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
  17 #include "util/record.h"
  18 #include <traceevent/event-parse.h>
  19 #include <api/fs/tracing_path.h>
  20 #include <bpf/bpf.h>
  21 #include "util/bpf_map.h"
  22 #include "util/rlimit.h"
  23 #include "builtin.h"
  24 #include "util/cgroup.h"
  25 #include "util/color.h"
  26 #include "util/config.h"
  27 #include "util/debug.h"
  28 #include "util/dso.h"
  29 #include "util/env.h"
  30 #include "util/event.h"
  31 #include "util/synthetic-events.h"
  32 #include "util/evlist.h"
  33 #include "util/evswitch.h"
  34 #include <subcmd/pager.h>
  35 #include <subcmd/exec-cmd.h>
  36 #include "util/machine.h"
  37 #include "util/map.h"
  38 #include "util/symbol.h"
  39 #include "util/path.h"
  40 #include "util/session.h"
  41 #include "util/thread.h"
  42 #include <subcmd/parse-options.h>
  43 #include "util/strlist.h"
  44 #include "util/intlist.h"
  45 #include "util/thread_map.h"
  46 #include "util/stat.h"
  47 #include "util/tool.h"
  48 #include "util/util.h"
  49 #include "trace/beauty/beauty.h"
  50 #include "trace-event.h"
  51 #include "util/parse-events.h"
  52 #include "util/bpf-loader.h"
  53 #include "callchain.h"
  54 #include "print_binary.h"
  55 #include "string2.h"
  56 #include "syscalltbl.h"
  57 #include "rb_resort.h"
  58 #include "../perf.h"
  59
  60 #include <errno.h>
  61 #include <inttypes.h>
  62 #include <poll.h>
  63 #include <signal.h>
  64 #include <stdlib.h>
  65 #include <string.h>
  66 #include <linux/err.h>
  67 #include <linux/filter.h>
  68 #include <linux/kernel.h>
  69 #include <linux/random.h>
  70 #include <linux/stringify.h>
  71 #include <linux/time64.h>
  72 #include <linux/zalloc.h>
  73 #include <fcntl.h>
  74 #include <sys/sysmacros.h>
  75
  76 #include <linux/ctype.h>
  77
  78 #ifndef O_CLOEXEC
  79 # define O_CLOEXEC              02000000
  80 #endif
  81
  82 #ifndef F_LINUX_SPECIFIC_BASE
  83 # define F_LINUX_SPECIFIC_BASE  1024
  84 #endif
  85
  86 struct trace {
  87         struct perf_tool        tool;
  88         struct syscalltbl       *sctbl;
  89         struct {
  90                 struct syscall  *table;
  91                 struct bpf_map  *map;
  92                 struct { // per syscall BPF_MAP_TYPE_PROG_ARRAY
  93                         struct bpf_map  *sys_enter,
  94                                         *sys_exit;
  95                 }               prog_array;
  96                 struct {
  97                         struct evsel *sys_enter,
  98                                           *sys_exit,
  99                                           *augmented;
 100                 }               events;
 101                 struct bpf_program *unaugmented_prog;
 102         } syscalls;
 103         struct {
 104                 struct bpf_map *map;
 105         } dump;
 106         struct record_opts      opts;
 107         struct evlist   *evlist;
 108         struct machine          *host;
 109         struct thread           *current;
 110         struct bpf_object       *bpf_obj;
 111         struct cgroup           *cgroup;
 112         u64                     base_time;
 113         FILE                    *output;
 114         unsigned long           nr_events;
 115         unsigned long           nr_events_printed;
 116         unsigned long           max_events;
 117         struct evswitch         evswitch;
 118         struct strlist          *ev_qualifier;
 119         struct {
 120                 size_t          nr;
 121                 int             *entries;
 122         }                       ev_qualifier_ids;
 123         struct {
 124                 size_t          nr;
 125                 pid_t           *entries;
 126                 struct bpf_map  *map;
 127         }                       filter_pids;
 128         double                  duration_filter;
 129         double                  runtime_ms;
 130         struct {
 131                 u64             vfs_getname,
 132                                 proc_getname;
 133         } stats;
 134         unsigned int            max_stack;
 135         unsigned int            min_stack;
 136         int                     raw_augmented_syscalls_args_size;
 137         bool                    raw_augmented_syscalls;
 138         bool                    fd_path_disabled;
 139         bool                    sort_events;
 140         bool                    not_ev_qualifier;
 141         bool                    live;
 142         bool                    full_time;
 143         bool                    sched;
 144         bool                    multiple_threads;
 145         bool                    summary;
 146         bool                    summary_only;
 147         bool                    failure_only;
 148         bool                    show_comm;
 149         bool                    print_sample;
 150         bool                    show_tool_stats;
 151         bool                    trace_syscalls;
 152         bool                    kernel_syscallchains;
 153         s16                     args_alignment;
 154         bool                    show_tstamp;
 155         bool                    show_duration;
 156         bool                    show_zeros;
 157         bool                    show_arg_names;
 158         bool                    show_string_prefix;
 159         bool                    force;
 160         bool                    vfs_getname;
 161         int                     trace_pgfaults;
 162         struct {
 163                 struct ordered_events   data;
 164                 u64                     last;
 165         } oe;
 166 };
 167
 168 struct tp_field {
 169         int offset;
 170         union {
 171                 u64 (*integer)(struct tp_field *field, struct perf_sample *sample);
 172                 void *(*pointer)(struct tp_field *field, struct perf_sample *sample);
 173         };
 174 };
 175
 176 #define TP_UINT_FIELD(bits) \
 177 static u64 tp_field__u##bits(struct tp_field *field, struct perf_sample *sample) \
 178 { \
 179         u##bits value; \
 180         memcpy(&value, sample->raw_data + field->offset, sizeof(value)); \
 181         return value;  \
 182 }
 183
 184 TP_UINT_FIELD(8);
 185 TP_UINT_FIELD(16);
 186 TP_UINT_FIELD(32);
 187 TP_UINT_FIELD(64);
 188
 189 #define TP_UINT_FIELD__SWAPPED(bits) \
 190 static u64 tp_field__swapped_u##bits(struct tp_field *field, struct perf_sample *sample) \
 191 { \
 192         u##bits value; \
 193         memcpy(&value, sample->raw_data + field->offset, sizeof(value)); \
 194         return bswap_##bits(value);\
 195 }
 196
 197 TP_UINT_FIELD__SWAPPED(16);
 198 TP_UINT_FIELD__SWAPPED(32);
 199 TP_UINT_FIELD__SWAPPED(64);
 200
 201 static int __tp_field__init_uint(struct tp_field *field, int size, int offset, bool needs_swap)
 202 {
 203         field->offset = offset;
 204
 205         switch (size) {
 206         case 1:
 207                 field->integer = tp_field__u8;
 208                 break;
 209         case 2:
 210                 field->integer = needs_swap ? tp_field__swapped_u16 : tp_field__u16;
 211                 break;
 212         case 4:
 213                 field->integer = needs_swap ? tp_field__swapped_u32 : tp_field__u32;
 214                 break;
 215         case 8:
 216                 field->integer = needs_swap ? tp_field__swapped_u64 : tp_field__u64;
 217                 break;
 218         default:
 219                 return -1;
 220         }
 221
 222         return 0;
 223 }
 224
 225 static int tp_field__init_uint(struct tp_field *field, struct tep_format_field *format_field, bool needs_swap)
 226 {
 227         return __tp_field__init_uint(field, format_field->size, format_field->offset, needs_swap);
 228 }
 229
 230 static void *tp_field__ptr(struct tp_field *field, struct perf_sample *sample)
 231 {
 232         return sample->raw_data + field->offset;
 233 }
 234
 235 static int __tp_field__init_ptr(struct tp_field *field, int offset)
 236 {
 237         field->offset = offset;
 238         field->pointer = tp_field__ptr;
 239         return 0;
 240 }
 241
 242 static int tp_field__init_ptr(struct tp_field *field, struct tep_format_field *format_field)
 243 {
 244         return __tp_field__init_ptr(field, format_field->offset);
 245 }
 246
 247 struct syscall_tp {
 248         struct tp_field id;
 249         union {
 250                 struct tp_field args, ret;
 251         };
 252 };
 253
 254 static int perf_evsel__init_tp_uint_field(struct evsel *evsel,
 255                                           struct tp_field *field,
 256                                           const char *name)
 257 {
 258         struct tep_format_field *format_field = perf_evsel__field(evsel, name);
 259
 260         if (format_field == NULL)
 261                 return -1;
 262
 263         return tp_field__init_uint(field, format_field, evsel->needs_swap);
 264 }
 265
 266 #define perf_evsel__init_sc_tp_uint_field(evsel, name) \
 267         ({ struct syscall_tp *sc = evsel->priv;\
 268            perf_evsel__init_tp_uint_field(evsel, &sc->name, #name); })
 269
 270 static int perf_evsel__init_tp_ptr_field(struct evsel *evsel,
 271                                          struct tp_field *field,
 272                                          const char *name)
 273 {
 274         struct tep_format_field *format_field = perf_evsel__field(evsel, name);
 275
 276         if (format_field == NULL)
 277                 return -1;
 278
 279         return tp_field__init_ptr(field, format_field);
 280 }
 281
 282 #define perf_evsel__init_sc_tp_ptr_field(evsel, name) \
 283         ({ struct syscall_tp *sc = evsel->priv;\
 284            perf_evsel__init_tp_ptr_field(evsel, &sc->name, #name); })
 285
 286 static void evsel__delete_priv(struct evsel *evsel)
 287 {
 288         zfree(&evsel->priv);
 289         evsel__delete(evsel);
 290 }
 291
 292 static int perf_evsel__init_syscall_tp(struct evsel *evsel)
 293 {
 294         struct syscall_tp *sc = evsel->priv = malloc(sizeof(struct syscall_tp));
 295
 296         if (evsel->priv != NULL) {
 297                 if (perf_evsel__init_tp_uint_field(evsel, &sc->id, "__syscall_nr") &&
 298                     perf_evsel__init_tp_uint_field(evsel, &sc->id, "nr"))
 299                         goto out_delete;
 300                 return 0;
 301         }
 302
 303         return -ENOMEM;
 304 out_delete:
 305         zfree(&evsel->priv);
 306         return -ENOENT;
 307 }
 308
 309 static int perf_evsel__init_augmented_syscall_tp(struct evsel *evsel, struct evsel *tp)
 310 {
 311         struct syscall_tp *sc = evsel->priv = malloc(sizeof(struct syscall_tp));
 312
 313         if (evsel->priv != NULL) {
 314                 struct tep_format_field *syscall_id = perf_evsel__field(tp, "id");
 315                 if (syscall_id == NULL)
 316                         syscall_id = perf_evsel__field(tp, "__syscall_nr");
 317                 if (syscall_id == NULL)
 318                         goto out_delete;
 319                 if (__tp_field__init_uint(&sc->id, syscall_id->size, syscall_id->offset, evsel->needs_swap))
 320                         goto out_delete;
 321
 322                 return 0;
 323         }
 324
 325         return -ENOMEM;
 326 out_delete:
 327         zfree(&evsel->priv);
 328         return -EINVAL;
 329 }
 330
 331 static int perf_evsel__init_augmented_syscall_tp_args(struct evsel *evsel)
 332 {
 333         struct syscall_tp *sc = evsel->priv;
 334
 335         return __tp_field__init_ptr(&sc->args, sc->id.offset + sizeof(u64));
 336 }
 337
 338 static int perf_evsel__init_augmented_syscall_tp_ret(struct evsel *evsel)
 339 {
 340         struct syscall_tp *sc = evsel->priv;
 341
 342         return __tp_field__init_uint(&sc->ret, sizeof(u64), sc->id.offset + sizeof(u64), evsel->needs_swap);
 343 }
 344
 345 static int perf_evsel__init_raw_syscall_tp(struct evsel *evsel, void *handler)
 346 {
 347         evsel->priv = malloc(sizeof(struct syscall_tp));
 348         if (evsel->priv != NULL) {
 349                 if (perf_evsel__init_sc_tp_uint_field(evsel, id))
 350                         goto out_delete;
 351
 352                 evsel->handler = handler;
 353                 return 0;
 354         }
 355
 356         return -ENOMEM;
 357
 358 out_delete:
 359         zfree(&evsel->priv);
 360         return -ENOENT;
 361 }
 362
 363 static struct evsel *perf_evsel__raw_syscall_newtp(const char *direction, void *handler)
 364 {
 365         struct evsel *evsel = perf_evsel__newtp("raw_syscalls", direction);
 366
 367         /* older kernel (e.g., RHEL6) use syscalls:{enter,exit} */
 368         if (IS_ERR(evsel))
 369                 evsel = perf_evsel__newtp("syscalls", direction);
 370
 371         if (IS_ERR(evsel))
 372                 return NULL;
 373
 374         if (perf_evsel__init_raw_syscall_tp(evsel, handler))
 375                 goto out_delete;
 376
 377         return evsel;
 378
 379 out_delete:
 380         evsel__delete_priv(evsel);
 381         return NULL;
 382 }
 383
 384 #define perf_evsel__sc_tp_uint(evsel, name, sample) \
 385         ({ struct syscall_tp *fields = evsel->priv; \
 386            fields->name.integer(&fields->name, sample); })
 387
 388 #define perf_evsel__sc_tp_ptr(evsel, name, sample) \
 389         ({ struct syscall_tp *fields = evsel->priv; \
 390            fields->name.pointer(&fields->name, sample); })
 391
 392 size_t strarray__scnprintf(struct strarray *sa, char *bf, size_t size, const char *intfmt, bool show_prefix, int val)
 393 {
 394         int idx = val - sa->offset;
 395
 396         if (idx < 0 || idx >= sa->nr_entries || sa->entries[idx] == NULL) {
 397                 size_t printed = scnprintf(bf, size, intfmt, val);
 398                 if (show_prefix)
 399                         printed += scnprintf(bf + printed, size - printed, " /* %s??? */", sa->prefix);
 400                 return printed;
 401         }
 402
 403         return scnprintf(bf, size, "%s%s", show_prefix ? sa->prefix : "", sa->entries[idx]);
 404 }
 405
 406 static size_t __syscall_arg__scnprintf_strarray(char *bf, size_t size,
 407                                                 const char *intfmt,
 408                                                 struct syscall_arg *arg)
 409 {
 410         return strarray__scnprintf(arg->parm, bf, size, intfmt, arg->show_string_prefix, arg->val);
 411 }
 412
 413 static size_t syscall_arg__scnprintf_strarray(char *bf, size_t size,
 414                                               struct syscall_arg *arg)
 415 {
 416         return __syscall_arg__scnprintf_strarray(bf, size, "%d", arg);
 417 }
 418
 419 #define SCA_STRARRAY syscall_arg__scnprintf_strarray
 420
 421 size_t syscall_arg__scnprintf_strarray_flags(char *bf, size_t size, struct syscall_arg *arg)
 422 {
 423         return strarray__scnprintf_flags(arg->parm, bf, size, arg->show_string_prefix, arg->val);
 424 }
 425
 426 size_t strarrays__scnprintf(struct strarrays *sas, char *bf, size_t size, const char *intfmt, bool show_prefix, int val)
 427 {
 428         size_t printed;
 429         int i;
 430
 431         for (i = 0; i < sas->nr_entries; ++i) {
 432                 struct strarray *sa = sas->entries[i];
 433                 int idx = val - sa->offset;
 434
 435                 if (idx >= 0 && idx < sa->nr_entries) {
 436                         if (sa->entries[idx] == NULL)
 437                                 break;
 438                         return scnprintf(bf, size, "%s%s", show_prefix ? sa->prefix : "", sa->entries[idx]);
 439                 }
 440         }
 441
 442         printed = scnprintf(bf, size, intfmt, val);
 443         if (show_prefix)
 444                 printed += scnprintf(bf + printed, size - printed, " /* %s??? */", sas->entries[0]->prefix);
 445         return printed;
 446 }
 447
 448 size_t syscall_arg__scnprintf_strarrays(char *bf, size_t size,
 449                                         struct syscall_arg *arg)
 450 {
 451         return strarrays__scnprintf(arg->parm, bf, size, "%d", arg->show_string_prefix, arg->val);
 452 }
 453
 454 #ifndef AT_FDCWD
 455 #define AT_FDCWD        -100
 456 #endif
 457
 458 static size_t syscall_arg__scnprintf_fd_at(char *bf, size_t size,
 459                                            struct syscall_arg *arg)
 460 {
 461         int fd = arg->val;
 462         const char *prefix = "AT_FD";
 463
 464         if (fd == AT_FDCWD)
 465                 return scnprintf(bf, size, "%s%s", arg->show_string_prefix ? prefix : "", "CWD");
 466
 467         return syscall_arg__scnprintf_fd(bf, size, arg);
 468 }
 469
 470 #define SCA_FDAT syscall_arg__scnprintf_fd_at
 471
 472 static size_t syscall_arg__scnprintf_close_fd(char *bf, size_t size,
 473                                               struct syscall_arg *arg);
 474
 475 #define SCA_CLOSE_FD syscall_arg__scnprintf_close_fd
 476
 477 size_t syscall_arg__scnprintf_hex(char *bf, size_t size, struct syscall_arg *arg)
 478 {
 479         return scnprintf(bf, size, "%#lx", arg->val);
 480 }
 481
 482 size_t syscall_arg__scnprintf_ptr(char *bf, size_t size, struct syscall_arg *arg)
 483 {
 484         if (arg->val == 0)
 485                 return scnprintf(bf, size, "NULL");
 486         return syscall_arg__scnprintf_hex(bf, size, arg);
 487 }
 488
 489 size_t syscall_arg__scnprintf_int(char *bf, size_t size, struct syscall_arg *arg)
 490 {
 491         return scnprintf(bf, size, "%d", arg->val);
 492 }
 493
 494 size_t syscall_arg__scnprintf_long(char *bf, size_t size, struct syscall_arg *arg)
 495 {
 496         return scnprintf(bf, size, "%ld", arg->val);
 497 }
 498
 499 static const char *bpf_cmd[] = {
 500         "MAP_CREATE", "MAP_LOOKUP_ELEM", "MAP_UPDATE_ELEM", "MAP_DELETE_ELEM",
 501         "MAP_GET_NEXT_KEY", "PROG_LOAD",
 502 };
 503 static DEFINE_STRARRAY(bpf_cmd, "BPF_");
 504
 505 static const char *fsmount_flags[] = {
 506         [1] = "CLOEXEC",
 507 };
 508 static DEFINE_STRARRAY(fsmount_flags, "FSMOUNT_");
 509
 510 #include "trace/beauty/generated/fsconfig_arrays.c"
 511
 512 static DEFINE_STRARRAY(fsconfig_cmds, "FSCONFIG_");
 513
 514 static const char *epoll_ctl_ops[] = { "ADD", "DEL", "MOD", };
 515 static DEFINE_STRARRAY_OFFSET(epoll_ctl_ops, "EPOLL_CTL_", 1);
 516
 517 static const char *itimers[] = { "REAL", "VIRTUAL", "PROF", };
 518 static DEFINE_STRARRAY(itimers, "ITIMER_");
 519
 520 static const char *keyctl_options[] = {
 521         "GET_KEYRING_ID", "JOIN_SESSION_KEYRING", "UPDATE", "REVOKE", "CHOWN",
 522         "SETPERM", "DESCRIBE", "CLEAR", "LINK", "UNLINK", "SEARCH", "READ",
 523         "INSTANTIATE", "NEGATE", "SET_REQKEY_KEYRING", "SET_TIMEOUT",
 524         "ASSUME_AUTHORITY", "GET_SECURITY", "SESSION_TO_PARENT", "REJECT",
 525         "INSTANTIATE_IOV", "INVALIDATE", "GET_PERSISTENT",
 526 };
 527 static DEFINE_STRARRAY(keyctl_options, "KEYCTL_");
 528
 529 static const char *whences[] = { "SET", "CUR", "END",
 530 #ifdef SEEK_DATA
 531 "DATA",
 532 #endif
 533 #ifdef SEEK_HOLE
 534 "HOLE",
 535 #endif
 536 };
 537 static DEFINE_STRARRAY(whences, "SEEK_");
 538
 539 static const char *fcntl_cmds[] = {
 540         "DUPFD", "GETFD", "SETFD", "GETFL", "SETFL", "GETLK", "SETLK",
 541         "SETLKW", "SETOWN", "GETOWN", "SETSIG", "GETSIG", "GETLK64",
 542         "SETLK64", "SETLKW64", "SETOWN_EX", "GETOWN_EX",
 543         "GETOWNER_UIDS",
 544 };
 545 static DEFINE_STRARRAY(fcntl_cmds, "F_");
 546
 547 static const char *fcntl_linux_specific_cmds[] = {
 548         "SETLEASE", "GETLEASE", "NOTIFY", [5] = "CANCELLK", "DUPFD_CLOEXEC",
 549         "SETPIPE_SZ", "GETPIPE_SZ", "ADD_SEALS", "GET_SEALS",
 550         "GET_RW_HINT", "SET_RW_HINT", "GET_FILE_RW_HINT", "SET_FILE_RW_HINT",
 551 };
 552
 553 static DEFINE_STRARRAY_OFFSET(fcntl_linux_specific_cmds, "F_", F_LINUX_SPECIFIC_BASE);
 554
 555 static struct strarray *fcntl_cmds_arrays[] = {
 556         &strarray__fcntl_cmds,
 557         &strarray__fcntl_linux_specific_cmds,
 558 };
 559
 560 static DEFINE_STRARRAYS(fcntl_cmds_arrays);
 561
 562 static const char *rlimit_resources[] = {
 563         "CPU", "FSIZE", "DATA", "STACK", "CORE", "RSS", "NPROC", "NOFILE",
 564         "MEMLOCK", "AS", "LOCKS", "SIGPENDING", "MSGQUEUE", "NICE", "RTPRIO",
 565         "RTTIME",
 566 };
 567 static DEFINE_STRARRAY(rlimit_resources, "RLIMIT_");
 568
 569 static const char *sighow[] = { "BLOCK", "UNBLOCK", "SETMASK", };
 570 static DEFINE_STRARRAY(sighow, "SIG_");
 571
 572 static const char *clockid[] = {
 573         "REALTIME", "MONOTONIC", "PROCESS_CPUTIME_ID", "THREAD_CPUTIME_ID",
 574         "MONOTONIC_RAW", "REALTIME_COARSE", "MONOTONIC_COARSE", "BOOTTIME",
 575         "REALTIME_ALARM", "BOOTTIME_ALARM", "SGI_CYCLE", "TAI"
 576 };
