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 = { [0] = { .scnprintf = SCA_FILENAME, /* dev_name */ },
 731                    [3] = { .scnprintf = SCA_MOUNT_FLAGS, /* flags */
 732                            .mask_val  = SCAMV_MOUNT_FLAGS, /* flags */ }, }, },
 733         { .name     = "mprotect",
 734           .arg = { [0] = { .scnprintf = SCA_HEX,        /* start */ },
 735                    [2] = { .scnprintf = SCA_MMAP_PROT,  /* prot */ }, }, },
 736         { .name     = "mq_unlink",
 737           .arg = { [0] = { .scnprintf = SCA_FILENAME, /* u_name */ }, }, },
 738         { .name     = "mremap",     .hexret = true,
 739           .arg = { [0] = { .scnprintf = SCA_HEX,          /* addr */ },
 740                    [3] = { .scnprintf = SCA_MREMAP_FLAGS, /* flags */ },
 741                    [4] = { .scnprintf = SCA_HEX,          /* new_addr */ }, }, },
 742         { .name     = "munlock",
 743           .arg = { [0] = { .scnprintf = SCA_HEX, /* addr */ }, }, },
 744         { .name     = "munmap",
 745           .arg = { [0] = { .scnprintf = SCA_HEX, /* addr */ }, }, },
 746         { .name     = "name_to_handle_at",
 747           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
 748         { .name     = "newfstatat",
 749           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
 750         { .name     = "open",
 751           .arg = { [1] = { .scnprintf = SCA_OPEN_FLAGS, /* flags */ }, }, },
 752         { .name     = "open_by_handle_at",
 753           .arg = { [0] = { .scnprintf = SCA_FDAT,       /* dfd */ },
 754                    [2] = { .scnprintf = SCA_OPEN_FLAGS, /* flags */ }, }, },
 755         { .name     = "openat",
 756           .arg = { [0] = { .scnprintf = SCA_FDAT,       /* dfd */ },
 757                    [2] = { .scnprintf = SCA_OPEN_FLAGS, /* flags */ }, }, },
 758         { .name     = "perf_event_open",
 759           .arg = { [2] = { .scnprintf = SCA_INT,        /* cpu */ },
 760                    [3] = { .scnprintf = SCA_FD,         /* group_fd */ },
 761                    [4] = { .scnprintf = SCA_PERF_FLAGS, /* flags */ }, }, },
 762         { .name     = "pipe2",
 763           .arg = { [1] = { .scnprintf = SCA_PIPE_FLAGS, /* flags */ }, }, },
 764         { .name     = "pkey_alloc",
 765           .arg = { [1] = { .scnprintf = SCA_PKEY_ALLOC_ACCESS_RIGHTS,   /* access_rights */ }, }, },
 766         { .name     = "pkey_free",
 767           .arg = { [0] = { .scnprintf = SCA_INT,        /* key */ }, }, },
 768         { .name     = "pkey_mprotect",
 769           .arg = { [0] = { .scnprintf = SCA_HEX,        /* start */ },
 770                    [2] = { .scnprintf = SCA_MMAP_PROT,  /* prot */ },
 771                    [3] = { .scnprintf = SCA_INT,        /* pkey */ }, }, },
 772         { .name     = "poll", .timeout = true, },
 773         { .name     = "ppoll", .timeout = true, },
 774         { .name     = "prctl", .alias = "arch_prctl",
 775           .arg = { [0] = { .scnprintf = SCA_PRCTL_OPTION, /* option */ },
 776                    [1] = { .scnprintf = SCA_PRCTL_ARG2, /* arg2 */ },
 777                    [2] = { .scnprintf = SCA_PRCTL_ARG3, /* arg3 */ }, }, },
 778         { .name     = "pread", .alias = "pread64", },
 779         { .name     = "preadv", .alias = "pread", },
 780         { .name     = "prlimit64",
 781           .arg = { [1] = STRARRAY(resource, rlimit_resources), }, },
 782         { .name     = "pwrite", .alias = "pwrite64", },
 783         { .name     = "readlinkat",
 784           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
 785         { .name     = "recvfrom",
 786           .arg = { [3] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
 787         { .name     = "recvmmsg",
 788           .arg = { [3] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
 789         { .name     = "recvmsg",
 790           .arg = { [2] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
 791         { .name     = "renameat",
 792           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
 793         { .name     = "rt_sigaction",
 794           .arg = { [0] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
 795         { .name     = "rt_sigprocmask",
 796           .arg = { [0] = STRARRAY(how, sighow), }, },
 797         { .name     = "rt_sigqueueinfo",
 798           .arg = { [1] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
 799         { .name     = "rt_tgsigqueueinfo",
 800           .arg = { [2] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
 801         { .name     = "sched_setscheduler",
 802           .arg = { [1] = { .scnprintf = SCA_SCHED_POLICY, /* policy */ }, }, },
 803         { .name     = "seccomp",
 804           .arg = { [0] = { .scnprintf = SCA_SECCOMP_OP,    /* op */ },
 805                    [1] = { .scnprintf = SCA_SECCOMP_FLAGS, /* flags */ }, }, },
 806         { .name     = "select", .timeout = true, },
 807         { .name     = "sendmmsg",
 808           .arg = { [3] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
 809         { .name     = "sendmsg",
 810           .arg = { [2] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
 811         { .name     = "sendto",
 812           .arg = { [3] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ },
 813                    [4] = { .scnprintf = SCA_SOCKADDR, /* addr */ }, }, },
 814         { .name     = "set_tid_address", .errpid = true, },
 815         { .name     = "setitimer",
 816           .arg = { [0] = STRARRAY(which, itimers), }, },
 817         { .name     = "setrlimit",
 818           .arg = { [0] = STRARRAY(resource, rlimit_resources), }, },
 819         { .name     = "socket",
 820           .arg = { [0] = STRARRAY(family, socket_families),
 821                    [1] = { .scnprintf = SCA_SK_TYPE, /* type */ },
 822                    [2] = { .scnprintf = SCA_SK_PROTO, /* protocol */ }, }, },
 823         { .name     = "socketpair",
 824           .arg = { [0] = STRARRAY(family, socket_families),
 825                    [1] = { .scnprintf = SCA_SK_TYPE, /* type */ },
 826                    [2] = { .scnprintf = SCA_SK_PROTO, /* protocol */ }, }, },
 827         { .name     = "stat", .alias = "newstat", },
 828         { .name     = "statx",
 829           .arg = { [0] = { .scnprintf = SCA_FDAT,        /* fdat */ },
 830                    [2] = { .scnprintf = SCA_STATX_FLAGS, /* flags */ } ,
 831                    [3] = { .scnprintf = SCA_STATX_MASK,  /* mask */ }, }, },
 832         { .name     = "swapoff",
 833           .arg = { [0] = { .scnprintf = SCA_FILENAME, /* specialfile */ }, }, },
 834         { .name     = "swapon",
 835           .arg = { [0] = { .scnprintf = SCA_FILENAME, /* specialfile */ }, }, },
 836         { .name     = "symlinkat",
 837           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
 838         { .name     = "tgkill",
 839           .arg = { [2] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
 840         { .name     = "tkill",
 841           .arg = { [1] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
 842         { .name     = "umount2", .alias = "umount",
 843           .arg = { [0] = { .scnprintf = SCA_FILENAME, /* name */ }, }, },
 844         { .name     = "uname", .alias = "newuname", },
 845         { .name     = "unlinkat",
 846           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
 847         { .name     = "utimensat",
 848           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dirfd */ }, }, },
 849         { .name     = "wait4",      .errpid = true,
 850           .arg = { [2] = { .scnprintf = SCA_WAITID_OPTIONS, /* options */ }, }, },
 851         { .name     = "waitid",     .errpid = true,
 852           .arg = { [3] = { .scnprintf = SCA_WAITID_OPTIONS, /* options */ }, }, },
 853 };
 854
 855 static int syscall_fmt__cmp(const void *name, const void *fmtp)
 856 {
 857         const struct syscall_fmt *fmt = fmtp;
 858         return strcmp(name, fmt->name);
 859 }
 860
 861 static struct syscall_fmt *syscall_fmt__find(const char *name)
 862 {
 863         const int nmemb = ARRAY_SIZE(syscall_fmts);
 864         return bsearch(name, syscall_fmts, nmemb, sizeof(struct syscall_fmt), syscall_fmt__cmp);
 865 }
 866
 867 static struct syscall_fmt *syscall_fmt__find_by_alias(const char *alias)
 868 {
 869         int i, nmemb = ARRAY_SIZE(syscall_fmts);
 870
 871         for (i = 0; i < nmemb; ++i) {
 872                 if (syscall_fmts[i].alias && strcmp(syscall_fmts[i].alias, alias) == 0)
 873                         return &syscall_fmts[i];
 874         }
 875
 876         return NULL;
 877 }
 878
 879 /*
 880  * is_exit: is this "exit" or "exit_group"?
 881  * is_open: is this "open" or "openat"? To associate the fd returned in sys_exit with the pathname in sys_enter.
 882  * args_size: sum of the sizes of the syscall arguments, anything after that is augmented stuff: pathname for openat, etc.
 883  */
 884 struct syscall {
 885         struct tep_event_format *tp_format;
 886         int                 nr_args;
 887         int                 args_size;
 888         bool                is_exit;
 889         bool                is_open;
 890         struct tep_format_field *args;
 891         const char          *name;
 892         struct syscall_fmt  *fmt;
 893         struct syscall_arg_fmt *arg_fmt;
 894 };
 895
 896 /*
 897  * We need to have this 'calculated' boolean because in some cases we really
 898  * don't know what is the duration of a syscall, for instance, when we start
 899  * a session and some threads are waiting for a syscall to finish, say 'poll',
 900  * in which case all we can do is to print "( ? ) for duration and for the
 901  * start timestamp.
 902  */
 903 static size_t fprintf_duration(unsigned long t, bool calculated, FILE *fp)
 904 {
 905         double duration = (double)t / NSEC_PER_MSEC;
 906         size_t printed = fprintf(fp, "(");
 907
 908         if (!calculated)
 909                 printed += fprintf(fp, "         ");
 910         else if (duration >= 1.0)
 911                 printed += color_fprintf(fp, PERF_COLOR_RED, "%6.3f ms", duration);
 912         else if (duration >= 0.01)
 913                 printed += color_fprintf(fp, PERF_COLOR_YELLOW, "%6.3f ms", duration);
 914         else
 915                 printed += color_fprintf(fp, PERF_COLOR_NORMAL, "%6.3f ms", duration);
 916         return printed + fprintf(fp, "): ");
 917 }
 918
 919 /**
 920  * filename.ptr: The filename char pointer that will be vfs_getname'd
 921  * filename.entry_str_pos: Where to insert the string translated from
 922  *                         filename.ptr by the vfs_getname tracepoint/kprobe.
 923  * ret_scnprintf: syscall args may set this to a different syscall return
 924  *                formatter, for instance, fcntl may return fds, file flags, etc.
