Re: [PATCH v4 00/10] Add Kernel Concurrency Sanitizer (KCSAN)

[Marco Elver]

On Tue, 19 Nov 2019, Marco Elver wrote:

> On Tue, 19 Nov 2019 at 21:13, Qian Cai <cai@xxxxxx> wrote:
> >
> > On Thu, 2019-11-14 at 19:02 +0100, 'Marco Elver' via kasan-dev wrote:
> > > This is the patch-series for the Kernel Concurrency Sanitizer (KCSAN).
> > > KCSAN is a sampling watchpoint-based *data race detector*. More details
> > > are included in **Documentation/dev-tools/kcsan.rst**. This patch-series
> > > only enables KCSAN for x86, but we expect adding support for other
> > > architectures is relatively straightforward (we are aware of
> > > experimental ARM64 and POWER support).
> >
> > This does not allow the system to boot. Just hang forever at the end.
> >
> > https://cailca.github.io/files/dmesg.txt
> >
> > the config (dselect KASAN and select KCSAN with default options):
> >
> > https://raw.githubusercontent.com/cailca/linux-mm/master/x86.config
> 
> Thanks! That config enables lots of other debug code. I could
> reproduce the hang. It's related to CONFIG_PROVE_LOCKING etc.
> 
> The problem is definitely not the fact that kcsan_setup_watchpoint
> disables interrupts (tested by removing that code). Although lockdep
> still complains here, and looking at the code in kcsan/core.c, I just
> can't see how local_irq_restore cannot be called before returning (in
> the stacktrace you provided, there is no kcsan function), and
> interrupts should always be re-enabled. (Interrupts are only disabled
> during delay in kcsan_setup_watchpoint.)
> 
> What I also notice is that this happens when the console starts
> getting spammed with data-race reports (presumably because some extra
> debug code has lots of data races according to KCSAN).
> 
> My guess is that some of the extra debug logic enabled in that config
> is incompatible with KCSAN. However, so far I cannot tell where
> exactly the problem is. For now the work-around would be not using
> KCSAN with these extra debug options.  I will investigate more, but
> nothing obviously wrong stands out..

It seems that due to spinlock_debug.c containing data races, the console
gets spammed with reports. However, it's also possible to encounter
deadlock, e.g.  printk lock -> spinlock_debug -> KCSAN detects data race
-> kcsan_print_report() -> printk lock -> deadlock.

So the best thing is to fix the data races in spinlock_debug. I will
send a patch separately for you to test.

The issue that lockdep still reports inconsistency in IRQ flags tracing
I cannot yet say what the problem is. It seems that lockdep IRQ flags
tracing may have an issue with KCSAN for numerous reasons: let's say
lockdep and IRQ flags tracing code is instrumented, which then calls
into KCSAN, which disables/enables interrupts, but due to tracing calls
back into lockdep code. In other words, there may be some recursion
which corrupts hardirqs_enabled.

Thanks,
-- Marco