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 "util/mmap.h"
  35 #include <subcmd/pager.h>
  36 #include <subcmd/exec-cmd.h>
  37 #include "util/machine.h"
  38 #include "util/map.h"
  39 #include "util/symbol.h"
  40 #include "util/path.h"
  41 #include "util/session.h"
  42 #include "util/thread.h"
  43 #include <subcmd/parse-options.h>
  44 #include "util/strlist.h"
  45 #include "util/intlist.h"
  46 #include "util/thread_map.h"
  47 #include "util/stat.h"
  48 #include "util/tool.h"
  49 #include "util/util.h"
  50 #include "trace/beauty/beauty.h"
  51 #include "trace-event.h"
  52 #include "util/parse-events.h"
  53 #include "util/bpf-loader.h"
  54 #include "callchain.h"
  55 #include "print_binary.h"
  56 #include "string2.h"
  57 #include "syscalltbl.h"
  58 #include "rb_resort.h"
  59 #include "../perf.h"
  60
  61 #include <errno.h>
  62 #include <inttypes.h>
  63 #include <poll.h>
  64 #include <signal.h>
  65 #include <stdlib.h>
  66 #include <string.h>
  67 #include <linux/err.h>
  68 #include <linux/filter.h>
  69 #include <linux/kernel.h>
  70 #include <linux/random.h>
  71 #include <linux/stringify.h>
  72 #include <linux/time64.h>
  73 #include <linux/zalloc.h>
  74 #include <fcntl.h>
  75 #include <sys/sysmacros.h>
  76
  77 #include <linux/ctype.h>
  78
  79 #ifndef O_CLOEXEC
  80 # define O_CLOEXEC              02000000
  81 #endif
  82
  83 #ifndef F_LINUX_SPECIFIC_BASE
  84 # define F_LINUX_SPECIFIC_BASE  1024
  85 #endif
  86
  87 struct trace {
  88         struct perf_tool        tool;
  89         struct syscalltbl       *sctbl;
  90         struct {
  91                 struct syscall  *table;
  92                 struct bpf_map  *map;
  93                 struct { // per syscall BPF_MAP_TYPE_PROG_ARRAY
  94                         struct bpf_map  *sys_enter,
  95                                         *sys_exit;
  96                 }               prog_array;
  97                 struct {
  98                         struct evsel *sys_enter,
  99                                           *sys_exit,
 100                                           *augmented;
 101                 }               events;
 102                 struct bpf_program *unaugmented_prog;
 103         } syscalls;
 104         struct {
 105                 struct bpf_map *map;
 106         } dump;
 107         struct record_opts      opts;
 108         struct evlist   *evlist;
 109         struct machine          *host;
 110         struct thread           *current;
 111         struct bpf_object       *bpf_obj;
 112         struct cgroup           *cgroup;
 113         u64                     base_time;
 114         FILE                    *output;
 115         unsigned long           nr_events;
 116         unsigned long           nr_events_printed;
 117         unsigned long           max_events;
 118         struct evswitch         evswitch;
 119         struct strlist          *ev_qualifier;
 120         struct {
 121                 size_t          nr;
 122                 int             *entries;
 123         }                       ev_qualifier_ids;
 124         struct {
 125                 size_t          nr;
 126                 pid_t           *entries;
 127                 struct bpf_map  *map;
 128         }                       filter_pids;
 129         double                  duration_filter;
 130         double                  runtime_ms;
 131         struct {
 132                 u64             vfs_getname,
 133                                 proc_getname;
 134         } stats;
 135         unsigned int            max_stack;
 136         unsigned int            min_stack;
 137         int                     raw_augmented_syscalls_args_size;
 138         bool                    raw_augmented_syscalls;
 139         bool                    fd_path_disabled;
 140         bool                    sort_events;
 141         bool                    not_ev_qualifier;
 142         bool                    live;
 143         bool                    full_time;
 144         bool                    sched;
 145         bool                    multiple_threads;
 146         bool                    summary;
 147         bool                    summary_only;
 148         bool                    failure_only;
 149         bool                    show_comm;
 150         bool                    print_sample;
 151         bool                    show_tool_stats;
 152         bool                    trace_syscalls;
 153         bool                    kernel_syscallchains;
 154         s16                     args_alignment;
 155         bool                    show_tstamp;
 156         bool                    show_duration;
 157         bool                    show_zeros;
 158         bool                    show_arg_names;
 159         bool                    show_string_prefix;
 160         bool                    force;
 161         bool                    vfs_getname;
 162         int                     trace_pgfaults;
 163         struct {
 164                 struct ordered_events   data;
 165                 u64                     last;
 166         } oe;
 167 };
 168
 169 struct tp_field {
 170         int offset;
 171         union {
 172                 u64 (*integer)(struct tp_field *field, struct perf_sample *sample);
 173                 void *(*pointer)(struct tp_field *field, struct perf_sample *sample);
 174         };
 175 };
 176
 177 #define TP_UINT_FIELD(bits) \
 178 static u64 tp_field__u##bits(struct tp_field *field, struct perf_sample *sample) \
 179 { \
 180         u##bits value; \
 181         memcpy(&value, sample->raw_data + field->offset, sizeof(value)); \
 182         return value;  \
 183 }
 184
 185 TP_UINT_FIELD(8);
 186 TP_UINT_FIELD(16);
 187 TP_UINT_FIELD(32);
 188 TP_UINT_FIELD(64);
 189
 190 #define TP_UINT_FIELD__SWAPPED(bits) \
 191 static u64 tp_field__swapped_u##bits(struct tp_field *field, struct perf_sample *sample) \
 192 { \
 193         u##bits value; \
 194         memcpy(&value, sample->raw_data + field->offset, sizeof(value)); \
 195         return bswap_##bits(value);\
 196 }
 197
 198 TP_UINT_FIELD__SWAPPED(16);
 199 TP_UINT_FIELD__SWAPPED(32);
 200 TP_UINT_FIELD__SWAPPED(64);
 201
 202 static int __tp_field__init_uint(struct tp_field *field, int size, int offset, bool needs_swap)
 203 {
 204         field->offset = offset;
 205
 206         switch (size) {
 207         case 1:
 208                 field->integer = tp_field__u8;
 209                 break;
 210         case 2:
 211                 field->integer = needs_swap ? tp_field__swapped_u16 : tp_field__u16;
 212                 break;
 213         case 4:
 214                 field->integer = needs_swap ? tp_field__swapped_u32 : tp_field__u32;
 215                 break;
 216         case 8:
 217                 field->integer = needs_swap ? tp_field__swapped_u64 : tp_field__u64;
 218                 break;
 219         default:
 220                 return -1;
 221         }
 222
 223         return 0;
 224 }
 225
 226 static int tp_field__init_uint(struct tp_field *field, struct tep_format_field *format_field, bool needs_swap)
 227 {
 228         return __tp_field__init_uint(field, format_field->size, format_field->offset, needs_swap);
 229 }
 230
 231 static void *tp_field__ptr(struct tp_field *field, struct perf_sample *sample)
 232 {
 233         return sample->raw_data + field->offset;
 234 }
 235
 236 static int __tp_field__init_ptr(struct tp_field *field, int offset)
 237 {
 238         field->offset = offset;
 239         field->pointer = tp_field__ptr;
 240         return 0;
 241 }
 242
 243 static int tp_field__init_ptr(struct tp_field *field, struct tep_format_field *format_field)
 244 {
 245         return __tp_field__init_ptr(field, format_field->offset);
 246 }
 247
 248 struct syscall_tp {
 249         struct tp_field id;
 250         union {
 251                 struct tp_field args, ret;
 252         };
 253 };
 254
 255 static int perf_evsel__init_tp_uint_field(struct evsel *evsel,
 256                                           struct tp_field *field,
 257                                           const char *name)
 258 {
 259         struct tep_format_field *format_field = perf_evsel__field(evsel, name);
 260
 261         if (format_field == NULL)
 262                 return -1;
 263
 264         return tp_field__init_uint(field, format_field, evsel->needs_swap);
 265 }
 266
 267 #define perf_evsel__init_sc_tp_uint_field(evsel, name) \
 268         ({ struct syscall_tp *sc = evsel->priv;\
 269            perf_evsel__init_tp_uint_field(evsel, &sc->name, #name); })
 270
 271 static int perf_evsel__init_tp_ptr_field(struct evsel *evsel,
 272                                          struct tp_field *field,
 273                                          const char *name)
 274 {
 275         struct tep_format_field *format_field = perf_evsel__field(evsel, name);
 276
 277         if (format_field == NULL)
 278                 return -1;
 279
 280         return tp_field__init_ptr(field, format_field);
 281 }
 282
 283 #define perf_evsel__init_sc_tp_ptr_field(evsel, name) \
 284         ({ struct syscall_tp *sc = evsel->priv;\
 285            perf_evsel__init_tp_ptr_field(evsel, &sc->name, #name); })
 286
 287 static void evsel__delete_priv(struct evsel *evsel)
 288 {
 289         zfree(&evsel->priv);
 290         evsel__delete(evsel);
 291 }
 292
 293 static int perf_evsel__init_syscall_tp(struct evsel *evsel)
 294 {
 295         struct syscall_tp *sc = evsel->priv = malloc(sizeof(struct syscall_tp));
 296
 297         if (evsel->priv != NULL) {
 298                 if (perf_evsel__init_tp_uint_field(evsel, &sc->id, "__syscall_nr") &&
 299                     perf_evsel__init_tp_uint_field(evsel, &sc->id, "nr"))
 300                         goto out_delete;
 301                 return 0;
 302         }
 303
 304         return -ENOMEM;
 305 out_delete:
 306         zfree(&evsel->priv);
 307         return -ENOENT;
 308 }
 309
 310 static int perf_evsel__init_augmented_syscall_tp(struct evsel *evsel, struct evsel *tp)
 311 {
 312         struct syscall_tp *sc = evsel->priv = malloc(sizeof(struct syscall_tp));
 313
 314         if (evsel->priv != NULL) {
 315                 struct tep_format_field *syscall_id = perf_evsel__field(tp, "id");
 316                 if (syscall_id == NULL)
 317                         syscall_id = perf_evsel__field(tp, "__syscall_nr");
 318                 if (syscall_id == NULL)
 319                         goto out_delete;
 320                 if (__tp_field__init_uint(&sc->id, syscall_id->size, syscall_id->offset, evsel->needs_swap))
 321                         goto out_delete;
 322
 323                 return 0;
 324         }
 325
 326         return -ENOMEM;
 327 out_delete:
 328         zfree(&evsel->priv);
 329         return -EINVAL;
 330 }
 331
 332 static int perf_evsel__init_augmented_syscall_tp_args(struct evsel *evsel)
 333 {
 334         struct syscall_tp *sc = evsel->priv;
 335
 336         return __tp_field__init_ptr(&sc->args, sc->id.offset + sizeof(u64));
 337 }
 338
 339 static int perf_evsel__init_augmented_syscall_tp_ret(struct evsel *evsel)
 340 {
 341         struct syscall_tp *sc = evsel->priv;
 342
 343         return __tp_field__init_uint(&sc->ret, sizeof(u64), sc->id.offset + sizeof(u64), evsel->needs_swap);
 344 }
 345
 346 static int perf_evsel__init_raw_syscall_tp(struct evsel *evsel, void *handler)
 347 {
 348         evsel->priv = malloc(sizeof(struct syscall_tp));
 349         if (evsel->priv != NULL) {
 350                 if (perf_evsel__init_sc_tp_uint_field(evsel, id))
 351                         goto out_delete;
 352
 353                 evsel->handler = handler;
 354                 return 0;
 355         }
 356
 357         return -ENOMEM;
 358
 359 out_delete:
 360         zfree(&evsel->priv);
 361         return -ENOENT;
 362 }
 363
 364 static struct evsel *perf_evsel__raw_syscall_newtp(const char *direction, void *handler)
 365 {
 366         struct evsel *evsel = perf_evsel__newtp("raw_syscalls", direction);
 367
 368         /* older kernel (e.g., RHEL6) use syscalls:{enter,exit} */
 369         if (IS_ERR(evsel))
 370                 evsel = perf_evsel__newtp("syscalls", direction);
 371
 372         if (IS_ERR(evsel))
 373                 return NULL;
 374
 375         if (perf_evsel__init_raw_syscall_tp(evsel, handler))
 376                 goto out_delete;
 377
 378         return evsel;
 379
 380 out_delete:
 381         evsel__delete_priv(evsel);
 382         return NULL;
 383 }
 384
 385 #define perf_evsel__sc_tp_uint(evsel, name, sample) \
 386         ({ struct syscall_tp *fields = evsel->priv; \
 387            fields->name.integer(&fields->name, sample); })
 388
 389 #define perf_evsel__sc_tp_ptr(evsel, name, sample) \
 390         ({ struct syscall_tp *fields = evsel->priv; \
 391            fields->name.pointer(&fields->name, sample); })
 392
 393 size_t strarray__scnprintf(struct strarray *sa, char *bf, size_t size, const char *intfmt, bool show_prefix, int val)
 394 {
 395         int idx = val - sa->offset;
 396
 397         if (idx < 0 || idx >= sa->nr_entries || sa->entries[idx] == NULL) {
 398                 size_t printed = scnprintf(bf, size, intfmt, val);
 399                 if (show_prefix)
 400                         printed += scnprintf(bf + printed, size - printed, " /* %s??? */", sa->prefix);
 401                 return printed;
 402         }
 403
 404         return scnprintf(bf, size, "%s%s", show_prefix ? sa->prefix : "", sa->entries[idx]);
 405 }
 406
 407 static size_t __syscall_arg__scnprintf_strarray(char *bf, size_t size,
 408                                                 const char *intfmt,
 409                                                 struct syscall_arg *arg)
 410 {
 411         return strarray__scnprintf(arg->parm, bf, size, intfmt, arg->show_string_prefix, arg->val);
 412 }
 413
 414 static size_t syscall_arg__scnprintf_strarray(char *bf, size_t size,
 415                                               struct syscall_arg *arg)
 416 {
 417         return __syscall_arg__scnprintf_strarray(bf, size, "%d", arg);
 418 }
 419
 420 #define SCA_STRARRAY syscall_arg__scnprintf_strarray
 421
 422 size_t syscall_arg__scnprintf_strarray_flags(char *bf, size_t size, struct syscall_arg *arg)
 423 {
 424         return strarray__scnprintf_flags(arg->parm, bf, size, arg->show_string_prefix, arg->val);
 425 }
 426
 427 size_t strarrays__scnprintf(struct strarrays *sas, char *bf, size_t size, const char *intfmt, bool show_prefix, int val)
 428 {
 429         size_t printed;
 430         int i;
 431
 432         for (i = 0; i < sas->nr_entries; ++i) {
 433                 struct strarray *sa = sas->entries[i];
 434                 int idx = val - sa->offset;
 435
 436                 if (idx >= 0 && idx < sa->nr_entries) {
 437                         if (sa->entries[idx] == NULL)
 438                                 break;
 439                         return scnprintf(bf, size, "%s%s", show_prefix ? sa->prefix : "", sa->entries[idx]);
 440                 }
 441         }
 442
 443         printed = scnprintf(bf, size, intfmt, val);
 444         if (show_prefix)
 445                 printed += scnprintf(bf + printed, size - printed, " /* %s??? */", sas->entries[0]->prefix);
 446         return printed;
 447 }
 448
 449 size_t syscall_arg__scnprintf_strarrays(char *bf, size_t size,
 450                                         struct syscall_arg *arg)
 451 {
 452         return strarrays__scnprintf(arg->parm, bf, size, "%d", arg->show_string_prefix, arg->val);
 453 }
 454
 455 #ifndef AT_FDCWD
 456 #define AT_FDCWD        -100
 457 #endif
 458
 459 static size_t syscall_arg__scnprintf_fd_at(char *bf, size_t size,
 460                                            struct syscall_arg *arg)
 461 {
 462         int fd = arg->val;
 463         const char *prefix = "AT_FD";
 464
 465         if (fd == AT_FDCWD)
 466                 return scnprintf(bf, size, "%s%s", arg->show_string_prefix ? prefix : "", "CWD");
 467
 468         return syscall_arg__scnprintf_fd(bf, size, arg);
 469 }
 470
 471 #define SCA_FDAT syscall_arg__scnprintf_fd_at
 472
 473 static size_t syscall_arg__scnprintf_close_fd(char *bf, size_t size,
 474                                               struct syscall_arg *arg);
 475
 476 #define SCA_CLOSE_FD syscall_arg__scnprintf_close_fd
 477
 478 size_t syscall_arg__scnprintf_hex(char *bf, size_t size, struct syscall_arg *arg)
 479 {
 480         return scnprintf(bf, size, "%#lx", arg->val);
 481 }
 482
 483 size_t syscall_arg__scnprintf_ptr(char *bf, size_t size, struct syscall_arg *arg)
 484 {
 485         if (arg->val == 0)
 486                 return scnprintf(bf, size, "NULL");
 487         return syscall_arg__scnprintf_hex(bf, size, arg);
 488 }
 489
 490 size_t syscall_arg__scnprintf_int(char *bf, size_t size, struct syscall_arg *arg)
 491 {
 492         return scnprintf(bf, size, "%d", arg->val);
 493 }
 494
 495 size_t syscall_arg__scnprintf_long(char *bf, size_t size, struct syscall_arg *arg)
 496 {
 497         return scnprintf(bf, size, "%ld", arg->val);
 498 }
 499
 500 static const char *bpf_cmd[] = {
 501         "MAP_CREATE", "MAP_LOOKUP_ELEM", "MAP_UPDATE_ELEM", "MAP_DELETE_ELEM",
 502         "MAP_GET_NEXT_KEY", "PROG_LOAD",
 503 };
 504 static DEFINE_STRARRAY(bpf_cmd, "BPF_");
 505
 506 static const char *fsmount_flags[] = {
 507         [1] = "CLOEXEC",
 508 };
 509 static DEFINE_STRARRAY(fsmount_flags, "FSMOUNT_");
 510
 511 #include "trace/beauty/generated/fsconfig_arrays.c"
 512
 513 static DEFINE_STRARRAY(fsconfig_cmds, "FSCONFIG_");
 514
 515 static const char *epoll_ctl_ops[] = { "ADD", "DEL", "MOD", };
 516 static DEFINE_STRARRAY_OFFSET(epoll_ctl_ops, "EPOLL_CTL_", 1);
 517
 518 static const char *itimers[] = { "REAL", "VIRTUAL", "PROF", };
 519 static DEFINE_STRARRAY(itimers, "ITIMER_");
 520
 521 static const char *keyctl_options[] = {
 522         "GET_KEYRING_ID", "JOIN_SESSION_KEYRING", "UPDATE", "REVOKE", "CHOWN",
 523         "SETPERM", "DESCRIBE", "CLEAR", "LINK", "UNLINK", "SEARCH", "READ",
 524         "INSTANTIATE", "NEGATE", "SET_REQKEY_KEYRING", "SET_TIMEOUT",
 525         "ASSUME_AUTHORITY", "GET_SECURITY", "SESSION_TO_PARENT", "REJECT",
 526         "INSTANTIATE_IOV", "INVALIDATE", "GET_PERSISTENT",
 527 };
 528 static DEFINE_STRARRAY(keyctl_options, "KEYCTL_");
 529
 530 static const char *whences[] = { "SET", "CUR", "END",
 531 #ifdef SEEK_DATA
 532 "DATA",
 533 #endif
 534 #ifdef SEEK_HOLE
 535 "HOLE",
 536 #endif
 537 };
 538 static DEFINE_STRARRAY(whences, "SEEK_");
 539
 540 static const char *fcntl_cmds[] = {
 541         "DUPFD", "GETFD", "SETFD", "GETFL", "SETFL", "GETLK", "SETLK",
 542         "SETLKW", "SETOWN", "GETOWN", "SETSIG", "GETSIG", "GETLK64",
 543         "SETLK64", "SETLKW64", "SETOWN_EX", "GETOWN_EX",
 544         "GETOWNER_UIDS",
 545 };
 546 static DEFINE_STRARRAY(fcntl_cmds, "F_");
 547
 548 static const char *fcntl_linux_specific_cmds[] = {
 549         "SETLEASE", "GETLEASE", "NOTIFY", [5] = "CANCELLK", "DUPFD_CLOEXEC",
 550         "SETPIPE_SZ", "GETPIPE_SZ", "ADD_SEALS", "GET_SEALS",
 551         "GET_RW_HINT", "SET_RW_HINT", "GET_FILE_RW_HINT", "SET_FILE_RW_HINT",
 552 };
 553
 554 static DEFINE_STRARRAY_OFFSET(fcntl_linux_specific_cmds, "F_", F_LINUX_SPECIFIC_BASE);
 555
 556 static struct strarray *fcntl_cmds_arrays[] = {
 557         &strarray__fcntl_cmds,
 558         &strarray__fcntl_linux_specific_cmds,
 559 };
 560
 561 static DEFINE_STRARRAYS(fcntl_cmds_arrays);
 562
 563 static const char *rlimit_resources[] = {
 564         "CPU", "FSIZE", "DATA", "STACK", "CORE", "RSS", "NPROC", "NOFILE",
 565         "MEMLOCK", "AS", "LOCKS", "SIGPENDING", "MSGQUEUE", "NICE", "RTPRIO",
 566         "RTTIME",
 567 };
 568 static DEFINE_STRARRAY(rlimit_resources, "RLIMIT_");
 569
 570 static const char *sighow[] = { "BLOCK", "UNBLOCK", "SETMASK", };
 571 static DEFINE_STRARRAY(sighow, "SIG_");
 572
 573 static const char *clockid[] = {
 574         "REALTIME", "MONOTONIC", "PROCESS_CPUTIME_ID", "THREAD_CPUTIME_ID",
 575         "MONOTONIC_RAW", "REALTIME_COARSE", "MONOTONIC_COARSE", "BOOTTIME",
 576         "REALTIME_ALARM", "BOOTTIME_ALARM", "SGI_CYCLE", "TAI"
 577 };
 578 static DEFINE_STRARRAY(clockid, "CLOCK_");
 579
 580 static size_t syscall_arg__scnprintf_access_mode(char *bf, size_t size,
 581                                                  struct syscall_arg *arg)
 582 {
 583         bool show_prefix = arg->show_string_prefix;
 584         const char *suffix = "_OK";
 585         size_t printed = 0;
 586         int mode = arg->val;
 587
 588         if (mode == F_OK) /* 0 */
 589                 return scnprintf(bf, size, "F%s", show_prefix ? suffix : "");
 590 #define P_MODE(n) \
 591         if (mode & n##_OK) { \
 592                 printed += scnprintf(bf + printed, size - printed, "%s%s", #n, show_prefix ? suffix : ""); \
 593                 mode &= ~n##_OK; \
 594         }
 595
 596         P_MODE(R);
 597         P_MODE(W);
 598         P_MODE(X);
 599 #undef P_MODE
 600
 601         if (mode)
 602                 printed += scnprintf(bf + printed, size - printed, "|%#x", mode);
