This looks fine to me. I agree that the comment could be less terse, perhaps something like

Looks reasonable to me. Good idea.

Works for now. If I have time over the break I’ll take a look at whether gcc could easily be extended to apply this attribute to variables where it makes sense.

I think bpf jit would need a change to work with this — a kernel equivalent to this mprotect() call for the userspace case.

Looks correct to me per the spec.

Add comments.

Change generation number to 32 bit.

In D8041#167776, @kib wrote:

The bit which I do not like about the patch is the use of uint64_t for generation. Why cannot it be uint32_t ?

In D8098#167666, @kib wrote:

The main point is that bwait() is performed by threads which are not allowed to modify buffer until B_DONE is observed. Then, as was noted, mtxpool provides the exclusion between g_up thread (executing bdone()) and whatever threads happen to do bwait().

Add PCPU field for i386.

A very tiny microoptimization for invlrng_handler, while I'm there.

Move the read of smp_tlb_generation before serializing instructions.

In D8041#166681, @kib wrote:

Are you sure about 2 cores ? Might be 4 ?

Remove unneeded wmb().

In D8041#166525, @kib wrote:

Do you have any measurements that demostrate the improvement due to the patch ? I do not object against this change going in if e.g. even micromeasurements like hwpmc or intel pcm tool show reduction of inter-core traffic.