tools/perf/util/evsel.c

   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
  18 void perf_evsel__init(struct perf_evsel *evsel,
  19                       struct perf_event_attr *attr, int idx)
  20 {
  21         evsel->idx         = idx;
  22         evsel->attr        = *attr;
  23         INIT_LIST_HEAD(&evsel->node);
  24 }
  25
  26 struct perf_evsel *perf_evsel__new(struct perf_event_attr *attr, int idx)
  27 {
  28         struct perf_evsel *evsel = zalloc(sizeof(*evsel));
  29
  30         if (evsel != NULL)
  31                 perf_evsel__init(evsel, attr, idx);
  32
  33         return evsel;
  34 }
  35
  36 int perf_evsel__alloc_fd(struct perf_evsel *evsel, int ncpus, int nthreads)
  37 {
  38         evsel->fd = xyarray__new(ncpus, nthreads, sizeof(int));
  39         return evsel->fd != NULL ? 0 : -ENOMEM;
  40 }
  41
  42 int perf_evsel__alloc_id(struct perf_evsel *evsel, int ncpus, int nthreads)
  43 {
  44         evsel->id = xyarray__new(ncpus, nthreads, sizeof(struct perf_sample_id));
  45         return evsel->id != NULL ? 0 : -ENOMEM;
  46 }
  47
  48 int perf_evsel__alloc_counts(struct perf_evsel *evsel, int ncpus)
  49 {
  50         evsel->counts = zalloc((sizeof(*evsel->counts) +
  51                                 (ncpus * sizeof(struct perf_counts_values))));
  52         return evsel->counts != NULL ? 0 : -ENOMEM;
  53 }
  54
  55 void perf_evsel__free_fd(struct perf_evsel *evsel)
  56 {
  57         xyarray__delete(evsel->fd);
  58         evsel->fd = NULL;
  59 }
  60
  61 void perf_evsel__free_id(struct perf_evsel *evsel)
  62 {
  63         xyarray__delete(evsel->id);
  64         evsel->id = NULL;
  65 }
  66
  67 void perf_evsel__close_fd(struct perf_evsel *evsel, int ncpus, int nthreads)
  68 {
  69         int cpu, thread;
  70
  71         for (cpu = 0; cpu < ncpus; cpu++)
  72                 for (thread = 0; thread < nthreads; ++thread) {
  73                         close(FD(evsel, cpu, thread));
  74                         FD(evsel, cpu, thread) = -1;
  75                 }
  76 }
  77
  78 void perf_evsel__exit(struct perf_evsel *evsel)
  79 {
  80         assert(list_empty(&evsel->node));
  81         xyarray__delete(evsel->fd);
  82         xyarray__delete(evsel->id);
  83 }
  84
  85 void perf_evsel__delete(struct perf_evsel *evsel)
  86 {
  87         perf_evsel__exit(evsel);
  88         free(evsel);
  89 }
  90
  91 int __perf_evsel__read_on_cpu(struct perf_evsel *evsel,
  92                               int cpu, int thread, bool scale)
  93 {
  94         struct perf_counts_values count;
  95         size_t nv = scale ? 3 : 1;
  96
  97         if (FD(evsel, cpu, thread) < 0)
  98                 return -EINVAL;
  99
 100         if (evsel->counts == NULL && perf_evsel__alloc_counts(evsel, cpu + 1) < 0)
 101                 return -ENOMEM;
 102
 103         if (readn(FD(evsel, cpu, thread), &count, nv * sizeof(u64)) < 0)
 104                 return -errno;
 105
 106         if (scale) {
 107                 if (count.run == 0)
 108                         count.val = 0;
 109                 else if (count.run < count.ena)
 110                         count.val = (u64)((double)count.val * count.ena / count.run + 0.5);
 111         } else
 112                 count.ena = count.run = 0;
 113
 114         evsel->counts->cpu[cpu] = count;
 115         return 0;
 116 }
 117
 118 int __perf_evsel__read(struct perf_evsel *evsel,
 119                        int ncpus, int nthreads, bool scale)
 120 {
 121         size_t nv = scale ? 3 : 1;
 122         int cpu, thread;
 123         struct perf_counts_values *aggr = &evsel->counts->aggr, count;
 124
 125         aggr->val = aggr->ena = aggr->run = 0;
 126
 127         for (cpu = 0; cpu < ncpus; cpu++) {
 128                 for (thread = 0; thread < nthreads; thread++) {
 129                         if (FD(evsel, cpu, thread) < 0)
