[RFC PATCH v3 0/3] Async nvme passthrough over io_uring

[Kanchan Joshi]

On Thu, Mar 18, 2021 at 7:28 AM Jens Axboe <axboe at kernel.dk> wrote:
>
> On 3/17/21 3:31 AM, Kanchan Joshi wrote:
> > On Tue, Mar 16, 2021 at 9:31 PM Jens Axboe <axboe at kernel.dk> wrote:
> >>
> >> On 3/16/21 8:01 AM, Kanchan Joshi wrote:
> >>> This series adds async passthrough capability for nvme block-dev over
> >>> iouring_cmd. The patches are on top of Jens uring-cmd branch:
> >>> https://git.kernel.dk/cgit/linux-block/log/?h=io_uring-fops.v3
> >>>
> >>> Application is expected to allocate passthrough command structure, set
> >>> it up traditionally, and pass its address via "block_uring_cmd->addr".
> >>> On completion, CQE is posted with completion-status after any ioctl
> >>> specific buffer/field update.
> >>
> >> Do you have a test app? I'd be curious to try and add support for this
> >> to t/io_uring from fio just to run some perf numbers.
> >
> > Yes Jens. Need to do a couple of things to make it public, will post it today.

Please see if this is accessible to you -
https://github.com/joshkan/fio/commit/6c18653bc87015213a18c23d56518d4daf21b191

I run it on nvme device with the extra option "-uring_cmd=1". And pit
passthru read/write against regular uring read/write.
While write perf looks fine, I notice higher-qd reads going tad-bit
low which is puzzling.
But I need to test more to see if this is about my test-env (including
the added test-code) itself.

It will be great if you could, at some point in future, have a look at
this test or spin off what you already had in mind.

> Sounds good! I commented on 1/3, I think it can be simplified and
> cleaned up quite a bit, which is great. Then let's base it on top of v4
> that I posted, let me know if you run into any issues with that.

Yes, will move to V4,  thanks!