Re: [PATCH v3] NFSD: Implement NFSD_IO_DIRECT for NFS WRITE

[Mike Snitzer <snitzer@kernel.org>]

On Fri, Oct 17, 2025 at 05:54:00PM -0400, Chuck Lever wrote:
> On 10/17/25 5:03 PM, Mike Snitzer wrote:
> > On Fri, Oct 17, 2025 at 10:13:14AM -0400, Chuck Lever wrote:
> >> On 10/15/25 6:13 PM, Mike Snitzer wrote:
> 
> >>> +	*cnt = 0;
> >>> +	for (int i = 0; i < n_iters; i++) {
> >>> +		if (iter_is_dio_aligned[i])
> >>> +			kiocb->ki_flags |= IOCB_DIRECT;
> >>> +		else
> >>> +			kiocb->ki_flags &= ~IOCB_DIRECT;
> >>> +
> >>> +		host_err = vfs_iocb_iter_write(file, kiocb, &iter[i]);
> >>> +		if (host_err < 0) {
> >>> +			/*
> >>> +			 * VFS will return -ENOTBLK if DIO WRITE fails to
> >>> +			 * invalidate the page cache. Retry using buffered IO.
> >>> +			 */
> >>
> >> I'm debating with myself whether, now that NFSD is using DIO, nfserrno
> >> should get a mapping from ENOTBLK to nfserr_serverfault or nfserr_io,
> >> simply as a defensive measure.
> >>
> >> I kind of like the idea that we get a warning in nfserrno: "hey, I
> >> didn't expect ENOTBLK here" so I'm leaning towards leaving it as is
> >> for now.
> > 
> > While ENOTBLK isn't expected, it is a real possibility from the MM
> > subsystem's inability invalidate the page cache ("for reasons").
> > 
> > So it isn't that NFSD would've done anything wrong.. it just needs to
> > deal with the possibility (as should any other DIO write in kernel).
>
> Well my point is that if the MM/VFS can return this on a buffered write,
> should NFSD be prepared to deal with it in other places besides this
> path? I think it's handling ENOTBLK correctly here. Just wondering about
> elsewhere.
> 
> This is something we can set aside and worry about after merge. It
> actually has more to do with the existing parts of NFSD, not the new
> direct write code path.

I'm only aware of MM returning ENOTBLK in response to a DIO write that
it couldn't invalidate the page cache for.

So existing NFSD code shouldn't be exposed AFAIK.  But yeah, worth
another look.

> I'll repost the "stable_how" patch and this one together in a new
> series, for Christoph, and pull that series into nfsd-testing.

Awesome, thanks.