lkml.org 
[lkml]   [2012]   [Jun]   [1]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH] x86: Avoid intermixing cpu dump_stack output on multi-processor systems
On Tue, May 29, 2012 at 07:54:07PM -0400, Don Zickus wrote:
> On Tue, May 29, 2012 at 06:11:35PM -0500, Russ Anderson wrote:
> > > > In this case, I am just using the hardware NMI, which sends the NMI
> > > > signal to each logical cpu. Since each cpu receives the NMI at nearly
> > > > the exact same time, they end up in dump_stack() at the same time.
> > > > Without some form of locking, trace lines from different cpus end
> > > > up intermixed, making it impossible to tell what any individual
> > > > cpu was doing.
> > >
> > > I forgot the original reasons for having the NMI go to each CPU instead of
> > > just the boot CPU (commit 78c06176), but it seems like if you revert that
> > > patch and have the nmi handler just call trigger_all_cpu_backtrace()
> > > instead (which does stack trace locking for pretty output), that would
> > > solve your problem, no? That locking is safe because it is only called in
> > > the NMI context.
> >
> > We want NMI to hit all the cpus at the same time to get a coherent
> > snapshot of what is happening in the system at one point in time.
> > Sending an IPI one cpu at a time skews the results, and doesn't
>
> Oh, I thought it was broadcasting, but I see the apic_uv code serializes
> it. Though getting all those hardware locks in the nmi handler has to be
> time consuming? But I know you guys did some tricks to speed that up.
>
> > really solve the problem of multiple cpus going into dump_stack()
> > at the same time. NMI isn't the only possible caller of dump_stack().
>
> I am curious, your NMI handler has locking wrapped around dump_stack,
> shouldn't that serialize the output the way you want it? Why isn't that
> working?

Yes, you're right, it does. It is working. I'd forgotten that
the community kernel has uv_nmi_lock in uv_handle_nmi. Must
be working too much with distro kernels. :-) But that doesn't
help for all the other code paths than call dump_stack.


> > FWIW, "Wait for up to 10 seconds for all CPUs to do the backtrace" on
> > a 4096 cpu system isn't long enough. :-)
>
> Good point. :-)
>
> >
> > > Whereas the lock you are proposing can be called in a mixture of NMI and
> > > IRQ which could cause deadlocks I believe.
> >
> > Since this is a lock just around the dump_stack printk, would
> > checking for forward progress and a timeout to catch any possible
> > deadlock be sufficient? In the unlikely case of a deadlock the
> > lock gets broken and some of the cpu backtraces get intermixed.
> > That is still a huge improvement over the current case where
> > all of the backtraces get intermixed.
>
> I saw your new patch based on Frederick's input. It seems to take care of
> deadlock situations though you run into the starving lock problem that
> ticketed spinlocks solved. Which is why I am curious why moving the
> locking one layer up to the NMI handler (which is where it is currently),
> didn't fix your problem.

Locking in dump_stack would remove the need for uv_nmi_lock.

--
Russ Anderson, OS RAS/Partitioning Project Lead
SGI - Silicon Graphics Inc rja@sgi.com


\
 
 \ /
  Last update: 2012-06-02 01:21    [W:1.944 / U:0.192 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site