Re: [PATCH] ocfs2: fix circular locking dependency in ocfs2_acquire_dquot

[Joseph Qi]

On 2025/12/28 01:42, Szymon Wilczek wrote:
> Move ocfs2_extend_no_holes() to execute before ocfs2_lock_global_qf() to
> fix a circular locking dependency reported by syzbot.
> 
> The issue occurs because ocfs2_extend_no_holes() internally calls
> ocfs2_extend_allocation() which starts a transaction (acquiring
> sb_start_intwrite). When called while holding the global quota file
> lock, this conflicts with mount-time operations that acquire

It seems the following locking sequence is fine:
ocfs2_lock_global_qf -> start_trans

So could you please elaborate more?

> sb_internal first, creating the following circular dependency:
> 
>   sb_internal -> ocfs2_sysfile_lock_key -> ocfs2_quota_ip_alloc_sem_key
> 
> By moving the quota file extension before acquiring the global quota
> file lock, we ensure that any internal transactions complete before
> quota locks are held, breaking the circular dependency.
> 
> Reported-by: syzbot+51244a05705883616c95@xxxxxxxxxxxxxxxxxxxxxxxxx
> Tested-by: syzbot+51244a05705883616c95@xxxxxxxxxxxxxxxxxxxxxxxxx
> Closes: https://syzkaller.appspot.com/bug?extid=51244a05705883616c95
> Signed-off-by: Szymon Wilczek <swilczek.lx@xxxxxxxxx>
> ---
>  fs/ocfs2/quota_global.c | 26 ++++++++++++++------------
>  1 file changed, 14 insertions(+), 12 deletions(-)
> 
> diff --git a/fs/ocfs2/quota_global.c b/fs/ocfs2/quota_global.c
> index e85b1ccf81be..136aaaae27f3 100644
> --- a/fs/ocfs2/quota_global.c
> +++ b/fs/ocfs2/quota_global.c
> @@ -821,6 +821,19 @@ static int ocfs2_acquire_dquot(struct dquot *dquot)
>  	trace_ocfs2_acquire_dquot(from_kqid(&init_user_ns, dquot->dq_id),
>  				  type);
>  	mutex_lock(&dquot->dq_lock);
> +	/*
> +	 * Extend global quota file before acquiring global qf lock to avoid
> +	 * lock inversion with sb_internal (via ocfs2_start_trans).
> +	 */
> +	if (need_alloc) {
> +		WARN_ON(journal_current_handle());
> +		status = ocfs2_extend_no_holes(gqinode, NULL,
> +			i_size_read(gqinode) + (need_alloc << sb->s_blocksize_bits),
> +			i_size_read(gqinode));
> +		if (status < 0)
> +			goto out;
> +	}
> +
>  	/*
>  	 * We need an exclusive lock, because we're going to update use count
>  	 * and instantiate possibly new dquot structure
> @@ -843,19 +856,8 @@ static int ocfs2_acquire_dquot(struct dquot *dquot)
>  	OCFS2_DQUOT(dquot)->dq_use_count++;
>  	OCFS2_DQUOT(dquot)->dq_origspace = dquot->dq_dqb.dqb_curspace;
>  	OCFS2_DQUOT(dquot)->dq_originodes = dquot->dq_dqb.dqb_curinodes;
> -	if (!dquot->dq_off) {	/* No real quota entry? */
> +	if (!dquot->dq_off)	/* No real quota entry? */
>  		ex = 1;
> -		/*
> -		 * Add blocks to quota file before we start a transaction since
> -		 * locking allocators ranks above a transaction start
> -		 */
> -		WARN_ON(journal_current_handle());
> -		status = ocfs2_extend_no_holes(gqinode, NULL,
> -			i_size_read(gqinode) + (need_alloc << sb->s_blocksize_bits),
> -			i_size_read(gqinode));
> -		if (status < 0)
> -			goto out_dq;
> -	}
>  
>  	handle = ocfs2_start_trans(osb,
>  				   ocfs2_calc_global_qinit_credits(sb, type));

BTW, even if your analysis is right, here also calls ocfs2_start_trans().

Thanks,
Joseph