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 #define RAW_SYSCALL_ARGS_NUM    6
  91
  92 /*
  93  * strtoul: Go from a string to a value, i.e. for msr: MSR_FS_BASE to 0xc0000100
  94  */
  95 struct syscall_arg_fmt {
  96         size_t     (*scnprintf)(char *bf, size_t size, struct syscall_arg *arg);
  97         bool       (*strtoul)(char *bf, size_t size, struct syscall_arg *arg, u64 *val);
  98         unsigned long (*mask_val)(struct syscall_arg *arg, unsigned long val);
  99         void       *parm;
 100         const char *name;
 101         u16        nr_entries; // for arrays
 102         bool       show_zero;
 103 };
 104
 105 struct syscall_fmt {
 106         const char *name;
 107         const char *alias;
 108         struct {
 109                 const char *sys_enter,
 110                            *sys_exit;
 111         }          bpf_prog_name;
 112         struct syscall_arg_fmt arg[RAW_SYSCALL_ARGS_NUM];
 113         u8         nr_args;
 114         bool       errpid;
 115         bool       timeout;
 116         bool       hexret;
 117 };
 118
 119 struct trace {
 120         struct perf_tool        tool;
 121         struct syscalltbl       *sctbl;
 122         struct {
 123                 struct syscall  *table;
 124                 struct bpf_map  *map;
 125                 struct { // per syscall BPF_MAP_TYPE_PROG_ARRAY
 126                         struct bpf_map  *sys_enter,
 127                                         *sys_exit;
 128                 }               prog_array;
 129                 struct {
 130                         struct evsel *sys_enter,
 131                                           *sys_exit,
 132                                           *augmented;
 133                 }               events;
 134                 struct bpf_program *unaugmented_prog;
 135         } syscalls;
 136         struct {
 137                 struct bpf_map *map;
 138         } dump;
 139         struct record_opts      opts;
 140         struct evlist   *evlist;
 141         struct machine          *host;
 142         struct thread           *current;
 143         struct bpf_object       *bpf_obj;
 144         struct cgroup           *cgroup;
 145         u64                     base_time;
 146         FILE                    *output;
 147         unsigned long           nr_events;
 148         unsigned long           nr_events_printed;
 149         unsigned long           max_events;
 150         struct evswitch         evswitch;
 151         struct strlist          *ev_qualifier;
 152         struct {
 153                 size_t          nr;
 154                 int             *entries;
 155         }                       ev_qualifier_ids;
 156         struct {
 157                 size_t          nr;
 158                 pid_t           *entries;
 159                 struct bpf_map  *map;
 160         }                       filter_pids;
 161         double                  duration_filter;
 162         double                  runtime_ms;
 163         struct {
 164                 u64             vfs_getname,
 165                                 proc_getname;
 166         } stats;
 167         unsigned int            max_stack;
 168         unsigned int            min_stack;
 169         int                     raw_augmented_syscalls_args_size;
 170         bool                    raw_augmented_syscalls;
 171         bool                    fd_path_disabled;
 172         bool                    sort_events;
 173         bool                    not_ev_qualifier;
 174         bool                    live;
 175         bool                    full_time;
 176         bool                    sched;
 177         bool                    multiple_threads;
 178         bool                    summary;
 179         bool                    summary_only;
 180         bool                    errno_summary;
 181         bool                    failure_only;
 182         bool                    show_comm;
 183         bool                    print_sample;
 184         bool                    show_tool_stats;
 185         bool                    trace_syscalls;
 186         bool                    libtraceevent_print;
 187         bool                    kernel_syscallchains;
 188         s16                     args_alignment;
 189         bool                    show_tstamp;
 190         bool                    show_duration;
 191         bool                    show_zeros;
 192         bool                    show_arg_names;
 193         bool                    show_string_prefix;
 194         bool                    force;
 195         bool                    vfs_getname;
 196         int                     trace_pgfaults;
 197         char                    *perfconfig_events;
 198         struct {
 199                 struct ordered_events   data;
 200                 u64                     last;
 201         } oe;
 202 };
 203
 204 struct tp_field {
 205         int offset;
 206         union {
 207                 u64 (*integer)(struct tp_field *field, struct perf_sample *sample);
 208                 void *(*pointer)(struct tp_field *field, struct perf_sample *sample);
 209         };
 210 };
 211
 212 #define TP_UINT_FIELD(bits) \
 213 static u64 tp_field__u##bits(struct tp_field *field, struct perf_sample *sample) \
 214 { \
 215         u##bits value; \
 216         memcpy(&value, sample->raw_data + field->offset, sizeof(value)); \
 217         return value;  \
 218 }
 219
 220 TP_UINT_FIELD(8);
 221 TP_UINT_FIELD(16);
 222 TP_UINT_FIELD(32);
 223 TP_UINT_FIELD(64);
 224
 225 #define TP_UINT_FIELD__SWAPPED(bits) \
 226 static u64 tp_field__swapped_u##bits(struct tp_field *field, struct perf_sample *sample) \
 227 { \
 228         u##bits value; \
 229         memcpy(&value, sample->raw_data + field->offset, sizeof(value)); \
 230         return bswap_##bits(value);\
 231 }
 232
 233 TP_UINT_FIELD__SWAPPED(16);
 234 TP_UINT_FIELD__SWAPPED(32);
 235 TP_UINT_FIELD__SWAPPED(64);
 236
 237 static int __tp_field__init_uint(struct tp_field *field, int size, int offset, bool needs_swap)
 238 {
 239         field->offset = offset;
 240
 241         switch (size) {
 242         case 1:
 243                 field->integer = tp_field__u8;
 244                 break;
 245         case 2:
 246                 field->integer = needs_swap ? tp_field__swapped_u16 : tp_field__u16;
 247                 break;
 248         case 4:
 249                 field->integer = needs_swap ? tp_field__swapped_u32 : tp_field__u32;
 250                 break;
 251         case 8:
 252                 field->integer = needs_swap ? tp_field__swapped_u64 : tp_field__u64;
 253                 break;
 254         default:
 255                 return -1;
 256         }
 257
 258         return 0;
 259 }
 260
 261 static int tp_field__init_uint(struct tp_field *field, struct tep_format_field *format_field, bool needs_swap)
 262 {
 263         return __tp_field__init_uint(field, format_field->size, format_field->offset, needs_swap);
 264 }
 265
 266 static void *tp_field__ptr(struct tp_field *field, struct perf_sample *sample)
 267 {
 268         return sample->raw_data + field->offset;
 269 }
 270
 271 static int __tp_field__init_ptr(struct tp_field *field, int offset)
 272 {
 273         field->offset = offset;
 274         field->pointer = tp_field__ptr;
 275         return 0;
 276 }
 277
 278 static int tp_field__init_ptr(struct tp_field *field, struct tep_format_field *format_field)
 279 {
 280         return __tp_field__init_ptr(field, format_field->offset);
 281 }
 282
 283 struct syscall_tp {
 284         struct tp_field id;
 285         union {
 286                 struct tp_field args, ret;
 287         };
 288 };
 289
 290 /*
 291  * The evsel->priv as used by 'perf trace'
 292  * sc:  for raw_syscalls:sys_{enter,exit} and syscalls:sys_{enter,exit}_SYSCALLNAME
 293  * fmt: for all the other tracepoints
 294  */
 295 struct evsel_trace {
 296         struct syscall_tp       sc;
 297         struct syscall_arg_fmt  *fmt;
 298 };
 299
 300 static struct evsel_trace *evsel_trace__new(void)
 301 {
 302         return zalloc(sizeof(struct evsel_trace));
 303 }
 304
 305 static void evsel_trace__delete(struct evsel_trace *et)
 306 {
 307         if (et == NULL)
 308                 return;
 309
 310         zfree(&et->fmt);
 311         free(et);
 312 }
 313
 314 /*
 315  * Used with raw_syscalls:sys_{enter,exit} and with the
 316  * syscalls:sys_{enter,exit}_SYSCALL tracepoints
 317  */
 318 static inline struct syscall_tp *__evsel__syscall_tp(struct evsel *evsel)
 319 {
 320         struct evsel_trace *et = evsel->priv;
 321
 322         return &et->sc;
 323 }
 324
 325 static struct syscall_tp *evsel__syscall_tp(struct evsel *evsel)
 326 {
 327         if (evsel->priv == NULL) {
 328                 evsel->priv = evsel_trace__new();
 329                 if (evsel->priv == NULL)
 330                         return NULL;
 331         }
 332
 333         return __evsel__syscall_tp(evsel);
 334 }
 335
 336 /*
 337  * Used with all the other tracepoints.
 338  */
 339 static inline struct syscall_arg_fmt *__evsel__syscall_arg_fmt(struct evsel *evsel)
 340 {
 341         struct evsel_trace *et = evsel->priv;
 342
 343         return et->fmt;
 344 }
 345
 346 static struct syscall_arg_fmt *evsel__syscall_arg_fmt(struct evsel *evsel)
 347 {
 348         struct evsel_trace *et = evsel->priv;
 349
 350         if (evsel->priv == NULL) {
 351                 et = evsel->priv = evsel_trace__new();
 352
 353                 if (et == NULL)
 354                         return NULL;
 355         }
 356
 357         if (et->fmt == NULL) {
 358                 et->fmt = calloc(evsel->tp_format->format.nr_fields, sizeof(struct syscall_arg_fmt));
 359                 if (et->fmt == NULL)
 360                         goto out_delete;
 361         }
 362
 363         return __evsel__syscall_arg_fmt(evsel);
 364
 365 out_delete:
 366         evsel_trace__delete(evsel->priv);
 367         evsel->priv = NULL;
 368         return NULL;
 369 }
 370
 371 static int evsel__init_tp_uint_field(struct evsel *evsel, struct tp_field *field, const char *name)
 372 {
 373         struct tep_format_field *format_field = evsel__field(evsel, name);
 374
 375         if (format_field == NULL)
 376                 return -1;
 377
 378         return tp_field__init_uint(field, format_field, evsel->needs_swap);
 379 }
 380
 381 #define perf_evsel__init_sc_tp_uint_field(evsel, name) \
 382         ({ struct syscall_tp *sc = __evsel__syscall_tp(evsel);\
 383            evsel__init_tp_uint_field(evsel, &sc->name, #name); })
 384
 385 static int evsel__init_tp_ptr_field(struct evsel *evsel, struct tp_field *field, const char *name)
 386 {
 387         struct tep_format_field *format_field = evsel__field(evsel, name);
 388
 389         if (format_field == NULL)
 390                 return -1;
 391
 392         return tp_field__init_ptr(field, format_field);
 393 }
 394
 395 #define perf_evsel__init_sc_tp_ptr_field(evsel, name) \
 396         ({ struct syscall_tp *sc = __evsel__syscall_tp(evsel);\
 397            evsel__init_tp_ptr_field(evsel, &sc->name, #name); })
 398
 399 static void evsel__delete_priv(struct evsel *evsel)
 400 {
 401         zfree(&evsel->priv);
 402         evsel__delete(evsel);
 403 }
 404
 405 static int evsel__init_syscall_tp(struct evsel *evsel)
 406 {
 407         struct syscall_tp *sc = evsel__syscall_tp(evsel);
 408
 409         if (sc != NULL) {
 410                 if (evsel__init_tp_uint_field(evsel, &sc->id, "__syscall_nr") &&
 411                     evsel__init_tp_uint_field(evsel, &sc->id, "nr"))
 412                         return -ENOENT;
 413                 return 0;
 414         }
 415
 416         return -ENOMEM;
 417 }
 418
 419 static int evsel__init_augmented_syscall_tp(struct evsel *evsel, struct evsel *tp)
 420 {
 421         struct syscall_tp *sc = evsel__syscall_tp(evsel);
 422
 423         if (sc != NULL) {
 424                 struct tep_format_field *syscall_id = evsel__field(tp, "id");
 425                 if (syscall_id == NULL)
 426                         syscall_id = evsel__field(tp, "__syscall_nr");
 427                 if (syscall_id == NULL ||
 428                     __tp_field__init_uint(&sc->id, syscall_id->size, syscall_id->offset, evsel->needs_swap))
 429                         return -EINVAL;
 430
 431                 return 0;
 432         }
 433
 434         return -ENOMEM;
 435 }
 436
 437 static int evsel__init_augmented_syscall_tp_args(struct evsel *evsel)
 438 {
 439         struct syscall_tp *sc = __evsel__syscall_tp(evsel);
 440
 441         return __tp_field__init_ptr(&sc->args, sc->id.offset + sizeof(u64));
 442 }
 443
 444 static int evsel__init_augmented_syscall_tp_ret(struct evsel *evsel)
 445 {
 446         struct syscall_tp *sc = __evsel__syscall_tp(evsel);
 447
 448         return __tp_field__init_uint(&sc->ret, sizeof(u64), sc->id.offset + sizeof(u64), evsel->needs_swap);
 449 }
 450
 451 static int evsel__init_raw_syscall_tp(struct evsel *evsel, void *handler)
 452 {
 453         if (evsel__syscall_tp(evsel) != NULL) {
 454                 if (perf_evsel__init_sc_tp_uint_field(evsel, id))
 455                         return -ENOENT;
 456
 457                 evsel->handler = handler;
 458                 return 0;
 459         }
 460
 461         return -ENOMEM;
 462 }
 463
 464 static struct evsel *perf_evsel__raw_syscall_newtp(const char *direction, void *handler)
 465 {
 466         struct evsel *evsel = evsel__newtp("raw_syscalls", direction);
 467
 468         /* older kernel (e.g., RHEL6) use syscalls:{enter,exit} */
 469         if (IS_ERR(evsel))
 470                 evsel = evsel__newtp("syscalls", direction);
 471
 472         if (IS_ERR(evsel))
 473                 return NULL;
 474
 475         if (evsel__init_raw_syscall_tp(evsel, handler))
 476                 goto out_delete;
 477
 478         return evsel;
 479
 480 out_delete:
 481         evsel__delete_priv(evsel);
 482         return NULL;
 483 }
 484
 485 #define perf_evsel__sc_tp_uint(evsel, name, sample) \
 486         ({ struct syscall_tp *fields = __evsel__syscall_tp(evsel); \
 487            fields->name.integer(&fields->name, sample); })
 488
 489 #define perf_evsel__sc_tp_ptr(evsel, name, sample) \
 490         ({ struct syscall_tp *fields = __evsel__syscall_tp(evsel); \
 491            fields->name.pointer(&fields->name, sample); })
 492
 493 size_t strarray__scnprintf_suffix(struct strarray *sa, char *bf, size_t size, const char *intfmt, bool show_suffix, int val)
 494 {
 495         int idx = val - sa->offset;
 496
 497         if (idx < 0 || idx >= sa->nr_entries || sa->entries[idx] == NULL) {
 498                 size_t printed = scnprintf(bf, size, intfmt, val);
 499                 if (show_suffix)
 500                         printed += scnprintf(bf + printed, size - printed, " /* %s??? */", sa->prefix);
 501                 return printed;
 502         }
 503
 504         return scnprintf(bf, size, "%s%s", sa->entries[idx], show_suffix ? sa->prefix : "");
 505 }
 506
 507 size_t strarray__scnprintf(struct strarray *sa, char *bf, size_t size, const char *intfmt, bool show_prefix, int val)
 508 {
 509         int idx = val - sa->offset;
 510
 511         if (idx < 0 || idx >= sa->nr_entries || sa->entries[idx] == NULL) {
 512                 size_t printed = scnprintf(bf, size, intfmt, val);
 513                 if (show_prefix)
 514                         printed += scnprintf(bf + printed, size - printed, " /* %s??? */", sa->prefix);
 515                 return printed;
 516         }
 517
 518         return scnprintf(bf, size, "%s%s", show_prefix ? sa->prefix : "", sa->entries[idx]);
 519 }
 520
 521 static size_t __syscall_arg__scnprintf_strarray(char *bf, size_t size,
 522                                                 const char *intfmt,
 523                                                 struct syscall_arg *arg)
 524 {
 525         return strarray__scnprintf(arg->parm, bf, size, intfmt, arg->show_string_prefix, arg->val);
 526 }
 527
 528 static size_t syscall_arg__scnprintf_strarray(char *bf, size_t size,
 529                                               struct syscall_arg *arg)
 530 {
 531         return __syscall_arg__scnprintf_strarray(bf, size, "%d", arg);
 532 }
 533
 534 #define SCA_STRARRAY syscall_arg__scnprintf_strarray
 535
 536 bool syscall_arg__strtoul_strarray(char *bf, size_t size, struct syscall_arg *arg, u64 *ret)
 537 {
 538         return strarray__strtoul(arg->parm, bf, size, ret);
 539 }
 540
 541 bool syscall_arg__strtoul_strarray_flags(char *bf, size_t size, struct syscall_arg *arg, u64 *ret)
 542 {
 543         return strarray__strtoul_flags(arg->parm, bf, size, ret);
 544 }
 545
 546 bool syscall_arg__strtoul_strarrays(char *bf, size_t size, struct syscall_arg *arg, u64 *ret)
 547 {
 548         return strarrays__strtoul(arg->parm, bf, size, ret);
 549 }
 550
 551 size_t syscall_arg__scnprintf_strarray_flags(char *bf, size_t size, struct syscall_arg *arg)
 552 {
 553         return strarray__scnprintf_flags(arg->parm, bf, size, arg->show_string_prefix, arg->val);
 554 }
 555
 556 size_t strarrays__scnprintf(struct strarrays *sas, char *bf, size_t size, const char *intfmt, bool show_prefix, int val)
 557 {
 558         size_t printed;
 559         int i;
 560
 561         for (i = 0; i < sas->nr_entries; ++i) {
 562                 struct strarray *sa = sas->entries[i];
 563                 int idx = val - sa->offset;
 564
 565                 if (idx >= 0 && idx < sa->nr_entries) {
 566                         if (sa->entries[idx] == NULL)
 567                                 break;
 568                         return scnprintf(bf, size, "%s%s", show_prefix ? sa->prefix : "", sa->entries[idx]);
 569                 }
 570         }
 571
 572         printed = scnprintf(bf, size, intfmt, val);
 573         if (show_prefix)
 574                 printed += scnprintf(bf + printed, size - printed, " /* %s??? */", sas->entries[0]->prefix);
 575         return printed;
 576 }
 577
 578 bool strarray__strtoul(struct strarray *sa, char *bf, size_t size, u64 *ret)
 579 {
 580         int i;
 581
 582         for (i = 0; i < sa->nr_entries; ++i) {
 583                 if (sa->entries[i] && strncmp(sa->entries[i], bf, size) == 0 && sa->entries[i][size] == '\0') {
 584                         *ret = sa->offset + i;
 585                         return true;
 586                 }
 587         }
 588
 589         return false;
 590 }
 591
 592 bool strarray__strtoul_flags(struct strarray *sa, char *bf, size_t size, u64 *ret)
 593 {
 594         u64 val = 0;
 595         char *tok = bf, *sep, *end;
 596
 597         *ret = 0;
 598
 599         while (size != 0) {
 600                 int toklen = size;
 601
 602                 sep = memchr(tok, '|', size);
 603                 if (sep != NULL) {
 604                         size -= sep - tok + 1;
 605
 606                         end = sep - 1;
 607                         while (end > tok && isspace(*end))
 608                                 --end;
 609
 610                         toklen = end - tok + 1;
 611                 }
 612
 613                 while (isspace(*tok))
 614                         ++tok;
 615
 616                 if (isalpha(*tok) || *tok == '_') {
 617                         if (!strarray__strtoul(sa, tok, toklen, &val))
 618                                 return false;
 619                 } else {
 620                         bool is_hexa = tok[0] == 0 && (tok[1] = 'x' || tok[1] == 'X');
 621
 622                         val = strtoul(tok, NULL, is_hexa ? 16 : 0);
 623                 }
 624
 625                 *ret |= (1 << (val - 1));
 626
 627                 if (sep == NULL)
 628                         break;
 629                 tok = sep + 1;
 630         }
 631
 632         return true;
 633 }
 634
 635 bool strarrays__strtoul(struct strarrays *sas, char *bf, size_t size, u64 *ret)
 636 {
 637         int i;
 638
 639         for (i = 0; i < sas->nr_entries; ++i) {
 640                 struct strarray *sa = sas->entries[i];
 641
 642                 if (strarray__strtoul(sa, bf, size, ret))
 643                         return true;
 644         }
 645
 646         return false;
 647 }
 648
 649 size_t syscall_arg__scnprintf_strarrays(char *bf, size_t size,
 650                                         struct syscall_arg *arg)
 651 {
 652         return strarrays__scnprintf(arg->parm, bf, size, "%d", arg->show_string_prefix, arg->val);
 653 }
 654
 655 #ifndef AT_FDCWD
 656 #define AT_FDCWD        -100
 657 #endif
 658
 659 static size_t syscall_arg__scnprintf_fd_at(char *bf, size_t size,
 660                                            struct syscall_arg *arg)
 661 {
 662         int fd = arg->val;
 663         const char *prefix = "AT_FD";
 664
 665         if (fd == AT_FDCWD)
 666                 return scnprintf(bf, size, "%s%s", arg->show_string_prefix ? prefix : "", "CWD");
 667
 668         return syscall_arg__scnprintf_fd(bf, size, arg);
 669 }
 670
 671 #define SCA_FDAT syscall_arg__scnprintf_fd_at
 672
 673 static size_t syscall_arg__scnprintf_close_fd(char *bf, size_t size,
 674                                               struct syscall_arg *arg);
 675
 676 #define SCA_CLOSE_FD syscall_arg__scnprintf_close_fd
 677
 678 size_t syscall_arg__scnprintf_hex(char *bf, size_t size, struct syscall_arg *arg)
 679 {
 680         return scnprintf(bf, size, "%#lx", arg->val);
 681 }
 682
 683 size_t syscall_arg__scnprintf_ptr(char *bf, size_t size, struct syscall_arg *arg)
 684 {
 685         if (arg->val == 0)
 686                 return scnprintf(bf, size, "NULL");
 687         return syscall_arg__scnprintf_hex(bf, size, arg);
 688 }
 689
 690 size_t syscall_arg__scnprintf_int(char *bf, size_t size, struct syscall_arg *arg)
 691 {
 692         return scnprintf(bf, size, "%d", arg->val);
 693 }
 694
 695 size_t syscall_arg__scnprintf_long(char *bf, size_t size, struct syscall_arg *arg)
 696 {
 697         return scnprintf(bf, size, "%ld", arg->val);
 698 }
 699
 700 static size_t syscall_arg__scnprintf_char_array(char *bf, size_t size, struct syscall_arg *arg)
 701 {
 702         // XXX Hey, maybe for sched:sched_switch prev/next comm fields we can
 703         //     fill missing comms using thread__set_comm()...
 704         //     here or in a special syscall_arg__scnprintf_pid_sched_tp...
