bcachefs: Use deferred btree updates for inode updates

Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

5 files changed, 44 insertions, 18 deletions

diff --git a/fs/bcachefs/btree_iter.h b/fs/bcachefs/btree_iter.h
index 52e0e003153b..a64ed6d32175 100644
--- a/fs/bcachefs/btree_iter.h
+++ b/fs/bcachefs/btree_iter.h
@@ -104,6 +104,7 @@ void bch2_btree_node_iter_fix(struct btree_iter *, struct btree *,
 			      unsigned, unsigned);
 
 int bch2_btree_iter_unlock(struct btree_iter *);
+bool bch2_btree_iter_relock(struct btree_iter *);
 
 bool __bch2_btree_iter_upgrade(struct btree_iter *, unsigned);
 bool __bch2_btree_iter_upgrade_nounlock(struct btree_iter *, unsigned);
diff --git a/fs/bcachefs/btree_locking.h b/fs/bcachefs/btree_locking.h
index 48b50e066186..c036cd0458a4 100644
--- a/fs/bcachefs/btree_locking.h
+++ b/fs/bcachefs/btree_locking.h
@@ -203,8 +203,6 @@ static inline bool bch2_btree_node_relock(struct btree_iter *iter,
 		__bch2_btree_node_relock(iter, level);
 }
 
-bool bch2_btree_iter_relock(struct btree_iter *);
-
 void bch2_btree_node_unlock_write(struct btree *, struct btree_iter *);
 
 void __bch2_btree_node_lock_write(struct btree *, struct btree_iter *);
diff --git a/fs/bcachefs/fs-io.c b/fs/bcachefs/fs-io.c
index 7681cfbc6bed..f8657baf0521 100644
--- a/fs/bcachefs/fs-io.c
+++ b/fs/bcachefs/fs-io.c
@@ -287,11 +287,11 @@ static int bch2_extent_update(struct btree_trans *trans,
 			      bool direct,
 			      s64 *total_delta)
 {
-	struct btree_iter *inode_iter = NULL;
 	struct bch_inode_unpacked inode_u;
 	struct bkey_inode_buf inode_p;
 	bool allocating = false;
 	bool extended = false;
+	bool inode_locked = false;
 	s64 i_sectors_delta;
 	int ret;
 
@@ -314,16 +314,20 @@ static int bch2_extent_update(struct btree_trans *trans,
 	/* XXX: inode->i_size locking */
 	if (i_sectors_delta ||
 	    new_i_size > inode->ei_inode.bi_size) {
-		inode_iter = bch2_trans_get_iter(trans,
-			BTREE_ID_INODES,
-			POS(k->k.p.inode, 0),
-			BTREE_ITER_SLOTS|BTREE_ITER_INTENT);
-		if (IS_ERR(inode_iter))
-			return PTR_ERR(inode_iter);
+		bch2_btree_iter_unlock(extent_iter);
+		mutex_lock(&inode->ei_update_lock);
 
-		ret = bch2_btree_iter_traverse(inode_iter);
-		if (ret)
-			goto err;
+		if (!bch2_btree_iter_relock(extent_iter)) {
+			mutex_unlock(&inode->ei_update_lock);
+			return -EINTR;
+		}
+
+		inode_locked = true;
+
+		if (!inode->ei_inode_update)
+			inode->ei_inode_update =
+				bch2_deferred_update_alloc(trans->c,
+							BTREE_ID_INODES, 64);
 
 		inode_u = inode->ei_inode;
 		inode_u.bi_sectors += i_sectors_delta;
@@ -337,7 +341,8 @@ static int bch2_extent_update(struct btree_trans *trans,
 
 		bch2_inode_pack(&inode_p, &inode_u);
 		bch2_trans_update(trans,
-			BTREE_INSERT_ENTRY(inode_iter, &inode_p.inode.k_i));
+			BTREE_INSERT_DEFERRED(inode->ei_inode_update,
+					      &inode_p.inode.k_i));
 	}
 
 	ret = bch2_trans_commit(trans, disk_res,
@@ -371,13 +376,15 @@ static int bch2_extent_update(struct btree_trans *trans,
 	if (total_delta)
 		*total_delta += i_sectors_delta;
 err:
-	if (!IS_ERR_OR_NULL(inode_iter))
-		bch2_trans_iter_put(trans, inode_iter);
+	if (inode_locked)
+		mutex_unlock(&inode->ei_update_lock);
+
 	return ret;
 }
 
 static int bchfs_write_index_update(struct bch_write_op *wop)
 {
+	struct bch_fs *c = wop->c;
 	struct bchfs_write_op *op = container_of(wop,
 				struct bchfs_write_op, op);
 	struct quota_res *quota_res = op->is_dio
@@ -392,7 +399,7 @@ static int bchfs_write_index_update(struct bch_write_op *wop)
 
