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 <byteswap.h>
  11 #include <errno.h>
  12 #include <inttypes.h>
  13 #include <linux/bitops.h>
  14 #include <api/fs/fs.h>
  15 #include <api/fs/tracing_path.h>
  16 #include <traceevent/event-parse.h>
  17 #include <linux/hw_breakpoint.h>
  18 #include <linux/perf_event.h>
  19 #include <linux/compiler.h>
  20 #include <linux/err.h>
  21 #include <sys/ioctl.h>
  22 #include <sys/resource.h>
  23 #include <sys/types.h>
  24 #include <dirent.h>
  25 #include "asm/bug.h"
  26 #include "callchain.h"
  27 #include "cgroup.h"
  28 #include "event.h"
  29 #include "evsel.h"
  30 #include "evlist.h"
  31 #include "util.h"
  32 #include "cpumap.h"
  33 #include "thread_map.h"
  34 #include "target.h"
  35 #include "perf_regs.h"
  36 #include "debug.h"
  37 #include "trace-event.h"
  38 #include "stat.h"
  39 #include "util/parse-branch-options.h"
  40
  41 #include "sane_ctype.h"
  42
  43 static struct {
  44         bool sample_id_all;
  45         bool exclude_guest;
  46         bool mmap2;
  47         bool cloexec;
  48         bool clockid;
  49         bool clockid_wrong;
  50         bool lbr_flags;
  51         bool write_backward;
  52         bool group_read;
  53 } perf_missing_features;
  54
  55 static clockid_t clockid;
  56
  57 static int perf_evsel__no_extra_init(struct perf_evsel *evsel __maybe_unused)
  58 {
  59         return 0;
  60 }
  61
  62 void __weak test_attr__ready(void) { }
  63
  64 static void perf_evsel__no_extra_fini(struct perf_evsel *evsel __maybe_unused)
  65 {
  66 }
  67
  68 static struct {
  69         size_t  size;
  70         int     (*init)(struct perf_evsel *evsel);
  71         void    (*fini)(struct perf_evsel *evsel);
  72 } perf_evsel__object = {
  73         .size = sizeof(struct perf_evsel),
  74         .init = perf_evsel__no_extra_init,
  75         .fini = perf_evsel__no_extra_fini,
  76 };
  77
  78 int perf_evsel__object_config(size_t object_size,
  79                               int (*init)(struct perf_evsel *evsel),
  80                               void (*fini)(struct perf_evsel *evsel))
  81 {
  82
  83         if (object_size == 0)
  84                 goto set_methods;
  85
  86         if (perf_evsel__object.size > object_size)
  87                 return -EINVAL;
  88
  89         perf_evsel__object.size = object_size;
  90
  91 set_methods:
  92         if (init != NULL)
  93                 perf_evsel__object.init = init;
  94
  95         if (fini != NULL)
  96                 perf_evsel__object.fini = fini;
  97
  98         return 0;
  99 }
 100
 101 #define FD(e, x, y) (*(int *)xyarray__entry(e->fd, x, y))
 102
 103 int __perf_evsel__sample_size(u64 sample_type)
 104 {
 105         u64 mask = sample_type & PERF_SAMPLE_MASK;
 106         int size = 0;
 107         int i;
 108
 109         for (i = 0; i < 64; i++) {
 110                 if (mask & (1ULL << i))
 111                         size++;
 112         }
 113
 114         size *= sizeof(u64);
 115
 116         return size;
 117 }
 118
 119 /**
 120  * __perf_evsel__calc_id_pos - calculate id_pos.
 121  * @sample_type: sample type
 122  *
 123  * This function returns the position of the event id (PERF_SAMPLE_ID or
 124  * PERF_SAMPLE_IDENTIFIER) in a sample event i.e. in the array of struct
 125  * sample_event.
 126  */
 127 static int __perf_evsel__calc_id_pos(u64 sample_type)
 128 {
 129         int idx = 0;
 130
 131         if (sample_type & PERF_SAMPLE_IDENTIFIER)
 132                 return 0;
 133
 134         if (!(sample_type & PERF_SAMPLE_ID))
 135                 return -1;
 136
 137         if (sample_type & PERF_SAMPLE_IP)
 138                 idx += 1;
 139
 140         if (sample_type & PERF_SAMPLE_TID)
 141                 idx += 1;
 142
 143         if (sample_type & PERF_SAMPLE_TIME)
 144                 idx += 1;
 145
 146         if (sample_type & PERF_SAMPLE_ADDR)
 147                 idx += 1;
 148
 149         return idx;
 150 }
 151
 152 /**
 153  * __perf_evsel__calc_is_pos - calculate is_pos.
 154  * @sample_type: sample type
 155  *
 156  * This function returns the position (counting backwards) of the event id
 157  * (PERF_SAMPLE_ID or PERF_SAMPLE_IDENTIFIER) in a non-sample event i.e. if
 158  * sample_id_all is used there is an id sample appended to non-sample events.
 159  */
 160 static int __perf_evsel__calc_is_pos(u64 sample_type)
 161 {
 162         int idx = 1;
 163
 164         if (sample_type & PERF_SAMPLE_IDENTIFIER)
 165                 return 1;
 166
 167         if (!(sample_type & PERF_SAMPLE_ID))
 168                 return -1;
 169
 170         if (sample_type & PERF_SAMPLE_CPU)
 171                 idx += 1;
 172
 173         if (sample_type & PERF_SAMPLE_STREAM_ID)
 174                 idx += 1;
 175
 176         return idx;
 177 }
 178
 179 void perf_evsel__calc_id_pos(struct perf_evsel *evsel)
 180 {
 181         evsel->id_pos = __perf_evsel__calc_id_pos(evsel->attr.sample_type);
 182         evsel->is_pos = __perf_evsel__calc_is_pos(evsel->attr.sample_type);
 183 }
 184
 185 void __perf_evsel__set_sample_bit(struct perf_evsel *evsel,
 186                                   enum perf_event_sample_format bit)
 187 {
 188         if (!(evsel->attr.sample_type & bit)) {
 189                 evsel->attr.sample_type |= bit;
 190                 evsel->sample_size += sizeof(u64);
 191                 perf_evsel__calc_id_pos(evsel);
 192         }
 193 }
 194
 195 void __perf_evsel__reset_sample_bit(struct perf_evsel *evsel,
 196                                     enum perf_event_sample_format bit)
 197 {
 198         if (evsel->attr.sample_type & bit) {
 199                 evsel->attr.sample_type &= ~bit;
 200                 evsel->sample_size -= sizeof(u64);
 201                 perf_evsel__calc_id_pos(evsel);
 202         }
 203 }
 204
 205 void perf_evsel__set_sample_id(struct perf_evsel *evsel,
 206                                bool can_sample_identifier)
 207 {
 208         if (can_sample_identifier) {
 209                 perf_evsel__reset_sample_bit(evsel, ID);
 210                 perf_evsel__set_sample_bit(evsel, IDENTIFIER);
 211         } else {
 212                 perf_evsel__set_sample_bit(evsel, ID);
 213         }
 214         evsel->attr.read_format |= PERF_FORMAT_ID;
 215 }
 216
 217 /**
 218  * perf_evsel__is_function_event - Return whether given evsel is a function
 219  * trace event
 220  *
 221  * @evsel - evsel selector to be tested
 222  *
 223  * Return %true if event is function trace event
 224  */
 225 bool perf_evsel__is_function_event(struct perf_evsel *evsel)
 226 {
 227 #define FUNCTION_EVENT "ftrace:function"
 228
 229         return evsel->name &&
 230                !strncmp(FUNCTION_EVENT, evsel->name, sizeof(FUNCTION_EVENT));
 231
 232 #undef FUNCTION_EVENT
 233 }
 234
 235 void perf_evsel__init(struct perf_evsel *evsel,
 236                       struct perf_event_attr *attr, int idx)
 237 {
 238         evsel->idx         = idx;
 239         evsel->tracking    = !idx;
 240         evsel->attr        = *attr;
 241         evsel->leader      = evsel;
 242         evsel->unit        = "";
 243         evsel->scale       = 1.0;
 244         evsel->evlist      = NULL;
 245         evsel->bpf_fd      = -1;
 246         INIT_LIST_HEAD(&evsel->node);
 247         INIT_LIST_HEAD(&evsel->config_terms);
 248         perf_evsel__object.init(evsel);
 249         evsel->sample_size = __perf_evsel__sample_size(attr->sample_type);
 250         perf_evsel__calc_id_pos(evsel);
 251         evsel->cmdline_group_boundary = false;
 252         evsel->metric_expr   = NULL;
 253         evsel->metric_name   = NULL;
 254         evsel->metric_events = NULL;
 255         evsel->collect_stat  = false;
 256 }
 257
 258 struct perf_evsel *perf_evsel__new_idx(struct perf_event_attr *attr, int idx)
 259 {
 260         struct perf_evsel *evsel = zalloc(perf_evsel__object.size);
 261
 262         if (evsel != NULL)
 263                 perf_evsel__init(evsel, attr, idx);
 264
 265         if (perf_evsel__is_bpf_output(evsel)) {
 266                 evsel->attr.sample_type |= (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
 267                                             PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
 268                 evsel->attr.sample_period = 1;
 269         }
 270
 271         return evsel;
 272 }
 273
 274 static bool perf_event_can_profile_kernel(void)
 275 {
 276         return geteuid() == 0 || perf_event_paranoid() == -1;
 277 }
 278
 279 struct perf_evsel *perf_evsel__new_cycles(bool precise)
 280 {
 281         struct perf_event_attr attr = {
 282                 .type   = PERF_TYPE_HARDWARE,
 283                 .config = PERF_COUNT_HW_CPU_CYCLES,
 284                 .exclude_kernel = !perf_event_can_profile_kernel(),
 285         };
 286         struct perf_evsel *evsel;
 287
 288         event_attr_init(&attr);
 289
 290         if (!precise)
 291                 goto new_event;
 292         /*
 293          * Unnamed union member, not supported as struct member named
 294          * initializer in older compilers such as gcc 4.4.7
 295          *
 296          * Just for probing the precise_ip:
 297          */
 298         attr.sample_period = 1;
 299
 300         perf_event_attr__set_max_precise_ip(&attr);
 301         /*
 302          * Now let the usual logic to set up the perf_event_attr defaults
 303          * to kick in when we return and before perf_evsel__open() is called.
 304          */
 305         attr.sample_period = 0;
 306 new_event:
 307         evsel = perf_evsel__new(&attr);
 308         if (evsel == NULL)
 309                 goto out;
 310
 311         /* use asprintf() because free(evsel) assumes name is allocated */
 312         if (asprintf(&evsel->name, "cycles%s%s%.*s",
 313                      (attr.precise_ip || attr.exclude_kernel) ? ":" : "",
 314                      attr.exclude_kernel ? "u" : "",
 315                      attr.precise_ip ? attr.precise_ip + 1 : 0, "ppp") < 0)
 316                 goto error_free;
 317 out:
 318         return evsel;
 319 error_free:
 320         perf_evsel__delete(evsel);
 321         evsel = NULL;
 322         goto out;
 323 }
 324
 325 /*
 326  * Returns pointer with encoded error via <linux/err.h> interface.
