tools/perf/builtin-trace.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * builtin-trace.c
   4  *
   5  * Builtin 'trace' command:
   6  *
   7  * Display a continuously updated trace of any workload, CPU, specific PID,
   8  * system wide, etc.  Default format is loosely strace like, but any other
   9  * event may be specified using --event.
  10  *
  11  * Copyright (C) 2012, 2013, 2014, 2015 Red Hat Inc, Arnaldo Carvalho de Melo <acme@redhat.com>
  12  *
  13  * Initially based on the 'trace' prototype by Thomas Gleixner:
  14  *
  15  * http://lwn.net/Articles/415728/ ("Announcing a new utility: 'trace'")
  16  */
  17
  18 #include <traceevent/event-parse.h>
  19 #include <api/fs/tracing_path.h>
  20 #include <bpf/bpf.h>
  21 #include "util/bpf_map.h"
  22 #include "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 <linux/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 {
  91                 struct bpf_map *map;
  92         } dump;
  93         struct record_opts      opts;
  94         struct perf_evlist      *evlist;
  95         struct machine          *host;
  96         struct thread           *current;
  97         struct cgroup           *cgroup;
  98         u64                     base_time;
  99         FILE                    *output;
 100         unsigned long           nr_events;
 101         unsigned long           nr_events_printed;
 102         unsigned long           max_events;
 103         struct strlist          *ev_qualifier;
 104         struct {
 105                 size_t          nr;
 106                 int             *entries;
 107         }                       ev_qualifier_ids;
 108         struct {
 109                 size_t          nr;
 110                 pid_t           *entries;
 111                 struct bpf_map  *map;
 112         }                       filter_pids;
 113         double                  duration_filter;
 114         double                  runtime_ms;
 115         struct {
 116                 u64             vfs_getname,
 117                                 proc_getname;
 118         } stats;
 119         unsigned int            max_stack;
 120         unsigned int            min_stack;
 121         int                     raw_augmented_syscalls_args_size;
 122         bool                    raw_augmented_syscalls;
 123         bool                    sort_events;
 124         bool                    not_ev_qualifier;
 125         bool                    live;
 126         bool                    full_time;
 127         bool                    sched;
 128         bool                    multiple_threads;
 129         bool                    summary;
 130         bool                    summary_only;
 131         bool                    failure_only;
 132         bool                    show_comm;
 133         bool                    print_sample;
 134         bool                    show_tool_stats;
 135         bool                    trace_syscalls;
 136         bool                    kernel_syscallchains;
 137         s16                     args_alignment;
 138         bool                    show_tstamp;
 139         bool                    show_duration;
 140         bool                    show_zeros;
 141         bool                    show_arg_names;
 142         bool                    show_string_prefix;
 143         bool                    force;
 144         bool                    vfs_getname;
 145         int                     trace_pgfaults;
 146         struct {
 147                 struct ordered_events   data;
 148                 u64                     last;
 149         } oe;
 150 };
 151
 152 struct tp_field {
 153         int offset;
 154         union {
 155                 u64 (*integer)(struct tp_field *field, struct perf_sample *sample);
 156                 void *(*pointer)(struct tp_field *field, struct perf_sample *sample);
 157         };
 158 };
 159
 160 #define TP_UINT_FIELD(bits) \
 161 static u64 tp_field__u##bits(struct tp_field *field, struct perf_sample *sample) \
 162 { \
 163         u##bits value; \
 164         memcpy(&value, sample->raw_data + field->offset, sizeof(value)); \
 165         return value;  \
 166 }
 167
 168 TP_UINT_FIELD(8);
 169 TP_UINT_FIELD(16);
 170 TP_UINT_FIELD(32);
 171 TP_UINT_FIELD(64);
 172
 173 #define TP_UINT_FIELD__SWAPPED(bits) \
 174 static u64 tp_field__swapped_u##bits(struct tp_field *field, struct perf_sample *sample) \
 175 { \
 176         u##bits value; \
 177         memcpy(&value, sample->raw_data + field->offset, sizeof(value)); \
 178         return bswap_##bits(value);\
 179 }
 180
 181 TP_UINT_FIELD__SWAPPED(16);
 182 TP_UINT_FIELD__SWAPPED(32);
 183 TP_UINT_FIELD__SWAPPED(64);
 184
 185 static int __tp_field__init_uint(struct tp_field *field, int size, int offset, bool needs_swap)
 186 {
 187         field->offset = offset;
 188
 189         switch (size) {
 190         case 1:
 191                 field->integer = tp_field__u8;
 192                 break;
 193         case 2:
 194                 field->integer = needs_swap ? tp_field__swapped_u16 : tp_field__u16;
 195                 break;
 196         case 4:
 197                 field->integer = needs_swap ? tp_field__swapped_u32 : tp_field__u32;
 198                 break;
 199         case 8:
 200                 field->integer = needs_swap ? tp_field__swapped_u64 : tp_field__u64;
 201                 break;
 202         default:
 203                 return -1;
 204         }
 205
 206         return 0;
 207 }
 208
 209 static int tp_field__init_uint(struct tp_field *field, struct tep_format_field *format_field, bool needs_swap)
 210 {
 211         return __tp_field__init_uint(field, format_field->size, format_field->offset, needs_swap);
 212 }
 213
 214 static void *tp_field__ptr(struct tp_field *field, struct perf_sample *sample)
 215 {
 216         return sample->raw_data + field->offset;
 217 }
 218
 219 static int __tp_field__init_ptr(struct tp_field *field, int offset)
 220 {
 221         field->offset = offset;
 222         field->pointer = tp_field__ptr;
 223         return 0;
 224 }
 225
 226 static int tp_field__init_ptr(struct tp_field *field, struct tep_format_field *format_field)
 227 {
 228         return __tp_field__init_ptr(field, format_field->offset);
 229 }
 230
 231 struct syscall_tp {
 232         struct tp_field id;
 233         union {
 234                 struct tp_field args, ret;
 235         };
 236 };
 237
 238 static int perf_evsel__init_tp_uint_field(struct perf_evsel *evsel,
 239                                           struct tp_field *field,
 240                                           const char *name)
 241 {
 242         struct tep_format_field *format_field = perf_evsel__field(evsel, name);
 243
 244         if (format_field == NULL)
 245                 return -1;
 246
 247         return tp_field__init_uint(field, format_field, evsel->needs_swap);
 248 }
 249
 250 #define perf_evsel__init_sc_tp_uint_field(evsel, name) \
 251         ({ struct syscall_tp *sc = evsel->priv;\
 252            perf_evsel__init_tp_uint_field(evsel, &sc->name, #name); })
 253
 254 static int perf_evsel__init_tp_ptr_field(struct perf_evsel *evsel,
 255                                          struct tp_field *field,
 256                                          const char *name)
 257 {
 258         struct tep_format_field *format_field = perf_evsel__field(evsel, name);
 259
 260         if (format_field == NULL)
 261                 return -1;
 262
 263         return tp_field__init_ptr(field, format_field);
 264 }
 265
 266 #define perf_evsel__init_sc_tp_ptr_field(evsel, name) \
 267         ({ struct syscall_tp *sc = evsel->priv;\
 268            perf_evsel__init_tp_ptr_field(evsel, &sc->name, #name); })
 269
 270 static void perf_evsel__delete_priv(struct perf_evsel *evsel)
 271 {
 272         zfree(&evsel->priv);
 273         perf_evsel__delete(evsel);
 274 }
 275
 276 static int perf_evsel__init_syscall_tp(struct perf_evsel *evsel)
 277 {
 278         struct syscall_tp *sc = evsel->priv = malloc(sizeof(struct syscall_tp));
 279
 280         if (evsel->priv != NULL) {
 281                 if (perf_evsel__init_tp_uint_field(evsel, &sc->id, "__syscall_nr") &&
 282                     perf_evsel__init_tp_uint_field(evsel, &sc->id, "nr"))
 283                         goto out_delete;
 284                 return 0;
 285         }
 286
 287         return -ENOMEM;
 288 out_delete:
 289         zfree(&evsel->priv);
 290         return -ENOENT;
 291 }
 292
 293 static int perf_evsel__init_augmented_syscall_tp(struct perf_evsel *evsel, struct perf_evsel *tp)
 294 {
 295         struct syscall_tp *sc = evsel->priv = malloc(sizeof(struct syscall_tp));
 296
 297         if (evsel->priv != NULL) {
 298                 struct tep_format_field *syscall_id = perf_evsel__field(tp, "id");
 299                 if (syscall_id == NULL)
 300                         syscall_id = perf_evsel__field(tp, "__syscall_nr");
 301                 if (syscall_id == NULL)
 302                         goto out_delete;
 303                 if (__tp_field__init_uint(&sc->id, syscall_id->size, syscall_id->offset, evsel->needs_swap))
 304                         goto out_delete;
 305
 306                 return 0;
 307         }
 308
 309         return -ENOMEM;
 310 out_delete:
 311         zfree(&evsel->priv);
 312         return -EINVAL;
 313 }
 314
 315 static int perf_evsel__init_augmented_syscall_tp_args(struct perf_evsel *evsel)
 316 {
 317         struct syscall_tp *sc = evsel->priv;
 318
 319         return __tp_field__init_ptr(&sc->args, sc->id.offset + sizeof(u64));
 320 }
 321
 322 static int perf_evsel__init_augmented_syscall_tp_ret(struct perf_evsel *evsel)
 323 {
 324         struct syscall_tp *sc = evsel->priv;
 325
 326         return __tp_field__init_uint(&sc->ret, sizeof(u64), sc->id.offset + sizeof(u64), evsel->needs_swap);
 327 }
 328
 329 static int perf_evsel__init_raw_syscall_tp(struct perf_evsel *evsel, void *handler)
 330 {
 331         evsel->priv = malloc(sizeof(struct syscall_tp));
 332         if (evsel->priv != NULL) {
 333                 if (perf_evsel__init_sc_tp_uint_field(evsel, id))
 334                         goto out_delete;
 335
 336                 evsel->handler = handler;
 337                 return 0;
 338         }
 339
 340         return -ENOMEM;
 341
 342 out_delete:
 343         zfree(&evsel->priv);
 344         return -ENOENT;
 345 }
 346
 347 static struct perf_evsel *perf_evsel__raw_syscall_newtp(const char *direction, void *handler)
 348 {
 349         struct perf_evsel *evsel = perf_evsel__newtp("raw_syscalls", direction);
 350
 351         /* older kernel (e.g., RHEL6) use syscalls:{enter,exit} */
 352         if (IS_ERR(evsel))
 353                 evsel = perf_evsel__newtp("syscalls", direction);
 354
 355         if (IS_ERR(evsel))
 356                 return NULL;
 357
 358         if (perf_evsel__init_raw_syscall_tp(evsel, handler))
 359                 goto out_delete;
 360
 361         return evsel;
 362
 363 out_delete:
 364         perf_evsel__delete_priv(evsel);
 365         return NULL;
 366 }
 367
 368 #define perf_evsel__sc_tp_uint(evsel, name, sample) \
 369         ({ struct syscall_tp *fields = evsel->priv; \
 370            fields->name.integer(&fields->name, sample); })
 371
 372 #define perf_evsel__sc_tp_ptr(evsel, name, sample) \
 373         ({ struct syscall_tp *fields = evsel->priv; \
 374            fields->name.pointer(&fields->name, sample); })
 375
 376 size_t strarray__scnprintf(struct strarray *sa, char *bf, size_t size, const char *intfmt, bool show_prefix, int val)
 377 {
 378         int idx = val - sa->offset;
 379
 380         if (idx < 0 || idx >= sa->nr_entries || sa->entries[idx] == NULL) {
 381                 size_t printed = scnprintf(bf, size, intfmt, val);
 382                 if (show_prefix)
 383                         printed += scnprintf(bf + printed, size - printed, " /* %s??? */", sa->prefix);
 384                 return printed;
 385         }
 386
 387         return scnprintf(bf, size, "%s%s", show_prefix ? sa->prefix : "", sa->entries[idx]);
 388 }
 389
 390 static size_t __syscall_arg__scnprintf_strarray(char *bf, size_t size,
 391                                                 const char *intfmt,
 392                                                 struct syscall_arg *arg)
 393 {
 394         return strarray__scnprintf(arg->parm, bf, size, intfmt, arg->show_string_prefix, arg->val);
 395 }
 396
 397 static size_t syscall_arg__scnprintf_strarray(char *bf, size_t size,
 398                                               struct syscall_arg *arg)
 399 {
 400         return __syscall_arg__scnprintf_strarray(bf, size, "%d", arg);
 401 }
 402
 403 #define SCA_STRARRAY syscall_arg__scnprintf_strarray
 404
 405 size_t syscall_arg__scnprintf_strarray_flags(char *bf, size_t size, struct syscall_arg *arg)
 406 {
 407         return strarray__scnprintf_flags(arg->parm, bf, size, arg->show_string_prefix, arg->val);
 408 }
 409
 410 size_t strarrays__scnprintf(struct strarrays *sas, char *bf, size_t size, const char *intfmt, bool show_prefix, int val)
 411 {
 412         size_t printed;
 413         int i;
 414
 415         for (i = 0; i < sas->nr_entries; ++i) {
 416                 struct strarray *sa = sas->entries[i];
 417                 int idx = val - sa->offset;
 418
 419                 if (idx >= 0 && idx < sa->nr_entries) {
 420                         if (sa->entries[idx] == NULL)
 421                                 break;
 422                         return scnprintf(bf, size, "%s%s", show_prefix ? sa->prefix : "", sa->entries[idx]);
 423                 }
 424         }
 425
 426         printed = scnprintf(bf, size, intfmt, val);
 427         if (show_prefix)
 428                 printed += scnprintf(bf + printed, size - printed, " /* %s??? */", sas->entries[0]->prefix);
 429         return printed;
 430 }
 431
 432 size_t syscall_arg__scnprintf_strarrays(char *bf, size_t size,
 433                                         struct syscall_arg *arg)
 434 {
 435         return strarrays__scnprintf(arg->parm, bf, size, "%d", arg->show_string_prefix, arg->val);
 436 }
 437
 438 #ifndef AT_FDCWD
 439 #define AT_FDCWD        -100
 440 #endif
 441
 442 static size_t syscall_arg__scnprintf_fd_at(char *bf, size_t size,
 443                                            struct syscall_arg *arg)
 444 {
 445         int fd = arg->val;
 446         const char *prefix = "AT_FD";
 447
 448         if (fd == AT_FDCWD)
 449                 return scnprintf(bf, size, "%s%s", arg->show_string_prefix ? prefix : "", "CWD");
 450
 451         return syscall_arg__scnprintf_fd(bf, size, arg);
 452 }
 453
 454 #define SCA_FDAT syscall_arg__scnprintf_fd_at
 455
 456 static size_t syscall_arg__scnprintf_close_fd(char *bf, size_t size,
 457                                               struct syscall_arg *arg);
 458
 459 #define SCA_CLOSE_FD syscall_arg__scnprintf_close_fd
 460
 461 size_t syscall_arg__scnprintf_hex(char *bf, size_t size, struct syscall_arg *arg)
 462 {
 463         return scnprintf(bf, size, "%#lx", arg->val);
 464 }
 465
 466 size_t syscall_arg__scnprintf_ptr(char *bf, size_t size, struct syscall_arg *arg)
 467 {
 468         if (arg->val == 0)
 469                 return scnprintf(bf, size, "NULL");
 470         return syscall_arg__scnprintf_hex(bf, size, arg);
 471 }
 472
 473 size_t syscall_arg__scnprintf_int(char *bf, size_t size, struct syscall_arg *arg)
 474 {
 475         return scnprintf(bf, size, "%d", arg->val);
 476 }
 477
 478 size_t syscall_arg__scnprintf_long(char *bf, size_t size, struct syscall_arg *arg)
 479 {
 480         return scnprintf(bf, size, "%ld", arg->val);
 481 }
 482
 483 static const char *bpf_cmd[] = {
 484         "MAP_CREATE", "MAP_LOOKUP_ELEM", "MAP_UPDATE_ELEM", "MAP_DELETE_ELEM",
 485         "MAP_GET_NEXT_KEY", "PROG_LOAD",
 486 };
 487 static DEFINE_STRARRAY(bpf_cmd, "BPF_");
 488
 489 static const char *fsmount_flags[] = {
 490         [1] = "CLOEXEC",
 491 };
 492 static DEFINE_STRARRAY(fsmount_flags, "FSMOUNT_");
 493
 494 #include "trace/beauty/generated/fsconfig_arrays.c"
 495
 496 static DEFINE_STRARRAY(fsconfig_cmds, "FSCONFIG_");
 497
 498 static const char *epoll_ctl_ops[] = { "ADD", "DEL", "MOD", };
 499 static DEFINE_STRARRAY_OFFSET(epoll_ctl_ops, "EPOLL_CTL_", 1);
 500
 501 static const char *itimers[] = { "REAL", "VIRTUAL", "PROF", };
 502 static DEFINE_STRARRAY(itimers, "ITIMER_");
 503
 504 static const char *keyctl_options[] = {
 505         "GET_KEYRING_ID", "JOIN_SESSION_KEYRING", "UPDATE", "REVOKE", "CHOWN",
 506         "SETPERM", "DESCRIBE", "CLEAR", "LINK", "UNLINK", "SEARCH", "READ",
 507         "INSTANTIATE", "NEGATE", "SET_REQKEY_KEYRING", "SET_TIMEOUT",
 508         "ASSUME_AUTHORITY", "GET_SECURITY", "SESSION_TO_PARENT", "REJECT",
 509         "INSTANTIATE_IOV", "INVALIDATE", "GET_PERSISTENT",
 510 };
 511 static DEFINE_STRARRAY(keyctl_options, "KEYCTL_");
 512
 513 static const char *whences[] = { "SET", "CUR", "END",
 514 #ifdef SEEK_DATA
 515 "DATA",
 516 #endif
 517 #ifdef SEEK_HOLE
 518 "HOLE",
 519 #endif
 520 };
 521 static DEFINE_STRARRAY(whences, "SEEK_");
 522
 523 static const char *fcntl_cmds[] = {
 524         "DUPFD", "GETFD", "SETFD", "GETFL", "SETFL", "GETLK", "SETLK",
 525         "SETLKW", "SETOWN", "GETOWN", "SETSIG", "GETSIG", "GETLK64",
 526         "SETLK64", "SETLKW64", "SETOWN_EX", "GETOWN_EX",
 527         "GETOWNER_UIDS",
 528 };
 529 static DEFINE_STRARRAY(fcntl_cmds, "F_");
 530
 531 static const char *fcntl_linux_specific_cmds[] = {
 532         "SETLEASE", "GETLEASE", "NOTIFY", [5] = "CANCELLK", "DUPFD_CLOEXEC",
 533         "SETPIPE_SZ", "GETPIPE_SZ", "ADD_SEALS", "GET_SEALS",
 534         "GET_RW_HINT", "SET_RW_HINT", "GET_FILE_RW_HINT", "SET_FILE_RW_HINT",
 535 };
 536
 537 static DEFINE_STRARRAY_OFFSET(fcntl_linux_specific_cmds, "F_", F_LINUX_SPECIFIC_BASE);
 538
 539 static struct strarray *fcntl_cmds_arrays[] = {
 540         &strarray__fcntl_cmds,
 541         &strarray__fcntl_linux_specific_cmds,
 542 };
 543
 544 static DEFINE_STRARRAYS(fcntl_cmds_arrays);
 545
 546 static const char *rlimit_resources[] = {
 547         "CPU", "FSIZE", "DATA", "STACK", "CORE", "RSS", "NPROC", "NOFILE",
 548         "MEMLOCK", "AS", "LOCKS", "SIGPENDING", "MSGQUEUE", "NICE", "RTPRIO",
 549         "RTTIME",
 550 };
 551 static DEFINE_STRARRAY(rlimit_resources, "RLIMIT_");
 552
 553 static const char *sighow[] = { "BLOCK", "UNBLOCK", "SETMASK", };
 554 static DEFINE_STRARRAY(sighow, "SIG_");
 555
 556 static const char *clockid[] = {
 557         "REALTIME", "MONOTONIC", "PROCESS_CPUTIME_ID", "THREAD_CPUTIME_ID",
 558         "MONOTONIC_RAW", "REALTIME_COARSE", "MONOTONIC_COARSE", "BOOTTIME",
 559         "REALTIME_ALARM", "BOOTTIME_ALARM", "SGI_CYCLE", "TAI"
 560 };
 561 static DEFINE_STRARRAY(clockid, "CLOCK_");
 562
 563 static size_t syscall_arg__scnprintf_access_mode(char *bf, size_t size,
 564                                                  struct syscall_arg *arg)
 565 {
 566         bool show_prefix = arg->show_string_prefix;
 567         const char *suffix = "_OK";
 568         size_t printed = 0;
 569         int mode = arg->val;
 570
 571         if (mode == F_OK) /* 0 */
 572                 return scnprintf(bf, size, "F%s", show_prefix ? suffix : "");
 573 #define P_MODE(n) \
 574         if (mode & n##_OK) { \
 575                 printed += scnprintf(bf + printed, size - printed, "%s%s", #n, show_prefix ? suffix : ""); \
