tools/perf/builtin-trace.c

   1 #include <traceevent/event-parse.h>
   2 #include "builtin.h"
   3 #include "util/color.h"
   4 #include "util/debug.h"
   5 #include "util/evlist.h"
   6 #include "util/machine.h"
   7 #include "util/session.h"
   8 #include "util/thread.h"
   9 #include "util/parse-options.h"
  10 #include "util/strlist.h"
  11 #include "util/intlist.h"
  12 #include "util/thread_map.h"
  13
  14 #include <libaudit.h>
  15 #include <stdlib.h>
  16 #include <sys/mman.h>
  17
  18 static size_t syscall_arg__scnprintf_hex(char *bf, size_t size, unsigned long arg)
  19 {
  20         return scnprintf(bf, size, "%#lx", arg);
  21 }
  22
  23 #define SCA_HEX syscall_arg__scnprintf_hex
  24
  25 static size_t syscall_arg__scnprintf_mmap_prot(char *bf, size_t size, unsigned long arg)
  26 {
  27         int printed = 0, prot = arg;
  28
  29         if (prot == PROT_NONE)
  30                 return scnprintf(bf, size, "NONE");
  31 #define P_MMAP_PROT(n) \
  32         if (prot & PROT_##n) { \
  33                 printed += scnprintf(bf + printed, size - printed, "%s%s", printed ? "|" : "", #n); \
  34                 prot &= ~PROT_##n; \
  35         }
  36
  37         P_MMAP_PROT(EXEC);
  38         P_MMAP_PROT(READ);
  39         P_MMAP_PROT(WRITE);
  40 #ifdef PROT_SEM
  41         P_MMAP_PROT(SEM);
  42 #endif
  43         P_MMAP_PROT(GROWSDOWN);
  44         P_MMAP_PROT(GROWSUP);
  45 #undef P_MMAP_PROT
  46
  47         if (prot)
  48                 printed += scnprintf(bf + printed, size - printed, "%s%#x", printed ? "|" : "", prot);
  49
  50         return printed;
  51 }
  52
  53 #define SCA_MMAP_PROT syscall_arg__scnprintf_mmap_prot
  54
  55 static size_t syscall_arg__scnprintf_mmap_flags(char *bf, size_t size, unsigned long arg)
  56 {
  57         int printed = 0, flags = arg;
  58
  59 #define P_MMAP_FLAG(n) \
  60         if (flags & MAP_##n) { \
  61                 printed += scnprintf(bf + printed, size - printed, "%s%s", printed ? "|" : "", #n); \
  62                 flags &= ~MAP_##n; \
  63         }
  64
  65         P_MMAP_FLAG(SHARED);
  66         P_MMAP_FLAG(PRIVATE);
  67         P_MMAP_FLAG(32BIT);
  68         P_MMAP_FLAG(ANONYMOUS);
  69         P_MMAP_FLAG(DENYWRITE);
  70         P_MMAP_FLAG(EXECUTABLE);
  71         P_MMAP_FLAG(FILE);
  72         P_MMAP_FLAG(FIXED);
  73         P_MMAP_FLAG(GROWSDOWN);
  74 #ifdef MAP_HUGETLB
  75         P_MMAP_FLAG(HUGETLB);
  76 #endif
  77         P_MMAP_FLAG(LOCKED);
  78         P_MMAP_FLAG(NONBLOCK);
  79         P_MMAP_FLAG(NORESERVE);
  80         P_MMAP_FLAG(POPULATE);
  81         P_MMAP_FLAG(STACK);
  82 #ifdef MAP_UNINITIALIZED
  83         P_MMAP_FLAG(UNINITIALIZED);
  84 #endif
  85 #undef P_MMAP_FLAG
  86
  87         if (flags)
  88                 printed += scnprintf(bf + printed, size - printed, "%s%#x", printed ? "|" : "", flags);
  89
  90         return printed;
  91 }
  92
  93 #define SCA_MMAP_FLAGS syscall_arg__scnprintf_mmap_flags
  94
  95 static size_t syscall_arg__scnprintf_madvise_behavior(char *bf, size_t size, unsigned long arg)
  96 {
  97         int behavior = arg;
  98
  99         switch (behavior) {
 100 #define P_MADV_BHV(n) case MADV_##n: return scnprintf(bf, size, #n)
 101         P_MADV_BHV(NORMAL);
 102         P_MADV_BHV(RANDOM);
 103         P_MADV_BHV(SEQUENTIAL);
 104         P_MADV_BHV(WILLNEED);
 105         P_MADV_BHV(DONTNEED);
 106         P_MADV_BHV(REMOVE);
 107         P_MADV_BHV(DONTFORK);
 108         P_MADV_BHV(DOFORK);
 109         P_MADV_BHV(HWPOISON);
 110 #ifdef MADV_SOFT_OFFLINE
 111         P_MADV_BHV(SOFT_OFFLINE);
 112 #endif
 113         P_MADV_BHV(MERGEABLE);
 114         P_MADV_BHV(UNMERGEABLE);
 115 #ifdef MADV_HUGEPAGE
 116         P_MADV_BHV(HUGEPAGE);
 117 #endif
 118 #ifdef MADV_NOHUGEPAGE
 119         P_MADV_BHV(NOHUGEPAGE);
 120 #endif
 121 #ifdef MADV_DONTDUMP
 122         P_MADV_BHV(DONTDUMP);
 123 #endif
 124 #ifdef MADV_DODUMP
 125         P_MADV_BHV(DODUMP);
 126 #endif
 127 #undef P_MADV_PHV
 128         default: break;
 129         }
 130
 131         return scnprintf(bf, size, "%#x", behavior);
 132 }
 133
 134 #define SCA_MADV_BHV syscall_arg__scnprintf_madvise_behavior
 135
 136 static struct syscall_fmt {
 137         const char *name;
 138         const char *alias;
 139         size_t     (*arg_scnprintf[6])(char *bf, size_t size, unsigned long arg);
 140         bool       errmsg;
 141         bool       timeout;
 142         bool       hexret;
 143 } syscall_fmts[] = {
 144         { .name     = "access",     .errmsg = true, },
 145         { .name     = "arch_prctl", .errmsg = true, .alias = "prctl", },
 146         { .name     = "brk",        .hexret = true,
 147           .arg_scnprintf = { [0] = SCA_HEX, /* brk */ }, },
 148         { .name     = "mmap",       .hexret = true, },
 149         { .name     = "connect",    .errmsg = true, },
 150         { .name     = "fstat",      .errmsg = true, .alias = "newfstat", },
 151         { .name     = "fstatat",    .errmsg = true, .alias = "newfstatat", },
 152         { .name     = "futex",      .errmsg = true, },
 153         { .name     = "ioctl",      .errmsg = true,
 154           .arg_scnprintf = { [2] = SCA_HEX, /* arg */ }, },
