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 "util/pmu.h"
  22 #include "util/pmu-hybrid.h"
  23 #include <linux/err.h>
  24
  25 #define MEM_OPERATION_LOAD      0x1
  26 #define MEM_OPERATION_STORE     0x2
  27
  28 struct perf_mem {
  29         struct perf_tool        tool;
  30         char const              *input_name;
  31         bool                    hide_unresolved;
  32         bool                    dump_raw;
  33         bool                    force;
  34         bool                    phys_addr;
  35         bool                    data_page_size;
  36         int                     operation;
  37         const char              *cpu_list;
  38         DECLARE_BITMAP(cpu_bitmap, MAX_NR_CPUS);
  39 };
  40
  41 static int parse_record_events(const struct option *opt,
  42                                const char *str, int unset __maybe_unused)
  43 {
  44         struct perf_mem *mem = *(struct perf_mem **)opt->value;
  45
  46         if (!strcmp(str, "list")) {
  47                 perf_mem_events__list();
  48                 exit(0);
  49         }
  50         if (perf_mem_events__parse(str))
  51                 exit(-1);
  52
  53         mem->operation = 0;
  54         return 0;
  55 }
  56
  57 static const char * const __usage[] = {
  58         "perf mem record [<options>] [<command>]",
  59         "perf mem record [<options>] -- <command> [<options>]",
  60         NULL
  61 };
  62
  63 static const char * const *record_mem_usage = __usage;
  64
  65 static int __cmd_record(int argc, const char **argv, struct perf_mem *mem)
  66 {
  67         int rec_argc, i = 0, j, tmp_nr = 0;
  68         int start, end;
  69         const char **rec_argv;
  70         char **rec_tmp;
  71         int ret;
  72         bool all_user = false, all_kernel = false;
  73         struct perf_mem_event *e;
  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         if (perf_mem_events__init()) {
  87                 pr_err("failed: memory events not supported\n");
  88                 return -1;
  89         }
  90
  91         argc = parse_options(argc, argv, options, record_mem_usage,
  92                              PARSE_OPT_KEEP_UNKNOWN);
  93
  94         if (!perf_pmu__has_hybrid())
  95                 rec_argc = argc + 9; /* max number of arguments */
  96         else
  97                 rec_argc = argc + 9 * perf_pmu__hybrid_pmu_num();
  98
  99         rec_argv = calloc(rec_argc + 1, sizeof(char *));
 100         if (!rec_argv)
 101                 return -1;
 102
 103         /*
 104          * Save the allocated event name strings.
 105          */
 106         rec_tmp = calloc(rec_argc + 1, sizeof(char *));
 107         if (!rec_tmp) {
 108                 free(rec_argv);
 109                 return -1;
 110         }
 111
 112         rec_argv[i++] = "record";
 113
 114         e = perf_mem_events__ptr(PERF_MEM_EVENTS__LOAD_STORE);
 115
 116         /*
 117          * The load and store operations are required, use the event
 118          * PERF_MEM_EVENTS__LOAD_STORE if it is supported.
