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                 memset(table + trace->sctbl->syscalls.max_id, 0, (id - trace->sctbl->syscalls.max_id) * sizeof(*sc));
1766
1767                 trace->syscalls.table         = table;
1768                 trace->sctbl->syscalls.max_id = id;
1769         }
1770 #endif
1771         sc = trace->syscalls.table + id;
1772         if (sc->nonexistent)
1773                 return 0;
1774
1775         if (name == NULL) {
1776                 sc->nonexistent = true;
1777                 return 0;
1778         }
1779
1780         sc->name = name;
1781         sc->fmt  = syscall_fmt__find(sc->name);
1782
1783         snprintf(tp_name, sizeof(tp_name), "sys_enter_%s", sc->name);
1784         sc->tp_format = trace_event__tp_format("syscalls", tp_name);
1785
1786         if (IS_ERR(sc->tp_format) && sc->fmt && sc->fmt->alias) {
1787                 snprintf(tp_name, sizeof(tp_name), "sys_enter_%s", sc->fmt->alias);
1788                 sc->tp_format = trace_event__tp_format("syscalls", tp_name);
1789         }
1790
1791         if (syscall__alloc_arg_fmts(sc, IS_ERR(sc->tp_format) ? 6 : sc->tp_format->format.nr_fields))
1792                 return -ENOMEM;
1793
1794         if (IS_ERR(sc->tp_format))
1795                 return PTR_ERR(sc->tp_format);
1796
1797         sc->args = sc->tp_format->format.fields;
1798         /*
1799          * We need to check and discard the first variable '__syscall_nr'
1800          * or 'nr' that mean the syscall number. It is needless here.
1801          * So drop '__syscall_nr' or 'nr' field but does not exist on older kernels.
1802          */
1803         if (sc->args && (!strcmp(sc->args->name, "__syscall_nr") || !strcmp(sc->args->name, "nr"))) {
1804                 sc->args = sc->args->next;
1805                 --sc->nr_args;
1806         }
1807
1808         sc->is_exit = !strcmp(name, "exit_group") || !strcmp(name, "exit");
1809         sc->is_open = !strcmp(name, "open") || !strcmp(name, "openat");
1810
1811         return syscall__set_arg_fmts(sc);
1812 }
1813
1814 static int evsel__init_tp_arg_scnprintf(struct evsel *evsel)
1815 {
1816         struct syscall_arg_fmt *fmt = evsel__syscall_arg_fmt(evsel);
1817
1818         if (fmt != NULL) {
1819                 syscall_arg_fmt__init_array(fmt, evsel->tp_format->format.fields);
1820                 return 0;
1821         }
1822
1823         return -ENOMEM;
1824 }
1825
1826 static int intcmp(const void *a, const void *b)
1827 {
1828         const int *one = a, *another = b;
1829
1830         return *one - *another;
1831 }
1832
1833 static int trace__validate_ev_qualifier(struct trace *trace)
1834 {
1835         int err = 0;
1836         bool printed_invalid_prefix = false;
1837         struct str_node *pos;
1838         size_t nr_used = 0, nr_allocated = strlist__nr_entries(trace->ev_qualifier);
1839
1840         trace->ev_qualifier_ids.entries = malloc(nr_allocated *
1841                                                  sizeof(trace->ev_qualifier_ids.entries[0]));
1842
1843         if (trace->ev_qualifier_ids.entries == NULL) {
1844                 fputs("Error:\tNot enough memory for allocating events qualifier ids\n",
1845                        trace->output);
1846                 err = -EINVAL;
1847                 goto out;
1848         }
1849
1850         strlist__for_each_entry(pos, trace->ev_qualifier) {
1851                 const char *sc = pos->s;
1852                 int id = syscalltbl__id(trace->sctbl, sc), match_next = -1;
1853
1854                 if (id < 0) {
1855                         id = syscalltbl__strglobmatch_first(trace->sctbl, sc, &match_next);
1856                         if (id >= 0)
1857                                 goto matches;
1858
1859                         if (!printed_invalid_prefix) {
1860                                 pr_debug("Skipping unknown syscalls: ");
1861                                 printed_invalid_prefix = true;
1862                         } else {
1863                                 pr_debug(", ");
1864                         }
1865
1866                         pr_debug("%s", sc);
1867                         continue;
1868                 }
1869 matches:
1870                 trace->ev_qualifier_ids.entries[nr_used++] = id;
1871                 if (match_next == -1)
1872                         continue;
1873
1874                 while (1) {
1875                         id = syscalltbl__strglobmatch_next(trace->sctbl, sc, &match_next);
1876                         if (id < 0)
1877                                 break;
1878                         if (nr_allocated == nr_used) {
1879                                 void *entries;
1880
1881                                 nr_allocated += 8;
1882                                 entries = realloc(trace->ev_qualifier_ids.entries,
1883                                                   nr_allocated * sizeof(trace->ev_qualifier_ids.entries[0]));
1884                                 if (entries == NULL) {
1885                                         err = -ENOMEM;
1886                                         fputs("\nError:\t Not enough memory for parsing\n", trace->output);
1887                                         goto out_free;
1888                                 }
1889                                 trace->ev_qualifier_ids.entries = entries;
1890                         }
1891                         trace->ev_qualifier_ids.entries[nr_used++] = id;
1892                 }
1893         }
1894
1895         trace->ev_qualifier_ids.nr = nr_used;
1896         qsort(trace->ev_qualifier_ids.entries, nr_used, sizeof(int), intcmp);
1897 out:
1898         if (printed_invalid_prefix)
1899                 pr_debug("\n");
1900         return err;
1901 out_free:
1902         zfree(&trace->ev_qualifier_ids.entries);
1903         trace->ev_qualifier_ids.nr = 0;
1904         goto out;
1905 }
1906
1907 static __maybe_unused bool trace__syscall_enabled(struct trace *trace, int id)
1908 {
1909         bool in_ev_qualifier;
1910
1911         if (trace->ev_qualifier_ids.nr == 0)
1912                 return true;
1913
1914         in_ev_qualifier = bsearch(&id, trace->ev_qualifier_ids.entries,
1915                                   trace->ev_qualifier_ids.nr, sizeof(int), intcmp) != NULL;
1916
1917         if (in_ev_qualifier)
1918                return !trace->not_ev_qualifier;
1919
1920         return trace->not_ev_qualifier;
1921 }
1922
1923 /*
1924  * args is to be interpreted as a series of longs but we need to handle
1925  * 8-byte unaligned accesses. args points to raw_data within the event
1926  * and raw_data is guaranteed to be 8-byte unaligned because it is
1927  * preceded by raw_size which is a u32. So we need to copy args to a temp
1928  * variable to read it. Most notably this avoids extended load instructions
1929  * on unaligned addresses
1930  */
1931 unsigned long syscall_arg__val(struct syscall_arg *arg, u8 idx)
1932 {
1933         unsigned long val;
1934         unsigned char *p = arg->args + sizeof(unsigned long) * idx;
1935
1936         memcpy(&val, p, sizeof(val));
1937         return val;
1938 }
1939
1940 static size_t syscall__scnprintf_name(struct syscall *sc, char *bf, size_t size,
1941                                       struct syscall_arg *arg)
1942 {
1943         if (sc->arg_fmt && sc->arg_fmt[arg->idx].name)
1944                 return scnprintf(bf, size, "%s: ", sc->arg_fmt[arg->idx].name);
1945
1946         return scnprintf(bf, size, "arg%d: ", arg->idx);
1947 }
1948
1949 /*
1950  * Check if the value is in fact zero, i.e. mask whatever needs masking, such
1951  * as mount 'flags' argument that needs ignoring some magic flag, see comment
1952  * in tools/perf/trace/beauty/mount_flags.c
1953  */
1954 static unsigned long syscall_arg_fmt__mask_val(struct syscall_arg_fmt *fmt, struct syscall_arg *arg, unsigned long val)
1955 {
1956         if (fmt && fmt->mask_val)
1957                 return fmt->mask_val(arg, val);
1958
1959         return val;
1960 }
1961
1962 static size_t syscall_arg_fmt__scnprintf_val(struct syscall_arg_fmt *fmt, char *bf, size_t size,
1963                                              struct syscall_arg *arg, unsigned long val)
1964 {
1965         if (fmt && fmt->scnprintf) {
1966                 arg->val = val;
1967                 if (fmt->parm)
1968                         arg->parm = fmt->parm;
1969                 return fmt->scnprintf(bf, size, arg);
1970         }
1971         return scnprintf(bf, size, "%ld", val);
1972 }
1973
1974 static size_t syscall__scnprintf_args(struct syscall *sc, char *bf, size_t size,
1975                                       unsigned char *args, void *augmented_args, int augmented_args_size,
1976                                       struct trace *trace, struct thread *thread)
1977 {
1978         size_t printed = 0;
1979         unsigned long val;
1980         u8 bit = 1;
1981         struct syscall_arg arg = {
1982                 .args   = args,
1983                 .augmented = {
1984                         .size = augmented_args_size,
1985                         .args = augmented_args,
1986                 },
1987                 .idx    = 0,
1988                 .mask   = 0,
1989                 .trace  = trace,
1990                 .thread = thread,
1991                 .show_string_prefix = trace->show_string_prefix,
1992         };
1993         struct thread_trace *ttrace = thread__priv(thread);
1994
1995         /*
1996          * Things like fcntl will set this in its 'cmd' formatter to pick the
1997          * right formatter for the return value (an fd? file flags?), which is
1998          * not needed for syscalls that always return a given type, say an fd.
1999          */
2000         ttrace->ret_scnprintf = NULL;
2001
2002         if (sc->args != NULL) {
2003                 struct tep_format_field *field;
2004
2005                 for (field = sc->args; field;
2006                      field = field->next, ++arg.idx, bit <<= 1) {
2007                         if (arg.mask & bit)
2008                                 continue;
2009
2010                         arg.fmt = &sc->arg_fmt[arg.idx];
2011                         val = syscall_arg__val(&arg, arg.idx);
2012                         /*
2013                          * Some syscall args need some mask, most don't and
2014                          * return val untouched.
2015                          */
2016                         val = syscall_arg_fmt__mask_val(&sc->arg_fmt[arg.idx], &arg, val);
2017
2018                         /*
2019                          * Suppress this argument if its value is zero and
2020                          * and we don't have a string associated in an
2021                          * strarray for it.
2022                          */
2023                         if (val == 0 &&
2024                             !trace->show_zeros &&
2025                             !(sc->arg_fmt &&
2026                               (sc->arg_fmt[arg.idx].show_zero ||
2027                                sc->arg_fmt[arg.idx].scnprintf == SCA_STRARRAY ||
2028                                sc->arg_fmt[arg.idx].scnprintf == SCA_STRARRAYS) &&
2029                               sc->arg_fmt[arg.idx].parm))
2030                                 continue;
2031
2032                         printed += scnprintf(bf + printed, size - printed, "%s", printed ? ", " : "");
2033
2034                         if (trace->show_arg_names)
2035                                 printed += scnprintf(bf + printed, size - printed, "%s: ", field->name);
2036
2037                         printed += syscall_arg_fmt__scnprintf_val(&sc->arg_fmt[arg.idx],
2038                                                                   bf + printed, size - printed, &arg, val);
2039                 }
2040         } else if (IS_ERR(sc->tp_format)) {
2041                 /*
2042                  * If we managed to read the tracepoint /format file, then we
2043                  * may end up not having any args, like with gettid(), so only
2044                  * print the raw args when we didn't manage to read it.
2045                  */
2046                 while (arg.idx < sc->nr_args) {
2047                         if (arg.mask & bit)
2048                                 goto next_arg;
2049                         val = syscall_arg__val(&arg, arg.idx);
2050                         if (printed)
2051                                 printed += scnprintf(bf + printed, size - printed, ", ");
2052                         printed += syscall__scnprintf_name(sc, bf + printed, size - printed, &arg);
2053                         printed += syscall_arg_fmt__scnprintf_val(&sc->arg_fmt[arg.idx], bf + printed, size - printed, &arg, val);
2054 next_arg:
2055                         ++arg.idx;
2056                         bit <<= 1;
2057                 }
2058         }
2059
2060         return printed;
2061 }
2062
2063 typedef int (*tracepoint_handler)(struct trace *trace, struct evsel *evsel,
2064                                   union perf_event *event,
2065                                   struct perf_sample *sample);
2066
2067 static struct syscall *trace__syscall_info(struct trace *trace,
2068                                            struct evsel *evsel, int id)
2069 {
2070         int err = 0;
2071
2072         if (id < 0) {
2073
2074                 /*
2075                  * XXX: Noticed on x86_64, reproduced as far back as 3.0.36, haven't tried
2076                  * before that, leaving at a higher verbosity level till that is
2077                  * explained. Reproduced with plain ftrace with:
2078                  *
2079                  * echo 1 > /t/events/raw_syscalls/sys_exit/enable
2080                  * grep "NR -1 " /t/trace_pipe
2081                  *
2082                  * After generating some load on the machine.
2083                  */
2084                 if (verbose > 1) {
2085                         static u64 n;
2086                         fprintf(trace->output, "Invalid syscall %d id, skipping (%s, %" PRIu64 ") ...\n",
2087                                 id, evsel__name(evsel), ++n);
2088                 }
2089                 return NULL;
2090         }
2091
2092         err = -EINVAL;
2093
2094 #ifdef HAVE_SYSCALL_TABLE_SUPPORT
2095         if (id > trace->sctbl->syscalls.max_id) {
2096 #else
2097         if (id >= trace->sctbl->syscalls.max_id) {
2098                 /*
2099                  * With libaudit we don't know beforehand what is the max_id,
2100                  * so we let trace__read_syscall_info() figure that out as we
2101                  * go on reading syscalls.
2102                  */
2103                 err = trace__read_syscall_info(trace, id);
2104                 if (err)
2105 #endif
2106                 goto out_cant_read;
2107         }
2108
2109         if ((trace->syscalls.table == NULL || trace->syscalls.table[id].name == NULL) &&
2110             (err = trace__read_syscall_info(trace, id)) != 0)
2111                 goto out_cant_read;
2112
2113         if (trace->syscalls.table[id].name == NULL) {
2114                 if (trace->syscalls.table[id].nonexistent)
2115                         return NULL;
2116                 goto out_cant_read;
2117         }
2118
2119         return &trace->syscalls.table[id];
2120
2121 out_cant_read:
2122         if (verbose > 0) {
2123                 char sbuf[STRERR_BUFSIZE];
2124                 fprintf(trace->output, "Problems reading syscall %d: %d (%s)", id, -err, str_error_r(-err, sbuf, sizeof(sbuf)));
2125                 if (id <= trace->sctbl->syscalls.max_id && trace->syscalls.table[id].name != NULL)
2126                         fprintf(trace->output, "(%s)", trace->syscalls.table[id].name);
2127                 fputs(" information\n", trace->output);
2128         }
2129         return NULL;
2130 }
2131
2132 struct syscall_stats {
2133         struct stats stats;
2134         u64          nr_failures;
2135         int          max_errno;
2136         u32          *errnos;
2137 };
2138
2139 static void thread__update_stats(struct thread *thread, struct thread_trace *ttrace,
2140                                  int id, struct perf_sample *sample, long err, bool errno_summary)
2141 {
2142         struct int_node *inode;
2143         struct syscall_stats *stats;
2144         u64 duration = 0;
2145
2146         inode = intlist__findnew(ttrace->syscall_stats, id);
2147         if (inode == NULL)
2148                 return;
2149
2150         stats = inode->priv;
2151         if (stats == NULL) {
2152                 stats = malloc(sizeof(*stats));
2153                 if (stats == NULL)
2154                         return;
2155
2156                 stats->nr_failures = 0;
2157                 stats->max_errno   = 0;
2158                 stats->errnos      = NULL;
2159                 init_stats(&stats->stats);
2160                 inode->priv = stats;
2161         }
2162
2163         if (ttrace->entry_time && sample->time > ttrace->entry_time)
2164                 duration = sample->time - ttrace->entry_time;
2165
2166         update_stats(&stats->stats, duration);
2167
2168         if (err < 0) {
2169                 ++stats->nr_failures;
2170
2171                 if (!errno_summary)
2172                         return;
2173
2174                 err = -err;
2175                 if (err > stats->max_errno) {
2176                         u32 *new_errnos = realloc(stats->errnos, err * sizeof(u32));
2177
2178                         if (new_errnos) {
2179                                 memset(new_errnos + stats->max_errno, 0, (err - stats->max_errno) * sizeof(u32));
2180                         } else {
2181                                 pr_debug("Not enough memory for errno stats for thread \"%s\"(%d/%d), results will be incomplete\n",
2182                                          thread__comm_str(thread), thread->pid_, thread->tid);
2183                                 return;
2184                         }
2185
2186                         stats->errnos = new_errnos;
2187                         stats->max_errno = err;
2188                 }
2189
2190                 ++stats->errnos[err - 1];
2191         }
2192 }
2193
2194 static int trace__printf_interrupted_entry(struct trace *trace)
2195 {
2196         struct thread_trace *ttrace;
2197         size_t printed;
2198         int len;
2199
2200         if (trace->failure_only || trace->current == NULL)
2201                 return 0;
2202
2203         ttrace = thread__priv(trace->current);
2204
2205         if (!ttrace->entry_pending)
2206                 return 0;
2207
2208         printed  = trace__fprintf_entry_head(trace, trace->current, 0, false, ttrace->entry_time, trace->output);
2209         printed += len = fprintf(trace->output, "%s)", ttrace->entry_str);
2210
2211         if (len < trace->args_alignment - 4)
2212                 printed += fprintf(trace->output, "%-*s", trace->args_alignment - 4 - len, " ");
2213
2214         printed += fprintf(trace->output, " ...\n");
2215
2216         ttrace->entry_pending = false;
2217         ++trace->nr_events_printed;
2218
2219         return printed;
2220 }
2221
2222 static int trace__fprintf_sample(struct trace *trace, struct evsel *evsel,
2223                                  struct perf_sample *sample, struct thread *thread)
2224 {
2225         int printed = 0;
2226
2227         if (trace->print_sample) {
2228                 double ts = (double)sample->time / NSEC_PER_MSEC;
2229
2230                 printed += fprintf(trace->output, "%22s %10.3f %s %d/%d [%d]\n",
2231                                    evsel__name(evsel), ts,
2232                                    thread__comm_str(thread),
2233                                    sample->pid, sample->tid, sample->cpu);
2234         }
2235
2236         return printed;
2237 }
2238
2239 static void *syscall__augmented_args(struct syscall *sc, struct perf_sample *sample, int *augmented_args_size, int raw_augmented_args_size)
2240 {
2241         void *augmented_args = NULL;
2242         /*
2243          * For now with BPF raw_augmented we hook into raw_syscalls:sys_enter
2244          * and there we get all 6 syscall args plus the tracepoint common fields
2245          * that gets calculated at the start and the syscall_nr (another long).
