Re: [PATCH] perf/x86: fix wrong assumption that LBR is only useful for sampling events

[Liang, Kan]

On 2024-09-05 4:33 p.m., Andrii Nakryiko wrote:
> On Thu, Sep 5, 2024 at 1:29 PM Liang, Kan <kan.liang@xxxxxxxxxxxxxxx> wrote:
>>
>>
>>
>> On 2024-09-05 4:22 p.m., Andrii Nakryiko wrote:
>>> On Thu, Sep 5, 2024 at 12:21 PM Liang, Kan <kan.liang@xxxxxxxxxxxxxxx> wrote:
>>>>
>>>>
>>>>
>>>> On 2024-09-05 2:00 p.m., Andrii Nakryiko wrote:
>>>>> It's incorrect to assume that LBR can/should only be used with sampling
>>>>> events. BPF subsystem provides bpf_get_branch_snapshot() BPF helper,
>>>>> which expects a properly setup and activated perf event which allows
>>>>> kernel to capture LBR data.
>>>>>
>>>>> For instance, retsnoop tool ([0]) makes an extensive use of this
>>>>> functionality and sets up perf event as follows:
>>>>>
>>>>>       struct perf_event_attr attr;
>>>>>
>>>>>       memset(&attr, 0, sizeof(attr));
>>>>>       attr.size = sizeof(attr);
>>>>>       attr.type = PERF_TYPE_HARDWARE;
>>>>>       attr.config = PERF_COUNT_HW_CPU_CYCLES;
>>>>>       attr.sample_type = PERF_SAMPLE_BRANCH_STACK;
>>>>>       attr.branch_sample_type = PERF_SAMPLE_BRANCH_KERNEL;
>>>>>
>>>>> Commit referenced in Fixes tag broke this setup by making invalid assumption
>>>>> that LBR is useful only for sampling events. Remove that assumption.
>>>>>
>>>>> Note, earlier we removed a similar assumption on AMD side of LBR support,
>>>>> see [1] for details.
>>>>>
>>>>>   [0] https://github.com/anakryiko/retsnoop
>>>>>   [1] 9794563d4d05 ("perf/x86/amd: Don't reject non-sampling events with configured LBR")
>>>>>
>>>>> Cc: stable@xxxxxxxxxxxxxxx # 6.8+
>>>>> Fixes: 85846b27072d ("perf/x86: Add PERF_X86_EVENT_NEEDS_BRANCH_STACK flag")
>>>>> Signed-off-by: Andrii Nakryiko <andrii@xxxxxxxxxx>
>>>>> ---
>>>>>  arch/x86/events/intel/core.c | 2 +-
>>>>>  1 file changed, 1 insertion(+), 1 deletion(-)
>>>>>
>>>>> diff --git a/arch/x86/events/intel/core.c b/arch/x86/events/intel/core.c
>>>>> index 9e519d8a810a..f82a342b8852 100644
>>>>> --- a/arch/x86/events/intel/core.c
>>>>> +++ b/arch/x86/events/intel/core.c
>>>>> @@ -3972,7 +3972,7 @@ static int intel_pmu_hw_config(struct perf_event *event)
>>>>>                       x86_pmu.pebs_aliases(event);
>>>>>       }
>>>>>
>>>>> -     if (needs_branch_stack(event) && is_sampling_event(event))
>>>>> +     if (needs_branch_stack(event))
>>>>>               event->hw.flags  |= PERF_X86_EVENT_NEEDS_BRANCH_STACK;
>>>>
>>>> To limit the LBR for a sampling event is to avoid unnecessary branch
>>>> stack setup for a counting event in the sample read. The above change
>>>> should break the sample read case.
>>>>
>>>> How about the below patch (not test)? Is it good enough for the BPF usage?
>>>>
>>>> diff --git a/arch/x86/events/intel/core.c b/arch/x86/events/intel/core.c
>>>> index 0c9c2706d4ec..8d67cbda916b 100644
>>>> --- a/arch/x86/events/intel/core.c
>>>> +++ b/arch/x86/events/intel/core.c
>>>> @@ -3972,8 +3972,12 @@ static int intel_pmu_hw_config(struct perf_event
>>>> *event)
>>>>                 x86_pmu.pebs_aliases(event);
>>>>         }
>>>>
>>>> -       if (needs_branch_stack(event) && is_sampling_event(event))
>>>> -               event->hw.flags  |= PERF_X86_EVENT_NEEDS_BRANCH_STACK;
>>>> +       if (needs_branch_stack(event)) {
>>>> +               /* Avoid branch stack setup for counting events in SAMPLE READ */
>>>> +               if (is_sampling_event(event) ||
>>>> +                   !(event->attr.sample_type & PERF_SAMPLE_READ))
>>>> +                       event->hw.flags  |= PERF_X86_EVENT_NEEDS_BRANCH_STACK;
>>>> +       }
>>>>
>>>
>>> I'm sure it will be fine for my use case, as I set only
>>> PERF_SAMPLE_BRANCH_STACK.
>>>
>>> But I'll leave it up to perf subsystem experts to decide if this
>>> condition makes sense, because looking at what PERF_SAMPLE_READ is:
>>>
>>>           PERF_SAMPLE_READ
>>>                  Record counter values for all events in a group,
>>>                  not just the group leader.
>>>
>>> It's not clear why this would disable LBR, if specified.
>>
>> It only disables the counting event with SAMPLE_READ, since LBR is only
>> read in the sampling event's overflow.
>>
> 
> Ok, sounds good! Would you like to send a proper patch with your
> proposed changes?

The patch has been posted. Please give it a try.
https://lore.kernel.org/lkml/20240909155848.326640-1-kan.liang@xxxxxxxxxxxxxxx/

Thanks,
Kan
> 
>> Thanks,
>> Kan
>>>
>>>>         if (branch_sample_counters(event)) {
>>>>                 struct perf_event *leader, *sibling;
>>>>
>>>>
>>>> Thanks,
>>>> Kan
>>>>>
>>>>>       if (branch_sample_counters(event)) {
>