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         const char *arch_name = perf_env__arch(env);
2474
2475         return arch_syscalls__strerrno(arch_name, err);
2476 }
2477
2478 static int trace__sys_exit(struct trace *trace, struct evsel *evsel,
2479                            union perf_event *event __maybe_unused,
2480                            struct perf_sample *sample)
2481 {
2482         long ret;
2483         u64 duration = 0;
2484         bool duration_calculated = false;
2485         struct thread *thread;
2486         int id = perf_evsel__sc_tp_uint(evsel, id, sample), err = -1, callchain_ret = 0, printed = 0;
2487         int alignment = trace->args_alignment;
2488         struct syscall *sc = trace__syscall_info(trace, evsel, id);
2489         struct thread_trace *ttrace;
2490
2491         if (sc == NULL)
2492                 return -1;
2493
2494         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2495         ttrace = thread__trace(thread, trace->output);
2496         if (ttrace == NULL)
2497                 goto out_put;
2498
2499         trace__fprintf_sample(trace, evsel, sample, thread);
2500
2501         ret = perf_evsel__sc_tp_uint(evsel, ret, sample);
2502
2503         if (trace->summary)
2504                 thread__update_stats(thread, ttrace, id, sample, ret, trace->errno_summary);
2505
2506         if (!trace->fd_path_disabled && sc->is_open && ret >= 0 && ttrace->filename.pending_open) {
2507                 trace__set_fd_pathname(thread, ret, ttrace->filename.name);
2508                 ttrace->filename.pending_open = false;
2509                 ++trace->stats.vfs_getname;
2510         }
2511
2512         if (ttrace->entry_time) {
2513                 duration = sample->time - ttrace->entry_time;
2514                 if (trace__filter_duration(trace, duration))
2515                         goto out;
2516                 duration_calculated = true;
2517         } else if (trace->duration_filter)
2518                 goto out;
2519
2520         if (sample->callchain) {
2521                 struct callchain_cursor *cursor = get_tls_callchain_cursor();
2522
2523                 callchain_ret = trace__resolve_callchain(trace, evsel, sample, cursor);
2524                 if (callchain_ret == 0) {
2525                         if (cursor->nr < trace->min_stack)
2526                                 goto out;
2527                         callchain_ret = 1;
2528                 }
2529         }
2530
2531         if (trace->summary_only || (ret >= 0 && trace->failure_only))
2532                 goto out;
2533
2534         trace__fprintf_entry_head(trace, thread, duration, duration_calculated, ttrace->entry_time, trace->output);
2535
2536         if (ttrace->entry_pending) {
2537                 printed = fprintf(trace->output, "%s", ttrace->entry_str);
2538         } else {
2539                 printed += fprintf(trace->output, " ... [");
2540                 color_fprintf(trace->output, PERF_COLOR_YELLOW, "continued");
2541                 printed += 9;
2542                 printed += fprintf(trace->output, "]: %s()", sc->name);
2543         }
2544
2545         printed++; /* the closing ')' */
2546
2547         if (alignment > printed)
2548                 alignment -= printed;
2549         else
2550                 alignment = 0;
2551
2552         fprintf(trace->output, ")%*s= ", alignment, " ");
2553
2554         if (sc->fmt == NULL) {
2555                 if (ret < 0)
2556                         goto errno_print;
2557 signed_print:
2558                 fprintf(trace->output, "%ld", ret);
2559         } else if (ret < 0) {
2560 errno_print: {
2561                 char bf[STRERR_BUFSIZE];
2562                 const char *emsg = str_error_r(-ret, bf, sizeof(bf)),
2563                            *e = errno_to_name(evsel, -ret);
2564
2565                 fprintf(trace->output, "-1 %s (%s)", e, emsg);
2566         }
2567         } else if (ret == 0 && sc->fmt->timeout)
2568                 fprintf(trace->output, "0 (Timeout)");
2569         else if (ttrace->ret_scnprintf) {
2570                 char bf[1024];
2571                 struct syscall_arg arg = {
2572                         .val    = ret,
2573                         .thread = thread,
2574                         .trace  = trace,
2575                 };
2576                 ttrace->ret_scnprintf(bf, sizeof(bf), &arg);
2577                 ttrace->ret_scnprintf = NULL;
2578                 fprintf(trace->output, "%s", bf);
2579         } else if (sc->fmt->hexret)
2580                 fprintf(trace->output, "%#lx", ret);
2581         else if (sc->fmt->errpid) {
2582                 struct thread *child = machine__find_thread(trace->host, ret, ret);
2583
2584                 if (child != NULL) {
2585                         fprintf(trace->output, "%ld", ret);
2586                         if (thread__comm_set(child))
2587                                 fprintf(trace->output, " (%s)", thread__comm_str(child));
2588                         thread__put(child);
2589                 }
2590         } else
2591                 goto signed_print;
2592
2593         fputc('\n', trace->output);
2594
2595         /*
2596          * We only consider an 'event' for the sake of --max-events a non-filtered
2597          * sys_enter + sys_exit and other tracepoint events.
2598          */
2599         if (++trace->nr_events_printed == trace->max_events && trace->max_events != ULONG_MAX)
2600                 interrupted = true;
2601
2602         if (callchain_ret > 0)
2603                 trace__fprintf_callchain(trace, sample);
2604         else if (callchain_ret < 0)
2605                 pr_err("Problem processing %s callchain, skipping...\n", evsel__name(evsel));
2606 out:
2607         ttrace->entry_pending = false;
2608         err = 0;
2609 out_put:
2610         thread__put(thread);
2611         return err;
2612 }
2613
2614 static int trace__vfs_getname(struct trace *trace, struct evsel *evsel,
2615                               union perf_event *event __maybe_unused,
2616                               struct perf_sample *sample)
2617 {
2618         struct thread *thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2619         struct thread_trace *ttrace;
2620         size_t filename_len, entry_str_len, to_move;
2621         ssize_t remaining_space;
2622         char *pos;
2623         const char *filename = evsel__rawptr(evsel, sample, "pathname");
2624
2625         if (!thread)
2626                 goto out;
2627
2628         ttrace = thread__priv(thread);
2629         if (!ttrace)
2630                 goto out_put;
2631
2632         filename_len = strlen(filename);
2633         if (filename_len == 0)
2634                 goto out_put;
2635
2636         if (ttrace->filename.namelen < filename_len) {
2637                 char *f = realloc(ttrace->filename.name, filename_len + 1);
2638
2639                 if (f == NULL)
2640                         goto out_put;
2641
2642                 ttrace->filename.namelen = filename_len;
2643                 ttrace->filename.name = f;
2644         }
2645
2646         strcpy(ttrace->filename.name, filename);
2647         ttrace->filename.pending_open = true;
2648
2649         if (!ttrace->filename.ptr)
2650                 goto out_put;
2651
2652         entry_str_len = strlen(ttrace->entry_str);
2653         remaining_space = trace__entry_str_size - entry_str_len - 1; /* \0 */
2654         if (remaining_space <= 0)
2655                 goto out_put;
2656
2657         if (filename_len > (size_t)remaining_space) {
2658                 filename += filename_len - remaining_space;
2659                 filename_len = remaining_space;
2660         }
2661
2662         to_move = entry_str_len - ttrace->filename.entry_str_pos + 1; /* \0 */
2663         pos = ttrace->entry_str + ttrace->filename.entry_str_pos;
2664         memmove(pos + filename_len, pos, to_move);
2665         memcpy(pos, filename, filename_len);
2666
2667         ttrace->filename.ptr = 0;
2668         ttrace->filename.entry_str_pos = 0;
2669 out_put:
2670         thread__put(thread);
2671 out:
2672         return 0;
2673 }
2674
2675 static int trace__sched_stat_runtime(struct trace *trace, struct evsel *evsel,
2676                                      union perf_event *event __maybe_unused,
2677                                      struct perf_sample *sample)
2678 {
2679         u64 runtime = evsel__intval(evsel, sample, "runtime");
2680         double runtime_ms = (double)runtime / NSEC_PER_MSEC;
2681         struct thread *thread = machine__findnew_thread(trace->host,
2682                                                         sample->pid,
2683                                                         sample->tid);
2684         struct thread_trace *ttrace = thread__trace(thread, trace->output);
2685
2686         if (ttrace == NULL)
2687                 goto out_dump;
2688
2689         ttrace->runtime_ms += runtime_ms;
2690         trace->runtime_ms += runtime_ms;
2691 out_put:
2692         thread__put(thread);
2693         return 0;
2694
2695 out_dump:
2696         fprintf(trace->output, "%s: comm=%s,pid=%u,runtime=%" PRIu64 ",vruntime=%" PRIu64 ")\n",
2697                evsel->name,
2698                evsel__strval(evsel, sample, "comm"),
2699                (pid_t)evsel__intval(evsel, sample, "pid"),
2700                runtime,
2701                evsel__intval(evsel, sample, "vruntime"));
2702         goto out_put;
2703 }
2704
2705 static int bpf_output__printer(enum binary_printer_ops op,
2706                                unsigned int val, void *extra __maybe_unused, FILE *fp)
2707 {
2708         unsigned char ch = (unsigned char)val;
2709
2710         switch (op) {
2711         case BINARY_PRINT_CHAR_DATA:
2712                 return fprintf(fp, "%c", isprint(ch) ? ch : '.');
2713         case BINARY_PRINT_DATA_BEGIN:
2714         case BINARY_PRINT_LINE_BEGIN:
2715         case BINARY_PRINT_ADDR:
2716         case BINARY_PRINT_NUM_DATA:
2717         case BINARY_PRINT_NUM_PAD:
2718         case BINARY_PRINT_SEP:
2719         case BINARY_PRINT_CHAR_PAD:
2720         case BINARY_PRINT_LINE_END:
2721         case BINARY_PRINT_DATA_END:
