tools/perf/builtin-trace.c

   1 /*
   2  * builtin-trace.c
   3  *
   4  * Builtin 'trace' command:
   5  *
   6  * Display a continuously updated trace of any workload, CPU, specific PID,
   7  * system wide, etc.  Default format is loosely strace like, but any other
   8  * event may be specified using --event.
   9  *
  10  * Copyright (C) 2012, 2013, 2014, 2015 Red Hat Inc, Arnaldo Carvalho de Melo <acme@redhat.com>
  11  *
  12  * Initially based on the 'trace' prototype by Thomas Gleixner:
  13  *
  14  * http://lwn.net/Articles/415728/ ("Announcing a new utility: 'trace'")
  15  */
  16
  17 #include "util/record.h"
  18 #include <traceevent/event-parse.h>
  19 #include <api/fs/tracing_path.h>
  20 #include <bpf/bpf.h>
  21 #include "util/bpf_map.h"
  22 #include "util/rlimit.h"
  23 #include "builtin.h"
  24 #include "util/cgroup.h"
  25 #include "util/color.h"
  26 #include "util/config.h"
  27 #include "util/debug.h"
  28 #include "util/dso.h"
  29 #include "util/env.h"
  30 #include "util/event.h"
  31 #include "util/evsel.h"
  32 #include "util/evsel_fprintf.h"
  33 #include "util/synthetic-events.h"
  34 #include "util/evlist.h"
  35 #include "util/evswitch.h"
  36 #include "util/mmap.h"
  37 #include <subcmd/pager.h>
  38 #include <subcmd/exec-cmd.h>
  39 #include "util/machine.h"
  40 #include "util/map.h"
  41 #include "util/symbol.h"
  42 #include "util/path.h"
  43 #include "util/session.h"
  44 #include "util/thread.h"
  45 #include <subcmd/parse-options.h>
  46 #include "util/strlist.h"
  47 #include "util/intlist.h"
  48 #include "util/thread_map.h"
  49 #include "util/stat.h"
  50 #include "util/tool.h"
  51 #include "util/util.h"
  52 #include "trace/beauty/beauty.h"
  53 #include "trace-event.h"
  54 #include "util/parse-events.h"
  55 #include "util/bpf-loader.h"
  56 #include "callchain.h"
  57 #include "print_binary.h"
  58 #include "string2.h"
  59 #include "syscalltbl.h"
  60 #include "rb_resort.h"
  61 #include "../perf.h"
  62
  63 #include <errno.h>
  64 #include <inttypes.h>
  65 #include <poll.h>
  66 #include <signal.h>
  67 #include <stdlib.h>
  68 #include <string.h>
  69 #include <linux/err.h>
  70 #include <linux/filter.h>
  71 #include <linux/kernel.h>
  72 #include <linux/random.h>
  73 #include <linux/stringify.h>
  74 #include <linux/time64.h>
  75 #include <linux/zalloc.h>
  76 #include <fcntl.h>
  77 #include <sys/sysmacros.h>
  78
  79 #include <linux/ctype.h>
  80
  81 #ifndef O_CLOEXEC
  82 # define O_CLOEXEC              02000000
  83 #endif
  84
  85 #ifndef F_LINUX_SPECIFIC_BASE
  86 # define F_LINUX_SPECIFIC_BASE  1024
  87 #endif
  88
  89 struct trace {
  90         struct perf_tool        tool;
  91         struct syscalltbl       *sctbl;
  92         struct {
  93                 struct syscall  *table;
  94                 struct bpf_map  *map;
  95                 struct { // per syscall BPF_MAP_TYPE_PROG_ARRAY
  96                         struct bpf_map  *sys_enter,
  97                                         *sys_exit;
  98                 }               prog_array;
  99                 struct {
 100                         struct evsel *sys_enter,
 101                                           *sys_exit,
 102                                           *augmented;
 103                 }               events;
 104                 struct bpf_program *unaugmented_prog;
 105         } syscalls;
 106         struct {
 107                 struct bpf_map *map;
 108         } dump;
 109         struct record_opts      opts;
 110         struct evlist   *evlist;
 111         struct machine          *host;
 112         struct thread           *current;
 113         struct bpf_object       *bpf_obj;
 114         struct cgroup           *cgroup;
 115         u64                     base_time;
 116         FILE                    *output;
 117         unsigned long           nr_events;
 118         unsigned long           nr_events_printed;
 119         unsigned long           max_events;
 120         struct evswitch         evswitch;
 121         struct strlist          *ev_qualifier;
 122         struct {
 123                 size_t          nr;
 124                 int             *entries;
 125         }                       ev_qualifier_ids;
 126         struct {
 127                 size_t          nr;
 128                 pid_t           *entries;
 129                 struct bpf_map  *map;
 130         }                       filter_pids;
 131         double                  duration_filter;
 132         double                  runtime_ms;
 133         struct {
 134                 u64             vfs_getname,
 135                                 proc_getname;
 136         } stats;
 137         unsigned int            max_stack;
 138         unsigned int            min_stack;
 139         int                     raw_augmented_syscalls_args_size;
 140         bool                    raw_augmented_syscalls;
 141         bool                    fd_path_disabled;
 142         bool                    sort_events;
 143         bool                    not_ev_qualifier;
 144         bool                    live;
 145         bool                    full_time;
 146         bool                    sched;
 147         bool                    multiple_threads;
 148         bool                    summary;
 149         bool                    summary_only;
 150         bool                    failure_only;
 151         bool                    show_comm;
 152         bool                    print_sample;
 153         bool                    show_tool_stats;
 154         bool                    trace_syscalls;
 155         bool                    kernel_syscallchains;
 156         s16                     args_alignment;
 157         bool                    show_tstamp;
 158         bool                    show_duration;
 159         bool                    show_zeros;
 160         bool                    show_arg_names;
 161         bool                    show_string_prefix;
 162         bool                    force;
 163         bool                    vfs_getname;
 164         int                     trace_pgfaults;
 165         struct {
 166                 struct ordered_events   data;
 167                 u64                     last;
 168         } oe;
 169 };
 170
 171 struct tp_field {
 172         int offset;
 173         union {
 174                 u64 (*integer)(struct tp_field *field, struct perf_sample *sample);
 175                 void *(*pointer)(struct tp_field *field, struct perf_sample *sample);
 176         };
 177 };
 178
 179 #define TP_UINT_FIELD(bits) \
 180 static u64 tp_field__u##bits(struct tp_field *field, struct perf_sample *sample) \
 181 { \
 182         u##bits value; \
 183         memcpy(&value, sample->raw_data + field->offset, sizeof(value)); \
 184         return value;  \
 185 }
 186
 187 TP_UINT_FIELD(8);
 188 TP_UINT_FIELD(16);
 189 TP_UINT_FIELD(32);
 190 TP_UINT_FIELD(64);
 191
 192 #define TP_UINT_FIELD__SWAPPED(bits) \
 193 static u64 tp_field__swapped_u##bits(struct tp_field *field, struct perf_sample *sample) \
 194 { \
 195         u##bits value; \
 196         memcpy(&value, sample->raw_data + field->offset, sizeof(value)); \
 197         return bswap_##bits(value);\
 198 }
 199
 200 TP_UINT_FIELD__SWAPPED(16);
 201 TP_UINT_FIELD__SWAPPED(32);
 202 TP_UINT_FIELD__SWAPPED(64);
 203
 204 static int __tp_field__init_uint(struct tp_field *field, int size, int offset, bool needs_swap)
 205 {
 206         field->offset = offset;
 207
 208         switch (size) {
 209         case 1:
 210                 field->integer = tp_field__u8;
 211                 break;
 212         case 2:
 213                 field->integer = needs_swap ? tp_field__swapped_u16 : tp_field__u16;
 214                 break;
 215         case 4:
 216                 field->integer = needs_swap ? tp_field__swapped_u32 : tp_field__u32;
 217                 break;
 218         case 8:
 219                 field->integer = needs_swap ? tp_field__swapped_u64 : tp_field__u64;
 220                 break;
 221         default:
 222                 return -1;
 223         }
 224
 225         return 0;
 226 }
 227
 228 static int tp_field__init_uint(struct tp_field *field, struct tep_format_field *format_field, bool needs_swap)
 229 {
 230         return __tp_field__init_uint(field, format_field->size, format_field->offset, needs_swap);
 231 }
 232
 233 static void *tp_field__ptr(struct tp_field *field, struct perf_sample *sample)
 234 {
 235         return sample->raw_data + field->offset;
 236 }
 237
 238 static int __tp_field__init_ptr(struct tp_field *field, int offset)
 239 {
 240         field->offset = offset;
 241         field->pointer = tp_field__ptr;
 242         return 0;
 243 }
 244
 245 static int tp_field__init_ptr(struct tp_field *field, struct tep_format_field *format_field)
 246 {
 247         return __tp_field__init_ptr(field, format_field->offset);
 248 }
 249
 250 struct syscall_tp {
 251         struct tp_field id;
 252         union {
 253                 struct tp_field args, ret;
 254         };
 255 };
 256
 257 static int perf_evsel__init_tp_uint_field(struct evsel *evsel,
 258                                           struct tp_field *field,
 259                                           const char *name)
 260 {
 261         struct tep_format_field *format_field = perf_evsel__field(evsel, name);
 262
 263         if (format_field == NULL)
 264                 return -1;
 265
 266         return tp_field__init_uint(field, format_field, evsel->needs_swap);
 267 }
 268
 269 #define perf_evsel__init_sc_tp_uint_field(evsel, name) \
 270         ({ struct syscall_tp *sc = evsel->priv;\
 271            perf_evsel__init_tp_uint_field(evsel, &sc->name, #name); })
 272
 273 static int perf_evsel__init_tp_ptr_field(struct evsel *evsel,
 274                                          struct tp_field *field,
 275                                          const char *name)
 276 {
 277         struct tep_format_field *format_field = perf_evsel__field(evsel, name);
 278
 279         if (format_field == NULL)
 280                 return -1;
 281
 282         return tp_field__init_ptr(field, format_field);
 283 }
 284
 285 #define perf_evsel__init_sc_tp_ptr_field(evsel, name) \
 286         ({ struct syscall_tp *sc = evsel->priv;\
 287            perf_evsel__init_tp_ptr_field(evsel, &sc->name, #name); })
 288
 289 static void evsel__delete_priv(struct evsel *evsel)
 290 {
 291         zfree(&evsel->priv);
 292         evsel__delete(evsel);
 293 }
 294
 295 static int perf_evsel__init_syscall_tp(struct evsel *evsel)
 296 {
 297         struct syscall_tp *sc = evsel->priv = malloc(sizeof(struct syscall_tp));
 298
 299         if (evsel->priv != NULL) {
 300                 if (perf_evsel__init_tp_uint_field(evsel, &sc->id, "__syscall_nr") &&
 301                     perf_evsel__init_tp_uint_field(evsel, &sc->id, "nr"))
 302                         goto out_delete;
 303                 return 0;
 304         }
 305
 306         return -ENOMEM;
 307 out_delete:
 308         zfree(&evsel->priv);
 309         return -ENOENT;
 310 }
 311
 312 static int perf_evsel__init_augmented_syscall_tp(struct evsel *evsel, struct evsel *tp)
 313 {
 314         struct syscall_tp *sc = evsel->priv = malloc(sizeof(struct syscall_tp));
 315
 316         if (evsel->priv != NULL) {
 317                 struct tep_format_field *syscall_id = perf_evsel__field(tp, "id");
 318                 if (syscall_id == NULL)
 319                         syscall_id = perf_evsel__field(tp, "__syscall_nr");
 320                 if (syscall_id == NULL)
 321                         goto out_delete;
 322                 if (__tp_field__init_uint(&sc->id, syscall_id->size, syscall_id->offset, evsel->needs_swap))
 323                         goto out_delete;
 324
 325                 return 0;
 326         }
 327
 328         return -ENOMEM;
 329 out_delete:
 330         zfree(&evsel->priv);
 331         return -EINVAL;
 332 }
 333
 334 static int perf_evsel__init_augmented_syscall_tp_args(struct evsel *evsel)
 335 {
 336         struct syscall_tp *sc = evsel->priv;
 337
 338         return __tp_field__init_ptr(&sc->args, sc->id.offset + sizeof(u64));
 339 }
 340
 341 static int perf_evsel__init_augmented_syscall_tp_ret(struct evsel *evsel)
 342 {
 343         struct syscall_tp *sc = evsel->priv;
 344
 345         return __tp_field__init_uint(&sc->ret, sizeof(u64), sc->id.offset + sizeof(u64), evsel->needs_swap);
 346 }
 347
 348 static int perf_evsel__init_raw_syscall_tp(struct evsel *evsel, void *handler)
 349 {
 350         evsel->priv = malloc(sizeof(struct syscall_tp));
 351         if (evsel->priv != NULL) {
 352                 if (perf_evsel__init_sc_tp_uint_field(evsel, id))
 353                         goto out_delete;
 354
 355                 evsel->handler = handler;
 356                 return 0;
 357         }
 358
 359         return -ENOMEM;
 360
 361 out_delete:
 362         zfree(&evsel->priv);
 363         return -ENOENT;
 364 }
 365
 366 static struct evsel *perf_evsel__raw_syscall_newtp(const char *direction, void *handler)
 367 {
 368         struct evsel *evsel = perf_evsel__newtp("raw_syscalls", direction);
 369
 370         /* older kernel (e.g., RHEL6) use syscalls:{enter,exit} */
 371         if (IS_ERR(evsel))
 372                 evsel = perf_evsel__newtp("syscalls", direction);
 373
 374         if (IS_ERR(evsel))
 375                 return NULL;
 376
 377         if (perf_evsel__init_raw_syscall_tp(evsel, handler))
 378                 goto out_delete;
 379
 380         return evsel;
 381
 382 out_delete:
 383         evsel__delete_priv(evsel);
 384         return NULL;
 385 }
 386
 387 #define perf_evsel__sc_tp_uint(evsel, name, sample) \
 388         ({ struct syscall_tp *fields = evsel->priv; \
 389            fields->name.integer(&fields->name, sample); })
 390
 391 #define perf_evsel__sc_tp_ptr(evsel, name, sample) \
 392         ({ struct syscall_tp *fields = evsel->priv; \
 393            fields->name.pointer(&fields->name, sample); })
 394
 395 size_t strarray__scnprintf(struct strarray *sa, char *bf, size_t size, const char *intfmt, bool show_prefix, int val)
 396 {
 397         int idx = val - sa->offset;
 398
 399         if (idx < 0 || idx >= sa->nr_entries || sa->entries[idx] == NULL) {
 400                 size_t printed = scnprintf(bf, size, intfmt, val);
 401                 if (show_prefix)
 402                         printed += scnprintf(bf + printed, size - printed, " /* %s??? */", sa->prefix);
 403                 return printed;
 404         }
 405
 406         return scnprintf(bf, size, "%s%s", show_prefix ? sa->prefix : "", sa->entries[idx]);
 407 }
 408
 409 static size_t __syscall_arg__scnprintf_strarray(char *bf, size_t size,
 410                                                 const char *intfmt,
 411                                                 struct syscall_arg *arg)
 412 {
 413         return strarray__scnprintf(arg->parm, bf, size, intfmt, arg->show_string_prefix, arg->val);
 414 }
 415
 416 static size_t syscall_arg__scnprintf_strarray(char *bf, size_t size,
 417                                               struct syscall_arg *arg)
 418 {
 419         return __syscall_arg__scnprintf_strarray(bf, size, "%d", arg);
 420 }
 421
 422 #define SCA_STRARRAY syscall_arg__scnprintf_strarray
 423
 424 size_t syscall_arg__scnprintf_strarray_flags(char *bf, size_t size, struct syscall_arg *arg)
 425 {
 426         return strarray__scnprintf_flags(arg->parm, bf, size, arg->show_string_prefix, arg->val);
 427 }
 428
 429 size_t strarrays__scnprintf(struct strarrays *sas, char *bf, size_t size, const char *intfmt, bool show_prefix, int val)
 430 {
 431         size_t printed;
 432         int i;
 433
 434         for (i = 0; i < sas->nr_entries; ++i) {
 435                 struct strarray *sa = sas->entries[i];
 436                 int idx = val - sa->offset;
 437
 438                 if (idx >= 0 && idx < sa->nr_entries) {
 439                         if (sa->entries[idx] == NULL)
 440                                 break;
 441                         return scnprintf(bf, size, "%s%s", show_prefix ? sa->prefix : "", sa->entries[idx]);
 442                 }
 443         }
 444
 445         printed = scnprintf(bf, size, intfmt, val);
 446         if (show_prefix)
 447                 printed += scnprintf(bf + printed, size - printed, " /* %s??? */", sas->entries[0]->prefix);
 448         return printed;
 449 }
 450
 451 size_t syscall_arg__scnprintf_strarrays(char *bf, size_t size,
 452                                         struct syscall_arg *arg)
 453 {
 454         return strarrays__scnprintf(arg->parm, bf, size, "%d", arg->show_string_prefix, arg->val);
 455 }
 456
 457 #ifndef AT_FDCWD
 458 #define AT_FDCWD        -100
 459 #endif
 460
 461 static size_t syscall_arg__scnprintf_fd_at(char *bf, size_t size,
 462                                            struct syscall_arg *arg)
 463 {
 464         int fd = arg->val;
 465         const char *prefix = "AT_FD";
 466
 467         if (fd == AT_FDCWD)
 468                 return scnprintf(bf, size, "%s%s", arg->show_string_prefix ? prefix : "", "CWD");
 469
 470         return syscall_arg__scnprintf_fd(bf, size, arg);
 471 }
 472
 473 #define SCA_FDAT syscall_arg__scnprintf_fd_at
 474
 475 static size_t syscall_arg__scnprintf_close_fd(char *bf, size_t size,
 476                                               struct syscall_arg *arg);
 477
 478 #define SCA_CLOSE_FD syscall_arg__scnprintf_close_fd
 479
 480 size_t syscall_arg__scnprintf_hex(char *bf, size_t size, struct syscall_arg *arg)
 481 {
 482         return scnprintf(bf, size, "%#lx", arg->val);
 483 }
 484
 485 size_t syscall_arg__scnprintf_ptr(char *bf, size_t size, struct syscall_arg *arg)
 486 {
 487         if (arg->val == 0)
 488                 return scnprintf(bf, size, "NULL");
 489         return syscall_arg__scnprintf_hex(bf, size, arg);
 490 }
 491
 492 size_t syscall_arg__scnprintf_int(char *bf, size_t size, struct syscall_arg *arg)
 493 {
 494         return scnprintf(bf, size, "%d", arg->val);
 495 }
 496
 497 size_t syscall_arg__scnprintf_long(char *bf, size_t size, struct syscall_arg *arg)
 498 {
 499         return scnprintf(bf, size, "%ld", arg->val);
 500 }
 501
 502 static const char *bpf_cmd[] = {
 503         "MAP_CREATE", "MAP_LOOKUP_ELEM", "MAP_UPDATE_ELEM", "MAP_DELETE_ELEM",
 504         "MAP_GET_NEXT_KEY", "PROG_LOAD",
 505 };
 506 static DEFINE_STRARRAY(bpf_cmd, "BPF_");
 507
 508 static const char *fsmount_flags[] = {
 509         [1] = "CLOEXEC",
 510 };
 511 static DEFINE_STRARRAY(fsmount_flags, "FSMOUNT_");
 512
 513 #include "trace/beauty/generated/fsconfig_arrays.c"
 514
 515 static DEFINE_STRARRAY(fsconfig_cmds, "FSCONFIG_");
 516
 517 static const char *epoll_ctl_ops[] = { "ADD", "DEL", "MOD", };
 518 static DEFINE_STRARRAY_OFFSET(epoll_ctl_ops, "EPOLL_CTL_", 1);
 519
 520 static const char *itimers[] = { "REAL", "VIRTUAL", "PROF", };
 521 static DEFINE_STRARRAY(itimers, "ITIMER_");
 522
 523 static const char *keyctl_options[] = {
 524         "GET_KEYRING_ID", "JOIN_SESSION_KEYRING", "UPDATE", "REVOKE", "CHOWN",
 525         "SETPERM", "DESCRIBE", "CLEAR", "LINK", "UNLINK", "SEARCH", "READ",
 526         "INSTANTIATE", "NEGATE", "SET_REQKEY_KEYRING", "SET_TIMEOUT",
 527         "ASSUME_AUTHORITY", "GET_SECURITY", "SESSION_TO_PARENT", "REJECT",
 528         "INSTANTIATE_IOV", "INVALIDATE", "GET_PERSISTENT",
 529 };
 530 static DEFINE_STRARRAY(keyctl_options, "KEYCTL_");
 531
 532 static const char *whences[] = { "SET", "CUR", "END",
 533 #ifdef SEEK_DATA
 534 "DATA",
 535 #endif
 536 #ifdef SEEK_HOLE
 537 "HOLE",
 538 #endif
 539 };
 540 static DEFINE_STRARRAY(whences, "SEEK_");
 541
 542 static const char *fcntl_cmds[] = {
 543         "DUPFD", "GETFD", "SETFD", "GETFL", "SETFL", "GETLK", "SETLK",
 544         "SETLKW", "SETOWN", "GETOWN", "SETSIG", "GETSIG", "GETLK64",
 545         "SETLK64", "SETLKW64", "SETOWN_EX", "GETOWN_EX",
 546         "GETOWNER_UIDS",
 547 };
 548 static DEFINE_STRARRAY(fcntl_cmds, "F_");
 549
 550 static const char *fcntl_linux_specific_cmds[] = {
 551         "SETLEASE", "GETLEASE", "NOTIFY", [5] = "CANCELLK", "DUPFD_CLOEXEC",
 552         "SETPIPE_SZ", "GETPIPE_SZ", "ADD_SEALS", "GET_SEALS",
 553         "GET_RW_HINT", "SET_RW_HINT", "GET_FILE_RW_HINT", "SET_FILE_RW_HINT",
 554 };
 555
 556 static DEFINE_STRARRAY_OFFSET(fcntl_linux_specific_cmds, "F_", F_LINUX_SPECIFIC_BASE);
 557
 558 static struct strarray *fcntl_cmds_arrays[] = {
 559         &strarray__fcntl_cmds,
 560         &strarray__fcntl_linux_specific_cmds,
 561 };
 562
 563 static DEFINE_STRARRAYS(fcntl_cmds_arrays);
 564
 565 static const char *rlimit_resources[] = {
 566         "CPU", "FSIZE", "DATA", "STACK", "CORE", "RSS", "NPROC", "NOFILE",
 567         "MEMLOCK", "AS", "LOCKS", "SIGPENDING", "MSGQUEUE", "NICE", "RTPRIO",
 568         "RTTIME",
 569 };
 570 static DEFINE_STRARRAY(rlimit_resources, "RLIMIT_");
 571
 572 static const char *sighow[] = { "BLOCK", "UNBLOCK", "SETMASK", };