 576                 mode &= ~n##_OK; \
 577         }
 578
 579         P_MODE(R);
 580         P_MODE(W);
 581         P_MODE(X);
 582 #undef P_MODE
 583
 584         if (mode)
 585                 printed += scnprintf(bf + printed, size - printed, "|%#x", mode);
 586
 587         return printed;
 588 }
 589
 590 #define SCA_ACCMODE syscall_arg__scnprintf_access_mode
 591
 592 static size_t syscall_arg__scnprintf_filename(char *bf, size_t size,
 593                                               struct syscall_arg *arg);
 594
 595 #define SCA_FILENAME syscall_arg__scnprintf_filename
 596
 597 static size_t syscall_arg__scnprintf_pipe_flags(char *bf, size_t size,
 598                                                 struct syscall_arg *arg)
 599 {
 600         bool show_prefix = arg->show_string_prefix;
 601         const char *prefix = "O_";
 602         int printed = 0, flags = arg->val;
 603
 604 #define P_FLAG(n) \
 605         if (flags & O_##n) { \
 606                 printed += scnprintf(bf + printed, size - printed, "%s%s%s", printed ? "|" : "", show_prefix ? prefix : "", #n); \
 607                 flags &= ~O_##n; \
 608         }
 609
 610         P_FLAG(CLOEXEC);
 611         P_FLAG(NONBLOCK);
 612 #undef P_FLAG
 613
 614         if (flags)
 615                 printed += scnprintf(bf + printed, size - printed, "%s%#x", printed ? "|" : "", flags);
 616
 617         return printed;
 618 }
 619
 620 #define SCA_PIPE_FLAGS syscall_arg__scnprintf_pipe_flags
 621
 622 #ifndef GRND_NONBLOCK
 623 #define GRND_NONBLOCK   0x0001
 624 #endif
 625 #ifndef GRND_RANDOM
 626 #define GRND_RANDOM     0x0002
 627 #endif
 628
 629 static size_t syscall_arg__scnprintf_getrandom_flags(char *bf, size_t size,
 630                                                    struct syscall_arg *arg)
 631 {
 632         bool show_prefix = arg->show_string_prefix;
 633         const char *prefix = "GRND_";
 634         int printed = 0, flags = arg->val;
 635
 636 #define P_FLAG(n) \
 637         if (flags & GRND_##n) { \
 638                 printed += scnprintf(bf + printed, size - printed, "%s%s%s", printed ? "|" : "", show_prefix ? prefix : "", #n); \
 639                 flags &= ~GRND_##n; \
 640         }
 641
 642         P_FLAG(RANDOM);
 643         P_FLAG(NONBLOCK);
 644 #undef P_FLAG
 645
 646         if (flags)
 647                 printed += scnprintf(bf + printed, size - printed, "%s%#x", printed ? "|" : "", flags);
 648
 649         return printed;
 650 }
 651
 652 #define SCA_GETRANDOM_FLAGS syscall_arg__scnprintf_getrandom_flags
 653
 654 #define STRARRAY(name, array) \
 655           { .scnprintf  = SCA_STRARRAY, \
 656             .parm       = &strarray__##array, }
 657
 658 #define STRARRAY_FLAGS(name, array) \
 659           { .scnprintf  = SCA_STRARRAY_FLAGS, \
 660             .parm       = &strarray__##array, }
 661
 662 #include "trace/beauty/arch_errno_names.c"
 663 #include "trace/beauty/eventfd.c"
 664 #include "trace/beauty/futex_op.c"
 665 #include "trace/beauty/futex_val3.c"
 666 #include "trace/beauty/mmap.c"
 667 #include "trace/beauty/mode_t.c"
 668 #include "trace/beauty/msg_flags.c"
 669 #include "trace/beauty/open_flags.c"
 670 #include "trace/beauty/perf_event_open.c"
 671 #include "trace/beauty/pid.c"
 672 #include "trace/beauty/sched_policy.c"
 673 #include "trace/beauty/seccomp.c"
 674 #include "trace/beauty/signum.c"
 675 #include "trace/beauty/socket_type.c"
 676 #include "trace/beauty/waitid_options.c"
 677
 678 struct syscall_arg_fmt {
 679         size_t     (*scnprintf)(char *bf, size_t size, struct syscall_arg *arg);
 680         unsigned long (*mask_val)(struct syscall_arg *arg, unsigned long val);
 681         void       *parm;
 682         const char *name;
 683         bool       show_zero;
 684 };
 685
 686 static struct syscall_fmt {
 687         const char *name;
 688         const char *alias;
 689         struct syscall_arg_fmt arg[6];
 690         u8         nr_args;
 691         bool       errpid;
 692         bool       timeout;
 693         bool       hexret;
 694 } syscall_fmts[] = {
 695         { .name     = "access",
 696           .arg = { [1] = { .scnprintf = SCA_ACCMODE,  /* mode */ }, }, },
 697         { .name     = "arch_prctl",
 698           .arg = { [0] = { .scnprintf = SCA_X86_ARCH_PRCTL_CODE, /* code */ },
 699                    [1] = { .scnprintf = SCA_PTR, /* arg2 */ }, }, },
 700         { .name     = "bind",
 701           .arg = { [1] = { .scnprintf = SCA_SOCKADDR, /* umyaddr */ }, }, },
 702         { .name     = "bpf",
 703           .arg = { [0] = STRARRAY(cmd, bpf_cmd), }, },
 704         { .name     = "brk",        .hexret = true,
 705           .arg = { [0] = { .scnprintf = SCA_PTR, /* brk */ }, }, },
 706         { .name     = "clock_gettime",
 707           .arg = { [0] = STRARRAY(clk_id, clockid), }, },
 708         { .name     = "clone",      .errpid = true, .nr_args = 5,
 709           .arg = { [0] = { .name = "flags",         .scnprintf = SCA_CLONE_FLAGS, },
 710                    [1] = { .name = "child_stack",   .scnprintf = SCA_HEX, },
 711                    [2] = { .name = "parent_tidptr", .scnprintf = SCA_HEX, },
 712                    [3] = { .name = "child_tidptr",  .scnprintf = SCA_HEX, },
 713                    [4] = { .name = "tls",           .scnprintf = SCA_HEX, }, }, },
 714         { .name     = "close",
 715           .arg = { [0] = { .scnprintf = SCA_CLOSE_FD, /* fd */ }, }, },
 716         { .name     = "connect",
 717           .arg = { [1] = { .scnprintf = SCA_SOCKADDR, /* servaddr */ }, }, },
 718         { .name     = "epoll_ctl",
 719           .arg = { [1] = STRARRAY(op, epoll_ctl_ops), }, },
 720         { .name     = "eventfd2",
 721           .arg = { [1] = { .scnprintf = SCA_EFD_FLAGS, /* flags */ }, }, },
 722         { .name     = "fchmodat",
 723           .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
 724         { .name     = "fchownat",
 725           .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
 726         { .name     = "fcntl",
 727           .arg = { [1] = { .scnprintf = SCA_FCNTL_CMD, /* cmd */
 728                            .parm      = &strarrays__fcntl_cmds_arrays,
 729                            .show_zero = true, },
 730                    [2] = { .scnprintf =  SCA_FCNTL_ARG, /* arg */ }, }, },
 731         { .name     = "flock",
 732           .arg = { [1] = { .scnprintf = SCA_FLOCK, /* cmd */ }, }, },
 733         { .name     = "fsconfig",
 734           .arg = { [1] = STRARRAY(cmd, fsconfig_cmds), }, },
 735         { .name     = "fsmount",
 736           .arg = { [1] = STRARRAY_FLAGS(flags, fsmount_flags),
 737                    [2] = { .scnprintf = SCA_FSMOUNT_ATTR_FLAGS, /* attr_flags */ }, }, },
 738         { .name     = "fspick",
 739           .arg = { [0] = { .scnprintf = SCA_FDAT,         /* dfd */ },
 740                    [1] = { .scnprintf = SCA_FILENAME,     /* path */ },
 741                    [2] = { .scnprintf = SCA_FSPICK_FLAGS, /* flags */ }, }, },
 742         { .name     = "fstat", .alias = "newfstat", },
 743         { .name     = "fstatat", .alias = "newfstatat", },
 744         { .name     = "futex",
 745           .arg = { [1] = { .scnprintf = SCA_FUTEX_OP, /* op */ },
 746                    [5] = { .scnprintf = SCA_FUTEX_VAL3, /* val3 */ }, }, },
 747         { .name     = "futimesat",
 748           .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
 749         { .name     = "getitimer",
 750           .arg = { [0] = STRARRAY(which, itimers), }, },
 751         { .name     = "getpid",     .errpid = true, },
 752         { .name     = "getpgid",    .errpid = true, },
 753         { .name     = "getppid",    .errpid = true, },
 754         { .name     = "getrandom",
 755           .arg = { [2] = { .scnprintf = SCA_GETRANDOM_FLAGS, /* flags */ }, }, },
 756         { .name     = "getrlimit",
 757           .arg = { [0] = STRARRAY(resource, rlimit_resources), }, },
 758         { .name     = "gettid",     .errpid = true, },
 759         { .name     = "ioctl",
 760           .arg = {
 761 #if defined(__i386__) || defined(__x86_64__)
 762 /*
 763  * FIXME: Make this available to all arches.
 764  */
 765                    [1] = { .scnprintf = SCA_IOCTL_CMD, /* cmd */ },
 766                    [2] = { .scnprintf = SCA_HEX, /* arg */ }, }, },
 767 #else
 768                    [2] = { .scnprintf = SCA_HEX, /* arg */ }, }, },
 769 #endif
 770         { .name     = "kcmp",       .nr_args = 5,
 771           .arg = { [0] = { .name = "pid1",      .scnprintf = SCA_PID, },
 772                    [1] = { .name = "pid2",      .scnprintf = SCA_PID, },
 773                    [2] = { .name = "type",      .scnprintf = SCA_KCMP_TYPE, },
 774                    [3] = { .name = "idx1",      .scnprintf = SCA_KCMP_IDX, },
 775                    [4] = { .name = "idx2",      .scnprintf = SCA_KCMP_IDX, }, }, },
 776         { .name     = "keyctl",
 777           .arg = { [0] = STRARRAY(option, keyctl_options), }, },
 778         { .name     = "kill",
 779           .arg = { [1] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
 780         { .name     = "linkat",
 781           .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
 782         { .name     = "lseek",
 783           .arg = { [2] = STRARRAY(whence, whences), }, },
 784         { .name     = "lstat", .alias = "newlstat", },
 785         { .name     = "madvise",
 786           .arg = { [0] = { .scnprintf = SCA_HEX,      /* start */ },
 787                    [2] = { .scnprintf = SCA_MADV_BHV, /* behavior */ }, }, },
 788         { .name     = "mkdirat",
 789           .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
 790         { .name     = "mknodat",
 791           .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
 792         { .name     = "mmap",       .hexret = true,
 793 /* The standard mmap maps to old_mmap on s390x */
 794 #if defined(__s390x__)
 795         .alias = "old_mmap",
 796 #endif
 797           .arg = { [2] = { .scnprintf = SCA_MMAP_PROT,  /* prot */ },
 798                    [3] = { .scnprintf = SCA_MMAP_FLAGS, /* flags */ },
 799                    [5] = { .scnprintf = SCA_HEX,        /* offset */ }, }, },
 800         { .name     = "mount",
 801           .arg = { [0] = { .scnprintf = SCA_FILENAME, /* dev_name */ },
 802                    [3] = { .scnprintf = SCA_MOUNT_FLAGS, /* flags */
 803                            .mask_val  = SCAMV_MOUNT_FLAGS, /* flags */ }, }, },
 804         { .name     = "move_mount",
 805           .arg = { [0] = { .scnprintf = SCA_FDAT,       /* from_dfd */ },
 806                    [1] = { .scnprintf = SCA_FILENAME, /* from_pathname */ },
 807                    [2] = { .scnprintf = SCA_FDAT,       /* to_dfd */ },
 808                    [3] = { .scnprintf = SCA_FILENAME, /* to_pathname */ },
 809                    [4] = { .scnprintf = SCA_MOVE_MOUNT_FLAGS, /* flags */ }, }, },
 810         { .name     = "mprotect",
 811           .arg = { [0] = { .scnprintf = SCA_HEX,        /* start */ },
 812                    [2] = { .scnprintf = SCA_MMAP_PROT,  /* prot */ }, }, },
 813         { .name     = "mq_unlink",
 814           .arg = { [0] = { .scnprintf = SCA_FILENAME, /* u_name */ }, }, },
 815         { .name     = "mremap",     .hexret = true,
 816           .arg = { [3] = { .scnprintf = SCA_MREMAP_FLAGS, /* flags */ }, }, },
 817         { .name     = "name_to_handle_at",
 818           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
 819         { .name     = "newfstatat",
 820           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
 821         { .name     = "open",
 822           .arg = { [1] = { .scnprintf = SCA_OPEN_FLAGS, /* flags */ }, }, },
 823         { .name     = "open_by_handle_at",
 824           .arg = { [0] = { .scnprintf = SCA_FDAT,       /* dfd */ },
 825                    [2] = { .scnprintf = SCA_OPEN_FLAGS, /* flags */ }, }, },
 826         { .name     = "openat",
 827           .arg = { [0] = { .scnprintf = SCA_FDAT,       /* dfd */ },
 828                    [2] = { .scnprintf = SCA_OPEN_FLAGS, /* flags */ }, }, },
 829         { .name     = "perf_event_open",
 830           .arg = { [2] = { .scnprintf = SCA_INT,        /* cpu */ },
 831                    [3] = { .scnprintf = SCA_FD,         /* group_fd */ },
 832                    [4] = { .scnprintf = SCA_PERF_FLAGS, /* flags */ }, }, },
 833         { .name     = "pipe2",
 834           .arg = { [1] = { .scnprintf = SCA_PIPE_FLAGS, /* flags */ }, }, },
 835         { .name     = "pkey_alloc",
 836           .arg = { [1] = { .scnprintf = SCA_PKEY_ALLOC_ACCESS_RIGHTS,   /* access_rights */ }, }, },
 837         { .name     = "pkey_free",
 838           .arg = { [0] = { .scnprintf = SCA_INT,        /* key */ }, }, },
 839         { .name     = "pkey_mprotect",
 840           .arg = { [0] = { .scnprintf = SCA_HEX,        /* start */ },
 841                    [2] = { .scnprintf = SCA_MMAP_PROT,  /* prot */ },
 842                    [3] = { .scnprintf = SCA_INT,        /* pkey */ }, }, },
 843         { .name     = "poll", .timeout = true, },
 844         { .name     = "ppoll", .timeout = true, },
 845         { .name     = "prctl",
 846           .arg = { [0] = { .scnprintf = SCA_PRCTL_OPTION, /* option */ },
 847                    [1] = { .scnprintf = SCA_PRCTL_ARG2, /* arg2 */ },
 848                    [2] = { .scnprintf = SCA_PRCTL_ARG3, /* arg3 */ }, }, },
 849         { .name     = "pread", .alias = "pread64", },
 850         { .name     = "preadv", .alias = "pread", },
 851         { .name     = "prlimit64",
 852           .arg = { [1] = STRARRAY(resource, rlimit_resources), }, },
 853         { .name     = "pwrite", .alias = "pwrite64", },
 854         { .name     = "readlinkat",
 855           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
 856         { .name     = "recvfrom",
 857           .arg = { [3] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
 858         { .name     = "recvmmsg",
 859           .arg = { [3] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
 860         { .name     = "recvmsg",
 861           .arg = { [2] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
 862         { .name     = "renameat",
 863           .arg = { [0] = { .scnprintf = SCA_FDAT, /* olddirfd */ },
 864                    [2] = { .scnprintf = SCA_FDAT, /* newdirfd */ }, }, },
 865         { .name     = "renameat2",
 866           .arg = { [0] = { .scnprintf = SCA_FDAT, /* olddirfd */ },
 867                    [2] = { .scnprintf = SCA_FDAT, /* newdirfd */ },
 868                    [4] = { .scnprintf = SCA_RENAMEAT2_FLAGS, /* flags */ }, }, },
 869         { .name     = "rt_sigaction",
 870           .arg = { [0] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
 871         { .name     = "rt_sigprocmask",
 872           .arg = { [0] = STRARRAY(how, sighow), }, },
 873         { .name     = "rt_sigqueueinfo",
 874           .arg = { [1] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
 875         { .name     = "rt_tgsigqueueinfo",
 876           .arg = { [2] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
 877         { .name     = "sched_setscheduler",
 878           .arg = { [1] = { .scnprintf = SCA_SCHED_POLICY, /* policy */ }, }, },
 879         { .name     = "seccomp",
 880           .arg = { [0] = { .scnprintf = SCA_SECCOMP_OP,    /* op */ },
 881                    [1] = { .scnprintf = SCA_SECCOMP_FLAGS, /* flags */ }, }, },
 882         { .name     = "select", .timeout = true, },
 883         { .name     = "sendmmsg",
 884           .arg = { [3] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
 885         { .name     = "sendmsg",
 886           .arg = { [2] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
 887         { .name     = "sendto",
 888           .arg = { [3] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ },
 889                    [4] = { .scnprintf = SCA_SOCKADDR, /* addr */ }, }, },
 890         { .name     = "set_tid_address", .errpid = true, },
 891         { .name     = "setitimer",
 892           .arg = { [0] = STRARRAY(which, itimers), }, },
 893         { .name     = "setrlimit",
 894           .arg = { [0] = STRARRAY(resource, rlimit_resources), }, },
 895         { .name     = "socket",
 896           .arg = { [0] = STRARRAY(family, socket_families),
 897                    [1] = { .scnprintf = SCA_SK_TYPE, /* type */ },
 898                    [2] = { .scnprintf = SCA_SK_PROTO, /* protocol */ }, }, },
 899         { .name     = "socketpair",
 900           .arg = { [0] = STRARRAY(family, socket_families),
 901                    [1] = { .scnprintf = SCA_SK_TYPE, /* type */ },
 902                    [2] = { .scnprintf = SCA_SK_PROTO, /* protocol */ }, }, },
 903         { .name     = "stat", .alias = "newstat", },
 904         { .name     = "statx",
 905           .arg = { [0] = { .scnprintf = SCA_FDAT,        /* fdat */ },
 906                    [2] = { .scnprintf = SCA_STATX_FLAGS, /* flags */ } ,
 907                    [3] = { .scnprintf = SCA_STATX_MASK,  /* mask */ }, }, },
 908         { .name     = "swapoff",
 909           .arg = { [0] = { .scnprintf = SCA_FILENAME, /* specialfile */ }, }, },
 910         { .name     = "swapon",
 911           .arg = { [0] = { .scnprintf = SCA_FILENAME, /* specialfile */ }, }, },
 912         { .name     = "symlinkat",
 913           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
 914         { .name     = "sync_file_range",
 915           .arg = { [3] = { .scnprintf = SCA_SYNC_FILE_RANGE_FLAGS, /* flags */ }, }, },
 916         { .name     = "tgkill",
 917           .arg = { [2] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
 918         { .name     = "tkill",
 919           .arg = { [1] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
 920         { .name     = "umount2", .alias = "umount",
 921           .arg = { [0] = { .scnprintf = SCA_FILENAME, /* name */ }, }, },
 922         { .name     = "uname", .alias = "newuname", },
 923         { .name     = "unlinkat",
 924           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
 925         { .name     = "utimensat",
 926           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dirfd */ }, }, },
 927         { .name     = "wait4",      .errpid = true,
 928           .arg = { [2] = { .scnprintf = SCA_WAITID_OPTIONS, /* options */ }, }, },
 929         { .name     = "waitid",     .errpid = true,
 930           .arg = { [3] = { .scnprintf = SCA_WAITID_OPTIONS, /* options */ }, }, },
 931 };
 932
 933 static int syscall_fmt__cmp(const void *name, const void *fmtp)
 934 {
 935         const struct syscall_fmt *fmt = fmtp;
 936         return strcmp(name, fmt->name);
 937 }
 938
 939 static struct syscall_fmt *syscall_fmt__find(const char *name)
 940 {
 941         const int nmemb = ARRAY_SIZE(syscall_fmts);
 942         return bsearch(name, syscall_fmts, nmemb, sizeof(struct syscall_fmt), syscall_fmt__cmp);
 943 }
 944
 945 static struct syscall_fmt *syscall_fmt__find_by_alias(const char *alias)
 946 {
 947         int i, nmemb = ARRAY_SIZE(syscall_fmts);
 948
 949         for (i = 0; i < nmemb; ++i) {
 950                 if (syscall_fmts[i].alias && strcmp(syscall_fmts[i].alias, alias) == 0)
 951                         return &syscall_fmts[i];
 952         }
 953
 954         return NULL;
 955 }
 956
 957 /*
 958  * is_exit: is this "exit" or "exit_group"?
