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         close_cgroup(evsel->cgrp);
  89         free(evsel->name);
  90         free(evsel);
  91 }
  92
  93 int __perf_evsel__read_on_cpu(struct perf_evsel *evsel,
  94                               int cpu, int thread, bool scale)
  95 {
  96         struct perf_counts_values count;
  97         size_t nv = scale ? 3 : 1;
  98
  99         if (FD(evsel, cpu, thread) < 0)
 100                 return -EINVAL;
 101
 102         if (evsel->counts == NULL && perf_evsel__alloc_counts(evsel, cpu + 1) < 0)
 103                 return -ENOMEM;
 104
 105         if (readn(FD(evsel, cpu, thread), &count, nv * sizeof(u64)) < 0)
 106                 return -errno;
 107
 108         if (scale) {
 109                 if (count.run == 0)
 110                         count.val = 0;
 111                 else if (count.run < count.ena)
 112                         count.val = (u64)((double)count.val * count.ena / count.run + 0.5);
 113         } else
 114                 count.ena = count.run = 0;
 115
 116         evsel->counts->cpu[cpu] = count;
 117         return 0;
 118 }
 119
 120 int __perf_evsel__read(struct perf_evsel *evsel,
 121                        int ncpus, int nthreads, bool scale)
 122 {
 123         size_t nv = scale ? 3 : 1;
 124         int cpu, thread;
 125         struct perf_counts_values *aggr = &evsel->counts->aggr, count;
 126
 127         aggr->val = aggr->ena = aggr->run = 0;
 128
 129         for (cpu = 0; cpu < ncpus; cpu++) {
 130                 for (thread = 0; thread < nthreads; thread++) {
 131                         if (FD(evsel, cpu, thread) < 0)
 132                                 continue;
 133
 134                         if (readn(FD(evsel, cpu, thread),
 135                                   &count, nv * sizeof(u64)) < 0)
 136                                 return -errno;
 137
 138                         aggr->val += count.val;
 139                         if (scale) {
 140                                 aggr->ena += count.ena;
 141                                 aggr->run += count.run;
 142                         }
 143                 }
 144         }
 145
 146         evsel->counts->scaled = 0;
 147         if (scale) {
 148                 if (aggr->run == 0) {
 149                         evsel->counts->scaled = -1;
 150                         aggr->val = 0;
 151                         return 0;
 152                 }
 153
 154                 if (aggr->run < aggr->ena) {
 155                         evsel->counts->scaled = 1;
 156                         aggr->val = (u64)((double)aggr->val * aggr->ena / aggr->run + 0.5);
 157                 }
 158         } else
 159                 aggr->ena = aggr->run = 0;
 160
 161         return 0;
 162 }
 163
 164 static int __perf_evsel__open(struct perf_evsel *evsel, struct cpu_map *cpus,
 165                               struct thread_map *threads, bool group, bool inherit)
 166 {
 167         int cpu, thread;
 168         unsigned long flags = 0;
 169         int pid = -1;
 170
 171         if (evsel->fd == NULL &&
 172             perf_evsel__alloc_fd(evsel, cpus->nr, threads->nr) < 0)
 173                 return -1;
 174
 175         if (evsel->cgrp) {
 176                 flags = PERF_FLAG_PID_CGROUP;
 177                 pid = evsel->cgrp->fd;
 178         }
 179
 180         for (cpu = 0; cpu < cpus->nr; cpu++) {
 181                 int group_fd = -1;
 182                 /*
 183                  * Don't allow mmap() of inherited per-task counters. This
 184                  * would create a performance issue due to all children writing
 185                  * to the same buffer.
 186                  *
 187                  * FIXME:
 188                  * Proper fix is not to pass 'inherit' to perf_evsel__open*,
 189                  * but a 'flags' parameter, with 'group' folded there as well,
 190                  * then introduce a PERF_O_{MMAP,GROUP,INHERIT} enum, and if
 191                  * O_MMAP is set, emit a warning if cpu < 0 and O_INHERIT is
 192                  * set. Lets go for the minimal fix first tho.
