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 <api/fs/tracing_path.h>
  19 #ifdef HAVE_LIBBPF_SUPPORT
  20 #include <bpf/bpf.h>
  21 #include <bpf/libbpf.h>
  22 #ifdef HAVE_BPF_SKEL
  23 #include "bpf_skel/augmented_raw_syscalls.skel.h"
  24 #endif
  25 #endif
  26 #include "util/bpf_map.h"
  27 #include "util/rlimit.h"
  28 #include "builtin.h"
  29 #include "util/cgroup.h"
  30 #include "util/color.h"
  31 #include "util/config.h"
  32 #include "util/debug.h"
  33 #include "util/dso.h"
  34 #include "util/env.h"
  35 #include "util/event.h"
  36 #include "util/evsel.h"
  37 #include "util/evsel_fprintf.h"
  38 #include "util/synthetic-events.h"
  39 #include "util/evlist.h"
  40 #include "util/evswitch.h"
  41 #include "util/mmap.h"
  42 #include <subcmd/pager.h>
  43 #include <subcmd/exec-cmd.h>
  44 #include "util/machine.h"
  45 #include "util/map.h"
  46 #include "util/symbol.h"
  47 #include "util/path.h"
  48 #include "util/session.h"
  49 #include "util/thread.h"
  50 #include <subcmd/parse-options.h>
  51 #include "util/strlist.h"
  52 #include "util/intlist.h"
  53 #include "util/thread_map.h"
  54 #include "util/stat.h"
  55 #include "util/tool.h"
  56 #include "util/util.h"
  57 #include "trace/beauty/beauty.h"
  58 #include "trace-event.h"
  59 #include "util/parse-events.h"
  60 #include "util/tracepoint.h"
  61 #include "callchain.h"
  62 #include "print_binary.h"
  63 #include "string2.h"
  64 #include "syscalltbl.h"
  65 #include "rb_resort.h"
  66 #include "../perf.h"
  67
  68 #include <errno.h>
  69 #include <inttypes.h>
  70 #include <poll.h>
  71 #include <signal.h>
  72 #include <stdlib.h>
  73 #include <string.h>
  74 #include <linux/err.h>
  75 #include <linux/filter.h>
  76 #include <linux/kernel.h>
  77 #include <linux/random.h>
  78 #include <linux/stringify.h>
  79 #include <linux/time64.h>
  80 #include <linux/zalloc.h>
  81 #include <fcntl.h>
  82 #include <sys/sysmacros.h>
  83
  84 #include <linux/ctype.h>
  85 #include <perf/mmap.h>
  86
  87 #ifdef HAVE_LIBTRACEEVENT
  88 #include <traceevent/event-parse.h>
  89 #endif
  90
  91 #ifndef O_CLOEXEC
  92 # define O_CLOEXEC              02000000
  93 #endif
  94
  95 #ifndef F_LINUX_SPECIFIC_BASE
  96 # define F_LINUX_SPECIFIC_BASE  1024
  97 #endif
  98
  99 #define RAW_SYSCALL_ARGS_NUM    6
 100
 101 /*
 102  * strtoul: Go from a string to a value, i.e. for msr: MSR_FS_BASE to 0xc0000100
 103  */
 104 struct syscall_arg_fmt {
 105         size_t     (*scnprintf)(char *bf, size_t size, struct syscall_arg *arg);
 106         bool       (*strtoul)(char *bf, size_t size, struct syscall_arg *arg, u64 *val);
 107         unsigned long (*mask_val)(struct syscall_arg *arg, unsigned long val);
 108         void       *parm;
 109         const char *name;
 110         u16        nr_entries; // for arrays
 111         bool       show_zero;
 112 };
 113
 114 struct syscall_fmt {
 115         const char *name;
 116         const char *alias;
 117         struct {
 118                 const char *sys_enter,
 119                            *sys_exit;
 120         }          bpf_prog_name;
 121         struct syscall_arg_fmt arg[RAW_SYSCALL_ARGS_NUM];
 122         u8         nr_args;
 123         bool       errpid;
 124         bool       timeout;
 125         bool       hexret;
 126 };
 127
 128 struct trace {
 129         struct perf_tool        tool;
 130         struct syscalltbl       *sctbl;
 131         struct {
 132                 struct syscall  *table;
 133                 struct {
 134                         struct evsel *sys_enter,
 135                                 *sys_exit,
 136                                 *bpf_output;
 137                 }               events;
 138         } syscalls;
 139 #ifdef HAVE_BPF_SKEL
 140         struct augmented_raw_syscalls_bpf *skel;
 141 #endif
 142         struct record_opts      opts;
 143         struct evlist   *evlist;
 144         struct machine          *host;
 145         struct thread           *current;
 146         struct cgroup           *cgroup;
 147         u64                     base_time;
 148         FILE                    *output;
 149         unsigned long           nr_events;
 150         unsigned long           nr_events_printed;
 151         unsigned long           max_events;
 152         struct evswitch         evswitch;
 153         struct strlist          *ev_qualifier;
 154         struct {
 155                 size_t          nr;
 156                 int             *entries;
 157         }                       ev_qualifier_ids;
 158         struct {
 159                 size_t          nr;
 160                 pid_t           *entries;
 161                 struct bpf_map  *map;
 162         }                       filter_pids;
 163         double                  duration_filter;
 164         double                  runtime_ms;
 165         struct {
 166                 u64             vfs_getname,
 167                                 proc_getname;
 168         } stats;
 169         unsigned int            max_stack;
 170         unsigned int            min_stack;
 171         int                     raw_augmented_syscalls_args_size;
 172         bool                    raw_augmented_syscalls;
 173         bool                    fd_path_disabled;
 174         bool                    sort_events;
 175         bool                    not_ev_qualifier;
 176         bool                    live;
 177         bool                    full_time;
 178         bool                    sched;
 179         bool                    multiple_threads;
 180         bool                    summary;
 181         bool                    summary_only;
 182         bool                    errno_summary;
 183         bool                    failure_only;
 184         bool                    show_comm;
 185         bool                    print_sample;
 186         bool                    show_tool_stats;
 187         bool                    trace_syscalls;
 188         bool                    libtraceevent_print;
 189         bool                    kernel_syscallchains;
 190         s16                     args_alignment;
 191         bool                    show_tstamp;
 192         bool                    show_duration;
 193         bool                    show_zeros;
 194         bool                    show_arg_names;
 195         bool                    show_string_prefix;
 196         bool                    force;
 197         bool                    vfs_getname;
 198         int                     trace_pgfaults;
 199         char                    *perfconfig_events;
 200         struct {
 201                 struct ordered_events   data;
 202                 u64                     last;
 203         } oe;
 204 };
 205
 206 struct tp_field {
 207         int offset;
 208         union {
 209                 u64 (*integer)(struct tp_field *field, struct perf_sample *sample);
 210                 void *(*pointer)(struct tp_field *field, struct perf_sample *sample);
 211         };
 212 };
 213
 214 #define TP_UINT_FIELD(bits) \
 215 static u64 tp_field__u##bits(struct tp_field *field, struct perf_sample *sample) \
 216 { \
 217         u##bits value; \
 218         memcpy(&value, sample->raw_data + field->offset, sizeof(value)); \
 219         return value;  \
 220 }
 221
 222 TP_UINT_FIELD(8);
 223 TP_UINT_FIELD(16);
 224 TP_UINT_FIELD(32);
 225 TP_UINT_FIELD(64);
 226
 227 #define TP_UINT_FIELD__SWAPPED(bits) \
 228 static u64 tp_field__swapped_u##bits(struct tp_field *field, struct perf_sample *sample) \
 229 { \
 230         u##bits value; \
 231         memcpy(&value, sample->raw_data + field->offset, sizeof(value)); \
 232         return bswap_##bits(value);\
 233 }
 234
 235 TP_UINT_FIELD__SWAPPED(16);
 236 TP_UINT_FIELD__SWAPPED(32);
 237 TP_UINT_FIELD__SWAPPED(64);
 238
 239 static int __tp_field__init_uint(struct tp_field *field, int size, int offset, bool needs_swap)
 240 {
 241         field->offset = offset;
 242
 243         switch (size) {
 244         case 1:
 245                 field->integer = tp_field__u8;
 246                 break;
 247         case 2:
 248                 field->integer = needs_swap ? tp_field__swapped_u16 : tp_field__u16;
 249                 break;
 250         case 4:
 251                 field->integer = needs_swap ? tp_field__swapped_u32 : tp_field__u32;
 252                 break;
 253         case 8:
 254                 field->integer = needs_swap ? tp_field__swapped_u64 : tp_field__u64;
 255                 break;
 256         default:
 257                 return -1;
 258         }
 259
 260         return 0;
 261 }
 262
 263 static int tp_field__init_uint(struct tp_field *field, struct tep_format_field *format_field, bool needs_swap)
 264 {
 265         return __tp_field__init_uint(field, format_field->size, format_field->offset, needs_swap);
 266 }
 267
 268 static void *tp_field__ptr(struct tp_field *field, struct perf_sample *sample)
 269 {
 270         return sample->raw_data + field->offset;
 271 }
 272
 273 static int __tp_field__init_ptr(struct tp_field *field, int offset)
 274 {
 275         field->offset = offset;
 276         field->pointer = tp_field__ptr;
 277         return 0;
 278 }
 279
 280 static int tp_field__init_ptr(struct tp_field *field, struct tep_format_field *format_field)
 281 {
 282         return __tp_field__init_ptr(field, format_field->offset);
 283 }
 284
 285 struct syscall_tp {
 286         struct tp_field id;
 287         union {
 288                 struct tp_field args, ret;
 289         };
 290 };
 291
 292 /*
 293  * The evsel->priv as used by 'perf trace'
 294  * sc:  for raw_syscalls:sys_{enter,exit} and syscalls:sys_{enter,exit}_SYSCALLNAME
 295  * fmt: for all the other tracepoints
 296  */
 297 struct evsel_trace {
 298         struct syscall_tp       sc;
 299         struct syscall_arg_fmt  *fmt;
 300 };
 301
 302 static struct evsel_trace *evsel_trace__new(void)
 303 {
 304         return zalloc(sizeof(struct evsel_trace));
 305 }
 306
 307 static void evsel_trace__delete(struct evsel_trace *et)
 308 {
 309         if (et == NULL)
 310                 return;
 311
 312         zfree(&et->fmt);
 313         free(et);
 314 }
 315
 316 /*
 317  * Used with raw_syscalls:sys_{enter,exit} and with the
 318  * syscalls:sys_{enter,exit}_SYSCALL tracepoints
 319  */
 320 static inline struct syscall_tp *__evsel__syscall_tp(struct evsel *evsel)
 321 {
 322         struct evsel_trace *et = evsel->priv;
 323
 324         return &et->sc;
 325 }
 326
 327 static struct syscall_tp *evsel__syscall_tp(struct evsel *evsel)
 328 {
 329         if (evsel->priv == NULL) {
 330                 evsel->priv = evsel_trace__new();
 331                 if (evsel->priv == NULL)
 332                         return NULL;
 333         }
 334
 335         return __evsel__syscall_tp(evsel);
 336 }
 337
 338 /*
 339  * Used with all the other tracepoints.
 340  */
 341 static inline struct syscall_arg_fmt *__evsel__syscall_arg_fmt(struct evsel *evsel)
 342 {
 343         struct evsel_trace *et = evsel->priv;
 344
 345         return et->fmt;
 346 }
 347
 348 static struct syscall_arg_fmt *evsel__syscall_arg_fmt(struct evsel *evsel)
 349 {
 350         struct evsel_trace *et = evsel->priv;
 351
 352         if (evsel->priv == NULL) {
 353                 et = evsel->priv = evsel_trace__new();
 354
 355                 if (et == NULL)
 356                         return NULL;
 357         }
 358
 359         if (et->fmt == NULL) {
 360                 et->fmt = calloc(evsel->tp_format->format.nr_fields, sizeof(struct syscall_arg_fmt));
 361                 if (et->fmt == NULL)
 362                         goto out_delete;
 363         }
 364
 365         return __evsel__syscall_arg_fmt(evsel);
 366
 367 out_delete:
 368         evsel_trace__delete(evsel->priv);
 369         evsel->priv = NULL;
 370         return NULL;
 371 }
 372
 373 static int evsel__init_tp_uint_field(struct evsel *evsel, struct tp_field *field, const char *name)
 374 {
 375         struct tep_format_field *format_field = evsel__field(evsel, name);
 376
 377         if (format_field == NULL)
 378                 return -1;
 379
 380         return tp_field__init_uint(field, format_field, evsel->needs_swap);
 381 }
 382
 383 #define perf_evsel__init_sc_tp_uint_field(evsel, name) \
 384         ({ struct syscall_tp *sc = __evsel__syscall_tp(evsel);\
 385            evsel__init_tp_uint_field(evsel, &sc->name, #name); })
 386
 387 static int evsel__init_tp_ptr_field(struct evsel *evsel, struct tp_field *field, const char *name)
 388 {
 389         struct tep_format_field *format_field = evsel__field(evsel, name);
 390
 391         if (format_field == NULL)
 392                 return -1;
 393
 394         return tp_field__init_ptr(field, format_field);
 395 }
 396
 397 #define perf_evsel__init_sc_tp_ptr_field(evsel, name) \
 398         ({ struct syscall_tp *sc = __evsel__syscall_tp(evsel);\
 399            evsel__init_tp_ptr_field(evsel, &sc->name, #name); })
 400
 401 static void evsel__delete_priv(struct evsel *evsel)
 402 {
 403         zfree(&evsel->priv);
 404         evsel__delete(evsel);
 405 }
 406
 407 static int evsel__init_syscall_tp(struct evsel *evsel)
 408 {
 409         struct syscall_tp *sc = evsel__syscall_tp(evsel);
 410
 411         if (sc != NULL) {
 412                 if (evsel__init_tp_uint_field(evsel, &sc->id, "__syscall_nr") &&
 413                     evsel__init_tp_uint_field(evsel, &sc->id, "nr"))
 414                         return -ENOENT;
 415
 416                 return 0;
 417         }
 418
 419         return -ENOMEM;
 420 }
 421
 422 static int evsel__init_augmented_syscall_tp(struct evsel *evsel, struct evsel *tp)
 423 {
 424         struct syscall_tp *sc = evsel__syscall_tp(evsel);
 425
 426         if (sc != NULL) {
 427                 struct tep_format_field *syscall_id = evsel__field(tp, "id");
 428                 if (syscall_id == NULL)
 429                         syscall_id = evsel__field(tp, "__syscall_nr");
 430                 if (syscall_id == NULL ||
 431                     __tp_field__init_uint(&sc->id, syscall_id->size, syscall_id->offset, evsel->needs_swap))
 432                         return -EINVAL;
 433
 434                 return 0;
 435         }
 436
 437         return -ENOMEM;
 438 }
 439
 440 static int evsel__init_augmented_syscall_tp_args(struct evsel *evsel)
 441 {
 442         struct syscall_tp *sc = __evsel__syscall_tp(evsel);
 443
 444         return __tp_field__init_ptr(&sc->args, sc->id.offset + sizeof(u64));
 445 }
 446
 447 static int evsel__init_augmented_syscall_tp_ret(struct evsel *evsel)
 448 {
 449         struct syscall_tp *sc = __evsel__syscall_tp(evsel);
 450
 451         return __tp_field__init_uint(&sc->ret, sizeof(u64), sc->id.offset + sizeof(u64), evsel->needs_swap);
 452 }
 453
 454 static int evsel__init_raw_syscall_tp(struct evsel *evsel, void *handler)
 455 {
 456         if (evsel__syscall_tp(evsel) != NULL) {
 457                 if (perf_evsel__init_sc_tp_uint_field(evsel, id))
 458                         return -ENOENT;
 459
 460                 evsel->handler = handler;
 461                 return 0;
 462         }
 463
 464         return -ENOMEM;
 465 }
 466
 467 static struct evsel *perf_evsel__raw_syscall_newtp(const char *direction, void *handler)
 468 {
 469         struct evsel *evsel = evsel__newtp("raw_syscalls", direction);
 470
 471         /* older kernel (e.g., RHEL6) use syscalls:{enter,exit} */
 472         if (IS_ERR(evsel))
 473                 evsel = evsel__newtp("syscalls", direction);
 474
 475         if (IS_ERR(evsel))
 476                 return NULL;
 477
 478         if (evsel__init_raw_syscall_tp(evsel, handler))
 479                 goto out_delete;
 480
 481         return evsel;
 482
 483 out_delete:
 484         evsel__delete_priv(evsel);
 485         return NULL;
 486 }
 487
 488 #define perf_evsel__sc_tp_uint(evsel, name, sample) \
 489         ({ struct syscall_tp *fields = __evsel__syscall_tp(evsel); \
 490            fields->name.integer(&fields->name, sample); })
 491
 492 #define perf_evsel__sc_tp_ptr(evsel, name, sample) \
 493         ({ struct syscall_tp *fields = __evsel__syscall_tp(evsel); \
 494            fields->name.pointer(&fields->name, sample); })
 495
 496 size_t strarray__scnprintf_suffix(struct strarray *sa, char *bf, size_t size, const char *intfmt, bool show_suffix, int val)
 497 {
 498         int idx = val - sa->offset;
 499
 500         if (idx < 0 || idx >= sa->nr_entries || sa->entries[idx] == NULL) {
 501                 size_t printed = scnprintf(bf, size, intfmt, val);
 502                 if (show_suffix)
 503                         printed += scnprintf(bf + printed, size - printed, " /* %s??? */", sa->prefix);
 504                 return printed;
 505         }
 506
 507         return scnprintf(bf, size, "%s%s", sa->entries[idx], show_suffix ? sa->prefix : "");
 508 }
 509
 510 size_t strarray__scnprintf(struct strarray *sa, char *bf, size_t size, const char *intfmt, bool show_prefix, int val)
 511 {
 512         int idx = val - sa->offset;
 513
 514         if (idx < 0 || idx >= sa->nr_entries || sa->entries[idx] == NULL) {
 515                 size_t printed = scnprintf(bf, size, intfmt, val);
 516                 if (show_prefix)
 517                         printed += scnprintf(bf + printed, size - printed, " /* %s??? */", sa->prefix);
 518                 return printed;
 519         }
 520
 521         return scnprintf(bf, size, "%s%s", show_prefix ? sa->prefix : "", sa->entries[idx]);
 522 }
 523
 524 static size_t __syscall_arg__scnprintf_strarray(char *bf, size_t size,
 525                                                 const char *intfmt,
 526                                                 struct syscall_arg *arg)
 527 {
 528         return strarray__scnprintf(arg->parm, bf, size, intfmt, arg->show_string_prefix, arg->val);
 529 }
 530
 531 static size_t syscall_arg__scnprintf_strarray(char *bf, size_t size,
 532                                               struct syscall_arg *arg)
 533 {
 534         return __syscall_arg__scnprintf_strarray(bf, size, "%d", arg);
 535 }
 536
 537 #define SCA_STRARRAY syscall_arg__scnprintf_strarray
 538
 539 bool syscall_arg__strtoul_strarray(char *bf, size_t size, struct syscall_arg *arg, u64 *ret)
 540 {
 541         return strarray__strtoul(arg->parm, bf, size, ret);
 542 }
 543
 544 bool syscall_arg__strtoul_strarray_flags(char *bf, size_t size, struct syscall_arg *arg, u64 *ret)
 545 {
 546         return strarray__strtoul_flags(arg->parm, bf, size, ret);
 547 }
 548
 549 bool syscall_arg__strtoul_strarrays(char *bf, size_t size, struct syscall_arg *arg, u64 *ret)
 550 {
 551         return strarrays__strtoul(arg->parm, bf, size, ret);
 552 }
 553
 554 size_t syscall_arg__scnprintf_strarray_flags(char *bf, size_t size, struct syscall_arg *arg)
 555 {
 556         return strarray__scnprintf_flags(arg->parm, bf, size, arg->show_string_prefix, arg->val);
 557 }
 558
 559 size_t strarrays__scnprintf(struct strarrays *sas, char *bf, size_t size, const char *intfmt, bool show_prefix, int val)
 560 {
 561         size_t printed;
 562         int i;
 563
 564         for (i = 0; i < sas->nr_entries; ++i) {
 565                 struct strarray *sa = sas->entries[i];
 566                 int idx = val - sa->offset;
 567
 568                 if (idx >= 0 && idx < sa->nr_entries) {
 569                         if (sa->entries[idx] == NULL)
 570                                 break;
 571                         return scnprintf(bf, size, "%s%s", show_prefix ? sa->prefix : "", sa->entries[idx]);
 572                 }
 573         }
 574
 575         printed = scnprintf(bf, size, intfmt, val);
 576         if (show_prefix)
 577                 printed += scnprintf(bf + printed, size - printed, " /* %s??? */", sas->entries[0]->prefix);
 578         return printed;
 579 }
 580
 581 bool strarray__strtoul(struct strarray *sa, char *bf, size_t size, u64 *ret)
 582 {
 583         int i;
 584
 585         for (i = 0; i < sa->nr_entries; ++i) {
 586                 if (sa->entries[i] && strncmp(sa->entries[i], bf, size) == 0 && sa->entries[i][size] == '\0') {
 587                         *ret = sa->offset + i;
 588                         return true;
 589                 }
 590         }
 591
 592         return false;
 593 }
 594
 595 bool strarray__strtoul_flags(struct strarray *sa, char *bf, size_t size, u64 *ret)
 596 {
 597         u64 val = 0;
 598         char *tok = bf, *sep, *end;
 599
 600         *ret = 0;
 601
 602         while (size != 0) {
 603                 int toklen = size;
 604
 605                 sep = memchr(tok, '|', size);
 606                 if (sep != NULL) {
 607                         size -= sep - tok + 1;
 608
 609                         end = sep - 1;
 610                         while (end > tok && isspace(*end))
 611                                 --end;
 612
 613                         toklen = end - tok + 1;
 614                 }
 615
 616                 while (isspace(*tok))
 617                         ++tok;
 618
 619                 if (isalpha(*tok) || *tok == '_') {
 620                         if (!strarray__strtoul(sa, tok, toklen, &val))
 621                                 return false;
 622                 } else
 623                         val = strtoul(tok, NULL, 0);
 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", "OBJ_PIN", "OBJ_GET", "PROG_ATTACH",
 713         "PROG_DETACH", "PROG_TEST_RUN", "PROG_GET_NEXT_ID", "MAP_GET_NEXT_ID",
 714         "PROG_GET_FD_BY_ID", "MAP_GET_FD_BY_ID", "OBJ_GET_INFO_BY_FD",
 715         "PROG_QUERY", "RAW_TRACEPOINT_OPEN", "BTF_LOAD", "BTF_GET_FD_BY_ID",
 716         "TASK_FD_QUERY", "MAP_LOOKUP_AND_DELETE_ELEM", "MAP_FREEZE",
 717         "BTF_GET_NEXT_ID", "MAP_LOOKUP_BATCH", "MAP_LOOKUP_AND_DELETE_BATCH",
 718         "MAP_UPDATE_BATCH", "MAP_DELETE_BATCH", "LINK_CREATE", "LINK_UPDATE",
 719         "LINK_GET_FD_BY_ID", "LINK_GET_NEXT_ID", "ENABLE_STATS", "ITER_CREATE",
 720         "LINK_DETACH", "PROG_BIND_MAP",
 721 };
 722 static DEFINE_STRARRAY(bpf_cmd, "BPF_");
 723
 724 static const char *fsmount_flags[] = {
 725         [1] = "CLOEXEC",
 726 };
 727 static DEFINE_STRARRAY(fsmount_flags, "FSMOUNT_");
 728
 729 #include "trace/beauty/generated/fsconfig_arrays.c"
 730
 731 static DEFINE_STRARRAY(fsconfig_cmds, "FSCONFIG_");
 732
 733 static const char *epoll_ctl_ops[] = { "ADD", "DEL", "MOD", };
 734 static DEFINE_STRARRAY_OFFSET(epoll_ctl_ops, "EPOLL_CTL_", 1);
 735
 736 static const char *itimers[] = { "REAL", "VIRTUAL", "PROF", };
 737 static DEFINE_STRARRAY(itimers, "ITIMER_");
 738
 739 static const char *keyctl_options[] = {
 740         "GET_KEYRING_ID", "JOIN_SESSION_KEYRING", "UPDATE", "REVOKE", "CHOWN",
 741         "SETPERM", "DESCRIBE", "CLEAR", "LINK", "UNLINK", "SEARCH", "READ",
 742         "INSTANTIATE", "NEGATE", "SET_REQKEY_KEYRING", "SET_TIMEOUT",
 743         "ASSUME_AUTHORITY", "GET_SECURITY", "SESSION_TO_PARENT", "REJECT",
 744         "INSTANTIATE_IOV", "INVALIDATE", "GET_PERSISTENT",
 745 };
 746 static DEFINE_STRARRAY(keyctl_options, "KEYCTL_");
 747
 748 static const char *whences[] = { "SET", "CUR", "END",
 749 #ifdef SEEK_DATA
 750 "DATA",
 751 #endif
 752 #ifdef SEEK_HOLE
 753 "HOLE",
 754 #endif
 755 };
 756 static DEFINE_STRARRAY(whences, "SEEK_");
 757
 758 static const char *fcntl_cmds[] = {
 759         "DUPFD", "GETFD", "SETFD", "GETFL", "SETFL", "GETLK", "SETLK",
 760         "SETLKW", "SETOWN", "GETOWN", "SETSIG", "GETSIG", "GETLK64",
 761         "SETLK64", "SETLKW64", "SETOWN_EX", "GETOWN_EX",
 762         "GETOWNER_UIDS",
 763 };
 764 static DEFINE_STRARRAY(fcntl_cmds, "F_");
 765
 766 static const char *fcntl_linux_specific_cmds[] = {
 767         "SETLEASE", "GETLEASE", "NOTIFY", [5] = "CANCELLK", "DUPFD_CLOEXEC",