 959  * is_open: is this "open" or "openat"? To associate the fd returned in sys_exit with the pathname in sys_enter.
 960  * args_size: sum of the sizes of the syscall arguments, anything after that is augmented stuff: pathname for openat, etc.
 961  */
 962 struct syscall {
 963         struct tep_event    *tp_format;
 964         int                 nr_args;
 965         int                 args_size;
 966         bool                is_exit;
 967         bool                is_open;
 968         struct tep_format_field *args;
 969         const char          *name;
 970         struct syscall_fmt  *fmt;
 971         struct syscall_arg_fmt *arg_fmt;
 972 };
 973
 974 /*
 975  * Must match what is in the BPF program:
 976  *
 977  * tools/perf/examples/bpf/augmented_raw_syscalls.c
 978  */
 979 struct bpf_map_syscall_entry {
 980         bool    enabled;
 981         u16     string_args_len[6];
 982 };
 983
 984 /*
 985  * We need to have this 'calculated' boolean because in some cases we really
 986  * don't know what is the duration of a syscall, for instance, when we start
 987  * a session and some threads are waiting for a syscall to finish, say 'poll',
 988  * in which case all we can do is to print "( ? ) for duration and for the
 989  * start timestamp.
 990  */
 991 static size_t fprintf_duration(unsigned long t, bool calculated, FILE *fp)
 992 {
 993         double duration = (double)t / NSEC_PER_MSEC;
 994         size_t printed = fprintf(fp, "(");
 995
 996         if (!calculated)
 997                 printed += fprintf(fp, "         ");
 998         else if (duration >= 1.0)
 999                 printed += color_fprintf(fp, PERF_COLOR_RED, "%6.3f ms", duration);
1000         else if (duration >= 0.01)
1001                 printed += color_fprintf(fp, PERF_COLOR_YELLOW, "%6.3f ms", duration);
1002         else
1003                 printed += color_fprintf(fp, PERF_COLOR_NORMAL, "%6.3f ms", duration);
1004         return printed + fprintf(fp, "): ");
1005 }
1006
1007 /**
1008  * filename.ptr: The filename char pointer that will be vfs_getname'd
1009  * filename.entry_str_pos: Where to insert the string translated from
1010  *                         filename.ptr by the vfs_getname tracepoint/kprobe.
1011  * ret_scnprintf: syscall args may set this to a different syscall return
1012  *                formatter, for instance, fcntl may return fds, file flags, etc.
1013  */
1014 struct thread_trace {
1015         u64               entry_time;
1016         bool              entry_pending;
1017         unsigned long     nr_events;
1018         unsigned long     pfmaj, pfmin;
1019         char              *entry_str;
1020         double            runtime_ms;
1021         size_t            (*ret_scnprintf)(char *bf, size_t size, struct syscall_arg *arg);
1022         struct {
1023                 unsigned long ptr;
1024                 short int     entry_str_pos;
1025                 bool          pending_open;
1026                 unsigned int  namelen;
1027                 char          *name;
1028         } filename;
1029         struct {
1030                 int           max;
1031                 struct file   *table;
1032         } files;
1033
1034         struct intlist *syscall_stats;
1035 };
1036
1037 static struct thread_trace *thread_trace__new(void)
1038 {
1039         struct thread_trace *ttrace =  zalloc(sizeof(struct thread_trace));
1040
1041         if (ttrace) {
1042                 ttrace->files.max = -1;
1043                 ttrace->syscall_stats = intlist__new(NULL);
1044         }
1045
1046         return ttrace;
1047 }
1048
1049 static struct thread_trace *thread__trace(struct thread *thread, FILE *fp)
1050 {
1051         struct thread_trace *ttrace;
1052
1053         if (thread == NULL)
1054                 goto fail;
1055
1056         if (thread__priv(thread) == NULL)
1057                 thread__set_priv(thread, thread_trace__new());
1058
1059         if (thread__priv(thread) == NULL)
1060                 goto fail;
1061
1062         ttrace = thread__priv(thread);
1063         ++ttrace->nr_events;
1064
1065         return ttrace;
1066 fail:
1067         color_fprintf(fp, PERF_COLOR_RED,
1068                       "WARNING: not enough memory, dropping samples!\n");
1069         return NULL;
1070 }
1071
1072
1073 void syscall_arg__set_ret_scnprintf(struct syscall_arg *arg,
1074                                     size_t (*ret_scnprintf)(char *bf, size_t size, struct syscall_arg *arg))
1075 {
1076         struct thread_trace *ttrace = thread__priv(arg->thread);
1077
1078         ttrace->ret_scnprintf = ret_scnprintf;
1079 }
1080
1081 #define TRACE_PFMAJ             (1 << 0)
1082 #define TRACE_PFMIN             (1 << 1)
1083
1084 static const size_t trace__entry_str_size = 2048;
1085
1086 static struct file *thread_trace__files_entry(struct thread_trace *ttrace, int fd)
1087 {
1088         if (fd < 0)
1089                 return NULL;
1090
1091         if (fd > ttrace->files.max) {
1092                 struct file *nfiles = realloc(ttrace->files.table, (fd + 1) * sizeof(struct file));
1093
1094                 if (nfiles == NULL)
1095                         return NULL;
1096
1097                 if (ttrace->files.max != -1) {
1098                         memset(nfiles + ttrace->files.max + 1, 0,
1099                                (fd - ttrace->files.max) * sizeof(struct file));
1100                 } else {
1101                         memset(nfiles, 0, (fd + 1) * sizeof(struct file));
1102                 }
1103
1104                 ttrace->files.table = nfiles;
1105                 ttrace->files.max   = fd;
1106         }
1107
1108         return ttrace->files.table + fd;
1109 }
1110
1111 struct file *thread__files_entry(struct thread *thread, int fd)
1112 {
1113         return thread_trace__files_entry(thread__priv(thread), fd);
1114 }
1115
1116 static int trace__set_fd_pathname(struct thread *thread, int fd, const char *pathname)
1117 {
1118         struct thread_trace *ttrace = thread__priv(thread);
1119         struct file *file = thread_trace__files_entry(ttrace, fd);
1120
1121         if (file != NULL) {
1122                 struct stat st;
1123                 if (stat(pathname, &st) == 0)
1124                         file->dev_maj = major(st.st_rdev);
1125                 file->pathname = strdup(pathname);
1126                 if (file->pathname)
1127                         return 0;
1128         }
1129
1130         return -1;
1131 }
1132
1133 static int thread__read_fd_path(struct thread *thread, int fd)
1134 {
1135         char linkname[PATH_MAX], pathname[PATH_MAX];
1136         struct stat st;
1137         int ret;
1138
1139         if (thread->pid_ == thread->tid) {
1140                 scnprintf(linkname, sizeof(linkname),
1141                           "/proc/%d/fd/%d", thread->pid_, fd);
1142         } else {
1143                 scnprintf(linkname, sizeof(linkname),
1144                           "/proc/%d/task/%d/fd/%d", thread->pid_, thread->tid, fd);
1145         }
1146
1147         if (lstat(linkname, &st) < 0 || st.st_size + 1 > (off_t)sizeof(pathname))
1148                 return -1;
1149
1150         ret = readlink(linkname, pathname, sizeof(pathname));
1151
1152         if (ret < 0 || ret > st.st_size)
1153                 return -1;
1154
1155         pathname[ret] = '\0';
1156         return trace__set_fd_pathname(thread, fd, pathname);
1157 }
1158
1159 static const char *thread__fd_path(struct thread *thread, int fd,
1160                                    struct trace *trace)
1161 {
1162         struct thread_trace *ttrace = thread__priv(thread);
1163
1164         if (ttrace == NULL)
1165                 return NULL;
1166
1167         if (fd < 0)
1168                 return NULL;
1169
1170         if ((fd > ttrace->files.max || ttrace->files.table[fd].pathname == NULL)) {
1171                 if (!trace->live)
1172                         return NULL;
1173                 ++trace->stats.proc_getname;
1174                 if (thread__read_fd_path(thread, fd))
1175                         return NULL;
1176         }
1177
1178         return ttrace->files.table[fd].pathname;
1179 }
1180
1181 size_t syscall_arg__scnprintf_fd(char *bf, size_t size, struct syscall_arg *arg)
1182 {
1183         int fd = arg->val;
1184         size_t printed = scnprintf(bf, size, "%d", fd);
1185         const char *path = thread__fd_path(arg->thread, fd, arg->trace);
1186
1187         if (path)
1188                 printed += scnprintf(bf + printed, size - printed, "<%s>", path);
1189
1190         return printed;
1191 }
1192
1193 size_t pid__scnprintf_fd(struct trace *trace, pid_t pid, int fd, char *bf, size_t size)
1194 {
1195         size_t printed = scnprintf(bf, size, "%d", fd);
1196         struct thread *thread = machine__find_thread(trace->host, pid, pid);
1197
1198         if (thread) {
1199                 const char *path = thread__fd_path(thread, fd, trace);
1200
1201                 if (path)
1202                         printed += scnprintf(bf + printed, size - printed, "<%s>", path);
1203
1204                 thread__put(thread);
1205         }
1206
1207         return printed;
1208 }
1209
1210 static size_t syscall_arg__scnprintf_close_fd(char *bf, size_t size,
1211                                               struct syscall_arg *arg)
1212 {
1213         int fd = arg->val;
1214         size_t printed = syscall_arg__scnprintf_fd(bf, size, arg);
1215         struct thread_trace *ttrace = thread__priv(arg->thread);
1216
1217         if (ttrace && fd >= 0 && fd <= ttrace->files.max)
1218                 zfree(&ttrace->files.table[fd].pathname);
1219
1220         return printed;
1221 }
1222
1223 static void thread__set_filename_pos(struct thread *thread, const char *bf,
1224                                      unsigned long ptr)
1225 {
1226         struct thread_trace *ttrace = thread__priv(thread);
1227
1228         ttrace->filename.ptr = ptr;
1229         ttrace->filename.entry_str_pos = bf - ttrace->entry_str;
1230 }
1231
1232 static size_t syscall_arg__scnprintf_augmented_string(struct syscall_arg *arg, char *bf, size_t size)
1233 {
1234         struct augmented_arg *augmented_arg = arg->augmented.args;
1235         size_t printed = scnprintf(bf, size, "\"%.*s\"", augmented_arg->size, augmented_arg->value);
1236         /*
1237          * So that the next arg with a payload can consume its augmented arg, i.e. for rename* syscalls
1238          * we would have two strings, each prefixed by its size.
1239          */
1240         int consumed = sizeof(*augmented_arg) + augmented_arg->size;
1241
1242         arg->augmented.args = ((void *)arg->augmented.args) + consumed;
1243         arg->augmented.size -= consumed;
1244
1245         return printed;
1246 }
1247
1248 static size_t syscall_arg__scnprintf_filename(char *bf, size_t size,
1249                                               struct syscall_arg *arg)
1250 {
1251         unsigned long ptr = arg->val;
1252
1253         if (arg->augmented.args)
1254                 return syscall_arg__scnprintf_augmented_string(arg, bf, size);
1255
1256         if (!arg->trace->vfs_getname)
1257                 return scnprintf(bf, size, "%#x", ptr);
1258
1259         thread__set_filename_pos(arg->thread, bf, ptr);
1260         return 0;
1261 }
1262
1263 static bool trace__filter_duration(struct trace *trace, double t)
1264 {
1265         return t < (trace->duration_filter * NSEC_PER_MSEC);
1266 }
1267
1268 static size_t __trace__fprintf_tstamp(struct trace *trace, u64 tstamp, FILE *fp)
1269 {
1270         double ts = (double)(tstamp - trace->base_time) / NSEC_PER_MSEC;
1271
1272         return fprintf(fp, "%10.3f ", ts);
1273 }
1274
1275 /*
1276  * We're handling tstamp=0 as an undefined tstamp, i.e. like when we are
1277  * using ttrace->entry_time for a thread that receives a sys_exit without
1278  * first having received a sys_enter ("poll" issued before tracing session
1279  * starts, lost sys_enter exit due to ring buffer overflow).
1280  */
1281 static size_t trace__fprintf_tstamp(struct trace *trace, u64 tstamp, FILE *fp)
1282 {
1283         if (tstamp > 0)
1284                 return __trace__fprintf_tstamp(trace, tstamp, fp);
1285
1286         return fprintf(fp, "         ? ");
1287 }
1288
1289 static bool done = false;
1290 static bool interrupted = false;
1291
1292 static void sig_handler(int sig)
1293 {
1294         done = true;
1295         interrupted = sig == SIGINT;
1296 }
1297
1298 static size_t trace__fprintf_comm_tid(struct trace *trace, struct thread *thread, FILE *fp)
1299 {
1300         size_t printed = 0;
1301
1302         if (trace->multiple_threads) {
1303                 if (trace->show_comm)
1304                         printed += fprintf(fp, "%.14s/", thread__comm_str(thread));
1305                 printed += fprintf(fp, "%d ", thread->tid);
1306         }
1307
1308         return printed;
1309 }
1310
1311 static size_t trace__fprintf_entry_head(struct trace *trace, struct thread *thread,
1312                                         u64 duration, bool duration_calculated, u64 tstamp, FILE *fp)
1313 {
1314         size_t printed = 0;
1315
1316         if (trace->show_tstamp)
1317                 printed = trace__fprintf_tstamp(trace, tstamp, fp);
1318         if (trace->show_duration)
1319                 printed += fprintf_duration(duration, duration_calculated, fp);
1320         return printed + trace__fprintf_comm_tid(trace, thread, fp);
1321 }
1322
1323 static int trace__process_event(struct trace *trace, struct machine *machine,
1324                                 union perf_event *event, struct perf_sample *sample)
1325 {
1326         int ret = 0;
1327
1328         switch (event->header.type) {
1329         case PERF_RECORD_LOST:
1330                 color_fprintf(trace->output, PERF_COLOR_RED,
1331                               "LOST %" PRIu64 " events!\n", event->lost.lost);
1332                 ret = machine__process_lost_event(machine, event, sample);
1333                 break;
1334         default:
1335                 ret = machine__process_event(machine, event, sample);
1336                 break;
1337         }
1338
1339         return ret;
1340 }
1341
1342 static int trace__tool_process(struct perf_tool *tool,
1343                                union perf_event *event,
1344                                struct perf_sample *sample,
1345                                struct machine *machine)
1346 {
1347         struct trace *trace = container_of(tool, struct trace, tool);
1348         return trace__process_event(trace, machine, event, sample);
1349 }
1350
1351 static char *trace__machine__resolve_kernel_addr(void *vmachine, unsigned long long *addrp, char **modp)
1352 {
1353         struct machine *machine = vmachine;
1354
1355         if (machine->kptr_restrict_warned)
1356                 return NULL;
1357
1358         if (symbol_conf.kptr_restrict) {
1359                 pr_warning("Kernel address maps (/proc/{kallsyms,modules}) are restricted.\n\n"
1360                            "Check /proc/sys/kernel/kptr_restrict.\n\n"
1361                            "Kernel samples will not be resolved.\n");
1362                 machine->kptr_restrict_warned = true;
1363                 return NULL;
1364         }
1365
1366         return machine__resolve_kernel_addr(vmachine, addrp, modp);
1367 }
1368
1369 static int trace__symbols_init(struct trace *trace, struct perf_evlist *evlist)
1370 {
1371         int err = symbol__init(NULL);
1372
1373         if (err)
1374                 return err;
1375
1376         trace->host = machine__new_host();
1377         if (trace->host == NULL)
1378                 return -ENOMEM;
1379
1380         err = trace_event__register_resolver(trace->host, trace__machine__resolve_kernel_addr);
1381         if (err < 0)
1382                 goto out;
1383
1384         err = __machine__synthesize_threads(trace->host, &trace->tool, &trace->opts.target,
1385                                             evlist->threads, trace__tool_process, false,
1386                                             1);
1387 out:
1388         if (err)
1389                 symbol__exit();
1390
1391         return err;
1392 }
1393
1394 static void trace__symbols__exit(struct trace *trace)
1395 {
1396         machine__exit(trace->host);
1397         trace->host = NULL;
1398
1399         symbol__exit();
1400 }
1401
1402 static int syscall__alloc_arg_fmts(struct syscall *sc, int nr_args)
1403 {
1404         int idx;
1405
1406         if (nr_args == 6 && sc->fmt && sc->fmt->nr_args != 0)
1407                 nr_args = sc->fmt->nr_args;
1408
1409         sc->arg_fmt = calloc(nr_args, sizeof(*sc->arg_fmt));
1410         if (sc->arg_fmt == NULL)
1411                 return -1;
1412
1413         for (idx = 0; idx < nr_args; ++idx) {
1414                 if (sc->fmt)
1415                         sc->arg_fmt[idx] = sc->fmt->arg[idx];
1416         }
1417
1418         sc->nr_args = nr_args;
1419         return 0;
1420 }
1421
1422 static int syscall__set_arg_fmts(struct syscall *sc)
1423 {
1424         struct tep_format_field *field, *last_field = NULL;
1425         int idx = 0, len;
1426
1427         for (field = sc->args; field; field = field->next, ++idx) {
1428                 last_field = field;
1429
1430                 if (sc->fmt && sc->fmt->arg[idx].scnprintf)
1431                         continue;
1432
1433                 len = strlen(field->name);
1434
1435                 if (strcmp(field->type, "const char *") == 0 &&
1436                     ((len >= 4 && strcmp(field->name + len - 4, "name") == 0) ||
1437                      strstr(field->name, "path") != NULL))
1438                         sc->arg_fmt[idx].scnprintf = SCA_FILENAME;
1439                 else if ((field->flags & TEP_FIELD_IS_POINTER) || strstr(field->name, "addr"))
1440                         sc->arg_fmt[idx].scnprintf = SCA_PTR;
1441                 else if (strcmp(field->type, "pid_t") == 0)
1442                         sc->arg_fmt[idx].scnprintf = SCA_PID;
1443                 else if (strcmp(field->type, "umode_t") == 0)
1444                         sc->arg_fmt[idx].scnprintf = SCA_MODE_T;
1445                 else if ((strcmp(field->type, "int") == 0 ||
1446                           strcmp(field->type, "unsigned int") == 0 ||
1447                           strcmp(field->type, "long") == 0) &&
1448                          len >= 2 && strcmp(field->name + len - 2, "fd") == 0) {
1449                         /*
1450                          * /sys/kernel/tracing/events/syscalls/sys_enter*
1451                          * egrep 'field:.*fd;' .../format|sed -r 's/.*field:([a-z ]+) [a-z_]*fd.+/\1/g'|sort|uniq -c
1452                          * 65 int
1453                          * 23 unsigned int
1454                          * 7 unsigned long
1455                          */
1456                         sc->arg_fmt[idx].scnprintf = SCA_FD;
1457                 }
1458         }
1459
1460         if (last_field)
1461                 sc->args_size = last_field->offset + last_field->size;
1462
1463         return 0;
1464 }
1465
1466 static int trace__read_syscall_info(struct trace *trace, int id)
1467 {
1468         char tp_name[128];
1469         struct syscall *sc;
1470         const char *name = syscalltbl__name(trace->sctbl, id);
1471
1472         if (name == NULL)
1473                 return -1;
1474
1475         if (id > trace->syscalls.max) {
1476                 struct syscall *nsyscalls = realloc(trace->syscalls.table, (id + 1) * sizeof(*sc));
1477
1478                 if (nsyscalls == NULL)
1479                         return -1;
1480
1481                 if (trace->syscalls.max != -1) {
1482                         memset(nsyscalls + trace->syscalls.max + 1, 0,
1483                                (id - trace->syscalls.max) * sizeof(*sc));
1484                 } else {
1485                         memset(nsyscalls, 0, (id + 1) * sizeof(*sc));
1486                 }
1487
1488                 trace->syscalls.table = nsyscalls;
1489                 trace->syscalls.max   = id;
1490         }
1491
1492         sc = trace->syscalls.table + id;
1493         sc->name = name;
1494
1495         sc->fmt  = syscall_fmt__find(sc->name);
1496
1497         snprintf(tp_name, sizeof(tp_name), "sys_enter_%s", sc->name);
1498         sc->tp_format = trace_event__tp_format("syscalls", tp_name);
1499
1500         if (IS_ERR(sc->tp_format) && sc->fmt && sc->fmt->alias) {
1501                 snprintf(tp_name, sizeof(tp_name), "sys_enter_%s", sc->fmt->alias);
1502                 sc->tp_format = trace_event__tp_format("syscalls", tp_name);
1503         }
1504
1505         if (syscall__alloc_arg_fmts(sc, IS_ERR(sc->tp_format) ? 6 : sc->tp_format->format.nr_fields))
1506                 return -1;
1507
1508         if (IS_ERR(sc->tp_format))
1509                 return -1;
1510
1511         sc->args = sc->tp_format->format.fields;
1512         /*
1513          * We need to check and discard the first variable '__syscall_nr'
1514          * or 'nr' that mean the syscall number. It is needless here.