 925  */
 926 struct thread_trace {
 927         u64               entry_time;
 928         bool              entry_pending;
 929         unsigned long     nr_events;
 930         unsigned long     pfmaj, pfmin;
 931         char              *entry_str;
 932         double            runtime_ms;
 933         size_t            (*ret_scnprintf)(char *bf, size_t size, struct syscall_arg *arg);
 934         struct {
 935                 unsigned long ptr;
 936                 short int     entry_str_pos;
 937                 bool          pending_open;
 938                 unsigned int  namelen;
 939                 char          *name;
 940         } filename;
 941         struct {
 942                 int       max;
 943                 char      **table;
 944         } paths;
 945
 946         struct intlist *syscall_stats;
 947 };
 948
 949 static struct thread_trace *thread_trace__new(void)
 950 {
 951         struct thread_trace *ttrace =  zalloc(sizeof(struct thread_trace));
 952
 953         if (ttrace)
 954                 ttrace->paths.max = -1;
 955
 956         ttrace->syscall_stats = intlist__new(NULL);
 957
 958         return ttrace;
 959 }
 960
 961 static struct thread_trace *thread__trace(struct thread *thread, FILE *fp)
 962 {
 963         struct thread_trace *ttrace;
 964
 965         if (thread == NULL)
 966                 goto fail;
 967
 968         if (thread__priv(thread) == NULL)
 969                 thread__set_priv(thread, thread_trace__new());
 970
 971         if (thread__priv(thread) == NULL)
 972                 goto fail;
 973
 974         ttrace = thread__priv(thread);
 975         ++ttrace->nr_events;
 976
 977         return ttrace;
 978 fail:
 979         color_fprintf(fp, PERF_COLOR_RED,
 980                       "WARNING: not enough memory, dropping samples!\n");
 981         return NULL;
 982 }
 983
 984
 985 void syscall_arg__set_ret_scnprintf(struct syscall_arg *arg,
 986                                     size_t (*ret_scnprintf)(char *bf, size_t size, struct syscall_arg *arg))
 987 {
 988         struct thread_trace *ttrace = thread__priv(arg->thread);
 989
 990         ttrace->ret_scnprintf = ret_scnprintf;
 991 }
 992
 993 #define TRACE_PFMAJ             (1 << 0)
 994 #define TRACE_PFMIN             (1 << 1)
 995
 996 static const size_t trace__entry_str_size = 2048;
 997
 998 static int trace__set_fd_pathname(struct thread *thread, int fd, const char *pathname)
 999 {
1000         struct thread_trace *ttrace = thread__priv(thread);
1001
1002         if (fd > ttrace->paths.max) {
1003                 char **npath = realloc(ttrace->paths.table, (fd + 1) * sizeof(char *));
1004
1005                 if (npath == NULL)
1006                         return -1;
1007
1008                 if (ttrace->paths.max != -1) {
1009                         memset(npath + ttrace->paths.max + 1, 0,
1010                                (fd - ttrace->paths.max) * sizeof(char *));
1011                 } else {
1012                         memset(npath, 0, (fd + 1) * sizeof(char *));
1013                 }
1014
1015                 ttrace->paths.table = npath;
1016                 ttrace->paths.max   = fd;
1017         }
1018
1019         ttrace->paths.table[fd] = strdup(pathname);
1020
1021         return ttrace->paths.table[fd] != NULL ? 0 : -1;
1022 }
1023
1024 static int thread__read_fd_path(struct thread *thread, int fd)
1025 {
1026         char linkname[PATH_MAX], pathname[PATH_MAX];
1027         struct stat st;
1028         int ret;
1029
1030         if (thread->pid_ == thread->tid) {
1031                 scnprintf(linkname, sizeof(linkname),
1032                           "/proc/%d/fd/%d", thread->pid_, fd);
1033         } else {
1034                 scnprintf(linkname, sizeof(linkname),
1035                           "/proc/%d/task/%d/fd/%d", thread->pid_, thread->tid, fd);
1036         }
1037
1038         if (lstat(linkname, &st) < 0 || st.st_size + 1 > (off_t)sizeof(pathname))
1039                 return -1;
1040
1041         ret = readlink(linkname, pathname, sizeof(pathname));
1042
1043         if (ret < 0 || ret > st.st_size)
1044                 return -1;
1045
1046         pathname[ret] = '\0';
1047         return trace__set_fd_pathname(thread, fd, pathname);
1048 }
1049
1050 static const char *thread__fd_path(struct thread *thread, int fd,
1051                                    struct trace *trace)
1052 {
1053         struct thread_trace *ttrace = thread__priv(thread);
1054
1055         if (ttrace == NULL)
1056                 return NULL;
1057
1058         if (fd < 0)
1059                 return NULL;
1060
1061         if ((fd > ttrace->paths.max || ttrace->paths.table[fd] == NULL)) {
1062                 if (!trace->live)
1063                         return NULL;
1064                 ++trace->stats.proc_getname;
1065                 if (thread__read_fd_path(thread, fd))
1066                         return NULL;
1067         }
1068
1069         return ttrace->paths.table[fd];
1070 }
1071
1072 size_t syscall_arg__scnprintf_fd(char *bf, size_t size, struct syscall_arg *arg)
1073 {
1074         int fd = arg->val;
1075         size_t printed = scnprintf(bf, size, "%d", fd);
1076         const char *path = thread__fd_path(arg->thread, fd, arg->trace);
1077
1078         if (path)
1079                 printed += scnprintf(bf + printed, size - printed, "<%s>", path);
1080
1081         return printed;
1082 }
1083
1084 size_t pid__scnprintf_fd(struct trace *trace, pid_t pid, int fd, char *bf, size_t size)
1085 {
1086         size_t printed = scnprintf(bf, size, "%d", fd);
1087         struct thread *thread = machine__find_thread(trace->host, pid, pid);
1088
1089         if (thread) {
1090                 const char *path = thread__fd_path(thread, fd, trace);
1091
1092                 if (path)
1093                         printed += scnprintf(bf + printed, size - printed, "<%s>", path);
1094
1095                 thread__put(thread);
1096         }
1097
1098         return printed;
1099 }
1100
1101 static size_t syscall_arg__scnprintf_close_fd(char *bf, size_t size,
1102                                               struct syscall_arg *arg)
1103 {
1104         int fd = arg->val;
1105         size_t printed = syscall_arg__scnprintf_fd(bf, size, arg);
1106         struct thread_trace *ttrace = thread__priv(arg->thread);
1107
1108         if (ttrace && fd >= 0 && fd <= ttrace->paths.max)
1109                 zfree(&ttrace->paths.table[fd]);
1110
1111         return printed;
1112 }
1113
1114 static void thread__set_filename_pos(struct thread *thread, const char *bf,
1115                                      unsigned long ptr)
1116 {
1117         struct thread_trace *ttrace = thread__priv(thread);
1118
1119         ttrace->filename.ptr = ptr;
1120         ttrace->filename.entry_str_pos = bf - ttrace->entry_str;
1121 }
1122
1123 static size_t syscall_arg__scnprintf_augmented_string(struct syscall_arg *arg, char *bf, size_t size)
1124 {
1125         struct augmented_arg *augmented_arg = arg->augmented.args;
1126
1127         return scnprintf(bf, size, "%.*s", augmented_arg->size, augmented_arg->value);
1128 }
1129
1130 static size_t syscall_arg__scnprintf_filename(char *bf, size_t size,
1131                                               struct syscall_arg *arg)
1132 {
1133         unsigned long ptr = arg->val;
1134
1135         if (arg->augmented.args)
1136                 return syscall_arg__scnprintf_augmented_string(arg, bf, size);
1137
1138         if (!arg->trace->vfs_getname)
1139                 return scnprintf(bf, size, "%#x", ptr);
1140
1141         thread__set_filename_pos(arg->thread, bf, ptr);
1142         return 0;
1143 }
1144
1145 static bool trace__filter_duration(struct trace *trace, double t)
1146 {
1147         return t < (trace->duration_filter * NSEC_PER_MSEC);
1148 }
1149
1150 static size_t __trace__fprintf_tstamp(struct trace *trace, u64 tstamp, FILE *fp)
1151 {
1152         double ts = (double)(tstamp - trace->base_time) / NSEC_PER_MSEC;
1153
1154         return fprintf(fp, "%10.3f ", ts);
1155 }
1156
1157 /*
1158  * We're handling tstamp=0 as an undefined tstamp, i.e. like when we are
1159  * using ttrace->entry_time for a thread that receives a sys_exit without
1160  * first having received a sys_enter ("poll" issued before tracing session
1161  * starts, lost sys_enter exit due to ring buffer overflow).
1162  */
1163 static size_t trace__fprintf_tstamp(struct trace *trace, u64 tstamp, FILE *fp)
1164 {
1165         if (tstamp > 0)
1166                 return __trace__fprintf_tstamp(trace, tstamp, fp);
1167
1168         return fprintf(fp, "         ? ");
1169 }
1170
1171 static bool done = false;
1172 static bool interrupted = false;
1173
1174 static void sig_handler(int sig)
1175 {
1176         done = true;
1177         interrupted = sig == SIGINT;
1178 }
1179
1180 static size_t trace__fprintf_comm_tid(struct trace *trace, struct thread *thread, FILE *fp)
1181 {
1182         size_t printed = 0;
1183
1184         if (trace->multiple_threads) {
1185                 if (trace->show_comm)
1186                         printed += fprintf(fp, "%.14s/", thread__comm_str(thread));
1187                 printed += fprintf(fp, "%d ", thread->tid);
1188         }
1189
1190         return printed;
1191 }
1192
1193 static size_t trace__fprintf_entry_head(struct trace *trace, struct thread *thread,
1194                                         u64 duration, bool duration_calculated, u64 tstamp, FILE *fp)
1195 {
1196         size_t printed = trace__fprintf_tstamp(trace, tstamp, fp);
1197         printed += fprintf_duration(duration, duration_calculated, fp);
1198         return printed + trace__fprintf_comm_tid(trace, thread, fp);
1199 }
1200
1201 static int trace__process_event(struct trace *trace, struct machine *machine,
1202                                 union perf_event *event, struct perf_sample *sample)
1203 {
1204         int ret = 0;
1205
1206         switch (event->header.type) {
1207         case PERF_RECORD_LOST:
1208                 color_fprintf(trace->output, PERF_COLOR_RED,
1209                               "LOST %" PRIu64 " events!\n", event->lost.lost);
1210                 ret = machine__process_lost_event(machine, event, sample);
1211                 break;
1212         default:
1213                 ret = machine__process_event(machine, event, sample);
1214                 break;
1215         }
1216
1217         return ret;
1218 }
1219
1220 static int trace__tool_process(struct perf_tool *tool,
1221                                union perf_event *event,
1222                                struct perf_sample *sample,
1223                                struct machine *machine)
1224 {
1225         struct trace *trace = container_of(tool, struct trace, tool);
1226         return trace__process_event(trace, machine, event, sample);
1227 }
1228
1229 static char *trace__machine__resolve_kernel_addr(void *vmachine, unsigned long long *addrp, char **modp)
1230 {
1231         struct machine *machine = vmachine;
1232
1233         if (machine->kptr_restrict_warned)
1234                 return NULL;
1235
1236         if (symbol_conf.kptr_restrict) {
1237                 pr_warning("Kernel address maps (/proc/{kallsyms,modules}) are restricted.\n\n"
1238                            "Check /proc/sys/kernel/kptr_restrict.\n\n"
1239                            "Kernel samples will not be resolved.\n");
1240                 machine->kptr_restrict_warned = true;
1241                 return NULL;
1242         }
1243
1244         return machine__resolve_kernel_addr(vmachine, addrp, modp);
1245 }
1246
1247 static int trace__symbols_init(struct trace *trace, struct perf_evlist *evlist)
1248 {
1249         int err = symbol__init(NULL);
1250
1251         if (err)
1252                 return err;
1253
1254         trace->host = machine__new_host();
1255         if (trace->host == NULL)
1256                 return -ENOMEM;
1257
1258         err = trace_event__register_resolver(trace->host, trace__machine__resolve_kernel_addr);
1259         if (err < 0)
1260                 goto out;
1261
1262         err = __machine__synthesize_threads(trace->host, &trace->tool, &trace->opts.target,
1263                                             evlist->threads, trace__tool_process, false,
1264                                             trace->opts.proc_map_timeout, 1);
1265 out:
1266         if (err)
1267                 symbol__exit();
1268
1269         return err;
1270 }
1271
1272 static void trace__symbols__exit(struct trace *trace)
1273 {
1274         machine__exit(trace->host);
1275         trace->host = NULL;
1276
1277         symbol__exit();
1278 }
1279
1280 static int syscall__alloc_arg_fmts(struct syscall *sc, int nr_args)
1281 {
1282         int idx;
1283
1284         if (nr_args == 6 && sc->fmt && sc->fmt->nr_args != 0)
1285                 nr_args = sc->fmt->nr_args;
1286
1287         sc->arg_fmt = calloc(nr_args, sizeof(*sc->arg_fmt));
1288         if (sc->arg_fmt == NULL)
1289                 return -1;
1290
1291         for (idx = 0; idx < nr_args; ++idx) {
1292                 if (sc->fmt)
1293                         sc->arg_fmt[idx] = sc->fmt->arg[idx];
1294         }
1295
1296         sc->nr_args = nr_args;
1297         return 0;
1298 }
1299
1300 static int syscall__set_arg_fmts(struct syscall *sc)
1301 {
1302         struct tep_format_field *field, *last_field = NULL;
1303         int idx = 0, len;
1304
1305         for (field = sc->args; field; field = field->next, ++idx) {
1306                 last_field = field;
1307
1308                 if (sc->fmt && sc->fmt->arg[idx].scnprintf)
1309                         continue;
1310
1311                 if (strcmp(field->type, "const char *") == 0 &&
1312                          (strcmp(field->name, "filename") == 0 ||
1313                           strcmp(field->name, "path") == 0 ||
1314                           strcmp(field->name, "pathname") == 0))
1315                         sc->arg_fmt[idx].scnprintf = SCA_FILENAME;
1316                 else if (field->flags & TEP_FIELD_IS_POINTER)
1317                         sc->arg_fmt[idx].scnprintf = syscall_arg__scnprintf_hex;
1318                 else if (strcmp(field->type, "pid_t") == 0)
1319                         sc->arg_fmt[idx].scnprintf = SCA_PID;
1320                 else if (strcmp(field->type, "umode_t") == 0)
1321                         sc->arg_fmt[idx].scnprintf = SCA_MODE_T;
1322                 else if ((strcmp(field->type, "int") == 0 ||
1323                           strcmp(field->type, "unsigned int") == 0 ||
1324                           strcmp(field->type, "long") == 0) &&
1325                          (len = strlen(field->name)) >= 2 &&
1326                          strcmp(field->name + len - 2, "fd") == 0) {
1327                         /*
1328                          * /sys/kernel/tracing/events/syscalls/sys_enter*
1329                          * egrep 'field:.*fd;' .../format|sed -r 's/.*field:([a-z ]+) [a-z_]*fd.+/\1/g'|sort|uniq -c
1330                          * 65 int
1331                          * 23 unsigned int
1332                          * 7 unsigned long
1333                          */
1334                         sc->arg_fmt[idx].scnprintf = SCA_FD;
1335                 }
1336         }
1337
1338         if (last_field)
1339                 sc->args_size = last_field->offset + last_field->size;
1340
1341         return 0;
1342 }
1343
1344 static int trace__read_syscall_info(struct trace *trace, int id)
1345 {
1346         char tp_name[128];
1347         struct syscall *sc;
1348         const char *name = syscalltbl__name(trace->sctbl, id);
1349
1350         if (name == NULL)
1351                 return -1;
1352
1353         if (id > trace->syscalls.max) {
1354                 struct syscall *nsyscalls = realloc(trace->syscalls.table, (id + 1) * sizeof(*sc));
1355
1356                 if (nsyscalls == NULL)
1357                         return -1;
1358
1359                 if (trace->syscalls.max != -1) {
1360                         memset(nsyscalls + trace->syscalls.max + 1, 0,
1361                                (id - trace->syscalls.max) * sizeof(*sc));
1362                 } else {
1363                         memset(nsyscalls, 0, (id + 1) * sizeof(*sc));
1364                 }
1365
1366                 trace->syscalls.table = nsyscalls;
1367                 trace->syscalls.max   = id;
1368         }
1369
1370         sc = trace->syscalls.table + id;
1371         sc->name = name;
1372
1373         sc->fmt  = syscall_fmt__find(sc->name);
1374
1375         snprintf(tp_name, sizeof(tp_name), "sys_enter_%s", sc->name);
1376         sc->tp_format = trace_event__tp_format("syscalls", tp_name);
1377
1378         if (IS_ERR(sc->tp_format) && sc->fmt && sc->fmt->alias) {
1379                 snprintf(tp_name, sizeof(tp_name), "sys_enter_%s", sc->fmt->alias);
1380                 sc->tp_format = trace_event__tp_format("syscalls", tp_name);
1381         }
1382
1383         if (syscall__alloc_arg_fmts(sc, IS_ERR(sc->tp_format) ? 6 : sc->tp_format->format.nr_fields))
1384                 return -1;
1385
1386         if (IS_ERR(sc->tp_format))
1387                 return -1;
1388
1389         sc->args = sc->tp_format->format.fields;
1390         /*
1391          * We need to check and discard the first variable '__syscall_nr'
1392          * or 'nr' that mean the syscall number. It is needless here.
1393          * So drop '__syscall_nr' or 'nr' field but does not exist on older kernels.