 768         "SETPIPE_SZ", "GETPIPE_SZ", "ADD_SEALS", "GET_SEALS",
 769         "GET_RW_HINT", "SET_RW_HINT", "GET_FILE_RW_HINT", "SET_FILE_RW_HINT",
 770 };
 771
 772 static DEFINE_STRARRAY_OFFSET(fcntl_linux_specific_cmds, "F_", F_LINUX_SPECIFIC_BASE);
 773
 774 static struct strarray *fcntl_cmds_arrays[] = {
 775         &strarray__fcntl_cmds,
 776         &strarray__fcntl_linux_specific_cmds,
 777 };
 778
 779 static DEFINE_STRARRAYS(fcntl_cmds_arrays);
 780
 781 static const char *rlimit_resources[] = {
 782         "CPU", "FSIZE", "DATA", "STACK", "CORE", "RSS", "NPROC", "NOFILE",
 783         "MEMLOCK", "AS", "LOCKS", "SIGPENDING", "MSGQUEUE", "NICE", "RTPRIO",
 784         "RTTIME",
 785 };
 786 static DEFINE_STRARRAY(rlimit_resources, "RLIMIT_");
 787
 788 static const char *sighow[] = { "BLOCK", "UNBLOCK", "SETMASK", };
 789 static DEFINE_STRARRAY(sighow, "SIG_");
 790
 791 static const char *clockid[] = {
 792         "REALTIME", "MONOTONIC", "PROCESS_CPUTIME_ID", "THREAD_CPUTIME_ID",
 793         "MONOTONIC_RAW", "REALTIME_COARSE", "MONOTONIC_COARSE", "BOOTTIME",
 794         "REALTIME_ALARM", "BOOTTIME_ALARM", "SGI_CYCLE", "TAI"
 795 };
 796 static DEFINE_STRARRAY(clockid, "CLOCK_");
 797
 798 static size_t syscall_arg__scnprintf_access_mode(char *bf, size_t size,
 799                                                  struct syscall_arg *arg)
 800 {
 801         bool show_prefix = arg->show_string_prefix;
 802         const char *suffix = "_OK";
 803         size_t printed = 0;
 804         int mode = arg->val;
 805
 806         if (mode == F_OK) /* 0 */
 807                 return scnprintf(bf, size, "F%s", show_prefix ? suffix : "");
 808 #define P_MODE(n) \
 809         if (mode & n##_OK) { \
 810                 printed += scnprintf(bf + printed, size - printed, "%s%s", #n, show_prefix ? suffix : ""); \
 811                 mode &= ~n##_OK; \
 812         }
 813
 814         P_MODE(R);
 815         P_MODE(W);
 816         P_MODE(X);
 817 #undef P_MODE
 818
 819         if (mode)
 820                 printed += scnprintf(bf + printed, size - printed, "|%#x", mode);
 821
 822         return printed;
 823 }
 824
 825 #define SCA_ACCMODE syscall_arg__scnprintf_access_mode
 826
 827 static size_t syscall_arg__scnprintf_filename(char *bf, size_t size,
 828                                               struct syscall_arg *arg);
 829
 830 #define SCA_FILENAME syscall_arg__scnprintf_filename
 831
 832 static size_t syscall_arg__scnprintf_pipe_flags(char *bf, size_t size,
 833                                                 struct syscall_arg *arg)
 834 {
 835         bool show_prefix = arg->show_string_prefix;
 836         const char *prefix = "O_";
 837         int printed = 0, flags = arg->val;
 838
 839 #define P_FLAG(n) \
 840         if (flags & O_##n) { \
 841                 printed += scnprintf(bf + printed, size - printed, "%s%s%s", printed ? "|" : "", show_prefix ? prefix : "", #n); \
 842                 flags &= ~O_##n; \
 843         }
 844
 845         P_FLAG(CLOEXEC);
 846         P_FLAG(NONBLOCK);
 847 #undef P_FLAG
 848
 849         if (flags)
 850                 printed += scnprintf(bf + printed, size - printed, "%s%#x", printed ? "|" : "", flags);
 851
 852         return printed;
 853 }
 854
 855 #define SCA_PIPE_FLAGS syscall_arg__scnprintf_pipe_flags
 856
 857 #ifndef GRND_NONBLOCK
 858 #define GRND_NONBLOCK   0x0001
 859 #endif
 860 #ifndef GRND_RANDOM
 861 #define GRND_RANDOM     0x0002
 862 #endif
 863
 864 static size_t syscall_arg__scnprintf_getrandom_flags(char *bf, size_t size,
 865                                                    struct syscall_arg *arg)
 866 {
 867         bool show_prefix = arg->show_string_prefix;
 868         const char *prefix = "GRND_";
 869         int printed = 0, flags = arg->val;
 870
 871 #define P_FLAG(n) \
 872         if (flags & GRND_##n) { \
 873                 printed += scnprintf(bf + printed, size - printed, "%s%s%s", printed ? "|" : "", show_prefix ? prefix : "", #n); \
 874                 flags &= ~GRND_##n; \
 875         }
 876
 877         P_FLAG(RANDOM);
 878         P_FLAG(NONBLOCK);
 879 #undef P_FLAG
 880
 881         if (flags)
 882                 printed += scnprintf(bf + printed, size - printed, "%s%#x", printed ? "|" : "", flags);
 883
 884         return printed;
 885 }
 886
 887 #define SCA_GETRANDOM_FLAGS syscall_arg__scnprintf_getrandom_flags
 888
 889 #define STRARRAY(name, array) \
 890           { .scnprintf  = SCA_STRARRAY, \
 891             .strtoul    = STUL_STRARRAY, \
 892             .parm       = &strarray__##array, }
 893
 894 #define STRARRAY_FLAGS(name, array) \
 895           { .scnprintf  = SCA_STRARRAY_FLAGS, \
 896             .strtoul    = STUL_STRARRAY_FLAGS, \
 897             .parm       = &strarray__##array, }
 898
 899 #include "trace/beauty/arch_errno_names.c"
 900 #include "trace/beauty/eventfd.c"
 901 #include "trace/beauty/futex_op.c"
 902 #include "trace/beauty/futex_val3.c"
 903 #include "trace/beauty/mmap.c"
 904 #include "trace/beauty/mode_t.c"
 905 #include "trace/beauty/msg_flags.c"
 906 #include "trace/beauty/open_flags.c"
 907 #include "trace/beauty/perf_event_open.c"
 908 #include "trace/beauty/pid.c"
 909 #include "trace/beauty/sched_policy.c"
 910 #include "trace/beauty/seccomp.c"
 911 #include "trace/beauty/signum.c"
 912 #include "trace/beauty/socket_type.c"
 913 #include "trace/beauty/waitid_options.c"
 914
 915 static const struct syscall_fmt syscall_fmts[] = {
 916         { .name     = "access",
 917           .arg = { [1] = { .scnprintf = SCA_ACCMODE,  /* mode */ }, }, },
 918         { .name     = "arch_prctl",
 919           .arg = { [0] = { .scnprintf = SCA_X86_ARCH_PRCTL_CODE, /* code */ },
 920                    [1] = { .scnprintf = SCA_PTR, /* arg2 */ }, }, },
 921         { .name     = "bind",
 922           .arg = { [0] = { .scnprintf = SCA_INT, /* fd */ },
 923                    [1] = { .scnprintf = SCA_SOCKADDR, /* umyaddr */ },
 924                    [2] = { .scnprintf = SCA_INT, /* addrlen */ }, }, },
 925         { .name     = "bpf",
 926           .arg = { [0] = STRARRAY(cmd, bpf_cmd), }, },
 927         { .name     = "brk",        .hexret = true,
 928           .arg = { [0] = { .scnprintf = SCA_PTR, /* brk */ }, }, },
 929         { .name     = "clock_gettime",
 930           .arg = { [0] = STRARRAY(clk_id, clockid), }, },
 931         { .name     = "clock_nanosleep",
 932           .arg = { [2] = { .scnprintf = SCA_TIMESPEC,  /* rqtp */ }, }, },
 933         { .name     = "clone",      .errpid = true, .nr_args = 5,
 934           .arg = { [0] = { .name = "flags",         .scnprintf = SCA_CLONE_FLAGS, },
 935                    [1] = { .name = "child_stack",   .scnprintf = SCA_HEX, },
 936                    [2] = { .name = "parent_tidptr", .scnprintf = SCA_HEX, },
 937                    [3] = { .name = "child_tidptr",  .scnprintf = SCA_HEX, },
 938                    [4] = { .name = "tls",           .scnprintf = SCA_HEX, }, }, },
 939         { .name     = "close",
 940           .arg = { [0] = { .scnprintf = SCA_CLOSE_FD, /* fd */ }, }, },
 941         { .name     = "connect",
 942           .arg = { [0] = { .scnprintf = SCA_INT, /* fd */ },
 943                    [1] = { .scnprintf = SCA_SOCKADDR, /* servaddr */ },
 944                    [2] = { .scnprintf = SCA_INT, /* addrlen */ }, }, },
 945         { .name     = "epoll_ctl",
 946           .arg = { [1] = STRARRAY(op, epoll_ctl_ops), }, },
 947         { .name     = "eventfd2",
 948           .arg = { [1] = { .scnprintf = SCA_EFD_FLAGS, /* flags */ }, }, },
 949         { .name     = "fchmodat",
 950           .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
 951         { .name     = "fchownat",
 952           .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
 953         { .name     = "fcntl",
 954           .arg = { [1] = { .scnprintf = SCA_FCNTL_CMD,  /* cmd */
 955                            .strtoul   = STUL_STRARRAYS,
 956                            .parm      = &strarrays__fcntl_cmds_arrays,
 957                            .show_zero = true, },
 958                    [2] = { .scnprintf =  SCA_FCNTL_ARG, /* arg */ }, }, },
 959         { .name     = "flock",
 960           .arg = { [1] = { .scnprintf = SCA_FLOCK, /* cmd */ }, }, },
 961         { .name     = "fsconfig",
 962           .arg = { [1] = STRARRAY(cmd, fsconfig_cmds), }, },
 963         { .name     = "fsmount",
 964           .arg = { [1] = STRARRAY_FLAGS(flags, fsmount_flags),
 965                    [2] = { .scnprintf = SCA_FSMOUNT_ATTR_FLAGS, /* attr_flags */ }, }, },
 966         { .name     = "fspick",
 967           .arg = { [0] = { .scnprintf = SCA_FDAT,         /* dfd */ },
 968                    [1] = { .scnprintf = SCA_FILENAME,     /* path */ },
 969                    [2] = { .scnprintf = SCA_FSPICK_FLAGS, /* flags */ }, }, },
 970         { .name     = "fstat", .alias = "newfstat", },
 971         { .name     = "fstatat", .alias = "newfstatat", },
 972         { .name     = "futex",
 973           .arg = { [1] = { .scnprintf = SCA_FUTEX_OP, /* op */ },
 974                    [5] = { .scnprintf = SCA_FUTEX_VAL3, /* val3 */ }, }, },
 975         { .name     = "futimesat",
 976           .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
 977         { .name     = "getitimer",
 978           .arg = { [0] = STRARRAY(which, itimers), }, },
 979         { .name     = "getpid",     .errpid = true, },
 980         { .name     = "getpgid",    .errpid = true, },
 981         { .name     = "getppid",    .errpid = true, },
 982         { .name     = "getrandom",
 983           .arg = { [2] = { .scnprintf = SCA_GETRANDOM_FLAGS, /* flags */ }, }, },
 984         { .name     = "getrlimit",
 985           .arg = { [0] = STRARRAY(resource, rlimit_resources), }, },
 986         { .name     = "getsockopt",
 987           .arg = { [1] = STRARRAY(level, socket_level), }, },
 988         { .name     = "gettid",     .errpid = true, },
 989         { .name     = "ioctl",
 990           .arg = {
 991 #if defined(__i386__) || defined(__x86_64__)
 992 /*
 993  * FIXME: Make this available to all arches.
 994  */
 995                    [1] = { .scnprintf = SCA_IOCTL_CMD, /* cmd */ },
 996                    [2] = { .scnprintf = SCA_HEX, /* arg */ }, }, },
 997 #else
 998                    [2] = { .scnprintf = SCA_HEX, /* arg */ }, }, },
 999 #endif
1000         { .name     = "kcmp",       .nr_args = 5,
1001           .arg = { [0] = { .name = "pid1",      .scnprintf = SCA_PID, },
1002                    [1] = { .name = "pid2",      .scnprintf = SCA_PID, },
1003                    [2] = { .name = "type",      .scnprintf = SCA_KCMP_TYPE, },
1004                    [3] = { .name = "idx1",      .scnprintf = SCA_KCMP_IDX, },
1005                    [4] = { .name = "idx2",      .scnprintf = SCA_KCMP_IDX, }, }, },
1006         { .name     = "keyctl",
1007           .arg = { [0] = STRARRAY(option, keyctl_options), }, },
1008         { .name     = "kill",
1009           .arg = { [1] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
1010         { .name     = "linkat",
1011           .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
1012         { .name     = "lseek",
1013           .arg = { [2] = STRARRAY(whence, whences), }, },
1014         { .name     = "lstat", .alias = "newlstat", },
1015         { .name     = "madvise",
1016           .arg = { [0] = { .scnprintf = SCA_HEX,      /* start */ },
1017                    [2] = { .scnprintf = SCA_MADV_BHV, /* behavior */ }, }, },
1018         { .name     = "mkdirat",
1019           .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
1020         { .name     = "mknodat",
1021           .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
1022         { .name     = "mmap",       .hexret = true,
1023 /* The standard mmap maps to old_mmap on s390x */
1024 #if defined(__s390x__)
1025         .alias = "old_mmap",
1026 #endif
1027           .arg = { [2] = { .scnprintf = SCA_MMAP_PROT,  /* prot */ },
1028                    [3] = { .scnprintf = SCA_MMAP_FLAGS, /* flags */
1029                            .strtoul   = STUL_STRARRAY_FLAGS,
1030                            .parm      = &strarray__mmap_flags, },
1031                    [5] = { .scnprintf = SCA_HEX,        /* offset */ }, }, },
1032         { .name     = "mount",
1033           .arg = { [0] = { .scnprintf = SCA_FILENAME, /* dev_name */ },
1034                    [3] = { .scnprintf = SCA_MOUNT_FLAGS, /* flags */
1035                            .mask_val  = SCAMV_MOUNT_FLAGS, /* flags */ }, }, },
1036         { .name     = "move_mount",
1037           .arg = { [0] = { .scnprintf = SCA_FDAT,       /* from_dfd */ },
1038                    [1] = { .scnprintf = SCA_FILENAME, /* from_pathname */ },
1039                    [2] = { .scnprintf = SCA_FDAT,       /* to_dfd */ },
1040                    [3] = { .scnprintf = SCA_FILENAME, /* to_pathname */ },
1041                    [4] = { .scnprintf = SCA_MOVE_MOUNT_FLAGS, /* flags */ }, }, },
1042         { .name     = "mprotect",
1043           .arg = { [0] = { .scnprintf = SCA_HEX,        /* start */ },
1044                    [2] = { .scnprintf = SCA_MMAP_PROT,  /* prot */ }, }, },
1045         { .name     = "mq_unlink",
1046           .arg = { [0] = { .scnprintf = SCA_FILENAME, /* u_name */ }, }, },
1047         { .name     = "mremap",     .hexret = true,
1048           .arg = { [3] = { .scnprintf = SCA_MREMAP_FLAGS, /* flags */ }, }, },
1049         { .name     = "name_to_handle_at",
1050           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
1051         { .name     = "newfstatat",
1052           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
1053         { .name     = "open",
1054           .arg = { [1] = { .scnprintf = SCA_OPEN_FLAGS, /* flags */ }, }, },
1055         { .name     = "open_by_handle_at",
1056           .arg = { [0] = { .scnprintf = SCA_FDAT,       /* dfd */ },
1057                    [2] = { .scnprintf = SCA_OPEN_FLAGS, /* flags */ }, }, },
1058         { .name     = "openat",
1059           .arg = { [0] = { .scnprintf = SCA_FDAT,       /* dfd */ },
1060                    [2] = { .scnprintf = SCA_OPEN_FLAGS, /* flags */ }, }, },
1061         { .name     = "perf_event_open",
1062           .arg = { [0] = { .scnprintf = SCA_PERF_ATTR,  /* attr */ },
1063                    [2] = { .scnprintf = SCA_INT,        /* cpu */ },
1064                    [3] = { .scnprintf = SCA_FD,         /* group_fd */ },
1065                    [4] = { .scnprintf = SCA_PERF_FLAGS, /* flags */ }, }, },
1066         { .name     = "pipe2",
1067           .arg = { [1] = { .scnprintf = SCA_PIPE_FLAGS, /* flags */ }, }, },
1068         { .name     = "pkey_alloc",
1069           .arg = { [1] = { .scnprintf = SCA_PKEY_ALLOC_ACCESS_RIGHTS,   /* access_rights */ }, }, },
1070         { .name     = "pkey_free",
1071           .arg = { [0] = { .scnprintf = SCA_INT,        /* key */ }, }, },
1072         { .name     = "pkey_mprotect",
1073           .arg = { [0] = { .scnprintf = SCA_HEX,        /* start */ },
1074                    [2] = { .scnprintf = SCA_MMAP_PROT,  /* prot */ },
1075                    [3] = { .scnprintf = SCA_INT,        /* pkey */ }, }, },
1076         { .name     = "poll", .timeout = true, },
1077         { .name     = "ppoll", .timeout = true, },
1078         { .name     = "prctl",
1079           .arg = { [0] = { .scnprintf = SCA_PRCTL_OPTION, /* option */
1080                            .strtoul   = STUL_STRARRAY,
1081                            .parm      = &strarray__prctl_options, },
1082                    [1] = { .scnprintf = SCA_PRCTL_ARG2, /* arg2 */ },
1083                    [2] = { .scnprintf = SCA_PRCTL_ARG3, /* arg3 */ }, }, },
1084         { .name     = "pread", .alias = "pread64", },
1085         { .name     = "preadv", .alias = "pread", },
1086         { .name     = "prlimit64",
1087           .arg = { [1] = STRARRAY(resource, rlimit_resources), }, },
1088         { .name     = "pwrite", .alias = "pwrite64", },
1089         { .name     = "readlinkat",
1090           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
1091         { .name     = "recvfrom",
1092           .arg = { [3] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
1093         { .name     = "recvmmsg",
1094           .arg = { [3] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
1095         { .name     = "recvmsg",
1096           .arg = { [2] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
1097         { .name     = "renameat",
1098           .arg = { [0] = { .scnprintf = SCA_FDAT, /* olddirfd */ },
1099                    [2] = { .scnprintf = SCA_FDAT, /* newdirfd */ }, }, },
1100         { .name     = "renameat2",
1101           .arg = { [0] = { .scnprintf = SCA_FDAT, /* olddirfd */ },
1102                    [2] = { .scnprintf = SCA_FDAT, /* newdirfd */ },
1103                    [4] = { .scnprintf = SCA_RENAMEAT2_FLAGS, /* flags */ }, }, },
1104         { .name     = "rt_sigaction",
1105           .arg = { [0] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
1106         { .name     = "rt_sigprocmask",
1107           .arg = { [0] = STRARRAY(how, sighow), }, },
1108         { .name     = "rt_sigqueueinfo",
1109           .arg = { [1] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
1110         { .name     = "rt_tgsigqueueinfo",
1111           .arg = { [2] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
1112         { .name     = "sched_setscheduler",
1113           .arg = { [1] = { .scnprintf = SCA_SCHED_POLICY, /* policy */ }, }, },
1114         { .name     = "seccomp",
1115           .arg = { [0] = { .scnprintf = SCA_SECCOMP_OP,    /* op */ },
1116                    [1] = { .scnprintf = SCA_SECCOMP_FLAGS, /* flags */ }, }, },
1117         { .name     = "select", .timeout = true, },
1118         { .name     = "sendfile", .alias = "sendfile64", },
1119         { .name     = "sendmmsg",
1120           .arg = { [3] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
1121         { .name     = "sendmsg",
1122           .arg = { [2] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
1123         { .name     = "sendto",
1124           .arg = { [3] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ },
1125                    [4] = { .scnprintf = SCA_SOCKADDR, /* addr */ }, }, },
1126         { .name     = "set_tid_address", .errpid = true, },
1127         { .name     = "setitimer",
1128           .arg = { [0] = STRARRAY(which, itimers), }, },
1129         { .name     = "setrlimit",
1130           .arg = { [0] = STRARRAY(resource, rlimit_resources), }, },
1131         { .name     = "setsockopt",
1132           .arg = { [1] = STRARRAY(level, socket_level), }, },
1133         { .name     = "socket",
1134           .arg = { [0] = STRARRAY(family, socket_families),
1135                    [1] = { .scnprintf = SCA_SK_TYPE, /* type */ },
1136                    [2] = { .scnprintf = SCA_SK_PROTO, /* protocol */ }, }, },
1137         { .name     = "socketpair",
1138           .arg = { [0] = STRARRAY(family, socket_families),
1139                    [1] = { .scnprintf = SCA_SK_TYPE, /* type */ },
1140                    [2] = { .scnprintf = SCA_SK_PROTO, /* protocol */ }, }, },
1141         { .name     = "stat", .alias = "newstat", },
1142         { .name     = "statx",
1143           .arg = { [0] = { .scnprintf = SCA_FDAT,        /* fdat */ },
1144                    [2] = { .scnprintf = SCA_STATX_FLAGS, /* flags */ } ,
1145                    [3] = { .scnprintf = SCA_STATX_MASK,  /* mask */ }, }, },
1146         { .name     = "swapoff",
1147           .arg = { [0] = { .scnprintf = SCA_FILENAME, /* specialfile */ }, }, },
1148         { .name     = "swapon",
1149           .arg = { [0] = { .scnprintf = SCA_FILENAME, /* specialfile */ }, }, },
1150         { .name     = "symlinkat",
1151           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
1152         { .name     = "sync_file_range",
1153           .arg = { [3] = { .scnprintf = SCA_SYNC_FILE_RANGE_FLAGS, /* flags */ }, }, },
1154         { .name     = "tgkill",
1155           .arg = { [2] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
1156         { .name     = "tkill",
1157           .arg = { [1] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
1158         { .name     = "umount2", .alias = "umount",
1159           .arg = { [0] = { .scnprintf = SCA_FILENAME, /* name */ }, }, },
1160         { .name     = "uname", .alias = "newuname", },
1161         { .name     = "unlinkat",
1162           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
1163         { .name     = "utimensat",
1164           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dirfd */ }, }, },
1165         { .name     = "wait4",      .errpid = true,
1166           .arg = { [2] = { .scnprintf = SCA_WAITID_OPTIONS, /* options */ }, }, },
1167         { .name     = "waitid",     .errpid = true,
1168           .arg = { [3] = { .scnprintf = SCA_WAITID_OPTIONS, /* options */ }, }, },
1169 };
1170
1171 static int syscall_fmt__cmp(const void *name, const void *fmtp)
1172 {
1173         const struct syscall_fmt *fmt = fmtp;
1174         return strcmp(name, fmt->name);
1175 }
1176
1177 static const struct syscall_fmt *__syscall_fmt__find(const struct syscall_fmt *fmts,
1178                                                      const int nmemb,
1179                                                      const char *name)
1180 {
1181         return bsearch(name, fmts, nmemb, sizeof(struct syscall_fmt), syscall_fmt__cmp);
1182 }
1183
1184 static const struct syscall_fmt *syscall_fmt__find(const char *name)
1185 {
1186         const int nmemb = ARRAY_SIZE(syscall_fmts);
1187         return __syscall_fmt__find(syscall_fmts, nmemb, name);
1188 }
1189
1190 static const struct syscall_fmt *__syscall_fmt__find_by_alias(const struct syscall_fmt *fmts,
1191                                                               const int nmemb, const char *alias)
1192 {
1193         int i;
1194
1195         for (i = 0; i < nmemb; ++i) {
1196                 if (fmts[i].alias && strcmp(fmts[i].alias, alias) == 0)
1197                         return &fmts[i];
1198         }
1199
1200         return NULL;
1201 }
1202
1203 static const struct syscall_fmt *syscall_fmt__find_by_alias(const char *alias)
1204 {
1205         const int nmemb = ARRAY_SIZE(syscall_fmts);
1206         return __syscall_fmt__find_by_alias(syscall_fmts, nmemb, alias);
1207 }
1208
1209 /*
1210  * is_exit: is this "exit" or "exit_group"?
1211  * is_open: is this "open" or "openat"? To associate the fd returned in sys_exit with the pathname in sys_enter.
1212  * args_size: sum of the sizes of the syscall arguments, anything after that is augmented stuff: pathname for openat, etc.
1213  * nonexistent: Just a hole in the syscall table, syscall id not allocated
1214  */
1215 struct syscall {
1216         struct tep_event    *tp_format;
1217         int                 nr_args;
1218         int                 args_size;
1219         struct {
1220                 struct bpf_program *sys_enter,
1221                                    *sys_exit;
1222         }                   bpf_prog;
1223         bool                is_exit;
1224         bool                is_open;
1225         bool                nonexistent;
1226         struct tep_format_field *args;
1227         const char          *name;
1228         const struct syscall_fmt  *fmt;
1229         struct syscall_arg_fmt *arg_fmt;
1230 };
1231
1232 /*
1233  * We need to have this 'calculated' boolean because in some cases we really
1234  * don't know what is the duration of a syscall, for instance, when we start
1235  * a session and some threads are waiting for a syscall to finish, say 'poll',
1236  * in which case all we can do is to print "( ? ) for duration and for the
1237  * start timestamp.
