* [PATCH RFC v1 01/20] KVM: x86: Rename register accessors to be GPR-specific
2025-11-10 18:01 [PATCH RFC v1 00/20] KVM: x86: Support APX feature for guests Chang S. Bae
@ 2025-11-10 18:01 ` Chang S. Bae
2025-11-10 18:01 ` [PATCH RFC v1 02/20] KVM: x86: Refactor GPR accessors to differentiate register access types Chang S. Bae
` (20 subsequent siblings)
21 siblings, 0 replies; 55+ messages in thread
From: Chang S. Bae @ 2025-11-10 18:01 UTC (permalink / raw)
To: kvm, linux-kernel; +Cc: pbonzini, seanjc, chao.gao, zhao1.liu, chang.seok.bae
Refactor the VCPU register state accessors to make them explicitly
GPR-only.
The existing register accessors operate on the cached VCPU register
state. That cache holds GPRs and RIP. RIP has its own interface already.
This renaming clarifies GPR access only.
No functional changes intended.
Signed-off-by: Chang S. Bae <chang.seok.bae@intel.com>
---
arch/x86/kvm/svm/svm.c | 8 ++++----
arch/x86/kvm/vmx/nested.c | 20 ++++++++++----------
arch/x86/kvm/vmx/vmx.c | 12 ++++++------
arch/x86/kvm/x86.c | 10 +++++-----
arch/x86/kvm/x86.h | 5 ++---
arch/x86/kvm/xen.c | 2 +-
6 files changed, 28 insertions(+), 29 deletions(-)
diff --git a/arch/x86/kvm/svm/svm.c b/arch/x86/kvm/svm/svm.c
index 153c12dbf3eb..3aa2c37754ef 100644
--- a/arch/x86/kvm/svm/svm.c
+++ b/arch/x86/kvm/svm/svm.c
@@ -2473,7 +2473,7 @@ static int cr_interception(struct kvm_vcpu *vcpu)
err = 0;
if (cr >= 16) { /* mov to cr */
cr -= 16;
- val = kvm_register_read(vcpu, reg);
+ val = kvm_gpr_read(vcpu, reg);
trace_kvm_cr_write(cr, val);
switch (cr) {
case 0:
@@ -2519,7 +2519,7 @@ static int cr_interception(struct kvm_vcpu *vcpu)
kvm_queue_exception(vcpu, UD_VECTOR);
return 1;
}
- kvm_register_write(vcpu, reg, val);
+ kvm_gpr_write(vcpu, reg, val);
trace_kvm_cr_read(cr, val);
}
return kvm_complete_insn_gp(vcpu, err);
@@ -2591,9 +2591,9 @@ static int dr_interception(struct kvm_vcpu *vcpu)
dr = svm->vmcb->control.exit_code - SVM_EXIT_READ_DR0;
if (dr >= 16) { /* mov to DRn */
dr -= 16;
- err = kvm_set_dr(vcpu, dr, kvm_register_read(vcpu, reg));
+ err = kvm_set_dr(vcpu, dr, kvm_gpr_read(vcpu, reg));
} else {
- kvm_register_write(vcpu, reg, kvm_get_dr(vcpu, dr));
+ kvm_gpr_write(vcpu, reg, kvm_get_dr(vcpu, dr));
}
return kvm_complete_insn_gp(vcpu, err);
diff --git a/arch/x86/kvm/vmx/nested.c b/arch/x86/kvm/vmx/nested.c
index 76271962cb70..47a941989787 100644
--- a/arch/x86/kvm/vmx/nested.c
+++ b/arch/x86/kvm/vmx/nested.c
@@ -5325,9 +5325,9 @@ int get_vmx_mem_address(struct kvm_vcpu *vcpu, unsigned long exit_qualification,
else if (addr_size == 0)
off = (gva_t)sign_extend64(off, 15);
if (base_is_valid)
- off += kvm_register_read(vcpu, base_reg);
+ off += kvm_gpr_read(vcpu, base_reg);
if (index_is_valid)
- off += kvm_register_read(vcpu, index_reg) << scaling;
+ off += kvm_gpr_read(vcpu, index_reg) << scaling;
vmx_get_segment(vcpu, &s, seg_reg);
/*
@@ -5719,7 +5719,7 @@ static int handle_vmread(struct kvm_vcpu *vcpu)
return 1;
/* Decode instruction info and find the field to read */
- field = kvm_register_read(vcpu, (((instr_info) >> 28) & 0xf));
+ field = kvm_gpr_read(vcpu, (((instr_info) >> 28) & 0xf));
if (!nested_vmx_is_evmptr12_valid(vmx)) {
/*
@@ -5768,7 +5768,7 @@ static int handle_vmread(struct kvm_vcpu *vcpu)
* on the guest's mode (32 or 64 bit), not on the given field's length.
*/
if (instr_info & BIT(10)) {
- kvm_register_write(vcpu, (((instr_info) >> 3) & 0xf), value);
+ kvm_gpr_write(vcpu, (((instr_info) >> 3) & 0xf), value);
} else {
len = is_64_bit_mode(vcpu) ? 8 : 4;
if (get_vmx_mem_address(vcpu, exit_qualification,
@@ -5842,7 +5842,7 @@ static int handle_vmwrite(struct kvm_vcpu *vcpu)
return nested_vmx_failInvalid(vcpu);
if (instr_info & BIT(10))
- value = kvm_register_read(vcpu, (((instr_info) >> 3) & 0xf));
+ value = kvm_gpr_read(vcpu, (((instr_info) >> 3) & 0xf));
else {
len = is_64_bit_mode(vcpu) ? 8 : 4;
if (get_vmx_mem_address(vcpu, exit_qualification,
@@ -5853,7 +5853,7 @@ static int handle_vmwrite(struct kvm_vcpu *vcpu)
return kvm_handle_memory_failure(vcpu, r, &e);
}
- field = kvm_register_read(vcpu, (((instr_info) >> 28) & 0xf));
+ field = kvm_gpr_read(vcpu, (((instr_info) >> 28) & 0xf));
offset = get_vmcs12_field_offset(field);
if (offset < 0)
@@ -6051,7 +6051,7 @@ static int handle_invept(struct kvm_vcpu *vcpu)
vmx_instruction_info = vmcs_read32(VMX_INSTRUCTION_INFO);
gpr_index = vmx_get_instr_info_reg2(vmx_instruction_info);
- type = kvm_register_read(vcpu, gpr_index);
+ type = kvm_gpr_read(vcpu, gpr_index);
types = (vmx->nested.msrs.ept_caps >> VMX_EPT_EXTENT_SHIFT) & 6;
@@ -6132,7 +6132,7 @@ static int handle_invvpid(struct kvm_vcpu *vcpu)
vmx_instruction_info = vmcs_read32(VMX_INSTRUCTION_INFO);
gpr_index = vmx_get_instr_info_reg2(vmx_instruction_info);
- type = kvm_register_read(vcpu, gpr_index);
+ type = kvm_gpr_read(vcpu, gpr_index);
types = (vmx->nested.msrs.vpid_caps &
VMX_VPID_EXTENT_SUPPORTED_MASK) >> 8;
@@ -6406,7 +6406,7 @@ static bool nested_vmx_exit_handled_cr(struct kvm_vcpu *vcpu,
switch ((exit_qualification >> 4) & 3) {
case 0: /* mov to cr */
reg = (exit_qualification >> 8) & 15;
- val = kvm_register_read(vcpu, reg);
+ val = kvm_gpr_read(vcpu, reg);
switch (cr) {
case 0:
if (vmcs12->cr0_guest_host_mask &
@@ -6492,7 +6492,7 @@ static bool nested_vmx_exit_handled_vmcs_access(struct kvm_vcpu *vcpu,
/* Decode instruction info and find the field to access */
vmx_instruction_info = vmcs_read32(VMX_INSTRUCTION_INFO);
- field = kvm_register_read(vcpu, (((vmx_instruction_info) >> 28) & 0xf));
+ field = kvm_gpr_read(vcpu, (((vmx_instruction_info) >> 28) & 0xf));
/* Out-of-range fields always cause a VM exit from L2 to L1 */
if (field >> 15)
diff --git a/arch/x86/kvm/vmx/vmx.c b/arch/x86/kvm/vmx/vmx.c
index f87c216d976d..c7d38f7692cf 100644
--- a/arch/x86/kvm/vmx/vmx.c
+++ b/arch/x86/kvm/vmx/vmx.c
@@ -5462,7 +5462,7 @@ static int handle_cr(struct kvm_vcpu *vcpu)
reg = (exit_qualification >> 8) & 15;
switch ((exit_qualification >> 4) & 3) {
case 0: /* mov to cr */
- val = kvm_register_read(vcpu, reg);
+ val = kvm_gpr_read(vcpu, reg);
trace_kvm_cr_write(cr, val);
switch (cr) {
case 0:
@@ -5504,12 +5504,12 @@ static int handle_cr(struct kvm_vcpu *vcpu)
WARN_ON_ONCE(enable_unrestricted_guest);
val = kvm_read_cr3(vcpu);
- kvm_register_write(vcpu, reg, val);
+ kvm_gpr_write(vcpu, reg, val);
trace_kvm_cr_read(cr, val);
return kvm_skip_emulated_instruction(vcpu);
case 8:
val = kvm_get_cr8(vcpu);
- kvm_register_write(vcpu, reg, val);
+ kvm_gpr_write(vcpu, reg, val);
trace_kvm_cr_read(cr, val);
return kvm_skip_emulated_instruction(vcpu);
}
@@ -5579,10 +5579,10 @@ static int handle_dr(struct kvm_vcpu *vcpu)
reg = DEBUG_REG_ACCESS_REG(exit_qualification);
if (exit_qualification & TYPE_MOV_FROM_DR) {
- kvm_register_write(vcpu, reg, kvm_get_dr(vcpu, dr));
+ kvm_gpr_write(vcpu, reg, kvm_get_dr(vcpu, dr));
err = 0;
} else {
- err = kvm_set_dr(vcpu, dr, kvm_register_read(vcpu, reg));
+ err = kvm_set_dr(vcpu, dr, kvm_gpr_read(vcpu, reg));
}
out:
@@ -5941,7 +5941,7 @@ static int handle_invpcid(struct kvm_vcpu *vcpu)
vmx_instruction_info = vmcs_read32(VMX_INSTRUCTION_INFO);
gpr_index = vmx_get_instr_info_reg2(vmx_instruction_info);
- type = kvm_register_read(vcpu, gpr_index);
+ type = kvm_gpr_read(vcpu, gpr_index);
/* According to the Intel instruction reference, the memory operand
* is read even if it isn't needed (e.g., for type==all)
diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
index b4b5d2d09634..603057ea7421 100644
--- a/arch/x86/kvm/x86.c
+++ b/arch/x86/kvm/x86.c
@@ -2084,8 +2084,8 @@ static int complete_fast_rdmsr(struct kvm_vcpu *vcpu)
static int complete_fast_rdmsr_imm(struct kvm_vcpu *vcpu)
{
if (!vcpu->run->msr.error)
- kvm_register_write(vcpu, vcpu->arch.cui_rdmsr_imm_reg,
- vcpu->run->msr.data);
+ kvm_gpr_write(vcpu, vcpu->arch.cui_rdmsr_imm_reg,
+ vcpu->run->msr.data);
return complete_fast_msr_access(vcpu);
}
@@ -2139,7 +2139,7 @@ static int __kvm_emulate_rdmsr(struct kvm_vcpu *vcpu, u32 msr, int reg,
kvm_rax_write(vcpu, data & -1u);
kvm_rdx_write(vcpu, (data >> 32) & -1u);
} else {
- kvm_register_write(vcpu, reg, data);
+ kvm_gpr_write(vcpu, reg, data);
}
} else {
/* MSR read failed? See if we should ask user space */
@@ -2197,7 +2197,7 @@ EXPORT_SYMBOL_FOR_KVM_INTERNAL(kvm_emulate_wrmsr);
int kvm_emulate_wrmsr_imm(struct kvm_vcpu *vcpu, u32 msr, int reg)
{
- return __kvm_emulate_wrmsr(vcpu, msr, kvm_register_read(vcpu, reg));
+ return __kvm_emulate_wrmsr(vcpu, msr, kvm_gpr_read(vcpu, reg));
}
EXPORT_SYMBOL_FOR_KVM_INTERNAL(kvm_emulate_wrmsr_imm);
@@ -2301,7 +2301,7 @@ EXPORT_SYMBOL_FOR_KVM_INTERNAL(handle_fastpath_wrmsr);
fastpath_t handle_fastpath_wrmsr_imm(struct kvm_vcpu *vcpu, u32 msr, int reg)
{
- return __handle_fastpath_wrmsr(vcpu, msr, kvm_register_read(vcpu, reg));
+ return __handle_fastpath_wrmsr(vcpu, msr, kvm_gpr_read(vcpu, reg));
}
EXPORT_SYMBOL_FOR_KVM_INTERNAL(handle_fastpath_wrmsr_imm);
diff --git a/arch/x86/kvm/x86.h b/arch/x86/kvm/x86.h
index f3dc77f006f9..4edadd64d3d5 100644
--- a/arch/x86/kvm/x86.h
+++ b/arch/x86/kvm/x86.h
@@ -400,15 +400,14 @@ static inline bool vcpu_match_mmio_gpa(struct kvm_vcpu *vcpu, gpa_t gpa)
return false;
}
-static inline unsigned long kvm_register_read(struct kvm_vcpu *vcpu, int reg)
+static inline unsigned long kvm_gpr_read(struct kvm_vcpu *vcpu, int reg)
{
unsigned long val = kvm_register_read_raw(vcpu, reg);
return is_64_bit_mode(vcpu) ? val : (u32)val;
}
-static inline void kvm_register_write(struct kvm_vcpu *vcpu,
- int reg, unsigned long val)
+static inline void kvm_gpr_write(struct kvm_vcpu *vcpu, int reg, unsigned long val)
{
if (!is_64_bit_mode(vcpu))
val = (u32)val;
diff --git a/arch/x86/kvm/xen.c b/arch/x86/kvm/xen.c
index d6b2a665b499..c9700dc88bb1 100644
--- a/arch/x86/kvm/xen.c
+++ b/arch/x86/kvm/xen.c
@@ -1679,7 +1679,7 @@ int kvm_xen_hypercall(struct kvm_vcpu *vcpu)
bool handled = false;
u8 cpl;
- input = (u64)kvm_register_read(vcpu, VCPU_REGS_RAX);
+ input = (u64)kvm_gpr_read(vcpu, VCPU_REGS_RAX);
/* Hyper-V hypercalls get bit 31 set in EAX */
if ((input & 0x80000000) &&
--
2.51.0
^ permalink raw reply related [flat|nested] 55+ messages in thread* [PATCH RFC v1 02/20] KVM: x86: Refactor GPR accessors to differentiate register access types
2025-11-10 18:01 [PATCH RFC v1 00/20] KVM: x86: Support APX feature for guests Chang S. Bae
2025-11-10 18:01 ` [PATCH RFC v1 01/20] KVM: x86: Rename register accessors to be GPR-specific Chang S. Bae
@ 2025-11-10 18:01 ` Chang S. Bae
2025-11-11 18:08 ` Paolo Bonzini
2025-11-11 18:11 ` Paolo Bonzini
2025-11-10 18:01 ` [PATCH RFC v1 03/20] KVM: x86: Implement accessors for extended GPRs Chang S. Bae
` (19 subsequent siblings)
21 siblings, 2 replies; 55+ messages in thread
From: Chang S. Bae @ 2025-11-10 18:01 UTC (permalink / raw)
To: kvm, linux-kernel; +Cc: pbonzini, seanjc, chao.gao, zhao1.liu, chang.seok.bae
Refactor the GPR accessors to introduce internal helpers to distinguish
between legacy and extended registers.
x86 CPUs introduce additional GPRs, but those registers will initially
remain unused in the kernel and will not be saved in KVM register cache
on every VM exit. Guest states are expected to remain live in hardware
registers.
This abstraction layer centralizes the selection of access methods,
providing a unified interface. For now, the EGPR accessors are
placeholders to be implemented later.
Signed-off-by: Chang S. Bae <chang.seok.bae@intel.com>
---
arch/x86/include/asm/kvm_host.h | 18 ++++++++++++
arch/x86/include/asm/kvm_vcpu_regs.h | 16 ++++++++++
arch/x86/kvm/fpu.h | 6 ++++
arch/x86/kvm/x86.h | 44 ++++++++++++++++++++++++++--
4 files changed, 82 insertions(+), 2 deletions(-)
diff --git a/arch/x86/include/asm/kvm_host.h b/arch/x86/include/asm/kvm_host.h
index 48598d017d6f..940f83c121cf 100644
--- a/arch/x86/include/asm/kvm_host.h
+++ b/arch/x86/include/asm/kvm_host.h
@@ -212,6 +212,24 @@ enum {
VCPU_SREG_GS,
VCPU_SREG_TR,
VCPU_SREG_LDTR,
+#ifdef CONFIG_X86_64
+ VCPU_XREG_R16 = __VCPU_XREG_R16,
+ VCPU_XREG_R17 = __VCPU_XREG_R17,
+ VCPU_XREG_R18 = __VCPU_XREG_R18,
+ VCPU_XREG_R19 = __VCPU_XREG_R19,
+ VCPU_XREG_R20 = __VCPU_XREG_R20,
+ VCPU_XREG_R21 = __VCPU_XREG_R21,
+ VCPU_XREG_R22 = __VCPU_XREG_R22,
+ VCPU_XREG_R23 = __VCPU_XREG_R23,
+ VCPU_XREG_R24 = __VCPU_XREG_R24,
+ VCPU_XREG_R25 = __VCPU_XREG_R25,
+ VCPU_XREG_R26 = __VCPU_XREG_R26,
+ VCPU_XREG_R27 = __VCPU_XREG_R27,
+ VCPU_XREG_R28 = __VCPU_XREG_R28,
+ VCPU_XREG_R29 = __VCPU_XREG_R29,
+ VCPU_XREG_R30 = __VCPU_XREG_R30,
+ VCPU_XREG_R31 = __VCPU_XREG_R31,
+#endif
};
enum exit_fastpath_completion {
diff --git a/arch/x86/include/asm/kvm_vcpu_regs.h b/arch/x86/include/asm/kvm_vcpu_regs.h
index 1af2cb59233b..dd0cc171f405 100644
--- a/arch/x86/include/asm/kvm_vcpu_regs.h
+++ b/arch/x86/include/asm/kvm_vcpu_regs.h
@@ -20,6 +20,22 @@
#define __VCPU_REGS_R13 13
#define __VCPU_REGS_R14 14
#define __VCPU_REGS_R15 15
+#define __VCPU_XREG_R16 16
+#define __VCPU_XREG_R17 17
+#define __VCPU_XREG_R18 18
+#define __VCPU_XREG_R19 19
+#define __VCPU_XREG_R20 20
+#define __VCPU_XREG_R21 21
+#define __VCPU_XREG_R22 22
+#define __VCPU_XREG_R23 23
+#define __VCPU_XREG_R24 24
+#define __VCPU_XREG_R25 25
+#define __VCPU_XREG_R26 26
+#define __VCPU_XREG_R27 27
+#define __VCPU_XREG_R28 28
+#define __VCPU_XREG_R29 29
+#define __VCPU_XREG_R30 30
+#define __VCPU_XREG_R31 31
#endif
#endif /* _ASM_X86_KVM_VCPU_REGS_H */
diff --git a/arch/x86/kvm/fpu.h b/arch/x86/kvm/fpu.h
index 3ba12888bf66..159239b3a651 100644
--- a/arch/x86/kvm/fpu.h
+++ b/arch/x86/kvm/fpu.h
@@ -4,6 +4,7 @@
#define __KVM_FPU_H_
#include <asm/fpu/api.h>
+#include <asm/kvm_vcpu_regs.h>
typedef u32 __attribute__((vector_size(16))) sse128_t;
#define __sse128_u union { sse128_t vec; u64 as_u64[2]; u32 as_u32[4]; }
@@ -137,4 +138,9 @@ static inline void kvm_write_mmx_reg(int reg, const u64 *data)
kvm_fpu_put();
}
+#ifdef CONFIG_X86_64
+static inline unsigned long kvm_read_egpr(int reg) { return 0; }
+static inline void kvm_write_egpr(int reg, unsigned long data) { }
+#endif
+
#endif
diff --git a/arch/x86/kvm/x86.h b/arch/x86/kvm/x86.h
index 4edadd64d3d5..74ae8f12b5a1 100644
--- a/arch/x86/kvm/x86.h
+++ b/arch/x86/kvm/x86.h
@@ -400,9 +400,49 @@ static inline bool vcpu_match_mmio_gpa(struct kvm_vcpu *vcpu, gpa_t gpa)
return false;
}
+#ifdef CONFIG_X86_64
+static inline unsigned long _kvm_gpr_read(struct kvm_vcpu *vcpu, int reg)
+{
+ switch (reg) {
+ case VCPU_REGS_RAX ... VCPU_REGS_R15:
+ return kvm_register_read_raw(vcpu, reg);
+ case VCPU_XREG_R16 ... VCPU_XREG_R31:
+ return kvm_read_egpr(reg);
+ default:
+ WARN_ON_ONCE(1);
+ }
+
+ return 0;
+}
+
+static inline void _kvm_gpr_write(struct kvm_vcpu *vcpu, int reg, unsigned long val)
+{
+ switch (reg) {
+ case VCPU_REGS_RAX ... VCPU_REGS_R15:
+ kvm_register_write_raw(vcpu, reg, val);
+ break;
+ case VCPU_XREG_R16 ... VCPU_XREG_R31:
+ kvm_write_egpr(reg, val);
+ break;
+ default:
+ WARN_ON_ONCE(1);
+ }
+}
+#else
+static inline unsigned long _kvm_gpr_read(struct kvm_vcpu *vcpu, int reg)
+{
+ return kvm_register_read_raw(vcpu, reg);
+}
+
+static inline void _kvm_gpr_write(struct kvm_vcpu *vcpu, int reg, unsigned long val)
+{
+ kvm_register_write_raw(vcpu, reg, val);
+}
+#endif
+
static inline unsigned long kvm_gpr_read(struct kvm_vcpu *vcpu, int reg)
{
- unsigned long val = kvm_register_read_raw(vcpu, reg);
+ unsigned long val = _kvm_gpr_read(vcpu, reg);
return is_64_bit_mode(vcpu) ? val : (u32)val;
}
@@ -411,7 +451,7 @@ static inline void kvm_gpr_write(struct kvm_vcpu *vcpu, int reg, unsigned long v
{
if (!is_64_bit_mode(vcpu))
val = (u32)val;
- return kvm_register_write_raw(vcpu, reg, val);
+ _kvm_gpr_write(vcpu, reg, val);
}
static inline bool kvm_check_has_quirk(struct kvm *kvm, u64 quirk)
--
2.51.0
^ permalink raw reply related [flat|nested] 55+ messages in thread* Re: [PATCH RFC v1 02/20] KVM: x86: Refactor GPR accessors to differentiate register access types
2025-11-10 18:01 ` [PATCH RFC v1 02/20] KVM: x86: Refactor GPR accessors to differentiate register access types Chang S. Bae
@ 2025-11-11 18:08 ` Paolo Bonzini
2025-11-13 23:19 ` Chang S. Bae
2025-11-11 18:11 ` Paolo Bonzini
1 sibling, 1 reply; 55+ messages in thread
From: Paolo Bonzini @ 2025-11-11 18:08 UTC (permalink / raw)
To: Chang S. Bae, kvm, linux-kernel; +Cc: seanjc, chao.gao, zhao1.liu
On 11/10/25 19:01, Chang S. Bae wrote:
> Refactor the GPR accessors to introduce internal helpers to distinguish
> between legacy and extended registers.
>
> x86 CPUs introduce additional GPRs, but those registers will initially
> remain unused in the kernel and will not be saved in KVM register cache
> on every VM exit. Guest states are expected to remain live in hardware
> registers.
>
> This abstraction layer centralizes the selection of access methods,
> providing a unified interface. For now, the EGPR accessors are
> placeholders to be implemented later.
>
> Signed-off-by: Chang S. Bae <chang.seok.bae@intel.com>
> ---
> arch/x86/include/asm/kvm_host.h | 18 ++++++++++++
> arch/x86/include/asm/kvm_vcpu_regs.h | 16 ++++++++++
> arch/x86/kvm/fpu.h | 6 ++++
> arch/x86/kvm/x86.h | 44 ++++++++++++++++++++++++++--
> 4 files changed, 82 insertions(+), 2 deletions(-)
>
> diff --git a/arch/x86/include/asm/kvm_host.h b/arch/x86/include/asm/kvm_host.h
> index 48598d017d6f..940f83c121cf 100644
> --- a/arch/x86/include/asm/kvm_host.h
> +++ b/arch/x86/include/asm/kvm_host.h
> @@ -212,6 +212,24 @@ enum {
> VCPU_SREG_GS,
> VCPU_SREG_TR,
> VCPU_SREG_LDTR,
> +#ifdef CONFIG_X86_64
> + VCPU_XREG_R16 = __VCPU_XREG_R16,
> + VCPU_XREG_R17 = __VCPU_XREG_R17,
> + VCPU_XREG_R18 = __VCPU_XREG_R18,
> + VCPU_XREG_R19 = __VCPU_XREG_R19,
> + VCPU_XREG_R20 = __VCPU_XREG_R20,
> + VCPU_XREG_R21 = __VCPU_XREG_R21,
> + VCPU_XREG_R22 = __VCPU_XREG_R22,
> + VCPU_XREG_R23 = __VCPU_XREG_R23,
> + VCPU_XREG_R24 = __VCPU_XREG_R24,
> + VCPU_XREG_R25 = __VCPU_XREG_R25,
> + VCPU_XREG_R26 = __VCPU_XREG_R26,
> + VCPU_XREG_R27 = __VCPU_XREG_R27,
> + VCPU_XREG_R28 = __VCPU_XREG_R28,
> + VCPU_XREG_R29 = __VCPU_XREG_R29,
> + VCPU_XREG_R30 = __VCPU_XREG_R30,
> + VCPU_XREG_R31 = __VCPU_XREG_R31,
> +#endif
> };
>
> enum exit_fastpath_completion {
> diff --git a/arch/x86/include/asm/kvm_vcpu_regs.h b/arch/x86/include/asm/kvm_vcpu_regs.h
> index 1af2cb59233b..dd0cc171f405 100644
> --- a/arch/x86/include/asm/kvm_vcpu_regs.h
> +++ b/arch/x86/include/asm/kvm_vcpu_regs.h
> @@ -20,6 +20,22 @@
> #define __VCPU_REGS_R13 13
> #define __VCPU_REGS_R14 14
> #define __VCPU_REGS_R15 15
> +#define __VCPU_XREG_R16 16
> +#define __VCPU_XREG_R17 17
> +#define __VCPU_XREG_R18 18
> +#define __VCPU_XREG_R19 19
> +#define __VCPU_XREG_R20 20
> +#define __VCPU_XREG_R21 21
> +#define __VCPU_XREG_R22 22
> +#define __VCPU_XREG_R23 23
> +#define __VCPU_XREG_R24 24
> +#define __VCPU_XREG_R25 25
> +#define __VCPU_XREG_R26 26
> +#define __VCPU_XREG_R27 27
> +#define __VCPU_XREG_R28 28
> +#define __VCPU_XREG_R29 29
> +#define __VCPU_XREG_R30 30
> +#define __VCPU_XREG_R31 31
> #endif
>
> #endif /* _ASM_X86_KVM_VCPU_REGS_H */
> diff --git a/arch/x86/kvm/fpu.h b/arch/x86/kvm/fpu.h
> index 3ba12888bf66..159239b3a651 100644
> --- a/arch/x86/kvm/fpu.h
> +++ b/arch/x86/kvm/fpu.h
> @@ -4,6 +4,7 @@
> #define __KVM_FPU_H_
>
> #include <asm/fpu/api.h>
> +#include <asm/kvm_vcpu_regs.h>
>
> typedef u32 __attribute__((vector_size(16))) sse128_t;
> #define __sse128_u union { sse128_t vec; u64 as_u64[2]; u32 as_u32[4]; }
> @@ -137,4 +138,9 @@ static inline void kvm_write_mmx_reg(int reg, const u64 *data)
> kvm_fpu_put();
> }
>
> +#ifdef CONFIG_X86_64
> +static inline unsigned long kvm_read_egpr(int reg) { return 0; }
> +static inline void kvm_write_egpr(int reg, unsigned long data) { }
Do not inline these, they're quite large. Leave them in x86.c.
Also please add a KVM_APX Kconfig symbol and add "select KVM_APX if
X86_64" to KVM_INTEL.
This way, AMD and 32-bit use the same logic to elide all the EGPR code.
Paolo
> +#endif
> +
> #endif
> diff --git a/arch/x86/kvm/x86.h b/arch/x86/kvm/x86.h
> index 4edadd64d3d5..74ae8f12b5a1 100644
> --- a/arch/x86/kvm/x86.h
> +++ b/arch/x86/kvm/x86.h
> @@ -400,9 +400,49 @@ static inline bool vcpu_match_mmio_gpa(struct kvm_vcpu *vcpu, gpa_t gpa)
> return false;
> }
>
> +#ifdef CONFIG_X86_64
> +static inline unsigned long _kvm_gpr_read(struct kvm_vcpu *vcpu, int reg)
> +{
> + switch (reg) {
> + case VCPU_REGS_RAX ... VCPU_REGS_R15:
> + return kvm_register_read_raw(vcpu, reg);
> + case VCPU_XREG_R16 ... VCPU_XREG_R31:
> + return kvm_read_egpr(reg);
> + default:
> + WARN_ON_ONCE(1);
> + }
> +
> + return 0;
> +}
> +
> +static inline void _kvm_gpr_write(struct kvm_vcpu *vcpu, int reg, unsigned long val)
> +{
> + switch (reg) {
> + case VCPU_REGS_RAX ... VCPU_REGS_R15:
> + kvm_register_write_raw(vcpu, reg, val);
> + break;
> + case VCPU_XREG_R16 ... VCPU_XREG_R31:
> + kvm_write_egpr(reg, val);
> + break;
> + default:
> + WARN_ON_ONCE(1);
> + }
> +}
> +#else
> +static inline unsigned long _kvm_gpr_read(struct kvm_vcpu *vcpu, int reg)
> +{
> + return kvm_register_read_raw(vcpu, reg);
> +}
> +
> +static inline void _kvm_gpr_write(struct kvm_vcpu *vcpu, int reg, unsigned long val)
> +{
> + kvm_register_write_raw(vcpu, reg, val);
> +}
> +#endif
> +
> static inline unsigned long kvm_gpr_read(struct kvm_vcpu *vcpu, int reg)
> {
> - unsigned long val = kvm_register_read_raw(vcpu, reg);
> + unsigned long val = _kvm_gpr_read(vcpu, reg);
>
> return is_64_bit_mode(vcpu) ? val : (u32)val;
> }
> @@ -411,7 +451,7 @@ static inline void kvm_gpr_write(struct kvm_vcpu *vcpu, int reg, unsigned long v
> {
> if (!is_64_bit_mode(vcpu))
> val = (u32)val;
> - return kvm_register_write_raw(vcpu, reg, val);
> + _kvm_gpr_write(vcpu, reg, val);
> }
>
> static inline bool kvm_check_has_quirk(struct kvm *kvm, u64 quirk)
^ permalink raw reply [flat|nested] 55+ messages in thread* Re: [PATCH RFC v1 02/20] KVM: x86: Refactor GPR accessors to differentiate register access types
2025-11-11 18:08 ` Paolo Bonzini
@ 2025-11-13 23:19 ` Chang S. Bae
0 siblings, 0 replies; 55+ messages in thread
From: Chang S. Bae @ 2025-11-13 23:19 UTC (permalink / raw)
To: Paolo Bonzini, kvm, linux-kernel; +Cc: seanjc, chao.gao, zhao1.liu
[-- Attachment #1: Type: text/plain, Size: 349 bytes --]
On 11/11/2025 10:08 AM, Paolo Bonzini wrote:
>
> Do not inline these, they're quite large. Leave them in x86.c.
