Re: [PATCH v9 03/10] perf/x86: Add constraint to create guest LBR event without hw counter

["Xu, Like" <like.xu@intel.com>]

Hi Peter,

First of all, thanks for your comments!

On 2020/4/10 0:37, Peter Zijlstra wrote:
>> diff --git a/arch/x86/events/core.c b/arch/x86/events/core.c
>> index 3bb738f5a472..e919187a0751 100644
>> --- a/arch/x86/events/core.c
>> +++ b/arch/x86/events/core.c
>> @@ -74,7 +74,8 @@ u64 x86_perf_event_update(struct perf_event *event)
>>   	int idx = hwc->idx;
>>   	u64 delta;
>>   
>> -	if (idx == INTEL_PMC_IDX_FIXED_BTS)
>> +	if ((idx == INTEL_PMC_IDX_FIXED_BTS) ||
>> +		(idx == INTEL_PMC_IDX_FIXED_VLBR))
>>   		return 0;
>>   
>>   	/*
>> @@ -1102,7 +1103,8 @@ static inline void x86_assign_hw_event(struct perf_event *event,
>>   	hwc->last_cpu = smp_processor_id();
>>   	hwc->last_tag = ++cpuc->tags[i];
>>   
>> -	if (hwc->idx == INTEL_PMC_IDX_FIXED_BTS) {
>> +	if ((hwc->idx == INTEL_PMC_IDX_FIXED_BTS) ||
>> +		(hwc->idx == INTEL_PMC_IDX_FIXED_VLBR)) {
>>   		hwc->config_base = 0;
>>   		hwc->event_base	= 0;
>>   	} else if (hwc->idx >= INTEL_PMC_IDX_FIXED) {
>> @@ -1233,7 +1235,8 @@ int x86_perf_event_set_period(struct perf_event *event)
>>   	s64 period = hwc->sample_period;
>>   	int ret = 0, idx = hwc->idx;
>>   
>> -	if (idx == INTEL_PMC_IDX_FIXED_BTS)
>> +	if ((idx == INTEL_PMC_IDX_FIXED_BTS) ||
>> +		(idx == INTEL_PMC_IDX_FIXED_VLBR))
>>   		return 0;
>>   
>>   	/*
> That seems unfortunate; can that be >= INTEL_PMC_IDX_FIXED_BTS ? If so,
> that probably wants a comment with the definitions.
>
> Or otherwise check for !hwc->event_base. That should be 0 for both these
> things.
Yes, the !hwc->event_base looks good to me.
>
>> diff --git a/arch/x86/events/intel/core.c b/arch/x86/events/intel/core.c
>> index 3be51aa06e67..901c82032f4a 100644
>> --- a/arch/x86/events/intel/core.c
>> +++ b/arch/x86/events/intel/core.c
>> @@ -2157,6 +2157,9 @@ static void intel_pmu_disable_event(struct perf_event *event)
>>   		return;
>>   	}
>>   
>> +	if (unlikely(hwc->idx == INTEL_PMC_IDX_FIXED_VLBR))
>> +		return;
>> +
> Please check code-gen to see if you can cut down on brancher here;
> there's 4 cases:
>
>   - vlbr
>   - bts
>   - fixed
>   - gp
>
> perhaps you can write it like so:
>
> (also see https://lkml.kernel.org/r/20190828090217.GN2386@hirez.programming.kicks-ass.net )
>
> static void intel_pmu_enable_event(struct perf_event *event)
> {
> 	...
> 	int idx = hwx->idx;
>
> 	if (idx < INTEL_PMC_IDX_FIXED) {
> 		intel_set_masks(event, idx);
> 		__x86_pmu_enable_event(hwc, ARCH_PERFMON_EVENTSEL_ENABLE);
> 	} else if (idx < INTEL_PMC_IDX_FIXED_BTS) {
> 		intel_set_masks(event, idx);
> 		intel_pmu_enable_fixed(event);
> 	} else if (idx == INTEL_PMC_IDX_FIXED_BTS) {
> 		intel_pmu_enable_bts(hwc->config);
> 	}
>
> 	/* nothing for INTEL_PMC_IDX_FIXED_VLBR */
> }
>
> That should sort the branches in order of: gp,fixed,bts,vlbr

Note the current order is: bts, pebs, fixed, gp.

Sure, let me try to refactor it in this way.
>
>>   	cpuc->intel_ctrl_guest_mask &= ~(1ull << hwc->idx);
>>   	cpuc->intel_ctrl_host_mask &= ~(1ull << hwc->idx);
>>   	cpuc->intel_cp_status &= ~(1ull << hwc->idx);
>> @@ -2241,6 +2244,9 @@ static void intel_pmu_enable_event(struct perf_event *event)
>>   		return;
>>   	}
>>   
>> +	if (unlikely(hwc->idx == INTEL_PMC_IDX_FIXED_VLBR))
>> +		return;
>> +
>>   	if (event->attr.exclude_host)
>>   		cpuc->intel_ctrl_guest_mask |= (1ull << hwc->idx);
>>   	if (event->attr.exclude_guest)
> idem.
idem.
>
>> @@ -2595,6 +2601,15 @@ intel_bts_constraints(struct perf_event *event)
>>   	return NULL;
>>   }
>>   
>> +static struct event_constraint *
>> +intel_guest_event_constraints(struct perf_event *event)
>> +{
>> +	if (unlikely(is_guest_lbr_event(event)))
>> +		return &guest_lbr_constraint;
>> +
>> +	return NULL;
>> +}
> This is a mis-nomer, it isn't just any guest_event

Sure,  I'll rename it to intel_guest_lbr_event_constraints()
instead of using it as a unified interface to get all of guest event 
constraints.

