Re: [PATCH v2] xfs: don't nest transactions when scanning for eofblocks

[Allison Henderson <allison.henderson@oracle.com>]

On 2/19/21 10:23 AM, Darrick J. Wong wrote:
> From: Darrick J. Wong <djwong@kernel.org>
> 
> Brian Foster reported a lockdep warning on xfs/167:
> 
> ============================================
> WARNING: possible recursive locking detected
> 5.11.0-rc4 #35 Tainted: G        W I
> --------------------------------------------
> fsstress/17733 is trying to acquire lock:
> ffff8e0fd1d90650 (sb_internal){++++}-{0:0}, at: xfs_free_eofblocks+0x104/0x1d0 [xfs]
> 
> but task is already holding lock:
> ffff8e0fd1d90650 (sb_internal){++++}-{0:0}, at: xfs_trans_alloc_inode+0x5f/0x160 [xfs]
> 
> stack backtrace:
> CPU: 38 PID: 17733 Comm: fsstress Tainted: G        W I       5.11.0-rc4 #35
> Hardware name: Dell Inc. PowerEdge R740/01KPX8, BIOS 1.6.11 11/20/2018
> Call Trace:
>   dump_stack+0x8b/0xb0
>   __lock_acquire.cold+0x159/0x2ab
>   lock_acquire+0x116/0x370
>   xfs_trans_alloc+0x1ad/0x310 [xfs]
>   xfs_free_eofblocks+0x104/0x1d0 [xfs]
>   xfs_blockgc_scan_inode+0x24/0x60 [xfs]
>   xfs_inode_walk_ag+0x202/0x4b0 [xfs]
>   xfs_inode_walk+0x66/0xc0 [xfs]
>   xfs_trans_alloc+0x160/0x310 [xfs]
>   xfs_trans_alloc_inode+0x5f/0x160 [xfs]
>   xfs_alloc_file_space+0x105/0x300 [xfs]
>   xfs_file_fallocate+0x270/0x460 [xfs]
>   vfs_fallocate+0x14d/0x3d0
>   __x64_sys_fallocate+0x3e/0x70
>   do_syscall_64+0x33/0x40
>   entry_SYSCALL_64_after_hwframe+0x44/0xa9
> 
> The cause of this is the new code that spurs a scan to garbage collect
> speculative preallocations if we fail to reserve enough blocks while
> allocating a transaction.  While the warning itself is a fairly benign
> lockdep complaint, it does expose a potential livelock if the rwsem
> behavior ever changes with regards to nesting read locks when someone's
> waiting for a write lock.
> 
> Fix this by freeing the transaction and jumping back to xfs_trans_alloc
> like this patch in the V4 submission[1].
> 
> [1] https://urldefense.com/v3/__https://lore.kernel.org/linux-xfs/161142798066.2171939.9311024588681972086.stgit@magnolia/__;!!GqivPVa7Brio!ONLN-B-M-uqN8aJAN8zMDHDQZ6wwDyF4BSpjkT9j3mV2Zxe5zVD0vgjTWvPFRO2tzEpN$
> 
> Fixes: a1a7d05a0576 ("xfs: flush speculative space allocations when we run out of space")
> Reported-by: Brian Foster <bfoster@redhat.com>
> Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Ok, makes sense
Reviewed-by: Allison Henderson <allison.henderson@oracle.com>

> ---
> v2: fix commit message
> ---
>   fs/xfs/xfs_trans.c |   13 ++++++++++---
>   1 file changed, 10 insertions(+), 3 deletions(-)
> 
> diff --git a/fs/xfs/xfs_trans.c b/fs/xfs/xfs_trans.c
> index 44f72c09c203..377f3961d7ed 100644
> --- a/fs/xfs/xfs_trans.c
> +++ b/fs/xfs/xfs_trans.c
> @@ -260,6 +260,7 @@ xfs_trans_alloc(
>   	struct xfs_trans	**tpp)
>   {
>   	struct xfs_trans	*tp;
> +	bool			want_retry = true;
>   	int			error;
>   
>   	/*
> @@ -267,6 +268,7 @@ xfs_trans_alloc(
>   	 * GFP_NOFS allocation context so that we avoid lockdep false positives
>   	 * by doing GFP_KERNEL allocations inside sb_start_intwrite().
>   	 */
> +retry:
>   	tp = kmem_cache_zalloc(xfs_trans_zone, GFP_KERNEL | __GFP_NOFAIL);
>   	if (!(flags & XFS_TRANS_NO_WRITECOUNT))
>   		sb_start_intwrite(mp->m_super);
> @@ -289,7 +291,9 @@ xfs_trans_alloc(
>   	tp->t_firstblock = NULLFSBLOCK;
>   
>   	error = xfs_trans_reserve(tp, resp, blocks, rtextents);
> -	if (error == -ENOSPC) {
> +	if (error == -ENOSPC && want_retry) {
> +		xfs_trans_cancel(tp);
> +
>   		/*
>   		 * We weren't able to reserve enough space for the transaction.
>   		 * Flush the other speculative space allocations to free space.
> @@ -297,8 +301,11 @@ xfs_trans_alloc(
>   		 * other locks.
>   		 */
>   		error = xfs_blockgc_free_space(mp, NULL);
> -		if (!error)
> -			error = xfs_trans_reserve(tp, resp, blocks, rtextents);
> +		if (error)
> +			return error;
> +
> +		want_retry = false;
> +		goto retry;
>   	}
>   	if (error) {
>   		xfs_trans_cancel(tp);
>