 119          */
 120         if (e->tag &&
 121             (mem->operation & MEM_OPERATION_LOAD) &&
 122             (mem->operation & MEM_OPERATION_STORE)) {
 123                 e->record = true;
 124         } else {
 125                 if (mem->operation & MEM_OPERATION_LOAD) {
 126                         e = perf_mem_events__ptr(PERF_MEM_EVENTS__LOAD);
 127                         e->record = true;
 128                 }
 129
 130                 if (mem->operation & MEM_OPERATION_STORE) {
 131                         e = perf_mem_events__ptr(PERF_MEM_EVENTS__STORE);
 132                         e->record = true;
 133                 }
 134         }
 135
 136         e = perf_mem_events__ptr(PERF_MEM_EVENTS__LOAD);
 137         if (e->record)
 138                 rec_argv[i++] = "-W";
 139
 140         rec_argv[i++] = "-d";
 141
 142         if (mem->phys_addr)
 143                 rec_argv[i++] = "--phys-data";
 144
 145         if (mem->data_page_size)
 146                 rec_argv[i++] = "--data-page-size";
 147
 148         start = i;
 149         ret = perf_mem_events__record_args(rec_argv, &i, rec_tmp, &tmp_nr);
 150         if (ret)
 151                 goto out;
 152         end = i;
 153
 154         if (all_user)
 155                 rec_argv[i++] = "--all-user";
 156
 157         if (all_kernel)
 158                 rec_argv[i++] = "--all-kernel";
 159
 160         for (j = 0; j < argc; j++, i++)
 161                 rec_argv[i] = argv[j];
 162
 163         if (verbose > 0) {
 164                 pr_debug("calling: record ");
 165
 166                 for (j = start; j < end; j++)
 167                         pr_debug("%s ", rec_argv[j]);
 168
 169                 pr_debug("\n");
 170         }
 171
 172         ret = cmd_record(i, rec_argv);
 173 out:
 174         for (i = 0; i < tmp_nr; i++)
 175                 free(rec_tmp[i]);
 176
 177         free(rec_tmp);
 178         free(rec_argv);
 179         return ret;
 180 }
 181
 182 static int
 183 dump_raw_samples(struct perf_tool *tool,
 184                  union perf_event *event,
 185                  struct perf_sample *sample,
 186                  struct machine *machine)
 187 {
 188         struct perf_mem *mem = container_of(tool, struct perf_mem, tool);
 189         struct addr_location al;
 190         const char *fmt, *field_sep;
 191         char str[PAGE_SIZE_NAME_LEN];
 192
 193         if (machine__resolve(machine, &al, sample) < 0) {
 194                 fprintf(stderr, "problem processing %d event, skipping it.\n",
 195                                 event->header.type);
 196                 return -1;
 197         }
 198
 199         if (al.filtered || (mem->hide_unresolved && al.sym == NULL))
 200                 goto out_put;
 201
 202         if (al.map != NULL)
 203                 al.map->dso->hit = 1;
 204
 205         field_sep = symbol_conf.field_sep;
 206         if (field_sep) {
 207                 fmt = "%d%s%d%s0x%"PRIx64"%s0x%"PRIx64"%s";
 208         } else {
 209                 fmt = "%5d%s%5d%s0x%016"PRIx64"%s0x016%"PRIx64"%s";
 210                 symbol_conf.field_sep = " ";
 211         }
 212         printf(fmt,
 213                 sample->pid,
 214                 symbol_conf.field_sep,
 215                 sample->tid,
 216                 symbol_conf.field_sep,
 217                 sample->ip,
 218                 symbol_conf.field_sep,
 219                 sample->addr,
 220                 symbol_conf.field_sep);
 221
 222         if (mem->phys_addr) {
 223                 printf("0x%016"PRIx64"%s",
 224                         sample->phys_addr,
 225                         symbol_conf.field_sep);
 226         }
 227
 228         if (mem->data_page_size) {
 229                 printf("%s%s",
 230                         get_page_size_name(sample->data_page_size, str),
 231                         symbol_conf.field_sep);
 232         }
 233
 234         if (field_sep)
 235                 fmt = "%"PRIu64"%s0x%"PRIx64"%s%s:%s\n";
 236         else
 237                 fmt = "%5"PRIu64"%s0x%06"PRIx64"%s%s:%s\n";
 238
 239         printf(fmt,
 240                 sample->weight,
 241                 symbol_conf.field_sep,
 242                 sample->data_src,
 243                 symbol_conf.field_sep,
 244                 al.map ? (al.map->dso ? al.map->dso->long_name : "???") : "???",
 245                 al.sym ? al.sym->name : "???");
 246 out_put:
 247         addr_location__put(&al);
 248         return 0;
 249 }
 250
 251 static int process_sample_event(struct perf_tool *tool,
 252                                 union perf_event *event,
 253                                 struct perf_sample *sample,
 254                                 struct evsel *evsel __maybe_unused,