 603
 604         return printed;
 605 }
 606
 607 #define SCA_ACCMODE syscall_arg__scnprintf_access_mode
 608
 609 static size_t syscall_arg__scnprintf_filename(char *bf, size_t size,
 610                                               struct syscall_arg *arg);
 611
 612 #define SCA_FILENAME syscall_arg__scnprintf_filename
 613
 614 static size_t syscall_arg__scnprintf_pipe_flags(char *bf, size_t size,
 615                                                 struct syscall_arg *arg)
 616 {
 617         bool show_prefix = arg->show_string_prefix;
 618         const char *prefix = "O_";
 619         int printed = 0, flags = arg->val;
 620
 621 #define P_FLAG(n) \
 622         if (flags & O_##n) { \
 623                 printed += scnprintf(bf + printed, size - printed, "%s%s%s", printed ? "|" : "", show_prefix ? prefix : "", #n); \
 624                 flags &= ~O_##n; \
 625         }
 626
 627         P_FLAG(CLOEXEC);
 628         P_FLAG(NONBLOCK);
 629 #undef P_FLAG
 630
 631         if (flags)
 632                 printed += scnprintf(bf + printed, size - printed, "%s%#x", printed ? "|" : "", flags);
 633
 634         return printed;
 635 }
 636
 637 #define SCA_PIPE_FLAGS syscall_arg__scnprintf_pipe_flags
 638
 639 #ifndef GRND_NONBLOCK
 640 #define GRND_NONBLOCK   0x0001
 641 #endif
 642 #ifndef GRND_RANDOM
 643 #define GRND_RANDOM     0x0002
 644 #endif
 645
 646 static size_t syscall_arg__scnprintf_getrandom_flags(char *bf, size_t size,
 647                                                    struct syscall_arg *arg)
 648 {
 649         bool show_prefix = arg->show_string_prefix;
 650         const char *prefix = "GRND_";
 651         int printed = 0, flags = arg->val;
 652
 653 #define P_FLAG(n) \
 654         if (flags & GRND_##n) { \
 655                 printed += scnprintf(bf + printed, size - printed, "%s%s%s", printed ? "|" : "", show_prefix ? prefix : "", #n); \
 656                 flags &= ~GRND_##n; \
 657         }
 658
 659         P_FLAG(RANDOM);
 660         P_FLAG(NONBLOCK);
 661 #undef P_FLAG
 662
 663         if (flags)
 664                 printed += scnprintf(bf + printed, size - printed, "%s%#x", printed ? "|" : "", flags);
 665
 666         return printed;
 667 }
 668
 669 #define SCA_GETRANDOM_FLAGS syscall_arg__scnprintf_getrandom_flags
 670
 671 #define STRARRAY(name, array) \
 672           { .scnprintf  = SCA_STRARRAY, \
 673             .parm       = &strarray__##array, }
 674
 675 #define STRARRAY_FLAGS(name, array) \
 676           { .scnprintf  = SCA_STRARRAY_FLAGS, \
 677             .parm       = &strarray__##array, }
 678
 679 #include "trace/beauty/arch_errno_names.c"
 680 #include "trace/beauty/eventfd.c"
 681 #include "trace/beauty/futex_op.c"
 682 #include "trace/beauty/futex_val3.c"
 683 #include "trace/beauty/mmap.c"
 684 #include "trace/beauty/mode_t.c"
 685 #include "trace/beauty/msg_flags.c"
 686 #include "trace/beauty/open_flags.c"
 687 #include "trace/beauty/perf_event_open.c"
 688 #include "trace/beauty/pid.c"
 689 #include "trace/beauty/sched_policy.c"
 690 #include "trace/beauty/seccomp.c"
 691 #include "trace/beauty/signum.c"
 692 #include "trace/beauty/socket_type.c"
 693 #include "trace/beauty/waitid_options.c"
 694
 695 struct syscall_arg_fmt {
 696         size_t     (*scnprintf)(char *bf, size_t size, struct syscall_arg *arg);
 697         unsigned long (*mask_val)(struct syscall_arg *arg, unsigned long val);
 698         void       *parm;
 699         const char *name;
 700         bool       show_zero;
 701 };
 702
 703 static struct syscall_fmt {
 704         const char *name;
 705         const char *alias;
 706         struct {
 707                 const char *sys_enter,
 708                            *sys_exit;
 709         }          bpf_prog_name;
 710         struct syscall_arg_fmt arg[6];
 711         u8         nr_args;
 712         bool       errpid;
 713         bool       timeout;
 714         bool       hexret;
 715 } syscall_fmts[] = {
 716         { .name     = "access",
 717           .arg = { [1] = { .scnprintf = SCA_ACCMODE,  /* mode */ }, }, },
 718         { .name     = "arch_prctl",
 719           .arg = { [0] = { .scnprintf = SCA_X86_ARCH_PRCTL_CODE, /* code */ },
 720                    [1] = { .scnprintf = SCA_PTR, /* arg2 */ }, }, },
 721         { .name     = "bind",
 722           .arg = { [0] = { .scnprintf = SCA_INT, /* fd */ },
 723                    [1] = { .scnprintf = SCA_SOCKADDR, /* umyaddr */ },
 724                    [2] = { .scnprintf = SCA_INT, /* addrlen */ }, }, },
 725         { .name     = "bpf",
 726           .arg = { [0] = STRARRAY(cmd, bpf_cmd), }, },
 727         { .name     = "brk",        .hexret = true,
 728           .arg = { [0] = { .scnprintf = SCA_PTR, /* brk */ }, }, },
 729         { .name     = "clock_gettime",
 730           .arg = { [0] = STRARRAY(clk_id, clockid), }, },
 731         { .name     = "clone",      .errpid = true, .nr_args = 5,
 732           .arg = { [0] = { .name = "flags",         .scnprintf = SCA_CLONE_FLAGS, },
 733                    [1] = { .name = "child_stack",   .scnprintf = SCA_HEX, },
 734                    [2] = { .name = "parent_tidptr", .scnprintf = SCA_HEX, },
 735                    [3] = { .name = "child_tidptr",  .scnprintf = SCA_HEX, },
 736                    [4] = { .name = "tls",           .scnprintf = SCA_HEX, }, }, },
 737         { .name     = "close",
 738           .arg = { [0] = { .scnprintf = SCA_CLOSE_FD, /* fd */ }, }, },
 739         { .name     = "connect",
 740           .arg = { [0] = { .scnprintf = SCA_INT, /* fd */ },
 741                    [1] = { .scnprintf = SCA_SOCKADDR, /* servaddr */ },
 742                    [2] = { .scnprintf = SCA_INT, /* addrlen */ }, }, },
 743         { .name     = "epoll_ctl",
 744           .arg = { [1] = STRARRAY(op, epoll_ctl_ops), }, },
 745         { .name     = "eventfd2",
 746           .arg = { [1] = { .scnprintf = SCA_EFD_FLAGS, /* flags */ }, }, },
 747         { .name     = "fchmodat",
 748           .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
 749         { .name     = "fchownat",
 750           .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
 751         { .name     = "fcntl",
 752           .arg = { [1] = { .scnprintf = SCA_FCNTL_CMD, /* cmd */
 753                            .parm      = &strarrays__fcntl_cmds_arrays,
 754                            .show_zero = true, },
 755                    [2] = { .scnprintf =  SCA_FCNTL_ARG, /* arg */ }, }, },
 756         { .name     = "flock",
 757           .arg = { [1] = { .scnprintf = SCA_FLOCK, /* cmd */ }, }, },
 758         { .name     = "fsconfig",
 759           .arg = { [1] = STRARRAY(cmd, fsconfig_cmds), }, },
 760         { .name     = "fsmount",
 761           .arg = { [1] = STRARRAY_FLAGS(flags, fsmount_flags),
 762                    [2] = { .scnprintf = SCA_FSMOUNT_ATTR_FLAGS, /* attr_flags */ }, }, },
 763         { .name     = "fspick",
 764           .arg = { [0] = { .scnprintf = SCA_FDAT,         /* dfd */ },
 765                    [1] = { .scnprintf = SCA_FILENAME,     /* path */ },
 766                    [2] = { .scnprintf = SCA_FSPICK_FLAGS, /* flags */ }, }, },
 767         { .name     = "fstat", .alias = "newfstat", },
 768         { .name     = "fstatat", .alias = "newfstatat", },
 769         { .name     = "futex",
 770           .arg = { [1] = { .scnprintf = SCA_FUTEX_OP, /* op */ },
 771                    [5] = { .scnprintf = SCA_FUTEX_VAL3, /* val3 */ }, }, },
 772         { .name     = "futimesat",
 773           .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
 774         { .name     = "getitimer",
 775           .arg = { [0] = STRARRAY(which, itimers), }, },
 776         { .name     = "getpid",     .errpid = true, },
 777         { .name     = "getpgid",    .errpid = true, },
 778         { .name     = "getppid",    .errpid = true, },
 779         { .name     = "getrandom",
 780           .arg = { [2] = { .scnprintf = SCA_GETRANDOM_FLAGS, /* flags */ }, }, },
 781         { .name     = "getrlimit",
 782           .arg = { [0] = STRARRAY(resource, rlimit_resources), }, },
 783         { .name     = "gettid",     .errpid = true, },
 784         { .name     = "ioctl",
 785           .arg = {
 786 #if defined(__i386__) || defined(__x86_64__)
 787 /*
 788  * FIXME: Make this available to all arches.
 789  */
 790                    [1] = { .scnprintf = SCA_IOCTL_CMD, /* cmd */ },
 791                    [2] = { .scnprintf = SCA_HEX, /* arg */ }, }, },
 792 #else
 793                    [2] = { .scnprintf = SCA_HEX, /* arg */ }, }, },
 794 #endif
 795         { .name     = "kcmp",       .nr_args = 5,
 796           .arg = { [0] = { .name = "pid1",      .scnprintf = SCA_PID, },
 797                    [1] = { .name = "pid2",      .scnprintf = SCA_PID, },
 798                    [2] = { .name = "type",      .scnprintf = SCA_KCMP_TYPE, },
 799                    [3] = { .name = "idx1",      .scnprintf = SCA_KCMP_IDX, },
 800                    [4] = { .name = "idx2",      .scnprintf = SCA_KCMP_IDX, }, }, },
 801         { .name     = "keyctl",
 802           .arg = { [0] = STRARRAY(option, keyctl_options), }, },
 803         { .name     = "kill",
 804           .arg = { [1] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
 805         { .name     = "linkat",
 806           .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
 807         { .name     = "lseek",
 808           .arg = { [2] = STRARRAY(whence, whences), }, },
 809         { .name     = "lstat", .alias = "newlstat", },
 810         { .name     = "madvise",
 811           .arg = { [0] = { .scnprintf = SCA_HEX,      /* start */ },
 812                    [2] = { .scnprintf = SCA_MADV_BHV, /* behavior */ }, }, },
 813         { .name     = "mkdirat",
 814           .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
 815         { .name     = "mknodat",
 816           .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
 817         { .name     = "mmap",       .hexret = true,
 818 /* The standard mmap maps to old_mmap on s390x */
 819 #if defined(__s390x__)
 820         .alias = "old_mmap",
 821 #endif
 822           .arg = { [2] = { .scnprintf = SCA_MMAP_PROT,  /* prot */ },
 823                    [3] = { .scnprintf = SCA_MMAP_FLAGS, /* flags */ },
 824                    [5] = { .scnprintf = SCA_HEX,        /* offset */ }, }, },
 825         { .name     = "mount",
 826           .arg = { [0] = { .scnprintf = SCA_FILENAME, /* dev_name */ },
 827                    [3] = { .scnprintf = SCA_MOUNT_FLAGS, /* flags */
 828                            .mask_val  = SCAMV_MOUNT_FLAGS, /* flags */ }, }, },
 829         { .name     = "move_mount",
 830           .arg = { [0] = { .scnprintf = SCA_FDAT,       /* from_dfd */ },
 831                    [1] = { .scnprintf = SCA_FILENAME, /* from_pathname */ },
 832                    [2] = { .scnprintf = SCA_FDAT,       /* to_dfd */ },
 833                    [3] = { .scnprintf = SCA_FILENAME, /* to_pathname */ },
 834                    [4] = { .scnprintf = SCA_MOVE_MOUNT_FLAGS, /* flags */ }, }, },
 835         { .name     = "mprotect",
 836           .arg = { [0] = { .scnprintf = SCA_HEX,        /* start */ },
 837                    [2] = { .scnprintf = SCA_MMAP_PROT,  /* prot */ }, }, },
 838         { .name     = "mq_unlink",
 839           .arg = { [0] = { .scnprintf = SCA_FILENAME, /* u_name */ }, }, },
 840         { .name     = "mremap",     .hexret = true,
 841           .arg = { [3] = { .scnprintf = SCA_MREMAP_FLAGS, /* flags */ }, }, },
 842         { .name     = "name_to_handle_at",
 843           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
 844         { .name     = "newfstatat",
 845           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
 846         { .name     = "open",
 847           .arg = { [1] = { .scnprintf = SCA_OPEN_FLAGS, /* flags */ }, }, },
 848         { .name     = "open_by_handle_at",
 849           .arg = { [0] = { .scnprintf = SCA_FDAT,       /* dfd */ },
 850                    [2] = { .scnprintf = SCA_OPEN_FLAGS, /* flags */ }, }, },
 851         { .name     = "openat",
 852           .arg = { [0] = { .scnprintf = SCA_FDAT,       /* dfd */ },
 853                    [2] = { .scnprintf = SCA_OPEN_FLAGS, /* flags */ }, }, },
 854         { .name     = "perf_event_open",
 855           .arg = { [2] = { .scnprintf = SCA_INT,        /* cpu */ },
 856                    [3] = { .scnprintf = SCA_FD,         /* group_fd */ },
 857                    [4] = { .scnprintf = SCA_PERF_FLAGS, /* flags */ }, }, },
 858         { .name     = "pipe2",
 859           .arg = { [1] = { .scnprintf = SCA_PIPE_FLAGS, /* flags */ }, }, },
 860         { .name     = "pkey_alloc",
 861           .arg = { [1] = { .scnprintf = SCA_PKEY_ALLOC_ACCESS_RIGHTS,   /* access_rights */ }, }, },
 862         { .name     = "pkey_free",
 863           .arg = { [0] = { .scnprintf = SCA_INT,        /* key */ }, }, },
 864         { .name     = "pkey_mprotect",
 865           .arg = { [0] = { .scnprintf = SCA_HEX,        /* start */ },
 866                    [2] = { .scnprintf = SCA_MMAP_PROT,  /* prot */ },
 867                    [3] = { .scnprintf = SCA_INT,        /* pkey */ }, }, },
 868         { .name     = "poll", .timeout = true, },
 869         { .name     = "ppoll", .timeout = true, },
 870         { .name     = "prctl",
 871           .arg = { [0] = { .scnprintf = SCA_PRCTL_OPTION, /* option */ },
 872                    [1] = { .scnprintf = SCA_PRCTL_ARG2, /* arg2 */ },
 873                    [2] = { .scnprintf = SCA_PRCTL_ARG3, /* arg3 */ }, }, },
 874         { .name     = "pread", .alias = "pread64", },
 875         { .name     = "preadv", .alias = "pread", },
 876         { .name     = "prlimit64",
 877           .arg = { [1] = STRARRAY(resource, rlimit_resources), }, },
 878         { .name     = "pwrite", .alias = "pwrite64", },
 879         { .name     = "readlinkat",
 880           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
 881         { .name     = "recvfrom",
 882           .arg = { [3] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
 883         { .name     = "recvmmsg",
 884           .arg = { [3] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
 885         { .name     = "recvmsg",
 886           .arg = { [2] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
 887         { .name     = "renameat",
 888           .arg = { [0] = { .scnprintf = SCA_FDAT, /* olddirfd */ },
 889                    [2] = { .scnprintf = SCA_FDAT, /* newdirfd */ }, }, },
 890         { .name     = "renameat2",
 891           .arg = { [0] = { .scnprintf = SCA_FDAT, /* olddirfd */ },
 892                    [2] = { .scnprintf = SCA_FDAT, /* newdirfd */ },
 893                    [4] = { .scnprintf = SCA_RENAMEAT2_FLAGS, /* flags */ }, }, },
 894         { .name     = "rt_sigaction",
 895           .arg = { [0] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
 896         { .name     = "rt_sigprocmask",
 897           .arg = { [0] = STRARRAY(how, sighow), }, },
 898         { .name     = "rt_sigqueueinfo",
 899           .arg = { [1] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
 900         { .name     = "rt_tgsigqueueinfo",
 901           .arg = { [2] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
 902         { .name     = "sched_setscheduler",
 903           .arg = { [1] = { .scnprintf = SCA_SCHED_POLICY, /* policy */ }, }, },
 904         { .name     = "seccomp",
 905           .arg = { [0] = { .scnprintf = SCA_SECCOMP_OP,    /* op */ },
 906                    [1] = { .scnprintf = SCA_SECCOMP_FLAGS, /* flags */ }, }, },
 907         { .name     = "select", .timeout = true, },
 908         { .name     = "sendfile", .alias = "sendfile64", },
 909         { .name     = "sendmmsg",
 910           .arg = { [3] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
 911         { .name     = "sendmsg",
 912           .arg = { [2] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
 913         { .name     = "sendto",
 914           .arg = { [3] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ },
 915                    [4] = { .scnprintf = SCA_SOCKADDR, /* addr */ }, }, },
 916         { .name     = "set_tid_address", .errpid = true, },
 917         { .name     = "setitimer",
 918           .arg = { [0] = STRARRAY(which, itimers), }, },
 919         { .name     = "setrlimit",
 920           .arg = { [0] = STRARRAY(resource, rlimit_resources), }, },
 921         { .name     = "socket",
 922           .arg = { [0] = STRARRAY(family, socket_families),
 923                    [1] = { .scnprintf = SCA_SK_TYPE, /* type */ },
 924                    [2] = { .scnprintf = SCA_SK_PROTO, /* protocol */ }, }, },
 925         { .name     = "socketpair",
 926           .arg = { [0] = STRARRAY(family, socket_families),
 927                    [1] = { .scnprintf = SCA_SK_TYPE, /* type */ },
 928                    [2] = { .scnprintf = SCA_SK_PROTO, /* protocol */ }, }, },
 929         { .name     = "stat", .alias = "newstat", },
 930         { .name     = "statx",
 931           .arg = { [0] = { .scnprintf = SCA_FDAT,        /* fdat */ },
 932                    [2] = { .scnprintf = SCA_STATX_FLAGS, /* flags */ } ,
 933                    [3] = { .scnprintf = SCA_STATX_MASK,  /* mask */ }, }, },
 934         { .name     = "swapoff",
 935           .arg = { [0] = { .scnprintf = SCA_FILENAME, /* specialfile */ }, }, },
 936         { .name     = "swapon",
 937           .arg = { [0] = { .scnprintf = SCA_FILENAME, /* specialfile */ }, }, },
 938         { .name     = "symlinkat",
 939           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
 940         { .name     = "sync_file_range",
 941           .arg = { [3] = { .scnprintf = SCA_SYNC_FILE_RANGE_FLAGS, /* flags */ }, }, },
 942         { .name     = "tgkill",
 943           .arg = { [2] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
 944         { .name     = "tkill",
 945           .arg = { [1] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
 946         { .name     = "umount2", .alias = "umount",
 947           .arg = { [0] = { .scnprintf = SCA_FILENAME, /* name */ }, }, },
 948         { .name     = "uname", .alias = "newuname", },
 949         { .name     = "unlinkat",
 950           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
 951         { .name     = "utimensat",
 952           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dirfd */ }, }, },
 953         { .name     = "wait4",      .errpid = true,
 954           .arg = { [2] = { .scnprintf = SCA_WAITID_OPTIONS, /* options */ }, }, },
 955         { .name     = "waitid",     .errpid = true,
 956           .arg = { [3] = { .scnprintf = SCA_WAITID_OPTIONS, /* options */ }, }, },
 957 };
 958
 959 static int syscall_fmt__cmp(const void *name, const void *fmtp)
 960 {
 961         const struct syscall_fmt *fmt = fmtp;
 962         return strcmp(name, fmt->name);
 963 }
 964
 965 static struct syscall_fmt *syscall_fmt__find(const char *name)
 966 {
 967         const int nmemb = ARRAY_SIZE(syscall_fmts);
 968         return bsearch(name, syscall_fmts, nmemb, sizeof(struct syscall_fmt), syscall_fmt__cmp);
 969 }
 970
 971 static struct syscall_fmt *syscall_fmt__find_by_alias(const char *alias)
 972 {
 973         int i, nmemb = ARRAY_SIZE(syscall_fmts);
 974
 975         for (i = 0; i < nmemb; ++i) {
 976                 if (syscall_fmts[i].alias && strcmp(syscall_fmts[i].alias, alias) == 0)
 977                         return &syscall_fmts[i];
 978         }
 979
 980         return NULL;
 981 }
 982
 983 /*
 984  * is_exit: is this "exit" or "exit_group"?
 985  * is_open: is this "open" or "openat"? To associate the fd returned in sys_exit with the pathname in sys_enter.
 986  * args_size: sum of the sizes of the syscall arguments, anything after that is augmented stuff: pathname for openat, etc.
 987  * nonexistent: Just a hole in the syscall table, syscall id not allocated
 988  */
 989 struct syscall {
 990         struct tep_event    *tp_format;
 991         int                 nr_args;
 992         int                 args_size;
 993         struct {
 994                 struct bpf_program *sys_enter,
 995                                    *sys_exit;
 996         }                   bpf_prog;
 997         bool                is_exit;
 998         bool                is_open;
 999         bool                nonexistent;
1000         struct tep_format_field *args;
1001         const char          *name;
1002         struct syscall_fmt  *fmt;
1003         struct syscall_arg_fmt *arg_fmt;
1004 };
1005
1006 /*
1007  * Must match what is in the BPF program:
1008  *
1009  * tools/perf/examples/bpf/augmented_raw_syscalls.c
1010  */
1011 struct bpf_map_syscall_entry {
1012         bool    enabled;
1013         u16     string_args_len[6];
1014 };
1015
1016 /*
1017  * We need to have this 'calculated' boolean because in some cases we really
1018  * don't know what is the duration of a syscall, for instance, when we start
1019  * a session and some threads are waiting for a syscall to finish, say 'poll',
1020  * in which case all we can do is to print "( ? ) for duration and for the
1021  * start timestamp.