2722         default:
2723                 break;
2724         }
2725
2726         return 0;
2727 }
2728
2729 static void bpf_output__fprintf(struct trace *trace,
2730                                 struct perf_sample *sample)
2731 {
2732         binary__fprintf(sample->raw_data, sample->raw_size, 8,
2733                         bpf_output__printer, NULL, trace->output);
2734         ++trace->nr_events_printed;
2735 }
2736
2737 static size_t trace__fprintf_tp_fields(struct trace *trace, struct evsel *evsel, struct perf_sample *sample,
2738                                        struct thread *thread, void *augmented_args, int augmented_args_size)
2739 {
2740         char bf[2048];
2741         size_t size = sizeof(bf);
2742         struct tep_format_field *field = evsel->tp_format->format.fields;
2743         struct syscall_arg_fmt *arg = __evsel__syscall_arg_fmt(evsel);
2744         size_t printed = 0;
2745         unsigned long val;
2746         u8 bit = 1;
2747         struct syscall_arg syscall_arg = {
2748                 .augmented = {
2749                         .size = augmented_args_size,
2750                         .args = augmented_args,
2751                 },
2752                 .idx    = 0,
2753                 .mask   = 0,
2754                 .trace  = trace,
2755                 .thread = thread,
2756                 .show_string_prefix = trace->show_string_prefix,
2757         };
2758
2759         for (; field && arg; field = field->next, ++syscall_arg.idx, bit <<= 1, ++arg) {
2760                 if (syscall_arg.mask & bit)
2761                         continue;
2762
2763                 syscall_arg.len = 0;
2764                 syscall_arg.fmt = arg;
2765                 if (field->flags & TEP_FIELD_IS_ARRAY) {
2766                         int offset = field->offset;
2767
2768                         if (field->flags & TEP_FIELD_IS_DYNAMIC) {
2769                                 offset = format_field__intval(field, sample, evsel->needs_swap);
2770                                 syscall_arg.len = offset >> 16;
2771                                 offset &= 0xffff;
2772                                 if (tep_field_is_relative(field->flags))
2773                                         offset += field->offset + field->size;
2774                         }
2775
2776                         val = (uintptr_t)(sample->raw_data + offset);
2777                 } else
2778                         val = format_field__intval(field, sample, evsel->needs_swap);
2779                 /*
2780                  * Some syscall args need some mask, most don't and
2781                  * return val untouched.
2782                  */
2783                 val = syscall_arg_fmt__mask_val(arg, &syscall_arg, val);
2784
2785                 /*
2786                  * Suppress this argument if its value is zero and
2787                  * we don't have a string associated in an
2788                  * strarray for it.
2789                  */
2790                 if (val == 0 &&
2791                     !trace->show_zeros &&
2792                     !((arg->show_zero ||
2793                        arg->scnprintf == SCA_STRARRAY ||
2794                        arg->scnprintf == SCA_STRARRAYS) &&
2795                       arg->parm))
2796                         continue;
2797
2798                 printed += scnprintf(bf + printed, size - printed, "%s", printed ? ", " : "");
2799
2800                 if (trace->show_arg_names)
2801                         printed += scnprintf(bf + printed, size - printed, "%s: ", field->name);
2802
2803                 printed += syscall_arg_fmt__scnprintf_val(arg, bf + printed, size - printed, &syscall_arg, val);
2804         }
2805
2806         return printed + fprintf(trace->output, "%s", bf);
2807 }
2808
2809 static int trace__event_handler(struct trace *trace, struct evsel *evsel,
2810                                 union perf_event *event __maybe_unused,
2811                                 struct perf_sample *sample)
2812 {
2813         struct thread *thread;
2814         int callchain_ret = 0;
2815         /*
2816          * Check if we called perf_evsel__disable(evsel) due to, for instance,
2817          * this event's max_events having been hit and this is an entry coming
2818          * from the ring buffer that we should discard, since the max events
2819          * have already been considered/printed.
2820          */
2821         if (evsel->disabled)
2822                 return 0;
2823
2824         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2825
2826         if (sample->callchain) {
2827                 struct callchain_cursor *cursor = get_tls_callchain_cursor();
2828
2829                 callchain_ret = trace__resolve_callchain(trace, evsel, sample, cursor);
2830                 if (callchain_ret == 0) {
2831                         if (cursor->nr < trace->min_stack)
2832                                 goto out;
2833                         callchain_ret = 1;
2834                 }
2835         }
2836
2837         trace__printf_interrupted_entry(trace);
2838         trace__fprintf_tstamp(trace, sample->time, trace->output);
2839
2840         if (trace->trace_syscalls && trace->show_duration)
2841                 fprintf(trace->output, "(         ): ");
2842
2843         if (thread)
2844                 trace__fprintf_comm_tid(trace, thread, trace->output);
2845
2846         if (evsel == trace->syscalls.events.bpf_output) {
2847                 int id = perf_evsel__sc_tp_uint(evsel, id, sample);
2848                 struct syscall *sc = trace__syscall_info(trace, evsel, id);
2849
2850                 if (sc) {
2851                         fprintf(trace->output, "%s(", sc->name);
2852                         trace__fprintf_sys_enter(trace, evsel, sample);
2853                         fputc(')', trace->output);
2854                         goto newline;
2855                 }
2856
2857                 /*
2858                  * XXX: Not having the associated syscall info or not finding/adding
2859                  *      the thread should never happen, but if it does...
2860                  *      fall thru and print it as a bpf_output event.
2861                  */
2862         }
2863
2864         fprintf(trace->output, "%s(", evsel->name);
2865
2866         if (evsel__is_bpf_output(evsel)) {
2867                 bpf_output__fprintf(trace, sample);
2868         } else if (evsel->tp_format) {
2869                 if (strncmp(evsel->tp_format->name, "sys_enter_", 10) ||
2870                     trace__fprintf_sys_enter(trace, evsel, sample)) {
2871                         if (trace->libtraceevent_print) {
2872                                 event_format__fprintf(evsel->tp_format, sample->cpu,
2873                                                       sample->raw_data, sample->raw_size,
2874                                                       trace->output);
2875                         } else {
2876                                 trace__fprintf_tp_fields(trace, evsel, sample, thread, NULL, 0);
2877                         }
2878                 }
2879         }
2880
2881 newline:
2882         fprintf(trace->output, ")\n");
2883
2884         if (callchain_ret > 0)
2885                 trace__fprintf_callchain(trace, sample);
2886         else if (callchain_ret < 0)
2887                 pr_err("Problem processing %s callchain, skipping...\n", evsel__name(evsel));
2888
2889         ++trace->nr_events_printed;
2890
2891         if (evsel->max_events != ULONG_MAX && ++evsel->nr_events_printed == evsel->max_events) {
2892                 evsel__disable(evsel);
2893                 evsel__close(evsel);
2894         }
2895 out:
2896         thread__put(thread);
2897         return 0;
2898 }
2899
2900 static void print_location(FILE *f, struct perf_sample *sample,
2901                            struct addr_location *al,
2902                            bool print_dso, bool print_sym)
2903 {
2904
2905         if ((verbose > 0 || print_dso) && al->map)
2906                 fprintf(f, "%s@", map__dso(al->map)->long_name);
2907
2908         if ((verbose > 0 || print_sym) && al->sym)
2909                 fprintf(f, "%s+0x%" PRIx64, al->sym->name,
2910                         al->addr - al->sym->start);
2911         else if (al->map)
2912                 fprintf(f, "0x%" PRIx64, al->addr);
2913         else
2914                 fprintf(f, "0x%" PRIx64, sample->addr);
2915 }
2916
2917 static int trace__pgfault(struct trace *trace,
2918                           struct evsel *evsel,
2919                           union perf_event *event __maybe_unused,
2920                           struct perf_sample *sample)
2921 {
2922         struct thread *thread;
2923         struct addr_location al;
2924         char map_type = 'd';
2925         struct thread_trace *ttrace;
2926         int err = -1;
2927         int callchain_ret = 0;
2928
2929         addr_location__init(&al);
2930         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2931
2932         if (sample->callchain) {
2933                 struct callchain_cursor *cursor = get_tls_callchain_cursor();
2934
2935                 callchain_ret = trace__resolve_callchain(trace, evsel, sample, cursor);
2936                 if (callchain_ret == 0) {
2937                         if (cursor->nr < trace->min_stack)
2938                                 goto out_put;
2939                         callchain_ret = 1;
2940                 }
2941         }
2942
2943         ttrace = thread__trace(thread, trace->output);
2944         if (ttrace == NULL)
2945                 goto out_put;
2946
2947         if (evsel->core.attr.config == PERF_COUNT_SW_PAGE_FAULTS_MAJ)
2948                 ttrace->pfmaj++;
2949         else
2950                 ttrace->pfmin++;
2951
2952         if (trace->summary_only)
2953                 goto out;
2954
2955         thread__find_symbol(thread, sample->cpumode, sample->ip, &al);
2956
2957         trace__fprintf_entry_head(trace, thread, 0, true, sample->time, trace->output);
2958
2959         fprintf(trace->output, "%sfault [",
2960                 evsel->core.attr.config == PERF_COUNT_SW_PAGE_FAULTS_MAJ ?