2246          * So we check if that is the case and if so don't look after the
2247          * sc->args_size but always after the full raw_syscalls:sys_enter payload,
2248          * which is fixed.
2249          *
2250          * We'll revisit this later to pass s->args_size to the BPF augmenter
2251          * (now tools/perf/examples/bpf/augmented_raw_syscalls.c, so that it
2252          * copies only what we need for each syscall, like what happens when we
2253          * use syscalls:sys_enter_NAME, so that we reduce the kernel/userspace
2254          * traffic to just what is needed for each syscall.
2255          */
2256         int args_size = raw_augmented_args_size ?: sc->args_size;
2257
2258         *augmented_args_size = sample->raw_size - args_size;
2259         if (*augmented_args_size > 0)
2260                 augmented_args = sample->raw_data + args_size;
2261
2262         return augmented_args;
2263 }
2264
2265 static int trace__sys_enter(struct trace *trace, struct evsel *evsel,
2266                             union perf_event *event __maybe_unused,
2267                             struct perf_sample *sample)
2268 {
2269         char *msg;
2270         void *args;
2271         int printed = 0;
2272         struct thread *thread;
2273         int id = perf_evsel__sc_tp_uint(evsel, id, sample), err = -1;
2274         int augmented_args_size = 0;
2275         void *augmented_args = NULL;
2276         struct syscall *sc = trace__syscall_info(trace, evsel, id);
2277         struct thread_trace *ttrace;
2278
2279         if (sc == NULL)
2280                 return -1;
2281
2282         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2283         ttrace = thread__trace(thread, trace->output);
2284         if (ttrace == NULL)
2285                 goto out_put;
2286
2287         trace__fprintf_sample(trace, evsel, sample, thread);
2288
2289         args = perf_evsel__sc_tp_ptr(evsel, args, sample);
2290
2291         if (ttrace->entry_str == NULL) {
2292                 ttrace->entry_str = malloc(trace__entry_str_size);
2293                 if (!ttrace->entry_str)
2294                         goto out_put;
2295         }
2296
2297         if (!(trace->duration_filter || trace->summary_only || trace->min_stack))
2298                 trace__printf_interrupted_entry(trace);
2299         /*
2300          * If this is raw_syscalls.sys_enter, then it always comes with the 6 possible
2301          * arguments, even if the syscall being handled, say "openat", uses only 4 arguments
2302          * this breaks syscall__augmented_args() check for augmented args, as we calculate
2303          * syscall->args_size using each syscalls:sys_enter_NAME tracefs format file,
2304          * so when handling, say the openat syscall, we end up getting 6 args for the
2305          * raw_syscalls:sys_enter event, when we expected just 4, we end up mistakenly
2306          * thinking that the extra 2 u64 args are the augmented filename, so just check
2307          * here and avoid using augmented syscalls when the evsel is the raw_syscalls one.
2308          */
2309         if (evsel != trace->syscalls.events.sys_enter)
2310                 augmented_args = syscall__augmented_args(sc, sample, &augmented_args_size, trace->raw_augmented_syscalls_args_size);
2311         ttrace->entry_time = sample->time;
2312         msg = ttrace->entry_str;
2313         printed += scnprintf(msg + printed, trace__entry_str_size - printed, "%s(", sc->name);
2314
2315         printed += syscall__scnprintf_args(sc, msg + printed, trace__entry_str_size - printed,
2316                                            args, augmented_args, augmented_args_size, trace, thread);
2317
2318         if (sc->is_exit) {
2319                 if (!(trace->duration_filter || trace->summary_only || trace->failure_only || trace->min_stack)) {
2320                         int alignment = 0;
2321
2322                         trace__fprintf_entry_head(trace, thread, 0, false, ttrace->entry_time, trace->output);
2323                         printed = fprintf(trace->output, "%s)", ttrace->entry_str);
2324                         if (trace->args_alignment > printed)
2325                                 alignment = trace->args_alignment - printed;
2326                         fprintf(trace->output, "%*s= ?\n", alignment, " ");
2327                 }
2328         } else {
2329                 ttrace->entry_pending = true;
2330                 /* See trace__vfs_getname & trace__sys_exit */
2331                 ttrace->filename.pending_open = false;
2332         }
2333
2334         if (trace->current != thread) {
2335                 thread__put(trace->current);
2336                 trace->current = thread__get(thread);
2337         }
2338         err = 0;
2339 out_put:
2340         thread__put(thread);
2341         return err;
2342 }
2343
2344 static int trace__fprintf_sys_enter(struct trace *trace, struct evsel *evsel,
2345                                     struct perf_sample *sample)
2346 {
2347         struct thread_trace *ttrace;
2348         struct thread *thread;
2349         int id = perf_evsel__sc_tp_uint(evsel, id, sample), err = -1;
2350         struct syscall *sc = trace__syscall_info(trace, evsel, id);
2351         char msg[1024];
2352         void *args, *augmented_args = NULL;
2353         int augmented_args_size;
2354
2355         if (sc == NULL)
2356                 return -1;
2357
2358         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2359         ttrace = thread__trace(thread, trace->output);
2360         /*
2361          * We need to get ttrace just to make sure it is there when syscall__scnprintf_args()
2362          * and the rest of the beautifiers accessing it via struct syscall_arg touches it.
2363          */
2364         if (ttrace == NULL)
2365                 goto out_put;
2366
2367         args = perf_evsel__sc_tp_ptr(evsel, args, sample);
2368         augmented_args = syscall__augmented_args(sc, sample, &augmented_args_size, trace->raw_augmented_syscalls_args_size);
2369         syscall__scnprintf_args(sc, msg, sizeof(msg), args, augmented_args, augmented_args_size, trace, thread);
2370         fprintf(trace->output, "%s", msg);
2371         err = 0;
2372 out_put:
2373         thread__put(thread);
2374         return err;
2375 }
2376
2377 static int trace__resolve_callchain(struct trace *trace, struct evsel *evsel,
2378                                     struct perf_sample *sample,
2379                                     struct callchain_cursor *cursor)
2380 {
2381         struct addr_location al;
2382         int max_stack = evsel->core.attr.sample_max_stack ?
2383                         evsel->core.attr.sample_max_stack :
2384                         trace->max_stack;
2385         int err;
2386
2387         if (machine__resolve(trace->host, &al, sample) < 0)
2388                 return -1;
2389
2390         err = thread__resolve_callchain(al.thread, cursor, evsel, sample, NULL, NULL, max_stack);
2391         addr_location__put(&al);
2392         return err;
2393 }
2394
2395 static int trace__fprintf_callchain(struct trace *trace, struct perf_sample *sample)
2396 {
2397         /* TODO: user-configurable print_opts */
2398         const unsigned int print_opts = EVSEL__PRINT_SYM |
2399                                         EVSEL__PRINT_DSO |
2400                                         EVSEL__PRINT_UNKNOWN_AS_ADDR;
2401
2402         return sample__fprintf_callchain(sample, 38, print_opts, &callchain_cursor, symbol_conf.bt_stop_list, trace->output);
2403 }
2404
2405 static const char *errno_to_name(struct evsel *evsel, int err)
2406 {
2407         struct perf_env *env = evsel__env(evsel);
2408         const char *arch_name = perf_env__arch(env);
2409
2410         return arch_syscalls__strerrno(arch_name, err);
2411 }
2412
2413 static int trace__sys_exit(struct trace *trace, struct evsel *evsel,
2414                            union perf_event *event __maybe_unused,
2415                            struct perf_sample *sample)
2416 {
2417         long ret;
2418         u64 duration = 0;
2419         bool duration_calculated = false;
2420         struct thread *thread;
2421         int id = perf_evsel__sc_tp_uint(evsel, id, sample), err = -1, callchain_ret = 0, printed = 0;
2422         int alignment = trace->args_alignment;
2423         struct syscall *sc = trace__syscall_info(trace, evsel, id);
2424         struct thread_trace *ttrace;
2425
2426         if (sc == NULL)
2427                 return -1;
2428
2429         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2430         ttrace = thread__trace(thread, trace->output);
2431         if (ttrace == NULL)
2432                 goto out_put;
2433
2434         trace__fprintf_sample(trace, evsel, sample, thread);
2435
2436         ret = perf_evsel__sc_tp_uint(evsel, ret, sample);
2437
2438         if (trace->summary)
2439                 thread__update_stats(thread, ttrace, id, sample, ret, trace->errno_summary);
2440
2441         if (!trace->fd_path_disabled && sc->is_open && ret >= 0 && ttrace->filename.pending_open) {
2442                 trace__set_fd_pathname(thread, ret, ttrace->filename.name);
2443                 ttrace->filename.pending_open = false;
2444                 ++trace->stats.vfs_getname;
2445         }
2446
2447         if (ttrace->entry_time) {
2448                 duration = sample->time - ttrace->entry_time;
2449                 if (trace__filter_duration(trace, duration))
2450                         goto out;
2451                 duration_calculated = true;
2452         } else if (trace->duration_filter)
2453                 goto out;
2454
2455         if (sample->callchain) {
2456                 callchain_ret = trace__resolve_callchain(trace, evsel, sample, &callchain_cursor);
2457                 if (callchain_ret == 0) {
2458                         if (callchain_cursor.nr < trace->min_stack)
2459                                 goto out;
2460                         callchain_ret = 1;
2461                 }
2462         }
2463
2464         if (trace->summary_only || (ret >= 0 && trace->failure_only))
2465                 goto out;
2466
2467         trace__fprintf_entry_head(trace, thread, duration, duration_calculated, ttrace->entry_time, trace->output);
2468
2469         if (ttrace->entry_pending) {
2470                 printed = fprintf(trace->output, "%s", ttrace->entry_str);
2471         } else {
2472                 printed += fprintf(trace->output, " ... [");
2473                 color_fprintf(trace->output, PERF_COLOR_YELLOW, "continued");
2474                 printed += 9;
2475                 printed += fprintf(trace->output, "]: %s()", sc->name);
2476         }
2477
2478         printed++; /* the closing ')' */
2479
2480         if (alignment > printed)
2481                 alignment -= printed;
2482         else
2483                 alignment = 0;
2484
2485         fprintf(trace->output, ")%*s= ", alignment, " ");
2486
2487         if (sc->fmt == NULL) {
2488                 if (ret < 0)
2489                         goto errno_print;
2490 signed_print:
2491                 fprintf(trace->output, "%ld", ret);
2492         } else if (ret < 0) {
2493 errno_print: {
2494                 char bf[STRERR_BUFSIZE];
2495                 const char *emsg = str_error_r(-ret, bf, sizeof(bf)),
2496                            *e = errno_to_name(evsel, -ret);
2497
2498                 fprintf(trace->output, "-1 %s (%s)", e, emsg);
2499         }
2500         } else if (ret == 0 && sc->fmt->timeout)
2501                 fprintf(trace->output, "0 (Timeout)");
2502         else if (ttrace->ret_scnprintf) {
2503                 char bf[1024];
2504                 struct syscall_arg arg = {
2505                         .val    = ret,
2506                         .thread = thread,
2507                         .trace  = trace,
2508                 };
2509                 ttrace->ret_scnprintf(bf, sizeof(bf), &arg);
2510                 ttrace->ret_scnprintf = NULL;
2511                 fprintf(trace->output, "%s", bf);
2512         } else if (sc->fmt->hexret)
2513                 fprintf(trace->output, "%#lx", ret);
2514         else if (sc->fmt->errpid) {
2515                 struct thread *child = machine__find_thread(trace->host, ret, ret);
2516
2517                 if (child != NULL) {
2518                         fprintf(trace->output, "%ld", ret);
2519                         if (child->comm_set)
2520                                 fprintf(trace->output, " (%s)", thread__comm_str(child));
2521                         thread__put(child);
2522                 }
2523         } else
2524                 goto signed_print;
2525
2526         fputc('\n', trace->output);
2527
2528         /*
2529          * We only consider an 'event' for the sake of --max-events a non-filtered
2530          * sys_enter + sys_exit and other tracepoint events.
