selftests/bpf: Add BPF trampoline performance test
authorAlexei Starovoitov <ast@kernel.org>
Fri, 22 Nov 2019 01:15:15 +0000 (17:15 -0800)
committerAlexei Starovoitov <ast@kernel.org>
Mon, 25 Nov 2019 00:58:46 +0000 (16:58 -0800)
Add a test that benchmarks different ways of attaching BPF program to a kernel function.
Here are the results for 2.4Ghz x86 cpu on a kernel without mitigations:
$ ./test_progs -n 49 -v|grep events
task_rename base 2743K events per sec
task_rename kprobe 2419K events per sec
task_rename kretprobe 1876K events per sec
task_rename raw_tp 2578K events per sec
task_rename fentry 2710K events per sec
task_rename fexit 2685K events per sec

On a kernel with retpoline:
$ ./test_progs -n 49 -v|grep events
task_rename base 2401K events per sec
task_rename kprobe 1930K events per sec
task_rename kretprobe 1485K events per sec
task_rename raw_tp 2053K events per sec
task_rename fentry 2351K events per sec
task_rename fexit 2185K events per sec

All 5 approaches:
- kprobe/kretprobe in __set_task_comm()
- raw tracepoint in trace_task_rename()
- fentry/fexit in __set_task_comm()
are roughly equivalent.

__set_task_comm() by itself is quite fast, so any extra instructions add up.
Until BPF trampoline was introduced the fastest mechanism was raw tracepoint.
kprobe via ftrace was second best. kretprobe is slow due to trap. New
fentry/fexit methods via BPF trampoline are clearly the fastest and the
difference is more pronounced with retpoline on, since BPF trampoline doesn't
use indirect jumps.

Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: John Fastabend <john.fastabend@gmail.com>
Link: https://lore.kernel.org/bpf/20191122011515.255371-1-ast@kernel.org
tools/testing/selftests/bpf/prog_tests/test_overhead.c [new file with mode: 0644]
tools/testing/selftests/bpf/progs/test_overhead.c [new file with mode: 0644]

