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/trace-event.h"
  11 #include "util/tool.h"
  12 #include "util/session.h"
  13 #include "util/data.h"
  14 #include "util/map_symbol.h"
  15 #include "util/mem-events.h"
  16 #include "util/debug.h"
  17 #include "util/dso.h"
  18 #include "util/map.h"
  19 #include "util/symbol.h"
  20 #include <linux/err.h>
  21
  22 #define MEM_OPERATION_LOAD      0x1
  23 #define MEM_OPERATION_STORE     0x2
  24
  25 struct perf_mem {
  26         struct perf_tool        tool;
  27         char const              *input_name;
  28         bool                    hide_unresolved;
  29         bool                    dump_raw;
  30         bool                    force;
  31         bool                    phys_addr;
  32         int                     operation;
  33         const char              *cpu_list;
  34         DECLARE_BITMAP(cpu_bitmap, MAX_NR_CPUS);
  35 };
  36
  37 static int parse_record_events(const struct option *opt,
  38                                const char *str, int unset __maybe_unused)
  39 {
  40         struct perf_mem *mem = *(struct perf_mem **)opt->value;
  41
  42         if (!strcmp(str, "list")) {
  43                 perf_mem_events__list();
  44                 exit(0);
  45         }
  46         if (perf_mem_events__parse(str))
  47                 exit(-1);
  48
  49         mem->operation = 0;
  50         return 0;
  51 }
  52
  53 static const char * const __usage[] = {
  54         "perf mem record [<options>] [<command>]",
  55         "perf mem record [<options>] -- <command> [<options>]",
  56         NULL
  57 };
  58
  59 static const char * const *record_mem_usage = __usage;
  60
  61 static int __cmd_record(int argc, const char **argv, struct perf_mem *mem)
  62 {
  63         int rec_argc, i = 0, j;
  64         const char **rec_argv;
  65         int ret;
  66         bool all_user = false, all_kernel = false;
  67         struct perf_mem_event *e;
  68         struct option options[] = {
  69         OPT_CALLBACK('e', "event", &mem, "event",
  70                      "event selector. use 'perf mem record -e list' to list available events",
  71                      parse_record_events),
  72         OPT_UINTEGER(0, "ldlat", &perf_mem_events__loads_ldlat, "mem-loads latency"),
  73         OPT_INCR('v', "verbose", &verbose,
  74                  "be more verbose (show counter open errors, etc)"),
  75         OPT_BOOLEAN('U', "all-user", &all_user, "collect only user level data"),
  76         OPT_BOOLEAN('K', "all-kernel", &all_kernel, "collect only kernel level data"),
  77         OPT_END()
  78         };
  79
  80         if (perf_mem_events__init()) {
  81                 pr_err("failed: memory events not supported\n");
  82                 return -1;
  83         }
  84
  85         argc = parse_options(argc, argv, options, record_mem_usage,
  86                              PARSE_OPT_KEEP_UNKNOWN);
  87
  88         rec_argc = argc + 9; /* max number of arguments */
  89         rec_argv = calloc(rec_argc + 1, sizeof(char *));
  90         if (!rec_argv)
  91                 return -1;
  92
  93         rec_argv[i++] = "record";
  94
  95         e = perf_mem_events__ptr(PERF_MEM_EVENTS__LOAD_STORE);
  96
  97         /*
  98          * The load and store operations are required, use the event
  99          * PERF_MEM_EVENTS__LOAD_STORE if it is supported.
