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", "OBJ_PIN", "OBJ_GET", "PROG_ATTACH",
 711         "PROG_DETACH", "PROG_TEST_RUN", "PROG_GET_NEXT_ID", "MAP_GET_NEXT_ID",
 712         "PROG_GET_FD_BY_ID", "MAP_GET_FD_BY_ID", "OBJ_GET_INFO_BY_FD",
 713         "PROG_QUERY", "RAW_TRACEPOINT_OPEN", "BTF_LOAD", "BTF_GET_FD_BY_ID",
 714         "TASK_FD_QUERY", "MAP_LOOKUP_AND_DELETE_ELEM", "MAP_FREEZE",
 715         "BTF_GET_NEXT_ID", "MAP_LOOKUP_BATCH", "MAP_LOOKUP_AND_DELETE_BATCH",
 716         "MAP_UPDATE_BATCH", "MAP_DELETE_BATCH", "LINK_CREATE", "LINK_UPDATE",
 717         "LINK_GET_FD_BY_ID", "LINK_GET_NEXT_ID", "ENABLE_STATS", "ITER_CREATE",
 718         "LINK_DETACH", "PROG_BIND_MAP",
 719 };
 720 static DEFINE_STRARRAY(bpf_cmd, "BPF_");
 721
 722 static const char *fsmount_flags[] = {
 723         [1] = "CLOEXEC",
 724 };
 725 static DEFINE_STRARRAY(fsmount_flags, "FSMOUNT_");
 726
 727 #include "trace/beauty/generated/fsconfig_arrays.c"
 728
 729 static DEFINE_STRARRAY(fsconfig_cmds, "FSCONFIG_");
 730
 731 static const char *epoll_ctl_ops[] = { "ADD", "DEL", "MOD", };
 732 static DEFINE_STRARRAY_OFFSET(epoll_ctl_ops, "EPOLL_CTL_", 1);
 733
 734 static const char *itimers[] = { "REAL", "VIRTUAL", "PROF", };
 735 static DEFINE_STRARRAY(itimers, "ITIMER_");
 736
 737 static const char *keyctl_options[] = {
 738         "GET_KEYRING_ID", "JOIN_SESSION_KEYRING", "UPDATE", "REVOKE", "CHOWN",
 739         "SETPERM", "DESCRIBE", "CLEAR", "LINK", "UNLINK", "SEARCH", "READ",
 740         "INSTANTIATE", "NEGATE", "SET_REQKEY_KEYRING", "SET_TIMEOUT",
 741         "ASSUME_AUTHORITY", "GET_SECURITY", "SESSION_TO_PARENT", "REJECT",
 742         "INSTANTIATE_IOV", "INVALIDATE", "GET_PERSISTENT",
 743 };
 744 static DEFINE_STRARRAY(keyctl_options, "KEYCTL_");
 745
 746 static const char *whences[] = { "SET", "CUR", "END",
 747 #ifdef SEEK_DATA
 748 "DATA",
 749 #endif
 750 #ifdef SEEK_HOLE
 751 "HOLE",
 752 #endif
 753 };
 754 static DEFINE_STRARRAY(whences, "SEEK_");
 755
 756 static const char *fcntl_cmds[] = {
 757         "DUPFD", "GETFD", "SETFD", "GETFL", "SETFL", "GETLK", "SETLK",
 758         "SETLKW", "SETOWN", "GETOWN", "SETSIG", "GETSIG", "GETLK64",
 759         "SETLK64", "SETLKW64", "SETOWN_EX", "GETOWN_EX",
 760         "GETOWNER_UIDS",
 761 };
 762 static DEFINE_STRARRAY(fcntl_cmds, "F_");
 763
 764 static const char *fcntl_linux_specific_cmds[] = {
 765         "SETLEASE", "GETLEASE", "NOTIFY", [5] = "CANCELLK", "DUPFD_CLOEXEC",
 766         "SETPIPE_SZ", "GETPIPE_SZ", "ADD_SEALS", "GET_SEALS",
 767         "GET_RW_HINT", "SET_RW_HINT", "GET_FILE_RW_HINT", "SET_FILE_RW_HINT",
 768 };
 769
 770 static DEFINE_STRARRAY_OFFSET(fcntl_linux_specific_cmds, "F_", F_LINUX_SPECIFIC_BASE);
 771
 772 static struct strarray *fcntl_cmds_arrays[] = {
 773         &strarray__fcntl_cmds,
 774         &strarray__fcntl_linux_specific_cmds,
 775 };
 776
 777 static DEFINE_STRARRAYS(fcntl_cmds_arrays);
 778
 779 static const char *rlimit_resources[] = {
 780         "CPU", "FSIZE", "DATA", "STACK", "CORE", "RSS", "NPROC", "NOFILE",
 781         "MEMLOCK", "AS", "LOCKS", "SIGPENDING", "MSGQUEUE", "NICE", "RTPRIO",
 782         "RTTIME",
 783 };
 784 static DEFINE_STRARRAY(rlimit_resources, "RLIMIT_");
 785
 786 static const char *sighow[] = { "BLOCK", "UNBLOCK", "SETMASK", };
 787 static DEFINE_STRARRAY(sighow, "SIG_");
 788
 789 static const char *clockid[] = {
 790         "REALTIME", "MONOTONIC", "PROCESS_CPUTIME_ID", "THREAD_CPUTIME_ID",
 791         "MONOTONIC_RAW", "REALTIME_COARSE", "MONOTONIC_COARSE", "BOOTTIME",
 792         "REALTIME_ALARM", "BOOTTIME_ALARM", "SGI_CYCLE", "TAI"
 793 };
 794 static DEFINE_STRARRAY(clockid, "CLOCK_");
 795
 796 static size_t syscall_arg__scnprintf_access_mode(char *bf, size_t size,
 797                                                  struct syscall_arg *arg)
 798 {
 799         bool show_prefix = arg->show_string_prefix;
 800         const char *suffix = "_OK";
 801         size_t printed = 0;
 802         int mode = arg->val;
 803
 804         if (mode == F_OK) /* 0 */
 805                 return scnprintf(bf, size, "F%s", show_prefix ? suffix : "");
 806 #define P_MODE(n) \
 807         if (mode & n##_OK) { \
 808                 printed += scnprintf(bf + printed, size - printed, "%s%s", #n, show_prefix ? suffix : ""); \
 809                 mode &= ~n##_OK; \
 810         }
 811
 812         P_MODE(R);
 813         P_MODE(W);
 814         P_MODE(X);
 815 #undef P_MODE
 816
 817         if (mode)
 818                 printed += scnprintf(bf + printed, size - printed, "|%#x", mode);
 819
 820         return printed;
 821 }
 822
 823 #define SCA_ACCMODE syscall_arg__scnprintf_access_mode
 824
 825 static size_t syscall_arg__scnprintf_filename(char *bf, size_t size,
 826                                               struct syscall_arg *arg);
 827
 828 #define SCA_FILENAME syscall_arg__scnprintf_filename
 829
 830 static size_t syscall_arg__scnprintf_pipe_flags(char *bf, size_t size,
 831                                                 struct syscall_arg *arg)
 832 {
 833         bool show_prefix = arg->show_string_prefix;
 834         const char *prefix = "O_";
 835         int printed = 0, flags = arg->val;
 836
 837 #define P_FLAG(n) \
 838         if (flags & O_##n) { \
 839                 printed += scnprintf(bf + printed, size - printed, "%s%s%s", printed ? "|" : "", show_prefix ? prefix : "", #n); \
 840                 flags &= ~O_##n; \
 841         }
 842
 843         P_FLAG(CLOEXEC);
 844         P_FLAG(NONBLOCK);
 845 #undef P_FLAG
 846
 847         if (flags)
 848                 printed += scnprintf(bf + printed, size - printed, "%s%#x", printed ? "|" : "", flags);
 849
 850         return printed;
 851 }
 852
 853 #define SCA_PIPE_FLAGS syscall_arg__scnprintf_pipe_flags
 854
 855 #ifndef GRND_NONBLOCK
 856 #define GRND_NONBLOCK   0x0001
 857 #endif
 858 #ifndef GRND_RANDOM
 859 #define GRND_RANDOM     0x0002
 860 #endif
 861
 862 static size_t syscall_arg__scnprintf_getrandom_flags(char *bf, size_t size,
 863                                                    struct syscall_arg *arg)
 864 {
 865         bool show_prefix = arg->show_string_prefix;
 866         const char *prefix = "GRND_";
 867         int printed = 0, flags = arg->val;
 868
 869 #define P_FLAG(n) \
 870         if (flags & GRND_##n) { \
 871                 printed += scnprintf(bf + printed, size - printed, "%s%s%s", printed ? "|" : "", show_prefix ? prefix : "", #n); \
 872                 flags &= ~GRND_##n; \
 873         }
 874
 875         P_FLAG(RANDOM);
 876         P_FLAG(NONBLOCK);
 877 #undef P_FLAG
 878
 879         if (flags)
 880                 printed += scnprintf(bf + printed, size - printed, "%s%#x", printed ? "|" : "", flags);
 881
 882         return printed;
 883 }
 884
 885 #define SCA_GETRANDOM_FLAGS syscall_arg__scnprintf_getrandom_flags
 886
 887 #define STRARRAY(name, array) \
 888           { .scnprintf  = SCA_STRARRAY, \
 889             .strtoul    = STUL_STRARRAY, \
 890             .parm       = &strarray__##array, }
 891
 892 #define STRARRAY_FLAGS(name, array) \
 893           { .scnprintf  = SCA_STRARRAY_FLAGS, \
 894             .strtoul    = STUL_STRARRAY_FLAGS, \
 895             .parm       = &strarray__##array, }
 896
 897 #include "trace/beauty/arch_errno_names.c"
 898 #include "trace/beauty/eventfd.c"
 899 #include "trace/beauty/futex_op.c"
 900 #include "trace/beauty/futex_val3.c"
 901 #include "trace/beauty/mmap.c"
 902 #include "trace/beauty/mode_t.c"
 903 #include "trace/beauty/msg_flags.c"
 904 #include "trace/beauty/open_flags.c"
 905 #include "trace/beauty/perf_event_open.c"
 906 #include "trace/beauty/pid.c"
 907 #include "trace/beauty/sched_policy.c"
 908 #include "trace/beauty/seccomp.c"
 909 #include "trace/beauty/signum.c"
 910 #include "trace/beauty/socket_type.c"
 911 #include "trace/beauty/waitid_options.c"
 912
 913 static struct syscall_fmt syscall_fmts[] = {
 914         { .name     = "access",
 915           .arg = { [1] = { .scnprintf = SCA_ACCMODE,  /* mode */ }, }, },
 916         { .name     = "arch_prctl",
 917           .arg = { [0] = { .scnprintf = SCA_X86_ARCH_PRCTL_CODE, /* code */ },
 918                    [1] = { .scnprintf = SCA_PTR, /* arg2 */ }, }, },
 919         { .name     = "bind",
 920           .arg = { [0] = { .scnprintf = SCA_INT, /* fd */ },
 921                    [1] = { .scnprintf = SCA_SOCKADDR, /* umyaddr */ },
 922                    [2] = { .scnprintf = SCA_INT, /* addrlen */ }, }, },
 923         { .name     = "bpf",
 924           .arg = { [0] = STRARRAY(cmd, bpf_cmd), }, },
 925         { .name     = "brk",        .hexret = true,
 926           .arg = { [0] = { .scnprintf = SCA_PTR, /* brk */ }, }, },
 927         { .name     = "clock_gettime",
 928           .arg = { [0] = STRARRAY(clk_id, clockid), }, },
 929         { .name     = "clone",      .errpid = true, .nr_args = 5,
 930           .arg = { [0] = { .name = "flags",         .scnprintf = SCA_CLONE_FLAGS, },
 931                    [1] = { .name = "child_stack",   .scnprintf = SCA_HEX, },
 932                    [2] = { .name = "parent_tidptr", .scnprintf = SCA_HEX, },
 933                    [3] = { .name = "child_tidptr",  .scnprintf = SCA_HEX, },
 934                    [4] = { .name = "tls",           .scnprintf = SCA_HEX, }, }, },
 935         { .name     = "close",
 936           .arg = { [0] = { .scnprintf = SCA_CLOSE_FD, /* fd */ }, }, },
 937         { .name     = "connect",
 938           .arg = { [0] = { .scnprintf = SCA_INT, /* fd */ },
 939                    [1] = { .scnprintf = SCA_SOCKADDR, /* servaddr */ },
 940                    [2] = { .scnprintf = SCA_INT, /* addrlen */ }, }, },
 941         { .name     = "epoll_ctl",
 942           .arg = { [1] = STRARRAY(op, epoll_ctl_ops), }, },
 943         { .name     = "eventfd2",
 944           .arg = { [1] = { .scnprintf = SCA_EFD_FLAGS, /* flags */ }, }, },
 945         { .name     = "fchmodat",
 946           .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
 947         { .name     = "fchownat",
 948           .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
 949         { .name     = "fcntl",
 950           .arg = { [1] = { .scnprintf = SCA_FCNTL_CMD,  /* cmd */
 951                            .strtoul   = STUL_STRARRAYS,
 952                            .parm      = &strarrays__fcntl_cmds_arrays,
 953                            .show_zero = true, },
 954                    [2] = { .scnprintf =  SCA_FCNTL_ARG, /* arg */ }, }, },
 955         { .name     = "flock",
 956           .arg = { [1] = { .scnprintf = SCA_FLOCK, /* cmd */ }, }, },
 957         { .name     = "fsconfig",
 958           .arg = { [1] = STRARRAY(cmd, fsconfig_cmds), }, },
 959         { .name     = "fsmount",
 960           .arg = { [1] = STRARRAY_FLAGS(flags, fsmount_flags),
 961                    [2] = { .scnprintf = SCA_FSMOUNT_ATTR_FLAGS, /* attr_flags */ }, }, },
 962         { .name     = "fspick",
 963           .arg = { [0] = { .scnprintf = SCA_FDAT,         /* dfd */ },
 964                    [1] = { .scnprintf = SCA_FILENAME,     /* path */ },
 965                    [2] = { .scnprintf = SCA_FSPICK_FLAGS, /* flags */ }, }, },
 966         { .name     = "fstat", .alias = "newfstat", },
 967         { .name     = "fstatat", .alias = "newfstatat", },
 968         { .name     = "futex",
 969           .arg = { [1] = { .scnprintf = SCA_FUTEX_OP, /* op */ },
 970                    [5] = { .scnprintf = SCA_FUTEX_VAL3, /* val3 */ }, }, },
 971         { .name     = "futimesat",
 972           .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
 973         { .name     = "getitimer",
 974           .arg = { [0] = STRARRAY(which, itimers), }, },
 975         { .name     = "getpid",     .errpid = true, },
 976         { .name     = "getpgid",    .errpid = true, },
 977         { .name     = "getppid",    .errpid = true, },
 978         { .name     = "getrandom",
 979           .arg = { [2] = { .scnprintf = SCA_GETRANDOM_FLAGS, /* flags */ }, }, },
 980         { .name     = "getrlimit",
 981           .arg = { [0] = STRARRAY(resource, rlimit_resources), }, },
 982         { .name     = "gettid",     .errpid = true, },
 983         { .name     = "ioctl",
 984           .arg = {
 985 #if defined(__i386__) || defined(__x86_64__)
 986 /*
 987  * FIXME: Make this available to all arches.
 988  */
 989                    [1] = { .scnprintf = SCA_IOCTL_CMD, /* cmd */ },
 990                    [2] = { .scnprintf = SCA_HEX, /* arg */ }, }, },
 991 #else
 992                    [2] = { .scnprintf = SCA_HEX, /* arg */ }, }, },
 993 #endif
 994         { .name     = "kcmp",       .nr_args = 5,
 995           .arg = { [0] = { .name = "pid1",      .scnprintf = SCA_PID, },
 996                    [1] = { .name = "pid2",      .scnprintf = SCA_PID, },
 997                    [2] = { .name = "type",      .scnprintf = SCA_KCMP_TYPE, },
 998                    [3] = { .name = "idx1",      .scnprintf = SCA_KCMP_IDX, },
 999                    [4] = { .name = "idx2",      .scnprintf = SCA_KCMP_IDX, }, }, },
1000         { .name     = "keyctl",
1001           .arg = { [0] = STRARRAY(option, keyctl_options), }, },
1002         { .name     = "kill",
1003           .arg = { [1] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
1004         { .name     = "linkat",
1005           .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
1006         { .name     = "lseek",
1007           .arg = { [2] = STRARRAY(whence, whences), }, },
1008         { .name     = "lstat", .alias = "newlstat", },
1009         { .name     = "madvise",
1010           .arg = { [0] = { .scnprintf = SCA_HEX,      /* start */ },
1011                    [2] = { .scnprintf = SCA_MADV_BHV, /* behavior */ }, }, },
1012         { .name     = "mkdirat",
1013           .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
1014         { .name     = "mknodat",
1015           .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
1016         { .name     = "mmap",       .hexret = true,
1017 /* The standard mmap maps to old_mmap on s390x */
1018 #if defined(__s390x__)
1019         .alias = "old_mmap",
1020 #endif
1021           .arg = { [2] = { .scnprintf = SCA_MMAP_PROT,  /* prot */ },
1022                    [3] = { .scnprintf = SCA_MMAP_FLAGS, /* flags */
1023                            .strtoul   = STUL_STRARRAY_FLAGS,
1024                            .parm      = &strarray__mmap_flags, },
1025                    [5] = { .scnprintf = SCA_HEX,        /* offset */ }, }, },
1026         { .name     = "mount",
1027           .arg = { [0] = { .scnprintf = SCA_FILENAME, /* dev_name */ },
1028                    [3] = { .scnprintf = SCA_MOUNT_FLAGS, /* flags */
1029                            .mask_val  = SCAMV_MOUNT_FLAGS, /* flags */ }, }, },
1030         { .name     = "move_mount",
1031           .arg = { [0] = { .scnprintf = SCA_FDAT,       /* from_dfd */ },
1032                    [1] = { .scnprintf = SCA_FILENAME, /* from_pathname */ },
1033                    [2] = { .scnprintf = SCA_FDAT,       /* to_dfd */ },
1034                    [3] = { .scnprintf = SCA_FILENAME, /* to_pathname */ },
1035                    [4] = { .scnprintf = SCA_MOVE_MOUNT_FLAGS, /* flags */ }, }, },
1036         { .name     = "mprotect",
1037           .arg = { [0] = { .scnprintf = SCA_HEX,        /* start */ },
1038                    [2] = { .scnprintf = SCA_MMAP_PROT,  /* prot */ }, }, },
1039         { .name     = "mq_unlink",
1040           .arg = { [0] = { .scnprintf = SCA_FILENAME, /* u_name */ }, }, },
1041         { .name     = "mremap",     .hexret = true,
1042           .arg = { [3] = { .scnprintf = SCA_MREMAP_FLAGS, /* flags */ }, }, },
1043         { .name     = "name_to_handle_at",
1044           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
1045         { .name     = "newfstatat",
1046           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
1047         { .name     = "open",
1048           .arg = { [1] = { .scnprintf = SCA_OPEN_FLAGS, /* flags */ }, }, },
1049         { .name     = "open_by_handle_at",
1050           .arg = { [0] = { .scnprintf = SCA_FDAT,       /* dfd */ },
1051                    [2] = { .scnprintf = SCA_OPEN_FLAGS, /* flags */ }, }, },
1052         { .name     = "openat",
1053           .arg = { [0] = { .scnprintf = SCA_FDAT,       /* dfd */ },
1054                    [2] = { .scnprintf = SCA_OPEN_FLAGS, /* flags */ }, }, },
1055         { .name     = "perf_event_open",
1056           .arg = { [2] = { .scnprintf = SCA_INT,        /* cpu */ },
1057                    [3] = { .scnprintf = SCA_FD,         /* group_fd */ },
1058                    [4] = { .scnprintf = SCA_PERF_FLAGS, /* flags */ }, }, },
1059         { .name     = "pipe2",
1060           .arg = { [1] = { .scnprintf = SCA_PIPE_FLAGS, /* flags */ }, }, },
1061         { .name     = "pkey_alloc",
1062           .arg = { [1] = { .scnprintf = SCA_PKEY_ALLOC_ACCESS_RIGHTS,   /* access_rights */ }, }, },
1063         { .name     = "pkey_free",
1064           .arg = { [0] = { .scnprintf = SCA_INT,        /* key */ }, }, },
1065         { .name     = "pkey_mprotect",
1066           .arg = { [0] = { .scnprintf = SCA_HEX,        /* start */ },
1067                    [2] = { .scnprintf = SCA_MMAP_PROT,  /* prot */ },
1068                    [3] = { .scnprintf = SCA_INT,        /* pkey */ }, }, },
1069         { .name     = "poll", .timeout = true, },
1070         { .name     = "ppoll", .timeout = true, },
1071         { .name     = "prctl",
1072           .arg = { [0] = { .scnprintf = SCA_PRCTL_OPTION, /* option */
1073                            .strtoul   = STUL_STRARRAY,
1074                            .parm      = &strarray__prctl_options, },
1075                    [1] = { .scnprintf = SCA_PRCTL_ARG2, /* arg2 */ },
1076                    [2] = { .scnprintf = SCA_PRCTL_ARG3, /* arg3 */ }, }, },
1077         { .name     = "pread", .alias = "pread64", },
1078         { .name     = "preadv", .alias = "pread", },
1079         { .name     = "prlimit64",
1080           .arg = { [1] = STRARRAY(resource, rlimit_resources), }, },
1081         { .name     = "pwrite", .alias = "pwrite64", },
1082         { .name     = "readlinkat",
1083           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
1084         { .name     = "recvfrom",
1085           .arg = { [3] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
1086         { .name     = "recvmmsg",
1087           .arg = { [3] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
1088         { .name     = "recvmsg",
1089           .arg = { [2] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
1090         { .name     = "renameat",
1091           .arg = { [0] = { .scnprintf = SCA_FDAT, /* olddirfd */ },
1092                    [2] = { .scnprintf = SCA_FDAT, /* newdirfd */ }, }, },
1093         { .name     = "renameat2",
1094           .arg = { [0] = { .scnprintf = SCA_FDAT, /* olddirfd */ },
1095                    [2] = { .scnprintf = SCA_FDAT, /* newdirfd */ },
1096                    [4] = { .scnprintf = SCA_RENAMEAT2_FLAGS, /* flags */ }, }, },
1097         { .name     = "rt_sigaction",
1098           .arg = { [0] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
1099         { .name     = "rt_sigprocmask",
1100           .arg = { [0] = STRARRAY(how, sighow), }, },
1101         { .name     = "rt_sigqueueinfo",
1102           .arg = { [1] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
1103         { .name     = "rt_tgsigqueueinfo",
1104           .arg = { [2] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
1105         { .name     = "sched_setscheduler",
1106           .arg = { [1] = { .scnprintf = SCA_SCHED_POLICY, /* policy */ }, }, },
1107         { .name     = "seccomp",
1108           .arg = { [0] = { .scnprintf = SCA_SECCOMP_OP,    /* op */ },
1109                    [1] = { .scnprintf = SCA_SECCOMP_FLAGS, /* flags */ }, }, },
1110         { .name     = "select", .timeout = true, },
1111         { .name     = "sendfile", .alias = "sendfile64", },
1112         { .name     = "sendmmsg",
1113           .arg = { [3] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
1114         { .name     = "sendmsg",
1115           .arg = { [2] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
1116         { .name     = "sendto",
1117           .arg = { [3] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ },
1118                    [4] = { .scnprintf = SCA_SOCKADDR, /* addr */ }, }, },
1119         { .name     = "set_tid_address", .errpid = true, },
1120         { .name     = "setitimer",
1121           .arg = { [0] = STRARRAY(which, itimers), }, },
1122         { .name     = "setrlimit",
1123           .arg = { [0] = STRARRAY(resource, rlimit_resources), }, },
1124         { .name     = "socket",
1125           .arg = { [0] = STRARRAY(family, socket_families),
1126                    [1] = { .scnprintf = SCA_SK_TYPE, /* type */ },
1127                    [2] = { .scnprintf = SCA_SK_PROTO, /* protocol */ }, }, },
1128         { .name     = "socketpair",
1129           .arg = { [0] = STRARRAY(family, socket_families),
1130                    [1] = { .scnprintf = SCA_SK_TYPE, /* type */ },
1131                    [2] = { .scnprintf = SCA_SK_PROTO, /* protocol */ }, }, },
1132         { .name     = "stat", .alias = "newstat", },
1133         { .name     = "statx",
1134           .arg = { [0] = { .scnprintf = SCA_FDAT,        /* fdat */ },
1135                    [2] = { .scnprintf = SCA_STATX_FLAGS, /* flags */ } ,
1136                    [3] = { .scnprintf = SCA_STATX_MASK,  /* mask */ }, }, },
1137         { .name     = "swapoff",
1138           .arg = { [0] = { .scnprintf = SCA_FILENAME, /* specialfile */ }, }, },
1139         { .name     = "swapon",
1140           .arg = { [0] = { .scnprintf = SCA_FILENAME, /* specialfile */ }, }, },
1141         { .name     = "symlinkat",
1142           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
1143         { .name     = "sync_file_range",
1144           .arg = { [3] = { .scnprintf = SCA_SYNC_FILE_RANGE_FLAGS, /* flags */ }, }, },
1145         { .name     = "tgkill",
1146           .arg = { [2] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
1147         { .name     = "tkill",
1148           .arg = { [1] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
1149         { .name     = "umount2", .alias = "umount",
1150           .arg = { [0] = { .scnprintf = SCA_FILENAME, /* name */ }, }, },
1151         { .name     = "uname", .alias = "newuname", },
1152         { .name     = "unlinkat",
1153           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
1154         { .name     = "utimensat",
1155           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dirfd */ }, }, },
1156         { .name     = "wait4",      .errpid = true,
1157           .arg = { [2] = { .scnprintf = SCA_WAITID_OPTIONS, /* options */ }, }, },
1158         { .name     = "waitid",     .errpid = true,
1159           .arg = { [3] = { .scnprintf = SCA_WAITID_OPTIONS, /* options */ }, }, },
1160 };
1161
1162 static int syscall_fmt__cmp(const void *name, const void *fmtp)
1163 {
1164         const struct syscall_fmt *fmt = fmtp;
1165         return strcmp(name, fmt->name);
1166 }
1167
1168 static struct syscall_fmt *__syscall_fmt__find(struct syscall_fmt *fmts, const int nmemb, const char *name)
1169 {
1170         return bsearch(name, fmts, nmemb, sizeof(struct syscall_fmt), syscall_fmt__cmp);
1171 }
1172
1173 static struct syscall_fmt *syscall_fmt__find(const char *name)
1174 {
1175         const int nmemb = ARRAY_SIZE(syscall_fmts);
1176         return __syscall_fmt__find(syscall_fmts, nmemb, name);
1177 }
1178
1179 static struct syscall_fmt *__syscall_fmt__find_by_alias(struct syscall_fmt *fmts, const int nmemb, const char *alias)
1180 {
1181         int i;
1182
1183         for (i = 0; i < nmemb; ++i) {
1184                 if (fmts[i].alias && strcmp(fmts[i].alias, alias) == 0)
1185                         return &fmts[i];
1186         }
1187
1188         return NULL;
1189 }
1190
1191 static struct syscall_fmt *syscall_fmt__find_by_alias(const char *alias)
1192 {
1193         const int nmemb = ARRAY_SIZE(syscall_fmts);
1194         return __syscall_fmt__find_by_alias(syscall_fmts, nmemb, alias);
1195 }
1196
1197 /*
1198  * is_exit: is this "exit" or "exit_group"?