1022  */
1023 static size_t fprintf_duration(unsigned long t, bool calculated, FILE *fp)
1024 {
1025         double duration = (double)t / NSEC_PER_MSEC;
1026         size_t printed = fprintf(fp, "(");
1027
1028         if (!calculated)
1029                 printed += fprintf(fp, "         ");
1030         else if (duration >= 1.0)
1031                 printed += color_fprintf(fp, PERF_COLOR_RED, "%6.3f ms", duration);
1032         else if (duration >= 0.01)
1033                 printed += color_fprintf(fp, PERF_COLOR_YELLOW, "%6.3f ms", duration);
1034         else
1035                 printed += color_fprintf(fp, PERF_COLOR_NORMAL, "%6.3f ms", duration);
1036         return printed + fprintf(fp, "): ");
1037 }
1038
1039 /**
1040  * filename.ptr: The filename char pointer that will be vfs_getname'd
1041  * filename.entry_str_pos: Where to insert the string translated from
1042  *                         filename.ptr by the vfs_getname tracepoint/kprobe.
1043  * ret_scnprintf: syscall args may set this to a different syscall return
1044  *                formatter, for instance, fcntl may return fds, file flags, etc.
1045  */
1046 struct thread_trace {
1047         u64               entry_time;
1048         bool              entry_pending;
1049         unsigned long     nr_events;
1050         unsigned long     pfmaj, pfmin;
1051         char              *entry_str;
1052         double            runtime_ms;
1053         size_t            (*ret_scnprintf)(char *bf, size_t size, struct syscall_arg *arg);
1054         struct {
1055                 unsigned long ptr;
1056                 short int     entry_str_pos;
1057                 bool          pending_open;
1058                 unsigned int  namelen;
1059                 char          *name;
1060         } filename;
1061         struct {
1062                 int           max;
1063                 struct file   *table;
1064         } files;
1065
1066         struct intlist *syscall_stats;
1067 };
1068
1069 static struct thread_trace *thread_trace__new(void)
1070 {
1071         struct thread_trace *ttrace =  zalloc(sizeof(struct thread_trace));
1072
1073         if (ttrace) {
1074                 ttrace->files.max = -1;
1075                 ttrace->syscall_stats = intlist__new(NULL);
1076         }
1077
1078         return ttrace;
1079 }
1080
1081 static struct thread_trace *thread__trace(struct thread *thread, FILE *fp)
1082 {
1083         struct thread_trace *ttrace;
1084
1085         if (thread == NULL)
1086                 goto fail;
1087
1088         if (thread__priv(thread) == NULL)
1089                 thread__set_priv(thread, thread_trace__new());
1090
1091         if (thread__priv(thread) == NULL)
1092                 goto fail;
1093
1094         ttrace = thread__priv(thread);
1095         ++ttrace->nr_events;
1096
1097         return ttrace;
1098 fail:
1099         color_fprintf(fp, PERF_COLOR_RED,
1100                       "WARNING: not enough memory, dropping samples!\n");
1101         return NULL;
1102 }
1103
1104
1105 void syscall_arg__set_ret_scnprintf(struct syscall_arg *arg,
1106                                     size_t (*ret_scnprintf)(char *bf, size_t size, struct syscall_arg *arg))
1107 {
1108         struct thread_trace *ttrace = thread__priv(arg->thread);
1109
1110         ttrace->ret_scnprintf = ret_scnprintf;
1111 }
1112
1113 #define TRACE_PFMAJ             (1 << 0)
1114 #define TRACE_PFMIN             (1 << 1)
1115
1116 static const size_t trace__entry_str_size = 2048;
1117
1118 static struct file *thread_trace__files_entry(struct thread_trace *ttrace, int fd)
1119 {
1120         if (fd < 0)
1121                 return NULL;
1122
1123         if (fd > ttrace->files.max) {
1124                 struct file *nfiles = realloc(ttrace->files.table, (fd + 1) * sizeof(struct file));
1125
1126                 if (nfiles == NULL)
1127                         return NULL;
1128
1129                 if (ttrace->files.max != -1) {
1130                         memset(nfiles + ttrace->files.max + 1, 0,
1131                                (fd - ttrace->files.max) * sizeof(struct file));
1132                 } else {
1133                         memset(nfiles, 0, (fd + 1) * sizeof(struct file));
1134                 }
1135
1136                 ttrace->files.table = nfiles;
1137                 ttrace->files.max   = fd;
1138         }
1139
1140         return ttrace->files.table + fd;
1141 }
1142
1143 struct file *thread__files_entry(struct thread *thread, int fd)
1144 {
1145         return thread_trace__files_entry(thread__priv(thread), fd);
1146 }
1147
1148 static int trace__set_fd_pathname(struct thread *thread, int fd, const char *pathname)
1149 {
1150         struct thread_trace *ttrace = thread__priv(thread);
1151         struct file *file = thread_trace__files_entry(ttrace, fd);
1152
1153         if (file != NULL) {
1154                 struct stat st;
1155                 if (stat(pathname, &st) == 0)
1156                         file->dev_maj = major(st.st_rdev);
1157                 file->pathname = strdup(pathname);
1158                 if (file->pathname)
1159                         return 0;
1160         }
1161
1162         return -1;
1163 }
1164
1165 static int thread__read_fd_path(struct thread *thread, int fd)
1166 {
1167         char linkname[PATH_MAX], pathname[PATH_MAX];
1168         struct stat st;
1169         int ret;
1170
1171         if (thread->pid_ == thread->tid) {
1172                 scnprintf(linkname, sizeof(linkname),
1173                           "/proc/%d/fd/%d", thread->pid_, fd);
1174         } else {
1175                 scnprintf(linkname, sizeof(linkname),
1176                           "/proc/%d/task/%d/fd/%d", thread->pid_, thread->tid, fd);
1177         }
1178
1179         if (lstat(linkname, &st) < 0 || st.st_size + 1 > (off_t)sizeof(pathname))
1180                 return -1;
1181
1182         ret = readlink(linkname, pathname, sizeof(pathname));
1183
1184         if (ret < 0 || ret > st.st_size)
1185                 return -1;
1186
1187         pathname[ret] = '\0';
1188         return trace__set_fd_pathname(thread, fd, pathname);
1189 }
1190
1191 static const char *thread__fd_path(struct thread *thread, int fd,
1192                                    struct trace *trace)
1193 {
1194         struct thread_trace *ttrace = thread__priv(thread);
1195
1196         if (ttrace == NULL || trace->fd_path_disabled)
1197                 return NULL;
1198
1199         if (fd < 0)
1200                 return NULL;
1201
1202         if ((fd > ttrace->files.max || ttrace->files.table[fd].pathname == NULL)) {
1203                 if (!trace->live)
1204                         return NULL;
1205                 ++trace->stats.proc_getname;
1206                 if (thread__read_fd_path(thread, fd))
1207                         return NULL;
1208         }
1209
1210         return ttrace->files.table[fd].pathname;
1211 }
1212
1213 size_t syscall_arg__scnprintf_fd(char *bf, size_t size, struct syscall_arg *arg)
1214 {
1215         int fd = arg->val;
1216         size_t printed = scnprintf(bf, size, "%d", fd);
1217         const char *path = thread__fd_path(arg->thread, fd, arg->trace);
1218
1219         if (path)
1220                 printed += scnprintf(bf + printed, size - printed, "<%s>", path);
1221
1222         return printed;
1223 }
1224
1225 size_t pid__scnprintf_fd(struct trace *trace, pid_t pid, int fd, char *bf, size_t size)
1226 {
1227         size_t printed = scnprintf(bf, size, "%d", fd);
1228         struct thread *thread = machine__find_thread(trace->host, pid, pid);
1229
1230         if (thread) {
1231                 const char *path = thread__fd_path(thread, fd, trace);
1232
1233                 if (path)
1234                         printed += scnprintf(bf + printed, size - printed, "<%s>", path);
1235
1236                 thread__put(thread);
1237         }
1238
1239         return printed;
1240 }
1241
1242 static size_t syscall_arg__scnprintf_close_fd(char *bf, size_t size,
1243                                               struct syscall_arg *arg)
1244 {
1245         int fd = arg->val;
1246         size_t printed = syscall_arg__scnprintf_fd(bf, size, arg);
1247         struct thread_trace *ttrace = thread__priv(arg->thread);
1248
1249         if (ttrace && fd >= 0 && fd <= ttrace->files.max)
1250                 zfree(&ttrace->files.table[fd].pathname);
1251
1252         return printed;
1253 }
1254
1255 static void thread__set_filename_pos(struct thread *thread, const char *bf,
1256                                      unsigned long ptr)
1257 {
1258         struct thread_trace *ttrace = thread__priv(thread);
1259
1260         ttrace->filename.ptr = ptr;
1261         ttrace->filename.entry_str_pos = bf - ttrace->entry_str;
1262 }
1263
1264 static size_t syscall_arg__scnprintf_augmented_string(struct syscall_arg *arg, char *bf, size_t size)
1265 {
1266         struct augmented_arg *augmented_arg = arg->augmented.args;
1267         size_t printed = scnprintf(bf, size, "\"%.*s\"", augmented_arg->size, augmented_arg->value);
1268         /*
1269          * So that the next arg with a payload can consume its augmented arg, i.e. for rename* syscalls
1270          * we would have two strings, each prefixed by its size.
1271          */
1272         int consumed = sizeof(*augmented_arg) + augmented_arg->size;
1273
1274         arg->augmented.args = ((void *)arg->augmented.args) + consumed;
1275         arg->augmented.size -= consumed;
1276
1277         return printed;
1278 }
1279
1280 static size_t syscall_arg__scnprintf_filename(char *bf, size_t size,
1281                                               struct syscall_arg *arg)
1282 {
1283         unsigned long ptr = arg->val;
1284
1285         if (arg->augmented.args)
1286                 return syscall_arg__scnprintf_augmented_string(arg, bf, size);
1287
1288         if (!arg->trace->vfs_getname)
1289                 return scnprintf(bf, size, "%#x", ptr);
1290
1291         thread__set_filename_pos(arg->thread, bf, ptr);
1292         return 0;
1293 }
1294
1295 static bool trace__filter_duration(struct trace *trace, double t)
1296 {
1297         return t < (trace->duration_filter * NSEC_PER_MSEC);
1298 }
1299
1300 static size_t __trace__fprintf_tstamp(struct trace *trace, u64 tstamp, FILE *fp)
1301 {
1302         double ts = (double)(tstamp - trace->base_time) / NSEC_PER_MSEC;
1303
1304         return fprintf(fp, "%10.3f ", ts);
1305 }
1306
1307 /*
1308  * We're handling tstamp=0 as an undefined tstamp, i.e. like when we are
1309  * using ttrace->entry_time for a thread that receives a sys_exit without
1310  * first having received a sys_enter ("poll" issued before tracing session
1311  * starts, lost sys_enter exit due to ring buffer overflow).
1312  */
1313 static size_t trace__fprintf_tstamp(struct trace *trace, u64 tstamp, FILE *fp)
1314 {
1315         if (tstamp > 0)
1316                 return __trace__fprintf_tstamp(trace, tstamp, fp);
1317
1318         return fprintf(fp, "         ? ");
1319 }
1320
1321 static bool done = false;
1322 static bool interrupted = false;
1323
1324 static void sig_handler(int sig)
1325 {
1326         done = true;
1327         interrupted = sig == SIGINT;
1328 }
1329
1330 static size_t trace__fprintf_comm_tid(struct trace *trace, struct thread *thread, FILE *fp)
1331 {
1332         size_t printed = 0;
1333
1334         if (trace->multiple_threads) {
1335                 if (trace->show_comm)
1336                         printed += fprintf(fp, "%.14s/", thread__comm_str(thread));
1337                 printed += fprintf(fp, "%d ", thread->tid);
1338         }
1339
1340         return printed;
1341 }
1342
1343 static size_t trace__fprintf_entry_head(struct trace *trace, struct thread *thread,
1344                                         u64 duration, bool duration_calculated, u64 tstamp, FILE *fp)
1345 {
1346         size_t printed = 0;
1347
1348         if (trace->show_tstamp)
1349                 printed = trace__fprintf_tstamp(trace, tstamp, fp);
1350         if (trace->show_duration)
1351                 printed += fprintf_duration(duration, duration_calculated, fp);
1352         return printed + trace__fprintf_comm_tid(trace, thread, fp);
1353 }
1354
1355 static int trace__process_event(struct trace *trace, struct machine *machine,
1356                                 union perf_event *event, struct perf_sample *sample)
1357 {
1358         int ret = 0;
1359
1360         switch (event->header.type) {
1361         case PERF_RECORD_LOST:
1362                 color_fprintf(trace->output, PERF_COLOR_RED,
1363                               "LOST %" PRIu64 " events!\n", event->lost.lost);
1364                 ret = machine__process_lost_event(machine, event, sample);
1365                 break;
1366         default:
1367                 ret = machine__process_event(machine, event, sample);
1368                 break;
1369         }
1370
1371         return ret;
1372 }
1373
1374 static int trace__tool_process(struct perf_tool *tool,
1375                                union perf_event *event,
1376                                struct perf_sample *sample,
1377                                struct machine *machine)
1378 {
1379         struct trace *trace = container_of(tool, struct trace, tool);
1380         return trace__process_event(trace, machine, event, sample);
1381 }
1382
1383 static char *trace__machine__resolve_kernel_addr(void *vmachine, unsigned long long *addrp, char **modp)
1384 {
1385         struct machine *machine = vmachine;
1386
1387         if (machine->kptr_restrict_warned)
1388                 return NULL;
1389
1390         if (symbol_conf.kptr_restrict) {
1391                 pr_warning("Kernel address maps (/proc/{kallsyms,modules}) are restricted.\n\n"
1392                            "Check /proc/sys/kernel/kptr_restrict and /proc/sys/kernel/perf_event_paranoid.\n\n"
1393                            "Kernel samples will not be resolved.\n");
1394                 machine->kptr_restrict_warned = true;
1395                 return NULL;
1396         }
1397
1398         return machine__resolve_kernel_addr(vmachine, addrp, modp);
1399 }
1400
1401 static int trace__symbols_init(struct trace *trace, struct evlist *evlist)
1402 {
1403         int err = symbol__init(NULL);
1404
1405         if (err)
1406                 return err;
1407
1408         trace->host = machine__new_host();
1409         if (trace->host == NULL)
1410                 return -ENOMEM;
1411
1412         err = trace_event__register_resolver(trace->host, trace__machine__resolve_kernel_addr);
1413         if (err < 0)
1414                 goto out;
1415
1416         err = __machine__synthesize_threads(trace->host, &trace->tool, &trace->opts.target,
1417                                             evlist->core.threads, trace__tool_process, false,
1418                                             1);
1419 out:
1420         if (err)
1421                 symbol__exit();
1422
1423         return err;
1424 }
1425
1426 static void trace__symbols__exit(struct trace *trace)
1427 {
1428         machine__exit(trace->host);
1429         trace->host = NULL;
1430
1431         symbol__exit();
1432 }
1433
1434 static int syscall__alloc_arg_fmts(struct syscall *sc, int nr_args)
1435 {
1436         int idx;
1437
1438         if (nr_args == 6 && sc->fmt && sc->fmt->nr_args != 0)
1439                 nr_args = sc->fmt->nr_args;
1440
1441         sc->arg_fmt = calloc(nr_args, sizeof(*sc->arg_fmt));
1442         if (sc->arg_fmt == NULL)
1443                 return -1;
1444
1445         for (idx = 0; idx < nr_args; ++idx) {
1446                 if (sc->fmt)
1447                         sc->arg_fmt[idx] = sc->fmt->arg[idx];
1448         }
1449
1450         sc->nr_args = nr_args;
1451         return 0;
1452 }
1453
1454 static int syscall__set_arg_fmts(struct syscall *sc)
1455 {
1456         struct tep_format_field *field, *last_field = NULL;
1457         int idx = 0, len;
1458
1459         for (field = sc->args; field; field = field->next, ++idx) {
1460                 last_field = field;
1461
1462                 if (sc->fmt && sc->fmt->arg[idx].scnprintf)
1463                         continue;
1464
1465                 len = strlen(field->name);
1466
1467                 if (strcmp(field->type, "const char *") == 0 &&
1468                     ((len >= 4 && strcmp(field->name + len - 4, "name") == 0) ||
1469                      strstr(field->name, "path") != NULL))
1470                         sc->arg_fmt[idx].scnprintf = SCA_FILENAME;
1471                 else if ((field->flags & TEP_FIELD_IS_POINTER) || strstr(field->name, "addr"))
1472                         sc->arg_fmt[idx].scnprintf = SCA_PTR;
1473                 else if (strcmp(field->type, "pid_t") == 0)
1474                         sc->arg_fmt[idx].scnprintf = SCA_PID;
1475                 else if (strcmp(field->type, "umode_t") == 0)
1476                         sc->arg_fmt[idx].scnprintf = SCA_MODE_T;
1477                 else if ((strcmp(field->type, "int") == 0 ||
1478                           strcmp(field->type, "unsigned int") == 0 ||
1479                           strcmp(field->type, "long") == 0) &&
1480                          len >= 2 && strcmp(field->name + len - 2, "fd") == 0) {
1481                         /*
1482                          * /sys/kernel/tracing/events/syscalls/sys_enter*
1483                          * egrep 'field:.*fd;' .../format|sed -r 's/.*field:([a-z ]+) [a-z_]*fd.+/\1/g'|sort|uniq -c
1484                          * 65 int
1485                          * 23 unsigned int
1486                          * 7 unsigned long
1487                          */
1488                         sc->arg_fmt[idx].scnprintf = SCA_FD;
1489                 }
1490         }
1491
1492         if (last_field)
1493                 sc->args_size = last_field->offset + last_field->size;
1494
1495         return 0;
1496 }
1497
1498 static int trace__read_syscall_info(struct trace *trace, int id)
1499 {
1500         char tp_name[128];
1501         struct syscall *sc;
1502         const char *name = syscalltbl__name(trace->sctbl, id);
1503
1504         if (trace->syscalls.table == NULL) {
1505                 trace->syscalls.table = calloc(trace->sctbl->syscalls.max_id + 1, sizeof(*sc));
1506                 if (trace->syscalls.table == NULL)
1507                         return -ENOMEM;
1508         }
1509
1510         sc = trace->syscalls.table + id;
1511         if (sc->nonexistent)
1512                 return 0;
1513
1514         if (name == NULL) {
1515                 sc->nonexistent = true;
1516                 return 0;
1517         }
1518
1519         sc->name = name;
1520         sc->fmt  = syscall_fmt__find(sc->name);
1521
1522         snprintf(tp_name, sizeof(tp_name), "sys_enter_%s", sc->name);
1523         sc->tp_format = trace_event__tp_format("syscalls", tp_name);
1524
1525         if (IS_ERR(sc->tp_format) && sc->fmt && sc->fmt->alias) {
1526                 snprintf(tp_name, sizeof(tp_name), "sys_enter_%s", sc->fmt->alias);
1527                 sc->tp_format = trace_event__tp_format("syscalls", tp_name);
1528         }
1529
1530         if (syscall__alloc_arg_fmts(sc, IS_ERR(sc->tp_format) ? 6 : sc->tp_format->format.nr_fields))
1531                 return -ENOMEM;
1532
1533         if (IS_ERR(sc->tp_format))
1534                 return PTR_ERR(sc->tp_format);
1535
1536         sc->args = sc->tp_format->format.fields;
1537         /*
1538          * We need to check and discard the first variable '__syscall_nr'
1539          * or 'nr' that mean the syscall number. It is needless here.
1540          * So drop '__syscall_nr' or 'nr' field but does not exist on older kernels.