1238  */
1239 static size_t fprintf_duration(unsigned long t, bool calculated, FILE *fp)
1240 {
1241         double duration = (double)t / NSEC_PER_MSEC;
1242         size_t printed = fprintf(fp, "(");
1243
1244         if (!calculated)
1245                 printed += fprintf(fp, "         ");
1246         else if (duration >= 1.0)
1247                 printed += color_fprintf(fp, PERF_COLOR_RED, "%6.3f ms", duration);
1248         else if (duration >= 0.01)
1249                 printed += color_fprintf(fp, PERF_COLOR_YELLOW, "%6.3f ms", duration);
1250         else
1251                 printed += color_fprintf(fp, PERF_COLOR_NORMAL, "%6.3f ms", duration);
1252         return printed + fprintf(fp, "): ");
1253 }
1254
1255 /**
1256  * filename.ptr: The filename char pointer that will be vfs_getname'd
1257  * filename.entry_str_pos: Where to insert the string translated from
1258  *                         filename.ptr by the vfs_getname tracepoint/kprobe.
1259  * ret_scnprintf: syscall args may set this to a different syscall return
1260  *                formatter, for instance, fcntl may return fds, file flags, etc.
1261  */
1262 struct thread_trace {
1263         u64               entry_time;
1264         bool              entry_pending;
1265         unsigned long     nr_events;
1266         unsigned long     pfmaj, pfmin;
1267         char              *entry_str;
1268         double            runtime_ms;
1269         size_t            (*ret_scnprintf)(char *bf, size_t size, struct syscall_arg *arg);
1270         struct {
1271                 unsigned long ptr;
1272                 short int     entry_str_pos;
1273                 bool          pending_open;
1274                 unsigned int  namelen;
1275                 char          *name;
1276         } filename;
1277         struct {
1278                 int           max;
1279                 struct file   *table;
1280         } files;
1281
1282         struct intlist *syscall_stats;
1283 };
1284
1285 static struct thread_trace *thread_trace__new(void)
1286 {
1287         struct thread_trace *ttrace =  zalloc(sizeof(struct thread_trace));
1288
1289         if (ttrace) {
1290                 ttrace->files.max = -1;
1291                 ttrace->syscall_stats = intlist__new(NULL);
1292         }
1293
1294         return ttrace;
1295 }
1296
1297 static void thread_trace__free_files(struct thread_trace *ttrace);
1298
1299 static void thread_trace__delete(void *pttrace)
1300 {
1301         struct thread_trace *ttrace = pttrace;
1302
1303         if (!ttrace)
1304                 return;
1305
1306         intlist__delete(ttrace->syscall_stats);
1307         ttrace->syscall_stats = NULL;
1308         thread_trace__free_files(ttrace);
1309         zfree(&ttrace->entry_str);
1310         free(ttrace);
1311 }
1312
1313 static struct thread_trace *thread__trace(struct thread *thread, FILE *fp)
1314 {
1315         struct thread_trace *ttrace;
1316
1317         if (thread == NULL)
1318                 goto fail;
1319
1320         if (thread__priv(thread) == NULL)
1321                 thread__set_priv(thread, thread_trace__new());
1322
1323         if (thread__priv(thread) == NULL)
1324                 goto fail;
1325
1326         ttrace = thread__priv(thread);
1327         ++ttrace->nr_events;
1328
1329         return ttrace;
1330 fail:
1331         color_fprintf(fp, PERF_COLOR_RED,
1332                       "WARNING: not enough memory, dropping samples!\n");
1333         return NULL;
1334 }
1335
1336
1337 void syscall_arg__set_ret_scnprintf(struct syscall_arg *arg,
1338                                     size_t (*ret_scnprintf)(char *bf, size_t size, struct syscall_arg *arg))
1339 {
1340         struct thread_trace *ttrace = thread__priv(arg->thread);
1341
1342         ttrace->ret_scnprintf = ret_scnprintf;
1343 }
1344
1345 #define TRACE_PFMAJ             (1 << 0)
1346 #define TRACE_PFMIN             (1 << 1)
1347
1348 static const size_t trace__entry_str_size = 2048;
1349
1350 static void thread_trace__free_files(struct thread_trace *ttrace)
1351 {
1352         for (int i = 0; i < ttrace->files.max; ++i) {
1353                 struct file *file = ttrace->files.table + i;
1354                 zfree(&file->pathname);
1355         }
1356
1357         zfree(&ttrace->files.table);
1358         ttrace->files.max  = -1;
1359 }
1360
1361 static struct file *thread_trace__files_entry(struct thread_trace *ttrace, int fd)
1362 {
1363         if (fd < 0)
1364                 return NULL;
1365
1366         if (fd > ttrace->files.max) {
1367                 struct file *nfiles = realloc(ttrace->files.table, (fd + 1) * sizeof(struct file));
1368
1369                 if (nfiles == NULL)
1370                         return NULL;
1371
1372                 if (ttrace->files.max != -1) {
1373                         memset(nfiles + ttrace->files.max + 1, 0,
1374                                (fd - ttrace->files.max) * sizeof(struct file));
1375                 } else {
1376                         memset(nfiles, 0, (fd + 1) * sizeof(struct file));
1377                 }
1378
1379                 ttrace->files.table = nfiles;
1380                 ttrace->files.max   = fd;
1381         }
1382
1383         return ttrace->files.table + fd;
1384 }
1385
1386 struct file *thread__files_entry(struct thread *thread, int fd)
1387 {
1388         return thread_trace__files_entry(thread__priv(thread), fd);
1389 }
1390
1391 static int trace__set_fd_pathname(struct thread *thread, int fd, const char *pathname)
1392 {
1393         struct thread_trace *ttrace = thread__priv(thread);
1394         struct file *file = thread_trace__files_entry(ttrace, fd);
1395
1396         if (file != NULL) {
1397                 struct stat st;
1398                 if (stat(pathname, &st) == 0)
1399                         file->dev_maj = major(st.st_rdev);
1400                 file->pathname = strdup(pathname);
1401                 if (file->pathname)
1402                         return 0;
1403         }
1404
1405         return -1;
1406 }
1407
1408 static int thread__read_fd_path(struct thread *thread, int fd)
1409 {
1410         char linkname[PATH_MAX], pathname[PATH_MAX];
1411         struct stat st;
1412         int ret;
1413
1414         if (thread__pid(thread) == thread__tid(thread)) {
1415                 scnprintf(linkname, sizeof(linkname),
1416                           "/proc/%d/fd/%d", thread__pid(thread), fd);
1417         } else {
1418                 scnprintf(linkname, sizeof(linkname),
1419                           "/proc/%d/task/%d/fd/%d",
1420                           thread__pid(thread), thread__tid(thread), fd);
1421         }
1422
1423         if (lstat(linkname, &st) < 0 || st.st_size + 1 > (off_t)sizeof(pathname))
1424                 return -1;
1425
1426         ret = readlink(linkname, pathname, sizeof(pathname));
1427
1428         if (ret < 0 || ret > st.st_size)
1429                 return -1;
1430
1431         pathname[ret] = '\0';
1432         return trace__set_fd_pathname(thread, fd, pathname);
1433 }
1434
1435 static const char *thread__fd_path(struct thread *thread, int fd,
1436                                    struct trace *trace)
1437 {
1438         struct thread_trace *ttrace = thread__priv(thread);
1439
1440         if (ttrace == NULL || trace->fd_path_disabled)
1441                 return NULL;
1442
1443         if (fd < 0)
1444                 return NULL;
1445
1446         if ((fd > ttrace->files.max || ttrace->files.table[fd].pathname == NULL)) {
1447                 if (!trace->live)
1448                         return NULL;
1449                 ++trace->stats.proc_getname;
1450                 if (thread__read_fd_path(thread, fd))
1451                         return NULL;
1452         }
1453
1454         return ttrace->files.table[fd].pathname;
1455 }
1456
1457 size_t syscall_arg__scnprintf_fd(char *bf, size_t size, struct syscall_arg *arg)
1458 {
1459         int fd = arg->val;
1460         size_t printed = scnprintf(bf, size, "%d", fd);
1461         const char *path = thread__fd_path(arg->thread, fd, arg->trace);
1462
1463         if (path)
1464                 printed += scnprintf(bf + printed, size - printed, "<%s>", path);
1465
1466         return printed;
1467 }
1468
1469 size_t pid__scnprintf_fd(struct trace *trace, pid_t pid, int fd, char *bf, size_t size)
1470 {
1471         size_t printed = scnprintf(bf, size, "%d", fd);
1472         struct thread *thread = machine__find_thread(trace->host, pid, pid);
1473
1474         if (thread) {
1475                 const char *path = thread__fd_path(thread, fd, trace);
1476
1477                 if (path)
1478                         printed += scnprintf(bf + printed, size - printed, "<%s>", path);
1479
1480                 thread__put(thread);
1481         }
1482
1483         return printed;
1484 }
1485
1486 static size_t syscall_arg__scnprintf_close_fd(char *bf, size_t size,
1487                                               struct syscall_arg *arg)
1488 {
1489         int fd = arg->val;
1490         size_t printed = syscall_arg__scnprintf_fd(bf, size, arg);
1491         struct thread_trace *ttrace = thread__priv(arg->thread);
1492
1493         if (ttrace && fd >= 0 && fd <= ttrace->files.max)
1494                 zfree(&ttrace->files.table[fd].pathname);
1495
1496         return printed;
1497 }
1498
1499 static void thread__set_filename_pos(struct thread *thread, const char *bf,
1500                                      unsigned long ptr)
1501 {
1502         struct thread_trace *ttrace = thread__priv(thread);
1503
1504         ttrace->filename.ptr = ptr;
1505         ttrace->filename.entry_str_pos = bf - ttrace->entry_str;
1506 }
1507
1508 static size_t syscall_arg__scnprintf_augmented_string(struct syscall_arg *arg, char *bf, size_t size)
1509 {
1510         struct augmented_arg *augmented_arg = arg->augmented.args;
1511         size_t printed = scnprintf(bf, size, "\"%.*s\"", augmented_arg->size, augmented_arg->value);
1512         /*
1513          * So that the next arg with a payload can consume its augmented arg, i.e. for rename* syscalls
1514          * we would have two strings, each prefixed by its size.
1515          */
1516         int consumed = sizeof(*augmented_arg) + augmented_arg->size;
1517
1518         arg->augmented.args = ((void *)arg->augmented.args) + consumed;
1519         arg->augmented.size -= consumed;
1520
1521         return printed;
1522 }
1523
1524 static size_t syscall_arg__scnprintf_filename(char *bf, size_t size,
1525                                               struct syscall_arg *arg)
1526 {
1527         unsigned long ptr = arg->val;
1528
1529         if (arg->augmented.args)
1530                 return syscall_arg__scnprintf_augmented_string(arg, bf, size);
1531
1532         if (!arg->trace->vfs_getname)
1533                 return scnprintf(bf, size, "%#x", ptr);
1534
1535         thread__set_filename_pos(arg->thread, bf, ptr);
1536         return 0;
1537 }
1538
1539 static bool trace__filter_duration(struct trace *trace, double t)
1540 {
1541         return t < (trace->duration_filter * NSEC_PER_MSEC);
1542 }
1543
1544 static size_t __trace__fprintf_tstamp(struct trace *trace, u64 tstamp, FILE *fp)
1545 {
1546         double ts = (double)(tstamp - trace->base_time) / NSEC_PER_MSEC;
1547
1548         return fprintf(fp, "%10.3f ", ts);
1549 }
1550
1551 /*
1552  * We're handling tstamp=0 as an undefined tstamp, i.e. like when we are
1553  * using ttrace->entry_time for a thread that receives a sys_exit without
1554  * first having received a sys_enter ("poll" issued before tracing session
1555  * starts, lost sys_enter exit due to ring buffer overflow).
1556  */
1557 static size_t trace__fprintf_tstamp(struct trace *trace, u64 tstamp, FILE *fp)
1558 {
1559         if (tstamp > 0)
1560                 return __trace__fprintf_tstamp(trace, tstamp, fp);
1561
1562         return fprintf(fp, "         ? ");
1563 }
1564
1565 static pid_t workload_pid = -1;
1566 static volatile sig_atomic_t done = false;
1567 static volatile sig_atomic_t interrupted = false;
1568
1569 static void sighandler_interrupt(int sig __maybe_unused)
1570 {
1571         done = interrupted = true;
1572 }
1573
1574 static void sighandler_chld(int sig __maybe_unused, siginfo_t *info,
1575                             void *context __maybe_unused)
1576 {
1577         if (info->si_pid == workload_pid)
1578                 done = true;
1579 }
1580
1581 static size_t trace__fprintf_comm_tid(struct trace *trace, struct thread *thread, FILE *fp)
1582 {
1583         size_t printed = 0;
1584
1585         if (trace->multiple_threads) {
1586                 if (trace->show_comm)
1587                         printed += fprintf(fp, "%.14s/", thread__comm_str(thread));
1588                 printed += fprintf(fp, "%d ", thread__tid(thread));
1589         }
1590
1591         return printed;
1592 }
1593
1594 static size_t trace__fprintf_entry_head(struct trace *trace, struct thread *thread,
1595                                         u64 duration, bool duration_calculated, u64 tstamp, FILE *fp)
1596 {
1597         size_t printed = 0;
1598
1599         if (trace->show_tstamp)
1600                 printed = trace__fprintf_tstamp(trace, tstamp, fp);
1601         if (trace->show_duration)
1602                 printed += fprintf_duration(duration, duration_calculated, fp);
1603         return printed + trace__fprintf_comm_tid(trace, thread, fp);
1604 }
1605
1606 static int trace__process_event(struct trace *trace, struct machine *machine,
1607                                 union perf_event *event, struct perf_sample *sample)
1608 {
1609         int ret = 0;
1610
1611         switch (event->header.type) {
1612         case PERF_RECORD_LOST:
1613                 color_fprintf(trace->output, PERF_COLOR_RED,
1614                               "LOST %" PRIu64 " events!\n", event->lost.lost);
1615                 ret = machine__process_lost_event(machine, event, sample);
1616                 break;
1617         default:
1618                 ret = machine__process_event(machine, event, sample);
1619                 break;
1620         }
1621
1622         return ret;
1623 }
1624
1625 static int trace__tool_process(struct perf_tool *tool,
1626                                union perf_event *event,
1627                                struct perf_sample *sample,
1628                                struct machine *machine)
1629 {
1630         struct trace *trace = container_of(tool, struct trace, tool);
1631         return trace__process_event(trace, machine, event, sample);
1632 }
1633
1634 static char *trace__machine__resolve_kernel_addr(void *vmachine, unsigned long long *addrp, char **modp)
1635 {
1636         struct machine *machine = vmachine;
1637
1638         if (machine->kptr_restrict_warned)
1639                 return NULL;
1640
1641         if (symbol_conf.kptr_restrict) {
1642                 pr_warning("Kernel address maps (/proc/{kallsyms,modules}) are restricted.\n\n"
1643                            "Check /proc/sys/kernel/kptr_restrict and /proc/sys/kernel/perf_event_paranoid.\n\n"
1644                            "Kernel samples will not be resolved.\n");
1645                 machine->kptr_restrict_warned = true;
1646                 return NULL;
1647         }
1648
1649         return machine__resolve_kernel_addr(vmachine, addrp, modp);
1650 }
1651
1652 static int trace__symbols_init(struct trace *trace, struct evlist *evlist)
1653 {
1654         int err = symbol__init(NULL);
1655
1656         if (err)
1657                 return err;
1658
1659         trace->host = machine__new_host();
1660         if (trace->host == NULL)
1661                 return -ENOMEM;
1662
1663         thread__set_priv_destructor(thread_trace__delete);
1664
1665         err = trace_event__register_resolver(trace->host, trace__machine__resolve_kernel_addr);
1666         if (err < 0)
1667                 goto out;
1668
1669         err = __machine__synthesize_threads(trace->host, &trace->tool, &trace->opts.target,
1670                                             evlist->core.threads, trace__tool_process,
1671                                             true, false, 1);
1672 out:
1673         if (err)
1674                 symbol__exit();
1675
1676         return err;
1677 }
1678
1679 static void trace__symbols__exit(struct trace *trace)
1680 {
1681         machine__exit(trace->host);
1682         trace->host = NULL;
1683
1684         symbol__exit();
1685 }
1686
1687 static int syscall__alloc_arg_fmts(struct syscall *sc, int nr_args)
1688 {
1689         int idx;
1690
1691         if (nr_args == RAW_SYSCALL_ARGS_NUM && sc->fmt && sc->fmt->nr_args != 0)
1692                 nr_args = sc->fmt->nr_args;
1693
1694         sc->arg_fmt = calloc(nr_args, sizeof(*sc->arg_fmt));
1695         if (sc->arg_fmt == NULL)
1696                 return -1;
1697
1698         for (idx = 0; idx < nr_args; ++idx) {
1699                 if (sc->fmt)
1700                         sc->arg_fmt[idx] = sc->fmt->arg[idx];
1701         }
1702
1703         sc->nr_args = nr_args;
1704         return 0;
1705 }
1706
1707 static const struct syscall_arg_fmt syscall_arg_fmts__by_name[] = {
1708         { .name = "msr",        .scnprintf = SCA_X86_MSR,         .strtoul = STUL_X86_MSR,         },
1709         { .name = "vector",     .scnprintf = SCA_X86_IRQ_VECTORS, .strtoul = STUL_X86_IRQ_VECTORS, },
1710 };
1711
1712 static int syscall_arg_fmt__cmp(const void *name, const void *fmtp)
1713 {
1714        const struct syscall_arg_fmt *fmt = fmtp;
1715        return strcmp(name, fmt->name);
1716 }
1717
1718 static const struct syscall_arg_fmt *
1719 __syscall_arg_fmt__find_by_name(const struct syscall_arg_fmt *fmts, const int nmemb,
1720                                 const char *name)
1721 {
1722        return bsearch(name, fmts, nmemb, sizeof(struct syscall_arg_fmt), syscall_arg_fmt__cmp);
1723 }
1724
1725 static const struct syscall_arg_fmt *syscall_arg_fmt__find_by_name(const char *name)
1726 {
1727        const int nmemb = ARRAY_SIZE(syscall_arg_fmts__by_name);
1728        return __syscall_arg_fmt__find_by_name(syscall_arg_fmts__by_name, nmemb, name);
1729 }
1730
1731 static struct tep_format_field *
1732 syscall_arg_fmt__init_array(struct syscall_arg_fmt *arg, struct tep_format_field *field)
1733 {
1734         struct tep_format_field *last_field = NULL;
1735         int len;
1736
1737         for (; field; field = field->next, ++arg) {
1738                 last_field = field;
1739
1740                 if (arg->scnprintf)
1741                         continue;
1742
1743                 len = strlen(field->name);
1744
1745                 if (strcmp(field->type, "const char *") == 0 &&
1746                     ((len >= 4 && strcmp(field->name + len - 4, "name") == 0) ||
1747                      strstr(field->name, "path") != NULL))
1748                         arg->scnprintf = SCA_FILENAME;
1749                 else if ((field->flags & TEP_FIELD_IS_POINTER) || strstr(field->name, "addr"))
1750                         arg->scnprintf = SCA_PTR;
1751                 else if (strcmp(field->type, "pid_t") == 0)
1752                         arg->scnprintf = SCA_PID;
1753                 else if (strcmp(field->type, "umode_t") == 0)
1754                         arg->scnprintf = SCA_MODE_T;
1755                 else if ((field->flags & TEP_FIELD_IS_ARRAY) && strstr(field->type, "char")) {
1756                         arg->scnprintf = SCA_CHAR_ARRAY;
1757                         arg->nr_entries = field->arraylen;
1758                 } else if ((strcmp(field->type, "int") == 0 ||
1759                           strcmp(field->type, "unsigned int") == 0 ||
1760                           strcmp(field->type, "long") == 0) &&
1761                          len >= 2 && strcmp(field->name + len - 2, "fd") == 0) {
1762                         /*
1763                          * /sys/kernel/tracing/events/syscalls/sys_enter*
1764                          * grep -E 'field:.*fd;' .../format|sed -r 's/.*field:([a-z ]+) [a-z_]*fd.+/\1/g'|sort|uniq -c
1765                          * 65 int
1766                          * 23 unsigned int
1767                          * 7 unsigned long
1768                          */
1769                         arg->scnprintf = SCA_FD;
1770                 } else {
1771                         const struct syscall_arg_fmt *fmt =
1772                                 syscall_arg_fmt__find_by_name(field->name);
1773
1774                         if (fmt) {
1775                                 arg->scnprintf = fmt->scnprintf;
1776                                 arg->strtoul   = fmt->strtoul;
1777                         }
1778                 }
1779         }
1780
1781         return last_field;
1782 }
1783
1784 static int syscall__set_arg_fmts(struct syscall *sc)
1785 {
1786         struct tep_format_field *last_field = syscall_arg_fmt__init_array(sc->arg_fmt, sc->args);
1787
1788         if (last_field)
1789                 sc->args_size = last_field->offset + last_field->size;
1790
1791         return 0;
1792 }
1793
1794 static int trace__read_syscall_info(struct trace *trace, int id)
1795 {
1796         char tp_name[128];
1797         struct syscall *sc;
1798         const char *name = syscalltbl__name(trace->sctbl, id);
1799
1800 #ifdef HAVE_SYSCALL_TABLE_SUPPORT
1801         if (trace->syscalls.table == NULL) {
1802                 trace->syscalls.table = calloc(trace->sctbl->syscalls.max_id + 1, sizeof(*sc));
1803                 if (trace->syscalls.table == NULL)
1804                         return -ENOMEM;
1805         }
1806 #else
1807         if (id > trace->sctbl->syscalls.max_id || (id == 0 && trace->syscalls.table == NULL)) {
1808                 // When using libaudit we don't know beforehand what is the max syscall id
1809                 struct syscall *table = realloc(trace->syscalls.table, (id + 1) * sizeof(*sc));
1810
1811                 if (table == NULL)
1812                         return -ENOMEM;
1813
1814                 // Need to memset from offset 0 and +1 members if brand new
1815                 if (trace->syscalls.table == NULL)
1816                         memset(table, 0, (id + 1) * sizeof(*sc));
1817                 else
1818                         memset(table + trace->sctbl->syscalls.max_id + 1, 0, (id - trace->sctbl->syscalls.max_id) * sizeof(*sc));
1819
1820                 trace->syscalls.table         = table;
1821                 trace->sctbl->syscalls.max_id = id;
1822         }
1823 #endif
1824         sc = trace->syscalls.table + id;
1825         if (sc->nonexistent)
1826                 return -EEXIST;
1827
1828         if (name == NULL) {
1829                 sc->nonexistent = true;
1830                 return -EEXIST;
1831         }
1832
1833         sc->name = name;
1834         sc->fmt  = syscall_fmt__find(sc->name);
1835
1836         snprintf(tp_name, sizeof(tp_name), "sys_enter_%s", sc->name);
1837         sc->tp_format = trace_event__tp_format("syscalls", tp_name);
1838
1839         if (IS_ERR(sc->tp_format) && sc->fmt && sc->fmt->alias) {
1840                 snprintf(tp_name, sizeof(tp_name), "sys_enter_%s", sc->fmt->alias);
1841                 sc->tp_format = trace_event__tp_format("syscalls", tp_name);
1842         }
1843
1844         /*
1845          * Fails to read trace point format via sysfs node, so the trace point
1846          * doesn't exist.  Set the 'nonexistent' flag as true.
1847          */
1848         if (IS_ERR(sc->tp_format)) {
1849                 sc->nonexistent = true;
1850                 return PTR_ERR(sc->tp_format);
1851         }
1852
1853         if (syscall__alloc_arg_fmts(sc, IS_ERR(sc->tp_format) ?
1854                                         RAW_SYSCALL_ARGS_NUM : sc->tp_format->format.nr_fields))
1855                 return -ENOMEM;
1856
1857         sc->args = sc->tp_format->format.fields;
1858         /*
1859          * We need to check and discard the first variable '__syscall_nr'
1860          * or 'nr' that mean the syscall number. It is needless here.
1861          * So drop '__syscall_nr' or 'nr' field but does not exist on older kernels.
