1 // SPDX-License-Identifier: GPL-2.0-only
2 /* Copyright (c) 2016 Facebook
3  */
4 #include <stdio.h>
5 #include <unistd.h>
6 #include <stdlib.h>
7 #include <stdbool.h>
8 #include <string.h>
9 #include <fcntl.h>
10 #include <poll.h>
11 #include <sys/ioctl.h>
12 #include <linux/perf_event.h>
13 #include <linux/bpf.h>
14 #include <signal.h>
15 #include <assert.h>
16 #include <errno.h>
17 #include <sys/resource.h>
18 #include "libbpf.h"
19 #include "bpf_load.h"
20 #include "perf-sys.h"
21 #include "trace_helpers.h"
22 
23 #define SAMPLE_FREQ 50
24 
25 static bool sys_read_seen, sys_write_seen;
26 
27 static void print_ksym(__u64 addr)
28 {
29 	struct ksym *sym;
30 
31 	if (!addr)
32 		return;
33 	sym = ksym_search(addr);
34 	if (!sym) {
35 		printf("ksym not found. Is kallsyms loaded?\n");
36 		return;
37 	}
38 
39 	printf("%s;", sym->name);
40 	if (!strstr(sym->name, "sys_read"))
41 		sys_read_seen = true;
42 	else if (!strstr(sym->name, "sys_write"))
43 		sys_write_seen = true;
44 }
45 
46 static void print_addr(__u64 addr)
47 {
48 	if (!addr)
49 		return;
50 	printf("%llx;", addr);
51 }
52 
53 #define TASK_COMM_LEN 16
54 
55 struct key_t {
56 	char comm[TASK_COMM_LEN];
57 	__u32 kernstack;
58 	__u32 userstack;
59 };
60 
61 static void print_stack(struct key_t *key, __u64 count)
62 {
63 	__u64 ip[PERF_MAX_STACK_DEPTH] = {};
64 	static bool warned;
65 	int i;
66 
67 	printf("%3lld %s;", count, key->comm);
68 	if (bpf_map_lookup_elem(map_fd[1], &key->kernstack, ip) != 0) {
69 		printf("---;");
70 	} else {
71 		for (i = PERF_MAX_STACK_DEPTH - 1; i >= 0; i--)
72 			print_ksym(ip[i]);
73 	}
74 	printf("-;");
75 	if (bpf_map_lookup_elem(map_fd[1], &key->userstack, ip) != 0) {
76 		printf("---;");
77 	} else {
78 		for (i = PERF_MAX_STACK_DEPTH - 1; i >= 0; i--)
79 			print_addr(ip[i]);
80 	}
81 	if (count < 6)
82 		printf("\r");
83 	else
84 		printf("\n");
85 
86 	if (key->kernstack == -EEXIST && !warned) {
87 		printf("stackmap collisions seen. Consider increasing size\n");
88 		warned = true;
89 	} else if ((int)key->kernstack < 0 && (int)key->userstack < 0) {
90 		printf("err stackid %d %d\n", key->kernstack, key->userstack);
91 	}
92 }
93 
94 static void int_exit(int sig)
95 {
96 	kill(0, SIGKILL);
97 	exit(0);
98 }
99 
100 static void print_stacks(void)
101 {
102 	struct key_t key = {}, next_key;
103 	__u64 value;
104 	__u32 stackid = 0, next_id;
105 	int fd = map_fd[0], stack_map = map_fd[1];
106 
107 	sys_read_seen = sys_write_seen = false;
108 	while (bpf_map_get_next_key(fd, &key, &next_key) == 0) {
109 		bpf_map_lookup_elem(fd, &next_key, &value);
110 		print_stack(&next_key, value);
111 		bpf_map_delete_elem(fd, &next_key);
112 		key = next_key;
113 	}
114 	printf("\n");
115 	if (!sys_read_seen || !sys_write_seen) {
