[PATCH v2] hardlockup: detect hard lockups without NMIs using secondary cpus

[Colin Cross]

On Mon, Jan 14, 2013 at 6:48 PM, Frederic Weisbecker <fweisbec at gmail.com> wrote:
> 2013/1/15 Colin Cross <ccross at android.com>:
>> On Mon, Jan 14, 2013 at 4:25 PM, Frederic Weisbecker <fweisbec at gmail.com> wrote:
>>> 2013/1/15 Colin Cross <ccross at android.com>:
>>>> On Mon, Jan 14, 2013 at 4:13 PM, Frederic Weisbecker <fweisbec at gmail.com> wrote:
>>>>> I believe this is pretty much what the RCU stall detector does
>>>>> already: checks for other CPUs being responsive. The only difference
>>>>> is on how it checks that. For RCU it's about checking for CPUs
>>>>> reporting quiescent states when requested to do so. In your case it's
>>>>> about ensuring the hrtimer interrupt is well handled.
>>>>>
>>>>> One thing you can do is to enqueue an RCU callback (cal_rcu()) every
>>>>> minute so you can force other CPUs to report quiescent states
>>>>> periodically and thus check for lockups.
>>>>
>>>> That's a good point, I'll take a look at using that.  A minute is too
>>>> long, some SoCs have maximum HW watchdog periods of under 30 seconds,
>>>> but a call_rcu every 10-20 seconds might be sufficient.
>>>
>>> Sure. And you can tune CONFIG_RCU_CPU_STALL_TIMEOUT accordingly.
>>
>> After considering this, I think the hrtimer watchdog is more useful.
>> RCU stalls are not usually panic events, and I wouldn't want to add a
>> panic on every RCU stall.  The lack of stack traces on the affected
>> cpu makes a panic important.  I'm planning to add an ARM DBGPCSR panic
>> handler, which will be able to dump the PC of a stuck cpu even if it
>> is not responding to interrupts.  kexec or kgdb on panic might also
>> allow some inspection of the stack on stuck cpu.
>>
>> Failing to process interrupts is a much more serious event than an RCU
>> stall, and being able to detect them separately may be very valuable
>> for debugging.
>
> RCU stalls can happen for different reasons: softlockup (failure to
> schedule another task), hardlockup (failure to process interrupts), or
> a bug in RCU itself. But if you have a hardlockup, it will report it.

It will report it, but it will report it in the same way that it
reports a less serious issue, and in this case with zero debugging
information since the affected cpu won't dump its backtrace.  Better
than nothing, but not as useful as a panic can be.

> Now why do you need a panic in any case? I don't know DBGPCSR, is this
> a breakpoint register? How do you plan to use it remotely from the CPU
> that detects the lockup?

Panics can trigger extra debugging tools, like my previous examples
kexec and kgdb.

DBGPCSR is the "DeBuG Program Counter Sampling Register".  It is a
memory mapped register available on many (all?) ARM Cortex cpus that
returns a recent PC value for the cpu.  I have used it along with this
patch, and it produces very useful information.