xref: /aosp_15_r20/external/bcc/libbpf-tools/runqlen.c (revision 387f9dfdfa2baef462e92476d413c7bc2470293e)
1 // SPDX-License-Identifier: (LGPL-2.1 OR BSD-2-Clause)
2 // Copyright (c) 2020 Wenbo Zhang
3 //
4 // Based on runqlen(8) from BCC by Brendan Gregg.
5 // 11-Sep-2020   Wenbo Zhang   Created this.
6 #include <argp.h>
7 #include <signal.h>
8 #include <stdio.h>
9 #include <stdlib.h>
10 #include <string.h>
11 #include <unistd.h>
12 #include <time.h>
13 #include <linux/perf_event.h>
14 #include <asm/unistd.h>
15 #include <bpf/libbpf.h>
16 #include <bpf/bpf.h>
17 #include "runqlen.h"
18 #include "runqlen.skel.h"
19 #include "btf_helpers.h"
20 #include "trace_helpers.h"
21 
22 #define max(x, y) ({				 \
23 	typeof(x) _max1 = (x);			 \
24 	typeof(y) _max2 = (y);			 \
25 	(void) (&_max1 == &_max2);		 \
26 	_max1 > _max2 ? _max1 : _max2; })
27 
28 struct env {
29 	bool per_cpu;
30 	bool runqocc;
31 	bool timestamp;
32 	bool host;
33 	time_t interval;
34 	int freq;
35 	int times;
36 	bool verbose;
37 } env = {
38 	.interval = 99999999,
39 	.times = 99999999,
40 	.freq = 99,
41 };
42 
43 static volatile bool exiting;
44 
45 const char *argp_program_version = "runqlen 0.1";
46 const char *argp_program_bug_address =
47 	"https://github.com/iovisor/bcc/tree/master/libbpf-tools";
48 const char argp_program_doc[] =
49 "Summarize scheduler run queue length as a histogram.\n"
50 "\n"
51 "USAGE: runqlen [--help] [-C] [-O] [-T] [-f FREQUENCY] [interval] [count]\n"
52 "\n"
53 "EXAMPLES:\n"
54 "    runqlen         # summarize run queue length as a histogram\n"
55 "    runqlen 1 10    # print 1 second summaries, 10 times\n"
56 "    runqlen -T 1    # 1s summaries and timestamps\n"
57 "    runqlen -O      # report run queue occupancy\n"
58 "    runqlen -C      # show each CPU separately\n"
59 "    runqlen -H      # show nr_running from host's rq instead of cfs_rq\n"
60 "    runqlen -f 199  # sample at 199HZ\n";
61 
62 static const struct argp_option opts[] = {
63 	{ "cpus", 'C', NULL, 0, "Print output for each CPU separately" },
64 	{ "frequency", 'f', "FREQUENCY", 0, "Sample with a certain frequency" },
65 	{ "runqocc", 'O', NULL, 0, "Report run queue occupancy" },
66 	{ "timestamp", 'T', NULL, 0, "Include timestamp on output" },
67 	{ "verbose", 'v', NULL, 0, "Verbose debug output" },
68 	{ "host", 'H', NULL, 0, "Report nr_running from host's rq" },
69 	{ NULL, 'h', NULL, OPTION_HIDDEN, "Show the full help" },
70 	{},
71 };
72 
parse_arg(int key,char * arg,struct argp_state * state)73 static error_t parse_arg(int key, char *arg, struct argp_state *state)
74 {
75 	static int pos_args;
76 
77 	switch (key) {
78 	case 'h':
79 		argp_state_help(state, stderr, ARGP_HELP_STD_HELP);
80 		break;
81 	case 'v':
82 		env.verbose = true;
83 		break;
84 	case 'C':
85 		env.per_cpu = true;
86 		break;
87 	case 'O':
88 		env.runqocc = true;
89 		break;
90 	case 'T':
91 		env.timestamp = true;
92 		break;
93 	case 'H':
94 		env.host = true;
