tools/perf/builtin-trace.c

   1 /*
   2  * builtin-trace.c
   3  *
   4  * Builtin 'trace' command:
   5  *
   6  * Display a continuously updated trace of any workload, CPU, specific PID,
   7  * system wide, etc.  Default format is loosely strace like, but any other
   8  * event may be specified using --event.
   9  *
  10  * Copyright (C) 2012, 2013, 2014, 2015 Red Hat Inc, Arnaldo Carvalho de Melo <acme@redhat.com>
  11  *
  12  * Initially based on the 'trace' prototype by Thomas Gleixner:
  13  *
  14  * http://lwn.net/Articles/415728/ ("Announcing a new utility: 'trace'")
  15  *
  16  * Released under the GPL v2. (and only v2, not any later version)
  17  */
  18
  19 #include <traceevent/event-parse.h>
  20 #include <api/fs/tracing_path.h>
  21 #include "builtin.h"
  22 #include "util/cgroup.h"
  23 #include "util/color.h"
  24 #include "util/debug.h"
  25 #include "util/env.h"
  26 #include "util/event.h"
  27 #include "util/evlist.h"
  28 #include <subcmd/exec-cmd.h>
  29 #include "util/machine.h"
  30 #include "util/path.h"
  31 #include "util/session.h"
  32 #include "util/thread.h"
  33 #include <subcmd/parse-options.h>
  34 #include "util/strlist.h"
  35 #include "util/intlist.h"
  36 #include "util/thread_map.h"
  37 #include "util/stat.h"
  38 #include "trace/beauty/beauty.h"
  39 #include "trace-event.h"
  40 #include "util/parse-events.h"
  41 #include "util/bpf-loader.h"
  42 #include "callchain.h"
  43 #include "print_binary.h"
  44 #include "string2.h"
  45 #include "syscalltbl.h"
  46 #include "rb_resort.h"
  47
  48 #include <errno.h>
  49 #include <inttypes.h>
  50 #include <poll.h>
  51 #include <signal.h>
  52 #include <stdlib.h>
  53 #include <string.h>
  54 #include <linux/err.h>
  55 #include <linux/filter.h>
  56 #include <linux/kernel.h>
  57 #include <linux/random.h>
  58 #include <linux/stringify.h>
  59 #include <linux/time64.h>
  60 #include <fcntl.h>
  61
  62 #include "sane_ctype.h"
  63
  64 #ifndef O_CLOEXEC
  65 # define O_CLOEXEC              02000000
  66 #endif
  67
  68 #ifndef F_LINUX_SPECIFIC_BASE
  69 # define F_LINUX_SPECIFIC_BASE  1024
  70 #endif
  71
  72 struct trace {
  73         struct perf_tool        tool;
  74         struct syscalltbl       *sctbl;
  75         struct {
  76                 int             max;
  77                 struct syscall  *table;
  78                 struct {
  79                         struct perf_evsel *sys_enter,
  80                                           *sys_exit,
  81                                           *augmented;
  82                 }               events;
  83         } syscalls;
  84         struct record_opts      opts;
  85         struct perf_evlist      *evlist;
  86         struct machine          *host;
  87         struct thread           *current;
  88         struct cgroup           *cgroup;
  89         u64                     base_time;
  90         FILE                    *output;
  91         unsigned long           nr_events;
  92         unsigned long           nr_events_printed;
  93         unsigned long           max_events;
  94         struct strlist          *ev_qualifier;
  95         struct {
  96                 size_t          nr;
  97                 int             *entries;
  98         }                       ev_qualifier_ids;
  99         struct {
 100                 size_t          nr;
 101                 pid_t           *entries;
 102         }                       filter_pids;
 103         double                  duration_filter;
 104         double                  runtime_ms;
 105         struct {
 106                 u64             vfs_getname,
 107                                 proc_getname;
 108         } stats;
 109         unsigned int            max_stack;
 110         unsigned int            min_stack;
 111         bool                    not_ev_qualifier;
 112         bool                    live;
 113         bool                    full_time;
 114         bool                    sched;
 115         bool                    multiple_threads;
 116         bool                    summary;
 117         bool                    summary_only;
 118         bool                    failure_only;
 119         bool                    show_comm;
 120         bool                    print_sample;
 121         bool                    show_tool_stats;
 122         bool                    trace_syscalls;
 123         bool                    kernel_syscallchains;
 124         bool                    force;
 125         bool                    vfs_getname;
 126         int                     trace_pgfaults;
 127 };
 128
 129 struct tp_field {
 130         int offset;
 131         union {
 132                 u64 (*integer)(struct tp_field *field, struct perf_sample *sample);
 133                 void *(*pointer)(struct tp_field *field, struct perf_sample *sample);
 134         };
 135 };
 136
 137 #define TP_UINT_FIELD(bits) \
 138 static u64 tp_field__u##bits(struct tp_field *field, struct perf_sample *sample) \
 139 { \
 140         u##bits value; \
 141         memcpy(&value, sample->raw_data + field->offset, sizeof(value)); \
 142         return value;  \
 143 }
 144
 145 TP_UINT_FIELD(8);
 146 TP_UINT_FIELD(16);
 147 TP_UINT_FIELD(32);
 148 TP_UINT_FIELD(64);
 149
 150 #define TP_UINT_FIELD__SWAPPED(bits) \
 151 static u64 tp_field__swapped_u##bits(struct tp_field *field, struct perf_sample *sample) \
 152 { \
 153         u##bits value; \
 154         memcpy(&value, sample->raw_data + field->offset, sizeof(value)); \
 155         return bswap_##bits(value);\
 156 }
 157
 158 TP_UINT_FIELD__SWAPPED(16);
 159 TP_UINT_FIELD__SWAPPED(32);
 160 TP_UINT_FIELD__SWAPPED(64);
 161
 162 static int __tp_field__init_uint(struct tp_field *field, int size, int offset, bool needs_swap)
 163 {
 164         field->offset = offset;
 165
 166         switch (size) {
 167         case 1:
 168                 field->integer = tp_field__u8;
 169                 break;
 170         case 2:
 171                 field->integer = needs_swap ? tp_field__swapped_u16 : tp_field__u16;
 172                 break;
 173         case 4:
 174                 field->integer = needs_swap ? tp_field__swapped_u32 : tp_field__u32;
 175                 break;
 176         case 8:
 177                 field->integer = needs_swap ? tp_field__swapped_u64 : tp_field__u64;
 178                 break;
 179         default:
 180                 return -1;
 181         }
 182
 183         return 0;
 184 }
 185
 186 static int tp_field__init_uint(struct tp_field *field, struct tep_format_field *format_field, bool needs_swap)
 187 {
 188         return __tp_field__init_uint(field, format_field->size, format_field->offset, needs_swap);
 189 }
 190
 191 static void *tp_field__ptr(struct tp_field *field, struct perf_sample *sample)
 192 {
 193         return sample->raw_data + field->offset;
 194 }
 195
 196 static int __tp_field__init_ptr(struct tp_field *field, int offset)
 197 {
 198         field->offset = offset;
 199         field->pointer = tp_field__ptr;
 200         return 0;
 201 }
 202
 203 static int tp_field__init_ptr(struct tp_field *field, struct tep_format_field *format_field)
 204 {
 205         return __tp_field__init_ptr(field, format_field->offset);
 206 }
 207
 208 struct syscall_tp {
 209         struct tp_field id;
 210         union {
 211                 struct tp_field args, ret;
 212         };
 213 };
 214
 215 static int perf_evsel__init_tp_uint_field(struct perf_evsel *evsel,
 216                                           struct tp_field *field,
 217                                           const char *name)
 218 {
 219         struct tep_format_field *format_field = perf_evsel__field(evsel, name);
 220
 221         if (format_field == NULL)
 222                 return -1;
 223
 224         return tp_field__init_uint(field, format_field, evsel->needs_swap);
 225 }
 226
 227 #define perf_evsel__init_sc_tp_uint_field(evsel, name) \
 228         ({ struct syscall_tp *sc = evsel->priv;\
 229            perf_evsel__init_tp_uint_field(evsel, &sc->name, #name); })
 230
 231 static int perf_evsel__init_tp_ptr_field(struct perf_evsel *evsel,
 232                                          struct tp_field *field,
 233                                          const char *name)
 234 {
 235         struct tep_format_field *format_field = perf_evsel__field(evsel, name);
 236
 237         if (format_field == NULL)
 238                 return -1;
 239
 240         return tp_field__init_ptr(field, format_field);
 241 }
 242
 243 #define perf_evsel__init_sc_tp_ptr_field(evsel, name) \
 244         ({ struct syscall_tp *sc = evsel->priv;\
 245            perf_evsel__init_tp_ptr_field(evsel, &sc->name, #name); })
 246
 247 static void perf_evsel__delete_priv(struct perf_evsel *evsel)
 248 {
 249         zfree(&evsel->priv);
 250         perf_evsel__delete(evsel);
 251 }
 252
 253 static int perf_evsel__init_syscall_tp(struct perf_evsel *evsel)
 254 {
 255         struct syscall_tp *sc = evsel->priv = malloc(sizeof(struct syscall_tp));
 256
 257         if (evsel->priv != NULL) {
 258                 if (perf_evsel__init_tp_uint_field(evsel, &sc->id, "__syscall_nr"))
 259                         goto out_delete;
 260                 return 0;
 261         }
 262
 263         return -ENOMEM;
 264 out_delete:
 265         zfree(&evsel->priv);
 266         return -ENOENT;
 267 }
 268
 269 static int perf_evsel__init_augmented_syscall_tp(struct perf_evsel *evsel)
 270 {
 271         struct syscall_tp *sc = evsel->priv = malloc(sizeof(struct syscall_tp));
 272
 273         if (evsel->priv != NULL) {       /* field, sizeof_field, offsetof_field */
 274                 if (__tp_field__init_uint(&sc->id, sizeof(long), sizeof(long long), evsel->needs_swap))
 275                         goto out_delete;
 276
 277                 return 0;
 278         }
 279
 280         return -ENOMEM;
 281 out_delete:
 282         zfree(&evsel->priv);
 283         return -EINVAL;
 284 }
 285
 286 static int perf_evsel__init_augmented_syscall_tp_args(struct perf_evsel *evsel)
 287 {
 288         struct syscall_tp *sc = evsel->priv;
 289
 290         return __tp_field__init_ptr(&sc->args, sc->id.offset + sizeof(u64));
 291 }
 292
 293 static int perf_evsel__init_augmented_syscall_tp_ret(struct perf_evsel *evsel)
 294 {
 295         struct syscall_tp *sc = evsel->priv;
 296
 297         return __tp_field__init_uint(&sc->ret, sizeof(u64), sc->id.offset + sizeof(u64), evsel->needs_swap);
 298 }
 299
 300 static int perf_evsel__init_raw_syscall_tp(struct perf_evsel *evsel, void *handler)
 301 {
 302         evsel->priv = malloc(sizeof(struct syscall_tp));
 303         if (evsel->priv != NULL) {
 304                 if (perf_evsel__init_sc_tp_uint_field(evsel, id))
 305                         goto out_delete;
 306
 307                 evsel->handler = handler;
 308                 return 0;
 309         }
 310
 311         return -ENOMEM;
 312
 313 out_delete:
 314         zfree(&evsel->priv);
 315         return -ENOENT;
 316 }
 317
 318 static struct perf_evsel *perf_evsel__raw_syscall_newtp(const char *direction, void *handler)
 319 {
 320         struct perf_evsel *evsel = perf_evsel__newtp("raw_syscalls", direction);
 321
 322         /* older kernel (e.g., RHEL6) use syscalls:{enter,exit} */
 323         if (IS_ERR(evsel))
 324                 evsel = perf_evsel__newtp("syscalls", direction);
 325
 326         if (IS_ERR(evsel))
 327                 return NULL;
 328
 329         if (perf_evsel__init_raw_syscall_tp(evsel, handler))
 330                 goto out_delete;
 331
 332         return evsel;
 333
 334 out_delete:
 335         perf_evsel__delete_priv(evsel);
 336         return NULL;
 337 }
 338
 339 #define perf_evsel__sc_tp_uint(evsel, name, sample) \
 340         ({ struct syscall_tp *fields = evsel->priv; \
 341            fields->name.integer(&fields->name, sample); })
 342
 343 #define perf_evsel__sc_tp_ptr(evsel, name, sample) \
 344         ({ struct syscall_tp *fields = evsel->priv; \
 345            fields->name.pointer(&fields->name, sample); })
 346
 347 size_t strarray__scnprintf(struct strarray *sa, char *bf, size_t size, const char *intfmt, int val)
 348 {
 349         int idx = val - sa->offset;
 350
 351         if (idx < 0 || idx >= sa->nr_entries || sa->entries[idx] == NULL)
 352                 return scnprintf(bf, size, intfmt, val);
 353
 354         return scnprintf(bf, size, "%s", sa->entries[idx]);
 355 }
 356
 357 static size_t __syscall_arg__scnprintf_strarray(char *bf, size_t size,
 358                                                 const char *intfmt,
 359                                                 struct syscall_arg *arg)
 360 {
 361         return strarray__scnprintf(arg->parm, bf, size, intfmt, arg->val);
 362 }
 363
 364 static size_t syscall_arg__scnprintf_strarray(char *bf, size_t size,
 365                                               struct syscall_arg *arg)
 366 {
 367         return __syscall_arg__scnprintf_strarray(bf, size, "%d", arg);
 368 }
 369
 370 #define SCA_STRARRAY syscall_arg__scnprintf_strarray
 371
 372 struct strarrays {
 373         int             nr_entries;
 374         struct strarray **entries;
 375 };
 376
 377 #define DEFINE_STRARRAYS(array) struct strarrays strarrays__##array = { \
 378         .nr_entries = ARRAY_SIZE(array), \
 379         .entries = array, \
 380 }
 381
 382 size_t syscall_arg__scnprintf_strarrays(char *bf, size_t size,
 383                                         struct syscall_arg *arg)
 384 {
 385         struct strarrays *sas = arg->parm;
 386         int i;
 387
 388         for (i = 0; i < sas->nr_entries; ++i) {
 389                 struct strarray *sa = sas->entries[i];
 390                 int idx = arg->val - sa->offset;
 391
 392                 if (idx >= 0 && idx < sa->nr_entries) {
 393                         if (sa->entries[idx] == NULL)
 394                                 break;
 395                         return scnprintf(bf, size, "%s", sa->entries[idx]);
 396                 }
 397         }
 398
 399         return scnprintf(bf, size, "%d", arg->val);
 400 }
 401
 402 #ifndef AT_FDCWD
 403 #define AT_FDCWD        -100
 404 #endif
 405
 406 static size_t syscall_arg__scnprintf_fd_at(char *bf, size_t size,
 407                                            struct syscall_arg *arg)
 408 {
 409         int fd = arg->val;
 410
 411         if (fd == AT_FDCWD)
 412                 return scnprintf(bf, size, "CWD");
 413
 414         return syscall_arg__scnprintf_fd(bf, size, arg);
 415 }
 416
 417 #define SCA_FDAT syscall_arg__scnprintf_fd_at
 418
 419 static size_t syscall_arg__scnprintf_close_fd(char *bf, size_t size,
 420                                               struct syscall_arg *arg);
 421
 422 #define SCA_CLOSE_FD syscall_arg__scnprintf_close_fd
 423
 424 size_t syscall_arg__scnprintf_hex(char *bf, size_t size, struct syscall_arg *arg)
 425 {
 426         return scnprintf(bf, size, "%#lx", arg->val);
 427 }
 428
 429 size_t syscall_arg__scnprintf_int(char *bf, size_t size, struct syscall_arg *arg)
 430 {
 431         return scnprintf(bf, size, "%d", arg->val);
 432 }
 433
 434 size_t syscall_arg__scnprintf_long(char *bf, size_t size, struct syscall_arg *arg)
 435 {
 436         return scnprintf(bf, size, "%ld", arg->val);
 437 }
 438
 439 static const char *bpf_cmd[] = {
 440         "MAP_CREATE", "MAP_LOOKUP_ELEM", "MAP_UPDATE_ELEM", "MAP_DELETE_ELEM",
 441         "MAP_GET_NEXT_KEY", "PROG_LOAD",
 442 };
 443 static DEFINE_STRARRAY(bpf_cmd);
 444
 445 static const char *epoll_ctl_ops[] = { "ADD", "DEL", "MOD", };
 446 static DEFINE_STRARRAY_OFFSET(epoll_ctl_ops, 1);
 447
 448 static const char *itimers[] = { "REAL", "VIRTUAL", "PROF", };
 449 static DEFINE_STRARRAY(itimers);
 450
 451 static const char *keyctl_options[] = {
 452         "GET_KEYRING_ID", "JOIN_SESSION_KEYRING", "UPDATE", "REVOKE", "CHOWN",
 453         "SETPERM", "DESCRIBE", "CLEAR", "LINK", "UNLINK", "SEARCH", "READ",
 454         "INSTANTIATE", "NEGATE", "SET_REQKEY_KEYRING", "SET_TIMEOUT",
 455         "ASSUME_AUTHORITY", "GET_SECURITY", "SESSION_TO_PARENT", "REJECT",
 456         "INSTANTIATE_IOV", "INVALIDATE", "GET_PERSISTENT",
 457 };
 458 static DEFINE_STRARRAY(keyctl_options);
 459
 460 static const char *whences[] = { "SET", "CUR", "END",
 461 #ifdef SEEK_DATA
 462 "DATA",
 463 #endif
 464 #ifdef SEEK_HOLE
 465 "HOLE",
 466 #endif
 467 };
 468 static DEFINE_STRARRAY(whences);
 469
 470 static const char *fcntl_cmds[] = {
 471         "DUPFD", "GETFD", "SETFD", "GETFL", "SETFL", "GETLK", "SETLK",
 472         "SETLKW", "SETOWN", "GETOWN", "SETSIG", "GETSIG", "GETLK64",
 473         "SETLK64", "SETLKW64", "SETOWN_EX", "GETOWN_EX",
 474         "GETOWNER_UIDS",
 475 };
 476 static DEFINE_STRARRAY(fcntl_cmds);
 477
 478 static const char *fcntl_linux_specific_cmds[] = {
 479         "SETLEASE", "GETLEASE", "NOTIFY", [5] = "CANCELLK", "DUPFD_CLOEXEC",
 480         "SETPIPE_SZ", "GETPIPE_SZ", "ADD_SEALS", "GET_SEALS",
 481         "GET_RW_HINT", "SET_RW_HINT", "GET_FILE_RW_HINT", "SET_FILE_RW_HINT",
 482 };
 483
 484 static DEFINE_STRARRAY_OFFSET(fcntl_linux_specific_cmds, F_LINUX_SPECIFIC_BASE);
 485
 486 static struct strarray *fcntl_cmds_arrays[] = {
 487         &strarray__fcntl_cmds,
 488         &strarray__fcntl_linux_specific_cmds,
 489 };
 490
 491 static DEFINE_STRARRAYS(fcntl_cmds_arrays);
 492
 493 static const char *rlimit_resources[] = {
 494         "CPU", "FSIZE", "DATA", "STACK", "CORE", "RSS", "NPROC", "NOFILE",
 495         "MEMLOCK", "AS", "LOCKS", "SIGPENDING", "MSGQUEUE", "NICE", "RTPRIO",
 496         "RTTIME",
 497 };
 498 static DEFINE_STRARRAY(rlimit_resources);
 499
 500 static const char *sighow[] = { "BLOCK", "UNBLOCK", "SETMASK", };
 501 static DEFINE_STRARRAY(sighow);
 502
 503 static const char *clockid[] = {
 504         "REALTIME", "MONOTONIC", "PROCESS_CPUTIME_ID", "THREAD_CPUTIME_ID",
 505         "MONOTONIC_RAW", "REALTIME_COARSE", "MONOTONIC_COARSE", "BOOTTIME",
