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 <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;
 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         if (mem->phys_addr) {
 190                 if (symbol_conf.field_sep) {
 191                         fmt = "%d%s%d%s0x%"PRIx64"%s0x%"PRIx64"%s0x%016"PRIx64
 192                               "%s%"PRIu64"%s0x%"PRIx64"%s%s:%s\n";
 193                 } else {
 194                         fmt = "%5d%s%5d%s0x%016"PRIx64"%s0x016%"PRIx64
 195                               "%s0x%016"PRIx64"%s%5"PRIu64"%s0x%06"PRIx64
 196                               "%s%s:%s\n";
 197                         symbol_conf.field_sep = " ";
 198                 }
 199
 200                 printf(fmt,
 201                         sample->pid,
 202                         symbol_conf.field_sep,
 203                         sample->tid,
 204                         symbol_conf.field_sep,
 205                         sample->ip,
 206                         symbol_conf.field_sep,
 207                         sample->addr,
 208                         symbol_conf.field_sep,
 209                         sample->phys_addr,
 210                         symbol_conf.field_sep,
 211                         sample->weight,
 212                         symbol_conf.field_sep,
 213                         sample->data_src,
 214                         symbol_conf.field_sep,
 215                         al.map ? (al.map->dso ? al.map->dso->long_name : "???") : "???",
 216                         al.sym ? al.sym->name : "???");
 217         } else {
 218                 if (symbol_conf.field_sep) {
 219                         fmt = "%d%s%d%s0x%"PRIx64"%s0x%"PRIx64"%s%"PRIu64
 220                               "%s0x%"PRIx64"%s%s:%s\n";
 221                 } else {
 222                         fmt = "%5d%s%5d%s0x%016"PRIx64"%s0x016%"PRIx64
 223                               "%s%5"PRIu64"%s0x%06"PRIx64"%s%s:%s\n";
 224                         symbol_conf.field_sep = " ";
 225                 }
 226
 227                 printf(fmt,
 228                         sample->pid,
 229                         symbol_conf.field_sep,
 230                         sample->tid,
 231                         symbol_conf.field_sep,
 232                         sample->ip,
 233                         symbol_conf.field_sep,
 234                         sample->addr,
 235                         symbol_conf.field_sep,
 236                         sample->weight,
 237                         symbol_conf.field_sep,
 238                         sample->data_src,
 239                         symbol_conf.field_sep,
 240                         al.map ? (al.map->dso ? al.map->dso->long_name : "???") : "???",
 241                         al.sym ? al.sym->name : "???");
 242         }
 243 out_put:
 244         addr_location__put(&al);
 245         return 0;
 246 }
 247
 248 static int process_sample_event(struct perf_tool *tool,
 249                                 union perf_event *event,
 250                                 struct perf_sample *sample,
 251                                 struct evsel *evsel __maybe_unused,
 252                                 struct machine *machine)
 253 {
 254         return dump_raw_samples(tool, event, sample, machine);
 255 }
 256
 257 static int report_raw_events(struct perf_mem *mem)
 258 {
 259         struct itrace_synth_opts itrace_synth_opts = {
 260                 .set = true,
 261                 .mem = true,    /* Only enable memory event */
 262                 .default_no_sample = true,
 263         };
 264
 265         struct perf_data data = {
 266                 .path  = input_name,
 267                 .mode  = PERF_DATA_MODE_READ,
 268                 .force = mem->force,
 269         };
 270         int ret;
 271         struct perf_session *session = perf_session__new(&data, false,
 272                                                          &mem->tool);
 273
 274         if (IS_ERR(session))
 275                 return PTR_ERR(session);
 276
 277         session->itrace_synth_opts = &itrace_synth_opts;
 278
 279         if (mem->cpu_list) {
