On Wed, 2020-01-22 at 14:20 +0000, Robert Milkowski wrote:
> > -----Original Message-----
> > From: Trond Myklebust <trondmy@xxxxxxxxxxxxxxx>
> > Sent: 17 January 2020 17:24
> > To: linux-nfs@xxxxxxxxxxxxxxx; rmilkowski@xxxxxxxxx
> > Cc: anna.schumaker@xxxxxxxxxx; linux-kernel@xxxxxxxxxxxxxxx;
> > chuck.lever@xxxxxxxxxx
> > Subject: Re: [PATCH v2] NFSv4: try lease recovery on
> > NFS4ERR_EXPIRED
> >
> > On Fri, 2020-01-17 at 16:12 +0000, Robert Milkowski wrote:
> > > Anyone please?
> > >
> > >
> > > -----Original Message-----
> > > From: Robert Milkowski <rmilkowski@xxxxxxxxx>
> > > Sent: 08 January 2020 21:48
> > > To: linux-nfs@xxxxxxxxxxxxxxx
> > > Cc: 'Trond Myklebust' <trondmy@xxxxxxxxxxxxxxx>; 'Chuck Lever'
> > > <chuck.lever@xxxxxxxxxx>; 'Anna Schumaker' <
> > > anna.schumaker@xxxxxxxxxx
> > > > ;
> > > linux-kernel@xxxxxxxxxxxxxxx
> > > Subject: [PATCH v2] NFSv4: try lease recovery on NFS4ERR_EXPIRED
> > >
> > > From: Robert Milkowski <rmilkowski@xxxxxxxxx>
> > >
> > > Currently, if an nfs server returns NFS4ERR_EXPIRED to open(),
> > > etc.
> > > we return EIO to applications without even trying to recover.
> > >
> > > Fixes: 272289a3df72 ("NFSv4: nfs4_do_handle_exception() handle
> > > revoke/expiry of a single stateid")
> > > Signed-off-by: Robert Milkowski <rmilkowski@xxxxxxxxx>
> > > ---
> > > fs/nfs/nfs4proc.c | 4 ++++
> > > 1 file changed, 4 insertions(+)
> > >
> > > diff --git a/fs/nfs/nfs4proc.c b/fs/nfs/nfs4proc.c index
> > > 76d3716..2478405
> > > 100644
> > > --- a/fs/nfs/nfs4proc.c
> > > +++ b/fs/nfs/nfs4proc.c
> > > @@ -481,6 +481,10 @@ static int nfs4_do_handle_exception(struct
> > > nfs_server *server,
> > > stateid);
> > > goto wait_on_recovery;
> > > }
> > > + if (state == NULL) {
> > > + nfs4_schedule_lease_recovery(clp);
> > > + goto wait_on_recovery;
> > > + }
> > > /* Fall through */
> > > case -NFS4ERR_OPENMODE:
> > > if (inode) {
> > > --
> > > 1.8.3.1
> > >
> > >
> >
> > Does this apply to any case other than NFS4ERR_EXPIRED in the
> > specific
> > case of nfs4_do_open()? I can't see that it does. It looks to me as
> > if
> > the open recovery routines already have their own handling of this
> > case.
>
> I only observed the issue with open(). After further
> review I think you are right and it only applies to nfs4_do_open().
>
>
> > If so, why not just add it as a special case in the nfs4_do_open()
> > error
> > handling? Otherwise this patch will end up overriding other generic
> > cases where we have an inode, but no open state.
> >
>
> Fair point.
> So perhaps, few lines further instead of:
>
> if (inode) {
> ...
> if (state == NULL) {
> break;
> }
>
> There should be:
>
> if (inode) {
> ...
> if (state == NULL) {
> nfs4_schedule_lease_recovery(clp);
> goto wait_on_recovery;
> }
>
>
>
> This way we know that inode cannot be null at this point, and it's a
> case where both inode and state are NULL.
> This would be a little bit more general in case we reach this point.
>
> But if you think it is better to move it to nfs4_do_open() then I've
> just tested the following patch:
>
> diff --git a/fs/nfs/nfs4proc.c b/fs/nfs/nfs4proc.c
> index 76d3716..b7c4044 100644
> --- a/fs/nfs/nfs4proc.c
> +++ b/fs/nfs/nfs4proc.c
> @@ -3187,6 +3187,11 @@ static struct nfs4_state *nfs4_do_open(struct
> inode *dir,
> exception.retry = 1;
> continue;
> }
> + if (status == -NFS4ERR_EXPIRED) {
> + nfs4_schedule_lease_recovery(server-
> >nfs_client);
> + exception.retry = 1;
> + continue;
> + }
> if (status == -EAGAIN) {
> /* We must have found a delegation */
> exception.retry = 1;
>
This looks like what I'm asking for, yes. That seems like the minimal
patch that addresses the problem you're describing.
--
Trond Myklebust
Linux NFS client maintainer, Hammerspace
trond.myklebust@xxxxxxxxxxxxxxx
