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 #include <perf/mmap.h>
  81
  82 #ifndef O_CLOEXEC
  83 # define O_CLOEXEC              02000000
  84 #endif
  85
  86 #ifndef F_LINUX_SPECIFIC_BASE
  87 # define F_LINUX_SPECIFIC_BASE  1024
  88 #endif
  89
  90 /*
  91  * strtoul: Go from a string to a value, i.e. for msr: MSR_FS_BASE to 0xc0000100
  92  */
  93 struct syscall_arg_fmt {
  94         size_t     (*scnprintf)(char *bf, size_t size, struct syscall_arg *arg);
  95         bool       (*strtoul)(char *bf, size_t size, struct syscall_arg *arg, u64 *val);
  96         unsigned long (*mask_val)(struct syscall_arg *arg, unsigned long val);
  97         void       *parm;
  98         const char *name;
  99         u16        nr_entries; // for arrays
 100         bool       show_zero;
 101 };
 102
 103 struct syscall_fmt {
 104         const char *name;
 105         const char *alias;
 106         struct {
 107                 const char *sys_enter,
 108                            *sys_exit;
 109         }          bpf_prog_name;
 110         struct syscall_arg_fmt arg[6];
 111         u8         nr_args;
 112         bool       errpid;
 113         bool       timeout;
 114         bool       hexret;
 115 };
 116
 117 struct trace {
 118         struct perf_tool        tool;
 119         struct syscalltbl       *sctbl;
 120         struct {
 121                 struct syscall  *table;
 122                 struct bpf_map  *map;
 123                 struct { // per syscall BPF_MAP_TYPE_PROG_ARRAY
 124                         struct bpf_map  *sys_enter,
 125                                         *sys_exit;
 126                 }               prog_array;
 127                 struct {
 128                         struct evsel *sys_enter,
 129                                           *sys_exit,
 130                                           *augmented;
 131                 }               events;
 132                 struct bpf_program *unaugmented_prog;
 133         } syscalls;
 134         struct {
 135                 struct bpf_map *map;
 136         } dump;
 137         struct record_opts      opts;
 138         struct evlist   *evlist;
 139         struct machine          *host;
 140         struct thread           *current;
 141         struct bpf_object       *bpf_obj;
 142         struct cgroup           *cgroup;
 143         u64                     base_time;
 144         FILE                    *output;
 145         unsigned long           nr_events;
 146         unsigned long           nr_events_printed;
 147         unsigned long           max_events;
 148         struct evswitch         evswitch;
 149         struct strlist          *ev_qualifier;
 150         struct {
 151                 size_t          nr;
 152                 int             *entries;
 153         }                       ev_qualifier_ids;
 154         struct {
 155                 size_t          nr;
 156                 pid_t           *entries;
 157                 struct bpf_map  *map;
 158         }                       filter_pids;
 159         double                  duration_filter;
 160         double                  runtime_ms;
 161         struct {
 162                 u64             vfs_getname,
 163                                 proc_getname;
 164         } stats;
 165         unsigned int            max_stack;
 166         unsigned int            min_stack;
 167         int                     raw_augmented_syscalls_args_size;
 168         bool                    raw_augmented_syscalls;
 169         bool                    fd_path_disabled;
 170         bool                    sort_events;
 171         bool                    not_ev_qualifier;
 172         bool                    live;
 173         bool                    full_time;
 174         bool                    sched;
 175         bool                    multiple_threads;
 176         bool                    summary;
 177         bool                    summary_only;
 178         bool                    errno_summary;
 179         bool                    failure_only;
 180         bool                    show_comm;
 181         bool                    print_sample;
 182         bool                    show_tool_stats;
 183         bool                    trace_syscalls;
 184         bool                    libtraceevent_print;
 185         bool                    kernel_syscallchains;
 186         s16                     args_alignment;
 187         bool                    show_tstamp;
 188         bool                    show_duration;
 189         bool                    show_zeros;
 190         bool                    show_arg_names;
 191         bool                    show_string_prefix;
 192         bool                    force;
 193         bool                    vfs_getname;
 194         int                     trace_pgfaults;
 195         char                    *perfconfig_events;
 196         struct {
 197                 struct ordered_events   data;
 198                 u64                     last;
 199         } oe;
 200 };
 201
 202 struct tp_field {
 203         int offset;
 204         union {
 205                 u64 (*integer)(struct tp_field *field, struct perf_sample *sample);
 206                 void *(*pointer)(struct tp_field *field, struct perf_sample *sample);
 207         };
 208 };
 209
 210 #define TP_UINT_FIELD(bits) \
 211 static u64 tp_field__u##bits(struct tp_field *field, struct perf_sample *sample) \
 212 { \
 213         u##bits value; \
 214         memcpy(&value, sample->raw_data + field->offset, sizeof(value)); \
 215         return value;  \
 216 }
 217
 218 TP_UINT_FIELD(8);
 219 TP_UINT_FIELD(16);
 220 TP_UINT_FIELD(32);
 221 TP_UINT_FIELD(64);
 222
 223 #define TP_UINT_FIELD__SWAPPED(bits) \
 224 static u64 tp_field__swapped_u##bits(struct tp_field *field, struct perf_sample *sample) \
 225 { \
 226         u##bits value; \
 227         memcpy(&value, sample->raw_data + field->offset, sizeof(value)); \
 228         return bswap_##bits(value);\
 229 }
 230
 231 TP_UINT_FIELD__SWAPPED(16);
 232 TP_UINT_FIELD__SWAPPED(32);
 233 TP_UINT_FIELD__SWAPPED(64);
 234
 235 static int __tp_field__init_uint(struct tp_field *field, int size, int offset, bool needs_swap)
 236 {
 237         field->offset = offset;
 238
 239         switch (size) {
 240         case 1:
 241                 field->integer = tp_field__u8;
 242                 break;
 243         case 2:
 244                 field->integer = needs_swap ? tp_field__swapped_u16 : tp_field__u16;
 245                 break;
 246         case 4:
 247                 field->integer = needs_swap ? tp_field__swapped_u32 : tp_field__u32;
 248                 break;
 249         case 8:
 250                 field->integer = needs_swap ? tp_field__swapped_u64 : tp_field__u64;
 251                 break;
 252         default:
 253                 return -1;
 254         }
 255
 256         return 0;
 257 }
 258
 259 static int tp_field__init_uint(struct tp_field *field, struct tep_format_field *format_field, bool needs_swap)
 260 {
 261         return __tp_field__init_uint(field, format_field->size, format_field->offset, needs_swap);
 262 }
 263
 264 static void *tp_field__ptr(struct tp_field *field, struct perf_sample *sample)
 265 {
 266         return sample->raw_data + field->offset;
 267 }
 268
 269 static int __tp_field__init_ptr(struct tp_field *field, int offset)
 270 {
 271         field->offset = offset;
 272         field->pointer = tp_field__ptr;
 273         return 0;
 274 }
 275
 276 static int tp_field__init_ptr(struct tp_field *field, struct tep_format_field *format_field)
 277 {
 278         return __tp_field__init_ptr(field, format_field->offset);
 279 }
 280
 281 struct syscall_tp {
 282         struct tp_field id;
 283         union {
 284                 struct tp_field args, ret;
 285         };
 286 };
 287
 288 /*
 289  * The evsel->priv as used by 'perf trace'
 290  * sc:  for raw_syscalls:sys_{enter,exit} and syscalls:sys_{enter,exit}_SYSCALLNAME
 291  * fmt: for all the other tracepoints
 292  */
 293 struct evsel_trace {
 294         struct syscall_tp       sc;
 295         struct syscall_arg_fmt  *fmt;
 296 };
 297
 298 static struct evsel_trace *evsel_trace__new(void)
 299 {
 300         return zalloc(sizeof(struct evsel_trace));
 301 }
 302
 303 static void evsel_trace__delete(struct evsel_trace *et)
 304 {
 305         if (et == NULL)
 306                 return;
 307
 308         zfree(&et->fmt);
 309         free(et);
 310 }
 311
 312 /*
 313  * Used with raw_syscalls:sys_{enter,exit} and with the
 314  * syscalls:sys_{enter,exit}_SYSCALL tracepoints
 315  */
 316 static inline struct syscall_tp *__evsel__syscall_tp(struct evsel *evsel)
 317 {
 318         struct evsel_trace *et = evsel->priv;
 319
 320         return &et->sc;
 321 }
 322
 323 static struct syscall_tp *evsel__syscall_tp(struct evsel *evsel)
 324 {
 325         if (evsel->priv == NULL) {
 326                 evsel->priv = evsel_trace__new();
 327                 if (evsel->priv == NULL)
 328                         return NULL;
 329         }
 330
 331         return __evsel__syscall_tp(evsel);
 332 }
 333
 334 /*
 335  * Used with all the other tracepoints.
 336  */
 337 static inline struct syscall_arg_fmt *__evsel__syscall_arg_fmt(struct evsel *evsel)
 338 {
 339         struct evsel_trace *et = evsel->priv;
 340
 341         return et->fmt;
 342 }
 343
 344 static struct syscall_arg_fmt *evsel__syscall_arg_fmt(struct evsel *evsel)
 345 {
 346         struct evsel_trace *et = evsel->priv;
 347
 348         if (evsel->priv == NULL) {
 349                 et = evsel->priv = evsel_trace__new();
 350
 351                 if (et == NULL)
 352                         return NULL;
 353         }
 354
 355         if (et->fmt == NULL) {
 356                 et->fmt = calloc(evsel->tp_format->format.nr_fields, sizeof(struct syscall_arg_fmt));
 357                 if (et->fmt == NULL)
 358                         goto out_delete;
 359         }
 360
 361         return __evsel__syscall_arg_fmt(evsel);
 362
 363 out_delete:
 364         evsel_trace__delete(evsel->priv);
 365         evsel->priv = NULL;
 366         return NULL;
 367 }
 368
 369 static int evsel__init_tp_uint_field(struct evsel *evsel, struct tp_field *field, const char *name)
 370 {
 371         struct tep_format_field *format_field = evsel__field(evsel, name);
 372
 373         if (format_field == NULL)
 374                 return -1;
 375
 376         return tp_field__init_uint(field, format_field, evsel->needs_swap);
 377 }
 378
 379 #define perf_evsel__init_sc_tp_uint_field(evsel, name) \
 380         ({ struct syscall_tp *sc = __evsel__syscall_tp(evsel);\
 381            evsel__init_tp_uint_field(evsel, &sc->name, #name); })
 382
 383 static int evsel__init_tp_ptr_field(struct evsel *evsel, struct tp_field *field, const char *name)
 384 {
 385         struct tep_format_field *format_field = evsel__field(evsel, name);
 386
 387         if (format_field == NULL)
 388                 return -1;
 389
 390         return tp_field__init_ptr(field, format_field);
 391 }
 392
 393 #define perf_evsel__init_sc_tp_ptr_field(evsel, name) \
 394         ({ struct syscall_tp *sc = __evsel__syscall_tp(evsel);\
 395            evsel__init_tp_ptr_field(evsel, &sc->name, #name); })
 396
 397 static void evsel__delete_priv(struct evsel *evsel)
 398 {
 399         zfree(&evsel->priv);
 400         evsel__delete(evsel);
 401 }
 402
 403 static int evsel__init_syscall_tp(struct evsel *evsel)
 404 {
 405         struct syscall_tp *sc = evsel__syscall_tp(evsel);
 406
 407         if (sc != NULL) {
 408                 if (evsel__init_tp_uint_field(evsel, &sc->id, "__syscall_nr") &&
 409                     evsel__init_tp_uint_field(evsel, &sc->id, "nr"))
 410                         return -ENOENT;
 411                 return 0;
 412         }
 413
 414         return -ENOMEM;
 415 }
 416
 417 static int evsel__init_augmented_syscall_tp(struct evsel *evsel, struct evsel *tp)
 418 {
 419         struct syscall_tp *sc = evsel__syscall_tp(evsel);
 420
 421         if (sc != NULL) {
 422                 struct tep_format_field *syscall_id = evsel__field(tp, "id");
 423                 if (syscall_id == NULL)
 424                         syscall_id = evsel__field(tp, "__syscall_nr");
 425                 if (syscall_id == NULL ||
 426                     __tp_field__init_uint(&sc->id, syscall_id->size, syscall_id->offset, evsel->needs_swap))
 427                         return -EINVAL;
 428
 429                 return 0;
 430         }
 431
 432         return -ENOMEM;
 433 }
 434
 435 static int evsel__init_augmented_syscall_tp_args(struct evsel *evsel)
 436 {
 437         struct syscall_tp *sc = __evsel__syscall_tp(evsel);
 438
 439         return __tp_field__init_ptr(&sc->args, sc->id.offset + sizeof(u64));
 440 }
 441
 442 static int evsel__init_augmented_syscall_tp_ret(struct evsel *evsel)
 443 {
 444         struct syscall_tp *sc = __evsel__syscall_tp(evsel);
 445
 446         return __tp_field__init_uint(&sc->ret, sizeof(u64), sc->id.offset + sizeof(u64), evsel->needs_swap);
 447 }
 448
 449 static int evsel__init_raw_syscall_tp(struct evsel *evsel, void *handler)
 450 {
 451         if (evsel__syscall_tp(evsel) != NULL) {
 452                 if (perf_evsel__init_sc_tp_uint_field(evsel, id))
 453                         return -ENOENT;
 454
 455                 evsel->handler = handler;
 456                 return 0;
 457         }
 458
 459         return -ENOMEM;
 460 }
 461
 462 static struct evsel *perf_evsel__raw_syscall_newtp(const char *direction, void *handler)
 463 {
 464         struct evsel *evsel = evsel__newtp("raw_syscalls", direction);
 465
 466         /* older kernel (e.g., RHEL6) use syscalls:{enter,exit} */
 467         if (IS_ERR(evsel))
 468                 evsel = evsel__newtp("syscalls", direction);
 469
 470         if (IS_ERR(evsel))
 471                 return NULL;
 472
 473         if (evsel__init_raw_syscall_tp(evsel, handler))
 474                 goto out_delete;
 475
 476         return evsel;
 477
 478 out_delete:
 479         evsel__delete_priv(evsel);
 480         return NULL;
 481 }
 482
 483 #define perf_evsel__sc_tp_uint(evsel, name, sample) \
 484         ({ struct syscall_tp *fields = __evsel__syscall_tp(evsel); \
 485            fields->name.integer(&fields->name, sample); })
 486
 487 #define perf_evsel__sc_tp_ptr(evsel, name, sample) \
 488         ({ struct syscall_tp *fields = __evsel__syscall_tp(evsel); \
 489            fields->name.pointer(&fields->name, sample); })
 490
 491 size_t strarray__scnprintf_suffix(struct strarray *sa, char *bf, size_t size, const char *intfmt, bool show_suffix, int val)
 492 {
 493         int idx = val - sa->offset;
 494
 495         if (idx < 0 || idx >= sa->nr_entries || sa->entries[idx] == NULL) {
 496                 size_t printed = scnprintf(bf, size, intfmt, val);
 497                 if (show_suffix)
 498                         printed += scnprintf(bf + printed, size - printed, " /* %s??? */", sa->prefix);
 499                 return printed;
 500         }
 501
 502         return scnprintf(bf, size, "%s%s", sa->entries[idx], show_suffix ? sa->prefix : "");
 503 }
 504
 505 size_t strarray__scnprintf(struct strarray *sa, char *bf, size_t size, const char *intfmt, bool show_prefix, int val)
 506 {
 507         int idx = val - sa->offset;
 508
 509         if (idx < 0 || idx >= sa->nr_entries || sa->entries[idx] == NULL) {
 510                 size_t printed = scnprintf(bf, size, intfmt, val);
 511                 if (show_prefix)
 512                         printed += scnprintf(bf + printed, size - printed, " /* %s??? */", sa->prefix);
 513                 return printed;
 514         }
 515
 516         return scnprintf(bf, size, "%s%s", show_prefix ? sa->prefix : "", sa->entries[idx]);
 517 }
 518
 519 static size_t __syscall_arg__scnprintf_strarray(char *bf, size_t size,
 520                                                 const char *intfmt,
 521                                                 struct syscall_arg *arg)
 522 {
 523         return strarray__scnprintf(arg->parm, bf, size, intfmt, arg->show_string_prefix, arg->val);
 524 }
 525
 526 static size_t syscall_arg__scnprintf_strarray(char *bf, size_t size,
 527                                               struct syscall_arg *arg)
 528 {
 529         return __syscall_arg__scnprintf_strarray(bf, size, "%d", arg);
 530 }
 531
 532 #define SCA_STRARRAY syscall_arg__scnprintf_strarray
 533
 534 bool syscall_arg__strtoul_strarray(char *bf, size_t size, struct syscall_arg *arg, u64 *ret)
 535 {
 536         return strarray__strtoul(arg->parm, bf, size, ret);
 537 }
 538
 539 bool syscall_arg__strtoul_strarray_flags(char *bf, size_t size, struct syscall_arg *arg, u64 *ret)
 540 {
 541         return strarray__strtoul_flags(arg->parm, bf, size, ret);
 542 }
 543
 544 bool syscall_arg__strtoul_strarrays(char *bf, size_t size, struct syscall_arg *arg, u64 *ret)
 545 {
 546         return strarrays__strtoul(arg->parm, bf, size, ret);
 547 }
 548
 549 size_t syscall_arg__scnprintf_strarray_flags(char *bf, size_t size, struct syscall_arg *arg)
 550 {
 551         return strarray__scnprintf_flags(arg->parm, bf, size, arg->show_string_prefix, arg->val);
 552 }
 553
 554 size_t strarrays__scnprintf(struct strarrays *sas, char *bf, size_t size, const char *intfmt, bool show_prefix, int val)
 555 {
 556         size_t printed;
 557         int i;
 558
 559         for (i = 0; i < sas->nr_entries; ++i) {
 560                 struct strarray *sa = sas->entries[i];
 561                 int idx = val - sa->offset;
 562
 563                 if (idx >= 0 && idx < sa->nr_entries) {
 564                         if (sa->entries[idx] == NULL)
 565                                 break;
 566                         return scnprintf(bf, size, "%s%s", show_prefix ? sa->prefix : "", sa->entries[idx]);
 567                 }
 568         }
 569
 570         printed = scnprintf(bf, size, intfmt, val);
 571         if (show_prefix)
 572                 printed += scnprintf(bf + printed, size - printed, " /* %s??? */", sas->entries[0]->prefix);
 573         return printed;
 574 }
 575
 576 bool strarray__strtoul(struct strarray *sa, char *bf, size_t size, u64 *ret)
 577 {
 578         int i;
 579
 580         for (i = 0; i < sa->nr_entries; ++i) {
 581                 if (sa->entries[i] && strncmp(sa->entries[i], bf, size) == 0 && sa->entries[i][size] == '\0') {
 582                         *ret = sa->offset + i;
 583                         return true;
 584                 }
 585         }
 586
 587         return false;
 588 }
 589
 590 bool strarray__strtoul_flags(struct strarray *sa, char *bf, size_t size, u64 *ret)
 591 {
 592         u64 val = 0;
 593         char *tok = bf, *sep, *end;
 594
 595         *ret = 0;
 596
 597         while (size != 0) {
 598                 int toklen = size;
 599
 600                 sep = memchr(tok, '|', size);
 601                 if (sep != NULL) {
 602                         size -= sep - tok + 1;
 603
 604                         end = sep - 1;
 605                         while (end > tok && isspace(*end))
 606                                 --end;
 607
 608                         toklen = end - tok + 1;
 609                 }
 610
 611                 while (isspace(*tok))
 612                         ++tok;
 613
 614                 if (isalpha(*tok) || *tok == '_') {
 615                         if (!strarray__strtoul(sa, tok, toklen, &val))
 616                                 return false;
 617                 } else {
 618                         bool is_hexa = tok[0] == 0 && (tok[1] = 'x' || tok[1] == 'X');
 619
 620                         val = strtoul(tok, NULL, is_hexa ? 16 : 0);
 621                 }
 622
 623                 *ret |= (1 << (val - 1));
 624
 625                 if (sep == NULL)
 626                         break;
 627                 tok = sep + 1;
 628         }
 629
 630         return true;
 631 }
 632
 633 bool strarrays__strtoul(struct strarrays *sas, char *bf, size_t size, u64 *ret)
 634 {
 635         int i;
 636
 637         for (i = 0; i < sas->nr_entries; ++i) {
 638                 struct strarray *sa = sas->entries[i];
 639
 640                 if (strarray__strtoul(sa, bf, size, ret))
 641                         return true;
 642         }
 643
 644         return false;
 645 }
 646
 647 size_t syscall_arg__scnprintf_strarrays(char *bf, size_t size,
 648                                         struct syscall_arg *arg)
 649 {
 650         return strarrays__scnprintf(arg->parm, bf, size, "%d", arg->show_string_prefix, arg->val);
 651 }
 652
 653 #ifndef AT_FDCWD
 654 #define AT_FDCWD        -100
 655 #endif
 656
 657 static size_t syscall_arg__scnprintf_fd_at(char *bf, size_t size,
 658                                            struct syscall_arg *arg)
 659 {
 660         int fd = arg->val;
 661         const char *prefix = "AT_FD";
 662
 663         if (fd == AT_FDCWD)
 664                 return scnprintf(bf, size, "%s%s", arg->show_string_prefix ? prefix : "", "CWD");
 665
 666         return syscall_arg__scnprintf_fd(bf, size, arg);
 667 }
 668
 669 #define SCA_FDAT syscall_arg__scnprintf_fd_at
 670
 671 static size_t syscall_arg__scnprintf_close_fd(char *bf, size_t size,
 672                                               struct syscall_arg *arg);
 673
 674 #define SCA_CLOSE_FD syscall_arg__scnprintf_close_fd
 675
 676 size_t syscall_arg__scnprintf_hex(char *bf, size_t size, struct syscall_arg *arg)
 677 {
 678         return scnprintf(bf, size, "%#lx", arg->val);
 679 }
 680
 681 size_t syscall_arg__scnprintf_ptr(char *bf, size_t size, struct syscall_arg *arg)
 682 {
 683         if (arg->val == 0)
 684                 return scnprintf(bf, size, "NULL");
 685         return syscall_arg__scnprintf_hex(bf, size, arg);
 686 }
 687
 688 size_t syscall_arg__scnprintf_int(char *bf, size_t size, struct syscall_arg *arg)
 689 {
 690         return scnprintf(bf, size, "%d", arg->val);
 691 }
 692
 693 size_t syscall_arg__scnprintf_long(char *bf, size_t size, struct syscall_arg *arg)
 694 {
 695         return scnprintf(bf, size, "%ld", arg->val);
 696 }
 697
 698 static size_t syscall_arg__scnprintf_char_array(char *bf, size_t size, struct syscall_arg *arg)
 699 {
 700         // XXX Hey, maybe for sched:sched_switch prev/next comm fields we can
 701         //     fill missing comms using thread__set_comm()...
 702         //     here or in a special syscall_arg__scnprintf_pid_sched_tp...
