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