>
> Also please add a KVM_APX Kconfig symbol and add "select KVM_APX if
> X86_64" to KVM_INTEL.
>
> This way, AMD and 32-bit use the same logic to elide all the EGPR code.
Yeah, that makes sense. Here is diff on top of this patch:
[-- Attachment #2: PATCH2.diff --]
[-- Type: text/plain, Size: 3739 bytes --]
diff --git a/arch/x86/kvm/Kconfig b/arch/x86/kvm/Kconfig
index 278f08194ec8..e0fb460b6d37 100644
--- a/arch/x86/kvm/Kconfig
+++ b/arch/x86/kvm/Kconfig
@@ -93,10 +93,14 @@ config KVM_SW_PROTECTED_VM
If unsure, say "N".
+config KVM_APX
+ bool
+
config KVM_INTEL
tristate "KVM for Intel (and compatible) processors support"
depends on KVM && IA32_FEAT_CTL
select X86_FRED if X86_64
+ select KVM_APX if x86_64
help
Provides support for KVM on processors equipped with Intel's VT
extensions, a.k.a. Virtual Machine Extensions (VMX).
diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
index 603057ea7421..5060afc8b4f8 100644
--- a/arch/x86/kvm/x86.c
+++ b/arch/x86/kvm/x86.c
@@ -1259,6 +1259,38 @@ static inline u64 kvm_guest_supported_xfd(struct kvm_vcpu *vcpu)
}
#endif
+#ifdef CONFIG_KVM_APX
+unsigned long kvm_gpr_read_raw(struct kvm_vcpu *vcpu, int reg)
+{
+ switch (reg) {
+ case VCPU_REGS_RAX ... VCPU_REGS_R15:
+ return kvm_register_read_raw(vcpu, reg);
+ case VCPU_XREG_R16 ... VCPU_XREG_R31:
+ return kvm_read_egpr(reg);
+ default:
+ WARN_ON_ONCE(1);
+ }
+
+ return 0;
+}
+EXPORT_SYMBOL_FOR_KVM_INTERNAL(kvm_gpr_read_raw);
+
+void kvm_gpr_write_raw(struct kvm_vcpu *vcpu, int reg, unsigned long val)
+{
+ switch (reg) {
+ case VCPU_REGS_RAX ... VCPU_REGS_R15:
+ kvm_register_write_raw(vcpu, reg, val);
+ break;
+ case VCPU_XREG_R16 ... VCPU_XREG_R31:
+ kvm_write_egpr(reg, val);
+ break;
+ default:
+ WARN_ON_ONCE(1);
+ }
+}
+EXPORT_SYMBOL_FOR_KVM_INTERNAL(kvm_gpr_write_raw);
+#endif
+
int __kvm_set_xcr(struct kvm_vcpu *vcpu, u32 index, u64 xcr)
{
u64 xcr0 = xcr;
diff --git a/arch/x86/kvm/x86.h b/arch/x86/kvm/x86.h
index 74ae8f12b5a1..4e23c915f855 100644
--- a/arch/x86/kvm/x86.h
+++ b/arch/x86/kvm/x86.h
@@ -400,41 +400,16 @@ static inline bool vcpu_match_mmio_gpa(struct kvm_vcpu *vcpu, gpa_t gpa)
return false;
}
-#ifdef CONFIG_X86_64
-static inline unsigned long _kvm_gpr_read(struct kvm_vcpu *vcpu, int reg)
-{
- switch (reg) {
- case VCPU_REGS_RAX ... VCPU_REGS_R15:
- return kvm_register_read_raw(vcpu, reg);
- case VCPU_XREG_R16 ... VCPU_XREG_R31:
- return kvm_read_egpr(reg);
- default:
- WARN_ON_ONCE(1);
- }
-
- return 0;
-}
-
-static inline void _kvm_gpr_write(struct kvm_vcpu *vcpu, int reg, unsigned long val)
-{
- switch (reg) {
- case VCPU_REGS_RAX ... VCPU_REGS_R15:
- kvm_register_write_raw(vcpu, reg, val);
- break;
- case VCPU_XREG_R16 ... VCPU_XREG_R31:
- kvm_write_egpr(reg, val);
- break;
- default:
- WARN_ON_ONCE(1);
- }
-}
+#ifdef CONFIG_KVM_APX
+unsigned long kvm_gpr_read_raw(struct kvm_vcpu *vcpu, int reg);
+void kvm_gpr_write_raw(struct kvm_vcpu *vcpu, int reg, unsigned long val);
#else
-static inline unsigned long _kvm_gpr_read(struct kvm_vcpu *vcpu, int reg)
+static inline unsigned long kvm_gpr_read_raw(struct kvm_vcpu *vcpu, int reg)
{
return kvm_register_read_raw(vcpu, reg);
}
-static inline void _kvm_gpr_write(struct kvm_vcpu *vcpu, int reg, unsigned long val)
+static inline void kvm_gpr_write_raw(struct kvm_vcpu *vcpu, int reg, unsigned long val)
{
kvm_register_write_raw(vcpu, reg, val);
}
@@ -442,7 +417,7 @@ static inline void _kvm_gpr_write(struct kvm_vcpu *vcpu, int reg, unsigned long
static inline unsigned long kvm_gpr_read(struct kvm_vcpu *vcpu, int reg)
{
- unsigned long val = _kvm_gpr_read(vcpu, reg);
+ unsigned long val = kvm_gpr_read_raw(vcpu, reg);
return is_64_bit_mode(vcpu) ? val : (u32)val;
}
@@ -451,7 +426,7 @@ static inline void kvm_gpr_write(struct kvm_vcpu *vcpu, int reg, unsigned long v
{
if (!is_64_bit_mode(vcpu))
val = (u32)val;
- _kvm_gpr_write(vcpu, reg, val);
+ kvm_gpr_write_raw(vcpu, reg, val);
}
static inline bool kvm_check_has_quirk(struct kvm *kvm, u64 quirk)
^ permalink raw reply related [flat|nested] 55+ messages in thread
* Re: [PATCH RFC v1 02/20] KVM: x86: Refactor GPR accessors to differentiate register access types
2025-11-10 18:01 ` [PATCH RFC v1 02/20] KVM: x86: Refactor GPR accessors to differentiate register access types Chang S. Bae
2025-11-11 18:08 ` Paolo Bonzini
@ 2025-11-11 18:11 ` Paolo Bonzini
2025-11-13 23:18 ` Chang S. Bae
1 sibling, 1 reply; 55+ messages in thread
From: Paolo Bonzini @ 2025-11-11 18:11 UTC (permalink / raw)
To: Chang S. Bae, kvm, linux-kernel; +Cc: seanjc, chao.gao, zhao1.liu
On 11/10/25 19:01, Chang S. Bae wrote:
> Refactor the GPR accessors to introduce internal helpers to distinguish
> between legacy and extended registers.
>
> x86 CPUs introduce additional GPRs, but those registers will initially
> remain unused in the kernel and will not be saved in KVM register cache
> on every VM exit. Guest states are expected to remain live in hardware
> registers.
>
> This abstraction layer centralizes the selection of access methods,
> providing a unified interface. For now, the EGPR accessors are
> placeholders to be implemented later.
>
> Signed-off-by: Chang S. Bae <chang.seok.bae@intel.com>
> ---
> arch/x86/include/asm/kvm_host.h | 18 ++++++++++++
> arch/x86/include/asm/kvm_vcpu_regs.h | 16 ++++++++++
> arch/x86/kvm/fpu.h | 6 ++++
> arch/x86/kvm/x86.h | 44 ++++++++++++++++++++++++++--
> 4 files changed, 82 insertions(+), 2 deletions(-)
>
> diff --git a/arch/x86/include/asm/kvm_host.h b/arch/x86/include/asm/kvm_host.h
> index 48598d017d6f..940f83c121cf 100644
> --- a/arch/x86/include/asm/kvm_host.h
> +++ b/arch/x86/include/asm/kvm_host.h
> @@ -212,6 +212,24 @@ enum {
> VCPU_SREG_GS,
> VCPU_SREG_TR,
> VCPU_SREG_LDTR,
> +#ifdef CONFIG_X86_64
> + VCPU_XREG_R16 = __VCPU_XREG_R16,
> + VCPU_XREG_R17 = __VCPU_XREG_R17,
> + VCPU_XREG_R18 = __VCPU_XREG_R18,
> + VCPU_XREG_R19 = __VCPU_XREG_R19,
> + VCPU_XREG_R20 = __VCPU_XREG_R20,
> + VCPU_XREG_R21 = __VCPU_XREG_R21,
> + VCPU_XREG_R22 = __VCPU_XREG_R22,
> + VCPU_XREG_R23 = __VCPU_XREG_R23,
> + VCPU_XREG_R24 = __VCPU_XREG_R24,
> + VCPU_XREG_R25 = __VCPU_XREG_R25,
> + VCPU_XREG_R26 = __VCPU_XREG_R26,
> + VCPU_XREG_R27 = __VCPU_XREG_R27,
> + VCPU_XREG_R28 = __VCPU_XREG_R28,
> + VCPU_XREG_R29 = __VCPU_XREG_R29,
> + VCPU_XREG_R30 = __VCPU_XREG_R30,
> + VCPU_XREG_R31 = __VCPU_XREG_R31,
> +#endif
> };
>
> enum exit_fastpath_completion {
> diff --git a/arch/x86/include/asm/kvm_vcpu_regs.h b/arch/x86/include/asm/kvm_vcpu_regs.h
> index 1af2cb59233b..dd0cc171f405 100644
> --- a/arch/x86/include/asm/kvm_vcpu_regs.h
> +++ b/arch/x86/include/asm/kvm_vcpu_regs.h
> @@ -20,6 +20,22 @@
> #define __VCPU_REGS_R13 13
> #define __VCPU_REGS_R14 14
> #define __VCPU_REGS_R15 15
> +#define __VCPU_XREG_R16 16
> +#define __VCPU_XREG_R17 17
> +#define __VCPU_XREG_R18 18
> +#define __VCPU_XREG_R19 19
> +#define __VCPU_XREG_R20 20
> +#define __VCPU_XREG_R21 21
> +#define __VCPU_XREG_R22 22
> +#define __VCPU_XREG_R23 23
> +#define __VCPU_XREG_R24 24
> +#define __VCPU_XREG_R25 25
> +#define __VCPU_XREG_R26 26
> +#define __VCPU_XREG_R27 27
> +#define __VCPU_XREG_R28 28
> +#define __VCPU_XREG_R29 29
> +#define __VCPU_XREG_R30 30
> +#define __VCPU_XREG_R31 31
> #endif
>
> #endif /* _ASM_X86_KVM_VCPU_REGS_H */
> diff --git a/arch/x86/kvm/fpu.h b/arch/x86/kvm/fpu.h
> index 3ba12888bf66..159239b3a651 100644
> --- a/arch/x86/kvm/fpu.h
> +++ b/arch/x86/kvm/fpu.h
> @@ -4,6 +4,7 @@
> #define __KVM_FPU_H_
>
> #include <asm/fpu/api.h>
> +#include <asm/kvm_vcpu_regs.h>
>
> typedef u32 __attribute__((vector_size(16))) sse128_t;
> #define __sse128_u union { sse128_t vec; u64 as_u64[2]; u32 as_u32[4]; }
> @@ -137,4 +138,9 @@ static inline void kvm_write_mmx_reg(int reg, const u64 *data)
> kvm_fpu_put();
> }
>
> +#ifdef CONFIG_X86_64
> +static inline unsigned long kvm_read_egpr(int reg) { return 0; }
> +static inline void kvm_write_egpr(int reg, unsigned long data) { }
> +#endif
> +
> #endif
> diff --git a/arch/x86/kvm/x86.h b/arch/x86/kvm/x86.h
> index 4edadd64d3d5..74ae8f12b5a1 100644
> --- a/arch/x86/kvm/x86.h
> +++ b/arch/x86/kvm/x86.h
> @@ -400,9 +400,49 @@ static inline bool vcpu_match_mmio_gpa(struct kvm_vcpu *vcpu, gpa_t gpa)
> return false;
> }
>
> +#ifdef CONFIG_X86_64
> +static inline unsigned long _kvm_gpr_read(struct kvm_vcpu *vcpu, int reg)
> +{
> + switch (reg) {
> + case VCPU_REGS_RAX ... VCPU_REGS_R15:
> + return kvm_register_read_raw(vcpu, reg);
> + case VCPU_XREG_R16 ... VCPU_XREG_R31:
> + return kvm_read_egpr(reg);
> + default:
> + WARN_ON_ONCE(1);
> + }
> +
> + return 0;
> +}
> +
> +static inline void _kvm_gpr_write(struct kvm_vcpu *vcpu, int reg, unsigned long val)
> +{
> + switch (reg) {
> + case VCPU_REGS_RAX ... VCPU_REGS_R15:
> + kvm_register_write_raw(vcpu, reg, val);
> + break;
> + case VCPU_XREG_R16 ... VCPU_XREG_R31:
> + kvm_write_egpr(reg, val);
> + break;
> + default:
> + WARN_ON_ONCE(1);
> + }
> +}
> +#else
> +static inline unsigned long _kvm_gpr_read(struct kvm_vcpu *vcpu, int reg)
> +{
> + return kvm_register_read_raw(vcpu, reg);
> +}
Please leave these as kvm_gpr_{read,write}_raw. It's easier to review
than the leading underscore. (It's not hard to undo if you use "git
format-patch --stdout", do a mass replace on the resulting patch file,
and then reapply the patch with "git am").
Paolo
> +static inline void _kvm_gpr_write(struct kvm_vcpu *vcpu, int reg, unsigned long val)
> +{
> + kvm_register_write_raw(vcpu, reg, val);
> +}
> +#endif
> +
> static inline unsigned long kvm_gpr_read(struct kvm_vcpu *vcpu, int reg)
> {
> - unsigned long val = kvm_register_read_raw(vcpu, reg);
> + unsigned long val = _kvm_gpr_read(vcpu, reg);
>
> return is_64_bit_mode(vcpu) ? val : (u32)val;
> }
> @@ -411,7 +451,7 @@ static inline void kvm_gpr_write(struct kvm_vcpu *vcpu, int reg, unsigned long v
> {
> if (!is_64_bit_mode(vcpu))
> val = (u32)val;
> - return kvm_register_write_raw(vcpu, reg, val);
> + _kvm_gpr_write(vcpu, reg, val);
> }
>
> static inline bool kvm_check_has_quirk(struct kvm *kvm, u64 quirk)
^ permalink raw reply [flat|nested] 55+ messages in thread* Re: [PATCH RFC v1 02/20] KVM: x86: Refactor GPR accessors to differentiate register access types
2025-11-11 18:11 ` Paolo Bonzini
@ 2025-11-13 23:18 ` Chang S. Bae
0 siblings, 0 replies; 55+ messages in thread
From: Chang S. Bae @ 2025-11-13 23:18 UTC (permalink / raw)
To: Paolo Bonzini, kvm, linux-kernel; +Cc: seanjc, chao.gao, zhao1.liu
On 11/11/2025 10:11 AM, Paolo Bonzini wrote:
>
> Please leave these as kvm_gpr_{read,write}_raw. It's easier to review
> than the leading underscore. (It's not hard to undo if you use "git
> format-patch --stdout", do a mass replace on the resulting patch file,
> and then reapply the patch with "git am").
Thanks for the tip. Yeah, I saw that naming convention. Fixed.
^ permalink raw reply [flat|nested] 55+ messages in thread
* [PATCH RFC v1 03/20] KVM: x86: Implement accessors for extended GPRs
2025-11-10 18:01 [PATCH RFC v1 00/20] KVM: x86: Support APX feature for guests Chang S. Bae
2025-11-10 18:01 ` [PATCH RFC v1 01/20] KVM: x86: Rename register accessors to be GPR-specific Chang S. Bae
2025-11-10 18:01 ` [PATCH RFC v1 02/20] KVM: x86: Refactor GPR accessors to differentiate register access types Chang S. Bae
@ 2025-11-10 18:01 ` Chang S. Bae
2025-11-10 18:01 ` [PATCH RFC v1 04/20] KVM: VMX: Introduce unified instruction info structure Chang S. Bae
` (18 subsequent siblings)
21 siblings, 0 replies; 55+ messages in thread
From: Chang S. Bae @ 2025-11-10 18:01 UTC (permalink / raw)
To: kvm, linux-kernel; +Cc: pbonzini, seanjc, chao.gao, zhao1.liu, chang.seok.bae
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 Extentions (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@intel.com>
---
RFC note:
There may be alternative options for EGPR access. If the EGPR state is
saved in the guest fpstate, KVM could read or write it there instead.
However, since EGPR-related VM exits are expected to be rare, adding
extra complexity and overhead at this stage doesn’t seem worthwhile.
---
arch/x86/kvm/fpu.h | 80 ++++++++++++++++++++++++++++++++++++++++++++--
1 file changed, 78 insertions(+), 2 deletions(-)
diff --git a/arch/x86/kvm/fpu.h b/arch/x86/kvm/fpu.h
index 159239b3a651..aa35bdf1a073 100644
--- a/arch/x86/kvm/fpu.h
+++ b/arch/x86/kvm/fpu.h
@@ -96,6 +96,61 @@ static inline void _kvm_write_mmx_reg(int reg, const u64 *data)
}
}
+#ifdef CONFIG_X86_64
+/*
+ * Accessors for extended general-purpose registers. binutils >= 2.43 can
+ * recognize those register symbols.
+ */
+
+static inline void _kvm_read_egpr(int reg, unsigned long *data)
+{
+ /* 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;
+ default: BUG();
+ }
+}
+
+static inline void _kvm_write_egpr(int reg, unsigned long *data)
+{
+ /* 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();
+ }
+}
+#endif
+
static inline void kvm_fpu_get(void)
{
fpregs_lock();
@@ -139,8 +194,29 @@ static inline void kvm_write_mmx_reg(int reg, const u64 *data)
}
#ifdef CONFIG_X86_64
-static inline unsigned long kvm_read_egpr(int reg) { return 0; }
-static inline void kvm_write_egpr(int reg, unsigned long data) { }
+static inline unsigned long kvm_read_egpr(int reg)
+{
+ unsigned long data;
+
+ if (WARN_ON_ONCE(!cpu_has_xfeatures(XFEATURE_MASK_APX, NULL)))
+ return 0;
+
+ kvm_fpu_get();
+ _kvm_read_egpr(reg, &data);
+ kvm_fpu_put();
+
+ return data;
+}
+
+static inline void kvm_write_egpr(int reg, unsigned long data)
+{
+ if (WARN_ON_ONCE(!cpu_has_xfeatures(XFEATURE_MASK_APX, NULL)))
+ return;
+
+ kvm_fpu_get();
+ _kvm_write_egpr(reg, &data);
+ kvm_fpu_put();
+}
#endif
#endif
--
2.51.0
^ permalink raw reply related [flat|nested] 55+ messages in thread* [PATCH RFC v1 04/20] KVM: VMX: Introduce unified instruction info structure
2025-11-10 18:01 [PATCH RFC v1 00/20] KVM: x86: Support APX feature for guests Chang S. Bae
` (2 preceding siblings ...)
2025-11-10 18:01 ` [PATCH RFC v1 03/20] KVM: x86: Implement accessors for extended GPRs Chang S. Bae
@ 2025-11-10 18:01 ` Chang S. Bae
2025-11-10 18:01 ` [PATCH RFC v1 05/20] KVM: VMX: Refactor instruction information retrieval Chang S. Bae
` (17 subsequent siblings)
21 siblings, 0 replies; 55+ messages in thread
From: Chang S. Bae @ 2025-11-10 18:01 UTC (permalink / raw)
To: kvm, linux-kernel; +Cc: pbonzini, seanjc, chao.gao, zhao1.liu, chang.seok.bae
Define a unified data structure that can represent both the legacy and
extended VMX instruction information formats.
VMX provides per-instruction metadata for VM exits to help decode the
attributes of the instruction that triggered the exit. The legacy format,
however, only supports up to 16 GPRs and thus cannot represent EGPRs. To
support these new registers, VMX introduces an extended 64-bit layout.
Instead of maintaining separate storage for each format, a single
union structure makes the overall handling simple. The field names are
consistent across both layouts. While the presence of certain fields
depends on the instruction type, the offsets remain fixed within each
format.
Signed-off-by: Chang S. Bae <chang.seok.bae@intel.com>
---
arch/x86/kvm/vmx/vmx.h | 61 ++++++++++++++++++++++++++++++++++++++++++
1 file changed, 61 insertions(+)
diff --git a/arch/x86/kvm/vmx/vmx.h b/arch/x86/kvm/vmx/vmx.h
index ea93121029f9..c358aca7253c 100644
--- a/arch/x86/kvm/vmx/vmx.h
+++ b/arch/x86/kvm/vmx/vmx.h
@@ -311,6 +311,67 @@ struct kvm_vmx {
u64 *pid_table;
};
+/*
+ * 32-bit layout of the legacy instruction information field. This format
+ * supports the 16 legacy GPRs.
+ */
+struct base_insn_info {
+ u32 scale : 2; /* Scaling factor */
+ u32 reserved1 : 1;
+ u32 reg1 : 4; /* First register index */
+ u32 asize : 3; /* Address size */
+ u32 is_reg : 1; /* 0: memory, 1: register */
+ u32 osize : 2; /* Operand size */
+ u32 reserved2 : 2;
+ u32 seg : 3; /* Segment register index */
+ u32 index : 4; /* Index register index */
+ u32 index_invalid : 1; /* 0: valid, 1: invalid */
+ u32 base : 4; /* Base register index */
+ u32 base_invalid : 1; /* 0: valid, 1: invalid */
+ u32 reg2 : 4; /* Second register index */
+};
+
+/*
+ * 64-bit layout of the extended instruction information field, which
+ * supports EGPRs.
+ */
+struct ext_insn_info {
+ u64 scale : 2; /* Scaling factor */
+ u64 asize : 2; /* Address size */
+ u64 is_reg : 1; /* 0: memory, 1: register */
+ u64 osize : 2; /* Operand size */
+ u64 seg : 3; /* Segment register index */
+ u64 index_invalid : 1; /* 0: valid, 1: invalid */
+ u64 base_invalid : 1; /* 0: valid, 1: invalid */
+ u64 reserved1 : 4;
+ u64 reg1 : 5; /* First register index */
+ u64 reserved2 : 3;
+ u64 index : 5; /* Index register index */
+ u64 reserved3 : 3;
+ u64 base : 5; /* Base register index */
+ u64 reserved4 : 3;
+ u64 reg2 : 5; /* Second register index */
+ u64 reserved5 : 19;
+};
+
+/* Union for accessing either the legacy or extended format. */
+union insn_info {
+ struct base_insn_info base;
+ struct ext_insn_info ext;
+ u32 word;
+ u64 dword;
+};
+
+/*
+ * Wrapper structure combining the instruction info and a flag indicating
+ * whether the extended layout is in use.
+ */
+struct vmx_insn_info {
+ /* true if using the extended layout */
+ bool extended;
+ union insn_info info;
+};
+
static __always_inline struct vcpu_vt *to_vt(struct kvm_vcpu *vcpu)
{
return &(container_of(vcpu, struct vcpu_vmx, vcpu)->vt);
--
2.51.0
^ permalink raw reply related [flat|nested] 55+ messages in thread* [PATCH RFC v1 05/20] KVM: VMX: Refactor instruction information retrieval
2025-11-10 18:01 [PATCH RFC v1 00/20] KVM: x86: Support APX feature for guests Chang S. Bae
` (3 preceding siblings ...)
2025-11-10 18:01 ` [PATCH RFC v1 04/20] KVM: VMX: Introduce unified instruction info structure Chang S. Bae
@ 2025-11-10 18:01 ` Chang S. Bae
2025-11-10 18:01 ` [PATCH RFC v1 06/20] KVM: VMX: Refactor GPR index retrieval from exit qualification Chang S. Bae
` (16 subsequent siblings)
21 siblings, 0 replies; 55+ messages in thread
From: Chang S. Bae @ 2025-11-10 18:01 UTC (permalink / raw)
To: kvm, linux-kernel; +Cc: pbonzini, seanjc, chao.gao, zhao1.liu, chang.seok.bae
Introduce helpers to convert and extract exited instruction attributes,
preparing for EGPR support and deprecating some existing helpers.
Previously, VMX exit handlers directly decoded the raw VMCS field,
resulting in duplicated logic and assumption tied to the legacy layout.
With the unified structure, handlers can convert raw data into a
structure form and access each instruction attribute by field name.
The helper will later determine the format based on the VCPU
configuration. For now, there is no functional change since only the
legacy layout is used.
Signed-off-by: Chang S. Bae <chang.seok.bae@intel.com>
---
RFC note:
Macro and variable naming may still evolve depending on
maintainer/reviewer preferences.
---
arch/x86/kvm/vmx/nested.c | 73 +++++++++++++++++++--------------------
arch/x86/kvm/vmx/nested.h | 2 +-
arch/x86/kvm/vmx/vmx.c | 14 ++++----
arch/x86/kvm/vmx/vmx.h | 23 ++++++------
4 files changed, 57 insertions(+), 55 deletions(-)
diff --git a/arch/x86/kvm/vmx/nested.c b/arch/x86/kvm/vmx/nested.c
index 47a941989787..4b883ded6c4b 100644
--- a/arch/x86/kvm/vmx/nested.c
+++ b/arch/x86/kvm/vmx/nested.c
@@ -5289,7 +5289,7 @@ static void nested_vmx_triple_fault(struct kvm_vcpu *vcpu)
* #UD, #GP, or #SS.
*/
int get_vmx_mem_address(struct kvm_vcpu *vcpu, unsigned long exit_qualification,
- u32 vmx_instruction_info, bool wr, int len, gva_t *ret)
+ struct vmx_insn_info info, bool wr, int len, gva_t *ret)
{
gva_t off;
bool exn;
@@ -5303,14 +5303,14 @@ int get_vmx_mem_address(struct kvm_vcpu *vcpu, unsigned long exit_qualification,
* For how an actual address is calculated from all these components,
* refer to Vol. 1, "Operand Addressing".
*/
- int scaling = vmx_instruction_info & 3;
- int addr_size = (vmx_instruction_info >> 7) & 7;
- bool is_reg = vmx_instruction_info & (1u << 10);
- int seg_reg = (vmx_instruction_info >> 15) & 7;
- int index_reg = (vmx_instruction_info >> 18) & 0xf;
- bool index_is_valid = !(vmx_instruction_info & (1u << 22));
- int base_reg = (vmx_instruction_info >> 23) & 0xf;
- bool base_is_valid = !(vmx_instruction_info & (1u << 27));
+ int scaling = insn_attr(info, scale);
+ int addr_size = insn_attr(info, asize);
+ bool is_reg = insn_attr(info, is_reg);
+ int seg_reg = insn_attr(info, seg);
+ int index_reg = insn_attr(info, index);
+ bool index_is_valid = !insn_attr(info, index_invalid);
+ int base_reg = insn_attr(info, base);
+ bool base_is_valid = !insn_attr(info, base_invalid);
if (is_reg) {
kvm_queue_exception(vcpu, UD_VECTOR);
@@ -5421,7 +5421,7 @@ static int nested_vmx_get_vmptr(struct kvm_vcpu *vcpu, gpa_t *vmpointer,
int r;
if (get_vmx_mem_address(vcpu, vmx_get_exit_qual(vcpu),
- vmcs_read32(VMX_INSTRUCTION_INFO), false,
+ vmx_get_insn_info(vcpu), false,
sizeof(*vmpointer), &gva)) {
*ret = 1;
return -EINVAL;
@@ -5706,7 +5706,7 @@ static int handle_vmread(struct kvm_vcpu *vcpu)
struct vmcs12 *vmcs12 = is_guest_mode(vcpu) ? get_shadow_vmcs12(vcpu)
: get_vmcs12(vcpu);
unsigned long exit_qualification = vmx_get_exit_qual(vcpu);
- u32 instr_info = vmcs_read32(VMX_INSTRUCTION_INFO);
+ struct vmx_insn_info info = vmx_get_insn_info(vcpu);
struct vcpu_vmx *vmx = to_vmx(vcpu);
struct x86_exception e;
unsigned long field;
@@ -5719,7 +5719,7 @@ static int handle_vmread(struct kvm_vcpu *vcpu)
return 1;
/* Decode instruction info and find the field to read */
- field = kvm_gpr_read(vcpu, (((instr_info) >> 28) & 0xf));
+ field = kvm_gpr_read(vcpu, insn_attr(info, reg2));
if (!nested_vmx_is_evmptr12_valid(vmx)) {
/*
@@ -5767,12 +5767,12 @@ static int handle_vmread(struct kvm_vcpu *vcpu)
* Note that the number of bits actually copied is 32 or 64 depending
* on the guest's mode (32 or 64 bit), not on the given field's length.
