Re: smp_call_function_single lockups

[Linus Torvalds]

On Mon, Mar 30, 2015 at 8:15 PM, Chris J Arges
<chris.j.arges@xxxxxxxxxxxxx> wrote:
>
> I've been able to repro with your patch and observed the WARN_ON when booting a
> VM on affected hardware and non affected hardware:

Ok, interesting. So the whole "we try to do an APIC ACK with the ISR
bit clear" seems to be a real issue.

> I modified the posted patch with the following:
> -               WARN_ON_ONCE(!(v & 1));
> +               WARN(!(v & 1), "ack_APIC_irq: vector = %0x\n", vector);

Yes, makes sense, although I'm not sure what the vector translations
end up being. See below.

> And it showed vector = 1b when booting. However, when I run the reproducer on
> an affected machine I get the following WARNs before the hang:

Ok, so the boot-time thing I think happens because a device irq
happens but goes away immediately because the CPU that triggers it
also clears it immediately in the device initialization code, so it's
a level-triggered interrupt that goes away "on its own".

But vector 0x1b seems odd. I thought we mapped external interrupts to
0x20+ (FIRST_EXTERNAL_VECTOR). Ingo/Peter? Is there any sane interface
to look up the percpu apic vector data?

Chris, since this is repeatable for you, can you do

        int irq;
        irq = __this_cpu_read(vector_irq[vector]);

and print that out too? That *should* show the actual hardware irq,
although there are a few magic cases too (-1/-2 mean special things)

But the fact that you get the warning before the hang is much more interesting.

> [   36.301299] WARNING: CPU: 0 PID: 0 at ./arch/x86/include/asm/apic.h:444 apic_ack_edge+0x93/0xa0()
> [   36.301301] ack_APIC_irq: vector = e1

Is this repeatable? Does it happen before *every* hang, or at least
often enough to be a good pattern?

> [   40.430533] ack_APIC_irq: vector = 22
>
> So vector = e1 then 22 before the hang.

Is it always the same ones? I assume that on different machines the
vector allocations would be different, but is it consistent on any
particular machine? That's assuming the whole warning is consistent at
all before the hang, of course.

> Anyway, maybe this sheds some more light on this issue. I can reproduce this at
> will, so let me know of other experiments to do.

Somebody else who knows the apic needs to also take a look, but I'd
love to hear what the actual hardware interrupt is (from that
"vector_irq[vector]" thing above.

I'm not recognizing 0xe1 as any of the hardcoded SMP vectors (they are
0xf0-0xff), so it sounds like an external one. But that then requires
the whole mapping table thing.

Ingo/Peter/Jiang - is there anything else useful we could print out? I
worry about the irq movement code. Can we add printk's to when an irq
is chasing from one CPU to another and doing that "move_in_progress"
thing? I've always been scared of that code.

                              Linus
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/