Received-SPF: pass (google.com: domain of linux-nfs+bounces-1601-linux.lists.archive=gmail.com@vger.kernel.org designates 147.75.199.223 as permitted sender) client-ip=147.75.199.223;
Precedence: bulk
MIME-Version: 1.0
References: <20240129154750.1245317-1-dwysocha@redhat.com> <1526807.1706548521@warthog.procyon.org.uk>
 <CALF+zOm9QBkf3PbchuVzkgYW1aEJWFFQ5JF0Y-_6BbPJke8CHw@mail.gmail.com> <dd504acbdfb0b926e54318fd6ea8f9d43af73634.camel@kernel.org>
In-Reply-To: <dd504acbdfb0b926e54318fd6ea8f9d43af73634.camel@kernel.org>
From: David Wysochanski <dwysocha@redhat.com>
Date: Tue, 30 Jan 2024 09:56:06 -0500
Message-ID: <CALF+zOnR1Hu-M=N7+ALcNicbVvEO=G5XN0roigxps15Wj0O8uA@mail.gmail.com>
Subject: Re: [PATCH] NFS: Fix nfs_netfs_issue_read() xarray locking for
 writeback interrupt
To: Jeff Layton <jlayton@kernel.org>
Cc: David Howells <dhowells@redhat.com>, Anna Schumaker <anna.schumaker@netapp.com>, 
	Trond Myklebust <trond.myklebust@hammerspace.com>, linux-nfs@vger.kernel.org, 
	linux-cachefs@redhat.com
Content-Type: text/plain; charset="UTF-8"
Content-Transfer-Encoding: quoted-printable

On Mon, Jan 29, 2024 at 12:44=E2=80=AFPM Jeff Layton <jlayton@kernel.org> w=
rote:
>
> On Mon, 2024-01-29 at 12:34 -0500, David Wysochanski wrote:
> > On Mon, Jan 29, 2024 at 12:15=E2=80=AFPM David Howells <dhowells@redhat=
.com> wrote:
> > >
> > > Dave Wysochanski <dwysocha@redhat.com> wrote:
> > >
> > > > -     xas_lock(&xas);
> > > > +     xas_lock_irqsave(&xas, flags);
> > > >       xas_for_each(&xas, page, last) {
> > >
> > > You probably want to use RCU, not xas_lock().  The pages are locked a=
nd so
> > > cannot be evicted from the xarray.
> > >
> >
> > I tried RCU originally and ran into a problem because NFS can schedule
> > (see comment on line 328 below)
> >
> > 326         xas_lock_irqsave(&xas, flags);
> > 327         xas_for_each(&xas, page, last) {
> > 328                 /* nfs_read_add_folio() may schedule() due to pNFS
> > layout and other RPCs  */
> > 329                 xas_pause(&xas);
> > 330                 xas_unlock_irqrestore(&xas, flags);
> > 331                 err =3D nfs_read_add_folio(&pgio, ctx, page_folio(p=
age));
> > 332                 if (err < 0) {
> > 333                         netfs->error =3D err;
> > 334                         goto out;
> > 335                 }
> > 336                 xas_lock_irqsave(&xas, flags);
> > 337         }
> > 338         xas_unlock_irqrestore(&xas, flags);
> >
>
> Looking at it more closely, I think you might want to just use
> xa_for_each_start(). That will do the traversal using the rcu_read_lock
> under the hood, and you should be able to block on every iteration.
>
Thanks Jeff.  Yes, I agree after looking at this further, this is a
good approach, and much cleaner.  I'll work on a v2 patch (actually
with xa_for_each_range as you suggested off list) and send after
a bit of testing -- so far, so good.

FWIW, my original usage of RCU was outside the whole loop.
I ran into problems due to nfs_read_add_folio().