1199  * is_open: is this "open" or "openat"? To associate the fd returned in sys_exit with the pathname in sys_enter.
1200  * args_size: sum of the sizes of the syscall arguments, anything after that is augmented stuff: pathname for openat, etc.
1201  * nonexistent: Just a hole in the syscall table, syscall id not allocated
1202  */
1203 struct syscall {
1204         struct tep_event    *tp_format;
1205         int                 nr_args;
1206         int                 args_size;
1207         struct {
1208                 struct bpf_program *sys_enter,
1209                                    *sys_exit;
1210         }                   bpf_prog;
1211         bool                is_exit;
1212         bool                is_open;
1213         bool                nonexistent;
1214         struct tep_format_field *args;
1215         const char          *name;
1216         struct syscall_fmt  *fmt;
1217         struct syscall_arg_fmt *arg_fmt;
1218 };
1219
1220 /*
1221  * Must match what is in the BPF program:
1222  *
1223  * tools/perf/examples/bpf/augmented_raw_syscalls.c
1224  */
1225 struct bpf_map_syscall_entry {
1226         bool    enabled;
1227         u16     string_args_len[6];
1228 };
1229
1230 /*
1231  * We need to have this 'calculated' boolean because in some cases we really
1232  * don't know what is the duration of a syscall, for instance, when we start
1233  * a session and some threads are waiting for a syscall to finish, say 'poll',
1234  * in which case all we can do is to print "( ? ) for duration and for the
1235  * start timestamp.
1236  */
1237 static size_t fprintf_duration(unsigned long t, bool calculated, FILE *fp)
1238 {
1239         double duration = (double)t / NSEC_PER_MSEC;
1240         size_t printed = fprintf(fp, "(");
1241
1242         if (!calculated)
1243                 printed += fprintf(fp, "         ");
1244         else if (duration >= 1.0)
1245                 printed += color_fprintf(fp, PERF_COLOR_RED, "%6.3f ms", duration);
1246         else if (duration >= 0.01)
1247                 printed += color_fprintf(fp, PERF_COLOR_YELLOW, "%6.3f ms", duration);
1248         else
1249                 printed += color_fprintf(fp, PERF_COLOR_NORMAL, "%6.3f ms", duration);
1250         return printed + fprintf(fp, "): ");
1251 }
1252
1253 /**
1254  * filename.ptr: The filename char pointer that will be vfs_getname'd
1255  * filename.entry_str_pos: Where to insert the string translated from
1256  *                         filename.ptr by the vfs_getname tracepoint/kprobe.
1257  * ret_scnprintf: syscall args may set this to a different syscall return
1258  *                formatter, for instance, fcntl may return fds, file flags, etc.
1259  */
1260 struct thread_trace {
1261         u64               entry_time;
1262         bool              entry_pending;
1263         unsigned long     nr_events;
1264         unsigned long     pfmaj, pfmin;
1265         char              *entry_str;
1266         double            runtime_ms;
1267         size_t            (*ret_scnprintf)(char *bf, size_t size, struct syscall_arg *arg);
1268         struct {
1269                 unsigned long ptr;
1270                 short int     entry_str_pos;
1271                 bool          pending_open;
1272                 unsigned int  namelen;
1273                 char          *name;
1274         } filename;
1275         struct {
1276                 int           max;
1277                 struct file   *table;
1278         } files;
1279
1280         struct intlist *syscall_stats;
1281 };
1282
1283 static struct thread_trace *thread_trace__new(void)
1284 {
1285         struct thread_trace *ttrace =  zalloc(sizeof(struct thread_trace));
1286
1287         if (ttrace) {
1288                 ttrace->files.max = -1;
1289                 ttrace->syscall_stats = intlist__new(NULL);
1290         }
1291
1292         return ttrace;
1293 }
1294
1295 static struct thread_trace *thread__trace(struct thread *thread, FILE *fp)
1296 {
1297         struct thread_trace *ttrace;
1298
1299         if (thread == NULL)
1300                 goto fail;
1301
1302         if (thread__priv(thread) == NULL)
1303                 thread__set_priv(thread, thread_trace__new());
1304
1305         if (thread__priv(thread) == NULL)
1306                 goto fail;
1307
1308         ttrace = thread__priv(thread);
1309         ++ttrace->nr_events;
1310
1311         return ttrace;
1312 fail:
1313         color_fprintf(fp, PERF_COLOR_RED,
1314                       "WARNING: not enough memory, dropping samples!\n");
1315         return NULL;
1316 }
1317
1318
1319 void syscall_arg__set_ret_scnprintf(struct syscall_arg *arg,
1320                                     size_t (*ret_scnprintf)(char *bf, size_t size, struct syscall_arg *arg))
1321 {
1322         struct thread_trace *ttrace = thread__priv(arg->thread);
1323
1324         ttrace->ret_scnprintf = ret_scnprintf;
1325 }
1326
1327 #define TRACE_PFMAJ             (1 << 0)
1328 #define TRACE_PFMIN             (1 << 1)
1329
1330 static const size_t trace__entry_str_size = 2048;
1331
1332 static struct file *thread_trace__files_entry(struct thread_trace *ttrace, int fd)
1333 {
1334         if (fd < 0)
1335                 return NULL;
1336
1337         if (fd > ttrace->files.max) {
1338                 struct file *nfiles = realloc(ttrace->files.table, (fd + 1) * sizeof(struct file));
1339
1340                 if (nfiles == NULL)
1341                         return NULL;
1342
1343                 if (ttrace->files.max != -1) {
1344                         memset(nfiles + ttrace->files.max + 1, 0,
1345                                (fd - ttrace->files.max) * sizeof(struct file));
1346                 } else {
1347                         memset(nfiles, 0, (fd + 1) * sizeof(struct file));
1348                 }
1349
1350                 ttrace->files.table = nfiles;
1351                 ttrace->files.max   = fd;
1352         }
1353
1354         return ttrace->files.table + fd;
1355 }
1356
1357 struct file *thread__files_entry(struct thread *thread, int fd)
1358 {
1359         return thread_trace__files_entry(thread__priv(thread), fd);
1360 }
1361
1362 static int trace__set_fd_pathname(struct thread *thread, int fd, const char *pathname)
1363 {
1364         struct thread_trace *ttrace = thread__priv(thread);
1365         struct file *file = thread_trace__files_entry(ttrace, fd);
1366
1367         if (file != NULL) {
1368                 struct stat st;
1369                 if (stat(pathname, &st) == 0)
1370                         file->dev_maj = major(st.st_rdev);
1371                 file->pathname = strdup(pathname);
1372                 if (file->pathname)
1373                         return 0;
1374         }
1375
1376         return -1;
1377 }
1378
1379 static int thread__read_fd_path(struct thread *thread, int fd)
1380 {
1381         char linkname[PATH_MAX], pathname[PATH_MAX];
1382         struct stat st;
1383         int ret;
1384
1385         if (thread->pid_ == thread->tid) {
1386                 scnprintf(linkname, sizeof(linkname),
1387                           "/proc/%d/fd/%d", thread->pid_, fd);
1388         } else {
1389                 scnprintf(linkname, sizeof(linkname),
1390                           "/proc/%d/task/%d/fd/%d", thread->pid_, thread->tid, fd);
1391         }
1392
1393         if (lstat(linkname, &st) < 0 || st.st_size + 1 > (off_t)sizeof(pathname))
1394                 return -1;
1395
1396         ret = readlink(linkname, pathname, sizeof(pathname));
1397
1398         if (ret < 0 || ret > st.st_size)
1399                 return -1;
1400
1401         pathname[ret] = '\0';
1402         return trace__set_fd_pathname(thread, fd, pathname);
1403 }
1404
1405 static const char *thread__fd_path(struct thread *thread, int fd,
1406                                    struct trace *trace)
1407 {
1408         struct thread_trace *ttrace = thread__priv(thread);
1409
1410         if (ttrace == NULL || trace->fd_path_disabled)
1411                 return NULL;
1412
1413         if (fd < 0)
1414                 return NULL;
1415
1416         if ((fd > ttrace->files.max || ttrace->files.table[fd].pathname == NULL)) {
1417                 if (!trace->live)
1418                         return NULL;
1419                 ++trace->stats.proc_getname;
1420                 if (thread__read_fd_path(thread, fd))
1421                         return NULL;
1422         }
1423
1424         return ttrace->files.table[fd].pathname;
1425 }
1426
1427 size_t syscall_arg__scnprintf_fd(char *bf, size_t size, struct syscall_arg *arg)
1428 {
1429         int fd = arg->val;
1430         size_t printed = scnprintf(bf, size, "%d", fd);
1431         const char *path = thread__fd_path(arg->thread, fd, arg->trace);
1432
1433         if (path)
1434                 printed += scnprintf(bf + printed, size - printed, "<%s>", path);
1435
1436         return printed;
1437 }
1438
1439 size_t pid__scnprintf_fd(struct trace *trace, pid_t pid, int fd, char *bf, size_t size)
1440 {
1441         size_t printed = scnprintf(bf, size, "%d", fd);
1442         struct thread *thread = machine__find_thread(trace->host, pid, pid);
1443
1444         if (thread) {
1445                 const char *path = thread__fd_path(thread, fd, trace);
1446
1447                 if (path)
1448                         printed += scnprintf(bf + printed, size - printed, "<%s>", path);
1449
1450                 thread__put(thread);
1451         }
1452
1453         return printed;
1454 }
1455
1456 static size_t syscall_arg__scnprintf_close_fd(char *bf, size_t size,
1457                                               struct syscall_arg *arg)
1458 {
1459         int fd = arg->val;
1460         size_t printed = syscall_arg__scnprintf_fd(bf, size, arg);
1461         struct thread_trace *ttrace = thread__priv(arg->thread);
1462
1463         if (ttrace && fd >= 0 && fd <= ttrace->files.max)
1464                 zfree(&ttrace->files.table[fd].pathname);
1465
1466         return printed;
1467 }
1468
1469 static void thread__set_filename_pos(struct thread *thread, const char *bf,
1470                                      unsigned long ptr)
1471 {
1472         struct thread_trace *ttrace = thread__priv(thread);
1473
1474         ttrace->filename.ptr = ptr;
1475         ttrace->filename.entry_str_pos = bf - ttrace->entry_str;
1476 }
1477
1478 static size_t syscall_arg__scnprintf_augmented_string(struct syscall_arg *arg, char *bf, size_t size)
1479 {
1480         struct augmented_arg *augmented_arg = arg->augmented.args;
1481         size_t printed = scnprintf(bf, size, "\"%.*s\"", augmented_arg->size, augmented_arg->value);
1482         /*
1483          * So that the next arg with a payload can consume its augmented arg, i.e. for rename* syscalls
1484          * we would have two strings, each prefixed by its size.
1485          */
1486         int consumed = sizeof(*augmented_arg) + augmented_arg->size;
1487
1488         arg->augmented.args = ((void *)arg->augmented.args) + consumed;
1489         arg->augmented.size -= consumed;
1490
1491         return printed;
1492 }
1493
1494 static size_t syscall_arg__scnprintf_filename(char *bf, size_t size,
1495                                               struct syscall_arg *arg)
1496 {
1497         unsigned long ptr = arg->val;
1498
1499         if (arg->augmented.args)
1500                 return syscall_arg__scnprintf_augmented_string(arg, bf, size);
1501
1502         if (!arg->trace->vfs_getname)
1503                 return scnprintf(bf, size, "%#x", ptr);
1504
1505         thread__set_filename_pos(arg->thread, bf, ptr);
1506         return 0;
1507 }
1508
1509 static bool trace__filter_duration(struct trace *trace, double t)
1510 {
1511         return t < (trace->duration_filter * NSEC_PER_MSEC);
1512 }
1513
1514 static size_t __trace__fprintf_tstamp(struct trace *trace, u64 tstamp, FILE *fp)
1515 {
1516         double ts = (double)(tstamp - trace->base_time) / NSEC_PER_MSEC;
1517
1518         return fprintf(fp, "%10.3f ", ts);
1519 }
1520
1521 /*
1522  * We're handling tstamp=0 as an undefined tstamp, i.e. like when we are
1523  * using ttrace->entry_time for a thread that receives a sys_exit without
1524  * first having received a sys_enter ("poll" issued before tracing session
1525  * starts, lost sys_enter exit due to ring buffer overflow).
1526  */
1527 static size_t trace__fprintf_tstamp(struct trace *trace, u64 tstamp, FILE *fp)
1528 {
1529         if (tstamp > 0)
1530                 return __trace__fprintf_tstamp(trace, tstamp, fp);
1531
1532         return fprintf(fp, "         ? ");
1533 }
1534
1535 static bool done = false;
1536 static bool interrupted = false;
1537
1538 static void sig_handler(int sig)
1539 {
1540         done = true;
1541         interrupted = sig == SIGINT;
1542 }
1543
1544 static size_t trace__fprintf_comm_tid(struct trace *trace, struct thread *thread, FILE *fp)
1545 {
1546         size_t printed = 0;
1547
1548         if (trace->multiple_threads) {
1549                 if (trace->show_comm)
1550                         printed += fprintf(fp, "%.14s/", thread__comm_str(thread));
1551                 printed += fprintf(fp, "%d ", thread->tid);
1552         }
1553
1554         return printed;
1555 }
1556
1557 static size_t trace__fprintf_entry_head(struct trace *trace, struct thread *thread,
1558                                         u64 duration, bool duration_calculated, u64 tstamp, FILE *fp)
1559 {
1560         size_t printed = 0;
1561
1562         if (trace->show_tstamp)
1563                 printed = trace__fprintf_tstamp(trace, tstamp, fp);
1564         if (trace->show_duration)
1565                 printed += fprintf_duration(duration, duration_calculated, fp);
1566         return printed + trace__fprintf_comm_tid(trace, thread, fp);
1567 }
1568
1569 static int trace__process_event(struct trace *trace, struct machine *machine,
1570                                 union perf_event *event, struct perf_sample *sample)
1571 {
1572         int ret = 0;
1573
1574         switch (event->header.type) {
1575         case PERF_RECORD_LOST:
1576                 color_fprintf(trace->output, PERF_COLOR_RED,
1577                               "LOST %" PRIu64 " events!\n", event->lost.lost);
1578                 ret = machine__process_lost_event(machine, event, sample);
1579                 break;
1580         default:
1581                 ret = machine__process_event(machine, event, sample);
1582                 break;
1583         }
1584
1585         return ret;
1586 }
1587
1588 static int trace__tool_process(struct perf_tool *tool,
1589                                union perf_event *event,
1590                                struct perf_sample *sample,
1591                                struct machine *machine)
1592 {
1593         struct trace *trace = container_of(tool, struct trace, tool);
1594         return trace__process_event(trace, machine, event, sample);
1595 }
1596
1597 static char *trace__machine__resolve_kernel_addr(void *vmachine, unsigned long long *addrp, char **modp)
1598 {
1599         struct machine *machine = vmachine;
1600
1601         if (machine->kptr_restrict_warned)
1602                 return NULL;
1603
1604         if (symbol_conf.kptr_restrict) {
1605                 pr_warning("Kernel address maps (/proc/{kallsyms,modules}) are restricted.\n\n"
1606                            "Check /proc/sys/kernel/kptr_restrict and /proc/sys/kernel/perf_event_paranoid.\n\n"
1607                            "Kernel samples will not be resolved.\n");
1608                 machine->kptr_restrict_warned = true;
1609                 return NULL;
1610         }
1611
1612         return machine__resolve_kernel_addr(vmachine, addrp, modp);
1613 }
1614
1615 static int trace__symbols_init(struct trace *trace, struct evlist *evlist)
1616 {
1617         int err = symbol__init(NULL);
1618
1619         if (err)
1620                 return err;
1621
1622         trace->host = machine__new_host();
1623         if (trace->host == NULL)
1624                 return -ENOMEM;
1625
1626         err = trace_event__register_resolver(trace->host, trace__machine__resolve_kernel_addr);
1627         if (err < 0)
1628                 goto out;
1629
1630         err = __machine__synthesize_threads(trace->host, &trace->tool, &trace->opts.target,
1631                                             evlist->core.threads, trace__tool_process, false,
1632                                             1);
1633 out:
1634         if (err)
1635                 symbol__exit();
1636
1637         return err;
1638 }
1639
1640 static void trace__symbols__exit(struct trace *trace)
1641 {
1642         machine__exit(trace->host);
1643         trace->host = NULL;
1644
1645         symbol__exit();
1646 }
1647
1648 static int syscall__alloc_arg_fmts(struct syscall *sc, int nr_args)
1649 {
1650         int idx;
1651
1652         if (nr_args == 6 && sc->fmt && sc->fmt->nr_args != 0)
1653                 nr_args = sc->fmt->nr_args;
1654
1655         sc->arg_fmt = calloc(nr_args, sizeof(*sc->arg_fmt));
1656         if (sc->arg_fmt == NULL)
1657                 return -1;
1658
1659         for (idx = 0; idx < nr_args; ++idx) {
1660                 if (sc->fmt)
1661                         sc->arg_fmt[idx] = sc->fmt->arg[idx];
1662         }
1663
1664         sc->nr_args = nr_args;
1665         return 0;
1666 }
1667
1668 static struct syscall_arg_fmt syscall_arg_fmts__by_name[] = {
1669         { .name = "msr",        .scnprintf = SCA_X86_MSR,         .strtoul = STUL_X86_MSR,         },
1670         { .name = "vector",     .scnprintf = SCA_X86_IRQ_VECTORS, .strtoul = STUL_X86_IRQ_VECTORS, },
1671 };
1672
1673 static int syscall_arg_fmt__cmp(const void *name, const void *fmtp)
1674 {
1675        const struct syscall_arg_fmt *fmt = fmtp;
1676        return strcmp(name, fmt->name);
1677 }
1678
1679 static struct syscall_arg_fmt *
1680 __syscall_arg_fmt__find_by_name(struct syscall_arg_fmt *fmts, const int nmemb, const char *name)
1681 {
1682        return bsearch(name, fmts, nmemb, sizeof(struct syscall_arg_fmt), syscall_arg_fmt__cmp);
1683 }
1684
1685 static struct syscall_arg_fmt *syscall_arg_fmt__find_by_name(const char *name)
1686 {
1687        const int nmemb = ARRAY_SIZE(syscall_arg_fmts__by_name);
1688        return __syscall_arg_fmt__find_by_name(syscall_arg_fmts__by_name, nmemb, name);
1689 }
1690
1691 static struct tep_format_field *
1692 syscall_arg_fmt__init_array(struct syscall_arg_fmt *arg, struct tep_format_field *field)
1693 {
1694         struct tep_format_field *last_field = NULL;
1695         int len;
1696
1697         for (; field; field = field->next, ++arg) {
1698                 last_field = field;
1699
1700                 if (arg->scnprintf)
1701                         continue;
1702
1703                 len = strlen(field->name);
1704
1705                 if (strcmp(field->type, "const char *") == 0 &&
1706                     ((len >= 4 && strcmp(field->name + len - 4, "name") == 0) ||
1707                      strstr(field->name, "path") != NULL))
1708                         arg->scnprintf = SCA_FILENAME;
1709                 else if ((field->flags & TEP_FIELD_IS_POINTER) || strstr(field->name, "addr"))
1710                         arg->scnprintf = SCA_PTR;
1711                 else if (strcmp(field->type, "pid_t") == 0)
1712                         arg->scnprintf = SCA_PID;
1713                 else if (strcmp(field->type, "umode_t") == 0)
1714                         arg->scnprintf = SCA_MODE_T;
1715                 else if ((field->flags & TEP_FIELD_IS_ARRAY) && strstr(field->type, "char")) {
1716                         arg->scnprintf = SCA_CHAR_ARRAY;
1717                         arg->nr_entries = field->arraylen;
1718                 } else if ((strcmp(field->type, "int") == 0 ||
1719                           strcmp(field->type, "unsigned int") == 0 ||
1720                           strcmp(field->type, "long") == 0) &&
1721                          len >= 2 && strcmp(field->name + len - 2, "fd") == 0) {
1722                         /*
1723                          * /sys/kernel/tracing/events/syscalls/sys_enter*
1724                          * egrep 'field:.*fd;' .../format|sed -r 's/.*field:([a-z ]+) [a-z_]*fd.+/\1/g'|sort|uniq -c
1725                          * 65 int
1726                          * 23 unsigned int
1727                          * 7 unsigned long
1728                          */
1729                         arg->scnprintf = SCA_FD;
1730                } else {
1731                         struct syscall_arg_fmt *fmt = syscall_arg_fmt__find_by_name(field->name);
1732
1733                         if (fmt) {
1734                                 arg->scnprintf = fmt->scnprintf;
1735                                 arg->strtoul   = fmt->strtoul;
1736                         }
1737                 }
1738         }
1739
1740         return last_field;
1741 }
1742
1743 static int syscall__set_arg_fmts(struct syscall *sc)
1744 {
1745         struct tep_format_field *last_field = syscall_arg_fmt__init_array(sc->arg_fmt, sc->args);
1746
1747         if (last_field)
1748                 sc->args_size = last_field->offset + last_field->size;
1749
1750         return 0;
1751 }
1752
1753 static int trace__read_syscall_info(struct trace *trace, int id)
1754 {
1755         char tp_name[128];
1756         struct syscall *sc;
1757         const char *name = syscalltbl__name(trace->sctbl, id);
1758
1759 #ifdef HAVE_SYSCALL_TABLE_SUPPORT
1760         if (trace->syscalls.table == NULL) {
1761                 trace->syscalls.table = calloc(trace->sctbl->syscalls.max_id + 1, sizeof(*sc));
1762                 if (trace->syscalls.table == NULL)
1763                         return -ENOMEM;
1764         }
1765 #else
1766         if (id > trace->sctbl->syscalls.max_id || (id == 0 && trace->syscalls.table == NULL)) {
1767                 // When using libaudit we don't know beforehand what is the max syscall id
1768                 struct syscall *table = realloc(trace->syscalls.table, (id + 1) * sizeof(*sc));
1769
1770                 if (table == NULL)
1771                         return -ENOMEM;
1772
1773                 // Need to memset from offset 0 and +1 members if brand new
1774                 if (trace->syscalls.table == NULL)
1775                         memset(table, 0, (id + 1) * sizeof(*sc));
1776                 else
1777                         memset(table + trace->sctbl->syscalls.max_id + 1, 0, (id - trace->sctbl->syscalls.max_id) * sizeof(*sc));
1778
1779                 trace->syscalls.table         = table;
1780                 trace->sctbl->syscalls.max_id = id;
1781         }
1782 #endif
1783         sc = trace->syscalls.table + id;
1784         if (sc->nonexistent)
1785                 return 0;
1786
1787         if (name == NULL) {
1788                 sc->nonexistent = true;
1789                 return 0;
1790         }
1791
1792         sc->name = name;
1793         sc->fmt  = syscall_fmt__find(sc->name);
1794
1795         snprintf(tp_name, sizeof(tp_name), "sys_enter_%s", sc->name);
1796         sc->tp_format = trace_event__tp_format("syscalls", tp_name);
1797
1798         if (IS_ERR(sc->tp_format) && sc->fmt && sc->fmt->alias) {
1799                 snprintf(tp_name, sizeof(tp_name), "sys_enter_%s", sc->fmt->alias);
1800                 sc->tp_format = trace_event__tp_format("syscalls", tp_name);
1801         }
1802
1803         if (syscall__alloc_arg_fmts(sc, IS_ERR(sc->tp_format) ? 6 : sc->tp_format->format.nr_fields))
1804                 return -ENOMEM;
1805
1806         if (IS_ERR(sc->tp_format))
1807                 return PTR_ERR(sc->tp_format);
1808
1809         sc->args = sc->tp_format->format.fields;
1810         /*
1811          * We need to check and discard the first variable '__syscall_nr'
1812          * or 'nr' that mean the syscall number. It is needless here.