95 		break;
96 	case 'f':
97 		errno = 0;
98 		env.freq = strtol(arg, NULL, 10);
99 		if (errno || env.freq <= 0) {
100 			fprintf(stderr, "Invalid freq (in hz): %s\n", arg);
101 			argp_usage(state);
102 		}
103 		break;
104 	case ARGP_KEY_ARG:
105 		errno = 0;
106 		if (pos_args == 0) {
107 			env.interval = strtol(arg, NULL, 10);
108 			if (errno) {
109 				fprintf(stderr, "invalid internal\n");
110 				argp_usage(state);
111 			}
112 		} else if (pos_args == 1) {
113 			env.times = strtol(arg, NULL, 10);
114 			if (errno) {
115 				fprintf(stderr, "invalid times\n");
116 				argp_usage(state);
117 			}
118 		} else {
119 			fprintf(stderr,
120 				"unrecognized positional argument: %s\n", arg);
121 			argp_usage(state);
122 		}
123 		pos_args++;
124 		break;
125 	default:
126 		return ARGP_ERR_UNKNOWN;
127 	}
128 	return 0;
129 }
130 
131 static int nr_cpus;
132 
open_and_attach_perf_event(int freq,struct bpf_program * prog,struct bpf_link * links[])133 static int open_and_attach_perf_event(int freq, struct bpf_program *prog,
134 				struct bpf_link *links[])
135 {
136 	struct perf_event_attr attr = {
137 		.type = PERF_TYPE_SOFTWARE,
138 		.freq = 1,
139 		.sample_period = freq,
140 		.config = PERF_COUNT_SW_CPU_CLOCK,
141 	};
142 	int i, fd;
143 
144 	for (i = 0; i < nr_cpus; i++) {
145 		fd = syscall(__NR_perf_event_open, &attr, -1, i, -1, 0);
146 		if (fd < 0) {
147 			/* Ignore CPU that is offline */
148 			if (errno == ENODEV)
149 				continue;
150 			fprintf(stderr, "failed to init perf sampling: %s\n",
151 				strerror(errno));
152 			return -1;
153 		}
154 		links[i] = bpf_program__attach_perf_event(prog, fd);
155 		if (!links[i]) {
156 			fprintf(stderr, "failed to attach perf event on cpu: %d\n", i);
157 			close(fd);
158 			return -1;
159 		}
160 	}
161 
162 	return 0;
163 }
164 
libbpf_print_fn(enum libbpf_print_level level,const char * format,va_list args)165 static int libbpf_print_fn(enum libbpf_print_level level, const char *format, va_list args)
166 {
167 	if (level == LIBBPF_DEBUG && !env.verbose)
168 		return 0;
169 	return vfprintf(stderr, format, args);
170 }
171 
sig_handler(int sig)172 static void sig_handler(int sig)
173 {
174 	exiting = true;
175 }
176 
177 static struct hist zero;
178 
print_runq_occupancy(struct runqlen_bpf__bss * bss)179 static void print_runq_occupancy(struct runqlen_bpf__bss *bss)
180 {
181 	struct hist hist;
182 	int slot, i = 0;
183 	float runqocc;
184 
185 	do {
186 		__u64 samples, idle = 0, queued = 0;
187 
188 		hist = bss->hists[i];
189 		bss->hists[i] = zero;
190 		for (slot = 0; slot < MAX_SLOTS; slot++) {
191 			__u64 val = hist.slots[slot];
192 
193 			if (slot == 0)
194 				idle += val;
195 			else
196 				queued += val;
197 		}
198 		samples = idle + queued;
199 		runqocc = queued * 1.0 / max(1ULL, samples);
200 		if (env.per_cpu)
201 			printf("runqocc, CPU %-3d %6.2f%%\n", i,
202 				100 * runqocc);
203 		else
204 			printf("runqocc: %0.2f%%\n", 100 * runqocc);
