[fedora-java] RE: rssowl: libgcj bugs that need fixing for it torun

[Robin Green]

On Thu, 28 Jul 2005 17:02:35 -0700, Boehm, Hans wrote:
> There is some code in earlier versions of the
> GC to do thread registration, but it's very platform specific
> and thus ugly.  I think there wasn't a real facility for Linux.
> 
> The tricky part of doing this in general is that the GC needs
> to know the stack bounds for the newly registered thread.  It
> can find the hot end, but the cold end is often hard.  GC7
> addresses this by providing two ways to get the cold stack
> end:
> 
> 1) A generic mechanism that just takes the address of a local.
> The collector knows how to implement that everywhere.  We just
> provide a function that calls back one of your functions f with
> a stack address that's guaranteed to be "below" f.  Since this
> is not the actual base of the stack, the GC ends up tracing
> pointers only in "new" frames.
> 
> 2) A separate routine that tries to discover the stack base
> in a platform dependent way.  It may fail.  (And currently
> usually does.)  I think that for Linux, pthread_getattr_np
> works for most threads, though perhaps not the main one.
> (The thread pointer also probably works in many cases.)
> 
> I'm not sure the JNI primitives can be implemented in terms
> of (1).  Certainly if you use CNI that has different semantics,
> in that the GC doesn't see pointers "below" you on the stack.

On the contrary, I believe that JNI's AttachCurrentThread can
and should be implemented in terms of something like (1).
AttachCurrentThread (which allows you to set up a thread to call
Java from C) is part of the Invocation API, which does not appear
to make any special garbage collection guarantees about references
held by the native caller. It is my understanding that any
references to Java objects held by the native caller must be
explicitly registered as JNI global references, and therefore the
stack below the Java part does not have to be - and indeed should
*not* be - considered. It should not be considered because this
could cause spurious leaks or hide wrong code. (However, it's probably
OK to include it if doing a temporary workaround.)

However, I think some setup would need to be done *every* time a call
was made from native->Java code at "top level" (i.e. where the native
caller was not called by Java code on this thread), because
the cold end of the Java part of the stack can be in a different
place each time.

I am going to take a different tack and try to implement something like my
earlier proposal:

>> If so, the only other thing I can think of is to spawn a new 
>> registered thread instead of calling AttachCurrentThread, and 
>> somehow translate all C->Java invocations on the unregistered 
>> thread into inter-thread calls onto the new registered 
>> thread. In other words, keep all Java code on a separate thread. Yuck.

-- 
Robin

P.S. Cacao and Kaffe "avoid" this problem because they simply refuse
to provide a working AttachCurrentThread, so callbacks from native threads
can't work at all - which is, obviously, not a workaround.