tools/perf/builtin-trace.c

   1 /*
   2  * builtin-trace.c
   3  *
   4  * Builtin 'trace' command:
   5  *
   6  * Display a continuously updated trace of any workload, CPU, specific PID,
   7  * system wide, etc.  Default format is loosely strace like, but any other
   8  * event may be specified using --event.
   9  *
  10  * Copyright (C) 2012, 2013, 2014, 2015 Red Hat Inc, Arnaldo Carvalho de Melo <acme@redhat.com>
  11  *
  12  * Initially based on the 'trace' prototype by Thomas Gleixner:
  13  *
  14  * http://lwn.net/Articles/415728/ ("Announcing a new utility: 'trace'")
  15  *
  16  * Released under the GPL v2. (and only v2, not any later version)
  17  */
  18
  19 #include <traceevent/event-parse.h>
  20 #include <api/fs/tracing_path.h>
  21 #include <bpf/bpf.h>
  22 #include "builtin.h"
  23 #include "util/cgroup.h"
  24 #include "util/color.h"
  25 #include "util/config.h"
  26 #include "util/debug.h"
  27 #include "util/env.h"
  28 #include "util/event.h"
  29 #include "util/evlist.h"
  30 #include <subcmd/exec-cmd.h>
  31 #include "util/machine.h"
  32 #include "util/map.h"
  33 #include "util/symbol.h"
  34 #include "util/path.h"
  35 #include "util/session.h"
  36 #include "util/thread.h"
  37 #include <subcmd/parse-options.h>
  38 #include "util/strlist.h"
  39 #include "util/intlist.h"
  40 #include "util/thread_map.h"
  41 #include "util/stat.h"
  42 #include "trace/beauty/beauty.h"
  43 #include "trace-event.h"
  44 #include "util/parse-events.h"
  45 #include "util/bpf-loader.h"
  46 #include "callchain.h"
  47 #include "print_binary.h"
  48 #include "string2.h"
  49 #include "syscalltbl.h"
  50 #include "rb_resort.h"
  51
  52 #include <errno.h>
  53 #include <inttypes.h>
  54 #include <poll.h>
  55 #include <signal.h>
  56 #include <stdlib.h>
  57 #include <string.h>
  58 #include <linux/err.h>
  59 #include <linux/filter.h>
  60 #include <linux/kernel.h>
  61 #include <linux/random.h>
  62 #include <linux/stringify.h>
  63 #include <linux/time64.h>
  64 #include <fcntl.h>
  65 #include <sys/sysmacros.h>
  66
  67 #include "sane_ctype.h"
  68
  69 #ifndef O_CLOEXEC
  70 # define O_CLOEXEC              02000000
  71 #endif
  72
  73 #ifndef F_LINUX_SPECIFIC_BASE
  74 # define F_LINUX_SPECIFIC_BASE  1024
  75 #endif
  76
  77 struct trace {
  78         struct perf_tool        tool;
  79         struct syscalltbl       *sctbl;
  80         struct {
  81                 int             max;
  82                 struct syscall  *table;
  83                 struct bpf_map  *map;
  84                 struct {
  85                         struct perf_evsel *sys_enter,
  86                                           *sys_exit,
  87                                           *augmented;
  88                 }               events;
  89         } syscalls;
  90         struct record_opts      opts;
  91         struct perf_evlist      *evlist;
  92         struct machine          *host;
  93         struct thread           *current;
  94         struct cgroup           *cgroup;
  95         u64                     base_time;
  96         FILE                    *output;
  97         unsigned long           nr_events;
  98         unsigned long           nr_events_printed;
  99         unsigned long           max_events;
 100         struct strlist          *ev_qualifier;
 101         struct {
 102                 size_t          nr;
 103                 int             *entries;
 104         }                       ev_qualifier_ids;
 105         struct {
 106                 size_t          nr;
 107                 pid_t           *entries;
 108                 struct bpf_map  *map;
 109         }                       filter_pids;
 110         double                  duration_filter;
 111         double                  runtime_ms;
 112         struct {
 113                 u64             vfs_getname,
 114                                 proc_getname;
 115         } stats;
 116         unsigned int            max_stack;
 117         unsigned int            min_stack;
 118         int                     raw_augmented_syscalls_args_size;
 119         bool                    raw_augmented_syscalls;
 120         bool                    sort_events;
 121         bool                    not_ev_qualifier;
 122         bool                    live;
 123         bool                    full_time;
 124         bool                    sched;
 125         bool                    multiple_threads;
 126         bool                    summary;
 127         bool                    summary_only;
 128         bool                    failure_only;
 129         bool                    show_comm;
 130         bool                    print_sample;
 131         bool                    show_tool_stats;
 132         bool                    trace_syscalls;
 133         bool                    kernel_syscallchains;
 134         s16                     args_alignment;
 135         bool                    show_tstamp;
 136         bool                    show_duration;
 137         bool                    show_zeros;
 138         bool                    show_arg_names;
 139         bool                    show_string_prefix;
 140         bool                    force;
 141         bool                    vfs_getname;
 142         int                     trace_pgfaults;
 143         struct {
 144                 struct ordered_events   data;
 145                 u64                     last;
 146         } oe;
 147 };
 148
 149 struct tp_field {
 150         int offset;
 151         union {
 152                 u64 (*integer)(struct tp_field *field, struct perf_sample *sample);
 153                 void *(*pointer)(struct tp_field *field, struct perf_sample *sample);
 154         };
 155 };
 156
 157 #define TP_UINT_FIELD(bits) \
 158 static u64 tp_field__u##bits(struct tp_field *field, struct perf_sample *sample) \
 159 { \
 160         u##bits value; \
 161         memcpy(&value, sample->raw_data + field->offset, sizeof(value)); \
 162         return value;  \
 163 }
 164
 165 TP_UINT_FIELD(8);
 166 TP_UINT_FIELD(16);
 167 TP_UINT_FIELD(32);
 168 TP_UINT_FIELD(64);
 169
 170 #define TP_UINT_FIELD__SWAPPED(bits) \
 171 static u64 tp_field__swapped_u##bits(struct tp_field *field, struct perf_sample *sample) \
 172 { \
 173         u##bits value; \
 174         memcpy(&value, sample->raw_data + field->offset, sizeof(value)); \
 175         return bswap_##bits(value);\
 176 }
 177
 178 TP_UINT_FIELD__SWAPPED(16);
 179 TP_UINT_FIELD__SWAPPED(32);
 180 TP_UINT_FIELD__SWAPPED(64);
 181
 182 static int __tp_field__init_uint(struct tp_field *field, int size, int offset, bool needs_swap)
 183 {
 184         field->offset = offset;
 185
 186         switch (size) {
 187         case 1:
 188                 field->integer = tp_field__u8;
 189                 break;
 190         case 2:
 191                 field->integer = needs_swap ? tp_field__swapped_u16 : tp_field__u16;
 192                 break;
 193         case 4:
 194                 field->integer = needs_swap ? tp_field__swapped_u32 : tp_field__u32;
 195                 break;
 196         case 8:
 197                 field->integer = needs_swap ? tp_field__swapped_u64 : tp_field__u64;
 198                 break;
 199         default:
 200                 return -1;
 201         }
 202
 203         return 0;
 204 }
 205
 206 static int tp_field__init_uint(struct tp_field *field, struct tep_format_field *format_field, bool needs_swap)
 207 {
 208         return __tp_field__init_uint(field, format_field->size, format_field->offset, needs_swap);
 209 }
 210
 211 static void *tp_field__ptr(struct tp_field *field, struct perf_sample *sample)
 212 {
 213         return sample->raw_data + field->offset;
 214 }
 215
 216 static int __tp_field__init_ptr(struct tp_field *field, int offset)
 217 {
 218         field->offset = offset;
 219         field->pointer = tp_field__ptr;
 220         return 0;
 221 }
 222
 223 static int tp_field__init_ptr(struct tp_field *field, struct tep_format_field *format_field)
 224 {
 225         return __tp_field__init_ptr(field, format_field->offset);
 226 }
 227
 228 struct syscall_tp {
 229         struct tp_field id;
 230         union {
 231                 struct tp_field args, ret;
 232         };
 233 };
 234
 235 static int perf_evsel__init_tp_uint_field(struct perf_evsel *evsel,
 236                                           struct tp_field *field,
 237                                           const char *name)
 238 {
 239         struct tep_format_field *format_field = perf_evsel__field(evsel, name);
 240
 241         if (format_field == NULL)
 242                 return -1;
 243
 244         return tp_field__init_uint(field, format_field, evsel->needs_swap);
 245 }
 246
 247 #define perf_evsel__init_sc_tp_uint_field(evsel, name) \
 248         ({ struct syscall_tp *sc = evsel->priv;\
 249            perf_evsel__init_tp_uint_field(evsel, &sc->name, #name); })
 250
 251 static int perf_evsel__init_tp_ptr_field(struct perf_evsel *evsel,
 252                                          struct tp_field *field,
 253                                          const char *name)
 254 {
 255         struct tep_format_field *format_field = perf_evsel__field(evsel, name);
 256
 257         if (format_field == NULL)
 258                 return -1;
 259
 260         return tp_field__init_ptr(field, format_field);
 261 }
 262
 263 #define perf_evsel__init_sc_tp_ptr_field(evsel, name) \
 264         ({ struct syscall_tp *sc = evsel->priv;\
 265            perf_evsel__init_tp_ptr_field(evsel, &sc->name, #name); })
 266
 267 static void perf_evsel__delete_priv(struct perf_evsel *evsel)
 268 {
 269         zfree(&evsel->priv);
 270         perf_evsel__delete(evsel);
 271 }
 272
 273 static int perf_evsel__init_syscall_tp(struct perf_evsel *evsel)
 274 {
 275         struct syscall_tp *sc = evsel->priv = malloc(sizeof(struct syscall_tp));
 276
 277         if (evsel->priv != NULL) {
 278                 if (perf_evsel__init_tp_uint_field(evsel, &sc->id, "__syscall_nr") &&
 279                     perf_evsel__init_tp_uint_field(evsel, &sc->id, "nr"))
 280                         goto out_delete;
 281                 return 0;
 282         }
 283
 284         return -ENOMEM;
 285 out_delete:
 286         zfree(&evsel->priv);
 287         return -ENOENT;
 288 }
 289
 290 static int perf_evsel__init_augmented_syscall_tp(struct perf_evsel *evsel, struct perf_evsel *tp)
 291 {
 292         struct syscall_tp *sc = evsel->priv = malloc(sizeof(struct syscall_tp));
 293
 294         if (evsel->priv != NULL) {
 295                 struct tep_format_field *syscall_id = perf_evsel__field(tp, "id");
 296                 if (syscall_id == NULL)
 297                         syscall_id = perf_evsel__field(tp, "__syscall_nr");
 298                 if (syscall_id == NULL)
 299                         goto out_delete;
 300                 if (__tp_field__init_uint(&sc->id, syscall_id->size, syscall_id->offset, evsel->needs_swap))
 301                         goto out_delete;
 302
 303                 return 0;
 304         }
 305
 306         return -ENOMEM;
 307 out_delete:
 308         zfree(&evsel->priv);
 309         return -EINVAL;
 310 }
 311
 312 static int perf_evsel__init_augmented_syscall_tp_args(struct perf_evsel *evsel)
 313 {
 314         struct syscall_tp *sc = evsel->priv;
 315
 316         return __tp_field__init_ptr(&sc->args, sc->id.offset + sizeof(u64));
 317 }
 318
 319 static int perf_evsel__init_augmented_syscall_tp_ret(struct perf_evsel *evsel)
 320 {
 321         struct syscall_tp *sc = evsel->priv;
 322
 323         return __tp_field__init_uint(&sc->ret, sizeof(u64), sc->id.offset + sizeof(u64), evsel->needs_swap);
 324 }
 325
 326 static int perf_evsel__init_raw_syscall_tp(struct perf_evsel *evsel, void *handler)
 327 {
 328         evsel->priv = malloc(sizeof(struct syscall_tp));
 329         if (evsel->priv != NULL) {
 330                 if (perf_evsel__init_sc_tp_uint_field(evsel, id))
 331                         goto out_delete;
 332
 333                 evsel->handler = handler;
 334                 return 0;
 335         }
 336
 337         return -ENOMEM;
 338
 339 out_delete:
 340         zfree(&evsel->priv);
 341         return -ENOENT;
 342 }
 343
 344 static struct perf_evsel *perf_evsel__raw_syscall_newtp(const char *direction, void *handler)
 345 {
 346         struct perf_evsel *evsel = perf_evsel__newtp("raw_syscalls", direction);
 347
 348         /* older kernel (e.g., RHEL6) use syscalls:{enter,exit} */
 349         if (IS_ERR(evsel))
 350                 evsel = perf_evsel__newtp("syscalls", direction);
 351
 352         if (IS_ERR(evsel))
 353                 return NULL;
 354
 355         if (perf_evsel__init_raw_syscall_tp(evsel, handler))
 356                 goto out_delete;
 357
 358         return evsel;
 359
 360 out_delete:
 361         perf_evsel__delete_priv(evsel);
 362         return NULL;
 363 }
 364
 365 #define perf_evsel__sc_tp_uint(evsel, name, sample) \
 366         ({ struct syscall_tp *fields = evsel->priv; \
 367            fields->name.integer(&fields->name, sample); })
 368
 369 #define perf_evsel__sc_tp_ptr(evsel, name, sample) \
 370         ({ struct syscall_tp *fields = evsel->priv; \
 371            fields->name.pointer(&fields->name, sample); })
 372
 373 size_t strarray__scnprintf(struct strarray *sa, char *bf, size_t size, const char *intfmt, bool show_prefix, int val)
 374 {
 375         int idx = val - sa->offset;
 376
 377         if (idx < 0 || idx >= sa->nr_entries || sa->entries[idx] == NULL) {
 378                 size_t printed = scnprintf(bf, size, intfmt, val);
 379                 if (show_prefix)
 380                         printed += scnprintf(bf + printed, size - printed, " /* %s??? */", sa->prefix);
 381                 return printed;
 382         }
 383
 384         return scnprintf(bf, size, "%s%s", show_prefix ? sa->prefix : "", sa->entries[idx]);
 385 }
 386
 387 static size_t __syscall_arg__scnprintf_strarray(char *bf, size_t size,
 388                                                 const char *intfmt,
 389                                                 struct syscall_arg *arg)
 390 {
 391         return strarray__scnprintf(arg->parm, bf, size, intfmt, arg->show_string_prefix, arg->val);
 392 }
 393
 394 static size_t syscall_arg__scnprintf_strarray(char *bf, size_t size,
 395                                               struct syscall_arg *arg)
 396 {
 397         return __syscall_arg__scnprintf_strarray(bf, size, "%d", arg);
 398 }
 399
 400 #define SCA_STRARRAY syscall_arg__scnprintf_strarray
 401
 402 size_t strarrays__scnprintf(struct strarrays *sas, char *bf, size_t size, const char *intfmt, bool show_prefix, int val)
 403 {
 404         size_t printed;
 405         int i;
 406
 407         for (i = 0; i < sas->nr_entries; ++i) {
 408                 struct strarray *sa = sas->entries[i];
 409                 int idx = val - sa->offset;
 410
 411                 if (idx >= 0 && idx < sa->nr_entries) {
 412                         if (sa->entries[idx] == NULL)
 413                                 break;
 414                         return scnprintf(bf, size, "%s%s", show_prefix ? sa->prefix : "", sa->entries[idx]);
 415                 }
 416         }
 417
 418         printed = scnprintf(bf, size, intfmt, val);
 419         if (show_prefix)
 420                 printed += scnprintf(bf + printed, size - printed, " /* %s??? */", sas->entries[0]->prefix);
 421         return printed;
 422 }
 423
 424 size_t syscall_arg__scnprintf_strarrays(char *bf, size_t size,
 425                                         struct syscall_arg *arg)
 426 {
 427         return strarrays__scnprintf(arg->parm, bf, size, "%d", arg->show_string_prefix, arg->val);
 428 }
 429
 430 #ifndef AT_FDCWD
 431 #define AT_FDCWD        -100
 432 #endif
 433
 434 static size_t syscall_arg__scnprintf_fd_at(char *bf, size_t size,
 435                                            struct syscall_arg *arg)
 436 {
 437         int fd = arg->val;
 438         const char *prefix = "AT_FD";
 439
 440         if (fd == AT_FDCWD)
 441                 return scnprintf(bf, size, "%s%s", arg->show_string_prefix ? prefix : "", "CWD");
 442
 443         return syscall_arg__scnprintf_fd(bf, size, arg);
 444 }
 445
 446 #define SCA_FDAT syscall_arg__scnprintf_fd_at
 447
 448 static size_t syscall_arg__scnprintf_close_fd(char *bf, size_t size,
 449                                               struct syscall_arg *arg);
 450
 451 #define SCA_CLOSE_FD syscall_arg__scnprintf_close_fd
 452
 453 size_t syscall_arg__scnprintf_hex(char *bf, size_t size, struct syscall_arg *arg)
 454 {
 455         return scnprintf(bf, size, "%#lx", arg->val);
 456 }
 457
 458 size_t syscall_arg__scnprintf_ptr(char *bf, size_t size, struct syscall_arg *arg)
 459 {
 460         if (arg->val == 0)
 461                 return scnprintf(bf, size, "NULL");
 462         return syscall_arg__scnprintf_hex(bf, size, arg);
 463 }
 464
 465 size_t syscall_arg__scnprintf_int(char *bf, size_t size, struct syscall_arg *arg)
 466 {
 467         return scnprintf(bf, size, "%d", arg->val);
 468 }
 469
 470 size_t syscall_arg__scnprintf_long(char *bf, size_t size, struct syscall_arg *arg)
 471 {
 472         return scnprintf(bf, size, "%ld", arg->val);
 473 }
 474
 475 static const char *bpf_cmd[] = {
 476         "MAP_CREATE", "MAP_LOOKUP_ELEM", "MAP_UPDATE_ELEM", "MAP_DELETE_ELEM",
 477         "MAP_GET_NEXT_KEY", "PROG_LOAD",
 478 };
 479 static DEFINE_STRARRAY(bpf_cmd, "BPF_");
 480
 481 static const char *epoll_ctl_ops[] = { "ADD", "DEL", "MOD", };
 482 static DEFINE_STRARRAY_OFFSET(epoll_ctl_ops, "EPOLL_CTL_", 1);
 483
 484 static const char *itimers[] = { "REAL", "VIRTUAL", "PROF", };
 485 static DEFINE_STRARRAY(itimers, "ITIMER_");
 486
 487 static const char *keyctl_options[] = {
 488         "GET_KEYRING_ID", "JOIN_SESSION_KEYRING", "UPDATE", "REVOKE", "CHOWN",
 489         "SETPERM", "DESCRIBE", "CLEAR", "LINK", "UNLINK", "SEARCH", "READ",
 490         "INSTANTIATE", "NEGATE", "SET_REQKEY_KEYRING", "SET_TIMEOUT",
 491         "ASSUME_AUTHORITY", "GET_SECURITY", "SESSION_TO_PARENT", "REJECT",
 492         "INSTANTIATE_IOV", "INVALIDATE", "GET_PERSISTENT",
 493 };
 494 static DEFINE_STRARRAY(keyctl_options, "KEYCTL_");
 495
 496 static const char *whences[] = { "SET", "CUR", "END",
 497 #ifdef SEEK_DATA
 498 "DATA",
 499 #endif
 500 #ifdef SEEK_HOLE
 501 "HOLE",
 502 #endif
 503 };
 504 static DEFINE_STRARRAY(whences, "SEEK_");
 505
 506 static const char *fcntl_cmds[] = {
 507         "DUPFD", "GETFD", "SETFD", "GETFL", "SETFL", "GETLK", "SETLK",
 508         "SETLKW", "SETOWN", "GETOWN", "SETSIG", "GETSIG", "GETLK64",
 509         "SETLK64", "SETLKW64", "SETOWN_EX", "GETOWN_EX",
 510         "GETOWNER_UIDS",
 511 };
 512 static DEFINE_STRARRAY(fcntl_cmds, "F_");
 513
 514 static const char *fcntl_linux_specific_cmds[] = {
 515         "SETLEASE", "GETLEASE", "NOTIFY", [5] = "CANCELLK", "DUPFD_CLOEXEC",
 516         "SETPIPE_SZ", "GETPIPE_SZ", "ADD_SEALS", "GET_SEALS",
 517         "GET_RW_HINT", "SET_RW_HINT", "GET_FILE_RW_HINT", "SET_FILE_RW_HINT",
 518 };
 519
 520 static DEFINE_STRARRAY_OFFSET(fcntl_linux_specific_cmds, "F_", F_LINUX_SPECIFIC_BASE);
 521
 522 static struct strarray *fcntl_cmds_arrays[] = {
 523         &strarray__fcntl_cmds,
 524         &strarray__fcntl_linux_specific_cmds,
 525 };
 526
 527 static DEFINE_STRARRAYS(fcntl_cmds_arrays);
