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