tools/perf/builtin-mem.c

   1 #include <inttypes.h>
   2 #include <sys/types.h>
   3 #include <sys/stat.h>
   4 #include <unistd.h>
   5 #include "builtin.h"
   6 #include "perf.h"
   7
   8 #include <subcmd/parse-options.h>
   9 #include "util/trace-event.h"
  10 #include "util/tool.h"
  11 #include "util/session.h"
  12 #include "util/data.h"
  13 #include "util/mem-events.h"
  14 #include "util/debug.h"
  15 #include "util/symbol.h"
  16
  17 #define MEM_OPERATION_LOAD      0x1
  18 #define MEM_OPERATION_STORE     0x2
  19
  20 struct perf_mem {
  21         struct perf_tool        tool;
  22         char const              *input_name;
  23         bool                    hide_unresolved;
  24         bool                    dump_raw;
  25         bool                    force;
  26         bool                    phys_addr;
  27         int                     operation;
  28         const char              *cpu_list;
  29         DECLARE_BITMAP(cpu_bitmap, MAX_NR_CPUS);
  30 };
  31
  32 static int parse_record_events(const struct option *opt,
  33                                const char *str, int unset __maybe_unused)
  34 {
  35         struct perf_mem *mem = *(struct perf_mem **)opt->value;
  36         int j;
  37
  38         if (strcmp(str, "list")) {
  39                 if (!perf_mem_events__parse(str)) {
  40                         mem->operation = 0;
  41                         return 0;
  42                 }
  43                 exit(-1);
  44         }
  45
  46         for (j = 0; j < PERF_MEM_EVENTS__MAX; j++) {
  47                 struct perf_mem_event *e = &perf_mem_events[j];
  48
  49                 fprintf(stderr, "%-13s%-*s%s\n",
  50                         e->tag,
  51                         verbose > 0 ? 25 : 0,
  52                         verbose > 0 ? perf_mem_events__name(j) : "",
  53                         e->supported ? ": available" : "");
  54         }
  55         exit(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;
  69         const char **rec_argv;
  70         int ret;
  71         bool all_user = false, all_kernel = false;
  72         struct option options[] = {
  73         OPT_CALLBACK('e', "event", &mem, "event",
  74                      "event selector. use 'perf mem record -e list' to list available events",
  75                      parse_record_events),
  76         OPT_UINTEGER(0, "ldlat", &perf_mem_events__loads_ldlat, "mem-loads latency"),
  77         OPT_INCR('v', "verbose", &verbose,
  78                  "be more verbose (show counter open errors, etc)"),
  79         OPT_BOOLEAN('U', "all-user", &all_user, "collect only user level data"),
  80         OPT_BOOLEAN('K', "all-kernel", &all_kernel, "collect only kernel level data"),
  81         OPT_END()
  82         };
  83
  84         argc = parse_options(argc, argv, options, record_mem_usage,
  85                              PARSE_OPT_STOP_AT_NON_OPTION);
  86
  87         rec_argc = argc + 9; /* max number of arguments */
  88         rec_argv = calloc(rec_argc + 1, sizeof(char *));
  89         if (!rec_argv)
  90                 return -1;
  91
  92         rec_argv[i++] = "record";
  93
  94         if (mem->operation & MEM_OPERATION_LOAD)
  95                 perf_mem_events[PERF_MEM_EVENTS__LOAD].record = true;
  96
  97         if (mem->operation & MEM_OPERATION_STORE)
  98                 perf_mem_events[PERF_MEM_EVENTS__STORE].record = true;
  99
 100         if (perf_mem_events[PERF_MEM_EVENTS__LOAD].record)
 101                 rec_argv[i++] = "-W";
 102
 103         rec_argv[i++] = "-d";
 104
 105         if (mem->phys_addr)
 106                 rec_argv[i++] = "--phys-data";
 107
 108         for (j = 0; j < PERF_MEM_EVENTS__MAX; j++) {
 109                 if (!perf_mem_events[j].record)
 110                         continue;
 111
 112                 if (!perf_mem_events[j].supported) {
 113                         pr_err("failed: event '%s' not supported\n",
 114                                perf_mem_events__name(j));
 115                         free(rec_argv);
 116                         return -1;
 117                 }
 118
 119                 rec_argv[i++] = "-e";
 120                 rec_argv[i++] = perf_mem_events__name(j);
 121         };
 122
 123         if (all_user)
 124                 rec_argv[i++] = "--all-user";
 125