 155         { .name     = "lstat",      .errmsg = true, .alias = "newlstat", },
 156         { .name     = "madvise",    .errmsg = true,
 157           .arg_scnprintf = { [0] = SCA_HEX,      /* start */
 158                              [2] = SCA_MADV_BHV, /* behavior */ }, },
 159         { .name     = "mmap",       .hexret = true,
 160           .arg_scnprintf = { [0] = SCA_HEX,       /* addr */
 161                              [2] = SCA_MMAP_PROT, /* prot */
 162                              [3] = SCA_MMAP_FLAGS, /* flags */ }, },
 163         { .name     = "mprotect",   .errmsg = true,
 164           .arg_scnprintf = { [0] = SCA_HEX, /* start */
 165                              [2] = SCA_MMAP_PROT, /* prot */ }, },
 166         { .name     = "mremap",     .hexret = true,
 167           .arg_scnprintf = { [0] = SCA_HEX, /* addr */
 168                              [4] = SCA_HEX, /* new_addr */ }, },
 169         { .name     = "munmap",     .errmsg = true,
 170           .arg_scnprintf = { [0] = SCA_HEX, /* addr */ }, },
 171         { .name     = "open",       .errmsg = true, },
 172         { .name     = "poll",       .errmsg = true, .timeout = true, },
 173         { .name     = "ppoll",      .errmsg = true, .timeout = true, },
 174         { .name     = "pread",      .errmsg = true, .alias = "pread64", },
 175         { .name     = "pwrite",     .errmsg = true, .alias = "pwrite64", },
 176         { .name     = "read",       .errmsg = true, },
 177         { .name     = "recvfrom",   .errmsg = true, },
 178         { .name     = "select",     .errmsg = true, .timeout = true, },
 179         { .name     = "socket",     .errmsg = true, },
 180         { .name     = "stat",       .errmsg = true, .alias = "newstat", },
 181         { .name     = "uname",      .errmsg = true, .alias = "newuname", },
 182 };
 183
 184 static int syscall_fmt__cmp(const void *name, const void *fmtp)
 185 {
 186         const struct syscall_fmt *fmt = fmtp;
 187         return strcmp(name, fmt->name);
 188 }
 189
 190 static struct syscall_fmt *syscall_fmt__find(const char *name)
 191 {
 192         const int nmemb = ARRAY_SIZE(syscall_fmts);
 193         return bsearch(name, syscall_fmts, nmemb, sizeof(struct syscall_fmt), syscall_fmt__cmp);
 194 }
 195
 196 struct syscall {
 197         struct event_format *tp_format;
 198         const char          *name;
 199         bool                filtered;
 200         struct syscall_fmt  *fmt;
 201         size_t              (**arg_scnprintf)(char *bf, size_t size, unsigned long arg);
 202 };
 203
 204 static size_t fprintf_duration(unsigned long t, FILE *fp)
 205 {
 206         double duration = (double)t / NSEC_PER_MSEC;
 207         size_t printed = fprintf(fp, "(");
 208
 209         if (duration >= 1.0)
 210                 printed += color_fprintf(fp, PERF_COLOR_RED, "%6.3f ms", duration);
 211         else if (duration >= 0.01)
 212                 printed += color_fprintf(fp, PERF_COLOR_YELLOW, "%6.3f ms", duration);
 213         else
 214                 printed += color_fprintf(fp, PERF_COLOR_NORMAL, "%6.3f ms", duration);
 215         return printed + fprintf(fp, "): ");
 216 }
 217
 218 struct thread_trace {
 219         u64               entry_time;
 220         u64               exit_time;
 221         bool              entry_pending;
 222         unsigned long     nr_events;
 223         char              *entry_str;
 224         double            runtime_ms;
 225 };
 226
 227 static struct thread_trace *thread_trace__new(void)
 228 {
 229         return zalloc(sizeof(struct thread_trace));
 230 }
 231
 232 static struct thread_trace *thread__trace(struct thread *thread, FILE *fp)
 233 {
 234         struct thread_trace *ttrace;
 235
 236         if (thread == NULL)
 237                 goto fail;
 238
 239         if (thread->priv == NULL)
 240                 thread->priv = thread_trace__new();
 241
 242         if (thread->priv == NULL)
 243                 goto fail;
 244
 245         ttrace = thread->priv;
 246         ++ttrace->nr_events;
 247
 248         return ttrace;
 249 fail:
 250         color_fprintf(fp, PERF_COLOR_RED,
 251                       "WARNING: not enough memory, dropping samples!\n");
 252         return NULL;
 253 }
 254
 255 struct trace {
 256         struct perf_tool        tool;
 257         int                     audit_machine;
 258         struct {
 259                 int             max;
 260                 struct syscall  *table;
 261         } syscalls;
 262         struct perf_record_opts opts;
 263         struct machine          host;
 264         u64                     base_time;
 265         FILE                    *output;
 266         unsigned long           nr_events;
 267         struct strlist          *ev_qualifier;
 268         bool                    not_ev_qualifier;
 269         struct intlist          *tid_list;
 270         struct intlist          *pid_list;
 271         bool                    sched;
 272         bool                    multiple_threads;
 273         double                  duration_filter;
 274         double                  runtime_ms;
 275 };
 276
 277 static bool trace__filter_duration(struct trace *trace, double t)
