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	if (count < 6)
 79		printf("\r");
 80	else
 81		printf("\n");
 82
 83	if (key->kernstack == -EEXIST && !warned) {
 84		printf("stackmap collisions seen. Consider increasing size\n");
 85		warned = true;
 86	} else if ((int)key->kernstack < 0 && (int)key->userstack < 0) {
 87		printf("err stackid %d %d\n", key->kernstack, key->userstack);
 88	}
 89}
 90
 91static void int_exit(int sig)
 92{
 93	kill(0, SIGKILL);
 94	exit(0);
 95}
 96
 97static void print_stacks(void)
 98{
 99	struct key_t key = {}, next_key;
100	__u64 value;
101	__u32 stackid = 0, next_id;
102	int fd = map_fd[0], stack_map = map_fd[1];
103
104	sys_read_seen = sys_write_seen = false;
105	while (bpf_map_get_next_key(fd, &key, &next_key) == 0) {
106		bpf_map_lookup_elem(fd, &next_key, &value);
107		print_stack(&next_key, value);
108		bpf_map_delete_elem(fd, &next_key);
109		key = next_key;
110	}
111	printf("\n");
112	if (!sys_read_seen || !sys_write_seen) {
113		printf("BUG kernel stack doesn't contain sys_read() and sys_write()\n");
114		int_exit(0);
115	}
116
117	/* clear stack map */
118	while (bpf_map_get_next_key(stack_map, &stackid, &next_id) == 0) {
119		bpf_map_delete_elem(stack_map, &next_id);
120		stackid = next_id;
121	}
122}
123
124static void test_perf_event_all_cpu(struct perf_event_attr *attr)
125{
126	int nr_cpus = sysconf(_SC_NPROCESSORS_CONF);
127	int *pmu_fd = malloc(nr_cpus * sizeof(int));
128	int i, error = 0;
129
130	/* system wide perf event, no need to inherit */
131	attr->inherit = 0;
132
133	/* open perf_event on all cpus */
134	for (i = 0; i < nr_cpus; i++) {
135		pmu_fd[i] = sys_perf_event_open(attr, -1, i, -1, 0);
136		if (pmu_fd[i] < 0) {
137			printf("sys_perf_event_open failed\n");
138			error = 1;
139			goto all_cpu_err;
140		}
141		assert(ioctl(pmu_fd[i], PERF_EVENT_IOC_SET_BPF, prog_fd[0]) == 0);
142		assert(ioctl(pmu_fd[i], PERF_EVENT_IOC_ENABLE) == 0);
143	}
144	system("dd if=/dev/zero of=/dev/null count=5000k status=none");
145	print_stacks();
146all_cpu_err:
147	for (i--; i >= 0; i--) {
148		ioctl(pmu_fd[i], PERF_EVENT_IOC_DISABLE);
149		close(pmu_fd[i]);
150	}
151	free(pmu_fd);
152	if (error)
153		int_exit(0);
154}
155
156static void test_perf_event_task(struct perf_event_attr *attr)
157{
158	int pmu_fd;
159
160	/* per task perf event, enable inherit so the "dd ..." command can be traced properly.
161	 * Enabling inherit will cause bpf_perf_prog_read_time helper failure.