1862          */
1863         if (sc->args && (!strcmp(sc->args->name, "__syscall_nr") || !strcmp(sc->args->name, "nr"))) {
1864                 sc->args = sc->args->next;
1865                 --sc->nr_args;
1866         }
1867
1868         sc->is_exit = !strcmp(name, "exit_group") || !strcmp(name, "exit");
1869         sc->is_open = !strcmp(name, "open") || !strcmp(name, "openat");
1870
1871         return syscall__set_arg_fmts(sc);
1872 }
1873
1874 static int evsel__init_tp_arg_scnprintf(struct evsel *evsel)
1875 {
1876         struct syscall_arg_fmt *fmt = evsel__syscall_arg_fmt(evsel);
1877
1878         if (fmt != NULL) {
1879                 syscall_arg_fmt__init_array(fmt, evsel->tp_format->format.fields);
1880                 return 0;
1881         }
1882
1883         return -ENOMEM;
1884 }
1885
1886 static int intcmp(const void *a, const void *b)
1887 {
1888         const int *one = a, *another = b;
1889
1890         return *one - *another;
1891 }
1892
1893 static int trace__validate_ev_qualifier(struct trace *trace)
1894 {
1895         int err = 0;
1896         bool printed_invalid_prefix = false;
1897         struct str_node *pos;
1898         size_t nr_used = 0, nr_allocated = strlist__nr_entries(trace->ev_qualifier);
1899
1900         trace->ev_qualifier_ids.entries = malloc(nr_allocated *
1901                                                  sizeof(trace->ev_qualifier_ids.entries[0]));
1902
1903         if (trace->ev_qualifier_ids.entries == NULL) {
1904                 fputs("Error:\tNot enough memory for allocating events qualifier ids\n",
1905                        trace->output);
1906                 err = -EINVAL;
1907                 goto out;
1908         }
1909
1910         strlist__for_each_entry(pos, trace->ev_qualifier) {
1911                 const char *sc = pos->s;
1912                 int id = syscalltbl__id(trace->sctbl, sc), match_next = -1;
1913
1914                 if (id < 0) {
1915                         id = syscalltbl__strglobmatch_first(trace->sctbl, sc, &match_next);
1916                         if (id >= 0)
1917                                 goto matches;
1918
1919                         if (!printed_invalid_prefix) {
1920                                 pr_debug("Skipping unknown syscalls: ");
1921                                 printed_invalid_prefix = true;
1922                         } else {
1923                                 pr_debug(", ");
1924                         }
1925
1926                         pr_debug("%s", sc);
1927                         continue;
1928                 }
1929 matches:
1930                 trace->ev_qualifier_ids.entries[nr_used++] = id;
1931                 if (match_next == -1)
1932                         continue;
1933
1934                 while (1) {
1935                         id = syscalltbl__strglobmatch_next(trace->sctbl, sc, &match_next);
1936                         if (id < 0)
1937                                 break;
1938                         if (nr_allocated == nr_used) {
1939                                 void *entries;
1940
1941                                 nr_allocated += 8;
1942                                 entries = realloc(trace->ev_qualifier_ids.entries,
1943                                                   nr_allocated * sizeof(trace->ev_qualifier_ids.entries[0]));
1944                                 if (entries == NULL) {
1945                                         err = -ENOMEM;
1946                                         fputs("\nError:\t Not enough memory for parsing\n", trace->output);
1947                                         goto out_free;
1948                                 }
1949                                 trace->ev_qualifier_ids.entries = entries;
1950                         }
1951                         trace->ev_qualifier_ids.entries[nr_used++] = id;
1952                 }
1953         }
1954
1955         trace->ev_qualifier_ids.nr = nr_used;
1956         qsort(trace->ev_qualifier_ids.entries, nr_used, sizeof(int), intcmp);
1957 out:
1958         if (printed_invalid_prefix)
1959                 pr_debug("\n");
1960         return err;
1961 out_free:
1962         zfree(&trace->ev_qualifier_ids.entries);
1963         trace->ev_qualifier_ids.nr = 0;
1964         goto out;
1965 }
1966
1967 static __maybe_unused bool trace__syscall_enabled(struct trace *trace, int id)
1968 {
1969         bool in_ev_qualifier;
1970
1971         if (trace->ev_qualifier_ids.nr == 0)
1972                 return true;
1973
1974         in_ev_qualifier = bsearch(&id, trace->ev_qualifier_ids.entries,
1975                                   trace->ev_qualifier_ids.nr, sizeof(int), intcmp) != NULL;
1976
1977         if (in_ev_qualifier)
1978                return !trace->not_ev_qualifier;
1979
1980         return trace->not_ev_qualifier;
1981 }
1982
1983 /*
1984  * args is to be interpreted as a series of longs but we need to handle
1985  * 8-byte unaligned accesses. args points to raw_data within the event
1986  * and raw_data is guaranteed to be 8-byte unaligned because it is
1987  * preceded by raw_size which is a u32. So we need to copy args to a temp
1988  * variable to read it. Most notably this avoids extended load instructions
1989  * on unaligned addresses
1990  */
1991 unsigned long syscall_arg__val(struct syscall_arg *arg, u8 idx)
1992 {
1993         unsigned long val;
1994         unsigned char *p = arg->args + sizeof(unsigned long) * idx;
1995
1996         memcpy(&val, p, sizeof(val));
1997         return val;
1998 }
1999
2000 static size_t syscall__scnprintf_name(struct syscall *sc, char *bf, size_t size,
2001                                       struct syscall_arg *arg)
2002 {
2003         if (sc->arg_fmt && sc->arg_fmt[arg->idx].name)
2004                 return scnprintf(bf, size, "%s: ", sc->arg_fmt[arg->idx].name);
2005
2006         return scnprintf(bf, size, "arg%d: ", arg->idx);
2007 }
2008
2009 /*
2010  * Check if the value is in fact zero, i.e. mask whatever needs masking, such
2011  * as mount 'flags' argument that needs ignoring some magic flag, see comment
2012  * in tools/perf/trace/beauty/mount_flags.c
2013  */
2014 static unsigned long syscall_arg_fmt__mask_val(struct syscall_arg_fmt *fmt, struct syscall_arg *arg, unsigned long val)
2015 {
2016         if (fmt && fmt->mask_val)
2017                 return fmt->mask_val(arg, val);
2018
2019         return val;
2020 }
2021
2022 static size_t syscall_arg_fmt__scnprintf_val(struct syscall_arg_fmt *fmt, char *bf, size_t size,
2023                                              struct syscall_arg *arg, unsigned long val)
2024 {
2025         if (fmt && fmt->scnprintf) {
2026                 arg->val = val;
2027                 if (fmt->parm)
2028                         arg->parm = fmt->parm;
2029                 return fmt->scnprintf(bf, size, arg);
2030         }
2031         return scnprintf(bf, size, "%ld", val);
2032 }
2033
2034 static size_t syscall__scnprintf_args(struct syscall *sc, char *bf, size_t size,
2035                                       unsigned char *args, void *augmented_args, int augmented_args_size,
2036                                       struct trace *trace, struct thread *thread)
2037 {
2038         size_t printed = 0;
2039         unsigned long val;
2040         u8 bit = 1;
2041         struct syscall_arg arg = {
2042                 .args   = args,
2043                 .augmented = {
2044                         .size = augmented_args_size,
2045                         .args = augmented_args,
2046                 },
2047                 .idx    = 0,
2048                 .mask   = 0,
2049                 .trace  = trace,
2050                 .thread = thread,
2051                 .show_string_prefix = trace->show_string_prefix,
2052         };
2053         struct thread_trace *ttrace = thread__priv(thread);
2054
2055         /*
2056          * Things like fcntl will set this in its 'cmd' formatter to pick the
2057          * right formatter for the return value (an fd? file flags?), which is
2058          * not needed for syscalls that always return a given type, say an fd.
2059          */
2060         ttrace->ret_scnprintf = NULL;
2061
2062         if (sc->args != NULL) {
2063                 struct tep_format_field *field;
2064
2065                 for (field = sc->args; field;
2066                      field = field->next, ++arg.idx, bit <<= 1) {
2067                         if (arg.mask & bit)
2068                                 continue;
2069
2070                         arg.fmt = &sc->arg_fmt[arg.idx];
2071                         val = syscall_arg__val(&arg, arg.idx);
2072                         /*
2073                          * Some syscall args need some mask, most don't and
2074                          * return val untouched.
2075                          */
2076                         val = syscall_arg_fmt__mask_val(&sc->arg_fmt[arg.idx], &arg, val);
2077
2078                         /*
2079                          * Suppress this argument if its value is zero and
2080                          * and we don't have a string associated in an
2081                          * strarray for it.
2082                          */
2083                         if (val == 0 &&
2084                             !trace->show_zeros &&
2085                             !(sc->arg_fmt &&
2086                               (sc->arg_fmt[arg.idx].show_zero ||
2087                                sc->arg_fmt[arg.idx].scnprintf == SCA_STRARRAY ||
2088                                sc->arg_fmt[arg.idx].scnprintf == SCA_STRARRAYS) &&
2089                               sc->arg_fmt[arg.idx].parm))
2090                                 continue;
2091
2092                         printed += scnprintf(bf + printed, size - printed, "%s", printed ? ", " : "");
2093
2094                         if (trace->show_arg_names)
2095                                 printed += scnprintf(bf + printed, size - printed, "%s: ", field->name);
2096
2097                         printed += syscall_arg_fmt__scnprintf_val(&sc->arg_fmt[arg.idx],
2098                                                                   bf + printed, size - printed, &arg, val);
2099                 }
2100         } else if (IS_ERR(sc->tp_format)) {
2101                 /*
2102                  * If we managed to read the tracepoint /format file, then we
2103                  * may end up not having any args, like with gettid(), so only
2104                  * print the raw args when we didn't manage to read it.
2105                  */
2106                 while (arg.idx < sc->nr_args) {
2107                         if (arg.mask & bit)
2108                                 goto next_arg;
2109                         val = syscall_arg__val(&arg, arg.idx);
2110                         if (printed)
2111                                 printed += scnprintf(bf + printed, size - printed, ", ");
2112                         printed += syscall__scnprintf_name(sc, bf + printed, size - printed, &arg);
2113                         printed += syscall_arg_fmt__scnprintf_val(&sc->arg_fmt[arg.idx], bf + printed, size - printed, &arg, val);
2114 next_arg:
2115                         ++arg.idx;
2116                         bit <<= 1;
2117                 }
2118         }
2119
2120         return printed;
2121 }
2122
2123 typedef int (*tracepoint_handler)(struct trace *trace, struct evsel *evsel,
2124                                   union perf_event *event,
2125                                   struct perf_sample *sample);
2126
2127 static struct syscall *trace__syscall_info(struct trace *trace,
2128                                            struct evsel *evsel, int id)
2129 {
2130         int err = 0;
2131
2132         if (id < 0) {
2133
2134                 /*
2135                  * XXX: Noticed on x86_64, reproduced as far back as 3.0.36, haven't tried
2136                  * before that, leaving at a higher verbosity level till that is
2137                  * explained. Reproduced with plain ftrace with:
2138                  *
2139                  * echo 1 > /t/events/raw_syscalls/sys_exit/enable
2140                  * grep "NR -1 " /t/trace_pipe
2141                  *
2142                  * After generating some load on the machine.
2143                  */
2144                 if (verbose > 1) {
2145                         static u64 n;
2146                         fprintf(trace->output, "Invalid syscall %d id, skipping (%s, %" PRIu64 ") ...\n",
2147                                 id, evsel__name(evsel), ++n);
2148                 }
2149                 return NULL;
2150         }
2151
2152         err = -EINVAL;
2153
2154 #ifdef HAVE_SYSCALL_TABLE_SUPPORT
2155         if (id > trace->sctbl->syscalls.max_id) {
2156 #else
2157         if (id >= trace->sctbl->syscalls.max_id) {
2158                 /*
2159                  * With libaudit we don't know beforehand what is the max_id,
2160                  * so we let trace__read_syscall_info() figure that out as we
2161                  * go on reading syscalls.
2162                  */
2163                 err = trace__read_syscall_info(trace, id);
2164                 if (err)
2165 #endif
2166                 goto out_cant_read;
2167         }
2168
2169         if ((trace->syscalls.table == NULL || trace->syscalls.table[id].name == NULL) &&
2170             (err = trace__read_syscall_info(trace, id)) != 0)
2171                 goto out_cant_read;
2172
2173         if (trace->syscalls.table && trace->syscalls.table[id].nonexistent)
2174                 goto out_cant_read;
2175
2176         return &trace->syscalls.table[id];
2177
2178 out_cant_read:
2179         if (verbose > 0) {
2180                 char sbuf[STRERR_BUFSIZE];
2181                 fprintf(trace->output, "Problems reading syscall %d: %d (%s)", id, -err, str_error_r(-err, sbuf, sizeof(sbuf)));
2182                 if (id <= trace->sctbl->syscalls.max_id && trace->syscalls.table[id].name != NULL)
2183                         fprintf(trace->output, "(%s)", trace->syscalls.table[id].name);
2184                 fputs(" information\n", trace->output);
2185         }
2186         return NULL;
2187 }
2188
2189 struct syscall_stats {
2190         struct stats stats;
2191         u64          nr_failures;
2192         int          max_errno;
2193         u32          *errnos;
2194 };
2195
2196 static void thread__update_stats(struct thread *thread, struct thread_trace *ttrace,
2197                                  int id, struct perf_sample *sample, long err, bool errno_summary)
2198 {
2199         struct int_node *inode;
2200         struct syscall_stats *stats;
2201         u64 duration = 0;
2202
2203         inode = intlist__findnew(ttrace->syscall_stats, id);
2204         if (inode == NULL)
2205                 return;
2206
2207         stats = inode->priv;
2208         if (stats == NULL) {
2209                 stats = zalloc(sizeof(*stats));
2210                 if (stats == NULL)
2211                         return;
2212
2213                 init_stats(&stats->stats);
2214                 inode->priv = stats;
2215         }
2216
2217         if (ttrace->entry_time && sample->time > ttrace->entry_time)
2218                 duration = sample->time - ttrace->entry_time;
2219
2220         update_stats(&stats->stats, duration);
2221
2222         if (err < 0) {
2223                 ++stats->nr_failures;
2224
2225                 if (!errno_summary)
2226                         return;
2227
2228                 err = -err;
2229                 if (err > stats->max_errno) {
2230                         u32 *new_errnos = realloc(stats->errnos, err * sizeof(u32));
2231
2232                         if (new_errnos) {
2233                                 memset(new_errnos + stats->max_errno, 0, (err - stats->max_errno) * sizeof(u32));
2234                         } else {
2235                                 pr_debug("Not enough memory for errno stats for thread \"%s\"(%d/%d), results will be incomplete\n",
2236                                          thread__comm_str(thread), thread__pid(thread),
2237                                          thread__tid(thread));
2238                                 return;
2239                         }
2240
2241                         stats->errnos = new_errnos;
2242                         stats->max_errno = err;
2243                 }
2244
2245                 ++stats->errnos[err - 1];
2246         }
2247 }
2248
2249 static int trace__printf_interrupted_entry(struct trace *trace)
2250 {
2251         struct thread_trace *ttrace;
2252         size_t printed;
2253         int len;
2254
2255         if (trace->failure_only || trace->current == NULL)
2256                 return 0;
2257
2258         ttrace = thread__priv(trace->current);
2259
2260         if (!ttrace->entry_pending)
2261                 return 0;
2262
2263         printed  = trace__fprintf_entry_head(trace, trace->current, 0, false, ttrace->entry_time, trace->output);
2264         printed += len = fprintf(trace->output, "%s)", ttrace->entry_str);
2265
2266         if (len < trace->args_alignment - 4)
2267                 printed += fprintf(trace->output, "%-*s", trace->args_alignment - 4 - len, " ");
2268
2269         printed += fprintf(trace->output, " ...\n");
2270
2271         ttrace->entry_pending = false;
2272         ++trace->nr_events_printed;
2273
2274         return printed;
2275 }
2276
2277 static int trace__fprintf_sample(struct trace *trace, struct evsel *evsel,
2278                                  struct perf_sample *sample, struct thread *thread)
2279 {
2280         int printed = 0;
2281
2282         if (trace->print_sample) {
2283                 double ts = (double)sample->time / NSEC_PER_MSEC;
2284
2285                 printed += fprintf(trace->output, "%22s %10.3f %s %d/%d [%d]\n",
2286                                    evsel__name(evsel), ts,
2287                                    thread__comm_str(thread),
2288                                    sample->pid, sample->tid, sample->cpu);
2289         }
2290
2291         return printed;
2292 }
2293
2294 static void *syscall__augmented_args(struct syscall *sc, struct perf_sample *sample, int *augmented_args_size, int raw_augmented_args_size)
2295 {
2296         void *augmented_args = NULL;
2297         /*
2298          * For now with BPF raw_augmented we hook into raw_syscalls:sys_enter
2299          * and there we get all 6 syscall args plus the tracepoint common fields
2300          * that gets calculated at the start and the syscall_nr (another long).
2301          * So we check if that is the case and if so don't look after the
2302          * sc->args_size but always after the full raw_syscalls:sys_enter payload,
2303          * which is fixed.
2304          *
2305          * We'll revisit this later to pass s->args_size to the BPF augmenter
2306          * (now tools/perf/examples/bpf/augmented_raw_syscalls.c, so that it
2307          * copies only what we need for each syscall, like what happens when we
2308          * use syscalls:sys_enter_NAME, so that we reduce the kernel/userspace
2309          * traffic to just what is needed for each syscall.
2310          */
2311         int args_size = raw_augmented_args_size ?: sc->args_size;
2312
2313         *augmented_args_size = sample->raw_size - args_size;
2314         if (*augmented_args_size > 0)
2315                 augmented_args = sample->raw_data + args_size;
2316
2317         return augmented_args;
2318 }
2319
2320 static void syscall__exit(struct syscall *sc)
2321 {
2322         if (!sc)
2323                 return;
2324
2325         zfree(&sc->arg_fmt);
2326 }
2327
2328 static int trace__sys_enter(struct trace *trace, struct evsel *evsel,
2329                             union perf_event *event __maybe_unused,
2330                             struct perf_sample *sample)
2331 {
2332         char *msg;
2333         void *args;
2334         int printed = 0;
2335         struct thread *thread;
2336         int id = perf_evsel__sc_tp_uint(evsel, id, sample), err = -1;
2337         int augmented_args_size = 0;
2338         void *augmented_args = NULL;
2339         struct syscall *sc = trace__syscall_info(trace, evsel, id);
2340         struct thread_trace *ttrace;
2341
2342         if (sc == NULL)
2343                 return -1;
2344
2345         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2346         ttrace = thread__trace(thread, trace->output);
2347         if (ttrace == NULL)
2348                 goto out_put;
2349
2350         trace__fprintf_sample(trace, evsel, sample, thread);
2351
2352         args = perf_evsel__sc_tp_ptr(evsel, args, sample);
2353
2354         if (ttrace->entry_str == NULL) {
2355                 ttrace->entry_str = malloc(trace__entry_str_size);
2356                 if (!ttrace->entry_str)
2357                         goto out_put;
2358         }
2359
2360         if (!(trace->duration_filter || trace->summary_only || trace->min_stack))
2361                 trace__printf_interrupted_entry(trace);
2362         /*
2363          * If this is raw_syscalls.sys_enter, then it always comes with the 6 possible
2364          * arguments, even if the syscall being handled, say "openat", uses only 4 arguments
2365          * this breaks syscall__augmented_args() check for augmented args, as we calculate
2366          * syscall->args_size using each syscalls:sys_enter_NAME tracefs format file,
2367          * so when handling, say the openat syscall, we end up getting 6 args for the
2368          * raw_syscalls:sys_enter event, when we expected just 4, we end up mistakenly
2369          * thinking that the extra 2 u64 args are the augmented filename, so just check
2370          * here and avoid using augmented syscalls when the evsel is the raw_syscalls one.
2371          */
2372         if (evsel != trace->syscalls.events.sys_enter)
2373                 augmented_args = syscall__augmented_args(sc, sample, &augmented_args_size, trace->raw_augmented_syscalls_args_size);
2374         ttrace->entry_time = sample->time;
2375         msg = ttrace->entry_str;
2376         printed += scnprintf(msg + printed, trace__entry_str_size - printed, "%s(", sc->name);
2377
2378         printed += syscall__scnprintf_args(sc, msg + printed, trace__entry_str_size - printed,
2379                                            args, augmented_args, augmented_args_size, trace, thread);
2380
2381         if (sc->is_exit) {
2382                 if (!(trace->duration_filter || trace->summary_only || trace->failure_only || trace->min_stack)) {
2383                         int alignment = 0;
2384
2385                         trace__fprintf_entry_head(trace, thread, 0, false, ttrace->entry_time, trace->output);
2386                         printed = fprintf(trace->output, "%s)", ttrace->entry_str);
2387                         if (trace->args_alignment > printed)
2388                                 alignment = trace->args_alignment - printed;
2389                         fprintf(trace->output, "%*s= ?\n", alignment, " ");
2390                 }
2391         } else {
2392                 ttrace->entry_pending = true;
2393                 /* See trace__vfs_getname & trace__sys_exit */
2394                 ttrace->filename.pending_open = false;
2395         }
2396
2397         if (trace->current != thread) {
2398                 thread__put(trace->current);
2399                 trace->current = thread__get(thread);
2400         }
2401         err = 0;
2402 out_put:
2403         thread__put(thread);
2404         return err;
2405 }
2406
2407 static int trace__fprintf_sys_enter(struct trace *trace, struct evsel *evsel,
2408                                     struct perf_sample *sample)
2409 {
2410         struct thread_trace *ttrace;
2411         struct thread *thread;
2412         int id = perf_evsel__sc_tp_uint(evsel, id, sample), err = -1;
2413         struct syscall *sc = trace__syscall_info(trace, evsel, id);
2414         char msg[1024];
2415         void *args, *augmented_args = NULL;
2416         int augmented_args_size;
2417
2418         if (sc == NULL)
2419                 return -1;
2420
2421         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2422         ttrace = thread__trace(thread, trace->output);
2423         /*
2424          * We need to get ttrace just to make sure it is there when syscall__scnprintf_args()
2425          * and the rest of the beautifiers accessing it via struct syscall_arg touches it.
2426          */
2427         if (ttrace == NULL)
2428                 goto out_put;
2429
2430         args = perf_evsel__sc_tp_ptr(evsel, args, sample);
2431         augmented_args = syscall__augmented_args(sc, sample, &augmented_args_size, trace->raw_augmented_syscalls_args_size);
2432         syscall__scnprintf_args(sc, msg, sizeof(msg), args, augmented_args, augmented_args_size, trace, thread);
2433         fprintf(trace->output, "%s", msg);
2434         err = 0;
2435 out_put:
2436         thread__put(thread);
2437         return err;
2438 }
2439
2440 static int trace__resolve_callchain(struct trace *trace, struct evsel *evsel,
2441                                     struct perf_sample *sample,
2442                                     struct callchain_cursor *cursor)
2443 {
2444         struct addr_location al;
2445         int max_stack = evsel->core.attr.sample_max_stack ?
2446                         evsel->core.attr.sample_max_stack :
2447                         trace->max_stack;
2448         int err = -1;
2449
2450         addr_location__init(&al);
2451         if (machine__resolve(trace->host, &al, sample) < 0)
2452                 goto out;
2453
2454         err = thread__resolve_callchain(al.thread, cursor, evsel, sample, NULL, NULL, max_stack);
2455 out:
2456         addr_location__exit(&al);
2457         return err;
2458 }
2459
2460 static int trace__fprintf_callchain(struct trace *trace, struct perf_sample *sample)
2461 {
2462         /* TODO: user-configurable print_opts */
2463         const unsigned int print_opts = EVSEL__PRINT_SYM |
2464                                         EVSEL__PRINT_DSO |
2465                                         EVSEL__PRINT_UNKNOWN_AS_ADDR;
2466
2467         return sample__fprintf_callchain(sample, 38, print_opts, get_tls_callchain_cursor(), symbol_conf.bt_stop_list, trace->output);
2468 }
2469
2470 static const char *errno_to_name(struct evsel *evsel, int err)
2471 {
2472         struct perf_env *env = evsel__env(evsel);
2473
2474         return perf_env__arch_strerrno(env, err);
2475 }
2476
2477 static int trace__sys_exit(struct trace *trace, struct evsel *evsel,
2478                            union perf_event *event __maybe_unused,
2479                            struct perf_sample *sample)
2480 {
2481         long ret;
2482         u64 duration = 0;
2483         bool duration_calculated = false;
2484         struct thread *thread;
2485         int id = perf_evsel__sc_tp_uint(evsel, id, sample), err = -1, callchain_ret = 0, printed = 0;
2486         int alignment = trace->args_alignment;
2487         struct syscall *sc = trace__syscall_info(trace, evsel, id);
2488         struct thread_trace *ttrace;
2489
2490         if (sc == NULL)
2491                 return -1;
2492
2493         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2494         ttrace = thread__trace(thread, trace->output);
2495         if (ttrace == NULL)
2496                 goto out_put;
2497
2498         trace__fprintf_sample(trace, evsel, sample, thread);
2499
2500         ret = perf_evsel__sc_tp_uint(evsel, ret, sample);
2501
2502         if (trace->summary)
2503                 thread__update_stats(thread, ttrace, id, sample, ret, trace->errno_summary);
2504
2505         if (!trace->fd_path_disabled && sc->is_open && ret >= 0 && ttrace->filename.pending_open) {
2506                 trace__set_fd_pathname(thread, ret, ttrace->filename.name);
2507                 ttrace->filename.pending_open = false;
2508                 ++trace->stats.vfs_getname;
2509         }
2510
2511         if (ttrace->entry_time) {
2512                 duration = sample->time - ttrace->entry_time;
2513                 if (trace__filter_duration(trace, duration))
2514                         goto out;
2515                 duration_calculated = true;
2516         } else if (trace->duration_filter)
2517                 goto out;
2518
2519         if (sample->callchain) {
2520                 struct callchain_cursor *cursor = get_tls_callchain_cursor();
2521
2522                 callchain_ret = trace__resolve_callchain(trace, evsel, sample, cursor);
2523                 if (callchain_ret == 0) {
2524                         if (cursor->nr < trace->min_stack)
2525                                 goto out;
2526                         callchain_ret = 1;
2527                 }
2528         }
2529
2530         if (trace->summary_only || (ret >= 0 && trace->failure_only))
2531                 goto out;
2532
2533         trace__fprintf_entry_head(trace, thread, duration, duration_calculated, ttrace->entry_time, trace->output);
2534
2535         if (ttrace->entry_pending) {
2536                 printed = fprintf(trace->output, "%s", ttrace->entry_str);
2537         } else {
2538                 printed += fprintf(trace->output, " ... [");
2539                 color_fprintf(trace->output, PERF_COLOR_YELLOW, "continued");
2540                 printed += 9;
2541                 printed += fprintf(trace->output, "]: %s()", sc->name);
2542         }
2543
2544         printed++; /* the closing ')' */
2545
2546         if (alignment > printed)
2547                 alignment -= printed;
2548         else
2549                 alignment = 0;
2550
2551         fprintf(trace->output, ")%*s= ", alignment, " ");
2552
2553         if (sc->fmt == NULL) {
2554                 if (ret < 0)
2555                         goto errno_print;
2556 signed_print:
2557                 fprintf(trace->output, "%ld", ret);
2558         } else if (ret < 0) {
2559 errno_print: {
2560                 char bf[STRERR_BUFSIZE];
2561                 const char *emsg = str_error_r(-ret, bf, sizeof(bf)),
2562                            *e = errno_to_name(evsel, -ret);
2563
2564                 fprintf(trace->output, "-1 %s (%s)", e, emsg);
2565         }
2566         } else if (ret == 0 && sc->fmt->timeout)
2567                 fprintf(trace->output, "0 (Timeout)");
2568         else if (ttrace->ret_scnprintf) {
2569                 char bf[1024];
2570                 struct syscall_arg arg = {
2571                         .val    = ret,
2572                         .thread = thread,
2573                         .trace  = trace,
2574                 };
2575                 ttrace->ret_scnprintf(bf, sizeof(bf), &arg);
2576                 ttrace->ret_scnprintf = NULL;
2577                 fprintf(trace->output, "%s", bf);
2578         } else if (sc->fmt->hexret)
2579                 fprintf(trace->output, "%#lx", ret);
2580         else if (sc->fmt->errpid) {
2581                 struct thread *child = machine__find_thread(trace->host, ret, ret);
2582
2583                 if (child != NULL) {
2584                         fprintf(trace->output, "%ld", ret);
2585                         if (thread__comm_set(child))
2586                                 fprintf(trace->output, " (%s)", thread__comm_str(child));
2587                         thread__put(child);
2588                 }
2589         } else
2590                 goto signed_print;
2591
2592         fputc('\n', trace->output);
2593
2594         /*
2595          * We only consider an 'event' for the sake of --max-events a non-filtered
2596          * sys_enter + sys_exit and other tracepoint events.
