On Tue, Jan 10, 2023, Mingwei Zhang wrote:
> After tilerelease instruction, AMX tiles are in INIT state. According to
> Intel SDM vol 1. 13.10: "If RFBM[i] = 1, XSTATE_BV[i] is set to the
> value of XINUSE[i].", XSTATE_BV[18] should be cleared after xsavec.
>
> On the other hand, according to Intel SDM vol 1. 13.4.3: "If XCOMP_BV[i] =
> 1, state component i is located at a byte offset locationI from the base
> address of the XSAVE area". Since at the time of xsavec, XCR0[18] is set
> indicating AMX tile data component is still enabled, xcomp_bv[18] should be
> set.
>
> Complete the checks by adding the assert to xcomp_bv[18] after xsavec.
>
> Signed-off-by: Mingwei Zhang <mizhang@xxxxxxxxxx>
> ---
> tools/testing/selftests/kvm/x86_64/amx_test.c | 30 +++++++++++++++++--
> 1 file changed, 27 insertions(+), 3 deletions(-)
>
> diff --git a/tools/testing/selftests/kvm/x86_64/amx_test.c b/tools/testing/selftests/kvm/x86_64/amx_test.c
> index b2369f956fea..18203e399e9d 100644
> --- a/tools/testing/selftests/kvm/x86_64/amx_test.c
> +++ b/tools/testing/selftests/kvm/x86_64/amx_test.c
> @@ -41,6 +41,12 @@
>
> #define XSAVE_HDR_OFFSET 512
>
> +struct xstate_header {
> + u64 xfeatures;
> + u64 xcomp_bv;
> + u64 reserved[6];
> +} __packed;
I definitely like the idea of using a struct, but let's go all the way, i.e. add
a mostly-functional "struct xstate" too so that we don't to do pointer arithmetic.
I don't think it makes sense to copy+paste from the kernel since I highly doubt
anyone is going to write an x87 test, so maybe this?
struct xstate_header {
u64 xstate_bv;
u64 xcomp_bv;
u64 reserved[6];
} __attribute__((packed));
struct xstate {
u8 i387[512];
struct xstate_header header;
u8 extended_state_area[0];
} __attribute__ ((packed, aligned (64)));
> struct xsave_data {
> u8 area[XSAVE_SIZE];
Ewww. Not your code. The existing code declares XSAVE_SIZE bytes, but allocates
3 4KiB pages. It took me a bit of starting to realize TILE_SIZE is 1KiB, not 4KiB.
Can you tack on a patch do something like:
@@ -244,7 +230,7 @@ int main(int argc, char *argv[])
struct kvm_run *run;
struct kvm_x86_state *state;
int xsave_restore_size;
- vm_vaddr_t amx_cfg, tiledata, xsavedata;
+ vm_vaddr_t amx_cfg, tiledata, xstate;
struct ucall uc;
u32 amx_offset;
int stage, ret;
@@ -284,10 +270,10 @@ int main(int argc, char *argv[])
tiledata = vm_vaddr_alloc_pages(vm, 2);
memset(addr_gva2hva(vm, tiledata), rand() | 1, 2 * getpagesize());
- /* xsave data for guest_code */
- xsavedata = vm_vaddr_alloc_pages(vm, 3);
- memset(addr_gva2hva(vm, xsavedata), 0, 3 * getpagesize());
- vcpu_args_set(vcpu, 3, amx_cfg, tiledata, xsavedata);
+ /* XSAVE state for guest_code */
+ xstate = vm_vaddr_alloc_pages(vm, DIV_ROUND_UP(XSAVE_SIZE, PAGE_SIZE));
+ memset(addr_gva2hva(vm, xstate), 0, DIV_ROUND_UP(XSAVE_SIZE, PAGE_SIZE));
+ vcpu_args_set(vcpu, 3, amx_cfg, tiledata, xstate);
for (stage = 1; ; stage++) {
vcpu_run(vcpu);
> } __aligned(64);
> @@ -160,12 +166,26 @@ static void set_tilecfg(struct tile_config *cfg)
>
> static void set_xstatebv(void *data, uint64_t bv)
> {
> - *(uint64_t *)(data + XSAVE_HDR_OFFSET) = bv;
> + struct xstate_header *header =
> + (struct xstate_header *)(data + XSAVE_HDR_OFFSET);
> +
> + header->xfeatures = bv;
> }
>
> static u64 get_xstatebv(void *data)
> {
> - return *(u64 *)(data + XSAVE_HDR_OFFSET);
> + struct xstate_header *header =
> + (struct xstate_header *)(data + XSAVE_HDR_OFFSET);
> +
> + return header->xfeatures;
> +}
> +
> +static u64 get_xcompbv(void *data)
> +{
> + struct xstate_header *header =
> + (struct xstate_header *)(data + XSAVE_HDR_OFFSET);
> +
> + return header->xcomp_bv;
> }
If we add a "full" struct, these ugly wrappers go away, e.g. as untested patches
that you can claim as your own if you test 'em and write changelogs :-)
---
.../selftests/kvm/include/x86_64/processor.h | 12 +++++++