 506         "REALTIME_ALARM", "BOOTTIME_ALARM", "SGI_CYCLE", "TAI"
 507 };
 508 static DEFINE_STRARRAY(clockid);
 509
 510 static size_t syscall_arg__scnprintf_access_mode(char *bf, size_t size,
 511                                                  struct syscall_arg *arg)
 512 {
 513         size_t printed = 0;
 514         int mode = arg->val;
 515
 516         if (mode == F_OK) /* 0 */
 517                 return scnprintf(bf, size, "F");
 518 #define P_MODE(n) \
 519         if (mode & n##_OK) { \
 520                 printed += scnprintf(bf + printed, size - printed, "%s", #n); \
 521                 mode &= ~n##_OK; \
 522         }
 523
 524         P_MODE(R);
 525         P_MODE(W);
 526         P_MODE(X);
 527 #undef P_MODE
 528
 529         if (mode)
 530                 printed += scnprintf(bf + printed, size - printed, "|%#x", mode);
 531
 532         return printed;
 533 }
 534
 535 #define SCA_ACCMODE syscall_arg__scnprintf_access_mode
 536
 537 static size_t syscall_arg__scnprintf_filename(char *bf, size_t size,
 538                                               struct syscall_arg *arg);
 539
 540 #define SCA_FILENAME syscall_arg__scnprintf_filename
 541
 542 static size_t syscall_arg__scnprintf_pipe_flags(char *bf, size_t size,
 543                                                 struct syscall_arg *arg)
 544 {
 545         int printed = 0, flags = arg->val;
 546
 547 #define P_FLAG(n) \
 548         if (flags & O_##n) { \
 549                 printed += scnprintf(bf + printed, size - printed, "%s%s", printed ? "|" : "", #n); \
 550                 flags &= ~O_##n; \
 551         }
 552
 553         P_FLAG(CLOEXEC);
 554         P_FLAG(NONBLOCK);
 555 #undef P_FLAG
 556
 557         if (flags)
 558                 printed += scnprintf(bf + printed, size - printed, "%s%#x", printed ? "|" : "", flags);
 559
 560         return printed;
 561 }
 562
 563 #define SCA_PIPE_FLAGS syscall_arg__scnprintf_pipe_flags
 564
 565 #ifndef GRND_NONBLOCK
 566 #define GRND_NONBLOCK   0x0001
 567 #endif
 568 #ifndef GRND_RANDOM
 569 #define GRND_RANDOM     0x0002
 570 #endif
 571
 572 static size_t syscall_arg__scnprintf_getrandom_flags(char *bf, size_t size,
 573                                                    struct syscall_arg *arg)
 574 {
 575         int printed = 0, flags = arg->val;
 576
 577 #define P_FLAG(n) \
 578         if (flags & GRND_##n) { \
 579                 printed += scnprintf(bf + printed, size - printed, "%s%s", printed ? "|" : "", #n); \
 580                 flags &= ~GRND_##n; \
 581         }
 582
 583         P_FLAG(RANDOM);
 584         P_FLAG(NONBLOCK);
 585 #undef P_FLAG
 586
 587         if (flags)
 588                 printed += scnprintf(bf + printed, size - printed, "%s%#x", printed ? "|" : "", flags);
 589
 590         return printed;
 591 }
 592
 593 #define SCA_GETRANDOM_FLAGS syscall_arg__scnprintf_getrandom_flags
 594
 595 #define STRARRAY(name, array) \
 596           { .scnprintf  = SCA_STRARRAY, \
 597             .parm       = &strarray__##array, }
 598
 599 #include "trace/beauty/arch_errno_names.c"
 600 #include "trace/beauty/eventfd.c"
 601 #include "trace/beauty/futex_op.c"
 602 #include "trace/beauty/futex_val3.c"
 603 #include "trace/beauty/mmap.c"
 604 #include "trace/beauty/mode_t.c"
 605 #include "trace/beauty/msg_flags.c"
 606 #include "trace/beauty/open_flags.c"
 607 #include "trace/beauty/perf_event_open.c"
 608 #include "trace/beauty/pid.c"
 609 #include "trace/beauty/sched_policy.c"
 610 #include "trace/beauty/seccomp.c"
 611 #include "trace/beauty/signum.c"
 612 #include "trace/beauty/socket_type.c"
 613 #include "trace/beauty/waitid_options.c"
 614
 615 struct syscall_arg_fmt {
 616         size_t     (*scnprintf)(char *bf, size_t size, struct syscall_arg *arg);
 617         unsigned long (*mask_val)(struct syscall_arg *arg, unsigned long val);
 618         void       *parm;
 619         const char *name;
 620         bool       show_zero;
 621 };
 622
 623 static struct syscall_fmt {
 624         const char *name;
 625         const char *alias;
 626         struct syscall_arg_fmt arg[6];
 627         u8         nr_args;
 628         bool       errpid;
 629         bool       timeout;
 630         bool       hexret;
 631 } syscall_fmts[] = {
 632         { .name     = "access",
 633           .arg = { [1] = { .scnprintf = SCA_ACCMODE,  /* mode */ }, }, },
 634         { .name     = "bind",
 635           .arg = { [1] = { .scnprintf = SCA_SOCKADDR, /* umyaddr */ }, }, },
 636         { .name     = "bpf",
 637           .arg = { [0] = STRARRAY(cmd, bpf_cmd), }, },
 638         { .name     = "brk",        .hexret = true,
 639           .arg = { [0] = { .scnprintf = SCA_HEX, /* brk */ }, }, },
 640         { .name     = "clock_gettime",
 641           .arg = { [0] = STRARRAY(clk_id, clockid), }, },
 642         { .name     = "clone",      .errpid = true, .nr_args = 5,
 643           .arg = { [0] = { .name = "flags",         .scnprintf = SCA_CLONE_FLAGS, },
 644                    [1] = { .name = "child_stack",   .scnprintf = SCA_HEX, },
 645                    [2] = { .name = "parent_tidptr", .scnprintf = SCA_HEX, },
 646                    [3] = { .name = "child_tidptr",  .scnprintf = SCA_HEX, },
 647                    [4] = { .name = "tls",           .scnprintf = SCA_HEX, }, }, },
 648         { .name     = "close",
 649           .arg = { [0] = { .scnprintf = SCA_CLOSE_FD, /* fd */ }, }, },
 650         { .name     = "connect",
 651           .arg = { [1] = { .scnprintf = SCA_SOCKADDR, /* servaddr */ }, }, },
 652         { .name     = "epoll_ctl",
 653           .arg = { [1] = STRARRAY(op, epoll_ctl_ops), }, },
 654         { .name     = "eventfd2",
 655           .arg = { [1] = { .scnprintf = SCA_EFD_FLAGS, /* flags */ }, }, },
 656         { .name     = "fchmodat",
 657           .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
 658         { .name     = "fchownat",
 659           .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
 660         { .name     = "fcntl",
 661           .arg = { [1] = { .scnprintf = SCA_FCNTL_CMD, /* cmd */
 662                            .parm      = &strarrays__fcntl_cmds_arrays,
 663                            .show_zero = true, },
 664                    [2] = { .scnprintf =  SCA_FCNTL_ARG, /* arg */ }, }, },
 665         { .name     = "flock",
 666           .arg = { [1] = { .scnprintf = SCA_FLOCK, /* cmd */ }, }, },
 667         { .name     = "fstat", .alias = "newfstat", },
 668         { .name     = "fstatat", .alias = "newfstatat", },
 669         { .name     = "futex",
 670           .arg = { [1] = { .scnprintf = SCA_FUTEX_OP, /* op */ },
 671                    [5] = { .scnprintf = SCA_FUTEX_VAL3, /* val3 */ }, }, },
 672         { .name     = "futimesat",
 673           .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
 674         { .name     = "getitimer",
 675           .arg = { [0] = STRARRAY(which, itimers), }, },
 676         { .name     = "getpid",     .errpid = true, },
 677         { .name     = "getpgid",    .errpid = true, },
 678         { .name     = "getppid",    .errpid = true, },
 679         { .name     = "getrandom",
 680           .arg = { [2] = { .scnprintf = SCA_GETRANDOM_FLAGS, /* flags */ }, }, },
 681         { .name     = "getrlimit",
 682           .arg = { [0] = STRARRAY(resource, rlimit_resources), }, },
 683         { .name     = "gettid",     .errpid = true, },
 684         { .name     = "ioctl",
 685           .arg = {
 686 #if defined(__i386__) || defined(__x86_64__)
 687 /*
 688  * FIXME: Make this available to all arches.
 689  */
 690                    [1] = { .scnprintf = SCA_IOCTL_CMD, /* cmd */ },
 691                    [2] = { .scnprintf = SCA_HEX, /* arg */ }, }, },
 692 #else
 693                    [2] = { .scnprintf = SCA_HEX, /* arg */ }, }, },
 694 #endif
 695         { .name     = "kcmp",       .nr_args = 5,
 696           .arg = { [0] = { .name = "pid1",      .scnprintf = SCA_PID, },
 697                    [1] = { .name = "pid2",      .scnprintf = SCA_PID, },
 698                    [2] = { .name = "type",      .scnprintf = SCA_KCMP_TYPE, },
 699                    [3] = { .name = "idx1",      .scnprintf = SCA_KCMP_IDX, },
 700                    [4] = { .name = "idx2",      .scnprintf = SCA_KCMP_IDX, }, }, },
 701         { .name     = "keyctl",
 702           .arg = { [0] = STRARRAY(option, keyctl_options), }, },
 703         { .name     = "kill",
 704           .arg = { [1] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
 705         { .name     = "linkat",
 706           .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
 707         { .name     = "lseek",
 708           .arg = { [2] = STRARRAY(whence, whences), }, },
 709         { .name     = "lstat", .alias = "newlstat", },
 710         { .name     = "madvise",
 711           .arg = { [0] = { .scnprintf = SCA_HEX,      /* start */ },
 712                    [2] = { .scnprintf = SCA_MADV_BHV, /* behavior */ }, }, },
 713         { .name     = "mkdirat",
 714           .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
 715         { .name     = "mknodat",
 716           .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
 717         { .name     = "mlock",
 718           .arg = { [0] = { .scnprintf = SCA_HEX, /* addr */ }, }, },
 719         { .name     = "mlockall",
 720           .arg = { [0] = { .scnprintf = SCA_HEX, /* addr */ }, }, },
 721         { .name     = "mmap",       .hexret = true,
 722 /* The standard mmap maps to old_mmap on s390x */
 723 #if defined(__s390x__)
 724         .alias = "old_mmap",
 725 #endif
 726           .arg = { [0] = { .scnprintf = SCA_HEX,        /* addr */ },
 727                    [2] = { .scnprintf = SCA_MMAP_PROT,  /* prot */ },
 728                    [3] = { .scnprintf = SCA_MMAP_FLAGS, /* flags */ }, }, },
 729         { .name     = "mount",
 730           .arg = { [3] = { .scnprintf = SCA_MOUNT_FLAGS, /* flags */
 731                            .mask_val  = SCAMV_MOUNT_FLAGS, /* flags */ }, }, },
 732         { .name     = "mprotect",
 733           .arg = { [0] = { .scnprintf = SCA_HEX,        /* start */ },
 734                    [2] = { .scnprintf = SCA_MMAP_PROT,  /* prot */ }, }, },
 735         { .name     = "mq_unlink",
 736           .arg = { [0] = { .scnprintf = SCA_FILENAME, /* u_name */ }, }, },
 737         { .name     = "mremap",     .hexret = true,
 738           .arg = { [0] = { .scnprintf = SCA_HEX,          /* addr */ },
 739                    [3] = { .scnprintf = SCA_MREMAP_FLAGS, /* flags */ },
 740                    [4] = { .scnprintf = SCA_HEX,          /* new_addr */ }, }, },
 741         { .name     = "munlock",
 742           .arg = { [0] = { .scnprintf = SCA_HEX, /* addr */ }, }, },
 743         { .name     = "munmap",
 744           .arg = { [0] = { .scnprintf = SCA_HEX, /* addr */ }, }, },
 745         { .name     = "name_to_handle_at",
 746           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
 747         { .name     = "newfstatat",
 748           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
 749         { .name     = "open",
 750           .arg = { [1] = { .scnprintf = SCA_OPEN_FLAGS, /* flags */ }, }, },
 751         { .name     = "open_by_handle_at",
 752           .arg = { [0] = { .scnprintf = SCA_FDAT,       /* dfd */ },
 753                    [2] = { .scnprintf = SCA_OPEN_FLAGS, /* flags */ }, }, },
 754         { .name     = "openat",
 755           .arg = { [0] = { .scnprintf = SCA_FDAT,       /* dfd */ },
 756                    [2] = { .scnprintf = SCA_OPEN_FLAGS, /* flags */ }, }, },
 757         { .name     = "perf_event_open",
 758           .arg = { [2] = { .scnprintf = SCA_INT,        /* cpu */ },
 759                    [3] = { .scnprintf = SCA_FD,         /* group_fd */ },
 760                    [4] = { .scnprintf = SCA_PERF_FLAGS, /* flags */ }, }, },
 761         { .name     = "pipe2",
 762           .arg = { [1] = { .scnprintf = SCA_PIPE_FLAGS, /* flags */ }, }, },
 763         { .name     = "pkey_alloc",
 764           .arg = { [1] = { .scnprintf = SCA_PKEY_ALLOC_ACCESS_RIGHTS,   /* access_rights */ }, }, },
 765         { .name     = "pkey_free",
 766           .arg = { [0] = { .scnprintf = SCA_INT,        /* key */ }, }, },
 767         { .name     = "pkey_mprotect",
 768           .arg = { [0] = { .scnprintf = SCA_HEX,        /* start */ },
 769                    [2] = { .scnprintf = SCA_MMAP_PROT,  /* prot */ },
 770                    [3] = { .scnprintf = SCA_INT,        /* pkey */ }, }, },
 771         { .name     = "poll", .timeout = true, },
 772         { .name     = "ppoll", .timeout = true, },
 773         { .name     = "prctl", .alias = "arch_prctl",
 774           .arg = { [0] = { .scnprintf = SCA_PRCTL_OPTION, /* option */ },
 775                    [1] = { .scnprintf = SCA_PRCTL_ARG2, /* arg2 */ },
 776                    [2] = { .scnprintf = SCA_PRCTL_ARG3, /* arg3 */ }, }, },
 777         { .name     = "pread", .alias = "pread64", },
 778         { .name     = "preadv", .alias = "pread", },
 779         { .name     = "prlimit64",
 780           .arg = { [1] = STRARRAY(resource, rlimit_resources), }, },
 781         { .name     = "pwrite", .alias = "pwrite64", },
 782         { .name     = "readlinkat",
 783           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
 784         { .name     = "recvfrom",
 785           .arg = { [3] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
 786         { .name     = "recvmmsg",
 787           .arg = { [3] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
 788         { .name     = "recvmsg",
 789           .arg = { [2] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
 790         { .name     = "renameat",
 791           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
 792         { .name     = "rt_sigaction",
 793           .arg = { [0] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
 794         { .name     = "rt_sigprocmask",
 795           .arg = { [0] = STRARRAY(how, sighow), }, },
 796         { .name     = "rt_sigqueueinfo",
 797           .arg = { [1] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
 798         { .name     = "rt_tgsigqueueinfo",
 799           .arg = { [2] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
 800         { .name     = "sched_setscheduler",
 801           .arg = { [1] = { .scnprintf = SCA_SCHED_POLICY, /* policy */ }, }, },
 802         { .name     = "seccomp",
 803           .arg = { [0] = { .scnprintf = SCA_SECCOMP_OP,    /* op */ },
 804                    [1] = { .scnprintf = SCA_SECCOMP_FLAGS, /* flags */ }, }, },
 805         { .name     = "select", .timeout = true, },
 806         { .name     = "sendmmsg",
 807           .arg = { [3] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
 808         { .name     = "sendmsg",
 809           .arg = { [2] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
 810         { .name     = "sendto",
 811           .arg = { [3] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ },
 812                    [4] = { .scnprintf = SCA_SOCKADDR, /* addr */ }, }, },
 813         { .name     = "set_tid_address", .errpid = true, },
 814         { .name     = "setitimer",
 815           .arg = { [0] = STRARRAY(which, itimers), }, },
 816         { .name     = "setrlimit",
 817           .arg = { [0] = STRARRAY(resource, rlimit_resources), }, },
 818         { .name     = "socket",
 819           .arg = { [0] = STRARRAY(family, socket_families),
 820                    [1] = { .scnprintf = SCA_SK_TYPE, /* type */ },
 821                    [2] = { .scnprintf = SCA_SK_PROTO, /* protocol */ }, }, },
 822         { .name     = "socketpair",
 823           .arg = { [0] = STRARRAY(family, socket_families),
 824                    [1] = { .scnprintf = SCA_SK_TYPE, /* type */ },
 825                    [2] = { .scnprintf = SCA_SK_PROTO, /* protocol */ }, }, },
 826         { .name     = "stat", .alias = "newstat", },
 827         { .name     = "statx",
 828           .arg = { [0] = { .scnprintf = SCA_FDAT,        /* fdat */ },
 829                    [2] = { .scnprintf = SCA_STATX_FLAGS, /* flags */ } ,
 830                    [3] = { .scnprintf = SCA_STATX_MASK,  /* mask */ }, }, },
 831         { .name     = "swapoff",
 832           .arg = { [0] = { .scnprintf = SCA_FILENAME, /* specialfile */ }, }, },
 833         { .name     = "swapon",
 834           .arg = { [0] = { .scnprintf = SCA_FILENAME, /* specialfile */ }, }, },
 835         { .name     = "symlinkat",
 836           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
 837         { .name     = "tgkill",
 838           .arg = { [2] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
 839         { .name     = "tkill",
 840           .arg = { [1] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
 841         { .name     = "umount2", .alias = "umount", },
 842         { .name     = "uname", .alias = "newuname", },
 843         { .name     = "unlinkat",
 844           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
 845         { .name     = "utimensat",
 846           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dirfd */ }, }, },
 847         { .name     = "wait4",      .errpid = true,
 848           .arg = { [2] = { .scnprintf = SCA_WAITID_OPTIONS, /* options */ }, }, },
 849         { .name     = "waitid",     .errpid = true,
 850           .arg = { [3] = { .scnprintf = SCA_WAITID_OPTIONS, /* options */ }, }, },
 851 };
 852
 853 static int syscall_fmt__cmp(const void *name, const void *fmtp)
 854 {
 855         const struct syscall_fmt *fmt = fmtp;
 856         return strcmp(name, fmt->name);
 857 }
 858
 859 static struct syscall_fmt *syscall_fmt__find(const char *name)
 860 {
 861         const int nmemb = ARRAY_SIZE(syscall_fmts);
 862         return bsearch(name, syscall_fmts, nmemb, sizeof(struct syscall_fmt), syscall_fmt__cmp);
 863 }
 864
 865 /*
 866  * is_exit: is this "exit" or "exit_group"?
