Le 09/10/2018 à 13:16, Nicholas Piggin a écrit :

On Tue, 9 Oct 2018 09:36:18 +0000
Christophe Leroy [off-list ref] wrote:

quoted

On 10/09/2018 05:30 AM, Nicholas Piggin wrote:

quoted

On Tue, 9 Oct 2018 06:46:30 +0200
Christophe LEROY [off-list ref] wrote:
   

quoted

Le 09/10/2018 à 06:32, Nicholas Piggin a écrit :

quoted

On Mon, 8 Oct 2018 17:39:11 +0200
Christophe LEROY [off-list ref] wrote:
      

quoted

Hi Nick,

Le 19/07/2017 à 08:59, Nicholas Piggin a écrit :

quoted

Use nmi_enter similarly to system reset interrupts. This uses NMI
printk NMI buffers and turns off various debugging facilities that
helps avoid tripping on ourselves or other CPUs.

Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
---
     arch/powerpc/kernel/traps.c | 9 ++++++---
     1 file changed, 6 insertions(+), 3 deletions(-)

diff --git a/arch/powerpc/kernel/traps.c b/arch/powerpc/kernel/traps.c
index 2849c4f50324..6d31f9d7c333 100644
--- a/arch/powerpc/kernel/traps.c
+++ b/arch/powerpc/kernel/traps.c

@@ -789,8 +789,10 @@ int machine_check_generic(struct pt_regs *regs)
     
     void machine_check_exception(struct pt_regs *regs)
     {
-	enum ctx_state prev_state = exception_enter();
     	int recover = 0;
+	bool nested = in_nmi();
+	if (!nested)
+		nmi_enter();

This alters preempt_count, then when die() is called
in_interrupt() returns true allthough the trap didn't happen in
interrupt, so oops_end() panics for "fatal exception in interrupt"
instead of gently sending SIGBUS the faulting app.

Thanks for tracking that down.
      

quoted

Any idea on how to fix this ?

I would say we have to deliver the sigbus by hand.

       if ((user_mode(regs)))
           _exception(SIGBUS, regs, BUS_MCEERR_AR, regs->nip);
       else
           die("Machine check", regs, SIGBUS);
      

And what about all the other things done by 'die()' ?

And what if it is a kernel thread ?

In one of my boards, I have a kernel thread regularly checking the HW,
and if it gets a machine check I expect it to gently stop and the die
notification to be delivered to all registered notifiers.

Until before this patch, it was working well.

I guess the alternative is we could check regs->trap for machine
check in the die test. Complication is having to account for MCE
in an interrupt handler.

         if (in_interrupt()) {
                  if (!IS_MCHECK_EXC(regs) || (irq_count() - (NMI_OFFSET + HARDIRQ_OFFSET)))
                      panic("Fatal exception in interrupt");
         }

Something like that might work for you? We needs a ppc64 macro for the
MCE, and can probably add something like in_nmi_from_interrupt() for
the second part of the test.

Don't know, I'm away from home on business trip so I won't be able to
test anything before next week. However it looks more or less like a
hack, doesn't it ?

I thought it seemed okay (with the right functions added). Actually it
could be a bit nicer to do this, then it works generally :

          if (in_interrupt()) {
                   if (!in_nmi() || in_nmi_from_interrupt())
                       panic("Fatal exception in interrupt");
          }

quoted

What about the following ?

Hmm, in some ways maybe it's nicer. One complication is I would like the
same thing to be available for platform specific machine check
handlers, so then you need to pass is_in_interrupt to them. Which you
can do without any problem... But is it cleaner than the above?

For me it looks cleaner than twiddle the preempt_count depending on 
whether we were or not already in nmi() .

Let's draft something and see what it looks like.

I guess one advantage of yours is that a BUG somewhere in the NMI path
will panic the system. Or is that a disadvantage?

Why would it panic the system more than now ? And is it an issue at all 
? Doesn't BUG() panic in any case ?

Christophe