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 size_t syscall_arg__scnprintf_open_flags(char *bf, size_t size,
 210                                                unsigned long arg, u8 *arg_mask)
 211 {
 212         int printed = 0, flags = arg;
 213
 214         if (!(flags & O_CREAT))
 215                 *arg_mask |= 1 << 2; /* Mask the mode parm */
 216
 217         if (flags == 0)
 218                 return scnprintf(bf, size, "RDONLY");
 219 #define P_FLAG(n) \
 220         if (flags & O_##n) { \
 221                 printed += scnprintf(bf + printed, size - printed, "%s%s", printed ? "|" : "", #n); \
 222                 flags &= ~O_##n; \
 223         }
 224
 225         P_FLAG(APPEND);
 226         P_FLAG(ASYNC);
 227         P_FLAG(CLOEXEC);
 228         P_FLAG(CREAT);
 229         P_FLAG(DIRECT);
 230         P_FLAG(DIRECTORY);
 231         P_FLAG(EXCL);
 232         P_FLAG(LARGEFILE);
 233         P_FLAG(NOATIME);
 234         P_FLAG(NOCTTY);
 235 #ifdef O_NONBLOCK
 236         P_FLAG(NONBLOCK);
 237 #elif O_NDELAY
 238         P_FLAG(NDELAY);
 239 #endif
 240 #ifdef O_PATH
 241         P_FLAG(PATH);
 242 #endif
 243         P_FLAG(RDWR);
 244 #ifdef O_DSYNC
 245         if ((flags & O_SYNC) == O_SYNC)
 246                 printed += scnprintf(bf + printed, size - printed, "%s%s", printed ? "|" : "", "SYNC");
 247         else {
 248                 P_FLAG(DSYNC);
 249         }
 250 #else
 251         P_FLAG(SYNC);
 252 #endif
 253         P_FLAG(TRUNC);
 254         P_FLAG(WRONLY);
 255 #undef P_FLAG
 256
 257         if (flags)
 258                 printed += scnprintf(bf + printed, size - printed, "%s%#x", printed ? "|" : "", flags);
 259
 260         return printed;
 261 }
 262
 263 #define SCA_OPEN_FLAGS syscall_arg__scnprintf_open_flags
 264
 265 static struct syscall_fmt {
 266         const char *name;
 267         const char *alias;
 268         size_t     (*arg_scnprintf[6])(char *bf, size_t size, unsigned long arg, u8 *arg_mask);
 269         bool       errmsg;
 270         bool       timeout;
 271         bool       hexret;
 272 } syscall_fmts[] = {
 273         { .name     = "access",     .errmsg = true, },
 274         { .name     = "arch_prctl", .errmsg = true, .alias = "prctl", },
 275         { .name     = "brk",        .hexret = true,
 276           .arg_scnprintf = { [0] = SCA_HEX, /* brk */ }, },
 277         { .name     = "mmap",       .hexret = true, },
 278         { .name     = "connect",    .errmsg = true, },
 279         { .name     = "fstat",      .errmsg = true, .alias = "newfstat", },
 280         { .name     = "fstatat",    .errmsg = true, .alias = "newfstatat", },
 281         { .name     = "futex",      .errmsg = true,
 282           .arg_scnprintf = { [1] = SCA_FUTEX_OP, /* op */ }, },
 283         { .name     = "ioctl",      .errmsg = true,
 284           .arg_scnprintf = { [2] = SCA_HEX, /* arg */ }, },
 285         { .name     = "lseek",      .errmsg = true,
 286           .arg_scnprintf = { [2] = SCA_WHENCE, /* whence */ }, },
 287         { .name     = "lstat",      .errmsg = true, .alias = "newlstat", },
 288         { .name     = "madvise",    .errmsg = true,
 289           .arg_scnprintf = { [0] = SCA_HEX,      /* start */
 290                              [2] = SCA_MADV_BHV, /* behavior */ }, },
 291         { .name     = "mmap",       .hexret = true,
 292           .arg_scnprintf = { [0] = SCA_HEX,       /* addr */
 293                              [2] = SCA_MMAP_PROT, /* prot */
 294                              [3] = SCA_MMAP_FLAGS, /* flags */ }, },
 295         { .name     = "mprotect",   .errmsg = true,
 296           .arg_scnprintf = { [0] = SCA_HEX, /* start */
 297                              [2] = SCA_MMAP_PROT, /* prot */ }, },
 298         { .name     = "mremap",     .hexret = true,
 299           .arg_scnprintf = { [0] = SCA_HEX, /* addr */
 300                              [4] = SCA_HEX, /* new_addr */ }, },
 301         { .name     = "munmap",     .errmsg = true,
 302           .arg_scnprintf = { [0] = SCA_HEX, /* addr */ }, },
 303         { .name     = "open",       .errmsg = true,
 304           .arg_scnprintf = { [1] = SCA_OPEN_FLAGS, /* flags */ }, },
 305         { .name     = "poll",       .errmsg = true, .timeout = true, },
 306         { .name     = "ppoll",      .errmsg = true, .timeout = true, },
 307         { .name     = "pread",      .errmsg = true, .alias = "pread64", },
 308         { .name     = "pwrite",     .errmsg = true, .alias = "pwrite64", },
