tools/perf/builtin-trace.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * builtin-trace.c
   4  *
   5  * Builtin 'trace' command:
   6  *
   7  * Display a continuously updated trace of any workload, CPU, specific PID,
   8  * system wide, etc.  Default format is loosely strace like, but any other
   9  * event may be specified using --event.
  10  *
  11  * Copyright (C) 2012, 2013, 2014, 2015 Red Hat Inc, Arnaldo Carvalho de Melo <acme@redhat.com>
  12  *
  13  * Initially based on the 'trace' prototype by Thomas Gleixner:
  14  *
  15  * http://lwn.net/Articles/415728/ ("Announcing a new utility: 'trace'")
  16  */
  17
  18 #include <traceevent/event-parse.h>
  19 #include <api/fs/tracing_path.h>
  20 #include <bpf/bpf.h>
  21 #include "util/bpf_map.h"
  22 #include "builtin.h"
  23 #include "util/cgroup.h"
  24 #include "util/color.h"
  25 #include "util/config.h"
  26 #include "util/debug.h"
  27 #include "util/env.h"
  28 #include "util/event.h"
  29 #include "util/evlist.h"
  30 #include <subcmd/exec-cmd.h>
  31 #include "util/machine.h"
  32 #include "util/map.h"
  33 #include "util/symbol.h"
  34 #include "util/path.h"
  35 #include "util/session.h"
  36 #include "util/thread.h"
  37 #include <subcmd/parse-options.h>
  38 #include "util/strlist.h"
  39 #include "util/intlist.h"
  40 #include "util/thread_map.h"
  41 #include "util/stat.h"
  42 #include "trace/beauty/beauty.h"
  43 #include "trace-event.h"
  44 #include "util/parse-events.h"
  45 #include "util/bpf-loader.h"
  46 #include "callchain.h"
  47 #include "print_binary.h"
  48 #include "string2.h"
  49 #include "syscalltbl.h"
  50 #include "rb_resort.h"
  51
  52 #include <errno.h>
  53 #include <inttypes.h>
  54 #include <poll.h>
  55 #include <signal.h>
  56 #include <stdlib.h>
  57 #include <string.h>
  58 #include <linux/err.h>
  59 #include <linux/filter.h>
  60 #include <linux/kernel.h>
  61 #include <linux/random.h>
  62 #include <linux/stringify.h>
  63 #include <linux/time64.h>
  64 #include <fcntl.h>
  65 #include <sys/sysmacros.h>
  66
  67 #include "sane_ctype.h"
  68
  69 #ifndef O_CLOEXEC
  70 # define O_CLOEXEC              02000000
  71 #endif
  72
  73 #ifndef F_LINUX_SPECIFIC_BASE
  74 # define F_LINUX_SPECIFIC_BASE  1024
  75 #endif
  76
  77 struct trace {
  78         struct perf_tool        tool;
  79         struct syscalltbl       *sctbl;
  80         struct {
  81                 int             max;
  82                 struct syscall  *table;
  83                 struct bpf_map  *map;
  84                 struct {
  85                         struct perf_evsel *sys_enter,
  86                                           *sys_exit,
  87                                           *augmented;
  88                 }               events;
  89         } syscalls;
  90         struct {
  91                 struct bpf_map *map;
  92         } dump;
  93         struct record_opts      opts;
  94         struct perf_evlist      *evlist;
  95         struct machine          *host;
  96         struct thread           *current;
  97         struct cgroup           *cgroup;
  98         u64                     base_time;
  99         FILE                    *output;
 100         unsigned long           nr_events;
 101         unsigned long           nr_events_printed;
 102         unsigned long           max_events;
 103         struct strlist          *ev_qualifier;
 104         struct {
 105                 size_t          nr;
 106                 int             *entries;
 107         }                       ev_qualifier_ids;
 108         struct {
 109                 size_t          nr;
 110                 pid_t           *entries;
 111                 struct bpf_map  *map;
 112         }                       filter_pids;
 113         double                  duration_filter;
 114         double                  runtime_ms;
 115         struct {
 116                 u64             vfs_getname,
 117                                 proc_getname;
 118         } stats;
 119         unsigned int            max_stack;
 120         unsigned int            min_stack;
 121         int                     raw_augmented_syscalls_args_size;
 122         bool                    raw_augmented_syscalls;
 123         bool                    sort_events;
 124         bool                    not_ev_qualifier;
 125         bool                    live;
 126         bool                    full_time;
 127         bool                    sched;
 128         bool                    multiple_threads;
 129         bool                    summary;
 130         bool                    summary_only;
 131         bool                    failure_only;
 132         bool                    show_comm;
 133         bool                    print_sample;
 134         bool                    show_tool_stats;
 135         bool                    trace_syscalls;
 136         bool                    kernel_syscallchains;
 137         s16                     args_alignment;
 138         bool                    show_tstamp;
 139         bool                    show_duration;
 140         bool                    show_zeros;
 141         bool                    show_arg_names;
 142         bool                    show_string_prefix;
 143         bool                    force;
 144         bool                    vfs_getname;
 145         int                     trace_pgfaults;
 146         struct {
 147                 struct ordered_events   data;
 148                 u64                     last;
 149         } oe;
 150 };
 151
 152 struct tp_field {
 153         int offset;
 154         union {
 155                 u64 (*integer)(struct tp_field *field, struct perf_sample *sample);
 156                 void *(*pointer)(struct tp_field *field, struct perf_sample *sample);
 157         };
 158 };
 159
 160 #define TP_UINT_FIELD(bits) \
 161 static u64 tp_field__u##bits(struct tp_field *field, struct perf_sample *sample) \
 162 { \
 163         u##bits value; \
 164         memcpy(&value, sample->raw_data + field->offset, sizeof(value)); \
 165         return value;  \
 166 }
 167
 168 TP_UINT_FIELD(8);
 169 TP_UINT_FIELD(16);
 170 TP_UINT_FIELD(32);
 171 TP_UINT_FIELD(64);
 172
 173 #define TP_UINT_FIELD__SWAPPED(bits) \
 174 static u64 tp_field__swapped_u##bits(struct tp_field *field, struct perf_sample *sample) \
 175 { \
 176         u##bits value; \
 177         memcpy(&value, sample->raw_data + field->offset, sizeof(value)); \
 178         return bswap_##bits(value);\
 179 }
 180
 181 TP_UINT_FIELD__SWAPPED(16);
 182 TP_UINT_FIELD__SWAPPED(32);
 183 TP_UINT_FIELD__SWAPPED(64);
 184
 185 static int __tp_field__init_uint(struct tp_field *field, int size, int offset, bool needs_swap)
 186 {
 187         field->offset = offset;
 188
 189         switch (size) {
 190         case 1:
 191                 field->integer = tp_field__u8;
 192                 break;
 193         case 2:
 194                 field->integer = needs_swap ? tp_field__swapped_u16 : tp_field__u16;
 195                 break;
 196         case 4:
 197                 field->integer = needs_swap ? tp_field__swapped_u32 : tp_field__u32;
 198                 break;
 199         case 8:
 200                 field->integer = needs_swap ? tp_field__swapped_u64 : tp_field__u64;
 201                 break;
 202         default:
 203                 return -1;
 204         }
 205
 206         return 0;
 207 }
 208
 209 static int tp_field__init_uint(struct tp_field *field, struct tep_format_field *format_field, bool needs_swap)
 210 {
 211         return __tp_field__init_uint(field, format_field->size, format_field->offset, needs_swap);
 212 }
 213
 214 static void *tp_field__ptr(struct tp_field *field, struct perf_sample *sample)
 215 {
 216         return sample->raw_data + field->offset;
 217 }
 218
 219 static int __tp_field__init_ptr(struct tp_field *field, int offset)
 220 {
 221         field->offset = offset;
 222         field->pointer = tp_field__ptr;
 223         return 0;
 224 }
 225
 226 static int tp_field__init_ptr(struct tp_field *field, struct tep_format_field *format_field)
 227 {
 228         return __tp_field__init_ptr(field, format_field->offset);
 229 }
 230
 231 struct syscall_tp {
 232         struct tp_field id;
 233         union {
 234                 struct tp_field args, ret;
 235         };
 236 };
 237
 238 static int perf_evsel__init_tp_uint_field(struct perf_evsel *evsel,
 239                                           struct tp_field *field,
 240                                           const char *name)
 241 {
 242         struct tep_format_field *format_field = perf_evsel__field(evsel, name);
 243
 244         if (format_field == NULL)
 245                 return -1;
 246
 247         return tp_field__init_uint(field, format_field, evsel->needs_swap);
 248 }
 249
 250 #define perf_evsel__init_sc_tp_uint_field(evsel, name) \
 251         ({ struct syscall_tp *sc = evsel->priv;\
 252            perf_evsel__init_tp_uint_field(evsel, &sc->name, #name); })
 253
 254 static int perf_evsel__init_tp_ptr_field(struct perf_evsel *evsel,
 255                                          struct tp_field *field,
 256                                          const char *name)
 257 {
 258         struct tep_format_field *format_field = perf_evsel__field(evsel, name);
 259
 260         if (format_field == NULL)
 261                 return -1;
 262
 263         return tp_field__init_ptr(field, format_field);
 264 }
 265
 266 #define perf_evsel__init_sc_tp_ptr_field(evsel, name) \
 267         ({ struct syscall_tp *sc = evsel->priv;\
 268            perf_evsel__init_tp_ptr_field(evsel, &sc->name, #name); })
 269
 270 static void perf_evsel__delete_priv(struct perf_evsel *evsel)
 271 {
 272         zfree(&evsel->priv);
 273         perf_evsel__delete(evsel);
 274 }
 275
 276 static int perf_evsel__init_syscall_tp(struct perf_evsel *evsel)
 277 {
 278         struct syscall_tp *sc = evsel->priv = malloc(sizeof(struct syscall_tp));
 279
 280         if (evsel->priv != NULL) {
 281                 if (perf_evsel__init_tp_uint_field(evsel, &sc->id, "__syscall_nr") &&
 282                     perf_evsel__init_tp_uint_field(evsel, &sc->id, "nr"))
 283                         goto out_delete;
 284                 return 0;
 285         }
 286
 287         return -ENOMEM;
 288 out_delete:
 289         zfree(&evsel->priv);
 290         return -ENOENT;
 291 }
 292
 293 static int perf_evsel__init_augmented_syscall_tp(struct perf_evsel *evsel, struct perf_evsel *tp)
 294 {
 295         struct syscall_tp *sc = evsel->priv = malloc(sizeof(struct syscall_tp));
 296
 297         if (evsel->priv != NULL) {
 298                 struct tep_format_field *syscall_id = perf_evsel__field(tp, "id");
 299                 if (syscall_id == NULL)
 300                         syscall_id = perf_evsel__field(tp, "__syscall_nr");
 301                 if (syscall_id == NULL)
 302                         goto out_delete;
 303                 if (__tp_field__init_uint(&sc->id, syscall_id->size, syscall_id->offset, evsel->needs_swap))
 304                         goto out_delete;
 305
 306                 return 0;
 307         }
 308
 309         return -ENOMEM;
 310 out_delete:
 311         zfree(&evsel->priv);
 312         return -EINVAL;
 313 }
 314
 315 static int perf_evsel__init_augmented_syscall_tp_args(struct perf_evsel *evsel)
 316 {
 317         struct syscall_tp *sc = evsel->priv;
 318
 319         return __tp_field__init_ptr(&sc->args, sc->id.offset + sizeof(u64));
 320 }
 321
 322 static int perf_evsel__init_augmented_syscall_tp_ret(struct perf_evsel *evsel)
 323 {
 324         struct syscall_tp *sc = evsel->priv;
 325
 326         return __tp_field__init_uint(&sc->ret, sizeof(u64), sc->id.offset + sizeof(u64), evsel->needs_swap);
 327 }
 328
 329 static int perf_evsel__init_raw_syscall_tp(struct perf_evsel *evsel, void *handler)
 330 {
 331         evsel->priv = malloc(sizeof(struct syscall_tp));
 332         if (evsel->priv != NULL) {
 333                 if (perf_evsel__init_sc_tp_uint_field(evsel, id))
 334                         goto out_delete;
 335
 336                 evsel->handler = handler;
 337                 return 0;
 338         }
 339
 340         return -ENOMEM;
 341
 342 out_delete:
 343         zfree(&evsel->priv);
 344         return -ENOENT;
 345 }
 346
 347 static struct perf_evsel *perf_evsel__raw_syscall_newtp(const char *direction, void *handler)
 348 {
 349         struct perf_evsel *evsel = perf_evsel__newtp("raw_syscalls", direction);
 350
 351         /* older kernel (e.g., RHEL6) use syscalls:{enter,exit} */
 352         if (IS_ERR(evsel))
 353                 evsel = perf_evsel__newtp("syscalls", direction);
 354
 355         if (IS_ERR(evsel))
 356                 return NULL;
 357
 358         if (perf_evsel__init_raw_syscall_tp(evsel, handler))
 359                 goto out_delete;
 360
 361         return evsel;
 362
 363 out_delete:
 364         perf_evsel__delete_priv(evsel);
 365         return NULL;
 366 }
 367
 368 #define perf_evsel__sc_tp_uint(evsel, name, sample) \
 369         ({ struct syscall_tp *fields = evsel->priv; \
 370            fields->name.integer(&fields->name, sample); })
 371
 372 #define perf_evsel__sc_tp_ptr(evsel, name, sample) \
 373         ({ struct syscall_tp *fields = evsel->priv; \
 374            fields->name.pointer(&fields->name, sample); })
 375
 376 size_t strarray__scnprintf(struct strarray *sa, char *bf, size_t size, const char *intfmt, bool show_prefix, int val)
 377 {
 378         int idx = val - sa->offset;
 379
 380         if (idx < 0 || idx >= sa->nr_entries || sa->entries[idx] == NULL) {
 381                 size_t printed = scnprintf(bf, size, intfmt, val);
 382                 if (show_prefix)
 383                         printed += scnprintf(bf + printed, size - printed, " /* %s??? */", sa->prefix);
 384                 return printed;
 385         }
 386
 387         return scnprintf(bf, size, "%s%s", show_prefix ? sa->prefix : "", sa->entries[idx]);
 388 }
 389
 390 static size_t __syscall_arg__scnprintf_strarray(char *bf, size_t size,
 391                                                 const char *intfmt,
 392                                                 struct syscall_arg *arg)
 393 {
 394         return strarray__scnprintf(arg->parm, bf, size, intfmt, arg->show_string_prefix, arg->val);
 395 }
 396
 397 static size_t syscall_arg__scnprintf_strarray(char *bf, size_t size,
 398                                               struct syscall_arg *arg)
 399 {
 400         return __syscall_arg__scnprintf_strarray(bf, size, "%d", arg);
 401 }
 402
 403 #define SCA_STRARRAY syscall_arg__scnprintf_strarray
 404
 405 size_t strarrays__scnprintf(struct strarrays *sas, char *bf, size_t size, const char *intfmt, bool show_prefix, int val)
 406 {
 407         size_t printed;
 408         int i;
 409
 410         for (i = 0; i < sas->nr_entries; ++i) {
 411                 struct strarray *sa = sas->entries[i];
 412                 int idx = val - sa->offset;
 413
 414                 if (idx >= 0 && idx < sa->nr_entries) {
 415                         if (sa->entries[idx] == NULL)
 416                                 break;
 417                         return scnprintf(bf, size, "%s%s", show_prefix ? sa->prefix : "", sa->entries[idx]);
 418                 }
 419         }
 420
 421         printed = scnprintf(bf, size, intfmt, val);
 422         if (show_prefix)
 423                 printed += scnprintf(bf + printed, size - printed, " /* %s??? */", sas->entries[0]->prefix);
 424         return printed;
 425 }
 426
 427 size_t syscall_arg__scnprintf_strarrays(char *bf, size_t size,
 428                                         struct syscall_arg *arg)
 429 {
 430         return strarrays__scnprintf(arg->parm, bf, size, "%d", arg->show_string_prefix, arg->val);
 431 }
 432
 433 #ifndef AT_FDCWD
 434 #define AT_FDCWD        -100
 435 #endif
 436
 437 static size_t syscall_arg__scnprintf_fd_at(char *bf, size_t size,
 438                                            struct syscall_arg *arg)
 439 {
 440         int fd = arg->val;
 441         const char *prefix = "AT_FD";
 442
 443         if (fd == AT_FDCWD)
 444                 return scnprintf(bf, size, "%s%s", arg->show_string_prefix ? prefix : "", "CWD");
 445
 446         return syscall_arg__scnprintf_fd(bf, size, arg);
 447 }
 448
 449 #define SCA_FDAT syscall_arg__scnprintf_fd_at
 450
 451 static size_t syscall_arg__scnprintf_close_fd(char *bf, size_t size,
 452                                               struct syscall_arg *arg);
 453
 454 #define SCA_CLOSE_FD syscall_arg__scnprintf_close_fd
 455
 456 size_t syscall_arg__scnprintf_hex(char *bf, size_t size, struct syscall_arg *arg)
 457 {
 458         return scnprintf(bf, size, "%#lx", arg->val);
 459 }
 460
 461 size_t syscall_arg__scnprintf_ptr(char *bf, size_t size, struct syscall_arg *arg)
 462 {
 463         if (arg->val == 0)
 464                 return scnprintf(bf, size, "NULL");
 465         return syscall_arg__scnprintf_hex(bf, size, arg);
 466 }
 467
 468 size_t syscall_arg__scnprintf_int(char *bf, size_t size, struct syscall_arg *arg)
 469 {
 470         return scnprintf(bf, size, "%d", arg->val);
 471 }
 472
 473 size_t syscall_arg__scnprintf_long(char *bf, size_t size, struct syscall_arg *arg)
 474 {
 475         return scnprintf(bf, size, "%ld", arg->val);
 476 }
 477
 478 static const char *bpf_cmd[] = {
 479         "MAP_CREATE", "MAP_LOOKUP_ELEM", "MAP_UPDATE_ELEM", "MAP_DELETE_ELEM",
 480         "MAP_GET_NEXT_KEY", "PROG_LOAD",
 481 };
 482 static DEFINE_STRARRAY(bpf_cmd, "BPF_");
 483
 484 static const char *epoll_ctl_ops[] = { "ADD", "DEL", "MOD", };
 485 static DEFINE_STRARRAY_OFFSET(epoll_ctl_ops, "EPOLL_CTL_", 1);
 486
 487 static const char *itimers[] = { "REAL", "VIRTUAL", "PROF", };
 488 static DEFINE_STRARRAY(itimers, "ITIMER_");
 489
 490 static const char *keyctl_options[] = {
 491         "GET_KEYRING_ID", "JOIN_SESSION_KEYRING", "UPDATE", "REVOKE", "CHOWN",
 492         "SETPERM", "DESCRIBE", "CLEAR", "LINK", "UNLINK", "SEARCH", "READ",
 493         "INSTANTIATE", "NEGATE", "SET_REQKEY_KEYRING", "SET_TIMEOUT",
 494         "ASSUME_AUTHORITY", "GET_SECURITY", "SESSION_TO_PARENT", "REJECT",
 495         "INSTANTIATE_IOV", "INVALIDATE", "GET_PERSISTENT",
 496 };
 497 static DEFINE_STRARRAY(keyctl_options, "KEYCTL_");
 498
 499 static const char *whences[] = { "SET", "CUR", "END",
 500 #ifdef SEEK_DATA
 501 "DATA",
 502 #endif
 503 #ifdef SEEK_HOLE
 504 "HOLE",
 505 #endif
 506 };
 507 static DEFINE_STRARRAY(whences, "SEEK_");
 508
 509 static const char *fcntl_cmds[] = {
 510         "DUPFD", "GETFD", "SETFD", "GETFL", "SETFL", "GETLK", "SETLK",
 511         "SETLKW", "SETOWN", "GETOWN", "SETSIG", "GETSIG", "GETLK64",
 512         "SETLK64", "SETLKW64", "SETOWN_EX", "GETOWN_EX",
 513         "GETOWNER_UIDS",
 514 };
 515 static DEFINE_STRARRAY(fcntl_cmds, "F_");
 516
 517 static const char *fcntl_linux_specific_cmds[] = {
 518         "SETLEASE", "GETLEASE", "NOTIFY", [5] = "CANCELLK", "DUPFD_CLOEXEC",
 519         "SETPIPE_SZ", "GETPIPE_SZ", "ADD_SEALS", "GET_SEALS",
 520         "GET_RW_HINT", "SET_RW_HINT", "GET_FILE_RW_HINT", "SET_FILE_RW_HINT",
 521 };
 522
 523 static DEFINE_STRARRAY_OFFSET(fcntl_linux_specific_cmds, "F_", F_LINUX_SPECIFIC_BASE);
 524
 525 static struct strarray *fcntl_cmds_arrays[] = {
 526         &strarray__fcntl_cmds,
 527         &strarray__fcntl_linux_specific_cmds,
 528 };
 529
