Re: [PATCH] watchdog: print all locks on a softlock

[Don Zickus <dzickus@redhat.com>]

On Thu, May 01, 2014 at 10:09:01PM +0200, Frederic Weisbecker wrote:
> On Thu, May 01, 2014 at 03:17:20PM -0400, Don Zickus wrote:
> > On Thu, May 01, 2014 at 02:55:35PM -0400, Eric Paris wrote:
> > > If the CPU hits a softlockup this patch will also have it print the
> > > information about all locks being held on the system.  This might help
> > > determine if a lock is being held too long leading to this problem.
> > 
> > I am not sure this helps you.  A softlockup is the result of pre-emption
> > disabled, ie the scheduler not being called after 60 seconds.  Holding a
> > lock does not disable pre-emption usually.  So I don't think this is going
> > to add anything.
> > 
> > Are you trying to debug a hung task?  The the hung_task thread checks to
> > see if a task hasn't scheduled in 2 minutes or so.  That could be the
> > result of long lock (but that output already dumps the lockdep stuff).
> 
> There may be some deadlocks that lockdep doesn't detect yet. 2 example:
> 
> 1) spinlock <-> IPI dependency
> 
> 
>         CPU 0                                            CPU 1
>         --------------------------------------------------------
>         spin_lock_irq(A)
>         smp_send_function_single_async(CPU 1, func)
>                                                          //IPI
>                                                          func {
>                                                             spin_lock(1)
>                                                          }
> 
> But this should be resolved with a virtual lock on the IPI functions.
> I should try that.
> 
> 2) rwlock <-> IPI
> 
>         CPU 0                                            CPU 1
>         --------------------------------------------------------
>         read_lock(A)
>                                                          write_lock_irq(A)
>         smp_send_function_single(CPU 1, func)
>                                                          //IPI never happens

The hardlockup detector would go off here.  And dumping all the cpus in
the system (something we don't do today), would show this scenario.  I see
this scenario a lot during page flushes on RHEL (a lot being once every
other month or so).

Cheers,
Don

> 
> This one is much trickier.
> 
> Anyway those are the only scenario I know of but there may be more. When possible
> we want to extend lockdep to detect new scenarios of deadlock but we don't have the
> guarantee that it can detect everything.
> 
> So, could be useful...