1515          * So drop '__syscall_nr' or 'nr' field but does not exist on older kernels.
1516          */
1517         if (sc->args && (!strcmp(sc->args->name, "__syscall_nr") || !strcmp(sc->args->name, "nr"))) {
1518                 sc->args = sc->args->next;
1519                 --sc->nr_args;
1520         }
1521
1522         sc->is_exit = !strcmp(name, "exit_group") || !strcmp(name, "exit");
1523         sc->is_open = !strcmp(name, "open") || !strcmp(name, "openat");
1524
1525         return syscall__set_arg_fmts(sc);
1526 }
1527
1528 static int trace__validate_ev_qualifier(struct trace *trace)
1529 {
1530         int err = 0;
1531         bool printed_invalid_prefix = false;
1532         struct str_node *pos;
1533         size_t nr_used = 0, nr_allocated = strlist__nr_entries(trace->ev_qualifier);
1534
1535         trace->ev_qualifier_ids.entries = malloc(nr_allocated *
1536                                                  sizeof(trace->ev_qualifier_ids.entries[0]));
1537
1538         if (trace->ev_qualifier_ids.entries == NULL) {
1539                 fputs("Error:\tNot enough memory for allocating events qualifier ids\n",
1540                        trace->output);
1541                 err = -EINVAL;
1542                 goto out;
1543         }
1544
1545         strlist__for_each_entry(pos, trace->ev_qualifier) {
1546                 const char *sc = pos->s;
1547                 int id = syscalltbl__id(trace->sctbl, sc), match_next = -1;
1548
1549                 if (id < 0) {
1550                         id = syscalltbl__strglobmatch_first(trace->sctbl, sc, &match_next);
1551                         if (id >= 0)
1552                                 goto matches;
1553
1554                         if (!printed_invalid_prefix) {
1555                                 pr_debug("Skipping unknown syscalls: ");
1556                                 printed_invalid_prefix = true;
1557                         } else {
1558                                 pr_debug(", ");
1559                         }
1560
1561                         pr_debug("%s", sc);
1562                         continue;
1563                 }
1564 matches:
1565                 trace->ev_qualifier_ids.entries[nr_used++] = id;
1566                 if (match_next == -1)
1567                         continue;
1568
1569                 while (1) {
1570                         id = syscalltbl__strglobmatch_next(trace->sctbl, sc, &match_next);
1571                         if (id < 0)
1572                                 break;
1573                         if (nr_allocated == nr_used) {
1574                                 void *entries;
1575
1576                                 nr_allocated += 8;
1577                                 entries = realloc(trace->ev_qualifier_ids.entries,
1578                                                   nr_allocated * sizeof(trace->ev_qualifier_ids.entries[0]));
1579                                 if (entries == NULL) {
1580                                         err = -ENOMEM;
1581                                         fputs("\nError:\t Not enough memory for parsing\n", trace->output);
1582                                         goto out_free;
1583                                 }
1584                                 trace->ev_qualifier_ids.entries = entries;
1585                         }
1586                         trace->ev_qualifier_ids.entries[nr_used++] = id;
1587                 }
1588         }
1589
1590         trace->ev_qualifier_ids.nr = nr_used;
1591 out:
1592         if (printed_invalid_prefix)
1593                 pr_debug("\n");
1594         return err;
1595 out_free:
1596         zfree(&trace->ev_qualifier_ids.entries);
1597         trace->ev_qualifier_ids.nr = 0;
1598         goto out;
1599 }
1600
1601 /*
1602  * args is to be interpreted as a series of longs but we need to handle
1603  * 8-byte unaligned accesses. args points to raw_data within the event
1604  * and raw_data is guaranteed to be 8-byte unaligned because it is
1605  * preceded by raw_size which is a u32. So we need to copy args to a temp
1606  * variable to read it. Most notably this avoids extended load instructions
1607  * on unaligned addresses
1608  */
1609 unsigned long syscall_arg__val(struct syscall_arg *arg, u8 idx)
1610 {
1611         unsigned long val;
1612         unsigned char *p = arg->args + sizeof(unsigned long) * idx;
1613
1614         memcpy(&val, p, sizeof(val));
1615         return val;
1616 }
1617
1618 static size_t syscall__scnprintf_name(struct syscall *sc, char *bf, size_t size,
1619                                       struct syscall_arg *arg)
1620 {
1621         if (sc->arg_fmt && sc->arg_fmt[arg->idx].name)
1622                 return scnprintf(bf, size, "%s: ", sc->arg_fmt[arg->idx].name);
1623
1624         return scnprintf(bf, size, "arg%d: ", arg->idx);
1625 }
1626
1627 /*
1628  * Check if the value is in fact zero, i.e. mask whatever needs masking, such
1629  * as mount 'flags' argument that needs ignoring some magic flag, see comment
1630  * in tools/perf/trace/beauty/mount_flags.c
1631  */
1632 static unsigned long syscall__mask_val(struct syscall *sc, struct syscall_arg *arg, unsigned long val)
1633 {
1634         if (sc->arg_fmt && sc->arg_fmt[arg->idx].mask_val)
1635                 return sc->arg_fmt[arg->idx].mask_val(arg, val);
1636
1637         return val;
1638 }
1639
1640 static size_t syscall__scnprintf_val(struct syscall *sc, char *bf, size_t size,
1641                                      struct syscall_arg *arg, unsigned long val)
1642 {
1643         if (sc->arg_fmt && sc->arg_fmt[arg->idx].scnprintf) {
1644                 arg->val = val;
1645                 if (sc->arg_fmt[arg->idx].parm)
1646                         arg->parm = sc->arg_fmt[arg->idx].parm;
1647                 return sc->arg_fmt[arg->idx].scnprintf(bf, size, arg);
1648         }
1649         return scnprintf(bf, size, "%ld", val);
1650 }
1651
1652 static size_t syscall__scnprintf_args(struct syscall *sc, char *bf, size_t size,
1653                                       unsigned char *args, void *augmented_args, int augmented_args_size,
1654                                       struct trace *trace, struct thread *thread)
1655 {
1656         size_t printed = 0;
1657         unsigned long val;
1658         u8 bit = 1;
1659         struct syscall_arg arg = {
1660                 .args   = args,
1661                 .augmented = {
1662                         .size = augmented_args_size,
1663                         .args = augmented_args,
1664                 },
1665                 .idx    = 0,
1666                 .mask   = 0,
1667                 .trace  = trace,
1668                 .thread = thread,
1669                 .show_string_prefix = trace->show_string_prefix,
1670         };
1671         struct thread_trace *ttrace = thread__priv(thread);
1672
1673         /*
1674          * Things like fcntl will set this in its 'cmd' formatter to pick the
1675          * right formatter for the return value (an fd? file flags?), which is
1676          * not needed for syscalls that always return a given type, say an fd.
1677          */
1678         ttrace->ret_scnprintf = NULL;
1679
1680         if (sc->args != NULL) {
1681                 struct tep_format_field *field;
1682
1683                 for (field = sc->args; field;
1684                      field = field->next, ++arg.idx, bit <<= 1) {
1685                         if (arg.mask & bit)
1686                                 continue;
1687
1688                         val = syscall_arg__val(&arg, arg.idx);
1689                         /*
1690                          * Some syscall args need some mask, most don't and
1691                          * return val untouched.
1692                          */
1693                         val = syscall__mask_val(sc, &arg, val);
1694
1695                         /*
1696                          * Suppress this argument if its value is zero and
1697                          * and we don't have a string associated in an
1698                          * strarray for it.
1699                          */
1700                         if (val == 0 &&
1701                             !trace->show_zeros &&
1702                             !(sc->arg_fmt &&
1703                               (sc->arg_fmt[arg.idx].show_zero ||
1704                                sc->arg_fmt[arg.idx].scnprintf == SCA_STRARRAY ||
1705                                sc->arg_fmt[arg.idx].scnprintf == SCA_STRARRAYS) &&
1706                               sc->arg_fmt[arg.idx].parm))
1707                                 continue;
1708
1709                         printed += scnprintf(bf + printed, size - printed, "%s", printed ? ", " : "");
1710
1711                         if (trace->show_arg_names)
1712                                 printed += scnprintf(bf + printed, size - printed, "%s: ", field->name);
1713
1714                         printed += syscall__scnprintf_val(sc, bf + printed, size - printed, &arg, val);
1715                 }
1716         } else if (IS_ERR(sc->tp_format)) {
1717                 /*
1718                  * If we managed to read the tracepoint /format file, then we
1719                  * may end up not having any args, like with gettid(), so only
1720                  * print the raw args when we didn't manage to read it.
1721                  */
1722                 while (arg.idx < sc->nr_args) {
1723                         if (arg.mask & bit)
1724                                 goto next_arg;
1725                         val = syscall_arg__val(&arg, arg.idx);
1726                         if (printed)
1727                                 printed += scnprintf(bf + printed, size - printed, ", ");
1728                         printed += syscall__scnprintf_name(sc, bf + printed, size - printed, &arg);
1729                         printed += syscall__scnprintf_val(sc, bf + printed, size - printed, &arg, val);
1730 next_arg:
1731                         ++arg.idx;
1732                         bit <<= 1;
1733                 }
1734         }
1735
1736         return printed;
1737 }
1738
1739 typedef int (*tracepoint_handler)(struct trace *trace, struct perf_evsel *evsel,
1740                                   union perf_event *event,
1741                                   struct perf_sample *sample);
1742
1743 static struct syscall *trace__syscall_info(struct trace *trace,
1744                                            struct perf_evsel *evsel, int id)
1745 {
1746
1747         if (id < 0) {
1748
1749                 /*
1750                  * XXX: Noticed on x86_64, reproduced as far back as 3.0.36, haven't tried
1751                  * before that, leaving at a higher verbosity level till that is
1752                  * explained. Reproduced with plain ftrace with:
1753                  *
1754                  * echo 1 > /t/events/raw_syscalls/sys_exit/enable
1755                  * grep "NR -1 " /t/trace_pipe
1756                  *
1757                  * After generating some load on the machine.
1758                  */
1759                 if (verbose > 1) {
1760                         static u64 n;
1761                         fprintf(trace->output, "Invalid syscall %d id, skipping (%s, %" PRIu64 ") ...\n",
1762                                 id, perf_evsel__name(evsel), ++n);
1763                 }
1764                 return NULL;
1765         }
1766
1767         if ((id > trace->syscalls.max || trace->syscalls.table[id].name == NULL) &&
1768             trace__read_syscall_info(trace, id))
1769                 goto out_cant_read;
1770
1771         if ((id > trace->syscalls.max || trace->syscalls.table[id].name == NULL))
1772                 goto out_cant_read;
1773
1774         return &trace->syscalls.table[id];
1775
1776 out_cant_read:
1777         if (verbose > 0) {
1778                 fprintf(trace->output, "Problems reading syscall %d", id);
1779                 if (id <= trace->syscalls.max && trace->syscalls.table[id].name != NULL)
1780                         fprintf(trace->output, "(%s)", trace->syscalls.table[id].name);
1781                 fputs(" information\n", trace->output);
1782         }
1783         return NULL;
1784 }
1785
1786 static void thread__update_stats(struct thread_trace *ttrace,
1787                                  int id, struct perf_sample *sample)
1788 {
1789         struct int_node *inode;
1790         struct stats *stats;
1791         u64 duration = 0;
1792
1793         inode = intlist__findnew(ttrace->syscall_stats, id);
1794         if (inode == NULL)
1795                 return;
1796
1797         stats = inode->priv;
1798         if (stats == NULL) {
1799                 stats = malloc(sizeof(struct stats));
1800                 if (stats == NULL)
1801                         return;
1802                 init_stats(stats);
1803                 inode->priv = stats;
1804         }
1805
1806         if (ttrace->entry_time && sample->time > ttrace->entry_time)
1807                 duration = sample->time - ttrace->entry_time;
1808
1809         update_stats(stats, duration);
1810 }
1811
1812 static int trace__printf_interrupted_entry(struct trace *trace)
1813 {
1814         struct thread_trace *ttrace;
1815         size_t printed;
1816         int len;
1817
1818         if (trace->failure_only || trace->current == NULL)
1819                 return 0;
1820
1821         ttrace = thread__priv(trace->current);
1822
1823         if (!ttrace->entry_pending)
1824                 return 0;
1825
1826         printed  = trace__fprintf_entry_head(trace, trace->current, 0, false, ttrace->entry_time, trace->output);
1827         printed += len = fprintf(trace->output, "%s)", ttrace->entry_str);
1828
1829         if (len < trace->args_alignment - 4)
1830                 printed += fprintf(trace->output, "%-*s", trace->args_alignment - 4 - len, " ");
1831
1832         printed += fprintf(trace->output, " ...\n");
1833
1834         ttrace->entry_pending = false;
1835         ++trace->nr_events_printed;
1836
1837         return printed;
1838 }
1839
1840 static int trace__fprintf_sample(struct trace *trace, struct perf_evsel *evsel,
1841                                  struct perf_sample *sample, struct thread *thread)
1842 {
1843         int printed = 0;
1844
1845         if (trace->print_sample) {
1846                 double ts = (double)sample->time / NSEC_PER_MSEC;
1847
1848                 printed += fprintf(trace->output, "%22s %10.3f %s %d/%d [%d]\n",
1849                                    perf_evsel__name(evsel), ts,
1850                                    thread__comm_str(thread),
1851                                    sample->pid, sample->tid, sample->cpu);
1852         }
1853
1854         return printed;
1855 }
1856
1857 static void *syscall__augmented_args(struct syscall *sc, struct perf_sample *sample, int *augmented_args_size, int raw_augmented_args_size)
1858 {
1859         void *augmented_args = NULL;
1860         /*
1861          * For now with BPF raw_augmented we hook into raw_syscalls:sys_enter
1862          * and there we get all 6 syscall args plus the tracepoint common fields
1863          * that gets calculated at the start and the syscall_nr (another long).
1864          * So we check if that is the case and if so don't look after the
1865          * sc->args_size but always after the full raw_syscalls:sys_enter payload,
1866          * which is fixed.
1867          *
1868          * We'll revisit this later to pass s->args_size to the BPF augmenter
1869          * (now tools/perf/examples/bpf/augmented_raw_syscalls.c, so that it
1870          * copies only what we need for each syscall, like what happens when we
1871          * use syscalls:sys_enter_NAME, so that we reduce the kernel/userspace
1872          * traffic to just what is needed for each syscall.
1873          */
1874         int args_size = raw_augmented_args_size ?: sc->args_size;
1875
1876         *augmented_args_size = sample->raw_size - args_size;
1877         if (*augmented_args_size > 0)
1878                 augmented_args = sample->raw_data + args_size;
1879
1880         return augmented_args;
1881 }
1882
1883 static int trace__sys_enter(struct trace *trace, struct perf_evsel *evsel,
1884                             union perf_event *event __maybe_unused,
1885                             struct perf_sample *sample)
1886 {
1887         char *msg;
1888         void *args;
1889         int printed = 0;
1890         struct thread *thread;
1891         int id = perf_evsel__sc_tp_uint(evsel, id, sample), err = -1;
1892         int augmented_args_size = 0;
1893         void *augmented_args = NULL;
1894         struct syscall *sc = trace__syscall_info(trace, evsel, id);
1895         struct thread_trace *ttrace;
1896
1897         if (sc == NULL)
1898                 return -1;
1899
1900         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
1901         ttrace = thread__trace(thread, trace->output);
1902         if (ttrace == NULL)
1903                 goto out_put;
1904
1905         trace__fprintf_sample(trace, evsel, sample, thread);
1906
1907         args = perf_evsel__sc_tp_ptr(evsel, args, sample);
1908
1909         if (ttrace->entry_str == NULL) {
1910                 ttrace->entry_str = malloc(trace__entry_str_size);
1911                 if (!ttrace->entry_str)
1912                         goto out_put;
1913         }
1914
1915         if (!(trace->duration_filter || trace->summary_only || trace->min_stack))
1916                 trace__printf_interrupted_entry(trace);
1917         /*
1918          * If this is raw_syscalls.sys_enter, then it always comes with the 6 possible
1919          * arguments, even if the syscall being handled, say "openat", uses only 4 arguments
1920          * this breaks syscall__augmented_args() check for augmented args, as we calculate
1921          * syscall->args_size using each syscalls:sys_enter_NAME tracefs format file,
1922          * so when handling, say the openat syscall, we end up getting 6 args for the
1923          * raw_syscalls:sys_enter event, when we expected just 4, we end up mistakenly
1924          * thinking that the extra 2 u64 args are the augmented filename, so just check
1925          * here and avoid using augmented syscalls when the evsel is the raw_syscalls one.
1926          */
1927         if (evsel != trace->syscalls.events.sys_enter)
1928                 augmented_args = syscall__augmented_args(sc, sample, &augmented_args_size, trace->raw_augmented_syscalls_args_size);
1929         ttrace->entry_time = sample->time;
1930         msg = ttrace->entry_str;
1931         printed += scnprintf(msg + printed, trace__entry_str_size - printed, "%s(", sc->name);
1932
1933         printed += syscall__scnprintf_args(sc, msg + printed, trace__entry_str_size - printed,
1934                                            args, augmented_args, augmented_args_size, trace, thread);
1935
1936         if (sc->is_exit) {
1937                 if (!(trace->duration_filter || trace->summary_only || trace->failure_only || trace->min_stack)) {
1938                         int alignment = 0;
1939
1940                         trace__fprintf_entry_head(trace, thread, 0, false, ttrace->entry_time, trace->output);
1941                         printed = fprintf(trace->output, "%s)", ttrace->entry_str);
1942                         if (trace->args_alignment > printed)
1943                                 alignment = trace->args_alignment - printed;
1944                         fprintf(trace->output, "%*s= ?\n", alignment, " ");
1945                 }
1946         } else {
1947                 ttrace->entry_pending = true;
1948                 /* See trace__vfs_getname & trace__sys_exit */
1949                 ttrace->filename.pending_open = false;
1950         }
1951
1952         if (trace->current != thread) {
1953                 thread__put(trace->current);
1954                 trace->current = thread__get(thread);
1955         }
1956         err = 0;
1957 out_put:
1958         thread__put(thread);
1959         return err;
1960 }
1961
1962 static int trace__fprintf_sys_enter(struct trace *trace, struct perf_evsel *evsel,
1963                                     struct perf_sample *sample)
1964 {
1965         struct thread_trace *ttrace;
1966         struct thread *thread;
1967         int id = perf_evsel__sc_tp_uint(evsel, id, sample), err = -1;
1968         struct syscall *sc = trace__syscall_info(trace, evsel, id);
1969         char msg[1024];
1970         void *args, *augmented_args = NULL;
1971         int augmented_args_size;
1972
1973         if (sc == NULL)
1974                 return -1;
1975
1976         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
1977         ttrace = thread__trace(thread, trace->output);
1978         /*
1979          * We need to get ttrace just to make sure it is there when syscall__scnprintf_args()
1980          * and the rest of the beautifiers accessing it via struct syscall_arg touches it.
