tools/perf/builtin-trace.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * builtin-trace.c
   4  *
   5  * Builtin 'trace' command:
   6  *
   7  * Display a continuously updated trace of any workload, CPU, specific PID,
   8  * system wide, etc.  Default format is loosely strace like, but any other
   9  * event may be specified using --event.
  10  *
  11  * Copyright (C) 2012, 2013, 2014, 2015 Red Hat Inc, Arnaldo Carvalho de Melo <acme@redhat.com>
  12  *
  13  * Initially based on the 'trace' prototype by Thomas Gleixner:
  14  *
  15  * http://lwn.net/Articles/415728/ ("Announcing a new utility: 'trace'")
  16  */
  17
  18 #include <traceevent/event-parse.h>
  19 #include <api/fs/tracing_path.h>
  20 #include <bpf/bpf.h>
  21 #include "util/bpf_map.h"
  22 #include "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/env.h"
  29 #include "util/event.h"
  30 #include "util/evlist.h"
  31 #include <subcmd/exec-cmd.h>
  32 #include "util/machine.h"
  33 #include "util/map.h"
  34 #include "util/symbol.h"
  35 #include "util/path.h"
  36 #include "util/session.h"
  37 #include "util/thread.h"
  38 #include <subcmd/parse-options.h>
  39 #include "util/strlist.h"
  40 #include "util/intlist.h"
  41 #include "util/thread_map.h"
  42 #include "util/stat.h"
  43 #include "trace/beauty/beauty.h"
  44 #include "trace-event.h"
  45 #include "util/parse-events.h"
  46 #include "util/bpf-loader.h"
  47 #include "callchain.h"
  48 #include "print_binary.h"
  49 #include "string2.h"
  50 #include "syscalltbl.h"
  51 #include "rb_resort.h"
  52
  53 #include <errno.h>
  54 #include <inttypes.h>
  55 #include <poll.h>
  56 #include <signal.h>
  57 #include <stdlib.h>
  58 #include <string.h>
  59 #include <linux/err.h>
  60 #include <linux/filter.h>
  61 #include <linux/kernel.h>
  62 #include <linux/random.h>
  63 #include <linux/stringify.h>
  64 #include <linux/time64.h>
  65 #include <linux/zalloc.h>
  66 #include <fcntl.h>
  67 #include <sys/sysmacros.h>
  68
  69 #include <linux/ctype.h>
  70
  71 #ifndef O_CLOEXEC
  72 # define O_CLOEXEC              02000000
  73 #endif
  74
  75 #ifndef F_LINUX_SPECIFIC_BASE
  76 # define F_LINUX_SPECIFIC_BASE  1024
  77 #endif
  78
  79 struct trace {
  80         struct perf_tool        tool;
  81         struct syscalltbl       *sctbl;
  82         struct {
  83                 int             max;
  84                 struct syscall  *table;
  85                 struct bpf_map  *map;
  86                 struct {
  87                         struct perf_evsel *sys_enter,
  88                                           *sys_exit,
  89                                           *augmented;
  90                 }               events;
  91         } syscalls;
  92         struct {
  93                 struct bpf_map *map;
  94         } dump;
  95         struct record_opts      opts;
  96         struct perf_evlist      *evlist;
  97         struct machine          *host;
  98         struct thread           *current;
  99         struct cgroup           *cgroup;
 100         u64                     base_time;
 101         FILE                    *output;
 102         unsigned long           nr_events;
 103         unsigned long           nr_events_printed;
 104         unsigned long           max_events;
 105         struct strlist          *ev_qualifier;
 106         struct {
 107                 size_t          nr;
 108                 int             *entries;
 109         }                       ev_qualifier_ids;
 110         struct {
 111                 size_t          nr;
 112                 pid_t           *entries;
 113                 struct bpf_map  *map;
 114         }                       filter_pids;
 115         double                  duration_filter;
 116         double                  runtime_ms;
 117         struct {
 118                 u64             vfs_getname,
 119                                 proc_getname;
 120         } stats;
 121         unsigned int            max_stack;
 122         unsigned int            min_stack;
 123         int                     raw_augmented_syscalls_args_size;
 124         bool                    raw_augmented_syscalls;
 125         bool                    sort_events;
 126         bool                    not_ev_qualifier;
 127         bool                    live;
 128         bool                    full_time;
 129         bool                    sched;
 130         bool                    multiple_threads;
 131         bool                    summary;
 132         bool                    summary_only;
 133         bool                    failure_only;
 134         bool                    show_comm;
 135         bool                    print_sample;
 136         bool                    show_tool_stats;
 137         bool                    trace_syscalls;
 138         bool                    kernel_syscallchains;
 139         s16                     args_alignment;
 140         bool                    show_tstamp;
 141         bool                    show_duration;
 142         bool                    show_zeros;
 143         bool                    show_arg_names;
 144         bool                    show_string_prefix;
 145         bool                    force;
 146         bool                    vfs_getname;
 147         int                     trace_pgfaults;
 148         struct {
 149                 struct ordered_events   data;
 150                 u64                     last;
 151         } oe;
 152 };
 153
 154 struct tp_field {
 155         int offset;
 156         union {
 157                 u64 (*integer)(struct tp_field *field, struct perf_sample *sample);
 158                 void *(*pointer)(struct tp_field *field, struct perf_sample *sample);
 159         };
 160 };
 161
 162 #define TP_UINT_FIELD(bits) \
 163 static u64 tp_field__u##bits(struct tp_field *field, struct perf_sample *sample) \
 164 { \
 165         u##bits value; \
 166         memcpy(&value, sample->raw_data + field->offset, sizeof(value)); \
 167         return value;  \
 168 }
 169
 170 TP_UINT_FIELD(8);
 171 TP_UINT_FIELD(16);
 172 TP_UINT_FIELD(32);
 173 TP_UINT_FIELD(64);
 174
 175 #define TP_UINT_FIELD__SWAPPED(bits) \
 176 static u64 tp_field__swapped_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 bswap_##bits(value);\
 181 }
 182
 183 TP_UINT_FIELD__SWAPPED(16);
 184 TP_UINT_FIELD__SWAPPED(32);
 185 TP_UINT_FIELD__SWAPPED(64);
 186
 187 static int __tp_field__init_uint(struct tp_field *field, int size, int offset, bool needs_swap)
 188 {
 189         field->offset = offset;
 190
 191         switch (size) {
 192         case 1:
 193                 field->integer = tp_field__u8;
 194                 break;
 195         case 2:
 196                 field->integer = needs_swap ? tp_field__swapped_u16 : tp_field__u16;
 197                 break;
 198         case 4:
 199                 field->integer = needs_swap ? tp_field__swapped_u32 : tp_field__u32;
 200                 break;
 201         case 8:
 202                 field->integer = needs_swap ? tp_field__swapped_u64 : tp_field__u64;
 203                 break;
 204         default:
 205                 return -1;
 206         }
 207
 208         return 0;
 209 }
 210
 211 static int tp_field__init_uint(struct tp_field *field, struct tep_format_field *format_field, bool needs_swap)
 212 {
 213         return __tp_field__init_uint(field, format_field->size, format_field->offset, needs_swap);
 214 }
 215
 216 static void *tp_field__ptr(struct tp_field *field, struct perf_sample *sample)
 217 {
 218         return sample->raw_data + field->offset;
 219 }
 220
 221 static int __tp_field__init_ptr(struct tp_field *field, int offset)
 222 {
 223         field->offset = offset;
 224         field->pointer = tp_field__ptr;
 225         return 0;
 226 }
 227
 228 static int tp_field__init_ptr(struct tp_field *field, struct tep_format_field *format_field)
 229 {
 230         return __tp_field__init_ptr(field, format_field->offset);
 231 }
 232
 233 struct syscall_tp {
 234         struct tp_field id;
 235         union {
 236                 struct tp_field args, ret;
 237         };
 238 };
 239
 240 static int perf_evsel__init_tp_uint_field(struct perf_evsel *evsel,
 241                                           struct tp_field *field,
 242                                           const char *name)
 243 {
 244         struct tep_format_field *format_field = perf_evsel__field(evsel, name);
 245
 246         if (format_field == NULL)
 247                 return -1;
 248
 249         return tp_field__init_uint(field, format_field, evsel->needs_swap);
 250 }
 251
 252 #define perf_evsel__init_sc_tp_uint_field(evsel, name) \
 253         ({ struct syscall_tp *sc = evsel->priv;\
 254            perf_evsel__init_tp_uint_field(evsel, &sc->name, #name); })
 255
 256 static int perf_evsel__init_tp_ptr_field(struct perf_evsel *evsel,
 257                                          struct tp_field *field,
 258                                          const char *name)
 259 {
 260         struct tep_format_field *format_field = perf_evsel__field(evsel, name);
 261
 262         if (format_field == NULL)
 263                 return -1;
 264
 265         return tp_field__init_ptr(field, format_field);
 266 }
 267
 268 #define perf_evsel__init_sc_tp_ptr_field(evsel, name) \
 269         ({ struct syscall_tp *sc = evsel->priv;\
 270            perf_evsel__init_tp_ptr_field(evsel, &sc->name, #name); })
 271
 272 static void perf_evsel__delete_priv(struct perf_evsel *evsel)
 273 {
 274         zfree(&evsel->priv);
 275         perf_evsel__delete(evsel);
 276 }
 277
 278 static int perf_evsel__init_syscall_tp(struct perf_evsel *evsel)
 279 {
 280         struct syscall_tp *sc = evsel->priv = malloc(sizeof(struct syscall_tp));
 281
 282         if (evsel->priv != NULL) {
 283                 if (perf_evsel__init_tp_uint_field(evsel, &sc->id, "__syscall_nr") &&
 284                     perf_evsel__init_tp_uint_field(evsel, &sc->id, "nr"))
 285                         goto out_delete;
 286                 return 0;
 287         }
 288
 289         return -ENOMEM;
 290 out_delete:
 291         zfree(&evsel->priv);
 292         return -ENOENT;
 293 }
 294
 295 static int perf_evsel__init_augmented_syscall_tp(struct perf_evsel *evsel, struct perf_evsel *tp)
 296 {
 297         struct syscall_tp *sc = evsel->priv = malloc(sizeof(struct syscall_tp));
 298
 299         if (evsel->priv != NULL) {
 300                 struct tep_format_field *syscall_id = perf_evsel__field(tp, "id");
 301                 if (syscall_id == NULL)
 302                         syscall_id = perf_evsel__field(tp, "__syscall_nr");
 303                 if (syscall_id == NULL)
 304                         goto out_delete;
 305                 if (__tp_field__init_uint(&sc->id, syscall_id->size, syscall_id->offset, evsel->needs_swap))
 306                         goto out_delete;
 307
 308                 return 0;
 309         }
 310
 311         return -ENOMEM;
 312 out_delete:
 313         zfree(&evsel->priv);
 314         return -EINVAL;
 315 }
 316
 317 static int perf_evsel__init_augmented_syscall_tp_args(struct perf_evsel *evsel)
 318 {
 319         struct syscall_tp *sc = evsel->priv;
 320
 321         return __tp_field__init_ptr(&sc->args, sc->id.offset + sizeof(u64));
 322 }
 323
 324 static int perf_evsel__init_augmented_syscall_tp_ret(struct perf_evsel *evsel)
 325 {
 326         struct syscall_tp *sc = evsel->priv;
 327
 328         return __tp_field__init_uint(&sc->ret, sizeof(u64), sc->id.offset + sizeof(u64), evsel->needs_swap);
 329 }
 330
 331 static int perf_evsel__init_raw_syscall_tp(struct perf_evsel *evsel, void *handler)
 332 {
 333         evsel->priv = malloc(sizeof(struct syscall_tp));
 334         if (evsel->priv != NULL) {
 335                 if (perf_evsel__init_sc_tp_uint_field(evsel, id))
 336                         goto out_delete;
 337
 338                 evsel->handler = handler;
 339                 return 0;
 340         }
 341
 342         return -ENOMEM;
 343
 344 out_delete:
 345         zfree(&evsel->priv);
 346         return -ENOENT;
 347 }
 348
 349 static struct perf_evsel *perf_evsel__raw_syscall_newtp(const char *direction, void *handler)
 350 {
 351         struct perf_evsel *evsel = perf_evsel__newtp("raw_syscalls", direction);
 352
 353         /* older kernel (e.g., RHEL6) use syscalls:{enter,exit} */
 354         if (IS_ERR(evsel))
 355                 evsel = perf_evsel__newtp("syscalls", direction);
 356
 357         if (IS_ERR(evsel))
 358                 return NULL;
 359
 360         if (perf_evsel__init_raw_syscall_tp(evsel, handler))
 361                 goto out_delete;
 362
 363         return evsel;
 364
 365 out_delete:
 366         perf_evsel__delete_priv(evsel);
 367         return NULL;
 368 }
 369
 370 #define perf_evsel__sc_tp_uint(evsel, name, sample) \
 371         ({ struct syscall_tp *fields = evsel->priv; \
 372            fields->name.integer(&fields->name, sample); })
 373
 374 #define perf_evsel__sc_tp_ptr(evsel, name, sample) \
 375         ({ struct syscall_tp *fields = evsel->priv; \
 376            fields->name.pointer(&fields->name, sample); })
 377
 378 size_t strarray__scnprintf(struct strarray *sa, char *bf, size_t size, const char *intfmt, bool show_prefix, int val)
 379 {
 380         int idx = val - sa->offset;
 381
 382         if (idx < 0 || idx >= sa->nr_entries || sa->entries[idx] == NULL) {
 383                 size_t printed = scnprintf(bf, size, intfmt, val);
 384                 if (show_prefix)
 385                         printed += scnprintf(bf + printed, size - printed, " /* %s??? */", sa->prefix);
 386                 return printed;
 387         }
 388
 389         return scnprintf(bf, size, "%s%s", show_prefix ? sa->prefix : "", sa->entries[idx]);
 390 }
 391
 392 static size_t __syscall_arg__scnprintf_strarray(char *bf, size_t size,
 393                                                 const char *intfmt,
 394                                                 struct syscall_arg *arg)
 395 {
 396         return strarray__scnprintf(arg->parm, bf, size, intfmt, arg->show_string_prefix, arg->val);
 397 }
 398
 399 static size_t syscall_arg__scnprintf_strarray(char *bf, size_t size,
 400                                               struct syscall_arg *arg)
 401 {
 402         return __syscall_arg__scnprintf_strarray(bf, size, "%d", arg);
 403 }
 404
 405 #define SCA_STRARRAY syscall_arg__scnprintf_strarray
 406
 407 size_t syscall_arg__scnprintf_strarray_flags(char *bf, size_t size, struct syscall_arg *arg)
 408 {
 409         return strarray__scnprintf_flags(arg->parm, bf, size, arg->show_string_prefix, arg->val);
 410 }
 411
 412 size_t strarrays__scnprintf(struct strarrays *sas, char *bf, size_t size, const char *intfmt, bool show_prefix, int val)
 413 {
 414         size_t printed;
 415         int i;
 416
 417         for (i = 0; i < sas->nr_entries; ++i) {
 418                 struct strarray *sa = sas->entries[i];
 419                 int idx = val - sa->offset;
 420
 421                 if (idx >= 0 && idx < sa->nr_entries) {
 422                         if (sa->entries[idx] == NULL)
 423                                 break;
 424                         return scnprintf(bf, size, "%s%s", show_prefix ? sa->prefix : "", sa->entries[idx]);
 425                 }
 426         }
 427
 428         printed = scnprintf(bf, size, intfmt, val);
 429         if (show_prefix)
 430                 printed += scnprintf(bf + printed, size - printed, " /* %s??? */", sas->entries[0]->prefix);
 431         return printed;
 432 }
 433
 434 size_t syscall_arg__scnprintf_strarrays(char *bf, size_t size,
 435                                         struct syscall_arg *arg)
 436 {
 437         return strarrays__scnprintf(arg->parm, bf, size, "%d", arg->show_string_prefix, arg->val);
 438 }
 439
 440 #ifndef AT_FDCWD
 441 #define AT_FDCWD        -100
 442 #endif
 443
 444 static size_t syscall_arg__scnprintf_fd_at(char *bf, size_t size,
 445                                            struct syscall_arg *arg)
 446 {
 447         int fd = arg->val;
 448         const char *prefix = "AT_FD";
 449
 450         if (fd == AT_FDCWD)
 451                 return scnprintf(bf, size, "%s%s", arg->show_string_prefix ? prefix : "", "CWD");
 452
 453         return syscall_arg__scnprintf_fd(bf, size, arg);
 454 }
 455
 456 #define SCA_FDAT syscall_arg__scnprintf_fd_at
 457
 458 static size_t syscall_arg__scnprintf_close_fd(char *bf, size_t size,
 459                                               struct syscall_arg *arg);
 460
 461 #define SCA_CLOSE_FD syscall_arg__scnprintf_close_fd
 462
 463 size_t syscall_arg__scnprintf_hex(char *bf, size_t size, struct syscall_arg *arg)
 464 {
 465         return scnprintf(bf, size, "%#lx", arg->val);
 466 }
 467
 468 size_t syscall_arg__scnprintf_ptr(char *bf, size_t size, struct syscall_arg *arg)
 469 {
 470         if (arg->val == 0)
 471                 return scnprintf(bf, size, "NULL");
 472         return syscall_arg__scnprintf_hex(bf, size, arg);
 473 }
 474
 475 size_t syscall_arg__scnprintf_int(char *bf, size_t size, struct syscall_arg *arg)
 476 {
 477         return scnprintf(bf, size, "%d", arg->val);
 478 }
 479
 480 size_t syscall_arg__scnprintf_long(char *bf, size_t size, struct syscall_arg *arg)
 481 {
 482         return scnprintf(bf, size, "%ld", arg->val);
 483 }
 484
 485 static const char *bpf_cmd[] = {
 486         "MAP_CREATE", "MAP_LOOKUP_ELEM", "MAP_UPDATE_ELEM", "MAP_DELETE_ELEM",
 487         "MAP_GET_NEXT_KEY", "PROG_LOAD",
 488 };
 489 static DEFINE_STRARRAY(bpf_cmd, "BPF_");
 490
 491 static const char *fsmount_flags[] = {
 492         [1] = "CLOEXEC",
 493 };
 494 static DEFINE_STRARRAY(fsmount_flags, "FSMOUNT_");
 495
 496 #include "trace/beauty/generated/fsconfig_arrays.c"
 497
 498 static DEFINE_STRARRAY(fsconfig_cmds, "FSCONFIG_");
 499
 500 static const char *epoll_ctl_ops[] = { "ADD", "DEL", "MOD", };
 501 static DEFINE_STRARRAY_OFFSET(epoll_ctl_ops, "EPOLL_CTL_", 1);
 502
 503 static const char *itimers[] = { "REAL", "VIRTUAL", "PROF", };
 504 static DEFINE_STRARRAY(itimers, "ITIMER_");
 505
 506 static const char *keyctl_options[] = {
 507         "GET_KEYRING_ID", "JOIN_SESSION_KEYRING", "UPDATE", "REVOKE", "CHOWN",
 508         "SETPERM", "DESCRIBE", "CLEAR", "LINK", "UNLINK", "SEARCH", "READ",
 509         "INSTANTIATE", "NEGATE", "SET_REQKEY_KEYRING", "SET_TIMEOUT",
 510         "ASSUME_AUTHORITY", "GET_SECURITY", "SESSION_TO_PARENT", "REJECT",
 511         "INSTANTIATE_IOV", "INVALIDATE", "GET_PERSISTENT",
 512 };
 513 static DEFINE_STRARRAY(keyctl_options, "KEYCTL_");
 514
 515 static const char *whences[] = { "SET", "CUR", "END",
 516 #ifdef SEEK_DATA
 517 "DATA",
 518 #endif
 519 #ifdef SEEK_HOLE
 520 "HOLE",
 521 #endif
 522 };
 523 static DEFINE_STRARRAY(whences, "SEEK_");
 524
 525 static const char *fcntl_cmds[] = {
 526         "DUPFD", "GETFD", "SETFD", "GETFL", "SETFL", "GETLK", "SETLK",
 527         "SETLKW", "SETOWN", "GETOWN", "SETSIG", "GETSIG", "GETLK64",
 528         "SETLK64", "SETLKW64", "SETOWN_EX", "GETOWN_EX",
 529         "GETOWNER_UIDS",
 530 };
 531 static DEFINE_STRARRAY(fcntl_cmds, "F_");
 532
 533 static const char *fcntl_linux_specific_cmds[] = {
 534         "SETLEASE", "GETLEASE", "NOTIFY", [5] = "CANCELLK", "DUPFD_CLOEXEC",
 535         "SETPIPE_SZ", "GETPIPE_SZ", "ADD_SEALS", "GET_SEALS",
 536         "GET_RW_HINT", "SET_RW_HINT", "GET_FILE_RW_HINT", "SET_FILE_RW_HINT",
 537 };
 538
 539 static DEFINE_STRARRAY_OFFSET(fcntl_linux_specific_cmds, "F_", F_LINUX_SPECIFIC_BASE);
 540
 541 static struct strarray *fcntl_cmds_arrays[] = {
 542         &strarray__fcntl_cmds,
 543         &strarray__fcntl_linux_specific_cmds,
 544 };
 545
 546 static DEFINE_STRARRAYS(fcntl_cmds_arrays);
 547
 548 static const char *rlimit_resources[] = {
 549         "CPU", "FSIZE", "DATA", "STACK", "CORE", "RSS", "NPROC", "NOFILE",
 550         "MEMLOCK", "AS", "LOCKS", "SIGPENDING", "MSGQUEUE", "NICE", "RTPRIO",
 551         "RTTIME",
 552 };
 553 static DEFINE_STRARRAY(rlimit_resources, "RLIMIT_");
 554
 555 static const char *sighow[] = { "BLOCK", "UNBLOCK", "SETMASK", };
 556 static DEFINE_STRARRAY(sighow, "SIG_");
 557
 558 static const char *clockid[] = {
