tools/perf/builtin-trace.c

   1 /*
   2  * builtin-trace.c
   3  *
   4  * Builtin 'trace' command:
   5  *
   6  * Display a continuously updated trace of any workload, CPU, specific PID,
   7  * system wide, etc.  Default format is loosely strace like, but any other
   8  * event may be specified using --event.
   9  *
  10  * Copyright (C) 2012, 2013, 2014, 2015 Red Hat Inc, Arnaldo Carvalho de Melo <acme@redhat.com>
  11  *
  12  * Initially based on the 'trace' prototype by Thomas Gleixner:
  13  *
  14  * http://lwn.net/Articles/415728/ ("Announcing a new utility: 'trace'")
  15  *
  16  * Released under the GPL v2. (and only v2, not any later version)
  17  */
  18
  19 #include <traceevent/event-parse.h>
  20 #include <api/fs/tracing_path.h>
  21 #include <bpf/bpf.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/path.h"
  33 #include "util/session.h"
  34 #include "util/thread.h"
  35 #include <subcmd/parse-options.h>
  36 #include "util/strlist.h"
  37 #include "util/intlist.h"
  38 #include "util/thread_map.h"
  39 #include "util/stat.h"
  40 #include "trace/beauty/beauty.h"
  41 #include "trace-event.h"
  42 #include "util/parse-events.h"
  43 #include "util/bpf-loader.h"
  44 #include "callchain.h"
  45 #include "print_binary.h"
  46 #include "string2.h"
  47 #include "syscalltbl.h"
  48 #include "rb_resort.h"
  49
  50 #include <errno.h>
  51 #include <inttypes.h>
  52 #include <poll.h>
  53 #include <signal.h>
  54 #include <stdlib.h>
  55 #include <string.h>
  56 #include <linux/err.h>
  57 #include <linux/filter.h>
  58 #include <linux/kernel.h>
  59 #include <linux/random.h>
  60 #include <linux/stringify.h>
  61 #include <linux/time64.h>
  62 #include <fcntl.h>
  63 #include <sys/sysmacros.h>
  64
  65 #include "sane_ctype.h"
  66
  67 #ifndef O_CLOEXEC
  68 # define O_CLOEXEC              02000000
  69 #endif
  70
  71 #ifndef F_LINUX_SPECIFIC_BASE
  72 # define F_LINUX_SPECIFIC_BASE  1024
  73 #endif
  74
  75 struct trace {
  76         struct perf_tool        tool;
  77         struct syscalltbl       *sctbl;
  78         struct {
  79                 int             max;
  80                 struct syscall  *table;
  81                 struct bpf_map  *map;
  82                 struct {
  83                         struct perf_evsel *sys_enter,
  84                                           *sys_exit,
  85                                           *augmented;
  86                 }               events;
  87         } syscalls;
  88         struct record_opts      opts;
  89         struct perf_evlist      *evlist;
  90         struct machine          *host;
  91         struct thread           *current;
  92         struct cgroup           *cgroup;
  93         u64                     base_time;
  94         FILE                    *output;
  95         unsigned long           nr_events;
  96         unsigned long           nr_events_printed;
  97         unsigned long           max_events;
  98         struct strlist          *ev_qualifier;
  99         struct {
 100                 size_t          nr;
 101                 int             *entries;
 102         }                       ev_qualifier_ids;
 103         struct {
 104                 size_t          nr;
 105                 pid_t           *entries;
 106                 struct bpf_map  *map;
 107         }                       filter_pids;
 108         double                  duration_filter;
 109         double                  runtime_ms;
 110         struct {
 111                 u64             vfs_getname,
 112                                 proc_getname;
 113         } stats;
 114         unsigned int            max_stack;
 115         unsigned int            min_stack;
 116         int                     raw_augmented_syscalls_args_size;
 117         bool                    raw_augmented_syscalls;
 118         bool                    sort_events;
 119         bool                    not_ev_qualifier;
 120         bool                    live;
 121         bool                    full_time;
 122         bool                    sched;
 123         bool                    multiple_threads;
 124         bool                    summary;
 125         bool                    summary_only;
 126         bool                    failure_only;
 127         bool                    show_comm;
 128         bool                    print_sample;
 129         bool                    show_tool_stats;
 130         bool                    trace_syscalls;
 131         bool                    kernel_syscallchains;
 132         s16                     args_alignment;
 133         bool                    show_tstamp;
 134         bool                    show_duration;
 135         bool                    show_zeros;
 136         bool                    show_arg_names;
 137         bool                    show_string_prefix;
 138         bool                    force;
 139         bool                    vfs_getname;
 140         int                     trace_pgfaults;
 141         struct {
 142                 struct ordered_events   data;
 143                 u64                     last;
 144         } oe;
 145 };
 146
 147 struct tp_field {
 148         int offset;
 149         union {
 150                 u64 (*integer)(struct tp_field *field, struct perf_sample *sample);
 151                 void *(*pointer)(struct tp_field *field, struct perf_sample *sample);
 152         };
 153 };
 154
 155 #define TP_UINT_FIELD(bits) \
 156 static u64 tp_field__u##bits(struct tp_field *field, struct perf_sample *sample) \
 157 { \
 158         u##bits value; \
 159         memcpy(&value, sample->raw_data + field->offset, sizeof(value)); \
 160         return value;  \
 161 }
 162
 163 TP_UINT_FIELD(8);
 164 TP_UINT_FIELD(16);
 165 TP_UINT_FIELD(32);
 166 TP_UINT_FIELD(64);
 167
 168 #define TP_UINT_FIELD__SWAPPED(bits) \
 169 static u64 tp_field__swapped_u##bits(struct tp_field *field, struct perf_sample *sample) \
 170 { \
 171         u##bits value; \
 172         memcpy(&value, sample->raw_data + field->offset, sizeof(value)); \
 173         return bswap_##bits(value);\
 174 }
 175
 176 TP_UINT_FIELD__SWAPPED(16);
 177 TP_UINT_FIELD__SWAPPED(32);
 178 TP_UINT_FIELD__SWAPPED(64);
 179
 180 static int __tp_field__init_uint(struct tp_field *field, int size, int offset, bool needs_swap)
 181 {
 182         field->offset = offset;
 183
 184         switch (size) {
 185         case 1:
 186                 field->integer = tp_field__u8;
 187                 break;
 188         case 2:
 189                 field->integer = needs_swap ? tp_field__swapped_u16 : tp_field__u16;
 190                 break;
 191         case 4:
 192                 field->integer = needs_swap ? tp_field__swapped_u32 : tp_field__u32;
 193                 break;
 194         case 8:
 195                 field->integer = needs_swap ? tp_field__swapped_u64 : tp_field__u64;
 196                 break;
 197         default:
 198                 return -1;
 199         }
 200
 201         return 0;
 202 }
 203
 204 static int tp_field__init_uint(struct tp_field *field, struct tep_format_field *format_field, bool needs_swap)
 205 {
 206         return __tp_field__init_uint(field, format_field->size, format_field->offset, needs_swap);
 207 }
 208
 209 static void *tp_field__ptr(struct tp_field *field, struct perf_sample *sample)
 210 {
 211         return sample->raw_data + field->offset;
 212 }
 213
 214 static int __tp_field__init_ptr(struct tp_field *field, int offset)
 215 {
 216         field->offset = offset;
 217         field->pointer = tp_field__ptr;
 218         return 0;
 219 }
 220
 221 static int tp_field__init_ptr(struct tp_field *field, struct tep_format_field *format_field)
 222 {
 223         return __tp_field__init_ptr(field, format_field->offset);
 224 }
 225
 226 struct syscall_tp {
 227         struct tp_field id;
 228         union {
 229                 struct tp_field args, ret;
 230         };
 231 };
 232
 233 static int perf_evsel__init_tp_uint_field(struct perf_evsel *evsel,
 234                                           struct tp_field *field,
 235                                           const char *name)
 236 {
 237         struct tep_format_field *format_field = perf_evsel__field(evsel, name);
 238
 239         if (format_field == NULL)
 240                 return -1;
 241
 242         return tp_field__init_uint(field, format_field, evsel->needs_swap);
 243 }
 244
 245 #define perf_evsel__init_sc_tp_uint_field(evsel, name) \
 246         ({ struct syscall_tp *sc = evsel->priv;\
 247            perf_evsel__init_tp_uint_field(evsel, &sc->name, #name); })
 248
 249 static int perf_evsel__init_tp_ptr_field(struct perf_evsel *evsel,
 250                                          struct tp_field *field,
 251                                          const char *name)
 252 {
 253         struct tep_format_field *format_field = perf_evsel__field(evsel, name);
 254
 255         if (format_field == NULL)
 256                 return -1;
 257
 258         return tp_field__init_ptr(field, format_field);
 259 }
 260
 261 #define perf_evsel__init_sc_tp_ptr_field(evsel, name) \
 262         ({ struct syscall_tp *sc = evsel->priv;\
 263            perf_evsel__init_tp_ptr_field(evsel, &sc->name, #name); })
 264
 265 static void perf_evsel__delete_priv(struct perf_evsel *evsel)
 266 {
 267         zfree(&evsel->priv);
 268         perf_evsel__delete(evsel);
 269 }
 270
 271 static int perf_evsel__init_syscall_tp(struct perf_evsel *evsel)
 272 {
 273         struct syscall_tp *sc = evsel->priv = malloc(sizeof(struct syscall_tp));
 274
 275         if (evsel->priv != NULL) {
 276                 if (perf_evsel__init_tp_uint_field(evsel, &sc->id, "__syscall_nr") &&
 277                     perf_evsel__init_tp_uint_field(evsel, &sc->id, "nr"))
 278                         goto out_delete;
 279                 return 0;
 280         }
 281
 282         return -ENOMEM;
 283 out_delete:
 284         zfree(&evsel->priv);
 285         return -ENOENT;
 286 }
 287
 288 static int perf_evsel__init_augmented_syscall_tp(struct perf_evsel *evsel, struct perf_evsel *tp)
 289 {
 290         struct syscall_tp *sc = evsel->priv = malloc(sizeof(struct syscall_tp));
 291
 292         if (evsel->priv != NULL) {
 293                 struct tep_format_field *syscall_id = perf_evsel__field(tp, "id");
 294                 if (syscall_id == NULL)
 295                         syscall_id = perf_evsel__field(tp, "__syscall_nr");
 296                 if (syscall_id == NULL)
 297                         goto out_delete;
 298                 if (__tp_field__init_uint(&sc->id, syscall_id->size, syscall_id->offset, evsel->needs_swap))
 299                         goto out_delete;
 300
 301                 return 0;
 302         }
 303
 304         return -ENOMEM;
 305 out_delete:
 306         zfree(&evsel->priv);
 307         return -EINVAL;
 308 }
 309
 310 static int perf_evsel__init_augmented_syscall_tp_args(struct perf_evsel *evsel)
 311 {
 312         struct syscall_tp *sc = evsel->priv;
 313
 314         return __tp_field__init_ptr(&sc->args, sc->id.offset + sizeof(u64));
 315 }
 316
 317 static int perf_evsel__init_augmented_syscall_tp_ret(struct perf_evsel *evsel)
 318 {
 319         struct syscall_tp *sc = evsel->priv;
 320
 321         return __tp_field__init_uint(&sc->ret, sizeof(u64), sc->id.offset + sizeof(u64), evsel->needs_swap);
 322 }
 323
 324 static int perf_evsel__init_raw_syscall_tp(struct perf_evsel *evsel, void *handler)
 325 {
 326         evsel->priv = malloc(sizeof(struct syscall_tp));
 327         if (evsel->priv != NULL) {
 328                 if (perf_evsel__init_sc_tp_uint_field(evsel, id))
 329                         goto out_delete;
 330
 331                 evsel->handler = handler;
 332                 return 0;
 333         }
 334
 335         return -ENOMEM;
 336
 337 out_delete:
 338         zfree(&evsel->priv);
 339         return -ENOENT;
 340 }
 341
 342 static struct perf_evsel *perf_evsel__raw_syscall_newtp(const char *direction, void *handler)
 343 {
 344         struct perf_evsel *evsel = perf_evsel__newtp("raw_syscalls", direction);
 345
 346         /* older kernel (e.g., RHEL6) use syscalls:{enter,exit} */
 347         if (IS_ERR(evsel))
 348                 evsel = perf_evsel__newtp("syscalls", direction);
 349
 350         if (IS_ERR(evsel))
 351                 return NULL;
 352
 353         if (perf_evsel__init_raw_syscall_tp(evsel, handler))
 354                 goto out_delete;
 355
 356         return evsel;
 357
 358 out_delete:
 359         perf_evsel__delete_priv(evsel);
 360         return NULL;
 361 }
 362
 363 #define perf_evsel__sc_tp_uint(evsel, name, sample) \
 364         ({ struct syscall_tp *fields = evsel->priv; \
 365            fields->name.integer(&fields->name, sample); })
 366
 367 #define perf_evsel__sc_tp_ptr(evsel, name, sample) \
 368         ({ struct syscall_tp *fields = evsel->priv; \
 369            fields->name.pointer(&fields->name, sample); })
 370
 371 size_t strarray__scnprintf(struct strarray *sa, char *bf, size_t size, const char *intfmt, bool show_prefix, int val)
 372 {
 373         int idx = val - sa->offset;
 374
 375         if (idx < 0 || idx >= sa->nr_entries || sa->entries[idx] == NULL) {
 376                 size_t printed = scnprintf(bf, size, intfmt, val);
 377                 if (show_prefix)
 378                         printed += scnprintf(bf + printed, size - printed, " /* %s??? */", sa->prefix);
 379                 return printed;
 380         }
 381
 382         return scnprintf(bf, size, "%s%s", show_prefix ? sa->prefix : "", sa->entries[idx]);
 383 }
 384
 385 static size_t __syscall_arg__scnprintf_strarray(char *bf, size_t size,
 386                                                 const char *intfmt,
 387                                                 struct syscall_arg *arg)
 388 {
 389         return strarray__scnprintf(arg->parm, bf, size, intfmt, arg->show_string_prefix, arg->val);
 390 }
 391
 392 static size_t syscall_arg__scnprintf_strarray(char *bf, size_t size,
 393                                               struct syscall_arg *arg)
 394 {
 395         return __syscall_arg__scnprintf_strarray(bf, size, "%d", arg);
 396 }
 397
 398 #define SCA_STRARRAY syscall_arg__scnprintf_strarray
 399
 400 size_t strarrays__scnprintf(struct strarrays *sas, char *bf, size_t size, const char *intfmt, bool show_prefix, int val)
 401 {
 402         size_t printed;
 403         int i;
 404
 405         for (i = 0; i < sas->nr_entries; ++i) {
 406                 struct strarray *sa = sas->entries[i];
 407                 int idx = val - sa->offset;
 408
 409                 if (idx >= 0 && idx < sa->nr_entries) {
 410                         if (sa->entries[idx] == NULL)
 411                                 break;
 412                         return scnprintf(bf, size, "%s%s", show_prefix ? sa->prefix : "", sa->entries[idx]);
 413                 }
 414         }
 415
 416         printed = scnprintf(bf, size, intfmt, val);
 417         if (show_prefix)
 418                 printed += scnprintf(bf + printed, size - printed, " /* %s??? */", sas->entries[0]->prefix);
 419         return printed;
 420 }
 421
 422 size_t syscall_arg__scnprintf_strarrays(char *bf, size_t size,
 423                                         struct syscall_arg *arg)
 424 {
 425         return strarrays__scnprintf(arg->parm, bf, size, "%d", arg->show_string_prefix, arg->val);
 426 }
 427
 428 #ifndef AT_FDCWD
 429 #define AT_FDCWD        -100
 430 #endif
 431
 432 static size_t syscall_arg__scnprintf_fd_at(char *bf, size_t size,
 433                                            struct syscall_arg *arg)
 434 {
 435         int fd = arg->val;
 436         const char *prefix = "AT_FD";
 437
 438         if (fd == AT_FDCWD)
 439                 return scnprintf(bf, size, "%s%s", arg->show_string_prefix ? prefix : "", "CWD");
 440
 441         return syscall_arg__scnprintf_fd(bf, size, arg);
 442 }
 443
 444 #define SCA_FDAT syscall_arg__scnprintf_fd_at
 445
 446 static size_t syscall_arg__scnprintf_close_fd(char *bf, size_t size,
 447                                               struct syscall_arg *arg);
 448
 449 #define SCA_CLOSE_FD syscall_arg__scnprintf_close_fd
 450
 451 size_t syscall_arg__scnprintf_hex(char *bf, size_t size, struct syscall_arg *arg)
 452 {
 453         return scnprintf(bf, size, "%#lx", arg->val);
 454 }
 455
 456 size_t syscall_arg__scnprintf_ptr(char *bf, size_t size, struct syscall_arg *arg)
 457 {
 458         if (arg->val == 0)
 459                 return scnprintf(bf, size, "NULL");
 460         return syscall_arg__scnprintf_hex(bf, size, arg);
 461 }
 462
 463 size_t syscall_arg__scnprintf_int(char *bf, size_t size, struct syscall_arg *arg)
 464 {
 465         return scnprintf(bf, size, "%d", arg->val);
 466 }
 467
 468 size_t syscall_arg__scnprintf_long(char *bf, size_t size, struct syscall_arg *arg)
 469 {
 470         return scnprintf(bf, size, "%ld", arg->val);
 471 }
 472
 473 static const char *bpf_cmd[] = {
 474         "MAP_CREATE", "MAP_LOOKUP_ELEM", "MAP_UPDATE_ELEM", "MAP_DELETE_ELEM",
 475         "MAP_GET_NEXT_KEY", "PROG_LOAD",
 476 };
 477 static DEFINE_STRARRAY(bpf_cmd, "BPF_");
 478
 479 static const char *epoll_ctl_ops[] = { "ADD", "DEL", "MOD", };
 480 static DEFINE_STRARRAY_OFFSET(epoll_ctl_ops, "EPOLL_CTL_", 1);
 481
 482 static const char *itimers[] = { "REAL", "VIRTUAL", "PROF", };
 483 static DEFINE_STRARRAY(itimers, "ITIMER_");
 484
 485 static const char *keyctl_options[] = {
 486         "GET_KEYRING_ID", "JOIN_SESSION_KEYRING", "UPDATE", "REVOKE", "CHOWN",
 487         "SETPERM", "DESCRIBE", "CLEAR", "LINK", "UNLINK", "SEARCH", "READ",
 488         "INSTANTIATE", "NEGATE", "SET_REQKEY_KEYRING", "SET_TIMEOUT",
 489         "ASSUME_AUTHORITY", "GET_SECURITY", "SESSION_TO_PARENT", "REJECT",
 490         "INSTANTIATE_IOV", "INVALIDATE", "GET_PERSISTENT",
 491 };
 492 static DEFINE_STRARRAY(keyctl_options, "KEYCTL_");
 493
 494 static const char *whences[] = { "SET", "CUR", "END",
 495 #ifdef SEEK_DATA
 496 "DATA",
 497 #endif
 498 #ifdef SEEK_HOLE
 499 "HOLE",
 500 #endif
 501 };
 502 static DEFINE_STRARRAY(whences, "SEEK_");
 503
 504 static const char *fcntl_cmds[] = {
 505         "DUPFD", "GETFD", "SETFD", "GETFL", "SETFL", "GETLK", "SETLK",
 506         "SETLKW", "SETOWN", "GETOWN", "SETSIG", "GETSIG", "GETLK64",
 507         "SETLK64", "SETLKW64", "SETOWN_EX", "GETOWN_EX",
 508         "GETOWNER_UIDS",
 509 };
 510 static DEFINE_STRARRAY(fcntl_cmds, "F_");
 511
 512 static const char *fcntl_linux_specific_cmds[] = {
 513         "SETLEASE", "GETLEASE", "NOTIFY", [5] = "CANCELLK", "DUPFD_CLOEXEC",
 514         "SETPIPE_SZ", "GETPIPE_SZ", "ADD_SEALS", "GET_SEALS",
 515         "GET_RW_HINT", "SET_RW_HINT", "GET_FILE_RW_HINT", "SET_FILE_RW_HINT",
 516 };
 517
 518 static DEFINE_STRARRAY_OFFSET(fcntl_linux_specific_cmds, "F_", F_LINUX_SPECIFIC_BASE);
 519
 520 static struct strarray *fcntl_cmds_arrays[] = {
 521         &strarray__fcntl_cmds,
 522         &strarray__fcntl_linux_specific_cmds,
 523 };
 524
 525 static DEFINE_STRARRAYS(fcntl_cmds_arrays);
 526
 527 static const char *rlimit_resources[] = {
 528         "CPU", "FSIZE", "DATA", "STACK", "CORE", "RSS", "NPROC", "NOFILE",