 867  * is_open: is this "open" or "openat"? To associate the fd returned in sys_exit with the pathname in sys_enter.
 868  * args_size: sum of the sizes of the syscall arguments, anything after that is augmented stuff: pathname for openat, etc.
 869  */
 870 struct syscall {
 871         struct tep_event_format *tp_format;
 872         int                 nr_args;
 873         int                 args_size;
 874         bool                is_exit;
 875         bool                is_open;
 876         struct tep_format_field *args;
 877         const char          *name;
 878         struct syscall_fmt  *fmt;
 879         struct syscall_arg_fmt *arg_fmt;
 880 };
 881
 882 /*
 883  * We need to have this 'calculated' boolean because in some cases we really
 884  * don't know what is the duration of a syscall, for instance, when we start
 885  * a session and some threads are waiting for a syscall to finish, say 'poll',
 886  * in which case all we can do is to print "( ? ) for duration and for the
 887  * start timestamp.
 888  */
 889 static size_t fprintf_duration(unsigned long t, bool calculated, FILE *fp)
 890 {
 891         double duration = (double)t / NSEC_PER_MSEC;
 892         size_t printed = fprintf(fp, "(");
 893
 894         if (!calculated)
 895                 printed += fprintf(fp, "         ");
 896         else if (duration >= 1.0)
 897                 printed += color_fprintf(fp, PERF_COLOR_RED, "%6.3f ms", duration);
 898         else if (duration >= 0.01)
 899                 printed += color_fprintf(fp, PERF_COLOR_YELLOW, "%6.3f ms", duration);
 900         else
 901                 printed += color_fprintf(fp, PERF_COLOR_NORMAL, "%6.3f ms", duration);
 902         return printed + fprintf(fp, "): ");
 903 }
 904
 905 /**
 906  * filename.ptr: The filename char pointer that will be vfs_getname'd
 907  * filename.entry_str_pos: Where to insert the string translated from
 908  *                         filename.ptr by the vfs_getname tracepoint/kprobe.
 909  * ret_scnprintf: syscall args may set this to a different syscall return
 910  *                formatter, for instance, fcntl may return fds, file flags, etc.
 911  */
 912 struct thread_trace {
 913         u64               entry_time;
 914         bool              entry_pending;
 915         unsigned long     nr_events;
 916         unsigned long     pfmaj, pfmin;
 917         char              *entry_str;
 918         double            runtime_ms;
 919         size_t            (*ret_scnprintf)(char *bf, size_t size, struct syscall_arg *arg);
 920         struct {
 921                 unsigned long ptr;
 922                 short int     entry_str_pos;
 923                 bool          pending_open;
 924                 unsigned int  namelen;
 925                 char          *name;
 926         } filename;
 927         struct {
 928                 int       max;
 929                 char      **table;
 930         } paths;
 931
 932         struct intlist *syscall_stats;
 933 };
 934
 935 static struct thread_trace *thread_trace__new(void)
 936 {
 937         struct thread_trace *ttrace =  zalloc(sizeof(struct thread_trace));
 938
 939         if (ttrace)
 940                 ttrace->paths.max = -1;
 941
 942         ttrace->syscall_stats = intlist__new(NULL);
 943
 944         return ttrace;
 945 }
 946
 947 static struct thread_trace *thread__trace(struct thread *thread, FILE *fp)
 948 {
 949         struct thread_trace *ttrace;
 950
 951         if (thread == NULL)
 952                 goto fail;
 953
 954         if (thread__priv(thread) == NULL)
 955                 thread__set_priv(thread, thread_trace__new());
 956
 957         if (thread__priv(thread) == NULL)
 958                 goto fail;
 959
 960         ttrace = thread__priv(thread);
 961         ++ttrace->nr_events;
 962
 963         return ttrace;
 964 fail:
 965         color_fprintf(fp, PERF_COLOR_RED,
 966                       "WARNING: not enough memory, dropping samples!\n");
 967         return NULL;
 968 }
 969
 970
 971 void syscall_arg__set_ret_scnprintf(struct syscall_arg *arg,
 972                                     size_t (*ret_scnprintf)(char *bf, size_t size, struct syscall_arg *arg))
 973 {
 974         struct thread_trace *ttrace = thread__priv(arg->thread);
 975
 976         ttrace->ret_scnprintf = ret_scnprintf;
 977 }
 978
 979 #define TRACE_PFMAJ             (1 << 0)
 980 #define TRACE_PFMIN             (1 << 1)
 981
 982 static const size_t trace__entry_str_size = 2048;
 983
 984 static int trace__set_fd_pathname(struct thread *thread, int fd, const char *pathname)
 985 {
 986         struct thread_trace *ttrace = thread__priv(thread);
 987
 988         if (fd > ttrace->paths.max) {
 989                 char **npath = realloc(ttrace->paths.table, (fd + 1) * sizeof(char *));
 990
 991                 if (npath == NULL)
 992                         return -1;
 993
 994                 if (ttrace->paths.max != -1) {
 995                         memset(npath + ttrace->paths.max + 1, 0,
 996                                (fd - ttrace->paths.max) * sizeof(char *));
 997                 } else {
 998                         memset(npath, 0, (fd + 1) * sizeof(char *));
 999                 }
1000
1001                 ttrace->paths.table = npath;
1002                 ttrace->paths.max   = fd;
1003         }
1004
1005         ttrace->paths.table[fd] = strdup(pathname);
1006
1007         return ttrace->paths.table[fd] != NULL ? 0 : -1;
1008 }
1009
1010 static int thread__read_fd_path(struct thread *thread, int fd)
1011 {
1012         char linkname[PATH_MAX], pathname[PATH_MAX];
1013         struct stat st;
1014         int ret;
1015
1016         if (thread->pid_ == thread->tid) {
1017                 scnprintf(linkname, sizeof(linkname),
1018                           "/proc/%d/fd/%d", thread->pid_, fd);
1019         } else {
1020                 scnprintf(linkname, sizeof(linkname),
1021                           "/proc/%d/task/%d/fd/%d", thread->pid_, thread->tid, fd);
1022         }
1023
1024         if (lstat(linkname, &st) < 0 || st.st_size + 1 > (off_t)sizeof(pathname))
1025                 return -1;
1026
1027         ret = readlink(linkname, pathname, sizeof(pathname));
1028
1029         if (ret < 0 || ret > st.st_size)
1030                 return -1;
1031
1032         pathname[ret] = '\0';
1033         return trace__set_fd_pathname(thread, fd, pathname);
1034 }
1035
1036 static const char *thread__fd_path(struct thread *thread, int fd,
1037                                    struct trace *trace)
1038 {
1039         struct thread_trace *ttrace = thread__priv(thread);
1040
1041         if (ttrace == NULL)
1042                 return NULL;
1043
1044         if (fd < 0)
1045                 return NULL;
1046
1047         if ((fd > ttrace->paths.max || ttrace->paths.table[fd] == NULL)) {
1048                 if (!trace->live)
1049                         return NULL;
1050                 ++trace->stats.proc_getname;
1051                 if (thread__read_fd_path(thread, fd))
1052                         return NULL;
1053         }
1054
1055         return ttrace->paths.table[fd];
1056 }
1057
1058 size_t syscall_arg__scnprintf_fd(char *bf, size_t size, struct syscall_arg *arg)
1059 {
1060         int fd = arg->val;
1061         size_t printed = scnprintf(bf, size, "%d", fd);
1062         const char *path = thread__fd_path(arg->thread, fd, arg->trace);
1063
1064         if (path)
1065                 printed += scnprintf(bf + printed, size - printed, "<%s>", path);
1066
1067         return printed;
1068 }
1069
1070 size_t pid__scnprintf_fd(struct trace *trace, pid_t pid, int fd, char *bf, size_t size)
1071 {
1072         size_t printed = scnprintf(bf, size, "%d", fd);
1073         struct thread *thread = machine__find_thread(trace->host, pid, pid);
1074
1075         if (thread) {
1076                 const char *path = thread__fd_path(thread, fd, trace);
1077
1078                 if (path)
1079                         printed += scnprintf(bf + printed, size - printed, "<%s>", path);
1080
1081                 thread__put(thread);
1082         }
1083
1084         return printed;
1085 }
1086
1087 static size_t syscall_arg__scnprintf_close_fd(char *bf, size_t size,
1088                                               struct syscall_arg *arg)
1089 {
1090         int fd = arg->val;
1091         size_t printed = syscall_arg__scnprintf_fd(bf, size, arg);
1092         struct thread_trace *ttrace = thread__priv(arg->thread);
1093
1094         if (ttrace && fd >= 0 && fd <= ttrace->paths.max)
1095                 zfree(&ttrace->paths.table[fd]);
1096
1097         return printed;
1098 }
1099
1100 static void thread__set_filename_pos(struct thread *thread, const char *bf,
1101                                      unsigned long ptr)
1102 {
1103         struct thread_trace *ttrace = thread__priv(thread);
1104
1105         ttrace->filename.ptr = ptr;
1106         ttrace->filename.entry_str_pos = bf - ttrace->entry_str;
1107 }
1108
1109 static size_t syscall_arg__scnprintf_augmented_string(struct syscall_arg *arg, char *bf, size_t size)
1110 {
1111         struct augmented_arg *augmented_arg = arg->augmented.args;
1112
1113         return scnprintf(bf, size, "%.*s", augmented_arg->size, augmented_arg->value);
1114 }
1115
1116 static size_t syscall_arg__scnprintf_filename(char *bf, size_t size,
1117                                               struct syscall_arg *arg)
1118 {
1119         unsigned long ptr = arg->val;
1120
1121         if (arg->augmented.args)
1122                 return syscall_arg__scnprintf_augmented_string(arg, bf, size);
1123
1124         if (!arg->trace->vfs_getname)
1125                 return scnprintf(bf, size, "%#x", ptr);
1126
1127         thread__set_filename_pos(arg->thread, bf, ptr);
1128         return 0;
1129 }
1130
1131 static bool trace__filter_duration(struct trace *trace, double t)
1132 {
1133         return t < (trace->duration_filter * NSEC_PER_MSEC);
1134 }
1135
1136 static size_t __trace__fprintf_tstamp(struct trace *trace, u64 tstamp, FILE *fp)
1137 {
1138         double ts = (double)(tstamp - trace->base_time) / NSEC_PER_MSEC;
1139
1140         return fprintf(fp, "%10.3f ", ts);
1141 }
1142
1143 /*
1144  * We're handling tstamp=0 as an undefined tstamp, i.e. like when we are
1145  * using ttrace->entry_time for a thread that receives a sys_exit without
1146  * first having received a sys_enter ("poll" issued before tracing session
1147  * starts, lost sys_enter exit due to ring buffer overflow).
1148  */
1149 static size_t trace__fprintf_tstamp(struct trace *trace, u64 tstamp, FILE *fp)
1150 {
1151         if (tstamp > 0)
1152                 return __trace__fprintf_tstamp(trace, tstamp, fp);
1153
1154         return fprintf(fp, "         ? ");
1155 }
1156
1157 static bool done = false;
1158 static bool interrupted = false;
1159
1160 static void sig_handler(int sig)
1161 {
1162         done = true;
1163         interrupted = sig == SIGINT;
1164 }
1165
1166 static size_t trace__fprintf_comm_tid(struct trace *trace, struct thread *thread, FILE *fp)
1167 {
1168         size_t printed = 0;
1169
1170         if (trace->multiple_threads) {
1171                 if (trace->show_comm)
1172                         printed += fprintf(fp, "%.14s/", thread__comm_str(thread));
1173                 printed += fprintf(fp, "%d ", thread->tid);
1174         }
1175
1176         return printed;
1177 }
1178
1179 static size_t trace__fprintf_entry_head(struct trace *trace, struct thread *thread,
1180                                         u64 duration, bool duration_calculated, u64 tstamp, FILE *fp)
1181 {
1182         size_t printed = trace__fprintf_tstamp(trace, tstamp, fp);
1183         printed += fprintf_duration(duration, duration_calculated, fp);
1184         return printed + trace__fprintf_comm_tid(trace, thread, fp);
1185 }
1186
1187 static int trace__process_event(struct trace *trace, struct machine *machine,
1188                                 union perf_event *event, struct perf_sample *sample)
1189 {
1190         int ret = 0;
1191
1192         switch (event->header.type) {
1193         case PERF_RECORD_LOST:
1194                 color_fprintf(trace->output, PERF_COLOR_RED,
1195                               "LOST %" PRIu64 " events!\n", event->lost.lost);
1196                 ret = machine__process_lost_event(machine, event, sample);
1197                 break;
1198         default:
1199                 ret = machine__process_event(machine, event, sample);
1200                 break;
1201         }
1202
1203         return ret;
1204 }
1205
1206 static int trace__tool_process(struct perf_tool *tool,
1207                                union perf_event *event,
1208                                struct perf_sample *sample,
1209                                struct machine *machine)
1210 {
1211         struct trace *trace = container_of(tool, struct trace, tool);
1212         return trace__process_event(trace, machine, event, sample);
1213 }
1214
1215 static char *trace__machine__resolve_kernel_addr(void *vmachine, unsigned long long *addrp, char **modp)
1216 {
1217         struct machine *machine = vmachine;
1218
1219         if (machine->kptr_restrict_warned)
1220                 return NULL;
1221
1222         if (symbol_conf.kptr_restrict) {
1223                 pr_warning("Kernel address maps (/proc/{kallsyms,modules}) are restricted.\n\n"
1224                            "Check /proc/sys/kernel/kptr_restrict.\n\n"
1225                            "Kernel samples will not be resolved.\n");
1226                 machine->kptr_restrict_warned = true;
1227                 return NULL;
1228         }
1229
1230         return machine__resolve_kernel_addr(vmachine, addrp, modp);
1231 }
1232
1233 static int trace__symbols_init(struct trace *trace, struct perf_evlist *evlist)
1234 {
1235         int err = symbol__init(NULL);
1236
1237         if (err)
1238                 return err;
1239
1240         trace->host = machine__new_host();
1241         if (trace->host == NULL)
1242                 return -ENOMEM;
1243
1244         err = trace_event__register_resolver(trace->host, trace__machine__resolve_kernel_addr);
1245         if (err < 0)
1246                 goto out;
1247
1248         err = __machine__synthesize_threads(trace->host, &trace->tool, &trace->opts.target,
1249                                             evlist->threads, trace__tool_process, false,
1250                                             trace->opts.proc_map_timeout, 1);
1251 out:
1252         if (err)
1253                 symbol__exit();
1254
1255         return err;
1256 }
1257
1258 static void trace__symbols__exit(struct trace *trace)
1259 {
1260         machine__exit(trace->host);
1261         trace->host = NULL;
1262
1263         symbol__exit();
1264 }
1265
1266 static int syscall__alloc_arg_fmts(struct syscall *sc, int nr_args)
1267 {
1268         int idx;
1269
1270         if (nr_args == 6 && sc->fmt && sc->fmt->nr_args != 0)
1271                 nr_args = sc->fmt->nr_args;
1272
1273         sc->arg_fmt = calloc(nr_args, sizeof(*sc->arg_fmt));
1274         if (sc->arg_fmt == NULL)
1275                 return -1;
1276
1277         for (idx = 0; idx < nr_args; ++idx) {
1278                 if (sc->fmt)
1279                         sc->arg_fmt[idx] = sc->fmt->arg[idx];
1280         }
1281
1282         sc->nr_args = nr_args;
1283         return 0;
1284 }
1285
1286 static int syscall__set_arg_fmts(struct syscall *sc)
1287 {
1288         struct tep_format_field *field, *last_field = NULL;
1289         int idx = 0, len;
1290
1291         for (field = sc->args; field; field = field->next, ++idx) {
1292                 last_field = field;
1293
1294                 if (sc->fmt && sc->fmt->arg[idx].scnprintf)
1295                         continue;
1296
1297                 if (strcmp(field->type, "const char *") == 0 &&
1298                          (strcmp(field->name, "filename") == 0 ||
1299                           strcmp(field->name, "path") == 0 ||
1300                           strcmp(field->name, "pathname") == 0))
1301                         sc->arg_fmt[idx].scnprintf = SCA_FILENAME;
1302                 else if (field->flags & TEP_FIELD_IS_POINTER)
1303                         sc->arg_fmt[idx].scnprintf = syscall_arg__scnprintf_hex;
1304                 else if (strcmp(field->type, "pid_t") == 0)
1305                         sc->arg_fmt[idx].scnprintf = SCA_PID;
1306                 else if (strcmp(field->type, "umode_t") == 0)
1307                         sc->arg_fmt[idx].scnprintf = SCA_MODE_T;
1308                 else if ((strcmp(field->type, "int") == 0 ||
1309                           strcmp(field->type, "unsigned int") == 0 ||
1310                           strcmp(field->type, "long") == 0) &&
1311                          (len = strlen(field->name)) >= 2 &&
1312                          strcmp(field->name + len - 2, "fd") == 0) {
1313                         /*
1314                          * /sys/kernel/tracing/events/syscalls/sys_enter*
1315                          * egrep 'field:.*fd;' .../format|sed -r 's/.*field:([a-z ]+) [a-z_]*fd.+/\1/g'|sort|uniq -c
1316                          * 65 int
1317                          * 23 unsigned int
1318                          * 7 unsigned long
1319                          */
1320                         sc->arg_fmt[idx].scnprintf = SCA_FD;
1321                 }
1322         }
1323
1324         if (last_field)
1325                 sc->args_size = last_field->offset + last_field->size;
1326
1327         return 0;
1328 }
1329
1330 static int trace__read_syscall_info(struct trace *trace, int id)
1331 {
1332         char tp_name[128];
1333         struct syscall *sc;
1334         const char *name = syscalltbl__name(trace->sctbl, id);
1335
1336         if (name == NULL)
1337                 return -1;
1338
1339         if (id > trace->syscalls.max) {
1340                 struct syscall *nsyscalls = realloc(trace->syscalls.table, (id + 1) * sizeof(*sc));
1341
1342                 if (nsyscalls == NULL)
1343                         return -1;
1344
1345                 if (trace->syscalls.max != -1) {
1346                         memset(nsyscalls + trace->syscalls.max + 1, 0,
1347                                (id - trace->syscalls.max) * sizeof(*sc));
1348                 } else {
1349                         memset(nsyscalls, 0, (id + 1) * sizeof(*sc));
1350                 }
1351
1352                 trace->syscalls.table = nsyscalls;
1353                 trace->syscalls.max   = id;
1354         }
1355
1356         sc = trace->syscalls.table + id;
1357         sc->name = name;
1358
1359         sc->fmt  = syscall_fmt__find(sc->name);
1360
1361         snprintf(tp_name, sizeof(tp_name), "sys_enter_%s", sc->name);
1362         sc->tp_format = trace_event__tp_format("syscalls", tp_name);
1363
1364         if (IS_ERR(sc->tp_format) && sc->fmt && sc->fmt->alias) {
1365                 snprintf(tp_name, sizeof(tp_name), "sys_enter_%s", sc->fmt->alias);
1366                 sc->tp_format = trace_event__tp_format("syscalls", tp_name);
1367         }
1368
1369         if (syscall__alloc_arg_fmts(sc, IS_ERR(sc->tp_format) ? 6 : sc->tp_format->format.nr_fields))
1370                 return -1;
1371
1372         if (IS_ERR(sc->tp_format))
1373                 return -1;
1374
1375         sc->args = sc->tp_format->format.fields;
1376         /*
1377          * We need to check and discard the first variable '__syscall_nr'
1378          * or 'nr' that mean the syscall number. It is needless here.