 530 static DEFINE_STRARRAYS(fcntl_cmds_arrays);
 531
 532 static const char *rlimit_resources[] = {
 533         "CPU", "FSIZE", "DATA", "STACK", "CORE", "RSS", "NPROC", "NOFILE",
 534         "MEMLOCK", "AS", "LOCKS", "SIGPENDING", "MSGQUEUE", "NICE", "RTPRIO",
 535         "RTTIME",
 536 };
 537 static DEFINE_STRARRAY(rlimit_resources, "RLIMIT_");
 538
 539 static const char *sighow[] = { "BLOCK", "UNBLOCK", "SETMASK", };
 540 static DEFINE_STRARRAY(sighow, "SIG_");
 541
 542 static const char *clockid[] = {
 543         "REALTIME", "MONOTONIC", "PROCESS_CPUTIME_ID", "THREAD_CPUTIME_ID",
 544         "MONOTONIC_RAW", "REALTIME_COARSE", "MONOTONIC_COARSE", "BOOTTIME",
 545         "REALTIME_ALARM", "BOOTTIME_ALARM", "SGI_CYCLE", "TAI"
 546 };
 547 static DEFINE_STRARRAY(clockid, "CLOCK_");
 548
 549 static size_t syscall_arg__scnprintf_access_mode(char *bf, size_t size,
 550                                                  struct syscall_arg *arg)
 551 {
 552         bool show_prefix = arg->show_string_prefix;
 553         const char *suffix = "_OK";
 554         size_t printed = 0;
 555         int mode = arg->val;
 556
 557         if (mode == F_OK) /* 0 */
 558                 return scnprintf(bf, size, "F%s", show_prefix ? suffix : "");
 559 #define P_MODE(n) \
 560         if (mode & n##_OK) { \
 561                 printed += scnprintf(bf + printed, size - printed, "%s%s", #n, show_prefix ? suffix : ""); \
 562                 mode &= ~n##_OK; \
 563         }
 564
 565         P_MODE(R);
 566         P_MODE(W);
 567         P_MODE(X);
 568 #undef P_MODE
 569
 570         if (mode)
 571                 printed += scnprintf(bf + printed, size - printed, "|%#x", mode);
 572
 573         return printed;
 574 }
 575
 576 #define SCA_ACCMODE syscall_arg__scnprintf_access_mode
 577
 578 static size_t syscall_arg__scnprintf_filename(char *bf, size_t size,
 579                                               struct syscall_arg *arg);
 580
 581 #define SCA_FILENAME syscall_arg__scnprintf_filename
 582
 583 static size_t syscall_arg__scnprintf_pipe_flags(char *bf, size_t size,
 584                                                 struct syscall_arg *arg)
 585 {
 586         bool show_prefix = arg->show_string_prefix;
 587         const char *prefix = "O_";
 588         int printed = 0, flags = arg->val;
 589
 590 #define P_FLAG(n) \
 591         if (flags & O_##n) { \
 592                 printed += scnprintf(bf + printed, size - printed, "%s%s%s", printed ? "|" : "", show_prefix ? prefix : "", #n); \
 593                 flags &= ~O_##n; \
 594         }
 595
 596         P_FLAG(CLOEXEC);
 597         P_FLAG(NONBLOCK);
 598 #undef P_FLAG
 599
 600         if (flags)
 601                 printed += scnprintf(bf + printed, size - printed, "%s%#x", printed ? "|" : "", flags);
 602
 603         return printed;
 604 }
 605
 606 #define SCA_PIPE_FLAGS syscall_arg__scnprintf_pipe_flags
 607
 608 #ifndef GRND_NONBLOCK
 609 #define GRND_NONBLOCK   0x0001
 610 #endif
 611 #ifndef GRND_RANDOM
 612 #define GRND_RANDOM     0x0002
 613 #endif
 614
 615 static size_t syscall_arg__scnprintf_getrandom_flags(char *bf, size_t size,
 616                                                    struct syscall_arg *arg)
 617 {
 618         bool show_prefix = arg->show_string_prefix;
 619         const char *prefix = "GRND_";
 620         int printed = 0, flags = arg->val;
 621
 622 #define P_FLAG(n) \
 623         if (flags & GRND_##n) { \
 624                 printed += scnprintf(bf + printed, size - printed, "%s%s%s", printed ? "|" : "", show_prefix ? prefix : "", #n); \
 625                 flags &= ~GRND_##n; \
 626         }
 627
 628         P_FLAG(RANDOM);
 629         P_FLAG(NONBLOCK);
 630 #undef P_FLAG
 631
 632         if (flags)
 633                 printed += scnprintf(bf + printed, size - printed, "%s%#x", printed ? "|" : "", flags);
 634
 635         return printed;
 636 }
 637
 638 #define SCA_GETRANDOM_FLAGS syscall_arg__scnprintf_getrandom_flags
 639
 640 #define STRARRAY(name, array) \
 641           { .scnprintf  = SCA_STRARRAY, \
 642             .parm       = &strarray__##array, }
 643
 644 #include "trace/beauty/arch_errno_names.c"
 645 #include "trace/beauty/eventfd.c"
 646 #include "trace/beauty/futex_op.c"
 647 #include "trace/beauty/futex_val3.c"
 648 #include "trace/beauty/mmap.c"
 649 #include "trace/beauty/mode_t.c"
 650 #include "trace/beauty/msg_flags.c"
 651 #include "trace/beauty/open_flags.c"
 652 #include "trace/beauty/perf_event_open.c"
 653 #include "trace/beauty/pid.c"
 654 #include "trace/beauty/sched_policy.c"
 655 #include "trace/beauty/seccomp.c"
 656 #include "trace/beauty/signum.c"
 657 #include "trace/beauty/socket_type.c"
 658 #include "trace/beauty/waitid_options.c"
 659
 660 struct syscall_arg_fmt {
 661         size_t     (*scnprintf)(char *bf, size_t size, struct syscall_arg *arg);
 662         unsigned long (*mask_val)(struct syscall_arg *arg, unsigned long val);
 663         void       *parm;
 664         const char *name;
 665         bool       show_zero;
 666 };
 667
 668 static struct syscall_fmt {
 669         const char *name;
 670         const char *alias;
 671         struct syscall_arg_fmt arg[6];
 672         u8         nr_args;
 673         bool       errpid;
 674         bool       timeout;
 675         bool       hexret;
 676 } syscall_fmts[] = {
 677         { .name     = "access",
 678           .arg = { [1] = { .scnprintf = SCA_ACCMODE,  /* mode */ }, }, },
 679         { .name     = "arch_prctl",
 680           .arg = { [0] = { .scnprintf = SCA_X86_ARCH_PRCTL_CODE, /* code */ },
 681                    [1] = { .scnprintf = SCA_PTR, /* arg2 */ }, }, },
 682         { .name     = "bind",
 683           .arg = { [1] = { .scnprintf = SCA_SOCKADDR, /* umyaddr */ }, }, },
 684         { .name     = "bpf",
 685           .arg = { [0] = STRARRAY(cmd, bpf_cmd), }, },
 686         { .name     = "brk",        .hexret = true,
 687           .arg = { [0] = { .scnprintf = SCA_PTR, /* brk */ }, }, },
 688         { .name     = "clock_gettime",
 689           .arg = { [0] = STRARRAY(clk_id, clockid), }, },
 690         { .name     = "clone",      .errpid = true, .nr_args = 5,
 691           .arg = { [0] = { .name = "flags",         .scnprintf = SCA_CLONE_FLAGS, },
 692                    [1] = { .name = "child_stack",   .scnprintf = SCA_HEX, },
 693                    [2] = { .name = "parent_tidptr", .scnprintf = SCA_HEX, },
 694                    [3] = { .name = "child_tidptr",  .scnprintf = SCA_HEX, },
 695                    [4] = { .name = "tls",           .scnprintf = SCA_HEX, }, }, },
 696         { .name     = "close",
 697           .arg = { [0] = { .scnprintf = SCA_CLOSE_FD, /* fd */ }, }, },
 698         { .name     = "connect",
 699           .arg = { [1] = { .scnprintf = SCA_SOCKADDR, /* servaddr */ }, }, },
 700         { .name     = "epoll_ctl",
 701           .arg = { [1] = STRARRAY(op, epoll_ctl_ops), }, },
 702         { .name     = "eventfd2",
 703           .arg = { [1] = { .scnprintf = SCA_EFD_FLAGS, /* flags */ }, }, },
 704         { .name     = "fchmodat",
 705           .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
 706         { .name     = "fchownat",
 707           .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
 708         { .name     = "fcntl",
 709           .arg = { [1] = { .scnprintf = SCA_FCNTL_CMD, /* cmd */
 710                            .parm      = &strarrays__fcntl_cmds_arrays,
 711                            .show_zero = true, },
 712                    [2] = { .scnprintf =  SCA_FCNTL_ARG, /* arg */ }, }, },
 713         { .name     = "flock",
 714           .arg = { [1] = { .scnprintf = SCA_FLOCK, /* cmd */ }, }, },
 715         { .name     = "fstat", .alias = "newfstat", },
 716         { .name     = "fstatat", .alias = "newfstatat", },
 717         { .name     = "futex",
 718           .arg = { [1] = { .scnprintf = SCA_FUTEX_OP, /* op */ },
 719                    [5] = { .scnprintf = SCA_FUTEX_VAL3, /* val3 */ }, }, },
 720         { .name     = "futimesat",
 721           .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
 722         { .name     = "getitimer",
 723           .arg = { [0] = STRARRAY(which, itimers), }, },
 724         { .name     = "getpid",     .errpid = true, },
 725         { .name     = "getpgid",    .errpid = true, },
 726         { .name     = "getppid",    .errpid = true, },
 727         { .name     = "getrandom",
 728           .arg = { [2] = { .scnprintf = SCA_GETRANDOM_FLAGS, /* flags */ }, }, },
 729         { .name     = "getrlimit",
 730           .arg = { [0] = STRARRAY(resource, rlimit_resources), }, },
 731         { .name     = "gettid",     .errpid = true, },
 732         { .name     = "ioctl",
 733           .arg = {
 734 #if defined(__i386__) || defined(__x86_64__)
 735 /*
 736  * FIXME: Make this available to all arches.
 737  */
 738                    [1] = { .scnprintf = SCA_IOCTL_CMD, /* cmd */ },
 739                    [2] = { .scnprintf = SCA_HEX, /* arg */ }, }, },
 740 #else
 741                    [2] = { .scnprintf = SCA_HEX, /* arg */ }, }, },
 742 #endif
 743         { .name     = "kcmp",       .nr_args = 5,
 744           .arg = { [0] = { .name = "pid1",      .scnprintf = SCA_PID, },
 745                    [1] = { .name = "pid2",      .scnprintf = SCA_PID, },
 746                    [2] = { .name = "type",      .scnprintf = SCA_KCMP_TYPE, },
 747                    [3] = { .name = "idx1",      .scnprintf = SCA_KCMP_IDX, },
 748                    [4] = { .name = "idx2",      .scnprintf = SCA_KCMP_IDX, }, }, },
 749         { .name     = "keyctl",
 750           .arg = { [0] = STRARRAY(option, keyctl_options), }, },
 751         { .name     = "kill",
 752           .arg = { [1] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
 753         { .name     = "linkat",
 754           .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
 755         { .name     = "lseek",
 756           .arg = { [2] = STRARRAY(whence, whences), }, },
 757         { .name     = "lstat", .alias = "newlstat", },
 758         { .name     = "madvise",
 759           .arg = { [0] = { .scnprintf = SCA_HEX,      /* start */ },
 760                    [2] = { .scnprintf = SCA_MADV_BHV, /* behavior */ }, }, },
 761         { .name     = "mkdirat",
 762           .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
 763         { .name     = "mknodat",
 764           .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
 765         { .name     = "mmap",       .hexret = true,
 766 /* The standard mmap maps to old_mmap on s390x */
 767 #if defined(__s390x__)
 768         .alias = "old_mmap",
 769 #endif
 770           .arg = { [2] = { .scnprintf = SCA_MMAP_PROT,  /* prot */ },
 771                    [3] = { .scnprintf = SCA_MMAP_FLAGS, /* flags */ },
 772                    [5] = { .scnprintf = SCA_HEX,        /* offset */ }, }, },
 773         { .name     = "mount",
 774           .arg = { [0] = { .scnprintf = SCA_FILENAME, /* dev_name */ },
 775                    [3] = { .scnprintf = SCA_MOUNT_FLAGS, /* flags */
 776                            .mask_val  = SCAMV_MOUNT_FLAGS, /* flags */ }, }, },
 777         { .name     = "mprotect",
 778           .arg = { [0] = { .scnprintf = SCA_HEX,        /* start */ },
 779                    [2] = { .scnprintf = SCA_MMAP_PROT,  /* prot */ }, }, },
 780         { .name     = "mq_unlink",
 781           .arg = { [0] = { .scnprintf = SCA_FILENAME, /* u_name */ }, }, },
 782         { .name     = "mremap",     .hexret = true,
 783           .arg = { [3] = { .scnprintf = SCA_MREMAP_FLAGS, /* flags */ }, }, },
 784         { .name     = "name_to_handle_at",
 785           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
 786         { .name     = "newfstatat",
 787           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
 788         { .name     = "open",
 789           .arg = { [1] = { .scnprintf = SCA_OPEN_FLAGS, /* flags */ }, }, },
 790         { .name     = "open_by_handle_at",
 791           .arg = { [0] = { .scnprintf = SCA_FDAT,       /* dfd */ },
 792                    [2] = { .scnprintf = SCA_OPEN_FLAGS, /* flags */ }, }, },
 793         { .name     = "openat",
 794           .arg = { [0] = { .scnprintf = SCA_FDAT,       /* dfd */ },
 795                    [2] = { .scnprintf = SCA_OPEN_FLAGS, /* flags */ }, }, },
 796         { .name     = "perf_event_open",
 797           .arg = { [2] = { .scnprintf = SCA_INT,        /* cpu */ },
 798                    [3] = { .scnprintf = SCA_FD,         /* group_fd */ },
 799                    [4] = { .scnprintf = SCA_PERF_FLAGS, /* flags */ }, }, },
 800         { .name     = "pipe2",
 801           .arg = { [1] = { .scnprintf = SCA_PIPE_FLAGS, /* flags */ }, }, },
 802         { .name     = "pkey_alloc",
 803           .arg = { [1] = { .scnprintf = SCA_PKEY_ALLOC_ACCESS_RIGHTS,   /* access_rights */ }, }, },
 804         { .name     = "pkey_free",
 805           .arg = { [0] = { .scnprintf = SCA_INT,        /* key */ }, }, },
 806         { .name     = "pkey_mprotect",
 807           .arg = { [0] = { .scnprintf = SCA_HEX,        /* start */ },
 808                    [2] = { .scnprintf = SCA_MMAP_PROT,  /* prot */ },
 809                    [3] = { .scnprintf = SCA_INT,        /* pkey */ }, }, },
 810         { .name     = "poll", .timeout = true, },
 811         { .name     = "ppoll", .timeout = true, },
 812         { .name     = "prctl",
 813           .arg = { [0] = { .scnprintf = SCA_PRCTL_OPTION, /* option */ },
 814                    [1] = { .scnprintf = SCA_PRCTL_ARG2, /* arg2 */ },
 815                    [2] = { .scnprintf = SCA_PRCTL_ARG3, /* arg3 */ }, }, },
 816         { .name     = "pread", .alias = "pread64", },
 817         { .name     = "preadv", .alias = "pread", },
 818         { .name     = "prlimit64",
 819           .arg = { [1] = STRARRAY(resource, rlimit_resources), }, },
 820         { .name     = "pwrite", .alias = "pwrite64", },
 821         { .name     = "readlinkat",
 822           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
 823         { .name     = "recvfrom",
 824           .arg = { [3] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
 825         { .name     = "recvmmsg",
 826           .arg = { [3] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
 827         { .name     = "recvmsg",
 828           .arg = { [2] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
 829         { .name     = "renameat",
 830           .arg = { [0] = { .scnprintf = SCA_FDAT, /* olddirfd */ },
 831                    [2] = { .scnprintf = SCA_FDAT, /* newdirfd */ }, }, },
 832         { .name     = "renameat2",
 833           .arg = { [0] = { .scnprintf = SCA_FDAT, /* olddirfd */ },
 834                    [2] = { .scnprintf = SCA_FDAT, /* newdirfd */ },
 835                    [4] = { .scnprintf = SCA_RENAMEAT2_FLAGS, /* flags */ }, }, },
 836         { .name     = "rt_sigaction",
 837           .arg = { [0] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
 838         { .name     = "rt_sigprocmask",
 839           .arg = { [0] = STRARRAY(how, sighow), }, },
 840         { .name     = "rt_sigqueueinfo",
 841           .arg = { [1] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
 842         { .name     = "rt_tgsigqueueinfo",
 843           .arg = { [2] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
 844         { .name     = "sched_setscheduler",
 845           .arg = { [1] = { .scnprintf = SCA_SCHED_POLICY, /* policy */ }, }, },
 846         { .name     = "seccomp",
 847           .arg = { [0] = { .scnprintf = SCA_SECCOMP_OP,    /* op */ },
 848                    [1] = { .scnprintf = SCA_SECCOMP_FLAGS, /* flags */ }, }, },
 849         { .name     = "select", .timeout = true, },
 850         { .name     = "sendmmsg",
 851           .arg = { [3] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
 852         { .name     = "sendmsg",
 853           .arg = { [2] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
 854         { .name     = "sendto",
 855           .arg = { [3] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ },
 856                    [4] = { .scnprintf = SCA_SOCKADDR, /* addr */ }, }, },
 857         { .name     = "set_tid_address", .errpid = true, },
 858         { .name     = "setitimer",
 859           .arg = { [0] = STRARRAY(which, itimers), }, },
 860         { .name     = "setrlimit",
 861           .arg = { [0] = STRARRAY(resource, rlimit_resources), }, },
 862         { .name     = "socket",
 863           .arg = { [0] = STRARRAY(family, socket_families),
 864                    [1] = { .scnprintf = SCA_SK_TYPE, /* type */ },
 865                    [2] = { .scnprintf = SCA_SK_PROTO, /* protocol */ }, }, },
 866         { .name     = "socketpair",
 867           .arg = { [0] = STRARRAY(family, socket_families),
 868                    [1] = { .scnprintf = SCA_SK_TYPE, /* type */ },
 869                    [2] = { .scnprintf = SCA_SK_PROTO, /* protocol */ }, }, },
 870         { .name     = "stat", .alias = "newstat", },
 871         { .name     = "statx",
 872           .arg = { [0] = { .scnprintf = SCA_FDAT,        /* fdat */ },
 873                    [2] = { .scnprintf = SCA_STATX_FLAGS, /* flags */ } ,
 874                    [3] = { .scnprintf = SCA_STATX_MASK,  /* mask */ }, }, },
 875         { .name     = "swapoff",
 876           .arg = { [0] = { .scnprintf = SCA_FILENAME, /* specialfile */ }, }, },
 877         { .name     = "swapon",
 878           .arg = { [0] = { .scnprintf = SCA_FILENAME, /* specialfile */ }, }, },
 879         { .name     = "symlinkat",
 880           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
 881         { .name     = "tgkill",
 882           .arg = { [2] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
 883         { .name     = "tkill",
 884           .arg = { [1] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
 885         { .name     = "umount2", .alias = "umount",
 886           .arg = { [0] = { .scnprintf = SCA_FILENAME, /* name */ }, }, },
 887         { .name     = "uname", .alias = "newuname", },
 888         { .name     = "unlinkat",
 889           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
 890         { .name     = "utimensat",
 891           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dirfd */ }, }, },
 892         { .name     = "wait4",      .errpid = true,
 893           .arg = { [2] = { .scnprintf = SCA_WAITID_OPTIONS, /* options */ }, }, },
 894         { .name     = "waitid",     .errpid = true,
 895           .arg = { [3] = { .scnprintf = SCA_WAITID_OPTIONS, /* options */ }, }, },
 896 };
 897
 898 static int syscall_fmt__cmp(const void *name, const void *fmtp)
 899 {
 900         const struct syscall_fmt *fmt = fmtp;
 901         return strcmp(name, fmt->name);
 902 }
 903
 904 static struct syscall_fmt *syscall_fmt__find(const char *name)
 905 {
 906         const int nmemb = ARRAY_SIZE(syscall_fmts);
 907         return bsearch(name, syscall_fmts, nmemb, sizeof(struct syscall_fmt), syscall_fmt__cmp);
 908 }
 909
 910 static struct syscall_fmt *syscall_fmt__find_by_alias(const char *alias)
 911 {
 912         int i, nmemb = ARRAY_SIZE(syscall_fmts);
 913
 914         for (i = 0; i < nmemb; ++i) {
 915                 if (syscall_fmts[i].alias && strcmp(syscall_fmts[i].alias, alias) == 0)
 916                         return &syscall_fmts[i];
 917         }
 918
 919         return NULL;
 920 }
 921
 922 /*
 923  * is_exit: is this "exit" or "exit_group"?