 529         "MEMLOCK", "AS", "LOCKS", "SIGPENDING", "MSGQUEUE", "NICE", "RTPRIO",
 530         "RTTIME",
 531 };
 532 static DEFINE_STRARRAY(rlimit_resources, "RLIMIT_");
 533
 534 static const char *sighow[] = { "BLOCK", "UNBLOCK", "SETMASK", };
 535 static DEFINE_STRARRAY(sighow, "SIG_");
 536
 537 static const char *clockid[] = {
 538         "REALTIME", "MONOTONIC", "PROCESS_CPUTIME_ID", "THREAD_CPUTIME_ID",
 539         "MONOTONIC_RAW", "REALTIME_COARSE", "MONOTONIC_COARSE", "BOOTTIME",
 540         "REALTIME_ALARM", "BOOTTIME_ALARM", "SGI_CYCLE", "TAI"
 541 };
 542 static DEFINE_STRARRAY(clockid, "CLOCK_");
 543
 544 static size_t syscall_arg__scnprintf_access_mode(char *bf, size_t size,
 545                                                  struct syscall_arg *arg)
 546 {
 547         bool show_prefix = arg->show_string_prefix;
 548         const char *suffix = "_OK";
 549         size_t printed = 0;
 550         int mode = arg->val;
 551
 552         if (mode == F_OK) /* 0 */
 553                 return scnprintf(bf, size, "F%s", show_prefix ? suffix : "");
 554 #define P_MODE(n) \
 555         if (mode & n##_OK) { \
 556                 printed += scnprintf(bf + printed, size - printed, "%s%s", #n, show_prefix ? suffix : ""); \
 557                 mode &= ~n##_OK; \
 558         }
 559
 560         P_MODE(R);
 561         P_MODE(W);
 562         P_MODE(X);
 563 #undef P_MODE
 564
 565         if (mode)
 566                 printed += scnprintf(bf + printed, size - printed, "|%#x", mode);
 567
 568         return printed;
 569 }
 570
 571 #define SCA_ACCMODE syscall_arg__scnprintf_access_mode
 572
 573 static size_t syscall_arg__scnprintf_filename(char *bf, size_t size,
 574                                               struct syscall_arg *arg);
 575
 576 #define SCA_FILENAME syscall_arg__scnprintf_filename
 577
 578 static size_t syscall_arg__scnprintf_pipe_flags(char *bf, size_t size,
 579                                                 struct syscall_arg *arg)
 580 {
 581         bool show_prefix = arg->show_string_prefix;
 582         const char *prefix = "O_";
 583         int printed = 0, flags = arg->val;
 584
 585 #define P_FLAG(n) \
 586         if (flags & O_##n) { \
 587                 printed += scnprintf(bf + printed, size - printed, "%s%s%s", printed ? "|" : "", show_prefix ? prefix : "", #n); \
 588                 flags &= ~O_##n; \
 589         }
 590
 591         P_FLAG(CLOEXEC);
 592         P_FLAG(NONBLOCK);
 593 #undef P_FLAG
 594
 595         if (flags)
 596                 printed += scnprintf(bf + printed, size - printed, "%s%#x", printed ? "|" : "", flags);
 597
 598         return printed;
 599 }
 600
 601 #define SCA_PIPE_FLAGS syscall_arg__scnprintf_pipe_flags
 602
 603 #ifndef GRND_NONBLOCK
 604 #define GRND_NONBLOCK   0x0001
 605 #endif
 606 #ifndef GRND_RANDOM
 607 #define GRND_RANDOM     0x0002
 608 #endif
 609
 610 static size_t syscall_arg__scnprintf_getrandom_flags(char *bf, size_t size,
 611                                                    struct syscall_arg *arg)
 612 {
 613         bool show_prefix = arg->show_string_prefix;
 614         const char *prefix = "GRND_";
 615         int printed = 0, flags = arg->val;
 616
 617 #define P_FLAG(n) \
 618         if (flags & GRND_##n) { \
 619                 printed += scnprintf(bf + printed, size - printed, "%s%s%s", printed ? "|" : "", show_prefix ? prefix : "", #n); \
 620                 flags &= ~GRND_##n; \
 621         }
 622
 623         P_FLAG(RANDOM);
 624         P_FLAG(NONBLOCK);
 625 #undef P_FLAG
 626
 627         if (flags)
 628                 printed += scnprintf(bf + printed, size - printed, "%s%#x", printed ? "|" : "", flags);
 629
 630         return printed;
 631 }
 632
 633 #define SCA_GETRANDOM_FLAGS syscall_arg__scnprintf_getrandom_flags
 634
 635 #define STRARRAY(name, array) \
 636           { .scnprintf  = SCA_STRARRAY, \
 637             .parm       = &strarray__##array, }
 638
 639 #include "trace/beauty/arch_errno_names.c"
 640 #include "trace/beauty/eventfd.c"
 641 #include "trace/beauty/futex_op.c"
 642 #include "trace/beauty/futex_val3.c"
 643 #include "trace/beauty/mmap.c"
 644 #include "trace/beauty/mode_t.c"
 645 #include "trace/beauty/msg_flags.c"
 646 #include "trace/beauty/open_flags.c"
 647 #include "trace/beauty/perf_event_open.c"
 648 #include "trace/beauty/pid.c"
 649 #include "trace/beauty/sched_policy.c"
 650 #include "trace/beauty/seccomp.c"
 651 #include "trace/beauty/signum.c"
 652 #include "trace/beauty/socket_type.c"
 653 #include "trace/beauty/waitid_options.c"
 654
 655 struct syscall_arg_fmt {
 656         size_t     (*scnprintf)(char *bf, size_t size, struct syscall_arg *arg);
 657         unsigned long (*mask_val)(struct syscall_arg *arg, unsigned long val);
 658         void       *parm;
 659         const char *name;
 660         bool       show_zero;
 661 };
 662
 663 static struct syscall_fmt {
 664         const char *name;
 665         const char *alias;
 666         struct syscall_arg_fmt arg[6];
 667         u8         nr_args;
 668         bool       errpid;
 669         bool       timeout;
 670         bool       hexret;
 671 } syscall_fmts[] = {
 672         { .name     = "access",
 673           .arg = { [1] = { .scnprintf = SCA_ACCMODE,  /* mode */ }, }, },
 674         { .name     = "arch_prctl",
 675           .arg = { [0] = { .scnprintf = SCA_X86_ARCH_PRCTL_CODE, /* code */ },
 676                    [1] = { .scnprintf = SCA_PTR, /* arg2 */ }, }, },
 677         { .name     = "bind",
 678           .arg = { [1] = { .scnprintf = SCA_SOCKADDR, /* umyaddr */ }, }, },
 679         { .name     = "bpf",
 680           .arg = { [0] = STRARRAY(cmd, bpf_cmd), }, },
 681         { .name     = "brk",        .hexret = true,
 682           .arg = { [0] = { .scnprintf = SCA_PTR, /* brk */ }, }, },
 683         { .name     = "clock_gettime",
 684           .arg = { [0] = STRARRAY(clk_id, clockid), }, },
 685         { .name     = "clone",      .errpid = true, .nr_args = 5,
 686           .arg = { [0] = { .name = "flags",         .scnprintf = SCA_CLONE_FLAGS, },
 687                    [1] = { .name = "child_stack",   .scnprintf = SCA_HEX, },
 688                    [2] = { .name = "parent_tidptr", .scnprintf = SCA_HEX, },
 689                    [3] = { .name = "child_tidptr",  .scnprintf = SCA_HEX, },
 690                    [4] = { .name = "tls",           .scnprintf = SCA_HEX, }, }, },
 691         { .name     = "close",
 692           .arg = { [0] = { .scnprintf = SCA_CLOSE_FD, /* fd */ }, }, },
 693         { .name     = "connect",
 694           .arg = { [1] = { .scnprintf = SCA_SOCKADDR, /* servaddr */ }, }, },
 695         { .name     = "epoll_ctl",
 696           .arg = { [1] = STRARRAY(op, epoll_ctl_ops), }, },
 697         { .name     = "eventfd2",
 698           .arg = { [1] = { .scnprintf = SCA_EFD_FLAGS, /* flags */ }, }, },
 699         { .name     = "fchmodat",
 700           .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
 701         { .name     = "fchownat",
 702           .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
 703         { .name     = "fcntl",
 704           .arg = { [1] = { .scnprintf = SCA_FCNTL_CMD, /* cmd */
 705                            .parm      = &strarrays__fcntl_cmds_arrays,
 706                            .show_zero = true, },
 707                    [2] = { .scnprintf =  SCA_FCNTL_ARG, /* arg */ }, }, },
 708         { .name     = "flock",
 709           .arg = { [1] = { .scnprintf = SCA_FLOCK, /* cmd */ }, }, },
 710         { .name     = "fstat", .alias = "newfstat", },
 711         { .name     = "fstatat", .alias = "newfstatat", },
 712         { .name     = "futex",
 713           .arg = { [1] = { .scnprintf = SCA_FUTEX_OP, /* op */ },
 714                    [5] = { .scnprintf = SCA_FUTEX_VAL3, /* val3 */ }, }, },
 715         { .name     = "futimesat",
 716           .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
 717         { .name     = "getitimer",
 718           .arg = { [0] = STRARRAY(which, itimers), }, },
 719         { .name     = "getpid",     .errpid = true, },
 720         { .name     = "getpgid",    .errpid = true, },
 721         { .name     = "getppid",    .errpid = true, },
 722         { .name     = "getrandom",
 723           .arg = { [2] = { .scnprintf = SCA_GETRANDOM_FLAGS, /* flags */ }, }, },
 724         { .name     = "getrlimit",
 725           .arg = { [0] = STRARRAY(resource, rlimit_resources), }, },
 726         { .name     = "gettid",     .errpid = true, },
 727         { .name     = "ioctl",
 728           .arg = {
 729 #if defined(__i386__) || defined(__x86_64__)
 730 /*
 731  * FIXME: Make this available to all arches.
 732  */
 733                    [1] = { .scnprintf = SCA_IOCTL_CMD, /* cmd */ },
 734                    [2] = { .scnprintf = SCA_HEX, /* arg */ }, }, },
 735 #else
 736                    [2] = { .scnprintf = SCA_HEX, /* arg */ }, }, },
 737 #endif
 738         { .name     = "kcmp",       .nr_args = 5,
 739           .arg = { [0] = { .name = "pid1",      .scnprintf = SCA_PID, },
 740                    [1] = { .name = "pid2",      .scnprintf = SCA_PID, },
 741                    [2] = { .name = "type",      .scnprintf = SCA_KCMP_TYPE, },
 742                    [3] = { .name = "idx1",      .scnprintf = SCA_KCMP_IDX, },
 743                    [4] = { .name = "idx2",      .scnprintf = SCA_KCMP_IDX, }, }, },
 744         { .name     = "keyctl",
 745           .arg = { [0] = STRARRAY(option, keyctl_options), }, },
 746         { .name     = "kill",
 747           .arg = { [1] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
 748         { .name     = "linkat",
 749           .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
 750         { .name     = "lseek",
 751           .arg = { [2] = STRARRAY(whence, whences), }, },
 752         { .name     = "lstat", .alias = "newlstat", },
 753         { .name     = "madvise",
 754           .arg = { [0] = { .scnprintf = SCA_HEX,      /* start */ },
 755                    [2] = { .scnprintf = SCA_MADV_BHV, /* behavior */ }, }, },
 756         { .name     = "mkdirat",
 757           .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
 758         { .name     = "mknodat",
 759           .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
 760         { .name     = "mmap",       .hexret = true,
 761 /* The standard mmap maps to old_mmap on s390x */
 762 #if defined(__s390x__)
 763         .alias = "old_mmap",
 764 #endif
 765           .arg = { [2] = { .scnprintf = SCA_MMAP_PROT,  /* prot */ },
 766                    [3] = { .scnprintf = SCA_MMAP_FLAGS, /* flags */ },
 767                    [5] = { .scnprintf = SCA_HEX,        /* offset */ }, }, },
 768         { .name     = "mount",
 769           .arg = { [0] = { .scnprintf = SCA_FILENAME, /* dev_name */ },
 770                    [3] = { .scnprintf = SCA_MOUNT_FLAGS, /* flags */
 771                            .mask_val  = SCAMV_MOUNT_FLAGS, /* flags */ }, }, },
 772         { .name     = "mprotect",
 773           .arg = { [0] = { .scnprintf = SCA_HEX,        /* start */ },
 774                    [2] = { .scnprintf = SCA_MMAP_PROT,  /* prot */ }, }, },
 775         { .name     = "mq_unlink",
 776           .arg = { [0] = { .scnprintf = SCA_FILENAME, /* u_name */ }, }, },
 777         { .name     = "mremap",     .hexret = true,
 778           .arg = { [3] = { .scnprintf = SCA_MREMAP_FLAGS, /* flags */ }, }, },
 779         { .name     = "name_to_handle_at",
 780           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
 781         { .name     = "newfstatat",
 782           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
 783         { .name     = "open",
 784           .arg = { [1] = { .scnprintf = SCA_OPEN_FLAGS, /* flags */ }, }, },
 785         { .name     = "open_by_handle_at",
 786           .arg = { [0] = { .scnprintf = SCA_FDAT,       /* dfd */ },
 787                    [2] = { .scnprintf = SCA_OPEN_FLAGS, /* flags */ }, }, },
 788         { .name     = "openat",
 789           .arg = { [0] = { .scnprintf = SCA_FDAT,       /* dfd */ },
 790                    [2] = { .scnprintf = SCA_OPEN_FLAGS, /* flags */ }, }, },
 791         { .name     = "perf_event_open",
 792           .arg = { [2] = { .scnprintf = SCA_INT,        /* cpu */ },
 793                    [3] = { .scnprintf = SCA_FD,         /* group_fd */ },
 794                    [4] = { .scnprintf = SCA_PERF_FLAGS, /* flags */ }, }, },
 795         { .name     = "pipe2",
 796           .arg = { [1] = { .scnprintf = SCA_PIPE_FLAGS, /* flags */ }, }, },
 797         { .name     = "pkey_alloc",
 798           .arg = { [1] = { .scnprintf = SCA_PKEY_ALLOC_ACCESS_RIGHTS,   /* access_rights */ }, }, },
 799         { .name     = "pkey_free",
 800           .arg = { [0] = { .scnprintf = SCA_INT,        /* key */ }, }, },
 801         { .name     = "pkey_mprotect",
 802           .arg = { [0] = { .scnprintf = SCA_HEX,        /* start */ },
 803                    [2] = { .scnprintf = SCA_MMAP_PROT,  /* prot */ },
 804                    [3] = { .scnprintf = SCA_INT,        /* pkey */ }, }, },
 805         { .name     = "poll", .timeout = true, },
 806         { .name     = "ppoll", .timeout = true, },
 807         { .name     = "prctl",
 808           .arg = { [0] = { .scnprintf = SCA_PRCTL_OPTION, /* option */ },
 809                    [1] = { .scnprintf = SCA_PRCTL_ARG2, /* arg2 */ },
 810                    [2] = { .scnprintf = SCA_PRCTL_ARG3, /* arg3 */ }, }, },
 811         { .name     = "pread", .alias = "pread64", },
 812         { .name     = "preadv", .alias = "pread", },
 813         { .name     = "prlimit64",
 814           .arg = { [1] = STRARRAY(resource, rlimit_resources), }, },
 815         { .name     = "pwrite", .alias = "pwrite64", },
 816         { .name     = "readlinkat",
 817           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
 818         { .name     = "recvfrom",
 819           .arg = { [3] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
 820         { .name     = "recvmmsg",
 821           .arg = { [3] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
 822         { .name     = "recvmsg",
 823           .arg = { [2] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
 824         { .name     = "renameat",
 825           .arg = { [0] = { .scnprintf = SCA_FDAT, /* olddirfd */ },
 826                    [2] = { .scnprintf = SCA_FDAT, /* newdirfd */ }, }, },
 827         { .name     = "renameat2",
 828           .arg = { [0] = { .scnprintf = SCA_FDAT, /* olddirfd */ },
 829                    [2] = { .scnprintf = SCA_FDAT, /* newdirfd */ },
 830                    [4] = { .scnprintf = SCA_RENAMEAT2_FLAGS, /* flags */ }, }, },
 831         { .name     = "rt_sigaction",
 832           .arg = { [0] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
 833         { .name     = "rt_sigprocmask",
 834           .arg = { [0] = STRARRAY(how, sighow), }, },
 835         { .name     = "rt_sigqueueinfo",
 836           .arg = { [1] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
 837         { .name     = "rt_tgsigqueueinfo",
 838           .arg = { [2] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
 839         { .name     = "sched_setscheduler",
 840           .arg = { [1] = { .scnprintf = SCA_SCHED_POLICY, /* policy */ }, }, },
 841         { .name     = "seccomp",
 842           .arg = { [0] = { .scnprintf = SCA_SECCOMP_OP,    /* op */ },
 843                    [1] = { .scnprintf = SCA_SECCOMP_FLAGS, /* flags */ }, }, },
 844         { .name     = "select", .timeout = true, },
 845         { .name     = "sendmmsg",
 846           .arg = { [3] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
 847         { .name     = "sendmsg",
 848           .arg = { [2] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
 849         { .name     = "sendto",
 850           .arg = { [3] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ },
 851                    [4] = { .scnprintf = SCA_SOCKADDR, /* addr */ }, }, },
 852         { .name     = "set_tid_address", .errpid = true, },
 853         { .name     = "setitimer",
 854           .arg = { [0] = STRARRAY(which, itimers), }, },
 855         { .name     = "setrlimit",
 856           .arg = { [0] = STRARRAY(resource, rlimit_resources), }, },
 857         { .name     = "socket",
 858           .arg = { [0] = STRARRAY(family, socket_families),
 859                    [1] = { .scnprintf = SCA_SK_TYPE, /* type */ },
 860                    [2] = { .scnprintf = SCA_SK_PROTO, /* protocol */ }, }, },
 861         { .name     = "socketpair",
 862           .arg = { [0] = STRARRAY(family, socket_families),
 863                    [1] = { .scnprintf = SCA_SK_TYPE, /* type */ },
 864                    [2] = { .scnprintf = SCA_SK_PROTO, /* protocol */ }, }, },
 865         { .name     = "stat", .alias = "newstat", },
 866         { .name     = "statx",
 867           .arg = { [0] = { .scnprintf = SCA_FDAT,        /* fdat */ },
 868                    [2] = { .scnprintf = SCA_STATX_FLAGS, /* flags */ } ,
 869                    [3] = { .scnprintf = SCA_STATX_MASK,  /* mask */ }, }, },
 870         { .name     = "swapoff",
 871           .arg = { [0] = { .scnprintf = SCA_FILENAME, /* specialfile */ }, }, },
 872         { .name     = "swapon",
 873           .arg = { [0] = { .scnprintf = SCA_FILENAME, /* specialfile */ }, }, },
 874         { .name     = "symlinkat",
 875           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
 876         { .name     = "tgkill",
 877           .arg = { [2] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
 878         { .name     = "tkill",
 879           .arg = { [1] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
 880         { .name     = "umount2", .alias = "umount",
 881           .arg = { [0] = { .scnprintf = SCA_FILENAME, /* name */ }, }, },
 882         { .name     = "uname", .alias = "newuname", },
 883         { .name     = "unlinkat",
 884           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
 885         { .name     = "utimensat",
 886           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dirfd */ }, }, },
 887         { .name     = "wait4",      .errpid = true,
 888           .arg = { [2] = { .scnprintf = SCA_WAITID_OPTIONS, /* options */ }, }, },
 889         { .name     = "waitid",     .errpid = true,
 890           .arg = { [3] = { .scnprintf = SCA_WAITID_OPTIONS, /* options */ }, }, },
 891 };
 892
 893 static int syscall_fmt__cmp(const void *name, const void *fmtp)
 894 {
 895         const struct syscall_fmt *fmt = fmtp;
 896         return strcmp(name, fmt->name);
 897 }
 898
 899 static struct syscall_fmt *syscall_fmt__find(const char *name)
 900 {
 901         const int nmemb = ARRAY_SIZE(syscall_fmts);
 902         return bsearch(name, syscall_fmts, nmemb, sizeof(struct syscall_fmt), syscall_fmt__cmp);
 903 }
 904
 905 static struct syscall_fmt *syscall_fmt__find_by_alias(const char *alias)
 906 {
 907         int i, nmemb = ARRAY_SIZE(syscall_fmts);
 908
 909         for (i = 0; i < nmemb; ++i) {
 910                 if (syscall_fmts[i].alias && strcmp(syscall_fmts[i].alias, alias) == 0)
 911                         return &syscall_fmts[i];
 912         }
 913
 914         return NULL;
 915 }
 916
 917 /*
 918  * is_exit: is this "exit" or "exit_group"?
