[PATCH] ocfs2: fix lock acquisition order in refcounttree

[PATCH] ocfs2: fix lock acquisition order in refcounttree

[Diogo Jahchan Koike]

Acquiring the locks in refcounttree should follow
the ip_alloc --> ip_xattr ordering, as done by multiple
code paths in ocfs2; otherwise, we risk an ABBA deadlock
(i.e in the start transaction path).

Reported-by: syzbot+1fed2de07d8e11a3ec1b@syzkaller.appspotmail.com
Closes: https://syzkaller.appspot.com/bug?extid=1fed2de07d8e11a3ec1b
Tested-by: syzbot+1fed2de07d8e11a3ec1b@syzkaller.appspotmail.com
Reviewed-by: Tetsuo Handa <penguin-kernel@i-love.sakura.ne.jp>
Signed-off-by: Diogo Jahchan Koike <djahchankoike@gmail.com>
---
 fs/ocfs2/refcounttree.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/fs/ocfs2/refcounttree.c b/fs/ocfs2/refcounttree.c
index 8f732742b26e..c8467b92b64e 100644
--- a/fs/ocfs2/refcounttree.c
+++ b/fs/ocfs2/refcounttree.c
@@ -928,8 +928,8 @@ int ocfs2_try_remove_refcount_tree(struct inode *inode,
 	struct ocfs2_inode_info *oi = OCFS2_I(inode);
 	struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
 
-	down_write(&oi->ip_xattr_sem);
 	down_write(&oi->ip_alloc_sem);
+	down_write(&oi->ip_xattr_sem);
 
 	if (oi->ip_clusters)
 		goto out;
@@ -945,8 +945,8 @@ int ocfs2_try_remove_refcount_tree(struct inode *inode,
 	if (ret)
 		mlog_errno(ret);
 out:
-	up_write(&oi->ip_alloc_sem);
 	up_write(&oi->ip_xattr_sem);
+	up_write(&oi->ip_alloc_sem);
 	return 0;
 }
 
-- 
2.43.0

Re: [PATCH] ocfs2: fix lock acquisition order in refcounttree

[Tetsuo Handa]

On 2025/07/15 7:15, Diogo Jahchan Koike wrote:
> Acquiring the locks in refcounttree should follow
> the ip_alloc --> ip_xattr ordering, as done by multiple
> code paths in ocfs2; otherwise, we risk an ABBA deadlock
> (i.e in the start transaction path).

I noticed that ocfs2_reflink() in the same file wants similar change.

	down_write(&OCFS2_I(inode)->ip_xattr_sem);
	down_write(&OCFS2_I(inode)->ip_alloc_sem);
	error = __ocfs2_reflink(old_dentry, old_bh,
				new_orphan_inode, preserve);
	up_write(&OCFS2_I(inode)->ip_alloc_sem);
	up_write(&OCFS2_I(inode)->ip_xattr_sem);

Re: [PATCH] ocfs2: fix lock acquisition order in refcounttree

[Tetsuo Handa]

On 2025/07/15 11:51, Tetsuo Handa wrote:
> On 2025/07/15 7:15, Diogo Jahchan Koike wrote:
>> Acquiring the locks in refcounttree should follow
>> the ip_alloc --> ip_xattr ordering, as done by multiple
>> code paths in ocfs2; otherwise, we risk an ABBA deadlock
>> (i.e in the start transaction path).
> 
> I noticed that ocfs2_reflink() in the same file wants similar change.
> 
> 	down_write(&OCFS2_I(inode)->ip_xattr_sem);
> 	down_write(&OCFS2_I(inode)->ip_alloc_sem);
> 	error = __ocfs2_reflink(old_dentry, old_bh,
> 				new_orphan_inode, preserve);
> 	up_write(&OCFS2_I(inode)->ip_alloc_sem);
> 	up_write(&OCFS2_I(inode)->ip_xattr_sem);
> 

Moreover, I noticed that e.g. ocfs2_xattr_set_handle() firstly acquires
ip_xatr_sem and then ocfs2_xattr_ibody_find() might acquire ip_alloc_sem.

Diogo, where do you see the ip_alloc --> ip_xattr ordering?

Unless we unify to either ip_alloc --> ip_xattr ordering or
ip_xattr --> ip_alloc ordering (or replace ip_xattr with ip_alloc),
this patch simply changes the location of lockdep warning?

Re: [PATCH] ocfs2: fix lock acquisition order in refcounttree

[Tetsuo Handa]

On 2025/07/18 22:54, Tetsuo Handa wrote:
> On 2025/07/15 11:51, Tetsuo Handa wrote:
>> On 2025/07/15 7:15, Diogo Jahchan Koike wrote:
>>> Acquiring the locks in refcounttree should follow
>>> the ip_alloc --> ip_xattr ordering, as done by multiple
>>> code paths in ocfs2; otherwise, we risk an ABBA deadlock
>>> (i.e in the start transaction path).
>>
>> I noticed that ocfs2_reflink() in the same file wants similar change.
>>
>> 	down_write(&OCFS2_I(inode)->ip_xattr_sem);
>> 	down_write(&OCFS2_I(inode)->ip_alloc_sem);
>> 	error = __ocfs2_reflink(old_dentry, old_bh,
>> 				new_orphan_inode, preserve);
>> 	up_write(&OCFS2_I(inode)->ip_alloc_sem);
>> 	up_write(&OCFS2_I(inode)->ip_xattr_sem);
>>
> 
> Moreover, I noticed that e.g. ocfs2_xattr_set_handle() firstly acquires
> ip_xatr_sem and then ocfs2_xattr_ibody_find() might acquire ip_alloc_sem.
> 
> Diogo, where do you see the ip_alloc --> ip_xattr ordering?
> 
> Unless we unify to either ip_alloc --> ip_xattr ordering or
> ip_xattr --> ip_alloc ordering (or replace ip_xattr with ip_alloc),
> this patch simply changes the location of lockdep warning?
> 

Since I couldn't find direct ip_alloc --> ip_xattr ordering,
I tried effectively replacing ip_xattr with ip_alloc at
https://lkml.kernel.org/r/687be24a.a70a0220.693ce.0092.GAE@google.com
and got sb_internal --> ip_alloc v.s. ip_alloc --> sb_internal
ordering problem, as with other lockdep reports in ocfs2 subsystem
at https://syzkaller.appspot.com/upstream/s/ocfs2 .