1981          */
1982         if (ttrace == NULL)
1983                 goto out_put;
1984
1985         args = perf_evsel__sc_tp_ptr(evsel, args, sample);
1986         augmented_args = syscall__augmented_args(sc, sample, &augmented_args_size, trace->raw_augmented_syscalls_args_size);
1987         syscall__scnprintf_args(sc, msg, sizeof(msg), args, augmented_args, augmented_args_size, trace, thread);
1988         fprintf(trace->output, "%s", msg);
1989         err = 0;
1990 out_put:
1991         thread__put(thread);
1992         return err;
1993 }
1994
1995 static int trace__resolve_callchain(struct trace *trace, struct perf_evsel *evsel,
1996                                     struct perf_sample *sample,
1997                                     struct callchain_cursor *cursor)
1998 {
1999         struct addr_location al;
2000         int max_stack = evsel->attr.sample_max_stack ?
2001                         evsel->attr.sample_max_stack :
2002                         trace->max_stack;
2003         int err;
2004
2005         if (machine__resolve(trace->host, &al, sample) < 0)
2006                 return -1;
2007
2008         err = thread__resolve_callchain(al.thread, cursor, evsel, sample, NULL, NULL, max_stack);
2009         addr_location__put(&al);
2010         return err;
2011 }
2012
2013 static int trace__fprintf_callchain(struct trace *trace, struct perf_sample *sample)
2014 {
2015         /* TODO: user-configurable print_opts */
2016         const unsigned int print_opts = EVSEL__PRINT_SYM |
2017                                         EVSEL__PRINT_DSO |
2018                                         EVSEL__PRINT_UNKNOWN_AS_ADDR;
2019
2020         return sample__fprintf_callchain(sample, 38, print_opts, &callchain_cursor, trace->output);
2021 }
2022
2023 static const char *errno_to_name(struct perf_evsel *evsel, int err)
2024 {
2025         struct perf_env *env = perf_evsel__env(evsel);
2026         const char *arch_name = perf_env__arch(env);
2027
2028         return arch_syscalls__strerrno(arch_name, err);
2029 }
2030
2031 static int trace__sys_exit(struct trace *trace, struct perf_evsel *evsel,
2032                            union perf_event *event __maybe_unused,
2033                            struct perf_sample *sample)
2034 {
2035         long ret;
2036         u64 duration = 0;
2037         bool duration_calculated = false;
2038         struct thread *thread;
2039         int id = perf_evsel__sc_tp_uint(evsel, id, sample), err = -1, callchain_ret = 0, printed = 0;
2040         int alignment = trace->args_alignment;
2041         struct syscall *sc = trace__syscall_info(trace, evsel, id);
2042         struct thread_trace *ttrace;
2043
2044         if (sc == NULL)
2045                 return -1;
2046
2047         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2048         ttrace = thread__trace(thread, trace->output);
2049         if (ttrace == NULL)
2050                 goto out_put;
2051
2052         trace__fprintf_sample(trace, evsel, sample, thread);
2053
2054         if (trace->summary)
2055                 thread__update_stats(ttrace, id, sample);
2056
2057         ret = perf_evsel__sc_tp_uint(evsel, ret, sample);
2058
2059         if (sc->is_open && ret >= 0 && ttrace->filename.pending_open) {
2060                 trace__set_fd_pathname(thread, ret, ttrace->filename.name);
2061                 ttrace->filename.pending_open = false;
2062                 ++trace->stats.vfs_getname;
2063         }
2064
2065         if (ttrace->entry_time) {
2066                 duration = sample->time - ttrace->entry_time;
2067                 if (trace__filter_duration(trace, duration))
2068                         goto out;
2069                 duration_calculated = true;
2070         } else if (trace->duration_filter)
2071                 goto out;
2072
2073         if (sample->callchain) {
2074                 callchain_ret = trace__resolve_callchain(trace, evsel, sample, &callchain_cursor);
2075                 if (callchain_ret == 0) {
2076                         if (callchain_cursor.nr < trace->min_stack)
2077                                 goto out;
2078                         callchain_ret = 1;
2079                 }
2080         }
2081
2082         if (trace->summary_only || (ret >= 0 && trace->failure_only))
2083                 goto out;
2084
2085         trace__fprintf_entry_head(trace, thread, duration, duration_calculated, ttrace->entry_time, trace->output);
2086
2087         if (ttrace->entry_pending) {
2088                 printed = fprintf(trace->output, "%s", ttrace->entry_str);
2089         } else {
2090                 printed += fprintf(trace->output, " ... [");
2091                 color_fprintf(trace->output, PERF_COLOR_YELLOW, "continued");
2092                 printed += 9;
2093                 printed += fprintf(trace->output, "]: %s()", sc->name);
2094         }
2095
2096         printed++; /* the closing ')' */
2097
2098         if (alignment > printed)
2099                 alignment -= printed;
2100         else
2101                 alignment = 0;
2102
2103         fprintf(trace->output, ")%*s= ", alignment, " ");
2104
2105         if (sc->fmt == NULL) {
2106                 if (ret < 0)
2107                         goto errno_print;
2108 signed_print:
2109                 fprintf(trace->output, "%ld", ret);
2110         } else if (ret < 0) {
2111 errno_print: {
2112                 char bf[STRERR_BUFSIZE];
2113                 const char *emsg = str_error_r(-ret, bf, sizeof(bf)),
2114                            *e = errno_to_name(evsel, -ret);
2115
2116                 fprintf(trace->output, "-1 %s (%s)", e, emsg);
2117         }
2118         } else if (ret == 0 && sc->fmt->timeout)
2119                 fprintf(trace->output, "0 (Timeout)");
2120         else if (ttrace->ret_scnprintf) {
2121                 char bf[1024];
2122                 struct syscall_arg arg = {
2123                         .val    = ret,
2124                         .thread = thread,
2125                         .trace  = trace,
2126                 };
2127                 ttrace->ret_scnprintf(bf, sizeof(bf), &arg);
2128                 ttrace->ret_scnprintf = NULL;
2129                 fprintf(trace->output, "%s", bf);
2130         } else if (sc->fmt->hexret)
2131                 fprintf(trace->output, "%#lx", ret);
2132         else if (sc->fmt->errpid) {
2133                 struct thread *child = machine__find_thread(trace->host, ret, ret);
2134
2135                 if (child != NULL) {
2136                         fprintf(trace->output, "%ld", ret);
2137                         if (child->comm_set)
2138                                 fprintf(trace->output, " (%s)", thread__comm_str(child));
2139                         thread__put(child);
2140                 }
2141         } else
2142                 goto signed_print;
2143
2144         fputc('\n', trace->output);
2145
2146         /*
2147          * We only consider an 'event' for the sake of --max-events a non-filtered
2148          * sys_enter + sys_exit and other tracepoint events.
2149          */
2150         if (++trace->nr_events_printed == trace->max_events && trace->max_events != ULONG_MAX)
2151                 interrupted = true;
2152
2153         if (callchain_ret > 0)
2154                 trace__fprintf_callchain(trace, sample);
2155         else if (callchain_ret < 0)
2156                 pr_err("Problem processing %s callchain, skipping...\n", perf_evsel__name(evsel));
2157 out:
2158         ttrace->entry_pending = false;
2159         err = 0;
2160 out_put:
2161         thread__put(thread);
2162         return err;
2163 }
2164
2165 static int trace__vfs_getname(struct trace *trace, struct perf_evsel *evsel,
2166                               union perf_event *event __maybe_unused,
2167                               struct perf_sample *sample)
2168 {
2169         struct thread *thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2170         struct thread_trace *ttrace;
2171         size_t filename_len, entry_str_len, to_move;
2172         ssize_t remaining_space;
2173         char *pos;
2174         const char *filename = perf_evsel__rawptr(evsel, sample, "pathname");
2175
2176         if (!thread)
2177                 goto out;
2178
2179         ttrace = thread__priv(thread);
2180         if (!ttrace)
2181                 goto out_put;
2182
2183         filename_len = strlen(filename);
2184         if (filename_len == 0)
2185                 goto out_put;
2186
2187         if (ttrace->filename.namelen < filename_len) {
2188                 char *f = realloc(ttrace->filename.name, filename_len + 1);
2189
2190                 if (f == NULL)
2191                         goto out_put;
2192
2193                 ttrace->filename.namelen = filename_len;
2194                 ttrace->filename.name = f;
2195         }
2196
2197         strcpy(ttrace->filename.name, filename);
2198         ttrace->filename.pending_open = true;
2199
2200         if (!ttrace->filename.ptr)
2201                 goto out_put;
2202
2203         entry_str_len = strlen(ttrace->entry_str);
2204         remaining_space = trace__entry_str_size - entry_str_len - 1; /* \0 */
2205         if (remaining_space <= 0)
2206                 goto out_put;
2207
2208         if (filename_len > (size_t)remaining_space) {
2209                 filename += filename_len - remaining_space;
2210                 filename_len = remaining_space;
2211         }
2212
2213         to_move = entry_str_len - ttrace->filename.entry_str_pos + 1; /* \0 */
2214         pos = ttrace->entry_str + ttrace->filename.entry_str_pos;
2215         memmove(pos + filename_len, pos, to_move);
2216         memcpy(pos, filename, filename_len);
2217
2218         ttrace->filename.ptr = 0;
2219         ttrace->filename.entry_str_pos = 0;
2220 out_put:
2221         thread__put(thread);
2222 out:
2223         return 0;
2224 }
2225
2226 static int trace__sched_stat_runtime(struct trace *trace, struct perf_evsel *evsel,
2227                                      union perf_event *event __maybe_unused,
2228                                      struct perf_sample *sample)
2229 {
2230         u64 runtime = perf_evsel__intval(evsel, sample, "runtime");
2231         double runtime_ms = (double)runtime / NSEC_PER_MSEC;
2232         struct thread *thread = machine__findnew_thread(trace->host,
2233                                                         sample->pid,
2234                                                         sample->tid);
2235         struct thread_trace *ttrace = thread__trace(thread, trace->output);
2236
2237         if (ttrace == NULL)
2238                 goto out_dump;
2239
2240         ttrace->runtime_ms += runtime_ms;
2241         trace->runtime_ms += runtime_ms;
2242 out_put:
2243         thread__put(thread);
2244         return 0;
2245
2246 out_dump:
2247         fprintf(trace->output, "%s: comm=%s,pid=%u,runtime=%" PRIu64 ",vruntime=%" PRIu64 ")\n",
2248                evsel->name,
2249                perf_evsel__strval(evsel, sample, "comm"),
2250                (pid_t)perf_evsel__intval(evsel, sample, "pid"),
2251                runtime,
2252                perf_evsel__intval(evsel, sample, "vruntime"));
2253         goto out_put;
2254 }
2255
2256 static int bpf_output__printer(enum binary_printer_ops op,
2257                                unsigned int val, void *extra __maybe_unused, FILE *fp)
2258 {
2259         unsigned char ch = (unsigned char)val;
2260
2261         switch (op) {
2262         case BINARY_PRINT_CHAR_DATA:
2263                 return fprintf(fp, "%c", isprint(ch) ? ch : '.');
2264         case BINARY_PRINT_DATA_BEGIN:
2265         case BINARY_PRINT_LINE_BEGIN:
2266         case BINARY_PRINT_ADDR:
2267         case BINARY_PRINT_NUM_DATA:
2268         case BINARY_PRINT_NUM_PAD:
2269         case BINARY_PRINT_SEP:
2270         case BINARY_PRINT_CHAR_PAD:
2271         case BINARY_PRINT_LINE_END:
2272         case BINARY_PRINT_DATA_END:
2273         default:
2274                 break;
2275         }
2276
2277         return 0;
2278 }
2279
2280 static void bpf_output__fprintf(struct trace *trace,
2281                                 struct perf_sample *sample)
2282 {
2283         binary__fprintf(sample->raw_data, sample->raw_size, 8,
2284                         bpf_output__printer, NULL, trace->output);
2285         ++trace->nr_events_printed;
2286 }
2287
2288 static int trace__event_handler(struct trace *trace, struct perf_evsel *evsel,
2289                                 union perf_event *event __maybe_unused,
2290                                 struct perf_sample *sample)
2291 {
2292         struct thread *thread;
2293         int callchain_ret = 0;
2294         /*
2295          * Check if we called perf_evsel__disable(evsel) due to, for instance,
2296          * this event's max_events having been hit and this is an entry coming
2297          * from the ring buffer that we should discard, since the max events
2298          * have already been considered/printed.
2299          */
2300         if (evsel->disabled)
2301                 return 0;
2302
2303         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2304
2305         if (sample->callchain) {
2306                 callchain_ret = trace__resolve_callchain(trace, evsel, sample, &callchain_cursor);
2307                 if (callchain_ret == 0) {
2308                         if (callchain_cursor.nr < trace->min_stack)
2309                                 goto out;
2310                         callchain_ret = 1;
2311                 }
2312         }
2313
2314         trace__printf_interrupted_entry(trace);
2315         trace__fprintf_tstamp(trace, sample->time, trace->output);
2316
2317         if (trace->trace_syscalls && trace->show_duration)
2318                 fprintf(trace->output, "(         ): ");
2319
2320         if (thread)
2321                 trace__fprintf_comm_tid(trace, thread, trace->output);
2322
2323         if (evsel == trace->syscalls.events.augmented) {
2324                 int id = perf_evsel__sc_tp_uint(evsel, id, sample);
2325                 struct syscall *sc = trace__syscall_info(trace, evsel, id);
2326
2327                 if (sc) {
2328                         fprintf(trace->output, "%s(", sc->name);
2329                         trace__fprintf_sys_enter(trace, evsel, sample);
2330                         fputc(')', trace->output);
2331                         goto newline;
2332                 }
2333
2334                 /*
2335                  * XXX: Not having the associated syscall info or not finding/adding
2336                  *      the thread should never happen, but if it does...
2337                  *      fall thru and print it as a bpf_output event.
2338                  */
2339         }
2340
2341         fprintf(trace->output, "%s:", evsel->name);
2342
2343         if (perf_evsel__is_bpf_output(evsel)) {
2344                 bpf_output__fprintf(trace, sample);
2345         } else if (evsel->tp_format) {
2346                 if (strncmp(evsel->tp_format->name, "sys_enter_", 10) ||
2347                     trace__fprintf_sys_enter(trace, evsel, sample)) {
2348                         event_format__fprintf(evsel->tp_format, sample->cpu,
2349                                               sample->raw_data, sample->raw_size,
2350                                               trace->output);
2351                         ++trace->nr_events_printed;
2352
2353                         if (evsel->max_events != ULONG_MAX && ++evsel->nr_events_printed == evsel->max_events) {
2354                                 perf_evsel__disable(evsel);
2355                                 perf_evsel__close(evsel);
2356                         }
2357                 }
2358         }
2359
2360 newline:
2361         fprintf(trace->output, "\n");
2362
2363         if (callchain_ret > 0)
2364                 trace__fprintf_callchain(trace, sample);
2365         else if (callchain_ret < 0)
2366                 pr_err("Problem processing %s callchain, skipping...\n", perf_evsel__name(evsel));
2367 out:
2368         thread__put(thread);
2369         return 0;
2370 }
2371
2372 static void print_location(FILE *f, struct perf_sample *sample,
2373                            struct addr_location *al,
2374                            bool print_dso, bool print_sym)
2375 {
2376
2377         if ((verbose > 0 || print_dso) && al->map)
2378                 fprintf(f, "%s@", al->map->dso->long_name);
2379
2380         if ((verbose > 0 || print_sym) && al->sym)
2381                 fprintf(f, "%s+0x%" PRIx64, al->sym->name,
2382                         al->addr - al->sym->start);
2383         else if (al->map)
2384                 fprintf(f, "0x%" PRIx64, al->addr);
2385         else
2386                 fprintf(f, "0x%" PRIx64, sample->addr);
2387 }
2388
2389 static int trace__pgfault(struct trace *trace,
2390                           struct perf_evsel *evsel,
2391                           union perf_event *event __maybe_unused,
2392                           struct perf_sample *sample)
2393 {
2394         struct thread *thread;
2395         struct addr_location al;
2396         char map_type = 'd';
2397         struct thread_trace *ttrace;
2398         int err = -1;
2399         int callchain_ret = 0;
2400
2401         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2402
2403         if (sample->callchain) {
2404                 callchain_ret = trace__resolve_callchain(trace, evsel, sample, &callchain_cursor);
2405                 if (callchain_ret == 0) {
2406                         if (callchain_cursor.nr < trace->min_stack)
2407                                 goto out_put;
2408                         callchain_ret = 1;
2409                 }
2410         }
2411
2412         ttrace = thread__trace(thread, trace->output);
2413         if (ttrace == NULL)
2414                 goto out_put;
2415
2416         if (evsel->attr.config == PERF_COUNT_SW_PAGE_FAULTS_MAJ)
2417                 ttrace->pfmaj++;
2418         else
2419                 ttrace->pfmin++;
2420
2421         if (trace->summary_only)
2422                 goto out;
2423
2424         thread__find_symbol(thread, sample->cpumode, sample->ip, &al);
2425
2426         trace__fprintf_entry_head(trace, thread, 0, true, sample->time, trace->output);
2427
2428         fprintf(trace->output, "%sfault [",
2429                 evsel->attr.config == PERF_COUNT_SW_PAGE_FAULTS_MAJ ?
2430                 "maj" : "min");
2431
2432         print_location(trace->output, sample, &al, false, true);
2433
2434         fprintf(trace->output, "] => ");
2435
2436         thread__find_symbol(thread, sample->cpumode, sample->addr, &al);
2437
2438         if (!al.map) {
2439                 thread__find_symbol(thread, sample->cpumode, sample->addr, &al);
2440
2441                 if (al.map)
2442                         map_type = 'x';
2443                 else
2444                         map_type = '?';
2445         }
2446
2447         print_location(trace->output, sample, &al, true, false);
2448
2449         fprintf(trace->output, " (%c%c)\n", map_type, al.level);
2450
2451         if (callchain_ret > 0)
2452                 trace__fprintf_callchain(trace, sample);
2453         else if (callchain_ret < 0)
2454                 pr_err("Problem processing %s callchain, skipping...\n", perf_evsel__name(evsel));
2455
2456         ++trace->nr_events_printed;
2457 out:
2458         err = 0;
2459 out_put:
2460         thread__put(thread);
2461         return err;
2462 }
2463
2464 static void trace__set_base_time(struct trace *trace,
2465                                  struct perf_evsel *evsel,
2466                                  struct perf_sample *sample)
2467 {
2468         /*
2469          * BPF events were not setting PERF_SAMPLE_TIME, so be more robust
2470          * and don't use sample->time unconditionally, we may end up having
2471          * some other event in the future without PERF_SAMPLE_TIME for good
2472          * reason, i.e. we may not be interested in its timestamps, just in
2473          * it taking place, picking some piece of information when it
2474          * appears in our event stream (vfs_getname comes to mind).