 924  * is_open: is this "open" or "openat"? To associate the fd returned in sys_exit with the pathname in sys_enter.
 925  * args_size: sum of the sizes of the syscall arguments, anything after that is augmented stuff: pathname for openat, etc.
 926  */
 927 struct syscall {
 928         struct tep_event    *tp_format;
 929         int                 nr_args;
 930         int                 args_size;
 931         bool                is_exit;
 932         bool                is_open;
 933         struct tep_format_field *args;
 934         const char          *name;
 935         struct syscall_fmt  *fmt;
 936         struct syscall_arg_fmt *arg_fmt;
 937 };
 938
 939 struct bpf_map_syscall_entry {
 940         bool    enabled;
 941 };
 942
 943 /*
 944  * We need to have this 'calculated' boolean because in some cases we really
 945  * don't know what is the duration of a syscall, for instance, when we start
 946  * a session and some threads are waiting for a syscall to finish, say 'poll',
 947  * in which case all we can do is to print "( ? ) for duration and for the
 948  * start timestamp.
 949  */
 950 static size_t fprintf_duration(unsigned long t, bool calculated, FILE *fp)
 951 {
 952         double duration = (double)t / NSEC_PER_MSEC;
 953         size_t printed = fprintf(fp, "(");
 954
 955         if (!calculated)
 956                 printed += fprintf(fp, "         ");
 957         else if (duration >= 1.0)
 958                 printed += color_fprintf(fp, PERF_COLOR_RED, "%6.3f ms", duration);
 959         else if (duration >= 0.01)
 960                 printed += color_fprintf(fp, PERF_COLOR_YELLOW, "%6.3f ms", duration);
 961         else
 962                 printed += color_fprintf(fp, PERF_COLOR_NORMAL, "%6.3f ms", duration);
 963         return printed + fprintf(fp, "): ");
 964 }
 965
 966 /**
 967  * filename.ptr: The filename char pointer that will be vfs_getname'd
 968  * filename.entry_str_pos: Where to insert the string translated from
 969  *                         filename.ptr by the vfs_getname tracepoint/kprobe.
 970  * ret_scnprintf: syscall args may set this to a different syscall return
 971  *                formatter, for instance, fcntl may return fds, file flags, etc.
 972  */
 973 struct thread_trace {
 974         u64               entry_time;
 975         bool              entry_pending;
 976         unsigned long     nr_events;
 977         unsigned long     pfmaj, pfmin;
 978         char              *entry_str;
 979         double            runtime_ms;
 980         size_t            (*ret_scnprintf)(char *bf, size_t size, struct syscall_arg *arg);
 981         struct {
 982                 unsigned long ptr;
 983                 short int     entry_str_pos;
 984                 bool          pending_open;
 985                 unsigned int  namelen;
 986                 char          *name;
 987         } filename;
 988         struct {
 989                 int           max;
 990                 struct file   *table;
 991         } files;
 992
 993         struct intlist *syscall_stats;
 994 };
 995
 996 static struct thread_trace *thread_trace__new(void)
 997 {
 998         struct thread_trace *ttrace =  zalloc(sizeof(struct thread_trace));
 999
1000         if (ttrace)
1001                 ttrace->files.max = -1;
1002
1003         ttrace->syscall_stats = intlist__new(NULL);
1004
1005         return ttrace;
1006 }
1007
1008 static struct thread_trace *thread__trace(struct thread *thread, FILE *fp)
1009 {
1010         struct thread_trace *ttrace;
1011
1012         if (thread == NULL)
1013                 goto fail;
1014
1015         if (thread__priv(thread) == NULL)
1016                 thread__set_priv(thread, thread_trace__new());
1017
1018         if (thread__priv(thread) == NULL)
1019                 goto fail;
1020
1021         ttrace = thread__priv(thread);
1022         ++ttrace->nr_events;
1023
1024         return ttrace;
1025 fail:
1026         color_fprintf(fp, PERF_COLOR_RED,
1027                       "WARNING: not enough memory, dropping samples!\n");
1028         return NULL;
1029 }
1030
1031
1032 void syscall_arg__set_ret_scnprintf(struct syscall_arg *arg,
1033                                     size_t (*ret_scnprintf)(char *bf, size_t size, struct syscall_arg *arg))
1034 {
1035         struct thread_trace *ttrace = thread__priv(arg->thread);
1036
1037         ttrace->ret_scnprintf = ret_scnprintf;
1038 }
1039
1040 #define TRACE_PFMAJ             (1 << 0)
1041 #define TRACE_PFMIN             (1 << 1)
1042
1043 static const size_t trace__entry_str_size = 2048;
1044
1045 static struct file *thread_trace__files_entry(struct thread_trace *ttrace, int fd)
1046 {
1047         if (fd < 0)
1048                 return NULL;
1049
1050         if (fd > ttrace->files.max) {
1051                 struct file *nfiles = realloc(ttrace->files.table, (fd + 1) * sizeof(struct file));
1052
1053                 if (nfiles == NULL)
1054                         return NULL;
1055
1056                 if (ttrace->files.max != -1) {
1057                         memset(nfiles + ttrace->files.max + 1, 0,
1058                                (fd - ttrace->files.max) * sizeof(struct file));
1059                 } else {
1060                         memset(nfiles, 0, (fd + 1) * sizeof(struct file));
1061                 }
1062
1063                 ttrace->files.table = nfiles;
1064                 ttrace->files.max   = fd;
1065         }
1066
1067         return ttrace->files.table + fd;
1068 }
1069
1070 struct file *thread__files_entry(struct thread *thread, int fd)
1071 {
1072         return thread_trace__files_entry(thread__priv(thread), fd);
1073 }
1074
1075 static int trace__set_fd_pathname(struct thread *thread, int fd, const char *pathname)
1076 {
1077         struct thread_trace *ttrace = thread__priv(thread);
1078         struct file *file = thread_trace__files_entry(ttrace, fd);
1079
1080         if (file != NULL) {
1081                 struct stat st;
1082                 if (stat(pathname, &st) == 0)
1083                         file->dev_maj = major(st.st_rdev);
1084                 file->pathname = strdup(pathname);
1085                 if (file->pathname)
1086                         return 0;
1087         }
1088
1089         return -1;
1090 }
1091
1092 static int thread__read_fd_path(struct thread *thread, int fd)
1093 {
1094         char linkname[PATH_MAX], pathname[PATH_MAX];
1095         struct stat st;
1096         int ret;
1097
1098         if (thread->pid_ == thread->tid) {
1099                 scnprintf(linkname, sizeof(linkname),
1100                           "/proc/%d/fd/%d", thread->pid_, fd);
1101         } else {
1102                 scnprintf(linkname, sizeof(linkname),
1103                           "/proc/%d/task/%d/fd/%d", thread->pid_, thread->tid, fd);
1104         }
1105
1106         if (lstat(linkname, &st) < 0 || st.st_size + 1 > (off_t)sizeof(pathname))
1107                 return -1;
1108
1109         ret = readlink(linkname, pathname, sizeof(pathname));
1110
1111         if (ret < 0 || ret > st.st_size)
1112                 return -1;
1113
1114         pathname[ret] = '\0';
1115         return trace__set_fd_pathname(thread, fd, pathname);
1116 }
1117
1118 static const char *thread__fd_path(struct thread *thread, int fd,
1119                                    struct trace *trace)
1120 {
1121         struct thread_trace *ttrace = thread__priv(thread);
1122
1123         if (ttrace == NULL)
1124                 return NULL;
1125
1126         if (fd < 0)
1127                 return NULL;
1128
1129         if ((fd > ttrace->files.max || ttrace->files.table[fd].pathname == NULL)) {
1130                 if (!trace->live)
1131                         return NULL;
1132                 ++trace->stats.proc_getname;
1133                 if (thread__read_fd_path(thread, fd))
1134                         return NULL;
1135         }
1136
1137         return ttrace->files.table[fd].pathname;
1138 }
1139
1140 size_t syscall_arg__scnprintf_fd(char *bf, size_t size, struct syscall_arg *arg)
1141 {
1142         int fd = arg->val;
1143         size_t printed = scnprintf(bf, size, "%d", fd);
1144         const char *path = thread__fd_path(arg->thread, fd, arg->trace);
1145
1146         if (path)
1147                 printed += scnprintf(bf + printed, size - printed, "<%s>", path);
1148
1149         return printed;
1150 }
1151
1152 size_t pid__scnprintf_fd(struct trace *trace, pid_t pid, int fd, char *bf, size_t size)
1153 {
1154         size_t printed = scnprintf(bf, size, "%d", fd);
1155         struct thread *thread = machine__find_thread(trace->host, pid, pid);
1156
1157         if (thread) {
1158                 const char *path = thread__fd_path(thread, fd, trace);
1159
1160                 if (path)
1161                         printed += scnprintf(bf + printed, size - printed, "<%s>", path);
1162
1163                 thread__put(thread);
1164         }
1165
1166         return printed;
1167 }
1168
1169 static size_t syscall_arg__scnprintf_close_fd(char *bf, size_t size,
1170                                               struct syscall_arg *arg)
1171 {
1172         int fd = arg->val;
1173         size_t printed = syscall_arg__scnprintf_fd(bf, size, arg);
1174         struct thread_trace *ttrace = thread__priv(arg->thread);
1175
1176         if (ttrace && fd >= 0 && fd <= ttrace->files.max)
1177                 zfree(&ttrace->files.table[fd].pathname);
1178
1179         return printed;
1180 }
1181
1182 static void thread__set_filename_pos(struct thread *thread, const char *bf,
1183                                      unsigned long ptr)
1184 {
1185         struct thread_trace *ttrace = thread__priv(thread);
1186
1187         ttrace->filename.ptr = ptr;
1188         ttrace->filename.entry_str_pos = bf - ttrace->entry_str;
1189 }
1190
1191 static size_t syscall_arg__scnprintf_augmented_string(struct syscall_arg *arg, char *bf, size_t size)
1192 {
1193         struct augmented_arg *augmented_arg = arg->augmented.args;
1194
1195         return scnprintf(bf, size, "\"%.*s\"", augmented_arg->size, augmented_arg->value);
1196 }
1197
1198 static size_t syscall_arg__scnprintf_filename(char *bf, size_t size,
1199                                               struct syscall_arg *arg)
1200 {
1201         unsigned long ptr = arg->val;
1202
1203         if (arg->augmented.args)
1204                 return syscall_arg__scnprintf_augmented_string(arg, bf, size);
1205
1206         if (!arg->trace->vfs_getname)
1207                 return scnprintf(bf, size, "%#x", ptr);
1208
1209         thread__set_filename_pos(arg->thread, bf, ptr);
1210         return 0;
1211 }
1212
1213 static bool trace__filter_duration(struct trace *trace, double t)
1214 {
1215         return t < (trace->duration_filter * NSEC_PER_MSEC);
1216 }
1217
1218 static size_t __trace__fprintf_tstamp(struct trace *trace, u64 tstamp, FILE *fp)
1219 {
1220         double ts = (double)(tstamp - trace->base_time) / NSEC_PER_MSEC;
1221
1222         return fprintf(fp, "%10.3f ", ts);
1223 }
1224
1225 /*
1226  * We're handling tstamp=0 as an undefined tstamp, i.e. like when we are
1227  * using ttrace->entry_time for a thread that receives a sys_exit without
1228  * first having received a sys_enter ("poll" issued before tracing session
1229  * starts, lost sys_enter exit due to ring buffer overflow).
1230  */
1231 static size_t trace__fprintf_tstamp(struct trace *trace, u64 tstamp, FILE *fp)
1232 {
1233         if (tstamp > 0)
1234                 return __trace__fprintf_tstamp(trace, tstamp, fp);
1235
1236         return fprintf(fp, "         ? ");
1237 }
1238
1239 static bool done = false;
1240 static bool interrupted = false;
1241
1242 static void sig_handler(int sig)
1243 {
1244         done = true;
1245         interrupted = sig == SIGINT;
1246 }
1247
1248 static size_t trace__fprintf_comm_tid(struct trace *trace, struct thread *thread, FILE *fp)
1249 {
1250         size_t printed = 0;
1251
1252         if (trace->multiple_threads) {
1253                 if (trace->show_comm)
1254                         printed += fprintf(fp, "%.14s/", thread__comm_str(thread));
1255                 printed += fprintf(fp, "%d ", thread->tid);
1256         }
1257
1258         return printed;
1259 }
1260
1261 static size_t trace__fprintf_entry_head(struct trace *trace, struct thread *thread,
1262                                         u64 duration, bool duration_calculated, u64 tstamp, FILE *fp)
1263 {
1264         size_t printed = 0;
1265
1266         if (trace->show_tstamp)
1267                 printed = trace__fprintf_tstamp(trace, tstamp, fp);
1268         if (trace->show_duration)
1269                 printed += fprintf_duration(duration, duration_calculated, fp);
1270         return printed + trace__fprintf_comm_tid(trace, thread, fp);
1271 }
1272
1273 static int trace__process_event(struct trace *trace, struct machine *machine,
1274                                 union perf_event *event, struct perf_sample *sample)
1275 {
1276         int ret = 0;
1277
1278         switch (event->header.type) {
1279         case PERF_RECORD_LOST:
1280                 color_fprintf(trace->output, PERF_COLOR_RED,
1281                               "LOST %" PRIu64 " events!\n", event->lost.lost);
1282                 ret = machine__process_lost_event(machine, event, sample);
1283                 break;
1284         default:
1285                 ret = machine__process_event(machine, event, sample);
1286                 break;
1287         }
1288
1289         return ret;
1290 }
1291
1292 static int trace__tool_process(struct perf_tool *tool,
1293                                union perf_event *event,
1294                                struct perf_sample *sample,
1295                                struct machine *machine)
1296 {
1297         struct trace *trace = container_of(tool, struct trace, tool);
1298         return trace__process_event(trace, machine, event, sample);
1299 }
1300
1301 static char *trace__machine__resolve_kernel_addr(void *vmachine, unsigned long long *addrp, char **modp)
1302 {
1303         struct machine *machine = vmachine;
1304
1305         if (machine->kptr_restrict_warned)
1306                 return NULL;
1307
1308         if (symbol_conf.kptr_restrict) {
1309                 pr_warning("Kernel address maps (/proc/{kallsyms,modules}) are restricted.\n\n"
1310                            "Check /proc/sys/kernel/kptr_restrict.\n\n"
1311                            "Kernel samples will not be resolved.\n");
1312                 machine->kptr_restrict_warned = true;
1313                 return NULL;
1314         }
1315
1316         return machine__resolve_kernel_addr(vmachine, addrp, modp);
1317 }
1318
1319 static int trace__symbols_init(struct trace *trace, struct perf_evlist *evlist)
1320 {
1321         int err = symbol__init(NULL);
1322
1323         if (err)
1324                 return err;
1325
1326         trace->host = machine__new_host();
1327         if (trace->host == NULL)
1328                 return -ENOMEM;
1329
1330         err = trace_event__register_resolver(trace->host, trace__machine__resolve_kernel_addr);
1331         if (err < 0)
1332                 goto out;
1333
1334         err = __machine__synthesize_threads(trace->host, &trace->tool, &trace->opts.target,
1335                                             evlist->threads, trace__tool_process, false,
1336                                             1);
1337 out:
1338         if (err)
1339                 symbol__exit();
1340
1341         return err;
1342 }
1343
1344 static void trace__symbols__exit(struct trace *trace)
1345 {
1346         machine__exit(trace->host);
1347         trace->host = NULL;
1348
1349         symbol__exit();
1350 }
1351
1352 static int syscall__alloc_arg_fmts(struct syscall *sc, int nr_args)
1353 {
1354         int idx;
1355
1356         if (nr_args == 6 && sc->fmt && sc->fmt->nr_args != 0)
1357                 nr_args = sc->fmt->nr_args;
1358
1359         sc->arg_fmt = calloc(nr_args, sizeof(*sc->arg_fmt));
1360         if (sc->arg_fmt == NULL)
1361                 return -1;
1362
1363         for (idx = 0; idx < nr_args; ++idx) {
1364                 if (sc->fmt)
1365                         sc->arg_fmt[idx] = sc->fmt->arg[idx];
1366         }
1367
1368         sc->nr_args = nr_args;
1369         return 0;
1370 }
1371
1372 static int syscall__set_arg_fmts(struct syscall *sc)
1373 {
1374         struct tep_format_field *field, *last_field = NULL;
1375         int idx = 0, len;
1376
1377         for (field = sc->args; field; field = field->next, ++idx) {
1378                 last_field = field;
1379
1380                 if (sc->fmt && sc->fmt->arg[idx].scnprintf)
1381                         continue;
1382
1383                 if (strcmp(field->type, "const char *") == 0 &&
1384                          (strcmp(field->name, "filename") == 0 ||
1385                           strcmp(field->name, "path") == 0 ||
1386                           strcmp(field->name, "pathname") == 0))
1387                         sc->arg_fmt[idx].scnprintf = SCA_FILENAME;
1388                 else if ((field->flags & TEP_FIELD_IS_POINTER) || strstr(field->name, "addr"))
1389                         sc->arg_fmt[idx].scnprintf = SCA_PTR;
1390                 else if (strcmp(field->type, "pid_t") == 0)
1391                         sc->arg_fmt[idx].scnprintf = SCA_PID;
1392                 else if (strcmp(field->type, "umode_t") == 0)
1393                         sc->arg_fmt[idx].scnprintf = SCA_MODE_T;
1394                 else if ((strcmp(field->type, "int") == 0 ||
1395                           strcmp(field->type, "unsigned int") == 0 ||
1396                           strcmp(field->type, "long") == 0) &&
1397                          (len = strlen(field->name)) >= 2 &&
1398                          strcmp(field->name + len - 2, "fd") == 0) {
1399                         /*
1400                          * /sys/kernel/tracing/events/syscalls/sys_enter*
1401                          * egrep 'field:.*fd;' .../format|sed -r 's/.*field:([a-z ]+) [a-z_]*fd.+/\1/g'|sort|uniq -c
1402                          * 65 int
1403                          * 23 unsigned int
1404                          * 7 unsigned long
1405                          */
1406                         sc->arg_fmt[idx].scnprintf = SCA_FD;
1407                 }
1408         }
1409
1410         if (last_field)
1411                 sc->args_size = last_field->offset + last_field->size;
1412
1413         return 0;
1414 }
1415
1416 static int trace__read_syscall_info(struct trace *trace, int id)
1417 {
1418         char tp_name[128];
1419         struct syscall *sc;
1420         const char *name = syscalltbl__name(trace->sctbl, id);
1421
1422         if (name == NULL)
1423                 return -1;
1424
1425         if (id > trace->syscalls.max) {
1426                 struct syscall *nsyscalls = realloc(trace->syscalls.table, (id + 1) * sizeof(*sc));
1427
1428                 if (nsyscalls == NULL)
1429                         return -1;
1430
1431                 if (trace->syscalls.max != -1) {
1432                         memset(nsyscalls + trace->syscalls.max + 1, 0,
1433                                (id - trace->syscalls.max) * sizeof(*sc));
1434                 } else {
1435                         memset(nsyscalls, 0, (id + 1) * sizeof(*sc));
1436                 }
1437
1438                 trace->syscalls.table = nsyscalls;
1439                 trace->syscalls.max   = id;
1440         }
1441
1442         sc = trace->syscalls.table + id;
1443         sc->name = name;
1444
1445         sc->fmt  = syscall_fmt__find(sc->name);
1446
1447         snprintf(tp_name, sizeof(tp_name), "sys_enter_%s", sc->name);
1448         sc->tp_format = trace_event__tp_format("syscalls", tp_name);
1449
1450         if (IS_ERR(sc->tp_format) && sc->fmt && sc->fmt->alias) {
1451                 snprintf(tp_name, sizeof(tp_name), "sys_enter_%s", sc->fmt->alias);
1452                 sc->tp_format = trace_event__tp_format("syscalls", tp_name);
1453         }
1454
1455         if (syscall__alloc_arg_fmts(sc, IS_ERR(sc->tp_format) ? 6 : sc->tp_format->format.nr_fields))
1456                 return -1;
1457
1458         if (IS_ERR(sc->tp_format))
1459                 return -1;
1460
1461         sc->args = sc->tp_format->format.fields;
1462         /*
1463          * We need to check and discard the first variable '__syscall_nr'
1464          * or 'nr' that mean the syscall number. It is needless here.
