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