Oh - this is on dev channel btw. Attempting to attach my trace, but I've encountered yet another bug wherein traces are not saving properly. :/

There we go!

Deleted: trace.json.gz 2.3 MB

trace.json.gz 2.3 MB Download

Adding a speed person too, as I think this could use more eyes.

This isn't a v8 GC, it's an oilpan GC (I think). CCing some oilpan folks.

Can you take another trace with the blink_gc category enabled?

Yeah, though it doesn't really add any more data other than probably confirming blink GC within invokePreFinalizers.

Deleted: trace-blink-gc.json.gz 2.8 MB

trace-blink-gc.json.gz 2.8 MB Download

I can also at least confirm that it appears to be limited to M65+. No success attempting to repro in beta or stable.

Why is the CPU time so low? Looks like it's getting consistently descheduled somehow.

I am confused by the fact that performIdleGC CPU is consistent with wall time, yet invokePreFinalizers CPU time is not - despite essentially spanning the full extent of performIdleGC. Is that even possible?

I'm not seeing what you describe. performIdleGC also appears to have tiny CPU time.

Title			ThreadState::performIdleGC
Category		blink_gc
User Friendly Category	other
Start			5,432.827 ms
Wall Duration		8,888.681 ms
CPU Duration		7.961 ms
Self Time		0.268 ms
CPU Self Time		0.267 ms

You are correct. I'm just bad at reading.

So if it's getting descheduled a bunch... IO or lock contention?

There's something media related happening:
src_file	"../../media/base/serial_runner.cc"
src_func	"RunOnTaskRunner"

It's also descheduled though.

The CPU usage graph at the top makes it look like there isn't much going on.

FWIW invokePreFinalizers should be very lightweight.

Who is the right owner for this?

Presumably someone with context on Oilpan GC?

FWIW it seems like this has disappeared since my last dev channel update.

Plase reopen if this is a still a problem.

Component: Blink>MemoryAllocator>GarbageCollection is the right one and should be triaged (we are improving the process right now).