Re: linuxthreads and signals

[Corey Minyard <minyard at acm dot org>]

george anzinger wrote:

> dank@kegel.com wrote:
>  
> > Now that the Linux kernel has added the "thread group" feature
> > in an effort to allow standard Posix thread signal semantics, is
> > anyone updating Linuxthreads to use it?
> >
> > The NGPT project is using that feature to good advantage, but
> > they're also adding user-level threading, so they're not what I'm after.
> > I'm asking about just plain old Linuxthreads (the one included in glibc)
> > fixed up to have Posix signal semantics.
> >
> > Thanks,
> > Dan
> >    
> A bit more info.  In order to use this, the
> CLONE_THREADGROUP flag has to be used.  This changes a few
> more things....
>  
This is somewhat complicated to implement due to the way the the 
CLONE_THREAD flag handles signals.  All the signal go to the first 
thread in the thread group, but to meet POSIX semantics the signal must 
go to a thread that doesn't have the signal blocked.  So you basically 
have to have one thread that receives all the signals and distributes 
them out to the other threads.  This means pretty big changes to 
linuxthread signal handling and sigaction() and buddies.  It's not a 
small job.

I know, because I have implemented a threads package that uses 
CLONE_THREAD (see http://ssthreads.sourceforge.net).  It would have been 
nice if the NGPT kernel patch had implemented the POSIX semantics for 
signal delivery in the kernel, it would make it a breeze to modify 
Linuxthreads to work properly and delivery of siginfo information would 
then work properly.

-Corey