*/
- if (instr_info & BIT(10)) {
- kvm_gpr_write(vcpu, (((instr_info) >> 3) & 0xf), value);
+ if (insn_attr(info, is_reg)) {
+ kvm_gpr_write(vcpu, insn_attr(info, reg1), value);
} else {
len = is_64_bit_mode(vcpu) ? 8 : 4;
if (get_vmx_mem_address(vcpu, exit_qualification,
- instr_info, true, len, &gva))
+ info, true, len, &gva))
return 1;
/* _system ok, nested_vmx_check_permission has verified cpl=0 */
r = kvm_write_guest_virt_system(vcpu, gva, &value, len, &e);
@@ -5812,7 +5812,7 @@ static int handle_vmwrite(struct kvm_vcpu *vcpu)
struct vmcs12 *vmcs12 = is_guest_mode(vcpu) ? get_shadow_vmcs12(vcpu)
: get_vmcs12(vcpu);
unsigned long exit_qualification = vmx_get_exit_qual(vcpu);
- u32 instr_info = vmcs_read32(VMX_INSTRUCTION_INFO);
+ struct vmx_insn_info info = vmx_get_insn_info(vcpu);
struct vcpu_vmx *vmx = to_vmx(vcpu);
struct x86_exception e;
unsigned long field;
@@ -5841,19 +5841,19 @@ static int handle_vmwrite(struct kvm_vcpu *vcpu)
get_vmcs12(vcpu)->vmcs_link_pointer == INVALID_GPA))
return nested_vmx_failInvalid(vcpu);
- if (instr_info & BIT(10))
- value = kvm_gpr_read(vcpu, (((instr_info) >> 3) & 0xf));
+ if (insn_attr(info, is_reg))
+ value = kvm_gpr_read(vcpu, insn_attr(info, reg1));
else {
len = is_64_bit_mode(vcpu) ? 8 : 4;
if (get_vmx_mem_address(vcpu, exit_qualification,
- instr_info, false, len, &gva))
+ info, false, len, &gva))
return 1;
r = kvm_read_guest_virt(vcpu, gva, &value, len, &e);
if (r != X86EMUL_CONTINUE)
return kvm_handle_memory_failure(vcpu, r, &e);
}
- field = kvm_gpr_read(vcpu, (((instr_info) >> 28) & 0xf));
+ field = kvm_gpr_read(vcpu, insn_attr(info, reg2));
offset = get_vmcs12_field_offset(field);
if (offset < 0)
@@ -6001,7 +6001,7 @@ static int handle_vmptrld(struct kvm_vcpu *vcpu)
static int handle_vmptrst(struct kvm_vcpu *vcpu)
{
unsigned long exit_qual = vmx_get_exit_qual(vcpu);
- u32 instr_info = vmcs_read32(VMX_INSTRUCTION_INFO);
+ struct vmx_insn_info info = vmx_get_insn_info(vcpu);
gpa_t current_vmptr = to_vmx(vcpu)->nested.current_vmptr;
struct x86_exception e;
gva_t gva;
@@ -6013,7 +6013,7 @@ static int handle_vmptrst(struct kvm_vcpu *vcpu)
if (unlikely(nested_vmx_is_evmptr12_valid(to_vmx(vcpu))))
return 1;
- if (get_vmx_mem_address(vcpu, exit_qual, instr_info,
+ if (get_vmx_mem_address(vcpu, exit_qual, info,
true, sizeof(gpa_t), &gva))
return 1;
/* *_system ok, nested_vmx_check_permission has verified cpl=0 */
@@ -6029,15 +6029,16 @@ static int handle_vmptrst(struct kvm_vcpu *vcpu)
static int handle_invept(struct kvm_vcpu *vcpu)
{
struct vcpu_vmx *vmx = to_vmx(vcpu);
- u32 vmx_instruction_info, types;
unsigned long type, roots_to_free;
+ struct vmx_insn_info info;
struct kvm_mmu *mmu;
gva_t gva;
struct x86_exception e;
struct {
u64 eptp, gpa;
} operand;
- int i, r, gpr_index;
+ u32 types;
+ int i, r;
if (!(vmx->nested.msrs.secondary_ctls_high &
SECONDARY_EXEC_ENABLE_EPT) ||
@@ -6049,9 +6050,8 @@ static int handle_invept(struct kvm_vcpu *vcpu)
if (!nested_vmx_check_permission(vcpu))
return 1;
- vmx_instruction_info = vmcs_read32(VMX_INSTRUCTION_INFO);
- gpr_index = vmx_get_instr_info_reg2(vmx_instruction_info);
- type = kvm_gpr_read(vcpu, gpr_index);
+ info = vmx_get_insn_info(vcpu);
+ type = kvm_gpr_read(vcpu, insn_attr(info, reg2));
types = (vmx->nested.msrs.ept_caps >> VMX_EPT_EXTENT_SHIFT) & 6;
@@ -6062,7 +6062,7 @@ static int handle_invept(struct kvm_vcpu *vcpu)
* operand is read even if it isn't needed (e.g., for type==global)
*/
if (get_vmx_mem_address(vcpu, vmx_get_exit_qual(vcpu),
- vmx_instruction_info, false, sizeof(operand), &gva))
+ info, false, sizeof(operand), &gva))
return 1;
r = kvm_read_guest_virt(vcpu, gva, &operand, sizeof(operand), &e);
if (r != X86EMUL_CONTINUE)
@@ -6109,7 +6109,7 @@ static int handle_invept(struct kvm_vcpu *vcpu)
static int handle_invvpid(struct kvm_vcpu *vcpu)
{
struct vcpu_vmx *vmx = to_vmx(vcpu);
- u32 vmx_instruction_info;
+ struct vmx_insn_info info;
unsigned long type, types;
gva_t gva;
struct x86_exception e;
@@ -6118,7 +6118,7 @@ static int handle_invvpid(struct kvm_vcpu *vcpu)
u64 gla;
} operand;
u16 vpid02;
- int r, gpr_index;
+ int r;
if (!(vmx->nested.msrs.secondary_ctls_high &
SECONDARY_EXEC_ENABLE_VPID) ||
@@ -6130,9 +6130,8 @@ static int handle_invvpid(struct kvm_vcpu *vcpu)
if (!nested_vmx_check_permission(vcpu))
return 1;
- vmx_instruction_info = vmcs_read32(VMX_INSTRUCTION_INFO);
- gpr_index = vmx_get_instr_info_reg2(vmx_instruction_info);
- type = kvm_gpr_read(vcpu, gpr_index);
+ info = vmx_get_insn_info(vcpu);
+ type = kvm_gpr_read(vcpu, insn_attr(info, reg2));
types = (vmx->nested.msrs.vpid_caps &
VMX_VPID_EXTENT_SUPPORTED_MASK) >> 8;
@@ -6145,7 +6144,7 @@ static int handle_invvpid(struct kvm_vcpu *vcpu)
* operand is read even if it isn't needed (e.g., for type==global)
*/
if (get_vmx_mem_address(vcpu, vmx_get_exit_qual(vcpu),
- vmx_instruction_info, false, sizeof(operand), &gva))
+ info, false, sizeof(operand), &gva))
return 1;
r = kvm_read_guest_virt(vcpu, gva, &operand, sizeof(operand), &e);
if (r != X86EMUL_CONTINUE)
@@ -6483,7 +6482,7 @@ static bool nested_vmx_exit_handled_encls(struct kvm_vcpu *vcpu,
static bool nested_vmx_exit_handled_vmcs_access(struct kvm_vcpu *vcpu,
struct vmcs12 *vmcs12, gpa_t bitmap)
{
- u32 vmx_instruction_info;
+ struct vmx_insn_info info;
unsigned long field;
u8 b;
@@ -6491,8 +6490,8 @@ static bool nested_vmx_exit_handled_vmcs_access(struct kvm_vcpu *vcpu,
return true;
/* Decode instruction info and find the field to access */
- vmx_instruction_info = vmcs_read32(VMX_INSTRUCTION_INFO);
- field = kvm_gpr_read(vcpu, (((vmx_instruction_info) >> 28) & 0xf));
+ info = vmx_get_insn_info(vcpu);
+ field = kvm_gpr_read(vcpu, insn_attr(info, reg2));
/* Out-of-range fields always cause a VM exit from L2 to L1 */
if (field >> 15)
diff --git a/arch/x86/kvm/vmx/nested.h b/arch/x86/kvm/vmx/nested.h
index 983484d42ebf..e54f4e7b3664 100644
--- a/arch/x86/kvm/vmx/nested.h
+++ b/arch/x86/kvm/vmx/nested.h
@@ -50,7 +50,7 @@ void nested_sync_vmcs12_to_shadow(struct kvm_vcpu *vcpu);
int vmx_set_vmx_msr(struct kvm_vcpu *vcpu, u32 msr_index, u64 data);
int vmx_get_vmx_msr(struct nested_vmx_msrs *msrs, u32 msr_index, u64 *pdata);
int get_vmx_mem_address(struct kvm_vcpu *vcpu, unsigned long exit_qualification,
- u32 vmx_instruction_info, bool wr, int len, gva_t *ret);
+ struct vmx_insn_info info, bool wr, int len, gva_t *ret);
void nested_mark_vmcs12_pages_dirty(struct kvm_vcpu *vcpu);
bool nested_vmx_check_io_bitmaps(struct kvm_vcpu *vcpu, unsigned int port,
int size);
diff --git a/arch/x86/kvm/vmx/vmx.c b/arch/x86/kvm/vmx/vmx.c
index c7d38f7692cf..dd8c9517c38c 100644
--- a/arch/x86/kvm/vmx/vmx.c
+++ b/arch/x86/kvm/vmx/vmx.c
@@ -5925,29 +5925,27 @@ static int handle_monitor_trap(struct kvm_vcpu *vcpu)
static int handle_invpcid(struct kvm_vcpu *vcpu)
{
- u32 vmx_instruction_info;
+ struct vmx_insn_info info;
unsigned long type;
gva_t gva;
struct {
u64 pcid;
u64 gla;
} operand;
- int gpr_index;
if (!guest_cpu_cap_has(vcpu, X86_FEATURE_INVPCID)) {
kvm_queue_exception(vcpu, UD_VECTOR);
return 1;
}
- vmx_instruction_info = vmcs_read32(VMX_INSTRUCTION_INFO);
- gpr_index = vmx_get_instr_info_reg2(vmx_instruction_info);
- type = kvm_gpr_read(vcpu, gpr_index);
+ info = vmx_get_insn_info(vcpu);
+ type = kvm_gpr_read(vcpu, insn_attr(info, reg2));
/* According to the Intel instruction reference, the memory operand
* is read even if it isn't needed (e.g., for type==all)
*/
if (get_vmx_mem_address(vcpu, vmx_get_exit_qual(vcpu),
- vmx_instruction_info, false,
+ info, false,
sizeof(operand), &gva))
return 1;
@@ -6084,7 +6082,9 @@ static int handle_notify(struct kvm_vcpu *vcpu)
static int vmx_get_msr_imm_reg(struct kvm_vcpu *vcpu)
{
- return vmx_get_instr_info_reg(vmcs_read32(VMX_INSTRUCTION_INFO));
+ struct vmx_insn_info info = vmx_get_insn_info(vcpu);
+
+ return insn_attr(info, reg1);
}
static int handle_rdmsr_imm(struct kvm_vcpu *vcpu)
diff --git a/arch/x86/kvm/vmx/vmx.h b/arch/x86/kvm/vmx/vmx.h
index c358aca7253c..a58d9187ed1d 100644
--- a/arch/x86/kvm/vmx/vmx.h
+++ b/arch/x86/kvm/vmx/vmx.h
@@ -372,6 +372,19 @@ struct vmx_insn_info {
union insn_info info;
};
+static inline struct vmx_insn_info vmx_get_insn_info(struct kvm_vcpu *vcpu __maybe_unused)
+{
+ struct vmx_insn_info insn;
+
+ insn.extended = false;
+ insn.info.word = vmcs_read32(VMX_INSTRUCTION_INFO);
+
+ return insn;
+}
+
+#define insn_attr(insn, attr) \
+ ((insn).extended ? (insn).info.ext.attr : (insn).info.base.attr)
+
static __always_inline struct vcpu_vt *to_vt(struct kvm_vcpu *vcpu)
{
return &(container_of(vcpu, struct vcpu_vmx, vcpu)->vt);
@@ -780,16 +793,6 @@ static inline bool vmx_guest_state_valid(struct kvm_vcpu *vcpu)
void dump_vmcs(struct kvm_vcpu *vcpu);
-static inline int vmx_get_instr_info_reg(u32 vmx_instr_info)
-{
- return (vmx_instr_info >> 3) & 0xf;
-}
-
-static inline int vmx_get_instr_info_reg2(u32 vmx_instr_info)
-{
- return (vmx_instr_info >> 28) & 0xf;
-}
-
static inline bool vmx_can_use_ipiv(struct kvm_vcpu *vcpu)
{
return lapic_in_kernel(vcpu) && enable_ipiv;
--
2.51.0
^ permalink raw reply related [flat|nested] 55+ messages in thread* [PATCH RFC v1 06/20] KVM: VMX: Refactor GPR index retrieval from exit qualification
2025-11-10 18:01 [PATCH RFC v1 00/20] KVM: x86: Support APX feature for guests Chang S. Bae
` (4 preceding siblings ...)
2025-11-10 18:01 ` [PATCH RFC v1 05/20] KVM: VMX: Refactor instruction information retrieval Chang S. Bae
@ 2025-11-10 18:01 ` Chang S. Bae
2025-11-10 18:01 ` [PATCH RFC v1 07/20] KVM: nVMX: Support the extended instruction info field Chang S. Bae
` (15 subsequent siblings)
21 siblings, 0 replies; 55+ messages in thread
From: Chang S. Bae @ 2025-11-10 18:01 UTC (permalink / raw)
To: kvm, linux-kernel; +Cc: pbonzini, seanjc, chao.gao, zhao1.liu, chang.seok.bae
Introduce a helper to extract the GPR index from the exit qualification
field.
Some VMX exit qualification, in addition to the VMX instruction info
field, encode a GPR index. With the introduction of EGPRs, this field is
extended by a previously reserved bit position.
This refactoring centralizes the logic so that future updates can handle
the extended GPR index without code duplication.
Since the VMCS exit qualification is cached in VCPU state, it is safe
for the helper to access it directly via the VCPU pointer. This argument
will also be used later to determine EGPR availability.
No functional change intended.
Signed-off-by: Chang S. Bae <chang.seok.bae@intel.com>
---
arch/x86/kvm/vmx/nested.c | 2 +-
arch/x86/kvm/vmx/vmx.c | 2 +-
arch/x86/kvm/vmx/vmx.h | 5 +++++
3 files changed, 7 insertions(+), 2 deletions(-)
diff --git a/arch/x86/kvm/vmx/nested.c b/arch/x86/kvm/vmx/nested.c
index 4b883ded6c4b..97ec8e594155 100644
--- a/arch/x86/kvm/vmx/nested.c
+++ b/arch/x86/kvm/vmx/nested.c
@@ -6404,7 +6404,7 @@ static bool nested_vmx_exit_handled_cr(struct kvm_vcpu *vcpu,
switch ((exit_qualification >> 4) & 3) {
case 0: /* mov to cr */
- reg = (exit_qualification >> 8) & 15;
+ reg = vmx_get_exit_qual_gpr(vcpu);
val = kvm_gpr_read(vcpu, reg);
switch (cr) {
case 0:
diff --git a/arch/x86/kvm/vmx/vmx.c b/arch/x86/kvm/vmx/vmx.c
index dd8c9517c38c..4405724cb874 100644
--- a/arch/x86/kvm/vmx/vmx.c
+++ b/arch/x86/kvm/vmx/vmx.c
@@ -5459,7 +5459,7 @@ static int handle_cr(struct kvm_vcpu *vcpu)
exit_qualification = vmx_get_exit_qual(vcpu);
cr = exit_qualification & 15;
- reg = (exit_qualification >> 8) & 15;
+ reg = vmx_get_exit_qual_gpr(vcpu);
switch ((exit_qualification >> 4) & 3) {
case 0: /* mov to cr */
val = kvm_gpr_read(vcpu, reg);
diff --git a/arch/x86/kvm/vmx/vmx.h b/arch/x86/kvm/vmx/vmx.h
index a58d9187ed1d..64a0772c883c 100644
--- a/arch/x86/kvm/vmx/vmx.h
+++ b/arch/x86/kvm/vmx/vmx.h
@@ -411,6 +411,11 @@ static __always_inline unsigned long vmx_get_exit_qual(struct kvm_vcpu *vcpu)
return vt->exit_qualification;
}
+static inline int vmx_get_exit_qual_gpr(struct kvm_vcpu *vcpu)
+{
+ return (vmx_get_exit_qual(vcpu) >> 8) & 0xf;
+}
+
static __always_inline u32 vmx_get_intr_info(struct kvm_vcpu *vcpu)
{
struct vcpu_vt *vt = to_vt(vcpu);
--
2.51.0
^ permalink raw reply related [flat|nested] 55+ messages in thread* [PATCH RFC v1 07/20] KVM: nVMX: Support the extended instruction info field
2025-11-10 18:01 [PATCH RFC v1 00/20] KVM: x86: Support APX feature for guests Chang S. Bae
` (5 preceding siblings ...)
2025-11-10 18:01 ` [PATCH RFC v1 06/20] KVM: VMX: Refactor GPR index retrieval from exit qualification Chang S. Bae
@ 2025-11-10 18:01 ` Chang S. Bae
2025-11-11 17:48 ` Paolo Bonzini
2025-11-10 18:01 ` [PATCH RFC v1 08/20] KVM: VMX: Support extended register index in exit handling Chang S. Bae
` (14 subsequent siblings)
21 siblings, 1 reply; 55+ messages in thread
From: Chang S. Bae @ 2025-11-10 18:01 UTC (permalink / raw)
To: kvm, linux-kernel; +Cc: pbonzini, seanjc, chao.gao, zhao1.liu, chang.seok.bae
Define the VMCS field offset for the extended instruction information and
handle it for nested VMX.
When EGPRs are available, VMX provides a new 64-bit field to extend the
legacy instruction information, allowing access to the higher register
indices. Then, nested VMX needs to propagate this field between L1 and
L2.
The EGPR checker will be implemented later.
Signed-off-by: Chang S. Bae <chang.seok.bae@intel.com>
---
RFC note:
During the draft, I brought up the offset definition initially for
non-nested VMX primarily. Then, I realized the switching helper affects
nVMX code anyway. Due to this dependency, this change is placed first
together with the offset definition.
---
arch/x86/include/asm/vmx.h | 2 ++
arch/x86/kvm/vmx/nested.c | 2 ++
arch/x86/kvm/vmx/vmcs12.c | 1 +
arch/x86/kvm/vmx/vmcs12.h | 3 ++-
arch/x86/kvm/vmx/vmx.h | 2 ++
5 files changed, 9 insertions(+), 1 deletion(-)
diff --git a/arch/x86/include/asm/vmx.h b/arch/x86/include/asm/vmx.h
index c85c50019523..ab0684948c56 100644
--- a/arch/x86/include/asm/vmx.h
+++ b/arch/x86/include/asm/vmx.h
@@ -264,6 +264,8 @@ enum vmcs_field {
PID_POINTER_TABLE_HIGH = 0x00002043,
GUEST_PHYSICAL_ADDRESS = 0x00002400,
GUEST_PHYSICAL_ADDRESS_HIGH = 0x00002401,
+ EXTENDED_INSTRUCTION_INFO = 0x00002406,
+ EXTENDED_INSTRUCTION_INFO_HIGH = 0x00002407,
VMCS_LINK_POINTER = 0x00002800,
VMCS_LINK_POINTER_HIGH = 0x00002801,
GUEST_IA32_DEBUGCTL = 0x00002802,
diff --git a/arch/x86/kvm/vmx/nested.c b/arch/x86/kvm/vmx/nested.c
index 97ec8e594155..3442610a6b70 100644
--- a/arch/x86/kvm/vmx/nested.c
+++ b/arch/x86/kvm/vmx/nested.c
@@ -4798,6 +4798,8 @@ static void prepare_vmcs12(struct kvm_vcpu *vcpu, struct vmcs12 *vmcs12,
vmcs12->vm_exit_intr_info = exit_intr_info;
vmcs12->vm_exit_instruction_len = exit_insn_len;
vmcs12->vmx_instruction_info = vmcs_read32(VMX_INSTRUCTION_INFO);
+ if (vmx_egpr_enabled(vcpu))
+ vmcs12->extended_instruction_info = vmcs_read64(EXTENDED_INSTRUCTION_INFO);
/*
* According to spec, there's no need to store the guest's
diff --git a/arch/x86/kvm/vmx/vmcs12.c b/arch/x86/kvm/vmx/vmcs12.c
index 4233b5ca9461..ea2b690a419e 100644
--- a/arch/x86/kvm/vmx/vmcs12.c
+++ b/arch/x86/kvm/vmx/vmcs12.c
@@ -53,6 +53,7 @@ const unsigned short vmcs12_field_offsets[] = {
FIELD64(XSS_EXIT_BITMAP, xss_exit_bitmap),
FIELD64(ENCLS_EXITING_BITMAP, encls_exiting_bitmap),
FIELD64(GUEST_PHYSICAL_ADDRESS, guest_physical_address),
+ FIELD64(EXTENDED_INSTRUCTION_INFO, extended_instruction_info),
FIELD64(VMCS_LINK_POINTER, vmcs_link_pointer),
FIELD64(GUEST_IA32_DEBUGCTL, guest_ia32_debugctl),
FIELD64(GUEST_IA32_PAT, guest_ia32_pat),
diff --git a/arch/x86/kvm/vmx/vmcs12.h b/arch/x86/kvm/vmx/vmcs12.h
index 4ad6b16525b9..2146e45aaade 100644
--- a/arch/x86/kvm/vmx/vmcs12.h
+++ b/arch/x86/kvm/vmx/vmcs12.h
@@ -71,7 +71,7 @@ struct __packed vmcs12 {
u64 pml_address;
u64 encls_exiting_bitmap;
u64 tsc_multiplier;
- u64 padding64[1]; /* room for future expansion */
+ u64 extended_instruction_info;
/*
* To allow migration of L1 (complete with its L2 guests) between
* machines of different natural widths (32 or 64 bit), we cannot have
@@ -261,6 +261,7 @@ static inline void vmx_check_vmcs12_offsets(void)
CHECK_OFFSET(pml_address, 312);
CHECK_OFFSET(encls_exiting_bitmap, 320);
CHECK_OFFSET(tsc_multiplier, 328);
+ CHECK_OFFSET(extended_instruction_info, 336);
CHECK_OFFSET(cr0_guest_host_mask, 344);
CHECK_OFFSET(cr4_guest_host_mask, 352);
CHECK_OFFSET(cr0_read_shadow, 360);
diff --git a/arch/x86/kvm/vmx/vmx.h b/arch/x86/kvm/vmx/vmx.h
index 64a0772c883c..b8da6ebc35dc 100644
--- a/arch/x86/kvm/vmx/vmx.h
+++ b/arch/x86/kvm/vmx/vmx.h
@@ -372,6 +372,8 @@ struct vmx_insn_info {
union insn_info info;
};
+static inline bool vmx_egpr_enabled(struct kvm_vcpu *vcpu __maybe_unused) { return false; }
+
static inline struct vmx_insn_info vmx_get_insn_info(struct kvm_vcpu *vcpu __maybe_unused)
{
struct vmx_insn_info insn;
--
2.51.0
^ permalink raw reply related [flat|nested] 55+ messages in thread* Re: [PATCH RFC v1 07/20] KVM: nVMX: Support the extended instruction info field
2025-11-10 18:01 ` [PATCH RFC v1 07/20] KVM: nVMX: Support the extended instruction info field Chang S. Bae
@ 2025-11-11 17:48 ` Paolo Bonzini
2025-11-12 1:54 ` Chao Gao
2025-11-13 23:20 ` Chang S. Bae
0 siblings, 2 replies; 55+ messages in thread
From: Paolo Bonzini @ 2025-11-11 17:48 UTC (permalink / raw)
To: Chang S. Bae, kvm, linux-kernel; +Cc: seanjc, chao.gao, zhao1.liu
On 11/10/25 19:01, Chang S. Bae wrote:
> Define the VMCS field offset for the extended instruction information and
> handle it for nested VMX.
>
> When EGPRs are available, VMX provides a new 64-bit field to extend the
> legacy instruction information, allowing access to the higher register
> indices. Then, nested VMX needs to propagate this field between L1 and
> L2.
>
> The EGPR checker will be implemented later.
>
> Signed-off-by: Chang S. Bae <chang.seok.bae@intel.com>
> ---
> RFC note:
> During the draft, I brought up the offset definition initially for
> non-nested VMX primarily. Then, I realized the switching helper affects
> nVMX code anyway. Due to this dependency, this change is placed first
> together with the offset definition.
> ---
> arch/x86/include/asm/vmx.h | 2 ++
> arch/x86/kvm/vmx/nested.c | 2 ++
> arch/x86/kvm/vmx/vmcs12.c | 1 +
> arch/x86/kvm/vmx/vmcs12.h | 3 ++-
> arch/x86/kvm/vmx/vmx.h | 2 ++
> 5 files changed, 9 insertions(+), 1 deletion(-)
>
> diff --git a/arch/x86/include/asm/vmx.h b/arch/x86/include/asm/vmx.h
> index c85c50019523..ab0684948c56 100644
> --- a/arch/x86/include/asm/vmx.h
> +++ b/arch/x86/include/asm/vmx.h
> @@ -264,6 +264,8 @@ enum vmcs_field {
> PID_POINTER_TABLE_HIGH = 0x00002043,
> GUEST_PHYSICAL_ADDRESS = 0x00002400,
> GUEST_PHYSICAL_ADDRESS_HIGH = 0x00002401,
> + EXTENDED_INSTRUCTION_INFO = 0x00002406,
> + EXTENDED_INSTRUCTION_INFO_HIGH = 0x00002407,
> VMCS_LINK_POINTER = 0x00002800,
> VMCS_LINK_POINTER_HIGH = 0x00002801,
> GUEST_IA32_DEBUGCTL = 0x00002802,
> diff --git a/arch/x86/kvm/vmx/nested.c b/arch/x86/kvm/vmx/nested.c
> index 97ec8e594155..3442610a6b70 100644
> --- a/arch/x86/kvm/vmx/nested.c
> +++ b/arch/x86/kvm/vmx/nested.c
> @@ -4798,6 +4798,8 @@ static void prepare_vmcs12(struct kvm_vcpu *vcpu, struct vmcs12 *vmcs12,
> vmcs12->vm_exit_intr_info = exit_intr_info;
> vmcs12->vm_exit_instruction_len = exit_insn_len;
> vmcs12->vmx_instruction_info = vmcs_read32(VMX_INSTRUCTION_INFO);
> + if (vmx_egpr_enabled(vcpu))
> + vmcs12->extended_instruction_info = vmcs_read64(EXTENDED_INSTRUCTION_INFO);
From patch 17:
+static inline bool vmx_egpr_enabled(struct kvm_vcpu *vcpu)
+{
+ return vcpu->arch.xcr0 & XFEATURE_MASK_APX && is_64_bit_mode(vcpu);
+}
but here you must not check XCR0, the extended instruction information
field is always available. The spec says "A non-Intel® APX enabled VMM
is free to continue using the legacy definition of the field, since lack
of Intel® APX enabling will guarantee that regIDs are only 4-bits,
maximum" but you can also use the extended instruction information field
if you want. So, I'd make this also static_cpu_has(X86_FEATURE_APX).
Paolo
>
> /*
> * According to spec, there's no need to store the guest's
> diff --git a/arch/x86/kvm/vmx/vmcs12.c b/arch/x86/kvm/vmx/vmcs12.c
> index 4233b5ca9461..ea2b690a419e 100644
> --- a/arch/x86/kvm/vmx/vmcs12.c
> +++ b/arch/x86/kvm/vmx/vmcs12.c
> @@ -53,6 +53,7 @@ const unsigned short vmcs12_field_offsets[] = {
> FIELD64(XSS_EXIT_BITMAP, xss_exit_bitmap),
> FIELD64(ENCLS_EXITING_BITMAP, encls_exiting_bitmap),
> FIELD64(GUEST_PHYSICAL_ADDRESS, guest_physical_address),
> + FIELD64(EXTENDED_INSTRUCTION_INFO, extended_instruction_info),
> FIELD64(VMCS_LINK_POINTER, vmcs_link_pointer),
> FIELD64(GUEST_IA32_DEBUGCTL, guest_ia32_debugctl),
> FIELD64(GUEST_IA32_PAT, guest_ia32_pat),
> diff --git a/arch/x86/kvm/vmx/vmcs12.h b/arch/x86/kvm/vmx/vmcs12.h
> index 4ad6b16525b9..2146e45aaade 100644
> --- a/arch/x86/kvm/vmx/vmcs12.h
> +++ b/arch/x86/kvm/vmx/vmcs12.h
> @@ -71,7 +71,7 @@ struct __packed vmcs12 {
> u64 pml_address;
> u64 encls_exiting_bitmap;
> u64 tsc_multiplier;
> - u64 padding64[1]; /* room for future expansion */
> + u64 extended_instruction_info;
> /*
> * To allow migration of L1 (complete with its L2 guests) between
> * machines of different natural widths (32 or 64 bit), we cannot have
> @@ -261,6 +261,7 @@ static inline void vmx_check_vmcs12_offsets(void)
> CHECK_OFFSET(pml_address, 312);
> CHECK_OFFSET(encls_exiting_bitmap, 320);
> CHECK_OFFSET(tsc_multiplier, 328);
> + CHECK_OFFSET(extended_instruction_info, 336);
> CHECK_OFFSET(cr0_guest_host_mask, 344);
> CHECK_OFFSET(cr4_guest_host_mask, 352);
> CHECK_OFFSET(cr0_read_shadow, 360);
> diff --git a/arch/x86/kvm/vmx/vmx.h b/arch/x86/kvm/vmx/vmx.h
> index 64a0772c883c..b8da6ebc35dc 100644
> --- a/arch/x86/kvm/vmx/vmx.h
> +++ b/arch/x86/kvm/vmx/vmx.h
> @@ -372,6 +372,8 @@ struct vmx_insn_info {
> union insn_info info;
> };
>
> +static inline bool vmx_egpr_enabled(struct kvm_vcpu *vcpu __maybe_unused) { return false; }
> +
> static inline struct vmx_insn_info vmx_get_insn_info(struct kvm_vcpu *vcpu __maybe_unused)
> {
> struct vmx_insn_info insn;
^ permalink raw reply [flat|nested] 55+ messages in thread* Re: [PATCH RFC v1 07/20] KVM: nVMX: Support the extended instruction info field
2025-11-11 17:48 ` Paolo Bonzini
@ 2025-11-12 1:54 ` Chao Gao
2025-11-13 23:21 ` Chang S. Bae
2025-11-17 23:29 ` Paolo Bonzini
2025-11-13 23:20 ` Chang S. Bae
1 sibling, 2 replies; 55+ messages in thread
From: Chao Gao @ 2025-11-12 1:54 UTC (permalink / raw)
To: Paolo Bonzini; +Cc: Chang S. Bae, kvm, linux-kernel, seanjc, zhao1.liu
>> --- a/arch/x86/kvm/vmx/nested.c
>> +++ b/arch/x86/kvm/vmx/nested.c
>> @@ -4798,6 +4798,8 @@ static void prepare_vmcs12(struct kvm_vcpu *vcpu, struct vmcs12 *vmcs12,
>> vmcs12->vm_exit_intr_info = exit_intr_info;
>> vmcs12->vm_exit_instruction_len = exit_insn_len;
>> vmcs12->vmx_instruction_info = vmcs_read32(VMX_INSTRUCTION_INFO);
>> + if (vmx_egpr_enabled(vcpu))
>> + vmcs12->extended_instruction_info = vmcs_read64(EXTENDED_INSTRUCTION_INFO);
>
>From patch 17:
>
>+static inline bool vmx_egpr_enabled(struct kvm_vcpu *vcpu)
>+{
>+ return vcpu->arch.xcr0 & XFEATURE_MASK_APX && is_64_bit_mode(vcpu);
>+}
>
>but here you must not check XCR0, the extended instruction information field
>is always available. The spec says "A non-Intel® APX enabled VMM is free to
>continue using the legacy definition of the field, since lack of Intel® APX
>enabling will guarantee that regIDs are only 4-bits, maximum" but you can
>also use the extended instruction information field if you want. So, I'd
>make this also static_cpu_has(X86_FEATURE_APX).
Shouldn't we check guest's capabilities rather than host's,
i.e., guest_cpu_cap_has(X86_FEATURE_APX)?
^ permalink raw reply [flat|nested] 55+ messages in thread* Re: [PATCH RFC v1 07/20] KVM: nVMX: Support the extended instruction info field
2025-11-12 1:54 ` Chao Gao
@ 2025-11-13 23:21 ` Chang S. Bae
2025-11-17 23:29 ` Paolo Bonzini
1 sibling, 0 replies; 55+ messages in thread
From: Chang S. Bae @ 2025-11-13 23:21 UTC (permalink / raw)
To: Chao Gao, Paolo Bonzini; +Cc: kvm, linux-kernel, seanjc, zhao1.liu
On 11/11/2025 5:54 PM, Chao Gao wrote:
>
> Shouldn't we check guest's capabilities rather than host's,
>
> i.e., guest_cpu_cap_has(X86_FEATURE_APX)?
Perhaps adding a comment like this would help clarify it:
/*
* The APX enumeration guarantees the presence of the extended
* field. The host CPUID bit alone is sufficient to rely on it.
*/
static inline bool vmx_ext_insn_info_available(void) {
return static_cpu_has(X86_FEATURE_APX);
}
^ permalink raw reply [flat|nested] 55+ messages in thread* Re: [PATCH RFC v1 07/20] KVM: nVMX: Support the extended instruction info field
2025-11-12 1:54 ` Chao Gao
2025-11-13 23:21 ` Chang S. Bae
@ 2025-11-17 23:29 ` Paolo Bonzini
2025-11-18 1:39 ` Chao Gao
1 sibling, 1 reply; 55+ messages in thread
From: Paolo Bonzini @ 2025-11-17 23:29 UTC (permalink / raw)
To: Chao Gao
Cc: Chang S. Bae, kvm, Kernel Mailing List, Linux,
Sean Christopherson, Zhao Liu
Il mer 12 nov 2025, 02:54 Chao Gao <chao.gao@intel.com> ha scritto:
>
> Shouldn't we check guest's capabilities rather than host's,
>
> i.e., guest_cpu_cap_has(X86_FEATURE_APX)?
As the manual says, you're free to use the extended field if
available, and it's faster.