2961                 "maj" : "min");
2962
2963         print_location(trace->output, sample, &al, false, true);
2964
2965         fprintf(trace->output, "] => ");
2966
2967         thread__find_symbol(thread, sample->cpumode, sample->addr, &al);
2968
2969         if (!al.map) {
2970                 thread__find_symbol(thread, sample->cpumode, sample->addr, &al);
2971
2972                 if (al.map)
2973                         map_type = 'x';
2974                 else
2975                         map_type = '?';
2976         }
2977
2978         print_location(trace->output, sample, &al, true, false);
2979
2980         fprintf(trace->output, " (%c%c)\n", map_type, al.level);
2981
2982         if (callchain_ret > 0)
2983                 trace__fprintf_callchain(trace, sample);
2984         else if (callchain_ret < 0)
2985                 pr_err("Problem processing %s callchain, skipping...\n", evsel__name(evsel));
2986
2987         ++trace->nr_events_printed;
2988 out:
2989         err = 0;
2990 out_put:
2991         thread__put(thread);
2992         addr_location__exit(&al);
2993         return err;
2994 }
2995
2996 static void trace__set_base_time(struct trace *trace,
2997                                  struct evsel *evsel,
2998                                  struct perf_sample *sample)
2999 {
3000         /*
3001          * BPF events were not setting PERF_SAMPLE_TIME, so be more robust
3002          * and don't use sample->time unconditionally, we may end up having
3003          * some other event in the future without PERF_SAMPLE_TIME for good
3004          * reason, i.e. we may not be interested in its timestamps, just in
3005          * it taking place, picking some piece of information when it
3006          * appears in our event stream (vfs_getname comes to mind).
3007          */
3008         if (trace->base_time == 0 && !trace->full_time &&
3009             (evsel->core.attr.sample_type & PERF_SAMPLE_TIME))
3010                 trace->base_time = sample->time;
3011 }
3012
3013 static int trace__process_sample(struct perf_tool *tool,
3014                                  union perf_event *event,
3015                                  struct perf_sample *sample,
3016                                  struct evsel *evsel,
3017                                  struct machine *machine __maybe_unused)
3018 {
3019         struct trace *trace = container_of(tool, struct trace, tool);
3020         struct thread *thread;
3021         int err = 0;
3022
3023         tracepoint_handler handler = evsel->handler;
3024
3025         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
3026         if (thread && thread__is_filtered(thread))
3027                 goto out;
3028
3029         trace__set_base_time(trace, evsel, sample);
3030
3031         if (handler) {
3032                 ++trace->nr_events;
3033                 handler(trace, evsel, event, sample);
3034         }
3035 out:
3036         thread__put(thread);
3037         return err;
3038 }
3039
3040 static int trace__record(struct trace *trace, int argc, const char **argv)
3041 {
3042         unsigned int rec_argc, i, j;
3043         const char **rec_argv;
3044         const char * const record_args[] = {
3045                 "record",
3046                 "-R",
3047                 "-m", "1024",
3048                 "-c", "1",
3049         };
3050         pid_t pid = getpid();
3051         char *filter = asprintf__tp_filter_pids(1, &pid);
3052         const char * const sc_args[] = { "-e", };
3053         unsigned int sc_args_nr = ARRAY_SIZE(sc_args);
3054         const char * const majpf_args[] = { "-e", "major-faults" };
3055         unsigned int majpf_args_nr = ARRAY_SIZE(majpf_args);
3056         const char * const minpf_args[] = { "-e", "minor-faults" };
3057         unsigned int minpf_args_nr = ARRAY_SIZE(minpf_args);
3058         int err = -1;
3059
3060         /* +3 is for the event string below and the pid filter */
3061         rec_argc = ARRAY_SIZE(record_args) + sc_args_nr + 3 +
3062                 majpf_args_nr + minpf_args_nr + argc;
3063         rec_argv = calloc(rec_argc + 1, sizeof(char *));
3064
3065         if (rec_argv == NULL || filter == NULL)
3066                 goto out_free;
3067
3068         j = 0;
3069         for (i = 0; i < ARRAY_SIZE(record_args); i++)
3070                 rec_argv[j++] = record_args[i];
3071
3072         if (trace->trace_syscalls) {
3073                 for (i = 0; i < sc_args_nr; i++)
3074                         rec_argv[j++] = sc_args[i];
3075
3076                 /* event string may be different for older kernels - e.g., RHEL6 */
3077                 if (is_valid_tracepoint("raw_syscalls:sys_enter"))
3078                         rec_argv[j++] = "raw_syscalls:sys_enter,raw_syscalls:sys_exit";
3079                 else if (is_valid_tracepoint("syscalls:sys_enter"))
3080                         rec_argv[j++] = "syscalls:sys_enter,syscalls:sys_exit";
3081                 else {
3082                         pr_err("Neither raw_syscalls nor syscalls events exist.\n");
3083                         goto out_free;
3084                 }
3085         }
3086
3087         rec_argv[j++] = "--filter";
3088         rec_argv[j++] = filter;
3089
3090         if (trace->trace_pgfaults & TRACE_PFMAJ)
3091                 for (i = 0; i < majpf_args_nr; i++)
3092                         rec_argv[j++] = majpf_args[i];
3093
3094         if (trace->trace_pgfaults & TRACE_PFMIN)
3095                 for (i = 0; i < minpf_args_nr; i++)
3096                         rec_argv[j++] = minpf_args[i];
3097
3098         for (i = 0; i < (unsigned int)argc; i++)
3099                 rec_argv[j++] = argv[i];
3100
3101         err = cmd_record(j, rec_argv);
3102 out_free:
3103         free(filter);
3104         free(rec_argv);
3105         return err;
3106 }
3107
3108 static size_t trace__fprintf_thread_summary(struct trace *trace, FILE *fp);
3109
3110 static bool evlist__add_vfs_getname(struct evlist *evlist)
3111 {
3112         bool found = false;
3113         struct evsel *evsel, *tmp;
3114         struct parse_events_error err;
3115         int ret;
3116
3117         parse_events_error__init(&err);
3118         ret = parse_events(evlist, "probe:vfs_getname*", &err);
3119         parse_events_error__exit(&err);
3120         if (ret)
3121                 return false;
3122
3123         evlist__for_each_entry_safe(evlist, evsel, tmp) {
3124                 if (!strstarts(evsel__name(evsel), "probe:vfs_getname"))
3125                         continue;
3126
3127                 if (evsel__field(evsel, "pathname")) {
3128                         evsel->handler = trace__vfs_getname;
3129                         found = true;
3130                         continue;
3131                 }
3132
3133                 list_del_init(&evsel->core.node);
3134                 evsel->evlist = NULL;
3135                 evsel__delete(evsel);
3136         }
3137
3138         return found;
3139 }
3140
3141 static struct evsel *evsel__new_pgfault(u64 config)
3142 {
3143         struct evsel *evsel;
3144         struct perf_event_attr attr = {
3145                 .type = PERF_TYPE_SOFTWARE,
3146                 .mmap_data = 1,
3147         };
3148
3149         attr.config = config;
3150         attr.sample_period = 1;
3151
3152         event_attr_init(&attr);
3153
3154         evsel = evsel__new(&attr);
3155         if (evsel)
3156                 evsel->handler = trace__pgfault;
3157
3158         return evsel;
3159 }
3160
3161 static void evlist__free_syscall_tp_fields(struct evlist *evlist)
3162 {
3163         struct evsel *evsel;
3164
3165         evlist__for_each_entry(evlist, evsel) {
3166                 evsel_trace__delete(evsel->priv);
3167                 evsel->priv = NULL;
3168         }
3169 }
3170
3171 static void trace__handle_event(struct trace *trace, union perf_event *event, struct perf_sample *sample)
3172 {
3173         const u32 type = event->header.type;
3174         struct evsel *evsel;
3175
3176         if (type != PERF_RECORD_SAMPLE) {
3177                 trace__process_event(trace, trace->host, event, sample);
3178                 return;
3179         }
3180
3181         evsel = evlist__id2evsel(trace->evlist, sample->id);
3182         if (evsel == NULL) {
3183                 fprintf(trace->output, "Unknown tp ID %" PRIu64 ", skipping...\n", sample->id);
3184                 return;
3185         }
3186