 309         { .name     = "read",       .errmsg = true, },
 310         { .name     = "recvfrom",   .errmsg = true, },
 311         { .name     = "select",     .errmsg = true, .timeout = true, },
 312         { .name     = "socket",     .errmsg = true, },
 313         { .name     = "stat",       .errmsg = true, .alias = "newstat", },
 314         { .name     = "uname",      .errmsg = true, .alias = "newuname", },
 315 };
 316
 317 static int syscall_fmt__cmp(const void *name, const void *fmtp)
 318 {
 319         const struct syscall_fmt *fmt = fmtp;
 320         return strcmp(name, fmt->name);
 321 }
 322
 323 static struct syscall_fmt *syscall_fmt__find(const char *name)
 324 {
 325         const int nmemb = ARRAY_SIZE(syscall_fmts);
 326         return bsearch(name, syscall_fmts, nmemb, sizeof(struct syscall_fmt), syscall_fmt__cmp);
 327 }
 328
 329 struct syscall {
 330         struct event_format *tp_format;
 331         const char          *name;
 332         bool                filtered;
 333         struct syscall_fmt  *fmt;
 334         size_t              (**arg_scnprintf)(char *bf, size_t size,
 335                                               unsigned long arg, u8 *args_mask);
 336 };
 337
 338 static size_t fprintf_duration(unsigned long t, FILE *fp)
 339 {
 340         double duration = (double)t / NSEC_PER_MSEC;
 341         size_t printed = fprintf(fp, "(");
 342
 343         if (duration >= 1.0)
 344                 printed += color_fprintf(fp, PERF_COLOR_RED, "%6.3f ms", duration);
 345         else if (duration >= 0.01)
 346                 printed += color_fprintf(fp, PERF_COLOR_YELLOW, "%6.3f ms", duration);
 347         else
 348                 printed += color_fprintf(fp, PERF_COLOR_NORMAL, "%6.3f ms", duration);
 349         return printed + fprintf(fp, "): ");
 350 }
 351
 352 struct thread_trace {
 353         u64               entry_time;
 354         u64               exit_time;
 355         bool              entry_pending;
 356         unsigned long     nr_events;
 357         char              *entry_str;
 358         double            runtime_ms;
 359 };
 360
 361 static struct thread_trace *thread_trace__new(void)
 362 {
 363         return zalloc(sizeof(struct thread_trace));
 364 }
 365
 366 static struct thread_trace *thread__trace(struct thread *thread, FILE *fp)
 367 {
 368         struct thread_trace *ttrace;
 369
 370         if (thread == NULL)
 371                 goto fail;
 372
 373         if (thread->priv == NULL)
 374                 thread->priv = thread_trace__new();
 375
 376         if (thread->priv == NULL)
 377                 goto fail;
 378
 379         ttrace = thread->priv;
 380         ++ttrace->nr_events;
 381
 382         return ttrace;
 383 fail:
 384         color_fprintf(fp, PERF_COLOR_RED,
 385                       "WARNING: not enough memory, dropping samples!\n");
 386         return NULL;
 387 }
 388
 389 struct trace {
 390         struct perf_tool        tool;
 391         int                     audit_machine;
 392         struct {
 393                 int             max;
 394                 struct syscall  *table;
 395         } syscalls;
 396         struct perf_record_opts opts;
 397         struct machine          host;
 398         u64                     base_time;
 399         FILE                    *output;
 400         unsigned long           nr_events;
 401         struct strlist          *ev_qualifier;
 402         bool                    not_ev_qualifier;
 403         struct intlist          *tid_list;
 404         struct intlist          *pid_list;
 405         bool                    sched;
 406         bool                    multiple_threads;
 407         double                  duration_filter;
 408         double                  runtime_ms;
 409 };
 410
 411 static bool trace__filter_duration(struct trace *trace, double t)
 412 {
 413         return t < (trace->duration_filter * NSEC_PER_MSEC);
 414 }
 415
 416 static size_t trace__fprintf_tstamp(struct trace *trace, u64 tstamp, FILE *fp)
 417 {
 418         double ts = (double)(tstamp - trace->base_time) / NSEC_PER_MSEC;
 419
 420         return fprintf(fp, "%10.3f ", ts);
 421 }
 422
 423 static bool done = false;
 424
 425 static void sig_handler(int sig __maybe_unused)
 426 {
 427         done = true;
 428 }
 429
 430 static size_t trace__fprintf_entry_head(struct trace *trace, struct thread *thread,
 431                                         u64 duration, u64 tstamp, FILE *fp)
 432 {
 433         size_t printed = trace__fprintf_tstamp(trace, tstamp, fp);
