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

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: <[email protected]>
Date: Mon, 21 Jul 2025 19:40:58 +0900
Subject: Re: [PATCH] ocfs2: fix lock acquisition order in refcounttree
From: Tetsuo Handa <[email protected]>
To: Diogo Jahchan Koike <[email protected]>, Mark Fasheh <[email protected]>, Joel Becker <[email protected]>, Joseph Qi <[email protected]>, [email protected]
Cc: [email protected], [email protected]
References: <[email protected]> <[email protected]> <[email protected]>
In-Reply-To: <[email protected]>

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/[email protected]
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/[email protected] .

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