 130                                 continue;
 131
 132                         if (readn(FD(evsel, cpu, thread),
 133                                   &count, nv * sizeof(u64)) < 0)
 134                                 return -errno;
 135
 136                         aggr->val += count.val;
 137                         if (scale) {
 138                                 aggr->ena += count.ena;
 139                                 aggr->run += count.run;
 140                         }
 141                 }
 142         }
 143
 144         evsel->counts->scaled = 0;
 145         if (scale) {
 146                 if (aggr->run == 0) {
 147                         evsel->counts->scaled = -1;
 148                         aggr->val = 0;
 149                         return 0;
 150                 }
 151
 152                 if (aggr->run < aggr->ena) {
 153                         evsel->counts->scaled = 1;
 154                         aggr->val = (u64)((double)aggr->val * aggr->ena / aggr->run + 0.5);
 155                 }
 156         } else
 157                 aggr->ena = aggr->run = 0;
 158
 159         return 0;
 160 }
 161
 162 static int __perf_evsel__open(struct perf_evsel *evsel, struct cpu_map *cpus,
 163                               struct thread_map *threads, bool group, bool inherit)
 164 {
 165         int cpu, thread;
 166
 167         if (evsel->fd == NULL &&
 168             perf_evsel__alloc_fd(evsel, cpus->nr, threads->nr) < 0)
 169                 return -1;
 170
 171         for (cpu = 0; cpu < cpus->nr; cpu++) {
 172                 int group_fd = -1;
 173
 174                 evsel->attr.inherit = (cpus->map[cpu] < 0) && inherit;
 175
 176                 for (thread = 0; thread < threads->nr; thread++) {
 177                         FD(evsel, cpu, thread) = sys_perf_event_open(&evsel->attr,
 178                                                                      threads->map[thread],
 179                                                                      cpus->map[cpu],
 180                                                                      group_fd, 0);
 181                         if (FD(evsel, cpu, thread) < 0)
 182                                 goto out_close;
 183
 184                         if (group && group_fd == -1)
 185                                 group_fd = FD(evsel, cpu, thread);
 186                 }
 187         }
 188
 189         return 0;
 190
 191 out_close:
 192         do {
 193                 while (--thread >= 0) {
 194                         close(FD(evsel, cpu, thread));
 195                         FD(evsel, cpu, thread) = -1;
 196                 }
 197                 thread = threads->nr;
 198         } while (--cpu >= 0);
 199         return -1;
 200 }
 201
 202 static struct {
 203         struct cpu_map map;
 204         int cpus[1];
 205 } empty_cpu_map = {
 206         .map.nr = 1,
 207         .cpus   = { -1, },
 208 };
 209
 210 static struct {
 211         struct thread_map map;
 212         int threads[1];
 213 } empty_thread_map = {
 214         .map.nr  = 1,
 215         .threads = { -1, },
 216 };
 217
 218 int perf_evsel__open(struct perf_evsel *evsel, struct cpu_map *cpus,
 219                      struct thread_map *threads, bool group, bool inherit)
 220 {
 221         if (cpus == NULL) {
 222                 /* Work around old compiler warnings about strict aliasing */
 223                 cpus = &empty_cpu_map.map;
 224         }
 225
 226         if (threads == NULL)
 227                 threads = &empty_thread_map.map;
 228
 229         return __perf_evsel__open(evsel, cpus, threads, group, inherit);
 230 }
 231
 232 int perf_evsel__open_per_cpu(struct perf_evsel *evsel,
 233                              struct cpu_map *cpus, bool group, bool inherit)
 234 {
 235         return __perf_evsel__open(evsel, cpus, &empty_thread_map.map, group, inherit);
