Skip to contentSkip to conversation

Clean up AMX cpuid bits XTILE_CFG and XTILE_DATA

PATCH In review Aaron Lewis 2 participants 13 messages Dec 27, 2022 – Dec 28, 2022

Subsystems (4)
lore
Share & tools
lore
#0 lore raw

cover Clean up AMX cpuid bits XTILE_CFG and XTILE_DATA

When running a SPR guest without AMX enabled CPUID.(EAX=0DH,ECX=0):EAX.XTILE_CFG[bit-17] will be set and CPUID.(EAX=0DH,ECX=0):EAX.XTILE_DATA[bit-18] will be clear. While this is architecturally correct it can be a little awkward for userspace or a guest when using them. Instead of leaving the CPUID leaf in a half baked state, either clear them both or leave them both set.

Additionally, add testing to verify the CPUID isn't in such a state.

Aaron Lewis (3):
KVM: x86: Clear XTILE_CFG if XTILE_DATA is clear
KVM: selftests: Hoist XGETBV and XSETBV to make them more accessible
KVM: selftests: Add XCR0 Test

 arch/x86/kvm/cpuid.c                          |   4 +
 tools/testing/selftests/kvm/Makefile          |   1 +
 .../selftests/kvm/include/x86_64/processor.h  |  19 +++
 tools/testing/selftests/kvm/x86_64/amx_test.c |  24 +---
 .../selftests/kvm/x86_64/xcr0_cpuid_test.c    | 134 ++++++++++++++++++
 5 files changed, 161 insertions(+), 21 deletions(-)
 create mode 100644 tools/testing/selftests/kvm/x86_64/xcr0_cpuid_test.c
Signature · 1 line
2.39.0.314.g84b9a713c41-goog
#1 lore raw

PATCH 1/3 KVM: x86: Clear XTILE_CFG if XTILE_DATA is clear

Be good citizens by clearing both CPUID.(EAX=0DH,ECX=0):EAX.XTILECFG[bit-17] and CPUID.(EAX=0DH,ECX=0):EAX.XTILEDATA[bit-18] if they are both not set. That way userspace or a guest doesn't fail if it attempts to set XCR0 with the user xfeature bits, i.e. EDX:EAX of CPUID.(EAX=0DH,ECX=0).

---
 arch/x86/kvm/cpuid.c | 4 ++++
 1 file changed, 4 insertions(+)
arch/x86/kvm/cpuid.c +4 -0
index 0b5bf013fcb8e..2d9910847786a 100644
@@ -977,6 +977,10 @@ static inline int __do_cpuid_func(struct kvm_cpuid_array *array, u32 function)
 		u64 permitted_xcr0 = kvm_caps.supported_xcr0 & xstate_get_guest_group_perm();
 		u64 permitted_xss = kvm_caps.supported_xss;
 
+		if (!(permitted_xcr0 & XFEATURE_MASK_XTILE_CFG) ||
+		    !(permitted_xcr0 & XFEATURE_MASK_XTILE_DATA))
+			permitted_xcr0 &= ~XFEATURE_MASK_XTILE;
+
 		entry->eax &= permitted_xcr0;
 		entry->ebx = xstate_required_size(permitted_xcr0, false);
 		entry->ecx = entry->ebx;
Signature · 1 line
2.39.0.314.g84b9a713c41-goog
#2 lore raw

Re: [PATCH 1/3] KVM: x86: Clear XTILE_CFG if XTILE_DATA is clear

Quoted 29 lines · Aaron Lewis

Be good citizens by clearing both CPUID.(EAX=0DH,ECX=0):EAX.XTILECFG[bit-17] and CPUID.(EAX=0DH,ECX=0):EAX.XTILEDATA[bit-18] if they are both not set. That way userspace or a guest doesn't fail if it attempts to set XCR0 with the user xfeature bits, i.e. EDX:EAX of CPUID.(EAX=0DH,ECX=0).