2531          */
2532         if (++trace->nr_events_printed == trace->max_events && trace->max_events != ULONG_MAX)
2533                 interrupted = true;
2534
2535         if (callchain_ret > 0)
2536                 trace__fprintf_callchain(trace, sample);
2537         else if (callchain_ret < 0)
2538                 pr_err("Problem processing %s callchain, skipping...\n", evsel__name(evsel));
2539 out:
2540         ttrace->entry_pending = false;
2541         err = 0;
2542 out_put:
2543         thread__put(thread);
2544         return err;
2545 }
2546
2547 static int trace__vfs_getname(struct trace *trace, struct evsel *evsel,
2548                               union perf_event *event __maybe_unused,
2549                               struct perf_sample *sample)
2550 {
2551         struct thread *thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2552         struct thread_trace *ttrace;
2553         size_t filename_len, entry_str_len, to_move;
2554         ssize_t remaining_space;
2555         char *pos;
2556         const char *filename = evsel__rawptr(evsel, sample, "pathname");
2557
2558         if (!thread)
2559                 goto out;
2560
2561         ttrace = thread__priv(thread);
2562         if (!ttrace)
2563                 goto out_put;
2564
2565         filename_len = strlen(filename);
2566         if (filename_len == 0)
2567                 goto out_put;
2568
2569         if (ttrace->filename.namelen < filename_len) {
2570                 char *f = realloc(ttrace->filename.name, filename_len + 1);
2571
2572                 if (f == NULL)
2573                         goto out_put;
2574
2575                 ttrace->filename.namelen = filename_len;
2576                 ttrace->filename.name = f;
2577         }
2578
2579         strcpy(ttrace->filename.name, filename);
2580         ttrace->filename.pending_open = true;
2581
2582         if (!ttrace->filename.ptr)
2583                 goto out_put;
2584
2585         entry_str_len = strlen(ttrace->entry_str);
2586         remaining_space = trace__entry_str_size - entry_str_len - 1; /* \0 */
2587         if (remaining_space <= 0)
2588                 goto out_put;
2589
2590         if (filename_len > (size_t)remaining_space) {
2591                 filename += filename_len - remaining_space;
2592                 filename_len = remaining_space;
2593         }
2594
2595         to_move = entry_str_len - ttrace->filename.entry_str_pos + 1; /* \0 */
2596         pos = ttrace->entry_str + ttrace->filename.entry_str_pos;
2597         memmove(pos + filename_len, pos, to_move);
2598         memcpy(pos, filename, filename_len);
2599
2600         ttrace->filename.ptr = 0;
2601         ttrace->filename.entry_str_pos = 0;
2602 out_put:
2603         thread__put(thread);
2604 out:
2605         return 0;
2606 }
2607
2608 static int trace__sched_stat_runtime(struct trace *trace, struct evsel *evsel,
2609                                      union perf_event *event __maybe_unused,
2610                                      struct perf_sample *sample)
2611 {
2612         u64 runtime = evsel__intval(evsel, sample, "runtime");
2613         double runtime_ms = (double)runtime / NSEC_PER_MSEC;
2614         struct thread *thread = machine__findnew_thread(trace->host,
2615                                                         sample->pid,
2616                                                         sample->tid);
2617         struct thread_trace *ttrace = thread__trace(thread, trace->output);
2618
2619         if (ttrace == NULL)
2620                 goto out_dump;
2621
2622         ttrace->runtime_ms += runtime_ms;
2623         trace->runtime_ms += runtime_ms;
2624 out_put:
2625         thread__put(thread);
2626         return 0;
2627
2628 out_dump:
2629         fprintf(trace->output, "%s: comm=%s,pid=%u,runtime=%" PRIu64 ",vruntime=%" PRIu64 ")\n",
2630                evsel->name,
2631                evsel__strval(evsel, sample, "comm"),
2632                (pid_t)evsel__intval(evsel, sample, "pid"),
2633                runtime,
2634                evsel__intval(evsel, sample, "vruntime"));
2635         goto out_put;
2636 }
2637
2638 static int bpf_output__printer(enum binary_printer_ops op,
2639                                unsigned int val, void *extra __maybe_unused, FILE *fp)
2640 {
2641         unsigned char ch = (unsigned char)val;
2642
2643         switch (op) {
2644         case BINARY_PRINT_CHAR_DATA:
2645                 return fprintf(fp, "%c", isprint(ch) ? ch : '.');
2646         case BINARY_PRINT_DATA_BEGIN:
2647         case BINARY_PRINT_LINE_BEGIN:
2648         case BINARY_PRINT_ADDR:
2649         case BINARY_PRINT_NUM_DATA:
2650         case BINARY_PRINT_NUM_PAD:
2651         case BINARY_PRINT_SEP:
2652         case BINARY_PRINT_CHAR_PAD:
2653         case BINARY_PRINT_LINE_END:
2654         case BINARY_PRINT_DATA_END:
2655         default:
2656                 break;
2657         }
2658
2659         return 0;
2660 }
2661
2662 static void bpf_output__fprintf(struct trace *trace,
2663                                 struct perf_sample *sample)
2664 {
2665         binary__fprintf(sample->raw_data, sample->raw_size, 8,
2666                         bpf_output__printer, NULL, trace->output);
2667         ++trace->nr_events_printed;
2668 }
2669
2670 static size_t trace__fprintf_tp_fields(struct trace *trace, struct evsel *evsel, struct perf_sample *sample,
2671                                        struct thread *thread, void *augmented_args, int augmented_args_size)
2672 {
2673         char bf[2048];
2674         size_t size = sizeof(bf);
2675         struct tep_format_field *field = evsel->tp_format->format.fields;
2676         struct syscall_arg_fmt *arg = __evsel__syscall_arg_fmt(evsel);
2677         size_t printed = 0;
2678         unsigned long val;
2679         u8 bit = 1;
2680         struct syscall_arg syscall_arg = {
2681                 .augmented = {
2682                         .size = augmented_args_size,
2683                         .args = augmented_args,
2684                 },
2685                 .idx    = 0,
2686                 .mask   = 0,
2687                 .trace  = trace,
2688                 .thread = thread,
2689                 .show_string_prefix = trace->show_string_prefix,
2690         };
2691
2692         for (; field && arg; field = field->next, ++syscall_arg.idx, bit <<= 1, ++arg) {
2693                 if (syscall_arg.mask & bit)
2694                         continue;
2695
2696                 syscall_arg.len = 0;
2697                 syscall_arg.fmt = arg;
2698                 if (field->flags & TEP_FIELD_IS_ARRAY) {
2699                         int offset = field->offset;
2700
2701                         if (field->flags & TEP_FIELD_IS_DYNAMIC) {
2702                                 offset = format_field__intval(field, sample, evsel->needs_swap);
2703                                 syscall_arg.len = offset >> 16;
2704                                 offset &= 0xffff;
2705                         }
2706
2707                         val = (uintptr_t)(sample->raw_data + offset);
2708                 } else
2709                         val = format_field__intval(field, sample, evsel->needs_swap);
2710                 /*
2711                  * Some syscall args need some mask, most don't and
2712                  * return val untouched.
2713                  */
2714                 val = syscall_arg_fmt__mask_val(arg, &syscall_arg, val);
2715
2716                 /*
2717                  * Suppress this argument if its value is zero and
2718                  * and we don't have a string associated in an
2719                  * strarray for it.
2720                  */
2721                 if (val == 0 &&
2722                     !trace->show_zeros &&
2723                     !((arg->show_zero ||
2724                        arg->scnprintf == SCA_STRARRAY ||
2725                        arg->scnprintf == SCA_STRARRAYS) &&
2726                       arg->parm))
2727                         continue;
2728
2729                 printed += scnprintf(bf + printed, size - printed, "%s", printed ? ", " : "");
2730
2731                 /*
2732                  * XXX Perhaps we should have a show_tp_arg_names,
2733                  * leaving show_arg_names just for syscalls?
2734                  */
2735                 if (1 || trace->show_arg_names)
2736                         printed += scnprintf(bf + printed, size - printed, "%s: ", field->name);
2737
2738                 printed += syscall_arg_fmt__scnprintf_val(arg, bf + printed, size - printed, &syscall_arg, val);
2739         }
2740
2741         return printed + fprintf(trace->output, "%s", bf);
2742 }
2743
2744 static int trace__event_handler(struct trace *trace, struct evsel *evsel,
2745                                 union perf_event *event __maybe_unused,
2746                                 struct perf_sample *sample)
2747 {
2748         struct thread *thread;
2749         int callchain_ret = 0;
2750         /*
2751          * Check if we called perf_evsel__disable(evsel) due to, for instance,
2752          * this event's max_events having been hit and this is an entry coming
2753          * from the ring buffer that we should discard, since the max events
2754          * have already been considered/printed.
2755          */
2756         if (evsel->disabled)
2757                 return 0;
2758
2759         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2760
2761         if (sample->callchain) {
2762                 callchain_ret = trace__resolve_callchain(trace, evsel, sample, &callchain_cursor);
2763                 if (callchain_ret == 0) {
2764                         if (callchain_cursor.nr < trace->min_stack)
2765                                 goto out;
2766                         callchain_ret = 1;
2767                 }
2768         }
2769
2770         trace__printf_interrupted_entry(trace);
2771         trace__fprintf_tstamp(trace, sample->time, trace->output);
2772
2773         if (trace->trace_syscalls && trace->show_duration)
2774                 fprintf(trace->output, "(         ): ");
2775
2776         if (thread)
2777                 trace__fprintf_comm_tid(trace, thread, trace->output);
2778
2779         if (evsel == trace->syscalls.events.augmented) {
2780                 int id = perf_evsel__sc_tp_uint(evsel, id, sample);
2781                 struct syscall *sc = trace__syscall_info(trace, evsel, id);
2782
2783                 if (sc) {
2784                         fprintf(trace->output, "%s(", sc->name);
2785                         trace__fprintf_sys_enter(trace, evsel, sample);
2786                         fputc(')', trace->output);
2787                         goto newline;
2788                 }
2789
2790                 /*
2791                  * XXX: Not having the associated syscall info or not finding/adding
2792                  *      the thread should never happen, but if it does...
2793                  *      fall thru and print it as a bpf_output event.
2794                  */
2795         }
2796
2797         fprintf(trace->output, "%s(", evsel->name);
2798
2799         if (evsel__is_bpf_output(evsel)) {
2800                 bpf_output__fprintf(trace, sample);
2801         } else if (evsel->tp_format) {
2802                 if (strncmp(evsel->tp_format->name, "sys_enter_", 10) ||
2803                     trace__fprintf_sys_enter(trace, evsel, sample)) {
2804                         if (trace->libtraceevent_print) {
2805                                 event_format__fprintf(evsel->tp_format, sample->cpu,
2806                                                       sample->raw_data, sample->raw_size,
2807                                                       trace->output);
2808                         } else {
2809                                 trace__fprintf_tp_fields(trace, evsel, sample, thread, NULL, 0);
2810                         }
2811                 }
2812         }
2813
2814 newline:
2815         fprintf(trace->output, ")\n");
2816
2817         if (callchain_ret > 0)
2818                 trace__fprintf_callchain(trace, sample);
2819         else if (callchain_ret < 0)
2820                 pr_err("Problem processing %s callchain, skipping...\n", evsel__name(evsel));
2821
2822         ++trace->nr_events_printed;
2823
2824         if (evsel->max_events != ULONG_MAX && ++evsel->nr_events_printed == evsel->max_events) {
2825                 evsel__disable(evsel);
2826                 evsel__close(evsel);
2827         }
2828 out:
2829         thread__put(thread);
2830         return 0;
2831 }
2832
2833 static void print_location(FILE *f, struct perf_sample *sample,
2834                            struct addr_location *al,
2835                            bool print_dso, bool print_sym)
2836 {
2837
2838         if ((verbose > 0 || print_dso) && al->map)
2839                 fprintf(f, "%s@", al->map->dso->long_name);
2840
2841         if ((verbose > 0 || print_sym) && al->sym)
2842                 fprintf(f, "%s+0x%" PRIx64, al->sym->name,
2843                         al->addr - al->sym->start);
2844         else if (al->map)
2845                 fprintf(f, "0x%" PRIx64, al->addr);
2846         else
2847                 fprintf(f, "0x%" PRIx64, sample->addr);
2848 }
2849
2850 static int trace__pgfault(struct trace *trace,
2851                           struct evsel *evsel,
2852                           union perf_event *event __maybe_unused,
2853                           struct perf_sample *sample)
2854 {
2855         struct thread *thread;
2856         struct addr_location al;
2857         char map_type = 'd';
2858         struct thread_trace *ttrace;
2859         int err = -1;
2860         int callchain_ret = 0;
2861
2862         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2863
2864         if (sample->callchain) {
2865                 callchain_ret = trace__resolve_callchain(trace, evsel, sample, &callchain_cursor);
2866                 if (callchain_ret == 0) {
2867                         if (callchain_cursor.nr < trace->min_stack)
2868                                 goto out_put;
2869                         callchain_ret = 1;
2870                 }
2871         }
2872
2873         ttrace = thread__trace(thread, trace->output);
2874         if (ttrace == NULL)
2875                 goto out_put;
2876
2877         if (evsel->core.attr.config == PERF_COUNT_SW_PAGE_FAULTS_MAJ)
2878                 ttrace->pfmaj++;
2879         else
2880                 ttrace->pfmin++;
2881
2882         if (trace->summary_only)
2883                 goto out;
2884
2885         thread__find_symbol(thread, sample->cpumode, sample->ip, &al);
2886
2887         trace__fprintf_entry_head(trace, thread, 0, true, sample->time, trace->output);
2888
2889         fprintf(trace->output, "%sfault [",
2890                 evsel->core.attr.config == PERF_COUNT_SW_PAGE_FAULTS_MAJ ?
2891                 "maj" : "min");
2892
2893         print_location(trace->output, sample, &al, false, true);
2894
2895         fprintf(trace->output, "] => ");
2896
2897         thread__find_symbol(thread, sample->cpumode, sample->addr, &al);
2898
2899         if (!al.map) {
2900                 thread__find_symbol(thread, sample->cpumode, sample->addr, &al);
2901
2902                 if (al.map)
2903                         map_type = 'x';
2904                 else
2905                         map_type = '?';
2906         }
2907
2908         print_location(trace->output, sample, &al, true, false);
2909
2910         fprintf(trace->output, " (%c%c)\n", map_type, al.level);
2911
2912         if (callchain_ret > 0)
2913                 trace__fprintf_callchain(trace, sample);
2914         else if (callchain_ret < 0)
2915                 pr_err("Problem processing %s callchain, skipping...\n", evsel__name(evsel));
2916
2917         ++trace->nr_events_printed;
2918 out:
2919         err = 0;
2920 out_put:
2921         thread__put(thread);
2922         return err;
2923 }
2924
2925 static void trace__set_base_time(struct trace *trace,
2926                                  struct evsel *evsel,
2927                                  struct perf_sample *sample)
2928 {
2929         /*
2930          * BPF events were not setting PERF_SAMPLE_TIME, so be more robust
2931          * and don't use sample->time unconditionally, we may end up having
2932          * some other event in the future without PERF_SAMPLE_TIME for good
2933          * reason, i.e. we may not be interested in its timestamps, just in
2934          * it taking place, picking some piece of information when it
2935          * appears in our event stream (vfs_getname comes to mind).