3187         if (evswitch__discard(&trace->evswitch, evsel))
3188                 return;
3189
3190         trace__set_base_time(trace, evsel, sample);
3191
3192         if (evsel->core.attr.type == PERF_TYPE_TRACEPOINT &&
3193             sample->raw_data == NULL) {
3194                 fprintf(trace->output, "%s sample with no payload for tid: %d, cpu %d, raw_size=%d, skipping...\n",
3195                        evsel__name(evsel), sample->tid,
3196                        sample->cpu, sample->raw_size);
3197         } else {
3198                 tracepoint_handler handler = evsel->handler;
3199                 handler(trace, evsel, event, sample);
3200         }
3201
3202         if (trace->nr_events_printed >= trace->max_events && trace->max_events != ULONG_MAX)
3203                 interrupted = true;
3204 }
3205
3206 static int trace__add_syscall_newtp(struct trace *trace)
3207 {
3208         int ret = -1;
3209         struct evlist *evlist = trace->evlist;
3210         struct evsel *sys_enter, *sys_exit;
3211
3212         sys_enter = perf_evsel__raw_syscall_newtp("sys_enter", trace__sys_enter);
3213         if (sys_enter == NULL)
3214                 goto out;
3215
3216         if (perf_evsel__init_sc_tp_ptr_field(sys_enter, args))
3217                 goto out_delete_sys_enter;
3218
3219         sys_exit = perf_evsel__raw_syscall_newtp("sys_exit", trace__sys_exit);
3220         if (sys_exit == NULL)
3221                 goto out_delete_sys_enter;
3222
3223         if (perf_evsel__init_sc_tp_uint_field(sys_exit, ret))
3224                 goto out_delete_sys_exit;
3225
3226         evsel__config_callchain(sys_enter, &trace->opts, &callchain_param);
3227         evsel__config_callchain(sys_exit, &trace->opts, &callchain_param);
3228
3229         evlist__add(evlist, sys_enter);
3230         evlist__add(evlist, sys_exit);
3231
3232         if (callchain_param.enabled && !trace->kernel_syscallchains) {
3233                 /*
3234                  * We're interested only in the user space callchain
3235                  * leading to the syscall, allow overriding that for
3236                  * debugging reasons using --kernel_syscall_callchains
3237                  */
3238                 sys_exit->core.attr.exclude_callchain_kernel = 1;
3239         }
3240
3241         trace->syscalls.events.sys_enter = sys_enter;
3242         trace->syscalls.events.sys_exit  = sys_exit;
3243
3244         ret = 0;
3245 out:
3246         return ret;
3247
3248 out_delete_sys_exit:
3249         evsel__delete_priv(sys_exit);
3250 out_delete_sys_enter:
3251         evsel__delete_priv(sys_enter);
3252         goto out;
3253 }
3254
3255 static int trace__set_ev_qualifier_tp_filter(struct trace *trace)
3256 {
3257         int err = -1;
3258         struct evsel *sys_exit;
3259         char *filter = asprintf_expr_inout_ints("id", !trace->not_ev_qualifier,
3260                                                 trace->ev_qualifier_ids.nr,
3261                                                 trace->ev_qualifier_ids.entries);
3262
3263         if (filter == NULL)
3264                 goto out_enomem;
3265
3266         if (!evsel__append_tp_filter(trace->syscalls.events.sys_enter, filter)) {
3267                 sys_exit = trace->syscalls.events.sys_exit;
3268                 err = evsel__append_tp_filter(sys_exit, filter);
3269         }
3270
3271         free(filter);
3272 out:
3273         return err;
3274 out_enomem:
3275         errno = ENOMEM;
3276         goto out;
3277 }
3278
3279 #ifdef HAVE_BPF_SKEL
3280 static struct bpf_program *trace__find_bpf_program_by_title(struct trace *trace, const char *name)
3281 {
3282         struct bpf_program *pos, *prog = NULL;
3283         const char *sec_name;
3284
3285         if (trace->skel->obj == NULL)
3286                 return NULL;
3287
3288         bpf_object__for_each_program(pos, trace->skel->obj) {
3289                 sec_name = bpf_program__section_name(pos);
3290                 if (sec_name && !strcmp(sec_name, name)) {
3291                         prog = pos;
3292                         break;
3293                 }
3294         }
3295
3296         return prog;
3297 }
3298
3299 static struct bpf_program *trace__find_syscall_bpf_prog(struct trace *trace, struct syscall *sc,
3300                                                         const char *prog_name, const char *type)
3301 {
3302         struct bpf_program *prog;
3303
3304         if (prog_name == NULL) {
3305                 char default_prog_name[256];
3306                 scnprintf(default_prog_name, sizeof(default_prog_name), "tp/syscalls/sys_%s_%s", type, sc->name);
3307                 prog = trace__find_bpf_program_by_title(trace, default_prog_name);
3308                 if (prog != NULL)
3309                         goto out_found;
3310                 if (sc->fmt && sc->fmt->alias) {
3311                         scnprintf(default_prog_name, sizeof(default_prog_name), "tp/syscalls/sys_%s_%s", type, sc->fmt->alias);
3312                         prog = trace__find_bpf_program_by_title(trace, default_prog_name);
3313                         if (prog != NULL)
3314                                 goto out_found;
3315                 }
3316                 goto out_unaugmented;
3317         }
3318
3319         prog = trace__find_bpf_program_by_title(trace, prog_name);
3320
3321         if (prog != NULL) {
3322 out_found:
3323                 return prog;
3324         }
3325
3326         pr_debug("Couldn't find BPF prog \"%s\" to associate with syscalls:sys_%s_%s, not augmenting it\n",
3327                  prog_name, type, sc->name);
3328 out_unaugmented:
3329         return trace->skel->progs.syscall_unaugmented;
3330 }
3331
3332 static void trace__init_syscall_bpf_progs(struct trace *trace, int id)
3333 {
3334         struct syscall *sc = trace__syscall_info(trace, NULL, id);
3335
3336         if (sc == NULL)
3337                 return;
3338
3339         sc->bpf_prog.sys_enter = trace__find_syscall_bpf_prog(trace, sc, sc->fmt ? sc->fmt->bpf_prog_name.sys_enter : NULL, "enter");
3340         sc->bpf_prog.sys_exit  = trace__find_syscall_bpf_prog(trace, sc, sc->fmt ? sc->fmt->bpf_prog_name.sys_exit  : NULL,  "exit");
3341 }
3342
3343 static int trace__bpf_prog_sys_enter_fd(struct trace *trace, int id)
3344 {
3345         struct syscall *sc = trace__syscall_info(trace, NULL, id);
3346         return sc ? bpf_program__fd(sc->bpf_prog.sys_enter) : bpf_program__fd(trace->skel->progs.syscall_unaugmented);
3347 }
3348
3349 static int trace__bpf_prog_sys_exit_fd(struct trace *trace, int id)
3350 {
3351         struct syscall *sc = trace__syscall_info(trace, NULL, id);
3352         return sc ? bpf_program__fd(sc->bpf_prog.sys_exit) : bpf_program__fd(trace->skel->progs.syscall_unaugmented);
3353 }
3354
3355 static struct bpf_program *trace__find_usable_bpf_prog_entry(struct trace *trace, struct syscall *sc)
3356 {
3357         struct tep_format_field *field, *candidate_field;
3358         int id;
3359
3360         /*
3361          * We're only interested in syscalls that have a pointer:
3362          */
3363         for (field = sc->args; field; field = field->next) {
3364                 if (field->flags & TEP_FIELD_IS_POINTER)
3365                         goto try_to_find_pair;
3366         }
3367
3368         return NULL;
3369
3370 try_to_find_pair:
3371         for (id = 0; id < trace->sctbl->syscalls.nr_entries; ++id) {
3372                 struct syscall *pair = trace__syscall_info(trace, NULL, id);
3373                 struct bpf_program *pair_prog;
3374                 bool is_candidate = false;
3375
3376                 if (pair == NULL || pair == sc ||
3377                     pair->bpf_prog.sys_enter == trace->skel->progs.syscall_unaugmented)
3378                         continue;
3379
3380                 for (field = sc->args, candidate_field = pair->args;
3381                      field && candidate_field; field = field->next, candidate_field = candidate_field->next) {
3382                         bool is_pointer = field->flags & TEP_FIELD_IS_POINTER,
3383                              candidate_is_pointer = candidate_field->flags & TEP_FIELD_IS_POINTER;
3384
3385                         if (is_pointer) {
3386                                if (!candidate_is_pointer) {
3387                                         // The candidate just doesn't copies our pointer arg, might copy other pointers we want.
