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 #include <linux/futex.h>
  18
  19 static size_t syscall_arg__scnprintf_hex(char *bf, size_t size,
  20                                          unsigned long arg, u8 *arg_mask __maybe_unused)
  21 {
  22         return scnprintf(bf, size, "%#lx", arg);
  23 }
  24
  25 #define SCA_HEX syscall_arg__scnprintf_hex
  26
  27 static size_t syscall_arg__scnprintf_whence(char *bf, size_t size,
  28                                             unsigned long arg, u8 *arg_mask __maybe_unused)
  29 {
  30         int whence = arg;
  31
  32         switch (whence) {
  33 #define P_WHENCE(n) case SEEK_##n: return scnprintf(bf, size, #n)
  34         P_WHENCE(SET);
  35         P_WHENCE(CUR);
  36         P_WHENCE(END);
  37 #ifdef SEEK_DATA
  38         P_WHENCE(DATA);
  39 #endif
  40 #ifdef SEEK_HOLE
  41         P_WHENCE(HOLE);
  42 #endif
  43 #undef P_WHENCE
  44         default: break;
  45         }
  46
  47         return scnprintf(bf, size, "%#x", whence);
  48 }
  49
  50 #define SCA_WHENCE syscall_arg__scnprintf_whence
  51
  52 static size_t syscall_arg__scnprintf_mmap_prot(char *bf, size_t size,
  53                                                unsigned long arg, u8 *arg_mask __maybe_unused)
  54 {
  55         int printed = 0, prot = arg;
  56
  57         if (prot == PROT_NONE)
  58                 return scnprintf(bf, size, "NONE");
  59 #define P_MMAP_PROT(n) \
  60         if (prot & PROT_##n) { \
  61                 printed += scnprintf(bf + printed, size - printed, "%s%s", printed ? "|" : "", #n); \
  62                 prot &= ~PROT_##n; \
  63         }
  64
  65         P_MMAP_PROT(EXEC);
  66         P_MMAP_PROT(READ);
  67         P_MMAP_PROT(WRITE);
  68 #ifdef PROT_SEM
  69         P_MMAP_PROT(SEM);
  70 #endif
  71         P_MMAP_PROT(GROWSDOWN);
  72         P_MMAP_PROT(GROWSUP);
  73 #undef P_MMAP_PROT
  74
  75         if (prot)
  76                 printed += scnprintf(bf + printed, size - printed, "%s%#x", printed ? "|" : "", prot);
  77
  78         return printed;
  79 }
  80
  81 #define SCA_MMAP_PROT syscall_arg__scnprintf_mmap_prot
  82
  83 static size_t syscall_arg__scnprintf_mmap_flags(char *bf, size_t size,
  84                                                 unsigned long arg, u8 *arg_mask __maybe_unused)
  85 {
  86         int printed = 0, flags = arg;
  87
  88 #define P_MMAP_FLAG(n) \
  89         if (flags & MAP_##n) { \
  90                 printed += scnprintf(bf + printed, size - printed, "%s%s", printed ? "|" : "", #n); \
  91                 flags &= ~MAP_##n; \
  92         }
  93
  94         P_MMAP_FLAG(SHARED);
  95         P_MMAP_FLAG(PRIVATE);
  96         P_MMAP_FLAG(32BIT);
  97         P_MMAP_FLAG(ANONYMOUS);
  98         P_MMAP_FLAG(DENYWRITE);
  99         P_MMAP_FLAG(EXECUTABLE);
 100         P_MMAP_FLAG(FILE);
 101         P_MMAP_FLAG(FIXED);
 102         P_MMAP_FLAG(GROWSDOWN);
 103 #ifdef MAP_HUGETLB
 104         P_MMAP_FLAG(HUGETLB);
 105 #endif
 106         P_MMAP_FLAG(LOCKED);
 107         P_MMAP_FLAG(NONBLOCK);
 108         P_MMAP_FLAG(NORESERVE);
 109         P_MMAP_FLAG(POPULATE);
 110         P_MMAP_FLAG(STACK);
 111 #ifdef MAP_UNINITIALIZED
 112         P_MMAP_FLAG(UNINITIALIZED);
 113 #endif
 114 #undef P_MMAP_FLAG
 115
 116         if (flags)
 117                 printed += scnprintf(bf + printed, size - printed, "%s%#x", printed ? "|" : "", flags);
 118
 119         return printed;
 120 }
 121
 122 #define SCA_MMAP_FLAGS syscall_arg__scnprintf_mmap_flags
 123
 124 static size_t syscall_arg__scnprintf_madvise_behavior(char *bf, size_t size,
 125                                                       unsigned long arg, u8 *arg_mask __maybe_unused)
 126 {
 127         int behavior = arg;
 128
 129         switch (behavior) {
 130 #define P_MADV_BHV(n) case MADV_##n: return scnprintf(bf, size, #n)
 131         P_MADV_BHV(NORMAL);
 132         P_MADV_BHV(RANDOM);
 133         P_MADV_BHV(SEQUENTIAL);
 134         P_MADV_BHV(WILLNEED);
 135         P_MADV_BHV(DONTNEED);
 136         P_MADV_BHV(REMOVE);
 137         P_MADV_BHV(DONTFORK);
 138         P_MADV_BHV(DOFORK);
 139         P_MADV_BHV(HWPOISON);
 140 #ifdef MADV_SOFT_OFFLINE
 141         P_MADV_BHV(SOFT_OFFLINE);
 142 #endif
 143         P_MADV_BHV(MERGEABLE);
 144         P_MADV_BHV(UNMERGEABLE);
 145 #ifdef MADV_HUGEPAGE
 146         P_MADV_BHV(HUGEPAGE);
 147 #endif
 148 #ifdef MADV_NOHUGEPAGE
 149         P_MADV_BHV(NOHUGEPAGE);
 150 #endif
 151 #ifdef MADV_DONTDUMP
 152         P_MADV_BHV(DONTDUMP);
 153 #endif
 154 #ifdef MADV_DODUMP
 155         P_MADV_BHV(DODUMP);
 156 #endif
 157 #undef P_MADV_PHV
 158         default: break;
 159         }
 160
 161         return scnprintf(bf, size, "%#x", behavior);
 162 }
 163
 164 #define SCA_MADV_BHV syscall_arg__scnprintf_madvise_behavior
 165
 166 static size_t syscall_arg__scnprintf_futex_op(char *bf, size_t size, unsigned long arg, u8 *arg_mask)
 167 {
 168         enum syscall_futex_args {
 169                 SCF_UADDR   = (1 << 0),
 170                 SCF_OP      = (1 << 1),
 171                 SCF_VAL     = (1 << 2),
 172                 SCF_TIMEOUT = (1 << 3),
 173                 SCF_UADDR2  = (1 << 4),
 174                 SCF_VAL3    = (1 << 5),
 175         };
 176         int op = arg;
 177         int cmd = op & FUTEX_CMD_MASK;
 178         size_t printed = 0;
 179
 180         switch (cmd) {
 181 #define P_FUTEX_OP(n) case FUTEX_##n: printed = scnprintf(bf, size, #n);
 182         P_FUTEX_OP(WAIT);           *arg_mask |= SCF_VAL3|SCF_UADDR2;             break;
