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 = btf__load_vmlinux_btf();
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 /* btf_trace func proto has one more argument for the context */
120 if (t3 && btf_is_func_proto(t3) && btf_vlen(t3) == 5) {
121 /* new format: pass prev_state as 4th arg */
122 skel->rodata->has_prev_state = true;
123 }
124 }
125
off_cpu_prepare(struct evlist * evlist,struct target * target,struct record_opts * opts)126 int off_cpu_prepare(struct evlist *evlist, struct target *target,
127 struct record_opts *opts)
128 {
129 int err, fd, i;
130 int ncpus = 1, ntasks = 1, ncgrps = 1;
131 struct strlist *pid_slist = NULL;
132 struct str_node *pos;
133
134 if (off_cpu_config(evlist) < 0) {
135 pr_err("Failed to config off-cpu BPF event\n");
136 return -1;
137 }
138
139 skel = off_cpu_bpf__open();
140 if (!skel) {
141 pr_err("Failed to open off-cpu BPF skeleton\n");
142 return -1;
143 }
144
145 /* don't need to set cpu filter for system-wide mode */
146 if (target->cpu_list) {
147 ncpus = perf_cpu_map__nr(evlist->core.user_requested_cpus);
148 bpf_map__set_max_entries(skel->maps.cpu_filter, ncpus);
149 }
150
151 if (target->pid) {
152 pid_slist = strlist__new(target->pid, NULL);
153 if (!pid_slist) {
154 pr_err("Failed to create a strlist for pid\n");
155 return -1;
156 }
157
158 ntasks = 0;
159 strlist__for_each_entry(pos, pid_slist) {
160 char *end_ptr;
161 int pid = strtol(pos->s, &end_ptr, 10);
162
163 if (pid == INT_MIN || pid == INT_MAX ||
164 (*end_ptr != '\0' && *end_ptr != ','))
165 continue;
166
167 ntasks++;
168 }
169
170 if (ntasks < MAX_PROC)
171 ntasks = MAX_PROC;
172
173 bpf_map__set_max_entries(skel->maps.task_filter, ntasks);
174 } else if (target__has_task(target)) {
175 ntasks = perf_thread_map__nr(evlist->core.threads);
176 bpf_map__set_max_entries(skel->maps.task_filter, ntasks);
177 } else if (target__none(target)) {
178 bpf_map__set_max_entries(skel->maps.task_filter, MAX_PROC);
179 }
180
181 if (evlist__first(evlist)->cgrp) {
182 ncgrps = evlist->core.nr_entries - 1; /* excluding a dummy */
183 bpf_map__set_max_entries(skel->maps.cgroup_filter, ncgrps);
184
185 if (!cgroup_is_v2("perf_event"))
186 skel->rodata->uses_cgroup_v1 = true;
187 }
188
189 if (opts->record_cgroup) {
190 skel->rodata->needs_cgroup = true;
191
192 if (!cgroup_is_v2("perf_event"))
193 skel->rodata->uses_cgroup_v1 = true;
194 }
195
196 set_max_rlimit();
197 check_sched_switch_args();
198
199 err = off_cpu_bpf__load(skel);
200 if (err) {
201 pr_err("Failed to load off-cpu skeleton\n");
202 goto out;
203 }
204
205 if (target->cpu_list) {
206 u32 cpu;
207 u8 val = 1;
208
209 skel->bss->has_cpu = 1;
210 fd = bpf_map__fd(skel->maps.cpu_filter);
211
212 for (i = 0; i < ncpus; i++) {
213 cpu = perf_cpu_map__cpu(evlist->core.user_requested_cpus, i).cpu;
214 bpf_map_update_elem(fd, &cpu, &val, BPF_ANY);
215 }
216 }
217
218 if (target->pid) {
219 u8 val = 1;
220
221 skel->bss->has_task = 1;
222 skel->bss->uses_tgid = 1;
223 fd = bpf_map__fd(skel->maps.task_filter);
224
225 strlist__for_each_entry(pos, pid_slist) {
226 char *end_ptr;
227 u32 tgid;
228 int pid = strtol(pos->s, &end_ptr, 10);
229
230 if (pid == INT_MIN || pid == INT_MAX ||
231 (*end_ptr != '\0' && *end_ptr != ','))
232 continue;
233
234 tgid = pid;
235 bpf_map_update_elem(fd, &tgid, &val, BPF_ANY);
236 }
237 } else if (target__has_task(target)) {
238 u32 pid;
239 u8 val = 1;
240
241 skel->bss->has_task = 1;
242 fd = bpf_map__fd(skel->maps.task_filter);
243
244 for (i = 0; i < ntasks; i++) {
245 pid = perf_thread_map__pid(evlist->core.threads, i);
246 bpf_map_update_elem(fd, &pid, &val, BPF_ANY);
247 }
248 }
249
250 if (evlist__first(evlist)->cgrp) {
251 struct evsel *evsel;
