Re: [PATCH v3] xfs: don't hold buffer locks across sync transaction commit in xfs_sync_sb_buf

["Zhou, Yun" <yun.zhou@windriver.com>]

On 9/12/26 01:42, Darrick J. Wong wrote:
> CAUTION: This email comes from a non Wind River email account!
> Do not click links or open attachments unless you recognize the sender and know the content is safe.
> 
> On Fri, Jul 31, 2026 at 11:10:10AM +0800, Yun Zhou wrote:
>> xfs_sync_sb_buf() holds sb/rtsb buffer locks across a synchronous
>> xfs_trans_commit(), which flushes the CIL push workqueue internally.
>> If shutdown occurs during the CIL push, xfs_buf_item_unpin() needs to
>> lock these buffers to fail them, causing a deadlock:
>>
>>    setlabel:          holds buf lock -> flush_workqueue(xfs-cil)
>>    CIL push worker:   xfs_buf_item_unpin -> xfs_buf_lock(same buf)
>>
>> Remove the xfs_trans_bhold() calls so that commit releases the buffer
>> locks normally.  After the sync commit, re-acquire the buffers via
>> mp->m_sb_bp / mp->m_rtsb_bp for the on-disk writeback.
>>
>> Fixes: f7664b31975b ("xfs: implement online get/set fs label")
>> Reported-by: syzbot+837bcd54843dd6262f2f@syzkaller.appspotmail.com
>> Closes: https://syzkaller.appspot.com/bug?extid=837bcd54843dd6262f2f
>> Cc: stable@vger.kernel.org
>> Signed-off-by: Yun Zhou <yun.zhou@windriver.com>
>> Reviewed-by: Christoph Hellwig <hch@lst.de>
>> ---
>> v3:
>>   - Drop xfs_buf_hold/xfs_buf_relse, use lock/unlock directly (Christoph).
>>
>> v2:
>>   - Remove the bp variable and pass xfs_trans_getsb(tp) directly to
>>     xfs_log_rtsb() to fix compilation warnings when CONFIG_XFS_RT=n.
>>   - Convert xfs_log_rtsb() stub from macro to inline function to avoid
>>     the need for (void) casting (Christoph).
>> ---
>>   fs/xfs/libxfs/xfs_rtgroup.h |  6 +++++-
>>   fs/xfs/libxfs/xfs_sb.c      | 37 +++++++++++++++++--------------------
>>   2 files changed, 22 insertions(+), 21 deletions(-)
>>
>> diff --git a/fs/xfs/libxfs/xfs_rtgroup.h b/fs/xfs/libxfs/xfs_rtgroup.h
>> index c0b9f9f2c413..fca2eb74908c 100644
>> --- a/fs/xfs/libxfs/xfs_rtgroup.h
>> +++ b/fs/xfs/libxfs/xfs_rtgroup.h
>> @@ -359,7 +359,11 @@ static inline int xfs_initialize_rtgroups(struct xfs_mount *mp,
>>   # define xfs_rtgroup_unlock(rtg, gf)         ((void)0)
>>   # define xfs_rtgroup_trans_join(tp, rtg, gf) ((void)0)
>>   # define xfs_update_rtsb(bp, sb_bp)  ((void)0)
>> -# define xfs_log_rtsb(tp, sb_bp)     (NULL)
>> +static inline struct xfs_buf *xfs_log_rtsb(struct xfs_trans *tp,
>> +             const struct xfs_buf *sb_bp)
>> +{
>> +     return NULL;
>> +}
>>   # define xfs_rtgroup_get_geometry(rtg, rgeo) (-EOPNOTSUPP)
>>   #endif /* CONFIG_XFS_RT */
>>
>> diff --git a/fs/xfs/libxfs/xfs_sb.c b/fs/xfs/libxfs/xfs_sb.c
>> index 47322adb7690..2664728f2d8f 100644
>> --- a/fs/xfs/libxfs/xfs_sb.c
>> +++ b/fs/xfs/libxfs/xfs_sb.c
>> @@ -1470,36 +1470,33 @@ xfs_sync_sb_buf(
>>        bool                    update_rtsb)
>>   {
>>        struct xfs_trans        *tp;
>> -     struct xfs_buf          *bp;
>> -     struct xfs_buf          *rtsb_bp = NULL;
>>        int                     error;
>>
>>        error = xfs_trans_alloc(mp, &M_RES(mp)->tr_sb, 0, 0, 0, &tp);
>>        if (error)
>>                return error;
>>
>> -     bp = xfs_trans_getsb(tp);
>>        xfs_log_sb(tp);
>> -     xfs_trans_bhold(tp, bp);
>> -     if (update_rtsb) {
>> -             rtsb_bp = xfs_log_rtsb(tp, bp);
>> -             if (rtsb_bp)
>> -                     xfs_trans_bhold(tp, rtsb_bp);
>> -     }
>> +     if (update_rtsb)
>> +             xfs_log_rtsb(tp, xfs_trans_getsb(tp));
>>        xfs_trans_set_sync(tp);
>>        error = xfs_trans_commit(tp);
>>        if (error)
>> -             goto out;
>> -     /*
>> -      * write out the sb buffer to get the changes to disk
>> -      */
>> -     error = xfs_bwrite(bp);
>> -     if (!error && rtsb_bp)
>> -             error = xfs_bwrite(rtsb_bp);
>> -out:
>> -     if (rtsb_bp)
>> -             xfs_buf_relse(rtsb_bp);
>> -     xfs_buf_relse(bp);
>> +             return error;
>> +
>> +     /* Re-acquire and write the sb and rtsb to disk. */
>> +     xfs_buf_lock(mp->m_sb_bp);
> 
> FYI, this causes porting problems to xfsprogs' libxfs because userspace
> doesn't have m_sb_bp or m_rtsb_bp pointers in struct xfs_mount.  It's
> not a big deal to add them, but if anyone ever wants to call
> xfs_sync_sb_buf then we're going to have to figure out when to populate
> those pointers.
> 

You're right — the problem is referencing mp->m_sb_bp / mp->m_rtsb_bp
directly in libxfs. I can avoid that by going back to acquiring the
buffer early via xfs_trans_getsb() / xfs_log_rtsb() (as the original
code did), instead of touching the mount pointers.

To keep this safe for a userspace caller that may not hold a
mount-level reference to the sb buffer, we can take an explicit
xfs_buf_hold() after getsb so the buffer stays alive across the
commit. Unlike the original xfs_trans_bhold(), xfs_buf_hold() only
takes a reference and does not keep the buffer locked across the
commit, so it doesn't reintroduce the deadlock.

Alternatively, since xfs_ioc_setlabel() is currently the only caller
and it's kernel-only, we could just move xfs_sync_sb_buf() out of
libxfs into xfs_ioctl.c.

Since this is already in for-next, would you prefer an incremental 
follow-up patch, or should I resend as v4?

Thanks,
Yun