1541          */
1542         if (sc->args && (!strcmp(sc->args->name, "__syscall_nr") || !strcmp(sc->args->name, "nr"))) {
1543                 sc->args = sc->args->next;
1544                 --sc->nr_args;
1545         }
1546
1547         sc->is_exit = !strcmp(name, "exit_group") || !strcmp(name, "exit");
1548         sc->is_open = !strcmp(name, "open") || !strcmp(name, "openat");
1549
1550         return syscall__set_arg_fmts(sc);
1551 }
1552
1553 static int intcmp(const void *a, const void *b)
1554 {
1555         const int *one = a, *another = b;
1556
1557         return *one - *another;
1558 }
1559
1560 static int trace__validate_ev_qualifier(struct trace *trace)
1561 {
1562         int err = 0;
1563         bool printed_invalid_prefix = false;
1564         struct str_node *pos;
1565         size_t nr_used = 0, nr_allocated = strlist__nr_entries(trace->ev_qualifier);
1566
1567         trace->ev_qualifier_ids.entries = malloc(nr_allocated *
1568                                                  sizeof(trace->ev_qualifier_ids.entries[0]));
1569
1570         if (trace->ev_qualifier_ids.entries == NULL) {
1571                 fputs("Error:\tNot enough memory for allocating events qualifier ids\n",
1572                        trace->output);
1573                 err = -EINVAL;
1574                 goto out;
1575         }
1576
1577         strlist__for_each_entry(pos, trace->ev_qualifier) {
1578                 const char *sc = pos->s;
1579                 int id = syscalltbl__id(trace->sctbl, sc), match_next = -1;
1580
1581                 if (id < 0) {
1582                         id = syscalltbl__strglobmatch_first(trace->sctbl, sc, &match_next);
1583                         if (id >= 0)
1584                                 goto matches;
1585
1586                         if (!printed_invalid_prefix) {
1587                                 pr_debug("Skipping unknown syscalls: ");
1588                                 printed_invalid_prefix = true;
1589                         } else {
1590                                 pr_debug(", ");
1591                         }
1592
1593                         pr_debug("%s", sc);
1594                         continue;
1595                 }
1596 matches:
1597                 trace->ev_qualifier_ids.entries[nr_used++] = id;
1598                 if (match_next == -1)
1599                         continue;
1600
1601                 while (1) {
1602                         id = syscalltbl__strglobmatch_next(trace->sctbl, sc, &match_next);
1603                         if (id < 0)
1604                                 break;
1605                         if (nr_allocated == nr_used) {
1606                                 void *entries;
1607
1608                                 nr_allocated += 8;
1609                                 entries = realloc(trace->ev_qualifier_ids.entries,
1610                                                   nr_allocated * sizeof(trace->ev_qualifier_ids.entries[0]));
1611                                 if (entries == NULL) {
1612                                         err = -ENOMEM;
1613                                         fputs("\nError:\t Not enough memory for parsing\n", trace->output);
1614                                         goto out_free;
1615                                 }
1616                                 trace->ev_qualifier_ids.entries = entries;
1617                         }
1618                         trace->ev_qualifier_ids.entries[nr_used++] = id;
1619                 }
1620         }
1621
1622         trace->ev_qualifier_ids.nr = nr_used;
1623         qsort(trace->ev_qualifier_ids.entries, nr_used, sizeof(int), intcmp);
1624 out:
1625         if (printed_invalid_prefix)
1626                 pr_debug("\n");
1627         return err;
1628 out_free:
1629         zfree(&trace->ev_qualifier_ids.entries);
1630         trace->ev_qualifier_ids.nr = 0;
1631         goto out;
1632 }
1633
1634 static __maybe_unused bool trace__syscall_enabled(struct trace *trace, int id)
1635 {
1636         bool in_ev_qualifier;
1637
1638         if (trace->ev_qualifier_ids.nr == 0)
1639                 return true;
1640
1641         in_ev_qualifier = bsearch(&id, trace->ev_qualifier_ids.entries,
1642                                   trace->ev_qualifier_ids.nr, sizeof(int), intcmp) != NULL;
1643
1644         if (in_ev_qualifier)
1645                return !trace->not_ev_qualifier;
1646
1647         return trace->not_ev_qualifier;
1648 }
1649
1650 /*
1651  * args is to be interpreted as a series of longs but we need to handle
1652  * 8-byte unaligned accesses. args points to raw_data within the event
1653  * and raw_data is guaranteed to be 8-byte unaligned because it is
1654  * preceded by raw_size which is a u32. So we need to copy args to a temp
1655  * variable to read it. Most notably this avoids extended load instructions
1656  * on unaligned addresses
1657  */
1658 unsigned long syscall_arg__val(struct syscall_arg *arg, u8 idx)
1659 {
1660         unsigned long val;
1661         unsigned char *p = arg->args + sizeof(unsigned long) * idx;
1662
1663         memcpy(&val, p, sizeof(val));
1664         return val;
1665 }
1666
1667 static size_t syscall__scnprintf_name(struct syscall *sc, char *bf, size_t size,
1668                                       struct syscall_arg *arg)
1669 {
1670         if (sc->arg_fmt && sc->arg_fmt[arg->idx].name)
1671                 return scnprintf(bf, size, "%s: ", sc->arg_fmt[arg->idx].name);
1672
1673         return scnprintf(bf, size, "arg%d: ", arg->idx);
1674 }
1675
1676 /*
1677  * Check if the value is in fact zero, i.e. mask whatever needs masking, such
1678  * as mount 'flags' argument that needs ignoring some magic flag, see comment
1679  * in tools/perf/trace/beauty/mount_flags.c
1680  */
1681 static unsigned long syscall__mask_val(struct syscall *sc, struct syscall_arg *arg, unsigned long val)
1682 {
1683         if (sc->arg_fmt && sc->arg_fmt[arg->idx].mask_val)
1684                 return sc->arg_fmt[arg->idx].mask_val(arg, val);
1685
1686         return val;
1687 }
1688
1689 static size_t syscall__scnprintf_val(struct syscall *sc, char *bf, size_t size,
1690                                      struct syscall_arg *arg, unsigned long val)
1691 {
1692         if (sc->arg_fmt && sc->arg_fmt[arg->idx].scnprintf) {
1693                 arg->val = val;
1694                 if (sc->arg_fmt[arg->idx].parm)
1695                         arg->parm = sc->arg_fmt[arg->idx].parm;
1696                 return sc->arg_fmt[arg->idx].scnprintf(bf, size, arg);
1697         }
1698         return scnprintf(bf, size, "%ld", val);
1699 }
1700
1701 static size_t syscall__scnprintf_args(struct syscall *sc, char *bf, size_t size,
1702                                       unsigned char *args, void *augmented_args, int augmented_args_size,
1703                                       struct trace *trace, struct thread *thread)
1704 {
1705         size_t printed = 0;
1706         unsigned long val;
1707         u8 bit = 1;
1708         struct syscall_arg arg = {
1709                 .args   = args,
1710                 .augmented = {
1711                         .size = augmented_args_size,
1712                         .args = augmented_args,
1713                 },
1714                 .idx    = 0,
1715                 .mask   = 0,
1716                 .trace  = trace,
1717                 .thread = thread,
1718                 .show_string_prefix = trace->show_string_prefix,
1719         };
1720         struct thread_trace *ttrace = thread__priv(thread);
1721
1722         /*
1723          * Things like fcntl will set this in its 'cmd' formatter to pick the
1724          * right formatter for the return value (an fd? file flags?), which is
1725          * not needed for syscalls that always return a given type, say an fd.
1726          */
1727         ttrace->ret_scnprintf = NULL;
1728
1729         if (sc->args != NULL) {
1730                 struct tep_format_field *field;
1731
1732                 for (field = sc->args; field;
1733                      field = field->next, ++arg.idx, bit <<= 1) {
1734                         if (arg.mask & bit)
1735                                 continue;
1736
1737                         val = syscall_arg__val(&arg, arg.idx);
1738                         /*
1739                          * Some syscall args need some mask, most don't and
1740                          * return val untouched.
1741                          */
1742                         val = syscall__mask_val(sc, &arg, val);
1743
1744                         /*
1745                          * Suppress this argument if its value is zero and
1746                          * and we don't have a string associated in an
1747                          * strarray for it.
1748                          */
1749                         if (val == 0 &&
1750                             !trace->show_zeros &&
1751                             !(sc->arg_fmt &&
1752                               (sc->arg_fmt[arg.idx].show_zero ||
1753                                sc->arg_fmt[arg.idx].scnprintf == SCA_STRARRAY ||
1754                                sc->arg_fmt[arg.idx].scnprintf == SCA_STRARRAYS) &&
1755                               sc->arg_fmt[arg.idx].parm))
1756                                 continue;
1757
1758                         printed += scnprintf(bf + printed, size - printed, "%s", printed ? ", " : "");
1759
1760                         if (trace->show_arg_names)
1761                                 printed += scnprintf(bf + printed, size - printed, "%s: ", field->name);
1762
1763                         printed += syscall__scnprintf_val(sc, bf + printed, size - printed, &arg, val);
1764                 }
1765         } else if (IS_ERR(sc->tp_format)) {
1766                 /*
1767                  * If we managed to read the tracepoint /format file, then we
1768                  * may end up not having any args, like with gettid(), so only
1769                  * print the raw args when we didn't manage to read it.
1770                  */
1771                 while (arg.idx < sc->nr_args) {
1772                         if (arg.mask & bit)
1773                                 goto next_arg;
1774                         val = syscall_arg__val(&arg, arg.idx);
1775                         if (printed)
1776                                 printed += scnprintf(bf + printed, size - printed, ", ");
1777                         printed += syscall__scnprintf_name(sc, bf + printed, size - printed, &arg);
1778                         printed += syscall__scnprintf_val(sc, bf + printed, size - printed, &arg, val);
1779 next_arg:
1780                         ++arg.idx;
1781                         bit <<= 1;
1782                 }
1783         }
1784
1785         return printed;
1786 }
1787
1788 typedef int (*tracepoint_handler)(struct trace *trace, struct evsel *evsel,
1789                                   union perf_event *event,
1790                                   struct perf_sample *sample);
1791
1792 static struct syscall *trace__syscall_info(struct trace *trace,
1793                                            struct evsel *evsel, int id)
1794 {
1795         int err = 0;
1796
1797         if (id < 0) {
1798
1799                 /*
1800                  * XXX: Noticed on x86_64, reproduced as far back as 3.0.36, haven't tried
1801                  * before that, leaving at a higher verbosity level till that is
1802                  * explained. Reproduced with plain ftrace with:
1803                  *
1804                  * echo 1 > /t/events/raw_syscalls/sys_exit/enable
1805                  * grep "NR -1 " /t/trace_pipe
1806                  *
1807                  * After generating some load on the machine.
1808                  */
1809                 if (verbose > 1) {
1810                         static u64 n;
1811                         fprintf(trace->output, "Invalid syscall %d id, skipping (%s, %" PRIu64 ") ...\n",
1812                                 id, perf_evsel__name(evsel), ++n);
1813                 }
1814                 return NULL;
1815         }
1816
1817         err = -EINVAL;
1818
1819         if (id > trace->sctbl->syscalls.max_id)
1820                 goto out_cant_read;
1821
1822         if ((trace->syscalls.table == NULL || trace->syscalls.table[id].name == NULL) &&
1823             (err = trace__read_syscall_info(trace, id)) != 0)
1824                 goto out_cant_read;
1825
1826         if (trace->syscalls.table[id].name == NULL) {
1827                 if (trace->syscalls.table[id].nonexistent)
1828                         return NULL;
1829                 goto out_cant_read;
1830         }
1831
1832         return &trace->syscalls.table[id];
1833
1834 out_cant_read:
1835         if (verbose > 0) {
1836                 char sbuf[STRERR_BUFSIZE];
1837                 fprintf(trace->output, "Problems reading syscall %d: %d (%s)", id, -err, str_error_r(-err, sbuf, sizeof(sbuf)));
1838                 if (id <= trace->sctbl->syscalls.max_id && trace->syscalls.table[id].name != NULL)
1839                         fprintf(trace->output, "(%s)", trace->syscalls.table[id].name);
1840                 fputs(" information\n", trace->output);
1841         }
1842         return NULL;
1843 }
1844
1845 static void thread__update_stats(struct thread_trace *ttrace,
1846                                  int id, struct perf_sample *sample)
1847 {
1848         struct int_node *inode;
1849         struct stats *stats;
1850         u64 duration = 0;
1851
1852         inode = intlist__findnew(ttrace->syscall_stats, id);
1853         if (inode == NULL)
1854                 return;
1855
1856         stats = inode->priv;
1857         if (stats == NULL) {
1858                 stats = malloc(sizeof(struct stats));
1859                 if (stats == NULL)
1860                         return;
1861                 init_stats(stats);
1862                 inode->priv = stats;
1863         }
1864
1865         if (ttrace->entry_time && sample->time > ttrace->entry_time)
1866                 duration = sample->time - ttrace->entry_time;
1867
1868         update_stats(stats, duration);
1869 }
1870
1871 static int trace__printf_interrupted_entry(struct trace *trace)
1872 {
1873         struct thread_trace *ttrace;
1874         size_t printed;
1875         int len;
1876
1877         if (trace->failure_only || trace->current == NULL)
1878                 return 0;
1879
1880         ttrace = thread__priv(trace->current);
1881
1882         if (!ttrace->entry_pending)
1883                 return 0;
1884
1885         printed  = trace__fprintf_entry_head(trace, trace->current, 0, false, ttrace->entry_time, trace->output);
1886         printed += len = fprintf(trace->output, "%s)", ttrace->entry_str);
1887
1888         if (len < trace->args_alignment - 4)
1889                 printed += fprintf(trace->output, "%-*s", trace->args_alignment - 4 - len, " ");
1890
1891         printed += fprintf(trace->output, " ...\n");
1892
1893         ttrace->entry_pending = false;
1894         ++trace->nr_events_printed;
1895
1896         return printed;
1897 }
1898
1899 static int trace__fprintf_sample(struct trace *trace, struct evsel *evsel,
1900                                  struct perf_sample *sample, struct thread *thread)
1901 {
1902         int printed = 0;
1903
1904         if (trace->print_sample) {
1905                 double ts = (double)sample->time / NSEC_PER_MSEC;
1906
1907                 printed += fprintf(trace->output, "%22s %10.3f %s %d/%d [%d]\n",
1908                                    perf_evsel__name(evsel), ts,
1909                                    thread__comm_str(thread),
1910                                    sample->pid, sample->tid, sample->cpu);
1911         }
1912
1913         return printed;
1914 }
1915
1916 static void *syscall__augmented_args(struct syscall *sc, struct perf_sample *sample, int *augmented_args_size, int raw_augmented_args_size)
1917 {
1918         void *augmented_args = NULL;
1919         /*
1920          * For now with BPF raw_augmented we hook into raw_syscalls:sys_enter
1921          * and there we get all 6 syscall args plus the tracepoint common fields
1922          * that gets calculated at the start and the syscall_nr (another long).
1923          * So we check if that is the case and if so don't look after the
1924          * sc->args_size but always after the full raw_syscalls:sys_enter payload,
1925          * which is fixed.
1926          *
1927          * We'll revisit this later to pass s->args_size to the BPF augmenter
1928          * (now tools/perf/examples/bpf/augmented_raw_syscalls.c, so that it
1929          * copies only what we need for each syscall, like what happens when we
1930          * use syscalls:sys_enter_NAME, so that we reduce the kernel/userspace
1931          * traffic to just what is needed for each syscall.
1932          */
1933         int args_size = raw_augmented_args_size ?: sc->args_size;
1934
1935         *augmented_args_size = sample->raw_size - args_size;
1936         if (*augmented_args_size > 0)
1937                 augmented_args = sample->raw_data + args_size;
1938
1939         return augmented_args;
1940 }
1941
1942 static int trace__sys_enter(struct trace *trace, struct evsel *evsel,
1943                             union perf_event *event __maybe_unused,
1944                             struct perf_sample *sample)
1945 {
1946         char *msg;
1947         void *args;
1948         int printed = 0;
1949         struct thread *thread;
1950         int id = perf_evsel__sc_tp_uint(evsel, id, sample), err = -1;
1951         int augmented_args_size = 0;
1952         void *augmented_args = NULL;
1953         struct syscall *sc = trace__syscall_info(trace, evsel, id);
1954         struct thread_trace *ttrace;
1955
1956         if (sc == NULL)
1957                 return -1;
1958
1959         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
1960         ttrace = thread__trace(thread, trace->output);
1961         if (ttrace == NULL)
1962                 goto out_put;
1963
1964         trace__fprintf_sample(trace, evsel, sample, thread);
1965
1966         args = perf_evsel__sc_tp_ptr(evsel, args, sample);
1967
1968         if (ttrace->entry_str == NULL) {
1969                 ttrace->entry_str = malloc(trace__entry_str_size);
1970                 if (!ttrace->entry_str)
1971                         goto out_put;
1972         }
1973
1974         if (!(trace->duration_filter || trace->summary_only || trace->min_stack))
1975                 trace__printf_interrupted_entry(trace);
1976         /*
1977          * If this is raw_syscalls.sys_enter, then it always comes with the 6 possible
1978          * arguments, even if the syscall being handled, say "openat", uses only 4 arguments
1979          * this breaks syscall__augmented_args() check for augmented args, as we calculate
1980          * syscall->args_size using each syscalls:sys_enter_NAME tracefs format file,
1981          * so when handling, say the openat syscall, we end up getting 6 args for the
1982          * raw_syscalls:sys_enter event, when we expected just 4, we end up mistakenly
1983          * thinking that the extra 2 u64 args are the augmented filename, so just check
1984          * here and avoid using augmented syscalls when the evsel is the raw_syscalls one.
1985          */
1986         if (evsel != trace->syscalls.events.sys_enter)
1987                 augmented_args = syscall__augmented_args(sc, sample, &augmented_args_size, trace->raw_augmented_syscalls_args_size);
1988         ttrace->entry_time = sample->time;
1989         msg = ttrace->entry_str;
1990         printed += scnprintf(msg + printed, trace__entry_str_size - printed, "%s(", sc->name);
1991
1992         printed += syscall__scnprintf_args(sc, msg + printed, trace__entry_str_size - printed,
1993                                            args, augmented_args, augmented_args_size, trace, thread);
1994
1995         if (sc->is_exit) {
1996                 if (!(trace->duration_filter || trace->summary_only || trace->failure_only || trace->min_stack)) {
1997                         int alignment = 0;
1998
1999                         trace__fprintf_entry_head(trace, thread, 0, false, ttrace->entry_time, trace->output);
2000                         printed = fprintf(trace->output, "%s)", ttrace->entry_str);
2001                         if (trace->args_alignment > printed)
2002                                 alignment = trace->args_alignment - printed;
2003                         fprintf(trace->output, "%*s= ?\n", alignment, " ");
2004                 }
2005         } else {
2006                 ttrace->entry_pending = true;
2007                 /* See trace__vfs_getname & trace__sys_exit */
2008                 ttrace->filename.pending_open = false;
2009         }
2010
2011         if (trace->current != thread) {
2012                 thread__put(trace->current);
2013                 trace->current = thread__get(thread);
2014         }
2015         err = 0;
2016 out_put:
2017         thread__put(thread);
2018         return err;
2019 }
2020
2021 static int trace__fprintf_sys_enter(struct trace *trace, struct evsel *evsel,
2022                                     struct perf_sample *sample)
2023 {
2024         struct thread_trace *ttrace;
2025         struct thread *thread;
2026         int id = perf_evsel__sc_tp_uint(evsel, id, sample), err = -1;
2027         struct syscall *sc = trace__syscall_info(trace, evsel, id);
2028         char msg[1024];
2029         void *args, *augmented_args = NULL;
2030         int augmented_args_size;
2031
2032         if (sc == NULL)
2033                 return -1;
2034
2035         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2036         ttrace = thread__trace(thread, trace->output);
2037         /*
2038          * We need to get ttrace just to make sure it is there when syscall__scnprintf_args()
2039          * and the rest of the beautifiers accessing it via struct syscall_arg touches it.
2040          */
2041         if (ttrace == NULL)
2042                 goto out_put;
2043
2044         args = perf_evsel__sc_tp_ptr(evsel, args, sample);
2045         augmented_args = syscall__augmented_args(sc, sample, &augmented_args_size, trace->raw_augmented_syscalls_args_size);
2046         syscall__scnprintf_args(sc, msg, sizeof(msg), args, augmented_args, augmented_args_size, trace, thread);
2047         fprintf(trace->output, "%s", msg);
2048         err = 0;
2049 out_put:
2050         thread__put(thread);
2051         return err;
2052 }
2053
2054 static int trace__resolve_callchain(struct trace *trace, struct evsel *evsel,
2055                                     struct perf_sample *sample,
2056                                     struct callchain_cursor *cursor)
2057 {
2058         struct addr_location al;
2059         int max_stack = evsel->core.attr.sample_max_stack ?
2060                         evsel->core.attr.sample_max_stack :
2061                         trace->max_stack;
2062         int err;
2063
2064         if (machine__resolve(trace->host, &al, sample) < 0)
2065                 return -1;
2066
2067         err = thread__resolve_callchain(al.thread, cursor, evsel, sample, NULL, NULL, max_stack);
2068         addr_location__put(&al);
2069         return err;
2070 }
2071
2072 static int trace__fprintf_callchain(struct trace *trace, struct perf_sample *sample)
2073 {
2074         /* TODO: user-configurable print_opts */
2075         const unsigned int print_opts = EVSEL__PRINT_SYM |
2076                                         EVSEL__PRINT_DSO |
2077                                         EVSEL__PRINT_UNKNOWN_AS_ADDR;
2078
2079         return sample__fprintf_callchain(sample, 38, print_opts, &callchain_cursor, trace->output);
2080 }
2081
2082 static const char *errno_to_name(struct evsel *evsel, int err)
2083 {
2084         struct perf_env *env = perf_evsel__env(evsel);
2085         const char *arch_name = perf_env__arch(env);
2086
2087         return arch_syscalls__strerrno(arch_name, err);
2088 }
2089
2090 static int trace__sys_exit(struct trace *trace, struct evsel *evsel,
2091                            union perf_event *event __maybe_unused,
2092                            struct perf_sample *sample)
2093 {
2094         long ret;
2095         u64 duration = 0;
2096         bool duration_calculated = false;
2097         struct thread *thread;
2098         int id = perf_evsel__sc_tp_uint(evsel, id, sample), err = -1, callchain_ret = 0, printed = 0;
2099         int alignment = trace->args_alignment;
2100         struct syscall *sc = trace__syscall_info(trace, evsel, id);
2101         struct thread_trace *ttrace;
2102
2103         if (sc == NULL)
2104                 return -1;
2105
2106         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2107         ttrace = thread__trace(thread, trace->output);
2108         if (ttrace == NULL)
2109                 goto out_put;
2110
2111         trace__fprintf_sample(trace, evsel, sample, thread);
2112
2113         if (trace->summary)
2114                 thread__update_stats(ttrace, id, sample);
2115
2116         ret = perf_evsel__sc_tp_uint(evsel, ret, sample);
2117
2118         if (!trace->fd_path_disabled && sc->is_open && ret >= 0 && ttrace->filename.pending_open) {
2119                 trace__set_fd_pathname(thread, ret, ttrace->filename.name);
2120                 ttrace->filename.pending_open = false;
2121                 ++trace->stats.vfs_getname;
2122         }
2123
2124         if (ttrace->entry_time) {
2125                 duration = sample->time - ttrace->entry_time;
2126                 if (trace__filter_duration(trace, duration))
2127                         goto out;
2128                 duration_calculated = true;
2129         } else if (trace->duration_filter)
2130                 goto out;
2131
2132         if (sample->callchain) {
2133                 callchain_ret = trace__resolve_callchain(trace, evsel, sample, &callchain_cursor);
2134                 if (callchain_ret == 0) {
2135                         if (callchain_cursor.nr < trace->min_stack)
2136                                 goto out;
2137                         callchain_ret = 1;
2138                 }
2139         }
2140
2141         if (trace->summary_only || (ret >= 0 && trace->failure_only))
2142                 goto out;
2143
2144         trace__fprintf_entry_head(trace, thread, duration, duration_calculated, ttrace->entry_time, trace->output);
2145
2146         if (ttrace->entry_pending) {
2147                 printed = fprintf(trace->output, "%s", ttrace->entry_str);
2148         } else {
2149                 printed += fprintf(trace->output, " ... [");
2150                 color_fprintf(trace->output, PERF_COLOR_YELLOW, "continued");
2151                 printed += 9;
2152                 printed += fprintf(trace->output, "]: %s()", sc->name);
2153         }
2154
2155         printed++; /* the closing ')' */
2156
2157         if (alignment > printed)
2158                 alignment -= printed;
2159         else
2160                 alignment = 0;
2161
2162         fprintf(trace->output, ")%*s= ", alignment, " ");
2163
2164         if (sc->fmt == NULL) {
2165                 if (ret < 0)
2166                         goto errno_print;
2167 signed_print:
2168                 fprintf(trace->output, "%ld", ret);
2169         } else if (ret < 0) {
2170 errno_print: {
2171                 char bf[STRERR_BUFSIZE];
2172                 const char *emsg = str_error_r(-ret, bf, sizeof(bf)),
2173                            *e = errno_to_name(evsel, -ret);
2174
2175                 fprintf(trace->output, "-1 %s (%s)", e, emsg);
2176         }
2177         } else if (ret == 0 && sc->fmt->timeout)
2178                 fprintf(trace->output, "0 (Timeout)");
2179         else if (ttrace->ret_scnprintf) {
2180                 char bf[1024];
2181                 struct syscall_arg arg = {
2182                         .val    = ret,
2183                         .thread = thread,
2184                         .trace  = trace,
2185                 };
2186                 ttrace->ret_scnprintf(bf, sizeof(bf), &arg);
2187                 ttrace->ret_scnprintf = NULL;
2188                 fprintf(trace->output, "%s", bf);
2189         } else if (sc->fmt->hexret)
2190                 fprintf(trace->output, "%#lx", ret);
2191         else if (sc->fmt->errpid) {
2192                 struct thread *child = machine__find_thread(trace->host, ret, ret);
2193
2194                 if (child != NULL) {
2195                         fprintf(trace->output, "%ld", ret);
2196                         if (child->comm_set)
2197                                 fprintf(trace->output, " (%s)", thread__comm_str(child));
2198                         thread__put(child);
2199                 }
2200         } else
2201                 goto signed_print;
2202
2203         fputc('\n', trace->output);
2204
2205         /*
2206          * We only consider an 'event' for the sake of --max-events a non-filtered
2207          * sys_enter + sys_exit and other tracepoint events.
