Re: [RFC PATCH 5/6] bpf: Implement function bpf_read_pmu() that get the selected hardware PMU conuter

["Wangnan (F)" <wangnan0@huawei.com>]

On 2015/7/17 20:18, Peter Zijlstra wrote:
> On Fri, Jul 17, 2015 at 08:01:07PM +0800, Wangnan (F) wrote:
>>
>> On 2015/7/17 19:56, Peter Zijlstra wrote:
>>> On Fri, Jul 17, 2015 at 01:55:05PM +0200, Peter Zijlstra wrote:
>>>> On Fri, Jul 17, 2015 at 07:45:02PM +0800, Wangnan (F) wrote:
>>>>
>>>>>> Depends on what all you need, if you need full perf events to work then
>>>>>> yes perf_event_read_value() is your only option.
>>>>>>
>>>>>> But note that that requires scheduling, so you cannot actually use it
>>>>>> for tracing purposes etc..
>>>>> What you mean "full perf events"? Even with your code some event still not
>>>>> work?
>>>> The code I posted only works for events that do not have inherit set.
>>>> And only works from IRQ/NMI context for events that monitor the current
>>>> task or the current CPU (although that needs a little extra code still).
>>>>
>>>> Anything else and it does not work (correctly).
>>> Scratch that from NMI, for that to work we need more magic still.
>> The scheduling you said is caused by
>>
>> mutex_lock(&event->child_mutex)
>>
>> right?
>>
>> What about replacing it to mutex_trylock() and simply return an error
>> if it read from a BPF program?
> That is vile and unreliable.
>
> I think you really want to put very strict limits on what kind of events
> you accept, or create the events yourself.
>
I think we can check the limitation in BPF program. What about this:

  event must on current CPU or must be on current process. If not,
  bpf_read_pmu() should simply return an error.

With current design it is easy to implement, and users can still control
it through bpf map.

But what if we really want cross-cpu PMU accessing? Impossible?

Thank you.