1813          * So drop '__syscall_nr' or 'nr' field but does not exist on older kernels.
1814          */
1815         if (sc->args && (!strcmp(sc->args->name, "__syscall_nr") || !strcmp(sc->args->name, "nr"))) {
1816                 sc->args = sc->args->next;
1817                 --sc->nr_args;
1818         }
1819
1820         sc->is_exit = !strcmp(name, "exit_group") || !strcmp(name, "exit");
1821         sc->is_open = !strcmp(name, "open") || !strcmp(name, "openat");
1822
1823         return syscall__set_arg_fmts(sc);
1824 }
1825
1826 static int evsel__init_tp_arg_scnprintf(struct evsel *evsel)
1827 {
1828         struct syscall_arg_fmt *fmt = evsel__syscall_arg_fmt(evsel);
1829
1830         if (fmt != NULL) {
1831                 syscall_arg_fmt__init_array(fmt, evsel->tp_format->format.fields);
1832                 return 0;
1833         }
1834
1835         return -ENOMEM;
1836 }
1837
1838 static int intcmp(const void *a, const void *b)
1839 {
1840         const int *one = a, *another = b;
1841
1842         return *one - *another;
1843 }
1844
1845 static int trace__validate_ev_qualifier(struct trace *trace)
1846 {
1847         int err = 0;
1848         bool printed_invalid_prefix = false;
1849         struct str_node *pos;
1850         size_t nr_used = 0, nr_allocated = strlist__nr_entries(trace->ev_qualifier);
1851
1852         trace->ev_qualifier_ids.entries = malloc(nr_allocated *
1853                                                  sizeof(trace->ev_qualifier_ids.entries[0]));
1854
1855         if (trace->ev_qualifier_ids.entries == NULL) {
1856                 fputs("Error:\tNot enough memory for allocating events qualifier ids\n",
1857                        trace->output);
1858                 err = -EINVAL;
1859                 goto out;
1860         }
1861
1862         strlist__for_each_entry(pos, trace->ev_qualifier) {
1863                 const char *sc = pos->s;
1864                 int id = syscalltbl__id(trace->sctbl, sc), match_next = -1;
1865
1866                 if (id < 0) {
1867                         id = syscalltbl__strglobmatch_first(trace->sctbl, sc, &match_next);
1868                         if (id >= 0)
1869                                 goto matches;
1870
1871                         if (!printed_invalid_prefix) {
1872                                 pr_debug("Skipping unknown syscalls: ");
1873                                 printed_invalid_prefix = true;
1874                         } else {
1875                                 pr_debug(", ");
1876                         }
1877
1878                         pr_debug("%s", sc);
1879                         continue;
1880                 }
1881 matches:
1882                 trace->ev_qualifier_ids.entries[nr_used++] = id;
1883                 if (match_next == -1)
1884                         continue;
1885
1886                 while (1) {
1887                         id = syscalltbl__strglobmatch_next(trace->sctbl, sc, &match_next);
1888                         if (id < 0)
1889                                 break;
1890                         if (nr_allocated == nr_used) {
1891                                 void *entries;
1892
1893                                 nr_allocated += 8;
1894                                 entries = realloc(trace->ev_qualifier_ids.entries,
1895                                                   nr_allocated * sizeof(trace->ev_qualifier_ids.entries[0]));
1896                                 if (entries == NULL) {
1897                                         err = -ENOMEM;
1898                                         fputs("\nError:\t Not enough memory for parsing\n", trace->output);
1899                                         goto out_free;
1900                                 }
1901                                 trace->ev_qualifier_ids.entries = entries;
1902                         }
1903                         trace->ev_qualifier_ids.entries[nr_used++] = id;
1904                 }
1905         }
1906
1907         trace->ev_qualifier_ids.nr = nr_used;
1908         qsort(trace->ev_qualifier_ids.entries, nr_used, sizeof(int), intcmp);
1909 out:
1910         if (printed_invalid_prefix)
1911                 pr_debug("\n");
1912         return err;
1913 out_free:
1914         zfree(&trace->ev_qualifier_ids.entries);
1915         trace->ev_qualifier_ids.nr = 0;
1916         goto out;
1917 }
1918
1919 static __maybe_unused bool trace__syscall_enabled(struct trace *trace, int id)
1920 {
1921         bool in_ev_qualifier;
1922
1923         if (trace->ev_qualifier_ids.nr == 0)
1924                 return true;
1925
1926         in_ev_qualifier = bsearch(&id, trace->ev_qualifier_ids.entries,
1927                                   trace->ev_qualifier_ids.nr, sizeof(int), intcmp) != NULL;
1928
1929         if (in_ev_qualifier)
1930                return !trace->not_ev_qualifier;
1931
1932         return trace->not_ev_qualifier;
1933 }
1934
1935 /*
1936  * args is to be interpreted as a series of longs but we need to handle
1937  * 8-byte unaligned accesses. args points to raw_data within the event
1938  * and raw_data is guaranteed to be 8-byte unaligned because it is
1939  * preceded by raw_size which is a u32. So we need to copy args to a temp
1940  * variable to read it. Most notably this avoids extended load instructions
1941  * on unaligned addresses
1942  */
1943 unsigned long syscall_arg__val(struct syscall_arg *arg, u8 idx)
1944 {
1945         unsigned long val;
1946         unsigned char *p = arg->args + sizeof(unsigned long) * idx;
1947
1948         memcpy(&val, p, sizeof(val));
1949         return val;
1950 }
1951
1952 static size_t syscall__scnprintf_name(struct syscall *sc, char *bf, size_t size,
1953                                       struct syscall_arg *arg)
1954 {
1955         if (sc->arg_fmt && sc->arg_fmt[arg->idx].name)
1956                 return scnprintf(bf, size, "%s: ", sc->arg_fmt[arg->idx].name);
1957
1958         return scnprintf(bf, size, "arg%d: ", arg->idx);
1959 }
1960
1961 /*
1962  * Check if the value is in fact zero, i.e. mask whatever needs masking, such
1963  * as mount 'flags' argument that needs ignoring some magic flag, see comment
1964  * in tools/perf/trace/beauty/mount_flags.c
1965  */
1966 static unsigned long syscall_arg_fmt__mask_val(struct syscall_arg_fmt *fmt, struct syscall_arg *arg, unsigned long val)
1967 {
1968         if (fmt && fmt->mask_val)
1969                 return fmt->mask_val(arg, val);
1970
1971         return val;
1972 }
1973
1974 static size_t syscall_arg_fmt__scnprintf_val(struct syscall_arg_fmt *fmt, char *bf, size_t size,
1975                                              struct syscall_arg *arg, unsigned long val)
1976 {
1977         if (fmt && fmt->scnprintf) {
1978                 arg->val = val;
1979                 if (fmt->parm)
1980                         arg->parm = fmt->parm;
1981                 return fmt->scnprintf(bf, size, arg);
1982         }
1983         return scnprintf(bf, size, "%ld", val);
1984 }
1985
1986 static size_t syscall__scnprintf_args(struct syscall *sc, char *bf, size_t size,
1987                                       unsigned char *args, void *augmented_args, int augmented_args_size,
1988                                       struct trace *trace, struct thread *thread)
1989 {
1990         size_t printed = 0;
1991         unsigned long val;
1992         u8 bit = 1;
1993         struct syscall_arg arg = {
1994                 .args   = args,
1995                 .augmented = {
1996                         .size = augmented_args_size,
1997                         .args = augmented_args,
1998                 },
1999                 .idx    = 0,
2000                 .mask   = 0,
2001                 .trace  = trace,
2002                 .thread = thread,
2003                 .show_string_prefix = trace->show_string_prefix,
2004         };
2005         struct thread_trace *ttrace = thread__priv(thread);
2006
2007         /*
2008          * Things like fcntl will set this in its 'cmd' formatter to pick the
2009          * right formatter for the return value (an fd? file flags?), which is
2010          * not needed for syscalls that always return a given type, say an fd.
2011          */
2012         ttrace->ret_scnprintf = NULL;
2013
2014         if (sc->args != NULL) {
2015                 struct tep_format_field *field;
2016
2017                 for (field = sc->args; field;
2018                      field = field->next, ++arg.idx, bit <<= 1) {
2019                         if (arg.mask & bit)
2020                                 continue;
2021
2022                         arg.fmt = &sc->arg_fmt[arg.idx];
2023                         val = syscall_arg__val(&arg, arg.idx);
2024                         /*
2025                          * Some syscall args need some mask, most don't and
2026                          * return val untouched.
2027                          */
2028                         val = syscall_arg_fmt__mask_val(&sc->arg_fmt[arg.idx], &arg, val);
2029
2030                         /*
2031                          * Suppress this argument if its value is zero and
2032                          * and we don't have a string associated in an
2033                          * strarray for it.
2034                          */
2035                         if (val == 0 &&
2036                             !trace->show_zeros &&
2037                             !(sc->arg_fmt &&
2038                               (sc->arg_fmt[arg.idx].show_zero ||
2039                                sc->arg_fmt[arg.idx].scnprintf == SCA_STRARRAY ||
2040                                sc->arg_fmt[arg.idx].scnprintf == SCA_STRARRAYS) &&
2041                               sc->arg_fmt[arg.idx].parm))
2042                                 continue;
2043
2044                         printed += scnprintf(bf + printed, size - printed, "%s", printed ? ", " : "");
2045
2046                         if (trace->show_arg_names)
2047                                 printed += scnprintf(bf + printed, size - printed, "%s: ", field->name);
2048
2049                         printed += syscall_arg_fmt__scnprintf_val(&sc->arg_fmt[arg.idx],
2050                                                                   bf + printed, size - printed, &arg, val);
2051                 }
2052         } else if (IS_ERR(sc->tp_format)) {
2053                 /*
2054                  * If we managed to read the tracepoint /format file, then we
2055                  * may end up not having any args, like with gettid(), so only
2056                  * print the raw args when we didn't manage to read it.
2057                  */
2058                 while (arg.idx < sc->nr_args) {
2059                         if (arg.mask & bit)
2060                                 goto next_arg;
2061                         val = syscall_arg__val(&arg, arg.idx);
2062                         if (printed)
2063                                 printed += scnprintf(bf + printed, size - printed, ", ");
2064                         printed += syscall__scnprintf_name(sc, bf + printed, size - printed, &arg);
2065                         printed += syscall_arg_fmt__scnprintf_val(&sc->arg_fmt[arg.idx], bf + printed, size - printed, &arg, val);
2066 next_arg:
2067                         ++arg.idx;
2068                         bit <<= 1;
2069                 }
2070         }
2071
2072         return printed;
2073 }
2074
2075 typedef int (*tracepoint_handler)(struct trace *trace, struct evsel *evsel,
2076                                   union perf_event *event,
2077                                   struct perf_sample *sample);
2078
2079 static struct syscall *trace__syscall_info(struct trace *trace,
2080                                            struct evsel *evsel, int id)
2081 {
2082         int err = 0;
2083
2084         if (id < 0) {
2085
2086                 /*
2087                  * XXX: Noticed on x86_64, reproduced as far back as 3.0.36, haven't tried
2088                  * before that, leaving at a higher verbosity level till that is
2089                  * explained. Reproduced with plain ftrace with:
2090                  *
2091                  * echo 1 > /t/events/raw_syscalls/sys_exit/enable
2092                  * grep "NR -1 " /t/trace_pipe
2093                  *
2094                  * After generating some load on the machine.
2095                  */
2096                 if (verbose > 1) {
2097                         static u64 n;
2098                         fprintf(trace->output, "Invalid syscall %d id, skipping (%s, %" PRIu64 ") ...\n",
2099                                 id, evsel__name(evsel), ++n);
2100                 }
2101                 return NULL;
2102         }
2103
2104         err = -EINVAL;
2105
2106 #ifdef HAVE_SYSCALL_TABLE_SUPPORT
2107         if (id > trace->sctbl->syscalls.max_id) {
2108 #else
2109         if (id >= trace->sctbl->syscalls.max_id) {
2110                 /*
2111                  * With libaudit we don't know beforehand what is the max_id,
2112                  * so we let trace__read_syscall_info() figure that out as we
2113                  * go on reading syscalls.
2114                  */
2115                 err = trace__read_syscall_info(trace, id);
2116                 if (err)
2117 #endif
2118                 goto out_cant_read;
2119         }
2120
2121         if ((trace->syscalls.table == NULL || trace->syscalls.table[id].name == NULL) &&
2122             (err = trace__read_syscall_info(trace, id)) != 0)
2123                 goto out_cant_read;
2124
2125         if (trace->syscalls.table[id].name == NULL) {
2126                 if (trace->syscalls.table[id].nonexistent)
2127                         return NULL;
2128                 goto out_cant_read;
2129         }
2130
2131         return &trace->syscalls.table[id];
2132
2133 out_cant_read:
2134         if (verbose > 0) {
2135                 char sbuf[STRERR_BUFSIZE];
2136                 fprintf(trace->output, "Problems reading syscall %d: %d (%s)", id, -err, str_error_r(-err, sbuf, sizeof(sbuf)));
2137                 if (id <= trace->sctbl->syscalls.max_id && trace->syscalls.table[id].name != NULL)
2138                         fprintf(trace->output, "(%s)", trace->syscalls.table[id].name);
2139                 fputs(" information\n", trace->output);
2140         }
2141         return NULL;
2142 }
2143
2144 struct syscall_stats {
2145         struct stats stats;
2146         u64          nr_failures;
2147         int          max_errno;
2148         u32          *errnos;
2149 };
2150
2151 static void thread__update_stats(struct thread *thread, struct thread_trace *ttrace,
2152                                  int id, struct perf_sample *sample, long err, bool errno_summary)
2153 {
2154         struct int_node *inode;
2155         struct syscall_stats *stats;
2156         u64 duration = 0;
2157
2158         inode = intlist__findnew(ttrace->syscall_stats, id);
2159         if (inode == NULL)
2160                 return;
2161
2162         stats = inode->priv;
2163         if (stats == NULL) {
2164                 stats = malloc(sizeof(*stats));
2165                 if (stats == NULL)
2166                         return;
2167
2168                 stats->nr_failures = 0;
2169                 stats->max_errno   = 0;
2170                 stats->errnos      = NULL;
2171                 init_stats(&stats->stats);
2172                 inode->priv = stats;
2173         }
2174
2175         if (ttrace->entry_time && sample->time > ttrace->entry_time)
2176                 duration = sample->time - ttrace->entry_time;
2177
2178         update_stats(&stats->stats, duration);
2179
2180         if (err < 0) {
2181                 ++stats->nr_failures;
2182
2183                 if (!errno_summary)
2184                         return;
2185
2186                 err = -err;
2187                 if (err > stats->max_errno) {
2188                         u32 *new_errnos = realloc(stats->errnos, err * sizeof(u32));
2189
2190                         if (new_errnos) {
2191                                 memset(new_errnos + stats->max_errno, 0, (err - stats->max_errno) * sizeof(u32));
2192                         } else {
2193                                 pr_debug("Not enough memory for errno stats for thread \"%s\"(%d/%d), results will be incomplete\n",
2194                                          thread__comm_str(thread), thread->pid_, thread->tid);
2195                                 return;
2196                         }
2197
2198                         stats->errnos = new_errnos;
2199                         stats->max_errno = err;
2200                 }
2201
2202                 ++stats->errnos[err - 1];
2203         }
2204 }
2205
2206 static int trace__printf_interrupted_entry(struct trace *trace)
2207 {
2208         struct thread_trace *ttrace;
2209         size_t printed;
2210         int len;
2211
2212         if (trace->failure_only || trace->current == NULL)
2213                 return 0;
2214
2215         ttrace = thread__priv(trace->current);
2216
2217         if (!ttrace->entry_pending)
2218                 return 0;
2219
2220         printed  = trace__fprintf_entry_head(trace, trace->current, 0, false, ttrace->entry_time, trace->output);
2221         printed += len = fprintf(trace->output, "%s)", ttrace->entry_str);
2222
2223         if (len < trace->args_alignment - 4)
2224                 printed += fprintf(trace->output, "%-*s", trace->args_alignment - 4 - len, " ");
2225
2226         printed += fprintf(trace->output, " ...\n");
2227
2228         ttrace->entry_pending = false;
2229         ++trace->nr_events_printed;
2230
2231         return printed;
2232 }
2233
2234 static int trace__fprintf_sample(struct trace *trace, struct evsel *evsel,
2235                                  struct perf_sample *sample, struct thread *thread)
2236 {
2237         int printed = 0;
2238
2239         if (trace->print_sample) {
2240                 double ts = (double)sample->time / NSEC_PER_MSEC;
2241
2242                 printed += fprintf(trace->output, "%22s %10.3f %s %d/%d [%d]\n",
2243                                    evsel__name(evsel), ts,
2244                                    thread__comm_str(thread),
2245                                    sample->pid, sample->tid, sample->cpu);
2246         }
2247
2248         return printed;
2249 }
2250
2251 static void *syscall__augmented_args(struct syscall *sc, struct perf_sample *sample, int *augmented_args_size, int raw_augmented_args_size)
2252 {
2253         void *augmented_args = NULL;
2254         /*
2255          * For now with BPF raw_augmented we hook into raw_syscalls:sys_enter
2256          * and there we get all 6 syscall args plus the tracepoint common fields
2257          * that gets calculated at the start and the syscall_nr (another long).
2258          * So we check if that is the case and if so don't look after the
2259          * sc->args_size but always after the full raw_syscalls:sys_enter payload,
2260          * which is fixed.
2261          *
2262          * We'll revisit this later to pass s->args_size to the BPF augmenter
2263          * (now tools/perf/examples/bpf/augmented_raw_syscalls.c, so that it
2264          * copies only what we need for each syscall, like what happens when we
2265          * use syscalls:sys_enter_NAME, so that we reduce the kernel/userspace
2266          * traffic to just what is needed for each syscall.