 236 }
 237
 238 int perf_evsel__open_per_thread(struct perf_evsel *evsel,
 239                                 struct thread_map *threads, bool group, bool inherit)
 240 {
 241         return __perf_evsel__open(evsel, &empty_cpu_map.map, threads, group, inherit);
 242 }
 243
 244 static int perf_event__parse_id_sample(const union perf_event *event, u64 type,
 245                                        struct perf_sample *sample)
 246 {
 247         const u64 *array = event->sample.array;
 248
 249         array += ((event->header.size -
 250                    sizeof(event->header)) / sizeof(u64)) - 1;
 251
 252         if (type & PERF_SAMPLE_CPU) {
 253                 u32 *p = (u32 *)array;
 254                 sample->cpu = *p;
 255                 array--;
 256         }
 257
 258         if (type & PERF_SAMPLE_STREAM_ID) {
 259                 sample->stream_id = *array;
 260                 array--;
 261         }
 262
 263         if (type & PERF_SAMPLE_ID) {
 264                 sample->id = *array;
 265                 array--;
 266         }
 267
 268         if (type & PERF_SAMPLE_TIME) {
 269                 sample->time = *array;
 270                 array--;
 271         }
 272
 273         if (type & PERF_SAMPLE_TID) {
 274                 u32 *p = (u32 *)array;
 275                 sample->pid = p[0];
 276                 sample->tid = p[1];
 277         }
 278
 279         return 0;
 280 }
 281
 282 int perf_event__parse_sample(const union perf_event *event, u64 type,
 283                              bool sample_id_all, struct perf_sample *data)
 284 {
 285         const u64 *array;
 286
 287         data->cpu = data->pid = data->tid = -1;
 288         data->stream_id = data->id = data->time = -1ULL;
 289
 290         if (event->header.type != PERF_RECORD_SAMPLE) {
 291                 if (!sample_id_all)
 292                         return 0;
 293                 return perf_event__parse_id_sample(event, type, data);
 294         }
 295
 296         array = event->sample.array;
 297
 298         if (type & PERF_SAMPLE_IP) {
 299                 data->ip = event->ip.ip;
 300                 array++;
 301         }
 302
 303         if (type & PERF_SAMPLE_TID) {
 304                 u32 *p = (u32 *)array;
 305                 data->pid = p[0];
 306                 data->tid = p[1];
 307                 array++;
 308         }
 309
 310         if (type & PERF_SAMPLE_TIME) {
 311                 data->time = *array;
 312                 array++;
 313         }
 314
 315         if (type & PERF_SAMPLE_ADDR) {
 316                 data->addr = *array;
 317                 array++;
 318         }
 319
 320         data->id = -1ULL;
 321         if (type & PERF_SAMPLE_ID) {
 322                 data->id = *array;
 323                 array++;
 324         }
 325
 326         if (type & PERF_SAMPLE_STREAM_ID) {
 327                 data->stream_id = *array;
 328                 array++;
 329         }
 330
 331         if (type & PERF_SAMPLE_CPU) {
 332                 u32 *p = (u32 *)array;
 333                 data->cpu = *p;
 334                 array++;
 335         }
 336
 337         if (type & PERF_SAMPLE_PERIOD) {
 338                 data->period = *array;
 339                 array++;
 340         }
 341
 342         if (type & PERF_SAMPLE_READ) {
 343                 fprintf(stderr, "PERF_SAMPLE_READ is unsuported for now\n");
 344                 return -1;
 345         }
 346
 347         if (type & PERF_SAMPLE_CALLCHAIN) {
 348                 data->callchain = (struct ip_callchain *)array;
 349                 array += 1 + data->callchain->nr;
 350         }
 351
 352         if (type & PERF_SAMPLE_RAW) {
 353                 u32 *p = (u32 *)array;
 354                 data->raw_size = *p;
 355                 p++;
 356                 data->raw_data = p;
 357         }
 358
 359         return 0;
 360 }