661927 - API to force a precise GC (return to event loop, GC with no JS stack, then call a callback)

Status: RESOLVED FIXED

(Core :: XPConnect, enhancement)

Description

[Jesse Ruderman]

For quitWithLeakCheck (and perhaps other automated tests), I'd like to have a way to cause a GC to happen with no JS on the stack, so I don't have to worry about conservative GC false positives.

Comment 1

[Kyle Huey (Exited; not receiving bugmail, old account, do not use)]

Do we want to run other things on the event loop here, or do we want to just remove what's entrained on the stack from consideration?

Comment 2

[Josh Matthews [:jdm]]

Hmm, is the idea that you would just dispatch a runnable that would check if there's JS on the stack and redispatch itself if it couldn't gc yet?

Comment 3

[Andrew McCreight [:mccr8]]

I'd think that the function would take a JS function as an argument.  When run, it would blow away the JS stack, and invoke the function argument, which is being used as a continuation.  A JS-exec, if you prefer.

Comment 4

[Jesse Ruderman]

Comment 2 sounds right, given the possibility of nested event loops.

No synchronously "blowing away the stack", please :)

Comment 5

[Josh Matthews [:jdm]]

Sounds like fun.

Comment 6

[Kyle Huey (Exited; not receiving bugmail, old account, do not use)]

I would think for fuzzing that you'd want to be able to make GC happen "right now", or as close to right now as is possible.  That would mean that you wouldn't want to run extra things off the event loop, which comment 2 would do ...

Comment 7

[Andrew McCreight [:mccr8]]

Clearly Javascript needs a general call-cc mechanism.

Comment 8

[Jesse Ruderman]

I can already make GC happen "right now", and I use that for fuzzing. What I'm asking for here would be used for precise leak checks at the end of fuzzing, so it's ok if it's stuck on the end of the event queue.

Comment 9

[Josh Matthews [:jdm]]

Attachment: Add Cu.schedulePreciseGC to allow for a GC to run with no JS code running.

Comment 10

[Josh Matthews [:jdm]]

Attachment: Add Cu.schedulePreciseGC to allow for a GC to run with no JS code running.

Comment 11

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

How does the patch guarantee that there is no JS running?

Comment 12

[Josh Matthews [:jdm]]

I assumed that simply checking JS_IsRunning on the desired context would be enough. Should I instead acquire the xpconnect JS runtime, then iterate over all its contexts and do the same check?

Comment 13

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

My mistake. I missed the JS_IsRunning in the ::Run().
(I don't know what JS_IsRunning does)

Another question, couldn't you have some idl interface to implement the callback?

Comment 14

[Josh Matthews [:jdm]]

I could. I don't really see why that's desirable, however.

Comment 15

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

That way the callback would automatically support both function() {} and
{ foo: function() {} }. And using JS API is quite error prone.

Comment 16

[Blake Kaplan (:mrbkap) (inactive)]

Comment on attachment 541856 [details] [diff] [review]
Add Cu.schedulePreciseGC to allow for a GC to run with no JS code running.

I think I agree with smaug here. If you take an XPCOM interface, you won't have to deal with rooting or context pushing or anything else like that. I'm not sure if the JS_IsRunning check is sufficient. It's possible that other JS code could be running on a different context. On the other hand, I'm not sure if we have an easy way of finding that out...

Comment 17

[Josh Matthews [:jdm]]

> I'm not sure if the JS_IsRunning check is sufficient. It's possible that other
> JS code could be running on a different context. On the other hand, I'm not
> sure if we have an easy way of finding that out...

It's possible to grab the runtime service, get the JSRuntime and then use a context iterator to check every context therein. Would that be more suitable?

Comment 18

[Josh Matthews [:jdm]]

Attachment: Add Cu.schedulePreciseGC to allow for a GC to run with no JS code running.

Comment 19

[Blake Kaplan (:mrbkap) (inactive)]

Comment on attachment 542941 [details] [diff] [review]
Add Cu.schedulePreciseGC to allow for a GC to run with no JS code running.

Yeah, this is much cleaner.

One nit: your test addition to the makefile should use tabs instead of spaces for spacing. r=mrbkap with that.

Comment 20

[Josh Matthews [:jdm]]

http://hg.mozilla.org/integration/mozilla-inbound/rev/8dccc136f585

Comment 21

[Josh Matthews [:jdm]]

Backed out in http://hg.mozilla.org/integration/mozilla-inbound/rev/5773d613df94 due to persistent windows build failures.

xpccomponents.obj : error LNK2019: unresolved external symbol "struct JSContext * __cdecl js_ContextIterator(struct JSRuntime *,int,struct JSContext * *)" (?js_ContextIterator@@YAPAUJSContext@@PAUJSRuntime@@HPAPAU1@@Z) referenced in function "public: virtual unsigned int __stdcall PreciseGCRunnable::Run(void)" (?Run@PreciseGCRunnable@@UAGIXZ)
NEXT ERROR xul.dll : fatal error LNK1120: 1 unresolved externals

Comment 22

[Josh Matthews [:jdm]]

Attachment: Add Cu.schedulePreciseGC to allow for a GC to run with no JS code running.

Third time's the charm.

Comment 23

[Blake Kaplan (:mrbkap) (inactive)]

Comment on attachment 543136 [details] [diff] [review]
Add Cu.schedulePreciseGC to allow for a GC to run with no JS code running.

You don't want the request around the JS_GC call. Otherwise, we'll still scan the stack. Other than that, this looks good.

Comment 24

[Josh Matthews [:jdm]]

Attachment: Add Cu.schedulePreciseGC to allow for a GC to run with no JS code running.

Patch for landing.

Comment 25

[Dão Gottwald [:dao]]

http://hg.mozilla.org/mozilla-central/rev/7898841a922a

Comment 26

[Eric Shepherd [:sheppy]]

Documentation added:

https://developer.mozilla.org/en/Components.utils.schedulePreciseGC

Documentation updated:

https://developer.mozilla.org/en/Components.utils.forceGC
https://developer.mozilla.org/en/Components.utils
https://developer.mozilla.org/en/Components_object

Firefox 7 for developers has also been updated.