2936          */
2937         if (trace->base_time == 0 && !trace->full_time &&
2938             (evsel->core.attr.sample_type & PERF_SAMPLE_TIME))
2939                 trace->base_time = sample->time;
2940 }
2941
2942 static int trace__process_sample(struct perf_tool *tool,
2943                                  union perf_event *event,
2944                                  struct perf_sample *sample,
2945                                  struct evsel *evsel,
2946                                  struct machine *machine __maybe_unused)
2947 {
2948         struct trace *trace = container_of(tool, struct trace, tool);
2949         struct thread *thread;
2950         int err = 0;
2951
2952         tracepoint_handler handler = evsel->handler;
2953
2954         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2955         if (thread && thread__is_filtered(thread))
2956                 goto out;
2957
2958         trace__set_base_time(trace, evsel, sample);
2959
2960         if (handler) {
2961                 ++trace->nr_events;
2962                 handler(trace, evsel, event, sample);
2963         }
2964 out:
2965         thread__put(thread);
2966         return err;
2967 }
2968
2969 static int trace__record(struct trace *trace, int argc, const char **argv)
2970 {
2971         unsigned int rec_argc, i, j;
2972         const char **rec_argv;
2973         const char * const record_args[] = {
2974                 "record",
2975                 "-R",
2976                 "-m", "1024",
2977                 "-c", "1",
2978         };
2979         pid_t pid = getpid();
2980         char *filter = asprintf__tp_filter_pids(1, &pid);
2981         const char * const sc_args[] = { "-e", };
2982         unsigned int sc_args_nr = ARRAY_SIZE(sc_args);
2983         const char * const majpf_args[] = { "-e", "major-faults" };
2984         unsigned int majpf_args_nr = ARRAY_SIZE(majpf_args);
2985         const char * const minpf_args[] = { "-e", "minor-faults" };
2986         unsigned int minpf_args_nr = ARRAY_SIZE(minpf_args);
2987         int err = -1;
2988
2989         /* +3 is for the event string below and the pid filter */
2990         rec_argc = ARRAY_SIZE(record_args) + sc_args_nr + 3 +
2991                 majpf_args_nr + minpf_args_nr + argc;
2992         rec_argv = calloc(rec_argc + 1, sizeof(char *));
2993
2994         if (rec_argv == NULL || filter == NULL)
2995                 goto out_free;
2996
2997         j = 0;
2998         for (i = 0; i < ARRAY_SIZE(record_args); i++)
2999                 rec_argv[j++] = record_args[i];
3000
3001         if (trace->trace_syscalls) {
3002                 for (i = 0; i < sc_args_nr; i++)
3003                         rec_argv[j++] = sc_args[i];
3004
3005                 /* event string may be different for older kernels - e.g., RHEL6 */
3006                 if (is_valid_tracepoint("raw_syscalls:sys_enter"))
3007                         rec_argv[j++] = "raw_syscalls:sys_enter,raw_syscalls:sys_exit";
3008                 else if (is_valid_tracepoint("syscalls:sys_enter"))
3009                         rec_argv[j++] = "syscalls:sys_enter,syscalls:sys_exit";
3010                 else {
3011                         pr_err("Neither raw_syscalls nor syscalls events exist.\n");
3012                         goto out_free;
3013                 }
3014         }
3015
3016         rec_argv[j++] = "--filter";
3017         rec_argv[j++] = filter;
3018
3019         if (trace->trace_pgfaults & TRACE_PFMAJ)
3020                 for (i = 0; i < majpf_args_nr; i++)
3021                         rec_argv[j++] = majpf_args[i];
3022
3023         if (trace->trace_pgfaults & TRACE_PFMIN)
3024                 for (i = 0; i < minpf_args_nr; i++)
3025                         rec_argv[j++] = minpf_args[i];
3026
3027         for (i = 0; i < (unsigned int)argc; i++)
3028                 rec_argv[j++] = argv[i];
3029
3030         err = cmd_record(j, rec_argv);
3031 out_free:
3032         free(filter);
3033         free(rec_argv);
3034         return err;
3035 }
3036
3037 static size_t trace__fprintf_thread_summary(struct trace *trace, FILE *fp);
3038
3039 static bool evlist__add_vfs_getname(struct evlist *evlist)
3040 {
3041         bool found = false;
3042         struct evsel *evsel, *tmp;
3043         struct parse_events_error err;
3044         int ret;
3045
3046         bzero(&err, sizeof(err));
3047         ret = parse_events(evlist, "probe:vfs_getname*", &err);
3048         if (ret) {
3049                 free(err.str);
3050                 free(err.help);
3051                 free(err.first_str);
3052                 free(err.first_help);
3053                 return false;
3054         }
3055
3056         evlist__for_each_entry_safe(evlist, evsel, tmp) {
3057                 if (!strstarts(evsel__name(evsel), "probe:vfs_getname"))
3058                         continue;
3059
3060                 if (evsel__field(evsel, "pathname")) {
3061                         evsel->handler = trace__vfs_getname;
3062                         found = true;
3063                         continue;
3064                 }
3065
3066                 list_del_init(&evsel->core.node);
3067                 evsel->evlist = NULL;
3068                 evsel__delete(evsel);
3069         }
3070
3071         return found;
3072 }
3073
3074 static struct evsel *evsel__new_pgfault(u64 config)
3075 {
3076         struct evsel *evsel;
3077         struct perf_event_attr attr = {
3078                 .type = PERF_TYPE_SOFTWARE,
3079                 .mmap_data = 1,
3080         };
3081
3082         attr.config = config;
3083         attr.sample_period = 1;
3084
3085         event_attr_init(&attr);
3086
3087         evsel = evsel__new(&attr);
3088         if (evsel)
3089                 evsel->handler = trace__pgfault;
3090
3091         return evsel;
3092 }
3093
3094 static void trace__handle_event(struct trace *trace, union perf_event *event, struct perf_sample *sample)
3095 {
3096         const u32 type = event->header.type;
3097         struct evsel *evsel;
3098
3099         if (type != PERF_RECORD_SAMPLE) {
3100                 trace__process_event(trace, trace->host, event, sample);
3101                 return;
3102         }
3103
3104         evsel = perf_evlist__id2evsel(trace->evlist, sample->id);
3105         if (evsel == NULL) {
3106                 fprintf(trace->output, "Unknown tp ID %" PRIu64 ", skipping...\n", sample->id);
3107                 return;
3108         }
3109
3110         if (evswitch__discard(&trace->evswitch, evsel))
3111                 return;
3112
3113         trace__set_base_time(trace, evsel, sample);
3114
3115         if (evsel->core.attr.type == PERF_TYPE_TRACEPOINT &&
3116             sample->raw_data == NULL) {
3117                 fprintf(trace->output, "%s sample with no payload for tid: %d, cpu %d, raw_size=%d, skipping...\n",
3118                        evsel__name(evsel), sample->tid,
3119                        sample->cpu, sample->raw_size);
3120         } else {
3121                 tracepoint_handler handler = evsel->handler;
3122                 handler(trace, evsel, event, sample);
3123         }
3124
3125         if (trace->nr_events_printed >= trace->max_events && trace->max_events != ULONG_MAX)
3126                 interrupted = true;
3127 }
3128
3129 static int trace__add_syscall_newtp(struct trace *trace)
3130 {
3131         int ret = -1;
3132         struct evlist *evlist = trace->evlist;
3133         struct evsel *sys_enter, *sys_exit;
3134
3135         sys_enter = perf_evsel__raw_syscall_newtp("sys_enter", trace__sys_enter);
3136         if (sys_enter == NULL)
3137                 goto out;
3138
3139         if (perf_evsel__init_sc_tp_ptr_field(sys_enter, args))
3140                 goto out_delete_sys_enter;
3141
3142         sys_exit = perf_evsel__raw_syscall_newtp("sys_exit", trace__sys_exit);
3143         if (sys_exit == NULL)
3144                 goto out_delete_sys_enter;
3145
3146         if (perf_evsel__init_sc_tp_uint_field(sys_exit, ret))
3147                 goto out_delete_sys_exit;
3148
3149         evsel__config_callchain(sys_enter, &trace->opts, &callchain_param);
3150         evsel__config_callchain(sys_exit, &trace->opts, &callchain_param);
3151
3152         evlist__add(evlist, sys_enter);
3153         evlist__add(evlist, sys_exit);
3154
3155         if (callchain_param.enabled && !trace->kernel_syscallchains) {
3156                 /*
3157                  * We're interested only in the user space callchain
3158                  * leading to the syscall, allow overriding that for
3159                  * debugging reasons using --kernel_syscall_callchains
3160                  */
3161                 sys_exit->core.attr.exclude_callchain_kernel = 1;
3162         }
3163
3164         trace->syscalls.events.sys_enter = sys_enter;
3165         trace->syscalls.events.sys_exit  = sys_exit;
3166
3167         ret = 0;
3168 out:
3169         return ret;
3170
3171 out_delete_sys_exit:
3172         evsel__delete_priv(sys_exit);
3173 out_delete_sys_enter:
3174         evsel__delete_priv(sys_enter);
3175         goto out;
3176 }
3177
3178 static int trace__set_ev_qualifier_tp_filter(struct trace *trace)
3179 {
3180         int err = -1;
3181         struct evsel *sys_exit;
3182         char *filter = asprintf_expr_inout_ints("id", !trace->not_ev_qualifier,
3183                                                 trace->ev_qualifier_ids.nr,
3184                                                 trace->ev_qualifier_ids.entries);
3185
3186         if (filter == NULL)
3187                 goto out_enomem;
3188
3189         if (!evsel__append_tp_filter(trace->syscalls.events.sys_enter, filter)) {
3190                 sys_exit = trace->syscalls.events.sys_exit;
3191                 err = evsel__append_tp_filter(sys_exit, filter);
3192         }
3193
3194         free(filter);
3195 out:
3196         return err;
3197 out_enomem:
3198         errno = ENOMEM;
3199         goto out;
3200 }
3201
3202 #ifdef HAVE_LIBBPF_SUPPORT
3203 static struct bpf_map *trace__find_bpf_map_by_name(struct trace *trace, const char *name)
3204 {
3205         if (trace->bpf_obj == NULL)
3206                 return NULL;
3207
3208         return bpf_object__find_map_by_name(trace->bpf_obj, name);
3209 }
3210
3211 static void trace__set_bpf_map_filtered_pids(struct trace *trace)
3212 {
3213         trace->filter_pids.map = trace__find_bpf_map_by_name(trace, "pids_filtered");
3214 }
3215
3216 static void trace__set_bpf_map_syscalls(struct trace *trace)
3217 {
3218         trace->syscalls.map = trace__find_bpf_map_by_name(trace, "syscalls");
3219         trace->syscalls.prog_array.sys_enter = trace__find_bpf_map_by_name(trace, "syscalls_sys_enter");
3220         trace->syscalls.prog_array.sys_exit  = trace__find_bpf_map_by_name(trace, "syscalls_sys_exit");
3221 }
3222
3223 static struct bpf_program *trace__find_bpf_program_by_title(struct trace *trace, const char *name)
3224 {
3225         if (trace->bpf_obj == NULL)
3226                 return NULL;
3227
3228         return bpf_object__find_program_by_title(trace->bpf_obj, name);
3229 }
3230
3231 static struct bpf_program *trace__find_syscall_bpf_prog(struct trace *trace, struct syscall *sc,
3232                                                         const char *prog_name, const char *type)
3233 {
3234         struct bpf_program *prog;
3235
3236         if (prog_name == NULL) {
3237                 char default_prog_name[256];
3238                 scnprintf(default_prog_name, sizeof(default_prog_name), "!syscalls:sys_%s_%s", type, sc->name);
3239                 prog = trace__find_bpf_program_by_title(trace, default_prog_name);
3240                 if (prog != NULL)
3241                         goto out_found;
3242                 if (sc->fmt && sc->fmt->alias) {
3243                         scnprintf(default_prog_name, sizeof(default_prog_name), "!syscalls:sys_%s_%s", type, sc->fmt->alias);
3244                         prog = trace__find_bpf_program_by_title(trace, default_prog_name);
3245                         if (prog != NULL)
3246                                 goto out_found;
3247                 }
3248                 goto out_unaugmented;
3249         }
3250
3251         prog = trace__find_bpf_program_by_title(trace, prog_name);
3252
3253         if (prog != NULL) {
3254 out_found:
3255                 return prog;
3256         }
3257
3258         pr_debug("Couldn't find BPF prog \"%s\" to associate with syscalls:sys_%s_%s, not augmenting it\n",
3259                  prog_name, type, sc->name);
3260 out_unaugmented:
3261         return trace->syscalls.unaugmented_prog;
3262 }
3263
3264 static void trace__init_syscall_bpf_progs(struct trace *trace, int id)
3265 {
3266         struct syscall *sc = trace__syscall_info(trace, NULL, id);
3267
3268         if (sc == NULL)
3269                 return;
3270
3271         sc->bpf_prog.sys_enter = trace__find_syscall_bpf_prog(trace, sc, sc->fmt ? sc->fmt->bpf_prog_name.sys_enter : NULL, "enter");
3272         sc->bpf_prog.sys_exit  = trace__find_syscall_bpf_prog(trace, sc, sc->fmt ? sc->fmt->bpf_prog_name.sys_exit  : NULL,  "exit");
3273 }
3274
3275 static int trace__bpf_prog_sys_enter_fd(struct trace *trace, int id)
3276 {
3277         struct syscall *sc = trace__syscall_info(trace, NULL, id);
3278         return sc ? bpf_program__fd(sc->bpf_prog.sys_enter) : bpf_program__fd(trace->syscalls.unaugmented_prog);
3279 }
3280
3281 static int trace__bpf_prog_sys_exit_fd(struct trace *trace, int id)
3282 {
3283         struct syscall *sc = trace__syscall_info(trace, NULL, id);
3284         return sc ? bpf_program__fd(sc->bpf_prog.sys_exit) : bpf_program__fd(trace->syscalls.unaugmented_prog);
3285 }
3286
3287 static void trace__init_bpf_map_syscall_args(struct trace *trace, int id, struct bpf_map_syscall_entry *entry)
3288 {
3289         struct syscall *sc = trace__syscall_info(trace, NULL, id);
3290         int arg = 0;
3291
3292         if (sc == NULL)
3293                 goto out;
3294
3295         for (; arg < sc->nr_args; ++arg) {
3296                 entry->string_args_len[arg] = 0;
3297                 if (sc->arg_fmt[arg].scnprintf == SCA_FILENAME) {
3298                         /* Should be set like strace -s strsize */
3299                         entry->string_args_len[arg] = PATH_MAX;
3300                 }
3301         }
3302 out:
3303         for (; arg < 6; ++arg)
3304                 entry->string_args_len[arg] = 0;
3305 }
3306 static int trace__set_ev_qualifier_bpf_filter(struct trace *trace)
3307 {
3308         int fd = bpf_map__fd(trace->syscalls.map);
3309         struct bpf_map_syscall_entry value = {
3310                 .enabled = !trace->not_ev_qualifier,
3311         };
3312         int err = 0;
3313         size_t i;
3314
3315         for (i = 0; i < trace->ev_qualifier_ids.nr; ++i) {
3316                 int key = trace->ev_qualifier_ids.entries[i];
3317
3318                 if (value.enabled) {
3319                         trace__init_bpf_map_syscall_args(trace, key, &value);
3320                         trace__init_syscall_bpf_progs(trace, key);
3321                 }
3322
3323                 err = bpf_map_update_elem(fd, &key, &value, BPF_EXIST);
3324                 if (err)
3325                         break;
3326         }
3327
3328         return err;
3329 }
3330
3331 static int __trace__init_syscalls_bpf_map(struct trace *trace, bool enabled)
3332 {
3333         int fd = bpf_map__fd(trace->syscalls.map);
3334         struct bpf_map_syscall_entry value = {
3335                 .enabled = enabled,
3336         };
3337         int err = 0, key;
3338
3339         for (key = 0; key < trace->sctbl->syscalls.nr_entries; ++key) {
3340                 if (enabled)
3341                         trace__init_bpf_map_syscall_args(trace, key, &value);
3342
3343                 err = bpf_map_update_elem(fd, &key, &value, BPF_ANY);
3344                 if (err)
3345                         break;
3346         }
3347
3348         return err;
3349 }
3350
3351 static int trace__init_syscalls_bpf_map(struct trace *trace)
3352 {
3353         bool enabled = true;
3354
3355         if (trace->ev_qualifier_ids.nr)
3356                 enabled = trace->not_ev_qualifier;
3357
3358         return __trace__init_syscalls_bpf_map(trace, enabled);
3359 }
3360
3361 static struct bpf_program *trace__find_usable_bpf_prog_entry(struct trace *trace, struct syscall *sc)
3362 {
3363         struct tep_format_field *field, *candidate_field;
3364         int id;
3365
3366         /*
3367          * We're only interested in syscalls that have a pointer:
3368          */
3369         for (field = sc->args; field; field = field->next) {
3370                 if (field->flags & TEP_FIELD_IS_POINTER)
3371                         goto try_to_find_pair;
3372         }
3373
3374         return NULL;
3375
3376 try_to_find_pair:
3377         for (id = 0; id < trace->sctbl->syscalls.nr_entries; ++id) {
3378                 struct syscall *pair = trace__syscall_info(trace, NULL, id);
3379                 struct bpf_program *pair_prog;
3380                 bool is_candidate = false;
3381
3382                 if (pair == NULL || pair == sc ||
3383                     pair->bpf_prog.sys_enter == trace->syscalls.unaugmented_prog)
3384                         continue;
3385
3386                 for (field = sc->args, candidate_field = pair->args;
3387                      field && candidate_field; field = field->next, candidate_field = candidate_field->next) {
3388                         bool is_pointer = field->flags & TEP_FIELD_IS_POINTER,
3389                              candidate_is_pointer = candidate_field->flags & TEP_FIELD_IS_POINTER;
3390
3391                         if (is_pointer) {
3392                                if (!candidate_is_pointer) {
3393                                         // The candidate just doesn't copies our pointer arg, might copy other pointers we want.
3394                                         continue;
3395                                }
3396                         } else {
3397                                 if (candidate_is_pointer) {
3398                                         // The candidate might copy a pointer we don't have, skip it.
3399                                         goto next_candidate;
3400                                 }
3401                                 continue;
3402                         }
3403
3404                         if (strcmp(field->type, candidate_field->type))
3405                                 goto next_candidate;
3406
3407                         is_candidate = true;
3408                 }
3409
3410                 if (!is_candidate)
3411                         goto next_candidate;
3412
3413                 /*
3414                  * Check if the tentative pair syscall augmenter has more pointers, if it has,
3415                  * then it may be collecting that and we then can't use it, as it would collect
3416                  * more than what is common to the two syscalls.
3417                  */
3418                 if (candidate_field) {
3419                         for (candidate_field = candidate_field->next; candidate_field; candidate_field = candidate_field->next)
3420                                 if (candidate_field->flags & TEP_FIELD_IS_POINTER)
3421                                         goto next_candidate;
3422                 }
3423
3424                 pair_prog = pair->bpf_prog.sys_enter;
3425                 /*
3426                  * If the pair isn't enabled, then its bpf_prog.sys_enter will not
3427                  * have been searched for, so search it here and if it returns the
3428                  * unaugmented one, then ignore it, otherwise we'll reuse that BPF
3429                  * program for a filtered syscall on a non-filtered one.
3430                  *
3431                  * For instance, we have "!syscalls:sys_enter_renameat" and that is
3432                  * useful for "renameat2".
3433                  */
3434                 if (pair_prog == NULL) {
3435                         pair_prog = trace__find_syscall_bpf_prog(trace, pair, pair->fmt ? pair->fmt->bpf_prog_name.sys_enter : NULL, "enter");
3436                         if (pair_prog == trace->syscalls.unaugmented_prog)
3437                                 goto next_candidate;
3438                 }
3439
3440                 pr_debug("Reusing \"%s\" BPF sys_enter augmenter for \"%s\"\n", pair->name, sc->name);
3441                 return pair_prog;
3442         next_candidate:
3443                 continue;
3444         }
3445
3446         return NULL;
3447 }
3448
3449 static int trace__init_syscalls_bpf_prog_array_maps(struct trace *trace)
3450 {
3451         int map_enter_fd = bpf_map__fd(trace->syscalls.prog_array.sys_enter),
3452             map_exit_fd  = bpf_map__fd(trace->syscalls.prog_array.sys_exit);
3453         int err = 0, key;
3454
3455         for (key = 0; key < trace->sctbl->syscalls.nr_entries; ++key) {
3456                 int prog_fd;
3457
3458                 if (!trace__syscall_enabled(trace, key))
3459                         continue;
3460
3461                 trace__init_syscall_bpf_progs(trace, key);
3462
3463                 // It'll get at least the "!raw_syscalls:unaugmented"
3464                 prog_fd = trace__bpf_prog_sys_enter_fd(trace, key);
3465                 err = bpf_map_update_elem(map_enter_fd, &key, &prog_fd, BPF_ANY);
3466                 if (err)
3467                         break;
3468                 prog_fd = trace__bpf_prog_sys_exit_fd(trace, key);
3469                 err = bpf_map_update_elem(map_exit_fd, &key, &prog_fd, BPF_ANY);
3470                 if (err)
3471                         break;
3472         }
3473
3474         /*
3475          * Now lets do a second pass looking for enabled syscalls without
3476          * an augmenter that have a signature that is a superset of another
3477          * syscall with an augmenter so that we can auto-reuse it.