2267          */
2268         int args_size = raw_augmented_args_size ?: sc->args_size;
2269
2270         *augmented_args_size = sample->raw_size - args_size;
2271         if (*augmented_args_size > 0)
2272                 augmented_args = sample->raw_data + args_size;
2273
2274         return augmented_args;
2275 }
2276
2277 static void syscall__exit(struct syscall *sc)
2278 {
2279         if (!sc)
2280                 return;
2281
2282         free(sc->arg_fmt);
2283 }
2284
2285 static int trace__sys_enter(struct trace *trace, struct evsel *evsel,
2286                             union perf_event *event __maybe_unused,
2287                             struct perf_sample *sample)
2288 {
2289         char *msg;
2290         void *args;
2291         int printed = 0;
2292         struct thread *thread;
2293         int id = perf_evsel__sc_tp_uint(evsel, id, sample), err = -1;
2294         int augmented_args_size = 0;
2295         void *augmented_args = NULL;
2296         struct syscall *sc = trace__syscall_info(trace, evsel, id);
2297         struct thread_trace *ttrace;
2298
2299         if (sc == NULL)
2300                 return -1;
2301
2302         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2303         ttrace = thread__trace(thread, trace->output);
2304         if (ttrace == NULL)
2305                 goto out_put;
2306
2307         trace__fprintf_sample(trace, evsel, sample, thread);
2308
2309         args = perf_evsel__sc_tp_ptr(evsel, args, sample);
2310
2311         if (ttrace->entry_str == NULL) {
2312                 ttrace->entry_str = malloc(trace__entry_str_size);
2313                 if (!ttrace->entry_str)
2314                         goto out_put;
2315         }
2316
2317         if (!(trace->duration_filter || trace->summary_only || trace->min_stack))
2318                 trace__printf_interrupted_entry(trace);
2319         /*
2320          * If this is raw_syscalls.sys_enter, then it always comes with the 6 possible
2321          * arguments, even if the syscall being handled, say "openat", uses only 4 arguments
2322          * this breaks syscall__augmented_args() check for augmented args, as we calculate
2323          * syscall->args_size using each syscalls:sys_enter_NAME tracefs format file,
2324          * so when handling, say the openat syscall, we end up getting 6 args for the
2325          * raw_syscalls:sys_enter event, when we expected just 4, we end up mistakenly
2326          * thinking that the extra 2 u64 args are the augmented filename, so just check
2327          * here and avoid using augmented syscalls when the evsel is the raw_syscalls one.
2328          */
2329         if (evsel != trace->syscalls.events.sys_enter)
2330                 augmented_args = syscall__augmented_args(sc, sample, &augmented_args_size, trace->raw_augmented_syscalls_args_size);
2331         ttrace->entry_time = sample->time;
2332         msg = ttrace->entry_str;
2333         printed += scnprintf(msg + printed, trace__entry_str_size - printed, "%s(", sc->name);
2334
2335         printed += syscall__scnprintf_args(sc, msg + printed, trace__entry_str_size - printed,
2336                                            args, augmented_args, augmented_args_size, trace, thread);
2337
2338         if (sc->is_exit) {
2339                 if (!(trace->duration_filter || trace->summary_only || trace->failure_only || trace->min_stack)) {
2340                         int alignment = 0;
2341
2342                         trace__fprintf_entry_head(trace, thread, 0, false, ttrace->entry_time, trace->output);
2343                         printed = fprintf(trace->output, "%s)", ttrace->entry_str);
2344                         if (trace->args_alignment > printed)
2345                                 alignment = trace->args_alignment - printed;
2346                         fprintf(trace->output, "%*s= ?\n", alignment, " ");
2347                 }
2348         } else {
2349                 ttrace->entry_pending = true;
2350                 /* See trace__vfs_getname & trace__sys_exit */
2351                 ttrace->filename.pending_open = false;
2352         }
2353
2354         if (trace->current != thread) {
2355                 thread__put(trace->current);
2356                 trace->current = thread__get(thread);
2357         }
2358         err = 0;
2359 out_put:
2360         thread__put(thread);
2361         return err;
2362 }
2363
2364 static int trace__fprintf_sys_enter(struct trace *trace, struct evsel *evsel,
2365                                     struct perf_sample *sample)
2366 {
2367         struct thread_trace *ttrace;
2368         struct thread *thread;
2369         int id = perf_evsel__sc_tp_uint(evsel, id, sample), err = -1;
2370         struct syscall *sc = trace__syscall_info(trace, evsel, id);
2371         char msg[1024];
2372         void *args, *augmented_args = NULL;
2373         int augmented_args_size;
2374
2375         if (sc == NULL)
2376                 return -1;
2377
2378         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2379         ttrace = thread__trace(thread, trace->output);
2380         /*
2381          * We need to get ttrace just to make sure it is there when syscall__scnprintf_args()
2382          * and the rest of the beautifiers accessing it via struct syscall_arg touches it.
2383          */
2384         if (ttrace == NULL)
2385                 goto out_put;
2386
2387         args = perf_evsel__sc_tp_ptr(evsel, args, sample);
2388         augmented_args = syscall__augmented_args(sc, sample, &augmented_args_size, trace->raw_augmented_syscalls_args_size);
2389         syscall__scnprintf_args(sc, msg, sizeof(msg), args, augmented_args, augmented_args_size, trace, thread);
2390         fprintf(trace->output, "%s", msg);
2391         err = 0;
2392 out_put:
2393         thread__put(thread);
2394         return err;
2395 }
2396
2397 static int trace__resolve_callchain(struct trace *trace, struct evsel *evsel,
2398                                     struct perf_sample *sample,
2399                                     struct callchain_cursor *cursor)
2400 {
2401         struct addr_location al;
2402         int max_stack = evsel->core.attr.sample_max_stack ?
2403                         evsel->core.attr.sample_max_stack :
2404                         trace->max_stack;
2405         int err;
2406
2407         if (machine__resolve(trace->host, &al, sample) < 0)
2408                 return -1;
2409
2410         err = thread__resolve_callchain(al.thread, cursor, evsel, sample, NULL, NULL, max_stack);
2411         addr_location__put(&al);
2412         return err;
2413 }
2414
2415 static int trace__fprintf_callchain(struct trace *trace, struct perf_sample *sample)
2416 {
2417         /* TODO: user-configurable print_opts */
2418         const unsigned int print_opts = EVSEL__PRINT_SYM |
2419                                         EVSEL__PRINT_DSO |
2420                                         EVSEL__PRINT_UNKNOWN_AS_ADDR;
2421
2422         return sample__fprintf_callchain(sample, 38, print_opts, &callchain_cursor, symbol_conf.bt_stop_list, trace->output);
2423 }
2424
2425 static const char *errno_to_name(struct evsel *evsel, int err)
2426 {
2427         struct perf_env *env = evsel__env(evsel);
2428         const char *arch_name = perf_env__arch(env);
2429
2430         return arch_syscalls__strerrno(arch_name, err);
2431 }
2432
2433 static int trace__sys_exit(struct trace *trace, struct evsel *evsel,
2434                            union perf_event *event __maybe_unused,
2435                            struct perf_sample *sample)
2436 {
2437         long ret;
2438         u64 duration = 0;
2439         bool duration_calculated = false;
2440         struct thread *thread;
2441         int id = perf_evsel__sc_tp_uint(evsel, id, sample), err = -1, callchain_ret = 0, printed = 0;
2442         int alignment = trace->args_alignment;
2443         struct syscall *sc = trace__syscall_info(trace, evsel, id);
2444         struct thread_trace *ttrace;
2445
2446         if (sc == NULL)
2447                 return -1;
2448
2449         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2450         ttrace = thread__trace(thread, trace->output);
2451         if (ttrace == NULL)
2452                 goto out_put;
2453
2454         trace__fprintf_sample(trace, evsel, sample, thread);
2455
2456         ret = perf_evsel__sc_tp_uint(evsel, ret, sample);
2457
2458         if (trace->summary)
2459                 thread__update_stats(thread, ttrace, id, sample, ret, trace->errno_summary);
2460
2461         if (!trace->fd_path_disabled && sc->is_open && ret >= 0 && ttrace->filename.pending_open) {
2462                 trace__set_fd_pathname(thread, ret, ttrace->filename.name);
2463                 ttrace->filename.pending_open = false;
2464                 ++trace->stats.vfs_getname;
2465         }
2466
2467         if (ttrace->entry_time) {
2468                 duration = sample->time - ttrace->entry_time;
2469                 if (trace__filter_duration(trace, duration))
2470                         goto out;
2471                 duration_calculated = true;
2472         } else if (trace->duration_filter)
2473                 goto out;
2474
2475         if (sample->callchain) {
2476                 callchain_ret = trace__resolve_callchain(trace, evsel, sample, &callchain_cursor);
2477                 if (callchain_ret == 0) {
2478                         if (callchain_cursor.nr < trace->min_stack)
2479                                 goto out;
2480                         callchain_ret = 1;
2481                 }
2482         }
2483
2484         if (trace->summary_only || (ret >= 0 && trace->failure_only))
2485                 goto out;
2486
2487         trace__fprintf_entry_head(trace, thread, duration, duration_calculated, ttrace->entry_time, trace->output);
2488
2489         if (ttrace->entry_pending) {
2490                 printed = fprintf(trace->output, "%s", ttrace->entry_str);
2491         } else {
2492                 printed += fprintf(trace->output, " ... [");
2493                 color_fprintf(trace->output, PERF_COLOR_YELLOW, "continued");
2494                 printed += 9;
2495                 printed += fprintf(trace->output, "]: %s()", sc->name);
2496         }
2497
2498         printed++; /* the closing ')' */
2499
2500         if (alignment > printed)
2501                 alignment -= printed;
2502         else
2503                 alignment = 0;
2504
2505         fprintf(trace->output, ")%*s= ", alignment, " ");
2506
2507         if (sc->fmt == NULL) {
2508                 if (ret < 0)
2509                         goto errno_print;
2510 signed_print:
2511                 fprintf(trace->output, "%ld", ret);
2512         } else if (ret < 0) {
2513 errno_print: {
2514                 char bf[STRERR_BUFSIZE];
2515                 const char *emsg = str_error_r(-ret, bf, sizeof(bf)),
2516                            *e = errno_to_name(evsel, -ret);
2517
2518                 fprintf(trace->output, "-1 %s (%s)", e, emsg);
2519         }
2520         } else if (ret == 0 && sc->fmt->timeout)
2521                 fprintf(trace->output, "0 (Timeout)");
2522         else if (ttrace->ret_scnprintf) {
2523                 char bf[1024];
2524                 struct syscall_arg arg = {
2525                         .val    = ret,
2526                         .thread = thread,
2527                         .trace  = trace,
2528                 };
2529                 ttrace->ret_scnprintf(bf, sizeof(bf), &arg);
2530                 ttrace->ret_scnprintf = NULL;
2531                 fprintf(trace->output, "%s", bf);
2532         } else if (sc->fmt->hexret)
2533                 fprintf(trace->output, "%#lx", ret);
2534         else if (sc->fmt->errpid) {
2535                 struct thread *child = machine__find_thread(trace->host, ret, ret);
2536
2537                 if (child != NULL) {
2538                         fprintf(trace->output, "%ld", ret);
2539                         if (child->comm_set)
2540                                 fprintf(trace->output, " (%s)", thread__comm_str(child));
2541                         thread__put(child);
2542                 }
2543         } else
2544                 goto signed_print;
2545
2546         fputc('\n', trace->output);
2547
2548         /*
2549          * We only consider an 'event' for the sake of --max-events a non-filtered
2550          * sys_enter + sys_exit and other tracepoint events.
2551          */
2552         if (++trace->nr_events_printed == trace->max_events && trace->max_events != ULONG_MAX)
2553                 interrupted = true;
2554
2555         if (callchain_ret > 0)
2556                 trace__fprintf_callchain(trace, sample);
2557         else if (callchain_ret < 0)
2558                 pr_err("Problem processing %s callchain, skipping...\n", evsel__name(evsel));
2559 out:
2560         ttrace->entry_pending = false;
2561         err = 0;
2562 out_put:
2563         thread__put(thread);
2564         return err;
2565 }
2566
2567 static int trace__vfs_getname(struct trace *trace, struct evsel *evsel,
2568                               union perf_event *event __maybe_unused,
2569                               struct perf_sample *sample)
2570 {
2571         struct thread *thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2572         struct thread_trace *ttrace;
2573         size_t filename_len, entry_str_len, to_move;
2574         ssize_t remaining_space;
2575         char *pos;
2576         const char *filename = evsel__rawptr(evsel, sample, "pathname");
2577
2578         if (!thread)
2579                 goto out;
2580
2581         ttrace = thread__priv(thread);
2582         if (!ttrace)
2583                 goto out_put;
2584
2585         filename_len = strlen(filename);
2586         if (filename_len == 0)
2587                 goto out_put;
2588
2589         if (ttrace->filename.namelen < filename_len) {
2590                 char *f = realloc(ttrace->filename.name, filename_len + 1);
2591
2592                 if (f == NULL)
2593                         goto out_put;
2594
2595                 ttrace->filename.namelen = filename_len;
2596                 ttrace->filename.name = f;
2597         }
2598
2599         strcpy(ttrace->filename.name, filename);
2600         ttrace->filename.pending_open = true;
2601
2602         if (!ttrace->filename.ptr)
2603                 goto out_put;
2604
2605         entry_str_len = strlen(ttrace->entry_str);
2606         remaining_space = trace__entry_str_size - entry_str_len - 1; /* \0 */
2607         if (remaining_space <= 0)
2608                 goto out_put;
2609
2610         if (filename_len > (size_t)remaining_space) {
2611                 filename += filename_len - remaining_space;
2612                 filename_len = remaining_space;
2613         }
2614
2615         to_move = entry_str_len - ttrace->filename.entry_str_pos + 1; /* \0 */
2616         pos = ttrace->entry_str + ttrace->filename.entry_str_pos;
2617         memmove(pos + filename_len, pos, to_move);
2618         memcpy(pos, filename, filename_len);
2619
2620         ttrace->filename.ptr = 0;
2621         ttrace->filename.entry_str_pos = 0;
2622 out_put:
2623         thread__put(thread);
2624 out:
2625         return 0;
2626 }
2627
2628 static int trace__sched_stat_runtime(struct trace *trace, struct evsel *evsel,
2629                                      union perf_event *event __maybe_unused,
2630                                      struct perf_sample *sample)
2631 {
2632         u64 runtime = evsel__intval(evsel, sample, "runtime");
2633         double runtime_ms = (double)runtime / NSEC_PER_MSEC;
2634         struct thread *thread = machine__findnew_thread(trace->host,
2635                                                         sample->pid,
2636                                                         sample->tid);
2637         struct thread_trace *ttrace = thread__trace(thread, trace->output);
2638
2639         if (ttrace == NULL)
2640                 goto out_dump;
2641
2642         ttrace->runtime_ms += runtime_ms;
2643         trace->runtime_ms += runtime_ms;
2644 out_put:
2645         thread__put(thread);
2646         return 0;
2647
2648 out_dump:
2649         fprintf(trace->output, "%s: comm=%s,pid=%u,runtime=%" PRIu64 ",vruntime=%" PRIu64 ")\n",
2650                evsel->name,
2651                evsel__strval(evsel, sample, "comm"),
2652                (pid_t)evsel__intval(evsel, sample, "pid"),
2653                runtime,
2654                evsel__intval(evsel, sample, "vruntime"));
2655         goto out_put;
2656 }
2657
2658 static int bpf_output__printer(enum binary_printer_ops op,
2659                                unsigned int val, void *extra __maybe_unused, FILE *fp)
2660 {
2661         unsigned char ch = (unsigned char)val;
2662
2663         switch (op) {
2664         case BINARY_PRINT_CHAR_DATA:
2665                 return fprintf(fp, "%c", isprint(ch) ? ch : '.');
2666         case BINARY_PRINT_DATA_BEGIN:
2667         case BINARY_PRINT_LINE_BEGIN:
2668         case BINARY_PRINT_ADDR:
2669         case BINARY_PRINT_NUM_DATA:
2670         case BINARY_PRINT_NUM_PAD:
2671         case BINARY_PRINT_SEP:
2672         case BINARY_PRINT_CHAR_PAD:
2673         case BINARY_PRINT_LINE_END:
2674         case BINARY_PRINT_DATA_END:
2675         default:
2676                 break;
2677         }
2678
2679         return 0;
2680 }
2681
2682 static void bpf_output__fprintf(struct trace *trace,
2683                                 struct perf_sample *sample)
2684 {
2685         binary__fprintf(sample->raw_data, sample->raw_size, 8,
2686                         bpf_output__printer, NULL, trace->output);
2687         ++trace->nr_events_printed;
2688 }
2689
2690 static size_t trace__fprintf_tp_fields(struct trace *trace, struct evsel *evsel, struct perf_sample *sample,
2691                                        struct thread *thread, void *augmented_args, int augmented_args_size)
2692 {
2693         char bf[2048];
2694         size_t size = sizeof(bf);
2695         struct tep_format_field *field = evsel->tp_format->format.fields;
2696         struct syscall_arg_fmt *arg = __evsel__syscall_arg_fmt(evsel);
2697         size_t printed = 0;
2698         unsigned long val;
2699         u8 bit = 1;
2700         struct syscall_arg syscall_arg = {
2701                 .augmented = {
2702                         .size = augmented_args_size,
2703                         .args = augmented_args,
2704                 },
2705                 .idx    = 0,
2706                 .mask   = 0,
2707                 .trace  = trace,
2708                 .thread = thread,
2709                 .show_string_prefix = trace->show_string_prefix,
2710         };
2711
2712         for (; field && arg; field = field->next, ++syscall_arg.idx, bit <<= 1, ++arg) {
2713                 if (syscall_arg.mask & bit)
2714                         continue;
2715
2716                 syscall_arg.len = 0;
2717                 syscall_arg.fmt = arg;
2718                 if (field->flags & TEP_FIELD_IS_ARRAY) {
2719                         int offset = field->offset;
2720
2721                         if (field->flags & TEP_FIELD_IS_DYNAMIC) {
2722                                 offset = format_field__intval(field, sample, evsel->needs_swap);
2723                                 syscall_arg.len = offset >> 16;
2724                                 offset &= 0xffff;
2725                         }
2726
2727                         val = (uintptr_t)(sample->raw_data + offset);
2728                 } else
2729                         val = format_field__intval(field, sample, evsel->needs_swap);
2730                 /*
2731                  * Some syscall args need some mask, most don't and
2732                  * return val untouched.
2733                  */
2734                 val = syscall_arg_fmt__mask_val(arg, &syscall_arg, val);
2735
2736                 /*
2737                  * Suppress this argument if its value is zero and
2738                  * and we don't have a string associated in an
2739                  * strarray for it.
2740                  */
2741                 if (val == 0 &&
2742                     !trace->show_zeros &&
2743                     !((arg->show_zero ||
2744                        arg->scnprintf == SCA_STRARRAY ||
2745                        arg->scnprintf == SCA_STRARRAYS) &&
2746                       arg->parm))
2747                         continue;
2748
2749                 printed += scnprintf(bf + printed, size - printed, "%s", printed ? ", " : "");
2750
2751                 /*
2752                  * XXX Perhaps we should have a show_tp_arg_names,
2753                  * leaving show_arg_names just for syscalls?
2754                  */
2755                 if (1 || trace->show_arg_names)
2756                         printed += scnprintf(bf + printed, size - printed, "%s: ", field->name);
2757
2758                 printed += syscall_arg_fmt__scnprintf_val(arg, bf + printed, size - printed, &syscall_arg, val);
2759         }
2760
2761         return printed + fprintf(trace->output, "%s", bf);
2762 }
2763
2764 static int trace__event_handler(struct trace *trace, struct evsel *evsel,
2765                                 union perf_event *event __maybe_unused,
2766                                 struct perf_sample *sample)
2767 {
2768         struct thread *thread;
2769         int callchain_ret = 0;
2770         /*
2771          * Check if we called perf_evsel__disable(evsel) due to, for instance,
2772          * this event's max_events having been hit and this is an entry coming
2773          * from the ring buffer that we should discard, since the max events
2774          * have already been considered/printed.
2775          */
2776         if (evsel->disabled)
2777                 return 0;
2778
2779         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2780
2781         if (sample->callchain) {
2782                 callchain_ret = trace__resolve_callchain(trace, evsel, sample, &callchain_cursor);
2783                 if (callchain_ret == 0) {
2784                         if (callchain_cursor.nr < trace->min_stack)
2785                                 goto out;
2786                         callchain_ret = 1;
2787                 }
2788         }
2789
2790         trace__printf_interrupted_entry(trace);
2791         trace__fprintf_tstamp(trace, sample->time, trace->output);
2792
2793         if (trace->trace_syscalls && trace->show_duration)
2794                 fprintf(trace->output, "(         ): ");
2795
2796         if (thread)
2797                 trace__fprintf_comm_tid(trace, thread, trace->output);
2798
2799         if (evsel == trace->syscalls.events.augmented) {
2800                 int id = perf_evsel__sc_tp_uint(evsel, id, sample);
2801                 struct syscall *sc = trace__syscall_info(trace, evsel, id);
2802
2803                 if (sc) {
2804                         fprintf(trace->output, "%s(", sc->name);
2805                         trace__fprintf_sys_enter(trace, evsel, sample);
2806                         fputc(')', trace->output);
2807                         goto newline;
2808                 }
2809
2810                 /*
2811                  * XXX: Not having the associated syscall info or not finding/adding
2812                  *      the thread should never happen, but if it does...
2813                  *      fall thru and print it as a bpf_output event.
2814                  */
2815         }
2816
2817         fprintf(trace->output, "%s(", evsel->name);
2818
2819         if (evsel__is_bpf_output(evsel)) {
2820                 bpf_output__fprintf(trace, sample);
2821         } else if (evsel->tp_format) {
2822                 if (strncmp(evsel->tp_format->name, "sys_enter_", 10) ||
2823                     trace__fprintf_sys_enter(trace, evsel, sample)) {
2824                         if (trace->libtraceevent_print) {
2825                                 event_format__fprintf(evsel->tp_format, sample->cpu,
2826                                                       sample->raw_data, sample->raw_size,
2827                                                       trace->output);
2828                         } else {
2829                                 trace__fprintf_tp_fields(trace, evsel, sample, thread, NULL, 0);
2830                         }
2831                 }
2832         }
2833
2834 newline:
2835         fprintf(trace->output, ")\n");
2836
2837         if (callchain_ret > 0)
2838                 trace__fprintf_callchain(trace, sample);
2839         else if (callchain_ret < 0)
2840                 pr_err("Problem processing %s callchain, skipping...\n", evsel__name(evsel));
2841
2842         ++trace->nr_events_printed;
2843
2844         if (evsel->max_events != ULONG_MAX && ++evsel->nr_events_printed == evsel->max_events) {
2845                 evsel__disable(evsel);
2846                 evsel__close(evsel);
2847         }
2848 out:
2849         thread__put(thread);
2850         return 0;
2851 }
2852
2853 static void print_location(FILE *f, struct perf_sample *sample,
2854                            struct addr_location *al,
2855                            bool print_dso, bool print_sym)
2856 {
2857
2858         if ((verbose > 0 || print_dso) && al->map)
2859                 fprintf(f, "%s@", al->map->dso->long_name);
2860
2861         if ((verbose > 0 || print_sym) && al->sym)
2862                 fprintf(f, "%s+0x%" PRIx64, al->sym->name,
2863                         al->addr - al->sym->start);
2864         else if (al->map)
2865                 fprintf(f, "0x%" PRIx64, al->addr);
2866         else
2867                 fprintf(f, "0x%" PRIx64, sample->addr);
2868 }
2869
2870 static int trace__pgfault(struct trace *trace,
2871                           struct evsel *evsel,
2872                           union perf_event *event __maybe_unused,
2873                           struct perf_sample *sample)
2874 {
2875         struct thread *thread;
2876         struct addr_location al;
2877         char map_type = 'd';
2878         struct thread_trace *ttrace;
2879         int err = -1;
2880         int callchain_ret = 0;
2881
2882         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2883
2884         if (sample->callchain) {
2885                 callchain_ret = trace__resolve_callchain(trace, evsel, sample, &callchain_cursor);
2886                 if (callchain_ret == 0) {
2887                         if (callchain_cursor.nr < trace->min_stack)
2888                                 goto out_put;
2889                         callchain_ret = 1;
2890                 }
2891         }
2892
2893         ttrace = thread__trace(thread, trace->output);
2894         if (ttrace == NULL)
2895                 goto out_put;
2896
2897         if (evsel->core.attr.config == PERF_COUNT_SW_PAGE_FAULTS_MAJ)
2898                 ttrace->pfmaj++;
2899         else
2900                 ttrace->pfmin++;
2901
2902         if (trace->summary_only)
2903                 goto out;
2904
2905         thread__find_symbol(thread, sample->cpumode, sample->ip, &al);
2906
2907         trace__fprintf_entry_head(trace, thread, 0, true, sample->time, trace->output);
2908
2909         fprintf(trace->output, "%sfault [",
2910                 evsel->core.attr.config == PERF_COUNT_SW_PAGE_FAULTS_MAJ ?
