822031 - OOM caused by browser process can result in only the main process getting the low-memory event

Status: RESOLVED INVALID

(Firefox OS Graveyard :: General, defect)

Description

[Benoit Jacob [:bjacob] (mostly away)]

In bug 820217 we have a game that tries to load 280 M of WebGL textures, resulting in an OOM crash.

The problem is that while this OOM is caused by work done by the browser process, only the _main_ process receives a low-memory event.

How about saying that as soon at the main process receives a low-memory event, it should propagate it to all other processes immediately?

Should that be generalized to saying that as soon as _any_ process receives a low-memory event, it should propagate it to all other processes?

How should the propagation be achieved? A key requirement is it happening as fast as possible. Maybe a signal?

Comment 1

[Justin Lebar (not reading bugmail)]

> How about saying that as soon at the main process receives a low-memory event, it should propagate 
> it to all other processes immediately?

Each process listens for low memory events itself.  That's why the MemoryPressureRunnable at the top of GonkMemoryPressureRunnable fires a "low-memory-no-forward" event.  The "no-forward" part indicates that ContentParent shouldn't forward the event to the children.

We have each process listen for low-memory instead of having the parent forward the memory-pressure events down to its children specifically because we want to get these events out to the relevant processes as quickly as possible.

What I suspect is happening is that the browser is OOMing too quickly for it to notice that it's run out of memory before it crashes.  As such, I suspect this bug is invalid.

Comment 2

[Justin Lebar (not reading bugmail)]

> That's why the MemoryPressureRunnable at the top of GonkMemoryPressureRunnable

Sorry, that should be widget/gonk/GonkMemoryPressureMonitoring.cpp.

Comment 3

[Benoit Jacob [:bjacob] (mostly away)]

OK, that makes sense.

It would be nice if we could consistently ensure that processes get a low-memory notification early enough before getting killed, but that's a hard problem to solve.