Re: [PATCH v2 00/20] KVM: Use atomic SRCU for gfn-to-pfn cache, reinstate guest mode for x86 nesting
From: David Woodhouse
Date: Tue Nov 10 2026 - 14:55:07 EST
On Thu, 2026-11-05 at 02:00 +0000, David Woodhouse wrote:
> From: David Woodhouse <dwmw@xxxxxxxxxxxx>
>
> Based on v7.4-rc1 (kvm/next doesn't have the atomic SRCU flavour yet),
> this series:
>
> • Converts the GPC to use atomic SRCU, ditching the rwlock which was a
> scaling and latency problem, and which PREEMPT_RT hated because it
> turned the rwlock into a sleeping lock¹.
>
> • Fixes the pre-existing bugs which torture-testing that conversion
> shook out: a guest-reachable dirty ring overflow from the hypercall
> page write (128 pushes of the same gfn against 64 reserved entries,
> without PML), and a lost TLB flush request in record_steal_time().
>
> • Adds KVM_XEN_VCPU_ATTR_TYPE_WRITE_HYPERCALL_PAGE, for VMMs which
> intercept the guest's MSR write and need to replay it from host
> context².
>
> • Cleans up the intentional lack of dirty-tracking of Xen shinfo and
> vcpu_info pages (Sean).
>
> • Converts steal-time / preempted status to use gfn_to_pfn_cache.
> (Carsten)
>
I love it when sashiko-bot finds things which my own local runs don't.
I'm still iterating on the latter part of the series; I think it
probably makes sense to merge the first half (patches 1-9 above), and
the nested parts (below) can come later.
Sean, what do you think?
> • Reinstates guest-mode pinning: replacing the old GUEST_USES_PFN
> mode, letting nested state pages be cached and pinned while in
> guest context.
>
> • nVMX: Uses GPC for the L1 MSR bitmap, and guest mode for APIC
> access and vAPIC pages. (Fred)
>
> • nSVM: Uses GPC for vmcb12 / MSR-permissions / IO-permissions pages.
> This cuts the vmcb12 transition cost measurably (selftest included).
>
> • Re-posts KVM_REQ_GET_NESTED_STATE_PAGES on memslot updates so the
> pinned caches are revalidated lazily, riding out the transient
> KVM_MEMSLOT_INVALID window of an in-flight update via the pending
> request rather than ever entering L2 with a stale latched address.
>
> I had briefly split out some of the preliminary fixes and cleanups
> into a separate series which didn't depend on atomic SRCU. But I've
> joined it all back together now that atomic SRCU is merged. And I've
> ditched half of those 'fixes' as a bad idea anyway³.
>
>
> Tested: per-commit builds across the series; KVM selftests including
> the new invalidation, transition and race tests on 192-way AMD, and
> the applicable subset on Intel under PROVE_LOCKING + PROVE_RCU +
> DEBUG_ATOMIC_SLEEP (which catches the attribute-ioctl srcu fix on the
> very first xen_shinfo_test run when it's reverted; that config is now
> a standing part of the testing here). The new mmu_notifier race test
> asserts that every one of 5000 event channel deliveries into HVA-based
> caches survives continuous MADV_DONTNEED invalidation of the backing
> page, and soaks at 10/10.