Paolo
^ permalink raw reply [flat|nested] 55+ messages in thread
* Re: [PATCH RFC v1 07/20] KVM: nVMX: Support the extended instruction info field
2025-11-17 23:29 ` Paolo Bonzini
@ 2025-11-18 1:39 ` Chao Gao
2025-11-18 10:33 ` Paolo Bonzini
0 siblings, 1 reply; 55+ messages in thread
From: Chao Gao @ 2025-11-18 1:39 UTC (permalink / raw)
To: Paolo Bonzini
Cc: Chang S. Bae, kvm, Kernel Mailing List, Linux,
Sean Christopherson, Zhao Liu
On Tue, Nov 18, 2025 at 12:29:19AM +0100, Paolo Bonzini wrote:
>Il mer 12 nov 2025, 02:54 Chao Gao <chao.gao@intel.com> ha scritto:
>>
>> Shouldn't we check guest's capabilities rather than host's,
>>
>> i.e., guest_cpu_cap_has(X86_FEATURE_APX)?
>
>As the manual says, you're free to use the extended field if
>available, and it's faster.
The point is, from the guest's perspective, the field is available iff the vCPU
supports APX. KVM (L0) doesn't need to virtualize VMCS12's EII field if the vCPU
doesn't have APX.
For other call sites of vmx_egpr_enabled(), I agree we should just check host
capabilities.
^ permalink raw reply [flat|nested] 55+ messages in thread
* Re: [PATCH RFC v1 07/20] KVM: nVMX: Support the extended instruction info field
2025-11-18 1:39 ` Chao Gao
@ 2025-11-18 10:33 ` Paolo Bonzini
0 siblings, 0 replies; 55+ messages in thread
From: Paolo Bonzini @ 2025-11-18 10:33 UTC (permalink / raw)
To: Chao Gao
Cc: Chang S. Bae, kvm, Kernel Mailing List, Linux,
Sean Christopherson, Zhao Liu
On Tue, Nov 18, 2025 at 2:39 AM Chao Gao <chao.gao@intel.com> wrote:
>
> On Tue, Nov 18, 2025 at 12:29:19AM +0100, Paolo Bonzini wrote:
> >Il mer 12 nov 2025, 02:54 Chao Gao <chao.gao@intel.com> ha scritto:
> >>
> >> Shouldn't we check guest's capabilities rather than host's,
> >>
> >> i.e., guest_cpu_cap_has(X86_FEATURE_APX)?
> >
> >As the manual says, you're free to use the extended field if
> >available, and it's faster.
>
> The point is, from the guest's perspective, the field is available iff the vCPU
> supports APX. KVM (L0) doesn't need to virtualize VMCS12's EII field if the vCPU
> doesn't have APX.
Well, it would be faster to just do it. But you're right, checking the
guest CPUID is consistent with other code, for example
if (nested_cpu_has_vid(vmcs12))
vmcs12->guest_intr_status = vmcs_read16(GUEST_INTR_STATUS);
Paolo
^ permalink raw reply [flat|nested] 55+ messages in thread
* Re: [PATCH RFC v1 07/20] KVM: nVMX: Support the extended instruction info field
2025-11-11 17:48 ` Paolo Bonzini
2025-11-12 1:54 ` Chao Gao
@ 2025-11-13 23:20 ` Chang S. Bae
1 sibling, 0 replies; 55+ messages in thread
From: Chang S. Bae @ 2025-11-13 23:20 UTC (permalink / raw)
To: Paolo Bonzini, kvm, linux-kernel; +Cc: seanjc, chao.gao, zhao1.liu
On 11/11/2025 9:48 AM, Paolo Bonzini wrote:
>
> but here you must not check XCR0, the extended instruction information
> field is always available. The spec says "A non-Intel® APX enabled VMM
> is free to continue using the legacy definition of the field, since lack
> of Intel® APX enabling will guarantee that regIDs are only 4-bits,
> maximum" but you can also use the extended instruction information field
> if you want. So, I'd make this also static_cpu_has(X86_FEATURE_APX).
I just got confirmation from the hardware folks. The CPUID enumeration
alone is sufficient to indicate the presence and usability of the
extended instruction information field, regardless of the XCR0 state.
^ permalink raw reply [flat|nested] 55+ messages in thread
* [PATCH RFC v1 08/20] KVM: VMX: Support extended register index in exit handling
2025-11-10 18:01 [PATCH RFC v1 00/20] KVM: x86: Support APX feature for guests Chang S. Bae
` (6 preceding siblings ...)
2025-11-10 18:01 ` [PATCH RFC v1 07/20] KVM: nVMX: Support the extended instruction info field Chang S. Bae
@ 2025-11-10 18:01 ` Chang S. Bae
2025-11-11 17:45 ` Paolo Bonzini
2025-11-10 18:01 ` [PATCH RFC v1 09/20] KVM: x86: Support EGPR accessing and tracking for instruction emulation Chang S. Bae
` (13 subsequent siblings)
21 siblings, 1 reply; 55+ messages in thread
From: Chang S. Bae @ 2025-11-10 18:01 UTC (permalink / raw)
To: kvm, linux-kernel; +Cc: pbonzini, seanjc, chao.gao, zhao1.liu, chang.seok.bae
Support to 5-bit register indices in VMCS fields when EGPRs are enabled.
Signed-off-by: Chang S. Bae <chang.seok.bae@intel.com>
---
RFC note:
The "chicken bit" (XCR0.APX) checker is intentionally deferred, as the
emulator in the next series will do a similar check. Consolidating the
XCR0 handling at the end keeps the logic clearer during the feature
exposition.
---
arch/x86/kvm/vmx/vmx.h | 16 ++++++++++++----
1 file changed, 12 insertions(+), 4 deletions(-)
diff --git a/arch/x86/kvm/vmx/vmx.h b/arch/x86/kvm/vmx/vmx.h
index b8da6ebc35dc..6cf1eb739caf 100644
--- a/arch/x86/kvm/vmx/vmx.h
+++ b/arch/x86/kvm/vmx/vmx.h
@@ -374,12 +374,17 @@ struct vmx_insn_info {
static inline bool vmx_egpr_enabled(struct kvm_vcpu *vcpu __maybe_unused) { return false; }
-static inline struct vmx_insn_info vmx_get_insn_info(struct kvm_vcpu *vcpu __maybe_unused)
+static inline struct vmx_insn_info vmx_get_insn_info(struct kvm_vcpu *vcpu)
{
struct vmx_insn_info insn;
- insn.extended = false;
- insn.info.word = vmcs_read32(VMX_INSTRUCTION_INFO);
+ if (vmx_egpr_enabled(vcpu)) {
+ insn.extended = true;
+ insn.info.dword = vmcs_read64(EXTENDED_INSTRUCTION_INFO);
+ } else {
+ insn.extended = false;
+ insn.info.word = vmcs_read32(VMX_INSTRUCTION_INFO);
+ }
return insn;
}
@@ -415,7 +420,10 @@ static __always_inline unsigned long vmx_get_exit_qual(struct kvm_vcpu *vcpu)
static inline int vmx_get_exit_qual_gpr(struct kvm_vcpu *vcpu)
{
- return (vmx_get_exit_qual(vcpu) >> 8) & 0xf;
+ if (vmx_egpr_enabled(vcpu))
+ return (vmx_get_exit_qual(vcpu) >> 8) & 0x1f;
+ else
+ return (vmx_get_exit_qual(vcpu) >> 8) & 0xf;
}
static __always_inline u32 vmx_get_intr_info(struct kvm_vcpu *vcpu)
--
2.51.0
^ permalink raw reply related [flat|nested] 55+ messages in thread* Re: [PATCH RFC v1 08/20] KVM: VMX: Support extended register index in exit handling
2025-11-10 18:01 ` [PATCH RFC v1 08/20] KVM: VMX: Support extended register index in exit handling Chang S. Bae
@ 2025-11-11 17:45 ` Paolo Bonzini
2025-11-13 23:22 ` Chang S. Bae
0 siblings, 1 reply; 55+ messages in thread
From: Paolo Bonzini @ 2025-11-11 17:45 UTC (permalink / raw)
To: Chang S. Bae, kvm, linux-kernel; +Cc: seanjc, chao.gao, zhao1.liu
On 11/10/25 19:01, Chang S. Bae wrote:
> Support to 5-bit register indices in VMCS fields when EGPRs are enabled.
>
> Signed-off-by: Chang S. Bae <chang.seok.bae@intel.com>
> ---
> RFC note:
> The "chicken bit" (XCR0.APX) checker is intentionally deferred, as the
> emulator in the next series will do a similar check. Consolidating the
> XCR0 handling at the end keeps the logic clearer during the feature
> exposition.
> ---
> arch/x86/kvm/vmx/vmx.h | 16 ++++++++++++----
> 1 file changed, 12 insertions(+), 4 deletions(-)
>
> diff --git a/arch/x86/kvm/vmx/vmx.h b/arch/x86/kvm/vmx/vmx.h
> index b8da6ebc35dc..6cf1eb739caf 100644
> --- a/arch/x86/kvm/vmx/vmx.h
> +++ b/arch/x86/kvm/vmx/vmx.h
> @@ -374,12 +374,17 @@ struct vmx_insn_info {
>
> static inline bool vmx_egpr_enabled(struct kvm_vcpu *vcpu __maybe_unused) { return false; }
>
> -static inline struct vmx_insn_info vmx_get_insn_info(struct kvm_vcpu *vcpu __maybe_unused)
> +static inline struct vmx_insn_info vmx_get_insn_info(struct kvm_vcpu *vcpu)
> {
> struct vmx_insn_info insn;
>
> - insn.extended = false;
> - insn.info.word = vmcs_read32(VMX_INSTRUCTION_INFO);
> + if (vmx_egpr_enabled(vcpu)) {
> + insn.extended = true;
> + insn.info.dword = vmcs_read64(EXTENDED_INSTRUCTION_INFO);
> + } else {
> + insn.extended = false;
> + insn.info.word = vmcs_read32(VMX_INSTRUCTION_INFO);
> + }
Could this use static_cpu_has(X86_FEATURE_APX) instead, which is more
efficient (avoids a runtime test)?
> return insn;
> }
> @@ -415,7 +420,10 @@ static __always_inline unsigned long vmx_get_exit_qual(struct kvm_vcpu *vcpu)
>
> static inline int vmx_get_exit_qual_gpr(struct kvm_vcpu *vcpu)
> {
> - return (vmx_get_exit_qual(vcpu) >> 8) & 0xf;
> + if (vmx_egpr_enabled(vcpu))
> + return (vmx_get_exit_qual(vcpu) >> 8) & 0x1f;
> + else
> + return (vmx_get_exit_qual(vcpu) >> 8) & 0xf;
Can this likewise mask against 0x1f, unconditionally?
Paolo
> }
>
> static __always_inline u32 vmx_get_intr_info(struct kvm_vcpu *vcpu)
^ permalink raw reply [flat|nested] 55+ messages in thread* Re: [PATCH RFC v1 08/20] KVM: VMX: Support extended register index in exit handling
2025-11-11 17:45 ` Paolo Bonzini
@ 2025-11-13 23:22 ` Chang S. Bae
2025-11-13 23:40 ` Paolo Bonzini
0 siblings, 1 reply; 55+ messages in thread
From: Chang S. Bae @ 2025-11-13 23:22 UTC (permalink / raw)
To: Paolo Bonzini, kvm, linux-kernel; +Cc: seanjc, chao.gao, zhao1.liu
On 11/11/2025 9:45 AM, Paolo Bonzini wrote:
> On 11/10/25 19:01, Chang S. Bae wrote:
>>
>> -static inline struct vmx_insn_info vmx_get_insn_info(struct kvm_vcpu
>> *vcpu __maybe_unused)
>> +static inline struct vmx_insn_info vmx_get_insn_info(struct kvm_vcpu
>> *vcpu)
>> {
>> struct vmx_insn_info insn;
>> - insn.extended = false;
>> - insn.info.word = vmcs_read32(VMX_INSTRUCTION_INFO);
>> + if (vmx_egpr_enabled(vcpu)) {
>> + insn.extended = true;
>> + insn.info.dword = vmcs_read64(EXTENDED_INSTRUCTION_INFO);
>> + } else {
>> + insn.extended = false;
>> + insn.info.word = vmcs_read32(VMX_INSTRUCTION_INFO);
>> + }
>
> Could this use static_cpu_has(X86_FEATURE_APX) instead, which is more
> efficient (avoids a runtime test)?
Yes, for the same reason mentioned in patch7.
>> @@ -415,7 +420,10 @@ static __always_inline unsigned long
>> vmx_get_exit_qual(struct kvm_vcpu *vcpu)
>> static inline int vmx_get_exit_qual_gpr(struct kvm_vcpu *vcpu)
>> {
>> - return (vmx_get_exit_qual(vcpu) >> 8) & 0xf;
>> + if (vmx_egpr_enabled(vcpu))
>> + return (vmx_get_exit_qual(vcpu) >> 8) & 0x1f;
>> + else
>> + return (vmx_get_exit_qual(vcpu) >> 8) & 0xf;
>
> Can this likewise mask against 0x1f, unconditionally?
It looks like the behavior of that previously-undefined bit is not
guaranteed -- there's no architectural promise that the bit will always
read as zero. So in this case, I think it's still safer to rely on the
enumeration.
Perhaps adding a comment like this would clarify the intent:
/*
* Bit 12 was previously undefined, so its value is not guaranteed to
* be zero. Only rely on the full 5-bit with the extension.
*/
if (vmx_ext_insn_info_available())
...
^ permalink raw reply [flat|nested] 55+ messages in thread* Re: [PATCH RFC v1 08/20] KVM: VMX: Support extended register index in exit handling
2025-11-13 23:22 ` Chang S. Bae
@ 2025-11-13 23:40 ` Paolo Bonzini
0 siblings, 0 replies; 55+ messages in thread
From: Paolo Bonzini @ 2025-11-13 23:40 UTC (permalink / raw)
To: Chang S. Bae, kvm, linux-kernel; +Cc: seanjc, chao.gao, zhao1.liu
On 11/14/25 00:22, Chang S. Bae wrote:
>>> @@ -415,7 +420,10 @@ static __always_inline unsigned long
>>> vmx_get_exit_qual(struct kvm_vcpu *vcpu)
>>> static inline int vmx_get_exit_qual_gpr(struct kvm_vcpu *vcpu)
>>> {
>>> - return (vmx_get_exit_qual(vcpu) >> 8) & 0xf;
>>> + if (vmx_egpr_enabled(vcpu))
>>> + return (vmx_get_exit_qual(vcpu) >> 8) & 0x1f;
>>> + else
>>> + return (vmx_get_exit_qual(vcpu) >> 8) & 0xf;
>>
>> Can this likewise mask against 0x1f, unconditionally?
>
> It looks like the behavior of that previously-undefined bit is not
> guaranteed -- there's no architectural promise that the bit will always
> read as zero. So in this case, I think it's still safer to rely on the
> enumeration.
The manual says "VM Exit qualification is extended in-place"... This
suggests that the wording of the SDM, "Not currently defined", really
means "currently zero, but may not be *in the future*".
It is indeed safer, but it's a bit sad too. :) Let's at least use
either static_cpu_has() or a global __read_mostly variable, instead of a
more expensive test using the vcpu.
Paolo
^ permalink raw reply [flat|nested] 55+ messages in thread
* [PATCH RFC v1 09/20] KVM: x86: Support EGPR accessing and tracking for instruction emulation
2025-11-10 18:01 [PATCH RFC v1 00/20] KVM: x86: Support APX feature for guests Chang S. Bae
` (7 preceding siblings ...)
2025-11-10 18:01 ` [PATCH RFC v1 08/20] KVM: VMX: Support extended register index in exit handling Chang S. Bae
@ 2025-11-10 18:01 ` Chang S. Bae
2025-11-10 18:01 ` [PATCH RFC v1 10/20] KVM: x86: Refactor REX prefix handling in " Chang S. Bae
` (12 subsequent siblings)
21 siblings, 0 replies; 55+ messages in thread
From: Chang S. Bae @ 2025-11-10 18:01 UTC (permalink / raw)
To: kvm, linux-kernel; +Cc: pbonzini, seanjc, chao.gao, zhao1.liu, chang.seok.bae
Extend the emulator context and GPR accessors to handle EGPRs before
adding support for REX2-prefixed instructions.
Now the KVM GPR accessors can handle EGPRs. Then, the emulator can
uniformly cache and track all GPRs without requiring separate handling.
Signed-off-by: Chang S. Bae <chang.seok.bae@intel.com>
---
arch/x86/kvm/kvm_emulate.h | 10 +++++-----
arch/x86/kvm/x86.c | 4 ++--
2 files changed, 7 insertions(+), 7 deletions(-)
diff --git a/arch/x86/kvm/kvm_emulate.h b/arch/x86/kvm/kvm_emulate.h
index 7b5ddb787a25..153c70ea5561 100644
--- a/arch/x86/kvm/kvm_emulate.h
+++ b/arch/x86/kvm/kvm_emulate.h
@@ -105,13 +105,13 @@ struct x86_instruction_info {
struct x86_emulate_ops {
void (*vm_bugged)(struct x86_emulate_ctxt *ctxt);
/*
- * read_gpr: read a general purpose register (rax - r15)
+ * read_gpr: read a general purpose register (rax - r31)
*
* @reg: gpr number.
*/
ulong (*read_gpr)(struct x86_emulate_ctxt *ctxt, unsigned reg);
/*
- * write_gpr: write a general purpose register (rax - r15)
+ * write_gpr: write a general purpose register (rax - r31)
*
* @reg: gpr number.
* @val: value to write.
@@ -312,7 +312,7 @@ typedef void (*fastop_t)(struct fastop *);
* a ModRM or SIB byte.
*/
#ifdef CONFIG_X86_64
-#define NR_EMULATOR_GPRS 16
+#define NR_EMULATOR_GPRS 32
#else
#define NR_EMULATOR_GPRS 8
#endif
@@ -361,9 +361,9 @@ struct x86_emulate_ctxt {
u8 lock_prefix;
u8 rep_prefix;
/* bitmaps of registers in _regs[] that can be read */
- u16 regs_valid;
+ u32 regs_valid;
/* bitmaps of registers in _regs[] that have been written */
- u16 regs_dirty;
+ u32 regs_dirty;
/* modrm */
u8 modrm;
u8 modrm_mod;
diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
index 603057ea7421..338986a5a3ae 100644
--- a/arch/x86/kvm/x86.c
+++ b/arch/x86/kvm/x86.c
@@ -8812,12 +8812,12 @@ static bool emulator_guest_cpuid_is_intel_compatible(struct x86_emulate_ctxt *ct
static ulong emulator_read_gpr(struct x86_emulate_ctxt *ctxt, unsigned reg)
{
- return kvm_register_read_raw(emul_to_vcpu(ctxt), reg);
+ return _kvm_gpr_read(emul_to_vcpu(ctxt), reg);
}
static void emulator_write_gpr(struct x86_emulate_ctxt *ctxt, unsigned reg, ulong val)
{
- kvm_register_write_raw(emul_to_vcpu(ctxt), reg, val);
+ _kvm_gpr_write(emul_to_vcpu(ctxt), reg, val);
}
static void emulator_set_nmi_mask(struct x86_emulate_ctxt *ctxt, bool masked)
--
2.51.0
^ permalink raw reply related [flat|nested] 55+ messages in thread* [PATCH RFC v1 10/20] KVM: x86: Refactor REX prefix handling in instruction emulation
2025-11-10 18:01 [PATCH RFC v1 00/20] KVM: x86: Support APX feature for guests Chang S. Bae
` (8 preceding siblings ...)
2025-11-10 18:01 ` [PATCH RFC v1 09/20] KVM: x86: Support EGPR accessing and tracking for instruction emulation Chang S. Bae
@ 2025-11-10 18:01 ` Chang S. Bae
2025-11-11 18:17 ` Paolo Bonzini
2025-11-10 18:01 ` [PATCH RFC v1 11/20] KVM: x86: Refactor opcode table lookup " Chang S. Bae
` (11 subsequent siblings)
21 siblings, 1 reply; 55+ messages in thread
From: Chang S. Bae @ 2025-11-10 18:01 UTC (permalink / raw)
To: kvm, linux-kernel; +Cc: pbonzini, seanjc, chao.gao, zhao1.liu, chang.seok.bae
Restructure how to represent and interpret REX fields. Specifically,
* Repurpose the existing rex_prefix field to identify the prefix type
* Introduce a new union to hold both REX and REX2 bitfields
* Update decoder logic to interpret the unified data type
Historically, REX used the upper four bits of a signle byte as a fixed
identifier, with the lower bits encoded. REX2 extends this to two bytes.
The first byte identifies the prefix, and the second encodes additional
bits, preserving compatibility with legacy REX encoding.
Previously, the emulator stored the REX byte as-is, which cannot capture
REX2 semantics. This refactor prepares for REX2 decoding while preserving
current behavior.
No functional changes intended.
Signed-off-by: Chang S. Bae <chang.seok.bae@intel.com>
---
arch/x86/kvm/emulate.c | 33 ++++++++++++++++++---------------
arch/x86/kvm/kvm_emulate.h | 31 ++++++++++++++++++++++++++++++-
2 files changed, 48 insertions(+), 16 deletions(-)
diff --git a/arch/x86/kvm/emulate.c b/arch/x86/kvm/emulate.c
index 4e3da5b497b8..763fbd139242 100644
--- a/arch/x86/kvm/emulate.c
+++ b/arch/x86/kvm/emulate.c
@@ -924,7 +924,7 @@ static void *decode_register(struct x86_emulate_ctxt *ctxt, u8 modrm_reg,
int byteop)
{
void *p;
- int highbyte_regs = (ctxt->rex_prefix == 0) && byteop;
+ int highbyte_regs = (ctxt->rex_prefix == REX_NONE) && byteop;
if (highbyte_regs && modrm_reg >= 4 && modrm_reg < 8)
p = (unsigned char *)reg_rmw(ctxt, modrm_reg & 3) + 1;
@@ -1080,10 +1080,12 @@ static void decode_register_operand(struct x86_emulate_ctxt *ctxt,
{
unsigned int reg;
- if (ctxt->d & ModRM)
+ if (ctxt->d & ModRM) {
reg = ctxt->modrm_reg;
- else
- reg = (ctxt->b & 7) | ((ctxt->rex_prefix & 1) << 3);
+ } else {
+ reg = (ctxt->b & 7) |
+ (ctxt->rex.bits.b3 * BIT(3));
+ }
if (ctxt->d & Sse) {
op->type = OP_XMM;
@@ -1122,9 +1124,9 @@ static int decode_modrm(struct x86_emulate_ctxt *ctxt,
int rc = X86EMUL_CONTINUE;
ulong modrm_ea = 0;
- ctxt->modrm_reg = ((ctxt->rex_prefix << 1) & 8); /* REX.R */
- index_reg = (ctxt->rex_prefix << 2) & 8; /* REX.X */
- base_reg = (ctxt->rex_prefix << 3) & 8; /* REX.B */
+ ctxt->modrm_reg = ctxt->rex.bits.r3 * BIT(3);
+ index_reg = ctxt->rex.bits.x3 * BIT(3);
+ base_reg = ctxt->rex.bits.b3 * BIT(3);
ctxt->modrm_mod = (ctxt->modrm & 0xc0) >> 6;
ctxt->modrm_reg |= (ctxt->modrm & 0x38) >> 3;
@@ -2466,7 +2468,7 @@ static int em_sysexit(struct x86_emulate_ctxt *ctxt)
setup_syscalls_segments(&cs, &ss);
- if ((ctxt->rex_prefix & 0x8) != 0x0)
+ if (ctxt->rex.bits.w)
usermode = X86EMUL_MODE_PROT64;
else
usermode = X86EMUL_MODE_PROT32;
@@ -4851,7 +4853,8 @@ int x86_decode_insn(struct x86_emulate_ctxt *ctxt, void *insn, int insn_len, int
case 0x40 ... 0x4f: /* REX */
if (mode != X86EMUL_MODE_PROT64)
goto done_prefixes;
- ctxt->rex_prefix = ctxt->b;
+ ctxt->rex_prefix = REX_PREFIX;
+ ctxt->rex.raw = 0x0f & ctxt->b;
continue;
case 0xf0: /* LOCK */
ctxt->lock_prefix = 1;
@@ -4865,15 +4868,14 @@ int x86_decode_insn(struct x86_emulate_ctxt *ctxt, void *insn, int insn_len, int
}
/* Any legacy prefix after a REX prefix nullifies its effect. */
-
- ctxt->rex_prefix = 0;
+ ctxt->rex_prefix = REX_NONE;
+ ctxt->rex.raw = 0;
}
done_prefixes:
- /* REX prefix. */
- if (ctxt->rex_prefix & 8)
- ctxt->op_bytes = 8; /* REX.W */
+ if (ctxt->rex.bits.w)
+ ctxt->op_bytes = 8;
/* Opcode byte(s). */
opcode = opcode_table[ctxt->b];
@@ -5137,7 +5139,8 @@ void init_decode_cache(struct x86_emulate_ctxt *ctxt)
{
/* Clear fields that are set conditionally but read without a guard. */
ctxt->rip_relative = false;
- ctxt->rex_prefix = 0;
+ ctxt->rex_prefix = REX_NONE;
+ ctxt->rex.raw = 0;
ctxt->lock_prefix = 0;
ctxt->rep_prefix = 0;
ctxt->regs_valid = 0;
diff --git a/arch/x86/kvm/kvm_emulate.h b/arch/x86/kvm/kvm_emulate.h
index 153c70ea5561..b285299ebfa4 100644
--- a/arch/x86/kvm/kvm_emulate.h
+++ b/arch/x86/kvm/kvm_emulate.h
@@ -317,6 +317,32 @@ typedef void (*fastop_t)(struct fastop *);
#define NR_EMULATOR_GPRS 8
#endif
+/*
+ * REX prefix type to distinguish between no prefix, legacy REX, REX2,
+ * or an invalid REX2 sequence.
+ */
+enum rex_type {
+ REX_NONE,
+ REX_PREFIX,
+ REX2_PREFIX,
+ REX2_INVALID
+};
+
+/* Unified representation for REX/REX2 prefix bits */
+union rex_field {
+ struct {
+ u8 b3 :1, /* REX2.B3 or REX.B */
+ x3 :1, /* REX2.X3 or REX.X */
+ r3 :1, /* REX2.R3 or REX.R */
+ w :1, /* REX2.W or REX.W */
+ b4 :1, /* REX2.B4 */
+ x4 :1, /* REX2.X4 */
+ r4 :1, /* REX2.R4 */
+ m0 :1; /* REX2.M0 */
+ } bits;
+ u8 raw;
+};
+
struct x86_emulate_ctxt {
void *vcpu;
const struct x86_emulate_ops *ops;
@@ -357,7 +383,10 @@ struct x86_emulate_ctxt {
int (*check_perm)(struct x86_emulate_ctxt *ctxt);
bool rip_relative;
- u8 rex_prefix;
+ /* Type of REX prefix (none, REX, REX2) */
+ enum rex_type rex_prefix;
+ /* Rex bits */
+ union rex_field rex;
u8 lock_prefix;
u8 rep_prefix;
/* bitmaps of registers in _regs[] that can be read */
--
2.51.0
^ permalink raw reply related [flat|nested] 55+ messages in thread* Re: [PATCH RFC v1 10/20] KVM: x86: Refactor REX prefix handling in instruction emulation
2025-11-10 18:01 ` [PATCH RFC v1 10/20] KVM: x86: Refactor REX prefix handling in " Chang S. Bae
@ 2025-11-11 18:17 ` Paolo Bonzini
2025-11-13 23:23 ` Chang S. Bae
0 siblings, 1 reply; 55+ messages in thread
From: Paolo Bonzini @ 2025-11-11 18:17 UTC (permalink / raw)
To: Chang S. Bae, kvm, linux-kernel; +Cc: seanjc, chao.gao, zhao1.liu
On 11/10/25 19:01, Chang S. Bae wrote:
> Restructure how to represent and interpret REX fields. Specifically,
>
> * Repurpose the existing rex_prefix field to identify the prefix type
> * Introduce a new union to hold both REX and REX2 bitfields
> * Update decoder logic to interpret the unified data type
>
> Historically, REX used the upper four bits of a signle byte as a fixed
> identifier, with the lower bits encoded. REX2 extends this to two bytes.
> The first byte identifies the prefix, and the second encodes additional
> bits, preserving compatibility with legacy REX encoding.
>
> Previously, the emulator stored the REX byte as-is, which cannot capture
> REX2 semantics. This refactor prepares for REX2 decoding while preserving
> current behavior.
>
> No functional changes intended.
>
> Signed-off-by: Chang S. Bae <chang.seok.bae@intel.com>
Good idea; having to add 0x40 to rex_prefix when decoding VEX was too
mysterious, I'll include something like this in my VEX patches. Here
is the commit message I came up with:
commit fc8aa5c45d558393069a1c89b7a64e059b8f9418
Author: Chang S. Bae <chang.seok.bae@intel.com>
Date: Mon Nov 10 18:01:21 2025 +0000
KVM: x86: Refactor REX prefix handling in instruction emulation
Restructure how to represent and interpret REX fields, preparing
for handling of VEX and REX2.
REX uses the upper four bits of a single byte as a fixed identifier,
and the lower four bits containing the data. VEX and REX2 extend this so
that the first byte identifies the prefix and the rest encode additional
bits; and while VEX only has the same four data bits as REX, eight zero
bits are a valid value for the data bits of REX2. So, stop storing the
REX byte as-is. Instead, store only the low bits of the REX prefix and
track separately whether a REX-like prefix wasused.
No functional changes intended.
