791062 - GC: exact rooting for JSD

Status: RESOLVED FIXED

(Core :: JavaScript Engine, defect)

Comment 1

[David Mandelin [:dmandelin]]

We should check in with Honza and robcee before actually doing anything.

(Btw, description is missing. I realize the title is exactly what it says on the tin but a bug like this wants a rationale, and maybe something about the scope or approach if appropriate.)

Comment 2

[Terrence Cole [:terrence]]

Fair enough.

In order to turn on exact rooting in the browser without disabling JSD and thus breaking Firebug, we will need to port JSD to use the new exact rooting APIs.  

This will probably be as simple as replacing all of the JSObject*/Value/jsid/etc with Rooted versions and adding Handles until everything compiles again.  On the other hand, if JSD is doing something strange this could take longer potentially introduce bugs.

Comment 3

[Steve Fink [:sfink] [:s:]]

JSD maintains shadow copies in its own heap of nearly every type of gcthing there is, and connects them via hashtable lookups on pointer values. Up until now, it has relied on JS_AddNamedRoot and other variants to make this all work. It also keeps weak pointers to at least some of these objects, eg JSScripts, and relies on a debug hook callback to be invoked during finalization to clean them up.

Exact rooting JSD is feasible, but it is not a minor project. I'm not sure how it compares in magnitude to the project of porting Firebug to the new Debugger API, but that needs to be done regardless.

We could also consider having debug mode disable moving GC.

Comment 4

[Terrence Cole [:terrence]]

Hurm.  Well then.  Sadly, we will need this to even turn on exact rooting: it's not like we can just partially disable that.  The JSD situation is rather closer to being a schedule risk for generational GC than I thought.

:Honza, :robcee, what is the status of the Firebug port to JSD2?  Is there anything we can do to make this happen sooner, rather than later?

Comment 5

[Jan Honza Odvarko [:Honza] (always need-info? me)]

I started on JSD2 porting to Firebug several weeks ago and I think I am doing good progress. However there is still *a lot* what to do and Firebug needs months to be JSD2 ready (being ready at the end of this year would be rather huge success).

Honza

Comment 6

[Terrence Cole [:terrence]]

(In reply to Jan Honza Odvarko from comment #5)
> I started on JSD2 porting to Firebug several weeks ago and I think I am
> doing good progress.

It's awesome to hear that this work has started!

> However there is still *a lot* what to do and Firebug
> needs months to be JSD2 ready (being ready at the end of this year would be
> rather huge success).

We'd like to be fully transitioned over to exact rooting before then, but not by a huge amount.  It looks like we'll still need some way to deal with JSD in the meantime.  As long as work to get firebug off JSD is going to finish in about the right timeframe, we can use almost arbitrarily ugly hacks to bridge the gap.

Comment 7

[Terrence Cole [:terrence]]

Attachment: wip

While investigating, I put in some rooting for testing purposes. This might be useful as a base when we go forward with rooting this.

Comment 8

[(no longer active)]

I've done most of the rooting work necessary for the JS debugger and all of those patches should be on inbound now.  I did not know how to fix the remaining rooting hazards.

Comment 9

[Boris Zbarsky [:bzbarsky]]

So what remains is:

1)  An on-stack JSPropertyDesc that needs to be rooted.  I think this could most simply be
    done via a RootedJSPropertyDesc that inherits from both JSPropertyDesc and
    CustomAutoRooter.
2)  The glob_proper thing in jsds_SyncFilter.  The biggest problem there is lack of a
    JSContext to root it with....

Comment 10

[Terrence Cole [:terrence]]

We're all done here.