162	 */
163	attr->inherit = 1;
164
165	/* open task bound event */
166	pmu_fd = sys_perf_event_open(attr, 0, -1, -1, 0);
167	if (pmu_fd < 0) {
168		printf("sys_perf_event_open failed\n");
169		int_exit(0);
170	}
171	assert(ioctl(pmu_fd, PERF_EVENT_IOC_SET_BPF, prog_fd[0]) == 0);
172	assert(ioctl(pmu_fd, PERF_EVENT_IOC_ENABLE) == 0);
173	system("dd if=/dev/zero of=/dev/null count=5000k status=none");
174	print_stacks();
175	ioctl(pmu_fd, PERF_EVENT_IOC_DISABLE);
176	close(pmu_fd);
177}
178
179static void test_bpf_perf_event(void)
180{
181	struct perf_event_attr attr_type_hw = {
182		.sample_freq = SAMPLE_FREQ,
183		.freq = 1,
184		.type = PERF_TYPE_HARDWARE,
185		.config = PERF_COUNT_HW_CPU_CYCLES,
186	};
187	struct perf_event_attr attr_type_sw = {
188		.sample_freq = SAMPLE_FREQ,
189		.freq = 1,
190		.type = PERF_TYPE_SOFTWARE,
191		.config = PERF_COUNT_SW_CPU_CLOCK,
192	};
193	struct perf_event_attr attr_hw_cache_l1d = {
194		.sample_freq = SAMPLE_FREQ,
195		.freq = 1,
196		.type = PERF_TYPE_HW_CACHE,
197		.config =
198			PERF_COUNT_HW_CACHE_L1D |
199			(PERF_COUNT_HW_CACHE_OP_READ << 8) |
200			(PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16),
201	};
202	struct perf_event_attr attr_hw_cache_branch_miss = {
203		.sample_freq = SAMPLE_FREQ,
204		.freq = 1,
205		.type = PERF_TYPE_HW_CACHE,
206		.config =
207			PERF_COUNT_HW_CACHE_BPU |
208			(PERF_COUNT_HW_CACHE_OP_READ << 8) |
209			(PERF_COUNT_HW_CACHE_RESULT_MISS << 16),
210	};
211	struct perf_event_attr attr_type_raw = {
212		.sample_freq = SAMPLE_FREQ,
213		.freq = 1,
214		.type = PERF_TYPE_RAW,
215		/* Intel Instruction Retired */
216		.config = 0xc0,
217	};
218	struct perf_event_attr attr_type_raw_lock_load = {
219		.sample_freq = SAMPLE_FREQ,
220		.freq = 1,
221		.type = PERF_TYPE_RAW,
222		/* Intel MEM_UOPS_RETIRED.LOCK_LOADS */
223		.config = 0x21d0,
224		/* Request to record lock address from PEBS */
225		.sample_type = PERF_SAMPLE_ADDR,
226		/* Record address value requires precise event */
227		.precise_ip = 2,
228	};
229
230	printf("Test HW_CPU_CYCLES\n");
231	test_perf_event_all_cpu(&attr_type_hw);
232	test_perf_event_task(&attr_type_hw);
233
234	printf("Test SW_CPU_CLOCK\n");
235	test_perf_event_all_cpu(&attr_type_sw);
236	test_perf_event_task(&attr_type_sw);
237
238	printf("Test HW_CACHE_L1D\n");
239	test_perf_event_all_cpu(&attr_hw_cache_l1d);
240	test_perf_event_task(&attr_hw_cache_l1d);
241
242	printf("Test HW_CACHE_BPU\n");
243	test_perf_event_all_cpu(&attr_hw_cache_branch_miss);
244	test_perf_event_task(&attr_hw_cache_branch_miss);
245
246	printf("Test Instruction Retired\n");
247	test_perf_event_all_cpu(&attr_type_raw);
248	test_perf_event_task(&attr_type_raw);
249
250	printf("Test Lock Load\n");
251	test_perf_event_all_cpu(&attr_type_raw_lock_load);
252	test_perf_event_task(&attr_type_raw_lock_load);
253
254	printf("*** PASS ***\n");
255}
256
257
258int main(int argc, char **argv)
259{
260	struct rlimit r = {RLIM_INFINITY, RLIM_INFINITY};
261	char filename[256];
262
263	snprintf(filename, sizeof(filename), "%s_kern.o", argv[0]);
264	setrlimit(RLIMIT_MEMLOCK, &r);
265
266	signal(SIGINT, int_exit);
267	signal(SIGTERM, int_exit);
268
269	if (load_kallsyms()) {
270		printf("failed to process /proc/kallsyms\n");
271		return 1;
272	}
273
274	if (load_bpf_file(filename)) {
275		printf("%s", bpf_log_buf);
276		return 2;
277	}
278
279	if (fork() == 0) {
280		read_trace_pipe();
281		return 0;
282	}
283	test_bpf_perf_event();
284	int_exit(0);
285	return 0;
286}