[9fans] apic fail

[9fans] apic fail

[Richard Miller]

The change to /sys/src/9/pc/apic.c made on 13 July makes my system
(dual-processor atom) crash soon after the ethernet (rtl8169) is
enabled.  Symptom: either a silent hang, or "panic: freeb: ref -1"
called from etheriq (simultaneously on both cpus).

Is one interrupt being dispatched to two cpus, I wonder?

Re: [9fans] apic fail

[Charles Forsyth]

>Is one interrupt being dispatched to two cpus, I wonder?

even so, that should be fine, but ether8169.c doesn't seem to make much use of the rlock it declares.

Re: [9fans] apic fail

[erik quanstrom]

On Mon Jul 19 17:26:13 EDT 2010, forsyth@terzarima.net wrote:
> >Is one interrupt being dispatched to two cpus, I wonder?
>
> even so, that should be fine, but ether8169.c doesn't seem to make much use of the rlock it declares.

really?   i thought a bedrock assumption of pci interrupts is
that once an interrupt vector is asserted, it can't reassert until
the os has acked that vector with EOI.

it may be that the same interrupt source is mapped to different
vectors on the different processors.  that would be an interesting
world to live in.

- erik

Re: [9fans] apic fail

[Richard Miller]

> Is one interrupt being dispatched to two cpus, I wonder?

The answer appears to be yes.  I bracketed rtl8169interrupt()
with these lines:

static void
rtl8169interrupt(Ureg*, void* arg)
{
  ...
  static Ref twice;
  if(incref(&twice) != 1) iprint("cpu%d: I see everything twice!\n", m->machno);
  ...
  decref(&twice);
}

The result on rebooting was:

cpu0: I see everything twice!
cpu1: I see everything twice!
cpu0: I see everything twice!
cpu0: I see everything twice!
...

many times over.

As a side effect this stops the kernel from crashing, because while one cpu
is doing its iprint, the other can get on with its interrupt routine safely.