1379          * So drop '__syscall_nr' or 'nr' field but does not exist on older kernels.
1380          */
1381         if (sc->args && (!strcmp(sc->args->name, "__syscall_nr") || !strcmp(sc->args->name, "nr"))) {
1382                 sc->args = sc->args->next;
1383                 --sc->nr_args;
1384         }
1385
1386         sc->is_exit = !strcmp(name, "exit_group") || !strcmp(name, "exit");
1387         sc->is_open = !strcmp(name, "open") || !strcmp(name, "openat");
1388
1389         return syscall__set_arg_fmts(sc);
1390 }
1391
1392 static int trace__validate_ev_qualifier(struct trace *trace)
1393 {
1394         int err = 0, i;
1395         size_t nr_allocated;
1396         struct str_node *pos;
1397
1398         trace->ev_qualifier_ids.nr = strlist__nr_entries(trace->ev_qualifier);
1399         trace->ev_qualifier_ids.entries = malloc(trace->ev_qualifier_ids.nr *
1400                                                  sizeof(trace->ev_qualifier_ids.entries[0]));
1401
1402         if (trace->ev_qualifier_ids.entries == NULL) {
1403                 fputs("Error:\tNot enough memory for allocating events qualifier ids\n",
1404                        trace->output);
1405                 err = -EINVAL;
1406                 goto out;
1407         }
1408
1409         nr_allocated = trace->ev_qualifier_ids.nr;
1410         i = 0;
1411
1412         strlist__for_each_entry(pos, trace->ev_qualifier) {
1413                 const char *sc = pos->s;
1414                 int id = syscalltbl__id(trace->sctbl, sc), match_next = -1;
1415
1416                 if (id < 0) {
1417                         id = syscalltbl__strglobmatch_first(trace->sctbl, sc, &match_next);
1418                         if (id >= 0)
1419                                 goto matches;
1420
1421                         if (err == 0) {
1422                                 fputs("Error:\tInvalid syscall ", trace->output);
1423                                 err = -EINVAL;
1424                         } else {
1425                                 fputs(", ", trace->output);
1426                         }
1427
1428                         fputs(sc, trace->output);
1429                 }
1430 matches:
1431                 trace->ev_qualifier_ids.entries[i++] = id;
1432                 if (match_next == -1)
1433                         continue;
1434
1435                 while (1) {
1436                         id = syscalltbl__strglobmatch_next(trace->sctbl, sc, &match_next);
1437                         if (id < 0)
1438                                 break;
1439                         if (nr_allocated == trace->ev_qualifier_ids.nr) {
1440                                 void *entries;
1441
1442                                 nr_allocated += 8;
1443                                 entries = realloc(trace->ev_qualifier_ids.entries,
1444                                                   nr_allocated * sizeof(trace->ev_qualifier_ids.entries[0]));
1445                                 if (entries == NULL) {
1446                                         err = -ENOMEM;
1447                                         fputs("\nError:\t Not enough memory for parsing\n", trace->output);
1448                                         goto out_free;
1449                                 }
1450                                 trace->ev_qualifier_ids.entries = entries;
1451                         }
1452                         trace->ev_qualifier_ids.nr++;
1453                         trace->ev_qualifier_ids.entries[i++] = id;
1454                 }
1455         }
1456
1457         if (err < 0) {
1458                 fputs("\nHint:\ttry 'perf list syscalls:sys_enter_*'"
1459                       "\nHint:\tand: 'man syscalls'\n", trace->output);
1460 out_free:
1461                 zfree(&trace->ev_qualifier_ids.entries);
1462                 trace->ev_qualifier_ids.nr = 0;
1463         }
1464 out:
1465         return err;
1466 }
1467
1468 /*
1469  * args is to be interpreted as a series of longs but we need to handle
1470  * 8-byte unaligned accesses. args points to raw_data within the event
1471  * and raw_data is guaranteed to be 8-byte unaligned because it is
1472  * preceded by raw_size which is a u32. So we need to copy args to a temp
1473  * variable to read it. Most notably this avoids extended load instructions
1474  * on unaligned addresses
1475  */
1476 unsigned long syscall_arg__val(struct syscall_arg *arg, u8 idx)
1477 {
1478         unsigned long val;
1479         unsigned char *p = arg->args + sizeof(unsigned long) * idx;
1480
1481         memcpy(&val, p, sizeof(val));
1482         return val;
1483 }
1484
1485 static size_t syscall__scnprintf_name(struct syscall *sc, char *bf, size_t size,
1486                                       struct syscall_arg *arg)
1487 {
1488         if (sc->arg_fmt && sc->arg_fmt[arg->idx].name)
1489                 return scnprintf(bf, size, "%s: ", sc->arg_fmt[arg->idx].name);
1490
1491         return scnprintf(bf, size, "arg%d: ", arg->idx);
1492 }
1493
1494 /*
1495  * Check if the value is in fact zero, i.e. mask whatever needs masking, such
1496  * as mount 'flags' argument that needs ignoring some magic flag, see comment
1497  * in tools/perf/trace/beauty/mount_flags.c
1498  */
1499 static unsigned long syscall__mask_val(struct syscall *sc, struct syscall_arg *arg, unsigned long val)
1500 {
1501         if (sc->arg_fmt && sc->arg_fmt[arg->idx].mask_val)
1502                 return sc->arg_fmt[arg->idx].mask_val(arg, val);
1503
1504         return val;
1505 }
1506
1507 static size_t syscall__scnprintf_val(struct syscall *sc, char *bf, size_t size,
1508                                      struct syscall_arg *arg, unsigned long val)
1509 {
1510         if (sc->arg_fmt && sc->arg_fmt[arg->idx].scnprintf) {
1511                 arg->val = val;
1512                 if (sc->arg_fmt[arg->idx].parm)
1513                         arg->parm = sc->arg_fmt[arg->idx].parm;
1514                 return sc->arg_fmt[arg->idx].scnprintf(bf, size, arg);
1515         }
1516         return scnprintf(bf, size, "%ld", val);
1517 }
1518
1519 static size_t syscall__scnprintf_args(struct syscall *sc, char *bf, size_t size,
1520                                       unsigned char *args, void *augmented_args, int augmented_args_size,
1521                                       struct trace *trace, struct thread *thread)
1522 {
1523         size_t printed = 0;
1524         unsigned long val;
1525         u8 bit = 1;
1526         struct syscall_arg arg = {
1527                 .args   = args,
1528                 .augmented = {
1529                         .size = augmented_args_size,
1530                         .args = augmented_args,
1531                 },
1532                 .idx    = 0,
1533                 .mask   = 0,
1534                 .trace  = trace,
1535                 .thread = thread,
1536         };
1537         struct thread_trace *ttrace = thread__priv(thread);
1538
1539         /*
1540          * Things like fcntl will set this in its 'cmd' formatter to pick the
1541          * right formatter for the return value (an fd? file flags?), which is
1542          * not needed for syscalls that always return a given type, say an fd.
1543          */
1544         ttrace->ret_scnprintf = NULL;
1545
1546         if (sc->args != NULL) {
1547                 struct tep_format_field *field;
1548
1549                 for (field = sc->args; field;
1550                      field = field->next, ++arg.idx, bit <<= 1) {
1551                         if (arg.mask & bit)
1552                                 continue;
1553
1554                         val = syscall_arg__val(&arg, arg.idx);
1555                         /*
1556                          * Some syscall args need some mask, most don't and
1557                          * return val untouched.
1558                          */
1559                         val = syscall__mask_val(sc, &arg, val);
1560
1561                         /*
1562                          * Suppress this argument if its value is zero and
1563                          * and we don't have a string associated in an
1564                          * strarray for it.
1565                          */
1566                         if (val == 0 &&
1567                             !(sc->arg_fmt &&
1568                               (sc->arg_fmt[arg.idx].show_zero ||
1569                                sc->arg_fmt[arg.idx].scnprintf == SCA_STRARRAY ||
1570                                sc->arg_fmt[arg.idx].scnprintf == SCA_STRARRAYS) &&
1571                               sc->arg_fmt[arg.idx].parm))
1572                                 continue;
1573
1574                         printed += scnprintf(bf + printed, size - printed,
1575                                              "%s%s: ", printed ? ", " : "", field->name);
1576                         printed += syscall__scnprintf_val(sc, bf + printed, size - printed, &arg, val);
1577                 }
1578         } else if (IS_ERR(sc->tp_format)) {
1579                 /*
1580                  * If we managed to read the tracepoint /format file, then we
1581                  * may end up not having any args, like with gettid(), so only
1582                  * print the raw args when we didn't manage to read it.
1583                  */
1584                 while (arg.idx < sc->nr_args) {
1585                         if (arg.mask & bit)
1586                                 goto next_arg;
1587                         val = syscall_arg__val(&arg, arg.idx);
1588                         if (printed)
1589                                 printed += scnprintf(bf + printed, size - printed, ", ");
1590                         printed += syscall__scnprintf_name(sc, bf + printed, size - printed, &arg);
1591                         printed += syscall__scnprintf_val(sc, bf + printed, size - printed, &arg, val);
1592 next_arg:
1593                         ++arg.idx;
1594                         bit <<= 1;
1595                 }
1596         }
1597
1598         return printed;
1599 }
1600
1601 typedef int (*tracepoint_handler)(struct trace *trace, struct perf_evsel *evsel,
1602                                   union perf_event *event,
1603                                   struct perf_sample *sample);
1604
1605 static struct syscall *trace__syscall_info(struct trace *trace,
1606                                            struct perf_evsel *evsel, int id)
1607 {
1608
1609         if (id < 0) {
1610
1611                 /*
1612                  * XXX: Noticed on x86_64, reproduced as far back as 3.0.36, haven't tried
1613                  * before that, leaving at a higher verbosity level till that is
1614                  * explained. Reproduced with plain ftrace with:
1615                  *
1616                  * echo 1 > /t/events/raw_syscalls/sys_exit/enable
1617                  * grep "NR -1 " /t/trace_pipe
1618                  *
1619                  * After generating some load on the machine.
1620                  */
1621                 if (verbose > 1) {
1622                         static u64 n;
1623                         fprintf(trace->output, "Invalid syscall %d id, skipping (%s, %" PRIu64 ") ...\n",
1624                                 id, perf_evsel__name(evsel), ++n);
1625                 }
1626                 return NULL;
1627         }
1628
1629         if ((id > trace->syscalls.max || trace->syscalls.table[id].name == NULL) &&
1630             trace__read_syscall_info(trace, id))
1631                 goto out_cant_read;
1632
1633         if ((id > trace->syscalls.max || trace->syscalls.table[id].name == NULL))
1634                 goto out_cant_read;
1635
1636         return &trace->syscalls.table[id];
1637
1638 out_cant_read:
1639         if (verbose > 0) {
1640                 fprintf(trace->output, "Problems reading syscall %d", id);
1641                 if (id <= trace->syscalls.max && trace->syscalls.table[id].name != NULL)
1642                         fprintf(trace->output, "(%s)", trace->syscalls.table[id].name);
1643                 fputs(" information\n", trace->output);
1644         }
1645         return NULL;
1646 }
1647
1648 static void thread__update_stats(struct thread_trace *ttrace,
1649                                  int id, struct perf_sample *sample)
1650 {
1651         struct int_node *inode;
1652         struct stats *stats;
1653         u64 duration = 0;
1654
1655         inode = intlist__findnew(ttrace->syscall_stats, id);
1656         if (inode == NULL)
1657                 return;
1658
1659         stats = inode->priv;
1660         if (stats == NULL) {
1661                 stats = malloc(sizeof(struct stats));
1662                 if (stats == NULL)
1663                         return;
1664                 init_stats(stats);
1665                 inode->priv = stats;
1666         }
1667
1668         if (ttrace->entry_time && sample->time > ttrace->entry_time)
1669                 duration = sample->time - ttrace->entry_time;
1670
1671         update_stats(stats, duration);
1672 }
1673
1674 static int trace__printf_interrupted_entry(struct trace *trace)
1675 {
1676         struct thread_trace *ttrace;
1677         size_t printed;
1678
1679         if (trace->failure_only || trace->current == NULL)
1680                 return 0;
1681
1682         ttrace = thread__priv(trace->current);
1683
1684         if (!ttrace->entry_pending)
1685                 return 0;
1686
1687         printed  = trace__fprintf_entry_head(trace, trace->current, 0, false, ttrace->entry_time, trace->output);
1688         printed += fprintf(trace->output, "%-70s) ...\n", ttrace->entry_str);
1689         ttrace->entry_pending = false;
1690
1691         ++trace->nr_events_printed;
1692
1693         return printed;
1694 }
1695
1696 static int trace__fprintf_sample(struct trace *trace, struct perf_evsel *evsel,
1697                                  struct perf_sample *sample, struct thread *thread)
1698 {
1699         int printed = 0;
1700
1701         if (trace->print_sample) {
1702                 double ts = (double)sample->time / NSEC_PER_MSEC;
1703
1704                 printed += fprintf(trace->output, "%22s %10.3f %s %d/%d [%d]\n",
1705                                    perf_evsel__name(evsel), ts,
1706                                    thread__comm_str(thread),
1707                                    sample->pid, sample->tid, sample->cpu);
1708         }
1709
1710         return printed;
1711 }
1712
1713 static void *syscall__augmented_args(struct syscall *sc, struct perf_sample *sample, int *augmented_args_size)
1714 {
1715         void *augmented_args = NULL;
1716
1717         *augmented_args_size = sample->raw_size - sc->args_size;
1718         if (*augmented_args_size > 0)
1719                 augmented_args = sample->raw_data + sc->args_size;
1720
1721         return augmented_args;
1722 }
1723
1724 static int trace__sys_enter(struct trace *trace, struct perf_evsel *evsel,
1725                             union perf_event *event __maybe_unused,
1726                             struct perf_sample *sample)
1727 {
1728         char *msg;
1729         void *args;
1730         size_t printed = 0;
1731         struct thread *thread;
1732         int id = perf_evsel__sc_tp_uint(evsel, id, sample), err = -1;
1733         int augmented_args_size = 0;
1734         void *augmented_args = NULL;
1735         struct syscall *sc = trace__syscall_info(trace, evsel, id);
1736         struct thread_trace *ttrace;
1737
1738         if (sc == NULL)
1739                 return -1;
1740
1741         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
1742         ttrace = thread__trace(thread, trace->output);
1743         if (ttrace == NULL)
1744                 goto out_put;
1745
1746         trace__fprintf_sample(trace, evsel, sample, thread);
1747
1748         args = perf_evsel__sc_tp_ptr(evsel, args, sample);
1749
1750         if (ttrace->entry_str == NULL) {
1751                 ttrace->entry_str = malloc(trace__entry_str_size);
1752                 if (!ttrace->entry_str)
1753                         goto out_put;
1754         }
1755
1756         if (!(trace->duration_filter || trace->summary_only || trace->min_stack))
1757                 trace__printf_interrupted_entry(trace);
1758         /*
1759          * If this is raw_syscalls.sys_enter, then it always comes with the 6 possible
1760          * arguments, even if the syscall being handled, say "openat", uses only 4 arguments
1761          * this breaks syscall__augmented_args() check for augmented args, as we calculate
1762          * syscall->args_size using each syscalls:sys_enter_NAME tracefs format file,
1763          * so when handling, say the openat syscall, we end up getting 6 args for the
1764          * raw_syscalls:sys_enter event, when we expected just 4, we end up mistakenly
1765          * thinking that the extra 2 u64 args are the augmented filename, so just check
1766          * here and avoid using augmented syscalls when the evsel is the raw_syscalls one.
