1/* Copyright (c) 2016 Facebook
  2 *
  3 * This program is free software; you can redistribute it and/or
  4 * modify it under the terms of version 2 of the GNU General Public
  5 * License as published by the Free Software Foundation.
  6 */
  7#include <stdio.h>
  8#include <unistd.h>
  9#include <stdlib.h>
 10#include <stdbool.h>
 11#include <string.h>
 12#include <fcntl.h>
 13#include <poll.h>
 14#include <sys/ioctl.h>
 15#include <linux/perf_event.h>
 16#include <linux/bpf.h>
 17#include <signal.h>
 18#include <assert.h>
 19#include <errno.h>
 20#include <sys/resource.h>
 21#include "libbpf.h"
 22#include "bpf_load.h"
 23#include "perf-sys.h"
 24
 25#define SAMPLE_FREQ 50
 26
 27static bool sys_read_seen, sys_write_seen;
 28
 29static void print_ksym(__u64 addr)
 30{
 31	struct ksym *sym;
 32
 33	if (!addr)
 34		return;
 35	sym = ksym_search(addr);
 36	printf("%s;", sym->name);
 37	if (!strcmp(sym->name, "sys_read"))
 38		sys_read_seen = true;
 39	else if (!strcmp(sym->name, "sys_write"))
 40		sys_write_seen = true;
 41}
 42
 43static void print_addr(__u64 addr)
 44{
 45	if (!addr)
 46		return;
 47	printf("%llx;", addr);
 48}
 49
 50#define TASK_COMM_LEN 16
 51
 52struct key_t {
 53	char comm[TASK_COMM_LEN];
 54	__u32 kernstack;
 55	__u32 userstack;
 56};
 57
 58static void print_stack(struct key_t *key, __u64 count)
 59{
 60	__u64 ip[PERF_MAX_STACK_DEPTH] = {};
 61	static bool warned;
 62	int i;
 63
 64	printf("%3lld %s;", count, key->comm);
 65	if (bpf_map_lookup_elem(map_fd[1], &key->kernstack, ip) != 0) {
 66		printf("---;");
 67	} else {
 68		for (i = PERF_MAX_STACK_DEPTH - 1; i >= 0; i--)
 69			print_ksym(ip[i]);
 70	}
 71	printf("-;");
 72	if (bpf_map_lookup_elem(map_fd[1], &key->userstack, ip) != 0) {
 73		printf("---;");
 74	} else {
 75		for (i = PERF_MAX_STACK_DEPTH - 1; i >= 0; i--)
 76			print_addr(ip[i]);
 77	}
 78	printf("\n");
 79
 80	if (key->kernstack == -EEXIST && !warned) {
 81		printf("stackmap collisions seen. Consider increasing size\n");
 82		warned = true;
 83	} else if ((int)key->kernstack < 0 && (int)key->userstack < 0) {
 84		printf("err stackid %d %d\n", key->kernstack, key->userstack);
 85	}
 86}
 87
 88static void int_exit(int sig)
 89{
 90	kill(0, SIGKILL);
 91	exit(0);
 92}
 93
 94static void print_stacks(void)
 95{
 96	struct key_t key = {}, next_key;
 97	__u64 value;
 98	__u32 stackid = 0, next_id;
 99	int fd = map_fd[0], stack_map = map_fd[1];
100
101	sys_read_seen = sys_write_seen = false;
102	while (bpf_map_get_next_key(fd, &key, &next_key) == 0) {
103		bpf_map_lookup_elem(fd, &next_key, &value);
104		print_stack(&next_key, value);
105		bpf_map_delete_elem(fd, &next_key);
106		key = next_key;
107	}
108
109	if (!sys_read_seen || !sys_write_seen) {
110		printf("BUG kernel stack doesn't contain sys_read() and sys_write()\n");
111		int_exit(0);
112	}
113
114	/* clear stack map */
115	while (bpf_map_get_next_key(stack_map, &stackid, &next_id) == 0) {
116		bpf_map_delete_elem(stack_map, &next_id);
117		stackid = next_id;
118	}
119}
120
121static void test_perf_event_all_cpu(struct perf_event_attr *attr)
122{
123	int nr_cpus = sysconf(_SC_NPROCESSORS_CONF);
124	int *pmu_fd = malloc(nr_cpus * sizeof(int));
125	int i;
126
127	/* open perf_event on all cpus */
128	for (i = 0; i < nr_cpus; i++) {
129		pmu_fd[i] = sys_perf_event_open(attr, -1, i, -1, 0);
130		if (pmu_fd[i] < 0) {
131			printf("sys_perf_event_open failed\n");
132			goto all_cpu_err;
133		}
134		assert(ioctl(pmu_fd[i], PERF_EVENT_IOC_SET_BPF, prog_fd[0]) == 0);
135		assert(ioctl(pmu_fd[i], PERF_EVENT_IOC_ENABLE, 0) == 0);
136	}
137	system("dd if=/dev/zero of=/dev/null count=5000k");
138	print_stacks();
139all_cpu_err:
140	for (i--; i >= 0; i--)
141		close(pmu_fd[i]);
142	free(pmu_fd);
143}
144
145static void test_perf_event_task(struct perf_event_attr *attr)
146{
147	int pmu_fd;
148
149	/* open task bound event */
150	pmu_fd = sys_perf_event_open(attr, 0, -1, -1, 0);
151	if (pmu_fd < 0) {
152		printf("sys_perf_event_open failed\n");
153		return;
154	}
155	assert(ioctl(pmu_fd, PERF_EVENT_IOC_SET_BPF, prog_fd[0]) == 0);
156	assert(ioctl(pmu_fd, PERF_EVENT_IOC_ENABLE, 0) == 0);
157	system("dd if=/dev/zero of=/dev/null count=5000k");
158	print_stacks();
159	close(pmu_fd);
160}
161
162static void test_bpf_perf_event(void)
163{
164	struct perf_event_attr attr_type_hw = {
165		.sample_freq = SAMPLE_FREQ,
166		.freq = 1,
167		.type = PERF_TYPE_HARDWARE,
168		.config = PERF_COUNT_HW_CPU_CYCLES,
169		.inherit = 1,
170	};
171	struct perf_event_attr attr_type_sw = {
172		.sample_freq = SAMPLE_FREQ,
173		.freq = 1,
174		.type = PERF_TYPE_SOFTWARE,
175		.config = PERF_COUNT_SW_CPU_CLOCK,
176		.inherit = 1,
177	};
178
179	test_perf_event_all_cpu(&attr_type_hw);
180	test_perf_event_task(&attr_type_hw);
181	test_perf_event_all_cpu(&attr_type_sw);
182	test_perf_event_task(&attr_type_sw);
183}
184
185
186int main(int argc, char **argv)
187{
188	struct rlimit r = {RLIM_INFINITY, RLIM_INFINITY};
189	char filename[256];
190
191	snprintf(filename, sizeof(filename), "%s_kern.o", argv[0]);
192	setrlimit(RLIMIT_MEMLOCK, &r);
193
194	signal(SIGINT, int_exit);
195
196	if (load_kallsyms()) {
197		printf("failed to process /proc/kallsyms\n");
198		return 1;
199	}
200
201	if (load_bpf_file(filename)) {
202		printf("%s", bpf_log_buf);
203		return 2;
204	}
205
206	if (fork() == 0) {
207		read_trace_pipe();
208		return 0;
209	}
210	test_bpf_perf_event();
211
212	int_exit(0);
213	return 0;
214}