Re: [PATCH] x86_64, asm: Work around AMD SYSRET SS descriptor attribute issue
From: Borislav Petkov
Date: Sun May 03 2015 - 07:51:48 EST
On Thu, Apr 30, 2015 at 02:39:07PM -0700, H. Peter Anvin wrote:
> This is the microbenchmark I used.
>
> For the record, Intel's intention going forward is that 0F 1F will
> always be as fast or faster than any other alternative.
It looks like this is the case on AMD too.
So I took your benchmark and made it to measure all sizes of K8 and P6
NOPs. Also I'm doing 10^6 iterations and taking the minimum. The results
speak for themselves, especially from 5-byte NOPs onwards where we have
to repeat the K8 NOP but still can use a single P6 NOP.
And I'm going to move all relevant AMD hw to use the P6 NOPs for the
alternatives.
Unless I've done something wrong, of course. Please double-check, I'm
attaching the microbenchmark too.
Anyway, here's a patch:
---
From: Borislav Petkov <bp@xxxxxxx>
Date: Sat, 2 May 2015 23:55:40 +0200
Subject: [PATCH] x86/alternatives: Switch AMD F15h and later to the P6 NOPs
Software optimization guides for both F15h and F16h cite those NOPs as
the optimal ones. A microbenchmark confirms that actually even older
families are better with the single-insn NOPs so switch to them for the
alternatives.
Cycles count below includes the loop overhead of the measurement but
that overhead is the same with all runs.
F10h, revE:
-----------
Running NOP tests, 1000 NOPs x 1000000 repetitions
K8:
90 288.212282 cycles
66 90 288.220840 cycles
66 66 90 288.219447 cycles
66 66 66 90 288.223204 cycles
66 66 90 66 90 571.393424 cycles
66 66 90 66 66 90 571.374919 cycles
66 66 66 90 66 66 90 572.249281 cycles
66 66 66 90 66 66 66 90 571.388651 cycles
P6:
90 288.214193 cycles
66 90 288.225550 cycles
0f 1f 00 288.224441 cycles
0f 1f 40 00 288.225030 cycles
0f 1f 44 00 00 288.233558 cycles
66 0f 1f 44 00 00 324.792342 cycles
0f 1f 80 00 00 00 00 325.657462 cycles
0f 1f 84 00 00 00 00 00 430.246643 cycles
F14h:
----
Running NOP tests, 1000 NOPs x 1000000 repetitions
K8:
90 510.404890 cycles
66 90 510.432117 cycles
66 66 90 510.561858 cycles
66 66 66 90 510.541865 cycles
66 66 90 66 90 1014.192782 cycles
66 66 90 66 66 90 1014.226546 cycles
66 66 66 90 66 66 90 1014.334299 cycles
66 66 66 90 66 66 66 90 1014.381205 cycles
P6:
90 510.436710 cycles
66 90 510.448229 cycles
0f 1f 00 510.545100 cycles
0f 1f 40 00 510.502792 cycles
0f 1f 44 00 00 510.589517 cycles
66 0f 1f 44 00 00 510.611462 cycles
0f 1f 80 00 00 00 00 511.166794 cycles
0f 1f 84 00 00 00 00 00 511.651641 cycles
F15h:
-----
Running NOP tests, 1000 NOPs x 1000000 repetitions
K8:
90 243.128396 cycles
66 90 243.129883 cycles
66 66 90 243.131631 cycles
66 66 66 90 242.499324 cycles
66 66 90 66 90 481.829083 cycles
66 66 90 66 66 90 481.884413 cycles
66 66 66 90 66 66 90 481.851446 cycles
66 66 66 90 66 66 66 90 481.409220 cycles
P6:
90 243.127026 cycles
66 90 243.130711 cycles
0f 1f 00 243.122747 cycles
0f 1f 40 00 242.497617 cycles
0f 1f 44 00 00 245.354461 cycles
66 0f 1f 44 00 00 361.930417 cycles
0f 1f 80 00 00 00 00 362.844944 cycles