1767          */
1768         if (evsel != trace->syscalls.events.sys_enter)
1769                 augmented_args = syscall__augmented_args(sc, sample, &augmented_args_size);
1770         ttrace->entry_time = sample->time;
1771         msg = ttrace->entry_str;
1772         printed += scnprintf(msg + printed, trace__entry_str_size - printed, "%s(", sc->name);
1773
1774         printed += syscall__scnprintf_args(sc, msg + printed, trace__entry_str_size - printed,
1775                                            args, augmented_args, augmented_args_size, trace, thread);
1776
1777         if (sc->is_exit) {
1778                 if (!(trace->duration_filter || trace->summary_only || trace->failure_only || trace->min_stack)) {
1779                         trace__fprintf_entry_head(trace, thread, 0, false, ttrace->entry_time, trace->output);
1780                         fprintf(trace->output, "%-70s)\n", ttrace->entry_str);
1781                 }
1782         } else {
1783                 ttrace->entry_pending = true;
1784                 /* See trace__vfs_getname & trace__sys_exit */
1785                 ttrace->filename.pending_open = false;
1786         }
1787
1788         if (trace->current != thread) {
1789                 thread__put(trace->current);
1790                 trace->current = thread__get(thread);
1791         }
1792         err = 0;
1793 out_put:
1794         thread__put(thread);
1795         return err;
1796 }
1797
1798 static int trace__fprintf_sys_enter(struct trace *trace, struct perf_evsel *evsel,
1799                                     struct perf_sample *sample)
1800 {
1801         struct thread_trace *ttrace;
1802         struct thread *thread;
1803         int id = perf_evsel__sc_tp_uint(evsel, id, sample), err = -1;
1804         struct syscall *sc = trace__syscall_info(trace, evsel, id);
1805         char msg[1024];
1806         void *args, *augmented_args = NULL;
1807         int augmented_args_size;
1808
1809         if (sc == NULL)
1810                 return -1;
1811
1812         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
1813         ttrace = thread__trace(thread, trace->output);
1814         /*
1815          * We need to get ttrace just to make sure it is there when syscall__scnprintf_args()
1816          * and the rest of the beautifiers accessing it via struct syscall_arg touches it.
1817          */
1818         if (ttrace == NULL)
1819                 goto out_put;
1820
1821         args = perf_evsel__sc_tp_ptr(evsel, args, sample);
1822         augmented_args = syscall__augmented_args(sc, sample, &augmented_args_size);
1823         syscall__scnprintf_args(sc, msg, sizeof(msg), args, augmented_args, augmented_args_size, trace, thread);
1824         fprintf(trace->output, "%s", msg);
1825         err = 0;
1826 out_put:
1827         thread__put(thread);
1828         return err;
1829 }
1830
1831 static int trace__resolve_callchain(struct trace *trace, struct perf_evsel *evsel,
1832                                     struct perf_sample *sample,
1833                                     struct callchain_cursor *cursor)
1834 {
1835         struct addr_location al;
1836         int max_stack = evsel->attr.sample_max_stack ?
1837                         evsel->attr.sample_max_stack :
1838                         trace->max_stack;
1839         int err;
1840
1841         if (machine__resolve(trace->host, &al, sample) < 0)
1842                 return -1;
1843
1844         err = thread__resolve_callchain(al.thread, cursor, evsel, sample, NULL, NULL, max_stack);
1845         addr_location__put(&al);
1846         return err;
1847 }
1848
1849 static int trace__fprintf_callchain(struct trace *trace, struct perf_sample *sample)
1850 {
1851         /* TODO: user-configurable print_opts */
1852         const unsigned int print_opts = EVSEL__PRINT_SYM |
1853                                         EVSEL__PRINT_DSO |
1854                                         EVSEL__PRINT_UNKNOWN_AS_ADDR;
1855
1856         return sample__fprintf_callchain(sample, 38, print_opts, &callchain_cursor, trace->output);
1857 }
1858
1859 static const char *errno_to_name(struct perf_evsel *evsel, int err)
1860 {
1861         struct perf_env *env = perf_evsel__env(evsel);
1862         const char *arch_name = perf_env__arch(env);
1863
1864         return arch_syscalls__strerrno(arch_name, err);
1865 }
1866
1867 static int trace__sys_exit(struct trace *trace, struct perf_evsel *evsel,
1868                            union perf_event *event __maybe_unused,
1869                            struct perf_sample *sample)
1870 {
1871         long ret;
1872         u64 duration = 0;
1873         bool duration_calculated = false;
1874         struct thread *thread;
1875         int id = perf_evsel__sc_tp_uint(evsel, id, sample), err = -1, callchain_ret = 0;
1876         struct syscall *sc = trace__syscall_info(trace, evsel, id);
1877         struct thread_trace *ttrace;
1878
1879         if (sc == NULL)
1880                 return -1;
1881
1882         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
1883         ttrace = thread__trace(thread, trace->output);
1884         if (ttrace == NULL)
1885                 goto out_put;
1886
1887         trace__fprintf_sample(trace, evsel, sample, thread);
1888
1889         if (trace->summary)
1890                 thread__update_stats(ttrace, id, sample);
1891
1892         ret = perf_evsel__sc_tp_uint(evsel, ret, sample);
1893
1894         if (sc->is_open && ret >= 0 && ttrace->filename.pending_open) {
1895                 trace__set_fd_pathname(thread, ret, ttrace->filename.name);
1896                 ttrace->filename.pending_open = false;
1897                 ++trace->stats.vfs_getname;
1898         }
1899
1900         if (ttrace->entry_time) {
1901                 duration = sample->time - ttrace->entry_time;
1902                 if (trace__filter_duration(trace, duration))
1903                         goto out;
1904                 duration_calculated = true;
1905         } else if (trace->duration_filter)
1906                 goto out;
1907
1908         if (sample->callchain) {
1909                 callchain_ret = trace__resolve_callchain(trace, evsel, sample, &callchain_cursor);
1910                 if (callchain_ret == 0) {
1911                         if (callchain_cursor.nr < trace->min_stack)
1912                                 goto out;
1913                         callchain_ret = 1;
1914                 }
1915         }
1916
1917         if (trace->summary_only || (ret >= 0 && trace->failure_only))
1918                 goto out;
1919
1920         trace__fprintf_entry_head(trace, thread, duration, duration_calculated, ttrace->entry_time, trace->output);
1921
1922         if (ttrace->entry_pending) {
1923                 fprintf(trace->output, "%-70s", ttrace->entry_str);
1924         } else {
1925                 fprintf(trace->output, " ... [");
1926                 color_fprintf(trace->output, PERF_COLOR_YELLOW, "continued");
1927                 fprintf(trace->output, "]: %s()", sc->name);
1928         }
1929
1930         if (sc->fmt == NULL) {
1931                 if (ret < 0)
1932                         goto errno_print;
1933 signed_print:
1934                 fprintf(trace->output, ") = %ld", ret);
1935         } else if (ret < 0) {
1936 errno_print: {
1937                 char bf[STRERR_BUFSIZE];
1938                 const char *emsg = str_error_r(-ret, bf, sizeof(bf)),
1939                            *e = errno_to_name(evsel, -ret);
1940
1941                 fprintf(trace->output, ") = -1 %s %s", e, emsg);
1942         }
1943         } else if (ret == 0 && sc->fmt->timeout)
1944                 fprintf(trace->output, ") = 0 Timeout");
1945         else if (ttrace->ret_scnprintf) {
1946                 char bf[1024];
1947                 struct syscall_arg arg = {
1948                         .val    = ret,
1949                         .thread = thread,
1950                         .trace  = trace,
1951                 };
1952                 ttrace->ret_scnprintf(bf, sizeof(bf), &arg);
1953                 ttrace->ret_scnprintf = NULL;
1954                 fprintf(trace->output, ") = %s", bf);
1955         } else if (sc->fmt->hexret)
1956                 fprintf(trace->output, ") = %#lx", ret);
1957         else if (sc->fmt->errpid) {
1958                 struct thread *child = machine__find_thread(trace->host, ret, ret);
1959
1960                 if (child != NULL) {
1961                         fprintf(trace->output, ") = %ld", ret);
1962                         if (child->comm_set)
1963                                 fprintf(trace->output, " (%s)", thread__comm_str(child));
1964                         thread__put(child);
1965                 }
1966         } else
1967                 goto signed_print;
1968
1969         fputc('\n', trace->output);
1970
1971         /*
1972          * We only consider an 'event' for the sake of --max-events a non-filtered
1973          * sys_enter + sys_exit and other tracepoint events.
1974          */
1975         if (++trace->nr_events_printed == trace->max_events && trace->max_events != ULONG_MAX)
1976                 interrupted = true;
1977
1978         if (callchain_ret > 0)
1979                 trace__fprintf_callchain(trace, sample);
1980         else if (callchain_ret < 0)
1981                 pr_err("Problem processing %s callchain, skipping...\n", perf_evsel__name(evsel));
1982 out:
1983         ttrace->entry_pending = false;
1984         err = 0;
1985 out_put:
1986         thread__put(thread);
1987         return err;
1988 }
1989
1990 static int trace__vfs_getname(struct trace *trace, struct perf_evsel *evsel,
1991                               union perf_event *event __maybe_unused,
1992                               struct perf_sample *sample)
1993 {
1994         struct thread *thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
1995         struct thread_trace *ttrace;
1996         size_t filename_len, entry_str_len, to_move;
1997         ssize_t remaining_space;
1998         char *pos;
1999         const char *filename = perf_evsel__rawptr(evsel, sample, "pathname");
2000
2001         if (!thread)
2002                 goto out;
2003
2004         ttrace = thread__priv(thread);
2005         if (!ttrace)
2006                 goto out_put;
2007
2008         filename_len = strlen(filename);
2009         if (filename_len == 0)
2010                 goto out_put;
2011
2012         if (ttrace->filename.namelen < filename_len) {
2013                 char *f = realloc(ttrace->filename.name, filename_len + 1);
2014
2015                 if (f == NULL)
2016                         goto out_put;
2017
2018                 ttrace->filename.namelen = filename_len;
2019                 ttrace->filename.name = f;
2020         }
2021
2022         strcpy(ttrace->filename.name, filename);
2023         ttrace->filename.pending_open = true;
2024
2025         if (!ttrace->filename.ptr)
2026                 goto out_put;
2027
2028         entry_str_len = strlen(ttrace->entry_str);
2029         remaining_space = trace__entry_str_size - entry_str_len - 1; /* \0 */
2030         if (remaining_space <= 0)
2031                 goto out_put;
2032
2033         if (filename_len > (size_t)remaining_space) {
2034                 filename += filename_len - remaining_space;
2035                 filename_len = remaining_space;
2036         }
2037
2038         to_move = entry_str_len - ttrace->filename.entry_str_pos + 1; /* \0 */
2039         pos = ttrace->entry_str + ttrace->filename.entry_str_pos;
2040         memmove(pos + filename_len, pos, to_move);
2041         memcpy(pos, filename, filename_len);
2042
2043         ttrace->filename.ptr = 0;
2044         ttrace->filename.entry_str_pos = 0;
2045 out_put:
2046         thread__put(thread);
2047 out:
2048         return 0;
2049 }
2050
2051 static int trace__sched_stat_runtime(struct trace *trace, struct perf_evsel *evsel,
2052                                      union perf_event *event __maybe_unused,
2053                                      struct perf_sample *sample)
2054 {
2055         u64 runtime = perf_evsel__intval(evsel, sample, "runtime");
2056         double runtime_ms = (double)runtime / NSEC_PER_MSEC;
2057         struct thread *thread = machine__findnew_thread(trace->host,
2058                                                         sample->pid,
2059                                                         sample->tid);
2060         struct thread_trace *ttrace = thread__trace(thread, trace->output);
2061
2062         if (ttrace == NULL)
2063                 goto out_dump;
2064
2065         ttrace->runtime_ms += runtime_ms;
2066         trace->runtime_ms += runtime_ms;
2067 out_put:
2068         thread__put(thread);
2069         return 0;
2070
2071 out_dump:
2072         fprintf(trace->output, "%s: comm=%s,pid=%u,runtime=%" PRIu64 ",vruntime=%" PRIu64 ")\n",
2073                evsel->name,
2074                perf_evsel__strval(evsel, sample, "comm"),
2075                (pid_t)perf_evsel__intval(evsel, sample, "pid"),
2076                runtime,
2077                perf_evsel__intval(evsel, sample, "vruntime"));
2078         goto out_put;
2079 }
2080
2081 static int bpf_output__printer(enum binary_printer_ops op,
2082                                unsigned int val, void *extra __maybe_unused, FILE *fp)
2083 {
2084         unsigned char ch = (unsigned char)val;
2085
2086         switch (op) {
2087         case BINARY_PRINT_CHAR_DATA:
2088                 return fprintf(fp, "%c", isprint(ch) ? ch : '.');
2089         case BINARY_PRINT_DATA_BEGIN:
2090         case BINARY_PRINT_LINE_BEGIN:
2091         case BINARY_PRINT_ADDR:
2092         case BINARY_PRINT_NUM_DATA:
2093         case BINARY_PRINT_NUM_PAD:
2094         case BINARY_PRINT_SEP:
2095         case BINARY_PRINT_CHAR_PAD:
2096         case BINARY_PRINT_LINE_END:
2097         case BINARY_PRINT_DATA_END:
2098         default:
2099                 break;
2100         }
2101
2102         return 0;
2103 }
2104
2105 static void bpf_output__fprintf(struct trace *trace,
2106                                 struct perf_sample *sample)
2107 {
2108         binary__fprintf(sample->raw_data, sample->raw_size, 8,
2109                         bpf_output__printer, NULL, trace->output);
2110         ++trace->nr_events_printed;
2111 }
2112
2113 static int trace__event_handler(struct trace *trace, struct perf_evsel *evsel,
2114                                 union perf_event *event __maybe_unused,
2115                                 struct perf_sample *sample)
2116 {
2117         struct thread *thread;
2118         int callchain_ret = 0;
2119         /*
2120          * Check if we called perf_evsel__disable(evsel) due to, for instance,
2121          * this event's max_events having been hit and this is an entry coming
2122          * from the ring buffer that we should discard, since the max events
2123          * have already been considered/printed.
2124          */
2125         if (evsel->disabled)
2126                 return 0;
2127
2128         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2129
2130         if (sample->callchain) {
2131                 callchain_ret = trace__resolve_callchain(trace, evsel, sample, &callchain_cursor);
2132                 if (callchain_ret == 0) {
2133                         if (callchain_cursor.nr < trace->min_stack)
2134                                 goto out;
2135                         callchain_ret = 1;
2136                 }
2137         }
2138
2139         trace__printf_interrupted_entry(trace);
2140         trace__fprintf_tstamp(trace, sample->time, trace->output);
2141
2142         if (trace->trace_syscalls)
2143                 fprintf(trace->output, "(         ): ");
2144
2145         if (thread)
2146                 trace__fprintf_comm_tid(trace, thread, trace->output);
2147
2148         if (evsel == trace->syscalls.events.augmented) {
2149                 int id = perf_evsel__sc_tp_uint(evsel, id, sample);
2150                 struct syscall *sc = trace__syscall_info(trace, evsel, id);
2151
2152                 if (sc) {
2153                         fprintf(trace->output, "%s(", sc->name);
2154                         trace__fprintf_sys_enter(trace, evsel, sample);
2155                         fputc(')', trace->output);
2156                         goto newline;
2157                 }
2158
2159                 /*
2160                  * XXX: Not having the associated syscall info or not finding/adding
2161                  *      the thread should never happen, but if it does...
2162                  *      fall thru and print it as a bpf_output event.
