Re: [PATCH 0/6] KVM: Optimize MMU notifier's THP page invalidation -v2

[Alexander Graf <agraf@suse.de>]

On 21.06.2012, at 10:48, Takuya Yoshikawa wrote:

> I added further optimization based on Avi's advice and my rmap_pde work.
> - patch [5-6]
> 
> The new test result was impressively good, see below, and THP page
> invalidation was more than 5 times faster on my x86 machine.
> 
> I also noticed that GUI's responsiveness on the host during swapping the
> guest's memory improved a lot.
> 
> Before:
>  ...
>  19.852 us  |  __mmu_notifier_invalidate_range_start();
>  28.033 us  |  __mmu_notifier_invalidate_range_start();
>  19.066 us  |  __mmu_notifier_invalidate_range_start();
>  44.715 us  |  __mmu_notifier_invalidate_range_start();
>  31.613 us  |  __mmu_notifier_invalidate_range_start();
>  20.659 us  |  __mmu_notifier_invalidate_range_start();
>  19.979 us  |  __mmu_notifier_invalidate_range_start();
>  20.416 us  |  __mmu_notifier_invalidate_range_start();
>  20.632 us  |  __mmu_notifier_invalidate_range_start();
>  22.316 us  |  __mmu_notifier_invalidate_range_start();
>  ...
> 
> After:
>  ...
>  4.089 us   |  __mmu_notifier_invalidate_range_start();
>  4.096 us   |  __mmu_notifier_invalidate_range_start();
>  3.560 us   |  __mmu_notifier_invalidate_range_start();
>  3.376 us   |  __mmu_notifier_invalidate_range_start();
>  3.772 us   |  __mmu_notifier_invalidate_range_start();
>  3.353 us   |  __mmu_notifier_invalidate_range_start();
>  3.332 us   |  __mmu_notifier_invalidate_range_start();
>  3.332 us   |  __mmu_notifier_invalidate_range_start();
>  3.332 us   |  __mmu_notifier_invalidate_range_start();
>  3.337 us   |  __mmu_notifier_invalidate_range_start();
>  ...
> 
> Takuya Yoshikawa (6):
>  KVM: MMU: Use __gfn_to_rmap() to clean up kvm_handle_hva()
>  KVM: Introduce hva_to_gfn_memslot() for kvm_handle_hva()
>  KVM: MMU: Make kvm_handle_hva() handle range of addresses
>  KVM: Introduce kvm_unmap_hva_range() for kvm_mmu_notifier_invalidate_range_start()
>  KVM: Separate rmap_pde from kvm_lpage_info->write_count
>  KVM: MMU: Avoid handling same rmap_pde in kvm_handle_hva_range()
> 
> arch/powerpc/include/asm/kvm_host.h |    2 +
> arch/powerpc/kvm/book3s_64_mmu_hv.c |   42 +++++++++++++++-----
> arch/x86/include/asm/kvm_host.h     |    3 +-
> arch/x86/kvm/mmu.c                  |   72 +++++++++++++++++++++++------------
> arch/x86/kvm/x86.c                  |   13 ++++++-
> include/linux/kvm_host.h            |    8 ++++
> virt/kvm/kvm_main.c                 |    3 +-
> 7 files changed, 104 insertions(+), 39 deletions(-)

Paulus, please review the book3s_hv parts more closely. I didn't spot anything obviously wrong, but rmap magic is still slightly puzzling to me, so me reviewing there doesn't mean much :).


Alex