 193                  */
 194                 evsel->attr.inherit = (cpus->map[cpu] >= 0) && inherit;
 195
 196                 for (thread = 0; thread < threads->nr; thread++) {
 197
 198                         if (!evsel->cgrp)
 199                                 pid = threads->map[thread];
 200
 201                         FD(evsel, cpu, thread) = sys_perf_event_open(&evsel->attr,
 202                                                                      pid,
 203                                                                      cpus->map[cpu],
 204                                                                      group_fd, flags);
 205                         if (FD(evsel, cpu, thread) < 0)
 206                                 goto out_close;
 207
 208                         if (group && group_fd == -1)
 209                                 group_fd = FD(evsel, cpu, thread);
 210                 }
 211         }
 212
 213         return 0;
 214
 215 out_close:
 216         do {
 217                 while (--thread >= 0) {
 218                         close(FD(evsel, cpu, thread));
 219                         FD(evsel, cpu, thread) = -1;
 220                 }
 221                 thread = threads->nr;
 222         } while (--cpu >= 0);
 223         return -1;
 224 }
 225
 226 static struct {
 227         struct cpu_map map;
 228         int cpus[1];
 229 } empty_cpu_map = {
 230         .map.nr = 1,
 231         .cpus   = { -1, },
 232 };
 233
 234 static struct {
 235         struct thread_map map;
 236         int threads[1];
 237 } empty_thread_map = {
 238         .map.nr  = 1,
 239         .threads = { -1, },
 240 };
 241
 242 int perf_evsel__open(struct perf_evsel *evsel, struct cpu_map *cpus,
 243                      struct thread_map *threads, bool group, bool inherit)
 244 {
 245         if (cpus == NULL) {
 246                 /* Work around old compiler warnings about strict aliasing */
 247                 cpus = &empty_cpu_map.map;
 248         }
 249
 250         if (threads == NULL)
 251                 threads = &empty_thread_map.map;
 252
 253         return __perf_evsel__open(evsel, cpus, threads, group, inherit);
 254 }
 255
 256 int perf_evsel__open_per_cpu(struct perf_evsel *evsel,
 257                              struct cpu_map *cpus, bool group, bool inherit)
 258 {
 259         return __perf_evsel__open(evsel, cpus, &empty_thread_map.map, group, inherit);
 260 }
 261
 262 int perf_evsel__open_per_thread(struct perf_evsel *evsel,
 263                                 struct thread_map *threads, bool group, bool inherit)
 264 {
 265         return __perf_evsel__open(evsel, &empty_cpu_map.map, threads, group, inherit);
 266 }
 267
 268 static int perf_event__parse_id_sample(const union perf_event *event, u64 type,
 269                                        struct perf_sample *sample)
 270 {
 271         const u64 *array = event->sample.array;
 272
 273         array += ((event->header.size -
 274                    sizeof(event->header)) / sizeof(u64)) - 1;
 275
 276         if (type & PERF_SAMPLE_CPU) {
 277                 u32 *p = (u32 *)array;
 278                 sample->cpu = *p;
 279                 array--;
 280         }
 281
 282         if (type & PERF_SAMPLE_STREAM_ID) {
 283                 sample->stream_id = *array;
 284                 array--;
 285         }
 286
 287         if (type & PERF_SAMPLE_ID) {
 288                 sample->id = *array;
 289                 array--;
 290         }
 291
 292         if (type & PERF_SAMPLE_TIME) {
 293                 sample->time = *array;
 294                 array--;
 295         }
 296
 297         if (type & PERF_SAMPLE_TID) {
 298                 u32 *p = (u32 *)array;
 299                 sample->pid = p[0];
 300                 sample->tid = p[1];
 301         }
 302
 303         return 0;
 304 }
 305
 306 int perf_event__parse_sample(const union perf_event *event, u64 type,
 307                              bool sample_id_all, struct perf_sample *data)
 308 {
 309         const u64 *array;
 310
 311         data->cpu = data->pid = data->tid = -1;
 312         data->stream_id = data->id = data->time = -1ULL;
 313
 314         if (event->header.type != PERF_RECORD_SAMPLE) {
 315                 if (!sample_id_all)
 316                         return 0;
 317                 return perf_event__parse_id_sample(event, type, data);
 318         }
 319
 320         array = event->sample.array;
 321
 322         if (type & PERF_SAMPLE_IP) {
 323                 data->ip = event->ip.ip;
 324                 array++;
 325         }
 326
 327         if (type & PERF_SAMPLE_TID) {
 328                 u32 *p = (u32 *)array;
 329                 data->pid = p[0];
 330                 data->tid = p[1];
 331                 array++;
 332         }
 333
 334         if (type & PERF_SAMPLE_TIME) {
 335                 data->time = *array;
 336                 array++;
 337         }
 338
 339         if (type & PERF_SAMPLE_ADDR) {
 340                 data->addr = *array;
 341                 array++;
 342         }
 343
 344         data->id = -1ULL;
 345         if (type & PERF_SAMPLE_ID) {
 346                 data->id = *array;
 347                 array++;
 348         }
 349
 350         if (type & PERF_SAMPLE_STREAM_ID) {
 351                 data->stream_id = *array;
 352                 array++;
 353         }
 354
 355         if (type & PERF_SAMPLE_CPU) {
 356                 u32 *p = (u32 *)array;
 357                 data->cpu = *p;
 358                 array++;
 359         }
 360
 361         if (type & PERF_SAMPLE_PERIOD) {
 362                 data->period = *array;
 363                 array++;
 364         }
 365
 366         if (type & PERF_SAMPLE_READ) {
 367                 fprintf(stderr, "PERF_SAMPLE_READ is unsuported for now\n");
 368                 return -1;
 369         }
 370
 371         if (type & PERF_SAMPLE_CALLCHAIN) {
 372                 data->callchain = (struct ip_callchain *)array;
 373                 array += 1 + data->callchain->nr;
 374         }
 375
 376         if (type & PERF_SAMPLE_RAW) {
 377                 u32 *p = (u32 *)array;
 378                 data->raw_size = *p;
 379                 p++;
 380                 data->raw_data = p;
 381         }
 382
 383         return 0;
 384 }