 528
 529 static const char *rlimit_resources[] = {
 530         "CPU", "FSIZE", "DATA", "STACK", "CORE", "RSS", "NPROC", "NOFILE",
 531         "MEMLOCK", "AS", "LOCKS", "SIGPENDING", "MSGQUEUE", "NICE", "RTPRIO",
 532         "RTTIME",
 533 };
 534 static DEFINE_STRARRAY(rlimit_resources, "RLIMIT_");
 535
 536 static const char *sighow[] = { "BLOCK", "UNBLOCK", "SETMASK", };
 537 static DEFINE_STRARRAY(sighow, "SIG_");
 538
 539 static const char *clockid[] = {
 540         "REALTIME", "MONOTONIC", "PROCESS_CPUTIME_ID", "THREAD_CPUTIME_ID",
 541         "MONOTONIC_RAW", "REALTIME_COARSE", "MONOTONIC_COARSE", "BOOTTIME",
 542         "REALTIME_ALARM", "BOOTTIME_ALARM", "SGI_CYCLE", "TAI"
 543 };
 544 static DEFINE_STRARRAY(clockid, "CLOCK_");
 545
 546 static size_t syscall_arg__scnprintf_access_mode(char *bf, size_t size,
 547                                                  struct syscall_arg *arg)
 548 {
 549         bool show_prefix = arg->show_string_prefix;
 550         const char *suffix = "_OK";
 551         size_t printed = 0;
 552         int mode = arg->val;
 553
 554         if (mode == F_OK) /* 0 */
 555                 return scnprintf(bf, size, "F%s", show_prefix ? suffix : "");
 556 #define P_MODE(n) \
 557         if (mode & n##_OK) { \
 558                 printed += scnprintf(bf + printed, size - printed, "%s%s", #n, show_prefix ? suffix : ""); \
 559                 mode &= ~n##_OK; \
 560         }
 561
 562         P_MODE(R);
 563         P_MODE(W);
 564         P_MODE(X);
 565 #undef P_MODE
 566
 567         if (mode)
 568                 printed += scnprintf(bf + printed, size - printed, "|%#x", mode);
 569
 570         return printed;
 571 }
 572
 573 #define SCA_ACCMODE syscall_arg__scnprintf_access_mode
 574
 575 static size_t syscall_arg__scnprintf_filename(char *bf, size_t size,
 576                                               struct syscall_arg *arg);
 577
 578 #define SCA_FILENAME syscall_arg__scnprintf_filename
 579
 580 static size_t syscall_arg__scnprintf_pipe_flags(char *bf, size_t size,
 581                                                 struct syscall_arg *arg)
 582 {
 583         bool show_prefix = arg->show_string_prefix;
 584         const char *prefix = "O_";
 585         int printed = 0, flags = arg->val;
 586
 587 #define P_FLAG(n) \
 588         if (flags & O_##n) { \
 589                 printed += scnprintf(bf + printed, size - printed, "%s%s%s", printed ? "|" : "", show_prefix ? prefix : "", #n); \
 590                 flags &= ~O_##n; \
 591         }
 592
 593         P_FLAG(CLOEXEC);
 594         P_FLAG(NONBLOCK);
 595 #undef P_FLAG
 596
 597         if (flags)
 598                 printed += scnprintf(bf + printed, size - printed, "%s%#x", printed ? "|" : "", flags);
 599
 600         return printed;
 601 }
 602
 603 #define SCA_PIPE_FLAGS syscall_arg__scnprintf_pipe_flags
 604
 605 #ifndef GRND_NONBLOCK
 606 #define GRND_NONBLOCK   0x0001
 607 #endif
 608 #ifndef GRND_RANDOM
 609 #define GRND_RANDOM     0x0002
 610 #endif
 611
 612 static size_t syscall_arg__scnprintf_getrandom_flags(char *bf, size_t size,
 613                                                    struct syscall_arg *arg)
 614 {
 615         bool show_prefix = arg->show_string_prefix;
 616         const char *prefix = "GRND_";
 617         int printed = 0, flags = arg->val;
 618
 619 #define P_FLAG(n) \
 620         if (flags & GRND_##n) { \
 621                 printed += scnprintf(bf + printed, size - printed, "%s%s%s", printed ? "|" : "", show_prefix ? prefix : "", #n); \
 622                 flags &= ~GRND_##n; \
 623         }
 624
 625         P_FLAG(RANDOM);
 626         P_FLAG(NONBLOCK);
 627 #undef P_FLAG
 628
 629         if (flags)
 630                 printed += scnprintf(bf + printed, size - printed, "%s%#x", printed ? "|" : "", flags);
 631
 632         return printed;
 633 }
 634
 635 #define SCA_GETRANDOM_FLAGS syscall_arg__scnprintf_getrandom_flags
 636
 637 #define STRARRAY(name, array) \
 638           { .scnprintf  = SCA_STRARRAY, \
 639             .parm       = &strarray__##array, }
 640
 641 #include "trace/beauty/arch_errno_names.c"
 642 #include "trace/beauty/eventfd.c"
 643 #include "trace/beauty/futex_op.c"
 644 #include "trace/beauty/futex_val3.c"
 645 #include "trace/beauty/mmap.c"
 646 #include "trace/beauty/mode_t.c"
 647 #include "trace/beauty/msg_flags.c"
 648 #include "trace/beauty/open_flags.c"
 649 #include "trace/beauty/perf_event_open.c"
 650 #include "trace/beauty/pid.c"
 651 #include "trace/beauty/sched_policy.c"
 652 #include "trace/beauty/seccomp.c"
 653 #include "trace/beauty/signum.c"
 654 #include "trace/beauty/socket_type.c"
 655 #include "trace/beauty/waitid_options.c"
 656
 657 struct syscall_arg_fmt {
 658         size_t     (*scnprintf)(char *bf, size_t size, struct syscall_arg *arg);
 659         unsigned long (*mask_val)(struct syscall_arg *arg, unsigned long val);
 660         void       *parm;
 661         const char *name;
 662         bool       show_zero;
 663 };
 664
 665 static struct syscall_fmt {
 666         const char *name;
 667         const char *alias;
 668         struct syscall_arg_fmt arg[6];
 669         u8         nr_args;
 670         bool       errpid;
 671         bool       timeout;
 672         bool       hexret;
 673 } syscall_fmts[] = {
 674         { .name     = "access",
 675           .arg = { [1] = { .scnprintf = SCA_ACCMODE,  /* mode */ }, }, },
 676         { .name     = "arch_prctl",
 677           .arg = { [0] = { .scnprintf = SCA_X86_ARCH_PRCTL_CODE, /* code */ },
 678                    [1] = { .scnprintf = SCA_PTR, /* arg2 */ }, }, },
 679         { .name     = "bind",
 680           .arg = { [1] = { .scnprintf = SCA_SOCKADDR, /* umyaddr */ }, }, },
 681         { .name     = "bpf",
 682           .arg = { [0] = STRARRAY(cmd, bpf_cmd), }, },
 683         { .name     = "brk",        .hexret = true,
 684           .arg = { [0] = { .scnprintf = SCA_PTR, /* brk */ }, }, },
 685         { .name     = "clock_gettime",
 686           .arg = { [0] = STRARRAY(clk_id, clockid), }, },
 687         { .name     = "clone",      .errpid = true, .nr_args = 5,
 688           .arg = { [0] = { .name = "flags",         .scnprintf = SCA_CLONE_FLAGS, },
 689                    [1] = { .name = "child_stack",   .scnprintf = SCA_HEX, },
 690                    [2] = { .name = "parent_tidptr", .scnprintf = SCA_HEX, },
 691                    [3] = { .name = "child_tidptr",  .scnprintf = SCA_HEX, },
 692                    [4] = { .name = "tls",           .scnprintf = SCA_HEX, }, }, },
 693         { .name     = "close",
 694           .arg = { [0] = { .scnprintf = SCA_CLOSE_FD, /* fd */ }, }, },
 695         { .name     = "connect",
 696           .arg = { [1] = { .scnprintf = SCA_SOCKADDR, /* servaddr */ }, }, },
 697         { .name     = "epoll_ctl",
 698           .arg = { [1] = STRARRAY(op, epoll_ctl_ops), }, },
 699         { .name     = "eventfd2",
 700           .arg = { [1] = { .scnprintf = SCA_EFD_FLAGS, /* flags */ }, }, },
 701         { .name     = "fchmodat",
 702           .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
 703         { .name     = "fchownat",
 704           .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
 705         { .name     = "fcntl",
 706           .arg = { [1] = { .scnprintf = SCA_FCNTL_CMD, /* cmd */
 707                            .parm      = &strarrays__fcntl_cmds_arrays,
 708                            .show_zero = true, },
 709                    [2] = { .scnprintf =  SCA_FCNTL_ARG, /* arg */ }, }, },
 710         { .name     = "flock",
 711           .arg = { [1] = { .scnprintf = SCA_FLOCK, /* cmd */ }, }, },
 712         { .name     = "fstat", .alias = "newfstat", },
 713         { .name     = "fstatat", .alias = "newfstatat", },
 714         { .name     = "futex",
 715           .arg = { [1] = { .scnprintf = SCA_FUTEX_OP, /* op */ },
 716                    [5] = { .scnprintf = SCA_FUTEX_VAL3, /* val3 */ }, }, },
 717         { .name     = "futimesat",
 718           .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
 719         { .name     = "getitimer",
 720           .arg = { [0] = STRARRAY(which, itimers), }, },
 721         { .name     = "getpid",     .errpid = true, },
 722         { .name     = "getpgid",    .errpid = true, },
 723         { .name     = "getppid",    .errpid = true, },
 724         { .name     = "getrandom",
 725           .arg = { [2] = { .scnprintf = SCA_GETRANDOM_FLAGS, /* flags */ }, }, },
 726         { .name     = "getrlimit",
 727           .arg = { [0] = STRARRAY(resource, rlimit_resources), }, },
 728         { .name     = "gettid",     .errpid = true, },
 729         { .name     = "ioctl",
 730           .arg = {
 731 #if defined(__i386__) || defined(__x86_64__)
 732 /*
 733  * FIXME: Make this available to all arches.
 734  */
 735                    [1] = { .scnprintf = SCA_IOCTL_CMD, /* cmd */ },
 736                    [2] = { .scnprintf = SCA_HEX, /* arg */ }, }, },
 737 #else
 738                    [2] = { .scnprintf = SCA_HEX, /* arg */ }, }, },
 739 #endif
 740         { .name     = "kcmp",       .nr_args = 5,
 741           .arg = { [0] = { .name = "pid1",      .scnprintf = SCA_PID, },
 742                    [1] = { .name = "pid2",      .scnprintf = SCA_PID, },
 743                    [2] = { .name = "type",      .scnprintf = SCA_KCMP_TYPE, },
 744                    [3] = { .name = "idx1",      .scnprintf = SCA_KCMP_IDX, },
 745                    [4] = { .name = "idx2",      .scnprintf = SCA_KCMP_IDX, }, }, },
 746         { .name     = "keyctl",
 747           .arg = { [0] = STRARRAY(option, keyctl_options), }, },
 748         { .name     = "kill",
 749           .arg = { [1] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
 750         { .name     = "linkat",
 751           .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
 752         { .name     = "lseek",
 753           .arg = { [2] = STRARRAY(whence, whences), }, },
 754         { .name     = "lstat", .alias = "newlstat", },
 755         { .name     = "madvise",
 756           .arg = { [0] = { .scnprintf = SCA_HEX,      /* start */ },
 757                    [2] = { .scnprintf = SCA_MADV_BHV, /* behavior */ }, }, },
 758         { .name     = "mkdirat",
 759           .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
 760         { .name     = "mknodat",
 761           .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
 762         { .name     = "mmap",       .hexret = true,
 763 /* The standard mmap maps to old_mmap on s390x */
 764 #if defined(__s390x__)
 765         .alias = "old_mmap",
 766 #endif
 767           .arg = { [2] = { .scnprintf = SCA_MMAP_PROT,  /* prot */ },
 768                    [3] = { .scnprintf = SCA_MMAP_FLAGS, /* flags */ },
 769                    [5] = { .scnprintf = SCA_HEX,        /* offset */ }, }, },
 770         { .name     = "mount",
 771           .arg = { [0] = { .scnprintf = SCA_FILENAME, /* dev_name */ },
 772                    [3] = { .scnprintf = SCA_MOUNT_FLAGS, /* flags */
 773                            .mask_val  = SCAMV_MOUNT_FLAGS, /* flags */ }, }, },
 774         { .name     = "mprotect",
 775           .arg = { [0] = { .scnprintf = SCA_HEX,        /* start */ },
 776                    [2] = { .scnprintf = SCA_MMAP_PROT,  /* prot */ }, }, },
 777         { .name     = "mq_unlink",
 778           .arg = { [0] = { .scnprintf = SCA_FILENAME, /* u_name */ }, }, },
 779         { .name     = "mremap",     .hexret = true,
 780           .arg = { [3] = { .scnprintf = SCA_MREMAP_FLAGS, /* flags */ }, }, },
 781         { .name     = "name_to_handle_at",
 782           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
 783         { .name     = "newfstatat",
 784           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
 785         { .name     = "open",
 786           .arg = { [1] = { .scnprintf = SCA_OPEN_FLAGS, /* flags */ }, }, },
 787         { .name     = "open_by_handle_at",
 788           .arg = { [0] = { .scnprintf = SCA_FDAT,       /* dfd */ },
 789                    [2] = { .scnprintf = SCA_OPEN_FLAGS, /* flags */ }, }, },
 790         { .name     = "openat",
 791           .arg = { [0] = { .scnprintf = SCA_FDAT,       /* dfd */ },
 792                    [2] = { .scnprintf = SCA_OPEN_FLAGS, /* flags */ }, }, },
 793         { .name     = "perf_event_open",
 794           .arg = { [2] = { .scnprintf = SCA_INT,        /* cpu */ },
 795                    [3] = { .scnprintf = SCA_FD,         /* group_fd */ },
 796                    [4] = { .scnprintf = SCA_PERF_FLAGS, /* flags */ }, }, },
 797         { .name     = "pipe2",
 798           .arg = { [1] = { .scnprintf = SCA_PIPE_FLAGS, /* flags */ }, }, },
 799         { .name     = "pkey_alloc",
 800           .arg = { [1] = { .scnprintf = SCA_PKEY_ALLOC_ACCESS_RIGHTS,   /* access_rights */ }, }, },
 801         { .name     = "pkey_free",
 802           .arg = { [0] = { .scnprintf = SCA_INT,        /* key */ }, }, },
 803         { .name     = "pkey_mprotect",
 804           .arg = { [0] = { .scnprintf = SCA_HEX,        /* start */ },
 805                    [2] = { .scnprintf = SCA_MMAP_PROT,  /* prot */ },
 806                    [3] = { .scnprintf = SCA_INT,        /* pkey */ }, }, },
 807         { .name     = "poll", .timeout = true, },
 808         { .name     = "ppoll", .timeout = true, },
 809         { .name     = "prctl",
 810           .arg = { [0] = { .scnprintf = SCA_PRCTL_OPTION, /* option */ },
 811                    [1] = { .scnprintf = SCA_PRCTL_ARG2, /* arg2 */ },
 812                    [2] = { .scnprintf = SCA_PRCTL_ARG3, /* arg3 */ }, }, },
 813         { .name     = "pread", .alias = "pread64", },
 814         { .name     = "preadv", .alias = "pread", },
 815         { .name     = "prlimit64",
 816           .arg = { [1] = STRARRAY(resource, rlimit_resources), }, },
 817         { .name     = "pwrite", .alias = "pwrite64", },
 818         { .name     = "readlinkat",
 819           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
 820         { .name     = "recvfrom",
 821           .arg = { [3] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
 822         { .name     = "recvmmsg",
 823           .arg = { [3] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
 824         { .name     = "recvmsg",
 825           .arg = { [2] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
 826         { .name     = "renameat",
 827           .arg = { [0] = { .scnprintf = SCA_FDAT, /* olddirfd */ },
 828                    [2] = { .scnprintf = SCA_FDAT, /* newdirfd */ }, }, },
 829         { .name     = "renameat2",
 830           .arg = { [0] = { .scnprintf = SCA_FDAT, /* olddirfd */ },
 831                    [2] = { .scnprintf = SCA_FDAT, /* newdirfd */ },
 832                    [4] = { .scnprintf = SCA_RENAMEAT2_FLAGS, /* flags */ }, }, },
 833         { .name     = "rt_sigaction",
 834           .arg = { [0] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
 835         { .name     = "rt_sigprocmask",
 836           .arg = { [0] = STRARRAY(how, sighow), }, },
 837         { .name     = "rt_sigqueueinfo",
 838           .arg = { [1] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
 839         { .name     = "rt_tgsigqueueinfo",
 840           .arg = { [2] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
 841         { .name     = "sched_setscheduler",
 842           .arg = { [1] = { .scnprintf = SCA_SCHED_POLICY, /* policy */ }, }, },
 843         { .name     = "seccomp",
 844           .arg = { [0] = { .scnprintf = SCA_SECCOMP_OP,    /* op */ },
 845                    [1] = { .scnprintf = SCA_SECCOMP_FLAGS, /* flags */ }, }, },
 846         { .name     = "select", .timeout = true, },
 847         { .name     = "sendmmsg",
 848           .arg = { [3] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
 849         { .name     = "sendmsg",
 850           .arg = { [2] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
 851         { .name     = "sendto",
 852           .arg = { [3] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ },
 853                    [4] = { .scnprintf = SCA_SOCKADDR, /* addr */ }, }, },
 854         { .name     = "set_tid_address", .errpid = true, },
 855         { .name     = "setitimer",
 856           .arg = { [0] = STRARRAY(which, itimers), }, },
 857         { .name     = "setrlimit",
 858           .arg = { [0] = STRARRAY(resource, rlimit_resources), }, },
 859         { .name     = "socket",
 860           .arg = { [0] = STRARRAY(family, socket_families),
 861                    [1] = { .scnprintf = SCA_SK_TYPE, /* type */ },
 862                    [2] = { .scnprintf = SCA_SK_PROTO, /* protocol */ }, }, },
 863         { .name     = "socketpair",
 864           .arg = { [0] = STRARRAY(family, socket_families),
 865                    [1] = { .scnprintf = SCA_SK_TYPE, /* type */ },
 866                    [2] = { .scnprintf = SCA_SK_PROTO, /* protocol */ }, }, },
 867         { .name     = "stat", .alias = "newstat", },
 868         { .name     = "statx",
 869           .arg = { [0] = { .scnprintf = SCA_FDAT,        /* fdat */ },
 870                    [2] = { .scnprintf = SCA_STATX_FLAGS, /* flags */ } ,
 871                    [3] = { .scnprintf = SCA_STATX_MASK,  /* mask */ }, }, },
 872         { .name     = "swapoff",
 873           .arg = { [0] = { .scnprintf = SCA_FILENAME, /* specialfile */ }, }, },
 874         { .name     = "swapon",
 875           .arg = { [0] = { .scnprintf = SCA_FILENAME, /* specialfile */ }, }, },
 876         { .name     = "symlinkat",
 877           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
 878         { .name     = "tgkill",
 879           .arg = { [2] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
 880         { .name     = "tkill",
 881           .arg = { [1] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
 882         { .name     = "umount2", .alias = "umount",
 883           .arg = { [0] = { .scnprintf = SCA_FILENAME, /* name */ }, }, },
 884         { .name     = "uname", .alias = "newuname", },
 885         { .name     = "unlinkat",
 886           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
 887         { .name     = "utimensat",
 888           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dirfd */ }, }, },
 889         { .name     = "wait4",      .errpid = true,
 890           .arg = { [2] = { .scnprintf = SCA_WAITID_OPTIONS, /* options */ }, }, },
 891         { .name     = "waitid",     .errpid = true,
 892           .arg = { [3] = { .scnprintf = SCA_WAITID_OPTIONS, /* options */ }, }, },
 893 };
 894
 895 static int syscall_fmt__cmp(const void *name, const void *fmtp)
 896 {
 897         const struct syscall_fmt *fmt = fmtp;
 898         return strcmp(name, fmt->name);
 899 }
 900
 901 static struct syscall_fmt *syscall_fmt__find(const char *name)
 902 {
 903         const int nmemb = ARRAY_SIZE(syscall_fmts);
 904         return bsearch(name, syscall_fmts, nmemb, sizeof(struct syscall_fmt), syscall_fmt__cmp);
 905 }
 906
 907 static struct syscall_fmt *syscall_fmt__find_by_alias(const char *alias)
 908 {
 909         int i, nmemb = ARRAY_SIZE(syscall_fmts);
 910
 911         for (i = 0; i < nmemb; ++i) {
 912                 if (syscall_fmts[i].alias && strcmp(syscall_fmts[i].alias, alias) == 0)
 913                         return &syscall_fmts[i];
 914         }
 915
 916         return NULL;
 917 }
 918
 919 /*
 920  * is_exit: is this "exit" or "exit_group"?
