[M3devel] chosing SIG_SUSPEND?

[Tony Hosking]

On 23 Jan 2009, at 02:09, Jay wrote:

> I noticed in some of the code..I think for SegV, Quit,
> etc., not for thread suspension, that if the signal
> is already set to other than SIG_DFL, Modula-3 refrains
> from changing it.
>
> So, I was thinking, would it be reasonable to iterate through, say
> SIGUSR1, SIGUSR2, and SIGRTMIN through SIGRTMAX, and
> use the "first" one that isn't set to SIG_DFL?
> By "first", I don't mean to imply what order to check in,
> probably my statement has the order backwards.
> The point is to check them and not just hijack them.

Maybe.  I would argue the opposite: that M3 apps that use signals  
should avoid using the one dedicated to thread suspension.

> OR, is there another way to do this?
>
> Doesn't or can't the Modula-3 code "check a flag"
> "every so often" and then voluntarily yield?
> Of course, the allocator would check the flag.

This is a reasonable approach for Modula-3-compiled code: the polling  
points are called GC-safe-points.  You just need to insert poll-points  
on backward branches and calls.  The real problem is stopping threads  
that are off in native code and system calls.   We could wrap every  
native call with a code to save GC-relevant state before going native,  
and then check the flag on return from native in case a GC is in  
progress.  Many Java implementations use that approach.

> Perhaps I just really need to get "that paper" into my head?
> Understand the issue around "other threads running in
> tight computation loops for a long time", and therefore not checking
> on yielding?
>
> Is the signal mechanism actually preemptive anyway?

Preemptive in what sense?  Depending on the OS, signals tend to get  
delivered at system calls or thread context-switches.

>
>
>  - Jay
>
>
>
>
>
>
> > From: hosking at cs.purdue.edu
> > To: jay.krell at cornell.edu
> > Date: Wed, 21 Jan 2009 09:42:38 -0500
> > CC: m3devel at elegosoft.com
> > Subject: Re: [M3devel] chosing SIG_SUSPEND?
> >
> > Yes, we want a signal no-one else is using to stop the threads.
> >
> > On 21 Jan 2009, at 08:37, Jay wrote:
> >
> > >
> > > Solaris at least currently has SIGRTMAX.
> > > I want to switch RTSignalC.c to "something" here.
> > > I can preserve compat, like
> > >
> > > #ifdef __sun
> > > SIGUSR2
> > > #else..
> > >
> > > or..?
> > >
> > > I could do
> > > #ifdef SIGUSR2
> > > #define SIG_SUSPEND SIGUSR2
> > > #elif defined(SIGRTMAX)
> > > #define SIG_SUSPEND SIGRTMAX
> > > #else
> > > #error
> > > #endif
> > >
> > > I'll go with a compatible version for now and you can change it if
> > > you want.
> > > I'll test on Linux/x86, FreeBSD (AMD64, but hopefully there is  
> gcc -
> > > m32),
> > > Linux/PPC, Solaris, Cygwin, Darwin/PPC (uh, actually, no, Darwin  
> isn't
> > > really relevant, but I'll make sure the result does what is  
> intended.)
> > >
> > > It looks like SIGRTMAX is a function call on Solaris.
> > >
> > > This stuff is like broken, right?
> > > I mean, there's a small number of signals and there's no  
> arbitration.
> > > People just take them over and hope nobody else cares.
> > >
> > >
> > > There's no way to just queue a function call to a thread, in
> > > portable/general?
> > > Windows has QueueUserAPC or SuspendThread (SuspendThread is  
> dangerous,
> > > thread could be doing anything), QueueUserAPC only interrupts at
> > > certain times.
> > >
> > >
> > > - Jay
> > >
> > >
> > > ----------------------------------------
> > >> From: hosking at cs.purdue.edu
> > >> To: jay.krell at cornell.edu
> > >> Date: Wed, 21 Jan 2009 03:42:01 +1100
> > >> CC: m3devel at elegosoft.com
> > >> Subject: Re: [M3devel] chosing SIG_SUSPEND?
> > >>
> > >> Choose SIGRTMAX first if defined, then SIGUSR2. I don't thing  
> NSIG-1
> > >> is reliable, but just works on some systems.
> > >>
> > >> On 21 Jan 2009, at 01:12, Jay wrote:
> > >>
> > >>>
> > >>> What is the algorithm for chosing SIG_SUSPEND?
> > >>>
> > >>> Something like:
> > >>>
> > >>> #include
> > >>>
> > >>> #ifdef __APPLE__
> > >>> /* nothing -- SIG_SUSPEND not used */
> > >>> #elif defined(NSIG)
> > >>> #define SIG_SUSPEND (NSIG - 1)
> > >>> #elif defined(_NSIG)
> > >>> #define SIG_SUSPEND (_NSIG - 1)
> > >>> #elif defined(SIGRTMAX)
> > >>> #define SIG_SUSPEND SIGRTMAX
> > >>> #else
> > >>> #define SIG_SUSPEND SIGUSR2
> > >>> #endif
> > >>>
> > >>> ?
> > >>>
> > >>> Whatever it is, I think it should be in RTSignalC.c.
> > >>>
> > >>> - Jay
> > >>
> >
>

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://m3lists.elegosoft.com/pipermail/m3devel/attachments/20090123/99ab75e6/attachment-0002.html>