Signed-off-by: Chang S. Bae <chang.seok.bae@intel.com>
Message-ID: <20251110180131.28264-11-chang.seok.bae@intel.com>
[Extracted from APX series; removed bitfields and REX2-specific
definitions. - Paolo]
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Paolo
> ---
> arch/x86/kvm/emulate.c | 33 ++++++++++++++++++---------------
> arch/x86/kvm/kvm_emulate.h | 31 ++++++++++++++++++++++++++++++-
> 2 files changed, 48 insertions(+), 16 deletions(-)
>
> diff --git a/arch/x86/kvm/emulate.c b/arch/x86/kvm/emulate.c
> index 4e3da5b497b8..763fbd139242 100644
> --- a/arch/x86/kvm/emulate.c
> +++ b/arch/x86/kvm/emulate.c
> @@ -924,7 +924,7 @@ static void *decode_register(struct x86_emulate_ctxt *ctxt, u8 modrm_reg,
> int byteop)
> {
> void *p;
> - int highbyte_regs = (ctxt->rex_prefix == 0) && byteop;
> + int highbyte_regs = (ctxt->rex_prefix == REX_NONE) && byteop;
>
> if (highbyte_regs && modrm_reg >= 4 && modrm_reg < 8)
> p = (unsigned char *)reg_rmw(ctxt, modrm_reg & 3) + 1;
> @@ -1080,10 +1080,12 @@ static void decode_register_operand(struct x86_emulate_ctxt *ctxt,
> {
> unsigned int reg;
>
> - if (ctxt->d & ModRM)
> + if (ctxt->d & ModRM) {
> reg = ctxt->modrm_reg;
> - else
> - reg = (ctxt->b & 7) | ((ctxt->rex_prefix & 1) << 3);
> + } else {
> + reg = (ctxt->b & 7) |
> + (ctxt->rex.bits.b3 * BIT(3));
> + }
>
> if (ctxt->d & Sse) {
> op->type = OP_XMM;
> @@ -1122,9 +1124,9 @@ static int decode_modrm(struct x86_emulate_ctxt *ctxt,
> int rc = X86EMUL_CONTINUE;
> ulong modrm_ea = 0;
>
> - ctxt->modrm_reg = ((ctxt->rex_prefix << 1) & 8); /* REX.R */
> - index_reg = (ctxt->rex_prefix << 2) & 8; /* REX.X */
> - base_reg = (ctxt->rex_prefix << 3) & 8; /* REX.B */
> + ctxt->modrm_reg = ctxt->rex.bits.r3 * BIT(3);
> + index_reg = ctxt->rex.bits.x3 * BIT(3);
> + base_reg = ctxt->rex.bits.b3 * BIT(3);
>
> ctxt->modrm_mod = (ctxt->modrm & 0xc0) >> 6;
> ctxt->modrm_reg |= (ctxt->modrm & 0x38) >> 3;
> @@ -2466,7 +2468,7 @@ static int em_sysexit(struct x86_emulate_ctxt *ctxt)
>
> setup_syscalls_segments(&cs, &ss);
>
> - if ((ctxt->rex_prefix & 0x8) != 0x0)
> + if (ctxt->rex.bits.w)
> usermode = X86EMUL_MODE_PROT64;
> else
> usermode = X86EMUL_MODE_PROT32;
> @@ -4851,7 +4853,8 @@ int x86_decode_insn(struct x86_emulate_ctxt *ctxt, void *insn, int insn_len, int
> case 0x40 ... 0x4f: /* REX */
> if (mode != X86EMUL_MODE_PROT64)
> goto done_prefixes;
> - ctxt->rex_prefix = ctxt->b;
> + ctxt->rex_prefix = REX_PREFIX;
> + ctxt->rex.raw = 0x0f & ctxt->b;
> continue;
> case 0xf0: /* LOCK */
> ctxt->lock_prefix = 1;
> @@ -4865,15 +4868,14 @@ int x86_decode_insn(struct x86_emulate_ctxt *ctxt, void *insn, int insn_len, int
> }
>
> /* Any legacy prefix after a REX prefix nullifies its effect. */
> -
> - ctxt->rex_prefix = 0;
> + ctxt->rex_prefix = REX_NONE;
> + ctxt->rex.raw = 0;
> }
>
> done_prefixes:
>
> - /* REX prefix. */
> - if (ctxt->rex_prefix & 8)
> - ctxt->op_bytes = 8; /* REX.W */
> + if (ctxt->rex.bits.w)
> + ctxt->op_bytes = 8;
>
> /* Opcode byte(s). */
> opcode = opcode_table[ctxt->b];
> @@ -5137,7 +5139,8 @@ void init_decode_cache(struct x86_emulate_ctxt *ctxt)
> {
> /* Clear fields that are set conditionally but read without a guard. */
> ctxt->rip_relative = false;
> - ctxt->rex_prefix = 0;
> + ctxt->rex_prefix = REX_NONE;
> + ctxt->rex.raw = 0;
> ctxt->lock_prefix = 0;
> ctxt->rep_prefix = 0;
> ctxt->regs_valid = 0;
> diff --git a/arch/x86/kvm/kvm_emulate.h b/arch/x86/kvm/kvm_emulate.h
> index 153c70ea5561..b285299ebfa4 100644
> --- a/arch/x86/kvm/kvm_emulate.h
> +++ b/arch/x86/kvm/kvm_emulate.h
> @@ -317,6 +317,32 @@ typedef void (*fastop_t)(struct fastop *);
> #define NR_EMULATOR_GPRS 8
> #endif
>
> +/*
> + * REX prefix type to distinguish between no prefix, legacy REX, REX2,
> + * or an invalid REX2 sequence.
> + */
> +enum rex_type {
> + REX_NONE,
> + REX_PREFIX,
> + REX2_PREFIX,
> + REX2_INVALID
> +};
> +
> +/* Unified representation for REX/REX2 prefix bits */
> +union rex_field {
> + struct {
> + u8 b3 :1, /* REX2.B3 or REX.B */
> + x3 :1, /* REX2.X3 or REX.X */
> + r3 :1, /* REX2.R3 or REX.R */
> + w :1, /* REX2.W or REX.W */
> + b4 :1, /* REX2.B4 */
> + x4 :1, /* REX2.X4 */
> + r4 :1, /* REX2.R4 */
> + m0 :1; /* REX2.M0 */
> + } bits;
> + u8 raw;
> +};
> +
> struct x86_emulate_ctxt {
> void *vcpu;
> const struct x86_emulate_ops *ops;
> @@ -357,7 +383,10 @@ struct x86_emulate_ctxt {
> int (*check_perm)(struct x86_emulate_ctxt *ctxt);
>
> bool rip_relative;
> - u8 rex_prefix;
> + /* Type of REX prefix (none, REX, REX2) */
> + enum rex_type rex_prefix;
> + /* Rex bits */
> + union rex_field rex;
> u8 lock_prefix;
> u8 rep_prefix;
> /* bitmaps of registers in _regs[] that can be read */
^ permalink raw reply [flat|nested] 55+ messages in thread* Re: [PATCH RFC v1 10/20] KVM: x86: Refactor REX prefix handling in instruction emulation
2025-11-11 18:17 ` Paolo Bonzini
@ 2025-11-13 23:23 ` Chang S. Bae
0 siblings, 0 replies; 55+ messages in thread
From: Chang S. Bae @ 2025-11-13 23:23 UTC (permalink / raw)
To: Paolo Bonzini, kvm, linux-kernel; +Cc: seanjc, chao.gao, zhao1.liu
On 11/11/2025 10:17 AM, Paolo Bonzini wrote:
>
> commit fc8aa5c45d558393069a1c89b7a64e059b8f9418
> Author: Chang S. Bae <chang.seok.bae@intel.com>
> Date: Mon Nov 10 18:01:21 2025 +0000
>
> KVM: x86: Refactor REX prefix handling in instruction emulation
> Restructure how to represent and interpret REX fields, preparing
> for handling of VEX and REX2.
> REX uses the upper four bits of a single byte as a fixed identifier,
> and the lower four bits containing the data. VEX and REX2 extend
> this so
> that the first byte identifies the prefix and the rest encode
> additional
> bits; and while VEX only has the same four data bits as REX, eight
> zero
> bits are a valid value for the data bits of REX2. So, stop storing
> the
> REX byte as-is. Instead, store only the low bits of the REX prefix
> and
> track separately whether a REX-like prefix wasused.
> No functional changes intended.
> Signed-off-by: Chang S. Bae <chang.seok.bae@intel.com>
> Message-ID: <20251110180131.28264-11-chang.seok.bae@intel.com>
> [Extracted from APX series; removed bitfields and REX2-specific
> definitions. - Paolo]
> Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Sure, I think it's good to have consistent handling like this for these
extended prefixes across the board. Thanks!
^ permalink raw reply [flat|nested] 55+ messages in thread
* [PATCH RFC v1 11/20] KVM: x86: Refactor opcode table lookup in instruction emulation
2025-11-10 18:01 [PATCH RFC v1 00/20] KVM: x86: Support APX feature for guests Chang S. Bae
` (9 preceding siblings ...)
2025-11-10 18:01 ` [PATCH RFC v1 10/20] KVM: x86: Refactor REX prefix handling in " Chang S. Bae
@ 2025-11-10 18:01 ` Chang S. Bae
2025-11-11 16:55 ` Paolo Bonzini
2025-11-10 18:01 ` [PATCH RFC v1 12/20] KVM: x86: Support REX2-extended register index in the decoder Chang S. Bae
` (10 subsequent siblings)
21 siblings, 1 reply; 55+ messages in thread
From: Chang S. Bae @ 2025-11-10 18:01 UTC (permalink / raw)
To: kvm, linux-kernel; +Cc: pbonzini, seanjc, chao.gao, zhao1.liu, chang.seok.bae
Refactor opcode lookup to clearly separate handling of different byte
sequences and prefix types, in preparation for REX2 support.
The decoder begins with a one-byte opcode table by default and falls
through to other tables on escape bytes, but the logic is intertwined and
hard to extend.
REX2 introduces a dedicated bit in its payload byte to indicate which
opcode table to use. To accommodate this mapping bit, the existing lookup
path needs to be restructured.
No functional changes intended.
Signed-off-by: Chang S. Bae <chang.seok.bae@intel.com>
---
arch/x86/kvm/emulate.c | 19 +++++++++++--------
1 file changed, 11 insertions(+), 8 deletions(-)
diff --git a/arch/x86/kvm/emulate.c b/arch/x86/kvm/emulate.c
index 763fbd139242..9c98843094a1 100644
--- a/arch/x86/kvm/emulate.c
+++ b/arch/x86/kvm/emulate.c
@@ -4773,7 +4773,6 @@ int x86_decode_insn(struct x86_emulate_ctxt *ctxt, void *insn, int insn_len, int
ctxt->_eip = ctxt->eip;
ctxt->fetch.ptr = ctxt->fetch.data;
ctxt->fetch.end = ctxt->fetch.data + insn_len;
- ctxt->opcode_len = 1;
ctxt->intercept = x86_intercept_none;
if (insn_len > 0)
memcpy(ctxt->fetch.data, insn, insn_len);
@@ -4877,20 +4876,24 @@ int x86_decode_insn(struct x86_emulate_ctxt *ctxt, void *insn, int insn_len, int
if (ctxt->rex.bits.w)
ctxt->op_bytes = 8;
- /* Opcode byte(s). */
- opcode = opcode_table[ctxt->b];
- /* Two-byte opcode? */
+ /* Determine opcode byte(s): */
if (ctxt->b == 0x0f) {
- ctxt->opcode_len = 2;
+ /* Escape byte: start two-byte opcode sequence */
ctxt->b = insn_fetch(u8, ctxt);
- opcode = twobyte_table[ctxt->b];
-
- /* 0F_38 opcode map */
if (ctxt->b == 0x38) {
+ /* Three-byte opcode */
ctxt->opcode_len = 3;
ctxt->b = insn_fetch(u8, ctxt);
opcode = opcode_map_0f_38[ctxt->b];
+ } else {
+ /* Two-byte opcode */
+ ctxt->opcode_len = 2;
+ opcode = twobyte_table[ctxt->b];
}
+ } else {
+ /* Single-byte opcode */
+ ctxt->opcode_len = 1;
+ opcode = opcode_table[ctxt->b];
}
ctxt->d = opcode.flags;
--
2.51.0
^ permalink raw reply related [flat|nested] 55+ messages in thread* Re: [PATCH RFC v1 11/20] KVM: x86: Refactor opcode table lookup in instruction emulation
2025-11-10 18:01 ` [PATCH RFC v1 11/20] KVM: x86: Refactor opcode table lookup " Chang S. Bae
@ 2025-11-11 16:55 ` Paolo Bonzini
2025-11-13 23:24 ` Chang S. Bae
0 siblings, 1 reply; 55+ messages in thread
From: Paolo Bonzini @ 2025-11-11 16:55 UTC (permalink / raw)
To: Chang S. Bae, kvm, linux-kernel; +Cc: seanjc, chao.gao, zhao1.liu
On 11/10/25 19:01, Chang S. Bae wrote:
> Refactor opcode lookup to clearly separate handling of different byte
> sequences and prefix types, in preparation for REX2 support.
>
> The decoder begins with a one-byte opcode table by default and falls
> through to other tables on escape bytes, but the logic is intertwined and
> hard to extend.
>
> REX2 introduces a dedicated bit in its payload byte to indicate which
> opcode table to use. To accommodate this mapping bit, the existing lookup
> path needs to be restructured.
>
> No functional changes intended.
>
> Signed-off-by: Chang S. Bae <chang.seok.bae@intel.com>
> ---
> arch/x86/kvm/emulate.c | 19 +++++++++++--------
> 1 file changed, 11 insertions(+), 8 deletions(-)
>
> diff --git a/arch/x86/kvm/emulate.c b/arch/x86/kvm/emulate.c
> index 763fbd139242..9c98843094a1 100644
> --- a/arch/x86/kvm/emulate.c
> +++ b/arch/x86/kvm/emulate.c
> @@ -4773,7 +4773,6 @@ int x86_decode_insn(struct x86_emulate_ctxt *ctxt, void *insn, int insn_len, int
> ctxt->_eip = ctxt->eip;
> ctxt->fetch.ptr = ctxt->fetch.data;
> ctxt->fetch.end = ctxt->fetch.data + insn_len;
> - ctxt->opcode_len = 1;
> ctxt->intercept = x86_intercept_none;
> if (insn_len > 0)
> memcpy(ctxt->fetch.data, insn, insn_len);
> @@ -4877,20 +4876,24 @@ int x86_decode_insn(struct x86_emulate_ctxt *ctxt, void *insn, int insn_len, int
> if (ctxt->rex.bits.w)
> ctxt->op_bytes = 8;
>
> - /* Opcode byte(s). */
> - opcode = opcode_table[ctxt->b];
> - /* Two-byte opcode? */
> + /* Determine opcode byte(s): */
> if (ctxt->b == 0x0f) {
> - ctxt->opcode_len = 2;
> + /* Escape byte: start two-byte opcode sequence */
> ctxt->b = insn_fetch(u8, ctxt);
> - opcode = twobyte_table[ctxt->b];
> -
> - /* 0F_38 opcode map */
> if (ctxt->b == 0x38) {
> + /* Three-byte opcode */
> ctxt->opcode_len = 3;
> ctxt->b = insn_fetch(u8, ctxt);
> opcode = opcode_map_0f_38[ctxt->b];
> + } else {
> + /* Two-byte opcode */
> + ctxt->opcode_len = 2;
> + opcode = twobyte_table[ctxt->b];
> }
> + } else {
> + /* Single-byte opcode */
> + ctxt->opcode_len = 1;
> + opcode = opcode_table[ctxt->b];
> }
> ctxt->d = opcode.flags;
>
This will also conflict with the VEX patches, overall I think all 10-12
patches can be merged in one.
Paolo
^ permalink raw reply [flat|nested] 55+ messages in thread* Re: [PATCH RFC v1 11/20] KVM: x86: Refactor opcode table lookup in instruction emulation
2025-11-11 16:55 ` Paolo Bonzini
@ 2025-11-13 23:24 ` Chang S. Bae
0 siblings, 0 replies; 55+ messages in thread
From: Chang S. Bae @ 2025-11-13 23:24 UTC (permalink / raw)
To: Paolo Bonzini, kvm, linux-kernel; +Cc: seanjc, chao.gao, zhao1.liu
On 11/11/2025 8:55 AM, Paolo Bonzini wrote:
>
> This will also conflict with the VEX patches, overall I think all 10-12
> patches can be merged in one.
I initially split these into micro-patches to make the first round of
review easier. But yes, now that the series is shaping up, folding those
preparatory pieces together makes perfect sense to me.
^ permalink raw reply [flat|nested] 55+ messages in thread
* [PATCH RFC v1 12/20] KVM: x86: Support REX2-extended register index in the decoder
2025-11-10 18:01 [PATCH RFC v1 00/20] KVM: x86: Support APX feature for guests Chang S. Bae
` (10 preceding siblings ...)
2025-11-10 18:01 ` [PATCH RFC v1 11/20] KVM: x86: Refactor opcode table lookup " Chang S. Bae
@ 2025-11-10 18:01 ` Chang S. Bae
2025-11-11 16:53 ` Paolo Bonzini
2025-11-11 16:53 ` Paolo Bonzini
2025-11-10 18:01 ` [PATCH RFC v1 13/20] KVM: x86: Add REX2 opcode tables to the instruction decoder Chang S. Bae
` (9 subsequent siblings)
21 siblings, 2 replies; 55+ messages in thread
From: Chang S. Bae @ 2025-11-10 18:01 UTC (permalink / raw)
To: kvm, linux-kernel; +Cc: pbonzini, seanjc, chao.gao, zhao1.liu, chang.seok.bae
Update register index decoding to account for the additional bit fields
introduced by the REX2 prefix.
Both ModR/M and opcode register decoding paths now consider the extended
index bits (R4, X4, B4) in addition to the legacy REX bits (R3, X3, B3).
Signed-off-by: Chang S. Bae <chang.seok.bae@intel.com>
---
arch/x86/kvm/emulate.c | 12 ++++++++----
1 file changed, 8 insertions(+), 4 deletions(-)
diff --git a/arch/x86/kvm/emulate.c b/arch/x86/kvm/emulate.c
index 9c98843094a1..ed3a8c0bca20 100644
--- a/arch/x86/kvm/emulate.c
+++ b/arch/x86/kvm/emulate.c
@@ -1084,7 +1084,8 @@ static void decode_register_operand(struct x86_emulate_ctxt *ctxt,
reg = ctxt->modrm_reg;
} else {
reg = (ctxt->b & 7) |
- (ctxt->rex.bits.b3 * BIT(3));
+ (ctxt->rex.bits.b3 * BIT(3)) |
+ (ctxt->rex.bits.b4 * BIT(4));
}
if (ctxt->d & Sse) {
@@ -1124,9 +1125,12 @@ static int decode_modrm(struct x86_emulate_ctxt *ctxt,
int rc = X86EMUL_CONTINUE;
ulong modrm_ea = 0;
- ctxt->modrm_reg = ctxt->rex.bits.r3 * BIT(3);
- index_reg = ctxt->rex.bits.x3 * BIT(3);
- base_reg = ctxt->rex.bits.b3 * BIT(3);
+ ctxt->modrm_reg = (ctxt->rex.bits.r3 * BIT(3)) |
+ (ctxt->rex.bits.r4 * BIT(4));
+ index_reg = (ctxt->rex.bits.x3 * BIT(3)) |
+ (ctxt->rex.bits.x4 * BIT(4));
+ base_reg = (ctxt->rex.bits.b3 * BIT(3)) |
+ (ctxt->rex.bits.b4 * BIT(4));
ctxt->modrm_mod = (ctxt->modrm & 0xc0) >> 6;
ctxt->modrm_reg |= (ctxt->modrm & 0x38) >> 3;
--
2.51.0
^ permalink raw reply related [flat|nested] 55+ messages in thread* Re: [PATCH RFC v1 12/20] KVM: x86: Support REX2-extended register index in the decoder
2025-11-10 18:01 ` [PATCH RFC v1 12/20] KVM: x86: Support REX2-extended register index in the decoder Chang S. Bae
@ 2025-11-11 16:53 ` Paolo Bonzini
2025-11-13 23:26 ` Chang S. Bae
2025-11-11 16:53 ` Paolo Bonzini
1 sibling, 1 reply; 55+ messages in thread
From: Paolo Bonzini @ 2025-11-11 16:53 UTC (permalink / raw)
To: Chang S. Bae, kvm, linux-kernel; +Cc: seanjc, chao.gao, zhao1.liu
On 11/10/25 19:01, Chang S. Bae wrote:
> Update register index decoding to account for the additional bit fields
> introduced by the REX2 prefix.
>
> Both ModR/M and opcode register decoding paths now consider the extended
> index bits (R4, X4, B4) in addition to the legacy REX bits (R3, X3, B3).
>
> Signed-off-by: Chang S. Bae <chang.seok.bae@intel.com>
Replying here for both patches 10 and 12, because I think you can merge
them in one, I'd prefer to avoid bitfields.
You only need a single enum:
enum {
REX_B = 1,
REX_X = 2,
REX_R = 4,
REX_W = 8,
REX_M = 0x80,
};
for REX_W/REX_M you access them directly, while for RXB you go through a
function:
static inline rex_get_rxb(u8 rex, u8 fld)
{
BUILD_BUG_ON(!__builtin_constant_p(fld));
BUILD_BUG_ON(fld != REX_B && fld != REX_X && fld != REX_R);
rex >>= ffs(fld) - 1; // bits 0+4
return (rex & 1) + (rex & 0x10 ? 8 : 0);
}
> } else {
> reg = (ctxt->b & 7) |
> - (ctxt->rex.bits.b3 * BIT(3));
> + (ctxt->rex.bits.b3 * BIT(3)) |
> + (ctxt->rex.bits.b4 * BIT(4));
+ rex_get_rxb(ctxt->rex, REX_B);
and likewise everywhere else.
> }
>
> if (ctxt->d & Sse) {
> @@ -1124,9 +1125,12 @@ static int decode_modrm(struct x86_emulate_ctxt *ctxt,
> int rc = X86EMUL_CONTINUE;
> ulong modrm_ea = 0;
>
> - ctxt->modrm_reg = ctxt->rex.bits.r3 * BIT(3);
> - index_reg = ctxt->rex.bits.x3 * BIT(3);
> - base_reg = ctxt->rex.bits.b3 * BIT(3);
> + ctxt->modrm_reg = (ctxt->rex.bits.r3 * BIT(3)) |
> + (ctxt->rex.bits.r4 * BIT(4));
> + index_reg = (ctxt->rex.bits.x3 * BIT(3)) |
> + (ctxt->rex.bits.x4 * BIT(4));
> + base_reg = (ctxt->rex.bits.b3 * BIT(3)) |
> + (ctxt->rex.bits.b4 * BIT(4));
>
> ctxt->modrm_mod = (ctxt->modrm & 0xc0) >> 6;
> ctxt->modrm_reg |= (ctxt->modrm & 0x38) >> 3;
^ permalink raw reply [flat|nested] 55+ messages in thread* Re: [PATCH RFC v1 12/20] KVM: x86: Support REX2-extended register index in the decoder
2025-11-11 16:53 ` Paolo Bonzini
@ 2025-11-13 23:26 ` Chang S. Bae
0 siblings, 0 replies; 55+ messages in thread
From: Chang S. Bae @ 2025-11-13 23:26 UTC (permalink / raw)
To: Paolo Bonzini, kvm, linux-kernel; +Cc: seanjc, chao.gao, zhao1.liu
On 11/11/2025 8:53 AM, Paolo Bonzini wrote:
>
> Replying here for both patches 10 and 12, because I think you can merge
> them in one,
Done
> I'd prefer to avoid bitfields.
>
> You only need a single enum:
>
> enum {
> REX_B = 1,
> REX_X = 2,
> REX_R = 4,
> REX_W = 8,
> REX_M = 0x80,
> };
>
> for REX_W/REX_M you access them directly, while for RXB you go through a
> function:
>
> static inline rex_get_rxb(u8 rex, u8 fld)
> {
> BUILD_BUG_ON(!__builtin_constant_p(fld));
> BUILD_BUG_ON(fld != REX_B && fld != REX_X && fld != REX_R);
>
> rex >>= ffs(fld) - 1; // bits 0+4
> return (rex & 1) + (rex & 0x10 ? 8 : 0);
> }
>
>> } else {
>> reg = (ctxt->b & 7) |
>> - (ctxt->rex.bits.b3 * BIT(3));
>> + (ctxt->rex.bits.b3 * BIT(3)) |
>> + (ctxt->rex.bits.b4 * BIT(4));
>
> + rex_get_rxb(ctxt->rex, REX_B);
>
> and likewise everywhere else.
Neat! This looks much better. Thanks for the suggestion.
^ permalink raw reply [flat|nested] 55+ messages in thread
* Re: [PATCH RFC v1 12/20] KVM: x86: Support REX2-extended register index in the decoder
2025-11-10 18:01 ` [PATCH RFC v1 12/20] KVM: x86: Support REX2-extended register index in the decoder Chang S. Bae
2025-11-11 16:53 ` Paolo Bonzini
@ 2025-11-11 16:53 ` Paolo Bonzini
1 sibling, 0 replies; 55+ messages in thread
From: Paolo Bonzini @ 2025-11-11 16:53 UTC (permalink / raw)
To: Chang S. Bae, kvm, linux-kernel; +Cc: seanjc, chao.gao, zhao1.liu
On 11/10/25 19:01, Chang S. Bae wrote:
> Update register index decoding to account for the additional bit fields
> introduced by the REX2 prefix.
>
> Both ModR/M and opcode register decoding paths now consider the extended
> index bits (R4, X4, B4) in addition to the legacy REX bits (R3, X3, B3).
>
> Signed-off-by: Chang S. Bae <chang.seok.bae@intel.com>
Replying here for both patches 10 and 12, because I think you can merge
them in one, I'd prefer to avoid bitfields.
You only need a single enum:
enum {
REX_B = 1,
REX_X = 2,
REX_R = 4,
REX_W = 8,
REX_M = 0x80,
};
for REX_W/REX_M you access them directly, while for RXB you go through a
function:
static inline rex_get_rxb(u8 rex, u8 fld)
{
BUILD_BUG_ON(!__builtin_constant_p(fld));
BUILD_BUG_ON(fld != REX_B && fld != REX_X && fld != REX_R);
rex >>= ffs(fld) - 1; // bits 0+4
return (rex & 1) + (rex & 0x10 ? 8 : 0);
}
> } else {
> reg = (ctxt->b & 7) |
> - (ctxt->rex.bits.b3 * BIT(3));
> + (ctxt->rex.bits.b3 * BIT(3)) |
> + (ctxt->rex.bits.b4 * BIT(4));
+ rex_get_rxb(ctxt->rex, REX_B);
and likewise everywhere else.
> }
>
> if (ctxt->d & Sse) {
> @@ -1124,9 +1125,12 @@ static int decode_modrm(struct x86_emulate_ctxt *ctxt,
> int rc = X86EMUL_CONTINUE;
> ulong modrm_ea = 0;
>
> - ctxt->modrm_reg = ctxt->rex.bits.r3 * BIT(3);
> - index_reg = ctxt->rex.bits.x3 * BIT(3);
> - base_reg = ctxt->rex.bits.b3 * BIT(3);
> + ctxt->modrm_reg = (ctxt->rex.bits.r3 * BIT(3)) |
> + (ctxt->rex.bits.r4 * BIT(4));
> + index_reg = (ctxt->rex.bits.x3 * BIT(3)) |
> + (ctxt->rex.bits.x4 * BIT(4));
> + base_reg = (ctxt->rex.bits.b3 * BIT(3)) |
> + (ctxt->rex.bits.b4 * BIT(4));
>
> ctxt->modrm_mod = (ctxt->modrm & 0xc0) >> 6;
> ctxt->modrm_reg |= (ctxt->modrm & 0x38) >> 3;
^ permalink raw reply [flat|nested] 55+ messages in thread
* [PATCH RFC v1 13/20] KVM: x86: Add REX2 opcode tables to the instruction decoder
2025-11-10 18:01 [PATCH RFC v1 00/20] KVM: x86: Support APX feature for guests Chang S. Bae
` (11 preceding siblings ...)
2025-11-10 18:01 ` [PATCH RFC v1 12/20] KVM: x86: Support REX2-extended register index in the decoder Chang S. Bae
@ 2025-11-10 18:01 ` Chang S. Bae
2025-11-10 18:01 ` [PATCH RFC v1 14/20] KVM: x86: Emulate REX2-prefixed 64-bit absolute jump Chang S. Bae
` (8 subsequent siblings)
21 siblings, 0 replies; 55+ messages in thread
From: Chang S. Bae @ 2025-11-10 18:01 UTC (permalink / raw)
To: kvm, linux-kernel; +Cc: pbonzini, seanjc, chao.gao, zhao1.liu, chang.seok.bae
Extend the decoder to find REX2-prefixed opcodes by introducing dedicated
REX2 opcode tables. During initialization, clone the legacy opcode tables
and patch entries that differ under REX2.
Although most REX2-prefixed opcodes follow the legacy tables, some differ
for instructions that do not reference extended register bits or are
newly introduced under REX2. Using separate tables simplifies the
lookup logic and allows efficient patching of exceptions.
The EGPR checker will be implemented later.
Signed-off-by: Chang S. Bae <chang.seok.bae@intel.com>
---
RFC note:
The lookup logic could be separated from the table population, but
keeping the user of the tables close to their initialization helps
clarify the purpose of the new table. If this becomes hard to follow,
splitting the lookup separately can be an option.
---
arch/x86/kvm/emulate.c | 73 +++++++++++++++++++++++++++++++++++++-
arch/x86/kvm/kvm_emulate.h | 2 ++
arch/x86/kvm/x86.c | 1 +
3 files changed, 75 insertions(+), 1 deletion(-)
diff --git a/arch/x86/kvm/emulate.c b/arch/x86/kvm/emulate.c
index ed3a8c0bca20..58879a31abcd 100644
--- a/arch/x86/kvm/emulate.c
+++ b/arch/x86/kvm/emulate.c
@@ -4475,6 +4475,19 @@ static const struct opcode opcode_map_0f_38[256] = {
N, N, X4(N), X8(N)
};
+/*
+ * REX2 opcode tables.
+ *
+ * REX2-prefixed opcodes mostly follow the legacy tables but differ slightly
+ * for instructions that do not use R/X/B register bits. Initialize the REX2
+ * tables by copying the legacy ones, then mark mismatched rows as undefined.
+ */
+static struct opcode rex2_opcode_table[256] __ro_after_init;
+static struct opcode rex2_twobyte_table[256] __ro_after_init;
+
+static const struct opcode undefined = D(Undefined);
+static const struct opcode notimpl = N;
+
#undef D
#undef N
#undef G
@@ -4761,6 +4774,11 @@ static int decode_operand(struct x86_emulate_ctxt *ctxt, struct operand *op,
return rc;
}
+static inline bool emul_egpr_enabled(struct x86_emulate_ctxt *ctxt __maybe_unused)
+{
+ return false;
+}
+
int x86_decode_insn(struct x86_emulate_ctxt *ctxt, void *insn, int insn_len, int emulation_type)
{
int rc = X86EMUL_CONTINUE;
@@ -4881,7 +4899,24 @@ int x86_decode_insn(struct x86_emulate_ctxt *ctxt, void *insn, int insn_len, int
ctxt->op_bytes = 8;
/* Determine opcode byte(s): */
- if (ctxt->b == 0x0f) {
+ if (ctxt->rex_prefix == REX2_INVALID) {
+ /*
+ * A REX2 prefix was detected, but the prefix decoder
+ * found invalid byte sequence.
+ */
+ opcode = undefined;
+ } else if (ctxt->rex_prefix == REX2_PREFIX) {
+ /* REX2 prefix is only valid when EGPRs are enabled. */
+ if (!emul_egpr_enabled(ctxt)) {
+ opcode = undefined;
+ } else if (ctxt->rex.bits.m0) {
+ ctxt->opcode_len = 2;
+ opcode = rex2_twobyte_table[ctxt->b];
+ } else {
+ ctxt->opcode_len = 1;
+ opcode = rex2_opcode_table[ctxt->b];
+ }
+ } else if (ctxt->b == 0x0f) {
/* Escape byte: start two-byte opcode sequence */
ctxt->b = insn_fetch(u8, ctxt);
if (ctxt->b == 0x38) {
@@ -5526,3 +5561,39 @@ bool emulator_can_use_gpa(struct x86_emulate_ctxt *ctxt)
return true;
}
+
+static void undefine_row(struct opcode *row)
+{
+ struct opcode *ptr = row;
+ int i;
+
+ /* Clear 16 entries per row */
+ for (i = 0; i < 0x10; i++, ptr++)
+ *ptr = undefined;
+}
+
+/*
+ * Populate REX2 opcode table:
+ *
+ * REX2-prefixed opcodes mostly reuse the legacy layout, except for those that
+ * neither reference extended register bits nor are newly introduced under the
+ * REX2 prefix. Initialize both single- and two-byte tables by cloning the
+ * legacy versions, then patch the table for some exceptions.
+ */
+void __init kvm_init_rex2_opcode_table(void)
+{
+ /* Copy legacy tables: */
+ memcpy(rex2_opcode_table, opcode_table, sizeof(opcode_table));
+ memcpy(rex2_twobyte_table, twobyte_table, sizeof(twobyte_table));
+
+ /* Undefine reserved opcode ranges: */
+ undefine_row(&rex2_opcode_table[0x40]);
+ undefine_row(&rex2_opcode_table[0x70]);
+ undefine_row(&rex2_opcode_table[0xa0]);
+ undefine_row(&rex2_opcode_table[0xe0]);
+ undefine_row(&rex2_twobyte_table[0x30]);
+ undefine_row(&rex2_twobyte_table[0x80]);
+
+ /* Mark opcode not yet implemented: */
+ rex2_opcode_table[0xa1] = notimpl;
+}
diff --git a/arch/x86/kvm/kvm_emulate.h b/arch/x86/kvm/kvm_emulate.h
index b285299ebfa4..cc16211d61f6 100644
--- a/arch/x86/kvm/kvm_emulate.h
+++ b/arch/x86/kvm/kvm_emulate.h
@@ -589,4 +589,6 @@ static inline ulong *reg_rmw(struct x86_emulate_ctxt *ctxt, unsigned nr)
return reg_write(ctxt, nr);
}
+void __init kvm_init_rex2_opcode_table(void);
+
#endif /* _ASM_X86_KVM_X86_EMULATE_H */
diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
index 338986a5a3ae..4c8c2fc3bda6 100644
--- a/arch/x86/kvm/x86.c
+++ b/arch/x86/kvm/x86.c
@@ -14354,6 +14354,7 @@ EXPORT_TRACEPOINT_SYMBOL_GPL(kvm_rmp_fault);
static int __init kvm_x86_init(void)
{
kvm_init_xstate_sizes();
+ kvm_init_rex2_opcode_table();
kvm_mmu_x86_module_init();
mitigate_smt_rsb &= boot_cpu_has_bug(X86_BUG_SMT_RSB) && cpu_smt_possible();
--
2.51.0
^ permalink raw reply related [flat|nested] 55+ messages in thread* [PATCH RFC v1 14/20] KVM: x86: Emulate REX2-prefixed 64-bit absolute jump
2025-11-10 18:01 [PATCH RFC v1 00/20] KVM: x86: Support APX feature for guests Chang S. Bae
` (12 preceding siblings ...)
