[RFC PATCH v3 2/3] nvme: keep nvme_command instead of pointer to it

[Keith Busch]

On Wed, Mar 17, 2021 at 03:08:09PM +0530, Kanchan Joshi wrote:
> On Tue, Mar 16, 2021 at 10:53 PM Keith Busch <kbusch at kernel.org> wrote:
> >
> > On Tue, Mar 16, 2021 at 07:31:25PM +0530, Kanchan Joshi wrote:
> > > nvme_req structure originally contained a pointer to nvme_command.
> > > Change nvme_req structure to keep the command itself.
> > > This helps in avoiding hot-path memory-allocation for async-passthrough.
> >
> > I have a slightly different take on how to handle pre-allocated
> > passthrough commands. Every transport except PCI already preallocates a
> > 'struct nvme_command' within the pdu, so allocating another one looks
> > redundant. Also, it does consume quite a bit of memory for something
> > that is used only for the passthrough case.
> >
> > I think we can solve both concerns by always using the PDU nvme_command
> > rather than have the transport drivers provide it. I just sent the patch
> > here if you can take a look. It tested fine on PCI and loop (haven't
> > tested any other transports).
> >
> >  http://lists.infradead.org/pipermail/linux-nvme/2021-March/023711.html
> 
> Sounds fine, thanks for the patch, looking at it.
> Which kernel you used for these. 'Patch 2' doesn't  apply cleanly.

I used nvme-5.13 as my starting point.

 http://git.infradead.org/nvme.git/shortlog/refs/heads/nvme-5.13