 	BUG_ON(k->k.p.inode != inode->v.i_ino);
 
-	bch2_trans_init(&trans, wop->c);
+	bch2_trans_init(&trans, c);
 	bch2_trans_preload_iters(&trans);
 
 	iter = bch2_trans_get_iter(&trans,
diff --git a/fs/bcachefs/fs.c b/fs/bcachefs/fs.c
index 02c7543e40c8..5f93ea76785f 100644
--- a/fs/bcachefs/fs.c
+++ b/fs/bcachefs/fs.c
@@ -156,12 +156,18 @@ int __must_check bch2_write_inode_trans(struct btree_trans *trans,
 				inode_set_fn set,
 				void *p)
 {
+	struct bch_fs *c = trans->c;
 	struct btree_iter *iter;
 	struct bkey_inode_buf *inode_p;
 	int ret;
 
 	lockdep_assert_held(&inode->ei_update_lock);
 
+	/* XXX: Don't do this with btree locks held */
+	if (!inode->ei_inode_update)
+		inode->ei_inode_update =
+			bch2_deferred_update_alloc(c, BTREE_ID_INODES, 64);
+#if 0
 	iter = bch2_trans_get_iter(trans, BTREE_ID_INODES,
 			POS(inode->v.i_ino, 0),
 			BTREE_ITER_SLOTS|BTREE_ITER_INTENT);
@@ -172,7 +178,7 @@ int __must_check bch2_write_inode_trans(struct btree_trans *trans,
 	ret = bch2_btree_iter_traverse(iter);
 	if (ret)
 		return ret;
-
+#endif
 	*inode_u = inode->ei_inode;
 
 	if (set) {
@@ -186,7 +192,15 @@ int __must_check bch2_write_inode_trans(struct btree_trans *trans,
 		return PTR_ERR(inode_p);
 
 	bch2_inode_pack(inode_p, inode_u);
-	bch2_trans_update(trans, BTREE_INSERT_ENTRY(iter, &inode_p->inode.k_i));
+
+	if (!inode->ei_inode_update)
+		bch2_trans_update(trans,
+			BTREE_INSERT_ENTRY(iter, &inode_p->inode.k_i));
+	else
+		bch2_trans_update(trans,
+			BTREE_INSERT_DEFERRED(inode->ei_inode_update,
+					      &inode_p->inode.k_i));
+
 	return 0;
 }
 
@@ -1431,6 +1445,7 @@ static struct inode *bch2_alloc_inode(struct super_block *sb)
 	mutex_init(&inode->ei_update_lock);
 	pagecache_lock_init(&inode->ei_pagecache_lock);
 	mutex_init(&inode->ei_quota_lock);
+	inode->ei_inode_update = NULL;
 	inode->ei_journal_seq = 0;
 
 	return &inode->v;
@@ -1494,6 +1509,10 @@ static void bch2_evict_inode(struct inode *vinode)
 
 	BUG_ON(!is_bad_inode(&inode->v) && inode->ei_quota_reserved);
 
+	if (inode->ei_inode_update)
+		bch2_deferred_update_free(c, inode->ei_inode_update);
+	inode->ei_inode_update = NULL;
+
 	if (!inode->v.i_nlink && !is_bad_inode(&inode->v)) {
 		bch2_quota_acct(c, inode->ei_qid, Q_SPC, -((s64) inode->v.i_blocks),
 				KEY_TYPE_QUOTA_WARN);
diff --git a/fs/bcachefs/fs.h b/fs/bcachefs/fs.h
index f949cd0d2a68..b9a8a9bc3e90 100644
--- a/fs/bcachefs/fs.h
+++ b/fs/bcachefs/fs.h
@@ -34,6 +34,7 @@ struct bch_inode_info {
 	struct inode		v;
 
 	struct mutex		ei_update_lock;
+	struct deferred_update	*ei_inode_update;
 	u64			ei_journal_seq;
 	u64			ei_quota_reserved;
 	unsigned long		ei_last_dirtied;