在 2020/8/31 上午4:32, Alexander Duyck 写道: >> Right that the different level to fix this problem, but narrow the cmpxchg >> comparsion is still needed and helpful. > What I was getting at though is that I am not sure that is the case. > Normally I believe we are always holding the zone lock when updating > the migrate type. The skip flag is a one-off operation that could > easily be addressed by changing the logic to use atomic_and or > atomic_or for the cases where we are updating single bit flags and > setting the mask value to all 1's or all 0's. So adding this extra > complexity which only really applies to the skip bit may not provide > much value, especially as there are a number of possible paths that > don't use the skip bit anyway. Using atomic bit set is only helpful for skip bit. but migrate bits are still do the false sharing sync large. And actully, for this issue on cmpxchg, it's cross arch issue which it's better addressed in kernel. I see much of cmpxchg is using in kernel, but no one mentioned the shortage on this. On this pointer, this problem should be resovled in kernel. Thanks Alex