 434         printed += fprintf_duration(duration, fp);
 435
 436         if (trace->multiple_threads)
 437                 printed += fprintf(fp, "%d ", thread->tid);
 438
 439         return printed;
 440 }
 441
 442 static int trace__process_event(struct trace *trace, struct machine *machine,
 443                                 union perf_event *event)
 444 {
 445         int ret = 0;
 446
 447         switch (event->header.type) {
 448         case PERF_RECORD_LOST:
 449                 color_fprintf(trace->output, PERF_COLOR_RED,
 450                               "LOST %" PRIu64 " events!\n", event->lost.lost);
 451                 ret = machine__process_lost_event(machine, event);
 452         default:
 453                 ret = machine__process_event(machine, event);
 454                 break;
 455         }
 456
 457         return ret;
 458 }
 459
 460 static int trace__tool_process(struct perf_tool *tool,
 461                                union perf_event *event,
 462                                struct perf_sample *sample __maybe_unused,
 463                                struct machine *machine)
 464 {
 465         struct trace *trace = container_of(tool, struct trace, tool);
 466         return trace__process_event(trace, machine, event);
 467 }
 468
 469 static int trace__symbols_init(struct trace *trace, struct perf_evlist *evlist)
 470 {
 471         int err = symbol__init();
 472
 473         if (err)
 474                 return err;
 475
 476         machine__init(&trace->host, "", HOST_KERNEL_ID);
 477         machine__create_kernel_maps(&trace->host);
 478
 479         if (perf_target__has_task(&trace->opts.target)) {
 480                 err = perf_event__synthesize_thread_map(&trace->tool, evlist->threads,
 481                                                         trace__tool_process,
 482                                                         &trace->host);
 483         } else {
 484                 err = perf_event__synthesize_threads(&trace->tool, trace__tool_process,
 485                                                      &trace->host);
 486         }
 487
 488         if (err)
 489                 symbol__exit();
 490
 491         return err;
 492 }
 493
 494 static int syscall__set_arg_fmts(struct syscall *sc)
 495 {
 496         struct format_field *field;
 497         int idx = 0;
 498
 499         sc->arg_scnprintf = calloc(sc->tp_format->format.nr_fields - 1, sizeof(void *));
 500         if (sc->arg_scnprintf == NULL)
 501                 return -1;
 502
 503         for (field = sc->tp_format->format.fields->next; field; field = field->next) {
 504                 if (sc->fmt && sc->fmt->arg_scnprintf[idx])
 505                         sc->arg_scnprintf[idx] = sc->fmt->arg_scnprintf[idx];
 506                 else if (field->flags & FIELD_IS_POINTER)
 507                         sc->arg_scnprintf[idx] = syscall_arg__scnprintf_hex;
 508                 ++idx;
 509         }
 510
 511         return 0;
 512 }
 513
 514 static int trace__read_syscall_info(struct trace *trace, int id)
 515 {
 516         char tp_name[128];
 517         struct syscall *sc;
 518         const char *name = audit_syscall_to_name(id, trace->audit_machine);
 519
 520         if (name == NULL)
 521                 return -1;
 522
 523         if (id > trace->syscalls.max) {
 524                 struct syscall *nsyscalls = realloc(trace->syscalls.table, (id + 1) * sizeof(*sc));
 525
 526                 if (nsyscalls == NULL)
 527                         return -1;
 528
 529                 if (trace->syscalls.max != -1) {
 530                         memset(nsyscalls + trace->syscalls.max + 1, 0,
 531                                (id - trace->syscalls.max) * sizeof(*sc));
 532                 } else {
 533                         memset(nsyscalls, 0, (id + 1) * sizeof(*sc));
 534                 }
 535
 536                 trace->syscalls.table = nsyscalls;
 537                 trace->syscalls.max   = id;
 538         }
 539
 540         sc = trace->syscalls.table + id;
 541         sc->name = name;
 542
 543         if (trace->ev_qualifier) {
 544                 bool in = strlist__find(trace->ev_qualifier, name) != NULL;
 545
 546                 if (!(in ^ trace->not_ev_qualifier)) {
 547                         sc->filtered = true;
 548                         /*
 549                          * No need to do read tracepoint information since this will be
 550                          * filtered out.