3388                                         continue;
3389                                }
3390                         } else {
3391                                 if (candidate_is_pointer) {
3392                                         // The candidate might copy a pointer we don't have, skip it.
3393                                         goto next_candidate;
3394                                 }
3395                                 continue;
3396                         }
3397
3398                         if (strcmp(field->type, candidate_field->type))
3399                                 goto next_candidate;
3400
3401                         is_candidate = true;
3402                 }
3403
3404                 if (!is_candidate)
3405                         goto next_candidate;
3406
3407                 /*
3408                  * Check if the tentative pair syscall augmenter has more pointers, if it has,
3409                  * then it may be collecting that and we then can't use it, as it would collect
3410                  * more than what is common to the two syscalls.
3411                  */
3412                 if (candidate_field) {
3413                         for (candidate_field = candidate_field->next; candidate_field; candidate_field = candidate_field->next)
3414                                 if (candidate_field->flags & TEP_FIELD_IS_POINTER)
3415                                         goto next_candidate;
3416                 }
3417
3418                 pair_prog = pair->bpf_prog.sys_enter;
3419                 /*
3420                  * If the pair isn't enabled, then its bpf_prog.sys_enter will not
3421                  * have been searched for, so search it here and if it returns the
3422                  * unaugmented one, then ignore it, otherwise we'll reuse that BPF
3423                  * program for a filtered syscall on a non-filtered one.
3424                  *
3425                  * For instance, we have "!syscalls:sys_enter_renameat" and that is
3426                  * useful for "renameat2".
3427                  */
3428                 if (pair_prog == NULL) {
3429                         pair_prog = trace__find_syscall_bpf_prog(trace, pair, pair->fmt ? pair->fmt->bpf_prog_name.sys_enter : NULL, "enter");
3430                         if (pair_prog == trace->skel->progs.syscall_unaugmented)
3431                                 goto next_candidate;
3432                 }
3433
3434                 pr_debug("Reusing \"%s\" BPF sys_enter augmenter for \"%s\"\n", pair->name, sc->name);
3435                 return pair_prog;
3436         next_candidate:
3437                 continue;
3438         }
3439
3440         return NULL;
3441 }
3442
3443 static int trace__init_syscalls_bpf_prog_array_maps(struct trace *trace)
3444 {
3445         int map_enter_fd = bpf_map__fd(trace->skel->maps.syscalls_sys_enter);
3446         int map_exit_fd  = bpf_map__fd(trace->skel->maps.syscalls_sys_exit);
3447         int err = 0, key;
3448
3449         for (key = 0; key < trace->sctbl->syscalls.nr_entries; ++key) {
3450                 int prog_fd;
3451
3452                 if (!trace__syscall_enabled(trace, key))
3453                         continue;
3454
3455                 trace__init_syscall_bpf_progs(trace, key);
3456
3457                 // It'll get at least the "!raw_syscalls:unaugmented"
3458                 prog_fd = trace__bpf_prog_sys_enter_fd(trace, key);
3459                 err = bpf_map_update_elem(map_enter_fd, &key, &prog_fd, BPF_ANY);
3460                 if (err)
3461                         break;
3462                 prog_fd = trace__bpf_prog_sys_exit_fd(trace, key);
3463                 err = bpf_map_update_elem(map_exit_fd, &key, &prog_fd, BPF_ANY);
3464                 if (err)
3465                         break;
3466         }
3467
3468         /*
3469          * Now lets do a second pass looking for enabled syscalls without
3470          * an augmenter that have a signature that is a superset of another
3471          * syscall with an augmenter so that we can auto-reuse it.
3472          *
3473          * I.e. if we have an augmenter for the "open" syscall that has
3474          * this signature:
3475          *
3476          *   int open(const char *pathname, int flags, mode_t mode);
3477          *
3478          * I.e. that will collect just the first string argument, then we
3479          * can reuse it for the 'creat' syscall, that has this signature:
3480          *
3481          *   int creat(const char *pathname, mode_t mode);
3482          *
3483          * and for:
3484          *
3485          *   int stat(const char *pathname, struct stat *statbuf);
3486          *   int lstat(const char *pathname, struct stat *statbuf);
3487          *
3488          * Because the 'open' augmenter will collect the first arg as a string,
3489          * and leave alone all the other args, which already helps with
3490          * beautifying 'stat' and 'lstat''s pathname arg.
3491          *
3492          * Then, in time, when 'stat' gets an augmenter that collects both
3493          * first and second arg (this one on the raw_syscalls:sys_exit prog
3494          * array tail call, then that one will be used.
3495          */
3496         for (key = 0; key < trace->sctbl->syscalls.nr_entries; ++key) {
3497                 struct syscall *sc = trace__syscall_info(trace, NULL, key);
3498                 struct bpf_program *pair_prog;
3499                 int prog_fd;
3500
3501                 if (sc == NULL || sc->bpf_prog.sys_enter == NULL)
3502                         continue;
3503
3504                 /*
3505                  * For now we're just reusing the sys_enter prog, and if it
3506                  * already has an augmenter, we don't need to find one.
3507                  */
3508                 if (sc->bpf_prog.sys_enter != trace->skel->progs.syscall_unaugmented)
3509                         continue;
3510
3511                 /*
3512                  * Look at all the other syscalls for one that has a signature
3513                  * that is close enough that we can share:
3514                  */
3515                 pair_prog = trace__find_usable_bpf_prog_entry(trace, sc);
3516                 if (pair_prog == NULL)
3517                         continue;
3518
3519                 sc->bpf_prog.sys_enter = pair_prog;
3520
3521                 /*
3522                  * Update the BPF_MAP_TYPE_PROG_SHARED for raw_syscalls:sys_enter
3523                  * with the fd for the program we're reusing:
3524                  */
3525                 prog_fd = bpf_program__fd(sc->bpf_prog.sys_enter);
3526                 err = bpf_map_update_elem(map_enter_fd, &key, &prog_fd, BPF_ANY);
3527                 if (err)
3528                         break;
3529         }
3530
3531         return err;
3532 }
3533 #endif // HAVE_BPF_SKEL
3534
3535 static int trace__set_ev_qualifier_filter(struct trace *trace)
3536 {
3537         if (trace->syscalls.events.sys_enter)
3538                 return trace__set_ev_qualifier_tp_filter(trace);
3539         return 0;
3540 }
3541
3542 static int bpf_map__set_filter_pids(struct bpf_map *map __maybe_unused,
3543                                     size_t npids __maybe_unused, pid_t *pids __maybe_unused)
3544 {
3545         int err = 0;
3546 #ifdef HAVE_LIBBPF_SUPPORT
3547         bool value = true;
3548         int map_fd = bpf_map__fd(map);
3549         size_t i;
3550
3551         for (i = 0; i < npids; ++i) {
3552                 err = bpf_map_update_elem(map_fd, &pids[i], &value, BPF_ANY);
3553                 if (err)
3554                         break;
3555         }
3556 #endif
3557         return err;
3558 }
3559
3560 static int trace__set_filter_loop_pids(struct trace *trace)
3561 {
3562         unsigned int nr = 1, err;
3563         pid_t pids[32] = {
3564                 getpid(),
3565         };
3566         struct thread *thread = machine__find_thread(trace->host, pids[0], pids[0]);
3567
3568         while (thread && nr < ARRAY_SIZE(pids)) {
3569                 struct thread *parent = machine__find_thread(trace->host,
3570                                                              thread__ppid(thread),
3571                                                              thread__ppid(thread));
3572
3573                 if (parent == NULL)
3574                         break;
3575
3576                 if (!strcmp(thread__comm_str(parent), "sshd") ||
3577                     strstarts(thread__comm_str(parent), "gnome-terminal")) {
3578                         pids[nr++] = thread__tid(parent);
3579                         break;
3580                 }
3581                 thread = parent;
3582         }
3583
3584         err = evlist__append_tp_filter_pids(trace->evlist, nr, pids);
3585         if (!err && trace->filter_pids.map)
3586                 err = bpf_map__set_filter_pids(trace->filter_pids.map, nr, pids);
3587
3588         return err;
3589 }
3590
3591 static int trace__set_filter_pids(struct trace *trace)
3592 {
3593         int err = 0;
3594         /*
3595          * Better not use !target__has_task() here because we need to cover the
3596          * case where no threads were specified in the command line, but a
3597          * workload was, and in that case we will fill in the thread_map when
3598          * we fork the workload in evlist__prepare_workload.