---
 arch/x86/kvm/cpuid.c | 4 ++++
 1 file changed, 4 insertions(+)
arch/x86/kvm/cpuid.c +4 -1
index 0b5bf013fcb8e..2d9910847786a 100644
@@ -977,6 +977,10 @@ static inline int __do_cpuid_func(struct kvm_cpuid_array *array, u32 function)
                u64 permitted_xcr0 = kvm_caps.supported_xcr0 & xstate_get_guest_group_perm();
                u64 permitted_xss = kvm_caps.supported_xss;
 
+               if (!(permitted_xcr0 & XFEATURE_MASK_XTILE_CFG) ||
+                   !(permitted_xcr0 & XFEATURE_MASK_XTILE_DATA))
+                       permitted_xcr0 &= ~XFEATURE_MASK_XTILE;
+
                entry->eax &= permitted_xcr0;
                entry->ebx = xstate_required_size(permitted_xcr0, false);
                entry->ecx = entry->ebx;
--
Quoted 2 lines

2.39.0.314.g84b9a713c41-goog

Two questions:

1) Under what circumstances would this happen?
2) Shouldn't we also clear XFEATURE_MASK_CFG if both bits are not set?

#3 lore raw

Re: [PATCH 1/3] KVM: x86: Clear XTILE_CFG if XTILE_DATA is clear

Quoted 21 lines
Quoted 16 lines
arch/x86/kvm/cpuid.c +4 -1
index 0b5bf013fcb8e..2d9910847786a 100644
@@ -977,6 +977,10 @@ static inline int __do_cpuid_func(struct kvm_cpuid_array *array, u32 function)
                u64 permitted_xcr0 = kvm_caps.supported_xcr0 & xstate_get_guest_group_perm();
                u64 permitted_xss = kvm_caps.supported_xss;
 
+               if (!(permitted_xcr0 & XFEATURE_MASK_XTILE_CFG) ||
+                   !(permitted_xcr0 & XFEATURE_MASK_XTILE_DATA))
+                       permitted_xcr0 &= ~XFEATURE_MASK_XTILE;
+
                entry->eax &= permitted_xcr0;
                entry->ebx = xstate_required_size(permitted_xcr0, false);
                entry->ecx = entry->ebx;
--
Quoted 2 lines

2.39.0.314.g84b9a713c41-goog

Two questions:

Quoted 1 line

1) Under what circumstances would this happen?

This would happen if userspace hasn't opted in to using AMX via arch_prctl().

Quoted 1 line

2) Shouldn't we also clear XFEATURE_MASK_CFG if both bits are not set?

Both CFG and DATA are cleared with XFEATURE_MASK_XTILE. It defines both.

Reviewed-by
#4 lore raw

Re: [PATCH 1/3] KVM: x86: Clear XTILE_CFG if XTILE_DATA is clear

Quoted 22 lines · Aaron Lewis
Quoted 22 lines
Quoted 17 lines
arch/x86/kvm/cpuid.c +4 -1
index 0b5bf013fcb8e..2d9910847786a 100644
@@ -977,6 +977,10 @@ static inline int __do_cpuid_func(struct kvm_cpuid_array *array, u32 function)
                u64 permitted_xcr0 = kvm_caps.supported_xcr0 & xstate_get_guest_group_perm();
                u64 permitted_xss = kvm_caps.supported_xss;
 
+               if (!(permitted_xcr0 & XFEATURE_MASK_XTILE_CFG) ||
+                   !(permitted_xcr0 & XFEATURE_MASK_XTILE_DATA))
+                       permitted_xcr0 &= ~XFEATURE_MASK_XTILE;
+
                entry->eax &= permitted_xcr0;
                entry->ebx = xstate_required_size(permitted_xcr0, false);
                entry->ecx = entry->ebx;
--
Quoted 2 lines

2.39.0.314.g84b9a713c41-goog

Two questions:

Quoted 5 lines
Quoted 1 line

1) Under what circumstances would this happen?

This would happen if userspace hasn't opted in to using AMX via arch_prctl().

Quoted 1 line

2) Shouldn't we also clear XFEATURE_MASK_CFG if both bits are not set?

Both CFG and DATA are cleared with XFEATURE_MASK_XTILE. It defines both.

Doh!

#5 lore raw

Re: [PATCH 1/3] KVM: x86: Clear XTILE_CFG if XTILE_DATA is clear

Quoted 1 line · Jim Mattson

Could you change the shortlog to clarify that both feature bits are cleared if either one is clear?