 705         return scnprintf(bf, size, "\"%-.*s\"", arg->fmt->nr_entries ?: arg->len, arg->val);
 706 }
 707
 708 #define SCA_CHAR_ARRAY syscall_arg__scnprintf_char_array
 709
 710 static const char *bpf_cmd[] = {
 711         "MAP_CREATE", "MAP_LOOKUP_ELEM", "MAP_UPDATE_ELEM", "MAP_DELETE_ELEM",
 712         "MAP_GET_NEXT_KEY", "PROG_LOAD",
 713 };
 714 static DEFINE_STRARRAY(bpf_cmd, "BPF_");
 715
 716 static const char *fsmount_flags[] = {
 717         [1] = "CLOEXEC",
 718 };
 719 static DEFINE_STRARRAY(fsmount_flags, "FSMOUNT_");
 720
 721 #include "trace/beauty/generated/fsconfig_arrays.c"
 722
 723 static DEFINE_STRARRAY(fsconfig_cmds, "FSCONFIG_");
 724
 725 static const char *epoll_ctl_ops[] = { "ADD", "DEL", "MOD", };
 726 static DEFINE_STRARRAY_OFFSET(epoll_ctl_ops, "EPOLL_CTL_", 1);
 727
 728 static const char *itimers[] = { "REAL", "VIRTUAL", "PROF", };
 729 static DEFINE_STRARRAY(itimers, "ITIMER_");
 730
 731 static const char *keyctl_options[] = {
 732         "GET_KEYRING_ID", "JOIN_SESSION_KEYRING", "UPDATE", "REVOKE", "CHOWN",
 733         "SETPERM", "DESCRIBE", "CLEAR", "LINK", "UNLINK", "SEARCH", "READ",
 734         "INSTANTIATE", "NEGATE", "SET_REQKEY_KEYRING", "SET_TIMEOUT",
 735         "ASSUME_AUTHORITY", "GET_SECURITY", "SESSION_TO_PARENT", "REJECT",
 736         "INSTANTIATE_IOV", "INVALIDATE", "GET_PERSISTENT",
 737 };
 738 static DEFINE_STRARRAY(keyctl_options, "KEYCTL_");
 739
 740 static const char *whences[] = { "SET", "CUR", "END",
 741 #ifdef SEEK_DATA
 742 "DATA",
 743 #endif
 744 #ifdef SEEK_HOLE
 745 "HOLE",
 746 #endif
 747 };
 748 static DEFINE_STRARRAY(whences, "SEEK_");
 749
 750 static const char *fcntl_cmds[] = {
 751         "DUPFD", "GETFD", "SETFD", "GETFL", "SETFL", "GETLK", "SETLK",
 752         "SETLKW", "SETOWN", "GETOWN", "SETSIG", "GETSIG", "GETLK64",
 753         "SETLK64", "SETLKW64", "SETOWN_EX", "GETOWN_EX",
 754         "GETOWNER_UIDS",
 755 };
 756 static DEFINE_STRARRAY(fcntl_cmds, "F_");
 757
 758 static const char *fcntl_linux_specific_cmds[] = {
 759         "SETLEASE", "GETLEASE", "NOTIFY", [5] = "CANCELLK", "DUPFD_CLOEXEC",
 760         "SETPIPE_SZ", "GETPIPE_SZ", "ADD_SEALS", "GET_SEALS",
 761         "GET_RW_HINT", "SET_RW_HINT", "GET_FILE_RW_HINT", "SET_FILE_RW_HINT",
 762 };
 763
 764 static DEFINE_STRARRAY_OFFSET(fcntl_linux_specific_cmds, "F_", F_LINUX_SPECIFIC_BASE);
 765
 766 static struct strarray *fcntl_cmds_arrays[] = {
 767         &strarray__fcntl_cmds,
 768         &strarray__fcntl_linux_specific_cmds,
 769 };
 770
 771 static DEFINE_STRARRAYS(fcntl_cmds_arrays);
 772
 773 static const char *rlimit_resources[] = {
 774         "CPU", "FSIZE", "DATA", "STACK", "CORE", "RSS", "NPROC", "NOFILE",
 775         "MEMLOCK", "AS", "LOCKS", "SIGPENDING", "MSGQUEUE", "NICE", "RTPRIO",
 776         "RTTIME",
 777 };
 778 static DEFINE_STRARRAY(rlimit_resources, "RLIMIT_");
 779
 780 static const char *sighow[] = { "BLOCK", "UNBLOCK", "SETMASK", };
 781 static DEFINE_STRARRAY(sighow, "SIG_");
 782
 783 static const char *clockid[] = {
 784         "REALTIME", "MONOTONIC", "PROCESS_CPUTIME_ID", "THREAD_CPUTIME_ID",
 785         "MONOTONIC_RAW", "REALTIME_COARSE", "MONOTONIC_COARSE", "BOOTTIME",
 786         "REALTIME_ALARM", "BOOTTIME_ALARM", "SGI_CYCLE", "TAI"
 787 };
 788 static DEFINE_STRARRAY(clockid, "CLOCK_");
 789
 790 static size_t syscall_arg__scnprintf_access_mode(char *bf, size_t size,
 791                                                  struct syscall_arg *arg)
 792 {
 793         bool show_prefix = arg->show_string_prefix;
 794         const char *suffix = "_OK";
 795         size_t printed = 0;
 796         int mode = arg->val;
 797
 798         if (mode == F_OK) /* 0 */
 799                 return scnprintf(bf, size, "F%s", show_prefix ? suffix : "");
 800 #define P_MODE(n) \
 801         if (mode & n##_OK) { \
 802                 printed += scnprintf(bf + printed, size - printed, "%s%s", #n, show_prefix ? suffix : ""); \
 803                 mode &= ~n##_OK; \
 804         }
 805
 806         P_MODE(R);
 807         P_MODE(W);
 808         P_MODE(X);
 809 #undef P_MODE
 810
 811         if (mode)
 812                 printed += scnprintf(bf + printed, size - printed, "|%#x", mode);
 813
 814         return printed;
 815 }
 816
 817 #define SCA_ACCMODE syscall_arg__scnprintf_access_mode
 818
 819 static size_t syscall_arg__scnprintf_filename(char *bf, size_t size,
 820                                               struct syscall_arg *arg);
 821
 822 #define SCA_FILENAME syscall_arg__scnprintf_filename
 823
 824 static size_t syscall_arg__scnprintf_pipe_flags(char *bf, size_t size,
 825                                                 struct syscall_arg *arg)
 826 {
 827         bool show_prefix = arg->show_string_prefix;
 828         const char *prefix = "O_";
 829         int printed = 0, flags = arg->val;
 830
 831 #define P_FLAG(n) \
 832         if (flags & O_##n) { \
 833                 printed += scnprintf(bf + printed, size - printed, "%s%s%s", printed ? "|" : "", show_prefix ? prefix : "", #n); \
 834                 flags &= ~O_##n; \
 835         }
 836
 837         P_FLAG(CLOEXEC);
 838         P_FLAG(NONBLOCK);
 839 #undef P_FLAG
 840
 841         if (flags)
 842                 printed += scnprintf(bf + printed, size - printed, "%s%#x", printed ? "|" : "", flags);
 843
 844         return printed;
 845 }
 846
 847 #define SCA_PIPE_FLAGS syscall_arg__scnprintf_pipe_flags
 848
 849 #ifndef GRND_NONBLOCK
 850 #define GRND_NONBLOCK   0x0001
 851 #endif
 852 #ifndef GRND_RANDOM
 853 #define GRND_RANDOM     0x0002
 854 #endif
 855
 856 static size_t syscall_arg__scnprintf_getrandom_flags(char *bf, size_t size,
 857                                                    struct syscall_arg *arg)
 858 {
 859         bool show_prefix = arg->show_string_prefix;
 860         const char *prefix = "GRND_";
 861         int printed = 0, flags = arg->val;
 862
 863 #define P_FLAG(n) \
 864         if (flags & GRND_##n) { \
 865                 printed += scnprintf(bf + printed, size - printed, "%s%s%s", printed ? "|" : "", show_prefix ? prefix : "", #n); \
 866                 flags &= ~GRND_##n; \
 867         }
 868
 869         P_FLAG(RANDOM);
 870         P_FLAG(NONBLOCK);
 871 #undef P_FLAG
 872
 873         if (flags)
 874                 printed += scnprintf(bf + printed, size - printed, "%s%#x", printed ? "|" : "", flags);
 875
 876         return printed;
 877 }
 878
 879 #define SCA_GETRANDOM_FLAGS syscall_arg__scnprintf_getrandom_flags
 880
 881 #define STRARRAY(name, array) \
 882           { .scnprintf  = SCA_STRARRAY, \
 883             .strtoul    = STUL_STRARRAY, \
 884             .parm       = &strarray__##array, }
 885
 886 #define STRARRAY_FLAGS(name, array) \
 887           { .scnprintf  = SCA_STRARRAY_FLAGS, \
 888             .strtoul    = STUL_STRARRAY_FLAGS, \
 889             .parm       = &strarray__##array, }
 890
 891 #include "trace/beauty/arch_errno_names.c"
 892 #include "trace/beauty/eventfd.c"
 893 #include "trace/beauty/futex_op.c"
 894 #include "trace/beauty/futex_val3.c"
 895 #include "trace/beauty/mmap.c"
 896 #include "trace/beauty/mode_t.c"
 897 #include "trace/beauty/msg_flags.c"
 898 #include "trace/beauty/open_flags.c"
 899 #include "trace/beauty/perf_event_open.c"
 900 #include "trace/beauty/pid.c"
 901 #include "trace/beauty/sched_policy.c"
 902 #include "trace/beauty/seccomp.c"
 903 #include "trace/beauty/signum.c"
 904 #include "trace/beauty/socket_type.c"
 905 #include "trace/beauty/waitid_options.c"
 906
 907 static struct syscall_fmt syscall_fmts[] = {
 908         { .name     = "access",
 909           .arg = { [1] = { .scnprintf = SCA_ACCMODE,  /* mode */ }, }, },
 910         { .name     = "arch_prctl",
 911           .arg = { [0] = { .scnprintf = SCA_X86_ARCH_PRCTL_CODE, /* code */ },
 912                    [1] = { .scnprintf = SCA_PTR, /* arg2 */ }, }, },
 913         { .name     = "bind",
 914           .arg = { [0] = { .scnprintf = SCA_INT, /* fd */ },
 915                    [1] = { .scnprintf = SCA_SOCKADDR, /* umyaddr */ },
 916                    [2] = { .scnprintf = SCA_INT, /* addrlen */ }, }, },
 917         { .name     = "bpf",
 918           .arg = { [0] = STRARRAY(cmd, bpf_cmd), }, },
 919         { .name     = "brk",        .hexret = true,
 920           .arg = { [0] = { .scnprintf = SCA_PTR, /* brk */ }, }, },
 921         { .name     = "clock_gettime",
 922           .arg = { [0] = STRARRAY(clk_id, clockid), }, },
 923         { .name     = "clone",      .errpid = true, .nr_args = 5,
 924           .arg = { [0] = { .name = "flags",         .scnprintf = SCA_CLONE_FLAGS, },
 925                    [1] = { .name = "child_stack",   .scnprintf = SCA_HEX, },
 926                    [2] = { .name = "parent_tidptr", .scnprintf = SCA_HEX, },
 927                    [3] = { .name = "child_tidptr",  .scnprintf = SCA_HEX, },
 928                    [4] = { .name = "tls",           .scnprintf = SCA_HEX, }, }, },
 929         { .name     = "close",
 930           .arg = { [0] = { .scnprintf = SCA_CLOSE_FD, /* fd */ }, }, },
 931         { .name     = "connect",
 932           .arg = { [0] = { .scnprintf = SCA_INT, /* fd */ },
 933                    [1] = { .scnprintf = SCA_SOCKADDR, /* servaddr */ },
 934                    [2] = { .scnprintf = SCA_INT, /* addrlen */ }, }, },
 935         { .name     = "epoll_ctl",
 936           .arg = { [1] = STRARRAY(op, epoll_ctl_ops), }, },
 937         { .name     = "eventfd2",
 938           .arg = { [1] = { .scnprintf = SCA_EFD_FLAGS, /* flags */ }, }, },
 939         { .name     = "fchmodat",
 940           .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
 941         { .name     = "fchownat",
 942           .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
 943         { .name     = "fcntl",
 944           .arg = { [1] = { .scnprintf = SCA_FCNTL_CMD,  /* cmd */
 945                            .strtoul   = STUL_STRARRAYS,
 946                            .parm      = &strarrays__fcntl_cmds_arrays,
 947                            .show_zero = true, },
 948                    [2] = { .scnprintf =  SCA_FCNTL_ARG, /* arg */ }, }, },
 949         { .name     = "flock",
 950           .arg = { [1] = { .scnprintf = SCA_FLOCK, /* cmd */ }, }, },
 951         { .name     = "fsconfig",
 952           .arg = { [1] = STRARRAY(cmd, fsconfig_cmds), }, },
 953         { .name     = "fsmount",
 954           .arg = { [1] = STRARRAY_FLAGS(flags, fsmount_flags),
 955                    [2] = { .scnprintf = SCA_FSMOUNT_ATTR_FLAGS, /* attr_flags */ }, }, },
 956         { .name     = "fspick",
 957           .arg = { [0] = { .scnprintf = SCA_FDAT,         /* dfd */ },
 958                    [1] = { .scnprintf = SCA_FILENAME,     /* path */ },
 959                    [2] = { .scnprintf = SCA_FSPICK_FLAGS, /* flags */ }, }, },
 960         { .name     = "fstat", .alias = "newfstat", },
 961         { .name     = "fstatat", .alias = "newfstatat", },
 962         { .name     = "futex",
 963           .arg = { [1] = { .scnprintf = SCA_FUTEX_OP, /* op */ },
 964                    [5] = { .scnprintf = SCA_FUTEX_VAL3, /* val3 */ }, }, },
 965         { .name     = "futimesat",
 966           .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
 967         { .name     = "getitimer",
 968           .arg = { [0] = STRARRAY(which, itimers), }, },
 969         { .name     = "getpid",     .errpid = true, },
 970         { .name     = "getpgid",    .errpid = true, },
 971         { .name     = "getppid",    .errpid = true, },
 972         { .name     = "getrandom",
 973           .arg = { [2] = { .scnprintf = SCA_GETRANDOM_FLAGS, /* flags */ }, }, },
 974         { .name     = "getrlimit",
 975           .arg = { [0] = STRARRAY(resource, rlimit_resources), }, },
 976         { .name     = "gettid",     .errpid = true, },
 977         { .name     = "ioctl",
 978           .arg = {
 979 #if defined(__i386__) || defined(__x86_64__)
 980 /*
 981  * FIXME: Make this available to all arches.
 982  */
 983                    [1] = { .scnprintf = SCA_IOCTL_CMD, /* cmd */ },
 984                    [2] = { .scnprintf = SCA_HEX, /* arg */ }, }, },
 985 #else
 986                    [2] = { .scnprintf = SCA_HEX, /* arg */ }, }, },
 987 #endif
 988         { .name     = "kcmp",       .nr_args = 5,
 989           .arg = { [0] = { .name = "pid1",      .scnprintf = SCA_PID, },
 990                    [1] = { .name = "pid2",      .scnprintf = SCA_PID, },
 991                    [2] = { .name = "type",      .scnprintf = SCA_KCMP_TYPE, },
 992                    [3] = { .name = "idx1",      .scnprintf = SCA_KCMP_IDX, },
 993                    [4] = { .name = "idx2",      .scnprintf = SCA_KCMP_IDX, }, }, },
 994         { .name     = "keyctl",
 995           .arg = { [0] = STRARRAY(option, keyctl_options), }, },
 996         { .name     = "kill",
 997           .arg = { [1] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
 998         { .name     = "linkat",
 999           .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
1000         { .name     = "lseek",
1001           .arg = { [2] = STRARRAY(whence, whences), }, },
1002         { .name     = "lstat", .alias = "newlstat", },
1003         { .name     = "madvise",
1004           .arg = { [0] = { .scnprintf = SCA_HEX,      /* start */ },
1005                    [2] = { .scnprintf = SCA_MADV_BHV, /* behavior */ }, }, },
1006         { .name     = "mkdirat",
1007           .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
1008         { .name     = "mknodat",
1009           .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
1010         { .name     = "mmap",       .hexret = true,
1011 /* The standard mmap maps to old_mmap on s390x */
1012 #if defined(__s390x__)
1013         .alias = "old_mmap",
1014 #endif
1015           .arg = { [2] = { .scnprintf = SCA_MMAP_PROT,  /* prot */ },
1016                    [3] = { .scnprintf = SCA_MMAP_FLAGS, /* flags */
1017                            .strtoul   = STUL_STRARRAY_FLAGS,
1018                            .parm      = &strarray__mmap_flags, },
1019                    [5] = { .scnprintf = SCA_HEX,        /* offset */ }, }, },
1020         { .name     = "mount",
1021           .arg = { [0] = { .scnprintf = SCA_FILENAME, /* dev_name */ },
1022                    [3] = { .scnprintf = SCA_MOUNT_FLAGS, /* flags */
1023                            .mask_val  = SCAMV_MOUNT_FLAGS, /* flags */ }, }, },
1024         { .name     = "move_mount",
1025           .arg = { [0] = { .scnprintf = SCA_FDAT,       /* from_dfd */ },
1026                    [1] = { .scnprintf = SCA_FILENAME, /* from_pathname */ },
1027                    [2] = { .scnprintf = SCA_FDAT,       /* to_dfd */ },
1028                    [3] = { .scnprintf = SCA_FILENAME, /* to_pathname */ },
1029                    [4] = { .scnprintf = SCA_MOVE_MOUNT_FLAGS, /* flags */ }, }, },
1030         { .name     = "mprotect",
1031           .arg = { [0] = { .scnprintf = SCA_HEX,        /* start */ },
1032                    [2] = { .scnprintf = SCA_MMAP_PROT,  /* prot */ }, }, },
1033         { .name     = "mq_unlink",
1034           .arg = { [0] = { .scnprintf = SCA_FILENAME, /* u_name */ }, }, },
1035         { .name     = "mremap",     .hexret = true,
1036           .arg = { [3] = { .scnprintf = SCA_MREMAP_FLAGS, /* flags */ }, }, },
1037         { .name     = "name_to_handle_at",
1038           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
1039         { .name     = "newfstatat",
1040           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
1041         { .name     = "open",
1042           .arg = { [1] = { .scnprintf = SCA_OPEN_FLAGS, /* flags */ }, }, },
1043         { .name     = "open_by_handle_at",
1044           .arg = { [0] = { .scnprintf = SCA_FDAT,       /* dfd */ },
1045                    [2] = { .scnprintf = SCA_OPEN_FLAGS, /* flags */ }, }, },
1046         { .name     = "openat",
1047           .arg = { [0] = { .scnprintf = SCA_FDAT,       /* dfd */ },
1048                    [2] = { .scnprintf = SCA_OPEN_FLAGS, /* flags */ }, }, },
1049         { .name     = "perf_event_open",
1050           .arg = { [2] = { .scnprintf = SCA_INT,        /* cpu */ },
1051                    [3] = { .scnprintf = SCA_FD,         /* group_fd */ },
1052                    [4] = { .scnprintf = SCA_PERF_FLAGS, /* flags */ }, }, },
1053         { .name     = "pipe2",
1054           .arg = { [1] = { .scnprintf = SCA_PIPE_FLAGS, /* flags */ }, }, },
1055         { .name     = "pkey_alloc",
1056           .arg = { [1] = { .scnprintf = SCA_PKEY_ALLOC_ACCESS_RIGHTS,   /* access_rights */ }, }, },
1057         { .name     = "pkey_free",
1058           .arg = { [0] = { .scnprintf = SCA_INT,        /* key */ }, }, },
1059         { .name     = "pkey_mprotect",
1060           .arg = { [0] = { .scnprintf = SCA_HEX,        /* start */ },
1061                    [2] = { .scnprintf = SCA_MMAP_PROT,  /* prot */ },
1062                    [3] = { .scnprintf = SCA_INT,        /* pkey */ }, }, },
1063         { .name     = "poll", .timeout = true, },
1064         { .name     = "ppoll", .timeout = true, },
1065         { .name     = "prctl",
1066           .arg = { [0] = { .scnprintf = SCA_PRCTL_OPTION, /* option */
1067                            .strtoul   = STUL_STRARRAY,
1068                            .parm      = &strarray__prctl_options, },
1069                    [1] = { .scnprintf = SCA_PRCTL_ARG2, /* arg2 */ },
1070                    [2] = { .scnprintf = SCA_PRCTL_ARG3, /* arg3 */ }, }, },
1071         { .name     = "pread", .alias = "pread64", },
1072         { .name     = "preadv", .alias = "pread", },
1073         { .name     = "prlimit64",
1074           .arg = { [1] = STRARRAY(resource, rlimit_resources), }, },
1075         { .name     = "pwrite", .alias = "pwrite64", },
1076         { .name     = "readlinkat",
1077           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
1078         { .name     = "recvfrom",
1079           .arg = { [3] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
1080         { .name     = "recvmmsg",
1081           .arg = { [3] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
1082         { .name     = "recvmsg",
1083           .arg = { [2] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
1084         { .name     = "renameat",
1085           .arg = { [0] = { .scnprintf = SCA_FDAT, /* olddirfd */ },
1086                    [2] = { .scnprintf = SCA_FDAT, /* newdirfd */ }, }, },
1087         { .name     = "renameat2",
1088           .arg = { [0] = { .scnprintf = SCA_FDAT, /* olddirfd */ },
1089                    [2] = { .scnprintf = SCA_FDAT, /* newdirfd */ },
1090                    [4] = { .scnprintf = SCA_RENAMEAT2_FLAGS, /* flags */ }, }, },
1091         { .name     = "rt_sigaction",
1092           .arg = { [0] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
1093         { .name     = "rt_sigprocmask",
1094           .arg = { [0] = STRARRAY(how, sighow), }, },
1095         { .name     = "rt_sigqueueinfo",
1096           .arg = { [1] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
1097         { .name     = "rt_tgsigqueueinfo",
1098           .arg = { [2] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
1099         { .name     = "sched_setscheduler",
1100           .arg = { [1] = { .scnprintf = SCA_SCHED_POLICY, /* policy */ }, }, },
1101         { .name     = "seccomp",
1102           .arg = { [0] = { .scnprintf = SCA_SECCOMP_OP,    /* op */ },
1103                    [1] = { .scnprintf = SCA_SECCOMP_FLAGS, /* flags */ }, }, },
1104         { .name     = "select", .timeout = true, },
1105         { .name     = "sendfile", .alias = "sendfile64", },
1106         { .name     = "sendmmsg",
1107           .arg = { [3] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
1108         { .name     = "sendmsg",
1109           .arg = { [2] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
1110         { .name     = "sendto",
1111           .arg = { [3] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ },
1112                    [4] = { .scnprintf = SCA_SOCKADDR, /* addr */ }, }, },
1113         { .name     = "set_tid_address", .errpid = true, },
1114         { .name     = "setitimer",
1115           .arg = { [0] = STRARRAY(which, itimers), }, },
1116         { .name     = "setrlimit",
1117           .arg = { [0] = STRARRAY(resource, rlimit_resources), }, },
1118         { .name     = "socket",
1119           .arg = { [0] = STRARRAY(family, socket_families),
1120                    [1] = { .scnprintf = SCA_SK_TYPE, /* type */ },
1121                    [2] = { .scnprintf = SCA_SK_PROTO, /* protocol */ }, }, },
1122         { .name     = "socketpair",
1123           .arg = { [0] = STRARRAY(family, socket_families),
1124                    [1] = { .scnprintf = SCA_SK_TYPE, /* type */ },
1125                    [2] = { .scnprintf = SCA_SK_PROTO, /* protocol */ }, }, },
1126         { .name     = "stat", .alias = "newstat", },
1127         { .name     = "statx",
1128           .arg = { [0] = { .scnprintf = SCA_FDAT,        /* fdat */ },
1129                    [2] = { .scnprintf = SCA_STATX_FLAGS, /* flags */ } ,
1130                    [3] = { .scnprintf = SCA_STATX_MASK,  /* mask */ }, }, },
1131         { .name     = "swapoff",
1132           .arg = { [0] = { .scnprintf = SCA_FILENAME, /* specialfile */ }, }, },
1133         { .name     = "swapon",
1134           .arg = { [0] = { .scnprintf = SCA_FILENAME, /* specialfile */ }, }, },
1135         { .name     = "symlinkat",
1136           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
1137         { .name     = "sync_file_range",
1138           .arg = { [3] = { .scnprintf = SCA_SYNC_FILE_RANGE_FLAGS, /* flags */ }, }, },
1139         { .name     = "tgkill",
1140           .arg = { [2] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
1141         { .name     = "tkill",
1142           .arg = { [1] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
1143         { .name     = "umount2", .alias = "umount",
1144           .arg = { [0] = { .scnprintf = SCA_FILENAME, /* name */ }, }, },
1145         { .name     = "uname", .alias = "newuname", },
1146         { .name     = "unlinkat",
1147           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
1148         { .name     = "utimensat",
1149           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dirfd */ }, }, },
1150         { .name     = "wait4",      .errpid = true,
1151           .arg = { [2] = { .scnprintf = SCA_WAITID_OPTIONS, /* options */ }, }, },
1152         { .name     = "waitid",     .errpid = true,
1153           .arg = { [3] = { .scnprintf = SCA_WAITID_OPTIONS, /* options */ }, }, },
1154 };
1155
1156 static int syscall_fmt__cmp(const void *name, const void *fmtp)
1157 {
1158         const struct syscall_fmt *fmt = fmtp;
1159         return strcmp(name, fmt->name);
1160 }
1161
1162 static struct syscall_fmt *__syscall_fmt__find(struct syscall_fmt *fmts, const int nmemb, const char *name)
1163 {
1164         return bsearch(name, fmts, nmemb, sizeof(struct syscall_fmt), syscall_fmt__cmp);
1165 }
1166
1167 static struct syscall_fmt *syscall_fmt__find(const char *name)
1168 {
1169         const int nmemb = ARRAY_SIZE(syscall_fmts);
1170         return __syscall_fmt__find(syscall_fmts, nmemb, name);
1171 }
1172
1173 static struct syscall_fmt *__syscall_fmt__find_by_alias(struct syscall_fmt *fmts, const int nmemb, const char *alias)
1174 {
1175         int i;
1176
1177         for (i = 0; i < nmemb; ++i) {
1178                 if (fmts[i].alias && strcmp(fmts[i].alias, alias) == 0)
1179                         return &fmts[i];
1180         }
1181
1182         return NULL;
1183 }
1184
1185 static struct syscall_fmt *syscall_fmt__find_by_alias(const char *alias)
1186 {
1187         const int nmemb = ARRAY_SIZE(syscall_fmts);
1188         return __syscall_fmt__find_by_alias(syscall_fmts, nmemb, alias);
1189 }
1190
1191 /*
1192  * is_exit: is this "exit" or "exit_group"?
1193  * is_open: is this "open" or "openat"? To associate the fd returned in sys_exit with the pathname in sys_enter.
1194  * args_size: sum of the sizes of the syscall arguments, anything after that is augmented stuff: pathname for openat, etc.
1195  * nonexistent: Just a hole in the syscall table, syscall id not allocated
1196  */
1197 struct syscall {
1198         struct tep_event    *tp_format;
1199         int                 nr_args;
1200         int                 args_size;
1201         struct {
1202                 struct bpf_program *sys_enter,
1203                                    *sys_exit;
1204         }                   bpf_prog;
1205         bool                is_exit;
1206         bool                is_open;
1207         bool                nonexistent;
1208         struct tep_format_field *args;
1209         const char          *name;
1210         struct syscall_fmt  *fmt;
1211         struct syscall_arg_fmt *arg_fmt;
1212 };
1213
1214 /*
1215  * Must match what is in the BPF program:
1216  *
1217  * tools/perf/examples/bpf/augmented_raw_syscalls.c
1218  */
1219 struct bpf_map_syscall_entry {
1220         bool    enabled;
1221         u16     string_args_len[RAW_SYSCALL_ARGS_NUM];
1222 };
1223
1224 /*
1225  * We need to have this 'calculated' boolean because in some cases we really
1226  * don't know what is the duration of a syscall, for instance, when we start
1227  * a session and some threads are waiting for a syscall to finish, say 'poll',
1228  * in which case all we can do is to print "( ? ) for duration and for the
1229  * start timestamp.
1230  */
1231 static size_t fprintf_duration(unsigned long t, bool calculated, FILE *fp)
1232 {
1233         double duration = (double)t / NSEC_PER_MSEC;
1234         size_t printed = fprintf(fp, "(");
1235
1236         if (!calculated)
1237                 printed += fprintf(fp, "         ");
1238         else if (duration >= 1.0)
1239                 printed += color_fprintf(fp, PERF_COLOR_RED, "%6.3f ms", duration);
1240         else if (duration >= 0.01)
1241                 printed += color_fprintf(fp, PERF_COLOR_YELLOW, "%6.3f ms", duration);
1242         else
1243                 printed += color_fprintf(fp, PERF_COLOR_NORMAL, "%6.3f ms", duration);
1244         return printed + fprintf(fp, "): ");
1245 }
1246
1247 /**
1248  * filename.ptr: The filename char pointer that will be vfs_getname'd
1249  * filename.entry_str_pos: Where to insert the string translated from
1250  *                         filename.ptr by the vfs_getname tracepoint/kprobe.