2597          */
2598         if (++trace->nr_events_printed == trace->max_events && trace->max_events != ULONG_MAX)
2599                 interrupted = true;
2600
2601         if (callchain_ret > 0)
2602                 trace__fprintf_callchain(trace, sample);
2603         else if (callchain_ret < 0)
2604                 pr_err("Problem processing %s callchain, skipping...\n", evsel__name(evsel));
2605 out:
2606         ttrace->entry_pending = false;
2607         err = 0;
2608 out_put:
2609         thread__put(thread);
2610         return err;
2611 }
2612
2613 static int trace__vfs_getname(struct trace *trace, struct evsel *evsel,
2614                               union perf_event *event __maybe_unused,
2615                               struct perf_sample *sample)
2616 {
2617         struct thread *thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2618         struct thread_trace *ttrace;
2619         size_t filename_len, entry_str_len, to_move;
2620         ssize_t remaining_space;
2621         char *pos;
2622         const char *filename = evsel__rawptr(evsel, sample, "pathname");
2623
2624         if (!thread)
2625                 goto out;
2626
2627         ttrace = thread__priv(thread);
2628         if (!ttrace)
2629                 goto out_put;
2630
2631         filename_len = strlen(filename);
2632         if (filename_len == 0)
2633                 goto out_put;
2634
2635         if (ttrace->filename.namelen < filename_len) {
2636                 char *f = realloc(ttrace->filename.name, filename_len + 1);
2637
2638                 if (f == NULL)
2639                         goto out_put;
2640
2641                 ttrace->filename.namelen = filename_len;
2642                 ttrace->filename.name = f;
2643         }
2644
2645         strcpy(ttrace->filename.name, filename);
2646         ttrace->filename.pending_open = true;
2647
2648         if (!ttrace->filename.ptr)
2649                 goto out_put;
2650
2651         entry_str_len = strlen(ttrace->entry_str);
2652         remaining_space = trace__entry_str_size - entry_str_len - 1; /* \0 */
2653         if (remaining_space <= 0)
2654                 goto out_put;
2655
2656         if (filename_len > (size_t)remaining_space) {
2657                 filename += filename_len - remaining_space;
2658                 filename_len = remaining_space;
2659         }
2660
2661         to_move = entry_str_len - ttrace->filename.entry_str_pos + 1; /* \0 */
2662         pos = ttrace->entry_str + ttrace->filename.entry_str_pos;
2663         memmove(pos + filename_len, pos, to_move);
2664         memcpy(pos, filename, filename_len);
2665
2666         ttrace->filename.ptr = 0;
2667         ttrace->filename.entry_str_pos = 0;
2668 out_put:
2669         thread__put(thread);
2670 out:
2671         return 0;
2672 }
2673
2674 static int trace__sched_stat_runtime(struct trace *trace, struct evsel *evsel,
2675                                      union perf_event *event __maybe_unused,
2676                                      struct perf_sample *sample)
2677 {
2678         u64 runtime = evsel__intval(evsel, sample, "runtime");
2679         double runtime_ms = (double)runtime / NSEC_PER_MSEC;
2680         struct thread *thread = machine__findnew_thread(trace->host,
2681                                                         sample->pid,
2682                                                         sample->tid);
2683         struct thread_trace *ttrace = thread__trace(thread, trace->output);
2684
2685         if (ttrace == NULL)
2686                 goto out_dump;
2687
2688         ttrace->runtime_ms += runtime_ms;
2689         trace->runtime_ms += runtime_ms;
2690 out_put:
2691         thread__put(thread);
2692         return 0;
2693
2694 out_dump:
2695         fprintf(trace->output, "%s: comm=%s,pid=%u,runtime=%" PRIu64 ",vruntime=%" PRIu64 ")\n",
2696                evsel->name,
2697                evsel__strval(evsel, sample, "comm"),
2698                (pid_t)evsel__intval(evsel, sample, "pid"),
2699                runtime,
2700                evsel__intval(evsel, sample, "vruntime"));
2701         goto out_put;
2702 }
2703
2704 static int bpf_output__printer(enum binary_printer_ops op,
2705                                unsigned int val, void *extra __maybe_unused, FILE *fp)
2706 {
2707         unsigned char ch = (unsigned char)val;
2708
2709         switch (op) {
2710         case BINARY_PRINT_CHAR_DATA:
2711                 return fprintf(fp, "%c", isprint(ch) ? ch : '.');
2712         case BINARY_PRINT_DATA_BEGIN:
2713         case BINARY_PRINT_LINE_BEGIN:
2714         case BINARY_PRINT_ADDR:
2715         case BINARY_PRINT_NUM_DATA:
2716         case BINARY_PRINT_NUM_PAD:
2717         case BINARY_PRINT_SEP:
2718         case BINARY_PRINT_CHAR_PAD:
2719         case BINARY_PRINT_LINE_END:
2720         case BINARY_PRINT_DATA_END:
2721         default:
2722                 break;
2723         }
2724
2725         return 0;
2726 }
2727
2728 static void bpf_output__fprintf(struct trace *trace,
2729                                 struct perf_sample *sample)
2730 {
2731         binary__fprintf(sample->raw_data, sample->raw_size, 8,
2732                         bpf_output__printer, NULL, trace->output);
2733         ++trace->nr_events_printed;
2734 }
2735
2736 static size_t trace__fprintf_tp_fields(struct trace *trace, struct evsel *evsel, struct perf_sample *sample,
2737                                        struct thread *thread, void *augmented_args, int augmented_args_size)
2738 {
2739         char bf[2048];
2740         size_t size = sizeof(bf);
2741         struct tep_format_field *field = evsel->tp_format->format.fields;
2742         struct syscall_arg_fmt *arg = __evsel__syscall_arg_fmt(evsel);
2743         size_t printed = 0;
2744         unsigned long val;
2745         u8 bit = 1;
2746         struct syscall_arg syscall_arg = {
2747                 .augmented = {
2748                         .size = augmented_args_size,
2749                         .args = augmented_args,
2750                 },
2751                 .idx    = 0,
2752                 .mask   = 0,
2753                 .trace  = trace,
2754                 .thread = thread,
2755                 .show_string_prefix = trace->show_string_prefix,
2756         };
2757
2758         for (; field && arg; field = field->next, ++syscall_arg.idx, bit <<= 1, ++arg) {
2759                 if (syscall_arg.mask & bit)
2760                         continue;
2761
2762                 syscall_arg.len = 0;
2763                 syscall_arg.fmt = arg;
2764                 if (field->flags & TEP_FIELD_IS_ARRAY) {
2765                         int offset = field->offset;
2766
2767                         if (field->flags & TEP_FIELD_IS_DYNAMIC) {
2768                                 offset = format_field__intval(field, sample, evsel->needs_swap);
2769                                 syscall_arg.len = offset >> 16;
2770                                 offset &= 0xffff;
2771                                 if (tep_field_is_relative(field->flags))
2772                                         offset += field->offset + field->size;
2773                         }
2774
2775                         val = (uintptr_t)(sample->raw_data + offset);
2776                 } else
2777                         val = format_field__intval(field, sample, evsel->needs_swap);
2778                 /*
2779                  * Some syscall args need some mask, most don't and
2780                  * return val untouched.
2781                  */
2782                 val = syscall_arg_fmt__mask_val(arg, &syscall_arg, val);
2783
2784                 /*
2785                  * Suppress this argument if its value is zero and
2786                  * we don't have a string associated in an
2787                  * strarray for it.
2788                  */
2789                 if (val == 0 &&
2790                     !trace->show_zeros &&
2791                     !((arg->show_zero ||
2792                        arg->scnprintf == SCA_STRARRAY ||
2793                        arg->scnprintf == SCA_STRARRAYS) &&
2794                       arg->parm))
2795                         continue;
2796
2797                 printed += scnprintf(bf + printed, size - printed, "%s", printed ? ", " : "");
2798
2799                 if (trace->show_arg_names)
2800                         printed += scnprintf(bf + printed, size - printed, "%s: ", field->name);
2801
2802                 printed += syscall_arg_fmt__scnprintf_val(arg, bf + printed, size - printed, &syscall_arg, val);
2803         }
2804
2805         return printed + fprintf(trace->output, "%s", bf);
2806 }
2807
2808 static int trace__event_handler(struct trace *trace, struct evsel *evsel,
2809                                 union perf_event *event __maybe_unused,
2810                                 struct perf_sample *sample)
2811 {
2812         struct thread *thread;
2813         int callchain_ret = 0;
2814         /*
2815          * Check if we called perf_evsel__disable(evsel) due to, for instance,
2816          * this event's max_events having been hit and this is an entry coming
2817          * from the ring buffer that we should discard, since the max events
2818          * have already been considered/printed.
2819          */
2820         if (evsel->disabled)
2821                 return 0;
2822
2823         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2824
2825         if (sample->callchain) {
2826                 struct callchain_cursor *cursor = get_tls_callchain_cursor();
2827
2828                 callchain_ret = trace__resolve_callchain(trace, evsel, sample, cursor);
2829                 if (callchain_ret == 0) {
2830                         if (cursor->nr < trace->min_stack)
2831                                 goto out;
2832                         callchain_ret = 1;
2833                 }
2834         }
2835
2836         trace__printf_interrupted_entry(trace);
2837         trace__fprintf_tstamp(trace, sample->time, trace->output);
2838
2839         if (trace->trace_syscalls && trace->show_duration)
2840                 fprintf(trace->output, "(         ): ");
2841
2842         if (thread)
2843                 trace__fprintf_comm_tid(trace, thread, trace->output);
2844
2845         if (evsel == trace->syscalls.events.bpf_output) {
2846                 int id = perf_evsel__sc_tp_uint(evsel, id, sample);
2847                 struct syscall *sc = trace__syscall_info(trace, evsel, id);
2848
2849                 if (sc) {
2850                         fprintf(trace->output, "%s(", sc->name);
2851                         trace__fprintf_sys_enter(trace, evsel, sample);
2852                         fputc(')', trace->output);
2853                         goto newline;
2854                 }
2855
2856                 /*
2857                  * XXX: Not having the associated syscall info or not finding/adding
2858                  *      the thread should never happen, but if it does...
2859                  *      fall thru and print it as a bpf_output event.
2860                  */
2861         }
2862
2863         fprintf(trace->output, "%s(", evsel->name);
2864
2865         if (evsel__is_bpf_output(evsel)) {
2866                 bpf_output__fprintf(trace, sample);
2867         } else if (evsel->tp_format) {
2868                 if (strncmp(evsel->tp_format->name, "sys_enter_", 10) ||
2869                     trace__fprintf_sys_enter(trace, evsel, sample)) {
2870                         if (trace->libtraceevent_print) {
2871                                 event_format__fprintf(evsel->tp_format, sample->cpu,
2872                                                       sample->raw_data, sample->raw_size,
2873                                                       trace->output);
2874                         } else {
2875                                 trace__fprintf_tp_fields(trace, evsel, sample, thread, NULL, 0);
2876                         }
2877                 }
2878         }
2879
2880 newline:
2881         fprintf(trace->output, ")\n");
2882
2883         if (callchain_ret > 0)
2884                 trace__fprintf_callchain(trace, sample);
2885         else if (callchain_ret < 0)
2886                 pr_err("Problem processing %s callchain, skipping...\n", evsel__name(evsel));
2887
2888         ++trace->nr_events_printed;
2889
2890         if (evsel->max_events != ULONG_MAX && ++evsel->nr_events_printed == evsel->max_events) {
2891                 evsel__disable(evsel);
2892                 evsel__close(evsel);
2893         }
2894 out:
2895         thread__put(thread);
2896         return 0;
2897 }
2898
2899 static void print_location(FILE *f, struct perf_sample *sample,
2900                            struct addr_location *al,
2901                            bool print_dso, bool print_sym)
2902 {
2903
2904         if ((verbose > 0 || print_dso) && al->map)
2905                 fprintf(f, "%s@", map__dso(al->map)->long_name);
2906
2907         if ((verbose > 0 || print_sym) && al->sym)
2908                 fprintf(f, "%s+0x%" PRIx64, al->sym->name,
2909                         al->addr - al->sym->start);
2910         else if (al->map)
2911                 fprintf(f, "0x%" PRIx64, al->addr);
2912         else
2913                 fprintf(f, "0x%" PRIx64, sample->addr);
2914 }
2915
2916 static int trace__pgfault(struct trace *trace,
2917                           struct evsel *evsel,
2918                           union perf_event *event __maybe_unused,
2919                           struct perf_sample *sample)
2920 {
2921         struct thread *thread;
2922         struct addr_location al;
2923         char map_type = 'd';
2924         struct thread_trace *ttrace;
2925         int err = -1;
2926         int callchain_ret = 0;
2927
2928         addr_location__init(&al);
2929         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2930
2931         if (sample->callchain) {
2932                 struct callchain_cursor *cursor = get_tls_callchain_cursor();
2933
2934                 callchain_ret = trace__resolve_callchain(trace, evsel, sample, cursor);
2935                 if (callchain_ret == 0) {
2936                         if (cursor->nr < trace->min_stack)
2937                                 goto out_put;
2938                         callchain_ret = 1;
2939                 }
2940         }
2941
2942         ttrace = thread__trace(thread, trace->output);
2943         if (ttrace == NULL)
2944                 goto out_put;
2945
2946         if (evsel->core.attr.config == PERF_COUNT_SW_PAGE_FAULTS_MAJ)
2947                 ttrace->pfmaj++;
2948         else
2949                 ttrace->pfmin++;
2950
2951         if (trace->summary_only)
2952                 goto out;
2953
2954         thread__find_symbol(thread, sample->cpumode, sample->ip, &al);
2955
2956         trace__fprintf_entry_head(trace, thread, 0, true, sample->time, trace->output);
2957
2958         fprintf(trace->output, "%sfault [",
2959                 evsel->core.attr.config == PERF_COUNT_SW_PAGE_FAULTS_MAJ ?
2960                 "maj" : "min");
2961
2962         print_location(trace->output, sample, &al, false, true);
2963
2964         fprintf(trace->output, "] => ");
2965
2966         thread__find_symbol(thread, sample->cpumode, sample->addr, &al);
2967
2968         if (!al.map) {
2969                 thread__find_symbol(thread, sample->cpumode, sample->addr, &al);
2970
2971                 if (al.map)
2972                         map_type = 'x';
2973                 else
2974                         map_type = '?';
2975         }
2976
2977         print_location(trace->output, sample, &al, true, false);
2978
2979         fprintf(trace->output, " (%c%c)\n", map_type, al.level);
2980
2981         if (callchain_ret > 0)
2982                 trace__fprintf_callchain(trace, sample);
2983         else if (callchain_ret < 0)
2984                 pr_err("Problem processing %s callchain, skipping...\n", evsel__name(evsel));
2985
2986         ++trace->nr_events_printed;
2987 out:
2988         err = 0;
2989 out_put:
2990         thread__put(thread);
2991         addr_location__exit(&al);
2992         return err;
2993 }
2994
2995 static void trace__set_base_time(struct trace *trace,
2996                                  struct evsel *evsel,
2997                                  struct perf_sample *sample)
2998 {
2999         /*
3000          * BPF events were not setting PERF_SAMPLE_TIME, so be more robust
3001          * and don't use sample->time unconditionally, we may end up having
3002          * some other event in the future without PERF_SAMPLE_TIME for good
3003          * reason, i.e. we may not be interested in its timestamps, just in
3004          * it taking place, picking some piece of information when it
3005          * appears in our event stream (vfs_getname comes to mind).
3006          */
3007         if (trace->base_time == 0 && !trace->full_time &&
3008             (evsel->core.attr.sample_type & PERF_SAMPLE_TIME))
3009                 trace->base_time = sample->time;
3010 }
3011
3012 static int trace__process_sample(struct perf_tool *tool,
3013                                  union perf_event *event,
3014                                  struct perf_sample *sample,
3015                                  struct evsel *evsel,
3016                                  struct machine *machine __maybe_unused)
3017 {
3018         struct trace *trace = container_of(tool, struct trace, tool);
3019         struct thread *thread;
3020         int err = 0;
3021
3022         tracepoint_handler handler = evsel->handler;
3023
3024         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
3025         if (thread && thread__is_filtered(thread))
3026                 goto out;
3027
3028         trace__set_base_time(trace, evsel, sample);
3029
3030         if (handler) {
3031                 ++trace->nr_events;
3032                 handler(trace, evsel, event, sample);
3033         }
3034 out:
3035         thread__put(thread);
3036         return err;
3037 }
3038
3039 static int trace__record(struct trace *trace, int argc, const char **argv)
3040 {
3041         unsigned int rec_argc, i, j;
3042         const char **rec_argv;
3043         const char * const record_args[] = {
3044                 "record",
3045                 "-R",
3046                 "-m", "1024",
3047                 "-c", "1",
3048         };
3049         pid_t pid = getpid();
3050         char *filter = asprintf__tp_filter_pids(1, &pid);
3051         const char * const sc_args[] = { "-e", };
3052         unsigned int sc_args_nr = ARRAY_SIZE(sc_args);
3053         const char * const majpf_args[] = { "-e", "major-faults" };
3054         unsigned int majpf_args_nr = ARRAY_SIZE(majpf_args);
3055         const char * const minpf_args[] = { "-e", "minor-faults" };
3056         unsigned int minpf_args_nr = ARRAY_SIZE(minpf_args);
3057         int err = -1;
3058
3059         /* +3 is for the event string below and the pid filter */
3060         rec_argc = ARRAY_SIZE(record_args) + sc_args_nr + 3 +
3061                 majpf_args_nr + minpf_args_nr + argc;
3062         rec_argv = calloc(rec_argc + 1, sizeof(char *));
3063
3064         if (rec_argv == NULL || filter == NULL)
3065                 goto out_free;
3066
3067         j = 0;
3068         for (i = 0; i < ARRAY_SIZE(record_args); i++)
3069                 rec_argv[j++] = record_args[i];
3070
3071         if (trace->trace_syscalls) {
3072                 for (i = 0; i < sc_args_nr; i++)
3073                         rec_argv[j++] = sc_args[i];
3074
3075                 /* event string may be different for older kernels - e.g., RHEL6 */
3076                 if (is_valid_tracepoint("raw_syscalls:sys_enter"))
3077                         rec_argv[j++] = "raw_syscalls:sys_enter,raw_syscalls:sys_exit";
3078                 else if (is_valid_tracepoint("syscalls:sys_enter"))
3079                         rec_argv[j++] = "syscalls:sys_enter,syscalls:sys_exit";
3080                 else {
3081                         pr_err("Neither raw_syscalls nor syscalls events exist.\n");
3082                         goto out_free;
3083                 }
3084         }
3085
3086         rec_argv[j++] = "--filter";
3087         rec_argv[j++] = filter;
3088
3089         if (trace->trace_pgfaults & TRACE_PFMAJ)
3090                 for (i = 0; i < majpf_args_nr; i++)
3091                         rec_argv[j++] = majpf_args[i];
3092
3093         if (trace->trace_pgfaults & TRACE_PFMIN)
3094                 for (i = 0; i < minpf_args_nr; i++)
3095                         rec_argv[j++] = minpf_args[i];
3096
3097         for (i = 0; i < (unsigned int)argc; i++)
3098                 rec_argv[j++] = argv[i];
3099
3100         err = cmd_record(j, rec_argv);
3101 out_free:
3102         free(filter);
3103         free(rec_argv);
3104         return err;
3105 }
3106
3107 static size_t trace__fprintf_thread_summary(struct trace *trace, FILE *fp);
3108
3109 static bool evlist__add_vfs_getname(struct evlist *evlist)
3110 {
3111         bool found = false;
3112         struct evsel *evsel, *tmp;
3113         struct parse_events_error err;
3114         int ret;
3115
3116         parse_events_error__init(&err);
3117         ret = parse_events(evlist, "probe:vfs_getname*", &err);
3118         parse_events_error__exit(&err);
3119         if (ret)
3120                 return false;
3121
3122         evlist__for_each_entry_safe(evlist, evsel, tmp) {
3123                 if (!strstarts(evsel__name(evsel), "probe:vfs_getname"))
3124                         continue;
3125
3126                 if (evsel__field(evsel, "pathname")) {
3127                         evsel->handler = trace__vfs_getname;
3128                         found = true;
3129                         continue;
3130                 }
3131
3132                 list_del_init(&evsel->core.node);
3133                 evsel->evlist = NULL;
3134                 evsel__delete(evsel);
3135         }
3136
3137         return found;
3138 }
3139
3140 static struct evsel *evsel__new_pgfault(u64 config)
3141 {
3142         struct evsel *evsel;
3143         struct perf_event_attr attr = {
3144                 .type = PERF_TYPE_SOFTWARE,
3145                 .mmap_data = 1,
3146         };
3147
3148         attr.config = config;
3149         attr.sample_period = 1;
3150
3151         event_attr_init(&attr);
3152
3153         evsel = evsel__new(&attr);
3154         if (evsel)
3155                 evsel->handler = trace__pgfault;
3156
3157         return evsel;
3158 }
3159
3160 static void evlist__free_syscall_tp_fields(struct evlist *evlist)
3161 {
3162         struct evsel *evsel;
3163
3164         evlist__for_each_entry(evlist, evsel) {
3165                 evsel_trace__delete(evsel->priv);
3166                 evsel->priv = NULL;
3167         }
3168 }
3169
3170 static void trace__handle_event(struct trace *trace, union perf_event *event, struct perf_sample *sample)
3171 {
3172         const u32 type = event->header.type;
3173         struct evsel *evsel;
3174
3175         if (type != PERF_RECORD_SAMPLE) {
3176                 trace__process_event(trace, trace->host, event, sample);
3177                 return;
3178         }
3179
3180         evsel = evlist__id2evsel(trace->evlist, sample->id);
3181         if (evsel == NULL) {
3182                 fprintf(trace->output, "Unknown tp ID %" PRIu64 ", skipping...\n", sample->id);
3183                 return;
3184         }
3185
3186         if (evswitch__discard(&trace->evswitch, evsel))
3187                 return;
3188
3189         trace__set_base_time(trace, evsel, sample);
3190
3191         if (evsel->core.attr.type == PERF_TYPE_TRACEPOINT &&
3192             sample->raw_data == NULL) {
3193                 fprintf(trace->output, "%s sample with no payload for tid: %d, cpu %d, raw_size=%d, skipping...\n",
3194                        evsel__name(evsel), sample->tid,
3195                        sample->cpu, sample->raw_size);
3196         } else {
3197                 tracepoint_handler handler = evsel->handler;
3198                 handler(trace, evsel, event, sample);
3199         }
3200
3201         if (trace->nr_events_printed >= trace->max_events && trace->max_events != ULONG_MAX)
3202                 interrupted = true;
3203 }
3204
3205 static int trace__add_syscall_newtp(struct trace *trace)
3206 {
3207         int ret = -1;
3208         struct evlist *evlist = trace->evlist;
3209         struct evsel *sys_enter, *sys_exit;
3210
3211         sys_enter = perf_evsel__raw_syscall_newtp("sys_enter", trace__sys_enter);
3212         if (sys_enter == NULL)
3213                 goto out;
3214
3215         if (perf_evsel__init_sc_tp_ptr_field(sys_enter, args))
3216                 goto out_delete_sys_enter;
3217
3218         sys_exit = perf_evsel__raw_syscall_newtp("sys_exit", trace__sys_exit);
3219         if (sys_exit == NULL)
3220                 goto out_delete_sys_enter;
3221
3222         if (perf_evsel__init_sc_tp_uint_field(sys_exit, ret))
3223                 goto out_delete_sys_exit;
3224
3225         evsel__config_callchain(sys_enter, &trace->opts, &callchain_param);
3226         evsel__config_callchain(sys_exit, &trace->opts, &callchain_param);
3227
3228         evlist__add(evlist, sys_enter);
3229         evlist__add(evlist, sys_exit);
3230
3231         if (callchain_param.enabled && !trace->kernel_syscallchains) {
3232                 /*
3233                  * We're interested only in the user space callchain
3234                  * leading to the syscall, allow overriding that for
3235                  * debugging reasons using --kernel_syscall_callchains
3236                  */
3237                 sys_exit->core.attr.exclude_callchain_kernel = 1;
3238         }
3239
3240         trace->syscalls.events.sys_enter = sys_enter;
3241         trace->syscalls.events.sys_exit  = sys_exit;
3242
3243         ret = 0;
3244 out:
3245         return ret;
3246
3247 out_delete_sys_exit:
3248         evsel__delete_priv(sys_exit);
3249 out_delete_sys_enter:
3250         evsel__delete_priv(sys_enter);
3251         goto out;
3252 }
3253
3254 static int trace__set_ev_qualifier_tp_filter(struct trace *trace)
3255 {
3256         int err = -1;
3257         struct evsel *sys_exit;
3258         char *filter = asprintf_expr_inout_ints("id", !trace->not_ev_qualifier,
3259                                                 trace->ev_qualifier_ids.nr,
3260                                                 trace->ev_qualifier_ids.entries);
3261
3262         if (filter == NULL)
3263                 goto out_enomem;
3264
3265         if (!evsel__append_tp_filter(trace->syscalls.events.sys_enter, filter)) {
3266                 sys_exit = trace->syscalls.events.sys_exit;
3267                 err = evsel__append_tp_filter(sys_exit, filter);
3268         }
3269
3270         free(filter);
3271 out:
3272         return err;
3273 out_enomem:
3274         errno = ENOMEM;
3275         goto out;
3276 }
3277
3278 #ifdef HAVE_BPF_SKEL
3279 static struct bpf_program *trace__find_bpf_program_by_title(struct trace *trace, const char *name)
3280 {
3281         struct bpf_program *pos, *prog = NULL;
3282         const char *sec_name;
3283
3284         if (trace->skel->obj == NULL)
3285                 return NULL;
3286
3287         bpf_object__for_each_program(pos, trace->skel->obj) {
3288                 sec_name = bpf_program__section_name(pos);
3289                 if (sec_name && !strcmp(sec_name, name)) {
3290                         prog = pos;
3291                         break;
3292                 }
3293         }
3294
3295         return prog;
3296 }
3297
3298 static struct bpf_program *trace__find_syscall_bpf_prog(struct trace *trace, struct syscall *sc,
3299                                                         const char *prog_name, const char *type)
3300 {
3301         struct bpf_program *prog;
3302
3303         if (prog_name == NULL) {
3304                 char default_prog_name[256];
3305                 scnprintf(default_prog_name, sizeof(default_prog_name), "tp/syscalls/sys_%s_%s", type, sc->name);
3306                 prog = trace__find_bpf_program_by_title(trace, default_prog_name);
3307                 if (prog != NULL)
3308                         goto out_found;
3309                 if (sc->fmt && sc->fmt->alias) {
3310                         scnprintf(default_prog_name, sizeof(default_prog_name), "tp/syscalls/sys_%s_%s", type, sc->fmt->alias);
3311                         prog = trace__find_bpf_program_by_title(trace, default_prog_name);
3312                         if (prog != NULL)
3313                                 goto out_found;
3314                 }
3315                 goto out_unaugmented;
3316         }
3317
3318         prog = trace__find_bpf_program_by_title(trace, prog_name);
3319
3320         if (prog != NULL) {
3321 out_found:
3322                 return prog;
3323         }
3324
3325         pr_debug("Couldn't find BPF prog \"%s\" to associate with syscalls:sys_%s_%s, not augmenting it\n",
3326                  prog_name, type, sc->name);
3327 out_unaugmented:
3328         return trace->skel->progs.syscall_unaugmented;
3329 }
3330
3331 static void trace__init_syscall_bpf_progs(struct trace *trace, int id)
3332 {
3333         struct syscall *sc = trace__syscall_info(trace, NULL, id);
3334
3335         if (sc == NULL)
3336                 return;
3337
3338         sc->bpf_prog.sys_enter = trace__find_syscall_bpf_prog(trace, sc, sc->fmt ? sc->fmt->bpf_prog_name.sys_enter : NULL, "enter");
3339         sc->bpf_prog.sys_exit  = trace__find_syscall_bpf_prog(trace, sc, sc->fmt ? sc->fmt->bpf_prog_name.sys_exit  : NULL,  "exit");
3340 }
3341
3342 static int trace__bpf_prog_sys_enter_fd(struct trace *trace, int id)
3343 {
3344         struct syscall *sc = trace__syscall_info(trace, NULL, id);
3345         return sc ? bpf_program__fd(sc->bpf_prog.sys_enter) : bpf_program__fd(trace->skel->progs.syscall_unaugmented);
3346 }
3347
3348 static int trace__bpf_prog_sys_exit_fd(struct trace *trace, int id)
3349 {
3350         struct syscall *sc = trace__syscall_info(trace, NULL, id);
3351         return sc ? bpf_program__fd(sc->bpf_prog.sys_exit) : bpf_program__fd(trace->skel->progs.syscall_unaugmented);
3352 }
3353
3354 static struct bpf_program *trace__find_usable_bpf_prog_entry(struct trace *trace, struct syscall *sc)
3355 {
3356         struct tep_format_field *field, *candidate_field;
3357         int id;
3358
3359         /*
3360          * We're only interested in syscalls that have a pointer:
3361          */
3362         for (field = sc->args; field; field = field->next) {
3363                 if (field->flags & TEP_FIELD_IS_POINTER)
3364                         goto try_to_find_pair;
3365         }
3366
3367         return NULL;
3368
3369 try_to_find_pair:
3370         for (id = 0; id < trace->sctbl->syscalls.nr_entries; ++id) {
3371                 struct syscall *pair = trace__syscall_info(trace, NULL, id);
3372                 struct bpf_program *pair_prog;
3373                 bool is_candidate = false;
3374
3375                 if (pair == NULL || pair == sc ||
3376                     pair->bpf_prog.sys_enter == trace->skel->progs.syscall_unaugmented)
3377                         continue;
3378
3379                 for (field = sc->args, candidate_field = pair->args;
3380                      field && candidate_field; field = field->next, candidate_field = candidate_field->next) {
3381                         bool is_pointer = field->flags & TEP_FIELD_IS_POINTER,
3382                              candidate_is_pointer = candidate_field->flags & TEP_FIELD_IS_POINTER;
3383
3384                         if (is_pointer) {
3385                                if (!candidate_is_pointer) {
3386                                         // The candidate just doesn't copies our pointer arg, might copy other pointers we want.