#6 lore raw

PATCH 2/3 KVM: selftests: Hoist XGETBV and XSETBV to make them more accessible

In reply to Aaron Lewis

The instructions XGETBV and XSETBV are useful to other tests. Move them to processor.h to make them available to be used more broadly.

No functional change intended.

---
 .../selftests/kvm/include/x86_64/processor.h  | 19 +++++++++++++++
 tools/testing/selftests/kvm/x86_64/amx_test.c | 24 +++----------------
 2 files changed, 22 insertions(+), 21 deletions(-)
tools/testing/selftests/kvm/include/x86_64/processor.h +19 -0
index b1a31de7108ac..34957137be375 100644
@@ -492,6 +492,25 @@ static inline void set_cr4(uint64_t val)
 	__asm__ __volatile__("mov %0, %%cr4" : : "r" (val) : "memory");
 }
 
+static inline u64 xgetbv(u32 index)
+{
+	u32 eax, edx;
+
+	__asm__ __volatile__("xgetbv;"
+		     : "=a" (eax), "=d" (edx)
+		     : "c" (index));
+	return eax | ((u64)edx << 32);
+}
+
+static inline void xsetbv(u32 index, u64 value)
+{
+	u32 eax = value;
+	u32 edx = value >> 32;
+
+	__asm__ __volatile__("xsetbv" :: "a" (eax), "d" (edx), "c" (index));
+}
+
+
 static inline struct desc_ptr get_gdt(void)
 {
 	struct desc_ptr gdt;
tools/testing/selftests/kvm/x86_64/amx_test.c +3 -21
index bd72c6eb3b670..4b733ad218313 100644
@@ -68,24 +68,6 @@ struct xtile_info {
 
 static struct xtile_info xtile;
 
-static inline u64 __xgetbv(u32 index)
-{
-	u32 eax, edx;
-
-	asm volatile("xgetbv;"
-		     : "=a" (eax), "=d" (edx)
-		     : "c" (index));
-	return eax + ((u64)edx << 32);
-}
-
-static inline void __xsetbv(u32 index, u64 value)
-{
-	u32 eax = value;
-	u32 edx = value >> 32;
-
-	asm volatile("xsetbv" :: "a" (eax), "d" (edx), "c" (index));
-}
-
 static inline void __ldtilecfg(void *cfg)
 {
 	asm volatile(".byte 0xc4,0xe2,0x78,0x49,0x00"
@@ -121,7 +103,7 @@ static inline void check_cpuid_xsave(void)
 
 static bool check_xsave_supports_xtile(void)
 {
-	return __xgetbv(0) & XFEATURE_MASK_XTILE;
+	return xgetbv(0) & XFEATURE_MASK_XTILE;
 }
 
 static void check_xtile_info(void)
@@ -177,9 +159,9 @@ static void init_regs(void)
 	cr4 |= X86_CR4_OSXSAVE;
 	set_cr4(cr4);
 
-	xcr0 = __xgetbv(0);
+	xcr0 = xgetbv(0);
 	xcr0 |= XFEATURE_MASK_XTILE;
-	__xsetbv(0x0, xcr0);
+	xsetbv(0x0, xcr0);
 }
 
 static void __attribute__((__flatten__)) guest_code(struct tile_config *amx_cfg,
Signature · 1 line
2.39.0.314.g84b9a713c41-goog
Reviewed-by on v1 2/3
#7 lore raw

Re: [PATCH 2/3] KVM: selftests: Hoist XGETBV and XSETBV to make them more accessible

Quoted 29 lines · Aaron Lewis

The instructions XGETBV and XSETBV are useful to other tests. Move them to processor.h to make them available to be used more broadly.

No functional change intended.

---
 .../selftests/kvm/include/x86_64/processor.h  | 19 +++++++++++++++
 tools/testing/selftests/kvm/x86_64/amx_test.c | 24 +++----------------
 2 files changed, 22 insertions(+), 21 deletions(-)
tools/testing/selftests/kvm/include/x86_64/processor.h +9 -0
index b1a31de7108ac..34957137be375 100644
@@ -492,6 +492,25 @@ static inline void set_cr4(uint64_t val)
        __asm__ __volatile__("mov %0, %%cr4" : : "r" (val) : "memory");
 }
 