 919  * is_open: is this "open" or "openat"? To associate the fd returned in sys_exit with the pathname in sys_enter.
 920  * args_size: sum of the sizes of the syscall arguments, anything after that is augmented stuff: pathname for openat, etc.
 921  */
 922 struct syscall {
 923         struct tep_event    *tp_format;
 924         int                 nr_args;
 925         int                 args_size;
 926         bool                is_exit;
 927         bool                is_open;
 928         struct tep_format_field *args;
 929         const char          *name;
 930         struct syscall_fmt  *fmt;
 931         struct syscall_arg_fmt *arg_fmt;
 932 };
 933
 934 struct bpf_map_syscall_entry {
 935         bool    enabled;
 936 };
 937
 938 /*
 939  * We need to have this 'calculated' boolean because in some cases we really
 940  * don't know what is the duration of a syscall, for instance, when we start
 941  * a session and some threads are waiting for a syscall to finish, say 'poll',
 942  * in which case all we can do is to print "( ? ) for duration and for the
 943  * start timestamp.
 944  */
 945 static size_t fprintf_duration(unsigned long t, bool calculated, FILE *fp)
 946 {
 947         double duration = (double)t / NSEC_PER_MSEC;
 948         size_t printed = fprintf(fp, "(");
 949
 950         if (!calculated)
 951                 printed += fprintf(fp, "         ");
 952         else if (duration >= 1.0)
 953                 printed += color_fprintf(fp, PERF_COLOR_RED, "%6.3f ms", duration);
 954         else if (duration >= 0.01)
 955                 printed += color_fprintf(fp, PERF_COLOR_YELLOW, "%6.3f ms", duration);
 956         else
 957                 printed += color_fprintf(fp, PERF_COLOR_NORMAL, "%6.3f ms", duration);
 958         return printed + fprintf(fp, "): ");
 959 }
 960
 961 /**
 962  * filename.ptr: The filename char pointer that will be vfs_getname'd
 963  * filename.entry_str_pos: Where to insert the string translated from
 964  *                         filename.ptr by the vfs_getname tracepoint/kprobe.
 965  * ret_scnprintf: syscall args may set this to a different syscall return
 966  *                formatter, for instance, fcntl may return fds, file flags, etc.
 967  */
 968 struct thread_trace {
 969         u64               entry_time;
 970         bool              entry_pending;
 971         unsigned long     nr_events;
 972         unsigned long     pfmaj, pfmin;
 973         char              *entry_str;
 974         double            runtime_ms;
 975         size_t            (*ret_scnprintf)(char *bf, size_t size, struct syscall_arg *arg);
 976         struct {
 977                 unsigned long ptr;
 978                 short int     entry_str_pos;
 979                 bool          pending_open;
 980                 unsigned int  namelen;
 981                 char          *name;
 982         } filename;
 983         struct {
 984                 int           max;
 985                 struct file   *table;
 986         } files;
 987
 988         struct intlist *syscall_stats;
 989 };
 990
 991 static struct thread_trace *thread_trace__new(void)
 992 {
 993         struct thread_trace *ttrace =  zalloc(sizeof(struct thread_trace));
 994
 995         if (ttrace)
 996                 ttrace->files.max = -1;
 997
 998         ttrace->syscall_stats = intlist__new(NULL);
 999
1000         return ttrace;
1001 }
1002
1003 static struct thread_trace *thread__trace(struct thread *thread, FILE *fp)
1004 {
1005         struct thread_trace *ttrace;
1006
1007         if (thread == NULL)
1008                 goto fail;
1009
1010         if (thread__priv(thread) == NULL)
1011                 thread__set_priv(thread, thread_trace__new());
1012
1013         if (thread__priv(thread) == NULL)
1014                 goto fail;
1015
1016         ttrace = thread__priv(thread);
1017         ++ttrace->nr_events;
1018
1019         return ttrace;
1020 fail:
1021         color_fprintf(fp, PERF_COLOR_RED,
1022                       "WARNING: not enough memory, dropping samples!\n");
1023         return NULL;
1024 }
1025
1026
1027 void syscall_arg__set_ret_scnprintf(struct syscall_arg *arg,
1028                                     size_t (*ret_scnprintf)(char *bf, size_t size, struct syscall_arg *arg))
1029 {
1030         struct thread_trace *ttrace = thread__priv(arg->thread);
1031
1032         ttrace->ret_scnprintf = ret_scnprintf;
1033 }
1034
1035 #define TRACE_PFMAJ             (1 << 0)
1036 #define TRACE_PFMIN             (1 << 1)
1037
1038 static const size_t trace__entry_str_size = 2048;
1039
1040 static struct file *thread_trace__files_entry(struct thread_trace *ttrace, int fd)
1041 {
1042         if (fd > ttrace->files.max) {
1043                 struct file *nfiles = realloc(ttrace->files.table, (fd + 1) * sizeof(struct file));
1044
1045                 if (nfiles == NULL)
1046                         return NULL;
1047
1048                 if (ttrace->files.max != -1) {
1049                         memset(nfiles + ttrace->files.max + 1, 0,
1050                                (fd - ttrace->files.max) * sizeof(struct file));
1051                 } else {
1052                         memset(nfiles, 0, (fd + 1) * sizeof(struct file));
1053                 }
1054
1055                 ttrace->files.table = nfiles;
1056                 ttrace->files.max   = fd;
1057         }
1058
1059         return ttrace->files.table + fd;
1060 }
1061
1062 struct file *thread__files_entry(struct thread *thread, int fd)
1063 {
1064         return thread_trace__files_entry(thread__priv(thread), fd);
1065 }
1066
1067 static int trace__set_fd_pathname(struct thread *thread, int fd, const char *pathname)
1068 {
1069         struct thread_trace *ttrace = thread__priv(thread);
1070         struct file *file = thread_trace__files_entry(ttrace, fd);
1071
1072         if (file != NULL) {
1073                 struct stat st;
1074                 if (stat(pathname, &st) == 0)
1075                         file->dev_maj = major(st.st_rdev);
1076                 file->pathname = strdup(pathname);
1077                 if (file->pathname)
1078                         return 0;
1079         }
1080
1081         return -1;
1082 }
1083
1084 static int thread__read_fd_path(struct thread *thread, int fd)
1085 {
1086         char linkname[PATH_MAX], pathname[PATH_MAX];
1087         struct stat st;
1088         int ret;
1089
1090         if (thread->pid_ == thread->tid) {
1091                 scnprintf(linkname, sizeof(linkname),
1092                           "/proc/%d/fd/%d", thread->pid_, fd);
1093         } else {
1094                 scnprintf(linkname, sizeof(linkname),
1095                           "/proc/%d/task/%d/fd/%d", thread->pid_, thread->tid, fd);
1096         }
1097
1098         if (lstat(linkname, &st) < 0 || st.st_size + 1 > (off_t)sizeof(pathname))
1099                 return -1;
1100
1101         ret = readlink(linkname, pathname, sizeof(pathname));
1102
1103         if (ret < 0 || ret > st.st_size)
1104                 return -1;
1105
1106         pathname[ret] = '\0';
1107         return trace__set_fd_pathname(thread, fd, pathname);
1108 }
1109
1110 static const char *thread__fd_path(struct thread *thread, int fd,
1111                                    struct trace *trace)
1112 {
1113         struct thread_trace *ttrace = thread__priv(thread);
1114
1115         if (ttrace == NULL)
1116                 return NULL;
1117
1118         if (fd < 0)
1119                 return NULL;
1120
1121         if ((fd > ttrace->files.max || ttrace->files.table[fd].pathname == NULL)) {
1122                 if (!trace->live)
1123                         return NULL;
1124                 ++trace->stats.proc_getname;
1125                 if (thread__read_fd_path(thread, fd))
1126                         return NULL;
1127         }
1128
1129         return ttrace->files.table[fd].pathname;
1130 }
1131
1132 size_t syscall_arg__scnprintf_fd(char *bf, size_t size, struct syscall_arg *arg)
1133 {
1134         int fd = arg->val;
1135         size_t printed = scnprintf(bf, size, "%d", fd);
1136         const char *path = thread__fd_path(arg->thread, fd, arg->trace);
1137
1138         if (path)
1139                 printed += scnprintf(bf + printed, size - printed, "<%s>", path);
1140
1141         return printed;
1142 }
1143
1144 size_t pid__scnprintf_fd(struct trace *trace, pid_t pid, int fd, char *bf, size_t size)
1145 {
1146         size_t printed = scnprintf(bf, size, "%d", fd);
1147         struct thread *thread = machine__find_thread(trace->host, pid, pid);
1148
1149         if (thread) {
1150                 const char *path = thread__fd_path(thread, fd, trace);
1151
1152                 if (path)
1153                         printed += scnprintf(bf + printed, size - printed, "<%s>", path);
1154
1155                 thread__put(thread);
1156         }
1157
1158         return printed;
1159 }
1160
1161 static size_t syscall_arg__scnprintf_close_fd(char *bf, size_t size,
1162                                               struct syscall_arg *arg)
1163 {
1164         int fd = arg->val;
1165         size_t printed = syscall_arg__scnprintf_fd(bf, size, arg);
1166         struct thread_trace *ttrace = thread__priv(arg->thread);
1167
1168         if (ttrace && fd >= 0 && fd <= ttrace->files.max)
1169                 zfree(&ttrace->files.table[fd].pathname);
1170
1171         return printed;
1172 }
1173
1174 static void thread__set_filename_pos(struct thread *thread, const char *bf,
1175                                      unsigned long ptr)
1176 {
1177         struct thread_trace *ttrace = thread__priv(thread);
1178
1179         ttrace->filename.ptr = ptr;
1180         ttrace->filename.entry_str_pos = bf - ttrace->entry_str;
1181 }
1182
1183 static size_t syscall_arg__scnprintf_augmented_string(struct syscall_arg *arg, char *bf, size_t size)
1184 {
1185         struct augmented_arg *augmented_arg = arg->augmented.args;
1186
1187         return scnprintf(bf, size, "\"%.*s\"", augmented_arg->size, augmented_arg->value);
1188 }
1189
1190 static size_t syscall_arg__scnprintf_filename(char *bf, size_t size,
1191                                               struct syscall_arg *arg)
1192 {
1193         unsigned long ptr = arg->val;
1194
1195         if (arg->augmented.args)
1196                 return syscall_arg__scnprintf_augmented_string(arg, bf, size);
1197
1198         if (!arg->trace->vfs_getname)
1199                 return scnprintf(bf, size, "%#x", ptr);
1200
1201         thread__set_filename_pos(arg->thread, bf, ptr);
1202         return 0;
1203 }
1204
1205 static bool trace__filter_duration(struct trace *trace, double t)
1206 {
1207         return t < (trace->duration_filter * NSEC_PER_MSEC);
1208 }
1209
1210 static size_t __trace__fprintf_tstamp(struct trace *trace, u64 tstamp, FILE *fp)
1211 {
1212         double ts = (double)(tstamp - trace->base_time) / NSEC_PER_MSEC;
1213
1214         return fprintf(fp, "%10.3f ", ts);
1215 }
1216
1217 /*
1218  * We're handling tstamp=0 as an undefined tstamp, i.e. like when we are
1219  * using ttrace->entry_time for a thread that receives a sys_exit without
1220  * first having received a sys_enter ("poll" issued before tracing session
1221  * starts, lost sys_enter exit due to ring buffer overflow).
1222  */
1223 static size_t trace__fprintf_tstamp(struct trace *trace, u64 tstamp, FILE *fp)
1224 {
1225         if (tstamp > 0)
1226                 return __trace__fprintf_tstamp(trace, tstamp, fp);
1227
1228         return fprintf(fp, "         ? ");
1229 }
1230
1231 static bool done = false;
1232 static bool interrupted = false;
1233
1234 static void sig_handler(int sig)
1235 {
1236         done = true;
1237         interrupted = sig == SIGINT;
1238 }
1239
1240 static size_t trace__fprintf_comm_tid(struct trace *trace, struct thread *thread, FILE *fp)
1241 {
1242         size_t printed = 0;
1243
1244         if (trace->multiple_threads) {
1245                 if (trace->show_comm)
1246                         printed += fprintf(fp, "%.14s/", thread__comm_str(thread));
1247                 printed += fprintf(fp, "%d ", thread->tid);
1248         }
1249
1250         return printed;
1251 }
1252
1253 static size_t trace__fprintf_entry_head(struct trace *trace, struct thread *thread,
1254                                         u64 duration, bool duration_calculated, u64 tstamp, FILE *fp)
1255 {
1256         size_t printed = 0;
1257
1258         if (trace->show_tstamp)
1259                 printed = trace__fprintf_tstamp(trace, tstamp, fp);
1260         if (trace->show_duration)
1261                 printed += fprintf_duration(duration, duration_calculated, fp);
1262         return printed + trace__fprintf_comm_tid(trace, thread, fp);
1263 }
1264
1265 static int trace__process_event(struct trace *trace, struct machine *machine,
1266                                 union perf_event *event, struct perf_sample *sample)
1267 {
1268         int ret = 0;
1269
1270         switch (event->header.type) {
1271         case PERF_RECORD_LOST:
1272                 color_fprintf(trace->output, PERF_COLOR_RED,
1273                               "LOST %" PRIu64 " events!\n", event->lost.lost);
1274                 ret = machine__process_lost_event(machine, event, sample);
1275                 break;
1276         default:
1277                 ret = machine__process_event(machine, event, sample);
1278                 break;
1279         }
1280
1281         return ret;
1282 }
1283
1284 static int trace__tool_process(struct perf_tool *tool,
1285                                union perf_event *event,
1286                                struct perf_sample *sample,
1287                                struct machine *machine)
1288 {
1289         struct trace *trace = container_of(tool, struct trace, tool);
1290         return trace__process_event(trace, machine, event, sample);
1291 }
1292
1293 static char *trace__machine__resolve_kernel_addr(void *vmachine, unsigned long long *addrp, char **modp)
1294 {
1295         struct machine *machine = vmachine;
1296
1297         if (machine->kptr_restrict_warned)
1298                 return NULL;
1299
1300         if (symbol_conf.kptr_restrict) {
1301                 pr_warning("Kernel address maps (/proc/{kallsyms,modules}) are restricted.\n\n"
1302                            "Check /proc/sys/kernel/kptr_restrict.\n\n"
1303                            "Kernel samples will not be resolved.\n");
1304                 machine->kptr_restrict_warned = true;
1305                 return NULL;
1306         }
1307
1308         return machine__resolve_kernel_addr(vmachine, addrp, modp);
1309 }
1310
1311 static int trace__symbols_init(struct trace *trace, struct perf_evlist *evlist)
1312 {
1313         int err = symbol__init(NULL);
1314
1315         if (err)
1316                 return err;
1317
1318         trace->host = machine__new_host();
1319         if (trace->host == NULL)
1320                 return -ENOMEM;
1321
1322         err = trace_event__register_resolver(trace->host, trace__machine__resolve_kernel_addr);
1323         if (err < 0)
1324                 goto out;
1325
1326         err = __machine__synthesize_threads(trace->host, &trace->tool, &trace->opts.target,
1327                                             evlist->threads, trace__tool_process, false,
1328                                             1);
1329 out:
1330         if (err)
1331                 symbol__exit();
1332
1333         return err;
1334 }
1335
1336 static void trace__symbols__exit(struct trace *trace)
1337 {
1338         machine__exit(trace->host);
1339         trace->host = NULL;
1340
1341         symbol__exit();
1342 }
1343
1344 static int syscall__alloc_arg_fmts(struct syscall *sc, int nr_args)
1345 {
1346         int idx;
1347
1348         if (nr_args == 6 && sc->fmt && sc->fmt->nr_args != 0)
1349                 nr_args = sc->fmt->nr_args;
1350
1351         sc->arg_fmt = calloc(nr_args, sizeof(*sc->arg_fmt));
1352         if (sc->arg_fmt == NULL)
1353                 return -1;
1354
1355         for (idx = 0; idx < nr_args; ++idx) {
1356                 if (sc->fmt)
1357                         sc->arg_fmt[idx] = sc->fmt->arg[idx];
1358         }
1359
1360         sc->nr_args = nr_args;
1361         return 0;
1362 }
1363
1364 static int syscall__set_arg_fmts(struct syscall *sc)
1365 {
1366         struct tep_format_field *field, *last_field = NULL;
1367         int idx = 0, len;
1368
1369         for (field = sc->args; field; field = field->next, ++idx) {
1370                 last_field = field;
1371
1372                 if (sc->fmt && sc->fmt->arg[idx].scnprintf)
1373                         continue;
1374
1375                 if (strcmp(field->type, "const char *") == 0 &&
1376                          (strcmp(field->name, "filename") == 0 ||
1377                           strcmp(field->name, "path") == 0 ||
1378                           strcmp(field->name, "pathname") == 0))
1379                         sc->arg_fmt[idx].scnprintf = SCA_FILENAME;
1380                 else if ((field->flags & TEP_FIELD_IS_POINTER) || strstr(field->name, "addr"))
1381                         sc->arg_fmt[idx].scnprintf = SCA_PTR;
1382                 else if (strcmp(field->type, "pid_t") == 0)
1383                         sc->arg_fmt[idx].scnprintf = SCA_PID;
1384                 else if (strcmp(field->type, "umode_t") == 0)
1385                         sc->arg_fmt[idx].scnprintf = SCA_MODE_T;
1386                 else if ((strcmp(field->type, "int") == 0 ||
1387                           strcmp(field->type, "unsigned int") == 0 ||
1388                           strcmp(field->type, "long") == 0) &&
1389                          (len = strlen(field->name)) >= 2 &&
1390                          strcmp(field->name + len - 2, "fd") == 0) {
1391                         /*
1392                          * /sys/kernel/tracing/events/syscalls/sys_enter*
1393                          * egrep 'field:.*fd;' .../format|sed -r 's/.*field:([a-z ]+) [a-z_]*fd.+/\1/g'|sort|uniq -c
1394                          * 65 int
1395                          * 23 unsigned int
1396                          * 7 unsigned long
1397                          */
1398                         sc->arg_fmt[idx].scnprintf = SCA_FD;
1399                 }
1400         }
1401
1402         if (last_field)
1403                 sc->args_size = last_field->offset + last_field->size;
1404
1405         return 0;
1406 }
1407
1408 static int trace__read_syscall_info(struct trace *trace, int id)
1409 {
1410         char tp_name[128];
1411         struct syscall *sc;
1412         const char *name = syscalltbl__name(trace->sctbl, id);
1413
1414         if (name == NULL)
1415                 return -1;
1416
1417         if (id > trace->syscalls.max) {
1418                 struct syscall *nsyscalls = realloc(trace->syscalls.table, (id + 1) * sizeof(*sc));
1419
1420                 if (nsyscalls == NULL)
1421                         return -1;
1422
1423                 if (trace->syscalls.max != -1) {
1424                         memset(nsyscalls + trace->syscalls.max + 1, 0,
1425                                (id - trace->syscalls.max) * sizeof(*sc));
1426                 } else {
1427                         memset(nsyscalls, 0, (id + 1) * sizeof(*sc));
1428                 }
1429
1430                 trace->syscalls.table = nsyscalls;
1431                 trace->syscalls.max   = id;
1432         }
1433
1434         sc = trace->syscalls.table + id;
1435         sc->name = name;
1436
1437         sc->fmt  = syscall_fmt__find(sc->name);
1438
1439         snprintf(tp_name, sizeof(tp_name), "sys_enter_%s", sc->name);
1440         sc->tp_format = trace_event__tp_format("syscalls", tp_name);
1441
1442         if (IS_ERR(sc->tp_format) && sc->fmt && sc->fmt->alias) {
1443                 snprintf(tp_name, sizeof(tp_name), "sys_enter_%s", sc->fmt->alias);
1444                 sc->tp_format = trace_event__tp_format("syscalls", tp_name);
1445         }
1446
1447         if (syscall__alloc_arg_fmts(sc, IS_ERR(sc->tp_format) ? 6 : sc->tp_format->format.nr_fields))
1448                 return -1;
1449
1450         if (IS_ERR(sc->tp_format))
1451                 return -1;
1452
1453         sc->args = sc->tp_format->format.fields;
1454         /*
1455          * We need to check and discard the first variable '__syscall_nr'
1456          * or 'nr' that mean the syscall number. It is needless here.