 577 static DEFINE_STRARRAY(clockid, "CLOCK_");
 578
 579 static size_t syscall_arg__scnprintf_access_mode(char *bf, size_t size,
 580                                                  struct syscall_arg *arg)
 581 {
 582         bool show_prefix = arg->show_string_prefix;
 583         const char *suffix = "_OK";
 584         size_t printed = 0;
 585         int mode = arg->val;
 586
 587         if (mode == F_OK) /* 0 */
 588                 return scnprintf(bf, size, "F%s", show_prefix ? suffix : "");
 589 #define P_MODE(n) \
 590         if (mode & n##_OK) { \
 591                 printed += scnprintf(bf + printed, size - printed, "%s%s", #n, show_prefix ? suffix : ""); \
 592                 mode &= ~n##_OK; \
 593         }
 594
 595         P_MODE(R);
 596         P_MODE(W);
 597         P_MODE(X);
 598 #undef P_MODE
 599
 600         if (mode)
 601                 printed += scnprintf(bf + printed, size - printed, "|%#x", mode);
 602
 603         return printed;
 604 }
 605
 606 #define SCA_ACCMODE syscall_arg__scnprintf_access_mode
 607
 608 static size_t syscall_arg__scnprintf_filename(char *bf, size_t size,
 609                                               struct syscall_arg *arg);
 610
 611 #define SCA_FILENAME syscall_arg__scnprintf_filename
 612
 613 static size_t syscall_arg__scnprintf_pipe_flags(char *bf, size_t size,
 614                                                 struct syscall_arg *arg)
 615 {
 616         bool show_prefix = arg->show_string_prefix;
 617         const char *prefix = "O_";
 618         int printed = 0, flags = arg->val;
 619
 620 #define P_FLAG(n) \
 621         if (flags & O_##n) { \
 622                 printed += scnprintf(bf + printed, size - printed, "%s%s%s", printed ? "|" : "", show_prefix ? prefix : "", #n); \
 623                 flags &= ~O_##n; \
 624         }
 625
 626         P_FLAG(CLOEXEC);
 627         P_FLAG(NONBLOCK);
 628 #undef P_FLAG
 629
 630         if (flags)
 631                 printed += scnprintf(bf + printed, size - printed, "%s%#x", printed ? "|" : "", flags);
 632
 633         return printed;
 634 }
 635
 636 #define SCA_PIPE_FLAGS syscall_arg__scnprintf_pipe_flags
 637
 638 #ifndef GRND_NONBLOCK
 639 #define GRND_NONBLOCK   0x0001
 640 #endif
 641 #ifndef GRND_RANDOM
 642 #define GRND_RANDOM     0x0002
 643 #endif
 644
 645 static size_t syscall_arg__scnprintf_getrandom_flags(char *bf, size_t size,
 646                                                    struct syscall_arg *arg)
 647 {
 648         bool show_prefix = arg->show_string_prefix;
 649         const char *prefix = "GRND_";
 650         int printed = 0, flags = arg->val;
 651
 652 #define P_FLAG(n) \
 653         if (flags & GRND_##n) { \
 654                 printed += scnprintf(bf + printed, size - printed, "%s%s%s", printed ? "|" : "", show_prefix ? prefix : "", #n); \
 655                 flags &= ~GRND_##n; \
 656         }
 657
 658         P_FLAG(RANDOM);
 659         P_FLAG(NONBLOCK);
 660 #undef P_FLAG
 661
 662         if (flags)
 663                 printed += scnprintf(bf + printed, size - printed, "%s%#x", printed ? "|" : "", flags);
 664
 665         return printed;
 666 }
 667
 668 #define SCA_GETRANDOM_FLAGS syscall_arg__scnprintf_getrandom_flags
 669
 670 #define STRARRAY(name, array) \
 671           { .scnprintf  = SCA_STRARRAY, \
 672             .parm       = &strarray__##array, }
 673
 674 #define STRARRAY_FLAGS(name, array) \
 675           { .scnprintf  = SCA_STRARRAY_FLAGS, \
 676             .parm       = &strarray__##array, }
 677
 678 #include "trace/beauty/arch_errno_names.c"
 679 #include "trace/beauty/eventfd.c"
 680 #include "trace/beauty/futex_op.c"
 681 #include "trace/beauty/futex_val3.c"
 682 #include "trace/beauty/mmap.c"
 683 #include "trace/beauty/mode_t.c"
 684 #include "trace/beauty/msg_flags.c"
 685 #include "trace/beauty/open_flags.c"
 686 #include "trace/beauty/perf_event_open.c"
 687 #include "trace/beauty/pid.c"
 688 #include "trace/beauty/sched_policy.c"
 689 #include "trace/beauty/seccomp.c"
 690 #include "trace/beauty/signum.c"
 691 #include "trace/beauty/socket_type.c"
 692 #include "trace/beauty/waitid_options.c"
 693
 694 struct syscall_arg_fmt {
 695         size_t     (*scnprintf)(char *bf, size_t size, struct syscall_arg *arg);
 696         unsigned long (*mask_val)(struct syscall_arg *arg, unsigned long val);
 697         void       *parm;
 698         const char *name;
 699         bool       show_zero;
 700 };
 701
 702 static struct syscall_fmt {
 703         const char *name;
 704         const char *alias;
 705         struct {
 706                 const char *sys_enter,
 707                            *sys_exit;
 708         }          bpf_prog_name;
 709         struct syscall_arg_fmt arg[6];
 710         u8         nr_args;
 711         bool       errpid;
 712         bool       timeout;
 713         bool       hexret;
 714 } syscall_fmts[] = {
 715         { .name     = "access",
 716           .arg = { [1] = { .scnprintf = SCA_ACCMODE,  /* mode */ }, }, },
 717         { .name     = "arch_prctl",
 718           .arg = { [0] = { .scnprintf = SCA_X86_ARCH_PRCTL_CODE, /* code */ },
 719                    [1] = { .scnprintf = SCA_PTR, /* arg2 */ }, }, },
 720         { .name     = "bind",
 721           .arg = { [0] = { .scnprintf = SCA_INT, /* fd */ },
 722                    [1] = { .scnprintf = SCA_SOCKADDR, /* umyaddr */ },
 723                    [2] = { .scnprintf = SCA_INT, /* addrlen */ }, }, },
 724         { .name     = "bpf",
 725           .arg = { [0] = STRARRAY(cmd, bpf_cmd), }, },
 726         { .name     = "brk",        .hexret = true,
 727           .arg = { [0] = { .scnprintf = SCA_PTR, /* brk */ }, }, },
 728         { .name     = "clock_gettime",
 729           .arg = { [0] = STRARRAY(clk_id, clockid), }, },
 730         { .name     = "clone",      .errpid = true, .nr_args = 5,
 731           .arg = { [0] = { .name = "flags",         .scnprintf = SCA_CLONE_FLAGS, },
 732                    [1] = { .name = "child_stack",   .scnprintf = SCA_HEX, },
 733                    [2] = { .name = "parent_tidptr", .scnprintf = SCA_HEX, },
 734                    [3] = { .name = "child_tidptr",  .scnprintf = SCA_HEX, },
 735                    [4] = { .name = "tls",           .scnprintf = SCA_HEX, }, }, },
 736         { .name     = "close",
 737           .arg = { [0] = { .scnprintf = SCA_CLOSE_FD, /* fd */ }, }, },
 738         { .name     = "connect",
 739           .arg = { [0] = { .scnprintf = SCA_INT, /* fd */ },
 740                    [1] = { .scnprintf = SCA_SOCKADDR, /* servaddr */ },
 741                    [2] = { .scnprintf = SCA_INT, /* addrlen */ }, }, },
 742         { .name     = "epoll_ctl",
 743           .arg = { [1] = STRARRAY(op, epoll_ctl_ops), }, },
 744         { .name     = "eventfd2",
 745           .arg = { [1] = { .scnprintf = SCA_EFD_FLAGS, /* flags */ }, }, },
 746         { .name     = "fchmodat",
 747           .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
 748         { .name     = "fchownat",
 749           .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
 750         { .name     = "fcntl",
 751           .arg = { [1] = { .scnprintf = SCA_FCNTL_CMD, /* cmd */
 752                            .parm      = &strarrays__fcntl_cmds_arrays,
 753                            .show_zero = true, },
 754                    [2] = { .scnprintf =  SCA_FCNTL_ARG, /* arg */ }, }, },
 755         { .name     = "flock",
 756           .arg = { [1] = { .scnprintf = SCA_FLOCK, /* cmd */ }, }, },
 757         { .name     = "fsconfig",
 758           .arg = { [1] = STRARRAY(cmd, fsconfig_cmds), }, },
 759         { .name     = "fsmount",
 760           .arg = { [1] = STRARRAY_FLAGS(flags, fsmount_flags),
 761                    [2] = { .scnprintf = SCA_FSMOUNT_ATTR_FLAGS, /* attr_flags */ }, }, },
 762         { .name     = "fspick",
 763           .arg = { [0] = { .scnprintf = SCA_FDAT,         /* dfd */ },
 764                    [1] = { .scnprintf = SCA_FILENAME,     /* path */ },
 765                    [2] = { .scnprintf = SCA_FSPICK_FLAGS, /* flags */ }, }, },
 766         { .name     = "fstat", .alias = "newfstat", },
 767         { .name     = "fstatat", .alias = "newfstatat", },
 768         { .name     = "futex",
 769           .arg = { [1] = { .scnprintf = SCA_FUTEX_OP, /* op */ },
 770                    [5] = { .scnprintf = SCA_FUTEX_VAL3, /* val3 */ }, }, },
 771         { .name     = "futimesat",
 772           .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
 773         { .name     = "getitimer",
 774           .arg = { [0] = STRARRAY(which, itimers), }, },
 775         { .name     = "getpid",     .errpid = true, },
 776         { .name     = "getpgid",    .errpid = true, },
 777         { .name     = "getppid",    .errpid = true, },
 778         { .name     = "getrandom",
 779           .arg = { [2] = { .scnprintf = SCA_GETRANDOM_FLAGS, /* flags */ }, }, },
 780         { .name     = "getrlimit",
 781           .arg = { [0] = STRARRAY(resource, rlimit_resources), }, },
 782         { .name     = "gettid",     .errpid = true, },
 783         { .name     = "ioctl",
 784           .arg = {
 785 #if defined(__i386__) || defined(__x86_64__)
 786 /*
 787  * FIXME: Make this available to all arches.
 788  */
 789                    [1] = { .scnprintf = SCA_IOCTL_CMD, /* cmd */ },
 790                    [2] = { .scnprintf = SCA_HEX, /* arg */ }, }, },
 791 #else
 792                    [2] = { .scnprintf = SCA_HEX, /* arg */ }, }, },
 793 #endif
 794         { .name     = "kcmp",       .nr_args = 5,
 795           .arg = { [0] = { .name = "pid1",      .scnprintf = SCA_PID, },
 796                    [1] = { .name = "pid2",      .scnprintf = SCA_PID, },
 797                    [2] = { .name = "type",      .scnprintf = SCA_KCMP_TYPE, },
 798                    [3] = { .name = "idx1",      .scnprintf = SCA_KCMP_IDX, },
 799                    [4] = { .name = "idx2",      .scnprintf = SCA_KCMP_IDX, }, }, },
 800         { .name     = "keyctl",
 801           .arg = { [0] = STRARRAY(option, keyctl_options), }, },
 802         { .name     = "kill",
 803           .arg = { [1] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
 804         { .name     = "linkat",
 805           .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
 806         { .name     = "lseek",
 807           .arg = { [2] = STRARRAY(whence, whences), }, },
 808         { .name     = "lstat", .alias = "newlstat", },
 809         { .name     = "madvise",
 810           .arg = { [0] = { .scnprintf = SCA_HEX,      /* start */ },
 811                    [2] = { .scnprintf = SCA_MADV_BHV, /* behavior */ }, }, },
 812         { .name     = "mkdirat",
 813           .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
 814         { .name     = "mknodat",
 815           .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
 816         { .name     = "mmap",       .hexret = true,
 817 /* The standard mmap maps to old_mmap on s390x */
 818 #if defined(__s390x__)
 819         .alias = "old_mmap",
 820 #endif
 821           .arg = { [2] = { .scnprintf = SCA_MMAP_PROT,  /* prot */ },
 822                    [3] = { .scnprintf = SCA_MMAP_FLAGS, /* flags */ },
 823                    [5] = { .scnprintf = SCA_HEX,        /* offset */ }, }, },
 824         { .name     = "mount",
 825           .arg = { [0] = { .scnprintf = SCA_FILENAME, /* dev_name */ },
 826                    [3] = { .scnprintf = SCA_MOUNT_FLAGS, /* flags */
 827                            .mask_val  = SCAMV_MOUNT_FLAGS, /* flags */ }, }, },
 828         { .name     = "move_mount",
 829           .arg = { [0] = { .scnprintf = SCA_FDAT,       /* from_dfd */ },
 830                    [1] = { .scnprintf = SCA_FILENAME, /* from_pathname */ },
 831                    [2] = { .scnprintf = SCA_FDAT,       /* to_dfd */ },
 832                    [3] = { .scnprintf = SCA_FILENAME, /* to_pathname */ },
 833                    [4] = { .scnprintf = SCA_MOVE_MOUNT_FLAGS, /* flags */ }, }, },
 834         { .name     = "mprotect",
 835           .arg = { [0] = { .scnprintf = SCA_HEX,        /* start */ },
 836                    [2] = { .scnprintf = SCA_MMAP_PROT,  /* prot */ }, }, },
 837         { .name     = "mq_unlink",
 838           .arg = { [0] = { .scnprintf = SCA_FILENAME, /* u_name */ }, }, },
 839         { .name     = "mremap",     .hexret = true,
 840           .arg = { [3] = { .scnprintf = SCA_MREMAP_FLAGS, /* flags */ }, }, },
 841         { .name     = "name_to_handle_at",
 842           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
 843         { .name     = "newfstatat",
 844           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
 845         { .name     = "open",
 846           .arg = { [1] = { .scnprintf = SCA_OPEN_FLAGS, /* flags */ }, }, },
 847         { .name     = "open_by_handle_at",
 848           .arg = { [0] = { .scnprintf = SCA_FDAT,       /* dfd */ },
 849                    [2] = { .scnprintf = SCA_OPEN_FLAGS, /* flags */ }, }, },
 850         { .name     = "openat",
 851           .arg = { [0] = { .scnprintf = SCA_FDAT,       /* dfd */ },
 852                    [2] = { .scnprintf = SCA_OPEN_FLAGS, /* flags */ }, }, },
 853         { .name     = "perf_event_open",
 854           .arg = { [2] = { .scnprintf = SCA_INT,        /* cpu */ },
 855                    [3] = { .scnprintf = SCA_FD,         /* group_fd */ },
 856                    [4] = { .scnprintf = SCA_PERF_FLAGS, /* flags */ }, }, },
 857         { .name     = "pipe2",
 858           .arg = { [1] = { .scnprintf = SCA_PIPE_FLAGS, /* flags */ }, }, },
 859         { .name     = "pkey_alloc",
 860           .arg = { [1] = { .scnprintf = SCA_PKEY_ALLOC_ACCESS_RIGHTS,   /* access_rights */ }, }, },
 861         { .name     = "pkey_free",
 862           .arg = { [0] = { .scnprintf = SCA_INT,        /* key */ }, }, },
 863         { .name     = "pkey_mprotect",
 864           .arg = { [0] = { .scnprintf = SCA_HEX,        /* start */ },
 865                    [2] = { .scnprintf = SCA_MMAP_PROT,  /* prot */ },
 866                    [3] = { .scnprintf = SCA_INT,        /* pkey */ }, }, },
 867         { .name     = "poll", .timeout = true, },
 868         { .name     = "ppoll", .timeout = true, },
 869         { .name     = "prctl",
 870           .arg = { [0] = { .scnprintf = SCA_PRCTL_OPTION, /* option */ },
 871                    [1] = { .scnprintf = SCA_PRCTL_ARG2, /* arg2 */ },
 872                    [2] = { .scnprintf = SCA_PRCTL_ARG3, /* arg3 */ }, }, },
 873         { .name     = "pread", .alias = "pread64", },
 874         { .name     = "preadv", .alias = "pread", },
 875         { .name     = "prlimit64",
 876           .arg = { [1] = STRARRAY(resource, rlimit_resources), }, },
 877         { .name     = "pwrite", .alias = "pwrite64", },
 878         { .name     = "readlinkat",
 879           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
 880         { .name     = "recvfrom",
 881           .arg = { [3] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
 882         { .name     = "recvmmsg",
 883           .arg = { [3] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
 884         { .name     = "recvmsg",
 885           .arg = { [2] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
 886         { .name     = "renameat",
 887           .arg = { [0] = { .scnprintf = SCA_FDAT, /* olddirfd */ },
 888                    [2] = { .scnprintf = SCA_FDAT, /* newdirfd */ }, }, },
 889         { .name     = "renameat2",
 890           .arg = { [0] = { .scnprintf = SCA_FDAT, /* olddirfd */ },
 891                    [2] = { .scnprintf = SCA_FDAT, /* newdirfd */ },
 892                    [4] = { .scnprintf = SCA_RENAMEAT2_FLAGS, /* flags */ }, }, },
 893         { .name     = "rt_sigaction",
 894           .arg = { [0] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
 895         { .name     = "rt_sigprocmask",
 896           .arg = { [0] = STRARRAY(how, sighow), }, },
 897         { .name     = "rt_sigqueueinfo",
 898           .arg = { [1] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
 899         { .name     = "rt_tgsigqueueinfo",
 900           .arg = { [2] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
 901         { .name     = "sched_setscheduler",
 902           .arg = { [1] = { .scnprintf = SCA_SCHED_POLICY, /* policy */ }, }, },
 903         { .name     = "seccomp",
 904           .arg = { [0] = { .scnprintf = SCA_SECCOMP_OP,    /* op */ },
 905                    [1] = { .scnprintf = SCA_SECCOMP_FLAGS, /* flags */ }, }, },
 906         { .name     = "select", .timeout = true, },
 907         { .name     = "sendfile", .alias = "sendfile64", },
 908         { .name     = "sendmmsg",
 909           .arg = { [3] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
 910         { .name     = "sendmsg",
 911           .arg = { [2] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
 912         { .name     = "sendto",
 913           .arg = { [3] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ },
 914                    [4] = { .scnprintf = SCA_SOCKADDR, /* addr */ }, }, },
 915         { .name     = "set_tid_address", .errpid = true, },
 916         { .name     = "setitimer",
 917           .arg = { [0] = STRARRAY(which, itimers), }, },
 918         { .name     = "setrlimit",
 919           .arg = { [0] = STRARRAY(resource, rlimit_resources), }, },
 920         { .name     = "socket",
 921           .arg = { [0] = STRARRAY(family, socket_families),
 922                    [1] = { .scnprintf = SCA_SK_TYPE, /* type */ },
 923                    [2] = { .scnprintf = SCA_SK_PROTO, /* protocol */ }, }, },
 924         { .name     = "socketpair",
 925           .arg = { [0] = STRARRAY(family, socket_families),
 926                    [1] = { .scnprintf = SCA_SK_TYPE, /* type */ },
 927                    [2] = { .scnprintf = SCA_SK_PROTO, /* protocol */ }, }, },
 928         { .name     = "stat", .alias = "newstat", },
 929         { .name     = "statx",
 930           .arg = { [0] = { .scnprintf = SCA_FDAT,        /* fdat */ },
 931                    [2] = { .scnprintf = SCA_STATX_FLAGS, /* flags */ } ,
 932                    [3] = { .scnprintf = SCA_STATX_MASK,  /* mask */ }, }, },
 933         { .name     = "swapoff",
 934           .arg = { [0] = { .scnprintf = SCA_FILENAME, /* specialfile */ }, }, },
 935         { .name     = "swapon",
 936           .arg = { [0] = { .scnprintf = SCA_FILENAME, /* specialfile */ }, }, },
 937         { .name     = "symlinkat",
 938           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
 939         { .name     = "sync_file_range",
 940           .arg = { [3] = { .scnprintf = SCA_SYNC_FILE_RANGE_FLAGS, /* flags */ }, }, },
 941         { .name     = "tgkill",
 942           .arg = { [2] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
 943         { .name     = "tkill",
 944           .arg = { [1] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
 945         { .name     = "umount2", .alias = "umount",
 946           .arg = { [0] = { .scnprintf = SCA_FILENAME, /* name */ }, }, },
 947         { .name     = "uname", .alias = "newuname", },
 948         { .name     = "unlinkat",
 949           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
 950         { .name     = "utimensat",
 951           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dirfd */ }, }, },
 952         { .name     = "wait4",      .errpid = true,
 953           .arg = { [2] = { .scnprintf = SCA_WAITID_OPTIONS, /* options */ }, }, },
 954         { .name     = "waitid",     .errpid = true,
 955           .arg = { [3] = { .scnprintf = SCA_WAITID_OPTIONS, /* options */ }, }, },
 956 };
 957
 958 static int syscall_fmt__cmp(const void *name, const void *fmtp)
 959 {
 960         const struct syscall_fmt *fmt = fmtp;
 961         return strcmp(name, fmt->name);
 962 }
 963
 964 static struct syscall_fmt *syscall_fmt__find(const char *name)
 965 {
 966         const int nmemb = ARRAY_SIZE(syscall_fmts);
 967         return bsearch(name, syscall_fmts, nmemb, sizeof(struct syscall_fmt), syscall_fmt__cmp);
 968 }
 969
 970 static struct syscall_fmt *syscall_fmt__find_by_alias(const char *alias)
 971 {
 972         int i, nmemb = ARRAY_SIZE(syscall_fmts);
 973
 974         for (i = 0; i < nmemb; ++i) {
 975                 if (syscall_fmts[i].alias && strcmp(syscall_fmts[i].alias, alias) == 0)
 976                         return &syscall_fmts[i];
 977         }
 978
 979         return NULL;
 980 }
 981
 982 /*
 983  * is_exit: is this "exit" or "exit_group"?
 984  * is_open: is this "open" or "openat"? To associate the fd returned in sys_exit with the pathname in sys_enter.
 985  * args_size: sum of the sizes of the syscall arguments, anything after that is augmented stuff: pathname for openat, etc.