>
> ¹ https://syzkaller.appspot.com/bug?extid=208f7f3e5f59c11aeb90
> ² https://lore.kernel.org/kvm/c855535b4262ecd41f67734d19e8f48a7f014c2a.camel@xxxxxxxxxxxxx/
> ³ https://lore.kernel.org/kvm/ac6b1891d6bdf3a855a5b840c462a6e7d464a146.camel@xxxxxxxxxxxxx/
>
> v2: https://git.infradead.org/users/dwmw2/linux.git/shortlog/refs/heads/gpc-srcu-v2
>
> • Absorb the fixes/cleanups which had been split out (and drop the
> delivery-retry half of them³, plus the SCHEDOP_poll fix which Sean
> is picking up separately)
> • Fix the mmu_notifier walk to wait out a grace period still owed by
> a concurrent mutator
> • Keep kvm->srcu in the Xen attribute ioctls for the memslot lookups
> • Pair smp_mb__after_atomic()/__after_spinlock() for the publish/veto
> store->load ordering
> • Nested: treat a refresh failure against a KVM_MEMSLOT_INVALID
> placeholder as transient; re-set vmcs02 controls after a lazy
> degradation heals; initialize caches once at vCPU creation
> • New: dirty ring overflow fix, WRITE_HYPERCALL_PAGE attribute,
> the two Xen race selftests
>
> v1: https://git.infradead.org/users/dwmw2/linux.git/shortlog/refs/heads/gpc-srcu
> https://lore.kernel.org/kvm/20260920211920.928306-1-dwmw2@xxxxxxxxxxxxx/
>
> Carsten Stollmaier (1):
> KVM: x86: Use gfn_to_pfn_cache for steal time / preempted status
>
> David Woodhouse (13):
> KVM: x86/xen: Fill the hypercall page with a single guest write
> KVM: x86/xen: Add KVM_XEN_VCPU_ATTR_TYPE_WRITE_HYPERCALL_PAGE
> KVM: x86: Request the guest TLB flush from record_steal_time()
> KVM: selftests: Actually deactivate the HVA-based shared_info cache in
> xen_shinfo_test
> KVM: pfncache: Use atomic SRCU for readers instead of a rwlock
> KVM: pfncache: Add guest-mode pinning (GUEST_USES_PFN successor)
> KVM: x86: Post KVM_REQ_GET_NESTED_STATE_PAGES on memslot updates
> KVM: nSVM: Use a gfn_to_pfn_cache for the vmcb12 page
> KVM: nSVM: Cache L1's MSR permissions map pages
> KVM: nSVM: Cache L1's IO permissions map pages
> KVM: selftests: Add nested transition benchmark
> KVM: selftests: Add Xen evtchn vs memslot-update race test
> KVM: selftests: Add Xen evtchn vs mmu_notifier race test for HVA
> caches
>
> Fred Griffoul (3):
> KVM: nVMX: Implement cache for L1 MSR bitmap
> KVM: nVMX: Use pinned pfncache for L1 APIC virtualization pages
> KVM: selftests: Add nested VMX APIC cache invalidation test
>
> Sean Christopherson (3):
> KVM: x86/xen: Extract delivery of event to vCPU into a separate helper
> KVM: x86/xen: Explicitly tag "shared info" page as never being dirty
> tracked
> KVM: x86/xen: Don't dirty track "vCPU info" page
>
> Documentation/virt/kvm/api.rst | 27 +-
> arch/x86/include/asm/kvm_host.h | 2 +-
> arch/x86/include/uapi/asm/kvm.h | 3 +
> arch/x86/kvm/msrs.c | 7 +-
> arch/x86/kvm/svm/nested.c | 174 ++++-
> arch/x86/kvm/svm/svm.c | 17 +
> arch/x86/kvm/svm/svm.h | 29 +
> arch/x86/kvm/vmx/nested.c | 517 +++++++++++--
> arch/x86/kvm/vmx/nested.h | 1 +
> arch/x86/kvm/vmx/vmx.c | 13 +-
> arch/x86/kvm/vmx/vmx.h | 21 +-
> arch/x86/kvm/x86.c | 209 ++++--
> arch/x86/kvm/x86.h | 33 +
> arch/x86/kvm/xen.c | 380 ++++++----
> include/linux/kvm_host.h | 121 +++-
> include/linux/kvm_types.h | 64 +-
> include/linux/srcu.h | 7 +
> tools/testing/selftests/kvm/Makefile.kvm | 4 +
> .../kvm/x86/nested_transition_bench.c | 210 ++++++
> .../selftests/kvm/x86/vmx_apic_update_test.c | 323 +++++++++
> .../kvm/x86/xen_evtchn_hva_notify_race_test.c | 188 +++++
> .../kvm/x86/xen_evtchn_memslot_race_test.c | 232 ++++++
> .../selftests/kvm/x86/xen_shinfo_test.c | 2 +-
> .../selftests/kvm/x86/xen_vmcall_test.c | 118 +++
> virt/kvm/kvm_main.c | 9 +
> virt/kvm/pfncache.c | 677 +++++++++++++++---
> 26 files changed, 2962 insertions(+), 426 deletions(-)
> create mode 100644 tools/testing/selftests/kvm/x86/nested_transition_bench.c
> create mode 100644 tools/testing/selftests/kvm/x86/vmx_apic_update_test.c
> create mode 100644 tools/testing/selftests/kvm/x86/xen_evtchn_hva_notify_race_test.c
> create mode 100644 tools/testing/selftests/kvm/x86/xen_evtchn_memslot_race_test.c
>
> base-commit: 6628e7d4c6b25c8d18f1345e3a20d21829aae7e4
Attachment:
smime.p7s
Description: S/MIME cryptographic signature