Re: [PATCH 0/2] io_uring/io-wq: respect cgroup cpusets

[MOESSBAUER, Felix]

On Tue, 2024-09-10 at 08:53 -0600, Jens Axboe wrote:
> On 9/10/24 8:33 AM, Felix Moessbauer wrote:
> > Hi,
> > 
> > this series continues the affinity cleanup work started in
> > io_uring/sqpoll. It has been tested against the liburing testsuite
> > (make runtests), whereby the read-mshot test always fails:
> > 
> >   Running test read-mshot.t
> >   Buffer ring register failed -22
> >   test_inc 0 0
> > failed                                                             
> >                                                              
> >   Test read-mshot.t failed with ret 1     
> > 
> > However, this test also fails on a non-patched linux-next @ 
> > bc83b4d1f086.
> 
> That sounds very odd... What liburing are you using? On old kernels
> where provided buffer rings aren't available the test should just
> skip,
> new ones it should pass. Only thing I can think of is that your
> liburing
> repo isn't current?

Hmm... I tested against
https://github.com/axboe/liburing/commit/74fefa1b51ee35a2014ca6e7667d7c10e9c5b06f

I'll redo the test against the unpatched kernel to be 100% sure that it
is not related to my patches. The -22 is likely an -EINVAL.

> 
> > The test wq-aff.t succeeds if at least cpu 0,1 are in the set and
> > fails otherwise. This is expected, as the test wants to pin on
> > these
> > cpus. I'll send a patch for liburing to skip that test in case this
> > pre-condition is not met.
> > 
> > Regarding backporting: I would like to backport these patches to
> > 6.1 as
> > well, as they affect our realtime applications. However, in-between
> > 6.1
> > and next there is a major change da64d6db3bd3 ("io_uring: One wqe
> > per
> > wq"), which makes the backport tricky. While I don't think we want
> > to
> > backport this change, would a dedicated backport of the two pinning
> > patches for the old multi-queue implementation have a chance to be
> > accepted?
> 
> Let's not backport that patch, just because it's pretty invasive.
> It's
> fine to have a separate backport patch of them for -stable, in this
> case
> we'll have one version for stable kernels new enough to have that
> change, and one for older versions. Thankfully not that many to care
> about.

Ok, that is fine for me. Then let's first get things right in this
series and then I'll send the backport.

Best regards,
Felix

> 

-- 
Siemens AG, Technology
Linux Expert Center