1465          * So drop '__syscall_nr' or 'nr' field but does not exist on older kernels.
1466          */
1467         if (sc->args && (!strcmp(sc->args->name, "__syscall_nr") || !strcmp(sc->args->name, "nr"))) {
1468                 sc->args = sc->args->next;
1469                 --sc->nr_args;
1470         }
1471
1472         sc->is_exit = !strcmp(name, "exit_group") || !strcmp(name, "exit");
1473         sc->is_open = !strcmp(name, "open") || !strcmp(name, "openat");
1474
1475         return syscall__set_arg_fmts(sc);
1476 }
1477
1478 static int trace__validate_ev_qualifier(struct trace *trace)
1479 {
1480         int err = 0, i;
1481         size_t nr_allocated;
1482         struct str_node *pos;
1483
1484         trace->ev_qualifier_ids.nr = strlist__nr_entries(trace->ev_qualifier);
1485         trace->ev_qualifier_ids.entries = malloc(trace->ev_qualifier_ids.nr *
1486                                                  sizeof(trace->ev_qualifier_ids.entries[0]));
1487
1488         if (trace->ev_qualifier_ids.entries == NULL) {
1489                 fputs("Error:\tNot enough memory for allocating events qualifier ids\n",
1490                        trace->output);
1491                 err = -EINVAL;
1492                 goto out;
1493         }
1494
1495         nr_allocated = trace->ev_qualifier_ids.nr;
1496         i = 0;
1497
1498         strlist__for_each_entry(pos, trace->ev_qualifier) {
1499                 const char *sc = pos->s;
1500                 int id = syscalltbl__id(trace->sctbl, sc), match_next = -1;
1501
1502                 if (id < 0) {
1503                         id = syscalltbl__strglobmatch_first(trace->sctbl, sc, &match_next);
1504                         if (id >= 0)
1505                                 goto matches;
1506
1507                         if (err == 0) {
1508                                 fputs("Error:\tInvalid syscall ", trace->output);
1509                                 err = -EINVAL;
1510                         } else {
1511                                 fputs(", ", trace->output);
1512                         }
1513
1514                         fputs(sc, trace->output);
1515                 }
1516 matches:
1517                 trace->ev_qualifier_ids.entries[i++] = id;
1518                 if (match_next == -1)
1519                         continue;
1520
1521                 while (1) {
1522                         id = syscalltbl__strglobmatch_next(trace->sctbl, sc, &match_next);
1523                         if (id < 0)
1524                                 break;
1525                         if (nr_allocated == trace->ev_qualifier_ids.nr) {
1526                                 void *entries;
1527
1528                                 nr_allocated += 8;
1529                                 entries = realloc(trace->ev_qualifier_ids.entries,
1530                                                   nr_allocated * sizeof(trace->ev_qualifier_ids.entries[0]));
1531                                 if (entries == NULL) {
1532                                         err = -ENOMEM;
1533                                         fputs("\nError:\t Not enough memory for parsing\n", trace->output);
1534                                         goto out_free;
1535                                 }
1536                                 trace->ev_qualifier_ids.entries = entries;
1537                         }
1538                         trace->ev_qualifier_ids.nr++;
1539                         trace->ev_qualifier_ids.entries[i++] = id;
1540                 }
1541         }
1542
1543         if (err < 0) {
1544                 fputs("\nHint:\ttry 'perf list syscalls:sys_enter_*'"
1545                       "\nHint:\tand: 'man syscalls'\n", trace->output);
1546 out_free:
1547                 zfree(&trace->ev_qualifier_ids.entries);
1548                 trace->ev_qualifier_ids.nr = 0;
1549         }
1550 out:
1551         return err;
1552 }
1553
1554 /*
1555  * args is to be interpreted as a series of longs but we need to handle
1556  * 8-byte unaligned accesses. args points to raw_data within the event
1557  * and raw_data is guaranteed to be 8-byte unaligned because it is
1558  * preceded by raw_size which is a u32. So we need to copy args to a temp
1559  * variable to read it. Most notably this avoids extended load instructions
1560  * on unaligned addresses
1561  */
1562 unsigned long syscall_arg__val(struct syscall_arg *arg, u8 idx)
1563 {
1564         unsigned long val;
1565         unsigned char *p = arg->args + sizeof(unsigned long) * idx;
1566
1567         memcpy(&val, p, sizeof(val));
1568         return val;
1569 }
1570
1571 static size_t syscall__scnprintf_name(struct syscall *sc, char *bf, size_t size,
1572                                       struct syscall_arg *arg)
1573 {
1574         if (sc->arg_fmt && sc->arg_fmt[arg->idx].name)
1575                 return scnprintf(bf, size, "%s: ", sc->arg_fmt[arg->idx].name);
1576
1577         return scnprintf(bf, size, "arg%d: ", arg->idx);
1578 }
1579
1580 /*
1581  * Check if the value is in fact zero, i.e. mask whatever needs masking, such
1582  * as mount 'flags' argument that needs ignoring some magic flag, see comment
1583  * in tools/perf/trace/beauty/mount_flags.c
1584  */
1585 static unsigned long syscall__mask_val(struct syscall *sc, struct syscall_arg *arg, unsigned long val)
1586 {
1587         if (sc->arg_fmt && sc->arg_fmt[arg->idx].mask_val)
1588                 return sc->arg_fmt[arg->idx].mask_val(arg, val);
1589
1590         return val;
1591 }
1592
1593 static size_t syscall__scnprintf_val(struct syscall *sc, char *bf, size_t size,
1594                                      struct syscall_arg *arg, unsigned long val)
1595 {
1596         if (sc->arg_fmt && sc->arg_fmt[arg->idx].scnprintf) {
1597                 arg->val = val;
1598                 if (sc->arg_fmt[arg->idx].parm)
1599                         arg->parm = sc->arg_fmt[arg->idx].parm;
1600                 return sc->arg_fmt[arg->idx].scnprintf(bf, size, arg);
1601         }
1602         return scnprintf(bf, size, "%ld", val);
1603 }
1604
1605 static size_t syscall__scnprintf_args(struct syscall *sc, char *bf, size_t size,
1606                                       unsigned char *args, void *augmented_args, int augmented_args_size,
1607                                       struct trace *trace, struct thread *thread)
1608 {
1609         size_t printed = 0;
1610         unsigned long val;
1611         u8 bit = 1;
1612         struct syscall_arg arg = {
1613                 .args   = args,
1614                 .augmented = {
1615                         .size = augmented_args_size,
1616                         .args = augmented_args,
1617                 },
1618                 .idx    = 0,
1619                 .mask   = 0,
1620                 .trace  = trace,
1621                 .thread = thread,
1622                 .show_string_prefix = trace->show_string_prefix,
1623         };
1624         struct thread_trace *ttrace = thread__priv(thread);
1625
1626         /*
1627          * Things like fcntl will set this in its 'cmd' formatter to pick the
1628          * right formatter for the return value (an fd? file flags?), which is
1629          * not needed for syscalls that always return a given type, say an fd.
1630          */
1631         ttrace->ret_scnprintf = NULL;
1632
1633         if (sc->args != NULL) {
1634                 struct tep_format_field *field;
1635
1636                 for (field = sc->args; field;
1637                      field = field->next, ++arg.idx, bit <<= 1) {
1638                         if (arg.mask & bit)
1639                                 continue;
1640
1641                         val = syscall_arg__val(&arg, arg.idx);
1642                         /*
1643                          * Some syscall args need some mask, most don't and
1644                          * return val untouched.
1645                          */
1646                         val = syscall__mask_val(sc, &arg, val);
1647
1648                         /*
1649                          * Suppress this argument if its value is zero and
1650                          * and we don't have a string associated in an
1651                          * strarray for it.
1652                          */
1653                         if (val == 0 &&
1654                             !trace->show_zeros &&
1655                             !(sc->arg_fmt &&
1656                               (sc->arg_fmt[arg.idx].show_zero ||
1657                                sc->arg_fmt[arg.idx].scnprintf == SCA_STRARRAY ||
1658                                sc->arg_fmt[arg.idx].scnprintf == SCA_STRARRAYS) &&
1659                               sc->arg_fmt[arg.idx].parm))
1660                                 continue;
1661
1662                         printed += scnprintf(bf + printed, size - printed, "%s", printed ? ", " : "");
1663
1664                         if (trace->show_arg_names)
1665                                 printed += scnprintf(bf + printed, size - printed, "%s: ", field->name);
1666
1667                         printed += syscall__scnprintf_val(sc, bf + printed, size - printed, &arg, val);
1668                 }
1669         } else if (IS_ERR(sc->tp_format)) {
1670                 /*
1671                  * If we managed to read the tracepoint /format file, then we
1672                  * may end up not having any args, like with gettid(), so only
1673                  * print the raw args when we didn't manage to read it.
1674                  */
1675                 while (arg.idx < sc->nr_args) {
1676                         if (arg.mask & bit)
1677                                 goto next_arg;
1678                         val = syscall_arg__val(&arg, arg.idx);
1679                         if (printed)
1680                                 printed += scnprintf(bf + printed, size - printed, ", ");
1681                         printed += syscall__scnprintf_name(sc, bf + printed, size - printed, &arg);
1682                         printed += syscall__scnprintf_val(sc, bf + printed, size - printed, &arg, val);
1683 next_arg:
1684                         ++arg.idx;
1685                         bit <<= 1;
1686                 }
1687         }
1688
1689         return printed;
1690 }
1691
1692 typedef int (*tracepoint_handler)(struct trace *trace, struct perf_evsel *evsel,
1693                                   union perf_event *event,
1694                                   struct perf_sample *sample);
1695
1696 static struct syscall *trace__syscall_info(struct trace *trace,
1697                                            struct perf_evsel *evsel, int id)
1698 {
1699
1700         if (id < 0) {
1701
1702                 /*
1703                  * XXX: Noticed on x86_64, reproduced as far back as 3.0.36, haven't tried
1704                  * before that, leaving at a higher verbosity level till that is
1705                  * explained. Reproduced with plain ftrace with:
1706                  *
1707                  * echo 1 > /t/events/raw_syscalls/sys_exit/enable
1708                  * grep "NR -1 " /t/trace_pipe
1709                  *
1710                  * After generating some load on the machine.
1711                  */
1712                 if (verbose > 1) {
1713                         static u64 n;
1714                         fprintf(trace->output, "Invalid syscall %d id, skipping (%s, %" PRIu64 ") ...\n",
1715                                 id, perf_evsel__name(evsel), ++n);
1716                 }
1717                 return NULL;
1718         }
1719
1720         if ((id > trace->syscalls.max || trace->syscalls.table[id].name == NULL) &&
1721             trace__read_syscall_info(trace, id))
1722                 goto out_cant_read;
1723
1724         if ((id > trace->syscalls.max || trace->syscalls.table[id].name == NULL))
1725                 goto out_cant_read;
1726
1727         return &trace->syscalls.table[id];
1728
1729 out_cant_read:
1730         if (verbose > 0) {
1731                 fprintf(trace->output, "Problems reading syscall %d", id);
1732                 if (id <= trace->syscalls.max && trace->syscalls.table[id].name != NULL)
1733                         fprintf(trace->output, "(%s)", trace->syscalls.table[id].name);
1734                 fputs(" information\n", trace->output);
1735         }
1736         return NULL;
1737 }
1738
1739 static void thread__update_stats(struct thread_trace *ttrace,
1740                                  int id, struct perf_sample *sample)
1741 {
1742         struct int_node *inode;
1743         struct stats *stats;
1744         u64 duration = 0;
1745
1746         inode = intlist__findnew(ttrace->syscall_stats, id);
1747         if (inode == NULL)
1748                 return;
1749
1750         stats = inode->priv;
1751         if (stats == NULL) {
1752                 stats = malloc(sizeof(struct stats));
1753                 if (stats == NULL)
1754                         return;
1755                 init_stats(stats);
1756                 inode->priv = stats;
1757         }
1758
1759         if (ttrace->entry_time && sample->time > ttrace->entry_time)
1760                 duration = sample->time - ttrace->entry_time;
1761
1762         update_stats(stats, duration);
1763 }
1764
1765 static int trace__printf_interrupted_entry(struct trace *trace)
1766 {
1767         struct thread_trace *ttrace;
1768         size_t printed;
1769         int len;
1770
1771         if (trace->failure_only || trace->current == NULL)
1772                 return 0;
1773
1774         ttrace = thread__priv(trace->current);
1775
1776         if (!ttrace->entry_pending)
1777                 return 0;
1778
1779         printed  = trace__fprintf_entry_head(trace, trace->current, 0, false, ttrace->entry_time, trace->output);
1780         printed += len = fprintf(trace->output, "%s)", ttrace->entry_str);
1781
1782         if (len < trace->args_alignment - 4)
1783                 printed += fprintf(trace->output, "%-*s", trace->args_alignment - 4 - len, " ");
1784
1785         printed += fprintf(trace->output, " ...\n");
1786
1787         ttrace->entry_pending = false;
1788         ++trace->nr_events_printed;
1789
1790         return printed;
1791 }
1792
1793 static int trace__fprintf_sample(struct trace *trace, struct perf_evsel *evsel,
1794                                  struct perf_sample *sample, struct thread *thread)
1795 {
1796         int printed = 0;
1797
1798         if (trace->print_sample) {
1799                 double ts = (double)sample->time / NSEC_PER_MSEC;
1800
1801                 printed += fprintf(trace->output, "%22s %10.3f %s %d/%d [%d]\n",
1802                                    perf_evsel__name(evsel), ts,
1803                                    thread__comm_str(thread),
1804                                    sample->pid, sample->tid, sample->cpu);
1805         }
1806
1807         return printed;
1808 }
1809
1810 static void *syscall__augmented_args(struct syscall *sc, struct perf_sample *sample, int *augmented_args_size, int raw_augmented_args_size)
1811 {
1812         void *augmented_args = NULL;
1813         /*
1814          * For now with BPF raw_augmented we hook into raw_syscalls:sys_enter
1815          * and there we get all 6 syscall args plus the tracepoint common fields
1816          * that gets calculated at the start and the syscall_nr (another long).
1817          * So we check if that is the case and if so don't look after the
1818          * sc->args_size but always after the full raw_syscalls:sys_enter payload,
1819          * which is fixed.
1820          *
1821          * We'll revisit this later to pass s->args_size to the BPF augmenter
1822          * (now tools/perf/examples/bpf/augmented_raw_syscalls.c, so that it
1823          * copies only what we need for each syscall, like what happens when we
1824          * use syscalls:sys_enter_NAME, so that we reduce the kernel/userspace
1825          * traffic to just what is needed for each syscall.
1826          */
1827         int args_size = raw_augmented_args_size ?: sc->args_size;
1828
1829         *augmented_args_size = sample->raw_size - args_size;
1830         if (*augmented_args_size > 0)
1831                 augmented_args = sample->raw_data + args_size;
1832
1833         return augmented_args;
1834 }
1835
1836 static int trace__sys_enter(struct trace *trace, struct perf_evsel *evsel,
1837                             union perf_event *event __maybe_unused,
1838                             struct perf_sample *sample)
1839 {
1840         char *msg;
1841         void *args;
1842         int printed = 0;
1843         struct thread *thread;
1844         int id = perf_evsel__sc_tp_uint(evsel, id, sample), err = -1;
1845         int augmented_args_size = 0;
1846         void *augmented_args = NULL;
1847         struct syscall *sc = trace__syscall_info(trace, evsel, id);
1848         struct thread_trace *ttrace;
1849
1850         if (sc == NULL)
1851                 return -1;
1852
1853         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
1854         ttrace = thread__trace(thread, trace->output);
1855         if (ttrace == NULL)
1856                 goto out_put;
1857
1858         trace__fprintf_sample(trace, evsel, sample, thread);
1859
1860         args = perf_evsel__sc_tp_ptr(evsel, args, sample);
1861
1862         if (ttrace->entry_str == NULL) {
1863                 ttrace->entry_str = malloc(trace__entry_str_size);
1864                 if (!ttrace->entry_str)
1865                         goto out_put;
1866         }
1867
1868         if (!(trace->duration_filter || trace->summary_only || trace->min_stack))
1869                 trace__printf_interrupted_entry(trace);
1870         /*
1871          * If this is raw_syscalls.sys_enter, then it always comes with the 6 possible
1872          * arguments, even if the syscall being handled, say "openat", uses only 4 arguments
1873          * this breaks syscall__augmented_args() check for augmented args, as we calculate
1874          * syscall->args_size using each syscalls:sys_enter_NAME tracefs format file,
1875          * so when handling, say the openat syscall, we end up getting 6 args for the
1876          * raw_syscalls:sys_enter event, when we expected just 4, we end up mistakenly
1877          * thinking that the extra 2 u64 args are the augmented filename, so just check
1878          * here and avoid using augmented syscalls when the evsel is the raw_syscalls one.
1879          */
1880         if (evsel != trace->syscalls.events.sys_enter)
1881                 augmented_args = syscall__augmented_args(sc, sample, &augmented_args_size, trace->raw_augmented_syscalls_args_size);
1882         ttrace->entry_time = sample->time;
1883         msg = ttrace->entry_str;
1884         printed += scnprintf(msg + printed, trace__entry_str_size - printed, "%s(", sc->name);
1885
1886         printed += syscall__scnprintf_args(sc, msg + printed, trace__entry_str_size - printed,
1887                                            args, augmented_args, augmented_args_size, trace, thread);
1888
1889         if (sc->is_exit) {
1890                 if (!(trace->duration_filter || trace->summary_only || trace->failure_only || trace->min_stack)) {
1891                         int alignment = 0;
1892
1893                         trace__fprintf_entry_head(trace, thread, 0, false, ttrace->entry_time, trace->output);
1894                         printed = fprintf(trace->output, "%s)", ttrace->entry_str);
1895                         if (trace->args_alignment > printed)
1896                                 alignment = trace->args_alignment - printed;
1897                         fprintf(trace->output, "%*s= ?\n", alignment, " ");
1898                 }
1899         } else {
1900                 ttrace->entry_pending = true;
1901                 /* See trace__vfs_getname & trace__sys_exit */
1902                 ttrace->filename.pending_open = false;
1903         }
1904
1905         if (trace->current != thread) {
1906                 thread__put(trace->current);
1907                 trace->current = thread__get(thread);
1908         }
1909         err = 0;
1910 out_put:
1911         thread__put(thread);
1912         return err;
1913 }
1914
1915 static int trace__fprintf_sys_enter(struct trace *trace, struct perf_evsel *evsel,
1916                                     struct perf_sample *sample)
1917 {
1918         struct thread_trace *ttrace;
1919         struct thread *thread;
1920         int id = perf_evsel__sc_tp_uint(evsel, id, sample), err = -1;
1921         struct syscall *sc = trace__syscall_info(trace, evsel, id);
1922         char msg[1024];
1923         void *args, *augmented_args = NULL;
1924         int augmented_args_size;
1925
1926         if (sc == NULL)
1927                 return -1;
1928
1929         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
1930         ttrace = thread__trace(thread, trace->output);
1931         /*
1932          * We need to get ttrace just to make sure it is there when syscall__scnprintf_args()
1933          * and the rest of the beautifiers accessing it via struct syscall_arg touches it.