 573 static DEFINE_STRARRAY(sighow, "SIG_");
 574
 575 static const char *clockid[] = {
 576         "REALTIME", "MONOTONIC", "PROCESS_CPUTIME_ID", "THREAD_CPUTIME_ID",
 577         "MONOTONIC_RAW", "REALTIME_COARSE", "MONOTONIC_COARSE", "BOOTTIME",
 578         "REALTIME_ALARM", "BOOTTIME_ALARM", "SGI_CYCLE", "TAI"
 579 };
 580 static DEFINE_STRARRAY(clockid, "CLOCK_");
 581
 582 static size_t syscall_arg__scnprintf_access_mode(char *bf, size_t size,
 583                                                  struct syscall_arg *arg)
 584 {
 585         bool show_prefix = arg->show_string_prefix;
 586         const char *suffix = "_OK";
 587         size_t printed = 0;
 588         int mode = arg->val;
 589
 590         if (mode == F_OK) /* 0 */
 591                 return scnprintf(bf, size, "F%s", show_prefix ? suffix : "");
 592 #define P_MODE(n) \
 593         if (mode & n##_OK) { \
 594                 printed += scnprintf(bf + printed, size - printed, "%s%s", #n, show_prefix ? suffix : ""); \
 595                 mode &= ~n##_OK; \
 596         }
 597
 598         P_MODE(R);
 599         P_MODE(W);
 600         P_MODE(X);
 601 #undef P_MODE
 602
 603         if (mode)
 604                 printed += scnprintf(bf + printed, size - printed, "|%#x", mode);
 605
 606         return printed;
 607 }
 608
 609 #define SCA_ACCMODE syscall_arg__scnprintf_access_mode
 610
 611 static size_t syscall_arg__scnprintf_filename(char *bf, size_t size,
 612                                               struct syscall_arg *arg);
 613
 614 #define SCA_FILENAME syscall_arg__scnprintf_filename
 615
 616 static size_t syscall_arg__scnprintf_pipe_flags(char *bf, size_t size,
 617                                                 struct syscall_arg *arg)
 618 {
 619         bool show_prefix = arg->show_string_prefix;
 620         const char *prefix = "O_";
 621         int printed = 0, flags = arg->val;
 622
 623 #define P_FLAG(n) \
 624         if (flags & O_##n) { \
 625                 printed += scnprintf(bf + printed, size - printed, "%s%s%s", printed ? "|" : "", show_prefix ? prefix : "", #n); \
 626                 flags &= ~O_##n; \
 627         }
 628
 629         P_FLAG(CLOEXEC);
 630         P_FLAG(NONBLOCK);
 631 #undef P_FLAG
 632
 633         if (flags)
 634                 printed += scnprintf(bf + printed, size - printed, "%s%#x", printed ? "|" : "", flags);
 635
 636         return printed;
 637 }
 638
 639 #define SCA_PIPE_FLAGS syscall_arg__scnprintf_pipe_flags
 640
 641 #ifndef GRND_NONBLOCK
 642 #define GRND_NONBLOCK   0x0001
 643 #endif
 644 #ifndef GRND_RANDOM
 645 #define GRND_RANDOM     0x0002
 646 #endif
 647
 648 static size_t syscall_arg__scnprintf_getrandom_flags(char *bf, size_t size,
 649                                                    struct syscall_arg *arg)
 650 {
 651         bool show_prefix = arg->show_string_prefix;
 652         const char *prefix = "GRND_";
 653         int printed = 0, flags = arg->val;
 654
 655 #define P_FLAG(n) \
 656         if (flags & GRND_##n) { \
 657                 printed += scnprintf(bf + printed, size - printed, "%s%s%s", printed ? "|" : "", show_prefix ? prefix : "", #n); \
 658                 flags &= ~GRND_##n; \
 659         }
 660
 661         P_FLAG(RANDOM);
 662         P_FLAG(NONBLOCK);
 663 #undef P_FLAG
 664
 665         if (flags)
 666                 printed += scnprintf(bf + printed, size - printed, "%s%#x", printed ? "|" : "", flags);
 667
 668         return printed;
 669 }
 670
 671 #define SCA_GETRANDOM_FLAGS syscall_arg__scnprintf_getrandom_flags
 672
 673 #define STRARRAY(name, array) \
 674           { .scnprintf  = SCA_STRARRAY, \
 675             .parm       = &strarray__##array, }
 676
 677 #define STRARRAY_FLAGS(name, array) \
 678           { .scnprintf  = SCA_STRARRAY_FLAGS, \
 679             .parm       = &strarray__##array, }
 680
 681 #include "trace/beauty/arch_errno_names.c"
 682 #include "trace/beauty/eventfd.c"
 683 #include "trace/beauty/futex_op.c"
 684 #include "trace/beauty/futex_val3.c"
 685 #include "trace/beauty/mmap.c"
 686 #include "trace/beauty/mode_t.c"
 687 #include "trace/beauty/msg_flags.c"
 688 #include "trace/beauty/open_flags.c"
 689 #include "trace/beauty/perf_event_open.c"
 690 #include "trace/beauty/pid.c"
 691 #include "trace/beauty/sched_policy.c"
 692 #include "trace/beauty/seccomp.c"
 693 #include "trace/beauty/signum.c"
 694 #include "trace/beauty/socket_type.c"
 695 #include "trace/beauty/waitid_options.c"
 696
 697 struct syscall_arg_fmt {
 698         size_t     (*scnprintf)(char *bf, size_t size, struct syscall_arg *arg);
 699         unsigned long (*mask_val)(struct syscall_arg *arg, unsigned long val);
 700         void       *parm;
 701         const char *name;
 702         bool       show_zero;
 703 };
 704
 705 static struct syscall_fmt {
 706         const char *name;
 707         const char *alias;
 708         struct {
 709                 const char *sys_enter,
 710                            *sys_exit;
 711         }          bpf_prog_name;
 712         struct syscall_arg_fmt arg[6];
 713         u8         nr_args;
 714         bool       errpid;
 715         bool       timeout;
 716         bool       hexret;
 717 } syscall_fmts[] = {
 718         { .name     = "access",
 719           .arg = { [1] = { .scnprintf = SCA_ACCMODE,  /* mode */ }, }, },
 720         { .name     = "arch_prctl",
 721           .arg = { [0] = { .scnprintf = SCA_X86_ARCH_PRCTL_CODE, /* code */ },
 722                    [1] = { .scnprintf = SCA_PTR, /* arg2 */ }, }, },
 723         { .name     = "bind",
 724           .arg = { [0] = { .scnprintf = SCA_INT, /* fd */ },
 725                    [1] = { .scnprintf = SCA_SOCKADDR, /* umyaddr */ },
 726                    [2] = { .scnprintf = SCA_INT, /* addrlen */ }, }, },
 727         { .name     = "bpf",
 728           .arg = { [0] = STRARRAY(cmd, bpf_cmd), }, },
 729         { .name     = "brk",        .hexret = true,
 730           .arg = { [0] = { .scnprintf = SCA_PTR, /* brk */ }, }, },
 731         { .name     = "clock_gettime",
 732           .arg = { [0] = STRARRAY(clk_id, clockid), }, },
 733         { .name     = "clone",      .errpid = true, .nr_args = 5,
 734           .arg = { [0] = { .name = "flags",         .scnprintf = SCA_CLONE_FLAGS, },
 735                    [1] = { .name = "child_stack",   .scnprintf = SCA_HEX, },
 736                    [2] = { .name = "parent_tidptr", .scnprintf = SCA_HEX, },
 737                    [3] = { .name = "child_tidptr",  .scnprintf = SCA_HEX, },
 738                    [4] = { .name = "tls",           .scnprintf = SCA_HEX, }, }, },
 739         { .name     = "close",
 740           .arg = { [0] = { .scnprintf = SCA_CLOSE_FD, /* fd */ }, }, },
 741         { .name     = "connect",
 742           .arg = { [0] = { .scnprintf = SCA_INT, /* fd */ },
 743                    [1] = { .scnprintf = SCA_SOCKADDR, /* servaddr */ },
 744                    [2] = { .scnprintf = SCA_INT, /* addrlen */ }, }, },
 745         { .name     = "epoll_ctl",
 746           .arg = { [1] = STRARRAY(op, epoll_ctl_ops), }, },
 747         { .name     = "eventfd2",
 748           .arg = { [1] = { .scnprintf = SCA_EFD_FLAGS, /* flags */ }, }, },
 749         { .name     = "fchmodat",
 750           .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
 751         { .name     = "fchownat",
 752           .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
 753         { .name     = "fcntl",
 754           .arg = { [1] = { .scnprintf = SCA_FCNTL_CMD, /* cmd */
 755                            .parm      = &strarrays__fcntl_cmds_arrays,
 756                            .show_zero = true, },
 757                    [2] = { .scnprintf =  SCA_FCNTL_ARG, /* arg */ }, }, },
 758         { .name     = "flock",
 759           .arg = { [1] = { .scnprintf = SCA_FLOCK, /* cmd */ }, }, },
 760         { .name     = "fsconfig",
 761           .arg = { [1] = STRARRAY(cmd, fsconfig_cmds), }, },
 762         { .name     = "fsmount",
 763           .arg = { [1] = STRARRAY_FLAGS(flags, fsmount_flags),
 764                    [2] = { .scnprintf = SCA_FSMOUNT_ATTR_FLAGS, /* attr_flags */ }, }, },
 765         { .name     = "fspick",
 766           .arg = { [0] = { .scnprintf = SCA_FDAT,         /* dfd */ },
 767                    [1] = { .scnprintf = SCA_FILENAME,     /* path */ },
 768                    [2] = { .scnprintf = SCA_FSPICK_FLAGS, /* flags */ }, }, },
 769         { .name     = "fstat", .alias = "newfstat", },
 770         { .name     = "fstatat", .alias = "newfstatat", },
 771         { .name     = "futex",
 772           .arg = { [1] = { .scnprintf = SCA_FUTEX_OP, /* op */ },
 773                    [5] = { .scnprintf = SCA_FUTEX_VAL3, /* val3 */ }, }, },
 774         { .name     = "futimesat",
 775           .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
 776         { .name     = "getitimer",
 777           .arg = { [0] = STRARRAY(which, itimers), }, },
 778         { .name     = "getpid",     .errpid = true, },
 779         { .name     = "getpgid",    .errpid = true, },
 780         { .name     = "getppid",    .errpid = true, },
 781         { .name     = "getrandom",
 782           .arg = { [2] = { .scnprintf = SCA_GETRANDOM_FLAGS, /* flags */ }, }, },
 783         { .name     = "getrlimit",
 784           .arg = { [0] = STRARRAY(resource, rlimit_resources), }, },
 785         { .name     = "gettid",     .errpid = true, },
 786         { .name     = "ioctl",
 787           .arg = {
 788 #if defined(__i386__) || defined(__x86_64__)
 789 /*
 790  * FIXME: Make this available to all arches.
 791  */
 792                    [1] = { .scnprintf = SCA_IOCTL_CMD, /* cmd */ },
 793                    [2] = { .scnprintf = SCA_HEX, /* arg */ }, }, },
 794 #else
 795                    [2] = { .scnprintf = SCA_HEX, /* arg */ }, }, },
 796 #endif
 797         { .name     = "kcmp",       .nr_args = 5,
 798           .arg = { [0] = { .name = "pid1",      .scnprintf = SCA_PID, },
 799                    [1] = { .name = "pid2",      .scnprintf = SCA_PID, },
 800                    [2] = { .name = "type",      .scnprintf = SCA_KCMP_TYPE, },
 801                    [3] = { .name = "idx1",      .scnprintf = SCA_KCMP_IDX, },
 802                    [4] = { .name = "idx2",      .scnprintf = SCA_KCMP_IDX, }, }, },
 803         { .name     = "keyctl",
 804           .arg = { [0] = STRARRAY(option, keyctl_options), }, },
 805         { .name     = "kill",
 806           .arg = { [1] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
 807         { .name     = "linkat",
 808           .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
 809         { .name     = "lseek",
 810           .arg = { [2] = STRARRAY(whence, whences), }, },
 811         { .name     = "lstat", .alias = "newlstat", },
 812         { .name     = "madvise",
 813           .arg = { [0] = { .scnprintf = SCA_HEX,      /* start */ },
 814                    [2] = { .scnprintf = SCA_MADV_BHV, /* behavior */ }, }, },
 815         { .name     = "mkdirat",
 816           .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
 817         { .name     = "mknodat",
 818           .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
 819         { .name     = "mmap",       .hexret = true,
 820 /* The standard mmap maps to old_mmap on s390x */
 821 #if defined(__s390x__)
 822         .alias = "old_mmap",
 823 #endif
 824           .arg = { [2] = { .scnprintf = SCA_MMAP_PROT,  /* prot */ },
 825                    [3] = { .scnprintf = SCA_MMAP_FLAGS, /* flags */ },
 826                    [5] = { .scnprintf = SCA_HEX,        /* offset */ }, }, },
 827         { .name     = "mount",
 828           .arg = { [0] = { .scnprintf = SCA_FILENAME, /* dev_name */ },
 829                    [3] = { .scnprintf = SCA_MOUNT_FLAGS, /* flags */
 830                            .mask_val  = SCAMV_MOUNT_FLAGS, /* flags */ }, }, },
 831         { .name     = "move_mount",
 832           .arg = { [0] = { .scnprintf = SCA_FDAT,       /* from_dfd */ },
 833                    [1] = { .scnprintf = SCA_FILENAME, /* from_pathname */ },
 834                    [2] = { .scnprintf = SCA_FDAT,       /* to_dfd */ },
 835                    [3] = { .scnprintf = SCA_FILENAME, /* to_pathname */ },
 836                    [4] = { .scnprintf = SCA_MOVE_MOUNT_FLAGS, /* flags */ }, }, },
 837         { .name     = "mprotect",
 838           .arg = { [0] = { .scnprintf = SCA_HEX,        /* start */ },
 839                    [2] = { .scnprintf = SCA_MMAP_PROT,  /* prot */ }, }, },
 840         { .name     = "mq_unlink",
 841           .arg = { [0] = { .scnprintf = SCA_FILENAME, /* u_name */ }, }, },
 842         { .name     = "mremap",     .hexret = true,
 843           .arg = { [3] = { .scnprintf = SCA_MREMAP_FLAGS, /* flags */ }, }, },
 844         { .name     = "name_to_handle_at",
 845           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
 846         { .name     = "newfstatat",
 847           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
 848         { .name     = "open",
 849           .arg = { [1] = { .scnprintf = SCA_OPEN_FLAGS, /* flags */ }, }, },
 850         { .name     = "open_by_handle_at",
 851           .arg = { [0] = { .scnprintf = SCA_FDAT,       /* dfd */ },
 852                    [2] = { .scnprintf = SCA_OPEN_FLAGS, /* flags */ }, }, },
 853         { .name     = "openat",
 854           .arg = { [0] = { .scnprintf = SCA_FDAT,       /* dfd */ },
 855                    [2] = { .scnprintf = SCA_OPEN_FLAGS, /* flags */ }, }, },
 856         { .name     = "perf_event_open",
 857           .arg = { [2] = { .scnprintf = SCA_INT,        /* cpu */ },
 858                    [3] = { .scnprintf = SCA_FD,         /* group_fd */ },
 859                    [4] = { .scnprintf = SCA_PERF_FLAGS, /* flags */ }, }, },
 860         { .name     = "pipe2",
 861           .arg = { [1] = { .scnprintf = SCA_PIPE_FLAGS, /* flags */ }, }, },
 862         { .name     = "pkey_alloc",
 863           .arg = { [1] = { .scnprintf = SCA_PKEY_ALLOC_ACCESS_RIGHTS,   /* access_rights */ }, }, },
 864         { .name     = "pkey_free",
 865           .arg = { [0] = { .scnprintf = SCA_INT,        /* key */ }, }, },
 866         { .name     = "pkey_mprotect",
 867           .arg = { [0] = { .scnprintf = SCA_HEX,        /* start */ },
 868                    [2] = { .scnprintf = SCA_MMAP_PROT,  /* prot */ },
 869                    [3] = { .scnprintf = SCA_INT,        /* pkey */ }, }, },
 870         { .name     = "poll", .timeout = true, },
 871         { .name     = "ppoll", .timeout = true, },
 872         { .name     = "prctl",
 873           .arg = { [0] = { .scnprintf = SCA_PRCTL_OPTION, /* option */ },
 874                    [1] = { .scnprintf = SCA_PRCTL_ARG2, /* arg2 */ },
 875                    [2] = { .scnprintf = SCA_PRCTL_ARG3, /* arg3 */ }, }, },
 876         { .name     = "pread", .alias = "pread64", },
 877         { .name     = "preadv", .alias = "pread", },
 878         { .name     = "prlimit64",
 879           .arg = { [1] = STRARRAY(resource, rlimit_resources), }, },
 880         { .name     = "pwrite", .alias = "pwrite64", },
 881         { .name     = "readlinkat",
 882           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
 883         { .name     = "recvfrom",
 884           .arg = { [3] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
 885         { .name     = "recvmmsg",
 886           .arg = { [3] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
 887         { .name     = "recvmsg",
 888           .arg = { [2] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
 889         { .name     = "renameat",
 890           .arg = { [0] = { .scnprintf = SCA_FDAT, /* olddirfd */ },
 891                    [2] = { .scnprintf = SCA_FDAT, /* newdirfd */ }, }, },
 892         { .name     = "renameat2",
 893           .arg = { [0] = { .scnprintf = SCA_FDAT, /* olddirfd */ },
 894                    [2] = { .scnprintf = SCA_FDAT, /* newdirfd */ },
 895                    [4] = { .scnprintf = SCA_RENAMEAT2_FLAGS, /* flags */ }, }, },
 896         { .name     = "rt_sigaction",
 897           .arg = { [0] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
 898         { .name     = "rt_sigprocmask",
 899           .arg = { [0] = STRARRAY(how, sighow), }, },
 900         { .name     = "rt_sigqueueinfo",
 901           .arg = { [1] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
 902         { .name     = "rt_tgsigqueueinfo",
 903           .arg = { [2] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
 904         { .name     = "sched_setscheduler",
 905           .arg = { [1] = { .scnprintf = SCA_SCHED_POLICY, /* policy */ }, }, },
 906         { .name     = "seccomp",
 907           .arg = { [0] = { .scnprintf = SCA_SECCOMP_OP,    /* op */ },
 908                    [1] = { .scnprintf = SCA_SECCOMP_FLAGS, /* flags */ }, }, },
 909         { .name     = "select", .timeout = true, },
 910         { .name     = "sendfile", .alias = "sendfile64", },
 911         { .name     = "sendmmsg",
 912           .arg = { [3] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
 913         { .name     = "sendmsg",
 914           .arg = { [2] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
 915         { .name     = "sendto",
 916           .arg = { [3] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ },
 917                    [4] = { .scnprintf = SCA_SOCKADDR, /* addr */ }, }, },
 918         { .name     = "set_tid_address", .errpid = true, },
 919         { .name     = "setitimer",
 920           .arg = { [0] = STRARRAY(which, itimers), }, },
 921         { .name     = "setrlimit",
 922           .arg = { [0] = STRARRAY(resource, rlimit_resources), }, },
 923         { .name     = "socket",
 924           .arg = { [0] = STRARRAY(family, socket_families),
 925                    [1] = { .scnprintf = SCA_SK_TYPE, /* type */ },
 926                    [2] = { .scnprintf = SCA_SK_PROTO, /* protocol */ }, }, },
 927         { .name     = "socketpair",
 928           .arg = { [0] = STRARRAY(family, socket_families),
 929                    [1] = { .scnprintf = SCA_SK_TYPE, /* type */ },
 930                    [2] = { .scnprintf = SCA_SK_PROTO, /* protocol */ }, }, },
 931         { .name     = "stat", .alias = "newstat", },
 932         { .name     = "statx",
 933           .arg = { [0] = { .scnprintf = SCA_FDAT,        /* fdat */ },
 934                    [2] = { .scnprintf = SCA_STATX_FLAGS, /* flags */ } ,
 935                    [3] = { .scnprintf = SCA_STATX_MASK,  /* mask */ }, }, },
 936         { .name     = "swapoff",
 937           .arg = { [0] = { .scnprintf = SCA_FILENAME, /* specialfile */ }, }, },
 938         { .name     = "swapon",
 939           .arg = { [0] = { .scnprintf = SCA_FILENAME, /* specialfile */ }, }, },
 940         { .name     = "symlinkat",
 941           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
 942         { .name     = "sync_file_range",
 943           .arg = { [3] = { .scnprintf = SCA_SYNC_FILE_RANGE_FLAGS, /* flags */ }, }, },
 944         { .name     = "tgkill",
 945           .arg = { [2] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
 946         { .name     = "tkill",
 947           .arg = { [1] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
 948         { .name     = "umount2", .alias = "umount",
 949           .arg = { [0] = { .scnprintf = SCA_FILENAME, /* name */ }, }, },
 950         { .name     = "uname", .alias = "newuname", },
 951         { .name     = "unlinkat",
 952           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
 953         { .name     = "utimensat",
 954           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dirfd */ }, }, },
 955         { .name     = "wait4",      .errpid = true,
 956           .arg = { [2] = { .scnprintf = SCA_WAITID_OPTIONS, /* options */ }, }, },
 957         { .name     = "waitid",     .errpid = true,
 958           .arg = { [3] = { .scnprintf = SCA_WAITID_OPTIONS, /* options */ }, }, },
 959 };
 960
 961 static int syscall_fmt__cmp(const void *name, const void *fmtp)
 962 {
 963         const struct syscall_fmt *fmt = fmtp;
 964         return strcmp(name, fmt->name);
 965 }
 966
 967 static struct syscall_fmt *syscall_fmt__find(const char *name)
 968 {
 969         const int nmemb = ARRAY_SIZE(syscall_fmts);
 970         return bsearch(name, syscall_fmts, nmemb, sizeof(struct syscall_fmt), syscall_fmt__cmp);
 971 }
 972
 973 static struct syscall_fmt *syscall_fmt__find_by_alias(const char *alias)
 974 {
 975         int i, nmemb = ARRAY_SIZE(syscall_fmts);
 976
 977         for (i = 0; i < nmemb; ++i) {
 978                 if (syscall_fmts[i].alias && strcmp(syscall_fmts[i].alias, alias) == 0)
 979                         return &syscall_fmts[i];
 980         }
 981
 982         return NULL;
 983 }
 984
 985 /*
 986  * is_exit: is this "exit" or "exit_group"?
 987  * is_open: is this "open" or "openat"? To associate the fd returned in sys_exit with the pathname in sys_enter.
 988  * args_size: sum of the sizes of the syscall arguments, anything after that is augmented stuff: pathname for openat, etc.