3599          */
3600         if (trace->filter_pids.nr > 0) {
3601                 err = evlist__append_tp_filter_pids(trace->evlist, trace->filter_pids.nr,
3602                                                     trace->filter_pids.entries);
3603                 if (!err && trace->filter_pids.map) {
3604                         err = bpf_map__set_filter_pids(trace->filter_pids.map, trace->filter_pids.nr,
3605                                                        trace->filter_pids.entries);
3606                 }
3607         } else if (perf_thread_map__pid(trace->evlist->core.threads, 0) == -1) {
3608                 err = trace__set_filter_loop_pids(trace);
3609         }
3610
3611         return err;
3612 }
3613
3614 static int __trace__deliver_event(struct trace *trace, union perf_event *event)
3615 {
3616         struct evlist *evlist = trace->evlist;
3617         struct perf_sample sample;
3618         int err = evlist__parse_sample(evlist, event, &sample);
3619
3620         if (err)
3621                 fprintf(trace->output, "Can't parse sample, err = %d, skipping...\n", err);
3622         else
3623                 trace__handle_event(trace, event, &sample);
3624
3625         return 0;
3626 }
3627
3628 static int __trace__flush_events(struct trace *trace)
3629 {
3630         u64 first = ordered_events__first_time(&trace->oe.data);
3631         u64 flush = trace->oe.last - NSEC_PER_SEC;
3632
3633         /* Is there some thing to flush.. */
3634         if (first && first < flush)
3635                 return ordered_events__flush_time(&trace->oe.data, flush);
3636
3637         return 0;
3638 }
3639
3640 static int trace__flush_events(struct trace *trace)
3641 {
3642         return !trace->sort_events ? 0 : __trace__flush_events(trace);
3643 }
3644
3645 static int trace__deliver_event(struct trace *trace, union perf_event *event)
3646 {
3647         int err;
3648
3649         if (!trace->sort_events)
3650                 return __trace__deliver_event(trace, event);
3651
3652         err = evlist__parse_sample_timestamp(trace->evlist, event, &trace->oe.last);
3653         if (err && err != -1)
3654                 return err;
3655
3656         err = ordered_events__queue(&trace->oe.data, event, trace->oe.last, 0, NULL);
3657         if (err)
3658                 return err;
3659
3660         return trace__flush_events(trace);
3661 }
3662
3663 static int ordered_events__deliver_event(struct ordered_events *oe,
3664                                          struct ordered_event *event)
3665 {
3666         struct trace *trace = container_of(oe, struct trace, oe.data);
3667
3668         return __trace__deliver_event(trace, event->event);
3669 }
3670
3671 static struct syscall_arg_fmt *evsel__find_syscall_arg_fmt_by_name(struct evsel *evsel, char *arg)
3672 {
3673         struct tep_format_field *field;
3674         struct syscall_arg_fmt *fmt = __evsel__syscall_arg_fmt(evsel);
3675
3676         if (evsel->tp_format == NULL || fmt == NULL)
3677                 return NULL;
3678
3679         for (field = evsel->tp_format->format.fields; field; field = field->next, ++fmt)
3680                 if (strcmp(field->name, arg) == 0)
3681                         return fmt;
3682
3683         return NULL;
3684 }
3685
3686 static int trace__expand_filter(struct trace *trace __maybe_unused, struct evsel *evsel)
3687 {
3688         char *tok, *left = evsel->filter, *new_filter = evsel->filter;
3689
3690         while ((tok = strpbrk(left, "=<>!")) != NULL) {
3691                 char *right = tok + 1, *right_end;
3692
3693                 if (*right == '=')
3694                         ++right;
3695
3696                 while (isspace(*right))
3697                         ++right;
3698
3699                 if (*right == '\0')
3700                         break;
3701
3702                 while (!isalpha(*left))
3703                         if (++left == tok) {
3704                                 /*
3705                                  * Bail out, can't find the name of the argument that is being
3706                                  * used in the filter, let it try to set this filter, will fail later.
3707                                  */
3708                                 return 0;
3709                         }
3710
3711                 right_end = right + 1;
3712                 while (isalnum(*right_end) || *right_end == '_' || *right_end == '|')
3713                         ++right_end;
3714
3715                 if (isalpha(*right)) {
3716                         struct syscall_arg_fmt *fmt;
3717                         int left_size = tok - left,
3718                             right_size = right_end - right;
3719                         char arg[128];
3720
3721                         while (isspace(left[left_size - 1]))
3722                                 --left_size;
3723
3724                         scnprintf(arg, sizeof(arg), "%.*s", left_size, left);
3725
3726                         fmt = evsel__find_syscall_arg_fmt_by_name(evsel, arg);
3727                         if (fmt == NULL) {
3728                                 pr_err("\"%s\" not found in \"%s\", can't set filter \"%s\"\n",
3729                                        arg, evsel->name, evsel->filter);
3730                                 return -1;
3731                         }
3732
3733                         pr_debug2("trying to expand \"%s\" \"%.*s\" \"%.*s\" -> ",
3734                                  arg, (int)(right - tok), tok, right_size, right);
3735
3736                         if (fmt->strtoul) {
3737                                 u64 val;
3738                                 struct syscall_arg syscall_arg = {
3739                                         .parm = fmt->parm,
3740                                 };
3741
3742                                 if (fmt->strtoul(right, right_size, &syscall_arg, &val)) {
3743                                         char *n, expansion[19];
3744                                         int expansion_lenght = scnprintf(expansion, sizeof(expansion), "%#" PRIx64, val);
3745                                         int expansion_offset = right - new_filter;
3746
3747                                         pr_debug("%s", expansion);
3748
3749                                         if (asprintf(&n, "%.*s%s%s", expansion_offset, new_filter, expansion, right_end) < 0) {
3750                                                 pr_debug(" out of memory!\n");
3751                                                 free(new_filter);
3752                                                 return -1;
3753                                         }
3754                                         if (new_filter != evsel->filter)
3755                                                 free(new_filter);
3756                                         left = n + expansion_offset + expansion_lenght;
3757                                         new_filter = n;
3758                                 } else {
3759                                         pr_err("\"%.*s\" not found for \"%s\" in \"%s\", can't set filter \"%s\"\n",
3760                                                right_size, right, arg, evsel->name, evsel->filter);
3761                                         return -1;
3762                                 }
3763                         } else {
3764                                 pr_err("No resolver (strtoul) for \"%s\" in \"%s\", can't set filter \"%s\"\n",
3765                                        arg, evsel->name, evsel->filter);
3766                                 return -1;
3767                         }
3768
3769                         pr_debug("\n");
3770                 } else {
3771                         left = right_end;
3772                 }
3773         }
3774
3775         if (new_filter != evsel->filter) {
3776                 pr_debug("New filter for %s: %s\n", evsel->name, new_filter);
3777                 evsel__set_filter(evsel, new_filter);
3778                 free(new_filter);
3779         }
3780
3781         return 0;
3782 }
3783
3784 static int trace__expand_filters(struct trace *trace, struct evsel **err_evsel)
3785 {
3786         struct evlist *evlist = trace->evlist;
3787         struct evsel *evsel;
3788
3789         evlist__for_each_entry(evlist, evsel) {
3790                 if (evsel->filter == NULL)
3791                         continue;
3792
3793                 if (trace__expand_filter(trace, evsel)) {
3794                         *err_evsel = evsel;
3795                         return -1;
3796                 }
3797         }
3798
3799         return 0;
3800 }
3801
3802 static int trace__run(struct trace *trace, int argc, const char **argv)
3803 {
3804         struct evlist *evlist = trace->evlist;
3805         struct evsel *evsel, *pgfault_maj = NULL, *pgfault_min = NULL;
3806         int err = -1, i;
3807         unsigned long before;
3808         const bool forks = argc > 0;
3809         bool draining = false;
3810
3811         trace->live = true;
3812
3813         if (!trace->raw_augmented_syscalls) {
3814                 if (trace->trace_syscalls && trace__add_syscall_newtp(trace))
3815                         goto out_error_raw_syscalls;
3816
3817                 if (trace->trace_syscalls)
3818                         trace->vfs_getname = evlist__add_vfs_getname(evlist);
3819         }
3820
3821         if ((trace->trace_pgfaults & TRACE_PFMAJ)) {
3822                 pgfault_maj = evsel__new_pgfault(PERF_COUNT_SW_PAGE_FAULTS_MAJ);
3823                 if (pgfault_maj == NULL)
3824                         goto out_error_mem;
3825                 evsel__config_callchain(pgfault_maj, &trace->opts, &callchain_param);
3826                 evlist__add(evlist, pgfault_maj);
3827         }
3828
3829         if ((trace->trace_pgfaults & TRACE_PFMIN)) {
3830                 pgfault_min = evsel__new_pgfault(PERF_COUNT_SW_PAGE_FAULTS_MIN);
3831                 if (pgfault_min == NULL)
3832                         goto out_error_mem;
3833                 evsel__config_callchain(pgfault_min, &trace->opts, &callchain_param);
3834                 evlist__add(evlist, pgfault_min);
3835         }
3836
3837         /* Enable ignoring missing threads when -u/-p option is defined. */
3838         trace->opts.ignore_missing_thread = trace->opts.target.uid != UINT_MAX || trace->opts.target.pid;
3839
3840         if (trace->sched &&
3841             evlist__add_newtp(evlist, "sched", "sched_stat_runtime", trace__sched_stat_runtime))
3842                 goto out_error_sched_stat_runtime;
3843         /*
3844          * If a global cgroup was set, apply it to all the events without an
3845          * explicit cgroup. I.e.:
3846          *
3847          *      trace -G A -e sched:*switch
3848          *
3849          * Will set all raw_syscalls:sys_{enter,exit}, pgfault, vfs_getname, etc
3850          * _and_ sched:sched_switch to the 'A' cgroup, while:
3851          *
3852          * trace -e sched:*switch -G A
3853          *
3854          * will only set the sched:sched_switch event to the 'A' cgroup, all the
3855          * other events (raw_syscalls:sys_{enter,exit}, etc are left "without"
3856          * a cgroup (on the root cgroup, sys wide, etc).