 327  */
 328 struct perf_evsel *perf_evsel__newtp_idx(const char *sys, const char *name, int idx)
 329 {
 330         struct perf_evsel *evsel = zalloc(perf_evsel__object.size);
 331         int err = -ENOMEM;
 332
 333         if (evsel == NULL) {
 334                 goto out_err;
 335         } else {
 336                 struct perf_event_attr attr = {
 337                         .type          = PERF_TYPE_TRACEPOINT,
 338                         .sample_type   = (PERF_SAMPLE_RAW | PERF_SAMPLE_TIME |
 339                                           PERF_SAMPLE_CPU | PERF_SAMPLE_PERIOD),
 340                 };
 341
 342                 if (asprintf(&evsel->name, "%s:%s", sys, name) < 0)
 343                         goto out_free;
 344
 345                 evsel->tp_format = trace_event__tp_format(sys, name);
 346                 if (IS_ERR(evsel->tp_format)) {
 347                         err = PTR_ERR(evsel->tp_format);
 348                         goto out_free;
 349                 }
 350
 351                 event_attr_init(&attr);
 352                 attr.config = evsel->tp_format->id;
 353                 attr.sample_period = 1;
 354                 perf_evsel__init(evsel, &attr, idx);
 355         }
 356
 357         return evsel;
 358
 359 out_free:
 360         zfree(&evsel->name);
 361         free(evsel);
 362 out_err:
 363         return ERR_PTR(err);
 364 }
 365
 366 const char *perf_evsel__hw_names[PERF_COUNT_HW_MAX] = {
 367         "cycles",
 368         "instructions",
 369         "cache-references",
 370         "cache-misses",
 371         "branches",
 372         "branch-misses",
 373         "bus-cycles",
 374         "stalled-cycles-frontend",
 375         "stalled-cycles-backend",
 376         "ref-cycles",
 377 };
 378
 379 static const char *__perf_evsel__hw_name(u64 config)
 380 {
 381         if (config < PERF_COUNT_HW_MAX && perf_evsel__hw_names[config])
 382                 return perf_evsel__hw_names[config];
 383
 384         return "unknown-hardware";
 385 }
 386
 387 static int perf_evsel__add_modifiers(struct perf_evsel *evsel, char *bf, size_t size)
 388 {
 389         int colon = 0, r = 0;
 390         struct perf_event_attr *attr = &evsel->attr;
 391         bool exclude_guest_default = false;
 392
 393 #define MOD_PRINT(context, mod) do {                                    \
 394                 if (!attr->exclude_##context) {                         \
 395                         if (!colon) colon = ++r;                        \
 396                         r += scnprintf(bf + r, size - r, "%c", mod);    \
 397                 } } while(0)
 398
 399         if (attr->exclude_kernel || attr->exclude_user || attr->exclude_hv) {
 400                 MOD_PRINT(kernel, 'k');
 401                 MOD_PRINT(user, 'u');
 402                 MOD_PRINT(hv, 'h');
 403                 exclude_guest_default = true;
 404         }
 405
 406         if (attr->precise_ip) {
 407                 if (!colon)
 408                         colon = ++r;
 409                 r += scnprintf(bf + r, size - r, "%.*s", attr->precise_ip, "ppp");
 410                 exclude_guest_default = true;
 411         }
 412
 413         if (attr->exclude_host || attr->exclude_guest == exclude_guest_default) {
 414                 MOD_PRINT(host, 'H');
 415                 MOD_PRINT(guest, 'G');
 416         }
 417 #undef MOD_PRINT
 418         if (colon)
 419                 bf[colon - 1] = ':';
 420         return r;
 421 }
 422
 423 static int perf_evsel__hw_name(struct perf_evsel *evsel, char *bf, size_t size)
 424 {
 425         int r = scnprintf(bf, size, "%s", __perf_evsel__hw_name(evsel->attr.config));
 426         return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
 427 }
 428
 429 const char *perf_evsel__sw_names[PERF_COUNT_SW_MAX] = {
 430         "cpu-clock",
 431         "task-clock",
 432         "page-faults",
 433         "context-switches",
 434         "cpu-migrations",
 435         "minor-faults",
 436         "major-faults",
 437         "alignment-faults",
 438         "emulation-faults",
 439         "dummy",
 440 };
 441
 442 static const char *__perf_evsel__sw_name(u64 config)
 443 {
 444         if (config < PERF_COUNT_SW_MAX && perf_evsel__sw_names[config])
 445                 return perf_evsel__sw_names[config];
 446         return "unknown-software";
 447 }
 448
 449 static int perf_evsel__sw_name(struct perf_evsel *evsel, char *bf, size_t size)
 450 {
 451         int r = scnprintf(bf, size, "%s", __perf_evsel__sw_name(evsel->attr.config));
 452         return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
 453 }
 454
 455 static int __perf_evsel__bp_name(char *bf, size_t size, u64 addr, u64 type)
 456 {
 457         int r;
 458
 459         r = scnprintf(bf, size, "mem:0x%" PRIx64 ":", addr);
 460
 461         if (type & HW_BREAKPOINT_R)
 462                 r += scnprintf(bf + r, size - r, "r");
 463
 464         if (type & HW_BREAKPOINT_W)
 465                 r += scnprintf(bf + r, size - r, "w");
 466
 467         if (type & HW_BREAKPOINT_X)
 468                 r += scnprintf(bf + r, size - r, "x");
 469
 470         return r;
 471 }
 472
 473 static int perf_evsel__bp_name(struct perf_evsel *evsel, char *bf, size_t size)
 474 {
 475         struct perf_event_attr *attr = &evsel->attr;
 476         int r = __perf_evsel__bp_name(bf, size, attr->bp_addr, attr->bp_type);
 477         return r + perf_evsel__add_modifiers(evsel, bf + r, size - r);
 478 }
 479
 480 const char *perf_evsel__hw_cache[PERF_COUNT_HW_CACHE_MAX]
 481                                 [PERF_EVSEL__MAX_ALIASES] = {
 482  { "L1-dcache", "l1-d",         "l1d",          "L1-data",              },
 483  { "L1-icache", "l1-i",         "l1i",          "L1-instruction",       },
 484  { "LLC",       "L2",                                                   },
 485  { "dTLB",      "d-tlb",        "Data-TLB",                             },
 486  { "iTLB",      "i-tlb",        "Instruction-TLB",                      },
 487  { "branch",    "branches",     "bpu",          "btb",          "bpc",  },
 488  { "node",                                                              },
 489 };
 490
 491 const char *perf_evsel__hw_cache_op[PERF_COUNT_HW_CACHE_OP_MAX]
 492                                    [PERF_EVSEL__MAX_ALIASES] = {
 493  { "load",      "loads",        "read",                                 },
 494  { "store",     "stores",       "write",                                },
 495  { "prefetch",  "prefetches",   "speculative-read", "speculative-load", },
 496 };
 497
 498 const char *perf_evsel__hw_cache_result[PERF_COUNT_HW_CACHE_RESULT_MAX]
 499                                        [PERF_EVSEL__MAX_ALIASES] = {
 500  { "refs",      "Reference",    "ops",          "access",               },
 501  { "misses",    "miss",                                                 },
 502 };
 503
 504 #define C(x)            PERF_COUNT_HW_CACHE_##x
 505 #define CACHE_READ      (1 << C(OP_READ))
 506 #define CACHE_WRITE     (1 << C(OP_WRITE))
 507 #define CACHE_PREFETCH  (1 << C(OP_PREFETCH))
 508 #define COP(x)          (1 << x)
 509
 510 /*
 511  * cache operartion stat
 512  * L1I : Read and prefetch only
 513  * ITLB and BPU : Read-only
 514  */
 515 static unsigned long perf_evsel__hw_cache_stat[C(MAX)] = {
 516  [C(L1D)]       = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 517  [C(L1I)]       = (CACHE_READ | CACHE_PREFETCH),
 518  [C(LL)]        = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 519  [C(DTLB)]      = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 520  [C(ITLB)]      = (CACHE_READ),
 521  [C(BPU)]       = (CACHE_READ),
 522  [C(NODE)]      = (CACHE_READ | CACHE_WRITE | CACHE_PREFETCH),
 523 };
 524
 525 bool perf_evsel__is_cache_op_valid(u8 type, u8 op)
 526 {
 527         if (perf_evsel__hw_cache_stat[type] & COP(op))
 528                 return true;    /* valid */
 529         else
 530                 return false;   /* invalid */
 531 }
 532
 533 int __perf_evsel__hw_cache_type_op_res_name(u8 type, u8 op, u8 result,
 534                                             char *bf, size_t size)
 535 {
 536         if (result) {
 537                 return scnprintf(bf, size, "%s-%s-%s", perf_evsel__hw_cache[type][0],
 538                                  perf_evsel__hw_cache_op[op][0],
 539                                  perf_evsel__hw_cache_result[result][0]);
 540         }
 541
 542         return scnprintf(bf, size, "%s-%s", perf_evsel__hw_cache[type][0],
 543                          perf_evsel__hw_cache_op[op][1]);
 544 }
 545
 546 static int __perf_evsel__hw_cache_name(u64 config, char *bf, size_t size)
 547 {
 548         u8 op, result, type = (config >>  0) & 0xff;
 549         const char *err = "unknown-ext-hardware-cache-type";
 550
 551         if (type >= PERF_COUNT_HW_CACHE_MAX)
 552                 goto out_err;
 553
 554         op = (config >>  8) & 0xff;
 555         err = "unknown-ext-hardware-cache-op";
 556         if (op >= PERF_COUNT_HW_CACHE_OP_MAX)
 557                 goto out_err;
 558
 559         result = (config >> 16) & 0xff;
 560         err = "unknown-ext-hardware-cache-result";
 561         if (result >= PERF_COUNT_HW_CACHE_RESULT_MAX)
 562                 goto out_err;
 563
 564         err = "invalid-cache";
 565         if (!perf_evsel__is_cache_op_valid(type, op))
 566                 goto out_err;
 567
 568         return __perf_evsel__hw_cache_type_op_res_name(type, op, result, bf, size);
 569 out_err:
 570         return scnprintf(bf, size, "%s", err);
 571 }
 572
 573 static int perf_evsel__hw_cache_name(struct perf_evsel *evsel, char *bf, size_t size)
 574 {
 575         int ret = __perf_evsel__hw_cache_name(evsel->attr.config, bf, size);
 576         return ret + perf_evsel__add_modifiers(evsel, bf + ret, size - ret);
 577 }
 578
 579 static int perf_evsel__raw_name(struct perf_evsel *evsel, char *bf, size_t size)
 580 {
 581         int ret = scnprintf(bf, size, "raw 0x%" PRIx64, evsel->attr.config);
 582         return ret + perf_evsel__add_modifiers(evsel, bf + ret, size - ret);
 583 }
 584
 585 const char *perf_evsel__name(struct perf_evsel *evsel)
 586 {
 587         char bf[128];
 588
 589         if (evsel->name)
 590                 return evsel->name;
 591
 592         switch (evsel->attr.type) {
 593         case PERF_TYPE_RAW:
 594                 perf_evsel__raw_name(evsel, bf, sizeof(bf));
 595                 break;
 596
 597         case PERF_TYPE_HARDWARE:
 598                 perf_evsel__hw_name(evsel, bf, sizeof(bf));
 599                 break;
 600
 601         case PERF_TYPE_HW_CACHE:
 602                 perf_evsel__hw_cache_name(evsel, bf, sizeof(bf));
 603                 break;
 604
 605         case PERF_TYPE_SOFTWARE:
 606                 perf_evsel__sw_name(evsel, bf, sizeof(bf));
 607                 break;
 608
 609         case PERF_TYPE_TRACEPOINT:
 610                 scnprintf(bf, sizeof(bf), "%s", "unknown tracepoint");
 611                 break;
 612
 613         case PERF_TYPE_BREAKPOINT:
 614                 perf_evsel__bp_name(evsel, bf, sizeof(bf));
 615                 break;
 616
 617         default:
 618                 scnprintf(bf, sizeof(bf), "unknown attr type: %d",
 619                           evsel->attr.type);
 620                 break;
 621         }
 622
 623         evsel->name = strdup(bf);
 624
 625         return evsel->name ?: "unknown";
 626 }
 627
 628 const char *perf_evsel__group_name(struct perf_evsel *evsel)
 629 {
 630         return evsel->group_name ?: "anon group";
 631 }
 632
 633 int perf_evsel__group_desc(struct perf_evsel *evsel, char *buf, size_t size)
 634 {
 635         int ret;
 636         struct perf_evsel *pos;
 637         const char *group_name = perf_evsel__group_name(evsel);
 638
 639         ret = scnprintf(buf, size, "%s", group_name);
 640
 641         ret += scnprintf(buf + ret, size - ret, " { %s",
 642                          perf_evsel__name(evsel));
 643
 644         for_each_group_member(pos, evsel)
 645                 ret += scnprintf(buf + ret, size - ret, ", %s",
 646                                  perf_evsel__name(pos));
 647
 648         ret += scnprintf(buf + ret, size - ret, " }");
 649
 650         return ret;
 651 }
 652
 653 void perf_evsel__config_callchain(struct perf_evsel *evsel,
 654                                   struct record_opts *opts,
 655                                   struct callchain_param *param)
 656 {
 657         bool function = perf_evsel__is_function_event(evsel);
 658         struct perf_event_attr *attr = &evsel->attr;
 659
 660         perf_evsel__set_sample_bit(evsel, CALLCHAIN);
 661
 662         attr->sample_max_stack = param->max_stack;
 663
 664         if (param->record_mode == CALLCHAIN_LBR) {
 665                 if (!opts->branch_stack) {
 666                         if (attr->exclude_user) {
 667                                 pr_warning("LBR callstack option is only available "
 668                                            "to get user callchain information. "
 669                                            "Falling back to framepointers.\n");
 670                         } else {
 671                                 perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
 672                                 attr->branch_sample_type = PERF_SAMPLE_BRANCH_USER |
 673                                                         PERF_SAMPLE_BRANCH_CALL_STACK |
 674                                                         PERF_SAMPLE_BRANCH_NO_CYCLES |
 675                                                         PERF_SAMPLE_BRANCH_NO_FLAGS;
 676                         }
 677                 } else
 678                          pr_warning("Cannot use LBR callstack with branch stack. "
 679                                     "Falling back to framepointers.\n");
 680         }
 681
 682         if (param->record_mode == CALLCHAIN_DWARF) {
 683                 if (!function) {
 684                         perf_evsel__set_sample_bit(evsel, REGS_USER);
 685                         perf_evsel__set_sample_bit(evsel, STACK_USER);
 686                         attr->sample_regs_user |= PERF_REGS_MASK;
 687                         attr->sample_stack_user = param->dump_size;
 688                         attr->exclude_callchain_user = 1;
 689                 } else {
 690                         pr_info("Cannot use DWARF unwind for function trace event,"
 691                                 " falling back to framepointers.\n");
 692                 }
 693         }
 694
 695         if (function) {
 696                 pr_info("Disabling user space callchains for function trace event.\n");
 697                 attr->exclude_callchain_user = 1;
 698         }
 699 }
 700
 701 static void
 702 perf_evsel__reset_callgraph(struct perf_evsel *evsel,
 703                             struct callchain_param *param)
 704 {
 705         struct perf_event_attr *attr = &evsel->attr;
 706
 707         perf_evsel__reset_sample_bit(evsel, CALLCHAIN);
 708         if (param->record_mode == CALLCHAIN_LBR) {
 709                 perf_evsel__reset_sample_bit(evsel, BRANCH_STACK);
 710                 attr->branch_sample_type &= ~(PERF_SAMPLE_BRANCH_USER |
 711                                               PERF_SAMPLE_BRANCH_CALL_STACK);
 712         }
 713         if (param->record_mode == CALLCHAIN_DWARF) {
 714                 perf_evsel__reset_sample_bit(evsel, REGS_USER);
 715                 perf_evsel__reset_sample_bit(evsel, STACK_USER);
 716         }
 717 }
 718
 719 static void apply_config_terms(struct perf_evsel *evsel,
 720                                struct record_opts *opts)
 721 {
 722         struct perf_evsel_config_term *term;
 723         struct list_head *config_terms = &evsel->config_terms;
 724         struct perf_event_attr *attr = &evsel->attr;
 725         struct callchain_param param;
 726         u32 dump_size = 0;
 727         int max_stack = 0;
 728         const char *callgraph_buf = NULL;
 729
 730         /* callgraph default */
 731         param.record_mode = callchain_param.record_mode;
 732
 733         list_for_each_entry(term, config_terms, list) {
 734                 switch (term->type) {
 735                 case PERF_EVSEL__CONFIG_TERM_PERIOD:
 736                         if (!(term->weak && opts->user_interval != ULLONG_MAX)) {
 737                                 attr->sample_period = term->val.period;
 738                                 attr->freq = 0;
 739                         }
 740                         break;
 741                 case PERF_EVSEL__CONFIG_TERM_FREQ:
 742                         if (!(term->weak && opts->user_freq != UINT_MAX)) {
 743                                 attr->sample_freq = term->val.freq;
 744                                 attr->freq = 1;
 745                         }
 746                         break;
 747                 case PERF_EVSEL__CONFIG_TERM_TIME:
 748                         if (term->val.time)
 749                                 perf_evsel__set_sample_bit(evsel, TIME);
 750                         else
 751                                 perf_evsel__reset_sample_bit(evsel, TIME);
 752                         break;
 753                 case PERF_EVSEL__CONFIG_TERM_CALLGRAPH:
 754                         callgraph_buf = term->val.callgraph;
 755                         break;
 756                 case PERF_EVSEL__CONFIG_TERM_BRANCH:
 757                         if (term->val.branch && strcmp(term->val.branch, "no")) {
 758                                 perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
 759                                 parse_branch_str(term->val.branch,
 760                                                  &attr->branch_sample_type);
 761                         } else
 762                                 perf_evsel__reset_sample_bit(evsel, BRANCH_STACK);
 763                         break;
 764                 case PERF_EVSEL__CONFIG_TERM_STACK_USER:
 765                         dump_size = term->val.stack_user;
 766                         break;
 767                 case PERF_EVSEL__CONFIG_TERM_MAX_STACK:
 768                         max_stack = term->val.max_stack;
 769                         break;
 770                 case PERF_EVSEL__CONFIG_TERM_INHERIT:
 771                         /*
 772                          * attr->inherit should has already been set by
 773                          * perf_evsel__config. If user explicitly set
 774                          * inherit using config terms, override global
 775                          * opt->no_inherit setting.