205 	} while (env.per_cpu && ++i < nr_cpus);
206 }
207 
print_linear_hists(struct runqlen_bpf__bss * bss)208 static void print_linear_hists(struct runqlen_bpf__bss *bss)
209 {
210 	struct hist hist;
211 	int i = 0;
212 
213 	do {
214 		hist = bss->hists[i];
215 		bss->hists[i] = zero;
216 		if (env.per_cpu)
217 			printf("cpu = %d\n", i);
218 		print_linear_hist(hist.slots, MAX_SLOTS, 0, 1, "runqlen");
219 	} while (env.per_cpu && ++i < nr_cpus);
220 }
221 
main(int argc,char ** argv)222 int main(int argc, char **argv)
223 {
224 	LIBBPF_OPTS(bpf_object_open_opts, open_opts);
225 	static const struct argp argp = {
226 		.options = opts,
227 		.parser = parse_arg,
228 		.doc = argp_program_doc,
229 	};
230 	struct bpf_link *links[MAX_CPU_NR] = {};
231 	struct runqlen_bpf *obj;
232 	struct tm *tm;
233 	char ts[32];
234 	int err, i;
235 	time_t t;
236 
237 	err = argp_parse(&argp, argc, argv, 0, NULL, NULL);
238 	if (err)
239 		return err;
240 
241 	libbpf_set_print(libbpf_print_fn);
242 
243 	nr_cpus = libbpf_num_possible_cpus();
244 	if (nr_cpus < 0) {
245 		printf("failed to get # of possible cpus: '%s'!\n",
246 		       strerror(-nr_cpus));
247 		return 1;
248 	}
249 	if (nr_cpus > MAX_CPU_NR) {
250 		fprintf(stderr, "the number of cpu cores is too big, please "
251 			"increase MAX_CPU_NR's value and recompile");
252 		return 1;
253 	}
254 
255 	err = ensure_core_btf(&open_opts);
256 	if (err) {
257 		fprintf(stderr, "failed to fetch necessary BTF for CO-RE: %s\n", strerror(-err));
258 		return 1;
259 	}
260 
261 	obj = runqlen_bpf__open_opts(&open_opts);
262 	if (!obj) {
263 		fprintf(stderr, "failed to open BPF object\n");
264 		return 1;
265 	}
266 
267 	/* initialize global data (filtering options) */
268 	obj->rodata->targ_per_cpu = env.per_cpu;
269 	obj->rodata->targ_host = env.host;
270 
271 	err = runqlen_bpf__load(obj);
272 	if (err) {
273 		fprintf(stderr, "failed to load BPF object: %d\n", err);
274 		goto cleanup;
275 	}
276 
277 	if (!obj->bss) {
278 		fprintf(stderr, "Memory-mapping BPF maps is supported starting from Linux 5.7, please upgrade.\n");
279 		goto cleanup;
280 	}
281 
282 	err = open_and_attach_perf_event(env.freq, obj->progs.do_sample, links);
283 	if (err)
284 		goto cleanup;
285 
286 	printf("Sampling run queue length... Hit Ctrl-C to end.\n");
287 
288 	signal(SIGINT, sig_handler);
289 
290 	while (1) {
291 		sleep(env.interval);
292 		printf("\n");
293 
294 		if (env.timestamp) {
295 			time(&t);
296 			tm = localtime(&t);
297 			strftime(ts, sizeof(ts), "%H:%M:%S", tm);
298 			printf("%-8s\n", ts);
299 		}
300 
301 		if (env.runqocc)
302 			print_runq_occupancy(obj->bss);
303 		else
304 			print_linear_hists(obj->bss);
305 
306 		if (exiting || --env.times == 0)
307 			break;
308 	}
309 
310 cleanup:
311 	for (i = 0; i < nr_cpus; i++)
312 		bpf_link__destroy(links[i]);
313 	runqlen_bpf__destroy(obj);
314 	cleanup_core_btf(&open_opts);
315 
316 	return err != 0;
317 }
318