 551                          */
 552                         return 0;
 553                 }
 554         }
 555
 556         sc->fmt  = syscall_fmt__find(sc->name);
 557
 558         snprintf(tp_name, sizeof(tp_name), "sys_enter_%s", sc->name);
 559         sc->tp_format = event_format__new("syscalls", tp_name);
 560
 561         if (sc->tp_format == NULL && sc->fmt && sc->fmt->alias) {
 562                 snprintf(tp_name, sizeof(tp_name), "sys_enter_%s", sc->fmt->alias);
 563                 sc->tp_format = event_format__new("syscalls", tp_name);
 564         }
 565
 566         if (sc->tp_format == NULL)
 567                 return -1;
 568
 569         return syscall__set_arg_fmts(sc);
 570 }
 571
 572 static size_t syscall__scnprintf_args(struct syscall *sc, char *bf, size_t size,
 573                                       unsigned long *args)
 574 {
 575         int i = 0;
 576         size_t printed = 0;
 577
 578         if (sc->tp_format != NULL) {
 579                 struct format_field *field;
 580                 u8 mask = 0, bit = 1;
 581
 582                 for (field = sc->tp_format->format.fields->next; field;
 583                      field = field->next, ++i, bit <<= 1) {
 584                         if (mask & bit)
 585                                 continue;
 586
 587                         printed += scnprintf(bf + printed, size - printed,
 588                                              "%s%s: ", printed ? ", " : "", field->name);
 589
 590                         if (sc->arg_scnprintf && sc->arg_scnprintf[i]) {
 591                                 printed += sc->arg_scnprintf[i](bf + printed, size - printed,
 592                                                                 args[i], &mask);
 593                         } else {
 594                                 printed += scnprintf(bf + printed, size - printed,
 595                                                      "%ld", args[i]);
 596                         }
 597                 }
 598         } else {
 599                 while (i < 6) {
 600                         printed += scnprintf(bf + printed, size - printed,
 601                                              "%sarg%d: %ld",
 602                                              printed ? ", " : "", i, args[i]);
 603                         ++i;
 604                 }
 605         }
 606
 607         return printed;
 608 }
 609
 610 typedef int (*tracepoint_handler)(struct trace *trace, struct perf_evsel *evsel,
 611                                   struct perf_sample *sample);
 612
 613 static struct syscall *trace__syscall_info(struct trace *trace,
 614                                            struct perf_evsel *evsel,
 615                                            struct perf_sample *sample)
 616 {
 617         int id = perf_evsel__intval(evsel, sample, "id");
 618
 619         if (id < 0) {
 620
 621                 /*
 622                  * XXX: Noticed on x86_64, reproduced as far back as 3.0.36, haven't tried
 623                  * before that, leaving at a higher verbosity level till that is
 624                  * explained. Reproduced with plain ftrace with:
 625                  *
 626                  * echo 1 > /t/events/raw_syscalls/sys_exit/enable
 627                  * grep "NR -1 " /t/trace_pipe
 628                  *
 629                  * After generating some load on the machine.
 630                  */
 631                 if (verbose > 1) {
 632                         static u64 n;
 633                         fprintf(trace->output, "Invalid syscall %d id, skipping (%s, %" PRIu64 ") ...\n",
 634                                 id, perf_evsel__name(evsel), ++n);
 635                 }
 636                 return NULL;
 637         }
 638
 639         if ((id > trace->syscalls.max || trace->syscalls.table[id].name == NULL) &&
 640             trace__read_syscall_info(trace, id))
 641                 goto out_cant_read;
 642
 643         if ((id > trace->syscalls.max || trace->syscalls.table[id].name == NULL))
 644                 goto out_cant_read;
 645
 646         return &trace->syscalls.table[id];
 647
 648 out_cant_read:
 649         if (verbose) {
 650                 fprintf(trace->output, "Problems reading syscall %d", id);
 651                 if (id <= trace->syscalls.max && trace->syscalls.table[id].name != NULL)
 652                         fprintf(trace->output, "(%s)", trace->syscalls.table[id].name);
 653                 fputs(" information\n", trace->output);
 654         }
 655         return NULL;
 656 }
 657
 658 static int trace__sys_enter(struct trace *trace, struct perf_evsel *evsel,
 659                             struct perf_sample *sample)
 660 {
 661         char *msg;
 662         void *args;
 663         size_t printed = 0;
 664         struct thread *thread;
 665         struct syscall *sc = trace__syscall_info(trace, evsel, sample);
 666         struct thread_trace *ttrace;
 667
 668         if (sc == NULL)
 669                 return -1;
 670
 671         if (sc->filtered)
 672                 return 0;
 673
 674         thread = machine__findnew_thread(&trace->host, sample->pid,
 675                                          sample->tid);
 676         ttrace = thread__trace(thread, trace->output);
 677         if (ttrace == NULL)
 678                 return -1;
 679
 680         args = perf_evsel__rawptr(evsel, sample, "args");
 681         if (args == NULL) {
 682                 fprintf(trace->output, "Problems reading syscall arguments\n");
 683                 return -1;
 684         }
 685
 686         ttrace = thread->priv;
 687