1394          */
1395         if (sc->args && (!strcmp(sc->args->name, "__syscall_nr") || !strcmp(sc->args->name, "nr"))) {
1396                 sc->args = sc->args->next;
1397                 --sc->nr_args;
1398         }
1399
1400         sc->is_exit = !strcmp(name, "exit_group") || !strcmp(name, "exit");
1401         sc->is_open = !strcmp(name, "open") || !strcmp(name, "openat");
1402
1403         return syscall__set_arg_fmts(sc);
1404 }
1405
1406 static int trace__validate_ev_qualifier(struct trace *trace)
1407 {
1408         int err = 0, i;
1409         size_t nr_allocated;
1410         struct str_node *pos;
1411
1412         trace->ev_qualifier_ids.nr = strlist__nr_entries(trace->ev_qualifier);
1413         trace->ev_qualifier_ids.entries = malloc(trace->ev_qualifier_ids.nr *
1414                                                  sizeof(trace->ev_qualifier_ids.entries[0]));
1415
1416         if (trace->ev_qualifier_ids.entries == NULL) {
1417                 fputs("Error:\tNot enough memory for allocating events qualifier ids\n",
1418                        trace->output);
1419                 err = -EINVAL;
1420                 goto out;
1421         }
1422
1423         nr_allocated = trace->ev_qualifier_ids.nr;
1424         i = 0;
1425
1426         strlist__for_each_entry(pos, trace->ev_qualifier) {
1427                 const char *sc = pos->s;
1428                 int id = syscalltbl__id(trace->sctbl, sc), match_next = -1;
1429
1430                 if (id < 0) {
1431                         id = syscalltbl__strglobmatch_first(trace->sctbl, sc, &match_next);
1432                         if (id >= 0)
1433                                 goto matches;
1434
1435                         if (err == 0) {
1436                                 fputs("Error:\tInvalid syscall ", trace->output);
1437                                 err = -EINVAL;
1438                         } else {
1439                                 fputs(", ", trace->output);
1440                         }
1441
1442                         fputs(sc, trace->output);
1443                 }
1444 matches:
1445                 trace->ev_qualifier_ids.entries[i++] = id;
1446                 if (match_next == -1)
1447                         continue;
1448
1449                 while (1) {
1450                         id = syscalltbl__strglobmatch_next(trace->sctbl, sc, &match_next);
1451                         if (id < 0)
1452                                 break;
1453                         if (nr_allocated == trace->ev_qualifier_ids.nr) {
1454                                 void *entries;
1455
1456                                 nr_allocated += 8;
1457                                 entries = realloc(trace->ev_qualifier_ids.entries,
1458                                                   nr_allocated * sizeof(trace->ev_qualifier_ids.entries[0]));
1459                                 if (entries == NULL) {
1460                                         err = -ENOMEM;
1461                                         fputs("\nError:\t Not enough memory for parsing\n", trace->output);
1462                                         goto out_free;
1463                                 }
1464                                 trace->ev_qualifier_ids.entries = entries;
1465                         }
1466                         trace->ev_qualifier_ids.nr++;
1467                         trace->ev_qualifier_ids.entries[i++] = id;
1468                 }
1469         }
1470
1471         if (err < 0) {
1472                 fputs("\nHint:\ttry 'perf list syscalls:sys_enter_*'"
1473                       "\nHint:\tand: 'man syscalls'\n", trace->output);
1474 out_free:
1475                 zfree(&trace->ev_qualifier_ids.entries);
1476                 trace->ev_qualifier_ids.nr = 0;
1477         }
1478 out:
1479         return err;
1480 }
1481
1482 /*
1483  * args is to be interpreted as a series of longs but we need to handle
1484  * 8-byte unaligned accesses. args points to raw_data within the event
1485  * and raw_data is guaranteed to be 8-byte unaligned because it is
1486  * preceded by raw_size which is a u32. So we need to copy args to a temp
1487  * variable to read it. Most notably this avoids extended load instructions
1488  * on unaligned addresses
1489  */
1490 unsigned long syscall_arg__val(struct syscall_arg *arg, u8 idx)
1491 {
1492         unsigned long val;
1493         unsigned char *p = arg->args + sizeof(unsigned long) * idx;
1494
1495         memcpy(&val, p, sizeof(val));
1496         return val;
1497 }
1498
1499 static size_t syscall__scnprintf_name(struct syscall *sc, char *bf, size_t size,
1500                                       struct syscall_arg *arg)
1501 {
1502         if (sc->arg_fmt && sc->arg_fmt[arg->idx].name)
1503                 return scnprintf(bf, size, "%s: ", sc->arg_fmt[arg->idx].name);
1504
1505         return scnprintf(bf, size, "arg%d: ", arg->idx);
1506 }
1507
1508 /*
1509  * Check if the value is in fact zero, i.e. mask whatever needs masking, such
1510  * as mount 'flags' argument that needs ignoring some magic flag, see comment
1511  * in tools/perf/trace/beauty/mount_flags.c
1512  */
1513 static unsigned long syscall__mask_val(struct syscall *sc, struct syscall_arg *arg, unsigned long val)
1514 {
1515         if (sc->arg_fmt && sc->arg_fmt[arg->idx].mask_val)
1516                 return sc->arg_fmt[arg->idx].mask_val(arg, val);
1517
1518         return val;
1519 }
1520
1521 static size_t syscall__scnprintf_val(struct syscall *sc, char *bf, size_t size,
1522                                      struct syscall_arg *arg, unsigned long val)
1523 {
1524         if (sc->arg_fmt && sc->arg_fmt[arg->idx].scnprintf) {
1525                 arg->val = val;
1526                 if (sc->arg_fmt[arg->idx].parm)
1527                         arg->parm = sc->arg_fmt[arg->idx].parm;
1528                 return sc->arg_fmt[arg->idx].scnprintf(bf, size, arg);
1529         }
1530         return scnprintf(bf, size, "%ld", val);
1531 }
1532
1533 static size_t syscall__scnprintf_args(struct syscall *sc, char *bf, size_t size,
1534                                       unsigned char *args, void *augmented_args, int augmented_args_size,
1535                                       struct trace *trace, struct thread *thread)
1536 {
1537         size_t printed = 0;
1538         unsigned long val;
1539         u8 bit = 1;
1540         struct syscall_arg arg = {
1541                 .args   = args,
1542                 .augmented = {
1543                         .size = augmented_args_size,
1544                         .args = augmented_args,
1545                 },
1546                 .idx    = 0,
1547                 .mask   = 0,
1548                 .trace  = trace,
1549                 .thread = thread,
1550         };
1551         struct thread_trace *ttrace = thread__priv(thread);
1552
1553         /*
1554          * Things like fcntl will set this in its 'cmd' formatter to pick the
1555          * right formatter for the return value (an fd? file flags?), which is
1556          * not needed for syscalls that always return a given type, say an fd.
1557          */
1558         ttrace->ret_scnprintf = NULL;
1559
1560         if (sc->args != NULL) {
1561                 struct tep_format_field *field;
1562
1563                 for (field = sc->args; field;
1564                      field = field->next, ++arg.idx, bit <<= 1) {
1565                         if (arg.mask & bit)
1566                                 continue;
1567
1568                         val = syscall_arg__val(&arg, arg.idx);
1569                         /*
1570                          * Some syscall args need some mask, most don't and
1571                          * return val untouched.
1572                          */
1573                         val = syscall__mask_val(sc, &arg, val);
1574
1575                         /*
1576                          * Suppress this argument if its value is zero and
1577                          * and we don't have a string associated in an
1578                          * strarray for it.
1579                          */
1580                         if (val == 0 &&
1581                             !(sc->arg_fmt &&
1582                               (sc->arg_fmt[arg.idx].show_zero ||
1583                                sc->arg_fmt[arg.idx].scnprintf == SCA_STRARRAY ||
1584                                sc->arg_fmt[arg.idx].scnprintf == SCA_STRARRAYS) &&
1585                               sc->arg_fmt[arg.idx].parm))
1586                                 continue;
1587
1588                         printed += scnprintf(bf + printed, size - printed,
1589                                              "%s%s: ", printed ? ", " : "", field->name);
1590                         printed += syscall__scnprintf_val(sc, bf + printed, size - printed, &arg, val);
1591                 }
1592         } else if (IS_ERR(sc->tp_format)) {
1593                 /*
1594                  * If we managed to read the tracepoint /format file, then we
1595                  * may end up not having any args, like with gettid(), so only
1596                  * print the raw args when we didn't manage to read it.
1597                  */
1598                 while (arg.idx < sc->nr_args) {
1599                         if (arg.mask & bit)
1600                                 goto next_arg;
1601                         val = syscall_arg__val(&arg, arg.idx);
1602                         if (printed)
1603                                 printed += scnprintf(bf + printed, size - printed, ", ");
1604                         printed += syscall__scnprintf_name(sc, bf + printed, size - printed, &arg);
1605                         printed += syscall__scnprintf_val(sc, bf + printed, size - printed, &arg, val);
1606 next_arg:
1607                         ++arg.idx;
1608                         bit <<= 1;
1609                 }
1610         }
1611
1612         return printed;
1613 }
1614
1615 typedef int (*tracepoint_handler)(struct trace *trace, struct perf_evsel *evsel,
1616                                   union perf_event *event,
1617                                   struct perf_sample *sample);
1618
1619 static struct syscall *trace__syscall_info(struct trace *trace,
1620                                            struct perf_evsel *evsel, int id)
1621 {
1622
1623         if (id < 0) {
1624
1625                 /*
1626                  * XXX: Noticed on x86_64, reproduced as far back as 3.0.36, haven't tried
1627                  * before that, leaving at a higher verbosity level till that is
1628                  * explained. Reproduced with plain ftrace with:
1629                  *
1630                  * echo 1 > /t/events/raw_syscalls/sys_exit/enable
1631                  * grep "NR -1 " /t/trace_pipe
1632                  *
1633                  * After generating some load on the machine.
1634                  */
1635                 if (verbose > 1) {
1636                         static u64 n;
1637                         fprintf(trace->output, "Invalid syscall %d id, skipping (%s, %" PRIu64 ") ...\n",
1638                                 id, perf_evsel__name(evsel), ++n);
1639                 }
1640                 return NULL;
1641         }
1642
1643         if ((id > trace->syscalls.max || trace->syscalls.table[id].name == NULL) &&
1644             trace__read_syscall_info(trace, id))
1645                 goto out_cant_read;
1646
1647         if ((id > trace->syscalls.max || trace->syscalls.table[id].name == NULL))
1648                 goto out_cant_read;
1649
1650         return &trace->syscalls.table[id];
1651
1652 out_cant_read:
1653         if (verbose > 0) {
1654                 fprintf(trace->output, "Problems reading syscall %d", id);
1655                 if (id <= trace->syscalls.max && trace->syscalls.table[id].name != NULL)
1656                         fprintf(trace->output, "(%s)", trace->syscalls.table[id].name);
1657                 fputs(" information\n", trace->output);
1658         }
1659         return NULL;
1660 }
1661
1662 static void thread__update_stats(struct thread_trace *ttrace,
1663                                  int id, struct perf_sample *sample)
1664 {
1665         struct int_node *inode;
1666         struct stats *stats;
1667         u64 duration = 0;
1668
1669         inode = intlist__findnew(ttrace->syscall_stats, id);
1670         if (inode == NULL)
1671                 return;
1672
1673         stats = inode->priv;
1674         if (stats == NULL) {
1675                 stats = malloc(sizeof(struct stats));
1676                 if (stats == NULL)
1677                         return;
1678                 init_stats(stats);
1679                 inode->priv = stats;
1680         }
1681
1682         if (ttrace->entry_time && sample->time > ttrace->entry_time)
1683                 duration = sample->time - ttrace->entry_time;
1684
1685         update_stats(stats, duration);
1686 }
1687
1688 static int trace__printf_interrupted_entry(struct trace *trace)
1689 {
1690         struct thread_trace *ttrace;
1691         size_t printed;
1692
1693         if (trace->failure_only || trace->current == NULL)
1694                 return 0;
1695
1696         ttrace = thread__priv(trace->current);
1697
1698         if (!ttrace->entry_pending)
1699                 return 0;
1700
1701         printed  = trace__fprintf_entry_head(trace, trace->current, 0, false, ttrace->entry_time, trace->output);
1702         printed += fprintf(trace->output, "%-70s) ...\n", ttrace->entry_str);
1703         ttrace->entry_pending = false;
1704
1705         ++trace->nr_events_printed;
1706
1707         return printed;
1708 }
1709
1710 static int trace__fprintf_sample(struct trace *trace, struct perf_evsel *evsel,
1711                                  struct perf_sample *sample, struct thread *thread)
1712 {
1713         int printed = 0;
1714
1715         if (trace->print_sample) {
1716                 double ts = (double)sample->time / NSEC_PER_MSEC;
1717
1718                 printed += fprintf(trace->output, "%22s %10.3f %s %d/%d [%d]\n",
1719                                    perf_evsel__name(evsel), ts,
1720                                    thread__comm_str(thread),
1721                                    sample->pid, sample->tid, sample->cpu);
1722         }
1723
1724         return printed;
1725 }
1726
1727 static void *syscall__augmented_args(struct syscall *sc, struct perf_sample *sample, int *augmented_args_size)
1728 {
1729         void *augmented_args = NULL;
1730
1731         *augmented_args_size = sample->raw_size - sc->args_size;
1732         if (*augmented_args_size > 0)
1733                 augmented_args = sample->raw_data + sc->args_size;
1734
1735         return augmented_args;
1736 }
1737
1738 static int trace__sys_enter(struct trace *trace, struct perf_evsel *evsel,
1739                             union perf_event *event __maybe_unused,
1740                             struct perf_sample *sample)
1741 {
1742         char *msg;
1743         void *args;
1744         size_t printed = 0;
1745         struct thread *thread;
1746         int id = perf_evsel__sc_tp_uint(evsel, id, sample), err = -1;
1747         int augmented_args_size = 0;
1748         void *augmented_args = NULL;
1749         struct syscall *sc = trace__syscall_info(trace, evsel, id);
1750         struct thread_trace *ttrace;
1751
1752         if (sc == NULL)
1753                 return -1;
1754
1755         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
1756         ttrace = thread__trace(thread, trace->output);
1757         if (ttrace == NULL)
1758                 goto out_put;
1759
1760         trace__fprintf_sample(trace, evsel, sample, thread);
1761
1762         args = perf_evsel__sc_tp_ptr(evsel, args, sample);
1763
1764         if (ttrace->entry_str == NULL) {
1765                 ttrace->entry_str = malloc(trace__entry_str_size);
1766                 if (!ttrace->entry_str)
1767                         goto out_put;
1768         }
1769
1770         if (!(trace->duration_filter || trace->summary_only || trace->min_stack))
1771                 trace__printf_interrupted_entry(trace);
1772         /*
1773          * If this is raw_syscalls.sys_enter, then it always comes with the 6 possible
1774          * arguments, even if the syscall being handled, say "openat", uses only 4 arguments
1775          * this breaks syscall__augmented_args() check for augmented args, as we calculate
1776          * syscall->args_size using each syscalls:sys_enter_NAME tracefs format file,
1777          * so when handling, say the openat syscall, we end up getting 6 args for the
1778          * raw_syscalls:sys_enter event, when we expected just 4, we end up mistakenly
1779          * thinking that the extra 2 u64 args are the augmented filename, so just check
1780          * here and avoid using augmented syscalls when the evsel is the raw_syscalls one.