 986  * nonexistent: Just a hole in the syscall table, syscall id not allocated
 987  */
 988 struct syscall {
 989         struct tep_event    *tp_format;
 990         int                 nr_args;
 991         int                 args_size;
 992         struct {
 993                 struct bpf_program *sys_enter,
 994                                    *sys_exit;
 995         }                   bpf_prog;
 996         bool                is_exit;
 997         bool                is_open;
 998         bool                nonexistent;
 999         struct tep_format_field *args;
1000         const char          *name;
1001         struct syscall_fmt  *fmt;
1002         struct syscall_arg_fmt *arg_fmt;
1003 };
1004
1005 /*
1006  * Must match what is in the BPF program:
1007  *
1008  * tools/perf/examples/bpf/augmented_raw_syscalls.c
1009  */
1010 struct bpf_map_syscall_entry {
1011         bool    enabled;
1012         u16     string_args_len[6];
1013 };
1014
1015 /*
1016  * We need to have this 'calculated' boolean because in some cases we really
1017  * don't know what is the duration of a syscall, for instance, when we start
1018  * a session and some threads are waiting for a syscall to finish, say 'poll',
1019  * in which case all we can do is to print "( ? ) for duration and for the
1020  * start timestamp.
1021  */
1022 static size_t fprintf_duration(unsigned long t, bool calculated, FILE *fp)
1023 {
1024         double duration = (double)t / NSEC_PER_MSEC;
1025         size_t printed = fprintf(fp, "(");
1026
1027         if (!calculated)
1028                 printed += fprintf(fp, "         ");
1029         else if (duration >= 1.0)
1030                 printed += color_fprintf(fp, PERF_COLOR_RED, "%6.3f ms", duration);
1031         else if (duration >= 0.01)
1032                 printed += color_fprintf(fp, PERF_COLOR_YELLOW, "%6.3f ms", duration);
1033         else
1034                 printed += color_fprintf(fp, PERF_COLOR_NORMAL, "%6.3f ms", duration);
1035         return printed + fprintf(fp, "): ");
1036 }
1037
1038 /**
1039  * filename.ptr: The filename char pointer that will be vfs_getname'd
1040  * filename.entry_str_pos: Where to insert the string translated from
1041  *                         filename.ptr by the vfs_getname tracepoint/kprobe.
1042  * ret_scnprintf: syscall args may set this to a different syscall return
1043  *                formatter, for instance, fcntl may return fds, file flags, etc.
1044  */
1045 struct thread_trace {
1046         u64               entry_time;
1047         bool              entry_pending;
1048         unsigned long     nr_events;
1049         unsigned long     pfmaj, pfmin;
1050         char              *entry_str;
1051         double            runtime_ms;
1052         size_t            (*ret_scnprintf)(char *bf, size_t size, struct syscall_arg *arg);
1053         struct {
1054                 unsigned long ptr;
1055                 short int     entry_str_pos;
1056                 bool          pending_open;
1057                 unsigned int  namelen;
1058                 char          *name;
1059         } filename;
1060         struct {
1061                 int           max;
1062                 struct file   *table;
1063         } files;
1064
1065         struct intlist *syscall_stats;
1066 };
1067
1068 static struct thread_trace *thread_trace__new(void)
1069 {
1070         struct thread_trace *ttrace =  zalloc(sizeof(struct thread_trace));
1071
1072         if (ttrace) {
1073                 ttrace->files.max = -1;
1074                 ttrace->syscall_stats = intlist__new(NULL);
1075         }
1076
1077         return ttrace;
1078 }
1079
1080 static struct thread_trace *thread__trace(struct thread *thread, FILE *fp)
1081 {
1082         struct thread_trace *ttrace;
1083
1084         if (thread == NULL)
1085                 goto fail;
1086
1087         if (thread__priv(thread) == NULL)
1088                 thread__set_priv(thread, thread_trace__new());
1089
1090         if (thread__priv(thread) == NULL)
1091                 goto fail;
1092
1093         ttrace = thread__priv(thread);
1094         ++ttrace->nr_events;
1095
1096         return ttrace;
1097 fail:
1098         color_fprintf(fp, PERF_COLOR_RED,
1099                       "WARNING: not enough memory, dropping samples!\n");
1100         return NULL;
1101 }
1102
1103
1104 void syscall_arg__set_ret_scnprintf(struct syscall_arg *arg,
1105                                     size_t (*ret_scnprintf)(char *bf, size_t size, struct syscall_arg *arg))
1106 {
1107         struct thread_trace *ttrace = thread__priv(arg->thread);
1108
1109         ttrace->ret_scnprintf = ret_scnprintf;
1110 }
1111
1112 #define TRACE_PFMAJ             (1 << 0)
1113 #define TRACE_PFMIN             (1 << 1)
1114
1115 static const size_t trace__entry_str_size = 2048;
1116
1117 static struct file *thread_trace__files_entry(struct thread_trace *ttrace, int fd)
1118 {
1119         if (fd < 0)
1120                 return NULL;
1121
1122         if (fd > ttrace->files.max) {
1123                 struct file *nfiles = realloc(ttrace->files.table, (fd + 1) * sizeof(struct file));
1124
1125                 if (nfiles == NULL)
1126                         return NULL;
1127
1128                 if (ttrace->files.max != -1) {
1129                         memset(nfiles + ttrace->files.max + 1, 0,
1130                                (fd - ttrace->files.max) * sizeof(struct file));
1131                 } else {
1132                         memset(nfiles, 0, (fd + 1) * sizeof(struct file));
1133                 }
1134
1135                 ttrace->files.table = nfiles;
1136                 ttrace->files.max   = fd;
1137         }
1138
1139         return ttrace->files.table + fd;
1140 }
1141
1142 struct file *thread__files_entry(struct thread *thread, int fd)
1143 {
1144         return thread_trace__files_entry(thread__priv(thread), fd);
1145 }
1146
1147 static int trace__set_fd_pathname(struct thread *thread, int fd, const char *pathname)
1148 {
1149         struct thread_trace *ttrace = thread__priv(thread);
1150         struct file *file = thread_trace__files_entry(ttrace, fd);
1151
1152         if (file != NULL) {
1153                 struct stat st;
1154                 if (stat(pathname, &st) == 0)
1155                         file->dev_maj = major(st.st_rdev);
1156                 file->pathname = strdup(pathname);
1157                 if (file->pathname)
1158                         return 0;
1159         }
1160
1161         return -1;
1162 }
1163
1164 static int thread__read_fd_path(struct thread *thread, int fd)
1165 {
1166         char linkname[PATH_MAX], pathname[PATH_MAX];
1167         struct stat st;
1168         int ret;
1169
1170         if (thread->pid_ == thread->tid) {
1171                 scnprintf(linkname, sizeof(linkname),
1172                           "/proc/%d/fd/%d", thread->pid_, fd);
1173         } else {
1174                 scnprintf(linkname, sizeof(linkname),
1175                           "/proc/%d/task/%d/fd/%d", thread->pid_, thread->tid, fd);
1176         }
1177
1178         if (lstat(linkname, &st) < 0 || st.st_size + 1 > (off_t)sizeof(pathname))
1179                 return -1;
1180
1181         ret = readlink(linkname, pathname, sizeof(pathname));
1182
1183         if (ret < 0 || ret > st.st_size)
1184                 return -1;
1185
1186         pathname[ret] = '\0';
1187         return trace__set_fd_pathname(thread, fd, pathname);
1188 }
1189
1190 static const char *thread__fd_path(struct thread *thread, int fd,
1191                                    struct trace *trace)
1192 {
1193         struct thread_trace *ttrace = thread__priv(thread);
1194
1195         if (ttrace == NULL || trace->fd_path_disabled)
1196                 return NULL;
1197
1198         if (fd < 0)
1199                 return NULL;
1200
1201         if ((fd > ttrace->files.max || ttrace->files.table[fd].pathname == NULL)) {
1202                 if (!trace->live)
1203                         return NULL;
1204                 ++trace->stats.proc_getname;
1205                 if (thread__read_fd_path(thread, fd))
1206                         return NULL;
1207         }
1208
1209         return ttrace->files.table[fd].pathname;
1210 }
1211
1212 size_t syscall_arg__scnprintf_fd(char *bf, size_t size, struct syscall_arg *arg)
1213 {
1214         int fd = arg->val;
1215         size_t printed = scnprintf(bf, size, "%d", fd);
1216         const char *path = thread__fd_path(arg->thread, fd, arg->trace);
1217
1218         if (path)
1219                 printed += scnprintf(bf + printed, size - printed, "<%s>", path);
1220
1221         return printed;
1222 }
1223
1224 size_t pid__scnprintf_fd(struct trace *trace, pid_t pid, int fd, char *bf, size_t size)
1225 {
1226         size_t printed = scnprintf(bf, size, "%d", fd);
1227         struct thread *thread = machine__find_thread(trace->host, pid, pid);
1228
1229         if (thread) {
1230                 const char *path = thread__fd_path(thread, fd, trace);
1231
1232                 if (path)
1233                         printed += scnprintf(bf + printed, size - printed, "<%s>", path);
1234
1235                 thread__put(thread);
1236         }
1237
1238         return printed;
1239 }
1240
1241 static size_t syscall_arg__scnprintf_close_fd(char *bf, size_t size,
1242                                               struct syscall_arg *arg)
1243 {
1244         int fd = arg->val;
1245         size_t printed = syscall_arg__scnprintf_fd(bf, size, arg);
1246         struct thread_trace *ttrace = thread__priv(arg->thread);
1247
1248         if (ttrace && fd >= 0 && fd <= ttrace->files.max)
1249                 zfree(&ttrace->files.table[fd].pathname);
1250
1251         return printed;
1252 }
1253
1254 static void thread__set_filename_pos(struct thread *thread, const char *bf,
1255                                      unsigned long ptr)
1256 {
1257         struct thread_trace *ttrace = thread__priv(thread);
1258
1259         ttrace->filename.ptr = ptr;
1260         ttrace->filename.entry_str_pos = bf - ttrace->entry_str;
1261 }
1262
1263 static size_t syscall_arg__scnprintf_augmented_string(struct syscall_arg *arg, char *bf, size_t size)
1264 {
1265         struct augmented_arg *augmented_arg = arg->augmented.args;
1266         size_t printed = scnprintf(bf, size, "\"%.*s\"", augmented_arg->size, augmented_arg->value);
1267         /*
1268          * So that the next arg with a payload can consume its augmented arg, i.e. for rename* syscalls
1269          * we would have two strings, each prefixed by its size.
1270          */
1271         int consumed = sizeof(*augmented_arg) + augmented_arg->size;
1272
1273         arg->augmented.args = ((void *)arg->augmented.args) + consumed;
1274         arg->augmented.size -= consumed;
1275
1276         return printed;
1277 }
1278
1279 static size_t syscall_arg__scnprintf_filename(char *bf, size_t size,
1280                                               struct syscall_arg *arg)
1281 {
1282         unsigned long ptr = arg->val;
1283
1284         if (arg->augmented.args)
1285                 return syscall_arg__scnprintf_augmented_string(arg, bf, size);
1286
1287         if (!arg->trace->vfs_getname)
1288                 return scnprintf(bf, size, "%#x", ptr);
1289
1290         thread__set_filename_pos(arg->thread, bf, ptr);
1291         return 0;
1292 }
1293
1294 static bool trace__filter_duration(struct trace *trace, double t)
1295 {
1296         return t < (trace->duration_filter * NSEC_PER_MSEC);
1297 }
1298
1299 static size_t __trace__fprintf_tstamp(struct trace *trace, u64 tstamp, FILE *fp)
1300 {
1301         double ts = (double)(tstamp - trace->base_time) / NSEC_PER_MSEC;
1302
1303         return fprintf(fp, "%10.3f ", ts);
1304 }
1305
1306 /*
1307  * We're handling tstamp=0 as an undefined tstamp, i.e. like when we are
1308  * using ttrace->entry_time for a thread that receives a sys_exit without
1309  * first having received a sys_enter ("poll" issued before tracing session
1310  * starts, lost sys_enter exit due to ring buffer overflow).
1311  */
1312 static size_t trace__fprintf_tstamp(struct trace *trace, u64 tstamp, FILE *fp)
1313 {
1314         if (tstamp > 0)
1315                 return __trace__fprintf_tstamp(trace, tstamp, fp);
1316
1317         return fprintf(fp, "         ? ");
1318 }
1319
1320 static bool done = false;
1321 static bool interrupted = false;
1322
1323 static void sig_handler(int sig)
1324 {
1325         done = true;
1326         interrupted = sig == SIGINT;
1327 }
1328
1329 static size_t trace__fprintf_comm_tid(struct trace *trace, struct thread *thread, FILE *fp)
1330 {
1331         size_t printed = 0;
1332
1333         if (trace->multiple_threads) {
1334                 if (trace->show_comm)
1335                         printed += fprintf(fp, "%.14s/", thread__comm_str(thread));
1336                 printed += fprintf(fp, "%d ", thread->tid);
1337         }
1338
1339         return printed;
1340 }
1341
1342 static size_t trace__fprintf_entry_head(struct trace *trace, struct thread *thread,
1343                                         u64 duration, bool duration_calculated, u64 tstamp, FILE *fp)
1344 {
1345         size_t printed = 0;
1346
1347         if (trace->show_tstamp)
1348                 printed = trace__fprintf_tstamp(trace, tstamp, fp);
1349         if (trace->show_duration)
1350                 printed += fprintf_duration(duration, duration_calculated, fp);
1351         return printed + trace__fprintf_comm_tid(trace, thread, fp);
1352 }
1353
1354 static int trace__process_event(struct trace *trace, struct machine *machine,
1355                                 union perf_event *event, struct perf_sample *sample)
1356 {
1357         int ret = 0;
1358
1359         switch (event->header.type) {
1360         case PERF_RECORD_LOST:
1361                 color_fprintf(trace->output, PERF_COLOR_RED,
1362                               "LOST %" PRIu64 " events!\n", event->lost.lost);
1363                 ret = machine__process_lost_event(machine, event, sample);
1364                 break;
1365         default:
1366                 ret = machine__process_event(machine, event, sample);
1367                 break;
1368         }
1369
1370         return ret;
1371 }
1372
1373 static int trace__tool_process(struct perf_tool *tool,
1374                                union perf_event *event,
1375                                struct perf_sample *sample,
1376                                struct machine *machine)
1377 {
1378         struct trace *trace = container_of(tool, struct trace, tool);
1379         return trace__process_event(trace, machine, event, sample);
1380 }
1381
1382 static char *trace__machine__resolve_kernel_addr(void *vmachine, unsigned long long *addrp, char **modp)
1383 {
1384         struct machine *machine = vmachine;
1385
1386         if (machine->kptr_restrict_warned)
1387                 return NULL;
1388
1389         if (symbol_conf.kptr_restrict) {
1390                 pr_warning("Kernel address maps (/proc/{kallsyms,modules}) are restricted.\n\n"
1391                            "Check /proc/sys/kernel/kptr_restrict and /proc/sys/kernel/perf_event_paranoid.\n\n"
1392                            "Kernel samples will not be resolved.\n");
1393                 machine->kptr_restrict_warned = true;
1394                 return NULL;
1395         }
1396
1397         return machine__resolve_kernel_addr(vmachine, addrp, modp);
1398 }
1399
1400 static int trace__symbols_init(struct trace *trace, struct evlist *evlist)
1401 {
1402         int err = symbol__init(NULL);
1403
1404         if (err)
1405                 return err;
1406
1407         trace->host = machine__new_host();
1408         if (trace->host == NULL)
1409                 return -ENOMEM;
1410
1411         err = trace_event__register_resolver(trace->host, trace__machine__resolve_kernel_addr);
1412         if (err < 0)
1413                 goto out;
1414
1415         err = __machine__synthesize_threads(trace->host, &trace->tool, &trace->opts.target,
1416                                             evlist->core.threads, trace__tool_process, false,
1417                                             1);
1418 out:
1419         if (err)
1420                 symbol__exit();
1421
1422         return err;
1423 }
1424
1425 static void trace__symbols__exit(struct trace *trace)
1426 {
1427         machine__exit(trace->host);
1428         trace->host = NULL;
1429
1430         symbol__exit();
1431 }
1432
1433 static int syscall__alloc_arg_fmts(struct syscall *sc, int nr_args)
1434 {
1435         int idx;
1436
1437         if (nr_args == 6 && sc->fmt && sc->fmt->nr_args != 0)
1438                 nr_args = sc->fmt->nr_args;
1439
1440         sc->arg_fmt = calloc(nr_args, sizeof(*sc->arg_fmt));
1441         if (sc->arg_fmt == NULL)
1442                 return -1;
1443
1444         for (idx = 0; idx < nr_args; ++idx) {
1445                 if (sc->fmt)
1446                         sc->arg_fmt[idx] = sc->fmt->arg[idx];
1447         }
1448
1449         sc->nr_args = nr_args;
1450         return 0;
1451 }
1452
1453 static int syscall__set_arg_fmts(struct syscall *sc)
1454 {
1455         struct tep_format_field *field, *last_field = NULL;
1456         int idx = 0, len;
1457
1458         for (field = sc->args; field; field = field->next, ++idx) {
1459                 last_field = field;
1460
1461                 if (sc->fmt && sc->fmt->arg[idx].scnprintf)
1462                         continue;
1463
1464                 len = strlen(field->name);
1465
1466                 if (strcmp(field->type, "const char *") == 0 &&
1467                     ((len >= 4 && strcmp(field->name + len - 4, "name") == 0) ||
1468                      strstr(field->name, "path") != NULL))
1469                         sc->arg_fmt[idx].scnprintf = SCA_FILENAME;
1470                 else if ((field->flags & TEP_FIELD_IS_POINTER) || strstr(field->name, "addr"))
1471                         sc->arg_fmt[idx].scnprintf = SCA_PTR;
1472                 else if (strcmp(field->type, "pid_t") == 0)
1473                         sc->arg_fmt[idx].scnprintf = SCA_PID;
1474                 else if (strcmp(field->type, "umode_t") == 0)
1475                         sc->arg_fmt[idx].scnprintf = SCA_MODE_T;
1476                 else if ((strcmp(field->type, "int") == 0 ||
1477                           strcmp(field->type, "unsigned int") == 0 ||
1478                           strcmp(field->type, "long") == 0) &&
1479                          len >= 2 && strcmp(field->name + len - 2, "fd") == 0) {
1480                         /*
1481                          * /sys/kernel/tracing/events/syscalls/sys_enter*
1482                          * egrep 'field:.*fd;' .../format|sed -r 's/.*field:([a-z ]+) [a-z_]*fd.+/\1/g'|sort|uniq -c
1483                          * 65 int
1484                          * 23 unsigned int
1485                          * 7 unsigned long
1486                          */
1487                         sc->arg_fmt[idx].scnprintf = SCA_FD;
1488                 }
1489         }
1490
1491         if (last_field)
1492                 sc->args_size = last_field->offset + last_field->size;
1493
1494         return 0;
1495 }
1496
1497 static int trace__read_syscall_info(struct trace *trace, int id)
1498 {
1499         char tp_name[128];
1500         struct syscall *sc;
1501         const char *name = syscalltbl__name(trace->sctbl, id);
1502
1503         if (trace->syscalls.table == NULL) {
1504                 trace->syscalls.table = calloc(trace->sctbl->syscalls.max_id + 1, sizeof(*sc));
1505                 if (trace->syscalls.table == NULL)
1506                         return -ENOMEM;
1507         }
1508
1509         sc = trace->syscalls.table + id;
1510         if (sc->nonexistent)
1511                 return 0;
1512
1513         if (name == NULL) {
1514                 sc->nonexistent = true;
1515                 return 0;
1516         }
1517
1518         sc->name = name;
1519         sc->fmt  = syscall_fmt__find(sc->name);
1520
1521         snprintf(tp_name, sizeof(tp_name), "sys_enter_%s", sc->name);
1522         sc->tp_format = trace_event__tp_format("syscalls", tp_name);
1523
1524         if (IS_ERR(sc->tp_format) && sc->fmt && sc->fmt->alias) {
1525                 snprintf(tp_name, sizeof(tp_name), "sys_enter_%s", sc->fmt->alias);
1526                 sc->tp_format = trace_event__tp_format("syscalls", tp_name);
1527         }
1528
1529         if (syscall__alloc_arg_fmts(sc, IS_ERR(sc->tp_format) ? 6 : sc->tp_format->format.nr_fields))
1530                 return -ENOMEM;
1531
1532         if (IS_ERR(sc->tp_format))
1533                 return PTR_ERR(sc->tp_format);
1534
1535         sc->args = sc->tp_format->format.fields;
1536         /*
1537          * We need to check and discard the first variable '__syscall_nr'
1538          * or 'nr' that mean the syscall number. It is needless here.
1539          * So drop '__syscall_nr' or 'nr' field but does not exist on older kernels.