 703         return scnprintf(bf, size, "\"%-.*s\"", arg->fmt->nr_entries ?: arg->len, arg->val);
 704 }
 705
 706 #define SCA_CHAR_ARRAY syscall_arg__scnprintf_char_array
 707
 708 static const char *bpf_cmd[] = {
 709         "MAP_CREATE", "MAP_LOOKUP_ELEM", "MAP_UPDATE_ELEM", "MAP_DELETE_ELEM",
 710         "MAP_GET_NEXT_KEY", "PROG_LOAD",
 711 };
 712 static DEFINE_STRARRAY(bpf_cmd, "BPF_");
 713
 714 static const char *fsmount_flags[] = {
 715         [1] = "CLOEXEC",
 716 };
 717 static DEFINE_STRARRAY(fsmount_flags, "FSMOUNT_");
 718
 719 #include "trace/beauty/generated/fsconfig_arrays.c"
 720
 721 static DEFINE_STRARRAY(fsconfig_cmds, "FSCONFIG_");
 722
 723 static const char *epoll_ctl_ops[] = { "ADD", "DEL", "MOD", };
 724 static DEFINE_STRARRAY_OFFSET(epoll_ctl_ops, "EPOLL_CTL_", 1);
 725
 726 static const char *itimers[] = { "REAL", "VIRTUAL", "PROF", };
 727 static DEFINE_STRARRAY(itimers, "ITIMER_");
 728
 729 static const char *keyctl_options[] = {
 730         "GET_KEYRING_ID", "JOIN_SESSION_KEYRING", "UPDATE", "REVOKE", "CHOWN",
 731         "SETPERM", "DESCRIBE", "CLEAR", "LINK", "UNLINK", "SEARCH", "READ",
 732         "INSTANTIATE", "NEGATE", "SET_REQKEY_KEYRING", "SET_TIMEOUT",
 733         "ASSUME_AUTHORITY", "GET_SECURITY", "SESSION_TO_PARENT", "REJECT",
 734         "INSTANTIATE_IOV", "INVALIDATE", "GET_PERSISTENT",
 735 };
 736 static DEFINE_STRARRAY(keyctl_options, "KEYCTL_");
 737
 738 static const char *whences[] = { "SET", "CUR", "END",
 739 #ifdef SEEK_DATA
 740 "DATA",
 741 #endif
 742 #ifdef SEEK_HOLE
 743 "HOLE",
 744 #endif
 745 };
 746 static DEFINE_STRARRAY(whences, "SEEK_");
 747
 748 static const char *fcntl_cmds[] = {
 749         "DUPFD", "GETFD", "SETFD", "GETFL", "SETFL", "GETLK", "SETLK",
 750         "SETLKW", "SETOWN", "GETOWN", "SETSIG", "GETSIG", "GETLK64",
 751         "SETLK64", "SETLKW64", "SETOWN_EX", "GETOWN_EX",
 752         "GETOWNER_UIDS",
 753 };
 754 static DEFINE_STRARRAY(fcntl_cmds, "F_");
 755
 756 static const char *fcntl_linux_specific_cmds[] = {
 757         "SETLEASE", "GETLEASE", "NOTIFY", [5] = "CANCELLK", "DUPFD_CLOEXEC",
 758         "SETPIPE_SZ", "GETPIPE_SZ", "ADD_SEALS", "GET_SEALS",
 759         "GET_RW_HINT", "SET_RW_HINT", "GET_FILE_RW_HINT", "SET_FILE_RW_HINT",
 760 };
 761
 762 static DEFINE_STRARRAY_OFFSET(fcntl_linux_specific_cmds, "F_", F_LINUX_SPECIFIC_BASE);
 763
 764 static struct strarray *fcntl_cmds_arrays[] = {
 765         &strarray__fcntl_cmds,
 766         &strarray__fcntl_linux_specific_cmds,
 767 };
 768
 769 static DEFINE_STRARRAYS(fcntl_cmds_arrays);
 770
 771 static const char *rlimit_resources[] = {
 772         "CPU", "FSIZE", "DATA", "STACK", "CORE", "RSS", "NPROC", "NOFILE",
 773         "MEMLOCK", "AS", "LOCKS", "SIGPENDING", "MSGQUEUE", "NICE", "RTPRIO",
 774         "RTTIME",
 775 };
 776 static DEFINE_STRARRAY(rlimit_resources, "RLIMIT_");
 777
 778 static const char *sighow[] = { "BLOCK", "UNBLOCK", "SETMASK", };
 779 static DEFINE_STRARRAY(sighow, "SIG_");
 780
 781 static const char *clockid[] = {
 782         "REALTIME", "MONOTONIC", "PROCESS_CPUTIME_ID", "THREAD_CPUTIME_ID",
 783         "MONOTONIC_RAW", "REALTIME_COARSE", "MONOTONIC_COARSE", "BOOTTIME",
 784         "REALTIME_ALARM", "BOOTTIME_ALARM", "SGI_CYCLE", "TAI"
 785 };
 786 static DEFINE_STRARRAY(clockid, "CLOCK_");
 787
 788 static size_t syscall_arg__scnprintf_access_mode(char *bf, size_t size,
 789                                                  struct syscall_arg *arg)
 790 {
 791         bool show_prefix = arg->show_string_prefix;
 792         const char *suffix = "_OK";
 793         size_t printed = 0;
 794         int mode = arg->val;
 795
 796         if (mode == F_OK) /* 0 */
 797                 return scnprintf(bf, size, "F%s", show_prefix ? suffix : "");
 798 #define P_MODE(n) \
 799         if (mode & n##_OK) { \
 800                 printed += scnprintf(bf + printed, size - printed, "%s%s", #n, show_prefix ? suffix : ""); \
 801                 mode &= ~n##_OK; \
 802         }
 803
 804         P_MODE(R);
 805         P_MODE(W);
 806         P_MODE(X);
 807 #undef P_MODE
 808
 809         if (mode)
 810                 printed += scnprintf(bf + printed, size - printed, "|%#x", mode);
 811
 812         return printed;
 813 }
 814
 815 #define SCA_ACCMODE syscall_arg__scnprintf_access_mode
 816
 817 static size_t syscall_arg__scnprintf_filename(char *bf, size_t size,
 818                                               struct syscall_arg *arg);
 819
 820 #define SCA_FILENAME syscall_arg__scnprintf_filename
 821
 822 static size_t syscall_arg__scnprintf_pipe_flags(char *bf, size_t size,
 823                                                 struct syscall_arg *arg)
 824 {
 825         bool show_prefix = arg->show_string_prefix;
 826         const char *prefix = "O_";
 827         int printed = 0, flags = arg->val;
 828
 829 #define P_FLAG(n) \
 830         if (flags & O_##n) { \
 831                 printed += scnprintf(bf + printed, size - printed, "%s%s%s", printed ? "|" : "", show_prefix ? prefix : "", #n); \
 832                 flags &= ~O_##n; \
 833         }
 834
 835         P_FLAG(CLOEXEC);
 836         P_FLAG(NONBLOCK);
 837 #undef P_FLAG
 838
 839         if (flags)
 840                 printed += scnprintf(bf + printed, size - printed, "%s%#x", printed ? "|" : "", flags);
 841
 842         return printed;
 843 }
 844
 845 #define SCA_PIPE_FLAGS syscall_arg__scnprintf_pipe_flags
 846
 847 #ifndef GRND_NONBLOCK
 848 #define GRND_NONBLOCK   0x0001
 849 #endif
 850 #ifndef GRND_RANDOM
 851 #define GRND_RANDOM     0x0002
 852 #endif
 853
 854 static size_t syscall_arg__scnprintf_getrandom_flags(char *bf, size_t size,
 855                                                    struct syscall_arg *arg)
 856 {
 857         bool show_prefix = arg->show_string_prefix;
 858         const char *prefix = "GRND_";
 859         int printed = 0, flags = arg->val;
 860
 861 #define P_FLAG(n) \
 862         if (flags & GRND_##n) { \
 863                 printed += scnprintf(bf + printed, size - printed, "%s%s%s", printed ? "|" : "", show_prefix ? prefix : "", #n); \
 864                 flags &= ~GRND_##n; \
 865         }
 866
 867         P_FLAG(RANDOM);
 868         P_FLAG(NONBLOCK);
 869 #undef P_FLAG
 870
 871         if (flags)
 872                 printed += scnprintf(bf + printed, size - printed, "%s%#x", printed ? "|" : "", flags);
 873
 874         return printed;
 875 }
 876
 877 #define SCA_GETRANDOM_FLAGS syscall_arg__scnprintf_getrandom_flags
 878
 879 #define STRARRAY(name, array) \
 880           { .scnprintf  = SCA_STRARRAY, \
 881             .strtoul    = STUL_STRARRAY, \
 882             .parm       = &strarray__##array, }
 883
 884 #define STRARRAY_FLAGS(name, array) \
 885           { .scnprintf  = SCA_STRARRAY_FLAGS, \
 886             .strtoul    = STUL_STRARRAY_FLAGS, \
 887             .parm       = &strarray__##array, }
 888
 889 #include "trace/beauty/arch_errno_names.c"
 890 #include "trace/beauty/eventfd.c"
 891 #include "trace/beauty/futex_op.c"
 892 #include "trace/beauty/futex_val3.c"
 893 #include "trace/beauty/mmap.c"
 894 #include "trace/beauty/mode_t.c"
 895 #include "trace/beauty/msg_flags.c"
 896 #include "trace/beauty/open_flags.c"
 897 #include "trace/beauty/perf_event_open.c"
 898 #include "trace/beauty/pid.c"
 899 #include "trace/beauty/sched_policy.c"
 900 #include "trace/beauty/seccomp.c"
 901 #include "trace/beauty/signum.c"
 902 #include "trace/beauty/socket_type.c"
 903 #include "trace/beauty/waitid_options.c"
 904
 905 static struct syscall_fmt syscall_fmts[] = {
 906         { .name     = "access",
 907           .arg = { [1] = { .scnprintf = SCA_ACCMODE,  /* mode */ }, }, },
 908         { .name     = "arch_prctl",
 909           .arg = { [0] = { .scnprintf = SCA_X86_ARCH_PRCTL_CODE, /* code */ },
 910                    [1] = { .scnprintf = SCA_PTR, /* arg2 */ }, }, },
 911         { .name     = "bind",
 912           .arg = { [0] = { .scnprintf = SCA_INT, /* fd */ },
 913                    [1] = { .scnprintf = SCA_SOCKADDR, /* umyaddr */ },
 914                    [2] = { .scnprintf = SCA_INT, /* addrlen */ }, }, },
 915         { .name     = "bpf",
 916           .arg = { [0] = STRARRAY(cmd, bpf_cmd), }, },
 917         { .name     = "brk",        .hexret = true,
 918           .arg = { [0] = { .scnprintf = SCA_PTR, /* brk */ }, }, },
 919         { .name     = "clock_gettime",
 920           .arg = { [0] = STRARRAY(clk_id, clockid), }, },
 921         { .name     = "clone",      .errpid = true, .nr_args = 5,
 922           .arg = { [0] = { .name = "flags",         .scnprintf = SCA_CLONE_FLAGS, },
 923                    [1] = { .name = "child_stack",   .scnprintf = SCA_HEX, },
 924                    [2] = { .name = "parent_tidptr", .scnprintf = SCA_HEX, },
 925                    [3] = { .name = "child_tidptr",  .scnprintf = SCA_HEX, },
 926                    [4] = { .name = "tls",           .scnprintf = SCA_HEX, }, }, },
 927         { .name     = "close",
 928           .arg = { [0] = { .scnprintf = SCA_CLOSE_FD, /* fd */ }, }, },
 929         { .name     = "connect",
 930           .arg = { [0] = { .scnprintf = SCA_INT, /* fd */ },
 931                    [1] = { .scnprintf = SCA_SOCKADDR, /* servaddr */ },
 932                    [2] = { .scnprintf = SCA_INT, /* addrlen */ }, }, },
 933         { .name     = "epoll_ctl",
 934           .arg = { [1] = STRARRAY(op, epoll_ctl_ops), }, },
 935         { .name     = "eventfd2",
 936           .arg = { [1] = { .scnprintf = SCA_EFD_FLAGS, /* flags */ }, }, },
 937         { .name     = "fchmodat",
 938           .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
 939         { .name     = "fchownat",
 940           .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
 941         { .name     = "fcntl",
 942           .arg = { [1] = { .scnprintf = SCA_FCNTL_CMD,  /* cmd */
 943                            .strtoul   = STUL_STRARRAYS,
 944                            .parm      = &strarrays__fcntl_cmds_arrays,
 945                            .show_zero = true, },
 946                    [2] = { .scnprintf =  SCA_FCNTL_ARG, /* arg */ }, }, },
 947         { .name     = "flock",
 948           .arg = { [1] = { .scnprintf = SCA_FLOCK, /* cmd */ }, }, },
 949         { .name     = "fsconfig",
 950           .arg = { [1] = STRARRAY(cmd, fsconfig_cmds), }, },
 951         { .name     = "fsmount",
 952           .arg = { [1] = STRARRAY_FLAGS(flags, fsmount_flags),
 953                    [2] = { .scnprintf = SCA_FSMOUNT_ATTR_FLAGS, /* attr_flags */ }, }, },
 954         { .name     = "fspick",
 955           .arg = { [0] = { .scnprintf = SCA_FDAT,         /* dfd */ },
 956                    [1] = { .scnprintf = SCA_FILENAME,     /* path */ },
 957                    [2] = { .scnprintf = SCA_FSPICK_FLAGS, /* flags */ }, }, },
 958         { .name     = "fstat", .alias = "newfstat", },
 959         { .name     = "fstatat", .alias = "newfstatat", },
 960         { .name     = "futex",
 961           .arg = { [1] = { .scnprintf = SCA_FUTEX_OP, /* op */ },
 962                    [5] = { .scnprintf = SCA_FUTEX_VAL3, /* val3 */ }, }, },
 963         { .name     = "futimesat",
 964           .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
 965         { .name     = "getitimer",
 966           .arg = { [0] = STRARRAY(which, itimers), }, },
 967         { .name     = "getpid",     .errpid = true, },
 968         { .name     = "getpgid",    .errpid = true, },
 969         { .name     = "getppid",    .errpid = true, },
 970         { .name     = "getrandom",
 971           .arg = { [2] = { .scnprintf = SCA_GETRANDOM_FLAGS, /* flags */ }, }, },
 972         { .name     = "getrlimit",
 973           .arg = { [0] = STRARRAY(resource, rlimit_resources), }, },
 974         { .name     = "gettid",     .errpid = true, },
 975         { .name     = "ioctl",
 976           .arg = {
 977 #if defined(__i386__) || defined(__x86_64__)
 978 /*
 979  * FIXME: Make this available to all arches.
 980  */
 981                    [1] = { .scnprintf = SCA_IOCTL_CMD, /* cmd */ },
 982                    [2] = { .scnprintf = SCA_HEX, /* arg */ }, }, },
 983 #else
 984                    [2] = { .scnprintf = SCA_HEX, /* arg */ }, }, },
 985 #endif
 986         { .name     = "kcmp",       .nr_args = 5,
 987           .arg = { [0] = { .name = "pid1",      .scnprintf = SCA_PID, },
 988                    [1] = { .name = "pid2",      .scnprintf = SCA_PID, },
 989                    [2] = { .name = "type",      .scnprintf = SCA_KCMP_TYPE, },
 990                    [3] = { .name = "idx1",      .scnprintf = SCA_KCMP_IDX, },
 991                    [4] = { .name = "idx2",      .scnprintf = SCA_KCMP_IDX, }, }, },
 992         { .name     = "keyctl",
 993           .arg = { [0] = STRARRAY(option, keyctl_options), }, },
 994         { .name     = "kill",
 995           .arg = { [1] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
 996         { .name     = "linkat",
 997           .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
 998         { .name     = "lseek",
 999           .arg = { [2] = STRARRAY(whence, whences), }, },
1000         { .name     = "lstat", .alias = "newlstat", },
1001         { .name     = "madvise",
1002           .arg = { [0] = { .scnprintf = SCA_HEX,      /* start */ },
1003                    [2] = { .scnprintf = SCA_MADV_BHV, /* behavior */ }, }, },
1004         { .name     = "mkdirat",
1005           .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
1006         { .name     = "mknodat",
1007           .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
1008         { .name     = "mmap",       .hexret = true,
1009 /* The standard mmap maps to old_mmap on s390x */
1010 #if defined(__s390x__)
1011         .alias = "old_mmap",
1012 #endif
1013           .arg = { [2] = { .scnprintf = SCA_MMAP_PROT,  /* prot */ },
1014                    [3] = { .scnprintf = SCA_MMAP_FLAGS, /* flags */
1015                            .strtoul   = STUL_STRARRAY_FLAGS,
1016                            .parm      = &strarray__mmap_flags, },
1017                    [5] = { .scnprintf = SCA_HEX,        /* offset */ }, }, },
1018         { .name     = "mount",
1019           .arg = { [0] = { .scnprintf = SCA_FILENAME, /* dev_name */ },
1020                    [3] = { .scnprintf = SCA_MOUNT_FLAGS, /* flags */
1021                            .mask_val  = SCAMV_MOUNT_FLAGS, /* flags */ }, }, },
1022         { .name     = "move_mount",
1023           .arg = { [0] = { .scnprintf = SCA_FDAT,       /* from_dfd */ },
1024                    [1] = { .scnprintf = SCA_FILENAME, /* from_pathname */ },
1025                    [2] = { .scnprintf = SCA_FDAT,       /* to_dfd */ },
1026                    [3] = { .scnprintf = SCA_FILENAME, /* to_pathname */ },
1027                    [4] = { .scnprintf = SCA_MOVE_MOUNT_FLAGS, /* flags */ }, }, },
1028         { .name     = "mprotect",
1029           .arg = { [0] = { .scnprintf = SCA_HEX,        /* start */ },
1030                    [2] = { .scnprintf = SCA_MMAP_PROT,  /* prot */ }, }, },
1031         { .name     = "mq_unlink",
1032           .arg = { [0] = { .scnprintf = SCA_FILENAME, /* u_name */ }, }, },
1033         { .name     = "mremap",     .hexret = true,
1034           .arg = { [3] = { .scnprintf = SCA_MREMAP_FLAGS, /* flags */ }, }, },
1035         { .name     = "name_to_handle_at",
1036           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
1037         { .name     = "newfstatat",
1038           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
1039         { .name     = "open",
1040           .arg = { [1] = { .scnprintf = SCA_OPEN_FLAGS, /* flags */ }, }, },
1041         { .name     = "open_by_handle_at",
1042           .arg = { [0] = { .scnprintf = SCA_FDAT,       /* dfd */ },
1043                    [2] = { .scnprintf = SCA_OPEN_FLAGS, /* flags */ }, }, },
1044         { .name     = "openat",
1045           .arg = { [0] = { .scnprintf = SCA_FDAT,       /* dfd */ },
1046                    [2] = { .scnprintf = SCA_OPEN_FLAGS, /* flags */ }, }, },
1047         { .name     = "perf_event_open",
1048           .arg = { [2] = { .scnprintf = SCA_INT,        /* cpu */ },
1049                    [3] = { .scnprintf = SCA_FD,         /* group_fd */ },
1050                    [4] = { .scnprintf = SCA_PERF_FLAGS, /* flags */ }, }, },
1051         { .name     = "pipe2",
1052           .arg = { [1] = { .scnprintf = SCA_PIPE_FLAGS, /* flags */ }, }, },
1053         { .name     = "pkey_alloc",
1054           .arg = { [1] = { .scnprintf = SCA_PKEY_ALLOC_ACCESS_RIGHTS,   /* access_rights */ }, }, },
1055         { .name     = "pkey_free",
1056           .arg = { [0] = { .scnprintf = SCA_INT,        /* key */ }, }, },
1057         { .name     = "pkey_mprotect",
1058           .arg = { [0] = { .scnprintf = SCA_HEX,        /* start */ },
1059                    [2] = { .scnprintf = SCA_MMAP_PROT,  /* prot */ },
1060                    [3] = { .scnprintf = SCA_INT,        /* pkey */ }, }, },
1061         { .name     = "poll", .timeout = true, },
1062         { .name     = "ppoll", .timeout = true, },
1063         { .name     = "prctl",
1064           .arg = { [0] = { .scnprintf = SCA_PRCTL_OPTION, /* option */
1065                            .strtoul   = STUL_STRARRAY,
1066                            .parm      = &strarray__prctl_options, },
1067                    [1] = { .scnprintf = SCA_PRCTL_ARG2, /* arg2 */ },
1068                    [2] = { .scnprintf = SCA_PRCTL_ARG3, /* arg3 */ }, }, },
1069         { .name     = "pread", .alias = "pread64", },
1070         { .name     = "preadv", .alias = "pread", },
1071         { .name     = "prlimit64",
1072           .arg = { [1] = STRARRAY(resource, rlimit_resources), }, },
1073         { .name     = "pwrite", .alias = "pwrite64", },
1074         { .name     = "readlinkat",
1075           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
1076         { .name     = "recvfrom",
1077           .arg = { [3] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
1078         { .name     = "recvmmsg",
1079           .arg = { [3] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
1080         { .name     = "recvmsg",
1081           .arg = { [2] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
1082         { .name     = "renameat",
1083           .arg = { [0] = { .scnprintf = SCA_FDAT, /* olddirfd */ },
1084                    [2] = { .scnprintf = SCA_FDAT, /* newdirfd */ }, }, },
1085         { .name     = "renameat2",
1086           .arg = { [0] = { .scnprintf = SCA_FDAT, /* olddirfd */ },
1087                    [2] = { .scnprintf = SCA_FDAT, /* newdirfd */ },
1088                    [4] = { .scnprintf = SCA_RENAMEAT2_FLAGS, /* flags */ }, }, },
1089         { .name     = "rt_sigaction",
1090           .arg = { [0] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
1091         { .name     = "rt_sigprocmask",
1092           .arg = { [0] = STRARRAY(how, sighow), }, },
1093         { .name     = "rt_sigqueueinfo",
1094           .arg = { [1] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
1095         { .name     = "rt_tgsigqueueinfo",
1096           .arg = { [2] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
1097         { .name     = "sched_setscheduler",
1098           .arg = { [1] = { .scnprintf = SCA_SCHED_POLICY, /* policy */ }, }, },
1099         { .name     = "seccomp",
1100           .arg = { [0] = { .scnprintf = SCA_SECCOMP_OP,    /* op */ },
1101                    [1] = { .scnprintf = SCA_SECCOMP_FLAGS, /* flags */ }, }, },
1102         { .name     = "select", .timeout = true, },
1103         { .name     = "sendfile", .alias = "sendfile64", },
1104         { .name     = "sendmmsg",
1105           .arg = { [3] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
1106         { .name     = "sendmsg",
1107           .arg = { [2] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
1108         { .name     = "sendto",
1109           .arg = { [3] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ },
1110                    [4] = { .scnprintf = SCA_SOCKADDR, /* addr */ }, }, },
1111         { .name     = "set_tid_address", .errpid = true, },
1112         { .name     = "setitimer",
1113           .arg = { [0] = STRARRAY(which, itimers), }, },
1114         { .name     = "setrlimit",
1115           .arg = { [0] = STRARRAY(resource, rlimit_resources), }, },
1116         { .name     = "socket",
1117           .arg = { [0] = STRARRAY(family, socket_families),
1118                    [1] = { .scnprintf = SCA_SK_TYPE, /* type */ },
1119                    [2] = { .scnprintf = SCA_SK_PROTO, /* protocol */ }, }, },
1120         { .name     = "socketpair",
1121           .arg = { [0] = STRARRAY(family, socket_families),
1122                    [1] = { .scnprintf = SCA_SK_TYPE, /* type */ },
1123                    [2] = { .scnprintf = SCA_SK_PROTO, /* protocol */ }, }, },
1124         { .name     = "stat", .alias = "newstat", },
1125         { .name     = "statx",
1126           .arg = { [0] = { .scnprintf = SCA_FDAT,        /* fdat */ },
1127                    [2] = { .scnprintf = SCA_STATX_FLAGS, /* flags */ } ,
1128                    [3] = { .scnprintf = SCA_STATX_MASK,  /* mask */ }, }, },
1129         { .name     = "swapoff",
1130           .arg = { [0] = { .scnprintf = SCA_FILENAME, /* specialfile */ }, }, },
1131         { .name     = "swapon",
1132           .arg = { [0] = { .scnprintf = SCA_FILENAME, /* specialfile */ }, }, },
1133         { .name     = "symlinkat",
1134           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
1135         { .name     = "sync_file_range",
1136           .arg = { [3] = { .scnprintf = SCA_SYNC_FILE_RANGE_FLAGS, /* flags */ }, }, },
1137         { .name     = "tgkill",
1138           .arg = { [2] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
1139         { .name     = "tkill",
1140           .arg = { [1] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
1141         { .name     = "umount2", .alias = "umount",
1142           .arg = { [0] = { .scnprintf = SCA_FILENAME, /* name */ }, }, },
1143         { .name     = "uname", .alias = "newuname", },
1144         { .name     = "unlinkat",
1145           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
1146         { .name     = "utimensat",
1147           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dirfd */ }, }, },
1148         { .name     = "wait4",      .errpid = true,
1149           .arg = { [2] = { .scnprintf = SCA_WAITID_OPTIONS, /* options */ }, }, },
1150         { .name     = "waitid",     .errpid = true,
1151           .arg = { [3] = { .scnprintf = SCA_WAITID_OPTIONS, /* options */ }, }, },
1152 };
1153
1154 static int syscall_fmt__cmp(const void *name, const void *fmtp)
1155 {
1156         const struct syscall_fmt *fmt = fmtp;
1157         return strcmp(name, fmt->name);
1158 }
1159
1160 static struct syscall_fmt *__syscall_fmt__find(struct syscall_fmt *fmts, const int nmemb, const char *name)
1161 {
1162         return bsearch(name, fmts, nmemb, sizeof(struct syscall_fmt), syscall_fmt__cmp);
1163 }
1164
1165 static struct syscall_fmt *syscall_fmt__find(const char *name)
1166 {
1167         const int nmemb = ARRAY_SIZE(syscall_fmts);
1168         return __syscall_fmt__find(syscall_fmts, nmemb, name);
1169 }
1170
1171 static struct syscall_fmt *__syscall_fmt__find_by_alias(struct syscall_fmt *fmts, const int nmemb, const char *alias)
1172 {
1173         int i;
1174
1175         for (i = 0; i < nmemb; ++i) {
1176                 if (fmts[i].alias && strcmp(fmts[i].alias, alias) == 0)
1177                         return &fmts[i];
1178         }
1179
1180         return NULL;
1181 }
1182
1183 static struct syscall_fmt *syscall_fmt__find_by_alias(const char *alias)
1184 {
1185         const int nmemb = ARRAY_SIZE(syscall_fmts);
1186         return __syscall_fmt__find_by_alias(syscall_fmts, nmemb, alias);
1187 }
1188
1189 /*
1190  * is_exit: is this "exit" or "exit_group"?
1191  * is_open: is this "open" or "openat"? To associate the fd returned in sys_exit with the pathname in sys_enter.
1192  * args_size: sum of the sizes of the syscall arguments, anything after that is augmented stuff: pathname for openat, etc.
1193  * nonexistent: Just a hole in the syscall table, syscall id not allocated
1194  */
1195 struct syscall {
1196         struct tep_event    *tp_format;
1197         int                 nr_args;
1198         int                 args_size;
1199         struct {
1200                 struct bpf_program *sys_enter,
1201                                    *sys_exit;
1202         }                   bpf_prog;
1203         bool                is_exit;
1204         bool                is_open;
1205         bool                nonexistent;
1206         struct tep_format_field *args;
1207         const char          *name;
1208         struct syscall_fmt  *fmt;
1209         struct syscall_arg_fmt *arg_fmt;
1210 };
1211
1212 /*
1213  * Must match what is in the BPF program:
1214  *
1215  * tools/perf/examples/bpf/augmented_raw_syscalls.c
1216  */
1217 struct bpf_map_syscall_entry {
1218         bool    enabled;
1219         u16     string_args_len[6];
1220 };
1221
1222 /*
1223  * We need to have this 'calculated' boolean because in some cases we really
1224  * don't know what is the duration of a syscall, for instance, when we start
1225  * a session and some threads are waiting for a syscall to finish, say 'poll',
1226  * in which case all we can do is to print "( ? ) for duration and for the
1227  * start timestamp.
1228  */
1229 static size_t fprintf_duration(unsigned long t, bool calculated, FILE *fp)
1230 {
1231         double duration = (double)t / NSEC_PER_MSEC;
1232         size_t printed = fprintf(fp, "(");
1233
1234         if (!calculated)
1235                 printed += fprintf(fp, "         ");
1236         else if (duration >= 1.0)
1237                 printed += color_fprintf(fp, PERF_COLOR_RED, "%6.3f ms", duration);
1238         else if (duration >= 0.01)
1239                 printed += color_fprintf(fp, PERF_COLOR_YELLOW, "%6.3f ms", duration);
1240         else
1241                 printed += color_fprintf(fp, PERF_COLOR_NORMAL, "%6.3f ms", duration);
1242         return printed + fprintf(fp, "): ");
1243 }
1244
1245 /**
1246  * filename.ptr: The filename char pointer that will be vfs_getname'd
1247  * filename.entry_str_pos: Where to insert the string translated from
1248  *                         filename.ptr by the vfs_getname tracepoint/kprobe.