>
>> +
>>   static int intel_alt_er(int idx, u64 config)
>>   {
>>   	int alt_idx = idx;
>> @@ -2785,6 +2800,10 @@ __intel_get_event_constraints(struct cpu_hw_events *cpuc, int idx,
>>   {
>>   	struct event_constraint *c;
>>   
>> +	c = intel_guest_event_constraints(event);
>> +	if (c)
>> +		return c;
>> +
>>   	c = intel_bts_constraints(event);
>>   	if (c)
>>   		return c;
>> diff --git a/arch/x86/events/perf_event.h b/arch/x86/events/perf_event.h
>> index 1025bc6eb04f..9a62264a3068 100644
>> --- a/arch/x86/events/perf_event.h
>> +++ b/arch/x86/events/perf_event.h
>> @@ -969,6 +969,20 @@ static inline bool intel_pmu_has_bts(struct perf_event *event)
>>   	return intel_pmu_has_bts_period(event, hwc->sample_period);
>>   }
>>   
>> +static inline bool is_guest_event(struct perf_event *event)
>> +{
>> +	if (event->attr.exclude_host && is_kernel_event(event))
>> +		return true;
>> +	return false;
>> +}
> I don't like this one, what if another in-kernel users generates an
> event with exclude_host set ?
Thanks for the clear attitude.

How about:
- remove the is_guest_event() to avoid potential misuse;
- move all checks into is_guest_lbr_event() and make it dedicated:

static inline bool is_guest_lbr_event(struct perf_event *event)
{
     if (is_kernel_event(event) &&
         event->attr.exclude_host && needs_branch_stack(event))
         return true;
     return false;
}

In this case, it's safe to generate an event with exclude_host set
and also use LBR to count guest or nothing for other in-kernel users
because the intel_guest_lbr_event_constraints() makes LBR exclusive.

For this generic usage, I may rename:
- is_guest_lbr_event() to is_lbr_no_counter_event();
- intel_guest_lbr_event_constraints() to 
intel_lbr_no_counter_event_constraints();

Is this acceptable to you？
If there is anything needs to be improved, please let me know.

>> @@ -989,6 +1003,7 @@ void release_ds_buffers(void);
>>   void reserve_ds_buffers(void);
>>   
>>   extern struct event_constraint bts_constraint;
>> +extern struct event_constraint guest_lbr_constraint;
>>   
>>   void intel_pmu_enable_bts(u64 config);
>>   
>> diff --git a/arch/x86/include/asm/perf_event.h b/arch/x86/include/asm/perf_event.h
>> index e018a1cf604c..674130aca75a 100644
>> --- a/arch/x86/include/asm/perf_event.h
>> +++ b/arch/x86/include/asm/perf_event.h
>> @@ -181,9 +181,19 @@ struct x86_pmu_capability {
>>   #define GLOBAL_STATUS_UNC_OVF				BIT_ULL(61)
>>   #define GLOBAL_STATUS_ASIF				BIT_ULL(60)
>>   #define GLOBAL_STATUS_COUNTERS_FROZEN			BIT_ULL(59)
>> -#define GLOBAL_STATUS_LBRS_FROZEN			BIT_ULL(58)
>> +#define GLOBAL_STATUS_LBRS_FROZEN_BIT			58
>> +#define GLOBAL_STATUS_LBRS_FROZEN			BIT_ULL(GLOBAL_STATUS_LBRS_FROZEN_BIT)
>>   #define GLOBAL_STATUS_TRACE_TOPAPMI			BIT_ULL(55)
>>   
>> +/*
>> + * We model guest LBR event tracing as another fixed-mode PMC like BTS.
>> + *
>> + * We choose bit 58 (LBRS_FROZEN_BIT) which is used to indicate that the LBR
>> + * stack is frozen on a hardware PMI request in the PERF_GLOBAL_STATUS msr,
>> + * and the 59th PMC counter (if any) is not supposed to use it as well.
> Is this saying that STATUS.58 should never be set? I don't really
> understand the language.
My fault, and let me make it more clearly:

We choose bit 58 because it's used to indicate LBR stack frozen state
not like other overflow conditions in the PERF_GLOBAL_STATUS msr,
and it will not be used for any actual fixed events.

>
>> + */
>> +#define INTEL_PMC_IDX_FIXED_VLBR	GLOBAL_STATUS_LBRS_FROZEN_BIT
>> +
>>   /*
>>    * Adaptive PEBS v4
>>    */