 280                 ret = perf_session__cpu_bitmap(session, mem->cpu_list,
 281                                                mem->cpu_bitmap);
 282                 if (ret < 0)
 283                         goto out_delete;
 284         }
 285
 286         ret = symbol__init(&session->header.env);
 287         if (ret < 0)
 288                 goto out_delete;
 289
 290         if (mem->phys_addr)
 291                 printf("# PID, TID, IP, ADDR, PHYS ADDR, LOCAL WEIGHT, DSRC, SYMBOL\n");
 292         else
 293                 printf("# PID, TID, IP, ADDR, LOCAL WEIGHT, DSRC, SYMBOL\n");
 294
 295         ret = perf_session__process_events(session);
 296
 297 out_delete:
 298         perf_session__delete(session);
 299         return ret;
 300 }
 301 static char *get_sort_order(struct perf_mem *mem)
 302 {
 303         bool has_extra_options = mem->phys_addr ? true : false;
 304         char sort[128];
 305
 306         /*
 307          * there is no weight (cost) associated with stores, so don't print
 308          * the column
 309          */
 310         if (!(mem->operation & MEM_OPERATION_LOAD)) {
 311                 strcpy(sort, "--sort=mem,sym,dso,symbol_daddr,"
 312                              "dso_daddr,tlb,locked");
 313         } else if (has_extra_options) {
 314                 strcpy(sort, "--sort=local_weight,mem,sym,dso,symbol_daddr,"
 315                              "dso_daddr,snoop,tlb,locked");
 316         } else
 317                 return NULL;
 318
 319         if (mem->phys_addr)
 320                 strcat(sort, ",phys_daddr");
 321
 322         return strdup(sort);
 323 }
 324
 325 static int report_events(int argc, const char **argv, struct perf_mem *mem)
 326 {
 327         const char **rep_argv;
 328         int ret, i = 0, j, rep_argc;
 329         char *new_sort_order;
 330
 331         if (mem->dump_raw)
 332                 return report_raw_events(mem);
 333
 334         rep_argc = argc + 3;
 335         rep_argv = calloc(rep_argc + 1, sizeof(char *));
 336         if (!rep_argv)
 337                 return -1;
 338
 339         rep_argv[i++] = "report";
 340         rep_argv[i++] = "--mem-mode";
 341         rep_argv[i++] = "-n"; /* display number of samples */
 342
 343         new_sort_order = get_sort_order(mem);
 344         if (new_sort_order)
 345                 rep_argv[i++] = new_sort_order;
 346
 347         for (j = 1; j < argc; j++, i++)
 348                 rep_argv[i] = argv[j];
 349
 350         ret = cmd_report(i, rep_argv);
 351         free(rep_argv);
 352         return ret;
 353 }
 354
 355 struct mem_mode {
 356         const char *name;
 357         int mode;
 358 };
 359
 360 #define MEM_OPT(n, m) \
 361         { .name = n, .mode = (m) }
 362
 363 #define MEM_END { .name = NULL }
 364
 365 static const struct mem_mode mem_modes[]={
 366         MEM_OPT("load", MEM_OPERATION_LOAD),
 367         MEM_OPT("store", MEM_OPERATION_STORE),
 368         MEM_END
 369 };
 370
 371 static int
 372 parse_mem_ops(const struct option *opt, const char *str, int unset)
 373 {
 374         int *mode = (int *)opt->value;
 375         const struct mem_mode *m;
 376         char *s, *os = NULL, *p;
 377         int ret = -1;
 378
 379         if (unset)
 380                 return 0;
 381
 382         /* str may be NULL in case no arg is passed to -t */
 383         if (str) {
 384                 /* because str is read-only */
 385                 s = os = strdup(str);
 386                 if (!s)
 387                         return -1;
 388
 389                 /* reset mode */
 390                 *mode = 0;
 391
 392                 for (;;) {
 393                         p = strchr(s, ',');
 394                         if (p)
 395                                 *p = '\0';
 396
 397                         for (m = mem_modes; m->name; m++) {
 398                                 if (!strcasecmp(s, m->name))
 399                                         break;
 400                         }
 401                         if (!m->name) {
 402                                 fprintf(stderr, "unknown sampling op %s,"
 403                                             " check man page\n", s);
