[PATCH v4 0/5] io_uring passthrough for nvme

[Kanchan Joshi]

On Thu, May 05, 2022 at 12:29:27PM -0600, Jens Axboe wrote:
>On 5/5/22 12:20 PM, Jens Axboe wrote:
>> On 5/5/22 12:06 AM, Kanchan Joshi wrote:
>>> This iteration is against io_uring-big-sqe brach (linux-block).
>>> On top of a739b2354 ("io_uring: enable CQE32").
>>>
>>> fio testing branch:
>>> https://protect2.fireeye.com/v1/url?k=b0d23f72-d1592a52-b0d3b43d-74fe485fb347-02541f801e3b5f5f&q=1&e=ef4bb07a-7707-4854-819c-98abcabb5d2d&u=https%3A%2F%2Fgithub.com%2Fjoshkan%2Ffio%2Ftree%2Fbig-cqe-pt.v4
>>
>> I folded in the suggested changes, the branch is here:
>>
>> https://protect2.fireeye.com/v1/url?k=40e9c3d0-2162d6f0-40e8489f-74fe485fb347-f8e801be1f796980&q=1&e=ef4bb07a-7707-4854-819c-98abcabb5d2d&u=https%3A%2F%2Fgit.kernel.dk%2Fcgit%2Flinux-block%2Flog%2F%3Fh%3Dfor-5.19%2Fio_uring-passthrough
>>
>> I'll try and run the fio test branch, but please take a look and see what
>> you think.
>
>Tested that fio branch and it works for me with what I had pushed out.
>Also tested explicit deferral of requests.

Thanks for sorting everything out! I could test this out now only :-(
Fio scripts ran fine (post refreshing SQE128/CQE32 flag values;repushed fio).

I think these two changes are needed in your branch:

1. since uring-cmd can be without large-cqe, we need to add that
condition in io_uring_cmd_done(). This change in patch 1 - 

diff --git a/fs/io_uring.c b/fs/io_uring.c
index 884f40f51536..c24469564ebc 100644
--- a/fs/io_uring.c
+++ b/fs/io_uring.c
@@ -4938,7 +4938,10 @@ void io_uring_cmd_done(struct io_uring_cmd *ioucmd, ssize_t ret, ssize_t res2)

        if (ret < 0)
                req_set_fail(req);
-       __io_req_complete32(req, 0, ret, 0, res2, 0);
+       if (req->ctx->flags & IORING_SETUP_CQE32)
+               __io_req_complete32(req, 0, ret, 0, res2, 0);
+       else
+               io_req_complete(req, ret);
 }
 EXPORT_SYMBOL_GPL(io_uring_cmd_done);

2. In the commit-message of patch 1, we should delete last line.
i.e.
"This operation can be issued only on the ring that is
setup with both IORING_SETUP_SQE128 and IORING_SETUP_CQE32 flags."

And/or we can move this commit-message of patch 4.

You can either take these changes in, or I can respin the series. LMK.