2208          */
2209         if (++trace->nr_events_printed == trace->max_events && trace->max_events != ULONG_MAX)
2210                 interrupted = true;
2211
2212         if (callchain_ret > 0)
2213                 trace__fprintf_callchain(trace, sample);
2214         else if (callchain_ret < 0)
2215                 pr_err("Problem processing %s callchain, skipping...\n", perf_evsel__name(evsel));
2216 out:
2217         ttrace->entry_pending = false;
2218         err = 0;
2219 out_put:
2220         thread__put(thread);
2221         return err;
2222 }
2223
2224 static int trace__vfs_getname(struct trace *trace, struct evsel *evsel,
2225                               union perf_event *event __maybe_unused,
2226                               struct perf_sample *sample)
2227 {
2228         struct thread *thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2229         struct thread_trace *ttrace;
2230         size_t filename_len, entry_str_len, to_move;
2231         ssize_t remaining_space;
2232         char *pos;
2233         const char *filename = perf_evsel__rawptr(evsel, sample, "pathname");
2234
2235         if (!thread)
2236                 goto out;
2237
2238         ttrace = thread__priv(thread);
2239         if (!ttrace)
2240                 goto out_put;
2241
2242         filename_len = strlen(filename);
2243         if (filename_len == 0)
2244                 goto out_put;
2245
2246         if (ttrace->filename.namelen < filename_len) {
2247                 char *f = realloc(ttrace->filename.name, filename_len + 1);
2248
2249                 if (f == NULL)
2250                         goto out_put;
2251
2252                 ttrace->filename.namelen = filename_len;
2253                 ttrace->filename.name = f;
2254         }
2255
2256         strcpy(ttrace->filename.name, filename);
2257         ttrace->filename.pending_open = true;
2258
2259         if (!ttrace->filename.ptr)
2260                 goto out_put;
2261
2262         entry_str_len = strlen(ttrace->entry_str);
2263         remaining_space = trace__entry_str_size - entry_str_len - 1; /* \0 */
2264         if (remaining_space <= 0)
2265                 goto out_put;
2266
2267         if (filename_len > (size_t)remaining_space) {
2268                 filename += filename_len - remaining_space;
2269                 filename_len = remaining_space;
2270         }
2271
2272         to_move = entry_str_len - ttrace->filename.entry_str_pos + 1; /* \0 */
2273         pos = ttrace->entry_str + ttrace->filename.entry_str_pos;
2274         memmove(pos + filename_len, pos, to_move);
2275         memcpy(pos, filename, filename_len);
2276
2277         ttrace->filename.ptr = 0;
2278         ttrace->filename.entry_str_pos = 0;
2279 out_put:
2280         thread__put(thread);
2281 out:
2282         return 0;
2283 }
2284
2285 static int trace__sched_stat_runtime(struct trace *trace, struct evsel *evsel,
2286                                      union perf_event *event __maybe_unused,
2287                                      struct perf_sample *sample)
2288 {
2289         u64 runtime = perf_evsel__intval(evsel, sample, "runtime");
2290         double runtime_ms = (double)runtime / NSEC_PER_MSEC;
2291         struct thread *thread = machine__findnew_thread(trace->host,
2292                                                         sample->pid,
2293                                                         sample->tid);
2294         struct thread_trace *ttrace = thread__trace(thread, trace->output);
2295
2296         if (ttrace == NULL)
2297                 goto out_dump;
2298
2299         ttrace->runtime_ms += runtime_ms;
2300         trace->runtime_ms += runtime_ms;
2301 out_put:
2302         thread__put(thread);
2303         return 0;
2304
2305 out_dump:
2306         fprintf(trace->output, "%s: comm=%s,pid=%u,runtime=%" PRIu64 ",vruntime=%" PRIu64 ")\n",
2307                evsel->name,
2308                perf_evsel__strval(evsel, sample, "comm"),
2309                (pid_t)perf_evsel__intval(evsel, sample, "pid"),
2310                runtime,
2311                perf_evsel__intval(evsel, sample, "vruntime"));
2312         goto out_put;
2313 }
2314
2315 static int bpf_output__printer(enum binary_printer_ops op,
2316                                unsigned int val, void *extra __maybe_unused, FILE *fp)
2317 {
2318         unsigned char ch = (unsigned char)val;
2319
2320         switch (op) {
2321         case BINARY_PRINT_CHAR_DATA:
2322                 return fprintf(fp, "%c", isprint(ch) ? ch : '.');
2323         case BINARY_PRINT_DATA_BEGIN:
2324         case BINARY_PRINT_LINE_BEGIN:
2325         case BINARY_PRINT_ADDR:
2326         case BINARY_PRINT_NUM_DATA:
2327         case BINARY_PRINT_NUM_PAD:
2328         case BINARY_PRINT_SEP:
2329         case BINARY_PRINT_CHAR_PAD:
2330         case BINARY_PRINT_LINE_END:
2331         case BINARY_PRINT_DATA_END:
2332         default:
2333                 break;
2334         }
2335
2336         return 0;
2337 }
2338
2339 static void bpf_output__fprintf(struct trace *trace,
2340                                 struct perf_sample *sample)
2341 {
2342         binary__fprintf(sample->raw_data, sample->raw_size, 8,
2343                         bpf_output__printer, NULL, trace->output);
2344         ++trace->nr_events_printed;
2345 }
2346
2347 static int trace__event_handler(struct trace *trace, struct evsel *evsel,
2348                                 union perf_event *event __maybe_unused,
2349                                 struct perf_sample *sample)
2350 {
2351         struct thread *thread;
2352         int callchain_ret = 0;
2353         /*
2354          * Check if we called perf_evsel__disable(evsel) due to, for instance,
2355          * this event's max_events having been hit and this is an entry coming
2356          * from the ring buffer that we should discard, since the max events
2357          * have already been considered/printed.
2358          */
2359         if (evsel->disabled)
2360                 return 0;
2361
2362         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2363
2364         if (sample->callchain) {
2365                 callchain_ret = trace__resolve_callchain(trace, evsel, sample, &callchain_cursor);
2366                 if (callchain_ret == 0) {
2367                         if (callchain_cursor.nr < trace->min_stack)
2368                                 goto out;
2369                         callchain_ret = 1;
2370                 }
2371         }
2372
2373         trace__printf_interrupted_entry(trace);
2374         trace__fprintf_tstamp(trace, sample->time, trace->output);
2375
2376         if (trace->trace_syscalls && trace->show_duration)
2377                 fprintf(trace->output, "(         ): ");
2378
2379         if (thread)
2380                 trace__fprintf_comm_tid(trace, thread, trace->output);
2381
2382         if (evsel == trace->syscalls.events.augmented) {
2383                 int id = perf_evsel__sc_tp_uint(evsel, id, sample);
2384                 struct syscall *sc = trace__syscall_info(trace, evsel, id);
2385
2386                 if (sc) {
2387                         fprintf(trace->output, "%s(", sc->name);
2388                         trace__fprintf_sys_enter(trace, evsel, sample);
2389                         fputc(')', trace->output);
2390                         goto newline;
2391                 }
2392
2393                 /*
2394                  * XXX: Not having the associated syscall info or not finding/adding
2395                  *      the thread should never happen, but if it does...
2396                  *      fall thru and print it as a bpf_output event.
2397                  */
2398         }
2399
2400         fprintf(trace->output, "%s:", evsel->name);
2401
2402         if (perf_evsel__is_bpf_output(evsel)) {
2403                 bpf_output__fprintf(trace, sample);
2404         } else if (evsel->tp_format) {
2405                 if (strncmp(evsel->tp_format->name, "sys_enter_", 10) ||
2406                     trace__fprintf_sys_enter(trace, evsel, sample)) {
2407                         event_format__fprintf(evsel->tp_format, sample->cpu,
2408                                               sample->raw_data, sample->raw_size,
2409                                               trace->output);
2410                         ++trace->nr_events_printed;
2411
2412                         if (evsel->max_events != ULONG_MAX && ++evsel->nr_events_printed == evsel->max_events) {
2413                                 evsel__disable(evsel);
2414                                 evsel__close(evsel);
2415                         }
2416                 }
2417         }
2418
2419 newline:
2420         fprintf(trace->output, "\n");
2421
2422         if (callchain_ret > 0)
2423                 trace__fprintf_callchain(trace, sample);
2424         else if (callchain_ret < 0)
2425                 pr_err("Problem processing %s callchain, skipping...\n", perf_evsel__name(evsel));
2426 out:
2427         thread__put(thread);
2428         return 0;
2429 }
2430
2431 static void print_location(FILE *f, struct perf_sample *sample,
2432                            struct addr_location *al,
2433                            bool print_dso, bool print_sym)
2434 {
2435
2436         if ((verbose > 0 || print_dso) && al->map)
2437                 fprintf(f, "%s@", al->map->dso->long_name);
2438
2439         if ((verbose > 0 || print_sym) && al->sym)
2440                 fprintf(f, "%s+0x%" PRIx64, al->sym->name,
2441                         al->addr - al->sym->start);
2442         else if (al->map)
2443                 fprintf(f, "0x%" PRIx64, al->addr);
2444         else
2445                 fprintf(f, "0x%" PRIx64, sample->addr);
2446 }
2447
2448 static int trace__pgfault(struct trace *trace,
2449                           struct evsel *evsel,
2450                           union perf_event *event __maybe_unused,
2451                           struct perf_sample *sample)
2452 {
2453         struct thread *thread;
2454         struct addr_location al;
2455         char map_type = 'd';
2456         struct thread_trace *ttrace;
2457         int err = -1;
2458         int callchain_ret = 0;
2459
2460         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2461
2462         if (sample->callchain) {
2463                 callchain_ret = trace__resolve_callchain(trace, evsel, sample, &callchain_cursor);
2464                 if (callchain_ret == 0) {
2465                         if (callchain_cursor.nr < trace->min_stack)
2466                                 goto out_put;
2467                         callchain_ret = 1;
2468                 }
2469         }
2470
2471         ttrace = thread__trace(thread, trace->output);
2472         if (ttrace == NULL)
2473                 goto out_put;
2474
2475         if (evsel->core.attr.config == PERF_COUNT_SW_PAGE_FAULTS_MAJ)
2476                 ttrace->pfmaj++;
2477         else
2478                 ttrace->pfmin++;
2479
2480         if (trace->summary_only)
2481                 goto out;
2482
2483         thread__find_symbol(thread, sample->cpumode, sample->ip, &al);
2484
2485         trace__fprintf_entry_head(trace, thread, 0, true, sample->time, trace->output);
2486
2487         fprintf(trace->output, "%sfault [",
2488                 evsel->core.attr.config == PERF_COUNT_SW_PAGE_FAULTS_MAJ ?
2489                 "maj" : "min");
2490
2491         print_location(trace->output, sample, &al, false, true);
2492
2493         fprintf(trace->output, "] => ");
2494
2495         thread__find_symbol(thread, sample->cpumode, sample->addr, &al);
2496
2497         if (!al.map) {
2498                 thread__find_symbol(thread, sample->cpumode, sample->addr, &al);
2499
2500                 if (al.map)
2501                         map_type = 'x';
2502                 else
2503                         map_type = '?';
2504         }
2505
2506         print_location(trace->output, sample, &al, true, false);
2507
2508         fprintf(trace->output, " (%c%c)\n", map_type, al.level);
2509
2510         if (callchain_ret > 0)
2511                 trace__fprintf_callchain(trace, sample);
2512         else if (callchain_ret < 0)
2513                 pr_err("Problem processing %s callchain, skipping...\n", perf_evsel__name(evsel));
2514
2515         ++trace->nr_events_printed;
2516 out:
2517         err = 0;
2518 out_put:
2519         thread__put(thread);
2520         return err;
2521 }
2522
2523 static void trace__set_base_time(struct trace *trace,
2524                                  struct evsel *evsel,
2525                                  struct perf_sample *sample)
2526 {
2527         /*
2528          * BPF events were not setting PERF_SAMPLE_TIME, so be more robust
2529          * and don't use sample->time unconditionally, we may end up having
2530          * some other event in the future without PERF_SAMPLE_TIME for good
2531          * reason, i.e. we may not be interested in its timestamps, just in
2532          * it taking place, picking some piece of information when it
2533          * appears in our event stream (vfs_getname comes to mind).
2534          */
2535         if (trace->base_time == 0 && !trace->full_time &&
2536             (evsel->core.attr.sample_type & PERF_SAMPLE_TIME))
2537                 trace->base_time = sample->time;
2538 }
2539
2540 static int trace__process_sample(struct perf_tool *tool,
2541                                  union perf_event *event,
2542                                  struct perf_sample *sample,
2543                                  struct evsel *evsel,
2544                                  struct machine *machine __maybe_unused)
2545 {
2546         struct trace *trace = container_of(tool, struct trace, tool);
2547         struct thread *thread;
2548         int err = 0;
2549
2550         tracepoint_handler handler = evsel->handler;
2551
2552         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2553         if (thread && thread__is_filtered(thread))
2554                 goto out;
2555
2556         trace__set_base_time(trace, evsel, sample);
2557
2558         if (handler) {
2559                 ++trace->nr_events;
2560                 handler(trace, evsel, event, sample);
2561         }
2562 out:
2563         thread__put(thread);
2564         return err;
2565 }
2566
2567 static int trace__record(struct trace *trace, int argc, const char **argv)
2568 {
2569         unsigned int rec_argc, i, j;
2570         const char **rec_argv;
2571         const char * const record_args[] = {
2572                 "record",
2573                 "-R",
2574                 "-m", "1024",
2575                 "-c", "1",
2576         };
2577
2578         const char * const sc_args[] = { "-e", };
2579         unsigned int sc_args_nr = ARRAY_SIZE(sc_args);
2580         const char * const majpf_args[] = { "-e", "major-faults" };
2581         unsigned int majpf_args_nr = ARRAY_SIZE(majpf_args);
2582         const char * const minpf_args[] = { "-e", "minor-faults" };
2583         unsigned int minpf_args_nr = ARRAY_SIZE(minpf_args);
2584
2585         /* +1 is for the event string below */
2586         rec_argc = ARRAY_SIZE(record_args) + sc_args_nr + 1 +
2587                 majpf_args_nr + minpf_args_nr + argc;
2588         rec_argv = calloc(rec_argc + 1, sizeof(char *));
2589
2590         if (rec_argv == NULL)
2591                 return -ENOMEM;
2592
2593         j = 0;
2594         for (i = 0; i < ARRAY_SIZE(record_args); i++)
2595                 rec_argv[j++] = record_args[i];
2596
2597         if (trace->trace_syscalls) {
2598                 for (i = 0; i < sc_args_nr; i++)
2599                         rec_argv[j++] = sc_args[i];
2600
2601                 /* event string may be different for older kernels - e.g., RHEL6 */
2602                 if (is_valid_tracepoint("raw_syscalls:sys_enter"))
2603                         rec_argv[j++] = "raw_syscalls:sys_enter,raw_syscalls:sys_exit";
2604                 else if (is_valid_tracepoint("syscalls:sys_enter"))
2605                         rec_argv[j++] = "syscalls:sys_enter,syscalls:sys_exit";
2606                 else {
2607                         pr_err("Neither raw_syscalls nor syscalls events exist.\n");
2608                         free(rec_argv);
2609                         return -1;
2610                 }
2611         }
2612
2613         if (trace->trace_pgfaults & TRACE_PFMAJ)
2614                 for (i = 0; i < majpf_args_nr; i++)
2615                         rec_argv[j++] = majpf_args[i];
2616
2617         if (trace->trace_pgfaults & TRACE_PFMIN)
2618                 for (i = 0; i < minpf_args_nr; i++)
2619                         rec_argv[j++] = minpf_args[i];
2620
2621         for (i = 0; i < (unsigned int)argc; i++)
2622                 rec_argv[j++] = argv[i];
2623
2624         return cmd_record(j, rec_argv);
2625 }
2626
2627 static size_t trace__fprintf_thread_summary(struct trace *trace, FILE *fp);
2628
2629 static bool evlist__add_vfs_getname(struct evlist *evlist)
2630 {
2631         bool found = false;
2632         struct evsel *evsel, *tmp;
2633         struct parse_events_error err = { .idx = 0, };
2634         int ret = parse_events(evlist, "probe:vfs_getname*", &err);
2635
2636         if (ret)
2637                 return false;
2638
2639         evlist__for_each_entry_safe(evlist, evsel, tmp) {
2640                 if (!strstarts(perf_evsel__name(evsel), "probe:vfs_getname"))
2641                         continue;
2642
2643                 if (perf_evsel__field(evsel, "pathname")) {
2644                         evsel->handler = trace__vfs_getname;
2645                         found = true;
2646                         continue;
2647                 }
2648
2649                 list_del_init(&evsel->core.node);
2650                 evsel->evlist = NULL;
2651                 evsel__delete(evsel);
2652         }
2653
2654         return found;
2655 }
2656
2657 static struct evsel *perf_evsel__new_pgfault(u64 config)
2658 {
2659         struct evsel *evsel;
2660         struct perf_event_attr attr = {
2661                 .type = PERF_TYPE_SOFTWARE,
2662                 .mmap_data = 1,
2663         };
2664
2665         attr.config = config;
2666         attr.sample_period = 1;
2667
2668         event_attr_init(&attr);
2669
2670         evsel = evsel__new(&attr);
2671         if (evsel)
2672                 evsel->handler = trace__pgfault;
2673
2674         return evsel;
2675 }
2676
2677 static void trace__handle_event(struct trace *trace, union perf_event *event, struct perf_sample *sample)
2678 {
2679         const u32 type = event->header.type;
2680         struct evsel *evsel;
2681
2682         if (type != PERF_RECORD_SAMPLE) {
2683                 trace__process_event(trace, trace->host, event, sample);
2684                 return;
2685         }
2686
2687         evsel = perf_evlist__id2evsel(trace->evlist, sample->id);
2688         if (evsel == NULL) {
2689                 fprintf(trace->output, "Unknown tp ID %" PRIu64 ", skipping...\n", sample->id);
2690                 return;
2691         }
2692
2693         if (evswitch__discard(&trace->evswitch, evsel))
2694                 return;
2695
2696         trace__set_base_time(trace, evsel, sample);
2697
2698         if (evsel->core.attr.type == PERF_TYPE_TRACEPOINT &&
2699             sample->raw_data == NULL) {
2700                 fprintf(trace->output, "%s sample with no payload for tid: %d, cpu %d, raw_size=%d, skipping...\n",
2701                        perf_evsel__name(evsel), sample->tid,
2702                        sample->cpu, sample->raw_size);
2703         } else {
2704                 tracepoint_handler handler = evsel->handler;
2705                 handler(trace, evsel, event, sample);
2706         }
2707
2708         if (trace->nr_events_printed >= trace->max_events && trace->max_events != ULONG_MAX)
2709                 interrupted = true;
2710 }
2711
2712 static int trace__add_syscall_newtp(struct trace *trace)
2713 {
2714         int ret = -1;
2715         struct evlist *evlist = trace->evlist;
2716         struct evsel *sys_enter, *sys_exit;
2717
2718         sys_enter = perf_evsel__raw_syscall_newtp("sys_enter", trace__sys_enter);
2719         if (sys_enter == NULL)
2720                 goto out;
2721
2722         if (perf_evsel__init_sc_tp_ptr_field(sys_enter, args))
2723                 goto out_delete_sys_enter;
2724
2725         sys_exit = perf_evsel__raw_syscall_newtp("sys_exit", trace__sys_exit);
2726         if (sys_exit == NULL)
2727                 goto out_delete_sys_enter;
2728
2729         if (perf_evsel__init_sc_tp_uint_field(sys_exit, ret))
2730                 goto out_delete_sys_exit;
2731
2732         perf_evsel__config_callchain(sys_enter, &trace->opts, &callchain_param);
2733         perf_evsel__config_callchain(sys_exit, &trace->opts, &callchain_param);
2734
2735         evlist__add(evlist, sys_enter);
2736         evlist__add(evlist, sys_exit);
2737
2738         if (callchain_param.enabled && !trace->kernel_syscallchains) {
2739                 /*
2740                  * We're interested only in the user space callchain
2741                  * leading to the syscall, allow overriding that for
2742                  * debugging reasons using --kernel_syscall_callchains
2743                  */
2744                 sys_exit->core.attr.exclude_callchain_kernel = 1;
2745         }
2746
2747         trace->syscalls.events.sys_enter = sys_enter;
2748         trace->syscalls.events.sys_exit  = sys_exit;
2749
2750         ret = 0;
2751 out:
2752         return ret;
2753
2754 out_delete_sys_exit:
2755         evsel__delete_priv(sys_exit);
2756 out_delete_sys_enter:
2757         evsel__delete_priv(sys_enter);
2758         goto out;
2759 }
2760
2761 static int trace__set_ev_qualifier_tp_filter(struct trace *trace)
2762 {
2763         int err = -1;
2764         struct evsel *sys_exit;
2765         char *filter = asprintf_expr_inout_ints("id", !trace->not_ev_qualifier,
2766                                                 trace->ev_qualifier_ids.nr,
2767                                                 trace->ev_qualifier_ids.entries);
2768
2769         if (filter == NULL)
2770                 goto out_enomem;
2771
2772         if (!perf_evsel__append_tp_filter(trace->syscalls.events.sys_enter,
2773                                           filter)) {
2774                 sys_exit = trace->syscalls.events.sys_exit;
2775                 err = perf_evsel__append_tp_filter(sys_exit, filter);
2776         }
2777
2778         free(filter);
2779 out:
2780         return err;
2781 out_enomem:
2782         errno = ENOMEM;
2783         goto out;
2784 }
2785
2786 #ifdef HAVE_LIBBPF_SUPPORT