1249  * ret_scnprintf: syscall args may set this to a different syscall return
1250  *                formatter, for instance, fcntl may return fds, file flags, etc.
1251  */
1252 struct thread_trace {
1253         u64               entry_time;
1254         bool              entry_pending;
1255         unsigned long     nr_events;
1256         unsigned long     pfmaj, pfmin;
1257         char              *entry_str;
1258         double            runtime_ms;
1259         size_t            (*ret_scnprintf)(char *bf, size_t size, struct syscall_arg *arg);
1260         struct {
1261                 unsigned long ptr;
1262                 short int     entry_str_pos;
1263                 bool          pending_open;
1264                 unsigned int  namelen;
1265                 char          *name;
1266         } filename;
1267         struct {
1268                 int           max;
1269                 struct file   *table;
1270         } files;
1271
1272         struct intlist *syscall_stats;
1273 };
1274
1275 static struct thread_trace *thread_trace__new(void)
1276 {
1277         struct thread_trace *ttrace =  zalloc(sizeof(struct thread_trace));
1278
1279         if (ttrace) {
1280                 ttrace->files.max = -1;
1281                 ttrace->syscall_stats = intlist__new(NULL);
1282         }
1283
1284         return ttrace;
1285 }
1286
1287 static struct thread_trace *thread__trace(struct thread *thread, FILE *fp)
1288 {
1289         struct thread_trace *ttrace;
1290
1291         if (thread == NULL)
1292                 goto fail;
1293
1294         if (thread__priv(thread) == NULL)
1295                 thread__set_priv(thread, thread_trace__new());
1296
1297         if (thread__priv(thread) == NULL)
1298                 goto fail;
1299
1300         ttrace = thread__priv(thread);
1301         ++ttrace->nr_events;
1302
1303         return ttrace;
1304 fail:
1305         color_fprintf(fp, PERF_COLOR_RED,
1306                       "WARNING: not enough memory, dropping samples!\n");
1307         return NULL;
1308 }
1309
1310
1311 void syscall_arg__set_ret_scnprintf(struct syscall_arg *arg,
1312                                     size_t (*ret_scnprintf)(char *bf, size_t size, struct syscall_arg *arg))
1313 {
1314         struct thread_trace *ttrace = thread__priv(arg->thread);
1315
1316         ttrace->ret_scnprintf = ret_scnprintf;
1317 }
1318
1319 #define TRACE_PFMAJ             (1 << 0)
1320 #define TRACE_PFMIN             (1 << 1)
1321
1322 static const size_t trace__entry_str_size = 2048;
1323
1324 static struct file *thread_trace__files_entry(struct thread_trace *ttrace, int fd)
1325 {
1326         if (fd < 0)
1327                 return NULL;
1328
1329         if (fd > ttrace->files.max) {
1330                 struct file *nfiles = realloc(ttrace->files.table, (fd + 1) * sizeof(struct file));
1331
1332                 if (nfiles == NULL)
1333                         return NULL;
1334
1335                 if (ttrace->files.max != -1) {
1336                         memset(nfiles + ttrace->files.max + 1, 0,
1337                                (fd - ttrace->files.max) * sizeof(struct file));
1338                 } else {
1339                         memset(nfiles, 0, (fd + 1) * sizeof(struct file));
1340                 }
1341
1342                 ttrace->files.table = nfiles;
1343                 ttrace->files.max   = fd;
1344         }
1345
1346         return ttrace->files.table + fd;
1347 }
1348
1349 struct file *thread__files_entry(struct thread *thread, int fd)
1350 {
1351         return thread_trace__files_entry(thread__priv(thread), fd);
1352 }
1353
1354 static int trace__set_fd_pathname(struct thread *thread, int fd, const char *pathname)
1355 {
1356         struct thread_trace *ttrace = thread__priv(thread);
1357         struct file *file = thread_trace__files_entry(ttrace, fd);
1358
1359         if (file != NULL) {
1360                 struct stat st;
1361                 if (stat(pathname, &st) == 0)
1362                         file->dev_maj = major(st.st_rdev);
1363                 file->pathname = strdup(pathname);
1364                 if (file->pathname)
1365                         return 0;
1366         }
1367
1368         return -1;
1369 }
1370
1371 static int thread__read_fd_path(struct thread *thread, int fd)
1372 {
1373         char linkname[PATH_MAX], pathname[PATH_MAX];
1374         struct stat st;
1375         int ret;
1376
1377         if (thread->pid_ == thread->tid) {
1378                 scnprintf(linkname, sizeof(linkname),
1379                           "/proc/%d/fd/%d", thread->pid_, fd);
1380         } else {
1381                 scnprintf(linkname, sizeof(linkname),
1382                           "/proc/%d/task/%d/fd/%d", thread->pid_, thread->tid, fd);
1383         }
1384
1385         if (lstat(linkname, &st) < 0 || st.st_size + 1 > (off_t)sizeof(pathname))
1386                 return -1;
1387
1388         ret = readlink(linkname, pathname, sizeof(pathname));
1389
1390         if (ret < 0 || ret > st.st_size)
1391                 return -1;
1392
1393         pathname[ret] = '\0';
1394         return trace__set_fd_pathname(thread, fd, pathname);
1395 }
1396
1397 static const char *thread__fd_path(struct thread *thread, int fd,
1398                                    struct trace *trace)
1399 {
1400         struct thread_trace *ttrace = thread__priv(thread);
1401
1402         if (ttrace == NULL || trace->fd_path_disabled)
1403                 return NULL;
1404
1405         if (fd < 0)
1406                 return NULL;
1407
1408         if ((fd > ttrace->files.max || ttrace->files.table[fd].pathname == NULL)) {
1409                 if (!trace->live)
1410                         return NULL;
1411                 ++trace->stats.proc_getname;
1412                 if (thread__read_fd_path(thread, fd))
1413                         return NULL;
1414         }
1415
1416         return ttrace->files.table[fd].pathname;
1417 }
1418
1419 size_t syscall_arg__scnprintf_fd(char *bf, size_t size, struct syscall_arg *arg)
1420 {
1421         int fd = arg->val;
1422         size_t printed = scnprintf(bf, size, "%d", fd);
1423         const char *path = thread__fd_path(arg->thread, fd, arg->trace);
1424
1425         if (path)
1426                 printed += scnprintf(bf + printed, size - printed, "<%s>", path);
1427
1428         return printed;
1429 }
1430
1431 size_t pid__scnprintf_fd(struct trace *trace, pid_t pid, int fd, char *bf, size_t size)
1432 {
1433         size_t printed = scnprintf(bf, size, "%d", fd);
1434         struct thread *thread = machine__find_thread(trace->host, pid, pid);
1435
1436         if (thread) {
1437                 const char *path = thread__fd_path(thread, fd, trace);
1438
1439                 if (path)
1440                         printed += scnprintf(bf + printed, size - printed, "<%s>", path);
1441
1442                 thread__put(thread);
1443         }
1444
1445         return printed;
1446 }
1447
1448 static size_t syscall_arg__scnprintf_close_fd(char *bf, size_t size,
1449                                               struct syscall_arg *arg)
1450 {
1451         int fd = arg->val;
1452         size_t printed = syscall_arg__scnprintf_fd(bf, size, arg);
1453         struct thread_trace *ttrace = thread__priv(arg->thread);
1454
1455         if (ttrace && fd >= 0 && fd <= ttrace->files.max)
1456                 zfree(&ttrace->files.table[fd].pathname);
1457
1458         return printed;
1459 }
1460
1461 static void thread__set_filename_pos(struct thread *thread, const char *bf,
1462                                      unsigned long ptr)
1463 {
1464         struct thread_trace *ttrace = thread__priv(thread);
1465
1466         ttrace->filename.ptr = ptr;
1467         ttrace->filename.entry_str_pos = bf - ttrace->entry_str;
1468 }
1469
1470 static size_t syscall_arg__scnprintf_augmented_string(struct syscall_arg *arg, char *bf, size_t size)
1471 {
1472         struct augmented_arg *augmented_arg = arg->augmented.args;
1473         size_t printed = scnprintf(bf, size, "\"%.*s\"", augmented_arg->size, augmented_arg->value);
1474         /*
1475          * So that the next arg with a payload can consume its augmented arg, i.e. for rename* syscalls
1476          * we would have two strings, each prefixed by its size.
1477          */
1478         int consumed = sizeof(*augmented_arg) + augmented_arg->size;
1479
1480         arg->augmented.args = ((void *)arg->augmented.args) + consumed;
1481         arg->augmented.size -= consumed;
1482
1483         return printed;
1484 }
1485
1486 static size_t syscall_arg__scnprintf_filename(char *bf, size_t size,
1487                                               struct syscall_arg *arg)
1488 {
1489         unsigned long ptr = arg->val;
1490
1491         if (arg->augmented.args)
1492                 return syscall_arg__scnprintf_augmented_string(arg, bf, size);
1493
1494         if (!arg->trace->vfs_getname)
1495                 return scnprintf(bf, size, "%#x", ptr);
1496
1497         thread__set_filename_pos(arg->thread, bf, ptr);
1498         return 0;
1499 }
1500
1501 static bool trace__filter_duration(struct trace *trace, double t)
1502 {
1503         return t < (trace->duration_filter * NSEC_PER_MSEC);
1504 }
1505
1506 static size_t __trace__fprintf_tstamp(struct trace *trace, u64 tstamp, FILE *fp)
1507 {
1508         double ts = (double)(tstamp - trace->base_time) / NSEC_PER_MSEC;
1509
1510         return fprintf(fp, "%10.3f ", ts);
1511 }
1512
1513 /*
1514  * We're handling tstamp=0 as an undefined tstamp, i.e. like when we are
1515  * using ttrace->entry_time for a thread that receives a sys_exit without
1516  * first having received a sys_enter ("poll" issued before tracing session
1517  * starts, lost sys_enter exit due to ring buffer overflow).
1518  */
1519 static size_t trace__fprintf_tstamp(struct trace *trace, u64 tstamp, FILE *fp)
1520 {
1521         if (tstamp > 0)
1522                 return __trace__fprintf_tstamp(trace, tstamp, fp);
1523
1524         return fprintf(fp, "         ? ");
1525 }
1526
1527 static bool done = false;
1528 static bool interrupted = false;
1529
1530 static void sig_handler(int sig)
1531 {
1532         done = true;
1533         interrupted = sig == SIGINT;
1534 }
1535
1536 static size_t trace__fprintf_comm_tid(struct trace *trace, struct thread *thread, FILE *fp)
1537 {
1538         size_t printed = 0;
1539
1540         if (trace->multiple_threads) {
1541                 if (trace->show_comm)
1542                         printed += fprintf(fp, "%.14s/", thread__comm_str(thread));
1543                 printed += fprintf(fp, "%d ", thread->tid);
1544         }
1545
1546         return printed;
1547 }
1548
1549 static size_t trace__fprintf_entry_head(struct trace *trace, struct thread *thread,
1550                                         u64 duration, bool duration_calculated, u64 tstamp, FILE *fp)
1551 {
1552         size_t printed = 0;
1553
1554         if (trace->show_tstamp)
1555                 printed = trace__fprintf_tstamp(trace, tstamp, fp);
1556         if (trace->show_duration)
1557                 printed += fprintf_duration(duration, duration_calculated, fp);
1558         return printed + trace__fprintf_comm_tid(trace, thread, fp);
1559 }
1560
1561 static int trace__process_event(struct trace *trace, struct machine *machine,
1562                                 union perf_event *event, struct perf_sample *sample)
1563 {
1564         int ret = 0;
1565
1566         switch (event->header.type) {
1567         case PERF_RECORD_LOST:
1568                 color_fprintf(trace->output, PERF_COLOR_RED,
1569                               "LOST %" PRIu64 " events!\n", event->lost.lost);
1570                 ret = machine__process_lost_event(machine, event, sample);
1571                 break;
1572         default:
1573                 ret = machine__process_event(machine, event, sample);
1574                 break;
1575         }
1576
1577         return ret;
1578 }
1579
1580 static int trace__tool_process(struct perf_tool *tool,
1581                                union perf_event *event,
1582                                struct perf_sample *sample,
1583                                struct machine *machine)
1584 {
1585         struct trace *trace = container_of(tool, struct trace, tool);
1586         return trace__process_event(trace, machine, event, sample);
1587 }
1588
1589 static char *trace__machine__resolve_kernel_addr(void *vmachine, unsigned long long *addrp, char **modp)
1590 {
1591         struct machine *machine = vmachine;
1592
1593         if (machine->kptr_restrict_warned)
1594                 return NULL;
1595
1596         if (symbol_conf.kptr_restrict) {
1597                 pr_warning("Kernel address maps (/proc/{kallsyms,modules}) are restricted.\n\n"
1598                            "Check /proc/sys/kernel/kptr_restrict and /proc/sys/kernel/perf_event_paranoid.\n\n"
1599                            "Kernel samples will not be resolved.\n");
1600                 machine->kptr_restrict_warned = true;
1601                 return NULL;
1602         }
1603
1604         return machine__resolve_kernel_addr(vmachine, addrp, modp);
1605 }
1606
1607 static int trace__symbols_init(struct trace *trace, struct evlist *evlist)
1608 {
1609         int err = symbol__init(NULL);
1610
1611         if (err)
1612                 return err;
1613
1614         trace->host = machine__new_host();
1615         if (trace->host == NULL)
1616                 return -ENOMEM;
1617
1618         err = trace_event__register_resolver(trace->host, trace__machine__resolve_kernel_addr);
1619         if (err < 0)
1620                 goto out;
1621
1622         err = __machine__synthesize_threads(trace->host, &trace->tool, &trace->opts.target,
1623                                             evlist->core.threads, trace__tool_process, false,
1624                                             1);
1625 out:
1626         if (err)
1627                 symbol__exit();
1628
1629         return err;
1630 }
1631
1632 static void trace__symbols__exit(struct trace *trace)
1633 {
1634         machine__exit(trace->host);
1635         trace->host = NULL;
1636
1637         symbol__exit();
1638 }
1639
1640 static int syscall__alloc_arg_fmts(struct syscall *sc, int nr_args)
1641 {
1642         int idx;
1643
1644         if (nr_args == 6 && sc->fmt && sc->fmt->nr_args != 0)
1645                 nr_args = sc->fmt->nr_args;
1646
1647         sc->arg_fmt = calloc(nr_args, sizeof(*sc->arg_fmt));
1648         if (sc->arg_fmt == NULL)
1649                 return -1;
1650
1651         for (idx = 0; idx < nr_args; ++idx) {
1652                 if (sc->fmt)
1653                         sc->arg_fmt[idx] = sc->fmt->arg[idx];
1654         }
1655
1656         sc->nr_args = nr_args;
1657         return 0;
1658 }
1659
1660 static struct syscall_arg_fmt syscall_arg_fmts__by_name[] = {
1661         { .name = "msr",        .scnprintf = SCA_X86_MSR,         .strtoul = STUL_X86_MSR,         },
1662         { .name = "vector",     .scnprintf = SCA_X86_IRQ_VECTORS, .strtoul = STUL_X86_IRQ_VECTORS, },
1663 };
1664
1665 static int syscall_arg_fmt__cmp(const void *name, const void *fmtp)
1666 {
1667        const struct syscall_arg_fmt *fmt = fmtp;
1668        return strcmp(name, fmt->name);
1669 }
1670
1671 static struct syscall_arg_fmt *
1672 __syscall_arg_fmt__find_by_name(struct syscall_arg_fmt *fmts, const int nmemb, const char *name)
1673 {
1674        return bsearch(name, fmts, nmemb, sizeof(struct syscall_arg_fmt), syscall_arg_fmt__cmp);
1675 }
1676
1677 static struct syscall_arg_fmt *syscall_arg_fmt__find_by_name(const char *name)
1678 {
1679        const int nmemb = ARRAY_SIZE(syscall_arg_fmts__by_name);
1680        return __syscall_arg_fmt__find_by_name(syscall_arg_fmts__by_name, nmemb, name);
1681 }
1682
1683 static struct tep_format_field *
1684 syscall_arg_fmt__init_array(struct syscall_arg_fmt *arg, struct tep_format_field *field)
1685 {
1686         struct tep_format_field *last_field = NULL;
1687         int len;
1688
1689         for (; field; field = field->next, ++arg) {
1690                 last_field = field;
1691
1692                 if (arg->scnprintf)
1693                         continue;
1694
1695                 len = strlen(field->name);
1696
1697                 if (strcmp(field->type, "const char *") == 0 &&
1698                     ((len >= 4 && strcmp(field->name + len - 4, "name") == 0) ||
1699                      strstr(field->name, "path") != NULL))
1700                         arg->scnprintf = SCA_FILENAME;
1701                 else if ((field->flags & TEP_FIELD_IS_POINTER) || strstr(field->name, "addr"))
1702                         arg->scnprintf = SCA_PTR;
1703                 else if (strcmp(field->type, "pid_t") == 0)
1704                         arg->scnprintf = SCA_PID;
1705                 else if (strcmp(field->type, "umode_t") == 0)
1706                         arg->scnprintf = SCA_MODE_T;
1707                 else if ((field->flags & TEP_FIELD_IS_ARRAY) && strstr(field->type, "char")) {
1708                         arg->scnprintf = SCA_CHAR_ARRAY;
1709                         arg->nr_entries = field->arraylen;
1710                 } else if ((strcmp(field->type, "int") == 0 ||
1711                           strcmp(field->type, "unsigned int") == 0 ||
1712                           strcmp(field->type, "long") == 0) &&
1713                          len >= 2 && strcmp(field->name + len - 2, "fd") == 0) {
1714                         /*
1715                          * /sys/kernel/tracing/events/syscalls/sys_enter*
1716                          * egrep 'field:.*fd;' .../format|sed -r 's/.*field:([a-z ]+) [a-z_]*fd.+/\1/g'|sort|uniq -c
1717                          * 65 int
1718                          * 23 unsigned int
1719                          * 7 unsigned long
1720                          */
1721                         arg->scnprintf = SCA_FD;
1722                } else {
1723                         struct syscall_arg_fmt *fmt = syscall_arg_fmt__find_by_name(field->name);
1724
1725                         if (fmt) {
1726                                 arg->scnprintf = fmt->scnprintf;
1727                                 arg->strtoul   = fmt->strtoul;
1728                         }
1729                 }
1730         }
1731
1732         return last_field;
1733 }
1734
1735 static int syscall__set_arg_fmts(struct syscall *sc)
1736 {
1737         struct tep_format_field *last_field = syscall_arg_fmt__init_array(sc->arg_fmt, sc->args);
1738
1739         if (last_field)
1740                 sc->args_size = last_field->offset + last_field->size;
1741
1742         return 0;
1743 }
1744
1745 static int trace__read_syscall_info(struct trace *trace, int id)
1746 {
1747         char tp_name[128];
1748         struct syscall *sc;
1749         const char *name = syscalltbl__name(trace->sctbl, id);
1750
1751 #ifdef HAVE_SYSCALL_TABLE_SUPPORT
1752         if (trace->syscalls.table == NULL) {
1753                 trace->syscalls.table = calloc(trace->sctbl->syscalls.max_id + 1, sizeof(*sc));
1754                 if (trace->syscalls.table == NULL)
1755                         return -ENOMEM;
1756         }
1757 #else
1758         if (id > trace->sctbl->syscalls.max_id || (id == 0 && trace->syscalls.table == NULL)) {
1759                 // When using libaudit we don't know beforehand what is the max syscall id
1760                 struct syscall *table = realloc(trace->syscalls.table, (id + 1) * sizeof(*sc));
1761
1762                 if (table == NULL)
1763                         return -ENOMEM;
1764
1765                 // Need to memset from offset 0 and +1 members if brand new
1766                 if (trace->syscalls.table == NULL)
1767                         memset(table, 0, (id + 1) * sizeof(*sc));
1768                 else
1769                         memset(table + trace->sctbl->syscalls.max_id + 1, 0, (id - trace->sctbl->syscalls.max_id) * sizeof(*sc));
1770
1771                 trace->syscalls.table         = table;
1772                 trace->sctbl->syscalls.max_id = id;
1773         }
1774 #endif
1775         sc = trace->syscalls.table + id;
1776         if (sc->nonexistent)
1777                 return 0;
1778
1779         if (name == NULL) {
1780                 sc->nonexistent = true;
1781                 return 0;
1782         }
1783
1784         sc->name = name;
1785         sc->fmt  = syscall_fmt__find(sc->name);
1786
1787         snprintf(tp_name, sizeof(tp_name), "sys_enter_%s", sc->name);
1788         sc->tp_format = trace_event__tp_format("syscalls", tp_name);
1789
1790         if (IS_ERR(sc->tp_format) && sc->fmt && sc->fmt->alias) {
1791                 snprintf(tp_name, sizeof(tp_name), "sys_enter_%s", sc->fmt->alias);
1792                 sc->tp_format = trace_event__tp_format("syscalls", tp_name);
1793         }
1794
1795         if (syscall__alloc_arg_fmts(sc, IS_ERR(sc->tp_format) ? 6 : sc->tp_format->format.nr_fields))
1796                 return -ENOMEM;
1797
1798         if (IS_ERR(sc->tp_format))
1799                 return PTR_ERR(sc->tp_format);
1800
1801         sc->args = sc->tp_format->format.fields;
1802         /*
1803          * We need to check and discard the first variable '__syscall_nr'
1804          * or 'nr' that mean the syscall number. It is needless here.
1805          * So drop '__syscall_nr' or 'nr' field but does not exist on older kernels.
1806          */
1807         if (sc->args && (!strcmp(sc->args->name, "__syscall_nr") || !strcmp(sc->args->name, "nr"))) {
1808                 sc->args = sc->args->next;
1809                 --sc->nr_args;
1810         }
1811
1812         sc->is_exit = !strcmp(name, "exit_group") || !strcmp(name, "exit");
1813         sc->is_open = !strcmp(name, "open") || !strcmp(name, "openat");
1814
1815         return syscall__set_arg_fmts(sc);
1816 }
1817
1818 static int evsel__init_tp_arg_scnprintf(struct evsel *evsel)
1819 {
1820         struct syscall_arg_fmt *fmt = evsel__syscall_arg_fmt(evsel);
1821
1822         if (fmt != NULL) {
1823                 syscall_arg_fmt__init_array(fmt, evsel->tp_format->format.fields);
1824                 return 0;
1825         }
1826
1827         return -ENOMEM;
1828 }
1829
1830 static int intcmp(const void *a, const void *b)
1831 {
1832         const int *one = a, *another = b;
1833
1834         return *one - *another;
1835 }
1836
1837 static int trace__validate_ev_qualifier(struct trace *trace)
1838 {
1839         int err = 0;
1840         bool printed_invalid_prefix = false;
1841         struct str_node *pos;
1842         size_t nr_used = 0, nr_allocated = strlist__nr_entries(trace->ev_qualifier);
1843
1844         trace->ev_qualifier_ids.entries = malloc(nr_allocated *
1845                                                  sizeof(trace->ev_qualifier_ids.entries[0]));
1846
1847         if (trace->ev_qualifier_ids.entries == NULL) {
1848                 fputs("Error:\tNot enough memory for allocating events qualifier ids\n",
1849                        trace->output);
1850                 err = -EINVAL;
1851                 goto out;
1852         }
1853
1854         strlist__for_each_entry(pos, trace->ev_qualifier) {
1855                 const char *sc = pos->s;
1856                 int id = syscalltbl__id(trace->sctbl, sc), match_next = -1;
1857
1858                 if (id < 0) {
1859                         id = syscalltbl__strglobmatch_first(trace->sctbl, sc, &match_next);
1860                         if (id >= 0)
1861                                 goto matches;
1862
1863                         if (!printed_invalid_prefix) {
1864                                 pr_debug("Skipping unknown syscalls: ");
1865                                 printed_invalid_prefix = true;
1866                         } else {
1867                                 pr_debug(", ");
1868                         }
1869
1870                         pr_debug("%s", sc);
1871                         continue;
1872                 }
1873 matches:
1874                 trace->ev_qualifier_ids.entries[nr_used++] = id;
1875                 if (match_next == -1)
1876                         continue;
1877
1878                 while (1) {
1879                         id = syscalltbl__strglobmatch_next(trace->sctbl, sc, &match_next);
1880                         if (id < 0)
1881                                 break;
1882                         if (nr_allocated == nr_used) {
1883                                 void *entries;
1884
1885                                 nr_allocated += 8;
1886                                 entries = realloc(trace->ev_qualifier_ids.entries,
1887                                                   nr_allocated * sizeof(trace->ev_qualifier_ids.entries[0]));
1888                                 if (entries == NULL) {
1889                                         err = -ENOMEM;
1890                                         fputs("\nError:\t Not enough memory for parsing\n", trace->output);
1891                                         goto out_free;
1892                                 }
1893                                 trace->ev_qualifier_ids.entries = entries;
1894                         }
1895                         trace->ev_qualifier_ids.entries[nr_used++] = id;
1896                 }
1897         }
1898
1899         trace->ev_qualifier_ids.nr = nr_used;
1900         qsort(trace->ev_qualifier_ids.entries, nr_used, sizeof(int), intcmp);
1901 out:
1902         if (printed_invalid_prefix)
1903                 pr_debug("\n");
1904         return err;
1905 out_free:
1906         zfree(&trace->ev_qualifier_ids.entries);
1907         trace->ev_qualifier_ids.nr = 0;
1908         goto out;
1909 }
1910
1911 static __maybe_unused bool trace__syscall_enabled(struct trace *trace, int id)
1912 {
1913         bool in_ev_qualifier;
1914
1915         if (trace->ev_qualifier_ids.nr == 0)
1916                 return true;
1917
1918         in_ev_qualifier = bsearch(&id, trace->ev_qualifier_ids.entries,
1919                                   trace->ev_qualifier_ids.nr, sizeof(int), intcmp) != NULL;
1920
1921         if (in_ev_qualifier)
1922                return !trace->not_ev_qualifier;
1923
1924         return trace->not_ev_qualifier;
1925 }
1926
1927 /*
1928  * args is to be interpreted as a series of longs but we need to handle
1929  * 8-byte unaligned accesses. args points to raw_data within the event
1930  * and raw_data is guaranteed to be 8-byte unaligned because it is
1931  * preceded by raw_size which is a u32. So we need to copy args to a temp
1932  * variable to read it. Most notably this avoids extended load instructions
1933  * on unaligned addresses
1934  */
1935 unsigned long syscall_arg__val(struct syscall_arg *arg, u8 idx)
1936 {
1937         unsigned long val;
1938         unsigned char *p = arg->args + sizeof(unsigned long) * idx;
1939
1940         memcpy(&val, p, sizeof(val));
1941         return val;
1942 }
1943
1944 static size_t syscall__scnprintf_name(struct syscall *sc, char *bf, size_t size,
1945                                       struct syscall_arg *arg)
1946 {
1947         if (sc->arg_fmt && sc->arg_fmt[arg->idx].name)
1948                 return scnprintf(bf, size, "%s: ", sc->arg_fmt[arg->idx].name);
1949
1950         return scnprintf(bf, size, "arg%d: ", arg->idx);
1951 }
1952
1953 /*
1954  * Check if the value is in fact zero, i.e. mask whatever needs masking, such
1955  * as mount 'flags' argument that needs ignoring some magic flag, see comment
1956  * in tools/perf/trace/beauty/mount_flags.c
1957  */
1958 static unsigned long syscall_arg_fmt__mask_val(struct syscall_arg_fmt *fmt, struct syscall_arg *arg, unsigned long val)
1959 {
1960         if (fmt && fmt->mask_val)
1961                 return fmt->mask_val(arg, val);
1962
1963         return val;
1964 }
1965
1966 static size_t syscall_arg_fmt__scnprintf_val(struct syscall_arg_fmt *fmt, char *bf, size_t size,
1967                                              struct syscall_arg *arg, unsigned long val)
1968 {
1969         if (fmt && fmt->scnprintf) {
1970                 arg->val = val;
1971                 if (fmt->parm)
1972                         arg->parm = fmt->parm;
1973                 return fmt->scnprintf(bf, size, arg);
1974         }
1975         return scnprintf(bf, size, "%ld", val);
1976 }
1977
1978 static size_t syscall__scnprintf_args(struct syscall *sc, char *bf, size_t size,
1979                                       unsigned char *args, void *augmented_args, int augmented_args_size,
1980                                       struct trace *trace, struct thread *thread)
1981 {
1982         size_t printed = 0;
1983         unsigned long val;
1984         u8 bit = 1;
1985         struct syscall_arg arg = {
1986                 .args   = args,
1987                 .augmented = {
1988                         .size = augmented_args_size,
1989                         .args = augmented_args,
1990                 },
1991                 .idx    = 0,
1992                 .mask   = 0,
1993                 .trace  = trace,
1994                 .thread = thread,
1995                 .show_string_prefix = trace->show_string_prefix,
1996         };
1997         struct thread_trace *ttrace = thread__priv(thread);
1998
1999         /*
2000          * Things like fcntl will set this in its 'cmd' formatter to pick the
2001          * right formatter for the return value (an fd? file flags?), which is
2002          * not needed for syscalls that always return a given type, say an fd.
