In ocfs2_move_extent(), tl_inode is currently locked after the global
bitmap inode. However, in ocfs2_flush_truncate_log(), the lock order
is reversed: tl_inode is locked first, followed by the global bitmap
inode.

This creates a classic ABBA deadlock scenario if two threads attempt
these operations concurrently and acquire the locks in different orders.

To prevent this, move the tl_inode locking earlier in
ocfs2_move_extent(), so that it always precedes the global bitmap
inode lock.

No functional changes beyond lock ordering.

Reported-by: syzbot+6bf948e47f9bac7aacfa@syzkaller.appspotmail.com
Closes: https://lore.kernel.org/all/67d5645c.050a0220.1dc86f.0004.GAE@google.com/
Signed-off-by: Ivan Pravdin <ipravdin.official@gmail.com>
---
v1 -> v2: Fixed unlocking order in ocfs2_move_extent.
v2 -> v3: Added missing out_unlock_tl_inode and renamed out_unlock_gb_mutex to
		  out_unlock_gb_inode and out_unlock_gb_inode to out_unlock.

 fs/ocfs2/move_extents.c | 19 +++++++++----------
 1 file changed, 9 insertions(+), 10 deletions(-)

diff --git a/fs/ocfs2/move_extents.c b/fs/ocfs2/move_extents.c
index 369c7d27befd..9d1fa8ecd97e 100644
--- a/fs/ocfs2/move_extents.c
+++ b/fs/ocfs2/move_extents.c
@@ -617,6 +617,8 @@ static int ocfs2_move_extent(struct ocfs2_move_extents_context *context,
 	 */
 	credits += OCFS2_INODE_UPDATE_CREDITS + 1;
 
+	inode_lock(tl_inode);
+
 	/*
 	 * ocfs2_move_extent() didn't reserve any clusters in lock_allocators()
 	 * logic, while we still need to lock the global_bitmap.
@@ -626,7 +628,7 @@ static int ocfs2_move_extent(struct ocfs2_move_extents_context *context,
 	if (!gb_inode) {
 		mlog(ML_ERROR, "unable to get global_bitmap inode\n");
 		ret = -EIO;
-		goto out;
+		goto out_unlock_tl_inode;
 	}
 
 	inode_lock(gb_inode);
@@ -634,16 +636,14 @@ static int ocfs2_move_extent(struct ocfs2_move_extents_context *context,
 	ret = ocfs2_inode_lock(gb_inode, &gb_bh, 1);
 	if (ret) {
 		mlog_errno(ret);
-		goto out_unlock_gb_mutex;
+		goto out_unlock_gb_inode;
 	}
 
-	inode_lock(tl_inode);
-
 	handle = ocfs2_start_trans(osb, credits);
 	if (IS_ERR(handle)) {
 		ret = PTR_ERR(handle);
 		mlog_errno(ret);
-		goto out_unlock_tl_inode;
+		goto out_unlock;
 	}
 
 	new_phys_blkno = ocfs2_clusters_to_blocks(inode->i_sb, *new_phys_cpos);
@@ -703,13 +703,12 @@ static int ocfs2_move_extent(struct ocfs2_move_extents_context *context,
 out_commit:
 	ocfs2_commit_trans(osb, handle);
 	brelse(gd_bh);
-
-out_unlock_tl_inode:
-	inode_unlock(tl_inode);
-
+out_unlock:
 	ocfs2_inode_unlock(gb_inode, 1);
-out_unlock_gb_mutex:
+out_unlock_gb_inode:
 	inode_unlock(gb_inode);
+out_unlock_tl_inode:
+	inode_unlock(tl_inode);
 	brelse(gb_bh);
 	iput(gb_inode);
 