 126         if (all_kernel)
 127                 rec_argv[i++] = "--all-kernel";
 128
 129         for (j = 0; j < argc; j++, i++)
 130                 rec_argv[i] = argv[j];
 131
 132         if (verbose > 0) {
 133                 pr_debug("calling: record ");
 134
 135                 while (rec_argv[j]) {
 136                         pr_debug("%s ", rec_argv[j]);
 137                         j++;
 138                 }
 139                 pr_debug("\n");
 140         }
 141
 142         ret = cmd_record(i, rec_argv);
 143         free(rec_argv);
 144         return ret;
 145 }
 146
 147 static int
 148 dump_raw_samples(struct perf_tool *tool,
 149                  union perf_event *event,
 150                  struct perf_sample *sample,
 151                  struct machine *machine)
 152 {
 153         struct perf_mem *mem = container_of(tool, struct perf_mem, tool);
 154         struct addr_location al;
 155         const char *fmt;
 156
 157         if (machine__resolve(machine, &al, sample) < 0) {
 158                 fprintf(stderr, "problem processing %d event, skipping it.\n",
 159                                 event->header.type);
 160                 return -1;
 161         }
 162
 163         if (al.filtered || (mem->hide_unresolved && al.sym == NULL))
 164                 goto out_put;
 165
 166         if (al.map != NULL)
 167                 al.map->dso->hit = 1;
 168
 169         if (mem->phys_addr) {
 170                 if (symbol_conf.field_sep) {
 171                         fmt = "%d%s%d%s0x%"PRIx64"%s0x%"PRIx64"%s0x%016"PRIx64
 172                               "%s%"PRIu64"%s0x%"PRIx64"%s%s:%s\n";
 173                 } else {
 174                         fmt = "%5d%s%5d%s0x%016"PRIx64"%s0x016%"PRIx64
 175                               "%s0x%016"PRIx64"%s%5"PRIu64"%s0x%06"PRIx64
 176                               "%s%s:%s\n";
 177                         symbol_conf.field_sep = " ";
 178                 }
 179
 180                 printf(fmt,
 181                         sample->pid,
 182                         symbol_conf.field_sep,
 183                         sample->tid,
 184                         symbol_conf.field_sep,
 185                         sample->ip,
 186                         symbol_conf.field_sep,
 187                         sample->addr,
 188                         symbol_conf.field_sep,
 189                         sample->phys_addr,
 190                         symbol_conf.field_sep,
 191                         sample->weight,
 192                         symbol_conf.field_sep,
 193                         sample->data_src,
 194                         symbol_conf.field_sep,
 195                         al.map ? (al.map->dso ? al.map->dso->long_name : "???") : "???",
 196                         al.sym ? al.sym->name : "???");
 197         } else {
 198                 if (symbol_conf.field_sep) {
 199                         fmt = "%d%s%d%s0x%"PRIx64"%s0x%"PRIx64"%s%"PRIu64
 200                               "%s0x%"PRIx64"%s%s:%s\n";
 201                 } else {
 202                         fmt = "%5d%s%5d%s0x%016"PRIx64"%s0x016%"PRIx64
 203                               "%s%5"PRIu64"%s0x%06"PRIx64"%s%s:%s\n";
 204                         symbol_conf.field_sep = " ";
 205                 }
 206
 207                 printf(fmt,
 208                         sample->pid,
 209                         symbol_conf.field_sep,
 210                         sample->tid,
 211                         symbol_conf.field_sep,
 212                         sample->ip,
 213                         symbol_conf.field_sep,
 214                         sample->addr,
 215                         symbol_conf.field_sep,
 216                         sample->weight,
 217                         symbol_conf.field_sep,
 218                         sample->data_src,
 219                         symbol_conf.field_sep,
 220                         al.map ? (al.map->dso ? al.map->dso->long_name : "???") : "???",
 221                         al.sym ? al.sym->name : "???");
 222         }
 223 out_put:
 224         addr_location__put(&al);
 225         return 0;
 226 }
 227
 228 static int process_sample_event(struct perf_tool *tool,
 229                                 union perf_event *event,
 230                                 struct perf_sample *sample,
