Re: [BUG] 4.4.262: infinite loop in futex_unlock_pi (EAGAIN loop)

[Joe Korty <joe.korty@concurrent-rt.com>]

On Wed, Jul 28, 2021 at 08:07:00AM +0200, Greg Kroah-Hartman wrote:
> On Tue, Jul 27, 2021 at 06:19:50PM -0400, Joe Korty wrote:
> > 
> >  [ Added missing people to the cc: as listed in MAINTAINERS ]
> > 
> > On Thu, Jul 22, 2021 at 04:11:41PM +0200, Greg Kroah-Hartman wrote:
> > > On Mon, Jul 19, 2021 at 12:24:18PM -0400, Joe Korty wrote:
> > > > [BUG] 4.4.262: infinite loop in futex_unlock_pi (EAGAIN loop)
> > > > 
> > > >    [ replicator, attached ]
> > > >    [ workaround patch that crudely clears the loop, attached ]
> > > >    [ 4.4.256 does _not_ have this problem, 4.4.262 is known to have it ]
> > > > 
> > > > When a certain, secure-site application is run on 4.4.262, it locks up and
> > > > is unkillable.  Crash(8) and sysrq backtraces show that the application
> > > > is looping in the kernel in futex_unlock_pi.
> > > > 
> > > > Between 4.4.256 and .257, 4.4 got this 4.12 patch backported into it:
> > > > 
> > > >    73d786b ("[PATCH] futex: Rework inconsistent rt_mutex/futex_q state")
> > > > 
> > > > This patch has the following comment:
> > > > 
> > > >    The only problem is that this breaks RT timeliness guarantees. That
> > > >    is, consider the following scenario:
> > > > 
> > > >       T1 and T2 are both pinned to CPU0. prio(T2) > prio(T1)
> > > > 
> > > >         CPU0
> > > > 
> > > >         T1
> > > >           lock_pi()
> > > >           queue_me()  <- Waiter is visible
> > > >    
> > > >         preemption
> > > > 
> > > >         T2
> > > >           unlock_pi()
> > > >             loops with -EAGAIN forever
> > > > 
> > > >     Which is undesirable for PI primitives. Future patches will rectify
> > > >     this.
> > > > 
> > > > This describes the situation exactly.  To prove, we developed a little
> > > > kernel patch that, on loop detection, puts a message into the kernel log for
> > > > just the first occurrence, keeps a count of the number of occurrences seen
> > > > since boot, and tries to break out of the loop via usleep_range(1000,1000).
> > > > Note that the patch is not really needed for replication.  It merely shows,
> > > > by 'fixing' the problem, that it really is the EAGAIN loop that triggers
> > > > the lockup.
> > > > 
> > > > Along with this patch, we submit a replicator.  Running this replicator
> > > > with this patch, it can be seen that 4.4.256 does not have the problem.
> > > > 4.4.267 and the latest 4.4, 4.4.275, do.  In addition, 4.9.274 (tested
> > > > w/o the patch) does not have the problem.
> > > > 
> > > > >From this pattern there may be some futex fixup patch that was ported
> > > > back into 4.9 but failed to make it to 4.4.
> > > 
> > > Odd, I can't seem to find anything that we missed.  Can you dig to see
> > > if there is something that we need to do here so we can resolve this?
> > > 
> > > thanks,
> > > greg k-h
> > 
> > 
> > Hi Greg,
> > 
> > 4.12 has these apparently-original patches:
> > 
> >   73d786b  futex: Rework inconsistent rt_mutex/futex_q state
> >   cfafcd1  futex: Rework futex_lock_pi() to use rt_mutex_*_proxy_lock()
> > 
> > I have verified that the first commit, 73d786b, introduces
> > the futex_unlock_pi infinite loop bug into 4.12.  I have
> > also verified that the last commit, cfafcd1, fixes the bug.
> > 
> > 4.9 has had both futex patches backported into it.
> > Verified that 4.9.276 does not suffer from the bug.
> > 
> > 4.4 has had the first patch backported, as 394fc49, but
> > not the last.  I have verified that building a kernel at
> > 394fc49 and at v4.4.276, the bug is seen, and at 394fc49^,
> > the bug is not present.
> > 
> > The missing commit, cfafcd1 in 4.12, is present in 4.9
> > as 13c98b0.  A visual spot-check of 13c98b0, as a patch,
> > with kernel/futex.c in 4.4.276 did not find any hunks of
> > 13c98b0 present in 4.4.276's kernel/futex.c.
> 
> Ok, so what do you recommend be done to resolve this?
> 
> thanks,
> greg k-h

I suppose we could either back out 394fc49 from 4.4, or
backport 13c98b0 from 4.9 to 4.4.  At the time I wrote
the above, I hadn't tried either approach yet.

Since then, I did a trial backport of 13c98b0 into 4.4.
All the changes to kernel/futex.c applied, none of the
changes to kernel/locking/rtmutex.c applied.  That implies
to me that we have at least one other patch that needs
finding-n-backporting before we can proceed.

I hope this doesn't turn into a Wack-a-Mole operation...

Joe