tools/testing/selftests/kvm/x86_64/amx_test.c | 36 ++++++-------------
2 files changed, 23 insertions(+), 25 deletions(-)
diff --git a/tools/testing/selftests/kvm/include/x86_64/processor.h b/tools/testing/selftests/kvm/include/x86_64/processor.h
index 53ffa43c90db..a7ce1fe8d70f 100644
--- a/tools/testing/selftests/kvm/include/x86_64/processor.h
+++ b/tools/testing/selftests/kvm/include/x86_64/processor.h
@@ -48,6 +48,18 @@ extern bool host_cpu_is_amd;
#define X86_CR4_SMAP (1ul << 21)
#define X86_CR4_PKE (1ul << 22)
+struct xstate_header {
+ u64 xstate_bv;
+ u64 xcomp_bv;
+ u64 reserved[6];
+} __attribute__((packed));
+
+struct xstate {
+ u8 i387[512];
+ struct xstate_header header;
+ u8 extended_state_area[0];
+} __attribute__ ((packed, aligned (64)));
+
/* Note, these are ordered alphabetically to match kvm_cpuid_entry2. Eww. */
enum cpuid_output_regs {
KVM_CPUID_EAX,
diff --git a/tools/testing/selftests/kvm/x86_64/amx_test.c b/tools/testing/selftests/kvm/x86_64/amx_test.c
index bd72c6eb3b67..d506821a5a26 100644
--- a/tools/testing/selftests/kvm/x86_64/amx_test.c
+++ b/tools/testing/selftests/kvm/x86_64/amx_test.c
@@ -41,10 +41,6 @@
#define XSAVE_HDR_OFFSET 512
-struct xsave_data {
- u8 area[XSAVE_SIZE];
-} __aligned(64);
-
struct tile_config {
u8 palette_id;
u8 start_row;
@@ -103,13 +99,13 @@ static inline void __tilerelease(void)
asm volatile(".byte 0xc4, 0xe2, 0x78, 0x49, 0xc0" ::);
}
-static inline void __xsavec(struct xsave_data *data, uint64_t rfbm)
+static inline void __xsavec(struct xstate *xstate, uint64_t rfbm)
{
uint32_t rfbm_lo = rfbm;
uint32_t rfbm_hi = rfbm >> 32;
asm volatile("xsavec (%%rdi)"
- : : "D" (data), "a" (rfbm_lo), "d" (rfbm_hi)
+ : : "D" (xstate), "a" (rfbm_lo), "d" (rfbm_hi)
: "memory");
}
@@ -158,16 +154,6 @@ static void set_tilecfg(struct tile_config *cfg)
}
}
-static void set_xstatebv(void *data, uint64_t bv)
-{
- *(uint64_t *)(data + XSAVE_HDR_OFFSET) = bv;
-}
-
-static u64 get_xstatebv(void *data)
-{
- return *(u64 *)(data + XSAVE_HDR_OFFSET);
-}
-
static void init_regs(void)
{
uint64_t cr4, xcr0;
@@ -184,7 +170,7 @@ static void init_regs(void)
static void __attribute__((__flatten__)) guest_code(struct tile_config *amx_cfg,
struct tile_data *tiledata,
- struct xsave_data *xsave_data)
+ struct xstate *xstate)
{
init_regs();
check_cpuid_xsave();
@@ -205,9 +191,9 @@ static void __attribute__((__flatten__)) guest_code(struct tile_config *amx_cfg,
__tilerelease();
GUEST_SYNC(5);
/* bit 18 not in the XCOMP_BV after xsavec() */
- set_xstatebv(xsave_data, XFEATURE_MASK_XTILEDATA);
- __xsavec(xsave_data, XFEATURE_MASK_XTILEDATA);
- GUEST_ASSERT((get_xstatebv(xsave_data) & XFEATURE_MASK_XTILEDATA) == 0);
+ xstate->header.xstate_bv = XFEATURE_MASK_XTILEDATA;
+ __xsavec(xstate, XFEATURE_MASK_XTILEDATA);
+ GUEST_ASSERT(!(xstate->header.xstate_bv & XFEATURE_MASK_XTILEDATA));
/* xfd=0x40000, disable amx tiledata */
wrmsr(MSR_IA32_XFD, XFEATURE_MASK_XTILEDATA);
@@ -244,7 +230,7 @@ int main(int argc, char *argv[])
struct kvm_run *run;
struct kvm_x86_state *state;
int xsave_restore_size;
- vm_vaddr_t amx_cfg, tiledata, xsavedata;
+ vm_vaddr_t amx_cfg, tiledata, xstate;
struct ucall uc;
u32 amx_offset;
int stage, ret;
@@ -284,10 +270,10 @@ int main(int argc, char *argv[])
tiledata = vm_vaddr_alloc_pages(vm, 2);
memset(addr_gva2hva(vm, tiledata), rand() | 1, 2 * getpagesize());
- /* xsave data for guest_code */
- xsavedata = vm_vaddr_alloc_pages(vm, 3);
- memset(addr_gva2hva(vm, xsavedata), 0, 3 * getpagesize());
- vcpu_args_set(vcpu, 3, amx_cfg, tiledata, xsavedata);
+ /* XSAVE state for guest_code */
+ xstate = vm_vaddr_alloc_pages(vm, DIV_ROUND_UP(XSAVE_SIZE, PAGE_SIZE));
+ memset(addr_gva2hva(vm, xstate), 0, DIV_ROUND_UP(XSAVE_SIZE, PAGE_SIZE));
+ vcpu_args_set(vcpu, 3, amx_cfg, tiledata, xstate);
for (stage = 1; ; stage++) {
vcpu_run(vcpu);
base-commit: 78332517a5dab54514ae719805eec218715de1fc
--
