This is the patch I'm using to evaluate the impact syscall user dispatch has on native syscall (syscalls not redirected to userspace) when enabled for the process and submiting syscalls though the unblocked dispatch selector. It works by running a step to define a baseline of the cost of executing sysinfo, then enabling SUD, and rerunning that step.
On my test machine, an AMD Ryzen 5 1500X, I have the following results with the latest version of syscall user dispatch patches.
root@olga:~# syscall_user_dispatch/sud_benchmark Calibrating test set to last ~5 seconds... test iterations = 37500000 Avg syscall time 134ns. Caught sys_ff00 trapped_call_count 1, native_call_count 0. Avg syscall time 147ns. Interception overhead: 9.7% (+13ns).
Signed-off-by: Gabriel Krisman Bertazi krisman@collabora.com Acked-by: Peter Zijlstra (Intel) peterz@infradead.org --- .../selftests/syscall_user_dispatch/Makefile | 2 +- .../syscall_user_dispatch/sud_benchmark.c | 200 ++++++++++++++++++ 2 files changed, 201 insertions(+), 1 deletion(-) create mode 100644 tools/testing/selftests/syscall_user_dispatch/sud_benchmark.c
diff --git a/tools/testing/selftests/syscall_user_dispatch/Makefile b/tools/testing/selftests/syscall_user_dispatch/Makefile index 8e15fa42bcda..03c120270953 100644 --- a/tools/testing/selftests/syscall_user_dispatch/Makefile +++ b/tools/testing/selftests/syscall_user_dispatch/Makefile @@ -5,5 +5,5 @@ LINUX_HDR_PATH = $(INSTALL_HDR_PATH)/include/
CFLAGS += -Wall -I$(LINUX_HDR_PATH)
-TEST_GEN_PROGS := sud_test +TEST_GEN_PROGS := sud_test sud_benchmark include ../lib.mk diff --git a/tools/testing/selftests/syscall_user_dispatch/sud_benchmark.c b/tools/testing/selftests/syscall_user_dispatch/sud_benchmark.c new file mode 100644 index 000000000000..6689f1183dbf --- /dev/null +++ b/tools/testing/selftests/syscall_user_dispatch/sud_benchmark.c @@ -0,0 +1,200 @@ +// SPDX-License-Identifier: GPL-2.0-only +/* + * Copyright (c) 2020 Collabora Ltd. + * + * Benchmark and test syscall user dispatch + */ + +#define _GNU_SOURCE +#include <stdio.h> +#include <string.h> +#include <stdlib.h> +#include <signal.h> +#include <errno.h> +#include <time.h> +#include <sys/time.h> +#include <unistd.h> +#include <sys/sysinfo.h> +#include <sys/prctl.h> +#include <sys/syscall.h> + +#ifndef PR_SET_SYSCALL_USER_DISPATCH +# define PR_SET_SYSCALL_USER_DISPATCH 59 +# define PR_SYS_DISPATCH_OFF 0 +# define PR_SYS_DISPATCH_ON 1 +#endif + +#ifdef __NR_syscalls +# define MAGIC_SYSCALL_1 (__NR_syscalls + 1) /* Bad Linux syscall number */ +#else +# define MAGIC_SYSCALL_1 (0xff00) /* Bad Linux syscall number */ +#endif + +/* + * To test returning from a sigsys with selector blocked, the test + * requires some per-architecture support (i.e. knowledge about the + * signal trampoline address). On i386, we know it is on the vdso, and + * a small trampoline is open-coded for x86_64. Other architectures + * that have a trampoline in the vdso will support TEST_BLOCKED_RETURN + * out of the box, but don't enable them until they support syscall user + * dispatch. + */ +#if defined(__x86_64__) || defined(__i386__) +#define TEST_BLOCKED_RETURN +#endif + +#ifdef __x86_64__ +void* (syscall_dispatcher_start)(void); +void* (syscall_dispatcher_end)(void); +#else +unsigned long syscall_dispatcher_start = 0; +unsigned long syscall_dispatcher_end = 0; +#endif + +unsigned long trapped_call_count = 0; +unsigned long native_call_count = 0; + +char selector; +#define SYSCALL_BLOCK (selector = PR_SYS_DISPATCH_ON) +#define SYSCALL_UNBLOCK (selector = PR_SYS_DISPATCH_OFF) + +#define CALIBRATION_STEP 100000 +#define CALIBRATE_TO_SECS 5 +int factor; + +static double one_sysinfo_step(void) +{ + struct timespec t1, t2; + int i; + struct sysinfo info; + + clock_gettime(CLOCK_MONOTONIC, &t1); + for (i = 0; i < CALIBRATION_STEP; i++) + sysinfo(&info); + clock_gettime(CLOCK_MONOTONIC, &t2); + return (t2.tv_sec - t1.tv_sec) + 1.0e-9 * (t2.tv_nsec - t1.tv_nsec); +} + +static void calibrate_set(void) +{ + double elapsed = 0; + + printf("Calibrating test set to last ~%d seconds...\n", CALIBRATE_TO_SECS); + + while (elapsed < 1) { + elapsed += one_sysinfo_step(); + factor += CALIBRATE_TO_SECS; + } + + printf("test iterations = %d\n", CALIBRATION_STEP * factor); +} + +static double perf_syscall(void) +{ + unsigned int i; + double partial = 0; + + for (i = 0; i < factor; ++i) + partial += one_sysinfo_step()/(CALIBRATION_STEP*factor); + return partial; +} + +static void handle_sigsys(int sig, siginfo_t *info, void *ucontext) +{ + char buf[1024]; + int len; + + SYSCALL_UNBLOCK; + + /* printf and friends are not signal-safe. */ + len = snprintf(buf, 1024, "Caught sys_%x\n", info->si_syscall); + write(1, buf, len); + + if (info->si_syscall == MAGIC_SYSCALL_1) + trapped_call_count++; + else + native_call_count++; + +#ifdef TEST_BLOCKED_RETURN + SYSCALL_BLOCK; +#endif + +#ifdef __x86_64__ + __asm__ volatile("movq $0xf, %rax"); + __asm__ volatile("leaveq"); + __asm__ volatile("add $0x8, %rsp"); + __asm__ volatile("syscall_dispatcher_start:"); + __asm__ volatile("syscall"); + __asm__ volatile("nop"); /* Landing pad within dispatcher area */ + __asm__ volatile("syscall_dispatcher_end:"); +#endif + +} + +int main(void) +{ + struct sigaction act; + double time1, time2; + int ret; + sigset_t mask; + + memset(&act, 0, sizeof(act)); + sigemptyset(&mask); + + act.sa_sigaction = handle_sigsys; + act.sa_flags = SA_SIGINFO; + act.sa_mask = mask; + + calibrate_set(); + + time1 = perf_syscall(); + printf("Avg syscall time %.0lfns.\n", time1 * 1.0e9); + + ret = sigaction(SIGSYS, &act, NULL); + if (ret) { + perror("Error sigaction:"); + exit(-1); + } + + fprintf(stderr, "Enabling syscall trapping.\n"); + + if (prctl(PR_SET_SYSCALL_USER_DISPATCH, PR_SYS_DISPATCH_ON, + syscall_dispatcher_start, + (syscall_dispatcher_end - syscall_dispatcher_start + 1), + &selector)) { + perror("prctl failed\n"); + exit(-1); + } + + SYSCALL_BLOCK; + syscall(MAGIC_SYSCALL_1); + +#ifdef TEST_BLOCKED_RETURN + if (selector == PR_SYS_DISPATCH_OFF) { + fprintf(stderr, "Failed to return with selector blocked.\n"); + exit(-1); + } +#endif + + SYSCALL_UNBLOCK; + + if (!trapped_call_count) { + fprintf(stderr, "syscall trapping does not work.\n"); + exit(-1); + } + + time2 = perf_syscall(); + + if (native_call_count) { + perror("syscall trapping intercepted more syscalls than expected\n"); + exit(-1); + } + + printf("trapped_call_count %lu, native_call_count %lu.\n", + trapped_call_count, native_call_count); + printf("Avg syscall time %.0lfns.\n", time2 * 1.0e9); + printf("Interception overhead: %.1lf%% (+%.0lfns).\n", + 100.0 * (time2 / time1 - 1.0), 1.0e9 * (time2 - time1)); + return 0; + +}