Re: [RFC][PATCH] perf: Rewrite core context handling

[Song Liu]

> On Oct 16, 2018, at 2:50 AM, Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
> 
> On Sat, Oct 13, 2018 at 08:31:37AM +0000, Song Liu wrote:
> 
>> The only suggestion I have right now is on which struct owns which
>> data:
>> 
>> 1. perf_cpu_context owns two perf_event_context: ctx and *task_ctx. 
>>   This is the same as right now. 
> 
>> 2. perf_event_context owns multiple perf_event_pmu_context: 
>>   One perf_event_pmu_context for software groups;
>>   One perf_event_pmu_context for each hardware PMU.
> 
> It does now already, right? Through the pmu_ctx_list we can, given an
> perf_event_context, find all associated perf_event_pmu_context's.

Yes, this is very similar to right now. It is related to #4, #5 below. 
With current patch, perf_cpu_pmu_context is more like the "owner" of
the per CPU perf_event_pmu_context. I feel it is more natural that 
perf_cpu_context is the owner of perf_event_pmu_context, while the
perf_cpu_pmu_context doesn't own anything. 

Again, the difference to current patch is very small. 

> 
>> 3. perf_event_pmu_context owns RB tree of events. Since we don't 
>>   need rotation across multiple hardware PMUs, the rotation is 
>>   within same perf_event_pmu_context.  
> 
> By keeping the RB trees in perf_event_context, we get bigger trees,
> which is more efficient (log(n+m) < log(n) + log(m))
> 
> Also, specifically, it means we only need a single merge sort /
> iteration to schedule in a full context, instead of (again) doing 'n' of
> them.
> 
> Also, given a context and a pmu, it is cheaper for finding the relevant
> events; this is needed for big.little for instance. Something the
> proposed patch doesn't fully flesh out.

Would it be faster if we add a perf_event_pmu_context pointer to the 
perf_event? I think a group on hw PMU-a should never know a group on
hw PMU-b. So some separation would make things simpler. 

Thanks,
Song