2163                  */
2164         }
2165
2166         fprintf(trace->output, "%s:", evsel->name);
2167
2168         if (perf_evsel__is_bpf_output(evsel)) {
2169                 bpf_output__fprintf(trace, sample);
2170         } else if (evsel->tp_format) {
2171                 if (strncmp(evsel->tp_format->name, "sys_enter_", 10) ||
2172                     trace__fprintf_sys_enter(trace, evsel, sample)) {
2173                         event_format__fprintf(evsel->tp_format, sample->cpu,
2174                                               sample->raw_data, sample->raw_size,
2175                                               trace->output);
2176                         ++trace->nr_events_printed;
2177
2178                         if (evsel->max_events != ULONG_MAX && ++evsel->nr_events_printed == evsel->max_events) {
2179                                 perf_evsel__disable(evsel);
2180                                 perf_evsel__close(evsel);
2181                         }
2182                 }
2183         }
2184
2185 newline:
2186         fprintf(trace->output, "\n");
2187
2188         if (callchain_ret > 0)
2189                 trace__fprintf_callchain(trace, sample);
2190         else if (callchain_ret < 0)
2191                 pr_err("Problem processing %s callchain, skipping...\n", perf_evsel__name(evsel));
2192 out:
2193         thread__put(thread);
2194         return 0;
2195 }
2196
2197 static void print_location(FILE *f, struct perf_sample *sample,
2198                            struct addr_location *al,
2199                            bool print_dso, bool print_sym)
2200 {
2201
2202         if ((verbose > 0 || print_dso) && al->map)
2203                 fprintf(f, "%s@", al->map->dso->long_name);
2204
2205         if ((verbose > 0 || print_sym) && al->sym)
2206                 fprintf(f, "%s+0x%" PRIx64, al->sym->name,
2207                         al->addr - al->sym->start);
2208         else if (al->map)
2209                 fprintf(f, "0x%" PRIx64, al->addr);
2210         else
2211                 fprintf(f, "0x%" PRIx64, sample->addr);
2212 }
2213
2214 static int trace__pgfault(struct trace *trace,
2215                           struct perf_evsel *evsel,
2216                           union perf_event *event __maybe_unused,
2217                           struct perf_sample *sample)
2218 {
2219         struct thread *thread;
2220         struct addr_location al;
2221         char map_type = 'd';
2222         struct thread_trace *ttrace;
2223         int err = -1;
2224         int callchain_ret = 0;
2225
2226         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2227
2228         if (sample->callchain) {
2229                 callchain_ret = trace__resolve_callchain(trace, evsel, sample, &callchain_cursor);
2230                 if (callchain_ret == 0) {
2231                         if (callchain_cursor.nr < trace->min_stack)
2232                                 goto out_put;
2233                         callchain_ret = 1;
2234                 }
2235         }
2236
2237         ttrace = thread__trace(thread, trace->output);
2238         if (ttrace == NULL)
2239                 goto out_put;
2240
2241         if (evsel->attr.config == PERF_COUNT_SW_PAGE_FAULTS_MAJ)
2242                 ttrace->pfmaj++;
2243         else
2244                 ttrace->pfmin++;
2245
2246         if (trace->summary_only)
2247                 goto out;
2248
2249         thread__find_symbol(thread, sample->cpumode, sample->ip, &al);
2250
2251         trace__fprintf_entry_head(trace, thread, 0, true, sample->time, trace->output);
2252
2253         fprintf(trace->output, "%sfault [",
2254                 evsel->attr.config == PERF_COUNT_SW_PAGE_FAULTS_MAJ ?
2255                 "maj" : "min");
2256
2257         print_location(trace->output, sample, &al, false, true);
2258
2259         fprintf(trace->output, "] => ");
2260
2261         thread__find_symbol(thread, sample->cpumode, sample->addr, &al);
2262
2263         if (!al.map) {
2264                 thread__find_symbol(thread, sample->cpumode, sample->addr, &al);
2265
2266                 if (al.map)
2267                         map_type = 'x';
2268                 else
2269                         map_type = '?';
2270         }
2271
2272         print_location(trace->output, sample, &al, true, false);
2273
2274         fprintf(trace->output, " (%c%c)\n", map_type, al.level);
2275
2276         if (callchain_ret > 0)
2277                 trace__fprintf_callchain(trace, sample);
2278         else if (callchain_ret < 0)
2279                 pr_err("Problem processing %s callchain, skipping...\n", perf_evsel__name(evsel));
2280
2281         ++trace->nr_events_printed;
2282 out:
2283         err = 0;
2284 out_put:
2285         thread__put(thread);
2286         return err;
2287 }
2288
2289 static void trace__set_base_time(struct trace *trace,
2290                                  struct perf_evsel *evsel,
2291                                  struct perf_sample *sample)
2292 {
2293         /*
2294          * BPF events were not setting PERF_SAMPLE_TIME, so be more robust
2295          * and don't use sample->time unconditionally, we may end up having
2296          * some other event in the future without PERF_SAMPLE_TIME for good
2297          * reason, i.e. we may not be interested in its timestamps, just in
2298          * it taking place, picking some piece of information when it
2299          * appears in our event stream (vfs_getname comes to mind).
2300          */
2301         if (trace->base_time == 0 && !trace->full_time &&
2302             (evsel->attr.sample_type & PERF_SAMPLE_TIME))
2303                 trace->base_time = sample->time;
2304 }
2305
2306 static int trace__process_sample(struct perf_tool *tool,
2307                                  union perf_event *event,
2308                                  struct perf_sample *sample,
2309                                  struct perf_evsel *evsel,
2310                                  struct machine *machine __maybe_unused)
2311 {
2312         struct trace *trace = container_of(tool, struct trace, tool);
2313         struct thread *thread;
2314         int err = 0;
2315
2316         tracepoint_handler handler = evsel->handler;
2317
2318         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2319         if (thread && thread__is_filtered(thread))
2320                 goto out;
2321
2322         trace__set_base_time(trace, evsel, sample);
2323
2324         if (handler) {
2325                 ++trace->nr_events;
2326                 handler(trace, evsel, event, sample);
2327         }
2328 out:
2329         thread__put(thread);
2330         return err;
2331 }
2332
2333 static int trace__record(struct trace *trace, int argc, const char **argv)
2334 {
2335         unsigned int rec_argc, i, j;
2336         const char **rec_argv;
2337         const char * const record_args[] = {
2338                 "record",
2339                 "-R",
2340                 "-m", "1024",
2341                 "-c", "1",
2342         };
2343
2344         const char * const sc_args[] = { "-e", };
2345         unsigned int sc_args_nr = ARRAY_SIZE(sc_args);
2346         const char * const majpf_args[] = { "-e", "major-faults" };
2347         unsigned int majpf_args_nr = ARRAY_SIZE(majpf_args);
2348         const char * const minpf_args[] = { "-e", "minor-faults" };
2349         unsigned int minpf_args_nr = ARRAY_SIZE(minpf_args);
2350
2351         /* +1 is for the event string below */
2352         rec_argc = ARRAY_SIZE(record_args) + sc_args_nr + 1 +
2353                 majpf_args_nr + minpf_args_nr + argc;
2354         rec_argv = calloc(rec_argc + 1, sizeof(char *));
2355
2356         if (rec_argv == NULL)
2357                 return -ENOMEM;
2358
2359         j = 0;
2360         for (i = 0; i < ARRAY_SIZE(record_args); i++)
2361                 rec_argv[j++] = record_args[i];
2362
2363         if (trace->trace_syscalls) {
2364                 for (i = 0; i < sc_args_nr; i++)
2365                         rec_argv[j++] = sc_args[i];
2366
2367                 /* event string may be different for older kernels - e.g., RHEL6 */
2368                 if (is_valid_tracepoint("raw_syscalls:sys_enter"))
2369                         rec_argv[j++] = "raw_syscalls:sys_enter,raw_syscalls:sys_exit";
2370                 else if (is_valid_tracepoint("syscalls:sys_enter"))
2371                         rec_argv[j++] = "syscalls:sys_enter,syscalls:sys_exit";
2372                 else {
2373                         pr_err("Neither raw_syscalls nor syscalls events exist.\n");
2374                         free(rec_argv);
2375                         return -1;
2376                 }
2377         }
2378
2379         if (trace->trace_pgfaults & TRACE_PFMAJ)
2380                 for (i = 0; i < majpf_args_nr; i++)
2381                         rec_argv[j++] = majpf_args[i];
2382
2383         if (trace->trace_pgfaults & TRACE_PFMIN)
2384                 for (i = 0; i < minpf_args_nr; i++)
2385                         rec_argv[j++] = minpf_args[i];
2386
2387         for (i = 0; i < (unsigned int)argc; i++)
2388                 rec_argv[j++] = argv[i];
2389
2390         return cmd_record(j, rec_argv);
2391 }
2392
2393 static size_t trace__fprintf_thread_summary(struct trace *trace, FILE *fp);
2394
2395 static bool perf_evlist__add_vfs_getname(struct perf_evlist *evlist)
2396 {
2397         struct perf_evsel *evsel = perf_evsel__newtp("probe", "vfs_getname");
2398
2399         if (IS_ERR(evsel))
2400                 return false;
2401
2402         if (perf_evsel__field(evsel, "pathname") == NULL) {
2403                 perf_evsel__delete(evsel);
2404                 return false;
2405         }
2406
2407         evsel->handler = trace__vfs_getname;
2408         perf_evlist__add(evlist, evsel);
2409         return true;
2410 }
2411
2412 static struct perf_evsel *perf_evsel__new_pgfault(u64 config)
2413 {
2414         struct perf_evsel *evsel;
2415         struct perf_event_attr attr = {
2416                 .type = PERF_TYPE_SOFTWARE,
2417                 .mmap_data = 1,
2418         };
2419
2420         attr.config = config;
2421         attr.sample_period = 1;
2422
2423         event_attr_init(&attr);
2424
2425         evsel = perf_evsel__new(&attr);
2426         if (evsel)
2427                 evsel->handler = trace__pgfault;
2428
2429         return evsel;
2430 }
2431
2432 static void trace__handle_event(struct trace *trace, union perf_event *event, struct perf_sample *sample)
2433 {
2434         const u32 type = event->header.type;
2435         struct perf_evsel *evsel;
2436
2437         if (type != PERF_RECORD_SAMPLE) {
2438                 trace__process_event(trace, trace->host, event, sample);
2439                 return;
2440         }
2441
2442         evsel = perf_evlist__id2evsel(trace->evlist, sample->id);
2443         if (evsel == NULL) {
2444                 fprintf(trace->output, "Unknown tp ID %" PRIu64 ", skipping...\n", sample->id);
2445                 return;
2446         }
2447
2448         trace__set_base_time(trace, evsel, sample);
2449
2450         if (evsel->attr.type == PERF_TYPE_TRACEPOINT &&
2451             sample->raw_data == NULL) {
2452                 fprintf(trace->output, "%s sample with no payload for tid: %d, cpu %d, raw_size=%d, skipping...\n",
2453                        perf_evsel__name(evsel), sample->tid,
2454                        sample->cpu, sample->raw_size);
2455         } else {
2456                 tracepoint_handler handler = evsel->handler;
2457                 handler(trace, evsel, event, sample);
2458         }
2459
2460         if (trace->nr_events_printed >= trace->max_events && trace->max_events != ULONG_MAX)
2461                 interrupted = true;
2462 }
2463
2464 static int trace__add_syscall_newtp(struct trace *trace)
2465 {
2466         int ret = -1;
2467         struct perf_evlist *evlist = trace->evlist;
2468         struct perf_evsel *sys_enter, *sys_exit;
2469
2470         sys_enter = perf_evsel__raw_syscall_newtp("sys_enter", trace__sys_enter);
2471         if (sys_enter == NULL)
2472                 goto out;
2473
2474         if (perf_evsel__init_sc_tp_ptr_field(sys_enter, args))
2475                 goto out_delete_sys_enter;
2476
2477         sys_exit = perf_evsel__raw_syscall_newtp("sys_exit", trace__sys_exit);
2478         if (sys_exit == NULL)
2479                 goto out_delete_sys_enter;
2480
2481         if (perf_evsel__init_sc_tp_uint_field(sys_exit, ret))
2482                 goto out_delete_sys_exit;
2483
2484         perf_evsel__config_callchain(sys_enter, &trace->opts, &callchain_param);
2485         perf_evsel__config_callchain(sys_exit, &trace->opts, &callchain_param);
2486
2487         perf_evlist__add(evlist, sys_enter);
2488         perf_evlist__add(evlist, sys_exit);
2489
2490         if (callchain_param.enabled && !trace->kernel_syscallchains) {
2491                 /*
2492                  * We're interested only in the user space callchain
2493                  * leading to the syscall, allow overriding that for
2494                  * debugging reasons using --kernel_syscall_callchains
2495                  */
2496                 sys_exit->attr.exclude_callchain_kernel = 1;
2497         }
2498
2499         trace->syscalls.events.sys_enter = sys_enter;
2500         trace->syscalls.events.sys_exit  = sys_exit;
2501
2502         ret = 0;
2503 out:
2504         return ret;
2505
2506 out_delete_sys_exit:
2507         perf_evsel__delete_priv(sys_exit);
2508 out_delete_sys_enter:
2509         perf_evsel__delete_priv(sys_enter);
2510         goto out;
2511 }
2512
2513 static int trace__set_ev_qualifier_filter(struct trace *trace)
2514 {
2515         int err = -1;
2516         struct perf_evsel *sys_exit;
2517         char *filter = asprintf_expr_inout_ints("id", !trace->not_ev_qualifier,
2518                                                 trace->ev_qualifier_ids.nr,
2519                                                 trace->ev_qualifier_ids.entries);
2520
2521         if (filter == NULL)
2522                 goto out_enomem;
2523
2524         if (!perf_evsel__append_tp_filter(trace->syscalls.events.sys_enter,
2525                                           filter)) {
2526                 sys_exit = trace->syscalls.events.sys_exit;
2527                 err = perf_evsel__append_tp_filter(sys_exit, filter);
2528         }
2529
2530         free(filter);
2531 out:
2532         return err;
2533 out_enomem:
2534         errno = ENOMEM;
2535         goto out;
2536 }
2537
2538 static int trace__set_filter_loop_pids(struct trace *trace)
2539 {
2540         unsigned int nr = 1;
2541         pid_t pids[32] = {
2542                 getpid(),
2543         };
2544         struct thread *thread = machine__find_thread(trace->host, pids[0], pids[0]);
2545
2546         while (thread && nr < ARRAY_SIZE(pids)) {
2547                 struct thread *parent = machine__find_thread(trace->host, thread->ppid, thread->ppid);
2548
2549                 if (parent == NULL)
2550                         break;
2551
2552                 if (!strcmp(thread__comm_str(parent), "sshd")) {
2553                         pids[nr++] = parent->tid;
2554                         break;
2555                 }
2556                 thread = parent;
2557         }
2558
2559         return perf_evlist__set_filter_pids(trace->evlist, nr, pids);
2560 }
2561
2562 static int trace__run(struct trace *trace, int argc, const char **argv)
2563 {
2564         struct perf_evlist *evlist = trace->evlist;
2565         struct perf_evsel *evsel, *pgfault_maj = NULL, *pgfault_min = NULL;
2566         int err = -1, i;
2567         unsigned long before;
2568         const bool forks = argc > 0;
2569         bool draining = false;
2570
2571         trace->live = true;
2572
2573         if (trace->trace_syscalls && trace__add_syscall_newtp(trace))
2574                 goto out_error_raw_syscalls;
2575
2576         if (trace->trace_syscalls)
2577                 trace->vfs_getname = perf_evlist__add_vfs_getname(evlist);
2578
2579         if ((trace->trace_pgfaults & TRACE_PFMAJ)) {
2580                 pgfault_maj = perf_evsel__new_pgfault(PERF_COUNT_SW_PAGE_FAULTS_MAJ);
2581                 if (pgfault_maj == NULL)
2582                         goto out_error_mem;
2583                 perf_evsel__config_callchain(pgfault_maj, &trace->opts, &callchain_param);
2584                 perf_evlist__add(evlist, pgfault_maj);
2585         }
2586
2587         if ((trace->trace_pgfaults & TRACE_PFMIN)) {
2588                 pgfault_min = perf_evsel__new_pgfault(PERF_COUNT_SW_PAGE_FAULTS_MIN);
2589                 if (pgfault_min == NULL)
2590                         goto out_error_mem;
2591                 perf_evsel__config_callchain(pgfault_min, &trace->opts, &callchain_param);
2592                 perf_evlist__add(evlist, pgfault_min);
2593         }
2594
2595         if (trace->sched &&
2596             perf_evlist__add_newtp(evlist, "sched", "sched_stat_runtime",
2597                                    trace__sched_stat_runtime))
2598                 goto out_error_sched_stat_runtime;
2599
2600         /*
2601          * If a global cgroup was set, apply it to all the events without an
2602          * explicit cgroup. I.e.:
2603          *
2604          *      trace -G A -e sched:*switch
2605          *
2606          * Will set all raw_syscalls:sys_{enter,exit}, pgfault, vfs_getname, etc
2607          * _and_ sched:sched_switch to the 'A' cgroup, while:
2608          *
2609          * trace -e sched:*switch -G A
2610          *
2611          * will only set the sched:sched_switch event to the 'A' cgroup, all the
2612          * other events (raw_syscalls:sys_{enter,exit}, etc are left "without"
2613          * a cgroup (on the root cgroup, sys wide, etc).
2614          *
2615          * Multiple cgroups:
2616          *
2617          * trace -G A -e sched:*switch -G B
2618          *
2619          * the syscall ones go to the 'A' cgroup, the sched:sched_switch goes
2620          * to the 'B' cgroup.
2621          *
2622          * evlist__set_default_cgroup() grabs a reference of the passed cgroup
2623          * only for the evsels still without a cgroup, i.e. evsel->cgroup == NULL.
2624          */
2625         if (trace->cgroup)
2626                 evlist__set_default_cgroup(trace->evlist, trace->cgroup);
2627
2628         err = perf_evlist__create_maps(evlist, &trace->opts.target);
2629         if (err < 0) {
2630                 fprintf(trace->output, "Problems parsing the target to trace, check your options!\n");
2631                 goto out_delete_evlist;
2632         }
2633
2634         err = trace__symbols_init(trace, evlist);
2635         if (err < 0) {
2636                 fprintf(trace->output, "Problems initializing symbol libraries!\n");
2637                 goto out_delete_evlist;
2638         }
2639
2640         perf_evlist__config(evlist, &trace->opts, &callchain_param);
2641
2642         signal(SIGCHLD, sig_handler);
2643         signal(SIGINT, sig_handler);
2644
2645         if (forks) {
2646                 err = perf_evlist__prepare_workload(evlist, &trace->opts.target,
2647                                                     argv, false, NULL);
2648                 if (err < 0) {
2649                         fprintf(trace->output, "Couldn't run the workload!\n");
2650                         goto out_delete_evlist;
2651                 }
2652         }
2653
2654         err = perf_evlist__open(evlist);
2655         if (err < 0)
2656                 goto out_error_open;
2657
2658         err = bpf__apply_obj_config();
2659         if (err) {
2660                 char errbuf[BUFSIZ];
2661
2662                 bpf__strerror_apply_obj_config(err, errbuf, sizeof(errbuf));
2663                 pr_err("ERROR: Apply config to BPF failed: %s\n",
2664                          errbuf);
2665                 goto out_error_open;
2666         }
2667
2668         /*
2669          * Better not use !target__has_task() here because we need to cover the
2670          * case where no threads were specified in the command line, but a
2671          * workload was, and in that case we will fill in the thread_map when
2672          * we fork the workload in perf_evlist__prepare_workload.
