1 /*
2 * Copyright (C) 2011, Red Hat Inc, Arnaldo Carvalho de Melo <acme@redhat.com>
3 *
4 * Parts came from builtin-{top,stat,record}.c, see those files for further
5 * copyright notes.
6 *
7 * Released under the GPL v2. (and only v2, not any later version)
8 */
9
10 #include "evsel.h"
11 #include "evlist.h"
12 #include "util.h"
13 #include "cpumap.h"
14 #include "thread_map.h"
15
16 #define FD(e, x, y) (*(int *)xyarray__entry(e->fd, x, y))
17
__perf_evsel__sample_size(u64 sample_type)18 int __perf_evsel__sample_size(u64 sample_type)
19 {
20 u64 mask = sample_type & PERF_SAMPLE_MASK;
21 int size = 0;
22 int i;
23
24 for (i = 0; i < 64; i++) {
25 if (mask & (1ULL << i))
26 size++;
27 }
28
29 size *= sizeof(u64);
30
31 return size;
32 }
33
perf_evsel__init(struct perf_evsel * evsel,struct perf_event_attr * attr,int idx)34 void perf_evsel__init(struct perf_evsel *evsel,
35 struct perf_event_attr *attr, int idx)
36 {
37 evsel->idx = idx;
38 evsel->attr = *attr;
39 INIT_LIST_HEAD(&evsel->node);
40 }
41
perf_evsel__new(struct perf_event_attr * attr,int idx)42 struct perf_evsel *perf_evsel__new(struct perf_event_attr *attr, int idx)
43 {
44 struct perf_evsel *evsel = zalloc(sizeof(*evsel));
45
46 if (evsel != NULL)
47 perf_evsel__init(evsel, attr, idx);
48
49 return evsel;
50 }
51
perf_evsel__alloc_fd(struct perf_evsel * evsel,int ncpus,int nthreads)52 int perf_evsel__alloc_fd(struct perf_evsel *evsel, int ncpus, int nthreads)
53 {
54 int cpu, thread;
55 evsel->fd = xyarray__new(ncpus, nthreads, sizeof(int));
56
57 if (evsel->fd) {
58 for (cpu = 0; cpu < ncpus; cpu++) {
59 for (thread = 0; thread < nthreads; thread++) {
60 FD(evsel, cpu, thread) = -1;
61 }
62 }
63 }
64
65 return evsel->fd != NULL ? 0 : -ENOMEM;
66 }
67
perf_evsel__alloc_id(struct perf_evsel * evsel,int ncpus,int nthreads)68 int perf_evsel__alloc_id(struct perf_evsel *evsel, int ncpus, int nthreads)
69 {
70 evsel->sample_id = xyarray__new(ncpus, nthreads, sizeof(struct perf_sample_id));
71 if (evsel->sample_id == NULL)
72 return -ENOMEM;
73
74 evsel->id = zalloc(ncpus * nthreads * sizeof(u64));
75 if (evsel->id == NULL) {
76 xyarray__delete(evsel->sample_id);
77 evsel->sample_id = NULL;
78 return -ENOMEM;
79 }
80
81 return 0;
82 }
83
perf_evsel__alloc_counts(struct perf_evsel * evsel,int ncpus)84 int perf_evsel__alloc_counts(struct perf_evsel *evsel, int ncpus)
85 {
86 evsel->counts = zalloc((sizeof(*evsel->counts) +
87 (ncpus * sizeof(struct perf_counts_values))));
88 return evsel->counts != NULL ? 0 : -ENOMEM;
89 }
90
perf_evsel__free_fd(struct perf_evsel * evsel)91 void perf_evsel__free_fd(struct perf_evsel *evsel)
92 {
93 xyarray__delete(evsel->fd);
94 evsel->fd = NULL;
95 }
96
perf_evsel__free_id(struct perf_evsel * evsel)97 void perf_evsel__free_id(struct perf_evsel *evsel)
98 {
99 xyarray__delete(evsel->sample_id);
100 evsel->sample_id = NULL;
101 free(evsel->id);
102 evsel->id = NULL;
103 }
104
perf_evsel__close_fd(struct perf_evsel * evsel,int ncpus,int nthreads)105 void perf_evsel__close_fd(struct perf_evsel *evsel, int ncpus, int nthreads)
106 {
107 int cpu, thread;
108
109 for (cpu = 0; cpu < ncpus; cpu++)
110 for (thread = 0; thread < nthreads; ++thread) {
111 close(FD(evsel, cpu, thread));
112 FD(evsel, cpu, thread) = -1;
113 }
114 }
115
perf_evsel__exit(struct perf_evsel * evsel)116 void perf_evsel__exit(struct perf_evsel *evsel)
117 {
118 assert(list_empty(&evsel->node));
119 xyarray__delete(evsel->fd);
120 xyarray__delete(evsel->sample_id);
121 free(evsel->id);
122 }
123
perf_evsel__delete(struct perf_evsel * evsel)124 void perf_evsel__delete(struct perf_evsel *evsel)
125 {
126 perf_evsel__exit(evsel);
127 close_cgroup(evsel->cgrp);
128 free(evsel->name);
129 free(evsel);
130 }
131
__perf_evsel__read_on_cpu(struct perf_evsel * evsel,int cpu,int thread,bool scale)132 int __perf_evsel__read_on_cpu(struct perf_evsel *evsel,
133 int cpu, int thread, bool scale)