 559         "REALTIME", "MONOTONIC", "PROCESS_CPUTIME_ID", "THREAD_CPUTIME_ID",
 560         "MONOTONIC_RAW", "REALTIME_COARSE", "MONOTONIC_COARSE", "BOOTTIME",
 561         "REALTIME_ALARM", "BOOTTIME_ALARM", "SGI_CYCLE", "TAI"
 562 };
 563 static DEFINE_STRARRAY(clockid, "CLOCK_");
 564
 565 static size_t syscall_arg__scnprintf_access_mode(char *bf, size_t size,
 566                                                  struct syscall_arg *arg)
 567 {
 568         bool show_prefix = arg->show_string_prefix;
 569         const char *suffix = "_OK";
 570         size_t printed = 0;
 571         int mode = arg->val;
 572
 573         if (mode == F_OK) /* 0 */
 574                 return scnprintf(bf, size, "F%s", show_prefix ? suffix : "");
 575 #define P_MODE(n) \
 576         if (mode & n##_OK) { \
 577                 printed += scnprintf(bf + printed, size - printed, "%s%s", #n, show_prefix ? suffix : ""); \
 578                 mode &= ~n##_OK; \
 579         }
 580
 581         P_MODE(R);
 582         P_MODE(W);
 583         P_MODE(X);
 584 #undef P_MODE
 585
 586         if (mode)
 587                 printed += scnprintf(bf + printed, size - printed, "|%#x", mode);
 588
 589         return printed;
 590 }
 591
 592 #define SCA_ACCMODE syscall_arg__scnprintf_access_mode
 593
 594 static size_t syscall_arg__scnprintf_filename(char *bf, size_t size,
 595                                               struct syscall_arg *arg);
 596
 597 #define SCA_FILENAME syscall_arg__scnprintf_filename
 598
 599 static size_t syscall_arg__scnprintf_pipe_flags(char *bf, size_t size,
 600                                                 struct syscall_arg *arg)
 601 {
 602         bool show_prefix = arg->show_string_prefix;
 603         const char *prefix = "O_";
 604         int printed = 0, flags = arg->val;
 605
 606 #define P_FLAG(n) \
 607         if (flags & O_##n) { \
 608                 printed += scnprintf(bf + printed, size - printed, "%s%s%s", printed ? "|" : "", show_prefix ? prefix : "", #n); \
 609                 flags &= ~O_##n; \
 610         }
 611
 612         P_FLAG(CLOEXEC);
 613         P_FLAG(NONBLOCK);
 614 #undef P_FLAG
 615
 616         if (flags)
 617                 printed += scnprintf(bf + printed, size - printed, "%s%#x", printed ? "|" : "", flags);
 618
 619         return printed;
 620 }
 621
 622 #define SCA_PIPE_FLAGS syscall_arg__scnprintf_pipe_flags
 623
 624 #ifndef GRND_NONBLOCK
 625 #define GRND_NONBLOCK   0x0001
 626 #endif
 627 #ifndef GRND_RANDOM
 628 #define GRND_RANDOM     0x0002
 629 #endif
 630
 631 static size_t syscall_arg__scnprintf_getrandom_flags(char *bf, size_t size,
 632                                                    struct syscall_arg *arg)
 633 {
 634         bool show_prefix = arg->show_string_prefix;
 635         const char *prefix = "GRND_";
 636         int printed = 0, flags = arg->val;
 637
 638 #define P_FLAG(n) \
 639         if (flags & GRND_##n) { \
 640                 printed += scnprintf(bf + printed, size - printed, "%s%s%s", printed ? "|" : "", show_prefix ? prefix : "", #n); \
 641                 flags &= ~GRND_##n; \
 642         }
 643
 644         P_FLAG(RANDOM);
 645         P_FLAG(NONBLOCK);
 646 #undef P_FLAG
 647
 648         if (flags)
 649                 printed += scnprintf(bf + printed, size - printed, "%s%#x", printed ? "|" : "", flags);
 650
 651         return printed;
 652 }
 653
 654 #define SCA_GETRANDOM_FLAGS syscall_arg__scnprintf_getrandom_flags
 655
 656 #define STRARRAY(name, array) \
 657           { .scnprintf  = SCA_STRARRAY, \
 658             .parm       = &strarray__##array, }
 659
 660 #define STRARRAY_FLAGS(name, array) \
 661           { .scnprintf  = SCA_STRARRAY_FLAGS, \
 662             .parm       = &strarray__##array, }
 663
 664 #include "trace/beauty/arch_errno_names.c"
 665 #include "trace/beauty/eventfd.c"
 666 #include "trace/beauty/futex_op.c"
 667 #include "trace/beauty/futex_val3.c"
 668 #include "trace/beauty/mmap.c"
 669 #include "trace/beauty/mode_t.c"
 670 #include "trace/beauty/msg_flags.c"
 671 #include "trace/beauty/open_flags.c"
 672 #include "trace/beauty/perf_event_open.c"
 673 #include "trace/beauty/pid.c"
 674 #include "trace/beauty/sched_policy.c"
 675 #include "trace/beauty/seccomp.c"
 676 #include "trace/beauty/signum.c"
 677 #include "trace/beauty/socket_type.c"
 678 #include "trace/beauty/waitid_options.c"
 679
 680 struct syscall_arg_fmt {
 681         size_t     (*scnprintf)(char *bf, size_t size, struct syscall_arg *arg);
 682         unsigned long (*mask_val)(struct syscall_arg *arg, unsigned long val);
 683         void       *parm;
 684         const char *name;
 685         bool       show_zero;
 686 };
 687
 688 static struct syscall_fmt {
 689         const char *name;
 690         const char *alias;
 691         struct syscall_arg_fmt arg[6];
 692         u8         nr_args;
 693         bool       errpid;
 694         bool       timeout;
 695         bool       hexret;
 696 } syscall_fmts[] = {
 697         { .name     = "access",
 698           .arg = { [1] = { .scnprintf = SCA_ACCMODE,  /* mode */ }, }, },
 699         { .name     = "arch_prctl",
 700           .arg = { [0] = { .scnprintf = SCA_X86_ARCH_PRCTL_CODE, /* code */ },
 701                    [1] = { .scnprintf = SCA_PTR, /* arg2 */ }, }, },
 702         { .name     = "bind",
 703           .arg = { [1] = { .scnprintf = SCA_SOCKADDR, /* umyaddr */ }, }, },
 704         { .name     = "bpf",
 705           .arg = { [0] = STRARRAY(cmd, bpf_cmd), }, },
 706         { .name     = "brk",        .hexret = true,
 707           .arg = { [0] = { .scnprintf = SCA_PTR, /* brk */ }, }, },
 708         { .name     = "clock_gettime",
 709           .arg = { [0] = STRARRAY(clk_id, clockid), }, },
 710         { .name     = "clone",      .errpid = true, .nr_args = 5,
 711           .arg = { [0] = { .name = "flags",         .scnprintf = SCA_CLONE_FLAGS, },
 712                    [1] = { .name = "child_stack",   .scnprintf = SCA_HEX, },
 713                    [2] = { .name = "parent_tidptr", .scnprintf = SCA_HEX, },
 714                    [3] = { .name = "child_tidptr",  .scnprintf = SCA_HEX, },
 715                    [4] = { .name = "tls",           .scnprintf = SCA_HEX, }, }, },
 716         { .name     = "close",
 717           .arg = { [0] = { .scnprintf = SCA_CLOSE_FD, /* fd */ }, }, },
 718         { .name     = "connect",
 719           .arg = { [1] = { .scnprintf = SCA_SOCKADDR, /* servaddr */ }, }, },
 720         { .name     = "epoll_ctl",
 721           .arg = { [1] = STRARRAY(op, epoll_ctl_ops), }, },
 722         { .name     = "eventfd2",
 723           .arg = { [1] = { .scnprintf = SCA_EFD_FLAGS, /* flags */ }, }, },
 724         { .name     = "fchmodat",
 725           .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
 726         { .name     = "fchownat",
 727           .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
 728         { .name     = "fcntl",
 729           .arg = { [1] = { .scnprintf = SCA_FCNTL_CMD, /* cmd */
 730                            .parm      = &strarrays__fcntl_cmds_arrays,
 731                            .show_zero = true, },
 732                    [2] = { .scnprintf =  SCA_FCNTL_ARG, /* arg */ }, }, },
 733         { .name     = "flock",
 734           .arg = { [1] = { .scnprintf = SCA_FLOCK, /* cmd */ }, }, },
 735         { .name     = "fsconfig",
 736           .arg = { [1] = STRARRAY(cmd, fsconfig_cmds), }, },
 737         { .name     = "fsmount",
 738           .arg = { [1] = STRARRAY_FLAGS(flags, fsmount_flags),
 739                    [2] = { .scnprintf = SCA_FSMOUNT_ATTR_FLAGS, /* attr_flags */ }, }, },
 740         { .name     = "fspick",
 741           .arg = { [0] = { .scnprintf = SCA_FDAT,         /* dfd */ },
 742                    [1] = { .scnprintf = SCA_FILENAME,     /* path */ },
 743                    [2] = { .scnprintf = SCA_FSPICK_FLAGS, /* flags */ }, }, },
 744         { .name     = "fstat", .alias = "newfstat", },
 745         { .name     = "fstatat", .alias = "newfstatat", },
 746         { .name     = "futex",
 747           .arg = { [1] = { .scnprintf = SCA_FUTEX_OP, /* op */ },
 748                    [5] = { .scnprintf = SCA_FUTEX_VAL3, /* val3 */ }, }, },
 749         { .name     = "futimesat",
 750           .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
 751         { .name     = "getitimer",
 752           .arg = { [0] = STRARRAY(which, itimers), }, },
 753         { .name     = "getpid",     .errpid = true, },
 754         { .name     = "getpgid",    .errpid = true, },
 755         { .name     = "getppid",    .errpid = true, },
 756         { .name     = "getrandom",
 757           .arg = { [2] = { .scnprintf = SCA_GETRANDOM_FLAGS, /* flags */ }, }, },
 758         { .name     = "getrlimit",
 759           .arg = { [0] = STRARRAY(resource, rlimit_resources), }, },
 760         { .name     = "gettid",     .errpid = true, },
 761         { .name     = "ioctl",
 762           .arg = {
 763 #if defined(__i386__) || defined(__x86_64__)
 764 /*
 765  * FIXME: Make this available to all arches.
 766  */
 767                    [1] = { .scnprintf = SCA_IOCTL_CMD, /* cmd */ },
 768                    [2] = { .scnprintf = SCA_HEX, /* arg */ }, }, },
 769 #else
 770                    [2] = { .scnprintf = SCA_HEX, /* arg */ }, }, },
 771 #endif
 772         { .name     = "kcmp",       .nr_args = 5,
 773           .arg = { [0] = { .name = "pid1",      .scnprintf = SCA_PID, },
 774                    [1] = { .name = "pid2",      .scnprintf = SCA_PID, },
 775                    [2] = { .name = "type",      .scnprintf = SCA_KCMP_TYPE, },
 776                    [3] = { .name = "idx1",      .scnprintf = SCA_KCMP_IDX, },
 777                    [4] = { .name = "idx2",      .scnprintf = SCA_KCMP_IDX, }, }, },
 778         { .name     = "keyctl",
 779           .arg = { [0] = STRARRAY(option, keyctl_options), }, },
 780         { .name     = "kill",
 781           .arg = { [1] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
 782         { .name     = "linkat",
 783           .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
 784         { .name     = "lseek",
 785           .arg = { [2] = STRARRAY(whence, whences), }, },
 786         { .name     = "lstat", .alias = "newlstat", },
 787         { .name     = "madvise",
 788           .arg = { [0] = { .scnprintf = SCA_HEX,      /* start */ },
 789                    [2] = { .scnprintf = SCA_MADV_BHV, /* behavior */ }, }, },
 790         { .name     = "mkdirat",
 791           .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
 792         { .name     = "mknodat",
 793           .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
 794         { .name     = "mmap",       .hexret = true,
 795 /* The standard mmap maps to old_mmap on s390x */
 796 #if defined(__s390x__)
 797         .alias = "old_mmap",
 798 #endif
 799           .arg = { [2] = { .scnprintf = SCA_MMAP_PROT,  /* prot */ },
 800                    [3] = { .scnprintf = SCA_MMAP_FLAGS, /* flags */ },
 801                    [5] = { .scnprintf = SCA_HEX,        /* offset */ }, }, },
 802         { .name     = "mount",
 803           .arg = { [0] = { .scnprintf = SCA_FILENAME, /* dev_name */ },
 804                    [3] = { .scnprintf = SCA_MOUNT_FLAGS, /* flags */
 805                            .mask_val  = SCAMV_MOUNT_FLAGS, /* flags */ }, }, },
 806         { .name     = "move_mount",
 807           .arg = { [0] = { .scnprintf = SCA_FDAT,       /* from_dfd */ },
 808                    [1] = { .scnprintf = SCA_FILENAME, /* from_pathname */ },
 809                    [2] = { .scnprintf = SCA_FDAT,       /* to_dfd */ },
 810                    [3] = { .scnprintf = SCA_FILENAME, /* to_pathname */ },
 811                    [4] = { .scnprintf = SCA_MOVE_MOUNT_FLAGS, /* flags */ }, }, },
 812         { .name     = "mprotect",
 813           .arg = { [0] = { .scnprintf = SCA_HEX,        /* start */ },
 814                    [2] = { .scnprintf = SCA_MMAP_PROT,  /* prot */ }, }, },
 815         { .name     = "mq_unlink",
 816           .arg = { [0] = { .scnprintf = SCA_FILENAME, /* u_name */ }, }, },
 817         { .name     = "mremap",     .hexret = true,
 818           .arg = { [3] = { .scnprintf = SCA_MREMAP_FLAGS, /* flags */ }, }, },
 819         { .name     = "name_to_handle_at",
 820           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
 821         { .name     = "newfstatat",
 822           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
 823         { .name     = "open",
 824           .arg = { [1] = { .scnprintf = SCA_OPEN_FLAGS, /* flags */ }, }, },
 825         { .name     = "open_by_handle_at",
 826           .arg = { [0] = { .scnprintf = SCA_FDAT,       /* dfd */ },
 827                    [2] = { .scnprintf = SCA_OPEN_FLAGS, /* flags */ }, }, },
 828         { .name     = "openat",
 829           .arg = { [0] = { .scnprintf = SCA_FDAT,       /* dfd */ },
 830                    [2] = { .scnprintf = SCA_OPEN_FLAGS, /* flags */ }, }, },
 831         { .name     = "perf_event_open",
 832           .arg = { [2] = { .scnprintf = SCA_INT,        /* cpu */ },
 833                    [3] = { .scnprintf = SCA_FD,         /* group_fd */ },
 834                    [4] = { .scnprintf = SCA_PERF_FLAGS, /* flags */ }, }, },
 835         { .name     = "pipe2",
 836           .arg = { [1] = { .scnprintf = SCA_PIPE_FLAGS, /* flags */ }, }, },
 837         { .name     = "pkey_alloc",
 838           .arg = { [1] = { .scnprintf = SCA_PKEY_ALLOC_ACCESS_RIGHTS,   /* access_rights */ }, }, },
 839         { .name     = "pkey_free",
 840           .arg = { [0] = { .scnprintf = SCA_INT,        /* key */ }, }, },
 841         { .name     = "pkey_mprotect",
 842           .arg = { [0] = { .scnprintf = SCA_HEX,        /* start */ },
 843                    [2] = { .scnprintf = SCA_MMAP_PROT,  /* prot */ },
 844                    [3] = { .scnprintf = SCA_INT,        /* pkey */ }, }, },
 845         { .name     = "poll", .timeout = true, },
 846         { .name     = "ppoll", .timeout = true, },
 847         { .name     = "prctl",
 848           .arg = { [0] = { .scnprintf = SCA_PRCTL_OPTION, /* option */ },
 849                    [1] = { .scnprintf = SCA_PRCTL_ARG2, /* arg2 */ },
 850                    [2] = { .scnprintf = SCA_PRCTL_ARG3, /* arg3 */ }, }, },
 851         { .name     = "pread", .alias = "pread64", },
 852         { .name     = "preadv", .alias = "pread", },
 853         { .name     = "prlimit64",
 854           .arg = { [1] = STRARRAY(resource, rlimit_resources), }, },
 855         { .name     = "pwrite", .alias = "pwrite64", },
 856         { .name     = "readlinkat",
 857           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
 858         { .name     = "recvfrom",
 859           .arg = { [3] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
 860         { .name     = "recvmmsg",
 861           .arg = { [3] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
 862         { .name     = "recvmsg",
 863           .arg = { [2] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
 864         { .name     = "renameat",
 865           .arg = { [0] = { .scnprintf = SCA_FDAT, /* olddirfd */ },
 866                    [2] = { .scnprintf = SCA_FDAT, /* newdirfd */ }, }, },
 867         { .name     = "renameat2",
 868           .arg = { [0] = { .scnprintf = SCA_FDAT, /* olddirfd */ },
 869                    [2] = { .scnprintf = SCA_FDAT, /* newdirfd */ },
 870                    [4] = { .scnprintf = SCA_RENAMEAT2_FLAGS, /* flags */ }, }, },
 871         { .name     = "rt_sigaction",
 872           .arg = { [0] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
 873         { .name     = "rt_sigprocmask",
 874           .arg = { [0] = STRARRAY(how, sighow), }, },
 875         { .name     = "rt_sigqueueinfo",
 876           .arg = { [1] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
 877         { .name     = "rt_tgsigqueueinfo",
 878           .arg = { [2] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
 879         { .name     = "sched_setscheduler",
 880           .arg = { [1] = { .scnprintf = SCA_SCHED_POLICY, /* policy */ }, }, },
 881         { .name     = "seccomp",
 882           .arg = { [0] = { .scnprintf = SCA_SECCOMP_OP,    /* op */ },
 883                    [1] = { .scnprintf = SCA_SECCOMP_FLAGS, /* flags */ }, }, },
 884         { .name     = "select", .timeout = true, },
 885         { .name     = "sendmmsg",
 886           .arg = { [3] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
 887         { .name     = "sendmsg",
 888           .arg = { [2] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
 889         { .name     = "sendto",
 890           .arg = { [3] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ },
 891                    [4] = { .scnprintf = SCA_SOCKADDR, /* addr */ }, }, },
 892         { .name     = "set_tid_address", .errpid = true, },
 893         { .name     = "setitimer",
 894           .arg = { [0] = STRARRAY(which, itimers), }, },
 895         { .name     = "setrlimit",
 896           .arg = { [0] = STRARRAY(resource, rlimit_resources), }, },
 897         { .name     = "socket",
 898           .arg = { [0] = STRARRAY(family, socket_families),
 899                    [1] = { .scnprintf = SCA_SK_TYPE, /* type */ },
 900                    [2] = { .scnprintf = SCA_SK_PROTO, /* protocol */ }, }, },
 901         { .name     = "socketpair",
 902           .arg = { [0] = STRARRAY(family, socket_families),
 903                    [1] = { .scnprintf = SCA_SK_TYPE, /* type */ },
 904                    [2] = { .scnprintf = SCA_SK_PROTO, /* protocol */ }, }, },
 905         { .name     = "stat", .alias = "newstat", },
 906         { .name     = "statx",
 907           .arg = { [0] = { .scnprintf = SCA_FDAT,        /* fdat */ },
 908                    [2] = { .scnprintf = SCA_STATX_FLAGS, /* flags */ } ,
 909                    [3] = { .scnprintf = SCA_STATX_MASK,  /* mask */ }, }, },
 910         { .name     = "swapoff",
 911           .arg = { [0] = { .scnprintf = SCA_FILENAME, /* specialfile */ }, }, },
 912         { .name     = "swapon",
 913           .arg = { [0] = { .scnprintf = SCA_FILENAME, /* specialfile */ }, }, },
 914         { .name     = "symlinkat",
 915           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
 916         { .name     = "sync_file_range",
 917           .arg = { [3] = { .scnprintf = SCA_SYNC_FILE_RANGE_FLAGS, /* flags */ }, }, },
 918         { .name     = "tgkill",
 919           .arg = { [2] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
 920         { .name     = "tkill",
 921           .arg = { [1] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
 922         { .name     = "umount2", .alias = "umount",
 923           .arg = { [0] = { .scnprintf = SCA_FILENAME, /* name */ }, }, },
 924         { .name     = "uname", .alias = "newuname", },
 925         { .name     = "unlinkat",
 926           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
 927         { .name     = "utimensat",
 928           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dirfd */ }, }, },
 929         { .name     = "wait4",      .errpid = true,
 930           .arg = { [2] = { .scnprintf = SCA_WAITID_OPTIONS, /* options */ }, }, },
 931         { .name     = "waitid",     .errpid = true,
 932           .arg = { [3] = { .scnprintf = SCA_WAITID_OPTIONS, /* options */ }, }, },
 933 };
 934
 935 static int syscall_fmt__cmp(const void *name, const void *fmtp)
 936 {
 937         const struct syscall_fmt *fmt = fmtp;
 938         return strcmp(name, fmt->name);
 939 }
 940
 941 static struct syscall_fmt *syscall_fmt__find(const char *name)
 942 {
 943         const int nmemb = ARRAY_SIZE(syscall_fmts);
 944         return bsearch(name, syscall_fmts, nmemb, sizeof(struct syscall_fmt), syscall_fmt__cmp);
 945 }
 946
 947 static struct syscall_fmt *syscall_fmt__find_by_alias(const char *alias)
 948 {
 949         int i, nmemb = ARRAY_SIZE(syscall_fmts);
 950
 951         for (i = 0; i < nmemb; ++i) {
 952                 if (syscall_fmts[i].alias && strcmp(syscall_fmts[i].alias, alias) == 0)
 953                         return &syscall_fmts[i];
 954         }
 955
 956         return NULL;
 957 }
 958
 959 /*
 960  * is_exit: is this "exit" or "exit_group"?