 776                          */
 777                         attr->inherit = term->val.inherit ? 1 : 0;
 778                         break;
 779                 case PERF_EVSEL__CONFIG_TERM_OVERWRITE:
 780                         attr->write_backward = term->val.overwrite ? 1 : 0;
 781                         break;
 782                 default:
 783                         break;
 784                 }
 785         }
 786
 787         /* User explicitly set per-event callgraph, clear the old setting and reset. */
 788         if ((callgraph_buf != NULL) || (dump_size > 0) || max_stack) {
 789                 if (max_stack) {
 790                         param.max_stack = max_stack;
 791                         if (callgraph_buf == NULL)
 792                                 callgraph_buf = "fp";
 793                 }
 794
 795                 /* parse callgraph parameters */
 796                 if (callgraph_buf != NULL) {
 797                         if (!strcmp(callgraph_buf, "no")) {
 798                                 param.enabled = false;
 799                                 param.record_mode = CALLCHAIN_NONE;
 800                         } else {
 801                                 param.enabled = true;
 802                                 if (parse_callchain_record(callgraph_buf, &param)) {
 803                                         pr_err("per-event callgraph setting for %s failed. "
 804                                                "Apply callgraph global setting for it\n",
 805                                                evsel->name);
 806                                         return;
 807                                 }
 808                         }
 809                 }
 810                 if (dump_size > 0) {
 811                         dump_size = round_up(dump_size, sizeof(u64));
 812                         param.dump_size = dump_size;
 813                 }
 814
 815                 /* If global callgraph set, clear it */
 816                 if (callchain_param.enabled)
 817                         perf_evsel__reset_callgraph(evsel, &callchain_param);
 818
 819                 /* set perf-event callgraph */
 820                 if (param.enabled)
 821                         perf_evsel__config_callchain(evsel, opts, &param);
 822         }
 823 }
 824
 825 /*
 826  * The enable_on_exec/disabled value strategy:
 827  *
 828  *  1) For any type of traced program:
 829  *    - all independent events and group leaders are disabled
 830  *    - all group members are enabled
 831  *
 832  *     Group members are ruled by group leaders. They need to
 833  *     be enabled, because the group scheduling relies on that.
 834  *
 835  *  2) For traced programs executed by perf:
 836  *     - all independent events and group leaders have
 837  *       enable_on_exec set
 838  *     - we don't specifically enable or disable any event during
 839  *       the record command
 840  *
 841  *     Independent events and group leaders are initially disabled
 842  *     and get enabled by exec. Group members are ruled by group
 843  *     leaders as stated in 1).
 844  *
 845  *  3) For traced programs attached by perf (pid/tid):
 846  *     - we specifically enable or disable all events during
 847  *       the record command
 848  *
 849  *     When attaching events to already running traced we
 850  *     enable/disable events specifically, as there's no
 851  *     initial traced exec call.
 852  */
 853 void perf_evsel__config(struct perf_evsel *evsel, struct record_opts *opts,
 854                         struct callchain_param *callchain)
 855 {
 856         struct perf_evsel *leader = evsel->leader;
 857         struct perf_event_attr *attr = &evsel->attr;
 858         int track = evsel->tracking;
 859         bool per_cpu = opts->target.default_per_cpu && !opts->target.per_thread;
 860
 861         attr->sample_id_all = perf_missing_features.sample_id_all ? 0 : 1;
 862         attr->inherit       = !opts->no_inherit;
 863         attr->write_backward = opts->overwrite ? 1 : 0;
 864
 865         perf_evsel__set_sample_bit(evsel, IP);
 866         perf_evsel__set_sample_bit(evsel, TID);
 867
 868         if (evsel->sample_read) {
 869                 perf_evsel__set_sample_bit(evsel, READ);
 870
 871                 /*
 872                  * We need ID even in case of single event, because
 873                  * PERF_SAMPLE_READ process ID specific data.
 874                  */
 875                 perf_evsel__set_sample_id(evsel, false);
 876
 877                 /*
 878                  * Apply group format only if we belong to group
 879                  * with more than one members.
 880                  */
 881                 if (leader->nr_members > 1) {
 882                         attr->read_format |= PERF_FORMAT_GROUP;
 883                         attr->inherit = 0;
 884                 }
 885         }
 886
 887         /*
 888          * We default some events to have a default interval. But keep
 889          * it a weak assumption overridable by the user.
 890          */
 891         if (!attr->sample_period || (opts->user_freq != UINT_MAX ||
 892                                      opts->user_interval != ULLONG_MAX)) {
 893                 if (opts->freq) {
 894                         perf_evsel__set_sample_bit(evsel, PERIOD);
 895                         attr->freq              = 1;
 896                         attr->sample_freq       = opts->freq;
 897                 } else {
 898                         attr->sample_period = opts->default_interval;
 899                 }
 900         }
 901
 902         /*
 903          * Disable sampling for all group members other
 904          * than leader in case leader 'leads' the sampling.
 905          */
 906         if ((leader != evsel) && leader->sample_read) {
 907                 attr->sample_freq   = 0;
 908                 attr->sample_period = 0;
 909         }
 910
 911         if (opts->no_samples)
 912                 attr->sample_freq = 0;
 913
 914         if (opts->inherit_stat) {
 915                 evsel->attr.read_format |=
 916                         PERF_FORMAT_TOTAL_TIME_ENABLED |
 917                         PERF_FORMAT_TOTAL_TIME_RUNNING |
 918                         PERF_FORMAT_ID;
 919                 attr->inherit_stat = 1;
 920         }
 921
 922         if (opts->sample_address) {
 923                 perf_evsel__set_sample_bit(evsel, ADDR);
 924                 attr->mmap_data = track;
 925         }
 926
 927         /*
 928          * We don't allow user space callchains for  function trace
 929          * event, due to issues with page faults while tracing page
 930          * fault handler and its overall trickiness nature.
 931          */
 932         if (perf_evsel__is_function_event(evsel))
 933                 evsel->attr.exclude_callchain_user = 1;
 934
 935         if (callchain && callchain->enabled && !evsel->no_aux_samples)
 936                 perf_evsel__config_callchain(evsel, opts, callchain);
 937
 938         if (opts->sample_intr_regs) {
 939                 attr->sample_regs_intr = opts->sample_intr_regs;
 940                 perf_evsel__set_sample_bit(evsel, REGS_INTR);
 941         }
 942
 943         if (opts->sample_user_regs) {
 944                 attr->sample_regs_user |= opts->sample_user_regs;
 945                 perf_evsel__set_sample_bit(evsel, REGS_USER);
 946         }
 947
 948         if (target__has_cpu(&opts->target) || opts->sample_cpu)
 949                 perf_evsel__set_sample_bit(evsel, CPU);
 950
 951         if (opts->period)
 952                 perf_evsel__set_sample_bit(evsel, PERIOD);
 953
 954         /*
 955          * When the user explicitly disabled time don't force it here.
 956          */
 957         if (opts->sample_time &&
 958             (!perf_missing_features.sample_id_all &&
 959             (!opts->no_inherit || target__has_cpu(&opts->target) || per_cpu ||
 960              opts->sample_time_set)))
 961                 perf_evsel__set_sample_bit(evsel, TIME);
 962
 963         if (opts->raw_samples && !evsel->no_aux_samples) {
 964                 perf_evsel__set_sample_bit(evsel, TIME);
 965                 perf_evsel__set_sample_bit(evsel, RAW);
 966                 perf_evsel__set_sample_bit(evsel, CPU);
 967         }
 968
 969         if (opts->sample_address)
 970                 perf_evsel__set_sample_bit(evsel, DATA_SRC);
 971
 972         if (opts->sample_phys_addr)
 973                 perf_evsel__set_sample_bit(evsel, PHYS_ADDR);
 974
 975         if (opts->no_buffering) {
 976                 attr->watermark = 0;
 977                 attr->wakeup_events = 1;
 978         }
 979         if (opts->branch_stack && !evsel->no_aux_samples) {
 980                 perf_evsel__set_sample_bit(evsel, BRANCH_STACK);
 981                 attr->branch_sample_type = opts->branch_stack;
 982         }
 983
 984         if (opts->sample_weight)
 985                 perf_evsel__set_sample_bit(evsel, WEIGHT);
 986
 987         attr->task  = track;
 988         attr->mmap  = track;
 989         attr->mmap2 = track && !perf_missing_features.mmap2;
 990         attr->comm  = track;
 991
 992         if (opts->record_namespaces)
 993                 attr->namespaces  = track;
 994
 995         if (opts->record_switch_events)
 996                 attr->context_switch = track;
 997
 998         if (opts->sample_transaction)
 999                 perf_evsel__set_sample_bit(evsel, TRANSACTION);
1000
1001         if (opts->running_time) {
1002                 evsel->attr.read_format |=
1003                         PERF_FORMAT_TOTAL_TIME_ENABLED |
1004                         PERF_FORMAT_TOTAL_TIME_RUNNING;
1005         }
1006
1007         /*
1008          * XXX see the function comment above
1009          *
1010          * Disabling only independent events or group leaders,
1011          * keeping group members enabled.
1012          */
1013         if (perf_evsel__is_group_leader(evsel))
1014                 attr->disabled = 1;
1015
1016         /*
1017          * Setting enable_on_exec for independent events and
1018          * group leaders for traced executed by perf.
1019          */
1020         if (target__none(&opts->target) && perf_evsel__is_group_leader(evsel) &&
1021                 !opts->initial_delay)
1022                 attr->enable_on_exec = 1;
1023
1024         if (evsel->immediate) {
1025                 attr->disabled = 0;
1026                 attr->enable_on_exec = 0;
1027         }
1028
1029         clockid = opts->clockid;
1030         if (opts->use_clockid) {
1031                 attr->use_clockid = 1;
1032                 attr->clockid = opts->clockid;
1033         }
1034
1035         if (evsel->precise_max)
1036                 perf_event_attr__set_max_precise_ip(attr);
1037
1038         if (opts->all_user) {
1039                 attr->exclude_kernel = 1;
1040                 attr->exclude_user   = 0;
1041         }
1042
1043         if (opts->all_kernel) {
1044                 attr->exclude_kernel = 0;
1045                 attr->exclude_user   = 1;
1046         }
1047
1048         /*
1049          * Apply event specific term settings,
1050          * it overloads any global configuration.
