Hi Robin,
On Tue, 20 Jun 2023, Robin Murphy wrote:
On 07/06/2023 9:31 am, Ilkka Koskinen wrote:
Some platforms may use e.g. different filtering mechanism and, thus,
may need different way to validate the events and group.
Signed-off-by: Ilkka Koskinen <ilkka@xxxxxxxxxxxxxxxxxxxxxx>
---
drivers/perf/arm_cspmu/arm_cspmu.c | 13 ++++++++++++-
drivers/perf/arm_cspmu/arm_cspmu.h | 4 ++++
2 files changed, 16 insertions(+), 1 deletion(-)
diff --git a/drivers/perf/arm_cspmu/arm_cspmu.c
b/drivers/perf/arm_cspmu/arm_cspmu.c
index 72ca4f56347c..9021d1878250 100644
--- a/drivers/perf/arm_cspmu/arm_cspmu.c
+++ b/drivers/perf/arm_cspmu/arm_cspmu.c
@@ -559,7 +559,7 @@ static void arm_cspmu_disable(struct pmu *pmu)
static int arm_cspmu_get_event_idx(struct arm_cspmu_hw_events *hw_events,
struct perf_event *event)
{
- int idx;
+ int idx, ret;
struct arm_cspmu *cspmu = to_arm_cspmu(event->pmu);
if (supports_cycle_counter(cspmu)) {
@@ -593,6 +593,12 @@ static int arm_cspmu_get_event_idx(struct
arm_cspmu_hw_events *hw_events,
if (idx >= cspmu->num_logical_ctrs)
return -EAGAIN;
+ if (cspmu->impl.ops.validate_event) {
+ ret = cspmu->impl.ops.validate_event(cspmu, event);
+ if (ret)
+ return ret;
+ }
+
set_bit(idx, hw_events->used_ctrs);
return idx;
@@ -618,6 +624,7 @@ static bool arm_cspmu_validate_event(struct pmu *pmu,
*/
static bool arm_cspmu_validate_group(struct perf_event *event)
{
+ struct arm_cspmu *cspmu = to_arm_cspmu(event->pmu);
struct perf_event *sibling, *leader = event->group_leader;
struct arm_cspmu_hw_events fake_hw_events;
@@ -635,6 +642,10 @@ static bool arm_cspmu_validate_group(struct
perf_event *event)
return false;
}
+ if (cspmu->impl.ops.validate_group &&
+ cspmu->impl.ops.validate_group(event))
+ return false;
Hmm, this means that any driver wanting to use it has to duplicate all the
group iteration logic, which isn't ideal. More than that, though, the way
you've implemented it in patch #4 I'm not sure even does anything, since it
only appears to be repeating the same checks that already happen in this
path:
arm_csmpu_validate_group()
arm_cspmu_validate_event()
arm_cspmu_get_event_idx()
ops.validate_event() -> ampere_cspmu_validate_params()
so there's no need for the ops.validate_group hook to just call
ampere_cspmu_validate_params() a second time when it's guaranteed to succeed
(because otherwise we'd have bailed out already).
Yeah, I took another look how the framework really does it and you're
absolutely correct, it's totally unnecessary.
I think what we want overall is an "is this event config valid at all" hook
from arm_cspmu_event_init() (which we don't really need to implement yet
unless you want to start sanity-checking your actual rank/bank/threshold
values), plus an "is this event schedulable in the given PMU context" hook
from arm_cspmu_get_event_idx(), which should serve for both group validation
via the fake context in event_init and actual scheduling in the real context
in add.
Ah, that's true. I can already verify the group event has the same
rank/bank/threshold settings as the group leader in ops.validate_event().
Thus, one hook seems enough.
I fix and rebase the patchset.
Cheers, Ilkka
Thanks,
Robin.
+
return arm_cspmu_validate_event(event->pmu, &fake_hw_events, event);
}
diff --git a/drivers/perf/arm_cspmu/arm_cspmu.h
b/drivers/perf/arm_cspmu/arm_cspmu.h
index f89ae2077164..291cedb196ea 100644
--- a/drivers/perf/arm_cspmu/arm_cspmu.h
+++ b/drivers/perf/arm_cspmu/arm_cspmu.h
@@ -106,6 +106,10 @@ struct arm_cspmu_impl_ops {
void (*set_ev_filter)(struct arm_cspmu *cspmu,
struct hw_perf_event *hwc,
u32 filter);
+ /* Implementation specific group validation */
+ int (*validate_group)(struct perf_event *event);
+ /* Implementation specific event validation */
+ int (*validate_event)(struct arm_cspmu *cspmu, struct perf_event
*new);
/* Hide/show unsupported events */
umode_t (*event_attr_is_visible)(struct kobject *kobj,
struct attribute *attr, int unused);