1457          * So drop '__syscall_nr' or 'nr' field but does not exist on older kernels.
1458          */
1459         if (sc->args && (!strcmp(sc->args->name, "__syscall_nr") || !strcmp(sc->args->name, "nr"))) {
1460                 sc->args = sc->args->next;
1461                 --sc->nr_args;
1462         }
1463
1464         sc->is_exit = !strcmp(name, "exit_group") || !strcmp(name, "exit");
1465         sc->is_open = !strcmp(name, "open") || !strcmp(name, "openat");
1466
1467         return syscall__set_arg_fmts(sc);
1468 }
1469
1470 static int trace__validate_ev_qualifier(struct trace *trace)
1471 {
1472         int err = 0, i;
1473         size_t nr_allocated;
1474         struct str_node *pos;
1475
1476         trace->ev_qualifier_ids.nr = strlist__nr_entries(trace->ev_qualifier);
1477         trace->ev_qualifier_ids.entries = malloc(trace->ev_qualifier_ids.nr *
1478                                                  sizeof(trace->ev_qualifier_ids.entries[0]));
1479
1480         if (trace->ev_qualifier_ids.entries == NULL) {
1481                 fputs("Error:\tNot enough memory for allocating events qualifier ids\n",
1482                        trace->output);
1483                 err = -EINVAL;
1484                 goto out;
1485         }
1486
1487         nr_allocated = trace->ev_qualifier_ids.nr;
1488         i = 0;
1489
1490         strlist__for_each_entry(pos, trace->ev_qualifier) {
1491                 const char *sc = pos->s;
1492                 int id = syscalltbl__id(trace->sctbl, sc), match_next = -1;
1493
1494                 if (id < 0) {
1495                         id = syscalltbl__strglobmatch_first(trace->sctbl, sc, &match_next);
1496                         if (id >= 0)
1497                                 goto matches;
1498
1499                         if (err == 0) {
1500                                 fputs("Error:\tInvalid syscall ", trace->output);
1501                                 err = -EINVAL;
1502                         } else {
1503                                 fputs(", ", trace->output);
1504                         }
1505
1506                         fputs(sc, trace->output);
1507                 }
1508 matches:
1509                 trace->ev_qualifier_ids.entries[i++] = id;
1510                 if (match_next == -1)
1511                         continue;
1512
1513                 while (1) {
1514                         id = syscalltbl__strglobmatch_next(trace->sctbl, sc, &match_next);
1515                         if (id < 0)
1516                                 break;
1517                         if (nr_allocated == trace->ev_qualifier_ids.nr) {
1518                                 void *entries;
1519
1520                                 nr_allocated += 8;
1521                                 entries = realloc(trace->ev_qualifier_ids.entries,
1522                                                   nr_allocated * sizeof(trace->ev_qualifier_ids.entries[0]));
1523                                 if (entries == NULL) {
1524                                         err = -ENOMEM;
1525                                         fputs("\nError:\t Not enough memory for parsing\n", trace->output);
1526                                         goto out_free;
1527                                 }
1528                                 trace->ev_qualifier_ids.entries = entries;
1529                         }
1530                         trace->ev_qualifier_ids.nr++;
1531                         trace->ev_qualifier_ids.entries[i++] = id;
1532                 }
1533         }
1534
1535         if (err < 0) {
1536                 fputs("\nHint:\ttry 'perf list syscalls:sys_enter_*'"
1537                       "\nHint:\tand: 'man syscalls'\n", trace->output);
1538 out_free:
1539                 zfree(&trace->ev_qualifier_ids.entries);
1540                 trace->ev_qualifier_ids.nr = 0;
1541         }
1542 out:
1543         return err;
1544 }
1545
1546 /*
1547  * args is to be interpreted as a series of longs but we need to handle
1548  * 8-byte unaligned accesses. args points to raw_data within the event
1549  * and raw_data is guaranteed to be 8-byte unaligned because it is
1550  * preceded by raw_size which is a u32. So we need to copy args to a temp
1551  * variable to read it. Most notably this avoids extended load instructions
1552  * on unaligned addresses
1553  */
1554 unsigned long syscall_arg__val(struct syscall_arg *arg, u8 idx)
1555 {
1556         unsigned long val;
1557         unsigned char *p = arg->args + sizeof(unsigned long) * idx;
1558
1559         memcpy(&val, p, sizeof(val));
1560         return val;
1561 }
1562
1563 static size_t syscall__scnprintf_name(struct syscall *sc, char *bf, size_t size,
1564                                       struct syscall_arg *arg)
1565 {
1566         if (sc->arg_fmt && sc->arg_fmt[arg->idx].name)
1567                 return scnprintf(bf, size, "%s: ", sc->arg_fmt[arg->idx].name);
1568
1569         return scnprintf(bf, size, "arg%d: ", arg->idx);
1570 }
1571
1572 /*
1573  * Check if the value is in fact zero, i.e. mask whatever needs masking, such
1574  * as mount 'flags' argument that needs ignoring some magic flag, see comment
1575  * in tools/perf/trace/beauty/mount_flags.c
1576  */
1577 static unsigned long syscall__mask_val(struct syscall *sc, struct syscall_arg *arg, unsigned long val)
1578 {
1579         if (sc->arg_fmt && sc->arg_fmt[arg->idx].mask_val)
1580                 return sc->arg_fmt[arg->idx].mask_val(arg, val);
1581
1582         return val;
1583 }
1584
1585 static size_t syscall__scnprintf_val(struct syscall *sc, char *bf, size_t size,
1586                                      struct syscall_arg *arg, unsigned long val)
1587 {
1588         if (sc->arg_fmt && sc->arg_fmt[arg->idx].scnprintf) {
1589                 arg->val = val;
1590                 if (sc->arg_fmt[arg->idx].parm)
1591                         arg->parm = sc->arg_fmt[arg->idx].parm;
1592                 return sc->arg_fmt[arg->idx].scnprintf(bf, size, arg);
1593         }
1594         return scnprintf(bf, size, "%ld", val);
1595 }
1596
1597 static size_t syscall__scnprintf_args(struct syscall *sc, char *bf, size_t size,
1598                                       unsigned char *args, void *augmented_args, int augmented_args_size,
1599                                       struct trace *trace, struct thread *thread)
1600 {
1601         size_t printed = 0;
1602         unsigned long val;
1603         u8 bit = 1;
1604         struct syscall_arg arg = {
1605                 .args   = args,
1606                 .augmented = {
1607                         .size = augmented_args_size,
1608                         .args = augmented_args,
1609                 },
1610                 .idx    = 0,
1611                 .mask   = 0,
1612                 .trace  = trace,
1613                 .thread = thread,
1614                 .show_string_prefix = trace->show_string_prefix,
1615         };
1616         struct thread_trace *ttrace = thread__priv(thread);
1617
1618         /*
1619          * Things like fcntl will set this in its 'cmd' formatter to pick the
1620          * right formatter for the return value (an fd? file flags?), which is
1621          * not needed for syscalls that always return a given type, say an fd.
1622          */
1623         ttrace->ret_scnprintf = NULL;
1624
1625         if (sc->args != NULL) {
1626                 struct tep_format_field *field;
1627
1628                 for (field = sc->args; field;
1629                      field = field->next, ++arg.idx, bit <<= 1) {
1630                         if (arg.mask & bit)
1631                                 continue;
1632
1633                         val = syscall_arg__val(&arg, arg.idx);
1634                         /*
1635                          * Some syscall args need some mask, most don't and
1636                          * return val untouched.
1637                          */
1638                         val = syscall__mask_val(sc, &arg, val);
1639
1640                         /*
1641                          * Suppress this argument if its value is zero and
1642                          * and we don't have a string associated in an
1643                          * strarray for it.
1644                          */
1645                         if (val == 0 &&
1646                             !trace->show_zeros &&
1647                             !(sc->arg_fmt &&
1648                               (sc->arg_fmt[arg.idx].show_zero ||
1649                                sc->arg_fmt[arg.idx].scnprintf == SCA_STRARRAY ||
1650                                sc->arg_fmt[arg.idx].scnprintf == SCA_STRARRAYS) &&
1651                               sc->arg_fmt[arg.idx].parm))
1652                                 continue;
1653
1654                         printed += scnprintf(bf + printed, size - printed, "%s", printed ? ", " : "");
1655
1656                         if (trace->show_arg_names)
1657                                 printed += scnprintf(bf + printed, size - printed, "%s: ", field->name);
1658
1659                         printed += syscall__scnprintf_val(sc, bf + printed, size - printed, &arg, val);
1660                 }
1661         } else if (IS_ERR(sc->tp_format)) {
1662                 /*
1663                  * If we managed to read the tracepoint /format file, then we
1664                  * may end up not having any args, like with gettid(), so only
1665                  * print the raw args when we didn't manage to read it.
1666                  */
1667                 while (arg.idx < sc->nr_args) {
1668                         if (arg.mask & bit)
1669                                 goto next_arg;
1670                         val = syscall_arg__val(&arg, arg.idx);
1671                         if (printed)
1672                                 printed += scnprintf(bf + printed, size - printed, ", ");
1673                         printed += syscall__scnprintf_name(sc, bf + printed, size - printed, &arg);
1674                         printed += syscall__scnprintf_val(sc, bf + printed, size - printed, &arg, val);
1675 next_arg:
1676                         ++arg.idx;
1677                         bit <<= 1;
1678                 }
1679         }
1680
1681         return printed;
1682 }
1683
1684 typedef int (*tracepoint_handler)(struct trace *trace, struct perf_evsel *evsel,
1685                                   union perf_event *event,
1686                                   struct perf_sample *sample);
1687
1688 static struct syscall *trace__syscall_info(struct trace *trace,
1689                                            struct perf_evsel *evsel, int id)
1690 {
1691
1692         if (id < 0) {
1693
1694                 /*
1695                  * XXX: Noticed on x86_64, reproduced as far back as 3.0.36, haven't tried
1696                  * before that, leaving at a higher verbosity level till that is
1697                  * explained. Reproduced with plain ftrace with:
1698                  *
1699                  * echo 1 > /t/events/raw_syscalls/sys_exit/enable
1700                  * grep "NR -1 " /t/trace_pipe
1701                  *
1702                  * After generating some load on the machine.
1703                  */
1704                 if (verbose > 1) {
1705                         static u64 n;
1706                         fprintf(trace->output, "Invalid syscall %d id, skipping (%s, %" PRIu64 ") ...\n",
1707                                 id, perf_evsel__name(evsel), ++n);
1708                 }
1709                 return NULL;
1710         }
1711
1712         if ((id > trace->syscalls.max || trace->syscalls.table[id].name == NULL) &&
1713             trace__read_syscall_info(trace, id))
1714                 goto out_cant_read;
1715
1716         if ((id > trace->syscalls.max || trace->syscalls.table[id].name == NULL))
1717                 goto out_cant_read;
1718
1719         return &trace->syscalls.table[id];
1720
1721 out_cant_read:
1722         if (verbose > 0) {
1723                 fprintf(trace->output, "Problems reading syscall %d", id);
1724                 if (id <= trace->syscalls.max && trace->syscalls.table[id].name != NULL)
1725                         fprintf(trace->output, "(%s)", trace->syscalls.table[id].name);
1726                 fputs(" information\n", trace->output);
1727         }
1728         return NULL;
1729 }
1730
1731 static void thread__update_stats(struct thread_trace *ttrace,
1732                                  int id, struct perf_sample *sample)
1733 {
1734         struct int_node *inode;
1735         struct stats *stats;
1736         u64 duration = 0;
1737
1738         inode = intlist__findnew(ttrace->syscall_stats, id);
1739         if (inode == NULL)
1740                 return;
1741
1742         stats = inode->priv;
1743         if (stats == NULL) {
1744                 stats = malloc(sizeof(struct stats));
1745                 if (stats == NULL)
1746                         return;
1747                 init_stats(stats);
1748                 inode->priv = stats;
1749         }
1750
1751         if (ttrace->entry_time && sample->time > ttrace->entry_time)
1752                 duration = sample->time - ttrace->entry_time;
1753
1754         update_stats(stats, duration);
1755 }
1756
1757 static int trace__printf_interrupted_entry(struct trace *trace)
1758 {
1759         struct thread_trace *ttrace;
1760         size_t printed;
1761         int len;
1762
1763         if (trace->failure_only || trace->current == NULL)
1764                 return 0;
1765
1766         ttrace = thread__priv(trace->current);
1767
1768         if (!ttrace->entry_pending)
1769                 return 0;
1770
1771         printed  = trace__fprintf_entry_head(trace, trace->current, 0, false, ttrace->entry_time, trace->output);
1772         printed += len = fprintf(trace->output, "%s)", ttrace->entry_str);
1773
1774         if (len < trace->args_alignment - 4)
1775                 printed += fprintf(trace->output, "%-*s", trace->args_alignment - 4 - len, " ");
1776
1777         printed += fprintf(trace->output, " ...\n");
1778
1779         ttrace->entry_pending = false;
1780         ++trace->nr_events_printed;
1781
1782         return printed;
1783 }
1784
1785 static int trace__fprintf_sample(struct trace *trace, struct perf_evsel *evsel,
1786                                  struct perf_sample *sample, struct thread *thread)
1787 {
1788         int printed = 0;
1789
1790         if (trace->print_sample) {
1791                 double ts = (double)sample->time / NSEC_PER_MSEC;
1792
1793                 printed += fprintf(trace->output, "%22s %10.3f %s %d/%d [%d]\n",
1794                                    perf_evsel__name(evsel), ts,
1795                                    thread__comm_str(thread),
1796                                    sample->pid, sample->tid, sample->cpu);
1797         }
1798
1799         return printed;
1800 }
1801
1802 static void *syscall__augmented_args(struct syscall *sc, struct perf_sample *sample, int *augmented_args_size, int raw_augmented_args_size)
1803 {
1804         void *augmented_args = NULL;
1805         /*
1806          * For now with BPF raw_augmented we hook into raw_syscalls:sys_enter
1807          * and there we get all 6 syscall args plus the tracepoint common fields
1808          * that gets calculated at the start and the syscall_nr (another long).
1809          * So we check if that is the case and if so don't look after the
1810          * sc->args_size but always after the full raw_syscalls:sys_enter payload,
1811          * which is fixed.
1812          *
1813          * We'll revisit this later to pass s->args_size to the BPF augmenter
1814          * (now tools/perf/examples/bpf/augmented_raw_syscalls.c, so that it
1815          * copies only what we need for each syscall, like what happens when we
1816          * use syscalls:sys_enter_NAME, so that we reduce the kernel/userspace
1817          * traffic to just what is needed for each syscall.
1818          */
1819         int args_size = raw_augmented_args_size ?: sc->args_size;
1820
1821         *augmented_args_size = sample->raw_size - args_size;
1822         if (*augmented_args_size > 0)
1823                 augmented_args = sample->raw_data + args_size;
1824
1825         return augmented_args;
1826 }
1827
1828 static int trace__sys_enter(struct trace *trace, struct perf_evsel *evsel,
1829                             union perf_event *event __maybe_unused,
1830                             struct perf_sample *sample)
1831 {
1832         char *msg;
1833         void *args;
1834         int printed = 0;
1835         struct thread *thread;
1836         int id = perf_evsel__sc_tp_uint(evsel, id, sample), err = -1;
1837         int augmented_args_size = 0;
1838         void *augmented_args = NULL;
1839         struct syscall *sc = trace__syscall_info(trace, evsel, id);
1840         struct thread_trace *ttrace;
1841
1842         if (sc == NULL)
1843                 return -1;
1844
1845         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
1846         ttrace = thread__trace(thread, trace->output);
1847         if (ttrace == NULL)
1848                 goto out_put;
1849
1850         trace__fprintf_sample(trace, evsel, sample, thread);
1851
1852         args = perf_evsel__sc_tp_ptr(evsel, args, sample);
1853
1854         if (ttrace->entry_str == NULL) {
1855                 ttrace->entry_str = malloc(trace__entry_str_size);
1856                 if (!ttrace->entry_str)
1857                         goto out_put;
1858         }
1859
1860         if (!(trace->duration_filter || trace->summary_only || trace->min_stack))
1861                 trace__printf_interrupted_entry(trace);
1862         /*
1863          * If this is raw_syscalls.sys_enter, then it always comes with the 6 possible
1864          * arguments, even if the syscall being handled, say "openat", uses only 4 arguments
1865          * this breaks syscall__augmented_args() check for augmented args, as we calculate
1866          * syscall->args_size using each syscalls:sys_enter_NAME tracefs format file,
1867          * so when handling, say the openat syscall, we end up getting 6 args for the
1868          * raw_syscalls:sys_enter event, when we expected just 4, we end up mistakenly
1869          * thinking that the extra 2 u64 args are the augmented filename, so just check
1870          * here and avoid using augmented syscalls when the evsel is the raw_syscalls one.
1871          */
1872         if (evsel != trace->syscalls.events.sys_enter)
1873                 augmented_args = syscall__augmented_args(sc, sample, &augmented_args_size, trace->raw_augmented_syscalls_args_size);
1874         ttrace->entry_time = sample->time;
1875         msg = ttrace->entry_str;
1876         printed += scnprintf(msg + printed, trace__entry_str_size - printed, "%s(", sc->name);
1877
1878         printed += syscall__scnprintf_args(sc, msg + printed, trace__entry_str_size - printed,
1879                                            args, augmented_args, augmented_args_size, trace, thread);
1880
1881         if (sc->is_exit) {
1882                 if (!(trace->duration_filter || trace->summary_only || trace->failure_only || trace->min_stack)) {
1883                         int alignment = 0;
1884
1885                         trace__fprintf_entry_head(trace, thread, 0, false, ttrace->entry_time, trace->output);
1886                         printed = fprintf(trace->output, "%s)", ttrace->entry_str);
1887                         if (trace->args_alignment > printed)
1888                                 alignment = trace->args_alignment - printed;
1889                         fprintf(trace->output, "%*s= ?\n", alignment, " ");
1890                 }
1891         } else {
1892                 ttrace->entry_pending = true;
1893                 /* See trace__vfs_getname & trace__sys_exit */
1894                 ttrace->filename.pending_open = false;
1895         }
1896
1897         if (trace->current != thread) {
1898                 thread__put(trace->current);
1899                 trace->current = thread__get(thread);
1900         }
1901         err = 0;
1902 out_put:
1903         thread__put(thread);
1904         return err;
1905 }
1906
1907 static int trace__fprintf_sys_enter(struct trace *trace, struct perf_evsel *evsel,
1908                                     struct perf_sample *sample)
1909 {
1910         struct thread_trace *ttrace;
1911         struct thread *thread;
1912         int id = perf_evsel__sc_tp_uint(evsel, id, sample), err = -1;
1913         struct syscall *sc = trace__syscall_info(trace, evsel, id);
1914         char msg[1024];
1915         void *args, *augmented_args = NULL;
1916         int augmented_args_size;
1917
1918         if (sc == NULL)
1919                 return -1;
1920
1921         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
1922         ttrace = thread__trace(thread, trace->output);
1923         /*
1924          * We need to get ttrace just to make sure it is there when syscall__scnprintf_args()
1925          * and the rest of the beautifiers accessing it via struct syscall_arg touches it.