 278 {
 279         return t < (trace->duration_filter * NSEC_PER_MSEC);
 280 }
 281
 282 static size_t trace__fprintf_tstamp(struct trace *trace, u64 tstamp, FILE *fp)
 283 {
 284         double ts = (double)(tstamp - trace->base_time) / NSEC_PER_MSEC;
 285
 286         return fprintf(fp, "%10.3f ", ts);
 287 }
 288
 289 static bool done = false;
 290
 291 static void sig_handler(int sig __maybe_unused)
 292 {
 293         done = true;
 294 }
 295
 296 static size_t trace__fprintf_entry_head(struct trace *trace, struct thread *thread,
 297                                         u64 duration, u64 tstamp, FILE *fp)
 298 {
 299         size_t printed = trace__fprintf_tstamp(trace, tstamp, fp);
 300         printed += fprintf_duration(duration, fp);
 301
 302         if (trace->multiple_threads)
 303                 printed += fprintf(fp, "%d ", thread->tid);
 304
 305         return printed;
 306 }
 307
 308 static int trace__process_event(struct trace *trace, struct machine *machine,
 309                                 union perf_event *event)
 310 {
 311         int ret = 0;
 312
 313         switch (event->header.type) {
 314         case PERF_RECORD_LOST:
 315                 color_fprintf(trace->output, PERF_COLOR_RED,
 316                               "LOST %" PRIu64 " events!\n", event->lost.lost);
 317                 ret = machine__process_lost_event(machine, event);
 318         default:
 319                 ret = machine__process_event(machine, event);
 320                 break;
 321         }
 322
 323         return ret;
 324 }
 325
 326 static int trace__tool_process(struct perf_tool *tool,
 327                                union perf_event *event,
 328                                struct perf_sample *sample __maybe_unused,
 329                                struct machine *machine)
 330 {
 331         struct trace *trace = container_of(tool, struct trace, tool);
 332         return trace__process_event(trace, machine, event);
 333 }
 334
 335 static int trace__symbols_init(struct trace *trace, struct perf_evlist *evlist)
 336 {
 337         int err = symbol__init();
 338
 339         if (err)
 340                 return err;
 341
 342         machine__init(&trace->host, "", HOST_KERNEL_ID);
 343         machine__create_kernel_maps(&trace->host);
 344
 345         if (perf_target__has_task(&trace->opts.target)) {
 346                 err = perf_event__synthesize_thread_map(&trace->tool, evlist->threads,
 347                                                         trace__tool_process,
 348                                                         &trace->host);
 349         } else {
 350                 err = perf_event__synthesize_threads(&trace->tool, trace__tool_process,
 351                                                      &trace->host);
 352         }
 353
 354         if (err)
 355                 symbol__exit();
 356
 357         return err;
 358 }
 359
 360 static int syscall__set_arg_fmts(struct syscall *sc)
 361 {
 362         struct format_field *field;
 363         int idx = 0;
 364
 365         sc->arg_scnprintf = calloc(sc->tp_format->format.nr_fields - 1, sizeof(void *));
 366         if (sc->arg_scnprintf == NULL)
 367                 return -1;
 368
 369         for (field = sc->tp_format->format.fields->next; field; field = field->next) {
 370                 if (sc->fmt && sc->fmt->arg_scnprintf[idx])
 371                         sc->arg_scnprintf[idx] = sc->fmt->arg_scnprintf[idx];
 372                 else if (field->flags & FIELD_IS_POINTER)
 373                         sc->arg_scnprintf[idx] = syscall_arg__scnprintf_hex;
 374                 ++idx;
 375         }
 376
 377         return 0;
 378 }
 379
 380 static int trace__read_syscall_info(struct trace *trace, int id)
 381 {
 382         char tp_name[128];
 383         struct syscall *sc;
 384         const char *name = audit_syscall_to_name(id, trace->audit_machine);
 385
 386         if (name == NULL)
 387                 return -1;
 388
 389         if (id > trace->syscalls.max) {
 390                 struct syscall *nsyscalls = realloc(trace->syscalls.table, (id + 1) * sizeof(*sc));
 391
 392                 if (nsyscalls == NULL)
 393                         return -1;
 394
 395                 if (trace->syscalls.max != -1) {
 396                         memset(nsyscalls + trace->syscalls.max + 1, 0,
 397                                (id - trace->syscalls.max) * sizeof(*sc));
 398                 } else {
 399                         memset(nsyscalls, 0, (id + 1) * sizeof(*sc));
 400                 }
 401
 402                 trace->syscalls.table = nsyscalls;
 403                 trace->syscalls.max   = id;
 404         }
 405
 406         sc = trace->syscalls.table + id;
 407         sc->name = name;
 408
 409         if (trace->ev_qualifier) {
 410                 bool in = strlist__find(trace->ev_qualifier, name) != NULL;
 411
 412                 if (!(in ^ trace->not_ev_qualifier)) {
 413                         sc->filtered = true;
 414                         /*
 415                          * No need to do read tracepoint information since this will be
 416                          * filtered out.