3387                                         continue;
3388                                }
3389                         } else {
3390                                 if (candidate_is_pointer) {
3391                                         // The candidate might copy a pointer we don't have, skip it.
3392                                         goto next_candidate;
3393                                 }
3394                                 continue;
3395                         }
3396
3397                         if (strcmp(field->type, candidate_field->type))
3398                                 goto next_candidate;
3399
3400                         /*
3401                          * This is limited in the BPF program but sys_write
3402                          * uses "const char *" for its "buf" arg so we need to
3403                          * use some heuristic that is kinda future proof...
3404                          */
3405                         if (strcmp(field->type, "const char *") == 0 &&
3406                             !(strstr(field->name, "name") ||
3407                               strstr(field->name, "path") ||
3408                               strstr(field->name, "file") ||
3409                               strstr(field->name, "root") ||
3410                               strstr(field->name, "description")))
3411                                 goto next_candidate;
3412
3413                         is_candidate = true;
3414                 }
3415
3416                 if (!is_candidate)
3417                         goto next_candidate;
3418
3419                 /*
3420                  * Check if the tentative pair syscall augmenter has more pointers, if it has,
3421                  * then it may be collecting that and we then can't use it, as it would collect
3422                  * more than what is common to the two syscalls.
3423                  */
3424                 if (candidate_field) {
3425                         for (candidate_field = candidate_field->next; candidate_field; candidate_field = candidate_field->next)
3426                                 if (candidate_field->flags & TEP_FIELD_IS_POINTER)
3427                                         goto next_candidate;
3428                 }
3429
3430                 pair_prog = pair->bpf_prog.sys_enter;
3431                 /*
3432                  * If the pair isn't enabled, then its bpf_prog.sys_enter will not
3433                  * have been searched for, so search it here and if it returns the
3434                  * unaugmented one, then ignore it, otherwise we'll reuse that BPF
3435                  * program for a filtered syscall on a non-filtered one.
3436                  *
3437                  * For instance, we have "!syscalls:sys_enter_renameat" and that is
3438                  * useful for "renameat2".
3439                  */
3440                 if (pair_prog == NULL) {
3441                         pair_prog = trace__find_syscall_bpf_prog(trace, pair, pair->fmt ? pair->fmt->bpf_prog_name.sys_enter : NULL, "enter");
3442                         if (pair_prog == trace->skel->progs.syscall_unaugmented)
3443                                 goto next_candidate;
3444                 }
3445
3446                 pr_debug("Reusing \"%s\" BPF sys_enter augmenter for \"%s\"\n", pair->name, sc->name);
3447                 return pair_prog;
3448         next_candidate:
3449                 continue;
3450         }
3451
3452         return NULL;
3453 }
3454
3455 static int trace__init_syscalls_bpf_prog_array_maps(struct trace *trace)
3456 {
3457         int map_enter_fd = bpf_map__fd(trace->skel->maps.syscalls_sys_enter);
3458         int map_exit_fd  = bpf_map__fd(trace->skel->maps.syscalls_sys_exit);
3459         int err = 0, key;
3460
3461         for (key = 0; key < trace->sctbl->syscalls.nr_entries; ++key) {
3462                 int prog_fd;
3463
3464                 if (!trace__syscall_enabled(trace, key))
3465                         continue;
3466
3467                 trace__init_syscall_bpf_progs(trace, key);
3468
3469                 // It'll get at least the "!raw_syscalls:unaugmented"
3470                 prog_fd = trace__bpf_prog_sys_enter_fd(trace, key);
3471                 err = bpf_map_update_elem(map_enter_fd, &key, &prog_fd, BPF_ANY);
3472                 if (err)
3473                         break;
3474                 prog_fd = trace__bpf_prog_sys_exit_fd(trace, key);
3475                 err = bpf_map_update_elem(map_exit_fd, &key, &prog_fd, BPF_ANY);
3476                 if (err)
3477                         break;
3478         }
3479
3480         /*
3481          * Now lets do a second pass looking for enabled syscalls without
3482          * an augmenter that have a signature that is a superset of another
3483          * syscall with an augmenter so that we can auto-reuse it.
3484          *
3485          * I.e. if we have an augmenter for the "open" syscall that has
3486          * this signature:
3487          *
3488          *   int open(const char *pathname, int flags, mode_t mode);
3489          *
3490          * I.e. that will collect just the first string argument, then we
3491          * can reuse it for the 'creat' syscall, that has this signature:
3492          *
3493          *   int creat(const char *pathname, mode_t mode);
3494          *
3495          * and for:
3496          *
3497          *   int stat(const char *pathname, struct stat *statbuf);
3498          *   int lstat(const char *pathname, struct stat *statbuf);
3499          *
3500          * Because the 'open' augmenter will collect the first arg as a string,
3501          * and leave alone all the other args, which already helps with
3502          * beautifying 'stat' and 'lstat''s pathname arg.
3503          *
3504          * Then, in time, when 'stat' gets an augmenter that collects both
3505          * first and second arg (this one on the raw_syscalls:sys_exit prog
3506          * array tail call, then that one will be used.
3507          */
3508         for (key = 0; key < trace->sctbl->syscalls.nr_entries; ++key) {
3509                 struct syscall *sc = trace__syscall_info(trace, NULL, key);
3510                 struct bpf_program *pair_prog;
3511                 int prog_fd;
3512
3513                 if (sc == NULL || sc->bpf_prog.sys_enter == NULL)
3514                         continue;
3515
3516                 /*
3517                  * For now we're just reusing the sys_enter prog, and if it
3518                  * already has an augmenter, we don't need to find one.
3519                  */
3520                 if (sc->bpf_prog.sys_enter != trace->skel->progs.syscall_unaugmented)
3521                         continue;
3522
3523                 /*
3524                  * Look at all the other syscalls for one that has a signature
3525                  * that is close enough that we can share:
3526                  */
3527                 pair_prog = trace__find_usable_bpf_prog_entry(trace, sc);
3528                 if (pair_prog == NULL)
3529                         continue;
3530
3531                 sc->bpf_prog.sys_enter = pair_prog;
3532
3533                 /*
3534                  * Update the BPF_MAP_TYPE_PROG_SHARED for raw_syscalls:sys_enter
3535                  * with the fd for the program we're reusing:
3536                  */
3537                 prog_fd = bpf_program__fd(sc->bpf_prog.sys_enter);
3538                 err = bpf_map_update_elem(map_enter_fd, &key, &prog_fd, BPF_ANY);
3539                 if (err)
3540                         break;
3541         }
3542
3543         return err;
3544 }
3545 #endif // HAVE_BPF_SKEL
3546
3547 static int trace__set_ev_qualifier_filter(struct trace *trace)
3548 {
3549         if (trace->syscalls.events.sys_enter)
3550                 return trace__set_ev_qualifier_tp_filter(trace);
3551         return 0;
3552 }
3553
3554 static int bpf_map__set_filter_pids(struct bpf_map *map __maybe_unused,
3555                                     size_t npids __maybe_unused, pid_t *pids __maybe_unused)
3556 {
3557         int err = 0;
3558 #ifdef HAVE_LIBBPF_SUPPORT
3559         bool value = true;
3560         int map_fd = bpf_map__fd(map);
3561         size_t i;
3562
3563         for (i = 0; i < npids; ++i) {
3564                 err = bpf_map_update_elem(map_fd, &pids[i], &value, BPF_ANY);
3565                 if (err)
3566                         break;
3567         }
3568 #endif
3569         return err;
3570 }
3571
3572 static int trace__set_filter_loop_pids(struct trace *trace)
3573 {
3574         unsigned int nr = 1, err;
3575         pid_t pids[32] = {
3576                 getpid(),
3577         };
3578         struct thread *thread = machine__find_thread(trace->host, pids[0], pids[0]);
3579
3580         while (thread && nr < ARRAY_SIZE(pids)) {
3581                 struct thread *parent = machine__find_thread(trace->host,
3582                                                              thread__ppid(thread),
3583                                                              thread__ppid(thread));
3584
3585                 if (parent == NULL)
3586                         break;
3587
3588                 if (!strcmp(thread__comm_str(parent), "sshd") ||
3589                     strstarts(thread__comm_str(parent), "gnome-terminal")) {
3590                         pids[nr++] = thread__tid(parent);
3591                         break;
3592                 }
3593                 thread = parent;
3594         }
3595
3596         err = evlist__append_tp_filter_pids(trace->evlist, nr, pids);
3597         if (!err && trace->filter_pids.map)
3598                 err = bpf_map__set_filter_pids(trace->filter_pids.map, nr, pids);
3599
3600         return err;
3601 }
3602
3603 static int trace__set_filter_pids(struct trace *trace)
3604 {
3605         int err = 0;
3606         /*
3607          * Better not use !target__has_task() here because we need to cover the
3608          * case where no threads were specified in the command line, but a
3609          * workload was, and in that case we will fill in the thread_map when
3610          * we fork the workload in evlist__prepare_workload.
3611          */
3612         if (trace->filter_pids.nr > 0) {
3613                 err = evlist__append_tp_filter_pids(trace->evlist, trace->filter_pids.nr,
3614                                                     trace->filter_pids.entries);
3615                 if (!err && trace->filter_pids.map) {
3616                         err = bpf_map__set_filter_pids(trace->filter_pids.map, trace->filter_pids.nr,
3617                                                        trace->filter_pids.entries);
3618                 }
3619         } else if (perf_thread_map__pid(trace->evlist->core.threads, 0) == -1) {
3620                 err = trace__set_filter_loop_pids(trace);
3621         }
3622
3623         return err;
3624 }
3625
3626 static int __trace__deliver_event(struct trace *trace, union perf_event *event)
3627 {
3628         struct evlist *evlist = trace->evlist;
3629         struct perf_sample sample;
3630         int err = evlist__parse_sample(evlist, event, &sample);
3631
3632         if (err)
3633                 fprintf(trace->output, "Can't parse sample, err = %d, skipping...\n", err);
3634         else
3635                 trace__handle_event(trace, event, &sample);
3636
3637         return 0;
3638 }
3639
3640 static int __trace__flush_events(struct trace *trace)
3641 {
3642         u64 first = ordered_events__first_time(&trace->oe.data);
3643         u64 flush = trace->oe.last - NSEC_PER_SEC;
3644
3645         /* Is there some thing to flush.. */
3646         if (first && first < flush)
3647                 return ordered_events__flush_time(&trace->oe.data, flush);
3648
3649         return 0;
3650 }
3651
3652 static int trace__flush_events(struct trace *trace)
3653 {
3654         return !trace->sort_events ? 0 : __trace__flush_events(trace);
3655 }
3656
3657 static int trace__deliver_event(struct trace *trace, union perf_event *event)
3658 {
3659         int err;
3660
3661         if (!trace->sort_events)
3662                 return __trace__deliver_event(trace, event);
3663
3664         err = evlist__parse_sample_timestamp(trace->evlist, event, &trace->oe.last);
3665         if (err && err != -1)
3666                 return err;
3667
3668         err = ordered_events__queue(&trace->oe.data, event, trace->oe.last, 0, NULL);
3669         if (err)
3670                 return err;
3671
3672         return trace__flush_events(trace);
3673 }
3674
3675 static int ordered_events__deliver_event(struct ordered_events *oe,
3676                                          struct ordered_event *event)
3677 {
3678         struct trace *trace = container_of(oe, struct trace, oe.data);
3679
3680         return __trace__deliver_event(trace, event->event);
3681 }
3682
3683 static struct syscall_arg_fmt *evsel__find_syscall_arg_fmt_by_name(struct evsel *evsel, char *arg)
3684 {
3685         struct tep_format_field *field;
3686         struct syscall_arg_fmt *fmt = __evsel__syscall_arg_fmt(evsel);
3687
3688         if (evsel->tp_format == NULL || fmt == NULL)
3689                 return NULL;
3690
3691         for (field = evsel->tp_format->format.fields; field; field = field->next, ++fmt)
3692                 if (strcmp(field->name, arg) == 0)
3693                         return fmt;
3694
3695         return NULL;
3696 }
3697
3698 static int trace__expand_filter(struct trace *trace __maybe_unused, struct evsel *evsel)
3699 {
3700         char *tok, *left = evsel->filter, *new_filter = evsel->filter;
3701
3702         while ((tok = strpbrk(left, "=<>!")) != NULL) {
3703                 char *right = tok + 1, *right_end;
3704
3705                 if (*right == '=')
3706                         ++right;
3707
3708                 while (isspace(*right))
3709                         ++right;
3710
3711                 if (*right == '\0')
3712                         break;
3713
3714                 while (!isalpha(*left))
3715                         if (++left == tok) {
3716                                 /*
3717                                  * Bail out, can't find the name of the argument that is being
3718                                  * used in the filter, let it try to set this filter, will fail later.
3719                                  */
3720                                 return 0;
3721                         }
3722
3723                 right_end = right + 1;
3724                 while (isalnum(*right_end) || *right_end == '_' || *right_end == '|')
3725                         ++right_end;
3726
3727                 if (isalpha(*right)) {
3728                         struct syscall_arg_fmt *fmt;
3729                         int left_size = tok - left,
3730                             right_size = right_end - right;
3731                         char arg[128];
3732
3733                         while (isspace(left[left_size - 1]))
3734                                 --left_size;
3735
3736                         scnprintf(arg, sizeof(arg), "%.*s", left_size, left);
3737
3738                         fmt = evsel__find_syscall_arg_fmt_by_name(evsel, arg);
3739                         if (fmt == NULL) {
3740                                 pr_err("\"%s\" not found in \"%s\", can't set filter \"%s\"\n",
3741                                        arg, evsel->name, evsel->filter);
3742                                 return -1;
3743                         }
3744
3745                         pr_debug2("trying to expand \"%s\" \"%.*s\" \"%.*s\" -> ",
3746                                  arg, (int)(right - tok), tok, right_size, right);
3747
3748                         if (fmt->strtoul) {
3749                                 u64 val;
3750                                 struct syscall_arg syscall_arg = {
3751                                         .parm = fmt->parm,
3752                                 };
3753
3754                                 if (fmt->strtoul(right, right_size, &syscall_arg, &val)) {
3755                                         char *n, expansion[19];
3756                                         int expansion_lenght = scnprintf(expansion, sizeof(expansion), "%#" PRIx64, val);
3757                                         int expansion_offset = right - new_filter;
3758
3759                                         pr_debug("%s", expansion);
3760
3761                                         if (asprintf(&n, "%.*s%s%s", expansion_offset, new_filter, expansion, right_end) < 0) {
3762                                                 pr_debug(" out of memory!\n");
3763                                                 free(new_filter);
3764                                                 return -1;
3765                                         }
3766                                         if (new_filter != evsel->filter)
3767                                                 free(new_filter);
3768                                         left = n + expansion_offset + expansion_lenght;
3769                                         new_filter = n;
3770                                 } else {
3771                                         pr_err("\"%.*s\" not found for \"%s\" in \"%s\", can't set filter \"%s\"\n",
3772                                                right_size, right, arg, evsel->name, evsel->filter);
3773                                         return -1;
3774                                 }
3775                         } else {
3776                                 pr_err("No resolver (strtoul) for \"%s\" in \"%s\", can't set filter \"%s\"\n",
3777                                        arg, evsel->name, evsel->filter);
3778                                 return -1;
3779                         }
3780
3781                         pr_debug("\n");
3782                 } else {
3783                         left = right_end;
3784                 }
3785         }
3786
3787         if (new_filter != evsel->filter) {
3788                 pr_debug("New filter for %s: %s\n", evsel->name, new_filter);
3789                 evsel__set_filter(evsel, new_filter);
3790                 free(new_filter);
3791         }
3792
3793         return 0;
3794 }
3795
3796 static int trace__expand_filters(struct trace *trace, struct evsel **err_evsel)
3797 {
3798         struct evlist *evlist = trace->evlist;
3799         struct evsel *evsel;
3800
3801         evlist__for_each_entry(evlist, evsel) {
3802                 if (evsel->filter == NULL)
3803                         continue;
3804
3805                 if (trace__expand_filter(trace, evsel)) {
3806                         *err_evsel = evsel;
3807                         return -1;
3808                 }
3809         }
3810
3811         return 0;
3812 }
3813
3814 static int trace__run(struct trace *trace, int argc, const char **argv)
3815 {
3816         struct evlist *evlist = trace->evlist;
3817         struct evsel *evsel, *pgfault_maj = NULL, *pgfault_min = NULL;
3818         int err = -1, i;
3819         unsigned long before;
3820         const bool forks = argc > 0;
3821         bool draining = false;
3822
3823         trace->live = true;
3824
3825         if (!trace->raw_augmented_syscalls) {
3826                 if (trace->trace_syscalls && trace__add_syscall_newtp(trace))
3827                         goto out_error_raw_syscalls;
3828
3829                 if (trace->trace_syscalls)
3830                         trace->vfs_getname = evlist__add_vfs_getname(evlist);
3831         }
3832
3833         if ((trace->trace_pgfaults & TRACE_PFMAJ)) {
3834                 pgfault_maj = evsel__new_pgfault(PERF_COUNT_SW_PAGE_FAULTS_MAJ);
3835                 if (pgfault_maj == NULL)
3836                         goto out_error_mem;
3837                 evsel__config_callchain(pgfault_maj, &trace->opts, &callchain_param);
3838                 evlist__add(evlist, pgfault_maj);
3839         }
3840
3841         if ((trace->trace_pgfaults & TRACE_PFMIN)) {
3842                 pgfault_min = evsel__new_pgfault(PERF_COUNT_SW_PAGE_FAULTS_MIN);
3843                 if (pgfault_min == NULL)
3844                         goto out_error_mem;
3845                 evsel__config_callchain(pgfault_min, &trace->opts, &callchain_param);
3846                 evlist__add(evlist, pgfault_min);
3847         }
3848
3849         /* Enable ignoring missing threads when -u/-p option is defined. */
3850         trace->opts.ignore_missing_thread = trace->opts.target.uid != UINT_MAX || trace->opts.target.pid;
3851
3852         if (trace->sched &&
3853             evlist__add_newtp(evlist, "sched", "sched_stat_runtime", trace__sched_stat_runtime))
3854                 goto out_error_sched_stat_runtime;
3855         /*
3856          * If a global cgroup was set, apply it to all the events without an
3857          * explicit cgroup. I.e.:
3858          *
3859          *      trace -G A -e sched:*switch
3860          *
3861          * Will set all raw_syscalls:sys_{enter,exit}, pgfault, vfs_getname, etc
3862          * _and_ sched:sched_switch to the 'A' cgroup, while:
3863          *
3864          * trace -e sched:*switch -G A
3865          *
3866          * will only set the sched:sched_switch event to the 'A' cgroup, all the
3867          * other events (raw_syscalls:sys_{enter,exit}, etc are left "without"
3868          * a cgroup (on the root cgroup, sys wide, etc).