 100          */
 101         if (e->tag &&
 102             (mem->operation & MEM_OPERATION_LOAD) &&
 103             (mem->operation & MEM_OPERATION_STORE)) {
 104                 e->record = true;
 105         } else {
 106                 if (mem->operation & MEM_OPERATION_LOAD) {
 107                         e = perf_mem_events__ptr(PERF_MEM_EVENTS__LOAD);
 108                         e->record = true;
 109                 }
 110
 111                 if (mem->operation & MEM_OPERATION_STORE) {
 112                         e = perf_mem_events__ptr(PERF_MEM_EVENTS__STORE);
 113                         e->record = true;
 114                 }
 115         }
 116
 117         e = perf_mem_events__ptr(PERF_MEM_EVENTS__LOAD);
 118         if (e->record)
 119                 rec_argv[i++] = "-W";
 120
 121         rec_argv[i++] = "-d";
 122
 123         if (mem->phys_addr)
 124                 rec_argv[i++] = "--phys-data";
 125
 126         for (j = 0; j < PERF_MEM_EVENTS__MAX; j++) {
 127                 e = perf_mem_events__ptr(j);
 128                 if (!e->record)
 129                         continue;
 130
 131                 if (!e->supported) {
 132                         pr_err("failed: event '%s' not supported\n",
 133                                perf_mem_events__name(j));
 134                         free(rec_argv);
 135                         return -1;
 136                 }
 137
 138                 rec_argv[i++] = "-e";
 139                 rec_argv[i++] = perf_mem_events__name(j);
 140         }
 141
 142         if (all_user)
 143                 rec_argv[i++] = "--all-user";
 144
 145         if (all_kernel)
 146                 rec_argv[i++] = "--all-kernel";
 147
 148         for (j = 0; j < argc; j++, i++)
 149                 rec_argv[i] = argv[j];
 150
 151         if (verbose > 0) {
 152                 pr_debug("calling: record ");
 153
 154                 while (rec_argv[j]) {
 155                         pr_debug("%s ", rec_argv[j]);
 156                         j++;
 157                 }
 158                 pr_debug("\n");
 159         }
 160
 161         ret = cmd_record(i, rec_argv);
 162         free(rec_argv);
 163         return ret;
 164 }
 165
 166 static int
 167 dump_raw_samples(struct perf_tool *tool,
 168                  union perf_event *event,
 169                  struct perf_sample *sample,
 170                  struct machine *machine)
 171 {
 172         struct perf_mem *mem = container_of(tool, struct perf_mem, tool);
 173         struct addr_location al;
 174         const char *fmt;
 175
 176         if (machine__resolve(machine, &al, sample) < 0) {
 177                 fprintf(stderr, "problem processing %d event, skipping it.\n",
 178                                 event->header.type);
 179                 return -1;
 180         }
 181
 182         if (al.filtered || (mem->hide_unresolved && al.sym == NULL))
 183                 goto out_put;
 184
 185         if (al.map != NULL)
 186                 al.map->dso->hit = 1;
 187
 188         if (mem->phys_addr) {
 189                 if (symbol_conf.field_sep) {
 190                         fmt = "%d%s%d%s0x%"PRIx64"%s0x%"PRIx64"%s0x%016"PRIx64
 191                               "%s%"PRIu64"%s0x%"PRIx64"%s%s:%s\n";
 192                 } else {
 193                         fmt = "%5d%s%5d%s0x%016"PRIx64"%s0x016%"PRIx64
 194                               "%s0x%016"PRIx64"%s%5"PRIu64"%s0x%06"PRIx64
 195                               "%s%s:%s\n";
 196                         symbol_conf.field_sep = " ";
 197                 }
 198
 199                 printf(fmt,
 200                         sample->pid,
 201                         symbol_conf.field_sep,
 202                         sample->tid,
 203                         symbol_conf.field_sep,
 204                         sample->ip,
 205                         symbol_conf.field_sep,
 206                         sample->addr,
 207                         symbol_conf.field_sep,
 208                         sample->phys_addr,
 209                         symbol_conf.field_sep,
 210                         sample->weight,
 211                         symbol_conf.field_sep,
 212                         sample->data_src,
 213                         symbol_conf.field_sep,
 214                         al.map ? (al.map->dso ? al.map->dso->long_name : "???") : "???",
 215                         al.sym ? al.sym->name : "???");
 216         } else {
 217                 if (symbol_conf.field_sep) {
 218                         fmt = "%d%s%d%s0x%"PRIx64"%s0x%"PRIx64"%s%"PRIu64
 219                               "%s0x%"PRIx64"%s%s:%s\n";
 220                 } else {
 221                         fmt = "%5d%s%5d%s0x%016"PRIx64"%s0x016%"PRIx64
 222                               "%s%5"PRIu64"%s0x%06"PRIx64"%s%s:%s\n";