 417                          */
 418                         return 0;
 419                 }
 420         }
 421
 422         sc->fmt  = syscall_fmt__find(sc->name);
 423
 424         snprintf(tp_name, sizeof(tp_name), "sys_enter_%s", sc->name);
 425         sc->tp_format = event_format__new("syscalls", tp_name);
 426
 427         if (sc->tp_format == NULL && sc->fmt && sc->fmt->alias) {
 428                 snprintf(tp_name, sizeof(tp_name), "sys_enter_%s", sc->fmt->alias);
 429                 sc->tp_format = event_format__new("syscalls", tp_name);
 430         }
 431
 432         if (sc->tp_format == NULL)
 433                 return -1;
 434
 435         return syscall__set_arg_fmts(sc);
 436 }
 437
 438 static size_t syscall__scnprintf_args(struct syscall *sc, char *bf, size_t size,
 439                                       unsigned long *args)
 440 {
 441         int i = 0;
 442         size_t printed = 0;
 443
 444         if (sc->tp_format != NULL) {
 445                 struct format_field *field;
 446
 447                 for (field = sc->tp_format->format.fields->next; field; field = field->next) {
 448                         printed += scnprintf(bf + printed, size - printed,
 449                                              "%s%s: ", printed ? ", " : "", field->name);
 450
 451                         if (sc->arg_scnprintf && sc->arg_scnprintf[i])
 452                                 printed += sc->arg_scnprintf[i](bf + printed, size - printed, args[i]);
 453                         else
 454                                 printed += scnprintf(bf + printed, size - printed,
 455                                                      "%ld", args[i]);
 456                        ++i;
 457                 }
 458         } else {
 459                 while (i < 6) {
 460                         printed += scnprintf(bf + printed, size - printed,
 461                                              "%sarg%d: %ld",
 462                                              printed ? ", " : "", i, args[i]);
 463                         ++i;
 464                 }
 465         }
 466
 467         return printed;
 468 }
 469
 470 typedef int (*tracepoint_handler)(struct trace *trace, struct perf_evsel *evsel,
 471                                   struct perf_sample *sample);
 472
 473 static struct syscall *trace__syscall_info(struct trace *trace,
 474                                            struct perf_evsel *evsel,
 475                                            struct perf_sample *sample)
 476 {
 477         int id = perf_evsel__intval(evsel, sample, "id");
 478
 479         if (id < 0) {
 480
 481                 /*
 482                  * XXX: Noticed on x86_64, reproduced as far back as 3.0.36, haven't tried
 483                  * before that, leaving at a higher verbosity level till that is
 484                  * explained. Reproduced with plain ftrace with:
 485                  *
 486                  * echo 1 > /t/events/raw_syscalls/sys_exit/enable
 487                  * grep "NR -1 " /t/trace_pipe
 488                  *
 489                  * After generating some load on the machine.
 490                  */
 491                 if (verbose > 1) {
 492                         static u64 n;
 493                         fprintf(trace->output, "Invalid syscall %d id, skipping (%s, %" PRIu64 ") ...\n",
 494                                 id, perf_evsel__name(evsel), ++n);
 495                 }
 496                 return NULL;
 497         }
 498
 499         if ((id > trace->syscalls.max || trace->syscalls.table[id].name == NULL) &&
 500             trace__read_syscall_info(trace, id))
 501                 goto out_cant_read;
 502
 503         if ((id > trace->syscalls.max || trace->syscalls.table[id].name == NULL))
 504                 goto out_cant_read;
 505
 506         return &trace->syscalls.table[id];
 507
 508 out_cant_read:
 509         if (verbose) {
 510                 fprintf(trace->output, "Problems reading syscall %d", id);
 511                 if (id <= trace->syscalls.max && trace->syscalls.table[id].name != NULL)
 512                         fprintf(trace->output, "(%s)", trace->syscalls.table[id].name);
 513                 fputs(" information\n", trace->output);
 514         }
 515         return NULL;
 516 }
 517
 518 static int trace__sys_enter(struct trace *trace, struct perf_evsel *evsel,
 519                             struct perf_sample *sample)
 520 {
 521         char *msg;
 522         void *args;
 523         size_t printed = 0;
 524         struct thread *thread;
 525         struct syscall *sc = trace__syscall_info(trace, evsel, sample);
 526         struct thread_trace *ttrace;
 527
 528         if (sc == NULL)
 529                 return -1;
 530
 531         if (sc->filtered)
 532                 return 0;
 533
 534         thread = machine__findnew_thread(&trace->host, sample->pid,
 535                                          sample->tid);
 536         ttrace = thread__trace(thread, trace->output);
 537         if (ttrace == NULL)
 538                 return -1;
 539
 540         args = perf_evsel__rawptr(evsel, sample, "args");
 541         if (args == NULL) {
 542                 fprintf(trace->output, "Problems reading syscall arguments\n");
 543                 return -1;
 544         }
 545
 546         ttrace = thread->priv;
 547
 548         if (ttrace->entry_str == NULL) {
 549                 ttrace->entry_str = malloc(1024);
 550                 if (!ttrace->entry_str)
 551                         return -1;
 552         }
 553