-- 
2.45.2

On 2025/7/8 08:35, Ivan Pravdin wrote:
> In ocfs2_move_extent(), tl_inode is currently locked after the global
> bitmap inode. However, in ocfs2_flush_truncate_log(), the lock order
> is reversed: tl_inode is locked first, followed by the global bitmap
> inode.
> 
> This creates a classic ABBA deadlock scenario if two threads attempt
> these operations concurrently and acquire the locks in different orders.
> 
> To prevent this, move the tl_inode locking earlier in
> ocfs2_move_extent(), so that it always precedes the global bitmap
> inode lock.
> 
> No functional changes beyond lock ordering.
> 
> Reported-by: syzbot+6bf948e47f9bac7aacfa@syzkaller.appspotmail.com
> Closes: https://lore.kernel.org/all/67d5645c.050a0220.1dc86f.0004.GAE@google.com/
> Signed-off-by: Ivan Pravdin <ipravdin.official@gmail.com>
> ---
> v1 -> v2: Fixed unlocking order in ocfs2_move_extent.
> v2 -> v3: Added missing out_unlock_tl_inode and renamed out_unlock_gb_mutex to
> 		  out_unlock_gb_inode and out_unlock_gb_inode to out_unlock.
> 
>  fs/ocfs2/move_extents.c | 19 +++++++++----------
>  1 file changed, 9 insertions(+), 10 deletions(-)
> 
> diff --git a/fs/ocfs2/move_extents.c b/fs/ocfs2/move_extents.c
> index 369c7d27befd..9d1fa8ecd97e 100644
> --- a/fs/ocfs2/move_extents.c
> +++ b/fs/ocfs2/move_extents.c
> @@ -617,6 +617,8 @@ static int ocfs2_move_extent(struct ocfs2_move_extents_context *context,
>  	 */
>  	credits += OCFS2_INODE_UPDATE_CREDITS + 1;
>  
> +	inode_lock(tl_inode);
> +
>  	/*
>  	 * ocfs2_move_extent() didn't reserve any clusters in lock_allocators()
>  	 * logic, while we still need to lock the global_bitmap.
> @@ -626,7 +628,7 @@ static int ocfs2_move_extent(struct ocfs2_move_extents_context *context,
>  	if (!gb_inode) {
>  		mlog(ML_ERROR, "unable to get global_bitmap inode\n");
>  		ret = -EIO;
> -		goto out;
> +		goto out_unlock_tl_inode;
>  	}
>  
>  	inode_lock(gb_inode);
> @@ -634,16 +636,14 @@ static int ocfs2_move_extent(struct ocfs2_move_extents_context *context,
>  	ret = ocfs2_inode_lock(gb_inode, &gb_bh, 1);
>  	if (ret) {
>  		mlog_errno(ret);
> -		goto out_unlock_gb_mutex;
> +		goto out_unlock_gb_inode;
>  	}
>  
> -	inode_lock(tl_inode);
> -
>  	handle = ocfs2_start_trans(osb, credits);
>  	if (IS_ERR(handle)) {
>  		ret = PTR_ERR(handle);
>  		mlog_errno(ret);
> -		goto out_unlock_tl_inode;
> +		goto out_unlock;
>  	}
>  
>  	new_phys_blkno = ocfs2_clusters_to_blocks(inode->i_sb, *new_phys_cpos);
> @@ -703,13 +703,12 @@ static int ocfs2_move_extent(struct ocfs2_move_extents_context *context,
>  out_commit:
>  	ocfs2_commit_trans(osb, handle);
>  	brelse(gd_bh);
> -
> -out_unlock_tl_inode:
> -	inode_unlock(tl_inode);
> -
> +out_unlock:
>  	ocfs2_inode_unlock(gb_inode, 1);
> -out_unlock_gb_mutex:
> +out_unlock_gb_inode:
>  	inode_unlock(gb_inode);
> +out_unlock_tl_inode:
> +	inode_unlock(tl_inode);

Ummm... I'd rather keep gb_inode cleanup together.

Thanks,
Joseph

>  	brelse(gb_bh);
>  	iput(gb_inode);
>