 921  * is_open: is this "open" or "openat"? To associate the fd returned in sys_exit with the pathname in sys_enter.
 922  * args_size: sum of the sizes of the syscall arguments, anything after that is augmented stuff: pathname for openat, etc.
 923  */
 924 struct syscall {
 925         struct tep_event    *tp_format;
 926         int                 nr_args;
 927         int                 args_size;
 928         bool                is_exit;
 929         bool                is_open;
 930         struct tep_format_field *args;
 931         const char          *name;
 932         struct syscall_fmt  *fmt;
 933         struct syscall_arg_fmt *arg_fmt;
 934 };
 935
 936 struct bpf_map_syscall_entry {
 937         bool    enabled;
 938 };
 939
 940 /*
 941  * We need to have this 'calculated' boolean because in some cases we really
 942  * don't know what is the duration of a syscall, for instance, when we start
 943  * a session and some threads are waiting for a syscall to finish, say 'poll',
 944  * in which case all we can do is to print "( ? ) for duration and for the
 945  * start timestamp.
 946  */
 947 static size_t fprintf_duration(unsigned long t, bool calculated, FILE *fp)
 948 {
 949         double duration = (double)t / NSEC_PER_MSEC;
 950         size_t printed = fprintf(fp, "(");
 951
 952         if (!calculated)
 953                 printed += fprintf(fp, "         ");
 954         else if (duration >= 1.0)
 955                 printed += color_fprintf(fp, PERF_COLOR_RED, "%6.3f ms", duration);
 956         else if (duration >= 0.01)
 957                 printed += color_fprintf(fp, PERF_COLOR_YELLOW, "%6.3f ms", duration);
 958         else
 959                 printed += color_fprintf(fp, PERF_COLOR_NORMAL, "%6.3f ms", duration);
 960         return printed + fprintf(fp, "): ");
 961 }
 962
 963 /**
 964  * filename.ptr: The filename char pointer that will be vfs_getname'd
 965  * filename.entry_str_pos: Where to insert the string translated from
 966  *                         filename.ptr by the vfs_getname tracepoint/kprobe.
 967  * ret_scnprintf: syscall args may set this to a different syscall return
 968  *                formatter, for instance, fcntl may return fds, file flags, etc.
 969  */
 970 struct thread_trace {
 971         u64               entry_time;
 972         bool              entry_pending;
 973         unsigned long     nr_events;
 974         unsigned long     pfmaj, pfmin;
 975         char              *entry_str;
 976         double            runtime_ms;
 977         size_t            (*ret_scnprintf)(char *bf, size_t size, struct syscall_arg *arg);
 978         struct {
 979                 unsigned long ptr;
 980                 short int     entry_str_pos;
 981                 bool          pending_open;
 982                 unsigned int  namelen;
 983                 char          *name;
 984         } filename;
 985         struct {
 986                 int           max;
 987                 struct file   *table;
 988         } files;
 989
 990         struct intlist *syscall_stats;
 991 };
 992
 993 static struct thread_trace *thread_trace__new(void)
 994 {
 995         struct thread_trace *ttrace =  zalloc(sizeof(struct thread_trace));
 996
 997         if (ttrace)
 998                 ttrace->files.max = -1;
 999
1000         ttrace->syscall_stats = intlist__new(NULL);
1001
1002         return ttrace;
1003 }
1004
1005 static struct thread_trace *thread__trace(struct thread *thread, FILE *fp)
1006 {
1007         struct thread_trace *ttrace;
1008
1009         if (thread == NULL)
1010                 goto fail;
1011
1012         if (thread__priv(thread) == NULL)
1013                 thread__set_priv(thread, thread_trace__new());
1014
1015         if (thread__priv(thread) == NULL)
1016                 goto fail;
1017
1018         ttrace = thread__priv(thread);
1019         ++ttrace->nr_events;
1020
1021         return ttrace;
1022 fail:
1023         color_fprintf(fp, PERF_COLOR_RED,
1024                       "WARNING: not enough memory, dropping samples!\n");
1025         return NULL;
1026 }
1027
1028
1029 void syscall_arg__set_ret_scnprintf(struct syscall_arg *arg,
1030                                     size_t (*ret_scnprintf)(char *bf, size_t size, struct syscall_arg *arg))
1031 {
1032         struct thread_trace *ttrace = thread__priv(arg->thread);
1033
1034         ttrace->ret_scnprintf = ret_scnprintf;
1035 }
1036
1037 #define TRACE_PFMAJ             (1 << 0)
1038 #define TRACE_PFMIN             (1 << 1)
1039
1040 static const size_t trace__entry_str_size = 2048;
1041
1042 static struct file *thread_trace__files_entry(struct thread_trace *ttrace, int fd)
1043 {
1044         if (fd < 0)
1045                 return NULL;
1046
1047         if (fd > ttrace->files.max) {
1048                 struct file *nfiles = realloc(ttrace->files.table, (fd + 1) * sizeof(struct file));
1049
1050                 if (nfiles == NULL)
1051                         return NULL;
1052
1053                 if (ttrace->files.max != -1) {
1054                         memset(nfiles + ttrace->files.max + 1, 0,
1055                                (fd - ttrace->files.max) * sizeof(struct file));
1056                 } else {
1057                         memset(nfiles, 0, (fd + 1) * sizeof(struct file));
1058                 }
1059
1060                 ttrace->files.table = nfiles;
1061                 ttrace->files.max   = fd;
1062         }
1063
1064         return ttrace->files.table + fd;
1065 }
1066
1067 struct file *thread__files_entry(struct thread *thread, int fd)
1068 {
1069         return thread_trace__files_entry(thread__priv(thread), fd);
1070 }
1071
1072 static int trace__set_fd_pathname(struct thread *thread, int fd, const char *pathname)
1073 {
1074         struct thread_trace *ttrace = thread__priv(thread);
1075         struct file *file = thread_trace__files_entry(ttrace, fd);
1076
1077         if (file != NULL) {
1078                 struct stat st;
1079                 if (stat(pathname, &st) == 0)
1080                         file->dev_maj = major(st.st_rdev);
1081                 file->pathname = strdup(pathname);
1082                 if (file->pathname)
1083                         return 0;
1084         }
1085
1086         return -1;
1087 }
1088
1089 static int thread__read_fd_path(struct thread *thread, int fd)
1090 {
1091         char linkname[PATH_MAX], pathname[PATH_MAX];
1092         struct stat st;
1093         int ret;
1094
1095         if (thread->pid_ == thread->tid) {
1096                 scnprintf(linkname, sizeof(linkname),
1097                           "/proc/%d/fd/%d", thread->pid_, fd);
1098         } else {
1099                 scnprintf(linkname, sizeof(linkname),
1100                           "/proc/%d/task/%d/fd/%d", thread->pid_, thread->tid, fd);
1101         }
1102
1103         if (lstat(linkname, &st) < 0 || st.st_size + 1 > (off_t)sizeof(pathname))
1104                 return -1;
1105
1106         ret = readlink(linkname, pathname, sizeof(pathname));
1107
1108         if (ret < 0 || ret > st.st_size)
1109                 return -1;
1110
1111         pathname[ret] = '\0';
1112         return trace__set_fd_pathname(thread, fd, pathname);
1113 }
1114
1115 static const char *thread__fd_path(struct thread *thread, int fd,
1116                                    struct trace *trace)
1117 {
1118         struct thread_trace *ttrace = thread__priv(thread);
1119
1120         if (ttrace == NULL)
1121                 return NULL;
1122
1123         if (fd < 0)
1124                 return NULL;
1125
1126         if ((fd > ttrace->files.max || ttrace->files.table[fd].pathname == NULL)) {
1127                 if (!trace->live)
1128                         return NULL;
1129                 ++trace->stats.proc_getname;
1130                 if (thread__read_fd_path(thread, fd))
1131                         return NULL;
1132         }
1133
1134         return ttrace->files.table[fd].pathname;
1135 }
1136
1137 size_t syscall_arg__scnprintf_fd(char *bf, size_t size, struct syscall_arg *arg)
1138 {
1139         int fd = arg->val;
1140         size_t printed = scnprintf(bf, size, "%d", fd);
1141         const char *path = thread__fd_path(arg->thread, fd, arg->trace);
1142
1143         if (path)
1144                 printed += scnprintf(bf + printed, size - printed, "<%s>", path);
1145
1146         return printed;
1147 }
1148
1149 size_t pid__scnprintf_fd(struct trace *trace, pid_t pid, int fd, char *bf, size_t size)
1150 {
1151         size_t printed = scnprintf(bf, size, "%d", fd);
1152         struct thread *thread = machine__find_thread(trace->host, pid, pid);
1153
1154         if (thread) {
1155                 const char *path = thread__fd_path(thread, fd, trace);
1156
1157                 if (path)
1158                         printed += scnprintf(bf + printed, size - printed, "<%s>", path);
1159
1160                 thread__put(thread);
1161         }
1162
1163         return printed;
1164 }
1165
1166 static size_t syscall_arg__scnprintf_close_fd(char *bf, size_t size,
1167                                               struct syscall_arg *arg)
1168 {
1169         int fd = arg->val;
1170         size_t printed = syscall_arg__scnprintf_fd(bf, size, arg);
1171         struct thread_trace *ttrace = thread__priv(arg->thread);
1172
1173         if (ttrace && fd >= 0 && fd <= ttrace->files.max)
1174                 zfree(&ttrace->files.table[fd].pathname);
1175
1176         return printed;
1177 }
1178
1179 static void thread__set_filename_pos(struct thread *thread, const char *bf,
1180                                      unsigned long ptr)
1181 {
1182         struct thread_trace *ttrace = thread__priv(thread);
1183
1184         ttrace->filename.ptr = ptr;
1185         ttrace->filename.entry_str_pos = bf - ttrace->entry_str;
1186 }
1187
1188 static size_t syscall_arg__scnprintf_augmented_string(struct syscall_arg *arg, char *bf, size_t size)
1189 {
1190         struct augmented_arg *augmented_arg = arg->augmented.args;
1191
1192         return scnprintf(bf, size, "\"%.*s\"", augmented_arg->size, augmented_arg->value);
1193 }
1194
1195 static size_t syscall_arg__scnprintf_filename(char *bf, size_t size,
1196                                               struct syscall_arg *arg)
1197 {
1198         unsigned long ptr = arg->val;
1199
1200         if (arg->augmented.args)
1201                 return syscall_arg__scnprintf_augmented_string(arg, bf, size);
1202
1203         if (!arg->trace->vfs_getname)
1204                 return scnprintf(bf, size, "%#x", ptr);
1205
1206         thread__set_filename_pos(arg->thread, bf, ptr);
1207         return 0;
1208 }
1209
1210 static bool trace__filter_duration(struct trace *trace, double t)
1211 {
1212         return t < (trace->duration_filter * NSEC_PER_MSEC);
1213 }
1214
1215 static size_t __trace__fprintf_tstamp(struct trace *trace, u64 tstamp, FILE *fp)
1216 {
1217         double ts = (double)(tstamp - trace->base_time) / NSEC_PER_MSEC;
1218
1219         return fprintf(fp, "%10.3f ", ts);
1220 }
1221
1222 /*
1223  * We're handling tstamp=0 as an undefined tstamp, i.e. like when we are
1224  * using ttrace->entry_time for a thread that receives a sys_exit without
1225  * first having received a sys_enter ("poll" issued before tracing session
1226  * starts, lost sys_enter exit due to ring buffer overflow).
1227  */
1228 static size_t trace__fprintf_tstamp(struct trace *trace, u64 tstamp, FILE *fp)
1229 {
1230         if (tstamp > 0)
1231                 return __trace__fprintf_tstamp(trace, tstamp, fp);
1232
1233         return fprintf(fp, "         ? ");
1234 }
1235
1236 static bool done = false;
1237 static bool interrupted = false;
1238
1239 static void sig_handler(int sig)
1240 {
1241         done = true;
1242         interrupted = sig == SIGINT;
1243 }
1244
1245 static size_t trace__fprintf_comm_tid(struct trace *trace, struct thread *thread, FILE *fp)
1246 {
1247         size_t printed = 0;
1248
1249         if (trace->multiple_threads) {
1250                 if (trace->show_comm)
1251                         printed += fprintf(fp, "%.14s/", thread__comm_str(thread));
1252                 printed += fprintf(fp, "%d ", thread->tid);
1253         }
1254
1255         return printed;
1256 }
1257
1258 static size_t trace__fprintf_entry_head(struct trace *trace, struct thread *thread,
1259                                         u64 duration, bool duration_calculated, u64 tstamp, FILE *fp)
1260 {
1261         size_t printed = 0;
1262
1263         if (trace->show_tstamp)
1264                 printed = trace__fprintf_tstamp(trace, tstamp, fp);
1265         if (trace->show_duration)
1266                 printed += fprintf_duration(duration, duration_calculated, fp);
1267         return printed + trace__fprintf_comm_tid(trace, thread, fp);
1268 }
1269
1270 static int trace__process_event(struct trace *trace, struct machine *machine,
1271                                 union perf_event *event, struct perf_sample *sample)
1272 {
1273         int ret = 0;
1274
1275         switch (event->header.type) {
1276         case PERF_RECORD_LOST:
1277                 color_fprintf(trace->output, PERF_COLOR_RED,
1278                               "LOST %" PRIu64 " events!\n", event->lost.lost);
1279                 ret = machine__process_lost_event(machine, event, sample);
1280                 break;
1281         default:
1282                 ret = machine__process_event(machine, event, sample);
1283                 break;
1284         }
1285
1286         return ret;
1287 }
1288
1289 static int trace__tool_process(struct perf_tool *tool,
1290                                union perf_event *event,
1291                                struct perf_sample *sample,
1292                                struct machine *machine)
1293 {
1294         struct trace *trace = container_of(tool, struct trace, tool);
1295         return trace__process_event(trace, machine, event, sample);
1296 }
1297
1298 static char *trace__machine__resolve_kernel_addr(void *vmachine, unsigned long long *addrp, char **modp)
1299 {
1300         struct machine *machine = vmachine;
1301
1302         if (machine->kptr_restrict_warned)
1303                 return NULL;
1304
1305         if (symbol_conf.kptr_restrict) {
1306                 pr_warning("Kernel address maps (/proc/{kallsyms,modules}) are restricted.\n\n"
1307                            "Check /proc/sys/kernel/kptr_restrict.\n\n"
1308                            "Kernel samples will not be resolved.\n");
1309                 machine->kptr_restrict_warned = true;
1310                 return NULL;
1311         }
1312
1313         return machine__resolve_kernel_addr(vmachine, addrp, modp);
1314 }
1315
1316 static int trace__symbols_init(struct trace *trace, struct perf_evlist *evlist)
1317 {
1318         int err = symbol__init(NULL);
1319
1320         if (err)
1321                 return err;
1322
1323         trace->host = machine__new_host();
1324         if (trace->host == NULL)
1325                 return -ENOMEM;
1326
1327         err = trace_event__register_resolver(trace->host, trace__machine__resolve_kernel_addr);
1328         if (err < 0)
1329                 goto out;
1330
1331         err = __machine__synthesize_threads(trace->host, &trace->tool, &trace->opts.target,
1332                                             evlist->threads, trace__tool_process, false,
1333                                             1);
1334 out:
1335         if (err)
1336                 symbol__exit();
1337
1338         return err;
1339 }
1340
1341 static void trace__symbols__exit(struct trace *trace)
1342 {
1343         machine__exit(trace->host);
1344         trace->host = NULL;
1345
1346         symbol__exit();
1347 }
1348
1349 static int syscall__alloc_arg_fmts(struct syscall *sc, int nr_args)
1350 {
1351         int idx;
1352
1353         if (nr_args == 6 && sc->fmt && sc->fmt->nr_args != 0)
1354                 nr_args = sc->fmt->nr_args;
1355
1356         sc->arg_fmt = calloc(nr_args, sizeof(*sc->arg_fmt));
1357         if (sc->arg_fmt == NULL)
1358                 return -1;
1359
1360         for (idx = 0; idx < nr_args; ++idx) {
1361                 if (sc->fmt)
1362                         sc->arg_fmt[idx] = sc->fmt->arg[idx];
1363         }
1364
1365         sc->nr_args = nr_args;
1366         return 0;
1367 }
1368
1369 static int syscall__set_arg_fmts(struct syscall *sc)
1370 {
1371         struct tep_format_field *field, *last_field = NULL;
1372         int idx = 0, len;
1373
1374         for (field = sc->args; field; field = field->next, ++idx) {
1375                 last_field = field;
1376
1377                 if (sc->fmt && sc->fmt->arg[idx].scnprintf)
1378                         continue;
1379
1380                 if (strcmp(field->type, "const char *") == 0 &&
1381                          (strcmp(field->name, "filename") == 0 ||
1382                           strcmp(field->name, "path") == 0 ||
1383                           strcmp(field->name, "pathname") == 0))
1384                         sc->arg_fmt[idx].scnprintf = SCA_FILENAME;
1385                 else if ((field->flags & TEP_FIELD_IS_POINTER) || strstr(field->name, "addr"))
1386                         sc->arg_fmt[idx].scnprintf = SCA_PTR;
1387                 else if (strcmp(field->type, "pid_t") == 0)
1388                         sc->arg_fmt[idx].scnprintf = SCA_PID;
1389                 else if (strcmp(field->type, "umode_t") == 0)
1390                         sc->arg_fmt[idx].scnprintf = SCA_MODE_T;
1391                 else if ((strcmp(field->type, "int") == 0 ||
1392                           strcmp(field->type, "unsigned int") == 0 ||
1393                           strcmp(field->type, "long") == 0) &&
1394                          (len = strlen(field->name)) >= 2 &&
1395                          strcmp(field->name + len - 2, "fd") == 0) {
1396                         /*
1397                          * /sys/kernel/tracing/events/syscalls/sys_enter*
1398                          * egrep 'field:.*fd;' .../format|sed -r 's/.*field:([a-z ]+) [a-z_]*fd.+/\1/g'|sort|uniq -c
1399                          * 65 int
1400                          * 23 unsigned int
1401                          * 7 unsigned long
1402                          */
1403                         sc->arg_fmt[idx].scnprintf = SCA_FD;
1404                 }
1405         }
1406
1407         if (last_field)
1408                 sc->args_size = last_field->offset + last_field->size;
1409
1410         return 0;
1411 }
1412
1413 static int trace__read_syscall_info(struct trace *trace, int id)
1414 {
1415         char tp_name[128];
1416         struct syscall *sc;
1417         const char *name = syscalltbl__name(trace->sctbl, id);
1418
1419         if (name == NULL)
1420                 return -1;
1421
1422         if (id > trace->syscalls.max) {
1423                 struct syscall *nsyscalls = realloc(trace->syscalls.table, (id + 1) * sizeof(*sc));
1424
1425                 if (nsyscalls == NULL)
1426                         return -1;
1427
1428                 if (trace->syscalls.max != -1) {
1429                         memset(nsyscalls + trace->syscalls.max + 1, 0,
1430                                (id - trace->syscalls.max) * sizeof(*sc));
1431                 } else {
1432                         memset(nsyscalls, 0, (id + 1) * sizeof(*sc));
1433                 }
1434
1435                 trace->syscalls.table = nsyscalls;
1436                 trace->syscalls.max   = id;
1437         }
1438
1439         sc = trace->syscalls.table + id;
1440         sc->name = name;
1441
1442         sc->fmt  = syscall_fmt__find(sc->name);
1443
1444         snprintf(tp_name, sizeof(tp_name), "sys_enter_%s", sc->name);
1445         sc->tp_format = trace_event__tp_format("syscalls", tp_name);
1446
1447         if (IS_ERR(sc->tp_format) && sc->fmt && sc->fmt->alias) {
1448                 snprintf(tp_name, sizeof(tp_name), "sys_enter_%s", sc->fmt->alias);
1449                 sc->tp_format = trace_event__tp_format("syscalls", tp_name);
1450         }
1451
1452         if (syscall__alloc_arg_fmts(sc, IS_ERR(sc->tp_format) ? 6 : sc->tp_format->format.nr_fields))
1453                 return -1;
1454
1455         if (IS_ERR(sc->tp_format))
1456                 return -1;
1457
1458         sc->args = sc->tp_format->format.fields;
1459         /*
1460          * We need to check and discard the first variable '__syscall_nr'
1461          * or 'nr' that mean the syscall number. It is needless here.