2911                 "maj" : "min");
2912
2913         print_location(trace->output, sample, &al, false, true);
2914
2915         fprintf(trace->output, "] => ");
2916
2917         thread__find_symbol(thread, sample->cpumode, sample->addr, &al);
2918
2919         if (!al.map) {
2920                 thread__find_symbol(thread, sample->cpumode, sample->addr, &al);
2921
2922                 if (al.map)
2923                         map_type = 'x';
2924                 else
2925                         map_type = '?';
2926         }
2927
2928         print_location(trace->output, sample, &al, true, false);
2929
2930         fprintf(trace->output, " (%c%c)\n", map_type, al.level);
2931
2932         if (callchain_ret > 0)
2933                 trace__fprintf_callchain(trace, sample);
2934         else if (callchain_ret < 0)
2935                 pr_err("Problem processing %s callchain, skipping...\n", evsel__name(evsel));
2936
2937         ++trace->nr_events_printed;
2938 out:
2939         err = 0;
2940 out_put:
2941         thread__put(thread);
2942         return err;
2943 }
2944
2945 static void trace__set_base_time(struct trace *trace,
2946                                  struct evsel *evsel,
2947                                  struct perf_sample *sample)
2948 {
2949         /*
2950          * BPF events were not setting PERF_SAMPLE_TIME, so be more robust
2951          * and don't use sample->time unconditionally, we may end up having
2952          * some other event in the future without PERF_SAMPLE_TIME for good
2953          * reason, i.e. we may not be interested in its timestamps, just in
2954          * it taking place, picking some piece of information when it
2955          * appears in our event stream (vfs_getname comes to mind).
2956          */
2957         if (trace->base_time == 0 && !trace->full_time &&
2958             (evsel->core.attr.sample_type & PERF_SAMPLE_TIME))
2959                 trace->base_time = sample->time;
2960 }
2961
2962 static int trace__process_sample(struct perf_tool *tool,
2963                                  union perf_event *event,
2964                                  struct perf_sample *sample,
2965                                  struct evsel *evsel,
2966                                  struct machine *machine __maybe_unused)
2967 {
2968         struct trace *trace = container_of(tool, struct trace, tool);
2969         struct thread *thread;
2970         int err = 0;
2971
2972         tracepoint_handler handler = evsel->handler;
2973
2974         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2975         if (thread && thread__is_filtered(thread))
2976                 goto out;
2977
2978         trace__set_base_time(trace, evsel, sample);
2979
2980         if (handler) {
2981                 ++trace->nr_events;
2982                 handler(trace, evsel, event, sample);
2983         }
2984 out:
2985         thread__put(thread);
2986         return err;
2987 }
2988
2989 static int trace__record(struct trace *trace, int argc, const char **argv)
2990 {
2991         unsigned int rec_argc, i, j;
2992         const char **rec_argv;
2993         const char * const record_args[] = {
2994                 "record",
2995                 "-R",
2996                 "-m", "1024",
2997                 "-c", "1",
2998         };
2999         pid_t pid = getpid();
3000         char *filter = asprintf__tp_filter_pids(1, &pid);
3001         const char * const sc_args[] = { "-e", };
3002         unsigned int sc_args_nr = ARRAY_SIZE(sc_args);
3003         const char * const majpf_args[] = { "-e", "major-faults" };
3004         unsigned int majpf_args_nr = ARRAY_SIZE(majpf_args);
3005         const char * const minpf_args[] = { "-e", "minor-faults" };
3006         unsigned int minpf_args_nr = ARRAY_SIZE(minpf_args);
3007         int err = -1;
3008
3009         /* +3 is for the event string below and the pid filter */
3010         rec_argc = ARRAY_SIZE(record_args) + sc_args_nr + 3 +
3011                 majpf_args_nr + minpf_args_nr + argc;
3012         rec_argv = calloc(rec_argc + 1, sizeof(char *));
3013
3014         if (rec_argv == NULL || filter == NULL)
3015                 goto out_free;
3016
3017         j = 0;
3018         for (i = 0; i < ARRAY_SIZE(record_args); i++)
3019                 rec_argv[j++] = record_args[i];
3020
3021         if (trace->trace_syscalls) {
3022                 for (i = 0; i < sc_args_nr; i++)
3023                         rec_argv[j++] = sc_args[i];
3024
3025                 /* event string may be different for older kernels - e.g., RHEL6 */
3026                 if (is_valid_tracepoint("raw_syscalls:sys_enter"))
3027                         rec_argv[j++] = "raw_syscalls:sys_enter,raw_syscalls:sys_exit";
3028                 else if (is_valid_tracepoint("syscalls:sys_enter"))
3029                         rec_argv[j++] = "syscalls:sys_enter,syscalls:sys_exit";
3030                 else {
3031                         pr_err("Neither raw_syscalls nor syscalls events exist.\n");
3032                         goto out_free;
3033                 }
3034         }
3035
3036         rec_argv[j++] = "--filter";
3037         rec_argv[j++] = filter;
3038
3039         if (trace->trace_pgfaults & TRACE_PFMAJ)
3040                 for (i = 0; i < majpf_args_nr; i++)
3041                         rec_argv[j++] = majpf_args[i];
3042
3043         if (trace->trace_pgfaults & TRACE_PFMIN)
3044                 for (i = 0; i < minpf_args_nr; i++)
3045                         rec_argv[j++] = minpf_args[i];
3046
3047         for (i = 0; i < (unsigned int)argc; i++)
3048                 rec_argv[j++] = argv[i];
3049
3050         err = cmd_record(j, rec_argv);
3051 out_free:
3052         free(filter);
3053         free(rec_argv);
3054         return err;
3055 }
3056
3057 static size_t trace__fprintf_thread_summary(struct trace *trace, FILE *fp);
3058
3059 static bool evlist__add_vfs_getname(struct evlist *evlist)
3060 {
3061         bool found = false;
3062         struct evsel *evsel, *tmp;
3063         struct parse_events_error err;
3064         int ret;
3065
3066         bzero(&err, sizeof(err));
3067         ret = parse_events(evlist, "probe:vfs_getname*", &err);
3068         if (ret) {
3069                 free(err.str);
3070                 free(err.help);
3071                 free(err.first_str);
3072                 free(err.first_help);
3073                 return false;
3074         }
3075
3076         evlist__for_each_entry_safe(evlist, evsel, tmp) {
3077                 if (!strstarts(evsel__name(evsel), "probe:vfs_getname"))
3078                         continue;
3079
3080                 if (evsel__field(evsel, "pathname")) {
3081                         evsel->handler = trace__vfs_getname;
3082                         found = true;
3083                         continue;
3084                 }
3085
3086                 list_del_init(&evsel->core.node);
3087                 evsel->evlist = NULL;
3088                 evsel__delete(evsel);
3089         }
3090
3091         return found;
3092 }
3093
3094 static struct evsel *evsel__new_pgfault(u64 config)
3095 {
3096         struct evsel *evsel;
3097         struct perf_event_attr attr = {
3098                 .type = PERF_TYPE_SOFTWARE,
3099                 .mmap_data = 1,
3100         };
3101
3102         attr.config = config;
3103         attr.sample_period = 1;
3104
3105         event_attr_init(&attr);
3106
3107         evsel = evsel__new(&attr);
3108         if (evsel)
3109                 evsel->handler = trace__pgfault;
3110
3111         return evsel;
3112 }
3113
3114 static void evlist__free_syscall_tp_fields(struct evlist *evlist)
3115 {
3116         struct evsel *evsel;
3117
3118         evlist__for_each_entry(evlist, evsel) {
3119                 struct evsel_trace *et = evsel->priv;
3120
3121                 if (!et || !evsel->tp_format || strcmp(evsel->tp_format->system, "syscalls"))
3122                         continue;
3123
3124                 free(et->fmt);
3125                 free(et);
3126         }
3127 }
3128
3129 static void trace__handle_event(struct trace *trace, union perf_event *event, struct perf_sample *sample)
3130 {
3131         const u32 type = event->header.type;
3132         struct evsel *evsel;
3133
3134         if (type != PERF_RECORD_SAMPLE) {
3135                 trace__process_event(trace, trace->host, event, sample);
3136                 return;
3137         }
3138
3139         evsel = evlist__id2evsel(trace->evlist, sample->id);
3140         if (evsel == NULL) {
3141                 fprintf(trace->output, "Unknown tp ID %" PRIu64 ", skipping...\n", sample->id);
3142                 return;
3143         }
3144
3145         if (evswitch__discard(&trace->evswitch, evsel))
3146                 return;
3147
3148         trace__set_base_time(trace, evsel, sample);
3149
3150         if (evsel->core.attr.type == PERF_TYPE_TRACEPOINT &&
3151             sample->raw_data == NULL) {
3152                 fprintf(trace->output, "%s sample with no payload for tid: %d, cpu %d, raw_size=%d, skipping...\n",
3153                        evsel__name(evsel), sample->tid,
3154                        sample->cpu, sample->raw_size);
3155         } else {
3156                 tracepoint_handler handler = evsel->handler;
3157                 handler(trace, evsel, event, sample);
3158         }
3159
3160         if (trace->nr_events_printed >= trace->max_events && trace->max_events != ULONG_MAX)
3161                 interrupted = true;
3162 }
3163
3164 static int trace__add_syscall_newtp(struct trace *trace)
3165 {
3166         int ret = -1;
3167         struct evlist *evlist = trace->evlist;
3168         struct evsel *sys_enter, *sys_exit;
3169
3170         sys_enter = perf_evsel__raw_syscall_newtp("sys_enter", trace__sys_enter);
3171         if (sys_enter == NULL)
3172                 goto out;
3173
3174         if (perf_evsel__init_sc_tp_ptr_field(sys_enter, args))
3175                 goto out_delete_sys_enter;
3176
3177         sys_exit = perf_evsel__raw_syscall_newtp("sys_exit", trace__sys_exit);
3178         if (sys_exit == NULL)
3179                 goto out_delete_sys_enter;
3180
3181         if (perf_evsel__init_sc_tp_uint_field(sys_exit, ret))
3182                 goto out_delete_sys_exit;
3183
3184         evsel__config_callchain(sys_enter, &trace->opts, &callchain_param);
3185         evsel__config_callchain(sys_exit, &trace->opts, &callchain_param);
3186
3187         evlist__add(evlist, sys_enter);
3188         evlist__add(evlist, sys_exit);
3189
3190         if (callchain_param.enabled && !trace->kernel_syscallchains) {
3191                 /*
3192                  * We're interested only in the user space callchain
3193                  * leading to the syscall, allow overriding that for
3194                  * debugging reasons using --kernel_syscall_callchains
3195                  */
3196                 sys_exit->core.attr.exclude_callchain_kernel = 1;
3197         }
3198
3199         trace->syscalls.events.sys_enter = sys_enter;
3200         trace->syscalls.events.sys_exit  = sys_exit;
3201
3202         ret = 0;
3203 out:
3204         return ret;
3205
3206 out_delete_sys_exit:
3207         evsel__delete_priv(sys_exit);
3208 out_delete_sys_enter:
3209         evsel__delete_priv(sys_enter);
3210         goto out;
3211 }
3212
3213 static int trace__set_ev_qualifier_tp_filter(struct trace *trace)
3214 {
3215         int err = -1;
3216         struct evsel *sys_exit;
3217         char *filter = asprintf_expr_inout_ints("id", !trace->not_ev_qualifier,
3218                                                 trace->ev_qualifier_ids.nr,
3219                                                 trace->ev_qualifier_ids.entries);
3220
3221         if (filter == NULL)
3222                 goto out_enomem;
3223
3224         if (!evsel__append_tp_filter(trace->syscalls.events.sys_enter, filter)) {
3225                 sys_exit = trace->syscalls.events.sys_exit;
3226                 err = evsel__append_tp_filter(sys_exit, filter);
3227         }
3228
3229         free(filter);
3230 out:
3231         return err;
3232 out_enomem:
3233         errno = ENOMEM;
3234         goto out;
3235 }
3236
3237 #ifdef HAVE_LIBBPF_SUPPORT
3238 static struct bpf_map *trace__find_bpf_map_by_name(struct trace *trace, const char *name)
3239 {
3240         if (trace->bpf_obj == NULL)
3241                 return NULL;
3242
3243         return bpf_object__find_map_by_name(trace->bpf_obj, name);
3244 }
3245
3246 static void trace__set_bpf_map_filtered_pids(struct trace *trace)
3247 {
3248         trace->filter_pids.map = trace__find_bpf_map_by_name(trace, "pids_filtered");
3249 }
3250
3251 static void trace__set_bpf_map_syscalls(struct trace *trace)
3252 {
3253         trace->syscalls.map = trace__find_bpf_map_by_name(trace, "syscalls");
3254         trace->syscalls.prog_array.sys_enter = trace__find_bpf_map_by_name(trace, "syscalls_sys_enter");
3255         trace->syscalls.prog_array.sys_exit  = trace__find_bpf_map_by_name(trace, "syscalls_sys_exit");
3256 }
3257
3258 static struct bpf_program *trace__find_bpf_program_by_title(struct trace *trace, const char *name)
3259 {
3260         if (trace->bpf_obj == NULL)
3261                 return NULL;
3262
3263         return bpf_object__find_program_by_title(trace->bpf_obj, name);
3264 }
3265
3266 static struct bpf_program *trace__find_syscall_bpf_prog(struct trace *trace, struct syscall *sc,
3267                                                         const char *prog_name, const char *type)
3268 {
3269         struct bpf_program *prog;
3270
3271         if (prog_name == NULL) {
3272                 char default_prog_name[256];
3273                 scnprintf(default_prog_name, sizeof(default_prog_name), "!syscalls:sys_%s_%s", type, sc->name);
3274                 prog = trace__find_bpf_program_by_title(trace, default_prog_name);
3275                 if (prog != NULL)
3276                         goto out_found;
3277                 if (sc->fmt && sc->fmt->alias) {
3278                         scnprintf(default_prog_name, sizeof(default_prog_name), "!syscalls:sys_%s_%s", type, sc->fmt->alias);
3279                         prog = trace__find_bpf_program_by_title(trace, default_prog_name);
3280                         if (prog != NULL)
3281                                 goto out_found;
3282                 }
3283                 goto out_unaugmented;
3284         }
3285
3286         prog = trace__find_bpf_program_by_title(trace, prog_name);
3287
3288         if (prog != NULL) {
3289 out_found:
3290                 return prog;
3291         }
3292
3293         pr_debug("Couldn't find BPF prog \"%s\" to associate with syscalls:sys_%s_%s, not augmenting it\n",
3294                  prog_name, type, sc->name);
3295 out_unaugmented:
3296         return trace->syscalls.unaugmented_prog;
3297 }
3298
3299 static void trace__init_syscall_bpf_progs(struct trace *trace, int id)
3300 {
3301         struct syscall *sc = trace__syscall_info(trace, NULL, id);
3302
3303         if (sc == NULL)
3304                 return;
3305
3306         sc->bpf_prog.sys_enter = trace__find_syscall_bpf_prog(trace, sc, sc->fmt ? sc->fmt->bpf_prog_name.sys_enter : NULL, "enter");
3307         sc->bpf_prog.sys_exit  = trace__find_syscall_bpf_prog(trace, sc, sc->fmt ? sc->fmt->bpf_prog_name.sys_exit  : NULL,  "exit");
3308 }
3309
3310 static int trace__bpf_prog_sys_enter_fd(struct trace *trace, int id)
3311 {
3312         struct syscall *sc = trace__syscall_info(trace, NULL, id);
3313         return sc ? bpf_program__fd(sc->bpf_prog.sys_enter) : bpf_program__fd(trace->syscalls.unaugmented_prog);
3314 }
3315
3316 static int trace__bpf_prog_sys_exit_fd(struct trace *trace, int id)
3317 {
3318         struct syscall *sc = trace__syscall_info(trace, NULL, id);
3319         return sc ? bpf_program__fd(sc->bpf_prog.sys_exit) : bpf_program__fd(trace->syscalls.unaugmented_prog);
3320 }
3321
3322 static void trace__init_bpf_map_syscall_args(struct trace *trace, int id, struct bpf_map_syscall_entry *entry)
3323 {
3324         struct syscall *sc = trace__syscall_info(trace, NULL, id);
3325         int arg = 0;
3326
3327         if (sc == NULL)
3328                 goto out;
3329
3330         for (; arg < sc->nr_args; ++arg) {
3331                 entry->string_args_len[arg] = 0;
3332                 if (sc->arg_fmt[arg].scnprintf == SCA_FILENAME) {
3333                         /* Should be set like strace -s strsize */
3334                         entry->string_args_len[arg] = PATH_MAX;
3335                 }
3336         }
3337 out:
3338         for (; arg < 6; ++arg)
3339                 entry->string_args_len[arg] = 0;
3340 }
3341 static int trace__set_ev_qualifier_bpf_filter(struct trace *trace)
3342 {
3343         int fd = bpf_map__fd(trace->syscalls.map);
3344         struct bpf_map_syscall_entry value = {
3345                 .enabled = !trace->not_ev_qualifier,
3346         };
3347         int err = 0;
3348         size_t i;
3349
3350         for (i = 0; i < trace->ev_qualifier_ids.nr; ++i) {
3351                 int key = trace->ev_qualifier_ids.entries[i];
3352
3353                 if (value.enabled) {
3354                         trace__init_bpf_map_syscall_args(trace, key, &value);
3355                         trace__init_syscall_bpf_progs(trace, key);
3356                 }
3357
3358                 err = bpf_map_update_elem(fd, &key, &value, BPF_EXIST);
3359                 if (err)
3360                         break;
3361         }
3362
3363         return err;
3364 }
3365
3366 static int __trace__init_syscalls_bpf_map(struct trace *trace, bool enabled)
3367 {
3368         int fd = bpf_map__fd(trace->syscalls.map);
3369         struct bpf_map_syscall_entry value = {
3370                 .enabled = enabled,
3371         };
3372         int err = 0, key;
3373
3374         for (key = 0; key < trace->sctbl->syscalls.nr_entries; ++key) {
3375                 if (enabled)
3376                         trace__init_bpf_map_syscall_args(trace, key, &value);
3377
3378                 err = bpf_map_update_elem(fd, &key, &value, BPF_ANY);
3379                 if (err)
3380                         break;
3381         }
3382
3383         return err;
3384 }
3385
3386 static int trace__init_syscalls_bpf_map(struct trace *trace)
3387 {
3388         bool enabled = true;
3389
3390         if (trace->ev_qualifier_ids.nr)
3391                 enabled = trace->not_ev_qualifier;
3392
3393         return __trace__init_syscalls_bpf_map(trace, enabled);
3394 }
3395
3396 static struct bpf_program *trace__find_usable_bpf_prog_entry(struct trace *trace, struct syscall *sc)
3397 {
3398         struct tep_format_field *field, *candidate_field;
3399         int id;
3400
3401         /*
3402          * We're only interested in syscalls that have a pointer:
3403          */
3404         for (field = sc->args; field; field = field->next) {
3405                 if (field->flags & TEP_FIELD_IS_POINTER)
3406                         goto try_to_find_pair;
3407         }
3408
3409         return NULL;
3410
3411 try_to_find_pair:
3412         for (id = 0; id < trace->sctbl->syscalls.nr_entries; ++id) {
3413                 struct syscall *pair = trace__syscall_info(trace, NULL, id);
3414                 struct bpf_program *pair_prog;
3415                 bool is_candidate = false;
3416
3417                 if (pair == NULL || pair == sc ||
3418                     pair->bpf_prog.sys_enter == trace->syscalls.unaugmented_prog)
3419                         continue;
3420
3421                 for (field = sc->args, candidate_field = pair->args;
3422                      field && candidate_field; field = field->next, candidate_field = candidate_field->next) {
3423                         bool is_pointer = field->flags & TEP_FIELD_IS_POINTER,
3424                              candidate_is_pointer = candidate_field->flags & TEP_FIELD_IS_POINTER;
3425
3426                         if (is_pointer) {
3427                                if (!candidate_is_pointer) {
3428                                         // The candidate just doesn't copies our pointer arg, might copy other pointers we want.
3429                                         continue;
3430                                }
3431                         } else {
3432                                 if (candidate_is_pointer) {
3433                                         // The candidate might copy a pointer we don't have, skip it.
3434                                         goto next_candidate;
3435                                 }
3436                                 continue;
3437                         }
3438
3439                         if (strcmp(field->type, candidate_field->type))
3440                                 goto next_candidate;
3441
3442                         is_candidate = true;
3443                 }
3444
3445                 if (!is_candidate)
3446                         goto next_candidate;
3447
3448                 /*
3449                  * Check if the tentative pair syscall augmenter has more pointers, if it has,
3450                  * then it may be collecting that and we then can't use it, as it would collect
3451                  * more than what is common to the two syscalls.
3452                  */
3453                 if (candidate_field) {
3454                         for (candidate_field = candidate_field->next; candidate_field; candidate_field = candidate_field->next)
3455                                 if (candidate_field->flags & TEP_FIELD_IS_POINTER)
3456                                         goto next_candidate;
3457                 }
3458
3459                 pair_prog = pair->bpf_prog.sys_enter;
3460                 /*
3461                  * If the pair isn't enabled, then its bpf_prog.sys_enter will not
3462                  * have been searched for, so search it here and if it returns the
3463                  * unaugmented one, then ignore it, otherwise we'll reuse that BPF
3464                  * program for a filtered syscall on a non-filtered one.
3465                  *
3466                  * For instance, we have "!syscalls:sys_enter_renameat" and that is
3467                  * useful for "renameat2".
