tools/perf/builtin-mem.c

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