1051          */
1052         apply_config_terms(evsel, opts);
1053
1054         evsel->ignore_missing_thread = opts->ignore_missing_thread;
1055 }
1056
1057 static int perf_evsel__alloc_fd(struct perf_evsel *evsel, int ncpus, int nthreads)
1058 {
1059         if (evsel->system_wide)
1060                 nthreads = 1;
1061
1062         evsel->fd = xyarray__new(ncpus, nthreads, sizeof(int));
1063
1064         if (evsel->fd) {
1065                 int cpu, thread;
1066                 for (cpu = 0; cpu < ncpus; cpu++) {
1067                         for (thread = 0; thread < nthreads; thread++) {
1068                                 FD(evsel, cpu, thread) = -1;
1069                         }
1070                 }
1071         }
1072
1073         return evsel->fd != NULL ? 0 : -ENOMEM;
1074 }
1075
1076 static int perf_evsel__run_ioctl(struct perf_evsel *evsel,
1077                           int ioc,  void *arg)
1078 {
1079         int cpu, thread;
1080
1081         for (cpu = 0; cpu < xyarray__max_x(evsel->fd); cpu++) {
1082                 for (thread = 0; thread < xyarray__max_y(evsel->fd); thread++) {
1083                         int fd = FD(evsel, cpu, thread),
1084                             err = ioctl(fd, ioc, arg);
1085
1086                         if (err)
1087                                 return err;
1088                 }
1089         }
1090
1091         return 0;
1092 }
1093
1094 int perf_evsel__apply_filter(struct perf_evsel *evsel, const char *filter)
1095 {
1096         return perf_evsel__run_ioctl(evsel,
1097                                      PERF_EVENT_IOC_SET_FILTER,
1098                                      (void *)filter);
1099 }
1100
1101 int perf_evsel__set_filter(struct perf_evsel *evsel, const char *filter)
1102 {
1103         char *new_filter = strdup(filter);
1104
1105         if (new_filter != NULL) {
1106                 free(evsel->filter);
1107                 evsel->filter = new_filter;
1108                 return 0;
1109         }
1110
1111         return -1;
1112 }
1113
1114 static int perf_evsel__append_filter(struct perf_evsel *evsel,
1115                                      const char *fmt, const char *filter)
1116 {
1117         char *new_filter;
1118
1119         if (evsel->filter == NULL)
1120                 return perf_evsel__set_filter(evsel, filter);
1121
1122         if (asprintf(&new_filter, fmt, evsel->filter, filter) > 0) {
1123                 free(evsel->filter);
1124                 evsel->filter = new_filter;
1125                 return 0;
1126         }
1127
1128         return -1;
1129 }
1130
1131 int perf_evsel__append_tp_filter(struct perf_evsel *evsel, const char *filter)
1132 {
1133         return perf_evsel__append_filter(evsel, "(%s) && (%s)", filter);
1134 }
1135
1136 int perf_evsel__append_addr_filter(struct perf_evsel *evsel, const char *filter)
1137 {
1138         return perf_evsel__append_filter(evsel, "%s,%s", filter);
1139 }
1140
1141 int perf_evsel__enable(struct perf_evsel *evsel)
1142 {
1143         return perf_evsel__run_ioctl(evsel,
1144                                      PERF_EVENT_IOC_ENABLE,
1145                                      0);
1146 }
1147
1148 int perf_evsel__disable(struct perf_evsel *evsel)
1149 {
1150         return perf_evsel__run_ioctl(evsel,
1151                                      PERF_EVENT_IOC_DISABLE,
1152                                      0);
1153 }
1154
1155 int perf_evsel__alloc_id(struct perf_evsel *evsel, int ncpus, int nthreads)
1156 {
1157         if (ncpus == 0 || nthreads == 0)
1158                 return 0;
1159
1160         if (evsel->system_wide)
1161                 nthreads = 1;
1162
1163         evsel->sample_id = xyarray__new(ncpus, nthreads, sizeof(struct perf_sample_id));
1164         if (evsel->sample_id == NULL)
1165                 return -ENOMEM;
1166
1167         evsel->id = zalloc(ncpus * nthreads * sizeof(u64));
1168         if (evsel->id == NULL) {
1169                 xyarray__delete(evsel->sample_id);
1170                 evsel->sample_id = NULL;
1171                 return -ENOMEM;
1172         }
1173
1174         return 0;
1175 }
1176
1177 static void perf_evsel__free_fd(struct perf_evsel *evsel)
1178 {
1179         xyarray__delete(evsel->fd);
1180         evsel->fd = NULL;
1181 }
1182
1183 static void perf_evsel__free_id(struct perf_evsel *evsel)
1184 {
1185         xyarray__delete(evsel->sample_id);
1186         evsel->sample_id = NULL;
1187         zfree(&evsel->id);
1188 }
1189
1190 static void perf_evsel__free_config_terms(struct perf_evsel *evsel)
1191 {
1192         struct perf_evsel_config_term *term, *h;
1193
1194         list_for_each_entry_safe(term, h, &evsel->config_terms, list) {
1195                 list_del(&term->list);
1196                 free(term);
1197         }
1198 }
1199
1200 void perf_evsel__close_fd(struct perf_evsel *evsel)
1201 {
1202         int cpu, thread;
1203
1204         for (cpu = 0; cpu < xyarray__max_x(evsel->fd); cpu++)
1205                 for (thread = 0; thread < xyarray__max_y(evsel->fd); ++thread) {
1206                         close(FD(evsel, cpu, thread));
1207                         FD(evsel, cpu, thread) = -1;
1208                 }
1209 }
1210
1211 void perf_evsel__exit(struct perf_evsel *evsel)
1212 {
1213         assert(list_empty(&evsel->node));
1214         assert(evsel->evlist == NULL);
1215         perf_evsel__free_fd(evsel);
1216         perf_evsel__free_id(evsel);
1217         perf_evsel__free_config_terms(evsel);
1218         close_cgroup(evsel->cgrp);
1219         cpu_map__put(evsel->cpus);
1220         cpu_map__put(evsel->own_cpus);
1221         thread_map__put(evsel->threads);
1222         zfree(&evsel->group_name);
1223         zfree(&evsel->name);
1224         perf_evsel__object.fini(evsel);
1225 }
1226
1227 void perf_evsel__delete(struct perf_evsel *evsel)
1228 {
1229         perf_evsel__exit(evsel);
1230         free(evsel);
1231 }
1232
1233 void perf_evsel__compute_deltas(struct perf_evsel *evsel, int cpu, int thread,
1234                                 struct perf_counts_values *count)
1235 {
1236         struct perf_counts_values tmp;
1237
1238         if (!evsel->prev_raw_counts)
1239                 return;
1240
1241         if (cpu == -1) {
1242                 tmp = evsel->prev_raw_counts->aggr;
1243                 evsel->prev_raw_counts->aggr = *count;
1244         } else {
1245                 tmp = *perf_counts(evsel->prev_raw_counts, cpu, thread);
1246                 *perf_counts(evsel->prev_raw_counts, cpu, thread) = *count;
1247         }
1248
1249         count->val = count->val - tmp.val;
1250         count->ena = count->ena - tmp.ena;
1251         count->run = count->run - tmp.run;
1252 }
1253
1254 void perf_counts_values__scale(struct perf_counts_values *count,
1255                                bool scale, s8 *pscaled)
1256 {
1257         s8 scaled = 0;
1258
1259         if (scale) {
1260                 if (count->run == 0) {
1261                         scaled = -1;
1262                         count->val = 0;
1263                 } else if (count->run < count->ena) {
1264                         scaled = 1;
1265                         count->val = (u64)((double) count->val * count->ena / count->run + 0.5);
1266                 }
1267         } else
1268                 count->ena = count->run = 0;
1269
1270         if (pscaled)
1271                 *pscaled = scaled;
1272 }
1273
1274 static int perf_evsel__read_size(struct perf_evsel *evsel)
1275 {
1276         u64 read_format = evsel->attr.read_format;
1277         int entry = sizeof(u64); /* value */
1278         int size = 0;
1279         int nr = 1;
1280
1281         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1282                 size += sizeof(u64);
1283
1284         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1285                 size += sizeof(u64);
1286
1287         if (read_format & PERF_FORMAT_ID)
1288                 entry += sizeof(u64);
1289
1290         if (read_format & PERF_FORMAT_GROUP) {
1291                 nr = evsel->nr_members;
1292                 size += sizeof(u64);
1293         }
1294
1295         size += entry * nr;
1296         return size;
1297 }
1298
1299 int perf_evsel__read(struct perf_evsel *evsel, int cpu, int thread,
1300                      struct perf_counts_values *count)
1301 {
1302         size_t size = perf_evsel__read_size(evsel);
1303
1304         memset(count, 0, sizeof(*count));
1305
1306         if (FD(evsel, cpu, thread) < 0)
1307                 return -EINVAL;
1308
1309         if (readn(FD(evsel, cpu, thread), count->values, size) <= 0)
1310                 return -errno;
1311
1312         return 0;
1313 }
1314
1315 static int
1316 perf_evsel__read_one(struct perf_evsel *evsel, int cpu, int thread)
1317 {
1318         struct perf_counts_values *count = perf_counts(evsel->counts, cpu, thread);
1319
1320         return perf_evsel__read(evsel, cpu, thread, count);
1321 }
1322
1323 static void
1324 perf_evsel__set_count(struct perf_evsel *counter, int cpu, int thread,
1325                       u64 val, u64 ena, u64 run)
1326 {
1327         struct perf_counts_values *count;
1328
1329         count = perf_counts(counter->counts, cpu, thread);
1330
1331         count->val    = val;
1332         count->ena    = ena;
1333         count->run    = run;
1334         count->loaded = true;
1335 }
1336
1337 static int
1338 perf_evsel__process_group_data(struct perf_evsel *leader,
1339                                int cpu, int thread, u64 *data)
1340 {
1341         u64 read_format = leader->attr.read_format;
1342         struct sample_read_value *v;
1343         u64 nr, ena = 0, run = 0, i;
1344
1345         nr = *data++;
1346
1347         if (nr != (u64) leader->nr_members)
1348                 return -EINVAL;
1349
1350         if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
1351                 ena = *data++;
1352
1353         if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
1354                 run = *data++;
1355
1356         v = (struct sample_read_value *) data;
1357
1358         perf_evsel__set_count(leader, cpu, thread,
1359                               v[0].value, ena, run);
1360
1361         for (i = 1; i < nr; i++) {
1362                 struct perf_evsel *counter;
1363
1364                 counter = perf_evlist__id2evsel(leader->evlist, v[i].id);
1365                 if (!counter)
1366                         return -EINVAL;
1367
1368                 perf_evsel__set_count(counter, cpu, thread,
1369                                       v[i].value, ena, run);
1370         }
1371
1372         return 0;
1373 }
1374
1375 static int
1376 perf_evsel__read_group(struct perf_evsel *leader, int cpu, int thread)
1377 {
1378         struct perf_stat_evsel *ps = leader->stats;
1379         u64 read_format = leader->attr.read_format;
1380         int size = perf_evsel__read_size(leader);
1381         u64 *data = ps->group_data;
1382
1383         if (!(read_format & PERF_FORMAT_ID))
1384                 return -EINVAL;
1385
1386         if (!perf_evsel__is_group_leader(leader))
1387                 return -EINVAL;
1388
1389         if (!data) {
1390                 data = zalloc(size);
1391                 if (!data)
1392                         return -ENOMEM;
1393
1394                 ps->group_data = data;
1395         }
1396
1397         if (FD(leader, cpu, thread) < 0)
1398                 return -EINVAL;
1399
1400         if (readn(FD(leader, cpu, thread), data, size) <= 0)
1401                 return -errno;
1402
1403         return perf_evsel__process_group_data(leader, cpu, thread, data);
1404 }
1405
1406 int perf_evsel__read_counter(struct perf_evsel *evsel, int cpu, int thread)
1407 {
1408         u64 read_format = evsel->attr.read_format;
1409
1410         if (read_format & PERF_FORMAT_GROUP)
1411                 return perf_evsel__read_group(evsel, cpu, thread);
1412         else
1413                 return perf_evsel__read_one(evsel, cpu, thread);
1414 }
1415
1416 int __perf_evsel__read_on_cpu(struct perf_evsel *evsel,
1417                               int cpu, int thread, bool scale)
1418 {
1419         struct perf_counts_values count;
1420         size_t nv = scale ? 3 : 1;
1421
1422         if (FD(evsel, cpu, thread) < 0)
1423                 return -EINVAL;
1424
1425         if (evsel->counts == NULL && perf_evsel__alloc_counts(evsel, cpu + 1, thread + 1) < 0)
1426                 return -ENOMEM;
1427
1428         if (readn(FD(evsel, cpu, thread), &count, nv * sizeof(u64)) <= 0)
1429                 return -errno;
1430
1431         perf_evsel__compute_deltas(evsel, cpu, thread, &count);
1432         perf_counts_values__scale(&count, scale, NULL);
1433         *perf_counts(evsel->counts, cpu, thread) = count;
1434         return 0;
1435 }
1436
1437 static int get_group_fd(struct perf_evsel *evsel, int cpu, int thread)
1438 {
1439         struct perf_evsel *leader = evsel->leader;
1440         int fd;
1441
1442         if (perf_evsel__is_group_leader(evsel))
1443                 return -1;
1444
1445         /*
1446          * Leader must be already processed/open,
1447          * if not it's a bug.
