Re: [patch] tls-2.5.31-D5

Jakub Jelinek (jakub@redhat.com)
Mon, 12 Aug 2002 11:21:55 -0400


On Mon, Aug 12, 2002 at 07:06:50PM +0200, Ingo Molnar wrote:
>
> okay, the attached patch does some more things:
>
> - moves the first two TLS entries and the user CS/DS entries on the same
> cacheline.
>
> - excludes CS/DS from the TLS space - Luca is right in that it only slows
> things down unnecesserily, and there is nothing that cannot be done by
> changing the %ds %cs selectors - and every cycle counts in the
> context-switch path.
>
> the only open issues are the number of TLSs supported. I'd vote for making
> them 4 and then we can inline the copy and make it unconditional, it will
> be 12 cycles to copy them all which alone is better than a branch miss. In
> this patch it's 2, thus the copying cost is 6 cycles.
>
> with 4 entries the 0x40 entry would be taken and APM has to move further
> up, and has to save/restore the 0x40 entry across BIOS calls.

As each supported TLS entry has its context-switch time cost, I think we
should stay at 2 supported TLS entries.
My understanding was that the GDT patches were written to optimize the
common case (all threaded apps using LDT and with the advent of __thread
support causing every single application to use LDT), with 2 TLS entries
where one is for libc/libpthread and the other one is for application
usage I think it is enough for 99.9% of apps. In the rare
case someone needs more, there is still LDT which offers 8192 entries.

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