 255                                 struct machine *machine)
 256 {
 257         return dump_raw_samples(tool, event, sample, machine);
 258 }
 259
 260 static int report_raw_events(struct perf_mem *mem)
 261 {
 262         struct itrace_synth_opts itrace_synth_opts = {
 263                 .set = true,
 264                 .mem = true,    /* Only enable memory event */
 265                 .default_no_sample = true,
 266         };
 267
 268         struct perf_data data = {
 269                 .path  = input_name,
 270                 .mode  = PERF_DATA_MODE_READ,
 271                 .force = mem->force,
 272         };
 273         int ret;
 274         struct perf_session *session = perf_session__new(&data, &mem->tool);
 275
 276         if (IS_ERR(session))
 277                 return PTR_ERR(session);
 278
 279         session->itrace_synth_opts = &itrace_synth_opts;
 280
 281         if (mem->cpu_list) {
 282                 ret = perf_session__cpu_bitmap(session, mem->cpu_list,
 283                                                mem->cpu_bitmap);
 284                 if (ret < 0)
 285                         goto out_delete;
 286         }
 287
 288         ret = symbol__init(&session->header.env);
 289         if (ret < 0)
 290                 goto out_delete;
 291
 292         printf("# PID, TID, IP, ADDR, ");
 293
 294         if (mem->phys_addr)
 295                 printf("PHYS ADDR, ");
 296
 297         if (mem->data_page_size)
 298                 printf("DATA PAGE SIZE, ");
 299
 300         printf("LOCAL WEIGHT, DSRC, SYMBOL\n");
 301
 302         ret = perf_session__process_events(session);
 303
 304 out_delete:
 305         perf_session__delete(session);
 306         return ret;
 307 }
 308 static char *get_sort_order(struct perf_mem *mem)
 309 {
 310         bool has_extra_options = (mem->phys_addr | mem->data_page_size) ? true : false;
 311         char sort[128];
 312
 313         /*
 314          * there is no weight (cost) associated with stores, so don't print
 315          * the column
 316          */
 317         if (!(mem->operation & MEM_OPERATION_LOAD)) {
 318                 strcpy(sort, "--sort=mem,sym,dso,symbol_daddr,"
 319                              "dso_daddr,tlb,locked");
 320         } else if (has_extra_options) {
 321                 strcpy(sort, "--sort=local_weight,mem,sym,dso,symbol_daddr,"
 322                              "dso_daddr,snoop,tlb,locked,blocked");
 323         } else
 324                 return NULL;
 325
 326         if (mem->phys_addr)
 327                 strcat(sort, ",phys_daddr");
 328
 329         if (mem->data_page_size)
 330                 strcat(sort, ",data_page_size");
 331
 332         return strdup(sort);
 333 }
 334
 335 static int report_events(int argc, const char **argv, struct perf_mem *mem)
 336 {
 337         const char **rep_argv;
 338         int ret, i = 0, j, rep_argc;
 339         char *new_sort_order;
 340
 341         if (mem->dump_raw)
 342                 return report_raw_events(mem);
 343
 344         rep_argc = argc + 3;
 345         rep_argv = calloc(rep_argc + 1, sizeof(char *));
 346         if (!rep_argv)
 347                 return -1;
 348
 349         rep_argv[i++] = "report";
 350         rep_argv[i++] = "--mem-mode";
 351         rep_argv[i++] = "-n"; /* display number of samples */
 352
 353         new_sort_order = get_sort_order(mem);
 354         if (new_sort_order)
 355                 rep_argv[i++] = new_sort_order;
 356
 357         for (j = 1; j < argc; j++, i++)
 358                 rep_argv[i] = argv[j];
 359
 360         ret = cmd_report(i, rep_argv);
 361         free(rep_argv);
 362         return ret;
 363 }
 364
 365 struct mem_mode {
 366         const char *name;
 367         int mode;
 368 };
 369
 370 #define MEM_OPT(n, m) \
 371         { .name = n, .mode = (m) }
 372
 373 #define MEM_END { .name = NULL }
 374
 375 static const struct mem_mode mem_modes[]={
 376         MEM_OPT("load", MEM_OPERATION_LOAD),
 377         MEM_OPT("store", MEM_OPERATION_STORE),
 378         MEM_END
 379 };
 380
 381 static int
 382 parse_mem_ops(const struct option *opt, const char *str, int unset)
 383 {
 384         int *mode = (int *)opt->value;
 385         const struct mem_mode *m;
 386         char *s, *os = NULL, *p;
 387         int ret = -1;
 388
 389         if (unset)
 390                 return 0;
 391
 392         /* str may be NULL in case no arg is passed to -t */
 393         if (str) {
 394                 /* because str is read-only */
 395                 s = os = strdup(str);
 396                 if (!s)
 397                         return -1;
 398
 399                 /* reset mode */
 400                 *mode = 0;
 401
 402                 for (;;) {
 403                         p = strchr(s, ',');