 183         P_FUTEX_OP(WAKE);           *arg_mask |= SCF_VAL3|SCF_UADDR2|SCF_TIMEOUT; break;
 184         P_FUTEX_OP(FD);             *arg_mask |= SCF_VAL3|SCF_UADDR2|SCF_TIMEOUT; break;
 185         P_FUTEX_OP(REQUEUE);        *arg_mask |= SCF_VAL3|SCF_TIMEOUT;            break;
 186         P_FUTEX_OP(CMP_REQUEUE);    *arg_mask |= SCF_TIMEOUT;                     break;
 187         P_FUTEX_OP(CMP_REQUEUE_PI); *arg_mask |= SCF_TIMEOUT;                     break;
 188         P_FUTEX_OP(WAKE_OP);                                                      break;
 189         P_FUTEX_OP(LOCK_PI);        *arg_mask |= SCF_VAL3|SCF_UADDR2|SCF_TIMEOUT; break;
 190         P_FUTEX_OP(UNLOCK_PI);      *arg_mask |= SCF_VAL3|SCF_UADDR2|SCF_TIMEOUT; break;
 191         P_FUTEX_OP(TRYLOCK_PI);     *arg_mask |= SCF_VAL3|SCF_UADDR2;             break;
 192         P_FUTEX_OP(WAIT_BITSET);    *arg_mask |= SCF_UADDR2;                      break;
 193         P_FUTEX_OP(WAKE_BITSET);    *arg_mask |= SCF_UADDR2;                      break;
 194         P_FUTEX_OP(WAIT_REQUEUE_PI);                                              break;
 195         default: printed = scnprintf(bf, size, "%#x", cmd);                       break;
 196         }
 197
 198         if (op & FUTEX_PRIVATE_FLAG)
 199                 printed += scnprintf(bf + printed, size - printed, "|PRIV");
 200
 201         if (op & FUTEX_CLOCK_REALTIME)
 202                 printed += scnprintf(bf + printed, size - printed, "|CLKRT");
 203
 204         return printed;
 205 }
 206
 207 #define SCA_FUTEX_OP  syscall_arg__scnprintf_futex_op
 208
 209 static struct syscall_fmt {
 210         const char *name;
 211         const char *alias;
 212         size_t     (*arg_scnprintf[6])(char *bf, size_t size, unsigned long arg, u8 *arg_mask);
 213         bool       errmsg;
 214         bool       timeout;
 215         bool       hexret;
 216 } syscall_fmts[] = {
 217         { .name     = "access",     .errmsg = true, },
 218         { .name     = "arch_prctl", .errmsg = true, .alias = "prctl", },
 219         { .name     = "brk",        .hexret = true,
 220           .arg_scnprintf = { [0] = SCA_HEX, /* brk */ }, },
 221         { .name     = "mmap",       .hexret = true, },
 222         { .name     = "connect",    .errmsg = true, },
 223         { .name     = "fstat",      .errmsg = true, .alias = "newfstat", },
 224         { .name     = "fstatat",    .errmsg = true, .alias = "newfstatat", },
 225         { .name     = "futex",      .errmsg = true,
 226           .arg_scnprintf = { [1] = SCA_FUTEX_OP, /* op */ }, },
 227         { .name     = "ioctl",      .errmsg = true,
 228           .arg_scnprintf = { [2] = SCA_HEX, /* arg */ }, },
 229         { .name     = "lseek",      .errmsg = true,
 230           .arg_scnprintf = { [2] = SCA_WHENCE, /* whence */ }, },
 231         { .name     = "lstat",      .errmsg = true, .alias = "newlstat", },
 232         { .name     = "madvise",    .errmsg = true,
 233           .arg_scnprintf = { [0] = SCA_HEX,      /* start */
 234                              [2] = SCA_MADV_BHV, /* behavior */ }, },
 235         { .name     = "mmap",       .hexret = true,
 236           .arg_scnprintf = { [0] = SCA_HEX,       /* addr */
 237                              [2] = SCA_MMAP_PROT, /* prot */
 238                              [3] = SCA_MMAP_FLAGS, /* flags */ }, },
 239         { .name     = "mprotect",   .errmsg = true,
 240           .arg_scnprintf = { [0] = SCA_HEX, /* start */
 241                              [2] = SCA_MMAP_PROT, /* prot */ }, },
 242         { .name     = "mremap",     .hexret = true,
 243           .arg_scnprintf = { [0] = SCA_HEX, /* addr */
 244                              [4] = SCA_HEX, /* new_addr */ }, },
 245         { .name     = "munmap",     .errmsg = true,
 246           .arg_scnprintf = { [0] = SCA_HEX, /* addr */ }, },
 247         { .name     = "open",       .errmsg = true, },
 248         { .name     = "poll",       .errmsg = true, .timeout = true, },
 249         { .name     = "ppoll",      .errmsg = true, .timeout = true, },
 250         { .name     = "pread",      .errmsg = true, .alias = "pread64", },
 251         { .name     = "pwrite",     .errmsg = true, .alias = "pwrite64", },
 252         { .name     = "read",       .errmsg = true, },
 253         { .name     = "recvfrom",   .errmsg = true, },
 254         { .name     = "select",     .errmsg = true, .timeout = true, },
 255         { .name     = "socket",     .errmsg = true, },
 256         { .name     = "stat",       .errmsg = true, .alias = "newstat", },
 257         { .name     = "uname",      .errmsg = true, .alias = "newuname", },
 258 };
 259
 260 static int syscall_fmt__cmp(const void *name, const void *fmtp)
 261 {
 262         const struct syscall_fmt *fmt = fmtp;
 263         return strcmp(name, fmt->name);
 264 }
 265
 266 static struct syscall_fmt *syscall_fmt__find(const char *name)
 267 {
 268         const int nmemb = ARRAY_SIZE(syscall_fmts);
 269         return bsearch(name, syscall_fmts, nmemb, sizeof(struct syscall_fmt), syscall_fmt__cmp);
 270 }
 271
 272 struct syscall {
 273         struct event_format *tp_format;
 274         const char          *name;
 275         bool                filtered;
 276         struct syscall_fmt  *fmt;
 277         size_t              (**arg_scnprintf)(char *bf, size_t size,
 278                                               unsigned long arg, u8 *args_mask);
 279 };
 280
 281 static size_t fprintf_duration(unsigned long t, FILE *fp)
 282 {
 283         double duration = (double)t / NSEC_PER_MSEC;