0f 1f 84 00 00 00 00 00 480.514948 cycles
F16h:
-----
Running NOP tests, 1000 NOPs x 1000000 repetitions
K8:
90 507.793298 cycles
66 90 507.789636 cycles
66 66 90 507.826490 cycles
66 66 66 90 507.859075 cycles
66 66 90 66 90 1008.663129 cycles
66 66 90 66 66 90 1008.696259 cycles
66 66 66 90 66 66 90 1008.692517 cycles
66 66 66 90 66 66 66 90 1008.755399 cycles
P6:
90 507.795232 cycles
66 90 507.794761 cycles
0f 1f 00 507.834901 cycles
0f 1f 40 00 507.822629 cycles
0f 1f 44 00 00 507.838493 cycles
66 0f 1f 44 00 00 507.908597 cycles
0f 1f 80 00 00 00 00 507.946417 cycles
0f 1f 84 00 00 00 00 00 507.954960 cycles
Signed-off-by: Borislav Petkov <bp@xxxxxxx>
Cc: Aravind Gopalakrishnan <aravind.gopalakrishnan@xxxxxxx>
---
arch/x86/kernel/alternative.c | 9 +++++++++
1 file changed, 9 insertions(+)
diff --git a/arch/x86/kernel/alternative.c b/arch/x86/kernel/alternative.c
index aef653193160..b0932c4341b3 100644
--- a/arch/x86/kernel/alternative.c
+++ b/arch/x86/kernel/alternative.c
@@ -227,6 +227,15 @@ void __init arch_init_ideal_nops(void)
#endif
}
break;
+
+ case X86_VENDOR_AMD:
+ if (boot_cpu_data.x86 > 0xf) {
+ ideal_nops = p6_nops;
+ return;
+ }
+
+ /* fall through */
+
default:
#ifdef CONFIG_X86_64
ideal_nops = k8_nops;
--
2.3.5
Modified benchmark:
---
#define _GNU_SOURCE
#include <stdio.h>
#include <stdlib.h>
#include <time.h>
#include <stdbool.h>
#include <sys/time.h>
typedef unsigned long long u64;
#define DECLARE_ARGS(val, low, high) unsigned low, high
#define EAX_EDX_VAL(val, low, high) ((low) | ((u64)(high) << 32))
#define EAX_EDX_ARGS(val, low, high) "a" (low), "d" (high)
#define EAX_EDX_RET(val, low, high) "=a" (low), "=d" (high)
static __always_inline unsigned long long rdtsc(void)
{
DECLARE_ARGS(val, low, high);
asm volatile("rdtsc" : EAX_EDX_RET(val, low, high));
return EAX_EDX_VAL(val, low, high);
}
static inline u64 read_tsc(void)
{
u64 ret;
asm volatile("mfence");
ret = rdtsc();
asm volatile("mfence");
return ret;
}
#define __stringify_1(x...) #x
#define __stringify(x...) __stringify_1(x)
#define GENERIC_NOP1 0x90
#define K8_NOP1 GENERIC_NOP1
#define K8_NOP2 0x66,K8_NOP1
#define K8_NOP3 0x66,K8_NOP2
#define K8_NOP4 0x66,K8_NOP3
#define K8_NOP5 K8_NOP3,K8_NOP2
#define K8_NOP6 K8_NOP3,K8_NOP3
#define K8_NOP7 K8_NOP4,K8_NOP3
#define K8_NOP8 K8_NOP4,K8_NOP4
#define P6_NOP3 0x0f,0x1f,0x00
#define P6_NOP4 0x0f,0x1f,0x40,0
#define P6_NOP5 0x0f,0x1f,0x44,0x00,0
#define P6_NOP6 0x66,0x0f,0x1f,0x44,0x00,0
#define P6_NOP7 0x0f,0x1f,0x80,0,0,0,0
#define P6_NOP8 0x0f,0x1f,0x84,0x00,0,0,0,0
#define BUILD_NOP(func, nop) \
static void func(void) \
{ \
asm volatile(".rept 1000\n" \
".byte " __stringify(nop) "\n" \
".endr"); \
}
/* single-byte NOP */
BUILD_NOP(k8_nop1, K8_NOP1)
/* 2-byte NOPs */
BUILD_NOP(k8_nop2, K8_NOP2)
/* 3-byte NOPs */
BUILD_NOP(k8_nop3, K8_NOP3)
BUILD_NOP(p6_nop3, P6_NOP3)
/* 4-byte NOPs */
BUILD_NOP(k8_nop4, K8_NOP4)
BUILD_NOP(p6_nop4, P6_NOP4)
/* 5-byte NOPs */
static void p6_nop5(void)
{
asm volatile(".rept 1000\n"
".byte 0x0f,0x1f,0x44,0x00,0x00\n"
".endr");
}
static void nop_k8(void)
{
asm volatile(".rept 1000\n"
".byte 0x66,0x66,0x66,0x66,0x90\n"
