Re: [PATCH] nfs: set block size according to pnfs_blksize first

[Trond Myklebust <trondmy@hammerspace.com>]

On Wed, 2021-06-16 at 22:06 +0800, Gao Xiang wrote:
> On Wed, Jun 16, 2021 at 01:47:13PM +0000, Trond Myklebust wrote:
> > On Wed, 2021-06-16 at 20:44 +0800, Gao Xiang wrote:
> > > When testing fstests with ext4 over nfs 4.2, I found generic/486
> > > failed. The root cause is that the length of its xattr value is
> > >   min(st_blksize * 3 / 4, XATTR_SIZE_MAX)
> > > 
> > > which is 4096 * 3 / 4 = 3072 for underlayfs ext4 rather than
> > > XATTR_SIZE_MAX = 65536 for nfs since the block size would be
> > > wsize
> > > (=131072) if bsize is not specified.
> > > 
> > > Let's use pnfs_blksize first instead of using wsize directly if
> > > bsize isn't specified. And the testcase itself can pass now.
> > > 
> > > Cc: Trond Myklebust <trond.myklebust@hammerspace.com>
> > > Cc: Anna Schumaker <anna.schumaker@netapp.com>
> > > Cc: Joseph Qi <joseph.qi@linux.alibaba.com>
> > > Signed-off-by: Gao Xiang <hsiangkao@linux.alibaba.com>
> > > ---
> > > Considering bsize is not specified, we might use pnfs_blksize
> > > directly first rather than wsize.
> > > 
> > >  fs/nfs/super.c | 8 ++++++--
> > >  1 file changed, 6 insertions(+), 2 deletions(-)
> > > 
> > > diff --git a/fs/nfs/super.c b/fs/nfs/super.c
> > > index fe58525cfed4..5015edf0cd9a 100644
> > > --- a/fs/nfs/super.c
> > > +++ b/fs/nfs/super.c
> > > @@ -1068,9 +1068,13 @@ static void nfs_fill_super(struct
> > > super_block
> > > *sb, struct nfs_fs_context *ctx)
> > >         snprintf(sb->s_id, sizeof(sb->s_id),
> > >                  "%u:%u", MAJOR(sb->s_dev), MINOR(sb->s_dev));
> > >  
> > > -       if (sb->s_blocksize == 0)
> > > -               sb->s_blocksize = nfs_block_bits(server->wsize,
> > > +       if (sb->s_blocksize == 0) {
> > > +               unsigned int blksize = server->pnfs_blksize ?
> > > +                       server->pnfs_blksize : server->wsize;
> > 
> > NACK. The pnfs block size is a layout driver-specific quantity, and
> > should not be used to substitute for the server-advertised block
> > size.
> > It only applies to I/O _if_ the client is holding a layout for a
> > specific file and is using pNFS to do I/O to that file.
> 
> Honestly, I'm not sure if it's ok as well.
> 
> > 
> > It has nothing to do with xattrs at all.
> 
> Yet my question is how to deal with generic/486, should we just skip
> the case directly? I cannot find some proper way to get underlayfs
> block size or real xattr value limit.
> 

RFC8276 provides no method for determining the xattr size limits. It
just notes that such limits may exist, and provides the error code
NFS4ERR_XATTR2BIG, that the server may use as a return value when those
limits are exceeded.

> For now, generic/486 will return ENOSPC at
> fsetxattr(fd, "user.world", value, 65536, XATTR_REPLACE);
> when testing new nfs4.2 xattr support.
> 

As noted above, the NFS server should really be returning
NFS4ERR_XATTR2BIG in this case, which the client, again, should be
transforming into -E2BIG. Where does ENOSPC come from?

> Thanks,
> Gao Xiang
> 
> > 
> > > +
> > > +               sb->s_blocksize = nfs_block_bits(blksize,
> > >                                                  &sb-
> > > > s_blocksize_bits);
> > > +       }
> > >  
> > >         nfs_super_set_maxbytes(sb, server->maxfilesize);
> > >         server->has_sec_mnt_opts = ctx->has_sec_mnt_opts;

-- 
Trond Myklebust
Linux NFS client maintainer, Hammerspace
trond.myklebust@hammerspace.com