2003          */
2004         ttrace->ret_scnprintf = NULL;
2005
2006         if (sc->args != NULL) {
2007                 struct tep_format_field *field;
2008
2009                 for (field = sc->args; field;
2010                      field = field->next, ++arg.idx, bit <<= 1) {
2011                         if (arg.mask & bit)
2012                                 continue;
2013
2014                         arg.fmt = &sc->arg_fmt[arg.idx];
2015                         val = syscall_arg__val(&arg, arg.idx);
2016                         /*
2017                          * Some syscall args need some mask, most don't and
2018                          * return val untouched.
2019                          */
2020                         val = syscall_arg_fmt__mask_val(&sc->arg_fmt[arg.idx], &arg, val);
2021
2022                         /*
2023                          * Suppress this argument if its value is zero and
2024                          * and we don't have a string associated in an
2025                          * strarray for it.
2026                          */
2027                         if (val == 0 &&
2028                             !trace->show_zeros &&
2029                             !(sc->arg_fmt &&
2030                               (sc->arg_fmt[arg.idx].show_zero ||
2031                                sc->arg_fmt[arg.idx].scnprintf == SCA_STRARRAY ||
2032                                sc->arg_fmt[arg.idx].scnprintf == SCA_STRARRAYS) &&
2033                               sc->arg_fmt[arg.idx].parm))
2034                                 continue;
2035
2036                         printed += scnprintf(bf + printed, size - printed, "%s", printed ? ", " : "");
2037
2038                         if (trace->show_arg_names)
2039                                 printed += scnprintf(bf + printed, size - printed, "%s: ", field->name);
2040
2041                         printed += syscall_arg_fmt__scnprintf_val(&sc->arg_fmt[arg.idx],
2042                                                                   bf + printed, size - printed, &arg, val);
2043                 }
2044         } else if (IS_ERR(sc->tp_format)) {
2045                 /*
2046                  * If we managed to read the tracepoint /format file, then we
2047                  * may end up not having any args, like with gettid(), so only
2048                  * print the raw args when we didn't manage to read it.
2049                  */
2050                 while (arg.idx < sc->nr_args) {
2051                         if (arg.mask & bit)
2052                                 goto next_arg;
2053                         val = syscall_arg__val(&arg, arg.idx);
2054                         if (printed)
2055                                 printed += scnprintf(bf + printed, size - printed, ", ");
2056                         printed += syscall__scnprintf_name(sc, bf + printed, size - printed, &arg);
2057                         printed += syscall_arg_fmt__scnprintf_val(&sc->arg_fmt[arg.idx], bf + printed, size - printed, &arg, val);
2058 next_arg:
2059                         ++arg.idx;
2060                         bit <<= 1;
2061                 }
2062         }
2063
2064         return printed;
2065 }
2066
2067 typedef int (*tracepoint_handler)(struct trace *trace, struct evsel *evsel,
2068                                   union perf_event *event,
2069                                   struct perf_sample *sample);
2070
2071 static struct syscall *trace__syscall_info(struct trace *trace,
2072                                            struct evsel *evsel, int id)
2073 {
2074         int err = 0;
2075
2076         if (id < 0) {
2077
2078                 /*
2079                  * XXX: Noticed on x86_64, reproduced as far back as 3.0.36, haven't tried
2080                  * before that, leaving at a higher verbosity level till that is
2081                  * explained. Reproduced with plain ftrace with:
2082                  *
2083                  * echo 1 > /t/events/raw_syscalls/sys_exit/enable
2084                  * grep "NR -1 " /t/trace_pipe
2085                  *
2086                  * After generating some load on the machine.
2087                  */
2088                 if (verbose > 1) {
2089                         static u64 n;
2090                         fprintf(trace->output, "Invalid syscall %d id, skipping (%s, %" PRIu64 ") ...\n",
2091                                 id, evsel__name(evsel), ++n);
2092                 }
2093                 return NULL;
2094         }
2095
2096         err = -EINVAL;
2097
2098 #ifdef HAVE_SYSCALL_TABLE_SUPPORT
2099         if (id > trace->sctbl->syscalls.max_id) {
2100 #else
2101         if (id >= trace->sctbl->syscalls.max_id) {
2102                 /*
2103                  * With libaudit we don't know beforehand what is the max_id,
2104                  * so we let trace__read_syscall_info() figure that out as we
2105                  * go on reading syscalls.
2106                  */
2107                 err = trace__read_syscall_info(trace, id);
2108                 if (err)
2109 #endif
2110                 goto out_cant_read;
2111         }
2112
2113         if ((trace->syscalls.table == NULL || trace->syscalls.table[id].name == NULL) &&
2114             (err = trace__read_syscall_info(trace, id)) != 0)
2115                 goto out_cant_read;
2116
2117         if (trace->syscalls.table[id].name == NULL) {
2118                 if (trace->syscalls.table[id].nonexistent)
2119                         return NULL;
2120                 goto out_cant_read;
2121         }
2122
2123         return &trace->syscalls.table[id];
2124
2125 out_cant_read:
2126         if (verbose > 0) {
2127                 char sbuf[STRERR_BUFSIZE];
2128                 fprintf(trace->output, "Problems reading syscall %d: %d (%s)", id, -err, str_error_r(-err, sbuf, sizeof(sbuf)));
2129                 if (id <= trace->sctbl->syscalls.max_id && trace->syscalls.table[id].name != NULL)
2130                         fprintf(trace->output, "(%s)", trace->syscalls.table[id].name);
2131                 fputs(" information\n", trace->output);
2132         }
2133         return NULL;
2134 }
2135
2136 struct syscall_stats {
2137         struct stats stats;
2138         u64          nr_failures;
2139         int          max_errno;
2140         u32          *errnos;
2141 };
2142
2143 static void thread__update_stats(struct thread *thread, struct thread_trace *ttrace,
2144                                  int id, struct perf_sample *sample, long err, bool errno_summary)
2145 {
2146         struct int_node *inode;
2147         struct syscall_stats *stats;
2148         u64 duration = 0;
2149
2150         inode = intlist__findnew(ttrace->syscall_stats, id);
2151         if (inode == NULL)
2152                 return;
2153
2154         stats = inode->priv;
2155         if (stats == NULL) {
2156                 stats = malloc(sizeof(*stats));
2157                 if (stats == NULL)
2158                         return;
2159
2160                 stats->nr_failures = 0;
2161                 stats->max_errno   = 0;
2162                 stats->errnos      = NULL;
2163                 init_stats(&stats->stats);
2164                 inode->priv = stats;
2165         }
2166
2167         if (ttrace->entry_time && sample->time > ttrace->entry_time)
2168                 duration = sample->time - ttrace->entry_time;
2169
2170         update_stats(&stats->stats, duration);
2171
2172         if (err < 0) {
2173                 ++stats->nr_failures;
2174
2175                 if (!errno_summary)
2176                         return;
2177
2178                 err = -err;
2179                 if (err > stats->max_errno) {
2180                         u32 *new_errnos = realloc(stats->errnos, err * sizeof(u32));
2181
2182                         if (new_errnos) {
2183                                 memset(new_errnos + stats->max_errno, 0, (err - stats->max_errno) * sizeof(u32));
2184                         } else {
2185                                 pr_debug("Not enough memory for errno stats for thread \"%s\"(%d/%d), results will be incomplete\n",
2186                                          thread__comm_str(thread), thread->pid_, thread->tid);
2187                                 return;
2188                         }
2189
2190                         stats->errnos = new_errnos;
2191                         stats->max_errno = err;
2192                 }
2193
2194                 ++stats->errnos[err - 1];
2195         }
2196 }
2197
2198 static int trace__printf_interrupted_entry(struct trace *trace)
2199 {
2200         struct thread_trace *ttrace;
2201         size_t printed;
2202         int len;
2203
2204         if (trace->failure_only || trace->current == NULL)
2205                 return 0;
2206
2207         ttrace = thread__priv(trace->current);
2208
2209         if (!ttrace->entry_pending)
2210                 return 0;
2211
2212         printed  = trace__fprintf_entry_head(trace, trace->current, 0, false, ttrace->entry_time, trace->output);
2213         printed += len = fprintf(trace->output, "%s)", ttrace->entry_str);
2214
2215         if (len < trace->args_alignment - 4)
2216                 printed += fprintf(trace->output, "%-*s", trace->args_alignment - 4 - len, " ");
2217
2218         printed += fprintf(trace->output, " ...\n");
2219
2220         ttrace->entry_pending = false;
2221         ++trace->nr_events_printed;
2222
2223         return printed;
2224 }
2225
2226 static int trace__fprintf_sample(struct trace *trace, struct evsel *evsel,
2227                                  struct perf_sample *sample, struct thread *thread)
2228 {
2229         int printed = 0;
2230
2231         if (trace->print_sample) {
2232                 double ts = (double)sample->time / NSEC_PER_MSEC;
2233
2234                 printed += fprintf(trace->output, "%22s %10.3f %s %d/%d [%d]\n",
2235                                    evsel__name(evsel), ts,
2236                                    thread__comm_str(thread),
2237                                    sample->pid, sample->tid, sample->cpu);
2238         }
2239
2240         return printed;
2241 }
2242
2243 static void *syscall__augmented_args(struct syscall *sc, struct perf_sample *sample, int *augmented_args_size, int raw_augmented_args_size)
2244 {
2245         void *augmented_args = NULL;
2246         /*
2247          * For now with BPF raw_augmented we hook into raw_syscalls:sys_enter
2248          * and there we get all 6 syscall args plus the tracepoint common fields
2249          * that gets calculated at the start and the syscall_nr (another long).
2250          * So we check if that is the case and if so don't look after the
2251          * sc->args_size but always after the full raw_syscalls:sys_enter payload,
2252          * which is fixed.
2253          *
2254          * We'll revisit this later to pass s->args_size to the BPF augmenter
2255          * (now tools/perf/examples/bpf/augmented_raw_syscalls.c, so that it
2256          * copies only what we need for each syscall, like what happens when we
2257          * use syscalls:sys_enter_NAME, so that we reduce the kernel/userspace
2258          * traffic to just what is needed for each syscall.
2259          */
2260         int args_size = raw_augmented_args_size ?: sc->args_size;
2261
2262         *augmented_args_size = sample->raw_size - args_size;
2263         if (*augmented_args_size > 0)
2264                 augmented_args = sample->raw_data + args_size;
2265
2266         return augmented_args;
2267 }
2268
2269 static void syscall__exit(struct syscall *sc)
2270 {
2271         if (!sc)
2272                 return;
2273
2274         free(sc->arg_fmt);
2275 }
2276
2277 static int trace__sys_enter(struct trace *trace, struct evsel *evsel,
2278                             union perf_event *event __maybe_unused,
2279                             struct perf_sample *sample)
2280 {
2281         char *msg;
2282         void *args;
2283         int printed = 0;
2284         struct thread *thread;
2285         int id = perf_evsel__sc_tp_uint(evsel, id, sample), err = -1;
2286         int augmented_args_size = 0;
2287         void *augmented_args = NULL;
2288         struct syscall *sc = trace__syscall_info(trace, evsel, id);
2289         struct thread_trace *ttrace;
2290
2291         if (sc == NULL)
2292                 return -1;
2293
2294         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2295         ttrace = thread__trace(thread, trace->output);
2296         if (ttrace == NULL)
2297                 goto out_put;
2298
2299         trace__fprintf_sample(trace, evsel, sample, thread);
2300
2301         args = perf_evsel__sc_tp_ptr(evsel, args, sample);
2302
2303         if (ttrace->entry_str == NULL) {
2304                 ttrace->entry_str = malloc(trace__entry_str_size);
2305                 if (!ttrace->entry_str)
2306                         goto out_put;
2307         }
2308
2309         if (!(trace->duration_filter || trace->summary_only || trace->min_stack))
2310                 trace__printf_interrupted_entry(trace);
2311         /*
2312          * If this is raw_syscalls.sys_enter, then it always comes with the 6 possible
2313          * arguments, even if the syscall being handled, say "openat", uses only 4 arguments
2314          * this breaks syscall__augmented_args() check for augmented args, as we calculate
2315          * syscall->args_size using each syscalls:sys_enter_NAME tracefs format file,
2316          * so when handling, say the openat syscall, we end up getting 6 args for the
2317          * raw_syscalls:sys_enter event, when we expected just 4, we end up mistakenly
2318          * thinking that the extra 2 u64 args are the augmented filename, so just check
2319          * here and avoid using augmented syscalls when the evsel is the raw_syscalls one.
2320          */
2321         if (evsel != trace->syscalls.events.sys_enter)
2322                 augmented_args = syscall__augmented_args(sc, sample, &augmented_args_size, trace->raw_augmented_syscalls_args_size);
2323         ttrace->entry_time = sample->time;
2324         msg = ttrace->entry_str;
2325         printed += scnprintf(msg + printed, trace__entry_str_size - printed, "%s(", sc->name);
2326
2327         printed += syscall__scnprintf_args(sc, msg + printed, trace__entry_str_size - printed,
2328                                            args, augmented_args, augmented_args_size, trace, thread);
2329
2330         if (sc->is_exit) {
2331                 if (!(trace->duration_filter || trace->summary_only || trace->failure_only || trace->min_stack)) {
2332                         int alignment = 0;
2333
2334                         trace__fprintf_entry_head(trace, thread, 0, false, ttrace->entry_time, trace->output);
2335                         printed = fprintf(trace->output, "%s)", ttrace->entry_str);
2336                         if (trace->args_alignment > printed)
2337                                 alignment = trace->args_alignment - printed;
2338                         fprintf(trace->output, "%*s= ?\n", alignment, " ");
2339                 }
2340         } else {
2341                 ttrace->entry_pending = true;
2342                 /* See trace__vfs_getname & trace__sys_exit */
2343                 ttrace->filename.pending_open = false;
2344         }
2345
2346         if (trace->current != thread) {
2347                 thread__put(trace->current);
2348                 trace->current = thread__get(thread);
2349         }
2350         err = 0;
2351 out_put:
2352         thread__put(thread);
2353         return err;
2354 }
2355
2356 static int trace__fprintf_sys_enter(struct trace *trace, struct evsel *evsel,
2357                                     struct perf_sample *sample)
2358 {
2359         struct thread_trace *ttrace;
2360         struct thread *thread;
2361         int id = perf_evsel__sc_tp_uint(evsel, id, sample), err = -1;
2362         struct syscall *sc = trace__syscall_info(trace, evsel, id);
2363         char msg[1024];
2364         void *args, *augmented_args = NULL;
2365         int augmented_args_size;
2366
2367         if (sc == NULL)
2368                 return -1;
2369
2370         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2371         ttrace = thread__trace(thread, trace->output);
2372         /*
2373          * We need to get ttrace just to make sure it is there when syscall__scnprintf_args()
2374          * and the rest of the beautifiers accessing it via struct syscall_arg touches it.
2375          */
2376         if (ttrace == NULL)
2377                 goto out_put;
2378
2379         args = perf_evsel__sc_tp_ptr(evsel, args, sample);
2380         augmented_args = syscall__augmented_args(sc, sample, &augmented_args_size, trace->raw_augmented_syscalls_args_size);
2381         syscall__scnprintf_args(sc, msg, sizeof(msg), args, augmented_args, augmented_args_size, trace, thread);
2382         fprintf(trace->output, "%s", msg);
2383         err = 0;
2384 out_put:
2385         thread__put(thread);
2386         return err;
2387 }
2388
2389 static int trace__resolve_callchain(struct trace *trace, struct evsel *evsel,
2390                                     struct perf_sample *sample,
2391                                     struct callchain_cursor *cursor)
2392 {
2393         struct addr_location al;
2394         int max_stack = evsel->core.attr.sample_max_stack ?
2395                         evsel->core.attr.sample_max_stack :
2396                         trace->max_stack;
2397         int err;
2398
2399         if (machine__resolve(trace->host, &al, sample) < 0)
2400                 return -1;
2401
2402         err = thread__resolve_callchain(al.thread, cursor, evsel, sample, NULL, NULL, max_stack);
2403         addr_location__put(&al);
2404         return err;
2405 }
2406
2407 static int trace__fprintf_callchain(struct trace *trace, struct perf_sample *sample)
2408 {
2409         /* TODO: user-configurable print_opts */
2410         const unsigned int print_opts = EVSEL__PRINT_SYM |
2411                                         EVSEL__PRINT_DSO |
2412                                         EVSEL__PRINT_UNKNOWN_AS_ADDR;
2413
2414         return sample__fprintf_callchain(sample, 38, print_opts, &callchain_cursor, symbol_conf.bt_stop_list, trace->output);
2415 }
2416
2417 static const char *errno_to_name(struct evsel *evsel, int err)
2418 {
2419         struct perf_env *env = evsel__env(evsel);
2420         const char *arch_name = perf_env__arch(env);
2421
2422         return arch_syscalls__strerrno(arch_name, err);
2423 }
2424
2425 static int trace__sys_exit(struct trace *trace, struct evsel *evsel,
2426                            union perf_event *event __maybe_unused,
2427                            struct perf_sample *sample)
2428 {
2429         long ret;
2430         u64 duration = 0;
2431         bool duration_calculated = false;
2432         struct thread *thread;
2433         int id = perf_evsel__sc_tp_uint(evsel, id, sample), err = -1, callchain_ret = 0, printed = 0;
2434         int alignment = trace->args_alignment;
2435         struct syscall *sc = trace__syscall_info(trace, evsel, id);
2436         struct thread_trace *ttrace;
2437
2438         if (sc == NULL)
2439                 return -1;
2440
2441         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2442         ttrace = thread__trace(thread, trace->output);
2443         if (ttrace == NULL)
2444                 goto out_put;
2445
2446         trace__fprintf_sample(trace, evsel, sample, thread);
2447
2448         ret = perf_evsel__sc_tp_uint(evsel, ret, sample);
2449
2450         if (trace->summary)
2451                 thread__update_stats(thread, ttrace, id, sample, ret, trace->errno_summary);
2452
2453         if (!trace->fd_path_disabled && sc->is_open && ret >= 0 && ttrace->filename.pending_open) {
2454                 trace__set_fd_pathname(thread, ret, ttrace->filename.name);
2455                 ttrace->filename.pending_open = false;
2456                 ++trace->stats.vfs_getname;
2457         }
2458
2459         if (ttrace->entry_time) {
2460                 duration = sample->time - ttrace->entry_time;
2461                 if (trace__filter_duration(trace, duration))
2462                         goto out;
2463                 duration_calculated = true;
2464         } else if (trace->duration_filter)
2465                 goto out;
2466
2467         if (sample->callchain) {
2468                 callchain_ret = trace__resolve_callchain(trace, evsel, sample, &callchain_cursor);
2469                 if (callchain_ret == 0) {
2470                         if (callchain_cursor.nr < trace->min_stack)
2471                                 goto out;
2472                         callchain_ret = 1;
2473                 }
2474         }
2475
2476         if (trace->summary_only || (ret >= 0 && trace->failure_only))
2477                 goto out;
2478
2479         trace__fprintf_entry_head(trace, thread, duration, duration_calculated, ttrace->entry_time, trace->output);
2480
2481         if (ttrace->entry_pending) {
2482                 printed = fprintf(trace->output, "%s", ttrace->entry_str);
2483         } else {
2484                 printed += fprintf(trace->output, " ... [");
2485                 color_fprintf(trace->output, PERF_COLOR_YELLOW, "continued");
2486                 printed += 9;
2487                 printed += fprintf(trace->output, "]: %s()", sc->name);
2488         }
2489
2490         printed++; /* the closing ')' */
2491
2492         if (alignment > printed)
2493                 alignment -= printed;
2494         else
2495                 alignment = 0;
2496
2497         fprintf(trace->output, ")%*s= ", alignment, " ");
2498
2499         if (sc->fmt == NULL) {
2500                 if (ret < 0)
2501                         goto errno_print;
2502 signed_print:
2503                 fprintf(trace->output, "%ld", ret);
2504         } else if (ret < 0) {
2505 errno_print: {
2506                 char bf[STRERR_BUFSIZE];
2507                 const char *emsg = str_error_r(-ret, bf, sizeof(bf)),
2508                            *e = errno_to_name(evsel, -ret);
2509
2510                 fprintf(trace->output, "-1 %s (%s)", e, emsg);
2511         }
2512         } else if (ret == 0 && sc->fmt->timeout)
2513                 fprintf(trace->output, "0 (Timeout)");
2514         else if (ttrace->ret_scnprintf) {
2515                 char bf[1024];
2516                 struct syscall_arg arg = {
2517                         .val    = ret,
2518                         .thread = thread,
2519                         .trace  = trace,
2520                 };
2521                 ttrace->ret_scnprintf(bf, sizeof(bf), &arg);
2522                 ttrace->ret_scnprintf = NULL;
2523                 fprintf(trace->output, "%s", bf);
2524         } else if (sc->fmt->hexret)
2525                 fprintf(trace->output, "%#lx", ret);
2526         else if (sc->fmt->errpid) {
2527                 struct thread *child = machine__find_thread(trace->host, ret, ret);
2528
2529                 if (child != NULL) {
2530                         fprintf(trace->output, "%ld", ret);
2531                         if (child->comm_set)
2532                                 fprintf(trace->output, " (%s)", thread__comm_str(child));
2533                         thread__put(child);
2534                 }
2535         } else
2536                 goto signed_print;
2537
2538         fputc('\n', trace->output);
2539
2540         /*
2541          * We only consider an 'event' for the sake of --max-events a non-filtered
2542          * sys_enter + sys_exit and other tracepoint events.