 989  * nonexistent: Just a hole in the syscall table, syscall id not allocated
 990  */
 991 struct syscall {
 992         struct tep_event    *tp_format;
 993         int                 nr_args;
 994         int                 args_size;
 995         struct {
 996                 struct bpf_program *sys_enter,
 997                                    *sys_exit;
 998         }                   bpf_prog;
 999         bool                is_exit;
1000         bool                is_open;
1001         bool                nonexistent;
1002         struct tep_format_field *args;
1003         const char          *name;
1004         struct syscall_fmt  *fmt;
1005         struct syscall_arg_fmt *arg_fmt;
1006 };
1007
1008 /*
1009  * Must match what is in the BPF program:
1010  *
1011  * tools/perf/examples/bpf/augmented_raw_syscalls.c
1012  */
1013 struct bpf_map_syscall_entry {
1014         bool    enabled;
1015         u16     string_args_len[6];
1016 };
1017
1018 /*
1019  * We need to have this 'calculated' boolean because in some cases we really
1020  * don't know what is the duration of a syscall, for instance, when we start
1021  * a session and some threads are waiting for a syscall to finish, say 'poll',
1022  * in which case all we can do is to print "( ? ) for duration and for the
1023  * start timestamp.
1024  */
1025 static size_t fprintf_duration(unsigned long t, bool calculated, FILE *fp)
1026 {
1027         double duration = (double)t / NSEC_PER_MSEC;
1028         size_t printed = fprintf(fp, "(");
1029
1030         if (!calculated)
1031                 printed += fprintf(fp, "         ");
1032         else if (duration >= 1.0)
1033                 printed += color_fprintf(fp, PERF_COLOR_RED, "%6.3f ms", duration);
1034         else if (duration >= 0.01)
1035                 printed += color_fprintf(fp, PERF_COLOR_YELLOW, "%6.3f ms", duration);
1036         else
1037                 printed += color_fprintf(fp, PERF_COLOR_NORMAL, "%6.3f ms", duration);
1038         return printed + fprintf(fp, "): ");
1039 }
1040
1041 /**
1042  * filename.ptr: The filename char pointer that will be vfs_getname'd
1043  * filename.entry_str_pos: Where to insert the string translated from
1044  *                         filename.ptr by the vfs_getname tracepoint/kprobe.
1045  * ret_scnprintf: syscall args may set this to a different syscall return
1046  *                formatter, for instance, fcntl may return fds, file flags, etc.
1047  */
1048 struct thread_trace {
1049         u64               entry_time;
1050         bool              entry_pending;
1051         unsigned long     nr_events;
1052         unsigned long     pfmaj, pfmin;
1053         char              *entry_str;
1054         double            runtime_ms;
1055         size_t            (*ret_scnprintf)(char *bf, size_t size, struct syscall_arg *arg);
1056         struct {
1057                 unsigned long ptr;
1058                 short int     entry_str_pos;
1059                 bool          pending_open;
1060                 unsigned int  namelen;
1061                 char          *name;
1062         } filename;
1063         struct {
1064                 int           max;
1065                 struct file   *table;
1066         } files;
1067
1068         struct intlist *syscall_stats;
1069 };
1070
1071 static struct thread_trace *thread_trace__new(void)
1072 {
1073         struct thread_trace *ttrace =  zalloc(sizeof(struct thread_trace));
1074
1075         if (ttrace) {
1076                 ttrace->files.max = -1;
1077                 ttrace->syscall_stats = intlist__new(NULL);
1078         }
1079
1080         return ttrace;
1081 }
1082
1083 static struct thread_trace *thread__trace(struct thread *thread, FILE *fp)
1084 {
1085         struct thread_trace *ttrace;
1086
1087         if (thread == NULL)
1088                 goto fail;
1089
1090         if (thread__priv(thread) == NULL)
1091                 thread__set_priv(thread, thread_trace__new());
1092
1093         if (thread__priv(thread) == NULL)
1094                 goto fail;
1095
1096         ttrace = thread__priv(thread);
1097         ++ttrace->nr_events;
1098
1099         return ttrace;
1100 fail:
1101         color_fprintf(fp, PERF_COLOR_RED,
1102                       "WARNING: not enough memory, dropping samples!\n");
1103         return NULL;
1104 }
1105
1106
1107 void syscall_arg__set_ret_scnprintf(struct syscall_arg *arg,
1108                                     size_t (*ret_scnprintf)(char *bf, size_t size, struct syscall_arg *arg))
1109 {
1110         struct thread_trace *ttrace = thread__priv(arg->thread);
1111
1112         ttrace->ret_scnprintf = ret_scnprintf;
1113 }
1114
1115 #define TRACE_PFMAJ             (1 << 0)
1116 #define TRACE_PFMIN             (1 << 1)
1117
1118 static const size_t trace__entry_str_size = 2048;
1119
1120 static struct file *thread_trace__files_entry(struct thread_trace *ttrace, int fd)
1121 {
1122         if (fd < 0)
1123                 return NULL;
1124
1125         if (fd > ttrace->files.max) {
1126                 struct file *nfiles = realloc(ttrace->files.table, (fd + 1) * sizeof(struct file));
1127
1128                 if (nfiles == NULL)
1129                         return NULL;
1130
1131                 if (ttrace->files.max != -1) {
1132                         memset(nfiles + ttrace->files.max + 1, 0,
1133                                (fd - ttrace->files.max) * sizeof(struct file));
1134                 } else {
1135                         memset(nfiles, 0, (fd + 1) * sizeof(struct file));
1136                 }
1137
1138                 ttrace->files.table = nfiles;
1139                 ttrace->files.max   = fd;
1140         }
1141
1142         return ttrace->files.table + fd;
1143 }
1144
1145 struct file *thread__files_entry(struct thread *thread, int fd)
1146 {
1147         return thread_trace__files_entry(thread__priv(thread), fd);
1148 }
1149
1150 static int trace__set_fd_pathname(struct thread *thread, int fd, const char *pathname)
1151 {
1152         struct thread_trace *ttrace = thread__priv(thread);
1153         struct file *file = thread_trace__files_entry(ttrace, fd);
1154
1155         if (file != NULL) {
1156                 struct stat st;
1157                 if (stat(pathname, &st) == 0)
1158                         file->dev_maj = major(st.st_rdev);
1159                 file->pathname = strdup(pathname);
1160                 if (file->pathname)
1161                         return 0;
1162         }
1163
1164         return -1;
1165 }
1166
1167 static int thread__read_fd_path(struct thread *thread, int fd)
1168 {
1169         char linkname[PATH_MAX], pathname[PATH_MAX];
1170         struct stat st;
1171         int ret;
1172
1173         if (thread->pid_ == thread->tid) {
1174                 scnprintf(linkname, sizeof(linkname),
1175                           "/proc/%d/fd/%d", thread->pid_, fd);
1176         } else {
1177                 scnprintf(linkname, sizeof(linkname),
1178                           "/proc/%d/task/%d/fd/%d", thread->pid_, thread->tid, fd);
1179         }
1180
1181         if (lstat(linkname, &st) < 0 || st.st_size + 1 > (off_t)sizeof(pathname))
1182                 return -1;
1183
1184         ret = readlink(linkname, pathname, sizeof(pathname));
1185
1186         if (ret < 0 || ret > st.st_size)
1187                 return -1;
1188
1189         pathname[ret] = '\0';
1190         return trace__set_fd_pathname(thread, fd, pathname);
1191 }
1192
1193 static const char *thread__fd_path(struct thread *thread, int fd,
1194                                    struct trace *trace)
1195 {
1196         struct thread_trace *ttrace = thread__priv(thread);
1197
1198         if (ttrace == NULL || trace->fd_path_disabled)
1199                 return NULL;
1200
1201         if (fd < 0)
1202                 return NULL;
1203
1204         if ((fd > ttrace->files.max || ttrace->files.table[fd].pathname == NULL)) {
1205                 if (!trace->live)
1206                         return NULL;
1207                 ++trace->stats.proc_getname;
1208                 if (thread__read_fd_path(thread, fd))
1209                         return NULL;
1210         }
1211
1212         return ttrace->files.table[fd].pathname;
1213 }
1214
1215 size_t syscall_arg__scnprintf_fd(char *bf, size_t size, struct syscall_arg *arg)
1216 {
1217         int fd = arg->val;
1218         size_t printed = scnprintf(bf, size, "%d", fd);
1219         const char *path = thread__fd_path(arg->thread, fd, arg->trace);
1220
1221         if (path)
1222                 printed += scnprintf(bf + printed, size - printed, "<%s>", path);
1223
1224         return printed;
1225 }
1226
1227 size_t pid__scnprintf_fd(struct trace *trace, pid_t pid, int fd, char *bf, size_t size)
1228 {
1229         size_t printed = scnprintf(bf, size, "%d", fd);
1230         struct thread *thread = machine__find_thread(trace->host, pid, pid);
1231
1232         if (thread) {
1233                 const char *path = thread__fd_path(thread, fd, trace);
1234
1235                 if (path)
1236                         printed += scnprintf(bf + printed, size - printed, "<%s>", path);
1237
1238                 thread__put(thread);
1239         }
1240
1241         return printed;
1242 }
1243
1244 static size_t syscall_arg__scnprintf_close_fd(char *bf, size_t size,
1245                                               struct syscall_arg *arg)
1246 {
1247         int fd = arg->val;
1248         size_t printed = syscall_arg__scnprintf_fd(bf, size, arg);
1249         struct thread_trace *ttrace = thread__priv(arg->thread);
1250
1251         if (ttrace && fd >= 0 && fd <= ttrace->files.max)
1252                 zfree(&ttrace->files.table[fd].pathname);
1253
1254         return printed;
1255 }
1256
1257 static void thread__set_filename_pos(struct thread *thread, const char *bf,
1258                                      unsigned long ptr)
1259 {
1260         struct thread_trace *ttrace = thread__priv(thread);
1261
1262         ttrace->filename.ptr = ptr;
1263         ttrace->filename.entry_str_pos = bf - ttrace->entry_str;
1264 }
1265
1266 static size_t syscall_arg__scnprintf_augmented_string(struct syscall_arg *arg, char *bf, size_t size)
1267 {
1268         struct augmented_arg *augmented_arg = arg->augmented.args;
1269         size_t printed = scnprintf(bf, size, "\"%.*s\"", augmented_arg->size, augmented_arg->value);
1270         /*
1271          * So that the next arg with a payload can consume its augmented arg, i.e. for rename* syscalls
1272          * we would have two strings, each prefixed by its size.
1273          */
1274         int consumed = sizeof(*augmented_arg) + augmented_arg->size;
1275
1276         arg->augmented.args = ((void *)arg->augmented.args) + consumed;
1277         arg->augmented.size -= consumed;
1278
1279         return printed;
1280 }
1281
1282 static size_t syscall_arg__scnprintf_filename(char *bf, size_t size,
1283                                               struct syscall_arg *arg)
1284 {
1285         unsigned long ptr = arg->val;
1286
1287         if (arg->augmented.args)
1288                 return syscall_arg__scnprintf_augmented_string(arg, bf, size);
1289
1290         if (!arg->trace->vfs_getname)
1291                 return scnprintf(bf, size, "%#x", ptr);
1292
1293         thread__set_filename_pos(arg->thread, bf, ptr);
1294         return 0;
1295 }
1296
1297 static bool trace__filter_duration(struct trace *trace, double t)
1298 {
1299         return t < (trace->duration_filter * NSEC_PER_MSEC);
1300 }
1301
1302 static size_t __trace__fprintf_tstamp(struct trace *trace, u64 tstamp, FILE *fp)
1303 {
1304         double ts = (double)(tstamp - trace->base_time) / NSEC_PER_MSEC;
1305
1306         return fprintf(fp, "%10.3f ", ts);
1307 }
1308
1309 /*
1310  * We're handling tstamp=0 as an undefined tstamp, i.e. like when we are
1311  * using ttrace->entry_time for a thread that receives a sys_exit without
1312  * first having received a sys_enter ("poll" issued before tracing session
1313  * starts, lost sys_enter exit due to ring buffer overflow).
1314  */
1315 static size_t trace__fprintf_tstamp(struct trace *trace, u64 tstamp, FILE *fp)
1316 {
1317         if (tstamp > 0)
1318                 return __trace__fprintf_tstamp(trace, tstamp, fp);
1319
1320         return fprintf(fp, "         ? ");
1321 }
1322
1323 static bool done = false;
1324 static bool interrupted = false;
1325
1326 static void sig_handler(int sig)
1327 {
1328         done = true;
1329         interrupted = sig == SIGINT;
1330 }
1331
1332 static size_t trace__fprintf_comm_tid(struct trace *trace, struct thread *thread, FILE *fp)
1333 {
1334         size_t printed = 0;
1335
1336         if (trace->multiple_threads) {
1337                 if (trace->show_comm)
1338                         printed += fprintf(fp, "%.14s/", thread__comm_str(thread));
1339                 printed += fprintf(fp, "%d ", thread->tid);
1340         }
1341
1342         return printed;
1343 }
1344
1345 static size_t trace__fprintf_entry_head(struct trace *trace, struct thread *thread,
1346                                         u64 duration, bool duration_calculated, u64 tstamp, FILE *fp)
1347 {
1348         size_t printed = 0;
1349
1350         if (trace->show_tstamp)
1351                 printed = trace__fprintf_tstamp(trace, tstamp, fp);
1352         if (trace->show_duration)
1353                 printed += fprintf_duration(duration, duration_calculated, fp);
1354         return printed + trace__fprintf_comm_tid(trace, thread, fp);
1355 }
1356
1357 static int trace__process_event(struct trace *trace, struct machine *machine,
1358                                 union perf_event *event, struct perf_sample *sample)
1359 {
1360         int ret = 0;
1361
1362         switch (event->header.type) {
1363         case PERF_RECORD_LOST:
1364                 color_fprintf(trace->output, PERF_COLOR_RED,
1365                               "LOST %" PRIu64 " events!\n", event->lost.lost);
1366                 ret = machine__process_lost_event(machine, event, sample);
1367                 break;
1368         default:
1369                 ret = machine__process_event(machine, event, sample);
1370                 break;
1371         }
1372
1373         return ret;
1374 }
1375
1376 static int trace__tool_process(struct perf_tool *tool,
1377                                union perf_event *event,
1378                                struct perf_sample *sample,
1379                                struct machine *machine)
1380 {
1381         struct trace *trace = container_of(tool, struct trace, tool);
1382         return trace__process_event(trace, machine, event, sample);
1383 }
1384
1385 static char *trace__machine__resolve_kernel_addr(void *vmachine, unsigned long long *addrp, char **modp)
1386 {
1387         struct machine *machine = vmachine;
1388
1389         if (machine->kptr_restrict_warned)
1390                 return NULL;
1391
1392         if (symbol_conf.kptr_restrict) {
1393                 pr_warning("Kernel address maps (/proc/{kallsyms,modules}) are restricted.\n\n"
1394                            "Check /proc/sys/kernel/kptr_restrict and /proc/sys/kernel/perf_event_paranoid.\n\n"
1395                            "Kernel samples will not be resolved.\n");
1396                 machine->kptr_restrict_warned = true;
1397                 return NULL;
1398         }
1399
1400         return machine__resolve_kernel_addr(vmachine, addrp, modp);
1401 }
1402
1403 static int trace__symbols_init(struct trace *trace, struct evlist *evlist)
1404 {
1405         int err = symbol__init(NULL);
1406
1407         if (err)
1408                 return err;
1409
1410         trace->host = machine__new_host();
1411         if (trace->host == NULL)
1412                 return -ENOMEM;
1413
1414         err = trace_event__register_resolver(trace->host, trace__machine__resolve_kernel_addr);
1415         if (err < 0)
1416                 goto out;
1417
1418         err = __machine__synthesize_threads(trace->host, &trace->tool, &trace->opts.target,
1419                                             evlist->core.threads, trace__tool_process, false,
1420                                             1);
1421 out:
1422         if (err)
1423                 symbol__exit();
1424
1425         return err;
1426 }
1427
1428 static void trace__symbols__exit(struct trace *trace)
1429 {
1430         machine__exit(trace->host);
1431         trace->host = NULL;
1432
1433         symbol__exit();
1434 }
1435
1436 static int syscall__alloc_arg_fmts(struct syscall *sc, int nr_args)
1437 {
1438         int idx;
1439
1440         if (nr_args == 6 && sc->fmt && sc->fmt->nr_args != 0)
1441                 nr_args = sc->fmt->nr_args;
1442
1443         sc->arg_fmt = calloc(nr_args, sizeof(*sc->arg_fmt));
1444         if (sc->arg_fmt == NULL)
1445                 return -1;
1446
1447         for (idx = 0; idx < nr_args; ++idx) {
1448                 if (sc->fmt)
1449                         sc->arg_fmt[idx] = sc->fmt->arg[idx];
1450         }
1451
1452         sc->nr_args = nr_args;
1453         return 0;
1454 }
1455
1456 static int syscall__set_arg_fmts(struct syscall *sc)
1457 {
1458         struct tep_format_field *field, *last_field = NULL;
1459         int idx = 0, len;
1460
1461         for (field = sc->args; field; field = field->next, ++idx) {
1462                 last_field = field;
1463
1464                 if (sc->fmt && sc->fmt->arg[idx].scnprintf)
1465                         continue;
1466
1467                 len = strlen(field->name);
1468
1469                 if (strcmp(field->type, "const char *") == 0 &&
1470                     ((len >= 4 && strcmp(field->name + len - 4, "name") == 0) ||
1471                      strstr(field->name, "path") != NULL))
1472                         sc->arg_fmt[idx].scnprintf = SCA_FILENAME;
1473                 else if ((field->flags & TEP_FIELD_IS_POINTER) || strstr(field->name, "addr"))
1474                         sc->arg_fmt[idx].scnprintf = SCA_PTR;
1475                 else if (strcmp(field->type, "pid_t") == 0)
1476                         sc->arg_fmt[idx].scnprintf = SCA_PID;
1477                 else if (strcmp(field->type, "umode_t") == 0)
1478                         sc->arg_fmt[idx].scnprintf = SCA_MODE_T;
1479                 else if ((strcmp(field->type, "int") == 0 ||
1480                           strcmp(field->type, "unsigned int") == 0 ||
1481                           strcmp(field->type, "long") == 0) &&
1482                          len >= 2 && strcmp(field->name + len - 2, "fd") == 0) {
1483                         /*
1484                          * /sys/kernel/tracing/events/syscalls/sys_enter*
1485                          * egrep 'field:.*fd;' .../format|sed -r 's/.*field:([a-z ]+) [a-z_]*fd.+/\1/g'|sort|uniq -c
1486                          * 65 int
1487                          * 23 unsigned int
1488                          * 7 unsigned long
1489                          */
1490                         sc->arg_fmt[idx].scnprintf = SCA_FD;
1491                 }
1492         }
1493
1494         if (last_field)
1495                 sc->args_size = last_field->offset + last_field->size;
1496
1497         return 0;
1498 }
1499
1500 static int trace__read_syscall_info(struct trace *trace, int id)
1501 {
1502         char tp_name[128];
1503         struct syscall *sc;
1504         const char *name = syscalltbl__name(trace->sctbl, id);
1505
1506         if (trace->syscalls.table == NULL) {
1507                 trace->syscalls.table = calloc(trace->sctbl->syscalls.max_id + 1, sizeof(*sc));
1508                 if (trace->syscalls.table == NULL)
1509                         return -ENOMEM;
1510         }
1511
1512         sc = trace->syscalls.table + id;
1513         if (sc->nonexistent)
1514                 return 0;
1515
1516         if (name == NULL) {
1517                 sc->nonexistent = true;
1518                 return 0;
1519         }
1520
1521         sc->name = name;
1522         sc->fmt  = syscall_fmt__find(sc->name);
1523
1524         snprintf(tp_name, sizeof(tp_name), "sys_enter_%s", sc->name);
1525         sc->tp_format = trace_event__tp_format("syscalls", tp_name);
1526
1527         if (IS_ERR(sc->tp_format) && sc->fmt && sc->fmt->alias) {
1528                 snprintf(tp_name, sizeof(tp_name), "sys_enter_%s", sc->fmt->alias);
1529                 sc->tp_format = trace_event__tp_format("syscalls", tp_name);
1530         }
1531
1532         if (syscall__alloc_arg_fmts(sc, IS_ERR(sc->tp_format) ? 6 : sc->tp_format->format.nr_fields))
1533                 return -ENOMEM;
1534
1535         if (IS_ERR(sc->tp_format))
1536                 return PTR_ERR(sc->tp_format);
1537
1538         sc->args = sc->tp_format->format.fields;
1539         /*
1540          * We need to check and discard the first variable '__syscall_nr'
1541          * or 'nr' that mean the syscall number. It is needless here.
1542          * So drop '__syscall_nr' or 'nr' field but does not exist on older kernels.