3478          *
3479          * I.e. if we have an augmenter for the "open" syscall that has
3480          * this signature:
3481          *
3482          *   int open(const char *pathname, int flags, mode_t mode);
3483          *
3484          * I.e. that will collect just the first string argument, then we
3485          * can reuse it for the 'creat' syscall, that has this signature:
3486          *
3487          *   int creat(const char *pathname, mode_t mode);
3488          *
3489          * and for:
3490          *
3491          *   int stat(const char *pathname, struct stat *statbuf);
3492          *   int lstat(const char *pathname, struct stat *statbuf);
3493          *
3494          * Because the 'open' augmenter will collect the first arg as a string,
3495          * and leave alone all the other args, which already helps with
3496          * beautifying 'stat' and 'lstat''s pathname arg.
3497          *
3498          * Then, in time, when 'stat' gets an augmenter that collects both
3499          * first and second arg (this one on the raw_syscalls:sys_exit prog
3500          * array tail call, then that one will be used.
3501          */
3502         for (key = 0; key < trace->sctbl->syscalls.nr_entries; ++key) {
3503                 struct syscall *sc = trace__syscall_info(trace, NULL, key);
3504                 struct bpf_program *pair_prog;
3505                 int prog_fd;
3506
3507                 if (sc == NULL || sc->bpf_prog.sys_enter == NULL)
3508                         continue;
3509
3510                 /*
3511                  * For now we're just reusing the sys_enter prog, and if it
3512                  * already has an augmenter, we don't need to find one.
3513                  */
3514                 if (sc->bpf_prog.sys_enter != trace->syscalls.unaugmented_prog)
3515                         continue;
3516
3517                 /*
3518                  * Look at all the other syscalls for one that has a signature
3519                  * that is close enough that we can share:
3520                  */
3521                 pair_prog = trace__find_usable_bpf_prog_entry(trace, sc);
3522                 if (pair_prog == NULL)
3523                         continue;
3524
3525                 sc->bpf_prog.sys_enter = pair_prog;
3526
3527                 /*
3528                  * Update the BPF_MAP_TYPE_PROG_SHARED for raw_syscalls:sys_enter
3529                  * with the fd for the program we're reusing:
3530                  */
3531                 prog_fd = bpf_program__fd(sc->bpf_prog.sys_enter);
3532                 err = bpf_map_update_elem(map_enter_fd, &key, &prog_fd, BPF_ANY);
3533                 if (err)
3534                         break;
3535         }
3536
3537
3538         return err;
3539 }
3540
3541 static void trace__delete_augmented_syscalls(struct trace *trace)
3542 {
3543         struct evsel *evsel, *tmp;
3544
3545         evlist__remove(trace->evlist, trace->syscalls.events.augmented);
3546         evsel__delete(trace->syscalls.events.augmented);
3547         trace->syscalls.events.augmented = NULL;
3548
3549         evlist__for_each_entry_safe(trace->evlist, tmp, evsel) {
3550                 if (evsel->bpf_obj == trace->bpf_obj) {
3551                         evlist__remove(trace->evlist, evsel);
3552                         evsel__delete(evsel);
3553                 }
3554
3555         }
3556
3557         bpf_object__close(trace->bpf_obj);
3558         trace->bpf_obj = NULL;
3559 }
3560 #else // HAVE_LIBBPF_SUPPORT
3561 static struct bpf_map *trace__find_bpf_map_by_name(struct trace *trace __maybe_unused,
3562                                                    const char *name __maybe_unused)
3563 {
3564         return NULL;
3565 }
3566
3567 static void trace__set_bpf_map_filtered_pids(struct trace *trace __maybe_unused)
3568 {
3569 }
3570
3571 static void trace__set_bpf_map_syscalls(struct trace *trace __maybe_unused)
3572 {
3573 }
3574
3575 static int trace__set_ev_qualifier_bpf_filter(struct trace *trace __maybe_unused)
3576 {
3577         return 0;
3578 }
3579
3580 static int trace__init_syscalls_bpf_map(struct trace *trace __maybe_unused)
3581 {
3582         return 0;
3583 }
3584
3585 static struct bpf_program *trace__find_bpf_program_by_title(struct trace *trace __maybe_unused,
3586                                                             const char *name __maybe_unused)
3587 {
3588         return NULL;
3589 }
3590
3591 static int trace__init_syscalls_bpf_prog_array_maps(struct trace *trace __maybe_unused)
3592 {
3593         return 0;
3594 }
3595
3596 static void trace__delete_augmented_syscalls(struct trace *trace __maybe_unused)
3597 {
3598 }
3599 #endif // HAVE_LIBBPF_SUPPORT
3600
3601 static bool trace__only_augmented_syscalls_evsels(struct trace *trace)
3602 {
3603         struct evsel *evsel;
3604
3605         evlist__for_each_entry(trace->evlist, evsel) {
3606                 if (evsel == trace->syscalls.events.augmented ||
3607                     evsel->bpf_obj == trace->bpf_obj)
3608                         continue;
3609
3610                 return false;
3611         }
3612
3613         return true;
3614 }
3615
3616 static int trace__set_ev_qualifier_filter(struct trace *trace)
3617 {
3618         if (trace->syscalls.map)
3619                 return trace__set_ev_qualifier_bpf_filter(trace);
3620         if (trace->syscalls.events.sys_enter)
3621                 return trace__set_ev_qualifier_tp_filter(trace);
3622         return 0;
3623 }
3624
3625 static int bpf_map__set_filter_pids(struct bpf_map *map __maybe_unused,
3626                                     size_t npids __maybe_unused, pid_t *pids __maybe_unused)
3627 {
3628         int err = 0;
3629 #ifdef HAVE_LIBBPF_SUPPORT
3630         bool value = true;
3631         int map_fd = bpf_map__fd(map);
3632         size_t i;
3633
3634         for (i = 0; i < npids; ++i) {
3635                 err = bpf_map_update_elem(map_fd, &pids[i], &value, BPF_ANY);
3636                 if (err)
3637                         break;
3638         }
3639 #endif
3640         return err;
3641 }
3642
3643 static int trace__set_filter_loop_pids(struct trace *trace)
3644 {
3645         unsigned int nr = 1, err;
3646         pid_t pids[32] = {
3647                 getpid(),
3648         };
3649         struct thread *thread = machine__find_thread(trace->host, pids[0], pids[0]);
3650
3651         while (thread && nr < ARRAY_SIZE(pids)) {
3652                 struct thread *parent = machine__find_thread(trace->host, thread->ppid, thread->ppid);
3653
3654                 if (parent == NULL)
3655                         break;
3656
3657                 if (!strcmp(thread__comm_str(parent), "sshd") ||
3658                     strstarts(thread__comm_str(parent), "gnome-terminal")) {
3659                         pids[nr++] = parent->tid;
3660                         break;
3661                 }
3662                 thread = parent;
3663         }
3664
3665         err = perf_evlist__append_tp_filter_pids(trace->evlist, nr, pids);
3666         if (!err && trace->filter_pids.map)
3667                 err = bpf_map__set_filter_pids(trace->filter_pids.map, nr, pids);
3668
3669         return err;
3670 }
3671
3672 static int trace__set_filter_pids(struct trace *trace)
3673 {
3674         int err = 0;
3675         /*
3676          * Better not use !target__has_task() here because we need to cover the
3677          * case where no threads were specified in the command line, but a
3678          * workload was, and in that case we will fill in the thread_map when
3679          * we fork the workload in perf_evlist__prepare_workload.
3680          */
3681         if (trace->filter_pids.nr > 0) {
3682                 err = perf_evlist__append_tp_filter_pids(trace->evlist, trace->filter_pids.nr,
3683                                                          trace->filter_pids.entries);
3684                 if (!err && trace->filter_pids.map) {
3685                         err = bpf_map__set_filter_pids(trace->filter_pids.map, trace->filter_pids.nr,
3686                                                        trace->filter_pids.entries);
3687                 }
3688         } else if (perf_thread_map__pid(trace->evlist->core.threads, 0) == -1) {
3689                 err = trace__set_filter_loop_pids(trace);
3690         }
3691
3692         return err;
3693 }
3694
3695 static int __trace__deliver_event(struct trace *trace, union perf_event *event)
3696 {
3697         struct evlist *evlist = trace->evlist;
3698         struct perf_sample sample;
3699         int err;
3700
3701         err = perf_evlist__parse_sample(evlist, event, &sample);
3702         if (err)
3703                 fprintf(trace->output, "Can't parse sample, err = %d, skipping...\n", err);
3704         else
3705                 trace__handle_event(trace, event, &sample);
3706
3707         return 0;
3708 }
3709
3710 static int __trace__flush_events(struct trace *trace)
3711 {
3712         u64 first = ordered_events__first_time(&trace->oe.data);
3713         u64 flush = trace->oe.last - NSEC_PER_SEC;
3714
3715         /* Is there some thing to flush.. */
3716         if (first && first < flush)
3717                 return ordered_events__flush_time(&trace->oe.data, flush);
3718
3719         return 0;
3720 }
3721
3722 static int trace__flush_events(struct trace *trace)
3723 {
3724         return !trace->sort_events ? 0 : __trace__flush_events(trace);
3725 }
3726
3727 static int trace__deliver_event(struct trace *trace, union perf_event *event)
3728 {
3729         int err;
3730
3731         if (!trace->sort_events)
3732                 return __trace__deliver_event(trace, event);
3733
3734         err = perf_evlist__parse_sample_timestamp(trace->evlist, event, &trace->oe.last);
3735         if (err && err != -1)
3736                 return err;
3737
3738         err = ordered_events__queue(&trace->oe.data, event, trace->oe.last, 0);
3739         if (err)
3740                 return err;
3741
3742         return trace__flush_events(trace);
3743 }
3744
3745 static int ordered_events__deliver_event(struct ordered_events *oe,
3746                                          struct ordered_event *event)
3747 {
3748         struct trace *trace = container_of(oe, struct trace, oe.data);
3749
3750         return __trace__deliver_event(trace, event->event);
3751 }
3752
3753 static struct syscall_arg_fmt *evsel__find_syscall_arg_fmt_by_name(struct evsel *evsel, char *arg)
3754 {
3755         struct tep_format_field *field;
3756         struct syscall_arg_fmt *fmt = __evsel__syscall_arg_fmt(evsel);
3757
3758         if (evsel->tp_format == NULL || fmt == NULL)
3759                 return NULL;
3760
3761         for (field = evsel->tp_format->format.fields; field; field = field->next, ++fmt)
3762                 if (strcmp(field->name, arg) == 0)
3763                         return fmt;
3764
3765         return NULL;
3766 }
3767
3768 static int trace__expand_filter(struct trace *trace __maybe_unused, struct evsel *evsel)
3769 {
3770         char *tok, *left = evsel->filter, *new_filter = evsel->filter;
3771
3772         while ((tok = strpbrk(left, "=<>!")) != NULL) {
3773                 char *right = tok + 1, *right_end;
3774
3775                 if (*right == '=')
3776                         ++right;
3777
3778                 while (isspace(*right))
3779                         ++right;
3780
3781                 if (*right == '\0')
3782                         break;
3783
3784                 while (!isalpha(*left))
3785                         if (++left == tok) {
3786                                 /*
3787                                  * Bail out, can't find the name of the argument that is being
3788                                  * used in the filter, let it try to set this filter, will fail later.
3789                                  */
3790                                 return 0;
3791                         }
3792
3793                 right_end = right + 1;
3794                 while (isalnum(*right_end) || *right_end == '_' || *right_end == '|')
3795                         ++right_end;
3796
3797                 if (isalpha(*right)) {
3798                         struct syscall_arg_fmt *fmt;
3799                         int left_size = tok - left,
3800                             right_size = right_end - right;
3801                         char arg[128];
3802
3803                         while (isspace(left[left_size - 1]))
3804                                 --left_size;
3805
3806                         scnprintf(arg, sizeof(arg), "%.*s", left_size, left);
3807
3808                         fmt = evsel__find_syscall_arg_fmt_by_name(evsel, arg);
3809                         if (fmt == NULL) {
3810                                 pr_err("\"%s\" not found in \"%s\", can't set filter \"%s\"\n",
3811                                        arg, evsel->name, evsel->filter);
3812                                 return -1;
3813                         }
3814
3815                         pr_debug2("trying to expand \"%s\" \"%.*s\" \"%.*s\" -> ",
3816                                  arg, (int)(right - tok), tok, right_size, right);
3817
3818                         if (fmt->strtoul) {
3819                                 u64 val;
3820                                 struct syscall_arg syscall_arg = {
3821                                         .parm = fmt->parm,
3822                                 };
3823
3824                                 if (fmt->strtoul(right, right_size, &syscall_arg, &val)) {
3825                                         char *n, expansion[19];
3826                                         int expansion_lenght = scnprintf(expansion, sizeof(expansion), "%#" PRIx64, val);
3827                                         int expansion_offset = right - new_filter;
3828
3829                                         pr_debug("%s", expansion);
3830
3831                                         if (asprintf(&n, "%.*s%s%s", expansion_offset, new_filter, expansion, right_end) < 0) {
3832                                                 pr_debug(" out of memory!\n");
3833                                                 free(new_filter);
3834                                                 return -1;
3835                                         }
3836                                         if (new_filter != evsel->filter)
3837                                                 free(new_filter);
3838                                         left = n + expansion_offset + expansion_lenght;
3839                                         new_filter = n;
3840                                 } else {
3841                                         pr_err("\"%.*s\" not found for \"%s\" in \"%s\", can't set filter \"%s\"\n",
3842                                                right_size, right, arg, evsel->name, evsel->filter);
3843                                         return -1;
3844                                 }
3845                         } else {
3846                                 pr_err("No resolver (strtoul) for \"%s\" in \"%s\", can't set filter \"%s\"\n",
3847                                        arg, evsel->name, evsel->filter);
3848                                 return -1;
3849                         }
3850
3851                         pr_debug("\n");
3852                 } else {
3853                         left = right_end;
3854                 }
3855         }
3856
3857         if (new_filter != evsel->filter) {
3858                 pr_debug("New filter for %s: %s\n", evsel->name, new_filter);
3859                 evsel__set_filter(evsel, new_filter);
3860                 free(new_filter);
3861         }
3862
3863         return 0;
3864 }
3865
3866 static int trace__expand_filters(struct trace *trace, struct evsel **err_evsel)
3867 {
3868         struct evlist *evlist = trace->evlist;
3869         struct evsel *evsel;
3870
3871         evlist__for_each_entry(evlist, evsel) {
3872                 if (evsel->filter == NULL)
3873                         continue;
3874
3875                 if (trace__expand_filter(trace, evsel)) {
3876                         *err_evsel = evsel;
3877                         return -1;
3878                 }
3879         }
3880
3881         return 0;
3882 }
3883
3884 static int trace__run(struct trace *trace, int argc, const char **argv)
3885 {
3886         struct evlist *evlist = trace->evlist;
3887         struct evsel *evsel, *pgfault_maj = NULL, *pgfault_min = NULL;
3888         int err = -1, i;
3889         unsigned long before;
3890         const bool forks = argc > 0;
3891         bool draining = false;
3892
3893         trace->live = true;
3894
3895         if (!trace->raw_augmented_syscalls) {
3896                 if (trace->trace_syscalls && trace__add_syscall_newtp(trace))
3897                         goto out_error_raw_syscalls;
3898
3899                 if (trace->trace_syscalls)
3900                         trace->vfs_getname = evlist__add_vfs_getname(evlist);
3901         }
3902
3903         if ((trace->trace_pgfaults & TRACE_PFMAJ)) {
3904                 pgfault_maj = evsel__new_pgfault(PERF_COUNT_SW_PAGE_FAULTS_MAJ);
3905                 if (pgfault_maj == NULL)
3906                         goto out_error_mem;
3907                 evsel__config_callchain(pgfault_maj, &trace->opts, &callchain_param);
3908                 evlist__add(evlist, pgfault_maj);
3909         }
3910
3911         if ((trace->trace_pgfaults & TRACE_PFMIN)) {
3912                 pgfault_min = evsel__new_pgfault(PERF_COUNT_SW_PAGE_FAULTS_MIN);
3913                 if (pgfault_min == NULL)
3914                         goto out_error_mem;
3915                 evsel__config_callchain(pgfault_min, &trace->opts, &callchain_param);
3916                 evlist__add(evlist, pgfault_min);
3917         }
3918
3919         if (trace->sched &&
3920             evlist__add_newtp(evlist, "sched", "sched_stat_runtime", trace__sched_stat_runtime))
3921                 goto out_error_sched_stat_runtime;
3922         /*
3923          * If a global cgroup was set, apply it to all the events without an
3924          * explicit cgroup. I.e.:
3925          *
3926          *      trace -G A -e sched:*switch
3927          *
3928          * Will set all raw_syscalls:sys_{enter,exit}, pgfault, vfs_getname, etc
3929          * _and_ sched:sched_switch to the 'A' cgroup, while:
3930          *
3931          * trace -e sched:*switch -G A
3932          *
3933          * will only set the sched:sched_switch event to the 'A' cgroup, all the
3934          * other events (raw_syscalls:sys_{enter,exit}, etc are left "without"
3935          * a cgroup (on the root cgroup, sys wide, etc).
