Re: Question and patch about spinlocks (x86)

[Davide Libenzi]

On Thu, 11 Oct 2001, Mark Zealey wrote:

> Just looking through at the spinlock assembly I noticed a few things which I
> think are bugs:
>
>  	"js 2f\n" \
>  	".section .text.lock,\"ax\"\n" \
>  	"2:\t" \
> 	"cmpb $0,%0\n\t" \
>  	"rep;nop\n\t" \
> 	"jle 2b\n\t" \
>  	"jmp 1b\n" \
>  	".previous"
>
> We do the cmp loop as a 'soft' check, as the lock operand locks the whole system
> bus, stopping the system for a while (as much as 70 cycles, I believe). However,
> I don't understand why it was put before the 'rep; nop' which just sets the
> processor to wait for a bit. Surely it would be better to test *after* we have
> waited, as then we have a better chance of it being correct.

The effect of the rep-nop is not to wait but to slow down the cpu to the
speed of the memory bus.
This to not overload ( due pipeline prefetch ) the memory controller with
requests that 1) will be useless coz the watched memory location can
change only at the membus speed 2) will have a big cost on loop exit due
the invalidation of a number>1 requests issued on the memory controller.
Beside this i kindly agree to move the pause before the cmp.
There should be a valid reason to not have followed the intel scheme but i
don't know why.

- Davide