Re: [PATCH] xfs: don't call into blockgc scan with freeze protection

[Dave Chinner <david@fromorbit.com>]

On Fri, Feb 19, 2021 at 08:09:32AM -0500, Brian Foster wrote:
> On Fri, Feb 19, 2021 at 03:56:58PM +1100, Dave Chinner wrote:
> > On Thu, Feb 18, 2021 at 07:23:09PM -0800, Darrick J. Wong wrote:
> > > On Thu, Feb 18, 2021 at 03:14:58PM -0500, Brian Foster wrote:
> > > > fstest xfs/167 produced a lockdep splat that complained about a
> > > > nested transaction acquiring freeze protection during an eofblocks
> > > > scan. Drop freeze protection around the block reclaim scan in the
> > > > transaction allocation code to avoid this problem.
> > > > 
> > > > Signed-off-by: Brian Foster <bfoster@redhat.com>
> ...
> > > >  fs/xfs/xfs_trans.c | 19 ++++++++++++++-----
> > > >  1 file changed, 14 insertions(+), 5 deletions(-)
> > > > 
> > > > diff --git a/fs/xfs/xfs_trans.c b/fs/xfs/xfs_trans.c
> > > > index 44f72c09c203..c32c62d3b77a 100644
> > > > --- a/fs/xfs/xfs_trans.c
> > > > +++ b/fs/xfs/xfs_trans.c
> ...
> > > > @@ -288,19 +289,27 @@ xfs_trans_alloc(
> > > >  	INIT_LIST_HEAD(&tp->t_dfops);
> > > >  	tp->t_firstblock = NULLFSBLOCK;
> > > >  
> > > > +retry:
> > > >  	error = xfs_trans_reserve(tp, resp, blocks, rtextents);
> > > > -	if (error == -ENOSPC) {
> > > > +	if (error == -ENOSPC && !retried) {
> > > >  		/*
> > > >  		 * We weren't able to reserve enough space for the transaction.
> > > >  		 * Flush the other speculative space allocations to free space.
> > > >  		 * Do not perform a synchronous scan because callers can hold
> > > >  		 * other locks.
> > > >  		 */
> > > > +		retried = true;
> > > > +		if (!(flags & XFS_TRANS_NO_WRITECOUNT))
> > > > +			sb_end_intwrite(mp->m_super);
> > > >  		error = xfs_blockgc_free_space(mp, NULL);
> > > > -		if (!error)
> > > > -			error = xfs_trans_reserve(tp, resp, blocks, rtextents);
> > > > -	}
> > > > -	if (error) {
> > > > +		if (error) {
> > > > +			kmem_cache_free(xfs_trans_zone, tp);
> > > > +			return error;
> > > > +		}
> > 
> > This seems dangerous to me. If xfs_trans_reserve() adds anything to
> > the transaction even if it fails, this will fail to free it. e.g.
> > xfs_log_reserve() call allocate a ticket and attach it to the
> > transaction and *then fail*. This code will now leak that ticket.
> > 
> 
> xfs_trans_reserve() ungrants the log ticket (which frees it, at least in
> the allocation case) and disassociates from the transaction on error, so
> I don't see how this causes any problems.

It ungrants the log ticket when it jumps to "undo_log" on error.
When xfs_log_reserve() fails, it jumps to "undo_blocks" and doesn't
ungrant the ticket. Hence potentially leaving an allocated ticket
attached to the transaction on error.  xfs_trans_cancel() handles
this just fine, just freeing the transaction doesn't.

Cheers,

Dave.
-- 
Dave Chinner
david@fromorbit.com