 554         ttrace->entry_time = sample->time;
 555         msg = ttrace->entry_str;
 556         printed += scnprintf(msg + printed, 1024 - printed, "%s(", sc->name);
 557
 558         printed += syscall__scnprintf_args(sc, msg + printed, 1024 - printed,  args);
 559
 560         if (!strcmp(sc->name, "exit_group") || !strcmp(sc->name, "exit")) {
 561                 if (!trace->duration_filter) {
 562                         trace__fprintf_entry_head(trace, thread, 1, sample->time, trace->output);
 563                         fprintf(trace->output, "%-70s\n", ttrace->entry_str);
 564                 }
 565         } else
 566                 ttrace->entry_pending = true;
 567
 568         return 0;
 569 }
 570
 571 static int trace__sys_exit(struct trace *trace, struct perf_evsel *evsel,
 572                            struct perf_sample *sample)
 573 {
 574         int ret;
 575         u64 duration = 0;
 576         struct thread *thread;
 577         struct syscall *sc = trace__syscall_info(trace, evsel, sample);
 578         struct thread_trace *ttrace;
 579
 580         if (sc == NULL)
 581                 return -1;
 582
 583         if (sc->filtered)
 584                 return 0;
 585
 586         thread = machine__findnew_thread(&trace->host, sample->pid,
 587                                          sample->tid);
 588         ttrace = thread__trace(thread, trace->output);
 589         if (ttrace == NULL)
 590                 return -1;
 591
 592         ret = perf_evsel__intval(evsel, sample, "ret");
 593
 594         ttrace = thread->priv;
 595
 596         ttrace->exit_time = sample->time;
 597
 598         if (ttrace->entry_time) {
 599                 duration = sample->time - ttrace->entry_time;
 600                 if (trace__filter_duration(trace, duration))
 601                         goto out;
 602         } else if (trace->duration_filter)
 603                 goto out;
 604
 605         trace__fprintf_entry_head(trace, thread, duration, sample->time, trace->output);
 606
 607         if (ttrace->entry_pending) {
 608                 fprintf(trace->output, "%-70s", ttrace->entry_str);
 609         } else {
 610                 fprintf(trace->output, " ... [");
 611                 color_fprintf(trace->output, PERF_COLOR_YELLOW, "continued");
 612                 fprintf(trace->output, "]: %s()", sc->name);
 613         }
 614
 615         if (sc->fmt == NULL) {
 616 signed_print:
 617                 fprintf(trace->output, ") = %d", ret);
 618         } else if (ret < 0 && sc->fmt->errmsg) {
 619                 char bf[256];
 620                 const char *emsg = strerror_r(-ret, bf, sizeof(bf)),
 621                            *e = audit_errno_to_name(-ret);
 622
 623                 fprintf(trace->output, ") = -1 %s %s", e, emsg);
 624         } else if (ret == 0 && sc->fmt->timeout)
 625                 fprintf(trace->output, ") = 0 Timeout");
 626         else if (sc->fmt->hexret)
 627                 fprintf(trace->output, ") = %#x", ret);
 628         else
 629                 goto signed_print;
 630
 631         fputc('\n', trace->output);
 632 out:
 633         ttrace->entry_pending = false;
 634
 635         return 0;
 636 }
 637
 638 static int trace__sched_stat_runtime(struct trace *trace, struct perf_evsel *evsel,
 639                                      struct perf_sample *sample)
 640 {
 641         u64 runtime = perf_evsel__intval(evsel, sample, "runtime");
 642         double runtime_ms = (double)runtime / NSEC_PER_MSEC;
 643         struct thread *thread = machine__findnew_thread(&trace->host,
 644                                                         sample->pid,
 645                                                         sample->tid);
 646         struct thread_trace *ttrace = thread__trace(thread, trace->output);
 647
 648         if (ttrace == NULL)
 649                 goto out_dump;
 650
 651         ttrace->runtime_ms += runtime_ms;
 652         trace->runtime_ms += runtime_ms;
 653         return 0;
 654
 655 out_dump:
 656         fprintf(trace->output, "%s: comm=%s,pid=%u,runtime=%" PRIu64 ",vruntime=%" PRIu64 ")\n",
 657                evsel->name,
 658                perf_evsel__strval(evsel, sample, "comm"),
 659                (pid_t)perf_evsel__intval(evsel, sample, "pid"),
 660                runtime,
 661                perf_evsel__intval(evsel, sample, "vruntime"));
 662         return 0;
 663 }
 664
 665 static bool skip_sample(struct trace *trace, struct perf_sample *sample)
 666 {
 667         if ((trace->pid_list && intlist__find(trace->pid_list, sample->pid)) ||
 668             (trace->tid_list && intlist__find(trace->tid_list, sample->tid)))
 669                 return false;
 670
 671         if (trace->pid_list || trace->tid_list)
 672                 return true;
 673
 674         return false;
 675 }
 676
 677 static int trace__process_sample(struct perf_tool *tool,
 678                                  union perf_event *event __maybe_unused,
 679                                  struct perf_sample *sample,
 680                                  struct perf_evsel *evsel,
 681                                  struct machine *machine __maybe_unused)
 682 {
 683         struct trace *trace = container_of(tool, struct trace, tool);
 684         int err = 0;
 685
 686         tracepoint_handler handler = evsel->handler.func;
 687
 688         if (skip_sample(trace, sample))
 689                 return 0;
 690
 691         if (trace->base_time == 0)
 692                 trace->base_time = sample->time;
 693
 694         if (handler)
 695                 handler(trace, evsel, sample);