1926          */
1927         if (ttrace == NULL)
1928                 goto out_put;
1929
1930         args = perf_evsel__sc_tp_ptr(evsel, args, sample);
1931         augmented_args = syscall__augmented_args(sc, sample, &augmented_args_size, trace->raw_augmented_syscalls_args_size);
1932         syscall__scnprintf_args(sc, msg, sizeof(msg), args, augmented_args, augmented_args_size, trace, thread);
1933         fprintf(trace->output, "%s", msg);
1934         err = 0;
1935 out_put:
1936         thread__put(thread);
1937         return err;
1938 }
1939
1940 static int trace__resolve_callchain(struct trace *trace, struct perf_evsel *evsel,
1941                                     struct perf_sample *sample,
1942                                     struct callchain_cursor *cursor)
1943 {
1944         struct addr_location al;
1945         int max_stack = evsel->attr.sample_max_stack ?
1946                         evsel->attr.sample_max_stack :
1947                         trace->max_stack;
1948         int err;
1949
1950         if (machine__resolve(trace->host, &al, sample) < 0)
1951                 return -1;
1952
1953         err = thread__resolve_callchain(al.thread, cursor, evsel, sample, NULL, NULL, max_stack);
1954         addr_location__put(&al);
1955         return err;
1956 }
1957
1958 static int trace__fprintf_callchain(struct trace *trace, struct perf_sample *sample)
1959 {
1960         /* TODO: user-configurable print_opts */
1961         const unsigned int print_opts = EVSEL__PRINT_SYM |
1962                                         EVSEL__PRINT_DSO |
1963                                         EVSEL__PRINT_UNKNOWN_AS_ADDR;
1964
1965         return sample__fprintf_callchain(sample, 38, print_opts, &callchain_cursor, trace->output);
1966 }
1967
1968 static const char *errno_to_name(struct perf_evsel *evsel, int err)
1969 {
1970         struct perf_env *env = perf_evsel__env(evsel);
1971         const char *arch_name = perf_env__arch(env);
1972
1973         return arch_syscalls__strerrno(arch_name, err);
1974 }
1975
1976 static int trace__sys_exit(struct trace *trace, struct perf_evsel *evsel,
1977                            union perf_event *event __maybe_unused,
1978                            struct perf_sample *sample)
1979 {
1980         long ret;
1981         u64 duration = 0;
1982         bool duration_calculated = false;
1983         struct thread *thread;
1984         int id = perf_evsel__sc_tp_uint(evsel, id, sample), err = -1, callchain_ret = 0, printed = 0;
1985         int alignment = trace->args_alignment;
1986         struct syscall *sc = trace__syscall_info(trace, evsel, id);
1987         struct thread_trace *ttrace;
1988
1989         if (sc == NULL)
1990                 return -1;
1991
1992         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
1993         ttrace = thread__trace(thread, trace->output);
1994         if (ttrace == NULL)
1995                 goto out_put;
1996
1997         trace__fprintf_sample(trace, evsel, sample, thread);
1998
1999         if (trace->summary)
2000                 thread__update_stats(ttrace, id, sample);
2001
2002         ret = perf_evsel__sc_tp_uint(evsel, ret, sample);
2003
2004         if (sc->is_open && ret >= 0 && ttrace->filename.pending_open) {
2005                 trace__set_fd_pathname(thread, ret, ttrace->filename.name);
2006                 ttrace->filename.pending_open = false;
2007                 ++trace->stats.vfs_getname;
2008         }
2009
2010         if (ttrace->entry_time) {
2011                 duration = sample->time - ttrace->entry_time;
2012                 if (trace__filter_duration(trace, duration))
2013                         goto out;
2014                 duration_calculated = true;
2015         } else if (trace->duration_filter)
2016                 goto out;
2017
2018         if (sample->callchain) {
2019                 callchain_ret = trace__resolve_callchain(trace, evsel, sample, &callchain_cursor);
2020                 if (callchain_ret == 0) {
2021                         if (callchain_cursor.nr < trace->min_stack)
2022                                 goto out;
2023                         callchain_ret = 1;
2024                 }
2025         }
2026
2027         if (trace->summary_only || (ret >= 0 && trace->failure_only))
2028                 goto out;
2029
2030         trace__fprintf_entry_head(trace, thread, duration, duration_calculated, ttrace->entry_time, trace->output);
2031
2032         if (ttrace->entry_pending) {
2033                 printed = fprintf(trace->output, "%s", ttrace->entry_str);
2034         } else {
2035                 printed += fprintf(trace->output, " ... [");
2036                 color_fprintf(trace->output, PERF_COLOR_YELLOW, "continued");
2037                 printed += 9;
2038                 printed += fprintf(trace->output, "]: %s()", sc->name);
2039         }
2040
2041         printed++; /* the closing ')' */
2042
2043         if (alignment > printed)
2044                 alignment -= printed;
2045         else
2046                 alignment = 0;
2047
2048         fprintf(trace->output, ")%*s= ", alignment, " ");
2049
2050         if (sc->fmt == NULL) {
2051                 if (ret < 0)
2052                         goto errno_print;
2053 signed_print:
2054                 fprintf(trace->output, "%ld", ret);
2055         } else if (ret < 0) {
2056 errno_print: {
2057                 char bf[STRERR_BUFSIZE];
2058                 const char *emsg = str_error_r(-ret, bf, sizeof(bf)),
2059                            *e = errno_to_name(evsel, -ret);
2060
2061                 fprintf(trace->output, "-1 %s (%s)", e, emsg);
2062         }
2063         } else if (ret == 0 && sc->fmt->timeout)
2064                 fprintf(trace->output, "0 (Timeout)");
2065         else if (ttrace->ret_scnprintf) {
2066                 char bf[1024];
2067                 struct syscall_arg arg = {
2068                         .val    = ret,
2069                         .thread = thread,
2070                         .trace  = trace,
2071                 };
2072                 ttrace->ret_scnprintf(bf, sizeof(bf), &arg);
2073                 ttrace->ret_scnprintf = NULL;
2074                 fprintf(trace->output, "%s", bf);
2075         } else if (sc->fmt->hexret)
2076                 fprintf(trace->output, "%#lx", ret);
2077         else if (sc->fmt->errpid) {
2078                 struct thread *child = machine__find_thread(trace->host, ret, ret);
2079
2080                 if (child != NULL) {
2081                         fprintf(trace->output, "%ld", ret);
2082                         if (child->comm_set)
2083                                 fprintf(trace->output, " (%s)", thread__comm_str(child));
2084                         thread__put(child);
2085                 }
2086         } else
2087                 goto signed_print;
2088
2089         fputc('\n', trace->output);
2090
2091         /*
2092          * We only consider an 'event' for the sake of --max-events a non-filtered
2093          * sys_enter + sys_exit and other tracepoint events.
2094          */
2095         if (++trace->nr_events_printed == trace->max_events && trace->max_events != ULONG_MAX)
2096                 interrupted = true;
2097
2098         if (callchain_ret > 0)
2099                 trace__fprintf_callchain(trace, sample);
2100         else if (callchain_ret < 0)
2101                 pr_err("Problem processing %s callchain, skipping...\n", perf_evsel__name(evsel));
2102 out:
2103         ttrace->entry_pending = false;
2104         err = 0;
2105 out_put:
2106         thread__put(thread);
2107         return err;
2108 }
2109
2110 static int trace__vfs_getname(struct trace *trace, struct perf_evsel *evsel,
2111                               union perf_event *event __maybe_unused,
2112                               struct perf_sample *sample)
2113 {
2114         struct thread *thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2115         struct thread_trace *ttrace;
2116         size_t filename_len, entry_str_len, to_move;
2117         ssize_t remaining_space;
2118         char *pos;
2119         const char *filename = perf_evsel__rawptr(evsel, sample, "pathname");
2120
2121         if (!thread)
2122                 goto out;
2123
2124         ttrace = thread__priv(thread);
2125         if (!ttrace)
2126                 goto out_put;
2127
2128         filename_len = strlen(filename);
2129         if (filename_len == 0)
2130                 goto out_put;
2131
2132         if (ttrace->filename.namelen < filename_len) {
2133                 char *f = realloc(ttrace->filename.name, filename_len + 1);
2134
2135                 if (f == NULL)
2136                         goto out_put;
2137
2138                 ttrace->filename.namelen = filename_len;
2139                 ttrace->filename.name = f;
2140         }
2141
2142         strcpy(ttrace->filename.name, filename);
2143         ttrace->filename.pending_open = true;
2144
2145         if (!ttrace->filename.ptr)
2146                 goto out_put;
2147
2148         entry_str_len = strlen(ttrace->entry_str);
2149         remaining_space = trace__entry_str_size - entry_str_len - 1; /* \0 */
2150         if (remaining_space <= 0)
2151                 goto out_put;
2152
2153         if (filename_len > (size_t)remaining_space) {
2154                 filename += filename_len - remaining_space;
2155                 filename_len = remaining_space;
2156         }
2157
2158         to_move = entry_str_len - ttrace->filename.entry_str_pos + 1; /* \0 */
2159         pos = ttrace->entry_str + ttrace->filename.entry_str_pos;
2160         memmove(pos + filename_len, pos, to_move);
2161         memcpy(pos, filename, filename_len);
2162
2163         ttrace->filename.ptr = 0;
2164         ttrace->filename.entry_str_pos = 0;
2165 out_put:
2166         thread__put(thread);
2167 out:
2168         return 0;
2169 }
2170
2171 static int trace__sched_stat_runtime(struct trace *trace, struct perf_evsel *evsel,
2172                                      union perf_event *event __maybe_unused,
2173                                      struct perf_sample *sample)
2174 {
2175         u64 runtime = perf_evsel__intval(evsel, sample, "runtime");
2176         double runtime_ms = (double)runtime / NSEC_PER_MSEC;
2177         struct thread *thread = machine__findnew_thread(trace->host,
2178                                                         sample->pid,
2179                                                         sample->tid);
2180         struct thread_trace *ttrace = thread__trace(thread, trace->output);
2181
2182         if (ttrace == NULL)
2183                 goto out_dump;
2184
2185         ttrace->runtime_ms += runtime_ms;
2186         trace->runtime_ms += runtime_ms;
2187 out_put:
2188         thread__put(thread);
2189         return 0;
2190
2191 out_dump:
2192         fprintf(trace->output, "%s: comm=%s,pid=%u,runtime=%" PRIu64 ",vruntime=%" PRIu64 ")\n",
2193                evsel->name,
2194                perf_evsel__strval(evsel, sample, "comm"),
2195                (pid_t)perf_evsel__intval(evsel, sample, "pid"),
2196                runtime,
2197                perf_evsel__intval(evsel, sample, "vruntime"));
2198         goto out_put;
2199 }
2200
2201 static int bpf_output__printer(enum binary_printer_ops op,
2202                                unsigned int val, void *extra __maybe_unused, FILE *fp)
2203 {
2204         unsigned char ch = (unsigned char)val;
2205
2206         switch (op) {
2207         case BINARY_PRINT_CHAR_DATA:
2208                 return fprintf(fp, "%c", isprint(ch) ? ch : '.');
2209         case BINARY_PRINT_DATA_BEGIN:
2210         case BINARY_PRINT_LINE_BEGIN:
2211         case BINARY_PRINT_ADDR:
2212         case BINARY_PRINT_NUM_DATA:
2213         case BINARY_PRINT_NUM_PAD:
2214         case BINARY_PRINT_SEP:
2215         case BINARY_PRINT_CHAR_PAD:
2216         case BINARY_PRINT_LINE_END:
2217         case BINARY_PRINT_DATA_END:
2218         default:
2219                 break;
2220         }
2221
2222         return 0;
2223 }
2224
2225 static void bpf_output__fprintf(struct trace *trace,
2226                                 struct perf_sample *sample)
2227 {
2228         binary__fprintf(sample->raw_data, sample->raw_size, 8,
2229                         bpf_output__printer, NULL, trace->output);
2230         ++trace->nr_events_printed;
2231 }
2232
2233 static int trace__event_handler(struct trace *trace, struct perf_evsel *evsel,
2234                                 union perf_event *event __maybe_unused,
2235                                 struct perf_sample *sample)
2236 {
2237         struct thread *thread;
2238         int callchain_ret = 0;
2239         /*
2240          * Check if we called perf_evsel__disable(evsel) due to, for instance,
2241          * this event's max_events having been hit and this is an entry coming
2242          * from the ring buffer that we should discard, since the max events
2243          * have already been considered/printed.
2244          */
2245         if (evsel->disabled)
2246                 return 0;
2247
2248         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2249
2250         if (sample->callchain) {
2251                 callchain_ret = trace__resolve_callchain(trace, evsel, sample, &callchain_cursor);
2252                 if (callchain_ret == 0) {
2253                         if (callchain_cursor.nr < trace->min_stack)
2254                                 goto out;
2255                         callchain_ret = 1;
2256                 }
2257         }
2258
2259         trace__printf_interrupted_entry(trace);
2260         trace__fprintf_tstamp(trace, sample->time, trace->output);
2261
2262         if (trace->trace_syscalls && trace->show_duration)
2263                 fprintf(trace->output, "(         ): ");
2264
2265         if (thread)
2266                 trace__fprintf_comm_tid(trace, thread, trace->output);
2267
2268         if (evsel == trace->syscalls.events.augmented) {
2269                 int id = perf_evsel__sc_tp_uint(evsel, id, sample);
2270                 struct syscall *sc = trace__syscall_info(trace, evsel, id);
2271
2272                 if (sc) {
2273                         fprintf(trace->output, "%s(", sc->name);
2274                         trace__fprintf_sys_enter(trace, evsel, sample);
2275                         fputc(')', trace->output);
2276                         goto newline;
2277                 }
2278
2279                 /*
2280                  * XXX: Not having the associated syscall info or not finding/adding
2281                  *      the thread should never happen, but if it does...
2282                  *      fall thru and print it as a bpf_output event.
2283                  */
2284         }
2285
2286         fprintf(trace->output, "%s:", evsel->name);
2287
2288         if (perf_evsel__is_bpf_output(evsel)) {
2289                 bpf_output__fprintf(trace, sample);
2290         } else if (evsel->tp_format) {
2291                 if (strncmp(evsel->tp_format->name, "sys_enter_", 10) ||
2292                     trace__fprintf_sys_enter(trace, evsel, sample)) {
2293                         event_format__fprintf(evsel->tp_format, sample->cpu,
2294                                               sample->raw_data, sample->raw_size,
2295                                               trace->output);
2296                         ++trace->nr_events_printed;
2297
2298                         if (evsel->max_events != ULONG_MAX && ++evsel->nr_events_printed == evsel->max_events) {
2299                                 perf_evsel__disable(evsel);
2300                                 perf_evsel__close(evsel);
2301                         }
2302                 }
2303         }
2304
2305 newline:
2306         fprintf(trace->output, "\n");
2307
2308         if (callchain_ret > 0)
2309                 trace__fprintf_callchain(trace, sample);
2310         else if (callchain_ret < 0)
2311                 pr_err("Problem processing %s callchain, skipping...\n", perf_evsel__name(evsel));
2312 out:
2313         thread__put(thread);
2314         return 0;
2315 }
2316
2317 static void print_location(FILE *f, struct perf_sample *sample,
2318                            struct addr_location *al,
2319                            bool print_dso, bool print_sym)
2320 {
2321
2322         if ((verbose > 0 || print_dso) && al->map)
2323                 fprintf(f, "%s@", al->map->dso->long_name);
2324
2325         if ((verbose > 0 || print_sym) && al->sym)
2326                 fprintf(f, "%s+0x%" PRIx64, al->sym->name,
2327                         al->addr - al->sym->start);
2328         else if (al->map)
2329                 fprintf(f, "0x%" PRIx64, al->addr);
2330         else
2331                 fprintf(f, "0x%" PRIx64, sample->addr);
2332 }
2333
2334 static int trace__pgfault(struct trace *trace,
2335                           struct perf_evsel *evsel,
2336                           union perf_event *event __maybe_unused,
2337                           struct perf_sample *sample)
2338 {
2339         struct thread *thread;
2340         struct addr_location al;
2341         char map_type = 'd';
2342         struct thread_trace *ttrace;
2343         int err = -1;
2344         int callchain_ret = 0;
2345
2346         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2347
2348         if (sample->callchain) {
2349                 callchain_ret = trace__resolve_callchain(trace, evsel, sample, &callchain_cursor);
2350                 if (callchain_ret == 0) {
2351                         if (callchain_cursor.nr < trace->min_stack)
2352                                 goto out_put;
2353                         callchain_ret = 1;
2354                 }
2355         }
2356
2357         ttrace = thread__trace(thread, trace->output);
2358         if (ttrace == NULL)
2359                 goto out_put;
2360
2361         if (evsel->attr.config == PERF_COUNT_SW_PAGE_FAULTS_MAJ)
2362                 ttrace->pfmaj++;
2363         else
2364                 ttrace->pfmin++;
2365
2366         if (trace->summary_only)
2367                 goto out;
2368
2369         thread__find_symbol(thread, sample->cpumode, sample->ip, &al);
2370
2371         trace__fprintf_entry_head(trace, thread, 0, true, sample->time, trace->output);
2372
2373         fprintf(trace->output, "%sfault [",
2374                 evsel->attr.config == PERF_COUNT_SW_PAGE_FAULTS_MAJ ?
2375                 "maj" : "min");
2376
2377         print_location(trace->output, sample, &al, false, true);
2378
2379         fprintf(trace->output, "] => ");
2380
2381         thread__find_symbol(thread, sample->cpumode, sample->addr, &al);
2382
2383         if (!al.map) {
2384                 thread__find_symbol(thread, sample->cpumode, sample->addr, &al);
2385
2386                 if (al.map)
2387                         map_type = 'x';
2388                 else
2389                         map_type = '?';
2390         }
2391
2392         print_location(trace->output, sample, &al, true, false);
2393
2394         fprintf(trace->output, " (%c%c)\n", map_type, al.level);
2395
2396         if (callchain_ret > 0)
2397                 trace__fprintf_callchain(trace, sample);
2398         else if (callchain_ret < 0)
2399                 pr_err("Problem processing %s callchain, skipping...\n", perf_evsel__name(evsel));
2400
2401         ++trace->nr_events_printed;
2402 out:
2403         err = 0;
2404 out_put:
2405         thread__put(thread);
2406         return err;
2407 }
2408
2409 static void trace__set_base_time(struct trace *trace,
2410                                  struct perf_evsel *evsel,
2411                                  struct perf_sample *sample)
2412 {
2413         /*
2414          * BPF events were not setting PERF_SAMPLE_TIME, so be more robust
2415          * and don't use sample->time unconditionally, we may end up having
2416          * some other event in the future without PERF_SAMPLE_TIME for good
2417          * reason, i.e. we may not be interested in its timestamps, just in
2418          * it taking place, picking some piece of information when it
2419          * appears in our event stream (vfs_getname comes to mind).
