1 // SPDX-License-Identifier: GPL-2.0
2 #include "util/bpf_counter.h"
3 #include "util/debug.h"
4 #include "util/evsel.h"
5 #include "util/evlist.h"
6 #include "util/off_cpu.h"
7 #include "util/perf-hooks.h"
8 #include "util/record.h"
9 #include "util/session.h"
10 #include "util/target.h"
11 #include "util/cpumap.h"
12 #include "util/thread_map.h"
13 #include "util/cgroup.h"
14 #include "util/strlist.h"
15 #include <bpf/bpf.h>
16 
17 #include "bpf_skel/off_cpu.skel.h"
18 
19 #define MAX_STACKS  32
20 #define MAX_PROC  4096
21 /* we don't need actual timestamp, just want to put the samples at last */
22 #define OFF_CPU_TIMESTAMP  (~0ull << 32)
23 
24 static struct off_cpu_bpf *skel;
25 
26 struct off_cpu_key {
27 	u32 pid;
28 	u32 tgid;
29 	u32 stack_id;
30 	u32 state;
31 	u64 cgroup_id;
32 };
33 
34 union off_cpu_data {
35 	struct perf_event_header hdr;
36 	u64 array[1024 / sizeof(u64)];
37 };
38 
off_cpu_config(struct evlist * evlist)39 static int off_cpu_config(struct evlist *evlist)
40 {
41 	struct evsel *evsel;
42 	struct perf_event_attr attr = {
43 		.type	= PERF_TYPE_SOFTWARE,
44 		.config = PERF_COUNT_SW_BPF_OUTPUT,
45 		.size	= sizeof(attr), /* to capture ABI version */
46 	};
47 	char *evname = strdup(OFFCPU_EVENT);
48 
49 	if (evname == NULL)
50 		return -ENOMEM;
51 
52 	evsel = evsel__new(&attr);
53 	if (!evsel) {
54 		free(evname);
55 		return -ENOMEM;
56 	}
57 
58 	evsel->core.attr.freq = 1;
59 	evsel->core.attr.sample_period = 1;
60 	/* off-cpu analysis depends on stack trace */
61 	evsel->core.attr.sample_type = PERF_SAMPLE_CALLCHAIN;
62 
63 	evlist__add(evlist, evsel);
64 
65 	free(evsel->name);
66 	evsel->name = evname;
67 
68 	return 0;
69 }
70 
off_cpu_start(void * arg)71 static void off_cpu_start(void *arg)
72 {
73 	struct evlist *evlist = arg;
74 
75 	/* update task filter for the given workload */
76 	if (!skel->bss->has_cpu && !skel->bss->has_task &&
77 	    perf_thread_map__pid(evlist->core.threads, 0) != -1) {
78 		int fd;
79 		u32 pid;
80 		u8 val = 1;
81 
82 		skel->bss->has_task = 1;
83 		skel->bss->uses_tgid = 1;
84 		fd = bpf_map__fd(skel->maps.task_filter);
85 		pid = perf_thread_map__pid(evlist->core.threads, 0);
86 		bpf_map_update_elem(fd, &pid, &val, BPF_ANY);
87 	}
88 
89 	skel->bss->enabled = 1;
90 }
91 
off_cpu_finish(void * arg __maybe_unused)92 static void off_cpu_finish(void *arg __maybe_unused)
93 {
94 	skel->bss->enabled = 0;
95 	off_cpu_bpf__destroy(skel);
96 }
97 
98 /* v5.18 kernel added prev_state arg, so it needs to check the signature */
check_sched_switch_args(void)99 static void check_sched_switch_args(void)
100 {
101 	const struct btf *btf = bpf_object__btf(skel->obj);
102 	const struct btf_type *t1, *t2, *t3;
103 	u32 type_id;
104 
105 	type_id = btf__find_by_name_kind(btf, "btf_trace_sched_switch",
106 					 BTF_KIND_TYPEDEF);
107 	if ((s32)type_id < 0)
108 		return;
109 
110 	t1 = btf__type_by_id(btf, type_id);
111 	if (t1 == NULL)
112 		return;
113 
114 	t2 = btf__type_by_id(btf, t1->type);
115 	if (t2 == NULL || !btf_is_ptr(t2))
116 		return;
117 
118 	t3 = btf__type_by_id(btf, t2->type);
119 	if (t3 && btf_is_func_proto(t3) && btf_vlen(t3) == 4) {
120 		/* new format: pass prev_state as 4th arg */
121 		skel->rodata->has_prev_state = true;
122 	}
123 }
124 
off_cpu_prepare(struct evlist * evlist,struct target * target,struct record_opts * opts)125 int off_cpu_prepare(struct evlist *evlist, struct target *target,
126 		    struct record_opts *opts)
127 {
128 	int err, fd, i;
129 	int ncpus = 1, ntasks = 1, ncgrps = 1;
130 	struct strlist *pid_slist = NULL;
131 	struct str_node *pos;
132 