In the team network driver, this kind of ordering issues has been
addressed by simplifying locking dependency at
https://lkml.kernel.org/r/20250623153147.3413631-1-sdf@fomichev.me .

ocfs2 developers, can you simplify locking dependency in ocfs2, by
killing several locks and reordering the locks?

Re: [PATCH] ocfs2: fix lock acquisition order in refcounttree

[Andrew Morton]

On Mon, 14 Jul 2025 19:15:32 -0300 Diogo Jahchan Koike <djahchankoike@gmail.com> wrote:

> Acquiring the locks in refcounttree should follow
> the ip_alloc --> ip_xattr ordering, as done by multiple
> code paths in ocfs2; otherwise, we risk an ABBA deadlock
> (i.e in the start transaction path).
> 
> Reported-by: syzbot+1fed2de07d8e11a3ec1b@syzkaller.appspotmail.com
> Closes: https://syzkaller.appspot.com/bug?extid=1fed2de07d8e11a3ec1b

But ocfs2_reflink() does

	down_write(&OCFS2_I(inode)->ip_xattr_sem);
	down_write(&OCFS2_I(inode)->ip_alloc_sem);

Re: [PATCH] ocfs2: fix lock acquisition order in refcounttree

[Tetsuo Handa]

On 2026/04/26 6:26, Andrew Morton wrote:
> But ocfs2_reflink() does
> 
> 	down_write(&OCFS2_I(inode)->ip_xattr_sem);
> 	down_write(&OCFS2_I(inode)->ip_alloc_sem);
> 
> 

Yes, we know. We are waiting for response from OCFS2 people on the below message.

-------- Forwarded Message --------
Message-ID: <dee1a0fe-7846-466a-a08b-4967929b5f6e@I-love.SAKURA.ne.jp>
Date: Mon, 21 Jul 2025 19:40:58 +0900
Subject: Re: [PATCH] ocfs2: fix lock acquisition order in refcounttree
From: Tetsuo Handa <penguin-kernel@I-love.SAKURA.ne.jp>
To: Diogo Jahchan Koike <djahchankoike@gmail.com>, Mark Fasheh <mark@fasheh.com>, Joel Becker <jlbec@evilplan.org>, Joseph Qi <joseph.qi@linux.alibaba.com>, ocfs2-devel@lists.linux.dev
Cc: syzbot+1fed2de07d8e11a3ec1b@syzkaller.appspotmail.com, linux-kernel@vger.kernel.org
References: <20250714221552.4853-1-djahchankoike@gmail.com> <5957f94d-2521-43d4-ba0b-7b98576167a4@I-love.SAKURA.ne.jp> <6e49d04a-4069-4cce-9f49-fd63983ae658@I-love.SAKURA.ne.jp>
In-Reply-To: <6e49d04a-4069-4cce-9f49-fd63983ae658@I-love.SAKURA.ne.jp>

On 2025/07/18 22:54, Tetsuo Handa wrote:
> On 2025/07/15 11:51, Tetsuo Handa wrote:
>> On 2025/07/15 7:15, Diogo Jahchan Koike wrote:
>>> Acquiring the locks in refcounttree should follow
>>> the ip_alloc --> ip_xattr ordering, as done by multiple
>>> code paths in ocfs2; otherwise, we risk an ABBA deadlock
>>> (i.e in the start transaction path).
>>
>> I noticed that ocfs2_reflink() in the same file wants similar change.
>>
>> 	down_write(&OCFS2_I(inode)->ip_xattr_sem);
>> 	down_write(&OCFS2_I(inode)->ip_alloc_sem);
>> 	error = __ocfs2_reflink(old_dentry, old_bh,
>> 				new_orphan_inode, preserve);
>> 	up_write(&OCFS2_I(inode)->ip_alloc_sem);
>> 	up_write(&OCFS2_I(inode)->ip_xattr_sem);
>>
> 
> Moreover, I noticed that e.g. ocfs2_xattr_set_handle() firstly acquires
> ip_xatr_sem and then ocfs2_xattr_ibody_find() might acquire ip_alloc_sem.
> 
> Diogo, where do you see the ip_alloc --> ip_xattr ordering?
> 
> Unless we unify to either ip_alloc --> ip_xattr ordering or
> ip_xattr --> ip_alloc ordering (or replace ip_xattr with ip_alloc),
> this patch simply changes the location of lockdep warning?
> 

Since I couldn't find direct ip_alloc --> ip_xattr ordering,
I tried effectively replacing ip_xattr with ip_alloc at
https://lkml.kernel.org/r/687be24a.a70a0220.693ce.0092.GAE@google.com
and got sb_internal --> ip_alloc v.s. ip_alloc --> sb_internal
ordering problem, as with other lockdep reports in ocfs2 subsystem
at https://syzkaller.appspot.com/upstream/s/ocfs2 .

In the team network driver, this kind of ordering issues has been
addressed by simplifying locking dependency at
https://lkml.kernel.org/r/20250623153147.3413631-1-sdf@fomichev.me .

ocfs2 developers, can you simplify locking dependency in ocfs2, by
killing several locks and reordering the locks?