 696
 697         return err;
 698 }
 699
 700 static bool
 701 perf_session__has_tp(struct perf_session *session, const char *name)
 702 {
 703         struct perf_evsel *evsel;
 704
 705         evsel = perf_evlist__find_tracepoint_by_name(session->evlist, name);
 706
 707         return evsel != NULL;
 708 }
 709
 710 static int parse_target_str(struct trace *trace)
 711 {
 712         if (trace->opts.target.pid) {
 713                 trace->pid_list = intlist__new(trace->opts.target.pid);
 714                 if (trace->pid_list == NULL) {
 715                         pr_err("Error parsing process id string\n");
 716                         return -EINVAL;
 717                 }
 718         }
 719
 720         if (trace->opts.target.tid) {
 721                 trace->tid_list = intlist__new(trace->opts.target.tid);
 722                 if (trace->tid_list == NULL) {
 723                         pr_err("Error parsing thread id string\n");
 724                         return -EINVAL;
 725                 }
 726         }
 727
 728         return 0;
 729 }
 730
 731 static int trace__run(struct trace *trace, int argc, const char **argv)
 732 {
 733         struct perf_evlist *evlist = perf_evlist__new();
 734         struct perf_evsel *evsel;
 735         int err = -1, i;
 736         unsigned long before;
 737         const bool forks = argc > 0;
 738
 739         if (evlist == NULL) {
 740                 fprintf(trace->output, "Not enough memory to run!\n");
 741                 goto out;
 742         }
 743
 744         if (perf_evlist__add_newtp(evlist, "raw_syscalls", "sys_enter", trace__sys_enter) ||
 745             perf_evlist__add_newtp(evlist, "raw_syscalls", "sys_exit", trace__sys_exit)) {
 746                 fprintf(trace->output, "Couldn't read the raw_syscalls tracepoints information!\n");
 747                 goto out_delete_evlist;
 748         }
 749
 750         if (trace->sched &&
 751             perf_evlist__add_newtp(evlist, "sched", "sched_stat_runtime",
 752                                    trace__sched_stat_runtime)) {
 753                 fprintf(trace->output, "Couldn't read the sched_stat_runtime tracepoint information!\n");
 754                 goto out_delete_evlist;
 755         }
 756
 757         err = perf_evlist__create_maps(evlist, &trace->opts.target);
 758         if (err < 0) {
 759                 fprintf(trace->output, "Problems parsing the target to trace, check your options!\n");
 760                 goto out_delete_evlist;
 761         }
 762
 763         err = trace__symbols_init(trace, evlist);
 764         if (err < 0) {
 765                 fprintf(trace->output, "Problems initializing symbol libraries!\n");
 766                 goto out_delete_maps;
 767         }
 768
 769         perf_evlist__config(evlist, &trace->opts);
 770
 771         signal(SIGCHLD, sig_handler);
 772         signal(SIGINT, sig_handler);
 773
 774         if (forks) {
 775                 err = perf_evlist__prepare_workload(evlist, &trace->opts.target,
 776                                                     argv, false, false);
 777                 if (err < 0) {
 778                         fprintf(trace->output, "Couldn't run the workload!\n");
 779                         goto out_delete_maps;
 780                 }
 781         }
 782
 783         err = perf_evlist__open(evlist);
 784         if (err < 0) {
 785                 fprintf(trace->output, "Couldn't create the events: %s\n", strerror(errno));
 786                 goto out_delete_maps;
 787         }
 788
 789         err = perf_evlist__mmap(evlist, UINT_MAX, false);
 790         if (err < 0) {
 791                 fprintf(trace->output, "Couldn't mmap the events: %s\n", strerror(errno));
 792                 goto out_close_evlist;
 793         }
 794
 795         perf_evlist__enable(evlist);
 796
 797         if (forks)
 798                 perf_evlist__start_workload(evlist);
 799
 800         trace->multiple_threads = evlist->threads->map[0] == -1 || evlist->threads->nr > 1;
 801 again:
 802         before = trace->nr_events;
 803
 804         for (i = 0; i < evlist->nr_mmaps; i++) {
 805                 union perf_event *event;
 806
 807                 while ((event = perf_evlist__mmap_read(evlist, i)) != NULL) {
 808                         const u32 type = event->header.type;
 809                         tracepoint_handler handler;
 810                         struct perf_sample sample;
 811
 812                         ++trace->nr_events;
 813
 814                         err = perf_evlist__parse_sample(evlist, event, &sample);
 815                         if (err) {
 816                                 fprintf(trace->output, "Can't parse sample, err = %d, skipping...\n", err);
 817                                 continue;
 818                         }
 819
 820                         if (trace->base_time == 0)
 821                                 trace->base_time = sample.time;
 822
 823                         if (type != PERF_RECORD_SAMPLE) {
 824                                 trace__process_event(trace, &trace->host, event);
 825                                 continue;
 826                         }
 827
 828                         evsel = perf_evlist__id2evsel(evlist, sample.id);
 829                         if (evsel == NULL) {
 830                                 fprintf(trace->output, "Unknown tp ID %" PRIu64 ", skipping...\n", sample.id);
 831                                 continue;
 832                         }
 833
 834                         if (sample.raw_data == NULL) {
