Re: [PATCH v4 7/8] lockdep: Change hardirq{s_enabled,_context} to per-cpu variables

[Peter Zijlstra <peterz@infradead.org>]

On Tue, Jun 23, 2020 at 07:59:57PM +0200, Marco Elver wrote:
> On Tue, Jun 23, 2020 at 06:37PM +0200, Peter Zijlstra wrote:
> > On Tue, Jun 23, 2020 at 06:13:21PM +0200, Ahmed S. Darwish wrote:
> > > Well, freshly merged code is using it. For example, KCSAN:
> > > 
> > >     => f1bc96210c6a ("kcsan: Make KCSAN compatible with lockdep")
> > >     => kernel/kcsan/report.c:
> > > 
> > >     void kcsan_report(...)
> > >     {
> > > 	...
> > >         /*
> > >          * With TRACE_IRQFLAGS, lockdep's IRQ trace state becomes corrupted if
> > >          * we do not turn off lockdep here; this could happen due to recursion
> > >          * into lockdep via KCSAN if we detect a race in utilities used by
> > >          * lockdep.
> > >          */
> > >         lockdep_off();
> > > 	...
> > >     }
> > 
> > Marco, do you remember what exactly happened there? Because I'm about to
> > wreck that. That is, I'm going to make TRACE_IRQFLAGS ignore
> > lockdep_off().
> 
> Yeah, I was trying to squash any kind of recursion:
> 
> 	lockdep -> other libs ->
> 		-> KCSAN
> 		-> print report
> 		-> dump stack, printk and friends
> 		-> lockdep -> other libs
> 			-> KCSAN ...
> 
> Some history:
> 
> * Initial patch to fix:
> 	https://lore.kernel.org/lkml/20200115162512.70807-1-elver@google.com/

That patch is weird; just :=n on lockdep.c should've cured that, the
rest is massive overkill.

> * KCSAN+lockdep+ftrace:
> 	https://lore.kernel.org/lkml/20200214211035.209972-1-elver@google.com/

That doesn't really have anything useful..

> lockdep now has KCSAN_SANITIZE := n, but we still need to ensure that
> there are no paths out of lockdep, or the IRQ flags tracing code, that
> might lead through other libs, through KCSAN, libs used to generate a
> report, and back to lockdep.
> 
> I never quite figured out the exact trace that led to corruption, but
> avoiding any kind of potential for recursion was the only thing that
> would avoid the check_flags() warnings.

Fair enough; I'll rip it all up and boot a KCSAN kernel, see what if
anything happens.