1 // SPDX-License-Identifier: GPL-2.0 2 #include <inttypes.h> 3 #include <sys/types.h> 4 #include <sys/stat.h> 5 #include <unistd.h> 6 #include "builtin.h" 7 #include "perf.h" 8 9 #include <subcmd/parse-options.h> 10 #include "util/auxtrace.h" 11 #include "util/trace-event.h" 12 #include "util/tool.h" 13 #include "util/session.h" 14 #include "util/data.h" 15 #include "util/map_symbol.h" 16 #include "util/mem-events.h" 17 #include "util/debug.h" 18 #include "util/dso.h" 19 #include "util/map.h" 20 #include "util/symbol.h" 21 #include "util/pmus.h" 22 #include "util/sample.h" 23 #include "util/sort.h" 24 #include "util/string2.h" 25 #include "util/util.h" 26 #include <linux/err.h> 27 28 #define MEM_OPERATION_LOAD 0x1 29 #define MEM_OPERATION_STORE 0x2 30 31 struct perf_mem { 32 struct perf_tool tool; 33 const char *input_name; 34 const char *sort_key; 35 bool hide_unresolved; 36 bool dump_raw; 37 bool force; 38 bool phys_addr; 39 bool data_page_size; 40 bool all_kernel; 41 bool all_user; 42 bool data_type; 43 int operation; 44 const char *cpu_list; 45 DECLARE_BITMAP(cpu_bitmap, MAX_NR_CPUS); 46 }; 47 48 static int parse_record_events(const struct option *opt, 49 const char *str, int unset __maybe_unused) 50 { 51 struct perf_mem *mem = (struct perf_mem *)opt->value; 52 struct perf_pmu *pmu; 53 54 pmu = perf_mem_events_find_pmu(); 55 if (!pmu) { 56 pr_err("failed: there is no PMU that supports perf mem\n"); 57 exit(-1); 58 } 59 60 if (!strcmp(str, "list")) { 61 perf_pmu__mem_events_list(pmu); 62 exit(0); 63 } 64 if (perf_pmu__mem_events_parse(pmu, str)) 65 exit(-1); 66 67 mem->operation = 0; 68 return 0; 69 } 70 71 static int __cmd_record(int argc, const char **argv, struct perf_mem *mem, 72 const struct option *options) 73 { 74 int rec_argc, i = 0, j; 75 int start, end; 76 const char **rec_argv; 77 int ret; 78 struct perf_mem_event *e; 79 struct perf_pmu *pmu; 80 const char * const record_usage[] = { 81 "perf mem record [<options>] [<command>]", 82 "perf mem record [<options>] -- <command> [<options>]", 83 NULL 84 }; 85 86 pmu = perf_mem_events_find_pmu(); 87 if (!pmu) { 88 pr_err("failed: no PMU supports the memory events\n"); 89 return -1; 90 } 91 92 if (perf_pmu__mem_events_init()) { 93 pr_err("failed: memory events not supported\n"); 94 return -1; 95 } 96 97 argc = parse_options(argc, argv, options, record_usage, 98 PARSE_OPT_KEEP_UNKNOWN); 99 100 /* Max number of arguments multiplied by number of PMUs that can support them. */ 101 rec_argc = argc + 9 * (perf_pmu__mem_events_num_mem_pmus(pmu) + 1); 102 103 if (mem->cpu_list) 104 rec_argc += 2; 105 106 rec_argv = calloc(rec_argc + 1, sizeof(char *)); 107 if (!rec_argv) 108 return -1; 109 110 rec_argv[i++] = "record"; 111 112 e = perf_pmu__mem_events_ptr(pmu, PERF_MEM_EVENTS__LOAD_STORE); 113 114 /* 115 * The load and store operations are required, use the event 116 * PERF_MEM_EVENTS__LOAD_STORE if it is supported. 117 */ 118 if (e->tag && 119 (mem->operation & MEM_OPERATION_LOAD) && 120 (mem->operation & MEM_OPERATION_STORE)) { 121 perf_mem_record[PERF_MEM_EVENTS__LOAD_STORE] = true; 122 rec_argv[i++] = "-W"; 123 } else { 124 if (mem->operation & MEM_OPERATION_LOAD) 125 perf_mem_record[PERF_MEM_EVENTS__LOAD] = true; 126 127 if (mem->operation & MEM_OPERATION_STORE) 128 perf_mem_record[PERF_MEM_EVENTS__STORE] = true; 129 } 130 131 if (perf_mem_record[PERF_MEM_EVENTS__LOAD]) 132 rec_argv[i++] = "-W"; 133 134 rec_argv[i++] = "-d"; 135 136 if (mem->phys_addr) 137 rec_argv[i++] = "--phys-data"; 138 139 if (mem->data_page_size) 140 rec_argv[i++] = "--data-page-size"; 141 142 start = i; 143 ret = perf_mem_events__record_args(rec_argv, &i); 144 