1251  * ret_scnprintf: syscall args may set this to a different syscall return
1252  *                formatter, for instance, fcntl may return fds, file flags, etc.
1253  */
1254 struct thread_trace {
1255         u64               entry_time;
1256         bool              entry_pending;
1257         unsigned long     nr_events;
1258         unsigned long     pfmaj, pfmin;
1259         char              *entry_str;
1260         double            runtime_ms;
1261         size_t            (*ret_scnprintf)(char *bf, size_t size, struct syscall_arg *arg);
1262         struct {
1263                 unsigned long ptr;
1264                 short int     entry_str_pos;
1265                 bool          pending_open;
1266                 unsigned int  namelen;
1267                 char          *name;
1268         } filename;
1269         struct {
1270                 int           max;
1271                 struct file   *table;
1272         } files;
1273
1274         struct intlist *syscall_stats;
1275 };
1276
1277 static struct thread_trace *thread_trace__new(void)
1278 {
1279         struct thread_trace *ttrace =  zalloc(sizeof(struct thread_trace));
1280
1281         if (ttrace) {
1282                 ttrace->files.max = -1;
1283                 ttrace->syscall_stats = intlist__new(NULL);
1284         }
1285
1286         return ttrace;
1287 }
1288
1289 static struct thread_trace *thread__trace(struct thread *thread, FILE *fp)
1290 {
1291         struct thread_trace *ttrace;
1292
1293         if (thread == NULL)
1294                 goto fail;
1295
1296         if (thread__priv(thread) == NULL)
1297                 thread__set_priv(thread, thread_trace__new());
1298
1299         if (thread__priv(thread) == NULL)
1300                 goto fail;
1301
1302         ttrace = thread__priv(thread);
1303         ++ttrace->nr_events;
1304
1305         return ttrace;
1306 fail:
1307         color_fprintf(fp, PERF_COLOR_RED,
1308                       "WARNING: not enough memory, dropping samples!\n");
1309         return NULL;
1310 }
1311
1312
1313 void syscall_arg__set_ret_scnprintf(struct syscall_arg *arg,
1314                                     size_t (*ret_scnprintf)(char *bf, size_t size, struct syscall_arg *arg))
1315 {
1316         struct thread_trace *ttrace = thread__priv(arg->thread);
1317
1318         ttrace->ret_scnprintf = ret_scnprintf;
1319 }
1320
1321 #define TRACE_PFMAJ             (1 << 0)
1322 #define TRACE_PFMIN             (1 << 1)
1323
1324 static const size_t trace__entry_str_size = 2048;
1325
1326 static struct file *thread_trace__files_entry(struct thread_trace *ttrace, int fd)
1327 {
1328         if (fd < 0)
1329                 return NULL;
1330
1331         if (fd > ttrace->files.max) {
1332                 struct file *nfiles = realloc(ttrace->files.table, (fd + 1) * sizeof(struct file));
1333
1334                 if (nfiles == NULL)
1335                         return NULL;
1336
1337                 if (ttrace->files.max != -1) {
1338                         memset(nfiles + ttrace->files.max + 1, 0,
1339                                (fd - ttrace->files.max) * sizeof(struct file));
1340                 } else {
1341                         memset(nfiles, 0, (fd + 1) * sizeof(struct file));
1342                 }
1343
1344                 ttrace->files.table = nfiles;
1345                 ttrace->files.max   = fd;
1346         }
1347
1348         return ttrace->files.table + fd;
1349 }
1350
1351 struct file *thread__files_entry(struct thread *thread, int fd)
1352 {
1353         return thread_trace__files_entry(thread__priv(thread), fd);
1354 }
1355
1356 static int trace__set_fd_pathname(struct thread *thread, int fd, const char *pathname)
1357 {
1358         struct thread_trace *ttrace = thread__priv(thread);
1359         struct file *file = thread_trace__files_entry(ttrace, fd);
1360
1361         if (file != NULL) {
1362                 struct stat st;
1363                 if (stat(pathname, &st) == 0)
1364                         file->dev_maj = major(st.st_rdev);
1365                 file->pathname = strdup(pathname);
1366                 if (file->pathname)
1367                         return 0;
1368         }
1369
1370         return -1;
1371 }
1372
1373 static int thread__read_fd_path(struct thread *thread, int fd)
1374 {
1375         char linkname[PATH_MAX], pathname[PATH_MAX];
1376         struct stat st;
1377         int ret;
1378
1379         if (thread->pid_ == thread->tid) {
1380                 scnprintf(linkname, sizeof(linkname),
1381                           "/proc/%d/fd/%d", thread->pid_, fd);
1382         } else {
1383                 scnprintf(linkname, sizeof(linkname),
1384                           "/proc/%d/task/%d/fd/%d", thread->pid_, thread->tid, fd);
1385         }
1386
1387         if (lstat(linkname, &st) < 0 || st.st_size + 1 > (off_t)sizeof(pathname))
1388                 return -1;
1389
1390         ret = readlink(linkname, pathname, sizeof(pathname));
1391
1392         if (ret < 0 || ret > st.st_size)
1393                 return -1;
1394
1395         pathname[ret] = '\0';
1396         return trace__set_fd_pathname(thread, fd, pathname);
1397 }
1398
1399 static const char *thread__fd_path(struct thread *thread, int fd,
1400                                    struct trace *trace)
1401 {
1402         struct thread_trace *ttrace = thread__priv(thread);
1403
1404         if (ttrace == NULL || trace->fd_path_disabled)
1405                 return NULL;
1406
1407         if (fd < 0)
1408                 return NULL;
1409
1410         if ((fd > ttrace->files.max || ttrace->files.table[fd].pathname == NULL)) {
1411                 if (!trace->live)
1412                         return NULL;
1413                 ++trace->stats.proc_getname;
1414                 if (thread__read_fd_path(thread, fd))
1415                         return NULL;
1416         }
1417
1418         return ttrace->files.table[fd].pathname;
1419 }
1420
1421 size_t syscall_arg__scnprintf_fd(char *bf, size_t size, struct syscall_arg *arg)
1422 {
1423         int fd = arg->val;
1424         size_t printed = scnprintf(bf, size, "%d", fd);
1425         const char *path = thread__fd_path(arg->thread, fd, arg->trace);
1426
1427         if (path)
1428                 printed += scnprintf(bf + printed, size - printed, "<%s>", path);
1429
1430         return printed;
1431 }
1432
1433 size_t pid__scnprintf_fd(struct trace *trace, pid_t pid, int fd, char *bf, size_t size)
1434 {
1435         size_t printed = scnprintf(bf, size, "%d", fd);
1436         struct thread *thread = machine__find_thread(trace->host, pid, pid);
1437
1438         if (thread) {
1439                 const char *path = thread__fd_path(thread, fd, trace);
1440
1441                 if (path)
1442                         printed += scnprintf(bf + printed, size - printed, "<%s>", path);
1443
1444                 thread__put(thread);
1445         }
1446
1447         return printed;
1448 }
1449
1450 static size_t syscall_arg__scnprintf_close_fd(char *bf, size_t size,
1451                                               struct syscall_arg *arg)
1452 {
1453         int fd = arg->val;
1454         size_t printed = syscall_arg__scnprintf_fd(bf, size, arg);
1455         struct thread_trace *ttrace = thread__priv(arg->thread);
1456
1457         if (ttrace && fd >= 0 && fd <= ttrace->files.max)
1458                 zfree(&ttrace->files.table[fd].pathname);
1459
1460         return printed;
1461 }
1462
1463 static void thread__set_filename_pos(struct thread *thread, const char *bf,
1464                                      unsigned long ptr)
1465 {
1466         struct thread_trace *ttrace = thread__priv(thread);
1467
1468         ttrace->filename.ptr = ptr;
1469         ttrace->filename.entry_str_pos = bf - ttrace->entry_str;
1470 }
1471
1472 static size_t syscall_arg__scnprintf_augmented_string(struct syscall_arg *arg, char *bf, size_t size)
1473 {
1474         struct augmented_arg *augmented_arg = arg->augmented.args;
1475         size_t printed = scnprintf(bf, size, "\"%.*s\"", augmented_arg->size, augmented_arg->value);
1476         /*
1477          * So that the next arg with a payload can consume its augmented arg, i.e. for rename* syscalls
1478          * we would have two strings, each prefixed by its size.
1479          */
1480         int consumed = sizeof(*augmented_arg) + augmented_arg->size;
1481
1482         arg->augmented.args = ((void *)arg->augmented.args) + consumed;
1483         arg->augmented.size -= consumed;
1484
1485         return printed;
1486 }
1487
1488 static size_t syscall_arg__scnprintf_filename(char *bf, size_t size,
1489                                               struct syscall_arg *arg)
1490 {
1491         unsigned long ptr = arg->val;
1492
1493         if (arg->augmented.args)
1494                 return syscall_arg__scnprintf_augmented_string(arg, bf, size);
1495
1496         if (!arg->trace->vfs_getname)
1497                 return scnprintf(bf, size, "%#x", ptr);
1498
1499         thread__set_filename_pos(arg->thread, bf, ptr);
1500         return 0;
1501 }
1502
1503 static bool trace__filter_duration(struct trace *trace, double t)
1504 {
1505         return t < (trace->duration_filter * NSEC_PER_MSEC);
1506 }
1507
1508 static size_t __trace__fprintf_tstamp(struct trace *trace, u64 tstamp, FILE *fp)
1509 {
1510         double ts = (double)(tstamp - trace->base_time) / NSEC_PER_MSEC;
1511
1512         return fprintf(fp, "%10.3f ", ts);
1513 }
1514
1515 /*
1516  * We're handling tstamp=0 as an undefined tstamp, i.e. like when we are
1517  * using ttrace->entry_time for a thread that receives a sys_exit without
1518  * first having received a sys_enter ("poll" issued before tracing session
1519  * starts, lost sys_enter exit due to ring buffer overflow).
1520  */
1521 static size_t trace__fprintf_tstamp(struct trace *trace, u64 tstamp, FILE *fp)
1522 {
1523         if (tstamp > 0)
1524                 return __trace__fprintf_tstamp(trace, tstamp, fp);
1525
1526         return fprintf(fp, "         ? ");
1527 }
1528
1529 static bool done = false;
1530 static bool interrupted = false;
1531
1532 static void sig_handler(int sig)
1533 {
1534         done = true;
1535         interrupted = sig == SIGINT;
1536 }
1537
1538 static size_t trace__fprintf_comm_tid(struct trace *trace, struct thread *thread, FILE *fp)
1539 {
1540         size_t printed = 0;
1541
1542         if (trace->multiple_threads) {
1543                 if (trace->show_comm)
1544                         printed += fprintf(fp, "%.14s/", thread__comm_str(thread));
1545                 printed += fprintf(fp, "%d ", thread->tid);
1546         }
1547
1548         return printed;
1549 }
1550
1551 static size_t trace__fprintf_entry_head(struct trace *trace, struct thread *thread,
1552                                         u64 duration, bool duration_calculated, u64 tstamp, FILE *fp)
1553 {
1554         size_t printed = 0;
1555
1556         if (trace->show_tstamp)
1557                 printed = trace__fprintf_tstamp(trace, tstamp, fp);
1558         if (trace->show_duration)
1559                 printed += fprintf_duration(duration, duration_calculated, fp);
1560         return printed + trace__fprintf_comm_tid(trace, thread, fp);
1561 }
1562
1563 static int trace__process_event(struct trace *trace, struct machine *machine,
1564                                 union perf_event *event, struct perf_sample *sample)
1565 {
1566         int ret = 0;
1567
1568         switch (event->header.type) {
1569         case PERF_RECORD_LOST:
1570                 color_fprintf(trace->output, PERF_COLOR_RED,
1571                               "LOST %" PRIu64 " events!\n", event->lost.lost);
1572                 ret = machine__process_lost_event(machine, event, sample);
1573                 break;
1574         default:
1575                 ret = machine__process_event(machine, event, sample);
1576                 break;
1577         }
1578
1579         return ret;
1580 }
1581
1582 static int trace__tool_process(struct perf_tool *tool,
1583                                union perf_event *event,
1584                                struct perf_sample *sample,
1585                                struct machine *machine)
1586 {
1587         struct trace *trace = container_of(tool, struct trace, tool);
1588         return trace__process_event(trace, machine, event, sample);
1589 }
1590
1591 static char *trace__machine__resolve_kernel_addr(void *vmachine, unsigned long long *addrp, char **modp)
1592 {
1593         struct machine *machine = vmachine;
1594
1595         if (machine->kptr_restrict_warned)
1596                 return NULL;
1597
1598         if (symbol_conf.kptr_restrict) {
1599                 pr_warning("Kernel address maps (/proc/{kallsyms,modules}) are restricted.\n\n"
1600                            "Check /proc/sys/kernel/kptr_restrict and /proc/sys/kernel/perf_event_paranoid.\n\n"
1601                            "Kernel samples will not be resolved.\n");
1602                 machine->kptr_restrict_warned = true;
1603                 return NULL;
1604         }
1605
1606         return machine__resolve_kernel_addr(vmachine, addrp, modp);
1607 }
1608
1609 static int trace__symbols_init(struct trace *trace, struct evlist *evlist)
1610 {
1611         int err = symbol__init(NULL);
1612
1613         if (err)
1614                 return err;
1615
1616         trace->host = machine__new_host();
1617         if (trace->host == NULL)
1618                 return -ENOMEM;
1619
1620         err = trace_event__register_resolver(trace->host, trace__machine__resolve_kernel_addr);
1621         if (err < 0)
1622                 goto out;
1623
1624         err = __machine__synthesize_threads(trace->host, &trace->tool, &trace->opts.target,
1625                                             evlist->core.threads, trace__tool_process, false,
1626                                             1);
1627 out:
1628         if (err)
1629                 symbol__exit();
1630
1631         return err;
1632 }
1633
1634 static void trace__symbols__exit(struct trace *trace)
1635 {
1636         machine__exit(trace->host);
1637         trace->host = NULL;
1638
1639         symbol__exit();
1640 }
1641
1642 static int syscall__alloc_arg_fmts(struct syscall *sc, int nr_args)
1643 {
1644         int idx;
1645
1646         if (nr_args == RAW_SYSCALL_ARGS_NUM && sc->fmt && sc->fmt->nr_args != 0)
1647                 nr_args = sc->fmt->nr_args;
1648
1649         sc->arg_fmt = calloc(nr_args, sizeof(*sc->arg_fmt));
1650         if (sc->arg_fmt == NULL)
1651                 return -1;
1652
1653         for (idx = 0; idx < nr_args; ++idx) {
1654                 if (sc->fmt)
1655                         sc->arg_fmt[idx] = sc->fmt->arg[idx];
1656         }
1657
1658         sc->nr_args = nr_args;
1659         return 0;
1660 }
1661
1662 static struct syscall_arg_fmt syscall_arg_fmts__by_name[] = {
1663         { .name = "msr",        .scnprintf = SCA_X86_MSR,         .strtoul = STUL_X86_MSR,         },
1664         { .name = "vector",     .scnprintf = SCA_X86_IRQ_VECTORS, .strtoul = STUL_X86_IRQ_VECTORS, },
1665 };
1666
1667 static int syscall_arg_fmt__cmp(const void *name, const void *fmtp)
1668 {
1669        const struct syscall_arg_fmt *fmt = fmtp;
1670        return strcmp(name, fmt->name);
1671 }
1672
1673 static struct syscall_arg_fmt *
1674 __syscall_arg_fmt__find_by_name(struct syscall_arg_fmt *fmts, const int nmemb, const char *name)
1675 {
1676        return bsearch(name, fmts, nmemb, sizeof(struct syscall_arg_fmt), syscall_arg_fmt__cmp);
1677 }
1678
1679 static struct syscall_arg_fmt *syscall_arg_fmt__find_by_name(const char *name)
1680 {
1681        const int nmemb = ARRAY_SIZE(syscall_arg_fmts__by_name);
1682        return __syscall_arg_fmt__find_by_name(syscall_arg_fmts__by_name, nmemb, name);
1683 }
1684
1685 static struct tep_format_field *
1686 syscall_arg_fmt__init_array(struct syscall_arg_fmt *arg, struct tep_format_field *field)
1687 {
1688         struct tep_format_field *last_field = NULL;
1689         int len;
1690
1691         for (; field; field = field->next, ++arg) {
1692                 last_field = field;
1693
1694                 if (arg->scnprintf)
1695                         continue;
1696
1697                 len = strlen(field->name);
1698
1699                 if (strcmp(field->type, "const char *") == 0 &&
1700                     ((len >= 4 && strcmp(field->name + len - 4, "name") == 0) ||
1701                      strstr(field->name, "path") != NULL))
1702                         arg->scnprintf = SCA_FILENAME;
1703                 else if ((field->flags & TEP_FIELD_IS_POINTER) || strstr(field->name, "addr"))
1704                         arg->scnprintf = SCA_PTR;
1705                 else if (strcmp(field->type, "pid_t") == 0)
1706                         arg->scnprintf = SCA_PID;
1707                 else if (strcmp(field->type, "umode_t") == 0)
1708                         arg->scnprintf = SCA_MODE_T;
1709                 else if ((field->flags & TEP_FIELD_IS_ARRAY) && strstr(field->type, "char")) {
1710                         arg->scnprintf = SCA_CHAR_ARRAY;
1711                         arg->nr_entries = field->arraylen;
1712                 } else if ((strcmp(field->type, "int") == 0 ||
1713                           strcmp(field->type, "unsigned int") == 0 ||
1714                           strcmp(field->type, "long") == 0) &&
1715                          len >= 2 && strcmp(field->name + len - 2, "fd") == 0) {
1716                         /*
1717                          * /sys/kernel/tracing/events/syscalls/sys_enter*
1718                          * egrep 'field:.*fd;' .../format|sed -r 's/.*field:([a-z ]+) [a-z_]*fd.+/\1/g'|sort|uniq -c
1719                          * 65 int
1720                          * 23 unsigned int
1721                          * 7 unsigned long
1722                          */
1723                         arg->scnprintf = SCA_FD;
1724                } else {
1725                         struct syscall_arg_fmt *fmt = syscall_arg_fmt__find_by_name(field->name);
1726
1727                         if (fmt) {
1728                                 arg->scnprintf = fmt->scnprintf;
1729                                 arg->strtoul   = fmt->strtoul;
1730                         }
1731                 }
1732         }
1733
1734         return last_field;
1735 }
1736
1737 static int syscall__set_arg_fmts(struct syscall *sc)
1738 {
1739         struct tep_format_field *last_field = syscall_arg_fmt__init_array(sc->arg_fmt, sc->args);
1740
1741         if (last_field)
1742                 sc->args_size = last_field->offset + last_field->size;
1743
1744         return 0;
1745 }
1746
1747 static int trace__read_syscall_info(struct trace *trace, int id)
1748 {
1749         char tp_name[128];
1750         struct syscall *sc;
1751         const char *name = syscalltbl__name(trace->sctbl, id);
1752
1753 #ifdef HAVE_SYSCALL_TABLE_SUPPORT
1754         if (trace->syscalls.table == NULL) {
1755                 trace->syscalls.table = calloc(trace->sctbl->syscalls.max_id + 1, sizeof(*sc));
1756                 if (trace->syscalls.table == NULL)
1757                         return -ENOMEM;
1758         }
1759 #else
1760         if (id > trace->sctbl->syscalls.max_id || (id == 0 && trace->syscalls.table == NULL)) {
1761                 // When using libaudit we don't know beforehand what is the max syscall id
1762                 struct syscall *table = realloc(trace->syscalls.table, (id + 1) * sizeof(*sc));
1763
1764                 if (table == NULL)
1765                         return -ENOMEM;
1766
1767                 // Need to memset from offset 0 and +1 members if brand new
1768                 if (trace->syscalls.table == NULL)
1769                         memset(table, 0, (id + 1) * sizeof(*sc));
1770                 else
1771                         memset(table + trace->sctbl->syscalls.max_id + 1, 0, (id - trace->sctbl->syscalls.max_id) * sizeof(*sc));
1772
1773                 trace->syscalls.table         = table;
1774                 trace->sctbl->syscalls.max_id = id;
1775         }
1776 #endif
1777         sc = trace->syscalls.table + id;
1778         if (sc->nonexistent)
1779                 return -EEXIST;
1780
1781         if (name == NULL) {
1782                 sc->nonexistent = true;
1783                 return -EEXIST;
1784         }
1785
1786         sc->name = name;
1787         sc->fmt  = syscall_fmt__find(sc->name);
1788
1789         snprintf(tp_name, sizeof(tp_name), "sys_enter_%s", sc->name);
1790         sc->tp_format = trace_event__tp_format("syscalls", tp_name);
1791
1792         if (IS_ERR(sc->tp_format) && sc->fmt && sc->fmt->alias) {
1793                 snprintf(tp_name, sizeof(tp_name), "sys_enter_%s", sc->fmt->alias);
1794                 sc->tp_format = trace_event__tp_format("syscalls", tp_name);
1795         }
1796
1797         /*
1798          * Fails to read trace point format via sysfs node, so the trace point
1799          * doesn't exist.  Set the 'nonexistent' flag as true.
1800          */
1801         if (IS_ERR(sc->tp_format)) {
1802                 sc->nonexistent = true;
1803                 return PTR_ERR(sc->tp_format);
1804         }
1805
1806         if (syscall__alloc_arg_fmts(sc, IS_ERR(sc->tp_format) ?
1807                                         RAW_SYSCALL_ARGS_NUM : sc->tp_format->format.nr_fields))
1808                 return -ENOMEM;
1809
1810         sc->args = sc->tp_format->format.fields;
1811         /*
1812          * We need to check and discard the first variable '__syscall_nr'
1813          * or 'nr' that mean the syscall number. It is needless here.
1814          * So drop '__syscall_nr' or 'nr' field but does not exist on older kernels.