1781          */
1782         if (evsel != trace->syscalls.events.sys_enter)
1783                 augmented_args = syscall__augmented_args(sc, sample, &augmented_args_size);
1784         ttrace->entry_time = sample->time;
1785         msg = ttrace->entry_str;
1786         printed += scnprintf(msg + printed, trace__entry_str_size - printed, "%s(", sc->name);
1787
1788         printed += syscall__scnprintf_args(sc, msg + printed, trace__entry_str_size - printed,
1789                                            args, augmented_args, augmented_args_size, trace, thread);
1790
1791         if (sc->is_exit) {
1792                 if (!(trace->duration_filter || trace->summary_only || trace->failure_only || trace->min_stack)) {
1793                         trace__fprintf_entry_head(trace, thread, 0, false, ttrace->entry_time, trace->output);
1794                         fprintf(trace->output, "%-70s)\n", ttrace->entry_str);
1795                 }
1796         } else {
1797                 ttrace->entry_pending = true;
1798                 /* See trace__vfs_getname & trace__sys_exit */
1799                 ttrace->filename.pending_open = false;
1800         }
1801
1802         if (trace->current != thread) {
1803                 thread__put(trace->current);
1804                 trace->current = thread__get(thread);
1805         }
1806         err = 0;
1807 out_put:
1808         thread__put(thread);
1809         return err;
1810 }
1811
1812 static int trace__fprintf_sys_enter(struct trace *trace, struct perf_evsel *evsel,
1813                                     struct perf_sample *sample)
1814 {
1815         struct thread_trace *ttrace;
1816         struct thread *thread;
1817         int id = perf_evsel__sc_tp_uint(evsel, id, sample), err = -1;
1818         struct syscall *sc = trace__syscall_info(trace, evsel, id);
1819         char msg[1024];
1820         void *args, *augmented_args = NULL;
1821         int augmented_args_size;
1822
1823         if (sc == NULL)
1824                 return -1;
1825
1826         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
1827         ttrace = thread__trace(thread, trace->output);
1828         /*
1829          * We need to get ttrace just to make sure it is there when syscall__scnprintf_args()
1830          * and the rest of the beautifiers accessing it via struct syscall_arg touches it.
1831          */
1832         if (ttrace == NULL)
1833                 goto out_put;
1834
1835         args = perf_evsel__sc_tp_ptr(evsel, args, sample);
1836         augmented_args = syscall__augmented_args(sc, sample, &augmented_args_size);
1837         syscall__scnprintf_args(sc, msg, sizeof(msg), args, augmented_args, augmented_args_size, trace, thread);
1838         fprintf(trace->output, "%s", msg);
1839         err = 0;
1840 out_put:
1841         thread__put(thread);
1842         return err;
1843 }
1844
1845 static int trace__resolve_callchain(struct trace *trace, struct perf_evsel *evsel,
1846                                     struct perf_sample *sample,
1847                                     struct callchain_cursor *cursor)
1848 {
1849         struct addr_location al;
1850         int max_stack = evsel->attr.sample_max_stack ?
1851                         evsel->attr.sample_max_stack :
1852                         trace->max_stack;
1853         int err;
1854
1855         if (machine__resolve(trace->host, &al, sample) < 0)
1856                 return -1;
1857
1858         err = thread__resolve_callchain(al.thread, cursor, evsel, sample, NULL, NULL, max_stack);
1859         addr_location__put(&al);
1860         return err;
1861 }
1862
1863 static int trace__fprintf_callchain(struct trace *trace, struct perf_sample *sample)
1864 {
1865         /* TODO: user-configurable print_opts */
1866         const unsigned int print_opts = EVSEL__PRINT_SYM |
1867                                         EVSEL__PRINT_DSO |
1868                                         EVSEL__PRINT_UNKNOWN_AS_ADDR;
1869
1870         return sample__fprintf_callchain(sample, 38, print_opts, &callchain_cursor, trace->output);
1871 }
1872
1873 static const char *errno_to_name(struct perf_evsel *evsel, int err)
1874 {
1875         struct perf_env *env = perf_evsel__env(evsel);
1876         const char *arch_name = perf_env__arch(env);
1877
1878         return arch_syscalls__strerrno(arch_name, err);
1879 }
1880
1881 static int trace__sys_exit(struct trace *trace, struct perf_evsel *evsel,
1882                            union perf_event *event __maybe_unused,
1883                            struct perf_sample *sample)
1884 {
1885         long ret;
1886         u64 duration = 0;
1887         bool duration_calculated = false;
1888         struct thread *thread;
1889         int id = perf_evsel__sc_tp_uint(evsel, id, sample), err = -1, callchain_ret = 0;
1890         struct syscall *sc = trace__syscall_info(trace, evsel, id);
1891         struct thread_trace *ttrace;
1892
1893         if (sc == NULL)
1894                 return -1;
1895
1896         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
1897         ttrace = thread__trace(thread, trace->output);
1898         if (ttrace == NULL)
1899                 goto out_put;
1900
1901         trace__fprintf_sample(trace, evsel, sample, thread);
1902
1903         if (trace->summary)
1904                 thread__update_stats(ttrace, id, sample);
1905
1906         ret = perf_evsel__sc_tp_uint(evsel, ret, sample);
1907
1908         if (sc->is_open && ret >= 0 && ttrace->filename.pending_open) {
1909                 trace__set_fd_pathname(thread, ret, ttrace->filename.name);
1910                 ttrace->filename.pending_open = false;
1911                 ++trace->stats.vfs_getname;
1912         }
1913
1914         if (ttrace->entry_time) {
1915                 duration = sample->time - ttrace->entry_time;
1916                 if (trace__filter_duration(trace, duration))
1917                         goto out;
1918                 duration_calculated = true;
1919         } else if (trace->duration_filter)
1920                 goto out;
1921
1922         if (sample->callchain) {
1923                 callchain_ret = trace__resolve_callchain(trace, evsel, sample, &callchain_cursor);
1924                 if (callchain_ret == 0) {
1925                         if (callchain_cursor.nr < trace->min_stack)
1926                                 goto out;
1927                         callchain_ret = 1;
1928                 }
1929         }
1930
1931         if (trace->summary_only || (ret >= 0 && trace->failure_only))
1932                 goto out;
1933
1934         trace__fprintf_entry_head(trace, thread, duration, duration_calculated, ttrace->entry_time, trace->output);
1935
1936         if (ttrace->entry_pending) {
1937                 fprintf(trace->output, "%-70s", ttrace->entry_str);
1938         } else {
1939                 fprintf(trace->output, " ... [");
1940                 color_fprintf(trace->output, PERF_COLOR_YELLOW, "continued");
1941                 fprintf(trace->output, "]: %s()", sc->name);
1942         }
1943
1944         if (sc->fmt == NULL) {
1945                 if (ret < 0)
1946                         goto errno_print;
1947 signed_print:
1948                 fprintf(trace->output, ") = %ld", ret);
1949         } else if (ret < 0) {
1950 errno_print: {
1951                 char bf[STRERR_BUFSIZE];
1952                 const char *emsg = str_error_r(-ret, bf, sizeof(bf)),
1953                            *e = errno_to_name(evsel, -ret);
1954
1955                 fprintf(trace->output, ") = -1 %s %s", e, emsg);
1956         }
1957         } else if (ret == 0 && sc->fmt->timeout)
1958                 fprintf(trace->output, ") = 0 Timeout");
1959         else if (ttrace->ret_scnprintf) {
1960                 char bf[1024];
1961                 struct syscall_arg arg = {
1962                         .val    = ret,
1963                         .thread = thread,
1964                         .trace  = trace,
1965                 };
1966                 ttrace->ret_scnprintf(bf, sizeof(bf), &arg);
1967                 ttrace->ret_scnprintf = NULL;
1968                 fprintf(trace->output, ") = %s", bf);
1969         } else if (sc->fmt->hexret)
1970                 fprintf(trace->output, ") = %#lx", ret);
1971         else if (sc->fmt->errpid) {
1972                 struct thread *child = machine__find_thread(trace->host, ret, ret);
1973
1974                 if (child != NULL) {
1975                         fprintf(trace->output, ") = %ld", ret);
1976                         if (child->comm_set)
1977                                 fprintf(trace->output, " (%s)", thread__comm_str(child));
1978                         thread__put(child);
1979                 }
1980         } else
1981                 goto signed_print;
1982
1983         fputc('\n', trace->output);
1984
1985         /*
1986          * We only consider an 'event' for the sake of --max-events a non-filtered
1987          * sys_enter + sys_exit and other tracepoint events.
1988          */
1989         if (++trace->nr_events_printed == trace->max_events && trace->max_events != ULONG_MAX)
1990                 interrupted = true;
1991
1992         if (callchain_ret > 0)
1993                 trace__fprintf_callchain(trace, sample);
1994         else if (callchain_ret < 0)
1995                 pr_err("Problem processing %s callchain, skipping...\n", perf_evsel__name(evsel));
1996 out:
1997         ttrace->entry_pending = false;
1998         err = 0;
1999 out_put:
2000         thread__put(thread);
2001         return err;
2002 }
2003
2004 static int trace__vfs_getname(struct trace *trace, struct perf_evsel *evsel,
2005                               union perf_event *event __maybe_unused,
2006                               struct perf_sample *sample)
2007 {
2008         struct thread *thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2009         struct thread_trace *ttrace;
2010         size_t filename_len, entry_str_len, to_move;
2011         ssize_t remaining_space;
2012         char *pos;
2013         const char *filename = perf_evsel__rawptr(evsel, sample, "pathname");
2014
2015         if (!thread)
2016                 goto out;
2017
2018         ttrace = thread__priv(thread);
2019         if (!ttrace)
2020                 goto out_put;
2021
2022         filename_len = strlen(filename);
2023         if (filename_len == 0)
2024                 goto out_put;
2025
2026         if (ttrace->filename.namelen < filename_len) {
2027                 char *f = realloc(ttrace->filename.name, filename_len + 1);
2028
2029                 if (f == NULL)
2030                         goto out_put;
2031
2032                 ttrace->filename.namelen = filename_len;
2033                 ttrace->filename.name = f;
2034         }
2035
2036         strcpy(ttrace->filename.name, filename);
2037         ttrace->filename.pending_open = true;
2038
2039         if (!ttrace->filename.ptr)
2040                 goto out_put;
2041
2042         entry_str_len = strlen(ttrace->entry_str);
2043         remaining_space = trace__entry_str_size - entry_str_len - 1; /* \0 */
2044         if (remaining_space <= 0)
2045                 goto out_put;
2046
2047         if (filename_len > (size_t)remaining_space) {
2048                 filename += filename_len - remaining_space;
2049                 filename_len = remaining_space;
2050         }
2051
2052         to_move = entry_str_len - ttrace->filename.entry_str_pos + 1; /* \0 */
2053         pos = ttrace->entry_str + ttrace->filename.entry_str_pos;
2054         memmove(pos + filename_len, pos, to_move);
2055         memcpy(pos, filename, filename_len);
2056
2057         ttrace->filename.ptr = 0;
2058         ttrace->filename.entry_str_pos = 0;
2059 out_put:
2060         thread__put(thread);
2061 out:
2062         return 0;
2063 }
2064
2065 static int trace__sched_stat_runtime(struct trace *trace, struct perf_evsel *evsel,
2066                                      union perf_event *event __maybe_unused,
2067                                      struct perf_sample *sample)
2068 {
2069         u64 runtime = perf_evsel__intval(evsel, sample, "runtime");
2070         double runtime_ms = (double)runtime / NSEC_PER_MSEC;
2071         struct thread *thread = machine__findnew_thread(trace->host,
2072                                                         sample->pid,
2073                                                         sample->tid);
2074         struct thread_trace *ttrace = thread__trace(thread, trace->output);
2075
2076         if (ttrace == NULL)
2077                 goto out_dump;
2078
2079         ttrace->runtime_ms += runtime_ms;
2080         trace->runtime_ms += runtime_ms;
2081 out_put:
2082         thread__put(thread);
2083         return 0;
2084
2085 out_dump:
2086         fprintf(trace->output, "%s: comm=%s,pid=%u,runtime=%" PRIu64 ",vruntime=%" PRIu64 ")\n",
2087                evsel->name,
2088                perf_evsel__strval(evsel, sample, "comm"),
2089                (pid_t)perf_evsel__intval(evsel, sample, "pid"),
2090                runtime,
2091                perf_evsel__intval(evsel, sample, "vruntime"));
2092         goto out_put;
2093 }
2094
2095 static int bpf_output__printer(enum binary_printer_ops op,
2096                                unsigned int val, void *extra __maybe_unused, FILE *fp)
2097 {
2098         unsigned char ch = (unsigned char)val;
2099
2100         switch (op) {
2101         case BINARY_PRINT_CHAR_DATA:
2102                 return fprintf(fp, "%c", isprint(ch) ? ch : '.');
2103         case BINARY_PRINT_DATA_BEGIN:
2104         case BINARY_PRINT_LINE_BEGIN:
2105         case BINARY_PRINT_ADDR:
2106         case BINARY_PRINT_NUM_DATA:
2107         case BINARY_PRINT_NUM_PAD:
2108         case BINARY_PRINT_SEP:
2109         case BINARY_PRINT_CHAR_PAD:
2110         case BINARY_PRINT_LINE_END:
2111         case BINARY_PRINT_DATA_END:
2112         default:
2113                 break;
2114         }
2115
2116         return 0;
2117 }
2118
2119 static void bpf_output__fprintf(struct trace *trace,
2120                                 struct perf_sample *sample)
2121 {
2122         binary__fprintf(sample->raw_data, sample->raw_size, 8,
2123                         bpf_output__printer, NULL, trace->output);
2124         ++trace->nr_events_printed;
2125 }
2126
2127 static int trace__event_handler(struct trace *trace, struct perf_evsel *evsel,
2128                                 union perf_event *event __maybe_unused,
2129                                 struct perf_sample *sample)
2130 {
2131         struct thread *thread;
2132         int callchain_ret = 0;
2133         /*
2134          * Check if we called perf_evsel__disable(evsel) due to, for instance,
2135          * this event's max_events having been hit and this is an entry coming
2136          * from the ring buffer that we should discard, since the max events
2137          * have already been considered/printed.