1540          */
1541         if (sc->args && (!strcmp(sc->args->name, "__syscall_nr") || !strcmp(sc->args->name, "nr"))) {
1542                 sc->args = sc->args->next;
1543                 --sc->nr_args;
1544         }
1545
1546         sc->is_exit = !strcmp(name, "exit_group") || !strcmp(name, "exit");
1547         sc->is_open = !strcmp(name, "open") || !strcmp(name, "openat");
1548
1549         return syscall__set_arg_fmts(sc);
1550 }
1551
1552 static int intcmp(const void *a, const void *b)
1553 {
1554         const int *one = a, *another = b;
1555
1556         return *one - *another;
1557 }
1558
1559 static int trace__validate_ev_qualifier(struct trace *trace)
1560 {
1561         int err = 0;
1562         bool printed_invalid_prefix = false;
1563         struct str_node *pos;
1564         size_t nr_used = 0, nr_allocated = strlist__nr_entries(trace->ev_qualifier);
1565
1566         trace->ev_qualifier_ids.entries = malloc(nr_allocated *
1567                                                  sizeof(trace->ev_qualifier_ids.entries[0]));
1568
1569         if (trace->ev_qualifier_ids.entries == NULL) {
1570                 fputs("Error:\tNot enough memory for allocating events qualifier ids\n",
1571                        trace->output);
1572                 err = -EINVAL;
1573                 goto out;
1574         }
1575
1576         strlist__for_each_entry(pos, trace->ev_qualifier) {
1577                 const char *sc = pos->s;
1578                 int id = syscalltbl__id(trace->sctbl, sc), match_next = -1;
1579
1580                 if (id < 0) {
1581                         id = syscalltbl__strglobmatch_first(trace->sctbl, sc, &match_next);
1582                         if (id >= 0)
1583                                 goto matches;
1584
1585                         if (!printed_invalid_prefix) {
1586                                 pr_debug("Skipping unknown syscalls: ");
1587                                 printed_invalid_prefix = true;
1588                         } else {
1589                                 pr_debug(", ");
1590                         }
1591
1592                         pr_debug("%s", sc);
1593                         continue;
1594                 }
1595 matches:
1596                 trace->ev_qualifier_ids.entries[nr_used++] = id;
1597                 if (match_next == -1)
1598                         continue;
1599
1600                 while (1) {
1601                         id = syscalltbl__strglobmatch_next(trace->sctbl, sc, &match_next);
1602                         if (id < 0)
1603                                 break;
1604                         if (nr_allocated == nr_used) {
1605                                 void *entries;
1606
1607                                 nr_allocated += 8;
1608                                 entries = realloc(trace->ev_qualifier_ids.entries,
1609                                                   nr_allocated * sizeof(trace->ev_qualifier_ids.entries[0]));
1610                                 if (entries == NULL) {
1611                                         err = -ENOMEM;
1612                                         fputs("\nError:\t Not enough memory for parsing\n", trace->output);
1613                                         goto out_free;
1614                                 }
1615                                 trace->ev_qualifier_ids.entries = entries;
1616                         }
1617                         trace->ev_qualifier_ids.entries[nr_used++] = id;
1618                 }
1619         }
1620
1621         trace->ev_qualifier_ids.nr = nr_used;
1622         qsort(trace->ev_qualifier_ids.entries, nr_used, sizeof(int), intcmp);
1623 out:
1624         if (printed_invalid_prefix)
1625                 pr_debug("\n");
1626         return err;
1627 out_free:
1628         zfree(&trace->ev_qualifier_ids.entries);
1629         trace->ev_qualifier_ids.nr = 0;
1630         goto out;
1631 }
1632
1633 static __maybe_unused bool trace__syscall_enabled(struct trace *trace, int id)
1634 {
1635         bool in_ev_qualifier;
1636
1637         if (trace->ev_qualifier_ids.nr == 0)
1638                 return true;
1639
1640         in_ev_qualifier = bsearch(&id, trace->ev_qualifier_ids.entries,
1641                                   trace->ev_qualifier_ids.nr, sizeof(int), intcmp) != NULL;
1642
1643         if (in_ev_qualifier)
1644                return !trace->not_ev_qualifier;
1645
1646         return trace->not_ev_qualifier;
1647 }
1648
1649 /*
1650  * args is to be interpreted as a series of longs but we need to handle
1651  * 8-byte unaligned accesses. args points to raw_data within the event
1652  * and raw_data is guaranteed to be 8-byte unaligned because it is
1653  * preceded by raw_size which is a u32. So we need to copy args to a temp
1654  * variable to read it. Most notably this avoids extended load instructions
1655  * on unaligned addresses
1656  */
1657 unsigned long syscall_arg__val(struct syscall_arg *arg, u8 idx)
1658 {
1659         unsigned long val;
1660         unsigned char *p = arg->args + sizeof(unsigned long) * idx;
1661
1662         memcpy(&val, p, sizeof(val));
1663         return val;
1664 }
1665
1666 static size_t syscall__scnprintf_name(struct syscall *sc, char *bf, size_t size,
1667                                       struct syscall_arg *arg)
1668 {
1669         if (sc->arg_fmt && sc->arg_fmt[arg->idx].name)
1670                 return scnprintf(bf, size, "%s: ", sc->arg_fmt[arg->idx].name);
1671
1672         return scnprintf(bf, size, "arg%d: ", arg->idx);
1673 }
1674
1675 /*
1676  * Check if the value is in fact zero, i.e. mask whatever needs masking, such
1677  * as mount 'flags' argument that needs ignoring some magic flag, see comment
1678  * in tools/perf/trace/beauty/mount_flags.c
1679  */
1680 static unsigned long syscall__mask_val(struct syscall *sc, struct syscall_arg *arg, unsigned long val)
1681 {
1682         if (sc->arg_fmt && sc->arg_fmt[arg->idx].mask_val)
1683                 return sc->arg_fmt[arg->idx].mask_val(arg, val);
1684
1685         return val;
1686 }
1687
1688 static size_t syscall__scnprintf_val(struct syscall *sc, char *bf, size_t size,
1689                                      struct syscall_arg *arg, unsigned long val)
1690 {
1691         if (sc->arg_fmt && sc->arg_fmt[arg->idx].scnprintf) {
1692                 arg->val = val;
1693                 if (sc->arg_fmt[arg->idx].parm)
1694                         arg->parm = sc->arg_fmt[arg->idx].parm;
1695                 return sc->arg_fmt[arg->idx].scnprintf(bf, size, arg);
1696         }
1697         return scnprintf(bf, size, "%ld", val);
1698 }
1699
1700 static size_t syscall__scnprintf_args(struct syscall *sc, char *bf, size_t size,
1701                                       unsigned char *args, void *augmented_args, int augmented_args_size,
1702                                       struct trace *trace, struct thread *thread)
1703 {
1704         size_t printed = 0;
1705         unsigned long val;
1706         u8 bit = 1;
1707         struct syscall_arg arg = {
1708                 .args   = args,
1709                 .augmented = {
1710                         .size = augmented_args_size,
1711                         .args = augmented_args,
1712                 },
1713                 .idx    = 0,
1714                 .mask   = 0,
1715                 .trace  = trace,
1716                 .thread = thread,
1717                 .show_string_prefix = trace->show_string_prefix,
1718         };
1719         struct thread_trace *ttrace = thread__priv(thread);
1720
1721         /*
1722          * Things like fcntl will set this in its 'cmd' formatter to pick the
1723          * right formatter for the return value (an fd? file flags?), which is
1724          * not needed for syscalls that always return a given type, say an fd.
1725          */
1726         ttrace->ret_scnprintf = NULL;
1727
1728         if (sc->args != NULL) {
1729                 struct tep_format_field *field;
1730
1731                 for (field = sc->args; field;
1732                      field = field->next, ++arg.idx, bit <<= 1) {
1733                         if (arg.mask & bit)
1734                                 continue;
1735
1736                         val = syscall_arg__val(&arg, arg.idx);
1737                         /*
1738                          * Some syscall args need some mask, most don't and
1739                          * return val untouched.
1740                          */
1741                         val = syscall__mask_val(sc, &arg, val);
1742
1743                         /*
1744                          * Suppress this argument if its value is zero and
1745                          * and we don't have a string associated in an
1746                          * strarray for it.
1747                          */
1748                         if (val == 0 &&
1749                             !trace->show_zeros &&
1750                             !(sc->arg_fmt &&
1751                               (sc->arg_fmt[arg.idx].show_zero ||
1752                                sc->arg_fmt[arg.idx].scnprintf == SCA_STRARRAY ||
1753                                sc->arg_fmt[arg.idx].scnprintf == SCA_STRARRAYS) &&
1754                               sc->arg_fmt[arg.idx].parm))
1755                                 continue;
1756
1757                         printed += scnprintf(bf + printed, size - printed, "%s", printed ? ", " : "");
1758
1759                         if (trace->show_arg_names)
1760                                 printed += scnprintf(bf + printed, size - printed, "%s: ", field->name);
1761
1762                         printed += syscall__scnprintf_val(sc, bf + printed, size - printed, &arg, val);
1763                 }
1764         } else if (IS_ERR(sc->tp_format)) {
1765                 /*
1766                  * If we managed to read the tracepoint /format file, then we
1767                  * may end up not having any args, like with gettid(), so only
1768                  * print the raw args when we didn't manage to read it.
1769                  */
1770                 while (arg.idx < sc->nr_args) {
1771                         if (arg.mask & bit)
1772                                 goto next_arg;
1773                         val = syscall_arg__val(&arg, arg.idx);
1774                         if (printed)
1775                                 printed += scnprintf(bf + printed, size - printed, ", ");
1776                         printed += syscall__scnprintf_name(sc, bf + printed, size - printed, &arg);
1777                         printed += syscall__scnprintf_val(sc, bf + printed, size - printed, &arg, val);
1778 next_arg:
1779                         ++arg.idx;
1780                         bit <<= 1;
1781                 }
1782         }
1783
1784         return printed;
1785 }
1786
1787 typedef int (*tracepoint_handler)(struct trace *trace, struct evsel *evsel,
1788                                   union perf_event *event,
1789                                   struct perf_sample *sample);
1790
1791 static struct syscall *trace__syscall_info(struct trace *trace,
1792                                            struct evsel *evsel, int id)
1793 {
1794         int err = 0;
1795
1796         if (id < 0) {
1797
1798                 /*
1799                  * XXX: Noticed on x86_64, reproduced as far back as 3.0.36, haven't tried
1800                  * before that, leaving at a higher verbosity level till that is
1801                  * explained. Reproduced with plain ftrace with:
1802                  *
1803                  * echo 1 > /t/events/raw_syscalls/sys_exit/enable
1804                  * grep "NR -1 " /t/trace_pipe
1805                  *
1806                  * After generating some load on the machine.
1807                  */
1808                 if (verbose > 1) {
1809                         static u64 n;
1810                         fprintf(trace->output, "Invalid syscall %d id, skipping (%s, %" PRIu64 ") ...\n",
1811                                 id, perf_evsel__name(evsel), ++n);
1812                 }
1813                 return NULL;
1814         }
1815
1816         err = -EINVAL;
1817
1818         if (id > trace->sctbl->syscalls.max_id)
1819                 goto out_cant_read;
1820
1821         if ((trace->syscalls.table == NULL || trace->syscalls.table[id].name == NULL) &&
1822             (err = trace__read_syscall_info(trace, id)) != 0)
1823                 goto out_cant_read;
1824
1825         if (trace->syscalls.table[id].name == NULL) {
1826                 if (trace->syscalls.table[id].nonexistent)
1827                         return NULL;
1828                 goto out_cant_read;
1829         }
1830
1831         return &trace->syscalls.table[id];
1832
1833 out_cant_read:
1834         if (verbose > 0) {
1835                 char sbuf[STRERR_BUFSIZE];
1836                 fprintf(trace->output, "Problems reading syscall %d: %d (%s)", id, -err, str_error_r(-err, sbuf, sizeof(sbuf)));
1837                 if (id <= trace->sctbl->syscalls.max_id && trace->syscalls.table[id].name != NULL)
1838                         fprintf(trace->output, "(%s)", trace->syscalls.table[id].name);
1839                 fputs(" information\n", trace->output);
1840         }
1841         return NULL;
1842 }
1843
1844 static void thread__update_stats(struct thread_trace *ttrace,
1845                                  int id, struct perf_sample *sample)
1846 {
1847         struct int_node *inode;
1848         struct stats *stats;
1849         u64 duration = 0;
1850
1851         inode = intlist__findnew(ttrace->syscall_stats, id);
1852         if (inode == NULL)
1853                 return;
1854
1855         stats = inode->priv;
1856         if (stats == NULL) {
1857                 stats = malloc(sizeof(struct stats));
1858                 if (stats == NULL)
1859                         return;
1860                 init_stats(stats);
1861                 inode->priv = stats;
1862         }
1863
1864         if (ttrace->entry_time && sample->time > ttrace->entry_time)
1865                 duration = sample->time - ttrace->entry_time;
1866
1867         update_stats(stats, duration);
1868 }
1869
1870 static int trace__printf_interrupted_entry(struct trace *trace)
1871 {
1872         struct thread_trace *ttrace;
1873         size_t printed;
1874         int len;
1875
1876         if (trace->failure_only || trace->current == NULL)
1877                 return 0;
1878
1879         ttrace = thread__priv(trace->current);
1880
1881         if (!ttrace->entry_pending)
1882                 return 0;
1883
1884         printed  = trace__fprintf_entry_head(trace, trace->current, 0, false, ttrace->entry_time, trace->output);
1885         printed += len = fprintf(trace->output, "%s)", ttrace->entry_str);
1886
1887         if (len < trace->args_alignment - 4)
1888                 printed += fprintf(trace->output, "%-*s", trace->args_alignment - 4 - len, " ");
1889
1890         printed += fprintf(trace->output, " ...\n");
1891
1892         ttrace->entry_pending = false;
1893         ++trace->nr_events_printed;
1894
1895         return printed;
1896 }
1897
1898 static int trace__fprintf_sample(struct trace *trace, struct evsel *evsel,
1899                                  struct perf_sample *sample, struct thread *thread)
1900 {
1901         int printed = 0;
1902
1903         if (trace->print_sample) {
1904                 double ts = (double)sample->time / NSEC_PER_MSEC;
1905
1906                 printed += fprintf(trace->output, "%22s %10.3f %s %d/%d [%d]\n",
1907                                    perf_evsel__name(evsel), ts,
1908                                    thread__comm_str(thread),
1909                                    sample->pid, sample->tid, sample->cpu);
1910         }
1911
1912         return printed;
1913 }
1914
1915 static void *syscall__augmented_args(struct syscall *sc, struct perf_sample *sample, int *augmented_args_size, int raw_augmented_args_size)
1916 {
1917         void *augmented_args = NULL;
1918         /*
1919          * For now with BPF raw_augmented we hook into raw_syscalls:sys_enter
1920          * and there we get all 6 syscall args plus the tracepoint common fields
1921          * that gets calculated at the start and the syscall_nr (another long).
1922          * So we check if that is the case and if so don't look after the
1923          * sc->args_size but always after the full raw_syscalls:sys_enter payload,
1924          * which is fixed.
1925          *
1926          * We'll revisit this later to pass s->args_size to the BPF augmenter
1927          * (now tools/perf/examples/bpf/augmented_raw_syscalls.c, so that it
1928          * copies only what we need for each syscall, like what happens when we
1929          * use syscalls:sys_enter_NAME, so that we reduce the kernel/userspace
1930          * traffic to just what is needed for each syscall.
1931          */
1932         int args_size = raw_augmented_args_size ?: sc->args_size;
1933
1934         *augmented_args_size = sample->raw_size - args_size;
1935         if (*augmented_args_size > 0)
1936                 augmented_args = sample->raw_data + args_size;
1937
1938         return augmented_args;
1939 }
1940
1941 static int trace__sys_enter(struct trace *trace, struct evsel *evsel,
1942                             union perf_event *event __maybe_unused,
1943                             struct perf_sample *sample)
1944 {
1945         char *msg;
1946         void *args;
1947         int printed = 0;
1948         struct thread *thread;
1949         int id = perf_evsel__sc_tp_uint(evsel, id, sample), err = -1;
1950         int augmented_args_size = 0;
1951         void *augmented_args = NULL;
1952         struct syscall *sc = trace__syscall_info(trace, evsel, id);
1953         struct thread_trace *ttrace;
1954
1955         if (sc == NULL)
1956                 return -1;
1957
1958         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
1959         ttrace = thread__trace(thread, trace->output);
1960         if (ttrace == NULL)
1961                 goto out_put;
1962
1963         trace__fprintf_sample(trace, evsel, sample, thread);
1964
1965         args = perf_evsel__sc_tp_ptr(evsel, args, sample);
1966
1967         if (ttrace->entry_str == NULL) {
1968                 ttrace->entry_str = malloc(trace__entry_str_size);
1969                 if (!ttrace->entry_str)
1970                         goto out_put;
1971         }
1972
1973         if (!(trace->duration_filter || trace->summary_only || trace->min_stack))
1974                 trace__printf_interrupted_entry(trace);
1975         /*
1976          * If this is raw_syscalls.sys_enter, then it always comes with the 6 possible
1977          * arguments, even if the syscall being handled, say "openat", uses only 4 arguments
1978          * this breaks syscall__augmented_args() check for augmented args, as we calculate
1979          * syscall->args_size using each syscalls:sys_enter_NAME tracefs format file,
1980          * so when handling, say the openat syscall, we end up getting 6 args for the
1981          * raw_syscalls:sys_enter event, when we expected just 4, we end up mistakenly
1982          * thinking that the extra 2 u64 args are the augmented filename, so just check
1983          * here and avoid using augmented syscalls when the evsel is the raw_syscalls one.
1984          */
1985         if (evsel != trace->syscalls.events.sys_enter)
1986                 augmented_args = syscall__augmented_args(sc, sample, &augmented_args_size, trace->raw_augmented_syscalls_args_size);
1987         ttrace->entry_time = sample->time;
1988         msg = ttrace->entry_str;
1989         printed += scnprintf(msg + printed, trace__entry_str_size - printed, "%s(", sc->name);
1990
1991         printed += syscall__scnprintf_args(sc, msg + printed, trace__entry_str_size - printed,
1992                                            args, augmented_args, augmented_args_size, trace, thread);
1993
1994         if (sc->is_exit) {
1995                 if (!(trace->duration_filter || trace->summary_only || trace->failure_only || trace->min_stack)) {
1996                         int alignment = 0;
1997
1998                         trace__fprintf_entry_head(trace, thread, 0, false, ttrace->entry_time, trace->output);
1999                         printed = fprintf(trace->output, "%s)", ttrace->entry_str);
2000                         if (trace->args_alignment > printed)
2001                                 alignment = trace->args_alignment - printed;
2002                         fprintf(trace->output, "%*s= ?\n", alignment, " ");
2003                 }
2004         } else {
2005                 ttrace->entry_pending = true;
2006                 /* See trace__vfs_getname & trace__sys_exit */
2007                 ttrace->filename.pending_open = false;
2008         }
2009
2010         if (trace->current != thread) {
2011                 thread__put(trace->current);
2012                 trace->current = thread__get(thread);
2013         }
2014         err = 0;
2015 out_put:
2016         thread__put(thread);
2017         return err;
2018 }
2019
2020 static int trace__fprintf_sys_enter(struct trace *trace, struct evsel *evsel,
2021                                     struct perf_sample *sample)
2022 {
2023         struct thread_trace *ttrace;
2024         struct thread *thread;
2025         int id = perf_evsel__sc_tp_uint(evsel, id, sample), err = -1;
2026         struct syscall *sc = trace__syscall_info(trace, evsel, id);
2027         char msg[1024];
2028         void *args, *augmented_args = NULL;
2029         int augmented_args_size;
2030
2031         if (sc == NULL)
2032                 return -1;
2033
2034         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2035         ttrace = thread__trace(thread, trace->output);
2036         /*
2037          * We need to get ttrace just to make sure it is there when syscall__scnprintf_args()
2038          * and the rest of the beautifiers accessing it via struct syscall_arg touches it.
2039          */
2040         if (ttrace == NULL)
2041                 goto out_put;
2042
2043         args = perf_evsel__sc_tp_ptr(evsel, args, sample);
2044         augmented_args = syscall__augmented_args(sc, sample, &augmented_args_size, trace->raw_augmented_syscalls_args_size);
2045         syscall__scnprintf_args(sc, msg, sizeof(msg), args, augmented_args, augmented_args_size, trace, thread);
2046         fprintf(trace->output, "%s", msg);
2047         err = 0;
2048 out_put:
2049         thread__put(thread);
2050         return err;
2051 }
2052
2053 static int trace__resolve_callchain(struct trace *trace, struct evsel *evsel,
2054                                     struct perf_sample *sample,
2055                                     struct callchain_cursor *cursor)
2056 {
2057         struct addr_location al;
2058         int max_stack = evsel->core.attr.sample_max_stack ?
2059                         evsel->core.attr.sample_max_stack :
2060                         trace->max_stack;
2061         int err;
2062
2063         if (machine__resolve(trace->host, &al, sample) < 0)
2064                 return -1;
2065
2066         err = thread__resolve_callchain(al.thread, cursor, evsel, sample, NULL, NULL, max_stack);
2067         addr_location__put(&al);
2068         return err;
2069 }
2070
2071 static int trace__fprintf_callchain(struct trace *trace, struct perf_sample *sample)
2072 {
2073         /* TODO: user-configurable print_opts */
2074         const unsigned int print_opts = EVSEL__PRINT_SYM |
2075                                         EVSEL__PRINT_DSO |
2076                                         EVSEL__PRINT_UNKNOWN_AS_ADDR;
2077
2078         return sample__fprintf_callchain(sample, 38, print_opts, &callchain_cursor, trace->output);
2079 }
2080
2081 static const char *errno_to_name(struct evsel *evsel, int err)
2082 {
2083         struct perf_env *env = perf_evsel__env(evsel);
2084         const char *arch_name = perf_env__arch(env);
2085
2086         return arch_syscalls__strerrno(arch_name, err);
2087 }
2088
2089 static int trace__sys_exit(struct trace *trace, struct evsel *evsel,
2090                            union perf_event *event __maybe_unused,
2091                            struct perf_sample *sample)
2092 {
2093         long ret;
2094         u64 duration = 0;
2095         bool duration_calculated = false;
2096         struct thread *thread;
2097         int id = perf_evsel__sc_tp_uint(evsel, id, sample), err = -1, callchain_ret = 0, printed = 0;
2098         int alignment = trace->args_alignment;
2099         struct syscall *sc = trace__syscall_info(trace, evsel, id);
2100         struct thread_trace *ttrace;
2101
2102         if (sc == NULL)
2103                 return -1;
2104
2105         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2106         ttrace = thread__trace(thread, trace->output);
2107         if (ttrace == NULL)
2108                 goto out_put;
2109
2110         trace__fprintf_sample(trace, evsel, sample, thread);
2111
2112         if (trace->summary)
2113                 thread__update_stats(ttrace, id, sample);
2114
2115         ret = perf_evsel__sc_tp_uint(evsel, ret, sample);
2116
2117         if (!trace->fd_path_disabled && sc->is_open && ret >= 0 && ttrace->filename.pending_open) {
2118                 trace__set_fd_pathname(thread, ret, ttrace->filename.name);
2119                 ttrace->filename.pending_open = false;
2120                 ++trace->stats.vfs_getname;
2121         }
2122
2123         if (ttrace->entry_time) {
2124                 duration = sample->time - ttrace->entry_time;
2125                 if (trace__filter_duration(trace, duration))
2126                         goto out;
2127                 duration_calculated = true;
2128         } else if (trace->duration_filter)
2129                 goto out;
2130
2131         if (sample->callchain) {
2132                 callchain_ret = trace__resolve_callchain(trace, evsel, sample, &callchain_cursor);
2133                 if (callchain_ret == 0) {
2134                         if (callchain_cursor.nr < trace->min_stack)
2135                                 goto out;
2136                         callchain_ret = 1;
2137                 }
2138         }
2139
2140         if (trace->summary_only || (ret >= 0 && trace->failure_only))
2141                 goto out;
2142
2143         trace__fprintf_entry_head(trace, thread, duration, duration_calculated, ttrace->entry_time, trace->output);
2144
2145         if (ttrace->entry_pending) {
2146                 printed = fprintf(trace->output, "%s", ttrace->entry_str);
2147         } else {
2148                 printed += fprintf(trace->output, " ... [");
2149                 color_fprintf(trace->output, PERF_COLOR_YELLOW, "continued");
2150                 printed += 9;
2151                 printed += fprintf(trace->output, "]: %s()", sc->name);
2152         }
2153
2154         printed++; /* the closing ')' */
2155
2156         if (alignment > printed)
2157                 alignment -= printed;
2158         else
2159                 alignment = 0;
2160
2161         fprintf(trace->output, ")%*s= ", alignment, " ");
2162
2163         if (sc->fmt == NULL) {
2164                 if (ret < 0)
2165                         goto errno_print;
2166 signed_print:
2167                 fprintf(trace->output, "%ld", ret);
2168         } else if (ret < 0) {
2169 errno_print: {
2170                 char bf[STRERR_BUFSIZE];
2171                 const char *emsg = str_error_r(-ret, bf, sizeof(bf)),
2172                            *e = errno_to_name(evsel, -ret);
2173
2174                 fprintf(trace->output, "-1 %s (%s)", e, emsg);
2175         }
2176         } else if (ret == 0 && sc->fmt->timeout)
2177                 fprintf(trace->output, "0 (Timeout)");
2178         else if (ttrace->ret_scnprintf) {
2179                 char bf[1024];
2180                 struct syscall_arg arg = {
2181                         .val    = ret,
2182                         .thread = thread,
2183                         .trace  = trace,
2184                 };
2185                 ttrace->ret_scnprintf(bf, sizeof(bf), &arg);
2186                 ttrace->ret_scnprintf = NULL;
2187                 fprintf(trace->output, "%s", bf);
2188         } else if (sc->fmt->hexret)
2189                 fprintf(trace->output, "%#lx", ret);
2190         else if (sc->fmt->errpid) {
2191                 struct thread *child = machine__find_thread(trace->host, ret, ret);
2192
2193                 if (child != NULL) {
2194                         fprintf(trace->output, "%ld", ret);
2195                         if (child->comm_set)
2196                                 fprintf(trace->output, " (%s)", thread__comm_str(child));
2197                         thread__put(child);
2198                 }
2199         } else
2200                 goto signed_print;
2201
2202         fputc('\n', trace->output);
2203
2204         /*
2205          * We only consider an 'event' for the sake of --max-events a non-filtered
2206          * sys_enter + sys_exit and other tracepoint events.