3468                  */
3469                 if (pair_prog == NULL) {
3470                         pair_prog = trace__find_syscall_bpf_prog(trace, pair, pair->fmt ? pair->fmt->bpf_prog_name.sys_enter : NULL, "enter");
3471                         if (pair_prog == trace->syscalls.unaugmented_prog)
3472                                 goto next_candidate;
3473                 }
3474
3475                 pr_debug("Reusing \"%s\" BPF sys_enter augmenter for \"%s\"\n", pair->name, sc->name);
3476                 return pair_prog;
3477         next_candidate:
3478                 continue;
3479         }
3480
3481         return NULL;
3482 }
3483
3484 static int trace__init_syscalls_bpf_prog_array_maps(struct trace *trace)
3485 {
3486         int map_enter_fd = bpf_map__fd(trace->syscalls.prog_array.sys_enter),
3487             map_exit_fd  = bpf_map__fd(trace->syscalls.prog_array.sys_exit);
3488         int err = 0, key;
3489
3490         for (key = 0; key < trace->sctbl->syscalls.nr_entries; ++key) {
3491                 int prog_fd;
3492
3493                 if (!trace__syscall_enabled(trace, key))
3494                         continue;
3495
3496                 trace__init_syscall_bpf_progs(trace, key);
3497
3498                 // It'll get at least the "!raw_syscalls:unaugmented"
3499                 prog_fd = trace__bpf_prog_sys_enter_fd(trace, key);
3500                 err = bpf_map_update_elem(map_enter_fd, &key, &prog_fd, BPF_ANY);
3501                 if (err)
3502                         break;
3503                 prog_fd = trace__bpf_prog_sys_exit_fd(trace, key);
3504                 err = bpf_map_update_elem(map_exit_fd, &key, &prog_fd, BPF_ANY);
3505                 if (err)
3506                         break;
3507         }
3508
3509         /*
3510          * Now lets do a second pass looking for enabled syscalls without
3511          * an augmenter that have a signature that is a superset of another
3512          * syscall with an augmenter so that we can auto-reuse it.
3513          *
3514          * I.e. if we have an augmenter for the "open" syscall that has
3515          * this signature:
3516          *
3517          *   int open(const char *pathname, int flags, mode_t mode);
3518          *
3519          * I.e. that will collect just the first string argument, then we
3520          * can reuse it for the 'creat' syscall, that has this signature:
3521          *
3522          *   int creat(const char *pathname, mode_t mode);
3523          *
3524          * and for:
3525          *
3526          *   int stat(const char *pathname, struct stat *statbuf);
3527          *   int lstat(const char *pathname, struct stat *statbuf);
3528          *
3529          * Because the 'open' augmenter will collect the first arg as a string,
3530          * and leave alone all the other args, which already helps with
3531          * beautifying 'stat' and 'lstat''s pathname arg.
3532          *
3533          * Then, in time, when 'stat' gets an augmenter that collects both
3534          * first and second arg (this one on the raw_syscalls:sys_exit prog
3535          * array tail call, then that one will be used.
3536          */
3537         for (key = 0; key < trace->sctbl->syscalls.nr_entries; ++key) {
3538                 struct syscall *sc = trace__syscall_info(trace, NULL, key);
3539                 struct bpf_program *pair_prog;
3540                 int prog_fd;
3541
3542                 if (sc == NULL || sc->bpf_prog.sys_enter == NULL)
3543                         continue;
3544
3545                 /*
3546                  * For now we're just reusing the sys_enter prog, and if it
3547                  * already has an augmenter, we don't need to find one.
3548                  */
3549                 if (sc->bpf_prog.sys_enter != trace->syscalls.unaugmented_prog)
3550                         continue;
3551
3552                 /*
3553                  * Look at all the other syscalls for one that has a signature
3554                  * that is close enough that we can share:
3555                  */
3556                 pair_prog = trace__find_usable_bpf_prog_entry(trace, sc);
3557                 if (pair_prog == NULL)
3558                         continue;
3559
3560                 sc->bpf_prog.sys_enter = pair_prog;
3561
3562                 /*
3563                  * Update the BPF_MAP_TYPE_PROG_SHARED for raw_syscalls:sys_enter
3564                  * with the fd for the program we're reusing:
3565                  */
3566                 prog_fd = bpf_program__fd(sc->bpf_prog.sys_enter);
3567                 err = bpf_map_update_elem(map_enter_fd, &key, &prog_fd, BPF_ANY);
3568                 if (err)
3569                         break;
3570         }
3571
3572
3573         return err;
3574 }
3575
3576 static void trace__delete_augmented_syscalls(struct trace *trace)
3577 {
3578         struct evsel *evsel, *tmp;
3579
3580         evlist__remove(trace->evlist, trace->syscalls.events.augmented);
3581         evsel__delete(trace->syscalls.events.augmented);
3582         trace->syscalls.events.augmented = NULL;
3583
3584         evlist__for_each_entry_safe(trace->evlist, tmp, evsel) {
3585                 if (evsel->bpf_obj == trace->bpf_obj) {
3586                         evlist__remove(trace->evlist, evsel);
3587                         evsel__delete(evsel);
3588                 }
3589
3590         }
3591
3592         bpf_object__close(trace->bpf_obj);
3593         trace->bpf_obj = NULL;
3594 }
3595 #else // HAVE_LIBBPF_SUPPORT
3596 static struct bpf_map *trace__find_bpf_map_by_name(struct trace *trace __maybe_unused,
3597                                                    const char *name __maybe_unused)
3598 {
3599         return NULL;
3600 }
3601
3602 static void trace__set_bpf_map_filtered_pids(struct trace *trace __maybe_unused)
3603 {
3604 }
3605
3606 static void trace__set_bpf_map_syscalls(struct trace *trace __maybe_unused)
3607 {
3608 }
3609
3610 static int trace__set_ev_qualifier_bpf_filter(struct trace *trace __maybe_unused)
3611 {
3612         return 0;
3613 }
3614
3615 static int trace__init_syscalls_bpf_map(struct trace *trace __maybe_unused)
3616 {
3617         return 0;
3618 }
3619
3620 static struct bpf_program *trace__find_bpf_program_by_title(struct trace *trace __maybe_unused,
3621                                                             const char *name __maybe_unused)
3622 {
3623         return NULL;
3624 }
3625
3626 static int trace__init_syscalls_bpf_prog_array_maps(struct trace *trace __maybe_unused)
3627 {
3628         return 0;
3629 }
3630
3631 static void trace__delete_augmented_syscalls(struct trace *trace __maybe_unused)
3632 {
3633 }
3634 #endif // HAVE_LIBBPF_SUPPORT
3635
3636 static bool trace__only_augmented_syscalls_evsels(struct trace *trace)
3637 {
3638         struct evsel *evsel;
3639
3640         evlist__for_each_entry(trace->evlist, evsel) {
3641                 if (evsel == trace->syscalls.events.augmented ||
3642                     evsel->bpf_obj == trace->bpf_obj)
3643                         continue;
3644
3645                 return false;
3646         }
3647
3648         return true;
3649 }
3650
3651 static int trace__set_ev_qualifier_filter(struct trace *trace)
3652 {
3653         if (trace->syscalls.map)
3654                 return trace__set_ev_qualifier_bpf_filter(trace);
3655         if (trace->syscalls.events.sys_enter)
3656                 return trace__set_ev_qualifier_tp_filter(trace);
3657         return 0;
3658 }
3659
3660 static int bpf_map__set_filter_pids(struct bpf_map *map __maybe_unused,
3661                                     size_t npids __maybe_unused, pid_t *pids __maybe_unused)
3662 {
3663         int err = 0;
3664 #ifdef HAVE_LIBBPF_SUPPORT
3665         bool value = true;
3666         int map_fd = bpf_map__fd(map);
3667         size_t i;
3668
3669         for (i = 0; i < npids; ++i) {
3670                 err = bpf_map_update_elem(map_fd, &pids[i], &value, BPF_ANY);
3671                 if (err)
3672                         break;
3673         }
3674 #endif
3675         return err;
3676 }
3677
3678 static int trace__set_filter_loop_pids(struct trace *trace)
3679 {
3680         unsigned int nr = 1, err;
3681         pid_t pids[32] = {
3682                 getpid(),
3683         };
3684         struct thread *thread = machine__find_thread(trace->host, pids[0], pids[0]);
3685
3686         while (thread && nr < ARRAY_SIZE(pids)) {
3687                 struct thread *parent = machine__find_thread(trace->host, thread->ppid, thread->ppid);
3688
3689                 if (parent == NULL)
3690                         break;
3691
3692                 if (!strcmp(thread__comm_str(parent), "sshd") ||
3693                     strstarts(thread__comm_str(parent), "gnome-terminal")) {
3694                         pids[nr++] = parent->tid;
3695                         break;
3696                 }
3697                 thread = parent;
3698         }
3699
3700         err = evlist__append_tp_filter_pids(trace->evlist, nr, pids);
3701         if (!err && trace->filter_pids.map)
3702                 err = bpf_map__set_filter_pids(trace->filter_pids.map, nr, pids);
3703
3704         return err;
3705 }
3706
3707 static int trace__set_filter_pids(struct trace *trace)
3708 {
3709         int err = 0;
3710         /*
3711          * Better not use !target__has_task() here because we need to cover the
3712          * case where no threads were specified in the command line, but a
3713          * workload was, and in that case we will fill in the thread_map when
3714          * we fork the workload in evlist__prepare_workload.
3715          */
3716         if (trace->filter_pids.nr > 0) {
3717                 err = evlist__append_tp_filter_pids(trace->evlist, trace->filter_pids.nr,
3718                                                     trace->filter_pids.entries);
3719                 if (!err && trace->filter_pids.map) {
3720                         err = bpf_map__set_filter_pids(trace->filter_pids.map, trace->filter_pids.nr,
3721                                                        trace->filter_pids.entries);
3722                 }
3723         } else if (perf_thread_map__pid(trace->evlist->core.threads, 0) == -1) {
3724                 err = trace__set_filter_loop_pids(trace);
3725         }
3726
3727         return err;
3728 }
3729
3730 static int __trace__deliver_event(struct trace *trace, union perf_event *event)
3731 {
3732         struct evlist *evlist = trace->evlist;
3733         struct perf_sample sample;
3734         int err = evlist__parse_sample(evlist, event, &sample);
3735
3736         if (err)
3737                 fprintf(trace->output, "Can't parse sample, err = %d, skipping...\n", err);
3738         else
3739                 trace__handle_event(trace, event, &sample);
3740
3741         return 0;
3742 }
3743
3744 static int __trace__flush_events(struct trace *trace)
3745 {
3746         u64 first = ordered_events__first_time(&trace->oe.data);
3747         u64 flush = trace->oe.last - NSEC_PER_SEC;
3748
3749         /* Is there some thing to flush.. */
3750         if (first && first < flush)
3751                 return ordered_events__flush_time(&trace->oe.data, flush);
3752
3753         return 0;
3754 }
3755
3756 static int trace__flush_events(struct trace *trace)
3757 {
3758         return !trace->sort_events ? 0 : __trace__flush_events(trace);
3759 }
3760
3761 static int trace__deliver_event(struct trace *trace, union perf_event *event)
3762 {
3763         int err;
3764
3765         if (!trace->sort_events)
3766                 return __trace__deliver_event(trace, event);
3767
3768         err = evlist__parse_sample_timestamp(trace->evlist, event, &trace->oe.last);
3769         if (err && err != -1)
3770                 return err;
3771
3772         err = ordered_events__queue(&trace->oe.data, event, trace->oe.last, 0);
3773         if (err)
3774                 return err;
3775
3776         return trace__flush_events(trace);
3777 }
3778
3779 static int ordered_events__deliver_event(struct ordered_events *oe,
3780                                          struct ordered_event *event)
3781 {
3782         struct trace *trace = container_of(oe, struct trace, oe.data);
3783
3784         return __trace__deliver_event(trace, event->event);
3785 }
3786
3787 static struct syscall_arg_fmt *evsel__find_syscall_arg_fmt_by_name(struct evsel *evsel, char *arg)
3788 {
3789         struct tep_format_field *field;
3790         struct syscall_arg_fmt *fmt = __evsel__syscall_arg_fmt(evsel);
3791
3792         if (evsel->tp_format == NULL || fmt == NULL)
3793                 return NULL;
3794
3795         for (field = evsel->tp_format->format.fields; field; field = field->next, ++fmt)
3796                 if (strcmp(field->name, arg) == 0)
3797                         return fmt;
3798
3799         return NULL;
3800 }
3801
3802 static int trace__expand_filter(struct trace *trace __maybe_unused, struct evsel *evsel)
3803 {
3804         char *tok, *left = evsel->filter, *new_filter = evsel->filter;
3805
3806         while ((tok = strpbrk(left, "=<>!")) != NULL) {
3807                 char *right = tok + 1, *right_end;
3808
3809                 if (*right == '=')
3810                         ++right;
3811
3812                 while (isspace(*right))
3813                         ++right;
3814
3815                 if (*right == '\0')
3816                         break;
3817
3818                 while (!isalpha(*left))
3819                         if (++left == tok) {
3820                                 /*
3821                                  * Bail out, can't find the name of the argument that is being
3822                                  * used in the filter, let it try to set this filter, will fail later.
3823                                  */
3824                                 return 0;
3825                         }
3826
3827                 right_end = right + 1;
3828                 while (isalnum(*right_end) || *right_end == '_' || *right_end == '|')
3829                         ++right_end;
3830
3831                 if (isalpha(*right)) {
3832                         struct syscall_arg_fmt *fmt;
3833                         int left_size = tok - left,
3834                             right_size = right_end - right;
3835                         char arg[128];
3836
3837                         while (isspace(left[left_size - 1]))
3838                                 --left_size;
3839
3840                         scnprintf(arg, sizeof(arg), "%.*s", left_size, left);
3841
3842                         fmt = evsel__find_syscall_arg_fmt_by_name(evsel, arg);
3843                         if (fmt == NULL) {
3844                                 pr_err("\"%s\" not found in \"%s\", can't set filter \"%s\"\n",
3845                                        arg, evsel->name, evsel->filter);
3846                                 return -1;
3847                         }
3848
3849                         pr_debug2("trying to expand \"%s\" \"%.*s\" \"%.*s\" -> ",
3850                                  arg, (int)(right - tok), tok, right_size, right);
3851
3852                         if (fmt->strtoul) {
3853                                 u64 val;
3854                                 struct syscall_arg syscall_arg = {
3855                                         .parm = fmt->parm,
3856                                 };
3857
3858                                 if (fmt->strtoul(right, right_size, &syscall_arg, &val)) {
3859                                         char *n, expansion[19];
3860                                         int expansion_lenght = scnprintf(expansion, sizeof(expansion), "%#" PRIx64, val);
3861                                         int expansion_offset = right - new_filter;
3862
3863                                         pr_debug("%s", expansion);
3864
3865                                         if (asprintf(&n, "%.*s%s%s", expansion_offset, new_filter, expansion, right_end) < 0) {
3866                                                 pr_debug(" out of memory!\n");
3867                                                 free(new_filter);
3868                                                 return -1;
3869                                         }
3870                                         if (new_filter != evsel->filter)
3871                                                 free(new_filter);
3872                                         left = n + expansion_offset + expansion_lenght;
3873                                         new_filter = n;
3874                                 } else {
3875                                         pr_err("\"%.*s\" not found for \"%s\" in \"%s\", can't set filter \"%s\"\n",
3876                                                right_size, right, arg, evsel->name, evsel->filter);
3877                                         return -1;
3878                                 }
3879                         } else {
3880                                 pr_err("No resolver (strtoul) for \"%s\" in \"%s\", can't set filter \"%s\"\n",
3881                                        arg, evsel->name, evsel->filter);
3882                                 return -1;
3883                         }
3884
3885                         pr_debug("\n");
3886                 } else {
3887                         left = right_end;
3888                 }
3889         }
3890
3891         if (new_filter != evsel->filter) {
3892                 pr_debug("New filter for %s: %s\n", evsel->name, new_filter);
3893                 evsel__set_filter(evsel, new_filter);
3894                 free(new_filter);
3895         }
3896
3897         return 0;
3898 }
3899
3900 static int trace__expand_filters(struct trace *trace, struct evsel **err_evsel)
3901 {
3902         struct evlist *evlist = trace->evlist;
3903         struct evsel *evsel;
3904
3905         evlist__for_each_entry(evlist, evsel) {
3906                 if (evsel->filter == NULL)
3907                         continue;
3908
3909                 if (trace__expand_filter(trace, evsel)) {
3910                         *err_evsel = evsel;
3911                         return -1;
3912                 }
3913         }
3914
3915         return 0;
3916 }
3917
3918 static int trace__run(struct trace *trace, int argc, const char **argv)
3919 {
3920         struct evlist *evlist = trace->evlist;
3921         struct evsel *evsel, *pgfault_maj = NULL, *pgfault_min = NULL;
3922         int err = -1, i;
3923         unsigned long before;
3924         const bool forks = argc > 0;
3925         bool draining = false;
3926
3927         trace->live = true;
3928
3929         if (!trace->raw_augmented_syscalls) {
3930                 if (trace->trace_syscalls && trace__add_syscall_newtp(trace))
3931                         goto out_error_raw_syscalls;
3932
3933                 if (trace->trace_syscalls)
3934                         trace->vfs_getname = evlist__add_vfs_getname(evlist);
3935         }
3936
3937         if ((trace->trace_pgfaults & TRACE_PFMAJ)) {
3938                 pgfault_maj = evsel__new_pgfault(PERF_COUNT_SW_PAGE_FAULTS_MAJ);
3939                 if (pgfault_maj == NULL)
3940                         goto out_error_mem;
3941                 evsel__config_callchain(pgfault_maj, &trace->opts, &callchain_param);
3942                 evlist__add(evlist, pgfault_maj);
3943         }
3944
3945         if ((trace->trace_pgfaults & TRACE_PFMIN)) {
3946                 pgfault_min = evsel__new_pgfault(PERF_COUNT_SW_PAGE_FAULTS_MIN);
3947                 if (pgfault_min == NULL)
3948                         goto out_error_mem;
3949                 evsel__config_callchain(pgfault_min, &trace->opts, &callchain_param);
3950                 evlist__add(evlist, pgfault_min);
3951         }
3952
3953         if (trace->sched &&
3954             evlist__add_newtp(evlist, "sched", "sched_stat_runtime", trace__sched_stat_runtime))
3955                 goto out_error_sched_stat_runtime;
3956         /*
3957          * If a global cgroup was set, apply it to all the events without an
3958          * explicit cgroup. I.e.:
3959          *
3960          *      trace -G A -e sched:*switch
3961          *
3962          * Will set all raw_syscalls:sys_{enter,exit}, pgfault, vfs_getname, etc
3963          * _and_ sched:sched_switch to the 'A' cgroup, while:
3964          *
3965          * trace -e sched:*switch -G A
3966          *
3967          * will only set the sched:sched_switch event to the 'A' cgroup, all the
3968          * other events (raw_syscalls:sys_{enter,exit}, etc are left "without"
3969          * a cgroup (on the root cgroup, sys wide, etc).
3970          *
3971          * Multiple cgroups:
3972          *
3973          * trace -G A -e sched:*switch -G B
3974          *
3975          * the syscall ones go to the 'A' cgroup, the sched:sched_switch goes
3976          * to the 'B' cgroup.
3977          *
3978          * evlist__set_default_cgroup() grabs a reference of the passed cgroup
3979          * only for the evsels still without a cgroup, i.e. evsel->cgroup == NULL.
3980          */
3981         if (trace->cgroup)
3982                 evlist__set_default_cgroup(trace->evlist, trace->cgroup);
3983
3984         err = evlist__create_maps(evlist, &trace->opts.target);
3985         if (err < 0) {
3986                 fprintf(trace->output, "Problems parsing the target to trace, check your options!\n");
3987                 goto out_delete_evlist;
3988         }
3989
3990         err = trace__symbols_init(trace, evlist);
3991         if (err < 0) {
3992                 fprintf(trace->output, "Problems initializing symbol libraries!\n");
3993                 goto out_delete_evlist;
3994         }
3995
3996         evlist__config(evlist, &trace->opts, &callchain_param);
3997
3998         if (forks) {
3999                 err = evlist__prepare_workload(evlist, &trace->opts.target, argv, false, NULL);
4000                 if (err < 0) {
4001                         fprintf(trace->output, "Couldn't run the workload!\n");
4002                         goto out_delete_evlist;
4003                 }
4004         }
4005
4006         err = evlist__open(evlist);
4007         if (err < 0)
4008                 goto out_error_open;
4009
4010         err = bpf__apply_obj_config();
4011         if (err) {
4012                 char errbuf[BUFSIZ];
4013
4014                 bpf__strerror_apply_obj_config(err, errbuf, sizeof(errbuf));
4015                 pr_err("ERROR: Apply config to BPF failed: %s\n",
4016                          errbuf);
4017                 goto out_error_open;
4018         }
4019
4020         err = trace__set_filter_pids(trace);
4021         if (err < 0)
4022                 goto out_error_mem;
4023
4024         if (trace->syscalls.map)
4025                 trace__init_syscalls_bpf_map(trace);
4026
4027         if (trace->syscalls.prog_array.sys_enter)
4028                 trace__init_syscalls_bpf_prog_array_maps(trace);
4029
4030         if (trace->ev_qualifier_ids.nr > 0) {
4031                 err = trace__set_ev_qualifier_filter(trace);
4032                 if (err < 0)
4033                         goto out_errno;
4034
4035                 if (trace->syscalls.events.sys_exit) {
4036                         pr_debug("event qualifier tracepoint filter: %s\n",
4037                                  trace->syscalls.events.sys_exit->filter);
4038                 }
4039         }
4040
4041         /*
4042          * If the "close" syscall is not traced, then we will not have the
4043          * opportunity to, in syscall_arg__scnprintf_close_fd() invalidate the
4044          * fd->pathname table and were ending up showing the last value set by
4045          * syscalls opening a pathname and associating it with a descriptor or
4046          * reading it from /proc/pid/fd/ in cases where that doesn't make
4047          * sense.
4048          *
4049          *  So just disable this beautifier (SCA_FD, SCA_FDAT) when 'close' is
4050          *  not in use.
4051          */
4052         trace->fd_path_disabled = !trace__syscall_enabled(trace, syscalltbl__id(trace->sctbl, "close"));
4053
4054         err = trace__expand_filters(trace, &evsel);
4055         if (err)
4056                 goto out_delete_evlist;
4057         err = evlist__apply_filters(evlist, &evsel);
4058         if (err < 0)
4059                 goto out_error_apply_filters;
4060
4061         if (trace->dump.map)
4062                 bpf_map__fprintf(trace->dump.map, trace->output);
4063
4064         err = evlist__mmap(evlist, trace->opts.mmap_pages);
4065         if (err < 0)
4066                 goto out_error_mmap;
4067
4068         if (!target__none(&trace->opts.target) && !trace->opts.initial_delay)
4069                 evlist__enable(evlist);
4070
4071         if (forks)
4072                 evlist__start_workload(evlist);
4073
4074         if (trace->opts.initial_delay) {
4075                 usleep(trace->opts.initial_delay * 1000);
4076                 evlist__enable(evlist);
4077         }
4078
4079         trace->multiple_threads = perf_thread_map__pid(evlist->core.threads, 0) == -1 ||
4080                                   evlist->core.threads->nr > 1 ||
4081                                   evlist__first(evlist)->core.attr.inherit;
4082
4083         /*
4084          * Now that we already used evsel->core.attr to ask the kernel to setup the
4085          * events, lets reuse evsel->core.attr.sample_max_stack as the limit in
4086          * trace__resolve_callchain(), allowing per-event max-stack settings
4087          * to override an explicitly set --max-stack global setting.