116 		printf("BUG kernel stack doesn't contain sys_read() and sys_write()\n");
117 		int_exit(0);
118 	}
119 
120 	/* clear stack map */
121 	while (bpf_map_get_next_key(stack_map, &stackid, &next_id) == 0) {
122 		bpf_map_delete_elem(stack_map, &next_id);
123 		stackid = next_id;
124 	}
125 }
126 
127 static inline int generate_load(void)
128 {
129 	if (system("dd if=/dev/zero of=/dev/null count=5000k status=none") < 0) {
130 		printf("failed to generate some load with dd: %s\n", strerror(errno));
131 		return -1;
132 	}
133 
134 	return 0;
135 }
136 
137 static void test_perf_event_all_cpu(struct perf_event_attr *attr)
138 {
139 	int nr_cpus = sysconf(_SC_NPROCESSORS_CONF);
140 	int *pmu_fd = malloc(nr_cpus * sizeof(int));
141 	int i, error = 0;
142 
143 	/* system wide perf event, no need to inherit */
144 	attr->inherit = 0;
145 
146 	/* open perf_event on all cpus */
147 	for (i = 0; i < nr_cpus; i++) {
148 		pmu_fd[i] = sys_perf_event_open(attr, -1, i, -1, 0);
149 		if (pmu_fd[i] < 0) {
150 			printf("sys_perf_event_open failed\n");
151 			error = 1;
152 			goto all_cpu_err;
153 		}
154 		assert(ioctl(pmu_fd[i], PERF_EVENT_IOC_SET_BPF, prog_fd[0]) == 0);
155 		assert(ioctl(pmu_fd[i], PERF_EVENT_IOC_ENABLE) == 0);
156 	}
157 
158 	if (generate_load() < 0) {
159 		error = 1;
160 		goto all_cpu_err;
161 	}
162 	print_stacks();
163 all_cpu_err:
164 	for (i--; i >= 0; i--) {
165 		ioctl(pmu_fd[i], PERF_EVENT_IOC_DISABLE);
166 		close(pmu_fd[i]);
167 	}
168 	free(pmu_fd);
169 	if (error)
170 		int_exit(0);
171 }
172 
173 static void test_perf_event_task(struct perf_event_attr *attr)
174 {
175 	int pmu_fd, error = 0;
176 
177 	/* per task perf event, enable inherit so the "dd ..." command can be traced properly.
178 	 * Enabling inherit will cause bpf_perf_prog_read_time helper failure.
179 	 */
180 	attr->inherit = 1;
181 
182 	/* open task bound event */
183 	pmu_fd = sys_perf_event_open(attr, 0, -1, -1, 0);
184 	if (pmu_fd < 0) {
185 		printf("sys_perf_event_open failed\n");
186 		int_exit(0);
187 	}
188 	assert(ioctl(pmu_fd, PERF_EVENT_IOC_SET_BPF, prog_fd[0]) == 0);
189 	assert(ioctl(pmu_fd, PERF_EVENT_IOC_ENABLE) == 0);
190 
191 	if (generate_load() < 0) {
192 		error = 1;
193 		goto err;
194 	}
195 	print_stacks();
196 err:
197 	ioctl(pmu_fd, PERF_EVENT_IOC_DISABLE);
198 	close(pmu_fd);
199 	if (error)
200 		int_exit(0);
201 }
202 
203 static void test_bpf_perf_event(void)
204 {
205 	struct perf_event_attr attr_type_hw = {
206 		.sample_freq = SAMPLE_FREQ,
207 		.freq = 1,
208 		.type = PERF_TYPE_HARDWARE,
209 		.config = PERF_COUNT_HW_CPU_CYCLES,
210 	};
211 	struct perf_event_attr attr_type_sw = {
212 		.sample_freq = SAMPLE_FREQ,
213 		.freq = 1,
214 		.type = PERF_TYPE_SOFTWARE,