 688         if (ttrace->entry_str == NULL) {
 689                 ttrace->entry_str = malloc(1024);
 690                 if (!ttrace->entry_str)
 691                         return -1;
 692         }
 693
 694         ttrace->entry_time = sample->time;
 695         msg = ttrace->entry_str;
 696         printed += scnprintf(msg + printed, 1024 - printed, "%s(", sc->name);
 697
 698         printed += syscall__scnprintf_args(sc, msg + printed, 1024 - printed,  args);
 699
 700         if (!strcmp(sc->name, "exit_group") || !strcmp(sc->name, "exit")) {
 701                 if (!trace->duration_filter) {
 702                         trace__fprintf_entry_head(trace, thread, 1, sample->time, trace->output);
 703                         fprintf(trace->output, "%-70s\n", ttrace->entry_str);
 704                 }
 705         } else
 706                 ttrace->entry_pending = true;
 707
 708         return 0;
 709 }
 710
 711 static int trace__sys_exit(struct trace *trace, struct perf_evsel *evsel,
 712                            struct perf_sample *sample)
 713 {
 714         int ret;
 715         u64 duration = 0;
 716         struct thread *thread;
 717         struct syscall *sc = trace__syscall_info(trace, evsel, sample);
 718         struct thread_trace *ttrace;
 719
 720         if (sc == NULL)
 721                 return -1;
 722
 723         if (sc->filtered)
 724                 return 0;
 725
 726         thread = machine__findnew_thread(&trace->host, sample->pid,
 727                                          sample->tid);
 728         ttrace = thread__trace(thread, trace->output);
 729         if (ttrace == NULL)
 730                 return -1;
 731
 732         ret = perf_evsel__intval(evsel, sample, "ret");
 733
 734         ttrace = thread->priv;
 735
 736         ttrace->exit_time = sample->time;
 737
 738         if (ttrace->entry_time) {
 739                 duration = sample->time - ttrace->entry_time;
 740                 if (trace__filter_duration(trace, duration))
 741                         goto out;
 742         } else if (trace->duration_filter)
 743                 goto out;
 744
 745         trace__fprintf_entry_head(trace, thread, duration, sample->time, trace->output);
 746
 747         if (ttrace->entry_pending) {
 748                 fprintf(trace->output, "%-70s", ttrace->entry_str);
 749         } else {
 750                 fprintf(trace->output, " ... [");
 751                 color_fprintf(trace->output, PERF_COLOR_YELLOW, "continued");
 752                 fprintf(trace->output, "]: %s()", sc->name);
 753         }
 754
 755         if (sc->fmt == NULL) {
 756 signed_print:
 757                 fprintf(trace->output, ") = %d", ret);
 758         } else if (ret < 0 && sc->fmt->errmsg) {
 759                 char bf[256];
 760                 const char *emsg = strerror_r(-ret, bf, sizeof(bf)),
 761                            *e = audit_errno_to_name(-ret);
 762
 763                 fprintf(trace->output, ") = -1 %s %s", e, emsg);
 764         } else if (ret == 0 && sc->fmt->timeout)
 765                 fprintf(trace->output, ") = 0 Timeout");
 766         else if (sc->fmt->hexret)
 767                 fprintf(trace->output, ") = %#x", ret);
 768         else
 769                 goto signed_print;
 770
 771         fputc('\n', trace->output);
 772 out:
 773         ttrace->entry_pending = false;
 774
 775         return 0;
 776 }
 777
 778 static int trace__sched_stat_runtime(struct trace *trace, struct perf_evsel *evsel,
 779                                      struct perf_sample *sample)
 780 {
 781         u64 runtime = perf_evsel__intval(evsel, sample, "runtime");
 782         double runtime_ms = (double)runtime / NSEC_PER_MSEC;
 783         struct thread *thread = machine__findnew_thread(&trace->host,
 784                                                         sample->pid,
 785                                                         sample->tid);
 786         struct thread_trace *ttrace = thread__trace(thread, trace->output);
 787
 788         if (ttrace == NULL)
 789                 goto out_dump;
 790
 791         ttrace->runtime_ms += runtime_ms;
 792         trace->runtime_ms += runtime_ms;
 793         return 0;
 794
 795 out_dump:
 796         fprintf(trace->output, "%s: comm=%s,pid=%u,runtime=%" PRIu64 ",vruntime=%" PRIu64 ")\n",
 797                evsel->name,
 798                perf_evsel__strval(evsel, sample, "comm"),
 799                (pid_t)perf_evsel__intval(evsel, sample, "pid"),
 800                runtime,
 801                perf_evsel__intval(evsel, sample, "vruntime"));
 802         return 0;
 803 }
 804
 805 static bool skip_sample(struct trace *trace, struct perf_sample *sample)
 806 {
 807         if ((trace->pid_list && intlist__find(trace->pid_list, sample->pid)) ||
 808             (trace->tid_list && intlist__find(trace->tid_list, sample->tid)))
 809                 return false;
 810
 811         if (trace->pid_list || trace->tid_list)
 812                 return true;
 813
 814         return false;
 815 }
 816
 817 static int trace__process_sample(struct perf_tool *tool,
 818                                  union perf_event *event __maybe_unused,
 819                                  struct perf_sample *sample,
 820                                  struct perf_evsel *evsel,
 821                                  struct machine *machine __maybe_unused)
 822 {
 823         struct trace *trace = container_of(tool, struct trace, tool);
 824         int err = 0;
 825
 826         tracepoint_handler handler = evsel->handler.func;
 827
 828         if (skip_sample(trace, sample))
 829                 return 0;
 830