+static inline u64 xgetbv(u32 index)
+{
+       u32 eax, edx;
+
+       __asm__ __volatile__("xgetbv;"
+                    : "=a" (eax), "=d" (edx)
+                    : "c" (index));
+       return eax | ((u64)edx << 32);
+}

tools/testing/selftests/kvm/include/x86_64/processor.h static inline void set_cr4(uint64_t val) · line 513

Quoted 10 lines
tools/testing/selftests/kvm/include/x86_64/processor.h +10 -0
+
+static inline void xsetbv(u32 index, u64 value)
+{
+       u32 eax = value;
+       u32 edx = value >> 32;
+
+       __asm__ __volatile__("xsetbv" :: "a" (eax), "d" (edx), "c" (index));
+}
+
+

Not your change, but shouldn't both of these asm statements have artificial "memory" clobbers, to prevent reordering?

#8 lore raw

PATCH 3/3 KVM: selftests: Add XCR0 Test

In reply to Aaron Lewis

Test that the user xfeature bits, EDX:EAX of CPUID.(EAX=0DH,ECX=0), don't set up userspace for failure.

Though it isn't architectural, test that the user xfeature bits aren't set in a half baked state that will cause a #GP if used when setting XCR0.

Check that the rules for XCR0 described in the SDM vol 1, section 13.3 ENABLING THE XSAVE FEATURE SET AND XSAVE-ENABLED FEATURES, are followed for the xfeature bits too.

---
 tools/testing/selftests/kvm/Makefile          |   1 +
 .../selftests/kvm/x86_64/xcr0_cpuid_test.c    | 134 ++++++++++++++++++
 2 files changed, 135 insertions(+)
 create mode 100644 tools/testing/selftests/kvm/x86_64/xcr0_cpuid_test.c
tools/testing/selftests/kvm/Makefile +1 -0
index 1750f91dd9362..e2e56c82b8a90 100644
@@ -104,6 +104,7 @@ TEST_GEN_PROGS_x86_64 += x86_64/vmx_tsc_adjust_test
 TEST_GEN_PROGS_x86_64 += x86_64/vmx_nested_tsc_scaling_test
 TEST_GEN_PROGS_x86_64 += x86_64/xapic_ipi_test
 TEST_GEN_PROGS_x86_64 += x86_64/xapic_state_test
+TEST_GEN_PROGS_x86_64 += x86_64/xcr0_cpuid_test
 TEST_GEN_PROGS_x86_64 += x86_64/xss_msr_test
 TEST_GEN_PROGS_x86_64 += x86_64/debug_regs
 TEST_GEN_PROGS_x86_64 += x86_64/tsc_msrs_test
tools/testing/selftests/kvm/x86_64/xcr0_cpuid_test.c new file +134 -0

134 lines not on this page · raw message

+}

Signature · 1 line
2.39.0.314.g84b9a713c41-goog
#9 lore raw

Re: [PATCH 3/3] KVM: selftests: Add XCR0 Test

Quoted 56 lines · Aaron Lewis

Test that the user xfeature bits, EDX:EAX of CPUID.(EAX=0DH,ECX=0), don't set up userspace for failure.

Though it isn't architectural, test that the user xfeature bits aren't set in a half baked state that will cause a #GP if used when setting XCR0.

Check that the rules for XCR0 described in the SDM vol 1, section 13.3 ENABLING THE XSAVE FEATURE SET AND XSAVE-ENABLED FEATURES, are followed for the xfeature bits too.