3936          *
3937          * Multiple cgroups:
3938          *
3939          * trace -G A -e sched:*switch -G B
3940          *
3941          * the syscall ones go to the 'A' cgroup, the sched:sched_switch goes
3942          * to the 'B' cgroup.
3943          *
3944          * evlist__set_default_cgroup() grabs a reference of the passed cgroup
3945          * only for the evsels still without a cgroup, i.e. evsel->cgroup == NULL.
3946          */
3947         if (trace->cgroup)
3948                 evlist__set_default_cgroup(trace->evlist, trace->cgroup);
3949
3950         err = perf_evlist__create_maps(evlist, &trace->opts.target);
3951         if (err < 0) {
3952                 fprintf(trace->output, "Problems parsing the target to trace, check your options!\n");
3953                 goto out_delete_evlist;
3954         }
3955
3956         err = trace__symbols_init(trace, evlist);
3957         if (err < 0) {
3958                 fprintf(trace->output, "Problems initializing symbol libraries!\n");
3959                 goto out_delete_evlist;
3960         }
3961
3962         perf_evlist__config(evlist, &trace->opts, &callchain_param);
3963
3964         signal(SIGCHLD, sig_handler);
3965         signal(SIGINT, sig_handler);
3966
3967         if (forks) {
3968                 err = perf_evlist__prepare_workload(evlist, &trace->opts.target,
3969                                                     argv, false, NULL);
3970                 if (err < 0) {
3971                         fprintf(trace->output, "Couldn't run the workload!\n");
3972                         goto out_delete_evlist;
3973                 }
3974         }
3975
3976         err = evlist__open(evlist);
3977         if (err < 0)
3978                 goto out_error_open;
3979
3980         err = bpf__apply_obj_config();
3981         if (err) {
3982                 char errbuf[BUFSIZ];
3983
3984                 bpf__strerror_apply_obj_config(err, errbuf, sizeof(errbuf));
3985                 pr_err("ERROR: Apply config to BPF failed: %s\n",
3986                          errbuf);
3987                 goto out_error_open;
3988         }
3989
3990         err = trace__set_filter_pids(trace);
3991         if (err < 0)
3992                 goto out_error_mem;
3993
3994         if (trace->syscalls.map)
3995                 trace__init_syscalls_bpf_map(trace);
3996
3997         if (trace->syscalls.prog_array.sys_enter)
3998                 trace__init_syscalls_bpf_prog_array_maps(trace);
3999
4000         if (trace->ev_qualifier_ids.nr > 0) {
4001                 err = trace__set_ev_qualifier_filter(trace);
4002                 if (err < 0)
4003                         goto out_errno;
4004
4005                 if (trace->syscalls.events.sys_exit) {
4006                         pr_debug("event qualifier tracepoint filter: %s\n",
4007                                  trace->syscalls.events.sys_exit->filter);
4008                 }
4009         }
4010
4011         /*
4012          * If the "close" syscall is not traced, then we will not have the
4013          * opportunity to, in syscall_arg__scnprintf_close_fd() invalidate the
4014          * fd->pathname table and were ending up showing the last value set by
4015          * syscalls opening a pathname and associating it with a descriptor or
4016          * reading it from /proc/pid/fd/ in cases where that doesn't make
4017          * sense.
4018          *
4019          *  So just disable this beautifier (SCA_FD, SCA_FDAT) when 'close' is
4020          *  not in use.
4021          */
4022         trace->fd_path_disabled = !trace__syscall_enabled(trace, syscalltbl__id(trace->sctbl, "close"));
4023
4024         err = trace__expand_filters(trace, &evsel);
4025         if (err)
4026                 goto out_delete_evlist;
4027         err = perf_evlist__apply_filters(evlist, &evsel);
4028         if (err < 0)
4029                 goto out_error_apply_filters;
4030
4031         if (trace->dump.map)
4032                 bpf_map__fprintf(trace->dump.map, trace->output);
4033
4034         err = evlist__mmap(evlist, trace->opts.mmap_pages);
4035         if (err < 0)
4036                 goto out_error_mmap;
4037
4038         if (!target__none(&trace->opts.target) && !trace->opts.initial_delay)
4039                 evlist__enable(evlist);
4040
4041         if (forks)
4042                 perf_evlist__start_workload(evlist);
4043
4044         if (trace->opts.initial_delay) {
4045                 usleep(trace->opts.initial_delay * 1000);
4046                 evlist__enable(evlist);
4047         }
4048
4049         trace->multiple_threads = perf_thread_map__pid(evlist->core.threads, 0) == -1 ||
4050                                   evlist->core.threads->nr > 1 ||
4051                                   evlist__first(evlist)->core.attr.inherit;
4052
4053         /*
4054          * Now that we already used evsel->core.attr to ask the kernel to setup the
4055          * events, lets reuse evsel->core.attr.sample_max_stack as the limit in
4056          * trace__resolve_callchain(), allowing per-event max-stack settings
4057          * to override an explicitly set --max-stack global setting.
4058          */
4059         evlist__for_each_entry(evlist, evsel) {
4060                 if (evsel__has_callchain(evsel) &&
4061                     evsel->core.attr.sample_max_stack == 0)
4062                         evsel->core.attr.sample_max_stack = trace->max_stack;
4063         }
4064 again:
4065         before = trace->nr_events;
4066
4067         for (i = 0; i < evlist->core.nr_mmaps; i++) {
4068                 union perf_event *event;
4069                 struct mmap *md;
4070
4071                 md = &evlist->mmap[i];
4072                 if (perf_mmap__read_init(&md->core) < 0)
4073                         continue;
4074
4075                 while ((event = perf_mmap__read_event(&md->core)) != NULL) {
4076                         ++trace->nr_events;
4077
4078                         err = trace__deliver_event(trace, event);
4079                         if (err)
4080                                 goto out_disable;
4081
4082                         perf_mmap__consume(&md->core);
4083
4084                         if (interrupted)
4085                                 goto out_disable;
4086
4087                         if (done && !draining) {
4088                                 evlist__disable(evlist);
4089                                 draining = true;
4090                         }
4091                 }
4092                 perf_mmap__read_done(&md->core);
4093         }
4094
4095         if (trace->nr_events == before) {
4096                 int timeout = done ? 100 : -1;
4097
4098                 if (!draining && evlist__poll(evlist, timeout) > 0) {
4099                         if (evlist__filter_pollfd(evlist, POLLERR | POLLHUP | POLLNVAL) == 0)
4100                                 draining = true;
4101
4102                         goto again;
4103                 } else {
4104                         if (trace__flush_events(trace))
4105                                 goto out_disable;
4106                 }
4107         } else {
4108                 goto again;
4109         }
4110
4111 out_disable:
4112         thread__zput(trace->current);
4113
4114         evlist__disable(evlist);
4115
4116         if (trace->sort_events)
4117                 ordered_events__flush(&trace->oe.data, OE_FLUSH__FINAL);
4118
4119         if (!err) {
4120                 if (trace->summary)
4121                         trace__fprintf_thread_summary(trace, trace->output);
4122
4123                 if (trace->show_tool_stats) {
4124                         fprintf(trace->output, "Stats:\n "
4125                                                " vfs_getname : %" PRIu64 "\n"
4126                                                " proc_getname: %" PRIu64 "\n",
4127                                 trace->stats.vfs_getname,
4128                                 trace->stats.proc_getname);
4129                 }
4130         }
4131
4132 out_delete_evlist:
4133         trace__symbols__exit(trace);
4134
4135         evlist__delete(evlist);
4136         cgroup__put(trace->cgroup);
4137         trace->evlist = NULL;
4138         trace->live = false;
4139         return err;
4140 {
4141         char errbuf[BUFSIZ];
4142
4143 out_error_sched_stat_runtime:
4144         tracing_path__strerror_open_tp(errno, errbuf, sizeof(errbuf), "sched", "sched_stat_runtime");
4145         goto out_error;
4146
4147 out_error_raw_syscalls:
4148         tracing_path__strerror_open_tp(errno, errbuf, sizeof(errbuf), "raw_syscalls", "sys_(enter|exit)");
4149         goto out_error;
4150
4151 out_error_mmap:
4152         evlist__strerror_mmap(evlist, errno, errbuf, sizeof(errbuf));
4153         goto out_error;
4154
4155 out_error_open:
4156         evlist__strerror_open(evlist, errno, errbuf, sizeof(errbuf));
4157
4158 out_error:
4159         fprintf(trace->output, "%s\n", errbuf);
4160         goto out_delete_evlist;
4161
4162 out_error_apply_filters:
4163         fprintf(trace->output,
4164                 "Failed to set filter \"%s\" on event %s with %d (%s)\n",
4165                 evsel->filter, evsel__name(evsel), errno,
4166                 str_error_r(errno, errbuf, sizeof(errbuf)));
4167         goto out_delete_evlist;
4168 }
4169 out_error_mem:
4170         fprintf(trace->output, "Not enough memory to run!\n");
4171         goto out_delete_evlist;
4172
4173 out_errno:
4174         fprintf(trace->output, "errno=%d,%s\n", errno, strerror(errno));
4175         goto out_delete_evlist;
4176 }
4177
4178 static int trace__replay(struct trace *trace)
4179 {
4180         const struct evsel_str_handler handlers[] = {
4181                 { "probe:vfs_getname",       trace__vfs_getname, },
4182         };
4183         struct perf_data data = {
4184                 .path  = input_name,
4185                 .mode  = PERF_DATA_MODE_READ,
4186                 .force = trace->force,
4187         };
4188         struct perf_session *session;
4189         struct evsel *evsel;
4190         int err = -1;
4191
4192         trace->tool.sample        = trace__process_sample;
4193         trace->tool.mmap          = perf_event__process_mmap;
4194         trace->tool.mmap2         = perf_event__process_mmap2;
4195         trace->tool.comm          = perf_event__process_comm;
4196         trace->tool.exit          = perf_event__process_exit;
4197         trace->tool.fork          = perf_event__process_fork;
4198         trace->tool.attr          = perf_event__process_attr;
4199         trace->tool.tracing_data  = perf_event__process_tracing_data;
4200         trace->tool.build_id      = perf_event__process_build_id;
4201         trace->tool.namespaces    = perf_event__process_namespaces;
4202
4203         trace->tool.ordered_events = true;
4204         trace->tool.ordering_requires_timestamps = true;
4205
4206         /* add tid to output */
4207         trace->multiple_threads = true;
4208
4209         session = perf_session__new(&data, false, &trace->tool);
4210         if (IS_ERR(session))
4211                 return PTR_ERR(session);
4212
4213         if (trace->opts.target.pid)
4214                 symbol_conf.pid_list_str = strdup(trace->opts.target.pid);
4215
4216         if (trace->opts.target.tid)
4217                 symbol_conf.tid_list_str = strdup(trace->opts.target.tid);
4218
4219         if (symbol__init(&session->header.env) < 0)
4220                 goto out;
4221
4222         trace->host = &session->machines.host;
4223
4224         err = perf_session__set_tracepoints_handlers(session, handlers);
4225         if (err)
4226                 goto out;
4227
4228         evsel = perf_evlist__find_tracepoint_by_name(session->evlist,
4229                                                      "raw_syscalls:sys_enter");
4230         /* older kernels have syscalls tp versus raw_syscalls */
4231         if (evsel == NULL)
4232                 evsel = perf_evlist__find_tracepoint_by_name(session->evlist,
4233                                                              "syscalls:sys_enter");
4234
4235         if (evsel &&
4236             (evsel__init_raw_syscall_tp(evsel, trace__sys_enter) < 0 ||
4237             perf_evsel__init_sc_tp_ptr_field(evsel, args))) {
4238                 pr_err("Error during initialize raw_syscalls:sys_enter event\n");
4239                 goto out;
4240         }
4241
4242         evsel = perf_evlist__find_tracepoint_by_name(session->evlist,
4243                                                      "raw_syscalls:sys_exit");
4244         if (evsel == NULL)
4245                 evsel = perf_evlist__find_tracepoint_by_name(session->evlist,
4246                                                              "syscalls:sys_exit");
4247         if (evsel &&
4248             (evsel__init_raw_syscall_tp(evsel, trace__sys_exit) < 0 ||
4249             perf_evsel__init_sc_tp_uint_field(evsel, ret))) {
4250                 pr_err("Error during initialize raw_syscalls:sys_exit event\n");
4251                 goto out;
4252         }
4253
4254         evlist__for_each_entry(session->evlist, evsel) {
4255                 if (evsel->core.attr.type == PERF_TYPE_SOFTWARE &&
4256                     (evsel->core.attr.config == PERF_COUNT_SW_PAGE_FAULTS_MAJ ||
4257                      evsel->core.attr.config == PERF_COUNT_SW_PAGE_FAULTS_MIN ||
4258                      evsel->core.attr.config == PERF_COUNT_SW_PAGE_FAULTS))
4259                         evsel->handler = trace__pgfault;
4260         }
4261
4262         setup_pager();
4263
4264         err = perf_session__process_events(session);
4265         if (err)
4266                 pr_err("Failed to process events, error %d", err);
4267
4268         else if (trace->summary)
4269                 trace__fprintf_thread_summary(trace, trace->output);
4270
4271 out:
4272         perf_session__delete(session);
4273
4274         return err;
4275 }
4276
4277 static size_t trace__fprintf_threads_header(FILE *fp)
4278 {
4279         size_t printed;
4280
4281         printed  = fprintf(fp, "\n Summary of events:\n\n");
4282
4283         return printed;
4284 }
4285
4286 DEFINE_RESORT_RB(syscall_stats, a->msecs > b->msecs,
4287         struct syscall_stats *stats;
4288         double               msecs;
4289         int                  syscall;
4290 )
4291 {
4292         struct int_node *source = rb_entry(nd, struct int_node, rb_node);
4293         struct syscall_stats *stats = source->priv;
4294
4295         entry->syscall = source->i;
4296         entry->stats   = stats;
4297         entry->msecs   = stats ? (u64)stats->stats.n * (avg_stats(&stats->stats) / NSEC_PER_MSEC) : 0;
4298 }
4299
4300 static size_t thread__dump_stats(struct thread_trace *ttrace,
4301                                  struct trace *trace, FILE *fp)
4302 {
4303         size_t printed = 0;
4304         struct syscall *sc;
4305         struct rb_node *nd;
4306         DECLARE_RESORT_RB_INTLIST(syscall_stats, ttrace->syscall_stats);
4307
4308         if (syscall_stats == NULL)
4309                 return 0;
4310
4311         printed += fprintf(fp, "\n");
4312
4313         printed += fprintf(fp, "   syscall            calls  errors  total       min       avg       max       stddev\n");
4314         printed += fprintf(fp, "                                     (msec)    (msec)    (msec)    (msec)        (%%)\n");
4315         printed += fprintf(fp, "   --------------- --------  ------ -------- --------- --------- ---------     ------\n");
4316
4317         resort_rb__for_each_entry(nd, syscall_stats) {
4318                 struct syscall_stats *stats = syscall_stats_entry->stats;
4319                 if (stats) {
4320                         double min = (double)(stats->stats.min) / NSEC_PER_MSEC;
4321                         double max = (double)(stats->stats.max) / NSEC_PER_MSEC;
4322                         double avg = avg_stats(&stats->stats);
4323                         double pct;
4324                         u64 n = (u64)stats->stats.n;
4325
4326                         pct = avg ? 100.0 * stddev_stats(&stats->stats) / avg : 0.0;
4327                         avg /= NSEC_PER_MSEC;
4328
4329                         sc = &trace->syscalls.table[syscall_stats_entry->syscall];
4330                         printed += fprintf(fp, "   %-15s", sc->name);
4331                         printed += fprintf(fp, " %8" PRIu64 " %6" PRIu64 " %9.3f %9.3f %9.3f",
4332                                            n, stats->nr_failures, syscall_stats_entry->msecs, min, avg);
4333                         printed += fprintf(fp, " %9.3f %9.2f%%\n", max, pct);
4334
4335                         if (trace->errno_summary && stats->nr_failures) {
4336                                 const char *arch_name = perf_env__arch(trace->host->env);
4337                                 int e;
4338
4339                                 for (e = 0; e < stats->max_errno; ++e) {
4340                                         if (stats->errnos[e] != 0)
4341                                                 fprintf(fp, "\t\t\t\t%s: %d\n", arch_syscalls__strerrno(arch_name, e + 1), stats->errnos[e]);
4342                                 }
4343                         }
4344                 }
4345         }
4346
4347         resort_rb__delete(syscall_stats);
4348         printed += fprintf(fp, "\n\n");
4349
4350         return printed;
4351 }
4352
4353 static size_t trace__fprintf_thread(FILE *fp, struct thread *thread, struct trace *trace)
4354 {
4355         size_t printed = 0;
4356         struct thread_trace *ttrace = thread__priv(thread);
4357         double ratio;
4358
4359         if (ttrace == NULL)
4360                 return 0;
4361
4362         ratio = (double)ttrace->nr_events / trace->nr_events * 100.0;
4363
4364         printed += fprintf(fp, " %s (%d), ", thread__comm_str(thread), thread->tid);
4365         printed += fprintf(fp, "%lu events, ", ttrace->nr_events);
4366         printed += fprintf(fp, "%.1f%%", ratio);
4367         if (ttrace->pfmaj)
4368                 printed += fprintf(fp, ", %lu majfaults", ttrace->pfmaj);
4369         if (ttrace->pfmin)
4370                 printed += fprintf(fp, ", %lu minfaults", ttrace->pfmin);
4371         if (trace->sched)
4372                 printed += fprintf(fp, ", %.3f msec\n", ttrace->runtime_ms);
4373         else if (fputc('\n', fp) != EOF)
4374                 ++printed;
4375
4376         printed += thread__dump_stats(ttrace, trace, fp);
4377
4378         return printed;
4379 }
4380
4381 static unsigned long thread__nr_events(struct thread_trace *ttrace)
4382 {
4383         return ttrace ? ttrace->nr_events : 0;
4384 }
4385
4386 DEFINE_RESORT_RB(threads, (thread__nr_events(a->thread->priv) < thread__nr_events(b->thread->priv)),
4387         struct thread *thread;
4388 )
4389 {
4390         entry->thread = rb_entry(nd, struct thread, rb_node);
4391 }
4392
4393 static size_t trace__fprintf_thread_summary(struct trace *trace, FILE *fp)
4394 {
4395         size_t printed = trace__fprintf_threads_header(fp);
4396         struct rb_node *nd;
4397         int i;
4398
4399         for (i = 0; i < THREADS__TABLE_SIZE; i++) {
4400                 DECLARE_RESORT_RB_MACHINE_THREADS(threads, trace->host, i);
4401
4402                 if (threads == NULL) {
4403                         fprintf(fp, "%s", "Error sorting output by nr_events!\n");
4404                         return 0;
4405                 }
4406
4407                 resort_rb__for_each_entry(nd, threads)
4408                         printed += trace__fprintf_thread(fp, threads_entry->thread, trace);
4409
4410                 resort_rb__delete(threads);
4411         }
4412         return printed;
4413 }
4414
4415 static int trace__set_duration(const struct option *opt, const char *str,
4416                                int unset __maybe_unused)
4417 {
4418         struct trace *trace = opt->value;
4419
4420         trace->duration_filter = atof(str);
4421         return 0;
4422 }
4423
4424 static int trace__set_filter_pids_from_option(const struct option *opt, const char *str,
4425                                               int unset __maybe_unused)
4426 {
4427         int ret = -1;
4428         size_t i;
4429         struct trace *trace = opt->value;
4430         /*
4431          * FIXME: introduce a intarray class, plain parse csv and create a
4432          * { int nr, int entries[] } struct...