 961  * is_open: is this "open" or "openat"? To associate the fd returned in sys_exit with the pathname in sys_enter.
 962  * args_size: sum of the sizes of the syscall arguments, anything after that is augmented stuff: pathname for openat, etc.
 963  */
 964 struct syscall {
 965         struct tep_event    *tp_format;
 966         int                 nr_args;
 967         int                 args_size;
 968         bool                is_exit;
 969         bool                is_open;
 970         struct tep_format_field *args;
 971         const char          *name;
 972         struct syscall_fmt  *fmt;
 973         struct syscall_arg_fmt *arg_fmt;
 974 };
 975
 976 /*
 977  * Must match what is in the BPF program:
 978  *
 979  * tools/perf/examples/bpf/augmented_raw_syscalls.c
 980  */
 981 struct bpf_map_syscall_entry {
 982         bool    enabled;
 983         u16     string_args_len[6];
 984 };
 985
 986 /*
 987  * We need to have this 'calculated' boolean because in some cases we really
 988  * don't know what is the duration of a syscall, for instance, when we start
 989  * a session and some threads are waiting for a syscall to finish, say 'poll',
 990  * in which case all we can do is to print "( ? ) for duration and for the
 991  * start timestamp.
 992  */
 993 static size_t fprintf_duration(unsigned long t, bool calculated, FILE *fp)
 994 {
 995         double duration = (double)t / NSEC_PER_MSEC;
 996         size_t printed = fprintf(fp, "(");
 997
 998         if (!calculated)
 999                 printed += fprintf(fp, "         ");
1000         else if (duration >= 1.0)
1001                 printed += color_fprintf(fp, PERF_COLOR_RED, "%6.3f ms", duration);
1002         else if (duration >= 0.01)
1003                 printed += color_fprintf(fp, PERF_COLOR_YELLOW, "%6.3f ms", duration);
1004         else
1005                 printed += color_fprintf(fp, PERF_COLOR_NORMAL, "%6.3f ms", duration);
1006         return printed + fprintf(fp, "): ");
1007 }
1008
1009 /**
1010  * filename.ptr: The filename char pointer that will be vfs_getname'd
1011  * filename.entry_str_pos: Where to insert the string translated from
1012  *                         filename.ptr by the vfs_getname tracepoint/kprobe.
1013  * ret_scnprintf: syscall args may set this to a different syscall return
1014  *                formatter, for instance, fcntl may return fds, file flags, etc.
1015  */
1016 struct thread_trace {
1017         u64               entry_time;
1018         bool              entry_pending;
1019         unsigned long     nr_events;
1020         unsigned long     pfmaj, pfmin;
1021         char              *entry_str;
1022         double            runtime_ms;
1023         size_t            (*ret_scnprintf)(char *bf, size_t size, struct syscall_arg *arg);
1024         struct {
1025                 unsigned long ptr;
1026                 short int     entry_str_pos;
1027                 bool          pending_open;
1028                 unsigned int  namelen;
1029                 char          *name;
1030         } filename;
1031         struct {
1032                 int           max;
1033                 struct file   *table;
1034         } files;
1035
1036         struct intlist *syscall_stats;
1037 };
1038
1039 static struct thread_trace *thread_trace__new(void)
1040 {
1041         struct thread_trace *ttrace =  zalloc(sizeof(struct thread_trace));
1042
1043         if (ttrace) {
1044                 ttrace->files.max = -1;
1045                 ttrace->syscall_stats = intlist__new(NULL);
1046         }
1047
1048         return ttrace;
1049 }
1050
1051 static struct thread_trace *thread__trace(struct thread *thread, FILE *fp)
1052 {
1053         struct thread_trace *ttrace;
1054
1055         if (thread == NULL)
1056                 goto fail;
1057
1058         if (thread__priv(thread) == NULL)
1059                 thread__set_priv(thread, thread_trace__new());
1060
1061         if (thread__priv(thread) == NULL)
1062                 goto fail;
1063
1064         ttrace = thread__priv(thread);
1065         ++ttrace->nr_events;
1066
1067         return ttrace;
1068 fail:
1069         color_fprintf(fp, PERF_COLOR_RED,
1070                       "WARNING: not enough memory, dropping samples!\n");
1071         return NULL;
1072 }
1073
1074
1075 void syscall_arg__set_ret_scnprintf(struct syscall_arg *arg,
1076                                     size_t (*ret_scnprintf)(char *bf, size_t size, struct syscall_arg *arg))
1077 {
1078         struct thread_trace *ttrace = thread__priv(arg->thread);
1079
1080         ttrace->ret_scnprintf = ret_scnprintf;
1081 }
1082
1083 #define TRACE_PFMAJ             (1 << 0)
1084 #define TRACE_PFMIN             (1 << 1)
1085
1086 static const size_t trace__entry_str_size = 2048;
1087
1088 static struct file *thread_trace__files_entry(struct thread_trace *ttrace, int fd)
1089 {
1090         if (fd < 0)
1091                 return NULL;
1092
1093         if (fd > ttrace->files.max) {
1094                 struct file *nfiles = realloc(ttrace->files.table, (fd + 1) * sizeof(struct file));
1095
1096                 if (nfiles == NULL)
1097                         return NULL;
1098
1099                 if (ttrace->files.max != -1) {
1100                         memset(nfiles + ttrace->files.max + 1, 0,
1101                                (fd - ttrace->files.max) * sizeof(struct file));
1102                 } else {
1103                         memset(nfiles, 0, (fd + 1) * sizeof(struct file));
1104                 }
1105
1106                 ttrace->files.table = nfiles;
1107                 ttrace->files.max   = fd;
1108         }
1109
1110         return ttrace->files.table + fd;
1111 }
1112
1113 struct file *thread__files_entry(struct thread *thread, int fd)
1114 {
1115         return thread_trace__files_entry(thread__priv(thread), fd);
1116 }
1117
1118 static int trace__set_fd_pathname(struct thread *thread, int fd, const char *pathname)
1119 {
1120         struct thread_trace *ttrace = thread__priv(thread);
1121         struct file *file = thread_trace__files_entry(ttrace, fd);
1122
1123         if (file != NULL) {
1124                 struct stat st;
1125                 if (stat(pathname, &st) == 0)
1126                         file->dev_maj = major(st.st_rdev);
1127                 file->pathname = strdup(pathname);
1128                 if (file->pathname)
1129                         return 0;
1130         }
1131
1132         return -1;
1133 }
1134
1135 static int thread__read_fd_path(struct thread *thread, int fd)
1136 {
1137         char linkname[PATH_MAX], pathname[PATH_MAX];
1138         struct stat st;
1139         int ret;
1140
1141         if (thread->pid_ == thread->tid) {
1142                 scnprintf(linkname, sizeof(linkname),
1143                           "/proc/%d/fd/%d", thread->pid_, fd);
1144         } else {
1145                 scnprintf(linkname, sizeof(linkname),
1146                           "/proc/%d/task/%d/fd/%d", thread->pid_, thread->tid, fd);
1147         }
1148
1149         if (lstat(linkname, &st) < 0 || st.st_size + 1 > (off_t)sizeof(pathname))
1150                 return -1;
1151
1152         ret = readlink(linkname, pathname, sizeof(pathname));
1153
1154         if (ret < 0 || ret > st.st_size)
1155                 return -1;
1156
1157         pathname[ret] = '\0';
1158         return trace__set_fd_pathname(thread, fd, pathname);
1159 }
1160
1161 static const char *thread__fd_path(struct thread *thread, int fd,
1162                                    struct trace *trace)
1163 {
1164         struct thread_trace *ttrace = thread__priv(thread);
1165
1166         if (ttrace == NULL)
1167                 return NULL;
1168
1169         if (fd < 0)
1170                 return NULL;
1171
1172         if ((fd > ttrace->files.max || ttrace->files.table[fd].pathname == NULL)) {
1173                 if (!trace->live)
1174                         return NULL;
1175                 ++trace->stats.proc_getname;
1176                 if (thread__read_fd_path(thread, fd))
1177                         return NULL;
1178         }
1179
1180         return ttrace->files.table[fd].pathname;
1181 }
1182
1183 size_t syscall_arg__scnprintf_fd(char *bf, size_t size, struct syscall_arg *arg)
1184 {
1185         int fd = arg->val;
1186         size_t printed = scnprintf(bf, size, "%d", fd);
1187         const char *path = thread__fd_path(arg->thread, fd, arg->trace);
1188
1189         if (path)
1190                 printed += scnprintf(bf + printed, size - printed, "<%s>", path);
1191
1192         return printed;
1193 }
1194
1195 size_t pid__scnprintf_fd(struct trace *trace, pid_t pid, int fd, char *bf, size_t size)
1196 {
1197         size_t printed = scnprintf(bf, size, "%d", fd);
1198         struct thread *thread = machine__find_thread(trace->host, pid, pid);
1199
1200         if (thread) {
1201                 const char *path = thread__fd_path(thread, fd, trace);
1202
1203                 if (path)
1204                         printed += scnprintf(bf + printed, size - printed, "<%s>", path);
1205
1206                 thread__put(thread);
1207         }
1208
1209         return printed;
1210 }
1211
1212 static size_t syscall_arg__scnprintf_close_fd(char *bf, size_t size,
1213                                               struct syscall_arg *arg)
1214 {
1215         int fd = arg->val;
1216         size_t printed = syscall_arg__scnprintf_fd(bf, size, arg);
1217         struct thread_trace *ttrace = thread__priv(arg->thread);
1218
1219         if (ttrace && fd >= 0 && fd <= ttrace->files.max)
1220                 zfree(&ttrace->files.table[fd].pathname);
1221
1222         return printed;
1223 }
1224
1225 static void thread__set_filename_pos(struct thread *thread, const char *bf,
1226                                      unsigned long ptr)
1227 {
1228         struct thread_trace *ttrace = thread__priv(thread);
1229
1230         ttrace->filename.ptr = ptr;
1231         ttrace->filename.entry_str_pos = bf - ttrace->entry_str;
1232 }
1233
1234 static size_t syscall_arg__scnprintf_augmented_string(struct syscall_arg *arg, char *bf, size_t size)
1235 {
1236         struct augmented_arg *augmented_arg = arg->augmented.args;
1237         size_t printed = scnprintf(bf, size, "\"%.*s\"", augmented_arg->size, augmented_arg->value);
1238         /*
1239          * So that the next arg with a payload can consume its augmented arg, i.e. for rename* syscalls
1240          * we would have two strings, each prefixed by its size.
1241          */
1242         int consumed = sizeof(*augmented_arg) + augmented_arg->size;
1243
1244         arg->augmented.args = ((void *)arg->augmented.args) + consumed;
1245         arg->augmented.size -= consumed;
1246
1247         return printed;
1248 }
1249
1250 static size_t syscall_arg__scnprintf_filename(char *bf, size_t size,
1251                                               struct syscall_arg *arg)
1252 {
1253         unsigned long ptr = arg->val;
1254
1255         if (arg->augmented.args)
1256                 return syscall_arg__scnprintf_augmented_string(arg, bf, size);
1257
1258         if (!arg->trace->vfs_getname)
1259                 return scnprintf(bf, size, "%#x", ptr);
1260
1261         thread__set_filename_pos(arg->thread, bf, ptr);
1262         return 0;
1263 }
1264
1265 static bool trace__filter_duration(struct trace *trace, double t)
1266 {
1267         return t < (trace->duration_filter * NSEC_PER_MSEC);
1268 }
1269
1270 static size_t __trace__fprintf_tstamp(struct trace *trace, u64 tstamp, FILE *fp)
1271 {
1272         double ts = (double)(tstamp - trace->base_time) / NSEC_PER_MSEC;
1273
1274         return fprintf(fp, "%10.3f ", ts);
1275 }
1276
1277 /*
1278  * We're handling tstamp=0 as an undefined tstamp, i.e. like when we are
1279  * using ttrace->entry_time for a thread that receives a sys_exit without
1280  * first having received a sys_enter ("poll" issued before tracing session
1281  * starts, lost sys_enter exit due to ring buffer overflow).
1282  */
1283 static size_t trace__fprintf_tstamp(struct trace *trace, u64 tstamp, FILE *fp)
1284 {
1285         if (tstamp > 0)
1286                 return __trace__fprintf_tstamp(trace, tstamp, fp);
1287
1288         return fprintf(fp, "         ? ");
1289 }
1290
1291 static bool done = false;
1292 static bool interrupted = false;
1293
1294 static void sig_handler(int sig)
1295 {
1296         done = true;
1297         interrupted = sig == SIGINT;
1298 }
1299
1300 static size_t trace__fprintf_comm_tid(struct trace *trace, struct thread *thread, FILE *fp)
1301 {
1302         size_t printed = 0;
1303
1304         if (trace->multiple_threads) {
1305                 if (trace->show_comm)
1306                         printed += fprintf(fp, "%.14s/", thread__comm_str(thread));
1307                 printed += fprintf(fp, "%d ", thread->tid);
1308         }
1309
1310         return printed;
1311 }
1312
1313 static size_t trace__fprintf_entry_head(struct trace *trace, struct thread *thread,
1314                                         u64 duration, bool duration_calculated, u64 tstamp, FILE *fp)
1315 {
1316         size_t printed = 0;
1317
1318         if (trace->show_tstamp)
1319                 printed = trace__fprintf_tstamp(trace, tstamp, fp);
1320         if (trace->show_duration)
1321                 printed += fprintf_duration(duration, duration_calculated, fp);
1322         return printed + trace__fprintf_comm_tid(trace, thread, fp);
1323 }
1324
1325 static int trace__process_event(struct trace *trace, struct machine *machine,
1326                                 union perf_event *event, struct perf_sample *sample)
1327 {
1328         int ret = 0;
1329
1330         switch (event->header.type) {
1331         case PERF_RECORD_LOST:
1332                 color_fprintf(trace->output, PERF_COLOR_RED,
1333                               "LOST %" PRIu64 " events!\n", event->lost.lost);
1334                 ret = machine__process_lost_event(machine, event, sample);
1335                 break;
1336         default:
1337                 ret = machine__process_event(machine, event, sample);
1338                 break;
1339         }
1340
1341         return ret;
1342 }
1343
1344 static int trace__tool_process(struct perf_tool *tool,
1345                                union perf_event *event,
1346                                struct perf_sample *sample,
1347                                struct machine *machine)
1348 {
1349         struct trace *trace = container_of(tool, struct trace, tool);
1350         return trace__process_event(trace, machine, event, sample);
1351 }
1352
1353 static char *trace__machine__resolve_kernel_addr(void *vmachine, unsigned long long *addrp, char **modp)
1354 {
1355         struct machine *machine = vmachine;
1356
1357         if (machine->kptr_restrict_warned)
1358                 return NULL;
1359
1360         if (symbol_conf.kptr_restrict) {
1361                 pr_warning("Kernel address maps (/proc/{kallsyms,modules}) are restricted.\n\n"
1362                            "Check /proc/sys/kernel/kptr_restrict.\n\n"
1363                            "Kernel samples will not be resolved.\n");
1364                 machine->kptr_restrict_warned = true;
1365                 return NULL;
1366         }
1367
1368         return machine__resolve_kernel_addr(vmachine, addrp, modp);
1369 }
1370
1371 static int trace__symbols_init(struct trace *trace, struct perf_evlist *evlist)
1372 {
1373         int err = symbol__init(NULL);
1374
1375         if (err)
1376                 return err;
1377
1378         trace->host = machine__new_host();
1379         if (trace->host == NULL)
1380                 return -ENOMEM;
1381
1382         err = trace_event__register_resolver(trace->host, trace__machine__resolve_kernel_addr);
1383         if (err < 0)
1384                 goto out;
1385
1386         err = __machine__synthesize_threads(trace->host, &trace->tool, &trace->opts.target,
1387                                             evlist->threads, trace__tool_process, false,
1388                                             1);
1389 out:
1390         if (err)
1391                 symbol__exit();
1392
1393         return err;
1394 }
1395
1396 static void trace__symbols__exit(struct trace *trace)
1397 {
1398         machine__exit(trace->host);
1399         trace->host = NULL;
1400
1401         symbol__exit();
1402 }
1403
1404 static int syscall__alloc_arg_fmts(struct syscall *sc, int nr_args)
1405 {
1406         int idx;
1407
1408         if (nr_args == 6 && sc->fmt && sc->fmt->nr_args != 0)
1409                 nr_args = sc->fmt->nr_args;
1410
1411         sc->arg_fmt = calloc(nr_args, sizeof(*sc->arg_fmt));
1412         if (sc->arg_fmt == NULL)
1413                 return -1;
1414
1415         for (idx = 0; idx < nr_args; ++idx) {
1416                 if (sc->fmt)
1417                         sc->arg_fmt[idx] = sc->fmt->arg[idx];
1418         }
1419
1420         sc->nr_args = nr_args;
1421         return 0;
1422 }
1423
1424 static int syscall__set_arg_fmts(struct syscall *sc)
1425 {
1426         struct tep_format_field *field, *last_field = NULL;
1427         int idx = 0, len;
1428
1429         for (field = sc->args; field; field = field->next, ++idx) {
1430                 last_field = field;
1431
1432                 if (sc->fmt && sc->fmt->arg[idx].scnprintf)
1433                         continue;
1434
1435                 len = strlen(field->name);
1436
1437                 if (strcmp(field->type, "const char *") == 0 &&
1438                     ((len >= 4 && strcmp(field->name + len - 4, "name") == 0) ||
1439                      strstr(field->name, "path") != NULL))
1440                         sc->arg_fmt[idx].scnprintf = SCA_FILENAME;
1441                 else if ((field->flags & TEP_FIELD_IS_POINTER) || strstr(field->name, "addr"))
1442                         sc->arg_fmt[idx].scnprintf = SCA_PTR;
1443                 else if (strcmp(field->type, "pid_t") == 0)
1444                         sc->arg_fmt[idx].scnprintf = SCA_PID;
1445                 else if (strcmp(field->type, "umode_t") == 0)
1446                         sc->arg_fmt[idx].scnprintf = SCA_MODE_T;
1447                 else if ((strcmp(field->type, "int") == 0 ||
1448                           strcmp(field->type, "unsigned int") == 0 ||
1449                           strcmp(field->type, "long") == 0) &&
1450                          len >= 2 && strcmp(field->name + len - 2, "fd") == 0) {
1451                         /*
1452                          * /sys/kernel/tracing/events/syscalls/sys_enter*
1453                          * egrep 'field:.*fd;' .../format|sed -r 's/.*field:([a-z ]+) [a-z_]*fd.+/\1/g'|sort|uniq -c
1454                          * 65 int
1455                          * 23 unsigned int
1456                          * 7 unsigned long
1457                          */
1458                         sc->arg_fmt[idx].scnprintf = SCA_FD;
1459                 }
1460         }
1461
1462         if (last_field)
1463                 sc->args_size = last_field->offset + last_field->size;
1464
1465         return 0;
1466 }
1467
1468 static int trace__read_syscall_info(struct trace *trace, int id)
1469 {
1470         char tp_name[128];
1471         struct syscall *sc;
1472         const char *name = syscalltbl__name(trace->sctbl, id);
1473
1474         if (name == NULL)
1475                 return -1;
1476
1477         if (id > trace->syscalls.max) {
1478                 struct syscall *nsyscalls = realloc(trace->syscalls.table, (id + 1) * sizeof(*sc));
1479
1480                 if (nsyscalls == NULL)
1481                         return -1;
1482
1483                 if (trace->syscalls.max != -1) {
1484                         memset(nsyscalls + trace->syscalls.max + 1, 0,
1485                                (id - trace->syscalls.max) * sizeof(*sc));
1486                 } else {
1487                         memset(nsyscalls, 0, (id + 1) * sizeof(*sc));
1488                 }
1489
1490                 trace->syscalls.table = nsyscalls;
1491                 trace->syscalls.max   = id;
1492         }
1493
1494         sc = trace->syscalls.table + id;
1495         sc->name = name;
1496
1497         sc->fmt  = syscall_fmt__find(sc->name);
1498
1499         snprintf(tp_name, sizeof(tp_name), "sys_enter_%s", sc->name);
1500         sc->tp_format = trace_event__tp_format("syscalls", tp_name);
1501
1502         if (IS_ERR(sc->tp_format) && sc->fmt && sc->fmt->alias) {
1503                 snprintf(tp_name, sizeof(tp_name), "sys_enter_%s", sc->fmt->alias);
1504                 sc->tp_format = trace_event__tp_format("syscalls", tp_name);
1505         }
1506
1507         if (syscall__alloc_arg_fmts(sc, IS_ERR(sc->tp_format) ? 6 : sc->tp_format->format.nr_fields))
1508                 return -1;
1509
1510         if (IS_ERR(sc->tp_format))
1511                 return -1;
1512
1513         sc->args = sc->tp_format->format.fields;
1514         /*
1515          * We need to check and discard the first variable '__syscall_nr'
1516          * or 'nr' that mean the syscall number. It is needless here.