2673          */
2674         if (trace->filter_pids.nr > 0)
2675                 err = perf_evlist__set_filter_pids(evlist, trace->filter_pids.nr, trace->filter_pids.entries);
2676         else if (thread_map__pid(evlist->threads, 0) == -1)
2677                 err = trace__set_filter_loop_pids(trace);
2678
2679         if (err < 0)
2680                 goto out_error_mem;
2681
2682         if (trace->ev_qualifier_ids.nr > 0) {
2683                 err = trace__set_ev_qualifier_filter(trace);
2684                 if (err < 0)
2685                         goto out_errno;
2686
2687                 pr_debug("event qualifier tracepoint filter: %s\n",
2688                          trace->syscalls.events.sys_exit->filter);
2689         }
2690
2691         err = perf_evlist__apply_filters(evlist, &evsel);
2692         if (err < 0)
2693                 goto out_error_apply_filters;
2694
2695         err = perf_evlist__mmap(evlist, trace->opts.mmap_pages);
2696         if (err < 0)
2697                 goto out_error_mmap;
2698
2699         if (!target__none(&trace->opts.target) && !trace->opts.initial_delay)
2700                 perf_evlist__enable(evlist);
2701
2702         if (forks)
2703                 perf_evlist__start_workload(evlist);
2704
2705         if (trace->opts.initial_delay) {
2706                 usleep(trace->opts.initial_delay * 1000);
2707                 perf_evlist__enable(evlist);
2708         }
2709
2710         trace->multiple_threads = thread_map__pid(evlist->threads, 0) == -1 ||
2711                                   evlist->threads->nr > 1 ||
2712                                   perf_evlist__first(evlist)->attr.inherit;
2713
2714         /*
2715          * Now that we already used evsel->attr to ask the kernel to setup the
2716          * events, lets reuse evsel->attr.sample_max_stack as the limit in
2717          * trace__resolve_callchain(), allowing per-event max-stack settings
2718          * to override an explicitely set --max-stack global setting.
2719          */
2720         evlist__for_each_entry(evlist, evsel) {
2721                 if (evsel__has_callchain(evsel) &&
2722                     evsel->attr.sample_max_stack == 0)
2723                         evsel->attr.sample_max_stack = trace->max_stack;
2724         }
2725 again:
2726         before = trace->nr_events;
2727
2728         for (i = 0; i < evlist->nr_mmaps; i++) {
2729                 union perf_event *event;
2730                 struct perf_mmap *md;
2731
2732                 md = &evlist->mmap[i];
2733                 if (perf_mmap__read_init(md) < 0)
2734                         continue;
2735
2736                 while ((event = perf_mmap__read_event(md)) != NULL) {
2737                         struct perf_sample sample;
2738
2739                         ++trace->nr_events;
2740
2741                         err = perf_evlist__parse_sample(evlist, event, &sample);
2742                         if (err) {
2743                                 fprintf(trace->output, "Can't parse sample, err = %d, skipping...\n", err);
2744                                 goto next_event;
2745                         }
2746
2747                         trace__handle_event(trace, event, &sample);
2748 next_event:
2749                         perf_mmap__consume(md);
2750
2751                         if (interrupted)
2752                                 goto out_disable;
2753
2754                         if (done && !draining) {
2755                                 perf_evlist__disable(evlist);
2756                                 draining = true;
2757                         }
2758                 }
2759                 perf_mmap__read_done(md);
2760         }
2761
2762         if (trace->nr_events == before) {
2763                 int timeout = done ? 100 : -1;
2764
2765                 if (!draining && perf_evlist__poll(evlist, timeout) > 0) {
2766                         if (perf_evlist__filter_pollfd(evlist, POLLERR | POLLHUP | POLLNVAL) == 0)
2767                                 draining = true;
2768
2769                         goto again;
2770                 }
2771         } else {
2772                 goto again;
2773         }
2774
2775 out_disable:
2776         thread__zput(trace->current);
2777
2778         perf_evlist__disable(evlist);
2779
2780         if (!err) {
2781                 if (trace->summary)
2782                         trace__fprintf_thread_summary(trace, trace->output);
2783
2784                 if (trace->show_tool_stats) {
2785                         fprintf(trace->output, "Stats:\n "
2786                                                " vfs_getname : %" PRIu64 "\n"
2787                                                " proc_getname: %" PRIu64 "\n",
2788                                 trace->stats.vfs_getname,
2789                                 trace->stats.proc_getname);
2790                 }
2791         }
2792
2793 out_delete_evlist:
2794         trace__symbols__exit(trace);
2795
2796         perf_evlist__delete(evlist);
2797         cgroup__put(trace->cgroup);
2798         trace->evlist = NULL;
2799         trace->live = false;
2800         return err;
2801 {
2802         char errbuf[BUFSIZ];
2803
2804 out_error_sched_stat_runtime:
2805         tracing_path__strerror_open_tp(errno, errbuf, sizeof(errbuf), "sched", "sched_stat_runtime");
2806         goto out_error;
2807
2808 out_error_raw_syscalls:
2809         tracing_path__strerror_open_tp(errno, errbuf, sizeof(errbuf), "raw_syscalls", "sys_(enter|exit)");
2810         goto out_error;
2811
2812 out_error_mmap:
2813         perf_evlist__strerror_mmap(evlist, errno, errbuf, sizeof(errbuf));
2814         goto out_error;
2815
2816 out_error_open:
2817         perf_evlist__strerror_open(evlist, errno, errbuf, sizeof(errbuf));
2818
2819 out_error:
2820         fprintf(trace->output, "%s\n", errbuf);
2821         goto out_delete_evlist;
2822
2823 out_error_apply_filters:
2824         fprintf(trace->output,
2825                 "Failed to set filter \"%s\" on event %s with %d (%s)\n",
2826                 evsel->filter, perf_evsel__name(evsel), errno,
2827                 str_error_r(errno, errbuf, sizeof(errbuf)));
2828         goto out_delete_evlist;
2829 }
2830 out_error_mem:
2831         fprintf(trace->output, "Not enough memory to run!\n");
2832         goto out_delete_evlist;
2833
2834 out_errno:
2835         fprintf(trace->output, "errno=%d,%s\n", errno, strerror(errno));
2836         goto out_delete_evlist;
2837 }
2838
2839 static int trace__replay(struct trace *trace)
2840 {
2841         const struct perf_evsel_str_handler handlers[] = {
2842                 { "probe:vfs_getname",       trace__vfs_getname, },
2843         };
2844         struct perf_data data = {
2845                 .file      = {
2846                         .path = input_name,
2847                 },
2848                 .mode      = PERF_DATA_MODE_READ,
2849                 .force     = trace->force,
2850         };
2851         struct perf_session *session;
2852         struct perf_evsel *evsel;
2853         int err = -1;
2854
2855         trace->tool.sample        = trace__process_sample;
2856         trace->tool.mmap          = perf_event__process_mmap;
2857         trace->tool.mmap2         = perf_event__process_mmap2;
2858         trace->tool.comm          = perf_event__process_comm;
2859         trace->tool.exit          = perf_event__process_exit;
2860         trace->tool.fork          = perf_event__process_fork;
2861         trace->tool.attr          = perf_event__process_attr;
2862         trace->tool.tracing_data  = perf_event__process_tracing_data;
2863         trace->tool.build_id      = perf_event__process_build_id;
2864         trace->tool.namespaces    = perf_event__process_namespaces;
2865
2866         trace->tool.ordered_events = true;
2867         trace->tool.ordering_requires_timestamps = true;
2868
2869         /* add tid to output */
2870         trace->multiple_threads = true;
2871
2872         session = perf_session__new(&data, false, &trace->tool);
2873         if (session == NULL)
2874                 return -1;
2875
2876         if (trace->opts.target.pid)
2877                 symbol_conf.pid_list_str = strdup(trace->opts.target.pid);
2878
2879         if (trace->opts.target.tid)
2880                 symbol_conf.tid_list_str = strdup(trace->opts.target.tid);
2881
2882         if (symbol__init(&session->header.env) < 0)
2883                 goto out;
2884
2885         trace->host = &session->machines.host;
2886
2887         err = perf_session__set_tracepoints_handlers(session, handlers);
2888         if (err)
2889                 goto out;
2890
2891         evsel = perf_evlist__find_tracepoint_by_name(session->evlist,
2892                                                      "raw_syscalls:sys_enter");
2893         /* older kernels have syscalls tp versus raw_syscalls */
2894         if (evsel == NULL)
2895                 evsel = perf_evlist__find_tracepoint_by_name(session->evlist,
2896                                                              "syscalls:sys_enter");
2897
2898         if (evsel &&
2899             (perf_evsel__init_raw_syscall_tp(evsel, trace__sys_enter) < 0 ||
2900             perf_evsel__init_sc_tp_ptr_field(evsel, args))) {
2901                 pr_err("Error during initialize raw_syscalls:sys_enter event\n");
2902                 goto out;
2903         }
2904
2905         evsel = perf_evlist__find_tracepoint_by_name(session->evlist,
2906                                                      "raw_syscalls:sys_exit");
2907         if (evsel == NULL)
2908                 evsel = perf_evlist__find_tracepoint_by_name(session->evlist,
2909                                                              "syscalls:sys_exit");
2910         if (evsel &&
2911             (perf_evsel__init_raw_syscall_tp(evsel, trace__sys_exit) < 0 ||
2912             perf_evsel__init_sc_tp_uint_field(evsel, ret))) {
2913                 pr_err("Error during initialize raw_syscalls:sys_exit event\n");
2914                 goto out;
2915         }
2916
2917         evlist__for_each_entry(session->evlist, evsel) {
2918                 if (evsel->attr.type == PERF_TYPE_SOFTWARE &&
2919                     (evsel->attr.config == PERF_COUNT_SW_PAGE_FAULTS_MAJ ||
2920                      evsel->attr.config == PERF_COUNT_SW_PAGE_FAULTS_MIN ||
2921                      evsel->attr.config == PERF_COUNT_SW_PAGE_FAULTS))
2922                         evsel->handler = trace__pgfault;
2923         }
2924
2925         setup_pager();
2926
2927         err = perf_session__process_events(session);
2928         if (err)
2929                 pr_err("Failed to process events, error %d", err);
2930
2931         else if (trace->summary)
2932                 trace__fprintf_thread_summary(trace, trace->output);
2933
2934 out:
2935         perf_session__delete(session);
2936
2937         return err;
2938 }
2939
2940 static size_t trace__fprintf_threads_header(FILE *fp)
2941 {
2942         size_t printed;
2943
2944         printed  = fprintf(fp, "\n Summary of events:\n\n");
2945
2946         return printed;
2947 }
2948
2949 DEFINE_RESORT_RB(syscall_stats, a->msecs > b->msecs,
2950         struct stats    *stats;
2951         double          msecs;
2952         int             syscall;
2953 )
2954 {
2955         struct int_node *source = rb_entry(nd, struct int_node, rb_node);
2956         struct stats *stats = source->priv;
2957
2958         entry->syscall = source->i;
2959         entry->stats   = stats;
2960         entry->msecs   = stats ? (u64)stats->n * (avg_stats(stats) / NSEC_PER_MSEC) : 0;
2961 }
2962
2963 static size_t thread__dump_stats(struct thread_trace *ttrace,
2964                                  struct trace *trace, FILE *fp)
2965 {
2966         size_t printed = 0;
2967         struct syscall *sc;
2968         struct rb_node *nd;
2969         DECLARE_RESORT_RB_INTLIST(syscall_stats, ttrace->syscall_stats);
2970
2971         if (syscall_stats == NULL)
2972                 return 0;
2973
2974         printed += fprintf(fp, "\n");
2975
2976         printed += fprintf(fp, "   syscall            calls    total       min       avg       max      stddev\n");
2977         printed += fprintf(fp, "                               (msec)    (msec)    (msec)    (msec)        (%%)\n");
2978         printed += fprintf(fp, "   --------------- -------- --------- --------- --------- ---------     ------\n");
2979
2980         resort_rb__for_each_entry(nd, syscall_stats) {
2981                 struct stats *stats = syscall_stats_entry->stats;
2982                 if (stats) {
2983                         double min = (double)(stats->min) / NSEC_PER_MSEC;
2984                         double max = (double)(stats->max) / NSEC_PER_MSEC;
2985                         double avg = avg_stats(stats);
2986                         double pct;
2987                         u64 n = (u64) stats->n;
2988
2989                         pct = avg ? 100.0 * stddev_stats(stats)/avg : 0.0;
2990                         avg /= NSEC_PER_MSEC;
2991
2992                         sc = &trace->syscalls.table[syscall_stats_entry->syscall];
2993                         printed += fprintf(fp, "   %-15s", sc->name);
2994                         printed += fprintf(fp, " %8" PRIu64 " %9.3f %9.3f %9.3f",
2995                                            n, syscall_stats_entry->msecs, min, avg);
2996                         printed += fprintf(fp, " %9.3f %9.2f%%\n", max, pct);
2997                 }
2998         }
2999
3000         resort_rb__delete(syscall_stats);
3001         printed += fprintf(fp, "\n\n");
3002
3003         return printed;
3004 }
3005
3006 static size_t trace__fprintf_thread(FILE *fp, struct thread *thread, struct trace *trace)
3007 {
3008         size_t printed = 0;
3009         struct thread_trace *ttrace = thread__priv(thread);
3010         double ratio;
3011
3012         if (ttrace == NULL)
3013                 return 0;
3014
3015         ratio = (double)ttrace->nr_events / trace->nr_events * 100.0;
3016
3017         printed += fprintf(fp, " %s (%d), ", thread__comm_str(thread), thread->tid);
3018         printed += fprintf(fp, "%lu events, ", ttrace->nr_events);
3019         printed += fprintf(fp, "%.1f%%", ratio);
3020         if (ttrace->pfmaj)
3021                 printed += fprintf(fp, ", %lu majfaults", ttrace->pfmaj);
3022         if (ttrace->pfmin)
3023                 printed += fprintf(fp, ", %lu minfaults", ttrace->pfmin);
3024         if (trace->sched)
3025                 printed += fprintf(fp, ", %.3f msec\n", ttrace->runtime_ms);
3026         else if (fputc('\n', fp) != EOF)
3027                 ++printed;
3028
3029         printed += thread__dump_stats(ttrace, trace, fp);
3030
3031         return printed;
3032 }
3033
3034 static unsigned long thread__nr_events(struct thread_trace *ttrace)
3035 {
3036         return ttrace ? ttrace->nr_events : 0;
3037 }
3038
3039 DEFINE_RESORT_RB(threads, (thread__nr_events(a->thread->priv) < thread__nr_events(b->thread->priv)),
3040         struct thread *thread;
3041 )
3042 {
3043         entry->thread = rb_entry(nd, struct thread, rb_node);
3044 }
3045
3046 static size_t trace__fprintf_thread_summary(struct trace *trace, FILE *fp)
3047 {
3048         size_t printed = trace__fprintf_threads_header(fp);
3049         struct rb_node *nd;
3050         int i;
3051
3052         for (i = 0; i < THREADS__TABLE_SIZE; i++) {
3053                 DECLARE_RESORT_RB_MACHINE_THREADS(threads, trace->host, i);
3054
3055                 if (threads == NULL) {
3056                         fprintf(fp, "%s", "Error sorting output by nr_events!\n");
3057                         return 0;
3058                 }
3059
3060                 resort_rb__for_each_entry(nd, threads)
3061                         printed += trace__fprintf_thread(fp, threads_entry->thread, trace);
3062
3063                 resort_rb__delete(threads);
3064         }
3065         return printed;
3066 }
3067
3068 static int trace__set_duration(const struct option *opt, const char *str,
3069                                int unset __maybe_unused)
3070 {
3071         struct trace *trace = opt->value;
3072
3073         trace->duration_filter = atof(str);
3074         return 0;
3075 }
3076
3077 static int trace__set_filter_pids(const struct option *opt, const char *str,
3078                                   int unset __maybe_unused)
3079 {
3080         int ret = -1;
3081         size_t i;
3082         struct trace *trace = opt->value;
3083         /*
3084          * FIXME: introduce a intarray class, plain parse csv and create a
3085          * { int nr, int entries[] } struct...