 284         size_t printed = fprintf(fp, "(");
 285
 286         if (duration >= 1.0)
 287                 printed += color_fprintf(fp, PERF_COLOR_RED, "%6.3f ms", duration);
 288         else if (duration >= 0.01)
 289                 printed += color_fprintf(fp, PERF_COLOR_YELLOW, "%6.3f ms", duration);
 290         else
 291                 printed += color_fprintf(fp, PERF_COLOR_NORMAL, "%6.3f ms", duration);
 292         return printed + fprintf(fp, "): ");
 293 }
 294
 295 struct thread_trace {
 296         u64               entry_time;
 297         u64               exit_time;
 298         bool              entry_pending;
 299         unsigned long     nr_events;
 300         char              *entry_str;
 301         double            runtime_ms;
 302 };
 303
 304 static struct thread_trace *thread_trace__new(void)
 305 {
 306         return zalloc(sizeof(struct thread_trace));
 307 }
 308
 309 static struct thread_trace *thread__trace(struct thread *thread, FILE *fp)
 310 {
 311         struct thread_trace *ttrace;
 312
 313         if (thread == NULL)
 314                 goto fail;
 315
 316         if (thread->priv == NULL)
 317                 thread->priv = thread_trace__new();
 318
 319         if (thread->priv == NULL)
 320                 goto fail;
 321
 322         ttrace = thread->priv;
 323         ++ttrace->nr_events;
 324
 325         return ttrace;
 326 fail:
 327         color_fprintf(fp, PERF_COLOR_RED,
 328                       "WARNING: not enough memory, dropping samples!\n");
 329         return NULL;
 330 }
 331
 332 struct trace {
 333         struct perf_tool        tool;
 334         int                     audit_machine;
 335         struct {
 336                 int             max;
 337                 struct syscall  *table;
 338         } syscalls;
 339         struct perf_record_opts opts;
 340         struct machine          host;
 341         u64                     base_time;
 342         FILE                    *output;
 343         unsigned long           nr_events;
 344         struct strlist          *ev_qualifier;
 345         bool                    not_ev_qualifier;
 346         struct intlist          *tid_list;
 347         struct intlist          *pid_list;
 348         bool                    sched;
 349         bool                    multiple_threads;
 350         double                  duration_filter;
 351         double                  runtime_ms;
 352 };
 353
 354 static bool trace__filter_duration(struct trace *trace, double t)
 355 {
 356         return t < (trace->duration_filter * NSEC_PER_MSEC);
 357 }
 358
 359 static size_t trace__fprintf_tstamp(struct trace *trace, u64 tstamp, FILE *fp)
 360 {
 361         double ts = (double)(tstamp - trace->base_time) / NSEC_PER_MSEC;
 362
 363         return fprintf(fp, "%10.3f ", ts);
 364 }
 365
 366 static bool done = false;
 367
 368 static void sig_handler(int sig __maybe_unused)
 369 {
 370         done = true;
 371 }
 372
 373 static size_t trace__fprintf_entry_head(struct trace *trace, struct thread *thread,
 374                                         u64 duration, u64 tstamp, FILE *fp)
 375 {
 376         size_t printed = trace__fprintf_tstamp(trace, tstamp, fp);
 377         printed += fprintf_duration(duration, fp);
 378
 379         if (trace->multiple_threads)
 380                 printed += fprintf(fp, "%d ", thread->tid);
 381
 382         return printed;
 383 }
 384
 385 static int trace__process_event(struct trace *trace, struct machine *machine,
 386                                 union perf_event *event)
 387 {
 388         int ret = 0;
 389
 390         switch (event->header.type) {
 391         case PERF_RECORD_LOST:
 392                 color_fprintf(trace->output, PERF_COLOR_RED,
 393                               "LOST %" PRIu64 " events!\n", event->lost.lost);
 394                 ret = machine__process_lost_event(machine, event);
 395         default:
 396                 ret = machine__process_event(machine, event);
 397                 break;
 398         }
 399
 400         return ret;
 401 }
 402
 403 static int trace__tool_process(struct perf_tool *tool,
 404                                union perf_event *event,
 405                                struct perf_sample *sample __maybe_unused,
 406                                struct machine *machine)
 407 {
 408         struct trace *trace = container_of(tool, struct trace, tool);
 409         return trace__process_event(trace, machine, event);
 410 }
 411
 412 static int trace__symbols_init(struct trace *trace, struct perf_evlist *evlist)
 413 {
 414         int err = symbol__init();
 415
 416         if (err)
 417                 return err;
 418
 419         machine__init(&trace->host, "", HOST_KERNEL_ID);
 420         machine__create_kernel_maps(&trace->host);
 421
 422         if (perf_target__has_task(&trace->opts.target)) {
 423                 err = perf_event__synthesize_thread_map(&trace->tool, evlist->threads,
 424                                                         trace__tool_process,
 425                                                         &trace->host);
 426         } else {
 427                 err = perf_event__synthesize_threads(&trace->tool, trace__tool_process,
 428                                                      &trace->host);
 429         }
 430
 431         if (err)
 432                 symbol__exit();
 433
 434         return err;
 435 }
 436
 437 static int syscall__set_arg_fmts(struct syscall *sc)
 438 {
 439         struct format_field *field;