1448          */
1449         BUG_ON(!leader->fd);
1450
1451         fd = FD(leader, cpu, thread);
1452         BUG_ON(fd == -1);
1453
1454         return fd;
1455 }
1456
1457 struct bit_names {
1458         int bit;
1459         const char *name;
1460 };
1461
1462 static void __p_bits(char *buf, size_t size, u64 value, struct bit_names *bits)
1463 {
1464         bool first_bit = true;
1465         int i = 0;
1466
1467         do {
1468                 if (value & bits[i].bit) {
1469                         buf += scnprintf(buf, size, "%s%s", first_bit ? "" : "|", bits[i].name);
1470                         first_bit = false;
1471                 }
1472         } while (bits[++i].name != NULL);
1473 }
1474
1475 static void __p_sample_type(char *buf, size_t size, u64 value)
1476 {
1477 #define bit_name(n) { PERF_SAMPLE_##n, #n }
1478         struct bit_names bits[] = {
1479                 bit_name(IP), bit_name(TID), bit_name(TIME), bit_name(ADDR),
1480                 bit_name(READ), bit_name(CALLCHAIN), bit_name(ID), bit_name(CPU),
1481                 bit_name(PERIOD), bit_name(STREAM_ID), bit_name(RAW),
1482                 bit_name(BRANCH_STACK), bit_name(REGS_USER), bit_name(STACK_USER),
1483                 bit_name(IDENTIFIER), bit_name(REGS_INTR), bit_name(DATA_SRC),
1484                 bit_name(WEIGHT), bit_name(PHYS_ADDR),
1485                 { .name = NULL, }
1486         };
1487 #undef bit_name
1488         __p_bits(buf, size, value, bits);
1489 }
1490
1491 static void __p_branch_sample_type(char *buf, size_t size, u64 value)
1492 {
1493 #define bit_name(n) { PERF_SAMPLE_BRANCH_##n, #n }
1494         struct bit_names bits[] = {
1495                 bit_name(USER), bit_name(KERNEL), bit_name(HV), bit_name(ANY),
1496                 bit_name(ANY_CALL), bit_name(ANY_RETURN), bit_name(IND_CALL),
1497                 bit_name(ABORT_TX), bit_name(IN_TX), bit_name(NO_TX),
1498                 bit_name(COND), bit_name(CALL_STACK), bit_name(IND_JUMP),
1499                 bit_name(CALL), bit_name(NO_FLAGS), bit_name(NO_CYCLES),
1500                 { .name = NULL, }
1501         };
1502 #undef bit_name
1503         __p_bits(buf, size, value, bits);
1504 }
1505
1506 static void __p_read_format(char *buf, size_t size, u64 value)
1507 {
1508 #define bit_name(n) { PERF_FORMAT_##n, #n }
1509         struct bit_names bits[] = {
1510                 bit_name(TOTAL_TIME_ENABLED), bit_name(TOTAL_TIME_RUNNING),
1511                 bit_name(ID), bit_name(GROUP),
1512                 { .name = NULL, }
1513         };
1514 #undef bit_name
1515         __p_bits(buf, size, value, bits);
1516 }
1517
1518 #define BUF_SIZE                1024
1519
1520 #define p_hex(val)              snprintf(buf, BUF_SIZE, "%#"PRIx64, (uint64_t)(val))
1521 #define p_unsigned(val)         snprintf(buf, BUF_SIZE, "%"PRIu64, (uint64_t)(val))
1522 #define p_signed(val)           snprintf(buf, BUF_SIZE, "%"PRId64, (int64_t)(val))
1523 #define p_sample_type(val)      __p_sample_type(buf, BUF_SIZE, val)
1524 #define p_branch_sample_type(val) __p_branch_sample_type(buf, BUF_SIZE, val)
1525 #define p_read_format(val)      __p_read_format(buf, BUF_SIZE, val)
1526
1527 #define PRINT_ATTRn(_n, _f, _p)                         \
1528 do {                                                    \
1529         if (attr->_f) {                                 \
1530                 _p(attr->_f);                           \
1531                 ret += attr__fprintf(fp, _n, buf, priv);\
1532         }                                               \
1533 } while (0)
1534
1535 #define PRINT_ATTRf(_f, _p)     PRINT_ATTRn(#_f, _f, _p)
1536
1537 int perf_event_attr__fprintf(FILE *fp, struct perf_event_attr *attr,
1538                              attr__fprintf_f attr__fprintf, void *priv)
1539 {
1540         char buf[BUF_SIZE];
1541         int ret = 0;
1542
1543         PRINT_ATTRf(type, p_unsigned);
1544         PRINT_ATTRf(size, p_unsigned);
1545         PRINT_ATTRf(config, p_hex);
1546         PRINT_ATTRn("{ sample_period, sample_freq }", sample_period, p_unsigned);
1547         PRINT_ATTRf(sample_type, p_sample_type);
1548         PRINT_ATTRf(read_format, p_read_format);
1549
1550         PRINT_ATTRf(disabled, p_unsigned);
1551         PRINT_ATTRf(inherit, p_unsigned);
1552         PRINT_ATTRf(pinned, p_unsigned);
1553         PRINT_ATTRf(exclusive, p_unsigned);
1554         PRINT_ATTRf(exclude_user, p_unsigned);
1555         PRINT_ATTRf(exclude_kernel, p_unsigned);
1556         PRINT_ATTRf(exclude_hv, p_unsigned);
1557         PRINT_ATTRf(exclude_idle, p_unsigned);
1558         PRINT_ATTRf(mmap, p_unsigned);
1559         PRINT_ATTRf(comm, p_unsigned);
1560         PRINT_ATTRf(freq, p_unsigned);
1561         PRINT_ATTRf(inherit_stat, p_unsigned);
1562         PRINT_ATTRf(enable_on_exec, p_unsigned);
1563         PRINT_ATTRf(task, p_unsigned);
1564         PRINT_ATTRf(watermark, p_unsigned);
1565         PRINT_ATTRf(precise_ip, p_unsigned);
1566         PRINT_ATTRf(mmap_data, p_unsigned);
1567         PRINT_ATTRf(sample_id_all, p_unsigned);
1568         PRINT_ATTRf(exclude_host, p_unsigned);
1569         PRINT_ATTRf(exclude_guest, p_unsigned);
1570         PRINT_ATTRf(exclude_callchain_kernel, p_unsigned);
1571         PRINT_ATTRf(exclude_callchain_user, p_unsigned);
1572         PRINT_ATTRf(mmap2, p_unsigned);
1573         PRINT_ATTRf(comm_exec, p_unsigned);
1574         PRINT_ATTRf(use_clockid, p_unsigned);
1575         PRINT_ATTRf(context_switch, p_unsigned);
1576         PRINT_ATTRf(write_backward, p_unsigned);
1577
1578         PRINT_ATTRn("{ wakeup_events, wakeup_watermark }", wakeup_events, p_unsigned);
1579         PRINT_ATTRf(bp_type, p_unsigned);
1580         PRINT_ATTRn("{ bp_addr, config1 }", bp_addr, p_hex);
1581         PRINT_ATTRn("{ bp_len, config2 }", bp_len, p_hex);
1582         PRINT_ATTRf(branch_sample_type, p_branch_sample_type);
1583         PRINT_ATTRf(sample_regs_user, p_hex);
1584         PRINT_ATTRf(sample_stack_user, p_unsigned);
1585         PRINT_ATTRf(clockid, p_signed);
1586         PRINT_ATTRf(sample_regs_intr, p_hex);
1587         PRINT_ATTRf(aux_watermark, p_unsigned);
1588         PRINT_ATTRf(sample_max_stack, p_unsigned);
1589
1590         return ret;
1591 }
1592
1593 static int __open_attr__fprintf(FILE *fp, const char *name, const char *val,
1594                                 void *priv __maybe_unused)
1595 {
1596         return fprintf(fp, "  %-32s %s\n", name, val);
1597 }
1598
1599 static bool ignore_missing_thread(struct perf_evsel *evsel,
1600                                   struct thread_map *threads,
1601                                   int thread, int err)
1602 {
1603         if (!evsel->ignore_missing_thread)
1604                 return false;
1605
1606         /* The system wide setup does not work with threads. */
1607         if (evsel->system_wide)
1608                 return false;
1609
1610         /* The -ESRCH is perf event syscall errno for pid's not found. */
1611         if (err != -ESRCH)
1612                 return false;
1613
1614         /* If there's only one thread, let it fail. */
1615         if (threads->nr == 1)
1616                 return false;
1617
1618         if (thread_map__remove(threads, thread))
1619                 return false;
1620
1621         pr_warning("WARNING: Ignored open failure for pid %d\n",
1622                    thread_map__pid(threads, thread));
1623         return true;
1624 }
1625
1626 int perf_evsel__open(struct perf_evsel *evsel, struct cpu_map *cpus,
1627                      struct thread_map *threads)
1628 {
1629         int cpu, thread, nthreads;
1630         unsigned long flags = PERF_FLAG_FD_CLOEXEC;
1631         int pid = -1, err;
1632         enum { NO_CHANGE, SET_TO_MAX, INCREASED_MAX } set_rlimit = NO_CHANGE;
1633
1634         if (perf_missing_features.write_backward && evsel->attr.write_backward)
1635                 return -EINVAL;
1636
1637         if (cpus == NULL) {
1638                 static struct cpu_map *empty_cpu_map;
1639
1640                 if (empty_cpu_map == NULL) {
1641                         empty_cpu_map = cpu_map__dummy_new();
1642                         if (empty_cpu_map == NULL)
1643                                 return -ENOMEM;
1644                 }
1645
1646                 cpus = empty_cpu_map;
1647         }
1648
1649         if (threads == NULL) {
1650                 static struct thread_map *empty_thread_map;
1651
1652                 if (empty_thread_map == NULL) {
1653                         empty_thread_map = thread_map__new_by_tid(-1);
1654                         if (empty_thread_map == NULL)
1655                                 return -ENOMEM;
1656                 }
1657
1658                 threads = empty_thread_map;
1659         }
1660
1661         if (evsel->system_wide)
1662                 nthreads = 1;
1663         else
1664                 nthreads = threads->nr;
1665
1666         if (evsel->fd == NULL &&
1667             perf_evsel__alloc_fd(evsel, cpus->nr, nthreads) < 0)
1668                 return -ENOMEM;
1669
1670         if (evsel->cgrp) {
1671                 flags |= PERF_FLAG_PID_CGROUP;
1672                 pid = evsel->cgrp->fd;
1673         }
1674
1675 fallback_missing_features:
1676         if (perf_missing_features.clockid_wrong)
1677                 evsel->attr.clockid = CLOCK_MONOTONIC; /* should always work */
1678         if (perf_missing_features.clockid) {
1679                 evsel->attr.use_clockid = 0;
1680                 evsel->attr.clockid = 0;
1681         }
1682         if (perf_missing_features.cloexec)
1683                 flags &= ~(unsigned long)PERF_FLAG_FD_CLOEXEC;
1684         if (perf_missing_features.mmap2)
1685                 evsel->attr.mmap2 = 0;
1686         if (perf_missing_features.exclude_guest)
1687                 evsel->attr.exclude_guest = evsel->attr.exclude_host = 0;
1688         if (perf_missing_features.lbr_flags)
1689                 evsel->attr.branch_sample_type &= ~(PERF_SAMPLE_BRANCH_NO_FLAGS |
1690                                      PERF_SAMPLE_BRANCH_NO_CYCLES);
1691         if (perf_missing_features.group_read && evsel->attr.inherit)
1692                 evsel->attr.read_format &= ~(PERF_FORMAT_GROUP|PERF_FORMAT_ID);
1693 retry_sample_id:
1694         if (perf_missing_features.sample_id_all)
1695                 evsel->attr.sample_id_all = 0;
1696
1697         if (verbose >= 2) {
1698                 fprintf(stderr, "%.60s\n", graph_dotted_line);
1699                 fprintf(stderr, "perf_event_attr:\n");
1700                 perf_event_attr__fprintf(stderr, &evsel->attr, __open_attr__fprintf, NULL);
1701                 fprintf(stderr, "%.60s\n", graph_dotted_line);
1702         }
1703
1704         for (cpu = 0; cpu < cpus->nr; cpu++) {
1705
1706                 for (thread = 0; thread < nthreads; thread++) {
1707                         int fd, group_fd;
1708
1709                         if (!evsel->cgrp && !evsel->system_wide)
1710                                 pid = thread_map__pid(threads, thread);
1711
1712                         group_fd = get_group_fd(evsel, cpu, thread);
1713 retry_open:
1714                         pr_debug2("sys_perf_event_open: pid %d  cpu %d  group_fd %d  flags %#lx",
1715                                   pid, cpus->map[cpu], group_fd, flags);
1716
1717                         test_attr__ready();
1718
1719                         fd = sys_perf_event_open(&evsel->attr, pid, cpus->map[cpu],
1720                                                  group_fd, flags);
1721
1722                         FD(evsel, cpu, thread) = fd;
1723
1724                         if (fd < 0) {
1725                                 err = -errno;
1726
1727                                 if (ignore_missing_thread(evsel, threads, thread, err)) {
1728                                         /*
1729                                          * We just removed 1 thread, so take a step
1730                                          * back on thread index and lower the upper
1731                                          * nthreads limit.
