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

[Christoph Hellwig]

On Wed, May 21, 2025 at 10:29:59PM -0600, Keith Busch wrote:
> On Wed, May 21, 2025 at 08:41:40PM -0700, Caleb Sander Mateos wrote:
> > 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. I'm sure
> > we're in the minority of NVMe controller vendors in aggressively
> > validating the NVMe command parameters, but it was unfortunate to
> > discover this change in Linux's behavior.
> 
> This is a fabrics target you're talking about? I assume so because pci
> would use PRP for a 4k payload, which doesn't encode transfer lengths.
> All the offending controllers were pci, so maybe we could have
> constrained the DSM over-allocation to that transport if we knew this
> was causing problems for fabrics.

Or constrain the workaround to PRPs which never encode the actual
length anyway (which probably is the source of the bugs).

The fact that NVMe went with this stupid PRP scheme instead of SGLs
still makes me angry..