Re: [PATCH v6 4/5] NFSD: Implement NFSD_IO_DIRECT for NFS WRITE

[Mike Snitzer <snitzer@kernel.org>]

On Wed, Oct 22, 2025 at 03:27:02PM -0400, Chuck Lever wrote:
> On 10/22/25 3:22 PM, Chuck Lever wrote:
> > @@ -1311,6 +1484,9 @@ nfsd_vfs_write(struct svc_rqst *rqstp, struct svc_fh *fhp,
> >  	if (sb->s_export_op)
> >  		exp_op_flags = sb->s_export_op->flags;
> >  
> > +	/* cel: UNSTABLE buffered WRITEs might want some form of throttling
> > +	 *	as well to prevent clients from pushing writes to us faster
> > +	 *	than they can be flushed onto durable storage */
> >  	if (test_bit(RQ_LOCAL, &rqstp->rq_flags) &&
> >  	    !(exp_op_flags & EXPORT_OP_REMOTE_FS)) {
> >  		/*
> 
> Note: the above is an open question. I don't intend to merge this patch
> with the above comment included.

Such a mechanism would benefit all NFSD_IO modes.

As-is, even NFSD_IO_DIRECT isn't sufficient to ensure client's forward
progress if NFS client application is using buffered IO -- because the
client itself can then enter page reclaim (when it exhausts its memory
because it doesn't get any back pressure).