| 1 | #include "builtin.h" |
| 2 | #include "perf.h" |
| 3 | |
| 4 | #include <subcmd/parse-options.h> |
| 5 | #include "util/trace-event.h" |
| 6 | #include "util/tool.h" |
| 7 | #include "util/session.h" |
| 8 | #include "util/data.h" |
| 9 | #include "util/mem-events.h" |
| 10 | #include "util/debug.h" |
| 11 | |
| 12 | #define MEM_OPERATION_LOAD 0x1 |
| 13 | #define MEM_OPERATION_STORE 0x2 |
| 14 | |
| 15 | struct perf_mem { |
| 16 | struct perf_tool tool; |
| 17 | char const *input_name; |
| 18 | bool hide_unresolved; |
| 19 | bool dump_raw; |
| 20 | bool force; |
| 21 | int operation; |
| 22 | const char *cpu_list; |
| 23 | DECLARE_BITMAP(cpu_bitmap, MAX_NR_CPUS); |
| 24 | }; |
| 25 | |
| 26 | static int parse_record_events(const struct option *opt, |
| 27 | const char *str, int unset __maybe_unused) |
| 28 | { |
| 29 | struct perf_mem *mem = *(struct perf_mem **)opt->value; |
| 30 | int j; |
| 31 | |
| 32 | if (strcmp(str, "list")) { |
| 33 | if (!perf_mem_events__parse(str)) { |
| 34 | mem->operation = 0; |
| 35 | return 0; |
| 36 | } |
| 37 | exit(-1); |
| 38 | } |
| 39 | |
| 40 | for (j = 0; j < PERF_MEM_EVENTS__MAX; j++) { |
| 41 | struct perf_mem_event *e = &perf_mem_events[j]; |
| 42 | |
| 43 | fprintf(stderr, "%-13s%-*s%s\n", |
| 44 | e->tag, |
| 45 | verbose ? 25 : 0, |
| 46 | verbose ? perf_mem_events__name(j) : "", |
| 47 | e->supported ? ": available" : ""); |
| 48 | } |
| 49 | exit(0); |
| 50 | } |
| 51 | |
| 52 | static const char * const __usage[] = { |
| 53 | "perf mem record [<options>] [<command>]", |
| 54 | "perf mem record [<options>] -- <command> [<options>]", |
| 55 | NULL |
| 56 | }; |
| 57 | |
| 58 | static const char * const *record_mem_usage = __usage; |
| 59 | |
| 60 | static int __cmd_record(int argc, const char **argv, struct perf_mem *mem) |
| 61 | { |
| 62 | int rec_argc, i = 0, j; |
| 63 | const char **rec_argv; |
| 64 | int ret; |
| 65 | bool all_user = false, all_kernel = false; |
| 66 | struct option options[] = { |
| 67 | OPT_CALLBACK('e', "event", &mem, "event", |
| 68 | "event selector. use 'perf mem record -e list' to list available events", |
| 69 | parse_record_events), |
| 70 | OPT_UINTEGER(0, "ldlat", &perf_mem_events__loads_ldlat, "mem-loads latency"), |
| 71 | OPT_INCR('v', "verbose", &verbose, |
| 72 | "be more verbose (show counter open errors, etc)"), |
| 73 | OPT_BOOLEAN('U', "--all-user", &all_user, "collect only user level data"), |
| 74 | OPT_BOOLEAN('K', "--all-kernel", &all_kernel, "collect only kernel level data"), |
| 75 | OPT_END() |
| 76 | }; |
| 77 | |
| 78 | argc = parse_options(argc, argv, options, record_mem_usage, |
| 79 | PARSE_OPT_STOP_AT_NON_OPTION); |
| 80 | |
| 81 | rec_argc = argc + 9; /* max number of arguments */ |
| 82 | rec_argv = calloc(rec_argc + 1, sizeof(char *)); |
| 83 | if (!rec_argv) |
| 84 | return -1; |
| 85 | |
| 86 | rec_argv[i++] = "record"; |
| 87 | |
| 88 | if (mem->operation & MEM_OPERATION_LOAD) |
| 89 | perf_mem_events[PERF_MEM_EVENTS__LOAD].record = true; |
| 90 | |
| 91 | if (perf_mem_events[PERF_MEM_EVENTS__LOAD].record) |
| 92 | rec_argv[i++] = "-W"; |
| 93 | |
| 94 | rec_argv[i++] = "-d"; |
