On Wed, Nov 27, 2019 at 10:47:36AM -0800, Sean Christopherson wrote:
> On Thu, Sep 26, 2019 at 04:18:02PM -0700, Ben Gardon wrote:
> > Replace the KVM MMU spinlock with a read/write lock so that some parts of
> > the MMU can be made more concurrent in future commits by switching some
> > write mode aquisitions to read mode. A read/write lock was chosen over
> > other synchronization options beause it has minimal initial impact: this
> > change simply changes all uses of the MMU spin lock to an MMU read/write
> > lock, in write mode. This change has no effect on the logic of the code
> > and only a small performance penalty.
> >
> > Other, more invasive options were considered for synchronizing access to
> > the paging structures. Sharding the MMU lock to protect 2MB chunks of
> > addresses, as the main MM does, would also work, however it makes
> > acquiring locks for operations on large regions of memory expensive.
> > Further, the parallel page fault handling algorithm introduced later in
> > this series does not require exclusive access to the region of memory
> > for which it is handling a fault.
> >
> > There are several disadvantages to the read/write lock approach:
> > 1. The reader/writer terminology does not apply well to MMU operations.
> > 2. Many operations require exclusive access to a region of memory
> > (often a memslot), but not all of memory. The read/write lock does not
> > facilitate this.
> > 3. Contention between readers and writers can still create problems in
> > the face of long running MMU operations.
> >
> > Despite these issues,the use of a read/write lock facilitates
> > substantial improvements over the monolithic locking scheme.
> >
> > Signed-off-by: Ben Gardon <bgardon@xxxxxxxxxx>
> > ---
> > arch/x86/kvm/mmu.c | 106 +++++++++++++++++++------------------
> > arch/x86/kvm/page_track.c | 8 +--
> > arch/x86/kvm/paging_tmpl.h | 8 +--
> > arch/x86/kvm/x86.c | 4 +-
> > include/linux/kvm_host.h | 3 +-
> > virt/kvm/kvm_main.c | 34 ++++++------
> > 6 files changed, 83 insertions(+), 80 deletions(-)
> >
> > diff --git a/arch/x86/kvm/mmu.c b/arch/x86/kvm/mmu.c
> > index 56587655aecb9..0311d18d9a995 100644
> > --- a/arch/x86/kvm/mmu.c
> > +++ b/arch/x86/kvm/mmu.c
> > @@ -2446,9 +2446,9 @@ static void mmu_sync_children(struct kvm_vcpu *vcpu,
> > flush |= kvm_sync_page(vcpu, sp, &invalid_list);
> > mmu_pages_clear_parents(&parents);
> > }
> > - if (need_resched() || spin_needbreak(&vcpu->kvm->mmu_lock)) {
>
> I gather there is no equivalent to spin_needbreak() for r/w locks? Is it
> something that can be added? Losing spinlock contention detection will
> negatively impact other flows, e.g. fast zapping all pages will no longer
> drop the lock to allow insertion of SPTEs into the new generation of MMU.
Just saw that fast zap is explicitly noted in the cover letter. Is there
anything beyond a spin_needbreak() implementation that's needed to support
fast zap?
> > + if (need_resched()) {
> > kvm_mmu_flush_or_zap(vcpu, &invalid_list, false, flush);
> > - cond_resched_lock(&vcpu->kvm->mmu_lock);
> > + cond_resched_rwlock_write(&vcpu->kvm->mmu_lock);
> > flush = false;
> > }
> > }