2475          */
2476         if (trace->base_time == 0 && !trace->full_time &&
2477             (evsel->attr.sample_type & PERF_SAMPLE_TIME))
2478                 trace->base_time = sample->time;
2479 }
2480
2481 static int trace__process_sample(struct perf_tool *tool,
2482                                  union perf_event *event,
2483                                  struct perf_sample *sample,
2484                                  struct perf_evsel *evsel,
2485                                  struct machine *machine __maybe_unused)
2486 {
2487         struct trace *trace = container_of(tool, struct trace, tool);
2488         struct thread *thread;
2489         int err = 0;
2490
2491         tracepoint_handler handler = evsel->handler;
2492
2493         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2494         if (thread && thread__is_filtered(thread))
2495                 goto out;
2496
2497         trace__set_base_time(trace, evsel, sample);
2498
2499         if (handler) {
2500                 ++trace->nr_events;
2501                 handler(trace, evsel, event, sample);
2502         }
2503 out:
2504         thread__put(thread);
2505         return err;
2506 }
2507
2508 static int trace__record(struct trace *trace, int argc, const char **argv)
2509 {
2510         unsigned int rec_argc, i, j;
2511         const char **rec_argv;
2512         const char * const record_args[] = {
2513                 "record",
2514                 "-R",
2515                 "-m", "1024",
2516                 "-c", "1",
2517         };
2518
2519         const char * const sc_args[] = { "-e", };
2520         unsigned int sc_args_nr = ARRAY_SIZE(sc_args);
2521         const char * const majpf_args[] = { "-e", "major-faults" };
2522         unsigned int majpf_args_nr = ARRAY_SIZE(majpf_args);
2523         const char * const minpf_args[] = { "-e", "minor-faults" };
2524         unsigned int minpf_args_nr = ARRAY_SIZE(minpf_args);
2525
2526         /* +1 is for the event string below */
2527         rec_argc = ARRAY_SIZE(record_args) + sc_args_nr + 1 +
2528                 majpf_args_nr + minpf_args_nr + argc;
2529         rec_argv = calloc(rec_argc + 1, sizeof(char *));
2530
2531         if (rec_argv == NULL)
2532                 return -ENOMEM;
2533
2534         j = 0;
2535         for (i = 0; i < ARRAY_SIZE(record_args); i++)
2536                 rec_argv[j++] = record_args[i];
2537
2538         if (trace->trace_syscalls) {
2539                 for (i = 0; i < sc_args_nr; i++)
2540                         rec_argv[j++] = sc_args[i];
2541
2542                 /* event string may be different for older kernels - e.g., RHEL6 */
2543                 if (is_valid_tracepoint("raw_syscalls:sys_enter"))
2544                         rec_argv[j++] = "raw_syscalls:sys_enter,raw_syscalls:sys_exit";
2545                 else if (is_valid_tracepoint("syscalls:sys_enter"))
2546                         rec_argv[j++] = "syscalls:sys_enter,syscalls:sys_exit";
2547                 else {
2548                         pr_err("Neither raw_syscalls nor syscalls events exist.\n");
2549                         free(rec_argv);
2550                         return -1;
2551                 }
2552         }
2553
2554         if (trace->trace_pgfaults & TRACE_PFMAJ)
2555                 for (i = 0; i < majpf_args_nr; i++)
2556                         rec_argv[j++] = majpf_args[i];
2557
2558         if (trace->trace_pgfaults & TRACE_PFMIN)
2559                 for (i = 0; i < minpf_args_nr; i++)
2560                         rec_argv[j++] = minpf_args[i];
2561
2562         for (i = 0; i < (unsigned int)argc; i++)
2563                 rec_argv[j++] = argv[i];
2564
2565         return cmd_record(j, rec_argv);
2566 }
2567
2568 static size_t trace__fprintf_thread_summary(struct trace *trace, FILE *fp);
2569
2570 static bool perf_evlist__add_vfs_getname(struct perf_evlist *evlist)
2571 {
2572         bool found = false;
2573         struct perf_evsel *evsel, *tmp;
2574         struct parse_events_error err = { .idx = 0, };
2575         int ret = parse_events(evlist, "probe:vfs_getname*", &err);
2576
2577         if (ret)
2578                 return false;
2579
2580         evlist__for_each_entry_safe(evlist, evsel, tmp) {
2581                 if (!strstarts(perf_evsel__name(evsel), "probe:vfs_getname"))
2582                         continue;
2583
2584                 if (perf_evsel__field(evsel, "pathname")) {
2585                         evsel->handler = trace__vfs_getname;
2586                         found = true;
2587                         continue;
2588                 }
2589
2590                 list_del_init(&evsel->node);
2591                 evsel->evlist = NULL;
2592                 perf_evsel__delete(evsel);
2593         }
2594
2595         return found;
2596 }
2597
2598 static struct perf_evsel *perf_evsel__new_pgfault(u64 config)
2599 {
2600         struct perf_evsel *evsel;
2601         struct perf_event_attr attr = {
2602                 .type = PERF_TYPE_SOFTWARE,
2603                 .mmap_data = 1,
2604         };
2605
2606         attr.config = config;
2607         attr.sample_period = 1;
2608
2609         event_attr_init(&attr);
2610
2611         evsel = perf_evsel__new(&attr);
2612         if (evsel)
2613                 evsel->handler = trace__pgfault;
2614
2615         return evsel;
2616 }
2617
2618 static void trace__handle_event(struct trace *trace, union perf_event *event, struct perf_sample *sample)
2619 {
2620         const u32 type = event->header.type;
2621         struct perf_evsel *evsel;
2622
2623         if (type != PERF_RECORD_SAMPLE) {
2624                 trace__process_event(trace, trace->host, event, sample);
2625                 return;
2626         }
2627
2628         evsel = perf_evlist__id2evsel(trace->evlist, sample->id);
2629         if (evsel == NULL) {
2630                 fprintf(trace->output, "Unknown tp ID %" PRIu64 ", skipping...\n", sample->id);
2631                 return;
2632         }
2633
2634         trace__set_base_time(trace, evsel, sample);
2635
2636         if (evsel->attr.type == PERF_TYPE_TRACEPOINT &&
2637             sample->raw_data == NULL) {
2638                 fprintf(trace->output, "%s sample with no payload for tid: %d, cpu %d, raw_size=%d, skipping...\n",
2639                        perf_evsel__name(evsel), sample->tid,
2640                        sample->cpu, sample->raw_size);
2641         } else {
2642                 tracepoint_handler handler = evsel->handler;
2643                 handler(trace, evsel, event, sample);
2644         }
2645
2646         if (trace->nr_events_printed >= trace->max_events && trace->max_events != ULONG_MAX)
2647                 interrupted = true;
2648 }
2649
2650 static int trace__add_syscall_newtp(struct trace *trace)
2651 {
2652         int ret = -1;
2653         struct perf_evlist *evlist = trace->evlist;
2654         struct perf_evsel *sys_enter, *sys_exit;
2655
2656         sys_enter = perf_evsel__raw_syscall_newtp("sys_enter", trace__sys_enter);
2657         if (sys_enter == NULL)
2658                 goto out;
2659
2660         if (perf_evsel__init_sc_tp_ptr_field(sys_enter, args))
2661                 goto out_delete_sys_enter;
2662
2663         sys_exit = perf_evsel__raw_syscall_newtp("sys_exit", trace__sys_exit);
2664         if (sys_exit == NULL)
2665                 goto out_delete_sys_enter;
2666
2667         if (perf_evsel__init_sc_tp_uint_field(sys_exit, ret))
2668                 goto out_delete_sys_exit;
2669
2670         perf_evsel__config_callchain(sys_enter, &trace->opts, &callchain_param);
2671         perf_evsel__config_callchain(sys_exit, &trace->opts, &callchain_param);
2672
2673         perf_evlist__add(evlist, sys_enter);
2674         perf_evlist__add(evlist, sys_exit);
2675
2676         if (callchain_param.enabled && !trace->kernel_syscallchains) {
2677                 /*
2678                  * We're interested only in the user space callchain
2679                  * leading to the syscall, allow overriding that for
2680                  * debugging reasons using --kernel_syscall_callchains
2681                  */
2682                 sys_exit->attr.exclude_callchain_kernel = 1;
2683         }
2684
2685         trace->syscalls.events.sys_enter = sys_enter;
2686         trace->syscalls.events.sys_exit  = sys_exit;
2687
2688         ret = 0;
2689 out:
2690         return ret;
2691
2692 out_delete_sys_exit:
2693         perf_evsel__delete_priv(sys_exit);
2694 out_delete_sys_enter:
2695         perf_evsel__delete_priv(sys_enter);
2696         goto out;
2697 }
2698
2699 static int trace__set_ev_qualifier_tp_filter(struct trace *trace)
2700 {
2701         int err = -1;
2702         struct perf_evsel *sys_exit;
2703         char *filter = asprintf_expr_inout_ints("id", !trace->not_ev_qualifier,
2704                                                 trace->ev_qualifier_ids.nr,
2705                                                 trace->ev_qualifier_ids.entries);
2706
2707         if (filter == NULL)
2708                 goto out_enomem;
2709
2710         if (!perf_evsel__append_tp_filter(trace->syscalls.events.sys_enter,
2711                                           filter)) {
2712                 sys_exit = trace->syscalls.events.sys_exit;
2713                 err = perf_evsel__append_tp_filter(sys_exit, filter);
2714         }
2715
2716         free(filter);
2717 out:
2718         return err;
2719 out_enomem:
2720         errno = ENOMEM;
2721         goto out;
2722 }
2723
2724 #ifdef HAVE_LIBBPF_SUPPORT
2725 static void trace__init_bpf_map_syscall_args(struct trace *trace, int id, struct bpf_map_syscall_entry *entry)
2726 {
2727         struct syscall *sc = trace__syscall_info(trace, NULL, id);
2728         int arg = 0;
2729
2730         if (sc == NULL)
2731                 goto out;
2732
2733         for (; arg < sc->nr_args; ++arg) {
2734                 entry->string_args_len[arg] = 0;
2735                 if (sc->arg_fmt[arg].scnprintf == SCA_FILENAME) {
2736                         /* Should be set like strace -s strsize */
2737                         entry->string_args_len[arg] = PATH_MAX;
2738                 }
2739         }
2740 out:
2741         for (; arg < 6; ++arg)
2742                 entry->string_args_len[arg] = 0;
2743 }
2744 static int trace__set_ev_qualifier_bpf_filter(struct trace *trace)
2745 {
2746         int fd = bpf_map__fd(trace->syscalls.map);
2747         struct bpf_map_syscall_entry value = {
2748                 .enabled = !trace->not_ev_qualifier,
2749         };
2750         int err = 0;
2751         size_t i;
2752
2753         for (i = 0; i < trace->ev_qualifier_ids.nr; ++i) {
2754                 int key = trace->ev_qualifier_ids.entries[i];
2755
2756                 if (value.enabled)
2757                         trace__init_bpf_map_syscall_args(trace, key, &value);
2758
2759                 err = bpf_map_update_elem(fd, &key, &value, BPF_EXIST);
2760                 if (err)
2761                         break;
2762         }
2763
2764         return err;
2765 }
2766
2767 static int __trace__init_syscalls_bpf_map(struct trace *trace, bool enabled)
2768 {
2769         int fd = bpf_map__fd(trace->syscalls.map);
2770         struct bpf_map_syscall_entry value = {
2771                 .enabled = enabled,
2772         };
2773         int err = 0, key;
2774
2775         for (key = 0; key < trace->sctbl->syscalls.nr_entries; ++key) {
2776                 if (enabled)
2777                         trace__init_bpf_map_syscall_args(trace, key, &value);
2778
2779                 err = bpf_map_update_elem(fd, &key, &value, BPF_ANY);
2780                 if (err)
2781                         break;
2782         }
2783
2784         return err;
2785 }
2786
2787 static int trace__init_syscalls_bpf_map(struct trace *trace)
2788 {
2789         bool enabled = true;
2790
2791         if (trace->ev_qualifier_ids.nr)
2792                 enabled = trace->not_ev_qualifier;
2793
2794         return __trace__init_syscalls_bpf_map(trace, enabled);
2795 }
2796 #else
2797 static int trace__set_ev_qualifier_bpf_filter(struct trace *trace __maybe_unused)
2798 {
2799         return 0;
2800 }
2801
2802 static int trace__init_syscalls_bpf_map(struct trace *trace __maybe_unused)
2803 {
2804         return 0;
2805 }
2806 #endif // HAVE_LIBBPF_SUPPORT
2807
2808 static int trace__set_ev_qualifier_filter(struct trace *trace)
2809 {
2810         if (trace->syscalls.map)
2811                 return trace__set_ev_qualifier_bpf_filter(trace);
2812         if (trace->syscalls.events.sys_enter)
2813                 return trace__set_ev_qualifier_tp_filter(trace);
2814         return 0;
2815 }
2816
2817 static int bpf_map__set_filter_pids(struct bpf_map *map __maybe_unused,
2818                                     size_t npids __maybe_unused, pid_t *pids __maybe_unused)
2819 {
2820         int err = 0;
2821 #ifdef HAVE_LIBBPF_SUPPORT
2822         bool value = true;
2823         int map_fd = bpf_map__fd(map);
2824         size_t i;
2825
2826         for (i = 0; i < npids; ++i) {
2827                 err = bpf_map_update_elem(map_fd, &pids[i], &value, BPF_ANY);
2828                 if (err)
2829                         break;
2830         }
2831 #endif
2832         return err;
2833 }
2834
2835 static int trace__set_filter_loop_pids(struct trace *trace)
2836 {
2837         unsigned int nr = 1, err;
2838         pid_t pids[32] = {
2839                 getpid(),
2840         };
2841         struct thread *thread = machine__find_thread(trace->host, pids[0], pids[0]);
2842
2843         while (thread && nr < ARRAY_SIZE(pids)) {
2844                 struct thread *parent = machine__find_thread(trace->host, thread->ppid, thread->ppid);
2845
2846                 if (parent == NULL)
2847                         break;
2848
2849                 if (!strcmp(thread__comm_str(parent), "sshd") ||
2850                     strstarts(thread__comm_str(parent), "gnome-terminal")) {
2851                         pids[nr++] = parent->tid;
2852                         break;
2853                 }
2854                 thread = parent;
2855         }
2856
2857         err = perf_evlist__set_tp_filter_pids(trace->evlist, nr, pids);
2858         if (!err && trace->filter_pids.map)
2859                 err = bpf_map__set_filter_pids(trace->filter_pids.map, nr, pids);
2860
2861         return err;
2862 }
2863
2864 static int trace__set_filter_pids(struct trace *trace)
2865 {
2866         int err = 0;
2867         /*
2868          * Better not use !target__has_task() here because we need to cover the
2869          * case where no threads were specified in the command line, but a
2870          * workload was, and in that case we will fill in the thread_map when
2871          * we fork the workload in perf_evlist__prepare_workload.
2872          */
2873         if (trace->filter_pids.nr > 0) {
2874                 err = perf_evlist__set_tp_filter_pids(trace->evlist, trace->filter_pids.nr,
2875                                                       trace->filter_pids.entries);
2876                 if (!err && trace->filter_pids.map) {
2877                         err = bpf_map__set_filter_pids(trace->filter_pids.map, trace->filter_pids.nr,
2878                                                        trace->filter_pids.entries);
2879                 }
2880         } else if (thread_map__pid(trace->evlist->threads, 0) == -1) {
2881                 err = trace__set_filter_loop_pids(trace);
2882         }
2883
2884         return err;
2885 }
2886
2887 static int __trace__deliver_event(struct trace *trace, union perf_event *event)
2888 {
2889         struct perf_evlist *evlist = trace->evlist;
2890         struct perf_sample sample;
2891         int err;
2892
2893         err = perf_evlist__parse_sample(evlist, event, &sample);
2894         if (err)
2895                 fprintf(trace->output, "Can't parse sample, err = %d, skipping...\n", err);
2896         else
2897                 trace__handle_event(trace, event, &sample);
2898
2899         return 0;
2900 }
2901
2902 static int __trace__flush_events(struct trace *trace)
2903 {
2904         u64 first = ordered_events__first_time(&trace->oe.data);
2905         u64 flush = trace->oe.last - NSEC_PER_SEC;
2906
2907         /* Is there some thing to flush.. */
2908         if (first && first < flush)
2909                 return ordered_events__flush_time(&trace->oe.data, flush);
2910
2911         return 0;
2912 }
2913
2914 static int trace__flush_events(struct trace *trace)
2915 {
2916         return !trace->sort_events ? 0 : __trace__flush_events(trace);
2917 }
2918
2919 static int trace__deliver_event(struct trace *trace, union perf_event *event)
2920 {
2921         int err;
2922
2923         if (!trace->sort_events)
2924                 return __trace__deliver_event(trace, event);
2925
2926         err = perf_evlist__parse_sample_timestamp(trace->evlist, event, &trace->oe.last);
2927         if (err && err != -1)
2928                 return err;
2929
2930         err = ordered_events__queue(&trace->oe.data, event, trace->oe.last, 0);
2931         if (err)
2932                 return err;
2933
2934         return trace__flush_events(trace);
2935 }
2936
2937 static int ordered_events__deliver_event(struct ordered_events *oe,
2938                                          struct ordered_event *event)
2939 {
2940         struct trace *trace = container_of(oe, struct trace, oe.data);
2941
2942         return __trace__deliver_event(trace, event->event);
2943 }
2944
2945 static int trace__run(struct trace *trace, int argc, const char **argv)
2946 {
2947         struct perf_evlist *evlist = trace->evlist;
2948         struct perf_evsel *evsel, *pgfault_maj = NULL, *pgfault_min = NULL;
2949         int err = -1, i;
2950         unsigned long before;
2951         const bool forks = argc > 0;
2952         bool draining = false;
2953
2954         trace->live = true;
2955
2956         if (!trace->raw_augmented_syscalls) {
2957                 if (trace->trace_syscalls && trace__add_syscall_newtp(trace))
2958                         goto out_error_raw_syscalls;
2959
2960                 if (trace->trace_syscalls)
2961                         trace->vfs_getname = perf_evlist__add_vfs_getname(evlist);
2962         }
2963
2964         if ((trace->trace_pgfaults & TRACE_PFMAJ)) {
2965                 pgfault_maj = perf_evsel__new_pgfault(PERF_COUNT_SW_PAGE_FAULTS_MAJ);
2966                 if (pgfault_maj == NULL)
2967                         goto out_error_mem;
2968                 perf_evsel__config_callchain(pgfault_maj, &trace->opts, &callchain_param);
2969                 perf_evlist__add(evlist, pgfault_maj);
2970         }
2971
2972         if ((trace->trace_pgfaults & TRACE_PFMIN)) {
2973                 pgfault_min = perf_evsel__new_pgfault(PERF_COUNT_SW_PAGE_FAULTS_MIN);
2974                 if (pgfault_min == NULL)
2975                         goto out_error_mem;
2976                 perf_evsel__config_callchain(pgfault_min, &trace->opts, &callchain_param);
2977                 perf_evlist__add(evlist, pgfault_min);
2978         }
2979
2980         if (trace->sched &&
2981             perf_evlist__add_newtp(evlist, "sched", "sched_stat_runtime",
2982                                    trace__sched_stat_runtime))
2983                 goto out_error_sched_stat_runtime;
2984
2985         /*
2986          * If a global cgroup was set, apply it to all the events without an
2987          * explicit cgroup. I.e.:
2988          *
2989          *      trace -G A -e sched:*switch
2990          *
2991          * Will set all raw_syscalls:sys_{enter,exit}, pgfault, vfs_getname, etc
2992          * _and_ sched:sched_switch to the 'A' cgroup, while:
2993          *
2994          * trace -e sched:*switch -G A
2995          *
2996          * will only set the sched:sched_switch event to the 'A' cgroup, all the
2997          * other events (raw_syscalls:sys_{enter,exit}, etc are left "without"
2998          * a cgroup (on the root cgroup, sys wide, etc).
2999          *
3000          * Multiple cgroups:
3001          *
3002          * trace -G A -e sched:*switch -G B
3003          *
3004          * the syscall ones go to the 'A' cgroup, the sched:sched_switch goes
3005          * to the 'B' cgroup.
3006          *
3007          * evlist__set_default_cgroup() grabs a reference of the passed cgroup
3008          * only for the evsels still without a cgroup, i.e. evsel->cgroup == NULL.
3009          */
3010         if (trace->cgroup)
3011                 evlist__set_default_cgroup(trace->evlist, trace->cgroup);
3012
3013         err = perf_evlist__create_maps(evlist, &trace->opts.target);
3014         if (err < 0) {
3015                 fprintf(trace->output, "Problems parsing the target to trace, check your options!\n");
3016                 goto out_delete_evlist;
3017         }
3018
3019         err = trace__symbols_init(trace, evlist);
3020         if (err < 0) {
3021                 fprintf(trace->output, "Problems initializing symbol libraries!\n");
3022                 goto out_delete_evlist;
3023         }
3024
3025         perf_evlist__config(evlist, &trace->opts, &callchain_param);
3026
3027         signal(SIGCHLD, sig_handler);
3028         signal(SIGINT, sig_handler);
3029
3030         if (forks) {
3031                 err = perf_evlist__prepare_workload(evlist, &trace->opts.target,
3032                                                     argv, false, NULL);
3033                 if (err < 0) {
3034                         fprintf(trace->output, "Couldn't run the workload!\n");
3035                         goto out_delete_evlist;
3036                 }
3037         }
3038
3039         err = perf_evlist__open(evlist);
3040         if (err < 0)
3041                 goto out_error_open;
3042
3043         err = bpf__apply_obj_config();
3044         if (err) {
3045                 char errbuf[BUFSIZ];
3046
3047                 bpf__strerror_apply_obj_config(err, errbuf, sizeof(errbuf));
3048                 pr_err("ERROR: Apply config to BPF failed: %s\n",
3049                          errbuf);
3050                 goto out_error_open;
3051         }
3052
3053         err = trace__set_filter_pids(trace);
3054         if (err < 0)
3055                 goto out_error_mem;
3056
3057         if (trace->syscalls.map)
3058                 trace__init_syscalls_bpf_map(trace);
3059
3060         if (trace->ev_qualifier_ids.nr > 0) {
3061                 err = trace__set_ev_qualifier_filter(trace);
3062                 if (err < 0)
3063                         goto out_errno;
3064
3065                 if (trace->syscalls.events.sys_exit) {
3066                         pr_debug("event qualifier tracepoint filter: %s\n",
3067                                  trace->syscalls.events.sys_exit->filter);
3068                 }
3069         }
3070
3071         err = perf_evlist__apply_filters(evlist, &evsel);
3072         if (err < 0)
3073                 goto out_error_apply_filters;
3074
3075         if (trace->dump.map)
3076                 bpf_map__fprintf(trace->dump.map, trace->output);
3077
3078         err = perf_evlist__mmap(evlist, trace->opts.mmap_pages);
3079         if (err < 0)
3080                 goto out_error_mmap;
3081
3082         if (!target__none(&trace->opts.target) && !trace->opts.initial_delay)
3083                 perf_evlist__enable(evlist);
3084
3085         if (forks)
3086                 perf_evlist__start_workload(evlist);
3087
3088         if (trace->opts.initial_delay) {
3089                 usleep(trace->opts.initial_delay * 1000);
3090                 perf_evlist__enable(evlist);
3091         }
3092
3093         trace->multiple_threads = thread_map__pid(evlist->threads, 0) == -1 ||
3094                                   evlist->threads->nr > 1 ||
3095                                   perf_evlist__first(evlist)->attr.inherit;
3096
3097         /*
3098          * Now that we already used evsel->attr to ask the kernel to setup the
3099          * events, lets reuse evsel->attr.sample_max_stack as the limit in
3100          * trace__resolve_callchain(), allowing per-event max-stack settings
3101          * to override an explicitly set --max-stack global setting.
