RE: Thread stack pointers

["Boehm, Hans" <hans dot boehm at hp dot com>]

Yes.  You can also look at
http://bdwgc.cvs.sourceforge.net/bdwgc/bdwgc/pthread_stop_world.c?revisi
on=1.4&view=markup

You probably also want to make sure that you pause the thread in the
signal handler until the collector has either finished tracing
completely, or has at least finished tracing the stack, depending on the
details of your algorithm.

Hans 

> -----Original Message-----
> From: java-owner@gcc.gnu.org [mailto:java-owner@gcc.gnu.org] 
> On Behalf Of David Daney
> Sent: Wednesday, March 21, 2007 3:15 PM
> To: Delvin Defoe
> Cc: java@gcc.gnu.org
> Subject: Re: Thread stack pointers
> 
> Delvin Defoe wrote:
> > Does this mean if I define sigaction act, set the handler 
> as follows, 
> > // for stack processing act.sa_handler = _otf_process_stack; 
> > sigemptyset(&act.sa_mask); act.sa_flags = 0;
> > 
> > signal the thread in the following manner  if 
> (tkill(m->_tid,i+1) == 
> > -1)
> >      fprintf (stderr, "***** tikll failed for thread  %d \n. *****",
> > m->_tid);
> > 
> > define the sa_handler function accordingly
> > 
> > _otf_process_stack (int signo)
> > 
> > the stack top for the thread that receives the signal can 
> be give by 
> > something like stack_top = &signo + sizeOf(pointer)?
> > 
> 
> That is essentially correct.  You don't have to add 
> sizeof(pointer) as 
> the signal handler has no references to any java objects.  All of the 
> valid references on the stack will be quite a bit above the 
> stack frame 
> of the signal handler.  The signal handler is just a convenient lower 
> bound.  If you want the real value you could extract it from 
> the saved 
> machine state in the signal handler's siginfo structure, but 
> that makes 
> the code very machine dependent.
> 
> > 
> > 
> > Thanks,
> > 
> > Sorry about my previous imcomplete post.
> > 
> > Delvin
> > On 3/21/07, David Daney <ddaney@avtrex.com> wrote:
> >> Delvin Defoe wrote:
> >> > Hi:
> >> >
> >> > I am implementing a multi-thread, multiprocessor garbage 
> collection
> >> > system for Java in gcc.  In the runtime system I need to 
> determine the
> >> > stack base and stack top pointers for each mutator in 
> order to scan
> >> > the stack for each mutator for pointers to live heap 
> objects.   Any
> >> > idea how I can determine or compute these pointers?
> >> >
> >> Generic answer:
> >> How about doing the same thing boehm-gc does?
> >>
> >> For linux, you would have each thread is register with the 
> GC when it is
> >> started.  In this registration function you find the address of
> >> something on the stack of that thread, then look in 
> /proc/self/maps to
> >> see what memory range contains that address.  That range 
> will be your
> >> stack bounds.
> >>
> >> To find the stack top, send a signal to a thread, and then 
> the stack top
> >> will be slightly above the address of an automatic variable of the
> >> signal handler function.
> >>
> >> David Daney
> >>
> > 
> > 
> 
>