2420          */
2421         if (trace->base_time == 0 && !trace->full_time &&
2422             (evsel->attr.sample_type & PERF_SAMPLE_TIME))
2423                 trace->base_time = sample->time;
2424 }
2425
2426 static int trace__process_sample(struct perf_tool *tool,
2427                                  union perf_event *event,
2428                                  struct perf_sample *sample,
2429                                  struct perf_evsel *evsel,
2430                                  struct machine *machine __maybe_unused)
2431 {
2432         struct trace *trace = container_of(tool, struct trace, tool);
2433         struct thread *thread;
2434         int err = 0;
2435
2436         tracepoint_handler handler = evsel->handler;
2437
2438         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2439         if (thread && thread__is_filtered(thread))
2440                 goto out;
2441
2442         trace__set_base_time(trace, evsel, sample);
2443
2444         if (handler) {
2445                 ++trace->nr_events;
2446                 handler(trace, evsel, event, sample);
2447         }
2448 out:
2449         thread__put(thread);
2450         return err;
2451 }
2452
2453 static int trace__record(struct trace *trace, int argc, const char **argv)
2454 {
2455         unsigned int rec_argc, i, j;
2456         const char **rec_argv;
2457         const char * const record_args[] = {
2458                 "record",
2459                 "-R",
2460                 "-m", "1024",
2461                 "-c", "1",
2462         };
2463
2464         const char * const sc_args[] = { "-e", };
2465         unsigned int sc_args_nr = ARRAY_SIZE(sc_args);
2466         const char * const majpf_args[] = { "-e", "major-faults" };
2467         unsigned int majpf_args_nr = ARRAY_SIZE(majpf_args);
2468         const char * const minpf_args[] = { "-e", "minor-faults" };
2469         unsigned int minpf_args_nr = ARRAY_SIZE(minpf_args);
2470
2471         /* +1 is for the event string below */
2472         rec_argc = ARRAY_SIZE(record_args) + sc_args_nr + 1 +
2473                 majpf_args_nr + minpf_args_nr + argc;
2474         rec_argv = calloc(rec_argc + 1, sizeof(char *));
2475
2476         if (rec_argv == NULL)
2477                 return -ENOMEM;
2478
2479         j = 0;
2480         for (i = 0; i < ARRAY_SIZE(record_args); i++)
2481                 rec_argv[j++] = record_args[i];
2482
2483         if (trace->trace_syscalls) {
2484                 for (i = 0; i < sc_args_nr; i++)
2485                         rec_argv[j++] = sc_args[i];
2486
2487                 /* event string may be different for older kernels - e.g., RHEL6 */
2488                 if (is_valid_tracepoint("raw_syscalls:sys_enter"))
2489                         rec_argv[j++] = "raw_syscalls:sys_enter,raw_syscalls:sys_exit";
2490                 else if (is_valid_tracepoint("syscalls:sys_enter"))
2491                         rec_argv[j++] = "syscalls:sys_enter,syscalls:sys_exit";
2492                 else {
2493                         pr_err("Neither raw_syscalls nor syscalls events exist.\n");
2494                         free(rec_argv);
2495                         return -1;
2496                 }
2497         }
2498
2499         if (trace->trace_pgfaults & TRACE_PFMAJ)
2500                 for (i = 0; i < majpf_args_nr; i++)
2501                         rec_argv[j++] = majpf_args[i];
2502
2503         if (trace->trace_pgfaults & TRACE_PFMIN)
2504                 for (i = 0; i < minpf_args_nr; i++)
2505                         rec_argv[j++] = minpf_args[i];
2506
2507         for (i = 0; i < (unsigned int)argc; i++)
2508                 rec_argv[j++] = argv[i];
2509
2510         return cmd_record(j, rec_argv);
2511 }
2512
2513 static size_t trace__fprintf_thread_summary(struct trace *trace, FILE *fp);
2514
2515 static bool perf_evlist__add_vfs_getname(struct perf_evlist *evlist)
2516 {
2517         struct perf_evsel *evsel = perf_evsel__newtp("probe", "vfs_getname");
2518
2519         if (IS_ERR(evsel))
2520                 return false;
2521
2522         if (perf_evsel__field(evsel, "pathname") == NULL) {
2523                 perf_evsel__delete(evsel);
2524                 return false;
2525         }
2526
2527         evsel->handler = trace__vfs_getname;
2528         perf_evlist__add(evlist, evsel);
2529         return true;
2530 }
2531
2532 static struct perf_evsel *perf_evsel__new_pgfault(u64 config)
2533 {
2534         struct perf_evsel *evsel;
2535         struct perf_event_attr attr = {
2536                 .type = PERF_TYPE_SOFTWARE,
2537                 .mmap_data = 1,
2538         };
2539
2540         attr.config = config;
2541         attr.sample_period = 1;
2542
2543         event_attr_init(&attr);
2544
2545         evsel = perf_evsel__new(&attr);
2546         if (evsel)
2547                 evsel->handler = trace__pgfault;
2548
2549         return evsel;
2550 }
2551
2552 static void trace__handle_event(struct trace *trace, union perf_event *event, struct perf_sample *sample)
2553 {
2554         const u32 type = event->header.type;
2555         struct perf_evsel *evsel;
2556
2557         if (type != PERF_RECORD_SAMPLE) {
2558                 trace__process_event(trace, trace->host, event, sample);
2559                 return;
2560         }
2561
2562         evsel = perf_evlist__id2evsel(trace->evlist, sample->id);
2563         if (evsel == NULL) {
2564                 fprintf(trace->output, "Unknown tp ID %" PRIu64 ", skipping...\n", sample->id);
2565                 return;
2566         }
2567
2568         trace__set_base_time(trace, evsel, sample);
2569
2570         if (evsel->attr.type == PERF_TYPE_TRACEPOINT &&
2571             sample->raw_data == NULL) {
2572                 fprintf(trace->output, "%s sample with no payload for tid: %d, cpu %d, raw_size=%d, skipping...\n",
2573                        perf_evsel__name(evsel), sample->tid,
2574                        sample->cpu, sample->raw_size);
2575         } else {
2576                 tracepoint_handler handler = evsel->handler;
2577                 handler(trace, evsel, event, sample);
2578         }
2579
2580         if (trace->nr_events_printed >= trace->max_events && trace->max_events != ULONG_MAX)
2581                 interrupted = true;
2582 }
2583
2584 static int trace__add_syscall_newtp(struct trace *trace)
2585 {
2586         int ret = -1;
2587         struct perf_evlist *evlist = trace->evlist;
2588         struct perf_evsel *sys_enter, *sys_exit;
2589
2590         sys_enter = perf_evsel__raw_syscall_newtp("sys_enter", trace__sys_enter);
2591         if (sys_enter == NULL)
2592                 goto out;
2593
2594         if (perf_evsel__init_sc_tp_ptr_field(sys_enter, args))
2595                 goto out_delete_sys_enter;
2596
2597         sys_exit = perf_evsel__raw_syscall_newtp("sys_exit", trace__sys_exit);
2598         if (sys_exit == NULL)
2599                 goto out_delete_sys_enter;
2600
2601         if (perf_evsel__init_sc_tp_uint_field(sys_exit, ret))
2602                 goto out_delete_sys_exit;
2603
2604         perf_evsel__config_callchain(sys_enter, &trace->opts, &callchain_param);
2605         perf_evsel__config_callchain(sys_exit, &trace->opts, &callchain_param);
2606
2607         perf_evlist__add(evlist, sys_enter);
2608         perf_evlist__add(evlist, sys_exit);
2609
2610         if (callchain_param.enabled && !trace->kernel_syscallchains) {
2611                 /*
2612                  * We're interested only in the user space callchain
2613                  * leading to the syscall, allow overriding that for
2614                  * debugging reasons using --kernel_syscall_callchains
2615                  */
2616                 sys_exit->attr.exclude_callchain_kernel = 1;
2617         }
2618
2619         trace->syscalls.events.sys_enter = sys_enter;
2620         trace->syscalls.events.sys_exit  = sys_exit;
2621
2622         ret = 0;
2623 out:
2624         return ret;
2625
2626 out_delete_sys_exit:
2627         perf_evsel__delete_priv(sys_exit);
2628 out_delete_sys_enter:
2629         perf_evsel__delete_priv(sys_enter);
2630         goto out;
2631 }
2632
2633 static int trace__set_ev_qualifier_tp_filter(struct trace *trace)
2634 {
2635         int err = -1;
2636         struct perf_evsel *sys_exit;
2637         char *filter = asprintf_expr_inout_ints("id", !trace->not_ev_qualifier,
2638                                                 trace->ev_qualifier_ids.nr,
2639                                                 trace->ev_qualifier_ids.entries);
2640
2641         if (filter == NULL)
2642                 goto out_enomem;
2643
2644         if (!perf_evsel__append_tp_filter(trace->syscalls.events.sys_enter,
2645                                           filter)) {
2646                 sys_exit = trace->syscalls.events.sys_exit;
2647                 err = perf_evsel__append_tp_filter(sys_exit, filter);
2648         }
2649
2650         free(filter);
2651 out:
2652         return err;
2653 out_enomem:
2654         errno = ENOMEM;
2655         goto out;
2656 }
2657
2658 #ifdef HAVE_LIBBPF_SUPPORT
2659 static int trace__set_ev_qualifier_bpf_filter(struct trace *trace)
2660 {
2661         int fd = bpf_map__fd(trace->syscalls.map);
2662         struct bpf_map_syscall_entry value = {
2663                 .enabled = !trace->not_ev_qualifier,
2664         };
2665         int err = 0;
2666         size_t i;
2667
2668         for (i = 0; i < trace->ev_qualifier_ids.nr; ++i) {
2669                 int key = trace->ev_qualifier_ids.entries[i];
2670
2671                 err = bpf_map_update_elem(fd, &key, &value, BPF_EXIST);
2672                 if (err)
2673                         break;
2674         }
2675
2676         return err;
2677 }
2678
2679 static int __trace__init_syscalls_bpf_map(struct trace *trace, bool enabled)
2680 {
2681         int fd = bpf_map__fd(trace->syscalls.map);
2682         struct bpf_map_syscall_entry value = {
2683                 .enabled = enabled,
2684         };
2685         int err = 0, key;
2686
2687         for (key = 0; key < trace->sctbl->syscalls.nr_entries; ++key) {
2688                 err = bpf_map_update_elem(fd, &key, &value, BPF_ANY);
2689                 if (err)
2690                         break;
2691         }
2692
2693         return err;
2694 }
2695
2696 static int trace__init_syscalls_bpf_map(struct trace *trace)
2697 {
2698         bool enabled = true;
2699
2700         if (trace->ev_qualifier_ids.nr)
2701                 enabled = trace->not_ev_qualifier;
2702
2703         return __trace__init_syscalls_bpf_map(trace, enabled);
2704 }
2705 #else
2706 static int trace__set_ev_qualifier_bpf_filter(struct trace *trace __maybe_unused)
2707 {
2708         return 0;
2709 }
2710
2711 static int trace__init_syscalls_bpf_map(struct trace *trace __maybe_unused)
2712 {
2713         return 0;
2714 }
2715 #endif // HAVE_LIBBPF_SUPPORT
2716
2717 static int trace__set_ev_qualifier_filter(struct trace *trace)
2718 {
2719         if (trace->syscalls.map)
2720                 return trace__set_ev_qualifier_bpf_filter(trace);
2721         if (trace->syscalls.events.sys_enter)
2722                 return trace__set_ev_qualifier_tp_filter(trace);
2723         return 0;
2724 }
2725
2726 static int bpf_map__set_filter_pids(struct bpf_map *map __maybe_unused,
2727                                     size_t npids __maybe_unused, pid_t *pids __maybe_unused)
2728 {
2729         int err = 0;
2730 #ifdef HAVE_LIBBPF_SUPPORT
2731         bool value = true;
2732         int map_fd = bpf_map__fd(map);
2733         size_t i;
2734
2735         for (i = 0; i < npids; ++i) {
2736                 err = bpf_map_update_elem(map_fd, &pids[i], &value, BPF_ANY);
2737                 if (err)
2738                         break;
2739         }
2740 #endif
2741         return err;
2742 }
2743
2744 static int trace__set_filter_loop_pids(struct trace *trace)
2745 {
2746         unsigned int nr = 1, err;
2747         pid_t pids[32] = {
2748                 getpid(),
2749         };
2750         struct thread *thread = machine__find_thread(trace->host, pids[0], pids[0]);
2751
2752         while (thread && nr < ARRAY_SIZE(pids)) {
2753                 struct thread *parent = machine__find_thread(trace->host, thread->ppid, thread->ppid);
2754
2755                 if (parent == NULL)
2756                         break;
2757
2758                 if (!strcmp(thread__comm_str(parent), "sshd")) {
2759                         pids[nr++] = parent->tid;
2760                         break;
2761                 }
2762                 thread = parent;
2763         }
2764
2765         err = perf_evlist__set_tp_filter_pids(trace->evlist, nr, pids);
2766         if (!err && trace->filter_pids.map)
2767                 err = bpf_map__set_filter_pids(trace->filter_pids.map, nr, pids);
2768
2769         return err;
2770 }
2771
2772 static int trace__set_filter_pids(struct trace *trace)
2773 {
2774         int err = 0;
2775         /*
2776          * Better not use !target__has_task() here because we need to cover the
2777          * case where no threads were specified in the command line, but a
2778          * workload was, and in that case we will fill in the thread_map when
2779          * we fork the workload in perf_evlist__prepare_workload.
2780          */
2781         if (trace->filter_pids.nr > 0) {
2782                 err = perf_evlist__set_tp_filter_pids(trace->evlist, trace->filter_pids.nr,
2783                                                       trace->filter_pids.entries);
2784                 if (!err && trace->filter_pids.map) {
2785                         err = bpf_map__set_filter_pids(trace->filter_pids.map, trace->filter_pids.nr,
2786                                                        trace->filter_pids.entries);
2787                 }
2788         } else if (thread_map__pid(trace->evlist->threads, 0) == -1) {
2789                 err = trace__set_filter_loop_pids(trace);
2790         }
2791
2792         return err;
2793 }
2794
2795 static int __trace__deliver_event(struct trace *trace, union perf_event *event)
2796 {
2797         struct perf_evlist *evlist = trace->evlist;
2798         struct perf_sample sample;
2799         int err;
2800
2801         err = perf_evlist__parse_sample(evlist, event, &sample);
2802         if (err)
2803                 fprintf(trace->output, "Can't parse sample, err = %d, skipping...\n", err);
2804         else
2805                 trace__handle_event(trace, event, &sample);
2806
2807         return 0;
2808 }
2809
2810 static int __trace__flush_events(struct trace *trace)
2811 {
2812         u64 first = ordered_events__first_time(&trace->oe.data);
2813         u64 flush = trace->oe.last - NSEC_PER_SEC;
2814
2815         /* Is there some thing to flush.. */
2816         if (first && first < flush)
2817                 return ordered_events__flush_time(&trace->oe.data, flush);
2818
2819         return 0;
2820 }
2821
2822 static int trace__flush_events(struct trace *trace)
2823 {
2824         return !trace->sort_events ? 0 : __trace__flush_events(trace);
2825 }
2826
2827 static int trace__deliver_event(struct trace *trace, union perf_event *event)
2828 {
2829         int err;
2830
2831         if (!trace->sort_events)
2832                 return __trace__deliver_event(trace, event);
2833
2834         err = perf_evlist__parse_sample_timestamp(trace->evlist, event, &trace->oe.last);
2835         if (err && err != -1)
2836                 return err;
2837
2838         err = ordered_events__queue(&trace->oe.data, event, trace->oe.last, 0);
2839         if (err)
2840                 return err;
2841
2842         return trace__flush_events(trace);
2843 }
2844
2845 static int ordered_events__deliver_event(struct ordered_events *oe,
2846                                          struct ordered_event *event)
2847 {
2848         struct trace *trace = container_of(oe, struct trace, oe.data);
2849
2850         return __trace__deliver_event(trace, event->event);
2851 }
2852
2853 static int trace__run(struct trace *trace, int argc, const char **argv)
2854 {
2855         struct perf_evlist *evlist = trace->evlist;
2856         struct perf_evsel *evsel, *pgfault_maj = NULL, *pgfault_min = NULL;
2857         int err = -1, i;
2858         unsigned long before;
2859         const bool forks = argc > 0;
2860         bool draining = false;
2861
2862         trace->live = true;
2863
2864         if (!trace->raw_augmented_syscalls) {
2865                 if (trace->trace_syscalls && trace__add_syscall_newtp(trace))
2866                         goto out_error_raw_syscalls;
2867
2868                 if (trace->trace_syscalls)
2869                         trace->vfs_getname = perf_evlist__add_vfs_getname(evlist);
2870         }
2871
2872         if ((trace->trace_pgfaults & TRACE_PFMAJ)) {
2873                 pgfault_maj = perf_evsel__new_pgfault(PERF_COUNT_SW_PAGE_FAULTS_MAJ);
2874                 if (pgfault_maj == NULL)
2875                         goto out_error_mem;
2876                 perf_evsel__config_callchain(pgfault_maj, &trace->opts, &callchain_param);
2877                 perf_evlist__add(evlist, pgfault_maj);
2878         }
2879
2880         if ((trace->trace_pgfaults & TRACE_PFMIN)) {
2881                 pgfault_min = perf_evsel__new_pgfault(PERF_COUNT_SW_PAGE_FAULTS_MIN);
2882                 if (pgfault_min == NULL)
2883                         goto out_error_mem;
2884                 perf_evsel__config_callchain(pgfault_min, &trace->opts, &callchain_param);
2885                 perf_evlist__add(evlist, pgfault_min);
2886         }
2887
2888         if (trace->sched &&
2889             perf_evlist__add_newtp(evlist, "sched", "sched_stat_runtime",
2890                                    trace__sched_stat_runtime))
2891                 goto out_error_sched_stat_runtime;
2892
2893         /*
2894          * If a global cgroup was set, apply it to all the events without an
2895          * explicit cgroup. I.e.:
2896          *
2897          *      trace -G A -e sched:*switch
2898          *
2899          * Will set all raw_syscalls:sys_{enter,exit}, pgfault, vfs_getname, etc
2900          * _and_ sched:sched_switch to the 'A' cgroup, while:
2901          *
2902          * trace -e sched:*switch -G A
2903          *
2904          * will only set the sched:sched_switch event to the 'A' cgroup, all the
2905          * other events (raw_syscalls:sys_{enter,exit}, etc are left "without"
2906          * a cgroup (on the root cgroup, sys wide, etc).
2907          *
2908          * Multiple cgroups:
2909          *
2910          * trace -G A -e sched:*switch -G B
2911          *
2912          * the syscall ones go to the 'A' cgroup, the sched:sched_switch goes
2913          * to the 'B' cgroup.
2914          *
2915          * evlist__set_default_cgroup() grabs a reference of the passed cgroup
2916          * only for the evsels still without a cgroup, i.e. evsel->cgroup == NULL.
2917          */
2918         if (trace->cgroup)
2919                 evlist__set_default_cgroup(trace->evlist, trace->cgroup);
2920
2921         err = perf_evlist__create_maps(evlist, &trace->opts.target);
2922         if (err < 0) {
2923                 fprintf(trace->output, "Problems parsing the target to trace, check your options!\n");
2924                 goto out_delete_evlist;
2925         }
2926
2927         err = trace__symbols_init(trace, evlist);
2928         if (err < 0) {
2929                 fprintf(trace->output, "Problems initializing symbol libraries!\n");
2930                 goto out_delete_evlist;
2931         }
2932
2933         perf_evlist__config(evlist, &trace->opts, &callchain_param);
2934
2935         signal(SIGCHLD, sig_handler);
2936         signal(SIGINT, sig_handler);
2937
2938         if (forks) {
2939                 err = perf_evlist__prepare_workload(evlist, &trace->opts.target,
2940                                                     argv, false, NULL);
2941                 if (err < 0) {
2942                         fprintf(trace->output, "Couldn't run the workload!\n");
2943                         goto out_delete_evlist;
2944                 }
2945         }
2946
2947         err = perf_evlist__open(evlist);
2948         if (err < 0)
2949                 goto out_error_open;
2950
2951         err = bpf__apply_obj_config();
2952         if (err) {
2953                 char errbuf[BUFSIZ];
2954
2955                 bpf__strerror_apply_obj_config(err, errbuf, sizeof(errbuf));
2956                 pr_err("ERROR: Apply config to BPF failed: %s\n",
2957                          errbuf);
2958                 goto out_error_open;
2959         }
2960
2961         err = trace__set_filter_pids(trace);
2962         if (err < 0)
2963                 goto out_error_mem;
2964
2965         if (trace->syscalls.map)
2966                 trace__init_syscalls_bpf_map(trace);
2967
2968         if (trace->ev_qualifier_ids.nr > 0) {
2969                 err = trace__set_ev_qualifier_filter(trace);
2970                 if (err < 0)
2971                         goto out_errno;
2972
2973                 if (trace->syscalls.events.sys_exit) {
2974                         pr_debug("event qualifier tracepoint filter: %s\n",
2975                                  trace->syscalls.events.sys_exit->filter);
2976                 }
2977         }
2978
2979         err = perf_evlist__apply_filters(evlist, &evsel);
2980         if (err < 0)
2981                 goto out_error_apply_filters;
2982
2983         err = perf_evlist__mmap(evlist, trace->opts.mmap_pages);
2984         if (err < 0)
2985                 goto out_error_mmap;
2986
2987         if (!target__none(&trace->opts.target) && !trace->opts.initial_delay)
2988                 perf_evlist__enable(evlist);
2989
2990         if (forks)
2991                 perf_evlist__start_workload(evlist);
2992
2993         if (trace->opts.initial_delay) {
2994                 usleep(trace->opts.initial_delay * 1000);
2995                 perf_evlist__enable(evlist);
2996         }
2997
2998         trace->multiple_threads = thread_map__pid(evlist->threads, 0) == -1 ||
2999                                   evlist->threads->nr > 1 ||
3000                                   perf_evlist__first(evlist)->attr.inherit;
3001
3002         /*
3003          * Now that we already used evsel->attr to ask the kernel to setup the
3004          * events, lets reuse evsel->attr.sample_max_stack as the limit in
3005          * trace__resolve_callchain(), allowing per-event max-stack settings
3006          * to override an explicitly set --max-stack global setting.
