Christoph Lameter wrote:
One way of making the light weight counters race free for x86_64 and i386 is to use local_t. With that those two platforms are fine. However, the others fall back to atomic operations.Maybe we could deal with that on per platform basis? Some platforms may want to switch the local_t implementation away from atomics to regular integers if preemption is not configured. Most commercial Linux distros ship with preempt off. So this would preserve the speed of light weight counters, while holding off the worst races. But it still could allow interrupts while the counter is being incremented and so it would not be race free. This would no longer allow the use of local_t for refcounts but I think those uses are not that performance criticaland we may just switch to atomic. Maybe I am just off in fantasyland. Andi?Another thing to investigate (at least on ia64) is how significant the impact of a fetchadd instruction is if none of the results are used. Maybe it is tolerable and we can stay with the existing implementation.On IA64 we we would trade an interrupt disable/ load / add / store /interrupt enable against one fetchadd instruction with this patch. How bad/good a trade is that?
On one hand, switching to local counters can be a good idea if they are not evicted from the caches by the usual NRU/LRU cache replacement... On the other hand, I do not think the ia64's fetchadd instruction is expensive. If your data is in L2, then it takes 11 clock cycles. I do not think the counters have got much chance to stay in L1. Anyway, L1 is write through, you'll need to copy the updated value back into L2. As the shortest L2 access takes 5 clock cycles... You need 2 of them. (Assuming a counter is always in L2.) And add interrupt disable/enable time... Regards, Zoltan Menyhart - : send the line "unsubscribe linux-ia64" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
![]() |