1815          */
1816         if (sc->args && (!strcmp(sc->args->name, "__syscall_nr") || !strcmp(sc->args->name, "nr"))) {
1817                 sc->args = sc->args->next;
1818                 --sc->nr_args;
1819         }
1820
1821         sc->is_exit = !strcmp(name, "exit_group") || !strcmp(name, "exit");
1822         sc->is_open = !strcmp(name, "open") || !strcmp(name, "openat");
1823
1824         return syscall__set_arg_fmts(sc);
1825 }
1826
1827 static int evsel__init_tp_arg_scnprintf(struct evsel *evsel)
1828 {
1829         struct syscall_arg_fmt *fmt = evsel__syscall_arg_fmt(evsel);
1830
1831         if (fmt != NULL) {
1832                 syscall_arg_fmt__init_array(fmt, evsel->tp_format->format.fields);
1833                 return 0;
1834         }
1835
1836         return -ENOMEM;
1837 }
1838
1839 static int intcmp(const void *a, const void *b)
1840 {
1841         const int *one = a, *another = b;
1842
1843         return *one - *another;
1844 }
1845
1846 static int trace__validate_ev_qualifier(struct trace *trace)
1847 {
1848         int err = 0;
1849         bool printed_invalid_prefix = false;
1850         struct str_node *pos;
1851         size_t nr_used = 0, nr_allocated = strlist__nr_entries(trace->ev_qualifier);
1852
1853         trace->ev_qualifier_ids.entries = malloc(nr_allocated *
1854                                                  sizeof(trace->ev_qualifier_ids.entries[0]));
1855
1856         if (trace->ev_qualifier_ids.entries == NULL) {
1857                 fputs("Error:\tNot enough memory for allocating events qualifier ids\n",
1858                        trace->output);
1859                 err = -EINVAL;
1860                 goto out;
1861         }
1862
1863         strlist__for_each_entry(pos, trace->ev_qualifier) {
1864                 const char *sc = pos->s;
1865                 int id = syscalltbl__id(trace->sctbl, sc), match_next = -1;
1866
1867                 if (id < 0) {
1868                         id = syscalltbl__strglobmatch_first(trace->sctbl, sc, &match_next);
1869                         if (id >= 0)
1870                                 goto matches;
1871
1872                         if (!printed_invalid_prefix) {
1873                                 pr_debug("Skipping unknown syscalls: ");
1874                                 printed_invalid_prefix = true;
1875                         } else {
1876                                 pr_debug(", ");
1877                         }
1878
1879                         pr_debug("%s", sc);
1880                         continue;
1881                 }
1882 matches:
1883                 trace->ev_qualifier_ids.entries[nr_used++] = id;
1884                 if (match_next == -1)
1885                         continue;
1886
1887                 while (1) {
1888                         id = syscalltbl__strglobmatch_next(trace->sctbl, sc, &match_next);
1889                         if (id < 0)
1890                                 break;
1891                         if (nr_allocated == nr_used) {
1892                                 void *entries;
1893
1894                                 nr_allocated += 8;
1895                                 entries = realloc(trace->ev_qualifier_ids.entries,
1896                                                   nr_allocated * sizeof(trace->ev_qualifier_ids.entries[0]));
1897                                 if (entries == NULL) {
1898                                         err = -ENOMEM;
1899                                         fputs("\nError:\t Not enough memory for parsing\n", trace->output);
1900                                         goto out_free;
1901                                 }
1902                                 trace->ev_qualifier_ids.entries = entries;
1903                         }
1904                         trace->ev_qualifier_ids.entries[nr_used++] = id;
1905                 }
1906         }
1907
1908         trace->ev_qualifier_ids.nr = nr_used;
1909         qsort(trace->ev_qualifier_ids.entries, nr_used, sizeof(int), intcmp);
1910 out:
1911         if (printed_invalid_prefix)
1912                 pr_debug("\n");
1913         return err;
1914 out_free:
1915         zfree(&trace->ev_qualifier_ids.entries);
1916         trace->ev_qualifier_ids.nr = 0;
1917         goto out;
1918 }
1919
1920 static __maybe_unused bool trace__syscall_enabled(struct trace *trace, int id)
1921 {
1922         bool in_ev_qualifier;
1923
1924         if (trace->ev_qualifier_ids.nr == 0)
1925                 return true;
1926
1927         in_ev_qualifier = bsearch(&id, trace->ev_qualifier_ids.entries,
1928                                   trace->ev_qualifier_ids.nr, sizeof(int), intcmp) != NULL;
1929
1930         if (in_ev_qualifier)
1931                return !trace->not_ev_qualifier;
1932
1933         return trace->not_ev_qualifier;
1934 }
1935
1936 /*
1937  * args is to be interpreted as a series of longs but we need to handle
1938  * 8-byte unaligned accesses. args points to raw_data within the event
1939  * and raw_data is guaranteed to be 8-byte unaligned because it is
1940  * preceded by raw_size which is a u32. So we need to copy args to a temp
1941  * variable to read it. Most notably this avoids extended load instructions
1942  * on unaligned addresses
1943  */
1944 unsigned long syscall_arg__val(struct syscall_arg *arg, u8 idx)
1945 {
1946         unsigned long val;
1947         unsigned char *p = arg->args + sizeof(unsigned long) * idx;
1948
1949         memcpy(&val, p, sizeof(val));
1950         return val;
1951 }
1952
1953 static size_t syscall__scnprintf_name(struct syscall *sc, char *bf, size_t size,
1954                                       struct syscall_arg *arg)
1955 {
1956         if (sc->arg_fmt && sc->arg_fmt[arg->idx].name)
1957                 return scnprintf(bf, size, "%s: ", sc->arg_fmt[arg->idx].name);
1958
1959         return scnprintf(bf, size, "arg%d: ", arg->idx);
1960 }
1961
1962 /*
1963  * Check if the value is in fact zero, i.e. mask whatever needs masking, such
1964  * as mount 'flags' argument that needs ignoring some magic flag, see comment
1965  * in tools/perf/trace/beauty/mount_flags.c
1966  */
1967 static unsigned long syscall_arg_fmt__mask_val(struct syscall_arg_fmt *fmt, struct syscall_arg *arg, unsigned long val)
1968 {
1969         if (fmt && fmt->mask_val)
1970                 return fmt->mask_val(arg, val);
1971
1972         return val;
1973 }
1974
1975 static size_t syscall_arg_fmt__scnprintf_val(struct syscall_arg_fmt *fmt, char *bf, size_t size,
1976                                              struct syscall_arg *arg, unsigned long val)
1977 {
1978         if (fmt && fmt->scnprintf) {
1979                 arg->val = val;
1980                 if (fmt->parm)
1981                         arg->parm = fmt->parm;
1982                 return fmt->scnprintf(bf, size, arg);
1983         }
1984         return scnprintf(bf, size, "%ld", val);
1985 }
1986
1987 static size_t syscall__scnprintf_args(struct syscall *sc, char *bf, size_t size,
1988                                       unsigned char *args, void *augmented_args, int augmented_args_size,
1989                                       struct trace *trace, struct thread *thread)
1990 {
1991         size_t printed = 0;
1992         unsigned long val;
1993         u8 bit = 1;
1994         struct syscall_arg arg = {
1995                 .args   = args,
1996                 .augmented = {
1997                         .size = augmented_args_size,
1998                         .args = augmented_args,
1999                 },
2000                 .idx    = 0,
2001                 .mask   = 0,
2002                 .trace  = trace,
2003                 .thread = thread,
2004                 .show_string_prefix = trace->show_string_prefix,
2005         };
2006         struct thread_trace *ttrace = thread__priv(thread);
2007
2008         /*
2009          * Things like fcntl will set this in its 'cmd' formatter to pick the
2010          * right formatter for the return value (an fd? file flags?), which is
2011          * not needed for syscalls that always return a given type, say an fd.
2012          */
2013         ttrace->ret_scnprintf = NULL;
2014
2015         if (sc->args != NULL) {
2016                 struct tep_format_field *field;
2017
2018                 for (field = sc->args; field;
2019                      field = field->next, ++arg.idx, bit <<= 1) {
2020                         if (arg.mask & bit)
2021                                 continue;
2022
2023                         arg.fmt = &sc->arg_fmt[arg.idx];
2024                         val = syscall_arg__val(&arg, arg.idx);
2025                         /*
2026                          * Some syscall args need some mask, most don't and
2027                          * return val untouched.
2028                          */
2029                         val = syscall_arg_fmt__mask_val(&sc->arg_fmt[arg.idx], &arg, val);
2030
2031                         /*
2032                          * Suppress this argument if its value is zero and
2033                          * and we don't have a string associated in an
2034                          * strarray for it.
2035                          */
2036                         if (val == 0 &&
2037                             !trace->show_zeros &&
2038                             !(sc->arg_fmt &&
2039                               (sc->arg_fmt[arg.idx].show_zero ||
2040                                sc->arg_fmt[arg.idx].scnprintf == SCA_STRARRAY ||
2041                                sc->arg_fmt[arg.idx].scnprintf == SCA_STRARRAYS) &&
2042                               sc->arg_fmt[arg.idx].parm))
2043                                 continue;
2044
2045                         printed += scnprintf(bf + printed, size - printed, "%s", printed ? ", " : "");
2046
2047                         if (trace->show_arg_names)
2048                                 printed += scnprintf(bf + printed, size - printed, "%s: ", field->name);
2049
2050                         printed += syscall_arg_fmt__scnprintf_val(&sc->arg_fmt[arg.idx],
2051                                                                   bf + printed, size - printed, &arg, val);
2052                 }
2053         } else if (IS_ERR(sc->tp_format)) {
2054                 /*
2055                  * If we managed to read the tracepoint /format file, then we
2056                  * may end up not having any args, like with gettid(), so only
2057                  * print the raw args when we didn't manage to read it.
2058                  */
2059                 while (arg.idx < sc->nr_args) {
2060                         if (arg.mask & bit)
2061                                 goto next_arg;
2062                         val = syscall_arg__val(&arg, arg.idx);
2063                         if (printed)
2064                                 printed += scnprintf(bf + printed, size - printed, ", ");
2065                         printed += syscall__scnprintf_name(sc, bf + printed, size - printed, &arg);
2066                         printed += syscall_arg_fmt__scnprintf_val(&sc->arg_fmt[arg.idx], bf + printed, size - printed, &arg, val);
2067 next_arg:
2068                         ++arg.idx;
2069                         bit <<= 1;
2070                 }
2071         }
2072
2073         return printed;
2074 }
2075
2076 typedef int (*tracepoint_handler)(struct trace *trace, struct evsel *evsel,
2077                                   union perf_event *event,
2078                                   struct perf_sample *sample);
2079
2080 static struct syscall *trace__syscall_info(struct trace *trace,
2081                                            struct evsel *evsel, int id)
2082 {
2083         int err = 0;
2084
2085         if (id < 0) {
2086
2087                 /*
2088                  * XXX: Noticed on x86_64, reproduced as far back as 3.0.36, haven't tried
2089                  * before that, leaving at a higher verbosity level till that is
2090                  * explained. Reproduced with plain ftrace with:
2091                  *
2092                  * echo 1 > /t/events/raw_syscalls/sys_exit/enable
2093                  * grep "NR -1 " /t/trace_pipe
2094                  *
2095                  * After generating some load on the machine.
2096                  */
2097                 if (verbose > 1) {
2098                         static u64 n;
2099                         fprintf(trace->output, "Invalid syscall %d id, skipping (%s, %" PRIu64 ") ...\n",
2100                                 id, evsel__name(evsel), ++n);
2101                 }
2102                 return NULL;
2103         }
2104
2105         err = -EINVAL;
2106
2107 #ifdef HAVE_SYSCALL_TABLE_SUPPORT
2108         if (id > trace->sctbl->syscalls.max_id) {
2109 #else
2110         if (id >= trace->sctbl->syscalls.max_id) {
2111                 /*
2112                  * With libaudit we don't know beforehand what is the max_id,
2113                  * so we let trace__read_syscall_info() figure that out as we
2114                  * go on reading syscalls.
2115                  */
2116                 err = trace__read_syscall_info(trace, id);
2117                 if (err)
2118 #endif
2119                 goto out_cant_read;
2120         }
2121
2122         if ((trace->syscalls.table == NULL || trace->syscalls.table[id].name == NULL) &&
2123             (err = trace__read_syscall_info(trace, id)) != 0)
2124                 goto out_cant_read;
2125
2126         if (trace->syscalls.table && trace->syscalls.table[id].nonexistent)
2127                 goto out_cant_read;
2128
2129         return &trace->syscalls.table[id];
2130
2131 out_cant_read:
2132         if (verbose > 0) {
2133                 char sbuf[STRERR_BUFSIZE];
2134                 fprintf(trace->output, "Problems reading syscall %d: %d (%s)", id, -err, str_error_r(-err, sbuf, sizeof(sbuf)));
2135                 if (id <= trace->sctbl->syscalls.max_id && trace->syscalls.table[id].name != NULL)
2136                         fprintf(trace->output, "(%s)", trace->syscalls.table[id].name);
2137                 fputs(" information\n", trace->output);
2138         }
2139         return NULL;
2140 }
2141
2142 struct syscall_stats {
2143         struct stats stats;
2144         u64          nr_failures;
2145         int          max_errno;
2146         u32          *errnos;
2147 };
2148
2149 static void thread__update_stats(struct thread *thread, struct thread_trace *ttrace,
2150                                  int id, struct perf_sample *sample, long err, bool errno_summary)
2151 {
2152         struct int_node *inode;
2153         struct syscall_stats *stats;
2154         u64 duration = 0;
2155
2156         inode = intlist__findnew(ttrace->syscall_stats, id);
2157         if (inode == NULL)
2158                 return;
2159
2160         stats = inode->priv;
2161         if (stats == NULL) {
2162                 stats = malloc(sizeof(*stats));
2163                 if (stats == NULL)
2164                         return;
2165
2166                 stats->nr_failures = 0;
2167                 stats->max_errno   = 0;
2168                 stats->errnos      = NULL;
2169                 init_stats(&stats->stats);
2170                 inode->priv = stats;
2171         }
2172
2173         if (ttrace->entry_time && sample->time > ttrace->entry_time)
2174                 duration = sample->time - ttrace->entry_time;
2175
2176         update_stats(&stats->stats, duration);
2177
2178         if (err < 0) {
2179                 ++stats->nr_failures;
2180
2181                 if (!errno_summary)
2182                         return;
2183
2184                 err = -err;
2185                 if (err > stats->max_errno) {
2186                         u32 *new_errnos = realloc(stats->errnos, err * sizeof(u32));
2187
2188                         if (new_errnos) {
2189                                 memset(new_errnos + stats->max_errno, 0, (err - stats->max_errno) * sizeof(u32));
2190                         } else {
2191                                 pr_debug("Not enough memory for errno stats for thread \"%s\"(%d/%d), results will be incomplete\n",
2192                                          thread__comm_str(thread), thread->pid_, thread->tid);
2193                                 return;
2194                         }
2195
2196                         stats->errnos = new_errnos;
2197                         stats->max_errno = err;
2198                 }
2199
2200                 ++stats->errnos[err - 1];
2201         }
2202 }
2203
2204 static int trace__printf_interrupted_entry(struct trace *trace)
2205 {
2206         struct thread_trace *ttrace;
2207         size_t printed;
2208         int len;
2209
2210         if (trace->failure_only || trace->current == NULL)
2211                 return 0;
2212
2213         ttrace = thread__priv(trace->current);
2214
2215         if (!ttrace->entry_pending)
2216                 return 0;
2217
2218         printed  = trace__fprintf_entry_head(trace, trace->current, 0, false, ttrace->entry_time, trace->output);
2219         printed += len = fprintf(trace->output, "%s)", ttrace->entry_str);
2220
2221         if (len < trace->args_alignment - 4)
2222                 printed += fprintf(trace->output, "%-*s", trace->args_alignment - 4 - len, " ");
2223
2224         printed += fprintf(trace->output, " ...\n");
2225
2226         ttrace->entry_pending = false;
2227         ++trace->nr_events_printed;
2228
2229         return printed;
2230 }
2231
2232 static int trace__fprintf_sample(struct trace *trace, struct evsel *evsel,
2233                                  struct perf_sample *sample, struct thread *thread)
2234 {
2235         int printed = 0;
2236
2237         if (trace->print_sample) {
2238                 double ts = (double)sample->time / NSEC_PER_MSEC;
2239
2240                 printed += fprintf(trace->output, "%22s %10.3f %s %d/%d [%d]\n",
2241                                    evsel__name(evsel), ts,
2242                                    thread__comm_str(thread),
2243                                    sample->pid, sample->tid, sample->cpu);
2244         }
2245
2246         return printed;
2247 }
2248
2249 static void *syscall__augmented_args(struct syscall *sc, struct perf_sample *sample, int *augmented_args_size, int raw_augmented_args_size)
2250 {
2251         void *augmented_args = NULL;
2252         /*
2253          * For now with BPF raw_augmented we hook into raw_syscalls:sys_enter
2254          * and there we get all 6 syscall args plus the tracepoint common fields
2255          * that gets calculated at the start and the syscall_nr (another long).
2256          * So we check if that is the case and if so don't look after the
2257          * sc->args_size but always after the full raw_syscalls:sys_enter payload,
2258          * which is fixed.
2259          *
2260          * We'll revisit this later to pass s->args_size to the BPF augmenter
2261          * (now tools/perf/examples/bpf/augmented_raw_syscalls.c, so that it
2262          * copies only what we need for each syscall, like what happens when we
2263          * use syscalls:sys_enter_NAME, so that we reduce the kernel/userspace
2264          * traffic to just what is needed for each syscall.
2265          */
2266         int args_size = raw_augmented_args_size ?: sc->args_size;
2267
2268         *augmented_args_size = sample->raw_size - args_size;
2269         if (*augmented_args_size > 0)
2270                 augmented_args = sample->raw_data + args_size;
2271
2272         return augmented_args;
2273 }
2274
2275 static int trace__sys_enter(struct trace *trace, struct evsel *evsel,
2276                             union perf_event *event __maybe_unused,
2277                             struct perf_sample *sample)
2278 {
2279         char *msg;
2280         void *args;
2281         int printed = 0;
2282         struct thread *thread;
2283         int id = perf_evsel__sc_tp_uint(evsel, id, sample), err = -1;
2284         int augmented_args_size = 0;
2285         void *augmented_args = NULL;
2286         struct syscall *sc = trace__syscall_info(trace, evsel, id);
2287         struct thread_trace *ttrace;
2288
2289         if (sc == NULL)
2290                 return -1;
2291
2292         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2293         ttrace = thread__trace(thread, trace->output);
2294         if (ttrace == NULL)
2295                 goto out_put;
2296
2297         trace__fprintf_sample(trace, evsel, sample, thread);
2298
2299         args = perf_evsel__sc_tp_ptr(evsel, args, sample);
2300
2301         if (ttrace->entry_str == NULL) {
2302                 ttrace->entry_str = malloc(trace__entry_str_size);
2303                 if (!ttrace->entry_str)
2304                         goto out_put;
2305         }
2306
2307         if (!(trace->duration_filter || trace->summary_only || trace->min_stack))
2308                 trace__printf_interrupted_entry(trace);
2309         /*
2310          * If this is raw_syscalls.sys_enter, then it always comes with the 6 possible
2311          * arguments, even if the syscall being handled, say "openat", uses only 4 arguments
2312          * this breaks syscall__augmented_args() check for augmented args, as we calculate
2313          * syscall->args_size using each syscalls:sys_enter_NAME tracefs format file,
2314          * so when handling, say the openat syscall, we end up getting 6 args for the
2315          * raw_syscalls:sys_enter event, when we expected just 4, we end up mistakenly
2316          * thinking that the extra 2 u64 args are the augmented filename, so just check
2317          * here and avoid using augmented syscalls when the evsel is the raw_syscalls one.
2318          */
2319         if (evsel != trace->syscalls.events.sys_enter)
2320                 augmented_args = syscall__augmented_args(sc, sample, &augmented_args_size, trace->raw_augmented_syscalls_args_size);
2321         ttrace->entry_time = sample->time;
2322         msg = ttrace->entry_str;
2323         printed += scnprintf(msg + printed, trace__entry_str_size - printed, "%s(", sc->name);
2324
2325         printed += syscall__scnprintf_args(sc, msg + printed, trace__entry_str_size - printed,
2326                                            args, augmented_args, augmented_args_size, trace, thread);
2327
2328         if (sc->is_exit) {
2329                 if (!(trace->duration_filter || trace->summary_only || trace->failure_only || trace->min_stack)) {
2330                         int alignment = 0;
2331
2332                         trace__fprintf_entry_head(trace, thread, 0, false, ttrace->entry_time, trace->output);
2333                         printed = fprintf(trace->output, "%s)", ttrace->entry_str);
2334                         if (trace->args_alignment > printed)
2335                                 alignment = trace->args_alignment - printed;
2336                         fprintf(trace->output, "%*s= ?\n", alignment, " ");
2337                 }
2338         } else {
2339                 ttrace->entry_pending = true;
2340                 /* See trace__vfs_getname & trace__sys_exit */
2341                 ttrace->filename.pending_open = false;
2342         }
2343
2344         if (trace->current != thread) {
2345                 thread__put(trace->current);
2346                 trace->current = thread__get(thread);
2347         }
2348         err = 0;
2349 out_put:
2350         thread__put(thread);
2351         return err;
2352 }
2353
2354 static int trace__fprintf_sys_enter(struct trace *trace, struct evsel *evsel,
2355                                     struct perf_sample *sample)
2356 {
2357         struct thread_trace *ttrace;
2358         struct thread *thread;
2359         int id = perf_evsel__sc_tp_uint(evsel, id, sample), err = -1;
2360         struct syscall *sc = trace__syscall_info(trace, evsel, id);
2361         char msg[1024];
2362         void *args, *augmented_args = NULL;
2363         int augmented_args_size;
2364
2365         if (sc == NULL)
2366                 return -1;
2367
2368         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2369         ttrace = thread__trace(thread, trace->output);
2370         /*
2371          * We need to get ttrace just to make sure it is there when syscall__scnprintf_args()
2372          * and the rest of the beautifiers accessing it via struct syscall_arg touches it.
2373          */
2374         if (ttrace == NULL)
2375                 goto out_put;
2376
2377         args = perf_evsel__sc_tp_ptr(evsel, args, sample);
2378         augmented_args = syscall__augmented_args(sc, sample, &augmented_args_size, trace->raw_augmented_syscalls_args_size);
2379         syscall__scnprintf_args(sc, msg, sizeof(msg), args, augmented_args, augmented_args_size, trace, thread);
2380         fprintf(trace->output, "%s", msg);
2381         err = 0;
2382 out_put:
2383         thread__put(thread);
2384         return err;
2385 }
2386
2387 static int trace__resolve_callchain(struct trace *trace, struct evsel *evsel,
2388                                     struct perf_sample *sample,
2389                                     struct callchain_cursor *cursor)
2390 {
2391         struct addr_location al;
2392         int max_stack = evsel->core.attr.sample_max_stack ?
2393                         evsel->core.attr.sample_max_stack :
2394                         trace->max_stack;
2395         int err;
2396
2397         if (machine__resolve(trace->host, &al, sample) < 0)
2398                 return -1;
2399
2400         err = thread__resolve_callchain(al.thread, cursor, evsel, sample, NULL, NULL, max_stack);
2401         addr_location__put(&al);
2402         return err;
2403 }
2404
2405 static int trace__fprintf_callchain(struct trace *trace, struct perf_sample *sample)
2406 {
2407         /* TODO: user-configurable print_opts */
2408         const unsigned int print_opts = EVSEL__PRINT_SYM |
2409                                         EVSEL__PRINT_DSO |
2410                                         EVSEL__PRINT_UNKNOWN_AS_ADDR;
2411
2412         return sample__fprintf_callchain(sample, 38, print_opts, &callchain_cursor, symbol_conf.bt_stop_list, trace->output);
2413 }
2414
2415 static const char *errno_to_name(struct evsel *evsel, int err)
2416 {
2417         struct perf_env *env = evsel__env(evsel);
2418         const char *arch_name = perf_env__arch(env);
2419
2420         return arch_syscalls__strerrno(arch_name, err);
2421 }
2422
2423 static int trace__sys_exit(struct trace *trace, struct evsel *evsel,
2424                            union perf_event *event __maybe_unused,
2425                            struct perf_sample *sample)
2426 {
2427         long ret;
2428         u64 duration = 0;
2429         bool duration_calculated = false;
2430         struct thread *thread;
2431         int id = perf_evsel__sc_tp_uint(evsel, id, sample), err = -1, callchain_ret = 0, printed = 0;
2432         int alignment = trace->args_alignment;
2433         struct syscall *sc = trace__syscall_info(trace, evsel, id);
2434         struct thread_trace *ttrace;
2435
2436         if (sc == NULL)
2437                 return -1;
2438
2439         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2440         ttrace = thread__trace(thread, trace->output);
2441         if (ttrace == NULL)
2442                 goto out_put;
2443
2444         trace__fprintf_sample(trace, evsel, sample, thread);
2445
2446         ret = perf_evsel__sc_tp_uint(evsel, ret, sample);
2447
2448         if (trace->summary)
2449                 thread__update_stats(thread, ttrace, id, sample, ret, trace->errno_summary);
2450
2451         if (!trace->fd_path_disabled && sc->is_open && ret >= 0 && ttrace->filename.pending_open) {
2452                 trace__set_fd_pathname(thread, ret, ttrace->filename.name);
2453                 ttrace->filename.pending_open = false;
2454                 ++trace->stats.vfs_getname;
2455         }
2456
2457         if (ttrace->entry_time) {
2458                 duration = sample->time - ttrace->entry_time;
2459                 if (trace__filter_duration(trace, duration))
2460                         goto out;
2461                 duration_calculated = true;
2462         } else if (trace->duration_filter)
2463                 goto out;
2464
2465         if (sample->callchain) {
2466                 callchain_ret = trace__resolve_callchain(trace, evsel, sample, &callchain_cursor);
2467                 if (callchain_ret == 0) {
2468                         if (callchain_cursor.nr < trace->min_stack)
2469                                 goto out;
2470                         callchain_ret = 1;
2471                 }
2472         }
2473
2474         if (trace->summary_only || (ret >= 0 && trace->failure_only))
2475                 goto out;
2476
2477         trace__fprintf_entry_head(trace, thread, duration, duration_calculated, ttrace->entry_time, trace->output);
2478
2479         if (ttrace->entry_pending) {
2480                 printed = fprintf(trace->output, "%s", ttrace->entry_str);
2481         } else {
2482                 printed += fprintf(trace->output, " ... [");
2483                 color_fprintf(trace->output, PERF_COLOR_YELLOW, "continued");
2484                 printed += 9;
2485                 printed += fprintf(trace->output, "]: %s()", sc->name);
2486         }
2487
2488         printed++; /* the closing ')' */
2489
2490         if (alignment > printed)
2491                 alignment -= printed;
2492         else
2493                 alignment = 0;
2494
2495         fprintf(trace->output, ")%*s= ", alignment, " ");
2496
2497         if (sc->fmt == NULL) {
2498                 if (ret < 0)
2499                         goto errno_print;
2500 signed_print:
2501                 fprintf(trace->output, "%ld", ret);
2502         } else if (ret < 0) {
2503 errno_print: {
2504                 char bf[STRERR_BUFSIZE];
2505                 const char *emsg = str_error_r(-ret, bf, sizeof(bf)),
2506                            *e = errno_to_name(evsel, -ret);
2507
2508                 fprintf(trace->output, "-1 %s (%s)", e, emsg);
2509         }
2510         } else if (ret == 0 && sc->fmt->timeout)
2511                 fprintf(trace->output, "0 (Timeout)");
2512         else if (ttrace->ret_scnprintf) {
2513                 char bf[1024];
2514                 struct syscall_arg arg = {
2515                         .val    = ret,
2516                         .thread = thread,
2517                         .trace  = trace,
2518                 };
2519                 ttrace->ret_scnprintf(bf, sizeof(bf), &arg);
2520                 ttrace->ret_scnprintf = NULL;
2521                 fprintf(trace->output, "%s", bf);
2522         } else if (sc->fmt->hexret)
2523                 fprintf(trace->output, "%#lx", ret);
2524         else if (sc->fmt->errpid) {
2525                 struct thread *child = machine__find_thread(trace->host, ret, ret);
2526
2527                 if (child != NULL) {
2528                         fprintf(trace->output, "%ld", ret);
2529                         if (child->comm_set)
2530                                 fprintf(trace->output, " (%s)", thread__comm_str(child));
2531                         thread__put(child);
2532                 }
2533         } else
2534                 goto signed_print;
2535
2536         fputc('\n', trace->output);
2537
2538         /*
2539          * We only consider an 'event' for the sake of --max-events a non-filtered
2540          * sys_enter + sys_exit and other tracepoint events.
