Re: [PATCH] x86/paravirt: Add missing noinstr to arch_local*() helpers

[Marco Elver <elver@google.com>]

On Fri, Aug 07, 2020 at 02:08PM +0200, Marco Elver wrote:
> On Fri, 7 Aug 2020 at 14:04, Jürgen Groß <jgross@suse.com> wrote:
> >
> > On 07.08.20 13:38, Marco Elver wrote:
> > > On Fri, Aug 07, 2020 at 12:35PM +0200, Jürgen Groß wrote:
> > >> On 07.08.20 11:50, Marco Elver wrote:
> > >>> On Fri, Aug 07, 2020 at 11:24AM +0200, Jürgen Groß wrote:
> > >>>> On 07.08.20 11:01, Marco Elver wrote:
> > >>>>> On Thu, 6 Aug 2020 at 18:06, Marco Elver <elver@google.com> wrote:
> > >>>>>> On Thu, 6 Aug 2020 at 15:17, Marco Elver <elver@google.com> wrote:
> > >>>>>>> On Thu, Aug 06, 2020 at 01:32PM +0200, peterz@infradead.org wrote:
> > >>>>>>>> On Thu, Aug 06, 2020 at 09:47:23AM +0200, Marco Elver wrote:
> > >>>>>>>>> Testing my hypothesis that raw then nested non-raw
> > >>>>>>>>> local_irq_save/restore() breaks IRQ state tracking -- see the reproducer
> > >>>>>>>>> below. This is at least 1 case I can think of that we're bound to hit.
> > >>>>>>> ...
> > >>>>>>>>
> > >>>>>>>> /me goes ponder things...
> > >>>>>>>>
> > >>>>>>>> How's something like this then?
> > >>>>>>>>
> > >>>>>>>> ---
> > >>>>>>>>     include/linux/sched.h |  3 ---
> > >>>>>>>>     kernel/kcsan/core.c   | 62 ++++++++++++++++++++++++++++++++++++---------------
> > >>>>>>>>     2 files changed, 44 insertions(+), 21 deletions(-)
> > >>>>>>>
> > >>>>>>> Thank you! That approach seems to pass syzbot (also with
> > >>>>>>> CONFIG_PARAVIRT) and kcsan-test tests.
> > >>>>>>>
> > >>>>>>> I had to modify it some, so that report.c's use of the restore logic
> > >>>>>>> works and not mess up the IRQ trace printed on KCSAN reports (with
> > >>>>>>> CONFIG_KCSAN_VERBOSE).
> > >>>>>>>
> > >>>>>>> I still need to fully convince myself all is well now and we don't end
> > >>>>>>> up with more fixes. :-) If it passes further testing, I'll send it as a
> > >>>>>>> real patch (I want to add you as Co-developed-by, but would need your
> > >>>>>>> Signed-off-by for the code you pasted, I think.)
> > >>>>>
> > >>>>> I let it run on syzbot through the night, and it's fine without
> > >>>>> PARAVIRT (see below). I have sent the patch (need your Signed-off-by
> > >>>>> as it's based on your code, thank you!):
> > >>>>> https://lkml.kernel.org/r/20200807090031.3506555-1-elver@google.com
> > >>>>>
> > >>>>>> With CONFIG_PARAVIRT=y (without the notrace->noinstr patch), I still
> > >>>>>> get lockdep DEBUG_LOCKS_WARN_ON(!lockdep_hardirqs_enabled()), although
> > >>>>>> it takes longer for syzbot to hit them. But I think that's expected
> > >>>>>> because we can still get the recursion that I pointed out, and will
> > >>>>>> need that patch.
> > >>>>>
> > >>>>> Never mind, I get these warnings even if I don't turn on KCSAN
> > >>>>> (CONFIG_KCSAN=n). Something else is going on with PARAVIRT=y that
> > >>>>> throws off IRQ state tracking. :-/
> > >>>>
> > >>>> What are the settings of CONFIG_PARAVIRT_XXL and
> > >>>> CONFIG_PARAVIRT_SPINLOCKS in this case?
> > >>>
> > >>> I attached a config.
> > >>>
> > >>>     $> grep PARAVIRT .config
> > >>>     CONFIG_PARAVIRT=y
> > >>>     CONFIG_PARAVIRT_XXL=y
> > >>>     # CONFIG_PARAVIRT_DEBUG is not set
> > >>>     CONFIG_PARAVIRT_SPINLOCKS=y
> > >>>     # CONFIG_PARAVIRT_TIME_ACCOUNTING is not set
> > >>>     CONFIG_PARAVIRT_CLOCK=y
> > >>
> > >> Anything special I need to do to reproduce the problem? Or would you be
> > >> willing to do some more rounds with different config settings?
> > >
> > > I can only test it with syzkaller, but that probably doesn't help if you
> > > don't already have it set up. It can't seem to find a C reproducer.
> > >
> > > I did some more rounds with different configs.
> > >
> > >> I think CONFIG_PARAVIRT_XXL shouldn't matter, but I'm not completely
> > >> sure about that. CONFIG_PARAVIRT_SPINLOCKS would be my primary suspect.
> > >
> > > Yes, PARAVIRT_XXL doesn't make a different. When disabling
> > > PARAVIRT_SPINLOCKS, however, the warnings go away.
> >
> > Thanks for testing!
> >
> > I take it you are doing the tests in a KVM guest?
> 
> Yes, correct.
> 
> > If so I have a gut feeling that the use of local_irq_save() and
> > local_irq_restore() in kvm_wait() might be fishy. I might be completely
> > wrong here, though.
> 
> Happy to help debug more, although I might need patches or pointers
> what to play with.
> 
> > BTW, I think Xen's variant of pv spinlocks is fine (no playing with IRQ
> > on/off).
> >
> > Hyper-V seems to do the same as KVM, and kicking another vcpu could be
> > problematic as well, as it is just using IPI.

I experimented a bit more, and the below patch seems to solve the
warnings. However, that was based on your pointer about kvm_wait(), and
I can't quite tell if it is the right solution.

My hypothesis here is simply that kvm_wait() may be called in a place
where we get the same case I mentioned to Peter,

	raw_local_irq_save(); /* or other IRQs off without tracing */
	...
	kvm_wait() /* IRQ state tracing gets confused */
	...
	raw_local_irq_restore();

and therefore, using raw variants in kvm_wait() works. It's also safe
because it doesn't call any other libraries that would result in corrupt
IRQ state AFAIK.

Thanks,
-- Marco

------ >8 ------

diff --git a/arch/x86/kernel/kvm.c b/arch/x86/kernel/kvm.c
index 233c77d056c9..1d412d1466f0 100644
--- a/arch/x86/kernel/kvm.c
+++ b/arch/x86/kernel/kvm.c
@@ -797,7 +797,7 @@ static void kvm_wait(u8 *ptr, u8 val)
 	if (in_nmi())
 		return;
 
-	local_irq_save(flags);
+	raw_local_irq_save(flags);
 
 	if (READ_ONCE(*ptr) != val)
 		goto out;
@@ -810,10 +810,10 @@ static void kvm_wait(u8 *ptr, u8 val)
 	if (arch_irqs_disabled_flags(flags))
 		halt();
 	else
-		safe_halt();
+		raw_safe_halt();
 
 out:
-	local_irq_restore(flags);
+	raw_local_irq_restore(flags);
 }
 
 #ifdef CONFIG_X86_32