Date:   Mon, 15 Nov 2021 15:45:48 -0800
From:   Ben Gardon <bgardon@...gle.com>
To:     linux-kernel@...r.kernel.org, kvm@...r.kernel.org
Cc:     Paolo Bonzini <pbonzini@...hat.com>, Peter Xu <peterx@...hat.com>,
        Sean Christopherson <seanjc@...gle.com>,
        Peter Shier <pshier@...gle.com>,
        David Matlack <dmatlack@...gle.com>,
        Mingwei Zhang <mizhang@...gle.com>,
        Yulei Zhang <yulei.kernel@...il.com>,
        Wanpeng Li <kernellwp@...il.com>,
        Xiao Guangrong <xiaoguangrong.eric@...il.com>,
        Kai Huang <kai.huang@...el.com>,
        Keqian Zhu <zhukeqian1@...wei.com>,
        David Hildenbrand <david@...hat.com>,
        Ben Gardon <bgardon@...gle.com>
Subject: [PATCH 00/15] Currently disabling dirty logging with the TDP MMU is
 extremely slow. On a 96 vCPU / 96G VM it takes ~45 seconds to disable dirty
 logging with the TDP MMU, as opposed to ~3.5 seconds with the legacy MMU.
 This series optimizes TLB flushes and introduces in-place large page
 promotion, to bring the disable dirty log time down to ~2 seconds.

Testing:
Ran KVM selftests and kvm-unit-tests on an Intel Skylake. This
series introduced no new failures.

Performance:
To collect these results I needed to apply Mingwei's patch
"selftests: KVM: align guest physical memory base address to 1GB"
https://lkml.org/lkml/2021/8/29/310
David Matlack is going to send out an updated version of that patch soon.

Without this series, TDP MMU:
> ./dirty_log_perf_test -v 96 -s anonymous_hugetlb_1gb
Test iterations: 2
Testing guest mode: PA-bits:ANY, VA-bits:48,  4K pages
guest physical test memory offset: 0x3fe7c0000000
Populate memory time: 10.966500447s
Enabling dirty logging time: 0.002068737s

Iteration 1 dirty memory time: 0.047556280s
Iteration 1 get dirty log time: 0.001253914s
Iteration 1 clear dirty log time: 0.049716661s
Iteration 2 dirty memory time: 3.679662016s
Iteration 2 get dirty log time: 0.000659546s
Iteration 2 clear dirty log time: 1.834329322s
Disabling dirty logging time: 45.738439510s
Get dirty log over 2 iterations took 0.001913460s. (Avg 0.000956730s/iteration)
Clear dirty log over 2 iterations took 1.884045983s. (Avg 0.942022991s/iteration)

Without this series, Legacy MMU:
> ./dirty_log_perf_test -v 96 -s anonymous_hugetlb_1gb
Test iterations: 2
Testing guest mode: PA-bits:ANY, VA-bits:48,  4K pages
guest physical test memory offset: 0x3fe7c0000000
Populate memory time: 12.664750666s
Enabling dirty logging time: 0.002025510s

Iteration 1 dirty memory time: 0.046240875s
Iteration 1 get dirty log time: 0.001864342s
Iteration 1 clear dirty log time: 0.170243637s
Iteration 2 dirty memory time: 31.571088701s
Iteration 2 get dirty log time: 0.000626245s
Iteration 2 clear dirty log time: 1.294817729s
Disabling dirty logging time: 3.566831573s
Get dirty log over 2 iterations took 0.002490587s. (Avg 0.001245293s/iteration)
Clear dirty log over 2 iterations took 1.465061366s. (Avg 0.732530683s/iteration)

With this series, TDP MMU:
(Updated since RFC. Pulling out patches 1-4 could have a performance impact.)
> ./dirty_log_perf_test -v 96 -s anonymous_hugetlb_1gb
Test iterations: 2
Testing guest mode: PA-bits:ANY, VA-bits:48,  4K pages
guest physical test memory offset: 0x3fe7c0000000
Populate memory time: 12.225242366s
Enabling dirty logging time: 0.002063442s

Iteration 1 dirty memory time: 0.047598123s
Iteration 1 get dirty log time: 0.001247702s
Iteration 1 clear dirty log time: 0.051062420s
Iteration 2 dirty memory time: 3.660439803s
Iteration 2 get dirty log time: 0.000736229s
Iteration 2 clear dirty log time: 1.043469951s
Disabling dirty logging time: 1.400549627s
Get dirty log over 2 iterations took 0.001983931s. (Avg 0.000991965s/iteration)
Clear dirty log over 2 iterations took 1.094532371s. (Avg 0.547266185s/iteration)

Patch breakdown:
Patches 1 eliminates extra TLB flushes while disabling dirty logging.
Patches 2-8 remove the need for a vCPU pointer to make_spte
Patches 9-14 are small refactors in perparation for patch 19
Patch 15 implements in-place largepage promotion when disabling dirty logging

Changelog:
RFC -> v1:
	Dropped the first 4 patches from the series. Patch 1 was sent
	separately, patches 2-4 will be taken over by Sean Christopherson.
	Incorporated David Matlack's Reviewed-by.

Ben Gardon (15):
  KVM: x86/mmu: Remove redundant flushes when disabling dirty logging
  KVM: x86/mmu: Introduce vcpu_make_spte
  KVM: x86/mmu: Factor wrprot for nested PML out of make_spte
  KVM: x86/mmu: Factor mt_mask out of make_spte
  KVM: x86/mmu: Remove need for a vcpu from
    kvm_slot_page_track_is_active
  KVM: x86/mmu: Remove need for a vcpu from mmu_try_to_unsync_pages
  KVM: x86/mmu: Factor shadow_zero_check out of make_spte
  KVM: x86/mmu: Replace vcpu argument with kvm pointer in make_spte
  KVM: x86/mmu: Factor out the meat of reset_tdp_shadow_zero_bits_mask
  KVM: x86/mmu: Propagate memslot const qualifier
  KVM: x86/MMU: Refactor vmx_get_mt_mask
  KVM: x86/mmu: Factor out part of vmx_get_mt_mask which does not depend
    on vcpu
  KVM: x86/mmu: Add try_get_mt_mask to x86_ops
  KVM: x86/mmu: Make kvm_is_mmio_pfn usable outside of spte.c
  KVM: x86/mmu: Promote pages in-place when disabling dirty logging

 arch/x86/include/asm/kvm-x86-ops.h    |  1 +
 arch/x86/include/asm/kvm_host.h       |  2 +
 arch/x86/include/asm/kvm_page_track.h |  6 +-
 arch/x86/kvm/mmu/mmu.c                | 45 +++++++------
 arch/x86/kvm/mmu/mmu_internal.h       |  6 +-
 arch/x86/kvm/mmu/page_track.c         |  8 +--
 arch/x86/kvm/mmu/paging_tmpl.h        |  6 +-
 arch/x86/kvm/mmu/spte.c               | 43 ++++++++----
 arch/x86/kvm/mmu/spte.h               | 17 +++--
 arch/x86/kvm/mmu/tdp_mmu.c            | 97 +++++++++++++++++++++------
 arch/x86/kvm/mmu/tdp_mmu.h            |  5 +-
 arch/x86/kvm/svm/svm.c                |  8 +++
 arch/x86/kvm/vmx/vmx.c                | 40 ++++++-----
 include/linux/kvm_host.h              | 10 +--
 virt/kvm/kvm_main.c                   | 12 ++--
 15 files changed, 205 insertions(+), 101 deletions(-)

-- 
2.34.0.rc1.387.gb447b232ab-goog

Powered by blists - more mailing lists