Re: [tip:perf/urgent] perf/core: Fix the perf_cpu_time_max_percent check

[Tan Xiaojun]

On 2017/2/25 17:51, Peter Zijlstra wrote:
> On Sat, Feb 25, 2017 at 04:10:37PM +0800, Tan Xiaojun wrote:
> 
>> Recently I was using perf_fuzzer for testing in Hisilicon
>> D03/D05(arm64, linux-4.10-rc1).
>>
>> As we know perf_fuzzer will write a random value to procfs interface
>> of perf event(like sysctl_perf_cpu_time_max_percent). The value may be
>> 0 or 100, and I get logs like below:
>>
>> ----------------------------------
>> [ 4046.358811] perf: Dynamic interrupt throttling disabled, can hang your system!
>> ----------------------------------
>>
>> Most of the time, there is no problem, and the perf_fuzzer test can
>> end without any warings or errors.  But there is a small probability
>> that triggers the RCU and watchdog (The log is attached at the end).
>> It hungs after local_irq_enable() in __do_softirq.
>>
>> I think this is due to the dynamic interrupt throttling disabled and
>> too many hardware interruptions come. So I limit the
>> sysctl_perf_cpu_time_max_percent can only be set 1 to 99 in the kernel
>> codes. I test more than 20 times in D03, and there are no errors or
>> warnings in the test.
>>
>> So I want to ask:
>>
>> 1)Is it a problem or not? (It has already given you a warning.)
>>
>> 2)If it is, where we will fix it more appropriate, perf_fuzzer(not set
>> 0 or 100) or kernel(limit 1 to 99), or maybe it is the bug of
>> hardware(too many hardware interruptions)?
> 
> I think the best would be if the fuzzer would not set 0,100, those are
> clearly 'unsafe' settings and you pretty much get to keep the pieces.
> 
> I would like to preserve these settings for people that 'know' what
> they're doing and are willing to take the risk, but clearly, when you
> take the guard-rails off, things can come apart.
> 

OK. It makes sense, I agree.

Thank you for your answer.
Xiaojun.

> .
>