133 	if (off_cpu_config(evlist) < 0) {
134 		pr_err("Failed to config off-cpu BPF event\n");
135 		return -1;
136 	}
137 
138 	skel = off_cpu_bpf__open();
139 	if (!skel) {
140 		pr_err("Failed to open off-cpu BPF skeleton\n");
141 		return -1;
142 	}
143 
144 	/* don't need to set cpu filter for system-wide mode */
145 	if (target->cpu_list) {
146 		ncpus = perf_cpu_map__nr(evlist->core.user_requested_cpus);
147 		bpf_map__set_max_entries(skel->maps.cpu_filter, ncpus);
148 	}
149 
150 	if (target->pid) {
151 		pid_slist = strlist__new(target->pid, NULL);
152 		if (!pid_slist) {
153 			pr_err("Failed to create a strlist for pid\n");
154 			return -1;
155 		}
156 
157 		ntasks = 0;
158 		strlist__for_each_entry(pos, pid_slist) {
159 			char *end_ptr;
160 			int pid = strtol(pos->s, &end_ptr, 10);
161 
162 			if (pid == INT_MIN || pid == INT_MAX ||
163 			    (*end_ptr != '\0' && *end_ptr != ','))
164 				continue;
165 
166 			ntasks++;
167 		}
168 
169 		if (ntasks < MAX_PROC)
170 			ntasks = MAX_PROC;
171 
172 		bpf_map__set_max_entries(skel->maps.task_filter, ntasks);
173 	} else if (target__has_task(target)) {
174 		ntasks = perf_thread_map__nr(evlist->core.threads);
175 		bpf_map__set_max_entries(skel->maps.task_filter, ntasks);
176 	} else if (target__none(target)) {
177 		bpf_map__set_max_entries(skel->maps.task_filter, MAX_PROC);
178 	}
179 
180 	if (evlist__first(evlist)->cgrp) {
181 		ncgrps = evlist->core.nr_entries - 1; /* excluding a dummy */
182 		bpf_map__set_max_entries(skel->maps.cgroup_filter, ncgrps);
183 
184 		if (!cgroup_is_v2("perf_event"))
185 			skel->rodata->uses_cgroup_v1 = true;
186 	}
187 
188 	if (opts->record_cgroup) {
189 		skel->rodata->needs_cgroup = true;
190 
191 		if (!cgroup_is_v2("perf_event"))
192 			skel->rodata->uses_cgroup_v1 = true;
193 	}
194 
195 	set_max_rlimit();
196 	check_sched_switch_args();
197 
198 	err = off_cpu_bpf__load(skel);
199 	if (err) {
200 		pr_err("Failed to load off-cpu skeleton\n");
201 		goto out;
202 	}
203 
204 	if (target->cpu_list) {
205 		u32 cpu;
206 		u8 val = 1;
207 
208 		skel->bss->has_cpu = 1;
209 		fd = bpf_map__fd(skel->maps.cpu_filter);
210 
211 		for (i = 0; i < ncpus; i++) {
212 			cpu = perf_cpu_map__cpu(evlist->core.user_requested_cpus, i).cpu;
213 			bpf_map_update_elem(fd, &cpu, &val, BPF_ANY);
214 		}
215 	}
216 
217 	if (target->pid) {
218 		u8 val = 1;
219 
220 		skel->bss->has_task = 1;
221 		skel->bss->uses_tgid = 1;
222 		fd = bpf_map__fd(skel->maps.task_filter);
223 
224 		strlist__for_each_entry(pos, pid_slist) {
225 			char *end_ptr;
226 			u32 tgid;
227 			int pid = strtol(pos->s, &end_ptr, 10);
228 
229 			if (pid == INT_MIN || pid == INT_MAX ||
230 			    (*end_ptr != '\0' && *end_ptr != ','))
231 				continue;
232 
233 			tgid = pid;
234 			bpf_map_update_elem(fd, &tgid, &val, BPF_ANY);
235 		}
236 	} else if (target__has_task(target)) {
237 		u32 pid;
238 		u8 val = 1;
239 
240 		skel->bss->has_task = 1;
241 		fd = bpf_map__fd(skel->maps.task_filter);
242 
243 		for (i = 0; i < ntasks; i++) {
244 			pid = perf_thread_map__pid(evlist->core.threads, i);
245 			bpf_map_update_elem(fd, &pid, &val, BPF_ANY);
246 		}
247 	}
248 
249 	if (evlist__first(evlist)->cgrp) {
250 		struct evsel *evsel;
251 		u8 val = 1;
252 
253 		skel->bss->has_cgroup = 1;
254 		fd = bpf_map__fd(skel->maps.cgroup_filter);
255 
256 		evlist__for_each_entry(evlist, evsel) {
257 			struct cgroup *cgrp = evsel->cgrp;
258 
259 			if (cgrp == NULL)
260 				continue;
261 
262 			if (!cgrp->id && read_cgroup_id(cgrp) < 0) {
263 				pr_err("Failed to read cgroup id of %s\n",
264 				       cgrp->name);
