RE: Bug? CQE.res = -EAGAIN with nvme multipath driver

["Haeuptle, Michael" <michael.haeuptle@hpe.com>]

Hey Jens, sorry for the late response.

I was unable to reproduce the issue with your branch. However, I didn't even hit the spot where same_thread_group check was removed.

We backported your changes to 6.1.119 and we did see that our original issue is fixed with your patches.

It seems to me that io_uring performance increased quite a bit in the latest kernel, judging from fio queue utilization of my workload. Maybe that's why I'm not hitting the place where same_thread_group was removed.

Your patch didn't cause any regression after 1d testing in my NVMF/RDMA & multipath setup. So, I think it would be good to get this patch on main.

-- Michael

-----Original Message-----
From: Jens Axboe <axboe@kernel.dk> 
Sent: Tuesday, January 7, 2025 11:27 AM
To: Haeuptle, Michael <michael.haeuptle@hpe.com>; io-uring@vger.kernel.org
Subject: Re: Bug? CQE.res = -EAGAIN with nvme multipath driver

On 1/7/25 11:24 AM, Haeuptle, Michael wrote:
> Thanks for the quick response!
> 
> When I remove that check on the 6.1.85 kernel version we're using, 
> then it seems that the user space program is losing IOs. I confirmed 
> this with fio. When we hit this issue, fio never completes and is 
> stuck.

That's because the io_uring logic assumes it happens inline via submission, and for your case it does not. Which is also why it gets failed. And hence setting the retry flag in that condition will do absolutely nothing, as nobody is there to see it.

> I can certainly try that later kernel with your fix, if you think 
> there are other changes that prevent losing IOs.

Please try the branch and see how it fares for you.

--
Jens Axboe