1543          */
1544         if (sc->args && (!strcmp(sc->args->name, "__syscall_nr") || !strcmp(sc->args->name, "nr"))) {
1545                 sc->args = sc->args->next;
1546                 --sc->nr_args;
1547         }
1548
1549         sc->is_exit = !strcmp(name, "exit_group") || !strcmp(name, "exit");
1550         sc->is_open = !strcmp(name, "open") || !strcmp(name, "openat");
1551
1552         return syscall__set_arg_fmts(sc);
1553 }
1554
1555 static int intcmp(const void *a, const void *b)
1556 {
1557         const int *one = a, *another = b;
1558
1559         return *one - *another;
1560 }
1561
1562 static int trace__validate_ev_qualifier(struct trace *trace)
1563 {
1564         int err = 0;
1565         bool printed_invalid_prefix = false;
1566         struct str_node *pos;
1567         size_t nr_used = 0, nr_allocated = strlist__nr_entries(trace->ev_qualifier);
1568
1569         trace->ev_qualifier_ids.entries = malloc(nr_allocated *
1570                                                  sizeof(trace->ev_qualifier_ids.entries[0]));
1571
1572         if (trace->ev_qualifier_ids.entries == NULL) {
1573                 fputs("Error:\tNot enough memory for allocating events qualifier ids\n",
1574                        trace->output);
1575                 err = -EINVAL;
1576                 goto out;
1577         }
1578
1579         strlist__for_each_entry(pos, trace->ev_qualifier) {
1580                 const char *sc = pos->s;
1581                 int id = syscalltbl__id(trace->sctbl, sc), match_next = -1;
1582
1583                 if (id < 0) {
1584                         id = syscalltbl__strglobmatch_first(trace->sctbl, sc, &match_next);
1585                         if (id >= 0)
1586                                 goto matches;
1587
1588                         if (!printed_invalid_prefix) {
1589                                 pr_debug("Skipping unknown syscalls: ");
1590                                 printed_invalid_prefix = true;
1591                         } else {
1592                                 pr_debug(", ");
1593                         }
1594
1595                         pr_debug("%s", sc);
1596                         continue;
1597                 }
1598 matches:
1599                 trace->ev_qualifier_ids.entries[nr_used++] = id;
1600                 if (match_next == -1)
1601                         continue;
1602
1603                 while (1) {
1604                         id = syscalltbl__strglobmatch_next(trace->sctbl, sc, &match_next);
1605                         if (id < 0)
1606                                 break;
1607                         if (nr_allocated == nr_used) {
1608                                 void *entries;
1609
1610                                 nr_allocated += 8;
1611                                 entries = realloc(trace->ev_qualifier_ids.entries,
1612                                                   nr_allocated * sizeof(trace->ev_qualifier_ids.entries[0]));
1613                                 if (entries == NULL) {
1614                                         err = -ENOMEM;
1615                                         fputs("\nError:\t Not enough memory for parsing\n", trace->output);
1616                                         goto out_free;
1617                                 }
1618                                 trace->ev_qualifier_ids.entries = entries;
1619                         }
1620                         trace->ev_qualifier_ids.entries[nr_used++] = id;
1621                 }
1622         }
1623
1624         trace->ev_qualifier_ids.nr = nr_used;
1625         qsort(trace->ev_qualifier_ids.entries, nr_used, sizeof(int), intcmp);
1626 out:
1627         if (printed_invalid_prefix)
1628                 pr_debug("\n");
1629         return err;
1630 out_free:
1631         zfree(&trace->ev_qualifier_ids.entries);
1632         trace->ev_qualifier_ids.nr = 0;
1633         goto out;
1634 }
1635
1636 static __maybe_unused bool trace__syscall_enabled(struct trace *trace, int id)
1637 {
1638         bool in_ev_qualifier;
1639
1640         if (trace->ev_qualifier_ids.nr == 0)
1641                 return true;
1642
1643         in_ev_qualifier = bsearch(&id, trace->ev_qualifier_ids.entries,
1644                                   trace->ev_qualifier_ids.nr, sizeof(int), intcmp) != NULL;
1645
1646         if (in_ev_qualifier)
1647                return !trace->not_ev_qualifier;
1648
1649         return trace->not_ev_qualifier;
1650 }
1651
1652 /*
1653  * args is to be interpreted as a series of longs but we need to handle
1654  * 8-byte unaligned accesses. args points to raw_data within the event
1655  * and raw_data is guaranteed to be 8-byte unaligned because it is
1656  * preceded by raw_size which is a u32. So we need to copy args to a temp
1657  * variable to read it. Most notably this avoids extended load instructions
1658  * on unaligned addresses
1659  */
1660 unsigned long syscall_arg__val(struct syscall_arg *arg, u8 idx)
1661 {
1662         unsigned long val;
1663         unsigned char *p = arg->args + sizeof(unsigned long) * idx;
1664
1665         memcpy(&val, p, sizeof(val));
1666         return val;
1667 }
1668
1669 static size_t syscall__scnprintf_name(struct syscall *sc, char *bf, size_t size,
1670                                       struct syscall_arg *arg)
1671 {
1672         if (sc->arg_fmt && sc->arg_fmt[arg->idx].name)
1673                 return scnprintf(bf, size, "%s: ", sc->arg_fmt[arg->idx].name);
1674
1675         return scnprintf(bf, size, "arg%d: ", arg->idx);
1676 }
1677
1678 /*
1679  * Check if the value is in fact zero, i.e. mask whatever needs masking, such
1680  * as mount 'flags' argument that needs ignoring some magic flag, see comment
1681  * in tools/perf/trace/beauty/mount_flags.c
1682  */
1683 static unsigned long syscall__mask_val(struct syscall *sc, struct syscall_arg *arg, unsigned long val)
1684 {
1685         if (sc->arg_fmt && sc->arg_fmt[arg->idx].mask_val)
1686                 return sc->arg_fmt[arg->idx].mask_val(arg, val);
1687
1688         return val;
1689 }
1690
1691 static size_t syscall__scnprintf_val(struct syscall *sc, char *bf, size_t size,
1692                                      struct syscall_arg *arg, unsigned long val)
1693 {
1694         if (sc->arg_fmt && sc->arg_fmt[arg->idx].scnprintf) {
1695                 arg->val = val;
1696                 if (sc->arg_fmt[arg->idx].parm)
1697                         arg->parm = sc->arg_fmt[arg->idx].parm;
1698                 return sc->arg_fmt[arg->idx].scnprintf(bf, size, arg);
1699         }
1700         return scnprintf(bf, size, "%ld", val);
1701 }
1702
1703 static size_t syscall__scnprintf_args(struct syscall *sc, char *bf, size_t size,
1704                                       unsigned char *args, void *augmented_args, int augmented_args_size,
1705                                       struct trace *trace, struct thread *thread)
1706 {
1707         size_t printed = 0;
1708         unsigned long val;
1709         u8 bit = 1;
1710         struct syscall_arg arg = {
1711                 .args   = args,
1712                 .augmented = {
1713                         .size = augmented_args_size,
1714                         .args = augmented_args,
1715                 },
1716                 .idx    = 0,
1717                 .mask   = 0,
1718                 .trace  = trace,
1719                 .thread = thread,
1720                 .show_string_prefix = trace->show_string_prefix,
1721         };
1722         struct thread_trace *ttrace = thread__priv(thread);
1723
1724         /*
1725          * Things like fcntl will set this in its 'cmd' formatter to pick the
1726          * right formatter for the return value (an fd? file flags?), which is
1727          * not needed for syscalls that always return a given type, say an fd.
1728          */
1729         ttrace->ret_scnprintf = NULL;
1730
1731         if (sc->args != NULL) {
1732                 struct tep_format_field *field;
1733
1734                 for (field = sc->args; field;
1735                      field = field->next, ++arg.idx, bit <<= 1) {
1736                         if (arg.mask & bit)
1737                                 continue;
1738
1739                         val = syscall_arg__val(&arg, arg.idx);
1740                         /*
1741                          * Some syscall args need some mask, most don't and
1742                          * return val untouched.
1743                          */
1744                         val = syscall__mask_val(sc, &arg, val);
1745
1746                         /*
1747                          * Suppress this argument if its value is zero and
1748                          * and we don't have a string associated in an
1749                          * strarray for it.
1750                          */
1751                         if (val == 0 &&
1752                             !trace->show_zeros &&
1753                             !(sc->arg_fmt &&
1754                               (sc->arg_fmt[arg.idx].show_zero ||
1755                                sc->arg_fmt[arg.idx].scnprintf == SCA_STRARRAY ||
1756                                sc->arg_fmt[arg.idx].scnprintf == SCA_STRARRAYS) &&
1757                               sc->arg_fmt[arg.idx].parm))
1758                                 continue;
1759
1760                         printed += scnprintf(bf + printed, size - printed, "%s", printed ? ", " : "");
1761
1762                         if (trace->show_arg_names)
1763                                 printed += scnprintf(bf + printed, size - printed, "%s: ", field->name);
1764
1765                         printed += syscall__scnprintf_val(sc, bf + printed, size - printed, &arg, val);
1766                 }
1767         } else if (IS_ERR(sc->tp_format)) {
1768                 /*
1769                  * If we managed to read the tracepoint /format file, then we
1770                  * may end up not having any args, like with gettid(), so only
1771                  * print the raw args when we didn't manage to read it.
1772                  */
1773                 while (arg.idx < sc->nr_args) {
1774                         if (arg.mask & bit)
1775                                 goto next_arg;
1776                         val = syscall_arg__val(&arg, arg.idx);
1777                         if (printed)
1778                                 printed += scnprintf(bf + printed, size - printed, ", ");
1779                         printed += syscall__scnprintf_name(sc, bf + printed, size - printed, &arg);
1780                         printed += syscall__scnprintf_val(sc, bf + printed, size - printed, &arg, val);
1781 next_arg:
1782                         ++arg.idx;
1783                         bit <<= 1;
1784                 }
1785         }
1786
1787         return printed;
1788 }
1789
1790 typedef int (*tracepoint_handler)(struct trace *trace, struct evsel *evsel,
1791                                   union perf_event *event,
1792                                   struct perf_sample *sample);
1793
1794 static struct syscall *trace__syscall_info(struct trace *trace,
1795                                            struct evsel *evsel, int id)
1796 {
1797         int err = 0;
1798
1799         if (id < 0) {
1800
1801                 /*
1802                  * XXX: Noticed on x86_64, reproduced as far back as 3.0.36, haven't tried
1803                  * before that, leaving at a higher verbosity level till that is
1804                  * explained. Reproduced with plain ftrace with:
1805                  *
1806                  * echo 1 > /t/events/raw_syscalls/sys_exit/enable
1807                  * grep "NR -1 " /t/trace_pipe
1808                  *
1809                  * After generating some load on the machine.
1810                  */
1811                 if (verbose > 1) {
1812                         static u64 n;
1813                         fprintf(trace->output, "Invalid syscall %d id, skipping (%s, %" PRIu64 ") ...\n",
1814                                 id, perf_evsel__name(evsel), ++n);
1815                 }
1816                 return NULL;
1817         }
1818
1819         err = -EINVAL;
1820
1821         if (id > trace->sctbl->syscalls.max_id)
1822                 goto out_cant_read;
1823
1824         if ((trace->syscalls.table == NULL || trace->syscalls.table[id].name == NULL) &&
1825             (err = trace__read_syscall_info(trace, id)) != 0)
1826                 goto out_cant_read;
1827
1828         if (trace->syscalls.table[id].name == NULL) {
1829                 if (trace->syscalls.table[id].nonexistent)
1830                         return NULL;
1831                 goto out_cant_read;
1832         }
1833
1834         return &trace->syscalls.table[id];
1835
1836 out_cant_read:
1837         if (verbose > 0) {
1838                 char sbuf[STRERR_BUFSIZE];
1839                 fprintf(trace->output, "Problems reading syscall %d: %d (%s)", id, -err, str_error_r(-err, sbuf, sizeof(sbuf)));
1840                 if (id <= trace->sctbl->syscalls.max_id && trace->syscalls.table[id].name != NULL)
1841                         fprintf(trace->output, "(%s)", trace->syscalls.table[id].name);
1842                 fputs(" information\n", trace->output);
1843         }
1844         return NULL;
1845 }
1846
1847 static void thread__update_stats(struct thread_trace *ttrace,
1848                                  int id, struct perf_sample *sample)
1849 {
1850         struct int_node *inode;
1851         struct stats *stats;
1852         u64 duration = 0;
1853
1854         inode = intlist__findnew(ttrace->syscall_stats, id);
1855         if (inode == NULL)
1856                 return;
1857
1858         stats = inode->priv;
1859         if (stats == NULL) {
1860                 stats = malloc(sizeof(struct stats));
1861                 if (stats == NULL)
1862                         return;
1863                 init_stats(stats);
1864                 inode->priv = stats;
1865         }
1866
1867         if (ttrace->entry_time && sample->time > ttrace->entry_time)
1868                 duration = sample->time - ttrace->entry_time;
1869
1870         update_stats(stats, duration);
1871 }
1872
1873 static int trace__printf_interrupted_entry(struct trace *trace)
1874 {
1875         struct thread_trace *ttrace;
1876         size_t printed;
1877         int len;
1878
1879         if (trace->failure_only || trace->current == NULL)
1880                 return 0;
1881
1882         ttrace = thread__priv(trace->current);
1883
1884         if (!ttrace->entry_pending)
1885                 return 0;
1886
1887         printed  = trace__fprintf_entry_head(trace, trace->current, 0, false, ttrace->entry_time, trace->output);
1888         printed += len = fprintf(trace->output, "%s)", ttrace->entry_str);
1889
1890         if (len < trace->args_alignment - 4)
1891                 printed += fprintf(trace->output, "%-*s", trace->args_alignment - 4 - len, " ");
1892
1893         printed += fprintf(trace->output, " ...\n");
1894
1895         ttrace->entry_pending = false;
1896         ++trace->nr_events_printed;
1897
1898         return printed;
1899 }
1900
1901 static int trace__fprintf_sample(struct trace *trace, struct evsel *evsel,
1902                                  struct perf_sample *sample, struct thread *thread)
1903 {
1904         int printed = 0;
1905
1906         if (trace->print_sample) {
1907                 double ts = (double)sample->time / NSEC_PER_MSEC;
1908
1909                 printed += fprintf(trace->output, "%22s %10.3f %s %d/%d [%d]\n",
1910                                    perf_evsel__name(evsel), ts,
1911                                    thread__comm_str(thread),
1912                                    sample->pid, sample->tid, sample->cpu);
1913         }
1914
1915         return printed;
1916 }
1917
1918 static void *syscall__augmented_args(struct syscall *sc, struct perf_sample *sample, int *augmented_args_size, int raw_augmented_args_size)
1919 {
1920         void *augmented_args = NULL;
1921         /*
1922          * For now with BPF raw_augmented we hook into raw_syscalls:sys_enter
1923          * and there we get all 6 syscall args plus the tracepoint common fields
1924          * that gets calculated at the start and the syscall_nr (another long).
1925          * So we check if that is the case and if so don't look after the
1926          * sc->args_size but always after the full raw_syscalls:sys_enter payload,
1927          * which is fixed.
1928          *
1929          * We'll revisit this later to pass s->args_size to the BPF augmenter
1930          * (now tools/perf/examples/bpf/augmented_raw_syscalls.c, so that it
1931          * copies only what we need for each syscall, like what happens when we
1932          * use syscalls:sys_enter_NAME, so that we reduce the kernel/userspace
1933          * traffic to just what is needed for each syscall.
1934          */
1935         int args_size = raw_augmented_args_size ?: sc->args_size;
1936
1937         *augmented_args_size = sample->raw_size - args_size;
1938         if (*augmented_args_size > 0)
1939                 augmented_args = sample->raw_data + args_size;
1940
1941         return augmented_args;
1942 }
1943
1944 static int trace__sys_enter(struct trace *trace, struct evsel *evsel,
1945                             union perf_event *event __maybe_unused,
1946                             struct perf_sample *sample)
1947 {
1948         char *msg;
1949         void *args;
1950         int printed = 0;
1951         struct thread *thread;
1952         int id = perf_evsel__sc_tp_uint(evsel, id, sample), err = -1;
1953         int augmented_args_size = 0;
1954         void *augmented_args = NULL;
1955         struct syscall *sc = trace__syscall_info(trace, evsel, id);
1956         struct thread_trace *ttrace;
1957
1958         if (sc == NULL)
1959                 return -1;
1960
1961         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
1962         ttrace = thread__trace(thread, trace->output);
1963         if (ttrace == NULL)
1964                 goto out_put;
1965
1966         trace__fprintf_sample(trace, evsel, sample, thread);
1967
1968         args = perf_evsel__sc_tp_ptr(evsel, args, sample);
1969
1970         if (ttrace->entry_str == NULL) {
1971                 ttrace->entry_str = malloc(trace__entry_str_size);
1972                 if (!ttrace->entry_str)
1973                         goto out_put;
1974         }
1975
1976         if (!(trace->duration_filter || trace->summary_only || trace->min_stack))
1977                 trace__printf_interrupted_entry(trace);
1978         /*
1979          * If this is raw_syscalls.sys_enter, then it always comes with the 6 possible
1980          * arguments, even if the syscall being handled, say "openat", uses only 4 arguments
1981          * this breaks syscall__augmented_args() check for augmented args, as we calculate
1982          * syscall->args_size using each syscalls:sys_enter_NAME tracefs format file,
1983          * so when handling, say the openat syscall, we end up getting 6 args for the
1984          * raw_syscalls:sys_enter event, when we expected just 4, we end up mistakenly
1985          * thinking that the extra 2 u64 args are the augmented filename, so just check
1986          * here and avoid using augmented syscalls when the evsel is the raw_syscalls one.
1987          */
1988         if (evsel != trace->syscalls.events.sys_enter)
1989                 augmented_args = syscall__augmented_args(sc, sample, &augmented_args_size, trace->raw_augmented_syscalls_args_size);
1990         ttrace->entry_time = sample->time;
1991         msg = ttrace->entry_str;
1992         printed += scnprintf(msg + printed, trace__entry_str_size - printed, "%s(", sc->name);
1993
1994         printed += syscall__scnprintf_args(sc, msg + printed, trace__entry_str_size - printed,
1995                                            args, augmented_args, augmented_args_size, trace, thread);
1996
1997         if (sc->is_exit) {
1998                 if (!(trace->duration_filter || trace->summary_only || trace->failure_only || trace->min_stack)) {
1999                         int alignment = 0;
2000
2001                         trace__fprintf_entry_head(trace, thread, 0, false, ttrace->entry_time, trace->output);
2002                         printed = fprintf(trace->output, "%s)", ttrace->entry_str);
2003                         if (trace->args_alignment > printed)
2004                                 alignment = trace->args_alignment - printed;
2005                         fprintf(trace->output, "%*s= ?\n", alignment, " ");
2006                 }
2007         } else {
2008                 ttrace->entry_pending = true;
2009                 /* See trace__vfs_getname & trace__sys_exit */
2010                 ttrace->filename.pending_open = false;
2011         }
2012
2013         if (trace->current != thread) {
2014                 thread__put(trace->current);
2015                 trace->current = thread__get(thread);
2016         }
2017         err = 0;
2018 out_put:
2019         thread__put(thread);
2020         return err;
2021 }
2022
2023 static int trace__fprintf_sys_enter(struct trace *trace, struct evsel *evsel,
2024                                     struct perf_sample *sample)
2025 {
2026         struct thread_trace *ttrace;
2027         struct thread *thread;
2028         int id = perf_evsel__sc_tp_uint(evsel, id, sample), err = -1;
2029         struct syscall *sc = trace__syscall_info(trace, evsel, id);
2030         char msg[1024];
2031         void *args, *augmented_args = NULL;
2032         int augmented_args_size;
2033
2034         if (sc == NULL)
2035                 return -1;
2036
2037         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2038         ttrace = thread__trace(thread, trace->output);
2039         /*
2040          * We need to get ttrace just to make sure it is there when syscall__scnprintf_args()
2041          * and the rest of the beautifiers accessing it via struct syscall_arg touches it.
2042          */
2043         if (ttrace == NULL)
2044                 goto out_put;
2045
2046         args = perf_evsel__sc_tp_ptr(evsel, args, sample);
2047         augmented_args = syscall__augmented_args(sc, sample, &augmented_args_size, trace->raw_augmented_syscalls_args_size);
2048         syscall__scnprintf_args(sc, msg, sizeof(msg), args, augmented_args, augmented_args_size, trace, thread);
2049         fprintf(trace->output, "%s", msg);
2050         err = 0;
2051 out_put:
2052         thread__put(thread);
2053         return err;
2054 }
2055
2056 static int trace__resolve_callchain(struct trace *trace, struct evsel *evsel,
2057                                     struct perf_sample *sample,
2058                                     struct callchain_cursor *cursor)
2059 {
2060         struct addr_location al;
2061         int max_stack = evsel->core.attr.sample_max_stack ?
2062                         evsel->core.attr.sample_max_stack :
2063                         trace->max_stack;
2064         int err;
2065
2066         if (machine__resolve(trace->host, &al, sample) < 0)
2067                 return -1;
2068
2069         err = thread__resolve_callchain(al.thread, cursor, evsel, sample, NULL, NULL, max_stack);
2070         addr_location__put(&al);
2071         return err;
2072 }
2073
2074 static int trace__fprintf_callchain(struct trace *trace, struct perf_sample *sample)
2075 {
2076         /* TODO: user-configurable print_opts */
2077         const unsigned int print_opts = EVSEL__PRINT_SYM |
2078                                         EVSEL__PRINT_DSO |
2079                                         EVSEL__PRINT_UNKNOWN_AS_ADDR;
2080
2081         return sample__fprintf_callchain(sample, 38, print_opts, &callchain_cursor, symbol_conf.bt_stop_list, trace->output);
2082 }
2083
2084 static const char *errno_to_name(struct evsel *evsel, int err)
2085 {
2086         struct perf_env *env = perf_evsel__env(evsel);
2087         const char *arch_name = perf_env__arch(env);
2088
2089         return arch_syscalls__strerrno(arch_name, err);
2090 }
2091
2092 static int trace__sys_exit(struct trace *trace, struct evsel *evsel,
2093                            union perf_event *event __maybe_unused,
2094                            struct perf_sample *sample)
2095 {
2096         long ret;
2097         u64 duration = 0;
2098         bool duration_calculated = false;
2099         struct thread *thread;
2100         int id = perf_evsel__sc_tp_uint(evsel, id, sample), err = -1, callchain_ret = 0, printed = 0;
2101         int alignment = trace->args_alignment;
2102         struct syscall *sc = trace__syscall_info(trace, evsel, id);
2103         struct thread_trace *ttrace;
2104
2105         if (sc == NULL)
2106                 return -1;
2107
2108         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2109         ttrace = thread__trace(thread, trace->output);
2110         if (ttrace == NULL)
2111                 goto out_put;
2112
2113         trace__fprintf_sample(trace, evsel, sample, thread);
2114
2115         if (trace->summary)
2116                 thread__update_stats(ttrace, id, sample);
2117
2118         ret = perf_evsel__sc_tp_uint(evsel, ret, sample);
2119
2120         if (!trace->fd_path_disabled && sc->is_open && ret >= 0 && ttrace->filename.pending_open) {
2121                 trace__set_fd_pathname(thread, ret, ttrace->filename.name);
2122                 ttrace->filename.pending_open = false;
2123                 ++trace->stats.vfs_getname;
2124         }
2125
2126         if (ttrace->entry_time) {
2127                 duration = sample->time - ttrace->entry_time;
2128                 if (trace__filter_duration(trace, duration))
2129                         goto out;
2130                 duration_calculated = true;
2131         } else if (trace->duration_filter)
2132                 goto out;
2133
2134         if (sample->callchain) {
2135                 callchain_ret = trace__resolve_callchain(trace, evsel, sample, &callchain_cursor);
2136                 if (callchain_ret == 0) {
2137                         if (callchain_cursor.nr < trace->min_stack)
2138                                 goto out;
2139                         callchain_ret = 1;
2140                 }
2141         }
2142
2143         if (trace->summary_only || (ret >= 0 && trace->failure_only))
2144                 goto out;
2145
2146         trace__fprintf_entry_head(trace, thread, duration, duration_calculated, ttrace->entry_time, trace->output);
2147
2148         if (ttrace->entry_pending) {
2149                 printed = fprintf(trace->output, "%s", ttrace->entry_str);
2150         } else {
2151                 printed += fprintf(trace->output, " ... [");
2152                 color_fprintf(trace->output, PERF_COLOR_YELLOW, "continued");
2153                 printed += 9;
2154                 printed += fprintf(trace->output, "]: %s()", sc->name);
2155         }
2156
2157         printed++; /* the closing ')' */
2158
2159         if (alignment > printed)
2160                 alignment -= printed;
2161         else
2162                 alignment = 0;
2163
2164         fprintf(trace->output, ")%*s= ", alignment, " ");
2165
2166         if (sc->fmt == NULL) {
2167                 if (ret < 0)
2168                         goto errno_print;
2169 signed_print:
2170                 fprintf(trace->output, "%ld", ret);
2171         } else if (ret < 0) {
2172 errno_print: {
2173                 char bf[STRERR_BUFSIZE];
2174                 const char *emsg = str_error_r(-ret, bf, sizeof(bf)),
2175                            *e = errno_to_name(evsel, -ret);
2176
2177                 fprintf(trace->output, "-1 %s (%s)", e, emsg);
2178         }
2179         } else if (ret == 0 && sc->fmt->timeout)
2180                 fprintf(trace->output, "0 (Timeout)");
2181         else if (ttrace->ret_scnprintf) {
2182                 char bf[1024];
2183                 struct syscall_arg arg = {
2184                         .val    = ret,
2185                         .thread = thread,
2186                         .trace  = trace,
2187                 };
2188                 ttrace->ret_scnprintf(bf, sizeof(bf), &arg);
2189                 ttrace->ret_scnprintf = NULL;
2190                 fprintf(trace->output, "%s", bf);
2191         } else if (sc->fmt->hexret)
2192                 fprintf(trace->output, "%#lx", ret);
2193         else if (sc->fmt->errpid) {
2194                 struct thread *child = machine__find_thread(trace->host, ret, ret);
2195
2196                 if (child != NULL) {
2197                         fprintf(trace->output, "%ld", ret);
2198                         if (child->comm_set)
2199                                 fprintf(trace->output, " (%s)", thread__comm_str(child));
2200                         thread__put(child);
2201                 }
2202         } else
2203                 goto signed_print;
2204
2205         fputc('\n', trace->output);
2206
2207         /*
2208          * We only consider an 'event' for the sake of --max-events a non-filtered
2209          * sys_enter + sys_exit and other tracepoint events.
