Re: Interesting scheduling times

Andi Kleen (ak@muc.de)
Sat, 19 Sep 1998 06:52:34 +0200


On Sat, Sep 19, 1998 at 06:21:45AM +0200, Linus Torvalds wrote:
>
>
> On 19 Sep 1998, Andi Kleen wrote:
> >
> > Lazy FPU saving is pretty useless as long as glibc initialises the FPU
> > in the startup code in every program. I do not know if you have seen Jakub
> > Jelinek's AT_FPUCW patch - which passes the initial FPU state in the ELF
> > auxilliary vector so that the startup code could decide if it needs to
> > initialise the state or not. Without that the FPU initialisation makes
> > lazy FPU saving useless.
>
> No, the lazy FPU save works perfectly fine even with FPU initialization.
> It does it lazily on a per-scheduling-quantum issue: if the process
> doesn't use the FPU within some quantum, it will not save the FPU state.

Thank you for the clarification. Seems I was misunderstanding the used
algorithm.

>
> As to the AT_FPUCW patch, I don't see the point in saving a constant value
> in the ELF headers. Linux initializes the FPU for each process at
> "exevce()" time, and glibc might as well depend on that instead of trying
> to do so itself.

The problem is only that Ulrich Drepper does not trust the kernel to always
keep the same value (this was the outcome of a discussion on l-k a few months
ago). If the constant value was passed he could verify it and satisfy his
paranoia @)

-Andi

-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.rutgers.edu
Please read the FAQ at http://www.tux.org/lkml/