Subj : Re: JS_THREADSAFE /  assertion failed on rt->gcRunning
To   : netscape.public.mozilla.jseng
From : smaraux
Date : Fri Mar 04 2005 01:58 am

Brendan Eich <brendan@meer.net> wrote in message news:<4227931D.8050804@meer.net>...
> sebastien maraux wrote:
> 
> > When I run my program, it crashes after a few seconds, on an assertion
> > failed to know if gc is running:
> > JS_ASSERT(!rt->gcRunning);
> 
> 
> You have written a finalizer that attempts to allocate a new GC-thing. 
> Don't do that.
> 

I don't think so, I have 2 custom finalizers (one for arrays and one
for single elements) and none call any JS_New... or
JS_ConstructObject. It basically only call JSGetPrivate and delete its
private.

That's why I was thinking of a JS_THREADSAFE issue, or something
related to code build.

> 
> > I have read following post, where you can read that building js with
> > JS_THREADSAFE and NSPR or a small replacement for it could fix my
> > problem :
> > 
> > http://groups.google.ca/groups?hl=en&lr=&ie=UTF-8&selm=7be32359.0403171719.2ef8666%40posting.google.com
> > 
> > However, I don't understand why js is built as multithreaded but
> > without this tag in its default configuration, if it's unsafe.
> 
> 
> I don't know what you mean.  SpiderMonkey can be built without 
> JS_THREADSAFE for use in single-threaded embeddings, and with 
> JS_THREADSAFE for multi-threaded embeddings -- in the latter case you 
> want NSPR or a workalike (I am not sure that the post you link to above 
> supplies a complete workalike).

I am under VC6, which as a default project with build setting to
"multithreaded dll", but without JS_THREADSAFE definition. I may try
to build a "single threaded DLL" project to see if it makes things
better.

> 
> > Moreover latest NSPR static binaries are not valid on the ftp (I can
> > link js to dll ones, but not static ones), and I don't want additional
> > dlls in my project.
> 
> 
> Pull the source from cvs.mozilla.org and build it yourself.

I am behind a very strict firewall which does not authorized cvs port
and I don't have any "socks" utilities (I know, that s*cks, it is a
common issue here : we have to stick to tarballs in most opensource
projects)

> 
> 
> > I also mention that however my environement is multithreaded, I am
> > using js only in a single thread of my program but it still crashes.
> 
> 
> Then why are you defining JS_THREADSAFE?  Oh, you're saying that the 
> builds we put up define it?  Those must be the library builds used with 
> the Mozilla suite, e.g. libmozjs.so on Linux.  Again you should built it 
> yourself to suit your own requirements.  We provide builds as a 
> convenience for developers, but we can't manage a bunch of different 
> configurations that meet all needs.
> 

I am using independent js-rc1.5a build (VC6 project comes without this
option, but with multithreaded settings)

> > Also, is there some detailed doc on how to use JS_BeginRequest /
> > JS_EndRequest other than js-c api reference ? I understand that it
> > should be placed to bracket maximum non blocking chunk of code, but
> > for example if I set members of an object from another , without
> > knowing their number in advance, can it be considered as non blocking
> > (or short enough) calls ? Or should I test for array size before
> > calling JS_BeginRequest. If their is too many elements, must I break
> > it in several smaller calls (this could be an issue as I have to
> > handle a lot of different arrays types).
> 
> 
> It depends on how long the loop you've written could run.  If you're 
> worried, maybe using an unsigned index and testing (i % 1024) == 0 or 
> similar to call JS_YieldRequest(cx) would be good (I made up 1024, you 
> should pick a better number).

Thank you, it makes it clear for me.

> 
> 
> > I'll now try older versions of NSPR, but I am afraid that it will
> > greatly increase my code footprint to link with NSP lib.
> 
> 
> Use a static libnspr4.a library you build yourself, and ld shouldn't 
> link more than you need.

OK, I don't know if vc6 knows how to link only used and not unused
functions in a single c file. But in a first time, if I could just get
rid of the assertion, this would be a good first step.

> 
> /be

Here is my call stack when the assertion fails :
I call a function which should update data of my script :

JS_CallFunctionValue(m_JSContext, m_GlobalObject, fval, 2, argv,
&rval);
-> js_InternalInvoke(JSContext *cx, JSObject *obj, jsval fval, uintN
flags,
                  uintN argc, jsval *argv, jsval *rval)
-> js_Invoke(JSContext *cx, uintN argc, uintN flags) //call a function
to move an object left
-> js_Interpret(cx, &v);
-> JSBool js_DefaultValue(JSContext *cx, JSObject *obj, JSType hint,
jsval *vp)
-> js_TryMethod(cx, obj, cx->runtime->atomState.toStringAtom, 0, NULL,
&v))
-> js_InternalInvoke(JSContext *cx, JSObject *obj, jsval fval, uintN
flags,
                  uintN argc, jsval *argv, jsval *rval)
-> js_Invoke() // call a function to draw moved object
-> (my function) fieldToString() // try to print its new position :
build a string from its vector position
-> JS_NewUCStringCopyZ
-> JS_NewString
-> js_AllocGCThing
-> assertion failed on JS_ASSERT(!rt->gcRunning);

If this can help verify that I am not in a finalize function.

.