| Message ID | 20250202-arm-brbe-v19-v19-2-1c1300802385@kernel.org (mailing list archive) |
|---|---|
| State | New |
| Headers | show |
| Series | arm64/perf: Enable branch stack sampling | expand |
On 2/3/25 06:12, Rob Herring (Arm) wrote: > From: Mark Rutland <mark.rutland@arm.com> > > Currently armpmu_add() tries to handle a newly-allocated counter having > a stale associated event, but this should not be possible, and if this A stale associated event ? Does that mean hw_events->events[idx] still points to a valid event even though counter idx has already been freed up and allocated to a new event. > were to happen the current mitigation is insufficient and potentially > expensive. It would be better to warn if we encounter the impossible > case. Makes sense. > > Calls to pmu::add() and pmu::del() are serialized by the core perf code, > and armpmu_del() clears the relevant slot in pmu_hw_events::events[] > before clearing the bit in pmu_hw_events::used_mask such that the > counter can be reallocated. Thus when armpmu_add() allocates a counter > index from pmu_hw_events::used_mask, it should not be possible to observe > a stale even in pmu_hw_events::events[] unless either > pmu_hw_events::used_mask or pmu_hw_events::events[] have been corrupted. > > If this were to happen, we'd end up with two events with the same > event->hw.idx, which would clash with each other during reprogramming, > deletion, etc, and produce bogus results. Add a WARN_ON_ONCE() for this > case so that we can detect if this ever occurs in practice. Agreed. > > That possiblity aside, there's no need to call arm_pmu::disable(event) s/possiblity/possibility > for the new event. The PMU reset code initialises the counter in a > disabled state, and armpmu_del() will disable the counter before it can > be reused. Remove the redundant disable. > > Signed-off-by: Mark Rutland <mark.rutland@arm.com> > Signed-off-by: Rob Herring (Arm) <robh@kernel.org> > --- > drivers/perf/arm_pmu.c | 8 +++----- > 1 file changed, 3 insertions(+), 5 deletions(-) > > diff --git a/drivers/perf/arm_pmu.c b/drivers/perf/arm_pmu.c > index 398cce3d76fc..2f33e69a8caf 100644 > --- a/drivers/perf/arm_pmu.c > +++ b/drivers/perf/arm_pmu.c > @@ -342,12 +342,10 @@ armpmu_add(struct perf_event *event, int flags) > if (idx < 0) > return idx; > > - /* > - * If there is an event in the counter we are going to use then make > - * sure it is disabled. > - */ > + /* The newly-allocated counter should be empty */ Should this comment also include what happens when two events some how end up using the same 'event->hw.idx' as mentioned in the commit message, just to make things clearer. > + WARN_ON_ONCE(hw_events->events[idx]); > + > event->hw.idx = idx; > - armpmu->disable(event); > hw_events->events[idx] = event; > > hwc->state = PERF_HES_STOPPED | PERF_HES_UPTODATE; > Otherwise LGTM. Reviewed-by: Anshuman Khandual <anshuman.khandual@arm.com>
diff --git a/drivers/perf/arm_pmu.c b/drivers/perf/arm_pmu.c index 398cce3d76fc..2f33e69a8caf 100644 --- a/drivers/perf/arm_pmu.c +++ b/drivers/perf/arm_pmu.c @@ -342,12 +342,10 @@ armpmu_add(struct perf_event *event, int flags) if (idx < 0) return idx; - /* - * If there is an event in the counter we are going to use then make - * sure it is disabled. - */ + /* The newly-allocated counter should be empty */ + WARN_ON_ONCE(hw_events->events[idx]); + event->hw.idx = idx; - armpmu->disable(event); hw_events->events[idx] = event; hwc->state = PERF_HES_STOPPED | PERF_HES_UPTODATE;