Re: [Patch v7 14/24] perf/x86/intel: Account lbr_pebs_users for arch-PEBS events

From: Mi, Dapeng

Date: Wed Nov 18 2026 - 01:28:54 EST



On 11/18/2026 12:10 PM, bot+bpf-ci@xxxxxxxxxx wrote:
>> perf/x86/intel: Account lbr_pebs_users for arch-PEBS events
> TL;DR:
> - On arch-PEBS, a non-PEBS branch-counter sibling event may report a stale
> branch stack with zero counters until a later patch fixes it; consider
> reordering
> - Fixes tag may point to the wrong commit
>
>> perf/x86/intel: Account lbr_pebs_users for arch-PEBS events
>>
>> lbr_pebs_users tracks how many PEBS events are currently using LBR.
>> It is updated for legacy PEBS events, but arch-PEBS events were missing
>> from that accounting.
>>
>> Add the missing lbr_pebs_users accounting for arch-PEBS as well.
>>
>> Fixes: d243d0bb64af ("perf/x86/intel: Initialize architectural PEBS")
>> [ ... ]
>> Signed-off-by: Dapeng Mi <dapeng1.mi@xxxxxxxxxxxxxxx>
> d243d0bb64af only parses the arch-PEBS CPUID leaves and adds the
> enable/disable callbacks, it does not make arch-PEBS use LBR. The
> ARCH_PEBS_LBR handling and PERF_SAMPLE_BRANCH_STACK in large_pebs_flags
> came with 52448a0a7390, which predates this series.
>
> Should the tag point there instead:
>
> Fixes: 52448a0a7390 ("perf/x86/intel: Setup PEBS data configuration and enable legacy groups")

Not exactly. The arch-PEBS feature including the arch-PEBS assisted LBR
snapshotting is supported by a bunch of patches instead of a single patch,
e.g., the commit d243d0bb64af enumerates if the arch-PEBS has the
capability to support LBR snapshot, it's a more appropriate and accurate
commit comparing with 52448a0a7390.


>
>> diff --git a/arch/x86/events/intel/lbr.c b/arch/x86/events/intel/lbr.c
>> --- a/arch/x86/events/intel/lbr.c
>> +++ b/arch/x86/events/intel/lbr.c
> [ ... ]
>
>> @@ -616,7 +616,7 @@ void intel_pmu_lbr_add(struct perf_event *event)
>> * be 'new'. Conversely, a new event can get installed through the
>> * context switch path for the first time.
>> */
>> - if (x86_pmu.intel_cap.pebs_baseline && event->attr.precise_ip > 0)
>> + if (intel_pmu_has_pebs_lbr(event->pmu) && event->attr.precise_ip > 0)
>> cpuc->lbr_pebs_users++;
>> perf_sched_cb_inc(event->pmu);
>> if (!cpuc->lbr_users++ && !event->total_time_running)
> Can this make a non-PEBS branch counter event report a stale branch stack
> on arch-PEBS platforms?
>
> With lbr_pebs_users now non-zero on arch-PEBS, the existing shortcut in
> intel_pmu_handle_irq() is taken:
>
> if (cpuc->lbr_users != cpuc->lbr_pebs_users)
> intel_pmu_lbr_read(cpu_lbr_stack(cpuc), cpuc->lbr_counters);
>
> That shortcut assumes every consumer of the LBR stack is counted in
> lbr_users. An event whose branch_sample_type is only
> PERF_SAMPLE_BRANCH_COUNTERS plus PLM bits has
> PERF_X86_EVENT_NEEDS_BRANCH_STACK cleared in intel_pmu_hw_config(), so
> intel_pmu_add_event() never calls intel_pmu_lbr_add() for it and it is not
> counted in lbr_users. It still has has_branch_stack() set and samples
> through the non-PEBS path in handle_pmi_common():
>
> if (has_branch_stack(event))
> intel_pmu_lbr_save_brstack(&data, cpuc, event);
>
> Consider a group on an arch-PEBS CPU with ARCH_PEBS_LBR and PMU_FL_BR_CNTR,
> with a precise leader using branch_type=any,counter and a non-precise
> sampling sibling using only branch_type=counter. After this patch
> lbr_users == 1 and lbr_pebs_users == 1. If only the sibling overflows,
> intel_pmu_lbr_read() is skipped, and intel_pmu_lbr_save_brstack() ->
> intel_pmu_lbr_counters_reorder() works on whatever cpu_lbr_stack(cpuc)
> last held. That is either nr == 0 after intel_pmu_lbr_reset(), or the
> leader's last PEBS-record snapshot from intel_pmu_store_pebs_lbrs(),
> whose raw counters in entries[i].reserved were already zeroed when the
> leader's sample was saved.
>
> So the sibling would report a stale branch stack with all branch counters
> zero, where before this patch it got the live LBR state. The same
> problem already exists for legacy adaptive PEBS (pebs_baseline), and this
> patch extends it to arch-PEBS.
>
> This looks to be addressed later in the series by 0479605ddd39
> ("perf/x86/intel: Account LBR user counts for pure branch counter
> events"), and 9edb70d448eb ("perf/x86/intel: Refresh LBR stack for
> non-PEBS LBR events when needed") replaces the lbr_users != lbr_pebs_users
> check at PMI entry with an on-demand re-read. Would it make sense to
> order 0479605ddd39 before this patch so the intermediate state does not
> regress?

Hmm, I don't think the patch order matters. If we put the commit
9edb70d448eb ahead before this patch, the whole arch-PEBS assisted LBR
snapshot won't be counted without this patch regardless of it's a pure
branch counter event. It seems worse. 

Thanks.



>
>
> ---
> AI reviewed your patch. Please fix the bug or email reply why it's not a bug.
> See: https://github.com/kernel-patches/vmtest/blob/master/ci/claude/README.md
>
> CI run summary: https://github.com/kernel-patches/bpf/actions/runs/44231525404