1934          */
1935         if (ttrace == NULL)
1936                 goto out_put;
1937
1938         args = perf_evsel__sc_tp_ptr(evsel, args, sample);
1939         augmented_args = syscall__augmented_args(sc, sample, &augmented_args_size, trace->raw_augmented_syscalls_args_size);
1940         syscall__scnprintf_args(sc, msg, sizeof(msg), args, augmented_args, augmented_args_size, trace, thread);
1941         fprintf(trace->output, "%s", msg);
1942         err = 0;
1943 out_put:
1944         thread__put(thread);
1945         return err;
1946 }
1947
1948 static int trace__resolve_callchain(struct trace *trace, struct perf_evsel *evsel,
1949                                     struct perf_sample *sample,
1950                                     struct callchain_cursor *cursor)
1951 {
1952         struct addr_location al;
1953         int max_stack = evsel->attr.sample_max_stack ?
1954                         evsel->attr.sample_max_stack :
1955                         trace->max_stack;
1956         int err;
1957
1958         if (machine__resolve(trace->host, &al, sample) < 0)
1959                 return -1;
1960
1961         err = thread__resolve_callchain(al.thread, cursor, evsel, sample, NULL, NULL, max_stack);
1962         addr_location__put(&al);
1963         return err;
1964 }
1965
1966 static int trace__fprintf_callchain(struct trace *trace, struct perf_sample *sample)
1967 {
1968         /* TODO: user-configurable print_opts */
1969         const unsigned int print_opts = EVSEL__PRINT_SYM |
1970                                         EVSEL__PRINT_DSO |
1971                                         EVSEL__PRINT_UNKNOWN_AS_ADDR;
1972
1973         return sample__fprintf_callchain(sample, 38, print_opts, &callchain_cursor, trace->output);
1974 }
1975
1976 static const char *errno_to_name(struct perf_evsel *evsel, int err)
1977 {
1978         struct perf_env *env = perf_evsel__env(evsel);
1979         const char *arch_name = perf_env__arch(env);
1980
1981         return arch_syscalls__strerrno(arch_name, err);
1982 }
1983
1984 static int trace__sys_exit(struct trace *trace, struct perf_evsel *evsel,
1985                            union perf_event *event __maybe_unused,
1986                            struct perf_sample *sample)
1987 {
1988         long ret;
1989         u64 duration = 0;
1990         bool duration_calculated = false;
1991         struct thread *thread;
1992         int id = perf_evsel__sc_tp_uint(evsel, id, sample), err = -1, callchain_ret = 0, printed = 0;
1993         int alignment = trace->args_alignment;
1994         struct syscall *sc = trace__syscall_info(trace, evsel, id);
1995         struct thread_trace *ttrace;
1996
1997         if (sc == NULL)
1998                 return -1;
1999
2000         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2001         ttrace = thread__trace(thread, trace->output);
2002         if (ttrace == NULL)
2003                 goto out_put;
2004
2005         trace__fprintf_sample(trace, evsel, sample, thread);
2006
2007         if (trace->summary)
2008                 thread__update_stats(ttrace, id, sample);
2009
2010         ret = perf_evsel__sc_tp_uint(evsel, ret, sample);
2011
2012         if (sc->is_open && ret >= 0 && ttrace->filename.pending_open) {
2013                 trace__set_fd_pathname(thread, ret, ttrace->filename.name);
2014                 ttrace->filename.pending_open = false;
2015                 ++trace->stats.vfs_getname;
2016         }
2017
2018         if (ttrace->entry_time) {
2019                 duration = sample->time - ttrace->entry_time;
2020                 if (trace__filter_duration(trace, duration))
2021                         goto out;
2022                 duration_calculated = true;
2023         } else if (trace->duration_filter)
2024                 goto out;
2025
2026         if (sample->callchain) {
2027                 callchain_ret = trace__resolve_callchain(trace, evsel, sample, &callchain_cursor);
2028                 if (callchain_ret == 0) {
2029                         if (callchain_cursor.nr < trace->min_stack)
2030                                 goto out;
2031                         callchain_ret = 1;
2032                 }
2033         }
2034
2035         if (trace->summary_only || (ret >= 0 && trace->failure_only))
2036                 goto out;
2037
2038         trace__fprintf_entry_head(trace, thread, duration, duration_calculated, ttrace->entry_time, trace->output);
2039
2040         if (ttrace->entry_pending) {
2041                 printed = fprintf(trace->output, "%s", ttrace->entry_str);
2042         } else {
2043                 printed += fprintf(trace->output, " ... [");
2044                 color_fprintf(trace->output, PERF_COLOR_YELLOW, "continued");
2045                 printed += 9;
2046                 printed += fprintf(trace->output, "]: %s()", sc->name);
2047         }
2048
2049         printed++; /* the closing ')' */
2050
2051         if (alignment > printed)
2052                 alignment -= printed;
2053         else
2054                 alignment = 0;
2055
2056         fprintf(trace->output, ")%*s= ", alignment, " ");
2057
2058         if (sc->fmt == NULL) {
2059                 if (ret < 0)
2060                         goto errno_print;
2061 signed_print:
2062                 fprintf(trace->output, "%ld", ret);
2063         } else if (ret < 0) {
2064 errno_print: {
2065                 char bf[STRERR_BUFSIZE];
2066                 const char *emsg = str_error_r(-ret, bf, sizeof(bf)),
2067                            *e = errno_to_name(evsel, -ret);
2068
2069                 fprintf(trace->output, "-1 %s (%s)", e, emsg);
2070         }
2071         } else if (ret == 0 && sc->fmt->timeout)
2072                 fprintf(trace->output, "0 (Timeout)");
2073         else if (ttrace->ret_scnprintf) {
2074                 char bf[1024];
2075                 struct syscall_arg arg = {
2076                         .val    = ret,
2077                         .thread = thread,
2078                         .trace  = trace,
2079                 };
2080                 ttrace->ret_scnprintf(bf, sizeof(bf), &arg);
2081                 ttrace->ret_scnprintf = NULL;
2082                 fprintf(trace->output, "%s", bf);
2083         } else if (sc->fmt->hexret)
2084                 fprintf(trace->output, "%#lx", ret);
2085         else if (sc->fmt->errpid) {
2086                 struct thread *child = machine__find_thread(trace->host, ret, ret);
2087
2088                 if (child != NULL) {
2089                         fprintf(trace->output, "%ld", ret);
2090                         if (child->comm_set)
2091                                 fprintf(trace->output, " (%s)", thread__comm_str(child));
2092                         thread__put(child);
2093                 }
2094         } else
2095                 goto signed_print;
2096
2097         fputc('\n', trace->output);
2098
2099         /*
2100          * We only consider an 'event' for the sake of --max-events a non-filtered
2101          * sys_enter + sys_exit and other tracepoint events.
2102          */
2103         if (++trace->nr_events_printed == trace->max_events && trace->max_events != ULONG_MAX)
2104                 interrupted = true;
2105
2106         if (callchain_ret > 0)
2107                 trace__fprintf_callchain(trace, sample);
2108         else if (callchain_ret < 0)
2109                 pr_err("Problem processing %s callchain, skipping...\n", perf_evsel__name(evsel));
2110 out:
2111         ttrace->entry_pending = false;
2112         err = 0;
2113 out_put:
2114         thread__put(thread);
2115         return err;
2116 }
2117
2118 static int trace__vfs_getname(struct trace *trace, struct perf_evsel *evsel,
2119                               union perf_event *event __maybe_unused,
2120                               struct perf_sample *sample)
2121 {
2122         struct thread *thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2123         struct thread_trace *ttrace;
2124         size_t filename_len, entry_str_len, to_move;
2125         ssize_t remaining_space;
2126         char *pos;
2127         const char *filename = perf_evsel__rawptr(evsel, sample, "pathname");
2128
2129         if (!thread)
2130                 goto out;
2131
2132         ttrace = thread__priv(thread);
2133         if (!ttrace)
2134                 goto out_put;
2135
2136         filename_len = strlen(filename);
2137         if (filename_len == 0)
2138                 goto out_put;
2139
2140         if (ttrace->filename.namelen < filename_len) {
2141                 char *f = realloc(ttrace->filename.name, filename_len + 1);
2142
2143                 if (f == NULL)
2144                         goto out_put;
2145
2146                 ttrace->filename.namelen = filename_len;
2147                 ttrace->filename.name = f;
2148         }
2149
2150         strcpy(ttrace->filename.name, filename);
2151         ttrace->filename.pending_open = true;
2152
2153         if (!ttrace->filename.ptr)
2154                 goto out_put;
2155
2156         entry_str_len = strlen(ttrace->entry_str);
2157         remaining_space = trace__entry_str_size - entry_str_len - 1; /* \0 */
2158         if (remaining_space <= 0)
2159                 goto out_put;
2160
2161         if (filename_len > (size_t)remaining_space) {
2162                 filename += filename_len - remaining_space;
2163                 filename_len = remaining_space;
2164         }
2165
2166         to_move = entry_str_len - ttrace->filename.entry_str_pos + 1; /* \0 */
2167         pos = ttrace->entry_str + ttrace->filename.entry_str_pos;
2168         memmove(pos + filename_len, pos, to_move);
2169         memcpy(pos, filename, filename_len);
2170
2171         ttrace->filename.ptr = 0;
2172         ttrace->filename.entry_str_pos = 0;
2173 out_put:
2174         thread__put(thread);
2175 out:
2176         return 0;
2177 }
2178
2179 static int trace__sched_stat_runtime(struct trace *trace, struct perf_evsel *evsel,
2180                                      union perf_event *event __maybe_unused,
2181                                      struct perf_sample *sample)
2182 {
2183         u64 runtime = perf_evsel__intval(evsel, sample, "runtime");
2184         double runtime_ms = (double)runtime / NSEC_PER_MSEC;
2185         struct thread *thread = machine__findnew_thread(trace->host,
2186                                                         sample->pid,
2187                                                         sample->tid);
2188         struct thread_trace *ttrace = thread__trace(thread, trace->output);
2189
2190         if (ttrace == NULL)
2191                 goto out_dump;
2192
2193         ttrace->runtime_ms += runtime_ms;
2194         trace->runtime_ms += runtime_ms;
2195 out_put:
2196         thread__put(thread);
2197         return 0;
2198
2199 out_dump:
2200         fprintf(trace->output, "%s: comm=%s,pid=%u,runtime=%" PRIu64 ",vruntime=%" PRIu64 ")\n",
2201                evsel->name,
2202                perf_evsel__strval(evsel, sample, "comm"),
2203                (pid_t)perf_evsel__intval(evsel, sample, "pid"),
2204                runtime,
2205                perf_evsel__intval(evsel, sample, "vruntime"));
2206         goto out_put;
2207 }
2208
2209 static int bpf_output__printer(enum binary_printer_ops op,
2210                                unsigned int val, void *extra __maybe_unused, FILE *fp)
2211 {
2212         unsigned char ch = (unsigned char)val;
2213
2214         switch (op) {
2215         case BINARY_PRINT_CHAR_DATA:
2216                 return fprintf(fp, "%c", isprint(ch) ? ch : '.');
2217         case BINARY_PRINT_DATA_BEGIN:
2218         case BINARY_PRINT_LINE_BEGIN:
2219         case BINARY_PRINT_ADDR:
2220         case BINARY_PRINT_NUM_DATA:
2221         case BINARY_PRINT_NUM_PAD:
2222         case BINARY_PRINT_SEP:
2223         case BINARY_PRINT_CHAR_PAD:
2224         case BINARY_PRINT_LINE_END:
2225         case BINARY_PRINT_DATA_END:
2226         default:
2227                 break;
2228         }
2229
2230         return 0;
2231 }
2232
2233 static void bpf_output__fprintf(struct trace *trace,
2234                                 struct perf_sample *sample)
2235 {
2236         binary__fprintf(sample->raw_data, sample->raw_size, 8,
2237                         bpf_output__printer, NULL, trace->output);
2238         ++trace->nr_events_printed;
2239 }
2240
2241 static int trace__event_handler(struct trace *trace, struct perf_evsel *evsel,
2242                                 union perf_event *event __maybe_unused,
2243                                 struct perf_sample *sample)
2244 {
2245         struct thread *thread;
2246         int callchain_ret = 0;
2247         /*
2248          * Check if we called perf_evsel__disable(evsel) due to, for instance,
2249          * this event's max_events having been hit and this is an entry coming
2250          * from the ring buffer that we should discard, since the max events
2251          * have already been considered/printed.
2252          */
2253         if (evsel->disabled)
2254                 return 0;
2255
2256         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2257
2258         if (sample->callchain) {
2259                 callchain_ret = trace__resolve_callchain(trace, evsel, sample, &callchain_cursor);
2260                 if (callchain_ret == 0) {
2261                         if (callchain_cursor.nr < trace->min_stack)
2262                                 goto out;
2263                         callchain_ret = 1;
2264                 }
2265         }
2266
2267         trace__printf_interrupted_entry(trace);
2268         trace__fprintf_tstamp(trace, sample->time, trace->output);
2269
2270         if (trace->trace_syscalls && trace->show_duration)
2271                 fprintf(trace->output, "(         ): ");
2272
2273         if (thread)
2274                 trace__fprintf_comm_tid(trace, thread, trace->output);
2275
2276         if (evsel == trace->syscalls.events.augmented) {
2277                 int id = perf_evsel__sc_tp_uint(evsel, id, sample);
2278                 struct syscall *sc = trace__syscall_info(trace, evsel, id);
2279
2280                 if (sc) {
2281                         fprintf(trace->output, "%s(", sc->name);
2282                         trace__fprintf_sys_enter(trace, evsel, sample);
2283                         fputc(')', trace->output);
2284                         goto newline;
2285                 }
2286
2287                 /*
2288                  * XXX: Not having the associated syscall info or not finding/adding
2289                  *      the thread should never happen, but if it does...
2290                  *      fall thru and print it as a bpf_output event.
2291                  */
2292         }
2293
2294         fprintf(trace->output, "%s:", evsel->name);
2295
2296         if (perf_evsel__is_bpf_output(evsel)) {
2297                 bpf_output__fprintf(trace, sample);
2298         } else if (evsel->tp_format) {
2299                 if (strncmp(evsel->tp_format->name, "sys_enter_", 10) ||
2300                     trace__fprintf_sys_enter(trace, evsel, sample)) {
2301                         event_format__fprintf(evsel->tp_format, sample->cpu,
2302                                               sample->raw_data, sample->raw_size,
2303                                               trace->output);
2304                         ++trace->nr_events_printed;
2305
2306                         if (evsel->max_events != ULONG_MAX && ++evsel->nr_events_printed == evsel->max_events) {
2307                                 perf_evsel__disable(evsel);
2308                                 perf_evsel__close(evsel);
2309                         }
2310                 }
2311         }
2312
2313 newline:
2314         fprintf(trace->output, "\n");
2315
2316         if (callchain_ret > 0)
2317                 trace__fprintf_callchain(trace, sample);
2318         else if (callchain_ret < 0)
2319                 pr_err("Problem processing %s callchain, skipping...\n", perf_evsel__name(evsel));
2320 out:
2321         thread__put(thread);
2322         return 0;
2323 }
2324
2325 static void print_location(FILE *f, struct perf_sample *sample,
2326                            struct addr_location *al,
2327                            bool print_dso, bool print_sym)
2328 {
2329
2330         if ((verbose > 0 || print_dso) && al->map)
2331                 fprintf(f, "%s@", al->map->dso->long_name);
2332
2333         if ((verbose > 0 || print_sym) && al->sym)
2334                 fprintf(f, "%s+0x%" PRIx64, al->sym->name,
2335                         al->addr - al->sym->start);
2336         else if (al->map)
2337                 fprintf(f, "0x%" PRIx64, al->addr);
2338         else
2339                 fprintf(f, "0x%" PRIx64, sample->addr);
2340 }
2341
2342 static int trace__pgfault(struct trace *trace,
2343                           struct perf_evsel *evsel,
2344                           union perf_event *event __maybe_unused,
2345                           struct perf_sample *sample)
2346 {
2347         struct thread *thread;
2348         struct addr_location al;
2349         char map_type = 'd';
2350         struct thread_trace *ttrace;
2351         int err = -1;
2352         int callchain_ret = 0;
2353
2354         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2355
2356         if (sample->callchain) {
2357                 callchain_ret = trace__resolve_callchain(trace, evsel, sample, &callchain_cursor);
2358                 if (callchain_ret == 0) {
2359                         if (callchain_cursor.nr < trace->min_stack)
2360                                 goto out_put;
2361                         callchain_ret = 1;
2362                 }
2363         }
2364
2365         ttrace = thread__trace(thread, trace->output);
2366         if (ttrace == NULL)
2367                 goto out_put;
2368
2369         if (evsel->attr.config == PERF_COUNT_SW_PAGE_FAULTS_MAJ)
2370                 ttrace->pfmaj++;
2371         else
2372                 ttrace->pfmin++;
2373
2374         if (trace->summary_only)
2375                 goto out;
2376
2377         thread__find_symbol(thread, sample->cpumode, sample->ip, &al);
2378
2379         trace__fprintf_entry_head(trace, thread, 0, true, sample->time, trace->output);
2380
2381         fprintf(trace->output, "%sfault [",
2382                 evsel->attr.config == PERF_COUNT_SW_PAGE_FAULTS_MAJ ?
2383                 "maj" : "min");
2384
2385         print_location(trace->output, sample, &al, false, true);
2386
2387         fprintf(trace->output, "] => ");
2388
2389         thread__find_symbol(thread, sample->cpumode, sample->addr, &al);
2390
2391         if (!al.map) {
2392                 thread__find_symbol(thread, sample->cpumode, sample->addr, &al);
2393
2394                 if (al.map)
2395                         map_type = 'x';
2396                 else
2397                         map_type = '?';
2398         }
2399
2400         print_location(trace->output, sample, &al, true, false);
2401
2402         fprintf(trace->output, " (%c%c)\n", map_type, al.level);
2403
2404         if (callchain_ret > 0)
2405                 trace__fprintf_callchain(trace, sample);
2406         else if (callchain_ret < 0)
2407                 pr_err("Problem processing %s callchain, skipping...\n", perf_evsel__name(evsel));
2408
2409         ++trace->nr_events_printed;
2410 out:
2411         err = 0;
2412 out_put:
2413         thread__put(thread);
2414         return err;
2415 }
2416
2417 static void trace__set_base_time(struct trace *trace,
2418                                  struct perf_evsel *evsel,
2419                                  struct perf_sample *sample)
2420 {
2421         /*
2422          * BPF events were not setting PERF_SAMPLE_TIME, so be more robust
2423          * and don't use sample->time unconditionally, we may end up having
2424          * some other event in the future without PERF_SAMPLE_TIME for good
2425          * reason, i.e. we may not be interested in its timestamps, just in
2426          * it taking place, picking some piece of information when it
2427          * appears in our event stream (vfs_getname comes to mind).