3007          */
3008         evlist__for_each_entry(evlist, evsel) {
3009                 if (evsel__has_callchain(evsel) &&
3010                     evsel->attr.sample_max_stack == 0)
3011                         evsel->attr.sample_max_stack = trace->max_stack;
3012         }
3013 again:
3014         before = trace->nr_events;
3015
3016         for (i = 0; i < evlist->nr_mmaps; i++) {
3017                 union perf_event *event;
3018                 struct perf_mmap *md;
3019
3020                 md = &evlist->mmap[i];
3021                 if (perf_mmap__read_init(md) < 0)
3022                         continue;
3023
3024                 while ((event = perf_mmap__read_event(md)) != NULL) {
3025                         ++trace->nr_events;
3026
3027                         err = trace__deliver_event(trace, event);
3028                         if (err)
3029                                 goto out_disable;
3030
3031                         perf_mmap__consume(md);
3032
3033                         if (interrupted)
3034                                 goto out_disable;
3035
3036                         if (done && !draining) {
3037                                 perf_evlist__disable(evlist);
3038                                 draining = true;
3039                         }
3040                 }
3041                 perf_mmap__read_done(md);
3042         }
3043
3044         if (trace->nr_events == before) {
3045                 int timeout = done ? 100 : -1;
3046
3047                 if (!draining && perf_evlist__poll(evlist, timeout) > 0) {
3048                         if (perf_evlist__filter_pollfd(evlist, POLLERR | POLLHUP | POLLNVAL) == 0)
3049                                 draining = true;
3050
3051                         goto again;
3052                 } else {
3053                         if (trace__flush_events(trace))
3054                                 goto out_disable;
3055                 }
3056         } else {
3057                 goto again;
3058         }
3059
3060 out_disable:
3061         thread__zput(trace->current);
3062
3063         perf_evlist__disable(evlist);
3064
3065         if (trace->sort_events)
3066                 ordered_events__flush(&trace->oe.data, OE_FLUSH__FINAL);
3067
3068         if (!err) {
3069                 if (trace->summary)
3070                         trace__fprintf_thread_summary(trace, trace->output);
3071
3072                 if (trace->show_tool_stats) {
3073                         fprintf(trace->output, "Stats:\n "
3074                                                " vfs_getname : %" PRIu64 "\n"
3075                                                " proc_getname: %" PRIu64 "\n",
3076                                 trace->stats.vfs_getname,
3077                                 trace->stats.proc_getname);
3078                 }
3079         }
3080
3081 out_delete_evlist:
3082         trace__symbols__exit(trace);
3083
3084         perf_evlist__delete(evlist);
3085         cgroup__put(trace->cgroup);
3086         trace->evlist = NULL;
3087         trace->live = false;
3088         return err;
3089 {
3090         char errbuf[BUFSIZ];
3091
3092 out_error_sched_stat_runtime:
3093         tracing_path__strerror_open_tp(errno, errbuf, sizeof(errbuf), "sched", "sched_stat_runtime");
3094         goto out_error;
3095
3096 out_error_raw_syscalls:
3097         tracing_path__strerror_open_tp(errno, errbuf, sizeof(errbuf), "raw_syscalls", "sys_(enter|exit)");
3098         goto out_error;
3099
3100 out_error_mmap:
3101         perf_evlist__strerror_mmap(evlist, errno, errbuf, sizeof(errbuf));
3102         goto out_error;
3103
3104 out_error_open:
3105         perf_evlist__strerror_open(evlist, errno, errbuf, sizeof(errbuf));
3106
3107 out_error:
3108         fprintf(trace->output, "%s\n", errbuf);
3109         goto out_delete_evlist;
3110
3111 out_error_apply_filters:
3112         fprintf(trace->output,
3113                 "Failed to set filter \"%s\" on event %s with %d (%s)\n",
3114                 evsel->filter, perf_evsel__name(evsel), errno,
3115                 str_error_r(errno, errbuf, sizeof(errbuf)));
3116         goto out_delete_evlist;
3117 }
3118 out_error_mem:
3119         fprintf(trace->output, "Not enough memory to run!\n");
3120         goto out_delete_evlist;
3121
3122 out_errno:
3123         fprintf(trace->output, "errno=%d,%s\n", errno, strerror(errno));
3124         goto out_delete_evlist;
3125 }
3126
3127 static int trace__replay(struct trace *trace)
3128 {
3129         const struct perf_evsel_str_handler handlers[] = {
3130                 { "probe:vfs_getname",       trace__vfs_getname, },
3131         };
3132         struct perf_data data = {
3133                 .file      = {
3134                         .path = input_name,
3135                 },
3136                 .mode      = PERF_DATA_MODE_READ,
3137                 .force     = trace->force,
3138         };
3139         struct perf_session *session;
3140         struct perf_evsel *evsel;
3141         int err = -1;
3142
3143         trace->tool.sample        = trace__process_sample;
3144         trace->tool.mmap          = perf_event__process_mmap;
3145         trace->tool.mmap2         = perf_event__process_mmap2;
3146         trace->tool.comm          = perf_event__process_comm;
3147         trace->tool.exit          = perf_event__process_exit;
3148         trace->tool.fork          = perf_event__process_fork;
3149         trace->tool.attr          = perf_event__process_attr;
3150         trace->tool.tracing_data  = perf_event__process_tracing_data;
3151         trace->tool.build_id      = perf_event__process_build_id;
3152         trace->tool.namespaces    = perf_event__process_namespaces;
3153
3154         trace->tool.ordered_events = true;
3155         trace->tool.ordering_requires_timestamps = true;
3156
3157         /* add tid to output */
3158         trace->multiple_threads = true;
3159
3160         session = perf_session__new(&data, false, &trace->tool);
3161         if (session == NULL)
3162                 return -1;
3163
3164         if (trace->opts.target.pid)
3165                 symbol_conf.pid_list_str = strdup(trace->opts.target.pid);
3166
3167         if (trace->opts.target.tid)
3168                 symbol_conf.tid_list_str = strdup(trace->opts.target.tid);
3169
3170         if (symbol__init(&session->header.env) < 0)
3171                 goto out;
3172
3173         trace->host = &session->machines.host;
3174
3175         err = perf_session__set_tracepoints_handlers(session, handlers);
3176         if (err)
3177                 goto out;
3178
3179         evsel = perf_evlist__find_tracepoint_by_name(session->evlist,
3180                                                      "raw_syscalls:sys_enter");
3181         /* older kernels have syscalls tp versus raw_syscalls */
3182         if (evsel == NULL)
3183                 evsel = perf_evlist__find_tracepoint_by_name(session->evlist,
3184                                                              "syscalls:sys_enter");
3185
3186         if (evsel &&
3187             (perf_evsel__init_raw_syscall_tp(evsel, trace__sys_enter) < 0 ||
3188             perf_evsel__init_sc_tp_ptr_field(evsel, args))) {
3189                 pr_err("Error during initialize raw_syscalls:sys_enter event\n");
3190                 goto out;
3191         }
3192
3193         evsel = perf_evlist__find_tracepoint_by_name(session->evlist,
3194                                                      "raw_syscalls:sys_exit");
3195         if (evsel == NULL)
3196                 evsel = perf_evlist__find_tracepoint_by_name(session->evlist,
3197                                                              "syscalls:sys_exit");
3198         if (evsel &&
3199             (perf_evsel__init_raw_syscall_tp(evsel, trace__sys_exit) < 0 ||
3200             perf_evsel__init_sc_tp_uint_field(evsel, ret))) {
3201                 pr_err("Error during initialize raw_syscalls:sys_exit event\n");
3202                 goto out;
3203         }
3204
3205         evlist__for_each_entry(session->evlist, evsel) {
3206                 if (evsel->attr.type == PERF_TYPE_SOFTWARE &&
3207                     (evsel->attr.config == PERF_COUNT_SW_PAGE_FAULTS_MAJ ||
3208                      evsel->attr.config == PERF_COUNT_SW_PAGE_FAULTS_MIN ||
3209                      evsel->attr.config == PERF_COUNT_SW_PAGE_FAULTS))
3210                         evsel->handler = trace__pgfault;
3211         }
3212
3213         setup_pager();
3214
3215         err = perf_session__process_events(session);
3216         if (err)
3217                 pr_err("Failed to process events, error %d", err);
3218
3219         else if (trace->summary)
3220                 trace__fprintf_thread_summary(trace, trace->output);
3221
3222 out:
3223         perf_session__delete(session);
3224
3225         return err;
3226 }
3227
3228 static size_t trace__fprintf_threads_header(FILE *fp)
3229 {
3230         size_t printed;
3231
3232         printed  = fprintf(fp, "\n Summary of events:\n\n");
3233
3234         return printed;
3235 }
3236
3237 DEFINE_RESORT_RB(syscall_stats, a->msecs > b->msecs,
3238         struct stats    *stats;
3239         double          msecs;
3240         int             syscall;
3241 )
3242 {
3243         struct int_node *source = rb_entry(nd, struct int_node, rb_node);
3244         struct stats *stats = source->priv;
3245
3246         entry->syscall = source->i;
3247         entry->stats   = stats;
3248         entry->msecs   = stats ? (u64)stats->n * (avg_stats(stats) / NSEC_PER_MSEC) : 0;
3249 }
3250
3251 static size_t thread__dump_stats(struct thread_trace *ttrace,
3252                                  struct trace *trace, FILE *fp)
3253 {
3254         size_t printed = 0;
3255         struct syscall *sc;
3256         struct rb_node *nd;
3257         DECLARE_RESORT_RB_INTLIST(syscall_stats, ttrace->syscall_stats);
3258
3259         if (syscall_stats == NULL)
3260                 return 0;
3261
3262         printed += fprintf(fp, "\n");
3263
3264         printed += fprintf(fp, "   syscall            calls    total       min       avg       max      stddev\n");
3265         printed += fprintf(fp, "                               (msec)    (msec)    (msec)    (msec)        (%%)\n");
3266         printed += fprintf(fp, "   --------------- -------- --------- --------- --------- ---------     ------\n");
3267
3268         resort_rb__for_each_entry(nd, syscall_stats) {
3269                 struct stats *stats = syscall_stats_entry->stats;
3270                 if (stats) {
3271                         double min = (double)(stats->min) / NSEC_PER_MSEC;
3272                         double max = (double)(stats->max) / NSEC_PER_MSEC;
3273                         double avg = avg_stats(stats);
3274                         double pct;
3275                         u64 n = (u64) stats->n;
3276
3277                         pct = avg ? 100.0 * stddev_stats(stats)/avg : 0.0;
3278                         avg /= NSEC_PER_MSEC;
3279
3280                         sc = &trace->syscalls.table[syscall_stats_entry->syscall];
3281                         printed += fprintf(fp, "   %-15s", sc->name);
3282                         printed += fprintf(fp, " %8" PRIu64 " %9.3f %9.3f %9.3f",
3283                                            n, syscall_stats_entry->msecs, min, avg);
3284                         printed += fprintf(fp, " %9.3f %9.2f%%\n", max, pct);
3285                 }
3286         }
3287
3288         resort_rb__delete(syscall_stats);
3289         printed += fprintf(fp, "\n\n");
3290
3291         return printed;
3292 }
3293
3294 static size_t trace__fprintf_thread(FILE *fp, struct thread *thread, struct trace *trace)
3295 {
3296         size_t printed = 0;
3297         struct thread_trace *ttrace = thread__priv(thread);
3298         double ratio;
3299
3300         if (ttrace == NULL)
3301                 return 0;
3302
3303         ratio = (double)ttrace->nr_events / trace->nr_events * 100.0;
3304
3305         printed += fprintf(fp, " %s (%d), ", thread__comm_str(thread), thread->tid);
3306         printed += fprintf(fp, "%lu events, ", ttrace->nr_events);
3307         printed += fprintf(fp, "%.1f%%", ratio);
3308         if (ttrace->pfmaj)
3309                 printed += fprintf(fp, ", %lu majfaults", ttrace->pfmaj);
3310         if (ttrace->pfmin)
3311                 printed += fprintf(fp, ", %lu minfaults", ttrace->pfmin);
3312         if (trace->sched)
3313                 printed += fprintf(fp, ", %.3f msec\n", ttrace->runtime_ms);
3314         else if (fputc('\n', fp) != EOF)
3315                 ++printed;
3316
3317         printed += thread__dump_stats(ttrace, trace, fp);
3318
3319         return printed;
3320 }
3321
3322 static unsigned long thread__nr_events(struct thread_trace *ttrace)
3323 {
3324         return ttrace ? ttrace->nr_events : 0;
3325 }
3326
3327 DEFINE_RESORT_RB(threads, (thread__nr_events(a->thread->priv) < thread__nr_events(b->thread->priv)),
3328         struct thread *thread;
3329 )
3330 {
3331         entry->thread = rb_entry(nd, struct thread, rb_node);
3332 }
3333
3334 static size_t trace__fprintf_thread_summary(struct trace *trace, FILE *fp)
3335 {
3336         size_t printed = trace__fprintf_threads_header(fp);
3337         struct rb_node *nd;
3338         int i;
3339
3340         for (i = 0; i < THREADS__TABLE_SIZE; i++) {
3341                 DECLARE_RESORT_RB_MACHINE_THREADS(threads, trace->host, i);
3342
3343                 if (threads == NULL) {
3344                         fprintf(fp, "%s", "Error sorting output by nr_events!\n");
3345                         return 0;
3346                 }
3347
3348                 resort_rb__for_each_entry(nd, threads)
3349                         printed += trace__fprintf_thread(fp, threads_entry->thread, trace);
3350
3351                 resort_rb__delete(threads);
3352         }
3353         return printed;
3354 }
3355
3356 static int trace__set_duration(const struct option *opt, const char *str,
3357                                int unset __maybe_unused)
3358 {
3359         struct trace *trace = opt->value;
3360
3361         trace->duration_filter = atof(str);
3362         return 0;
3363 }
3364
3365 static int trace__set_filter_pids_from_option(const struct option *opt, const char *str,
3366                                               int unset __maybe_unused)
3367 {
3368         int ret = -1;
3369         size_t i;
3370         struct trace *trace = opt->value;
3371         /*
3372          * FIXME: introduce a intarray class, plain parse csv and create a
3373          * { int nr, int entries[] } struct...
3374          */
3375         struct intlist *list = intlist__new(str);
3376
3377         if (list == NULL)
3378                 return -1;
3379
3380         i = trace->filter_pids.nr = intlist__nr_entries(list) + 1;
3381         trace->filter_pids.entries = calloc(i, sizeof(pid_t));
3382
3383         if (trace->filter_pids.entries == NULL)
3384                 goto out;
3385
3386         trace->filter_pids.entries[0] = getpid();
3387
3388         for (i = 1; i < trace->filter_pids.nr; ++i)
3389                 trace->filter_pids.entries[i] = intlist__entry(list, i - 1)->i;
3390
3391         intlist__delete(list);
3392         ret = 0;
3393 out:
3394         return ret;
3395 }
3396
3397 static int trace__open_output(struct trace *trace, const char *filename)
3398 {
3399         struct stat st;
3400
3401         if (!stat(filename, &st) && st.st_size) {
3402                 char oldname[PATH_MAX];
3403
3404                 scnprintf(oldname, sizeof(oldname), "%s.old", filename);
3405                 unlink(oldname);
3406                 rename(filename, oldname);
3407         }
3408
3409         trace->output = fopen(filename, "w");
3410
3411         return trace->output == NULL ? -errno : 0;
3412 }
3413
3414 static int parse_pagefaults(const struct option *opt, const char *str,
3415                             int unset __maybe_unused)
3416 {
3417         int *trace_pgfaults = opt->value;
3418
3419         if (strcmp(str, "all") == 0)
3420                 *trace_pgfaults |= TRACE_PFMAJ | TRACE_PFMIN;
3421         else if (strcmp(str, "maj") == 0)
3422                 *trace_pgfaults |= TRACE_PFMAJ;
3423         else if (strcmp(str, "min") == 0)
3424                 *trace_pgfaults |= TRACE_PFMIN;
3425         else
3426                 return -1;
3427
3428         return 0;
3429 }
3430
3431 static void evlist__set_evsel_handler(struct perf_evlist *evlist, void *handler)
3432 {
3433         struct perf_evsel *evsel;
3434
3435         evlist__for_each_entry(evlist, evsel)
3436                 evsel->handler = handler;
3437 }
3438
3439 static int evlist__set_syscall_tp_fields(struct perf_evlist *evlist)
3440 {
3441         struct perf_evsel *evsel;
3442
3443         evlist__for_each_entry(evlist, evsel) {
3444                 if (evsel->priv || !evsel->tp_format)
3445                         continue;
3446
3447                 if (strcmp(evsel->tp_format->system, "syscalls"))
3448                         continue;
3449
3450                 if (perf_evsel__init_syscall_tp(evsel))
3451                         return -1;
3452
3453                 if (!strncmp(evsel->tp_format->name, "sys_enter_", 10)) {
3454                         struct syscall_tp *sc = evsel->priv;
3455
3456                         if (__tp_field__init_ptr(&sc->args, sc->id.offset + sizeof(u64)))
3457                                 return -1;
3458                 } else if (!strncmp(evsel->tp_format->name, "sys_exit_", 9)) {
3459                         struct syscall_tp *sc = evsel->priv;
3460
3461                         if (__tp_field__init_uint(&sc->ret, sizeof(u64), sc->id.offset + sizeof(u64), evsel->needs_swap))
3462                                 return -1;
3463                 }
3464         }
3465
3466         return 0;
3467 }
3468
3469 /*
3470  * XXX: Hackish, just splitting the combined -e+--event (syscalls
3471  * (raw_syscalls:{sys_{enter,exit}} + events (tracepoints, HW, SW, etc) to use
3472  * existing facilities unchanged (trace->ev_qualifier + parse_options()).
3473  *
3474  * It'd be better to introduce a parse_options() variant that would return a
3475  * list with the terms it didn't match to an event...