252 u8 val = 1;
253
254 skel->bss->has_cgroup = 1;
255 fd = bpf_map__fd(skel->maps.cgroup_filter);
256
257 evlist__for_each_entry(evlist, evsel) {
258 struct cgroup *cgrp = evsel->cgrp;
259
260 if (cgrp == NULL)
261 continue;
262
263 if (!cgrp->id && read_cgroup_id(cgrp) < 0) {
264 pr_err("Failed to read cgroup id of %s\n",
265 cgrp->name);
266 goto out;
267 }
268
269 bpf_map_update_elem(fd, &cgrp->id, &val, BPF_ANY);
270 }
271 }
272
273 err = off_cpu_bpf__attach(skel);
274 if (err) {
275 pr_err("Failed to attach off-cpu BPF skeleton\n");
276 goto out;
277 }
278
279 if (perf_hooks__set_hook("record_start", off_cpu_start, evlist) ||
280 perf_hooks__set_hook("record_end", off_cpu_finish, evlist)) {
281 pr_err("Failed to attach off-cpu skeleton\n");
282 goto out;
283 }
284
285 return 0;
286
287 out:
288 off_cpu_bpf__destroy(skel);
289 return -1;
290 }
291
off_cpu_write(struct perf_session * session)292 int off_cpu_write(struct perf_session *session)
293 {
294 int bytes = 0, size;
295 int fd, stack;
296 u64 sample_type, val, sid = 0;
297 struct evsel *evsel;
298 struct perf_data_file *file = &session->data->file;
299 struct off_cpu_key prev, key;
300 union off_cpu_data data = {
301 .hdr = {
302 .type = PERF_RECORD_SAMPLE,
303 .misc = PERF_RECORD_MISC_USER,
304 },
305 };
306 u64 tstamp = OFF_CPU_TIMESTAMP;
307
308 skel->bss->enabled = 0;
309
310 evsel = evlist__find_evsel_by_str(session->evlist, OFFCPU_EVENT);
311 if (evsel == NULL) {
312 pr_err("%s evsel not found\n", OFFCPU_EVENT);
313 return 0;
314 }
315
316 sample_type = evsel->core.attr.sample_type;
317
318 if (sample_type & ~OFFCPU_SAMPLE_TYPES) {
319 pr_err("not supported sample type: %llx\n",
320 (unsigned long long)sample_type);
321 return -1;
322 }
323
324 if (sample_type & (PERF_SAMPLE_ID | PERF_SAMPLE_IDENTIFIER)) {
325 if (evsel->core.id)
326 sid = evsel->core.id[0];
327 }
328
329 fd = bpf_map__fd(skel->maps.off_cpu);
330 stack = bpf_map__fd(skel->maps.stacks);
331 memset(&prev, 0, sizeof(prev));
332
333 while (!bpf_map_get_next_key(fd, &prev, &key)) {
334 int n = 1; /* start from perf_event_header */
335 int ip_pos = -1;
336
337 bpf_map_lookup_elem(fd, &key, &val);
338
339 if (sample_type & PERF_SAMPLE_IDENTIFIER)
340 data.array[n++] = sid;
341 if (sample_type & PERF_SAMPLE_IP) {
342 ip_pos = n;
343 data.array[n++] = 0; /* will be updated */
344 }
345 if (sample_type & PERF_SAMPLE_TID)
346 data.array[n++] = (u64)key.pid << 32 | key.tgid;
347 if (sample_type & PERF_SAMPLE_TIME)
348 data.array[n++] = tstamp;
349 if (sample_type & PERF_SAMPLE_ID)
350 data.array[n++] = sid;
351 if (sample_type & PERF_SAMPLE_CPU)
352 data.array[n++] = 0;
353 if (sample_type & PERF_SAMPLE_PERIOD)
354 data.array[n++] = val;
355 if (sample_type & PERF_SAMPLE_CALLCHAIN) {
356 int len = 0;
357
358 /* data.array[n] is callchain->nr (updated later) */
359 data.array[n + 1] = PERF_CONTEXT_USER;
360 data.array[n + 2] = 0;
361
362 bpf_map_lookup_elem(stack, &key.stack_id, &data.array[n + 2]);
363 while (data.array[n + 2 + len])
364 len++;
365
366 /* update length of callchain */
367 data.array[n] = len + 1;
368
369 /* update sample ip with the first callchain entry */
370 if (ip_pos >= 0)
371 data.array[ip_pos] = data.array[n + 2];
372
373 /* calculate sample callchain data array length */
374 n += len + 2;
375 }
376 if (sample_type & PERF_SAMPLE_CGROUP)
377 data.array[n++] = key.cgroup_id;
378
379 size = n * sizeof(u64);
380 data.hdr.size = size;
381 bytes += size;
382
383 if (perf_data_file__write(file, &data, size) < 0) {
384 pr_err("failed to write perf data, error: %m\n");
385 return bytes;
386 }
387
388 prev = key;
389 /* increase dummy timestamp to sort later samples */
390 tstamp++;
391 }
392 return bytes;
393 }
394