if (ret) 145 goto out; 146 end = i; 147 148 if (mem->all_user) 149 rec_argv[i++] = "--all-user"; 150 151 if (mem->all_kernel) 152 rec_argv[i++] = "--all-kernel"; 153 154 if (mem->cpu_list) { 155 rec_argv[i++] = "-C"; 156 rec_argv[i++] = mem->cpu_list; 157 } 158 159 for (j = 0; j < argc; j++, i++) 160 rec_argv[i] = argv[j]; 161 162 if (verbose > 0) { 163 pr_debug("calling: record "); 164 165 for (j = start; j < end; j++) 166 pr_debug("%s ", rec_argv[j]); 167 168 pr_debug("\n"); 169 } 170 171 ret = cmd_record(i, rec_argv); 172 out: 173 free(rec_argv); 174 return ret; 175 } 176 177 static int 178 dump_raw_samples(const struct perf_tool *tool, 179 union perf_event *event, 180 struct perf_sample *sample, 181 struct machine *machine) 182 { 183 struct perf_mem *mem = container_of(tool, struct perf_mem, tool); 184 struct addr_location al; 185 const char *fmt, *field_sep; 186 char str[PAGE_SIZE_NAME_LEN]; 187 struct dso *dso = NULL; 188 189 addr_location__init(&al); 190 if (machine__resolve(machine, &al, sample) < 0) { 191 fprintf(stderr, "problem processing %d event, skipping it.\n", 192 event->header.type); 193 addr_location__exit(&al); 194 return -1; 195 } 196 197 if (al.filtered || (mem->hide_unresolved && al.sym == NULL)) 198 goto out_put; 199 200 if (al.map != NULL) { 201 dso = map__dso(al.map); 202 if (dso) 203 dso__set_hit(dso); 204 } 205 206 field_sep = symbol_conf.field_sep; 207 if (field_sep) { 208 fmt = "%d%s%d%s0x%"PRIx64"%s0x%"PRIx64"%s"; 209 } else { 210 fmt = "%5d%s%5d%s0x%016"PRIx64"%s0x016%"PRIx64"%s"; 211 symbol_conf.field_sep = " "; 212 } 213 printf(fmt, 214 sample->pid, 215 symbol_conf.field_sep, 216 sample->tid, 217 symbol_conf.field_sep, 218 sample->ip, 219 symbol_conf.field_sep, 220 sample->addr, 221 symbol_conf.field_sep); 222 223 if (mem->phys_addr) { 224 printf("0x%016"PRIx64"%s", 225 sample->phys_addr, 226 symbol_conf.field_sep); 227 } 228 229 if (mem->data_page_size) { 230 printf("%s%s", 231 get_page_size_name(sample->data_page_size, str), 232 symbol_conf.field_sep); 233 } 234 235 if (field_sep) 236 fmt = "%"PRIu64"%s0x%"PRIx64"%s%s:%s\n"; 237 else 238 fmt = "%5"PRIu64"%s0x%06"PRIx64"%s%s:%s\n"; 239 240 printf(fmt, 241 sample->weight, 242 symbol_conf.field_sep, 243 sample->data_src, 244 symbol_conf.field_sep, 245 dso ? dso__long_name(dso) : "???", 246 al.sym ? al.sym->name : "???"); 247 out_put: 248 addr_location__exit(&al); 249 return 0; 250 } 251 252 static int process_sample_event(const struct perf_tool *tool, 253 union perf_event *event, 254 struct perf_sample *sample, 255 struct evsel *evsel __maybe_unused, 256 struct machine *machine) 257 { 258 return dump_raw_samples(tool, event, sample, machine); 259 } 260 261 static int report_raw_events(struct perf_mem *mem) 262 { 263 struct itrace_synth_opts itrace_synth_opts = { 264 .set = true, 265 .mem = true, /* Only enable memory event */ 266 .default_no_sample = true, 267 }; 268 269 struct perf_data data = { 270 .path = input_name, 271 .mode = PERF_DATA_MODE_READ, 272 .force = mem->force, 273 }; 274 int ret; 275 struct perf_session *session; 276 277 perf_tool__init(&mem->tool, /*ordered_events=*/true); 278 mem->tool.sample = process_sample_event; 279 mem->tool.mmap = perf_event__process_mmap; 280 mem->tool.mmap2 = perf_event__process_mmap2; 281 mem->tool.comm = perf_event__process_comm; 282 mem->tool.lost = perf_event__process_lost; 283 mem->tool.fork = perf_event__process_fork; 284 mem->tool.attr = perf_event__process_attr; 285 mem->tool.build_id = perf_event__process_build_id; 286 mem->tool.namespaces = perf_event__process_namespaces; 287 mem->tool.auxtrace_info = perf_event__process_auxtrace_info; 288 mem->tool.auxtrace = perf_event__process_auxtrace; 289 mem->tool.auxtrace_error = perf_event__process_auxtrace_error; 290 291 session = perf_session__new(&data, &mem->tool); 292 293 if (IS_ERR(session)) 294 return PTR_ERR(session); 295 296 session->itrace_synth_opts = &itrace_synth_opts; 297 298 if (mem->cpu_list) { 299 ret = perf_session__cpu_bitmap(session, mem->cpu_list, 300 mem->cpu_bitmap); 301 if (ret < 0) 302 goto out_delete; 303 } 304 305 ret = symbol__init(&session->header.env); 306 if (ret < 0) 307 goto out_delete; 308 309 printf("# PID, TID, IP, ADDR, "); 310 311 if (mem->phys_addr) 312 printf("PHYS ADDR, "); 313 314 if (mem->data_page_size) 315 printf("DATA PAGE SIZE, "); 316 317 printf("LOCAL WEIGHT, DSRC, SYMBOL\n"); 318 319 ret = perf_session__process_events(session); 320 321 out_delete: 322 perf_session__delete(session); 323 return ret; 324 } 325 326 static char *get_sort_order(struct perf_mem *mem) 327 { 328 bool has_extra_options = (mem->phys_addr | mem->data_page_size) ? true : false; 329 char sort[128]; 330 331 if (mem->sort_key) 332 scnprintf(sort, sizeof(sort), "--sort=%s", mem->sort_key); 333 else if (mem->data_type) 334 strcpy(sort, "--sort=mem,snoop,tlb,type"); 335 /* 336 * there is no weight (cost) associated with stores, so don't print 337 * the column 338 */ 339 else if (!(mem->operation & MEM_OPERATION_LOAD)) { 340 strcpy(sort, "--sort=mem,sym,dso,symbol_daddr," 341 "dso_daddr,tlb,locked"); 342 } else if (has_extra_options) { 343 strcpy(sort, "--sort=local_weight,mem,sym,dso,symbol_daddr," 344 "dso_daddr,snoop,tlb,locked,blocked"); 345 } else 346 return NULL; 347 348 if (mem->phys_addr) 349 strcat(sort, ",phys_daddr"); 350 351 if (mem->data_page_size) 352 strcat(sort, ",data_page_size"); 353 354 /* make sure it has 'type' sort key even -s option is used */ 355 if (mem->data_type && !strstr(sort, "type")) 356 strcat(sort, ",type"); 357 358 return strdup(sort); 359 } 360 361 static int __cmd_report(int argc, const char **argv, struct perf_mem *mem, 362 const struct option *options) 363 { 364 const char **rep_argv; 365 int ret, i = 0, j, rep_argc; 366 char *new_sort_order; 367 const char * const report_usage[] = { 368 "perf mem report [<options>]", 369 NULL 370 }; 371 372 argc = parse_options(argc, argv, options, report_usage, 373 PARSE_OPT_KEEP_UNKNOWN); 374 375 if (mem->dump_raw) 376 return report_raw_events(mem); 377 378 rep_argc = argc + 3; 379 rep_argv = calloc(rep_argc + 1, sizeof(char *)); 380 if (!rep_argv) 381 return -1; 382 383 rep_argv[i++] = "report"; 384 rep_argv[i++] = "--mem-mode"; 385 rep_argv[i++] = "-n"; /* display number of samples */ 386 387 new_sort_order = get_sort_order(mem); 388 if (new_sort_order) 389 rep_argv[i++] = new_sort_order; 390 391 for (j = 0; j < argc; j++, i++) 392 rep_argv[i] = argv[j]; 393 394 ret = cmd_report(i, rep_argv); 395 free(new_sort_order); 396 free(rep_argv); 397 return ret; 398 } 399 400 struct mem_mode { 401 const char *name; 402 int mode; 403 }; 404 405 #define MEM_OPT(n, m) \ 406 { .name = n, .mode = (m) } 407 408 #define MEM_END { .name = NULL } 409 410 static const struct mem_mode mem_modes[]={ 411 MEM_OPT("load", MEM_OPERATION_LOAD), 412 MEM_OPT("store", MEM_OPERATION_STORE), 413 MEM_END 414 }; 415 416 static int 417 parse_mem_ops(const struct option *opt, const char *str, int unset) 418 { 419 int *mode = (int *)opt->value; 420 const struct mem_mode *m; 421 char *s, *os = NULL, *p; 422 int ret = -1; 423 424 if (unset) 425 return 0; 426 427 /* str may be NULL in case no arg is passed to -t */ 428 if (str) { 429 /* because str is read-only */ 430 s = os = strdup(str); 431 if (!s) 432 return -1; 433 434 /* reset mode */ 435 *mode = 0; 436 437 for (;;) { 438 p = strchr(s, ','); 439 if (p) 440 *p = '\0'; 441 442 for (m = mem_modes; m->name; m++) { 443 if (!strcasecmp(s, m->name)) 444 break; 445 } 446 if (!m->name) { 447 fprintf(stderr, "unknown sampling op %s," 448 " check man page\n", s); 449 goto error; 450 } 451 452 *mode |= m->mode; 453 454 if (!p) 455 break; 456 457 s = p + 1; 458 } 459 } 460 ret = 0; 461 462 if (*mode == 0) 463 *mode = MEM_OPERATION_LOAD; 464 error: 465 free(os); 466 return ret; 467 } 468 469 int cmd_mem(int argc, const char **argv) 470 { 471 struct stat st; 472 struct perf_mem mem = { 473 .input_name = "perf.data", 474 /* 475 * default to both load an store sampling 476 */ 477 .operation = MEM_OPERATION_LOAD | MEM_OPERATION_STORE, 478 }; 479 char *sort_order_help = sort_help("sort by key(s):", SORT_MODE__MEMORY); 480 const struct option mem_options[] = { 481 OPT_CALLBACK('t', "type", &mem.operation, 482 "type", "memory operations(load,store) Default load,store", 483 parse_mem_ops), 484 OPT_STRING('C', "cpu", &mem.cpu_list, "cpu", 485 "list of cpus to profile"), 486 OPT_BOOLEAN('f', "force", &mem.force, "don't complain, do it"), 487 OPT_INCR('v', "verbose", &verbose, 488 "be more verbose (show counter open errors, etc)"), 489 OPT_BOOLEAN('p', "phys-data", &mem.phys_addr, "Record/Report sample physical addresses"), 490 OPT_BOOLEAN(0, "data-page-size", &mem.data_page_size, "Record/Report sample data address page size"), 491 OPT_END() 492 }; 493 const struct option record_options[] = { 494 OPT_CALLBACK('e', "event", &mem, "event", 495 "event selector. use 'perf mem record -e list' to list available events", 496 parse_record_events), 497 OPT_UINTEGER(0, "ldlat", &perf_mem_events__loads_ldlat, "mem-loads latency"), 498 OPT_BOOLEAN('U', "all-user", &mem.all_user, "collect only user level data"), 499 OPT_BOOLEAN('K', "all-kernel", &mem.all_kernel, "collect only kernel level data"), 500 OPT_PARENT(mem_options) 501 }; 502 const struct option report_options[] = { 503 OPT_BOOLEAN('D', "dump-raw-samples", &mem.dump_raw, 504 "dump raw samples in ASCII"), 505 OPT_BOOLEAN('U', "hide-unresolved", &mem.hide_unresolved, 506 "Only display entries resolved to a symbol"), 507 OPT_STRING('i', "input", &input_name, "file", 508 "input file name"), 509 OPT_STRING_NOEMPTY('x', "field-separator", &symbol_conf.field_sep, 510 "separator", 511 "separator for columns, no spaces will be added" 512 " between columns '.' is reserved."), 513 OPT_STRING('s', "sort", &mem.sort_key, "key[,key2...]", 514 sort_order_help), 515 OPT_BOOLEAN('T', "type-profile", &mem.data_type, 516 "Show data-type profile result"), 517 OPT_PARENT(mem_options) 518 }; 519 const char *const mem_subcommands[] = { "record", "report", NULL }; 520 const char *mem_usage[] = { 521 NULL, 522 NULL 523 }; 524 525 argc = parse_options_subcommand(argc, argv, mem_options, mem_subcommands, 526 mem_usage, PARSE_OPT_STOP_AT_NON_OPTION); 527 528 if (!argc || !(strncmp(argv[0], "rec", 3) || mem.operation)) 529 usage_with_options(mem_usage, mem_options); 530 531 if (!mem.input_name || !strlen(mem.input_name)) { 532 if (!fstat(STDIN_FILENO, &st) && S_ISFIFO(st.st_mode)) 533 mem.input_name = "-"; 534 else 535 mem.input_name = "perf.data"; 536 } 537 538 if (strlen(argv[0]) > 2 && strstarts("record", argv[0])) 539 return __cmd_record(argc, argv, &mem, record_options); 540 else if (strlen(argv[0]) > 2 && strstarts("report", argv[0])) 541 return __cmd_report(argc, argv, &mem, report_options); 542 else 543 usage_with_options(mem_usage, mem_options); 544 545 /* free usage string allocated by parse_options_subcommand */ 546 free((void *)mem_usage[0]); 547 548 return 0; 549 } 550