1517          * So drop '__syscall_nr' or 'nr' field but does not exist on older kernels.
1518          */
1519         if (sc->args && (!strcmp(sc->args->name, "__syscall_nr") || !strcmp(sc->args->name, "nr"))) {
1520                 sc->args = sc->args->next;
1521                 --sc->nr_args;
1522         }
1523
1524         sc->is_exit = !strcmp(name, "exit_group") || !strcmp(name, "exit");
1525         sc->is_open = !strcmp(name, "open") || !strcmp(name, "openat");
1526
1527         return syscall__set_arg_fmts(sc);
1528 }
1529
1530 static int trace__validate_ev_qualifier(struct trace *trace)
1531 {
1532         int err = 0;
1533         bool printed_invalid_prefix = false;
1534         struct str_node *pos;
1535         size_t nr_used = 0, nr_allocated = strlist__nr_entries(trace->ev_qualifier);
1536
1537         trace->ev_qualifier_ids.entries = malloc(nr_allocated *
1538                                                  sizeof(trace->ev_qualifier_ids.entries[0]));
1539
1540         if (trace->ev_qualifier_ids.entries == NULL) {
1541                 fputs("Error:\tNot enough memory for allocating events qualifier ids\n",
1542                        trace->output);
1543                 err = -EINVAL;
1544                 goto out;
1545         }
1546
1547         strlist__for_each_entry(pos, trace->ev_qualifier) {
1548                 const char *sc = pos->s;
1549                 int id = syscalltbl__id(trace->sctbl, sc), match_next = -1;
1550
1551                 if (id < 0) {
1552                         id = syscalltbl__strglobmatch_first(trace->sctbl, sc, &match_next);
1553                         if (id >= 0)
1554                                 goto matches;
1555
1556                         if (!printed_invalid_prefix) {
1557                                 pr_debug("Skipping unknown syscalls: ");
1558                                 printed_invalid_prefix = true;
1559                         } else {
1560                                 pr_debug(", ");
1561                         }
1562
1563                         pr_debug("%s", sc);
1564                         continue;
1565                 }
1566 matches:
1567                 trace->ev_qualifier_ids.entries[nr_used++] = id;
1568                 if (match_next == -1)
1569                         continue;
1570
1571                 while (1) {
1572                         id = syscalltbl__strglobmatch_next(trace->sctbl, sc, &match_next);
1573                         if (id < 0)
1574                                 break;
1575                         if (nr_allocated == nr_used) {
1576                                 void *entries;
1577
1578                                 nr_allocated += 8;
1579                                 entries = realloc(trace->ev_qualifier_ids.entries,
1580                                                   nr_allocated * sizeof(trace->ev_qualifier_ids.entries[0]));
1581                                 if (entries == NULL) {
1582                                         err = -ENOMEM;
1583                                         fputs("\nError:\t Not enough memory for parsing\n", trace->output);
1584                                         goto out_free;
1585                                 }
1586                                 trace->ev_qualifier_ids.entries = entries;
1587                         }
1588                         trace->ev_qualifier_ids.entries[nr_used++] = id;
1589                 }
1590         }
1591
1592         trace->ev_qualifier_ids.nr = nr_used;
1593 out:
1594         if (printed_invalid_prefix)
1595                 pr_debug("\n");
1596         return err;
1597 out_free:
1598         zfree(&trace->ev_qualifier_ids.entries);
1599         trace->ev_qualifier_ids.nr = 0;
1600         goto out;
1601 }
1602
1603 /*
1604  * args is to be interpreted as a series of longs but we need to handle
1605  * 8-byte unaligned accesses. args points to raw_data within the event
1606  * and raw_data is guaranteed to be 8-byte unaligned because it is
1607  * preceded by raw_size which is a u32. So we need to copy args to a temp
1608  * variable to read it. Most notably this avoids extended load instructions
1609  * on unaligned addresses
1610  */
1611 unsigned long syscall_arg__val(struct syscall_arg *arg, u8 idx)
1612 {
1613         unsigned long val;
1614         unsigned char *p = arg->args + sizeof(unsigned long) * idx;
1615
1616         memcpy(&val, p, sizeof(val));
1617         return val;
1618 }
1619
1620 static size_t syscall__scnprintf_name(struct syscall *sc, char *bf, size_t size,
1621                                       struct syscall_arg *arg)
1622 {
1623         if (sc->arg_fmt && sc->arg_fmt[arg->idx].name)
1624                 return scnprintf(bf, size, "%s: ", sc->arg_fmt[arg->idx].name);
1625
1626         return scnprintf(bf, size, "arg%d: ", arg->idx);
1627 }
1628
1629 /*
1630  * Check if the value is in fact zero, i.e. mask whatever needs masking, such
1631  * as mount 'flags' argument that needs ignoring some magic flag, see comment
1632  * in tools/perf/trace/beauty/mount_flags.c
1633  */
1634 static unsigned long syscall__mask_val(struct syscall *sc, struct syscall_arg *arg, unsigned long val)
1635 {
1636         if (sc->arg_fmt && sc->arg_fmt[arg->idx].mask_val)
1637                 return sc->arg_fmt[arg->idx].mask_val(arg, val);
1638
1639         return val;
1640 }
1641
1642 static size_t syscall__scnprintf_val(struct syscall *sc, char *bf, size_t size,
1643                                      struct syscall_arg *arg, unsigned long val)
1644 {
1645         if (sc->arg_fmt && sc->arg_fmt[arg->idx].scnprintf) {
1646                 arg->val = val;
1647                 if (sc->arg_fmt[arg->idx].parm)
1648                         arg->parm = sc->arg_fmt[arg->idx].parm;
1649                 return sc->arg_fmt[arg->idx].scnprintf(bf, size, arg);
1650         }
1651         return scnprintf(bf, size, "%ld", val);
1652 }
1653
1654 static size_t syscall__scnprintf_args(struct syscall *sc, char *bf, size_t size,
1655                                       unsigned char *args, void *augmented_args, int augmented_args_size,
1656                                       struct trace *trace, struct thread *thread)
1657 {
1658         size_t printed = 0;
1659         unsigned long val;
1660         u8 bit = 1;
1661         struct syscall_arg arg = {
1662                 .args   = args,
1663                 .augmented = {
1664                         .size = augmented_args_size,
1665                         .args = augmented_args,
1666                 },
1667                 .idx    = 0,
1668                 .mask   = 0,
1669                 .trace  = trace,
1670                 .thread = thread,
1671                 .show_string_prefix = trace->show_string_prefix,
1672         };
1673         struct thread_trace *ttrace = thread__priv(thread);
1674
1675         /*
1676          * Things like fcntl will set this in its 'cmd' formatter to pick the
1677          * right formatter for the return value (an fd? file flags?), which is
1678          * not needed for syscalls that always return a given type, say an fd.
1679          */
1680         ttrace->ret_scnprintf = NULL;
1681
1682         if (sc->args != NULL) {
1683                 struct tep_format_field *field;
1684
1685                 for (field = sc->args; field;
1686                      field = field->next, ++arg.idx, bit <<= 1) {
1687                         if (arg.mask & bit)
1688                                 continue;
1689
1690                         val = syscall_arg__val(&arg, arg.idx);
1691                         /*
1692                          * Some syscall args need some mask, most don't and
1693                          * return val untouched.
1694                          */
1695                         val = syscall__mask_val(sc, &arg, val);
1696
1697                         /*
1698                          * Suppress this argument if its value is zero and
1699                          * and we don't have a string associated in an
1700                          * strarray for it.
1701                          */
1702                         if (val == 0 &&
1703                             !trace->show_zeros &&
1704                             !(sc->arg_fmt &&
1705                               (sc->arg_fmt[arg.idx].show_zero ||
1706                                sc->arg_fmt[arg.idx].scnprintf == SCA_STRARRAY ||
1707                                sc->arg_fmt[arg.idx].scnprintf == SCA_STRARRAYS) &&
1708                               sc->arg_fmt[arg.idx].parm))
1709                                 continue;
1710
1711                         printed += scnprintf(bf + printed, size - printed, "%s", printed ? ", " : "");
1712
1713                         if (trace->show_arg_names)
1714                                 printed += scnprintf(bf + printed, size - printed, "%s: ", field->name);
1715
1716                         printed += syscall__scnprintf_val(sc, bf + printed, size - printed, &arg, val);
1717                 }
1718         } else if (IS_ERR(sc->tp_format)) {
1719                 /*
1720                  * If we managed to read the tracepoint /format file, then we
1721                  * may end up not having any args, like with gettid(), so only
1722                  * print the raw args when we didn't manage to read it.
1723                  */
1724                 while (arg.idx < sc->nr_args) {
1725                         if (arg.mask & bit)
1726                                 goto next_arg;
1727                         val = syscall_arg__val(&arg, arg.idx);
1728                         if (printed)
1729                                 printed += scnprintf(bf + printed, size - printed, ", ");
1730                         printed += syscall__scnprintf_name(sc, bf + printed, size - printed, &arg);
1731                         printed += syscall__scnprintf_val(sc, bf + printed, size - printed, &arg, val);
1732 next_arg:
1733                         ++arg.idx;
1734                         bit <<= 1;
1735                 }
1736         }
1737
1738         return printed;
1739 }
1740
1741 typedef int (*tracepoint_handler)(struct trace *trace, struct perf_evsel *evsel,
1742                                   union perf_event *event,
1743                                   struct perf_sample *sample);
1744
1745 static struct syscall *trace__syscall_info(struct trace *trace,
1746                                            struct perf_evsel *evsel, int id)
1747 {
1748
1749         if (id < 0) {
1750
1751                 /*
1752                  * XXX: Noticed on x86_64, reproduced as far back as 3.0.36, haven't tried
1753                  * before that, leaving at a higher verbosity level till that is
1754                  * explained. Reproduced with plain ftrace with:
1755                  *
1756                  * echo 1 > /t/events/raw_syscalls/sys_exit/enable
1757                  * grep "NR -1 " /t/trace_pipe
1758                  *
1759                  * After generating some load on the machine.
1760                  */
1761                 if (verbose > 1) {
1762                         static u64 n;
1763                         fprintf(trace->output, "Invalid syscall %d id, skipping (%s, %" PRIu64 ") ...\n",
1764                                 id, perf_evsel__name(evsel), ++n);
1765                 }
1766                 return NULL;
1767         }
1768
1769         if ((id > trace->syscalls.max || trace->syscalls.table[id].name == NULL) &&
1770             trace__read_syscall_info(trace, id))
1771                 goto out_cant_read;
1772
1773         if ((id > trace->syscalls.max || trace->syscalls.table[id].name == NULL))
1774                 goto out_cant_read;
1775
1776         return &trace->syscalls.table[id];
1777
1778 out_cant_read:
1779         if (verbose > 0) {
1780                 fprintf(trace->output, "Problems reading syscall %d", id);
1781                 if (id <= trace->syscalls.max && trace->syscalls.table[id].name != NULL)
1782                         fprintf(trace->output, "(%s)", trace->syscalls.table[id].name);
1783                 fputs(" information\n", trace->output);
1784         }
1785         return NULL;
1786 }
1787
1788 static void thread__update_stats(struct thread_trace *ttrace,
1789                                  int id, struct perf_sample *sample)
1790 {
1791         struct int_node *inode;
1792         struct stats *stats;
1793         u64 duration = 0;
1794
1795         inode = intlist__findnew(ttrace->syscall_stats, id);
1796         if (inode == NULL)
1797                 return;
1798
1799         stats = inode->priv;
1800         if (stats == NULL) {
1801                 stats = malloc(sizeof(struct stats));
1802                 if (stats == NULL)
1803                         return;
1804                 init_stats(stats);
1805                 inode->priv = stats;
1806         }
1807
1808         if (ttrace->entry_time && sample->time > ttrace->entry_time)
1809                 duration = sample->time - ttrace->entry_time;
1810
1811         update_stats(stats, duration);
1812 }
1813
1814 static int trace__printf_interrupted_entry(struct trace *trace)
1815 {
1816         struct thread_trace *ttrace;
1817         size_t printed;
1818         int len;
1819
1820         if (trace->failure_only || trace->current == NULL)
1821                 return 0;
1822
1823         ttrace = thread__priv(trace->current);
1824
1825         if (!ttrace->entry_pending)
1826                 return 0;
1827
1828         printed  = trace__fprintf_entry_head(trace, trace->current, 0, false, ttrace->entry_time, trace->output);
1829         printed += len = fprintf(trace->output, "%s)", ttrace->entry_str);
1830
1831         if (len < trace->args_alignment - 4)
1832                 printed += fprintf(trace->output, "%-*s", trace->args_alignment - 4 - len, " ");
1833
1834         printed += fprintf(trace->output, " ...\n");
1835
1836         ttrace->entry_pending = false;
1837         ++trace->nr_events_printed;
1838
1839         return printed;
1840 }
1841
1842 static int trace__fprintf_sample(struct trace *trace, struct perf_evsel *evsel,
1843                                  struct perf_sample *sample, struct thread *thread)
1844 {
1845         int printed = 0;
1846
1847         if (trace->print_sample) {
1848                 double ts = (double)sample->time / NSEC_PER_MSEC;
1849
1850                 printed += fprintf(trace->output, "%22s %10.3f %s %d/%d [%d]\n",
1851                                    perf_evsel__name(evsel), ts,
1852                                    thread__comm_str(thread),
1853                                    sample->pid, sample->tid, sample->cpu);
1854         }
1855
1856         return printed;
1857 }
1858
1859 static void *syscall__augmented_args(struct syscall *sc, struct perf_sample *sample, int *augmented_args_size, int raw_augmented_args_size)
1860 {
1861         void *augmented_args = NULL;
1862         /*
1863          * For now with BPF raw_augmented we hook into raw_syscalls:sys_enter
1864          * and there we get all 6 syscall args plus the tracepoint common fields
1865          * that gets calculated at the start and the syscall_nr (another long).
1866          * So we check if that is the case and if so don't look after the
1867          * sc->args_size but always after the full raw_syscalls:sys_enter payload,
1868          * which is fixed.
1869          *
1870          * We'll revisit this later to pass s->args_size to the BPF augmenter
1871          * (now tools/perf/examples/bpf/augmented_raw_syscalls.c, so that it
1872          * copies only what we need for each syscall, like what happens when we
1873          * use syscalls:sys_enter_NAME, so that we reduce the kernel/userspace
1874          * traffic to just what is needed for each syscall.
1875          */
1876         int args_size = raw_augmented_args_size ?: sc->args_size;
1877
1878         *augmented_args_size = sample->raw_size - args_size;
1879         if (*augmented_args_size > 0)
1880                 augmented_args = sample->raw_data + args_size;
1881
1882         return augmented_args;
1883 }
1884
1885 static int trace__sys_enter(struct trace *trace, struct perf_evsel *evsel,
1886                             union perf_event *event __maybe_unused,
1887                             struct perf_sample *sample)
1888 {
1889         char *msg;
1890         void *args;
1891         int printed = 0;
1892         struct thread *thread;
1893         int id = perf_evsel__sc_tp_uint(evsel, id, sample), err = -1;
1894         int augmented_args_size = 0;
1895         void *augmented_args = NULL;
1896         struct syscall *sc = trace__syscall_info(trace, evsel, id);
1897         struct thread_trace *ttrace;
1898
1899         if (sc == NULL)
1900                 return -1;
1901
1902         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
1903         ttrace = thread__trace(thread, trace->output);
1904         if (ttrace == NULL)
1905                 goto out_put;
1906
1907         trace__fprintf_sample(trace, evsel, sample, thread);
1908
1909         args = perf_evsel__sc_tp_ptr(evsel, args, sample);
1910
1911         if (ttrace->entry_str == NULL) {
1912                 ttrace->entry_str = malloc(trace__entry_str_size);
1913                 if (!ttrace->entry_str)
1914                         goto out_put;
1915         }
1916
1917         if (!(trace->duration_filter || trace->summary_only || trace->min_stack))
1918                 trace__printf_interrupted_entry(trace);
1919         /*
1920          * If this is raw_syscalls.sys_enter, then it always comes with the 6 possible
1921          * arguments, even if the syscall being handled, say "openat", uses only 4 arguments
1922          * this breaks syscall__augmented_args() check for augmented args, as we calculate
1923          * syscall->args_size using each syscalls:sys_enter_NAME tracefs format file,
1924          * so when handling, say the openat syscall, we end up getting 6 args for the
1925          * raw_syscalls:sys_enter event, when we expected just 4, we end up mistakenly
1926          * thinking that the extra 2 u64 args are the augmented filename, so just check
1927          * here and avoid using augmented syscalls when the evsel is the raw_syscalls one.
1928          */
1929         if (evsel != trace->syscalls.events.sys_enter)
1930                 augmented_args = syscall__augmented_args(sc, sample, &augmented_args_size, trace->raw_augmented_syscalls_args_size);
1931         ttrace->entry_time = sample->time;
1932         msg = ttrace->entry_str;
1933         printed += scnprintf(msg + printed, trace__entry_str_size - printed, "%s(", sc->name);
1934
1935         printed += syscall__scnprintf_args(sc, msg + printed, trace__entry_str_size - printed,
1936                                            args, augmented_args, augmented_args_size, trace, thread);
1937
1938         if (sc->is_exit) {
1939                 if (!(trace->duration_filter || trace->summary_only || trace->failure_only || trace->min_stack)) {
1940                         int alignment = 0;
1941
1942                         trace__fprintf_entry_head(trace, thread, 0, false, ttrace->entry_time, trace->output);
1943                         printed = fprintf(trace->output, "%s)", ttrace->entry_str);
1944                         if (trace->args_alignment > printed)
1945                                 alignment = trace->args_alignment - printed;
1946                         fprintf(trace->output, "%*s= ?\n", alignment, " ");
1947                 }
1948         } else {
1949                 ttrace->entry_pending = true;
1950                 /* See trace__vfs_getname & trace__sys_exit */
1951                 ttrace->filename.pending_open = false;
1952         }
1953
1954         if (trace->current != thread) {
1955                 thread__put(trace->current);
1956                 trace->current = thread__get(thread);
1957         }
1958         err = 0;
1959 out_put:
1960         thread__put(thread);
1961         return err;
1962 }
1963
1964 static int trace__fprintf_sys_enter(struct trace *trace, struct perf_evsel *evsel,
1965                                     struct perf_sample *sample)
1966 {
1967         struct thread_trace *ttrace;
1968         struct thread *thread;
1969         int id = perf_evsel__sc_tp_uint(evsel, id, sample), err = -1;
1970         struct syscall *sc = trace__syscall_info(trace, evsel, id);
1971         char msg[1024];
1972         void *args, *augmented_args = NULL;
1973         int augmented_args_size;
1974
1975         if (sc == NULL)
1976                 return -1;
1977
1978         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
1979         ttrace = thread__trace(thread, trace->output);
1980         /*
1981          * We need to get ttrace just to make sure it is there when syscall__scnprintf_args()
1982          * and the rest of the beautifiers accessing it via struct syscall_arg touches it.
