Messages in this thread | | | Date | Wed, 24 Aug 2011 16:10:21 -0700 | From | Jeremy Fitzhardinge <> | Subject | Re: [PATCH 00/18] x86: Ticket lock + cmpxchg cleanup |
| |
On 08/24/2011 04:05 PM, H. Peter Anvin wrote: > On 08/24/2011 03:59 PM, Jeremy Fitzhardinge wrote: >> On 08/24/2011 03:53 PM, Linus Torvalds wrote: >>> On Wed, Aug 24, 2011 at 1:10 PM, Jeremy Fitzhardinge <jeremy@goop.org> wrote: >>>> Could we just kill SMP support for OOSTORE machines now? That would be >>>> the cleanest possible fix... >>> No it wouldn't. The asm version would *still* be cleaner than the "C >>> plus random barriers". >>> >>> It's not like the C version is "portable" in any case. >> If there's no need to have a locked instruction, then it could simply be: >> >> barrier(); >> lock->head++; >> barrier(); >> >> with no need for asm at all. >> > That's not guaranteed in any way to generate a locally atomic instruction.
Doesn't need to be. The final write needs to be locally atomic, but we assume that a lot.
J
| |