Re: [PATCH v2 03/16] KVM: x86: Implement accessors for extended GPRs

[Sean Christopherson]

On Mon, Jan 12, 2026, Chang S. Bae wrote:
> Add helpers to directly read and write EGPRs (R16–R31).
> 
> Unlike legacy GPRs, EGPRs are not cached in vcpu->arch.regs[]. Their
> contents remain live in hardware. If preempted, the EGPR state is
> preserved in the guest XSAVE buffer.
> 
> The Advanced Performance Extensions (APX) feature introduces EGPRs as an
> XSAVE-managed state component. The new helpers access the registers
> directly between kvm_fpu_get() and kvm_fpu_put().
> 
> Callers should ensure that EGPRs are enabled before using these helpers.
> 
> Signed-off-by: Chang S. Bae <chang.seok.bae@xxxxxxxxx>
> ---
> V1 -> V2: Move _kvm_read_egpr()/_kvm_write_egpr() to x86.c (Paolo)
> ---
>  arch/x86/kvm/x86.c | 70 +++++++++++++++++++++++++++++++++++++++++++++-
>  1 file changed, 69 insertions(+), 1 deletion(-)
> 
> diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
> index 9857b4d319ed..edac2ec11e2f 100644
> --- a/arch/x86/kvm/x86.c
> +++ b/arch/x86/kvm/x86.c
> @@ -1256,13 +1256,81 @@ static inline u64 kvm_guest_supported_xfd(struct kvm_vcpu *vcpu)
>  #endif
>  
>  #ifdef CONFIG_KVM_APX
> +/*
> + * Accessors for extended general-purpose registers. binutils >= 2.43 can
> + * recognize those register symbols.
> + */
> +
> +static void _kvm_read_egpr(int reg, unsigned long *data)

Double underscores please (ignore the bad prior art).  And please don't use an
out-param with a void return.  That's "necessary" for e.g. _kvm_write_avx_reg()
because the value is large, but this is just 64-bits.

> +{
> +	/* mov %r16..%r31, %rax */
> +	switch (reg) {
> +	case __VCPU_XREG_R16: asm(".byte 0xd5, 0x48, 0x89, 0xc0" : "=a"(*data)); break;
> +	case __VCPU_XREG_R17: asm(".byte 0xd5, 0x48, 0x89, 0xc8" : "=a"(*data)); break;
> +	case __VCPU_XREG_R18: asm(".byte 0xd5, 0x48, 0x89, 0xd0" : "=a"(*data)); break;
> +	case __VCPU_XREG_R19: asm(".byte 0xd5, 0x48, 0x89, 0xd8" : "=a"(*data)); break;
> +	case __VCPU_XREG_R20: asm(".byte 0xd5, 0x48, 0x89, 0xe0" : "=a"(*data)); break;
> +	case __VCPU_XREG_R21: asm(".byte 0xd5, 0x48, 0x89, 0xe8" : "=a"(*data)); break;
> +	case __VCPU_XREG_R22: asm(".byte 0xd5, 0x48, 0x89, 0xf0" : "=a"(*data)); break;
> +	case __VCPU_XREG_R23: asm(".byte 0xd5, 0x48, 0x89, 0xf8" : "=a"(*data)); break;
> +	case __VCPU_XREG_R24: asm(".byte 0xd5, 0x4c, 0x89, 0xc0" : "=a"(*data)); break;
> +	case __VCPU_XREG_R25: asm(".byte 0xd5, 0x4c, 0x89, 0xc8" : "=a"(*data)); break;
> +	case __VCPU_XREG_R26: asm(".byte 0xd5, 0x4c, 0x89, 0xd0" : "=a"(*data)); break;
> +	case __VCPU_XREG_R27: asm(".byte 0xd5, 0x4c, 0x89, 0xd8" : "=a"(*data)); break;
> +	case __VCPU_XREG_R28: asm(".byte 0xd5, 0x4c, 0x89, 0xe0" : "=a"(*data)); break;
> +	case __VCPU_XREG_R29: asm(".byte 0xd5, 0x4c, 0x89, 0xe8" : "=a"(*data)); break;
> +	case __VCPU_XREG_R30: asm(".byte 0xd5, 0x4c, 0x89, 0xf0" : "=a"(*data)); break;
> +	case __VCPU_XREG_R31: asm(".byte 0xd5, 0x4c, 0x89, 0xf8" : "=a"(*data)); break;

Oof, is this really the most effecient way to encode this?  I guess so since that's
what all the SIMD instruction do, but ugh. 

> +	default: BUG();
> +	}
> +}
> +
> +static void _kvm_write_egpr(int reg, unsigned long *data)

And then take a value, not a pointer.

> +{
> +	/* mov %rax, %r16...%r31*/
> +	switch (reg) {
> +	case __VCPU_XREG_R16: asm(".byte 0xd5, 0x18, 0x89, 0xc0" : : "a"(*data)); break;
> +	case __VCPU_XREG_R17: asm(".byte 0xd5, 0x18, 0x89, 0xc1" : : "a"(*data)); break;
> +	case __VCPU_XREG_R18: asm(".byte 0xd5, 0x18, 0x89, 0xc2" : : "a"(*data)); break;
> +	case __VCPU_XREG_R19: asm(".byte 0xd5, 0x18, 0x89, 0xc3" : : "a"(*data)); break;
> +	case __VCPU_XREG_R20: asm(".byte 0xd5, 0x18, 0x89, 0xc4" : : "a"(*data)); break;
> +	case __VCPU_XREG_R21: asm(".byte 0xd5, 0x18, 0x89, 0xc5" : : "a"(*data)); break;
> +	case __VCPU_XREG_R22: asm(".byte 0xd5, 0x18, 0x89, 0xc6" : : "a"(*data)); break;
> +	case __VCPU_XREG_R23: asm(".byte 0xd5, 0x18, 0x89, 0xc7" : : "a"(*data)); break;
> +	case __VCPU_XREG_R24: asm(".byte 0xd5, 0x19, 0x89, 0xc0" : : "a"(*data)); break;
> +	case __VCPU_XREG_R25: asm(".byte 0xd5, 0x19, 0x89, 0xc1" : : "a"(*data)); break;
> +	case __VCPU_XREG_R26: asm(".byte 0xd5, 0x19, 0x89, 0xc2" : : "a"(*data)); break;
> +	case __VCPU_XREG_R27: asm(".byte 0xd5, 0x19, 0x89, 0xc3" : : "a"(*data)); break;
> +	case __VCPU_XREG_R28: asm(".byte 0xd5, 0x19, 0x89, 0xc4" : : "a"(*data)); break;
> +	case __VCPU_XREG_R29: asm(".byte 0xd5, 0x19, 0x89, 0xc5" : : "a"(*data)); break;
> +	case __VCPU_XREG_R30: asm(".byte 0xd5, 0x19, 0x89, 0xc6" : : "a"(*data)); break;
> +	case __VCPU_XREG_R31: asm(".byte 0xd5, 0x19, 0x89, 0xc7" : : "a"(*data)); break;
> +	default: BUG();
> +	}
> +}