Re: [PATCH v2 bpf-next 1/3] perf: enable branch record for software events

[Song Liu <songliubraving@fb.com>]

> On Aug 30, 2021, at 3:43 AM, Peter Zijlstra <peterz@infradead.org> wrote:
> 
> On Thu, Aug 26, 2021 at 03:13:04PM -0700, Song Liu wrote:
> 
>> Some data on intel_pmu_lbr_disable_all() and perf_pmu_disable().
>> 
>> With this patch, when fexit program triggers, intel_pmu_lbr_disable_all is
>> used to stop the LBR, and the LBR is stopped after 6 extra branch records
>> (see the full trace below). If we replace intel_pmu_lbr_disable_all in
>> intel_pmu_snapshot_branch_stack() with perf_pmu_disable, the LBR is stopped
>> after 19 extra branch records. This is still acceptable for systems with 32
>> LBR entries. But for systems with fewer entries, all the entries before
>> fexit are flushed. Therefore, I suggest we take the short cut and stop LBR
>> asap.
>> 
>> 
>> LBR snapshot captured when we use intel_pmu_lbr_disable_all():
>> 
>> ID: 0 from intel_pmu_lbr_disable_all.part.10+37 to intel_pmu_lbr_disable_all.part.10+72
>> ID: 1 from intel_pmu_lbr_disable_all.part.10+33 to intel_pmu_lbr_disable_all.part.10+37
>> ID: 2 from intel_pmu_snapshot_branch_stack+51 to intel_pmu_lbr_disable_all.part.10+0
>> ID: 3 from __bpf_prog_enter+53 to intel_pmu_snapshot_branch_stack+0
>> ID: 4 from __bpf_prog_enter+8 to __bpf_prog_enter+38
>> ID: 5 from __brk_limit+473903158 to __bpf_prog_enter+0
>> ID: 6 from bpf_fexit_loop_test1+22 to __brk_limit+473903139
>> ID: 7 from bpf_fexit_loop_test1+20 to bpf_fexit_loop_test1+13
>> ID: 8 from bpf_fexit_loop_test1+20 to bpf_fexit_loop_test1+13
>> ID: 9 from bpf_fexit_loop_test1+20 to bpf_fexit_loop_test1+13
>> 
>> 
>> LBR snapshot captured when we use perf_pmu_disable():
>> 
>> ID: 0 from intel_pmu_lbr_disable_all+58 to intel_pmu_lbr_disable_all+93
>> ID: 1 from intel_pmu_lbr_disable_all+54 to intel_pmu_lbr_disable_all+58
>> ID: 2 from intel_pmu_disable_all+15 to intel_pmu_lbr_disable_all+0
>> ID: 3 from intel_pmu_pebs_disable_all+30 to intel_pmu_disable_all+15
>> ID: 4 from intel_pmu_disable_all+10 to intel_pmu_pebs_disable_all+0
>> ID: 5 from __intel_pmu_disable_all+49 to intel_pmu_disable_all+10
>> ID: 6 from intel_pmu_disable_all+5 to __intel_pmu_disable_all+0
>> ID: 7 from x86_pmu_disable+61 to intel_pmu_disable_all+0
>> ID: 8 from x86_pmu_disable+38 to x86_pmu_disable+41
>> ID: 9 from __x86_indirect_thunk_rax+16 to x86_pmu_disable+0
>> ID: 10 from __x86_indirect_thunk_rax+0 to __x86_indirect_thunk_rax+12
>> ID: 11 from perf_pmu_disable.part.122+4 to __x86_indirect_thunk_rax+0
>> ID: 12 from perf_pmu_disable+23 to perf_pmu_disable.part.122+0
>> ID: 13 from intel_pmu_snapshot_branch_stack+45 to perf_pmu_disable+0
>> ID: 14 from x86_get_pmu+35 to intel_pmu_snapshot_branch_stack+39
>> ID: 15 from intel_pmu_snapshot_branch_stack+34 to x86_get_pmu+0
>> ID: 16 from __bpf_prog_enter+53 to intel_pmu_snapshot_branch_stack+0
>> ID: 17 from __bpf_prog_enter+8 to __bpf_prog_enter+38
>> ID: 18 from __brk_limit+478056502 to __bpf_prog_enter+0