2428          */
2429         if (trace->base_time == 0 && !trace->full_time &&
2430             (evsel->attr.sample_type & PERF_SAMPLE_TIME))
2431                 trace->base_time = sample->time;
2432 }
2433
2434 static int trace__process_sample(struct perf_tool *tool,
2435                                  union perf_event *event,
2436                                  struct perf_sample *sample,
2437                                  struct perf_evsel *evsel,
2438                                  struct machine *machine __maybe_unused)
2439 {
2440         struct trace *trace = container_of(tool, struct trace, tool);
2441         struct thread *thread;
2442         int err = 0;
2443
2444         tracepoint_handler handler = evsel->handler;
2445
2446         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2447         if (thread && thread__is_filtered(thread))
2448                 goto out;
2449
2450         trace__set_base_time(trace, evsel, sample);
2451
2452         if (handler) {
2453                 ++trace->nr_events;
2454                 handler(trace, evsel, event, sample);
2455         }
2456 out:
2457         thread__put(thread);
2458         return err;
2459 }
2460
2461 static int trace__record(struct trace *trace, int argc, const char **argv)
2462 {
2463         unsigned int rec_argc, i, j;
2464         const char **rec_argv;
2465         const char * const record_args[] = {
2466                 "record",
2467                 "-R",
2468                 "-m", "1024",
2469                 "-c", "1",
2470         };
2471
2472         const char * const sc_args[] = { "-e", };
2473         unsigned int sc_args_nr = ARRAY_SIZE(sc_args);
2474         const char * const majpf_args[] = { "-e", "major-faults" };
2475         unsigned int majpf_args_nr = ARRAY_SIZE(majpf_args);
2476         const char * const minpf_args[] = { "-e", "minor-faults" };
2477         unsigned int minpf_args_nr = ARRAY_SIZE(minpf_args);
2478
2479         /* +1 is for the event string below */
2480         rec_argc = ARRAY_SIZE(record_args) + sc_args_nr + 1 +
2481                 majpf_args_nr + minpf_args_nr + argc;
2482         rec_argv = calloc(rec_argc + 1, sizeof(char *));
2483
2484         if (rec_argv == NULL)
2485                 return -ENOMEM;
2486
2487         j = 0;
2488         for (i = 0; i < ARRAY_SIZE(record_args); i++)
2489                 rec_argv[j++] = record_args[i];
2490
2491         if (trace->trace_syscalls) {
2492                 for (i = 0; i < sc_args_nr; i++)
2493                         rec_argv[j++] = sc_args[i];
2494
2495                 /* event string may be different for older kernels - e.g., RHEL6 */
2496                 if (is_valid_tracepoint("raw_syscalls:sys_enter"))
2497                         rec_argv[j++] = "raw_syscalls:sys_enter,raw_syscalls:sys_exit";
2498                 else if (is_valid_tracepoint("syscalls:sys_enter"))
2499                         rec_argv[j++] = "syscalls:sys_enter,syscalls:sys_exit";
2500                 else {
2501                         pr_err("Neither raw_syscalls nor syscalls events exist.\n");
2502                         free(rec_argv);
2503                         return -1;
2504                 }
2505         }
2506
2507         if (trace->trace_pgfaults & TRACE_PFMAJ)
2508                 for (i = 0; i < majpf_args_nr; i++)
2509                         rec_argv[j++] = majpf_args[i];
2510
2511         if (trace->trace_pgfaults & TRACE_PFMIN)
2512                 for (i = 0; i < minpf_args_nr; i++)
2513                         rec_argv[j++] = minpf_args[i];
2514
2515         for (i = 0; i < (unsigned int)argc; i++)
2516                 rec_argv[j++] = argv[i];
2517
2518         return cmd_record(j, rec_argv);
2519 }
2520
2521 static size_t trace__fprintf_thread_summary(struct trace *trace, FILE *fp);
2522
2523 static bool perf_evlist__add_vfs_getname(struct perf_evlist *evlist)
2524 {
2525         bool found = false;
2526         struct perf_evsel *evsel, *tmp;
2527         struct parse_events_error err = { .idx = 0, };
2528         int ret = parse_events(evlist, "probe:vfs_getname*", &err);
2529
2530         if (ret)
2531                 return false;
2532
2533         evlist__for_each_entry_safe(evlist, evsel, tmp) {
2534                 if (!strstarts(perf_evsel__name(evsel), "probe:vfs_getname"))
2535                         continue;
2536
2537                 if (perf_evsel__field(evsel, "pathname")) {
2538                         evsel->handler = trace__vfs_getname;
2539                         found = true;
2540                         continue;
2541                 }
2542
2543                 list_del_init(&evsel->node);
2544                 evsel->evlist = NULL;
2545                 perf_evsel__delete(evsel);
2546         }
2547
2548         return found;
2549 }
2550
2551 static struct perf_evsel *perf_evsel__new_pgfault(u64 config)
2552 {
2553         struct perf_evsel *evsel;
2554         struct perf_event_attr attr = {
2555                 .type = PERF_TYPE_SOFTWARE,
2556                 .mmap_data = 1,
2557         };
2558
2559         attr.config = config;
2560         attr.sample_period = 1;
2561
2562         event_attr_init(&attr);
2563
2564         evsel = perf_evsel__new(&attr);
2565         if (evsel)
2566                 evsel->handler = trace__pgfault;
2567
2568         return evsel;
2569 }
2570
2571 static void trace__handle_event(struct trace *trace, union perf_event *event, struct perf_sample *sample)
2572 {
2573         const u32 type = event->header.type;
2574         struct perf_evsel *evsel;
2575
2576         if (type != PERF_RECORD_SAMPLE) {
2577                 trace__process_event(trace, trace->host, event, sample);
2578                 return;
2579         }
2580
2581         evsel = perf_evlist__id2evsel(trace->evlist, sample->id);
2582         if (evsel == NULL) {
2583                 fprintf(trace->output, "Unknown tp ID %" PRIu64 ", skipping...\n", sample->id);
2584                 return;
2585         }
2586
2587         trace__set_base_time(trace, evsel, sample);
2588
2589         if (evsel->attr.type == PERF_TYPE_TRACEPOINT &&
2590             sample->raw_data == NULL) {
2591                 fprintf(trace->output, "%s sample with no payload for tid: %d, cpu %d, raw_size=%d, skipping...\n",
2592                        perf_evsel__name(evsel), sample->tid,
2593                        sample->cpu, sample->raw_size);
2594         } else {
2595                 tracepoint_handler handler = evsel->handler;
2596                 handler(trace, evsel, event, sample);
2597         }
2598
2599         if (trace->nr_events_printed >= trace->max_events && trace->max_events != ULONG_MAX)
2600                 interrupted = true;
2601 }
2602
2603 static int trace__add_syscall_newtp(struct trace *trace)
2604 {
2605         int ret = -1;
2606         struct perf_evlist *evlist = trace->evlist;
2607         struct perf_evsel *sys_enter, *sys_exit;
2608
2609         sys_enter = perf_evsel__raw_syscall_newtp("sys_enter", trace__sys_enter);
2610         if (sys_enter == NULL)
2611                 goto out;
2612
2613         if (perf_evsel__init_sc_tp_ptr_field(sys_enter, args))
2614                 goto out_delete_sys_enter;
2615
2616         sys_exit = perf_evsel__raw_syscall_newtp("sys_exit", trace__sys_exit);
2617         if (sys_exit == NULL)
2618                 goto out_delete_sys_enter;
2619
2620         if (perf_evsel__init_sc_tp_uint_field(sys_exit, ret))
2621                 goto out_delete_sys_exit;
2622
2623         perf_evsel__config_callchain(sys_enter, &trace->opts, &callchain_param);
2624         perf_evsel__config_callchain(sys_exit, &trace->opts, &callchain_param);
2625
2626         perf_evlist__add(evlist, sys_enter);
2627         perf_evlist__add(evlist, sys_exit);
2628
2629         if (callchain_param.enabled && !trace->kernel_syscallchains) {
2630                 /*
2631                  * We're interested only in the user space callchain
2632                  * leading to the syscall, allow overriding that for
2633                  * debugging reasons using --kernel_syscall_callchains
2634                  */
2635                 sys_exit->attr.exclude_callchain_kernel = 1;
2636         }
2637
2638         trace->syscalls.events.sys_enter = sys_enter;
2639         trace->syscalls.events.sys_exit  = sys_exit;
2640
2641         ret = 0;
2642 out:
2643         return ret;
2644
2645 out_delete_sys_exit:
2646         perf_evsel__delete_priv(sys_exit);
2647 out_delete_sys_enter:
2648         perf_evsel__delete_priv(sys_enter);
2649         goto out;
2650 }
2651
2652 static int trace__set_ev_qualifier_tp_filter(struct trace *trace)
2653 {
2654         int err = -1;
2655         struct perf_evsel *sys_exit;
2656         char *filter = asprintf_expr_inout_ints("id", !trace->not_ev_qualifier,
2657                                                 trace->ev_qualifier_ids.nr,
2658                                                 trace->ev_qualifier_ids.entries);
2659
2660         if (filter == NULL)
2661                 goto out_enomem;
2662
2663         if (!perf_evsel__append_tp_filter(trace->syscalls.events.sys_enter,
2664                                           filter)) {
2665                 sys_exit = trace->syscalls.events.sys_exit;
2666                 err = perf_evsel__append_tp_filter(sys_exit, filter);
2667         }
2668
2669         free(filter);
2670 out:
2671         return err;
2672 out_enomem:
2673         errno = ENOMEM;
2674         goto out;
2675 }
2676
2677 #ifdef HAVE_LIBBPF_SUPPORT
2678 static int trace__set_ev_qualifier_bpf_filter(struct trace *trace)
2679 {
2680         int fd = bpf_map__fd(trace->syscalls.map);
2681         struct bpf_map_syscall_entry value = {
2682                 .enabled = !trace->not_ev_qualifier,
2683         };
2684         int err = 0;
2685         size_t i;
2686
2687         for (i = 0; i < trace->ev_qualifier_ids.nr; ++i) {
2688                 int key = trace->ev_qualifier_ids.entries[i];
2689
2690                 err = bpf_map_update_elem(fd, &key, &value, BPF_EXIST);
2691                 if (err)
2692                         break;
2693         }
2694
2695         return err;
2696 }
2697
2698 static int __trace__init_syscalls_bpf_map(struct trace *trace, bool enabled)
2699 {
2700         int fd = bpf_map__fd(trace->syscalls.map);
2701         struct bpf_map_syscall_entry value = {
2702                 .enabled = enabled,
2703         };
2704         int err = 0, key;
2705
2706         for (key = 0; key < trace->sctbl->syscalls.nr_entries; ++key) {
2707                 err = bpf_map_update_elem(fd, &key, &value, BPF_ANY);
2708                 if (err)
2709                         break;
2710         }
2711
2712         return err;
2713 }
2714
2715 static int trace__init_syscalls_bpf_map(struct trace *trace)
2716 {
2717         bool enabled = true;
2718
2719         if (trace->ev_qualifier_ids.nr)
2720                 enabled = trace->not_ev_qualifier;
2721
2722         return __trace__init_syscalls_bpf_map(trace, enabled);
2723 }
2724 #else
2725 static int trace__set_ev_qualifier_bpf_filter(struct trace *trace __maybe_unused)
2726 {
2727         return 0;
2728 }
2729
2730 static int trace__init_syscalls_bpf_map(struct trace *trace __maybe_unused)
2731 {
2732         return 0;
2733 }
2734 #endif // HAVE_LIBBPF_SUPPORT
2735
2736 static int trace__set_ev_qualifier_filter(struct trace *trace)
2737 {
2738         if (trace->syscalls.map)
2739                 return trace__set_ev_qualifier_bpf_filter(trace);
2740         if (trace->syscalls.events.sys_enter)
2741                 return trace__set_ev_qualifier_tp_filter(trace);
2742         return 0;
2743 }
2744
2745 static int bpf_map__set_filter_pids(struct bpf_map *map __maybe_unused,
2746                                     size_t npids __maybe_unused, pid_t *pids __maybe_unused)
2747 {
2748         int err = 0;
2749 #ifdef HAVE_LIBBPF_SUPPORT
2750         bool value = true;
2751         int map_fd = bpf_map__fd(map);
2752         size_t i;
2753
2754         for (i = 0; i < npids; ++i) {
2755                 err = bpf_map_update_elem(map_fd, &pids[i], &value, BPF_ANY);
2756                 if (err)
2757                         break;
2758         }
2759 #endif
2760         return err;
2761 }
2762
2763 static int trace__set_filter_loop_pids(struct trace *trace)
2764 {
2765         unsigned int nr = 1, err;
2766         pid_t pids[32] = {
2767                 getpid(),
2768         };
2769         struct thread *thread = machine__find_thread(trace->host, pids[0], pids[0]);
2770
2771         while (thread && nr < ARRAY_SIZE(pids)) {
2772                 struct thread *parent = machine__find_thread(trace->host, thread->ppid, thread->ppid);
2773
2774                 if (parent == NULL)
2775                         break;
2776
2777                 if (!strcmp(thread__comm_str(parent), "sshd") ||
2778                     strstarts(thread__comm_str(parent), "gnome-terminal")) {
2779                         pids[nr++] = parent->tid;
2780                         break;
2781                 }
2782                 thread = parent;
2783         }
2784
2785         err = perf_evlist__set_tp_filter_pids(trace->evlist, nr, pids);
2786         if (!err && trace->filter_pids.map)
2787                 err = bpf_map__set_filter_pids(trace->filter_pids.map, nr, pids);
2788
2789         return err;
2790 }
2791
2792 static int trace__set_filter_pids(struct trace *trace)
2793 {
2794         int err = 0;
2795         /*
2796          * Better not use !target__has_task() here because we need to cover the
2797          * case where no threads were specified in the command line, but a
2798          * workload was, and in that case we will fill in the thread_map when
2799          * we fork the workload in perf_evlist__prepare_workload.
2800          */
2801         if (trace->filter_pids.nr > 0) {
2802                 err = perf_evlist__set_tp_filter_pids(trace->evlist, trace->filter_pids.nr,
2803                                                       trace->filter_pids.entries);
2804                 if (!err && trace->filter_pids.map) {
2805                         err = bpf_map__set_filter_pids(trace->filter_pids.map, trace->filter_pids.nr,
2806                                                        trace->filter_pids.entries);
2807                 }
2808         } else if (thread_map__pid(trace->evlist->threads, 0) == -1) {
2809                 err = trace__set_filter_loop_pids(trace);
2810         }
2811
2812         return err;
2813 }
2814
2815 static int __trace__deliver_event(struct trace *trace, union perf_event *event)
2816 {
2817         struct perf_evlist *evlist = trace->evlist;
2818         struct perf_sample sample;
2819         int err;
2820
2821         err = perf_evlist__parse_sample(evlist, event, &sample);
2822         if (err)
2823                 fprintf(trace->output, "Can't parse sample, err = %d, skipping...\n", err);
2824         else
2825                 trace__handle_event(trace, event, &sample);
2826
2827         return 0;
2828 }
2829
2830 static int __trace__flush_events(struct trace *trace)
2831 {
2832         u64 first = ordered_events__first_time(&trace->oe.data);
2833         u64 flush = trace->oe.last - NSEC_PER_SEC;
2834
2835         /* Is there some thing to flush.. */
2836         if (first && first < flush)
2837                 return ordered_events__flush_time(&trace->oe.data, flush);
2838
2839         return 0;
2840 }
2841
2842 static int trace__flush_events(struct trace *trace)
2843 {
2844         return !trace->sort_events ? 0 : __trace__flush_events(trace);
2845 }
2846
2847 static int trace__deliver_event(struct trace *trace, union perf_event *event)
2848 {
2849         int err;
2850
2851         if (!trace->sort_events)
2852                 return __trace__deliver_event(trace, event);
2853
2854         err = perf_evlist__parse_sample_timestamp(trace->evlist, event, &trace->oe.last);
2855         if (err && err != -1)
2856                 return err;
2857
2858         err = ordered_events__queue(&trace->oe.data, event, trace->oe.last, 0);
2859         if (err)
2860                 return err;
2861
2862         return trace__flush_events(trace);
2863 }
2864
2865 static int ordered_events__deliver_event(struct ordered_events *oe,
2866                                          struct ordered_event *event)
2867 {
2868         struct trace *trace = container_of(oe, struct trace, oe.data);
2869
2870         return __trace__deliver_event(trace, event->event);
2871 }
2872
2873 static int trace__run(struct trace *trace, int argc, const char **argv)
2874 {
2875         struct perf_evlist *evlist = trace->evlist;
2876         struct perf_evsel *evsel, *pgfault_maj = NULL, *pgfault_min = NULL;
2877         int err = -1, i;
2878         unsigned long before;
2879         const bool forks = argc > 0;
2880         bool draining = false;
2881
2882         trace->live = true;
2883
2884         if (!trace->raw_augmented_syscalls) {
2885                 if (trace->trace_syscalls && trace__add_syscall_newtp(trace))
2886                         goto out_error_raw_syscalls;
2887
2888                 if (trace->trace_syscalls)
2889                         trace->vfs_getname = perf_evlist__add_vfs_getname(evlist);
2890         }
2891
2892         if ((trace->trace_pgfaults & TRACE_PFMAJ)) {
2893                 pgfault_maj = perf_evsel__new_pgfault(PERF_COUNT_SW_PAGE_FAULTS_MAJ);
2894                 if (pgfault_maj == NULL)
2895                         goto out_error_mem;
2896                 perf_evsel__config_callchain(pgfault_maj, &trace->opts, &callchain_param);
2897                 perf_evlist__add(evlist, pgfault_maj);
2898         }
2899
2900         if ((trace->trace_pgfaults & TRACE_PFMIN)) {
2901                 pgfault_min = perf_evsel__new_pgfault(PERF_COUNT_SW_PAGE_FAULTS_MIN);
2902                 if (pgfault_min == NULL)
2903                         goto out_error_mem;
2904                 perf_evsel__config_callchain(pgfault_min, &trace->opts, &callchain_param);
2905                 perf_evlist__add(evlist, pgfault_min);
2906         }
2907
2908         if (trace->sched &&
2909             perf_evlist__add_newtp(evlist, "sched", "sched_stat_runtime",
2910                                    trace__sched_stat_runtime))
2911                 goto out_error_sched_stat_runtime;
2912
2913         /*
2914          * If a global cgroup was set, apply it to all the events without an
2915          * explicit cgroup. I.e.:
2916          *
2917          *      trace -G A -e sched:*switch
2918          *
2919          * Will set all raw_syscalls:sys_{enter,exit}, pgfault, vfs_getname, etc
2920          * _and_ sched:sched_switch to the 'A' cgroup, while:
2921          *
2922          * trace -e sched:*switch -G A
2923          *
2924          * will only set the sched:sched_switch event to the 'A' cgroup, all the
2925          * other events (raw_syscalls:sys_{enter,exit}, etc are left "without"
2926          * a cgroup (on the root cgroup, sys wide, etc).
2927          *
2928          * Multiple cgroups:
2929          *
2930          * trace -G A -e sched:*switch -G B
2931          *
2932          * the syscall ones go to the 'A' cgroup, the sched:sched_switch goes
2933          * to the 'B' cgroup.
2934          *
2935          * evlist__set_default_cgroup() grabs a reference of the passed cgroup
2936          * only for the evsels still without a cgroup, i.e. evsel->cgroup == NULL.
2937          */
2938         if (trace->cgroup)
2939                 evlist__set_default_cgroup(trace->evlist, trace->cgroup);
2940
2941         err = perf_evlist__create_maps(evlist, &trace->opts.target);
2942         if (err < 0) {
2943                 fprintf(trace->output, "Problems parsing the target to trace, check your options!\n");
2944                 goto out_delete_evlist;
2945         }
2946
2947         err = trace__symbols_init(trace, evlist);
2948         if (err < 0) {
2949                 fprintf(trace->output, "Problems initializing symbol libraries!\n");
2950                 goto out_delete_evlist;
2951         }
2952
2953         perf_evlist__config(evlist, &trace->opts, &callchain_param);
2954
2955         signal(SIGCHLD, sig_handler);
2956         signal(SIGINT, sig_handler);
2957
2958         if (forks) {
2959                 err = perf_evlist__prepare_workload(evlist, &trace->opts.target,
2960                                                     argv, false, NULL);
2961                 if (err < 0) {
2962                         fprintf(trace->output, "Couldn't run the workload!\n");
2963                         goto out_delete_evlist;
2964                 }
2965         }
2966
2967         err = perf_evlist__open(evlist);
2968         if (err < 0)
2969                 goto out_error_open;
2970
2971         err = bpf__apply_obj_config();
2972         if (err) {
2973                 char errbuf[BUFSIZ];
2974
2975                 bpf__strerror_apply_obj_config(err, errbuf, sizeof(errbuf));
2976                 pr_err("ERROR: Apply config to BPF failed: %s\n",
2977                          errbuf);
2978                 goto out_error_open;
2979         }
2980
2981         err = trace__set_filter_pids(trace);
2982         if (err < 0)
2983                 goto out_error_mem;
2984
2985         if (trace->syscalls.map)
2986                 trace__init_syscalls_bpf_map(trace);
2987
2988         if (trace->ev_qualifier_ids.nr > 0) {
2989                 err = trace__set_ev_qualifier_filter(trace);
2990                 if (err < 0)
2991                         goto out_errno;
2992
2993                 if (trace->syscalls.events.sys_exit) {
2994                         pr_debug("event qualifier tracepoint filter: %s\n",
2995                                  trace->syscalls.events.sys_exit->filter);
2996                 }
2997         }
2998
2999         err = perf_evlist__apply_filters(evlist, &evsel);
3000         if (err < 0)
3001                 goto out_error_apply_filters;
3002
3003         if (trace->dump.map)
3004                 bpf_map__fprintf(trace->dump.map, trace->output);
3005
3006         err = perf_evlist__mmap(evlist, trace->opts.mmap_pages);
3007         if (err < 0)
3008                 goto out_error_mmap;
3009
3010         if (!target__none(&trace->opts.target) && !trace->opts.initial_delay)
3011                 perf_evlist__enable(evlist);
3012
3013         if (forks)
3014                 perf_evlist__start_workload(evlist);
3015
3016         if (trace->opts.initial_delay) {
3017                 usleep(trace->opts.initial_delay * 1000);
3018                 perf_evlist__enable(evlist);
3019         }
3020
3021         trace->multiple_threads = thread_map__pid(evlist->threads, 0) == -1 ||
3022                                   evlist->threads->nr > 1 ||
3023                                   perf_evlist__first(evlist)->attr.inherit;
3024
3025         /*
3026          * Now that we already used evsel->attr to ask the kernel to setup the
3027          * events, lets reuse evsel->attr.sample_max_stack as the limit in
3028          * trace__resolve_callchain(), allowing per-event max-stack settings
3029          * to override an explicitly set --max-stack global setting.