134 {
135 struct perf_counts_values count;
136 size_t nv = scale ? 3 : 1;
137
138 if (FD(evsel, cpu, thread) < 0)
139 return -EINVAL;
140
141 if (evsel->counts == NULL && perf_evsel__alloc_counts(evsel, cpu + 1) < 0)
142 return -ENOMEM;
143
144 if (readn(FD(evsel, cpu, thread), &count, nv * sizeof(u64)) < 0)
145 return -errno;
146
147 if (scale) {
148 if (count.run == 0)
149 count.val = 0;
150 else if (count.run < count.ena)
151 count.val = (u64)((double)count.val * count.ena / count.run + 0.5);
152 } else
153 count.ena = count.run = 0;
154
155 evsel->counts->cpu[cpu] = count;
156 return 0;
157 }
158
__perf_evsel__read(struct perf_evsel * evsel,int ncpus,int nthreads,bool scale)159 int __perf_evsel__read(struct perf_evsel *evsel,
160 int ncpus, int nthreads, bool scale)
161 {
162 size_t nv = scale ? 3 : 1;
163 int cpu, thread;
164 struct perf_counts_values *aggr = &evsel->counts->aggr, count;
165
166 aggr->val = aggr->ena = aggr->run = 0;
167
168 for (cpu = 0; cpu < ncpus; cpu++) {
169 for (thread = 0; thread < nthreads; thread++) {
170 if (FD(evsel, cpu, thread) < 0)
171 continue;
172
173 if (readn(FD(evsel, cpu, thread),
174 &count, nv * sizeof(u64)) < 0)
175 return -errno;
176
177 aggr->val += count.val;
178 if (scale) {
179 aggr->ena += count.ena;
180 aggr->run += count.run;
181 }
182 }
183 }
184
185 evsel->counts->scaled = 0;
186 if (scale) {
187 if (aggr->run == 0) {
188 evsel->counts->scaled = -1;
189 aggr->val = 0;
190 return 0;
191 }
192
193 if (aggr->run < aggr->ena) {
194 evsel->counts->scaled = 1;
195 aggr->val = (u64)((double)aggr->val * aggr->ena / aggr->run + 0.5);
196 }
197 } else
198 aggr->ena = aggr->run = 0;
199
200 return 0;
201 }
202
__perf_evsel__open(struct perf_evsel * evsel,struct cpu_map * cpus,struct thread_map * threads,bool group)203 static int __perf_evsel__open(struct perf_evsel *evsel, struct cpu_map *cpus,
204 struct thread_map *threads, bool group)
205 {
206 /* ANDROID_CHANGE_BEGIN */
207 #ifndef __APPLE__
208 int cpu, thread;
209 unsigned long flags = 0;
210 int pid = -1;
211
212 if (evsel->fd == NULL &&
213 perf_evsel__alloc_fd(evsel, cpus->nr, threads->nr) < 0)
214 return -1;
215
216 if (evsel->cgrp) {
217 flags = PERF_FLAG_PID_CGROUP;
218 pid = evsel->cgrp->fd;
219 }
220
221 for (cpu = 0; cpu < cpus->nr; cpu++) {
222 int group_fd = -1;
223
224 for (thread = 0; thread < threads->nr; thread++) {
225
226 if (!evsel->cgrp)
227 pid = threads->map[thread];
228
229 FD(evsel, cpu, thread) = sys_perf_event_open(&evsel->attr,
230 pid,
231 cpus->map[cpu],
232 group_fd, flags);
233 if (FD(evsel, cpu, thread) < 0)
234 goto out_close;
235
236 if (group && group_fd == -1)
237 group_fd = FD(evsel, cpu, thread);
238 }
239 }
240
241 return 0;
242
243 out_close:
244 do {
245 while (--thread >= 0) {
246 close(FD(evsel, cpu, thread));
247 FD(evsel, cpu, thread) = -1;
248 }
249 thread = threads->nr;
250 } while (--cpu >= 0);
251 return -1;
252 #else
253 return -1;
254 #endif
255 /* ANDROID_CHANGE_END */
256 }
257
258 static struct {
259 struct cpu_map map;
260 int cpus[1];
261 } empty_cpu_map = {
262 .map.nr = 1,
263 .cpus = { -1, },
264 };
265
266 static struct {
267 struct thread_map map;
268 int threads[1];
269 } empty_thread_map = {
270 .map.nr = 1,
271 .threads = { -1, },
272 };
273
perf_evsel__open(struct perf_evsel * evsel,struct cpu_map * cpus,struct thread_map * threads,bool group)274 int perf_evsel__open(struct perf_evsel *evsel, struct cpu_map *cpus,
275 struct thread_map *threads, bool group)
276 {
277 if (cpus == NULL) {
278 /* Work around old compiler warnings about strict aliasing */
279 cpus = &empty_cpu_map.map;
280 }
281
282 if (threads == NULL)
283 threads = &empty_thread_map.map;
284
285 return __perf_evsel__open(evsel, cpus, threads, group);
286 }
287
perf_evsel__open_per_cpu(struct perf_evsel * evsel,struct cpu_map * cpus,bool group)288 int perf_evsel__open_per_cpu(struct perf_evsel *evsel,
