@@ -4752,6 +4752,9 @@ static void btrfs_uring_read_finished(struct io_uring_cmd *cmd, unsigned int iss
size_t page_offset;
ssize_t ret;
+ /* The inode lock has already been acquired in btrfs_uring_read_extent. */
+ btrfs_lockdep_inode_acquire(inode, i_rwsem);
+
if (priv->err) {
ret = priv->err;
goto out;
@@ -4860,6 +4863,13 @@ static int btrfs_uring_read_extent(struct kiocb *iocb, struct iov_iter *iter,
* and inode and freeing the allocations.
*/
+ /*
+ * We're returning to userspace with the inode lock held, and that's
+ * okay - it'll get unlocked in a worker thread. Call
+ * btrfs_lockdep_inode_release() to avoid confusing lockdep.
+ */
+ btrfs_lockdep_inode_release(inode, i_rwsem);
+
return -EIOCBQUEUED;
out_fail:
@@ -128,6 +128,16 @@ enum btrfs_lockdep_trans_states {
#define btrfs_lockdep_release(owner, lock) \
rwsem_release(&owner->lock##_map, _THIS_IP_)
+/*
+ * Used to account for the fact that when doing io_uring encoded I/O, we can
+ * return to userspace with the inode lock still held.
+ */
+#define btrfs_lockdep_inode_acquire(owner, lock) \
+ rwsem_acquire_read(&owner->vfs_inode.lock.dep_map, 0, 0, _THIS_IP_)
+
+#define btrfs_lockdep_inode_release(owner, lock) \
+ rwsem_release(&owner->vfs_inode.lock.dep_map, _THIS_IP_)
+
/*
* Macros for the transaction states wait events, similar to the generic wait
* event macros.
Lockdep doesn't like the fact that btrfs_uring_read_extent() returns to userspace still holding the inode lock, even though we release it once the I/O finishes. Add calls to rwsem_release() and rwsem_acquire_read() to work round this. Signed-off-by: Mark Harmstone <maharmstone@fb.com> Reported-by: Johannes Thumshirn <johannes.thumshirn@wdc.com> --- Changelog: v2: Add btrfs_lockdep_inode_acquire and btrfs_lockdep_inode_release helpers, and remove unneeded CONFIG_DEBUG_LOCK_ALLOC #ifdefs. fs/btrfs/ioctl.c | 10 ++++++++++ fs/btrfs/locking.h | 10 ++++++++++ 2 files changed, 20 insertions(+)