3030          */
3031         evlist__for_each_entry(evlist, evsel) {
3032                 if (evsel__has_callchain(evsel) &&
3033                     evsel->attr.sample_max_stack == 0)
3034                         evsel->attr.sample_max_stack = trace->max_stack;
3035         }
3036 again:
3037         before = trace->nr_events;
3038
3039         for (i = 0; i < evlist->nr_mmaps; i++) {
3040                 union perf_event *event;
3041                 struct perf_mmap *md;
3042
3043                 md = &evlist->mmap[i];
3044                 if (perf_mmap__read_init(md) < 0)
3045                         continue;
3046
3047                 while ((event = perf_mmap__read_event(md)) != NULL) {
3048                         ++trace->nr_events;
3049
3050                         err = trace__deliver_event(trace, event);
3051                         if (err)
3052                                 goto out_disable;
3053
3054                         perf_mmap__consume(md);
3055
3056                         if (interrupted)
3057                                 goto out_disable;
3058
3059                         if (done && !draining) {
3060                                 perf_evlist__disable(evlist);
3061                                 draining = true;
3062                         }
3063                 }
3064                 perf_mmap__read_done(md);
3065         }
3066
3067         if (trace->nr_events == before) {
3068                 int timeout = done ? 100 : -1;
3069
3070                 if (!draining && perf_evlist__poll(evlist, timeout) > 0) {
3071                         if (perf_evlist__filter_pollfd(evlist, POLLERR | POLLHUP | POLLNVAL) == 0)
3072                                 draining = true;
3073
3074                         goto again;
3075                 } else {
3076                         if (trace__flush_events(trace))
3077                                 goto out_disable;
3078                 }
3079         } else {
3080                 goto again;
3081         }
3082
3083 out_disable:
3084         thread__zput(trace->current);
3085
3086         perf_evlist__disable(evlist);
3087
3088         if (trace->sort_events)
3089                 ordered_events__flush(&trace->oe.data, OE_FLUSH__FINAL);
3090
3091         if (!err) {
3092                 if (trace->summary)
3093                         trace__fprintf_thread_summary(trace, trace->output);
3094
3095                 if (trace->show_tool_stats) {
3096                         fprintf(trace->output, "Stats:\n "
3097                                                " vfs_getname : %" PRIu64 "\n"
3098                                                " proc_getname: %" PRIu64 "\n",
3099                                 trace->stats.vfs_getname,
3100                                 trace->stats.proc_getname);
3101                 }
3102         }
3103
3104 out_delete_evlist:
3105         trace__symbols__exit(trace);
3106
3107         perf_evlist__delete(evlist);
3108         cgroup__put(trace->cgroup);
3109         trace->evlist = NULL;
3110         trace->live = false;
3111         return err;
3112 {
3113         char errbuf[BUFSIZ];
3114
3115 out_error_sched_stat_runtime:
3116         tracing_path__strerror_open_tp(errno, errbuf, sizeof(errbuf), "sched", "sched_stat_runtime");
3117         goto out_error;
3118
3119 out_error_raw_syscalls:
3120         tracing_path__strerror_open_tp(errno, errbuf, sizeof(errbuf), "raw_syscalls", "sys_(enter|exit)");
3121         goto out_error;
3122
3123 out_error_mmap:
3124         perf_evlist__strerror_mmap(evlist, errno, errbuf, sizeof(errbuf));
3125         goto out_error;
3126
3127 out_error_open:
3128         perf_evlist__strerror_open(evlist, errno, errbuf, sizeof(errbuf));
3129
3130 out_error:
3131         fprintf(trace->output, "%s\n", errbuf);
3132         goto out_delete_evlist;
3133
3134 out_error_apply_filters:
3135         fprintf(trace->output,
3136                 "Failed to set filter \"%s\" on event %s with %d (%s)\n",
3137                 evsel->filter, perf_evsel__name(evsel), errno,
3138                 str_error_r(errno, errbuf, sizeof(errbuf)));
3139         goto out_delete_evlist;
3140 }
3141 out_error_mem:
3142         fprintf(trace->output, "Not enough memory to run!\n");
3143         goto out_delete_evlist;
3144
3145 out_errno:
3146         fprintf(trace->output, "errno=%d,%s\n", errno, strerror(errno));
3147         goto out_delete_evlist;
3148 }
3149
3150 static int trace__replay(struct trace *trace)
3151 {
3152         const struct perf_evsel_str_handler handlers[] = {
3153                 { "probe:vfs_getname",       trace__vfs_getname, },
3154         };
3155         struct perf_data data = {
3156                 .path  = input_name,
3157                 .mode  = PERF_DATA_MODE_READ,
3158                 .force = trace->force,
3159         };
3160         struct perf_session *session;
3161         struct perf_evsel *evsel;
3162         int err = -1;
3163
3164         trace->tool.sample        = trace__process_sample;
3165         trace->tool.mmap          = perf_event__process_mmap;
3166         trace->tool.mmap2         = perf_event__process_mmap2;
3167         trace->tool.comm          = perf_event__process_comm;
3168         trace->tool.exit          = perf_event__process_exit;
3169         trace->tool.fork          = perf_event__process_fork;
3170         trace->tool.attr          = perf_event__process_attr;
3171         trace->tool.tracing_data  = perf_event__process_tracing_data;
3172         trace->tool.build_id      = perf_event__process_build_id;
3173         trace->tool.namespaces    = perf_event__process_namespaces;
3174
3175         trace->tool.ordered_events = true;
3176         trace->tool.ordering_requires_timestamps = true;
3177
3178         /* add tid to output */
3179         trace->multiple_threads = true;
3180
3181         session = perf_session__new(&data, false, &trace->tool);
3182         if (session == NULL)
3183                 return -1;
3184
3185         if (trace->opts.target.pid)
3186                 symbol_conf.pid_list_str = strdup(trace->opts.target.pid);
3187
3188         if (trace->opts.target.tid)
3189                 symbol_conf.tid_list_str = strdup(trace->opts.target.tid);
3190
3191         if (symbol__init(&session->header.env) < 0)
3192                 goto out;
3193
3194         trace->host = &session->machines.host;
3195
3196         err = perf_session__set_tracepoints_handlers(session, handlers);
3197         if (err)
3198                 goto out;
3199
3200         evsel = perf_evlist__find_tracepoint_by_name(session->evlist,
3201                                                      "raw_syscalls:sys_enter");
3202         /* older kernels have syscalls tp versus raw_syscalls */
3203         if (evsel == NULL)
3204                 evsel = perf_evlist__find_tracepoint_by_name(session->evlist,
3205                                                              "syscalls:sys_enter");
3206
3207         if (evsel &&
3208             (perf_evsel__init_raw_syscall_tp(evsel, trace__sys_enter) < 0 ||
3209             perf_evsel__init_sc_tp_ptr_field(evsel, args))) {
3210                 pr_err("Error during initialize raw_syscalls:sys_enter event\n");
3211                 goto out;
3212         }
3213
3214         evsel = perf_evlist__find_tracepoint_by_name(session->evlist,
3215                                                      "raw_syscalls:sys_exit");
3216         if (evsel == NULL)
3217                 evsel = perf_evlist__find_tracepoint_by_name(session->evlist,
3218                                                              "syscalls:sys_exit");
3219         if (evsel &&
3220             (perf_evsel__init_raw_syscall_tp(evsel, trace__sys_exit) < 0 ||
3221             perf_evsel__init_sc_tp_uint_field(evsel, ret))) {
3222                 pr_err("Error during initialize raw_syscalls:sys_exit event\n");
3223                 goto out;
3224         }
3225
3226         evlist__for_each_entry(session->evlist, evsel) {
3227                 if (evsel->attr.type == PERF_TYPE_SOFTWARE &&
3228                     (evsel->attr.config == PERF_COUNT_SW_PAGE_FAULTS_MAJ ||
3229                      evsel->attr.config == PERF_COUNT_SW_PAGE_FAULTS_MIN ||
3230                      evsel->attr.config == PERF_COUNT_SW_PAGE_FAULTS))
3231                         evsel->handler = trace__pgfault;
3232         }
3233
3234         setup_pager();
3235
3236         err = perf_session__process_events(session);
3237         if (err)
3238                 pr_err("Failed to process events, error %d", err);
3239
3240         else if (trace->summary)
3241                 trace__fprintf_thread_summary(trace, trace->output);
3242
3243 out:
3244         perf_session__delete(session);
3245
3246         return err;
3247 }
3248
3249 static size_t trace__fprintf_threads_header(FILE *fp)
3250 {
3251         size_t printed;
3252
3253         printed  = fprintf(fp, "\n Summary of events:\n\n");
3254
3255         return printed;
3256 }
3257
3258 DEFINE_RESORT_RB(syscall_stats, a->msecs > b->msecs,
3259         struct stats    *stats;
3260         double          msecs;
3261         int             syscall;
3262 )
3263 {
3264         struct int_node *source = rb_entry(nd, struct int_node, rb_node);
3265         struct stats *stats = source->priv;
3266
3267         entry->syscall = source->i;
3268         entry->stats   = stats;
3269         entry->msecs   = stats ? (u64)stats->n * (avg_stats(stats) / NSEC_PER_MSEC) : 0;
3270 }
3271
3272 static size_t thread__dump_stats(struct thread_trace *ttrace,
3273                                  struct trace *trace, FILE *fp)
3274 {
3275         size_t printed = 0;
3276         struct syscall *sc;
3277         struct rb_node *nd;
3278         DECLARE_RESORT_RB_INTLIST(syscall_stats, ttrace->syscall_stats);
3279
3280         if (syscall_stats == NULL)
3281                 return 0;
3282
3283         printed += fprintf(fp, "\n");
3284
3285         printed += fprintf(fp, "   syscall            calls    total       min       avg       max      stddev\n");
3286         printed += fprintf(fp, "                               (msec)    (msec)    (msec)    (msec)        (%%)\n");
3287         printed += fprintf(fp, "   --------------- -------- --------- --------- --------- ---------     ------\n");
3288
3289         resort_rb__for_each_entry(nd, syscall_stats) {
3290                 struct stats *stats = syscall_stats_entry->stats;
3291                 if (stats) {
3292                         double min = (double)(stats->min) / NSEC_PER_MSEC;
3293                         double max = (double)(stats->max) / NSEC_PER_MSEC;
3294                         double avg = avg_stats(stats);
3295                         double pct;
3296                         u64 n = (u64) stats->n;
3297
3298                         pct = avg ? 100.0 * stddev_stats(stats)/avg : 0.0;
3299                         avg /= NSEC_PER_MSEC;
3300
3301                         sc = &trace->syscalls.table[syscall_stats_entry->syscall];
3302                         printed += fprintf(fp, "   %-15s", sc->name);
3303                         printed += fprintf(fp, " %8" PRIu64 " %9.3f %9.3f %9.3f",
3304                                            n, syscall_stats_entry->msecs, min, avg);
3305                         printed += fprintf(fp, " %9.3f %9.2f%%\n", max, pct);
3306                 }
3307         }
3308
3309         resort_rb__delete(syscall_stats);
3310         printed += fprintf(fp, "\n\n");
3311
3312         return printed;
3313 }
3314
3315 static size_t trace__fprintf_thread(FILE *fp, struct thread *thread, struct trace *trace)
3316 {
3317         size_t printed = 0;
3318         struct thread_trace *ttrace = thread__priv(thread);
3319         double ratio;
3320
3321         if (ttrace == NULL)
3322                 return 0;
3323
3324         ratio = (double)ttrace->nr_events / trace->nr_events * 100.0;
3325
3326         printed += fprintf(fp, " %s (%d), ", thread__comm_str(thread), thread->tid);
3327         printed += fprintf(fp, "%lu events, ", ttrace->nr_events);
3328         printed += fprintf(fp, "%.1f%%", ratio);
3329         if (ttrace->pfmaj)
3330                 printed += fprintf(fp, ", %lu majfaults", ttrace->pfmaj);
3331         if (ttrace->pfmin)
3332                 printed += fprintf(fp, ", %lu minfaults", ttrace->pfmin);
3333         if (trace->sched)
3334                 printed += fprintf(fp, ", %.3f msec\n", ttrace->runtime_ms);
3335         else if (fputc('\n', fp) != EOF)
3336                 ++printed;
3337
3338         printed += thread__dump_stats(ttrace, trace, fp);
3339
3340         return printed;
3341 }
3342
3343 static unsigned long thread__nr_events(struct thread_trace *ttrace)
3344 {
3345         return ttrace ? ttrace->nr_events : 0;
3346 }
3347
3348 DEFINE_RESORT_RB(threads, (thread__nr_events(a->thread->priv) < thread__nr_events(b->thread->priv)),
3349         struct thread *thread;
3350 )
3351 {
3352         entry->thread = rb_entry(nd, struct thread, rb_node);
3353 }
3354
3355 static size_t trace__fprintf_thread_summary(struct trace *trace, FILE *fp)
3356 {
3357         size_t printed = trace__fprintf_threads_header(fp);
3358         struct rb_node *nd;
3359         int i;
3360
3361         for (i = 0; i < THREADS__TABLE_SIZE; i++) {
3362                 DECLARE_RESORT_RB_MACHINE_THREADS(threads, trace->host, i);
3363
3364                 if (threads == NULL) {
3365                         fprintf(fp, "%s", "Error sorting output by nr_events!\n");
3366                         return 0;
3367                 }
3368
3369                 resort_rb__for_each_entry(nd, threads)
3370                         printed += trace__fprintf_thread(fp, threads_entry->thread, trace);
3371
3372                 resort_rb__delete(threads);
3373         }
3374         return printed;
3375 }
3376
3377 static int trace__set_duration(const struct option *opt, const char *str,
3378                                int unset __maybe_unused)
3379 {
3380         struct trace *trace = opt->value;
3381
3382         trace->duration_filter = atof(str);
3383         return 0;
3384 }
3385
3386 static int trace__set_filter_pids_from_option(const struct option *opt, const char *str,
3387                                               int unset __maybe_unused)
3388 {
3389         int ret = -1;
3390         size_t i;
3391         struct trace *trace = opt->value;
3392         /*
3393          * FIXME: introduce a intarray class, plain parse csv and create a
3394          * { int nr, int entries[] } struct...
3395          */
3396         struct intlist *list = intlist__new(str);
3397
3398         if (list == NULL)
3399                 return -1;
3400
3401         i = trace->filter_pids.nr = intlist__nr_entries(list) + 1;
3402         trace->filter_pids.entries = calloc(i, sizeof(pid_t));
3403
3404         if (trace->filter_pids.entries == NULL)
3405                 goto out;
3406
3407         trace->filter_pids.entries[0] = getpid();
3408
3409         for (i = 1; i < trace->filter_pids.nr; ++i)
3410                 trace->filter_pids.entries[i] = intlist__entry(list, i - 1)->i;
3411
3412         intlist__delete(list);
3413         ret = 0;
3414 out:
3415         return ret;
3416 }
3417
3418 static int trace__open_output(struct trace *trace, const char *filename)
3419 {
3420         struct stat st;
3421
3422         if (!stat(filename, &st) && st.st_size) {
3423                 char oldname[PATH_MAX];
3424
3425                 scnprintf(oldname, sizeof(oldname), "%s.old", filename);
3426                 unlink(oldname);
3427                 rename(filename, oldname);
3428         }
3429
3430         trace->output = fopen(filename, "w");
3431
3432         return trace->output == NULL ? -errno : 0;
3433 }
3434
3435 static int parse_pagefaults(const struct option *opt, const char *str,
3436                             int unset __maybe_unused)
3437 {
3438         int *trace_pgfaults = opt->value;
3439
3440         if (strcmp(str, "all") == 0)
3441                 *trace_pgfaults |= TRACE_PFMAJ | TRACE_PFMIN;
3442         else if (strcmp(str, "maj") == 0)
3443                 *trace_pgfaults |= TRACE_PFMAJ;
3444         else if (strcmp(str, "min") == 0)
3445                 *trace_pgfaults |= TRACE_PFMIN;
3446         else
3447                 return -1;
3448
3449         return 0;
3450 }
3451
3452 static void evlist__set_evsel_handler(struct perf_evlist *evlist, void *handler)
3453 {
3454         struct perf_evsel *evsel;
3455
3456         evlist__for_each_entry(evlist, evsel)
3457                 evsel->handler = handler;
3458 }
3459
3460 static int evlist__set_syscall_tp_fields(struct perf_evlist *evlist)
3461 {
3462         struct perf_evsel *evsel;
3463
3464         evlist__for_each_entry(evlist, evsel) {
3465                 if (evsel->priv || !evsel->tp_format)
3466                         continue;
3467
3468                 if (strcmp(evsel->tp_format->system, "syscalls"))
3469                         continue;
3470
3471                 if (perf_evsel__init_syscall_tp(evsel))
3472                         return -1;
3473
3474                 if (!strncmp(evsel->tp_format->name, "sys_enter_", 10)) {
3475                         struct syscall_tp *sc = evsel->priv;
3476
3477                         if (__tp_field__init_ptr(&sc->args, sc->id.offset + sizeof(u64)))
3478                                 return -1;
3479                 } else if (!strncmp(evsel->tp_format->name, "sys_exit_", 9)) {
3480                         struct syscall_tp *sc = evsel->priv;
3481
3482                         if (__tp_field__init_uint(&sc->ret, sizeof(u64), sc->id.offset + sizeof(u64), evsel->needs_swap))
3483                                 return -1;
3484                 }
3485         }
3486
3487         return 0;
3488 }
3489
3490 /*
3491  * XXX: Hackish, just splitting the combined -e+--event (syscalls
3492  * (raw_syscalls:{sys_{enter,exit}} + events (tracepoints, HW, SW, etc) to use
3493  * existing facilities unchanged (trace->ev_qualifier + parse_options()).
3494  *
3495  * It'd be better to introduce a parse_options() variant that would return a
3496  * list with the terms it didn't match to an event...
