Re: [PATCH 10/11] xfs: make xfs_*read_agf return EAGAIN to ALLOC_FLAG_TRYLOCK callers

From: "Darrick J. Wong" <darrick.wong@oracle.com>
To: Christoph Hellwig <hch@infradead.org>
Cc: linux-xfs@vger.kernel.org
Subject: Re: [PATCH 10/11] xfs: make xfs_*read_agf return EAGAIN to ALLOC_FLAG_TRYLOCK callers
Date: Fri, 17 Jan 2020 15:05:56 -0800	[thread overview]
Message-ID: <20200117230556.GV8247@magnolia> (raw)
In-Reply-To: <20200117065957.GC26438@infradead.org>

On Thu, Jan 16, 2020 at 10:59:57PM -0800, Christoph Hellwig wrote:
> On Thu, Jan 16, 2020 at 10:24:41PM -0800, Darrick J. Wong wrote:
> > From: Darrick J. Wong <darrick.wong@oracle.com>
> > 
> > Refactor xfs_read_agf and xfs_alloc_read_agf to return EAGAIN if the
> > caller passed TRYLOCK and we weren't able to get the lock; and change
> > the callers to recognize this.
> > 
> > Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
> > ---
> >  fs/xfs/libxfs/xfs_alloc.c |   31 +++++++++++++++----------------
> >  fs/xfs/libxfs/xfs_bmap.c  |    9 +++++----
> >  fs/xfs/xfs_filestream.c   |   11 ++++++-----
> >  3 files changed, 26 insertions(+), 25 deletions(-)
> > 
> > 
> > diff --git a/fs/xfs/libxfs/xfs_alloc.c b/fs/xfs/libxfs/xfs_alloc.c
> > index 83273975df77..26f3e4db84e0 100644
> > --- a/fs/xfs/libxfs/xfs_alloc.c
> > +++ b/fs/xfs/libxfs/xfs_alloc.c
> > @@ -2502,13 +2502,15 @@ xfs_alloc_fix_freelist(
> >  
> >  	if (!pag->pagf_init) {
> >  		error = xfs_alloc_read_agf(mp, tp, args->agno, flags, &agbp);
> > -		if (error)
> > -			goto out_no_agbp;
> > -		if (!pag->pagf_init) {
> > +		if (error == -EAGAIN) {
> > +			/* Couldn't lock the AGF so skip this AG. */
> >  			ASSERT(flags & XFS_ALLOC_FLAG_TRYLOCK);
> >  			ASSERT(!(flags & XFS_ALLOC_FLAG_FREEING));
> > -			goto out_agbp_relse;
> > +			error = 0;
> > +			goto out_no_agbp;
> >  		}
> > +		if (error)
> > +			goto out_no_agbp;
> 
> I wonder if something like:
> 
> 		if (error) {
> 			if (error == -EAGAIN) {
> 				/* Couldn't lock the AGF so skip this AG. */
> 	 			ASSERT(flags & XFS_ALLOC_FLAG_TRYLOCK);
> 	 			ASSERT(!(flags & XFS_ALLOC_FLAG_FREEING));
> 				error = 0;
> 			}
> 			goto out_no_agbp;
> 		}
> 
> would be a little nicer here?

> > @@ -2533,13 +2535,15 @@ xfs_alloc_fix_freelist(
> >  	 */
> >  	if (!agbp) {
> >  		error = xfs_alloc_read_agf(mp, tp, args->agno, flags, &agbp);
> > -		if (error)
> > -			goto out_no_agbp;
> > -		if (!agbp) {
> > +		if (error == -EAGAIN) {
> > +			/* Couldn't lock the AGF so skip this AG. */
> >  			ASSERT(flags & XFS_ALLOC_FLAG_TRYLOCK);
> >  			ASSERT(!(flags & XFS_ALLOC_FLAG_FREEING));
> > +			error = 0;
> >  			goto out_no_agbp;
> >  		}
> > +		if (error)
> > +			goto out_no_agbp;
> >  	}
> 
> Same here.  Also shouldn't those asserts just move into
> xfs_alloc_read_agf or go away now that we have a proper return value
> and not the magic NULL buffer?

I'll move the assert into xfs_alloc_read_agf, so these all turn into:

if (error) {
	if (error == -EAGAIN)
		error = 0;
	goto next ag;
}

> 
> > +	error = xfs_alloc_read_agf(mp, tp, agno, flags, &bp);
> > +	if (error)
> >  		return error;
> > +	xfs_trans_brelse(tp, bp);
> >  	return 0;
> 
> Maybe simplify this further to:
> 
> 	error = xfs_alloc_read_agf(mp, tp, agno, flags, &bp);
> 	if (!error)
> 		xfs_trans_brelse(tp, bp);
> 	return error;

Ok.

> > @@ -2958,12 +2962,9 @@ xfs_read_agf(
> >  	trace_xfs_read_agf(mp, agno);
> >  
> >  	ASSERT(agno != NULLAGNUMBER);
> > -	error = xfs_trans_read_buf(
> > -			mp, tp, mp->m_ddev_targp,
> > +	error = xfs_trans_read_buf(mp, tp, mp->m_ddev_targp,
> 
> This hunk should probably go into the patch that changed the
> xfs_trans_read_buf return value instead.

Ok.

> > +		if (error == -EAGAIN) {
> > +			/* Couldn't lock the AGF, so skip this AG. */
> >  			*notinit = 1;
> > +			error = 0;
> >  			goto out;
> >  		}
> > +		if (error)
> > +			goto out;
> 
> Should probably be:
> 
> 		if (error) {
> 			if (error == -EAGAIN) {
> 				/* Couldn't lock the AGF, so skip this AG. */
> 	 			*notinit = 1;
> 				error = 0;
> 			}
>  			goto out;

Fixed.

--D