bcachefs: Use ei_update_lock consistently

This is prep work for using deferred btree updates for inode updates -
the way inodes are done now we're relying on btree locking for ei_inode
and ei_update_lock could probably be removed, but it'll actually be
needed when we switch to deferred updates.

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

2 files changed, 47 insertions, 3 deletions

diff --git a/fs/bcachefs/acl.c b/fs/bcachefs/acl.c
index c81e5365ec84..2856736f7224 100644
--- a/fs/bcachefs/acl.c
+++ b/fs/bcachefs/acl.c
@@ -304,13 +304,14 @@ int bch2_set_acl(struct mnt_idmap *idmap,
 	umode_t mode = inode->v.i_mode;
 	int ret;
 
+	mutex_lock(&inode->ei_update_lock);
+	bch2_trans_init(&trans, c);
+
 	if (type == ACL_TYPE_ACCESS && acl) {
 		ret = posix_acl_update_mode(idmap, &inode->v, &mode, &acl);
 		if (ret)
-			return ret;
+			goto err;
 	}
-
-	bch2_trans_init(&trans, c);
 retry:
 	bch2_trans_begin(&trans);
 
@@ -336,6 +337,7 @@ retry:
 	set_cached_acl(&inode->v, type, acl);
 err:
 	bch2_trans_exit(&trans);
+	mutex_unlock(&inode->ei_update_lock);
 
 	return ret;
 }
diff --git a/fs/bcachefs/fs.c b/fs/bcachefs/fs.c
index f10ee147d389..a4d82252bc49 100644
--- a/fs/bcachefs/fs.c
+++ b/fs/bcachefs/fs.c
@@ -51,6 +51,30 @@ static void journal_seq_copy(struct bch_inode_info *dst,
 	} while ((v = cmpxchg(&dst->ei_journal_seq, old, journal_seq)) != old);
 }
 
+static inline int ptrcmp(void *l, void *r)
+{
+	return (l > r) - (l < r);
+}
+
+#define __bch2_lock_inodes(_lock, ...)					\
+do {									\
+	struct bch_inode_info *a[] = { NULL, __VA_ARGS__ };		\
+	unsigned i;							\
+									\
+	bubble_sort(&a[1], ARRAY_SIZE(a) - 1 , ptrcmp);			\
+									\
+	for (i = ARRAY_SIZE(a) - 1; a[i]; --i)				\
+		if (a[i] != a[i - 1]) {					\
+			if (_lock)					\
+				mutex_lock_nested(&a[i]->ei_update_lock, i);\
+			else						\
+				mutex_unlock(&a[i]->ei_update_lock);	\
+		}							\
+} while (0)
+
+#define bch2_lock_inodes(...)	__bch2_lock_inodes(true, __VA_ARGS__)
+#define bch2_unlock_inodes(...)	__bch2_lock_inodes(false, __VA_ARGS__)
+
 static void __pagecache_lock_put(struct pagecache_lock *lock, long i)
 {
 	BUG_ON(atomic_long_read(&lock->v) == 0);
@@ -161,6 +185,8 @@ int __must_check bch2_write_inode_trans(struct btree_trans *trans,
 	struct bkey_inode_buf *inode_p;
 	int ret;
 
+	lockdep_assert_held(&inode->ei_update_lock);
+
 	iter = bch2_trans_get_iter(trans, BTREE_ID_INODES,
 			POS(inode->v.i_ino, 0),
 			BTREE_ITER_SLOTS|BTREE_ITER_INTENT);
@@ -422,6 +448,9 @@ out:
 	posix_acl_release(acl);
 	return inode;
 err_trans:
+	if (!tmpfile)
+		mutex_unlock(&dir->ei_update_lock);
+
 	bch2_trans_exit(&trans);
 	make_bad_inode(&inode->v);
 	iput(&inode->v);
@@ -490,6 +519,7 @@ static int __bch2_link(struct bch_fs *c,
 	struct bch_inode_unpacked inode_u;
 	int ret;
 
+	mutex_lock(&inode->ei_update_lock);
 	bch2_trans_init(&trans, c);
 retry:
 	bch2_trans_begin(&trans);
@@ -515,6 +545,7 @@ retry:
 		bch2_inode_update_after_write(c, inode, &inode_u, ATTR_CTIME);
 
 	bch2_trans_exit(&trans);
+	mutex_unlock(&inode->ei_update_lock);
 	return ret;
 }
 
@@ -575,6 +606,7 @@ static int bch2_unlink(struct inode *vdir, struct dentry *dentry)
 	struct btree_trans trans;
 	int ret;
 
+	bch2_lock_inodes(dir, inode);
 	bch2_trans_init(&trans, c);
 retry:
 	bch2_trans_begin(&trans);
@@ -607,6 +639,7 @@ retry:
 				      ATTR_MTIME);
 err:
 	bch2_trans_exit(&trans);
+	bch2_unlock_inodes(dir, inode);
 
 	return ret;
 }
@@ -771,6 +804,11 @@ static int bch2_rename2(struct mnt_idmap *idmap,
 			return ret;
 	}
 
+	bch2_lock_inodes(i.src_dir,
+			 i.dst_dir,
+			 i.src_inode,
+			 i.dst_inode);
+
 	bch2_trans_init(&trans, c);
 retry:
 	bch2_trans_begin(&trans);
@@ -818,6 +856,10 @@ retry:
 					      ATTR_CTIME);
 err:
 	bch2_trans_exit(&trans);
+	bch2_unlock_inodes(i.src_dir,
+			   i.dst_dir,
+			   i.src_inode,
+			   i.dst_inode);
 
 	return ret;
 }