3476  */
3477 static int trace__parse_events_option(const struct option *opt, const char *str,
3478                                       int unset __maybe_unused)
3479 {
3480         struct trace *trace = (struct trace *)opt->value;
3481         const char *s = str;
3482         char *sep = NULL, *lists[2] = { NULL, NULL, };
3483         int len = strlen(str) + 1, err = -1, list, idx;
3484         char *strace_groups_dir = system_path(STRACE_GROUPS_DIR);
3485         char group_name[PATH_MAX];
3486         struct syscall_fmt *fmt;
3487
3488         if (strace_groups_dir == NULL)
3489                 return -1;
3490
3491         if (*s == '!') {
3492                 ++s;
3493                 trace->not_ev_qualifier = true;
3494         }
3495
3496         while (1) {
3497                 if ((sep = strchr(s, ',')) != NULL)
3498                         *sep = '\0';
3499
3500                 list = 0;
3501                 if (syscalltbl__id(trace->sctbl, s) >= 0 ||
3502                     syscalltbl__strglobmatch_first(trace->sctbl, s, &idx) >= 0) {
3503                         list = 1;
3504                         goto do_concat;
3505                 }
3506
3507                 fmt = syscall_fmt__find_by_alias(s);
3508                 if (fmt != NULL) {
3509                         list = 1;
3510                         s = fmt->name;
3511                 } else {
3512                         path__join(group_name, sizeof(group_name), strace_groups_dir, s);
3513                         if (access(group_name, R_OK) == 0)
3514                                 list = 1;
3515                 }
3516 do_concat:
3517                 if (lists[list]) {
3518                         sprintf(lists[list] + strlen(lists[list]), ",%s", s);
3519                 } else {
3520                         lists[list] = malloc(len);
3521                         if (lists[list] == NULL)
3522                                 goto out;
3523                         strcpy(lists[list], s);
3524                 }
3525
3526                 if (!sep)
3527                         break;
3528
3529                 *sep = ',';
3530                 s = sep + 1;
3531         }
3532
3533         if (lists[1] != NULL) {
3534                 struct strlist_config slist_config = {
3535                         .dirname = strace_groups_dir,
3536                 };
3537
3538                 trace->ev_qualifier = strlist__new(lists[1], &slist_config);
3539                 if (trace->ev_qualifier == NULL) {
3540                         fputs("Not enough memory to parse event qualifier", trace->output);
3541                         goto out;
3542                 }
3543
3544                 if (trace__validate_ev_qualifier(trace))
3545                         goto out;
3546                 trace->trace_syscalls = true;
3547         }
3548
3549         err = 0;
3550
3551         if (lists[0]) {
3552                 struct option o = OPT_CALLBACK('e', "event", &trace->evlist, "event",
3553                                                "event selector. use 'perf list' to list available events",
3554                                                parse_events_option);
3555                 err = parse_events_option(&o, lists[0], 0);
3556         }
3557 out:
3558         if (sep)
3559                 *sep = ',';
3560
3561         return err;
3562 }
3563
3564 static int trace__parse_cgroups(const struct option *opt, const char *str, int unset)
3565 {
3566         struct trace *trace = opt->value;
3567
3568         if (!list_empty(&trace->evlist->entries))
3569                 return parse_cgroups(opt, str, unset);
3570
3571         trace->cgroup = evlist__findnew_cgroup(trace->evlist, str);
3572
3573         return 0;
3574 }
3575
3576 static struct bpf_map *bpf__find_map_by_name(const char *name)
3577 {
3578         struct bpf_object *obj, *tmp;
3579
3580         bpf_object__for_each_safe(obj, tmp) {
3581                 struct bpf_map *map = bpf_object__find_map_by_name(obj, name);
3582                 if (map)
3583                         return map;
3584
3585         }
3586
3587         return NULL;
3588 }
3589
3590 static void trace__set_bpf_map_filtered_pids(struct trace *trace)
3591 {
3592         trace->filter_pids.map = bpf__find_map_by_name("pids_filtered");
3593 }
3594
3595 static void trace__set_bpf_map_syscalls(struct trace *trace)
3596 {
3597         trace->syscalls.map = bpf__find_map_by_name("syscalls");
3598 }
3599
3600 static int trace__config(const char *var, const char *value, void *arg)
3601 {
3602         struct trace *trace = arg;
3603         int err = 0;
3604
3605         if (!strcmp(var, "trace.add_events")) {
3606                 struct option o = OPT_CALLBACK('e', "event", &trace->evlist, "event",
3607                                                "event selector. use 'perf list' to list available events",
3608                                                parse_events_option);
3609                 err = parse_events_option(&o, value, 0);
3610         } else if (!strcmp(var, "trace.show_timestamp")) {
3611                 trace->show_tstamp = perf_config_bool(var, value);
3612         } else if (!strcmp(var, "trace.show_duration")) {
3613                 trace->show_duration = perf_config_bool(var, value);
3614         } else if (!strcmp(var, "trace.show_arg_names")) {
3615                 trace->show_arg_names = perf_config_bool(var, value);
3616                 if (!trace->show_arg_names)
3617                         trace->show_zeros = true;
3618         } else if (!strcmp(var, "trace.show_zeros")) {
3619                 bool new_show_zeros = perf_config_bool(var, value);
3620                 if (!trace->show_arg_names && !new_show_zeros) {
3621                         pr_warning("trace.show_zeros has to be set when trace.show_arg_names=no\n");
3622                         goto out;
3623                 }
3624                 trace->show_zeros = new_show_zeros;
3625         } else if (!strcmp(var, "trace.show_prefix")) {
3626                 trace->show_string_prefix = perf_config_bool(var, value);
3627         } else if (!strcmp(var, "trace.no_inherit")) {
3628                 trace->opts.no_inherit = perf_config_bool(var, value);
3629         } else if (!strcmp(var, "trace.args_alignment")) {
3630                 int args_alignment = 0;
3631                 if (perf_config_int(&args_alignment, var, value) == 0)
3632                         trace->args_alignment = args_alignment;
3633         }
3634 out:
3635         return err;
3636 }
3637
3638 int cmd_trace(int argc, const char **argv)
3639 {
3640         const char *trace_usage[] = {
3641                 "perf trace [<options>] [<command>]",
3642                 "perf trace [<options>] -- <command> [<options>]",
3643                 "perf trace record [<options>] [<command>]",
3644                 "perf trace record [<options>] -- <command> [<options>]",
3645                 NULL
3646         };
3647         struct trace trace = {
3648                 .syscalls = {
3649                         . max = -1,
3650                 },
3651                 .opts = {
3652                         .target = {
3653                                 .uid       = UINT_MAX,
3654                                 .uses_mmap = true,
3655                         },
3656                         .user_freq     = UINT_MAX,
3657                         .user_interval = ULLONG_MAX,
3658                         .no_buffering  = true,
3659                         .mmap_pages    = UINT_MAX,
3660                 },
3661                 .output = stderr,
3662                 .show_comm = true,
3663                 .show_tstamp = true,
3664                 .show_duration = true,
3665                 .show_arg_names = true,
3666                 .args_alignment = 70,
3667                 .trace_syscalls = false,
3668                 .kernel_syscallchains = false,
3669                 .max_stack = UINT_MAX,
3670                 .max_events = ULONG_MAX,
3671         };
3672         const char *output_name = NULL;
3673         const struct option trace_options[] = {
3674         OPT_CALLBACK('e', "event", &trace, "event",
3675                      "event/syscall selector. use 'perf list' to list available events",
3676                      trace__parse_events_option),
3677         OPT_BOOLEAN(0, "comm", &trace.show_comm,
3678                     "show the thread COMM next to its id"),
3679         OPT_BOOLEAN(0, "tool_stats", &trace.show_tool_stats, "show tool stats"),
3680         OPT_CALLBACK(0, "expr", &trace, "expr", "list of syscalls/events to trace",
3681                      trace__parse_events_option),
3682         OPT_STRING('o', "output", &output_name, "file", "output file name"),
3683         OPT_STRING('i', "input", &input_name, "file", "Analyze events in file"),
3684         OPT_STRING('p', "pid", &trace.opts.target.pid, "pid",
3685                     "trace events on existing process id"),
3686         OPT_STRING('t', "tid", &trace.opts.target.tid, "tid",
3687                     "trace events on existing thread id"),
3688         OPT_CALLBACK(0, "filter-pids", &trace, "CSV list of pids",
3689                      "pids to filter (by the kernel)", trace__set_filter_pids_from_option),
3690         OPT_BOOLEAN('a', "all-cpus", &trace.opts.target.system_wide,
3691                     "system-wide collection from all CPUs"),
3692         OPT_STRING('C', "cpu", &trace.opts.target.cpu_list, "cpu",
3693                     "list of cpus to monitor"),
3694         OPT_BOOLEAN(0, "no-inherit", &trace.opts.no_inherit,
3695                     "child tasks do not inherit counters"),
3696         OPT_CALLBACK('m', "mmap-pages", &trace.opts.mmap_pages, "pages",
3697                      "number of mmap data pages",
3698                      perf_evlist__parse_mmap_pages),
3699         OPT_STRING('u', "uid", &trace.opts.target.uid_str, "user",
3700                    "user to profile"),
3701         OPT_CALLBACK(0, "duration", &trace, "float",
3702                      "show only events with duration > N.M ms",
3703                      trace__set_duration),
3704         OPT_BOOLEAN(0, "sched", &trace.sched, "show blocking scheduler events"),
3705         OPT_INCR('v', "verbose", &verbose, "be more verbose"),
3706         OPT_BOOLEAN('T', "time", &trace.full_time,
3707                     "Show full timestamp, not time relative to first start"),
3708         OPT_BOOLEAN(0, "failure", &trace.failure_only,
3709                     "Show only syscalls that failed"),
3710         OPT_BOOLEAN('s', "summary", &trace.summary_only,
3711                     "Show only syscall summary with statistics"),
3712         OPT_BOOLEAN('S', "with-summary", &trace.summary,
3713                     "Show all syscalls and summary with statistics"),
3714         OPT_CALLBACK_DEFAULT('F', "pf", &trace.trace_pgfaults, "all|maj|min",
3715                      "Trace pagefaults", parse_pagefaults, "maj"),
3716         OPT_BOOLEAN(0, "syscalls", &trace.trace_syscalls, "Trace syscalls"),
3717         OPT_BOOLEAN('f', "force", &trace.force, "don't complain, do it"),
3718         OPT_CALLBACK(0, "call-graph", &trace.opts,
3719                      "record_mode[,record_size]", record_callchain_help,
3720                      &record_parse_callchain_opt),
3721         OPT_BOOLEAN(0, "kernel-syscall-graph", &trace.kernel_syscallchains,
3722                     "Show the kernel callchains on the syscall exit path"),
3723         OPT_ULONG(0, "max-events", &trace.max_events,
3724                 "Set the maximum number of events to print, exit after that is reached. "),
3725         OPT_UINTEGER(0, "min-stack", &trace.min_stack,
3726                      "Set the minimum stack depth when parsing the callchain, "
3727                      "anything below the specified depth will be ignored."),
3728         OPT_UINTEGER(0, "max-stack", &trace.max_stack,
3729                      "Set the maximum stack depth when parsing the callchain, "
3730                      "anything beyond the specified depth will be ignored. "
3731                      "Default: kernel.perf_event_max_stack or " __stringify(PERF_MAX_STACK_DEPTH)),
3732         OPT_BOOLEAN(0, "sort-events", &trace.sort_events,
3733                         "Sort batch of events before processing, use if getting out of order events"),
3734         OPT_BOOLEAN(0, "print-sample", &trace.print_sample,
3735                         "print the PERF_RECORD_SAMPLE PERF_SAMPLE_ info, for debugging"),
3736         OPT_UINTEGER(0, "proc-map-timeout", &proc_map_timeout,
3737                         "per thread proc mmap processing timeout in ms"),
3738         OPT_CALLBACK('G', "cgroup", &trace, "name", "monitor event in cgroup name only",
3739                      trace__parse_cgroups),
3740         OPT_UINTEGER('D', "delay", &trace.opts.initial_delay,
3741                      "ms to wait before starting measurement after program "
3742                      "start"),
3743         OPT_END()
3744         };
3745         bool __maybe_unused max_stack_user_set = true;
3746         bool mmap_pages_user_set = true;
3747         struct perf_evsel *evsel;
3748         const char * const trace_subcommands[] = { "record", NULL };
3749         int err = -1;
3750         char bf[BUFSIZ];
3751
3752         signal(SIGSEGV, sighandler_dump_stack);
3753         signal(SIGFPE, sighandler_dump_stack);
3754
3755         trace.evlist = perf_evlist__new();
3756         trace.sctbl = syscalltbl__new();
3757
3758         if (trace.evlist == NULL || trace.sctbl == NULL) {
3759                 pr_err("Not enough memory to run!\n");
3760                 err = -ENOMEM;
3761                 goto out;
3762         }
3763
3764         err = perf_config(trace__config, &trace);
3765         if (err)
3766                 goto out;
3767
3768         argc = parse_options_subcommand(argc, argv, trace_options, trace_subcommands,
3769                                  trace_usage, PARSE_OPT_STOP_AT_NON_OPTION);
3770
3771         if ((nr_cgroups || trace.cgroup) && !trace.opts.target.system_wide) {
3772                 usage_with_options_msg(trace_usage, trace_options,
3773                                        "cgroup monitoring only available in system-wide mode");
3774         }
3775
3776         evsel = bpf__setup_output_event(trace.evlist, "__augmented_syscalls__");
3777         if (IS_ERR(evsel)) {
3778                 bpf__strerror_setup_output_event(trace.evlist, PTR_ERR(evsel), bf, sizeof(bf));
3779                 pr_err("ERROR: Setup trace syscalls enter failed: %s\n", bf);
3780                 goto out;
3781         }
3782
3783         if (evsel) {
3784                 trace.syscalls.events.augmented = evsel;
3785                 trace__set_bpf_map_filtered_pids(&trace);
3786                 trace__set_bpf_map_syscalls(&trace);
3787         }
3788
3789         err = bpf__setup_stdout(trace.evlist);
3790         if (err) {
3791                 bpf__strerror_setup_stdout(trace.evlist, err, bf, sizeof(bf));
3792                 pr_err("ERROR: Setup BPF stdout failed: %s\n", bf);
3793                 goto out;
3794         }
3795
3796         err = -1;
3797
3798         if (trace.trace_pgfaults) {
3799                 trace.opts.sample_address = true;
3800                 trace.opts.sample_time = true;
3801         }
3802
3803         if (trace.opts.mmap_pages == UINT_MAX)
3804                 mmap_pages_user_set = false;
3805
3806         if (trace.max_stack == UINT_MAX) {
3807                 trace.max_stack = input_name ? PERF_MAX_STACK_DEPTH : sysctl__max_stack();
3808                 max_stack_user_set = false;
3809         }
3810
3811 #ifdef HAVE_DWARF_UNWIND_SUPPORT
3812         if ((trace.min_stack || max_stack_user_set) && !callchain_param.enabled) {
3813                 record_opts__parse_callchain(&trace.opts, &callchain_param, "dwarf", false);
3814         }
3815 #endif
3816
3817         if (callchain_param.enabled) {
3818                 if (!mmap_pages_user_set && geteuid() == 0)
3819                         trace.opts.mmap_pages = perf_event_mlock_kb_in_pages() * 4;
3820
3821                 symbol_conf.use_callchain = true;
3822         }
3823
3824         if (trace.evlist->nr_entries > 0) {
3825                 evlist__set_evsel_handler(trace.evlist, trace__event_handler);
3826                 if (evlist__set_syscall_tp_fields(trace.evlist)) {
3827                         perror("failed to set syscalls:* tracepoint fields");
3828                         goto out;
3829                 }
3830         }
3831
3832         if (trace.sort_events) {
3833                 ordered_events__init(&trace.oe.data, ordered_events__deliver_event, &trace);
3834                 ordered_events__set_copy_on_queue(&trace.oe.data, true);
3835         }
3836
3837         /*
3838          * If we are augmenting syscalls, then combine what we put in the
3839          * __augmented_syscalls__ BPF map with what is in the
3840          * syscalls:sys_exit_FOO tracepoints, i.e. just like we do without BPF,
3841          * combining raw_syscalls:sys_enter with raw_syscalls:sys_exit.
3842          *
3843          * We'll switch to look at two BPF maps, one for sys_enter and the
3844          * other for sys_exit when we start augmenting the sys_exit paths with
3845          * buffers that are being copied from kernel to userspace, think 'read'
3846          * syscall.
3847          */
3848         if (trace.syscalls.events.augmented) {
3849                 evlist__for_each_entry(trace.evlist, evsel) {
3850                         bool raw_syscalls_sys_exit = strcmp(perf_evsel__name(evsel), "raw_syscalls:sys_exit") == 0;
3851
3852                         if (raw_syscalls_sys_exit) {
3853                                 trace.raw_augmented_syscalls = true;
3854                                 goto init_augmented_syscall_tp;
3855                         }
3856
3857                         if (strcmp(perf_evsel__name(evsel), "raw_syscalls:sys_enter") == 0) {
3858                                 struct perf_evsel *augmented = trace.syscalls.events.augmented;
3859                                 if (perf_evsel__init_augmented_syscall_tp(augmented, evsel) ||
3860                                     perf_evsel__init_augmented_syscall_tp_args(augmented))
3861                                         goto out;
3862                                 augmented->handler = trace__sys_enter;
3863                         }
3864
3865                         if (strstarts(perf_evsel__name(evsel), "syscalls:sys_exit_")) {
3866                                 struct syscall_tp *sc;
3867 init_augmented_syscall_tp:
3868                                 if (perf_evsel__init_augmented_syscall_tp(evsel, evsel))
3869                                         goto out;
3870                                 sc = evsel->priv;
3871                                 /*
3872                                  * For now with BPF raw_augmented we hook into
3873                                  * raw_syscalls:sys_enter and there we get all
3874                                  * 6 syscall args plus the tracepoint common
3875                                  * fields and the syscall_nr (another long).
3876                                  * So we check if that is the case and if so
3877                                  * don't look after the sc->args_size but
3878                                  * always after the full raw_syscalls:sys_enter
3879                                  * payload, which is fixed.
3880                                  *
3881                                  * We'll revisit this later to pass
3882                                  * s->args_size to the BPF augmenter (now
3883                                  * tools/perf/examples/bpf/augmented_raw_syscalls.c,
3884                                  * so that it copies only what we need for each
3885                                  * syscall, like what happens when we use
3886                                  * syscalls:sys_enter_NAME, so that we reduce
3887                                  * the kernel/userspace traffic to just what is
3888                                  * needed for each syscall.
3889                                  */
3890                                 if (trace.raw_augmented_syscalls)
3891                                         trace.raw_augmented_syscalls_args_size = (6 + 1) * sizeof(long) + sc->id.offset;
3892                                 perf_evsel__init_augmented_syscall_tp_ret(evsel);
3893                                 evsel->handler = trace__sys_exit;
3894                         }
3895                 }
3896         }
3897
3898         if ((argc >= 1) && (strcmp(argv[0], "record") == 0))
3899                 return trace__record(&trace, argc-1, &argv[1]);
3900
3901         /* summary_only implies summary option, but don't overwrite summary if set */
3902         if (trace.summary_only)
3903                 trace.summary = trace.summary_only;
3904
3905         if (!trace.trace_syscalls && !trace.trace_pgfaults &&
3906             trace.evlist->nr_entries == 0 /* Was --events used? */) {
3907                 trace.trace_syscalls = true;
3908         }
3909
3910         if (output_name != NULL) {
3911                 err = trace__open_output(&trace, output_name);
3912                 if (err < 0) {
3913                         perror("failed to create output file");
3914                         goto out;
3915                 }
3916         }
3917
3918         err = target__validate(&trace.opts.target);
3919         if (err) {
3920                 target__strerror(&trace.opts.target, err, bf, sizeof(bf));
3921                 fprintf(trace.output, "%s", bf);
3922                 goto out_close;
3923         }
3924
3925         err = target__parse_uid(&trace.opts.target);
3926         if (err) {
3927                 target__strerror(&trace.opts.target, err, bf, sizeof(bf));
3928                 fprintf(trace.output, "%s", bf);
3929                 goto out_close;
3930         }
3931
3932         if (!argc && target__none(&trace.opts.target))
3933                 trace.opts.target.system_wide = true;
3934
3935         if (input_name)
3936                 err = trace__replay(&trace);
3937         else
3938                 err = trace__run(&trace, argc, argv);
3939
3940 out_close:
3941         if (output_name != NULL)
3942                 fclose(trace.output);
3943 out:
3944         return err;
3945 }