3857          *
3858          * Multiple cgroups:
3859          *
3860          * trace -G A -e sched:*switch -G B
3861          *
3862          * the syscall ones go to the 'A' cgroup, the sched:sched_switch goes
3863          * to the 'B' cgroup.
3864          *
3865          * evlist__set_default_cgroup() grabs a reference of the passed cgroup
3866          * only for the evsels still without a cgroup, i.e. evsel->cgroup == NULL.
3867          */
3868         if (trace->cgroup)
3869                 evlist__set_default_cgroup(trace->evlist, trace->cgroup);
3870
3871         err = evlist__create_maps(evlist, &trace->opts.target);
3872         if (err < 0) {
3873                 fprintf(trace->output, "Problems parsing the target to trace, check your options!\n");
3874                 goto out_delete_evlist;
3875         }
3876
3877         err = trace__symbols_init(trace, evlist);
3878         if (err < 0) {
3879                 fprintf(trace->output, "Problems initializing symbol libraries!\n");
3880                 goto out_delete_evlist;
3881         }
3882
3883         evlist__config(evlist, &trace->opts, &callchain_param);
3884
3885         if (forks) {
3886                 err = evlist__prepare_workload(evlist, &trace->opts.target, argv, false, NULL);
3887                 if (err < 0) {
3888                         fprintf(trace->output, "Couldn't run the workload!\n");
3889                         goto out_delete_evlist;
3890                 }
3891                 workload_pid = evlist->workload.pid;
3892         }
3893
3894         err = evlist__open(evlist);
3895         if (err < 0)
3896                 goto out_error_open;
3897 #ifdef HAVE_BPF_SKEL
3898         {
3899                 struct perf_cpu cpu;
3900
3901                 /*
3902                  * Set up the __augmented_syscalls__ BPF map to hold for each
3903                  * CPU the bpf-output event's file descriptor.
3904                  */
3905                 perf_cpu_map__for_each_cpu(cpu, i, trace->syscalls.events.bpf_output->core.cpus) {
3906                         bpf_map__update_elem(trace->skel->maps.__augmented_syscalls__,
3907                                         &cpu.cpu, sizeof(int),
3908                                         xyarray__entry(trace->syscalls.events.bpf_output->core.fd,
3909                                                        cpu.cpu, 0),
3910                                         sizeof(__u32), BPF_ANY);
3911                 }
3912         }
3913 #endif
3914         err = trace__set_filter_pids(trace);
3915         if (err < 0)
3916                 goto out_error_mem;
3917
3918 #ifdef HAVE_BPF_SKEL
3919         if (trace->skel->progs.sys_enter)
3920                 trace__init_syscalls_bpf_prog_array_maps(trace);
3921 #endif
3922
3923         if (trace->ev_qualifier_ids.nr > 0) {
3924                 err = trace__set_ev_qualifier_filter(trace);
3925                 if (err < 0)
3926                         goto out_errno;
3927
3928                 if (trace->syscalls.events.sys_exit) {
3929                         pr_debug("event qualifier tracepoint filter: %s\n",
3930                                  trace->syscalls.events.sys_exit->filter);
3931                 }
3932         }
3933
3934         /*
3935          * If the "close" syscall is not traced, then we will not have the
3936          * opportunity to, in syscall_arg__scnprintf_close_fd() invalidate the
3937          * fd->pathname table and were ending up showing the last value set by
3938          * syscalls opening a pathname and associating it with a descriptor or
3939          * reading it from /proc/pid/fd/ in cases where that doesn't make
3940          * sense.
3941          *
3942          *  So just disable this beautifier (SCA_FD, SCA_FDAT) when 'close' is
3943          *  not in use.
3944          */
3945         trace->fd_path_disabled = !trace__syscall_enabled(trace, syscalltbl__id(trace->sctbl, "close"));
3946
3947         err = trace__expand_filters(trace, &evsel);
3948         if (err)
3949                 goto out_delete_evlist;
3950         err = evlist__apply_filters(evlist, &evsel);
3951         if (err < 0)
3952                 goto out_error_apply_filters;
3953
3954         err = evlist__mmap(evlist, trace->opts.mmap_pages);
3955         if (err < 0)
3956                 goto out_error_mmap;
3957
3958         if (!target__none(&trace->opts.target) && !trace->opts.target.initial_delay)
3959                 evlist__enable(evlist);
3960
3961         if (forks)
3962                 evlist__start_workload(evlist);
3963
3964         if (trace->opts.target.initial_delay) {
3965                 usleep(trace->opts.target.initial_delay * 1000);
3966                 evlist__enable(evlist);
3967         }
3968
3969         trace->multiple_threads = perf_thread_map__pid(evlist->core.threads, 0) == -1 ||
3970                 perf_thread_map__nr(evlist->core.threads) > 1 ||
3971                 evlist__first(evlist)->core.attr.inherit;
3972
3973         /*
3974          * Now that we already used evsel->core.attr to ask the kernel to setup the
3975          * events, lets reuse evsel->core.attr.sample_max_stack as the limit in
3976          * trace__resolve_callchain(), allowing per-event max-stack settings
3977          * to override an explicitly set --max-stack global setting.