1732                                          */
1733                                         nthreads--;
1734                                         thread--;
1735
1736                                         /* ... and pretend like nothing have happened. */
1737                                         err = 0;
1738                                         continue;
1739                                 }
1740
1741                                 pr_debug2("\nsys_perf_event_open failed, error %d\n",
1742                                           err);
1743                                 goto try_fallback;
1744                         }
1745
1746                         pr_debug2(" = %d\n", fd);
1747
1748                         if (evsel->bpf_fd >= 0) {
1749                                 int evt_fd = fd;
1750                                 int bpf_fd = evsel->bpf_fd;
1751
1752                                 err = ioctl(evt_fd,
1753                                             PERF_EVENT_IOC_SET_BPF,
1754                                             bpf_fd);
1755                                 if (err && errno != EEXIST) {
1756                                         pr_err("failed to attach bpf fd %d: %s\n",
1757                                                bpf_fd, strerror(errno));
1758                                         err = -EINVAL;
1759                                         goto out_close;
1760                                 }
1761                         }
1762
1763                         set_rlimit = NO_CHANGE;
1764
1765                         /*
1766                          * If we succeeded but had to kill clockid, fail and
1767                          * have perf_evsel__open_strerror() print us a nice
1768                          * error.
1769                          */
1770                         if (perf_missing_features.clockid ||
1771                             perf_missing_features.clockid_wrong) {
1772                                 err = -EINVAL;
1773                                 goto out_close;
1774                         }
1775                 }
1776         }
1777
1778         return 0;
1779
1780 try_fallback:
1781         /*
1782          * perf stat needs between 5 and 22 fds per CPU. When we run out
1783          * of them try to increase the limits.
1784          */
1785         if (err == -EMFILE && set_rlimit < INCREASED_MAX) {
1786                 struct rlimit l;
1787                 int old_errno = errno;
1788
1789                 if (getrlimit(RLIMIT_NOFILE, &l) == 0) {
1790                         if (set_rlimit == NO_CHANGE)
1791                                 l.rlim_cur = l.rlim_max;
1792                         else {
1793                                 l.rlim_cur = l.rlim_max + 1000;
1794                                 l.rlim_max = l.rlim_cur;
1795                         }
1796                         if (setrlimit(RLIMIT_NOFILE, &l) == 0) {
1797                                 set_rlimit++;
1798                                 errno = old_errno;
1799                                 goto retry_open;
1800                         }
1801                 }
1802                 errno = old_errno;
1803         }
1804
1805         if (err != -EINVAL || cpu > 0 || thread > 0)
1806                 goto out_close;
1807
1808         /*
1809          * Must probe features in the order they were added to the
1810          * perf_event_attr interface.
1811          */
1812         if (!perf_missing_features.write_backward && evsel->attr.write_backward) {
1813                 perf_missing_features.write_backward = true;
1814                 pr_debug2("switching off write_backward\n");
1815                 goto out_close;
1816         } else if (!perf_missing_features.clockid_wrong && evsel->attr.use_clockid) {
1817                 perf_missing_features.clockid_wrong = true;
1818                 pr_debug2("switching off clockid\n");
1819                 goto fallback_missing_features;
1820         } else if (!perf_missing_features.clockid && evsel->attr.use_clockid) {
1821                 perf_missing_features.clockid = true;
1822                 pr_debug2("switching off use_clockid\n");
1823                 goto fallback_missing_features;
1824         } else if (!perf_missing_features.cloexec && (flags & PERF_FLAG_FD_CLOEXEC)) {
1825                 perf_missing_features.cloexec = true;
1826                 pr_debug2("switching off cloexec flag\n");
1827                 goto fallback_missing_features;
1828         } else if (!perf_missing_features.mmap2 && evsel->attr.mmap2) {
1829                 perf_missing_features.mmap2 = true;
1830                 pr_debug2("switching off mmap2\n");
1831                 goto fallback_missing_features;
1832         } else if (!perf_missing_features.exclude_guest &&
1833                    (evsel->attr.exclude_guest || evsel->attr.exclude_host)) {
1834                 perf_missing_features.exclude_guest = true;
1835                 pr_debug2("switching off exclude_guest, exclude_host\n");
1836                 goto fallback_missing_features;
1837         } else if (!perf_missing_features.sample_id_all) {
1838                 perf_missing_features.sample_id_all = true;
1839                 pr_debug2("switching off sample_id_all\n");
1840                 goto retry_sample_id;
1841         } else if (!perf_missing_features.lbr_flags &&
1842                         (evsel->attr.branch_sample_type &
1843                          (PERF_SAMPLE_BRANCH_NO_CYCLES |
1844                           PERF_SAMPLE_BRANCH_NO_FLAGS))) {
1845                 perf_missing_features.lbr_flags = true;
1846                 pr_debug2("switching off branch sample type no (cycles/flags)\n");
1847                 goto fallback_missing_features;
1848         } else if (!perf_missing_features.group_read &&
1849                     evsel->attr.inherit &&
1850                    (evsel->attr.read_format & PERF_FORMAT_GROUP)) {
1851                 perf_missing_features.group_read = true;
1852                 pr_debug2("switching off group read\n");
1853                 goto fallback_missing_features;
1854         }
1855 out_close:
1856         do {
1857                 while (--thread >= 0) {
1858                         close(FD(evsel, cpu, thread));
1859                         FD(evsel, cpu, thread) = -1;
1860                 }
1861                 thread = nthreads;
1862         } while (--cpu >= 0);
1863         return err;
1864 }
1865
1866 void perf_evsel__close(struct perf_evsel *evsel)
1867 {
1868         if (evsel->fd == NULL)
1869                 return;
1870
1871         perf_evsel__close_fd(evsel);
1872         perf_evsel__free_fd(evsel);
1873 }
1874
1875 int perf_evsel__open_per_cpu(struct perf_evsel *evsel,
1876                              struct cpu_map *cpus)
1877 {
1878         return perf_evsel__open(evsel, cpus, NULL);
1879 }
1880
1881 int perf_evsel__open_per_thread(struct perf_evsel *evsel,
1882                                 struct thread_map *threads)
1883 {
1884         return perf_evsel__open(evsel, NULL, threads);
1885 }
1886
1887 static int perf_evsel__parse_id_sample(const struct perf_evsel *evsel,
1888                                        const union perf_event *event,
1889                                        struct perf_sample *sample)
1890 {
1891         u64 type = evsel->attr.sample_type;
1892         const u64 *array = event->sample.array;
1893         bool swapped = evsel->needs_swap;
1894         union u64_swap u;
1895
1896         array += ((event->header.size -
1897                    sizeof(event->header)) / sizeof(u64)) - 1;
1898
1899         if (type & PERF_SAMPLE_IDENTIFIER) {
1900                 sample->id = *array;
1901                 array--;
1902         }
1903
1904         if (type & PERF_SAMPLE_CPU) {
1905                 u.val64 = *array;
1906                 if (swapped) {
1907                         /* undo swap of u64, then swap on individual u32s */
1908                         u.val64 = bswap_64(u.val64);
1909                         u.val32[0] = bswap_32(u.val32[0]);
1910                 }
1911
1912                 sample->cpu = u.val32[0];
1913                 array--;
1914         }
1915
1916         if (type & PERF_SAMPLE_STREAM_ID) {
1917                 sample->stream_id = *array;
1918                 array--;
1919         }
1920
1921         if (type & PERF_SAMPLE_ID) {
1922                 sample->id = *array;
1923                 array--;
1924         }
1925
1926         if (type & PERF_SAMPLE_TIME) {
1927                 sample->time = *array;
1928                 array--;
1929         }
1930
1931         if (type & PERF_SAMPLE_TID) {
1932                 u.val64 = *array;
1933                 if (swapped) {
1934                         /* undo swap of u64, then swap on individual u32s */
1935                         u.val64 = bswap_64(u.val64);
1936                         u.val32[0] = bswap_32(u.val32[0]);
1937                         u.val32[1] = bswap_32(u.val32[1]);
1938                 }
1939
1940                 sample->pid = u.val32[0];
1941                 sample->tid = u.val32[1];
1942                 array--;
1943         }
1944
1945         return 0;
1946 }
1947
1948 static inline bool overflow(const void *endp, u16 max_size, const void *offset,
1949                             u64 size)
1950 {
1951         return size > max_size || offset + size > endp;
1952 }
1953
1954 #define OVERFLOW_CHECK(offset, size, max_size)                          \
1955         do {                                                            \
1956                 if (overflow(endp, (max_size), (offset), (size)))       \
1957                         return -EFAULT;                                 \
1958         } while (0)
1959
1960 #define OVERFLOW_CHECK_u64(offset) \
1961         OVERFLOW_CHECK(offset, sizeof(u64), sizeof(u64))
1962
1963 int perf_evsel__parse_sample(struct perf_evsel *evsel, union perf_event *event,
1964                              struct perf_sample *data)
1965 {
1966         u64 type = evsel->attr.sample_type;
1967         bool swapped = evsel->needs_swap;
1968         const u64 *array;
1969         u16 max_size = event->header.size;
1970         const void *endp = (void *)event + max_size;
1971         u64 sz;
1972
1973         /*
1974          * used for cross-endian analysis. See git commit 65014ab3
1975          * for why this goofiness is needed.
1976          */
1977         union u64_swap u;
1978
1979         memset(data, 0, sizeof(*data));
1980         data->cpu = data->pid = data->tid = -1;
1981         data->stream_id = data->id = data->time = -1ULL;
1982         data->period = evsel->attr.sample_period;
1983         data->cpumode = event->header.misc & PERF_RECORD_MISC_CPUMODE_MASK;
1984
1985         if (event->header.type != PERF_RECORD_SAMPLE) {
1986                 if (!evsel->attr.sample_id_all)
1987                         return 0;
1988                 return perf_evsel__parse_id_sample(evsel, event, data);
1989         }
1990
1991         array = event->sample.array;
1992
1993         /*
1994          * The evsel's sample_size is based on PERF_SAMPLE_MASK which includes
1995          * up to PERF_SAMPLE_PERIOD.  After that overflow() must be used to
1996          * check the format does not go past the end of the event.
