On Thu, Sep 24, 2009 at 11:06:51AM -0300, Marcelo Tosatti wrote:
> On Mon, Sep 21, 2009 at 08:37:18PM -0300, Marcelo Tosatti wrote:
> > Use two steps for memslot deletion: mark the slot invalid (which stops
> > instantiation of new shadow pages for that slot, but allows destruction),
> > then instantiate the new empty slot.
> >
> > Also simplifies kvm_handle_hva locking.
> >
> > Signed-off-by: Marcelo Tosatti <mtosatti@xxxxxxxxxx>
> >
>
> <snip>
>
> > - if (!npages)
> > + if (!npages) {
> > + slots = kzalloc(sizeof(struct kvm_memslots), GFP_KERNEL);
> > + if (!slots)
> > + goto out_free;
> > + memcpy(slots, kvm->memslots, sizeof(struct kvm_memslots));
> > + if (mem->slot >= slots->nmemslots)
> > + slots->nmemslots = mem->slot + 1;
> > + slots->memslots[mem->slot].flags |= KVM_MEMSLOT_INVALID;
> > +
> > + old_memslots = kvm->memslots;
> > + rcu_assign_pointer(kvm->memslots, slots);
> > + synchronize_srcu(&kvm->srcu);
> > + /* From this point no new shadow pages pointing to a deleted
> > + * memslot will be created.
> > + *
> > + * validation of sp->gfn happens in:
> > + * - gfn_to_hva (kvm_read_guest, gfn_to_pfn)
> > + * - kvm_is_visible_gfn (mmu_check_roots)
> > + */
> > kvm_arch_flush_shadow(kvm);
> > + kfree(old_memslots);
> > + }
> >
> > r = kvm_arch_prepare_memory_region(kvm, &new, old, user_alloc);
> > if (r)
> > goto out_free;
> >
> > - spin_lock(&kvm->mmu_lock);
> > - if (mem->slot >= kvm->memslots->nmemslots)
> > - kvm->memslots->nmemslots = mem->slot + 1;
> > +#ifdef CONFIG_DMAR
> > + /* map the pages in iommu page table */
> > + if (npages)
> > + r = kvm_iommu_map_pages(kvm, &new);
> > + if (r)
> > + goto out_free;
> > +#endif
> >
> > - *memslot = new;
> > - spin_unlock(&kvm->mmu_lock);
> > + slots = kzalloc(sizeof(struct kvm_memslots), GFP_KERNEL);
> > + if (!slots)
> > + goto out_free;
> > + memcpy(slots, kvm->memslots, sizeof(struct kvm_memslots));
> > + if (mem->slot >= slots->nmemslots)
> > + slots->nmemslots = mem->slot + 1;
> > +
> > + /* actual memory is freed via old in kvm_free_physmem_slot below */
> > + if (!npages) {
> > + new.rmap = NULL;
> > + new.dirty_bitmap = NULL;
> > + for (i = 0; i < KVM_NR_PAGE_SIZES - 1; ++i)
> > + new.lpage_info[i] = NULL;
> > + }
> > +
> > + slots->memslots[mem->slot] = new;
> > + old_memslots = kvm->memslots;
> > + rcu_assign_pointer(kvm->memslots, slots);
> > + synchronize_srcu(&kvm->srcu);
> >
> > kvm_arch_commit_memory_region(kvm, mem, old, user_alloc);
>
> Paul,
>
> There is a scenario where this path, which updates KVM memory slots, is
> called relatively often.
>
> Each synchronize_srcu() call takes about 10ms (avg 3ms per
> synchronize_sched call), so this is hurting us.
>
> Is this expected? Is there any possibility for synchronize_srcu()
> optimization?
>
> There are other sides we can work on, such as reducing the memory slot
> updates, but i'm wondering what can be done regarding SRCU itself.
This is expected behavior, but there is a possible fix currently
in mainline (Linus's git tree). The idea would be to create a
synchronize_srcu_expedited(), which starts with synchronize_srcu(), and
replaces the synchronize_sched() calls with synchronize_sched_expedited().
This could potentially reduce the overall synchronize_srcu() latency
to well under a microsecond. The price to be paid is that each instance
of synchronize_sched_expedited() IPIs all the online CPUs, and awakens
the migration thread on each.
Would this approach likely work for you?
Thanx, Paul
--
To unsubscribe from this list: send the line "unsubscribe kvm" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