2207          */
2208         if (++trace->nr_events_printed == trace->max_events && trace->max_events != ULONG_MAX)
2209                 interrupted = true;
2210
2211         if (callchain_ret > 0)
2212                 trace__fprintf_callchain(trace, sample);
2213         else if (callchain_ret < 0)
2214                 pr_err("Problem processing %s callchain, skipping...\n", perf_evsel__name(evsel));
2215 out:
2216         ttrace->entry_pending = false;
2217         err = 0;
2218 out_put:
2219         thread__put(thread);
2220         return err;
2221 }
2222
2223 static int trace__vfs_getname(struct trace *trace, struct evsel *evsel,
2224                               union perf_event *event __maybe_unused,
2225                               struct perf_sample *sample)
2226 {
2227         struct thread *thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2228         struct thread_trace *ttrace;
2229         size_t filename_len, entry_str_len, to_move;
2230         ssize_t remaining_space;
2231         char *pos;
2232         const char *filename = perf_evsel__rawptr(evsel, sample, "pathname");
2233
2234         if (!thread)
2235                 goto out;
2236
2237         ttrace = thread__priv(thread);
2238         if (!ttrace)
2239                 goto out_put;
2240
2241         filename_len = strlen(filename);
2242         if (filename_len == 0)
2243                 goto out_put;
2244
2245         if (ttrace->filename.namelen < filename_len) {
2246                 char *f = realloc(ttrace->filename.name, filename_len + 1);
2247
2248                 if (f == NULL)
2249                         goto out_put;
2250
2251                 ttrace->filename.namelen = filename_len;
2252                 ttrace->filename.name = f;
2253         }
2254
2255         strcpy(ttrace->filename.name, filename);
2256         ttrace->filename.pending_open = true;
2257
2258         if (!ttrace->filename.ptr)
2259                 goto out_put;
2260
2261         entry_str_len = strlen(ttrace->entry_str);
2262         remaining_space = trace__entry_str_size - entry_str_len - 1; /* \0 */
2263         if (remaining_space <= 0)
2264                 goto out_put;
2265
2266         if (filename_len > (size_t)remaining_space) {
2267                 filename += filename_len - remaining_space;
2268                 filename_len = remaining_space;
2269         }
2270
2271         to_move = entry_str_len - ttrace->filename.entry_str_pos + 1; /* \0 */
2272         pos = ttrace->entry_str + ttrace->filename.entry_str_pos;
2273         memmove(pos + filename_len, pos, to_move);
2274         memcpy(pos, filename, filename_len);
2275
2276         ttrace->filename.ptr = 0;
2277         ttrace->filename.entry_str_pos = 0;
2278 out_put:
2279         thread__put(thread);
2280 out:
2281         return 0;
2282 }
2283
2284 static int trace__sched_stat_runtime(struct trace *trace, struct evsel *evsel,
2285                                      union perf_event *event __maybe_unused,
2286                                      struct perf_sample *sample)
2287 {
2288         u64 runtime = perf_evsel__intval(evsel, sample, "runtime");
2289         double runtime_ms = (double)runtime / NSEC_PER_MSEC;
2290         struct thread *thread = machine__findnew_thread(trace->host,
2291                                                         sample->pid,
2292                                                         sample->tid);
2293         struct thread_trace *ttrace = thread__trace(thread, trace->output);
2294
2295         if (ttrace == NULL)
2296                 goto out_dump;
2297
2298         ttrace->runtime_ms += runtime_ms;
2299         trace->runtime_ms += runtime_ms;
2300 out_put:
2301         thread__put(thread);
2302         return 0;
2303
2304 out_dump:
2305         fprintf(trace->output, "%s: comm=%s,pid=%u,runtime=%" PRIu64 ",vruntime=%" PRIu64 ")\n",
2306                evsel->name,
2307                perf_evsel__strval(evsel, sample, "comm"),
2308                (pid_t)perf_evsel__intval(evsel, sample, "pid"),
2309                runtime,
2310                perf_evsel__intval(evsel, sample, "vruntime"));
2311         goto out_put;
2312 }
2313
2314 static int bpf_output__printer(enum binary_printer_ops op,
2315                                unsigned int val, void *extra __maybe_unused, FILE *fp)
2316 {
2317         unsigned char ch = (unsigned char)val;
2318
2319         switch (op) {
2320         case BINARY_PRINT_CHAR_DATA:
2321                 return fprintf(fp, "%c", isprint(ch) ? ch : '.');
2322         case BINARY_PRINT_DATA_BEGIN:
2323         case BINARY_PRINT_LINE_BEGIN:
2324         case BINARY_PRINT_ADDR:
2325         case BINARY_PRINT_NUM_DATA:
2326         case BINARY_PRINT_NUM_PAD:
2327         case BINARY_PRINT_SEP:
2328         case BINARY_PRINT_CHAR_PAD:
2329         case BINARY_PRINT_LINE_END:
2330         case BINARY_PRINT_DATA_END:
2331         default:
2332                 break;
2333         }
2334
2335         return 0;
2336 }
2337
2338 static void bpf_output__fprintf(struct trace *trace,
2339                                 struct perf_sample *sample)
2340 {
2341         binary__fprintf(sample->raw_data, sample->raw_size, 8,
2342                         bpf_output__printer, NULL, trace->output);
2343         ++trace->nr_events_printed;
2344 }
2345
2346 static int trace__event_handler(struct trace *trace, struct evsel *evsel,
2347                                 union perf_event *event __maybe_unused,
2348                                 struct perf_sample *sample)
2349 {
2350         struct thread *thread;
2351         int callchain_ret = 0;
2352         /*
2353          * Check if we called perf_evsel__disable(evsel) due to, for instance,
2354          * this event's max_events having been hit and this is an entry coming
2355          * from the ring buffer that we should discard, since the max events
2356          * have already been considered/printed.
2357          */
2358         if (evsel->disabled)
2359                 return 0;
2360
2361         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2362
2363         if (sample->callchain) {
2364                 callchain_ret = trace__resolve_callchain(trace, evsel, sample, &callchain_cursor);
2365                 if (callchain_ret == 0) {
2366                         if (callchain_cursor.nr < trace->min_stack)
2367                                 goto out;
2368                         callchain_ret = 1;
2369                 }
2370         }
2371
2372         trace__printf_interrupted_entry(trace);
2373         trace__fprintf_tstamp(trace, sample->time, trace->output);
2374
2375         if (trace->trace_syscalls && trace->show_duration)
2376                 fprintf(trace->output, "(         ): ");
2377
2378         if (thread)
2379                 trace__fprintf_comm_tid(trace, thread, trace->output);
2380
2381         if (evsel == trace->syscalls.events.augmented) {
2382                 int id = perf_evsel__sc_tp_uint(evsel, id, sample);
2383                 struct syscall *sc = trace__syscall_info(trace, evsel, id);
2384
2385                 if (sc) {
2386                         fprintf(trace->output, "%s(", sc->name);
2387                         trace__fprintf_sys_enter(trace, evsel, sample);
2388                         fputc(')', trace->output);
2389                         goto newline;
2390                 }
2391
2392                 /*
2393                  * XXX: Not having the associated syscall info or not finding/adding
2394                  *      the thread should never happen, but if it does...
2395                  *      fall thru and print it as a bpf_output event.
2396                  */
2397         }
2398
2399         fprintf(trace->output, "%s:", evsel->name);
2400
2401         if (perf_evsel__is_bpf_output(evsel)) {
2402                 bpf_output__fprintf(trace, sample);
2403         } else if (evsel->tp_format) {
2404                 if (strncmp(evsel->tp_format->name, "sys_enter_", 10) ||
2405                     trace__fprintf_sys_enter(trace, evsel, sample)) {
2406                         event_format__fprintf(evsel->tp_format, sample->cpu,
2407                                               sample->raw_data, sample->raw_size,
2408                                               trace->output);
2409                         ++trace->nr_events_printed;
2410
2411                         if (evsel->max_events != ULONG_MAX && ++evsel->nr_events_printed == evsel->max_events) {
2412                                 evsel__disable(evsel);
2413                                 evsel__close(evsel);
2414                         }
2415                 }
2416         }
2417
2418 newline:
2419         fprintf(trace->output, "\n");
2420
2421         if (callchain_ret > 0)
2422                 trace__fprintf_callchain(trace, sample);
2423         else if (callchain_ret < 0)
2424                 pr_err("Problem processing %s callchain, skipping...\n", perf_evsel__name(evsel));
2425 out:
2426         thread__put(thread);
2427         return 0;
2428 }
2429
2430 static void print_location(FILE *f, struct perf_sample *sample,
2431                            struct addr_location *al,
2432                            bool print_dso, bool print_sym)
2433 {
2434
2435         if ((verbose > 0 || print_dso) && al->map)
2436                 fprintf(f, "%s@", al->map->dso->long_name);
2437
2438         if ((verbose > 0 || print_sym) && al->sym)
2439                 fprintf(f, "%s+0x%" PRIx64, al->sym->name,
2440                         al->addr - al->sym->start);
2441         else if (al->map)
2442                 fprintf(f, "0x%" PRIx64, al->addr);
2443         else
2444                 fprintf(f, "0x%" PRIx64, sample->addr);
2445 }
2446
2447 static int trace__pgfault(struct trace *trace,
2448                           struct evsel *evsel,
2449                           union perf_event *event __maybe_unused,
2450                           struct perf_sample *sample)
2451 {
2452         struct thread *thread;
2453         struct addr_location al;
2454         char map_type = 'd';
2455         struct thread_trace *ttrace;
2456         int err = -1;
2457         int callchain_ret = 0;
2458
2459         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2460
2461         if (sample->callchain) {
2462                 callchain_ret = trace__resolve_callchain(trace, evsel, sample, &callchain_cursor);
2463                 if (callchain_ret == 0) {
2464                         if (callchain_cursor.nr < trace->min_stack)
2465                                 goto out_put;
2466                         callchain_ret = 1;
2467                 }
2468         }
2469
2470         ttrace = thread__trace(thread, trace->output);
2471         if (ttrace == NULL)
2472                 goto out_put;
2473
2474         if (evsel->core.attr.config == PERF_COUNT_SW_PAGE_FAULTS_MAJ)
2475                 ttrace->pfmaj++;
2476         else
2477                 ttrace->pfmin++;
2478
2479         if (trace->summary_only)
2480                 goto out;
2481
2482         thread__find_symbol(thread, sample->cpumode, sample->ip, &al);
2483
2484         trace__fprintf_entry_head(trace, thread, 0, true, sample->time, trace->output);
2485
2486         fprintf(trace->output, "%sfault [",
2487                 evsel->core.attr.config == PERF_COUNT_SW_PAGE_FAULTS_MAJ ?
2488                 "maj" : "min");
2489
2490         print_location(trace->output, sample, &al, false, true);
2491
2492         fprintf(trace->output, "] => ");
2493
2494         thread__find_symbol(thread, sample->cpumode, sample->addr, &al);
2495
2496         if (!al.map) {
2497                 thread__find_symbol(thread, sample->cpumode, sample->addr, &al);
2498
2499                 if (al.map)
2500                         map_type = 'x';
2501                 else
2502                         map_type = '?';
2503         }
2504
2505         print_location(trace->output, sample, &al, true, false);
2506
2507         fprintf(trace->output, " (%c%c)\n", map_type, al.level);
2508
2509         if (callchain_ret > 0)
2510                 trace__fprintf_callchain(trace, sample);
2511         else if (callchain_ret < 0)
2512                 pr_err("Problem processing %s callchain, skipping...\n", perf_evsel__name(evsel));
2513
2514         ++trace->nr_events_printed;
2515 out:
2516         err = 0;
2517 out_put:
2518         thread__put(thread);
2519         return err;
2520 }
2521
2522 static void trace__set_base_time(struct trace *trace,
2523                                  struct evsel *evsel,
2524                                  struct perf_sample *sample)
2525 {
2526         /*
2527          * BPF events were not setting PERF_SAMPLE_TIME, so be more robust
2528          * and don't use sample->time unconditionally, we may end up having
2529          * some other event in the future without PERF_SAMPLE_TIME for good
2530          * reason, i.e. we may not be interested in its timestamps, just in
2531          * it taking place, picking some piece of information when it
2532          * appears in our event stream (vfs_getname comes to mind).
2533          */
2534         if (trace->base_time == 0 && !trace->full_time &&
2535             (evsel->core.attr.sample_type & PERF_SAMPLE_TIME))
2536                 trace->base_time = sample->time;
2537 }
2538
2539 static int trace__process_sample(struct perf_tool *tool,
2540                                  union perf_event *event,
2541                                  struct perf_sample *sample,
2542                                  struct evsel *evsel,
2543                                  struct machine *machine __maybe_unused)
2544 {
2545         struct trace *trace = container_of(tool, struct trace, tool);
2546         struct thread *thread;
2547         int err = 0;
2548
2549         tracepoint_handler handler = evsel->handler;
2550
2551         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2552         if (thread && thread__is_filtered(thread))
2553                 goto out;
2554
2555         trace__set_base_time(trace, evsel, sample);
2556
2557         if (handler) {
2558                 ++trace->nr_events;
2559                 handler(trace, evsel, event, sample);
2560         }
2561 out:
2562         thread__put(thread);
2563         return err;
2564 }
2565
2566 static int trace__record(struct trace *trace, int argc, const char **argv)
2567 {
2568         unsigned int rec_argc, i, j;
2569         const char **rec_argv;
2570         const char * const record_args[] = {
2571                 "record",
2572                 "-R",
2573                 "-m", "1024",
2574                 "-c", "1",
2575         };
2576
2577         const char * const sc_args[] = { "-e", };
2578         unsigned int sc_args_nr = ARRAY_SIZE(sc_args);
2579         const char * const majpf_args[] = { "-e", "major-faults" };
2580         unsigned int majpf_args_nr = ARRAY_SIZE(majpf_args);
2581         const char * const minpf_args[] = { "-e", "minor-faults" };
2582         unsigned int minpf_args_nr = ARRAY_SIZE(minpf_args);
2583
2584         /* +1 is for the event string below */
2585         rec_argc = ARRAY_SIZE(record_args) + sc_args_nr + 1 +
2586                 majpf_args_nr + minpf_args_nr + argc;
2587         rec_argv = calloc(rec_argc + 1, sizeof(char *));
2588
2589         if (rec_argv == NULL)
2590                 return -ENOMEM;
2591
2592         j = 0;
2593         for (i = 0; i < ARRAY_SIZE(record_args); i++)
2594                 rec_argv[j++] = record_args[i];
2595
2596         if (trace->trace_syscalls) {
2597                 for (i = 0; i < sc_args_nr; i++)
2598                         rec_argv[j++] = sc_args[i];
2599
2600                 /* event string may be different for older kernels - e.g., RHEL6 */
2601                 if (is_valid_tracepoint("raw_syscalls:sys_enter"))
2602                         rec_argv[j++] = "raw_syscalls:sys_enter,raw_syscalls:sys_exit";
2603                 else if (is_valid_tracepoint("syscalls:sys_enter"))
2604                         rec_argv[j++] = "syscalls:sys_enter,syscalls:sys_exit";
2605                 else {
2606                         pr_err("Neither raw_syscalls nor syscalls events exist.\n");
2607                         free(rec_argv);
2608                         return -1;
2609                 }
2610         }
2611
2612         if (trace->trace_pgfaults & TRACE_PFMAJ)
2613                 for (i = 0; i < majpf_args_nr; i++)
2614                         rec_argv[j++] = majpf_args[i];
2615
2616         if (trace->trace_pgfaults & TRACE_PFMIN)
2617                 for (i = 0; i < minpf_args_nr; i++)
2618                         rec_argv[j++] = minpf_args[i];
2619
2620         for (i = 0; i < (unsigned int)argc; i++)
2621                 rec_argv[j++] = argv[i];
2622
2623         return cmd_record(j, rec_argv);
2624 }
2625
2626 static size_t trace__fprintf_thread_summary(struct trace *trace, FILE *fp);
2627
2628 static bool evlist__add_vfs_getname(struct evlist *evlist)
2629 {
2630         bool found = false;
2631         struct evsel *evsel, *tmp;
2632         struct parse_events_error err = { .idx = 0, };
2633         int ret = parse_events(evlist, "probe:vfs_getname*", &err);
2634
2635         if (ret)
2636                 return false;
2637
2638         evlist__for_each_entry_safe(evlist, evsel, tmp) {
2639                 if (!strstarts(perf_evsel__name(evsel), "probe:vfs_getname"))
2640                         continue;
2641
2642                 if (perf_evsel__field(evsel, "pathname")) {
2643                         evsel->handler = trace__vfs_getname;
2644                         found = true;
2645                         continue;
2646                 }
2647
2648                 list_del_init(&evsel->core.node);
2649                 evsel->evlist = NULL;
2650                 evsel__delete(evsel);
2651         }
2652
2653         return found;
2654 }
2655
2656 static struct evsel *perf_evsel__new_pgfault(u64 config)
2657 {
2658         struct evsel *evsel;
2659         struct perf_event_attr attr = {
2660                 .type = PERF_TYPE_SOFTWARE,
2661                 .mmap_data = 1,
2662         };
2663
2664         attr.config = config;
2665         attr.sample_period = 1;
2666
2667         event_attr_init(&attr);
2668
2669         evsel = evsel__new(&attr);
2670         if (evsel)
2671                 evsel->handler = trace__pgfault;
2672
2673         return evsel;
2674 }
2675
2676 static void trace__handle_event(struct trace *trace, union perf_event *event, struct perf_sample *sample)
2677 {
2678         const u32 type = event->header.type;
2679         struct evsel *evsel;
2680
2681         if (type != PERF_RECORD_SAMPLE) {
2682                 trace__process_event(trace, trace->host, event, sample);
2683                 return;
2684         }
2685
2686         evsel = perf_evlist__id2evsel(trace->evlist, sample->id);
2687         if (evsel == NULL) {
2688                 fprintf(trace->output, "Unknown tp ID %" PRIu64 ", skipping...\n", sample->id);
2689                 return;
2690         }
2691
2692         if (evswitch__discard(&trace->evswitch, evsel))
2693                 return;
2694
2695         trace__set_base_time(trace, evsel, sample);
2696
2697         if (evsel->core.attr.type == PERF_TYPE_TRACEPOINT &&
2698             sample->raw_data == NULL) {
2699                 fprintf(trace->output, "%s sample with no payload for tid: %d, cpu %d, raw_size=%d, skipping...\n",
2700                        perf_evsel__name(evsel), sample->tid,
2701                        sample->cpu, sample->raw_size);
2702         } else {
2703                 tracepoint_handler handler = evsel->handler;
2704                 handler(trace, evsel, event, sample);
2705         }
2706
2707         if (trace->nr_events_printed >= trace->max_events && trace->max_events != ULONG_MAX)
2708                 interrupted = true;
2709 }
2710
2711 static int trace__add_syscall_newtp(struct trace *trace)
2712 {
2713         int ret = -1;
2714         struct evlist *evlist = trace->evlist;
2715         struct evsel *sys_enter, *sys_exit;
2716
2717         sys_enter = perf_evsel__raw_syscall_newtp("sys_enter", trace__sys_enter);
2718         if (sys_enter == NULL)
2719                 goto out;
2720
2721         if (perf_evsel__init_sc_tp_ptr_field(sys_enter, args))
2722                 goto out_delete_sys_enter;
2723
2724         sys_exit = perf_evsel__raw_syscall_newtp("sys_exit", trace__sys_exit);
2725         if (sys_exit == NULL)
2726                 goto out_delete_sys_enter;
2727
2728         if (perf_evsel__init_sc_tp_uint_field(sys_exit, ret))
2729                 goto out_delete_sys_exit;
2730
2731         perf_evsel__config_callchain(sys_enter, &trace->opts, &callchain_param);
2732         perf_evsel__config_callchain(sys_exit, &trace->opts, &callchain_param);
2733
2734         evlist__add(evlist, sys_enter);
2735         evlist__add(evlist, sys_exit);
2736
2737         if (callchain_param.enabled && !trace->kernel_syscallchains) {
2738                 /*
2739                  * We're interested only in the user space callchain
2740                  * leading to the syscall, allow overriding that for
2741                  * debugging reasons using --kernel_syscall_callchains
2742                  */
2743                 sys_exit->core.attr.exclude_callchain_kernel = 1;
2744         }
2745
2746         trace->syscalls.events.sys_enter = sys_enter;
2747         trace->syscalls.events.sys_exit  = sys_exit;
2748
2749         ret = 0;
2750 out:
2751         return ret;
2752
2753 out_delete_sys_exit:
2754         evsel__delete_priv(sys_exit);
2755 out_delete_sys_enter:
2756         evsel__delete_priv(sys_enter);
2757         goto out;
2758 }
2759
2760 static int trace__set_ev_qualifier_tp_filter(struct trace *trace)
2761 {
2762         int err = -1;
2763         struct evsel *sys_exit;
2764         char *filter = asprintf_expr_inout_ints("id", !trace->not_ev_qualifier,
2765                                                 trace->ev_qualifier_ids.nr,
2766                                                 trace->ev_qualifier_ids.entries);
2767
2768         if (filter == NULL)
2769                 goto out_enomem;
2770
2771         if (!perf_evsel__append_tp_filter(trace->syscalls.events.sys_enter,
2772                                           filter)) {
2773                 sys_exit = trace->syscalls.events.sys_exit;
2774                 err = perf_evsel__append_tp_filter(sys_exit, filter);
2775         }
2776
2777         free(filter);
2778 out:
2779         return err;
2780 out_enomem:
2781         errno = ENOMEM;
2782         goto out;
2783 }
2784
2785 #ifdef HAVE_LIBBPF_SUPPORT
