Re: PI futexes + lock stealing woes

[Darren Hart]

On Fri, Dec 01, 2017 at 03:49:01PM -0600, Julia Cartwright wrote:
> On Fri, Dec 01, 2017 at 12:11:15PM -0800, Darren Hart wrote:
> > On Wed, Nov 29, 2017 at 11:56:05AM -0600, Julia Cartwright wrote:
> > 
> > > The actual kernel we've been testing is 4.9.33-rt23, w/ 153fbd1226fb3
> > > ("futex: Fix more put_pi_state() vs. exit_pi_state_list() races")
> > 
> > And this does not exhibit the behavior above, correct?
> 
> Sorry if I was unclear.  This combination _does_ exhibit this incorrect
> behavior.
> 
> > > cherry-picked w/ PREEMPT_RT_FULL.  However, it appears that this issue
> > > may affect v4.15-rc1?
> >
> > And this does?
> 
> I only meant that: as far as I can tell the affected codepaths are
> mostly the same between v4.9.33-rt23 and v4.15-rc1, as the futex
> reworking stuff was cherry-picked back.
> 

Can you compare the futex and rtmutex related histories and see if there
is possibly something missing from the backport? A diff from current
version of the relevant files would be worth doing as well. There is
enough subtly here, that I'd want to be as confident as possible that we
aren't missing something here.

> We haven't yet tried reproducing on v4.15-rc1, and aren't really at a
> place where we can do so quickly.  It's unclear whether or not
> PREEMPT_RT is required to reproduce.

Obviously reproducing on an unmodified upstream (RT or not) would be a
very valuable data point.

> 
> Thanks!
>    Julia
> 

-- 
Darren Hart
VMware Open Source Technology Center