[PATCH] ARM: imx6q: work around faulty PMU irq routing

[Dirk Behme]

On 29.04.2014 11:55, Lucas Stach wrote:
> Am Dienstag, den 29.04.2014, 13:28 +0800 schrieb Shawn Guo:
>> On Fri, Apr 25, 2014 at 11:13:25AM +0200, Lucas Stach wrote:
>>> Am Freitag, den 25.04.2014, 07:37 +0200 schrieb Dirk Behme:
>>>> On 24.04.2014 22:23, Lucas Stach wrote:
>>>>> The i.MX6 PMU has a design errata where the interrupts of all cores are
>>>>> wired together into a single irq line. To work around this we have to
>>>>> bounce the interrupt around all cores until we find the one where the PMU
>>>>> counter overflow has happened.
>>>>>
>>>>> This causes the perf measurements to be less accurate and we can't really
>>>>> handle the case where two cores fire a PMU irq at the same time. The
>>>>> implemented woraround makes perf at least somewhat useable on imx6 SoCs
>>>>> with more than one core.
>>>>>
> [...]
>>>> Do you have anything like a test case which shows that it works (at
>>>> least better) on a !single core with this patch? Compared to a
>>>> non-patched system?
>>>>
>>> Without this patch, running perf top completely kills the system on
>>> i.MX6q, most likely because of the sheer number of spurious interrupts
>>> hitting the system from 4 cores. Even the spurious killer doesn't work
>>> sometimes, so perf is completely busted right now.
>>>
>>> With this patch perf has to reduce the sample frequency in order to
>>> compensate the added irq latency, but at least we get some plausible
>>> numbers out. Though I won't take any blame if the amount of salt you
>>> have to apply while looking at those numbers is already a deadly
>>> dose. ;)
>>>
>>> I don't yet have any numbers on how accurate the measurement is, but at
>>> least things didn't look completely off.
>>
>> If it cannot provide correct/accurate data, I'd say let's not fake it
>> to, and just let it be completely broken there, so that people can be
>> aware of the brokenness, and not take inaccurate data as accurate one.
>>
> The data isn't bogus, it just isn't as accurate as it could be with a
> properly working PMU. I'll run some tests with a defined load on
> Solo/Quad to see how far the measurements are off. I'm fine with holding
> this patch until then.
>
> But the thing is this patch also fixes a serious userspace triggerable
> DoS on i.MX6q. Just running perf top completely locks up the system
> because of the sheer number of stray irqs. This isn't the case anymore
> with this patch applied.
>
> Maybe we can just print a warning into dmesg to make the users aware of
> the imprecise measurement.

Yes, I think this sounds like a good compromise :)

Best regards

Dirk