3497  */
3498 static int trace__parse_events_option(const struct option *opt, const char *str,
3499                                       int unset __maybe_unused)
3500 {
3501         struct trace *trace = (struct trace *)opt->value;
3502         const char *s = str;
3503         char *sep = NULL, *lists[2] = { NULL, NULL, };
3504         int len = strlen(str) + 1, err = -1, list, idx;
3505         char *strace_groups_dir = system_path(STRACE_GROUPS_DIR);
3506         char group_name[PATH_MAX];
3507         struct syscall_fmt *fmt;
3508
3509         if (strace_groups_dir == NULL)
3510                 return -1;
3511
3512         if (*s == '!') {
3513                 ++s;
3514                 trace->not_ev_qualifier = true;
3515         }
3516
3517         while (1) {
3518                 if ((sep = strchr(s, ',')) != NULL)
3519                         *sep = '\0';
3520
3521                 list = 0;
3522                 if (syscalltbl__id(trace->sctbl, s) >= 0 ||
3523                     syscalltbl__strglobmatch_first(trace->sctbl, s, &idx) >= 0) {
3524                         list = 1;
3525                         goto do_concat;
3526                 }
3527
3528                 fmt = syscall_fmt__find_by_alias(s);
3529                 if (fmt != NULL) {
3530                         list = 1;
3531                         s = fmt->name;
3532                 } else {
3533                         path__join(group_name, sizeof(group_name), strace_groups_dir, s);
3534                         if (access(group_name, R_OK) == 0)
3535                                 list = 1;
3536                 }
3537 do_concat:
3538                 if (lists[list]) {
3539                         sprintf(lists[list] + strlen(lists[list]), ",%s", s);
3540                 } else {
3541                         lists[list] = malloc(len);
3542                         if (lists[list] == NULL)
3543                                 goto out;
3544                         strcpy(lists[list], s);
3545                 }
3546
3547                 if (!sep)
3548                         break;
3549
3550                 *sep = ',';
3551                 s = sep + 1;
3552         }
3553
3554         if (lists[1] != NULL) {
3555                 struct strlist_config slist_config = {
3556                         .dirname = strace_groups_dir,
3557                 };
3558
3559                 trace->ev_qualifier = strlist__new(lists[1], &slist_config);
3560                 if (trace->ev_qualifier == NULL) {
3561                         fputs("Not enough memory to parse event qualifier", trace->output);
3562                         goto out;
3563                 }
3564
3565                 if (trace__validate_ev_qualifier(trace))
3566                         goto out;
3567                 trace->trace_syscalls = true;
3568         }
3569
3570         err = 0;
3571
3572         if (lists[0]) {
3573                 struct option o = OPT_CALLBACK('e', "event", &trace->evlist, "event",
3574                                                "event selector. use 'perf list' to list available events",
3575                                                parse_events_option);
3576                 err = parse_events_option(&o, lists[0], 0);
3577         }
3578 out:
3579         if (sep)
3580                 *sep = ',';
3581
3582         return err;
3583 }
3584
3585 static int trace__parse_cgroups(const struct option *opt, const char *str, int unset)
3586 {
3587         struct trace *trace = opt->value;
3588
3589         if (!list_empty(&trace->evlist->entries))
3590                 return parse_cgroups(opt, str, unset);
3591
3592         trace->cgroup = evlist__findnew_cgroup(trace->evlist, str);
3593
3594         return 0;
3595 }
3596
3597 static struct bpf_map *bpf__find_map_by_name(const char *name)
3598 {
3599         struct bpf_object *obj, *tmp;
3600
3601         bpf_object__for_each_safe(obj, tmp) {
3602                 struct bpf_map *map = bpf_object__find_map_by_name(obj, name);
3603                 if (map)
3604                         return map;
3605
3606         }
3607
3608         return NULL;
3609 }
3610
3611 static void trace__set_bpf_map_filtered_pids(struct trace *trace)
3612 {
3613         trace->filter_pids.map = bpf__find_map_by_name("pids_filtered");
3614 }
3615
3616 static void trace__set_bpf_map_syscalls(struct trace *trace)
3617 {
3618         trace->syscalls.map = bpf__find_map_by_name("syscalls");
3619 }
3620
3621 static int trace__config(const char *var, const char *value, void *arg)
3622 {
3623         struct trace *trace = arg;
3624         int err = 0;
3625
3626         if (!strcmp(var, "trace.add_events")) {
3627                 struct option o = OPT_CALLBACK('e', "event", &trace->evlist, "event",
3628                                                "event selector. use 'perf list' to list available events",
3629                                                parse_events_option);
3630                 err = parse_events_option(&o, value, 0);
3631         } else if (!strcmp(var, "trace.show_timestamp")) {
3632                 trace->show_tstamp = perf_config_bool(var, value);
3633         } else if (!strcmp(var, "trace.show_duration")) {
3634                 trace->show_duration = perf_config_bool(var, value);
3635         } else if (!strcmp(var, "trace.show_arg_names")) {
3636                 trace->show_arg_names = perf_config_bool(var, value);
3637                 if (!trace->show_arg_names)
3638                         trace->show_zeros = true;
3639         } else if (!strcmp(var, "trace.show_zeros")) {
3640                 bool new_show_zeros = perf_config_bool(var, value);
3641                 if (!trace->show_arg_names && !new_show_zeros) {
3642                         pr_warning("trace.show_zeros has to be set when trace.show_arg_names=no\n");
3643                         goto out;
3644                 }
3645                 trace->show_zeros = new_show_zeros;
3646         } else if (!strcmp(var, "trace.show_prefix")) {
3647                 trace->show_string_prefix = perf_config_bool(var, value);
3648         } else if (!strcmp(var, "trace.no_inherit")) {
3649                 trace->opts.no_inherit = perf_config_bool(var, value);
3650         } else if (!strcmp(var, "trace.args_alignment")) {
3651                 int args_alignment = 0;
3652                 if (perf_config_int(&args_alignment, var, value) == 0)
3653                         trace->args_alignment = args_alignment;
3654         }
3655 out:
3656         return err;
3657 }
3658
3659 int cmd_trace(int argc, const char **argv)
3660 {
3661         const char *trace_usage[] = {
3662                 "perf trace [<options>] [<command>]",
3663                 "perf trace [<options>] -- <command> [<options>]",
3664                 "perf trace record [<options>] [<command>]",
3665                 "perf trace record [<options>] -- <command> [<options>]",
3666                 NULL
3667         };
3668         struct trace trace = {
3669                 .syscalls = {
3670                         . max = -1,
3671                 },
3672                 .opts = {
3673                         .target = {
3674                                 .uid       = UINT_MAX,
3675                                 .uses_mmap = true,
3676                         },
3677                         .user_freq     = UINT_MAX,
3678                         .user_interval = ULLONG_MAX,
3679                         .no_buffering  = true,
3680                         .mmap_pages    = UINT_MAX,
3681                 },
3682                 .output = stderr,
3683                 .show_comm = true,
3684                 .show_tstamp = true,
3685                 .show_duration = true,
3686                 .show_arg_names = true,
3687                 .args_alignment = 70,
3688                 .trace_syscalls = false,
3689                 .kernel_syscallchains = false,
3690                 .max_stack = UINT_MAX,
3691                 .max_events = ULONG_MAX,
3692         };
3693         const char *map_dump_str = NULL;
3694         const char *output_name = NULL;
3695         const struct option trace_options[] = {
3696         OPT_CALLBACK('e', "event", &trace, "event",
3697                      "event/syscall selector. use 'perf list' to list available events",
3698                      trace__parse_events_option),
3699         OPT_BOOLEAN(0, "comm", &trace.show_comm,
3700                     "show the thread COMM next to its id"),
3701         OPT_BOOLEAN(0, "tool_stats", &trace.show_tool_stats, "show tool stats"),
3702         OPT_CALLBACK(0, "expr", &trace, "expr", "list of syscalls/events to trace",
3703                      trace__parse_events_option),
3704         OPT_STRING('o', "output", &output_name, "file", "output file name"),
3705         OPT_STRING('i', "input", &input_name, "file", "Analyze events in file"),
3706         OPT_STRING('p', "pid", &trace.opts.target.pid, "pid",
3707                     "trace events on existing process id"),
3708         OPT_STRING('t', "tid", &trace.opts.target.tid, "tid",
3709                     "trace events on existing thread id"),
3710         OPT_CALLBACK(0, "filter-pids", &trace, "CSV list of pids",
3711                      "pids to filter (by the kernel)", trace__set_filter_pids_from_option),
3712         OPT_BOOLEAN('a', "all-cpus", &trace.opts.target.system_wide,
3713                     "system-wide collection from all CPUs"),
3714         OPT_STRING('C', "cpu", &trace.opts.target.cpu_list, "cpu",
3715                     "list of cpus to monitor"),
3716         OPT_BOOLEAN(0, "no-inherit", &trace.opts.no_inherit,
3717                     "child tasks do not inherit counters"),
3718         OPT_CALLBACK('m', "mmap-pages", &trace.opts.mmap_pages, "pages",
3719                      "number of mmap data pages",
3720                      perf_evlist__parse_mmap_pages),
3721         OPT_STRING('u', "uid", &trace.opts.target.uid_str, "user",
3722                    "user to profile"),
3723         OPT_CALLBACK(0, "duration", &trace, "float",
3724                      "show only events with duration > N.M ms",
3725                      trace__set_duration),
3726 #ifdef HAVE_LIBBPF_SUPPORT
3727         OPT_STRING(0, "map-dump", &map_dump_str, "BPF map", "BPF map to periodically dump"),
3728 #endif
3729         OPT_BOOLEAN(0, "sched", &trace.sched, "show blocking scheduler events"),
3730         OPT_INCR('v', "verbose", &verbose, "be more verbose"),
3731         OPT_BOOLEAN('T', "time", &trace.full_time,
3732                     "Show full timestamp, not time relative to first start"),
3733         OPT_BOOLEAN(0, "failure", &trace.failure_only,
3734                     "Show only syscalls that failed"),
3735         OPT_BOOLEAN('s', "summary", &trace.summary_only,
3736                     "Show only syscall summary with statistics"),
3737         OPT_BOOLEAN('S', "with-summary", &trace.summary,
3738                     "Show all syscalls and summary with statistics"),
3739         OPT_CALLBACK_DEFAULT('F', "pf", &trace.trace_pgfaults, "all|maj|min",
3740                      "Trace pagefaults", parse_pagefaults, "maj"),
3741         OPT_BOOLEAN(0, "syscalls", &trace.trace_syscalls, "Trace syscalls"),
3742         OPT_BOOLEAN('f', "force", &trace.force, "don't complain, do it"),
3743         OPT_CALLBACK(0, "call-graph", &trace.opts,
3744                      "record_mode[,record_size]", record_callchain_help,
3745                      &record_parse_callchain_opt),
3746         OPT_BOOLEAN(0, "kernel-syscall-graph", &trace.kernel_syscallchains,
3747                     "Show the kernel callchains on the syscall exit path"),
3748         OPT_ULONG(0, "max-events", &trace.max_events,
3749                 "Set the maximum number of events to print, exit after that is reached. "),
3750         OPT_UINTEGER(0, "min-stack", &trace.min_stack,
3751                      "Set the minimum stack depth when parsing the callchain, "
3752                      "anything below the specified depth will be ignored."),
3753         OPT_UINTEGER(0, "max-stack", &trace.max_stack,
3754                      "Set the maximum stack depth when parsing the callchain, "
3755                      "anything beyond the specified depth will be ignored. "
3756                      "Default: kernel.perf_event_max_stack or " __stringify(PERF_MAX_STACK_DEPTH)),
3757         OPT_BOOLEAN(0, "sort-events", &trace.sort_events,
3758                         "Sort batch of events before processing, use if getting out of order events"),
3759         OPT_BOOLEAN(0, "print-sample", &trace.print_sample,
3760                         "print the PERF_RECORD_SAMPLE PERF_SAMPLE_ info, for debugging"),
3761         OPT_UINTEGER(0, "proc-map-timeout", &proc_map_timeout,
3762                         "per thread proc mmap processing timeout in ms"),
3763         OPT_CALLBACK('G', "cgroup", &trace, "name", "monitor event in cgroup name only",
3764                      trace__parse_cgroups),
3765         OPT_UINTEGER('D', "delay", &trace.opts.initial_delay,
3766                      "ms to wait before starting measurement after program "
3767                      "start"),
3768         OPT_END()
3769         };
3770         bool __maybe_unused max_stack_user_set = true;
3771         bool mmap_pages_user_set = true;
3772         struct perf_evsel *evsel;
3773         const char * const trace_subcommands[] = { "record", NULL };
3774         int err = -1;
3775         char bf[BUFSIZ];
3776
3777         signal(SIGSEGV, sighandler_dump_stack);
3778         signal(SIGFPE, sighandler_dump_stack);
3779
3780         trace.evlist = perf_evlist__new();
3781         trace.sctbl = syscalltbl__new();
3782
3783         if (trace.evlist == NULL || trace.sctbl == NULL) {
3784                 pr_err("Not enough memory to run!\n");
3785                 err = -ENOMEM;
3786                 goto out;
3787         }
3788
3789         err = perf_config(trace__config, &trace);
3790         if (err)
3791                 goto out;
3792
3793         argc = parse_options_subcommand(argc, argv, trace_options, trace_subcommands,
3794                                  trace_usage, PARSE_OPT_STOP_AT_NON_OPTION);
3795
3796         if ((nr_cgroups || trace.cgroup) && !trace.opts.target.system_wide) {
3797                 usage_with_options_msg(trace_usage, trace_options,
3798                                        "cgroup monitoring only available in system-wide mode");
3799         }
3800
3801         evsel = bpf__setup_output_event(trace.evlist, "__augmented_syscalls__");
3802         if (IS_ERR(evsel)) {
3803                 bpf__strerror_setup_output_event(trace.evlist, PTR_ERR(evsel), bf, sizeof(bf));
3804                 pr_err("ERROR: Setup trace syscalls enter failed: %s\n", bf);
3805                 goto out;
3806         }
3807
3808         if (evsel) {
3809                 trace.syscalls.events.augmented = evsel;
3810                 trace__set_bpf_map_filtered_pids(&trace);
3811                 trace__set_bpf_map_syscalls(&trace);
3812         }
3813
3814         err = bpf__setup_stdout(trace.evlist);
3815         if (err) {
3816                 bpf__strerror_setup_stdout(trace.evlist, err, bf, sizeof(bf));
3817                 pr_err("ERROR: Setup BPF stdout failed: %s\n", bf);
3818                 goto out;
3819         }
3820
3821         err = -1;
3822
3823         if (map_dump_str) {
3824                 trace.dump.map = bpf__find_map_by_name(map_dump_str);
3825                 if (trace.dump.map == NULL) {
3826                         pr_err("ERROR: BPF map \"%s\" not found\n", map_dump_str);
3827                         goto out;
3828                 }
3829         }
3830
3831         if (trace.trace_pgfaults) {
3832                 trace.opts.sample_address = true;
3833                 trace.opts.sample_time = true;
3834         }
3835
3836         if (trace.opts.mmap_pages == UINT_MAX)
3837                 mmap_pages_user_set = false;
3838
3839         if (trace.max_stack == UINT_MAX) {
3840                 trace.max_stack = input_name ? PERF_MAX_STACK_DEPTH : sysctl__max_stack();
3841                 max_stack_user_set = false;
3842         }
3843
3844 #ifdef HAVE_DWARF_UNWIND_SUPPORT
3845         if ((trace.min_stack || max_stack_user_set) && !callchain_param.enabled) {
3846                 record_opts__parse_callchain(&trace.opts, &callchain_param, "dwarf", false);
3847         }
3848 #endif
3849
3850         if (callchain_param.enabled) {
3851                 if (!mmap_pages_user_set && geteuid() == 0)
3852                         trace.opts.mmap_pages = perf_event_mlock_kb_in_pages() * 4;
3853
3854                 symbol_conf.use_callchain = true;
3855         }
3856
3857         if (trace.evlist->nr_entries > 0) {
3858                 evlist__set_evsel_handler(trace.evlist, trace__event_handler);
3859                 if (evlist__set_syscall_tp_fields(trace.evlist)) {
3860                         perror("failed to set syscalls:* tracepoint fields");
3861                         goto out;
3862                 }
3863         }
3864
3865         if (trace.sort_events) {
3866                 ordered_events__init(&trace.oe.data, ordered_events__deliver_event, &trace);
3867                 ordered_events__set_copy_on_queue(&trace.oe.data, true);
3868         }
3869
3870         /*
3871          * If we are augmenting syscalls, then combine what we put in the
3872          * __augmented_syscalls__ BPF map with what is in the
3873          * syscalls:sys_exit_FOO tracepoints, i.e. just like we do without BPF,
3874          * combining raw_syscalls:sys_enter with raw_syscalls:sys_exit.
3875          *
3876          * We'll switch to look at two BPF maps, one for sys_enter and the
3877          * other for sys_exit when we start augmenting the sys_exit paths with
3878          * buffers that are being copied from kernel to userspace, think 'read'
3879          * syscall.
3880          */
3881         if (trace.syscalls.events.augmented) {
3882                 evlist__for_each_entry(trace.evlist, evsel) {
3883                         bool raw_syscalls_sys_exit = strcmp(perf_evsel__name(evsel), "raw_syscalls:sys_exit") == 0;
3884
3885                         if (raw_syscalls_sys_exit) {
3886                                 trace.raw_augmented_syscalls = true;
3887                                 goto init_augmented_syscall_tp;
3888                         }
3889
3890                         if (trace.syscalls.events.augmented->priv == NULL &&
3891                             strstr(perf_evsel__name(evsel), "syscalls:sys_enter")) {
3892                                 struct perf_evsel *augmented = trace.syscalls.events.augmented;
3893                                 if (perf_evsel__init_augmented_syscall_tp(augmented, evsel) ||
3894                                     perf_evsel__init_augmented_syscall_tp_args(augmented))
3895                                         goto out;
3896                                 augmented->handler = trace__sys_enter;
3897                         }
3898
3899                         if (strstarts(perf_evsel__name(evsel), "syscalls:sys_exit_")) {
3900                                 struct syscall_tp *sc;
3901 init_augmented_syscall_tp:
3902                                 if (perf_evsel__init_augmented_syscall_tp(evsel, evsel))
3903                                         goto out;
3904                                 sc = evsel->priv;
3905                                 /*
3906                                  * For now with BPF raw_augmented we hook into
3907                                  * raw_syscalls:sys_enter and there we get all
3908                                  * 6 syscall args plus the tracepoint common
3909                                  * fields and the syscall_nr (another long).
3910                                  * So we check if that is the case and if so
3911                                  * don't look after the sc->args_size but
3912                                  * always after the full raw_syscalls:sys_enter
3913                                  * payload, which is fixed.
3914                                  *
3915                                  * We'll revisit this later to pass
3916                                  * s->args_size to the BPF augmenter (now
3917                                  * tools/perf/examples/bpf/augmented_raw_syscalls.c,
3918                                  * so that it copies only what we need for each
3919                                  * syscall, like what happens when we use
3920                                  * syscalls:sys_enter_NAME, so that we reduce
3921                                  * the kernel/userspace traffic to just what is
3922                                  * needed for each syscall.
3923                                  */
3924                                 if (trace.raw_augmented_syscalls)
3925                                         trace.raw_augmented_syscalls_args_size = (6 + 1) * sizeof(long) + sc->id.offset;
3926                                 perf_evsel__init_augmented_syscall_tp_ret(evsel);
3927                                 evsel->handler = trace__sys_exit;
3928                         }
3929                 }
3930         }
3931
3932         if ((argc >= 1) && (strcmp(argv[0], "record") == 0))
3933                 return trace__record(&trace, argc-1, &argv[1]);
3934
3935         /* summary_only implies summary option, but don't overwrite summary if set */
3936         if (trace.summary_only)
3937                 trace.summary = trace.summary_only;
3938
3939         if (!trace.trace_syscalls && !trace.trace_pgfaults &&
3940             trace.evlist->nr_entries == 0 /* Was --events used? */) {
3941                 trace.trace_syscalls = true;
3942         }
3943
3944         if (output_name != NULL) {
3945                 err = trace__open_output(&trace, output_name);
3946                 if (err < 0) {
3947                         perror("failed to create output file");
3948                         goto out;
3949                 }
3950         }
3951
3952         err = target__validate(&trace.opts.target);
3953         if (err) {
3954                 target__strerror(&trace.opts.target, err, bf, sizeof(bf));
3955                 fprintf(trace.output, "%s", bf);
3956                 goto out_close;
3957         }
3958
3959         err = target__parse_uid(&trace.opts.target);
3960         if (err) {
3961                 target__strerror(&trace.opts.target, err, bf, sizeof(bf));
3962                 fprintf(trace.output, "%s", bf);
3963                 goto out_close;
3964         }
3965
3966         if (!argc && target__none(&trace.opts.target))
3967                 trace.opts.target.system_wide = true;
3968
3969         if (input_name)
3970                 err = trace__replay(&trace);
3971         else
3972                 err = trace__run(&trace, argc, argv);
3973
3974 out_close:
3975         if (output_name != NULL)
3976                 fclose(trace.output);
3977 out:
3978         return err;
3979 }