Messages in this thread | | | Subject | Re: [PATCH] perf/x86: Consider pinned events for group validation | From | "Liang, Kan" <> | Date | Tue, 20 Aug 2019 10:52:57 -0400 |
| |
On 8/20/2019 10:10 AM, Peter Zijlstra wrote: > On Fri, Aug 16, 2019 at 10:49:10AM -0700, kan.liang@linux.intel.com wrote: >> From: Kan Liang <kan.liang@linux.intel.com> >> >> perf stat -M metrics relies on weak groups to reject unschedulable >> groups and run them as non-groups. >> This uses the group validation code in the kernel. Unfortunately >> that code doesn't take pinned events, such as the NMI watchdog, into >> account. So some groups can pass validation, but then later still >> never schedule. > > But if you first create the group and then a pinned event it 'works', > which is inconsistent and makes all this timing dependent.
I don't think so. The pinned event will be validated by validate_event(), which doesn't simulate the schedule. So the validation still pass, but the group still never schedule.
> >> @@ -2011,9 +2011,11 @@ static int validate_event(struct perf_event *event) >> */ >> static int validate_group(struct perf_event *event) >> { >> + struct cpu_hw_events *cpuc = this_cpu_ptr(&cpu_hw_events); >> struct perf_event *leader = event->group_leader; >> struct cpu_hw_events *fake_cpuc; >> - int ret = -EINVAL, n; >> + struct perf_event *pinned_event; >> + int ret = -EINVAL, n, i; >> >> fake_cpuc = allocate_fake_cpuc(); >> if (IS_ERR(fake_cpuc)) >> @@ -2033,6 +2035,24 @@ static int validate_group(struct perf_event *event) >> if (n < 0) >> goto out; >> >> + /* >> + * The new group must can be scheduled >> + * together with current pinned events. >> + * Otherwise, it will never get a chance >> + * to be scheduled later. > > That's wrapped short; also I don't think it is sufficient; what if you > happen to have a pinned event on CPU1 (and not others) and happen to run > validation for a new CPU1 event on CPUn ? >
The patch doesn't support this case. It is mentioned in the description. The patch doesn't intend to catch all possible cases that cannot be scheduled. I think it's impossible to catch all cases. We only want to improve the validate_group() a little bit to catch some common cases, e.g. NMI watchdog interacting with group.
> Also; per that same; it is broken, you're accessing the cpu-local cpuc > without serialization.
Do you mean accessing all cpuc serially? We only check the cpuc on current CPU here. It doesn't intend to access other cpuc.
Thanks, Kan
> >> + */ >> + for (i = 0; i < cpuc->n_events; i++) { >> + pinned_event = cpuc->event_list[i]; >> + if (WARN_ON_ONCE(!pinned_event)) >> + continue; >> + if (!pinned_event->attr.pinned) >> + continue; >> + fake_cpuc->n_events = n; >> + n = collect_events(fake_cpuc, pinned_event, false); >> + if (n < 0) >> + goto out; >> + } >> + >> fake_cpuc->n_events = 0; >> ret = x86_pmu.schedule_events(fake_cpuc, n, NULL); >> >> -- >> 2.7.4 >>
| |