3102          */
3103         evlist__for_each_entry(evlist, evsel) {
3104                 if (evsel__has_callchain(evsel) &&
3105                     evsel->attr.sample_max_stack == 0)
3106                         evsel->attr.sample_max_stack = trace->max_stack;
3107         }
3108 again:
3109         before = trace->nr_events;
3110
3111         for (i = 0; i < evlist->nr_mmaps; i++) {
3112                 union perf_event *event;
3113                 struct perf_mmap *md;
3114
3115                 md = &evlist->mmap[i];
3116                 if (perf_mmap__read_init(md) < 0)
3117                         continue;
3118
3119                 while ((event = perf_mmap__read_event(md)) != NULL) {
3120                         ++trace->nr_events;
3121
3122                         err = trace__deliver_event(trace, event);
3123                         if (err)
3124                                 goto out_disable;
3125
3126                         perf_mmap__consume(md);
3127
3128                         if (interrupted)
3129                                 goto out_disable;
3130
3131                         if (done && !draining) {
3132                                 perf_evlist__disable(evlist);
3133                                 draining = true;
3134                         }
3135                 }
3136                 perf_mmap__read_done(md);
3137         }
3138
3139         if (trace->nr_events == before) {
3140                 int timeout = done ? 100 : -1;
3141
3142                 if (!draining && perf_evlist__poll(evlist, timeout) > 0) {
3143                         if (perf_evlist__filter_pollfd(evlist, POLLERR | POLLHUP | POLLNVAL) == 0)
3144                                 draining = true;
3145
3146                         goto again;
3147                 } else {
3148                         if (trace__flush_events(trace))
3149                                 goto out_disable;
3150                 }
3151         } else {
3152                 goto again;
3153         }
3154
3155 out_disable:
3156         thread__zput(trace->current);
3157
3158         perf_evlist__disable(evlist);
3159
3160         if (trace->sort_events)
3161                 ordered_events__flush(&trace->oe.data, OE_FLUSH__FINAL);
3162
3163         if (!err) {
3164                 if (trace->summary)
3165                         trace__fprintf_thread_summary(trace, trace->output);
3166
3167                 if (trace->show_tool_stats) {
3168                         fprintf(trace->output, "Stats:\n "
3169                                                " vfs_getname : %" PRIu64 "\n"
3170                                                " proc_getname: %" PRIu64 "\n",
3171                                 trace->stats.vfs_getname,
3172                                 trace->stats.proc_getname);
3173                 }
3174         }
3175
3176 out_delete_evlist:
3177         trace__symbols__exit(trace);
3178
3179         perf_evlist__delete(evlist);
3180         cgroup__put(trace->cgroup);
3181         trace->evlist = NULL;
3182         trace->live = false;
3183         return err;
3184 {
3185         char errbuf[BUFSIZ];
3186
3187 out_error_sched_stat_runtime:
3188         tracing_path__strerror_open_tp(errno, errbuf, sizeof(errbuf), "sched", "sched_stat_runtime");
3189         goto out_error;
3190
3191 out_error_raw_syscalls:
3192         tracing_path__strerror_open_tp(errno, errbuf, sizeof(errbuf), "raw_syscalls", "sys_(enter|exit)");
3193         goto out_error;
3194
3195 out_error_mmap:
3196         perf_evlist__strerror_mmap(evlist, errno, errbuf, sizeof(errbuf));
3197         goto out_error;
3198
3199 out_error_open:
3200         perf_evlist__strerror_open(evlist, errno, errbuf, sizeof(errbuf));
3201
3202 out_error:
3203         fprintf(trace->output, "%s\n", errbuf);
3204         goto out_delete_evlist;
3205
3206 out_error_apply_filters:
3207         fprintf(trace->output,
3208                 "Failed to set filter \"%s\" on event %s with %d (%s)\n",
3209                 evsel->filter, perf_evsel__name(evsel), errno,
3210                 str_error_r(errno, errbuf, sizeof(errbuf)));
3211         goto out_delete_evlist;
3212 }
3213 out_error_mem:
3214         fprintf(trace->output, "Not enough memory to run!\n");
3215         goto out_delete_evlist;
3216
3217 out_errno:
3218         fprintf(trace->output, "errno=%d,%s\n", errno, strerror(errno));
3219         goto out_delete_evlist;
3220 }
3221
3222 static int trace__replay(struct trace *trace)
3223 {
3224         const struct perf_evsel_str_handler handlers[] = {
3225                 { "probe:vfs_getname",       trace__vfs_getname, },
3226         };
3227         struct perf_data data = {
3228                 .path  = input_name,
3229                 .mode  = PERF_DATA_MODE_READ,
3230                 .force = trace->force,
3231         };
3232         struct perf_session *session;
3233         struct perf_evsel *evsel;
3234         int err = -1;
3235
3236         trace->tool.sample        = trace__process_sample;
3237         trace->tool.mmap          = perf_event__process_mmap;
3238         trace->tool.mmap2         = perf_event__process_mmap2;
3239         trace->tool.comm          = perf_event__process_comm;
3240         trace->tool.exit          = perf_event__process_exit;
3241         trace->tool.fork          = perf_event__process_fork;
3242         trace->tool.attr          = perf_event__process_attr;
3243         trace->tool.tracing_data  = perf_event__process_tracing_data;
3244         trace->tool.build_id      = perf_event__process_build_id;
3245         trace->tool.namespaces    = perf_event__process_namespaces;
3246
3247         trace->tool.ordered_events = true;
3248         trace->tool.ordering_requires_timestamps = true;
3249
3250         /* add tid to output */
3251         trace->multiple_threads = true;
3252
3253         session = perf_session__new(&data, false, &trace->tool);
3254         if (session == NULL)
3255                 return -1;
3256
3257         if (trace->opts.target.pid)
3258                 symbol_conf.pid_list_str = strdup(trace->opts.target.pid);
3259
3260         if (trace->opts.target.tid)
3261                 symbol_conf.tid_list_str = strdup(trace->opts.target.tid);
3262
3263         if (symbol__init(&session->header.env) < 0)
3264                 goto out;
3265
3266         trace->host = &session->machines.host;
3267
3268         err = perf_session__set_tracepoints_handlers(session, handlers);
3269         if (err)
3270                 goto out;
3271
3272         evsel = perf_evlist__find_tracepoint_by_name(session->evlist,
3273                                                      "raw_syscalls:sys_enter");
3274         /* older kernels have syscalls tp versus raw_syscalls */
3275         if (evsel == NULL)
3276                 evsel = perf_evlist__find_tracepoint_by_name(session->evlist,
3277                                                              "syscalls:sys_enter");
3278
3279         if (evsel &&
3280             (perf_evsel__init_raw_syscall_tp(evsel, trace__sys_enter) < 0 ||
3281             perf_evsel__init_sc_tp_ptr_field(evsel, args))) {
3282                 pr_err("Error during initialize raw_syscalls:sys_enter event\n");
3283                 goto out;
3284         }
3285
3286         evsel = perf_evlist__find_tracepoint_by_name(session->evlist,
3287                                                      "raw_syscalls:sys_exit");
3288         if (evsel == NULL)
3289                 evsel = perf_evlist__find_tracepoint_by_name(session->evlist,
3290                                                              "syscalls:sys_exit");
3291         if (evsel &&
3292             (perf_evsel__init_raw_syscall_tp(evsel, trace__sys_exit) < 0 ||
3293             perf_evsel__init_sc_tp_uint_field(evsel, ret))) {
3294                 pr_err("Error during initialize raw_syscalls:sys_exit event\n");
3295                 goto out;
3296         }
3297
3298         evlist__for_each_entry(session->evlist, evsel) {
3299                 if (evsel->attr.type == PERF_TYPE_SOFTWARE &&
3300                     (evsel->attr.config == PERF_COUNT_SW_PAGE_FAULTS_MAJ ||
3301                      evsel->attr.config == PERF_COUNT_SW_PAGE_FAULTS_MIN ||
3302                      evsel->attr.config == PERF_COUNT_SW_PAGE_FAULTS))
3303                         evsel->handler = trace__pgfault;
3304         }
3305
3306         setup_pager();
3307
3308         err = perf_session__process_events(session);
3309         if (err)
3310                 pr_err("Failed to process events, error %d", err);
3311
3312         else if (trace->summary)
3313                 trace__fprintf_thread_summary(trace, trace->output);
3314
3315 out:
3316         perf_session__delete(session);
3317
3318         return err;
3319 }
3320
3321 static size_t trace__fprintf_threads_header(FILE *fp)
3322 {
3323         size_t printed;
3324
3325         printed  = fprintf(fp, "\n Summary of events:\n\n");
3326
3327         return printed;
3328 }
3329
3330 DEFINE_RESORT_RB(syscall_stats, a->msecs > b->msecs,
3331         struct stats    *stats;
3332         double          msecs;
3333         int             syscall;
3334 )
3335 {
3336         struct int_node *source = rb_entry(nd, struct int_node, rb_node);
3337         struct stats *stats = source->priv;
3338
3339         entry->syscall = source->i;
3340         entry->stats   = stats;
3341         entry->msecs   = stats ? (u64)stats->n * (avg_stats(stats) / NSEC_PER_MSEC) : 0;
3342 }
3343
3344 static size_t thread__dump_stats(struct thread_trace *ttrace,
3345                                  struct trace *trace, FILE *fp)
3346 {
3347         size_t printed = 0;
3348         struct syscall *sc;
3349         struct rb_node *nd;
3350         DECLARE_RESORT_RB_INTLIST(syscall_stats, ttrace->syscall_stats);
3351
3352         if (syscall_stats == NULL)
3353                 return 0;
3354
3355         printed += fprintf(fp, "\n");
3356
3357         printed += fprintf(fp, "   syscall            calls    total       min       avg       max      stddev\n");
3358         printed += fprintf(fp, "                               (msec)    (msec)    (msec)    (msec)        (%%)\n");
3359         printed += fprintf(fp, "   --------------- -------- --------- --------- --------- ---------     ------\n");
3360
3361         resort_rb__for_each_entry(nd, syscall_stats) {
3362                 struct stats *stats = syscall_stats_entry->stats;
3363                 if (stats) {
3364                         double min = (double)(stats->min) / NSEC_PER_MSEC;
3365                         double max = (double)(stats->max) / NSEC_PER_MSEC;
3366                         double avg = avg_stats(stats);
3367                         double pct;
3368                         u64 n = (u64) stats->n;
3369
3370                         pct = avg ? 100.0 * stddev_stats(stats)/avg : 0.0;
3371                         avg /= NSEC_PER_MSEC;
3372
3373                         sc = &trace->syscalls.table[syscall_stats_entry->syscall];
3374                         printed += fprintf(fp, "   %-15s", sc->name);
3375                         printed += fprintf(fp, " %8" PRIu64 " %9.3f %9.3f %9.3f",
3376                                            n, syscall_stats_entry->msecs, min, avg);
3377                         printed += fprintf(fp, " %9.3f %9.2f%%\n", max, pct);
3378                 }
3379         }
3380
3381         resort_rb__delete(syscall_stats);
3382         printed += fprintf(fp, "\n\n");
3383
3384         return printed;
3385 }
3386
3387 static size_t trace__fprintf_thread(FILE *fp, struct thread *thread, struct trace *trace)
3388 {
3389         size_t printed = 0;
3390         struct thread_trace *ttrace = thread__priv(thread);
3391         double ratio;
3392
3393         if (ttrace == NULL)
3394                 return 0;
3395
3396         ratio = (double)ttrace->nr_events / trace->nr_events * 100.0;
3397
3398         printed += fprintf(fp, " %s (%d), ", thread__comm_str(thread), thread->tid);
3399         printed += fprintf(fp, "%lu events, ", ttrace->nr_events);
3400         printed += fprintf(fp, "%.1f%%", ratio);
3401         if (ttrace->pfmaj)
3402                 printed += fprintf(fp, ", %lu majfaults", ttrace->pfmaj);
3403         if (ttrace->pfmin)
3404                 printed += fprintf(fp, ", %lu minfaults", ttrace->pfmin);
3405         if (trace->sched)
3406                 printed += fprintf(fp, ", %.3f msec\n", ttrace->runtime_ms);
3407         else if (fputc('\n', fp) != EOF)
3408                 ++printed;
3409
3410         printed += thread__dump_stats(ttrace, trace, fp);
3411
3412         return printed;
3413 }
3414
3415 static unsigned long thread__nr_events(struct thread_trace *ttrace)
3416 {
3417         return ttrace ? ttrace->nr_events : 0;
3418 }
3419
3420 DEFINE_RESORT_RB(threads, (thread__nr_events(a->thread->priv) < thread__nr_events(b->thread->priv)),
3421         struct thread *thread;
3422 )
3423 {
3424         entry->thread = rb_entry(nd, struct thread, rb_node);
3425 }
3426
3427 static size_t trace__fprintf_thread_summary(struct trace *trace, FILE *fp)
3428 {
3429         size_t printed = trace__fprintf_threads_header(fp);
3430         struct rb_node *nd;
3431         int i;
3432
3433         for (i = 0; i < THREADS__TABLE_SIZE; i++) {
3434                 DECLARE_RESORT_RB_MACHINE_THREADS(threads, trace->host, i);
3435
3436                 if (threads == NULL) {
3437                         fprintf(fp, "%s", "Error sorting output by nr_events!\n");
3438                         return 0;
3439                 }
3440
3441                 resort_rb__for_each_entry(nd, threads)
3442                         printed += trace__fprintf_thread(fp, threads_entry->thread, trace);
3443
3444                 resort_rb__delete(threads);
3445         }
3446         return printed;
3447 }
3448
3449 static int trace__set_duration(const struct option *opt, const char *str,
3450                                int unset __maybe_unused)
3451 {
3452         struct trace *trace = opt->value;
3453
3454         trace->duration_filter = atof(str);
3455         return 0;
3456 }
3457
3458 static int trace__set_filter_pids_from_option(const struct option *opt, const char *str,
3459                                               int unset __maybe_unused)
3460 {
3461         int ret = -1;
3462         size_t i;
3463         struct trace *trace = opt->value;
3464         /*
3465          * FIXME: introduce a intarray class, plain parse csv and create a
3466          * { int nr, int entries[] } struct...
3467          */
3468         struct intlist *list = intlist__new(str);
3469
3470         if (list == NULL)
3471                 return -1;
3472
3473         i = trace->filter_pids.nr = intlist__nr_entries(list) + 1;
3474         trace->filter_pids.entries = calloc(i, sizeof(pid_t));
3475
3476         if (trace->filter_pids.entries == NULL)
3477                 goto out;
3478
3479         trace->filter_pids.entries[0] = getpid();
3480
3481         for (i = 1; i < trace->filter_pids.nr; ++i)
3482                 trace->filter_pids.entries[i] = intlist__entry(list, i - 1)->i;
3483
3484         intlist__delete(list);
3485         ret = 0;
3486 out:
3487         return ret;
3488 }
3489
3490 static int trace__open_output(struct trace *trace, const char *filename)
3491 {
3492         struct stat st;
3493
3494         if (!stat(filename, &st) && st.st_size) {
3495                 char oldname[PATH_MAX];
3496
3497                 scnprintf(oldname, sizeof(oldname), "%s.old", filename);
3498                 unlink(oldname);
3499                 rename(filename, oldname);
3500         }
3501
3502         trace->output = fopen(filename, "w");
3503
3504         return trace->output == NULL ? -errno : 0;
3505 }
3506
3507 static int parse_pagefaults(const struct option *opt, const char *str,
3508                             int unset __maybe_unused)
3509 {
3510         int *trace_pgfaults = opt->value;
3511
3512         if (strcmp(str, "all") == 0)
3513                 *trace_pgfaults |= TRACE_PFMAJ | TRACE_PFMIN;
3514         else if (strcmp(str, "maj") == 0)
3515                 *trace_pgfaults |= TRACE_PFMAJ;
3516         else if (strcmp(str, "min") == 0)
3517                 *trace_pgfaults |= TRACE_PFMIN;
3518         else
3519                 return -1;
3520
3521         return 0;
3522 }
3523
3524 static void evlist__set_evsel_handler(struct perf_evlist *evlist, void *handler)
3525 {
3526         struct perf_evsel *evsel;
3527
3528         evlist__for_each_entry(evlist, evsel)
3529                 evsel->handler = handler;
3530 }
3531
3532 static int evlist__set_syscall_tp_fields(struct perf_evlist *evlist)
3533 {
3534         struct perf_evsel *evsel;
3535
3536         evlist__for_each_entry(evlist, evsel) {
3537                 if (evsel->priv || !evsel->tp_format)
3538                         continue;
3539
3540                 if (strcmp(evsel->tp_format->system, "syscalls"))
3541                         continue;
3542
3543                 if (perf_evsel__init_syscall_tp(evsel))
3544                         return -1;
3545
3546                 if (!strncmp(evsel->tp_format->name, "sys_enter_", 10)) {
3547                         struct syscall_tp *sc = evsel->priv;
3548
3549                         if (__tp_field__init_ptr(&sc->args, sc->id.offset + sizeof(u64)))
3550                                 return -1;
3551                 } else if (!strncmp(evsel->tp_format->name, "sys_exit_", 9)) {
3552                         struct syscall_tp *sc = evsel->priv;
3553
3554                         if (__tp_field__init_uint(&sc->ret, sizeof(u64), sc->id.offset + sizeof(u64), evsel->needs_swap))
3555                                 return -1;
3556                 }
3557         }
3558
3559         return 0;
3560 }
3561
3562 /*
3563  * XXX: Hackish, just splitting the combined -e+--event (syscalls
3564  * (raw_syscalls:{sys_{enter,exit}} + events (tracepoints, HW, SW, etc) to use
3565  * existing facilities unchanged (trace->ev_qualifier + parse_options()).
3566  *
3567  * It'd be better to introduce a parse_options() variant that would return a
3568  * list with the terms it didn't match to an event...