2210          */
2211         if (++trace->nr_events_printed == trace->max_events && trace->max_events != ULONG_MAX)
2212                 interrupted = true;
2213
2214         if (callchain_ret > 0)
2215                 trace__fprintf_callchain(trace, sample);
2216         else if (callchain_ret < 0)
2217                 pr_err("Problem processing %s callchain, skipping...\n", perf_evsel__name(evsel));
2218 out:
2219         ttrace->entry_pending = false;
2220         err = 0;
2221 out_put:
2222         thread__put(thread);
2223         return err;
2224 }
2225
2226 static int trace__vfs_getname(struct trace *trace, struct evsel *evsel,
2227                               union perf_event *event __maybe_unused,
2228                               struct perf_sample *sample)
2229 {
2230         struct thread *thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2231         struct thread_trace *ttrace;
2232         size_t filename_len, entry_str_len, to_move;
2233         ssize_t remaining_space;
2234         char *pos;
2235         const char *filename = perf_evsel__rawptr(evsel, sample, "pathname");
2236
2237         if (!thread)
2238                 goto out;
2239
2240         ttrace = thread__priv(thread);
2241         if (!ttrace)
2242                 goto out_put;
2243
2244         filename_len = strlen(filename);
2245         if (filename_len == 0)
2246                 goto out_put;
2247
2248         if (ttrace->filename.namelen < filename_len) {
2249                 char *f = realloc(ttrace->filename.name, filename_len + 1);
2250
2251                 if (f == NULL)
2252                         goto out_put;
2253
2254                 ttrace->filename.namelen = filename_len;
2255                 ttrace->filename.name = f;
2256         }
2257
2258         strcpy(ttrace->filename.name, filename);
2259         ttrace->filename.pending_open = true;
2260
2261         if (!ttrace->filename.ptr)
2262                 goto out_put;
2263
2264         entry_str_len = strlen(ttrace->entry_str);
2265         remaining_space = trace__entry_str_size - entry_str_len - 1; /* \0 */
2266         if (remaining_space <= 0)
2267                 goto out_put;
2268
2269         if (filename_len > (size_t)remaining_space) {
2270                 filename += filename_len - remaining_space;
2271                 filename_len = remaining_space;
2272         }
2273
2274         to_move = entry_str_len - ttrace->filename.entry_str_pos + 1; /* \0 */
2275         pos = ttrace->entry_str + ttrace->filename.entry_str_pos;
2276         memmove(pos + filename_len, pos, to_move);
2277         memcpy(pos, filename, filename_len);
2278
2279         ttrace->filename.ptr = 0;
2280         ttrace->filename.entry_str_pos = 0;
2281 out_put:
2282         thread__put(thread);
2283 out:
2284         return 0;
2285 }
2286
2287 static int trace__sched_stat_runtime(struct trace *trace, struct evsel *evsel,
2288                                      union perf_event *event __maybe_unused,
2289                                      struct perf_sample *sample)
2290 {
2291         u64 runtime = perf_evsel__intval(evsel, sample, "runtime");
2292         double runtime_ms = (double)runtime / NSEC_PER_MSEC;
2293         struct thread *thread = machine__findnew_thread(trace->host,
2294                                                         sample->pid,
2295                                                         sample->tid);
2296         struct thread_trace *ttrace = thread__trace(thread, trace->output);
2297
2298         if (ttrace == NULL)
2299                 goto out_dump;
2300
2301         ttrace->runtime_ms += runtime_ms;
2302         trace->runtime_ms += runtime_ms;
2303 out_put:
2304         thread__put(thread);
2305         return 0;
2306
2307 out_dump:
2308         fprintf(trace->output, "%s: comm=%s,pid=%u,runtime=%" PRIu64 ",vruntime=%" PRIu64 ")\n",
2309                evsel->name,
2310                perf_evsel__strval(evsel, sample, "comm"),
2311                (pid_t)perf_evsel__intval(evsel, sample, "pid"),
2312                runtime,
2313                perf_evsel__intval(evsel, sample, "vruntime"));
2314         goto out_put;
2315 }
2316
2317 static int bpf_output__printer(enum binary_printer_ops op,
2318                                unsigned int val, void *extra __maybe_unused, FILE *fp)
2319 {
2320         unsigned char ch = (unsigned char)val;
2321
2322         switch (op) {
2323         case BINARY_PRINT_CHAR_DATA:
2324                 return fprintf(fp, "%c", isprint(ch) ? ch : '.');
2325         case BINARY_PRINT_DATA_BEGIN:
2326         case BINARY_PRINT_LINE_BEGIN:
2327         case BINARY_PRINT_ADDR:
2328         case BINARY_PRINT_NUM_DATA:
2329         case BINARY_PRINT_NUM_PAD:
2330         case BINARY_PRINT_SEP:
2331         case BINARY_PRINT_CHAR_PAD:
2332         case BINARY_PRINT_LINE_END:
2333         case BINARY_PRINT_DATA_END:
2334         default:
2335                 break;
2336         }
2337
2338         return 0;
2339 }
2340
2341 static void bpf_output__fprintf(struct trace *trace,
2342                                 struct perf_sample *sample)
2343 {
2344         binary__fprintf(sample->raw_data, sample->raw_size, 8,
2345                         bpf_output__printer, NULL, trace->output);
2346         ++trace->nr_events_printed;
2347 }
2348
2349 static int trace__event_handler(struct trace *trace, struct evsel *evsel,
2350                                 union perf_event *event __maybe_unused,
2351                                 struct perf_sample *sample)
2352 {
2353         struct thread *thread;
2354         int callchain_ret = 0;
2355         /*
2356          * Check if we called perf_evsel__disable(evsel) due to, for instance,
2357          * this event's max_events having been hit and this is an entry coming
2358          * from the ring buffer that we should discard, since the max events
2359          * have already been considered/printed.
2360          */
2361         if (evsel->disabled)
2362                 return 0;
2363
2364         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2365
2366         if (sample->callchain) {
2367                 callchain_ret = trace__resolve_callchain(trace, evsel, sample, &callchain_cursor);
2368                 if (callchain_ret == 0) {
2369                         if (callchain_cursor.nr < trace->min_stack)
2370                                 goto out;
2371                         callchain_ret = 1;
2372                 }
2373         }
2374
2375         trace__printf_interrupted_entry(trace);
2376         trace__fprintf_tstamp(trace, sample->time, trace->output);
2377
2378         if (trace->trace_syscalls && trace->show_duration)
2379                 fprintf(trace->output, "(         ): ");
2380
2381         if (thread)
2382                 trace__fprintf_comm_tid(trace, thread, trace->output);
2383
2384         if (evsel == trace->syscalls.events.augmented) {
2385                 int id = perf_evsel__sc_tp_uint(evsel, id, sample);
2386                 struct syscall *sc = trace__syscall_info(trace, evsel, id);
2387
2388                 if (sc) {
2389                         fprintf(trace->output, "%s(", sc->name);
2390                         trace__fprintf_sys_enter(trace, evsel, sample);
2391                         fputc(')', trace->output);
2392                         goto newline;
2393                 }
2394
2395                 /*
2396                  * XXX: Not having the associated syscall info or not finding/adding
2397                  *      the thread should never happen, but if it does...
2398                  *      fall thru and print it as a bpf_output event.
2399                  */
2400         }
2401
2402         fprintf(trace->output, "%s:", evsel->name);
2403
2404         if (perf_evsel__is_bpf_output(evsel)) {
2405                 bpf_output__fprintf(trace, sample);
2406         } else if (evsel->tp_format) {
2407                 if (strncmp(evsel->tp_format->name, "sys_enter_", 10) ||
2408                     trace__fprintf_sys_enter(trace, evsel, sample)) {
2409                         event_format__fprintf(evsel->tp_format, sample->cpu,
2410                                               sample->raw_data, sample->raw_size,
2411                                               trace->output);
2412                         ++trace->nr_events_printed;
2413
2414                         if (evsel->max_events != ULONG_MAX && ++evsel->nr_events_printed == evsel->max_events) {
2415                                 evsel__disable(evsel);
2416                                 evsel__close(evsel);
2417                         }
2418                 }
2419         }
2420
2421 newline:
2422         fprintf(trace->output, "\n");
2423
2424         if (callchain_ret > 0)
2425                 trace__fprintf_callchain(trace, sample);
2426         else if (callchain_ret < 0)
2427                 pr_err("Problem processing %s callchain, skipping...\n", perf_evsel__name(evsel));
2428 out:
2429         thread__put(thread);
2430         return 0;
2431 }
2432
2433 static void print_location(FILE *f, struct perf_sample *sample,
2434                            struct addr_location *al,
2435                            bool print_dso, bool print_sym)
2436 {
2437
2438         if ((verbose > 0 || print_dso) && al->map)
2439                 fprintf(f, "%s@", al->map->dso->long_name);
2440
2441         if ((verbose > 0 || print_sym) && al->sym)
2442                 fprintf(f, "%s+0x%" PRIx64, al->sym->name,
2443                         al->addr - al->sym->start);
2444         else if (al->map)
2445                 fprintf(f, "0x%" PRIx64, al->addr);
2446         else
2447                 fprintf(f, "0x%" PRIx64, sample->addr);
2448 }
2449
2450 static int trace__pgfault(struct trace *trace,
2451                           struct evsel *evsel,
2452                           union perf_event *event __maybe_unused,
2453                           struct perf_sample *sample)
2454 {
2455         struct thread *thread;
2456         struct addr_location al;
2457         char map_type = 'd';
2458         struct thread_trace *ttrace;
2459         int err = -1;
2460         int callchain_ret = 0;
2461
2462         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2463
2464         if (sample->callchain) {
2465                 callchain_ret = trace__resolve_callchain(trace, evsel, sample, &callchain_cursor);
2466                 if (callchain_ret == 0) {
2467                         if (callchain_cursor.nr < trace->min_stack)
2468                                 goto out_put;
2469                         callchain_ret = 1;
2470                 }
2471         }
2472
2473         ttrace = thread__trace(thread, trace->output);
2474         if (ttrace == NULL)
2475                 goto out_put;
2476
2477         if (evsel->core.attr.config == PERF_COUNT_SW_PAGE_FAULTS_MAJ)
2478                 ttrace->pfmaj++;
2479         else
2480                 ttrace->pfmin++;
2481
2482         if (trace->summary_only)
2483                 goto out;
2484
2485         thread__find_symbol(thread, sample->cpumode, sample->ip, &al);
2486
2487         trace__fprintf_entry_head(trace, thread, 0, true, sample->time, trace->output);
2488
2489         fprintf(trace->output, "%sfault [",
2490                 evsel->core.attr.config == PERF_COUNT_SW_PAGE_FAULTS_MAJ ?
2491                 "maj" : "min");
2492
2493         print_location(trace->output, sample, &al, false, true);
2494
2495         fprintf(trace->output, "] => ");
2496
2497         thread__find_symbol(thread, sample->cpumode, sample->addr, &al);
2498
2499         if (!al.map) {
2500                 thread__find_symbol(thread, sample->cpumode, sample->addr, &al);
2501
2502                 if (al.map)
2503                         map_type = 'x';
2504                 else
2505                         map_type = '?';
2506         }
2507
2508         print_location(trace->output, sample, &al, true, false);
2509
2510         fprintf(trace->output, " (%c%c)\n", map_type, al.level);
2511
2512         if (callchain_ret > 0)
2513                 trace__fprintf_callchain(trace, sample);
2514         else if (callchain_ret < 0)
2515                 pr_err("Problem processing %s callchain, skipping...\n", perf_evsel__name(evsel));
2516
2517         ++trace->nr_events_printed;
2518 out:
2519         err = 0;
2520 out_put:
2521         thread__put(thread);
2522         return err;
2523 }
2524
2525 static void trace__set_base_time(struct trace *trace,
2526                                  struct evsel *evsel,
2527                                  struct perf_sample *sample)
2528 {
2529         /*
2530          * BPF events were not setting PERF_SAMPLE_TIME, so be more robust
2531          * and don't use sample->time unconditionally, we may end up having
2532          * some other event in the future without PERF_SAMPLE_TIME for good
2533          * reason, i.e. we may not be interested in its timestamps, just in
2534          * it taking place, picking some piece of information when it
2535          * appears in our event stream (vfs_getname comes to mind).
2536          */
2537         if (trace->base_time == 0 && !trace->full_time &&
2538             (evsel->core.attr.sample_type & PERF_SAMPLE_TIME))
2539                 trace->base_time = sample->time;
2540 }
2541
2542 static int trace__process_sample(struct perf_tool *tool,
2543                                  union perf_event *event,
2544                                  struct perf_sample *sample,
2545                                  struct evsel *evsel,
2546                                  struct machine *machine __maybe_unused)
2547 {
2548         struct trace *trace = container_of(tool, struct trace, tool);
2549         struct thread *thread;
2550         int err = 0;
2551
2552         tracepoint_handler handler = evsel->handler;
2553
2554         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2555         if (thread && thread__is_filtered(thread))
2556                 goto out;
2557
2558         trace__set_base_time(trace, evsel, sample);
2559
2560         if (handler) {
2561                 ++trace->nr_events;
2562                 handler(trace, evsel, event, sample);
2563         }
2564 out:
2565         thread__put(thread);
2566         return err;
2567 }
2568
2569 static int trace__record(struct trace *trace, int argc, const char **argv)
2570 {
2571         unsigned int rec_argc, i, j;
2572         const char **rec_argv;
2573         const char * const record_args[] = {
2574                 "record",
2575                 "-R",
2576                 "-m", "1024",
2577                 "-c", "1",
2578         };
2579
2580         const char * const sc_args[] = { "-e", };
2581         unsigned int sc_args_nr = ARRAY_SIZE(sc_args);
2582         const char * const majpf_args[] = { "-e", "major-faults" };
2583         unsigned int majpf_args_nr = ARRAY_SIZE(majpf_args);
2584         const char * const minpf_args[] = { "-e", "minor-faults" };
2585         unsigned int minpf_args_nr = ARRAY_SIZE(minpf_args);
2586
2587         /* +1 is for the event string below */
2588         rec_argc = ARRAY_SIZE(record_args) + sc_args_nr + 1 +
2589                 majpf_args_nr + minpf_args_nr + argc;
2590         rec_argv = calloc(rec_argc + 1, sizeof(char *));
2591
2592         if (rec_argv == NULL)
2593                 return -ENOMEM;
2594
2595         j = 0;
2596         for (i = 0; i < ARRAY_SIZE(record_args); i++)
2597                 rec_argv[j++] = record_args[i];
2598
2599         if (trace->trace_syscalls) {
2600                 for (i = 0; i < sc_args_nr; i++)
2601                         rec_argv[j++] = sc_args[i];
2602
2603                 /* event string may be different for older kernels - e.g., RHEL6 */
2604                 if (is_valid_tracepoint("raw_syscalls:sys_enter"))
2605                         rec_argv[j++] = "raw_syscalls:sys_enter,raw_syscalls:sys_exit";
2606                 else if (is_valid_tracepoint("syscalls:sys_enter"))
2607                         rec_argv[j++] = "syscalls:sys_enter,syscalls:sys_exit";
2608                 else {
2609                         pr_err("Neither raw_syscalls nor syscalls events exist.\n");
2610                         free(rec_argv);
2611                         return -1;
2612                 }
2613         }
2614
2615         if (trace->trace_pgfaults & TRACE_PFMAJ)
2616                 for (i = 0; i < majpf_args_nr; i++)
2617                         rec_argv[j++] = majpf_args[i];
2618
2619         if (trace->trace_pgfaults & TRACE_PFMIN)
2620                 for (i = 0; i < minpf_args_nr; i++)
2621                         rec_argv[j++] = minpf_args[i];
2622
2623         for (i = 0; i < (unsigned int)argc; i++)
2624                 rec_argv[j++] = argv[i];
2625
2626         return cmd_record(j, rec_argv);
2627 }
2628
2629 static size_t trace__fprintf_thread_summary(struct trace *trace, FILE *fp);
2630
2631 static bool evlist__add_vfs_getname(struct evlist *evlist)
2632 {
2633         bool found = false;
2634         struct evsel *evsel, *tmp;
2635         struct parse_events_error err = { .idx = 0, };
2636         int ret = parse_events(evlist, "probe:vfs_getname*", &err);
2637
2638         if (ret)
2639                 return false;
2640
2641         evlist__for_each_entry_safe(evlist, evsel, tmp) {
2642                 if (!strstarts(perf_evsel__name(evsel), "probe:vfs_getname"))
2643                         continue;
2644
2645                 if (perf_evsel__field(evsel, "pathname")) {
2646                         evsel->handler = trace__vfs_getname;
2647                         found = true;
2648                         continue;
2649                 }
2650
2651                 list_del_init(&evsel->core.node);
2652                 evsel->evlist = NULL;
2653                 evsel__delete(evsel);
2654         }
2655
2656         return found;
2657 }
2658
2659 static struct evsel *perf_evsel__new_pgfault(u64 config)
2660 {
2661         struct evsel *evsel;
2662         struct perf_event_attr attr = {
2663                 .type = PERF_TYPE_SOFTWARE,
2664                 .mmap_data = 1,
2665         };
2666
2667         attr.config = config;
2668         attr.sample_period = 1;
2669
2670         event_attr_init(&attr);
2671
2672         evsel = evsel__new(&attr);
2673         if (evsel)
2674                 evsel->handler = trace__pgfault;
2675
2676         return evsel;
2677 }
2678
2679 static void trace__handle_event(struct trace *trace, union perf_event *event, struct perf_sample *sample)
2680 {
2681         const u32 type = event->header.type;
2682         struct evsel *evsel;
2683
2684         if (type != PERF_RECORD_SAMPLE) {
2685                 trace__process_event(trace, trace->host, event, sample);
2686                 return;
2687         }
2688
2689         evsel = perf_evlist__id2evsel(trace->evlist, sample->id);
2690         if (evsel == NULL) {
2691                 fprintf(trace->output, "Unknown tp ID %" PRIu64 ", skipping...\n", sample->id);
2692                 return;
2693         }
2694
2695         if (evswitch__discard(&trace->evswitch, evsel))
2696                 return;
2697
2698         trace__set_base_time(trace, evsel, sample);
2699
2700         if (evsel->core.attr.type == PERF_TYPE_TRACEPOINT &&
2701             sample->raw_data == NULL) {
2702                 fprintf(trace->output, "%s sample with no payload for tid: %d, cpu %d, raw_size=%d, skipping...\n",
2703                        perf_evsel__name(evsel), sample->tid,
2704                        sample->cpu, sample->raw_size);
2705         } else {
2706                 tracepoint_handler handler = evsel->handler;
2707                 handler(trace, evsel, event, sample);
2708         }
2709
2710         if (trace->nr_events_printed >= trace->max_events && trace->max_events != ULONG_MAX)
2711                 interrupted = true;
2712 }
2713
2714 static int trace__add_syscall_newtp(struct trace *trace)
2715 {
2716         int ret = -1;
2717         struct evlist *evlist = trace->evlist;
2718         struct evsel *sys_enter, *sys_exit;
2719
2720         sys_enter = perf_evsel__raw_syscall_newtp("sys_enter", trace__sys_enter);
2721         if (sys_enter == NULL)
2722                 goto out;
2723
2724         if (perf_evsel__init_sc_tp_ptr_field(sys_enter, args))
2725                 goto out_delete_sys_enter;
2726
2727         sys_exit = perf_evsel__raw_syscall_newtp("sys_exit", trace__sys_exit);
2728         if (sys_exit == NULL)
2729                 goto out_delete_sys_enter;
2730
2731         if (perf_evsel__init_sc_tp_uint_field(sys_exit, ret))
2732                 goto out_delete_sys_exit;
2733
2734         perf_evsel__config_callchain(sys_enter, &trace->opts, &callchain_param);
2735         perf_evsel__config_callchain(sys_exit, &trace->opts, &callchain_param);
2736
2737         evlist__add(evlist, sys_enter);
2738         evlist__add(evlist, sys_exit);
2739
2740         if (callchain_param.enabled && !trace->kernel_syscallchains) {
2741                 /*
2742                  * We're interested only in the user space callchain
2743                  * leading to the syscall, allow overriding that for
2744                  * debugging reasons using --kernel_syscall_callchains
2745                  */
2746                 sys_exit->core.attr.exclude_callchain_kernel = 1;
2747         }
2748
2749         trace->syscalls.events.sys_enter = sys_enter;
2750         trace->syscalls.events.sys_exit  = sys_exit;
2751
2752         ret = 0;
2753 out:
2754         return ret;
2755
2756 out_delete_sys_exit:
2757         evsel__delete_priv(sys_exit);
2758 out_delete_sys_enter:
2759         evsel__delete_priv(sys_enter);
2760         goto out;
2761 }
2762
2763 static int trace__set_ev_qualifier_tp_filter(struct trace *trace)
2764 {
2765         int err = -1;
2766         struct evsel *sys_exit;
2767         char *filter = asprintf_expr_inout_ints("id", !trace->not_ev_qualifier,
2768                                                 trace->ev_qualifier_ids.nr,
2769                                                 trace->ev_qualifier_ids.entries);
2770
2771         if (filter == NULL)
2772                 goto out_enomem;
2773
2774         if (!perf_evsel__append_tp_filter(trace->syscalls.events.sys_enter,
2775                                           filter)) {
2776                 sys_exit = trace->syscalls.events.sys_exit;
2777                 err = perf_evsel__append_tp_filter(sys_exit, filter);
2778         }
2779
2780         free(filter);
2781 out:
2782         return err;
2783 out_enomem:
2784         errno = ENOMEM;
2785         goto out;
2786 }
2787
2788 #ifdef HAVE_LIBBPF_SUPPORT