2787 static struct bpf_program *trace__find_bpf_program_by_title(struct trace *trace, const char *name)
2788 {
2789         if (trace->bpf_obj == NULL)
2790                 return NULL;
2791
2792         return bpf_object__find_program_by_title(trace->bpf_obj, name);
2793 }
2794
2795 static struct bpf_program *trace__find_syscall_bpf_prog(struct trace *trace, struct syscall *sc,
2796                                                         const char *prog_name, const char *type)
2797 {
2798         struct bpf_program *prog;
2799
2800         if (prog_name == NULL) {
2801                 char default_prog_name[256];
2802                 scnprintf(default_prog_name, sizeof(default_prog_name), "!syscalls:sys_%s_%s", type, sc->name);
2803                 prog = trace__find_bpf_program_by_title(trace, default_prog_name);
2804                 if (prog != NULL)
2805                         goto out_found;
2806                 if (sc->fmt && sc->fmt->alias) {
2807                         scnprintf(default_prog_name, sizeof(default_prog_name), "!syscalls:sys_%s_%s", type, sc->fmt->alias);
2808                         prog = trace__find_bpf_program_by_title(trace, default_prog_name);
2809                         if (prog != NULL)
2810                                 goto out_found;
2811                 }
2812                 goto out_unaugmented;
2813         }
2814
2815         prog = trace__find_bpf_program_by_title(trace, prog_name);
2816
2817         if (prog != NULL) {
2818 out_found:
2819                 return prog;
2820         }
2821
2822         pr_debug("Couldn't find BPF prog \"%s\" to associate with syscalls:sys_%s_%s, not augmenting it\n",
2823                  prog_name, type, sc->name);
2824 out_unaugmented:
2825         return trace->syscalls.unaugmented_prog;
2826 }
2827
2828 static void trace__init_syscall_bpf_progs(struct trace *trace, int id)
2829 {
2830         struct syscall *sc = trace__syscall_info(trace, NULL, id);
2831
2832         if (sc == NULL)
2833                 return;
2834
2835         sc->bpf_prog.sys_enter = trace__find_syscall_bpf_prog(trace, sc, sc->fmt ? sc->fmt->bpf_prog_name.sys_enter : NULL, "enter");
2836         sc->bpf_prog.sys_exit  = trace__find_syscall_bpf_prog(trace, sc, sc->fmt ? sc->fmt->bpf_prog_name.sys_exit  : NULL,  "exit");
2837 }
2838
2839 static int trace__bpf_prog_sys_enter_fd(struct trace *trace, int id)
2840 {
2841         struct syscall *sc = trace__syscall_info(trace, NULL, id);
2842         return sc ? bpf_program__fd(sc->bpf_prog.sys_enter) : bpf_program__fd(trace->syscalls.unaugmented_prog);
2843 }
2844
2845 static int trace__bpf_prog_sys_exit_fd(struct trace *trace, int id)
2846 {
2847         struct syscall *sc = trace__syscall_info(trace, NULL, id);
2848         return sc ? bpf_program__fd(sc->bpf_prog.sys_exit) : bpf_program__fd(trace->syscalls.unaugmented_prog);
2849 }
2850
2851 static void trace__init_bpf_map_syscall_args(struct trace *trace, int id, struct bpf_map_syscall_entry *entry)
2852 {
2853         struct syscall *sc = trace__syscall_info(trace, NULL, id);
2854         int arg = 0;
2855
2856         if (sc == NULL)
2857                 goto out;
2858
2859         for (; arg < sc->nr_args; ++arg) {
2860                 entry->string_args_len[arg] = 0;
2861                 if (sc->arg_fmt[arg].scnprintf == SCA_FILENAME) {
2862                         /* Should be set like strace -s strsize */
2863                         entry->string_args_len[arg] = PATH_MAX;
2864                 }
2865         }
2866 out:
2867         for (; arg < 6; ++arg)
2868                 entry->string_args_len[arg] = 0;
2869 }
2870 static int trace__set_ev_qualifier_bpf_filter(struct trace *trace)
2871 {
2872         int fd = bpf_map__fd(trace->syscalls.map);
2873         struct bpf_map_syscall_entry value = {
2874                 .enabled = !trace->not_ev_qualifier,
2875         };
2876         int err = 0;
2877         size_t i;
2878
2879         for (i = 0; i < trace->ev_qualifier_ids.nr; ++i) {
2880                 int key = trace->ev_qualifier_ids.entries[i];
2881
2882                 if (value.enabled) {
2883                         trace__init_bpf_map_syscall_args(trace, key, &value);
2884                         trace__init_syscall_bpf_progs(trace, key);
2885                 }
2886
2887                 err = bpf_map_update_elem(fd, &key, &value, BPF_EXIST);
2888                 if (err)
2889                         break;
2890         }
2891
2892         return err;
2893 }
2894
2895 static int __trace__init_syscalls_bpf_map(struct trace *trace, bool enabled)
2896 {
2897         int fd = bpf_map__fd(trace->syscalls.map);
2898         struct bpf_map_syscall_entry value = {
2899                 .enabled = enabled,
2900         };
2901         int err = 0, key;
2902
2903         for (key = 0; key < trace->sctbl->syscalls.nr_entries; ++key) {
2904                 if (enabled)
2905                         trace__init_bpf_map_syscall_args(trace, key, &value);
2906
2907                 err = bpf_map_update_elem(fd, &key, &value, BPF_ANY);
2908                 if (err)
2909                         break;
2910         }
2911
2912         return err;
2913 }
2914
2915 static int trace__init_syscalls_bpf_map(struct trace *trace)
2916 {
2917         bool enabled = true;
2918
2919         if (trace->ev_qualifier_ids.nr)
2920                 enabled = trace->not_ev_qualifier;
2921
2922         return __trace__init_syscalls_bpf_map(trace, enabled);
2923 }
2924
2925 static struct bpf_program *trace__find_usable_bpf_prog_entry(struct trace *trace, struct syscall *sc)
2926 {
2927         struct tep_format_field *field, *candidate_field;
2928         int id;
2929
2930         /*
2931          * We're only interested in syscalls that have a pointer:
2932          */
2933         for (field = sc->args; field; field = field->next) {
2934                 if (field->flags & TEP_FIELD_IS_POINTER)
2935                         goto try_to_find_pair;
2936         }
2937
2938         return NULL;
2939
2940 try_to_find_pair:
2941         for (id = 0; id < trace->sctbl->syscalls.nr_entries; ++id) {
2942                 struct syscall *pair = trace__syscall_info(trace, NULL, id);
2943                 struct bpf_program *pair_prog;
2944                 bool is_candidate = false;
2945
2946                 if (pair == NULL || pair == sc ||
2947                     pair->bpf_prog.sys_enter == trace->syscalls.unaugmented_prog)
2948                         continue;
2949
2950                 for (field = sc->args, candidate_field = pair->args;
2951                      field && candidate_field; field = field->next, candidate_field = candidate_field->next) {
2952                         bool is_pointer = field->flags & TEP_FIELD_IS_POINTER,
2953                              candidate_is_pointer = candidate_field->flags & TEP_FIELD_IS_POINTER;
2954
2955                         if (is_pointer) {
2956                                if (!candidate_is_pointer) {
2957                                         // The candidate just doesn't copies our pointer arg, might copy other pointers we want.
2958                                         continue;
2959                                }
2960                         } else {
2961                                 if (candidate_is_pointer) {
2962                                         // The candidate might copy a pointer we don't have, skip it.
2963                                         goto next_candidate;
2964                                 }
2965                                 continue;
2966                         }
2967
2968                         if (strcmp(field->type, candidate_field->type))
2969                                 goto next_candidate;
2970
2971                         is_candidate = true;
2972                 }
2973
2974                 if (!is_candidate)
2975                         goto next_candidate;
2976
2977                 /*
2978                  * Check if the tentative pair syscall augmenter has more pointers, if it has,
2979                  * then it may be collecting that and we then can't use it, as it would collect
2980                  * more than what is common to the two syscalls.
2981                  */
2982                 if (candidate_field) {
2983                         for (candidate_field = candidate_field->next; candidate_field; candidate_field = candidate_field->next)
2984                                 if (candidate_field->flags & TEP_FIELD_IS_POINTER)
2985                                         goto next_candidate;
2986                 }
2987
2988                 pair_prog = pair->bpf_prog.sys_enter;
2989                 /*
2990                  * If the pair isn't enabled, then its bpf_prog.sys_enter will not
2991                  * have been searched for, so search it here and if it returns the
2992                  * unaugmented one, then ignore it, otherwise we'll reuse that BPF
2993                  * program for a filtered syscall on a non-filtered one.
2994                  *
2995                  * For instance, we have "!syscalls:sys_enter_renameat" and that is
2996                  * useful for "renameat2".
2997                  */
2998                 if (pair_prog == NULL) {
2999                         pair_prog = trace__find_syscall_bpf_prog(trace, pair, pair->fmt ? pair->fmt->bpf_prog_name.sys_enter : NULL, "enter");
3000                         if (pair_prog == trace->syscalls.unaugmented_prog)
3001                                 goto next_candidate;
3002                 }
3003
3004                 pr_debug("Reusing \"%s\" BPF sys_enter augmenter for \"%s\"\n", pair->name, sc->name);
3005                 return pair_prog;
3006         next_candidate:
3007                 continue;
3008         }
3009
3010         return NULL;
3011 }
3012
3013 static int trace__init_syscalls_bpf_prog_array_maps(struct trace *trace)
3014 {
3015         int map_enter_fd = bpf_map__fd(trace->syscalls.prog_array.sys_enter),
3016             map_exit_fd  = bpf_map__fd(trace->syscalls.prog_array.sys_exit);
3017         int err = 0, key;
3018
3019         for (key = 0; key < trace->sctbl->syscalls.nr_entries; ++key) {
3020                 int prog_fd;
3021
3022                 if (!trace__syscall_enabled(trace, key))
3023                         continue;
3024
3025                 trace__init_syscall_bpf_progs(trace, key);
3026
3027                 // It'll get at least the "!raw_syscalls:unaugmented"
3028                 prog_fd = trace__bpf_prog_sys_enter_fd(trace, key);
3029                 err = bpf_map_update_elem(map_enter_fd, &key, &prog_fd, BPF_ANY);
3030                 if (err)
3031                         break;
3032                 prog_fd = trace__bpf_prog_sys_exit_fd(trace, key);
3033                 err = bpf_map_update_elem(map_exit_fd, &key, &prog_fd, BPF_ANY);
3034                 if (err)
3035                         break;
3036         }
3037
3038         /*
3039          * Now lets do a second pass looking for enabled syscalls without
3040          * an augmenter that have a signature that is a superset of another
3041          * syscall with an augmenter so that we can auto-reuse it.
3042          *
3043          * I.e. if we have an augmenter for the "open" syscall that has
3044          * this signature:
3045          *
3046          *   int open(const char *pathname, int flags, mode_t mode);
3047          *
3048          * I.e. that will collect just the first string argument, then we
3049          * can reuse it for the 'creat' syscall, that has this signature:
3050          *
3051          *   int creat(const char *pathname, mode_t mode);
3052          *
3053          * and for:
3054          *
3055          *   int stat(const char *pathname, struct stat *statbuf);
3056          *   int lstat(const char *pathname, struct stat *statbuf);
3057          *
3058          * Because the 'open' augmenter will collect the first arg as a string,
3059          * and leave alone all the other args, which already helps with
3060          * beautifying 'stat' and 'lstat''s pathname arg.
3061          *
3062          * Then, in time, when 'stat' gets an augmenter that collects both
3063          * first and second arg (this one on the raw_syscalls:sys_exit prog
3064          * array tail call, then that one will be used.
3065          */
3066         for (key = 0; key < trace->sctbl->syscalls.nr_entries; ++key) {
3067                 struct syscall *sc = trace__syscall_info(trace, NULL, key);
3068                 struct bpf_program *pair_prog;
3069                 int prog_fd;
3070
3071                 if (sc == NULL || sc->bpf_prog.sys_enter == NULL)
3072                         continue;
3073
3074                 /*
3075                  * For now we're just reusing the sys_enter prog, and if it
3076                  * already has an augmenter, we don't need to find one.
3077                  */
3078                 if (sc->bpf_prog.sys_enter != trace->syscalls.unaugmented_prog)
3079                         continue;
3080
3081                 /*
3082                  * Look at all the other syscalls for one that has a signature
3083                  * that is close enough that we can share:
3084                  */
3085                 pair_prog = trace__find_usable_bpf_prog_entry(trace, sc);
3086                 if (pair_prog == NULL)
3087                         continue;
3088
3089                 sc->bpf_prog.sys_enter = pair_prog;
3090
3091                 /*
3092                  * Update the BPF_MAP_TYPE_PROG_SHARED for raw_syscalls:sys_enter
3093                  * with the fd for the program we're reusing:
3094                  */
3095                 prog_fd = bpf_program__fd(sc->bpf_prog.sys_enter);
3096                 err = bpf_map_update_elem(map_enter_fd, &key, &prog_fd, BPF_ANY);
3097                 if (err)
3098                         break;
3099         }
3100
3101
3102         return err;
3103 }
3104 #else
3105 static int trace__set_ev_qualifier_bpf_filter(struct trace *trace __maybe_unused)
3106 {
3107         return 0;
3108 }
3109
3110 static int trace__init_syscalls_bpf_map(struct trace *trace __maybe_unused)
3111 {
3112         return 0;
3113 }
3114
3115 static struct bpf_program *trace__find_bpf_program_by_title(struct trace *trace __maybe_unused,
3116                                                             const char *name __maybe_unused)
3117 {
3118         return NULL;
3119 }
3120
3121 static int trace__init_syscalls_bpf_prog_array_maps(struct trace *trace __maybe_unused)
3122 {
3123         return 0;
3124 }
3125 #endif // HAVE_LIBBPF_SUPPORT
3126
3127 static int trace__set_ev_qualifier_filter(struct trace *trace)
3128 {
3129         if (trace->syscalls.map)
3130                 return trace__set_ev_qualifier_bpf_filter(trace);
3131         if (trace->syscalls.events.sys_enter)
3132                 return trace__set_ev_qualifier_tp_filter(trace);
3133         return 0;
3134 }
3135
3136 static int bpf_map__set_filter_pids(struct bpf_map *map __maybe_unused,
3137                                     size_t npids __maybe_unused, pid_t *pids __maybe_unused)
3138 {
3139         int err = 0;
3140 #ifdef HAVE_LIBBPF_SUPPORT
3141         bool value = true;
3142         int map_fd = bpf_map__fd(map);
3143         size_t i;
3144
3145         for (i = 0; i < npids; ++i) {
3146                 err = bpf_map_update_elem(map_fd, &pids[i], &value, BPF_ANY);
3147                 if (err)
3148                         break;
3149         }
3150 #endif
3151         return err;
3152 }
3153
3154 static int trace__set_filter_loop_pids(struct trace *trace)
3155 {
3156         unsigned int nr = 1, err;
3157         pid_t pids[32] = {
3158                 getpid(),
3159         };
3160         struct thread *thread = machine__find_thread(trace->host, pids[0], pids[0]);
3161
3162         while (thread && nr < ARRAY_SIZE(pids)) {
3163                 struct thread *parent = machine__find_thread(trace->host, thread->ppid, thread->ppid);
3164
3165                 if (parent == NULL)
3166                         break;
3167
3168                 if (!strcmp(thread__comm_str(parent), "sshd") ||
3169                     strstarts(thread__comm_str(parent), "gnome-terminal")) {
3170                         pids[nr++] = parent->tid;
3171                         break;
3172                 }
3173                 thread = parent;
3174         }
3175
3176         err = perf_evlist__set_tp_filter_pids(trace->evlist, nr, pids);
3177         if (!err && trace->filter_pids.map)
3178                 err = bpf_map__set_filter_pids(trace->filter_pids.map, nr, pids);
3179
3180         return err;
3181 }
3182
3183 static int trace__set_filter_pids(struct trace *trace)
3184 {
3185         int err = 0;
3186         /*
3187          * Better not use !target__has_task() here because we need to cover the
3188          * case where no threads were specified in the command line, but a
3189          * workload was, and in that case we will fill in the thread_map when
3190          * we fork the workload in perf_evlist__prepare_workload.
3191          */
3192         if (trace->filter_pids.nr > 0) {
3193                 err = perf_evlist__set_tp_filter_pids(trace->evlist, trace->filter_pids.nr,
3194                                                       trace->filter_pids.entries);
3195                 if (!err && trace->filter_pids.map) {
3196                         err = bpf_map__set_filter_pids(trace->filter_pids.map, trace->filter_pids.nr,
3197                                                        trace->filter_pids.entries);
3198                 }
3199         } else if (perf_thread_map__pid(trace->evlist->core.threads, 0) == -1) {
3200                 err = trace__set_filter_loop_pids(trace);
3201         }
3202
3203         return err;
3204 }
3205
3206 static int __trace__deliver_event(struct trace *trace, union perf_event *event)
3207 {
3208         struct evlist *evlist = trace->evlist;
3209         struct perf_sample sample;
3210         int err;
3211
3212         err = perf_evlist__parse_sample(evlist, event, &sample);
3213         if (err)
3214                 fprintf(trace->output, "Can't parse sample, err = %d, skipping...\n", err);
3215         else
3216                 trace__handle_event(trace, event, &sample);
3217
3218         return 0;
3219 }
3220
3221 static int __trace__flush_events(struct trace *trace)
3222 {
3223         u64 first = ordered_events__first_time(&trace->oe.data);
3224         u64 flush = trace->oe.last - NSEC_PER_SEC;
3225
3226         /* Is there some thing to flush.. */
3227         if (first && first < flush)
3228                 return ordered_events__flush_time(&trace->oe.data, flush);
3229
3230         return 0;
3231 }
3232
3233 static int trace__flush_events(struct trace *trace)
3234 {
3235         return !trace->sort_events ? 0 : __trace__flush_events(trace);
3236 }
3237
3238 static int trace__deliver_event(struct trace *trace, union perf_event *event)
3239 {
3240         int err;
3241
3242         if (!trace->sort_events)
3243                 return __trace__deliver_event(trace, event);
3244
3245         err = perf_evlist__parse_sample_timestamp(trace->evlist, event, &trace->oe.last);
3246         if (err && err != -1)
3247                 return err;
3248
3249         err = ordered_events__queue(&trace->oe.data, event, trace->oe.last, 0);
3250         if (err)
3251                 return err;
3252
3253         return trace__flush_events(trace);
3254 }
3255
3256 static int ordered_events__deliver_event(struct ordered_events *oe,
3257                                          struct ordered_event *event)
3258 {
3259         struct trace *trace = container_of(oe, struct trace, oe.data);
3260
3261         return __trace__deliver_event(trace, event->event);
3262 }
3263
3264 static int trace__run(struct trace *trace, int argc, const char **argv)
3265 {
3266         struct evlist *evlist = trace->evlist;
3267         struct evsel *evsel, *pgfault_maj = NULL, *pgfault_min = NULL;
3268         int err = -1, i;
3269         unsigned long before;
3270         const bool forks = argc > 0;
3271         bool draining = false;
3272
3273         trace->live = true;
3274
3275         if (!trace->raw_augmented_syscalls) {
3276                 if (trace->trace_syscalls && trace__add_syscall_newtp(trace))
3277                         goto out_error_raw_syscalls;
3278
3279                 if (trace->trace_syscalls)
3280                         trace->vfs_getname = evlist__add_vfs_getname(evlist);
3281         }
3282
3283         if ((trace->trace_pgfaults & TRACE_PFMAJ)) {
3284                 pgfault_maj = perf_evsel__new_pgfault(PERF_COUNT_SW_PAGE_FAULTS_MAJ);
3285                 if (pgfault_maj == NULL)
3286                         goto out_error_mem;
3287                 perf_evsel__config_callchain(pgfault_maj, &trace->opts, &callchain_param);
3288                 evlist__add(evlist, pgfault_maj);
3289         }
3290
3291         if ((trace->trace_pgfaults & TRACE_PFMIN)) {
3292                 pgfault_min = perf_evsel__new_pgfault(PERF_COUNT_SW_PAGE_FAULTS_MIN);
3293                 if (pgfault_min == NULL)
3294                         goto out_error_mem;
3295                 perf_evsel__config_callchain(pgfault_min, &trace->opts, &callchain_param);
3296                 evlist__add(evlist, pgfault_min);
3297         }
3298
3299         if (trace->sched &&
3300             perf_evlist__add_newtp(evlist, "sched", "sched_stat_runtime",
3301                                    trace__sched_stat_runtime))
3302                 goto out_error_sched_stat_runtime;
3303
3304         /*
3305          * If a global cgroup was set, apply it to all the events without an
3306          * explicit cgroup. I.e.:
3307          *
3308          *      trace -G A -e sched:*switch
3309          *
3310          * Will set all raw_syscalls:sys_{enter,exit}, pgfault, vfs_getname, etc
3311          * _and_ sched:sched_switch to the 'A' cgroup, while:
3312          *
3313          * trace -e sched:*switch -G A
3314          *
3315          * will only set the sched:sched_switch event to the 'A' cgroup, all the
3316          * other events (raw_syscalls:sys_{enter,exit}, etc are left "without"
3317          * a cgroup (on the root cgroup, sys wide, etc).
3318          *
3319          * Multiple cgroups:
3320          *
3321          * trace -G A -e sched:*switch -G B
3322          *
3323          * the syscall ones go to the 'A' cgroup, the sched:sched_switch goes
3324          * to the 'B' cgroup.
3325          *
3326          * evlist__set_default_cgroup() grabs a reference of the passed cgroup
3327          * only for the evsels still without a cgroup, i.e. evsel->cgroup == NULL.