2543          */
2544         if (++trace->nr_events_printed == trace->max_events && trace->max_events != ULONG_MAX)
2545                 interrupted = true;
2546
2547         if (callchain_ret > 0)
2548                 trace__fprintf_callchain(trace, sample);
2549         else if (callchain_ret < 0)
2550                 pr_err("Problem processing %s callchain, skipping...\n", evsel__name(evsel));
2551 out:
2552         ttrace->entry_pending = false;
2553         err = 0;
2554 out_put:
2555         thread__put(thread);
2556         return err;
2557 }
2558
2559 static int trace__vfs_getname(struct trace *trace, struct evsel *evsel,
2560                               union perf_event *event __maybe_unused,
2561                               struct perf_sample *sample)
2562 {
2563         struct thread *thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2564         struct thread_trace *ttrace;
2565         size_t filename_len, entry_str_len, to_move;
2566         ssize_t remaining_space;
2567         char *pos;
2568         const char *filename = evsel__rawptr(evsel, sample, "pathname");
2569
2570         if (!thread)
2571                 goto out;
2572
2573         ttrace = thread__priv(thread);
2574         if (!ttrace)
2575                 goto out_put;
2576
2577         filename_len = strlen(filename);
2578         if (filename_len == 0)
2579                 goto out_put;
2580
2581         if (ttrace->filename.namelen < filename_len) {
2582                 char *f = realloc(ttrace->filename.name, filename_len + 1);
2583
2584                 if (f == NULL)
2585                         goto out_put;
2586
2587                 ttrace->filename.namelen = filename_len;
2588                 ttrace->filename.name = f;
2589         }
2590
2591         strcpy(ttrace->filename.name, filename);
2592         ttrace->filename.pending_open = true;
2593
2594         if (!ttrace->filename.ptr)
2595                 goto out_put;
2596
2597         entry_str_len = strlen(ttrace->entry_str);
2598         remaining_space = trace__entry_str_size - entry_str_len - 1; /* \0 */
2599         if (remaining_space <= 0)
2600                 goto out_put;
2601
2602         if (filename_len > (size_t)remaining_space) {
2603                 filename += filename_len - remaining_space;
2604                 filename_len = remaining_space;
2605         }
2606
2607         to_move = entry_str_len - ttrace->filename.entry_str_pos + 1; /* \0 */
2608         pos = ttrace->entry_str + ttrace->filename.entry_str_pos;
2609         memmove(pos + filename_len, pos, to_move);
2610         memcpy(pos, filename, filename_len);
2611
2612         ttrace->filename.ptr = 0;
2613         ttrace->filename.entry_str_pos = 0;
2614 out_put:
2615         thread__put(thread);
2616 out:
2617         return 0;
2618 }
2619
2620 static int trace__sched_stat_runtime(struct trace *trace, struct evsel *evsel,
2621                                      union perf_event *event __maybe_unused,
2622                                      struct perf_sample *sample)
2623 {
2624         u64 runtime = evsel__intval(evsel, sample, "runtime");
2625         double runtime_ms = (double)runtime / NSEC_PER_MSEC;
2626         struct thread *thread = machine__findnew_thread(trace->host,
2627                                                         sample->pid,
2628                                                         sample->tid);
2629         struct thread_trace *ttrace = thread__trace(thread, trace->output);
2630
2631         if (ttrace == NULL)
2632                 goto out_dump;
2633
2634         ttrace->runtime_ms += runtime_ms;
2635         trace->runtime_ms += runtime_ms;
2636 out_put:
2637         thread__put(thread);
2638         return 0;
2639
2640 out_dump:
2641         fprintf(trace->output, "%s: comm=%s,pid=%u,runtime=%" PRIu64 ",vruntime=%" PRIu64 ")\n",
2642                evsel->name,
2643                evsel__strval(evsel, sample, "comm"),
2644                (pid_t)evsel__intval(evsel, sample, "pid"),
2645                runtime,
2646                evsel__intval(evsel, sample, "vruntime"));
2647         goto out_put;
2648 }
2649
2650 static int bpf_output__printer(enum binary_printer_ops op,
2651                                unsigned int val, void *extra __maybe_unused, FILE *fp)
2652 {
2653         unsigned char ch = (unsigned char)val;
2654
2655         switch (op) {
2656         case BINARY_PRINT_CHAR_DATA:
2657                 return fprintf(fp, "%c", isprint(ch) ? ch : '.');
2658         case BINARY_PRINT_DATA_BEGIN:
2659         case BINARY_PRINT_LINE_BEGIN:
2660         case BINARY_PRINT_ADDR:
2661         case BINARY_PRINT_NUM_DATA:
2662         case BINARY_PRINT_NUM_PAD:
2663         case BINARY_PRINT_SEP:
2664         case BINARY_PRINT_CHAR_PAD:
2665         case BINARY_PRINT_LINE_END:
2666         case BINARY_PRINT_DATA_END:
2667         default:
2668                 break;
2669         }
2670
2671         return 0;
2672 }
2673
2674 static void bpf_output__fprintf(struct trace *trace,
2675                                 struct perf_sample *sample)
2676 {
2677         binary__fprintf(sample->raw_data, sample->raw_size, 8,
2678                         bpf_output__printer, NULL, trace->output);
2679         ++trace->nr_events_printed;
2680 }
2681
2682 static size_t trace__fprintf_tp_fields(struct trace *trace, struct evsel *evsel, struct perf_sample *sample,
2683                                        struct thread *thread, void *augmented_args, int augmented_args_size)
2684 {
2685         char bf[2048];
2686         size_t size = sizeof(bf);
2687         struct tep_format_field *field = evsel->tp_format->format.fields;
2688         struct syscall_arg_fmt *arg = __evsel__syscall_arg_fmt(evsel);
2689         size_t printed = 0;
2690         unsigned long val;
2691         u8 bit = 1;
2692         struct syscall_arg syscall_arg = {
2693                 .augmented = {
2694                         .size = augmented_args_size,
2695                         .args = augmented_args,
2696                 },
2697                 .idx    = 0,
2698                 .mask   = 0,
2699                 .trace  = trace,
2700                 .thread = thread,
2701                 .show_string_prefix = trace->show_string_prefix,
2702         };
2703
2704         for (; field && arg; field = field->next, ++syscall_arg.idx, bit <<= 1, ++arg) {
2705                 if (syscall_arg.mask & bit)
2706                         continue;
2707
2708                 syscall_arg.len = 0;
2709                 syscall_arg.fmt = arg;
2710                 if (field->flags & TEP_FIELD_IS_ARRAY) {
2711                         int offset = field->offset;
2712
2713                         if (field->flags & TEP_FIELD_IS_DYNAMIC) {
2714                                 offset = format_field__intval(field, sample, evsel->needs_swap);
2715                                 syscall_arg.len = offset >> 16;
2716                                 offset &= 0xffff;
2717                         }
2718
2719                         val = (uintptr_t)(sample->raw_data + offset);
2720                 } else
2721                         val = format_field__intval(field, sample, evsel->needs_swap);
2722                 /*
2723                  * Some syscall args need some mask, most don't and
2724                  * return val untouched.
2725                  */
2726                 val = syscall_arg_fmt__mask_val(arg, &syscall_arg, val);
2727
2728                 /*
2729                  * Suppress this argument if its value is zero and
2730                  * and we don't have a string associated in an
2731                  * strarray for it.
2732                  */
2733                 if (val == 0 &&
2734                     !trace->show_zeros &&
2735                     !((arg->show_zero ||
2736                        arg->scnprintf == SCA_STRARRAY ||
2737                        arg->scnprintf == SCA_STRARRAYS) &&
2738                       arg->parm))
2739                         continue;
2740
2741                 printed += scnprintf(bf + printed, size - printed, "%s", printed ? ", " : "");
2742
2743                 /*
2744                  * XXX Perhaps we should have a show_tp_arg_names,
2745                  * leaving show_arg_names just for syscalls?
2746                  */
2747                 if (1 || trace->show_arg_names)
2748                         printed += scnprintf(bf + printed, size - printed, "%s: ", field->name);
2749
2750                 printed += syscall_arg_fmt__scnprintf_val(arg, bf + printed, size - printed, &syscall_arg, val);
2751         }
2752
2753         return printed + fprintf(trace->output, "%s", bf);
2754 }
2755
2756 static int trace__event_handler(struct trace *trace, struct evsel *evsel,
2757                                 union perf_event *event __maybe_unused,
2758                                 struct perf_sample *sample)
2759 {
2760         struct thread *thread;
2761         int callchain_ret = 0;
2762         /*
2763          * Check if we called perf_evsel__disable(evsel) due to, for instance,
2764          * this event's max_events having been hit and this is an entry coming
2765          * from the ring buffer that we should discard, since the max events
2766          * have already been considered/printed.
2767          */
2768         if (evsel->disabled)
2769                 return 0;
2770
2771         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2772
2773         if (sample->callchain) {
2774                 callchain_ret = trace__resolve_callchain(trace, evsel, sample, &callchain_cursor);
2775                 if (callchain_ret == 0) {
2776                         if (callchain_cursor.nr < trace->min_stack)
2777                                 goto out;
2778                         callchain_ret = 1;
2779                 }
2780         }
2781
2782         trace__printf_interrupted_entry(trace);
2783         trace__fprintf_tstamp(trace, sample->time, trace->output);
2784
2785         if (trace->trace_syscalls && trace->show_duration)
2786                 fprintf(trace->output, "(         ): ");
2787
2788         if (thread)
2789                 trace__fprintf_comm_tid(trace, thread, trace->output);
2790
2791         if (evsel == trace->syscalls.events.augmented) {
2792                 int id = perf_evsel__sc_tp_uint(evsel, id, sample);
2793                 struct syscall *sc = trace__syscall_info(trace, evsel, id);
2794
2795                 if (sc) {
2796                         fprintf(trace->output, "%s(", sc->name);
2797                         trace__fprintf_sys_enter(trace, evsel, sample);
2798                         fputc(')', trace->output);
2799                         goto newline;
2800                 }
2801
2802                 /*
2803                  * XXX: Not having the associated syscall info or not finding/adding
2804                  *      the thread should never happen, but if it does...
2805                  *      fall thru and print it as a bpf_output event.
2806                  */
2807         }
2808
2809         fprintf(trace->output, "%s(", evsel->name);
2810
2811         if (evsel__is_bpf_output(evsel)) {
2812                 bpf_output__fprintf(trace, sample);
2813         } else if (evsel->tp_format) {
2814                 if (strncmp(evsel->tp_format->name, "sys_enter_", 10) ||
2815                     trace__fprintf_sys_enter(trace, evsel, sample)) {
2816                         if (trace->libtraceevent_print) {
2817                                 event_format__fprintf(evsel->tp_format, sample->cpu,
2818                                                       sample->raw_data, sample->raw_size,
2819                                                       trace->output);
2820                         } else {
2821                                 trace__fprintf_tp_fields(trace, evsel, sample, thread, NULL, 0);
2822                         }
2823                 }
2824         }
2825
2826 newline:
2827         fprintf(trace->output, ")\n");
2828
2829         if (callchain_ret > 0)
2830                 trace__fprintf_callchain(trace, sample);
2831         else if (callchain_ret < 0)
2832                 pr_err("Problem processing %s callchain, skipping...\n", evsel__name(evsel));
2833
2834         ++trace->nr_events_printed;
2835
2836         if (evsel->max_events != ULONG_MAX && ++evsel->nr_events_printed == evsel->max_events) {
2837                 evsel__disable(evsel);
2838                 evsel__close(evsel);
2839         }
2840 out:
2841         thread__put(thread);
2842         return 0;
2843 }
2844
2845 static void print_location(FILE *f, struct perf_sample *sample,
2846                            struct addr_location *al,
2847                            bool print_dso, bool print_sym)
2848 {
2849
2850         if ((verbose > 0 || print_dso) && al->map)
2851                 fprintf(f, "%s@", al->map->dso->long_name);
2852
2853         if ((verbose > 0 || print_sym) && al->sym)
2854                 fprintf(f, "%s+0x%" PRIx64, al->sym->name,
2855                         al->addr - al->sym->start);
2856         else if (al->map)
2857                 fprintf(f, "0x%" PRIx64, al->addr);
2858         else
2859                 fprintf(f, "0x%" PRIx64, sample->addr);
2860 }
2861
2862 static int trace__pgfault(struct trace *trace,
2863                           struct evsel *evsel,
2864                           union perf_event *event __maybe_unused,
2865                           struct perf_sample *sample)
2866 {
2867         struct thread *thread;
2868         struct addr_location al;
2869         char map_type = 'd';
2870         struct thread_trace *ttrace;
2871         int err = -1;
2872         int callchain_ret = 0;
2873
2874         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2875
2876         if (sample->callchain) {
2877                 callchain_ret = trace__resolve_callchain(trace, evsel, sample, &callchain_cursor);
2878                 if (callchain_ret == 0) {
2879                         if (callchain_cursor.nr < trace->min_stack)
2880                                 goto out_put;
2881                         callchain_ret = 1;
2882                 }
2883         }
2884
2885         ttrace = thread__trace(thread, trace->output);
2886         if (ttrace == NULL)
2887                 goto out_put;
2888
2889         if (evsel->core.attr.config == PERF_COUNT_SW_PAGE_FAULTS_MAJ)
2890                 ttrace->pfmaj++;
2891         else
2892                 ttrace->pfmin++;
2893
2894         if (trace->summary_only)
2895                 goto out;
2896
2897         thread__find_symbol(thread, sample->cpumode, sample->ip, &al);
2898
2899         trace__fprintf_entry_head(trace, thread, 0, true, sample->time, trace->output);
2900
2901         fprintf(trace->output, "%sfault [",
2902                 evsel->core.attr.config == PERF_COUNT_SW_PAGE_FAULTS_MAJ ?
2903                 "maj" : "min");
2904
2905         print_location(trace->output, sample, &al, false, true);
2906
2907         fprintf(trace->output, "] => ");
2908
2909         thread__find_symbol(thread, sample->cpumode, sample->addr, &al);
2910
2911         if (!al.map) {
2912                 thread__find_symbol(thread, sample->cpumode, sample->addr, &al);
2913
2914                 if (al.map)
2915                         map_type = 'x';
2916                 else
2917                         map_type = '?';
2918         }
2919
2920         print_location(trace->output, sample, &al, true, false);
2921
2922         fprintf(trace->output, " (%c%c)\n", map_type, al.level);
2923
2924         if (callchain_ret > 0)
2925                 trace__fprintf_callchain(trace, sample);
2926         else if (callchain_ret < 0)
2927                 pr_err("Problem processing %s callchain, skipping...\n", evsel__name(evsel));
2928
2929         ++trace->nr_events_printed;
2930 out:
2931         err = 0;
2932 out_put:
2933         thread__put(thread);
2934         return err;
2935 }
2936
2937 static void trace__set_base_time(struct trace *trace,
2938                                  struct evsel *evsel,
2939                                  struct perf_sample *sample)
2940 {
2941         /*
2942          * BPF events were not setting PERF_SAMPLE_TIME, so be more robust
2943          * and don't use sample->time unconditionally, we may end up having
2944          * some other event in the future without PERF_SAMPLE_TIME for good
2945          * reason, i.e. we may not be interested in its timestamps, just in
2946          * it taking place, picking some piece of information when it
2947          * appears in our event stream (vfs_getname comes to mind).
2948          */
2949         if (trace->base_time == 0 && !trace->full_time &&
2950             (evsel->core.attr.sample_type & PERF_SAMPLE_TIME))
2951                 trace->base_time = sample->time;
2952 }
2953
2954 static int trace__process_sample(struct perf_tool *tool,
2955                                  union perf_event *event,
2956                                  struct perf_sample *sample,
2957                                  struct evsel *evsel,
2958                                  struct machine *machine __maybe_unused)
2959 {
2960         struct trace *trace = container_of(tool, struct trace, tool);
2961         struct thread *thread;
2962         int err = 0;
2963
2964         tracepoint_handler handler = evsel->handler;
2965
2966         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2967         if (thread && thread__is_filtered(thread))
2968                 goto out;
2969
2970         trace__set_base_time(trace, evsel, sample);
2971
2972         if (handler) {
2973                 ++trace->nr_events;
2974                 handler(trace, evsel, event, sample);
2975         }
2976 out:
2977         thread__put(thread);
2978         return err;
2979 }
2980
2981 static int trace__record(struct trace *trace, int argc, const char **argv)
2982 {
2983         unsigned int rec_argc, i, j;
2984         const char **rec_argv;
2985         const char * const record_args[] = {
2986                 "record",
2987                 "-R",
2988                 "-m", "1024",
2989                 "-c", "1",
2990         };
2991         pid_t pid = getpid();
2992         char *filter = asprintf__tp_filter_pids(1, &pid);
2993         const char * const sc_args[] = { "-e", };
2994         unsigned int sc_args_nr = ARRAY_SIZE(sc_args);
2995         const char * const majpf_args[] = { "-e", "major-faults" };
2996         unsigned int majpf_args_nr = ARRAY_SIZE(majpf_args);
2997         const char * const minpf_args[] = { "-e", "minor-faults" };
2998         unsigned int minpf_args_nr = ARRAY_SIZE(minpf_args);
2999         int err = -1;
3000
3001         /* +3 is for the event string below and the pid filter */
3002         rec_argc = ARRAY_SIZE(record_args) + sc_args_nr + 3 +
3003                 majpf_args_nr + minpf_args_nr + argc;
3004         rec_argv = calloc(rec_argc + 1, sizeof(char *));
3005
3006         if (rec_argv == NULL || filter == NULL)
3007                 goto out_free;
3008
3009         j = 0;
3010         for (i = 0; i < ARRAY_SIZE(record_args); i++)
3011                 rec_argv[j++] = record_args[i];
3012
3013         if (trace->trace_syscalls) {
3014                 for (i = 0; i < sc_args_nr; i++)
3015                         rec_argv[j++] = sc_args[i];
3016
3017                 /* event string may be different for older kernels - e.g., RHEL6 */
3018                 if (is_valid_tracepoint("raw_syscalls:sys_enter"))
3019                         rec_argv[j++] = "raw_syscalls:sys_enter,raw_syscalls:sys_exit";
3020                 else if (is_valid_tracepoint("syscalls:sys_enter"))
3021                         rec_argv[j++] = "syscalls:sys_enter,syscalls:sys_exit";
3022                 else {
3023                         pr_err("Neither raw_syscalls nor syscalls events exist.\n");
3024                         goto out_free;
3025                 }
3026         }
3027
3028         rec_argv[j++] = "--filter";
3029         rec_argv[j++] = filter;
3030
3031         if (trace->trace_pgfaults & TRACE_PFMAJ)
3032                 for (i = 0; i < majpf_args_nr; i++)
3033                         rec_argv[j++] = majpf_args[i];
3034
3035         if (trace->trace_pgfaults & TRACE_PFMIN)
3036                 for (i = 0; i < minpf_args_nr; i++)
3037                         rec_argv[j++] = minpf_args[i];
3038
3039         for (i = 0; i < (unsigned int)argc; i++)
3040                 rec_argv[j++] = argv[i];
3041
3042         err = cmd_record(j, rec_argv);
3043 out_free:
3044         free(filter);
3045         free(rec_argv);
3046         return err;
3047 }
3048
3049 static size_t trace__fprintf_thread_summary(struct trace *trace, FILE *fp);
3050
3051 static bool evlist__add_vfs_getname(struct evlist *evlist)
3052 {
3053         bool found = false;
3054         struct evsel *evsel, *tmp;
3055         struct parse_events_error err;
3056         int ret;
3057
3058         bzero(&err, sizeof(err));
3059         ret = parse_events(evlist, "probe:vfs_getname*", &err);
3060         if (ret) {
3061                 free(err.str);
3062                 free(err.help);
3063                 free(err.first_str);
3064                 free(err.first_help);
3065                 return false;
3066         }
3067
3068         evlist__for_each_entry_safe(evlist, evsel, tmp) {
3069                 if (!strstarts(evsel__name(evsel), "probe:vfs_getname"))
3070                         continue;
3071
3072                 if (evsel__field(evsel, "pathname")) {
3073                         evsel->handler = trace__vfs_getname;
3074                         found = true;
3075                         continue;
3076                 }
3077
3078                 list_del_init(&evsel->core.node);
3079                 evsel->evlist = NULL;
3080                 evsel__delete(evsel);
3081         }
3082
3083         return found;
3084 }
3085
3086 static struct evsel *evsel__new_pgfault(u64 config)
3087 {
3088         struct evsel *evsel;
3089         struct perf_event_attr attr = {
3090                 .type = PERF_TYPE_SOFTWARE,
3091                 .mmap_data = 1,
3092         };
3093
3094         attr.config = config;
3095         attr.sample_period = 1;
3096
3097         event_attr_init(&attr);
3098
3099         evsel = evsel__new(&attr);
3100         if (evsel)
3101                 evsel->handler = trace__pgfault;
3102
3103         return evsel;
3104 }
3105
3106 static void evlist__free_syscall_tp_fields(struct evlist *evlist)
3107 {
3108         struct evsel *evsel;
3109
3110         evlist__for_each_entry(evlist, evsel) {
3111                 struct evsel_trace *et = evsel->priv;
3112
3113                 if (!et || !evsel->tp_format || strcmp(evsel->tp_format->system, "syscalls"))
3114                         continue;
3115
3116                 free(et->fmt);
3117                 free(et);
3118         }
3119 }
3120
3121 static void trace__handle_event(struct trace *trace, union perf_event *event, struct perf_sample *sample)
3122 {
3123         const u32 type = event->header.type;
3124         struct evsel *evsel;
3125
3126         if (type != PERF_RECORD_SAMPLE) {
3127                 trace__process_event(trace, trace->host, event, sample);
3128                 return;
3129         }
3130
3131         evsel = evlist__id2evsel(trace->evlist, sample->id);
3132         if (evsel == NULL) {
3133                 fprintf(trace->output, "Unknown tp ID %" PRIu64 ", skipping...\n", sample->id);
3134                 return;
3135         }
3136
3137         if (evswitch__discard(&trace->evswitch, evsel))
3138                 return;
3139
3140         trace__set_base_time(trace, evsel, sample);
3141
3142         if (evsel->core.attr.type == PERF_TYPE_TRACEPOINT &&
3143             sample->raw_data == NULL) {
3144                 fprintf(trace->output, "%s sample with no payload for tid: %d, cpu %d, raw_size=%d, skipping...\n",
3145                        evsel__name(evsel), sample->tid,
3146                        sample->cpu, sample->raw_size);
3147         } else {
3148                 tracepoint_handler handler = evsel->handler;
3149                 handler(trace, evsel, event, sample);
3150         }
3151
3152         if (trace->nr_events_printed >= trace->max_events && trace->max_events != ULONG_MAX)
3153                 interrupted = true;
3154 }
3155
3156 static int trace__add_syscall_newtp(struct trace *trace)
3157 {
3158         int ret = -1;
3159         struct evlist *evlist = trace->evlist;
3160         struct evsel *sys_enter, *sys_exit;
3161
3162         sys_enter = perf_evsel__raw_syscall_newtp("sys_enter", trace__sys_enter);
3163         if (sys_enter == NULL)
3164                 goto out;
3165
3166         if (perf_evsel__init_sc_tp_ptr_field(sys_enter, args))
3167                 goto out_delete_sys_enter;
3168
3169         sys_exit = perf_evsel__raw_syscall_newtp("sys_exit", trace__sys_exit);
3170         if (sys_exit == NULL)
3171                 goto out_delete_sys_enter;
3172
3173         if (perf_evsel__init_sc_tp_uint_field(sys_exit, ret))
3174                 goto out_delete_sys_exit;
3175
3176         evsel__config_callchain(sys_enter, &trace->opts, &callchain_param);
3177         evsel__config_callchain(sys_exit, &trace->opts, &callchain_param);
3178
3179         evlist__add(evlist, sys_enter);
3180         evlist__add(evlist, sys_exit);
3181
3182         if (callchain_param.enabled && !trace->kernel_syscallchains) {
3183                 /*
3184                  * We're interested only in the user space callchain
3185                  * leading to the syscall, allow overriding that for
3186                  * debugging reasons using --kernel_syscall_callchains
3187                  */
3188                 sys_exit->core.attr.exclude_callchain_kernel = 1;
3189         }
3190
3191         trace->syscalls.events.sys_enter = sys_enter;
3192         trace->syscalls.events.sys_exit  = sys_exit;
3193
3194         ret = 0;
3195 out:
3196         return ret;
3197
3198 out_delete_sys_exit:
3199         evsel__delete_priv(sys_exit);
3200 out_delete_sys_enter:
3201         evsel__delete_priv(sys_enter);
3202         goto out;
3203 }
3204
3205 static int trace__set_ev_qualifier_tp_filter(struct trace *trace)
3206 {
3207         int err = -1;
3208         struct evsel *sys_exit;
3209         char *filter = asprintf_expr_inout_ints("id", !trace->not_ev_qualifier,
3210                                                 trace->ev_qualifier_ids.nr,
3211                                                 trace->ev_qualifier_ids.entries);
3212
3213         if (filter == NULL)
3214                 goto out_enomem;
3215
3216         if (!evsel__append_tp_filter(trace->syscalls.events.sys_enter, filter)) {
3217                 sys_exit = trace->syscalls.events.sys_exit;
3218                 err = evsel__append_tp_filter(sys_exit, filter);
3219         }
3220
3221         free(filter);
3222 out:
3223         return err;
3224 out_enomem:
3225         errno = ENOMEM;
3226         goto out;
3227 }
3228
3229 #ifdef HAVE_LIBBPF_SUPPORT
3230 static struct bpf_map *trace__find_bpf_map_by_name(struct trace *trace, const char *name)
3231 {
3232         if (trace->bpf_obj == NULL)
3233                 return NULL;
3234
3235         return bpf_object__find_map_by_name(trace->bpf_obj, name);
3236 }
3237
3238 static void trace__set_bpf_map_filtered_pids(struct trace *trace)
3239 {
3240         trace->filter_pids.map = trace__find_bpf_map_by_name(trace, "pids_filtered");
3241 }
3242
3243 static void trace__set_bpf_map_syscalls(struct trace *trace)
3244 {
3245         trace->syscalls.map = trace__find_bpf_map_by_name(trace, "syscalls");
3246         trace->syscalls.prog_array.sys_enter = trace__find_bpf_map_by_name(trace, "syscalls_sys_enter");
3247         trace->syscalls.prog_array.sys_exit  = trace__find_bpf_map_by_name(trace, "syscalls_sys_exit");
3248 }
3249
3250 static struct bpf_program *trace__find_bpf_program_by_title(struct trace *trace, const char *name)
3251 {
3252         if (trace->bpf_obj == NULL)
3253                 return NULL;
3254
3255         return bpf_object__find_program_by_title(trace->bpf_obj, name);
3256 }
3257
3258 static struct bpf_program *trace__find_syscall_bpf_prog(struct trace *trace, struct syscall *sc,
3259                                                         const char *prog_name, const char *type)
3260 {
3261         struct bpf_program *prog;
3262
3263         if (prog_name == NULL) {
3264                 char default_prog_name[256];
3265                 scnprintf(default_prog_name, sizeof(default_prog_name), "!syscalls:sys_%s_%s", type, sc->name);
3266                 prog = trace__find_bpf_program_by_title(trace, default_prog_name);
3267                 if (prog != NULL)
3268                         goto out_found;
3269                 if (sc->fmt && sc->fmt->alias) {
3270                         scnprintf(default_prog_name, sizeof(default_prog_name), "!syscalls:sys_%s_%s", type, sc->fmt->alias);
3271                         prog = trace__find_bpf_program_by_title(trace, default_prog_name);
3272                         if (prog != NULL)
3273                                 goto out_found;
3274                 }
3275                 goto out_unaugmented;
3276         }
3277
3278         prog = trace__find_bpf_program_by_title(trace, prog_name);
3279
3280         if (prog != NULL) {
3281 out_found:
3282                 return prog;
3283         }
3284
3285         pr_debug("Couldn't find BPF prog \"%s\" to associate with syscalls:sys_%s_%s, not augmenting it\n",
3286                  prog_name, type, sc->name);
3287 out_unaugmented:
3288         return trace->syscalls.unaugmented_prog;
3289 }
3290
3291 static void trace__init_syscall_bpf_progs(struct trace *trace, int id)
3292 {
3293         struct syscall *sc = trace__syscall_info(trace, NULL, id);
3294
3295         if (sc == NULL)
3296                 return;
3297
3298         sc->bpf_prog.sys_enter = trace__find_syscall_bpf_prog(trace, sc, sc->fmt ? sc->fmt->bpf_prog_name.sys_enter : NULL, "enter");
3299         sc->bpf_prog.sys_exit  = trace__find_syscall_bpf_prog(trace, sc, sc->fmt ? sc->fmt->bpf_prog_name.sys_exit  : NULL,  "exit");
3300 }
3301
3302 static int trace__bpf_prog_sys_enter_fd(struct trace *trace, int id)
3303 {
3304         struct syscall *sc = trace__syscall_info(trace, NULL, id);
3305         return sc ? bpf_program__fd(sc->bpf_prog.sys_enter) : bpf_program__fd(trace->syscalls.unaugmented_prog);
3306 }
3307
3308 static int trace__bpf_prog_sys_exit_fd(struct trace *trace, int id)
3309 {
3310         struct syscall *sc = trace__syscall_info(trace, NULL, id);
3311         return sc ? bpf_program__fd(sc->bpf_prog.sys_exit) : bpf_program__fd(trace->syscalls.unaugmented_prog);
3312 }
3313
3314 static void trace__init_bpf_map_syscall_args(struct trace *trace, int id, struct bpf_map_syscall_entry *entry)
3315 {
3316         struct syscall *sc = trace__syscall_info(trace, NULL, id);
3317         int arg = 0;
3318
3319         if (sc == NULL)
3320                 goto out;
3321
3322         for (; arg < sc->nr_args; ++arg) {
3323                 entry->string_args_len[arg] = 0;
3324                 if (sc->arg_fmt[arg].scnprintf == SCA_FILENAME) {
3325                         /* Should be set like strace -s strsize */
3326                         entry->string_args_len[arg] = PATH_MAX;
3327                 }
3328         }
3329 out:
3330         for (; arg < 6; ++arg)
3331                 entry->string_args_len[arg] = 0;
3332 }
3333 static int trace__set_ev_qualifier_bpf_filter(struct trace *trace)
3334 {
3335         int fd = bpf_map__fd(trace->syscalls.map);
3336         struct bpf_map_syscall_entry value = {
3337                 .enabled = !trace->not_ev_qualifier,
3338         };
3339         int err = 0;
3340         size_t i;
3341
3342         for (i = 0; i < trace->ev_qualifier_ids.nr; ++i) {
3343                 int key = trace->ev_qualifier_ids.entries[i];
3344
3345                 if (value.enabled) {
3346                         trace__init_bpf_map_syscall_args(trace, key, &value);
3347                         trace__init_syscall_bpf_progs(trace, key);
3348                 }
3349
3350                 err = bpf_map_update_elem(fd, &key, &value, BPF_EXIST);
3351                 if (err)
3352                         break;
3353         }
3354
3355         return err;
3356 }
3357
3358 static int __trace__init_syscalls_bpf_map(struct trace *trace, bool enabled)
3359 {
3360         int fd = bpf_map__fd(trace->syscalls.map);
3361         struct bpf_map_syscall_entry value = {
3362                 .enabled = enabled,
3363         };
3364         int err = 0, key;
3365
3366         for (key = 0; key < trace->sctbl->syscalls.nr_entries; ++key) {
3367                 if (enabled)
3368                         trace__init_bpf_map_syscall_args(trace, key, &value);
3369
3370                 err = bpf_map_update_elem(fd, &key, &value, BPF_ANY);
3371                 if (err)
3372                         break;
3373         }
3374
3375         return err;
3376 }
3377
3378 static int trace__init_syscalls_bpf_map(struct trace *trace)
3379 {
3380         bool enabled = true;
3381
3382         if (trace->ev_qualifier_ids.nr)
3383                 enabled = trace->not_ev_qualifier;
3384
3385         return __trace__init_syscalls_bpf_map(trace, enabled);
3386 }
3387
3388 static struct bpf_program *trace__find_usable_bpf_prog_entry(struct trace *trace, struct syscall *sc)
3389 {
3390         struct tep_format_field *field, *candidate_field;
3391         int id;
3392
3393         /*
3394          * We're only interested in syscalls that have a pointer:
3395          */
3396         for (field = sc->args; field; field = field->next) {
3397                 if (field->flags & TEP_FIELD_IS_POINTER)
3398                         goto try_to_find_pair;
3399         }
3400
3401         return NULL;
3402
3403 try_to_find_pair:
3404         for (id = 0; id < trace->sctbl->syscalls.nr_entries; ++id) {
3405                 struct syscall *pair = trace__syscall_info(trace, NULL, id);
3406                 struct bpf_program *pair_prog;
3407                 bool is_candidate = false;
3408
3409                 if (pair == NULL || pair == sc ||
3410                     pair->bpf_prog.sys_enter == trace->syscalls.unaugmented_prog)
3411                         continue;
3412
3413                 for (field = sc->args, candidate_field = pair->args;
3414                      field && candidate_field; field = field->next, candidate_field = candidate_field->next) {
3415                         bool is_pointer = field->flags & TEP_FIELD_IS_POINTER,
3416                              candidate_is_pointer = candidate_field->flags & TEP_FIELD_IS_POINTER;
3417
3418                         if (is_pointer) {
3419                                if (!candidate_is_pointer) {
3420                                         // The candidate just doesn't copies our pointer arg, might copy other pointers we want.
