On Tue, Nov 14, 2017 at 02:28:56PM +0800, Wanpeng Li wrote:
> > - have the TLB invalidate handler do something like:
> >
> > state = READ_ONCE(src->preempted);
> > if (!(state & KVM_VCPU_IPI_PENDING))
> > return;
> >
> > local_flush_tlb();
> >
> > do {
> > } while (!try_cmpxchg(&src->preempted, &state,
> > state & ~KVM_VCPU_IPI_PENDING));
>
> There are a lot of cases handled by flush_tlb_func_remote() ->
> flush_tlb_function_common(), so I'm afraid to have hole.
Sure, just fix the handler to do what must be done. The above was merely
a sketch. The important part is to only clear IPI_PENDING after we do
the actual flushing, since the caller is waiting for that bit.
So flush_tlb_others() has two callers:
- arch_tlbbatch_flush() -- info::end = TLB_FLUSH_ALL
- flush_tlb_mm_range() -- info::mm = mm
native_flush_tlb_others() does smp_call_function_many(
.func=flush_tlb_func_remote) which in turn calls flush_tlb_func_common(
.local=false, .reason=TLB_REMOTE_SHOOTDOWN).
So something like:
struct flush_tlb_info info = {
.start = 0,
.end = TLB_FLUSH_ALL,
};
flush_tlb_func_common(&info, false, TLB_REMOTE_SHOOTDOWN);
should work for the new IPI. It 'upgrades' all ranges to full flushes,
but meh.