2138          */
2139         if (evsel->disabled)
2140                 return 0;
2141
2142         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2143
2144         if (sample->callchain) {
2145                 callchain_ret = trace__resolve_callchain(trace, evsel, sample, &callchain_cursor);
2146                 if (callchain_ret == 0) {
2147                         if (callchain_cursor.nr < trace->min_stack)
2148                                 goto out;
2149                         callchain_ret = 1;
2150                 }
2151         }
2152
2153         trace__printf_interrupted_entry(trace);
2154         trace__fprintf_tstamp(trace, sample->time, trace->output);
2155
2156         if (trace->trace_syscalls)
2157                 fprintf(trace->output, "(         ): ");
2158
2159         if (thread)
2160                 trace__fprintf_comm_tid(trace, thread, trace->output);
2161
2162         if (evsel == trace->syscalls.events.augmented) {
2163                 int id = perf_evsel__sc_tp_uint(evsel, id, sample);
2164                 struct syscall *sc = trace__syscall_info(trace, evsel, id);
2165
2166                 if (sc) {
2167                         fprintf(trace->output, "%s(", sc->name);
2168                         trace__fprintf_sys_enter(trace, evsel, sample);
2169                         fputc(')', trace->output);
2170                         goto newline;
2171                 }
2172
2173                 /*
2174                  * XXX: Not having the associated syscall info or not finding/adding
2175                  *      the thread should never happen, but if it does...
2176                  *      fall thru and print it as a bpf_output event.
2177                  */
2178         }
2179
2180         fprintf(trace->output, "%s:", evsel->name);
2181
2182         if (perf_evsel__is_bpf_output(evsel)) {
2183                 bpf_output__fprintf(trace, sample);
2184         } else if (evsel->tp_format) {
2185                 if (strncmp(evsel->tp_format->name, "sys_enter_", 10) ||
2186                     trace__fprintf_sys_enter(trace, evsel, sample)) {
2187                         event_format__fprintf(evsel->tp_format, sample->cpu,
2188                                               sample->raw_data, sample->raw_size,
2189                                               trace->output);
2190                         ++trace->nr_events_printed;
2191
2192                         if (evsel->max_events != ULONG_MAX && ++evsel->nr_events_printed == evsel->max_events) {
2193                                 perf_evsel__disable(evsel);
2194                                 perf_evsel__close(evsel);
2195                         }
2196                 }
2197         }
2198
2199 newline:
2200         fprintf(trace->output, "\n");
2201
2202         if (callchain_ret > 0)
2203                 trace__fprintf_callchain(trace, sample);
2204         else if (callchain_ret < 0)
2205                 pr_err("Problem processing %s callchain, skipping...\n", perf_evsel__name(evsel));
2206 out:
2207         thread__put(thread);
2208         return 0;
2209 }
2210
2211 static void print_location(FILE *f, struct perf_sample *sample,
2212                            struct addr_location *al,
2213                            bool print_dso, bool print_sym)
2214 {
2215
2216         if ((verbose > 0 || print_dso) && al->map)
2217                 fprintf(f, "%s@", al->map->dso->long_name);
2218
2219         if ((verbose > 0 || print_sym) && al->sym)
2220                 fprintf(f, "%s+0x%" PRIx64, al->sym->name,
2221                         al->addr - al->sym->start);
2222         else if (al->map)
2223                 fprintf(f, "0x%" PRIx64, al->addr);
2224         else
2225                 fprintf(f, "0x%" PRIx64, sample->addr);
2226 }
2227
2228 static int trace__pgfault(struct trace *trace,
2229                           struct perf_evsel *evsel,
2230                           union perf_event *event __maybe_unused,
2231                           struct perf_sample *sample)
2232 {
2233         struct thread *thread;
2234         struct addr_location al;
2235         char map_type = 'd';
2236         struct thread_trace *ttrace;
2237         int err = -1;
2238         int callchain_ret = 0;
2239
2240         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2241
2242         if (sample->callchain) {
2243                 callchain_ret = trace__resolve_callchain(trace, evsel, sample, &callchain_cursor);
2244                 if (callchain_ret == 0) {
2245                         if (callchain_cursor.nr < trace->min_stack)
2246                                 goto out_put;
2247                         callchain_ret = 1;
2248                 }
2249         }
2250
2251         ttrace = thread__trace(thread, trace->output);
2252         if (ttrace == NULL)
2253                 goto out_put;
2254
2255         if (evsel->attr.config == PERF_COUNT_SW_PAGE_FAULTS_MAJ)
2256                 ttrace->pfmaj++;
2257         else
2258                 ttrace->pfmin++;
2259
2260         if (trace->summary_only)
2261                 goto out;
2262
2263         thread__find_symbol(thread, sample->cpumode, sample->ip, &al);
2264
2265         trace__fprintf_entry_head(trace, thread, 0, true, sample->time, trace->output);
2266
2267         fprintf(trace->output, "%sfault [",
2268                 evsel->attr.config == PERF_COUNT_SW_PAGE_FAULTS_MAJ ?
2269                 "maj" : "min");
2270
2271         print_location(trace->output, sample, &al, false, true);
2272
2273         fprintf(trace->output, "] => ");
2274
2275         thread__find_symbol(thread, sample->cpumode, sample->addr, &al);
2276
2277         if (!al.map) {
2278                 thread__find_symbol(thread, sample->cpumode, sample->addr, &al);
2279
2280                 if (al.map)
2281                         map_type = 'x';
2282                 else
2283                         map_type = '?';
2284         }
2285
2286         print_location(trace->output, sample, &al, true, false);
2287
2288         fprintf(trace->output, " (%c%c)\n", map_type, al.level);
2289
2290         if (callchain_ret > 0)
2291                 trace__fprintf_callchain(trace, sample);
2292         else if (callchain_ret < 0)
2293                 pr_err("Problem processing %s callchain, skipping...\n", perf_evsel__name(evsel));
2294
2295         ++trace->nr_events_printed;
2296 out:
2297         err = 0;
2298 out_put:
2299         thread__put(thread);
2300         return err;
2301 }
2302
2303 static void trace__set_base_time(struct trace *trace,
2304                                  struct perf_evsel *evsel,
2305                                  struct perf_sample *sample)
2306 {
2307         /*
2308          * BPF events were not setting PERF_SAMPLE_TIME, so be more robust
2309          * and don't use sample->time unconditionally, we may end up having
2310          * some other event in the future without PERF_SAMPLE_TIME for good
2311          * reason, i.e. we may not be interested in its timestamps, just in
2312          * it taking place, picking some piece of information when it
2313          * appears in our event stream (vfs_getname comes to mind).
2314          */
2315         if (trace->base_time == 0 && !trace->full_time &&
2316             (evsel->attr.sample_type & PERF_SAMPLE_TIME))
2317                 trace->base_time = sample->time;
2318 }
2319
2320 static int trace__process_sample(struct perf_tool *tool,
2321                                  union perf_event *event,
2322                                  struct perf_sample *sample,
2323                                  struct perf_evsel *evsel,
2324                                  struct machine *machine __maybe_unused)
2325 {
2326         struct trace *trace = container_of(tool, struct trace, tool);
2327         struct thread *thread;
2328         int err = 0;
2329
2330         tracepoint_handler handler = evsel->handler;
2331
2332         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2333         if (thread && thread__is_filtered(thread))
2334                 goto out;
2335
2336         trace__set_base_time(trace, evsel, sample);
2337
2338         if (handler) {
2339                 ++trace->nr_events;
2340                 handler(trace, evsel, event, sample);
2341         }
2342 out:
2343         thread__put(thread);
2344         return err;
2345 }
2346
2347 static int trace__record(struct trace *trace, int argc, const char **argv)
2348 {
2349         unsigned int rec_argc, i, j;
2350         const char **rec_argv;
2351         const char * const record_args[] = {
2352                 "record",
2353                 "-R",
2354                 "-m", "1024",
2355                 "-c", "1",
2356         };
2357
2358         const char * const sc_args[] = { "-e", };
2359         unsigned int sc_args_nr = ARRAY_SIZE(sc_args);
2360         const char * const majpf_args[] = { "-e", "major-faults" };
2361         unsigned int majpf_args_nr = ARRAY_SIZE(majpf_args);
2362         const char * const minpf_args[] = { "-e", "minor-faults" };
2363         unsigned int minpf_args_nr = ARRAY_SIZE(minpf_args);
2364
2365         /* +1 is for the event string below */
2366         rec_argc = ARRAY_SIZE(record_args) + sc_args_nr + 1 +
2367                 majpf_args_nr + minpf_args_nr + argc;
2368         rec_argv = calloc(rec_argc + 1, sizeof(char *));
2369
2370         if (rec_argv == NULL)
2371                 return -ENOMEM;
2372
2373         j = 0;
2374         for (i = 0; i < ARRAY_SIZE(record_args); i++)
2375                 rec_argv[j++] = record_args[i];
2376
2377         if (trace->trace_syscalls) {
2378                 for (i = 0; i < sc_args_nr; i++)
2379                         rec_argv[j++] = sc_args[i];
2380
2381                 /* event string may be different for older kernels - e.g., RHEL6 */
2382                 if (is_valid_tracepoint("raw_syscalls:sys_enter"))
2383                         rec_argv[j++] = "raw_syscalls:sys_enter,raw_syscalls:sys_exit";
2384                 else if (is_valid_tracepoint("syscalls:sys_enter"))
2385                         rec_argv[j++] = "syscalls:sys_enter,syscalls:sys_exit";
2386                 else {
2387                         pr_err("Neither raw_syscalls nor syscalls events exist.\n");
2388                         free(rec_argv);
2389                         return -1;
2390                 }
2391         }
2392
2393         if (trace->trace_pgfaults & TRACE_PFMAJ)
2394                 for (i = 0; i < majpf_args_nr; i++)
2395                         rec_argv[j++] = majpf_args[i];
2396
2397         if (trace->trace_pgfaults & TRACE_PFMIN)
2398                 for (i = 0; i < minpf_args_nr; i++)
2399                         rec_argv[j++] = minpf_args[i];
2400
2401         for (i = 0; i < (unsigned int)argc; i++)
2402                 rec_argv[j++] = argv[i];
2403
2404         return cmd_record(j, rec_argv);
2405 }
2406
2407 static size_t trace__fprintf_thread_summary(struct trace *trace, FILE *fp);
2408
2409 static bool perf_evlist__add_vfs_getname(struct perf_evlist *evlist)
2410 {
2411         struct perf_evsel *evsel = perf_evsel__newtp("probe", "vfs_getname");
2412
2413         if (IS_ERR(evsel))
2414                 return false;
2415
2416         if (perf_evsel__field(evsel, "pathname") == NULL) {
2417                 perf_evsel__delete(evsel);
2418                 return false;
2419         }
2420
2421         evsel->handler = trace__vfs_getname;
2422         perf_evlist__add(evlist, evsel);
2423         return true;
2424 }
2425
2426 static struct perf_evsel *perf_evsel__new_pgfault(u64 config)
2427 {
2428         struct perf_evsel *evsel;
2429         struct perf_event_attr attr = {
2430                 .type = PERF_TYPE_SOFTWARE,
2431                 .mmap_data = 1,
2432         };
2433
2434         attr.config = config;
2435         attr.sample_period = 1;
2436
2437         event_attr_init(&attr);
2438
2439         evsel = perf_evsel__new(&attr);
2440         if (evsel)
2441                 evsel->handler = trace__pgfault;
2442
2443         return evsel;
2444 }
2445
2446 static void trace__handle_event(struct trace *trace, union perf_event *event, struct perf_sample *sample)
2447 {
2448         const u32 type = event->header.type;
2449         struct perf_evsel *evsel;
2450
2451         if (type != PERF_RECORD_SAMPLE) {
2452                 trace__process_event(trace, trace->host, event, sample);
2453                 return;
2454         }
2455
2456         evsel = perf_evlist__id2evsel(trace->evlist, sample->id);
2457         if (evsel == NULL) {
2458                 fprintf(trace->output, "Unknown tp ID %" PRIu64 ", skipping...\n", sample->id);
2459                 return;
2460         }
2461
2462         trace__set_base_time(trace, evsel, sample);
2463
2464         if (evsel->attr.type == PERF_TYPE_TRACEPOINT &&
2465             sample->raw_data == NULL) {
2466                 fprintf(trace->output, "%s sample with no payload for tid: %d, cpu %d, raw_size=%d, skipping...\n",
2467                        perf_evsel__name(evsel), sample->tid,
2468                        sample->cpu, sample->raw_size);
2469         } else {
2470                 tracepoint_handler handler = evsel->handler;
2471                 handler(trace, evsel, event, sample);
2472         }
2473
2474         if (trace->nr_events_printed >= trace->max_events && trace->max_events != ULONG_MAX)
2475                 interrupted = true;
2476 }
2477
2478 static int trace__add_syscall_newtp(struct trace *trace)
2479 {
2480         int ret = -1;
2481         struct perf_evlist *evlist = trace->evlist;
2482         struct perf_evsel *sys_enter, *sys_exit;
2483
2484         sys_enter = perf_evsel__raw_syscall_newtp("sys_enter", trace__sys_enter);
2485         if (sys_enter == NULL)
2486                 goto out;
2487
2488         if (perf_evsel__init_sc_tp_ptr_field(sys_enter, args))
2489                 goto out_delete_sys_enter;
2490
2491         sys_exit = perf_evsel__raw_syscall_newtp("sys_exit", trace__sys_exit);
2492         if (sys_exit == NULL)
2493                 goto out_delete_sys_enter;
