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