 231                                 struct perf_evsel *evsel __maybe_unused,
 232                                 struct machine *machine)
 233 {
 234         return dump_raw_samples(tool, event, sample, machine);
 235 }
 236
 237 static int report_raw_events(struct perf_mem *mem)
 238 {
 239         struct perf_data_file file = {
 240                 .path = input_name,
 241                 .mode = PERF_DATA_MODE_READ,
 242                 .force = mem->force,
 243         };
 244         int ret;
 245         struct perf_session *session = perf_session__new(&file, false,
 246                                                          &mem->tool);
 247
 248         if (session == NULL)
 249                 return -1;
 250
 251         if (mem->cpu_list) {
 252                 ret = perf_session__cpu_bitmap(session, mem->cpu_list,
 253                                                mem->cpu_bitmap);
 254                 if (ret < 0)
 255                         goto out_delete;
 256         }
 257
 258         ret = symbol__init(&session->header.env);
 259         if (ret < 0)
 260                 goto out_delete;
 261
 262         if (mem->phys_addr)
 263                 printf("# PID, TID, IP, ADDR, PHYS ADDR, LOCAL WEIGHT, DSRC, SYMBOL\n");
 264         else
 265                 printf("# PID, TID, IP, ADDR, LOCAL WEIGHT, DSRC, SYMBOL\n");
 266
 267         ret = perf_session__process_events(session);
 268
 269 out_delete:
 270         perf_session__delete(session);
 271         return ret;
 272 }
 273
 274 static int report_events(int argc, const char **argv, struct perf_mem *mem)
 275 {
 276         const char **rep_argv;
 277         int ret, i = 0, j, rep_argc;
 278
 279         if (mem->dump_raw)
 280                 return report_raw_events(mem);
 281
 282         rep_argc = argc + 3;
 283         rep_argv = calloc(rep_argc + 1, sizeof(char *));
 284         if (!rep_argv)
 285                 return -1;
 286
 287         rep_argv[i++] = "report";
 288         rep_argv[i++] = "--mem-mode";
 289         rep_argv[i++] = "-n"; /* display number of samples */
 290
 291         /*
 292          * there is no weight (cost) associated with stores, so don't print
 293          * the column
 294          */
 295         if (!(mem->operation & MEM_OPERATION_LOAD)) {
 296                 if (mem->phys_addr)
 297                         rep_argv[i++] = "--sort=mem,sym,dso,symbol_daddr,"
 298                                         "dso_daddr,tlb,locked,phys_daddr";
 299                 else
 300                         rep_argv[i++] = "--sort=mem,sym,dso,symbol_daddr,"
 301                                         "dso_daddr,tlb,locked";
 302         } else if (mem->phys_addr)
 303                 rep_argv[i++] = "--sort=local_weight,mem,sym,dso,symbol_daddr,"
 304                                 "dso_daddr,snoop,tlb,locked,phys_daddr";
 305
 306         for (j = 1; j < argc; j++, i++)
 307                 rep_argv[i] = argv[j];
 308
 309         ret = cmd_report(i, rep_argv);
 310         free(rep_argv);
 311         return ret;
 312 }
 313
 314 struct mem_mode {
 315         const char *name;
 316         int mode;
 317 };
 318
 319 #define MEM_OPT(n, m) \
 320         { .name = n, .mode = (m) }
 321
 322 #define MEM_END { .name = NULL }
 323
 324 static const struct mem_mode mem_modes[]={
 325         MEM_OPT("load", MEM_OPERATION_LOAD),
 326         MEM_OPT("store", MEM_OPERATION_STORE),
 327         MEM_END
 328 };
 329
 330 static int
 331 parse_mem_ops(const struct option *opt, const char *str, int unset)
 332 {
 333         int *mode = (int *)opt->value;
 334         const struct mem_mode *m;
 335         char *s, *os = NULL, *p;
 336         int ret = -1;
 337
 338         if (unset)
 339                 return 0;
 340
 341         /* str may be NULL in case no arg is passed to -t */
 342         if (str) {
 343                 /* because str is read-only */
 344                 s = os = strdup(str);
 345                 if (!s)
 346                         return -1;
 347
 348                 /* reset mode */
 349                 *mode = 0;
 350
 351                 for (;;) {
 352                         p = strchr(s, ',');
 353                         if (p)
 354                                 *p = '\0';
 355
 356                         for (m = mem_modes; m->name; m++) {
 357                                 if (!strcasecmp(s, m->name))