1462          * So drop '__syscall_nr' or 'nr' field but does not exist on older kernels.
1463          */
1464         if (sc->args && (!strcmp(sc->args->name, "__syscall_nr") || !strcmp(sc->args->name, "nr"))) {
1465                 sc->args = sc->args->next;
1466                 --sc->nr_args;
1467         }
1468
1469         sc->is_exit = !strcmp(name, "exit_group") || !strcmp(name, "exit");
1470         sc->is_open = !strcmp(name, "open") || !strcmp(name, "openat");
1471
1472         return syscall__set_arg_fmts(sc);
1473 }
1474
1475 static int trace__validate_ev_qualifier(struct trace *trace)
1476 {
1477         int err = 0, i;
1478         size_t nr_allocated;
1479         struct str_node *pos;
1480
1481         trace->ev_qualifier_ids.nr = strlist__nr_entries(trace->ev_qualifier);
1482         trace->ev_qualifier_ids.entries = malloc(trace->ev_qualifier_ids.nr *
1483                                                  sizeof(trace->ev_qualifier_ids.entries[0]));
1484
1485         if (trace->ev_qualifier_ids.entries == NULL) {
1486                 fputs("Error:\tNot enough memory for allocating events qualifier ids\n",
1487                        trace->output);
1488                 err = -EINVAL;
1489                 goto out;
1490         }
1491
1492         nr_allocated = trace->ev_qualifier_ids.nr;
1493         i = 0;
1494
1495         strlist__for_each_entry(pos, trace->ev_qualifier) {
1496                 const char *sc = pos->s;
1497                 int id = syscalltbl__id(trace->sctbl, sc), match_next = -1;
1498
1499                 if (id < 0) {
1500                         id = syscalltbl__strglobmatch_first(trace->sctbl, sc, &match_next);
1501                         if (id >= 0)
1502                                 goto matches;
1503
1504                         if (err == 0) {
1505                                 fputs("Error:\tInvalid syscall ", trace->output);
1506                                 err = -EINVAL;
1507                         } else {
1508                                 fputs(", ", trace->output);
1509                         }
1510
1511                         fputs(sc, trace->output);
1512                 }
1513 matches:
1514                 trace->ev_qualifier_ids.entries[i++] = id;
1515                 if (match_next == -1)
1516                         continue;
1517
1518                 while (1) {
1519                         id = syscalltbl__strglobmatch_next(trace->sctbl, sc, &match_next);
1520                         if (id < 0)
1521                                 break;
1522                         if (nr_allocated == trace->ev_qualifier_ids.nr) {
1523                                 void *entries;
1524
1525                                 nr_allocated += 8;
1526                                 entries = realloc(trace->ev_qualifier_ids.entries,
1527                                                   nr_allocated * sizeof(trace->ev_qualifier_ids.entries[0]));
1528                                 if (entries == NULL) {
1529                                         err = -ENOMEM;
1530                                         fputs("\nError:\t Not enough memory for parsing\n", trace->output);
1531                                         goto out_free;
1532                                 }
1533                                 trace->ev_qualifier_ids.entries = entries;
1534                         }
1535                         trace->ev_qualifier_ids.nr++;
1536                         trace->ev_qualifier_ids.entries[i++] = id;
1537                 }
1538         }
1539
1540         if (err < 0) {
1541                 fputs("\nHint:\ttry 'perf list syscalls:sys_enter_*'"
1542                       "\nHint:\tand: 'man syscalls'\n", trace->output);
1543 out_free:
1544                 zfree(&trace->ev_qualifier_ids.entries);
1545                 trace->ev_qualifier_ids.nr = 0;
1546         }
1547 out:
1548         return err;
1549 }
1550
1551 /*
1552  * args is to be interpreted as a series of longs but we need to handle
1553  * 8-byte unaligned accesses. args points to raw_data within the event
1554  * and raw_data is guaranteed to be 8-byte unaligned because it is
1555  * preceded by raw_size which is a u32. So we need to copy args to a temp
1556  * variable to read it. Most notably this avoids extended load instructions
1557  * on unaligned addresses
1558  */
1559 unsigned long syscall_arg__val(struct syscall_arg *arg, u8 idx)
1560 {
1561         unsigned long val;
1562         unsigned char *p = arg->args + sizeof(unsigned long) * idx;
1563
1564         memcpy(&val, p, sizeof(val));
1565         return val;
1566 }
1567
1568 static size_t syscall__scnprintf_name(struct syscall *sc, char *bf, size_t size,
1569                                       struct syscall_arg *arg)
1570 {
1571         if (sc->arg_fmt && sc->arg_fmt[arg->idx].name)
1572                 return scnprintf(bf, size, "%s: ", sc->arg_fmt[arg->idx].name);
1573
1574         return scnprintf(bf, size, "arg%d: ", arg->idx);
1575 }
1576
1577 /*
1578  * Check if the value is in fact zero, i.e. mask whatever needs masking, such
1579  * as mount 'flags' argument that needs ignoring some magic flag, see comment
1580  * in tools/perf/trace/beauty/mount_flags.c
1581  */
1582 static unsigned long syscall__mask_val(struct syscall *sc, struct syscall_arg *arg, unsigned long val)
1583 {
1584         if (sc->arg_fmt && sc->arg_fmt[arg->idx].mask_val)
1585                 return sc->arg_fmt[arg->idx].mask_val(arg, val);
1586
1587         return val;
1588 }
1589
1590 static size_t syscall__scnprintf_val(struct syscall *sc, char *bf, size_t size,
1591                                      struct syscall_arg *arg, unsigned long val)
1592 {
1593         if (sc->arg_fmt && sc->arg_fmt[arg->idx].scnprintf) {
1594                 arg->val = val;
1595                 if (sc->arg_fmt[arg->idx].parm)
1596                         arg->parm = sc->arg_fmt[arg->idx].parm;
1597                 return sc->arg_fmt[arg->idx].scnprintf(bf, size, arg);
1598         }
1599         return scnprintf(bf, size, "%ld", val);
1600 }
1601
1602 static size_t syscall__scnprintf_args(struct syscall *sc, char *bf, size_t size,
1603                                       unsigned char *args, void *augmented_args, int augmented_args_size,
1604                                       struct trace *trace, struct thread *thread)
1605 {
1606         size_t printed = 0;
1607         unsigned long val;
1608         u8 bit = 1;
1609         struct syscall_arg arg = {
1610                 .args   = args,
1611                 .augmented = {
1612                         .size = augmented_args_size,
1613                         .args = augmented_args,
1614                 },
1615                 .idx    = 0,
1616                 .mask   = 0,
1617                 .trace  = trace,
1618                 .thread = thread,
1619                 .show_string_prefix = trace->show_string_prefix,
1620         };
1621         struct thread_trace *ttrace = thread__priv(thread);
1622
1623         /*
1624          * Things like fcntl will set this in its 'cmd' formatter to pick the
1625          * right formatter for the return value (an fd? file flags?), which is
1626          * not needed for syscalls that always return a given type, say an fd.
1627          */
1628         ttrace->ret_scnprintf = NULL;
1629
1630         if (sc->args != NULL) {
1631                 struct tep_format_field *field;
1632
1633                 for (field = sc->args; field;
1634                      field = field->next, ++arg.idx, bit <<= 1) {
1635                         if (arg.mask & bit)
1636                                 continue;
1637
1638                         val = syscall_arg__val(&arg, arg.idx);
1639                         /*
1640                          * Some syscall args need some mask, most don't and
1641                          * return val untouched.
1642                          */
1643                         val = syscall__mask_val(sc, &arg, val);
1644
1645                         /*
1646                          * Suppress this argument if its value is zero and
1647                          * and we don't have a string associated in an
1648                          * strarray for it.
1649                          */
1650                         if (val == 0 &&
1651                             !trace->show_zeros &&
1652                             !(sc->arg_fmt &&
1653                               (sc->arg_fmt[arg.idx].show_zero ||
1654                                sc->arg_fmt[arg.idx].scnprintf == SCA_STRARRAY ||
1655                                sc->arg_fmt[arg.idx].scnprintf == SCA_STRARRAYS) &&
1656                               sc->arg_fmt[arg.idx].parm))
1657                                 continue;
1658
1659                         printed += scnprintf(bf + printed, size - printed, "%s", printed ? ", " : "");
1660
1661                         if (trace->show_arg_names)
1662                                 printed += scnprintf(bf + printed, size - printed, "%s: ", field->name);
1663
1664                         printed += syscall__scnprintf_val(sc, bf + printed, size - printed, &arg, val);
1665                 }
1666         } else if (IS_ERR(sc->tp_format)) {
1667                 /*
1668                  * If we managed to read the tracepoint /format file, then we
1669                  * may end up not having any args, like with gettid(), so only
1670                  * print the raw args when we didn't manage to read it.
1671                  */
1672                 while (arg.idx < sc->nr_args) {
1673                         if (arg.mask & bit)
1674                                 goto next_arg;
1675                         val = syscall_arg__val(&arg, arg.idx);
1676                         if (printed)
1677                                 printed += scnprintf(bf + printed, size - printed, ", ");
1678                         printed += syscall__scnprintf_name(sc, bf + printed, size - printed, &arg);
1679                         printed += syscall__scnprintf_val(sc, bf + printed, size - printed, &arg, val);
1680 next_arg:
1681                         ++arg.idx;
1682                         bit <<= 1;
1683                 }
1684         }
1685
1686         return printed;
1687 }
1688
1689 typedef int (*tracepoint_handler)(struct trace *trace, struct perf_evsel *evsel,
1690                                   union perf_event *event,
1691                                   struct perf_sample *sample);
1692
1693 static struct syscall *trace__syscall_info(struct trace *trace,
1694                                            struct perf_evsel *evsel, int id)
1695 {
1696
1697         if (id < 0) {
1698
1699                 /*
1700                  * XXX: Noticed on x86_64, reproduced as far back as 3.0.36, haven't tried
1701                  * before that, leaving at a higher verbosity level till that is
1702                  * explained. Reproduced with plain ftrace with:
1703                  *
1704                  * echo 1 > /t/events/raw_syscalls/sys_exit/enable
1705                  * grep "NR -1 " /t/trace_pipe
1706                  *
1707                  * After generating some load on the machine.
1708                  */
1709                 if (verbose > 1) {
1710                         static u64 n;
1711                         fprintf(trace->output, "Invalid syscall %d id, skipping (%s, %" PRIu64 ") ...\n",
1712                                 id, perf_evsel__name(evsel), ++n);
1713                 }
1714                 return NULL;
1715         }
1716
1717         if ((id > trace->syscalls.max || trace->syscalls.table[id].name == NULL) &&
1718             trace__read_syscall_info(trace, id))
1719                 goto out_cant_read;
1720
1721         if ((id > trace->syscalls.max || trace->syscalls.table[id].name == NULL))
1722                 goto out_cant_read;
1723
1724         return &trace->syscalls.table[id];
1725
1726 out_cant_read:
1727         if (verbose > 0) {
1728                 fprintf(trace->output, "Problems reading syscall %d", id);
1729                 if (id <= trace->syscalls.max && trace->syscalls.table[id].name != NULL)
1730                         fprintf(trace->output, "(%s)", trace->syscalls.table[id].name);
1731                 fputs(" information\n", trace->output);
1732         }
1733         return NULL;
1734 }
1735
1736 static void thread__update_stats(struct thread_trace *ttrace,
1737                                  int id, struct perf_sample *sample)
1738 {
1739         struct int_node *inode;
1740         struct stats *stats;
1741         u64 duration = 0;
1742
1743         inode = intlist__findnew(ttrace->syscall_stats, id);
1744         if (inode == NULL)
1745                 return;
1746
1747         stats = inode->priv;
1748         if (stats == NULL) {
1749                 stats = malloc(sizeof(struct stats));
1750                 if (stats == NULL)
1751                         return;
1752                 init_stats(stats);
1753                 inode->priv = stats;
1754         }
1755
1756         if (ttrace->entry_time && sample->time > ttrace->entry_time)
1757                 duration = sample->time - ttrace->entry_time;
1758
1759         update_stats(stats, duration);
1760 }
1761
1762 static int trace__printf_interrupted_entry(struct trace *trace)
1763 {
1764         struct thread_trace *ttrace;
1765         size_t printed;
1766         int len;
1767
1768         if (trace->failure_only || trace->current == NULL)
1769                 return 0;
1770
1771         ttrace = thread__priv(trace->current);
1772
1773         if (!ttrace->entry_pending)
1774                 return 0;
1775
1776         printed  = trace__fprintf_entry_head(trace, trace->current, 0, false, ttrace->entry_time, trace->output);
1777         printed += len = fprintf(trace->output, "%s)", ttrace->entry_str);
1778
1779         if (len < trace->args_alignment - 4)
1780                 printed += fprintf(trace->output, "%-*s", trace->args_alignment - 4 - len, " ");
1781
1782         printed += fprintf(trace->output, " ...\n");
1783
1784         ttrace->entry_pending = false;
1785         ++trace->nr_events_printed;
1786
1787         return printed;
1788 }
1789
1790 static int trace__fprintf_sample(struct trace *trace, struct perf_evsel *evsel,
1791                                  struct perf_sample *sample, struct thread *thread)
1792 {
1793         int printed = 0;
1794
1795         if (trace->print_sample) {
1796                 double ts = (double)sample->time / NSEC_PER_MSEC;
1797
1798                 printed += fprintf(trace->output, "%22s %10.3f %s %d/%d [%d]\n",
1799                                    perf_evsel__name(evsel), ts,
1800                                    thread__comm_str(thread),
1801                                    sample->pid, sample->tid, sample->cpu);
1802         }
1803
1804         return printed;
1805 }
1806
1807 static void *syscall__augmented_args(struct syscall *sc, struct perf_sample *sample, int *augmented_args_size, int raw_augmented_args_size)
1808 {
1809         void *augmented_args = NULL;
1810         /*
1811          * For now with BPF raw_augmented we hook into raw_syscalls:sys_enter
1812          * and there we get all 6 syscall args plus the tracepoint common fields
1813          * that gets calculated at the start and the syscall_nr (another long).
1814          * So we check if that is the case and if so don't look after the
1815          * sc->args_size but always after the full raw_syscalls:sys_enter payload,
1816          * which is fixed.
1817          *
1818          * We'll revisit this later to pass s->args_size to the BPF augmenter
1819          * (now tools/perf/examples/bpf/augmented_raw_syscalls.c, so that it
1820          * copies only what we need for each syscall, like what happens when we
1821          * use syscalls:sys_enter_NAME, so that we reduce the kernel/userspace
1822          * traffic to just what is needed for each syscall.
1823          */
1824         int args_size = raw_augmented_args_size ?: sc->args_size;
1825
1826         *augmented_args_size = sample->raw_size - args_size;
1827         if (*augmented_args_size > 0)
1828                 augmented_args = sample->raw_data + args_size;
1829
1830         return augmented_args;
1831 }
1832
1833 static int trace__sys_enter(struct trace *trace, struct perf_evsel *evsel,
1834                             union perf_event *event __maybe_unused,
1835                             struct perf_sample *sample)
1836 {
1837         char *msg;
1838         void *args;
1839         int printed = 0;
1840         struct thread *thread;
1841         int id = perf_evsel__sc_tp_uint(evsel, id, sample), err = -1;
1842         int augmented_args_size = 0;
1843         void *augmented_args = NULL;
1844         struct syscall *sc = trace__syscall_info(trace, evsel, id);
1845         struct thread_trace *ttrace;
1846
1847         if (sc == NULL)
1848                 return -1;
1849
1850         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
1851         ttrace = thread__trace(thread, trace->output);
1852         if (ttrace == NULL)
1853                 goto out_put;
1854
1855         trace__fprintf_sample(trace, evsel, sample, thread);
1856
1857         args = perf_evsel__sc_tp_ptr(evsel, args, sample);
1858
1859         if (ttrace->entry_str == NULL) {
1860                 ttrace->entry_str = malloc(trace__entry_str_size);
1861                 if (!ttrace->entry_str)
1862                         goto out_put;
1863         }
1864
1865         if (!(trace->duration_filter || trace->summary_only || trace->min_stack))
1866                 trace__printf_interrupted_entry(trace);
1867         /*
1868          * If this is raw_syscalls.sys_enter, then it always comes with the 6 possible
1869          * arguments, even if the syscall being handled, say "openat", uses only 4 arguments
1870          * this breaks syscall__augmented_args() check for augmented args, as we calculate
1871          * syscall->args_size using each syscalls:sys_enter_NAME tracefs format file,
1872          * so when handling, say the openat syscall, we end up getting 6 args for the
1873          * raw_syscalls:sys_enter event, when we expected just 4, we end up mistakenly
1874          * thinking that the extra 2 u64 args are the augmented filename, so just check
1875          * here and avoid using augmented syscalls when the evsel is the raw_syscalls one.
1876          */
1877         if (evsel != trace->syscalls.events.sys_enter)
1878                 augmented_args = syscall__augmented_args(sc, sample, &augmented_args_size, trace->raw_augmented_syscalls_args_size);
1879         ttrace->entry_time = sample->time;
1880         msg = ttrace->entry_str;
1881         printed += scnprintf(msg + printed, trace__entry_str_size - printed, "%s(", sc->name);
1882
1883         printed += syscall__scnprintf_args(sc, msg + printed, trace__entry_str_size - printed,
1884                                            args, augmented_args, augmented_args_size, trace, thread);
1885
1886         if (sc->is_exit) {
1887                 if (!(trace->duration_filter || trace->summary_only || trace->failure_only || trace->min_stack)) {
1888                         int alignment = 0;
1889
1890                         trace__fprintf_entry_head(trace, thread, 0, false, ttrace->entry_time, trace->output);
1891                         printed = fprintf(trace->output, "%s)", ttrace->entry_str);
1892                         if (trace->args_alignment > printed)
1893                                 alignment = trace->args_alignment - printed;
1894                         fprintf(trace->output, "%*s= ?\n", alignment, " ");
1895                 }
1896         } else {
1897                 ttrace->entry_pending = true;
1898                 /* See trace__vfs_getname & trace__sys_exit */
1899                 ttrace->filename.pending_open = false;
1900         }
1901
1902         if (trace->current != thread) {
1903                 thread__put(trace->current);
1904                 trace->current = thread__get(thread);
1905         }
1906         err = 0;
1907 out_put:
1908         thread__put(thread);
1909         return err;
1910 }
1911
1912 static int trace__fprintf_sys_enter(struct trace *trace, struct perf_evsel *evsel,
1913                                     struct perf_sample *sample)
1914 {
1915         struct thread_trace *ttrace;
1916         struct thread *thread;
1917         int id = perf_evsel__sc_tp_uint(evsel, id, sample), err = -1;
1918         struct syscall *sc = trace__syscall_info(trace, evsel, id);
1919         char msg[1024];
1920         void *args, *augmented_args = NULL;
1921         int augmented_args_size;
1922
1923         if (sc == NULL)
1924                 return -1;
1925
1926         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
1927         ttrace = thread__trace(thread, trace->output);
1928         /*
1929          * We need to get ttrace just to make sure it is there when syscall__scnprintf_args()
1930          * and the rest of the beautifiers accessing it via struct syscall_arg touches it.