 831         if (trace->base_time == 0)
 832                 trace->base_time = sample->time;
 833
 834         if (handler)
 835                 handler(trace, evsel, sample);
 836
 837         return err;
 838 }
 839
 840 static bool
 841 perf_session__has_tp(struct perf_session *session, const char *name)
 842 {
 843         struct perf_evsel *evsel;
 844
 845         evsel = perf_evlist__find_tracepoint_by_name(session->evlist, name);
 846
 847         return evsel != NULL;
 848 }
 849
 850 static int parse_target_str(struct trace *trace)
 851 {
 852         if (trace->opts.target.pid) {
 853                 trace->pid_list = intlist__new(trace->opts.target.pid);
 854                 if (trace->pid_list == NULL) {
 855                         pr_err("Error parsing process id string\n");
 856                         return -EINVAL;
 857                 }
 858         }
 859
 860         if (trace->opts.target.tid) {
 861                 trace->tid_list = intlist__new(trace->opts.target.tid);
 862                 if (trace->tid_list == NULL) {
 863                         pr_err("Error parsing thread id string\n");
 864                         return -EINVAL;
 865                 }
 866         }
 867
 868         return 0;
 869 }
 870
 871 static int trace__run(struct trace *trace, int argc, const char **argv)
 872 {
 873         struct perf_evlist *evlist = perf_evlist__new();
 874         struct perf_evsel *evsel;
 875         int err = -1, i;
 876         unsigned long before;
 877         const bool forks = argc > 0;
 878
 879         if (evlist == NULL) {
 880                 fprintf(trace->output, "Not enough memory to run!\n");
 881                 goto out;
 882         }
 883
 884         if (perf_evlist__add_newtp(evlist, "raw_syscalls", "sys_enter", trace__sys_enter) ||
 885             perf_evlist__add_newtp(evlist, "raw_syscalls", "sys_exit", trace__sys_exit)) {
 886                 fprintf(trace->output, "Couldn't read the raw_syscalls tracepoints information!\n");
 887                 goto out_delete_evlist;
 888         }
 889
 890         if (trace->sched &&
 891             perf_evlist__add_newtp(evlist, "sched", "sched_stat_runtime",
 892                                    trace__sched_stat_runtime)) {
 893                 fprintf(trace->output, "Couldn't read the sched_stat_runtime tracepoint information!\n");
 894                 goto out_delete_evlist;
 895         }
 896
 897         err = perf_evlist__create_maps(evlist, &trace->opts.target);
 898         if (err < 0) {
 899                 fprintf(trace->output, "Problems parsing the target to trace, check your options!\n");
 900                 goto out_delete_evlist;
 901         }
 902
 903         err = trace__symbols_init(trace, evlist);
 904         if (err < 0) {
 905                 fprintf(trace->output, "Problems initializing symbol libraries!\n");
 906                 goto out_delete_maps;
 907         }
 908
 909         perf_evlist__config(evlist, &trace->opts);
 910
 911         signal(SIGCHLD, sig_handler);
 912         signal(SIGINT, sig_handler);
 913
 914         if (forks) {
 915                 err = perf_evlist__prepare_workload(evlist, &trace->opts.target,
 916                                                     argv, false, false);
 917                 if (err < 0) {
 918                         fprintf(trace->output, "Couldn't run the workload!\n");
 919                         goto out_delete_maps;
 920                 }
 921         }
 922
 923         err = perf_evlist__open(evlist);
 924         if (err < 0) {
 925                 fprintf(trace->output, "Couldn't create the events: %s\n", strerror(errno));
 926                 goto out_delete_maps;
 927         }
 928
 929         err = perf_evlist__mmap(evlist, UINT_MAX, false);
 930         if (err < 0) {
 931                 fprintf(trace->output, "Couldn't mmap the events: %s\n", strerror(errno));
 932                 goto out_close_evlist;
 933         }
 934
 935         perf_evlist__enable(evlist);
 936
 937         if (forks)
 938                 perf_evlist__start_workload(evlist);
 939
 940         trace->multiple_threads = evlist->threads->map[0] == -1 || evlist->threads->nr > 1;
 941 again:
 942         before = trace->nr_events;
 943
 944         for (i = 0; i < evlist->nr_mmaps; i++) {
 945                 union perf_event *event;
 946
 947                 while ((event = perf_evlist__mmap_read(evlist, i)) != NULL) {
 948                         const u32 type = event->header.type;
 949                         tracepoint_handler handler;
 950                         struct perf_sample sample;
 951
 952                         ++trace->nr_events;
 953
 954                         err = perf_evlist__parse_sample(evlist, event, &sample);
 955                         if (err) {
 956                                 fprintf(trace->output, "Can't parse sample, err = %d, skipping...\n", err);
 957                                 continue;
 958                         }
 959
 960                         if (trace->base_time == 0)
 961                                 trace->base_time = sample.time;
 962
 963                         if (type != PERF_RECORD_SAMPLE) {
 964                                 trace__process_event(trace, &trace->host, event);
 965                                 continue;
 966                         }
 967
 968                         evsel = perf_evlist__id2evsel(evlist, sample.id);
 969                         if (evsel == NULL) {
 970                                 fprintf(trace->output, "Unknown tp ID %" PRIu64 ", skipping...\n", sample.id);
 971                                 continue;