>> ID: 19 from bpf_fexit_loop_test1+22 to __brk_limit+478056483
>> ID: 20 from bpf_fexit_loop_test1+20 to bpf_fexit_loop_test1+13
>> ID: 21 from bpf_fexit_loop_test1+20 to bpf_fexit_loop_test1+13
> 
> Well, if you're willing to do something like:
> 
>> diff --git a/arch/x86/events/intel/core.c b/arch/x86/events/intel/core.c
>> index ac6fd2dabf6a2..a29649e7241cc 100644
>> --- a/arch/x86/events/intel/core.c
>> +++ b/arch/x86/events/intel/core.c
>> @@ -6283,8 +6283,11 @@ __init int intel_pmu_init(void)
>> 			x86_pmu.lbr_nr = 0;
>> 	}
>> 
>> -	if (x86_pmu.lbr_nr)
>> +	if (x86_pmu.lbr_nr) {
>> 		pr_cont("%d-deep LBR, ", x86_pmu.lbr_nr);
> 
> 		if (x86_pmu.disable_all == intel_pmu_disable_all)
> 
>> +		static_call_update(perf_snapshot_branch_stack,
>> +				   intel_pmu_snapshot_branch_stack);
>> +	}
>> 
>> 	intel_pmu_check_extra_regs(x86_pmu.extra_regs);
>> 
>> diff --git a/arch/x86/events/intel/lbr.c b/arch/x86/events/intel/lbr.c
>> index 9e6d6eaeb4cb6..7d4fe1d6e79ff 100644
>> --- a/arch/x86/events/intel/lbr.c
>> +++ b/arch/x86/events/intel/lbr.c
>> @@ -1862,3 +1862,16 @@ EXPORT_SYMBOL_GPL(x86_perf_get_lbr);
>> struct event_constraint vlbr_constraint =
>> 	__EVENT_CONSTRAINT(INTEL_FIXED_VLBR_EVENT, (1ULL << INTEL_PMC_IDX_FIXED_VLBR),
>> 			  FIXED_EVENT_FLAGS, 1, 0, PERF_X86_EVENT_LBR_SELECT);
>> +
>> +int intel_pmu_snapshot_branch_stack(struct perf_branch_snapshot *br_snapshot)
>> +{
>> +	struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
>> +
>> +	intel_pmu_lbr_disable_all();
>> +	intel_pmu_lbr_read();
>> +	memcpy(br_snapshot->entries, cpuc->lbr_entries,
>> +	       sizeof(struct perf_branch_entry) * x86_pmu.lbr_nr);
>> +	br_snapshot->nr = x86_pmu.lbr_nr;
>> +	intel_pmu_lbr_enable_all(false);
>> +	return 0;
>> +}
> 
> Then the above can assume perfmon > v2 and we can either inline
> __intel_pmu_disable_all() or simply do the
> wrmsrl(MSR_CORE_PERF_GLOBAL_CTRL).

I think can do perfmon > v2 only. 

> 
> One thing that needs checking, intel_pmu_disable_all() also clears
> MSR_IA32_PEBS_ENABLE, is that really needed if we just want to inhibit
> PMIs ? That is, will the PEBS machinery still trigger PMI if GLOBAL_CTRL
> == 0 ?

Actually, can we do something like:

static void intel_pmu_disable_all(void)
{
        intel_pmu_lbr_disable_all();   /* moved to the beginning */
        __intel_pmu_disable_all();
        intel_pmu_pebs_disable_all();
}

int intel_pmu_snapshot_branch_stack(struct perf_branch_snapshot *br_snapshot)
{
        struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);

	
        intel_pmu_disable_all();   /* call full pmu_disable */
        intel_pmu_lbr_read();
        memcpy(br_snapshot->entries, cpuc->lbr_entries,
               sizeof(struct perf_branch_entry) * x86_pmu.lbr_nr);
        br_snapshot->nr = x86_pmu.lbr_nr;

        intel_pmu_enable_all(false);
        return 0;
}

In this way, we still call intel_pmu_disable_all(), but since LBR is disabled 
at the beginning of it, we would not flush too many LBR entries. 

Thanks,
Song