 404                         if (p)
 405                                 *p = '\0';
 406
 407                         for (m = mem_modes; m->name; m++) {
 408                                 if (!strcasecmp(s, m->name))
 409                                         break;
 410                         }
 411                         if (!m->name) {
 412                                 fprintf(stderr, "unknown sampling op %s,"
 413                                             " check man page\n", s);
 414                                 goto error;
 415                         }
 416
 417                         *mode |= m->mode;
 418
 419                         if (!p)
 420                                 break;
 421
 422                         s = p + 1;
 423                 }
 424         }
 425         ret = 0;
 426
 427         if (*mode == 0)
 428                 *mode = MEM_OPERATION_LOAD;
 429 error:
 430         free(os);
 431         return ret;
 432 }
 433
 434 int cmd_mem(int argc, const char **argv)
 435 {
 436         struct stat st;
 437         struct perf_mem mem = {
 438                 .tool = {
 439                         .sample         = process_sample_event,
 440                         .mmap           = perf_event__process_mmap,
 441                         .mmap2          = perf_event__process_mmap2,
 442                         .comm           = perf_event__process_comm,
 443                         .lost           = perf_event__process_lost,
 444                         .fork           = perf_event__process_fork,
 445                         .attr           = perf_event__process_attr,
 446                         .build_id       = perf_event__process_build_id,
 447                         .namespaces     = perf_event__process_namespaces,
 448                         .auxtrace_info  = perf_event__process_auxtrace_info,
 449                         .auxtrace       = perf_event__process_auxtrace,
 450                         .auxtrace_error = perf_event__process_auxtrace_error,
 451                         .ordered_events = true,
 452                 },
 453                 .input_name              = "perf.data",
 454                 /*
 455                  * default to both load an store sampling
 456                  */
 457                 .operation               = MEM_OPERATION_LOAD | MEM_OPERATION_STORE,
 458         };
 459         const struct option mem_options[] = {
 460         OPT_CALLBACK('t', "type", &mem.operation,
 461                    "type", "memory operations(load,store) Default load,store",
 462                     parse_mem_ops),
 463         OPT_BOOLEAN('D', "dump-raw-samples", &mem.dump_raw,
 464                     "dump raw samples in ASCII"),
 465         OPT_BOOLEAN('U', "hide-unresolved", &mem.hide_unresolved,
 466                     "Only display entries resolved to a symbol"),
 467         OPT_STRING('i', "input", &input_name, "file",
 468                    "input file name"),
 469         OPT_STRING('C', "cpu", &mem.cpu_list, "cpu",
 470                    "list of cpus to profile"),
 471         OPT_STRING_NOEMPTY('x', "field-separator", &symbol_conf.field_sep,
 472                    "separator",
 473                    "separator for columns, no spaces will be added"
 474                    " between columns '.' is reserved."),
 475         OPT_BOOLEAN('f', "force", &mem.force, "don't complain, do it"),
 476         OPT_BOOLEAN('p', "phys-data", &mem.phys_addr, "Record/Report sample physical addresses"),
 477         OPT_BOOLEAN(0, "data-page-size", &mem.data_page_size, "Record/Report sample data address page size"),
 478         OPT_END()
 479         };
 480         const char *const mem_subcommands[] = { "record", "report", NULL };
 481         const char *mem_usage[] = {
 482                 NULL,
 483                 NULL
 484         };
 485
 486         argc = parse_options_subcommand(argc, argv, mem_options, mem_subcommands,
 487                                         mem_usage, PARSE_OPT_KEEP_UNKNOWN);
 488
 489         if (!argc || !(strncmp(argv[0], "rec", 3) || mem.operation))
 490                 usage_with_options(mem_usage, mem_options);
 491
 492         if (!mem.input_name || !strlen(mem.input_name)) {
 493                 if (!fstat(STDIN_FILENO, &st) && S_ISFIFO(st.st_mode))
 494                         mem.input_name = "-";
 495                 else
 496                         mem.input_name = "perf.data";
 497         }
 498
 499         if (!strncmp(argv[0], "rec", 3))
 500                 return __cmd_record(argc, argv, &mem);
 501         else if (!strncmp(argv[0], "rep", 3))
 502                 return report_events(argc, argv, &mem);
 503         else
 504                 usage_with_options(mem_usage, mem_options);
 505
 506         return 0;
 507 }