2789 static struct bpf_program *trace__find_bpf_program_by_title(struct trace *trace, const char *name)
2790 {
2791         if (trace->bpf_obj == NULL)
2792                 return NULL;
2793
2794         return bpf_object__find_program_by_title(trace->bpf_obj, name);
2795 }
2796
2797 static struct bpf_program *trace__find_syscall_bpf_prog(struct trace *trace, struct syscall *sc,
2798                                                         const char *prog_name, const char *type)
2799 {
2800         struct bpf_program *prog;
2801
2802         if (prog_name == NULL) {
2803                 char default_prog_name[256];
2804                 scnprintf(default_prog_name, sizeof(default_prog_name), "!syscalls:sys_%s_%s", type, sc->name);
2805                 prog = trace__find_bpf_program_by_title(trace, default_prog_name);
2806                 if (prog != NULL)
2807                         goto out_found;
2808                 if (sc->fmt && sc->fmt->alias) {
2809                         scnprintf(default_prog_name, sizeof(default_prog_name), "!syscalls:sys_%s_%s", type, sc->fmt->alias);
2810                         prog = trace__find_bpf_program_by_title(trace, default_prog_name);
2811                         if (prog != NULL)
2812                                 goto out_found;
2813                 }
2814                 goto out_unaugmented;
2815         }
2816
2817         prog = trace__find_bpf_program_by_title(trace, prog_name);
2818
2819         if (prog != NULL) {
2820 out_found:
2821                 return prog;
2822         }
2823
2824         pr_debug("Couldn't find BPF prog \"%s\" to associate with syscalls:sys_%s_%s, not augmenting it\n",
2825                  prog_name, type, sc->name);
2826 out_unaugmented:
2827         return trace->syscalls.unaugmented_prog;
2828 }
2829
2830 static void trace__init_syscall_bpf_progs(struct trace *trace, int id)
2831 {
2832         struct syscall *sc = trace__syscall_info(trace, NULL, id);
2833
2834         if (sc == NULL)
2835                 return;
2836
2837         sc->bpf_prog.sys_enter = trace__find_syscall_bpf_prog(trace, sc, sc->fmt ? sc->fmt->bpf_prog_name.sys_enter : NULL, "enter");
2838         sc->bpf_prog.sys_exit  = trace__find_syscall_bpf_prog(trace, sc, sc->fmt ? sc->fmt->bpf_prog_name.sys_exit  : NULL,  "exit");
2839 }
2840
2841 static int trace__bpf_prog_sys_enter_fd(struct trace *trace, int id)
2842 {
2843         struct syscall *sc = trace__syscall_info(trace, NULL, id);
2844         return sc ? bpf_program__fd(sc->bpf_prog.sys_enter) : bpf_program__fd(trace->syscalls.unaugmented_prog);
2845 }
2846
2847 static int trace__bpf_prog_sys_exit_fd(struct trace *trace, int id)
2848 {
2849         struct syscall *sc = trace__syscall_info(trace, NULL, id);
2850         return sc ? bpf_program__fd(sc->bpf_prog.sys_exit) : bpf_program__fd(trace->syscalls.unaugmented_prog);
2851 }
2852
2853 static void trace__init_bpf_map_syscall_args(struct trace *trace, int id, struct bpf_map_syscall_entry *entry)
2854 {
2855         struct syscall *sc = trace__syscall_info(trace, NULL, id);
2856         int arg = 0;
2857
2858         if (sc == NULL)
2859                 goto out;
2860
2861         for (; arg < sc->nr_args; ++arg) {
2862                 entry->string_args_len[arg] = 0;
2863                 if (sc->arg_fmt[arg].scnprintf == SCA_FILENAME) {
2864                         /* Should be set like strace -s strsize */
2865                         entry->string_args_len[arg] = PATH_MAX;
2866                 }
2867         }
2868 out:
2869         for (; arg < 6; ++arg)
2870                 entry->string_args_len[arg] = 0;
2871 }
2872 static int trace__set_ev_qualifier_bpf_filter(struct trace *trace)
2873 {
2874         int fd = bpf_map__fd(trace->syscalls.map);
2875         struct bpf_map_syscall_entry value = {
2876                 .enabled = !trace->not_ev_qualifier,
2877         };
2878         int err = 0;
2879         size_t i;
2880
2881         for (i = 0; i < trace->ev_qualifier_ids.nr; ++i) {
2882                 int key = trace->ev_qualifier_ids.entries[i];
2883
2884                 if (value.enabled) {
2885                         trace__init_bpf_map_syscall_args(trace, key, &value);
2886                         trace__init_syscall_bpf_progs(trace, key);
2887                 }
2888
2889                 err = bpf_map_update_elem(fd, &key, &value, BPF_EXIST);
2890                 if (err)
2891                         break;
2892         }
2893
2894         return err;
2895 }
2896
2897 static int __trace__init_syscalls_bpf_map(struct trace *trace, bool enabled)
2898 {
2899         int fd = bpf_map__fd(trace->syscalls.map);
2900         struct bpf_map_syscall_entry value = {
2901                 .enabled = enabled,
2902         };
2903         int err = 0, key;
2904
2905         for (key = 0; key < trace->sctbl->syscalls.nr_entries; ++key) {
2906                 if (enabled)
2907                         trace__init_bpf_map_syscall_args(trace, key, &value);
2908
2909                 err = bpf_map_update_elem(fd, &key, &value, BPF_ANY);
2910                 if (err)
2911                         break;
2912         }
2913
2914         return err;
2915 }
2916
2917 static int trace__init_syscalls_bpf_map(struct trace *trace)
2918 {
2919         bool enabled = true;
2920
2921         if (trace->ev_qualifier_ids.nr)
2922                 enabled = trace->not_ev_qualifier;
2923
2924         return __trace__init_syscalls_bpf_map(trace, enabled);
2925 }
2926
2927 static struct bpf_program *trace__find_usable_bpf_prog_entry(struct trace *trace, struct syscall *sc)
2928 {
2929         struct tep_format_field *field, *candidate_field;
2930         int id;
2931
2932         /*
2933          * We're only interested in syscalls that have a pointer:
2934          */
2935         for (field = sc->args; field; field = field->next) {
2936                 if (field->flags & TEP_FIELD_IS_POINTER)
2937                         goto try_to_find_pair;
2938         }
2939
2940         return NULL;
2941
2942 try_to_find_pair:
2943         for (id = 0; id < trace->sctbl->syscalls.nr_entries; ++id) {
2944                 struct syscall *pair = trace__syscall_info(trace, NULL, id);
2945                 struct bpf_program *pair_prog;
2946                 bool is_candidate = false;
2947
2948                 if (pair == NULL || pair == sc ||
2949                     pair->bpf_prog.sys_enter == trace->syscalls.unaugmented_prog)
2950                         continue;
2951
2952                 for (field = sc->args, candidate_field = pair->args;
2953                      field && candidate_field; field = field->next, candidate_field = candidate_field->next) {
2954                         bool is_pointer = field->flags & TEP_FIELD_IS_POINTER,
2955                              candidate_is_pointer = candidate_field->flags & TEP_FIELD_IS_POINTER;
2956
2957                         if (is_pointer) {
2958                                if (!candidate_is_pointer) {
2959                                         // The candidate just doesn't copies our pointer arg, might copy other pointers we want.
2960                                         continue;
2961                                }
2962                         } else {
2963                                 if (candidate_is_pointer) {
2964                                         // The candidate might copy a pointer we don't have, skip it.
2965                                         goto next_candidate;
2966                                 }
2967                                 continue;
2968                         }
2969
2970                         if (strcmp(field->type, candidate_field->type))
2971                                 goto next_candidate;
2972
2973                         is_candidate = true;
2974                 }
2975
2976                 if (!is_candidate)
2977                         goto next_candidate;
2978
2979                 /*
2980                  * Check if the tentative pair syscall augmenter has more pointers, if it has,
2981                  * then it may be collecting that and we then can't use it, as it would collect
2982                  * more than what is common to the two syscalls.
2983                  */
2984                 if (candidate_field) {
2985                         for (candidate_field = candidate_field->next; candidate_field; candidate_field = candidate_field->next)
2986                                 if (candidate_field->flags & TEP_FIELD_IS_POINTER)
2987                                         goto next_candidate;
2988                 }
2989
2990                 pair_prog = pair->bpf_prog.sys_enter;
2991                 /*
2992                  * If the pair isn't enabled, then its bpf_prog.sys_enter will not
2993                  * have been searched for, so search it here and if it returns the
2994                  * unaugmented one, then ignore it, otherwise we'll reuse that BPF
2995                  * program for a filtered syscall on a non-filtered one.
2996                  *
2997                  * For instance, we have "!syscalls:sys_enter_renameat" and that is
2998                  * useful for "renameat2".
2999                  */
3000                 if (pair_prog == NULL) {
3001                         pair_prog = trace__find_syscall_bpf_prog(trace, pair, pair->fmt ? pair->fmt->bpf_prog_name.sys_enter : NULL, "enter");
3002                         if (pair_prog == trace->syscalls.unaugmented_prog)
3003                                 goto next_candidate;
3004                 }
3005
3006                 pr_debug("Reusing \"%s\" BPF sys_enter augmenter for \"%s\"\n", pair->name, sc->name);
3007                 return pair_prog;
3008         next_candidate:
3009                 continue;
3010         }
3011
3012         return NULL;
3013 }
3014
3015 static int trace__init_syscalls_bpf_prog_array_maps(struct trace *trace)
3016 {
3017         int map_enter_fd = bpf_map__fd(trace->syscalls.prog_array.sys_enter),
3018             map_exit_fd  = bpf_map__fd(trace->syscalls.prog_array.sys_exit);
3019         int err = 0, key;
3020
3021         for (key = 0; key < trace->sctbl->syscalls.nr_entries; ++key) {
3022                 int prog_fd;
3023
3024                 if (!trace__syscall_enabled(trace, key))
3025                         continue;
3026
3027                 trace__init_syscall_bpf_progs(trace, key);
3028
3029                 // It'll get at least the "!raw_syscalls:unaugmented"
3030                 prog_fd = trace__bpf_prog_sys_enter_fd(trace, key);
3031                 err = bpf_map_update_elem(map_enter_fd, &key, &prog_fd, BPF_ANY);
3032                 if (err)
3033                         break;
3034                 prog_fd = trace__bpf_prog_sys_exit_fd(trace, key);
3035                 err = bpf_map_update_elem(map_exit_fd, &key, &prog_fd, BPF_ANY);
3036                 if (err)
3037                         break;
3038         }
3039
3040         /*
3041          * Now lets do a second pass looking for enabled syscalls without
3042          * an augmenter that have a signature that is a superset of another
3043          * syscall with an augmenter so that we can auto-reuse it.
3044          *
3045          * I.e. if we have an augmenter for the "open" syscall that has
3046          * this signature:
3047          *
3048          *   int open(const char *pathname, int flags, mode_t mode);
3049          *
3050          * I.e. that will collect just the first string argument, then we
3051          * can reuse it for the 'creat' syscall, that has this signature:
3052          *
3053          *   int creat(const char *pathname, mode_t mode);
3054          *
3055          * and for:
3056          *
3057          *   int stat(const char *pathname, struct stat *statbuf);
3058          *   int lstat(const char *pathname, struct stat *statbuf);
3059          *
3060          * Because the 'open' augmenter will collect the first arg as a string,
3061          * and leave alone all the other args, which already helps with
3062          * beautifying 'stat' and 'lstat''s pathname arg.
3063          *
3064          * Then, in time, when 'stat' gets an augmenter that collects both
3065          * first and second arg (this one on the raw_syscalls:sys_exit prog
3066          * array tail call, then that one will be used.
3067          */
3068         for (key = 0; key < trace->sctbl->syscalls.nr_entries; ++key) {
3069                 struct syscall *sc = trace__syscall_info(trace, NULL, key);
3070                 struct bpf_program *pair_prog;
3071                 int prog_fd;
3072
3073                 if (sc == NULL || sc->bpf_prog.sys_enter == NULL)
3074                         continue;
3075
3076                 /*
3077                  * For now we're just reusing the sys_enter prog, and if it
3078                  * already has an augmenter, we don't need to find one.
3079                  */
3080                 if (sc->bpf_prog.sys_enter != trace->syscalls.unaugmented_prog)
3081                         continue;
3082
3083                 /*
3084                  * Look at all the other syscalls for one that has a signature
3085                  * that is close enough that we can share:
3086                  */
3087                 pair_prog = trace__find_usable_bpf_prog_entry(trace, sc);
3088                 if (pair_prog == NULL)
3089                         continue;
3090
3091                 sc->bpf_prog.sys_enter = pair_prog;
3092
3093                 /*
3094                  * Update the BPF_MAP_TYPE_PROG_SHARED for raw_syscalls:sys_enter
3095                  * with the fd for the program we're reusing:
3096                  */
3097                 prog_fd = bpf_program__fd(sc->bpf_prog.sys_enter);
3098                 err = bpf_map_update_elem(map_enter_fd, &key, &prog_fd, BPF_ANY);
3099                 if (err)
3100                         break;
3101         }
3102
3103
3104         return err;
3105 }
3106 #else
3107 static int trace__set_ev_qualifier_bpf_filter(struct trace *trace __maybe_unused)
3108 {
3109         return 0;
3110 }
3111
3112 static int trace__init_syscalls_bpf_map(struct trace *trace __maybe_unused)
3113 {
3114         return 0;
3115 }
3116
3117 static struct bpf_program *trace__find_bpf_program_by_title(struct trace *trace __maybe_unused,
3118                                                             const char *name __maybe_unused)
3119 {
3120         return NULL;
3121 }
3122
3123 static int trace__init_syscalls_bpf_prog_array_maps(struct trace *trace __maybe_unused)
3124 {
3125         return 0;
3126 }
3127 #endif // HAVE_LIBBPF_SUPPORT
3128
3129 static int trace__set_ev_qualifier_filter(struct trace *trace)
3130 {
3131         if (trace->syscalls.map)
3132                 return trace__set_ev_qualifier_bpf_filter(trace);
3133         if (trace->syscalls.events.sys_enter)
3134                 return trace__set_ev_qualifier_tp_filter(trace);
3135         return 0;
3136 }
3137
3138 static int bpf_map__set_filter_pids(struct bpf_map *map __maybe_unused,
3139                                     size_t npids __maybe_unused, pid_t *pids __maybe_unused)
3140 {
3141         int err = 0;
3142 #ifdef HAVE_LIBBPF_SUPPORT
3143         bool value = true;
3144         int map_fd = bpf_map__fd(map);
3145         size_t i;
3146
3147         for (i = 0; i < npids; ++i) {
3148                 err = bpf_map_update_elem(map_fd, &pids[i], &value, BPF_ANY);
3149                 if (err)
3150                         break;
3151         }
3152 #endif
3153         return err;
3154 }
3155
3156 static int trace__set_filter_loop_pids(struct trace *trace)
3157 {
3158         unsigned int nr = 1, err;
3159         pid_t pids[32] = {
3160                 getpid(),
3161         };
3162         struct thread *thread = machine__find_thread(trace->host, pids[0], pids[0]);
3163
3164         while (thread && nr < ARRAY_SIZE(pids)) {
3165                 struct thread *parent = machine__find_thread(trace->host, thread->ppid, thread->ppid);
3166
3167                 if (parent == NULL)
3168                         break;
3169
3170                 if (!strcmp(thread__comm_str(parent), "sshd") ||
3171                     strstarts(thread__comm_str(parent), "gnome-terminal")) {
3172                         pids[nr++] = parent->tid;
3173                         break;
3174                 }
3175                 thread = parent;
3176         }
3177
3178         err = perf_evlist__set_tp_filter_pids(trace->evlist, nr, pids);
3179         if (!err && trace->filter_pids.map)
3180                 err = bpf_map__set_filter_pids(trace->filter_pids.map, nr, pids);
3181
3182         return err;
3183 }
3184
3185 static int trace__set_filter_pids(struct trace *trace)
3186 {
3187         int err = 0;
3188         /*
3189          * Better not use !target__has_task() here because we need to cover the
3190          * case where no threads were specified in the command line, but a
3191          * workload was, and in that case we will fill in the thread_map when
3192          * we fork the workload in perf_evlist__prepare_workload.
3193          */
3194         if (trace->filter_pids.nr > 0) {
3195                 err = perf_evlist__set_tp_filter_pids(trace->evlist, trace->filter_pids.nr,
3196                                                       trace->filter_pids.entries);
3197                 if (!err && trace->filter_pids.map) {
3198                         err = bpf_map__set_filter_pids(trace->filter_pids.map, trace->filter_pids.nr,
3199                                                        trace->filter_pids.entries);
3200                 }
3201         } else if (perf_thread_map__pid(trace->evlist->core.threads, 0) == -1) {
3202                 err = trace__set_filter_loop_pids(trace);
3203         }
3204
3205         return err;
3206 }
3207
3208 static int __trace__deliver_event(struct trace *trace, union perf_event *event)
3209 {
3210         struct evlist *evlist = trace->evlist;
3211         struct perf_sample sample;
3212         int err;
3213
3214         err = perf_evlist__parse_sample(evlist, event, &sample);
3215         if (err)
3216                 fprintf(trace->output, "Can't parse sample, err = %d, skipping...\n", err);
3217         else
3218                 trace__handle_event(trace, event, &sample);
3219
3220         return 0;
3221 }
3222
3223 static int __trace__flush_events(struct trace *trace)
3224 {
3225         u64 first = ordered_events__first_time(&trace->oe.data);
3226         u64 flush = trace->oe.last - NSEC_PER_SEC;
3227
3228         /* Is there some thing to flush.. */
3229         if (first && first < flush)
3230                 return ordered_events__flush_time(&trace->oe.data, flush);
3231
3232         return 0;
3233 }
3234
3235 static int trace__flush_events(struct trace *trace)
3236 {
3237         return !trace->sort_events ? 0 : __trace__flush_events(trace);
3238 }
3239
3240 static int trace__deliver_event(struct trace *trace, union perf_event *event)
3241 {
3242         int err;
3243
3244         if (!trace->sort_events)
3245                 return __trace__deliver_event(trace, event);
3246
3247         err = perf_evlist__parse_sample_timestamp(trace->evlist, event, &trace->oe.last);
3248         if (err && err != -1)
3249                 return err;
3250
3251         err = ordered_events__queue(&trace->oe.data, event, trace->oe.last, 0);
3252         if (err)
3253                 return err;
3254
3255         return trace__flush_events(trace);
3256 }
3257
3258 static int ordered_events__deliver_event(struct ordered_events *oe,
3259                                          struct ordered_event *event)
3260 {
3261         struct trace *trace = container_of(oe, struct trace, oe.data);
3262
3263         return __trace__deliver_event(trace, event->event);
3264 }
3265
3266 static int trace__run(struct trace *trace, int argc, const char **argv)
3267 {
3268         struct evlist *evlist = trace->evlist;
3269         struct evsel *evsel, *pgfault_maj = NULL, *pgfault_min = NULL;
3270         int err = -1, i;
3271         unsigned long before;
3272         const bool forks = argc > 0;
3273         bool draining = false;
3274
3275         trace->live = true;
3276
3277         if (!trace->raw_augmented_syscalls) {
3278                 if (trace->trace_syscalls && trace__add_syscall_newtp(trace))
3279                         goto out_error_raw_syscalls;
3280
3281                 if (trace->trace_syscalls)
3282                         trace->vfs_getname = evlist__add_vfs_getname(evlist);
3283         }
3284
3285         if ((trace->trace_pgfaults & TRACE_PFMAJ)) {
3286                 pgfault_maj = perf_evsel__new_pgfault(PERF_COUNT_SW_PAGE_FAULTS_MAJ);
3287                 if (pgfault_maj == NULL)
3288                         goto out_error_mem;
3289                 perf_evsel__config_callchain(pgfault_maj, &trace->opts, &callchain_param);
3290                 evlist__add(evlist, pgfault_maj);
3291         }
3292
3293         if ((trace->trace_pgfaults & TRACE_PFMIN)) {
3294                 pgfault_min = perf_evsel__new_pgfault(PERF_COUNT_SW_PAGE_FAULTS_MIN);
3295                 if (pgfault_min == NULL)
3296                         goto out_error_mem;
3297                 perf_evsel__config_callchain(pgfault_min, &trace->opts, &callchain_param);
3298                 evlist__add(evlist, pgfault_min);
3299         }
3300
3301         if (trace->sched &&
3302             perf_evlist__add_newtp(evlist, "sched", "sched_stat_runtime",
3303                                    trace__sched_stat_runtime))
3304                 goto out_error_sched_stat_runtime;
3305
3306         /*
3307          * If a global cgroup was set, apply it to all the events without an
3308          * explicit cgroup. I.e.:
3309          *
3310          *      trace -G A -e sched:*switch
3311          *
3312          * Will set all raw_syscalls:sys_{enter,exit}, pgfault, vfs_getname, etc
3313          * _and_ sched:sched_switch to the 'A' cgroup, while:
3314          *
3315          * trace -e sched:*switch -G A
3316          *
3317          * will only set the sched:sched_switch event to the 'A' cgroup, all the
3318          * other events (raw_syscalls:sys_{enter,exit}, etc are left "without"
3319          * a cgroup (on the root cgroup, sys wide, etc).
3320          *
3321          * Multiple cgroups:
3322          *
3323          * trace -G A -e sched:*switch -G B
3324          *
3325          * the syscall ones go to the 'A' cgroup, the sched:sched_switch goes
3326          * to the 'B' cgroup.
3327          *
3328          * evlist__set_default_cgroup() grabs a reference of the passed cgroup
3329          * only for the evsels still without a cgroup, i.e. evsel->cgroup == NULL.