3869          *
3870          * Multiple cgroups:
3871          *
3872          * trace -G A -e sched:*switch -G B
3873          *
3874          * the syscall ones go to the 'A' cgroup, the sched:sched_switch goes
3875          * to the 'B' cgroup.
3876          *
3877          * evlist__set_default_cgroup() grabs a reference of the passed cgroup
3878          * only for the evsels still without a cgroup, i.e. evsel->cgroup == NULL.
3879          */
3880         if (trace->cgroup)
3881                 evlist__set_default_cgroup(trace->evlist, trace->cgroup);
3882
3883         err = evlist__create_maps(evlist, &trace->opts.target);
3884         if (err < 0) {
3885                 fprintf(trace->output, "Problems parsing the target to trace, check your options!\n");
3886                 goto out_delete_evlist;
3887         }
3888
3889         err = trace__symbols_init(trace, evlist);
3890         if (err < 0) {
3891                 fprintf(trace->output, "Problems initializing symbol libraries!\n");
3892                 goto out_delete_evlist;
3893         }
3894
3895         evlist__config(evlist, &trace->opts, &callchain_param);
3896
3897         if (forks) {
3898                 err = evlist__prepare_workload(evlist, &trace->opts.target, argv, false, NULL);
3899                 if (err < 0) {
3900                         fprintf(trace->output, "Couldn't run the workload!\n");
3901                         goto out_delete_evlist;
3902                 }
3903                 workload_pid = evlist->workload.pid;
3904         }
3905
3906         err = evlist__open(evlist);
3907         if (err < 0)
3908                 goto out_error_open;
3909 #ifdef HAVE_BPF_SKEL
3910         if (trace->syscalls.events.bpf_output) {
3911                 struct perf_cpu cpu;
3912
3913                 /*
3914                  * Set up the __augmented_syscalls__ BPF map to hold for each
3915                  * CPU the bpf-output event's file descriptor.
3916                  */
3917                 perf_cpu_map__for_each_cpu(cpu, i, trace->syscalls.events.bpf_output->core.cpus) {
3918                         bpf_map__update_elem(trace->skel->maps.__augmented_syscalls__,
3919                                         &cpu.cpu, sizeof(int),
3920                                         xyarray__entry(trace->syscalls.events.bpf_output->core.fd,
3921                                                        cpu.cpu, 0),
3922                                         sizeof(__u32), BPF_ANY);
3923                 }
3924         }
3925 #endif
3926         err = trace__set_filter_pids(trace);
3927         if (err < 0)
3928                 goto out_error_mem;
3929
3930 #ifdef HAVE_BPF_SKEL
3931         if (trace->skel && trace->skel->progs.sys_enter)
3932                 trace__init_syscalls_bpf_prog_array_maps(trace);
3933 #endif
3934
3935         if (trace->ev_qualifier_ids.nr > 0) {
3936                 err = trace__set_ev_qualifier_filter(trace);
3937                 if (err < 0)
3938                         goto out_errno;
3939
3940                 if (trace->syscalls.events.sys_exit) {
3941                         pr_debug("event qualifier tracepoint filter: %s\n",
3942                                  trace->syscalls.events.sys_exit->filter);
3943                 }
3944         }
3945
3946         /*
3947          * If the "close" syscall is not traced, then we will not have the
3948          * opportunity to, in syscall_arg__scnprintf_close_fd() invalidate the
3949          * fd->pathname table and were ending up showing the last value set by
3950          * syscalls opening a pathname and associating it with a descriptor or
3951          * reading it from /proc/pid/fd/ in cases where that doesn't make
3952          * sense.
3953          *
3954          *  So just disable this beautifier (SCA_FD, SCA_FDAT) when 'close' is
3955          *  not in use.
3956          */
3957         trace->fd_path_disabled = !trace__syscall_enabled(trace, syscalltbl__id(trace->sctbl, "close"));
3958
3959         err = trace__expand_filters(trace, &evsel);
3960         if (err)
3961                 goto out_delete_evlist;
3962         err = evlist__apply_filters(evlist, &evsel);
3963         if (err < 0)
3964                 goto out_error_apply_filters;
3965
3966         err = evlist__mmap(evlist, trace->opts.mmap_pages);
3967         if (err < 0)
3968                 goto out_error_mmap;
3969
3970         if (!target__none(&trace->opts.target) && !trace->opts.target.initial_delay)
3971                 evlist__enable(evlist);
3972
3973         if (forks)
3974                 evlist__start_workload(evlist);
3975
3976         if (trace->opts.target.initial_delay) {
3977                 usleep(trace->opts.target.initial_delay * 1000);
3978                 evlist__enable(evlist);
3979         }
3980
3981         trace->multiple_threads = perf_thread_map__pid(evlist->core.threads, 0) == -1 ||
3982                 perf_thread_map__nr(evlist->core.threads) > 1 ||
3983                 evlist__first(evlist)->core.attr.inherit;
3984
3985         /*
3986          * Now that we already used evsel->core.attr to ask the kernel to setup the
3987          * events, lets reuse evsel->core.attr.sample_max_stack as the limit in
3988          * trace__resolve_callchain(), allowing per-event max-stack settings
3989          * to override an explicitly set --max-stack global setting.
3990          */
3991         evlist__for_each_entry(evlist, evsel) {
3992                 if (evsel__has_callchain(evsel) &&
3993                     evsel->core.attr.sample_max_stack == 0)
3994                         evsel->core.attr.sample_max_stack = trace->max_stack;
3995         }
3996 again:
3997         before = trace->nr_events;
3998
3999         for (i = 0; i < evlist->core.nr_mmaps; i++) {
4000                 union perf_event *event;
4001                 struct mmap *md;
4002
4003                 md = &evlist->mmap[i];
4004                 if (perf_mmap__read_init(&md->core) < 0)
4005                         continue;
4006
4007                 while ((event = perf_mmap__read_event(&md->core)) != NULL) {
4008                         ++trace->nr_events;
4009
4010                         err = trace__deliver_event(trace, event);
4011                         if (err)
4012                                 goto out_disable;
4013
4014                         perf_mmap__consume(&md->core);
4015
4016                         if (interrupted)
4017                                 goto out_disable;
4018
4019                         if (done && !draining) {
4020                                 evlist__disable(evlist);
4021                                 draining = true;
4022                         }
4023                 }
4024                 perf_mmap__read_done(&md->core);
4025         }
4026
4027         if (trace->nr_events == before) {
4028                 int timeout = done ? 100 : -1;
4029
4030                 if (!draining && evlist__poll(evlist, timeout) > 0) {
4031                         if (evlist__filter_pollfd(evlist, POLLERR | POLLHUP | POLLNVAL) == 0)
4032                                 draining = true;
4033
4034                         goto again;
4035                 } else {
4036                         if (trace__flush_events(trace))
4037                                 goto out_disable;
4038                 }
4039         } else {
4040                 goto again;
4041         }
4042
4043 out_disable:
4044         thread__zput(trace->current);
4045
4046         evlist__disable(evlist);
4047
4048         if (trace->sort_events)
4049                 ordered_events__flush(&trace->oe.data, OE_FLUSH__FINAL);
4050
4051         if (!err) {
4052                 if (trace->summary)
4053                         trace__fprintf_thread_summary(trace, trace->output);
4054
4055                 if (trace->show_tool_stats) {
4056                         fprintf(trace->output, "Stats:\n "
4057                                                " vfs_getname : %" PRIu64 "\n"
4058                                                " proc_getname: %" PRIu64 "\n",
4059                                 trace->stats.vfs_getname,
4060                                 trace->stats.proc_getname);
4061                 }
4062         }
4063
4064 out_delete_evlist:
4065         trace__symbols__exit(trace);
4066         evlist__free_syscall_tp_fields(evlist);
4067         evlist__delete(evlist);
4068         cgroup__put(trace->cgroup);
4069         trace->evlist = NULL;
4070         trace->live = false;
4071         return err;
4072 {
4073         char errbuf[BUFSIZ];
4074
4075 out_error_sched_stat_runtime:
4076         tracing_path__strerror_open_tp(errno, errbuf, sizeof(errbuf), "sched", "sched_stat_runtime");
4077         goto out_error;
4078
4079 out_error_raw_syscalls:
4080         tracing_path__strerror_open_tp(errno, errbuf, sizeof(errbuf), "raw_syscalls", "sys_(enter|exit)");
4081         goto out_error;
4082
4083 out_error_mmap:
4084         evlist__strerror_mmap(evlist, errno, errbuf, sizeof(errbuf));
4085         goto out_error;
4086
4087 out_error_open:
4088         evlist__strerror_open(evlist, errno, errbuf, sizeof(errbuf));
4089
4090 out_error:
4091         fprintf(trace->output, "%s\n", errbuf);
4092         goto out_delete_evlist;
4093
4094 out_error_apply_filters:
4095         fprintf(trace->output,
4096                 "Failed to set filter \"%s\" on event %s with %d (%s)\n",
4097                 evsel->filter, evsel__name(evsel), errno,
4098                 str_error_r(errno, errbuf, sizeof(errbuf)));
4099         goto out_delete_evlist;
4100 }
4101 out_error_mem:
4102         fprintf(trace->output, "Not enough memory to run!\n");
4103         goto out_delete_evlist;
4104
4105 out_errno:
4106         fprintf(trace->output, "errno=%d,%s\n", errno, strerror(errno));
4107         goto out_delete_evlist;
4108 }
4109
4110 static int trace__replay(struct trace *trace)
4111 {
4112         const struct evsel_str_handler handlers[] = {
4113                 { "probe:vfs_getname",       trace__vfs_getname, },
4114         };
4115         struct perf_data data = {
4116                 .path  = input_name,
4117                 .mode  = PERF_DATA_MODE_READ,
4118                 .force = trace->force,
4119         };
4120         struct perf_session *session;
4121         struct evsel *evsel;
4122         int err = -1;
4123
4124         trace->tool.sample        = trace__process_sample;
4125         trace->tool.mmap          = perf_event__process_mmap;
4126         trace->tool.mmap2         = perf_event__process_mmap2;
4127         trace->tool.comm          = perf_event__process_comm;
4128         trace->tool.exit          = perf_event__process_exit;
4129         trace->tool.fork          = perf_event__process_fork;
4130         trace->tool.attr          = perf_event__process_attr;
4131         trace->tool.tracing_data  = perf_event__process_tracing_data;
4132         trace->tool.build_id      = perf_event__process_build_id;
4133         trace->tool.namespaces    = perf_event__process_namespaces;
4134
4135         trace->tool.ordered_events = true;
4136         trace->tool.ordering_requires_timestamps = true;
4137
4138         /* add tid to output */
4139         trace->multiple_threads = true;
4140
4141         session = perf_session__new(&data, &trace->tool);
4142         if (IS_ERR(session))
4143                 return PTR_ERR(session);
4144
4145         if (trace->opts.target.pid)
4146                 symbol_conf.pid_list_str = strdup(trace->opts.target.pid);
4147
4148         if (trace->opts.target.tid)
4149                 symbol_conf.tid_list_str = strdup(trace->opts.target.tid);
4150
4151         if (symbol__init(&session->header.env) < 0)
4152                 goto out;
4153
4154         trace->host = &session->machines.host;
4155
4156         err = perf_session__set_tracepoints_handlers(session, handlers);
4157         if (err)
4158                 goto out;
4159
4160         evsel = evlist__find_tracepoint_by_name(session->evlist, "raw_syscalls:sys_enter");
4161         trace->syscalls.events.sys_enter = evsel;
4162         /* older kernels have syscalls tp versus raw_syscalls */
4163         if (evsel == NULL)
4164                 evsel = evlist__find_tracepoint_by_name(session->evlist, "syscalls:sys_enter");
4165
4166         if (evsel &&
4167             (evsel__init_raw_syscall_tp(evsel, trace__sys_enter) < 0 ||
4168             perf_evsel__init_sc_tp_ptr_field(evsel, args))) {
4169                 pr_err("Error during initialize raw_syscalls:sys_enter event\n");
4170                 goto out;
4171         }
4172
4173         evsel = evlist__find_tracepoint_by_name(session->evlist, "raw_syscalls:sys_exit");
4174         trace->syscalls.events.sys_exit = evsel;
4175         if (evsel == NULL)
4176                 evsel = evlist__find_tracepoint_by_name(session->evlist, "syscalls:sys_exit");
4177         if (evsel &&
4178             (evsel__init_raw_syscall_tp(evsel, trace__sys_exit) < 0 ||
4179             perf_evsel__init_sc_tp_uint_field(evsel, ret))) {
4180                 pr_err("Error during initialize raw_syscalls:sys_exit event\n");
4181                 goto out;
4182         }
4183
4184         evlist__for_each_entry(session->evlist, evsel) {
4185                 if (evsel->core.attr.type == PERF_TYPE_SOFTWARE &&
4186                     (evsel->core.attr.config == PERF_COUNT_SW_PAGE_FAULTS_MAJ ||
4187                      evsel->core.attr.config == PERF_COUNT_SW_PAGE_FAULTS_MIN ||
4188                      evsel->core.attr.config == PERF_COUNT_SW_PAGE_FAULTS))
4189                         evsel->handler = trace__pgfault;
4190         }
4191
4192         setup_pager();
4193
4194         err = perf_session__process_events(session);
4195         if (err)
4196                 pr_err("Failed to process events, error %d", err);
4197
4198         else if (trace->summary)
4199                 trace__fprintf_thread_summary(trace, trace->output);
4200
4201 out:
4202         perf_session__delete(session);
4203
4204         return err;
4205 }
4206
4207 static size_t trace__fprintf_threads_header(FILE *fp)
4208 {
4209         size_t printed;
4210
4211         printed  = fprintf(fp, "\n Summary of events:\n\n");
4212
4213         return printed;
4214 }
4215
4216 DEFINE_RESORT_RB(syscall_stats, a->msecs > b->msecs,
4217         struct syscall_stats *stats;
4218         double               msecs;
4219         int                  syscall;
4220 )
4221 {
4222         struct int_node *source = rb_entry(nd, struct int_node, rb_node);
4223         struct syscall_stats *stats = source->priv;
4224
4225         entry->syscall = source->i;
4226         entry->stats   = stats;
4227         entry->msecs   = stats ? (u64)stats->stats.n * (avg_stats(&stats->stats) / NSEC_PER_MSEC) : 0;
4228 }
4229
4230 static size_t thread__dump_stats(struct thread_trace *ttrace,
4231                                  struct trace *trace, FILE *fp)
4232 {
4233         size_t printed = 0;
4234         struct syscall *sc;
4235         struct rb_node *nd;
4236         DECLARE_RESORT_RB_INTLIST(syscall_stats, ttrace->syscall_stats);
4237
4238         if (syscall_stats == NULL)
4239                 return 0;
4240
4241         printed += fprintf(fp, "\n");
4242
4243         printed += fprintf(fp, "   syscall            calls  errors  total       min       avg       max       stddev\n");
4244         printed += fprintf(fp, "                                     (msec)    (msec)    (msec)    (msec)        (%%)\n");
4245         printed += fprintf(fp, "   --------------- --------  ------ -------- --------- --------- ---------     ------\n");
4246
4247         resort_rb__for_each_entry(nd, syscall_stats) {
4248                 struct syscall_stats *stats = syscall_stats_entry->stats;
4249                 if (stats) {
4250                         double min = (double)(stats->stats.min) / NSEC_PER_MSEC;
4251                         double max = (double)(stats->stats.max) / NSEC_PER_MSEC;
4252                         double avg = avg_stats(&stats->stats);
4253                         double pct;
4254                         u64 n = (u64)stats->stats.n;
4255
4256                         pct = avg ? 100.0 * stddev_stats(&stats->stats) / avg : 0.0;
4257                         avg /= NSEC_PER_MSEC;
4258
4259                         sc = &trace->syscalls.table[syscall_stats_entry->syscall];
4260                         printed += fprintf(fp, "   %-15s", sc->name);
4261                         printed += fprintf(fp, " %8" PRIu64 " %6" PRIu64 " %9.3f %9.3f %9.3f",
4262                                            n, stats->nr_failures, syscall_stats_entry->msecs, min, avg);
4263                         printed += fprintf(fp, " %9.3f %9.2f%%\n", max, pct);
4264
4265                         if (trace->errno_summary && stats->nr_failures) {
4266                                 int e;
4267
4268                                 for (e = 0; e < stats->max_errno; ++e) {
4269                                         if (stats->errnos[e] != 0)
4270                                                 fprintf(fp, "\t\t\t\t%s: %d\n", perf_env__arch_strerrno(trace->host->env, e + 1), stats->errnos[e]);
4271                                 }
4272                         }
4273                 }
4274         }
4275
4276         resort_rb__delete(syscall_stats);
4277         printed += fprintf(fp, "\n\n");
4278
4279         return printed;
4280 }
4281
4282 static size_t trace__fprintf_thread(FILE *fp, struct thread *thread, struct trace *trace)
4283 {
4284         size_t printed = 0;
4285         struct thread_trace *ttrace = thread__priv(thread);
4286         double ratio;
4287
4288         if (ttrace == NULL)
4289                 return 0;
4290
4291         ratio = (double)ttrace->nr_events / trace->nr_events * 100.0;
4292
4293         printed += fprintf(fp, " %s (%d), ", thread__comm_str(thread), thread__tid(thread));
4294         printed += fprintf(fp, "%lu events, ", ttrace->nr_events);
4295         printed += fprintf(fp, "%.1f%%", ratio);
4296         if (ttrace->pfmaj)
4297                 printed += fprintf(fp, ", %lu majfaults", ttrace->pfmaj);
4298         if (ttrace->pfmin)
4299                 printed += fprintf(fp, ", %lu minfaults", ttrace->pfmin);
4300         if (trace->sched)
4301                 printed += fprintf(fp, ", %.3f msec\n", ttrace->runtime_ms);
4302         else if (fputc('\n', fp) != EOF)
4303                 ++printed;
4304
4305         printed += thread__dump_stats(ttrace, trace, fp);
4306
4307         return printed;
4308 }
4309
4310 static unsigned long thread__nr_events(struct thread_trace *ttrace)
4311 {
4312         return ttrace ? ttrace->nr_events : 0;
4313 }
4314
4315 DEFINE_RESORT_RB(threads,
4316                 (thread__nr_events(thread__priv(a->thread)) <
4317                  thread__nr_events(thread__priv(b->thread))),
4318         struct thread *thread;
4319 )
4320 {
4321         entry->thread = rb_entry(nd, struct thread_rb_node, rb_node)->thread;
4322 }
4323
4324 static size_t trace__fprintf_thread_summary(struct trace *trace, FILE *fp)
4325 {
4326         size_t printed = trace__fprintf_threads_header(fp);
4327         struct rb_node *nd;
4328         int i;
4329
4330         for (i = 0; i < THREADS__TABLE_SIZE; i++) {
4331                 DECLARE_RESORT_RB_MACHINE_THREADS(threads, trace->host, i);
4332
4333                 if (threads == NULL) {
4334                         fprintf(fp, "%s", "Error sorting output by nr_events!\n");
4335                         return 0;
4336                 }
4337
4338                 resort_rb__for_each_entry(nd, threads)
4339                         printed += trace__fprintf_thread(fp, threads_entry->thread, trace);
4340
4341                 resort_rb__delete(threads);
4342         }
4343         return printed;
4344 }
4345
4346 static int trace__set_duration(const struct option *opt, const char *str,
4347                                int unset __maybe_unused)
4348 {
4349         struct trace *trace = opt->value;
4350
4351         trace->duration_filter = atof(str);
4352         return 0;
4353 }
4354
4355 static int trace__set_filter_pids_from_option(const struct option *opt, const char *str,
4356                                               int unset __maybe_unused)
4357 {
4358         int ret = -1;
4359         size_t i;
4360         struct trace *trace = opt->value;
4361         /*
4362          * FIXME: introduce a intarray class, plain parse csv and create a
4363          * { int nr, int entries[] } struct...
4364          */
4365         struct intlist *list = intlist__new(str);
4366
4367         if (list == NULL)
4368                 return -1;
4369
4370         i = trace->filter_pids.nr = intlist__nr_entries(list) + 1;
4371         trace->filter_pids.entries = calloc(i, sizeof(pid_t));
4372
4373         if (trace->filter_pids.entries == NULL)
4374                 goto out;
4375
4376         trace->filter_pids.entries[0] = getpid();
4377
4378         for (i = 1; i < trace->filter_pids.nr; ++i)
4379                 trace->filter_pids.entries[i] = intlist__entry(list, i - 1)->i;
4380
4381         intlist__delete(list);
4382         ret = 0;
4383 out:
4384         return ret;
4385 }
4386
4387 static int trace__open_output(struct trace *trace, const char *filename)
4388 {
4389         struct stat st;
4390
4391         if (!stat(filename, &st) && st.st_size) {
4392                 char oldname[PATH_MAX];
4393
4394                 scnprintf(oldname, sizeof(oldname), "%s.old", filename);
4395                 unlink(oldname);
4396                 rename(filename, oldname);
4397         }
4398
4399         trace->output = fopen(filename, "w");
4400
4401         return trace->output == NULL ? -errno : 0;
4402 }
4403
4404 static int parse_pagefaults(const struct option *opt, const char *str,
4405                             int unset __maybe_unused)
4406 {
4407         int *trace_pgfaults = opt->value;
4408
4409         if (strcmp(str, "all") == 0)
4410                 *trace_pgfaults |= TRACE_PFMAJ | TRACE_PFMIN;
4411         else if (strcmp(str, "maj") == 0)
4412                 *trace_pgfaults |= TRACE_PFMAJ;
4413         else if (strcmp(str, "min") == 0)
4414                 *trace_pgfaults |= TRACE_PFMIN;
4415         else
4416                 return -1;
4417
4418         return 0;
4419 }
4420
4421 static void evlist__set_default_evsel_handler(struct evlist *evlist, void *handler)
4422 {
4423         struct evsel *evsel;
4424
4425         evlist__for_each_entry(evlist, evsel) {
4426                 if (evsel->handler == NULL)
4427                         evsel->handler = handler;
4428         }
4429 }
4430
4431 static void evsel__set_syscall_arg_fmt(struct evsel *evsel, const char *name)
4432 {
4433         struct syscall_arg_fmt *fmt = evsel__syscall_arg_fmt(evsel);
4434
4435         if (fmt) {
4436                 const struct syscall_fmt *scfmt = syscall_fmt__find(name);
4437
4438                 if (scfmt) {
4439                         int skip = 0;
4440
4441                         if (strcmp(evsel->tp_format->format.fields->name, "__syscall_nr") == 0 ||
4442                             strcmp(evsel->tp_format->format.fields->name, "nr") == 0)
4443                                 ++skip;
4444
4445                         memcpy(fmt + skip, scfmt->arg, (evsel->tp_format->format.nr_fields - skip) * sizeof(*fmt));
4446                 }
4447         }
4448 }
4449
4450 static int evlist__set_syscall_tp_fields(struct evlist *evlist)
4451 {
4452         struct evsel *evsel;
4453
4454         evlist__for_each_entry(evlist, evsel) {
4455                 if (evsel->priv || !evsel->tp_format)
4456                         continue;
4457
4458                 if (strcmp(evsel->tp_format->system, "syscalls")) {
4459                         evsel__init_tp_arg_scnprintf(evsel);
4460                         continue;
4461                 }
4462
4463                 if (evsel__init_syscall_tp(evsel))
4464                         return -1;
4465
4466                 if (!strncmp(evsel->tp_format->name, "sys_enter_", 10)) {
4467                         struct syscall_tp *sc = __evsel__syscall_tp(evsel);
4468
4469                         if (__tp_field__init_ptr(&sc->args, sc->id.offset + sizeof(u64)))
4470                                 return -1;
4471
4472                         evsel__set_syscall_arg_fmt(evsel, evsel->tp_format->name + sizeof("sys_enter_") - 1);
4473                 } else if (!strncmp(evsel->tp_format->name, "sys_exit_", 9)) {
4474                         struct syscall_tp *sc = __evsel__syscall_tp(evsel);
4475
4476                         if (__tp_field__init_uint(&sc->ret, sizeof(u64), sc->id.offset + sizeof(u64), evsel->needs_swap))
4477                                 return -1;
4478
4479                         evsel__set_syscall_arg_fmt(evsel, evsel->tp_format->name + sizeof("sys_exit_") - 1);
4480                 }
4481         }
4482
4483         return 0;
4484 }
4485
4486 /*
4487  * XXX: Hackish, just splitting the combined -e+--event (syscalls
4488  * (raw_syscalls:{sys_{enter,exit}} + events (tracepoints, HW, SW, etc) to use
4489  * existing facilities unchanged (trace->ev_qualifier + parse_options()).
4490  *
4491  * It'd be better to introduce a parse_options() variant that would return a
4492  * list with the terms it didn't match to an event...
