On Fri, 2006-09-22 at 15:24 -0700, Jeremy Fitzhardinge wrote:

Rusty Russell wrote:

quoted

This patch implements save/restore of %gs in the kernel, so it can be
used for per-cpu data.  This is not cheap, and we do it for UP as well
as SMP, which is stupid.  Benchmarks, anyone?
  

I measured the cost as adding 9 cycles to a null syscall on my Core Duo 
machine.  I have not explicitly measured it on other machines, but I run 
a number of other segment save/load tests on a wide range of machines, 
and didn't find much variability.

Oh, OK!  I had a belief that segment loading was expensive, perhaps I'm
off-base here.

I think saving/restoring %gs will still be necessary. There are a number 
of places in the kernel which expect to find the usermode %gs on the 
kernel stack frame, including context switch, ptrace, vm86, signal 
context, and maybe something else.  If you don't save it on the stack, 
then you need to have UP variations of %gs handling in all those other 
places, which is pretty messy.  Also, unless you want to have two 
definitions of struct_pt regs (which would add even more mess into 
ptrace), you'd still need to sub/add %esp in entry.S to skip over the 
%gs hole.  I don't think this UP microoptimisation would be worth enough 
to justify the mess it would cause elsewhere.

How does this version of the patch differ from mine?  Is it just my 
patch+Ingo's fix, or are there other changes?  I couldn't see anything 
from a quick read-over.

Yep, no substative changes.  s/__KERNEL_PDA/__KERNEL_PERCPU/, plus your
version had a "write_pda(pcurrent, next_p)" inserted in process.c's
__switch_to which belonged in a successor patch...

Thanks!
Rusty.
-- 
Help! Save Australia from the worst of the DMCA: http://linux.org.au/law