3330          */
3331         if (trace->cgroup)
3332                 evlist__set_default_cgroup(trace->evlist, trace->cgroup);
3333
3334         err = perf_evlist__create_maps(evlist, &trace->opts.target);
3335         if (err < 0) {
3336                 fprintf(trace->output, "Problems parsing the target to trace, check your options!\n");
3337                 goto out_delete_evlist;
3338         }
3339
3340         err = trace__symbols_init(trace, evlist);
3341         if (err < 0) {
3342                 fprintf(trace->output, "Problems initializing symbol libraries!\n");
3343                 goto out_delete_evlist;
3344         }
3345
3346         perf_evlist__config(evlist, &trace->opts, &callchain_param);
3347
3348         signal(SIGCHLD, sig_handler);
3349         signal(SIGINT, sig_handler);
3350
3351         if (forks) {
3352                 err = perf_evlist__prepare_workload(evlist, &trace->opts.target,
3353                                                     argv, false, NULL);
3354                 if (err < 0) {
3355                         fprintf(trace->output, "Couldn't run the workload!\n");
3356                         goto out_delete_evlist;
3357                 }
3358         }
3359
3360         err = evlist__open(evlist);
3361         if (err < 0)
3362                 goto out_error_open;
3363
3364         err = bpf__apply_obj_config();
3365         if (err) {
3366                 char errbuf[BUFSIZ];
3367
3368                 bpf__strerror_apply_obj_config(err, errbuf, sizeof(errbuf));
3369                 pr_err("ERROR: Apply config to BPF failed: %s\n",
3370                          errbuf);
3371                 goto out_error_open;
3372         }
3373
3374         err = trace__set_filter_pids(trace);
3375         if (err < 0)
3376                 goto out_error_mem;
3377
3378         if (trace->syscalls.map)
3379                 trace__init_syscalls_bpf_map(trace);
3380
3381         if (trace->syscalls.prog_array.sys_enter)
3382                 trace__init_syscalls_bpf_prog_array_maps(trace);
3383
3384         if (trace->ev_qualifier_ids.nr > 0) {
3385                 err = trace__set_ev_qualifier_filter(trace);
3386                 if (err < 0)
3387                         goto out_errno;
3388
3389                 if (trace->syscalls.events.sys_exit) {
3390                         pr_debug("event qualifier tracepoint filter: %s\n",
3391                                  trace->syscalls.events.sys_exit->filter);
3392                 }
3393         }
3394
3395         /*
3396          * If the "close" syscall is not traced, then we will not have the
3397          * opportunity to, in syscall_arg__scnprintf_close_fd() invalidate the
3398          * fd->pathname table and were ending up showing the last value set by
3399          * syscalls opening a pathname and associating it with a descriptor or
3400          * reading it from /proc/pid/fd/ in cases where that doesn't make
3401          * sense.
3402          *
3403          *  So just disable this beautifier (SCA_FD, SCA_FDAT) when 'close' is
3404          *  not in use.
3405          */
3406         trace->fd_path_disabled = !trace__syscall_enabled(trace, syscalltbl__id(trace->sctbl, "close"));
3407
3408         err = perf_evlist__apply_filters(evlist, &evsel);
3409         if (err < 0)
3410                 goto out_error_apply_filters;
3411
3412         if (trace->dump.map)
3413                 bpf_map__fprintf(trace->dump.map, trace->output);
3414
3415         err = evlist__mmap(evlist, trace->opts.mmap_pages);
3416         if (err < 0)
3417                 goto out_error_mmap;
3418
3419         if (!target__none(&trace->opts.target) && !trace->opts.initial_delay)
3420                 evlist__enable(evlist);
3421
3422         if (forks)
3423                 perf_evlist__start_workload(evlist);
3424
3425         if (trace->opts.initial_delay) {
3426                 usleep(trace->opts.initial_delay * 1000);
3427                 evlist__enable(evlist);
3428         }
3429
3430         trace->multiple_threads = perf_thread_map__pid(evlist->core.threads, 0) == -1 ||
3431                                   evlist->core.threads->nr > 1 ||
3432                                   evlist__first(evlist)->core.attr.inherit;
3433
3434         /*
3435          * Now that we already used evsel->core.attr to ask the kernel to setup the
3436          * events, lets reuse evsel->core.attr.sample_max_stack as the limit in
3437          * trace__resolve_callchain(), allowing per-event max-stack settings
3438          * to override an explicitly set --max-stack global setting.
3439          */
3440         evlist__for_each_entry(evlist, evsel) {
3441                 if (evsel__has_callchain(evsel) &&
3442                     evsel->core.attr.sample_max_stack == 0)
3443                         evsel->core.attr.sample_max_stack = trace->max_stack;
3444         }
3445 again:
3446         before = trace->nr_events;
3447
3448         for (i = 0; i < evlist->core.nr_mmaps; i++) {
3449                 union perf_event *event;
3450                 struct mmap *md;
3451
3452                 md = &evlist->mmap[i];
3453                 if (perf_mmap__read_init(md) < 0)
3454                         continue;
3455
3456                 while ((event = perf_mmap__read_event(md)) != NULL) {
3457                         ++trace->nr_events;
3458
3459                         err = trace__deliver_event(trace, event);
3460                         if (err)
3461                                 goto out_disable;
3462
3463                         perf_mmap__consume(md);
3464
3465                         if (interrupted)
3466                                 goto out_disable;
3467
3468                         if (done && !draining) {
3469                                 evlist__disable(evlist);
3470                                 draining = true;
3471                         }
3472                 }
3473                 perf_mmap__read_done(md);
3474         }
3475
3476         if (trace->nr_events == before) {
3477                 int timeout = done ? 100 : -1;
3478
3479                 if (!draining && evlist__poll(evlist, timeout) > 0) {
3480                         if (evlist__filter_pollfd(evlist, POLLERR | POLLHUP | POLLNVAL) == 0)
3481                                 draining = true;
3482
3483                         goto again;
3484                 } else {
3485                         if (trace__flush_events(trace))
3486                                 goto out_disable;
3487                 }
3488         } else {
3489                 goto again;
3490         }
3491
3492 out_disable:
3493         thread__zput(trace->current);
3494
3495         evlist__disable(evlist);
3496
3497         if (trace->sort_events)
3498                 ordered_events__flush(&trace->oe.data, OE_FLUSH__FINAL);
3499
3500         if (!err) {
3501                 if (trace->summary)
3502                         trace__fprintf_thread_summary(trace, trace->output);
3503
3504                 if (trace->show_tool_stats) {
3505                         fprintf(trace->output, "Stats:\n "
3506                                                " vfs_getname : %" PRIu64 "\n"
3507                                                " proc_getname: %" PRIu64 "\n",
3508                                 trace->stats.vfs_getname,
3509                                 trace->stats.proc_getname);
3510                 }
3511         }
3512
3513 out_delete_evlist:
3514         trace__symbols__exit(trace);
3515
3516         evlist__delete(evlist);
3517         cgroup__put(trace->cgroup);
3518         trace->evlist = NULL;
3519         trace->live = false;
3520         return err;
3521 {
3522         char errbuf[BUFSIZ];
3523
3524 out_error_sched_stat_runtime:
3525         tracing_path__strerror_open_tp(errno, errbuf, sizeof(errbuf), "sched", "sched_stat_runtime");
3526         goto out_error;
3527
3528 out_error_raw_syscalls:
3529         tracing_path__strerror_open_tp(errno, errbuf, sizeof(errbuf), "raw_syscalls", "sys_(enter|exit)");
3530         goto out_error;
3531
3532 out_error_mmap:
3533         perf_evlist__strerror_mmap(evlist, errno, errbuf, sizeof(errbuf));
3534         goto out_error;
3535
3536 out_error_open:
3537         perf_evlist__strerror_open(evlist, errno, errbuf, sizeof(errbuf));
3538
3539 out_error:
3540         fprintf(trace->output, "%s\n", errbuf);
3541         goto out_delete_evlist;
3542
3543 out_error_apply_filters:
3544         fprintf(trace->output,
3545                 "Failed to set filter \"%s\" on event %s with %d (%s)\n",
3546                 evsel->filter, perf_evsel__name(evsel), errno,
3547                 str_error_r(errno, errbuf, sizeof(errbuf)));
3548         goto out_delete_evlist;
3549 }
3550 out_error_mem:
3551         fprintf(trace->output, "Not enough memory to run!\n");
3552         goto out_delete_evlist;
3553
3554 out_errno:
3555         fprintf(trace->output, "errno=%d,%s\n", errno, strerror(errno));
3556         goto out_delete_evlist;
3557 }
3558
3559 static int trace__replay(struct trace *trace)
3560 {
3561         const struct evsel_str_handler handlers[] = {
3562                 { "probe:vfs_getname",       trace__vfs_getname, },
3563         };
3564         struct perf_data data = {
3565                 .path  = input_name,
3566                 .mode  = PERF_DATA_MODE_READ,
3567                 .force = trace->force,
3568         };
3569         struct perf_session *session;
3570         struct evsel *evsel;
3571         int err = -1;
3572
3573         trace->tool.sample        = trace__process_sample;
3574         trace->tool.mmap          = perf_event__process_mmap;
3575         trace->tool.mmap2         = perf_event__process_mmap2;
3576         trace->tool.comm          = perf_event__process_comm;
3577         trace->tool.exit          = perf_event__process_exit;
3578         trace->tool.fork          = perf_event__process_fork;
3579         trace->tool.attr          = perf_event__process_attr;
3580         trace->tool.tracing_data  = perf_event__process_tracing_data;
3581         trace->tool.build_id      = perf_event__process_build_id;
3582         trace->tool.namespaces    = perf_event__process_namespaces;
3583
3584         trace->tool.ordered_events = true;
3585         trace->tool.ordering_requires_timestamps = true;
3586
3587         /* add tid to output */
3588         trace->multiple_threads = true;
3589
3590         session = perf_session__new(&data, false, &trace->tool);
3591         if (IS_ERR(session))
3592                 return PTR_ERR(session);
3593
3594         if (trace->opts.target.pid)
3595                 symbol_conf.pid_list_str = strdup(trace->opts.target.pid);
3596
3597         if (trace->opts.target.tid)
3598                 symbol_conf.tid_list_str = strdup(trace->opts.target.tid);
3599
3600         if (symbol__init(&session->header.env) < 0)
3601                 goto out;
3602
3603         trace->host = &session->machines.host;
3604
3605         err = perf_session__set_tracepoints_handlers(session, handlers);
3606         if (err)
3607                 goto out;
3608
3609         evsel = perf_evlist__find_tracepoint_by_name(session->evlist,
3610                                                      "raw_syscalls:sys_enter");
3611         /* older kernels have syscalls tp versus raw_syscalls */
3612         if (evsel == NULL)
3613                 evsel = perf_evlist__find_tracepoint_by_name(session->evlist,
3614                                                              "syscalls:sys_enter");
3615
3616         if (evsel &&
3617             (perf_evsel__init_raw_syscall_tp(evsel, trace__sys_enter) < 0 ||
3618             perf_evsel__init_sc_tp_ptr_field(evsel, args))) {
3619                 pr_err("Error during initialize raw_syscalls:sys_enter event\n");
3620                 goto out;
3621         }
3622
3623         evsel = perf_evlist__find_tracepoint_by_name(session->evlist,
3624                                                      "raw_syscalls:sys_exit");
3625         if (evsel == NULL)
3626                 evsel = perf_evlist__find_tracepoint_by_name(session->evlist,
3627                                                              "syscalls:sys_exit");
3628         if (evsel &&
3629             (perf_evsel__init_raw_syscall_tp(evsel, trace__sys_exit) < 0 ||
3630             perf_evsel__init_sc_tp_uint_field(evsel, ret))) {
3631                 pr_err("Error during initialize raw_syscalls:sys_exit event\n");
3632                 goto out;
3633         }
3634
3635         evlist__for_each_entry(session->evlist, evsel) {
3636                 if (evsel->core.attr.type == PERF_TYPE_SOFTWARE &&
3637                     (evsel->core.attr.config == PERF_COUNT_SW_PAGE_FAULTS_MAJ ||
3638                      evsel->core.attr.config == PERF_COUNT_SW_PAGE_FAULTS_MIN ||
3639                      evsel->core.attr.config == PERF_COUNT_SW_PAGE_FAULTS))
3640                         evsel->handler = trace__pgfault;
3641         }
3642
3643         setup_pager();
3644
3645         err = perf_session__process_events(session);
3646         if (err)
3647                 pr_err("Failed to process events, error %d", err);
3648
3649         else if (trace->summary)
3650                 trace__fprintf_thread_summary(trace, trace->output);
3651
3652 out:
3653         perf_session__delete(session);
3654
3655         return err;
3656 }
3657
3658 static size_t trace__fprintf_threads_header(FILE *fp)
3659 {
3660         size_t printed;
3661
3662         printed  = fprintf(fp, "\n Summary of events:\n\n");
3663
3664         return printed;
3665 }
3666
3667 DEFINE_RESORT_RB(syscall_stats, a->msecs > b->msecs,
3668         struct stats    *stats;
3669         double          msecs;
3670         int             syscall;
3671 )
3672 {
3673         struct int_node *source = rb_entry(nd, struct int_node, rb_node);
3674         struct stats *stats = source->priv;
3675
3676         entry->syscall = source->i;
3677         entry->stats   = stats;
3678         entry->msecs   = stats ? (u64)stats->n * (avg_stats(stats) / NSEC_PER_MSEC) : 0;
3679 }
3680
3681 static size_t thread__dump_stats(struct thread_trace *ttrace,
3682                                  struct trace *trace, FILE *fp)
3683 {
3684         size_t printed = 0;
3685         struct syscall *sc;
3686         struct rb_node *nd;
3687         DECLARE_RESORT_RB_INTLIST(syscall_stats, ttrace->syscall_stats);
3688
3689         if (syscall_stats == NULL)
3690                 return 0;
3691
3692         printed += fprintf(fp, "\n");
3693
3694         printed += fprintf(fp, "   syscall            calls    total       min       avg       max      stddev\n");
3695         printed += fprintf(fp, "                               (msec)    (msec)    (msec)    (msec)        (%%)\n");
3696         printed += fprintf(fp, "   --------------- -------- --------- --------- --------- ---------     ------\n");
3697
3698         resort_rb__for_each_entry(nd, syscall_stats) {
3699                 struct stats *stats = syscall_stats_entry->stats;
3700                 if (stats) {
3701                         double min = (double)(stats->min) / NSEC_PER_MSEC;
3702                         double max = (double)(stats->max) / NSEC_PER_MSEC;
3703                         double avg = avg_stats(stats);
3704                         double pct;
3705                         u64 n = (u64) stats->n;
3706
3707                         pct = avg ? 100.0 * stddev_stats(stats)/avg : 0.0;
3708                         avg /= NSEC_PER_MSEC;
3709
3710                         sc = &trace->syscalls.table[syscall_stats_entry->syscall];
3711                         printed += fprintf(fp, "   %-15s", sc->name);
3712                         printed += fprintf(fp, " %8" PRIu64 " %9.3f %9.3f %9.3f",
3713                                            n, syscall_stats_entry->msecs, min, avg);
3714                         printed += fprintf(fp, " %9.3f %9.2f%%\n", max, pct);
3715                 }
3716         }
3717
3718         resort_rb__delete(syscall_stats);
3719         printed += fprintf(fp, "\n\n");
3720
3721         return printed;
3722 }
3723
3724 static size_t trace__fprintf_thread(FILE *fp, struct thread *thread, struct trace *trace)
3725 {
3726         size_t printed = 0;
3727         struct thread_trace *ttrace = thread__priv(thread);
3728         double ratio;
3729
3730         if (ttrace == NULL)
3731                 return 0;
3732
3733         ratio = (double)ttrace->nr_events / trace->nr_events * 100.0;
3734
3735         printed += fprintf(fp, " %s (%d), ", thread__comm_str(thread), thread->tid);
3736         printed += fprintf(fp, "%lu events, ", ttrace->nr_events);
3737         printed += fprintf(fp, "%.1f%%", ratio);
3738         if (ttrace->pfmaj)
3739                 printed += fprintf(fp, ", %lu majfaults", ttrace->pfmaj);
3740         if (ttrace->pfmin)
3741                 printed += fprintf(fp, ", %lu minfaults", ttrace->pfmin);
3742         if (trace->sched)
3743                 printed += fprintf(fp, ", %.3f msec\n", ttrace->runtime_ms);
3744         else if (fputc('\n', fp) != EOF)
3745                 ++printed;
3746
3747         printed += thread__dump_stats(ttrace, trace, fp);
3748
3749         return printed;
3750 }
3751
3752 static unsigned long thread__nr_events(struct thread_trace *ttrace)
3753 {
3754         return ttrace ? ttrace->nr_events : 0;
3755 }
3756
3757 DEFINE_RESORT_RB(threads, (thread__nr_events(a->thread->priv) < thread__nr_events(b->thread->priv)),
3758         struct thread *thread;
3759 )
3760 {
3761         entry->thread = rb_entry(nd, struct thread, rb_node);
3762 }
3763
3764 static size_t trace__fprintf_thread_summary(struct trace *trace, FILE *fp)
3765 {
3766         size_t printed = trace__fprintf_threads_header(fp);
3767         struct rb_node *nd;
3768         int i;
3769
3770         for (i = 0; i < THREADS__TABLE_SIZE; i++) {
3771                 DECLARE_RESORT_RB_MACHINE_THREADS(threads, trace->host, i);
3772
3773                 if (threads == NULL) {
3774                         fprintf(fp, "%s", "Error sorting output by nr_events!\n");
3775                         return 0;
3776                 }
3777
3778                 resort_rb__for_each_entry(nd, threads)
3779                         printed += trace__fprintf_thread(fp, threads_entry->thread, trace);
3780
3781                 resort_rb__delete(threads);
3782         }
3783         return printed;
3784 }
3785
3786 static int trace__set_duration(const struct option *opt, const char *str,
3787                                int unset __maybe_unused)
3788 {
3789         struct trace *trace = opt->value;
3790
3791         trace->duration_filter = atof(str);
3792         return 0;
3793 }
3794
3795 static int trace__set_filter_pids_from_option(const struct option *opt, const char *str,
3796                                               int unset __maybe_unused)
3797 {
3798         int ret = -1;
3799         size_t i;
3800         struct trace *trace = opt->value;
3801         /*
3802          * FIXME: introduce a intarray class, plain parse csv and create a
3803          * { int nr, int entries[] } struct...
3804          */
3805         struct intlist *list = intlist__new(str);
3806
3807         if (list == NULL)
3808                 return -1;
3809
3810         i = trace->filter_pids.nr = intlist__nr_entries(list) + 1;
3811         trace->filter_pids.entries = calloc(i, sizeof(pid_t));
3812
3813         if (trace->filter_pids.entries == NULL)
3814                 goto out;
3815
3816         trace->filter_pids.entries[0] = getpid();
3817
3818         for (i = 1; i < trace->filter_pids.nr; ++i)
3819                 trace->filter_pids.entries[i] = intlist__entry(list, i - 1)->i;
3820
3821         intlist__delete(list);
3822         ret = 0;
3823 out:
3824         return ret;
3825 }
3826
3827 static int trace__open_output(struct trace *trace, const char *filename)
3828 {
3829         struct stat st;
3830
3831         if (!stat(filename, &st) && st.st_size) {
3832                 char oldname[PATH_MAX];
3833
3834                 scnprintf(oldname, sizeof(oldname), "%s.old", filename);
3835                 unlink(oldname);
3836                 rename(filename, oldname);
3837         }
3838
3839         trace->output = fopen(filename, "w");
3840
3841         return trace->output == NULL ? -errno : 0;
3842 }
3843
3844 static int parse_pagefaults(const struct option *opt, const char *str,
3845                             int unset __maybe_unused)
3846 {
3847         int *trace_pgfaults = opt->value;
3848
3849         if (strcmp(str, "all") == 0)
3850                 *trace_pgfaults |= TRACE_PFMAJ | TRACE_PFMIN;
3851         else if (strcmp(str, "maj") == 0)
3852                 *trace_pgfaults |= TRACE_PFMAJ;
3853         else if (strcmp(str, "min") == 0)
3854                 *trace_pgfaults |= TRACE_PFMIN;
3855         else
3856                 return -1;
3857
3858         return 0;
3859 }
3860
3861 static void evlist__set_evsel_handler(struct evlist *evlist, void *handler)
3862 {
3863         struct evsel *evsel;
3864
3865         evlist__for_each_entry(evlist, evsel)
3866                 evsel->handler = handler;
3867 }
3868
3869 static int evlist__set_syscall_tp_fields(struct evlist *evlist)
3870 {
3871         struct evsel *evsel;
3872
3873         evlist__for_each_entry(evlist, evsel) {
3874                 if (evsel->priv || !evsel->tp_format)
3875                         continue;
3876
3877                 if (strcmp(evsel->tp_format->system, "syscalls"))
3878                         continue;
3879
3880                 if (perf_evsel__init_syscall_tp(evsel))
3881                         return -1;
3882
3883                 if (!strncmp(evsel->tp_format->name, "sys_enter_", 10)) {
3884                         struct syscall_tp *sc = evsel->priv;
3885
3886                         if (__tp_field__init_ptr(&sc->args, sc->id.offset + sizeof(u64)))
3887                                 return -1;
3888                 } else if (!strncmp(evsel->tp_format->name, "sys_exit_", 9)) {
3889                         struct syscall_tp *sc = evsel->priv;
3890
3891                         if (__tp_field__init_uint(&sc->ret, sizeof(u64), sc->id.offset + sizeof(u64), evsel->needs_swap))
3892                                 return -1;
3893                 }
3894         }
3895
3896         return 0;
3897 }
3898
3899 /*
3900  * XXX: Hackish, just splitting the combined -e+--event (syscalls
3901  * (raw_syscalls:{sys_{enter,exit}} + events (tracepoints, HW, SW, etc) to use
3902  * existing facilities unchanged (trace->ev_qualifier + parse_options()).
3903  *
3904  * It'd be better to introduce a parse_options() variant that would return a
3905  * list with the terms it didn't match to an event...