289 struct cpu_map *cpus, bool group)
290 {
291 return __perf_evsel__open(evsel, cpus, &empty_thread_map.map, group);
292 }
293
perf_evsel__open_per_thread(struct perf_evsel * evsel,struct thread_map * threads,bool group)294 int perf_evsel__open_per_thread(struct perf_evsel *evsel,
295 struct thread_map *threads, bool group)
296 {
297 return __perf_evsel__open(evsel, &empty_cpu_map.map, threads, group);
298 }
299
perf_event__parse_id_sample(const union perf_event * event,u64 type,struct perf_sample * sample)300 static int perf_event__parse_id_sample(const union perf_event *event, u64 type,
301 struct perf_sample *sample)
302 {
303 const u64 *array = event->sample.array;
304
305 array += ((event->header.size -
306 sizeof(event->header)) / sizeof(u64)) - 1;
307
308 if (type & PERF_SAMPLE_CPU) {
309 u32 *p = (u32 *)array;
310 sample->cpu = *p;
311 array--;
312 }
313
314 if (type & PERF_SAMPLE_STREAM_ID) {
315 sample->stream_id = *array;
316 array--;
317 }
318
319 if (type & PERF_SAMPLE_ID) {
320 sample->id = *array;
321 array--;
322 }
323
324 if (type & PERF_SAMPLE_TIME) {
325 sample->time = *array;
326 array--;
327 }
328
329 if (type & PERF_SAMPLE_TID) {
330 u32 *p = (u32 *)array;
331 sample->pid = p[0];
332 sample->tid = p[1];
333 }
334
335 return 0;
336 }
337
sample_overlap(const union perf_event * event,const void * offset,u64 size)338 static bool sample_overlap(const union perf_event *event,
339 const void *offset, u64 size)
340 {
341 const void *base = event;
342
343 if (offset + size > base + event->header.size)
344 return true;
345
346 return false;
347 }
348
perf_event__parse_sample(const union perf_event * event,u64 type,int sample_size,bool sample_id_all,struct perf_sample * data)349 int perf_event__parse_sample(const union perf_event *event, u64 type,
350 int sample_size, bool sample_id_all,
351 struct perf_sample *data)
352 {
353 const u64 *array;
354
355 data->cpu = data->pid = data->tid = -1;
356 data->stream_id = data->id = data->time = -1ULL;
357
358 if (event->header.type != PERF_RECORD_SAMPLE) {
359 if (!sample_id_all)
360 return 0;
361 return perf_event__parse_id_sample(event, type, data);
362 }
363
364 array = event->sample.array;
365
366 if (sample_size + sizeof(event->header) > event->header.size)
367 return -EFAULT;
368
369 if (type & PERF_SAMPLE_IP) {
370 data->ip = event->ip.ip;
371 array++;
372 }
373
374 if (type & PERF_SAMPLE_TID) {
375 u32 *p = (u32 *)array;
376 data->pid = p[0];
377 data->tid = p[1];
378 array++;
379 }
380
381 if (type & PERF_SAMPLE_TIME) {
382 data->time = *array;
383 array++;
384 }
385
386 if (type & PERF_SAMPLE_ADDR) {
387 data->addr = *array;
388 array++;
389 }
390
391 data->id = -1ULL;
392 if (type & PERF_SAMPLE_ID) {
393 data->id = *array;
394 array++;
395 }
396
397 if (type & PERF_SAMPLE_STREAM_ID) {
398 data->stream_id = *array;
399 array++;
400 }
401
402 if (type & PERF_SAMPLE_CPU) {
403 u32 *p = (u32 *)array;
404 data->cpu = *p;
405 array++;
406 }
407
408 if (type & PERF_SAMPLE_PERIOD) {
409 data->period = *array;
410 array++;
411 }
412
413 if (type & PERF_SAMPLE_READ) {
414 fprintf(stderr, "PERF_SAMPLE_READ is unsuported for now\n");
415 return -1;
416 }
417
418 if (type & PERF_SAMPLE_CALLCHAIN) {
419 if (sample_overlap(event, array, sizeof(data->callchain->nr)))
420 return -EFAULT;
421
422 data->callchain = (struct ip_callchain *)array;
423
424 if (sample_overlap(event, array, data->callchain->nr))
425 return -EFAULT;
426
427 array += 1 + data->callchain->nr;
428 }
429
430 if (type & PERF_SAMPLE_RAW) {
431 u32 *p = (u32 *)array;
432
433 if (sample_overlap(event, array, sizeof(u32)))
434 return -EFAULT;
435
436 data->raw_size = *p;
437 p++;
438
439 if (sample_overlap(event, p, data->raw_size))
440 return -EFAULT;
441
442 data->raw_data = p;
443 }
444
445 return 0;
446 }
447