+Vitaly for HyperV
On Thu, Feb 06, 2020 at 04:41:06PM -0500, Peter Xu wrote:
> On Thu, Feb 06, 2020 at 01:21:20PM -0800, Sean Christopherson wrote:
> > On Thu, Feb 06, 2020 at 03:02:00PM -0500, Peter Xu wrote:
> > > But that matters to this patch because if MIPS can use
> > > kvm_flush_remote_tlbs(), then we probably don't need this
> > > arch-specific hook any more and we can directly call
> > > kvm_flush_remote_tlbs() after sync dirty log when flush==true.
> >
> > Ya, the asid_flush_mask in kvm_vz_flush_shadow_all() is the only thing
> > that prevents calling kvm_flush_remote_tlbs() directly, but I have no
> > clue as to the important of that code.
>
> As said above I think the x86 lockdep is really not necessary, then
> considering MIPS could be the only one that will use the new hook
> introduced in this patch... Shall we figure that out first?
So I prepped a follow-up patch to make kvm_arch_dirty_log_tlb_flush() a
MIPS-only hook and use kvm_flush_remote_tlbs() directly for arm and x86,
but then I realized x86 *has* a hook to do a precise remote TLB flush.
There's even an existing kvm_flush_remote_tlbs_with_address() call on a
memslot, i.e. this exact scenario. So arguably, x86 should be using the
more precise flush and should keep kvm_arch_dirty_log_tlb_flush().
But, the hook is only used when KVM is running as an L1 on top of HyperV,
and I assume dirty logging isn't used much, if at all, for L1 KVM on
HyperV?
I see three options:
1. Make kvm_arch_dirty_log_tlb_flush() MIPS-only and call
kvm_flush_remote_tlbs() directly for arm and x86. Add comments to
explain when an arch should implement kvm_arch_dirty_log_tlb_flush().
2. Change x86 to use kvm_flush_remote_tlbs_with_address() when flushing
a memslot after the dirty log is grabbed by userspace.
3. Keep the resulting code as is, but add a comment in x86's
kvm_arch_dirty_log_tlb_flush() to explain why it uses
kvm_flush_remote_tlbs() instead of the with_address() variant.
I strongly prefer to (2) or (3), but I'll defer to Vitaly as to which of
those is preferable.
I don't like (1) because (a) it requires more lines code (well comments),
to explain why kvm_flush_remote_tlbs() is the default, and (b) it would
require even more comments, which would be x86-specific in generic KVM,
to explain why x86 doesn't use its with_address() flush, or we'd lost that
info altogether.