".endr");
}
BUILD_NOP(k8_nop5, K8_NOP5)
static void nop_lea(void)
{
#ifdef __x86_64__
asm volatile(".rept 1000\n"
".byte 0x48,0x8d,0x74,0x26,0x00\n"
".endr");
#else
asm volatile(".rept 1000\n"
".byte 0x3e,0x8d,0x74,0x26,0x00\n"
".endr");
#endif
}
static void nop_jmp5(void)
{
asm volatile(".rept 1000\n"
".byte 0xe9,0,0,0,0\n"
".endr");
}
static void nop_jmp2(void)
{
asm volatile(".rept 1000\n"
".byte 0xeb,3,0x90,0x90,0x90\n"
".endr");
}
static void nop_xchg(void)
{
asm volatile(".rept 1000\n"
".byte 0x66,0x66,0x66,0x87,0xc0\n"
".endr");
}
static void nop_mov(void)
{
asm volatile(".rept 1000\n"
".byte 0x66,0x66,0x66,0x89,0xc0\n"
".endr");
}
static void nop_fdisi(void)
{
asm volatile(".rept 1000\n"
".byte 0x66,0x66,0x66,0xdb,0xe1\n"
".endr");
}
static void nop_feni(void)
{
asm volatile(".rept 1000\n"
".byte 0x66,0x66,0x66,0xdb,0xe0\n"
".endr");
}
/* 6-byte NOPs */
BUILD_NOP(k8_nop6, K8_NOP6)
BUILD_NOP(p6_nop6, P6_NOP6)
/* 7-byte NOPs */
BUILD_NOP(k8_nop7, K8_NOP7)
BUILD_NOP(p6_nop7, P6_NOP7)
/* 8-byte NOPs */
BUILD_NOP(k8_nop8, K8_NOP8)
BUILD_NOP(p6_nop8, P6_NOP8)
struct test_list {
const char *name;
void (*func)(void);
};
static const struct test_list tests[] = {
{ "P6 NOPs (NOPL)", p6_nop5 },
{ "K8 NOPs (66 90)", nop_k8 },
{ "LEA", nop_lea },
{ "XCHG", nop_xchg },
{ "MOV", nop_mov },
{ "FDISI", nop_fdisi },
{ "FENI", nop_feni },
{ "E9 JMP", nop_jmp5 },
{ "EB JMP", nop_jmp2 },
{ NULL, NULL }
};
#define TIMES 30
static void benchmark(const struct test_list *test, const int reps, bool warmup)
{
u64 p1, p2, r;
double min = 10000000000;
int i, j;
for (j = 0; j < TIMES; j++) {
p1 = read_tsc();
for (i = 0; i < reps; i++)
test->func();
p2 = read_tsc();
r = (p2 - p1);
if (r < min)
min = r;
}
if (!warmup)
printf("%24s%15f cycles\n", test->name, min/reps);
}
static const struct test_list k8_nops[] = {
{ NULL, NULL },
{ "90", k8_nop1 },
{ "66 90", k8_nop2 },
{ "66 66 90", k8_nop3 },
{ "66 66 66 90", k8_nop4 },
{ "66 66 90 66 90", k8_nop5 },
{ "66 66 90 66 66 90", k8_nop6 },
{ "66 66 66 90 66 66 90", k8_nop7 },
{ "66 66 66 90 66 66 66 90", k8_nop8 },
{ NULL, NULL },
};
static const struct test_list f16h_nops[] = {
{ NULL, NULL },
{ "90", k8_nop1 },
{ "66 90", k8_nop2 },
{ "0f 1f 00", p6_nop3 },
{ "0f 1f 40 00", p6_nop4 },
{ "0f 1f 44 00 00", p6_nop5 },
{ "66 0f 1f 44 00 00", p6_nop6 },
{ "0f 1f 80 00 00 00 00", p6_nop7 },
{ "0f 1f 84 00 00 00 00 00", p6_nop8 },
{ NULL, NULL },
};
int main(void)
{
const int reps = 1000000;
const struct test_list *test;
int i;
printf("Running NOP tests, 1000 NOPs x %d repetitions\n\n", reps);
#if 0
for (test = tests; test->func; test++) {
benchmark(test, reps, true);
benchmark(test, reps, false);
}
#endif
printf("K8:\n");
for (i = 1; i < 9; i++) {
benchmark(&k8_nops[i], reps, true);
benchmark(&k8_nops[i], reps, false);
}
printf("\n");
printf("P6:\n");
for (i = 1; i < 9; i++) {
benchmark(&f16h_nops[i], reps, true);
benchmark(&f16h_nops[i], reps, false);
}
printf("\n");
return 0;
}
--
Regards/Gruss,
Boris.
ECO tip #101: Trim your mails when you reply.
--
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/