Question and patch about spinlocks (x86)
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 $0,%0\n\t" \ "rep;nop\n\t" \ "jle 2b\n\t" \
"jmp 1b\n" \
".previous"
We do the cmp loop as a 'soft' check, as the lock operand locks the whole system bus, stopping the system for a while (as much as 70 cycles, I believe). However, I don't understand why it was put before the 'rep; nop' which just sets the processor to wait for a bit. Surely it would be better to test *after* we have waited, as then we have a better chance of it being correct.
Any comments? Attached is a patch to fix it.
Signature · 9 lines
Mark Zealey (aka JALH on irc.openprojects.net: #zealos and many more) [email protected] [email protected] UL++++>$ G!>(GCM/GCS/GS/GM) dpu? s:-@ a16! C++++>$ P++++>+++++$ L+++>+++++$ !E---? W+++>$ N- !o? !w--- O? !M? !V? !PS !PE--@ PGP+? r++ !t---?@ !X---? !R- b+ !tv b+ DI+ D+? G+++ e>+++++ !h++* r!-- y-- (www.geekcode.com)
asm-i386/spinlock.h renamed from asm-i386/spinlock.h.old +1 -1
@@ -58,8 +58,8 @@"js 2f\n" \".section .text.lock,\"ax\"\n" \"2:\t" \ -"cmpb $0,%0\n\t" \"rep;nop\n\t" \ +"cmpb $0,%0\n\t" \"jle 2b\n\t" \"jmp 1b\n" \".previous"
Quoted 12 lines · Mark Zealey
Just looking through at the spinlock assembly I noticed a few things which I think are bugs:
asm-i386/spinlock.h renamed from asm-i386/spinlock.h.old +0
-0"js 2f\n" \".section .text.lock,\"ax\"\n" \"2:\t" \"cmpb $0,%0\n\t" \ "rep;nop\n\t" \ "jle 2b\n\t" \asm-i386/spinlock.h renamed from asm-i386/spinlock.h.old +0
-0"jmp 1b\n" \".previous"
Quoted 5 lines
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