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