265 				goto out;
266 			}
267 
268 			bpf_map_update_elem(fd, &cgrp->id, &val, BPF_ANY);
269 		}
270 	}
271 
272 	err = off_cpu_bpf__attach(skel);
273 	if (err) {
274 		pr_err("Failed to attach off-cpu BPF skeleton\n");
275 		goto out;
276 	}
277 
278 	if (perf_hooks__set_hook("record_start", off_cpu_start, evlist) ||
279 	    perf_hooks__set_hook("record_end", off_cpu_finish, evlist)) {
280 		pr_err("Failed to attach off-cpu skeleton\n");
281 		goto out;
282 	}
283 
284 	return 0;
285 
286 out:
287 	off_cpu_bpf__destroy(skel);
288 	return -1;
289 }
290 
off_cpu_write(struct perf_session * session)291 int off_cpu_write(struct perf_session *session)
292 {
293 	int bytes = 0, size;
294 	int fd, stack;
295 	u64 sample_type, val, sid = 0;
296 	struct evsel *evsel;
297 	struct perf_data_file *file = &session->data->file;
298 	struct off_cpu_key prev, key;
299 	union off_cpu_data data = {
300 		.hdr = {
301 			.type = PERF_RECORD_SAMPLE,
302 			.misc = PERF_RECORD_MISC_USER,
303 		},
304 	};
305 	u64 tstamp = OFF_CPU_TIMESTAMP;
306 
307 	skel->bss->enabled = 0;
308 
309 	evsel = evlist__find_evsel_by_str(session->evlist, OFFCPU_EVENT);
310 	if (evsel == NULL) {
311 		pr_err("%s evsel not found\n", OFFCPU_EVENT);
312 		return 0;
313 	}
314 
315 	sample_type = evsel->core.attr.sample_type;
316 
317 	if (sample_type & ~OFFCPU_SAMPLE_TYPES) {
318 		pr_err("not supported sample type: %llx\n",
319 		       (unsigned long long)sample_type);
320 		return -1;
321 	}
322 
323 	if (sample_type & (PERF_SAMPLE_ID | PERF_SAMPLE_IDENTIFIER)) {
324 		if (evsel->core.id)
325 			sid = evsel->core.id[0];
326 	}
327 
328 	fd = bpf_map__fd(skel->maps.off_cpu);
329 	stack = bpf_map__fd(skel->maps.stacks);
330 	memset(&prev, 0, sizeof(prev));
331 
332 	while (!bpf_map_get_next_key(fd, &prev, &key)) {
333 		int n = 1;  /* start from perf_event_header */
334 		int ip_pos = -1;
335 
336 		bpf_map_lookup_elem(fd, &key, &val);
337 
338 		if (sample_type & PERF_SAMPLE_IDENTIFIER)
339 			data.array[n++] = sid;
340 		if (sample_type & PERF_SAMPLE_IP) {
341 			ip_pos = n;
342 			data.array[n++] = 0;  /* will be updated */
343 		}
344 		if (sample_type & PERF_SAMPLE_TID)
345 			data.array[n++] = (u64)key.pid << 32 | key.tgid;
346 		if (sample_type & PERF_SAMPLE_TIME)
347 			data.array[n++] = tstamp;
348 		if (sample_type & PERF_SAMPLE_ID)
349 			data.array[n++] = sid;
350 		if (sample_type & PERF_SAMPLE_CPU)
351 			data.array[n++] = 0;
352 		if (sample_type & PERF_SAMPLE_PERIOD)
353 			data.array[n++] = val;
354 		if (sample_type & PERF_SAMPLE_CALLCHAIN) {
355 			int len = 0;
356 
357 			/* data.array[n] is callchain->nr (updated later) */
358 			data.array[n + 1] = PERF_CONTEXT_USER;
359 			data.array[n + 2] = 0;
360 
361 			bpf_map_lookup_elem(stack, &key.stack_id, &data.array[n + 2]);
362 			while (data.array[n + 2 + len])
363 				len++;
364 
365 			/* update length of callchain */
366 			data.array[n] = len + 1;
367 
368 			/* update sample ip with the first callchain entry */
369 			if (ip_pos >= 0)
370 				data.array[ip_pos] = data.array[n + 2];
371 
372 			/* calculate sample callchain data array length */
373 			n += len + 2;
374 		}
375 		if (sample_type & PERF_SAMPLE_CGROUP)
376 			data.array[n++] = key.cgroup_id;
377 
378 		size = n * sizeof(u64);
379 		data.hdr.size = size;
380 		bytes += size;
381 
382 		if (perf_data_file__write(file, &data, size) < 0) {
383 			pr_err("failed to write perf data, error: %m\n");
384 			return bytes;
385 		}
386 
387 		prev = key;
388 		/* increase dummy timestamp to sort later samples */
389 		tstamp++;
390 	}
391 	return bytes;
392 }
393