 440         int idx = 0;
 441
 442         sc->arg_scnprintf = calloc(sc->tp_format->format.nr_fields - 1, sizeof(void *));
 443         if (sc->arg_scnprintf == NULL)
 444                 return -1;
 445
 446         for (field = sc->tp_format->format.fields->next; field; field = field->next) {
 447                 if (sc->fmt && sc->fmt->arg_scnprintf[idx])
 448                         sc->arg_scnprintf[idx] = sc->fmt->arg_scnprintf[idx];
 449                 else if (field->flags & FIELD_IS_POINTER)
 450                         sc->arg_scnprintf[idx] = syscall_arg__scnprintf_hex;
 451                 ++idx;
 452         }
 453
 454         return 0;
 455 }
 456
 457 static int trace__read_syscall_info(struct trace *trace, int id)
 458 {
 459         char tp_name[128];
 460         struct syscall *sc;
 461         const char *name = audit_syscall_to_name(id, trace->audit_machine);
 462
 463         if (name == NULL)
 464                 return -1;
 465
 466         if (id > trace->syscalls.max) {
 467                 struct syscall *nsyscalls = realloc(trace->syscalls.table, (id + 1) * sizeof(*sc));
 468
 469                 if (nsyscalls == NULL)
 470                         return -1;
 471
 472                 if (trace->syscalls.max != -1) {
 473                         memset(nsyscalls + trace->syscalls.max + 1, 0,
 474                                (id - trace->syscalls.max) * sizeof(*sc));
 475                 } else {
 476                         memset(nsyscalls, 0, (id + 1) * sizeof(*sc));
 477                 }
 478
 479                 trace->syscalls.table = nsyscalls;
 480                 trace->syscalls.max   = id;
 481         }
 482
 483         sc = trace->syscalls.table + id;
 484         sc->name = name;
 485
 486         if (trace->ev_qualifier) {
 487                 bool in = strlist__find(trace->ev_qualifier, name) != NULL;
 488
 489                 if (!(in ^ trace->not_ev_qualifier)) {
 490                         sc->filtered = true;
 491                         /*
 492                          * No need to do read tracepoint information since this will be
 493                          * filtered out.
 494                          */
 495                         return 0;
 496                 }
 497         }
 498
 499         sc->fmt  = syscall_fmt__find(sc->name);
 500
 501         snprintf(tp_name, sizeof(tp_name), "sys_enter_%s", sc->name);
 502         sc->tp_format = event_format__new("syscalls", tp_name);
 503
 504         if (sc->tp_format == NULL && sc->fmt && sc->fmt->alias) {
 505                 snprintf(tp_name, sizeof(tp_name), "sys_enter_%s", sc->fmt->alias);
 506                 sc->tp_format = event_format__new("syscalls", tp_name);
 507         }
 508
 509         if (sc->tp_format == NULL)
 510                 return -1;
 511
 512         return syscall__set_arg_fmts(sc);
 513 }
 514
 515 static size_t syscall__scnprintf_args(struct syscall *sc, char *bf, size_t size,
 516                                       unsigned long *args)
 517 {
 518         int i = 0;
 519         size_t printed = 0;
 520
 521         if (sc->tp_format != NULL) {
 522                 struct format_field *field;
 523                 u8 mask = 0, bit = 1;
 524
 525                 for (field = sc->tp_format->format.fields->next; field;
 526                      field = field->next, ++i, bit <<= 1) {
 527                         if (mask & bit)
 528                                 continue;
 529
 530                         printed += scnprintf(bf + printed, size - printed,
 531                                              "%s%s: ", printed ? ", " : "", field->name);
 532
 533                         if (sc->arg_scnprintf && sc->arg_scnprintf[i]) {
 534                                 printed += sc->arg_scnprintf[i](bf + printed, size - printed,
 535                                                                 args[i], &mask);
 536                         } else {
 537                                 printed += scnprintf(bf + printed, size - printed,
 538                                                      "%ld", args[i]);
 539                         }
 540                 }
 541         } else {
 542                 while (i < 6) {
 543                         printed += scnprintf(bf + printed, size - printed,
 544                                              "%sarg%d: %ld",
 545                                              printed ? ", " : "", i, args[i]);
 546                         ++i;
 547                 }
 548         }
 549
 550         return printed;
 551 }
 552
 553 typedef int (*tracepoint_handler)(struct trace *trace, struct perf_evsel *evsel,
 554                                   struct perf_sample *sample);
 555
 556 static struct syscall *trace__syscall_info(struct trace *trace,
 557                                            struct perf_evsel *evsel,
 558                                            struct perf_sample *sample)
 559 {
 560         int id = perf_evsel__intval(evsel, sample, "id");
 561
 562         if (id < 0) {
 563
 564                 /*
 565                  * XXX: Noticed on x86_64, reproduced as far back as 3.0.36, haven't tried
 566                  * before that, leaving at a higher verbosity level till that is
 567                  * explained. Reproduced with plain ftrace with:
 568                  *
 569                  * echo 1 > /t/events/raw_syscalls/sys_exit/enable
 570                  * grep "NR -1 " /t/trace_pipe
 571                  *
 572                  * After generating some load on the machine.