| 95 | |
| 96 | for (j = 0; j < PERF_MEM_EVENTS__MAX; j++) { |
| 97 | if (!perf_mem_events[j].record) |
| 98 | continue; |
| 99 | |
| 100 | if (!perf_mem_events[j].supported) { |
| 101 | pr_err("failed: event '%s' not supported\n", |
| 102 | perf_mem_events__name(j)); |
| 103 | return -1; |
| 104 | } |
| 105 | |
| 106 | rec_argv[i++] = "-e"; |
| 107 | rec_argv[i++] = perf_mem_events__name(j); |
| 108 | }; |
| 109 | |
| 110 | if (all_user) |
| 111 | rec_argv[i++] = "--all-user"; |
| 112 | |
| 113 | if (all_kernel) |
| 114 | rec_argv[i++] = "--all-kernel"; |
| 115 | |
| 116 | for (j = 0; j < argc; j++, i++) |
| 117 | rec_argv[i] = argv[j]; |
| 118 | |
| 119 | if (verbose > 0) { |
| 120 | pr_debug("calling: record "); |
| 121 | |
| 122 | while (rec_argv[j]) { |
| 123 | pr_debug("%s ", rec_argv[j]); |
| 124 | j++; |
| 125 | } |
| 126 | pr_debug("\n"); |
| 127 | } |
| 128 | |
| 129 | ret = cmd_record(i, rec_argv, NULL); |
| 130 | free(rec_argv); |
| 131 | return ret; |
| 132 | } |
| 133 | |
| 134 | static int |
| 135 | dump_raw_samples(struct perf_tool *tool, |
| 136 | union perf_event *event, |
| 137 | struct perf_sample *sample, |
| 138 | struct machine *machine) |
| 139 | { |
| 140 | struct perf_mem *mem = container_of(tool, struct perf_mem, tool); |
| 141 | struct addr_location al; |
| 142 | const char *fmt; |
| 143 | |
| 144 | if (machine__resolve(machine, &al, sample) < 0) { |
| 145 | fprintf(stderr, "problem processing %d event, skipping it.\n", |
| 146 | event->header.type); |
| 147 | return -1; |
| 148 | } |
| 149 | |
| 150 | if (al.filtered || (mem->hide_unresolved && al.sym == NULL)) |
| 151 | goto out_put; |
| 152 | |
| 153 | if (al.map != NULL) |
| 154 | al.map->dso->hit = 1; |
| 155 | |
| 156 | if (symbol_conf.field_sep) { |
| 157 | fmt = "%d%s%d%s0x%"PRIx64"%s0x%"PRIx64"%s%"PRIu64 |
| 158 | "%s0x%"PRIx64"%s%s:%s\n"; |
| 159 | } else { |
| 160 | fmt = "%5d%s%5d%s0x%016"PRIx64"%s0x016%"PRIx64 |
| 161 | "%s%5"PRIu64"%s0x%06"PRIx64"%s%s:%s\n"; |
| 162 | symbol_conf.field_sep = " "; |
| 163 | } |
| 164 | |
| 165 | printf(fmt, |
| 166 | sample->pid, |
| 167 | symbol_conf.field_sep, |
| 168 | sample->tid, |
| 169 | symbol_conf.field_sep, |
| 170 | sample->ip, |
| 171 | symbol_conf.field_sep, |
| 172 | sample->addr, |
| 173 | symbol_conf.field_sep, |
| 174 | sample->weight, |
| 175 | symbol_conf.field_sep, |
| 176 | sample->data_src, |
| 177 | symbol_conf.field_sep, |
| 178 | al.map ? (al.map->dso ? al.map->dso->long_name : "???") : "???", |
| 179 | al.sym ? al.sym->name : "???"); |
| 180 | out_put: |
| 181 | addr_location__put(&al); |
| 182 | return 0; |
| 183 | } |
| 184 | |
| 185 | static int process_sample_event(struct perf_tool *tool, |
| 186 | union perf_event *event, |
| 187 | struct perf_sample *sample, |
| 188 | struct perf_evsel *evsel __maybe_unused, |
| 189 | struct machine *machine) |
| 190 | { |
| 191 | return dump_raw_samples(tool, event, sample, machine); |
| 192 | } |
| 193 | |
| 194 | static int report_raw_events(struct perf_mem *mem) |
| 195 | { |
| 196 | struct perf_data_file file = { |
| 197 | .path = input_name, |
| 198 | .mode = PERF_DATA_MODE_READ, |