2025-11-10 18:01 ` [PATCH RFC v1 13/20] KVM: x86: Add REX2 opcode tables to the instruction decoder Chang S. Bae
@ 2025-11-10 18:01 ` Chang S. Bae
2025-11-11 16:39 ` Paolo Bonzini
2025-11-10 18:01 ` [PATCH RFC v1 15/20] KVM: x86: Reject EVEX-prefix instructions in the emulator Chang S. Bae
` (7 subsequent siblings)
21 siblings, 1 reply; 55+ messages in thread
From: Chang S. Bae @ 2025-11-10 18:01 UTC (permalink / raw)
To: kvm, linux-kernel; +Cc: pbonzini, seanjc, chao.gao, zhao1.liu, chang.seok.bae
Add support for the new absolute jump, previously unimplemented.
This instruction has an unusual quirk: the REX2.W bit uses inverted
polarity. Unlike normal REX or REX2 semantics (where W=1 indicates a
64-bit operand size), this instruction uses W=0 to select an 8-byte
operand size.
The new InvertedWidthPolarity flag and its helper to interpret the
W bit correctly, avoiding special-case hacks in the emulator logic.
Since the ctxt->op_bytes depends on the instruction flags, the size
should be determined after the instruction lookup.
Signed-off-by: Chang S. Bae <chang.seok.bae@intel.com>
---
arch/x86/kvm/emulate.c | 27 ++++++++++++++++++++-------
1 file changed, 20 insertions(+), 7 deletions(-)
diff --git a/arch/x86/kvm/emulate.c b/arch/x86/kvm/emulate.c
index 58879a31abcd..03f8e007b14e 100644
--- a/arch/x86/kvm/emulate.c
+++ b/arch/x86/kvm/emulate.c
@@ -179,6 +179,7 @@
#define TwoMemOp ((u64)1 << 55) /* Instruction has two memory operand */
#define IsBranch ((u64)1 << 56) /* Instruction is considered a branch. */
#define ShadowStack ((u64)1 << 57) /* Instruction affects Shadow Stacks. */
+#define InvertedWidthPolarity ((u64)1 << 58) /* Instruction uses inverted REX2.W polarity */
#define DstXacc (DstAccLo | SrcAccHi | SrcWrite)
@@ -993,6 +994,16 @@ EM_ASM_2W(btc);
EM_ASM_2R(cmp, cmp_r);
+static inline bool is_64bit_operand_size(struct x86_emulate_ctxt *ctxt)
+{
+ /*
+ * Most instructions interpret REX.W=1 as 64-bit operand size.
+ * Some REX2 opcodes invert this logic.
+ */
+ return ctxt->d & InvertedWidthPolarity ?
+ ctxt->rex.bits.w == 0 : ctxt->rex.bits.w == 1;
+}
+
static int em_bsf_c(struct x86_emulate_ctxt *ctxt)
{
/* If src is zero, do not writeback, but update flags */
@@ -2472,7 +2483,7 @@ static int em_sysexit(struct x86_emulate_ctxt *ctxt)
setup_syscalls_segments(&cs, &ss);
- if (ctxt->rex.bits.w)
+ if (is_64bit_operand_size(ctxt))
usermode = X86EMUL_MODE_PROT64;
else
usermode = X86EMUL_MODE_PROT32;
@@ -4486,7 +4497,8 @@ static struct opcode rex2_opcode_table[256] __ro_after_init;
static struct opcode rex2_twobyte_table[256] __ro_after_init;
static const struct opcode undefined = D(Undefined);
-static const struct opcode notimpl = N;
+static const struct opcode pfx_d5_a1 = I(SrcImm64 | NearBranch | IsBranch | InvertedWidthPolarity, \
+ em_jmp_abs);
#undef D
#undef N
@@ -4543,6 +4555,7 @@ static bool is_ibt_instruction(struct x86_emulate_ctxt *ctxt)
return true;
case SrcNone:
case SrcImm:
+ case SrcImm64:
case SrcImmByte:
/*
* Note, ImmU16 is used only for the stack adjustment operand on ENTER
@@ -4895,9 +4908,6 @@ int x86_decode_insn(struct x86_emulate_ctxt *ctxt, void *insn, int insn_len, int
done_prefixes:
- if (ctxt->rex.bits.w)
- ctxt->op_bytes = 8;
-
/* Determine opcode byte(s): */
if (ctxt->rex_prefix == REX2_INVALID) {
/*
@@ -4936,6 +4946,9 @@ int x86_decode_insn(struct x86_emulate_ctxt *ctxt, void *insn, int insn_len, int
}
ctxt->d = opcode.flags;
+ if (is_64bit_operand_size(ctxt))
+ ctxt->op_bytes = 8;
+
if (ctxt->d & ModRM)
ctxt->modrm = insn_fetch(u8, ctxt);
@@ -5594,6 +5607,6 @@ void __init kvm_init_rex2_opcode_table(void)
undefine_row(&rex2_twobyte_table[0x30]);
undefine_row(&rex2_twobyte_table[0x80]);
- /* Mark opcode not yet implemented: */
- rex2_opcode_table[0xa1] = notimpl;
+ /* Define the REX2-specific absolute jump (0xA1) opcode */
+ rex2_opcode_table[0xa1] = pfx_d5_a1;
}
--
2.51.0
^ permalink raw reply related [flat|nested] 55+ messages in thread* Re: [PATCH RFC v1 14/20] KVM: x86: Emulate REX2-prefixed 64-bit absolute jump
2025-11-10 18:01 ` [PATCH RFC v1 14/20] KVM: x86: Emulate REX2-prefixed 64-bit absolute jump Chang S. Bae
@ 2025-11-11 16:39 ` Paolo Bonzini
2025-11-13 23:27 ` Chang S. Bae
0 siblings, 1 reply; 55+ messages in thread
From: Paolo Bonzini @ 2025-11-11 16:39 UTC (permalink / raw)
To: Chang S. Bae, kvm, linux-kernel; +Cc: seanjc, chao.gao, zhao1.liu
On 11/10/25 19:01, Chang S. Bae wrote:
> Add support for the new absolute jump, previously unimplemented.
>
> This instruction has an unusual quirk: the REX2.W bit uses inverted
> polarity. Unlike normal REX or REX2 semantics (where W=1 indicates a
> 64-bit operand size), this instruction uses W=0 to select an 8-byte
> operand size.
>
> The new InvertedWidthPolarity flag and its helper to interpret the
> W bit correctly, avoiding special-case hacks in the emulator logic.
>
> Since the ctxt->op_bytes depends on the instruction flags, the size
> should be determined after the instruction lookup.
>
> Signed-off-by: Chang S. Bae <chang.seok.bae@intel.com>
I think this is not needed. Emulation of non-memory operations, in
practice, is only needed to support big real mode on very old processors.
We can just add a NoRex bit and apply it to the six reows you touch in
patch 13.
Paolo
> ---
> arch/x86/kvm/emulate.c | 27 ++++++++++++++++++++-------
> 1 file changed, 20 insertions(+), 7 deletions(-)
>
> diff --git a/arch/x86/kvm/emulate.c b/arch/x86/kvm/emulate.c
> index 58879a31abcd..03f8e007b14e 100644
> --- a/arch/x86/kvm/emulate.c
> +++ b/arch/x86/kvm/emulate.c
> @@ -179,6 +179,7 @@
> #define TwoMemOp ((u64)1 << 55) /* Instruction has two memory operand */
> #define IsBranch ((u64)1 << 56) /* Instruction is considered a branch. */
> #define ShadowStack ((u64)1 << 57) /* Instruction affects Shadow Stacks. */
> +#define InvertedWidthPolarity ((u64)1 << 58) /* Instruction uses inverted REX2.W polarity */
>
> #define DstXacc (DstAccLo | SrcAccHi | SrcWrite)
>
> @@ -993,6 +994,16 @@ EM_ASM_2W(btc);
>
> EM_ASM_2R(cmp, cmp_r);
>
> +static inline bool is_64bit_operand_size(struct x86_emulate_ctxt *ctxt)
> +{
> + /*
> + * Most instructions interpret REX.W=1 as 64-bit operand size.
> + * Some REX2 opcodes invert this logic.
> + */
> + return ctxt->d & InvertedWidthPolarity ?
> + ctxt->rex.bits.w == 0 : ctxt->rex.bits.w == 1;
> +}
> +
> static int em_bsf_c(struct x86_emulate_ctxt *ctxt)
> {
> /* If src is zero, do not writeback, but update flags */
> @@ -2472,7 +2483,7 @@ static int em_sysexit(struct x86_emulate_ctxt *ctxt)
>
> setup_syscalls_segments(&cs, &ss);
>
> - if (ctxt->rex.bits.w)
> + if (is_64bit_operand_size(ctxt))
> usermode = X86EMUL_MODE_PROT64;
> else
> usermode = X86EMUL_MODE_PROT32;
> @@ -4486,7 +4497,8 @@ static struct opcode rex2_opcode_table[256] __ro_after_init;
> static struct opcode rex2_twobyte_table[256] __ro_after_init;
>
> static const struct opcode undefined = D(Undefined);
> -static const struct opcode notimpl = N;
> +static const struct opcode pfx_d5_a1 = I(SrcImm64 | NearBranch | IsBranch | InvertedWidthPolarity, \
> + em_jmp_abs);
>
> #undef D
> #undef N
> @@ -4543,6 +4555,7 @@ static bool is_ibt_instruction(struct x86_emulate_ctxt *ctxt)
> return true;
> case SrcNone:
> case SrcImm:
> + case SrcImm64:
> case SrcImmByte:
> /*
> * Note, ImmU16 is used only for the stack adjustment operand on ENTER
> @@ -4895,9 +4908,6 @@ int x86_decode_insn(struct x86_emulate_ctxt *ctxt, void *insn, int insn_len, int
>
> done_prefixes:
>
> - if (ctxt->rex.bits.w)
> - ctxt->op_bytes = 8;
> -
> /* Determine opcode byte(s): */
> if (ctxt->rex_prefix == REX2_INVALID) {
> /*
> @@ -4936,6 +4946,9 @@ int x86_decode_insn(struct x86_emulate_ctxt *ctxt, void *insn, int insn_len, int
> }
> ctxt->d = opcode.flags;
>
> + if (is_64bit_operand_size(ctxt))
> + ctxt->op_bytes = 8;
> +
> if (ctxt->d & ModRM)
> ctxt->modrm = insn_fetch(u8, ctxt);
>
> @@ -5594,6 +5607,6 @@ void __init kvm_init_rex2_opcode_table(void)
> undefine_row(&rex2_twobyte_table[0x30]);
> undefine_row(&rex2_twobyte_table[0x80]);
>
> - /* Mark opcode not yet implemented: */
> - rex2_opcode_table[0xa1] = notimpl;
> + /* Define the REX2-specific absolute jump (0xA1) opcode */
> + rex2_opcode_table[0xa1] = pfx_d5_a1;
> }
^ permalink raw reply [flat|nested] 55+ messages in thread* Re: [PATCH RFC v1 14/20] KVM: x86: Emulate REX2-prefixed 64-bit absolute jump
2025-11-11 16:39 ` Paolo Bonzini
@ 2025-11-13 23:27 ` Chang S. Bae
0 siblings, 0 replies; 55+ messages in thread
From: Chang S. Bae @ 2025-11-13 23:27 UTC (permalink / raw)
To: Paolo Bonzini, kvm, linux-kernel; +Cc: seanjc, chao.gao, zhao1.liu
[-- Attachment #1: Type: text/plain, Size: 211 bytes --]
On 11/11/2025 8:39 AM, Paolo Bonzini wrote:
>
> We can just add a NoRex bit and apply it to the six reows you touch in
> patch 13.
Yeah, I think that is much simple and better. Attached is the diff for
this:
[-- Attachment #2: PATCH14.diff --]
[-- Type: text/plain, Size: 6022 bytes --]
diff --git a/arch/x86/kvm/emulate.c b/arch/x86/kvm/emulate.c
index b2490e56cb71..5f36dbcec484 100644
--- a/arch/x86/kvm/emulate.c
+++ b/arch/x86/kvm/emulate.c
@@ -179,6 +179,7 @@
#define TwoMemOp ((u64)1 << 55) /* Instruction has two memory operand */
#define IsBranch ((u64)1 << 56) /* Instruction is considered a branch. */
#define ShadowStack ((u64)1 << 57) /* Instruction affects Shadow Stacks. */
+#define NoRex2 ((u64)1 << 58) /* Instruction has no use of REX2 prefix */
#define DstXacc (DstAccLo | SrcAccHi | SrcWrite)
@@ -4247,7 +4248,7 @@ static const struct opcode opcode_table[256] = {
/* 0x38 - 0x3F */
I6ALU(NoWrite, em_cmp), N, N,
/* 0x40 - 0x4F */
- X8(I(DstReg, em_inc)), X8(I(DstReg, em_dec)),
+ X8(I(DstReg | NoRex2, em_inc)), X8(I(DstReg | NoRex2, em_dec)),
/* 0x50 - 0x57 */
X8(I(SrcReg | Stack, em_push)),
/* 0x58 - 0x5F */
@@ -4265,7 +4266,7 @@ static const struct opcode opcode_table[256] = {
I2bvIP(DstDI | SrcDX | Mov | String | Unaligned, em_in, ins, check_perm_in), /* insb, insw/insd */
I2bvIP(SrcSI | DstDX | String, em_out, outs, check_perm_out), /* outsb, outsw/outsd */
/* 0x70 - 0x7F */
- X16(D(SrcImmByte | NearBranch | IsBranch)),
+ X16(D(SrcImmByte | NearBranch | IsBranch | NoRex2)),
/* 0x80 - 0x87 */
G(ByteOp | DstMem | SrcImm, group1),
G(DstMem | SrcImm, group1),
@@ -4289,15 +4290,15 @@ static const struct opcode opcode_table[256] = {
II(ImplicitOps | Stack, em_popf, popf),
I(ImplicitOps, em_sahf), I(ImplicitOps, em_lahf),
/* 0xA0 - 0xA7 */
- I2bv(DstAcc | SrcMem | Mov | MemAbs, em_mov),
- I2bv(DstMem | SrcAcc | Mov | MemAbs | PageTable, em_mov),
- I2bv(SrcSI | DstDI | Mov | String | TwoMemOp, em_mov),
- I2bv(SrcSI | DstDI | String | NoWrite | TwoMemOp, em_cmp_r),
+ I2bv(DstAcc | SrcMem | Mov | MemAbs | NoRex2, em_mov),
+ I2bv(DstMem | SrcAcc | Mov | MemAbs | PageTable | NoRex2, em_mov),
+ I2bv(SrcSI | DstDI | Mov | String | TwoMemOp | NoRex2, em_mov),
+ I2bv(SrcSI | DstDI | String | NoWrite | TwoMemOp | NoRex2, em_cmp_r),
/* 0xA8 - 0xAF */
- I2bv(DstAcc | SrcImm | NoWrite, em_test),
- I2bv(SrcAcc | DstDI | Mov | String, em_mov),
- I2bv(SrcSI | DstAcc | Mov | String, em_mov),
- I2bv(SrcAcc | DstDI | String | NoWrite, em_cmp_r),
+ I2bv(DstAcc | SrcImm | NoWrite | NoRex2, em_test),
+ I2bv(SrcAcc | DstDI | Mov | String | NoRex2, em_mov),
+ I2bv(SrcSI | DstAcc | Mov | String | NoRex2, em_mov),
+ I2bv(SrcAcc | DstDI | String | NoWrite | NoRex2, em_cmp_r),
/* 0xB0 - 0xB7 */
X8(I(ByteOp | DstReg | SrcImm | Mov, em_mov)),
/* 0xB8 - 0xBF */
@@ -4327,17 +4328,17 @@ static const struct opcode opcode_table[256] = {
/* 0xD8 - 0xDF */
N, E(0, &escape_d9), N, E(0, &escape_db), N, E(0, &escape_dd), N, N,
/* 0xE0 - 0xE7 */
- X3(I(SrcImmByte | NearBranch | IsBranch, em_loop)),
- I(SrcImmByte | NearBranch | IsBranch, em_jcxz),
- I2bvIP(SrcImmUByte | DstAcc, em_in, in, check_perm_in),
- I2bvIP(SrcAcc | DstImmUByte, em_out, out, check_perm_out),
+ X3(I(SrcImmByte | NearBranch | IsBranch | NoRex2, em_loop)),
+ I(SrcImmByte | NearBranch | IsBranch | NoRex2, em_jcxz),
+ I2bvIP(SrcImmUByte | DstAcc | NoRex2, em_in, in, check_perm_in),
+ I2bvIP(SrcAcc | DstImmUByte | NoRex2, em_out, out, check_perm_out),
/* 0xE8 - 0xEF */
- I(SrcImm | NearBranch | IsBranch | ShadowStack, em_call),
- D(SrcImm | ImplicitOps | NearBranch | IsBranch),
- I(SrcImmFAddr | No64 | IsBranch, em_jmp_far),
- D(SrcImmByte | ImplicitOps | NearBranch | IsBranch),
- I2bvIP(SrcDX | DstAcc, em_in, in, check_perm_in),
- I2bvIP(SrcAcc | DstDX, em_out, out, check_perm_out),
+ I(SrcImm | NearBranch | IsBranch | ShadowStack | NoRex2, em_call),
+ D(SrcImm | ImplicitOps | NearBranch | IsBranch | NoRex2),
+ I(SrcImmFAddr | No64 | IsBranch | NoRex2, em_jmp_far),
+ D(SrcImmByte | ImplicitOps | NearBranch | IsBranch | NoRex2),
+ I2bvIP(SrcDX | DstAcc | NoRex2, em_in, in, check_perm_in),
+ I2bvIP(SrcAcc | DstDX | NoRex2, em_out, out, check_perm_out),
/* 0xF0 - 0xF7 */
N, DI(ImplicitOps, icebp), N, N,
DI(ImplicitOps | Priv, hlt), D(ImplicitOps),
@@ -4378,12 +4379,12 @@ static const struct opcode twobyte_table[256] = {
N, GP(ModRM | DstMem | SrcReg | Mov | Sse, &pfx_0f_2b),
N, N, N, N,
/* 0x30 - 0x3F */
- II(ImplicitOps | Priv, em_wrmsr, wrmsr),
- IIP(ImplicitOps, em_rdtsc, rdtsc, check_rdtsc),
- II(ImplicitOps | Priv, em_rdmsr, rdmsr),
- IIP(ImplicitOps, em_rdpmc, rdpmc, check_rdpmc),
- I(ImplicitOps | EmulateOnUD | IsBranch | ShadowStack, em_sysenter),
- I(ImplicitOps | Priv | EmulateOnUD | IsBranch | ShadowStack, em_sysexit),
+ II(ImplicitOps | Priv | NoRex2, em_wrmsr, wrmsr),
+ IIP(ImplicitOps | NoRex2, em_rdtsc, rdtsc, check_rdtsc),
+ II(ImplicitOps | Priv | NoRex2, em_rdmsr, rdmsr),
+ IIP(ImplicitOps | NoRex2, em_rdpmc, rdpmc, check_rdpmc),
+ I(ImplicitOps | EmulateOnUD | IsBranch | ShadowStack | NoRex2, em_sysenter),
+ I(ImplicitOps | Priv | EmulateOnUD | IsBranch | ShadowStack | NoRex2, em_sysexit),
N, N,
N, N, N, N, N, N, N, N,
/* 0x40 - 0x4F */
@@ -4401,7 +4402,7 @@ static const struct opcode twobyte_table[256] = {
N, N, N, N,
N, N, N, GP(SrcReg | DstMem | ModRM | Mov, &pfx_0f_6f_0f_7f),
/* 0x80 - 0x8F */
- X16(D(SrcImm | NearBranch | IsBranch)),
+ X16(D(SrcImm | NearBranch | IsBranch | NoRex2)),
/* 0x90 - 0x9F */
X16(D(ByteOp | DstMem | SrcNone | ModRM| Mov)),
/* 0xA0 - 0xA7 */
@@ -4888,7 +4889,7 @@ int x86_decode_insn(struct x86_emulate_ctxt *ctxt, void *insn, int insn_len, int
if (ctxt->b == 0x0f) {
/* Escape byte: start two-byte opcode sequence */
ctxt->b = insn_fetch(u8, ctxt);
- if (ctxt->b == 0x38) {
+ if (ctxt->b == 0x38 && ctxt->rex_prefix != REX2_PREFIX) {
/* Three-byte opcode */
ctxt->opcode_len = 3;
ctxt->b = insn_fetch(u8, ctxt);
@@ -4905,6 +4906,9 @@ int x86_decode_insn(struct x86_emulate_ctxt *ctxt, void *insn, int insn_len, int
}
ctxt->d = opcode.flags;
+ if (ctxt->d & NoRex2 && ctxt->rex_prefix == REX2_PREFIX)
+ ctxt->d = Undefined;
+
if (ctxt->d & ModRM)
ctxt->modrm = insn_fetch(u8, ctxt);
^ permalink raw reply related [flat|nested] 55+ messages in thread
* [PATCH RFC v1 15/20] KVM: x86: Reject EVEX-prefix instructions in the emulator
2025-11-10 18:01 [PATCH RFC v1 00/20] KVM: x86: Support APX feature for guests Chang S. Bae
` (13 preceding siblings ...)
2025-11-10 18:01 ` [PATCH RFC v1 14/20] KVM: x86: Emulate REX2-prefixed 64-bit absolute jump Chang S. Bae
@ 2025-11-10 18:01 ` Chang S. Bae
2025-11-11 16:37 ` Paolo Bonzini
2025-11-10 18:01 ` [PATCH RFC v1 16/20] KVM: x86: Decode REX2 prefix " Chang S. Bae
` (6 subsequent siblings)
21 siblings, 1 reply; 55+ messages in thread
From: Chang S. Bae @ 2025-11-10 18:01 UTC (permalink / raw)
To: kvm, linux-kernel; +Cc: pbonzini, seanjc, chao.gao, zhao1.liu, chang.seok.bae
Explicitly mark EVEX-prefixed opcodes (0x62) as unsupported, clarifying
current decoding behavior.
While new prefixes like REX2 extend GPR handling, EVEX emulation should
be addressed separately once after VEX support is implemented.
Signed-off-by: Chang S. Bae <chang.seok.bae@intel.com>
---
arch/x86/kvm/emulate.c | 4 ++--
1 file changed, 2 insertions(+), 2 deletions(-)
diff --git a/arch/x86/kvm/emulate.c b/arch/x86/kvm/emulate.c
index 03f8e007b14e..9bd61ea496e5 100644
--- a/arch/x86/kvm/emulate.c
+++ b/arch/x86/kvm/emulate.c
@@ -4952,8 +4952,8 @@ int x86_decode_insn(struct x86_emulate_ctxt *ctxt, void *insn, int insn_len, int
if (ctxt->d & ModRM)
ctxt->modrm = insn_fetch(u8, ctxt);
- /* vex-prefix instructions are not implemented */
- if (ctxt->opcode_len == 1 && (ctxt->b == 0xc5 || ctxt->b == 0xc4) &&
+ /* VEX and EVEX-prefixed instructions are not implemented */
+ if (ctxt->opcode_len == 1 && (ctxt->b == 0xc5 || ctxt->b == 0xc4 || ctxt->b == 0x62) &&
(mode == X86EMUL_MODE_PROT64 || (ctxt->modrm & 0xc0) == 0xc0)) {
ctxt->d = NotImpl;
}
--
2.51.0
^ permalink raw reply related [flat|nested] 55+ messages in thread* Re: [PATCH RFC v1 15/20] KVM: x86: Reject EVEX-prefix instructions in the emulator
2025-11-10 18:01 ` [PATCH RFC v1 15/20] KVM: x86: Reject EVEX-prefix instructions in the emulator Chang S. Bae
@ 2025-11-11 16:37 ` Paolo Bonzini
2025-11-13 23:28 ` Chang S. Bae
0 siblings, 1 reply; 55+ messages in thread
From: Paolo Bonzini @ 2025-11-11 16:37 UTC (permalink / raw)
To: Chang S. Bae, kvm, linux-kernel; +Cc: seanjc, chao.gao, zhao1.liu
On 11/10/25 19:01, Chang S. Bae wrote:
> Explicitly mark EVEX-prefixed opcodes (0x62) as unsupported, clarifying
> current decoding behavior.
>
> While new prefixes like REX2 extend GPR handling, EVEX emulation should
> be addressed separately once after VEX support is implemented.
>
> Signed-off-by: Chang S. Bae <chang.seok.bae@intel.com>
> ---
> arch/x86/kvm/emulate.c | 4 ++--
> 1 file changed, 2 insertions(+), 2 deletions(-)
>
> diff --git a/arch/x86/kvm/emulate.c b/arch/x86/kvm/emulate.c
> index 03f8e007b14e..9bd61ea496e5 100644
> --- a/arch/x86/kvm/emulate.c
> +++ b/arch/x86/kvm/emulate.c
> @@ -4952,8 +4952,8 @@ int x86_decode_insn(struct x86_emulate_ctxt *ctxt, void *insn, int insn_len, int
> if (ctxt->d & ModRM)
> ctxt->modrm = insn_fetch(u8, ctxt);
>
> - /* vex-prefix instructions are not implemented */
> - if (ctxt->opcode_len == 1 && (ctxt->b == 0xc5 || ctxt->b == 0xc4) &&
> + /* VEX and EVEX-prefixed instructions are not implemented */
> + if (ctxt->opcode_len == 1 && (ctxt->b == 0xc5 || ctxt->b == 0xc4 || ctxt->b == 0x62) &&
> (mode == X86EMUL_MODE_PROT64 || (ctxt->modrm & 0xc0) == 0xc0)) {
> ctxt->d = NotImpl;
> }
VEX support is coming (will post tomorrow I think) so the patches around
decode are going to need changes, but nothing major.
Paolo
^ permalink raw reply [flat|nested] 55+ messages in thread
* [PATCH RFC v1 16/20] KVM: x86: Decode REX2 prefix in the emulator
2025-11-10 18:01 [PATCH RFC v1 00/20] KVM: x86: Support APX feature for guests Chang S. Bae
` (14 preceding siblings ...)
2025-11-10 18:01 ` [PATCH RFC v1 15/20] KVM: x86: Reject EVEX-prefix instructions in the emulator Chang S. Bae
@ 2025-11-10 18:01 ` Chang S. Bae
2025-11-11 17:55 ` Paolo Bonzini
2025-11-10 18:01 ` [PATCH RFC v1 17/20] KVM: x86: Prepare APX state setting in XCR0 Chang S. Bae
` (5 subsequent siblings)
21 siblings, 1 reply; 55+ messages in thread
From: Chang S. Bae @ 2025-11-10 18:01 UTC (permalink / raw)
To: kvm, linux-kernel; +Cc: pbonzini, seanjc, chao.gao, zhao1.liu, chang.seok.bae
Extend the instruction emulator to recognize and interpret the REX2
prefix byte. Also, detect and flag invalid prefix sequences after a REX2
prefix.
In the existing prefix-decoding loop,
* The loop exits when the first non-prefix byte is encountered.
* Any non-REX prefix clears previously recorded REX information.
For REX2, however, once a REX2 prefix is encountered, most subsequent
prefixes are invalid. So, each subsequent prefix needs to be validated
before continuing the loop.
Signed-off-by: Chang S. Bae <chang.seok.bae@intel.com>
---
RFC note:
The REX2 decoding itself is straightforward. The additional logic is
mainly to detect and handle invalid prefix sequences. If this seems
excessive, there is a chance to cut off this check since VMX would raise
'#UD' on such cases anyway.
---
arch/x86/kvm/emulate.c | 24 +++++++++++++++++++++++-
1 file changed, 23 insertions(+), 1 deletion(-)
diff --git a/arch/x86/kvm/emulate.c b/arch/x86/kvm/emulate.c
index 9bd61ea496e5..f9381a4055d6 100644
--- a/arch/x86/kvm/emulate.c
+++ b/arch/x86/kvm/emulate.c
@@ -4844,7 +4844,7 @@ int x86_decode_insn(struct x86_emulate_ctxt *ctxt, void *insn, int insn_len, int
ctxt->op_bytes = def_op_bytes;
ctxt->ad_bytes = def_ad_bytes;
- /* Legacy prefixes. */
+ /* Legacy and REX/REX2 prefixes. */
for (;;) {
switch (ctxt->b = insn_fetch(u8, ctxt)) {
case 0x66: /* operand-size override */
@@ -4887,9 +4887,20 @@ int x86_decode_insn(struct x86_emulate_ctxt *ctxt, void *insn, int insn_len, int
case 0x40 ... 0x4f: /* REX */
if (mode != X86EMUL_MODE_PROT64)
goto done_prefixes;
+ if (ctxt->rex_prefix == REX2_PREFIX)
+ break;
ctxt->rex_prefix = REX_PREFIX;
ctxt->rex.raw = 0x0f & ctxt->b;
continue;
+ case 0xd5: /* REX2 */
+ if (mode != X86EMUL_MODE_PROT64)
+ goto done_prefixes;
+ if (ctxt->rex_prefix == REX2_PREFIX &&
+ ctxt->rex.bits.m0 == 0)
+ break;
+ ctxt->rex_prefix = REX2_PREFIX;
+ ctxt->rex.raw = insn_fetch(u8, ctxt);
+ continue;
case 0xf0: /* LOCK */
ctxt->lock_prefix = 1;
break;
@@ -4901,6 +4912,17 @@ int x86_decode_insn(struct x86_emulate_ctxt *ctxt, void *insn, int insn_len, int
goto done_prefixes;
}
+ if (ctxt->rex_prefix == REX2_PREFIX) {
+ /*
+ * A legacy or REX prefix following a REX2 prefix
+ * forms an invalid byte sequences. Likewise,
+ * a second REX2 prefix following a REX2 prefix
+ * with M0=0 is invalid.
+ */
+ ctxt->rex_prefix = REX2_INVALID;
+ goto done_prefixes;
+ }
+
/* Any legacy prefix after a REX prefix nullifies its effect. */
ctxt->rex_prefix = REX_NONE;
ctxt->rex.raw = 0;
--
2.51.0
^ permalink raw reply related [flat|nested] 55+ messages in thread* Re: [PATCH RFC v1 16/20] KVM: x86: Decode REX2 prefix in the emulator
2025-11-10 18:01 ` [PATCH RFC v1 16/20] KVM: x86: Decode REX2 prefix " Chang S. Bae
@ 2025-11-11 17:55 ` Paolo Bonzini
2025-11-13 23:30 ` Chang S. Bae
0 siblings, 1 reply; 55+ messages in thread
From: Paolo Bonzini @ 2025-11-11 17:55 UTC (permalink / raw)
To: Chang S. Bae, kvm, linux-kernel; +Cc: seanjc, chao.gao, zhao1.liu
On 11/10/25 19:01, Chang S. Bae wrote:
> Extend the instruction emulator to recognize and interpret the REX2
> prefix byte. Also, detect and flag invalid prefix sequences after a REX2
> prefix.