3086          */
3087         struct intlist *list = intlist__new(str);
3088
3089         if (list == NULL)
3090                 return -1;
3091
3092         i = trace->filter_pids.nr = intlist__nr_entries(list) + 1;
3093         trace->filter_pids.entries = calloc(i, sizeof(pid_t));
3094
3095         if (trace->filter_pids.entries == NULL)
3096                 goto out;
3097
3098         trace->filter_pids.entries[0] = getpid();
3099
3100         for (i = 1; i < trace->filter_pids.nr; ++i)
3101                 trace->filter_pids.entries[i] = intlist__entry(list, i - 1)->i;
3102
3103         intlist__delete(list);
3104         ret = 0;
3105 out:
3106         return ret;
3107 }
3108
3109 static int trace__open_output(struct trace *trace, const char *filename)
3110 {
3111         struct stat st;
3112
3113         if (!stat(filename, &st) && st.st_size) {
3114                 char oldname[PATH_MAX];
3115
3116                 scnprintf(oldname, sizeof(oldname), "%s.old", filename);
3117                 unlink(oldname);
3118                 rename(filename, oldname);
3119         }
3120
3121         trace->output = fopen(filename, "w");
3122
3123         return trace->output == NULL ? -errno : 0;
3124 }
3125
3126 static int parse_pagefaults(const struct option *opt, const char *str,
3127                             int unset __maybe_unused)
3128 {
3129         int *trace_pgfaults = opt->value;
3130
3131         if (strcmp(str, "all") == 0)
3132                 *trace_pgfaults |= TRACE_PFMAJ | TRACE_PFMIN;
3133         else if (strcmp(str, "maj") == 0)
3134                 *trace_pgfaults |= TRACE_PFMAJ;
3135         else if (strcmp(str, "min") == 0)
3136                 *trace_pgfaults |= TRACE_PFMIN;
3137         else
3138                 return -1;
3139
3140         return 0;
3141 }
3142
3143 static void evlist__set_evsel_handler(struct perf_evlist *evlist, void *handler)
3144 {
3145         struct perf_evsel *evsel;
3146
3147         evlist__for_each_entry(evlist, evsel)
3148                 evsel->handler = handler;
3149 }
3150
3151 static int evlist__set_syscall_tp_fields(struct perf_evlist *evlist)
3152 {
3153         struct perf_evsel *evsel;
3154
3155         evlist__for_each_entry(evlist, evsel) {
3156                 if (evsel->priv || !evsel->tp_format)
3157                         continue;
3158
3159                 if (strcmp(evsel->tp_format->system, "syscalls"))
3160                         continue;
3161
3162                 if (perf_evsel__init_syscall_tp(evsel))
3163                         return -1;
3164
3165                 if (!strncmp(evsel->tp_format->name, "sys_enter_", 10)) {
3166                         struct syscall_tp *sc = evsel->priv;
3167
3168                         if (__tp_field__init_ptr(&sc->args, sc->id.offset + sizeof(u64)))
3169                                 return -1;
3170                 } else if (!strncmp(evsel->tp_format->name, "sys_exit_", 9)) {
3171                         struct syscall_tp *sc = evsel->priv;
3172
3173                         if (__tp_field__init_uint(&sc->ret, sizeof(u64), sc->id.offset + sizeof(u64), evsel->needs_swap))
3174                                 return -1;
3175                 }
3176         }
3177
3178         return 0;
3179 }
3180
3181 /*
3182  * XXX: Hackish, just splitting the combined -e+--event (syscalls
3183  * (raw_syscalls:{sys_{enter,exit}} + events (tracepoints, HW, SW, etc) to use
3184  * existing facilities unchanged (trace->ev_qualifier + parse_options()).
3185  *
3186  * It'd be better to introduce a parse_options() variant that would return a
3187  * list with the terms it didn't match to an event...
3188  */
3189 static int trace__parse_events_option(const struct option *opt, const char *str,
3190                                       int unset __maybe_unused)
3191 {
3192         struct trace *trace = (struct trace *)opt->value;
3193         const char *s = str;
3194         char *sep = NULL, *lists[2] = { NULL, NULL, };
3195         int len = strlen(str) + 1, err = -1, list, idx;
3196         char *strace_groups_dir = system_path(STRACE_GROUPS_DIR);
3197         char group_name[PATH_MAX];
3198
3199         if (strace_groups_dir == NULL)
3200                 return -1;
3201
3202         if (*s == '!') {
3203                 ++s;
3204                 trace->not_ev_qualifier = true;
3205         }
3206
3207         while (1) {
3208                 if ((sep = strchr(s, ',')) != NULL)
3209                         *sep = '\0';
3210
3211                 list = 0;
3212                 if (syscalltbl__id(trace->sctbl, s) >= 0 ||
3213                     syscalltbl__strglobmatch_first(trace->sctbl, s, &idx) >= 0) {
3214                         list = 1;
3215                 } else {
3216                         path__join(group_name, sizeof(group_name), strace_groups_dir, s);
3217                         if (access(group_name, R_OK) == 0)
3218                                 list = 1;
3219                 }
3220
3221                 if (lists[list]) {
3222                         sprintf(lists[list] + strlen(lists[list]), ",%s", s);
3223                 } else {
3224                         lists[list] = malloc(len);
3225                         if (lists[list] == NULL)
3226                                 goto out;
3227                         strcpy(lists[list], s);
3228                 }
3229
3230                 if (!sep)
3231                         break;
3232
3233                 *sep = ',';
3234                 s = sep + 1;
3235         }
3236
3237         if (lists[1] != NULL) {
3238                 struct strlist_config slist_config = {
3239                         .dirname = strace_groups_dir,
3240                 };
3241
3242                 trace->ev_qualifier = strlist__new(lists[1], &slist_config);
3243                 if (trace->ev_qualifier == NULL) {
3244                         fputs("Not enough memory to parse event qualifier", trace->output);
3245                         goto out;
3246                 }
3247
3248                 if (trace__validate_ev_qualifier(trace))
3249                         goto out;
3250                 trace->trace_syscalls = true;
3251         }
3252
3253         err = 0;
3254
3255         if (lists[0]) {
3256                 struct option o = OPT_CALLBACK('e', "event", &trace->evlist, "event",
3257                                                "event selector. use 'perf list' to list available events",
3258                                                parse_events_option);
3259                 err = parse_events_option(&o, lists[0], 0);
3260         }
3261 out:
3262         if (sep)
3263                 *sep = ',';
3264
3265         return err;
3266 }
3267
3268 static int trace__parse_cgroups(const struct option *opt, const char *str, int unset)
3269 {
3270         struct trace *trace = opt->value;
3271
3272         if (!list_empty(&trace->evlist->entries))
3273                 return parse_cgroups(opt, str, unset);
3274
3275         trace->cgroup = evlist__findnew_cgroup(trace->evlist, str);
3276
3277         return 0;
3278 }
3279
3280 int cmd_trace(int argc, const char **argv)
3281 {
3282         const char *trace_usage[] = {
3283                 "perf trace [<options>] [<command>]",
3284                 "perf trace [<options>] -- <command> [<options>]",
3285                 "perf trace record [<options>] [<command>]",
3286                 "perf trace record [<options>] -- <command> [<options>]",
3287                 NULL
3288         };
3289         struct trace trace = {
3290                 .syscalls = {
3291                         . max = -1,
3292                 },
3293                 .opts = {
3294                         .target = {
3295                                 .uid       = UINT_MAX,
3296                                 .uses_mmap = true,
3297                         },
3298                         .user_freq     = UINT_MAX,
3299                         .user_interval = ULLONG_MAX,
3300                         .no_buffering  = true,
3301                         .mmap_pages    = UINT_MAX,
3302                         .proc_map_timeout  = 500,
3303                 },
3304                 .output = stderr,
3305                 .show_comm = true,
3306                 .trace_syscalls = false,
3307                 .kernel_syscallchains = false,
3308                 .max_stack = UINT_MAX,
3309                 .max_events = ULONG_MAX,
3310         };
3311         const char *output_name = NULL;
3312         const struct option trace_options[] = {
3313         OPT_CALLBACK('e', "event", &trace, "event",
3314                      "event/syscall selector. use 'perf list' to list available events",
3315                      trace__parse_events_option),
3316         OPT_BOOLEAN(0, "comm", &trace.show_comm,
3317                     "show the thread COMM next to its id"),
3318         OPT_BOOLEAN(0, "tool_stats", &trace.show_tool_stats, "show tool stats"),
3319         OPT_CALLBACK(0, "expr", &trace, "expr", "list of syscalls/events to trace",
3320                      trace__parse_events_option),
3321         OPT_STRING('o', "output", &output_name, "file", "output file name"),
3322         OPT_STRING('i', "input", &input_name, "file", "Analyze events in file"),
3323         OPT_STRING('p', "pid", &trace.opts.target.pid, "pid",
3324                     "trace events on existing process id"),
3325         OPT_STRING('t', "tid", &trace.opts.target.tid, "tid",
3326                     "trace events on existing thread id"),
3327         OPT_CALLBACK(0, "filter-pids", &trace, "CSV list of pids",
3328                      "pids to filter (by the kernel)", trace__set_filter_pids),
3329         OPT_BOOLEAN('a', "all-cpus", &trace.opts.target.system_wide,
3330                     "system-wide collection from all CPUs"),
3331         OPT_STRING('C', "cpu", &trace.opts.target.cpu_list, "cpu",
3332                     "list of cpus to monitor"),
3333         OPT_BOOLEAN(0, "no-inherit", &trace.opts.no_inherit,
3334                     "child tasks do not inherit counters"),
3335         OPT_CALLBACK('m', "mmap-pages", &trace.opts.mmap_pages, "pages",
3336                      "number of mmap data pages",
3337                      perf_evlist__parse_mmap_pages),
3338         OPT_STRING('u', "uid", &trace.opts.target.uid_str, "user",
3339                    "user to profile"),
3340         OPT_CALLBACK(0, "duration", &trace, "float",
3341                      "show only events with duration > N.M ms",
3342                      trace__set_duration),
3343         OPT_BOOLEAN(0, "sched", &trace.sched, "show blocking scheduler events"),
3344         OPT_INCR('v', "verbose", &verbose, "be more verbose"),
3345         OPT_BOOLEAN('T', "time", &trace.full_time,
3346                     "Show full timestamp, not time relative to first start"),
3347         OPT_BOOLEAN(0, "failure", &trace.failure_only,
3348                     "Show only syscalls that failed"),
3349         OPT_BOOLEAN('s', "summary", &trace.summary_only,
3350                     "Show only syscall summary with statistics"),
3351         OPT_BOOLEAN('S', "with-summary", &trace.summary,
3352                     "Show all syscalls and summary with statistics"),
3353         OPT_CALLBACK_DEFAULT('F', "pf", &trace.trace_pgfaults, "all|maj|min",
3354                      "Trace pagefaults", parse_pagefaults, "maj"),
3355         OPT_BOOLEAN(0, "syscalls", &trace.trace_syscalls, "Trace syscalls"),
3356         OPT_BOOLEAN('f', "force", &trace.force, "don't complain, do it"),
3357         OPT_CALLBACK(0, "call-graph", &trace.opts,
3358                      "record_mode[,record_size]", record_callchain_help,
3359                      &record_parse_callchain_opt),
3360         OPT_BOOLEAN(0, "kernel-syscall-graph", &trace.kernel_syscallchains,
3361                     "Show the kernel callchains on the syscall exit path"),
3362         OPT_ULONG(0, "max-events", &trace.max_events,
3363                 "Set the maximum number of events to print, exit after that is reached. "),
3364         OPT_UINTEGER(0, "min-stack", &trace.min_stack,
3365                      "Set the minimum stack depth when parsing the callchain, "
3366                      "anything below the specified depth will be ignored."),
3367         OPT_UINTEGER(0, "max-stack", &trace.max_stack,
3368                      "Set the maximum stack depth when parsing the callchain, "
3369                      "anything beyond the specified depth will be ignored. "
3370                      "Default: kernel.perf_event_max_stack or " __stringify(PERF_MAX_STACK_DEPTH)),
3371         OPT_BOOLEAN(0, "print-sample", &trace.print_sample,
3372                         "print the PERF_RECORD_SAMPLE PERF_SAMPLE_ info, for debugging"),
3373         OPT_UINTEGER(0, "proc-map-timeout", &trace.opts.proc_map_timeout,
3374                         "per thread proc mmap processing timeout in ms"),
3375         OPT_CALLBACK('G', "cgroup", &trace, "name", "monitor event in cgroup name only",
3376                      trace__parse_cgroups),
3377         OPT_UINTEGER('D', "delay", &trace.opts.initial_delay,
3378                      "ms to wait before starting measurement after program "
3379                      "start"),
3380         OPT_END()
3381         };
3382         bool __maybe_unused max_stack_user_set = true;
3383         bool mmap_pages_user_set = true;
3384         struct perf_evsel *evsel;
3385         const char * const trace_subcommands[] = { "record", NULL };
3386         int err = -1;
3387         char bf[BUFSIZ];
3388
3389         signal(SIGSEGV, sighandler_dump_stack);
3390         signal(SIGFPE, sighandler_dump_stack);
3391
3392         trace.evlist = perf_evlist__new();
3393         trace.sctbl = syscalltbl__new();
3394
3395         if (trace.evlist == NULL || trace.sctbl == NULL) {
3396                 pr_err("Not enough memory to run!\n");
3397                 err = -ENOMEM;
3398                 goto out;
3399         }
3400
3401         argc = parse_options_subcommand(argc, argv, trace_options, trace_subcommands,
3402                                  trace_usage, PARSE_OPT_STOP_AT_NON_OPTION);
3403
3404         if ((nr_cgroups || trace.cgroup) && !trace.opts.target.system_wide) {
3405                 usage_with_options_msg(trace_usage, trace_options,
3406                                        "cgroup monitoring only available in system-wide mode");
3407         }
3408
3409         evsel = bpf__setup_output_event(trace.evlist, "__augmented_syscalls__");
3410         if (IS_ERR(evsel)) {
3411                 bpf__strerror_setup_output_event(trace.evlist, PTR_ERR(evsel), bf, sizeof(bf));
3412                 pr_err("ERROR: Setup trace syscalls enter failed: %s\n", bf);
3413                 goto out;
3414         }
3415
3416         if (evsel)
3417                 trace.syscalls.events.augmented = evsel;
3418
3419         err = bpf__setup_stdout(trace.evlist);
3420         if (err) {
3421                 bpf__strerror_setup_stdout(trace.evlist, err, bf, sizeof(bf));
3422                 pr_err("ERROR: Setup BPF stdout failed: %s\n", bf);
3423                 goto out;
3424         }
3425
3426         err = -1;
3427
3428         if (trace.trace_pgfaults) {
3429                 trace.opts.sample_address = true;
3430                 trace.opts.sample_time = true;
3431         }
3432
3433         if (trace.opts.mmap_pages == UINT_MAX)
3434                 mmap_pages_user_set = false;
3435
3436         if (trace.max_stack == UINT_MAX) {
3437                 trace.max_stack = input_name ? PERF_MAX_STACK_DEPTH : sysctl__max_stack();
3438                 max_stack_user_set = false;
3439         }
3440
3441 #ifdef HAVE_DWARF_UNWIND_SUPPORT
3442         if ((trace.min_stack || max_stack_user_set) && !callchain_param.enabled) {
3443                 record_opts__parse_callchain(&trace.opts, &callchain_param, "dwarf", false);
3444         }
3445 #endif
3446
3447         if (callchain_param.enabled) {
3448                 if (!mmap_pages_user_set && geteuid() == 0)
3449                         trace.opts.mmap_pages = perf_event_mlock_kb_in_pages() * 4;
3450
3451                 symbol_conf.use_callchain = true;
3452         }
3453
3454         if (trace.evlist->nr_entries > 0) {
3455                 evlist__set_evsel_handler(trace.evlist, trace__event_handler);
3456                 if (evlist__set_syscall_tp_fields(trace.evlist)) {
3457                         perror("failed to set syscalls:* tracepoint fields");
3458                         goto out;
3459                 }
3460         }
3461
3462         /*
3463          * If we are augmenting syscalls, then combine what we put in the
3464          * __augmented_syscalls__ BPF map with what is in the
3465          * syscalls:sys_exit_FOO tracepoints, i.e. just like we do without BPF,
3466          * combining raw_syscalls:sys_enter with raw_syscalls:sys_exit.
3467          *
3468          * We'll switch to look at two BPF maps, one for sys_enter and the
3469          * other for sys_exit when we start augmenting the sys_exit paths with
3470          * buffers that are being copied from kernel to userspace, think 'read'
3471          * syscall.
3472          */
3473         if (trace.syscalls.events.augmented) {
3474                 evsel = trace.syscalls.events.augmented;
3475
3476                 if (perf_evsel__init_augmented_syscall_tp(evsel) ||
3477                     perf_evsel__init_augmented_syscall_tp_args(evsel))
3478                         goto out;
3479                 evsel->handler = trace__sys_enter;
3480
3481                 evlist__for_each_entry(trace.evlist, evsel) {
3482                         if (strstarts(perf_evsel__name(evsel), "syscalls:sys_exit_")) {
3483                                 perf_evsel__init_augmented_syscall_tp(evsel);
3484                                 perf_evsel__init_augmented_syscall_tp_ret(evsel);
3485                                 evsel->handler = trace__sys_exit;
3486                         }
3487                 }
3488         }
3489
3490         if ((argc >= 1) && (strcmp(argv[0], "record") == 0))
3491                 return trace__record(&trace, argc-1, &argv[1]);
3492
3493         /* summary_only implies summary option, but don't overwrite summary if set */
3494         if (trace.summary_only)
3495                 trace.summary = trace.summary_only;
3496
3497         if (!trace.trace_syscalls && !trace.trace_pgfaults &&
3498             trace.evlist->nr_entries == 0 /* Was --events used? */) {
3499                 trace.trace_syscalls = true;
3500         }
3501
3502         if (output_name != NULL) {
3503                 err = trace__open_output(&trace, output_name);
3504                 if (err < 0) {
3505                         perror("failed to create output file");
3506                         goto out;
3507                 }
3508         }
3509
3510         err = target__validate(&trace.opts.target);
3511         if (err) {
3512                 target__strerror(&trace.opts.target, err, bf, sizeof(bf));
3513                 fprintf(trace.output, "%s", bf);
3514                 goto out_close;
3515         }
3516
3517         err = target__parse_uid(&trace.opts.target);
3518         if (err) {
3519                 target__strerror(&trace.opts.target, err, bf, sizeof(bf));
3520                 fprintf(trace.output, "%s", bf);
3521                 goto out_close;
3522         }
3523
3524         if (!argc && target__none(&trace.opts.target))
3525                 trace.opts.target.system_wide = true;
3526
3527         if (input_name)
3528                 err = trace__replay(&trace);
3529         else
3530                 err = trace__run(&trace, argc, argv);
3531
3532 out_close:
3533         if (output_name != NULL)
3534                 fclose(trace.output);
3535 out:
3536         return err;
3537 }