 573                  */
 574                 if (verbose > 1) {
 575                         static u64 n;
 576                         fprintf(trace->output, "Invalid syscall %d id, skipping (%s, %" PRIu64 ") ...\n",
 577                                 id, perf_evsel__name(evsel), ++n);
 578                 }
 579                 return NULL;
 580         }
 581
 582         if ((id > trace->syscalls.max || trace->syscalls.table[id].name == NULL) &&
 583             trace__read_syscall_info(trace, id))
 584                 goto out_cant_read;
 585
 586         if ((id > trace->syscalls.max || trace->syscalls.table[id].name == NULL))
 587                 goto out_cant_read;
 588
 589         return &trace->syscalls.table[id];
 590
 591 out_cant_read:
 592         if (verbose) {
 593                 fprintf(trace->output, "Problems reading syscall %d", id);
 594                 if (id <= trace->syscalls.max && trace->syscalls.table[id].name != NULL)
 595                         fprintf(trace->output, "(%s)", trace->syscalls.table[id].name);
 596                 fputs(" information\n", trace->output);
 597         }
 598         return NULL;
 599 }
 600
 601 static int trace__sys_enter(struct trace *trace, struct perf_evsel *evsel,
 602                             struct perf_sample *sample)
 603 {
 604         char *msg;
 605         void *args;
 606         size_t printed = 0;
 607         struct thread *thread;
 608         struct syscall *sc = trace__syscall_info(trace, evsel, sample);
 609         struct thread_trace *ttrace;
 610
 611         if (sc == NULL)
 612                 return -1;
 613
 614         if (sc->filtered)
 615                 return 0;
 616
 617         thread = machine__findnew_thread(&trace->host, sample->pid,
 618                                          sample->tid);
 619         ttrace = thread__trace(thread, trace->output);
 620         if (ttrace == NULL)
 621                 return -1;
 622
 623         args = perf_evsel__rawptr(evsel, sample, "args");
 624         if (args == NULL) {
 625                 fprintf(trace->output, "Problems reading syscall arguments\n");
 626                 return -1;
 627         }
 628
 629         ttrace = thread->priv;
 630
 631         if (ttrace->entry_str == NULL) {
 632                 ttrace->entry_str = malloc(1024);
 633                 if (!ttrace->entry_str)
 634                         return -1;
 635         }
 636
 637         ttrace->entry_time = sample->time;
 638         msg = ttrace->entry_str;
 639         printed += scnprintf(msg + printed, 1024 - printed, "%s(", sc->name);
 640
 641         printed += syscall__scnprintf_args(sc, msg + printed, 1024 - printed,  args);
 642
 643         if (!strcmp(sc->name, "exit_group") || !strcmp(sc->name, "exit")) {
 644                 if (!trace->duration_filter) {
 645                         trace__fprintf_entry_head(trace, thread, 1, sample->time, trace->output);
 646                         fprintf(trace->output, "%-70s\n", ttrace->entry_str);
 647                 }
 648         } else
 649                 ttrace->entry_pending = true;
 650
 651         return 0;
 652 }
 653
 654 static int trace__sys_exit(struct trace *trace, struct perf_evsel *evsel,
 655                            struct perf_sample *sample)
 656 {
 657         int ret;
 658         u64 duration = 0;
 659         struct thread *thread;
 660         struct syscall *sc = trace__syscall_info(trace, evsel, sample);
 661         struct thread_trace *ttrace;
 662
 663         if (sc == NULL)
 664                 return -1;
 665
 666         if (sc->filtered)
 667                 return 0;
 668
 669         thread = machine__findnew_thread(&trace->host, sample->pid,
 670                                          sample->tid);
 671         ttrace = thread__trace(thread, trace->output);
 672         if (ttrace == NULL)
 673                 return -1;
 674
 675         ret = perf_evsel__intval(evsel, sample, "ret");
 676
 677         ttrace = thread->priv;
 678
 679         ttrace->exit_time = sample->time;
 680
 681         if (ttrace->entry_time) {
 682                 duration = sample->time - ttrace->entry_time;
 683                 if (trace__filter_duration(trace, duration))
 684                         goto out;
 685         } else if (trace->duration_filter)
 686                 goto out;
 687
 688         trace__fprintf_entry_head(trace, thread, duration, sample->time, trace->output);
 689
 690         if (ttrace->entry_pending) {
 691                 fprintf(trace->output, "%-70s", ttrace->entry_str);
 692         } else {
 693                 fprintf(trace->output, " ... [");
 694                 color_fprintf(trace->output, PERF_COLOR_YELLOW, "continued");
 695                 fprintf(trace->output, "]: %s()", sc->name);
 696         }
 697
 698         if (sc->fmt == NULL) {
 699 signed_print:
 700                 fprintf(trace->output, ") = %d", ret);
 701         } else if (ret < 0 && sc->fmt->errmsg) {
 702                 char bf[256];
 703                 const char *emsg = strerror_r(-ret, bf, sizeof(bf)),
 704                            *e = audit_errno_to_name(-ret);
 705
 706                 fprintf(trace->output, ") = -1 %s %s", e, emsg);
 707         } else if (ret == 0 && sc->fmt->timeout)
 708                 fprintf(trace->output, ") = 0 Timeout");
 709         else if (sc->fmt->hexret)
 710                 fprintf(trace->output, ") = %#x", ret);
 711         else
 712                 goto signed_print;
 713
 714         fputc('\n', trace->output);
 715 out:
 716         ttrace->entry_pending = false;
 717
 718         return 0;
 719 }
 720
 721 static int trace__sched_stat_runtime(struct trace *trace, struct perf_evsel *evsel,
 722                                      struct perf_sample *sample)
 723 {
 724         u64 runtime = perf_evsel__intval(evsel, sample, "runtime");
 725         double runtime_ms = (double)runtime / NSEC_PER_MSEC;
 726         struct thread *thread = machine__findnew_thread(&trace->host,
 727                                                         sample->pid,
 728                                                         sample->tid);
 729         struct thread_trace *ttrace = thread__trace(thread, trace->output);
