Message-ID: <672c7858f622e78973296bc0a504114c7610b761.camel@gmail.com>
Subject: Re: [PATCH 1/4] pNFS: Ensure we return the error if someone kills a
 waiting layoutget
From:   Anna Schumaker <schumaker.anna@gmail.com>
To:     Trond Myklebust <trondmy@hammerspace.com>
Cc:     "linux-nfs@vger.kernel.org" <linux-nfs@vger.kernel.org>
Date:   Tue, 12 Mar 2019 17:22:50 -0400
In-Reply-To: <6057b85eea9d1b25189e2c1ead89112c91180ff0.camel@hammerspace.com>
References: <20180905180715.99485-1-trond.myklebust@hammerspace.com>
         <233f58676bee06fe87fae6c6ca04708a24716187.camel@netapp.com>
         <6057b85eea9d1b25189e2c1ead89112c91180ff0.camel@hammerspace.com>
Content-Type: text/plain; charset="UTF-8"
User-Agent: Evolution 3.30.5 
MIME-Version: 1.0
Content-Transfer-Encoding: 7bit
Sender: linux-nfs-owner@vger.kernel.org
Precedence: bulk

On Tue, 2019-03-12 at 20:12 +0000, Trond Myklebust wrote:
> On Tue, 2019-03-12 at 20:04 +0000, Schumaker, Anna wrote:
> > Hi Trond,
> > 
> > I'm seeing a hang when testing xfstests generic/013 on v4.1 with pNFS
> > after this
> > patch:
> > 
> > On Wed, 2018-09-05 at 14:07 -0400, Trond Myklebust wrote:
> > > If someone interrupts a wait on one or more outstanding layoutgets
> > > in
> > > pnfs_update_layout() then return the ERESTARTSYS/EINTR error.
> > > 
> > > Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>
> > > ---
> > >  fs/nfs/pnfs.c | 26 ++++++++++++++++----------
> > >  1 file changed, 16 insertions(+), 10 deletions(-)
> > > 
> > > diff --git a/fs/nfs/pnfs.c b/fs/nfs/pnfs.c
> > > index e8f232de484f..7d9a51e6b847 100644
> > > --- a/fs/nfs/pnfs.c
> > > +++ b/fs/nfs/pnfs.c
> > > @@ -1740,16 +1740,16 @@ static bool pnfs_within_mdsthreshold(struct
> > > nfs_open_context *ctx,
> > >  	return ret;
> > >  }
> > >  
> > > -static bool pnfs_prepare_to_retry_layoutget(struct pnfs_layout_hdr
> > > *lo)
> > > +static int pnfs_prepare_to_retry_layoutget(struct pnfs_layout_hdr
> > > *lo)
> > >  {
> > >  	/*
> > >  	 * send layoutcommit as it can hold up layoutreturn due to lseg
> > >  	 * reference
> > >  	 */
> > >  	pnfs_layoutcommit_inode(lo->plh_inode, false);
> > > -	return !wait_on_bit_action(&lo->plh_flags, NFS_LAYOUT_RETURN,
> > > +	return wait_on_bit_action(&lo->plh_flags, NFS_LAYOUT_RETURN,
> > >  				   nfs_wait_bit_killable,
> > > -				   TASK_UNINTERRUPTIBLE);
> > > +				   TASK_KILLABLE);
> > >  }
> > >  
> > >  static void nfs_layoutget_begin(struct pnfs_layout_hdr *lo)
> > > @@ -1830,7 +1830,9 @@ pnfs_update_layout(struct inode *ino,
> > >  	}
> > >  
> > >  lookup_again:
> > > -	nfs4_client_recover_expired_lease(clp);
> > > +	lseg = ERR_PTR(nfs4_client_recover_expired_lease(clp));
> > > +	if (IS_ERR(lseg))
> > > +		goto out;
> > >  	first = false;
> > >  	spin_lock(&ino->i_lock);
> > >  	lo = pnfs_find_alloc_layout(ino, ctx, gfp_flags);
> > > @@ -1863,9 +1865,9 @@ pnfs_update_layout(struct inode *ino,
> > >  	if (list_empty(&lo->plh_segs) &&
> > >  	    atomic_read(&lo->plh_outstanding) != 0) {
> > >  		spin_unlock(&ino->i_lock);
> > > -		if (wait_var_event_killable(&lo->plh_outstanding,
> > > -					atomic_read(&lo-
> > > > plh_outstanding) == 0
> > > -					|| !list_empty(&lo->plh_segs)))
> > > +		lseg = ERR_PTR(wait_var_event_killable(&lo-
> > > > plh_outstanding,
> > > +					atomic_read(&lo-
> > > > plh_outstanding)));
> > > +		if (IS_ERR(lseg) || !list_empty(&lo->plh_segs))
> > 
> > Was dropping the "== 0" condition attached to the atomic_read() here
> > a mistake?
> > I think what's happening is that my client is waiting for
> > plh_outstanding to be
> > anything other than 0 when there isn't any work left to do.
> 
> Yes. That's a bug. How about the following patch?

This patch works for me, but for some reason doing "!atomic_read()" takes 8
minutes longer to complete compared to doing "atomic_read() == 0".  I have not
run this multiple times to confirm that it's always the case.

Anna

> 
> 8<---------------------------------------------------
> From 400417b05f3ec0531544ca5f94e64d838d8b8849 Mon Sep 17 00:00:00 2001
> From: Trond Myklebust <trond.myklebust@hammerspace.com>
> Date: Tue, 12 Mar 2019 16:04:51 -0400
> Subject: [PATCH] pNFS: Fix a typo in pnfs_update_layout
> 
> We're supposed to wait for the outstanding layout count to go to zero,
> but that got lost somehow.
> 
> Fixes: d03360aaf5cca ("pNFS: Ensure we return the error if someone...")
> Reported-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
> Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>
> ---
>  fs/nfs/pnfs.c | 2 +-
>  1 file changed, 1 insertion(+), 1 deletion(-)
> 
> diff --git a/fs/nfs/pnfs.c b/fs/nfs/pnfs.c
> index 8247bd1634cb..7066cd7c7aff 100644
> --- a/fs/nfs/pnfs.c
> +++ b/fs/nfs/pnfs.c
> @@ -1889,7 +1889,7 @@ pnfs_update_layout(struct inode *ino,
>  	    atomic_read(&lo->plh_outstanding) != 0) {
>  		spin_unlock(&ino->i_lock);
>  		lseg = ERR_PTR(wait_var_event_killable(&lo->plh_outstanding,
> -					atomic_read(&lo->plh_outstanding)));
> +					!atomic_read(&lo->plh_outstanding)));
>  		if (IS_ERR(lseg) || !list_empty(&lo->plh_segs))
>  			goto out_put_layout_hdr;
>  		pnfs_put_layout_hdr(lo);
> -- 
> 2.20.1
> 
> -- 
> Trond Myklebust
> Linux NFS client maintainer, Hammerspace
> trond.myklebust@hammerspace.com
> 
>