2541          */
2542         if (++trace->nr_events_printed == trace->max_events && trace->max_events != ULONG_MAX)
2543                 interrupted = true;
2544
2545         if (callchain_ret > 0)
2546                 trace__fprintf_callchain(trace, sample);
2547         else if (callchain_ret < 0)
2548                 pr_err("Problem processing %s callchain, skipping...\n", evsel__name(evsel));
2549 out:
2550         ttrace->entry_pending = false;
2551         err = 0;
2552 out_put:
2553         thread__put(thread);
2554         return err;
2555 }
2556
2557 static int trace__vfs_getname(struct trace *trace, struct evsel *evsel,
2558                               union perf_event *event __maybe_unused,
2559                               struct perf_sample *sample)
2560 {
2561         struct thread *thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2562         struct thread_trace *ttrace;
2563         size_t filename_len, entry_str_len, to_move;
2564         ssize_t remaining_space;
2565         char *pos;
2566         const char *filename = evsel__rawptr(evsel, sample, "pathname");
2567
2568         if (!thread)
2569                 goto out;
2570
2571         ttrace = thread__priv(thread);
2572         if (!ttrace)
2573                 goto out_put;
2574
2575         filename_len = strlen(filename);
2576         if (filename_len == 0)
2577                 goto out_put;
2578
2579         if (ttrace->filename.namelen < filename_len) {
2580                 char *f = realloc(ttrace->filename.name, filename_len + 1);
2581
2582                 if (f == NULL)
2583                         goto out_put;
2584
2585                 ttrace->filename.namelen = filename_len;
2586                 ttrace->filename.name = f;
2587         }
2588
2589         strcpy(ttrace->filename.name, filename);
2590         ttrace->filename.pending_open = true;
2591
2592         if (!ttrace->filename.ptr)
2593                 goto out_put;
2594
2595         entry_str_len = strlen(ttrace->entry_str);
2596         remaining_space = trace__entry_str_size - entry_str_len - 1; /* \0 */
2597         if (remaining_space <= 0)
2598                 goto out_put;
2599
2600         if (filename_len > (size_t)remaining_space) {
2601                 filename += filename_len - remaining_space;
2602                 filename_len = remaining_space;
2603         }
2604
2605         to_move = entry_str_len - ttrace->filename.entry_str_pos + 1; /* \0 */
2606         pos = ttrace->entry_str + ttrace->filename.entry_str_pos;
2607         memmove(pos + filename_len, pos, to_move);
2608         memcpy(pos, filename, filename_len);
2609
2610         ttrace->filename.ptr = 0;
2611         ttrace->filename.entry_str_pos = 0;
2612 out_put:
2613         thread__put(thread);
2614 out:
2615         return 0;
2616 }
2617
2618 static int trace__sched_stat_runtime(struct trace *trace, struct evsel *evsel,
2619                                      union perf_event *event __maybe_unused,
2620                                      struct perf_sample *sample)
2621 {
2622         u64 runtime = evsel__intval(evsel, sample, "runtime");
2623         double runtime_ms = (double)runtime / NSEC_PER_MSEC;
2624         struct thread *thread = machine__findnew_thread(trace->host,
2625                                                         sample->pid,
2626                                                         sample->tid);
2627         struct thread_trace *ttrace = thread__trace(thread, trace->output);
2628
2629         if (ttrace == NULL)
2630                 goto out_dump;
2631
2632         ttrace->runtime_ms += runtime_ms;
2633         trace->runtime_ms += runtime_ms;
2634 out_put:
2635         thread__put(thread);
2636         return 0;
2637
2638 out_dump:
2639         fprintf(trace->output, "%s: comm=%s,pid=%u,runtime=%" PRIu64 ",vruntime=%" PRIu64 ")\n",
2640                evsel->name,
2641                evsel__strval(evsel, sample, "comm"),
2642                (pid_t)evsel__intval(evsel, sample, "pid"),
2643                runtime,
2644                evsel__intval(evsel, sample, "vruntime"));
2645         goto out_put;
2646 }
2647
2648 static int bpf_output__printer(enum binary_printer_ops op,
2649                                unsigned int val, void *extra __maybe_unused, FILE *fp)
2650 {
2651         unsigned char ch = (unsigned char)val;
2652
2653         switch (op) {
2654         case BINARY_PRINT_CHAR_DATA:
2655                 return fprintf(fp, "%c", isprint(ch) ? ch : '.');
2656         case BINARY_PRINT_DATA_BEGIN:
2657         case BINARY_PRINT_LINE_BEGIN:
2658         case BINARY_PRINT_ADDR:
2659         case BINARY_PRINT_NUM_DATA:
2660         case BINARY_PRINT_NUM_PAD:
2661         case BINARY_PRINT_SEP:
2662         case BINARY_PRINT_CHAR_PAD:
2663         case BINARY_PRINT_LINE_END:
2664         case BINARY_PRINT_DATA_END:
2665         default:
2666                 break;
2667         }
2668
2669         return 0;
2670 }
2671
2672 static void bpf_output__fprintf(struct trace *trace,
2673                                 struct perf_sample *sample)
2674 {
2675         binary__fprintf(sample->raw_data, sample->raw_size, 8,
2676                         bpf_output__printer, NULL, trace->output);
2677         ++trace->nr_events_printed;
2678 }
2679
2680 static size_t trace__fprintf_tp_fields(struct trace *trace, struct evsel *evsel, struct perf_sample *sample,
2681                                        struct thread *thread, void *augmented_args, int augmented_args_size)
2682 {
2683         char bf[2048];
2684         size_t size = sizeof(bf);
2685         struct tep_format_field *field = evsel->tp_format->format.fields;
2686         struct syscall_arg_fmt *arg = __evsel__syscall_arg_fmt(evsel);
2687         size_t printed = 0;
2688         unsigned long val;
2689         u8 bit = 1;
2690         struct syscall_arg syscall_arg = {
2691                 .augmented = {
2692                         .size = augmented_args_size,
2693                         .args = augmented_args,
2694                 },
2695                 .idx    = 0,
2696                 .mask   = 0,
2697                 .trace  = trace,
2698                 .thread = thread,
2699                 .show_string_prefix = trace->show_string_prefix,
2700         };
2701
2702         for (; field && arg; field = field->next, ++syscall_arg.idx, bit <<= 1, ++arg) {
2703                 if (syscall_arg.mask & bit)
2704                         continue;
2705
2706                 syscall_arg.len = 0;
2707                 syscall_arg.fmt = arg;
2708                 if (field->flags & TEP_FIELD_IS_ARRAY) {
2709                         int offset = field->offset;
2710
2711                         if (field->flags & TEP_FIELD_IS_DYNAMIC) {
2712                                 offset = format_field__intval(field, sample, evsel->needs_swap);
2713                                 syscall_arg.len = offset >> 16;
2714                                 offset &= 0xffff;
2715                         }
2716
2717                         val = (uintptr_t)(sample->raw_data + offset);
2718                 } else
2719                         val = format_field__intval(field, sample, evsel->needs_swap);
2720                 /*
2721                  * Some syscall args need some mask, most don't and
2722                  * return val untouched.
2723                  */
2724                 val = syscall_arg_fmt__mask_val(arg, &syscall_arg, val);
2725
2726                 /*
2727                  * Suppress this argument if its value is zero and
2728                  * and we don't have a string associated in an
2729                  * strarray for it.
2730                  */
2731                 if (val == 0 &&
2732                     !trace->show_zeros &&
2733                     !((arg->show_zero ||
2734                        arg->scnprintf == SCA_STRARRAY ||
2735                        arg->scnprintf == SCA_STRARRAYS) &&
2736                       arg->parm))
2737                         continue;
2738
2739                 printed += scnprintf(bf + printed, size - printed, "%s", printed ? ", " : "");
2740
2741                 /*
2742                  * XXX Perhaps we should have a show_tp_arg_names,
2743                  * leaving show_arg_names just for syscalls?
2744                  */
2745                 if (1 || trace->show_arg_names)
2746                         printed += scnprintf(bf + printed, size - printed, "%s: ", field->name);
2747
2748                 printed += syscall_arg_fmt__scnprintf_val(arg, bf + printed, size - printed, &syscall_arg, val);
2749         }
2750
2751         return printed + fprintf(trace->output, "%s", bf);
2752 }
2753
2754 static int trace__event_handler(struct trace *trace, struct evsel *evsel,
2755                                 union perf_event *event __maybe_unused,
2756                                 struct perf_sample *sample)
2757 {
2758         struct thread *thread;
2759         int callchain_ret = 0;
2760         /*
2761          * Check if we called perf_evsel__disable(evsel) due to, for instance,
2762          * this event's max_events having been hit and this is an entry coming
2763          * from the ring buffer that we should discard, since the max events
2764          * have already been considered/printed.
2765          */
2766         if (evsel->disabled)
2767                 return 0;
2768
2769         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2770
2771         if (sample->callchain) {
2772                 callchain_ret = trace__resolve_callchain(trace, evsel, sample, &callchain_cursor);
2773                 if (callchain_ret == 0) {
2774                         if (callchain_cursor.nr < trace->min_stack)
2775                                 goto out;
2776                         callchain_ret = 1;
2777                 }
2778         }
2779
2780         trace__printf_interrupted_entry(trace);
2781         trace__fprintf_tstamp(trace, sample->time, trace->output);
2782
2783         if (trace->trace_syscalls && trace->show_duration)
2784                 fprintf(trace->output, "(         ): ");
2785
2786         if (thread)
2787                 trace__fprintf_comm_tid(trace, thread, trace->output);
2788
2789         if (evsel == trace->syscalls.events.augmented) {
2790                 int id = perf_evsel__sc_tp_uint(evsel, id, sample);
2791                 struct syscall *sc = trace__syscall_info(trace, evsel, id);
2792
2793                 if (sc) {
2794                         fprintf(trace->output, "%s(", sc->name);
2795                         trace__fprintf_sys_enter(trace, evsel, sample);
2796                         fputc(')', trace->output);
2797                         goto newline;
2798                 }
2799
2800                 /*
2801                  * XXX: Not having the associated syscall info or not finding/adding
2802                  *      the thread should never happen, but if it does...
2803                  *      fall thru and print it as a bpf_output event.
2804                  */
2805         }
2806
2807         fprintf(trace->output, "%s(", evsel->name);
2808
2809         if (evsel__is_bpf_output(evsel)) {
2810                 bpf_output__fprintf(trace, sample);
2811         } else if (evsel->tp_format) {
2812                 if (strncmp(evsel->tp_format->name, "sys_enter_", 10) ||
2813                     trace__fprintf_sys_enter(trace, evsel, sample)) {
2814                         if (trace->libtraceevent_print) {
2815                                 event_format__fprintf(evsel->tp_format, sample->cpu,
2816                                                       sample->raw_data, sample->raw_size,
2817                                                       trace->output);
2818                         } else {
2819                                 trace__fprintf_tp_fields(trace, evsel, sample, thread, NULL, 0);
2820                         }
2821                 }
2822         }
2823
2824 newline:
2825         fprintf(trace->output, ")\n");
2826
2827         if (callchain_ret > 0)
2828                 trace__fprintf_callchain(trace, sample);
2829         else if (callchain_ret < 0)
2830                 pr_err("Problem processing %s callchain, skipping...\n", evsel__name(evsel));
2831
2832         ++trace->nr_events_printed;
2833
2834         if (evsel->max_events != ULONG_MAX && ++evsel->nr_events_printed == evsel->max_events) {
2835                 evsel__disable(evsel);
2836                 evsel__close(evsel);
2837         }
2838 out:
2839         thread__put(thread);
2840         return 0;
2841 }
2842
2843 static void print_location(FILE *f, struct perf_sample *sample,
2844                            struct addr_location *al,
2845                            bool print_dso, bool print_sym)
2846 {
2847
2848         if ((verbose > 0 || print_dso) && al->map)
2849                 fprintf(f, "%s@", al->map->dso->long_name);
2850
2851         if ((verbose > 0 || print_sym) && al->sym)
2852                 fprintf(f, "%s+0x%" PRIx64, al->sym->name,
2853                         al->addr - al->sym->start);
2854         else if (al->map)
2855                 fprintf(f, "0x%" PRIx64, al->addr);
2856         else
2857                 fprintf(f, "0x%" PRIx64, sample->addr);
2858 }
2859
2860 static int trace__pgfault(struct trace *trace,
2861                           struct evsel *evsel,
2862                           union perf_event *event __maybe_unused,
2863                           struct perf_sample *sample)
2864 {
2865         struct thread *thread;
2866         struct addr_location al;
2867         char map_type = 'd';
2868         struct thread_trace *ttrace;
2869         int err = -1;
2870         int callchain_ret = 0;
2871
2872         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2873
2874         if (sample->callchain) {
2875                 callchain_ret = trace__resolve_callchain(trace, evsel, sample, &callchain_cursor);
2876                 if (callchain_ret == 0) {
2877                         if (callchain_cursor.nr < trace->min_stack)
2878                                 goto out_put;
2879                         callchain_ret = 1;
2880                 }
2881         }
2882
2883         ttrace = thread__trace(thread, trace->output);
2884         if (ttrace == NULL)
2885                 goto out_put;
2886
2887         if (evsel->core.attr.config == PERF_COUNT_SW_PAGE_FAULTS_MAJ)
2888                 ttrace->pfmaj++;
2889         else
2890                 ttrace->pfmin++;
2891
2892         if (trace->summary_only)
2893                 goto out;
2894
2895         thread__find_symbol(thread, sample->cpumode, sample->ip, &al);
2896
2897         trace__fprintf_entry_head(trace, thread, 0, true, sample->time, trace->output);
2898
2899         fprintf(trace->output, "%sfault [",
2900                 evsel->core.attr.config == PERF_COUNT_SW_PAGE_FAULTS_MAJ ?
2901                 "maj" : "min");
2902
2903         print_location(trace->output, sample, &al, false, true);
2904
2905         fprintf(trace->output, "] => ");
2906
2907         thread__find_symbol(thread, sample->cpumode, sample->addr, &al);
2908
2909         if (!al.map) {
2910                 thread__find_symbol(thread, sample->cpumode, sample->addr, &al);
2911
2912                 if (al.map)
2913                         map_type = 'x';
2914                 else
2915                         map_type = '?';
2916         }
2917
2918         print_location(trace->output, sample, &al, true, false);
2919
2920         fprintf(trace->output, " (%c%c)\n", map_type, al.level);
2921
2922         if (callchain_ret > 0)
2923                 trace__fprintf_callchain(trace, sample);
2924         else if (callchain_ret < 0)
2925                 pr_err("Problem processing %s callchain, skipping...\n", evsel__name(evsel));
2926
2927         ++trace->nr_events_printed;
2928 out:
2929         err = 0;
2930 out_put:
2931         thread__put(thread);
2932         return err;
2933 }
2934
2935 static void trace__set_base_time(struct trace *trace,
2936                                  struct evsel *evsel,
2937                                  struct perf_sample *sample)
2938 {
2939         /*
2940          * BPF events were not setting PERF_SAMPLE_TIME, so be more robust
2941          * and don't use sample->time unconditionally, we may end up having
2942          * some other event in the future without PERF_SAMPLE_TIME for good
2943          * reason, i.e. we may not be interested in its timestamps, just in
2944          * it taking place, picking some piece of information when it
2945          * appears in our event stream (vfs_getname comes to mind).
2946          */
2947         if (trace->base_time == 0 && !trace->full_time &&
2948             (evsel->core.attr.sample_type & PERF_SAMPLE_TIME))
2949                 trace->base_time = sample->time;
2950 }
2951
2952 static int trace__process_sample(struct perf_tool *tool,
2953                                  union perf_event *event,
2954                                  struct perf_sample *sample,
2955                                  struct evsel *evsel,
2956                                  struct machine *machine __maybe_unused)
2957 {
2958         struct trace *trace = container_of(tool, struct trace, tool);
2959         struct thread *thread;
2960         int err = 0;
2961
2962         tracepoint_handler handler = evsel->handler;
2963
2964         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2965         if (thread && thread__is_filtered(thread))
2966                 goto out;
2967
2968         trace__set_base_time(trace, evsel, sample);
2969
2970         if (handler) {
2971                 ++trace->nr_events;
2972                 handler(trace, evsel, event, sample);
2973         }
2974 out:
2975         thread__put(thread);
2976         return err;
2977 }
2978
2979 static int trace__record(struct trace *trace, int argc, const char **argv)
2980 {
2981         unsigned int rec_argc, i, j;
2982         const char **rec_argv;
2983         const char * const record_args[] = {
2984                 "record",
2985                 "-R",
2986                 "-m", "1024",
2987                 "-c", "1",
2988         };
2989         pid_t pid = getpid();
2990         char *filter = asprintf__tp_filter_pids(1, &pid);
2991         const char * const sc_args[] = { "-e", };
2992         unsigned int sc_args_nr = ARRAY_SIZE(sc_args);
2993         const char * const majpf_args[] = { "-e", "major-faults" };
2994         unsigned int majpf_args_nr = ARRAY_SIZE(majpf_args);
2995         const char * const minpf_args[] = { "-e", "minor-faults" };
2996         unsigned int minpf_args_nr = ARRAY_SIZE(minpf_args);
2997         int err = -1;
2998
2999         /* +3 is for the event string below and the pid filter */
3000         rec_argc = ARRAY_SIZE(record_args) + sc_args_nr + 3 +
3001                 majpf_args_nr + minpf_args_nr + argc;
3002         rec_argv = calloc(rec_argc + 1, sizeof(char *));
3003
3004         if (rec_argv == NULL || filter == NULL)
3005                 goto out_free;
3006
3007         j = 0;
3008         for (i = 0; i < ARRAY_SIZE(record_args); i++)
3009                 rec_argv[j++] = record_args[i];
3010
3011         if (trace->trace_syscalls) {
3012                 for (i = 0; i < sc_args_nr; i++)
3013                         rec_argv[j++] = sc_args[i];
3014
3015                 /* event string may be different for older kernels - e.g., RHEL6 */
3016                 if (is_valid_tracepoint("raw_syscalls:sys_enter"))
3017                         rec_argv[j++] = "raw_syscalls:sys_enter,raw_syscalls:sys_exit";
3018                 else if (is_valid_tracepoint("syscalls:sys_enter"))
3019                         rec_argv[j++] = "syscalls:sys_enter,syscalls:sys_exit";
3020                 else {
3021                         pr_err("Neither raw_syscalls nor syscalls events exist.\n");
3022                         goto out_free;
3023                 }
3024         }
3025
3026         rec_argv[j++] = "--filter";
3027         rec_argv[j++] = filter;
3028
3029         if (trace->trace_pgfaults & TRACE_PFMAJ)
3030                 for (i = 0; i < majpf_args_nr; i++)
3031                         rec_argv[j++] = majpf_args[i];
3032
3033         if (trace->trace_pgfaults & TRACE_PFMIN)
3034                 for (i = 0; i < minpf_args_nr; i++)
3035                         rec_argv[j++] = minpf_args[i];
3036
3037         for (i = 0; i < (unsigned int)argc; i++)
3038                 rec_argv[j++] = argv[i];
3039
3040         err = cmd_record(j, rec_argv);
3041 out_free:
3042         free(filter);
3043         free(rec_argv);
3044         return err;
3045 }
3046
3047 static size_t trace__fprintf_thread_summary(struct trace *trace, FILE *fp);
3048
3049 static bool evlist__add_vfs_getname(struct evlist *evlist)
3050 {
3051         bool found = false;
3052         struct evsel *evsel, *tmp;
3053         struct parse_events_error err;
3054         int ret;
3055
3056         bzero(&err, sizeof(err));
3057         ret = parse_events(evlist, "probe:vfs_getname*", &err);
3058         if (ret) {
3059                 free(err.str);
3060                 free(err.help);
3061                 free(err.first_str);
3062                 free(err.first_help);
3063                 return false;
3064         }
3065
3066         evlist__for_each_entry_safe(evlist, evsel, tmp) {
3067                 if (!strstarts(evsel__name(evsel), "probe:vfs_getname"))
3068                         continue;
3069
3070                 if (evsel__field(evsel, "pathname")) {
3071                         evsel->handler = trace__vfs_getname;
3072                         found = true;
3073                         continue;
3074                 }
3075
3076                 list_del_init(&evsel->core.node);
3077                 evsel->evlist = NULL;
3078                 evsel__delete(evsel);
3079         }
3080
3081         return found;
3082 }
3083
3084 static struct evsel *evsel__new_pgfault(u64 config)
3085 {
3086         struct evsel *evsel;
3087         struct perf_event_attr attr = {
3088                 .type = PERF_TYPE_SOFTWARE,
3089                 .mmap_data = 1,
3090         };
3091
3092         attr.config = config;
3093         attr.sample_period = 1;
3094
3095         event_attr_init(&attr);
3096
3097         evsel = evsel__new(&attr);
3098         if (evsel)
3099                 evsel->handler = trace__pgfault;
3100
3101         return evsel;
3102 }
3103
3104 static void trace__handle_event(struct trace *trace, union perf_event *event, struct perf_sample *sample)
3105 {
3106         const u32 type = event->header.type;
3107         struct evsel *evsel;
3108
3109         if (type != PERF_RECORD_SAMPLE) {
3110                 trace__process_event(trace, trace->host, event, sample);
3111                 return;
3112         }
3113
3114         evsel = perf_evlist__id2evsel(trace->evlist, sample->id);
3115         if (evsel == NULL) {
3116                 fprintf(trace->output, "Unknown tp ID %" PRIu64 ", skipping...\n", sample->id);
3117                 return;
3118         }
3119
3120         if (evswitch__discard(&trace->evswitch, evsel))
3121                 return;
3122
3123         trace__set_base_time(trace, evsel, sample);
3124
3125         if (evsel->core.attr.type == PERF_TYPE_TRACEPOINT &&
3126             sample->raw_data == NULL) {
3127                 fprintf(trace->output, "%s sample with no payload for tid: %d, cpu %d, raw_size=%d, skipping...\n",
3128                        evsel__name(evsel), sample->tid,
3129                        sample->cpu, sample->raw_size);
3130         } else {
3131                 tracepoint_handler handler = evsel->handler;
3132                 handler(trace, evsel, event, sample);
3133         }
3134
3135         if (trace->nr_events_printed >= trace->max_events && trace->max_events != ULONG_MAX)
3136                 interrupted = true;
3137 }
3138
3139 static int trace__add_syscall_newtp(struct trace *trace)
3140 {
3141         int ret = -1;
3142         struct evlist *evlist = trace->evlist;
3143         struct evsel *sys_enter, *sys_exit;
3144
3145         sys_enter = perf_evsel__raw_syscall_newtp("sys_enter", trace__sys_enter);
3146         if (sys_enter == NULL)
3147                 goto out;
3148
3149         if (perf_evsel__init_sc_tp_ptr_field(sys_enter, args))
3150                 goto out_delete_sys_enter;
3151
3152         sys_exit = perf_evsel__raw_syscall_newtp("sys_exit", trace__sys_exit);
3153         if (sys_exit == NULL)
3154                 goto out_delete_sys_enter;
3155
3156         if (perf_evsel__init_sc_tp_uint_field(sys_exit, ret))
3157                 goto out_delete_sys_exit;
3158
3159         evsel__config_callchain(sys_enter, &trace->opts, &callchain_param);
3160         evsel__config_callchain(sys_exit, &trace->opts, &callchain_param);
3161
3162         evlist__add(evlist, sys_enter);
3163         evlist__add(evlist, sys_exit);
3164
3165         if (callchain_param.enabled && !trace->kernel_syscallchains) {
3166                 /*
3167                  * We're interested only in the user space callchain
3168                  * leading to the syscall, allow overriding that for
3169                  * debugging reasons using --kernel_syscall_callchains
3170                  */
3171                 sys_exit->core.attr.exclude_callchain_kernel = 1;
3172         }
3173
3174         trace->syscalls.events.sys_enter = sys_enter;
3175         trace->syscalls.events.sys_exit  = sys_exit;
3176
3177         ret = 0;
3178 out:
3179         return ret;
3180
3181 out_delete_sys_exit:
3182         evsel__delete_priv(sys_exit);
3183 out_delete_sys_enter:
3184         evsel__delete_priv(sys_enter);
3185         goto out;
3186 }
3187
3188 static int trace__set_ev_qualifier_tp_filter(struct trace *trace)
3189 {
3190         int err = -1;
3191         struct evsel *sys_exit;
3192         char *filter = asprintf_expr_inout_ints("id", !trace->not_ev_qualifier,
3193                                                 trace->ev_qualifier_ids.nr,
3194                                                 trace->ev_qualifier_ids.entries);
3195
3196         if (filter == NULL)
3197                 goto out_enomem;
3198
3199         if (!evsel__append_tp_filter(trace->syscalls.events.sys_enter, filter)) {
3200                 sys_exit = trace->syscalls.events.sys_exit;
3201                 err = evsel__append_tp_filter(sys_exit, filter);
3202         }
3203
3204         free(filter);
3205 out:
3206         return err;
3207 out_enomem:
3208         errno = ENOMEM;
3209         goto out;
3210 }
3211
3212 #ifdef HAVE_LIBBPF_SUPPORT
3213 static struct bpf_map *trace__find_bpf_map_by_name(struct trace *trace, const char *name)
3214 {
3215         if (trace->bpf_obj == NULL)
3216                 return NULL;
3217
3218         return bpf_object__find_map_by_name(trace->bpf_obj, name);
3219 }
3220
3221 static void trace__set_bpf_map_filtered_pids(struct trace *trace)
3222 {
3223         trace->filter_pids.map = trace__find_bpf_map_by_name(trace, "pids_filtered");
3224 }
3225
3226 static void trace__set_bpf_map_syscalls(struct trace *trace)
3227 {
3228         trace->syscalls.map = trace__find_bpf_map_by_name(trace, "syscalls");
3229         trace->syscalls.prog_array.sys_enter = trace__find_bpf_map_by_name(trace, "syscalls_sys_enter");
3230         trace->syscalls.prog_array.sys_exit  = trace__find_bpf_map_by_name(trace, "syscalls_sys_exit");
3231 }
3232
3233 static struct bpf_program *trace__find_bpf_program_by_title(struct trace *trace, const char *name)
3234 {
3235         if (trace->bpf_obj == NULL)
3236                 return NULL;
3237
3238         return bpf_object__find_program_by_title(trace->bpf_obj, name);
3239 }
3240
3241 static struct bpf_program *trace__find_syscall_bpf_prog(struct trace *trace, struct syscall *sc,
3242                                                         const char *prog_name, const char *type)
3243 {
3244         struct bpf_program *prog;
3245
3246         if (prog_name == NULL) {
3247                 char default_prog_name[256];
3248                 scnprintf(default_prog_name, sizeof(default_prog_name), "!syscalls:sys_%s_%s", type, sc->name);
3249                 prog = trace__find_bpf_program_by_title(trace, default_prog_name);
3250                 if (prog != NULL)
3251                         goto out_found;
3252                 if (sc->fmt && sc->fmt->alias) {
3253                         scnprintf(default_prog_name, sizeof(default_prog_name), "!syscalls:sys_%s_%s", type, sc->fmt->alias);
3254                         prog = trace__find_bpf_program_by_title(trace, default_prog_name);
3255                         if (prog != NULL)
3256                                 goto out_found;
3257                 }
3258                 goto out_unaugmented;
3259         }
3260
3261         prog = trace__find_bpf_program_by_title(trace, prog_name);
3262
3263         if (prog != NULL) {
3264 out_found:
3265                 return prog;
3266         }
3267
3268         pr_debug("Couldn't find BPF prog \"%s\" to associate with syscalls:sys_%s_%s, not augmenting it\n",
3269                  prog_name, type, sc->name);
3270 out_unaugmented:
3271         return trace->syscalls.unaugmented_prog;
3272 }
3273
3274 static void trace__init_syscall_bpf_progs(struct trace *trace, int id)
3275 {
3276         struct syscall *sc = trace__syscall_info(trace, NULL, id);
3277
3278         if (sc == NULL)
3279                 return;
3280
3281         sc->bpf_prog.sys_enter = trace__find_syscall_bpf_prog(trace, sc, sc->fmt ? sc->fmt->bpf_prog_name.sys_enter : NULL, "enter");
3282         sc->bpf_prog.sys_exit  = trace__find_syscall_bpf_prog(trace, sc, sc->fmt ? sc->fmt->bpf_prog_name.sys_exit  : NULL,  "exit");
3283 }
3284
3285 static int trace__bpf_prog_sys_enter_fd(struct trace *trace, int id)
3286 {
3287         struct syscall *sc = trace__syscall_info(trace, NULL, id);
3288         return sc ? bpf_program__fd(sc->bpf_prog.sys_enter) : bpf_program__fd(trace->syscalls.unaugmented_prog);
3289 }
3290
3291 static int trace__bpf_prog_sys_exit_fd(struct trace *trace, int id)
3292 {
3293         struct syscall *sc = trace__syscall_info(trace, NULL, id);
3294         return sc ? bpf_program__fd(sc->bpf_prog.sys_exit) : bpf_program__fd(trace->syscalls.unaugmented_prog);
3295 }
3296
3297 static void trace__init_bpf_map_syscall_args(struct trace *trace, int id, struct bpf_map_syscall_entry *entry)
3298 {
3299         struct syscall *sc = trace__syscall_info(trace, NULL, id);
3300         int arg = 0;
3301
3302         if (sc == NULL)
3303                 goto out;
3304
3305         for (; arg < sc->nr_args; ++arg) {
3306                 entry->string_args_len[arg] = 0;
3307                 if (sc->arg_fmt[arg].scnprintf == SCA_FILENAME) {
3308                         /* Should be set like strace -s strsize */
3309                         entry->string_args_len[arg] = PATH_MAX;
3310                 }
3311         }
3312 out:
3313         for (; arg < 6; ++arg)
3314                 entry->string_args_len[arg] = 0;
3315 }
3316 static int trace__set_ev_qualifier_bpf_filter(struct trace *trace)
3317 {
3318         int fd = bpf_map__fd(trace->syscalls.map);
3319         struct bpf_map_syscall_entry value = {
3320                 .enabled = !trace->not_ev_qualifier,
3321         };
3322         int err = 0;
3323         size_t i;
3324
3325         for (i = 0; i < trace->ev_qualifier_ids.nr; ++i) {
3326                 int key = trace->ev_qualifier_ids.entries[i];
3327
3328                 if (value.enabled) {
3329                         trace__init_bpf_map_syscall_args(trace, key, &value);
3330                         trace__init_syscall_bpf_progs(trace, key);
3331                 }
3332
3333                 err = bpf_map_update_elem(fd, &key, &value, BPF_EXIST);
3334                 if (err)
3335                         break;
3336         }
3337
3338         return err;
3339 }
3340
3341 static int __trace__init_syscalls_bpf_map(struct trace *trace, bool enabled)
3342 {
3343         int fd = bpf_map__fd(trace->syscalls.map);
3344         struct bpf_map_syscall_entry value = {
3345                 .enabled = enabled,
3346         };
3347         int err = 0, key;
3348
3349         for (key = 0; key < trace->sctbl->syscalls.nr_entries; ++key) {
3350                 if (enabled)
3351                         trace__init_bpf_map_syscall_args(trace, key, &value);
3352
3353                 err = bpf_map_update_elem(fd, &key, &value, BPF_ANY);
3354                 if (err)
3355                         break;
3356         }
3357
3358         return err;
3359 }
3360
3361 static int trace__init_syscalls_bpf_map(struct trace *trace)
3362 {
3363         bool enabled = true;
3364
3365         if (trace->ev_qualifier_ids.nr)
3366                 enabled = trace->not_ev_qualifier;
3367
3368         return __trace__init_syscalls_bpf_map(trace, enabled);
3369 }
3370
3371 static struct bpf_program *trace__find_usable_bpf_prog_entry(struct trace *trace, struct syscall *sc)
3372 {
3373         struct tep_format_field *field, *candidate_field;
3374         int id;
3375
3376         /*
3377          * We're only interested in syscalls that have a pointer:
3378          */
3379         for (field = sc->args; field; field = field->next) {
3380                 if (field->flags & TEP_FIELD_IS_POINTER)
3381                         goto try_to_find_pair;
3382         }
3383
3384         return NULL;
3385
3386 try_to_find_pair:
3387         for (id = 0; id < trace->sctbl->syscalls.nr_entries; ++id) {
3388                 struct syscall *pair = trace__syscall_info(trace, NULL, id);
3389                 struct bpf_program *pair_prog;
3390                 bool is_candidate = false;
3391
3392                 if (pair == NULL || pair == sc ||
3393                     pair->bpf_prog.sys_enter == trace->syscalls.unaugmented_prog)
3394                         continue;
3395
3396                 for (field = sc->args, candidate_field = pair->args;
3397                      field && candidate_field; field = field->next, candidate_field = candidate_field->next) {
3398                         bool is_pointer = field->flags & TEP_FIELD_IS_POINTER,
3399                              candidate_is_pointer = candidate_field->flags & TEP_FIELD_IS_POINTER;
3400
3401                         if (is_pointer) {
3402                                if (!candidate_is_pointer) {
3403                                         // The candidate just doesn't copies our pointer arg, might copy other pointers we want.