3906  */
3907 static int trace__parse_events_option(const struct option *opt, const char *str,
3908                                       int unset __maybe_unused)
3909 {
3910         struct trace *trace = (struct trace *)opt->value;
3911         const char *s = str;
3912         char *sep = NULL, *lists[2] = { NULL, NULL, };
3913         int len = strlen(str) + 1, err = -1, list, idx;
3914         char *strace_groups_dir = system_path(STRACE_GROUPS_DIR);
3915         char group_name[PATH_MAX];
3916         struct syscall_fmt *fmt;
3917
3918         if (strace_groups_dir == NULL)
3919                 return -1;
3920
3921         if (*s == '!') {
3922                 ++s;
3923                 trace->not_ev_qualifier = true;
3924         }
3925
3926         while (1) {
3927                 if ((sep = strchr(s, ',')) != NULL)
3928                         *sep = '\0';
3929
3930                 list = 0;
3931                 if (syscalltbl__id(trace->sctbl, s) >= 0 ||
3932                     syscalltbl__strglobmatch_first(trace->sctbl, s, &idx) >= 0) {
3933                         list = 1;
3934                         goto do_concat;
3935                 }
3936
3937                 fmt = syscall_fmt__find_by_alias(s);
3938                 if (fmt != NULL) {
3939                         list = 1;
3940                         s = fmt->name;
3941                 } else {
3942                         path__join(group_name, sizeof(group_name), strace_groups_dir, s);
3943                         if (access(group_name, R_OK) == 0)
3944                                 list = 1;
3945                 }
3946 do_concat:
3947                 if (lists[list]) {
3948                         sprintf(lists[list] + strlen(lists[list]), ",%s", s);
3949                 } else {
3950                         lists[list] = malloc(len);
3951                         if (lists[list] == NULL)
3952                                 goto out;
3953                         strcpy(lists[list], s);
3954                 }
3955
3956                 if (!sep)
3957                         break;
3958
3959                 *sep = ',';
3960                 s = sep + 1;
3961         }
3962
3963         if (lists[1] != NULL) {
3964                 struct strlist_config slist_config = {
3965                         .dirname = strace_groups_dir,
3966                 };
3967
3968                 trace->ev_qualifier = strlist__new(lists[1], &slist_config);
3969                 if (trace->ev_qualifier == NULL) {
3970                         fputs("Not enough memory to parse event qualifier", trace->output);
3971                         goto out;
3972                 }
3973
3974                 if (trace__validate_ev_qualifier(trace))
3975                         goto out;
3976                 trace->trace_syscalls = true;
3977         }
3978
3979         err = 0;
3980
3981         if (lists[0]) {
3982                 struct option o = OPT_CALLBACK('e', "event", &trace->evlist, "event",
3983                                                "event selector. use 'perf list' to list available events",
3984                                                parse_events_option);
3985                 err = parse_events_option(&o, lists[0], 0);
3986         }
3987 out:
3988         if (sep)
3989                 *sep = ',';
3990
3991         return err;
3992 }
3993
3994 static int trace__parse_cgroups(const struct option *opt, const char *str, int unset)
3995 {
3996         struct trace *trace = opt->value;
3997
3998         if (!list_empty(&trace->evlist->core.entries))
3999                 return parse_cgroups(opt, str, unset);
4000
4001         trace->cgroup = evlist__findnew_cgroup(trace->evlist, str);
4002
4003         return 0;
4004 }
4005
4006 static struct bpf_map *trace__find_bpf_map_by_name(struct trace *trace, const char *name)
4007 {
4008         if (trace->bpf_obj == NULL)
4009                 return NULL;
4010
4011         return bpf_object__find_map_by_name(trace->bpf_obj, name);
4012 }
4013
4014 static void trace__set_bpf_map_filtered_pids(struct trace *trace)
4015 {
4016         trace->filter_pids.map = trace__find_bpf_map_by_name(trace, "pids_filtered");
4017 }
4018
4019 static void trace__set_bpf_map_syscalls(struct trace *trace)
4020 {
4021         trace->syscalls.map = trace__find_bpf_map_by_name(trace, "syscalls");
4022         trace->syscalls.prog_array.sys_enter = trace__find_bpf_map_by_name(trace, "syscalls_sys_enter");
4023         trace->syscalls.prog_array.sys_exit  = trace__find_bpf_map_by_name(trace, "syscalls_sys_exit");
4024 }
4025
4026 static int trace__config(const char *var, const char *value, void *arg)
4027 {
4028         struct trace *trace = arg;
4029         int err = 0;
4030
4031         if (!strcmp(var, "trace.add_events")) {
4032                 struct option o = OPT_CALLBACK('e', "event", &trace->evlist, "event",
4033                                                "event selector. use 'perf list' to list available events",
4034                                                parse_events_option);
4035                 /*
4036                  * We can't propagate parse_event_option() return, as it is 1
4037                  * for failure while perf_config() expects -1.
4038                  */
4039                 if (parse_events_option(&o, value, 0))
4040                         err = -1;
4041         } else if (!strcmp(var, "trace.show_timestamp")) {
4042                 trace->show_tstamp = perf_config_bool(var, value);
4043         } else if (!strcmp(var, "trace.show_duration")) {
4044                 trace->show_duration = perf_config_bool(var, value);
4045         } else if (!strcmp(var, "trace.show_arg_names")) {
4046                 trace->show_arg_names = perf_config_bool(var, value);
4047                 if (!trace->show_arg_names)
4048                         trace->show_zeros = true;
4049         } else if (!strcmp(var, "trace.show_zeros")) {
4050                 bool new_show_zeros = perf_config_bool(var, value);
4051                 if (!trace->show_arg_names && !new_show_zeros) {
4052                         pr_warning("trace.show_zeros has to be set when trace.show_arg_names=no\n");
4053                         goto out;
4054                 }
4055                 trace->show_zeros = new_show_zeros;
4056         } else if (!strcmp(var, "trace.show_prefix")) {
4057                 trace->show_string_prefix = perf_config_bool(var, value);
4058         } else if (!strcmp(var, "trace.no_inherit")) {
4059                 trace->opts.no_inherit = perf_config_bool(var, value);
4060         } else if (!strcmp(var, "trace.args_alignment")) {
4061                 int args_alignment = 0;
4062                 if (perf_config_int(&args_alignment, var, value) == 0)
4063                         trace->args_alignment = args_alignment;
4064         }
4065 out:
4066         return err;
4067 }
4068
4069 int cmd_trace(int argc, const char **argv)
4070 {
4071         const char *trace_usage[] = {
4072                 "perf trace [<options>] [<command>]",
4073                 "perf trace [<options>] -- <command> [<options>]",
4074                 "perf trace record [<options>] [<command>]",
4075                 "perf trace record [<options>] -- <command> [<options>]",
4076                 NULL
4077         };
4078         struct trace trace = {
4079                 .opts = {
4080                         .target = {
4081                                 .uid       = UINT_MAX,
4082                                 .uses_mmap = true,
4083                         },
4084                         .user_freq     = UINT_MAX,
4085                         .user_interval = ULLONG_MAX,
4086                         .no_buffering  = true,
4087                         .mmap_pages    = UINT_MAX,
4088                 },
4089                 .output = stderr,
4090                 .show_comm = true,
4091                 .show_tstamp = true,
4092                 .show_duration = true,
4093                 .show_arg_names = true,
4094                 .args_alignment = 70,
4095                 .trace_syscalls = false,
4096                 .kernel_syscallchains = false,
4097                 .max_stack = UINT_MAX,
4098                 .max_events = ULONG_MAX,
4099         };
4100         const char *map_dump_str = NULL;
4101         const char *output_name = NULL;
4102         const struct option trace_options[] = {
4103         OPT_CALLBACK('e', "event", &trace, "event",
4104                      "event/syscall selector. use 'perf list' to list available events",
4105                      trace__parse_events_option),
4106         OPT_BOOLEAN(0, "comm", &trace.show_comm,
4107                     "show the thread COMM next to its id"),
4108         OPT_BOOLEAN(0, "tool_stats", &trace.show_tool_stats, "show tool stats"),
4109         OPT_CALLBACK(0, "expr", &trace, "expr", "list of syscalls/events to trace",
4110                      trace__parse_events_option),
4111         OPT_STRING('o', "output", &output_name, "file", "output file name"),
4112         OPT_STRING('i', "input", &input_name, "file", "Analyze events in file"),
4113         OPT_STRING('p', "pid", &trace.opts.target.pid, "pid",
4114                     "trace events on existing process id"),
4115         OPT_STRING('t', "tid", &trace.opts.target.tid, "tid",
4116                     "trace events on existing thread id"),
4117         OPT_CALLBACK(0, "filter-pids", &trace, "CSV list of pids",
4118                      "pids to filter (by the kernel)", trace__set_filter_pids_from_option),
4119         OPT_BOOLEAN('a', "all-cpus", &trace.opts.target.system_wide,
4120                     "system-wide collection from all CPUs"),
4121         OPT_STRING('C', "cpu", &trace.opts.target.cpu_list, "cpu",
4122                     "list of cpus to monitor"),
4123         OPT_BOOLEAN(0, "no-inherit", &trace.opts.no_inherit,
4124                     "child tasks do not inherit counters"),
4125         OPT_CALLBACK('m', "mmap-pages", &trace.opts.mmap_pages, "pages",
4126                      "number of mmap data pages",
4127                      perf_evlist__parse_mmap_pages),
4128         OPT_STRING('u', "uid", &trace.opts.target.uid_str, "user",
4129                    "user to profile"),
4130         OPT_CALLBACK(0, "duration", &trace, "float",
4131                      "show only events with duration > N.M ms",
4132                      trace__set_duration),
4133 #ifdef HAVE_LIBBPF_SUPPORT
4134         OPT_STRING(0, "map-dump", &map_dump_str, "BPF map", "BPF map to periodically dump"),
4135 #endif
4136         OPT_BOOLEAN(0, "sched", &trace.sched, "show blocking scheduler events"),
4137         OPT_INCR('v', "verbose", &verbose, "be more verbose"),
4138         OPT_BOOLEAN('T', "time", &trace.full_time,
4139                     "Show full timestamp, not time relative to first start"),
4140         OPT_BOOLEAN(0, "failure", &trace.failure_only,
4141                     "Show only syscalls that failed"),
4142         OPT_BOOLEAN('s', "summary", &trace.summary_only,
4143                     "Show only syscall summary with statistics"),
4144         OPT_BOOLEAN('S', "with-summary", &trace.summary,
4145                     "Show all syscalls and summary with statistics"),
4146         OPT_CALLBACK_DEFAULT('F', "pf", &trace.trace_pgfaults, "all|maj|min",
4147                      "Trace pagefaults", parse_pagefaults, "maj"),
4148         OPT_BOOLEAN(0, "syscalls", &trace.trace_syscalls, "Trace syscalls"),
4149         OPT_BOOLEAN('f', "force", &trace.force, "don't complain, do it"),
4150         OPT_CALLBACK(0, "call-graph", &trace.opts,
4151                      "record_mode[,record_size]", record_callchain_help,
4152                      &record_parse_callchain_opt),
4153         OPT_BOOLEAN(0, "kernel-syscall-graph", &trace.kernel_syscallchains,
4154                     "Show the kernel callchains on the syscall exit path"),
4155         OPT_ULONG(0, "max-events", &trace.max_events,
4156                 "Set the maximum number of events to print, exit after that is reached. "),
4157         OPT_UINTEGER(0, "min-stack", &trace.min_stack,
4158                      "Set the minimum stack depth when parsing the callchain, "
4159                      "anything below the specified depth will be ignored."),
4160         OPT_UINTEGER(0, "max-stack", &trace.max_stack,
4161                      "Set the maximum stack depth when parsing the callchain, "
4162                      "anything beyond the specified depth will be ignored. "
4163                      "Default: kernel.perf_event_max_stack or " __stringify(PERF_MAX_STACK_DEPTH)),
4164         OPT_BOOLEAN(0, "sort-events", &trace.sort_events,
4165                         "Sort batch of events before processing, use if getting out of order events"),
4166         OPT_BOOLEAN(0, "print-sample", &trace.print_sample,
4167                         "print the PERF_RECORD_SAMPLE PERF_SAMPLE_ info, for debugging"),
4168         OPT_UINTEGER(0, "proc-map-timeout", &proc_map_timeout,
4169                         "per thread proc mmap processing timeout in ms"),
4170         OPT_CALLBACK('G', "cgroup", &trace, "name", "monitor event in cgroup name only",
4171                      trace__parse_cgroups),
4172         OPT_UINTEGER('D', "delay", &trace.opts.initial_delay,
4173                      "ms to wait before starting measurement after program "
4174                      "start"),
4175         OPTS_EVSWITCH(&trace.evswitch),
4176         OPT_END()
4177         };
4178         bool __maybe_unused max_stack_user_set = true;
4179         bool mmap_pages_user_set = true;
4180         struct evsel *evsel;
4181         const char * const trace_subcommands[] = { "record", NULL };
4182         int err = -1;
4183         char bf[BUFSIZ];
4184
4185         signal(SIGSEGV, sighandler_dump_stack);
4186         signal(SIGFPE, sighandler_dump_stack);
4187
4188         trace.evlist = evlist__new();
4189         trace.sctbl = syscalltbl__new();
4190
4191         if (trace.evlist == NULL || trace.sctbl == NULL) {
4192                 pr_err("Not enough memory to run!\n");
4193                 err = -ENOMEM;
4194                 goto out;
4195         }
4196
4197         /*
4198          * Parsing .perfconfig may entail creating a BPF event, that may need
4199          * to create BPF maps, so bump RLIM_MEMLOCK as the default 64K setting
4200          * is too small. This affects just this process, not touching the
4201          * global setting. If it fails we'll get something in 'perf trace -v'
4202          * to help diagnose the problem.
4203          */
4204         rlimit__bump_memlock();
4205
4206         err = perf_config(trace__config, &trace);
4207         if (err)
4208                 goto out;
4209
4210         argc = parse_options_subcommand(argc, argv, trace_options, trace_subcommands,
4211                                  trace_usage, PARSE_OPT_STOP_AT_NON_OPTION);
4212
4213         if ((nr_cgroups || trace.cgroup) && !trace.opts.target.system_wide) {
4214                 usage_with_options_msg(trace_usage, trace_options,
4215                                        "cgroup monitoring only available in system-wide mode");
4216         }
4217
4218         evsel = bpf__setup_output_event(trace.evlist, "__augmented_syscalls__");
4219         if (IS_ERR(evsel)) {
4220                 bpf__strerror_setup_output_event(trace.evlist, PTR_ERR(evsel), bf, sizeof(bf));
4221                 pr_err("ERROR: Setup trace syscalls enter failed: %s\n", bf);
4222                 goto out;
4223         }
4224
4225         if (evsel) {
4226                 trace.syscalls.events.augmented = evsel;
4227
4228                 evsel = perf_evlist__find_tracepoint_by_name(trace.evlist, "raw_syscalls:sys_enter");
4229                 if (evsel == NULL) {
4230                         pr_err("ERROR: raw_syscalls:sys_enter not found in the augmented BPF object\n");
4231                         goto out;
4232                 }
4233
4234                 if (evsel->bpf_obj == NULL) {
4235                         pr_err("ERROR: raw_syscalls:sys_enter not associated to a BPF object\n");
4236                         goto out;
4237                 }
4238
4239                 trace.bpf_obj = evsel->bpf_obj;
4240
4241                 trace__set_bpf_map_filtered_pids(&trace);
4242                 trace__set_bpf_map_syscalls(&trace);
4243                 trace.syscalls.unaugmented_prog = trace__find_bpf_program_by_title(&trace, "!raw_syscalls:unaugmented");
4244         }
4245
4246         err = bpf__setup_stdout(trace.evlist);
4247         if (err) {
4248                 bpf__strerror_setup_stdout(trace.evlist, err, bf, sizeof(bf));
4249                 pr_err("ERROR: Setup BPF stdout failed: %s\n", bf);
4250                 goto out;
4251         }
4252
4253         err = -1;
4254
4255         if (map_dump_str) {
4256                 trace.dump.map = trace__find_bpf_map_by_name(&trace, map_dump_str);
4257                 if (trace.dump.map == NULL) {
4258                         pr_err("ERROR: BPF map \"%s\" not found\n", map_dump_str);
4259                         goto out;
4260                 }
4261         }
4262
4263         if (trace.trace_pgfaults) {
4264                 trace.opts.sample_address = true;
4265                 trace.opts.sample_time = true;
4266         }
4267
4268         if (trace.opts.mmap_pages == UINT_MAX)
4269                 mmap_pages_user_set = false;
4270
4271         if (trace.max_stack == UINT_MAX) {
4272                 trace.max_stack = input_name ? PERF_MAX_STACK_DEPTH : sysctl__max_stack();
4273                 max_stack_user_set = false;
4274         }
4275
4276 #ifdef HAVE_DWARF_UNWIND_SUPPORT
4277         if ((trace.min_stack || max_stack_user_set) && !callchain_param.enabled) {
4278                 record_opts__parse_callchain(&trace.opts, &callchain_param, "dwarf", false);
4279         }
4280 #endif
4281
4282         if (callchain_param.enabled) {
4283                 if (!mmap_pages_user_set && geteuid() == 0)
4284                         trace.opts.mmap_pages = perf_event_mlock_kb_in_pages() * 4;
4285
4286                 symbol_conf.use_callchain = true;
4287         }
4288
4289         if (trace.evlist->core.nr_entries > 0) {
4290                 evlist__set_evsel_handler(trace.evlist, trace__event_handler);
4291                 if (evlist__set_syscall_tp_fields(trace.evlist)) {
4292                         perror("failed to set syscalls:* tracepoint fields");
4293                         goto out;
4294                 }
4295         }
4296
4297         if (trace.sort_events) {
4298                 ordered_events__init(&trace.oe.data, ordered_events__deliver_event, &trace);
4299                 ordered_events__set_copy_on_queue(&trace.oe.data, true);
4300         }
4301
4302         /*
4303          * If we are augmenting syscalls, then combine what we put in the
4304          * __augmented_syscalls__ BPF map with what is in the
4305          * syscalls:sys_exit_FOO tracepoints, i.e. just like we do without BPF,
4306          * combining raw_syscalls:sys_enter with raw_syscalls:sys_exit.
4307          *
4308          * We'll switch to look at two BPF maps, one for sys_enter and the
4309          * other for sys_exit when we start augmenting the sys_exit paths with
4310          * buffers that are being copied from kernel to userspace, think 'read'
4311          * syscall.
4312          */
4313         if (trace.syscalls.events.augmented) {
4314                 evlist__for_each_entry(trace.evlist, evsel) {
4315                         bool raw_syscalls_sys_exit = strcmp(perf_evsel__name(evsel), "raw_syscalls:sys_exit") == 0;
4316
4317                         if (raw_syscalls_sys_exit) {
4318                                 trace.raw_augmented_syscalls = true;
4319                                 goto init_augmented_syscall_tp;
4320                         }
4321
4322                         if (trace.syscalls.events.augmented->priv == NULL &&
4323                             strstr(perf_evsel__name(evsel), "syscalls:sys_enter")) {
4324                                 struct evsel *augmented = trace.syscalls.events.augmented;
4325                                 if (perf_evsel__init_augmented_syscall_tp(augmented, evsel) ||
4326                                     perf_evsel__init_augmented_syscall_tp_args(augmented))
4327                                         goto out;
4328                                 /*
4329                                  * Augmented is __augmented_syscalls__ BPF_OUTPUT event
4330                                  * Above we made sure we can get from the payload the tp fields
4331                                  * that we get from syscalls:sys_enter tracefs format file.
4332                                  */
4333                                 augmented->handler = trace__sys_enter;
4334                                 /*
4335                                  * Now we do the same for the *syscalls:sys_enter event so that
4336                                  * if we handle it directly, i.e. if the BPF prog returns 0 so
4337                                  * as not to filter it, then we'll handle it just like we would
4338                                  * for the BPF_OUTPUT one:
4339                                  */
4340                                 if (perf_evsel__init_augmented_syscall_tp(evsel, evsel) ||
4341                                     perf_evsel__init_augmented_syscall_tp_args(evsel))
4342                                         goto out;
4343                                 evsel->handler = trace__sys_enter;
4344                         }
4345
4346                         if (strstarts(perf_evsel__name(evsel), "syscalls:sys_exit_")) {
4347                                 struct syscall_tp *sc;
4348 init_augmented_syscall_tp:
4349                                 if (perf_evsel__init_augmented_syscall_tp(evsel, evsel))
4350                                         goto out;
4351                                 sc = evsel->priv;
4352                                 /*
4353                                  * For now with BPF raw_augmented we hook into
4354                                  * raw_syscalls:sys_enter and there we get all
4355                                  * 6 syscall args plus the tracepoint common
4356                                  * fields and the syscall_nr (another long).
4357                                  * So we check if that is the case and if so
4358                                  * don't look after the sc->args_size but
4359                                  * always after the full raw_syscalls:sys_enter
4360                                  * payload, which is fixed.
4361                                  *
4362                                  * We'll revisit this later to pass
4363                                  * s->args_size to the BPF augmenter (now
4364                                  * tools/perf/examples/bpf/augmented_raw_syscalls.c,
4365                                  * so that it copies only what we need for each
4366                                  * syscall, like what happens when we use
4367                                  * syscalls:sys_enter_NAME, so that we reduce
4368                                  * the kernel/userspace traffic to just what is
4369                                  * needed for each syscall.
4370                                  */
4371                                 if (trace.raw_augmented_syscalls)
4372                                         trace.raw_augmented_syscalls_args_size = (6 + 1) * sizeof(long) + sc->id.offset;
4373                                 perf_evsel__init_augmented_syscall_tp_ret(evsel);
4374                                 evsel->handler = trace__sys_exit;
4375                         }
4376                 }
4377         }
4378
4379         if ((argc >= 1) && (strcmp(argv[0], "record") == 0))
4380                 return trace__record(&trace, argc-1, &argv[1]);
4381
4382         /* summary_only implies summary option, but don't overwrite summary if set */
4383         if (trace.summary_only)
4384                 trace.summary = trace.summary_only;
4385
4386         if (!trace.trace_syscalls && !trace.trace_pgfaults &&
4387             trace.evlist->core.nr_entries == 0 /* Was --events used? */) {
4388                 trace.trace_syscalls = true;
4389         }
4390
4391         if (output_name != NULL) {
4392                 err = trace__open_output(&trace, output_name);
4393                 if (err < 0) {
4394                         perror("failed to create output file");
4395                         goto out;
4396                 }
4397         }
4398
4399         err = evswitch__init(&trace.evswitch, trace.evlist, stderr);
4400         if (err)
4401                 goto out_close;
4402
4403         err = target__validate(&trace.opts.target);
4404         if (err) {
4405                 target__strerror(&trace.opts.target, err, bf, sizeof(bf));
4406                 fprintf(trace.output, "%s", bf);
4407                 goto out_close;
4408         }
4409
4410         err = target__parse_uid(&trace.opts.target);
4411         if (err) {
4412                 target__strerror(&trace.opts.target, err, bf, sizeof(bf));
4413                 fprintf(trace.output, "%s", bf);
4414                 goto out_close;
4415         }
4416
4417         if (!argc && target__none(&trace.opts.target))
4418                 trace.opts.target.system_wide = true;
4419
4420         if (input_name)
4421                 err = trace__replay(&trace);
4422         else
4423                 err = trace__run(&trace, argc, argv);
4424
4425 out_close:
4426         if (output_name != NULL)
4427                 fclose(trace.output);
4428 out:
4429         return err;
4430 }