2786 static struct bpf_program *trace__find_bpf_program_by_title(struct trace *trace, const char *name)
2787 {
2788         if (trace->bpf_obj == NULL)
2789                 return NULL;
2790
2791         return bpf_object__find_program_by_title(trace->bpf_obj, name);
2792 }
2793
2794 static struct bpf_program *trace__find_syscall_bpf_prog(struct trace *trace, struct syscall *sc,
2795                                                         const char *prog_name, const char *type)
2796 {
2797         struct bpf_program *prog;
2798
2799         if (prog_name == NULL) {
2800                 char default_prog_name[256];
2801                 scnprintf(default_prog_name, sizeof(default_prog_name), "!syscalls:sys_%s_%s", type, sc->name);
2802                 prog = trace__find_bpf_program_by_title(trace, default_prog_name);
2803                 if (prog != NULL)
2804                         goto out_found;
2805                 if (sc->fmt && sc->fmt->alias) {
2806                         scnprintf(default_prog_name, sizeof(default_prog_name), "!syscalls:sys_%s_%s", type, sc->fmt->alias);
2807                         prog = trace__find_bpf_program_by_title(trace, default_prog_name);
2808                         if (prog != NULL)
2809                                 goto out_found;
2810                 }
2811                 goto out_unaugmented;
2812         }
2813
2814         prog = trace__find_bpf_program_by_title(trace, prog_name);
2815
2816         if (prog != NULL) {
2817 out_found:
2818                 return prog;
2819         }
2820
2821         pr_debug("Couldn't find BPF prog \"%s\" to associate with syscalls:sys_%s_%s, not augmenting it\n",
2822                  prog_name, type, sc->name);
2823 out_unaugmented:
2824         return trace->syscalls.unaugmented_prog;
2825 }
2826
2827 static void trace__init_syscall_bpf_progs(struct trace *trace, int id)
2828 {
2829         struct syscall *sc = trace__syscall_info(trace, NULL, id);
2830
2831         if (sc == NULL)
2832                 return;
2833
2834         sc->bpf_prog.sys_enter = trace__find_syscall_bpf_prog(trace, sc, sc->fmt ? sc->fmt->bpf_prog_name.sys_enter : NULL, "enter");
2835         sc->bpf_prog.sys_exit  = trace__find_syscall_bpf_prog(trace, sc, sc->fmt ? sc->fmt->bpf_prog_name.sys_exit  : NULL,  "exit");
2836 }
2837
2838 static int trace__bpf_prog_sys_enter_fd(struct trace *trace, int id)
2839 {
2840         struct syscall *sc = trace__syscall_info(trace, NULL, id);
2841         return sc ? bpf_program__fd(sc->bpf_prog.sys_enter) : bpf_program__fd(trace->syscalls.unaugmented_prog);
2842 }
2843
2844 static int trace__bpf_prog_sys_exit_fd(struct trace *trace, int id)
2845 {
2846         struct syscall *sc = trace__syscall_info(trace, NULL, id);
2847         return sc ? bpf_program__fd(sc->bpf_prog.sys_exit) : bpf_program__fd(trace->syscalls.unaugmented_prog);
2848 }
2849
2850 static void trace__init_bpf_map_syscall_args(struct trace *trace, int id, struct bpf_map_syscall_entry *entry)
2851 {
2852         struct syscall *sc = trace__syscall_info(trace, NULL, id);
2853         int arg = 0;
2854
2855         if (sc == NULL)
2856                 goto out;
2857
2858         for (; arg < sc->nr_args; ++arg) {
2859                 entry->string_args_len[arg] = 0;
2860                 if (sc->arg_fmt[arg].scnprintf == SCA_FILENAME) {
2861                         /* Should be set like strace -s strsize */
2862                         entry->string_args_len[arg] = PATH_MAX;
2863                 }
2864         }
2865 out:
2866         for (; arg < 6; ++arg)
2867                 entry->string_args_len[arg] = 0;
2868 }
2869 static int trace__set_ev_qualifier_bpf_filter(struct trace *trace)
2870 {
2871         int fd = bpf_map__fd(trace->syscalls.map);
2872         struct bpf_map_syscall_entry value = {
2873                 .enabled = !trace->not_ev_qualifier,
2874         };
2875         int err = 0;
2876         size_t i;
2877
2878         for (i = 0; i < trace->ev_qualifier_ids.nr; ++i) {
2879                 int key = trace->ev_qualifier_ids.entries[i];
2880
2881                 if (value.enabled) {
2882                         trace__init_bpf_map_syscall_args(trace, key, &value);
2883                         trace__init_syscall_bpf_progs(trace, key);
2884                 }
2885
2886                 err = bpf_map_update_elem(fd, &key, &value, BPF_EXIST);
2887                 if (err)
2888                         break;
2889         }
2890
2891         return err;
2892 }
2893
2894 static int __trace__init_syscalls_bpf_map(struct trace *trace, bool enabled)
2895 {
2896         int fd = bpf_map__fd(trace->syscalls.map);
2897         struct bpf_map_syscall_entry value = {
2898                 .enabled = enabled,
2899         };
2900         int err = 0, key;
2901
2902         for (key = 0; key < trace->sctbl->syscalls.nr_entries; ++key) {
2903                 if (enabled)
2904                         trace__init_bpf_map_syscall_args(trace, key, &value);
2905
2906                 err = bpf_map_update_elem(fd, &key, &value, BPF_ANY);
2907                 if (err)
2908                         break;
2909         }
2910
2911         return err;
2912 }
2913
2914 static int trace__init_syscalls_bpf_map(struct trace *trace)
2915 {
2916         bool enabled = true;
2917
2918         if (trace->ev_qualifier_ids.nr)
2919                 enabled = trace->not_ev_qualifier;
2920
2921         return __trace__init_syscalls_bpf_map(trace, enabled);
2922 }
2923
2924 static struct bpf_program *trace__find_usable_bpf_prog_entry(struct trace *trace, struct syscall *sc)
2925 {
2926         struct tep_format_field *field, *candidate_field;
2927         int id;
2928
2929         /*
2930          * We're only interested in syscalls that have a pointer:
2931          */
2932         for (field = sc->args; field; field = field->next) {
2933                 if (field->flags & TEP_FIELD_IS_POINTER)
2934                         goto try_to_find_pair;
2935         }
2936
2937         return NULL;
2938
2939 try_to_find_pair:
2940         for (id = 0; id < trace->sctbl->syscalls.nr_entries; ++id) {
2941                 struct syscall *pair = trace__syscall_info(trace, NULL, id);
2942                 struct bpf_program *pair_prog;
2943                 bool is_candidate = false;
2944
2945                 if (pair == NULL || pair == sc ||
2946                     pair->bpf_prog.sys_enter == trace->syscalls.unaugmented_prog)
2947                         continue;
2948
2949                 for (field = sc->args, candidate_field = pair->args;
2950                      field && candidate_field; field = field->next, candidate_field = candidate_field->next) {
2951                         bool is_pointer = field->flags & TEP_FIELD_IS_POINTER,
2952                              candidate_is_pointer = candidate_field->flags & TEP_FIELD_IS_POINTER;
2953
2954                         if (is_pointer) {
2955                                if (!candidate_is_pointer) {
2956                                         // The candidate just doesn't copies our pointer arg, might copy other pointers we want.
2957                                         continue;
2958                                }
2959                         } else {
2960                                 if (candidate_is_pointer) {
2961                                         // The candidate might copy a pointer we don't have, skip it.
2962                                         goto next_candidate;
2963                                 }
2964                                 continue;
2965                         }
2966
2967                         if (strcmp(field->type, candidate_field->type))
2968                                 goto next_candidate;
2969
2970                         is_candidate = true;
2971                 }
2972
2973                 if (!is_candidate)
2974                         goto next_candidate;
2975
2976                 /*
2977                  * Check if the tentative pair syscall augmenter has more pointers, if it has,
2978                  * then it may be collecting that and we then can't use it, as it would collect
2979                  * more than what is common to the two syscalls.
2980                  */
2981                 if (candidate_field) {
2982                         for (candidate_field = candidate_field->next; candidate_field; candidate_field = candidate_field->next)
2983                                 if (candidate_field->flags & TEP_FIELD_IS_POINTER)
2984                                         goto next_candidate;
2985                 }
2986
2987                 pair_prog = pair->bpf_prog.sys_enter;
2988                 /*
2989                  * If the pair isn't enabled, then its bpf_prog.sys_enter will not
2990                  * have been searched for, so search it here and if it returns the
2991                  * unaugmented one, then ignore it, otherwise we'll reuse that BPF
2992                  * program for a filtered syscall on a non-filtered one.
2993                  *
2994                  * For instance, we have "!syscalls:sys_enter_renameat" and that is
2995                  * useful for "renameat2".
2996                  */
2997                 if (pair_prog == NULL) {
2998                         pair_prog = trace__find_syscall_bpf_prog(trace, pair, pair->fmt ? pair->fmt->bpf_prog_name.sys_enter : NULL, "enter");
2999                         if (pair_prog == trace->syscalls.unaugmented_prog)
3000                                 goto next_candidate;
3001                 }
3002
3003                 pr_debug("Reusing \"%s\" BPF sys_enter augmenter for \"%s\"\n", pair->name, sc->name);
3004                 return pair_prog;
3005         next_candidate:
3006                 continue;
3007         }
3008
3009         return NULL;
3010 }
3011
3012 static int trace__init_syscalls_bpf_prog_array_maps(struct trace *trace)
3013 {
3014         int map_enter_fd = bpf_map__fd(trace->syscalls.prog_array.sys_enter),
3015             map_exit_fd  = bpf_map__fd(trace->syscalls.prog_array.sys_exit);
3016         int err = 0, key;
3017
3018         for (key = 0; key < trace->sctbl->syscalls.nr_entries; ++key) {
3019                 int prog_fd;
3020
3021                 if (!trace__syscall_enabled(trace, key))
3022                         continue;
3023
3024                 trace__init_syscall_bpf_progs(trace, key);
3025
3026                 // It'll get at least the "!raw_syscalls:unaugmented"
3027                 prog_fd = trace__bpf_prog_sys_enter_fd(trace, key);
3028                 err = bpf_map_update_elem(map_enter_fd, &key, &prog_fd, BPF_ANY);
3029                 if (err)
3030                         break;
3031                 prog_fd = trace__bpf_prog_sys_exit_fd(trace, key);
3032                 err = bpf_map_update_elem(map_exit_fd, &key, &prog_fd, BPF_ANY);
3033                 if (err)
3034                         break;
3035         }
3036
3037         /*
3038          * Now lets do a second pass looking for enabled syscalls without
3039          * an augmenter that have a signature that is a superset of another
3040          * syscall with an augmenter so that we can auto-reuse it.
3041          *
3042          * I.e. if we have an augmenter for the "open" syscall that has
3043          * this signature:
3044          *
3045          *   int open(const char *pathname, int flags, mode_t mode);
3046          *
3047          * I.e. that will collect just the first string argument, then we
3048          * can reuse it for the 'creat' syscall, that has this signature:
3049          *
3050          *   int creat(const char *pathname, mode_t mode);
3051          *
3052          * and for:
3053          *
3054          *   int stat(const char *pathname, struct stat *statbuf);
3055          *   int lstat(const char *pathname, struct stat *statbuf);
3056          *
3057          * Because the 'open' augmenter will collect the first arg as a string,
3058          * and leave alone all the other args, which already helps with
3059          * beautifying 'stat' and 'lstat''s pathname arg.
3060          *
3061          * Then, in time, when 'stat' gets an augmenter that collects both
3062          * first and second arg (this one on the raw_syscalls:sys_exit prog
3063          * array tail call, then that one will be used.
3064          */
3065         for (key = 0; key < trace->sctbl->syscalls.nr_entries; ++key) {
3066                 struct syscall *sc = trace__syscall_info(trace, NULL, key);
3067                 struct bpf_program *pair_prog;
3068                 int prog_fd;
3069
3070                 if (sc == NULL || sc->bpf_prog.sys_enter == NULL)
3071                         continue;
3072
3073                 /*
3074                  * For now we're just reusing the sys_enter prog, and if it
3075                  * already has an augmenter, we don't need to find one.
3076                  */
3077                 if (sc->bpf_prog.sys_enter != trace->syscalls.unaugmented_prog)
3078                         continue;
3079
3080                 /*
3081                  * Look at all the other syscalls for one that has a signature
3082                  * that is close enough that we can share:
3083                  */
3084                 pair_prog = trace__find_usable_bpf_prog_entry(trace, sc);
3085                 if (pair_prog == NULL)
3086                         continue;
3087
3088                 sc->bpf_prog.sys_enter = pair_prog;
3089
3090                 /*
3091                  * Update the BPF_MAP_TYPE_PROG_SHARED for raw_syscalls:sys_enter
3092                  * with the fd for the program we're reusing:
3093                  */
3094                 prog_fd = bpf_program__fd(sc->bpf_prog.sys_enter);
3095                 err = bpf_map_update_elem(map_enter_fd, &key, &prog_fd, BPF_ANY);
3096                 if (err)
3097                         break;
3098         }
3099
3100
3101         return err;
3102 }
3103 #else
3104 static int trace__set_ev_qualifier_bpf_filter(struct trace *trace __maybe_unused)
3105 {
3106         return 0;
3107 }
3108
3109 static int trace__init_syscalls_bpf_map(struct trace *trace __maybe_unused)
3110 {
3111         return 0;
3112 }
3113
3114 static struct bpf_program *trace__find_bpf_program_by_title(struct trace *trace __maybe_unused,
3115                                                             const char *name __maybe_unused)
3116 {
3117         return NULL;
3118 }
3119
3120 static int trace__init_syscalls_bpf_prog_array_maps(struct trace *trace __maybe_unused)
3121 {
3122         return 0;
3123 }
3124 #endif // HAVE_LIBBPF_SUPPORT
3125
3126 static int trace__set_ev_qualifier_filter(struct trace *trace)
3127 {
3128         if (trace->syscalls.map)
3129                 return trace__set_ev_qualifier_bpf_filter(trace);
3130         if (trace->syscalls.events.sys_enter)
3131                 return trace__set_ev_qualifier_tp_filter(trace);
3132         return 0;
3133 }
3134
3135 static int bpf_map__set_filter_pids(struct bpf_map *map __maybe_unused,
3136                                     size_t npids __maybe_unused, pid_t *pids __maybe_unused)
3137 {
3138         int err = 0;
3139 #ifdef HAVE_LIBBPF_SUPPORT
3140         bool value = true;
3141         int map_fd = bpf_map__fd(map);
3142         size_t i;
3143
3144         for (i = 0; i < npids; ++i) {
3145                 err = bpf_map_update_elem(map_fd, &pids[i], &value, BPF_ANY);
3146                 if (err)
3147                         break;
3148         }
3149 #endif
3150         return err;
3151 }
3152
3153 static int trace__set_filter_loop_pids(struct trace *trace)
3154 {
3155         unsigned int nr = 1, err;
3156         pid_t pids[32] = {
3157                 getpid(),
3158         };
3159         struct thread *thread = machine__find_thread(trace->host, pids[0], pids[0]);
3160
3161         while (thread && nr < ARRAY_SIZE(pids)) {
3162                 struct thread *parent = machine__find_thread(trace->host, thread->ppid, thread->ppid);
3163
3164                 if (parent == NULL)
3165                         break;
3166
3167                 if (!strcmp(thread__comm_str(parent), "sshd") ||
3168                     strstarts(thread__comm_str(parent), "gnome-terminal")) {
3169                         pids[nr++] = parent->tid;
3170                         break;
3171                 }
3172                 thread = parent;
3173         }
3174
3175         err = perf_evlist__set_tp_filter_pids(trace->evlist, nr, pids);
3176         if (!err && trace->filter_pids.map)
3177                 err = bpf_map__set_filter_pids(trace->filter_pids.map, nr, pids);
3178
3179         return err;
3180 }
3181
3182 static int trace__set_filter_pids(struct trace *trace)
3183 {
3184         int err = 0;
3185         /*
3186          * Better not use !target__has_task() here because we need to cover the
3187          * case where no threads were specified in the command line, but a
3188          * workload was, and in that case we will fill in the thread_map when
3189          * we fork the workload in perf_evlist__prepare_workload.
3190          */
3191         if (trace->filter_pids.nr > 0) {
3192                 err = perf_evlist__set_tp_filter_pids(trace->evlist, trace->filter_pids.nr,
3193                                                       trace->filter_pids.entries);
3194                 if (!err && trace->filter_pids.map) {
3195                         err = bpf_map__set_filter_pids(trace->filter_pids.map, trace->filter_pids.nr,
3196                                                        trace->filter_pids.entries);
3197                 }
3198         } else if (perf_thread_map__pid(trace->evlist->core.threads, 0) == -1) {
3199                 err = trace__set_filter_loop_pids(trace);
3200         }
3201
3202         return err;
3203 }
3204
3205 static int __trace__deliver_event(struct trace *trace, union perf_event *event)
3206 {
3207         struct evlist *evlist = trace->evlist;
3208         struct perf_sample sample;
3209         int err;
3210
3211         err = perf_evlist__parse_sample(evlist, event, &sample);
3212         if (err)
3213                 fprintf(trace->output, "Can't parse sample, err = %d, skipping...\n", err);
3214         else
3215                 trace__handle_event(trace, event, &sample);
3216
3217         return 0;
3218 }
3219
3220 static int __trace__flush_events(struct trace *trace)
3221 {
3222         u64 first = ordered_events__first_time(&trace->oe.data);
3223         u64 flush = trace->oe.last - NSEC_PER_SEC;
3224
3225         /* Is there some thing to flush.. */
3226         if (first && first < flush)
3227                 return ordered_events__flush_time(&trace->oe.data, flush);
3228
3229         return 0;
3230 }
3231
3232 static int trace__flush_events(struct trace *trace)
3233 {
3234         return !trace->sort_events ? 0 : __trace__flush_events(trace);
3235 }
3236
3237 static int trace__deliver_event(struct trace *trace, union perf_event *event)
3238 {
3239         int err;
3240
3241         if (!trace->sort_events)
3242                 return __trace__deliver_event(trace, event);
3243
3244         err = perf_evlist__parse_sample_timestamp(trace->evlist, event, &trace->oe.last);
3245         if (err && err != -1)
3246                 return err;
3247
3248         err = ordered_events__queue(&trace->oe.data, event, trace->oe.last, 0);
3249         if (err)
3250                 return err;
3251
3252         return trace__flush_events(trace);
3253 }
3254
3255 static int ordered_events__deliver_event(struct ordered_events *oe,
3256                                          struct ordered_event *event)
3257 {
3258         struct trace *trace = container_of(oe, struct trace, oe.data);
3259
3260         return __trace__deliver_event(trace, event->event);
3261 }
3262
3263 static int trace__run(struct trace *trace, int argc, const char **argv)
3264 {
3265         struct evlist *evlist = trace->evlist;
3266         struct evsel *evsel, *pgfault_maj = NULL, *pgfault_min = NULL;
3267         int err = -1, i;
3268         unsigned long before;
3269         const bool forks = argc > 0;
3270         bool draining = false;
3271
3272         trace->live = true;
3273
3274         if (!trace->raw_augmented_syscalls) {
3275                 if (trace->trace_syscalls && trace__add_syscall_newtp(trace))
3276                         goto out_error_raw_syscalls;
3277
3278                 if (trace->trace_syscalls)
3279                         trace->vfs_getname = evlist__add_vfs_getname(evlist);
3280         }
3281
3282         if ((trace->trace_pgfaults & TRACE_PFMAJ)) {
3283                 pgfault_maj = perf_evsel__new_pgfault(PERF_COUNT_SW_PAGE_FAULTS_MAJ);
3284                 if (pgfault_maj == NULL)
3285                         goto out_error_mem;
3286                 perf_evsel__config_callchain(pgfault_maj, &trace->opts, &callchain_param);
3287                 evlist__add(evlist, pgfault_maj);
3288         }
3289
3290         if ((trace->trace_pgfaults & TRACE_PFMIN)) {
3291                 pgfault_min = perf_evsel__new_pgfault(PERF_COUNT_SW_PAGE_FAULTS_MIN);
3292                 if (pgfault_min == NULL)
3293                         goto out_error_mem;
3294                 perf_evsel__config_callchain(pgfault_min, &trace->opts, &callchain_param);
3295                 evlist__add(evlist, pgfault_min);
3296         }
3297
3298         if (trace->sched &&
3299             perf_evlist__add_newtp(evlist, "sched", "sched_stat_runtime",
3300                                    trace__sched_stat_runtime))
3301                 goto out_error_sched_stat_runtime;
3302
3303         /*
3304          * If a global cgroup was set, apply it to all the events without an
3305          * explicit cgroup. I.e.:
3306          *
3307          *      trace -G A -e sched:*switch
3308          *
3309          * Will set all raw_syscalls:sys_{enter,exit}, pgfault, vfs_getname, etc
3310          * _and_ sched:sched_switch to the 'A' cgroup, while:
3311          *
3312          * trace -e sched:*switch -G A
3313          *
3314          * will only set the sched:sched_switch event to the 'A' cgroup, all the
3315          * other events (raw_syscalls:sys_{enter,exit}, etc are left "without"
3316          * a cgroup (on the root cgroup, sys wide, etc).
3317          *
3318          * Multiple cgroups:
3319          *
3320          * trace -G A -e sched:*switch -G B
3321          *
3322          * the syscall ones go to the 'A' cgroup, the sched:sched_switch goes
3323          * to the 'B' cgroup.
3324          *
3325          * evlist__set_default_cgroup() grabs a reference of the passed cgroup
3326          * only for the evsels still without a cgroup, i.e. evsel->cgroup == NULL.