1997          */
1998         if (evsel->sample_size + sizeof(event->header) > event->header.size)
1999                 return -EFAULT;
2000
2001         data->id = -1ULL;
2002         if (type & PERF_SAMPLE_IDENTIFIER) {
2003                 data->id = *array;
2004                 array++;
2005         }
2006
2007         if (type & PERF_SAMPLE_IP) {
2008                 data->ip = *array;
2009                 array++;
2010         }
2011
2012         if (type & PERF_SAMPLE_TID) {
2013                 u.val64 = *array;
2014                 if (swapped) {
2015                         /* undo swap of u64, then swap on individual u32s */
2016                         u.val64 = bswap_64(u.val64);
2017                         u.val32[0] = bswap_32(u.val32[0]);
2018                         u.val32[1] = bswap_32(u.val32[1]);
2019                 }
2020
2021                 data->pid = u.val32[0];
2022                 data->tid = u.val32[1];
2023                 array++;
2024         }
2025
2026         if (type & PERF_SAMPLE_TIME) {
2027                 data->time = *array;
2028                 array++;
2029         }
2030
2031         data->addr = 0;
2032         if (type & PERF_SAMPLE_ADDR) {
2033                 data->addr = *array;
2034                 array++;
2035         }
2036
2037         if (type & PERF_SAMPLE_ID) {
2038                 data->id = *array;
2039                 array++;
2040         }
2041
2042         if (type & PERF_SAMPLE_STREAM_ID) {
2043                 data->stream_id = *array;
2044                 array++;
2045         }
2046
2047         if (type & PERF_SAMPLE_CPU) {
2048
2049                 u.val64 = *array;
2050                 if (swapped) {
2051                         /* undo swap of u64, then swap on individual u32s */
2052                         u.val64 = bswap_64(u.val64);
2053                         u.val32[0] = bswap_32(u.val32[0]);
2054                 }
2055
2056                 data->cpu = u.val32[0];
2057                 array++;
2058         }
2059
2060         if (type & PERF_SAMPLE_PERIOD) {
2061                 data->period = *array;
2062                 array++;
2063         }
2064
2065         if (type & PERF_SAMPLE_READ) {
2066                 u64 read_format = evsel->attr.read_format;
2067
2068                 OVERFLOW_CHECK_u64(array);
2069                 if (read_format & PERF_FORMAT_GROUP)
2070                         data->read.group.nr = *array;
2071                 else
2072                         data->read.one.value = *array;
2073
2074                 array++;
2075
2076                 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
2077                         OVERFLOW_CHECK_u64(array);
2078                         data->read.time_enabled = *array;
2079                         array++;
2080                 }
2081
2082                 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
2083                         OVERFLOW_CHECK_u64(array);
2084                         data->read.time_running = *array;
2085                         array++;
2086                 }
2087
2088                 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
2089                 if (read_format & PERF_FORMAT_GROUP) {
2090                         const u64 max_group_nr = UINT64_MAX /
2091                                         sizeof(struct sample_read_value);
2092
2093                         if (data->read.group.nr > max_group_nr)
2094                                 return -EFAULT;
2095                         sz = data->read.group.nr *
2096                              sizeof(struct sample_read_value);
2097                         OVERFLOW_CHECK(array, sz, max_size);
2098                         data->read.group.values =
2099                                         (struct sample_read_value *)array;
2100                         array = (void *)array + sz;
2101                 } else {
2102                         OVERFLOW_CHECK_u64(array);
2103                         data->read.one.id = *array;
2104                         array++;
2105                 }
2106         }
2107
2108         if (type & PERF_SAMPLE_CALLCHAIN) {
2109                 const u64 max_callchain_nr = UINT64_MAX / sizeof(u64);
2110
2111                 OVERFLOW_CHECK_u64(array);
2112                 data->callchain = (struct ip_callchain *)array++;
2113                 if (data->callchain->nr > max_callchain_nr)
2114                         return -EFAULT;
2115                 sz = data->callchain->nr * sizeof(u64);
2116                 OVERFLOW_CHECK(array, sz, max_size);
2117                 array = (void *)array + sz;
2118         }
2119
2120         if (type & PERF_SAMPLE_RAW) {
2121                 OVERFLOW_CHECK_u64(array);
2122                 u.val64 = *array;
2123                 if (WARN_ONCE(swapped,
2124                               "Endianness of raw data not corrected!\n")) {
2125                         /* undo swap of u64, then swap on individual u32s */
2126                         u.val64 = bswap_64(u.val64);
2127                         u.val32[0] = bswap_32(u.val32[0]);
2128                         u.val32[1] = bswap_32(u.val32[1]);
2129                 }
2130                 data->raw_size = u.val32[0];
2131                 array = (void *)array + sizeof(u32);
2132
2133                 OVERFLOW_CHECK(array, data->raw_size, max_size);
2134                 data->raw_data = (void *)array;
2135                 array = (void *)array + data->raw_size;
2136         }
2137
2138         if (type & PERF_SAMPLE_BRANCH_STACK) {
2139                 const u64 max_branch_nr = UINT64_MAX /
2140                                           sizeof(struct branch_entry);
2141
2142                 OVERFLOW_CHECK_u64(array);
2143                 data->branch_stack = (struct branch_stack *)array++;
2144
2145                 if (data->branch_stack->nr > max_branch_nr)
2146                         return -EFAULT;
2147                 sz = data->branch_stack->nr * sizeof(struct branch_entry);
2148                 OVERFLOW_CHECK(array, sz, max_size);
2149                 array = (void *)array + sz;
2150         }
2151
2152         if (type & PERF_SAMPLE_REGS_USER) {
2153                 OVERFLOW_CHECK_u64(array);
2154                 data->user_regs.abi = *array;
2155                 array++;
2156
2157                 if (data->user_regs.abi) {
2158                         u64 mask = evsel->attr.sample_regs_user;
2159
2160                         sz = hweight_long(mask) * sizeof(u64);
2161                         OVERFLOW_CHECK(array, sz, max_size);
2162                         data->user_regs.mask = mask;
2163                         data->user_regs.regs = (u64 *)array;
2164                         array = (void *)array + sz;
2165                 }
2166         }
2167
2168         if (type & PERF_SAMPLE_STACK_USER) {
2169                 OVERFLOW_CHECK_u64(array);
2170                 sz = *array++;
2171
2172                 data->user_stack.offset = ((char *)(array - 1)
2173                                           - (char *) event);
2174
2175                 if (!sz) {
2176                         data->user_stack.size = 0;
2177                 } else {
2178                         OVERFLOW_CHECK(array, sz, max_size);
2179                         data->user_stack.data = (char *)array;
2180                         array = (void *)array + sz;
2181                         OVERFLOW_CHECK_u64(array);
2182                         data->user_stack.size = *array++;
2183                         if (WARN_ONCE(data->user_stack.size > sz,
2184                                       "user stack dump failure\n"))
2185                                 return -EFAULT;
2186                 }
2187         }
2188
2189         if (type & PERF_SAMPLE_WEIGHT) {
2190                 OVERFLOW_CHECK_u64(array);
2191                 data->weight = *array;
2192                 array++;
2193         }
2194
2195         data->data_src = PERF_MEM_DATA_SRC_NONE;
2196         if (type & PERF_SAMPLE_DATA_SRC) {
2197                 OVERFLOW_CHECK_u64(array);
2198                 data->data_src = *array;
2199                 array++;
2200         }
2201
2202         data->transaction = 0;
2203         if (type & PERF_SAMPLE_TRANSACTION) {
2204                 OVERFLOW_CHECK_u64(array);
2205                 data->transaction = *array;
2206                 array++;
2207         }
2208
2209         data->intr_regs.abi = PERF_SAMPLE_REGS_ABI_NONE;
2210         if (type & PERF_SAMPLE_REGS_INTR) {
2211                 OVERFLOW_CHECK_u64(array);
2212                 data->intr_regs.abi = *array;
2213                 array++;
2214
2215                 if (data->intr_regs.abi != PERF_SAMPLE_REGS_ABI_NONE) {
2216                         u64 mask = evsel->attr.sample_regs_intr;
2217
2218                         sz = hweight_long(mask) * sizeof(u64);
2219                         OVERFLOW_CHECK(array, sz, max_size);
2220                         data->intr_regs.mask = mask;
2221                         data->intr_regs.regs = (u64 *)array;
2222                         array = (void *)array + sz;
2223                 }
2224         }
2225
2226         data->phys_addr = 0;
2227         if (type & PERF_SAMPLE_PHYS_ADDR) {
2228                 data->phys_addr = *array;
2229                 array++;
2230         }
2231
2232         return 0;
2233 }
2234
2235 size_t perf_event__sample_event_size(const struct perf_sample *sample, u64 type,
2236                                      u64 read_format)
2237 {
2238         size_t sz, result = sizeof(struct sample_event);
2239
2240         if (type & PERF_SAMPLE_IDENTIFIER)
2241                 result += sizeof(u64);
2242
2243         if (type & PERF_SAMPLE_IP)
2244                 result += sizeof(u64);
2245
2246         if (type & PERF_SAMPLE_TID)
2247                 result += sizeof(u64);
2248
2249         if (type & PERF_SAMPLE_TIME)
2250                 result += sizeof(u64);
2251
2252         if (type & PERF_SAMPLE_ADDR)
2253                 result += sizeof(u64);
2254
2255         if (type & PERF_SAMPLE_ID)
2256                 result += sizeof(u64);
2257
2258         if (type & PERF_SAMPLE_STREAM_ID)
2259                 result += sizeof(u64);
2260
2261         if (type & PERF_SAMPLE_CPU)
2262                 result += sizeof(u64);
2263
2264         if (type & PERF_SAMPLE_PERIOD)
2265                 result += sizeof(u64);
2266
2267         if (type & PERF_SAMPLE_READ) {
2268                 result += sizeof(u64);
2269                 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED)
2270                         result += sizeof(u64);
2271                 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING)
2272                         result += sizeof(u64);
2273                 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
2274                 if (read_format & PERF_FORMAT_GROUP) {
2275                         sz = sample->read.group.nr *
2276                              sizeof(struct sample_read_value);
2277                         result += sz;
2278                 } else {
2279                         result += sizeof(u64);
2280                 }
2281         }
2282
2283         if (type & PERF_SAMPLE_CALLCHAIN) {
2284                 sz = (sample->callchain->nr + 1) * sizeof(u64);
2285                 result += sz;
2286         }
2287
2288         if (type & PERF_SAMPLE_RAW) {
2289                 result += sizeof(u32);
2290                 result += sample->raw_size;
2291         }
2292
2293         if (type & PERF_SAMPLE_BRANCH_STACK) {
2294                 sz = sample->branch_stack->nr * sizeof(struct branch_entry);
2295                 sz += sizeof(u64);
2296                 result += sz;
2297         }
2298
2299         if (type & PERF_SAMPLE_REGS_USER) {
2300                 if (sample->user_regs.abi) {
2301                         result += sizeof(u64);
2302                         sz = hweight_long(sample->user_regs.mask) * sizeof(u64);
2303                         result += sz;
2304                 } else {
2305                         result += sizeof(u64);
2306                 }
2307         }
2308
2309         if (type & PERF_SAMPLE_STACK_USER) {
2310                 sz = sample->user_stack.size;
2311                 result += sizeof(u64);
2312                 if (sz) {
2313                         result += sz;
2314                         result += sizeof(u64);
2315                 }
2316         }
2317
2318         if (type & PERF_SAMPLE_WEIGHT)
2319                 result += sizeof(u64);
2320
2321         if (type & PERF_SAMPLE_DATA_SRC)
2322                 result += sizeof(u64);
2323
2324         if (type & PERF_SAMPLE_TRANSACTION)
2325                 result += sizeof(u64);
2326
2327         if (type & PERF_SAMPLE_REGS_INTR) {
2328                 if (sample->intr_regs.abi) {
2329                         result += sizeof(u64);
2330                         sz = hweight_long(sample->intr_regs.mask) * sizeof(u64);
2331                         result += sz;
2332                 } else {
2333                         result += sizeof(u64);
2334                 }
2335         }
2336
2337         if (type & PERF_SAMPLE_PHYS_ADDR)
2338                 result += sizeof(u64);
2339
2340         return result;
2341 }
2342
2343 int perf_event__synthesize_sample(union perf_event *event, u64 type,
2344                                   u64 read_format,
2345                                   const struct perf_sample *sample,
2346                                   bool swapped)
2347 {
2348         u64 *array;
2349         size_t sz;
2350         /*
2351          * used for cross-endian analysis. See git commit 65014ab3
2352          * for why this goofiness is needed.