3569  */
3570 static int trace__parse_events_option(const struct option *opt, const char *str,
3571                                       int unset __maybe_unused)
3572 {
3573         struct trace *trace = (struct trace *)opt->value;
3574         const char *s = str;
3575         char *sep = NULL, *lists[2] = { NULL, NULL, };
3576         int len = strlen(str) + 1, err = -1, list, idx;
3577         char *strace_groups_dir = system_path(STRACE_GROUPS_DIR);
3578         char group_name[PATH_MAX];
3579         struct syscall_fmt *fmt;
3580
3581         if (strace_groups_dir == NULL)
3582                 return -1;
3583
3584         if (*s == '!') {
3585                 ++s;
3586                 trace->not_ev_qualifier = true;
3587         }
3588
3589         while (1) {
3590                 if ((sep = strchr(s, ',')) != NULL)
3591                         *sep = '\0';
3592
3593                 list = 0;
3594                 if (syscalltbl__id(trace->sctbl, s) >= 0 ||
3595                     syscalltbl__strglobmatch_first(trace->sctbl, s, &idx) >= 0) {
3596                         list = 1;
3597                         goto do_concat;
3598                 }
3599
3600                 fmt = syscall_fmt__find_by_alias(s);
3601                 if (fmt != NULL) {
3602                         list = 1;
3603                         s = fmt->name;
3604                 } else {
3605                         path__join(group_name, sizeof(group_name), strace_groups_dir, s);
3606                         if (access(group_name, R_OK) == 0)
3607                                 list = 1;
3608                 }
3609 do_concat:
3610                 if (lists[list]) {
3611                         sprintf(lists[list] + strlen(lists[list]), ",%s", s);
3612                 } else {
3613                         lists[list] = malloc(len);
3614                         if (lists[list] == NULL)
3615                                 goto out;
3616                         strcpy(lists[list], s);
3617                 }
3618
3619                 if (!sep)
3620                         break;
3621
3622                 *sep = ',';
3623                 s = sep + 1;
3624         }
3625
3626         if (lists[1] != NULL) {
3627                 struct strlist_config slist_config = {
3628                         .dirname = strace_groups_dir,
3629                 };
3630
3631                 trace->ev_qualifier = strlist__new(lists[1], &slist_config);
3632                 if (trace->ev_qualifier == NULL) {
3633                         fputs("Not enough memory to parse event qualifier", trace->output);
3634                         goto out;
3635                 }
3636
3637                 if (trace__validate_ev_qualifier(trace))
3638                         goto out;
3639                 trace->trace_syscalls = true;
3640         }
3641
3642         err = 0;
3643
3644         if (lists[0]) {
3645                 struct option o = OPT_CALLBACK('e', "event", &trace->evlist, "event",
3646                                                "event selector. use 'perf list' to list available events",
3647                                                parse_events_option);
3648                 err = parse_events_option(&o, lists[0], 0);
3649         }
3650 out:
3651         if (sep)
3652                 *sep = ',';
3653
3654         return err;
3655 }
3656
3657 static int trace__parse_cgroups(const struct option *opt, const char *str, int unset)
3658 {
3659         struct trace *trace = opt->value;
3660
3661         if (!list_empty(&trace->evlist->entries))
3662                 return parse_cgroups(opt, str, unset);
3663
3664         trace->cgroup = evlist__findnew_cgroup(trace->evlist, str);
3665
3666         return 0;
3667 }
3668
3669 static struct bpf_map *bpf__find_map_by_name(const char *name)
3670 {
3671         struct bpf_object *obj, *tmp;
3672
3673         bpf_object__for_each_safe(obj, tmp) {
3674                 struct bpf_map *map = bpf_object__find_map_by_name(obj, name);
3675                 if (map)
3676                         return map;
3677
3678         }
3679
3680         return NULL;
3681 }
3682
3683 static void trace__set_bpf_map_filtered_pids(struct trace *trace)
3684 {
3685         trace->filter_pids.map = bpf__find_map_by_name("pids_filtered");
3686 }
3687
3688 static void trace__set_bpf_map_syscalls(struct trace *trace)
3689 {
3690         trace->syscalls.map = bpf__find_map_by_name("syscalls");
3691 }
3692
3693 static int trace__config(const char *var, const char *value, void *arg)
3694 {
3695         struct trace *trace = arg;
3696         int err = 0;
3697
3698         if (!strcmp(var, "trace.add_events")) {
3699                 struct option o = OPT_CALLBACK('e', "event", &trace->evlist, "event",
3700                                                "event selector. use 'perf list' to list available events",
3701                                                parse_events_option);
3702                 /*
3703                  * We can't propagate parse_event_option() return, as it is 1
3704                  * for failure while perf_config() expects -1.
3705                  */
3706                 if (parse_events_option(&o, value, 0))
3707                         err = -1;
3708         } else if (!strcmp(var, "trace.show_timestamp")) {
3709                 trace->show_tstamp = perf_config_bool(var, value);
3710         } else if (!strcmp(var, "trace.show_duration")) {
3711                 trace->show_duration = perf_config_bool(var, value);
3712         } else if (!strcmp(var, "trace.show_arg_names")) {
3713                 trace->show_arg_names = perf_config_bool(var, value);
3714                 if (!trace->show_arg_names)
3715                         trace->show_zeros = true;
3716         } else if (!strcmp(var, "trace.show_zeros")) {
3717                 bool new_show_zeros = perf_config_bool(var, value);
3718                 if (!trace->show_arg_names && !new_show_zeros) {
3719                         pr_warning("trace.show_zeros has to be set when trace.show_arg_names=no\n");
3720                         goto out;
3721                 }
3722                 trace->show_zeros = new_show_zeros;
3723         } else if (!strcmp(var, "trace.show_prefix")) {
3724                 trace->show_string_prefix = perf_config_bool(var, value);
3725         } else if (!strcmp(var, "trace.no_inherit")) {
3726                 trace->opts.no_inherit = perf_config_bool(var, value);
3727         } else if (!strcmp(var, "trace.args_alignment")) {
3728                 int args_alignment = 0;
3729                 if (perf_config_int(&args_alignment, var, value) == 0)
3730                         trace->args_alignment = args_alignment;
3731         }
3732 out:
3733         return err;
3734 }
3735
3736 int cmd_trace(int argc, const char **argv)
3737 {
3738         const char *trace_usage[] = {
3739                 "perf trace [<options>] [<command>]",
3740                 "perf trace [<options>] -- <command> [<options>]",
3741                 "perf trace record [<options>] [<command>]",
3742                 "perf trace record [<options>] -- <command> [<options>]",
3743                 NULL
3744         };
3745         struct trace trace = {
3746                 .syscalls = {
3747                         . max = -1,
3748                 },
3749                 .opts = {
3750                         .target = {
3751                                 .uid       = UINT_MAX,
3752                                 .uses_mmap = true,
3753                         },
3754                         .user_freq     = UINT_MAX,
3755                         .user_interval = ULLONG_MAX,
3756                         .no_buffering  = true,
3757                         .mmap_pages    = UINT_MAX,
3758                 },
3759                 .output = stderr,
3760                 .show_comm = true,
3761                 .show_tstamp = true,
3762                 .show_duration = true,
3763                 .show_arg_names = true,
3764                 .args_alignment = 70,
3765                 .trace_syscalls = false,
3766                 .kernel_syscallchains = false,
3767                 .max_stack = UINT_MAX,
3768                 .max_events = ULONG_MAX,
3769         };
3770         const char *map_dump_str = NULL;
3771         const char *output_name = NULL;
3772         const struct option trace_options[] = {
3773         OPT_CALLBACK('e', "event", &trace, "event",
3774                      "event/syscall selector. use 'perf list' to list available events",
3775                      trace__parse_events_option),
3776         OPT_BOOLEAN(0, "comm", &trace.show_comm,
3777                     "show the thread COMM next to its id"),
3778         OPT_BOOLEAN(0, "tool_stats", &trace.show_tool_stats, "show tool stats"),
3779         OPT_CALLBACK(0, "expr", &trace, "expr", "list of syscalls/events to trace",
3780                      trace__parse_events_option),
3781         OPT_STRING('o', "output", &output_name, "file", "output file name"),
3782         OPT_STRING('i', "input", &input_name, "file", "Analyze events in file"),
3783         OPT_STRING('p', "pid", &trace.opts.target.pid, "pid",
3784                     "trace events on existing process id"),
3785         OPT_STRING('t', "tid", &trace.opts.target.tid, "tid",
3786                     "trace events on existing thread id"),
3787         OPT_CALLBACK(0, "filter-pids", &trace, "CSV list of pids",
3788                      "pids to filter (by the kernel)", trace__set_filter_pids_from_option),
3789         OPT_BOOLEAN('a', "all-cpus", &trace.opts.target.system_wide,
3790                     "system-wide collection from all CPUs"),
3791         OPT_STRING('C', "cpu", &trace.opts.target.cpu_list, "cpu",
3792                     "list of cpus to monitor"),
3793         OPT_BOOLEAN(0, "no-inherit", &trace.opts.no_inherit,
3794                     "child tasks do not inherit counters"),
3795         OPT_CALLBACK('m', "mmap-pages", &trace.opts.mmap_pages, "pages",
3796                      "number of mmap data pages",
3797                      perf_evlist__parse_mmap_pages),
3798         OPT_STRING('u', "uid", &trace.opts.target.uid_str, "user",
3799                    "user to profile"),
3800         OPT_CALLBACK(0, "duration", &trace, "float",
3801                      "show only events with duration > N.M ms",
3802                      trace__set_duration),
3803 #ifdef HAVE_LIBBPF_SUPPORT
3804         OPT_STRING(0, "map-dump", &map_dump_str, "BPF map", "BPF map to periodically dump"),
3805 #endif
3806         OPT_BOOLEAN(0, "sched", &trace.sched, "show blocking scheduler events"),
3807         OPT_INCR('v', "verbose", &verbose, "be more verbose"),
3808         OPT_BOOLEAN('T', "time", &trace.full_time,
3809                     "Show full timestamp, not time relative to first start"),
3810         OPT_BOOLEAN(0, "failure", &trace.failure_only,
3811                     "Show only syscalls that failed"),
3812         OPT_BOOLEAN('s', "summary", &trace.summary_only,
3813                     "Show only syscall summary with statistics"),
3814         OPT_BOOLEAN('S', "with-summary", &trace.summary,
3815                     "Show all syscalls and summary with statistics"),
3816         OPT_CALLBACK_DEFAULT('F', "pf", &trace.trace_pgfaults, "all|maj|min",
3817                      "Trace pagefaults", parse_pagefaults, "maj"),
3818         OPT_BOOLEAN(0, "syscalls", &trace.trace_syscalls, "Trace syscalls"),
3819         OPT_BOOLEAN('f', "force", &trace.force, "don't complain, do it"),
3820         OPT_CALLBACK(0, "call-graph", &trace.opts,
3821                      "record_mode[,record_size]", record_callchain_help,
3822                      &record_parse_callchain_opt),
3823         OPT_BOOLEAN(0, "kernel-syscall-graph", &trace.kernel_syscallchains,
3824                     "Show the kernel callchains on the syscall exit path"),
3825         OPT_ULONG(0, "max-events", &trace.max_events,
3826                 "Set the maximum number of events to print, exit after that is reached. "),
3827         OPT_UINTEGER(0, "min-stack", &trace.min_stack,
3828                      "Set the minimum stack depth when parsing the callchain, "
3829                      "anything below the specified depth will be ignored."),
3830         OPT_UINTEGER(0, "max-stack", &trace.max_stack,
3831                      "Set the maximum stack depth when parsing the callchain, "
3832                      "anything beyond the specified depth will be ignored. "
3833                      "Default: kernel.perf_event_max_stack or " __stringify(PERF_MAX_STACK_DEPTH)),
3834         OPT_BOOLEAN(0, "sort-events", &trace.sort_events,
3835                         "Sort batch of events before processing, use if getting out of order events"),
3836         OPT_BOOLEAN(0, "print-sample", &trace.print_sample,
3837                         "print the PERF_RECORD_SAMPLE PERF_SAMPLE_ info, for debugging"),
3838         OPT_UINTEGER(0, "proc-map-timeout", &proc_map_timeout,
3839                         "per thread proc mmap processing timeout in ms"),
3840         OPT_CALLBACK('G', "cgroup", &trace, "name", "monitor event in cgroup name only",
3841                      trace__parse_cgroups),
3842         OPT_UINTEGER('D', "delay", &trace.opts.initial_delay,
3843                      "ms to wait before starting measurement after program "
3844                      "start"),
3845         OPT_END()
3846         };
3847         bool __maybe_unused max_stack_user_set = true;
3848         bool mmap_pages_user_set = true;
3849         struct perf_evsel *evsel;
3850         const char * const trace_subcommands[] = { "record", NULL };
3851         int err = -1;
3852         char bf[BUFSIZ];
3853
3854         signal(SIGSEGV, sighandler_dump_stack);
3855         signal(SIGFPE, sighandler_dump_stack);
3856
3857         trace.evlist = perf_evlist__new();
3858         trace.sctbl = syscalltbl__new();
3859
3860         if (trace.evlist == NULL || trace.sctbl == NULL) {
3861                 pr_err("Not enough memory to run!\n");
3862                 err = -ENOMEM;
3863                 goto out;
3864         }
3865
3866         err = perf_config(trace__config, &trace);
3867         if (err)
3868                 goto out;
3869
3870         argc = parse_options_subcommand(argc, argv, trace_options, trace_subcommands,
3871                                  trace_usage, PARSE_OPT_STOP_AT_NON_OPTION);
3872
3873         if ((nr_cgroups || trace.cgroup) && !trace.opts.target.system_wide) {
3874                 usage_with_options_msg(trace_usage, trace_options,
3875                                        "cgroup monitoring only available in system-wide mode");
3876         }
3877
3878         evsel = bpf__setup_output_event(trace.evlist, "__augmented_syscalls__");
3879         if (IS_ERR(evsel)) {
3880                 bpf__strerror_setup_output_event(trace.evlist, PTR_ERR(evsel), bf, sizeof(bf));
3881                 pr_err("ERROR: Setup trace syscalls enter failed: %s\n", bf);
3882                 goto out;
3883         }
3884
3885         if (evsel) {
3886                 trace.syscalls.events.augmented = evsel;
3887                 trace__set_bpf_map_filtered_pids(&trace);
3888                 trace__set_bpf_map_syscalls(&trace);
3889         }
3890
3891         err = bpf__setup_stdout(trace.evlist);
3892         if (err) {
3893                 bpf__strerror_setup_stdout(trace.evlist, err, bf, sizeof(bf));
3894                 pr_err("ERROR: Setup BPF stdout failed: %s\n", bf);
3895                 goto out;
3896         }
3897
3898         err = -1;
3899
3900         if (map_dump_str) {
3901                 trace.dump.map = bpf__find_map_by_name(map_dump_str);
3902                 if (trace.dump.map == NULL) {
3903                         pr_err("ERROR: BPF map \"%s\" not found\n", map_dump_str);
3904                         goto out;
3905                 }
3906         }
3907
3908         if (trace.trace_pgfaults) {
3909                 trace.opts.sample_address = true;
3910                 trace.opts.sample_time = true;
3911         }
3912
3913         if (trace.opts.mmap_pages == UINT_MAX)
3914                 mmap_pages_user_set = false;
3915
3916         if (trace.max_stack == UINT_MAX) {
3917                 trace.max_stack = input_name ? PERF_MAX_STACK_DEPTH : sysctl__max_stack();
3918                 max_stack_user_set = false;
3919         }
3920
3921 #ifdef HAVE_DWARF_UNWIND_SUPPORT
3922         if ((trace.min_stack || max_stack_user_set) && !callchain_param.enabled) {
3923                 record_opts__parse_callchain(&trace.opts, &callchain_param, "dwarf", false);
3924         }
3925 #endif
3926
3927         if (callchain_param.enabled) {
3928                 if (!mmap_pages_user_set && geteuid() == 0)
3929                         trace.opts.mmap_pages = perf_event_mlock_kb_in_pages() * 4;
3930
3931                 symbol_conf.use_callchain = true;
3932         }
3933
3934         if (trace.evlist->nr_entries > 0) {
3935                 evlist__set_evsel_handler(trace.evlist, trace__event_handler);
3936                 if (evlist__set_syscall_tp_fields(trace.evlist)) {
3937                         perror("failed to set syscalls:* tracepoint fields");
3938                         goto out;
3939                 }
3940         }
3941
3942         if (trace.sort_events) {
3943                 ordered_events__init(&trace.oe.data, ordered_events__deliver_event, &trace);
3944                 ordered_events__set_copy_on_queue(&trace.oe.data, true);
3945         }
3946
3947         /*
3948          * If we are augmenting syscalls, then combine what we put in the
3949          * __augmented_syscalls__ BPF map with what is in the
3950          * syscalls:sys_exit_FOO tracepoints, i.e. just like we do without BPF,
3951          * combining raw_syscalls:sys_enter with raw_syscalls:sys_exit.
3952          *
3953          * We'll switch to look at two BPF maps, one for sys_enter and the
3954          * other for sys_exit when we start augmenting the sys_exit paths with
3955          * buffers that are being copied from kernel to userspace, think 'read'
3956          * syscall.
3957          */
3958         if (trace.syscalls.events.augmented) {
3959                 evlist__for_each_entry(trace.evlist, evsel) {
3960                         bool raw_syscalls_sys_exit = strcmp(perf_evsel__name(evsel), "raw_syscalls:sys_exit") == 0;
3961
3962                         if (raw_syscalls_sys_exit) {
3963                                 trace.raw_augmented_syscalls = true;
3964                                 goto init_augmented_syscall_tp;
3965                         }
3966
3967                         if (trace.syscalls.events.augmented->priv == NULL &&
3968                             strstr(perf_evsel__name(evsel), "syscalls:sys_enter")) {
3969                                 struct perf_evsel *augmented = trace.syscalls.events.augmented;
3970                                 if (perf_evsel__init_augmented_syscall_tp(augmented, evsel) ||
3971                                     perf_evsel__init_augmented_syscall_tp_args(augmented))
3972                                         goto out;
3973                                 augmented->handler = trace__sys_enter;
3974                         }
3975
3976                         if (strstarts(perf_evsel__name(evsel), "syscalls:sys_exit_")) {
3977                                 struct syscall_tp *sc;
3978 init_augmented_syscall_tp:
3979                                 if (perf_evsel__init_augmented_syscall_tp(evsel, evsel))
3980                                         goto out;
3981                                 sc = evsel->priv;
3982                                 /*
3983                                  * For now with BPF raw_augmented we hook into
3984                                  * raw_syscalls:sys_enter and there we get all
3985                                  * 6 syscall args plus the tracepoint common
3986                                  * fields and the syscall_nr (another long).
3987                                  * So we check if that is the case and if so
3988                                  * don't look after the sc->args_size but
3989                                  * always after the full raw_syscalls:sys_enter
3990                                  * payload, which is fixed.
3991                                  *
3992                                  * We'll revisit this later to pass
3993                                  * s->args_size to the BPF augmenter (now
3994                                  * tools/perf/examples/bpf/augmented_raw_syscalls.c,
3995                                  * so that it copies only what we need for each
3996                                  * syscall, like what happens when we use
3997                                  * syscalls:sys_enter_NAME, so that we reduce
3998                                  * the kernel/userspace traffic to just what is
3999                                  * needed for each syscall.
4000                                  */
4001                                 if (trace.raw_augmented_syscalls)
4002                                         trace.raw_augmented_syscalls_args_size = (6 + 1) * sizeof(long) + sc->id.offset;
4003                                 perf_evsel__init_augmented_syscall_tp_ret(evsel);
4004                                 evsel->handler = trace__sys_exit;
4005                         }
4006                 }
4007         }
4008
4009         if ((argc >= 1) && (strcmp(argv[0], "record") == 0))
4010                 return trace__record(&trace, argc-1, &argv[1]);
4011
4012         /* summary_only implies summary option, but don't overwrite summary if set */
4013         if (trace.summary_only)
4014                 trace.summary = trace.summary_only;
4015
4016         if (!trace.trace_syscalls && !trace.trace_pgfaults &&
4017             trace.evlist->nr_entries == 0 /* Was --events used? */) {
4018                 trace.trace_syscalls = true;
4019         }
4020
4021         if (output_name != NULL) {
4022                 err = trace__open_output(&trace, output_name);
4023                 if (err < 0) {
4024                         perror("failed to create output file");
4025                         goto out;
4026                 }
4027         }
4028
4029         err = target__validate(&trace.opts.target);
4030         if (err) {
4031                 target__strerror(&trace.opts.target, err, bf, sizeof(bf));
4032                 fprintf(trace.output, "%s", bf);
4033                 goto out_close;
4034         }
4035
4036         err = target__parse_uid(&trace.opts.target);
4037         if (err) {
4038                 target__strerror(&trace.opts.target, err, bf, sizeof(bf));
4039                 fprintf(trace.output, "%s", bf);
4040                 goto out_close;
4041         }
4042
4043         if (!argc && target__none(&trace.opts.target))
4044                 trace.opts.target.system_wide = true;
4045
4046         if (input_name)
4047                 err = trace__replay(&trace);
4048         else
4049                 err = trace__run(&trace, argc, argv);
4050
4051 out_close:
4052         if (output_name != NULL)
4053                 fclose(trace.output);
4054 out:
4055         return err;
4056 }