3327          */
3328         if (trace->cgroup)
3329                 evlist__set_default_cgroup(trace->evlist, trace->cgroup);
3330
3331         err = perf_evlist__create_maps(evlist, &trace->opts.target);
3332         if (err < 0) {
3333                 fprintf(trace->output, "Problems parsing the target to trace, check your options!\n");
3334                 goto out_delete_evlist;
3335         }
3336
3337         err = trace__symbols_init(trace, evlist);
3338         if (err < 0) {
3339                 fprintf(trace->output, "Problems initializing symbol libraries!\n");
3340                 goto out_delete_evlist;
3341         }
3342
3343         perf_evlist__config(evlist, &trace->opts, &callchain_param);
3344
3345         signal(SIGCHLD, sig_handler);
3346         signal(SIGINT, sig_handler);
3347
3348         if (forks) {
3349                 err = perf_evlist__prepare_workload(evlist, &trace->opts.target,
3350                                                     argv, false, NULL);
3351                 if (err < 0) {
3352                         fprintf(trace->output, "Couldn't run the workload!\n");
3353                         goto out_delete_evlist;
3354                 }
3355         }
3356
3357         err = evlist__open(evlist);
3358         if (err < 0)
3359                 goto out_error_open;
3360
3361         err = bpf__apply_obj_config();
3362         if (err) {
3363                 char errbuf[BUFSIZ];
3364
3365                 bpf__strerror_apply_obj_config(err, errbuf, sizeof(errbuf));
3366                 pr_err("ERROR: Apply config to BPF failed: %s\n",
3367                          errbuf);
3368                 goto out_error_open;
3369         }
3370
3371         err = trace__set_filter_pids(trace);
3372         if (err < 0)
3373                 goto out_error_mem;
3374
3375         if (trace->syscalls.map)
3376                 trace__init_syscalls_bpf_map(trace);
3377
3378         if (trace->syscalls.prog_array.sys_enter)
3379                 trace__init_syscalls_bpf_prog_array_maps(trace);
3380
3381         if (trace->ev_qualifier_ids.nr > 0) {
3382                 err = trace__set_ev_qualifier_filter(trace);
3383                 if (err < 0)
3384                         goto out_errno;
3385
3386                 if (trace->syscalls.events.sys_exit) {
3387                         pr_debug("event qualifier tracepoint filter: %s\n",
3388                                  trace->syscalls.events.sys_exit->filter);
3389                 }
3390         }
3391
3392         /*
3393          * If the "close" syscall is not traced, then we will not have the
3394          * opportunity to, in syscall_arg__scnprintf_close_fd() invalidate the
3395          * fd->pathname table and were ending up showing the last value set by
3396          * syscalls opening a pathname and associating it with a descriptor or
3397          * reading it from /proc/pid/fd/ in cases where that doesn't make
3398          * sense.
3399          *
3400          *  So just disable this beautifier (SCA_FD, SCA_FDAT) when 'close' is
3401          *  not in use.
3402          */
3403         trace->fd_path_disabled = !trace__syscall_enabled(trace, syscalltbl__id(trace->sctbl, "close"));
3404
3405         err = perf_evlist__apply_filters(evlist, &evsel);
3406         if (err < 0)
3407                 goto out_error_apply_filters;
3408
3409         if (trace->dump.map)
3410                 bpf_map__fprintf(trace->dump.map, trace->output);
3411
3412         err = evlist__mmap(evlist, trace->opts.mmap_pages);
3413         if (err < 0)
3414                 goto out_error_mmap;
3415
3416         if (!target__none(&trace->opts.target) && !trace->opts.initial_delay)
3417                 evlist__enable(evlist);
3418
3419         if (forks)
3420                 perf_evlist__start_workload(evlist);
3421
3422         if (trace->opts.initial_delay) {
3423                 usleep(trace->opts.initial_delay * 1000);
3424                 evlist__enable(evlist);
3425         }
3426
3427         trace->multiple_threads = perf_thread_map__pid(evlist->core.threads, 0) == -1 ||
3428                                   evlist->core.threads->nr > 1 ||
3429                                   perf_evlist__first(evlist)->core.attr.inherit;
3430
3431         /*
3432          * Now that we already used evsel->core.attr to ask the kernel to setup the
3433          * events, lets reuse evsel->core.attr.sample_max_stack as the limit in
3434          * trace__resolve_callchain(), allowing per-event max-stack settings
3435          * to override an explicitly set --max-stack global setting.
3436          */
3437         evlist__for_each_entry(evlist, evsel) {
3438                 if (evsel__has_callchain(evsel) &&
3439                     evsel->core.attr.sample_max_stack == 0)
3440                         evsel->core.attr.sample_max_stack = trace->max_stack;
3441         }
3442 again:
3443         before = trace->nr_events;
3444
3445         for (i = 0; i < evlist->nr_mmaps; i++) {
3446                 union perf_event *event;
3447                 struct mmap *md;
3448
3449                 md = &evlist->mmap[i];
3450                 if (perf_mmap__read_init(md) < 0)
3451                         continue;
3452
3453                 while ((event = perf_mmap__read_event(md)) != NULL) {
3454                         ++trace->nr_events;
3455
3456                         err = trace__deliver_event(trace, event);
3457                         if (err)
3458                                 goto out_disable;
3459
3460                         perf_mmap__consume(md);
3461
3462                         if (interrupted)
3463                                 goto out_disable;
3464
3465                         if (done && !draining) {
3466                                 evlist__disable(evlist);
3467                                 draining = true;
3468                         }
3469                 }
3470                 perf_mmap__read_done(md);
3471         }
3472
3473         if (trace->nr_events == before) {
3474                 int timeout = done ? 100 : -1;
3475
3476                 if (!draining && perf_evlist__poll(evlist, timeout) > 0) {
3477                         if (perf_evlist__filter_pollfd(evlist, POLLERR | POLLHUP | POLLNVAL) == 0)
3478                                 draining = true;
3479
3480                         goto again;
3481                 } else {
3482                         if (trace__flush_events(trace))
3483                                 goto out_disable;
3484                 }
3485         } else {
3486                 goto again;
3487         }
3488
3489 out_disable:
3490         thread__zput(trace->current);
3491
3492         evlist__disable(evlist);
3493
3494         if (trace->sort_events)
3495                 ordered_events__flush(&trace->oe.data, OE_FLUSH__FINAL);
3496
3497         if (!err) {
3498                 if (trace->summary)
3499                         trace__fprintf_thread_summary(trace, trace->output);
3500
3501                 if (trace->show_tool_stats) {
3502                         fprintf(trace->output, "Stats:\n "
3503                                                " vfs_getname : %" PRIu64 "\n"
3504                                                " proc_getname: %" PRIu64 "\n",
3505                                 trace->stats.vfs_getname,
3506                                 trace->stats.proc_getname);
3507                 }
3508         }
3509
3510 out_delete_evlist:
3511         trace__symbols__exit(trace);
3512
3513         evlist__delete(evlist);
3514         cgroup__put(trace->cgroup);
3515         trace->evlist = NULL;
3516         trace->live = false;
3517         return err;
3518 {
3519         char errbuf[BUFSIZ];
3520
3521 out_error_sched_stat_runtime:
3522         tracing_path__strerror_open_tp(errno, errbuf, sizeof(errbuf), "sched", "sched_stat_runtime");
3523         goto out_error;
3524
3525 out_error_raw_syscalls:
3526         tracing_path__strerror_open_tp(errno, errbuf, sizeof(errbuf), "raw_syscalls", "sys_(enter|exit)");
3527         goto out_error;
3528
3529 out_error_mmap:
3530         perf_evlist__strerror_mmap(evlist, errno, errbuf, sizeof(errbuf));
3531         goto out_error;
3532
3533 out_error_open:
3534         perf_evlist__strerror_open(evlist, errno, errbuf, sizeof(errbuf));
3535
3536 out_error:
3537         fprintf(trace->output, "%s\n", errbuf);
3538         goto out_delete_evlist;
3539
3540 out_error_apply_filters:
3541         fprintf(trace->output,
3542                 "Failed to set filter \"%s\" on event %s with %d (%s)\n",
3543                 evsel->filter, perf_evsel__name(evsel), errno,
3544                 str_error_r(errno, errbuf, sizeof(errbuf)));
3545         goto out_delete_evlist;
3546 }
3547 out_error_mem:
3548         fprintf(trace->output, "Not enough memory to run!\n");
3549         goto out_delete_evlist;
3550
3551 out_errno:
3552         fprintf(trace->output, "errno=%d,%s\n", errno, strerror(errno));
3553         goto out_delete_evlist;
3554 }
3555
3556 static int trace__replay(struct trace *trace)
3557 {
3558         const struct evsel_str_handler handlers[] = {
3559                 { "probe:vfs_getname",       trace__vfs_getname, },
3560         };
3561         struct perf_data data = {
3562                 .path  = input_name,
3563                 .mode  = PERF_DATA_MODE_READ,
3564                 .force = trace->force,
3565         };
3566         struct perf_session *session;
3567         struct evsel *evsel;
3568         int err = -1;
3569
3570         trace->tool.sample        = trace__process_sample;
3571         trace->tool.mmap          = perf_event__process_mmap;
3572         trace->tool.mmap2         = perf_event__process_mmap2;
3573         trace->tool.comm          = perf_event__process_comm;
3574         trace->tool.exit          = perf_event__process_exit;
3575         trace->tool.fork          = perf_event__process_fork;
3576         trace->tool.attr          = perf_event__process_attr;
3577         trace->tool.tracing_data  = perf_event__process_tracing_data;
3578         trace->tool.build_id      = perf_event__process_build_id;
3579         trace->tool.namespaces    = perf_event__process_namespaces;
3580
3581         trace->tool.ordered_events = true;
3582         trace->tool.ordering_requires_timestamps = true;
3583
3584         /* add tid to output */
3585         trace->multiple_threads = true;
3586
3587         session = perf_session__new(&data, false, &trace->tool);
3588         if (IS_ERR(session))
3589                 return PTR_ERR(session);
3590
3591         if (trace->opts.target.pid)
3592                 symbol_conf.pid_list_str = strdup(trace->opts.target.pid);
3593
3594         if (trace->opts.target.tid)
3595                 symbol_conf.tid_list_str = strdup(trace->opts.target.tid);
3596
3597         if (symbol__init(&session->header.env) < 0)
3598                 goto out;
3599
3600         trace->host = &session->machines.host;
3601
3602         err = perf_session__set_tracepoints_handlers(session, handlers);
3603         if (err)
3604                 goto out;
3605
3606         evsel = perf_evlist__find_tracepoint_by_name(session->evlist,
3607                                                      "raw_syscalls:sys_enter");
3608         /* older kernels have syscalls tp versus raw_syscalls */
3609         if (evsel == NULL)
3610                 evsel = perf_evlist__find_tracepoint_by_name(session->evlist,
3611                                                              "syscalls:sys_enter");
3612
3613         if (evsel &&
3614             (perf_evsel__init_raw_syscall_tp(evsel, trace__sys_enter) < 0 ||
3615             perf_evsel__init_sc_tp_ptr_field(evsel, args))) {
3616                 pr_err("Error during initialize raw_syscalls:sys_enter event\n");
3617                 goto out;
3618         }
3619
3620         evsel = perf_evlist__find_tracepoint_by_name(session->evlist,
3621                                                      "raw_syscalls:sys_exit");
3622         if (evsel == NULL)
3623                 evsel = perf_evlist__find_tracepoint_by_name(session->evlist,
3624                                                              "syscalls:sys_exit");
3625         if (evsel &&
3626             (perf_evsel__init_raw_syscall_tp(evsel, trace__sys_exit) < 0 ||
3627             perf_evsel__init_sc_tp_uint_field(evsel, ret))) {
3628                 pr_err("Error during initialize raw_syscalls:sys_exit event\n");
3629                 goto out;
3630         }
3631
3632         evlist__for_each_entry(session->evlist, evsel) {
3633                 if (evsel->core.attr.type == PERF_TYPE_SOFTWARE &&
3634                     (evsel->core.attr.config == PERF_COUNT_SW_PAGE_FAULTS_MAJ ||
3635                      evsel->core.attr.config == PERF_COUNT_SW_PAGE_FAULTS_MIN ||
3636                      evsel->core.attr.config == PERF_COUNT_SW_PAGE_FAULTS))
3637                         evsel->handler = trace__pgfault;
3638         }
3639
3640         setup_pager();
3641
3642         err = perf_session__process_events(session);
3643         if (err)
3644                 pr_err("Failed to process events, error %d", err);
3645
3646         else if (trace->summary)
3647                 trace__fprintf_thread_summary(trace, trace->output);
3648
3649 out:
3650         perf_session__delete(session);
3651
3652         return err;
3653 }
3654
3655 static size_t trace__fprintf_threads_header(FILE *fp)
3656 {
3657         size_t printed;
3658
3659         printed  = fprintf(fp, "\n Summary of events:\n\n");
3660
3661         return printed;
3662 }
3663
3664 DEFINE_RESORT_RB(syscall_stats, a->msecs > b->msecs,
3665         struct stats    *stats;
3666         double          msecs;
3667         int             syscall;
3668 )
3669 {
3670         struct int_node *source = rb_entry(nd, struct int_node, rb_node);
3671         struct stats *stats = source->priv;
3672
3673         entry->syscall = source->i;
3674         entry->stats   = stats;
3675         entry->msecs   = stats ? (u64)stats->n * (avg_stats(stats) / NSEC_PER_MSEC) : 0;
3676 }
3677
3678 static size_t thread__dump_stats(struct thread_trace *ttrace,
3679                                  struct trace *trace, FILE *fp)
3680 {
3681         size_t printed = 0;
3682         struct syscall *sc;
3683         struct rb_node *nd;
3684         DECLARE_RESORT_RB_INTLIST(syscall_stats, ttrace->syscall_stats);
3685
3686         if (syscall_stats == NULL)
3687                 return 0;
3688
3689         printed += fprintf(fp, "\n");
3690
3691         printed += fprintf(fp, "   syscall            calls    total       min       avg       max      stddev\n");
3692         printed += fprintf(fp, "                               (msec)    (msec)    (msec)    (msec)        (%%)\n");
3693         printed += fprintf(fp, "   --------------- -------- --------- --------- --------- ---------     ------\n");
3694
3695         resort_rb__for_each_entry(nd, syscall_stats) {
3696                 struct stats *stats = syscall_stats_entry->stats;
3697                 if (stats) {
3698                         double min = (double)(stats->min) / NSEC_PER_MSEC;
3699                         double max = (double)(stats->max) / NSEC_PER_MSEC;
3700                         double avg = avg_stats(stats);
3701                         double pct;
3702                         u64 n = (u64) stats->n;
3703
3704                         pct = avg ? 100.0 * stddev_stats(stats)/avg : 0.0;
3705                         avg /= NSEC_PER_MSEC;
3706
3707                         sc = &trace->syscalls.table[syscall_stats_entry->syscall];
3708                         printed += fprintf(fp, "   %-15s", sc->name);
3709                         printed += fprintf(fp, " %8" PRIu64 " %9.3f %9.3f %9.3f",
3710                                            n, syscall_stats_entry->msecs, min, avg);
3711                         printed += fprintf(fp, " %9.3f %9.2f%%\n", max, pct);
3712                 }
3713         }
3714
3715         resort_rb__delete(syscall_stats);
3716         printed += fprintf(fp, "\n\n");
3717
3718         return printed;
3719 }
3720
3721 static size_t trace__fprintf_thread(FILE *fp, struct thread *thread, struct trace *trace)
3722 {
3723         size_t printed = 0;
3724         struct thread_trace *ttrace = thread__priv(thread);
3725         double ratio;
3726
3727         if (ttrace == NULL)
3728                 return 0;
3729
3730         ratio = (double)ttrace->nr_events / trace->nr_events * 100.0;
3731
3732         printed += fprintf(fp, " %s (%d), ", thread__comm_str(thread), thread->tid);
3733         printed += fprintf(fp, "%lu events, ", ttrace->nr_events);
3734         printed += fprintf(fp, "%.1f%%", ratio);
3735         if (ttrace->pfmaj)
3736                 printed += fprintf(fp, ", %lu majfaults", ttrace->pfmaj);
3737         if (ttrace->pfmin)
3738                 printed += fprintf(fp, ", %lu minfaults", ttrace->pfmin);
3739         if (trace->sched)
3740                 printed += fprintf(fp, ", %.3f msec\n", ttrace->runtime_ms);
3741         else if (fputc('\n', fp) != EOF)
3742                 ++printed;
3743
3744         printed += thread__dump_stats(ttrace, trace, fp);
3745
3746         return printed;
3747 }
3748
3749 static unsigned long thread__nr_events(struct thread_trace *ttrace)
3750 {
3751         return ttrace ? ttrace->nr_events : 0;
3752 }
3753
3754 DEFINE_RESORT_RB(threads, (thread__nr_events(a->thread->priv) < thread__nr_events(b->thread->priv)),
3755         struct thread *thread;
3756 )
3757 {
3758         entry->thread = rb_entry(nd, struct thread, rb_node);
3759 }
3760
3761 static size_t trace__fprintf_thread_summary(struct trace *trace, FILE *fp)
3762 {
3763         size_t printed = trace__fprintf_threads_header(fp);
3764         struct rb_node *nd;
3765         int i;
3766
3767         for (i = 0; i < THREADS__TABLE_SIZE; i++) {
3768                 DECLARE_RESORT_RB_MACHINE_THREADS(threads, trace->host, i);
3769
3770                 if (threads == NULL) {
3771                         fprintf(fp, "%s", "Error sorting output by nr_events!\n");
3772                         return 0;
3773                 }
3774
3775                 resort_rb__for_each_entry(nd, threads)
3776                         printed += trace__fprintf_thread(fp, threads_entry->thread, trace);
3777
3778                 resort_rb__delete(threads);
3779         }
3780         return printed;
3781 }
3782
3783 static int trace__set_duration(const struct option *opt, const char *str,
3784                                int unset __maybe_unused)
3785 {
3786         struct trace *trace = opt->value;
3787
3788         trace->duration_filter = atof(str);
3789         return 0;
3790 }
3791
3792 static int trace__set_filter_pids_from_option(const struct option *opt, const char *str,
3793                                               int unset __maybe_unused)
3794 {
3795         int ret = -1;
3796         size_t i;
3797         struct trace *trace = opt->value;
3798         /*
3799          * FIXME: introduce a intarray class, plain parse csv and create a
3800          * { int nr, int entries[] } struct...
3801          */
3802         struct intlist *list = intlist__new(str);
3803
3804         if (list == NULL)
3805                 return -1;
3806
3807         i = trace->filter_pids.nr = intlist__nr_entries(list) + 1;
3808         trace->filter_pids.entries = calloc(i, sizeof(pid_t));
3809
3810         if (trace->filter_pids.entries == NULL)
3811                 goto out;
3812
3813         trace->filter_pids.entries[0] = getpid();
3814
3815         for (i = 1; i < trace->filter_pids.nr; ++i)
3816                 trace->filter_pids.entries[i] = intlist__entry(list, i - 1)->i;
3817
3818         intlist__delete(list);
3819         ret = 0;
3820 out:
3821         return ret;
3822 }
3823
3824 static int trace__open_output(struct trace *trace, const char *filename)
3825 {
3826         struct stat st;
3827
3828         if (!stat(filename, &st) && st.st_size) {
3829                 char oldname[PATH_MAX];
3830
3831                 scnprintf(oldname, sizeof(oldname), "%s.old", filename);
3832                 unlink(oldname);
3833                 rename(filename, oldname);
3834         }
3835
3836         trace->output = fopen(filename, "w");
3837
3838         return trace->output == NULL ? -errno : 0;
3839 }
3840
3841 static int parse_pagefaults(const struct option *opt, const char *str,
3842                             int unset __maybe_unused)
3843 {
3844         int *trace_pgfaults = opt->value;
3845
3846         if (strcmp(str, "all") == 0)
3847                 *trace_pgfaults |= TRACE_PFMAJ | TRACE_PFMIN;
3848         else if (strcmp(str, "maj") == 0)
3849                 *trace_pgfaults |= TRACE_PFMAJ;
3850         else if (strcmp(str, "min") == 0)
3851                 *trace_pgfaults |= TRACE_PFMIN;
3852         else
3853                 return -1;
3854
3855         return 0;
3856 }
3857
3858 static void evlist__set_evsel_handler(struct evlist *evlist, void *handler)
3859 {
3860         struct evsel *evsel;
3861
3862         evlist__for_each_entry(evlist, evsel)
3863                 evsel->handler = handler;
3864 }
3865
3866 static int evlist__set_syscall_tp_fields(struct evlist *evlist)
3867 {
3868         struct evsel *evsel;
3869
3870         evlist__for_each_entry(evlist, evsel) {
3871                 if (evsel->priv || !evsel->tp_format)
3872                         continue;
3873
3874                 if (strcmp(evsel->tp_format->system, "syscalls"))
3875                         continue;
3876
3877                 if (perf_evsel__init_syscall_tp(evsel))
3878                         return -1;
3879
3880                 if (!strncmp(evsel->tp_format->name, "sys_enter_", 10)) {
3881                         struct syscall_tp *sc = evsel->priv;
3882
3883                         if (__tp_field__init_ptr(&sc->args, sc->id.offset + sizeof(u64)))
3884                                 return -1;
3885                 } else if (!strncmp(evsel->tp_format->name, "sys_exit_", 9)) {
3886                         struct syscall_tp *sc = evsel->priv;
3887
3888                         if (__tp_field__init_uint(&sc->ret, sizeof(u64), sc->id.offset + sizeof(u64), evsel->needs_swap))
3889                                 return -1;
3890                 }
3891         }
3892
3893         return 0;
3894 }
3895
3896 /*
3897  * XXX: Hackish, just splitting the combined -e+--event (syscalls
3898  * (raw_syscalls:{sys_{enter,exit}} + events (tracepoints, HW, SW, etc) to use
3899  * existing facilities unchanged (trace->ev_qualifier + parse_options()).
3900  *
3901  * It'd be better to introduce a parse_options() variant that would return a
3902  * list with the terms it didn't match to an event...
