true;
+			goto out;
+		default:
+			BUG();
 		}
 	}
 
@@ -373,7 +407,6 @@ static inline int do_btree_insert_at(struct btree_insert *trans,
 			break;
 		case BTREE_INSERT_JOURNAL_RES_FULL:
 		case BTREE_INSERT_NEED_TRAVERSE:
-		case BTREE_INSERT_NEED_RESCHED:
 			ret = -EINTR;
 			break;
 		case BTREE_INSERT_BTREE_NODE_FULL:
@@ -383,10 +416,6 @@ static inline int do_btree_insert_at(struct btree_insert *trans,
 		case BTREE_INSERT_ENOSPC:
 			ret = -ENOSPC;
 			break;
-		case BTREE_INSERT_NEED_GC_LOCK:
-			ret = -EINTR;
-			*cycle_gc_lock = true;
-			break;
 		default:
 			BUG();
 		}

fs/bcachefs/extents.c

@@ -1113,8 +1113,6 @@ static bool bch2_extent_merge_inline(struct bch_fs *,
 				     struct bkey_packed *,
 				     bool);
 
-#define MAX_LOCK_HOLD_TIME	(5 * NSEC_PER_MSEC)
-
 static enum btree_insert_ret
 extent_insert_should_stop(struct extent_insert_state *s)
 {
@@ -1287,23 +1285,41 @@ extent_insert_advance_pos(struct extent_insert_state *s, struct bkey_s_c k)
 	return __extent_insert_advance_pos(s, next_pos, k);
 }
 
-static enum btree_insert_ret
-extent_insert_check_split_compressed(struct extent_insert_state *s,
-				     struct bkey_s_c k,
-				     enum bch_extent_overlap overlap)
+enum btree_insert_ret
+bch2_extent_can_insert(struct btree_insert *trans,
+		       struct btree_insert_entry *insert,
+		       unsigned *u64s)
 {
-	struct bch_fs *c = s->trans->c;
-	unsigned sectors;
+	struct btree_iter_level *l = &insert->iter->l[0];
+	struct btree_node_iter node_iter = l->iter;
+	enum bch_extent_overlap overlap;
+	struct bkey_packed *_k;
+	struct bkey unpacked;
+	struct bkey_s_c k;
+	int sectors;
+
+	_k = bch2_btree_node_iter_peek_filter(&node_iter, l->b,
+					      KEY_TYPE_DISCARD);
+	if (!_k)
+		return BTREE_INSERT_OK;
+
+	k = bkey_disassemble(l->b, _k, &unpacked);
+
+	overlap = bch2_extent_overlap(&insert->k->k, k.k);
+
+	/* account for having to split existing extent: */
+	if (overlap == BCH_EXTENT_OVERLAP_MIDDLE)
+		*u64s += _k->u64s;
 
 	if (overlap == BCH_EXTENT_OVERLAP_MIDDLE &&
 	    (sectors = bch2_extent_is_compressed(k))) {
 		int flags = BCH_DISK_RESERVATION_BTREE_LOCKS_HELD;
 
-		if (s->trans->flags & BTREE_INSERT_NOFAIL)
+		if (trans->flags & BTREE_INSERT_NOFAIL)
 			flags |= BCH_DISK_RESERVATION_NOFAIL;
 
-		switch (bch2_disk_reservation_add(c,
-				s->trans->disk_res,
+		switch (bch2_disk_reservation_add(trans->c,
+				trans->disk_res,
 				sectors * bch2_extent_nr_dirty_ptrs(k),
 				flags)) {
 		case 0:
@@ -1471,10 +1487,6 @@ __bch2_delete_fixup_extent(struct extent_insert_state *s)
 
 		overlap = bch2_extent_overlap(&insert->k, k.k);
 
-		ret = extent_insert_check_split_compressed(s, k.s_c, overlap);
-		if (ret)
-			break;
-
 		ret = extent_insert_advance_pos(s, k.s_c);
 		if (ret)
 			break;
@@ -1550,10 +1562,6 @@ __bch2_insert_fixup_extent(struct extent_insert_state *s)
 
 		overlap = bch2_extent_overlap(&insert->k, k.k);
 
-		ret = extent_insert_check_split_compressed(s, k.s_c, overlap);
-		if (ret)
-			break;
-
 		if (!k.k->size)
 			goto squash;
 

fs/bcachefs/extents.h

@@ -63,8 +63,10 @@ int bch2_extent_pick_ptr(struct bch_fs *, struct bkey_s_c,
 			 struct extent_pick_ptr *);
 
 enum btree_insert_ret
-bch2_insert_fixup_extent(struct btree_insert *,
-			struct btree_insert_entry *);
+bch2_extent_can_insert(struct btree_insert *, struct btree_insert_entry *,
+		       unsigned *);
+enum btree_insert_ret
+bch2_insert_fixup_extent(struct btree_insert *, struct btree_insert_entry *);
 
 bool bch2_extent_normalize(struct bch_fs *, struct bkey_s);
 void bch2_extent_mark_replicas_cached(struct bch_fs *, struct bkey_s_extent,