1983          */
1984         if (ttrace == NULL)
1985                 goto out_put;
1986
1987         args = perf_evsel__sc_tp_ptr(evsel, args, sample);
1988         augmented_args = syscall__augmented_args(sc, sample, &augmented_args_size, trace->raw_augmented_syscalls_args_size);
1989         syscall__scnprintf_args(sc, msg, sizeof(msg), args, augmented_args, augmented_args_size, trace, thread);
1990         fprintf(trace->output, "%s", msg);
1991         err = 0;
1992 out_put:
1993         thread__put(thread);
1994         return err;
1995 }
1996
1997 static int trace__resolve_callchain(struct trace *trace, struct perf_evsel *evsel,
1998                                     struct perf_sample *sample,
1999                                     struct callchain_cursor *cursor)
2000 {
2001         struct addr_location al;
2002         int max_stack = evsel->attr.sample_max_stack ?
2003                         evsel->attr.sample_max_stack :
2004                         trace->max_stack;
2005         int err;
2006
2007         if (machine__resolve(trace->host, &al, sample) < 0)
2008                 return -1;
2009
2010         err = thread__resolve_callchain(al.thread, cursor, evsel, sample, NULL, NULL, max_stack);
2011         addr_location__put(&al);
2012         return err;
2013 }
2014
2015 static int trace__fprintf_callchain(struct trace *trace, struct perf_sample *sample)
2016 {
2017         /* TODO: user-configurable print_opts */
2018         const unsigned int print_opts = EVSEL__PRINT_SYM |
2019                                         EVSEL__PRINT_DSO |
2020                                         EVSEL__PRINT_UNKNOWN_AS_ADDR;
2021
2022         return sample__fprintf_callchain(sample, 38, print_opts, &callchain_cursor, trace->output);
2023 }
2024
2025 static const char *errno_to_name(struct perf_evsel *evsel, int err)
2026 {
2027         struct perf_env *env = perf_evsel__env(evsel);
2028         const char *arch_name = perf_env__arch(env);
2029
2030         return arch_syscalls__strerrno(arch_name, err);
2031 }
2032
2033 static int trace__sys_exit(struct trace *trace, struct perf_evsel *evsel,
2034                            union perf_event *event __maybe_unused,
2035                            struct perf_sample *sample)
2036 {
2037         long ret;
2038         u64 duration = 0;
2039         bool duration_calculated = false;
2040         struct thread *thread;
2041         int id = perf_evsel__sc_tp_uint(evsel, id, sample), err = -1, callchain_ret = 0, printed = 0;
2042         int alignment = trace->args_alignment;
2043         struct syscall *sc = trace__syscall_info(trace, evsel, id);
2044         struct thread_trace *ttrace;
2045
2046         if (sc == NULL)
2047                 return -1;
2048
2049         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2050         ttrace = thread__trace(thread, trace->output);
2051         if (ttrace == NULL)
2052                 goto out_put;
2053
2054         trace__fprintf_sample(trace, evsel, sample, thread);
2055
2056         if (trace->summary)
2057                 thread__update_stats(ttrace, id, sample);
2058
2059         ret = perf_evsel__sc_tp_uint(evsel, ret, sample);
2060
2061         if (sc->is_open && ret >= 0 && ttrace->filename.pending_open) {
2062                 trace__set_fd_pathname(thread, ret, ttrace->filename.name);
2063                 ttrace->filename.pending_open = false;
2064                 ++trace->stats.vfs_getname;
2065         }
2066
2067         if (ttrace->entry_time) {
2068                 duration = sample->time - ttrace->entry_time;
2069                 if (trace__filter_duration(trace, duration))
2070                         goto out;
2071                 duration_calculated = true;
2072         } else if (trace->duration_filter)
2073                 goto out;
2074
2075         if (sample->callchain) {
2076                 callchain_ret = trace__resolve_callchain(trace, evsel, sample, &callchain_cursor);
2077                 if (callchain_ret == 0) {
2078                         if (callchain_cursor.nr < trace->min_stack)
2079                                 goto out;
2080                         callchain_ret = 1;
2081                 }
2082         }
2083
2084         if (trace->summary_only || (ret >= 0 && trace->failure_only))
2085                 goto out;
2086
2087         trace__fprintf_entry_head(trace, thread, duration, duration_calculated, ttrace->entry_time, trace->output);
2088
2089         if (ttrace->entry_pending) {
2090                 printed = fprintf(trace->output, "%s", ttrace->entry_str);
2091         } else {
2092                 printed += fprintf(trace->output, " ... [");
2093                 color_fprintf(trace->output, PERF_COLOR_YELLOW, "continued");
2094                 printed += 9;
2095                 printed += fprintf(trace->output, "]: %s()", sc->name);
2096         }
2097
2098         printed++; /* the closing ')' */
2099
2100         if (alignment > printed)
2101                 alignment -= printed;
2102         else
2103                 alignment = 0;
2104
2105         fprintf(trace->output, ")%*s= ", alignment, " ");
2106
2107         if (sc->fmt == NULL) {
2108                 if (ret < 0)
2109                         goto errno_print;
2110 signed_print:
2111                 fprintf(trace->output, "%ld", ret);
2112         } else if (ret < 0) {
2113 errno_print: {
2114                 char bf[STRERR_BUFSIZE];
2115                 const char *emsg = str_error_r(-ret, bf, sizeof(bf)),
2116                            *e = errno_to_name(evsel, -ret);
2117
2118                 fprintf(trace->output, "-1 %s (%s)", e, emsg);
2119         }
2120         } else if (ret == 0 && sc->fmt->timeout)
2121                 fprintf(trace->output, "0 (Timeout)");
2122         else if (ttrace->ret_scnprintf) {
2123                 char bf[1024];
2124                 struct syscall_arg arg = {
2125                         .val    = ret,
2126                         .thread = thread,
2127                         .trace  = trace,
2128                 };
2129                 ttrace->ret_scnprintf(bf, sizeof(bf), &arg);
2130                 ttrace->ret_scnprintf = NULL;
2131                 fprintf(trace->output, "%s", bf);
2132         } else if (sc->fmt->hexret)
2133                 fprintf(trace->output, "%#lx", ret);
2134         else if (sc->fmt->errpid) {
2135                 struct thread *child = machine__find_thread(trace->host, ret, ret);
2136
2137                 if (child != NULL) {
2138                         fprintf(trace->output, "%ld", ret);
2139                         if (child->comm_set)
2140                                 fprintf(trace->output, " (%s)", thread__comm_str(child));
2141                         thread__put(child);
2142                 }
2143         } else
2144                 goto signed_print;
2145
2146         fputc('\n', trace->output);
2147
2148         /*
2149          * We only consider an 'event' for the sake of --max-events a non-filtered
2150          * sys_enter + sys_exit and other tracepoint events.
2151          */
2152         if (++trace->nr_events_printed == trace->max_events && trace->max_events != ULONG_MAX)
2153                 interrupted = true;
2154
2155         if (callchain_ret > 0)
2156                 trace__fprintf_callchain(trace, sample);
2157         else if (callchain_ret < 0)
2158                 pr_err("Problem processing %s callchain, skipping...\n", perf_evsel__name(evsel));
2159 out:
2160         ttrace->entry_pending = false;
2161         err = 0;
2162 out_put:
2163         thread__put(thread);
2164         return err;
2165 }
2166
2167 static int trace__vfs_getname(struct trace *trace, struct perf_evsel *evsel,
2168                               union perf_event *event __maybe_unused,
2169                               struct perf_sample *sample)
2170 {
2171         struct thread *thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2172         struct thread_trace *ttrace;
2173         size_t filename_len, entry_str_len, to_move;
2174         ssize_t remaining_space;
2175         char *pos;
2176         const char *filename = perf_evsel__rawptr(evsel, sample, "pathname");
2177
2178         if (!thread)
2179                 goto out;
2180
2181         ttrace = thread__priv(thread);
2182         if (!ttrace)
2183                 goto out_put;
2184
2185         filename_len = strlen(filename);
2186         if (filename_len == 0)
2187                 goto out_put;
2188
2189         if (ttrace->filename.namelen < filename_len) {
2190                 char *f = realloc(ttrace->filename.name, filename_len + 1);
2191
2192                 if (f == NULL)
2193                         goto out_put;
2194
2195                 ttrace->filename.namelen = filename_len;
2196                 ttrace->filename.name = f;
2197         }
2198
2199         strcpy(ttrace->filename.name, filename);
2200         ttrace->filename.pending_open = true;
2201
2202         if (!ttrace->filename.ptr)
2203                 goto out_put;
2204
2205         entry_str_len = strlen(ttrace->entry_str);
2206         remaining_space = trace__entry_str_size - entry_str_len - 1; /* \0 */
2207         if (remaining_space <= 0)
2208                 goto out_put;
2209
2210         if (filename_len > (size_t)remaining_space) {
2211                 filename += filename_len - remaining_space;
2212                 filename_len = remaining_space;
2213         }
2214
2215         to_move = entry_str_len - ttrace->filename.entry_str_pos + 1; /* \0 */
2216         pos = ttrace->entry_str + ttrace->filename.entry_str_pos;
2217         memmove(pos + filename_len, pos, to_move);
2218         memcpy(pos, filename, filename_len);
2219
2220         ttrace->filename.ptr = 0;
2221         ttrace->filename.entry_str_pos = 0;
2222 out_put:
2223         thread__put(thread);
2224 out:
2225         return 0;
2226 }
2227
2228 static int trace__sched_stat_runtime(struct trace *trace, struct perf_evsel *evsel,
2229                                      union perf_event *event __maybe_unused,
2230                                      struct perf_sample *sample)
2231 {
2232         u64 runtime = perf_evsel__intval(evsel, sample, "runtime");
2233         double runtime_ms = (double)runtime / NSEC_PER_MSEC;
2234         struct thread *thread = machine__findnew_thread(trace->host,
2235                                                         sample->pid,
2236                                                         sample->tid);
2237         struct thread_trace *ttrace = thread__trace(thread, trace->output);
2238
2239         if (ttrace == NULL)
2240                 goto out_dump;
2241
2242         ttrace->runtime_ms += runtime_ms;
2243         trace->runtime_ms += runtime_ms;
2244 out_put:
2245         thread__put(thread);
2246         return 0;
2247
2248 out_dump:
2249         fprintf(trace->output, "%s: comm=%s,pid=%u,runtime=%" PRIu64 ",vruntime=%" PRIu64 ")\n",
2250                evsel->name,
2251                perf_evsel__strval(evsel, sample, "comm"),
2252                (pid_t)perf_evsel__intval(evsel, sample, "pid"),
2253                runtime,
2254                perf_evsel__intval(evsel, sample, "vruntime"));
2255         goto out_put;
2256 }
2257
2258 static int bpf_output__printer(enum binary_printer_ops op,
2259                                unsigned int val, void *extra __maybe_unused, FILE *fp)
2260 {
2261         unsigned char ch = (unsigned char)val;
2262
2263         switch (op) {
2264         case BINARY_PRINT_CHAR_DATA:
2265                 return fprintf(fp, "%c", isprint(ch) ? ch : '.');
2266         case BINARY_PRINT_DATA_BEGIN:
2267         case BINARY_PRINT_LINE_BEGIN:
2268         case BINARY_PRINT_ADDR:
2269         case BINARY_PRINT_NUM_DATA:
2270         case BINARY_PRINT_NUM_PAD:
2271         case BINARY_PRINT_SEP:
2272         case BINARY_PRINT_CHAR_PAD:
2273         case BINARY_PRINT_LINE_END:
2274         case BINARY_PRINT_DATA_END:
2275         default:
2276                 break;
2277         }
2278
2279         return 0;
2280 }
2281
2282 static void bpf_output__fprintf(struct trace *trace,
2283                                 struct perf_sample *sample)
2284 {
2285         binary__fprintf(sample->raw_data, sample->raw_size, 8,
2286                         bpf_output__printer, NULL, trace->output);
2287         ++trace->nr_events_printed;
2288 }
2289
2290 static int trace__event_handler(struct trace *trace, struct perf_evsel *evsel,
2291                                 union perf_event *event __maybe_unused,
2292                                 struct perf_sample *sample)
2293 {
2294         struct thread *thread;
2295         int callchain_ret = 0;
2296         /*
2297          * Check if we called perf_evsel__disable(evsel) due to, for instance,
2298          * this event's max_events having been hit and this is an entry coming
2299          * from the ring buffer that we should discard, since the max events
2300          * have already been considered/printed.
2301          */
2302         if (evsel->disabled)
2303                 return 0;
2304
2305         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2306
2307         if (sample->callchain) {
2308                 callchain_ret = trace__resolve_callchain(trace, evsel, sample, &callchain_cursor);
2309                 if (callchain_ret == 0) {
2310                         if (callchain_cursor.nr < trace->min_stack)
2311                                 goto out;
2312                         callchain_ret = 1;
2313                 }
2314         }
2315
2316         trace__printf_interrupted_entry(trace);
2317         trace__fprintf_tstamp(trace, sample->time, trace->output);
2318
2319         if (trace->trace_syscalls && trace->show_duration)
2320                 fprintf(trace->output, "(         ): ");
2321
2322         if (thread)
2323                 trace__fprintf_comm_tid(trace, thread, trace->output);
2324
2325         if (evsel == trace->syscalls.events.augmented) {
2326                 int id = perf_evsel__sc_tp_uint(evsel, id, sample);
2327                 struct syscall *sc = trace__syscall_info(trace, evsel, id);
2328
2329                 if (sc) {
2330                         fprintf(trace->output, "%s(", sc->name);
2331                         trace__fprintf_sys_enter(trace, evsel, sample);
2332                         fputc(')', trace->output);
2333                         goto newline;
2334                 }
2335
2336                 /*
2337                  * XXX: Not having the associated syscall info or not finding/adding
2338                  *      the thread should never happen, but if it does...
2339                  *      fall thru and print it as a bpf_output event.
2340                  */
2341         }
2342
2343         fprintf(trace->output, "%s:", evsel->name);
2344
2345         if (perf_evsel__is_bpf_output(evsel)) {
2346                 bpf_output__fprintf(trace, sample);
2347         } else if (evsel->tp_format) {
2348                 if (strncmp(evsel->tp_format->name, "sys_enter_", 10) ||
2349                     trace__fprintf_sys_enter(trace, evsel, sample)) {
2350                         event_format__fprintf(evsel->tp_format, sample->cpu,
2351                                               sample->raw_data, sample->raw_size,
2352                                               trace->output);
2353                         ++trace->nr_events_printed;
2354
2355                         if (evsel->max_events != ULONG_MAX && ++evsel->nr_events_printed == evsel->max_events) {
2356                                 perf_evsel__disable(evsel);
2357                                 perf_evsel__close(evsel);
2358                         }
2359                 }
2360         }
2361
2362 newline:
2363         fprintf(trace->output, "\n");
2364
2365         if (callchain_ret > 0)
2366                 trace__fprintf_callchain(trace, sample);
2367         else if (callchain_ret < 0)
2368                 pr_err("Problem processing %s callchain, skipping...\n", perf_evsel__name(evsel));
2369 out:
2370         thread__put(thread);
2371         return 0;
2372 }
2373
2374 static void print_location(FILE *f, struct perf_sample *sample,
2375                            struct addr_location *al,
2376                            bool print_dso, bool print_sym)
2377 {
2378
2379         if ((verbose > 0 || print_dso) && al->map)
2380                 fprintf(f, "%s@", al->map->dso->long_name);
2381
2382         if ((verbose > 0 || print_sym) && al->sym)
2383                 fprintf(f, "%s+0x%" PRIx64, al->sym->name,
2384                         al->addr - al->sym->start);
2385         else if (al->map)
2386                 fprintf(f, "0x%" PRIx64, al->addr);
2387         else
2388                 fprintf(f, "0x%" PRIx64, sample->addr);
2389 }
2390
2391 static int trace__pgfault(struct trace *trace,
2392                           struct perf_evsel *evsel,
2393                           union perf_event *event __maybe_unused,
2394                           struct perf_sample *sample)
2395 {
2396         struct thread *thread;
2397         struct addr_location al;
2398         char map_type = 'd';
2399         struct thread_trace *ttrace;
2400         int err = -1;
2401         int callchain_ret = 0;
2402
2403         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2404
2405         if (sample->callchain) {
2406                 callchain_ret = trace__resolve_callchain(trace, evsel, sample, &callchain_cursor);
2407                 if (callchain_ret == 0) {
2408                         if (callchain_cursor.nr < trace->min_stack)
2409                                 goto out_put;
2410                         callchain_ret = 1;
2411                 }
2412         }
2413
2414         ttrace = thread__trace(thread, trace->output);
2415         if (ttrace == NULL)
2416                 goto out_put;
2417
2418         if (evsel->attr.config == PERF_COUNT_SW_PAGE_FAULTS_MAJ)
2419                 ttrace->pfmaj++;
2420         else
2421                 ttrace->pfmin++;
2422
2423         if (trace->summary_only)
2424                 goto out;
2425
2426         thread__find_symbol(thread, sample->cpumode, sample->ip, &al);
2427
2428         trace__fprintf_entry_head(trace, thread, 0, true, sample->time, trace->output);
2429
2430         fprintf(trace->output, "%sfault [",
2431                 evsel->attr.config == PERF_COUNT_SW_PAGE_FAULTS_MAJ ?
2432                 "maj" : "min");
2433
2434         print_location(trace->output, sample, &al, false, true);
2435
2436         fprintf(trace->output, "] => ");
2437
2438         thread__find_symbol(thread, sample->cpumode, sample->addr, &al);
2439
2440         if (!al.map) {
2441                 thread__find_symbol(thread, sample->cpumode, sample->addr, &al);
2442
2443                 if (al.map)
2444                         map_type = 'x';
2445                 else
2446                         map_type = '?';
2447         }
2448
2449         print_location(trace->output, sample, &al, true, false);
2450
2451         fprintf(trace->output, " (%c%c)\n", map_type, al.level);
2452
2453         if (callchain_ret > 0)
2454                 trace__fprintf_callchain(trace, sample);
2455         else if (callchain_ret < 0)
2456                 pr_err("Problem processing %s callchain, skipping...\n", perf_evsel__name(evsel));
2457
2458         ++trace->nr_events_printed;
2459 out:
2460         err = 0;
2461 out_put:
2462         thread__put(thread);
2463         return err;
2464 }
2465
2466 static void trace__set_base_time(struct trace *trace,
2467                                  struct perf_evsel *evsel,
2468                                  struct perf_sample *sample)
2469 {
2470         /*
2471          * BPF events were not setting PERF_SAMPLE_TIME, so be more robust
2472          * and don't use sample->time unconditionally, we may end up having
2473          * some other event in the future without PERF_SAMPLE_TIME for good
2474          * reason, i.e. we may not be interested in its timestamps, just in
2475          * it taking place, picking some piece of information when it
2476          * appears in our event stream (vfs_getname comes to mind).