 358                                         break;
 359                         }
 360                         if (!m->name) {
 361                                 fprintf(stderr, "unknown sampling op %s,"
 362                                             " check man page\n", s);
 363                                 goto error;
 364                         }
 365
 366                         *mode |= m->mode;
 367
 368                         if (!p)
 369                                 break;
 370
 371                         s = p + 1;
 372                 }
 373         }
 374         ret = 0;
 375
 376         if (*mode == 0)
 377                 *mode = MEM_OPERATION_LOAD;
 378 error:
 379         free(os);
 380         return ret;
 381 }
 382
 383 int cmd_mem(int argc, const char **argv)
 384 {
 385         struct stat st;
 386         struct perf_mem mem = {
 387                 .tool = {
 388                         .sample         = process_sample_event,
 389                         .mmap           = perf_event__process_mmap,
 390                         .mmap2          = perf_event__process_mmap2,
 391                         .comm           = perf_event__process_comm,
 392                         .lost           = perf_event__process_lost,
 393                         .fork           = perf_event__process_fork,
 394                         .build_id       = perf_event__process_build_id,
 395                         .namespaces     = perf_event__process_namespaces,
 396                         .ordered_events = true,
 397                 },
 398                 .input_name              = "perf.data",
 399                 /*
 400                  * default to both load an store sampling
 401                  */
 402                 .operation               = MEM_OPERATION_LOAD | MEM_OPERATION_STORE,
 403         };
 404         const struct option mem_options[] = {
 405         OPT_CALLBACK('t', "type", &mem.operation,
 406                    "type", "memory operations(load,store) Default load,store",
 407                     parse_mem_ops),
 408         OPT_BOOLEAN('D', "dump-raw-samples", &mem.dump_raw,
 409                     "dump raw samples in ASCII"),
 410         OPT_BOOLEAN('U', "hide-unresolved", &mem.hide_unresolved,
 411                     "Only display entries resolved to a symbol"),
 412         OPT_STRING('i', "input", &input_name, "file",
 413                    "input file name"),
 414         OPT_STRING('C', "cpu", &mem.cpu_list, "cpu",
 415                    "list of cpus to profile"),
 416         OPT_STRING_NOEMPTY('x', "field-separator", &symbol_conf.field_sep,
 417                    "separator",
 418                    "separator for columns, no spaces will be added"
 419                    " between columns '.' is reserved."),
 420         OPT_BOOLEAN('f', "force", &mem.force, "don't complain, do it"),
 421         OPT_BOOLEAN('p', "phys-data", &mem.phys_addr, "Record/Report sample physical addresses"),
 422         OPT_END()
 423         };
 424         const char *const mem_subcommands[] = { "record", "report", NULL };
 425         const char *mem_usage[] = {
 426                 NULL,
 427                 NULL
 428         };
 429
 430         if (perf_mem_events__init()) {
 431                 pr_err("failed: memory events not supported\n");
 432                 return -1;
 433         }
 434
 435         argc = parse_options_subcommand(argc, argv, mem_options, mem_subcommands,
 436                                         mem_usage, PARSE_OPT_STOP_AT_NON_OPTION);
 437
 438         if (!argc || !(strncmp(argv[0], "rec", 3) || mem.operation))
 439                 usage_with_options(mem_usage, mem_options);
 440
 441         if (!mem.input_name || !strlen(mem.input_name)) {
 442                 if (!fstat(STDIN_FILENO, &st) && S_ISFIFO(st.st_mode))
 443                         mem.input_name = "-";
 444                 else
 445                         mem.input_name = "perf.data";
 446         }
 447
 448         if (!strncmp(argv[0], "rec", 3))
 449                 return __cmd_record(argc, argv, &mem);
 450         else if (!strncmp(argv[0], "rep", 3))
 451                 return report_events(argc, argv, &mem);
 452         else
 453                 usage_with_options(mem_usage, mem_options);
 454
 455         return 0;
 456 }