---
 tools/testing/selftests/kvm/Makefile          |   1 +
 .../selftests/kvm/x86_64/xcr0_cpuid_test.c    | 134 ++++++++++++++++++
 2 files changed, 135 insertions(+)
 create mode 100644 tools/testing/selftests/kvm/x86_64/xcr0_cpuid_test.c
tools/testing/selftests/kvm/Makefile +1 -0
index 1750f91dd9362..e2e56c82b8a90 100644
@@ -104,6 +104,7 @@ TEST_GEN_PROGS_x86_64 += x86_64/vmx_tsc_adjust_test
 TEST_GEN_PROGS_x86_64 += x86_64/vmx_nested_tsc_scaling_test
 TEST_GEN_PROGS_x86_64 += x86_64/xapic_ipi_test
 TEST_GEN_PROGS_x86_64 += x86_64/xapic_state_test
+TEST_GEN_PROGS_x86_64 += x86_64/xcr0_cpuid_test
 TEST_GEN_PROGS_x86_64 += x86_64/xss_msr_test
 TEST_GEN_PROGS_x86_64 += x86_64/debug_regs
 TEST_GEN_PROGS_x86_64 += x86_64/tsc_msrs_test
tools/testing/selftests/kvm/x86_64/xcr0_cpuid_test.c new file +19 -0
new file mode 100644
index 0000000000000..644791ff5c48b
@@ -0,0 +1,134 @@
+// SPDX-License-Identifier: GPL-2.0
+/*
+ * XCR0 cpuid test
+ *
+ * Copyright (C) 2022, Google LLC.
+ */
+
+#include <fcntl.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include <sys/ioctl.h>
+
+#include "test_util.h"
+
+#include "kvm_util.h"
+#include "processor.h"
+
+#define XFEATURE_MASK_SSE              (1ul << 1)

tools/testing/selftests/kvm/x86_64/xcr0_cpuid_test.c line 29

Quoted 10 lines
tools/testing/selftests/kvm/x86_64/xcr0_cpuid_test.c new file +10 -0
+#define XFEATURE_MASK_YMM              (1ul << 2)
+#define XFEATURE_MASK_BNDREGS          (1ul << 3)
+#define XFEATURE_MASK_BNDCSR           (1ul << 4)
+#define XFEATURE_MASK_OPMASK           (1ul << 5)
+#define XFEATURE_MASK_ZMM_Hi256                (1ul << 6)
+#define XFEATURE_MASK_Hi16_ZMM         (1ul << 7)
+#define XFEATURE_MASK_XTILECFG         (1ul << 17)
+#define XFEATURE_MASK_XTILEDATA                (1ul << 18)
+#define XFEATURE_MASK_XTILE            (XFEATURE_MASK_XTILECFG | \
+                                        XFEATURE_MASK_XTILEDATA)

With XSETBV hoisted into processor.h, shouldn't these macros be more widely available as well?

Quoted 37 lines
tools/testing/selftests/kvm/x86_64/xcr0_cpuid_test.c new file +37 -0
+static uint64_t get_supported_user_xfeatures(void)
+{
+       uint32_t a, b, c, d;
+
+       cpuid(0xd, &a, &b, &c, &d);
+
+       return a | ((uint64_t)d << 32);
+}
+
+static void guest_code(void)
+{
+       uint64_t xcr0_rest;
+       uint64_t supported_xcr0;
+       uint64_t xfeature_mask;
+       uint64_t supported_state;
+
+       set_cr4(get_cr4() | X86_CR4_OSXSAVE);
+
+       xcr0_rest = xgetbv(0);
+       supported_xcr0 = get_supported_user_xfeatures();
+
+       GUEST_ASSERT(xcr0_rest == 1ul);
+
+       /* Check AVX */
+       xfeature_mask = XFEATURE_MASK_SSE | XFEATURE_MASK_YMM;
+       supported_state = supported_xcr0 & xfeature_mask;
+       GUEST_ASSERT(supported_state != XFEATURE_MASK_YMM);
+
+       /* Check MPX */
+       xfeature_mask = XFEATURE_MASK_BNDREGS | XFEATURE_MASK_BNDCSR;
+       supported_state = supported_xcr0 & xfeature_mask;
+       GUEST_ASSERT((supported_state == xfeature_mask) ||
+                    (supported_state == 0ul));
+
+       /* Check AVX-512 */
+       xfeature_mask = XFEATURE_MASK_OPMASK |
+                       XFEATURE_MASK_ZMM_Hi256 |

tools/testing/selftests/kvm/x86_64/xcr0_cpuid_test.c line 77

Quoted 10 lines
tools/testing/selftests/kvm/x86_64/xcr0_cpuid_test.c new file +10 -0
+                       XFEATURE_MASK_Hi16_ZMM;
+       supported_state = supported_xcr0 & xfeature_mask;
+       GUEST_ASSERT((supported_state == xfeature_mask) ||
+                    (supported_state == 0ul));
+
+       /* Check AMX */
+       xfeature_mask = XFEATURE_MASK_XTILE;
+       supported_state = supported_xcr0 & xfeature_mask;
+       GUEST_ASSERT((supported_state == xfeature_mask) ||
+                    (supported_state == 0ul));

In this series, you've added code to __do_cpuid_func() to ensure that XFEATURE_MASK_XTILECFG and XFEATURE_MASK_XTILEDATA can't be set unless the other is set. Do we need to do something similar for AVX-512 and MPX?