3903  */
3904 static int trace__parse_events_option(const struct option *opt, const char *str,
3905                                       int unset __maybe_unused)
3906 {
3907         struct trace *trace = (struct trace *)opt->value;
3908         const char *s = str;
3909         char *sep = NULL, *lists[2] = { NULL, NULL, };
3910         int len = strlen(str) + 1, err = -1, list, idx;
3911         char *strace_groups_dir = system_path(STRACE_GROUPS_DIR);
3912         char group_name[PATH_MAX];
3913         struct syscall_fmt *fmt;
3914
3915         if (strace_groups_dir == NULL)
3916                 return -1;
3917
3918         if (*s == '!') {
3919                 ++s;
3920                 trace->not_ev_qualifier = true;
3921         }
3922
3923         while (1) {
3924                 if ((sep = strchr(s, ',')) != NULL)
3925                         *sep = '\0';
3926
3927                 list = 0;
3928                 if (syscalltbl__id(trace->sctbl, s) >= 0 ||
3929                     syscalltbl__strglobmatch_first(trace->sctbl, s, &idx) >= 0) {
3930                         list = 1;
3931                         goto do_concat;
3932                 }
3933
3934                 fmt = syscall_fmt__find_by_alias(s);
3935                 if (fmt != NULL) {
3936                         list = 1;
3937                         s = fmt->name;
3938                 } else {
3939                         path__join(group_name, sizeof(group_name), strace_groups_dir, s);
3940                         if (access(group_name, R_OK) == 0)
3941                                 list = 1;
3942                 }
3943 do_concat:
3944                 if (lists[list]) {
3945                         sprintf(lists[list] + strlen(lists[list]), ",%s", s);
3946                 } else {
3947                         lists[list] = malloc(len);
3948                         if (lists[list] == NULL)
3949                                 goto out;
3950                         strcpy(lists[list], s);
3951                 }
3952
3953                 if (!sep)
3954                         break;
3955
3956                 *sep = ',';
3957                 s = sep + 1;
3958         }
3959
3960         if (lists[1] != NULL) {
3961                 struct strlist_config slist_config = {
3962                         .dirname = strace_groups_dir,
3963                 };
3964
3965                 trace->ev_qualifier = strlist__new(lists[1], &slist_config);
3966                 if (trace->ev_qualifier == NULL) {
3967                         fputs("Not enough memory to parse event qualifier", trace->output);
3968                         goto out;
3969                 }
3970
3971                 if (trace__validate_ev_qualifier(trace))
3972                         goto out;
3973                 trace->trace_syscalls = true;
3974         }
3975
3976         err = 0;
3977
3978         if (lists[0]) {
3979                 struct option o = OPT_CALLBACK('e', "event", &trace->evlist, "event",
3980                                                "event selector. use 'perf list' to list available events",
3981                                                parse_events_option);
3982                 err = parse_events_option(&o, lists[0], 0);
3983         }
3984 out:
3985         if (sep)
3986                 *sep = ',';
3987
3988         return err;
3989 }
3990
3991 static int trace__parse_cgroups(const struct option *opt, const char *str, int unset)
3992 {
3993         struct trace *trace = opt->value;
3994
3995         if (!list_empty(&trace->evlist->core.entries))
3996                 return parse_cgroups(opt, str, unset);
3997
3998         trace->cgroup = evlist__findnew_cgroup(trace->evlist, str);
3999
4000         return 0;
4001 }
4002
4003 static struct bpf_map *trace__find_bpf_map_by_name(struct trace *trace, const char *name)
4004 {
4005         if (trace->bpf_obj == NULL)
4006                 return NULL;
4007
4008         return bpf_object__find_map_by_name(trace->bpf_obj, name);
4009 }
4010
4011 static void trace__set_bpf_map_filtered_pids(struct trace *trace)
4012 {
4013         trace->filter_pids.map = trace__find_bpf_map_by_name(trace, "pids_filtered");
4014 }
4015
4016 static void trace__set_bpf_map_syscalls(struct trace *trace)
4017 {
4018         trace->syscalls.map = trace__find_bpf_map_by_name(trace, "syscalls");
4019         trace->syscalls.prog_array.sys_enter = trace__find_bpf_map_by_name(trace, "syscalls_sys_enter");
4020         trace->syscalls.prog_array.sys_exit  = trace__find_bpf_map_by_name(trace, "syscalls_sys_exit");
4021 }
4022
4023 static int trace__config(const char *var, const char *value, void *arg)
4024 {
4025         struct trace *trace = arg;
4026         int err = 0;
4027
4028         if (!strcmp(var, "trace.add_events")) {
4029                 struct option o = OPT_CALLBACK('e', "event", &trace->evlist, "event",
4030                                                "event selector. use 'perf list' to list available events",
4031                                                parse_events_option);
4032                 /*
4033                  * We can't propagate parse_event_option() return, as it is 1
4034                  * for failure while perf_config() expects -1.
4035                  */
4036                 if (parse_events_option(&o, value, 0))
4037                         err = -1;
4038         } else if (!strcmp(var, "trace.show_timestamp")) {
4039                 trace->show_tstamp = perf_config_bool(var, value);
4040         } else if (!strcmp(var, "trace.show_duration")) {
4041                 trace->show_duration = perf_config_bool(var, value);
4042         } else if (!strcmp(var, "trace.show_arg_names")) {
4043                 trace->show_arg_names = perf_config_bool(var, value);
4044                 if (!trace->show_arg_names)
4045                         trace->show_zeros = true;
4046         } else if (!strcmp(var, "trace.show_zeros")) {
4047                 bool new_show_zeros = perf_config_bool(var, value);
4048                 if (!trace->show_arg_names && !new_show_zeros) {
4049                         pr_warning("trace.show_zeros has to be set when trace.show_arg_names=no\n");
4050                         goto out;
4051                 }
4052                 trace->show_zeros = new_show_zeros;
4053         } else if (!strcmp(var, "trace.show_prefix")) {
4054                 trace->show_string_prefix = perf_config_bool(var, value);
4055         } else if (!strcmp(var, "trace.no_inherit")) {
4056                 trace->opts.no_inherit = perf_config_bool(var, value);
4057         } else if (!strcmp(var, "trace.args_alignment")) {
4058                 int args_alignment = 0;
4059                 if (perf_config_int(&args_alignment, var, value) == 0)
4060                         trace->args_alignment = args_alignment;
4061         }
4062 out:
4063         return err;
4064 }
4065
4066 int cmd_trace(int argc, const char **argv)
4067 {
4068         const char *trace_usage[] = {
4069                 "perf trace [<options>] [<command>]",
4070                 "perf trace [<options>] -- <command> [<options>]",
4071                 "perf trace record [<options>] [<command>]",
4072                 "perf trace record [<options>] -- <command> [<options>]",
4073                 NULL
4074         };
4075         struct trace trace = {
4076                 .opts = {
4077                         .target = {
4078                                 .uid       = UINT_MAX,
4079                                 .uses_mmap = true,
4080                         },
4081                         .user_freq     = UINT_MAX,
4082                         .user_interval = ULLONG_MAX,
4083                         .no_buffering  = true,
4084                         .mmap_pages    = UINT_MAX,
4085                 },
4086                 .output = stderr,
4087                 .show_comm = true,
4088                 .show_tstamp = true,
4089                 .show_duration = true,
4090                 .show_arg_names = true,
4091                 .args_alignment = 70,
4092                 .trace_syscalls = false,
4093                 .kernel_syscallchains = false,
4094                 .max_stack = UINT_MAX,
4095                 .max_events = ULONG_MAX,
4096         };
4097         const char *map_dump_str = NULL;
4098         const char *output_name = NULL;
4099         const struct option trace_options[] = {
4100         OPT_CALLBACK('e', "event", &trace, "event",
4101                      "event/syscall selector. use 'perf list' to list available events",
4102                      trace__parse_events_option),
4103         OPT_BOOLEAN(0, "comm", &trace.show_comm,
4104                     "show the thread COMM next to its id"),
4105         OPT_BOOLEAN(0, "tool_stats", &trace.show_tool_stats, "show tool stats"),
4106         OPT_CALLBACK(0, "expr", &trace, "expr", "list of syscalls/events to trace",
4107                      trace__parse_events_option),
4108         OPT_STRING('o', "output", &output_name, "file", "output file name"),
4109         OPT_STRING('i', "input", &input_name, "file", "Analyze events in file"),
4110         OPT_STRING('p', "pid", &trace.opts.target.pid, "pid",
4111                     "trace events on existing process id"),
4112         OPT_STRING('t', "tid", &trace.opts.target.tid, "tid",
4113                     "trace events on existing thread id"),
4114         OPT_CALLBACK(0, "filter-pids", &trace, "CSV list of pids",
4115                      "pids to filter (by the kernel)", trace__set_filter_pids_from_option),
4116         OPT_BOOLEAN('a', "all-cpus", &trace.opts.target.system_wide,
4117                     "system-wide collection from all CPUs"),
4118         OPT_STRING('C', "cpu", &trace.opts.target.cpu_list, "cpu",
4119                     "list of cpus to monitor"),
4120         OPT_BOOLEAN(0, "no-inherit", &trace.opts.no_inherit,
4121                     "child tasks do not inherit counters"),
4122         OPT_CALLBACK('m', "mmap-pages", &trace.opts.mmap_pages, "pages",
4123                      "number of mmap data pages",
4124                      perf_evlist__parse_mmap_pages),
4125         OPT_STRING('u', "uid", &trace.opts.target.uid_str, "user",
4126                    "user to profile"),
4127         OPT_CALLBACK(0, "duration", &trace, "float",
4128                      "show only events with duration > N.M ms",
4129                      trace__set_duration),
4130 #ifdef HAVE_LIBBPF_SUPPORT
4131         OPT_STRING(0, "map-dump", &map_dump_str, "BPF map", "BPF map to periodically dump"),
4132 #endif
4133         OPT_BOOLEAN(0, "sched", &trace.sched, "show blocking scheduler events"),
4134         OPT_INCR('v', "verbose", &verbose, "be more verbose"),
4135         OPT_BOOLEAN('T', "time", &trace.full_time,
4136                     "Show full timestamp, not time relative to first start"),
4137         OPT_BOOLEAN(0, "failure", &trace.failure_only,
4138                     "Show only syscalls that failed"),
4139         OPT_BOOLEAN('s', "summary", &trace.summary_only,
4140                     "Show only syscall summary with statistics"),
4141         OPT_BOOLEAN('S', "with-summary", &trace.summary,
4142                     "Show all syscalls and summary with statistics"),
4143         OPT_CALLBACK_DEFAULT('F', "pf", &trace.trace_pgfaults, "all|maj|min",
4144                      "Trace pagefaults", parse_pagefaults, "maj"),
4145         OPT_BOOLEAN(0, "syscalls", &trace.trace_syscalls, "Trace syscalls"),
4146         OPT_BOOLEAN('f', "force", &trace.force, "don't complain, do it"),
4147         OPT_CALLBACK(0, "call-graph", &trace.opts,
4148                      "record_mode[,record_size]", record_callchain_help,
4149                      &record_parse_callchain_opt),
4150         OPT_BOOLEAN(0, "kernel-syscall-graph", &trace.kernel_syscallchains,
4151                     "Show the kernel callchains on the syscall exit path"),
4152         OPT_ULONG(0, "max-events", &trace.max_events,
4153                 "Set the maximum number of events to print, exit after that is reached. "),
4154         OPT_UINTEGER(0, "min-stack", &trace.min_stack,
4155                      "Set the minimum stack depth when parsing the callchain, "
4156                      "anything below the specified depth will be ignored."),
4157         OPT_UINTEGER(0, "max-stack", &trace.max_stack,
4158                      "Set the maximum stack depth when parsing the callchain, "
4159                      "anything beyond the specified depth will be ignored. "
4160                      "Default: kernel.perf_event_max_stack or " __stringify(PERF_MAX_STACK_DEPTH)),
4161         OPT_BOOLEAN(0, "sort-events", &trace.sort_events,
4162                         "Sort batch of events before processing, use if getting out of order events"),
4163         OPT_BOOLEAN(0, "print-sample", &trace.print_sample,
4164                         "print the PERF_RECORD_SAMPLE PERF_SAMPLE_ info, for debugging"),
4165         OPT_UINTEGER(0, "proc-map-timeout", &proc_map_timeout,
4166                         "per thread proc mmap processing timeout in ms"),
4167         OPT_CALLBACK('G', "cgroup", &trace, "name", "monitor event in cgroup name only",
4168                      trace__parse_cgroups),
4169         OPT_UINTEGER('D', "delay", &trace.opts.initial_delay,
4170                      "ms to wait before starting measurement after program "
4171                      "start"),
4172         OPTS_EVSWITCH(&trace.evswitch),
4173         OPT_END()
4174         };
4175         bool __maybe_unused max_stack_user_set = true;
4176         bool mmap_pages_user_set = true;
4177         struct evsel *evsel;
4178         const char * const trace_subcommands[] = { "record", NULL };
4179         int err = -1;
4180         char bf[BUFSIZ];
4181
4182         signal(SIGSEGV, sighandler_dump_stack);
4183         signal(SIGFPE, sighandler_dump_stack);
4184
4185         trace.evlist = evlist__new();
4186         trace.sctbl = syscalltbl__new();
4187
4188         if (trace.evlist == NULL || trace.sctbl == NULL) {
4189                 pr_err("Not enough memory to run!\n");
4190                 err = -ENOMEM;
4191                 goto out;
4192         }
4193
4194         /*
4195          * Parsing .perfconfig may entail creating a BPF event, that may need
4196          * to create BPF maps, so bump RLIM_MEMLOCK as the default 64K setting
4197          * is too small. This affects just this process, not touching the
4198          * global setting. If it fails we'll get something in 'perf trace -v'
4199          * to help diagnose the problem.
4200          */
4201         rlimit__bump_memlock();
4202
4203         err = perf_config(trace__config, &trace);
4204         if (err)
4205                 goto out;
4206
4207         argc = parse_options_subcommand(argc, argv, trace_options, trace_subcommands,
4208                                  trace_usage, PARSE_OPT_STOP_AT_NON_OPTION);
4209
4210         if ((nr_cgroups || trace.cgroup) && !trace.opts.target.system_wide) {
4211                 usage_with_options_msg(trace_usage, trace_options,
4212                                        "cgroup monitoring only available in system-wide mode");
4213         }
4214
4215         evsel = bpf__setup_output_event(trace.evlist, "__augmented_syscalls__");
4216         if (IS_ERR(evsel)) {
4217                 bpf__strerror_setup_output_event(trace.evlist, PTR_ERR(evsel), bf, sizeof(bf));
4218                 pr_err("ERROR: Setup trace syscalls enter failed: %s\n", bf);
4219                 goto out;
4220         }
4221
4222         if (evsel) {
4223                 trace.syscalls.events.augmented = evsel;
4224
4225                 evsel = perf_evlist__find_tracepoint_by_name(trace.evlist, "raw_syscalls:sys_enter");
4226                 if (evsel == NULL) {
4227                         pr_err("ERROR: raw_syscalls:sys_enter not found in the augmented BPF object\n");
4228                         goto out;
4229                 }
4230
4231                 if (evsel->bpf_obj == NULL) {
4232                         pr_err("ERROR: raw_syscalls:sys_enter not associated to a BPF object\n");
4233                         goto out;
4234                 }
4235
4236                 trace.bpf_obj = evsel->bpf_obj;
4237
4238                 trace__set_bpf_map_filtered_pids(&trace);
4239                 trace__set_bpf_map_syscalls(&trace);
4240                 trace.syscalls.unaugmented_prog = trace__find_bpf_program_by_title(&trace, "!raw_syscalls:unaugmented");
4241         }
4242
4243         err = bpf__setup_stdout(trace.evlist);
4244         if (err) {
4245                 bpf__strerror_setup_stdout(trace.evlist, err, bf, sizeof(bf));
4246                 pr_err("ERROR: Setup BPF stdout failed: %s\n", bf);
4247                 goto out;
4248         }
4249
4250         err = -1;
4251
4252         if (map_dump_str) {
4253                 trace.dump.map = trace__find_bpf_map_by_name(&trace, map_dump_str);
4254                 if (trace.dump.map == NULL) {
4255                         pr_err("ERROR: BPF map \"%s\" not found\n", map_dump_str);
4256                         goto out;
4257                 }
4258         }
4259
4260         if (trace.trace_pgfaults) {
4261                 trace.opts.sample_address = true;
4262                 trace.opts.sample_time = true;
4263         }
4264
4265         if (trace.opts.mmap_pages == UINT_MAX)
4266                 mmap_pages_user_set = false;
4267
4268         if (trace.max_stack == UINT_MAX) {
4269                 trace.max_stack = input_name ? PERF_MAX_STACK_DEPTH : sysctl__max_stack();
4270                 max_stack_user_set = false;
4271         }
4272
4273 #ifdef HAVE_DWARF_UNWIND_SUPPORT
4274         if ((trace.min_stack || max_stack_user_set) && !callchain_param.enabled) {
4275                 record_opts__parse_callchain(&trace.opts, &callchain_param, "dwarf", false);
4276         }
4277 #endif
4278
4279         if (callchain_param.enabled) {
4280                 if (!mmap_pages_user_set && geteuid() == 0)
4281                         trace.opts.mmap_pages = perf_event_mlock_kb_in_pages() * 4;
4282
4283                 symbol_conf.use_callchain = true;
4284         }
4285
4286         if (trace.evlist->core.nr_entries > 0) {
4287                 evlist__set_evsel_handler(trace.evlist, trace__event_handler);
4288                 if (evlist__set_syscall_tp_fields(trace.evlist)) {
4289                         perror("failed to set syscalls:* tracepoint fields");
4290                         goto out;
4291                 }
4292         }
4293
4294         if (trace.sort_events) {
4295                 ordered_events__init(&trace.oe.data, ordered_events__deliver_event, &trace);
4296                 ordered_events__set_copy_on_queue(&trace.oe.data, true);
4297         }
4298
4299         /*
4300          * If we are augmenting syscalls, then combine what we put in the
4301          * __augmented_syscalls__ BPF map with what is in the
4302          * syscalls:sys_exit_FOO tracepoints, i.e. just like we do without BPF,
4303          * combining raw_syscalls:sys_enter with raw_syscalls:sys_exit.
4304          *
4305          * We'll switch to look at two BPF maps, one for sys_enter and the
4306          * other for sys_exit when we start augmenting the sys_exit paths with
4307          * buffers that are being copied from kernel to userspace, think 'read'
4308          * syscall.
4309          */
4310         if (trace.syscalls.events.augmented) {
4311                 evlist__for_each_entry(trace.evlist, evsel) {
4312                         bool raw_syscalls_sys_exit = strcmp(perf_evsel__name(evsel), "raw_syscalls:sys_exit") == 0;
4313
4314                         if (raw_syscalls_sys_exit) {
4315                                 trace.raw_augmented_syscalls = true;
4316                                 goto init_augmented_syscall_tp;
4317                         }
4318
4319                         if (trace.syscalls.events.augmented->priv == NULL &&
4320                             strstr(perf_evsel__name(evsel), "syscalls:sys_enter")) {
4321                                 struct evsel *augmented = trace.syscalls.events.augmented;
4322                                 if (perf_evsel__init_augmented_syscall_tp(augmented, evsel) ||
4323                                     perf_evsel__init_augmented_syscall_tp_args(augmented))
4324                                         goto out;
4325                                 /*
4326                                  * Augmented is __augmented_syscalls__ BPF_OUTPUT event
4327                                  * Above we made sure we can get from the payload the tp fields
4328                                  * that we get from syscalls:sys_enter tracefs format file.
4329                                  */
4330                                 augmented->handler = trace__sys_enter;
4331                                 /*
4332                                  * Now we do the same for the *syscalls:sys_enter event so that
4333                                  * if we handle it directly, i.e. if the BPF prog returns 0 so
4334                                  * as not to filter it, then we'll handle it just like we would
4335                                  * for the BPF_OUTPUT one:
4336                                  */
4337                                 if (perf_evsel__init_augmented_syscall_tp(evsel, evsel) ||
4338                                     perf_evsel__init_augmented_syscall_tp_args(evsel))
4339                                         goto out;
4340                                 evsel->handler = trace__sys_enter;
4341                         }
4342
4343                         if (strstarts(perf_evsel__name(evsel), "syscalls:sys_exit_")) {
4344                                 struct syscall_tp *sc;
4345 init_augmented_syscall_tp:
4346                                 if (perf_evsel__init_augmented_syscall_tp(evsel, evsel))
4347                                         goto out;
4348                                 sc = evsel->priv;
4349                                 /*
4350                                  * For now with BPF raw_augmented we hook into
4351                                  * raw_syscalls:sys_enter and there we get all
4352                                  * 6 syscall args plus the tracepoint common
4353                                  * fields and the syscall_nr (another long).
4354                                  * So we check if that is the case and if so
4355                                  * don't look after the sc->args_size but
4356                                  * always after the full raw_syscalls:sys_enter
4357                                  * payload, which is fixed.
4358                                  *
4359                                  * We'll revisit this later to pass
4360                                  * s->args_size to the BPF augmenter (now
4361                                  * tools/perf/examples/bpf/augmented_raw_syscalls.c,
4362                                  * so that it copies only what we need for each
4363                                  * syscall, like what happens when we use
4364                                  * syscalls:sys_enter_NAME, so that we reduce
4365                                  * the kernel/userspace traffic to just what is
4366                                  * needed for each syscall.
4367                                  */
4368                                 if (trace.raw_augmented_syscalls)
4369                                         trace.raw_augmented_syscalls_args_size = (6 + 1) * sizeof(long) + sc->id.offset;
4370                                 perf_evsel__init_augmented_syscall_tp_ret(evsel);
4371                                 evsel->handler = trace__sys_exit;
4372                         }
4373                 }
4374         }
4375
4376         if ((argc >= 1) && (strcmp(argv[0], "record") == 0))
4377                 return trace__record(&trace, argc-1, &argv[1]);
4378
4379         /* summary_only implies summary option, but don't overwrite summary if set */
4380         if (trace.summary_only)
4381                 trace.summary = trace.summary_only;
4382
4383         if (!trace.trace_syscalls && !trace.trace_pgfaults &&
4384             trace.evlist->core.nr_entries == 0 /* Was --events used? */) {
4385                 trace.trace_syscalls = true;
4386         }
4387
4388         if (output_name != NULL) {
4389                 err = trace__open_output(&trace, output_name);
4390                 if (err < 0) {
4391                         perror("failed to create output file");
4392                         goto out;
4393                 }
4394         }
4395
4396         err = evswitch__init(&trace.evswitch, trace.evlist, stderr);
4397         if (err)
4398                 goto out_close;
4399
4400         err = target__validate(&trace.opts.target);
4401         if (err) {
4402                 target__strerror(&trace.opts.target, err, bf, sizeof(bf));
4403                 fprintf(trace.output, "%s", bf);
4404                 goto out_close;
4405         }
4406
4407         err = target__parse_uid(&trace.opts.target);
4408         if (err) {
4409                 target__strerror(&trace.opts.target, err, bf, sizeof(bf));
4410                 fprintf(trace.output, "%s", bf);
4411                 goto out_close;
4412         }
4413
4414         if (!argc && target__none(&trace.opts.target))
4415                 trace.opts.target.system_wide = true;
4416
4417         if (input_name)
4418                 err = trace__replay(&trace);
4419         else
4420                 err = trace__run(&trace, argc, argv);
4421
4422 out_close:
4423         if (output_name != NULL)
4424                 fclose(trace.output);
4425 out:
4426         return err;
4427 }