 223                         symbol_conf.field_sep = " ";
 224                 }
 225
 226                 printf(fmt,
 227                         sample->pid,
 228                         symbol_conf.field_sep,
 229                         sample->tid,
 230                         symbol_conf.field_sep,
 231                         sample->ip,
 232                         symbol_conf.field_sep,
 233                         sample->addr,
 234                         symbol_conf.field_sep,
 235                         sample->weight,
 236                         symbol_conf.field_sep,
 237                         sample->data_src,
 238                         symbol_conf.field_sep,
 239                         al.map ? (al.map->dso ? al.map->dso->long_name : "???") : "???",
 240                         al.sym ? al.sym->name : "???");
 241         }
 242 out_put:
 243         addr_location__put(&al);
 244         return 0;
 245 }
 246
 247 static int process_sample_event(struct perf_tool *tool,
 248                                 union perf_event *event,
 249                                 struct perf_sample *sample,
 250                                 struct evsel *evsel __maybe_unused,
 251                                 struct machine *machine)
 252 {
 253         return dump_raw_samples(tool, event, sample, machine);
 254 }
 255
 256 static int report_raw_events(struct perf_mem *mem)
 257 {
 258         struct perf_data data = {
 259                 .path  = input_name,
 260                 .mode  = PERF_DATA_MODE_READ,
 261                 .force = mem->force,
 262         };
 263         int ret;
 264         struct perf_session *session = perf_session__new(&data, false,
 265                                                          &mem->tool);
 266
 267         if (IS_ERR(session))
 268                 return PTR_ERR(session);
 269
 270         if (mem->cpu_list) {
 271                 ret = perf_session__cpu_bitmap(session, mem->cpu_list,
 272                                                mem->cpu_bitmap);
 273                 if (ret < 0)
 274                         goto out_delete;
 275         }
 276
 277         ret = symbol__init(&session->header.env);
 278         if (ret < 0)
 279                 goto out_delete;
 280
 281         if (mem->phys_addr)
 282                 printf("# PID, TID, IP, ADDR, PHYS ADDR, LOCAL WEIGHT, DSRC, SYMBOL\n");
 283         else
 284                 printf("# PID, TID, IP, ADDR, LOCAL WEIGHT, DSRC, SYMBOL\n");
 285
 286         ret = perf_session__process_events(session);
 287
 288 out_delete:
 289         perf_session__delete(session);
 290         return ret;
 291 }
 292
 293 static int report_events(int argc, const char **argv, struct perf_mem *mem)
 294 {
 295         const char **rep_argv;
 296         int ret, i = 0, j, rep_argc;
 297
 298         if (mem->dump_raw)
 299                 return report_raw_events(mem);
 300
 301         rep_argc = argc + 3;
 302         rep_argv = calloc(rep_argc + 1, sizeof(char *));
 303         if (!rep_argv)
 304                 return -1;
 305
 306         rep_argv[i++] = "report";
 307         rep_argv[i++] = "--mem-mode";
 308         rep_argv[i++] = "-n"; /* display number of samples */
 309
 310         /*
 311          * there is no weight (cost) associated with stores, so don't print
 312          * the column
 313          */
 314         if (!(mem->operation & MEM_OPERATION_LOAD)) {
 315                 if (mem->phys_addr)
 316                         rep_argv[i++] = "--sort=mem,sym,dso,symbol_daddr,"
 317                                         "dso_daddr,tlb,locked,phys_daddr";
 318                 else
 319                         rep_argv[i++] = "--sort=mem,sym,dso,symbol_daddr,"
 320                                         "dso_daddr,tlb,locked";
 321         } else if (mem->phys_addr)
 322                 rep_argv[i++] = "--sort=local_weight,mem,sym,dso,symbol_daddr,"
 323                                 "dso_daddr,snoop,tlb,locked,phys_daddr";
 324
 325         for (j = 1; j < argc; j++, i++)
 326                 rep_argv[i] = argv[j];
 327
 328         ret = cmd_report(i, rep_argv);
 329         free(rep_argv);
 330         return ret;
 331 }
 332
 333 struct mem_mode {
 334         const char *name;
 335         int mode;
 336 };
 337
 338 #define MEM_OPT(n, m) \
 339         { .name = n, .mode = (m) }
 340
 341 #define MEM_END { .name = NULL }
 342
 343 static const struct mem_mode mem_modes[]={
 344         MEM_OPT("load", MEM_OPERATION_LOAD),
 345         MEM_OPT("store", MEM_OPERATION_STORE),
 346         MEM_END
 347 };
 348
 349 static int
 350 parse_mem_ops(const struct option *opt, const char *str, int unset)
 351 {
 352         int *mode = (int *)opt->value;
 353         const struct mem_mode *m;
 354         char *s, *os = NULL, *p;
 355         int ret = -1;
 356
