Re: missing kvm smp tlb flush in invlpg

[Andrea Arcangeli <aarcange@redhat.com>]

On Sun, Mar 15, 2009 at 12:35:48PM +0200, Avi Kivity wrote:
> Izik pointed out that for invlpg, the guest is responsible for smp tlb 
> flushes, and mmu notifiers will protect against pageout.

How will mmu notifier protect against pageout if the spte is already
invalid and removed from the rmapp chain? mmu notifier will search the
rmapp chain and it'll find nothing, it'll do nothing, so then the page
will be freed under the other cpus without no ipi flushing their VT
tlbs.

All that mmu notifier does is to protect against pageout until the
mmu_lock is released. So that you can send a single ipi to the other
physical cpus after a flood of rmap_remove, without having to do the
array of pages like arch/x86/include/asm/tlb.h.

This because if the VM was in the process of swapping out that page
while we were inside the mmu_lock protected critical section, the mmu
notifier will force the swap path to take the vcpu->kvm->mmu_lock
first for each kvm instance registered with the mmu notifier. But
after taking that lock, the mmu notifier will do nothing if
rmap_remove already run before the mmu_lock was released (like in this
case). The mmu_lock is just to stop temporarily the swap, so that it
waits the ipi is delivered to all cpus before proceeding freeing the
page. It's up to the kvm code that takes the lock to flush the tlb of
any other running guest, before it is allowed to release the mmu_lock
as far as I can tell.