 972                         }
 973
 974                         if (sample.raw_data == NULL) {
 975                                 fprintf(trace->output, "%s sample with no payload for tid: %d, cpu %d, raw_size=%d, skipping...\n",
 976                                        perf_evsel__name(evsel), sample.tid,
 977                                        sample.cpu, sample.raw_size);
 978                                 continue;
 979                         }
 980
 981                         handler = evsel->handler.func;
 982                         handler(trace, evsel, &sample);
 983                 }
 984         }
 985
 986         if (trace->nr_events == before) {
 987                 if (done)
 988                         goto out_unmap_evlist;
 989
 990                 poll(evlist->pollfd, evlist->nr_fds, -1);
 991         }
 992
 993         if (done)
 994                 perf_evlist__disable(evlist);
 995
 996         goto again;
 997
 998 out_unmap_evlist:
 999         perf_evlist__munmap(evlist);
1000 out_close_evlist:
1001         perf_evlist__close(evlist);
1002 out_delete_maps:
1003         perf_evlist__delete_maps(evlist);
1004 out_delete_evlist:
1005         perf_evlist__delete(evlist);
1006 out:
1007         return err;
1008 }
1009
1010 static int trace__replay(struct trace *trace)
1011 {
1012         const struct perf_evsel_str_handler handlers[] = {
1013                 { "raw_syscalls:sys_enter",  trace__sys_enter, },
1014                 { "raw_syscalls:sys_exit",   trace__sys_exit, },
1015         };
1016
1017         struct perf_session *session;
1018         int err = -1;
1019
1020         trace->tool.sample        = trace__process_sample;
1021         trace->tool.mmap          = perf_event__process_mmap;
1022         trace->tool.comm          = perf_event__process_comm;
1023         trace->tool.exit          = perf_event__process_exit;
1024         trace->tool.fork          = perf_event__process_fork;
1025         trace->tool.attr          = perf_event__process_attr;
1026         trace->tool.tracing_data = perf_event__process_tracing_data;
1027         trace->tool.build_id      = perf_event__process_build_id;
1028
1029         trace->tool.ordered_samples = true;
1030         trace->tool.ordering_requires_timestamps = true;
1031
1032         /* add tid to output */
1033         trace->multiple_threads = true;
1034
1035         if (symbol__init() < 0)
1036                 return -1;
1037
1038         session = perf_session__new(input_name, O_RDONLY, 0, false,
1039                                     &trace->tool);
1040         if (session == NULL)
1041                 return -ENOMEM;
1042
1043         err = perf_session__set_tracepoints_handlers(session, handlers);
1044         if (err)
1045                 goto out;
1046
1047         if (!perf_session__has_tp(session, "raw_syscalls:sys_enter")) {
1048                 pr_err("Data file does not have raw_syscalls:sys_enter events\n");
1049                 goto out;
1050         }
1051
1052         if (!perf_session__has_tp(session, "raw_syscalls:sys_exit")) {
1053                 pr_err("Data file does not have raw_syscalls:sys_exit events\n");
1054                 goto out;
1055         }
1056
1057         err = parse_target_str(trace);
1058         if (err != 0)
1059                 goto out;
1060
1061         setup_pager();
1062
1063         err = perf_session__process_events(session, &trace->tool);
1064         if (err)
1065                 pr_err("Failed to process events, error %d", err);
1066
1067 out:
1068         perf_session__delete(session);
1069
1070         return err;
1071 }
1072
1073 static size_t trace__fprintf_threads_header(FILE *fp)
1074 {
1075         size_t printed;
1076
1077         printed  = fprintf(fp, "\n _____________________________________________________________________\n");
1078         printed += fprintf(fp," __)    Summary of events    (__\n\n");
1079         printed += fprintf(fp,"              [ task - pid ]     [ events ] [ ratio ]  [ runtime ]\n");
1080         printed += fprintf(fp," _____________________________________________________________________\n\n");
1081
1082         return printed;
1083 }
1084
1085 static size_t trace__fprintf_thread_summary(struct trace *trace, FILE *fp)
1086 {
1087         size_t printed = trace__fprintf_threads_header(fp);
1088         struct rb_node *nd;
1089
1090         for (nd = rb_first(&trace->host.threads); nd; nd = rb_next(nd)) {
1091                 struct thread *thread = rb_entry(nd, struct thread, rb_node);
1092                 struct thread_trace *ttrace = thread->priv;
1093                 const char *color;
1094                 double ratio;
1095
1096                 if (ttrace == NULL)
1097                         continue;
1098
1099                 ratio = (double)ttrace->nr_events / trace->nr_events * 100.0;
1100
1101                 color = PERF_COLOR_NORMAL;
1102                 if (ratio > 50.0)
1103                         color = PERF_COLOR_RED;
1104                 else if (ratio > 25.0)
1105                         color = PERF_COLOR_GREEN;
1106                 else if (ratio > 5.0)
1107                         color = PERF_COLOR_YELLOW;
1108
1109                 printed += color_fprintf(fp, color, "%20s", thread->comm);
1110                 printed += fprintf(fp, " - %-5d :%11lu   [", thread->tid, ttrace->nr_events);
1111                 printed += color_fprintf(fp, color, "%5.1f%%", ratio);
1112                 printed += fprintf(fp, " ] %10.3f ms\n", ttrace->runtime_ms);
1113         }
1114
1115         return printed;
1116 }
1117
1118 static int trace__set_duration(const struct option *opt, const char *str,