3404                                         continue;
3405                                }
3406                         } else {
3407                                 if (candidate_is_pointer) {
3408                                         // The candidate might copy a pointer we don't have, skip it.
3409                                         goto next_candidate;
3410                                 }
3411                                 continue;
3412                         }
3413
3414                         if (strcmp(field->type, candidate_field->type))
3415                                 goto next_candidate;
3416
3417                         is_candidate = true;
3418                 }
3419
3420                 if (!is_candidate)
3421                         goto next_candidate;
3422
3423                 /*
3424                  * Check if the tentative pair syscall augmenter has more pointers, if it has,
3425                  * then it may be collecting that and we then can't use it, as it would collect
3426                  * more than what is common to the two syscalls.
3427                  */
3428                 if (candidate_field) {
3429                         for (candidate_field = candidate_field->next; candidate_field; candidate_field = candidate_field->next)
3430                                 if (candidate_field->flags & TEP_FIELD_IS_POINTER)
3431                                         goto next_candidate;
3432                 }
3433
3434                 pair_prog = pair->bpf_prog.sys_enter;
3435                 /*
3436                  * If the pair isn't enabled, then its bpf_prog.sys_enter will not
3437                  * have been searched for, so search it here and if it returns the
3438                  * unaugmented one, then ignore it, otherwise we'll reuse that BPF
3439                  * program for a filtered syscall on a non-filtered one.
3440                  *
3441                  * For instance, we have "!syscalls:sys_enter_renameat" and that is
3442                  * useful for "renameat2".
3443                  */
3444                 if (pair_prog == NULL) {
3445                         pair_prog = trace__find_syscall_bpf_prog(trace, pair, pair->fmt ? pair->fmt->bpf_prog_name.sys_enter : NULL, "enter");
3446                         if (pair_prog == trace->syscalls.unaugmented_prog)
3447                                 goto next_candidate;
3448                 }
3449
3450                 pr_debug("Reusing \"%s\" BPF sys_enter augmenter for \"%s\"\n", pair->name, sc->name);
3451                 return pair_prog;
3452         next_candidate:
3453                 continue;
3454         }
3455
3456         return NULL;
3457 }
3458
3459 static int trace__init_syscalls_bpf_prog_array_maps(struct trace *trace)
3460 {
3461         int map_enter_fd = bpf_map__fd(trace->syscalls.prog_array.sys_enter),
3462             map_exit_fd  = bpf_map__fd(trace->syscalls.prog_array.sys_exit);
3463         int err = 0, key;
3464
3465         for (key = 0; key < trace->sctbl->syscalls.nr_entries; ++key) {
3466                 int prog_fd;
3467
3468                 if (!trace__syscall_enabled(trace, key))
3469                         continue;
3470
3471                 trace__init_syscall_bpf_progs(trace, key);
3472
3473                 // It'll get at least the "!raw_syscalls:unaugmented"
3474                 prog_fd = trace__bpf_prog_sys_enter_fd(trace, key);
3475                 err = bpf_map_update_elem(map_enter_fd, &key, &prog_fd, BPF_ANY);
3476                 if (err)
3477                         break;
3478                 prog_fd = trace__bpf_prog_sys_exit_fd(trace, key);
3479                 err = bpf_map_update_elem(map_exit_fd, &key, &prog_fd, BPF_ANY);
3480                 if (err)
3481                         break;
3482         }
3483
3484         /*
3485          * Now lets do a second pass looking for enabled syscalls without
3486          * an augmenter that have a signature that is a superset of another
3487          * syscall with an augmenter so that we can auto-reuse it.
3488          *
3489          * I.e. if we have an augmenter for the "open" syscall that has
3490          * this signature:
3491          *
3492          *   int open(const char *pathname, int flags, mode_t mode);
3493          *
3494          * I.e. that will collect just the first string argument, then we
3495          * can reuse it for the 'creat' syscall, that has this signature:
3496          *
3497          *   int creat(const char *pathname, mode_t mode);
3498          *
3499          * and for:
3500          *
3501          *   int stat(const char *pathname, struct stat *statbuf);
3502          *   int lstat(const char *pathname, struct stat *statbuf);
3503          *
3504          * Because the 'open' augmenter will collect the first arg as a string,
3505          * and leave alone all the other args, which already helps with
3506          * beautifying 'stat' and 'lstat''s pathname arg.
3507          *
3508          * Then, in time, when 'stat' gets an augmenter that collects both
3509          * first and second arg (this one on the raw_syscalls:sys_exit prog
3510          * array tail call, then that one will be used.
3511          */
3512         for (key = 0; key < trace->sctbl->syscalls.nr_entries; ++key) {
3513                 struct syscall *sc = trace__syscall_info(trace, NULL, key);
3514                 struct bpf_program *pair_prog;
3515                 int prog_fd;
3516
3517                 if (sc == NULL || sc->bpf_prog.sys_enter == NULL)
3518                         continue;
3519
3520                 /*
3521                  * For now we're just reusing the sys_enter prog, and if it
3522                  * already has an augmenter, we don't need to find one.
3523                  */
3524                 if (sc->bpf_prog.sys_enter != trace->syscalls.unaugmented_prog)
3525                         continue;
3526
3527                 /*
3528                  * Look at all the other syscalls for one that has a signature
3529                  * that is close enough that we can share:
3530                  */
3531                 pair_prog = trace__find_usable_bpf_prog_entry(trace, sc);
3532                 if (pair_prog == NULL)
3533                         continue;
3534
3535                 sc->bpf_prog.sys_enter = pair_prog;
3536
3537                 /*
3538                  * Update the BPF_MAP_TYPE_PROG_SHARED for raw_syscalls:sys_enter
3539                  * with the fd for the program we're reusing:
3540                  */
3541                 prog_fd = bpf_program__fd(sc->bpf_prog.sys_enter);
3542                 err = bpf_map_update_elem(map_enter_fd, &key, &prog_fd, BPF_ANY);
3543                 if (err)
3544                         break;
3545         }
3546
3547
3548         return err;
3549 }
3550
3551 static void trace__delete_augmented_syscalls(struct trace *trace)
3552 {
3553         struct evsel *evsel, *tmp;
3554
3555         evlist__remove(trace->evlist, trace->syscalls.events.augmented);
3556         evsel__delete(trace->syscalls.events.augmented);
3557         trace->syscalls.events.augmented = NULL;
3558
3559         evlist__for_each_entry_safe(trace->evlist, tmp, evsel) {
3560                 if (evsel->bpf_obj == trace->bpf_obj) {
3561                         evlist__remove(trace->evlist, evsel);
3562                         evsel__delete(evsel);
3563                 }
3564
3565         }
3566
3567         bpf_object__close(trace->bpf_obj);
3568         trace->bpf_obj = NULL;
3569 }
3570 #else // HAVE_LIBBPF_SUPPORT
3571 static struct bpf_map *trace__find_bpf_map_by_name(struct trace *trace __maybe_unused,
3572                                                    const char *name __maybe_unused)
3573 {
3574         return NULL;
3575 }
3576
3577 static void trace__set_bpf_map_filtered_pids(struct trace *trace __maybe_unused)
3578 {
3579 }
3580
3581 static void trace__set_bpf_map_syscalls(struct trace *trace __maybe_unused)
3582 {
3583 }
3584
3585 static int trace__set_ev_qualifier_bpf_filter(struct trace *trace __maybe_unused)
3586 {
3587         return 0;
3588 }
3589
3590 static int trace__init_syscalls_bpf_map(struct trace *trace __maybe_unused)
3591 {
3592         return 0;
3593 }
3594
3595 static struct bpf_program *trace__find_bpf_program_by_title(struct trace *trace __maybe_unused,
3596                                                             const char *name __maybe_unused)
3597 {
3598         return NULL;
3599 }
3600
3601 static int trace__init_syscalls_bpf_prog_array_maps(struct trace *trace __maybe_unused)
3602 {
3603         return 0;
3604 }
3605
3606 static void trace__delete_augmented_syscalls(struct trace *trace __maybe_unused)
3607 {
3608 }
3609 #endif // HAVE_LIBBPF_SUPPORT
3610
3611 static bool trace__only_augmented_syscalls_evsels(struct trace *trace)
3612 {
3613         struct evsel *evsel;
3614
3615         evlist__for_each_entry(trace->evlist, evsel) {
3616                 if (evsel == trace->syscalls.events.augmented ||
3617                     evsel->bpf_obj == trace->bpf_obj)
3618                         continue;
3619
3620                 return false;
3621         }
3622
3623         return true;
3624 }
3625
3626 static int trace__set_ev_qualifier_filter(struct trace *trace)
3627 {
3628         if (trace->syscalls.map)
3629                 return trace__set_ev_qualifier_bpf_filter(trace);
3630         if (trace->syscalls.events.sys_enter)
3631                 return trace__set_ev_qualifier_tp_filter(trace);
3632         return 0;
3633 }
3634
3635 static int bpf_map__set_filter_pids(struct bpf_map *map __maybe_unused,
3636                                     size_t npids __maybe_unused, pid_t *pids __maybe_unused)
3637 {
3638         int err = 0;
3639 #ifdef HAVE_LIBBPF_SUPPORT
3640         bool value = true;
3641         int map_fd = bpf_map__fd(map);
3642         size_t i;
3643
3644         for (i = 0; i < npids; ++i) {
3645                 err = bpf_map_update_elem(map_fd, &pids[i], &value, BPF_ANY);
3646                 if (err)
3647                         break;
3648         }
3649 #endif
3650         return err;
3651 }
3652
3653 static int trace__set_filter_loop_pids(struct trace *trace)
3654 {
3655         unsigned int nr = 1, err;
3656         pid_t pids[32] = {
3657                 getpid(),
3658         };
3659         struct thread *thread = machine__find_thread(trace->host, pids[0], pids[0]);
3660
3661         while (thread && nr < ARRAY_SIZE(pids)) {
3662                 struct thread *parent = machine__find_thread(trace->host, thread->ppid, thread->ppid);
3663
3664                 if (parent == NULL)
3665                         break;
3666
3667                 if (!strcmp(thread__comm_str(parent), "sshd") ||
3668                     strstarts(thread__comm_str(parent), "gnome-terminal")) {
3669                         pids[nr++] = parent->tid;
3670                         break;
3671                 }
3672                 thread = parent;
3673         }
3674
3675         err = perf_evlist__append_tp_filter_pids(trace->evlist, nr, pids);
3676         if (!err && trace->filter_pids.map)
3677                 err = bpf_map__set_filter_pids(trace->filter_pids.map, nr, pids);
3678
3679         return err;
3680 }
3681
3682 static int trace__set_filter_pids(struct trace *trace)
3683 {
3684         int err = 0;
3685         /*
3686          * Better not use !target__has_task() here because we need to cover the
3687          * case where no threads were specified in the command line, but a
3688          * workload was, and in that case we will fill in the thread_map when
3689          * we fork the workload in perf_evlist__prepare_workload.
3690          */
3691         if (trace->filter_pids.nr > 0) {
3692                 err = perf_evlist__append_tp_filter_pids(trace->evlist, trace->filter_pids.nr,
3693                                                          trace->filter_pids.entries);
3694                 if (!err && trace->filter_pids.map) {
3695                         err = bpf_map__set_filter_pids(trace->filter_pids.map, trace->filter_pids.nr,
3696                                                        trace->filter_pids.entries);
3697                 }
3698         } else if (perf_thread_map__pid(trace->evlist->core.threads, 0) == -1) {
3699                 err = trace__set_filter_loop_pids(trace);
3700         }
3701
3702         return err;
3703 }
3704
3705 static int __trace__deliver_event(struct trace *trace, union perf_event *event)
3706 {
3707         struct evlist *evlist = trace->evlist;
3708         struct perf_sample sample;
3709         int err;
3710
3711         err = perf_evlist__parse_sample(evlist, event, &sample);
3712         if (err)
3713                 fprintf(trace->output, "Can't parse sample, err = %d, skipping...\n", err);
3714         else
3715                 trace__handle_event(trace, event, &sample);
3716
3717         return 0;
3718 }
3719
3720 static int __trace__flush_events(struct trace *trace)
3721 {
3722         u64 first = ordered_events__first_time(&trace->oe.data);
3723         u64 flush = trace->oe.last - NSEC_PER_SEC;
3724
3725         /* Is there some thing to flush.. */
3726         if (first && first < flush)
3727                 return ordered_events__flush_time(&trace->oe.data, flush);
3728
3729         return 0;
3730 }
3731
3732 static int trace__flush_events(struct trace *trace)
3733 {
3734         return !trace->sort_events ? 0 : __trace__flush_events(trace);
3735 }
3736
3737 static int trace__deliver_event(struct trace *trace, union perf_event *event)
3738 {
3739         int err;
3740
3741         if (!trace->sort_events)
3742                 return __trace__deliver_event(trace, event);
3743
3744         err = perf_evlist__parse_sample_timestamp(trace->evlist, event, &trace->oe.last);
3745         if (err && err != -1)
3746                 return err;
3747
3748         err = ordered_events__queue(&trace->oe.data, event, trace->oe.last, 0);
3749         if (err)
3750                 return err;
3751
3752         return trace__flush_events(trace);
3753 }
3754
3755 static int ordered_events__deliver_event(struct ordered_events *oe,
3756                                          struct ordered_event *event)
3757 {
3758         struct trace *trace = container_of(oe, struct trace, oe.data);
3759
3760         return __trace__deliver_event(trace, event->event);
3761 }
3762
3763 static struct syscall_arg_fmt *evsel__find_syscall_arg_fmt_by_name(struct evsel *evsel, char *arg)
3764 {
3765         struct tep_format_field *field;
3766         struct syscall_arg_fmt *fmt = __evsel__syscall_arg_fmt(evsel);
3767
3768         if (evsel->tp_format == NULL || fmt == NULL)
3769                 return NULL;
3770
3771         for (field = evsel->tp_format->format.fields; field; field = field->next, ++fmt)
3772                 if (strcmp(field->name, arg) == 0)
3773                         return fmt;
3774
3775         return NULL;
3776 }
3777
3778 static int trace__expand_filter(struct trace *trace __maybe_unused, struct evsel *evsel)
3779 {
3780         char *tok, *left = evsel->filter, *new_filter = evsel->filter;
3781
3782         while ((tok = strpbrk(left, "=<>!")) != NULL) {
3783                 char *right = tok + 1, *right_end;
3784
3785                 if (*right == '=')
3786                         ++right;
3787
3788                 while (isspace(*right))
3789                         ++right;
3790
3791                 if (*right == '\0')
3792                         break;
3793
3794                 while (!isalpha(*left))
3795                         if (++left == tok) {
3796                                 /*
3797                                  * Bail out, can't find the name of the argument that is being
3798                                  * used in the filter, let it try to set this filter, will fail later.
3799                                  */
3800                                 return 0;
3801                         }
3802
3803                 right_end = right + 1;
3804                 while (isalnum(*right_end) || *right_end == '_' || *right_end == '|')
3805                         ++right_end;
3806
3807                 if (isalpha(*right)) {
3808                         struct syscall_arg_fmt *fmt;
3809                         int left_size = tok - left,
3810                             right_size = right_end - right;
3811                         char arg[128];
3812
3813                         while (isspace(left[left_size - 1]))
3814                                 --left_size;
3815
3816                         scnprintf(arg, sizeof(arg), "%.*s", left_size, left);
3817
3818                         fmt = evsel__find_syscall_arg_fmt_by_name(evsel, arg);
3819                         if (fmt == NULL) {
3820                                 pr_err("\"%s\" not found in \"%s\", can't set filter \"%s\"\n",
3821                                        arg, evsel->name, evsel->filter);
3822                                 return -1;
3823                         }
3824
3825                         pr_debug2("trying to expand \"%s\" \"%.*s\" \"%.*s\" -> ",
3826                                  arg, (int)(right - tok), tok, right_size, right);
3827
3828                         if (fmt->strtoul) {
3829                                 u64 val;
3830                                 struct syscall_arg syscall_arg = {
3831                                         .parm = fmt->parm,
3832                                 };
3833
3834                                 if (fmt->strtoul(right, right_size, &syscall_arg, &val)) {
3835                                         char *n, expansion[19];
3836                                         int expansion_lenght = scnprintf(expansion, sizeof(expansion), "%#" PRIx64, val);
3837                                         int expansion_offset = right - new_filter;
3838
3839                                         pr_debug("%s", expansion);
3840
3841                                         if (asprintf(&n, "%.*s%s%s", expansion_offset, new_filter, expansion, right_end) < 0) {
3842                                                 pr_debug(" out of memory!\n");
3843                                                 free(new_filter);
3844                                                 return -1;
3845                                         }
3846                                         if (new_filter != evsel->filter)
3847                                                 free(new_filter);
3848                                         left = n + expansion_offset + expansion_lenght;
3849                                         new_filter = n;
3850                                 } else {
3851                                         pr_err("\"%.*s\" not found for \"%s\" in \"%s\", can't set filter \"%s\"\n",
3852                                                right_size, right, arg, evsel->name, evsel->filter);
3853                                         return -1;
3854                                 }
3855                         } else {
3856                                 pr_err("No resolver (strtoul) for \"%s\" in \"%s\", can't set filter \"%s\"\n",
3857                                        arg, evsel->name, evsel->filter);
3858                                 return -1;
3859                         }
3860
3861                         pr_debug("\n");
3862                 } else {
3863                         left = right_end;
3864                 }
3865         }
3866
3867         if (new_filter != evsel->filter) {
3868                 pr_debug("New filter for %s: %s\n", evsel->name, new_filter);
3869                 evsel__set_filter(evsel, new_filter);
3870                 free(new_filter);
3871         }
3872
3873         return 0;
3874 }
3875
3876 static int trace__expand_filters(struct trace *trace, struct evsel **err_evsel)
3877 {
3878         struct evlist *evlist = trace->evlist;
3879         struct evsel *evsel;
3880
3881         evlist__for_each_entry(evlist, evsel) {
3882                 if (evsel->filter == NULL)
3883                         continue;
3884
3885                 if (trace__expand_filter(trace, evsel)) {
3886                         *err_evsel = evsel;
3887                         return -1;
3888                 }
3889         }
3890
3891         return 0;
3892 }
3893
3894 static int trace__run(struct trace *trace, int argc, const char **argv)
3895 {
3896         struct evlist *evlist = trace->evlist;
3897         struct evsel *evsel, *pgfault_maj = NULL, *pgfault_min = NULL;
3898         int err = -1, i;
3899         unsigned long before;
3900         const bool forks = argc > 0;
3901         bool draining = false;
3902
3903         trace->live = true;
3904
3905         if (!trace->raw_augmented_syscalls) {
3906                 if (trace->trace_syscalls && trace__add_syscall_newtp(trace))
3907                         goto out_error_raw_syscalls;
3908
3909                 if (trace->trace_syscalls)
3910                         trace->vfs_getname = evlist__add_vfs_getname(evlist);
3911         }
3912
3913         if ((trace->trace_pgfaults & TRACE_PFMAJ)) {
3914                 pgfault_maj = evsel__new_pgfault(PERF_COUNT_SW_PAGE_FAULTS_MAJ);
3915                 if (pgfault_maj == NULL)
3916                         goto out_error_mem;
3917                 evsel__config_callchain(pgfault_maj, &trace->opts, &callchain_param);
3918                 evlist__add(evlist, pgfault_maj);
3919         }
3920
3921         if ((trace->trace_pgfaults & TRACE_PFMIN)) {
3922                 pgfault_min = evsel__new_pgfault(PERF_COUNT_SW_PAGE_FAULTS_MIN);
3923                 if (pgfault_min == NULL)
3924                         goto out_error_mem;
3925                 evsel__config_callchain(pgfault_min, &trace->opts, &callchain_param);
3926                 evlist__add(evlist, pgfault_min);
3927         }
3928
3929         if (trace->sched &&
3930             evlist__add_newtp(evlist, "sched", "sched_stat_runtime", trace__sched_stat_runtime))
3931                 goto out_error_sched_stat_runtime;
3932         /*
3933          * If a global cgroup was set, apply it to all the events without an
3934          * explicit cgroup. I.e.:
3935          *
3936          *      trace -G A -e sched:*switch
3937          *
3938          * Will set all raw_syscalls:sys_{enter,exit}, pgfault, vfs_getname, etc
3939          * _and_ sched:sched_switch to the 'A' cgroup, while:
3940          *
3941          * trace -e sched:*switch -G A
3942          *
3943          * will only set the sched:sched_switch event to the 'A' cgroup, all the
3944          * other events (raw_syscalls:sys_{enter,exit}, etc are left "without"
3945          * a cgroup (on the root cgroup, sys wide, etc).