2494
2495         if (perf_evsel__init_sc_tp_uint_field(sys_exit, ret))
2496                 goto out_delete_sys_exit;
2497
2498         perf_evsel__config_callchain(sys_enter, &trace->opts, &callchain_param);
2499         perf_evsel__config_callchain(sys_exit, &trace->opts, &callchain_param);
2500
2501         perf_evlist__add(evlist, sys_enter);
2502         perf_evlist__add(evlist, sys_exit);
2503
2504         if (callchain_param.enabled && !trace->kernel_syscallchains) {
2505                 /*
2506                  * We're interested only in the user space callchain
2507                  * leading to the syscall, allow overriding that for
2508                  * debugging reasons using --kernel_syscall_callchains
2509                  */
2510                 sys_exit->attr.exclude_callchain_kernel = 1;
2511         }
2512
2513         trace->syscalls.events.sys_enter = sys_enter;
2514         trace->syscalls.events.sys_exit  = sys_exit;
2515
2516         ret = 0;
2517 out:
2518         return ret;
2519
2520 out_delete_sys_exit:
2521         perf_evsel__delete_priv(sys_exit);
2522 out_delete_sys_enter:
2523         perf_evsel__delete_priv(sys_enter);
2524         goto out;
2525 }
2526
2527 static int trace__set_ev_qualifier_filter(struct trace *trace)
2528 {
2529         int err = -1;
2530         struct perf_evsel *sys_exit;
2531         char *filter = asprintf_expr_inout_ints("id", !trace->not_ev_qualifier,
2532                                                 trace->ev_qualifier_ids.nr,
2533                                                 trace->ev_qualifier_ids.entries);
2534
2535         if (filter == NULL)
2536                 goto out_enomem;
2537
2538         if (!perf_evsel__append_tp_filter(trace->syscalls.events.sys_enter,
2539                                           filter)) {
2540                 sys_exit = trace->syscalls.events.sys_exit;
2541                 err = perf_evsel__append_tp_filter(sys_exit, filter);
2542         }
2543
2544         free(filter);
2545 out:
2546         return err;
2547 out_enomem:
2548         errno = ENOMEM;
2549         goto out;
2550 }
2551
2552 static int trace__set_filter_loop_pids(struct trace *trace)
2553 {
2554         unsigned int nr = 1;
2555         pid_t pids[32] = {
2556                 getpid(),
2557         };
2558         struct thread *thread = machine__find_thread(trace->host, pids[0], pids[0]);
2559
2560         while (thread && nr < ARRAY_SIZE(pids)) {
2561                 struct thread *parent = machine__find_thread(trace->host, thread->ppid, thread->ppid);
2562
2563                 if (parent == NULL)
2564                         break;
2565
2566                 if (!strcmp(thread__comm_str(parent), "sshd")) {
2567                         pids[nr++] = parent->tid;
2568                         break;
2569                 }
2570                 thread = parent;
2571         }
2572
2573         return perf_evlist__set_filter_pids(trace->evlist, nr, pids);
2574 }
2575
2576 static int trace__run(struct trace *trace, int argc, const char **argv)
2577 {
2578         struct perf_evlist *evlist = trace->evlist;
2579         struct perf_evsel *evsel, *pgfault_maj = NULL, *pgfault_min = NULL;
2580         int err = -1, i;
2581         unsigned long before;
2582         const bool forks = argc > 0;
2583         bool draining = false;
2584
2585         trace->live = true;
2586
2587         if (trace->trace_syscalls && trace__add_syscall_newtp(trace))
2588                 goto out_error_raw_syscalls;
2589
2590         if (trace->trace_syscalls)
2591                 trace->vfs_getname = perf_evlist__add_vfs_getname(evlist);
2592
2593         if ((trace->trace_pgfaults & TRACE_PFMAJ)) {
2594                 pgfault_maj = perf_evsel__new_pgfault(PERF_COUNT_SW_PAGE_FAULTS_MAJ);
2595                 if (pgfault_maj == NULL)
2596                         goto out_error_mem;
2597                 perf_evsel__config_callchain(pgfault_maj, &trace->opts, &callchain_param);
2598                 perf_evlist__add(evlist, pgfault_maj);
2599         }
2600
2601         if ((trace->trace_pgfaults & TRACE_PFMIN)) {
2602                 pgfault_min = perf_evsel__new_pgfault(PERF_COUNT_SW_PAGE_FAULTS_MIN);
2603                 if (pgfault_min == NULL)
2604                         goto out_error_mem;
2605                 perf_evsel__config_callchain(pgfault_min, &trace->opts, &callchain_param);
2606                 perf_evlist__add(evlist, pgfault_min);
2607         }
2608
2609         if (trace->sched &&
2610             perf_evlist__add_newtp(evlist, "sched", "sched_stat_runtime",
2611                                    trace__sched_stat_runtime))
2612                 goto out_error_sched_stat_runtime;
2613
2614         /*
2615          * If a global cgroup was set, apply it to all the events without an
2616          * explicit cgroup. I.e.:
2617          *
2618          *      trace -G A -e sched:*switch
2619          *
2620          * Will set all raw_syscalls:sys_{enter,exit}, pgfault, vfs_getname, etc
2621          * _and_ sched:sched_switch to the 'A' cgroup, while:
2622          *
2623          * trace -e sched:*switch -G A
2624          *
2625          * will only set the sched:sched_switch event to the 'A' cgroup, all the
2626          * other events (raw_syscalls:sys_{enter,exit}, etc are left "without"
2627          * a cgroup (on the root cgroup, sys wide, etc).
2628          *
2629          * Multiple cgroups:
2630          *
2631          * trace -G A -e sched:*switch -G B
2632          *
2633          * the syscall ones go to the 'A' cgroup, the sched:sched_switch goes
2634          * to the 'B' cgroup.
2635          *
2636          * evlist__set_default_cgroup() grabs a reference of the passed cgroup
2637          * only for the evsels still without a cgroup, i.e. evsel->cgroup == NULL.
2638          */
2639         if (trace->cgroup)
2640                 evlist__set_default_cgroup(trace->evlist, trace->cgroup);
2641
2642         err = perf_evlist__create_maps(evlist, &trace->opts.target);
2643         if (err < 0) {
2644                 fprintf(trace->output, "Problems parsing the target to trace, check your options!\n");
2645                 goto out_delete_evlist;
2646         }
2647
2648         err = trace__symbols_init(trace, evlist);
2649         if (err < 0) {
2650                 fprintf(trace->output, "Problems initializing symbol libraries!\n");
2651                 goto out_delete_evlist;
2652         }
2653
2654         perf_evlist__config(evlist, &trace->opts, &callchain_param);
2655
2656         signal(SIGCHLD, sig_handler);
2657         signal(SIGINT, sig_handler);
2658
2659         if (forks) {
2660                 err = perf_evlist__prepare_workload(evlist, &trace->opts.target,
2661                                                     argv, false, NULL);
2662                 if (err < 0) {
2663                         fprintf(trace->output, "Couldn't run the workload!\n");
2664                         goto out_delete_evlist;
2665                 }
2666         }
2667
2668         err = perf_evlist__open(evlist);
2669         if (err < 0)
2670                 goto out_error_open;
2671
2672         err = bpf__apply_obj_config();
2673         if (err) {
2674                 char errbuf[BUFSIZ];
2675
2676                 bpf__strerror_apply_obj_config(err, errbuf, sizeof(errbuf));
2677                 pr_err("ERROR: Apply config to BPF failed: %s\n",
2678                          errbuf);
2679                 goto out_error_open;
2680         }
2681
2682         /*
2683          * Better not use !target__has_task() here because we need to cover the
2684          * case where no threads were specified in the command line, but a
2685          * workload was, and in that case we will fill in the thread_map when
2686          * we fork the workload in perf_evlist__prepare_workload.
2687          */
2688         if (trace->filter_pids.nr > 0)
2689                 err = perf_evlist__set_filter_pids(evlist, trace->filter_pids.nr, trace->filter_pids.entries);
2690         else if (thread_map__pid(evlist->threads, 0) == -1)
2691                 err = trace__set_filter_loop_pids(trace);
2692
2693         if (err < 0)
2694                 goto out_error_mem;
2695
2696         if (trace->ev_qualifier_ids.nr > 0) {
2697                 err = trace__set_ev_qualifier_filter(trace);
2698                 if (err < 0)
2699                         goto out_errno;
2700
2701                 pr_debug("event qualifier tracepoint filter: %s\n",
2702                          trace->syscalls.events.sys_exit->filter);
2703         }
2704
2705         err = perf_evlist__apply_filters(evlist, &evsel);
2706         if (err < 0)
2707                 goto out_error_apply_filters;
2708
2709         err = perf_evlist__mmap(evlist, trace->opts.mmap_pages);
2710         if (err < 0)
2711                 goto out_error_mmap;
2712
2713         if (!target__none(&trace->opts.target) && !trace->opts.initial_delay)
2714                 perf_evlist__enable(evlist);
2715
2716         if (forks)
2717                 perf_evlist__start_workload(evlist);
2718
2719         if (trace->opts.initial_delay) {
2720                 usleep(trace->opts.initial_delay * 1000);
2721                 perf_evlist__enable(evlist);
2722         }
2723
2724         trace->multiple_threads = thread_map__pid(evlist->threads, 0) == -1 ||
2725                                   evlist->threads->nr > 1 ||
2726                                   perf_evlist__first(evlist)->attr.inherit;
2727
2728         /*
2729          * Now that we already used evsel->attr to ask the kernel to setup the
2730          * events, lets reuse evsel->attr.sample_max_stack as the limit in
2731          * trace__resolve_callchain(), allowing per-event max-stack settings
2732          * to override an explicitely set --max-stack global setting.
2733          */
2734         evlist__for_each_entry(evlist, evsel) {
2735                 if (evsel__has_callchain(evsel) &&
2736                     evsel->attr.sample_max_stack == 0)
2737                         evsel->attr.sample_max_stack = trace->max_stack;
2738         }
2739 again:
2740         before = trace->nr_events;
2741
2742         for (i = 0; i < evlist->nr_mmaps; i++) {
2743                 union perf_event *event;
2744                 struct perf_mmap *md;
2745
2746                 md = &evlist->mmap[i];
2747                 if (perf_mmap__read_init(md) < 0)
2748                         continue;
2749
2750                 while ((event = perf_mmap__read_event(md)) != NULL) {
2751                         struct perf_sample sample;
2752
2753                         ++trace->nr_events;
2754
2755                         err = perf_evlist__parse_sample(evlist, event, &sample);
2756                         if (err) {
2757                                 fprintf(trace->output, "Can't parse sample, err = %d, skipping...\n", err);
2758                                 goto next_event;
2759                         }
2760
2761                         trace__handle_event(trace, event, &sample);
2762 next_event:
2763                         perf_mmap__consume(md);
2764
2765                         if (interrupted)
2766                                 goto out_disable;
2767
2768                         if (done && !draining) {
2769                                 perf_evlist__disable(evlist);
2770                                 draining = true;
2771                         }
2772                 }
2773                 perf_mmap__read_done(md);
2774         }
2775
2776         if (trace->nr_events == before) {
2777                 int timeout = done ? 100 : -1;
2778
2779                 if (!draining && perf_evlist__poll(evlist, timeout) > 0) {
2780                         if (perf_evlist__filter_pollfd(evlist, POLLERR | POLLHUP | POLLNVAL) == 0)
2781                                 draining = true;
2782
2783                         goto again;
2784                 }
2785         } else {
2786                 goto again;
2787         }
2788
2789 out_disable:
2790         thread__zput(trace->current);
2791
2792         perf_evlist__disable(evlist);
2793
2794         if (!err) {
2795                 if (trace->summary)
2796                         trace__fprintf_thread_summary(trace, trace->output);
2797
2798                 if (trace->show_tool_stats) {
2799                         fprintf(trace->output, "Stats:\n "
2800                                                " vfs_getname : %" PRIu64 "\n"
2801                                                " proc_getname: %" PRIu64 "\n",
2802                                 trace->stats.vfs_getname,
2803                                 trace->stats.proc_getname);
2804                 }
2805         }
2806
2807 out_delete_evlist:
2808         trace__symbols__exit(trace);
2809
2810         perf_evlist__delete(evlist);
2811         cgroup__put(trace->cgroup);
2812         trace->evlist = NULL;
2813         trace->live = false;
2814         return err;
2815 {
2816         char errbuf[BUFSIZ];
2817
2818 out_error_sched_stat_runtime:
2819         tracing_path__strerror_open_tp(errno, errbuf, sizeof(errbuf), "sched", "sched_stat_runtime");
2820         goto out_error;
2821
2822 out_error_raw_syscalls:
2823         tracing_path__strerror_open_tp(errno, errbuf, sizeof(errbuf), "raw_syscalls", "sys_(enter|exit)");
2824         goto out_error;
2825
2826 out_error_mmap:
2827         perf_evlist__strerror_mmap(evlist, errno, errbuf, sizeof(errbuf));
2828         goto out_error;
2829
2830 out_error_open:
2831         perf_evlist__strerror_open(evlist, errno, errbuf, sizeof(errbuf));
2832
2833 out_error:
2834         fprintf(trace->output, "%s\n", errbuf);
2835         goto out_delete_evlist;
2836
2837 out_error_apply_filters:
2838         fprintf(trace->output,
2839                 "Failed to set filter \"%s\" on event %s with %d (%s)\n",
2840                 evsel->filter, perf_evsel__name(evsel), errno,
2841                 str_error_r(errno, errbuf, sizeof(errbuf)));
2842         goto out_delete_evlist;