4493  */
4494 static int trace__parse_events_option(const struct option *opt, const char *str,
4495                                       int unset __maybe_unused)
4496 {
4497         struct trace *trace = (struct trace *)opt->value;
4498         const char *s = str;
4499         char *sep = NULL, *lists[2] = { NULL, NULL, };
4500         int len = strlen(str) + 1, err = -1, list, idx;
4501         char *strace_groups_dir = system_path(STRACE_GROUPS_DIR);
4502         char group_name[PATH_MAX];
4503         const struct syscall_fmt *fmt;
4504
4505         if (strace_groups_dir == NULL)
4506                 return -1;
4507
4508         if (*s == '!') {
4509                 ++s;
4510                 trace->not_ev_qualifier = true;
4511         }
4512
4513         while (1) {
4514                 if ((sep = strchr(s, ',')) != NULL)
4515                         *sep = '\0';
4516
4517                 list = 0;
4518                 if (syscalltbl__id(trace->sctbl, s) >= 0 ||
4519                     syscalltbl__strglobmatch_first(trace->sctbl, s, &idx) >= 0) {
4520                         list = 1;
4521                         goto do_concat;
4522                 }
4523
4524                 fmt = syscall_fmt__find_by_alias(s);
4525                 if (fmt != NULL) {
4526                         list = 1;
4527                         s = fmt->name;
4528                 } else {
4529                         path__join(group_name, sizeof(group_name), strace_groups_dir, s);
4530                         if (access(group_name, R_OK) == 0)
4531                                 list = 1;
4532                 }
4533 do_concat:
4534                 if (lists[list]) {
4535                         sprintf(lists[list] + strlen(lists[list]), ",%s", s);
4536                 } else {
4537                         lists[list] = malloc(len);
4538                         if (lists[list] == NULL)
4539                                 goto out;
4540                         strcpy(lists[list], s);
4541                 }
4542
4543                 if (!sep)
4544                         break;
4545
4546                 *sep = ',';
4547                 s = sep + 1;
4548         }
4549
4550         if (lists[1] != NULL) {
4551                 struct strlist_config slist_config = {
4552                         .dirname = strace_groups_dir,
4553                 };
4554
4555                 trace->ev_qualifier = strlist__new(lists[1], &slist_config);
4556                 if (trace->ev_qualifier == NULL) {
4557                         fputs("Not enough memory to parse event qualifier", trace->output);
4558                         goto out;
4559                 }
4560
4561                 if (trace__validate_ev_qualifier(trace))
4562                         goto out;
4563                 trace->trace_syscalls = true;
4564         }
4565
4566         err = 0;
4567
4568         if (lists[0]) {
4569                 struct parse_events_option_args parse_events_option_args = {
4570                         .evlistp = &trace->evlist,
4571                 };
4572                 struct option o = {
4573                         .value = &parse_events_option_args,
4574                 };
4575                 err = parse_events_option(&o, lists[0], 0);
4576         }
4577 out:
4578         free(strace_groups_dir);
4579         free(lists[0]);
4580         free(lists[1]);
4581         if (sep)
4582                 *sep = ',';
4583
4584         return err;
4585 }
4586
4587 static int trace__parse_cgroups(const struct option *opt, const char *str, int unset)
4588 {
4589         struct trace *trace = opt->value;
4590
4591         if (!list_empty(&trace->evlist->core.entries)) {
4592                 struct option o = {
4593                         .value = &trace->evlist,
4594                 };
4595                 return parse_cgroups(&o, str, unset);
4596         }
4597         trace->cgroup = evlist__findnew_cgroup(trace->evlist, str);
4598
4599         return 0;
4600 }
4601
4602 static int trace__config(const char *var, const char *value, void *arg)
4603 {
4604         struct trace *trace = arg;
4605         int err = 0;
4606
4607         if (!strcmp(var, "trace.add_events")) {
4608                 trace->perfconfig_events = strdup(value);
4609                 if (trace->perfconfig_events == NULL) {
4610                         pr_err("Not enough memory for %s\n", "trace.add_events");
4611                         return -1;
4612                 }
4613         } else if (!strcmp(var, "trace.show_timestamp")) {
4614                 trace->show_tstamp = perf_config_bool(var, value);
4615         } else if (!strcmp(var, "trace.show_duration")) {
4616                 trace->show_duration = perf_config_bool(var, value);
4617         } else if (!strcmp(var, "trace.show_arg_names")) {
4618                 trace->show_arg_names = perf_config_bool(var, value);
4619                 if (!trace->show_arg_names)
4620                         trace->show_zeros = true;
4621         } else if (!strcmp(var, "trace.show_zeros")) {
4622                 bool new_show_zeros = perf_config_bool(var, value);
4623                 if (!trace->show_arg_names && !new_show_zeros) {
4624                         pr_warning("trace.show_zeros has to be set when trace.show_arg_names=no\n");
4625                         goto out;
4626                 }
4627                 trace->show_zeros = new_show_zeros;
4628         } else if (!strcmp(var, "trace.show_prefix")) {
4629                 trace->show_string_prefix = perf_config_bool(var, value);
4630         } else if (!strcmp(var, "trace.no_inherit")) {
4631                 trace->opts.no_inherit = perf_config_bool(var, value);
4632         } else if (!strcmp(var, "trace.args_alignment")) {
4633                 int args_alignment = 0;
4634                 if (perf_config_int(&args_alignment, var, value) == 0)
4635                         trace->args_alignment = args_alignment;
4636         } else if (!strcmp(var, "trace.tracepoint_beautifiers")) {
4637                 if (strcasecmp(value, "libtraceevent") == 0)
4638                         trace->libtraceevent_print = true;
4639                 else if (strcasecmp(value, "libbeauty") == 0)
4640                         trace->libtraceevent_print = false;
4641         }
4642 out:
4643         return err;
4644 }
4645
4646 static void trace__exit(struct trace *trace)
4647 {
4648         int i;
4649
4650         strlist__delete(trace->ev_qualifier);
4651         zfree(&trace->ev_qualifier_ids.entries);
4652         if (trace->syscalls.table) {
4653                 for (i = 0; i <= trace->sctbl->syscalls.max_id; i++)
4654                         syscall__exit(&trace->syscalls.table[i]);
4655                 zfree(&trace->syscalls.table);
4656         }
4657         syscalltbl__delete(trace->sctbl);
4658         zfree(&trace->perfconfig_events);
4659 }
4660
4661 #ifdef HAVE_BPF_SKEL
4662 static int bpf__setup_bpf_output(struct evlist *evlist)
4663 {
4664         int err = parse_event(evlist, "bpf-output/no-inherit=1,name=__augmented_syscalls__/");
4665
4666         if (err)
4667                 pr_debug("ERROR: failed to create the \"__augmented_syscalls__\" bpf-output event\n");
4668
4669         return err;
4670 }
4671 #endif
4672
4673 int cmd_trace(int argc, const char **argv)
4674 {
4675         const char *trace_usage[] = {
4676                 "perf trace [<options>] [<command>]",
4677                 "perf trace [<options>] -- <command> [<options>]",
4678                 "perf trace record [<options>] [<command>]",
4679                 "perf trace record [<options>] -- <command> [<options>]",
4680                 NULL
4681         };
4682         struct trace trace = {
4683                 .opts = {
4684                         .target = {
4685                                 .uid       = UINT_MAX,
4686                                 .uses_mmap = true,
4687                         },
4688                         .user_freq     = UINT_MAX,
4689                         .user_interval = ULLONG_MAX,
4690                         .no_buffering  = true,
4691                         .mmap_pages    = UINT_MAX,
4692                 },
4693                 .output = stderr,
4694                 .show_comm = true,
4695                 .show_tstamp = true,
4696                 .show_duration = true,
4697                 .show_arg_names = true,
4698                 .args_alignment = 70,
4699                 .trace_syscalls = false,
4700                 .kernel_syscallchains = false,
4701                 .max_stack = UINT_MAX,
4702                 .max_events = ULONG_MAX,
4703         };
4704         const char *output_name = NULL;
4705         const struct option trace_options[] = {
4706         OPT_CALLBACK('e', "event", &trace, "event",
4707                      "event/syscall selector. use 'perf list' to list available events",
4708                      trace__parse_events_option),
4709         OPT_CALLBACK(0, "filter", &trace.evlist, "filter",
4710                      "event filter", parse_filter),
4711         OPT_BOOLEAN(0, "comm", &trace.show_comm,
4712                     "show the thread COMM next to its id"),
4713         OPT_BOOLEAN(0, "tool_stats", &trace.show_tool_stats, "show tool stats"),
4714         OPT_CALLBACK(0, "expr", &trace, "expr", "list of syscalls/events to trace",
4715                      trace__parse_events_option),
4716         OPT_STRING('o', "output", &output_name, "file", "output file name"),
4717         OPT_STRING('i', "input", &input_name, "file", "Analyze events in file"),
4718         OPT_STRING('p', "pid", &trace.opts.target.pid, "pid",
4719                     "trace events on existing process id"),
4720         OPT_STRING('t', "tid", &trace.opts.target.tid, "tid",
4721                     "trace events on existing thread id"),
4722         OPT_CALLBACK(0, "filter-pids", &trace, "CSV list of pids",
4723                      "pids to filter (by the kernel)", trace__set_filter_pids_from_option),
4724         OPT_BOOLEAN('a', "all-cpus", &trace.opts.target.system_wide,
4725                     "system-wide collection from all CPUs"),
4726         OPT_STRING('C', "cpu", &trace.opts.target.cpu_list, "cpu",
4727                     "list of cpus to monitor"),
4728         OPT_BOOLEAN(0, "no-inherit", &trace.opts.no_inherit,
4729                     "child tasks do not inherit counters"),
4730         OPT_CALLBACK('m', "mmap-pages", &trace.opts.mmap_pages, "pages",
4731                      "number of mmap data pages", evlist__parse_mmap_pages),
4732         OPT_STRING('u', "uid", &trace.opts.target.uid_str, "user",
4733                    "user to profile"),
4734         OPT_CALLBACK(0, "duration", &trace, "float",
4735                      "show only events with duration > N.M ms",
4736                      trace__set_duration),
4737         OPT_BOOLEAN(0, "sched", &trace.sched, "show blocking scheduler events"),
4738         OPT_INCR('v', "verbose", &verbose, "be more verbose"),
4739         OPT_BOOLEAN('T', "time", &trace.full_time,
4740                     "Show full timestamp, not time relative to first start"),
4741         OPT_BOOLEAN(0, "failure", &trace.failure_only,
4742                     "Show only syscalls that failed"),
4743         OPT_BOOLEAN('s', "summary", &trace.summary_only,
4744                     "Show only syscall summary with statistics"),
4745         OPT_BOOLEAN('S', "with-summary", &trace.summary,
4746                     "Show all syscalls and summary with statistics"),
4747         OPT_BOOLEAN(0, "errno-summary", &trace.errno_summary,
4748                     "Show errno stats per syscall, use with -s or -S"),
4749         OPT_CALLBACK_DEFAULT('F', "pf", &trace.trace_pgfaults, "all|maj|min",
4750                      "Trace pagefaults", parse_pagefaults, "maj"),
4751         OPT_BOOLEAN(0, "syscalls", &trace.trace_syscalls, "Trace syscalls"),
4752         OPT_BOOLEAN('f', "force", &trace.force, "don't complain, do it"),
4753         OPT_CALLBACK(0, "call-graph", &trace.opts,
4754                      "record_mode[,record_size]", record_callchain_help,
4755                      &record_parse_callchain_opt),
4756         OPT_BOOLEAN(0, "libtraceevent_print", &trace.libtraceevent_print,
4757                     "Use libtraceevent to print the tracepoint arguments."),
4758         OPT_BOOLEAN(0, "kernel-syscall-graph", &trace.kernel_syscallchains,
4759                     "Show the kernel callchains on the syscall exit path"),
4760         OPT_ULONG(0, "max-events", &trace.max_events,
4761                 "Set the maximum number of events to print, exit after that is reached. "),
4762         OPT_UINTEGER(0, "min-stack", &trace.min_stack,
4763                      "Set the minimum stack depth when parsing the callchain, "
4764                      "anything below the specified depth will be ignored."),
4765         OPT_UINTEGER(0, "max-stack", &trace.max_stack,
4766                      "Set the maximum stack depth when parsing the callchain, "
4767                      "anything beyond the specified depth will be ignored. "
4768                      "Default: kernel.perf_event_max_stack or " __stringify(PERF_MAX_STACK_DEPTH)),
4769         OPT_BOOLEAN(0, "sort-events", &trace.sort_events,
4770                         "Sort batch of events before processing, use if getting out of order events"),
4771         OPT_BOOLEAN(0, "print-sample", &trace.print_sample,
4772                         "print the PERF_RECORD_SAMPLE PERF_SAMPLE_ info, for debugging"),
4773         OPT_UINTEGER(0, "proc-map-timeout", &proc_map_timeout,
4774                         "per thread proc mmap processing timeout in ms"),
4775         OPT_CALLBACK('G', "cgroup", &trace, "name", "monitor event in cgroup name only",
4776                      trace__parse_cgroups),
4777         OPT_INTEGER('D', "delay", &trace.opts.target.initial_delay,
4778                      "ms to wait before starting measurement after program "
4779                      "start"),
4780         OPTS_EVSWITCH(&trace.evswitch),
4781         OPT_END()
4782         };
4783         bool __maybe_unused max_stack_user_set = true;
4784         bool mmap_pages_user_set = true;
4785         struct evsel *evsel;
4786         const char * const trace_subcommands[] = { "record", NULL };
4787         int err = -1;
4788         char bf[BUFSIZ];
4789         struct sigaction sigchld_act;
4790
4791         signal(SIGSEGV, sighandler_dump_stack);
4792         signal(SIGFPE, sighandler_dump_stack);
4793         signal(SIGINT, sighandler_interrupt);
4794
4795         memset(&sigchld_act, 0, sizeof(sigchld_act));
4796         sigchld_act.sa_flags = SA_SIGINFO;
4797         sigchld_act.sa_sigaction = sighandler_chld;
4798         sigaction(SIGCHLD, &sigchld_act, NULL);
4799
4800         trace.evlist = evlist__new();
4801         trace.sctbl = syscalltbl__new();
4802
4803         if (trace.evlist == NULL || trace.sctbl == NULL) {
4804                 pr_err("Not enough memory to run!\n");
4805                 err = -ENOMEM;
4806                 goto out;
4807         }
4808
4809         /*
4810          * Parsing .perfconfig may entail creating a BPF event, that may need
4811          * to create BPF maps, so bump RLIM_MEMLOCK as the default 64K setting
4812          * is too small. This affects just this process, not touching the
4813          * global setting. If it fails we'll get something in 'perf trace -v'
4814          * to help diagnose the problem.
4815          */
4816         rlimit__bump_memlock();
4817
4818         err = perf_config(trace__config, &trace);
4819         if (err)
4820                 goto out;
4821
4822         argc = parse_options_subcommand(argc, argv, trace_options, trace_subcommands,
4823                                  trace_usage, PARSE_OPT_STOP_AT_NON_OPTION);
4824
4825         /*
4826          * Here we already passed thru trace__parse_events_option() and it has
4827          * already figured out if -e syscall_name, if not but if --event
4828          * foo:bar was used, the user is interested _just_ in those, say,
4829          * tracepoint events, not in the strace-like syscall-name-based mode.
4830          *
4831          * This is important because we need to check if strace-like mode is
4832          * needed to decided if we should filter out the eBPF
4833          * __augmented_syscalls__ code, if it is in the mix, say, via
4834          * .perfconfig trace.add_events, and filter those out.
4835          */
4836         if (!trace.trace_syscalls && !trace.trace_pgfaults &&
4837             trace.evlist->core.nr_entries == 0 /* Was --events used? */) {
4838                 trace.trace_syscalls = true;
4839         }
4840         /*
4841          * Now that we have --verbose figured out, lets see if we need to parse
4842          * events from .perfconfig, so that if those events fail parsing, say some
4843          * BPF program fails, then we'll be able to use --verbose to see what went
4844          * wrong in more detail.
4845          */
4846         if (trace.perfconfig_events != NULL) {
4847                 struct parse_events_error parse_err;
4848
4849                 parse_events_error__init(&parse_err);
4850                 err = parse_events(trace.evlist, trace.perfconfig_events, &parse_err);
4851                 if (err)
4852                         parse_events_error__print(&parse_err, trace.perfconfig_events);
4853                 parse_events_error__exit(&parse_err);
4854                 if (err)
4855                         goto out;
4856         }
4857
4858         if ((nr_cgroups || trace.cgroup) && !trace.opts.target.system_wide) {
4859                 usage_with_options_msg(trace_usage, trace_options,
4860                                        "cgroup monitoring only available in system-wide mode");
4861         }
4862
4863 #ifdef HAVE_BPF_SKEL
4864         if (!trace.trace_syscalls)
4865                 goto skip_augmentation;
4866
4867         trace.skel = augmented_raw_syscalls_bpf__open();
4868         if (!trace.skel) {
4869                 pr_debug("Failed to open augmented syscalls BPF skeleton");
4870         } else {
4871                 /*
4872                  * Disable attaching the BPF programs except for sys_enter and
4873                  * sys_exit that tail call into this as necessary.
4874                  */
4875                 struct bpf_program *prog;
4876
4877                 bpf_object__for_each_program(prog, trace.skel->obj) {
4878                         if (prog != trace.skel->progs.sys_enter && prog != trace.skel->progs.sys_exit)
4879                                 bpf_program__set_autoattach(prog, /*autoattach=*/false);
4880                 }
4881
4882                 err = augmented_raw_syscalls_bpf__load(trace.skel);
4883
4884                 if (err < 0) {
4885                         libbpf_strerror(err, bf, sizeof(bf));
4886                         pr_debug("Failed to load augmented syscalls BPF skeleton: %s\n", bf);
4887                 } else {
4888                         augmented_raw_syscalls_bpf__attach(trace.skel);
4889                         trace__add_syscall_newtp(&trace);
4890                 }
4891         }
4892
4893         err = bpf__setup_bpf_output(trace.evlist);
4894         if (err) {
4895                 libbpf_strerror(err, bf, sizeof(bf));
4896                 pr_err("ERROR: Setup BPF output event failed: %s\n", bf);
4897                 goto out;
4898         }
4899         trace.syscalls.events.bpf_output = evlist__last(trace.evlist);
4900         assert(!strcmp(evsel__name(trace.syscalls.events.bpf_output), "__augmented_syscalls__"));
4901 skip_augmentation:
4902 #endif
4903         err = -1;
4904
4905         if (trace.trace_pgfaults) {
4906                 trace.opts.sample_address = true;
4907                 trace.opts.sample_time = true;
4908         }
4909
4910         if (trace.opts.mmap_pages == UINT_MAX)
4911                 mmap_pages_user_set = false;
4912
4913         if (trace.max_stack == UINT_MAX) {
4914                 trace.max_stack = input_name ? PERF_MAX_STACK_DEPTH : sysctl__max_stack();
4915                 max_stack_user_set = false;
4916         }
4917
4918 #ifdef HAVE_DWARF_UNWIND_SUPPORT
4919         if ((trace.min_stack || max_stack_user_set) && !callchain_param.enabled) {
4920                 record_opts__parse_callchain(&trace.opts, &callchain_param, "dwarf", false);
4921         }
4922 #endif
4923
4924         if (callchain_param.enabled) {
4925                 if (!mmap_pages_user_set && geteuid() == 0)
4926                         trace.opts.mmap_pages = perf_event_mlock_kb_in_pages() * 4;
4927
4928                 symbol_conf.use_callchain = true;
4929         }
4930
4931         if (trace.evlist->core.nr_entries > 0) {
4932                 evlist__set_default_evsel_handler(trace.evlist, trace__event_handler);
4933                 if (evlist__set_syscall_tp_fields(trace.evlist)) {
4934                         perror("failed to set syscalls:* tracepoint fields");
4935                         goto out;
4936                 }
4937         }
4938
4939         if (trace.sort_events) {
4940                 ordered_events__init(&trace.oe.data, ordered_events__deliver_event, &trace);
4941                 ordered_events__set_copy_on_queue(&trace.oe.data, true);
4942         }
4943
4944         /*
4945          * If we are augmenting syscalls, then combine what we put in the
4946          * __augmented_syscalls__ BPF map with what is in the
4947          * syscalls:sys_exit_FOO tracepoints, i.e. just like we do without BPF,
4948          * combining raw_syscalls:sys_enter with raw_syscalls:sys_exit.
4949          *
4950          * We'll switch to look at two BPF maps, one for sys_enter and the
4951          * other for sys_exit when we start augmenting the sys_exit paths with
4952          * buffers that are being copied from kernel to userspace, think 'read'
4953          * syscall.
4954          */
4955         if (trace.syscalls.events.bpf_output) {
4956                 evlist__for_each_entry(trace.evlist, evsel) {
4957                         bool raw_syscalls_sys_exit = strcmp(evsel__name(evsel), "raw_syscalls:sys_exit") == 0;
4958
4959                         if (raw_syscalls_sys_exit) {
4960                                 trace.raw_augmented_syscalls = true;
4961                                 goto init_augmented_syscall_tp;
4962                         }
4963
4964                         if (trace.syscalls.events.bpf_output->priv == NULL &&
4965                             strstr(evsel__name(evsel), "syscalls:sys_enter")) {
4966                                 struct evsel *augmented = trace.syscalls.events.bpf_output;
4967                                 if (evsel__init_augmented_syscall_tp(augmented, evsel) ||
4968                                     evsel__init_augmented_syscall_tp_args(augmented))
4969                                         goto out;
4970                                 /*
4971                                  * Augmented is __augmented_syscalls__ BPF_OUTPUT event
4972                                  * Above we made sure we can get from the payload the tp fields
4973                                  * that we get from syscalls:sys_enter tracefs format file.
4974                                  */
4975                                 augmented->handler = trace__sys_enter;
4976                                 /*
4977                                  * Now we do the same for the *syscalls:sys_enter event so that
4978                                  * if we handle it directly, i.e. if the BPF prog returns 0 so
4979                                  * as not to filter it, then we'll handle it just like we would
4980                                  * for the BPF_OUTPUT one:
4981                                  */
4982                                 if (evsel__init_augmented_syscall_tp(evsel, evsel) ||
4983                                     evsel__init_augmented_syscall_tp_args(evsel))
4984                                         goto out;
4985                                 evsel->handler = trace__sys_enter;
4986                         }
4987
4988                         if (strstarts(evsel__name(evsel), "syscalls:sys_exit_")) {
4989                                 struct syscall_tp *sc;
4990 init_augmented_syscall_tp:
4991                                 if (evsel__init_augmented_syscall_tp(evsel, evsel))
4992                                         goto out;
4993                                 sc = __evsel__syscall_tp(evsel);
4994                                 /*
4995                                  * For now with BPF raw_augmented we hook into
4996                                  * raw_syscalls:sys_enter and there we get all
4997                                  * 6 syscall args plus the tracepoint common
4998                                  * fields and the syscall_nr (another long).
4999                                  * So we check if that is the case and if so
5000                                  * don't look after the sc->args_size but
5001                                  * always after the full raw_syscalls:sys_enter
5002                                  * payload, which is fixed.
5003                                  *
5004                                  * We'll revisit this later to pass
5005                                  * s->args_size to the BPF augmenter (now
5006                                  * tools/perf/examples/bpf/augmented_raw_syscalls.c,
5007                                  * so that it copies only what we need for each
5008                                  * syscall, like what happens when we use
5009                                  * syscalls:sys_enter_NAME, so that we reduce
5010                                  * the kernel/userspace traffic to just what is
5011                                  * needed for each syscall.
5012                                  */
5013                                 if (trace.raw_augmented_syscalls)
5014                                         trace.raw_augmented_syscalls_args_size = (6 + 1) * sizeof(long) + sc->id.offset;
5015                                 evsel__init_augmented_syscall_tp_ret(evsel);
5016                                 evsel->handler = trace__sys_exit;
5017                         }
5018                 }
5019         }
5020
5021         if ((argc >= 1) && (strcmp(argv[0], "record") == 0))
5022                 return trace__record(&trace, argc-1, &argv[1]);
5023
5024         /* Using just --errno-summary will trigger --summary */
5025         if (trace.errno_summary && !trace.summary && !trace.summary_only)
5026                 trace.summary_only = true;
5027
5028         /* summary_only implies summary option, but don't overwrite summary if set */
5029         if (trace.summary_only)
5030                 trace.summary = trace.summary_only;
5031
5032         if (output_name != NULL) {
5033                 err = trace__open_output(&trace, output_name);
5034                 if (err < 0) {
5035                         perror("failed to create output file");
5036                         goto out;
5037                 }
5038         }
5039
5040         err = evswitch__init(&trace.evswitch, trace.evlist, stderr);
5041         if (err)
5042                 goto out_close;
5043
5044         err = target__validate(&trace.opts.target);
5045         if (err) {
5046                 target__strerror(&trace.opts.target, err, bf, sizeof(bf));
5047                 fprintf(trace.output, "%s", bf);
5048                 goto out_close;
5049         }
5050
5051         err = target__parse_uid(&trace.opts.target);
5052         if (err) {
5053                 target__strerror(&trace.opts.target, err, bf, sizeof(bf));
5054                 fprintf(trace.output, "%s", bf);
5055                 goto out_close;
5056         }
5057
5058         if (!argc && target__none(&trace.opts.target))
5059                 trace.opts.target.system_wide = true;
5060
5061         if (input_name)
5062                 err = trace__replay(&trace);
5063         else
5064                 err = trace__run(&trace, argc, argv);
5065
5066 out_close:
5067         if (output_name != NULL)
5068                 fclose(trace.output);
5069 out:
5070         trace__exit(&trace);
5071 #ifdef HAVE_BPF_SKEL
5072         augmented_raw_syscalls_bpf__destroy(trace.skel);
5073 #endif
5074         return err;
5075 }