>
> In the existing prefix-decoding loop,
> * The loop exits when the first non-prefix byte is encountered.
> * Any non-REX prefix clears previously recorded REX information.
>
> For REX2, however, once a REX2 prefix is encountered, most subsequent
> prefixes are invalid. So, each subsequent prefix needs to be validated
> before continuing the loop.
>
> Signed-off-by: Chang S. Bae <chang.seok.bae@intel.com>
> ---
> RFC note:
> The REX2 decoding itself is straightforward. The additional logic is
> mainly to detect and handle invalid prefix sequences. If this seems
> excessive, there is a chance to cut off this check since VMX would raise
> '#UD' on such cases anyway.
> ---
> arch/x86/kvm/emulate.c | 24 +++++++++++++++++++++++-
> 1 file changed, 23 insertions(+), 1 deletion(-)
>
> diff --git a/arch/x86/kvm/emulate.c b/arch/x86/kvm/emulate.c
> index 9bd61ea496e5..f9381a4055d6 100644
> --- a/arch/x86/kvm/emulate.c
> +++ b/arch/x86/kvm/emulate.c
> @@ -4844,7 +4844,7 @@ int x86_decode_insn(struct x86_emulate_ctxt *ctxt, void *insn, int insn_len, int
> ctxt->op_bytes = def_op_bytes;
> ctxt->ad_bytes = def_ad_bytes;
>
> - /* Legacy prefixes. */
> + /* Legacy and REX/REX2 prefixes. */
> for (;;) {
> switch (ctxt->b = insn_fetch(u8, ctxt)) {
> case 0x66: /* operand-size override */
> @@ -4887,9 +4887,20 @@ int x86_decode_insn(struct x86_emulate_ctxt *ctxt, void *insn, int insn_len, int
> case 0x40 ... 0x4f: /* REX */
> if (mode != X86EMUL_MODE_PROT64)
> goto done_prefixes;
> + if (ctxt->rex_prefix == REX2_PREFIX)
> + break;
> ctxt->rex_prefix = REX_PREFIX;
> ctxt->rex.raw = 0x0f & ctxt->b;
> continue;
> + case 0xd5: /* REX2 */
> + if (mode != X86EMUL_MODE_PROT64)
> + goto done_prefixes;
Here you should also check
if (ctxt->rex_prefix == REX_PREFIX) {
ctxt->rex_prefix = REX2_INVALID;
goto done_prefixes;
}
> + if (ctxt->rex_prefix == REX2_PREFIX &&
> + ctxt->rex.bits.m0 == 0)
> + break;
> + ctxt->rex_prefix = REX2_PREFIX;
> + ctxt->rex.raw = insn_fetch(u8, ctxt);
> + continue;
After REX2 always comes the main opcode byte, so you can "goto
done_prefixes" here. Or even jump here already; in pseudocode:
ctxt->b = insn_fetch(u8, ctxt);
if (rex2 & REX_M)
goto decode_twobyte;
else
goto decode_onebyte;
...
if (ctxt->b == 0x0f) {
decode_twobyte:
...
if (ctxt->b == 0x38 && ctxt->rex_prefix != REX2_PREFIX)
...
} else {
decode_onebyte:
...
}
> case 0xf0: /* LOCK */
> ctxt->lock_prefix = 1;
> break;
> @@ -4901,6 +4912,17 @@ int x86_decode_insn(struct x86_emulate_ctxt *ctxt, void *insn, int insn_len, int
> goto done_prefixes;
> }
>
> + if (ctxt->rex_prefix == REX2_PREFIX) {
> + /*
> + * A legacy or REX prefix following a REX2 prefix
> + * forms an invalid byte sequences. Likewise,
> + * a second REX2 prefix following a REX2 prefix
> + * with M0=0 is invalid.
> + */
> + ctxt->rex_prefix = REX2_INVALID;
> + goto done_prefixes;
> + }
... and this is not needed.
Paolo
> /* Any legacy prefix after a REX prefix nullifies its effect. */
> ctxt->rex_prefix = REX_NONE;
> ctxt->rex.raw = 0;
^ permalink raw reply [flat|nested] 55+ messages in thread* Re: [PATCH RFC v1 16/20] KVM: x86: Decode REX2 prefix in the emulator
2025-11-11 17:55 ` Paolo Bonzini
@ 2025-11-13 23:30 ` Chang S. Bae
2025-11-13 23:34 ` Paolo Bonzini
2025-11-17 20:01 ` Chang S. Bae
0 siblings, 2 replies; 55+ messages in thread
From: Chang S. Bae @ 2025-11-13 23:30 UTC (permalink / raw)
To: Paolo Bonzini, kvm, linux-kernel; +Cc: seanjc, chao.gao, zhao1.liu
[-- Attachment #1: Type: text/plain, Size: 2564 bytes --]
On 11/11/2025 9:55 AM, Paolo Bonzini wrote:
> On 11/10/25 19:01, Chang S. Bae wrote:
>>
>> case 0x40 ... 0x4f: /* REX */
>> if (mode != X86EMUL_MODE_PROT64)
>> goto done_prefixes;
>> + if (ctxt->rex_prefix == REX2_PREFIX)
>> + break;
>> ctxt->rex_prefix = REX_PREFIX;
>> ctxt->rex.raw = 0x0f & ctxt->b;
>> continue;
>> + case 0xd5: /* REX2 */
>> + if (mode != X86EMUL_MODE_PROT64)
>> + goto done_prefixes;
> Here you should also check
>
> if (ctxt->rex_prefix == REX_PREFIX) {
> ctxt->rex_prefix = REX2_INVALID;
> goto done_prefixes;
> }
You're right. Section 3.1.2.1 states:
| A REX prefix (0x4*) immediately preceding the REX2 prefix is not
| allowed and triggers #UD.
Now I think REX2_INVALID would just add another condition to handle
later. Instead, for such invalid case, it might be simpler to mark the
opcode as undefined and jump all the way after the lookup. See the diff
-- please let me know if you dislike it.
>> + if (ctxt->rex_prefix == REX2_PREFIX &&
>> + ctxt->rex.bits.m0 == 0)
>> + break;
>> + ctxt->rex_prefix = REX2_PREFIX;
>> + ctxt->rex.raw = insn_fetch(u8, ctxt);
>> + continue;
> After REX2 always comes the main opcode byte, so you can "goto
> done_prefixes" here. Or even jump here already; in pseudocode:
>
> ctxt->b = insn_fetch(u8, ctxt);
> if (rex2 & REX_M)
> goto decode_twobyte;
> else
> goto decode_onebyte;
Yes, agreed. I think this makes the control flow more explicit.
>> + if (ctxt->rex_prefix == REX2_PREFIX) {
>> + /*
>> + * A legacy or REX prefix following a REX2 prefix
>> + * forms an invalid byte sequences. Likewise,
>> + * a second REX2 prefix following a REX2 prefix
>> + * with M0=0 is invalid.
>> + */
>> + ctxt->rex_prefix = REX2_INVALID;
>> + goto done_prefixes;
>> + }
>
> ... and this is not needed.
I really like that this can go away.
[-- Attachment #2: PATCH16.diff --]
[-- Type: text/plain, Size: 2699 bytes --]
diff --git a/arch/x86/kvm/emulate.c b/arch/x86/kvm/emulate.c
index b8a946cbd587..c62d21de14cb 100644
--- a/arch/x86/kvm/emulate.c
+++ b/arch/x86/kvm/emulate.c
@@ -4479,6 +4479,8 @@ static const struct opcode opcode_map_0f_38[256] = {
N, N, X4(N), X8(N)
};
+static const struct opcode undefined = D(Undefined);
+
#undef D
#undef N
#undef G
@@ -4765,6 +4767,11 @@ static int decode_operand(struct x86_emulate_ctxt *ctxt, struct operand *op,
return rc;
}
+static inline bool emul_egpr_enabled(struct x86_emulate_ctxt *ctxt __maybe_unused)
+{
+ return false;
+}
+
int x86_decode_insn(struct x86_emulate_ctxt *ctxt, void *insn, int insn_len, int emulation_type)
{
int rc = X86EMUL_CONTINUE;
@@ -4817,7 +4824,7 @@ int x86_decode_insn(struct x86_emulate_ctxt *ctxt, void *insn, int insn_len, int
ctxt->op_bytes = def_op_bytes;
ctxt->ad_bytes = def_ad_bytes;
- /* Legacy prefixes. */
+ /* Legacy and REX/REX2 prefixes. */
for (;;) {
switch (ctxt->b = insn_fetch(u8, ctxt)) {
case 0x66: /* operand-size override */
@@ -4860,9 +4867,29 @@ int x86_decode_insn(struct x86_emulate_ctxt *ctxt, void *insn, int insn_len, int
case 0x40 ... 0x4f: /* REX */
if (mode != X86EMUL_MODE_PROT64)
goto done_prefixes;
+ if (ctxt->rex_prefix == REX2_PREFIX) {
+ opcode = undefined;
+ goto decode_done;
+ }
ctxt->rex_prefix = REX_PREFIX;
ctxt->rex = 0x0f & ctxt->b;
continue;
+ case 0xd5: /* REX2 */
+ if (mode != X86EMUL_MODE_PROT64)
+ goto done_prefixes;
+ if ((ctxt->rex_prefix == REX2_PREFIX && (ctxt->rex & REX_M) == 0) ||
+ (ctxt->rex_prefix == REX_PREFIX) ||
+ (!emul_egpr_enabled(ctxt))) {
+ opcode = undefined;
+ goto decode_done;
+ }
+ ctxt->rex_prefix = REX2_PREFIX;
+ ctxt->rex = insn_fetch(u8, ctxt);
+ ctxt->b = insn_fetch(u8, ctxt);
+ if (ctxt->rex & REX_M)
+ goto decode_twobytes;
+ else
+ goto decode_onebyte;
case 0xf0: /* LOCK */
ctxt->lock_prefix = 1;
break;
@@ -4889,6 +4916,7 @@ int x86_decode_insn(struct x86_emulate_ctxt *ctxt, void *insn, int insn_len, int
if (ctxt->b == 0x0f) {
/* Escape byte: start two-byte opcode sequence */
ctxt->b = insn_fetch(u8, ctxt);
+decode_twobytes:
if (ctxt->b == 0x38 && ctxt->rex_prefix != REX2_PREFIX) {
/* Three-byte opcode */
ctxt->opcode_len = 3;
@@ -4900,10 +4928,12 @@ int x86_decode_insn(struct x86_emulate_ctxt *ctxt, void *insn, int insn_len, int
opcode = twobyte_table[ctxt->b];
}
} else {
+decode_onebyte:
/* Single-byte opcode */
ctxt->opcode_len = 1;
opcode = opcode_table[ctxt->b];
}
+decode_done:
ctxt->d = opcode.flags;
if (ctxt->d & NoRex2 && ctxt->rex_prefix == REX2_PREFIX)
^ permalink raw reply related [flat|nested] 55+ messages in thread* Re: [PATCH RFC v1 16/20] KVM: x86: Decode REX2 prefix in the emulator
2025-11-13 23:30 ` Chang S. Bae
@ 2025-11-13 23:34 ` Paolo Bonzini
2025-11-17 20:01 ` Chang S. Bae
1 sibling, 0 replies; 55+ messages in thread
From: Paolo Bonzini @ 2025-11-13 23:34 UTC (permalink / raw)
To: Chang S. Bae, kvm, linux-kernel; +Cc: seanjc, chao.gao, zhao1.liu
On 11/14/25 00:30, Chang S. Bae wrote:
> On 11/11/2025 9:55 AM, Paolo Bonzini wrote:
>> On 11/10/25 19:01, Chang S. Bae wrote:
>>>
>>> case 0x40 ... 0x4f: /* REX */
>>> if (mode != X86EMUL_MODE_PROT64)
>>> goto done_prefixes;
>>> + if (ctxt->rex_prefix == REX2_PREFIX)
>>> + break;
>>> ctxt->rex_prefix = REX_PREFIX;
>>> ctxt->rex.raw = 0x0f & ctxt->b;
>>> continue;
>>> + case 0xd5: /* REX2 */
>>> + if (mode != X86EMUL_MODE_PROT64)
>>> + goto done_prefixes;
>> Here you should also check
>>
>> if (ctxt->rex_prefix == REX_PREFIX) {
>> ctxt->rex_prefix = REX2_INVALID;
>> goto done_prefixes;
>> }
>
> You're right. Section 3.1.2.1 states:
> | A REX prefix (0x4*) immediately preceding the REX2 prefix is not
> | allowed and triggers #UD.
>
> Now I think REX2_INVALID would just add another condition to handle
> later. Instead, for such invalid case, it might be simpler to mark the
> opcode as undefined and jump all the way after the lookup. See the diff
> -- please let me know if you dislike it.
Yes, I also thought it was unnecessary but waited until we merged the
respective patches.
Paolo
^ permalink raw reply [flat|nested] 55+ messages in thread* Re: [PATCH RFC v1 16/20] KVM: x86: Decode REX2 prefix in the emulator
2025-11-13 23:30 ` Chang S. Bae
2025-11-13 23:34 ` Paolo Bonzini
@ 2025-11-17 20:01 ` Chang S. Bae
2025-11-17 23:33 ` Paolo Bonzini
1 sibling, 1 reply; 55+ messages in thread
From: Chang S. Bae @ 2025-11-17 20:01 UTC (permalink / raw)
To: Paolo Bonzini, kvm, linux-kernel; +Cc: seanjc, chao.gao, zhao1.liu
[-- Attachment #1: Type: text/plain, Size: 1968 bytes --]
On 11/13/2025 3:30 PM, Chang S. Bae wrote:
> On 11/11/2025 9:55 AM, Paolo Bonzini wrote:
>> On 11/10/25 19:01, Chang S. Bae wrote:
>>>
>>> case 0x40 ... 0x4f: /* REX */
>>> if (mode != X86EMUL_MODE_PROT64)
>>> goto done_prefixes;
>>> + if (ctxt->rex_prefix == REX2_PREFIX)
>>> + break;
>>> ctxt->rex_prefix = REX_PREFIX;
>>> ctxt->rex.raw = 0x0f & ctxt->b;
>>> continue;
>>> + case 0xd5: /* REX2 */
>>> + if (mode != X86EMUL_MODE_PROT64)
>>> + goto done_prefixes;
[...]
>>> + if (ctxt->rex_prefix == REX2_PREFIX &&
>>> + ctxt->rex.bits.m0 == 0)
>>> + break;
>>> + ctxt->rex_prefix = REX2_PREFIX;
>>> + ctxt->rex.raw = insn_fetch(u8, ctxt);
>>> + continue;
>> After REX2 always comes the main opcode byte, so you can "goto
>> done_prefixes" here. Or even jump here already; in pseudocode:
>>
>> ctxt->b = insn_fetch(u8, ctxt);
>> if (rex2 & REX_M)
>> goto decode_twobyte;
>> else
>> goto decode_onebyte;
>
> Yes, agreed. I think this makes the control flow more explicit.
While rebasing onto your VEX series, I noticed a couple of missings:
(1) Jumping directly to the decode path skips the ctxt->op_bytes
setup.
(2) It also removes the logic that detects the invalid sequence:
REX2->REX (unless intentional).
Perhaps it makes sense to simply continue prefix parsing. Then, at
'done_prefixes', we can check the M bit next to the W-bit check and jump
to the two-byte decode path.
I’ve attached a revised diff on top of the VEX series.
[-- Attachment #2: PATCH16_again.diff --]
[-- Type: text/plain, Size: 2615 bytes --]
diff --git a/arch/x86/kvm/emulate.c b/arch/x86/kvm/emulate.c
index 1b3da3ba26b8..3a66741b6c8c 100644
--- a/arch/x86/kvm/emulate.c
+++ b/arch/x86/kvm/emulate.c
@@ -245,6 +245,7 @@ enum {
REX_X = 2,
REX_R = 4,
REX_W = 8,
+ REX_M = 0x80,
};
static void writeback_registers(struct x86_emulate_ctxt *ctxt)
@@ -4849,6 +4850,18 @@ static int x86_decode_avx(struct x86_emulate_ctxt *ctxt,
return rc;
}
+static inline bool rex2_invalid(struct x86_emulate_ctxt *ctxt)
+{
+ const struct x86_emulate_ops *ops = ctxt->ops;
+ u64 xcr = 0;
+
+ return ctxt->rex_prefix == REX_PREFIX ||
+ (ctxt->rex_prefix == REX2_PREFIX && !(ctxt->rex_bits & REX_M)) ||
+ !(ops->get_cr(ctxt, 4) & X86_CR4_OSXSAVE) ||
+ ops->get_xcr(ctxt, 0, &xcr) ||
+ !(xcr & XFEATURE_MASK_APX);
+}
+
int x86_decode_insn(struct x86_emulate_ctxt *ctxt, void *insn, int insn_len, int emulation_type)
{
int rc = X86EMUL_CONTINUE;
@@ -4902,7 +4915,7 @@ int x86_decode_insn(struct x86_emulate_ctxt *ctxt, void *insn, int insn_len, int
ctxt->op_bytes = def_op_bytes;
ctxt->ad_bytes = def_ad_bytes;
- /* Legacy prefixes. */
+ /* Legacy and REX/REX2 prefixes. */
for (;;) {
switch (ctxt->b = insn_fetch(u8, ctxt)) {
case 0x66: /* operand-size override */
@@ -4945,9 +4958,23 @@ int x86_decode_insn(struct x86_emulate_ctxt *ctxt, void *insn, int insn_len, int
case 0x40 ... 0x4f: /* REX */
if (mode != X86EMUL_MODE_PROT64)
goto done_prefixes;
+ if (ctxt->rex_prefix == REX2_PREFIX) {
+ opcode = ud;
+ goto done_modrm;
+ }
ctxt->rex_prefix = REX_PREFIX;
ctxt->rex_bits = ctxt->b & 0xf;
continue;
+ case 0xd5: /* REX2 */
+ if (mode != X86EMUL_MODE_PROT64)
+ goto done_prefixes;
+ if (rex2_invalid(ctxt)) {
+ opcode = ud;
+ goto done_modrm;
+ }
+ ctxt->rex_prefix = REX2_PREFIX;
+ ctxt->rex_bits = insn_fetch(u8, ctxt);
+ continue;
case 0xf0: /* LOCK */
ctxt->lock_prefix = 1;
break;
@@ -4970,6 +4997,9 @@ int x86_decode_insn(struct x86_emulate_ctxt *ctxt, void *insn, int insn_len, int
if (ctxt->rex_bits & REX_W)
ctxt->op_bytes = 8;
+ if (ctxt->rex_bits & REX_M)
+ goto decode_twobytes;
+
/* Opcode byte(s). */
if (ctxt->b == 0xc4 || ctxt->b == 0xc5) {
/* VEX or LDS/LES */
@@ -4987,8 +5017,9 @@ int x86_decode_insn(struct x86_emulate_ctxt *ctxt, void *insn, int insn_len, int
goto done;
} else if (ctxt->b == 0x0f) {
/* Two- or three-byte opcode */
- ctxt->opcode_len = 2;
ctxt->b = insn_fetch(u8, ctxt);
+decode_twobytes:
+ ctxt->opcode_len = 2;
opcode = twobyte_table[ctxt->b];
/* 0F_38 opcode map */
^ permalink raw reply related [flat|nested] 55+ messages in thread* Re: [PATCH RFC v1 16/20] KVM: x86: Decode REX2 prefix in the emulator
2025-11-17 20:01 ` Chang S. Bae
@ 2025-11-17 23:33 ` Paolo Bonzini
0 siblings, 0 replies; 55+ messages in thread
From: Paolo Bonzini @ 2025-11-17 23:33 UTC (permalink / raw)
To: Chang S. Bae
Cc: kvm, Kernel Mailing List, Linux, Sean Christopherson, Gao, Chao,
Zhao Liu
Il lun 17 nov 2025, 21:01 Chang S. Bae <chang.seok.bae@intel.com> ha scritto:
>
> While rebasing onto your VEX series, I noticed a couple of missings:
>
> (1) Jumping directly to the decode path skips the ctxt->op_bytes
> setup.
> (2) It also removes the logic that detects the invalid sequence:
> REX2->REX (unless intentional).
>
> Perhaps it makes sense to simply continue prefix parsing. Then, at
> 'done_prefixes', we can check the M bit next to the W-bit check and jump
> to the two-byte decode path.
>
> I’ve attached a revised diff on top of the VEX series.
Yes, that works for me with one change---after REX2 is processed there
should be a "ctxt->b = insn_fetch(u8, ctxt); goto done_prefixes;"
because REX2 is always the last prefix.
This also means that checking "(ctxt->rex_prefix == REX2_PREFIX &&
!(ctxt->rex_bits & REX_M))" is unnecessary. Instead the second REX2
prefix's 0xd5 byte can be treated as a No64 opcode and will trigger
#UD. In fact this is what the manual says: "a REX prefix byte (0x4*),
a VEX prefix byte (0xC4 or 0xC5), an EVEX prefix byte (0x62), or
another REX2 prefix byte (0xD5) following a REX2 prefix with REX2.M0 =
0 must #UD, because none of those bytes is the opcode of a valid
instruction in legacy map 0 in 64-bit mode".
So all you need to do is add the No64 flag to entries 0x40...0x4F of
the opcode_table, and then "goto done_prefixes" will cover that
sentence nicely.
Paolo
^ permalink raw reply [flat|nested] 55+ messages in thread
* [PATCH RFC v1 17/20] KVM: x86: Prepare APX state setting in XCR0
2025-11-10 18:01 [PATCH RFC v1 00/20] KVM: x86: Support APX feature for guests Chang S. Bae
` (15 preceding siblings ...)
2025-11-10 18:01 ` [PATCH RFC v1 16/20] KVM: x86: Decode REX2 prefix " Chang S. Bae
@ 2025-11-10 18:01 ` Chang S. Bae
2025-11-11 16:59 ` Paolo Bonzini
2025-11-10 18:01 ` [PATCH RFC v1 18/20] KVM: x86: Expose APX foundational feature bit to guests Chang S. Bae
` (4 subsequent siblings)
21 siblings, 1 reply; 55+ messages in thread
From: Chang S. Bae @ 2025-11-10 18:01 UTC (permalink / raw)
To: kvm, linux-kernel; +Cc: pbonzini, seanjc, chao.gao, zhao1.liu, chang.seok.bae
Prepare the APX state enabling in XCR0 by implementing the previous
placeholders and ensuring its readiness.
APX introduces EGPRs, tracked as XSTATE component 19. Like other
XSAVE-managed states, EGPR availability is controlled through XCR0, and
the registers are accessible only in 64-bit mode.
At this point, only VMX supports EGPRs. SVM will require corresponding
extensions to handle EGPR indices.
The addition to the supported XCR0 mask should accompany guest CPUID
exposure, which will be done separately.
Signed-off-by: Chang S. Bae <chang.seok.bae@intel.com>
---
RFC note
Not all callers may need to validate the XCR0 bit -- maybe a capability
bit. However, every exit associated with EGPRs should already have that
control bit set in the first place. Checking it explicitly does not
charge additional cost, so I have this for consistency.
---
arch/x86/kvm/emulate.c | 9 +++++++--
arch/x86/kvm/kvm_cache_regs.h | 1 +
arch/x86/kvm/kvm_emulate.h | 1 +
arch/x86/kvm/svm/svm.c | 7 ++++++-
arch/x86/kvm/vmx/vmx.h | 9 ++++++++-
arch/x86/kvm/x86.c | 11 +++++++++++
6 files changed, 34 insertions(+), 4 deletions(-)
diff --git a/arch/x86/kvm/emulate.c b/arch/x86/kvm/emulate.c
index f9381a4055d6..ba3020e6f469 100644
--- a/arch/x86/kvm/emulate.c
+++ b/arch/x86/kvm/emulate.c
@@ -4787,9 +4787,14 @@ static int decode_operand(struct x86_emulate_ctxt *ctxt, struct operand *op,
return rc;
}
-static inline bool emul_egpr_enabled(struct x86_emulate_ctxt *ctxt __maybe_unused)
+/* EGPR availability is controlled by the APX feature bit in XCR0. */
+static inline bool emul_egpr_enabled(struct x86_emulate_ctxt *ctxt)
{
- return false;
+ u64 xcr0;
+
+ ctxt->ops->get_xcr(ctxt, XCR_XFEATURE_ENABLED_MASK, &xcr0);
+
+ return xcr0 & XFEATURE_MASK_APX;
}
int x86_decode_insn(struct x86_emulate_ctxt *ctxt, void *insn, int insn_len, int emulation_type)
diff --git a/arch/x86/kvm/kvm_cache_regs.h b/arch/x86/kvm/kvm_cache_regs.h
index 8ddb01191d6f..acdb3751317c 100644
--- a/arch/x86/kvm/kvm_cache_regs.h
+++ b/arch/x86/kvm/kvm_cache_regs.h
@@ -3,6 +3,7 @@
#define ASM_KVM_CACHE_REGS_H
#include <linux/kvm_host.h>
+#include <asm/fpu/xcr.h>
#define KVM_POSSIBLE_CR0_GUEST_BITS (X86_CR0_TS | X86_CR0_WP)
#define KVM_POSSIBLE_CR4_GUEST_BITS \
diff --git a/arch/x86/kvm/kvm_emulate.h b/arch/x86/kvm/kvm_emulate.h
index cc16211d61f6..673a82532c78 100644
--- a/arch/x86/kvm/kvm_emulate.h
+++ b/arch/x86/kvm/kvm_emulate.h
@@ -237,6 +237,7 @@ struct x86_emulate_ops {
bool (*is_smm)(struct x86_emulate_ctxt *ctxt);
int (*leave_smm)(struct x86_emulate_ctxt *ctxt);
void (*triple_fault)(struct x86_emulate_ctxt *ctxt);
+ int (*get_xcr)(struct x86_emulate_ctxt *ctxt, u32 index, u64 *xcr);
int (*set_xcr)(struct x86_emulate_ctxt *ctxt, u32 index, u64 xcr);
gva_t (*get_untagged_addr)(struct x86_emulate_ctxt *ctxt, gva_t addr,
diff --git a/arch/x86/kvm/svm/svm.c b/arch/x86/kvm/svm/svm.c
index 3aa2c37754ef..e6a082686000 100644
--- a/arch/x86/kvm/svm/svm.c
+++ b/arch/x86/kvm/svm/svm.c
@@ -5288,8 +5288,13 @@ static __init int svm_hardware_setup(void)
}
kvm_enable_efer_bits(EFER_NX);
+ /*
+ * APX introduces EGPRs, which require additional VMCB support.
+ * Disable APX until the necessary extensions are handled.
+ */
kvm_caps.supported_xcr0 &= ~(XFEATURE_MASK_BNDREGS |
- XFEATURE_MASK_BNDCSR);
+ XFEATURE_MASK_BNDCSR |
+ XFEATURE_MASK_APX);
if (boot_cpu_has(X86_FEATURE_FXSR_OPT))
kvm_enable_efer_bits(EFER_FFXSR);
diff --git a/arch/x86/kvm/vmx/vmx.h b/arch/x86/kvm/vmx/vmx.h
index 6cf1eb739caf..784aa0504dce 100644
--- a/arch/x86/kvm/vmx/vmx.h
+++ b/arch/x86/kvm/vmx/vmx.h
@@ -372,7 +372,14 @@ struct vmx_insn_info {
union insn_info info;
};
-static inline bool vmx_egpr_enabled(struct kvm_vcpu *vcpu __maybe_unused) { return false; }
+/*
+ * EGPR availability is controlled by the APX xfeature bit in XCR0 and is
+ * only accessible in 64-bit mode.
+ */
+static inline bool vmx_egpr_enabled(struct kvm_vcpu *vcpu)
+{
+ return vcpu->arch.xcr0 & XFEATURE_MASK_APX && is_64_bit_mode(vcpu);
+}
static inline struct vmx_insn_info vmx_get_insn_info(struct kvm_vcpu *vcpu)
{
diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
index 4c8c2fc3bda6..e087db0f4153 100644
--- a/arch/x86/kvm/x86.c
+++ b/arch/x86/kvm/x86.c
@@ -8843,6 +8843,16 @@ static void emulator_triple_fault(struct x86_emulate_ctxt *ctxt)
kvm_make_request(KVM_REQ_TRIPLE_FAULT, emul_to_vcpu(ctxt));
}
+static int emulator_get_xcr(struct x86_emulate_ctxt *ctxt, u32 index, u64 *xcr)
+{
+ /* Only support XCR_XFEATURE_ENABLED_MASK now */
+ if (index != XCR_XFEATURE_ENABLED_MASK)
+ return 1;
+
+ *xcr = emul_to_vcpu(ctxt)->arch.xcr0;
+ return 0;
+}
+
static int emulator_set_xcr(struct x86_emulate_ctxt *ctxt, u32 index, u64 xcr)
{
return __kvm_set_xcr(emul_to_vcpu(ctxt), index, xcr);
@@ -8915,6 +8925,7 @@ static const struct x86_emulate_ops emulate_ops = {
.is_smm = emulator_is_smm,
.leave_smm = emulator_leave_smm,
.triple_fault = emulator_triple_fault,
+ .get_xcr = emulator_get_xcr,
.set_xcr = emulator_set_xcr,
.get_untagged_addr = emulator_get_untagged_addr,
.is_canonical_addr = emulator_is_canonical_addr,
--
2.51.0
^ permalink raw reply related [flat|nested] 55+ messages in thread* Re: [PATCH RFC v1 17/20] KVM: x86: Prepare APX state setting in XCR0
2025-11-10 18:01 ` [PATCH RFC v1 17/20] KVM: x86: Prepare APX state setting in XCR0 Chang S. Bae
@ 2025-11-11 16:59 ` Paolo Bonzini
2025-11-13 23:32 ` Chang S. Bae
0 siblings, 1 reply; 55+ messages in thread
From: Paolo Bonzini @ 2025-11-11 16:59 UTC (permalink / raw)
To: Chang S. Bae, kvm, linux-kernel; +Cc: seanjc, chao.gao, zhao1.liu
On 11/10/25 19:01, Chang S. Bae wrote:
> Prepare the APX state enabling in XCR0 by implementing the previous
> placeholders and ensuring its readiness.
>
> APX introduces EGPRs, tracked as XSTATE component 19. Like other
> XSAVE-managed states, EGPR availability is controlled through XCR0, and
> the registers are accessible only in 64-bit mode.
>
> At this point, only VMX supports EGPRs. SVM will require corresponding
> extensions to handle EGPR indices.
>
> The addition to the supported XCR0 mask should accompany guest CPUID
> exposure, which will be done separately.
You should also adjust set_xcr to reject setting BNDREGS and APX
together (QEMU should also reject MPX and APX together in CPUID, but KVM
doesn't care enough about invalid CPUID configuration).
Paolo
>
> Signed-off-by: Chang S. Bae <chang.seok.bae@intel.com>
> ---
> RFC note
> Not all callers may need to validate the XCR0 bit -- maybe a capability
> bit. However, every exit associated with EGPRs should already have that
> control bit set in the first place. Checking it explicitly does not
> charge additional cost, so I have this for consistency.