2843 }
2844 out_error_mem:
2845         fprintf(trace->output, "Not enough memory to run!\n");
2846         goto out_delete_evlist;
2847
2848 out_errno:
2849         fprintf(trace->output, "errno=%d,%s\n", errno, strerror(errno));
2850         goto out_delete_evlist;
2851 }
2852
2853 static int trace__replay(struct trace *trace)
2854 {
2855         const struct perf_evsel_str_handler handlers[] = {
2856                 { "probe:vfs_getname",       trace__vfs_getname, },
2857         };
2858         struct perf_data data = {
2859                 .file      = {
2860                         .path = input_name,
2861                 },
2862                 .mode      = PERF_DATA_MODE_READ,
2863                 .force     = trace->force,
2864         };
2865         struct perf_session *session;
2866         struct perf_evsel *evsel;
2867         int err = -1;
2868
2869         trace->tool.sample        = trace__process_sample;
2870         trace->tool.mmap          = perf_event__process_mmap;
2871         trace->tool.mmap2         = perf_event__process_mmap2;
2872         trace->tool.comm          = perf_event__process_comm;
2873         trace->tool.exit          = perf_event__process_exit;
2874         trace->tool.fork          = perf_event__process_fork;
2875         trace->tool.attr          = perf_event__process_attr;
2876         trace->tool.tracing_data  = perf_event__process_tracing_data;
2877         trace->tool.build_id      = perf_event__process_build_id;
2878         trace->tool.namespaces    = perf_event__process_namespaces;
2879
2880         trace->tool.ordered_events = true;
2881         trace->tool.ordering_requires_timestamps = true;
2882
2883         /* add tid to output */
2884         trace->multiple_threads = true;
2885
2886         session = perf_session__new(&data, false, &trace->tool);
2887         if (session == NULL)
2888                 return -1;
2889
2890         if (trace->opts.target.pid)
2891                 symbol_conf.pid_list_str = strdup(trace->opts.target.pid);
2892
2893         if (trace->opts.target.tid)
2894                 symbol_conf.tid_list_str = strdup(trace->opts.target.tid);
2895
2896         if (symbol__init(&session->header.env) < 0)
2897                 goto out;
2898
2899         trace->host = &session->machines.host;
2900
2901         err = perf_session__set_tracepoints_handlers(session, handlers);
2902         if (err)
2903                 goto out;
2904
2905         evsel = perf_evlist__find_tracepoint_by_name(session->evlist,
2906                                                      "raw_syscalls:sys_enter");
2907         /* older kernels have syscalls tp versus raw_syscalls */
2908         if (evsel == NULL)
2909                 evsel = perf_evlist__find_tracepoint_by_name(session->evlist,
2910                                                              "syscalls:sys_enter");
2911
2912         if (evsel &&
2913             (perf_evsel__init_raw_syscall_tp(evsel, trace__sys_enter) < 0 ||
2914             perf_evsel__init_sc_tp_ptr_field(evsel, args))) {
2915                 pr_err("Error during initialize raw_syscalls:sys_enter event\n");
2916                 goto out;
2917         }
2918
2919         evsel = perf_evlist__find_tracepoint_by_name(session->evlist,
2920                                                      "raw_syscalls:sys_exit");
2921         if (evsel == NULL)
2922                 evsel = perf_evlist__find_tracepoint_by_name(session->evlist,
2923                                                              "syscalls:sys_exit");
2924         if (evsel &&
2925             (perf_evsel__init_raw_syscall_tp(evsel, trace__sys_exit) < 0 ||
2926             perf_evsel__init_sc_tp_uint_field(evsel, ret))) {
2927                 pr_err("Error during initialize raw_syscalls:sys_exit event\n");
2928                 goto out;
2929         }
2930
2931         evlist__for_each_entry(session->evlist, evsel) {
2932                 if (evsel->attr.type == PERF_TYPE_SOFTWARE &&
2933                     (evsel->attr.config == PERF_COUNT_SW_PAGE_FAULTS_MAJ ||
2934                      evsel->attr.config == PERF_COUNT_SW_PAGE_FAULTS_MIN ||
2935                      evsel->attr.config == PERF_COUNT_SW_PAGE_FAULTS))
2936                         evsel->handler = trace__pgfault;
2937         }
2938
2939         setup_pager();
2940
2941         err = perf_session__process_events(session);
2942         if (err)
2943                 pr_err("Failed to process events, error %d", err);
2944
2945         else if (trace->summary)
2946                 trace__fprintf_thread_summary(trace, trace->output);
2947
2948 out:
2949         perf_session__delete(session);
2950
2951         return err;
2952 }
2953
2954 static size_t trace__fprintf_threads_header(FILE *fp)
2955 {
2956         size_t printed;
2957
2958         printed  = fprintf(fp, "\n Summary of events:\n\n");
2959
2960         return printed;
2961 }
2962
2963 DEFINE_RESORT_RB(syscall_stats, a->msecs > b->msecs,
2964         struct stats    *stats;
2965         double          msecs;
2966         int             syscall;
2967 )
2968 {
2969         struct int_node *source = rb_entry(nd, struct int_node, rb_node);
2970         struct stats *stats = source->priv;
2971
2972         entry->syscall = source->i;
2973         entry->stats   = stats;
2974         entry->msecs   = stats ? (u64)stats->n * (avg_stats(stats) / NSEC_PER_MSEC) : 0;
2975 }
2976
2977 static size_t thread__dump_stats(struct thread_trace *ttrace,
2978                                  struct trace *trace, FILE *fp)
2979 {
2980         size_t printed = 0;
2981         struct syscall *sc;
2982         struct rb_node *nd;
2983         DECLARE_RESORT_RB_INTLIST(syscall_stats, ttrace->syscall_stats);
2984
2985         if (syscall_stats == NULL)
2986                 return 0;
2987
2988         printed += fprintf(fp, "\n");
2989
2990         printed += fprintf(fp, "   syscall            calls    total       min       avg       max      stddev\n");
2991         printed += fprintf(fp, "                               (msec)    (msec)    (msec)    (msec)        (%%)\n");
2992         printed += fprintf(fp, "   --------------- -------- --------- --------- --------- ---------     ------\n");
2993
2994         resort_rb__for_each_entry(nd, syscall_stats) {
2995                 struct stats *stats = syscall_stats_entry->stats;
2996                 if (stats) {
2997                         double min = (double)(stats->min) / NSEC_PER_MSEC;
2998                         double max = (double)(stats->max) / NSEC_PER_MSEC;
2999                         double avg = avg_stats(stats);
3000                         double pct;
3001                         u64 n = (u64) stats->n;
3002
3003                         pct = avg ? 100.0 * stddev_stats(stats)/avg : 0.0;
3004                         avg /= NSEC_PER_MSEC;
3005
3006                         sc = &trace->syscalls.table[syscall_stats_entry->syscall];
3007                         printed += fprintf(fp, "   %-15s", sc->name);
3008                         printed += fprintf(fp, " %8" PRIu64 " %9.3f %9.3f %9.3f",
3009                                            n, syscall_stats_entry->msecs, min, avg);
3010                         printed += fprintf(fp, " %9.3f %9.2f%%\n", max, pct);
3011                 }
3012         }
3013
3014         resort_rb__delete(syscall_stats);
3015         printed += fprintf(fp, "\n\n");
3016
3017         return printed;
3018 }
3019
3020 static size_t trace__fprintf_thread(FILE *fp, struct thread *thread, struct trace *trace)
3021 {
3022         size_t printed = 0;
3023         struct thread_trace *ttrace = thread__priv(thread);
3024         double ratio;
3025
3026         if (ttrace == NULL)
3027                 return 0;
3028
3029         ratio = (double)ttrace->nr_events / trace->nr_events * 100.0;
3030
3031         printed += fprintf(fp, " %s (%d), ", thread__comm_str(thread), thread->tid);
3032         printed += fprintf(fp, "%lu events, ", ttrace->nr_events);
3033         printed += fprintf(fp, "%.1f%%", ratio);
3034         if (ttrace->pfmaj)
3035                 printed += fprintf(fp, ", %lu majfaults", ttrace->pfmaj);
3036         if (ttrace->pfmin)
3037                 printed += fprintf(fp, ", %lu minfaults", ttrace->pfmin);
3038         if (trace->sched)
3039                 printed += fprintf(fp, ", %.3f msec\n", ttrace->runtime_ms);
3040         else if (fputc('\n', fp) != EOF)
3041                 ++printed;
3042
3043         printed += thread__dump_stats(ttrace, trace, fp);
3044
3045         return printed;
3046 }
3047
3048 static unsigned long thread__nr_events(struct thread_trace *ttrace)
3049 {
3050         return ttrace ? ttrace->nr_events : 0;
3051 }
3052
3053 DEFINE_RESORT_RB(threads, (thread__nr_events(a->thread->priv) < thread__nr_events(b->thread->priv)),
3054         struct thread *thread;
3055 )
3056 {
3057         entry->thread = rb_entry(nd, struct thread, rb_node);
3058 }
3059
3060 static size_t trace__fprintf_thread_summary(struct trace *trace, FILE *fp)
3061 {
3062         size_t printed = trace__fprintf_threads_header(fp);
3063         struct rb_node *nd;
3064         int i;
3065
3066         for (i = 0; i < THREADS__TABLE_SIZE; i++) {
3067                 DECLARE_RESORT_RB_MACHINE_THREADS(threads, trace->host, i);
3068
3069                 if (threads == NULL) {
3070                         fprintf(fp, "%s", "Error sorting output by nr_events!\n");
3071                         return 0;
3072                 }
3073
3074                 resort_rb__for_each_entry(nd, threads)
3075                         printed += trace__fprintf_thread(fp, threads_entry->thread, trace);
3076
3077                 resort_rb__delete(threads);
3078         }
3079         return printed;
3080 }
3081
3082 static int trace__set_duration(const struct option *opt, const char *str,
3083                                int unset __maybe_unused)
3084 {
3085         struct trace *trace = opt->value;
3086
3087         trace->duration_filter = atof(str);
3088         return 0;
3089 }
3090
3091 static int trace__set_filter_pids(const struct option *opt, const char *str,
3092                                   int unset __maybe_unused)
3093 {
3094         int ret = -1;
3095         size_t i;
3096         struct trace *trace = opt->value;
3097         /*
3098          * FIXME: introduce a intarray class, plain parse csv and create a
3099          * { int nr, int entries[] } struct...
3100          */
3101         struct intlist *list = intlist__new(str);
3102
3103         if (list == NULL)
3104                 return -1;
3105
3106         i = trace->filter_pids.nr = intlist__nr_entries(list) + 1;
3107         trace->filter_pids.entries = calloc(i, sizeof(pid_t));
3108
3109         if (trace->filter_pids.entries == NULL)
3110                 goto out;
3111
3112         trace->filter_pids.entries[0] = getpid();
3113
3114         for (i = 1; i < trace->filter_pids.nr; ++i)
3115                 trace->filter_pids.entries[i] = intlist__entry(list, i - 1)->i;
3116
3117         intlist__delete(list);
3118         ret = 0;
3119 out:
3120         return ret;
3121 }
3122
3123 static int trace__open_output(struct trace *trace, const char *filename)
3124 {
3125         struct stat st;
3126
3127         if (!stat(filename, &st) && st.st_size) {
3128                 char oldname[PATH_MAX];
3129
3130                 scnprintf(oldname, sizeof(oldname), "%s.old", filename);
3131                 unlink(oldname);
3132                 rename(filename, oldname);
3133         }
3134
3135         trace->output = fopen(filename, "w");
3136
3137         return trace->output == NULL ? -errno : 0;
3138 }
3139
3140 static int parse_pagefaults(const struct option *opt, const char *str,
3141                             int unset __maybe_unused)
3142 {
3143         int *trace_pgfaults = opt->value;
3144
3145         if (strcmp(str, "all") == 0)
3146                 *trace_pgfaults |= TRACE_PFMAJ | TRACE_PFMIN;
3147         else if (strcmp(str, "maj") == 0)
3148                 *trace_pgfaults |= TRACE_PFMAJ;
3149         else if (strcmp(str, "min") == 0)
3150                 *trace_pgfaults |= TRACE_PFMIN;
3151         else
3152                 return -1;
3153
3154         return 0;
3155 }
3156
3157 static void evlist__set_evsel_handler(struct perf_evlist *evlist, void *handler)
3158 {
3159         struct perf_evsel *evsel;
3160
3161         evlist__for_each_entry(evlist, evsel)
3162                 evsel->handler = handler;
3163 }
3164
3165 static int evlist__set_syscall_tp_fields(struct perf_evlist *evlist)
3166 {
3167         struct perf_evsel *evsel;
3168
3169         evlist__for_each_entry(evlist, evsel) {
3170                 if (evsel->priv || !evsel->tp_format)
3171                         continue;
3172
3173                 if (strcmp(evsel->tp_format->system, "syscalls"))
3174                         continue;
3175
3176                 if (perf_evsel__init_syscall_tp(evsel))
3177                         return -1;
3178
3179                 if (!strncmp(evsel->tp_format->name, "sys_enter_", 10)) {
3180                         struct syscall_tp *sc = evsel->priv;
3181
3182                         if (__tp_field__init_ptr(&sc->args, sc->id.offset + sizeof(u64)))
3183                                 return -1;
3184                 } else if (!strncmp(evsel->tp_format->name, "sys_exit_", 9)) {
3185                         struct syscall_tp *sc = evsel->priv;
3186
3187                         if (__tp_field__init_uint(&sc->ret, sizeof(u64), sc->id.offset + sizeof(u64), evsel->needs_swap))
3188                                 return -1;
3189                 }
3190         }
3191
3192         return 0;
3193 }
3194
3195 /*
3196  * XXX: Hackish, just splitting the combined -e+--event (syscalls
3197  * (raw_syscalls:{sys_{enter,exit}} + events (tracepoints, HW, SW, etc) to use
3198  * existing facilities unchanged (trace->ev_qualifier + parse_options()).
3199  *
3200  * It'd be better to introduce a parse_options() variant that would return a
3201  * list with the terms it didn't match to an event...