2477          */
2478         if (trace->base_time == 0 && !trace->full_time &&
2479             (evsel->attr.sample_type & PERF_SAMPLE_TIME))
2480                 trace->base_time = sample->time;
2481 }
2482
2483 static int trace__process_sample(struct perf_tool *tool,
2484                                  union perf_event *event,
2485                                  struct perf_sample *sample,
2486                                  struct perf_evsel *evsel,
2487                                  struct machine *machine __maybe_unused)
2488 {
2489         struct trace *trace = container_of(tool, struct trace, tool);
2490         struct thread *thread;
2491         int err = 0;
2492
2493         tracepoint_handler handler = evsel->handler;
2494
2495         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2496         if (thread && thread__is_filtered(thread))
2497                 goto out;
2498
2499         trace__set_base_time(trace, evsel, sample);
2500
2501         if (handler) {
2502                 ++trace->nr_events;
2503                 handler(trace, evsel, event, sample);
2504         }
2505 out:
2506         thread__put(thread);
2507         return err;
2508 }
2509
2510 static int trace__record(struct trace *trace, int argc, const char **argv)
2511 {
2512         unsigned int rec_argc, i, j;
2513         const char **rec_argv;
2514         const char * const record_args[] = {
2515                 "record",
2516                 "-R",
2517                 "-m", "1024",
2518                 "-c", "1",
2519         };
2520
2521         const char * const sc_args[] = { "-e", };
2522         unsigned int sc_args_nr = ARRAY_SIZE(sc_args);
2523         const char * const majpf_args[] = { "-e", "major-faults" };
2524         unsigned int majpf_args_nr = ARRAY_SIZE(majpf_args);
2525         const char * const minpf_args[] = { "-e", "minor-faults" };
2526         unsigned int minpf_args_nr = ARRAY_SIZE(minpf_args);
2527
2528         /* +1 is for the event string below */
2529         rec_argc = ARRAY_SIZE(record_args) + sc_args_nr + 1 +
2530                 majpf_args_nr + minpf_args_nr + argc;
2531         rec_argv = calloc(rec_argc + 1, sizeof(char *));
2532
2533         if (rec_argv == NULL)
2534                 return -ENOMEM;
2535
2536         j = 0;
2537         for (i = 0; i < ARRAY_SIZE(record_args); i++)
2538                 rec_argv[j++] = record_args[i];
2539
2540         if (trace->trace_syscalls) {
2541                 for (i = 0; i < sc_args_nr; i++)
2542                         rec_argv[j++] = sc_args[i];
2543
2544                 /* event string may be different for older kernels - e.g., RHEL6 */
2545                 if (is_valid_tracepoint("raw_syscalls:sys_enter"))
2546                         rec_argv[j++] = "raw_syscalls:sys_enter,raw_syscalls:sys_exit";
2547                 else if (is_valid_tracepoint("syscalls:sys_enter"))
2548                         rec_argv[j++] = "syscalls:sys_enter,syscalls:sys_exit";
2549                 else {
2550                         pr_err("Neither raw_syscalls nor syscalls events exist.\n");
2551                         free(rec_argv);
2552                         return -1;
2553                 }
2554         }
2555
2556         if (trace->trace_pgfaults & TRACE_PFMAJ)
2557                 for (i = 0; i < majpf_args_nr; i++)
2558                         rec_argv[j++] = majpf_args[i];
2559
2560         if (trace->trace_pgfaults & TRACE_PFMIN)
2561                 for (i = 0; i < minpf_args_nr; i++)
2562                         rec_argv[j++] = minpf_args[i];
2563
2564         for (i = 0; i < (unsigned int)argc; i++)
2565                 rec_argv[j++] = argv[i];
2566
2567         return cmd_record(j, rec_argv);
2568 }
2569
2570 static size_t trace__fprintf_thread_summary(struct trace *trace, FILE *fp);
2571
2572 static bool perf_evlist__add_vfs_getname(struct perf_evlist *evlist)
2573 {
2574         bool found = false;
2575         struct perf_evsel *evsel, *tmp;
2576         struct parse_events_error err = { .idx = 0, };
2577         int ret = parse_events(evlist, "probe:vfs_getname*", &err);
2578
2579         if (ret)
2580                 return false;
2581
2582         evlist__for_each_entry_safe(evlist, evsel, tmp) {
2583                 if (!strstarts(perf_evsel__name(evsel), "probe:vfs_getname"))
2584                         continue;
2585
2586                 if (perf_evsel__field(evsel, "pathname")) {
2587                         evsel->handler = trace__vfs_getname;
2588                         found = true;
2589                         continue;
2590                 }
2591
2592                 list_del_init(&evsel->node);
2593                 evsel->evlist = NULL;
2594                 perf_evsel__delete(evsel);
2595         }
2596
2597         return found;
2598 }
2599
2600 static struct perf_evsel *perf_evsel__new_pgfault(u64 config)
2601 {
2602         struct perf_evsel *evsel;
2603         struct perf_event_attr attr = {
2604                 .type = PERF_TYPE_SOFTWARE,
2605                 .mmap_data = 1,
2606         };
2607
2608         attr.config = config;
2609         attr.sample_period = 1;
2610
2611         event_attr_init(&attr);
2612
2613         evsel = perf_evsel__new(&attr);
2614         if (evsel)
2615                 evsel->handler = trace__pgfault;
2616
2617         return evsel;
2618 }
2619
2620 static void trace__handle_event(struct trace *trace, union perf_event *event, struct perf_sample *sample)
2621 {
2622         const u32 type = event->header.type;
2623         struct perf_evsel *evsel;
2624
2625         if (type != PERF_RECORD_SAMPLE) {
2626                 trace__process_event(trace, trace->host, event, sample);
2627                 return;
2628         }
2629
2630         evsel = perf_evlist__id2evsel(trace->evlist, sample->id);
2631         if (evsel == NULL) {
2632                 fprintf(trace->output, "Unknown tp ID %" PRIu64 ", skipping...\n", sample->id);
2633                 return;
2634         }
2635
2636         trace__set_base_time(trace, evsel, sample);
2637
2638         if (evsel->attr.type == PERF_TYPE_TRACEPOINT &&
2639             sample->raw_data == NULL) {
2640                 fprintf(trace->output, "%s sample with no payload for tid: %d, cpu %d, raw_size=%d, skipping...\n",
2641                        perf_evsel__name(evsel), sample->tid,
2642                        sample->cpu, sample->raw_size);
2643         } else {
2644                 tracepoint_handler handler = evsel->handler;
2645                 handler(trace, evsel, event, sample);
2646         }
2647
2648         if (trace->nr_events_printed >= trace->max_events && trace->max_events != ULONG_MAX)
2649                 interrupted = true;
2650 }
2651
2652 static int trace__add_syscall_newtp(struct trace *trace)
2653 {
2654         int ret = -1;
2655         struct perf_evlist *evlist = trace->evlist;
2656         struct perf_evsel *sys_enter, *sys_exit;
2657
2658         sys_enter = perf_evsel__raw_syscall_newtp("sys_enter", trace__sys_enter);
2659         if (sys_enter == NULL)
2660                 goto out;
2661
2662         if (perf_evsel__init_sc_tp_ptr_field(sys_enter, args))
2663                 goto out_delete_sys_enter;
2664
2665         sys_exit = perf_evsel__raw_syscall_newtp("sys_exit", trace__sys_exit);
2666         if (sys_exit == NULL)
2667                 goto out_delete_sys_enter;
2668
2669         if (perf_evsel__init_sc_tp_uint_field(sys_exit, ret))
2670                 goto out_delete_sys_exit;
2671
2672         perf_evsel__config_callchain(sys_enter, &trace->opts, &callchain_param);
2673         perf_evsel__config_callchain(sys_exit, &trace->opts, &callchain_param);
2674
2675         perf_evlist__add(evlist, sys_enter);
2676         perf_evlist__add(evlist, sys_exit);
2677
2678         if (callchain_param.enabled && !trace->kernel_syscallchains) {
2679                 /*
2680                  * We're interested only in the user space callchain
2681                  * leading to the syscall, allow overriding that for
2682                  * debugging reasons using --kernel_syscall_callchains
2683                  */
2684                 sys_exit->attr.exclude_callchain_kernel = 1;
2685         }
2686
2687         trace->syscalls.events.sys_enter = sys_enter;
2688         trace->syscalls.events.sys_exit  = sys_exit;
2689
2690         ret = 0;
2691 out:
2692         return ret;
2693
2694 out_delete_sys_exit:
2695         perf_evsel__delete_priv(sys_exit);
2696 out_delete_sys_enter:
2697         perf_evsel__delete_priv(sys_enter);
2698         goto out;
2699 }
2700
2701 static int trace__set_ev_qualifier_tp_filter(struct trace *trace)
2702 {
2703         int err = -1;
2704         struct perf_evsel *sys_exit;
2705         char *filter = asprintf_expr_inout_ints("id", !trace->not_ev_qualifier,
2706                                                 trace->ev_qualifier_ids.nr,
2707                                                 trace->ev_qualifier_ids.entries);
2708
2709         if (filter == NULL)
2710                 goto out_enomem;
2711
2712         if (!perf_evsel__append_tp_filter(trace->syscalls.events.sys_enter,
2713                                           filter)) {
2714                 sys_exit = trace->syscalls.events.sys_exit;
2715                 err = perf_evsel__append_tp_filter(sys_exit, filter);
2716         }
2717
2718         free(filter);
2719 out:
2720         return err;
2721 out_enomem:
2722         errno = ENOMEM;
2723         goto out;
2724 }
2725
2726 #ifdef HAVE_LIBBPF_SUPPORT
2727 static void trace__init_bpf_map_syscall_args(struct trace *trace, int id, struct bpf_map_syscall_entry *entry)
2728 {
2729         struct syscall *sc = trace__syscall_info(trace, NULL, id);
2730         int arg = 0;
2731
2732         if (sc == NULL)
2733                 goto out;
2734
2735         for (; arg < sc->nr_args; ++arg) {
2736                 entry->string_args_len[arg] = 0;
2737                 if (sc->arg_fmt[arg].scnprintf == SCA_FILENAME) {
2738                         /* Should be set like strace -s strsize */
2739                         entry->string_args_len[arg] = PATH_MAX;
2740                 }
2741         }
2742 out:
2743         for (; arg < 6; ++arg)
2744                 entry->string_args_len[arg] = 0;
2745 }
2746 static int trace__set_ev_qualifier_bpf_filter(struct trace *trace)
2747 {
2748         int fd = bpf_map__fd(trace->syscalls.map);
2749         struct bpf_map_syscall_entry value = {
2750                 .enabled = !trace->not_ev_qualifier,
2751         };
2752         int err = 0;
2753         size_t i;
2754
2755         for (i = 0; i < trace->ev_qualifier_ids.nr; ++i) {
2756                 int key = trace->ev_qualifier_ids.entries[i];
2757
2758                 if (value.enabled)
2759                         trace__init_bpf_map_syscall_args(trace, key, &value);
2760
2761                 err = bpf_map_update_elem(fd, &key, &value, BPF_EXIST);
2762                 if (err)
2763                         break;
2764         }
2765
2766         return err;
2767 }
2768
2769 static int __trace__init_syscalls_bpf_map(struct trace *trace, bool enabled)
2770 {
2771         int fd = bpf_map__fd(trace->syscalls.map);
2772         struct bpf_map_syscall_entry value = {
2773                 .enabled = enabled,
2774         };
2775         int err = 0, key;
2776
2777         for (key = 0; key < trace->sctbl->syscalls.nr_entries; ++key) {
2778                 if (enabled)
2779                         trace__init_bpf_map_syscall_args(trace, key, &value);
2780
2781                 err = bpf_map_update_elem(fd, &key, &value, BPF_ANY);
2782                 if (err)
2783                         break;
2784         }
2785
2786         return err;
2787 }
2788
2789 static int trace__init_syscalls_bpf_map(struct trace *trace)
2790 {
2791         bool enabled = true;
2792
2793         if (trace->ev_qualifier_ids.nr)
2794                 enabled = trace->not_ev_qualifier;
2795
2796         return __trace__init_syscalls_bpf_map(trace, enabled);
2797 }
2798 #else
2799 static int trace__set_ev_qualifier_bpf_filter(struct trace *trace __maybe_unused)
2800 {
2801         return 0;
2802 }
2803
2804 static int trace__init_syscalls_bpf_map(struct trace *trace __maybe_unused)
2805 {
2806         return 0;
2807 }
2808 #endif // HAVE_LIBBPF_SUPPORT
2809
2810 static int trace__set_ev_qualifier_filter(struct trace *trace)
2811 {
2812         if (trace->syscalls.map)
2813                 return trace__set_ev_qualifier_bpf_filter(trace);
2814         if (trace->syscalls.events.sys_enter)
2815                 return trace__set_ev_qualifier_tp_filter(trace);
2816         return 0;
2817 }
2818
2819 static int bpf_map__set_filter_pids(struct bpf_map *map __maybe_unused,
2820                                     size_t npids __maybe_unused, pid_t *pids __maybe_unused)
2821 {
2822         int err = 0;
2823 #ifdef HAVE_LIBBPF_SUPPORT
2824         bool value = true;
2825         int map_fd = bpf_map__fd(map);
2826         size_t i;
2827
2828         for (i = 0; i < npids; ++i) {
2829                 err = bpf_map_update_elem(map_fd, &pids[i], &value, BPF_ANY);
2830                 if (err)
2831                         break;
2832         }
2833 #endif
2834         return err;
2835 }
2836
2837 static int trace__set_filter_loop_pids(struct trace *trace)
2838 {
2839         unsigned int nr = 1, err;
2840         pid_t pids[32] = {
2841                 getpid(),
2842         };
2843         struct thread *thread = machine__find_thread(trace->host, pids[0], pids[0]);
2844
2845         while (thread && nr < ARRAY_SIZE(pids)) {
2846                 struct thread *parent = machine__find_thread(trace->host, thread->ppid, thread->ppid);
2847
2848                 if (parent == NULL)
2849                         break;
2850
2851                 if (!strcmp(thread__comm_str(parent), "sshd") ||
2852                     strstarts(thread__comm_str(parent), "gnome-terminal")) {
2853                         pids[nr++] = parent->tid;
2854                         break;
2855                 }
2856                 thread = parent;
2857         }
2858
2859         err = perf_evlist__set_tp_filter_pids(trace->evlist, nr, pids);
2860         if (!err && trace->filter_pids.map)
2861                 err = bpf_map__set_filter_pids(trace->filter_pids.map, nr, pids);
2862
2863         return err;
2864 }
2865
2866 static int trace__set_filter_pids(struct trace *trace)
2867 {
2868         int err = 0;
2869         /*
2870          * Better not use !target__has_task() here because we need to cover the
2871          * case where no threads were specified in the command line, but a
2872          * workload was, and in that case we will fill in the thread_map when
2873          * we fork the workload in perf_evlist__prepare_workload.
2874          */
2875         if (trace->filter_pids.nr > 0) {
2876                 err = perf_evlist__set_tp_filter_pids(trace->evlist, trace->filter_pids.nr,
2877                                                       trace->filter_pids.entries);
2878                 if (!err && trace->filter_pids.map) {
2879                         err = bpf_map__set_filter_pids(trace->filter_pids.map, trace->filter_pids.nr,
2880                                                        trace->filter_pids.entries);
2881                 }
2882         } else if (thread_map__pid(trace->evlist->threads, 0) == -1) {
2883                 err = trace__set_filter_loop_pids(trace);
2884         }
2885
2886         return err;
2887 }
2888
2889 static int __trace__deliver_event(struct trace *trace, union perf_event *event)
2890 {
2891         struct perf_evlist *evlist = trace->evlist;
2892         struct perf_sample sample;
2893         int err;
2894
2895         err = perf_evlist__parse_sample(evlist, event, &sample);
2896         if (err)
2897                 fprintf(trace->output, "Can't parse sample, err = %d, skipping...\n", err);
2898         else
2899                 trace__handle_event(trace, event, &sample);
2900
2901         return 0;
2902 }
2903
2904 static int __trace__flush_events(struct trace *trace)
2905 {
2906         u64 first = ordered_events__first_time(&trace->oe.data);
2907         u64 flush = trace->oe.last - NSEC_PER_SEC;
2908
2909         /* Is there some thing to flush.. */
2910         if (first && first < flush)
2911                 return ordered_events__flush_time(&trace->oe.data, flush);
2912
2913         return 0;
2914 }
2915
2916 static int trace__flush_events(struct trace *trace)
2917 {
2918         return !trace->sort_events ? 0 : __trace__flush_events(trace);
2919 }
2920
2921 static int trace__deliver_event(struct trace *trace, union perf_event *event)
2922 {
2923         int err;
2924
2925         if (!trace->sort_events)
2926                 return __trace__deliver_event(trace, event);
2927
2928         err = perf_evlist__parse_sample_timestamp(trace->evlist, event, &trace->oe.last);
2929         if (err && err != -1)
2930                 return err;
2931
2932         err = ordered_events__queue(&trace->oe.data, event, trace->oe.last, 0);
2933         if (err)
2934                 return err;
2935
2936         return trace__flush_events(trace);
2937 }
2938
2939 static int ordered_events__deliver_event(struct ordered_events *oe,
2940                                          struct ordered_event *event)
2941 {
2942         struct trace *trace = container_of(oe, struct trace, oe.data);
2943
2944         return __trace__deliver_event(trace, event->event);
2945 }
2946
2947 static int trace__run(struct trace *trace, int argc, const char **argv)
2948 {
2949         struct perf_evlist *evlist = trace->evlist;
2950         struct perf_evsel *evsel, *pgfault_maj = NULL, *pgfault_min = NULL;
2951         int err = -1, i;
2952         unsigned long before;
2953         const bool forks = argc > 0;
2954         bool draining = false;
2955
2956         trace->live = true;
2957
2958         if (!trace->raw_augmented_syscalls) {
2959                 if (trace->trace_syscalls && trace__add_syscall_newtp(trace))
2960                         goto out_error_raw_syscalls;
2961
2962                 if (trace->trace_syscalls)
2963                         trace->vfs_getname = perf_evlist__add_vfs_getname(evlist);
2964         }
2965
2966         if ((trace->trace_pgfaults & TRACE_PFMAJ)) {
2967                 pgfault_maj = perf_evsel__new_pgfault(PERF_COUNT_SW_PAGE_FAULTS_MAJ);
2968                 if (pgfault_maj == NULL)
2969                         goto out_error_mem;
2970                 perf_evsel__config_callchain(pgfault_maj, &trace->opts, &callchain_param);
2971                 perf_evlist__add(evlist, pgfault_maj);
2972         }
2973
2974         if ((trace->trace_pgfaults & TRACE_PFMIN)) {
2975                 pgfault_min = perf_evsel__new_pgfault(PERF_COUNT_SW_PAGE_FAULTS_MIN);
2976                 if (pgfault_min == NULL)
2977                         goto out_error_mem;
2978                 perf_evsel__config_callchain(pgfault_min, &trace->opts, &callchain_param);
2979                 perf_evlist__add(evlist, pgfault_min);
2980         }
2981
2982         if (trace->sched &&
2983             perf_evlist__add_newtp(evlist, "sched", "sched_stat_runtime",
2984                                    trace__sched_stat_runtime))
2985                 goto out_error_sched_stat_runtime;
2986
2987         /*
2988          * If a global cgroup was set, apply it to all the events without an
2989          * explicit cgroup. I.e.:
2990          *
2991          *      trace -G A -e sched:*switch
2992          *
2993          * Will set all raw_syscalls:sys_{enter,exit}, pgfault, vfs_getname, etc
2994          * _and_ sched:sched_switch to the 'A' cgroup, while:
2995          *
2996          * trace -e sched:*switch -G A
2997          *
2998          * will only set the sched:sched_switch event to the 'A' cgroup, all the
2999          * other events (raw_syscalls:sys_{enter,exit}, etc are left "without"
3000          * a cgroup (on the root cgroup, sys wide, etc).
3001          *
3002          * Multiple cgroups:
3003          *
3004          * trace -G A -e sched:*switch -G B
3005          *
3006          * the syscall ones go to the 'A' cgroup, the sched:sched_switch goes
3007          * to the 'B' cgroup.
3008          *
3009          * evlist__set_default_cgroup() grabs a reference of the passed cgroup
3010          * only for the evsels still without a cgroup, i.e. evsel->cgroup == NULL.