Quoted 60 lines
tools/testing/selftests/kvm/x86_64/xcr0_cpuid_test.c new file +57 -1
+       GUEST_SYNC(0);
+
+       xsetbv(0, supported_xcr0);
+
+       GUEST_DONE();
+}
+
+static void guest_gp_handler(struct ex_regs *regs)
+{
+       GUEST_ASSERT(!"Failed to set the supported xfeature bits in XCR0.");
+}
+
+int main(int argc, char *argv[])
+{
+       struct kvm_vcpu *vcpu;
+       struct kvm_run *run;
+       struct kvm_vm *vm;
+       struct ucall uc;
+
+       TEST_REQUIRE(kvm_cpu_has(X86_FEATURE_XSAVE));
+
+       /* Tell stdout not to buffer its content */
+       setbuf(stdout, NULL);
+
+       vm = vm_create_with_one_vcpu(&vcpu, guest_code);
+       run = vcpu->run;
+
+       vm_init_descriptor_tables(vm);
+       vcpu_init_descriptor_tables(vcpu);
+
+       while (1) {
+               vcpu_run(vcpu);
+
+               TEST_ASSERT(run->exit_reason == KVM_EXIT_IO,
+                           "Unexpected exit reason: %u (%s),\n",
+                           run->exit_reason,
+                           exit_reason_str(run->exit_reason));
+
+               switch (get_ucall(vcpu, &uc)) {
+               case UCALL_SYNC:
+                       vm_install_exception_handler(vm, GP_VECTOR,
+                                                    guest_gp_handler);
+                       break;
+               case UCALL_ABORT:
+                       REPORT_GUEST_ASSERT(uc);
+                       break;
+               case UCALL_DONE:
+                       goto done;
+               default:
+                       TEST_FAIL("Unknown ucall %lu", uc.cmd);
+               }
+       }
+
+done:
+       kvm_vm_free(vm);
+       return 0;
+}
--

2.39.0.314.g84b9a713c41-goog

#10 lore raw

Re: [PATCH 3/3] KVM: selftests: Add XCR0 Test

Quoted 9 lines
Quoted 9 lines
tools/testing/selftests/kvm/x86_64/xcr0_cpuid_test.c new file +9 -0
+#include "kvm_util.h"
+#include "processor.h"
+
+#define XFEATURE_MASK_SSE              (1ul << 1)
+#define XFEATURE_MASK_YMM              (1ul << 2)
+#define XFEATURE_MASK_BNDREGS          (1ul << 3)
+#define XFEATURE_MASK_BNDCSR           (1ul << 4)
+#define XFEATURE_MASK_OPMASK           (1ul << 5)
+#define XFEATURE_MASK_ZMM_Hi256                (1ul << 6)
Quoted 8 lines

tools/testing/selftests/kvm/x86_64/xcr0_cpuid_test.c line 29

Quoted 5 lines
tools/testing/selftests/kvm/x86_64/xcr0_cpuid_test.c new file +5 -0
+#define XFEATURE_MASK_Hi16_ZMM         (1ul << 7)
+#define XFEATURE_MASK_XTILECFG         (1ul << 17)
+#define XFEATURE_MASK_XTILEDATA                (1ul << 18)
+#define XFEATURE_MASK_XTILE            (XFEATURE_MASK_XTILECFG | \
+                                        XFEATURE_MASK_XTILEDATA)

With XSETBV hoisted into processor.h, shouldn't these macros be more widely available as well?

Sure, I'll hoist them up there too.

