Re: [PATCH v2 2/3] perf/ibs: Fix interface via core pmu events

[Ravi Bangoria]

Hi Peter,

On 12-Mar-23 8:24 PM, Peter Zijlstra wrote:
> On Thu, Mar 09, 2023 at 03:41:10PM +0530, Ravi Bangoria wrote:
>> diff --git a/arch/x86/events/amd/core.c b/arch/x86/events/amd/core.c
>> index 8c45b198b62f..81d67b899371 100644
>> --- a/arch/x86/events/amd/core.c
>> +++ b/arch/x86/events/amd/core.c
>> @@ -371,10 +371,15 @@ static inline int amd_has_nb(struct cpu_hw_events *cpuc)
>>  static int amd_pmu_hw_config(struct perf_event *event)
>>  {
>>  	int ret;
>> +	u64 dummy;
>>  
>> -	/* pass precise event sampling to ibs: */
>> -	if (event->attr.precise_ip && get_ibs_caps())
>> -		return -ENOENT;
>> +	if (event->attr.precise_ip) {
>> +		/* pass precise event sampling to ibs by returning -ESRCH */
>> +		if (get_ibs_caps() && !ibs_core_pmu_event(event, &dummy))
>> +			return -ESRCH;
>> +		else
>> +			return -ENOENT;
>> +	}
>>  
>>  	if (has_branch_stack(event) && !x86_pmu.lbr_nr)
>>  		return -EOPNOTSUPP;
>> diff --git a/kernel/events/core.c b/kernel/events/core.c
>> index f79fd8b87f75..e990c71ba34a 100644
>> --- a/kernel/events/core.c
>> +++ b/kernel/events/core.c
>> @@ -11639,18 +11639,26 @@ static struct pmu *perf_init_event(struct perf_event *event)
>>  			goto again;
>>  		}
>>  
>> +		/*
>> +		 * pmu->event_init() should return -ESRCH only when it
>> +		 * wants to forward the event to other pmu.
>> +		 */
>> +		if (ret == -ESRCH)
>> +			goto try_all;
>> +
>>  		if (ret)
>>  			pmu = ERR_PTR(ret);
>>  
>>  		goto unlock;
>>  	}
>>  
>> +try_all:
>>  	list_for_each_entry_rcu(pmu, &pmus, entry, lockdep_is_held(&pmus_srcu)) {
>>  		ret = perf_try_init_event(pmu, event);
>>  		if (!ret)
>>  			goto unlock;
>>  
>> -		if (ret != -ENOENT) {
>> +		if (ret != -ENOENT && ret != -ESRCH) {
>>  			pmu = ERR_PTR(ret);
>>  			goto unlock;
>>  		}
> 
> Urgh.. So amd_pmu_hw_config() knows what PMU it should be but has no
> real way to communicate this, so you make it try all of them again.
> 
> Now, we already have a gruesome hack in there, and I'm thikning you
> should use that instead of adding yet another one. Note:
> 
> 		if (ret == -ENOENT && event->attr.type != type && !extended_type) {
> 			type = event->attr.type;
> 			goto again;
> 
> So if you have amd_pmu_hw_config() do:
> 
> 	event->attr.type = ibs_pmu.type;
> 	return -ENOENT;
> 
> it should all just work no?

IBS driver needs to convert RAW pmu config to IBS config, which it does
based on original event->attr.type. See perf_ibs_precise_event(). This
logic will fail with event->attr.type overwrite.

> 
> And now thinking about this, I'm thinking we can clean up the whole
> swevent mess too, a little something like the below perhaps... Then it
> might just be possible to remove that list_for_each_entry_rcu()
> entirely.
> 
> Hmm?

I'll check this and revert back.

> 
> 
> ---
> diff --git a/kernel/events/core.c b/kernel/events/core.c
> index f79fd8b87f75..26130d1ca40b 100644
> --- a/kernel/events/core.c
> +++ b/kernel/events/core.c
> @@ -9951,6 +9951,9 @@ static void sw_perf_event_destroy(struct perf_event *event)
>  	swevent_hlist_put();
>  }
>  
> +static struct pmu perf_cpu_clock; /* fwd declaration */
> +static struct pmu perf_task_clock;
> +
>  static int perf_swevent_init(struct perf_event *event)
>  {
>  	u64 event_id = event->attr.config;
> @@ -9966,7 +9969,11 @@ static int perf_swevent_init(struct perf_event *event)
>  
>  	switch (event_id) {
>  	case PERF_COUNT_SW_CPU_CLOCK:
> +		event->attr.type = perf_cpu_clock.type;
> +		return -ENOENT;
> +
>  	case PERF_COUNT_SW_TASK_CLOCK:
> +		event->attr.type = perf_task_clock.type;
>  		return -ENOENT;
>  
>  	default:

Thanks,
Ravi