 404                                 goto error;
 405                         }
 406
 407                         *mode |= m->mode;
 408
 409                         if (!p)
 410                                 break;
 411
 412                         s = p + 1;
 413                 }
 414         }
 415         ret = 0;
 416
 417         if (*mode == 0)
 418                 *mode = MEM_OPERATION_LOAD;
 419 error:
 420         free(os);
 421         return ret;
 422 }
 423
 424 int cmd_mem(int argc, const char **argv)
 425 {
 426         struct stat st;
 427         struct perf_mem mem = {
 428                 .tool = {
 429                         .sample         = process_sample_event,
 430                         .mmap           = perf_event__process_mmap,
 431                         .mmap2          = perf_event__process_mmap2,
 432                         .comm           = perf_event__process_comm,
 433                         .lost           = perf_event__process_lost,
 434                         .fork           = perf_event__process_fork,
 435                         .attr           = perf_event__process_attr,
 436                         .build_id       = perf_event__process_build_id,
 437                         .namespaces     = perf_event__process_namespaces,
 438                         .auxtrace_info  = perf_event__process_auxtrace_info,
 439                         .auxtrace       = perf_event__process_auxtrace,
 440                         .auxtrace_error = perf_event__process_auxtrace_error,
 441                         .ordered_events = true,
 442                 },
 443                 .input_name              = "perf.data",
 444                 /*
 445                  * default to both load an store sampling
 446                  */
 447                 .operation               = MEM_OPERATION_LOAD | MEM_OPERATION_STORE,
 448         };
 449         const struct option mem_options[] = {
 450         OPT_CALLBACK('t', "type", &mem.operation,
 451                    "type", "memory operations(load,store) Default load,store",
 452                     parse_mem_ops),
 453         OPT_BOOLEAN('D', "dump-raw-samples", &mem.dump_raw,
 454                     "dump raw samples in ASCII"),
 455         OPT_BOOLEAN('U', "hide-unresolved", &mem.hide_unresolved,
 456                     "Only display entries resolved to a symbol"),
 457         OPT_STRING('i', "input", &input_name, "file",
 458                    "input file name"),
 459         OPT_STRING('C', "cpu", &mem.cpu_list, "cpu",
 460                    "list of cpus to profile"),
 461         OPT_STRING_NOEMPTY('x', "field-separator", &symbol_conf.field_sep,
 462                    "separator",
 463                    "separator for columns, no spaces will be added"
 464                    " between columns '.' is reserved."),
 465         OPT_BOOLEAN('f', "force", &mem.force, "don't complain, do it"),
 466         OPT_BOOLEAN('p', "phys-data", &mem.phys_addr, "Record/Report sample physical addresses"),
 467         OPT_END()
 468         };
 469         const char *const mem_subcommands[] = { "record", "report", NULL };
 470         const char *mem_usage[] = {
 471                 NULL,
 472                 NULL
 473         };
 474
 475         argc = parse_options_subcommand(argc, argv, mem_options, mem_subcommands,
 476                                         mem_usage, PARSE_OPT_KEEP_UNKNOWN);
 477
 478         if (!argc || !(strncmp(argv[0], "rec", 3) || mem.operation))
 479                 usage_with_options(mem_usage, mem_options);
 480
 481         if (!mem.input_name || !strlen(mem.input_name)) {
 482                 if (!fstat(STDIN_FILENO, &st) && S_ISFIFO(st.st_mode))
 483                         mem.input_name = "-";
 484                 else
 485                         mem.input_name = "perf.data";
 486         }
 487
 488         if (!strncmp(argv[0], "rec", 3))
 489                 return __cmd_record(argc, argv, &mem);
 490         else if (!strncmp(argv[0], "rep", 3))
 491                 return report_events(argc, argv, &mem);
 492         else
 493                 usage_with_options(mem_usage, mem_options);
 494
 495         return 0;
 496 }