 835                                 fprintf(trace->output, "%s sample with no payload for tid: %d, cpu %d, raw_size=%d, skipping...\n",
 836                                        perf_evsel__name(evsel), sample.tid,
 837                                        sample.cpu, sample.raw_size);
 838                                 continue;
 839                         }
 840
 841                         handler = evsel->handler.func;
 842                         handler(trace, evsel, &sample);
 843                 }
 844         }
 845
 846         if (trace->nr_events == before) {
 847                 if (done)
 848                         goto out_unmap_evlist;
 849
 850                 poll(evlist->pollfd, evlist->nr_fds, -1);
 851         }
 852
 853         if (done)
 854                 perf_evlist__disable(evlist);
 855
 856         goto again;
 857
 858 out_unmap_evlist:
 859         perf_evlist__munmap(evlist);
 860 out_close_evlist:
 861         perf_evlist__close(evlist);
 862 out_delete_maps:
 863         perf_evlist__delete_maps(evlist);
 864 out_delete_evlist:
 865         perf_evlist__delete(evlist);
 866 out:
 867         return err;
 868 }
 869
 870 static int trace__replay(struct trace *trace)
 871 {
 872         const struct perf_evsel_str_handler handlers[] = {
 873                 { "raw_syscalls:sys_enter",  trace__sys_enter, },
 874                 { "raw_syscalls:sys_exit",   trace__sys_exit, },
 875         };
 876
 877         struct perf_session *session;
 878         int err = -1;
 879
 880         trace->tool.sample        = trace__process_sample;
 881         trace->tool.mmap          = perf_event__process_mmap;
 882         trace->tool.comm          = perf_event__process_comm;
 883         trace->tool.exit          = perf_event__process_exit;
 884         trace->tool.fork          = perf_event__process_fork;
 885         trace->tool.attr          = perf_event__process_attr;
 886         trace->tool.tracing_data = perf_event__process_tracing_data;
 887         trace->tool.build_id      = perf_event__process_build_id;
 888
 889         trace->tool.ordered_samples = true;
 890         trace->tool.ordering_requires_timestamps = true;
 891
 892         /* add tid to output */
 893         trace->multiple_threads = true;
 894
 895         if (symbol__init() < 0)
 896                 return -1;
 897
 898         session = perf_session__new(input_name, O_RDONLY, 0, false,
 899                                     &trace->tool);
 900         if (session == NULL)
 901                 return -ENOMEM;
 902
 903         err = perf_session__set_tracepoints_handlers(session, handlers);
 904         if (err)
 905                 goto out;
 906
 907         if (!perf_session__has_tp(session, "raw_syscalls:sys_enter")) {
 908                 pr_err("Data file does not have raw_syscalls:sys_enter events\n");
 909                 goto out;
 910         }
 911
 912         if (!perf_session__has_tp(session, "raw_syscalls:sys_exit")) {
 913                 pr_err("Data file does not have raw_syscalls:sys_exit events\n");
 914                 goto out;
 915         }
 916
 917         err = parse_target_str(trace);
 918         if (err != 0)
 919                 goto out;
 920
 921         setup_pager();
 922
 923         err = perf_session__process_events(session, &trace->tool);
 924         if (err)
 925                 pr_err("Failed to process events, error %d", err);
 926
 927 out:
 928         perf_session__delete(session);
 929
 930         return err;
 931 }
 932
 933 static size_t trace__fprintf_threads_header(FILE *fp)
 934 {
 935         size_t printed;
 936
 937         printed  = fprintf(fp, "\n _____________________________________________________________________\n");
 938         printed += fprintf(fp," __)    Summary of events    (__\n\n");
 939         printed += fprintf(fp,"              [ task - pid ]     [ events ] [ ratio ]  [ runtime ]\n");
 940         printed += fprintf(fp," _____________________________________________________________________\n\n");
 941
 942         return printed;
 943 }
 944
 945 static size_t trace__fprintf_thread_summary(struct trace *trace, FILE *fp)
 946 {
 947         size_t printed = trace__fprintf_threads_header(fp);
 948         struct rb_node *nd;
 949
 950         for (nd = rb_first(&trace->host.threads); nd; nd = rb_next(nd)) {
 951                 struct thread *thread = rb_entry(nd, struct thread, rb_node);
 952                 struct thread_trace *ttrace = thread->priv;
 953                 const char *color;
 954                 double ratio;
 955
 956                 if (ttrace == NULL)
 957                         continue;
 958
 959                 ratio = (double)ttrace->nr_events / trace->nr_events * 100.0;
 960
 961                 color = PERF_COLOR_NORMAL;
 962                 if (ratio > 50.0)
 963                         color = PERF_COLOR_RED;
 964                 else if (ratio > 25.0)
 965                         color = PERF_COLOR_GREEN;
 966                 else if (ratio > 5.0)
 967                         color = PERF_COLOR_YELLOW;
 968
 969                 printed += color_fprintf(fp, color, "%20s", thread->comm);
 970                 printed += fprintf(fp, " - %-5d :%11lu   [", thread->tid, ttrace->nr_events);
 971                 printed += color_fprintf(fp, color, "%5.1f%%", ratio);
 972                 printed += fprintf(fp, " ] %10.3f ms\n", ttrace->runtime_ms);
 973         }
 974
 975         return printed;
 976 }
 977
 978 static int trace__set_duration(const struct option *opt, const char *str,
