Overview
========
        Defer pinning of the guest's pages until nested page faults occur
        to improve startup time and reduce memory pressure for SEV guests.

        Cease paying the computational cost of pinning all pages when an
        encrypted region is registered, before it is known if they will be 
accessed.

        Cease creating the memory pressure due to  pinning all pages when an
        encrypted region is registered before, it is known if they will be 
accessed.

Timing Results
==========
All timings are done by hand with and Android stopwatch app

SEV guest size(GiB)          |  4 |  8 | 16 | 32 | 60 |
without patch series(sec)    |  2 |  3 |  4 |  8 | 14 |
with patch series (sec)      |  1 |  1 |  1 |  1 |  1 |

Patch Series Changelog
==============
        1 -> 2:
                * mmio checks move from sev_pin_page() to caller and the
                  set_spte_notify* symbols are renamed to pin_page().

                * sev_get_page() warns on failures which should not happen.
                
                * kvm->srcu is used to avoid any possibility of a race in
                  sev_launch_update_data()
Applies To:
===========
        This patch applies top of this commit from the <next> branch of
        the kvm tree:
            c34b26b98cac   Tianjia Zhang : KVM: MIPS: clean up redundant 
'kvm_run' parameters

eric van tassell (4):
  KVM:MMU: Introduce the pin_page() callback
  KVM:SVM: Implement pin_page support
  KVM:SVM: Pin sev_launch_update_data() pages via sev_get_page()
  KVM:SVM: Remove struct enc_region and associated pinned page tracking.

 arch/x86/include/asm/kvm_host.h |   3 +
 arch/x86/kvm/mmu/mmu.c          |  30 +++-
 arch/x86/kvm/mmu/paging_tmpl.h  |  27 ++--
 arch/x86/kvm/svm/sev.c          | 238 +++++++++++++++++---------------
 arch/x86/kvm/svm/svm.c          |   2 +
 arch/x86/kvm/svm/svm.h          |   4 +-
 6 files changed, 174 insertions(+), 130 deletions(-)

-- 
2.17.1

Reply via email to