4433          */
4434         struct intlist *list = intlist__new(str);
4435
4436         if (list == NULL)
4437                 return -1;
4438
4439         i = trace->filter_pids.nr = intlist__nr_entries(list) + 1;
4440         trace->filter_pids.entries = calloc(i, sizeof(pid_t));
4441
4442         if (trace->filter_pids.entries == NULL)
4443                 goto out;
4444
4445         trace->filter_pids.entries[0] = getpid();
4446
4447         for (i = 1; i < trace->filter_pids.nr; ++i)
4448                 trace->filter_pids.entries[i] = intlist__entry(list, i - 1)->i;
4449
4450         intlist__delete(list);
4451         ret = 0;
4452 out:
4453         return ret;
4454 }
4455
4456 static int trace__open_output(struct trace *trace, const char *filename)
4457 {
4458         struct stat st;
4459
4460         if (!stat(filename, &st) && st.st_size) {
4461                 char oldname[PATH_MAX];
4462
4463                 scnprintf(oldname, sizeof(oldname), "%s.old", filename);
4464                 unlink(oldname);
4465                 rename(filename, oldname);
4466         }
4467
4468         trace->output = fopen(filename, "w");
4469
4470         return trace->output == NULL ? -errno : 0;
4471 }
4472
4473 static int parse_pagefaults(const struct option *opt, const char *str,
4474                             int unset __maybe_unused)
4475 {
4476         int *trace_pgfaults = opt->value;
4477
4478         if (strcmp(str, "all") == 0)
4479                 *trace_pgfaults |= TRACE_PFMAJ | TRACE_PFMIN;
4480         else if (strcmp(str, "maj") == 0)
4481                 *trace_pgfaults |= TRACE_PFMAJ;
4482         else if (strcmp(str, "min") == 0)
4483                 *trace_pgfaults |= TRACE_PFMIN;
4484         else
4485                 return -1;
4486
4487         return 0;
4488 }
4489
4490 static void evlist__set_default_evsel_handler(struct evlist *evlist, void *handler)
4491 {
4492         struct evsel *evsel;
4493
4494         evlist__for_each_entry(evlist, evsel) {
4495                 if (evsel->handler == NULL)
4496                         evsel->handler = handler;
4497         }
4498 }
4499
4500 static void evsel__set_syscall_arg_fmt(struct evsel *evsel, const char *name)
4501 {
4502         struct syscall_arg_fmt *fmt = evsel__syscall_arg_fmt(evsel);
4503
4504         if (fmt) {
4505                 struct syscall_fmt *scfmt = syscall_fmt__find(name);
4506
4507                 if (scfmt) {
4508                         int skip = 0;
4509
4510                         if (strcmp(evsel->tp_format->format.fields->name, "__syscall_nr") == 0 ||
4511                             strcmp(evsel->tp_format->format.fields->name, "nr") == 0)
4512                                 ++skip;
4513
4514                         memcpy(fmt + skip, scfmt->arg, (evsel->tp_format->format.nr_fields - skip) * sizeof(*fmt));
4515                 }
4516         }
4517 }
4518
4519 static int evlist__set_syscall_tp_fields(struct evlist *evlist)
4520 {
4521         struct evsel *evsel;
4522
4523         evlist__for_each_entry(evlist, evsel) {
4524                 if (evsel->priv || !evsel->tp_format)
4525                         continue;
4526
4527                 if (strcmp(evsel->tp_format->system, "syscalls")) {
4528                         evsel__init_tp_arg_scnprintf(evsel);
4529                         continue;
4530                 }
4531
4532                 if (evsel__init_syscall_tp(evsel))
4533                         return -1;
4534
4535                 if (!strncmp(evsel->tp_format->name, "sys_enter_", 10)) {
4536                         struct syscall_tp *sc = __evsel__syscall_tp(evsel);
4537
4538                         if (__tp_field__init_ptr(&sc->args, sc->id.offset + sizeof(u64)))
4539                                 return -1;
4540
4541                         evsel__set_syscall_arg_fmt(evsel, evsel->tp_format->name + sizeof("sys_enter_") - 1);
4542                 } else if (!strncmp(evsel->tp_format->name, "sys_exit_", 9)) {
4543                         struct syscall_tp *sc = __evsel__syscall_tp(evsel);
4544
4545                         if (__tp_field__init_uint(&sc->ret, sizeof(u64), sc->id.offset + sizeof(u64), evsel->needs_swap))
4546                                 return -1;
4547
4548                         evsel__set_syscall_arg_fmt(evsel, evsel->tp_format->name + sizeof("sys_exit_") - 1);
4549                 }
4550         }
4551
4552         return 0;
4553 }
4554
4555 /*
4556  * XXX: Hackish, just splitting the combined -e+--event (syscalls
4557  * (raw_syscalls:{sys_{enter,exit}} + events (tracepoints, HW, SW, etc) to use
4558  * existing facilities unchanged (trace->ev_qualifier + parse_options()).
4559  *
4560  * It'd be better to introduce a parse_options() variant that would return a
4561  * list with the terms it didn't match to an event...
4562  */
4563 static int trace__parse_events_option(const struct option *opt, const char *str,
4564                                       int unset __maybe_unused)
4565 {
4566         struct trace *trace = (struct trace *)opt->value;
4567         const char *s = str;
4568         char *sep = NULL, *lists[2] = { NULL, NULL, };
4569         int len = strlen(str) + 1, err = -1, list, idx;
4570         char *strace_groups_dir = system_path(STRACE_GROUPS_DIR);
4571         char group_name[PATH_MAX];
4572         struct syscall_fmt *fmt;
4573
4574         if (strace_groups_dir == NULL)
4575                 return -1;
4576
4577         if (*s == '!') {
4578                 ++s;
4579                 trace->not_ev_qualifier = true;
4580         }
4581
4582         while (1) {
4583                 if ((sep = strchr(s, ',')) != NULL)
4584                         *sep = '\0';
4585
4586                 list = 0;
4587                 if (syscalltbl__id(trace->sctbl, s) >= 0 ||
4588                     syscalltbl__strglobmatch_first(trace->sctbl, s, &idx) >= 0) {
4589                         list = 1;
4590                         goto do_concat;
4591                 }
4592
4593                 fmt = syscall_fmt__find_by_alias(s);
4594                 if (fmt != NULL) {
4595                         list = 1;
4596                         s = fmt->name;
4597                 } else {
4598                         path__join(group_name, sizeof(group_name), strace_groups_dir, s);
4599                         if (access(group_name, R_OK) == 0)
4600                                 list = 1;
4601                 }
4602 do_concat:
4603                 if (lists[list]) {
4604                         sprintf(lists[list] + strlen(lists[list]), ",%s", s);
4605                 } else {
4606                         lists[list] = malloc(len);
4607                         if (lists[list] == NULL)
4608                                 goto out;
4609                         strcpy(lists[list], s);
4610                 }
4611
4612                 if (!sep)
4613                         break;
4614
4615                 *sep = ',';
4616                 s = sep + 1;
4617         }
4618
4619         if (lists[1] != NULL) {
4620                 struct strlist_config slist_config = {
4621                         .dirname = strace_groups_dir,
4622                 };
4623
4624                 trace->ev_qualifier = strlist__new(lists[1], &slist_config);
4625                 if (trace->ev_qualifier == NULL) {
4626                         fputs("Not enough memory to parse event qualifier", trace->output);
4627                         goto out;
4628                 }
4629
4630                 if (trace__validate_ev_qualifier(trace))
4631                         goto out;
4632                 trace->trace_syscalls = true;
4633         }
4634
4635         err = 0;
4636
4637         if (lists[0]) {
4638                 struct option o = OPT_CALLBACK('e', "event", &trace->evlist, "event",
4639                                                "event selector. use 'perf list' to list available events",
4640                                                parse_events_option);
4641                 err = parse_events_option(&o, lists[0], 0);
4642         }
4643 out:
4644         if (sep)
4645                 *sep = ',';
4646
4647         return err;
4648 }
4649
4650 static int trace__parse_cgroups(const struct option *opt, const char *str, int unset)
4651 {
4652         struct trace *trace = opt->value;
4653
4654         if (!list_empty(&trace->evlist->core.entries))
4655                 return parse_cgroups(opt, str, unset);
4656
4657         trace->cgroup = evlist__findnew_cgroup(trace->evlist, str);
4658
4659         return 0;
4660 }
4661
4662 static int trace__config(const char *var, const char *value, void *arg)
4663 {
4664         struct trace *trace = arg;
4665         int err = 0;
4666
4667         if (!strcmp(var, "trace.add_events")) {
4668                 trace->perfconfig_events = strdup(value);
4669                 if (trace->perfconfig_events == NULL) {
4670                         pr_err("Not enough memory for %s\n", "trace.add_events");
4671                         return -1;
4672                 }
4673         } else if (!strcmp(var, "trace.show_timestamp")) {
4674                 trace->show_tstamp = perf_config_bool(var, value);
4675         } else if (!strcmp(var, "trace.show_duration")) {
4676                 trace->show_duration = perf_config_bool(var, value);
4677         } else if (!strcmp(var, "trace.show_arg_names")) {
4678                 trace->show_arg_names = perf_config_bool(var, value);
4679                 if (!trace->show_arg_names)
4680                         trace->show_zeros = true;
4681         } else if (!strcmp(var, "trace.show_zeros")) {
4682                 bool new_show_zeros = perf_config_bool(var, value);
4683                 if (!trace->show_arg_names && !new_show_zeros) {
4684                         pr_warning("trace.show_zeros has to be set when trace.show_arg_names=no\n");
4685                         goto out;
4686                 }
4687                 trace->show_zeros = new_show_zeros;
4688         } else if (!strcmp(var, "trace.show_prefix")) {
4689                 trace->show_string_prefix = perf_config_bool(var, value);
4690         } else if (!strcmp(var, "trace.no_inherit")) {
4691                 trace->opts.no_inherit = perf_config_bool(var, value);
4692         } else if (!strcmp(var, "trace.args_alignment")) {
4693                 int args_alignment = 0;
4694                 if (perf_config_int(&args_alignment, var, value) == 0)
4695                         trace->args_alignment = args_alignment;
4696         } else if (!strcmp(var, "trace.tracepoint_beautifiers")) {
4697                 if (strcasecmp(value, "libtraceevent") == 0)
4698                         trace->libtraceevent_print = true;
4699                 else if (strcasecmp(value, "libbeauty") == 0)
4700                         trace->libtraceevent_print = false;
4701         }
4702 out:
4703         return err;
4704 }
4705
4706 int cmd_trace(int argc, const char **argv)
4707 {
4708         const char *trace_usage[] = {
4709                 "perf trace [<options>] [<command>]",
4710                 "perf trace [<options>] -- <command> [<options>]",
4711                 "perf trace record [<options>] [<command>]",
4712                 "perf trace record [<options>] -- <command> [<options>]",
4713                 NULL
4714         };
4715         struct trace trace = {
4716                 .opts = {
4717                         .target = {
4718                                 .uid       = UINT_MAX,
4719                                 .uses_mmap = true,
4720                         },
4721                         .user_freq     = UINT_MAX,
4722                         .user_interval = ULLONG_MAX,
4723                         .no_buffering  = true,
4724                         .mmap_pages    = UINT_MAX,
4725                 },
4726                 .output = stderr,
4727                 .show_comm = true,
4728                 .show_tstamp = true,
4729                 .show_duration = true,
4730                 .show_arg_names = true,
4731                 .args_alignment = 70,
4732                 .trace_syscalls = false,
4733                 .kernel_syscallchains = false,
4734                 .max_stack = UINT_MAX,
4735                 .max_events = ULONG_MAX,
4736         };
4737         const char *map_dump_str = NULL;
4738         const char *output_name = NULL;
4739         const struct option trace_options[] = {
4740         OPT_CALLBACK('e', "event", &trace, "event",
4741                      "event/syscall selector. use 'perf list' to list available events",
4742                      trace__parse_events_option),
4743         OPT_CALLBACK(0, "filter", &trace.evlist, "filter",
4744                      "event filter", parse_filter),
4745         OPT_BOOLEAN(0, "comm", &trace.show_comm,
4746                     "show the thread COMM next to its id"),
4747         OPT_BOOLEAN(0, "tool_stats", &trace.show_tool_stats, "show tool stats"),
4748         OPT_CALLBACK(0, "expr", &trace, "expr", "list of syscalls/events to trace",
4749                      trace__parse_events_option),
4750         OPT_STRING('o', "output", &output_name, "file", "output file name"),
4751         OPT_STRING('i', "input", &input_name, "file", "Analyze events in file"),
4752         OPT_STRING('p', "pid", &trace.opts.target.pid, "pid",
4753                     "trace events on existing process id"),
4754         OPT_STRING('t', "tid", &trace.opts.target.tid, "tid",
4755                     "trace events on existing thread id"),
4756         OPT_CALLBACK(0, "filter-pids", &trace, "CSV list of pids",
4757                      "pids to filter (by the kernel)", trace__set_filter_pids_from_option),
4758         OPT_BOOLEAN('a', "all-cpus", &trace.opts.target.system_wide,
4759                     "system-wide collection from all CPUs"),
4760         OPT_STRING('C', "cpu", &trace.opts.target.cpu_list, "cpu",
4761                     "list of cpus to monitor"),
4762         OPT_BOOLEAN(0, "no-inherit", &trace.opts.no_inherit,
4763                     "child tasks do not inherit counters"),
4764         OPT_CALLBACK('m', "mmap-pages", &trace.opts.mmap_pages, "pages",
4765                      "number of mmap data pages",
4766                      perf_evlist__parse_mmap_pages),
4767         OPT_STRING('u', "uid", &trace.opts.target.uid_str, "user",
4768                    "user to profile"),
4769         OPT_CALLBACK(0, "duration", &trace, "float",
4770                      "show only events with duration > N.M ms",
4771                      trace__set_duration),
4772 #ifdef HAVE_LIBBPF_SUPPORT
4773         OPT_STRING(0, "map-dump", &map_dump_str, "BPF map", "BPF map to periodically dump"),
4774 #endif
4775         OPT_BOOLEAN(0, "sched", &trace.sched, "show blocking scheduler events"),
4776         OPT_INCR('v', "verbose", &verbose, "be more verbose"),
4777         OPT_BOOLEAN('T', "time", &trace.full_time,
4778                     "Show full timestamp, not time relative to first start"),
4779         OPT_BOOLEAN(0, "failure", &trace.failure_only,
4780                     "Show only syscalls that failed"),
4781         OPT_BOOLEAN('s', "summary", &trace.summary_only,
4782                     "Show only syscall summary with statistics"),
4783         OPT_BOOLEAN('S', "with-summary", &trace.summary,
4784                     "Show all syscalls and summary with statistics"),
4785         OPT_BOOLEAN(0, "errno-summary", &trace.errno_summary,
4786                     "Show errno stats per syscall, use with -s or -S"),
4787         OPT_CALLBACK_DEFAULT('F', "pf", &trace.trace_pgfaults, "all|maj|min",
4788                      "Trace pagefaults", parse_pagefaults, "maj"),
4789         OPT_BOOLEAN(0, "syscalls", &trace.trace_syscalls, "Trace syscalls"),
4790         OPT_BOOLEAN('f', "force", &trace.force, "don't complain, do it"),
4791         OPT_CALLBACK(0, "call-graph", &trace.opts,
4792                      "record_mode[,record_size]", record_callchain_help,
4793                      &record_parse_callchain_opt),
4794         OPT_BOOLEAN(0, "libtraceevent_print", &trace.libtraceevent_print,
4795                     "Use libtraceevent to print the tracepoint arguments."),
4796         OPT_BOOLEAN(0, "kernel-syscall-graph", &trace.kernel_syscallchains,
4797                     "Show the kernel callchains on the syscall exit path"),
4798         OPT_ULONG(0, "max-events", &trace.max_events,
4799                 "Set the maximum number of events to print, exit after that is reached. "),
4800         OPT_UINTEGER(0, "min-stack", &trace.min_stack,
4801                      "Set the minimum stack depth when parsing the callchain, "
4802                      "anything below the specified depth will be ignored."),
4803         OPT_UINTEGER(0, "max-stack", &trace.max_stack,
4804                      "Set the maximum stack depth when parsing the callchain, "
4805                      "anything beyond the specified depth will be ignored. "
4806                      "Default: kernel.perf_event_max_stack or " __stringify(PERF_MAX_STACK_DEPTH)),
4807         OPT_BOOLEAN(0, "sort-events", &trace.sort_events,
4808                         "Sort batch of events before processing, use if getting out of order events"),
4809         OPT_BOOLEAN(0, "print-sample", &trace.print_sample,
4810                         "print the PERF_RECORD_SAMPLE PERF_SAMPLE_ info, for debugging"),
4811         OPT_UINTEGER(0, "proc-map-timeout", &proc_map_timeout,
4812                         "per thread proc mmap processing timeout in ms"),
4813         OPT_CALLBACK('G', "cgroup", &trace, "name", "monitor event in cgroup name only",
4814                      trace__parse_cgroups),
4815         OPT_INTEGER('D', "delay", &trace.opts.initial_delay,
4816                      "ms to wait before starting measurement after program "
4817                      "start"),
4818         OPTS_EVSWITCH(&trace.evswitch),
4819         OPT_END()
4820         };
4821         bool __maybe_unused max_stack_user_set = true;
4822         bool mmap_pages_user_set = true;
4823         struct evsel *evsel;
4824         const char * const trace_subcommands[] = { "record", NULL };
4825         int err = -1;
4826         char bf[BUFSIZ];
4827
4828         signal(SIGSEGV, sighandler_dump_stack);
4829         signal(SIGFPE, sighandler_dump_stack);
4830
4831         trace.evlist = evlist__new();
4832         trace.sctbl = syscalltbl__new();
4833
4834         if (trace.evlist == NULL || trace.sctbl == NULL) {
4835                 pr_err("Not enough memory to run!\n");
4836                 err = -ENOMEM;
4837                 goto out;
4838         }
4839
4840         /*
4841          * Parsing .perfconfig may entail creating a BPF event, that may need
4842          * to create BPF maps, so bump RLIM_MEMLOCK as the default 64K setting
4843          * is too small. This affects just this process, not touching the
4844          * global setting. If it fails we'll get something in 'perf trace -v'
4845          * to help diagnose the problem.