 979                                int unset __maybe_unused)
 980 {
 981         struct trace *trace = opt->value;
 982
 983         trace->duration_filter = atof(str);
 984         return 0;
 985 }
 986
 987 static int trace__open_output(struct trace *trace, const char *filename)
 988 {
 989         struct stat st;
 990
 991         if (!stat(filename, &st) && st.st_size) {
 992                 char oldname[PATH_MAX];
 993
 994                 scnprintf(oldname, sizeof(oldname), "%s.old", filename);
 995                 unlink(oldname);
 996                 rename(filename, oldname);
 997         }
 998
 999         trace->output = fopen(filename, "w");
1000
1001         return trace->output == NULL ? -errno : 0;
1002 }
1003
1004 int cmd_trace(int argc, const char **argv, const char *prefix __maybe_unused)
1005 {
1006         const char * const trace_usage[] = {
1007                 "perf trace [<options>] [<command>]",
1008                 "perf trace [<options>] -- <command> [<options>]",
1009                 NULL
1010         };
1011         struct trace trace = {
1012                 .audit_machine = audit_detect_machine(),
1013                 .syscalls = {
1014                         . max = -1,
1015                 },
1016                 .opts = {
1017                         .target = {
1018                                 .uid       = UINT_MAX,
1019                                 .uses_mmap = true,
1020                         },
1021                         .user_freq     = UINT_MAX,
1022                         .user_interval = ULLONG_MAX,
1023                         .no_delay      = true,
1024                         .mmap_pages    = 1024,
1025                 },
1026                 .output = stdout,
1027         };
1028         const char *output_name = NULL;
1029         const char *ev_qualifier_str = NULL;
1030         const struct option trace_options[] = {
1031         OPT_STRING('e', "expr", &ev_qualifier_str, "expr",
1032                     "list of events to trace"),
1033         OPT_STRING('o', "output", &output_name, "file", "output file name"),
1034         OPT_STRING('i', "input", &input_name, "file", "Analyze events in file"),
1035         OPT_STRING('p', "pid", &trace.opts.target.pid, "pid",
1036                     "trace events on existing process id"),
1037         OPT_STRING('t', "tid", &trace.opts.target.tid, "tid",
1038                     "trace events on existing thread id"),
1039         OPT_BOOLEAN('a', "all-cpus", &trace.opts.target.system_wide,
1040                     "system-wide collection from all CPUs"),
1041         OPT_STRING('C', "cpu", &trace.opts.target.cpu_list, "cpu",
1042                     "list of cpus to monitor"),
1043         OPT_BOOLEAN(0, "no-inherit", &trace.opts.no_inherit,
1044                     "child tasks do not inherit counters"),
1045         OPT_UINTEGER('m', "mmap-pages", &trace.opts.mmap_pages,
1046                      "number of mmap data pages"),
1047         OPT_STRING('u', "uid", &trace.opts.target.uid_str, "user",
1048                    "user to profile"),
1049         OPT_CALLBACK(0, "duration", &trace, "float",
1050                      "show only events with duration > N.M ms",
1051                      trace__set_duration),
1052         OPT_BOOLEAN(0, "sched", &trace.sched, "show blocking scheduler events"),
1053         OPT_INCR('v', "verbose", &verbose, "be more verbose"),
1054         OPT_END()
1055         };
1056         int err;
1057         char bf[BUFSIZ];
1058
1059         argc = parse_options(argc, argv, trace_options, trace_usage, 0);
1060
1061         if (output_name != NULL) {
1062                 err = trace__open_output(&trace, output_name);
1063                 if (err < 0) {
1064                         perror("failed to create output file");
1065                         goto out;
1066                 }
1067         }
1068
1069         if (ev_qualifier_str != NULL) {
1070                 const char *s = ev_qualifier_str;
1071
1072                 trace.not_ev_qualifier = *s == '!';
1073                 if (trace.not_ev_qualifier)
1074                         ++s;
1075                 trace.ev_qualifier = strlist__new(true, s);
1076                 if (trace.ev_qualifier == NULL) {
1077                         fputs("Not enough memory to parse event qualifier",
1078                               trace.output);
1079                         err = -ENOMEM;
1080                         goto out_close;
1081                 }
1082         }
1083
1084         err = perf_target__validate(&trace.opts.target);
1085         if (err) {
1086                 perf_target__strerror(&trace.opts.target, err, bf, sizeof(bf));
1087                 fprintf(trace.output, "%s", bf);
1088                 goto out_close;
1089         }
1090
1091         err = perf_target__parse_uid(&trace.opts.target);
1092         if (err) {
1093                 perf_target__strerror(&trace.opts.target, err, bf, sizeof(bf));
1094                 fprintf(trace.output, "%s", bf);
1095                 goto out_close;
1096         }
1097
1098         if (!argc && perf_target__none(&trace.opts.target))
1099                 trace.opts.target.system_wide = true;
1100
1101         if (input_name)
1102                 err = trace__replay(&trace);
1103         else
1104                 err = trace__run(&trace, argc, argv);
1105
1106         if (trace.sched && !err)
1107                 trace__fprintf_thread_summary(&trace, trace.output);
1108
1109 out_close:
1110         if (output_name != NULL)
1111                 fclose(trace.output);
1112 out:
1113         return err;
1114 }