[Linux-cachefs] [RFC PATCH v2 13/14] NFS: Call fscache_resize_cookie() when inode size changes due to setattr

[David Wysochanski]

On Thu, Jul 30, 2020 at 3:23 PM David Wysochanski <dwysocha at redhat.com> wrote:
>
> On Thu, Jul 30, 2020 at 2:39 PM Jeff Layton <jlayton at redhat.com> wrote:
> >
> > On Wed, 2020-07-29 at 10:12 -0400, Dave Wysochanski wrote:
> > > Handle truncate / setattr when fscache is enabled by calling
> > > fscache_resize_cookie().
> > >
> > > Signed-off-by: Dave Wysochanski <dwysocha at redhat.com>
> > > ---
> > >  fs/nfs/inode.c | 1 +
> > >  1 file changed, 1 insertion(+)
> > >
> > > diff --git a/fs/nfs/inode.c b/fs/nfs/inode.c
> > > index 45067303348c..6b814246d07d 100644
> > > --- a/fs/nfs/inode.c
> > > +++ b/fs/nfs/inode.c
> > > @@ -667,6 +667,7 @@ static int nfs_vmtruncate(struct inode * inode, loff_t offset)
> > >       spin_unlock(&inode->i_lock);
> > >       truncate_pagecache(inode, offset);
> > >       spin_lock(&inode->i_lock);
> > > +     fscache_resize_cookie(nfs_i_fscache(inode), i_size_read(inode));
> > >  out:
> > >       return err;
> > >  }
> >
> > truncate can happen even when you have no open file descriptors on the
> > file and therefore w/o the cookie being "used". In the ceph vmtruncate
> > handling code, I do an explicit use/unuse around this call. Do you need
> > to do the same here?
> > --
> > Jeff Layton <jlayton at redhat.com>
> >
>
> Actually I think the case you mention is covered by a patch that I've just
> added today on top of my v2 posting.
> This was the result of looking deeper into a few xfstest failures with
> NFSv4.2.  I think this covers the truncate without a file open:
>
> commit 91d6922df9390ca1c090911be6e5c5ab1a79ea83
> Author: Dave Wysochanski <dwysocha at redhat.com>
> Date:   Thu Jul 30 12:33:40 2020 -0400
>
>     NFS: Call fscache_invalidate() from nfs_invalidate_mapping()
>
>     Be sure to invalidate fscache cookie for any call to
>     nfs_invalidate_mapping().
>
>     This patch fixes the following xfstests on NFS4.x:
>       generic/240
>     as well as fixes the following xfstests on NFSv4.2:
>       generic/029 generic/030
>
>     Signed-off-by: Dave Wysochanski <dwysocha at redhat.com>
>
> diff --git a/fs/nfs/inode.c b/fs/nfs/inode.c
> index 6b814246d07d..62243ec05917 100644
> --- a/fs/nfs/inode.c
> +++ b/fs/nfs/inode.c
> @@ -1233,6 +1233,7 @@ static int nfs_invalidate_mapping(struct inode
> *inode, struct address_space *map
>         struct nfs_inode *nfsi = NFS_I(inode);
>         int ret;
>
> +       nfs_fscache_invalidate(inode, 0);
>         if (mapping->nrpages != 0) {
>                 if (S_ISREG(inode->i_mode)) {
>                         ret = nfs_sync_mapping(mapping);


Actually the above patch fixes truncates when a file is open, not the
case that Jeff mentions.

To be honest I'm not sure about needing a call to fscache_use/unuse_cookie()
around the call to fscache_resize_cookie().  If the cookie has a
refcount of 1 when it is created, and a file is never opened, so
we never call fscache_use_cookie(), what might happen inside
fscache_resize_cookie()?  The header on use_cookie() says
/*
 * Start using the cookie for I/O.  This prevents the backing object from being
 * reaped by VM pressure.
 */

But we're not using it for I/O in this case.
I will have to dig deeper to be sure, or maybe David H will elaborate further.