On 15/02/16 18:54, Rafael J. Wysocki wrote:

On Mon, Feb 15, 2016 at 7:49 PM, Marc Zyngier [off-list ref] wrote:

quoted

On 15/02/16 18:41, Rafael J. Wysocki wrote:

quoted

On Mon, Feb 15, 2016 at 6:05 PM, Guenter Roeck [off-list ref] wrote:

quoted

Rafael,

Hi,

Thanks for the report!

quoted

I see crashes in various arm qemu tests due to 'cpufreq: governor: Replace
timers with utilization update callbacks' with next-20160215. An example
crash log and bisect results are attached below.

Please let me know if there is anything I can do to help tracking down
the problem.

It looks like we've uncovered some nastiness in the arch ARM code (see below).

[cut]

quoted

[    1.340000] Unable to handle kernel NULL pointer dereference at virtual address 00000000
[    1.340000] pgd = c0204000
[    1.340000] [00000000] *pgd=00000000
[    1.340000] Internal error: Oops: 80000005 [#1] SMP ARM
[    1.340000] Modules linked in:
[    1.340000] CPU: 0 PID: 1 Comm: swapper/0 Not tainted 4.5.0-rc4-next-20160215 #1
[    1.340000] Hardware name: Generic OMAP3-GP (Flattened Device Tree)
[    1.340000] task: cb060000 ti: cb05a000 task.ti: cb05a000
[    1.340000] PC is at 0x0
[    1.340000] LR is at arch_send_call_function_single_ipi+0x34/0x38

Since this is ARM, arch_send_call_function_single_ipi() looks like this:

void arch_send_call_function_single_ipi(int cpu)
{
         smp_cross_call(cpumask_of(cpu), IPI_CALL_FUNC_SINGLE);
}

so I'm not sure how the NULL pointer deref is possible even.

The only thing coming to mind would be that cpumask_of(cpu) triggers
this, but I'm not sure how exactly that can happen.

I need help from somebody who knows how this low-level stuff works on ARM.

Given that OMAP3 is a UP system, there is zero chance that it has
registered the magic hook that delivers IPIs (its interrupt controller
is not even capable of doing so).

I don't really know the context, but IPIs on a UP system seem at best odd.

That would explain it, thanks.

So it looks like we should always use irq_work_queue() on UP even if
CONFIG_SMP is set, shouldn't we?

Something like that, yes. CONFIG_SMP is not an indication of an SMP
system anymore (we've even dropped the config option on arm64).

Hopefully num_possible_cpus() is reliable enough to let you do the right
thing...

Thanks,

	M.
-- 
Jazz is not dead. It just smells funny...