2353          */
2354         union u64_swap u;
2355
2356         array = event->sample.array;
2357
2358         if (type & PERF_SAMPLE_IDENTIFIER) {
2359                 *array = sample->id;
2360                 array++;
2361         }
2362
2363         if (type & PERF_SAMPLE_IP) {
2364                 *array = sample->ip;
2365                 array++;
2366         }
2367
2368         if (type & PERF_SAMPLE_TID) {
2369                 u.val32[0] = sample->pid;
2370                 u.val32[1] = sample->tid;
2371                 if (swapped) {
2372                         /*
2373                          * Inverse of what is done in perf_evsel__parse_sample
2374                          */
2375                         u.val32[0] = bswap_32(u.val32[0]);
2376                         u.val32[1] = bswap_32(u.val32[1]);
2377                         u.val64 = bswap_64(u.val64);
2378                 }
2379
2380                 *array = u.val64;
2381                 array++;
2382         }
2383
2384         if (type & PERF_SAMPLE_TIME) {
2385                 *array = sample->time;
2386                 array++;
2387         }
2388
2389         if (type & PERF_SAMPLE_ADDR) {
2390                 *array = sample->addr;
2391                 array++;
2392         }
2393
2394         if (type & PERF_SAMPLE_ID) {
2395                 *array = sample->id;
2396                 array++;
2397         }
2398
2399         if (type & PERF_SAMPLE_STREAM_ID) {
2400                 *array = sample->stream_id;
2401                 array++;
2402         }
2403
2404         if (type & PERF_SAMPLE_CPU) {
2405                 u.val32[0] = sample->cpu;
2406                 if (swapped) {
2407                         /*
2408                          * Inverse of what is done in perf_evsel__parse_sample
2409                          */
2410                         u.val32[0] = bswap_32(u.val32[0]);
2411                         u.val64 = bswap_64(u.val64);
2412                 }
2413                 *array = u.val64;
2414                 array++;
2415         }
2416
2417         if (type & PERF_SAMPLE_PERIOD) {
2418                 *array = sample->period;
2419                 array++;
2420         }
2421
2422         if (type & PERF_SAMPLE_READ) {
2423                 if (read_format & PERF_FORMAT_GROUP)
2424                         *array = sample->read.group.nr;
2425                 else
2426                         *array = sample->read.one.value;
2427                 array++;
2428
2429                 if (read_format & PERF_FORMAT_TOTAL_TIME_ENABLED) {
2430                         *array = sample->read.time_enabled;
2431                         array++;
2432                 }
2433
2434                 if (read_format & PERF_FORMAT_TOTAL_TIME_RUNNING) {
2435                         *array = sample->read.time_running;
2436                         array++;
2437                 }
2438
2439                 /* PERF_FORMAT_ID is forced for PERF_SAMPLE_READ */
2440                 if (read_format & PERF_FORMAT_GROUP) {
2441                         sz = sample->read.group.nr *
2442                              sizeof(struct sample_read_value);
2443                         memcpy(array, sample->read.group.values, sz);
2444                         array = (void *)array + sz;
2445                 } else {
2446                         *array = sample->read.one.id;
2447                         array++;
2448                 }
2449         }
2450
2451         if (type & PERF_SAMPLE_CALLCHAIN) {
2452                 sz = (sample->callchain->nr + 1) * sizeof(u64);
2453                 memcpy(array, sample->callchain, sz);
2454                 array = (void *)array + sz;
2455         }
2456
2457         if (type & PERF_SAMPLE_RAW) {
2458                 u.val32[0] = sample->raw_size;
2459                 if (WARN_ONCE(swapped,
2460                               "Endianness of raw data not corrected!\n")) {
2461                         /*
2462                          * Inverse of what is done in perf_evsel__parse_sample
2463                          */
2464                         u.val32[0] = bswap_32(u.val32[0]);
2465                         u.val32[1] = bswap_32(u.val32[1]);
2466                         u.val64 = bswap_64(u.val64);
2467                 }
2468                 *array = u.val64;
2469                 array = (void *)array + sizeof(u32);
2470
2471                 memcpy(array, sample->raw_data, sample->raw_size);
2472                 array = (void *)array + sample->raw_size;
2473         }
2474
2475         if (type & PERF_SAMPLE_BRANCH_STACK) {
2476                 sz = sample->branch_stack->nr * sizeof(struct branch_entry);
2477                 sz += sizeof(u64);
2478                 memcpy(array, sample->branch_stack, sz);
2479                 array = (void *)array + sz;
2480         }
2481
2482         if (type & PERF_SAMPLE_REGS_USER) {
2483                 if (sample->user_regs.abi) {
2484                         *array++ = sample->user_regs.abi;
2485                         sz = hweight_long(sample->user_regs.mask) * sizeof(u64);
2486                         memcpy(array, sample->user_regs.regs, sz);
2487                         array = (void *)array + sz;
2488                 } else {
2489                         *array++ = 0;
2490                 }
2491         }
2492
2493         if (type & PERF_SAMPLE_STACK_USER) {
2494                 sz = sample->user_stack.size;
2495                 *array++ = sz;
2496                 if (sz) {
2497                         memcpy(array, sample->user_stack.data, sz);
2498                         array = (void *)array + sz;
2499                         *array++ = sz;
2500                 }
2501         }
2502
2503         if (type & PERF_SAMPLE_WEIGHT) {
2504                 *array = sample->weight;
2505                 array++;
2506         }
2507
2508         if (type & PERF_SAMPLE_DATA_SRC) {
2509                 *array = sample->data_src;
2510                 array++;
2511         }
2512
2513         if (type & PERF_SAMPLE_TRANSACTION) {
2514                 *array = sample->transaction;
2515                 array++;
2516         }
2517
2518         if (type & PERF_SAMPLE_REGS_INTR) {
2519                 if (sample->intr_regs.abi) {
2520                         *array++ = sample->intr_regs.abi;
2521                         sz = hweight_long(sample->intr_regs.mask) * sizeof(u64);
2522                         memcpy(array, sample->intr_regs.regs, sz);
2523                         array = (void *)array + sz;
2524                 } else {
2525                         *array++ = 0;
2526                 }
2527         }
2528
2529         if (type & PERF_SAMPLE_PHYS_ADDR) {
2530                 *array = sample->phys_addr;
2531                 array++;
2532         }
2533
2534         return 0;
2535 }
2536
2537 struct format_field *perf_evsel__field(struct perf_evsel *evsel, const char *name)
2538 {
2539         return pevent_find_field(evsel->tp_format, name);
2540 }
2541
2542 void *perf_evsel__rawptr(struct perf_evsel *evsel, struct perf_sample *sample,
2543                          const char *name)
2544 {
2545         struct format_field *field = perf_evsel__field(evsel, name);
2546         int offset;
2547
2548         if (!field)
2549                 return NULL;
2550
2551         offset = field->offset;
2552
2553         if (field->flags & FIELD_IS_DYNAMIC) {
2554                 offset = *(int *)(sample->raw_data + field->offset);
2555                 offset &= 0xffff;
2556         }
2557
2558         return sample->raw_data + offset;
2559 }
2560
2561 u64 format_field__intval(struct format_field *field, struct perf_sample *sample,
2562                          bool needs_swap)
2563 {
2564         u64 value;
2565         void *ptr = sample->raw_data + field->offset;
2566
2567         switch (field->size) {
2568         case 1:
2569                 return *(u8 *)ptr;
2570         case 2:
2571                 value = *(u16 *)ptr;
2572                 break;
2573         case 4:
2574                 value = *(u32 *)ptr;
2575                 break;
2576         case 8:
2577                 memcpy(&value, ptr, sizeof(u64));
2578                 break;
2579         default:
2580                 return 0;
2581         }
2582
2583         if (!needs_swap)
2584                 return value;
2585
2586         switch (field->size) {
2587         case 2:
2588                 return bswap_16(value);
2589         case 4:
2590                 return bswap_32(value);
2591         case 8:
2592                 return bswap_64(value);
2593         default:
2594                 return 0;
2595         }
2596
2597         return 0;
2598 }
2599
2600 u64 perf_evsel__intval(struct perf_evsel *evsel, struct perf_sample *sample,
2601                        const char *name)
2602 {
2603         struct format_field *field = perf_evsel__field(evsel, name);
2604
2605         if (!field)
2606                 return 0;
2607
2608         return field ? format_field__intval(field, sample, evsel->needs_swap) : 0;
2609 }
2610
2611 bool perf_evsel__fallback(struct perf_evsel *evsel, int err,
2612                           char *msg, size_t msgsize)
2613 {
2614         int paranoid;
2615
2616         if ((err == ENOENT || err == ENXIO || err == ENODEV) &&
2617             evsel->attr.type   == PERF_TYPE_HARDWARE &&
2618             evsel->attr.config == PERF_COUNT_HW_CPU_CYCLES) {
2619                 /*
2620                  * If it's cycles then fall back to hrtimer based
2621                  * cpu-clock-tick sw counter, which is always available even if
2622                  * no PMU support.
2623                  *
2624                  * PPC returns ENXIO until 2.6.37 (behavior changed with commit
2625                  * b0a873e).
2626                  */
2627                 scnprintf(msg, msgsize, "%s",
2628 "The cycles event is not supported, trying to fall back to cpu-clock-ticks");
2629
2630                 evsel->attr.type   = PERF_TYPE_SOFTWARE;
2631                 evsel->attr.config = PERF_COUNT_SW_CPU_CLOCK;
2632
2633                 zfree(&evsel->name);
2634                 return true;
2635         } else if (err == EACCES && !evsel->attr.exclude_kernel &&
2636                    (paranoid = perf_event_paranoid()) > 1) {
2637                 const char *name = perf_evsel__name(evsel);
2638                 char *new_name;
2639
2640                 if (asprintf(&new_name, "%s%su", name, strchr(name, ':') ? "" : ":") < 0)
2641                         return false;
2642
2643                 if (evsel->name)
2644                         free(evsel->name);
2645                 evsel->name = new_name;
2646                 scnprintf(msg, msgsize,
2647 "kernel.perf_event_paranoid=%d, trying to fall back to excluding kernel samples", paranoid);
2648                 evsel->attr.exclude_kernel = 1;
2649
2650                 return true;
2651         }
2652
2653         return false;
2654 }
2655
2656 static bool find_process(const char *name)
2657 {
2658         size_t len = strlen(name);
2659         DIR *dir;
2660         struct dirent *d;
2661         int ret = -1;
2662
2663         dir = opendir(procfs__mountpoint());
2664         if (!dir)
2665                 return false;
2666
2667         /* Walk through the directory. */
2668         while (ret && (d = readdir(dir)) != NULL) {
2669                 char path[PATH_MAX];
2670                 char *data;
2671                 size_t size;
2672
2673                 if ((d->d_type != DT_DIR) ||
2674                      !strcmp(".", d->d_name) ||
2675                      !strcmp("..", d->d_name))
2676                         continue;
2677
2678                 scnprintf(path, sizeof(path), "%s/%s/comm",
2679                           procfs__mountpoint(), d->d_name);
2680
2681                 if (filename__read_str(path, &data, &size))
2682                         continue;
2683
2684                 ret = strncmp(name, data, len);
2685                 free(data);
2686         }
2687
2688         closedir(dir);
2689         return ret ? false : true;
2690 }
2691
2692 int perf_evsel__open_strerror(struct perf_evsel *evsel, struct target *target,
2693                               int err, char *msg, size_t size)
2694 {
2695         char sbuf[STRERR_BUFSIZE];
2696         int printed = 0;
2697
2698         switch (err) {
2699         case EPERM:
2700         case EACCES:
2701                 if (err == EPERM)
2702                         printed = scnprintf(msg, size,
2703                                 "No permission to enable %s event.\n\n",
2704                                 perf_evsel__name(evsel));
2705
2706                 return scnprintf(msg + printed, size - printed,
2707                  "You may not have permission to collect %sstats.\n\n"
2708                  "Consider tweaking /proc/sys/kernel/perf_event_paranoid,\n"
2709                  "which controls use of the performance events system by\n"
2710                  "unprivileged users (without CAP_SYS_ADMIN).\n\n"
2711                  "The current value is %d:\n\n"
2712                  "  -1: Allow use of (almost) all events by all users\n"
2713                  "      Ignore mlock limit after perf_event_mlock_kb without CAP_IPC_LOCK\n"
2714                  ">= 0: Disallow ftrace function tracepoint by users without CAP_SYS_ADMIN\n"
2715                  "      Disallow raw tracepoint access by users without CAP_SYS_ADMIN\n"
2716                  ">= 1: Disallow CPU event access by users without CAP_SYS_ADMIN\n"
2717                  ">= 2: Disallow kernel profiling by users without CAP_SYS_ADMIN\n\n"
2718                  "To make this setting permanent, edit /etc/sysctl.conf too, e.g.:\n\n"
2719                  "      kernel.perf_event_paranoid = -1\n" ,
2720                                  target->system_wide ? "system-wide " : "",
2721                                  perf_event_paranoid());
2722         case ENOENT:
2723                 return scnprintf(msg, size, "The %s event is not supported.",
2724                                  perf_evsel__name(evsel));
2725         case EMFILE:
2726                 return scnprintf(msg, size, "%s",
2727                          "Too many events are opened.\n"
2728                          "Probably the maximum number of open file descriptors has been reached.\n"
2729                          "Hint: Try again after reducing the number of events.\n"
2730                          "Hint: Try increasing the limit with 'ulimit -n <limit>'");
2731         case ENOMEM:
2732                 if ((evsel->attr.sample_type & PERF_SAMPLE_CALLCHAIN) != 0 &&
2733                     access("/proc/sys/kernel/perf_event_max_stack", F_OK) == 0)
2734                         return scnprintf(msg, size,
2735                                          "Not enough memory to setup event with callchain.\n"
2736                                          "Hint: Try tweaking /proc/sys/kernel/perf_event_max_stack\n"
2737                                          "Hint: Current value: %d", sysctl_perf_event_max_stack);
2738                 break;
2739         case ENODEV:
2740                 if (target->cpu_list)
2741                         return scnprintf(msg, size, "%s",
2742          "No such device - did you specify an out-of-range profile CPU?");
2743                 break;
2744         case EOPNOTSUPP:
2745                 if (evsel->attr.sample_period != 0)
2746                         return scnprintf(msg, size, "%s",
2747         "PMU Hardware doesn't support sampling/overflow-interrupts.");
2748                 if (evsel->attr.precise_ip)
2749                         return scnprintf(msg, size, "%s",
2750         "\'precise\' request may not be supported. Try removing 'p' modifier.");
2751 #if defined(__i386__) || defined(__x86_64__)
2752                 if (evsel->attr.type == PERF_TYPE_HARDWARE)
2753                         return scnprintf(msg, size, "%s",
2754         "No hardware sampling interrupt available.\n"
2755         "No APIC? If so then you can boot the kernel with the \"lapic\" boot parameter to force-enable it.");
2756 #endif
2757                 break;
2758         case EBUSY:
2759                 if (find_process("oprofiled"))
2760                         return scnprintf(msg, size,
2761         "The PMU counters are busy/taken by another profiler.\n"
2762         "We found oprofile daemon running, please stop it and try again.");
2763                 break;
2764         case EINVAL:
2765                 if (evsel->attr.write_backward && perf_missing_features.write_backward)
2766                         return scnprintf(msg, size, "Reading from overwrite event is not supported by this kernel.");
2767                 if (perf_missing_features.clockid)
2768                         return scnprintf(msg, size, "clockid feature not supported.");
2769                 if (perf_missing_features.clockid_wrong)
2770                         return scnprintf(msg, size, "wrong clockid (%d).", clockid);
2771                 break;
2772         default:
2773                 break;
2774         }
2775
2776         return scnprintf(msg, size,
2777         "The sys_perf_event_open() syscall returned with %d (%s) for event (%s).\n"
2778         "/bin/dmesg may provide additional information.\n"
2779         "No CONFIG_PERF_EVENTS=y kernel support configured?",
2780                          err, str_error_r(err, sbuf, sizeof(sbuf)),
2781                          perf_evsel__name(evsel));
2782 }
2783
2784 char *perf_evsel__env_arch(struct perf_evsel *evsel)
2785 {
2786         if (evsel && evsel->evlist && evsel->evlist->env)
2787                 return evsel->evlist->env->arch;
2788         return NULL;
2789 }
2790
2791 char *perf_evsel__env_cpuid(struct perf_evsel *evsel)
2792 {
2793         if (evsel && evsel->evlist && evsel->evlist->env)
2794                 return evsel->evlist->env->cpuid;
2795         return NULL;
2796 }