3421                                         continue;
3422                                }
3423                         } else {
3424                                 if (candidate_is_pointer) {
3425                                         // The candidate might copy a pointer we don't have, skip it.
3426                                         goto next_candidate;
3427                                 }
3428                                 continue;
3429                         }
3430
3431                         if (strcmp(field->type, candidate_field->type))
3432                                 goto next_candidate;
3433
3434                         is_candidate = true;
3435                 }
3436
3437                 if (!is_candidate)
3438                         goto next_candidate;
3439
3440                 /*
3441                  * Check if the tentative pair syscall augmenter has more pointers, if it has,
3442                  * then it may be collecting that and we then can't use it, as it would collect
3443                  * more than what is common to the two syscalls.
3444                  */
3445                 if (candidate_field) {
3446                         for (candidate_field = candidate_field->next; candidate_field; candidate_field = candidate_field->next)
3447                                 if (candidate_field->flags & TEP_FIELD_IS_POINTER)
3448                                         goto next_candidate;
3449                 }
3450
3451                 pair_prog = pair->bpf_prog.sys_enter;
3452                 /*
3453                  * If the pair isn't enabled, then its bpf_prog.sys_enter will not
3454                  * have been searched for, so search it here and if it returns the
3455                  * unaugmented one, then ignore it, otherwise we'll reuse that BPF
3456                  * program for a filtered syscall on a non-filtered one.
3457                  *
3458                  * For instance, we have "!syscalls:sys_enter_renameat" and that is
3459                  * useful for "renameat2".
3460                  */
3461                 if (pair_prog == NULL) {
3462                         pair_prog = trace__find_syscall_bpf_prog(trace, pair, pair->fmt ? pair->fmt->bpf_prog_name.sys_enter : NULL, "enter");
3463                         if (pair_prog == trace->syscalls.unaugmented_prog)
3464                                 goto next_candidate;
3465                 }
3466
3467                 pr_debug("Reusing \"%s\" BPF sys_enter augmenter for \"%s\"\n", pair->name, sc->name);
3468                 return pair_prog;
3469         next_candidate:
3470                 continue;
3471         }
3472
3473         return NULL;
3474 }
3475
3476 static int trace__init_syscalls_bpf_prog_array_maps(struct trace *trace)
3477 {
3478         int map_enter_fd = bpf_map__fd(trace->syscalls.prog_array.sys_enter),
3479             map_exit_fd  = bpf_map__fd(trace->syscalls.prog_array.sys_exit);
3480         int err = 0, key;
3481
3482         for (key = 0; key < trace->sctbl->syscalls.nr_entries; ++key) {
3483                 int prog_fd;
3484
3485                 if (!trace__syscall_enabled(trace, key))
3486                         continue;
3487
3488                 trace__init_syscall_bpf_progs(trace, key);
3489
3490                 // It'll get at least the "!raw_syscalls:unaugmented"
3491                 prog_fd = trace__bpf_prog_sys_enter_fd(trace, key);
3492                 err = bpf_map_update_elem(map_enter_fd, &key, &prog_fd, BPF_ANY);
3493                 if (err)
3494                         break;
3495                 prog_fd = trace__bpf_prog_sys_exit_fd(trace, key);
3496                 err = bpf_map_update_elem(map_exit_fd, &key, &prog_fd, BPF_ANY);
3497                 if (err)
3498                         break;
3499         }
3500
3501         /*
3502          * Now lets do a second pass looking for enabled syscalls without
3503          * an augmenter that have a signature that is a superset of another
3504          * syscall with an augmenter so that we can auto-reuse it.
3505          *
3506          * I.e. if we have an augmenter for the "open" syscall that has
3507          * this signature:
3508          *
3509          *   int open(const char *pathname, int flags, mode_t mode);
3510          *
3511          * I.e. that will collect just the first string argument, then we
3512          * can reuse it for the 'creat' syscall, that has this signature:
3513          *
3514          *   int creat(const char *pathname, mode_t mode);
3515          *
3516          * and for:
3517          *
3518          *   int stat(const char *pathname, struct stat *statbuf);
3519          *   int lstat(const char *pathname, struct stat *statbuf);
3520          *
3521          * Because the 'open' augmenter will collect the first arg as a string,
3522          * and leave alone all the other args, which already helps with
3523          * beautifying 'stat' and 'lstat''s pathname arg.
3524          *
3525          * Then, in time, when 'stat' gets an augmenter that collects both
3526          * first and second arg (this one on the raw_syscalls:sys_exit prog
3527          * array tail call, then that one will be used.
3528          */
3529         for (key = 0; key < trace->sctbl->syscalls.nr_entries; ++key) {
3530                 struct syscall *sc = trace__syscall_info(trace, NULL, key);
3531                 struct bpf_program *pair_prog;
3532                 int prog_fd;
3533
3534                 if (sc == NULL || sc->bpf_prog.sys_enter == NULL)
3535                         continue;
3536
3537                 /*
3538                  * For now we're just reusing the sys_enter prog, and if it
3539                  * already has an augmenter, we don't need to find one.
3540                  */
3541                 if (sc->bpf_prog.sys_enter != trace->syscalls.unaugmented_prog)
3542                         continue;
3543
3544                 /*
3545                  * Look at all the other syscalls for one that has a signature
3546                  * that is close enough that we can share:
3547                  */
3548                 pair_prog = trace__find_usable_bpf_prog_entry(trace, sc);
3549                 if (pair_prog == NULL)
3550                         continue;
3551
3552                 sc->bpf_prog.sys_enter = pair_prog;
3553
3554                 /*
3555                  * Update the BPF_MAP_TYPE_PROG_SHARED for raw_syscalls:sys_enter
3556                  * with the fd for the program we're reusing:
3557                  */
3558                 prog_fd = bpf_program__fd(sc->bpf_prog.sys_enter);
3559                 err = bpf_map_update_elem(map_enter_fd, &key, &prog_fd, BPF_ANY);
3560                 if (err)
3561                         break;
3562         }
3563
3564
3565         return err;
3566 }
3567
3568 static void trace__delete_augmented_syscalls(struct trace *trace)
3569 {
3570         struct evsel *evsel, *tmp;
3571
3572         evlist__remove(trace->evlist, trace->syscalls.events.augmented);
3573         evsel__delete(trace->syscalls.events.augmented);
3574         trace->syscalls.events.augmented = NULL;
3575
3576         evlist__for_each_entry_safe(trace->evlist, tmp, evsel) {
3577                 if (evsel->bpf_obj == trace->bpf_obj) {
3578                         evlist__remove(trace->evlist, evsel);
3579                         evsel__delete(evsel);
3580                 }
3581
3582         }
3583
3584         bpf_object__close(trace->bpf_obj);
3585         trace->bpf_obj = NULL;
3586 }
3587 #else // HAVE_LIBBPF_SUPPORT
3588 static struct bpf_map *trace__find_bpf_map_by_name(struct trace *trace __maybe_unused,
3589                                                    const char *name __maybe_unused)
3590 {
3591         return NULL;
3592 }
3593
3594 static void trace__set_bpf_map_filtered_pids(struct trace *trace __maybe_unused)
3595 {
3596 }
3597
3598 static void trace__set_bpf_map_syscalls(struct trace *trace __maybe_unused)
3599 {
3600 }
3601
3602 static int trace__set_ev_qualifier_bpf_filter(struct trace *trace __maybe_unused)
3603 {
3604         return 0;
3605 }
3606
3607 static int trace__init_syscalls_bpf_map(struct trace *trace __maybe_unused)
3608 {
3609         return 0;
3610 }
3611
3612 static struct bpf_program *trace__find_bpf_program_by_title(struct trace *trace __maybe_unused,
3613                                                             const char *name __maybe_unused)
3614 {
3615         return NULL;
3616 }
3617
3618 static int trace__init_syscalls_bpf_prog_array_maps(struct trace *trace __maybe_unused)
3619 {
3620         return 0;
3621 }
3622
3623 static void trace__delete_augmented_syscalls(struct trace *trace __maybe_unused)
3624 {
3625 }
3626 #endif // HAVE_LIBBPF_SUPPORT
3627
3628 static bool trace__only_augmented_syscalls_evsels(struct trace *trace)
3629 {
3630         struct evsel *evsel;
3631
3632         evlist__for_each_entry(trace->evlist, evsel) {
3633                 if (evsel == trace->syscalls.events.augmented ||
3634                     evsel->bpf_obj == trace->bpf_obj)
3635                         continue;
3636
3637                 return false;
3638         }
3639
3640         return true;
3641 }
3642
3643 static int trace__set_ev_qualifier_filter(struct trace *trace)
3644 {
3645         if (trace->syscalls.map)
3646                 return trace__set_ev_qualifier_bpf_filter(trace);
3647         if (trace->syscalls.events.sys_enter)
3648                 return trace__set_ev_qualifier_tp_filter(trace);
3649         return 0;
3650 }
3651
3652 static int bpf_map__set_filter_pids(struct bpf_map *map __maybe_unused,
3653                                     size_t npids __maybe_unused, pid_t *pids __maybe_unused)
3654 {
3655         int err = 0;
3656 #ifdef HAVE_LIBBPF_SUPPORT
3657         bool value = true;
3658         int map_fd = bpf_map__fd(map);
3659         size_t i;
3660
3661         for (i = 0; i < npids; ++i) {
3662                 err = bpf_map_update_elem(map_fd, &pids[i], &value, BPF_ANY);
3663                 if (err)
3664                         break;
3665         }
3666 #endif
3667         return err;
3668 }
3669
3670 static int trace__set_filter_loop_pids(struct trace *trace)
3671 {
3672         unsigned int nr = 1, err;
3673         pid_t pids[32] = {
3674                 getpid(),
3675         };
3676         struct thread *thread = machine__find_thread(trace->host, pids[0], pids[0]);
3677
3678         while (thread && nr < ARRAY_SIZE(pids)) {
3679                 struct thread *parent = machine__find_thread(trace->host, thread->ppid, thread->ppid);
3680
3681                 if (parent == NULL)
3682                         break;
3683
3684                 if (!strcmp(thread__comm_str(parent), "sshd") ||
3685                     strstarts(thread__comm_str(parent), "gnome-terminal")) {
3686                         pids[nr++] = parent->tid;
3687                         break;
3688                 }
3689                 thread = parent;
3690         }
3691
3692         err = evlist__append_tp_filter_pids(trace->evlist, nr, pids);
3693         if (!err && trace->filter_pids.map)
3694                 err = bpf_map__set_filter_pids(trace->filter_pids.map, nr, pids);
3695
3696         return err;
3697 }
3698
3699 static int trace__set_filter_pids(struct trace *trace)
3700 {
3701         int err = 0;
3702         /*
3703          * Better not use !target__has_task() here because we need to cover the
3704          * case where no threads were specified in the command line, but a
3705          * workload was, and in that case we will fill in the thread_map when
3706          * we fork the workload in evlist__prepare_workload.
3707          */
3708         if (trace->filter_pids.nr > 0) {
3709                 err = evlist__append_tp_filter_pids(trace->evlist, trace->filter_pids.nr,
3710                                                     trace->filter_pids.entries);
3711                 if (!err && trace->filter_pids.map) {
3712                         err = bpf_map__set_filter_pids(trace->filter_pids.map, trace->filter_pids.nr,
3713                                                        trace->filter_pids.entries);
3714                 }
3715         } else if (perf_thread_map__pid(trace->evlist->core.threads, 0) == -1) {
3716                 err = trace__set_filter_loop_pids(trace);
3717         }
3718
3719         return err;
3720 }
3721
3722 static int __trace__deliver_event(struct trace *trace, union perf_event *event)
3723 {
3724         struct evlist *evlist = trace->evlist;
3725         struct perf_sample sample;
3726         int err = evlist__parse_sample(evlist, event, &sample);
3727
3728         if (err)
3729                 fprintf(trace->output, "Can't parse sample, err = %d, skipping...\n", err);
3730         else
3731                 trace__handle_event(trace, event, &sample);
3732
3733         return 0;
3734 }
3735
3736 static int __trace__flush_events(struct trace *trace)
3737 {
3738         u64 first = ordered_events__first_time(&trace->oe.data);
3739         u64 flush = trace->oe.last - NSEC_PER_SEC;
3740
3741         /* Is there some thing to flush.. */
3742         if (first && first < flush)
3743                 return ordered_events__flush_time(&trace->oe.data, flush);
3744
3745         return 0;
3746 }
3747
3748 static int trace__flush_events(struct trace *trace)
3749 {
3750         return !trace->sort_events ? 0 : __trace__flush_events(trace);
3751 }
3752
3753 static int trace__deliver_event(struct trace *trace, union perf_event *event)
3754 {
3755         int err;
3756
3757         if (!trace->sort_events)
3758                 return __trace__deliver_event(trace, event);
3759
3760         err = evlist__parse_sample_timestamp(trace->evlist, event, &trace->oe.last);
3761         if (err && err != -1)
3762                 return err;
3763
3764         err = ordered_events__queue(&trace->oe.data, event, trace->oe.last, 0);
3765         if (err)
3766                 return err;
3767
3768         return trace__flush_events(trace);
3769 }
3770
3771 static int ordered_events__deliver_event(struct ordered_events *oe,
3772                                          struct ordered_event *event)
3773 {
3774         struct trace *trace = container_of(oe, struct trace, oe.data);
3775
3776         return __trace__deliver_event(trace, event->event);
3777 }
3778
3779 static struct syscall_arg_fmt *evsel__find_syscall_arg_fmt_by_name(struct evsel *evsel, char *arg)
3780 {
3781         struct tep_format_field *field;
3782         struct syscall_arg_fmt *fmt = __evsel__syscall_arg_fmt(evsel);
3783
3784         if (evsel->tp_format == NULL || fmt == NULL)
3785                 return NULL;
3786
3787         for (field = evsel->tp_format->format.fields; field; field = field->next, ++fmt)
3788                 if (strcmp(field->name, arg) == 0)
3789                         return fmt;
3790
3791         return NULL;
3792 }
3793
3794 static int trace__expand_filter(struct trace *trace __maybe_unused, struct evsel *evsel)
3795 {
3796         char *tok, *left = evsel->filter, *new_filter = evsel->filter;
3797
3798         while ((tok = strpbrk(left, "=<>!")) != NULL) {
3799                 char *right = tok + 1, *right_end;
3800
3801                 if (*right == '=')
3802                         ++right;
3803
3804                 while (isspace(*right))
3805                         ++right;
3806
3807                 if (*right == '\0')
3808                         break;
3809
3810                 while (!isalpha(*left))
3811                         if (++left == tok) {
3812                                 /*
3813                                  * Bail out, can't find the name of the argument that is being
3814                                  * used in the filter, let it try to set this filter, will fail later.
3815                                  */
3816                                 return 0;
3817                         }
3818
3819                 right_end = right + 1;
3820                 while (isalnum(*right_end) || *right_end == '_' || *right_end == '|')
3821                         ++right_end;
3822
3823                 if (isalpha(*right)) {
3824                         struct syscall_arg_fmt *fmt;
3825                         int left_size = tok - left,
3826                             right_size = right_end - right;
3827                         char arg[128];
3828
3829                         while (isspace(left[left_size - 1]))
3830                                 --left_size;
3831
3832                         scnprintf(arg, sizeof(arg), "%.*s", left_size, left);
3833
3834                         fmt = evsel__find_syscall_arg_fmt_by_name(evsel, arg);
3835                         if (fmt == NULL) {
3836                                 pr_err("\"%s\" not found in \"%s\", can't set filter \"%s\"\n",
3837                                        arg, evsel->name, evsel->filter);
3838                                 return -1;
3839                         }
3840
3841                         pr_debug2("trying to expand \"%s\" \"%.*s\" \"%.*s\" -> ",
3842                                  arg, (int)(right - tok), tok, right_size, right);
3843
3844                         if (fmt->strtoul) {
3845                                 u64 val;
3846                                 struct syscall_arg syscall_arg = {
3847                                         .parm = fmt->parm,
3848                                 };
3849
3850                                 if (fmt->strtoul(right, right_size, &syscall_arg, &val)) {
3851                                         char *n, expansion[19];
3852                                         int expansion_lenght = scnprintf(expansion, sizeof(expansion), "%#" PRIx64, val);
3853                                         int expansion_offset = right - new_filter;
3854
3855                                         pr_debug("%s", expansion);
3856
3857                                         if (asprintf(&n, "%.*s%s%s", expansion_offset, new_filter, expansion, right_end) < 0) {
3858                                                 pr_debug(" out of memory!\n");
3859                                                 free(new_filter);
3860                                                 return -1;
3861                                         }
3862                                         if (new_filter != evsel->filter)
3863                                                 free(new_filter);
3864                                         left = n + expansion_offset + expansion_lenght;
3865                                         new_filter = n;
3866                                 } else {
3867                                         pr_err("\"%.*s\" not found for \"%s\" in \"%s\", can't set filter \"%s\"\n",
3868                                                right_size, right, arg, evsel->name, evsel->filter);
3869                                         return -1;
3870                                 }
3871                         } else {
3872                                 pr_err("No resolver (strtoul) for \"%s\" in \"%s\", can't set filter \"%s\"\n",
3873                                        arg, evsel->name, evsel->filter);
3874                                 return -1;
3875                         }
3876
3877                         pr_debug("\n");
3878                 } else {
3879                         left = right_end;
3880                 }
3881         }
3882
3883         if (new_filter != evsel->filter) {
3884                 pr_debug("New filter for %s: %s\n", evsel->name, new_filter);
3885                 evsel__set_filter(evsel, new_filter);
3886                 free(new_filter);
3887         }
3888
3889         return 0;
3890 }
3891
3892 static int trace__expand_filters(struct trace *trace, struct evsel **err_evsel)
3893 {
3894         struct evlist *evlist = trace->evlist;
3895         struct evsel *evsel;
3896
3897         evlist__for_each_entry(evlist, evsel) {
3898                 if (evsel->filter == NULL)
3899                         continue;
3900
3901                 if (trace__expand_filter(trace, evsel)) {
3902                         *err_evsel = evsel;
3903                         return -1;
3904                 }
3905         }
3906
3907         return 0;
3908 }
3909
3910 static int trace__run(struct trace *trace, int argc, const char **argv)
3911 {
3912         struct evlist *evlist = trace->evlist;
3913         struct evsel *evsel, *pgfault_maj = NULL, *pgfault_min = NULL;
3914         int err = -1, i;
3915         unsigned long before;
3916         const bool forks = argc > 0;
3917         bool draining = false;
3918
3919         trace->live = true;
3920
3921         if (!trace->raw_augmented_syscalls) {
3922                 if (trace->trace_syscalls && trace__add_syscall_newtp(trace))
3923                         goto out_error_raw_syscalls;
3924
3925                 if (trace->trace_syscalls)
3926                         trace->vfs_getname = evlist__add_vfs_getname(evlist);
3927         }
3928
3929         if ((trace->trace_pgfaults & TRACE_PFMAJ)) {
3930                 pgfault_maj = evsel__new_pgfault(PERF_COUNT_SW_PAGE_FAULTS_MAJ);
3931                 if (pgfault_maj == NULL)
3932                         goto out_error_mem;
3933                 evsel__config_callchain(pgfault_maj, &trace->opts, &callchain_param);
3934                 evlist__add(evlist, pgfault_maj);
3935         }
3936
3937         if ((trace->trace_pgfaults & TRACE_PFMIN)) {
3938                 pgfault_min = evsel__new_pgfault(PERF_COUNT_SW_PAGE_FAULTS_MIN);
3939                 if (pgfault_min == NULL)
3940                         goto out_error_mem;
3941                 evsel__config_callchain(pgfault_min, &trace->opts, &callchain_param);
3942                 evlist__add(evlist, pgfault_min);
3943         }
3944
3945         if (trace->sched &&
3946             evlist__add_newtp(evlist, "sched", "sched_stat_runtime", trace__sched_stat_runtime))
3947                 goto out_error_sched_stat_runtime;
3948         /*
3949          * If a global cgroup was set, apply it to all the events without an
3950          * explicit cgroup. I.e.:
3951          *
3952          *      trace -G A -e sched:*switch
3953          *
3954          * Will set all raw_syscalls:sys_{enter,exit}, pgfault, vfs_getname, etc
3955          * _and_ sched:sched_switch to the 'A' cgroup, while:
3956          *
3957          * trace -e sched:*switch -G A
3958          *
3959          * will only set the sched:sched_switch event to the 'A' cgroup, all the
3960          * other events (raw_syscalls:sys_{enter,exit}, etc are left "without"
3961          * a cgroup (on the root cgroup, sys wide, etc).