 730
 731         if (ttrace == NULL)
 732                 goto out_dump;
 733
 734         ttrace->runtime_ms += runtime_ms;
 735         trace->runtime_ms += runtime_ms;
 736         return 0;
 737
 738 out_dump:
 739         fprintf(trace->output, "%s: comm=%s,pid=%u,runtime=%" PRIu64 ",vruntime=%" PRIu64 ")\n",
 740                evsel->name,
 741                perf_evsel__strval(evsel, sample, "comm"),
 742                (pid_t)perf_evsel__intval(evsel, sample, "pid"),
 743                runtime,
 744                perf_evsel__intval(evsel, sample, "vruntime"));
 745         return 0;
 746 }
 747
 748 static bool skip_sample(struct trace *trace, struct perf_sample *sample)
 749 {
 750         if ((trace->pid_list && intlist__find(trace->pid_list, sample->pid)) ||
 751             (trace->tid_list && intlist__find(trace->tid_list, sample->tid)))
 752                 return false;
 753
 754         if (trace->pid_list || trace->tid_list)
 755                 return true;
 756
 757         return false;
 758 }
 759
 760 static int trace__process_sample(struct perf_tool *tool,
 761                                  union perf_event *event __maybe_unused,
 762                                  struct perf_sample *sample,
 763                                  struct perf_evsel *evsel,
 764                                  struct machine *machine __maybe_unused)
 765 {
 766         struct trace *trace = container_of(tool, struct trace, tool);
 767         int err = 0;
 768
 769         tracepoint_handler handler = evsel->handler.func;
 770
 771         if (skip_sample(trace, sample))
 772                 return 0;
 773
 774         if (trace->base_time == 0)
 775                 trace->base_time = sample->time;
 776
 777         if (handler)
 778                 handler(trace, evsel, sample);
 779
 780         return err;
 781 }
 782
 783 static bool
 784 perf_session__has_tp(struct perf_session *session, const char *name)
 785 {
 786         struct perf_evsel *evsel;
 787
 788         evsel = perf_evlist__find_tracepoint_by_name(session->evlist, name);
 789
 790         return evsel != NULL;
 791 }
 792
 793 static int parse_target_str(struct trace *trace)
 794 {
 795         if (trace->opts.target.pid) {
 796                 trace->pid_list = intlist__new(trace->opts.target.pid);
 797                 if (trace->pid_list == NULL) {
 798                         pr_err("Error parsing process id string\n");
 799                         return -EINVAL;
 800                 }
 801         }
 802
 803         if (trace->opts.target.tid) {
 804                 trace->tid_list = intlist__new(trace->opts.target.tid);
 805                 if (trace->tid_list == NULL) {
 806                         pr_err("Error parsing thread id string\n");
 807                         return -EINVAL;
 808                 }
 809         }
 810
 811         return 0;
 812 }
 813
 814 static int trace__run(struct trace *trace, int argc, const char **argv)
 815 {
 816         struct perf_evlist *evlist = perf_evlist__new();
 817         struct perf_evsel *evsel;
 818         int err = -1, i;
 819         unsigned long before;
 820         const bool forks = argc > 0;
 821
 822         if (evlist == NULL) {
 823                 fprintf(trace->output, "Not enough memory to run!\n");
 824                 goto out;
 825         }
 826
 827         if (perf_evlist__add_newtp(evlist, "raw_syscalls", "sys_enter", trace__sys_enter) ||
 828             perf_evlist__add_newtp(evlist, "raw_syscalls", "sys_exit", trace__sys_exit)) {
 829                 fprintf(trace->output, "Couldn't read the raw_syscalls tracepoints information!\n");
 830                 goto out_delete_evlist;
 831         }
 832
 833         if (trace->sched &&
 834             perf_evlist__add_newtp(evlist, "sched", "sched_stat_runtime",
 835                                    trace__sched_stat_runtime)) {
 836                 fprintf(trace->output, "Couldn't read the sched_stat_runtime tracepoint information!\n");
 837                 goto out_delete_evlist;
 838         }
 839
 840         err = perf_evlist__create_maps(evlist, &trace->opts.target);
 841         if (err < 0) {
 842                 fprintf(trace->output, "Problems parsing the target to trace, check your options!\n");
 843                 goto out_delete_evlist;
 844         }
 845
 846         err = trace__symbols_init(trace, evlist);
 847         if (err < 0) {
 848                 fprintf(trace->output, "Problems initializing symbol libraries!\n");
 849                 goto out_delete_maps;
 850         }
 851
 852         perf_evlist__config(evlist, &trace->opts);
 853
 854         signal(SIGCHLD, sig_handler);
 855         signal(SIGINT, sig_handler);
 856
 857         if (forks) {
 858                 err = perf_evlist__prepare_workload(evlist, &trace->opts.target,
 859                                                     argv, false, false);
 860                 if (err < 0) {
 861                         fprintf(trace->output, "Couldn't run the workload!\n");
 862                         goto out_delete_maps;
 863                 }
 864         }
 865
 866         err = perf_evlist__open(evlist);
 867         if (err < 0) {
 868                 fprintf(trace->output, "Couldn't create the events: %s\n", strerror(errno));
 869                 goto out_delete_maps;
 870         }
 871
 872         err = perf_evlist__mmap(evlist, UINT_MAX, false);
 873         if (err < 0) {
 874                 fprintf(trace->output, "Couldn't mmap the events: %s\n", strerror(errno));
 875                 goto out_close_evlist;
 876         }
 877
 878         perf_evlist__enable(evlist);
 879
 880         if (forks)
 881                 perf_evlist__start_workload(evlist);
 882
 883         trace->multiple_threads = evlist->threads->map[0] == -1 || evlist->threads->nr > 1;
 884 again:
 885         before = trace->nr_events;
 886
 887         for (i = 0; i < evlist->nr_mmaps; i++) {
 888                 union perf_event *event;
 889
 890                 while ((event = perf_evlist__mmap_read(evlist, i)) != NULL) {
 891                         const u32 type = event->header.type;