4088          */
4089         evlist__for_each_entry(evlist, evsel) {
4090                 if (evsel__has_callchain(evsel) &&
4091                     evsel->core.attr.sample_max_stack == 0)
4092                         evsel->core.attr.sample_max_stack = trace->max_stack;
4093         }
4094 again:
4095         before = trace->nr_events;
4096
4097         for (i = 0; i < evlist->core.nr_mmaps; i++) {
4098                 union perf_event *event;
4099                 struct mmap *md;
4100
4101                 md = &evlist->mmap[i];
4102                 if (perf_mmap__read_init(&md->core) < 0)
4103                         continue;
4104
4105                 while ((event = perf_mmap__read_event(&md->core)) != NULL) {
4106                         ++trace->nr_events;
4107
4108                         err = trace__deliver_event(trace, event);
4109                         if (err)
4110                                 goto out_disable;
4111
4112                         perf_mmap__consume(&md->core);
4113
4114                         if (interrupted)
4115                                 goto out_disable;
4116
4117                         if (done && !draining) {
4118                                 evlist__disable(evlist);
4119                                 draining = true;
4120                         }
4121                 }
4122                 perf_mmap__read_done(&md->core);
4123         }
4124
4125         if (trace->nr_events == before) {
4126                 int timeout = done ? 100 : -1;
4127
4128                 if (!draining && evlist__poll(evlist, timeout) > 0) {
4129                         if (evlist__filter_pollfd(evlist, POLLERR | POLLHUP | POLLNVAL) == 0)
4130                                 draining = true;
4131
4132                         goto again;
4133                 } else {
4134                         if (trace__flush_events(trace))
4135                                 goto out_disable;
4136                 }
4137         } else {
4138                 goto again;
4139         }
4140
4141 out_disable:
4142         thread__zput(trace->current);
4143
4144         evlist__disable(evlist);
4145
4146         if (trace->sort_events)
4147                 ordered_events__flush(&trace->oe.data, OE_FLUSH__FINAL);
4148
4149         if (!err) {
4150                 if (trace->summary)
4151                         trace__fprintf_thread_summary(trace, trace->output);
4152
4153                 if (trace->show_tool_stats) {
4154                         fprintf(trace->output, "Stats:\n "
4155                                                " vfs_getname : %" PRIu64 "\n"
4156                                                " proc_getname: %" PRIu64 "\n",
4157                                 trace->stats.vfs_getname,
4158                                 trace->stats.proc_getname);
4159                 }
4160         }
4161
4162 out_delete_evlist:
4163         trace__symbols__exit(trace);
4164         evlist__free_syscall_tp_fields(evlist);
4165         evlist__delete(evlist);
4166         cgroup__put(trace->cgroup);
4167         trace->evlist = NULL;
4168         trace->live = false;
4169         return err;
4170 {
4171         char errbuf[BUFSIZ];
4172
4173 out_error_sched_stat_runtime:
4174         tracing_path__strerror_open_tp(errno, errbuf, sizeof(errbuf), "sched", "sched_stat_runtime");
4175         goto out_error;
4176
4177 out_error_raw_syscalls:
4178         tracing_path__strerror_open_tp(errno, errbuf, sizeof(errbuf), "raw_syscalls", "sys_(enter|exit)");
4179         goto out_error;
4180
4181 out_error_mmap:
4182         evlist__strerror_mmap(evlist, errno, errbuf, sizeof(errbuf));
4183         goto out_error;
4184
4185 out_error_open:
4186         evlist__strerror_open(evlist, errno, errbuf, sizeof(errbuf));
4187
4188 out_error:
4189         fprintf(trace->output, "%s\n", errbuf);
4190         goto out_delete_evlist;
4191
4192 out_error_apply_filters:
4193         fprintf(trace->output,
4194                 "Failed to set filter \"%s\" on event %s with %d (%s)\n",
4195                 evsel->filter, evsel__name(evsel), errno,
4196                 str_error_r(errno, errbuf, sizeof(errbuf)));
4197         goto out_delete_evlist;
4198 }
4199 out_error_mem:
4200         fprintf(trace->output, "Not enough memory to run!\n");
4201         goto out_delete_evlist;
4202
4203 out_errno:
4204         fprintf(trace->output, "errno=%d,%s\n", errno, strerror(errno));
4205         goto out_delete_evlist;
4206 }
4207
4208 static int trace__replay(struct trace *trace)
4209 {
4210         const struct evsel_str_handler handlers[] = {
4211                 { "probe:vfs_getname",       trace__vfs_getname, },
4212         };
4213         struct perf_data data = {
4214                 .path  = input_name,
4215                 .mode  = PERF_DATA_MODE_READ,
4216                 .force = trace->force,
4217         };
4218         struct perf_session *session;
4219         struct evsel *evsel;
4220         int err = -1;
4221
4222         trace->tool.sample        = trace__process_sample;
4223         trace->tool.mmap          = perf_event__process_mmap;
4224         trace->tool.mmap2         = perf_event__process_mmap2;
4225         trace->tool.comm          = perf_event__process_comm;
4226         trace->tool.exit          = perf_event__process_exit;
4227         trace->tool.fork          = perf_event__process_fork;
4228         trace->tool.attr          = perf_event__process_attr;
4229         trace->tool.tracing_data  = perf_event__process_tracing_data;
4230         trace->tool.build_id      = perf_event__process_build_id;
4231         trace->tool.namespaces    = perf_event__process_namespaces;
4232
4233         trace->tool.ordered_events = true;
4234         trace->tool.ordering_requires_timestamps = true;
4235
4236         /* add tid to output */
4237         trace->multiple_threads = true;
4238
4239         session = perf_session__new(&data, &trace->tool);
4240         if (IS_ERR(session))
4241                 return PTR_ERR(session);
4242
4243         if (trace->opts.target.pid)
4244                 symbol_conf.pid_list_str = strdup(trace->opts.target.pid);
4245
4246         if (trace->opts.target.tid)
4247                 symbol_conf.tid_list_str = strdup(trace->opts.target.tid);
4248
4249         if (symbol__init(&session->header.env) < 0)
4250                 goto out;
4251
4252         trace->host = &session->machines.host;
4253
4254         err = perf_session__set_tracepoints_handlers(session, handlers);
4255         if (err)
4256                 goto out;
4257
4258         evsel = evlist__find_tracepoint_by_name(session->evlist, "raw_syscalls:sys_enter");
4259         /* older kernels have syscalls tp versus raw_syscalls */
4260         if (evsel == NULL)
4261                 evsel = evlist__find_tracepoint_by_name(session->evlist, "syscalls:sys_enter");
4262
4263         if (evsel &&
4264             (evsel__init_raw_syscall_tp(evsel, trace__sys_enter) < 0 ||
4265             perf_evsel__init_sc_tp_ptr_field(evsel, args))) {
4266                 pr_err("Error during initialize raw_syscalls:sys_enter event\n");
4267                 goto out;
4268         }
4269
4270         evsel = evlist__find_tracepoint_by_name(session->evlist, "raw_syscalls:sys_exit");
4271         if (evsel == NULL)
4272                 evsel = evlist__find_tracepoint_by_name(session->evlist, "syscalls:sys_exit");
4273         if (evsel &&
4274             (evsel__init_raw_syscall_tp(evsel, trace__sys_exit) < 0 ||
4275             perf_evsel__init_sc_tp_uint_field(evsel, ret))) {
4276                 pr_err("Error during initialize raw_syscalls:sys_exit event\n");
4277                 goto out;
4278         }
4279
4280         evlist__for_each_entry(session->evlist, evsel) {
4281                 if (evsel->core.attr.type == PERF_TYPE_SOFTWARE &&
4282                     (evsel->core.attr.config == PERF_COUNT_SW_PAGE_FAULTS_MAJ ||
4283                      evsel->core.attr.config == PERF_COUNT_SW_PAGE_FAULTS_MIN ||
4284                      evsel->core.attr.config == PERF_COUNT_SW_PAGE_FAULTS))
4285                         evsel->handler = trace__pgfault;
4286         }
4287
4288         setup_pager();
4289
4290         err = perf_session__process_events(session);
4291         if (err)
4292                 pr_err("Failed to process events, error %d", err);
4293
4294         else if (trace->summary)
4295                 trace__fprintf_thread_summary(trace, trace->output);
4296
4297 out:
4298         perf_session__delete(session);
4299
4300         return err;
4301 }
4302
4303 static size_t trace__fprintf_threads_header(FILE *fp)
4304 {
4305         size_t printed;
4306
4307         printed  = fprintf(fp, "\n Summary of events:\n\n");
4308
4309         return printed;
4310 }
4311
4312 DEFINE_RESORT_RB(syscall_stats, a->msecs > b->msecs,
4313         struct syscall_stats *stats;
4314         double               msecs;
4315         int                  syscall;
4316 )
4317 {
4318         struct int_node *source = rb_entry(nd, struct int_node, rb_node);
4319         struct syscall_stats *stats = source->priv;
4320
4321         entry->syscall = source->i;
4322         entry->stats   = stats;
4323         entry->msecs   = stats ? (u64)stats->stats.n * (avg_stats(&stats->stats) / NSEC_PER_MSEC) : 0;
4324 }
4325
4326 static size_t thread__dump_stats(struct thread_trace *ttrace,
4327                                  struct trace *trace, FILE *fp)
4328 {
4329         size_t printed = 0;
4330         struct syscall *sc;
4331         struct rb_node *nd;
4332         DECLARE_RESORT_RB_INTLIST(syscall_stats, ttrace->syscall_stats);
4333
4334         if (syscall_stats == NULL)
4335                 return 0;
4336
4337         printed += fprintf(fp, "\n");
4338
4339         printed += fprintf(fp, "   syscall            calls  errors  total       min       avg       max       stddev\n");
4340         printed += fprintf(fp, "                                     (msec)    (msec)    (msec)    (msec)        (%%)\n");
4341         printed += fprintf(fp, "   --------------- --------  ------ -------- --------- --------- ---------     ------\n");
4342
4343         resort_rb__for_each_entry(nd, syscall_stats) {
4344                 struct syscall_stats *stats = syscall_stats_entry->stats;
4345                 if (stats) {
4346                         double min = (double)(stats->stats.min) / NSEC_PER_MSEC;
4347                         double max = (double)(stats->stats.max) / NSEC_PER_MSEC;
4348                         double avg = avg_stats(&stats->stats);
4349                         double pct;
4350                         u64 n = (u64)stats->stats.n;
4351
4352                         pct = avg ? 100.0 * stddev_stats(&stats->stats) / avg : 0.0;
4353                         avg /= NSEC_PER_MSEC;
4354
4355                         sc = &trace->syscalls.table[syscall_stats_entry->syscall];
4356                         printed += fprintf(fp, "   %-15s", sc->name);
4357                         printed += fprintf(fp, " %8" PRIu64 " %6" PRIu64 " %9.3f %9.3f %9.3f",
4358                                            n, stats->nr_failures, syscall_stats_entry->msecs, min, avg);
4359                         printed += fprintf(fp, " %9.3f %9.2f%%\n", max, pct);
4360
4361                         if (trace->errno_summary && stats->nr_failures) {
4362                                 const char *arch_name = perf_env__arch(trace->host->env);
4363                                 int e;
4364
4365                                 for (e = 0; e < stats->max_errno; ++e) {
4366                                         if (stats->errnos[e] != 0)
4367                                                 fprintf(fp, "\t\t\t\t%s: %d\n", arch_syscalls__strerrno(arch_name, e + 1), stats->errnos[e]);
4368                                 }
4369                         }
4370                 }
4371         }
4372
4373         resort_rb__delete(syscall_stats);
4374         printed += fprintf(fp, "\n\n");
4375
4376         return printed;
4377 }
4378
4379 static size_t trace__fprintf_thread(FILE *fp, struct thread *thread, struct trace *trace)
4380 {
4381         size_t printed = 0;
4382         struct thread_trace *ttrace = thread__priv(thread);
4383         double ratio;
4384
4385         if (ttrace == NULL)
4386                 return 0;
4387
4388         ratio = (double)ttrace->nr_events / trace->nr_events * 100.0;
4389
4390         printed += fprintf(fp, " %s (%d), ", thread__comm_str(thread), thread->tid);
4391         printed += fprintf(fp, "%lu events, ", ttrace->nr_events);
4392         printed += fprintf(fp, "%.1f%%", ratio);
4393         if (ttrace->pfmaj)
4394                 printed += fprintf(fp, ", %lu majfaults", ttrace->pfmaj);
4395         if (ttrace->pfmin)
4396                 printed += fprintf(fp, ", %lu minfaults", ttrace->pfmin);
4397         if (trace->sched)
4398                 printed += fprintf(fp, ", %.3f msec\n", ttrace->runtime_ms);
4399         else if (fputc('\n', fp) != EOF)
4400                 ++printed;
4401
4402         printed += thread__dump_stats(ttrace, trace, fp);
4403
4404         return printed;
4405 }
4406
4407 static unsigned long thread__nr_events(struct thread_trace *ttrace)
4408 {
4409         return ttrace ? ttrace->nr_events : 0;
4410 }
4411
4412 DEFINE_RESORT_RB(threads, (thread__nr_events(a->thread->priv) < thread__nr_events(b->thread->priv)),
4413         struct thread *thread;
4414 )
4415 {
4416         entry->thread = rb_entry(nd, struct thread, rb_node);
4417 }
4418
4419 static size_t trace__fprintf_thread_summary(struct trace *trace, FILE *fp)
4420 {
4421         size_t printed = trace__fprintf_threads_header(fp);
4422         struct rb_node *nd;
4423         int i;
4424
4425         for (i = 0; i < THREADS__TABLE_SIZE; i++) {
4426                 DECLARE_RESORT_RB_MACHINE_THREADS(threads, trace->host, i);
4427
4428                 if (threads == NULL) {
4429                         fprintf(fp, "%s", "Error sorting output by nr_events!\n");
4430                         return 0;
4431                 }
4432
4433                 resort_rb__for_each_entry(nd, threads)
4434                         printed += trace__fprintf_thread(fp, threads_entry->thread, trace);
4435
4436                 resort_rb__delete(threads);
4437         }
4438         return printed;
4439 }
4440
4441 static int trace__set_duration(const struct option *opt, const char *str,
4442                                int unset __maybe_unused)
4443 {
4444         struct trace *trace = opt->value;
4445
4446         trace->duration_filter = atof(str);
4447         return 0;
4448 }
4449
4450 static int trace__set_filter_pids_from_option(const struct option *opt, const char *str,
4451                                               int unset __maybe_unused)
4452 {
4453         int ret = -1;
4454         size_t i;
4455         struct trace *trace = opt->value;
4456         /*
4457          * FIXME: introduce a intarray class, plain parse csv and create a
4458          * { int nr, int entries[] } struct...
4459          */
4460         struct intlist *list = intlist__new(str);
4461
4462         if (list == NULL)
4463                 return -1;
4464
4465         i = trace->filter_pids.nr = intlist__nr_entries(list) + 1;
4466         trace->filter_pids.entries = calloc(i, sizeof(pid_t));
4467
4468         if (trace->filter_pids.entries == NULL)
4469                 goto out;
4470
4471         trace->filter_pids.entries[0] = getpid();
4472
4473         for (i = 1; i < trace->filter_pids.nr; ++i)
4474                 trace->filter_pids.entries[i] = intlist__entry(list, i - 1)->i;
4475
4476         intlist__delete(list);
4477         ret = 0;
4478 out:
4479         return ret;
4480 }
4481
4482 static int trace__open_output(struct trace *trace, const char *filename)
4483 {
4484         struct stat st;
4485
4486         if (!stat(filename, &st) && st.st_size) {
4487                 char oldname[PATH_MAX];
4488
4489                 scnprintf(oldname, sizeof(oldname), "%s.old", filename);
4490                 unlink(oldname);
4491                 rename(filename, oldname);
4492         }
4493
4494         trace->output = fopen(filename, "w");
4495
4496         return trace->output == NULL ? -errno : 0;
4497 }
4498
4499 static int parse_pagefaults(const struct option *opt, const char *str,
4500                             int unset __maybe_unused)
4501 {
4502         int *trace_pgfaults = opt->value;
4503
4504         if (strcmp(str, "all") == 0)
4505                 *trace_pgfaults |= TRACE_PFMAJ | TRACE_PFMIN;
4506         else if (strcmp(str, "maj") == 0)
4507                 *trace_pgfaults |= TRACE_PFMAJ;
4508         else if (strcmp(str, "min") == 0)
4509                 *trace_pgfaults |= TRACE_PFMIN;
4510         else
4511                 return -1;
4512
4513         return 0;
4514 }
4515
4516 static void evlist__set_default_evsel_handler(struct evlist *evlist, void *handler)
4517 {
4518         struct evsel *evsel;
4519
4520         evlist__for_each_entry(evlist, evsel) {
4521                 if (evsel->handler == NULL)
4522                         evsel->handler = handler;
4523         }
4524 }
4525
4526 static void evsel__set_syscall_arg_fmt(struct evsel *evsel, const char *name)
4527 {
4528         struct syscall_arg_fmt *fmt = evsel__syscall_arg_fmt(evsel);
4529
4530         if (fmt) {
4531                 struct syscall_fmt *scfmt = syscall_fmt__find(name);
4532
4533                 if (scfmt) {
4534                         int skip = 0;
4535
4536                         if (strcmp(evsel->tp_format->format.fields->name, "__syscall_nr") == 0 ||
4537                             strcmp(evsel->tp_format->format.fields->name, "nr") == 0)
4538                                 ++skip;
4539
4540                         memcpy(fmt + skip, scfmt->arg, (evsel->tp_format->format.nr_fields - skip) * sizeof(*fmt));
4541                 }
4542         }
4543 }
4544
4545 static int evlist__set_syscall_tp_fields(struct evlist *evlist)
4546 {
4547         struct evsel *evsel;
4548
4549         evlist__for_each_entry(evlist, evsel) {
4550                 if (evsel->priv || !evsel->tp_format)
4551                         continue;
4552
4553                 if (strcmp(evsel->tp_format->system, "syscalls")) {
4554                         evsel__init_tp_arg_scnprintf(evsel);
4555                         continue;
4556                 }
4557
4558                 if (evsel__init_syscall_tp(evsel))
4559                         return -1;
4560
4561                 if (!strncmp(evsel->tp_format->name, "sys_enter_", 10)) {
4562                         struct syscall_tp *sc = __evsel__syscall_tp(evsel);
4563
4564                         if (__tp_field__init_ptr(&sc->args, sc->id.offset + sizeof(u64)))
4565                                 return -1;
4566
4567                         evsel__set_syscall_arg_fmt(evsel, evsel->tp_format->name + sizeof("sys_enter_") - 1);
4568                 } else if (!strncmp(evsel->tp_format->name, "sys_exit_", 9)) {
4569                         struct syscall_tp *sc = __evsel__syscall_tp(evsel);
4570
4571                         if (__tp_field__init_uint(&sc->ret, sizeof(u64), sc->id.offset + sizeof(u64), evsel->needs_swap))
4572                                 return -1;
4573
4574                         evsel__set_syscall_arg_fmt(evsel, evsel->tp_format->name + sizeof("sys_exit_") - 1);
4575                 }
4576         }
4577
4578         return 0;
4579 }
4580
4581 /*
4582  * XXX: Hackish, just splitting the combined -e+--event (syscalls
4583  * (raw_syscalls:{sys_{enter,exit}} + events (tracepoints, HW, SW, etc) to use
4584  * existing facilities unchanged (trace->ev_qualifier + parse_options()).
4585  *
4586  * It'd be better to introduce a parse_options() variant that would return a
4587  * list with the terms it didn't match to an event...