3962          *
3963          * Multiple cgroups:
3964          *
3965          * trace -G A -e sched:*switch -G B
3966          *
3967          * the syscall ones go to the 'A' cgroup, the sched:sched_switch goes
3968          * to the 'B' cgroup.
3969          *
3970          * evlist__set_default_cgroup() grabs a reference of the passed cgroup
3971          * only for the evsels still without a cgroup, i.e. evsel->cgroup == NULL.
3972          */
3973         if (trace->cgroup)
3974                 evlist__set_default_cgroup(trace->evlist, trace->cgroup);
3975
3976         err = evlist__create_maps(evlist, &trace->opts.target);
3977         if (err < 0) {
3978                 fprintf(trace->output, "Problems parsing the target to trace, check your options!\n");
3979                 goto out_delete_evlist;
3980         }
3981
3982         err = trace__symbols_init(trace, evlist);
3983         if (err < 0) {
3984                 fprintf(trace->output, "Problems initializing symbol libraries!\n");
3985                 goto out_delete_evlist;
3986         }
3987
3988         evlist__config(evlist, &trace->opts, &callchain_param);
3989
3990         if (forks) {
3991                 err = evlist__prepare_workload(evlist, &trace->opts.target, argv, false, NULL);
3992                 if (err < 0) {
3993                         fprintf(trace->output, "Couldn't run the workload!\n");
3994                         goto out_delete_evlist;
3995                 }
3996         }
3997
3998         err = evlist__open(evlist);
3999         if (err < 0)
4000                 goto out_error_open;
4001
4002         err = bpf__apply_obj_config();
4003         if (err) {
4004                 char errbuf[BUFSIZ];
4005
4006                 bpf__strerror_apply_obj_config(err, errbuf, sizeof(errbuf));
4007                 pr_err("ERROR: Apply config to BPF failed: %s\n",
4008                          errbuf);
4009                 goto out_error_open;
4010         }
4011
4012         err = trace__set_filter_pids(trace);
4013         if (err < 0)
4014                 goto out_error_mem;
4015
4016         if (trace->syscalls.map)
4017                 trace__init_syscalls_bpf_map(trace);
4018
4019         if (trace->syscalls.prog_array.sys_enter)
4020                 trace__init_syscalls_bpf_prog_array_maps(trace);
4021
4022         if (trace->ev_qualifier_ids.nr > 0) {
4023                 err = trace__set_ev_qualifier_filter(trace);
4024                 if (err < 0)
4025                         goto out_errno;
4026
4027                 if (trace->syscalls.events.sys_exit) {
4028                         pr_debug("event qualifier tracepoint filter: %s\n",
4029                                  trace->syscalls.events.sys_exit->filter);
4030                 }
4031         }
4032
4033         /*
4034          * If the "close" syscall is not traced, then we will not have the
4035          * opportunity to, in syscall_arg__scnprintf_close_fd() invalidate the
4036          * fd->pathname table and were ending up showing the last value set by
4037          * syscalls opening a pathname and associating it with a descriptor or
4038          * reading it from /proc/pid/fd/ in cases where that doesn't make
4039          * sense.
4040          *
4041          *  So just disable this beautifier (SCA_FD, SCA_FDAT) when 'close' is
4042          *  not in use.
4043          */
4044         trace->fd_path_disabled = !trace__syscall_enabled(trace, syscalltbl__id(trace->sctbl, "close"));
4045
4046         err = trace__expand_filters(trace, &evsel);
4047         if (err)
4048                 goto out_delete_evlist;
4049         err = evlist__apply_filters(evlist, &evsel);
4050         if (err < 0)
4051                 goto out_error_apply_filters;
4052
4053         if (trace->dump.map)
4054                 bpf_map__fprintf(trace->dump.map, trace->output);
4055
4056         err = evlist__mmap(evlist, trace->opts.mmap_pages);
4057         if (err < 0)
4058                 goto out_error_mmap;
4059
4060         if (!target__none(&trace->opts.target) && !trace->opts.initial_delay)
4061                 evlist__enable(evlist);
4062
4063         if (forks)
4064                 evlist__start_workload(evlist);
4065
4066         if (trace->opts.initial_delay) {
4067                 usleep(trace->opts.initial_delay * 1000);
4068                 evlist__enable(evlist);
4069         }
4070
4071         trace->multiple_threads = perf_thread_map__pid(evlist->core.threads, 0) == -1 ||
4072                                   evlist->core.threads->nr > 1 ||
4073                                   evlist__first(evlist)->core.attr.inherit;
4074
4075         /*
4076          * Now that we already used evsel->core.attr to ask the kernel to setup the
4077          * events, lets reuse evsel->core.attr.sample_max_stack as the limit in
4078          * trace__resolve_callchain(), allowing per-event max-stack settings
4079          * to override an explicitly set --max-stack global setting.
4080          */
4081         evlist__for_each_entry(evlist, evsel) {
4082                 if (evsel__has_callchain(evsel) &&
4083                     evsel->core.attr.sample_max_stack == 0)
4084                         evsel->core.attr.sample_max_stack = trace->max_stack;
4085         }
4086 again:
4087         before = trace->nr_events;
4088
4089         for (i = 0; i < evlist->core.nr_mmaps; i++) {
4090                 union perf_event *event;
4091                 struct mmap *md;
4092
4093                 md = &evlist->mmap[i];
4094                 if (perf_mmap__read_init(&md->core) < 0)
4095                         continue;
4096
4097                 while ((event = perf_mmap__read_event(&md->core)) != NULL) {
4098                         ++trace->nr_events;
4099
4100                         err = trace__deliver_event(trace, event);
4101                         if (err)
4102                                 goto out_disable;
4103
4104                         perf_mmap__consume(&md->core);
4105
4106                         if (interrupted)
4107                                 goto out_disable;
4108
4109                         if (done && !draining) {
4110                                 evlist__disable(evlist);
4111                                 draining = true;
4112                         }
4113                 }
4114                 perf_mmap__read_done(&md->core);
4115         }
4116
4117         if (trace->nr_events == before) {
4118                 int timeout = done ? 100 : -1;
4119
4120                 if (!draining && evlist__poll(evlist, timeout) > 0) {
4121                         if (evlist__filter_pollfd(evlist, POLLERR | POLLHUP | POLLNVAL) == 0)
4122                                 draining = true;
4123
4124                         goto again;
4125                 } else {
4126                         if (trace__flush_events(trace))
4127                                 goto out_disable;
4128                 }
4129         } else {
4130                 goto again;
4131         }
4132
4133 out_disable:
4134         thread__zput(trace->current);
4135
4136         evlist__disable(evlist);
4137
4138         if (trace->sort_events)
4139                 ordered_events__flush(&trace->oe.data, OE_FLUSH__FINAL);
4140
4141         if (!err) {
4142                 if (trace->summary)
4143                         trace__fprintf_thread_summary(trace, trace->output);
4144
4145                 if (trace->show_tool_stats) {
4146                         fprintf(trace->output, "Stats:\n "
4147                                                " vfs_getname : %" PRIu64 "\n"
4148                                                " proc_getname: %" PRIu64 "\n",
4149                                 trace->stats.vfs_getname,
4150                                 trace->stats.proc_getname);
4151                 }
4152         }
4153
4154 out_delete_evlist:
4155         trace__symbols__exit(trace);
4156         evlist__free_syscall_tp_fields(evlist);
4157         evlist__delete(evlist);
4158         cgroup__put(trace->cgroup);
4159         trace->evlist = NULL;
4160         trace->live = false;
4161         return err;
4162 {
4163         char errbuf[BUFSIZ];
4164
4165 out_error_sched_stat_runtime:
4166         tracing_path__strerror_open_tp(errno, errbuf, sizeof(errbuf), "sched", "sched_stat_runtime");
4167         goto out_error;
4168
4169 out_error_raw_syscalls:
4170         tracing_path__strerror_open_tp(errno, errbuf, sizeof(errbuf), "raw_syscalls", "sys_(enter|exit)");
4171         goto out_error;
4172
4173 out_error_mmap:
4174         evlist__strerror_mmap(evlist, errno, errbuf, sizeof(errbuf));
4175         goto out_error;
4176
4177 out_error_open:
4178         evlist__strerror_open(evlist, errno, errbuf, sizeof(errbuf));
4179
4180 out_error:
4181         fprintf(trace->output, "%s\n", errbuf);
4182         goto out_delete_evlist;
4183
4184 out_error_apply_filters:
4185         fprintf(trace->output,
4186                 "Failed to set filter \"%s\" on event %s with %d (%s)\n",
4187                 evsel->filter, evsel__name(evsel), errno,
4188                 str_error_r(errno, errbuf, sizeof(errbuf)));
4189         goto out_delete_evlist;
4190 }
4191 out_error_mem:
4192         fprintf(trace->output, "Not enough memory to run!\n");
4193         goto out_delete_evlist;
4194
4195 out_errno:
4196         fprintf(trace->output, "errno=%d,%s\n", errno, strerror(errno));
4197         goto out_delete_evlist;
4198 }
4199
4200 static int trace__replay(struct trace *trace)
4201 {
4202         const struct evsel_str_handler handlers[] = {
4203                 { "probe:vfs_getname",       trace__vfs_getname, },
4204         };
4205         struct perf_data data = {
4206                 .path  = input_name,
4207                 .mode  = PERF_DATA_MODE_READ,
4208                 .force = trace->force,
4209         };
4210         struct perf_session *session;
4211         struct evsel *evsel;
4212         int err = -1;
4213
4214         trace->tool.sample        = trace__process_sample;
4215         trace->tool.mmap          = perf_event__process_mmap;
4216         trace->tool.mmap2         = perf_event__process_mmap2;
4217         trace->tool.comm          = perf_event__process_comm;
4218         trace->tool.exit          = perf_event__process_exit;
4219         trace->tool.fork          = perf_event__process_fork;
4220         trace->tool.attr          = perf_event__process_attr;
4221         trace->tool.tracing_data  = perf_event__process_tracing_data;
4222         trace->tool.build_id      = perf_event__process_build_id;
4223         trace->tool.namespaces    = perf_event__process_namespaces;
4224
4225         trace->tool.ordered_events = true;
4226         trace->tool.ordering_requires_timestamps = true;
4227
4228         /* add tid to output */
4229         trace->multiple_threads = true;
4230
4231         session = perf_session__new(&data, false, &trace->tool);
4232         if (IS_ERR(session))
4233                 return PTR_ERR(session);
4234
4235         if (trace->opts.target.pid)
4236                 symbol_conf.pid_list_str = strdup(trace->opts.target.pid);
4237
4238         if (trace->opts.target.tid)
4239                 symbol_conf.tid_list_str = strdup(trace->opts.target.tid);
4240
4241         if (symbol__init(&session->header.env) < 0)
4242                 goto out;
4243
4244         trace->host = &session->machines.host;
4245
4246         err = perf_session__set_tracepoints_handlers(session, handlers);
4247         if (err)
4248                 goto out;
4249
4250         evsel = evlist__find_tracepoint_by_name(session->evlist, "raw_syscalls:sys_enter");
4251         /* older kernels have syscalls tp versus raw_syscalls */
4252         if (evsel == NULL)
4253                 evsel = evlist__find_tracepoint_by_name(session->evlist, "syscalls:sys_enter");
4254
4255         if (evsel &&
4256             (evsel__init_raw_syscall_tp(evsel, trace__sys_enter) < 0 ||
4257             perf_evsel__init_sc_tp_ptr_field(evsel, args))) {
4258                 pr_err("Error during initialize raw_syscalls:sys_enter event\n");
4259                 goto out;
4260         }
4261
4262         evsel = evlist__find_tracepoint_by_name(session->evlist, "raw_syscalls:sys_exit");
4263         if (evsel == NULL)
4264                 evsel = evlist__find_tracepoint_by_name(session->evlist, "syscalls:sys_exit");
4265         if (evsel &&
4266             (evsel__init_raw_syscall_tp(evsel, trace__sys_exit) < 0 ||
4267             perf_evsel__init_sc_tp_uint_field(evsel, ret))) {
4268                 pr_err("Error during initialize raw_syscalls:sys_exit event\n");
4269                 goto out;
4270         }
4271
4272         evlist__for_each_entry(session->evlist, evsel) {
4273                 if (evsel->core.attr.type == PERF_TYPE_SOFTWARE &&
4274                     (evsel->core.attr.config == PERF_COUNT_SW_PAGE_FAULTS_MAJ ||
4275                      evsel->core.attr.config == PERF_COUNT_SW_PAGE_FAULTS_MIN ||
4276                      evsel->core.attr.config == PERF_COUNT_SW_PAGE_FAULTS))
4277                         evsel->handler = trace__pgfault;
4278         }
4279
4280         setup_pager();
4281
4282         err = perf_session__process_events(session);
4283         if (err)
4284                 pr_err("Failed to process events, error %d", err);
4285
4286         else if (trace->summary)
4287                 trace__fprintf_thread_summary(trace, trace->output);
4288
4289 out:
4290         perf_session__delete(session);
4291
4292         return err;
4293 }
4294
4295 static size_t trace__fprintf_threads_header(FILE *fp)
4296 {
4297         size_t printed;
4298
4299         printed  = fprintf(fp, "\n Summary of events:\n\n");
4300
4301         return printed;
4302 }
4303
4304 DEFINE_RESORT_RB(syscall_stats, a->msecs > b->msecs,
4305         struct syscall_stats *stats;
4306         double               msecs;
4307         int                  syscall;
4308 )
4309 {
4310         struct int_node *source = rb_entry(nd, struct int_node, rb_node);
4311         struct syscall_stats *stats = source->priv;
4312
4313         entry->syscall = source->i;
4314         entry->stats   = stats;
4315         entry->msecs   = stats ? (u64)stats->stats.n * (avg_stats(&stats->stats) / NSEC_PER_MSEC) : 0;
4316 }
4317
4318 static size_t thread__dump_stats(struct thread_trace *ttrace,
4319                                  struct trace *trace, FILE *fp)
4320 {
4321         size_t printed = 0;
4322         struct syscall *sc;
4323         struct rb_node *nd;
4324         DECLARE_RESORT_RB_INTLIST(syscall_stats, ttrace->syscall_stats);
4325
4326         if (syscall_stats == NULL)
4327                 return 0;
4328
4329         printed += fprintf(fp, "\n");
4330
4331         printed += fprintf(fp, "   syscall            calls  errors  total       min       avg       max       stddev\n");
4332         printed += fprintf(fp, "                                     (msec)    (msec)    (msec)    (msec)        (%%)\n");
4333         printed += fprintf(fp, "   --------------- --------  ------ -------- --------- --------- ---------     ------\n");
4334
4335         resort_rb__for_each_entry(nd, syscall_stats) {
4336                 struct syscall_stats *stats = syscall_stats_entry->stats;
4337                 if (stats) {
4338                         double min = (double)(stats->stats.min) / NSEC_PER_MSEC;
4339                         double max = (double)(stats->stats.max) / NSEC_PER_MSEC;
4340                         double avg = avg_stats(&stats->stats);
4341                         double pct;
4342                         u64 n = (u64)stats->stats.n;
4343
4344                         pct = avg ? 100.0 * stddev_stats(&stats->stats) / avg : 0.0;
4345                         avg /= NSEC_PER_MSEC;
4346
4347                         sc = &trace->syscalls.table[syscall_stats_entry->syscall];
4348                         printed += fprintf(fp, "   %-15s", sc->name);
4349                         printed += fprintf(fp, " %8" PRIu64 " %6" PRIu64 " %9.3f %9.3f %9.3f",
4350                                            n, stats->nr_failures, syscall_stats_entry->msecs, min, avg);
4351                         printed += fprintf(fp, " %9.3f %9.2f%%\n", max, pct);
4352
4353                         if (trace->errno_summary && stats->nr_failures) {
4354                                 const char *arch_name = perf_env__arch(trace->host->env);
4355                                 int e;
4356
4357                                 for (e = 0; e < stats->max_errno; ++e) {
4358                                         if (stats->errnos[e] != 0)
4359                                                 fprintf(fp, "\t\t\t\t%s: %d\n", arch_syscalls__strerrno(arch_name, e + 1), stats->errnos[e]);
4360                                 }
4361                         }
4362                 }
4363         }
4364
4365         resort_rb__delete(syscall_stats);
4366         printed += fprintf(fp, "\n\n");
4367
4368         return printed;
4369 }
4370
4371 static size_t trace__fprintf_thread(FILE *fp, struct thread *thread, struct trace *trace)
4372 {
4373         size_t printed = 0;
4374         struct thread_trace *ttrace = thread__priv(thread);
4375         double ratio;
4376
4377         if (ttrace == NULL)
4378                 return 0;
4379
4380         ratio = (double)ttrace->nr_events / trace->nr_events * 100.0;
4381
4382         printed += fprintf(fp, " %s (%d), ", thread__comm_str(thread), thread->tid);
4383         printed += fprintf(fp, "%lu events, ", ttrace->nr_events);
4384         printed += fprintf(fp, "%.1f%%", ratio);
4385         if (ttrace->pfmaj)
4386                 printed += fprintf(fp, ", %lu majfaults", ttrace->pfmaj);
4387         if (ttrace->pfmin)
4388                 printed += fprintf(fp, ", %lu minfaults", ttrace->pfmin);
4389         if (trace->sched)
4390                 printed += fprintf(fp, ", %.3f msec\n", ttrace->runtime_ms);
4391         else if (fputc('\n', fp) != EOF)
4392                 ++printed;
4393
4394         printed += thread__dump_stats(ttrace, trace, fp);
4395
4396         return printed;
4397 }
4398
4399 static unsigned long thread__nr_events(struct thread_trace *ttrace)
4400 {
4401         return ttrace ? ttrace->nr_events : 0;
4402 }
4403
4404 DEFINE_RESORT_RB(threads, (thread__nr_events(a->thread->priv) < thread__nr_events(b->thread->priv)),
4405         struct thread *thread;
4406 )
4407 {
4408         entry->thread = rb_entry(nd, struct thread, rb_node);
4409 }
4410
4411 static size_t trace__fprintf_thread_summary(struct trace *trace, FILE *fp)
4412 {
4413         size_t printed = trace__fprintf_threads_header(fp);
4414         struct rb_node *nd;
4415         int i;
4416
4417         for (i = 0; i < THREADS__TABLE_SIZE; i++) {
4418                 DECLARE_RESORT_RB_MACHINE_THREADS(threads, trace->host, i);
4419
4420                 if (threads == NULL) {
4421                         fprintf(fp, "%s", "Error sorting output by nr_events!\n");
4422                         return 0;
4423                 }
4424
4425                 resort_rb__for_each_entry(nd, threads)
4426                         printed += trace__fprintf_thread(fp, threads_entry->thread, trace);
4427
4428                 resort_rb__delete(threads);
4429         }
4430         return printed;
4431 }
4432
4433 static int trace__set_duration(const struct option *opt, const char *str,
4434                                int unset __maybe_unused)
4435 {
4436         struct trace *trace = opt->value;
4437
4438         trace->duration_filter = atof(str);
4439         return 0;
4440 }
4441
4442 static int trace__set_filter_pids_from_option(const struct option *opt, const char *str,
4443                                               int unset __maybe_unused)
4444 {
4445         int ret = -1;
4446         size_t i;
4447         struct trace *trace = opt->value;
4448         /*
4449          * FIXME: introduce a intarray class, plain parse csv and create a
4450          * { int nr, int entries[] } struct...
4451          */
4452         struct intlist *list = intlist__new(str);
4453
4454         if (list == NULL)
4455                 return -1;
4456
4457         i = trace->filter_pids.nr = intlist__nr_entries(list) + 1;
4458         trace->filter_pids.entries = calloc(i, sizeof(pid_t));
4459
4460         if (trace->filter_pids.entries == NULL)
4461                 goto out;
4462
4463         trace->filter_pids.entries[0] = getpid();
4464
4465         for (i = 1; i < trace->filter_pids.nr; ++i)
4466                 trace->filter_pids.entries[i] = intlist__entry(list, i - 1)->i;
4467
4468         intlist__delete(list);
4469         ret = 0;
4470 out:
4471         return ret;
4472 }
4473
4474 static int trace__open_output(struct trace *trace, const char *filename)
4475 {
4476         struct stat st;
4477
4478         if (!stat(filename, &st) && st.st_size) {
4479                 char oldname[PATH_MAX];
4480
4481                 scnprintf(oldname, sizeof(oldname), "%s.old", filename);
4482                 unlink(oldname);
4483                 rename(filename, oldname);
4484         }
4485
4486         trace->output = fopen(filename, "w");
4487
4488         return trace->output == NULL ? -errno : 0;
4489 }
4490
4491 static int parse_pagefaults(const struct option *opt, const char *str,
4492                             int unset __maybe_unused)
4493 {
4494         int *trace_pgfaults = opt->value;
4495
4496         if (strcmp(str, "all") == 0)
4497                 *trace_pgfaults |= TRACE_PFMAJ | TRACE_PFMIN;
4498         else if (strcmp(str, "maj") == 0)
4499                 *trace_pgfaults |= TRACE_PFMAJ;
4500         else if (strcmp(str, "min") == 0)
4501                 *trace_pgfaults |= TRACE_PFMIN;
4502         else
4503                 return -1;
4504
4505         return 0;
4506 }
4507
4508 static void evlist__set_default_evsel_handler(struct evlist *evlist, void *handler)
4509 {
4510         struct evsel *evsel;
4511
4512         evlist__for_each_entry(evlist, evsel) {
4513                 if (evsel->handler == NULL)
4514                         evsel->handler = handler;
4515         }
4516 }
4517
4518 static void evsel__set_syscall_arg_fmt(struct evsel *evsel, const char *name)
4519 {
4520         struct syscall_arg_fmt *fmt = evsel__syscall_arg_fmt(evsel);
4521
4522         if (fmt) {
4523                 struct syscall_fmt *scfmt = syscall_fmt__find(name);
4524
4525                 if (scfmt) {
4526                         int skip = 0;
4527
4528                         if (strcmp(evsel->tp_format->format.fields->name, "__syscall_nr") == 0 ||
4529                             strcmp(evsel->tp_format->format.fields->name, "nr") == 0)
4530                                 ++skip;
4531
4532                         memcpy(fmt + skip, scfmt->arg, (evsel->tp_format->format.nr_fields - skip) * sizeof(*fmt));
4533                 }
4534         }
4535 }
4536
4537 static int evlist__set_syscall_tp_fields(struct evlist *evlist)
4538 {
4539         struct evsel *evsel;
4540
4541         evlist__for_each_entry(evlist, evsel) {
4542                 if (evsel->priv || !evsel->tp_format)
4543                         continue;
4544
4545                 if (strcmp(evsel->tp_format->system, "syscalls")) {
4546                         evsel__init_tp_arg_scnprintf(evsel);
4547                         continue;
4548                 }
4549
4550                 if (evsel__init_syscall_tp(evsel))
4551                         return -1;
4552
4553                 if (!strncmp(evsel->tp_format->name, "sys_enter_", 10)) {
4554                         struct syscall_tp *sc = __evsel__syscall_tp(evsel);
4555
4556                         if (__tp_field__init_ptr(&sc->args, sc->id.offset + sizeof(u64)))
4557                                 return -1;
4558
4559                         evsel__set_syscall_arg_fmt(evsel, evsel->tp_format->name + sizeof("sys_enter_") - 1);
4560                 } else if (!strncmp(evsel->tp_format->name, "sys_exit_", 9)) {
4561                         struct syscall_tp *sc = __evsel__syscall_tp(evsel);
4562
4563                         if (__tp_field__init_uint(&sc->ret, sizeof(u64), sc->id.offset + sizeof(u64), evsel->needs_swap))
4564                                 return -1;
4565
4566                         evsel__set_syscall_arg_fmt(evsel, evsel->tp_format->name + sizeof("sys_exit_") - 1);
4567                 }
4568         }
4569
4570         return 0;
4571 }
4572
4573 /*
4574  * XXX: Hackish, just splitting the combined -e+--event (syscalls
4575  * (raw_syscalls:{sys_{enter,exit}} + events (tracepoints, HW, SW, etc) to use
4576  * existing facilities unchanged (trace->ev_qualifier + parse_options()).
4577  *
4578  * It'd be better to introduce a parse_options() variant that would return a
4579  * list with the terms it didn't match to an event...
