[PATCH 1/2] eal: fix failure race and behavior of thread create

[David Marchand]

On Thu, Mar 9, 2023 at 9:49 PM Tyler Retzlaff
<roretzla at linux.microsoft.com> wrote:
>
> On Thu, Mar 09, 2023 at 10:58:06AM +0100, Thomas Monjalon wrote:
> > 09/03/2023 10:17, David Marchand:
> > > On Tue, Mar 7, 2023 at 3:33 PM David Marchand <david.marchand at redhat.com> wrote:
> > > > On Thu, Mar 2, 2023 at 7:44 PM Tyler Retzlaff
> > > > <roretzla at linux.microsoft.com> wrote:
> > > > >
> > > > > In rte_thread_create setting affinity after pthread_create may fail.
> > > > > Such a failure should result in the entire rte_thread_create failing
> > > > > but doesn't.
> > > > >
> > > > > Additionally if there is a failure to set affinity a race exists where
> > > > > the creating thread will free ctx and depending on scheduling of the new
> > > > > thread it may also free ctx (double free).
> > > > >
> > > > > Resolve both of the above issues by using the pthread_setaffinity_np
> > > > > prior to thread creation to set the affinity of the created thread. By
> > > > > doing this no failure paths exist after pthread_create returns
> > > > > successfully.
> > > > >
> > > > > Fixes: ce6e911d20f6 ("eal: add thread lifetime API")
> > > > > Cc: stable at dpdk.org
> > > > > Cc: roretzla at linux.microsoft.com
> > > > >
> > > > > Signed-off-by: Tyler Retzlaff <roretzla at linux.microsoft.com>
> > > > Reviewed-by: David Marchand <david.marchand at redhat.com>
> > >
> > > Series applied, thanks.
> >
> > Unfortunately we cannot merge this patch
> > because it does not compile on Alpine Linux (musl libc):
> >
> > lib/eal/unix/rte_thread.c:160:31: error:
> > implicit declaration of function 'pthread_attr_setaffinity_np'
>
> i didn't get any CI failure for this. did i just miss it?

Count on me, I would have complained if there was a CI issue ;-).


>
> >
> > Is it possible to fix the race without using pthread_attr_setaffinity_np?
> >
>
> it seems we never allowed threads to be created with a set affinity when
> using pthread_create directly (that was portable to alpine linux).  for worker
> threads the start_routine is setting the affinity from the new thread.
>
> certainly we can make this work by doing the same thing, but we'll have
> to adjust the start routine wrapper to synchronize/wait for the new
> thread to set the affinity and if it fails terminate the new thread
> cleanly.
>
> i don't have a way to build for alpine linux or run the unit tests, does
> someone want to make the above suggested adjustment? or i can try and
> make a patch but someone else will have to carefully review and test.
>
> let me know how you'd like to proceed.

UNH is looking into re-enabling the Alpine job.


For the time being, if you have a github repository, I can propose a
quick patch using GHA:
https://github.com/david-marchand/dpdk/commit/ci

I had tested compilation with a previous version of this patch.
I just added running the unit tests (adding a checks: tests line in
the job matrix), let's see how it goes...
https://github.com/david-marchand/dpdk/actions/runs/4378715081/jobs/7663806639


-- 
David Marchand