 892                         tracepoint_handler handler;
 893                         struct perf_sample sample;
 894
 895                         ++trace->nr_events;
 896
 897                         err = perf_evlist__parse_sample(evlist, event, &sample);
 898                         if (err) {
 899                                 fprintf(trace->output, "Can't parse sample, err = %d, skipping...\n", err);
 900                                 continue;
 901                         }
 902
 903                         if (trace->base_time == 0)
 904                                 trace->base_time = sample.time;
 905
 906                         if (type != PERF_RECORD_SAMPLE) {
 907                                 trace__process_event(trace, &trace->host, event);
 908                                 continue;
 909                         }
 910
 911                         evsel = perf_evlist__id2evsel(evlist, sample.id);
 912                         if (evsel == NULL) {
 913                                 fprintf(trace->output, "Unknown tp ID %" PRIu64 ", skipping...\n", sample.id);
 914                                 continue;
 915                         }
 916
 917                         if (sample.raw_data == NULL) {
 918                                 fprintf(trace->output, "%s sample with no payload for tid: %d, cpu %d, raw_size=%d, skipping...\n",
 919                                        perf_evsel__name(evsel), sample.tid,
 920                                        sample.cpu, sample.raw_size);
 921                                 continue;
 922                         }
 923
 924                         handler = evsel->handler.func;
 925                         handler(trace, evsel, &sample);
 926                 }
 927         }
 928
 929         if (trace->nr_events == before) {
 930                 if (done)
 931                         goto out_unmap_evlist;
 932
 933                 poll(evlist->pollfd, evlist->nr_fds, -1);
 934         }
 935
 936         if (done)
 937                 perf_evlist__disable(evlist);
 938
 939         goto again;
 940
 941 out_unmap_evlist:
 942         perf_evlist__munmap(evlist);
 943 out_close_evlist:
 944         perf_evlist__close(evlist);
 945 out_delete_maps:
 946         perf_evlist__delete_maps(evlist);
 947 out_delete_evlist:
 948         perf_evlist__delete(evlist);
 949 out:
 950         return err;
 951 }
 952
 953 static int trace__replay(struct trace *trace)
 954 {
 955         const struct perf_evsel_str_handler handlers[] = {
 956                 { "raw_syscalls:sys_enter",  trace__sys_enter, },
 957                 { "raw_syscalls:sys_exit",   trace__sys_exit, },
 958         };
 959
 960         struct perf_session *session;
 961         int err = -1;
 962
 963         trace->tool.sample        = trace__process_sample;
 964         trace->tool.mmap          = perf_event__process_mmap;
 965         trace->tool.comm          = perf_event__process_comm;
 966         trace->tool.exit          = perf_event__process_exit;
 967         trace->tool.fork          = perf_event__process_fork;
 968         trace->tool.attr          = perf_event__process_attr;
 969         trace->tool.tracing_data = perf_event__process_tracing_data;
 970         trace->tool.build_id      = perf_event__process_build_id;
 971
 972         trace->tool.ordered_samples = true;
 973         trace->tool.ordering_requires_timestamps = true;
 974
 975         /* add tid to output */
 976         trace->multiple_threads = true;
 977
 978         if (symbol__init() < 0)
 979                 return -1;
 980
 981         session = perf_session__new(input_name, O_RDONLY, 0, false,
 982                                     &trace->tool);
 983         if (session == NULL)
 984                 return -ENOMEM;
 985
 986         err = perf_session__set_tracepoints_handlers(session, handlers);
 987         if (err)
 988                 goto out;
 989
 990         if (!perf_session__has_tp(session, "raw_syscalls:sys_enter")) {
 991                 pr_err("Data file does not have raw_syscalls:sys_enter events\n");
 992                 goto out;
 993         }
 994
 995         if (!perf_session__has_tp(session, "raw_syscalls:sys_exit")) {
 996                 pr_err("Data file does not have raw_syscalls:sys_exit events\n");
 997                 goto out;
 998         }
 999
1000         err = parse_target_str(trace);
1001         if (err != 0)
1002                 goto out;
1003
1004         setup_pager();
1005
1006         err = perf_session__process_events(session, &trace->tool);
1007         if (err)
1008                 pr_err("Failed to process events, error %d", err);
1009
1010 out:
1011         perf_session__delete(session);
1012
1013         return err;
1014 }
1015
1016 static size_t trace__fprintf_threads_header(FILE *fp)
1017 {
1018         size_t printed;
1019
1020         printed  = fprintf(fp, "\n _____________________________________________________________________\n");
1021         printed += fprintf(fp," __)    Summary of events    (__\n\n");
1022         printed += fprintf(fp,"              [ task - pid ]     [ events ] [ ratio ]  [ runtime ]\n");
1023         printed += fprintf(fp," _____________________________________________________________________\n\n");
1024
1025         return printed;
1026 }
1027
1028 static size_t trace__fprintf_thread_summary(struct trace *trace, FILE *fp)
1029 {
1030         size_t printed = trace__fprintf_threads_header(fp);
1031         struct rb_node *nd;
1032
1033         for (nd = rb_first(&trace->host.threads); nd; nd = rb_next(nd)) {
1034                 struct thread *thread = rb_entry(nd, struct thread, rb_node);
1035                 struct thread_trace *ttrace = thread->priv;
1036                 const char *color;
1037                 double ratio;
1038
1039                 if (ttrace == NULL)
1040                         continue;
1041
1042                 ratio = (double)ttrace->nr_events / trace->nr_events * 100.0;
1043
1044                 color = PERF_COLOR_NORMAL;
1045                 if (ratio > 50.0)
1046                         color = PERF_COLOR_RED;
1047                 else if (ratio > 25.0)
1048                         color = PERF_COLOR_GREEN;