1931          */
1932         if (ttrace == NULL)
1933                 goto out_put;
1934
1935         args = perf_evsel__sc_tp_ptr(evsel, args, sample);
1936         augmented_args = syscall__augmented_args(sc, sample, &augmented_args_size, trace->raw_augmented_syscalls_args_size);
1937         syscall__scnprintf_args(sc, msg, sizeof(msg), args, augmented_args, augmented_args_size, trace, thread);
1938         fprintf(trace->output, "%s", msg);
1939         err = 0;
1940 out_put:
1941         thread__put(thread);
1942         return err;
1943 }
1944
1945 static int trace__resolve_callchain(struct trace *trace, struct perf_evsel *evsel,
1946                                     struct perf_sample *sample,
1947                                     struct callchain_cursor *cursor)
1948 {
1949         struct addr_location al;
1950         int max_stack = evsel->attr.sample_max_stack ?
1951                         evsel->attr.sample_max_stack :
1952                         trace->max_stack;
1953         int err;
1954
1955         if (machine__resolve(trace->host, &al, sample) < 0)
1956                 return -1;
1957
1958         err = thread__resolve_callchain(al.thread, cursor, evsel, sample, NULL, NULL, max_stack);
1959         addr_location__put(&al);
1960         return err;
1961 }
1962
1963 static int trace__fprintf_callchain(struct trace *trace, struct perf_sample *sample)
1964 {
1965         /* TODO: user-configurable print_opts */
1966         const unsigned int print_opts = EVSEL__PRINT_SYM |
1967                                         EVSEL__PRINT_DSO |
1968                                         EVSEL__PRINT_UNKNOWN_AS_ADDR;
1969
1970         return sample__fprintf_callchain(sample, 38, print_opts, &callchain_cursor, trace->output);
1971 }
1972
1973 static const char *errno_to_name(struct perf_evsel *evsel, int err)
1974 {
1975         struct perf_env *env = perf_evsel__env(evsel);
1976         const char *arch_name = perf_env__arch(env);
1977
1978         return arch_syscalls__strerrno(arch_name, err);
1979 }
1980
1981 static int trace__sys_exit(struct trace *trace, struct perf_evsel *evsel,
1982                            union perf_event *event __maybe_unused,
1983                            struct perf_sample *sample)
1984 {
1985         long ret;
1986         u64 duration = 0;
1987         bool duration_calculated = false;
1988         struct thread *thread;
1989         int id = perf_evsel__sc_tp_uint(evsel, id, sample), err = -1, callchain_ret = 0, printed = 0;
1990         int alignment = trace->args_alignment;
1991         struct syscall *sc = trace__syscall_info(trace, evsel, id);
1992         struct thread_trace *ttrace;
1993
1994         if (sc == NULL)
1995                 return -1;
1996
1997         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
1998         ttrace = thread__trace(thread, trace->output);
1999         if (ttrace == NULL)
2000                 goto out_put;
2001
2002         trace__fprintf_sample(trace, evsel, sample, thread);
2003
2004         if (trace->summary)
2005                 thread__update_stats(ttrace, id, sample);
2006
2007         ret = perf_evsel__sc_tp_uint(evsel, ret, sample);
2008
2009         if (sc->is_open && ret >= 0 && ttrace->filename.pending_open) {
2010                 trace__set_fd_pathname(thread, ret, ttrace->filename.name);
2011                 ttrace->filename.pending_open = false;
2012                 ++trace->stats.vfs_getname;
2013         }
2014
2015         if (ttrace->entry_time) {
2016                 duration = sample->time - ttrace->entry_time;
2017                 if (trace__filter_duration(trace, duration))
2018                         goto out;
2019                 duration_calculated = true;
2020         } else if (trace->duration_filter)
2021                 goto out;
2022
2023         if (sample->callchain) {
2024                 callchain_ret = trace__resolve_callchain(trace, evsel, sample, &callchain_cursor);
2025                 if (callchain_ret == 0) {
2026                         if (callchain_cursor.nr < trace->min_stack)
2027                                 goto out;
2028                         callchain_ret = 1;
2029                 }
2030         }
2031
2032         if (trace->summary_only || (ret >= 0 && trace->failure_only))
2033                 goto out;
2034
2035         trace__fprintf_entry_head(trace, thread, duration, duration_calculated, ttrace->entry_time, trace->output);
2036
2037         if (ttrace->entry_pending) {
2038                 printed = fprintf(trace->output, "%s", ttrace->entry_str);
2039         } else {
2040                 printed += fprintf(trace->output, " ... [");
2041                 color_fprintf(trace->output, PERF_COLOR_YELLOW, "continued");
2042                 printed += 9;
2043                 printed += fprintf(trace->output, "]: %s()", sc->name);
2044         }
2045
2046         printed++; /* the closing ')' */
2047
2048         if (alignment > printed)
2049                 alignment -= printed;
2050         else
2051                 alignment = 0;
2052
2053         fprintf(trace->output, ")%*s= ", alignment, " ");
2054
2055         if (sc->fmt == NULL) {
2056                 if (ret < 0)
2057                         goto errno_print;
2058 signed_print:
2059                 fprintf(trace->output, "%ld", ret);
2060         } else if (ret < 0) {
2061 errno_print: {
2062                 char bf[STRERR_BUFSIZE];
2063                 const char *emsg = str_error_r(-ret, bf, sizeof(bf)),
2064                            *e = errno_to_name(evsel, -ret);
2065
2066                 fprintf(trace->output, "-1 %s (%s)", e, emsg);
2067         }
2068         } else if (ret == 0 && sc->fmt->timeout)
2069                 fprintf(trace->output, "0 (Timeout)");
2070         else if (ttrace->ret_scnprintf) {
2071                 char bf[1024];
2072                 struct syscall_arg arg = {
2073                         .val    = ret,
2074                         .thread = thread,
2075                         .trace  = trace,
2076                 };
2077                 ttrace->ret_scnprintf(bf, sizeof(bf), &arg);
2078                 ttrace->ret_scnprintf = NULL;
2079                 fprintf(trace->output, "%s", bf);
2080         } else if (sc->fmt->hexret)
2081                 fprintf(trace->output, "%#lx", ret);
2082         else if (sc->fmt->errpid) {
2083                 struct thread *child = machine__find_thread(trace->host, ret, ret);
2084
2085                 if (child != NULL) {
2086                         fprintf(trace->output, "%ld", ret);
2087                         if (child->comm_set)
2088                                 fprintf(trace->output, " (%s)", thread__comm_str(child));
2089                         thread__put(child);
2090                 }
2091         } else
2092                 goto signed_print;
2093
2094         fputc('\n', trace->output);
2095
2096         /*
2097          * We only consider an 'event' for the sake of --max-events a non-filtered
2098          * sys_enter + sys_exit and other tracepoint events.
2099          */
2100         if (++trace->nr_events_printed == trace->max_events && trace->max_events != ULONG_MAX)
2101                 interrupted = true;
2102
2103         if (callchain_ret > 0)
2104                 trace__fprintf_callchain(trace, sample);
2105         else if (callchain_ret < 0)
2106                 pr_err("Problem processing %s callchain, skipping...\n", perf_evsel__name(evsel));
2107 out:
2108         ttrace->entry_pending = false;
2109         err = 0;
2110 out_put:
2111         thread__put(thread);
2112         return err;
2113 }
2114
2115 static int trace__vfs_getname(struct trace *trace, struct perf_evsel *evsel,
2116                               union perf_event *event __maybe_unused,
2117                               struct perf_sample *sample)
2118 {
2119         struct thread *thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2120         struct thread_trace *ttrace;
2121         size_t filename_len, entry_str_len, to_move;
2122         ssize_t remaining_space;
2123         char *pos;
2124         const char *filename = perf_evsel__rawptr(evsel, sample, "pathname");
2125
2126         if (!thread)
2127                 goto out;
2128
2129         ttrace = thread__priv(thread);
2130         if (!ttrace)
2131                 goto out_put;
2132
2133         filename_len = strlen(filename);
2134         if (filename_len == 0)
2135                 goto out_put;
2136
2137         if (ttrace->filename.namelen < filename_len) {
2138                 char *f = realloc(ttrace->filename.name, filename_len + 1);
2139
2140                 if (f == NULL)
2141                         goto out_put;
2142
2143                 ttrace->filename.namelen = filename_len;
2144                 ttrace->filename.name = f;
2145         }
2146
2147         strcpy(ttrace->filename.name, filename);
2148         ttrace->filename.pending_open = true;
2149
2150         if (!ttrace->filename.ptr)
2151                 goto out_put;
2152
2153         entry_str_len = strlen(ttrace->entry_str);
2154         remaining_space = trace__entry_str_size - entry_str_len - 1; /* \0 */
2155         if (remaining_space <= 0)
2156                 goto out_put;
2157
2158         if (filename_len > (size_t)remaining_space) {
2159                 filename += filename_len - remaining_space;
2160                 filename_len = remaining_space;
2161         }
2162
2163         to_move = entry_str_len - ttrace->filename.entry_str_pos + 1; /* \0 */
2164         pos = ttrace->entry_str + ttrace->filename.entry_str_pos;
2165         memmove(pos + filename_len, pos, to_move);
2166         memcpy(pos, filename, filename_len);
2167
2168         ttrace->filename.ptr = 0;
2169         ttrace->filename.entry_str_pos = 0;
2170 out_put:
2171         thread__put(thread);
2172 out:
2173         return 0;
2174 }
2175
2176 static int trace__sched_stat_runtime(struct trace *trace, struct perf_evsel *evsel,
2177                                      union perf_event *event __maybe_unused,
2178                                      struct perf_sample *sample)
2179 {
2180         u64 runtime = perf_evsel__intval(evsel, sample, "runtime");
2181         double runtime_ms = (double)runtime / NSEC_PER_MSEC;
2182         struct thread *thread = machine__findnew_thread(trace->host,
2183                                                         sample->pid,
2184                                                         sample->tid);
2185         struct thread_trace *ttrace = thread__trace(thread, trace->output);
2186
2187         if (ttrace == NULL)
2188                 goto out_dump;
2189
2190         ttrace->runtime_ms += runtime_ms;
2191         trace->runtime_ms += runtime_ms;
2192 out_put:
2193         thread__put(thread);
2194         return 0;
2195
2196 out_dump:
2197         fprintf(trace->output, "%s: comm=%s,pid=%u,runtime=%" PRIu64 ",vruntime=%" PRIu64 ")\n",
2198                evsel->name,
2199                perf_evsel__strval(evsel, sample, "comm"),
2200                (pid_t)perf_evsel__intval(evsel, sample, "pid"),
2201                runtime,
2202                perf_evsel__intval(evsel, sample, "vruntime"));
2203         goto out_put;
2204 }
2205
2206 static int bpf_output__printer(enum binary_printer_ops op,
2207                                unsigned int val, void *extra __maybe_unused, FILE *fp)
2208 {
2209         unsigned char ch = (unsigned char)val;
2210
2211         switch (op) {
2212         case BINARY_PRINT_CHAR_DATA:
2213                 return fprintf(fp, "%c", isprint(ch) ? ch : '.');
2214         case BINARY_PRINT_DATA_BEGIN:
2215         case BINARY_PRINT_LINE_BEGIN:
2216         case BINARY_PRINT_ADDR:
2217         case BINARY_PRINT_NUM_DATA:
2218         case BINARY_PRINT_NUM_PAD:
2219         case BINARY_PRINT_SEP:
2220         case BINARY_PRINT_CHAR_PAD:
2221         case BINARY_PRINT_LINE_END:
2222         case BINARY_PRINT_DATA_END:
2223         default:
2224                 break;
2225         }
2226
2227         return 0;
2228 }
2229
2230 static void bpf_output__fprintf(struct trace *trace,
2231                                 struct perf_sample *sample)
2232 {
2233         binary__fprintf(sample->raw_data, sample->raw_size, 8,
2234                         bpf_output__printer, NULL, trace->output);
2235         ++trace->nr_events_printed;
2236 }
2237
2238 static int trace__event_handler(struct trace *trace, struct perf_evsel *evsel,
2239                                 union perf_event *event __maybe_unused,
2240                                 struct perf_sample *sample)
2241 {
2242         struct thread *thread;
2243         int callchain_ret = 0;
2244         /*
2245          * Check if we called perf_evsel__disable(evsel) due to, for instance,
2246          * this event's max_events having been hit and this is an entry coming
2247          * from the ring buffer that we should discard, since the max events
2248          * have already been considered/printed.
2249          */
2250         if (evsel->disabled)
2251                 return 0;
2252
2253         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2254
2255         if (sample->callchain) {
2256                 callchain_ret = trace__resolve_callchain(trace, evsel, sample, &callchain_cursor);
2257                 if (callchain_ret == 0) {
2258                         if (callchain_cursor.nr < trace->min_stack)
2259                                 goto out;
2260                         callchain_ret = 1;
2261                 }
2262         }
2263
2264         trace__printf_interrupted_entry(trace);
2265         trace__fprintf_tstamp(trace, sample->time, trace->output);
2266
2267         if (trace->trace_syscalls && trace->show_duration)
2268                 fprintf(trace->output, "(         ): ");
2269
2270         if (thread)
2271                 trace__fprintf_comm_tid(trace, thread, trace->output);
2272
2273         if (evsel == trace->syscalls.events.augmented) {
2274                 int id = perf_evsel__sc_tp_uint(evsel, id, sample);
2275                 struct syscall *sc = trace__syscall_info(trace, evsel, id);
2276
2277                 if (sc) {
2278                         fprintf(trace->output, "%s(", sc->name);
2279                         trace__fprintf_sys_enter(trace, evsel, sample);
2280                         fputc(')', trace->output);
2281                         goto newline;
2282                 }
2283
2284                 /*
2285                  * XXX: Not having the associated syscall info or not finding/adding
2286                  *      the thread should never happen, but if it does...
2287                  *      fall thru and print it as a bpf_output event.
2288                  */
2289         }
2290
2291         fprintf(trace->output, "%s:", evsel->name);
2292
2293         if (perf_evsel__is_bpf_output(evsel)) {
2294                 bpf_output__fprintf(trace, sample);
2295         } else if (evsel->tp_format) {
2296                 if (strncmp(evsel->tp_format->name, "sys_enter_", 10) ||
2297                     trace__fprintf_sys_enter(trace, evsel, sample)) {
2298                         event_format__fprintf(evsel->tp_format, sample->cpu,
2299                                               sample->raw_data, sample->raw_size,
2300                                               trace->output);
2301                         ++trace->nr_events_printed;
2302
2303                         if (evsel->max_events != ULONG_MAX && ++evsel->nr_events_printed == evsel->max_events) {
2304                                 perf_evsel__disable(evsel);
2305                                 perf_evsel__close(evsel);
2306                         }
2307                 }
2308         }
2309
2310 newline:
2311         fprintf(trace->output, "\n");
2312
2313         if (callchain_ret > 0)
2314                 trace__fprintf_callchain(trace, sample);
2315         else if (callchain_ret < 0)
2316                 pr_err("Problem processing %s callchain, skipping...\n", perf_evsel__name(evsel));
2317 out:
2318         thread__put(thread);
2319         return 0;
2320 }
2321
2322 static void print_location(FILE *f, struct perf_sample *sample,
2323                            struct addr_location *al,
2324                            bool print_dso, bool print_sym)
2325 {
2326
2327         if ((verbose > 0 || print_dso) && al->map)
2328                 fprintf(f, "%s@", al->map->dso->long_name);
2329
2330         if ((verbose > 0 || print_sym) && al->sym)
2331                 fprintf(f, "%s+0x%" PRIx64, al->sym->name,
2332                         al->addr - al->sym->start);
2333         else if (al->map)
2334                 fprintf(f, "0x%" PRIx64, al->addr);
2335         else
2336                 fprintf(f, "0x%" PRIx64, sample->addr);
2337 }
2338
2339 static int trace__pgfault(struct trace *trace,
2340                           struct perf_evsel *evsel,
2341                           union perf_event *event __maybe_unused,
2342                           struct perf_sample *sample)
2343 {
2344         struct thread *thread;
2345         struct addr_location al;
2346         char map_type = 'd';
2347         struct thread_trace *ttrace;
2348         int err = -1;
2349         int callchain_ret = 0;
2350
2351         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2352
2353         if (sample->callchain) {
2354                 callchain_ret = trace__resolve_callchain(trace, evsel, sample, &callchain_cursor);
2355                 if (callchain_ret == 0) {
2356                         if (callchain_cursor.nr < trace->min_stack)
2357                                 goto out_put;
2358                         callchain_ret = 1;
2359                 }
2360         }
2361
2362         ttrace = thread__trace(thread, trace->output);
2363         if (ttrace == NULL)
2364                 goto out_put;
2365
2366         if (evsel->attr.config == PERF_COUNT_SW_PAGE_FAULTS_MAJ)
2367                 ttrace->pfmaj++;
2368         else
2369                 ttrace->pfmin++;
2370
2371         if (trace->summary_only)
2372                 goto out;
2373
2374         thread__find_symbol(thread, sample->cpumode, sample->ip, &al);
2375
2376         trace__fprintf_entry_head(trace, thread, 0, true, sample->time, trace->output);
2377
2378         fprintf(trace->output, "%sfault [",
2379                 evsel->attr.config == PERF_COUNT_SW_PAGE_FAULTS_MAJ ?
2380                 "maj" : "min");
2381
2382         print_location(trace->output, sample, &al, false, true);
2383
2384         fprintf(trace->output, "] => ");
2385
2386         thread__find_symbol(thread, sample->cpumode, sample->addr, &al);
2387
2388         if (!al.map) {
2389                 thread__find_symbol(thread, sample->cpumode, sample->addr, &al);
2390
2391                 if (al.map)
2392                         map_type = 'x';
2393                 else
2394                         map_type = '?';
2395         }
2396
2397         print_location(trace->output, sample, &al, true, false);
2398
2399         fprintf(trace->output, " (%c%c)\n", map_type, al.level);
2400
2401         if (callchain_ret > 0)
2402                 trace__fprintf_callchain(trace, sample);
2403         else if (callchain_ret < 0)
2404                 pr_err("Problem processing %s callchain, skipping...\n", perf_evsel__name(evsel));
2405
2406         ++trace->nr_events_printed;
2407 out:
2408         err = 0;
2409 out_put:
2410         thread__put(thread);
2411         return err;
2412 }
2413
2414 static void trace__set_base_time(struct trace *trace,
2415                                  struct perf_evsel *evsel,
2416                                  struct perf_sample *sample)
2417 {
2418         /*
2419          * BPF events were not setting PERF_SAMPLE_TIME, so be more robust
2420          * and don't use sample->time unconditionally, we may end up having
2421          * some other event in the future without PERF_SAMPLE_TIME for good
2422          * reason, i.e. we may not be interested in its timestamps, just in
2423          * it taking place, picking some piece of information when it
2424          * appears in our event stream (vfs_getname comes to mind).
