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