3011          */
3012         if (trace->cgroup)
3013                 evlist__set_default_cgroup(trace->evlist, trace->cgroup);
3014
3015         err = perf_evlist__create_maps(evlist, &trace->opts.target);
3016         if (err < 0) {
3017                 fprintf(trace->output, "Problems parsing the target to trace, check your options!\n");
3018                 goto out_delete_evlist;
3019         }
3020
3021         err = trace__symbols_init(trace, evlist);
3022         if (err < 0) {
3023                 fprintf(trace->output, "Problems initializing symbol libraries!\n");
3024                 goto out_delete_evlist;
3025         }
3026
3027         perf_evlist__config(evlist, &trace->opts, &callchain_param);
3028
3029         signal(SIGCHLD, sig_handler);
3030         signal(SIGINT, sig_handler);
3031
3032         if (forks) {
3033                 err = perf_evlist__prepare_workload(evlist, &trace->opts.target,
3034                                                     argv, false, NULL);
3035                 if (err < 0) {
3036                         fprintf(trace->output, "Couldn't run the workload!\n");
3037                         goto out_delete_evlist;
3038                 }
3039         }
3040
3041         err = perf_evlist__open(evlist);
3042         if (err < 0)
3043                 goto out_error_open;
3044
3045         err = bpf__apply_obj_config();
3046         if (err) {
3047                 char errbuf[BUFSIZ];
3048
3049                 bpf__strerror_apply_obj_config(err, errbuf, sizeof(errbuf));
3050                 pr_err("ERROR: Apply config to BPF failed: %s\n",
3051                          errbuf);
3052                 goto out_error_open;
3053         }
3054
3055         err = trace__set_filter_pids(trace);
3056         if (err < 0)
3057                 goto out_error_mem;
3058
3059         if (trace->syscalls.map)
3060                 trace__init_syscalls_bpf_map(trace);
3061
3062         if (trace->ev_qualifier_ids.nr > 0) {
3063                 err = trace__set_ev_qualifier_filter(trace);
3064                 if (err < 0)
3065                         goto out_errno;
3066
3067                 if (trace->syscalls.events.sys_exit) {
3068                         pr_debug("event qualifier tracepoint filter: %s\n",
3069                                  trace->syscalls.events.sys_exit->filter);
3070                 }
3071         }
3072
3073         err = perf_evlist__apply_filters(evlist, &evsel);
3074         if (err < 0)
3075                 goto out_error_apply_filters;
3076
3077         if (trace->dump.map)
3078                 bpf_map__fprintf(trace->dump.map, trace->output);
3079
3080         err = perf_evlist__mmap(evlist, trace->opts.mmap_pages);
3081         if (err < 0)
3082                 goto out_error_mmap;
3083
3084         if (!target__none(&trace->opts.target) && !trace->opts.initial_delay)
3085                 perf_evlist__enable(evlist);
3086
3087         if (forks)
3088                 perf_evlist__start_workload(evlist);
3089
3090         if (trace->opts.initial_delay) {
3091                 usleep(trace->opts.initial_delay * 1000);
3092                 perf_evlist__enable(evlist);
3093         }
3094
3095         trace->multiple_threads = thread_map__pid(evlist->threads, 0) == -1 ||
3096                                   evlist->threads->nr > 1 ||
3097                                   perf_evlist__first(evlist)->attr.inherit;
3098
3099         /*
3100          * Now that we already used evsel->attr to ask the kernel to setup the
3101          * events, lets reuse evsel->attr.sample_max_stack as the limit in
3102          * trace__resolve_callchain(), allowing per-event max-stack settings
3103          * to override an explicitly set --max-stack global setting.
3104          */
3105         evlist__for_each_entry(evlist, evsel) {
3106                 if (evsel__has_callchain(evsel) &&
3107                     evsel->attr.sample_max_stack == 0)
3108                         evsel->attr.sample_max_stack = trace->max_stack;
3109         }
3110 again:
3111         before = trace->nr_events;
3112
3113         for (i = 0; i < evlist->nr_mmaps; i++) {
3114                 union perf_event *event;
3115                 struct perf_mmap *md;
3116
3117                 md = &evlist->mmap[i];
3118                 if (perf_mmap__read_init(md) < 0)
3119                         continue;
3120
3121                 while ((event = perf_mmap__read_event(md)) != NULL) {
3122                         ++trace->nr_events;
3123
3124                         err = trace__deliver_event(trace, event);
3125                         if (err)
3126                                 goto out_disable;
3127
3128                         perf_mmap__consume(md);
3129
3130                         if (interrupted)
3131                                 goto out_disable;
3132
3133                         if (done && !draining) {
3134                                 perf_evlist__disable(evlist);
3135                                 draining = true;
3136                         }
3137                 }
3138                 perf_mmap__read_done(md);
3139         }
3140
3141         if (trace->nr_events == before) {
3142                 int timeout = done ? 100 : -1;
3143
3144                 if (!draining && perf_evlist__poll(evlist, timeout) > 0) {
3145                         if (perf_evlist__filter_pollfd(evlist, POLLERR | POLLHUP | POLLNVAL) == 0)
3146                                 draining = true;
3147
3148                         goto again;
3149                 } else {
3150                         if (trace__flush_events(trace))
3151                                 goto out_disable;
3152                 }
3153         } else {
3154                 goto again;
3155         }
3156
3157 out_disable:
3158         thread__zput(trace->current);
3159
3160         perf_evlist__disable(evlist);
3161
3162         if (trace->sort_events)
3163                 ordered_events__flush(&trace->oe.data, OE_FLUSH__FINAL);
3164
3165         if (!err) {
3166                 if (trace->summary)
3167                         trace__fprintf_thread_summary(trace, trace->output);
3168
3169                 if (trace->show_tool_stats) {
3170                         fprintf(trace->output, "Stats:\n "
3171                                                " vfs_getname : %" PRIu64 "\n"
3172                                                " proc_getname: %" PRIu64 "\n",
3173                                 trace->stats.vfs_getname,
3174                                 trace->stats.proc_getname);
3175                 }
3176         }
3177
3178 out_delete_evlist:
3179         trace__symbols__exit(trace);
3180
3181         perf_evlist__delete(evlist);
3182         cgroup__put(trace->cgroup);
3183         trace->evlist = NULL;
3184         trace->live = false;
3185         return err;
3186 {
3187         char errbuf[BUFSIZ];
3188
3189 out_error_sched_stat_runtime:
3190         tracing_path__strerror_open_tp(errno, errbuf, sizeof(errbuf), "sched", "sched_stat_runtime");
3191         goto out_error;
3192
3193 out_error_raw_syscalls:
3194         tracing_path__strerror_open_tp(errno, errbuf, sizeof(errbuf), "raw_syscalls", "sys_(enter|exit)");
3195         goto out_error;
3196
3197 out_error_mmap:
3198         perf_evlist__strerror_mmap(evlist, errno, errbuf, sizeof(errbuf));
3199         goto out_error;
3200
3201 out_error_open:
3202         perf_evlist__strerror_open(evlist, errno, errbuf, sizeof(errbuf));
3203
3204 out_error:
3205         fprintf(trace->output, "%s\n", errbuf);
3206         goto out_delete_evlist;
3207
3208 out_error_apply_filters:
3209         fprintf(trace->output,
3210                 "Failed to set filter \"%s\" on event %s with %d (%s)\n",
3211                 evsel->filter, perf_evsel__name(evsel), errno,
3212                 str_error_r(errno, errbuf, sizeof(errbuf)));
3213         goto out_delete_evlist;
3214 }
3215 out_error_mem:
3216         fprintf(trace->output, "Not enough memory to run!\n");
3217         goto out_delete_evlist;
3218
3219 out_errno:
3220         fprintf(trace->output, "errno=%d,%s\n", errno, strerror(errno));
3221         goto out_delete_evlist;
3222 }
3223
3224 static int trace__replay(struct trace *trace)
3225 {
3226         const struct perf_evsel_str_handler handlers[] = {
3227                 { "probe:vfs_getname",       trace__vfs_getname, },
3228         };
3229         struct perf_data data = {
3230                 .path  = input_name,
3231                 .mode  = PERF_DATA_MODE_READ,
3232                 .force = trace->force,
3233         };
3234         struct perf_session *session;
3235         struct perf_evsel *evsel;
3236         int err = -1;
3237
3238         trace->tool.sample        = trace__process_sample;
3239         trace->tool.mmap          = perf_event__process_mmap;
3240         trace->tool.mmap2         = perf_event__process_mmap2;
3241         trace->tool.comm          = perf_event__process_comm;
3242         trace->tool.exit          = perf_event__process_exit;
3243         trace->tool.fork          = perf_event__process_fork;
3244         trace->tool.attr          = perf_event__process_attr;
3245         trace->tool.tracing_data  = perf_event__process_tracing_data;
3246         trace->tool.build_id      = perf_event__process_build_id;
3247         trace->tool.namespaces    = perf_event__process_namespaces;
3248
3249         trace->tool.ordered_events = true;
3250         trace->tool.ordering_requires_timestamps = true;
3251
3252         /* add tid to output */
3253         trace->multiple_threads = true;
3254
3255         session = perf_session__new(&data, false, &trace->tool);
3256         if (session == NULL)
3257                 return -1;
3258
3259         if (trace->opts.target.pid)
3260                 symbol_conf.pid_list_str = strdup(trace->opts.target.pid);
3261
3262         if (trace->opts.target.tid)
3263                 symbol_conf.tid_list_str = strdup(trace->opts.target.tid);
3264
3265         if (symbol__init(&session->header.env) < 0)
3266                 goto out;
3267
3268         trace->host = &session->machines.host;
3269
3270         err = perf_session__set_tracepoints_handlers(session, handlers);
3271         if (err)
3272                 goto out;
3273
3274         evsel = perf_evlist__find_tracepoint_by_name(session->evlist,
3275                                                      "raw_syscalls:sys_enter");
3276         /* older kernels have syscalls tp versus raw_syscalls */
3277         if (evsel == NULL)
3278                 evsel = perf_evlist__find_tracepoint_by_name(session->evlist,
3279                                                              "syscalls:sys_enter");
3280
3281         if (evsel &&
3282             (perf_evsel__init_raw_syscall_tp(evsel, trace__sys_enter) < 0 ||
3283             perf_evsel__init_sc_tp_ptr_field(evsel, args))) {
3284                 pr_err("Error during initialize raw_syscalls:sys_enter event\n");
3285                 goto out;
3286         }
3287
3288         evsel = perf_evlist__find_tracepoint_by_name(session->evlist,
3289                                                      "raw_syscalls:sys_exit");
3290         if (evsel == NULL)
3291                 evsel = perf_evlist__find_tracepoint_by_name(session->evlist,
3292                                                              "syscalls:sys_exit");
3293         if (evsel &&
3294             (perf_evsel__init_raw_syscall_tp(evsel, trace__sys_exit) < 0 ||
3295             perf_evsel__init_sc_tp_uint_field(evsel, ret))) {
3296                 pr_err("Error during initialize raw_syscalls:sys_exit event\n");
3297                 goto out;
3298         }
3299
3300         evlist__for_each_entry(session->evlist, evsel) {
3301                 if (evsel->attr.type == PERF_TYPE_SOFTWARE &&
3302                     (evsel->attr.config == PERF_COUNT_SW_PAGE_FAULTS_MAJ ||
3303                      evsel->attr.config == PERF_COUNT_SW_PAGE_FAULTS_MIN ||
3304                      evsel->attr.config == PERF_COUNT_SW_PAGE_FAULTS))
3305                         evsel->handler = trace__pgfault;
3306         }
3307
3308         setup_pager();
3309
3310         err = perf_session__process_events(session);
3311         if (err)
3312                 pr_err("Failed to process events, error %d", err);
3313
3314         else if (trace->summary)
3315                 trace__fprintf_thread_summary(trace, trace->output);
3316
3317 out:
3318         perf_session__delete(session);
3319
3320         return err;
3321 }
3322
3323 static size_t trace__fprintf_threads_header(FILE *fp)
3324 {
3325         size_t printed;
3326
3327         printed  = fprintf(fp, "\n Summary of events:\n\n");
3328
3329         return printed;
3330 }
3331
3332 DEFINE_RESORT_RB(syscall_stats, a->msecs > b->msecs,
3333         struct stats    *stats;
3334         double          msecs;
3335         int             syscall;
3336 )
3337 {
3338         struct int_node *source = rb_entry(nd, struct int_node, rb_node);
3339         struct stats *stats = source->priv;
3340
3341         entry->syscall = source->i;
3342         entry->stats   = stats;
3343         entry->msecs   = stats ? (u64)stats->n * (avg_stats(stats) / NSEC_PER_MSEC) : 0;
3344 }
3345
3346 static size_t thread__dump_stats(struct thread_trace *ttrace,
3347                                  struct trace *trace, FILE *fp)
3348 {
3349         size_t printed = 0;
3350         struct syscall *sc;
3351         struct rb_node *nd;
3352         DECLARE_RESORT_RB_INTLIST(syscall_stats, ttrace->syscall_stats);
3353
3354         if (syscall_stats == NULL)
3355                 return 0;
3356
3357         printed += fprintf(fp, "\n");
3358
3359         printed += fprintf(fp, "   syscall            calls    total       min       avg       max      stddev\n");
3360         printed += fprintf(fp, "                               (msec)    (msec)    (msec)    (msec)        (%%)\n");
3361         printed += fprintf(fp, "   --------------- -------- --------- --------- --------- ---------     ------\n");
3362
3363         resort_rb__for_each_entry(nd, syscall_stats) {
3364                 struct stats *stats = syscall_stats_entry->stats;
3365                 if (stats) {
3366                         double min = (double)(stats->min) / NSEC_PER_MSEC;
3367                         double max = (double)(stats->max) / NSEC_PER_MSEC;
3368                         double avg = avg_stats(stats);
3369                         double pct;
3370                         u64 n = (u64) stats->n;
3371
3372                         pct = avg ? 100.0 * stddev_stats(stats)/avg : 0.0;
3373                         avg /= NSEC_PER_MSEC;
3374
3375                         sc = &trace->syscalls.table[syscall_stats_entry->syscall];
3376                         printed += fprintf(fp, "   %-15s", sc->name);
3377                         printed += fprintf(fp, " %8" PRIu64 " %9.3f %9.3f %9.3f",
3378                                            n, syscall_stats_entry->msecs, min, avg);
3379                         printed += fprintf(fp, " %9.3f %9.2f%%\n", max, pct);
3380                 }
3381         }
3382
3383         resort_rb__delete(syscall_stats);
3384         printed += fprintf(fp, "\n\n");
3385
3386         return printed;
3387 }
3388
3389 static size_t trace__fprintf_thread(FILE *fp, struct thread *thread, struct trace *trace)
3390 {
3391         size_t printed = 0;
3392         struct thread_trace *ttrace = thread__priv(thread);
3393         double ratio;
3394
3395         if (ttrace == NULL)
3396                 return 0;
3397
3398         ratio = (double)ttrace->nr_events / trace->nr_events * 100.0;
3399
3400         printed += fprintf(fp, " %s (%d), ", thread__comm_str(thread), thread->tid);
3401         printed += fprintf(fp, "%lu events, ", ttrace->nr_events);
3402         printed += fprintf(fp, "%.1f%%", ratio);
3403         if (ttrace->pfmaj)
3404                 printed += fprintf(fp, ", %lu majfaults", ttrace->pfmaj);
3405         if (ttrace->pfmin)
3406                 printed += fprintf(fp, ", %lu minfaults", ttrace->pfmin);
3407         if (trace->sched)
3408                 printed += fprintf(fp, ", %.3f msec\n", ttrace->runtime_ms);
3409         else if (fputc('\n', fp) != EOF)
3410                 ++printed;
3411
3412         printed += thread__dump_stats(ttrace, trace, fp);
3413
3414         return printed;
3415 }
3416
3417 static unsigned long thread__nr_events(struct thread_trace *ttrace)
3418 {
3419         return ttrace ? ttrace->nr_events : 0;
3420 }
3421
3422 DEFINE_RESORT_RB(threads, (thread__nr_events(a->thread->priv) < thread__nr_events(b->thread->priv)),
3423         struct thread *thread;
3424 )
3425 {
3426         entry->thread = rb_entry(nd, struct thread, rb_node);
3427 }
3428
3429 static size_t trace__fprintf_thread_summary(struct trace *trace, FILE *fp)
3430 {
3431         size_t printed = trace__fprintf_threads_header(fp);
3432         struct rb_node *nd;
3433         int i;
3434
3435         for (i = 0; i < THREADS__TABLE_SIZE; i++) {
3436                 DECLARE_RESORT_RB_MACHINE_THREADS(threads, trace->host, i);
3437
3438                 if (threads == NULL) {
3439                         fprintf(fp, "%s", "Error sorting output by nr_events!\n");
3440                         return 0;
3441                 }
3442
3443                 resort_rb__for_each_entry(nd, threads)
3444                         printed += trace__fprintf_thread(fp, threads_entry->thread, trace);
3445
3446                 resort_rb__delete(threads);
3447         }
3448         return printed;
3449 }
3450
3451 static int trace__set_duration(const struct option *opt, const char *str,
3452                                int unset __maybe_unused)
3453 {
3454         struct trace *trace = opt->value;
3455
3456         trace->duration_filter = atof(str);
3457         return 0;
3458 }
3459
3460 static int trace__set_filter_pids_from_option(const struct option *opt, const char *str,
3461                                               int unset __maybe_unused)
3462 {
3463         int ret = -1;
3464         size_t i;
3465         struct trace *trace = opt->value;
3466         /*
3467          * FIXME: introduce a intarray class, plain parse csv and create a
3468          * { int nr, int entries[] } struct...
3469          */
3470         struct intlist *list = intlist__new(str);
3471
3472         if (list == NULL)
3473                 return -1;
3474
3475         i = trace->filter_pids.nr = intlist__nr_entries(list) + 1;
3476         trace->filter_pids.entries = calloc(i, sizeof(pid_t));
3477
3478         if (trace->filter_pids.entries == NULL)
3479                 goto out;
3480
3481         trace->filter_pids.entries[0] = getpid();
3482
3483         for (i = 1; i < trace->filter_pids.nr; ++i)
3484                 trace->filter_pids.entries[i] = intlist__entry(list, i - 1)->i;
3485
3486         intlist__delete(list);
3487         ret = 0;
3488 out:
3489         return ret;
3490 }
3491
3492 static int trace__open_output(struct trace *trace, const char *filename)
3493 {
3494         struct stat st;
3495
3496         if (!stat(filename, &st) && st.st_size) {
3497                 char oldname[PATH_MAX];
3498
3499                 scnprintf(oldname, sizeof(oldname), "%s.old", filename);
3500                 unlink(oldname);
3501                 rename(filename, oldname);
3502         }
3503
3504         trace->output = fopen(filename, "w");
3505
3506         return trace->output == NULL ? -errno : 0;
3507 }
3508
3509 static int parse_pagefaults(const struct option *opt, const char *str,
3510                             int unset __maybe_unused)
3511 {
3512         int *trace_pgfaults = opt->value;
3513
3514         if (strcmp(str, "all") == 0)
3515                 *trace_pgfaults |= TRACE_PFMAJ | TRACE_PFMIN;
3516         else if (strcmp(str, "maj") == 0)
3517                 *trace_pgfaults |= TRACE_PFMAJ;
3518         else if (strcmp(str, "min") == 0)
3519                 *trace_pgfaults |= TRACE_PFMIN;
3520         else
3521                 return -1;
3522
3523         return 0;
3524 }
3525
3526 static void evlist__set_evsel_handler(struct perf_evlist *evlist, void *handler)
3527 {
3528         struct perf_evsel *evsel;
3529
3530         evlist__for_each_entry(evlist, evsel)
3531                 evsel->handler = handler;
3532 }
3533
3534 static int evlist__set_syscall_tp_fields(struct perf_evlist *evlist)
3535 {
3536         struct perf_evsel *evsel;
3537
3538         evlist__for_each_entry(evlist, evsel) {
3539                 if (evsel->priv || !evsel->tp_format)
3540                         continue;
3541
3542                 if (strcmp(evsel->tp_format->system, "syscalls"))
3543                         continue;
3544
3545                 if (perf_evsel__init_syscall_tp(evsel))
3546                         return -1;
3547
3548                 if (!strncmp(evsel->tp_format->name, "sys_enter_", 10)) {
3549                         struct syscall_tp *sc = evsel->priv;
3550
3551                         if (__tp_field__init_ptr(&sc->args, sc->id.offset + sizeof(u64)))
3552                                 return -1;
3553                 } else if (!strncmp(evsel->tp_format->name, "sys_exit_", 9)) {
3554                         struct syscall_tp *sc = evsel->priv;
3555
3556                         if (__tp_field__init_uint(&sc->ret, sizeof(u64), sc->id.offset + sizeof(u64), evsel->needs_swap))
3557                                 return -1;
3558                 }
3559         }
3560
3561         return 0;
3562 }
3563
3564 /*
3565  * XXX: Hackish, just splitting the combined -e+--event (syscalls
3566  * (raw_syscalls:{sys_{enter,exit}} + events (tracepoints, HW, SW, etc) to use
3567  * existing facilities unchanged (trace->ev_qualifier + parse_options()).
3568  *
3569  * It'd be better to introduce a parse_options() variant that would return a
3570  * list with the terms it didn't match to an event...
