1690104 - New main thread hangs in RtlQueryPerformanceCounter / TimeStamp::Now()

Status: RESOLVED WORKSFORME

(Core :: Performance: General, defect)

Description

[Markus Stange [:mstange]]

The BHR dashboard at https://fqueze.github.io/hang-stats/ is showing a new hang on Windows that wasn't there before: A lot of time is spent getting timestamps using RtlQueryPerformanceCounter under JS::UnmarkGrayGCThingRecursively (called from EventListenerManager::MarkForCC).

Jon, is it possible that the patch in bug 1689140 caused us to query the current timestamp more frequently than before?

Comment 1

[Jon Coppeard (:jonco)]

JS::UnmarkGrayGCThingRecursively doesn't use SliceBudget so that's unlikely. I don't know why this would have changed.

Unfortunately the BHR dashboard isn't working at the moment. Do you have any idea when this started happening?

Comment 2

[BugBot [:suhaib / :marco/ :calixte]]

Set release status flags based on info from the regressing bug 1689140

Comment 3

[Markus Stange [:mstange]]

Florian, can you find out when this started?

Comment 4

[Jon Coppeard (:jonco)]

The full stack is:

    RtlQueryPerformanceCounter ntdll
    static mozilla::TimeStamp::NowUnfuzzed(bool) mozglue
    JS::UnmarkGrayGCThingRecursively(JS::GCCellPtr) xul
    static mozilla::dom::FragmentOrElement::CanSkip(nsINode*, bool) xul
    mozilla::CycleCollectedJSRuntime::UnmarkSkippableJSHolders() xul
    nsCCUncollectableMarker::Observe(nsISupports*, char const*, char16_t const*) xul
    nsObserverList::NotifyObservers(nsISupports*, char const*, char16_t const*) xul
    nsObserverService::NotifyObservers(nsISupports*, char const*, char16_t const*) xul
    nsObserverService::NotifyObservers cycle-collector-forget-skippable
    XPCJSRuntime::PrepareForForgetSkippable() xul
    nsCycleCollector_forgetSkippable(js::SliceBudget&, bool, bool) xul
    nsCycleCollector_forgetSkippable
    FireForgetSkippable(bool, mozilla::TimeStamp) xul
    CCRunnerFired(mozilla::TimeStamp) xul
    virtual bool std::_Func_impl_no_alloc<bool ( *)(class mozilla::TimeStamp),bool,class mozilla::TimeStamp>::_Do_call(class mozilla::TimeStamp &&) xul
    mozilla::IdleTaskRunner::Run() xul
    mozilla::TaskController::DoExecuteNextTaskOnlyMainThreadInternal(mozilla::detail::BaseAutoLock<mozilla::Mutex &> const&) xul

Comment 5

[Jon Coppeard (:jonco)]

JS::UnmarkGrayGCThingRecursively does result in a call to RtlQueryPerformanceCounter and always has. Maybe something has made the forgetSkippable happen more often?

Comment 6

[Andrew McCreight (out of office until 8/21) [:mccr8]]

Is it possible that the stack changed somehow? I'm not aware of anything that would have made us do forget skippable more often.

Olli, have their been any idle scheduling changes that might affect this?

Failing that, we might want to look at general cycle collector or ghost window metrics to see if anything regressed there. It could be a symptom of leaks which tend to tank CC performance.

Comment 7

[Olli Pettay [:smaug][[e-mail]]]

Hmm, do we have still the old bhr UI available? I think using that it was easy(er) to see regressions.

I'm not aware of changes to idle handling, unless pbone's patches have already landed.

Comment 8

[Florian Quèze [:florian]]

(In reply to Markus Stange [:mstange] from comment #3)

Florian, can you find out when this started?

The frequency is very low in current data: https://fqueze.github.io/hang-stats/#row=0&filter=NowUnfuzzed (20210213).

Do we still care about finding out when this started, or can we resolve as worksforme?

(In reply to Olli Pettay [:smaug] from comment #7)

Hmm, do we have still the old bhr UI available?

We still have the code at https://github.com/squarewave/bhr.html but the arewesmoothyet domain name has expired.

Comment 9

[Brendan Dahl [:bdahl]]

^ (trying to clear up some release engineering triage)

Comment 10

[Markus Stange [:mstange]]

(In reply to Florian Quèze [:florian] from comment #8)

The frequency is very low in current data: https://fqueze.github.io/hang-stats/#row=0&filter=NowUnfuzzed (20210213).

Ok, then let's close this bug.