Quoted 33 lines
Quoted 33 lines
tools/testing/selftests/kvm/x86_64/xcr0_cpuid_test.c new file +32 -0
+static uint64_t get_supported_user_xfeatures(void)
+{
+       uint32_t a, b, c, d;
+
+       cpuid(0xd, &a, &b, &c, &d);
+
+       return a | ((uint64_t)d << 32);
+}
+
+       GUEST_ASSERT(xcr0_rest == 1ul);
+
+       /* Check AVX */
+       xfeature_mask = XFEATURE_MASK_SSE | XFEATURE_MASK_YMM;
+       supported_state = supported_xcr0 & xfeature_mask;
+       GUEST_ASSERT(supported_state != XFEATURE_MASK_YMM);
+
+       /* Check MPX */
+       xfeature_mask = XFEATURE_MASK_BNDREGS | XFEATURE_MASK_BNDCSR;
+       supported_state = supported_xcr0 & xfeature_mask;
+       GUEST_ASSERT((supported_state == xfeature_mask) ||
+                    (supported_state == 0ul));
+
+       /* Check AVX-512 */
+       xfeature_mask = XFEATURE_MASK_OPMASK |
+                       XFEATURE_MASK_ZMM_Hi256 |
+                       XFEATURE_MASK_Hi16_ZMM;
+       supported_state = supported_xcr0 & xfeature_mask;
+       GUEST_ASSERT((supported_state == xfeature_mask) ||
+                    (supported_state == 0ul));
+
+       /* Check AMX */
+       xfeature_mask = XFEATURE_MASK_XTILE;
Quoted 8 lines

tools/testing/selftests/kvm/x86_64/xcr0_cpuid_test.c line 63

Quoted 3 lines
tools/testing/selftests/kvm/x86_64/xcr0_cpuid_test.c new file +3 -0
+       supported_state = supported_xcr0 & xfeature_mask;
+       GUEST_ASSERT((supported_state == xfeature_mask) ||
+                    (supported_state == 0ul));

In this series, you've added code to __do_cpuid_func() to ensure that XFEATURE_MASK_XTILECFG and XFEATURE_MASK_XTILEDATA can't be set unless the other is set. Do we need to do something similar for AVX-512 and MPX?

That does sound like a natural extension of this. I can add support for that in v2.

Quoted 13 lines
Quoted 13 lines
tools/testing/selftests/kvm/x86_64/xcr0_cpuid_test.c new file +12 -0
+       GUEST_SYNC(0);
+
+       xsetbv(0, supported_xcr0);
+
+       GUEST_DONE();
+}
+
+static void guest_gp_handler(struct ex_regs *regs)
+{
+       GUEST_ASSERT(!"Failed to set the supported xfeature bits in XCR0.");
+}
+
#11 lore raw
In reply to Aaron Lewis
Quoted 2 lines · Aaron Lewis

When running a SPR guest without AMX enabled

Can you clarify what "without AMX enabled" means? Do you mean that userspace hasn't opted in to AMX via arch_prctl()?

Quoted 24 lines

CPUID.(EAX=0DH,ECX=0):EAX.XTILE_CFG[bit-17] will be set and CPUID.(EAX=0DH,ECX=0):EAX.XTILE_DATA[bit-18] will be clear. While this is architecturally correct it can be a little awkward for userspace or a guest when using them. Instead of leaving the CPUID leaf in a half baked state, either clear them both or leave them both set.

Additionally, add testing to verify the CPUID isn't in such a state.

Aaron Lewis (3):
KVM: x86: Clear XTILE_CFG if XTILE_DATA is clear
KVM: selftests: Hoist XGETBV and XSETBV to make them more accessible
KVM: selftests: Add XCR0 Test

 arch/x86/kvm/cpuid.c                          |   4 +
 tools/testing/selftests/kvm/Makefile          |   1 +
 .../selftests/kvm/include/x86_64/processor.h  |  19 +++
 tools/testing/selftests/kvm/x86_64/amx_test.c |  24 +---
 .../selftests/kvm/x86_64/xcr0_cpuid_test.c    | 134 ++++++++++++++++++
 5 files changed, 161 insertions(+), 21 deletions(-)
 create mode 100644 tools/testing/selftests/kvm/x86_64/xcr0_cpuid_test.c

--
2.39.0.314.g84b9a713c41-goog

#12 lore raw
Quoted 8 lines · Jim Mattson
Quoted 2 lines · Aaron Lewis

When running a SPR guest without AMX enabled

Can you clarify what "without AMX enabled" means? Do you mean that userspace hasn't opted in to AMX via arch_prctl()?

Yes, exactly. I'll clarify in the follow-up.