> ---
> arch/x86/kvm/emulate.c | 9 +++++++--
> arch/x86/kvm/kvm_cache_regs.h | 1 +
> arch/x86/kvm/kvm_emulate.h | 1 +
> arch/x86/kvm/svm/svm.c | 7 ++++++-
> arch/x86/kvm/vmx/vmx.h | 9 ++++++++-
> arch/x86/kvm/x86.c | 11 +++++++++++
> 6 files changed, 34 insertions(+), 4 deletions(-)
>
> diff --git a/arch/x86/kvm/emulate.c b/arch/x86/kvm/emulate.c
> index f9381a4055d6..ba3020e6f469 100644
> --- a/arch/x86/kvm/emulate.c
> +++ b/arch/x86/kvm/emulate.c
> @@ -4787,9 +4787,14 @@ static int decode_operand(struct x86_emulate_ctxt *ctxt, struct operand *op,
> return rc;
> }
>
> -static inline bool emul_egpr_enabled(struct x86_emulate_ctxt *ctxt __maybe_unused)
> +/* EGPR availability is controlled by the APX feature bit in XCR0. */
> +static inline bool emul_egpr_enabled(struct x86_emulate_ctxt *ctxt)
> {
> - return false;
> + u64 xcr0;
> +
> + ctxt->ops->get_xcr(ctxt, XCR_XFEATURE_ENABLED_MASK, &xcr0);
> +
> + return xcr0 & XFEATURE_MASK_APX;
> }
>
> int x86_decode_insn(struct x86_emulate_ctxt *ctxt, void *insn, int insn_len, int emulation_type)
> diff --git a/arch/x86/kvm/kvm_cache_regs.h b/arch/x86/kvm/kvm_cache_regs.h
> index 8ddb01191d6f..acdb3751317c 100644
> --- a/arch/x86/kvm/kvm_cache_regs.h
> +++ b/arch/x86/kvm/kvm_cache_regs.h
> @@ -3,6 +3,7 @@
> #define ASM_KVM_CACHE_REGS_H
>
> #include <linux/kvm_host.h>
> +#include <asm/fpu/xcr.h>
>
> #define KVM_POSSIBLE_CR0_GUEST_BITS (X86_CR0_TS | X86_CR0_WP)
> #define KVM_POSSIBLE_CR4_GUEST_BITS \
> diff --git a/arch/x86/kvm/kvm_emulate.h b/arch/x86/kvm/kvm_emulate.h
> index cc16211d61f6..673a82532c78 100644
> --- a/arch/x86/kvm/kvm_emulate.h
> +++ b/arch/x86/kvm/kvm_emulate.h
> @@ -237,6 +237,7 @@ struct x86_emulate_ops {
> bool (*is_smm)(struct x86_emulate_ctxt *ctxt);
> int (*leave_smm)(struct x86_emulate_ctxt *ctxt);
> void (*triple_fault)(struct x86_emulate_ctxt *ctxt);
> + int (*get_xcr)(struct x86_emulate_ctxt *ctxt, u32 index, u64 *xcr);
> int (*set_xcr)(struct x86_emulate_ctxt *ctxt, u32 index, u64 xcr);
>
> gva_t (*get_untagged_addr)(struct x86_emulate_ctxt *ctxt, gva_t addr,
> diff --git a/arch/x86/kvm/svm/svm.c b/arch/x86/kvm/svm/svm.c
> index 3aa2c37754ef..e6a082686000 100644
> --- a/arch/x86/kvm/svm/svm.c
> +++ b/arch/x86/kvm/svm/svm.c
> @@ -5288,8 +5288,13 @@ static __init int svm_hardware_setup(void)
> }
> kvm_enable_efer_bits(EFER_NX);
>
> + /*
> + * APX introduces EGPRs, which require additional VMCB support.
> + * Disable APX until the necessary extensions are handled.
> + */
> kvm_caps.supported_xcr0 &= ~(XFEATURE_MASK_BNDREGS |
> - XFEATURE_MASK_BNDCSR);
> + XFEATURE_MASK_BNDCSR |
> + XFEATURE_MASK_APX);
>
> if (boot_cpu_has(X86_FEATURE_FXSR_OPT))
> kvm_enable_efer_bits(EFER_FFXSR);
> diff --git a/arch/x86/kvm/vmx/vmx.h b/arch/x86/kvm/vmx/vmx.h
> index 6cf1eb739caf..784aa0504dce 100644
> --- a/arch/x86/kvm/vmx/vmx.h
> +++ b/arch/x86/kvm/vmx/vmx.h
> @@ -372,7 +372,14 @@ struct vmx_insn_info {
> union insn_info info;
> };
>
> -static inline bool vmx_egpr_enabled(struct kvm_vcpu *vcpu __maybe_unused) { return false; }
> +/*
> + * EGPR availability is controlled by the APX xfeature bit in XCR0 and is
> + * only accessible in 64-bit mode.
> + */
> +static inline bool vmx_egpr_enabled(struct kvm_vcpu *vcpu)
> +{
> + return vcpu->arch.xcr0 & XFEATURE_MASK_APX && is_64_bit_mode(vcpu);
> +}
>
> static inline struct vmx_insn_info vmx_get_insn_info(struct kvm_vcpu *vcpu)
> {
> diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
> index 4c8c2fc3bda6..e087db0f4153 100644
> --- a/arch/x86/kvm/x86.c
> +++ b/arch/x86/kvm/x86.c
> @@ -8843,6 +8843,16 @@ static void emulator_triple_fault(struct x86_emulate_ctxt *ctxt)
> kvm_make_request(KVM_REQ_TRIPLE_FAULT, emul_to_vcpu(ctxt));
> }
>
> +static int emulator_get_xcr(struct x86_emulate_ctxt *ctxt, u32 index, u64 *xcr)
> +{
> + /* Only support XCR_XFEATURE_ENABLED_MASK now */
> + if (index != XCR_XFEATURE_ENABLED_MASK)
> + return 1;
> +
> + *xcr = emul_to_vcpu(ctxt)->arch.xcr0;
> + return 0;
> +}
> +
> static int emulator_set_xcr(struct x86_emulate_ctxt *ctxt, u32 index, u64 xcr)
> {
> return __kvm_set_xcr(emul_to_vcpu(ctxt), index, xcr);
> @@ -8915,6 +8925,7 @@ static const struct x86_emulate_ops emulate_ops = {
> .is_smm = emulator_is_smm,
> .leave_smm = emulator_leave_smm,
> .triple_fault = emulator_triple_fault,
> + .get_xcr = emulator_get_xcr,
> .set_xcr = emulator_set_xcr,
> .get_untagged_addr = emulator_get_untagged_addr,
> .is_canonical_addr = emulator_is_canonical_addr,
^ permalink raw reply [flat|nested] 55+ messages in thread* Re: [PATCH RFC v1 17/20] KVM: x86: Prepare APX state setting in XCR0
2025-11-11 16:59 ` Paolo Bonzini
@ 2025-11-13 23:32 ` Chang S. Bae
0 siblings, 0 replies; 55+ messages in thread
From: Chang S. Bae @ 2025-11-13 23:32 UTC (permalink / raw)
To: Paolo Bonzini, kvm, linux-kernel; +Cc: seanjc, chao.gao, zhao1.liu
On 11/11/2025 8:59 AM, Paolo Bonzini wrote:
>
> You should also adjust set_xcr to reject setting BNDREGS and APX
/* MPX and APX conflict in the non-compacted XSAVE format */
if (xcr0 & XFEATURE_MASK_BNDREGS && xcr0 & XFEATURE_MASK_APX)
return 1;
> together (QEMU should also reject MPX and APX together in CPUID, but KVM
> doesn't care enough about invalid CPUID configuration).
Zhao, please ensure this on your QEMU patches.
^ permalink raw reply [flat|nested] 55+ messages in thread
* [PATCH RFC v1 18/20] KVM: x86: Expose APX foundational feature bit to guests
2025-11-10 18:01 [PATCH RFC v1 00/20] KVM: x86: Support APX feature for guests Chang S. Bae
` (16 preceding siblings ...)
2025-11-10 18:01 ` [PATCH RFC v1 17/20] KVM: x86: Prepare APX state setting in XCR0 Chang S. Bae
@ 2025-11-10 18:01 ` Chang S. Bae
2025-11-10 18:01 ` [PATCH RFC v1 19/20] KVM: x86: Expose APX sub-features " Chang S. Bae
` (3 subsequent siblings)
21 siblings, 0 replies; 55+ messages in thread
From: Chang S. Bae @ 2025-11-10 18:01 UTC (permalink / raw)
To: kvm, linux-kernel
Cc: pbonzini, seanjc, chao.gao, zhao1.liu, chang.seok.bae, Peter Fang
From: Peter Fang <peter.fang@intel.com>
Add the APX xfeature bit to the list of supported XCR0 components and
expose the APX feature to guests. Define the APX CPUID feature bits and
update the maximum supported CPUID leaf to 0x29 to include the APX leaf.
On SVM systems, ensure that the feature is not advertised as EGPR support
is not yet supported.
No APX sub-features are enumerated yet. Those will be exposed in a
separate patch.
Signed-off-by: Peter Fang <peter.fang@intel.com>
Signed-off-by: Chang S. Bae <chang.seok.bae@intel.com>
---
Peter had figured out establishing this change by spotting the CPUID
maximum updates.
---
arch/x86/kvm/cpuid.c | 8 +++++++-
arch/x86/kvm/reverse_cpuid.h | 2 ++
arch/x86/kvm/svm/svm.c | 8 ++++++++
arch/x86/kvm/x86.c | 3 ++-
4 files changed, 19 insertions(+), 2 deletions(-)
diff --git a/arch/x86/kvm/cpuid.c b/arch/x86/kvm/cpuid.c
index 52524e0ca97f..b90e58f2a42f 100644
--- a/arch/x86/kvm/cpuid.c
+++ b/arch/x86/kvm/cpuid.c
@@ -1031,6 +1031,7 @@ void kvm_set_cpu_caps(void)
F(AVX_VNNI_INT16),
F(PREFETCHITI),
F(AVX10),
+ SCATTERED_F(APX),
);
kvm_cpu_cap_init(CPUID_7_2_EDX,
@@ -1393,7 +1394,7 @@ static inline int __do_cpuid_func(struct kvm_cpuid_array *array, u32 function)
switch (function) {
case 0:
/* Limited to the highest leaf implemented in KVM. */
- entry->eax = min(entry->eax, 0x24U);
+ entry->eax = min(entry->eax, 0x29U);
break;
case 1:
cpuid_entry_override(entry, CPUID_1_EDX);
@@ -1638,6 +1639,11 @@ static inline int __do_cpuid_func(struct kvm_cpuid_array *array, u32 function)
entry->edx = 0;
break;
}
+ case 0x29: {
+ /* No APX sub-features are supported yet */
+ entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
+ break;
+ }
case KVM_CPUID_SIGNATURE: {
const u32 *sigptr = (const u32 *)KVM_SIGNATURE;
entry->eax = KVM_CPUID_FEATURES;
diff --git a/arch/x86/kvm/reverse_cpuid.h b/arch/x86/kvm/reverse_cpuid.h
index 743ab25ba787..e9d9fb4070ca 100644
--- a/arch/x86/kvm/reverse_cpuid.h
+++ b/arch/x86/kvm/reverse_cpuid.h
@@ -35,6 +35,7 @@
#define X86_FEATURE_AVX_VNNI_INT16 KVM_X86_FEATURE(CPUID_7_1_EDX, 10)
#define X86_FEATURE_PREFETCHITI KVM_X86_FEATURE(CPUID_7_1_EDX, 14)
#define X86_FEATURE_AVX10 KVM_X86_FEATURE(CPUID_7_1_EDX, 19)
+#define KVM_X86_FEATURE_APX KVM_X86_FEATURE(CPUID_7_1_EDX, 21)
/* Intel-defined sub-features, CPUID level 0x00000007:2 (EDX) */
#define X86_FEATURE_INTEL_PSFD KVM_X86_FEATURE(CPUID_7_2_EDX, 0)
@@ -126,6 +127,7 @@ static __always_inline u32 __feature_translate(int x86_feature)
KVM_X86_TRANSLATE_FEATURE(SGX1);
KVM_X86_TRANSLATE_FEATURE(SGX2);
KVM_X86_TRANSLATE_FEATURE(SGX_EDECCSSA);
+ KVM_X86_TRANSLATE_FEATURE(APX);
KVM_X86_TRANSLATE_FEATURE(CONSTANT_TSC);
KVM_X86_TRANSLATE_FEATURE(PERFMON_V2);
KVM_X86_TRANSLATE_FEATURE(RRSBA_CTRL);
diff --git a/arch/x86/kvm/svm/svm.c b/arch/x86/kvm/svm/svm.c
index e6a082686000..da57f7506f88 100644
--- a/arch/x86/kvm/svm/svm.c
+++ b/arch/x86/kvm/svm/svm.c
@@ -5271,6 +5271,14 @@ static __init void svm_set_cpu_caps(void)
*/
kvm_cpu_cap_clear(X86_FEATURE_BUS_LOCK_DETECT);
kvm_cpu_cap_clear(X86_FEATURE_MSR_IMM);
+
+ /*
+ * If the APX xfeature bit is not supported, meaning that VMCB
+ * support for EGPRs is unavailable, then the APX feature should
+ * not be exposed to the guest.
+ */
+ if (!(kvm_caps.supported_xcr0 & XFEATURE_MASK_APX))
+ kvm_cpu_cap_clear(X86_FEATURE_APX);
}
static __init int svm_hardware_setup(void)
diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
index e087db0f4153..bcf8e95d88dc 100644
--- a/arch/x86/kvm/x86.c
+++ b/arch/x86/kvm/x86.c
@@ -217,7 +217,8 @@ static struct kvm_user_return_msrs __percpu *user_return_msrs;
#define KVM_SUPPORTED_XCR0 (XFEATURE_MASK_FP | XFEATURE_MASK_SSE \
| XFEATURE_MASK_YMM | XFEATURE_MASK_BNDREGS \
| XFEATURE_MASK_BNDCSR | XFEATURE_MASK_AVX512 \
- | XFEATURE_MASK_PKRU | XFEATURE_MASK_XTILE)
+ | XFEATURE_MASK_PKRU | XFEATURE_MASK_XTILE \
+ | XFEATURE_MASK_APX)
#define XFEATURE_MASK_CET_ALL (XFEATURE_MASK_CET_USER | XFEATURE_MASK_CET_KERNEL)
/*
--
2.51.0
^ permalink raw reply related [flat|nested] 55+ messages in thread* [PATCH RFC v1 19/20] KVM: x86: Expose APX sub-features to guests
2025-11-10 18:01 [PATCH RFC v1 00/20] KVM: x86: Support APX feature for guests Chang S. Bae
` (17 preceding siblings ...)
2025-11-10 18:01 ` [PATCH RFC v1 18/20] KVM: x86: Expose APX foundational feature bit to guests Chang S. Bae
@ 2025-11-10 18:01 ` Chang S. Bae
2025-11-10 18:01 ` [PATCH RFC v1 20/20] KVM: selftests: Add APX state handling and XCR0 sanity checks Chang S. Bae
` (2 subsequent siblings)
21 siblings, 0 replies; 55+ messages in thread
From: Chang S. Bae @ 2025-11-10 18:01 UTC (permalink / raw)
To: kvm, linux-kernel; +Cc: pbonzini, seanjc, chao.gao, zhao1.liu, chang.seok.bae
Add CPUID leaf 0x29 sub-leaf 0 to enumerate APX sub-features to guests.
This leaf currently defines the following sub-features:
* New Conditional Instructions (NCI)
* New Data Destination (NDD)
* Flags Suppression (NF)
The CPUID leaf is only exposed if the APX feature is enabled.
Signed-off-by: Chang S. Bae <chang.seok.bae@intel.com>
---
arch/x86/include/asm/kvm_host.h | 1 +
arch/x86/kvm/cpuid.c | 10 ++++++++--
arch/x86/kvm/reverse_cpuid.h | 4 ++++
3 files changed, 13 insertions(+), 2 deletions(-)
diff --git a/arch/x86/include/asm/kvm_host.h b/arch/x86/include/asm/kvm_host.h
index 940f83c121cf..763872080c64 100644
--- a/arch/x86/include/asm/kvm_host.h
+++ b/arch/x86/include/asm/kvm_host.h
@@ -794,6 +794,7 @@ enum kvm_only_cpuid_leafs {
CPUID_24_0_EBX,
CPUID_8000_0021_ECX,
CPUID_7_1_ECX,
+ CPUID_29_0_EBX,
NR_KVM_CPU_CAPS,
NKVMCAPINTS = NR_KVM_CPU_CAPS - NCAPINTS,
diff --git a/arch/x86/kvm/cpuid.c b/arch/x86/kvm/cpuid.c
index b90e58f2a42f..95c25de641ca 100644
--- a/arch/x86/kvm/cpuid.c
+++ b/arch/x86/kvm/cpuid.c
@@ -1063,6 +1063,10 @@ void kvm_set_cpu_caps(void)
F(AVX10_512),
);
+ kvm_cpu_cap_init(CPUID_29_0_EBX,
+ F(APX_NCI_NDD_NF),
+ );
+
kvm_cpu_cap_init(CPUID_8000_0001_ECX,
F(LAHF_LM),
F(CMP_LEGACY),
@@ -1640,8 +1644,10 @@ static inline int __do_cpuid_func(struct kvm_cpuid_array *array, u32 function)
break;
}
case 0x29: {
- /* No APX sub-features are supported yet */
- entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
+ if (!(kvm_caps.supported_xcr0 & XFEATURE_MASK_APX)) {
+ entry->eax = entry->ebx = entry->ecx = entry->edx = 0;
+ break;
+ }
break;
}
case KVM_CPUID_SIGNATURE: {
diff --git a/arch/x86/kvm/reverse_cpuid.h b/arch/x86/kvm/reverse_cpuid.h
index e9d9fb4070ca..a8eca23ee2d4 100644
--- a/arch/x86/kvm/reverse_cpuid.h
+++ b/arch/x86/kvm/reverse_cpuid.h
@@ -50,6 +50,9 @@
#define X86_FEATURE_AVX10_256 KVM_X86_FEATURE(CPUID_24_0_EBX, 17)
#define X86_FEATURE_AVX10_512 KVM_X86_FEATURE(CPUID_24_0_EBX, 18)
+/* Intel-defined sub-features, CPUID level 0x00000029:0 (EBX) */
+#define X86_FEATURE_APX_NCI_NDD_NF KVM_X86_FEATURE(CPUID_29_0_EBX, 0)
+
/* CPUID level 0x80000007 (EDX). */
#define KVM_X86_FEATURE_CONSTANT_TSC KVM_X86_FEATURE(CPUID_8000_0007_EDX, 8)
@@ -92,6 +95,7 @@ static const struct cpuid_reg reverse_cpuid[] = {
[CPUID_24_0_EBX] = { 0x24, 0, CPUID_EBX},
[CPUID_8000_0021_ECX] = {0x80000021, 0, CPUID_ECX},
[CPUID_7_1_ECX] = { 7, 1, CPUID_ECX},
+ [CPUID_29_0_EBX] = { 0x29, 0, CPUID_EBX},
};
/*
--
2.51.0
^ permalink raw reply related [flat|nested] 55+ messages in thread* [PATCH RFC v1 20/20] KVM: selftests: Add APX state handling and XCR0 sanity checks
2025-11-10 18:01 [PATCH RFC v1 00/20] KVM: x86: Support APX feature for guests Chang S. Bae
` (18 preceding siblings ...)
2025-11-10 18:01 ` [PATCH RFC v1 19/20] KVM: x86: Expose APX sub-features " Chang S. Bae
@ 2025-11-10 18:01 ` Chang S. Bae
2025-11-10 18:50 ` [PATCH RFC v1 00/20] KVM: x86: Support APX feature for guests Chang S. Bae
2025-11-11 18:14 ` Paolo Bonzini
21 siblings, 0 replies; 55+ messages in thread
From: Chang S. Bae @ 2025-11-10 18:01 UTC (permalink / raw)
To: kvm, linux-kernel; +Cc: pbonzini, seanjc, chao.gao, zhao1.liu, chang.seok.bae
Now that KVM exposes the APX feature to guests on APX-capable systems,
extend the selftests to validate XCR0 configuration and state management.
Since APX repurposes the XSAVE area previously used by MPX in the
non-compacted format, add a check to ensure that MPX states are not set
when APX is enabled.
Also, load non-init APX state data in the guest so that XSTATE_BV[APX] is
set, allowing validation of APX state testing.
Signed-off-by: Chang S. Bae <chang.seok.bae@intel.com>
---
.../selftests/kvm/include/x86/processor.h | 1 +
tools/testing/selftests/kvm/x86/state_test.c | 6 ++++++
.../selftests/kvm/x86/xcr0_cpuid_test.c | 20 +++++++++++++++++++
3 files changed, 27 insertions(+)
diff --git a/tools/testing/selftests/kvm/include/x86/processor.h b/tools/testing/selftests/kvm/include/x86/processor.h
index 51cd84b9ca66..dde7af40584e 100644
--- a/tools/testing/selftests/kvm/include/x86/processor.h
+++ b/tools/testing/selftests/kvm/include/x86/processor.h
@@ -88,6 +88,7 @@ struct xstate {
#define XFEATURE_MASK_LBR BIT_ULL(15)
#define XFEATURE_MASK_XTILE_CFG BIT_ULL(17)
#define XFEATURE_MASK_XTILE_DATA BIT_ULL(18)
+#define XFEATURE_MASK_APX BIT_ULL(19)
#define XFEATURE_MASK_AVX512 (XFEATURE_MASK_OPMASK | \
XFEATURE_MASK_ZMM_Hi256 | \
diff --git a/tools/testing/selftests/kvm/x86/state_test.c b/tools/testing/selftests/kvm/x86/state_test.c
index 141b7fc0c965..6d1dc575b22b 100644
--- a/tools/testing/selftests/kvm/x86/state_test.c
+++ b/tools/testing/selftests/kvm/x86/state_test.c
@@ -167,6 +167,12 @@ static void __attribute__((__flatten__)) guest_code(void *arg)
asm volatile ("vmovupd %0, %%zmm16" :: "m" (buffer));
}
+ if (supported_xcr0 & XFEATURE_MASK_APX) {
+ /* mov $0xcccccccc, %r16 */
+ asm volatile (".byte 0xd5, 0x18, 0xb8, 0xcc, 0xcc,"
+ "0xcc, 0xcc, 0x00, 0x00, 0x00, 0x00");
+ }
+
if (this_cpu_has(X86_FEATURE_MPX)) {
uint64_t bounds[2] = { 10, 0xffffffffull };
uint64_t output[2] = { };
diff --git a/tools/testing/selftests/kvm/x86/xcr0_cpuid_test.c b/tools/testing/selftests/kvm/x86/xcr0_cpuid_test.c
index d038c1571729..6e4f2f83c831 100644
--- a/tools/testing/selftests/kvm/x86/xcr0_cpuid_test.c
+++ b/tools/testing/selftests/kvm/x86/xcr0_cpuid_test.c
@@ -46,6 +46,21 @@ do { \
__supported, (xfeatures)); \
} while (0)
+/*
+ * Verify that mutually exclusive architectural features do not overlap.
+ * For example, APX and MPX must never be reported as supported together.
+ */
+#define ASSERT_XFEATURE_CONFLICT(supported_xcr0, xfeatures, conflicts) \
+do { \
+ uint64_t __supported = (supported_xcr0) & ((xfeatures) | (conflicts)); \
+ \
+ __GUEST_ASSERT((__supported & (xfeatures)) != (xfeatures) || \
+ !(__supported & (conflicts)), \
+ "supported = 0x%lx, xfeatures = 0x%llx, conflicts = 0x%llx", \
+ __supported, (xfeatures), (conflicts)); \
+} while (0)
+
+
static void guest_code(void)
{
uint64_t initial_xcr0;
@@ -79,6 +94,11 @@ static void guest_code(void)
ASSERT_ALL_OR_NONE_XFEATURE(supported_xcr0,
XFEATURE_MASK_XTILE);
+ /* Check APX by ensuring MPX is not exposed concurrently */
+ ASSERT_XFEATURE_CONFLICT(supported_xcr0,
+ XFEATURE_MASK_APX,
+ XFEATURE_MASK_BNDREGS | XFEATURE_MASK_BNDCSR);
+
vector = xsetbv_safe(0, XFEATURE_MASK_FP);
__GUEST_ASSERT(!vector,
"Expected success on XSETBV(FP), got %s",
--
2.51.0
^ permalink raw reply related [flat|nested] 55+ messages in thread* Re: [PATCH RFC v1 00/20] KVM: x86: Support APX feature for guests
2025-11-10 18:01 [PATCH RFC v1 00/20] KVM: x86: Support APX feature for guests Chang S. Bae
` (19 preceding siblings ...)
2025-11-10 18:01 ` [PATCH RFC v1 20/20] KVM: selftests: Add APX state handling and XCR0 sanity checks Chang S. Bae
@ 2025-11-10 18:50 ` Chang S. Bae
2025-11-11 18:14 ` Paolo Bonzini
21 siblings, 0 replies; 55+ messages in thread
From: Chang S. Bae @ 2025-11-10 18:50 UTC (permalink / raw)
To: kvm, linux-kernel; +Cc: pbonzini, seanjc, chao.gao, zhao1.liu
On 11/10/2025 10:01 AM, Chang S. Bae wrote:
>
> This series is based on the next branch of the KVM x86 tree [3] and is
> available at:
>
> git://github.com/intel/apx.git apx-kvm_rfc-v1
...
> base-commit: e9a6fb0bcdd7609be6969112f3fbfcce3b1d4a7c
I don't think it really matters at this point. But I just rebased on
6.18-rc5 during last minutes.
^ permalink raw reply [flat|nested] 55+ messages in thread* Re: [PATCH RFC v1 00/20] KVM: x86: Support APX feature for guests
2025-11-10 18:01 [PATCH RFC v1 00/20] KVM: x86: Support APX feature for guests Chang S. Bae
` (20 preceding siblings ...)
2025-11-10 18:50 ` [PATCH RFC v1 00/20] KVM: x86: Support APX feature for guests Chang S. Bae
@ 2025-11-11 18:14 ` Paolo Bonzini
21 siblings, 0 replies; 55+ messages in thread
From: Paolo Bonzini @ 2025-11-11 18:14 UTC (permalink / raw)
To: Chang S. Bae, kvm, linux-kernel; +Cc: seanjc, chao.gao, zhao1.liu
On 11/10/25 19:01, Chang S. Bae wrote:
> Hi all,
>
> This series is intended to initiate the enablement of APX support in KVM.
> The goal is to start discussions and, ideally, reach consensus on key
> decisions for the enabling path.
>
> == Introduction ==
>
> Intel introduces Advanced Performance Extensions (APX), which add 16
> additional 64-bit general-purpose registers (EGPRs: R16–R31).
>
> APX also introduces new instruction prefixes that extend register indices
> to 5 bits, providing additional encoding space. The feature specification
> [1] describes these extensions in detail.
>
> The specification deliberately scopes out some areas. For example,
> Sections 3.1.4.4.2–7 note that initialization and reset behaviors follow
> existing XSTATE conventions.
>
> That said, there are still some essential elements to consider at first.
>
> == Ingredients to Consider ==
>
> With guest exposure, the new registers will affect how KVM handles VM
> exits and manages guest state. So the extended register indices need to
> be properly interpreted.
>
> There are three relevant contexts where KVM must handle the new 5-bit
> register index:
>
> * Instruction Information in VMCS
>
> VMX provides instruction information through a dedicated field in the
> VMCS. However, the legacy format supports only 4-bit register indices,
> so a new 64-bit field was introduced to support 5-bit indices for
> EGPRs (see Table 3.11 in the spec).
>
> * Exit Qualification in VMCS
>
> This field also carries register index information. Fortunately, the
> current layout is not fully packed, so it appears feasible to extend
> index fields to 5 bits without structural conflict.
>
> * Instruction Emulator
>
> When exits are handled through instruction emulation, the emulator
> must decode REX2-prefixed instructions, which carry the additional
> bits encoding the extended indices and other new modifiers.
>
> Once handlers can identify EGPR-related indices, the next question is how
> to access and maintain their state.
>
> * Extended GPR State Management
>
> Current KVM behavior for legacy GPRs can be summarized as follows:
> (a) All legacy GPRs are saved on every VM exit and written back on
> re-entry. Most access operations go through KVM register cache.
> (b) The instruction emulator also maintains a temporary GPR cache
> during emulation, backed by the same KVM-managed accessor
> interface.
>
> For the new EGPRs, there are a few important differences:
> (a) They are general-purpose registers, but their state is
> XSAVE-managed, which makes them distinct from legacy GPRs.
> (b) The kernel itself currently does not use them, so there is no
> requirement to save EGPRs on every VM exit -- they can remain in
> hardware registers.
>
> The mechanism to read and write EGPR state will therefore be commonly
> needed by both VMX handlers and the instruction emulator.
>
> == Approaches to Support ==
>
> Given the above aspects, the first step is to build out a generalized GPR
> accessor framework. This constitutes the first part (PART1: PATCH1–3) of
> the series, laying the foundational infrastructure.
>
> * New EGPR Accessors (PATCH3)
>
> Since EGPRs are not yet clobbered by the host, they can be accessed
> directly through hardware, using the existing helpers kvm_fpu_get()
> and kvm_fpu_put().
>
> This model follows the same approach used for legacy FP state
> handling.
>
> The design choice is based on the following considerations:
> (a) Caching EGPR state on every VM exit would be unnecessary cost.
> (b) If EGPRs are updated during VM exit handling or instruction
> emulation, synchronizing them with the guest fpstate would require
> new interfaces to interact with x86 core logic, adding complexity.
>
> * Common GPR Access Interface --Unifying Legacy and Extended Accessors
> (PATCH2)
>
> Although legacy GPRs and EGPRs differ in how their state is accessed,
> that distinction does not justify maintaining separate interfaces. A
> unified accessor abstraction allows both legacy and extended registers
> to be accessed uniformly, simplifying usage for both exit handlers and
> the emulator.
>
> Returning to the remaining key ingredients -- VMX handlers and the
> instruction emulator -- the necessary updates break down into the
> following.
>
> * Extended Instruction Information Extraction (PART2: PATCH4–8)
>
> Currently, instruction-related VMCS fields are stored as 32-bit raw
> data and interpreted on site. Adding separate variable in this manner
> would increase code complexity.
>
> Thus, refactoring this logic into a proper data structure with clear
> semantics appears to be a sensible direction.
>
> * Instruction Emulator (PART3: PATCH9–16)
>
> As noted in Paolo’s earlier feedback [2], support for REX2-prefixed
> instruction emulation is required.
>
> While REX2 largely mirrors legacy opcode behavior, a few exceptions
> and new instructions introduce additional decoding and handling
> requirements.
>
> Conceptually, the changes are straightforward, though details are better
> handled directly in the patches.
>
> Finally, actual feature exposure and XCR0 management form the last stage
> of enabling (PART4: PATCH17-20), relatively small but with a few key
> constraints:
>
> * XCR0 updates and CPUID feature exposure must occur together (PATCH18).
> * The current enabling scope applies only to VMX (PATCH17-18).
>
> == Patchset ==
>
> As mentioned earlier, while the number of patches is relatively large,
> the series is organized into four logical parts for clarity and
> reviewability:
>
> * PART1, PATCH 01–03: GPR accessor refactoring (foundational)
> * PART2, PATCH 04–08: VMX support for EGPR index handling
> * PART3, PATCH 09–16: Instruction emulator support for REX2
> * PART4, PATCH 17–20: APX exposure and minor selftest updates
>
> Each patch includes an RFC note to provide context for reviewers.
>
> This series is based on the next branch of the KVM x86 tree [3] and is
> available at:
>
> git://github.com/intel/apx.git apx-kvm_rfc-v1
>
> == Testing ==
>
> Testing so far has focused on unit and synthetic coverage of relevant
> code paths using KVM selftests, both on legacy and APX systems.
>
> For decoder-related changes, additional testing was performed by invoking
> x86_decode_insn() from a pseudo driver with some test inputs to exercise
> REX2 and legacy decoding logic.
>
> == References ==
>
> [1] https://cdrdv2.intel.com/v1/dl/getContent/784266
> [2] https://lore.kernel.org/kvm/CABgObfaHp9bH783Kdwm_tMBHZk5zWCxD7R+RroB_Q_o5NWBVZg@mail.gmail.com/
> [3] https://github.com/kvm-x86/linux
>
>
> I would also give thanks to Chao and Zhao for helping me out in this
> KVM series.
There are a few bugs, but most of my remarks are really just stylistic.
The next version, if you base it on the emulator AVX support, should be
pretty close.
Paolo
^ permalink raw reply [flat|nested] 55+ messages in thread