3571  */
3572 static int trace__parse_events_option(const struct option *opt, const char *str,
3573                                       int unset __maybe_unused)
3574 {
3575         struct trace *trace = (struct trace *)opt->value;
3576         const char *s = str;
3577         char *sep = NULL, *lists[2] = { NULL, NULL, };
3578         int len = strlen(str) + 1, err = -1, list, idx;
3579         char *strace_groups_dir = system_path(STRACE_GROUPS_DIR);
3580         char group_name[PATH_MAX];
3581         struct syscall_fmt *fmt;
3582
3583         if (strace_groups_dir == NULL)
3584                 return -1;
3585
3586         if (*s == '!') {
3587                 ++s;
3588                 trace->not_ev_qualifier = true;
3589         }
3590
3591         while (1) {
3592                 if ((sep = strchr(s, ',')) != NULL)
3593                         *sep = '\0';
3594
3595                 list = 0;
3596                 if (syscalltbl__id(trace->sctbl, s) >= 0 ||
3597                     syscalltbl__strglobmatch_first(trace->sctbl, s, &idx) >= 0) {
3598                         list = 1;
3599                         goto do_concat;
3600                 }
3601
3602                 fmt = syscall_fmt__find_by_alias(s);
3603                 if (fmt != NULL) {
3604                         list = 1;
3605                         s = fmt->name;
3606                 } else {
3607                         path__join(group_name, sizeof(group_name), strace_groups_dir, s);
3608                         if (access(group_name, R_OK) == 0)
3609                                 list = 1;
3610                 }
3611 do_concat:
3612                 if (lists[list]) {
3613                         sprintf(lists[list] + strlen(lists[list]), ",%s", s);
3614                 } else {
3615                         lists[list] = malloc(len);
3616                         if (lists[list] == NULL)
3617                                 goto out;
3618                         strcpy(lists[list], s);
3619                 }
3620
3621                 if (!sep)
3622                         break;
3623
3624                 *sep = ',';
3625                 s = sep + 1;
3626         }
3627
3628         if (lists[1] != NULL) {
3629                 struct strlist_config slist_config = {
3630                         .dirname = strace_groups_dir,
3631                 };
3632
3633                 trace->ev_qualifier = strlist__new(lists[1], &slist_config);
3634                 if (trace->ev_qualifier == NULL) {
3635                         fputs("Not enough memory to parse event qualifier", trace->output);
3636                         goto out;
3637                 }
3638
3639                 if (trace__validate_ev_qualifier(trace))
3640                         goto out;
3641                 trace->trace_syscalls = true;
3642         }
3643
3644         err = 0;
3645
3646         if (lists[0]) {
3647                 struct option o = OPT_CALLBACK('e', "event", &trace->evlist, "event",
3648                                                "event selector. use 'perf list' to list available events",
3649                                                parse_events_option);
3650                 err = parse_events_option(&o, lists[0], 0);
3651         }
3652 out:
3653         if (sep)
3654                 *sep = ',';
3655
3656         return err;
3657 }
3658
3659 static int trace__parse_cgroups(const struct option *opt, const char *str, int unset)
3660 {
3661         struct trace *trace = opt->value;
3662
3663         if (!list_empty(&trace->evlist->entries))
3664                 return parse_cgroups(opt, str, unset);
3665
3666         trace->cgroup = evlist__findnew_cgroup(trace->evlist, str);
3667
3668         return 0;
3669 }
3670
3671 static struct bpf_map *bpf__find_map_by_name(const char *name)
3672 {
3673         struct bpf_object *obj, *tmp;
3674
3675         bpf_object__for_each_safe(obj, tmp) {
3676                 struct bpf_map *map = bpf_object__find_map_by_name(obj, name);
3677                 if (map)
3678                         return map;
3679
3680         }
3681
3682         return NULL;
3683 }
3684
3685 static void trace__set_bpf_map_filtered_pids(struct trace *trace)
3686 {
3687         trace->filter_pids.map = bpf__find_map_by_name("pids_filtered");
3688 }
3689
3690 static void trace__set_bpf_map_syscalls(struct trace *trace)
3691 {
3692         trace->syscalls.map = bpf__find_map_by_name("syscalls");
3693 }
3694
3695 static int trace__config(const char *var, const char *value, void *arg)
3696 {
3697         struct trace *trace = arg;
3698         int err = 0;
3699
3700         if (!strcmp(var, "trace.add_events")) {
3701                 struct option o = OPT_CALLBACK('e', "event", &trace->evlist, "event",
3702                                                "event selector. use 'perf list' to list available events",
3703                                                parse_events_option);
3704                 /*
3705                  * We can't propagate parse_event_option() return, as it is 1
3706                  * for failure while perf_config() expects -1.
3707                  */
3708                 if (parse_events_option(&o, value, 0))
3709                         err = -1;
3710         } else if (!strcmp(var, "trace.show_timestamp")) {
3711                 trace->show_tstamp = perf_config_bool(var, value);
3712         } else if (!strcmp(var, "trace.show_duration")) {
3713                 trace->show_duration = perf_config_bool(var, value);
3714         } else if (!strcmp(var, "trace.show_arg_names")) {
3715                 trace->show_arg_names = perf_config_bool(var, value);
3716                 if (!trace->show_arg_names)
3717                         trace->show_zeros = true;
3718         } else if (!strcmp(var, "trace.show_zeros")) {
3719                 bool new_show_zeros = perf_config_bool(var, value);
3720                 if (!trace->show_arg_names && !new_show_zeros) {
3721                         pr_warning("trace.show_zeros has to be set when trace.show_arg_names=no\n");
3722                         goto out;
3723                 }
3724                 trace->show_zeros = new_show_zeros;
3725         } else if (!strcmp(var, "trace.show_prefix")) {
3726                 trace->show_string_prefix = perf_config_bool(var, value);
3727         } else if (!strcmp(var, "trace.no_inherit")) {
3728                 trace->opts.no_inherit = perf_config_bool(var, value);
3729         } else if (!strcmp(var, "trace.args_alignment")) {
3730                 int args_alignment = 0;
3731                 if (perf_config_int(&args_alignment, var, value) == 0)
3732                         trace->args_alignment = args_alignment;
3733         }
3734 out:
3735         return err;
3736 }
3737
3738 int cmd_trace(int argc, const char **argv)
3739 {
3740         const char *trace_usage[] = {
3741                 "perf trace [<options>] [<command>]",
3742                 "perf trace [<options>] -- <command> [<options>]",
3743                 "perf trace record [<options>] [<command>]",
3744                 "perf trace record [<options>] -- <command> [<options>]",
3745                 NULL
3746         };
3747         struct trace trace = {
3748                 .syscalls = {
3749                         . max = -1,
3750                 },
3751                 .opts = {
3752                         .target = {
3753                                 .uid       = UINT_MAX,
3754                                 .uses_mmap = true,
3755                         },
3756                         .user_freq     = UINT_MAX,
3757                         .user_interval = ULLONG_MAX,
3758                         .no_buffering  = true,
3759                         .mmap_pages    = UINT_MAX,
3760                 },
3761                 .output = stderr,
3762                 .show_comm = true,
3763                 .show_tstamp = true,
3764                 .show_duration = true,
3765                 .show_arg_names = true,
3766                 .args_alignment = 70,
3767                 .trace_syscalls = false,
3768                 .kernel_syscallchains = false,
3769                 .max_stack = UINT_MAX,
3770                 .max_events = ULONG_MAX,
3771         };
3772         const char *map_dump_str = NULL;
3773         const char *output_name = NULL;
3774         const struct option trace_options[] = {
3775         OPT_CALLBACK('e', "event", &trace, "event",
3776                      "event/syscall selector. use 'perf list' to list available events",
3777                      trace__parse_events_option),
3778         OPT_BOOLEAN(0, "comm", &trace.show_comm,
3779                     "show the thread COMM next to its id"),
3780         OPT_BOOLEAN(0, "tool_stats", &trace.show_tool_stats, "show tool stats"),
3781         OPT_CALLBACK(0, "expr", &trace, "expr", "list of syscalls/events to trace",
3782                      trace__parse_events_option),
3783         OPT_STRING('o', "output", &output_name, "file", "output file name"),
3784         OPT_STRING('i', "input", &input_name, "file", "Analyze events in file"),
3785         OPT_STRING('p', "pid", &trace.opts.target.pid, "pid",
3786                     "trace events on existing process id"),
3787         OPT_STRING('t', "tid", &trace.opts.target.tid, "tid",
3788                     "trace events on existing thread id"),
3789         OPT_CALLBACK(0, "filter-pids", &trace, "CSV list of pids",
3790                      "pids to filter (by the kernel)", trace__set_filter_pids_from_option),
3791         OPT_BOOLEAN('a', "all-cpus", &trace.opts.target.system_wide,
3792                     "system-wide collection from all CPUs"),
3793         OPT_STRING('C', "cpu", &trace.opts.target.cpu_list, "cpu",
3794                     "list of cpus to monitor"),
3795         OPT_BOOLEAN(0, "no-inherit", &trace.opts.no_inherit,
3796                     "child tasks do not inherit counters"),
3797         OPT_CALLBACK('m', "mmap-pages", &trace.opts.mmap_pages, "pages",
3798                      "number of mmap data pages",
3799                      perf_evlist__parse_mmap_pages),
3800         OPT_STRING('u', "uid", &trace.opts.target.uid_str, "user",
3801                    "user to profile"),
3802         OPT_CALLBACK(0, "duration", &trace, "float",
3803                      "show only events with duration > N.M ms",
3804                      trace__set_duration),
3805 #ifdef HAVE_LIBBPF_SUPPORT
3806         OPT_STRING(0, "map-dump", &map_dump_str, "BPF map", "BPF map to periodically dump"),
3807 #endif
3808         OPT_BOOLEAN(0, "sched", &trace.sched, "show blocking scheduler events"),
3809         OPT_INCR('v', "verbose", &verbose, "be more verbose"),
3810         OPT_BOOLEAN('T', "time", &trace.full_time,
3811                     "Show full timestamp, not time relative to first start"),
3812         OPT_BOOLEAN(0, "failure", &trace.failure_only,
3813                     "Show only syscalls that failed"),
3814         OPT_BOOLEAN('s', "summary", &trace.summary_only,
3815                     "Show only syscall summary with statistics"),
3816         OPT_BOOLEAN('S', "with-summary", &trace.summary,
3817                     "Show all syscalls and summary with statistics"),
3818         OPT_CALLBACK_DEFAULT('F', "pf", &trace.trace_pgfaults, "all|maj|min",
3819                      "Trace pagefaults", parse_pagefaults, "maj"),
3820         OPT_BOOLEAN(0, "syscalls", &trace.trace_syscalls, "Trace syscalls"),
3821         OPT_BOOLEAN('f', "force", &trace.force, "don't complain, do it"),
3822         OPT_CALLBACK(0, "call-graph", &trace.opts,
3823                      "record_mode[,record_size]", record_callchain_help,
3824                      &record_parse_callchain_opt),
3825         OPT_BOOLEAN(0, "kernel-syscall-graph", &trace.kernel_syscallchains,
3826                     "Show the kernel callchains on the syscall exit path"),
3827         OPT_ULONG(0, "max-events", &trace.max_events,
3828                 "Set the maximum number of events to print, exit after that is reached. "),
3829         OPT_UINTEGER(0, "min-stack", &trace.min_stack,
3830                      "Set the minimum stack depth when parsing the callchain, "
3831                      "anything below the specified depth will be ignored."),
3832         OPT_UINTEGER(0, "max-stack", &trace.max_stack,
3833                      "Set the maximum stack depth when parsing the callchain, "
3834                      "anything beyond the specified depth will be ignored. "
3835                      "Default: kernel.perf_event_max_stack or " __stringify(PERF_MAX_STACK_DEPTH)),
3836         OPT_BOOLEAN(0, "sort-events", &trace.sort_events,
3837                         "Sort batch of events before processing, use if getting out of order events"),
3838         OPT_BOOLEAN(0, "print-sample", &trace.print_sample,
3839                         "print the PERF_RECORD_SAMPLE PERF_SAMPLE_ info, for debugging"),
3840         OPT_UINTEGER(0, "proc-map-timeout", &proc_map_timeout,
3841                         "per thread proc mmap processing timeout in ms"),
3842         OPT_CALLBACK('G', "cgroup", &trace, "name", "monitor event in cgroup name only",
3843                      trace__parse_cgroups),
3844         OPT_UINTEGER('D', "delay", &trace.opts.initial_delay,
3845                      "ms to wait before starting measurement after program "
3846                      "start"),
3847         OPT_END()
3848         };
3849         bool __maybe_unused max_stack_user_set = true;
3850         bool mmap_pages_user_set = true;
3851         struct perf_evsel *evsel;
3852         const char * const trace_subcommands[] = { "record", NULL };
3853         int err = -1;
3854         char bf[BUFSIZ];
3855
3856         signal(SIGSEGV, sighandler_dump_stack);
3857         signal(SIGFPE, sighandler_dump_stack);
3858
3859         trace.evlist = perf_evlist__new();
3860         trace.sctbl = syscalltbl__new();
3861
3862         if (trace.evlist == NULL || trace.sctbl == NULL) {
3863                 pr_err("Not enough memory to run!\n");
3864                 err = -ENOMEM;
3865                 goto out;
3866         }
3867
3868         /*
3869          * Parsing .perfconfig may entail creating a BPF event, that may need
3870          * to create BPF maps, so bump RLIM_MEMLOCK as the default 64K setting
3871          * is too small. This affects just this process, not touching the
3872          * global setting. If it fails we'll get something in 'perf trace -v'
3873          * to help diagnose the problem.
3874          */
3875         rlimit__bump_memlock();
3876
3877         err = perf_config(trace__config, &trace);
3878         if (err)
3879                 goto out;
3880
3881         argc = parse_options_subcommand(argc, argv, trace_options, trace_subcommands,
3882                                  trace_usage, PARSE_OPT_STOP_AT_NON_OPTION);
3883
3884         if ((nr_cgroups || trace.cgroup) && !trace.opts.target.system_wide) {
3885                 usage_with_options_msg(trace_usage, trace_options,
3886                                        "cgroup monitoring only available in system-wide mode");
3887         }
3888
3889         evsel = bpf__setup_output_event(trace.evlist, "__augmented_syscalls__");
3890         if (IS_ERR(evsel)) {
3891                 bpf__strerror_setup_output_event(trace.evlist, PTR_ERR(evsel), bf, sizeof(bf));
3892                 pr_err("ERROR: Setup trace syscalls enter failed: %s\n", bf);
3893                 goto out;
3894         }
3895
3896         if (evsel) {
3897                 trace.syscalls.events.augmented = evsel;
3898                 trace__set_bpf_map_filtered_pids(&trace);
3899                 trace__set_bpf_map_syscalls(&trace);
3900         }
3901
3902         err = bpf__setup_stdout(trace.evlist);
3903         if (err) {
3904                 bpf__strerror_setup_stdout(trace.evlist, err, bf, sizeof(bf));
3905                 pr_err("ERROR: Setup BPF stdout failed: %s\n", bf);
3906                 goto out;
3907         }
3908
3909         err = -1;
3910
3911         if (map_dump_str) {
3912                 trace.dump.map = bpf__find_map_by_name(map_dump_str);
3913                 if (trace.dump.map == NULL) {
3914                         pr_err("ERROR: BPF map \"%s\" not found\n", map_dump_str);
3915                         goto out;
3916                 }
3917         }
3918
3919         if (trace.trace_pgfaults) {
3920                 trace.opts.sample_address = true;
3921                 trace.opts.sample_time = true;
3922         }
3923
3924         if (trace.opts.mmap_pages == UINT_MAX)
3925                 mmap_pages_user_set = false;
3926
3927         if (trace.max_stack == UINT_MAX) {
3928                 trace.max_stack = input_name ? PERF_MAX_STACK_DEPTH : sysctl__max_stack();
3929                 max_stack_user_set = false;
3930         }
3931
3932 #ifdef HAVE_DWARF_UNWIND_SUPPORT
3933         if ((trace.min_stack || max_stack_user_set) && !callchain_param.enabled) {
3934                 record_opts__parse_callchain(&trace.opts, &callchain_param, "dwarf", false);
3935         }
3936 #endif
3937
3938         if (callchain_param.enabled) {
3939                 if (!mmap_pages_user_set && geteuid() == 0)
3940                         trace.opts.mmap_pages = perf_event_mlock_kb_in_pages() * 4;
3941
3942                 symbol_conf.use_callchain = true;
3943         }
3944
3945         if (trace.evlist->nr_entries > 0) {
3946                 evlist__set_evsel_handler(trace.evlist, trace__event_handler);
3947                 if (evlist__set_syscall_tp_fields(trace.evlist)) {
3948                         perror("failed to set syscalls:* tracepoint fields");
3949                         goto out;
3950                 }
3951         }
3952
3953         if (trace.sort_events) {
3954                 ordered_events__init(&trace.oe.data, ordered_events__deliver_event, &trace);
3955                 ordered_events__set_copy_on_queue(&trace.oe.data, true);
3956         }
3957
3958         /*
3959          * If we are augmenting syscalls, then combine what we put in the
3960          * __augmented_syscalls__ BPF map with what is in the
3961          * syscalls:sys_exit_FOO tracepoints, i.e. just like we do without BPF,
3962          * combining raw_syscalls:sys_enter with raw_syscalls:sys_exit.
3963          *
3964          * We'll switch to look at two BPF maps, one for sys_enter and the
3965          * other for sys_exit when we start augmenting the sys_exit paths with
3966          * buffers that are being copied from kernel to userspace, think 'read'
3967          * syscall.
3968          */
3969         if (trace.syscalls.events.augmented) {
3970                 evlist__for_each_entry(trace.evlist, evsel) {
3971                         bool raw_syscalls_sys_exit = strcmp(perf_evsel__name(evsel), "raw_syscalls:sys_exit") == 0;
3972
3973                         if (raw_syscalls_sys_exit) {
3974                                 trace.raw_augmented_syscalls = true;
3975                                 goto init_augmented_syscall_tp;
3976                         }
3977
3978                         if (trace.syscalls.events.augmented->priv == NULL &&
3979                             strstr(perf_evsel__name(evsel), "syscalls:sys_enter")) {
3980                                 struct perf_evsel *augmented = trace.syscalls.events.augmented;
3981                                 if (perf_evsel__init_augmented_syscall_tp(augmented, evsel) ||
3982                                     perf_evsel__init_augmented_syscall_tp_args(augmented))
3983                                         goto out;
3984                                 augmented->handler = trace__sys_enter;
3985                         }
3986
3987                         if (strstarts(perf_evsel__name(evsel), "syscalls:sys_exit_")) {
3988                                 struct syscall_tp *sc;
3989 init_augmented_syscall_tp:
3990                                 if (perf_evsel__init_augmented_syscall_tp(evsel, evsel))
3991                                         goto out;
3992                                 sc = evsel->priv;
3993                                 /*
3994                                  * For now with BPF raw_augmented we hook into
3995                                  * raw_syscalls:sys_enter and there we get all
3996                                  * 6 syscall args plus the tracepoint common
3997                                  * fields and the syscall_nr (another long).
3998                                  * So we check if that is the case and if so
3999                                  * don't look after the sc->args_size but
4000                                  * always after the full raw_syscalls:sys_enter
4001                                  * payload, which is fixed.
4002                                  *
4003                                  * We'll revisit this later to pass
4004                                  * s->args_size to the BPF augmenter (now
4005                                  * tools/perf/examples/bpf/augmented_raw_syscalls.c,
4006                                  * so that it copies only what we need for each
4007                                  * syscall, like what happens when we use
4008                                  * syscalls:sys_enter_NAME, so that we reduce
4009                                  * the kernel/userspace traffic to just what is
4010                                  * needed for each syscall.
4011                                  */
4012                                 if (trace.raw_augmented_syscalls)
4013                                         trace.raw_augmented_syscalls_args_size = (6 + 1) * sizeof(long) + sc->id.offset;
4014                                 perf_evsel__init_augmented_syscall_tp_ret(evsel);
4015                                 evsel->handler = trace__sys_exit;
4016                         }
4017                 }
4018         }
4019
4020         if ((argc >= 1) && (strcmp(argv[0], "record") == 0))
4021                 return trace__record(&trace, argc-1, &argv[1]);
4022
4023         /* summary_only implies summary option, but don't overwrite summary if set */
4024         if (trace.summary_only)
4025                 trace.summary = trace.summary_only;
4026
4027         if (!trace.trace_syscalls && !trace.trace_pgfaults &&
4028             trace.evlist->nr_entries == 0 /* Was --events used? */) {
4029                 trace.trace_syscalls = true;
4030         }
4031
4032         if (output_name != NULL) {
4033                 err = trace__open_output(&trace, output_name);
4034                 if (err < 0) {
4035                         perror("failed to create output file");
4036                         goto out;
4037                 }
4038         }
4039
4040         err = target__validate(&trace.opts.target);
4041         if (err) {
4042                 target__strerror(&trace.opts.target, err, bf, sizeof(bf));
4043                 fprintf(trace.output, "%s", bf);
4044                 goto out_close;
4045         }
4046
4047         err = target__parse_uid(&trace.opts.target);
4048         if (err) {
4049                 target__strerror(&trace.opts.target, err, bf, sizeof(bf));
4050                 fprintf(trace.output, "%s", bf);
4051                 goto out_close;
4052         }
4053
4054         if (!argc && target__none(&trace.opts.target))
4055                 trace.opts.target.system_wide = true;
4056
4057         if (input_name)
4058                 err = trace__replay(&trace);
4059         else
4060                 err = trace__run(&trace, argc, argv);
4061
4062 out_close:
4063         if (output_name != NULL)
4064                 fclose(trace.output);
4065 out:
4066         return err;
4067 }