| 199 | .force = mem->force, |
| 200 | }; |
| 201 | int ret; |
| 202 | struct perf_session *session = perf_session__new(&file, false, |
| 203 | &mem->tool); |
| 204 | |
| 205 | if (session == NULL) |
| 206 | return -1; |
| 207 | |
| 208 | if (mem->cpu_list) { |
| 209 | ret = perf_session__cpu_bitmap(session, mem->cpu_list, |
| 210 | mem->cpu_bitmap); |
| 211 | if (ret < 0) |
| 212 | goto out_delete; |
| 213 | } |
| 214 | |
| 215 | ret = symbol__init(&session->header.env); |
| 216 | if (ret < 0) |
| 217 | goto out_delete; |
| 218 | |
| 219 | printf("# PID, TID, IP, ADDR, LOCAL WEIGHT, DSRC, SYMBOL\n"); |
| 220 | |
| 221 | ret = perf_session__process_events(session); |
| 222 | |
| 223 | out_delete: |
| 224 | perf_session__delete(session); |
| 225 | return ret; |
| 226 | } |
| 227 | |
| 228 | static int report_events(int argc, const char **argv, struct perf_mem *mem) |
| 229 | { |
| 230 | const char **rep_argv; |
| 231 | int ret, i = 0, j, rep_argc; |
| 232 | |
| 233 | if (mem->dump_raw) |
| 234 | return report_raw_events(mem); |
| 235 | |
| 236 | rep_argc = argc + 3; |
| 237 | rep_argv = calloc(rep_argc + 1, sizeof(char *)); |
| 238 | if (!rep_argv) |
| 239 | return -1; |
| 240 | |
| 241 | rep_argv[i++] = "report"; |
| 242 | rep_argv[i++] = "--mem-mode"; |
| 243 | rep_argv[i++] = "-n"; /* display number of samples */ |
| 244 | |
| 245 | /* |
| 246 | * there is no weight (cost) associated with stores, so don't print |
| 247 | * the column |
| 248 | */ |
| 249 | if (!(mem->operation & MEM_OPERATION_LOAD)) |
| 250 | rep_argv[i++] = "--sort=mem,sym,dso,symbol_daddr," |
| 251 | "dso_daddr,tlb,locked"; |
| 252 | |
| 253 | for (j = 1; j < argc; j++, i++) |
| 254 | rep_argv[i] = argv[j]; |
| 255 | |
| 256 | ret = cmd_report(i, rep_argv, NULL); |
| 257 | free(rep_argv); |
| 258 | return ret; |
| 259 | } |
| 260 | |
| 261 | struct mem_mode { |
| 262 | const char *name; |
| 263 | int mode; |
| 264 | }; |
| 265 | |
| 266 | #define MEM_OPT(n, m) \ |
| 267 | { .name = n, .mode = (m) } |
| 268 | |
| 269 | #define MEM_END { .name = NULL } |
| 270 | |
| 271 | static const struct mem_mode mem_modes[]={ |
| 272 | MEM_OPT("load", MEM_OPERATION_LOAD), |
| 273 | MEM_OPT("store", MEM_OPERATION_STORE), |
| 274 | MEM_END |
| 275 | }; |
| 276 | |
| 277 | static int |
| 278 | parse_mem_ops(const struct option *opt, const char *str, int unset) |
| 279 | { |
| 280 | int *mode = (int *)opt->value; |
| 281 | const struct mem_mode *m; |
| 282 | char *s, *os = NULL, *p; |
| 283 | int ret = -1; |
| 284 | |
| 285 | if (unset) |
| 286 | return 0; |
| 287 | |
| 288 | /* str may be NULL in case no arg is passed to -t */ |
| 289 | if (str) { |
| 290 | /* because str is read-only */ |
| 291 | s = os = strdup(str); |
| 292 | if (!s) |
| 293 | return -1; |
| 294 | |
| 295 | /* reset mode */ |
| 296 | *mode = 0; |
| 297 | |
| 298 | for (;;) { |
| 299 | p = strchr(s, ','); |
| 300 | if (p) |
| 301 | *p = '\0'; |
| 302 | |
| 303 | for (m = mem_modes; m->name; m++) { |
| 304 | if (!strcasecmp(s, m->name)) |
| 305 | break; |
| 306 | } |
| 307 | if (!m->name) { |
| 308 | fprintf(stderr, "unknown sampling op %s," |
