cvs commit: src/sys/i386/isa npx.c

[David Xu]

On Sunday 28 May 2006 15:25, Bruce Evans wrote:
> On Sun, 28 May 2006, David Xu wrote:
> > davidxu     2006-05-28 04:40:45 UTC
> >
> >  FreeBSD src repository
> >
> >  Modified files:
> >    sys/i386/isa         npx.c
> >  Log:
> >  If parent thread never used FPU, the only work is to clear flag
>
> PCB_NPXINITDONE doesn't mean that a thread never used the FPU.  It
> means that the FPU state is not the default.
>
PCB_NPXINITDONE does mean thread used FPU, I don't know where you got
the information.

> >  PCB_NPXINITDONE for new thread and let trap code initialize it.
> >
> >  Revision  Changes    Path
> >  1.168     +6 -1      src/sys/i386/isa/npx.c
>
> Why do so much?  If PCB_NPXINITDONE is clear in the parent, then it
> is already clear in the child, since it has just been copied (the whole
> pcb has been copied).  It doesn't take a function call to check if
> PCB_INITDONE is clear in the parent.
>
I have already checked PCB_NPXINITDONE for parent thread, if it is not set,
npx_fork_thread will return from fast path.

> cpu_fork() calls npxsave() to force the FPU state to the pcb so that
> it is automatically copied at no extra cost when the whole pcb is
> copied.  This causes the entire FPU state to be inherited.  Why is
> cpu_set_upcall() different?  I the old difference was just an invalid
> "optimization" and the new difference is not good.  POSIX requires
> fork() to duplicate the process _exactly_ except for some things not
> including any FPU or even integer state, and FreeBSD implements this
> except for clobbering a few integer registers.  POSIX places fewer
> requirments on pthread_create() by listing things that are inherited
> and not requiring things not listed to be preserved; old versions of
> POSIX apparently didn't even require the FPU state to be preserved,
> but that was a bug and was fixed by aligning with SUS.  Since there
> is considerable experience that duplicating the whole FPU state in
> fork() doesn't cause problems, I think it wouldn't cause problems in
> pthread_create() either.  Anyway, POSIX must really mean that the
> whole FP environment must be inherited, so all that pthread_create()
> is permitted to do differently than fork() is scrubbing the data
> registers.
>

fork and pthread_create are different, cpu_set_upcall is for new thread when 
the process is threaded application, pthread_create executes code from
different point, while fork continues from the syscall, the new thread 
entry function DOES NOT understand pending exceptions, the new thread 
should be started with clean FPU state, it is not needed to mess new thread's
FPU state.

> Bruce