 357         if (unset)
 358                 return 0;
 359
 360         /* str may be NULL in case no arg is passed to -t */
 361         if (str) {
 362                 /* because str is read-only */
 363                 s = os = strdup(str);
 364                 if (!s)
 365                         return -1;
 366
 367                 /* reset mode */
 368                 *mode = 0;
 369
 370                 for (;;) {
 371                         p = strchr(s, ',');
 372                         if (p)
 373                                 *p = '\0';
 374
 375                         for (m = mem_modes; m->name; m++) {
 376                                 if (!strcasecmp(s, m->name))
 377                                         break;
 378                         }
 379                         if (!m->name) {
 380                                 fprintf(stderr, "unknown sampling op %s,"
 381                                             " check man page\n", s);
 382                                 goto error;
 383                         }
 384
 385                         *mode |= m->mode;
 386
 387                         if (!p)
 388                                 break;
 389
 390                         s = p + 1;
 391                 }
 392         }
 393         ret = 0;
 394
 395         if (*mode == 0)
 396                 *mode = MEM_OPERATION_LOAD;
 397 error:
 398         free(os);
 399         return ret;
 400 }
 401
 402 int cmd_mem(int argc, const char **argv)
 403 {
 404         struct stat st;
 405         struct perf_mem mem = {
 406                 .tool = {
 407                         .sample         = process_sample_event,
 408                         .mmap           = perf_event__process_mmap,
 409                         .mmap2          = perf_event__process_mmap2,
 410                         .comm           = perf_event__process_comm,
 411                         .lost           = perf_event__process_lost,
 412                         .fork           = perf_event__process_fork,
 413                         .build_id       = perf_event__process_build_id,
 414                         .namespaces     = perf_event__process_namespaces,
 415                         .ordered_events = true,
 416                 },
 417                 .input_name              = "perf.data",
 418                 /*
 419                  * default to both load an store sampling
 420                  */
 421                 .operation               = MEM_OPERATION_LOAD | MEM_OPERATION_STORE,
 422         };
 423         const struct option mem_options[] = {
 424         OPT_CALLBACK('t', "type", &mem.operation,
 425                    "type", "memory operations(load,store) Default load,store",
 426                     parse_mem_ops),
 427         OPT_BOOLEAN('D', "dump-raw-samples", &mem.dump_raw,
 428                     "dump raw samples in ASCII"),
 429         OPT_BOOLEAN('U', "hide-unresolved", &mem.hide_unresolved,
 430                     "Only display entries resolved to a symbol"),
 431         OPT_STRING('i', "input", &input_name, "file",
 432                    "input file name"),
 433         OPT_STRING('C', "cpu", &mem.cpu_list, "cpu",
 434                    "list of cpus to profile"),
 435         OPT_STRING_NOEMPTY('x', "field-separator", &symbol_conf.field_sep,
 436                    "separator",
 437                    "separator for columns, no spaces will be added"
 438                    " between columns '.' is reserved."),
 439         OPT_BOOLEAN('f', "force", &mem.force, "don't complain, do it"),
 440         OPT_BOOLEAN('p', "phys-data", &mem.phys_addr, "Record/Report sample physical addresses"),
 441         OPT_END()
 442         };
 443         const char *const mem_subcommands[] = { "record", "report", NULL };
 444         const char *mem_usage[] = {
 445                 NULL,
 446                 NULL
 447         };
 448
 449         argc = parse_options_subcommand(argc, argv, mem_options, mem_subcommands,
 450                                         mem_usage, PARSE_OPT_KEEP_UNKNOWN);
 451
 452         if (!argc || !(strncmp(argv[0], "rec", 3) || mem.operation))
 453                 usage_with_options(mem_usage, mem_options);
 454
 455         if (!mem.input_name || !strlen(mem.input_name)) {
 456                 if (!fstat(STDIN_FILENO, &st) && S_ISFIFO(st.st_mode))
 457                         mem.input_name = "-";
 458                 else
 459                         mem.input_name = "perf.data";
 460         }
 461
 462         if (!strncmp(argv[0], "rec", 3))
 463                 return __cmd_record(argc, argv, &mem);
 464         else if (!strncmp(argv[0], "rep", 3))
 465                 return report_events(argc, argv, &mem);
 466         else
 467                 usage_with_options(mem_usage, mem_options);
 468
 469         return 0;
 470 }