Use atomic_interrupt_fence() instead of bare __compiler_membar()

[kib]

Summary

for the which which definitely use membar to sync with interrupt handlers.

libc and rtld uses of __compiler_membar() seems to want compiler barriers proper.

There are two special cases:

1. kpilite sched_unpin_lite() fence after td_pinned decrement. I am not sure if we need it at all
2. x86/include/bus.h (not handled in the patch) where I am not sure why do we need seq_cst fence at all for bus_space_barrier()

[markj]

I'm not very familiar with bus_space_barrier(). I'm wondering if it's trying to handle the case where the resource is mapped WC?

sys/sys/kpilite.h

51–52

I suspect it is duplicated from critical_exit(). The same commit that added kpilite.h also added an inline critical_exit() to systm.h, and there the second fence is indeed required. I think you can just drop it here.

[kib]

In D28956#648518, @markj wrote:

I'm not very familiar with bus_space_barrier(). I'm wondering if it's trying to handle the case where the resource is mapped WC?

For WC fencing, locked op might be not enough. For instance, for fencing CLFLUSHOPT, SDM requires SFENCE. I think it is right to not touch bus.h in this commit.