> What I'd like to see however is optimising the return path from a syscall. Im
> convinced that we could have a "fast return path", and make need_resched, and
> signal delivery overwrite the return address on the stack to a slow return
> and that it would be a measurable win - Linus, we need to get syscalls down
> to 2uS not 3 ;)
current standings in the game of null syscalls contra Linus:
null syscall latency on a Pentium: 221 cycles
null syscall latency on an SMP Pentium: 570 cycles :((
(the last 3 cycles score was made by Linus in a pre2.0.4-ish patch to
entry.S)
This is the clean CPU latency, with no cache misses or cache
misalignments.
(can send code on request)
Ingo
ps. [ heh, i tried to optimize entry.S a bit, helped by a silly pentium simulator,
and the latency went promptly up to 270 cycles ... Linus, this is no
fun, you write too good code :) ]