1119                                int unset __maybe_unused)
1120 {
1121         struct trace *trace = opt->value;
1122
1123         trace->duration_filter = atof(str);
1124         return 0;
1125 }
1126
1127 static int trace__open_output(struct trace *trace, const char *filename)
1128 {
1129         struct stat st;
1130
1131         if (!stat(filename, &st) && st.st_size) {
1132                 char oldname[PATH_MAX];
1133
1134                 scnprintf(oldname, sizeof(oldname), "%s.old", filename);
1135                 unlink(oldname);
1136                 rename(filename, oldname);
1137         }
1138
1139         trace->output = fopen(filename, "w");
1140
1141         return trace->output == NULL ? -errno : 0;
1142 }
1143
1144 int cmd_trace(int argc, const char **argv, const char *prefix __maybe_unused)
1145 {
1146         const char * const trace_usage[] = {
1147                 "perf trace [<options>] [<command>]",
1148                 "perf trace [<options>] -- <command> [<options>]",
1149                 NULL
1150         };
1151         struct trace trace = {
1152                 .audit_machine = audit_detect_machine(),
1153                 .syscalls = {
1154                         . max = -1,
1155                 },
1156                 .opts = {
1157                         .target = {
1158                                 .uid       = UINT_MAX,
1159                                 .uses_mmap = true,
1160                         },
1161                         .user_freq     = UINT_MAX,
1162                         .user_interval = ULLONG_MAX,
1163                         .no_delay      = true,
1164                         .mmap_pages    = 1024,
1165                 },
1166                 .output = stdout,
1167         };
1168         const char *output_name = NULL;
1169         const char *ev_qualifier_str = NULL;
1170         const struct option trace_options[] = {
1171         OPT_STRING('e', "expr", &ev_qualifier_str, "expr",
1172                     "list of events to trace"),
1173         OPT_STRING('o', "output", &output_name, "file", "output file name"),
1174         OPT_STRING('i', "input", &input_name, "file", "Analyze events in file"),
1175         OPT_STRING('p', "pid", &trace.opts.target.pid, "pid",
1176                     "trace events on existing process id"),
1177         OPT_STRING('t', "tid", &trace.opts.target.tid, "tid",
1178                     "trace events on existing thread id"),
1179         OPT_BOOLEAN('a', "all-cpus", &trace.opts.target.system_wide,
1180                     "system-wide collection from all CPUs"),
1181         OPT_STRING('C', "cpu", &trace.opts.target.cpu_list, "cpu",
1182                     "list of cpus to monitor"),
1183         OPT_BOOLEAN(0, "no-inherit", &trace.opts.no_inherit,
1184                     "child tasks do not inherit counters"),
1185         OPT_UINTEGER('m', "mmap-pages", &trace.opts.mmap_pages,
1186                      "number of mmap data pages"),
1187         OPT_STRING('u', "uid", &trace.opts.target.uid_str, "user",
1188                    "user to profile"),
1189         OPT_CALLBACK(0, "duration", &trace, "float",
1190                      "show only events with duration > N.M ms",
1191                      trace__set_duration),
1192         OPT_BOOLEAN(0, "sched", &trace.sched, "show blocking scheduler events"),
1193         OPT_INCR('v', "verbose", &verbose, "be more verbose"),
1194         OPT_END()
1195         };
1196         int err;
1197         char bf[BUFSIZ];
1198
1199         argc = parse_options(argc, argv, trace_options, trace_usage, 0);
1200
1201         if (output_name != NULL) {
1202                 err = trace__open_output(&trace, output_name);
1203                 if (err < 0) {
1204                         perror("failed to create output file");
1205                         goto out;
1206                 }
1207         }
1208
1209         if (ev_qualifier_str != NULL) {
1210                 const char *s = ev_qualifier_str;
1211
1212                 trace.not_ev_qualifier = *s == '!';
1213                 if (trace.not_ev_qualifier)
1214                         ++s;
1215                 trace.ev_qualifier = strlist__new(true, s);
1216                 if (trace.ev_qualifier == NULL) {
1217                         fputs("Not enough memory to parse event qualifier",
1218                               trace.output);
1219                         err = -ENOMEM;
1220                         goto out_close;
1221                 }
1222         }
1223
1224         err = perf_target__validate(&trace.opts.target);
1225         if (err) {
1226                 perf_target__strerror(&trace.opts.target, err, bf, sizeof(bf));
1227                 fprintf(trace.output, "%s", bf);
1228                 goto out_close;
1229         }
1230
1231         err = perf_target__parse_uid(&trace.opts.target);
1232         if (err) {
1233                 perf_target__strerror(&trace.opts.target, err, bf, sizeof(bf));
1234                 fprintf(trace.output, "%s", bf);
1235                 goto out_close;
1236         }
1237
1238         if (!argc && perf_target__none(&trace.opts.target))
1239                 trace.opts.target.system_wide = true;
1240
1241         if (input_name)
1242                 err = trace__replay(&trace);
1243         else
1244                 err = trace__run(&trace, argc, argv);
1245
1246         if (trace.sched && !err)
1247                 trace__fprintf_thread_summary(&trace, trace.output);
1248
1249 out_close:
1250         if (output_name != NULL)
1251                 fclose(trace.output);
1252 out:
1253         return err;
1254 }