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