3946          *
3947          * Multiple cgroups:
3948          *
3949          * trace -G A -e sched:*switch -G B
3950          *
3951          * the syscall ones go to the 'A' cgroup, the sched:sched_switch goes
3952          * to the 'B' cgroup.
3953          *
3954          * evlist__set_default_cgroup() grabs a reference of the passed cgroup
3955          * only for the evsels still without a cgroup, i.e. evsel->cgroup == NULL.
3956          */
3957         if (trace->cgroup)
3958                 evlist__set_default_cgroup(trace->evlist, trace->cgroup);
3959
3960         err = perf_evlist__create_maps(evlist, &trace->opts.target);
3961         if (err < 0) {
3962                 fprintf(trace->output, "Problems parsing the target to trace, check your options!\n");
3963                 goto out_delete_evlist;
3964         }
3965
3966         err = trace__symbols_init(trace, evlist);
3967         if (err < 0) {
3968                 fprintf(trace->output, "Problems initializing symbol libraries!\n");
3969                 goto out_delete_evlist;
3970         }
3971
3972         perf_evlist__config(evlist, &trace->opts, &callchain_param);
3973
3974         signal(SIGCHLD, sig_handler);
3975         signal(SIGINT, sig_handler);
3976
3977         if (forks) {
3978                 err = perf_evlist__prepare_workload(evlist, &trace->opts.target,
3979                                                     argv, false, NULL);
3980                 if (err < 0) {
3981                         fprintf(trace->output, "Couldn't run the workload!\n");
3982                         goto out_delete_evlist;
3983                 }
3984         }
3985
3986         err = evlist__open(evlist);
3987         if (err < 0)
3988                 goto out_error_open;
3989
3990         err = bpf__apply_obj_config();
3991         if (err) {
3992                 char errbuf[BUFSIZ];
3993
3994                 bpf__strerror_apply_obj_config(err, errbuf, sizeof(errbuf));
3995                 pr_err("ERROR: Apply config to BPF failed: %s\n",
3996                          errbuf);
3997                 goto out_error_open;
3998         }
3999
4000         err = trace__set_filter_pids(trace);
4001         if (err < 0)
4002                 goto out_error_mem;
4003
4004         if (trace->syscalls.map)
4005                 trace__init_syscalls_bpf_map(trace);
4006
4007         if (trace->syscalls.prog_array.sys_enter)
4008                 trace__init_syscalls_bpf_prog_array_maps(trace);
4009
4010         if (trace->ev_qualifier_ids.nr > 0) {
4011                 err = trace__set_ev_qualifier_filter(trace);
4012                 if (err < 0)
4013                         goto out_errno;
4014
4015                 if (trace->syscalls.events.sys_exit) {
4016                         pr_debug("event qualifier tracepoint filter: %s\n",
4017                                  trace->syscalls.events.sys_exit->filter);
4018                 }
4019         }
4020
4021         /*
4022          * If the "close" syscall is not traced, then we will not have the
4023          * opportunity to, in syscall_arg__scnprintf_close_fd() invalidate the
4024          * fd->pathname table and were ending up showing the last value set by
4025          * syscalls opening a pathname and associating it with a descriptor or
4026          * reading it from /proc/pid/fd/ in cases where that doesn't make
4027          * sense.
4028          *
4029          *  So just disable this beautifier (SCA_FD, SCA_FDAT) when 'close' is
4030          *  not in use.
4031          */
4032         trace->fd_path_disabled = !trace__syscall_enabled(trace, syscalltbl__id(trace->sctbl, "close"));
4033
4034         err = trace__expand_filters(trace, &evsel);
4035         if (err)
4036                 goto out_delete_evlist;
4037         err = perf_evlist__apply_filters(evlist, &evsel);
4038         if (err < 0)
4039                 goto out_error_apply_filters;
4040
4041         if (trace->dump.map)
4042                 bpf_map__fprintf(trace->dump.map, trace->output);
4043
4044         err = evlist__mmap(evlist, trace->opts.mmap_pages);
4045         if (err < 0)
4046                 goto out_error_mmap;
4047
4048         if (!target__none(&trace->opts.target) && !trace->opts.initial_delay)
4049                 evlist__enable(evlist);
4050
4051         if (forks)
4052                 perf_evlist__start_workload(evlist);
4053
4054         if (trace->opts.initial_delay) {
4055                 usleep(trace->opts.initial_delay * 1000);
4056                 evlist__enable(evlist);
4057         }
4058
4059         trace->multiple_threads = perf_thread_map__pid(evlist->core.threads, 0) == -1 ||
4060                                   evlist->core.threads->nr > 1 ||
4061                                   evlist__first(evlist)->core.attr.inherit;
4062
4063         /*
4064          * Now that we already used evsel->core.attr to ask the kernel to setup the
4065          * events, lets reuse evsel->core.attr.sample_max_stack as the limit in
4066          * trace__resolve_callchain(), allowing per-event max-stack settings
4067          * to override an explicitly set --max-stack global setting.
4068          */
4069         evlist__for_each_entry(evlist, evsel) {
4070                 if (evsel__has_callchain(evsel) &&
4071                     evsel->core.attr.sample_max_stack == 0)
4072                         evsel->core.attr.sample_max_stack = trace->max_stack;
4073         }
4074 again:
4075         before = trace->nr_events;
4076
4077         for (i = 0; i < evlist->core.nr_mmaps; i++) {
4078                 union perf_event *event;
4079                 struct mmap *md;
4080
4081                 md = &evlist->mmap[i];
4082                 if (perf_mmap__read_init(&md->core) < 0)
4083                         continue;
4084
4085                 while ((event = perf_mmap__read_event(&md->core)) != NULL) {
4086                         ++trace->nr_events;
4087
4088                         err = trace__deliver_event(trace, event);
4089                         if (err)
4090                                 goto out_disable;
4091
4092                         perf_mmap__consume(&md->core);
4093
4094                         if (interrupted)
4095                                 goto out_disable;
4096
4097                         if (done && !draining) {
4098                                 evlist__disable(evlist);
4099                                 draining = true;
4100                         }
4101                 }
4102                 perf_mmap__read_done(&md->core);
4103         }
4104
4105         if (trace->nr_events == before) {
4106                 int timeout = done ? 100 : -1;
4107
4108                 if (!draining && evlist__poll(evlist, timeout) > 0) {
4109                         if (evlist__filter_pollfd(evlist, POLLERR | POLLHUP | POLLNVAL) == 0)
4110                                 draining = true;
4111
4112                         goto again;
4113                 } else {
4114                         if (trace__flush_events(trace))
4115                                 goto out_disable;
4116                 }
4117         } else {
4118                 goto again;
4119         }
4120
4121 out_disable:
4122         thread__zput(trace->current);
4123
4124         evlist__disable(evlist);
4125
4126         if (trace->sort_events)
4127                 ordered_events__flush(&trace->oe.data, OE_FLUSH__FINAL);
4128
4129         if (!err) {
4130                 if (trace->summary)
4131                         trace__fprintf_thread_summary(trace, trace->output);
4132
4133                 if (trace->show_tool_stats) {
4134                         fprintf(trace->output, "Stats:\n "
4135                                                " vfs_getname : %" PRIu64 "\n"
4136                                                " proc_getname: %" PRIu64 "\n",
4137                                 trace->stats.vfs_getname,
4138                                 trace->stats.proc_getname);
4139                 }
4140         }
4141
4142 out_delete_evlist:
4143         trace__symbols__exit(trace);
4144
4145         evlist__delete(evlist);
4146         cgroup__put(trace->cgroup);
4147         trace->evlist = NULL;
4148         trace->live = false;
4149         return err;
4150 {
4151         char errbuf[BUFSIZ];
4152
4153 out_error_sched_stat_runtime:
4154         tracing_path__strerror_open_tp(errno, errbuf, sizeof(errbuf), "sched", "sched_stat_runtime");
4155         goto out_error;
4156
4157 out_error_raw_syscalls:
4158         tracing_path__strerror_open_tp(errno, errbuf, sizeof(errbuf), "raw_syscalls", "sys_(enter|exit)");
4159         goto out_error;
4160
4161 out_error_mmap:
4162         evlist__strerror_mmap(evlist, errno, errbuf, sizeof(errbuf));
4163         goto out_error;
4164
4165 out_error_open:
4166         evlist__strerror_open(evlist, errno, errbuf, sizeof(errbuf));
4167
4168 out_error:
4169         fprintf(trace->output, "%s\n", errbuf);
4170         goto out_delete_evlist;
4171
4172 out_error_apply_filters:
4173         fprintf(trace->output,
4174                 "Failed to set filter \"%s\" on event %s with %d (%s)\n",
4175                 evsel->filter, evsel__name(evsel), errno,
4176                 str_error_r(errno, errbuf, sizeof(errbuf)));
4177         goto out_delete_evlist;
4178 }
4179 out_error_mem:
4180         fprintf(trace->output, "Not enough memory to run!\n");
4181         goto out_delete_evlist;
4182
4183 out_errno:
4184         fprintf(trace->output, "errno=%d,%s\n", errno, strerror(errno));
4185         goto out_delete_evlist;
4186 }
4187
4188 static int trace__replay(struct trace *trace)
4189 {
4190         const struct evsel_str_handler handlers[] = {
4191                 { "probe:vfs_getname",       trace__vfs_getname, },
4192         };
4193         struct perf_data data = {
4194                 .path  = input_name,
4195                 .mode  = PERF_DATA_MODE_READ,
4196                 .force = trace->force,
4197         };
4198         struct perf_session *session;
4199         struct evsel *evsel;
4200         int err = -1;
4201
4202         trace->tool.sample        = trace__process_sample;
4203         trace->tool.mmap          = perf_event__process_mmap;
4204         trace->tool.mmap2         = perf_event__process_mmap2;
4205         trace->tool.comm          = perf_event__process_comm;
4206         trace->tool.exit          = perf_event__process_exit;
4207         trace->tool.fork          = perf_event__process_fork;
4208         trace->tool.attr          = perf_event__process_attr;
4209         trace->tool.tracing_data  = perf_event__process_tracing_data;
4210         trace->tool.build_id      = perf_event__process_build_id;
4211         trace->tool.namespaces    = perf_event__process_namespaces;
4212
4213         trace->tool.ordered_events = true;
4214         trace->tool.ordering_requires_timestamps = true;
4215
4216         /* add tid to output */
4217         trace->multiple_threads = true;
4218
4219         session = perf_session__new(&data, false, &trace->tool);
4220         if (IS_ERR(session))
4221                 return PTR_ERR(session);
4222
4223         if (trace->opts.target.pid)
4224                 symbol_conf.pid_list_str = strdup(trace->opts.target.pid);
4225
4226         if (trace->opts.target.tid)
4227                 symbol_conf.tid_list_str = strdup(trace->opts.target.tid);
4228
4229         if (symbol__init(&session->header.env) < 0)
4230                 goto out;
4231
4232         trace->host = &session->machines.host;
4233
4234         err = perf_session__set_tracepoints_handlers(session, handlers);
4235         if (err)
4236                 goto out;
4237
4238         evsel = perf_evlist__find_tracepoint_by_name(session->evlist,
4239                                                      "raw_syscalls:sys_enter");
4240         /* older kernels have syscalls tp versus raw_syscalls */
4241         if (evsel == NULL)
4242                 evsel = perf_evlist__find_tracepoint_by_name(session->evlist,
4243                                                              "syscalls:sys_enter");
4244
4245         if (evsel &&
4246             (evsel__init_raw_syscall_tp(evsel, trace__sys_enter) < 0 ||
4247             perf_evsel__init_sc_tp_ptr_field(evsel, args))) {
4248                 pr_err("Error during initialize raw_syscalls:sys_enter event\n");
4249                 goto out;
4250         }
4251
4252         evsel = perf_evlist__find_tracepoint_by_name(session->evlist,
4253                                                      "raw_syscalls:sys_exit");
4254         if (evsel == NULL)
4255                 evsel = perf_evlist__find_tracepoint_by_name(session->evlist,
4256                                                              "syscalls:sys_exit");
4257         if (evsel &&
4258             (evsel__init_raw_syscall_tp(evsel, trace__sys_exit) < 0 ||
4259             perf_evsel__init_sc_tp_uint_field(evsel, ret))) {
4260                 pr_err("Error during initialize raw_syscalls:sys_exit event\n");
4261                 goto out;
4262         }
4263
4264         evlist__for_each_entry(session->evlist, evsel) {
4265                 if (evsel->core.attr.type == PERF_TYPE_SOFTWARE &&
4266                     (evsel->core.attr.config == PERF_COUNT_SW_PAGE_FAULTS_MAJ ||
4267                      evsel->core.attr.config == PERF_COUNT_SW_PAGE_FAULTS_MIN ||
4268                      evsel->core.attr.config == PERF_COUNT_SW_PAGE_FAULTS))
4269                         evsel->handler = trace__pgfault;
4270         }
4271
4272         setup_pager();
4273
4274         err = perf_session__process_events(session);
4275         if (err)
4276                 pr_err("Failed to process events, error %d", err);
4277
4278         else if (trace->summary)
4279                 trace__fprintf_thread_summary(trace, trace->output);
4280
4281 out:
4282         perf_session__delete(session);
4283
4284         return err;
4285 }
4286
4287 static size_t trace__fprintf_threads_header(FILE *fp)
4288 {
4289         size_t printed;
4290
4291         printed  = fprintf(fp, "\n Summary of events:\n\n");
4292
4293         return printed;
4294 }
4295
4296 DEFINE_RESORT_RB(syscall_stats, a->msecs > b->msecs,
4297         struct syscall_stats *stats;
4298         double               msecs;
4299         int                  syscall;
4300 )
4301 {
4302         struct int_node *source = rb_entry(nd, struct int_node, rb_node);
4303         struct syscall_stats *stats = source->priv;
4304
4305         entry->syscall = source->i;
4306         entry->stats   = stats;
4307         entry->msecs   = stats ? (u64)stats->stats.n * (avg_stats(&stats->stats) / NSEC_PER_MSEC) : 0;
4308 }
4309
4310 static size_t thread__dump_stats(struct thread_trace *ttrace,
4311                                  struct trace *trace, FILE *fp)
4312 {
4313         size_t printed = 0;
4314         struct syscall *sc;
4315         struct rb_node *nd;
4316         DECLARE_RESORT_RB_INTLIST(syscall_stats, ttrace->syscall_stats);
4317
4318         if (syscall_stats == NULL)
4319                 return 0;
4320
4321         printed += fprintf(fp, "\n");
4322
4323         printed += fprintf(fp, "   syscall            calls  errors  total       min       avg       max       stddev\n");
4324         printed += fprintf(fp, "                                     (msec)    (msec)    (msec)    (msec)        (%%)\n");
4325         printed += fprintf(fp, "   --------------- --------  ------ -------- --------- --------- ---------     ------\n");
4326
4327         resort_rb__for_each_entry(nd, syscall_stats) {
4328                 struct syscall_stats *stats = syscall_stats_entry->stats;
4329                 if (stats) {
4330                         double min = (double)(stats->stats.min) / NSEC_PER_MSEC;
4331                         double max = (double)(stats->stats.max) / NSEC_PER_MSEC;
4332                         double avg = avg_stats(&stats->stats);
4333                         double pct;
4334                         u64 n = (u64)stats->stats.n;
4335
4336                         pct = avg ? 100.0 * stddev_stats(&stats->stats) / avg : 0.0;
4337                         avg /= NSEC_PER_MSEC;
4338
4339                         sc = &trace->syscalls.table[syscall_stats_entry->syscall];
4340                         printed += fprintf(fp, "   %-15s", sc->name);
4341                         printed += fprintf(fp, " %8" PRIu64 " %6" PRIu64 " %9.3f %9.3f %9.3f",
4342                                            n, stats->nr_failures, syscall_stats_entry->msecs, min, avg);
4343                         printed += fprintf(fp, " %9.3f %9.2f%%\n", max, pct);
4344
4345                         if (trace->errno_summary && stats->nr_failures) {
4346                                 const char *arch_name = perf_env__arch(trace->host->env);
4347                                 int e;
4348
4349                                 for (e = 0; e < stats->max_errno; ++e) {
4350                                         if (stats->errnos[e] != 0)
4351                                                 fprintf(fp, "\t\t\t\t%s: %d\n", arch_syscalls__strerrno(arch_name, e + 1), stats->errnos[e]);
4352                                 }
4353                         }
4354                 }
4355         }
4356
4357         resort_rb__delete(syscall_stats);
4358         printed += fprintf(fp, "\n\n");
4359
4360         return printed;
4361 }
4362
4363 static size_t trace__fprintf_thread(FILE *fp, struct thread *thread, struct trace *trace)
4364 {
4365         size_t printed = 0;
4366         struct thread_trace *ttrace = thread__priv(thread);
4367         double ratio;
4368
4369         if (ttrace == NULL)
4370                 return 0;
4371
4372         ratio = (double)ttrace->nr_events / trace->nr_events * 100.0;
4373
4374         printed += fprintf(fp, " %s (%d), ", thread__comm_str(thread), thread->tid);
4375         printed += fprintf(fp, "%lu events, ", ttrace->nr_events);
4376         printed += fprintf(fp, "%.1f%%", ratio);
4377         if (ttrace->pfmaj)
4378                 printed += fprintf(fp, ", %lu majfaults", ttrace->pfmaj);
4379         if (ttrace->pfmin)
4380                 printed += fprintf(fp, ", %lu minfaults", ttrace->pfmin);
4381         if (trace->sched)
4382                 printed += fprintf(fp, ", %.3f msec\n", ttrace->runtime_ms);
4383         else if (fputc('\n', fp) != EOF)
4384                 ++printed;
4385
4386         printed += thread__dump_stats(ttrace, trace, fp);
4387
4388         return printed;
4389 }
4390
4391 static unsigned long thread__nr_events(struct thread_trace *ttrace)
4392 {
4393         return ttrace ? ttrace->nr_events : 0;
4394 }
4395
4396 DEFINE_RESORT_RB(threads, (thread__nr_events(a->thread->priv) < thread__nr_events(b->thread->priv)),
4397         struct thread *thread;
4398 )
4399 {
4400         entry->thread = rb_entry(nd, struct thread, rb_node);
4401 }
4402
4403 static size_t trace__fprintf_thread_summary(struct trace *trace, FILE *fp)
4404 {
4405         size_t printed = trace__fprintf_threads_header(fp);
4406         struct rb_node *nd;
4407         int i;
4408
4409         for (i = 0; i < THREADS__TABLE_SIZE; i++) {
4410                 DECLARE_RESORT_RB_MACHINE_THREADS(threads, trace->host, i);
4411
4412                 if (threads == NULL) {
4413                         fprintf(fp, "%s", "Error sorting output by nr_events!\n");
4414                         return 0;
4415                 }
4416
4417                 resort_rb__for_each_entry(nd, threads)
4418                         printed += trace__fprintf_thread(fp, threads_entry->thread, trace);
4419
4420                 resort_rb__delete(threads);
4421         }
4422         return printed;
4423 }
4424
4425 static int trace__set_duration(const struct option *opt, const char *str,
4426                                int unset __maybe_unused)
4427 {
4428         struct trace *trace = opt->value;
4429
4430         trace->duration_filter = atof(str);
4431         return 0;
4432 }
4433
4434 static int trace__set_filter_pids_from_option(const struct option *opt, const char *str,
4435                                               int unset __maybe_unused)
4436 {
4437         int ret = -1;
4438         size_t i;
4439         struct trace *trace = opt->value;
4440         /*
4441          * FIXME: introduce a intarray class, plain parse csv and create a
4442          * { int nr, int entries[] } struct...
4443          */
4444         struct intlist *list = intlist__new(str);
4445
4446         if (list == NULL)
4447                 return -1;
4448
4449         i = trace->filter_pids.nr = intlist__nr_entries(list) + 1;
4450         trace->filter_pids.entries = calloc(i, sizeof(pid_t));
4451
4452         if (trace->filter_pids.entries == NULL)
4453                 goto out;
4454
4455         trace->filter_pids.entries[0] = getpid();
4456
4457         for (i = 1; i < trace->filter_pids.nr; ++i)
4458                 trace->filter_pids.entries[i] = intlist__entry(list, i - 1)->i;
4459
4460         intlist__delete(list);
4461         ret = 0;
4462 out:
4463         return ret;
4464 }
4465
4466 static int trace__open_output(struct trace *trace, const char *filename)
4467 {
4468         struct stat st;
4469
4470         if (!stat(filename, &st) && st.st_size) {
4471                 char oldname[PATH_MAX];
4472
4473                 scnprintf(oldname, sizeof(oldname), "%s.old", filename);
4474                 unlink(oldname);
4475                 rename(filename, oldname);
4476         }
4477
4478         trace->output = fopen(filename, "w");
4479
4480         return trace->output == NULL ? -errno : 0;
4481 }
4482
4483 static int parse_pagefaults(const struct option *opt, const char *str,
4484                             int unset __maybe_unused)
4485 {
4486         int *trace_pgfaults = opt->value;
4487
4488         if (strcmp(str, "all") == 0)
4489                 *trace_pgfaults |= TRACE_PFMAJ | TRACE_PFMIN;
4490         else if (strcmp(str, "maj") == 0)
4491                 *trace_pgfaults |= TRACE_PFMAJ;
4492         else if (strcmp(str, "min") == 0)
4493                 *trace_pgfaults |= TRACE_PFMIN;
4494         else
4495                 return -1;
4496
4497         return 0;
4498 }
4499
4500 static void evlist__set_default_evsel_handler(struct evlist *evlist, void *handler)
4501 {
4502         struct evsel *evsel;
4503
4504         evlist__for_each_entry(evlist, evsel) {
4505                 if (evsel->handler == NULL)
4506                         evsel->handler = handler;
4507         }
4508 }
4509
4510 static void evsel__set_syscall_arg_fmt(struct evsel *evsel, const char *name)
4511 {
4512         struct syscall_arg_fmt *fmt = evsel__syscall_arg_fmt(evsel);
4513
4514         if (fmt) {
4515                 struct syscall_fmt *scfmt = syscall_fmt__find(name);
4516
4517                 if (scfmt) {
4518                         int skip = 0;
4519
4520                         if (strcmp(evsel->tp_format->format.fields->name, "__syscall_nr") == 0 ||
4521                             strcmp(evsel->tp_format->format.fields->name, "nr") == 0)
4522                                 ++skip;
4523
4524                         memcpy(fmt + skip, scfmt->arg, (evsel->tp_format->format.nr_fields - skip) * sizeof(*fmt));
4525                 }
4526         }
4527 }
4528
4529 static int evlist__set_syscall_tp_fields(struct evlist *evlist)
4530 {
4531         struct evsel *evsel;
4532
4533         evlist__for_each_entry(evlist, evsel) {
4534                 if (evsel->priv || !evsel->tp_format)
4535                         continue;
4536
4537                 if (strcmp(evsel->tp_format->system, "syscalls")) {
4538                         evsel__init_tp_arg_scnprintf(evsel);
4539                         continue;
4540                 }
4541
4542                 if (evsel__init_syscall_tp(evsel))
4543                         return -1;
4544
4545                 if (!strncmp(evsel->tp_format->name, "sys_enter_", 10)) {
4546                         struct syscall_tp *sc = __evsel__syscall_tp(evsel);
4547
4548                         if (__tp_field__init_ptr(&sc->args, sc->id.offset + sizeof(u64)))
4549                                 return -1;
4550
4551                         evsel__set_syscall_arg_fmt(evsel, evsel->tp_format->name + sizeof("sys_enter_") - 1);
4552                 } else if (!strncmp(evsel->tp_format->name, "sys_exit_", 9)) {
4553                         struct syscall_tp *sc = __evsel__syscall_tp(evsel);
4554
4555                         if (__tp_field__init_uint(&sc->ret, sizeof(u64), sc->id.offset + sizeof(u64), evsel->needs_swap))
4556                                 return -1;
4557
4558                         evsel__set_syscall_arg_fmt(evsel, evsel->tp_format->name + sizeof("sys_exit_") - 1);
4559                 }
4560         }
4561
4562         return 0;
4563 }
4564
4565 /*
4566  * XXX: Hackish, just splitting the combined -e+--event (syscalls
4567  * (raw_syscalls:{sys_{enter,exit}} + events (tracepoints, HW, SW, etc) to use
4568  * existing facilities unchanged (trace->ev_qualifier + parse_options()).
4569  *
4570  * It'd be better to introduce a parse_options() variant that would return a
4571  * list with the terms it didn't match to an event...
