Re: [PATCH] KVM: x86/mmu: optimizing the code in mmu_try_to_unsync_pages

[Yun Lu]

On 2022/5/20 下午10:47, Sean Christopherson wrote:

> On Fri, May 20, 2022, Yuan Yao wrote:
> > On Fri, May 20, 2022 at 02:09:07PM +0800, Yun Lu wrote:
> > > There is no need to check can_unsync and prefetch in the loop
> > > every time, just move this check before the loop.
> > >
> > > Signed-off-by: Yun Lu <luyun@xxxxxxxxxx>
> > > ---
> > >   arch/x86/kvm/mmu/mmu.c | 12 ++++++------
> > >   1 file changed, 6 insertions(+), 6 deletions(-)
> > >
> > > diff --git a/arch/x86/kvm/mmu/mmu.c b/arch/x86/kvm/mmu/mmu.c
> > > index 311e4e1d7870..e51e7735adca 100644
> > > --- a/arch/x86/kvm/mmu/mmu.c
> > > +++ b/arch/x86/kvm/mmu/mmu.c
> > > @@ -2534,6 +2534,12 @@ int mmu_try_to_unsync_pages(struct kvm *kvm, const struct kvm_memory_slot *slot,
> > >   	if (kvm_slot_page_track_is_active(kvm, slot, gfn, KVM_PAGE_TRACK_WRITE))
> > >   		return -EPERM;
> > >
> > > +	if (!can_unsync)
> > > +		return -EPERM;
> > > +
> > > +	if (prefetch)
> > > +		return -EEXIST;
> > > +
> > >   	/*
> > >   	 * The page is not write-tracked, mark existing shadow pages unsync
> > >   	 * unless KVM is synchronizing an unsync SP (can_unsync = false).  In
> > > @@ -2541,15 +2547,9 @@ int mmu_try_to_unsync_pages(struct kvm *kvm, const struct kvm_memory_slot *slot,
> > >   	 * allowing shadow pages to become unsync (writable by the guest).
> > >   	 */
> > >   	for_each_gfn_indirect_valid_sp(kvm, sp, gfn) {
> > > -		if (!can_unsync)
> > > -			return -EPERM;
> > > -
> > >   		if (sp->unsync)
> > >   			continue;
> > >
> > > -		if (prefetch)
> > > -			return -EEXIST;
> > > -
> > Consider the case that for_each_gfn_indirect_valid_sp() loop is
> > not triggered, means the gfn is not MMU page table page:
> >
> > The old behavior when : return 0;
> > The new behavior with this change: returrn -EPERM / -EEXIST;
> >
> > It at least breaks FNAME(sync_page) -> make_spte(prefetch = true, can_unsync = false)
> > which removes PT_WRITABLE_MASK from last level mapping unexpectedly.
> Yep, the flags should be queried if and only if there's at least one valid, indirect
> SP for th gfn.  And querying whether there's such a SP is quite expesnive and requires
> looping over a list, so checking every iteration of the loop is far cheaper.  E.g. each
> check is a single uop on modern CPUs as both gcc and clang are smart enough to stash
> the flags in registers so that there's no reload from memory on each loop.  And that
> also means the CPU can more than likely correctly predict subsequent iterations.
OK, it's my mistake.  Thanks for your answers.