Message-Id: <20210730220455.26054-1-peterx@redhat.com>
Date:   Fri, 30 Jul 2021 18:04:48 -0400
From:   Peter Xu <peterx@...hat.com>
To:     kvm@...r.kernel.org, linux-kernel@...r.kernel.org
Cc:     Vitaly Kuznetsov <vkuznets@...hat.com>,
        Sean Christopherson <seanjc@...gle.com>, peterx@...hat.com,
        Maxim Levitsky <mlevitsk@...hat.com>,
        Paolo Bonzini <pbonzini@...hat.com>
Subject: [PATCH v3 0/7] KVM: X86: Some light optimizations on rmap logic

Major change to v3 is to address comments from Sean.

Since I retested the two relevant patches and the numbers changed slightly, I
updated the numbers in the two optimization patches to reflect that.  In the
latest measurement the 3->15 slots change showed more effect on the speedup.
Summary:

        Vanilla:      473.90 (+-5.93%)
        3->15 slots:  366.10 (+-4.94%)
        Add counter:  351.00 (+-3.70%)

All the numbers are also updated in the commit messages.

To apply the series upon kvm/queue, below patches should be replaced by the
corresponding patches in this v3:

        KVM: X86: MMU: Tune PTE_LIST_EXT to be bigger
        KVM: X86: Optimize pte_list_desc with per-array counter
        KVM: X86: Optimize zapping rmap

The 1st oneliner patch needs to be replaced because the commit message is
updated with the new numbers so to align all the numbers, the 2nd-3rd patches
are for addressing Sean's comments and also with the new numbers.

I didn't repost the initial two patches because they're already in kvm/queue
and they'll be identical in content.  Please have a look, thanks.

v2: https://lore.kernel.org/kvm/20210625153214.43106-1-peterx@redhat.com/
v1: https://lore.kernel.org/kvm/20210624181356.10235-1-peterx@redhat.com/

-- original cover letter --

All things started from patch 1, which introduced a new statistic to keep "max
rmap entry count per vm".  At that time I was just curious about how many rmap
is there normally for a guest, and it surprised me a bit.

For TDP mappings it's all fine as mostly rmap of a page is either 0 or 1
depending on faulted or not.  It turns out with EPT=N there seems to be a huge
number of pages that can have tens or hundreds of rmap entries even for an idle
guest.  Then I continued with the rest.

To understand better on "how much of those pages", I did patch 2-6 which
introduced the idea of per-arch per-vm debugfs nodes, and added a debug file to
do statistics for rmap, which is similar to kvm_arch_create_vcpu_debugfs() but
for vm not vcpu.

I did notice this should be the clean approach as I also see other archs
randomly create some per-vm debugfs nodes there:

---8<---
*** arch/arm64/kvm/vgic/vgic-debug.c:
vgic_debug_init[274]           debugfs_create_file("vgic-state", 0444, kvm->debugfs_dentry, kvm,

*** arch/powerpc/kvm/book3s_64_mmu_hv.c:
kvmppc_mmu_debugfs_init[2115]  debugfs_create_file("htab", 0400, kvm->arch.debugfs_dir, kvm,

*** arch/powerpc/kvm/book3s_64_mmu_radix.c:
kvmhv_radix_debugfs_init[1434] debugfs_create_file("radix", 0400, kvm->arch.debugfs_dir, kvm,

*** arch/powerpc/kvm/book3s_hv.c:
debugfs_vcpu_init[2395]        debugfs_create_file("timings", 0444, vcpu->arch.debugfs_dir, vcpu,

*** arch/powerpc/kvm/book3s_xics.c:
xics_debugfs_init[1027]        xics->dentry = debugfs_create_file(name, 0444, powerpc_debugfs_root,

*** arch/powerpc/kvm/book3s_xive.c:
xive_debugfs_init[2236]        xive->dentry = debugfs_create_file(name, S_IRUGO, powerpc_debugfs_root,

*** arch/powerpc/kvm/timing.c:
kvmppc_create_vcpu_debugfs[214] debugfs_file = debugfs_create_file(dbg_fname, 0666, kvm_debugfs_dir,
---8<---

PPC even has its own per-vm dir for that.  I think if patch 2-6 can be
considered to be accepted then the next thing to consider is to merge all these
usages to be under the same existing per-vm dentry with their per-arch hooks
introduced.

The last 3 patches (patch 7-9) are a few optimizations of existing rmap logic.
The major test case I used is rmap_fork [1], however it's not really the ideal
one to show their effect for sure as that test I wrote covers both
rmap_add/remove, while I don't have good idea on optimizing rmap_remove without
changing the array structure or adding much overhead (e.g. sort the array, or
making a tree-like structure somehow to replace the array list).  However it
already shows some benefit with those changes, so I post them out.

Applying patch 7-8 will bring a summary of 38% perf boost when I fork 500
childs with the test I used.  Didn't run perf test on patch 9.  More in the
commit log.

Please review, thanks.

[1] https://github.com/xzpeter/clibs/commit/825436f825453de2ea5aaee4bdb1c92281efe5b3

Peter Xu (7):
  KVM: Allow to have arch-specific per-vm debugfs files
  KVM: X86: Introduce pte_list_count() helper
  KVM: X86: Introduce kvm_mmu_slot_lpages() helpers
  KVM: X86: Introduce mmu_rmaps_stat per-vm debugfs file
  KVM: X86: MMU: Tune PTE_LIST_EXT to be bigger
  KVM: X86: Optimize pte_list_desc with per-array counter
  KVM: X86: Optimize zapping rmap

 arch/x86/kvm/mmu/mmu.c          |  98 +++++++++++++++++-------
 arch/x86/kvm/mmu/mmu_internal.h |   1 +
 arch/x86/kvm/x86.c              | 130 +++++++++++++++++++++++++++++++-
 include/linux/kvm_host.h        |   1 +
 virt/kvm/kvm_main.c             |  20 ++++-
 5 files changed, 221 insertions(+), 29 deletions(-)

-- 
2.31.1

Powered by blists - more mailing lists