[PATCH 3/5] nvme: add support for copy offload

[Christoph Hellwig]

On Wed, May 21, 2025 at 08:41:40PM -0700, Caleb Sander Mateos wrote:
> On Wed, May 21, 2025 at 8:23 PM Keith Busch <kbusch at kernel.org> wrote:
> >
> > On Wed, May 21, 2025 at 05:51:03PM -0700, Caleb Sander Mateos wrote:
> > > On Wed, May 21, 2025 at 5:47 PM Caleb Sander Mateos
> > > >
> > > > alloc_size should be sizeof(*range) * i? Otherwise this exceeds the
> > > > amount of data used by the Copy command, which not all controllers
> > > > support (see bit LLDTS of SGLS in the Identify Controller data
> > > > structure). We have seen the same behavior with Dataset Management
> > > > (always specifying 4 KB of data), which also passes the maximum size
> > > > of the allocation to bvec_set_virt().
> > >
> > > I see that was added in commit 530436c45ef2e ("nvme: Discard
> > > workaround for non-conformant devices"). I would rather wait for
> > > evidence of non-conformant devices supporting Copy before implementing
> > > the same spec-noncompliant workaround. It could be a quirk if
> > > necessary.
> >
> > Right, that's exactly why I didn't bother allocating tighter to what the
> > command actually needs. The number of devices that would have needed a
> > DSM quirk for the full 4k was untenable, so making the quirk behavior
> > the default was a sane compromise. I suppose Copy is a more enterprisey
> > feature in comparison, so maybe we can count on devices doing dma
> > correctly?
> 
> For the record, that change broke Linux hosts sending DSM commands to
> our NVMe controller, which was validating that the SGL length exactly
> matches the number of data bytes implied by the command.

Next time please send a bug report and/or fix ASAP when you see
such changes.