3328          */
3329         if (trace->cgroup)
3330                 evlist__set_default_cgroup(trace->evlist, trace->cgroup);
3331
3332         err = perf_evlist__create_maps(evlist, &trace->opts.target);
3333         if (err < 0) {
3334                 fprintf(trace->output, "Problems parsing the target to trace, check your options!\n");
3335                 goto out_delete_evlist;
3336         }
3337
3338         err = trace__symbols_init(trace, evlist);
3339         if (err < 0) {
3340                 fprintf(trace->output, "Problems initializing symbol libraries!\n");
3341                 goto out_delete_evlist;
3342         }
3343
3344         perf_evlist__config(evlist, &trace->opts, &callchain_param);
3345
3346         signal(SIGCHLD, sig_handler);
3347         signal(SIGINT, sig_handler);
3348
3349         if (forks) {
3350                 err = perf_evlist__prepare_workload(evlist, &trace->opts.target,
3351                                                     argv, false, NULL);
3352                 if (err < 0) {
3353                         fprintf(trace->output, "Couldn't run the workload!\n");
3354                         goto out_delete_evlist;
3355                 }
3356         }
3357
3358         err = evlist__open(evlist);
3359         if (err < 0)
3360                 goto out_error_open;
3361
3362         err = bpf__apply_obj_config();
3363         if (err) {
3364                 char errbuf[BUFSIZ];
3365
3366                 bpf__strerror_apply_obj_config(err, errbuf, sizeof(errbuf));
3367                 pr_err("ERROR: Apply config to BPF failed: %s\n",
3368                          errbuf);
3369                 goto out_error_open;
3370         }
3371
3372         err = trace__set_filter_pids(trace);
3373         if (err < 0)
3374                 goto out_error_mem;
3375
3376         if (trace->syscalls.map)
3377                 trace__init_syscalls_bpf_map(trace);
3378
3379         if (trace->syscalls.prog_array.sys_enter)
3380                 trace__init_syscalls_bpf_prog_array_maps(trace);
3381
3382         if (trace->ev_qualifier_ids.nr > 0) {
3383                 err = trace__set_ev_qualifier_filter(trace);
3384                 if (err < 0)
3385                         goto out_errno;
3386
3387                 if (trace->syscalls.events.sys_exit) {
3388                         pr_debug("event qualifier tracepoint filter: %s\n",
3389                                  trace->syscalls.events.sys_exit->filter);
3390                 }
3391         }
3392
3393         /*
3394          * If the "close" syscall is not traced, then we will not have the
3395          * opportunity to, in syscall_arg__scnprintf_close_fd() invalidate the
3396          * fd->pathname table and were ending up showing the last value set by
3397          * syscalls opening a pathname and associating it with a descriptor or
3398          * reading it from /proc/pid/fd/ in cases where that doesn't make
3399          * sense.
3400          *
3401          *  So just disable this beautifier (SCA_FD, SCA_FDAT) when 'close' is
3402          *  not in use.
3403          */
3404         trace->fd_path_disabled = !trace__syscall_enabled(trace, syscalltbl__id(trace->sctbl, "close"));
3405
3406         err = perf_evlist__apply_filters(evlist, &evsel);
3407         if (err < 0)
3408                 goto out_error_apply_filters;
3409
3410         if (trace->dump.map)
3411                 bpf_map__fprintf(trace->dump.map, trace->output);
3412
3413         err = evlist__mmap(evlist, trace->opts.mmap_pages);
3414         if (err < 0)
3415                 goto out_error_mmap;
3416
3417         if (!target__none(&trace->opts.target) && !trace->opts.initial_delay)
3418                 evlist__enable(evlist);
3419
3420         if (forks)
3421                 perf_evlist__start_workload(evlist);
3422
3423         if (trace->opts.initial_delay) {
3424                 usleep(trace->opts.initial_delay * 1000);
3425                 evlist__enable(evlist);
3426         }
3427
3428         trace->multiple_threads = perf_thread_map__pid(evlist->core.threads, 0) == -1 ||
3429                                   evlist->core.threads->nr > 1 ||
3430                                   perf_evlist__first(evlist)->core.attr.inherit;
3431
3432         /*
3433          * Now that we already used evsel->core.attr to ask the kernel to setup the
3434          * events, lets reuse evsel->core.attr.sample_max_stack as the limit in
3435          * trace__resolve_callchain(), allowing per-event max-stack settings
3436          * to override an explicitly set --max-stack global setting.
3437          */
3438         evlist__for_each_entry(evlist, evsel) {
3439                 if (evsel__has_callchain(evsel) &&
3440                     evsel->core.attr.sample_max_stack == 0)
3441                         evsel->core.attr.sample_max_stack = trace->max_stack;
3442         }
3443 again:
3444         before = trace->nr_events;
3445
3446         for (i = 0; i < evlist->core.nr_mmaps; i++) {
3447                 union perf_event *event;
3448                 struct mmap *md;
3449
3450                 md = &evlist->mmap[i];
3451                 if (perf_mmap__read_init(md) < 0)
3452                         continue;
3453
3454                 while ((event = perf_mmap__read_event(md)) != NULL) {
3455                         ++trace->nr_events;
3456
3457                         err = trace__deliver_event(trace, event);
3458                         if (err)
3459                                 goto out_disable;
3460
3461                         perf_mmap__consume(md);
3462
3463                         if (interrupted)
3464                                 goto out_disable;
3465
3466                         if (done && !draining) {
3467                                 evlist__disable(evlist);
3468                                 draining = true;
3469                         }
3470                 }
3471                 perf_mmap__read_done(md);
3472         }
3473
3474         if (trace->nr_events == before) {
3475                 int timeout = done ? 100 : -1;
3476
3477                 if (!draining && perf_evlist__poll(evlist, timeout) > 0) {
3478                         if (perf_evlist__filter_pollfd(evlist, POLLERR | POLLHUP | POLLNVAL) == 0)
3479                                 draining = true;
3480
3481                         goto again;
3482                 } else {
3483                         if (trace__flush_events(trace))
3484                                 goto out_disable;
3485                 }
3486         } else {
3487                 goto again;
3488         }
3489
3490 out_disable:
3491         thread__zput(trace->current);
3492
3493         evlist__disable(evlist);
3494
3495         if (trace->sort_events)
3496                 ordered_events__flush(&trace->oe.data, OE_FLUSH__FINAL);
3497
3498         if (!err) {
3499                 if (trace->summary)
3500                         trace__fprintf_thread_summary(trace, trace->output);
3501
3502                 if (trace->show_tool_stats) {
3503                         fprintf(trace->output, "Stats:\n "
3504                                                " vfs_getname : %" PRIu64 "\n"
3505                                                " proc_getname: %" PRIu64 "\n",
3506                                 trace->stats.vfs_getname,
3507                                 trace->stats.proc_getname);
3508                 }
3509         }
3510
3511 out_delete_evlist:
3512         trace__symbols__exit(trace);
3513
3514         evlist__delete(evlist);
3515         cgroup__put(trace->cgroup);
3516         trace->evlist = NULL;
3517         trace->live = false;
3518         return err;
3519 {
3520         char errbuf[BUFSIZ];
3521
3522 out_error_sched_stat_runtime:
3523         tracing_path__strerror_open_tp(errno, errbuf, sizeof(errbuf), "sched", "sched_stat_runtime");
3524         goto out_error;
3525
3526 out_error_raw_syscalls:
3527         tracing_path__strerror_open_tp(errno, errbuf, sizeof(errbuf), "raw_syscalls", "sys_(enter|exit)");
3528         goto out_error;
3529
3530 out_error_mmap:
3531         perf_evlist__strerror_mmap(evlist, errno, errbuf, sizeof(errbuf));
3532         goto out_error;
3533
3534 out_error_open:
3535         perf_evlist__strerror_open(evlist, errno, errbuf, sizeof(errbuf));
3536
3537 out_error:
3538         fprintf(trace->output, "%s\n", errbuf);
3539         goto out_delete_evlist;
3540
3541 out_error_apply_filters:
3542         fprintf(trace->output,
3543                 "Failed to set filter \"%s\" on event %s with %d (%s)\n",
3544                 evsel->filter, perf_evsel__name(evsel), errno,
3545                 str_error_r(errno, errbuf, sizeof(errbuf)));
3546         goto out_delete_evlist;
3547 }
3548 out_error_mem:
3549         fprintf(trace->output, "Not enough memory to run!\n");
3550         goto out_delete_evlist;
3551
3552 out_errno:
3553         fprintf(trace->output, "errno=%d,%s\n", errno, strerror(errno));
3554         goto out_delete_evlist;
3555 }
3556
3557 static int trace__replay(struct trace *trace)
3558 {
3559         const struct evsel_str_handler handlers[] = {
3560                 { "probe:vfs_getname",       trace__vfs_getname, },
3561         };
3562         struct perf_data data = {
3563                 .path  = input_name,
3564                 .mode  = PERF_DATA_MODE_READ,
3565                 .force = trace->force,
3566         };
3567         struct perf_session *session;
3568         struct evsel *evsel;
3569         int err = -1;
3570
3571         trace->tool.sample        = trace__process_sample;
3572         trace->tool.mmap          = perf_event__process_mmap;
3573         trace->tool.mmap2         = perf_event__process_mmap2;
3574         trace->tool.comm          = perf_event__process_comm;
3575         trace->tool.exit          = perf_event__process_exit;
3576         trace->tool.fork          = perf_event__process_fork;
3577         trace->tool.attr          = perf_event__process_attr;
3578         trace->tool.tracing_data  = perf_event__process_tracing_data;
3579         trace->tool.build_id      = perf_event__process_build_id;
3580         trace->tool.namespaces    = perf_event__process_namespaces;
3581
3582         trace->tool.ordered_events = true;
3583         trace->tool.ordering_requires_timestamps = true;
3584
3585         /* add tid to output */
3586         trace->multiple_threads = true;
3587
3588         session = perf_session__new(&data, false, &trace->tool);
3589         if (IS_ERR(session))
3590                 return PTR_ERR(session);
3591
3592         if (trace->opts.target.pid)
3593                 symbol_conf.pid_list_str = strdup(trace->opts.target.pid);
3594
3595         if (trace->opts.target.tid)
3596                 symbol_conf.tid_list_str = strdup(trace->opts.target.tid);
3597
3598         if (symbol__init(&session->header.env) < 0)
3599                 goto out;
3600
3601         trace->host = &session->machines.host;
3602
3603         err = perf_session__set_tracepoints_handlers(session, handlers);
3604         if (err)
3605                 goto out;
3606
3607         evsel = perf_evlist__find_tracepoint_by_name(session->evlist,
3608                                                      "raw_syscalls:sys_enter");
3609         /* older kernels have syscalls tp versus raw_syscalls */
3610         if (evsel == NULL)
3611                 evsel = perf_evlist__find_tracepoint_by_name(session->evlist,
3612                                                              "syscalls:sys_enter");
3613
3614         if (evsel &&
3615             (perf_evsel__init_raw_syscall_tp(evsel, trace__sys_enter) < 0 ||
3616             perf_evsel__init_sc_tp_ptr_field(evsel, args))) {
3617                 pr_err("Error during initialize raw_syscalls:sys_enter event\n");
3618                 goto out;
3619         }
3620
3621         evsel = perf_evlist__find_tracepoint_by_name(session->evlist,
3622                                                      "raw_syscalls:sys_exit");
3623         if (evsel == NULL)
3624                 evsel = perf_evlist__find_tracepoint_by_name(session->evlist,
3625                                                              "syscalls:sys_exit");
3626         if (evsel &&
3627             (perf_evsel__init_raw_syscall_tp(evsel, trace__sys_exit) < 0 ||
3628             perf_evsel__init_sc_tp_uint_field(evsel, ret))) {
3629                 pr_err("Error during initialize raw_syscalls:sys_exit event\n");
3630                 goto out;
3631         }
3632
3633         evlist__for_each_entry(session->evlist, evsel) {
3634                 if (evsel->core.attr.type == PERF_TYPE_SOFTWARE &&
3635                     (evsel->core.attr.config == PERF_COUNT_SW_PAGE_FAULTS_MAJ ||
3636                      evsel->core.attr.config == PERF_COUNT_SW_PAGE_FAULTS_MIN ||
3637                      evsel->core.attr.config == PERF_COUNT_SW_PAGE_FAULTS))
3638                         evsel->handler = trace__pgfault;
3639         }
3640
3641         setup_pager();
3642
3643         err = perf_session__process_events(session);
3644         if (err)
3645                 pr_err("Failed to process events, error %d", err);
3646
3647         else if (trace->summary)
3648                 trace__fprintf_thread_summary(trace, trace->output);
3649
3650 out:
3651         perf_session__delete(session);
3652
3653         return err;
3654 }
3655
3656 static size_t trace__fprintf_threads_header(FILE *fp)
3657 {
3658         size_t printed;
3659
3660         printed  = fprintf(fp, "\n Summary of events:\n\n");
3661
3662         return printed;
3663 }
3664
3665 DEFINE_RESORT_RB(syscall_stats, a->msecs > b->msecs,
3666         struct stats    *stats;
3667         double          msecs;
3668         int             syscall;
3669 )
3670 {
3671         struct int_node *source = rb_entry(nd, struct int_node, rb_node);
3672         struct stats *stats = source->priv;
3673
3674         entry->syscall = source->i;
3675         entry->stats   = stats;
3676         entry->msecs   = stats ? (u64)stats->n * (avg_stats(stats) / NSEC_PER_MSEC) : 0;
3677 }
3678
3679 static size_t thread__dump_stats(struct thread_trace *ttrace,
3680                                  struct trace *trace, FILE *fp)
3681 {
3682         size_t printed = 0;
3683         struct syscall *sc;
3684         struct rb_node *nd;
3685         DECLARE_RESORT_RB_INTLIST(syscall_stats, ttrace->syscall_stats);
3686
3687         if (syscall_stats == NULL)
3688                 return 0;
3689
3690         printed += fprintf(fp, "\n");
3691
3692         printed += fprintf(fp, "   syscall            calls    total       min       avg       max      stddev\n");
3693         printed += fprintf(fp, "                               (msec)    (msec)    (msec)    (msec)        (%%)\n");
3694         printed += fprintf(fp, "   --------------- -------- --------- --------- --------- ---------     ------\n");
3695
3696         resort_rb__for_each_entry(nd, syscall_stats) {
3697                 struct stats *stats = syscall_stats_entry->stats;
3698                 if (stats) {
3699                         double min = (double)(stats->min) / NSEC_PER_MSEC;
3700                         double max = (double)(stats->max) / NSEC_PER_MSEC;
3701                         double avg = avg_stats(stats);
3702                         double pct;
3703                         u64 n = (u64) stats->n;
3704
3705                         pct = avg ? 100.0 * stddev_stats(stats)/avg : 0.0;
3706                         avg /= NSEC_PER_MSEC;
3707
3708                         sc = &trace->syscalls.table[syscall_stats_entry->syscall];
3709                         printed += fprintf(fp, "   %-15s", sc->name);
3710                         printed += fprintf(fp, " %8" PRIu64 " %9.3f %9.3f %9.3f",
3711                                            n, syscall_stats_entry->msecs, min, avg);
3712                         printed += fprintf(fp, " %9.3f %9.2f%%\n", max, pct);
3713                 }
3714         }
3715
3716         resort_rb__delete(syscall_stats);
3717         printed += fprintf(fp, "\n\n");
3718
3719         return printed;
3720 }
3721
3722 static size_t trace__fprintf_thread(FILE *fp, struct thread *thread, struct trace *trace)
3723 {
3724         size_t printed = 0;
3725         struct thread_trace *ttrace = thread__priv(thread);
3726         double ratio;
3727
3728         if (ttrace == NULL)
3729                 return 0;
3730
3731         ratio = (double)ttrace->nr_events / trace->nr_events * 100.0;
3732
3733         printed += fprintf(fp, " %s (%d), ", thread__comm_str(thread), thread->tid);
3734         printed += fprintf(fp, "%lu events, ", ttrace->nr_events);
3735         printed += fprintf(fp, "%.1f%%", ratio);
3736         if (ttrace->pfmaj)
3737                 printed += fprintf(fp, ", %lu majfaults", ttrace->pfmaj);
3738         if (ttrace->pfmin)
3739                 printed += fprintf(fp, ", %lu minfaults", ttrace->pfmin);
3740         if (trace->sched)
3741                 printed += fprintf(fp, ", %.3f msec\n", ttrace->runtime_ms);
3742         else if (fputc('\n', fp) != EOF)
3743                 ++printed;
3744
3745         printed += thread__dump_stats(ttrace, trace, fp);
3746
3747         return printed;
3748 }
3749
3750 static unsigned long thread__nr_events(struct thread_trace *ttrace)
3751 {
3752         return ttrace ? ttrace->nr_events : 0;
3753 }
3754
3755 DEFINE_RESORT_RB(threads, (thread__nr_events(a->thread->priv) < thread__nr_events(b->thread->priv)),
3756         struct thread *thread;
3757 )
3758 {
3759         entry->thread = rb_entry(nd, struct thread, rb_node);
3760 }
3761
3762 static size_t trace__fprintf_thread_summary(struct trace *trace, FILE *fp)
3763 {
3764         size_t printed = trace__fprintf_threads_header(fp);
3765         struct rb_node *nd;
3766         int i;
3767
3768         for (i = 0; i < THREADS__TABLE_SIZE; i++) {
3769                 DECLARE_RESORT_RB_MACHINE_THREADS(threads, trace->host, i);
3770
3771                 if (threads == NULL) {
3772                         fprintf(fp, "%s", "Error sorting output by nr_events!\n");
3773                         return 0;
3774                 }
3775
3776                 resort_rb__for_each_entry(nd, threads)
3777                         printed += trace__fprintf_thread(fp, threads_entry->thread, trace);
3778
3779                 resort_rb__delete(threads);
3780         }
3781         return printed;
3782 }
3783
3784 static int trace__set_duration(const struct option *opt, const char *str,
3785                                int unset __maybe_unused)
3786 {
3787         struct trace *trace = opt->value;
3788
3789         trace->duration_filter = atof(str);
3790         return 0;
3791 }
3792
3793 static int trace__set_filter_pids_from_option(const struct option *opt, const char *str,
3794                                               int unset __maybe_unused)
3795 {
3796         int ret = -1;
3797         size_t i;
3798         struct trace *trace = opt->value;
3799         /*
3800          * FIXME: introduce a intarray class, plain parse csv and create a
3801          * { int nr, int entries[] } struct...
3802          */
3803         struct intlist *list = intlist__new(str);
3804
3805         if (list == NULL)
3806                 return -1;
3807
3808         i = trace->filter_pids.nr = intlist__nr_entries(list) + 1;
3809         trace->filter_pids.entries = calloc(i, sizeof(pid_t));
3810
3811         if (trace->filter_pids.entries == NULL)
3812                 goto out;
3813
3814         trace->filter_pids.entries[0] = getpid();
3815
3816         for (i = 1; i < trace->filter_pids.nr; ++i)
3817                 trace->filter_pids.entries[i] = intlist__entry(list, i - 1)->i;
3818
3819         intlist__delete(list);
3820         ret = 0;
3821 out:
3822         return ret;
3823 }
3824
3825 static int trace__open_output(struct trace *trace, const char *filename)
3826 {
3827         struct stat st;
3828
3829         if (!stat(filename, &st) && st.st_size) {
3830                 char oldname[PATH_MAX];
3831
3832                 scnprintf(oldname, sizeof(oldname), "%s.old", filename);
3833                 unlink(oldname);
3834                 rename(filename, oldname);
3835         }
3836
3837         trace->output = fopen(filename, "w");
3838
3839         return trace->output == NULL ? -errno : 0;
3840 }
3841
3842 static int parse_pagefaults(const struct option *opt, const char *str,
3843                             int unset __maybe_unused)
3844 {
3845         int *trace_pgfaults = opt->value;
3846
3847         if (strcmp(str, "all") == 0)
3848                 *trace_pgfaults |= TRACE_PFMAJ | TRACE_PFMIN;
3849         else if (strcmp(str, "maj") == 0)
3850                 *trace_pgfaults |= TRACE_PFMAJ;
3851         else if (strcmp(str, "min") == 0)
3852                 *trace_pgfaults |= TRACE_PFMIN;
3853         else
3854                 return -1;
3855
3856         return 0;
3857 }
3858
3859 static void evlist__set_evsel_handler(struct evlist *evlist, void *handler)
3860 {
3861         struct evsel *evsel;
3862
3863         evlist__for_each_entry(evlist, evsel)
3864                 evsel->handler = handler;
3865 }
3866
3867 static int evlist__set_syscall_tp_fields(struct evlist *evlist)
3868 {
3869         struct evsel *evsel;
3870
3871         evlist__for_each_entry(evlist, evsel) {
3872                 if (evsel->priv || !evsel->tp_format)
3873                         continue;
3874
3875                 if (strcmp(evsel->tp_format->system, "syscalls"))
3876                         continue;
3877
3878                 if (perf_evsel__init_syscall_tp(evsel))
3879                         return -1;
3880
3881                 if (!strncmp(evsel->tp_format->name, "sys_enter_", 10)) {
3882                         struct syscall_tp *sc = evsel->priv;
3883
3884                         if (__tp_field__init_ptr(&sc->args, sc->id.offset + sizeof(u64)))
3885                                 return -1;
3886                 } else if (!strncmp(evsel->tp_format->name, "sys_exit_", 9)) {
3887                         struct syscall_tp *sc = evsel->priv;
3888
3889                         if (__tp_field__init_uint(&sc->ret, sizeof(u64), sc->id.offset + sizeof(u64), evsel->needs_swap))
3890                                 return -1;
3891                 }
3892         }
3893
3894         return 0;
3895 }
3896
3897 /*
3898  * XXX: Hackish, just splitting the combined -e+--event (syscalls
3899  * (raw_syscalls:{sys_{enter,exit}} + events (tracepoints, HW, SW, etc) to use
3900  * existing facilities unchanged (trace->ev_qualifier + parse_options()).
3901  *
3902  * It'd be better to introduce a parse_options() variant that would return a
3903  * list with the terms it didn't match to an event...