1049                 else if (ratio > 5.0)
1050                         color = PERF_COLOR_YELLOW;
1051
1052                 printed += color_fprintf(fp, color, "%20s", thread->comm);
1053                 printed += fprintf(fp, " - %-5d :%11lu   [", thread->tid, ttrace->nr_events);
1054                 printed += color_fprintf(fp, color, "%5.1f%%", ratio);
1055                 printed += fprintf(fp, " ] %10.3f ms\n", ttrace->runtime_ms);
1056         }
1057
1058         return printed;
1059 }
1060
1061 static int trace__set_duration(const struct option *opt, const char *str,
1062                                int unset __maybe_unused)
1063 {
1064         struct trace *trace = opt->value;
1065
1066         trace->duration_filter = atof(str);
1067         return 0;
1068 }
1069
1070 static int trace__open_output(struct trace *trace, const char *filename)
1071 {
1072         struct stat st;
1073
1074         if (!stat(filename, &st) && st.st_size) {
1075                 char oldname[PATH_MAX];
1076
1077                 scnprintf(oldname, sizeof(oldname), "%s.old", filename);
1078                 unlink(oldname);
1079                 rename(filename, oldname);
1080         }
1081
1082         trace->output = fopen(filename, "w");
1083
1084         return trace->output == NULL ? -errno : 0;
1085 }
1086
1087 int cmd_trace(int argc, const char **argv, const char *prefix __maybe_unused)
1088 {
1089         const char * const trace_usage[] = {
1090                 "perf trace [<options>] [<command>]",
1091                 "perf trace [<options>] -- <command> [<options>]",
1092                 NULL
1093         };
1094         struct trace trace = {
1095                 .audit_machine = audit_detect_machine(),
1096                 .syscalls = {
1097                         . max = -1,
1098                 },
1099                 .opts = {
1100                         .target = {
1101                                 .uid       = UINT_MAX,
1102                                 .uses_mmap = true,
1103                         },
1104                         .user_freq     = UINT_MAX,
1105                         .user_interval = ULLONG_MAX,
1106                         .no_delay      = true,
1107                         .mmap_pages    = 1024,
1108                 },
1109                 .output = stdout,
1110         };
1111         const char *output_name = NULL;
1112         const char *ev_qualifier_str = NULL;
1113         const struct option trace_options[] = {
1114         OPT_STRING('e', "expr", &ev_qualifier_str, "expr",
1115                     "list of events to trace"),
1116         OPT_STRING('o', "output", &output_name, "file", "output file name"),
1117         OPT_STRING('i', "input", &input_name, "file", "Analyze events in file"),
1118         OPT_STRING('p', "pid", &trace.opts.target.pid, "pid",
1119                     "trace events on existing process id"),
1120         OPT_STRING('t', "tid", &trace.opts.target.tid, "tid",
1121                     "trace events on existing thread id"),
1122         OPT_BOOLEAN('a', "all-cpus", &trace.opts.target.system_wide,
1123                     "system-wide collection from all CPUs"),
1124         OPT_STRING('C', "cpu", &trace.opts.target.cpu_list, "cpu",
1125                     "list of cpus to monitor"),
1126         OPT_BOOLEAN(0, "no-inherit", &trace.opts.no_inherit,
1127                     "child tasks do not inherit counters"),
1128         OPT_UINTEGER('m', "mmap-pages", &trace.opts.mmap_pages,
1129                      "number of mmap data pages"),
1130         OPT_STRING('u', "uid", &trace.opts.target.uid_str, "user",
1131                    "user to profile"),
1132         OPT_CALLBACK(0, "duration", &trace, "float",
1133                      "show only events with duration > N.M ms",
1134                      trace__set_duration),
1135         OPT_BOOLEAN(0, "sched", &trace.sched, "show blocking scheduler events"),
1136         OPT_INCR('v', "verbose", &verbose, "be more verbose"),
1137         OPT_END()
1138         };
1139         int err;
1140         char bf[BUFSIZ];
1141
1142         argc = parse_options(argc, argv, trace_options, trace_usage, 0);
1143
1144         if (output_name != NULL) {
1145                 err = trace__open_output(&trace, output_name);
1146                 if (err < 0) {
1147                         perror("failed to create output file");
1148                         goto out;
1149                 }
1150         }
1151
1152         if (ev_qualifier_str != NULL) {
1153                 const char *s = ev_qualifier_str;
1154
1155                 trace.not_ev_qualifier = *s == '!';
1156                 if (trace.not_ev_qualifier)
1157                         ++s;
1158                 trace.ev_qualifier = strlist__new(true, s);
1159                 if (trace.ev_qualifier == NULL) {
1160                         fputs("Not enough memory to parse event qualifier",
1161                               trace.output);
1162                         err = -ENOMEM;
1163                         goto out_close;
1164                 }
1165         }
1166
1167         err = perf_target__validate(&trace.opts.target);
1168         if (err) {
1169                 perf_target__strerror(&trace.opts.target, err, bf, sizeof(bf));
1170                 fprintf(trace.output, "%s", bf);
1171                 goto out_close;
1172         }
1173
1174         err = perf_target__parse_uid(&trace.opts.target);
1175         if (err) {
1176                 perf_target__strerror(&trace.opts.target, err, bf, sizeof(bf));
1177                 fprintf(trace.output, "%s", bf);
1178                 goto out_close;
1179         }
1180
1181         if (!argc && perf_target__none(&trace.opts.target))
1182                 trace.opts.target.system_wide = true;
1183
1184         if (input_name)
1185                 err = trace__replay(&trace);
1186         else
1187                 err = trace__run(&trace, argc, argv);
1188
1189         if (trace.sched && !err)
1190                 trace__fprintf_thread_summary(&trace, trace.output);
1191
1192 out_close:
1193         if (output_name != NULL)
1194                 fclose(trace.output);
1195 out:
1196         return err;
1197 }