2425          */
2426         if (trace->base_time == 0 && !trace->full_time &&
2427             (evsel->attr.sample_type & PERF_SAMPLE_TIME))
2428                 trace->base_time = sample->time;
2429 }
2430
2431 static int trace__process_sample(struct perf_tool *tool,
2432                                  union perf_event *event,
2433                                  struct perf_sample *sample,
2434                                  struct perf_evsel *evsel,
2435                                  struct machine *machine __maybe_unused)
2436 {
2437         struct trace *trace = container_of(tool, struct trace, tool);
2438         struct thread *thread;
2439         int err = 0;
2440
2441         tracepoint_handler handler = evsel->handler;
2442
2443         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2444         if (thread && thread__is_filtered(thread))
2445                 goto out;
2446
2447         trace__set_base_time(trace, evsel, sample);
2448
2449         if (handler) {
2450                 ++trace->nr_events;
2451                 handler(trace, evsel, event, sample);
2452         }
2453 out:
2454         thread__put(thread);
2455         return err;
2456 }
2457
2458 static int trace__record(struct trace *trace, int argc, const char **argv)
2459 {
2460         unsigned int rec_argc, i, j;
2461         const char **rec_argv;
2462         const char * const record_args[] = {
2463                 "record",
2464                 "-R",
2465                 "-m", "1024",
2466                 "-c", "1",
2467         };
2468
2469         const char * const sc_args[] = { "-e", };
2470         unsigned int sc_args_nr = ARRAY_SIZE(sc_args);
2471         const char * const majpf_args[] = { "-e", "major-faults" };
2472         unsigned int majpf_args_nr = ARRAY_SIZE(majpf_args);
2473         const char * const minpf_args[] = { "-e", "minor-faults" };
2474         unsigned int minpf_args_nr = ARRAY_SIZE(minpf_args);
2475
2476         /* +1 is for the event string below */
2477         rec_argc = ARRAY_SIZE(record_args) + sc_args_nr + 1 +
2478                 majpf_args_nr + minpf_args_nr + argc;
2479         rec_argv = calloc(rec_argc + 1, sizeof(char *));
2480
2481         if (rec_argv == NULL)
2482                 return -ENOMEM;
2483
2484         j = 0;
2485         for (i = 0; i < ARRAY_SIZE(record_args); i++)
2486                 rec_argv[j++] = record_args[i];
2487
2488         if (trace->trace_syscalls) {
2489                 for (i = 0; i < sc_args_nr; i++)
2490                         rec_argv[j++] = sc_args[i];
2491
2492                 /* event string may be different for older kernels - e.g., RHEL6 */
2493                 if (is_valid_tracepoint("raw_syscalls:sys_enter"))
2494                         rec_argv[j++] = "raw_syscalls:sys_enter,raw_syscalls:sys_exit";
2495                 else if (is_valid_tracepoint("syscalls:sys_enter"))
2496                         rec_argv[j++] = "syscalls:sys_enter,syscalls:sys_exit";
2497                 else {
2498                         pr_err("Neither raw_syscalls nor syscalls events exist.\n");
2499                         free(rec_argv);
2500                         return -1;
2501                 }
2502         }
2503
2504         if (trace->trace_pgfaults & TRACE_PFMAJ)
2505                 for (i = 0; i < majpf_args_nr; i++)
2506                         rec_argv[j++] = majpf_args[i];
2507
2508         if (trace->trace_pgfaults & TRACE_PFMIN)
2509                 for (i = 0; i < minpf_args_nr; i++)
2510                         rec_argv[j++] = minpf_args[i];
2511
2512         for (i = 0; i < (unsigned int)argc; i++)
2513                 rec_argv[j++] = argv[i];
2514
2515         return cmd_record(j, rec_argv);
2516 }
2517
2518 static size_t trace__fprintf_thread_summary(struct trace *trace, FILE *fp);
2519
2520 static bool perf_evlist__add_vfs_getname(struct perf_evlist *evlist)
2521 {
2522         bool found = false;
2523         struct perf_evsel *evsel, *tmp;
2524         struct parse_events_error err = { .idx = 0, };
2525         int ret = parse_events(evlist, "probe:vfs_getname*", &err);
2526
2527         if (ret)
2528                 return false;
2529
2530         evlist__for_each_entry_safe(evlist, evsel, tmp) {
2531                 if (!strstarts(perf_evsel__name(evsel), "probe:vfs_getname"))
2532                         continue;
2533
2534                 if (perf_evsel__field(evsel, "pathname")) {
2535                         evsel->handler = trace__vfs_getname;
2536                         found = true;
2537                         continue;
2538                 }
2539
2540                 list_del_init(&evsel->node);
2541                 evsel->evlist = NULL;
2542                 perf_evsel__delete(evsel);
2543         }
2544
2545         return found;
2546 }
2547
2548 static struct perf_evsel *perf_evsel__new_pgfault(u64 config)
2549 {
2550         struct perf_evsel *evsel;
2551         struct perf_event_attr attr = {
2552                 .type = PERF_TYPE_SOFTWARE,
2553                 .mmap_data = 1,
2554         };
2555
2556         attr.config = config;
2557         attr.sample_period = 1;
2558
2559         event_attr_init(&attr);
2560
2561         evsel = perf_evsel__new(&attr);
2562         if (evsel)
2563                 evsel->handler = trace__pgfault;
2564
2565         return evsel;
2566 }
2567
2568 static void trace__handle_event(struct trace *trace, union perf_event *event, struct perf_sample *sample)
2569 {
2570         const u32 type = event->header.type;
2571         struct perf_evsel *evsel;
2572
2573         if (type != PERF_RECORD_SAMPLE) {
2574                 trace__process_event(trace, trace->host, event, sample);
2575                 return;
2576         }
2577
2578         evsel = perf_evlist__id2evsel(trace->evlist, sample->id);
2579         if (evsel == NULL) {
2580                 fprintf(trace->output, "Unknown tp ID %" PRIu64 ", skipping...\n", sample->id);
2581                 return;
2582         }
2583
2584         trace__set_base_time(trace, evsel, sample);
2585
2586         if (evsel->attr.type == PERF_TYPE_TRACEPOINT &&
2587             sample->raw_data == NULL) {
2588                 fprintf(trace->output, "%s sample with no payload for tid: %d, cpu %d, raw_size=%d, skipping...\n",
2589                        perf_evsel__name(evsel), sample->tid,
2590                        sample->cpu, sample->raw_size);
2591         } else {
2592                 tracepoint_handler handler = evsel->handler;
2593                 handler(trace, evsel, event, sample);
2594         }
2595
2596         if (trace->nr_events_printed >= trace->max_events && trace->max_events != ULONG_MAX)
2597                 interrupted = true;
2598 }
2599
2600 static int trace__add_syscall_newtp(struct trace *trace)
2601 {
2602         int ret = -1;
2603         struct perf_evlist *evlist = trace->evlist;
2604         struct perf_evsel *sys_enter, *sys_exit;
2605
2606         sys_enter = perf_evsel__raw_syscall_newtp("sys_enter", trace__sys_enter);
2607         if (sys_enter == NULL)
2608                 goto out;
2609
2610         if (perf_evsel__init_sc_tp_ptr_field(sys_enter, args))
2611                 goto out_delete_sys_enter;
2612
2613         sys_exit = perf_evsel__raw_syscall_newtp("sys_exit", trace__sys_exit);
2614         if (sys_exit == NULL)
2615                 goto out_delete_sys_enter;
2616
2617         if (perf_evsel__init_sc_tp_uint_field(sys_exit, ret))
2618                 goto out_delete_sys_exit;
2619
2620         perf_evsel__config_callchain(sys_enter, &trace->opts, &callchain_param);
2621         perf_evsel__config_callchain(sys_exit, &trace->opts, &callchain_param);
2622
2623         perf_evlist__add(evlist, sys_enter);
2624         perf_evlist__add(evlist, sys_exit);
2625
2626         if (callchain_param.enabled && !trace->kernel_syscallchains) {
2627                 /*
2628                  * We're interested only in the user space callchain
2629                  * leading to the syscall, allow overriding that for
2630                  * debugging reasons using --kernel_syscall_callchains
2631                  */
2632                 sys_exit->attr.exclude_callchain_kernel = 1;
2633         }
2634
2635         trace->syscalls.events.sys_enter = sys_enter;
2636         trace->syscalls.events.sys_exit  = sys_exit;
2637
2638         ret = 0;
2639 out:
2640         return ret;
2641
2642 out_delete_sys_exit:
2643         perf_evsel__delete_priv(sys_exit);
2644 out_delete_sys_enter:
2645         perf_evsel__delete_priv(sys_enter);
2646         goto out;
2647 }
2648
2649 static int trace__set_ev_qualifier_tp_filter(struct trace *trace)
2650 {
2651         int err = -1;
2652         struct perf_evsel *sys_exit;
2653         char *filter = asprintf_expr_inout_ints("id", !trace->not_ev_qualifier,
2654                                                 trace->ev_qualifier_ids.nr,
2655                                                 trace->ev_qualifier_ids.entries);
2656
2657         if (filter == NULL)
2658                 goto out_enomem;
2659
2660         if (!perf_evsel__append_tp_filter(trace->syscalls.events.sys_enter,
2661                                           filter)) {
2662                 sys_exit = trace->syscalls.events.sys_exit;
2663                 err = perf_evsel__append_tp_filter(sys_exit, filter);
2664         }
2665
2666         free(filter);
2667 out:
2668         return err;
2669 out_enomem:
2670         errno = ENOMEM;
2671         goto out;
2672 }
2673
2674 #ifdef HAVE_LIBBPF_SUPPORT
2675 static int trace__set_ev_qualifier_bpf_filter(struct trace *trace)
2676 {
2677         int fd = bpf_map__fd(trace->syscalls.map);
2678         struct bpf_map_syscall_entry value = {
2679                 .enabled = !trace->not_ev_qualifier,
2680         };
2681         int err = 0;
2682         size_t i;
2683
2684         for (i = 0; i < trace->ev_qualifier_ids.nr; ++i) {
2685                 int key = trace->ev_qualifier_ids.entries[i];
2686
2687                 err = bpf_map_update_elem(fd, &key, &value, BPF_EXIST);
2688                 if (err)
2689                         break;
2690         }
2691
2692         return err;
2693 }
2694
2695 static int __trace__init_syscalls_bpf_map(struct trace *trace, bool enabled)
2696 {
2697         int fd = bpf_map__fd(trace->syscalls.map);
2698         struct bpf_map_syscall_entry value = {
2699                 .enabled = enabled,
2700         };
2701         int err = 0, key;
2702
2703         for (key = 0; key < trace->sctbl->syscalls.nr_entries; ++key) {
2704                 err = bpf_map_update_elem(fd, &key, &value, BPF_ANY);
2705                 if (err)
2706                         break;
2707         }
2708
2709         return err;
2710 }
2711
2712 static int trace__init_syscalls_bpf_map(struct trace *trace)
2713 {
2714         bool enabled = true;
2715
2716         if (trace->ev_qualifier_ids.nr)
2717                 enabled = trace->not_ev_qualifier;
2718
2719         return __trace__init_syscalls_bpf_map(trace, enabled);
2720 }
2721 #else
2722 static int trace__set_ev_qualifier_bpf_filter(struct trace *trace __maybe_unused)
2723 {
2724         return 0;
2725 }
2726
2727 static int trace__init_syscalls_bpf_map(struct trace *trace __maybe_unused)
2728 {
2729         return 0;
2730 }
2731 #endif // HAVE_LIBBPF_SUPPORT
2732
2733 static int trace__set_ev_qualifier_filter(struct trace *trace)
2734 {
2735         if (trace->syscalls.map)
2736                 return trace__set_ev_qualifier_bpf_filter(trace);
2737         if (trace->syscalls.events.sys_enter)
2738                 return trace__set_ev_qualifier_tp_filter(trace);
2739         return 0;
2740 }
2741
2742 static int bpf_map__set_filter_pids(struct bpf_map *map __maybe_unused,
2743                                     size_t npids __maybe_unused, pid_t *pids __maybe_unused)
2744 {
2745         int err = 0;
2746 #ifdef HAVE_LIBBPF_SUPPORT
2747         bool value = true;
2748         int map_fd = bpf_map__fd(map);
2749         size_t i;
2750
2751         for (i = 0; i < npids; ++i) {
2752                 err = bpf_map_update_elem(map_fd, &pids[i], &value, BPF_ANY);
2753                 if (err)
2754                         break;
2755         }
2756 #endif
2757         return err;
2758 }
2759
2760 static int trace__set_filter_loop_pids(struct trace *trace)
2761 {
2762         unsigned int nr = 1, err;
2763         pid_t pids[32] = {
2764                 getpid(),
2765         };
2766         struct thread *thread = machine__find_thread(trace->host, pids[0], pids[0]);
2767
2768         while (thread && nr < ARRAY_SIZE(pids)) {
2769                 struct thread *parent = machine__find_thread(trace->host, thread->ppid, thread->ppid);
2770
2771                 if (parent == NULL)
2772                         break;
2773
2774                 if (!strcmp(thread__comm_str(parent), "sshd") ||
2775                     strstarts(thread__comm_str(parent), "gnome-terminal")) {
2776                         pids[nr++] = parent->tid;
2777                         break;
2778                 }
2779                 thread = parent;
2780         }
2781
2782         err = perf_evlist__set_tp_filter_pids(trace->evlist, nr, pids);
2783         if (!err && trace->filter_pids.map)
2784                 err = bpf_map__set_filter_pids(trace->filter_pids.map, nr, pids);
2785
2786         return err;
2787 }
2788
2789 static int trace__set_filter_pids(struct trace *trace)
2790 {
2791         int err = 0;
2792         /*
2793          * Better not use !target__has_task() here because we need to cover the
2794          * case where no threads were specified in the command line, but a
2795          * workload was, and in that case we will fill in the thread_map when
2796          * we fork the workload in perf_evlist__prepare_workload.
2797          */
2798         if (trace->filter_pids.nr > 0) {
2799                 err = perf_evlist__set_tp_filter_pids(trace->evlist, trace->filter_pids.nr,
2800                                                       trace->filter_pids.entries);
2801                 if (!err && trace->filter_pids.map) {
2802                         err = bpf_map__set_filter_pids(trace->filter_pids.map, trace->filter_pids.nr,
2803                                                        trace->filter_pids.entries);
2804                 }
2805         } else if (thread_map__pid(trace->evlist->threads, 0) == -1) {
2806                 err = trace__set_filter_loop_pids(trace);
2807         }
2808
2809         return err;
2810 }
2811
2812 static int __trace__deliver_event(struct trace *trace, union perf_event *event)
2813 {
2814         struct perf_evlist *evlist = trace->evlist;
2815         struct perf_sample sample;
2816         int err;
2817
2818         err = perf_evlist__parse_sample(evlist, event, &sample);
2819         if (err)
2820                 fprintf(trace->output, "Can't parse sample, err = %d, skipping...\n", err);
2821         else
2822                 trace__handle_event(trace, event, &sample);
2823
2824         return 0;
2825 }
2826
2827 static int __trace__flush_events(struct trace *trace)
2828 {
2829         u64 first = ordered_events__first_time(&trace->oe.data);
2830         u64 flush = trace->oe.last - NSEC_PER_SEC;
2831
2832         /* Is there some thing to flush.. */
2833         if (first && first < flush)
2834                 return ordered_events__flush_time(&trace->oe.data, flush);
2835
2836         return 0;
2837 }
2838
2839 static int trace__flush_events(struct trace *trace)
2840 {
2841         return !trace->sort_events ? 0 : __trace__flush_events(trace);
2842 }
2843
2844 static int trace__deliver_event(struct trace *trace, union perf_event *event)
2845 {
2846         int err;
2847
2848         if (!trace->sort_events)
2849                 return __trace__deliver_event(trace, event);
2850
2851         err = perf_evlist__parse_sample_timestamp(trace->evlist, event, &trace->oe.last);
2852         if (err && err != -1)
2853                 return err;
2854
2855         err = ordered_events__queue(&trace->oe.data, event, trace->oe.last, 0);
2856         if (err)
2857                 return err;
2858
2859         return trace__flush_events(trace);
2860 }
2861
2862 static int ordered_events__deliver_event(struct ordered_events *oe,
2863                                          struct ordered_event *event)
2864 {
2865         struct trace *trace = container_of(oe, struct trace, oe.data);
2866
2867         return __trace__deliver_event(trace, event->event);
2868 }
2869
2870 static int trace__run(struct trace *trace, int argc, const char **argv)
2871 {
2872         struct perf_evlist *evlist = trace->evlist;
2873         struct perf_evsel *evsel, *pgfault_maj = NULL, *pgfault_min = NULL;
2874         int err = -1, i;
2875         unsigned long before;
2876         const bool forks = argc > 0;
2877         bool draining = false;
2878
2879         trace->live = true;
2880
2881         if (!trace->raw_augmented_syscalls) {
2882                 if (trace->trace_syscalls && trace__add_syscall_newtp(trace))
2883                         goto out_error_raw_syscalls;
2884
2885                 if (trace->trace_syscalls)
2886                         trace->vfs_getname = perf_evlist__add_vfs_getname(evlist);
2887         }
2888
2889         if ((trace->trace_pgfaults & TRACE_PFMAJ)) {
2890                 pgfault_maj = perf_evsel__new_pgfault(PERF_COUNT_SW_PAGE_FAULTS_MAJ);
2891                 if (pgfault_maj == NULL)
2892                         goto out_error_mem;
2893                 perf_evsel__config_callchain(pgfault_maj, &trace->opts, &callchain_param);
2894                 perf_evlist__add(evlist, pgfault_maj);
2895         }
2896
2897         if ((trace->trace_pgfaults & TRACE_PFMIN)) {
2898                 pgfault_min = perf_evsel__new_pgfault(PERF_COUNT_SW_PAGE_FAULTS_MIN);
2899                 if (pgfault_min == NULL)
2900                         goto out_error_mem;
2901                 perf_evsel__config_callchain(pgfault_min, &trace->opts, &callchain_param);
2902                 perf_evlist__add(evlist, pgfault_min);
2903         }
2904
2905         if (trace->sched &&
2906             perf_evlist__add_newtp(evlist, "sched", "sched_stat_runtime",
2907                                    trace__sched_stat_runtime))
2908                 goto out_error_sched_stat_runtime;
2909
2910         /*
2911          * If a global cgroup was set, apply it to all the events without an
2912          * explicit cgroup. I.e.:
2913          *
2914          *      trace -G A -e sched:*switch
2915          *
2916          * Will set all raw_syscalls:sys_{enter,exit}, pgfault, vfs_getname, etc
2917          * _and_ sched:sched_switch to the 'A' cgroup, while:
2918          *
2919          * trace -e sched:*switch -G A
2920          *
2921          * will only set the sched:sched_switch event to the 'A' cgroup, all the
2922          * other events (raw_syscalls:sys_{enter,exit}, etc are left "without"
2923          * a cgroup (on the root cgroup, sys wide, etc).
2924          *
2925          * Multiple cgroups:
2926          *
2927          * trace -G A -e sched:*switch -G B
2928          *
2929          * the syscall ones go to the 'A' cgroup, the sched:sched_switch goes
2930          * to the 'B' cgroup.
2931          *
2932          * evlist__set_default_cgroup() grabs a reference of the passed cgroup
2933          * only for the evsels still without a cgroup, i.e. evsel->cgroup == NULL.