diff --git a/tools/testing/selftests/bpf/prog_tests/test_overhead.c b/tools/testing/selftests/bpf/prog_tests/test_overhead.c
new file mode 100644 (file)
index 0000000..c32aa28
--- /dev/null
@@ -0,0 +1,142 @@
+// SPDX-License-Identifier: GPL-2.0-only
+/* Copyright (c) 2019 Facebook */
+#define _GNU_SOURCE
+#include <sched.h>
+#include <test_progs.h>
+
+#define MAX_CNT 100000
+
+static __u64 time_get_ns(void)
+{
+       struct timespec ts;
+
+       clock_gettime(CLOCK_MONOTONIC, &ts);
+       return ts.tv_sec * 1000000000ull + ts.tv_nsec;
+}
+
+static int test_task_rename(const char *prog)
+{
+       int i, fd, duration = 0, err;
+       char buf[] = "test\n";
+       __u64 start_time;
+
+       fd = open("/proc/self/comm", O_WRONLY|O_TRUNC);
+       if (CHECK(fd < 0, "open /proc", "err %d", errno))
+               return -1;
+       start_time = time_get_ns();
+       for (i = 0; i < MAX_CNT; i++) {
+               err = write(fd, buf, sizeof(buf));
+               if (err < 0) {
+                       CHECK(err < 0, "task rename", "err %d", errno);
+                       close(fd);
+                       return -1;
+               }
+       }
+       printf("task_rename %s\t%lluK events per sec\n", prog,
+              MAX_CNT * 1000000ll / (time_get_ns() - start_time));
+       close(fd);
+       return 0;
+}
+
+static void test_run(const char *prog)
+{
+       test_task_rename(prog);
+}
+
+static void setaffinity(void)
+{
+       cpu_set_t cpuset;
+       int cpu = 0;
+
+       CPU_ZERO(&cpuset);
+       CPU_SET(cpu, &cpuset);
+       sched_setaffinity(0, sizeof(cpuset), &cpuset);
+}
+
+void test_test_overhead(void)
+{
+       const char *kprobe_name = "kprobe/__set_task_comm";
+       const char *kretprobe_name = "kretprobe/__set_task_comm";
+       const char *raw_tp_name = "raw_tp/task_rename";
+       const char *fentry_name = "fentry/__set_task_comm";
+       const char *fexit_name = "fexit/__set_task_comm";
+       const char *kprobe_func = "__set_task_comm";
+       struct bpf_program *kprobe_prog, *kretprobe_prog, *raw_tp_prog;
+       struct bpf_program *fentry_prog, *fexit_prog;
+       struct bpf_object *obj;
+       struct bpf_link *link;
+       int err, duration = 0;
+
+       obj = bpf_object__open_file("./test_overhead.o", NULL);
+       if (CHECK(IS_ERR(obj), "obj_open_file", "err %ld\n", PTR_ERR(obj)))
+               return;
+
+       kprobe_prog = bpf_object__find_program_by_title(obj, kprobe_name);
+       if (CHECK(!kprobe_prog, "find_probe",
+                 "prog '%s' not found\n", kprobe_name))
+               goto cleanup;
+       kretprobe_prog = bpf_object__find_program_by_title(obj, kretprobe_name);
+       if (CHECK(!kretprobe_prog, "find_probe",
+                 "prog '%s' not found\n", kretprobe_name))
+               goto cleanup;
+       raw_tp_prog = bpf_object__find_program_by_title(obj, raw_tp_name);
+       if (CHECK(!raw_tp_prog, "find_probe",
+                 "prog '%s' not found\n", raw_tp_name))
+               goto cleanup;
+       fentry_prog = bpf_object__find_program_by_title(obj, fentry_name);
+       if (CHECK(!fentry_prog, "find_probe",
+                 "prog '%s' not found\n", fentry_name))
+               goto cleanup;
+       fexit_prog = bpf_object__find_program_by_title(obj, fexit_name);
+       if (CHECK(!fexit_prog, "find_probe",
+                 "prog '%s' not found\n", fexit_name))
+               goto cleanup;
+
+       err = bpf_object__load(obj);
+       if (CHECK(err, "obj_load", "err %d\n", err))
+               goto cleanup;
+
+       setaffinity();
+
+       /* base line run */
+       test_run("base");
+
+       /* attach kprobe */
+       link = bpf_program__attach_kprobe(kprobe_prog, false /* retprobe */,
+                                         kprobe_func);
+       if (CHECK(IS_ERR(link), "attach_kprobe", "err %ld\n", PTR_ERR(link)))
+               goto cleanup;
+       test_run("kprobe");
+       bpf_link__destroy(link);
+
+       /* attach kretprobe */
+       link = bpf_program__attach_kprobe(kretprobe_prog, true /* retprobe */,
+                                         kprobe_func);
+       if (CHECK(IS_ERR(link), "attach kretprobe", "err %ld\n", PTR_ERR(link)))
+               goto cleanup;
+       test_run("kretprobe");
+       bpf_link__destroy(link);
+
+       /* attach raw_tp */
+       link = bpf_program__attach_raw_tracepoint(raw_tp_prog, "task_rename");
+       if (CHECK(IS_ERR(link), "attach fentry", "err %ld\n", PTR_ERR(link)))
+               goto cleanup;
+       test_run("raw_tp");
+       bpf_link__destroy(link);
+
+       /* attach fentry */
+       link = bpf_program__attach_trace(fentry_prog);
+       if (CHECK(IS_ERR(link), "attach fentry", "err %ld\n", PTR_ERR(link)))
+               goto cleanup;
+       test_run("fentry");
+       bpf_link__destroy(link);
+
+       /* attach fexit */
+       link = bpf_program__attach_trace(fexit_prog);
+       if (CHECK(IS_ERR(link), "attach fexit", "err %ld\n", PTR_ERR(link)))
+               goto cleanup;
+       test_run("fexit");
+       bpf_link__destroy(link);
+cleanup:
+       bpf_object__close(obj);
+}
diff --git a/tools/testing/selftests/bpf/progs/test_overhead.c b/tools/testing/selftests/bpf/progs/test_overhead.c
new file mode 100644 (file)
index 0000000..ef06b26
--- /dev/null
@@ -0,0 +1,43 @@
+// SPDX-License-Identifier: GPL-2.0
+/* Copyright (c) 2019 Facebook */
+#include <linux/bpf.h>
+#include "bpf_helpers.h"
+#include "bpf_tracing.h"
+
+SEC("kprobe/__set_task_comm")
+int prog1(struct pt_regs *ctx)
+{
+       return 0;
+}
+
+SEC("kretprobe/__set_task_comm")
+int prog2(struct pt_regs *ctx)
+{
+       return 0;
+}
+
+SEC("raw_tp/task_rename")
+int prog3(struct bpf_raw_tracepoint_args *ctx)
+{
+       return 0;
+}
+
+struct __set_task_comm_args {
+       struct task_struct *tsk;
+       const char *buf;
+       ku8 exec;
+};
+
+SEC("fentry/__set_task_comm")
+int prog4(struct __set_task_comm_args *ctx)
+{
+       return 0;
+}
+
+SEC("fexit/__set_task_comm")
+int prog5(struct __set_task_comm_args *ctx)
+{
+       return 0;
+}
+
+char _license[] SEC("license") = "GPL";