4572  */
4573 static int trace__parse_events_option(const struct option *opt, const char *str,
4574                                       int unset __maybe_unused)
4575 {
4576         struct trace *trace = (struct trace *)opt->value;
4577         const char *s = str;
4578         char *sep = NULL, *lists[2] = { NULL, NULL, };
4579         int len = strlen(str) + 1, err = -1, list, idx;
4580         char *strace_groups_dir = system_path(STRACE_GROUPS_DIR);
4581         char group_name[PATH_MAX];
4582         struct syscall_fmt *fmt;
4583
4584         if (strace_groups_dir == NULL)
4585                 return -1;
4586
4587         if (*s == '!') {
4588                 ++s;
4589                 trace->not_ev_qualifier = true;
4590         }
4591
4592         while (1) {
4593                 if ((sep = strchr(s, ',')) != NULL)
4594                         *sep = '\0';
4595
4596                 list = 0;
4597                 if (syscalltbl__id(trace->sctbl, s) >= 0 ||
4598                     syscalltbl__strglobmatch_first(trace->sctbl, s, &idx) >= 0) {
4599                         list = 1;
4600                         goto do_concat;
4601                 }
4602
4603                 fmt = syscall_fmt__find_by_alias(s);
4604                 if (fmt != NULL) {
4605                         list = 1;
4606                         s = fmt->name;
4607                 } else {
4608                         path__join(group_name, sizeof(group_name), strace_groups_dir, s);
4609                         if (access(group_name, R_OK) == 0)
4610                                 list = 1;
4611                 }
4612 do_concat:
4613                 if (lists[list]) {
4614                         sprintf(lists[list] + strlen(lists[list]), ",%s", s);
4615                 } else {
4616                         lists[list] = malloc(len);
4617                         if (lists[list] == NULL)
4618                                 goto out;
4619                         strcpy(lists[list], s);
4620                 }
4621
4622                 if (!sep)
4623                         break;
4624
4625                 *sep = ',';
4626                 s = sep + 1;
4627         }
4628
4629         if (lists[1] != NULL) {
4630                 struct strlist_config slist_config = {
4631                         .dirname = strace_groups_dir,
4632                 };
4633
4634                 trace->ev_qualifier = strlist__new(lists[1], &slist_config);
4635                 if (trace->ev_qualifier == NULL) {
4636                         fputs("Not enough memory to parse event qualifier", trace->output);
4637                         goto out;
4638                 }
4639
4640                 if (trace__validate_ev_qualifier(trace))
4641                         goto out;
4642                 trace->trace_syscalls = true;
4643         }
4644
4645         err = 0;
4646
4647         if (lists[0]) {
4648                 struct option o = {
4649                         .value = &trace->evlist,
4650                 };
4651                 err = parse_events_option(&o, lists[0], 0);
4652         }
4653 out:
4654         if (sep)
4655                 *sep = ',';
4656
4657         return err;
4658 }
4659
4660 static int trace__parse_cgroups(const struct option *opt, const char *str, int unset)
4661 {
4662         struct trace *trace = opt->value;
4663
4664         if (!list_empty(&trace->evlist->core.entries)) {
4665                 struct option o = {
4666                         .value = &trace->evlist,
4667                 };
4668                 return parse_cgroups(&o, str, unset);
4669         }
4670         trace->cgroup = evlist__findnew_cgroup(trace->evlist, str);
4671
4672         return 0;
4673 }
4674
4675 static int trace__config(const char *var, const char *value, void *arg)
4676 {
4677         struct trace *trace = arg;
4678         int err = 0;
4679
4680         if (!strcmp(var, "trace.add_events")) {
4681                 trace->perfconfig_events = strdup(value);
4682                 if (trace->perfconfig_events == NULL) {
4683                         pr_err("Not enough memory for %s\n", "trace.add_events");
4684                         return -1;
4685                 }
4686         } else if (!strcmp(var, "trace.show_timestamp")) {
4687                 trace->show_tstamp = perf_config_bool(var, value);
4688         } else if (!strcmp(var, "trace.show_duration")) {
4689                 trace->show_duration = perf_config_bool(var, value);
4690         } else if (!strcmp(var, "trace.show_arg_names")) {
4691                 trace->show_arg_names = perf_config_bool(var, value);
4692                 if (!trace->show_arg_names)
4693                         trace->show_zeros = true;
4694         } else if (!strcmp(var, "trace.show_zeros")) {
4695                 bool new_show_zeros = perf_config_bool(var, value);
4696                 if (!trace->show_arg_names && !new_show_zeros) {
4697                         pr_warning("trace.show_zeros has to be set when trace.show_arg_names=no\n");
4698                         goto out;
4699                 }
4700                 trace->show_zeros = new_show_zeros;
4701         } else if (!strcmp(var, "trace.show_prefix")) {
4702                 trace->show_string_prefix = perf_config_bool(var, value);
4703         } else if (!strcmp(var, "trace.no_inherit")) {
4704                 trace->opts.no_inherit = perf_config_bool(var, value);
4705         } else if (!strcmp(var, "trace.args_alignment")) {
4706                 int args_alignment = 0;
4707                 if (perf_config_int(&args_alignment, var, value) == 0)
4708                         trace->args_alignment = args_alignment;
4709         } else if (!strcmp(var, "trace.tracepoint_beautifiers")) {
4710                 if (strcasecmp(value, "libtraceevent") == 0)
4711                         trace->libtraceevent_print = true;
4712                 else if (strcasecmp(value, "libbeauty") == 0)
4713                         trace->libtraceevent_print = false;
4714         }
4715 out:
4716         return err;
4717 }
4718
4719 int cmd_trace(int argc, const char **argv)
4720 {
4721         const char *trace_usage[] = {
4722                 "perf trace [<options>] [<command>]",
4723                 "perf trace [<options>] -- <command> [<options>]",
4724                 "perf trace record [<options>] [<command>]",
4725                 "perf trace record [<options>] -- <command> [<options>]",
4726                 NULL
4727         };
4728         struct trace trace = {
4729                 .opts = {
4730                         .target = {
4731                                 .uid       = UINT_MAX,
4732                                 .uses_mmap = true,
4733                         },
4734                         .user_freq     = UINT_MAX,
4735                         .user_interval = ULLONG_MAX,
4736                         .no_buffering  = true,
4737                         .mmap_pages    = UINT_MAX,
4738                 },
4739                 .output = stderr,
4740                 .show_comm = true,
4741                 .show_tstamp = true,
4742                 .show_duration = true,
4743                 .show_arg_names = true,
4744                 .args_alignment = 70,
4745                 .trace_syscalls = false,
4746                 .kernel_syscallchains = false,
4747                 .max_stack = UINT_MAX,
4748                 .max_events = ULONG_MAX,
4749         };
4750         const char *map_dump_str = NULL;
4751         const char *output_name = NULL;
4752         const struct option trace_options[] = {
4753         OPT_CALLBACK('e', "event", &trace, "event",
4754                      "event/syscall selector. use 'perf list' to list available events",
4755                      trace__parse_events_option),
4756         OPT_CALLBACK(0, "filter", &trace.evlist, "filter",
4757                      "event filter", parse_filter),
4758         OPT_BOOLEAN(0, "comm", &trace.show_comm,
4759                     "show the thread COMM next to its id"),
4760         OPT_BOOLEAN(0, "tool_stats", &trace.show_tool_stats, "show tool stats"),
4761         OPT_CALLBACK(0, "expr", &trace, "expr", "list of syscalls/events to trace",
4762                      trace__parse_events_option),
4763         OPT_STRING('o', "output", &output_name, "file", "output file name"),
4764         OPT_STRING('i', "input", &input_name, "file", "Analyze events in file"),
4765         OPT_STRING('p', "pid", &trace.opts.target.pid, "pid",
4766                     "trace events on existing process id"),
4767         OPT_STRING('t', "tid", &trace.opts.target.tid, "tid",
4768                     "trace events on existing thread id"),
4769         OPT_CALLBACK(0, "filter-pids", &trace, "CSV list of pids",
4770                      "pids to filter (by the kernel)", trace__set_filter_pids_from_option),
4771         OPT_BOOLEAN('a', "all-cpus", &trace.opts.target.system_wide,
4772                     "system-wide collection from all CPUs"),
4773         OPT_STRING('C', "cpu", &trace.opts.target.cpu_list, "cpu",
4774                     "list of cpus to monitor"),
4775         OPT_BOOLEAN(0, "no-inherit", &trace.opts.no_inherit,
4776                     "child tasks do not inherit counters"),
4777         OPT_CALLBACK('m', "mmap-pages", &trace.opts.mmap_pages, "pages",
4778                      "number of mmap data pages",
4779                      perf_evlist__parse_mmap_pages),
4780         OPT_STRING('u', "uid", &trace.opts.target.uid_str, "user",
4781                    "user to profile"),
4782         OPT_CALLBACK(0, "duration", &trace, "float",
4783                      "show only events with duration > N.M ms",
4784                      trace__set_duration),
4785 #ifdef HAVE_LIBBPF_SUPPORT
4786         OPT_STRING(0, "map-dump", &map_dump_str, "BPF map", "BPF map to periodically dump"),
4787 #endif
4788         OPT_BOOLEAN(0, "sched", &trace.sched, "show blocking scheduler events"),
4789         OPT_INCR('v', "verbose", &verbose, "be more verbose"),
4790         OPT_BOOLEAN('T', "time", &trace.full_time,
4791                     "Show full timestamp, not time relative to first start"),
4792         OPT_BOOLEAN(0, "failure", &trace.failure_only,
4793                     "Show only syscalls that failed"),
4794         OPT_BOOLEAN('s', "summary", &trace.summary_only,
4795                     "Show only syscall summary with statistics"),
4796         OPT_BOOLEAN('S', "with-summary", &trace.summary,
4797                     "Show all syscalls and summary with statistics"),
4798         OPT_BOOLEAN(0, "errno-summary", &trace.errno_summary,
4799                     "Show errno stats per syscall, use with -s or -S"),
4800         OPT_CALLBACK_DEFAULT('F', "pf", &trace.trace_pgfaults, "all|maj|min",
4801                      "Trace pagefaults", parse_pagefaults, "maj"),
4802         OPT_BOOLEAN(0, "syscalls", &trace.trace_syscalls, "Trace syscalls"),
4803         OPT_BOOLEAN('f', "force", &trace.force, "don't complain, do it"),
4804         OPT_CALLBACK(0, "call-graph", &trace.opts,
4805                      "record_mode[,record_size]", record_callchain_help,
4806                      &record_parse_callchain_opt),
4807         OPT_BOOLEAN(0, "libtraceevent_print", &trace.libtraceevent_print,
4808                     "Use libtraceevent to print the tracepoint arguments."),
4809         OPT_BOOLEAN(0, "kernel-syscall-graph", &trace.kernel_syscallchains,
4810                     "Show the kernel callchains on the syscall exit path"),
4811         OPT_ULONG(0, "max-events", &trace.max_events,
4812                 "Set the maximum number of events to print, exit after that is reached. "),
4813         OPT_UINTEGER(0, "min-stack", &trace.min_stack,
4814                      "Set the minimum stack depth when parsing the callchain, "
4815                      "anything below the specified depth will be ignored."),
4816         OPT_UINTEGER(0, "max-stack", &trace.max_stack,
4817                      "Set the maximum stack depth when parsing the callchain, "
4818                      "anything beyond the specified depth will be ignored. "
4819                      "Default: kernel.perf_event_max_stack or " __stringify(PERF_MAX_STACK_DEPTH)),
4820         OPT_BOOLEAN(0, "sort-events", &trace.sort_events,
4821                         "Sort batch of events before processing, use if getting out of order events"),
4822         OPT_BOOLEAN(0, "print-sample", &trace.print_sample,
4823                         "print the PERF_RECORD_SAMPLE PERF_SAMPLE_ info, for debugging"),
4824         OPT_UINTEGER(0, "proc-map-timeout", &proc_map_timeout,
4825                         "per thread proc mmap processing timeout in ms"),
4826         OPT_CALLBACK('G', "cgroup", &trace, "name", "monitor event in cgroup name only",
4827                      trace__parse_cgroups),
4828         OPT_INTEGER('D', "delay", &trace.opts.initial_delay,
4829                      "ms to wait before starting measurement after program "
4830                      "start"),
4831         OPTS_EVSWITCH(&trace.evswitch),
4832         OPT_END()
4833         };
4834         bool __maybe_unused max_stack_user_set = true;
4835         bool mmap_pages_user_set = true;
4836         struct evsel *evsel;
4837         const char * const trace_subcommands[] = { "record", NULL };
4838         int err = -1;
4839         char bf[BUFSIZ];
4840
4841         signal(SIGSEGV, sighandler_dump_stack);
4842         signal(SIGFPE, sighandler_dump_stack);
4843
4844         trace.evlist = evlist__new();
4845         trace.sctbl = syscalltbl__new();
4846
4847         if (trace.evlist == NULL || trace.sctbl == NULL) {
4848                 pr_err("Not enough memory to run!\n");
4849                 err = -ENOMEM;
4850                 goto out;
4851         }
4852
4853         /*
4854          * Parsing .perfconfig may entail creating a BPF event, that may need
4855          * to create BPF maps, so bump RLIM_MEMLOCK as the default 64K setting
4856          * is too small. This affects just this process, not touching the
4857          * global setting. If it fails we'll get something in 'perf trace -v'
4858          * to help diagnose the problem.
4859          */
4860         rlimit__bump_memlock();
4861
4862         err = perf_config(trace__config, &trace);
4863         if (err)
4864                 goto out;
4865
4866         argc = parse_options_subcommand(argc, argv, trace_options, trace_subcommands,
4867                                  trace_usage, PARSE_OPT_STOP_AT_NON_OPTION);
4868
4869         /*
4870          * Here we already passed thru trace__parse_events_option() and it has
4871          * already figured out if -e syscall_name, if not but if --event
4872          * foo:bar was used, the user is interested _just_ in those, say,
4873          * tracepoint events, not in the strace-like syscall-name-based mode.
4874          *
4875          * This is important because we need to check if strace-like mode is
4876          * needed to decided if we should filter out the eBPF
4877          * __augmented_syscalls__ code, if it is in the mix, say, via
4878          * .perfconfig trace.add_events, and filter those out.
4879          */
4880         if (!trace.trace_syscalls && !trace.trace_pgfaults &&
4881             trace.evlist->core.nr_entries == 0 /* Was --events used? */) {
4882                 trace.trace_syscalls = true;
4883         }
4884         /*
4885          * Now that we have --verbose figured out, lets see if we need to parse
4886          * events from .perfconfig, so that if those events fail parsing, say some
4887          * BPF program fails, then we'll be able to use --verbose to see what went
4888          * wrong in more detail.
4889          */
4890         if (trace.perfconfig_events != NULL) {
4891                 struct parse_events_error parse_err;
4892
4893                 bzero(&parse_err, sizeof(parse_err));
4894                 err = parse_events(trace.evlist, trace.perfconfig_events, &parse_err);
4895                 if (err) {
4896                         parse_events_print_error(&parse_err, trace.perfconfig_events);
4897                         goto out;
4898                 }
4899         }
4900
4901         if ((nr_cgroups || trace.cgroup) && !trace.opts.target.system_wide) {
4902                 usage_with_options_msg(trace_usage, trace_options,
4903                                        "cgroup monitoring only available in system-wide mode");
4904         }
4905
4906         evsel = bpf__setup_output_event(trace.evlist, "__augmented_syscalls__");
4907         if (IS_ERR(evsel)) {
4908                 bpf__strerror_setup_output_event(trace.evlist, PTR_ERR(evsel), bf, sizeof(bf));
4909                 pr_err("ERROR: Setup trace syscalls enter failed: %s\n", bf);
4910                 goto out;
4911         }
4912
4913         if (evsel) {
4914                 trace.syscalls.events.augmented = evsel;
4915
4916                 evsel = perf_evlist__find_tracepoint_by_name(trace.evlist, "raw_syscalls:sys_enter");
4917                 if (evsel == NULL) {
4918                         pr_err("ERROR: raw_syscalls:sys_enter not found in the augmented BPF object\n");
4919                         goto out;
4920                 }
4921
4922                 if (evsel->bpf_obj == NULL) {
4923                         pr_err("ERROR: raw_syscalls:sys_enter not associated to a BPF object\n");
4924                         goto out;
4925                 }
4926
4927                 trace.bpf_obj = evsel->bpf_obj;
4928
4929                 /*
4930                  * If we have _just_ the augmenter event but don't have a
4931                  * explicit --syscalls, then assume we want all strace-like
4932                  * syscalls:
4933                  */
4934                 if (!trace.trace_syscalls && trace__only_augmented_syscalls_evsels(&trace))
4935                         trace.trace_syscalls = true;
4936                 /*
4937                  * So, if we have a syscall augmenter, but trace_syscalls, aka
4938                  * strace-like syscall tracing is not set, then we need to trow
4939                  * away the augmenter, i.e. all the events that were created
4940                  * from that BPF object file.
4941                  *
4942                  * This is more to fix the current .perfconfig trace.add_events
4943                  * style of setting up the strace-like eBPF based syscall point
4944                  * payload augmenter.
4945                  *
4946                  * All this complexity will be avoided by adding an alternative
4947                  * to trace.add_events in the form of
4948                  * trace.bpf_augmented_syscalls, that will be only parsed if we
4949                  * need it.
4950                  *
4951                  * .perfconfig trace.add_events is still useful if we want, for
4952                  * instance, have msr_write.msr in some .perfconfig profile based
4953                  * 'perf trace --config determinism.profile' mode, where for some
4954                  * particular goal/workload type we want a set of events and
4955                  * output mode (with timings, etc) instead of having to add
4956                  * all via the command line.
4957                  *
4958                  * Also --config to specify an alternate .perfconfig file needs
4959                  * to be implemented.
4960                  */
4961                 if (!trace.trace_syscalls) {
4962                         trace__delete_augmented_syscalls(&trace);
4963                 } else {
4964                         trace__set_bpf_map_filtered_pids(&trace);
4965                         trace__set_bpf_map_syscalls(&trace);
4966                         trace.syscalls.unaugmented_prog = trace__find_bpf_program_by_title(&trace, "!raw_syscalls:unaugmented");
4967                 }
4968         }
4969
4970         err = bpf__setup_stdout(trace.evlist);
4971         if (err) {
4972                 bpf__strerror_setup_stdout(trace.evlist, err, bf, sizeof(bf));
4973                 pr_err("ERROR: Setup BPF stdout failed: %s\n", bf);
4974                 goto out;
4975         }
4976
4977         err = -1;
4978
4979         if (map_dump_str) {
4980                 trace.dump.map = trace__find_bpf_map_by_name(&trace, map_dump_str);
4981                 if (trace.dump.map == NULL) {
4982                         pr_err("ERROR: BPF map \"%s\" not found\n", map_dump_str);
4983                         goto out;
4984                 }
4985         }
4986
4987         if (trace.trace_pgfaults) {
4988                 trace.opts.sample_address = true;
4989                 trace.opts.sample_time = true;
4990         }
4991
4992         if (trace.opts.mmap_pages == UINT_MAX)
4993                 mmap_pages_user_set = false;
4994
4995         if (trace.max_stack == UINT_MAX) {
4996                 trace.max_stack = input_name ? PERF_MAX_STACK_DEPTH : sysctl__max_stack();
4997                 max_stack_user_set = false;
4998         }
4999
5000 #ifdef HAVE_DWARF_UNWIND_SUPPORT
5001         if ((trace.min_stack || max_stack_user_set) && !callchain_param.enabled) {
5002                 record_opts__parse_callchain(&trace.opts, &callchain_param, "dwarf", false);
5003         }
5004 #endif
5005
5006         if (callchain_param.enabled) {
5007                 if (!mmap_pages_user_set && geteuid() == 0)
5008                         trace.opts.mmap_pages = perf_event_mlock_kb_in_pages() * 4;
5009
5010                 symbol_conf.use_callchain = true;
5011         }
5012
5013         if (trace.evlist->core.nr_entries > 0) {
5014                 evlist__set_default_evsel_handler(trace.evlist, trace__event_handler);
5015                 if (evlist__set_syscall_tp_fields(trace.evlist)) {
5016                         perror("failed to set syscalls:* tracepoint fields");
5017                         goto out;
5018                 }
5019         }
5020
5021         if (trace.sort_events) {
5022                 ordered_events__init(&trace.oe.data, ordered_events__deliver_event, &trace);
5023                 ordered_events__set_copy_on_queue(&trace.oe.data, true);
5024         }
5025
5026         /*
5027          * If we are augmenting syscalls, then combine what we put in the
5028          * __augmented_syscalls__ BPF map with what is in the
5029          * syscalls:sys_exit_FOO tracepoints, i.e. just like we do without BPF,
5030          * combining raw_syscalls:sys_enter with raw_syscalls:sys_exit.
5031          *
5032          * We'll switch to look at two BPF maps, one for sys_enter and the
5033          * other for sys_exit when we start augmenting the sys_exit paths with
5034          * buffers that are being copied from kernel to userspace, think 'read'
5035          * syscall.
5036          */
5037         if (trace.syscalls.events.augmented) {
5038                 evlist__for_each_entry(trace.evlist, evsel) {
5039                         bool raw_syscalls_sys_exit = strcmp(evsel__name(evsel), "raw_syscalls:sys_exit") == 0;
5040
5041                         if (raw_syscalls_sys_exit) {
5042                                 trace.raw_augmented_syscalls = true;
5043                                 goto init_augmented_syscall_tp;
5044                         }
5045
5046                         if (trace.syscalls.events.augmented->priv == NULL &&
5047                             strstr(evsel__name(evsel), "syscalls:sys_enter")) {
5048                                 struct evsel *augmented = trace.syscalls.events.augmented;
5049                                 if (evsel__init_augmented_syscall_tp(augmented, evsel) ||
5050                                     evsel__init_augmented_syscall_tp_args(augmented))
5051                                         goto out;
5052                                 /*
5053                                  * Augmented is __augmented_syscalls__ BPF_OUTPUT event
5054                                  * Above we made sure we can get from the payload the tp fields
5055                                  * that we get from syscalls:sys_enter tracefs format file.
5056                                  */
5057                                 augmented->handler = trace__sys_enter;
5058                                 /*
5059                                  * Now we do the same for the *syscalls:sys_enter event so that
5060                                  * if we handle it directly, i.e. if the BPF prog returns 0 so
5061                                  * as not to filter it, then we'll handle it just like we would
5062                                  * for the BPF_OUTPUT one:
5063                                  */
5064                                 if (evsel__init_augmented_syscall_tp(evsel, evsel) ||
5065                                     evsel__init_augmented_syscall_tp_args(evsel))
5066                                         goto out;
5067                                 evsel->handler = trace__sys_enter;
5068                         }
5069
5070                         if (strstarts(evsel__name(evsel), "syscalls:sys_exit_")) {
5071                                 struct syscall_tp *sc;
5072 init_augmented_syscall_tp:
5073                                 if (evsel__init_augmented_syscall_tp(evsel, evsel))
5074                                         goto out;
5075                                 sc = __evsel__syscall_tp(evsel);
5076                                 /*
5077                                  * For now with BPF raw_augmented we hook into
5078                                  * raw_syscalls:sys_enter and there we get all
5079                                  * 6 syscall args plus the tracepoint common
5080                                  * fields and the syscall_nr (another long).
5081                                  * So we check if that is the case and if so
5082                                  * don't look after the sc->args_size but
5083                                  * always after the full raw_syscalls:sys_enter
5084                                  * payload, which is fixed.
5085                                  *
5086                                  * We'll revisit this later to pass
5087                                  * s->args_size to the BPF augmenter (now
5088                                  * tools/perf/examples/bpf/augmented_raw_syscalls.c,
5089                                  * so that it copies only what we need for each
5090                                  * syscall, like what happens when we use
5091                                  * syscalls:sys_enter_NAME, so that we reduce
5092                                  * the kernel/userspace traffic to just what is
5093                                  * needed for each syscall.
5094                                  */
5095                                 if (trace.raw_augmented_syscalls)
5096                                         trace.raw_augmented_syscalls_args_size = (6 + 1) * sizeof(long) + sc->id.offset;
5097                                 evsel__init_augmented_syscall_tp_ret(evsel);
5098                                 evsel->handler = trace__sys_exit;
5099                         }
5100                 }
5101         }
5102
5103         if ((argc >= 1) && (strcmp(argv[0], "record") == 0))
5104                 return trace__record(&trace, argc-1, &argv[1]);
5105
5106         /* Using just --errno-summary will trigger --summary */
5107         if (trace.errno_summary && !trace.summary && !trace.summary_only)
5108                 trace.summary_only = true;
5109
5110         /* summary_only implies summary option, but don't overwrite summary if set */
5111         if (trace.summary_only)
5112                 trace.summary = trace.summary_only;
5113
5114         if (output_name != NULL) {
5115                 err = trace__open_output(&trace, output_name);
5116                 if (err < 0) {
5117                         perror("failed to create output file");
5118                         goto out;
5119                 }
5120         }
5121
5122         err = evswitch__init(&trace.evswitch, trace.evlist, stderr);
5123         if (err)
5124                 goto out_close;
5125
5126         err = target__validate(&trace.opts.target);
5127         if (err) {
5128                 target__strerror(&trace.opts.target, err, bf, sizeof(bf));
5129                 fprintf(trace.output, "%s", bf);
5130                 goto out_close;
5131         }
5132
5133         err = target__parse_uid(&trace.opts.target);
5134         if (err) {
5135                 target__strerror(&trace.opts.target, err, bf, sizeof(bf));
5136                 fprintf(trace.output, "%s", bf);
5137                 goto out_close;
5138         }
5139
5140         if (!argc && target__none(&trace.opts.target))
5141                 trace.opts.target.system_wide = true;
5142
5143         if (input_name)
5144                 err = trace__replay(&trace);
5145         else
5146                 err = trace__run(&trace, argc, argv);
5147
5148 out_close:
5149         if (output_name != NULL)
5150                 fclose(trace.output);
5151 out:
5152         zfree(&trace.perfconfig_events);
5153         return err;
5154 }