3202  */
3203 static int trace__parse_events_option(const struct option *opt, const char *str,
3204                                       int unset __maybe_unused)
3205 {
3206         struct trace *trace = (struct trace *)opt->value;
3207         const char *s = str;
3208         char *sep = NULL, *lists[2] = { NULL, NULL, };
3209         int len = strlen(str) + 1, err = -1, list, idx;
3210         char *strace_groups_dir = system_path(STRACE_GROUPS_DIR);
3211         char group_name[PATH_MAX];
3212         struct syscall_fmt *fmt;
3213
3214         if (strace_groups_dir == NULL)
3215                 return -1;
3216
3217         if (*s == '!') {
3218                 ++s;
3219                 trace->not_ev_qualifier = true;
3220         }
3221
3222         while (1) {
3223                 if ((sep = strchr(s, ',')) != NULL)
3224                         *sep = '\0';
3225
3226                 list = 0;
3227                 if (syscalltbl__id(trace->sctbl, s) >= 0 ||
3228                     syscalltbl__strglobmatch_first(trace->sctbl, s, &idx) >= 0) {
3229                         list = 1;
3230                         goto do_concat;
3231                 }
3232
3233                 fmt = syscall_fmt__find_by_alias(s);
3234                 if (fmt != NULL) {
3235                         list = 1;
3236                         s = fmt->name;
3237                 } else {
3238                         path__join(group_name, sizeof(group_name), strace_groups_dir, s);
3239                         if (access(group_name, R_OK) == 0)
3240                                 list = 1;
3241                 }
3242 do_concat:
3243                 if (lists[list]) {
3244                         sprintf(lists[list] + strlen(lists[list]), ",%s", s);
3245                 } else {
3246                         lists[list] = malloc(len);
3247                         if (lists[list] == NULL)
3248                                 goto out;
3249                         strcpy(lists[list], s);
3250                 }
3251
3252                 if (!sep)
3253                         break;
3254
3255                 *sep = ',';
3256                 s = sep + 1;
3257         }
3258
3259         if (lists[1] != NULL) {
3260                 struct strlist_config slist_config = {
3261                         .dirname = strace_groups_dir,
3262                 };
3263
3264                 trace->ev_qualifier = strlist__new(lists[1], &slist_config);
3265                 if (trace->ev_qualifier == NULL) {
3266                         fputs("Not enough memory to parse event qualifier", trace->output);
3267                         goto out;
3268                 }
3269
3270                 if (trace__validate_ev_qualifier(trace))
3271                         goto out;
3272                 trace->trace_syscalls = true;
3273         }
3274
3275         err = 0;
3276
3277         if (lists[0]) {
3278                 struct option o = OPT_CALLBACK('e', "event", &trace->evlist, "event",
3279                                                "event selector. use 'perf list' to list available events",
3280                                                parse_events_option);
3281                 err = parse_events_option(&o, lists[0], 0);
3282         }
3283 out:
3284         if (sep)
3285                 *sep = ',';
3286
3287         return err;
3288 }
3289
3290 static int trace__parse_cgroups(const struct option *opt, const char *str, int unset)
3291 {
3292         struct trace *trace = opt->value;
3293
3294         if (!list_empty(&trace->evlist->entries))
3295                 return parse_cgroups(opt, str, unset);
3296
3297         trace->cgroup = evlist__findnew_cgroup(trace->evlist, str);
3298
3299         return 0;
3300 }
3301
3302 int cmd_trace(int argc, const char **argv)
3303 {
3304         const char *trace_usage[] = {
3305                 "perf trace [<options>] [<command>]",
3306                 "perf trace [<options>] -- <command> [<options>]",
3307                 "perf trace record [<options>] [<command>]",
3308                 "perf trace record [<options>] -- <command> [<options>]",
3309                 NULL
3310         };
3311         struct trace trace = {
3312                 .syscalls = {
3313                         . max = -1,
3314                 },
3315                 .opts = {
3316                         .target = {
3317                                 .uid       = UINT_MAX,
3318                                 .uses_mmap = true,
3319                         },
3320                         .user_freq     = UINT_MAX,
3321                         .user_interval = ULLONG_MAX,
3322                         .no_buffering  = true,
3323                         .mmap_pages    = UINT_MAX,
3324                         .proc_map_timeout  = 500,
3325                 },
3326                 .output = stderr,
3327                 .show_comm = true,
3328                 .trace_syscalls = false,
3329                 .kernel_syscallchains = false,
3330                 .max_stack = UINT_MAX,
3331                 .max_events = ULONG_MAX,
3332         };
3333         const char *output_name = NULL;
3334         const struct option trace_options[] = {
3335         OPT_CALLBACK('e', "event", &trace, "event",
3336                      "event/syscall selector. use 'perf list' to list available events",
3337                      trace__parse_events_option),
3338         OPT_BOOLEAN(0, "comm", &trace.show_comm,
3339                     "show the thread COMM next to its id"),
3340         OPT_BOOLEAN(0, "tool_stats", &trace.show_tool_stats, "show tool stats"),
3341         OPT_CALLBACK(0, "expr", &trace, "expr", "list of syscalls/events to trace",
3342                      trace__parse_events_option),
3343         OPT_STRING('o', "output", &output_name, "file", "output file name"),
3344         OPT_STRING('i', "input", &input_name, "file", "Analyze events in file"),
3345         OPT_STRING('p', "pid", &trace.opts.target.pid, "pid",
3346                     "trace events on existing process id"),
3347         OPT_STRING('t', "tid", &trace.opts.target.tid, "tid",
3348                     "trace events on existing thread id"),
3349         OPT_CALLBACK(0, "filter-pids", &trace, "CSV list of pids",
3350                      "pids to filter (by the kernel)", trace__set_filter_pids),
3351         OPT_BOOLEAN('a', "all-cpus", &trace.opts.target.system_wide,
3352                     "system-wide collection from all CPUs"),
3353         OPT_STRING('C', "cpu", &trace.opts.target.cpu_list, "cpu",
3354                     "list of cpus to monitor"),
3355         OPT_BOOLEAN(0, "no-inherit", &trace.opts.no_inherit,
3356                     "child tasks do not inherit counters"),
3357         OPT_CALLBACK('m', "mmap-pages", &trace.opts.mmap_pages, "pages",
3358                      "number of mmap data pages",
3359                      perf_evlist__parse_mmap_pages),
3360         OPT_STRING('u', "uid", &trace.opts.target.uid_str, "user",
3361                    "user to profile"),
3362         OPT_CALLBACK(0, "duration", &trace, "float",
3363                      "show only events with duration > N.M ms",
3364                      trace__set_duration),
3365         OPT_BOOLEAN(0, "sched", &trace.sched, "show blocking scheduler events"),
3366         OPT_INCR('v', "verbose", &verbose, "be more verbose"),
3367         OPT_BOOLEAN('T', "time", &trace.full_time,
3368                     "Show full timestamp, not time relative to first start"),
3369         OPT_BOOLEAN(0, "failure", &trace.failure_only,
3370                     "Show only syscalls that failed"),
3371         OPT_BOOLEAN('s', "summary", &trace.summary_only,
3372                     "Show only syscall summary with statistics"),
3373         OPT_BOOLEAN('S', "with-summary", &trace.summary,
3374                     "Show all syscalls and summary with statistics"),
3375         OPT_CALLBACK_DEFAULT('F', "pf", &trace.trace_pgfaults, "all|maj|min",
3376                      "Trace pagefaults", parse_pagefaults, "maj"),
3377         OPT_BOOLEAN(0, "syscalls", &trace.trace_syscalls, "Trace syscalls"),
3378         OPT_BOOLEAN('f', "force", &trace.force, "don't complain, do it"),
3379         OPT_CALLBACK(0, "call-graph", &trace.opts,
3380                      "record_mode[,record_size]", record_callchain_help,
3381                      &record_parse_callchain_opt),
3382         OPT_BOOLEAN(0, "kernel-syscall-graph", &trace.kernel_syscallchains,
3383                     "Show the kernel callchains on the syscall exit path"),
3384         OPT_ULONG(0, "max-events", &trace.max_events,
3385                 "Set the maximum number of events to print, exit after that is reached. "),
3386         OPT_UINTEGER(0, "min-stack", &trace.min_stack,
3387                      "Set the minimum stack depth when parsing the callchain, "
3388                      "anything below the specified depth will be ignored."),
3389         OPT_UINTEGER(0, "max-stack", &trace.max_stack,
3390                      "Set the maximum stack depth when parsing the callchain, "
3391                      "anything beyond the specified depth will be ignored. "
3392                      "Default: kernel.perf_event_max_stack or " __stringify(PERF_MAX_STACK_DEPTH)),
3393         OPT_BOOLEAN(0, "print-sample", &trace.print_sample,
3394                         "print the PERF_RECORD_SAMPLE PERF_SAMPLE_ info, for debugging"),
3395         OPT_UINTEGER(0, "proc-map-timeout", &trace.opts.proc_map_timeout,
3396                         "per thread proc mmap processing timeout in ms"),
3397         OPT_CALLBACK('G', "cgroup", &trace, "name", "monitor event in cgroup name only",
3398                      trace__parse_cgroups),
3399         OPT_UINTEGER('D', "delay", &trace.opts.initial_delay,
3400                      "ms to wait before starting measurement after program "
3401                      "start"),
3402         OPT_END()
3403         };
3404         bool __maybe_unused max_stack_user_set = true;
3405         bool mmap_pages_user_set = true;
3406         struct perf_evsel *evsel;
3407         const char * const trace_subcommands[] = { "record", NULL };
3408         int err = -1;
3409         char bf[BUFSIZ];
3410
3411         signal(SIGSEGV, sighandler_dump_stack);
3412         signal(SIGFPE, sighandler_dump_stack);
3413
3414         trace.evlist = perf_evlist__new();
3415         trace.sctbl = syscalltbl__new();
3416
3417         if (trace.evlist == NULL || trace.sctbl == NULL) {
3418                 pr_err("Not enough memory to run!\n");
3419                 err = -ENOMEM;
3420                 goto out;
3421         }
3422
3423         argc = parse_options_subcommand(argc, argv, trace_options, trace_subcommands,
3424                                  trace_usage, PARSE_OPT_STOP_AT_NON_OPTION);
3425
3426         if ((nr_cgroups || trace.cgroup) && !trace.opts.target.system_wide) {
3427                 usage_with_options_msg(trace_usage, trace_options,
3428                                        "cgroup monitoring only available in system-wide mode");
3429         }
3430
3431         evsel = bpf__setup_output_event(trace.evlist, "__augmented_syscalls__");
3432         if (IS_ERR(evsel)) {
3433                 bpf__strerror_setup_output_event(trace.evlist, PTR_ERR(evsel), bf, sizeof(bf));
3434                 pr_err("ERROR: Setup trace syscalls enter failed: %s\n", bf);
3435                 goto out;
3436         }
3437
3438         if (evsel)
3439                 trace.syscalls.events.augmented = evsel;
3440
3441         err = bpf__setup_stdout(trace.evlist);
3442         if (err) {
3443                 bpf__strerror_setup_stdout(trace.evlist, err, bf, sizeof(bf));
3444                 pr_err("ERROR: Setup BPF stdout failed: %s\n", bf);
3445                 goto out;
3446         }
3447
3448         err = -1;
3449
3450         if (trace.trace_pgfaults) {
3451                 trace.opts.sample_address = true;
3452                 trace.opts.sample_time = true;
3453         }
3454
3455         if (trace.opts.mmap_pages == UINT_MAX)
3456                 mmap_pages_user_set = false;
3457
3458         if (trace.max_stack == UINT_MAX) {
3459                 trace.max_stack = input_name ? PERF_MAX_STACK_DEPTH : sysctl__max_stack();
3460                 max_stack_user_set = false;
3461         }
3462
3463 #ifdef HAVE_DWARF_UNWIND_SUPPORT
3464         if ((trace.min_stack || max_stack_user_set) && !callchain_param.enabled) {
3465                 record_opts__parse_callchain(&trace.opts, &callchain_param, "dwarf", false);
3466         }
3467 #endif
3468
3469         if (callchain_param.enabled) {
3470                 if (!mmap_pages_user_set && geteuid() == 0)
3471                         trace.opts.mmap_pages = perf_event_mlock_kb_in_pages() * 4;
3472
3473                 symbol_conf.use_callchain = true;
3474         }
3475
3476         if (trace.evlist->nr_entries > 0) {
3477                 evlist__set_evsel_handler(trace.evlist, trace__event_handler);
3478                 if (evlist__set_syscall_tp_fields(trace.evlist)) {
3479                         perror("failed to set syscalls:* tracepoint fields");
3480                         goto out;
3481                 }
3482         }
3483
3484         /*
3485          * If we are augmenting syscalls, then combine what we put in the
3486          * __augmented_syscalls__ BPF map with what is in the
3487          * syscalls:sys_exit_FOO tracepoints, i.e. just like we do without BPF,
3488          * combining raw_syscalls:sys_enter with raw_syscalls:sys_exit.
3489          *
3490          * We'll switch to look at two BPF maps, one for sys_enter and the
3491          * other for sys_exit when we start augmenting the sys_exit paths with
3492          * buffers that are being copied from kernel to userspace, think 'read'
3493          * syscall.
3494          */
3495         if (trace.syscalls.events.augmented) {
3496                 evsel = trace.syscalls.events.augmented;
3497
3498                 if (perf_evsel__init_augmented_syscall_tp(evsel) ||
3499                     perf_evsel__init_augmented_syscall_tp_args(evsel))
3500                         goto out;
3501                 evsel->handler = trace__sys_enter;
3502
3503                 evlist__for_each_entry(trace.evlist, evsel) {
3504                         if (strstarts(perf_evsel__name(evsel), "syscalls:sys_exit_")) {
3505                                 perf_evsel__init_augmented_syscall_tp(evsel);
3506                                 perf_evsel__init_augmented_syscall_tp_ret(evsel);
3507                                 evsel->handler = trace__sys_exit;
3508                         }
3509                 }
3510         }
3511
3512         if ((argc >= 1) && (strcmp(argv[0], "record") == 0))
3513                 return trace__record(&trace, argc-1, &argv[1]);
3514
3515         /* summary_only implies summary option, but don't overwrite summary if set */
3516         if (trace.summary_only)
3517                 trace.summary = trace.summary_only;
3518
3519         if (!trace.trace_syscalls && !trace.trace_pgfaults &&
3520             trace.evlist->nr_entries == 0 /* Was --events used? */) {
3521                 trace.trace_syscalls = true;
3522         }
3523
3524         if (output_name != NULL) {
3525                 err = trace__open_output(&trace, output_name);
3526                 if (err < 0) {
3527                         perror("failed to create output file");
3528                         goto out;
3529                 }
3530         }
3531
3532         err = target__validate(&trace.opts.target);
3533         if (err) {
3534                 target__strerror(&trace.opts.target, err, bf, sizeof(bf));
3535                 fprintf(trace.output, "%s", bf);
3536                 goto out_close;
3537         }
3538
3539         err = target__parse_uid(&trace.opts.target);
3540         if (err) {
3541                 target__strerror(&trace.opts.target, err, bf, sizeof(bf));
3542                 fprintf(trace.output, "%s", bf);
3543                 goto out_close;
3544         }
3545
3546         if (!argc && target__none(&trace.opts.target))
3547                 trace.opts.target.system_wide = true;
3548
3549         if (input_name)
3550                 err = trace__replay(&trace);
3551         else
3552                 err = trace__run(&trace, argc, argv);
3553
3554 out_close:
3555         if (output_name != NULL)
3556                 fclose(trace.output);
3557 out:
3558         return err;
3559 }