Re: [PATCH v4 26/30] x86/svm: clear the VMCB clean bits in svm_ctxt_switch_from()

[Ross Lagerwall <ross.lagerwall@xxxxxxxxxx>]

On 10/19/26 8:42 PM, George Dunlap wrote:
> A VMCB's clean bits must all be clear when its vCPU next runs after a
> deschedule, as other VMCBs run on the pCPU in the meantime.
> svm_ctxt_switch_to() clears them as the vCPU is scheduled in; clear them
> in svm_ctxt_switch_from() instead, as it is scheduled out.
>
> This is equivalent.  While the vCPU is descheduled, accessors reaching
> its VMCB only ever clear clean bits: svm_vmexit_handler(), which sets
> them all, runs only in the vCPU's own context.  So do the nested-SVM
> paths that change the active VMCB, nsvm_vcpu_hostrestore() and
> nsvm_vcpu_vmentry(), and they clear the incoming VMCB's clean bits
> themselves.
>
> This prepares for moving VMCBs off always-mapped memory:
> svm_ctxt_switch_to() runs before write_ptbase(), where the incoming
> vCPU's VMCB will not be mapped.
>
> Assisted-by: Claude Code:claude-fable-5, Claude Code:claude-opus-5-5
> Signed-off-by: George Dunlap <gwd@xxxxxxxxxxxxxx>
> ---
> Changes in v4:
> - New in this version.
> ---
>   xen/arch/x86/hvm/svm/svm.c | 4 ++--
>   1 file changed, 2 insertions(+), 2 deletions(-)
>
> diff --git a/xen/arch/x86/hvm/svm/svm.c b/xen/arch/x86/hvm/svm/svm.c
> index 241843d8c0..4e19df9865 100644
> --- a/xen/arch/x86/hvm/svm/svm.c
> +++ b/xen/arch/x86/hvm/svm/svm.c
> @@ -919,6 +919,8 @@ static void cf_check svm_ctxt_switch_from(struct vcpu *v)
>       svm_save_dr(v);
>       svm_tsc_ratio_save(v);
>   
> +    v->arch.hvm.svm.vmcb->cleanbits.raw = 0;
> +
>       svm_sync_vmcb(v, vmcb_needs_vmload);
>       svm_vmload_pa(per_cpu(host_vmcb, cpu));
>   
> @@ -938,7 +940,6 @@ static void cf_check svm_ctxt_switch_from(struct vcpu *v)
>   
>   static void cf_check svm_ctxt_switch_to(struct vcpu *v)
>   {
> -    struct vmcb_struct *vmcb = v->arch.hvm.svm.vmcb;
>       int cpu = smp_processor_id();
>   
>       /*
> @@ -956,7 +957,6 @@ static void cf_check svm_ctxt_switch_to(struct vcpu *v)
>   
>       svm_restore_dr(v);
>   
> -    vmcb->cleanbits.raw = 0;
>       svm_tsc_ratio_load(v);
>   
>       if ( cpu_has_msr_tsc_aux )
FYI I posted "x86/svm: Don't clear clean bits on every context switch" a few
weeks ago which moves it into svm_do_resume(). I think that is preferable for
the reasons mentioned in the patch and IIUC would still be after the vCPU's
VMCB is mapped.

Ross