Re: [PATCH] perf/x86/intel: Fix n_metric for the canceled group

From: Liang, Kan
Date: Fri Oct 02 2020 - 09:16:22 EST




On 10/2/2020 7:02 AM, Peter Zijlstra wrote:
On Wed, Sep 30, 2020 at 07:29:35AM -0700, kan.liang@xxxxxxxxxxxxxxx wrote:
From: Kan Liang <kan.liang@xxxxxxxxxxxxxxx>

When a group that has TopDown members is failed to be scheduled, any
later TopDown groups will not return valid values.

Here is an example.

A background perf that occupies all the GP counters and the fixed
counter 1.
$perf stat -e "{cycles,cycles,cycles,cycles,cycles,cycles,cycles,
cycles,cycles}:D" -a

A user monitors a TopDown group. It works well, because the fixed
counter 3 and the PERF_METRICS are available.
$perf stat -x, --topdown -- ./workload
retiring,bad speculation,frontend bound,backend bound,
18.0,16.1,40.4,25.5,

Then the user tries to monitor a group that has TopDown members.
Because of the cycles event, the group is failed to be scheduled.
$perf stat -x, -e '{slots,topdown-retiring,topdown-be-bound,
topdown-fe-bound,topdown-bad-spec,cycles}'
-- ./workload
<not counted>,,slots,0,0.00,,
<not counted>,,topdown-retiring,0,0.00,,
<not counted>,,topdown-be-bound,0,0.00,,
<not counted>,,topdown-fe-bound,0,0.00,,
<not counted>,,topdown-bad-spec,0,0.00,,
<not counted>,,cycles,0,0.00,,

The user tries to monitor a TopDown group again. It doesn't work anymore.
$perf stat -x, --topdown -- ./workload

,,,,,

In a txn, cancel_txn() is to truncate the event_list for a canceled
group and update the number of events added in this transaction.
However, the number of TopDown events added in this transaction is not
updated. The kernel will probably fail to add new Topdown events.

Check if the canceled group has Topdown events. If so, subtract the
TopDown events from n_metric accordingly.

Fixes: 7b2c05a15d29 ("perf/x86/intel: Generic support for hardware TopDown metrics")
Reported-by: Andi Kleen <ak@xxxxxxxxxxxxxxx>
Reviewed-by: Andi Kleen <ak@xxxxxxxxxxxxxxx>
Signed-off-by: Kan Liang <kan.liang@xxxxxxxxxxxxxxx>
---
arch/x86/events/core.c | 10 ++++++++++
1 file changed, 10 insertions(+)

diff --git a/arch/x86/events/core.c b/arch/x86/events/core.c
index 0f3d01562ded..4cb3ccbe2d62 100644
--- a/arch/x86/events/core.c
+++ b/arch/x86/events/core.c
@@ -2017,6 +2017,7 @@ static void x86_pmu_cancel_txn(struct pmu *pmu)
{
unsigned int txn_flags;
struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events);
+ int i;
WARN_ON_ONCE(!cpuc->txn_flags); /* no txn in flight */
@@ -2031,6 +2032,15 @@ static void x86_pmu_cancel_txn(struct pmu *pmu)
*/
__this_cpu_sub(cpu_hw_events.n_added, __this_cpu_read(cpu_hw_events.n_txn));
__this_cpu_sub(cpu_hw_events.n_events, __this_cpu_read(cpu_hw_events.n_txn));
+
+ /* Subtract Topdown events in the canceled group from n_metric */
+ if (x86_pmu.intel_cap.perf_metrics && cpuc->n_metric) {
+ for (i = 0; i < cpuc->n_txn; i++) {
+ if (is_metric_event(cpuc->event_list[i + cpuc->n_events]))
+ __this_cpu_dec(cpu_hw_events.n_metric);
+ }
+ WARN_ON_ONCE(__this_cpu_read(cpu_hw_events.n_metric) < 0);
+ }
perf_pmu_enable(pmu);
}


Urgh, I'd much rather we add n_txn_metric. But also, while looking at
this, don't we have the same problem with n_pair ?

Something like this perhaps...


Sure. For the perf metric, the below patch fixes the problem.

Tested-by: Kan Liang <kan.liang@xxxxxxxxxxxxxxx>

Thanks,
Kan

---
diff --git a/arch/x86/events/core.c b/arch/x86/events/core.c
index 757e49755e7c..9b7792c0b6fb 100644
--- a/arch/x86/events/core.c
+++ b/arch/x86/events/core.c
@@ -1066,6 +1066,7 @@ static int add_nr_metric_event(struct cpu_hw_events *cpuc,
if (cpuc->n_metric == INTEL_TD_METRIC_NUM)
return -EINVAL;
cpuc->n_metric++;
+ cpuc->n_txn_metric++;
}
return 0;
@@ -1089,8 +1090,10 @@ static int collect_event(struct cpu_hw_events *cpuc, struct perf_event *event,
return -EINVAL;
cpuc->event_list[n] = event;
- if (is_counter_pair(&event->hw))
+ if (is_counter_pair(&event->hw)) {
cpuc->n_pair++;
+ cpuc->n_txn_pair++;
+ }
return 0;
}
@@ -2062,6 +2065,8 @@ static void x86_pmu_start_txn(struct pmu *pmu, unsigned int txn_flags)
perf_pmu_disable(pmu);
__this_cpu_write(cpu_hw_events.n_txn, 0);
+ __this_cpu_write(cpu_hw_events.n_txn_metric, 0);
+ __this_cpu_write(cpu_hw_events.n_txn_pair, 0);
}
/*
@@ -2087,6 +2092,8 @@ static void x86_pmu_cancel_txn(struct pmu *pmu)
*/
__this_cpu_sub(cpu_hw_events.n_added, __this_cpu_read(cpu_hw_events.n_txn));
__this_cpu_sub(cpu_hw_events.n_events, __this_cpu_read(cpu_hw_events.n_txn));
+ __this_cpu_sub(cpu_hw_events.n_metric, __this_cpu_read(cpu_hw_events.n_txn_metric));
+ __this_cpu_sub(cpu_hw_events.n_pair, __this_cpu_read(cpu_hw_events.n_txn_pair));
perf_pmu_enable(pmu);
}
diff --git a/arch/x86/events/perf_event.h b/arch/x86/events/perf_event.h
index 345442410a4d..6348105b6d30 100644
--- a/arch/x86/events/perf_event.h
+++ b/arch/x86/events/perf_event.h
@@ -235,6 +235,8 @@ struct cpu_hw_events {
they've never been enabled yet */
int n_txn; /* the # last events in the below arrays;
added in the current transaction */
+ int n_txn_metric;
+ int n_txn_pair;
int assign[X86_PMC_IDX_MAX]; /* event to counter assignment */
u64 tags[X86_PMC_IDX_MAX];