3904  */
3905 static int trace__parse_events_option(const struct option *opt, const char *str,
3906                                       int unset __maybe_unused)
3907 {
3908         struct trace *trace = (struct trace *)opt->value;
3909         const char *s = str;
3910         char *sep = NULL, *lists[2] = { NULL, NULL, };
3911         int len = strlen(str) + 1, err = -1, list, idx;
3912         char *strace_groups_dir = system_path(STRACE_GROUPS_DIR);
3913         char group_name[PATH_MAX];
3914         struct syscall_fmt *fmt;
3915
3916         if (strace_groups_dir == NULL)
3917                 return -1;
3918
3919         if (*s == '!') {
3920                 ++s;
3921                 trace->not_ev_qualifier = true;
3922         }
3923
3924         while (1) {
3925                 if ((sep = strchr(s, ',')) != NULL)
3926                         *sep = '\0';
3927
3928                 list = 0;
3929                 if (syscalltbl__id(trace->sctbl, s) >= 0 ||
3930                     syscalltbl__strglobmatch_first(trace->sctbl, s, &idx) >= 0) {
3931                         list = 1;
3932                         goto do_concat;
3933                 }
3934
3935                 fmt = syscall_fmt__find_by_alias(s);
3936                 if (fmt != NULL) {
3937                         list = 1;
3938                         s = fmt->name;
3939                 } else {
3940                         path__join(group_name, sizeof(group_name), strace_groups_dir, s);
3941                         if (access(group_name, R_OK) == 0)
3942                                 list = 1;
3943                 }
3944 do_concat:
3945                 if (lists[list]) {
3946                         sprintf(lists[list] + strlen(lists[list]), ",%s", s);
3947                 } else {
3948                         lists[list] = malloc(len);
3949                         if (lists[list] == NULL)
3950                                 goto out;
3951                         strcpy(lists[list], s);
3952                 }
3953
3954                 if (!sep)
3955                         break;
3956
3957                 *sep = ',';
3958                 s = sep + 1;
3959         }
3960
3961         if (lists[1] != NULL) {
3962                 struct strlist_config slist_config = {
3963                         .dirname = strace_groups_dir,
3964                 };
3965
3966                 trace->ev_qualifier = strlist__new(lists[1], &slist_config);
3967                 if (trace->ev_qualifier == NULL) {
3968                         fputs("Not enough memory to parse event qualifier", trace->output);
3969                         goto out;
3970                 }
3971
3972                 if (trace__validate_ev_qualifier(trace))
3973                         goto out;
3974                 trace->trace_syscalls = true;
3975         }
3976
3977         err = 0;
3978
3979         if (lists[0]) {
3980                 struct option o = OPT_CALLBACK('e', "event", &trace->evlist, "event",
3981                                                "event selector. use 'perf list' to list available events",
3982                                                parse_events_option);
3983                 err = parse_events_option(&o, lists[0], 0);
3984         }
3985 out:
3986         if (sep)
3987                 *sep = ',';
3988
3989         return err;
3990 }
3991
3992 static int trace__parse_cgroups(const struct option *opt, const char *str, int unset)
3993 {
3994         struct trace *trace = opt->value;
3995
3996         if (!list_empty(&trace->evlist->core.entries))
3997                 return parse_cgroups(opt, str, unset);
3998
3999         trace->cgroup = evlist__findnew_cgroup(trace->evlist, str);
4000
4001         return 0;
4002 }
4003
4004 static struct bpf_map *trace__find_bpf_map_by_name(struct trace *trace, const char *name)
4005 {
4006         if (trace->bpf_obj == NULL)
4007                 return NULL;
4008
4009         return bpf_object__find_map_by_name(trace->bpf_obj, name);
4010 }
4011
4012 static void trace__set_bpf_map_filtered_pids(struct trace *trace)
4013 {
4014         trace->filter_pids.map = trace__find_bpf_map_by_name(trace, "pids_filtered");
4015 }
4016
4017 static void trace__set_bpf_map_syscalls(struct trace *trace)
4018 {
4019         trace->syscalls.map = trace__find_bpf_map_by_name(trace, "syscalls");
4020         trace->syscalls.prog_array.sys_enter = trace__find_bpf_map_by_name(trace, "syscalls_sys_enter");
4021         trace->syscalls.prog_array.sys_exit  = trace__find_bpf_map_by_name(trace, "syscalls_sys_exit");
4022 }
4023
4024 static int trace__config(const char *var, const char *value, void *arg)
4025 {
4026         struct trace *trace = arg;
4027         int err = 0;
4028
4029         if (!strcmp(var, "trace.add_events")) {
4030                 struct option o = OPT_CALLBACK('e', "event", &trace->evlist, "event",
4031                                                "event selector. use 'perf list' to list available events",
4032                                                parse_events_option);
4033                 /*
4034                  * We can't propagate parse_event_option() return, as it is 1
4035                  * for failure while perf_config() expects -1.
4036                  */
4037                 if (parse_events_option(&o, value, 0))
4038                         err = -1;
4039         } else if (!strcmp(var, "trace.show_timestamp")) {
4040                 trace->show_tstamp = perf_config_bool(var, value);
4041         } else if (!strcmp(var, "trace.show_duration")) {
4042                 trace->show_duration = perf_config_bool(var, value);
4043         } else if (!strcmp(var, "trace.show_arg_names")) {
4044                 trace->show_arg_names = perf_config_bool(var, value);
4045                 if (!trace->show_arg_names)
4046                         trace->show_zeros = true;
4047         } else if (!strcmp(var, "trace.show_zeros")) {
4048                 bool new_show_zeros = perf_config_bool(var, value);
4049                 if (!trace->show_arg_names && !new_show_zeros) {
4050                         pr_warning("trace.show_zeros has to be set when trace.show_arg_names=no\n");
4051                         goto out;
4052                 }
4053                 trace->show_zeros = new_show_zeros;
4054         } else if (!strcmp(var, "trace.show_prefix")) {
4055                 trace->show_string_prefix = perf_config_bool(var, value);
4056         } else if (!strcmp(var, "trace.no_inherit")) {
4057                 trace->opts.no_inherit = perf_config_bool(var, value);
4058         } else if (!strcmp(var, "trace.args_alignment")) {
4059                 int args_alignment = 0;
4060                 if (perf_config_int(&args_alignment, var, value) == 0)
4061                         trace->args_alignment = args_alignment;
4062         }
4063 out:
4064         return err;
4065 }
4066
4067 int cmd_trace(int argc, const char **argv)
4068 {
4069         const char *trace_usage[] = {
4070                 "perf trace [<options>] [<command>]",
4071                 "perf trace [<options>] -- <command> [<options>]",
4072                 "perf trace record [<options>] [<command>]",
4073                 "perf trace record [<options>] -- <command> [<options>]",
4074                 NULL
4075         };
4076         struct trace trace = {
4077                 .opts = {
4078                         .target = {
4079                                 .uid       = UINT_MAX,
4080                                 .uses_mmap = true,
4081                         },
4082                         .user_freq     = UINT_MAX,
4083                         .user_interval = ULLONG_MAX,
4084                         .no_buffering  = true,
4085                         .mmap_pages    = UINT_MAX,
4086                 },
4087                 .output = stderr,
4088                 .show_comm = true,
4089                 .show_tstamp = true,
4090                 .show_duration = true,
4091                 .show_arg_names = true,
4092                 .args_alignment = 70,
4093                 .trace_syscalls = false,
4094                 .kernel_syscallchains = false,
4095                 .max_stack = UINT_MAX,
4096                 .max_events = ULONG_MAX,
4097         };
4098         const char *map_dump_str = NULL;
4099         const char *output_name = NULL;
4100         const struct option trace_options[] = {
4101         OPT_CALLBACK('e', "event", &trace, "event",
4102                      "event/syscall selector. use 'perf list' to list available events",
4103                      trace__parse_events_option),
4104         OPT_BOOLEAN(0, "comm", &trace.show_comm,
4105                     "show the thread COMM next to its id"),
4106         OPT_BOOLEAN(0, "tool_stats", &trace.show_tool_stats, "show tool stats"),
4107         OPT_CALLBACK(0, "expr", &trace, "expr", "list of syscalls/events to trace",
4108                      trace__parse_events_option),
4109         OPT_STRING('o', "output", &output_name, "file", "output file name"),
4110         OPT_STRING('i', "input", &input_name, "file", "Analyze events in file"),
4111         OPT_STRING('p', "pid", &trace.opts.target.pid, "pid",
4112                     "trace events on existing process id"),
4113         OPT_STRING('t', "tid", &trace.opts.target.tid, "tid",
4114                     "trace events on existing thread id"),
4115         OPT_CALLBACK(0, "filter-pids", &trace, "CSV list of pids",
4116                      "pids to filter (by the kernel)", trace__set_filter_pids_from_option),
4117         OPT_BOOLEAN('a', "all-cpus", &trace.opts.target.system_wide,
4118                     "system-wide collection from all CPUs"),
4119         OPT_STRING('C', "cpu", &trace.opts.target.cpu_list, "cpu",
4120                     "list of cpus to monitor"),
4121         OPT_BOOLEAN(0, "no-inherit", &trace.opts.no_inherit,
4122                     "child tasks do not inherit counters"),
4123         OPT_CALLBACK('m', "mmap-pages", &trace.opts.mmap_pages, "pages",
4124                      "number of mmap data pages",
4125                      perf_evlist__parse_mmap_pages),
4126         OPT_STRING('u', "uid", &trace.opts.target.uid_str, "user",
4127                    "user to profile"),
4128         OPT_CALLBACK(0, "duration", &trace, "float",
4129                      "show only events with duration > N.M ms",
4130                      trace__set_duration),
4131 #ifdef HAVE_LIBBPF_SUPPORT
4132         OPT_STRING(0, "map-dump", &map_dump_str, "BPF map", "BPF map to periodically dump"),
4133 #endif
4134         OPT_BOOLEAN(0, "sched", &trace.sched, "show blocking scheduler events"),
4135         OPT_INCR('v', "verbose", &verbose, "be more verbose"),
4136         OPT_BOOLEAN('T', "time", &trace.full_time,
4137                     "Show full timestamp, not time relative to first start"),
4138         OPT_BOOLEAN(0, "failure", &trace.failure_only,
4139                     "Show only syscalls that failed"),
4140         OPT_BOOLEAN('s', "summary", &trace.summary_only,
4141                     "Show only syscall summary with statistics"),
4142         OPT_BOOLEAN('S', "with-summary", &trace.summary,
4143                     "Show all syscalls and summary with statistics"),
4144         OPT_CALLBACK_DEFAULT('F', "pf", &trace.trace_pgfaults, "all|maj|min",
4145                      "Trace pagefaults", parse_pagefaults, "maj"),
4146         OPT_BOOLEAN(0, "syscalls", &trace.trace_syscalls, "Trace syscalls"),
4147         OPT_BOOLEAN('f', "force", &trace.force, "don't complain, do it"),
4148         OPT_CALLBACK(0, "call-graph", &trace.opts,
4149                      "record_mode[,record_size]", record_callchain_help,
4150                      &record_parse_callchain_opt),
4151         OPT_BOOLEAN(0, "kernel-syscall-graph", &trace.kernel_syscallchains,
4152                     "Show the kernel callchains on the syscall exit path"),
4153         OPT_ULONG(0, "max-events", &trace.max_events,
4154                 "Set the maximum number of events to print, exit after that is reached. "),
4155         OPT_UINTEGER(0, "min-stack", &trace.min_stack,
4156                      "Set the minimum stack depth when parsing the callchain, "
4157                      "anything below the specified depth will be ignored."),
4158         OPT_UINTEGER(0, "max-stack", &trace.max_stack,
4159                      "Set the maximum stack depth when parsing the callchain, "
4160                      "anything beyond the specified depth will be ignored. "
4161                      "Default: kernel.perf_event_max_stack or " __stringify(PERF_MAX_STACK_DEPTH)),
4162         OPT_BOOLEAN(0, "sort-events", &trace.sort_events,
4163                         "Sort batch of events before processing, use if getting out of order events"),
4164         OPT_BOOLEAN(0, "print-sample", &trace.print_sample,
4165                         "print the PERF_RECORD_SAMPLE PERF_SAMPLE_ info, for debugging"),
4166         OPT_UINTEGER(0, "proc-map-timeout", &proc_map_timeout,
4167                         "per thread proc mmap processing timeout in ms"),
4168         OPT_CALLBACK('G', "cgroup", &trace, "name", "monitor event in cgroup name only",
4169                      trace__parse_cgroups),
4170         OPT_UINTEGER('D', "delay", &trace.opts.initial_delay,
4171                      "ms to wait before starting measurement after program "
4172                      "start"),
4173         OPTS_EVSWITCH(&trace.evswitch),
4174         OPT_END()
4175         };
4176         bool __maybe_unused max_stack_user_set = true;
4177         bool mmap_pages_user_set = true;
4178         struct evsel *evsel;
4179         const char * const trace_subcommands[] = { "record", NULL };
4180         int err = -1;
4181         char bf[BUFSIZ];
4182
4183         signal(SIGSEGV, sighandler_dump_stack);
4184         signal(SIGFPE, sighandler_dump_stack);
4185
4186         trace.evlist = evlist__new();
4187         trace.sctbl = syscalltbl__new();
4188
4189         if (trace.evlist == NULL || trace.sctbl == NULL) {
4190                 pr_err("Not enough memory to run!\n");
4191                 err = -ENOMEM;
4192                 goto out;
4193         }
4194
4195         /*
4196          * Parsing .perfconfig may entail creating a BPF event, that may need
4197          * to create BPF maps, so bump RLIM_MEMLOCK as the default 64K setting
4198          * is too small. This affects just this process, not touching the
4199          * global setting. If it fails we'll get something in 'perf trace -v'
4200          * to help diagnose the problem.
4201          */
4202         rlimit__bump_memlock();
4203
4204         err = perf_config(trace__config, &trace);
4205         if (err)
4206                 goto out;
4207
4208         argc = parse_options_subcommand(argc, argv, trace_options, trace_subcommands,
4209                                  trace_usage, PARSE_OPT_STOP_AT_NON_OPTION);
4210
4211         if ((nr_cgroups || trace.cgroup) && !trace.opts.target.system_wide) {
4212                 usage_with_options_msg(trace_usage, trace_options,
4213                                        "cgroup monitoring only available in system-wide mode");
4214         }
4215
4216         evsel = bpf__setup_output_event(trace.evlist, "__augmented_syscalls__");
4217         if (IS_ERR(evsel)) {
4218                 bpf__strerror_setup_output_event(trace.evlist, PTR_ERR(evsel), bf, sizeof(bf));
4219                 pr_err("ERROR: Setup trace syscalls enter failed: %s\n", bf);
4220                 goto out;
4221         }
4222
4223         if (evsel) {
4224                 trace.syscalls.events.augmented = evsel;
4225
4226                 evsel = perf_evlist__find_tracepoint_by_name(trace.evlist, "raw_syscalls:sys_enter");
4227                 if (evsel == NULL) {
4228                         pr_err("ERROR: raw_syscalls:sys_enter not found in the augmented BPF object\n");
4229                         goto out;
4230                 }
4231
4232                 if (evsel->bpf_obj == NULL) {
4233                         pr_err("ERROR: raw_syscalls:sys_enter not associated to a BPF object\n");
4234                         goto out;
4235                 }
4236
4237                 trace.bpf_obj = evsel->bpf_obj;
4238
4239                 trace__set_bpf_map_filtered_pids(&trace);
4240                 trace__set_bpf_map_syscalls(&trace);
4241                 trace.syscalls.unaugmented_prog = trace__find_bpf_program_by_title(&trace, "!raw_syscalls:unaugmented");
4242         }
4243
4244         err = bpf__setup_stdout(trace.evlist);
4245         if (err) {
4246                 bpf__strerror_setup_stdout(trace.evlist, err, bf, sizeof(bf));
4247                 pr_err("ERROR: Setup BPF stdout failed: %s\n", bf);
4248                 goto out;
4249         }
4250
4251         err = -1;
4252
4253         if (map_dump_str) {
4254                 trace.dump.map = trace__find_bpf_map_by_name(&trace, map_dump_str);
4255                 if (trace.dump.map == NULL) {
4256                         pr_err("ERROR: BPF map \"%s\" not found\n", map_dump_str);
4257                         goto out;
4258                 }
4259         }
4260
4261         if (trace.trace_pgfaults) {
4262                 trace.opts.sample_address = true;
4263                 trace.opts.sample_time = true;
4264         }
4265
4266         if (trace.opts.mmap_pages == UINT_MAX)
4267                 mmap_pages_user_set = false;
4268
4269         if (trace.max_stack == UINT_MAX) {
4270                 trace.max_stack = input_name ? PERF_MAX_STACK_DEPTH : sysctl__max_stack();
4271                 max_stack_user_set = false;
4272         }
4273
4274 #ifdef HAVE_DWARF_UNWIND_SUPPORT
4275         if ((trace.min_stack || max_stack_user_set) && !callchain_param.enabled) {
4276                 record_opts__parse_callchain(&trace.opts, &callchain_param, "dwarf", false);
4277         }
4278 #endif
4279
4280         if (callchain_param.enabled) {
4281                 if (!mmap_pages_user_set && geteuid() == 0)
4282                         trace.opts.mmap_pages = perf_event_mlock_kb_in_pages() * 4;
4283
4284                 symbol_conf.use_callchain = true;
4285         }
4286
4287         if (trace.evlist->core.nr_entries > 0) {
4288                 evlist__set_evsel_handler(trace.evlist, trace__event_handler);
4289                 if (evlist__set_syscall_tp_fields(trace.evlist)) {
4290                         perror("failed to set syscalls:* tracepoint fields");
4291                         goto out;
4292                 }
4293         }
4294
4295         if (trace.sort_events) {
4296                 ordered_events__init(&trace.oe.data, ordered_events__deliver_event, &trace);
4297                 ordered_events__set_copy_on_queue(&trace.oe.data, true);
4298         }
4299
4300         /*
4301          * If we are augmenting syscalls, then combine what we put in the
4302          * __augmented_syscalls__ BPF map with what is in the
4303          * syscalls:sys_exit_FOO tracepoints, i.e. just like we do without BPF,
4304          * combining raw_syscalls:sys_enter with raw_syscalls:sys_exit.
4305          *
4306          * We'll switch to look at two BPF maps, one for sys_enter and the
4307          * other for sys_exit when we start augmenting the sys_exit paths with
4308          * buffers that are being copied from kernel to userspace, think 'read'
4309          * syscall.
4310          */
4311         if (trace.syscalls.events.augmented) {
4312                 evlist__for_each_entry(trace.evlist, evsel) {
4313                         bool raw_syscalls_sys_exit = strcmp(perf_evsel__name(evsel), "raw_syscalls:sys_exit") == 0;
4314
4315                         if (raw_syscalls_sys_exit) {
4316                                 trace.raw_augmented_syscalls = true;
4317                                 goto init_augmented_syscall_tp;
4318                         }
4319
4320                         if (trace.syscalls.events.augmented->priv == NULL &&
4321                             strstr(perf_evsel__name(evsel), "syscalls:sys_enter")) {
4322                                 struct evsel *augmented = trace.syscalls.events.augmented;
4323                                 if (perf_evsel__init_augmented_syscall_tp(augmented, evsel) ||
4324                                     perf_evsel__init_augmented_syscall_tp_args(augmented))
4325                                         goto out;
4326                                 /*
4327                                  * Augmented is __augmented_syscalls__ BPF_OUTPUT event
4328                                  * Above we made sure we can get from the payload the tp fields
4329                                  * that we get from syscalls:sys_enter tracefs format file.
4330                                  */
4331                                 augmented->handler = trace__sys_enter;
4332                                 /*
4333                                  * Now we do the same for the *syscalls:sys_enter event so that
4334                                  * if we handle it directly, i.e. if the BPF prog returns 0 so
4335                                  * as not to filter it, then we'll handle it just like we would
4336                                  * for the BPF_OUTPUT one:
4337                                  */
4338                                 if (perf_evsel__init_augmented_syscall_tp(evsel, evsel) ||
4339                                     perf_evsel__init_augmented_syscall_tp_args(evsel))
4340                                         goto out;
4341                                 evsel->handler = trace__sys_enter;
4342                         }
4343
4344                         if (strstarts(perf_evsel__name(evsel), "syscalls:sys_exit_")) {
4345                                 struct syscall_tp *sc;
4346 init_augmented_syscall_tp:
4347                                 if (perf_evsel__init_augmented_syscall_tp(evsel, evsel))
4348                                         goto out;
4349                                 sc = evsel->priv;
4350                                 /*
4351                                  * For now with BPF raw_augmented we hook into
4352                                  * raw_syscalls:sys_enter and there we get all
4353                                  * 6 syscall args plus the tracepoint common
4354                                  * fields and the syscall_nr (another long).
4355                                  * So we check if that is the case and if so
4356                                  * don't look after the sc->args_size but
4357                                  * always after the full raw_syscalls:sys_enter
4358                                  * payload, which is fixed.
4359                                  *
4360                                  * We'll revisit this later to pass
4361                                  * s->args_size to the BPF augmenter (now
4362                                  * tools/perf/examples/bpf/augmented_raw_syscalls.c,
4363                                  * so that it copies only what we need for each
4364                                  * syscall, like what happens when we use
4365                                  * syscalls:sys_enter_NAME, so that we reduce
4366                                  * the kernel/userspace traffic to just what is
4367                                  * needed for each syscall.
4368                                  */
4369                                 if (trace.raw_augmented_syscalls)
4370                                         trace.raw_augmented_syscalls_args_size = (6 + 1) * sizeof(long) + sc->id.offset;
4371                                 perf_evsel__init_augmented_syscall_tp_ret(evsel);
4372                                 evsel->handler = trace__sys_exit;
4373                         }
4374                 }
4375         }
4376
4377         if ((argc >= 1) && (strcmp(argv[0], "record") == 0))
4378                 return trace__record(&trace, argc-1, &argv[1]);
4379
4380         /* summary_only implies summary option, but don't overwrite summary if set */
4381         if (trace.summary_only)
4382                 trace.summary = trace.summary_only;
4383
4384         if (!trace.trace_syscalls && !trace.trace_pgfaults &&
4385             trace.evlist->core.nr_entries == 0 /* Was --events used? */) {
4386                 trace.trace_syscalls = true;
4387         }
4388
4389         if (output_name != NULL) {
4390                 err = trace__open_output(&trace, output_name);
4391                 if (err < 0) {
4392                         perror("failed to create output file");
4393                         goto out;
4394                 }
4395         }
4396
4397         err = evswitch__init(&trace.evswitch, trace.evlist, stderr);
4398         if (err)
4399                 goto out_close;
4400
4401         err = target__validate(&trace.opts.target);
4402         if (err) {
4403                 target__strerror(&trace.opts.target, err, bf, sizeof(bf));
4404                 fprintf(trace.output, "%s", bf);
4405                 goto out_close;
4406         }
4407
4408         err = target__parse_uid(&trace.opts.target);
4409         if (err) {
4410                 target__strerror(&trace.opts.target, err, bf, sizeof(bf));
4411                 fprintf(trace.output, "%s", bf);
4412                 goto out_close;
4413         }
4414
4415         if (!argc && target__none(&trace.opts.target))
4416                 trace.opts.target.system_wide = true;
4417
4418         if (input_name)
4419                 err = trace__replay(&trace);
4420         else
4421                 err = trace__run(&trace, argc, argv);
4422
4423 out_close:
4424         if (output_name != NULL)
4425                 fclose(trace.output);
4426 out:
4427         return err;
4428 }