mbox series

[0/3] KVM: x86: Use "standard" mmu_notifier hook for APIC page

Message ID 20230602011518.787006-1-seanjc@google.com
Headers
Series KVM: x86: Use "standard" mmu_notifier hook for APIC page |

Message

Sean Christopherson June 2, 2023, 1:15 a.m. UTC 
  Convert VMX's handling of mmu_notifier invalidations of the APIC-access page
from invalidate_range() to KVM's standard invalidate_range_{start,end}().

KVM (ab)uses invalidate_range() to fudge around not stalling vCPUs until
relevant in-flight invalidations complete.  Abusing invalidate_range() works,
but it requires one-off code in KVM, sets a bad precedent in KVM, and is
blocking improvements to mmu_notifier's definition of invalidate_range()
due to KVM's usage diverging wildly from the original intent of notifying
IOMMUs of changes to shared page tables.

Clean up the mess by hooking x86's implementation of kvm_unmap_gfn_range()
and stalling vCPUs by re-requesting KVM_REQ_APIC_PAGE_RELOAD until the
invalidation completes.

Sean Christopherson (3):
  KVM: VMX: Retry APIC-access page reload if invalidation is in-progress
  KVM: x86: Use standard mmu_notifier invalidate hooks for APIC access
    page
  KVM: x86/mmu: Trigger APIC-access page reload iff vendor code cares

 arch/x86/kvm/mmu/mmu.c   |  4 ++++
 arch/x86/kvm/vmx/vmx.c   | 50 ++++++++++++++++++++++++++++++++++++----
 arch/x86/kvm/x86.c       | 14 -----------
 include/linux/kvm_host.h |  3 ---
 virt/kvm/kvm_main.c      | 18 ---------------
 5 files changed, 49 insertions(+), 40 deletions(-)


base-commit: 39428f6ea9eace95011681628717062ff7f5eb5f

Comments

Paolo Bonzini June 5, 2023, 10:15 a.m. UTC | #1 
On 6/2/23 03:15, Sean Christopherson wrote:
> Convert VMX's handling of mmu_notifier invalidations of the APIC-access page
> from invalidate_range() to KVM's standard invalidate_range_{start,end}().
> 
> KVM (ab)uses invalidate_range() to fudge around not stalling vCPUs until
> relevant in-flight invalidations complete.  Abusing invalidate_range() works,
> but it requires one-off code in KVM, sets a bad precedent in KVM, and is
> blocking improvements to mmu_notifier's definition of invalidate_range()
> due to KVM's usage diverging wildly from the original intent of notifying
> IOMMUs of changes to shared page tables.
> 
> Clean up the mess by hooking x86's implementation of kvm_unmap_gfn_range()
> and stalling vCPUs by re-requesting KVM_REQ_APIC_PAGE_RELOAD until the
> invalidation completes.
> 
> Sean Christopherson (3):
>    KVM: VMX: Retry APIC-access page reload if invalidation is in-progress
>    KVM: x86: Use standard mmu_notifier invalidate hooks for APIC access
>      page
>    KVM: x86/mmu: Trigger APIC-access page reload iff vendor code cares
> 
>   arch/x86/kvm/mmu/mmu.c   |  4 ++++
>   arch/x86/kvm/vmx/vmx.c   | 50 ++++++++++++++++++++++++++++++++++++----
>   arch/x86/kvm/x86.c       | 14 -----------
>   include/linux/kvm_host.h |  3 ---
>   virt/kvm/kvm_main.c      | 18 ---------------
>   5 files changed, 49 insertions(+), 40 deletions(-)
> 
> 
> base-commit: 39428f6ea9eace95011681628717062ff7f5eb5f


Reviewed-by: Paolo Bonzini <pbonzini@redhat.com>

Paolo
Jason Gunthorpe June 6, 2023, 4:43 p.m. UTC | #2 
On Thu, Jun 01, 2023 at 06:15:15PM -0700, Sean Christopherson wrote:
> Convert VMX's handling of mmu_notifier invalidations of the APIC-access page
> from invalidate_range() to KVM's standard invalidate_range_{start,end}().
> 
> KVM (ab)uses invalidate_range() to fudge around not stalling vCPUs until
> relevant in-flight invalidations complete.  Abusing invalidate_range() works,
> but it requires one-off code in KVM, sets a bad precedent in KVM, and is
> blocking improvements to mmu_notifier's definition of invalidate_range()
> due to KVM's usage diverging wildly from the original intent of notifying
> IOMMUs of changes to shared page tables.
> 
> Clean up the mess by hooking x86's implementation of kvm_unmap_gfn_range()
> and stalling vCPUs by re-requesting KVM_REQ_APIC_PAGE_RELOAD until the
> invalidation completes.

I don't know much about kvm, but this looks like what I had in mind
and is a good way to use mmu notifiers

Thanks,
Jason
Sean Christopherson June 7, 2023, 12:55 a.m. UTC | #3 
On Thu, 01 Jun 2023 18:15:15 -0700, Sean Christopherson wrote:
> Convert VMX's handling of mmu_notifier invalidations of the APIC-access page
> from invalidate_range() to KVM's standard invalidate_range_{start,end}().
> 
> KVM (ab)uses invalidate_range() to fudge around not stalling vCPUs until
> relevant in-flight invalidations complete.  Abusing invalidate_range() works,
> but it requires one-off code in KVM, sets a bad precedent in KVM, and is
> blocking improvements to mmu_notifier's definition of invalidate_range()
> due to KVM's usage diverging wildly from the original intent of notifying
> IOMMUs of changes to shared page tables.
> 
> [...]

Applied to kvm-x86 vmx, thanks!

[1/3] KVM: VMX: Retry APIC-access page reload if invalidation is in-progress
      https://github.com/kvm-x86/linux/commit/878940b33d76
[2/3] KVM: x86: Use standard mmu_notifier invalidate hooks for APIC access page
      https://github.com/kvm-x86/linux/commit/0a8a5f2c8c26
[3/3] KVM: x86/mmu: Trigger APIC-access page reload iff vendor code cares
      https://github.com/kvm-x86/linux/commit/0a3869e14d4a

--
https://github.com/kvm-x86/linux/tree/next
https://github.com/kvm-x86/linux/tree/fixes