4580  */
4581 static int trace__parse_events_option(const struct option *opt, const char *str,
4582                                       int unset __maybe_unused)
4583 {
4584         struct trace *trace = (struct trace *)opt->value;
4585         const char *s = str;
4586         char *sep = NULL, *lists[2] = { NULL, NULL, };
4587         int len = strlen(str) + 1, err = -1, list, idx;
4588         char *strace_groups_dir = system_path(STRACE_GROUPS_DIR);
4589         char group_name[PATH_MAX];
4590         struct syscall_fmt *fmt;
4591
4592         if (strace_groups_dir == NULL)
4593                 return -1;
4594
4595         if (*s == '!') {
4596                 ++s;
4597                 trace->not_ev_qualifier = true;
4598         }
4599
4600         while (1) {
4601                 if ((sep = strchr(s, ',')) != NULL)
4602                         *sep = '\0';
4603
4604                 list = 0;
4605                 if (syscalltbl__id(trace->sctbl, s) >= 0 ||
4606                     syscalltbl__strglobmatch_first(trace->sctbl, s, &idx) >= 0) {
4607                         list = 1;
4608                         goto do_concat;
4609                 }
4610
4611                 fmt = syscall_fmt__find_by_alias(s);
4612                 if (fmt != NULL) {
4613                         list = 1;
4614                         s = fmt->name;
4615                 } else {
4616                         path__join(group_name, sizeof(group_name), strace_groups_dir, s);
4617                         if (access(group_name, R_OK) == 0)
4618                                 list = 1;
4619                 }
4620 do_concat:
4621                 if (lists[list]) {
4622                         sprintf(lists[list] + strlen(lists[list]), ",%s", s);
4623                 } else {
4624                         lists[list] = malloc(len);
4625                         if (lists[list] == NULL)
4626                                 goto out;
4627                         strcpy(lists[list], s);
4628                 }
4629
4630                 if (!sep)
4631                         break;
4632
4633                 *sep = ',';
4634                 s = sep + 1;
4635         }
4636
4637         if (lists[1] != NULL) {
4638                 struct strlist_config slist_config = {
4639                         .dirname = strace_groups_dir,
4640                 };
4641
4642                 trace->ev_qualifier = strlist__new(lists[1], &slist_config);
4643                 if (trace->ev_qualifier == NULL) {
4644                         fputs("Not enough memory to parse event qualifier", trace->output);
4645                         goto out;
4646                 }
4647
4648                 if (trace__validate_ev_qualifier(trace))
4649                         goto out;
4650                 trace->trace_syscalls = true;
4651         }
4652
4653         err = 0;
4654
4655         if (lists[0]) {
4656                 struct option o = {
4657                         .value = &trace->evlist,
4658                 };
4659                 err = parse_events_option(&o, lists[0], 0);
4660         }
4661 out:
4662         if (sep)
4663                 *sep = ',';
4664
4665         return err;
4666 }
4667
4668 static int trace__parse_cgroups(const struct option *opt, const char *str, int unset)
4669 {
4670         struct trace *trace = opt->value;
4671
4672         if (!list_empty(&trace->evlist->core.entries)) {
4673                 struct option o = {
4674                         .value = &trace->evlist,
4675                 };
4676                 return parse_cgroups(&o, str, unset);
4677         }
4678         trace->cgroup = evlist__findnew_cgroup(trace->evlist, str);
4679
4680         return 0;
4681 }
4682
4683 static int trace__config(const char *var, const char *value, void *arg)
4684 {
4685         struct trace *trace = arg;
4686         int err = 0;
4687
4688         if (!strcmp(var, "trace.add_events")) {
4689                 trace->perfconfig_events = strdup(value);
4690                 if (trace->perfconfig_events == NULL) {
4691                         pr_err("Not enough memory for %s\n", "trace.add_events");
4692                         return -1;
4693                 }
4694         } else if (!strcmp(var, "trace.show_timestamp")) {
4695                 trace->show_tstamp = perf_config_bool(var, value);
4696         } else if (!strcmp(var, "trace.show_duration")) {
4697                 trace->show_duration = perf_config_bool(var, value);
4698         } else if (!strcmp(var, "trace.show_arg_names")) {
4699                 trace->show_arg_names = perf_config_bool(var, value);
4700                 if (!trace->show_arg_names)
4701                         trace->show_zeros = true;
4702         } else if (!strcmp(var, "trace.show_zeros")) {
4703                 bool new_show_zeros = perf_config_bool(var, value);
4704                 if (!trace->show_arg_names && !new_show_zeros) {
4705                         pr_warning("trace.show_zeros has to be set when trace.show_arg_names=no\n");
4706                         goto out;
4707                 }
4708                 trace->show_zeros = new_show_zeros;
4709         } else if (!strcmp(var, "trace.show_prefix")) {
4710                 trace->show_string_prefix = perf_config_bool(var, value);
4711         } else if (!strcmp(var, "trace.no_inherit")) {
4712                 trace->opts.no_inherit = perf_config_bool(var, value);
4713         } else if (!strcmp(var, "trace.args_alignment")) {
4714                 int args_alignment = 0;
4715                 if (perf_config_int(&args_alignment, var, value) == 0)
4716                         trace->args_alignment = args_alignment;
4717         } else if (!strcmp(var, "trace.tracepoint_beautifiers")) {
4718                 if (strcasecmp(value, "libtraceevent") == 0)
4719                         trace->libtraceevent_print = true;
4720                 else if (strcasecmp(value, "libbeauty") == 0)
4721                         trace->libtraceevent_print = false;
4722         }
4723 out:
4724         return err;
4725 }
4726
4727 static void trace__exit(struct trace *trace)
4728 {
4729         int i;
4730
4731         strlist__delete(trace->ev_qualifier);
4732         free(trace->ev_qualifier_ids.entries);
4733         if (trace->syscalls.table) {
4734                 for (i = 0; i <= trace->sctbl->syscalls.max_id; i++)
4735                         syscall__exit(&trace->syscalls.table[i]);
4736                 free(trace->syscalls.table);
4737         }
4738         syscalltbl__delete(trace->sctbl);
4739         zfree(&trace->perfconfig_events);
4740 }
4741
4742 int cmd_trace(int argc, const char **argv)
4743 {
4744         const char *trace_usage[] = {
4745                 "perf trace [<options>] [<command>]",
4746                 "perf trace [<options>] -- <command> [<options>]",
4747                 "perf trace record [<options>] [<command>]",
4748                 "perf trace record [<options>] -- <command> [<options>]",
4749                 NULL
4750         };
4751         struct trace trace = {
4752                 .opts = {
4753                         .target = {
4754                                 .uid       = UINT_MAX,
4755                                 .uses_mmap = true,
4756                         },
4757                         .user_freq     = UINT_MAX,
4758                         .user_interval = ULLONG_MAX,
4759                         .no_buffering  = true,
4760                         .mmap_pages    = UINT_MAX,
4761                 },
4762                 .output = stderr,
4763                 .show_comm = true,
4764                 .show_tstamp = true,
4765                 .show_duration = true,
4766                 .show_arg_names = true,
4767                 .args_alignment = 70,
4768                 .trace_syscalls = false,
4769                 .kernel_syscallchains = false,
4770                 .max_stack = UINT_MAX,
4771                 .max_events = ULONG_MAX,
4772         };
4773         const char *map_dump_str = NULL;
4774         const char *output_name = NULL;
4775         const struct option trace_options[] = {
4776         OPT_CALLBACK('e', "event", &trace, "event",
4777                      "event/syscall selector. use 'perf list' to list available events",
4778                      trace__parse_events_option),
4779         OPT_CALLBACK(0, "filter", &trace.evlist, "filter",
4780                      "event filter", parse_filter),
4781         OPT_BOOLEAN(0, "comm", &trace.show_comm,
4782                     "show the thread COMM next to its id"),
4783         OPT_BOOLEAN(0, "tool_stats", &trace.show_tool_stats, "show tool stats"),
4784         OPT_CALLBACK(0, "expr", &trace, "expr", "list of syscalls/events to trace",
4785                      trace__parse_events_option),
4786         OPT_STRING('o', "output", &output_name, "file", "output file name"),
4787         OPT_STRING('i', "input", &input_name, "file", "Analyze events in file"),
4788         OPT_STRING('p', "pid", &trace.opts.target.pid, "pid",
4789                     "trace events on existing process id"),
4790         OPT_STRING('t', "tid", &trace.opts.target.tid, "tid",
4791                     "trace events on existing thread id"),
4792         OPT_CALLBACK(0, "filter-pids", &trace, "CSV list of pids",
4793                      "pids to filter (by the kernel)", trace__set_filter_pids_from_option),
4794         OPT_BOOLEAN('a', "all-cpus", &trace.opts.target.system_wide,
4795                     "system-wide collection from all CPUs"),
4796         OPT_STRING('C', "cpu", &trace.opts.target.cpu_list, "cpu",
4797                     "list of cpus to monitor"),
4798         OPT_BOOLEAN(0, "no-inherit", &trace.opts.no_inherit,
4799                     "child tasks do not inherit counters"),
4800         OPT_CALLBACK('m', "mmap-pages", &trace.opts.mmap_pages, "pages",
4801                      "number of mmap data pages", evlist__parse_mmap_pages),
4802         OPT_STRING('u', "uid", &trace.opts.target.uid_str, "user",
4803                    "user to profile"),
4804         OPT_CALLBACK(0, "duration", &trace, "float",
4805                      "show only events with duration > N.M ms",
4806                      trace__set_duration),
4807 #ifdef HAVE_LIBBPF_SUPPORT
4808         OPT_STRING(0, "map-dump", &map_dump_str, "BPF map", "BPF map to periodically dump"),
4809 #endif
4810         OPT_BOOLEAN(0, "sched", &trace.sched, "show blocking scheduler events"),
4811         OPT_INCR('v', "verbose", &verbose, "be more verbose"),
4812         OPT_BOOLEAN('T', "time", &trace.full_time,
4813                     "Show full timestamp, not time relative to first start"),
4814         OPT_BOOLEAN(0, "failure", &trace.failure_only,
4815                     "Show only syscalls that failed"),
4816         OPT_BOOLEAN('s', "summary", &trace.summary_only,
4817                     "Show only syscall summary with statistics"),
4818         OPT_BOOLEAN('S', "with-summary", &trace.summary,
4819                     "Show all syscalls and summary with statistics"),
4820         OPT_BOOLEAN(0, "errno-summary", &trace.errno_summary,
4821                     "Show errno stats per syscall, use with -s or -S"),
4822         OPT_CALLBACK_DEFAULT('F', "pf", &trace.trace_pgfaults, "all|maj|min",
4823                      "Trace pagefaults", parse_pagefaults, "maj"),
4824         OPT_BOOLEAN(0, "syscalls", &trace.trace_syscalls, "Trace syscalls"),
4825         OPT_BOOLEAN('f', "force", &trace.force, "don't complain, do it"),
4826         OPT_CALLBACK(0, "call-graph", &trace.opts,
4827                      "record_mode[,record_size]", record_callchain_help,
4828                      &record_parse_callchain_opt),
4829         OPT_BOOLEAN(0, "libtraceevent_print", &trace.libtraceevent_print,
4830                     "Use libtraceevent to print the tracepoint arguments."),
4831         OPT_BOOLEAN(0, "kernel-syscall-graph", &trace.kernel_syscallchains,
4832                     "Show the kernel callchains on the syscall exit path"),
4833         OPT_ULONG(0, "max-events", &trace.max_events,
4834                 "Set the maximum number of events to print, exit after that is reached. "),
4835         OPT_UINTEGER(0, "min-stack", &trace.min_stack,
4836                      "Set the minimum stack depth when parsing the callchain, "
4837                      "anything below the specified depth will be ignored."),
4838         OPT_UINTEGER(0, "max-stack", &trace.max_stack,
4839                      "Set the maximum stack depth when parsing the callchain, "
4840                      "anything beyond the specified depth will be ignored. "
4841                      "Default: kernel.perf_event_max_stack or " __stringify(PERF_MAX_STACK_DEPTH)),
4842         OPT_BOOLEAN(0, "sort-events", &trace.sort_events,
4843                         "Sort batch of events before processing, use if getting out of order events"),
4844         OPT_BOOLEAN(0, "print-sample", &trace.print_sample,
4845                         "print the PERF_RECORD_SAMPLE PERF_SAMPLE_ info, for debugging"),
4846         OPT_UINTEGER(0, "proc-map-timeout", &proc_map_timeout,
4847                         "per thread proc mmap processing timeout in ms"),
4848         OPT_CALLBACK('G', "cgroup", &trace, "name", "monitor event in cgroup name only",
4849                      trace__parse_cgroups),
4850         OPT_INTEGER('D', "delay", &trace.opts.initial_delay,
4851                      "ms to wait before starting measurement after program "
4852                      "start"),
4853         OPTS_EVSWITCH(&trace.evswitch),
4854         OPT_END()
4855         };
4856         bool __maybe_unused max_stack_user_set = true;
4857         bool mmap_pages_user_set = true;
4858         struct evsel *evsel;
4859         const char * const trace_subcommands[] = { "record", NULL };
4860         int err = -1;
4861         char bf[BUFSIZ];
4862
4863         signal(SIGSEGV, sighandler_dump_stack);
4864         signal(SIGFPE, sighandler_dump_stack);
4865         signal(SIGCHLD, sig_handler);
4866         signal(SIGINT, sig_handler);
4867
4868         trace.evlist = evlist__new();
4869         trace.sctbl = syscalltbl__new();
4870
4871         if (trace.evlist == NULL || trace.sctbl == NULL) {
4872                 pr_err("Not enough memory to run!\n");
4873                 err = -ENOMEM;
4874                 goto out;
4875         }
4876
4877         /*
4878          * Parsing .perfconfig may entail creating a BPF event, that may need
4879          * to create BPF maps, so bump RLIM_MEMLOCK as the default 64K setting
4880          * is too small. This affects just this process, not touching the
4881          * global setting. If it fails we'll get something in 'perf trace -v'
4882          * to help diagnose the problem.
4883          */
4884         rlimit__bump_memlock();
4885
4886         err = perf_config(trace__config, &trace);
4887         if (err)
4888                 goto out;
4889
4890         argc = parse_options_subcommand(argc, argv, trace_options, trace_subcommands,
4891                                  trace_usage, PARSE_OPT_STOP_AT_NON_OPTION);
4892
4893         /*
4894          * Here we already passed thru trace__parse_events_option() and it has
4895          * already figured out if -e syscall_name, if not but if --event
4896          * foo:bar was used, the user is interested _just_ in those, say,
4897          * tracepoint events, not in the strace-like syscall-name-based mode.
4898          *
4899          * This is important because we need to check if strace-like mode is
4900          * needed to decided if we should filter out the eBPF
4901          * __augmented_syscalls__ code, if it is in the mix, say, via
4902          * .perfconfig trace.add_events, and filter those out.
4903          */
4904         if (!trace.trace_syscalls && !trace.trace_pgfaults &&
4905             trace.evlist->core.nr_entries == 0 /* Was --events used? */) {
4906                 trace.trace_syscalls = true;
4907         }
4908         /*
4909          * Now that we have --verbose figured out, lets see if we need to parse
4910          * events from .perfconfig, so that if those events fail parsing, say some
4911          * BPF program fails, then we'll be able to use --verbose to see what went
4912          * wrong in more detail.
4913          */
4914         if (trace.perfconfig_events != NULL) {
4915                 struct parse_events_error parse_err;
4916
4917                 bzero(&parse_err, sizeof(parse_err));
4918                 err = parse_events(trace.evlist, trace.perfconfig_events, &parse_err);
4919                 if (err) {
4920                         parse_events_print_error(&parse_err, trace.perfconfig_events);
4921                         goto out;
4922                 }
4923         }
4924
4925         if ((nr_cgroups || trace.cgroup) && !trace.opts.target.system_wide) {
4926                 usage_with_options_msg(trace_usage, trace_options,
4927                                        "cgroup monitoring only available in system-wide mode");
4928         }
4929
4930         evsel = bpf__setup_output_event(trace.evlist, "__augmented_syscalls__");
4931         if (IS_ERR(evsel)) {
4932                 bpf__strerror_setup_output_event(trace.evlist, PTR_ERR(evsel), bf, sizeof(bf));
4933                 pr_err("ERROR: Setup trace syscalls enter failed: %s\n", bf);
4934                 goto out;
4935         }
4936
4937         if (evsel) {
4938                 trace.syscalls.events.augmented = evsel;
4939
4940                 evsel = evlist__find_tracepoint_by_name(trace.evlist, "raw_syscalls:sys_enter");
4941                 if (evsel == NULL) {
4942                         pr_err("ERROR: raw_syscalls:sys_enter not found in the augmented BPF object\n");
4943                         goto out;
4944                 }
4945
4946                 if (evsel->bpf_obj == NULL) {
4947                         pr_err("ERROR: raw_syscalls:sys_enter not associated to a BPF object\n");
4948                         goto out;
4949                 }
4950
4951                 trace.bpf_obj = evsel->bpf_obj;
4952
4953                 /*
4954                  * If we have _just_ the augmenter event but don't have a
4955                  * explicit --syscalls, then assume we want all strace-like
4956                  * syscalls:
4957                  */
4958                 if (!trace.trace_syscalls && trace__only_augmented_syscalls_evsels(&trace))
4959                         trace.trace_syscalls = true;
4960                 /*
4961                  * So, if we have a syscall augmenter, but trace_syscalls, aka
4962                  * strace-like syscall tracing is not set, then we need to trow
4963                  * away the augmenter, i.e. all the events that were created
4964                  * from that BPF object file.
4965                  *
4966                  * This is more to fix the current .perfconfig trace.add_events
4967                  * style of setting up the strace-like eBPF based syscall point
4968                  * payload augmenter.
4969                  *
4970                  * All this complexity will be avoided by adding an alternative
4971                  * to trace.add_events in the form of
4972                  * trace.bpf_augmented_syscalls, that will be only parsed if we
4973                  * need it.
4974                  *
4975                  * .perfconfig trace.add_events is still useful if we want, for
4976                  * instance, have msr_write.msr in some .perfconfig profile based
4977                  * 'perf trace --config determinism.profile' mode, where for some
4978                  * particular goal/workload type we want a set of events and
4979                  * output mode (with timings, etc) instead of having to add
4980                  * all via the command line.
4981                  *
4982                  * Also --config to specify an alternate .perfconfig file needs
4983                  * to be implemented.
4984                  */
4985                 if (!trace.trace_syscalls) {
4986                         trace__delete_augmented_syscalls(&trace);
4987                 } else {
4988                         trace__set_bpf_map_filtered_pids(&trace);
4989                         trace__set_bpf_map_syscalls(&trace);
4990                         trace.syscalls.unaugmented_prog = trace__find_bpf_program_by_title(&trace, "!raw_syscalls:unaugmented");
4991                 }
4992         }
4993
4994         err = bpf__setup_stdout(trace.evlist);
4995         if (err) {
4996                 bpf__strerror_setup_stdout(trace.evlist, err, bf, sizeof(bf));
4997                 pr_err("ERROR: Setup BPF stdout failed: %s\n", bf);
4998                 goto out;
4999         }
5000
5001         err = -1;
5002
5003         if (map_dump_str) {
5004                 trace.dump.map = trace__find_bpf_map_by_name(&trace, map_dump_str);
5005                 if (trace.dump.map == NULL) {
5006                         pr_err("ERROR: BPF map \"%s\" not found\n", map_dump_str);
5007                         goto out;
5008                 }
5009         }
5010
5011         if (trace.trace_pgfaults) {
5012                 trace.opts.sample_address = true;
5013                 trace.opts.sample_time = true;
5014         }
5015
5016         if (trace.opts.mmap_pages == UINT_MAX)
5017                 mmap_pages_user_set = false;
5018
5019         if (trace.max_stack == UINT_MAX) {
5020                 trace.max_stack = input_name ? PERF_MAX_STACK_DEPTH : sysctl__max_stack();
5021                 max_stack_user_set = false;
5022         }
5023
5024 #ifdef HAVE_DWARF_UNWIND_SUPPORT
5025         if ((trace.min_stack || max_stack_user_set) && !callchain_param.enabled) {
5026                 record_opts__parse_callchain(&trace.opts, &callchain_param, "dwarf", false);
5027         }
5028 #endif
5029
5030         if (callchain_param.enabled) {
5031                 if (!mmap_pages_user_set && geteuid() == 0)
5032                         trace.opts.mmap_pages = perf_event_mlock_kb_in_pages() * 4;
5033
5034                 symbol_conf.use_callchain = true;
5035         }
5036
5037         if (trace.evlist->core.nr_entries > 0) {
5038                 evlist__set_default_evsel_handler(trace.evlist, trace__event_handler);
5039                 if (evlist__set_syscall_tp_fields(trace.evlist)) {
5040                         perror("failed to set syscalls:* tracepoint fields");
5041                         goto out;
5042                 }
5043         }
5044
5045         if (trace.sort_events) {
5046                 ordered_events__init(&trace.oe.data, ordered_events__deliver_event, &trace);
5047                 ordered_events__set_copy_on_queue(&trace.oe.data, true);
5048         }
5049
5050         /*
5051          * If we are augmenting syscalls, then combine what we put in the
5052          * __augmented_syscalls__ BPF map with what is in the
5053          * syscalls:sys_exit_FOO tracepoints, i.e. just like we do without BPF,
5054          * combining raw_syscalls:sys_enter with raw_syscalls:sys_exit.
5055          *
5056          * We'll switch to look at two BPF maps, one for sys_enter and the
5057          * other for sys_exit when we start augmenting the sys_exit paths with
5058          * buffers that are being copied from kernel to userspace, think 'read'
5059          * syscall.
5060          */
5061         if (trace.syscalls.events.augmented) {
5062                 evlist__for_each_entry(trace.evlist, evsel) {
5063                         bool raw_syscalls_sys_exit = strcmp(evsel__name(evsel), "raw_syscalls:sys_exit") == 0;
5064
5065                         if (raw_syscalls_sys_exit) {
5066                                 trace.raw_augmented_syscalls = true;
5067                                 goto init_augmented_syscall_tp;
5068                         }
5069
5070                         if (trace.syscalls.events.augmented->priv == NULL &&
5071                             strstr(evsel__name(evsel), "syscalls:sys_enter")) {
5072                                 struct evsel *augmented = trace.syscalls.events.augmented;
5073                                 if (evsel__init_augmented_syscall_tp(augmented, evsel) ||
5074                                     evsel__init_augmented_syscall_tp_args(augmented))
5075                                         goto out;
5076                                 /*
5077                                  * Augmented is __augmented_syscalls__ BPF_OUTPUT event
5078                                  * Above we made sure we can get from the payload the tp fields
5079                                  * that we get from syscalls:sys_enter tracefs format file.
5080                                  */
5081                                 augmented->handler = trace__sys_enter;
5082                                 /*
5083                                  * Now we do the same for the *syscalls:sys_enter event so that
5084                                  * if we handle it directly, i.e. if the BPF prog returns 0 so
5085                                  * as not to filter it, then we'll handle it just like we would
5086                                  * for the BPF_OUTPUT one:
5087                                  */
5088                                 if (evsel__init_augmented_syscall_tp(evsel, evsel) ||
5089                                     evsel__init_augmented_syscall_tp_args(evsel))
5090                                         goto out;
5091                                 evsel->handler = trace__sys_enter;
5092                         }
5093
5094                         if (strstarts(evsel__name(evsel), "syscalls:sys_exit_")) {
5095                                 struct syscall_tp *sc;
5096 init_augmented_syscall_tp:
5097                                 if (evsel__init_augmented_syscall_tp(evsel, evsel))
5098                                         goto out;
5099                                 sc = __evsel__syscall_tp(evsel);
5100                                 /*
5101                                  * For now with BPF raw_augmented we hook into
5102                                  * raw_syscalls:sys_enter and there we get all
5103                                  * 6 syscall args plus the tracepoint common
5104                                  * fields and the syscall_nr (another long).
5105                                  * So we check if that is the case and if so
5106                                  * don't look after the sc->args_size but
5107                                  * always after the full raw_syscalls:sys_enter
5108                                  * payload, which is fixed.
5109                                  *
5110                                  * We'll revisit this later to pass
5111                                  * s->args_size to the BPF augmenter (now
5112                                  * tools/perf/examples/bpf/augmented_raw_syscalls.c,
5113                                  * so that it copies only what we need for each
5114                                  * syscall, like what happens when we use
5115                                  * syscalls:sys_enter_NAME, so that we reduce
5116                                  * the kernel/userspace traffic to just what is
5117                                  * needed for each syscall.
5118                                  */
5119                                 if (trace.raw_augmented_syscalls)
5120                                         trace.raw_augmented_syscalls_args_size = (6 + 1) * sizeof(long) + sc->id.offset;
5121                                 evsel__init_augmented_syscall_tp_ret(evsel);
5122                                 evsel->handler = trace__sys_exit;
5123                         }
5124                 }
5125         }
5126
5127         if ((argc >= 1) && (strcmp(argv[0], "record") == 0))
5128                 return trace__record(&trace, argc-1, &argv[1]);
5129
5130         /* Using just --errno-summary will trigger --summary */
5131         if (trace.errno_summary && !trace.summary && !trace.summary_only)
5132                 trace.summary_only = true;
5133
5134         /* summary_only implies summary option, but don't overwrite summary if set */
5135         if (trace.summary_only)
5136                 trace.summary = trace.summary_only;
5137
5138         if (output_name != NULL) {
5139                 err = trace__open_output(&trace, output_name);
5140                 if (err < 0) {
5141                         perror("failed to create output file");
5142                         goto out;
5143                 }
5144         }
5145
5146         err = evswitch__init(&trace.evswitch, trace.evlist, stderr);
5147         if (err)
5148                 goto out_close;
5149
5150         err = target__validate(&trace.opts.target);
5151         if (err) {
5152                 target__strerror(&trace.opts.target, err, bf, sizeof(bf));
5153                 fprintf(trace.output, "%s", bf);
5154                 goto out_close;
5155         }
5156
5157         err = target__parse_uid(&trace.opts.target);
5158         if (err) {
5159                 target__strerror(&trace.opts.target, err, bf, sizeof(bf));
5160                 fprintf(trace.output, "%s", bf);
5161                 goto out_close;
5162         }
5163
5164         if (!argc && target__none(&trace.opts.target))
5165                 trace.opts.target.system_wide = true;
5166
5167         if (input_name)
5168                 err = trace__replay(&trace);
5169         else
5170                 err = trace__run(&trace, argc, argv);
5171
5172 out_close:
5173         if (output_name != NULL)
5174                 fclose(trace.output);
5175 out:
5176         trace__exit(&trace);
5177         return err;
5178 }