4588  */
4589 static int trace__parse_events_option(const struct option *opt, const char *str,
4590                                       int unset __maybe_unused)
4591 {
4592         struct trace *trace = (struct trace *)opt->value;
4593         const char *s = str;
4594         char *sep = NULL, *lists[2] = { NULL, NULL, };
4595         int len = strlen(str) + 1, err = -1, list, idx;
4596         char *strace_groups_dir = system_path(STRACE_GROUPS_DIR);
4597         char group_name[PATH_MAX];
4598         struct syscall_fmt *fmt;
4599
4600         if (strace_groups_dir == NULL)
4601                 return -1;
4602
4603         if (*s == '!') {
4604                 ++s;
4605                 trace->not_ev_qualifier = true;
4606         }
4607
4608         while (1) {
4609                 if ((sep = strchr(s, ',')) != NULL)
4610                         *sep = '\0';
4611
4612                 list = 0;
4613                 if (syscalltbl__id(trace->sctbl, s) >= 0 ||
4614                     syscalltbl__strglobmatch_first(trace->sctbl, s, &idx) >= 0) {
4615                         list = 1;
4616                         goto do_concat;
4617                 }
4618
4619                 fmt = syscall_fmt__find_by_alias(s);
4620                 if (fmt != NULL) {
4621                         list = 1;
4622                         s = fmt->name;
4623                 } else {
4624                         path__join(group_name, sizeof(group_name), strace_groups_dir, s);
4625                         if (access(group_name, R_OK) == 0)
4626                                 list = 1;
4627                 }
4628 do_concat:
4629                 if (lists[list]) {
4630                         sprintf(lists[list] + strlen(lists[list]), ",%s", s);
4631                 } else {
4632                         lists[list] = malloc(len);
4633                         if (lists[list] == NULL)
4634                                 goto out;
4635                         strcpy(lists[list], s);
4636                 }
4637
4638                 if (!sep)
4639                         break;
4640
4641                 *sep = ',';
4642                 s = sep + 1;
4643         }
4644
4645         if (lists[1] != NULL) {
4646                 struct strlist_config slist_config = {
4647                         .dirname = strace_groups_dir,
4648                 };
4649
4650                 trace->ev_qualifier = strlist__new(lists[1], &slist_config);
4651                 if (trace->ev_qualifier == NULL) {
4652                         fputs("Not enough memory to parse event qualifier", trace->output);
4653                         goto out;
4654                 }
4655
4656                 if (trace__validate_ev_qualifier(trace))
4657                         goto out;
4658                 trace->trace_syscalls = true;
4659         }
4660
4661         err = 0;
4662
4663         if (lists[0]) {
4664                 struct option o = {
4665                         .value = &trace->evlist,
4666                 };
4667                 err = parse_events_option(&o, lists[0], 0);
4668         }
4669 out:
4670         free(strace_groups_dir);
4671         free(lists[0]);
4672         free(lists[1]);
4673         if (sep)
4674                 *sep = ',';
4675
4676         return err;
4677 }
4678
4679 static int trace__parse_cgroups(const struct option *opt, const char *str, int unset)
4680 {
4681         struct trace *trace = opt->value;
4682
4683         if (!list_empty(&trace->evlist->core.entries)) {
4684                 struct option o = {
4685                         .value = &trace->evlist,
4686                 };
4687                 return parse_cgroups(&o, str, unset);
4688         }
4689         trace->cgroup = evlist__findnew_cgroup(trace->evlist, str);
4690
4691         return 0;
4692 }
4693
4694 static int trace__config(const char *var, const char *value, void *arg)
4695 {
4696         struct trace *trace = arg;
4697         int err = 0;
4698
4699         if (!strcmp(var, "trace.add_events")) {
4700                 trace->perfconfig_events = strdup(value);
4701                 if (trace->perfconfig_events == NULL) {
4702                         pr_err("Not enough memory for %s\n", "trace.add_events");
4703                         return -1;
4704                 }
4705         } else if (!strcmp(var, "trace.show_timestamp")) {
4706                 trace->show_tstamp = perf_config_bool(var, value);
4707         } else if (!strcmp(var, "trace.show_duration")) {
4708                 trace->show_duration = perf_config_bool(var, value);
4709         } else if (!strcmp(var, "trace.show_arg_names")) {
4710                 trace->show_arg_names = perf_config_bool(var, value);
4711                 if (!trace->show_arg_names)
4712                         trace->show_zeros = true;
4713         } else if (!strcmp(var, "trace.show_zeros")) {
4714                 bool new_show_zeros = perf_config_bool(var, value);
4715                 if (!trace->show_arg_names && !new_show_zeros) {
4716                         pr_warning("trace.show_zeros has to be set when trace.show_arg_names=no\n");
4717                         goto out;
4718                 }
4719                 trace->show_zeros = new_show_zeros;
4720         } else if (!strcmp(var, "trace.show_prefix")) {
4721                 trace->show_string_prefix = perf_config_bool(var, value);
4722         } else if (!strcmp(var, "trace.no_inherit")) {
4723                 trace->opts.no_inherit = perf_config_bool(var, value);
4724         } else if (!strcmp(var, "trace.args_alignment")) {
4725                 int args_alignment = 0;
4726                 if (perf_config_int(&args_alignment, var, value) == 0)
4727                         trace->args_alignment = args_alignment;
4728         } else if (!strcmp(var, "trace.tracepoint_beautifiers")) {
4729                 if (strcasecmp(value, "libtraceevent") == 0)
4730                         trace->libtraceevent_print = true;
4731                 else if (strcasecmp(value, "libbeauty") == 0)
4732                         trace->libtraceevent_print = false;
4733         }
4734 out:
4735         return err;
4736 }
4737
4738 static void trace__exit(struct trace *trace)
4739 {
4740         int i;
4741
4742         strlist__delete(trace->ev_qualifier);
4743         free(trace->ev_qualifier_ids.entries);
4744         if (trace->syscalls.table) {
4745                 for (i = 0; i <= trace->sctbl->syscalls.max_id; i++)
4746                         syscall__exit(&trace->syscalls.table[i]);
4747                 free(trace->syscalls.table);
4748         }
4749         syscalltbl__delete(trace->sctbl);
4750         zfree(&trace->perfconfig_events);
4751 }
4752
4753 int cmd_trace(int argc, const char **argv)
4754 {
4755         const char *trace_usage[] = {
4756                 "perf trace [<options>] [<command>]",
4757                 "perf trace [<options>] -- <command> [<options>]",
4758                 "perf trace record [<options>] [<command>]",
4759                 "perf trace record [<options>] -- <command> [<options>]",
4760                 NULL
4761         };
4762         struct trace trace = {
4763                 .opts = {
4764                         .target = {
4765                                 .uid       = UINT_MAX,
4766                                 .uses_mmap = true,
4767                         },
4768                         .user_freq     = UINT_MAX,
4769                         .user_interval = ULLONG_MAX,
4770                         .no_buffering  = true,
4771                         .mmap_pages    = UINT_MAX,
4772                 },
4773                 .output = stderr,
4774                 .show_comm = true,
4775                 .show_tstamp = true,
4776                 .show_duration = true,
4777                 .show_arg_names = true,
4778                 .args_alignment = 70,
4779                 .trace_syscalls = false,
4780                 .kernel_syscallchains = false,
4781                 .max_stack = UINT_MAX,
4782                 .max_events = ULONG_MAX,
4783         };
4784         const char *map_dump_str = NULL;
4785         const char *output_name = NULL;
4786         const struct option trace_options[] = {
4787         OPT_CALLBACK('e', "event", &trace, "event",
4788                      "event/syscall selector. use 'perf list' to list available events",
4789                      trace__parse_events_option),
4790         OPT_CALLBACK(0, "filter", &trace.evlist, "filter",
4791                      "event filter", parse_filter),
4792         OPT_BOOLEAN(0, "comm", &trace.show_comm,
4793                     "show the thread COMM next to its id"),
4794         OPT_BOOLEAN(0, "tool_stats", &trace.show_tool_stats, "show tool stats"),
4795         OPT_CALLBACK(0, "expr", &trace, "expr", "list of syscalls/events to trace",
4796                      trace__parse_events_option),
4797         OPT_STRING('o', "output", &output_name, "file", "output file name"),
4798         OPT_STRING('i', "input", &input_name, "file", "Analyze events in file"),
4799         OPT_STRING('p', "pid", &trace.opts.target.pid, "pid",
4800                     "trace events on existing process id"),
4801         OPT_STRING('t', "tid", &trace.opts.target.tid, "tid",
4802                     "trace events on existing thread id"),
4803         OPT_CALLBACK(0, "filter-pids", &trace, "CSV list of pids",
4804                      "pids to filter (by the kernel)", trace__set_filter_pids_from_option),
4805         OPT_BOOLEAN('a', "all-cpus", &trace.opts.target.system_wide,
4806                     "system-wide collection from all CPUs"),
4807         OPT_STRING('C', "cpu", &trace.opts.target.cpu_list, "cpu",
4808                     "list of cpus to monitor"),
4809         OPT_BOOLEAN(0, "no-inherit", &trace.opts.no_inherit,
4810                     "child tasks do not inherit counters"),
4811         OPT_CALLBACK('m', "mmap-pages", &trace.opts.mmap_pages, "pages",
4812                      "number of mmap data pages", evlist__parse_mmap_pages),
4813         OPT_STRING('u', "uid", &trace.opts.target.uid_str, "user",
4814                    "user to profile"),
4815         OPT_CALLBACK(0, "duration", &trace, "float",
4816                      "show only events with duration > N.M ms",
4817                      trace__set_duration),
4818 #ifdef HAVE_LIBBPF_SUPPORT
4819         OPT_STRING(0, "map-dump", &map_dump_str, "BPF map", "BPF map to periodically dump"),
4820 #endif
4821         OPT_BOOLEAN(0, "sched", &trace.sched, "show blocking scheduler events"),
4822         OPT_INCR('v', "verbose", &verbose, "be more verbose"),
4823         OPT_BOOLEAN('T', "time", &trace.full_time,
4824                     "Show full timestamp, not time relative to first start"),
4825         OPT_BOOLEAN(0, "failure", &trace.failure_only,
4826                     "Show only syscalls that failed"),
4827         OPT_BOOLEAN('s', "summary", &trace.summary_only,
4828                     "Show only syscall summary with statistics"),
4829         OPT_BOOLEAN('S', "with-summary", &trace.summary,
4830                     "Show all syscalls and summary with statistics"),
4831         OPT_BOOLEAN(0, "errno-summary", &trace.errno_summary,
4832                     "Show errno stats per syscall, use with -s or -S"),
4833         OPT_CALLBACK_DEFAULT('F', "pf", &trace.trace_pgfaults, "all|maj|min",
4834                      "Trace pagefaults", parse_pagefaults, "maj"),
4835         OPT_BOOLEAN(0, "syscalls", &trace.trace_syscalls, "Trace syscalls"),
4836         OPT_BOOLEAN('f', "force", &trace.force, "don't complain, do it"),
4837         OPT_CALLBACK(0, "call-graph", &trace.opts,
4838                      "record_mode[,record_size]", record_callchain_help,
4839                      &record_parse_callchain_opt),
4840         OPT_BOOLEAN(0, "libtraceevent_print", &trace.libtraceevent_print,
4841                     "Use libtraceevent to print the tracepoint arguments."),
4842         OPT_BOOLEAN(0, "kernel-syscall-graph", &trace.kernel_syscallchains,
4843                     "Show the kernel callchains on the syscall exit path"),
4844         OPT_ULONG(0, "max-events", &trace.max_events,
4845                 "Set the maximum number of events to print, exit after that is reached. "),
4846         OPT_UINTEGER(0, "min-stack", &trace.min_stack,
4847                      "Set the minimum stack depth when parsing the callchain, "
4848                      "anything below the specified depth will be ignored."),
4849         OPT_UINTEGER(0, "max-stack", &trace.max_stack,
4850                      "Set the maximum stack depth when parsing the callchain, "
4851                      "anything beyond the specified depth will be ignored. "
4852                      "Default: kernel.perf_event_max_stack or " __stringify(PERF_MAX_STACK_DEPTH)),
4853         OPT_BOOLEAN(0, "sort-events", &trace.sort_events,
4854                         "Sort batch of events before processing, use if getting out of order events"),
4855         OPT_BOOLEAN(0, "print-sample", &trace.print_sample,
4856                         "print the PERF_RECORD_SAMPLE PERF_SAMPLE_ info, for debugging"),
4857         OPT_UINTEGER(0, "proc-map-timeout", &proc_map_timeout,
4858                         "per thread proc mmap processing timeout in ms"),
4859         OPT_CALLBACK('G', "cgroup", &trace, "name", "monitor event in cgroup name only",
4860                      trace__parse_cgroups),
4861         OPT_INTEGER('D', "delay", &trace.opts.initial_delay,
4862                      "ms to wait before starting measurement after program "
4863                      "start"),
4864         OPTS_EVSWITCH(&trace.evswitch),
4865         OPT_END()
4866         };
4867         bool __maybe_unused max_stack_user_set = true;
4868         bool mmap_pages_user_set = true;
4869         struct evsel *evsel;
4870         const char * const trace_subcommands[] = { "record", NULL };
4871         int err = -1;
4872         char bf[BUFSIZ];
4873
4874         signal(SIGSEGV, sighandler_dump_stack);
4875         signal(SIGFPE, sighandler_dump_stack);
4876         signal(SIGCHLD, sig_handler);
4877         signal(SIGINT, sig_handler);
4878
4879         trace.evlist = evlist__new();
4880         trace.sctbl = syscalltbl__new();
4881
4882         if (trace.evlist == NULL || trace.sctbl == NULL) {
4883                 pr_err("Not enough memory to run!\n");
4884                 err = -ENOMEM;
4885                 goto out;
4886         }
4887
4888         /*
4889          * Parsing .perfconfig may entail creating a BPF event, that may need
4890          * to create BPF maps, so bump RLIM_MEMLOCK as the default 64K setting
4891          * is too small. This affects just this process, not touching the
4892          * global setting. If it fails we'll get something in 'perf trace -v'
4893          * to help diagnose the problem.
4894          */
4895         rlimit__bump_memlock();
4896
4897         err = perf_config(trace__config, &trace);
4898         if (err)
4899                 goto out;
4900
4901         argc = parse_options_subcommand(argc, argv, trace_options, trace_subcommands,
4902                                  trace_usage, PARSE_OPT_STOP_AT_NON_OPTION);
4903
4904         /*
4905          * Here we already passed thru trace__parse_events_option() and it has
4906          * already figured out if -e syscall_name, if not but if --event
4907          * foo:bar was used, the user is interested _just_ in those, say,
4908          * tracepoint events, not in the strace-like syscall-name-based mode.
4909          *
4910          * This is important because we need to check if strace-like mode is
4911          * needed to decided if we should filter out the eBPF
4912          * __augmented_syscalls__ code, if it is in the mix, say, via
4913          * .perfconfig trace.add_events, and filter those out.
4914          */
4915         if (!trace.trace_syscalls && !trace.trace_pgfaults &&
4916             trace.evlist->core.nr_entries == 0 /* Was --events used? */) {
4917                 trace.trace_syscalls = true;
4918         }
4919         /*
4920          * Now that we have --verbose figured out, lets see if we need to parse
4921          * events from .perfconfig, so that if those events fail parsing, say some
4922          * BPF program fails, then we'll be able to use --verbose to see what went
4923          * wrong in more detail.
4924          */
4925         if (trace.perfconfig_events != NULL) {
4926                 struct parse_events_error parse_err;
4927
4928                 bzero(&parse_err, sizeof(parse_err));
4929                 err = parse_events(trace.evlist, trace.perfconfig_events, &parse_err);
4930                 if (err) {
4931                         parse_events_print_error(&parse_err, trace.perfconfig_events);
4932                         goto out;
4933                 }
4934         }
4935
4936         if ((nr_cgroups || trace.cgroup) && !trace.opts.target.system_wide) {
4937                 usage_with_options_msg(trace_usage, trace_options,
4938                                        "cgroup monitoring only available in system-wide mode");
4939         }
4940
4941         evsel = bpf__setup_output_event(trace.evlist, "__augmented_syscalls__");
4942         if (IS_ERR(evsel)) {
4943                 bpf__strerror_setup_output_event(trace.evlist, PTR_ERR(evsel), bf, sizeof(bf));
4944                 pr_err("ERROR: Setup trace syscalls enter failed: %s\n", bf);
4945                 goto out;
4946         }
4947
4948         if (evsel) {
4949                 trace.syscalls.events.augmented = evsel;
4950
4951                 evsel = evlist__find_tracepoint_by_name(trace.evlist, "raw_syscalls:sys_enter");
4952                 if (evsel == NULL) {
4953                         pr_err("ERROR: raw_syscalls:sys_enter not found in the augmented BPF object\n");
4954                         goto out;
4955                 }
4956
4957                 if (evsel->bpf_obj == NULL) {
4958                         pr_err("ERROR: raw_syscalls:sys_enter not associated to a BPF object\n");
4959                         goto out;
4960                 }
4961
4962                 trace.bpf_obj = evsel->bpf_obj;
4963
4964                 /*
4965                  * If we have _just_ the augmenter event but don't have a
4966                  * explicit --syscalls, then assume we want all strace-like
4967                  * syscalls:
4968                  */
4969                 if (!trace.trace_syscalls && trace__only_augmented_syscalls_evsels(&trace))
4970                         trace.trace_syscalls = true;
4971                 /*
4972                  * So, if we have a syscall augmenter, but trace_syscalls, aka
4973                  * strace-like syscall tracing is not set, then we need to trow
4974                  * away the augmenter, i.e. all the events that were created
4975                  * from that BPF object file.
4976                  *
4977                  * This is more to fix the current .perfconfig trace.add_events
4978                  * style of setting up the strace-like eBPF based syscall point
4979                  * payload augmenter.
4980                  *
4981                  * All this complexity will be avoided by adding an alternative
4982                  * to trace.add_events in the form of
4983                  * trace.bpf_augmented_syscalls, that will be only parsed if we
4984                  * need it.
4985                  *
4986                  * .perfconfig trace.add_events is still useful if we want, for
4987                  * instance, have msr_write.msr in some .perfconfig profile based
4988                  * 'perf trace --config determinism.profile' mode, where for some
4989                  * particular goal/workload type we want a set of events and
4990                  * output mode (with timings, etc) instead of having to add
4991                  * all via the command line.
4992                  *
4993                  * Also --config to specify an alternate .perfconfig file needs
4994                  * to be implemented.
4995                  */
4996                 if (!trace.trace_syscalls) {
4997                         trace__delete_augmented_syscalls(&trace);
4998                 } else {
4999                         trace__set_bpf_map_filtered_pids(&trace);
5000                         trace__set_bpf_map_syscalls(&trace);
5001                         trace.syscalls.unaugmented_prog = trace__find_bpf_program_by_title(&trace, "!raw_syscalls:unaugmented");
5002                 }
5003         }
5004
5005         err = bpf__setup_stdout(trace.evlist);
5006         if (err) {
5007                 bpf__strerror_setup_stdout(trace.evlist, err, bf, sizeof(bf));
5008                 pr_err("ERROR: Setup BPF stdout failed: %s\n", bf);
5009                 goto out;
5010         }
5011
5012         err = -1;
5013
5014         if (map_dump_str) {
5015                 trace.dump.map = trace__find_bpf_map_by_name(&trace, map_dump_str);
5016                 if (trace.dump.map == NULL) {
5017                         pr_err("ERROR: BPF map \"%s\" not found\n", map_dump_str);
5018                         goto out;
5019                 }
5020         }
5021
5022         if (trace.trace_pgfaults) {
5023                 trace.opts.sample_address = true;
5024                 trace.opts.sample_time = true;
5025         }
5026
5027         if (trace.opts.mmap_pages == UINT_MAX)
5028                 mmap_pages_user_set = false;
5029
5030         if (trace.max_stack == UINT_MAX) {
5031                 trace.max_stack = input_name ? PERF_MAX_STACK_DEPTH : sysctl__max_stack();
5032                 max_stack_user_set = false;
5033         }
5034
5035 #ifdef HAVE_DWARF_UNWIND_SUPPORT
5036         if ((trace.min_stack || max_stack_user_set) && !callchain_param.enabled) {
5037                 record_opts__parse_callchain(&trace.opts, &callchain_param, "dwarf", false);
5038         }
5039 #endif
5040
5041         if (callchain_param.enabled) {
5042                 if (!mmap_pages_user_set && geteuid() == 0)
5043                         trace.opts.mmap_pages = perf_event_mlock_kb_in_pages() * 4;
5044
5045                 symbol_conf.use_callchain = true;
5046         }
5047
5048         if (trace.evlist->core.nr_entries > 0) {
5049                 evlist__set_default_evsel_handler(trace.evlist, trace__event_handler);
5050                 if (evlist__set_syscall_tp_fields(trace.evlist)) {
5051                         perror("failed to set syscalls:* tracepoint fields");
5052                         goto out;
5053                 }
5054         }
5055
5056         if (trace.sort_events) {
5057                 ordered_events__init(&trace.oe.data, ordered_events__deliver_event, &trace);
5058                 ordered_events__set_copy_on_queue(&trace.oe.data, true);
5059         }
5060
5061         /*
5062          * If we are augmenting syscalls, then combine what we put in the
5063          * __augmented_syscalls__ BPF map with what is in the
5064          * syscalls:sys_exit_FOO tracepoints, i.e. just like we do without BPF,
5065          * combining raw_syscalls:sys_enter with raw_syscalls:sys_exit.
5066          *
5067          * We'll switch to look at two BPF maps, one for sys_enter and the
5068          * other for sys_exit when we start augmenting the sys_exit paths with
5069          * buffers that are being copied from kernel to userspace, think 'read'
5070          * syscall.
5071          */
5072         if (trace.syscalls.events.augmented) {
5073                 evlist__for_each_entry(trace.evlist, evsel) {
5074                         bool raw_syscalls_sys_exit = strcmp(evsel__name(evsel), "raw_syscalls:sys_exit") == 0;
5075
5076                         if (raw_syscalls_sys_exit) {
5077                                 trace.raw_augmented_syscalls = true;
5078                                 goto init_augmented_syscall_tp;
5079                         }
5080
5081                         if (trace.syscalls.events.augmented->priv == NULL &&
5082                             strstr(evsel__name(evsel), "syscalls:sys_enter")) {
5083                                 struct evsel *augmented = trace.syscalls.events.augmented;
5084                                 if (evsel__init_augmented_syscall_tp(augmented, evsel) ||
5085                                     evsel__init_augmented_syscall_tp_args(augmented))
5086                                         goto out;
5087                                 /*
5088                                  * Augmented is __augmented_syscalls__ BPF_OUTPUT event
5089                                  * Above we made sure we can get from the payload the tp fields
5090                                  * that we get from syscalls:sys_enter tracefs format file.
5091                                  */
5092                                 augmented->handler = trace__sys_enter;
5093                                 /*
5094                                  * Now we do the same for the *syscalls:sys_enter event so that
5095                                  * if we handle it directly, i.e. if the BPF prog returns 0 so
5096                                  * as not to filter it, then we'll handle it just like we would
5097                                  * for the BPF_OUTPUT one:
5098                                  */
5099                                 if (evsel__init_augmented_syscall_tp(evsel, evsel) ||
5100                                     evsel__init_augmented_syscall_tp_args(evsel))
5101                                         goto out;
5102                                 evsel->handler = trace__sys_enter;
5103                         }
5104
5105                         if (strstarts(evsel__name(evsel), "syscalls:sys_exit_")) {
5106                                 struct syscall_tp *sc;
5107 init_augmented_syscall_tp:
5108                                 if (evsel__init_augmented_syscall_tp(evsel, evsel))
5109                                         goto out;
5110                                 sc = __evsel__syscall_tp(evsel);
5111                                 /*
5112                                  * For now with BPF raw_augmented we hook into
5113                                  * raw_syscalls:sys_enter and there we get all
5114                                  * 6 syscall args plus the tracepoint common
5115                                  * fields and the syscall_nr (another long).
5116                                  * So we check if that is the case and if so
5117                                  * don't look after the sc->args_size but
5118                                  * always after the full raw_syscalls:sys_enter
5119                                  * payload, which is fixed.
5120                                  *
5121                                  * We'll revisit this later to pass
5122                                  * s->args_size to the BPF augmenter (now
5123                                  * tools/perf/examples/bpf/augmented_raw_syscalls.c,
5124                                  * so that it copies only what we need for each
5125                                  * syscall, like what happens when we use
5126                                  * syscalls:sys_enter_NAME, so that we reduce
5127                                  * the kernel/userspace traffic to just what is
5128                                  * needed for each syscall.
5129                                  */
5130                                 if (trace.raw_augmented_syscalls)
5131                                         trace.raw_augmented_syscalls_args_size = (6 + 1) * sizeof(long) + sc->id.offset;
5132                                 evsel__init_augmented_syscall_tp_ret(evsel);
5133                                 evsel->handler = trace__sys_exit;
5134                         }
5135                 }
5136         }
5137
5138         if ((argc >= 1) && (strcmp(argv[0], "record") == 0))
5139                 return trace__record(&trace, argc-1, &argv[1]);
5140
5141         /* Using just --errno-summary will trigger --summary */
5142         if (trace.errno_summary && !trace.summary && !trace.summary_only)
5143                 trace.summary_only = true;
5144
5145         /* summary_only implies summary option, but don't overwrite summary if set */
5146         if (trace.summary_only)
5147                 trace.summary = trace.summary_only;
5148
5149         if (output_name != NULL) {
5150                 err = trace__open_output(&trace, output_name);
5151                 if (err < 0) {
5152                         perror("failed to create output file");
5153                         goto out;
5154                 }
5155         }
5156
5157         err = evswitch__init(&trace.evswitch, trace.evlist, stderr);
5158         if (err)
5159                 goto out_close;
5160
5161         err = target__validate(&trace.opts.target);
5162         if (err) {
5163                 target__strerror(&trace.opts.target, err, bf, sizeof(bf));
5164                 fprintf(trace.output, "%s", bf);
5165                 goto out_close;
5166         }
5167
5168         err = target__parse_uid(&trace.opts.target);
5169         if (err) {
5170                 target__strerror(&trace.opts.target, err, bf, sizeof(bf));
5171                 fprintf(trace.output, "%s", bf);
5172                 goto out_close;
5173         }
5174
5175         if (!argc && target__none(&trace.opts.target))
5176                 trace.opts.target.system_wide = true;
5177
5178         if (input_name)
5179                 err = trace__replay(&trace);
5180         else
5181                 err = trace__run(&trace, argc, argv);
5182
5183 out_close:
5184         if (output_name != NULL)
5185                 fclose(trace.output);
5186 out:
5187         trace__exit(&trace);
5188         return err;
5189 }