4846          */
4847         rlimit__bump_memlock();
4848
4849         err = perf_config(trace__config, &trace);
4850         if (err)
4851                 goto out;
4852
4853         argc = parse_options_subcommand(argc, argv, trace_options, trace_subcommands,
4854                                  trace_usage, PARSE_OPT_STOP_AT_NON_OPTION);
4855
4856         /*
4857          * Here we already passed thru trace__parse_events_option() and it has
4858          * already figured out if -e syscall_name, if not but if --event
4859          * foo:bar was used, the user is interested _just_ in those, say,
4860          * tracepoint events, not in the strace-like syscall-name-based mode.
4861          *
4862          * This is important because we need to check if strace-like mode is
4863          * needed to decided if we should filter out the eBPF
4864          * __augmented_syscalls__ code, if it is in the mix, say, via
4865          * .perfconfig trace.add_events, and filter those out.
4866          */
4867         if (!trace.trace_syscalls && !trace.trace_pgfaults &&
4868             trace.evlist->core.nr_entries == 0 /* Was --events used? */) {
4869                 trace.trace_syscalls = true;
4870         }
4871         /*
4872          * Now that we have --verbose figured out, lets see if we need to parse
4873          * events from .perfconfig, so that if those events fail parsing, say some
4874          * BPF program fails, then we'll be able to use --verbose to see what went
4875          * wrong in more detail.
4876          */
4877         if (trace.perfconfig_events != NULL) {
4878                 struct parse_events_error parse_err;
4879
4880                 bzero(&parse_err, sizeof(parse_err));
4881                 err = parse_events(trace.evlist, trace.perfconfig_events, &parse_err);
4882                 if (err) {
4883                         parse_events_print_error(&parse_err, trace.perfconfig_events);
4884                         goto out;
4885                 }
4886         }
4887
4888         if ((nr_cgroups || trace.cgroup) && !trace.opts.target.system_wide) {
4889                 usage_with_options_msg(trace_usage, trace_options,
4890                                        "cgroup monitoring only available in system-wide mode");
4891         }
4892
4893         evsel = bpf__setup_output_event(trace.evlist, "__augmented_syscalls__");
4894         if (IS_ERR(evsel)) {
4895                 bpf__strerror_setup_output_event(trace.evlist, PTR_ERR(evsel), bf, sizeof(bf));
4896                 pr_err("ERROR: Setup trace syscalls enter failed: %s\n", bf);
4897                 goto out;
4898         }
4899
4900         if (evsel) {
4901                 trace.syscalls.events.augmented = evsel;
4902
4903                 evsel = perf_evlist__find_tracepoint_by_name(trace.evlist, "raw_syscalls:sys_enter");
4904                 if (evsel == NULL) {
4905                         pr_err("ERROR: raw_syscalls:sys_enter not found in the augmented BPF object\n");
4906                         goto out;
4907                 }
4908
4909                 if (evsel->bpf_obj == NULL) {
4910                         pr_err("ERROR: raw_syscalls:sys_enter not associated to a BPF object\n");
4911                         goto out;
4912                 }
4913
4914                 trace.bpf_obj = evsel->bpf_obj;
4915
4916                 /*
4917                  * If we have _just_ the augmenter event but don't have a
4918                  * explicit --syscalls, then assume we want all strace-like
4919                  * syscalls:
4920                  */
4921                 if (!trace.trace_syscalls && trace__only_augmented_syscalls_evsels(&trace))
4922                         trace.trace_syscalls = true;
4923                 /*
4924                  * So, if we have a syscall augmenter, but trace_syscalls, aka
4925                  * strace-like syscall tracing is not set, then we need to trow
4926                  * away the augmenter, i.e. all the events that were created
4927                  * from that BPF object file.
4928                  *
4929                  * This is more to fix the current .perfconfig trace.add_events
4930                  * style of setting up the strace-like eBPF based syscall point
4931                  * payload augmenter.
4932                  *
4933                  * All this complexity will be avoided by adding an alternative
4934                  * to trace.add_events in the form of
4935                  * trace.bpf_augmented_syscalls, that will be only parsed if we
4936                  * need it.
4937                  *
4938                  * .perfconfig trace.add_events is still useful if we want, for
4939                  * instance, have msr_write.msr in some .perfconfig profile based
4940                  * 'perf trace --config determinism.profile' mode, where for some
4941                  * particular goal/workload type we want a set of events and
4942                  * output mode (with timings, etc) instead of having to add
4943                  * all via the command line.
4944                  *
4945                  * Also --config to specify an alternate .perfconfig file needs
4946                  * to be implemented.
4947                  */
4948                 if (!trace.trace_syscalls) {
4949                         trace__delete_augmented_syscalls(&trace);
4950                 } else {
4951                         trace__set_bpf_map_filtered_pids(&trace);
4952                         trace__set_bpf_map_syscalls(&trace);
4953                         trace.syscalls.unaugmented_prog = trace__find_bpf_program_by_title(&trace, "!raw_syscalls:unaugmented");
4954                 }
4955         }
4956
4957         err = bpf__setup_stdout(trace.evlist);
4958         if (err) {
4959                 bpf__strerror_setup_stdout(trace.evlist, err, bf, sizeof(bf));
4960                 pr_err("ERROR: Setup BPF stdout failed: %s\n", bf);
4961                 goto out;
4962         }
4963
4964         err = -1;
4965
4966         if (map_dump_str) {
4967                 trace.dump.map = trace__find_bpf_map_by_name(&trace, map_dump_str);
4968                 if (trace.dump.map == NULL) {
4969                         pr_err("ERROR: BPF map \"%s\" not found\n", map_dump_str);
4970                         goto out;
4971                 }
4972         }
4973
4974         if (trace.trace_pgfaults) {
4975                 trace.opts.sample_address = true;
4976                 trace.opts.sample_time = true;
4977         }
4978
4979         if (trace.opts.mmap_pages == UINT_MAX)
4980                 mmap_pages_user_set = false;
4981
4982         if (trace.max_stack == UINT_MAX) {
4983                 trace.max_stack = input_name ? PERF_MAX_STACK_DEPTH : sysctl__max_stack();
4984                 max_stack_user_set = false;
4985         }
4986
4987 #ifdef HAVE_DWARF_UNWIND_SUPPORT
4988         if ((trace.min_stack || max_stack_user_set) && !callchain_param.enabled) {
4989                 record_opts__parse_callchain(&trace.opts, &callchain_param, "dwarf", false);
4990         }
4991 #endif
4992
4993         if (callchain_param.enabled) {
4994                 if (!mmap_pages_user_set && geteuid() == 0)
4995                         trace.opts.mmap_pages = perf_event_mlock_kb_in_pages() * 4;
4996
4997                 symbol_conf.use_callchain = true;
4998         }
4999
5000         if (trace.evlist->core.nr_entries > 0) {
5001                 evlist__set_default_evsel_handler(trace.evlist, trace__event_handler);
5002                 if (evlist__set_syscall_tp_fields(trace.evlist)) {
5003                         perror("failed to set syscalls:* tracepoint fields");
5004                         goto out;
5005                 }
5006         }
5007
5008         if (trace.sort_events) {
5009                 ordered_events__init(&trace.oe.data, ordered_events__deliver_event, &trace);
5010                 ordered_events__set_copy_on_queue(&trace.oe.data, true);
5011         }
5012
5013         /*
5014          * If we are augmenting syscalls, then combine what we put in the
5015          * __augmented_syscalls__ BPF map with what is in the
5016          * syscalls:sys_exit_FOO tracepoints, i.e. just like we do without BPF,
5017          * combining raw_syscalls:sys_enter with raw_syscalls:sys_exit.
5018          *
5019          * We'll switch to look at two BPF maps, one for sys_enter and the
5020          * other for sys_exit when we start augmenting the sys_exit paths with
5021          * buffers that are being copied from kernel to userspace, think 'read'
5022          * syscall.
5023          */
5024         if (trace.syscalls.events.augmented) {
5025                 evlist__for_each_entry(trace.evlist, evsel) {
5026                         bool raw_syscalls_sys_exit = strcmp(evsel__name(evsel), "raw_syscalls:sys_exit") == 0;
5027
5028                         if (raw_syscalls_sys_exit) {
5029                                 trace.raw_augmented_syscalls = true;
5030                                 goto init_augmented_syscall_tp;
5031                         }
5032
5033                         if (trace.syscalls.events.augmented->priv == NULL &&
5034                             strstr(evsel__name(evsel), "syscalls:sys_enter")) {
5035                                 struct evsel *augmented = trace.syscalls.events.augmented;
5036                                 if (evsel__init_augmented_syscall_tp(augmented, evsel) ||
5037                                     evsel__init_augmented_syscall_tp_args(augmented))
5038                                         goto out;
5039                                 /*
5040                                  * Augmented is __augmented_syscalls__ BPF_OUTPUT event
5041                                  * Above we made sure we can get from the payload the tp fields
5042                                  * that we get from syscalls:sys_enter tracefs format file.
5043                                  */
5044                                 augmented->handler = trace__sys_enter;
5045                                 /*
5046                                  * Now we do the same for the *syscalls:sys_enter event so that
5047                                  * if we handle it directly, i.e. if the BPF prog returns 0 so
5048                                  * as not to filter it, then we'll handle it just like we would
5049                                  * for the BPF_OUTPUT one:
5050                                  */
5051                                 if (evsel__init_augmented_syscall_tp(evsel, evsel) ||
5052                                     evsel__init_augmented_syscall_tp_args(evsel))
5053                                         goto out;
5054                                 evsel->handler = trace__sys_enter;
5055                         }
5056
5057                         if (strstarts(evsel__name(evsel), "syscalls:sys_exit_")) {
5058                                 struct syscall_tp *sc;
5059 init_augmented_syscall_tp:
5060                                 if (evsel__init_augmented_syscall_tp(evsel, evsel))
5061                                         goto out;
5062                                 sc = __evsel__syscall_tp(evsel);
5063                                 /*
5064                                  * For now with BPF raw_augmented we hook into
5065                                  * raw_syscalls:sys_enter and there we get all
5066                                  * 6 syscall args plus the tracepoint common
5067                                  * fields and the syscall_nr (another long).
5068                                  * So we check if that is the case and if so
5069                                  * don't look after the sc->args_size but
5070                                  * always after the full raw_syscalls:sys_enter
5071                                  * payload, which is fixed.
5072                                  *
5073                                  * We'll revisit this later to pass
5074                                  * s->args_size to the BPF augmenter (now
5075                                  * tools/perf/examples/bpf/augmented_raw_syscalls.c,
5076                                  * so that it copies only what we need for each
5077                                  * syscall, like what happens when we use
5078                                  * syscalls:sys_enter_NAME, so that we reduce
5079                                  * the kernel/userspace traffic to just what is
5080                                  * needed for each syscall.
5081                                  */
5082                                 if (trace.raw_augmented_syscalls)
5083                                         trace.raw_augmented_syscalls_args_size = (6 + 1) * sizeof(long) + sc->id.offset;
5084                                 evsel__init_augmented_syscall_tp_ret(evsel);
5085                                 evsel->handler = trace__sys_exit;
5086                         }
5087                 }
5088         }
5089
5090         if ((argc >= 1) && (strcmp(argv[0], "record") == 0))
5091                 return trace__record(&trace, argc-1, &argv[1]);
5092
5093         /* Using just --errno-summary will trigger --summary */
5094         if (trace.errno_summary && !trace.summary && !trace.summary_only)
5095                 trace.summary_only = true;
5096
5097         /* summary_only implies summary option, but don't overwrite summary if set */
5098         if (trace.summary_only)
5099                 trace.summary = trace.summary_only;
5100
5101         if (output_name != NULL) {
5102                 err = trace__open_output(&trace, output_name);
5103                 if (err < 0) {
5104                         perror("failed to create output file");
5105                         goto out;
5106                 }
5107         }
5108
5109         err = evswitch__init(&trace.evswitch, trace.evlist, stderr);
5110         if (err)
5111                 goto out_close;
5112
5113         err = target__validate(&trace.opts.target);
5114         if (err) {
5115                 target__strerror(&trace.opts.target, err, bf, sizeof(bf));
5116                 fprintf(trace.output, "%s", bf);
5117                 goto out_close;
5118         }
5119
5120         err = target__parse_uid(&trace.opts.target);
5121         if (err) {
5122                 target__strerror(&trace.opts.target, err, bf, sizeof(bf));
5123                 fprintf(trace.output, "%s", bf);
5124                 goto out_close;
5125         }
5126
5127         if (!argc && target__none(&trace.opts.target))
5128                 trace.opts.target.system_wide = true;
5129
5130         if (input_name)
5131                 err = trace__replay(&trace);
5132         else
5133                 err = trace__run(&trace, argc, argv);
5134
5135 out_close:
5136         if (output_name != NULL)
5137                 fclose(trace.output);
5138 out:
5139         zfree(&trace.perfconfig_events);
5140         return err;
5141 }