2934          */
2935         if (trace->cgroup)
2936                 evlist__set_default_cgroup(trace->evlist, trace->cgroup);
2937
2938         err = perf_evlist__create_maps(evlist, &trace->opts.target);
2939         if (err < 0) {
2940                 fprintf(trace->output, "Problems parsing the target to trace, check your options!\n");
2941                 goto out_delete_evlist;
2942         }
2943
2944         err = trace__symbols_init(trace, evlist);
2945         if (err < 0) {
2946                 fprintf(trace->output, "Problems initializing symbol libraries!\n");
2947                 goto out_delete_evlist;
2948         }
2949
2950         perf_evlist__config(evlist, &trace->opts, &callchain_param);
2951
2952         signal(SIGCHLD, sig_handler);
2953         signal(SIGINT, sig_handler);
2954
2955         if (forks) {
2956                 err = perf_evlist__prepare_workload(evlist, &trace->opts.target,
2957                                                     argv, false, NULL);
2958                 if (err < 0) {
2959                         fprintf(trace->output, "Couldn't run the workload!\n");
2960                         goto out_delete_evlist;
2961                 }
2962         }
2963
2964         err = perf_evlist__open(evlist);
2965         if (err < 0)
2966                 goto out_error_open;
2967
2968         err = bpf__apply_obj_config();
2969         if (err) {
2970                 char errbuf[BUFSIZ];
2971
2972                 bpf__strerror_apply_obj_config(err, errbuf, sizeof(errbuf));
2973                 pr_err("ERROR: Apply config to BPF failed: %s\n",
2974                          errbuf);
2975                 goto out_error_open;
2976         }
2977
2978         err = trace__set_filter_pids(trace);
2979         if (err < 0)
2980                 goto out_error_mem;
2981
2982         if (trace->syscalls.map)
2983                 trace__init_syscalls_bpf_map(trace);
2984
2985         if (trace->ev_qualifier_ids.nr > 0) {
2986                 err = trace__set_ev_qualifier_filter(trace);
2987                 if (err < 0)
2988                         goto out_errno;
2989
2990                 if (trace->syscalls.events.sys_exit) {
2991                         pr_debug("event qualifier tracepoint filter: %s\n",
2992                                  trace->syscalls.events.sys_exit->filter);
2993                 }
2994         }
2995
2996         err = perf_evlist__apply_filters(evlist, &evsel);
2997         if (err < 0)
2998                 goto out_error_apply_filters;
2999
3000         err = perf_evlist__mmap(evlist, trace->opts.mmap_pages);
3001         if (err < 0)
3002                 goto out_error_mmap;
3003
3004         if (!target__none(&trace->opts.target) && !trace->opts.initial_delay)
3005                 perf_evlist__enable(evlist);
3006
3007         if (forks)
3008                 perf_evlist__start_workload(evlist);
3009
3010         if (trace->opts.initial_delay) {
3011                 usleep(trace->opts.initial_delay * 1000);
3012                 perf_evlist__enable(evlist);
3013         }
3014
3015         trace->multiple_threads = thread_map__pid(evlist->threads, 0) == -1 ||
3016                                   evlist->threads->nr > 1 ||
3017                                   perf_evlist__first(evlist)->attr.inherit;
3018
3019         /*
3020          * Now that we already used evsel->attr to ask the kernel to setup the
3021          * events, lets reuse evsel->attr.sample_max_stack as the limit in
3022          * trace__resolve_callchain(), allowing per-event max-stack settings
3023          * to override an explicitly set --max-stack global setting.
3024          */
3025         evlist__for_each_entry(evlist, evsel) {
3026                 if (evsel__has_callchain(evsel) &&
3027                     evsel->attr.sample_max_stack == 0)
3028                         evsel->attr.sample_max_stack = trace->max_stack;
3029         }
3030 again:
3031         before = trace->nr_events;
3032
3033         for (i = 0; i < evlist->nr_mmaps; i++) {
3034                 union perf_event *event;
3035                 struct perf_mmap *md;
3036
3037                 md = &evlist->mmap[i];
3038                 if (perf_mmap__read_init(md) < 0)
3039                         continue;
3040
3041                 while ((event = perf_mmap__read_event(md)) != NULL) {
3042                         ++trace->nr_events;
3043
3044                         err = trace__deliver_event(trace, event);
3045                         if (err)
3046                                 goto out_disable;
3047
3048                         perf_mmap__consume(md);
3049
3050                         if (interrupted)
3051                                 goto out_disable;
3052
3053                         if (done && !draining) {
3054                                 perf_evlist__disable(evlist);
3055                                 draining = true;
3056                         }
3057                 }
3058                 perf_mmap__read_done(md);
3059         }
3060
3061         if (trace->nr_events == before) {
3062                 int timeout = done ? 100 : -1;
3063
3064                 if (!draining && perf_evlist__poll(evlist, timeout) > 0) {
3065                         if (perf_evlist__filter_pollfd(evlist, POLLERR | POLLHUP | POLLNVAL) == 0)
3066                                 draining = true;
3067
3068                         goto again;
3069                 } else {
3070                         if (trace__flush_events(trace))
3071                                 goto out_disable;
3072                 }
3073         } else {
3074                 goto again;
3075         }
3076
3077 out_disable:
3078         thread__zput(trace->current);
3079
3080         perf_evlist__disable(evlist);
3081
3082         if (trace->sort_events)
3083                 ordered_events__flush(&trace->oe.data, OE_FLUSH__FINAL);
3084
3085         if (!err) {
3086                 if (trace->summary)
3087                         trace__fprintf_thread_summary(trace, trace->output);
3088
3089                 if (trace->show_tool_stats) {
3090                         fprintf(trace->output, "Stats:\n "
3091                                                " vfs_getname : %" PRIu64 "\n"
3092                                                " proc_getname: %" PRIu64 "\n",
3093                                 trace->stats.vfs_getname,
3094                                 trace->stats.proc_getname);
3095                 }
3096         }
3097
3098 out_delete_evlist:
3099         trace__symbols__exit(trace);
3100
3101         perf_evlist__delete(evlist);
3102         cgroup__put(trace->cgroup);
3103         trace->evlist = NULL;
3104         trace->live = false;
3105         return err;
3106 {
3107         char errbuf[BUFSIZ];
3108
3109 out_error_sched_stat_runtime:
3110         tracing_path__strerror_open_tp(errno, errbuf, sizeof(errbuf), "sched", "sched_stat_runtime");
3111         goto out_error;
3112
3113 out_error_raw_syscalls:
3114         tracing_path__strerror_open_tp(errno, errbuf, sizeof(errbuf), "raw_syscalls", "sys_(enter|exit)");
3115         goto out_error;
3116
3117 out_error_mmap:
3118         perf_evlist__strerror_mmap(evlist, errno, errbuf, sizeof(errbuf));
3119         goto out_error;
3120
3121 out_error_open:
3122         perf_evlist__strerror_open(evlist, errno, errbuf, sizeof(errbuf));
3123
3124 out_error:
3125         fprintf(trace->output, "%s\n", errbuf);
3126         goto out_delete_evlist;
3127
3128 out_error_apply_filters:
3129         fprintf(trace->output,
3130                 "Failed to set filter \"%s\" on event %s with %d (%s)\n",
3131                 evsel->filter, perf_evsel__name(evsel), errno,
3132                 str_error_r(errno, errbuf, sizeof(errbuf)));
3133         goto out_delete_evlist;
3134 }
3135 out_error_mem:
3136         fprintf(trace->output, "Not enough memory to run!\n");
3137         goto out_delete_evlist;
3138
3139 out_errno:
3140         fprintf(trace->output, "errno=%d,%s\n", errno, strerror(errno));
3141         goto out_delete_evlist;
3142 }
3143
3144 static int trace__replay(struct trace *trace)
3145 {
3146         const struct perf_evsel_str_handler handlers[] = {
3147                 { "probe:vfs_getname",       trace__vfs_getname, },
3148         };
3149         struct perf_data data = {
3150                 .file      = {
3151                         .path = input_name,
3152                 },
3153                 .mode      = PERF_DATA_MODE_READ,
3154                 .force     = trace->force,
3155         };
3156         struct perf_session *session;
3157         struct perf_evsel *evsel;
3158         int err = -1;
3159
3160         trace->tool.sample        = trace__process_sample;
3161         trace->tool.mmap          = perf_event__process_mmap;
3162         trace->tool.mmap2         = perf_event__process_mmap2;
3163         trace->tool.comm          = perf_event__process_comm;
3164         trace->tool.exit          = perf_event__process_exit;
3165         trace->tool.fork          = perf_event__process_fork;
3166         trace->tool.attr          = perf_event__process_attr;
3167         trace->tool.tracing_data  = perf_event__process_tracing_data;
3168         trace->tool.build_id      = perf_event__process_build_id;
3169         trace->tool.namespaces    = perf_event__process_namespaces;
3170
3171         trace->tool.ordered_events = true;
3172         trace->tool.ordering_requires_timestamps = true;
3173
3174         /* add tid to output */
3175         trace->multiple_threads = true;
3176
3177         session = perf_session__new(&data, false, &trace->tool);
3178         if (session == NULL)
3179                 return -1;
3180
3181         if (trace->opts.target.pid)
3182                 symbol_conf.pid_list_str = strdup(trace->opts.target.pid);
3183
3184         if (trace->opts.target.tid)
3185                 symbol_conf.tid_list_str = strdup(trace->opts.target.tid);
3186
3187         if (symbol__init(&session->header.env) < 0)
3188                 goto out;
3189
3190         trace->host = &session->machines.host;
3191
3192         err = perf_session__set_tracepoints_handlers(session, handlers);
3193         if (err)
3194                 goto out;
3195
3196         evsel = perf_evlist__find_tracepoint_by_name(session->evlist,
3197                                                      "raw_syscalls:sys_enter");
3198         /* older kernels have syscalls tp versus raw_syscalls */
3199         if (evsel == NULL)
3200                 evsel = perf_evlist__find_tracepoint_by_name(session->evlist,
3201                                                              "syscalls:sys_enter");
3202
3203         if (evsel &&
3204             (perf_evsel__init_raw_syscall_tp(evsel, trace__sys_enter) < 0 ||
3205             perf_evsel__init_sc_tp_ptr_field(evsel, args))) {
3206                 pr_err("Error during initialize raw_syscalls:sys_enter event\n");
3207                 goto out;
3208         }
3209
3210         evsel = perf_evlist__find_tracepoint_by_name(session->evlist,
3211                                                      "raw_syscalls:sys_exit");
3212         if (evsel == NULL)
3213                 evsel = perf_evlist__find_tracepoint_by_name(session->evlist,
3214                                                              "syscalls:sys_exit");
3215         if (evsel &&
3216             (perf_evsel__init_raw_syscall_tp(evsel, trace__sys_exit) < 0 ||
3217             perf_evsel__init_sc_tp_uint_field(evsel, ret))) {
3218                 pr_err("Error during initialize raw_syscalls:sys_exit event\n");
3219                 goto out;
3220         }
3221
3222         evlist__for_each_entry(session->evlist, evsel) {
3223                 if (evsel->attr.type == PERF_TYPE_SOFTWARE &&
3224                     (evsel->attr.config == PERF_COUNT_SW_PAGE_FAULTS_MAJ ||
3225                      evsel->attr.config == PERF_COUNT_SW_PAGE_FAULTS_MIN ||
3226                      evsel->attr.config == PERF_COUNT_SW_PAGE_FAULTS))
3227                         evsel->handler = trace__pgfault;
3228         }
3229
3230         setup_pager();
3231
3232         err = perf_session__process_events(session);
3233         if (err)
3234                 pr_err("Failed to process events, error %d", err);
3235
3236         else if (trace->summary)
3237                 trace__fprintf_thread_summary(trace, trace->output);
3238
3239 out:
3240         perf_session__delete(session);
3241
3242         return err;
3243 }
3244
3245 static size_t trace__fprintf_threads_header(FILE *fp)
3246 {
3247         size_t printed;
3248
3249         printed  = fprintf(fp, "\n Summary of events:\n\n");
3250
3251         return printed;
3252 }
3253
3254 DEFINE_RESORT_RB(syscall_stats, a->msecs > b->msecs,
3255         struct stats    *stats;
3256         double          msecs;
3257         int             syscall;
3258 )
3259 {
3260         struct int_node *source = rb_entry(nd, struct int_node, rb_node);
3261         struct stats *stats = source->priv;
3262
3263         entry->syscall = source->i;
3264         entry->stats   = stats;
3265         entry->msecs   = stats ? (u64)stats->n * (avg_stats(stats) / NSEC_PER_MSEC) : 0;
3266 }
3267
3268 static size_t thread__dump_stats(struct thread_trace *ttrace,
3269                                  struct trace *trace, FILE *fp)
3270 {
3271         size_t printed = 0;
3272         struct syscall *sc;
3273         struct rb_node *nd;
3274         DECLARE_RESORT_RB_INTLIST(syscall_stats, ttrace->syscall_stats);
3275
3276         if (syscall_stats == NULL)
3277                 return 0;
3278
3279         printed += fprintf(fp, "\n");
3280
3281         printed += fprintf(fp, "   syscall            calls    total       min       avg       max      stddev\n");
3282         printed += fprintf(fp, "                               (msec)    (msec)    (msec)    (msec)        (%%)\n");
3283         printed += fprintf(fp, "   --------------- -------- --------- --------- --------- ---------     ------\n");
3284
3285         resort_rb__for_each_entry(nd, syscall_stats) {
3286                 struct stats *stats = syscall_stats_entry->stats;
3287                 if (stats) {
3288                         double min = (double)(stats->min) / NSEC_PER_MSEC;
3289                         double max = (double)(stats->max) / NSEC_PER_MSEC;
3290                         double avg = avg_stats(stats);
3291                         double pct;
3292                         u64 n = (u64) stats->n;
3293
3294                         pct = avg ? 100.0 * stddev_stats(stats)/avg : 0.0;
3295                         avg /= NSEC_PER_MSEC;
3296
3297                         sc = &trace->syscalls.table[syscall_stats_entry->syscall];
3298                         printed += fprintf(fp, "   %-15s", sc->name);
3299                         printed += fprintf(fp, " %8" PRIu64 " %9.3f %9.3f %9.3f",
3300                                            n, syscall_stats_entry->msecs, min, avg);
3301                         printed += fprintf(fp, " %9.3f %9.2f%%\n", max, pct);
3302                 }
3303         }
3304
3305         resort_rb__delete(syscall_stats);
3306         printed += fprintf(fp, "\n\n");
3307
3308         return printed;
3309 }
3310
3311 static size_t trace__fprintf_thread(FILE *fp, struct thread *thread, struct trace *trace)
3312 {
3313         size_t printed = 0;
3314         struct thread_trace *ttrace = thread__priv(thread);
3315         double ratio;
3316
3317         if (ttrace == NULL)
3318                 return 0;
3319
3320         ratio = (double)ttrace->nr_events / trace->nr_events * 100.0;
3321
3322         printed += fprintf(fp, " %s (%d), ", thread__comm_str(thread), thread->tid);
3323         printed += fprintf(fp, "%lu events, ", ttrace->nr_events);
3324         printed += fprintf(fp, "%.1f%%", ratio);
3325         if (ttrace->pfmaj)
3326                 printed += fprintf(fp, ", %lu majfaults", ttrace->pfmaj);
3327         if (ttrace->pfmin)
3328                 printed += fprintf(fp, ", %lu minfaults", ttrace->pfmin);
3329         if (trace->sched)
3330                 printed += fprintf(fp, ", %.3f msec\n", ttrace->runtime_ms);
3331         else if (fputc('\n', fp) != EOF)
3332                 ++printed;
3333
3334         printed += thread__dump_stats(ttrace, trace, fp);
3335
3336         return printed;
3337 }
3338
3339 static unsigned long thread__nr_events(struct thread_trace *ttrace)
3340 {
3341         return ttrace ? ttrace->nr_events : 0;
3342 }
3343
3344 DEFINE_RESORT_RB(threads, (thread__nr_events(a->thread->priv) < thread__nr_events(b->thread->priv)),
3345         struct thread *thread;
3346 )
3347 {
3348         entry->thread = rb_entry(nd, struct thread, rb_node);
3349 }
3350
3351 static size_t trace__fprintf_thread_summary(struct trace *trace, FILE *fp)
3352 {
3353         size_t printed = trace__fprintf_threads_header(fp);
3354         struct rb_node *nd;
3355         int i;
3356
3357         for (i = 0; i < THREADS__TABLE_SIZE; i++) {
3358                 DECLARE_RESORT_RB_MACHINE_THREADS(threads, trace->host, i);
3359
3360                 if (threads == NULL) {
3361                         fprintf(fp, "%s", "Error sorting output by nr_events!\n");
3362                         return 0;
3363                 }
3364
3365                 resort_rb__for_each_entry(nd, threads)
3366                         printed += trace__fprintf_thread(fp, threads_entry->thread, trace);
3367
3368                 resort_rb__delete(threads);
3369         }
3370         return printed;
3371 }
3372
3373 static int trace__set_duration(const struct option *opt, const char *str,
3374                                int unset __maybe_unused)
3375 {
3376         struct trace *trace = opt->value;
3377
3378         trace->duration_filter = atof(str);
3379         return 0;
3380 }
3381
3382 static int trace__set_filter_pids_from_option(const struct option *opt, const char *str,
3383                                               int unset __maybe_unused)
3384 {
3385         int ret = -1;
3386         size_t i;
3387         struct trace *trace = opt->value;
3388         /*
3389          * FIXME: introduce a intarray class, plain parse csv and create a
3390          * { int nr, int entries[] } struct...
3391          */
3392         struct intlist *list = intlist__new(str);
3393
3394         if (list == NULL)
3395                 return -1;
3396
3397         i = trace->filter_pids.nr = intlist__nr_entries(list) + 1;
3398         trace->filter_pids.entries = calloc(i, sizeof(pid_t));
3399
3400         if (trace->filter_pids.entries == NULL)
3401                 goto out;
3402
3403         trace->filter_pids.entries[0] = getpid();
3404
3405         for (i = 1; i < trace->filter_pids.nr; ++i)
3406                 trace->filter_pids.entries[i] = intlist__entry(list, i - 1)->i;
3407
3408         intlist__delete(list);
3409         ret = 0;
3410 out:
3411         return ret;
3412 }
3413
3414 static int trace__open_output(struct trace *trace, const char *filename)
3415 {
3416         struct stat st;
3417
3418         if (!stat(filename, &st) && st.st_size) {
3419                 char oldname[PATH_MAX];
3420
3421                 scnprintf(oldname, sizeof(oldname), "%s.old", filename);
3422                 unlink(oldname);
3423                 rename(filename, oldname);
3424         }
3425
3426         trace->output = fopen(filename, "w");
3427
3428         return trace->output == NULL ? -errno : 0;
3429 }
3430
3431 static int parse_pagefaults(const struct option *opt, const char *str,
3432                             int unset __maybe_unused)
3433 {
3434         int *trace_pgfaults = opt->value;
3435
3436         if (strcmp(str, "all") == 0)
3437                 *trace_pgfaults |= TRACE_PFMAJ | TRACE_PFMIN;
3438         else if (strcmp(str, "maj") == 0)
3439                 *trace_pgfaults |= TRACE_PFMAJ;
3440         else if (strcmp(str, "min") == 0)
3441                 *trace_pgfaults |= TRACE_PFMIN;
3442         else
3443                 return -1;
3444
3445         return 0;
3446 }
3447
3448 static void evlist__set_evsel_handler(struct perf_evlist *evlist, void *handler)
3449 {
3450         struct perf_evsel *evsel;
3451
3452         evlist__for_each_entry(evlist, evsel)
3453                 evsel->handler = handler;
3454 }
3455
3456 static int evlist__set_syscall_tp_fields(struct perf_evlist *evlist)
3457 {
3458         struct perf_evsel *evsel;
3459
3460         evlist__for_each_entry(evlist, evsel) {
3461                 if (evsel->priv || !evsel->tp_format)
3462                         continue;
3463
3464                 if (strcmp(evsel->tp_format->system, "syscalls"))
3465                         continue;
3466
3467                 if (perf_evsel__init_syscall_tp(evsel))
3468                         return -1;
3469
3470                 if (!strncmp(evsel->tp_format->name, "sys_enter_", 10)) {
3471                         struct syscall_tp *sc = evsel->priv;
3472
3473                         if (__tp_field__init_ptr(&sc->args, sc->id.offset + sizeof(u64)))
3474                                 return -1;
3475                 } else if (!strncmp(evsel->tp_format->name, "sys_exit_", 9)) {
3476                         struct syscall_tp *sc = evsel->priv;
3477
3478                         if (__tp_field__init_uint(&sc->ret, sizeof(u64), sc->id.offset + sizeof(u64), evsel->needs_swap))
3479                                 return -1;
3480                 }
3481         }
3482
3483         return 0;
3484 }
3485
3486 /*
3487  * XXX: Hackish, just splitting the combined -e+--event (syscalls
3488  * (raw_syscalls:{sys_{enter,exit}} + events (tracepoints, HW, SW, etc) to use
3489  * existing facilities unchanged (trace->ev_qualifier + parse_options()).
3490  *
3491  * It'd be better to introduce a parse_options() variant that would return a
3492  * list with the terms it didn't match to an event...