| 309 | " check man page\n", s); |
| 310 | goto error; |
| 311 | } |
| 312 | |
| 313 | *mode |= m->mode; |
| 314 | |
| 315 | if (!p) |
| 316 | break; |
| 317 | |
| 318 | s = p + 1; |
| 319 | } |
| 320 | } |
| 321 | ret = 0; |
| 322 | |
| 323 | if (*mode == 0) |
| 324 | *mode = MEM_OPERATION_LOAD; |
| 325 | error: |
| 326 | free(os); |
| 327 | return ret; |
| 328 | } |
| 329 | |
| 330 | int cmd_mem(int argc, const char **argv, const char *prefix __maybe_unused) |
| 331 | { |
| 332 | struct stat st; |
| 333 | struct perf_mem mem = { |
| 334 | .tool = { |
| 335 | .sample = process_sample_event, |
| 336 | .mmap = perf_event__process_mmap, |
| 337 | .mmap2 = perf_event__process_mmap2, |
| 338 | .comm = perf_event__process_comm, |
| 339 | .lost = perf_event__process_lost, |
| 340 | .fork = perf_event__process_fork, |
| 341 | .build_id = perf_event__process_build_id, |
| 342 | .ordered_events = true, |
| 343 | }, |
| 344 | .input_name = "perf.data", |
| 345 | /* |
| 346 | * default to both load an store sampling |
| 347 | */ |
| 348 | .operation = MEM_OPERATION_LOAD | MEM_OPERATION_STORE, |
| 349 | }; |
| 350 | const struct option mem_options[] = { |
| 351 | OPT_CALLBACK('t', "type", &mem.operation, |
| 352 | "type", "memory operations(load,store) Default load,store", |
| 353 | parse_mem_ops), |
| 354 | OPT_BOOLEAN('D', "dump-raw-samples", &mem.dump_raw, |
| 355 | "dump raw samples in ASCII"), |
| 356 | OPT_BOOLEAN('U', "hide-unresolved", &mem.hide_unresolved, |
| 357 | "Only display entries resolved to a symbol"), |
| 358 | OPT_STRING('i', "input", &input_name, "file", |
| 359 | "input file name"), |
| 360 | OPT_STRING('C', "cpu", &mem.cpu_list, "cpu", |
| 361 | "list of cpus to profile"), |
| 362 | OPT_STRING_NOEMPTY('x', "field-separator", &symbol_conf.field_sep, |
| 363 | "separator", |
| 364 | "separator for columns, no spaces will be added" |
| 365 | " between columns '.' is reserved."), |
| 366 | OPT_BOOLEAN('f', "force", &mem.force, "don't complain, do it"), |
| 367 | OPT_END() |
| 368 | }; |
| 369 | const char *const mem_subcommands[] = { "record", "report", NULL }; |
| 370 | const char *mem_usage[] = { |
| 371 | NULL, |
| 372 | NULL |
| 373 | }; |
| 374 | |
| 375 | if (perf_mem_events__init()) { |
| 376 | pr_err("failed: memory events not supported\n"); |
| 377 | return -1; |
| 378 | } |
| 379 | |
| 380 | argc = parse_options_subcommand(argc, argv, mem_options, mem_subcommands, |
| 381 | mem_usage, PARSE_OPT_STOP_AT_NON_OPTION); |
| 382 | |
| 383 | if (!argc || !(strncmp(argv[0], "rec", 3) || mem.operation)) |
| 384 | usage_with_options(mem_usage, mem_options); |
| 385 | |
| 386 | if (!mem.input_name || !strlen(mem.input_name)) { |
| 387 | if (!fstat(STDIN_FILENO, &st) && S_ISFIFO(st.st_mode)) |
| 388 | mem.input_name = "-"; |
| 389 | else |
| 390 | mem.input_name = "perf.data"; |
| 391 | } |
| 392 | |
| 393 | if (!strncmp(argv[0], "rec", 3)) |
| 394 | return __cmd_record(argc, argv, &mem); |
| 395 | else if (!strncmp(argv[0], "rep", 3)) |
| 396 | return report_events(argc, argv, &mem); |
| 397 | else |
| 398 | usage_with_options(mem_usage, mem_options); |
| 399 | |
| 400 | return 0; |
| 401 | } |