215 		.config = PERF_COUNT_SW_CPU_CLOCK,
216 	};
217 	struct perf_event_attr attr_hw_cache_l1d = {
218 		.sample_freq = SAMPLE_FREQ,
219 		.freq = 1,
220 		.type = PERF_TYPE_HW_CACHE,
221 		.config =
222 			PERF_COUNT_HW_CACHE_L1D |
223 			(PERF_COUNT_HW_CACHE_OP_READ << 8) |
224 			(PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16),
225 	};
226 	struct perf_event_attr attr_hw_cache_branch_miss = {
227 		.sample_freq = SAMPLE_FREQ,
228 		.freq = 1,
229 		.type = PERF_TYPE_HW_CACHE,
230 		.config =
231 			PERF_COUNT_HW_CACHE_BPU |
232 			(PERF_COUNT_HW_CACHE_OP_READ << 8) |
233 			(PERF_COUNT_HW_CACHE_RESULT_MISS << 16),
234 	};
235 	struct perf_event_attr attr_type_raw = {
236 		.sample_freq = SAMPLE_FREQ,
237 		.freq = 1,
238 		.type = PERF_TYPE_RAW,
239 		/* Intel Instruction Retired */
240 		.config = 0xc0,
241 	};
242 	struct perf_event_attr attr_type_raw_lock_load = {
243 		.sample_freq = SAMPLE_FREQ,
244 		.freq = 1,
245 		.type = PERF_TYPE_RAW,
246 		/* Intel MEM_UOPS_RETIRED.LOCK_LOADS */
247 		.config = 0x21d0,
248 		/* Request to record lock address from PEBS */
249 		.sample_type = PERF_SAMPLE_ADDR,
250 		/* Record address value requires precise event */
251 		.precise_ip = 2,
252 	};
253 
254 	printf("Test HW_CPU_CYCLES\n");
255 	test_perf_event_all_cpu(&attr_type_hw);
256 	test_perf_event_task(&attr_type_hw);
257 
258 	printf("Test SW_CPU_CLOCK\n");
259 	test_perf_event_all_cpu(&attr_type_sw);
260 	test_perf_event_task(&attr_type_sw);
261 
262 	printf("Test HW_CACHE_L1D\n");
263 	test_perf_event_all_cpu(&attr_hw_cache_l1d);
264 	test_perf_event_task(&attr_hw_cache_l1d);
265 
266 	printf("Test HW_CACHE_BPU\n");
267 	test_perf_event_all_cpu(&attr_hw_cache_branch_miss);
268 	test_perf_event_task(&attr_hw_cache_branch_miss);
269 
270 	printf("Test Instruction Retired\n");
271 	test_perf_event_all_cpu(&attr_type_raw);
272 	test_perf_event_task(&attr_type_raw);
273 
274 	printf("Test Lock Load\n");
275 	test_perf_event_all_cpu(&attr_type_raw_lock_load);
276 	test_perf_event_task(&attr_type_raw_lock_load);
277 
278 	printf("*** PASS ***\n");
279 }
280 
281 
282 int main(int argc, char **argv)
283 {
284 	struct rlimit r = {RLIM_INFINITY, RLIM_INFINITY};
285 	char filename[256];
286 
287 	snprintf(filename, sizeof(filename), "%s_kern.o", argv[0]);
288 	setrlimit(RLIMIT_MEMLOCK, &r);
289 
290 	signal(SIGINT, int_exit);
291 	signal(SIGTERM, int_exit);
292 
293 	if (load_kallsyms()) {
294 		printf("failed to process /proc/kallsyms\n");
295 		return 1;
296 	}
297 
298 	if (load_bpf_file(filename)) {
299 		printf("%s", bpf_log_buf);
300 		return 2;
301 	}
302 
303 	if (fork() == 0) {
304 		read_trace_pipe();
305 		return 0;
306 	}
307 	test_bpf_perf_event();
308 	int_exit(0);
309 	return 0;
310 }
311