Re: Panic and page fault in loop during handling NMI backtrace handler

[Frederic Weisbecker <fweisbec@gmail.com>]

On Tue, Oct 15, 2013 at 06:06:59AM -0700, Paul E. McKenney wrote:
> On Tue, Oct 15, 2013 at 02:59:15PM +0200, Frederic Weisbecker wrote:
> > On Tue, Oct 15, 2013 at 05:48:37AM -0700, Paul E. McKenney wrote:
> > > On Tue, Oct 15, 2013 at 02:37:17PM +0200, Peter Zijlstra wrote:
> > > > On Tue, Oct 15, 2013 at 02:18:53PM +0200, Frederic Weisbecker wrote:
> > > > > On Tue, Oct 15, 2013 at 02:01:04AM +0000, Liu, Chuansheng wrote:
> > > > > > We meet one issue that during trigger all CPU backtrace, but during in the NMI handler arch_trigger_all_cpu_backtrace_handler,
> > > > > > It hit the PAGE fault, then PAGE fault is in loop, at last the thread stack overflow, and system panic.
> > > > > > 
> > > > > > Anyone can give some help? Thanks.
> > > > > 
> > > > > Looks like we re-enter the fault several times. On x86-32, NMIs can
> > > > > fault if they dereference vmalloc'ed area. I wonder if the module thing
> > > > > we lookup in the NMI is stored on some vmalloc'ed area.
> > > > 
> > > > IIRC modules are indeed allocated using vmalloc. See module_alloc()
> > > > using vmalloc_exec()
> > > 
> > > This might then be a module that uses call_rcu(), but which does not have
> > > the needed rcu_barrier() in the module-exit function.
> > 
> > I rather believe it's due to the lazy paging of vmalloc area in x86-32. We had issues
> > like that in the past. For example that's the reason why we do an ad-hoc per-cpu
> > allocation on callchain buffers in perf rather than using alloc_percpu() which might
> > use vmalloc.
> 
> I must defer to your greater experience with this type of bug.

With some chances I'll be proved wrong. I hope, because that issue is not easily fixed.