3493  */
3494 static int trace__parse_events_option(const struct option *opt, const char *str,
3495                                       int unset __maybe_unused)
3496 {
3497         struct trace *trace = (struct trace *)opt->value;
3498         const char *s = str;
3499         char *sep = NULL, *lists[2] = { NULL, NULL, };
3500         int len = strlen(str) + 1, err = -1, list, idx;
3501         char *strace_groups_dir = system_path(STRACE_GROUPS_DIR);
3502         char group_name[PATH_MAX];
3503         struct syscall_fmt *fmt;
3504
3505         if (strace_groups_dir == NULL)
3506                 return -1;
3507
3508         if (*s == '!') {
3509                 ++s;
3510                 trace->not_ev_qualifier = true;
3511         }
3512
3513         while (1) {
3514                 if ((sep = strchr(s, ',')) != NULL)
3515                         *sep = '\0';
3516
3517                 list = 0;
3518                 if (syscalltbl__id(trace->sctbl, s) >= 0 ||
3519                     syscalltbl__strglobmatch_first(trace->sctbl, s, &idx) >= 0) {
3520                         list = 1;
3521                         goto do_concat;
3522                 }
3523
3524                 fmt = syscall_fmt__find_by_alias(s);
3525                 if (fmt != NULL) {
3526                         list = 1;
3527                         s = fmt->name;
3528                 } else {
3529                         path__join(group_name, sizeof(group_name), strace_groups_dir, s);
3530                         if (access(group_name, R_OK) == 0)
3531                                 list = 1;
3532                 }
3533 do_concat:
3534                 if (lists[list]) {
3535                         sprintf(lists[list] + strlen(lists[list]), ",%s", s);
3536                 } else {
3537                         lists[list] = malloc(len);
3538                         if (lists[list] == NULL)
3539                                 goto out;
3540                         strcpy(lists[list], s);
3541                 }
3542
3543                 if (!sep)
3544                         break;
3545
3546                 *sep = ',';
3547                 s = sep + 1;
3548         }
3549
3550         if (lists[1] != NULL) {
3551                 struct strlist_config slist_config = {
3552                         .dirname = strace_groups_dir,
3553                 };
3554
3555                 trace->ev_qualifier = strlist__new(lists[1], &slist_config);
3556                 if (trace->ev_qualifier == NULL) {
3557                         fputs("Not enough memory to parse event qualifier", trace->output);
3558                         goto out;
3559                 }
3560
3561                 if (trace__validate_ev_qualifier(trace))
3562                         goto out;
3563                 trace->trace_syscalls = true;
3564         }
3565
3566         err = 0;
3567
3568         if (lists[0]) {
3569                 struct option o = OPT_CALLBACK('e', "event", &trace->evlist, "event",
3570                                                "event selector. use 'perf list' to list available events",
3571                                                parse_events_option);
3572                 err = parse_events_option(&o, lists[0], 0);
3573         }
3574 out:
3575         if (sep)
3576                 *sep = ',';
3577
3578         return err;
3579 }
3580
3581 static int trace__parse_cgroups(const struct option *opt, const char *str, int unset)
3582 {
3583         struct trace *trace = opt->value;
3584
3585         if (!list_empty(&trace->evlist->entries))
3586                 return parse_cgroups(opt, str, unset);
3587
3588         trace->cgroup = evlist__findnew_cgroup(trace->evlist, str);
3589
3590         return 0;
3591 }
3592
3593 static struct bpf_map *bpf__find_map_by_name(const char *name)
3594 {
3595         struct bpf_object *obj, *tmp;
3596
3597         bpf_object__for_each_safe(obj, tmp) {
3598                 struct bpf_map *map = bpf_object__find_map_by_name(obj, name);
3599                 if (map)
3600                         return map;
3601
3602         }
3603
3604         return NULL;
3605 }
3606
3607 static void trace__set_bpf_map_filtered_pids(struct trace *trace)
3608 {
3609         trace->filter_pids.map = bpf__find_map_by_name("pids_filtered");
3610 }
3611
3612 static void trace__set_bpf_map_syscalls(struct trace *trace)
3613 {
3614         trace->syscalls.map = bpf__find_map_by_name("syscalls");
3615 }
3616
3617 static int trace__config(const char *var, const char *value, void *arg)
3618 {
3619         struct trace *trace = arg;
3620         int err = 0;
3621
3622         if (!strcmp(var, "trace.add_events")) {
3623                 struct option o = OPT_CALLBACK('e', "event", &trace->evlist, "event",
3624                                                "event selector. use 'perf list' to list available events",
3625                                                parse_events_option);
3626                 err = parse_events_option(&o, value, 0);
3627         } else if (!strcmp(var, "trace.show_timestamp")) {
3628                 trace->show_tstamp = perf_config_bool(var, value);
3629         } else if (!strcmp(var, "trace.show_duration")) {
3630                 trace->show_duration = perf_config_bool(var, value);
3631         } else if (!strcmp(var, "trace.show_arg_names")) {
3632                 trace->show_arg_names = perf_config_bool(var, value);
3633                 if (!trace->show_arg_names)
3634                         trace->show_zeros = true;
3635         } else if (!strcmp(var, "trace.show_zeros")) {
3636                 bool new_show_zeros = perf_config_bool(var, value);
3637                 if (!trace->show_arg_names && !new_show_zeros) {
3638                         pr_warning("trace.show_zeros has to be set when trace.show_arg_names=no\n");
3639                         goto out;
3640                 }
3641                 trace->show_zeros = new_show_zeros;
3642         } else if (!strcmp(var, "trace.show_prefix")) {
3643                 trace->show_string_prefix = perf_config_bool(var, value);
3644         } else if (!strcmp(var, "trace.no_inherit")) {
3645                 trace->opts.no_inherit = perf_config_bool(var, value);
3646         } else if (!strcmp(var, "trace.args_alignment")) {
3647                 int args_alignment = 0;
3648                 if (perf_config_int(&args_alignment, var, value) == 0)
3649                         trace->args_alignment = args_alignment;
3650         }
3651 out:
3652         return err;
3653 }
3654
3655 int cmd_trace(int argc, const char **argv)
3656 {
3657         const char *trace_usage[] = {
3658                 "perf trace [<options>] [<command>]",
3659                 "perf trace [<options>] -- <command> [<options>]",
3660                 "perf trace record [<options>] [<command>]",
3661                 "perf trace record [<options>] -- <command> [<options>]",
3662                 NULL
3663         };
3664         struct trace trace = {
3665                 .syscalls = {
3666                         . max = -1,
3667                 },
3668                 .opts = {
3669                         .target = {
3670                                 .uid       = UINT_MAX,
3671                                 .uses_mmap = true,
3672                         },
3673                         .user_freq     = UINT_MAX,
3674                         .user_interval = ULLONG_MAX,
3675                         .no_buffering  = true,
3676                         .mmap_pages    = UINT_MAX,
3677                 },
3678                 .output = stderr,
3679                 .show_comm = true,
3680                 .show_tstamp = true,
3681                 .show_duration = true,
3682                 .show_arg_names = true,
3683                 .args_alignment = 70,
3684                 .trace_syscalls = false,
3685                 .kernel_syscallchains = false,
3686                 .max_stack = UINT_MAX,
3687                 .max_events = ULONG_MAX,
3688         };
3689         const char *output_name = NULL;
3690         const struct option trace_options[] = {
3691         OPT_CALLBACK('e', "event", &trace, "event",
3692                      "event/syscall selector. use 'perf list' to list available events",
3693                      trace__parse_events_option),
3694         OPT_BOOLEAN(0, "comm", &trace.show_comm,
3695                     "show the thread COMM next to its id"),
3696         OPT_BOOLEAN(0, "tool_stats", &trace.show_tool_stats, "show tool stats"),
3697         OPT_CALLBACK(0, "expr", &trace, "expr", "list of syscalls/events to trace",
3698                      trace__parse_events_option),
3699         OPT_STRING('o', "output", &output_name, "file", "output file name"),
3700         OPT_STRING('i', "input", &input_name, "file", "Analyze events in file"),
3701         OPT_STRING('p', "pid", &trace.opts.target.pid, "pid",
3702                     "trace events on existing process id"),
3703         OPT_STRING('t', "tid", &trace.opts.target.tid, "tid",
3704                     "trace events on existing thread id"),
3705         OPT_CALLBACK(0, "filter-pids", &trace, "CSV list of pids",
3706                      "pids to filter (by the kernel)", trace__set_filter_pids_from_option),
3707         OPT_BOOLEAN('a', "all-cpus", &trace.opts.target.system_wide,
3708                     "system-wide collection from all CPUs"),
3709         OPT_STRING('C', "cpu", &trace.opts.target.cpu_list, "cpu",
3710                     "list of cpus to monitor"),
3711         OPT_BOOLEAN(0, "no-inherit", &trace.opts.no_inherit,
3712                     "child tasks do not inherit counters"),
3713         OPT_CALLBACK('m', "mmap-pages", &trace.opts.mmap_pages, "pages",
3714                      "number of mmap data pages",
3715                      perf_evlist__parse_mmap_pages),
3716         OPT_STRING('u', "uid", &trace.opts.target.uid_str, "user",
3717                    "user to profile"),
3718         OPT_CALLBACK(0, "duration", &trace, "float",
3719                      "show only events with duration > N.M ms",
3720                      trace__set_duration),
3721         OPT_BOOLEAN(0, "sched", &trace.sched, "show blocking scheduler events"),
3722         OPT_INCR('v', "verbose", &verbose, "be more verbose"),
3723         OPT_BOOLEAN('T', "time", &trace.full_time,
3724                     "Show full timestamp, not time relative to first start"),
3725         OPT_BOOLEAN(0, "failure", &trace.failure_only,
3726                     "Show only syscalls that failed"),
3727         OPT_BOOLEAN('s', "summary", &trace.summary_only,
3728                     "Show only syscall summary with statistics"),
3729         OPT_BOOLEAN('S', "with-summary", &trace.summary,
3730                     "Show all syscalls and summary with statistics"),
3731         OPT_CALLBACK_DEFAULT('F', "pf", &trace.trace_pgfaults, "all|maj|min",
3732                      "Trace pagefaults", parse_pagefaults, "maj"),
3733         OPT_BOOLEAN(0, "syscalls", &trace.trace_syscalls, "Trace syscalls"),
3734         OPT_BOOLEAN('f', "force", &trace.force, "don't complain, do it"),
3735         OPT_CALLBACK(0, "call-graph", &trace.opts,
3736                      "record_mode[,record_size]", record_callchain_help,
3737                      &record_parse_callchain_opt),
3738         OPT_BOOLEAN(0, "kernel-syscall-graph", &trace.kernel_syscallchains,
3739                     "Show the kernel callchains on the syscall exit path"),
3740         OPT_ULONG(0, "max-events", &trace.max_events,
3741                 "Set the maximum number of events to print, exit after that is reached. "),
3742         OPT_UINTEGER(0, "min-stack", &trace.min_stack,
3743                      "Set the minimum stack depth when parsing the callchain, "
3744                      "anything below the specified depth will be ignored."),
3745         OPT_UINTEGER(0, "max-stack", &trace.max_stack,
3746                      "Set the maximum stack depth when parsing the callchain, "
3747                      "anything beyond the specified depth will be ignored. "
3748                      "Default: kernel.perf_event_max_stack or " __stringify(PERF_MAX_STACK_DEPTH)),
3749         OPT_BOOLEAN(0, "sort-events", &trace.sort_events,
3750                         "Sort batch of events before processing, use if getting out of order events"),
3751         OPT_BOOLEAN(0, "print-sample", &trace.print_sample,
3752                         "print the PERF_RECORD_SAMPLE PERF_SAMPLE_ info, for debugging"),
3753         OPT_UINTEGER(0, "proc-map-timeout", &proc_map_timeout,
3754                         "per thread proc mmap processing timeout in ms"),
3755         OPT_CALLBACK('G', "cgroup", &trace, "name", "monitor event in cgroup name only",
3756                      trace__parse_cgroups),
3757         OPT_UINTEGER('D', "delay", &trace.opts.initial_delay,
3758                      "ms to wait before starting measurement after program "
3759                      "start"),
3760         OPT_END()
3761         };
3762         bool __maybe_unused max_stack_user_set = true;
3763         bool mmap_pages_user_set = true;
3764         struct perf_evsel *evsel;
3765         const char * const trace_subcommands[] = { "record", NULL };
3766         int err = -1;
3767         char bf[BUFSIZ];
3768
3769         signal(SIGSEGV, sighandler_dump_stack);
3770         signal(SIGFPE, sighandler_dump_stack);
3771
3772         trace.evlist = perf_evlist__new();
3773         trace.sctbl = syscalltbl__new();
3774
3775         if (trace.evlist == NULL || trace.sctbl == NULL) {
3776                 pr_err("Not enough memory to run!\n");
3777                 err = -ENOMEM;
3778                 goto out;
3779         }
3780
3781         err = perf_config(trace__config, &trace);
3782         if (err)
3783                 goto out;
3784
3785         argc = parse_options_subcommand(argc, argv, trace_options, trace_subcommands,
3786                                  trace_usage, PARSE_OPT_STOP_AT_NON_OPTION);
3787
3788         if ((nr_cgroups || trace.cgroup) && !trace.opts.target.system_wide) {
3789                 usage_with_options_msg(trace_usage, trace_options,
3790                                        "cgroup monitoring only available in system-wide mode");
3791         }
3792
3793         evsel = bpf__setup_output_event(trace.evlist, "__augmented_syscalls__");
3794         if (IS_ERR(evsel)) {
3795                 bpf__strerror_setup_output_event(trace.evlist, PTR_ERR(evsel), bf, sizeof(bf));
3796                 pr_err("ERROR: Setup trace syscalls enter failed: %s\n", bf);
3797                 goto out;
3798         }
3799
3800         if (evsel) {
3801                 trace.syscalls.events.augmented = evsel;
3802                 trace__set_bpf_map_filtered_pids(&trace);
3803                 trace__set_bpf_map_syscalls(&trace);
3804         }
3805
3806         err = bpf__setup_stdout(trace.evlist);
3807         if (err) {
3808                 bpf__strerror_setup_stdout(trace.evlist, err, bf, sizeof(bf));
3809                 pr_err("ERROR: Setup BPF stdout failed: %s\n", bf);
3810                 goto out;
3811         }
3812
3813         err = -1;
3814
3815         if (trace.trace_pgfaults) {
3816                 trace.opts.sample_address = true;
3817                 trace.opts.sample_time = true;
3818         }
3819
3820         if (trace.opts.mmap_pages == UINT_MAX)
3821                 mmap_pages_user_set = false;
3822
3823         if (trace.max_stack == UINT_MAX) {
3824                 trace.max_stack = input_name ? PERF_MAX_STACK_DEPTH : sysctl__max_stack();
3825                 max_stack_user_set = false;
3826         }
3827
3828 #ifdef HAVE_DWARF_UNWIND_SUPPORT
3829         if ((trace.min_stack || max_stack_user_set) && !callchain_param.enabled) {
3830                 record_opts__parse_callchain(&trace.opts, &callchain_param, "dwarf", false);
3831         }
3832 #endif
3833
3834         if (callchain_param.enabled) {
3835                 if (!mmap_pages_user_set && geteuid() == 0)
3836                         trace.opts.mmap_pages = perf_event_mlock_kb_in_pages() * 4;
3837
3838                 symbol_conf.use_callchain = true;
3839         }
3840
3841         if (trace.evlist->nr_entries > 0) {
3842                 evlist__set_evsel_handler(trace.evlist, trace__event_handler);
3843                 if (evlist__set_syscall_tp_fields(trace.evlist)) {
3844                         perror("failed to set syscalls:* tracepoint fields");
3845                         goto out;
3846                 }
3847         }
3848
3849         if (trace.sort_events) {
3850                 ordered_events__init(&trace.oe.data, ordered_events__deliver_event, &trace);
3851                 ordered_events__set_copy_on_queue(&trace.oe.data, true);
3852         }
3853
3854         /*
3855          * If we are augmenting syscalls, then combine what we put in the
3856          * __augmented_syscalls__ BPF map with what is in the
3857          * syscalls:sys_exit_FOO tracepoints, i.e. just like we do without BPF,
3858          * combining raw_syscalls:sys_enter with raw_syscalls:sys_exit.
3859          *
3860          * We'll switch to look at two BPF maps, one for sys_enter and the
3861          * other for sys_exit when we start augmenting the sys_exit paths with
3862          * buffers that are being copied from kernel to userspace, think 'read'
3863          * syscall.
3864          */
3865         if (trace.syscalls.events.augmented) {
3866                 evlist__for_each_entry(trace.evlist, evsel) {
3867                         bool raw_syscalls_sys_exit = strcmp(perf_evsel__name(evsel), "raw_syscalls:sys_exit") == 0;
3868
3869                         if (raw_syscalls_sys_exit) {
3870                                 trace.raw_augmented_syscalls = true;
3871                                 goto init_augmented_syscall_tp;
3872                         }
3873
3874                         if (trace.syscalls.events.augmented->priv == NULL &&
3875                             strstr(perf_evsel__name(evsel), "syscalls:sys_enter")) {
3876                                 struct perf_evsel *augmented = trace.syscalls.events.augmented;
3877                                 if (perf_evsel__init_augmented_syscall_tp(augmented, evsel) ||
3878                                     perf_evsel__init_augmented_syscall_tp_args(augmented))
3879                                         goto out;
3880                                 augmented->handler = trace__sys_enter;
3881                         }
3882
3883                         if (strstarts(perf_evsel__name(evsel), "syscalls:sys_exit_")) {
3884                                 struct syscall_tp *sc;
3885 init_augmented_syscall_tp:
3886                                 if (perf_evsel__init_augmented_syscall_tp(evsel, evsel))
3887                                         goto out;
3888                                 sc = evsel->priv;
3889                                 /*
3890                                  * For now with BPF raw_augmented we hook into
3891                                  * raw_syscalls:sys_enter and there we get all
3892                                  * 6 syscall args plus the tracepoint common
3893                                  * fields and the syscall_nr (another long).
3894                                  * So we check if that is the case and if so
3895                                  * don't look after the sc->args_size but
3896                                  * always after the full raw_syscalls:sys_enter
3897                                  * payload, which is fixed.
3898                                  *
3899                                  * We'll revisit this later to pass
3900                                  * s->args_size to the BPF augmenter (now
3901                                  * tools/perf/examples/bpf/augmented_raw_syscalls.c,
3902                                  * so that it copies only what we need for each
3903                                  * syscall, like what happens when we use
3904                                  * syscalls:sys_enter_NAME, so that we reduce
3905                                  * the kernel/userspace traffic to just what is
3906                                  * needed for each syscall.
3907                                  */
3908                                 if (trace.raw_augmented_syscalls)
3909                                         trace.raw_augmented_syscalls_args_size = (6 + 1) * sizeof(long) + sc->id.offset;
3910                                 perf_evsel__init_augmented_syscall_tp_ret(evsel);
3911                                 evsel->handler = trace__sys_exit;
3912                         }
3913                 }
3914         }
3915
3916         if ((argc >= 1) && (strcmp(argv[0], "record") == 0))
3917                 return trace__record(&trace, argc-1, &argv[1]);
3918
3919         /* summary_only implies summary option, but don't overwrite summary if set */
3920         if (trace.summary_only)
3921                 trace.summary = trace.summary_only;
3922
3923         if (!trace.trace_syscalls && !trace.trace_pgfaults &&
3924             trace.evlist->nr_entries == 0 /* Was --events used? */) {
3925                 trace.trace_syscalls = true;
3926         }
3927
3928         if (output_name != NULL) {
3929                 err = trace__open_output(&trace, output_name);
3930                 if (err < 0) {
3931                         perror("failed to create output file");
3932                         goto out;
3933                 }
3934         }
3935
3936         err = target__validate(&trace.opts.target);
3937         if (err) {
3938                 target__strerror(&trace.opts.target, err, bf, sizeof(bf));
3939                 fprintf(trace.output, "%s", bf);
3940                 goto out_close;
3941         }
3942
3943         err = target__parse_uid(&trace.opts.target);
3944         if (err) {
3945                 target__strerror(&trace.opts.target, err, bf, sizeof(bf));
3946                 fprintf(trace.output, "%s", bf);
3947                 goto out_close;
3948         }
3949
3950         if (!argc && target__none(&trace.opts.target))
3951                 trace.opts.target.system_wide = true;
3952
3953         if (input_name)
3954                 err = trace__replay(&trace);
3955         else
3956                 err = trace__run(&trace, argc, argv);
3957
3958 out_close:
3959         if (output_name != NULL)
3960                 fclose(trace.output);
3961 out:
3962         return err;
3963 }