3978          */
3979         evlist__for_each_entry(evlist, evsel) {
3980                 if (evsel__has_callchain(evsel) &&
3981                     evsel->core.attr.sample_max_stack == 0)
3982                         evsel->core.attr.sample_max_stack = trace->max_stack;
3983         }
3984 again:
3985         before = trace->nr_events;
3986
3987         for (i = 0; i < evlist->core.nr_mmaps; i++) {
3988                 union perf_event *event;
3989                 struct mmap *md;
3990
3991                 md = &evlist->mmap[i];
3992                 if (perf_mmap__read_init(&md->core) < 0)
3993                         continue;
3994
3995                 while ((event = perf_mmap__read_event(&md->core)) != NULL) {
3996                         ++trace->nr_events;
3997
3998                         err = trace__deliver_event(trace, event);
3999                         if (err)
4000                                 goto out_disable;
4001
4002                         perf_mmap__consume(&md->core);
4003
4004                         if (interrupted)
4005                                 goto out_disable;
4006
4007                         if (done && !draining) {
4008                                 evlist__disable(evlist);
4009                                 draining = true;
4010                         }
4011                 }
4012                 perf_mmap__read_done(&md->core);
4013         }
4014
4015         if (trace->nr_events == before) {
4016                 int timeout = done ? 100 : -1;
4017
4018                 if (!draining && evlist__poll(evlist, timeout) > 0) {
4019                         if (evlist__filter_pollfd(evlist, POLLERR | POLLHUP | POLLNVAL) == 0)
4020                                 draining = true;
4021
4022                         goto again;
4023                 } else {
4024                         if (trace__flush_events(trace))
4025                                 goto out_disable;
4026                 }
4027         } else {
4028                 goto again;
4029         }
4030
4031 out_disable:
4032         thread__zput(trace->current);
4033
4034         evlist__disable(evlist);
4035
4036         if (trace->sort_events)
4037                 ordered_events__flush(&trace->oe.data, OE_FLUSH__FINAL);
4038
4039         if (!err) {
4040                 if (trace->summary)
4041                         trace__fprintf_thread_summary(trace, trace->output);
4042
4043                 if (trace->show_tool_stats) {
4044                         fprintf(trace->output, "Stats:\n "
4045                                                " vfs_getname : %" PRIu64 "\n"
4046                                                " proc_getname: %" PRIu64 "\n",
4047                                 trace->stats.vfs_getname,
4048                                 trace->stats.proc_getname);
4049                 }
4050         }
4051
4052 out_delete_evlist:
4053         trace__symbols__exit(trace);
4054         evlist__free_syscall_tp_fields(evlist);
4055         evlist__delete(evlist);
4056         cgroup__put(trace->cgroup);
4057         trace->evlist = NULL;
4058         trace->live = false;
4059         return err;
4060 {
4061         char errbuf[BUFSIZ];
4062
4063 out_error_sched_stat_runtime:
4064         tracing_path__strerror_open_tp(errno, errbuf, sizeof(errbuf), "sched", "sched_stat_runtime");
4065         goto out_error;
4066
4067 out_error_raw_syscalls:
4068         tracing_path__strerror_open_tp(errno, errbuf, sizeof(errbuf), "raw_syscalls", "sys_(enter|exit)");
4069         goto out_error;
4070
4071 out_error_mmap:
4072         evlist__strerror_mmap(evlist, errno, errbuf, sizeof(errbuf));
4073         goto out_error;
4074
4075 out_error_open:
4076         evlist__strerror_open(evlist, errno, errbuf, sizeof(errbuf));
4077
4078 out_error:
4079         fprintf(trace->output, "%s\n", errbuf);
4080         goto out_delete_evlist;
4081
4082 out_error_apply_filters:
4083         fprintf(trace->output,
4084                 "Failed to set filter \"%s\" on event %s with %d (%s)\n",
4085                 evsel->filter, evsel__name(evsel), errno,
4086                 str_error_r(errno, errbuf, sizeof(errbuf)));
4087         goto out_delete_evlist;
4088 }
4089 out_error_mem:
4090         fprintf(trace->output, "Not enough memory to run!\n");
4091         goto out_delete_evlist;
4092
4093 out_errno:
4094         fprintf(trace->output, "errno=%d,%s\n", errno, strerror(errno));
4095         goto out_delete_evlist;
4096 }
4097
4098 static int trace__replay(struct trace *trace)
4099 {
4100         const struct evsel_str_handler handlers[] = {
4101                 { "probe:vfs_getname",       trace__vfs_getname, },
4102         };
4103         struct perf_data data = {
4104                 .path  = input_name,
4105                 .mode  = PERF_DATA_MODE_READ,
4106                 .force = trace->force,
4107         };
4108         struct perf_session *session;
4109         struct evsel *evsel;
4110         int err = -1;
4111
4112         trace->tool.sample        = trace__process_sample;
4113         trace->tool.mmap          = perf_event__process_mmap;
4114         trace->tool.mmap2         = perf_event__process_mmap2;
4115         trace->tool.comm          = perf_event__process_comm;
4116         trace->tool.exit          = perf_event__process_exit;
4117         trace->tool.fork          = perf_event__process_fork;
4118         trace->tool.attr          = perf_event__process_attr;
4119         trace->tool.tracing_data  = perf_event__process_tracing_data;
4120         trace->tool.build_id      = perf_event__process_build_id;
4121         trace->tool.namespaces    = perf_event__process_namespaces;
4122
4123         trace->tool.ordered_events = true;
4124         trace->tool.ordering_requires_timestamps = true;
4125
4126         /* add tid to output */
4127         trace->multiple_threads = true;
4128
4129         session = perf_session__new(&data, &trace->tool);
4130         if (IS_ERR(session))
4131                 return PTR_ERR(session);
4132
4133         if (trace->opts.target.pid)
4134                 symbol_conf.pid_list_str = strdup(trace->opts.target.pid);
4135
4136         if (trace->opts.target.tid)
4137                 symbol_conf.tid_list_str = strdup(trace->opts.target.tid);
4138
4139         if (symbol__init(&session->header.env) < 0)
4140                 goto out;
4141
4142         trace->host = &session->machines.host;
4143
4144         err = perf_session__set_tracepoints_handlers(session, handlers);
4145         if (err)
4146                 goto out;
4147
4148         evsel = evlist__find_tracepoint_by_name(session->evlist, "raw_syscalls:sys_enter");
4149         trace->syscalls.events.sys_enter = evsel;
4150         /* older kernels have syscalls tp versus raw_syscalls */
4151         if (evsel == NULL)
4152                 evsel = evlist__find_tracepoint_by_name(session->evlist, "syscalls:sys_enter");
4153
4154         if (evsel &&
4155             (evsel__init_raw_syscall_tp(evsel, trace__sys_enter) < 0 ||
4156             perf_evsel__init_sc_tp_ptr_field(evsel, args))) {
4157                 pr_err("Error during initialize raw_syscalls:sys_enter event\n");
4158                 goto out;
4159         }
4160
4161         evsel = evlist__find_tracepoint_by_name(session->evlist, "raw_syscalls:sys_exit");
4162         trace->syscalls.events.sys_exit = evsel;
4163         if (evsel == NULL)
4164                 evsel = evlist__find_tracepoint_by_name(session->evlist, "syscalls:sys_exit");
4165         if (evsel &&
4166             (evsel__init_raw_syscall_tp(evsel, trace__sys_exit) < 0 ||
4167             perf_evsel__init_sc_tp_uint_field(evsel, ret))) {
4168                 pr_err("Error during initialize raw_syscalls:sys_exit event\n");
4169                 goto out;
4170         }
4171
4172         evlist__for_each_entry(session->evlist, evsel) {
4173                 if (evsel->core.attr.type == PERF_TYPE_SOFTWARE &&
4174                     (evsel->core.attr.config == PERF_COUNT_SW_PAGE_FAULTS_MAJ ||
4175                      evsel->core.attr.config == PERF_COUNT_SW_PAGE_FAULTS_MIN ||
4176                      evsel->core.attr.config == PERF_COUNT_SW_PAGE_FAULTS))
4177                         evsel->handler = trace__pgfault;
4178         }
4179
4180         setup_pager();
4181
4182         err = perf_session__process_events(session);
4183         if (err)
4184                 pr_err("Failed to process events, error %d", err);
4185
4186         else if (trace->summary)
4187                 trace__fprintf_thread_summary(trace, trace->output);
4188
4189 out:
4190         perf_session__delete(session);
4191
4192         return err;
4193 }
4194
4195 static size_t trace__fprintf_threads_header(FILE *fp)
4196 {
4197         size_t printed;
4198
4199         printed  = fprintf(fp, "\n Summary of events:\n\n");
4200
4201         return printed;
4202 }
4203
4204 DEFINE_RESORT_RB(syscall_stats, a->msecs > b->msecs,
4205         struct syscall_stats *stats;
4206         double               msecs;
4207         int                  syscall;
4208 )
4209 {
4210         struct int_node *source = rb_entry(nd, struct int_node, rb_node);
4211         struct syscall_stats *stats = source->priv;
4212
4213         entry->syscall = source->i;
4214         entry->stats   = stats;
4215         entry->msecs   = stats ? (u64)stats->stats.n * (avg_stats(&stats->stats) / NSEC_PER_MSEC) : 0;
4216 }
4217
4218 static size_t thread__dump_stats(struct thread_trace *ttrace,
4219                                  struct trace *trace, FILE *fp)
4220 {
4221         size_t printed = 0;
4222         struct syscall *sc;
4223         struct rb_node *nd;
4224         DECLARE_RESORT_RB_INTLIST(syscall_stats, ttrace->syscall_stats);
4225
4226         if (syscall_stats == NULL)
4227                 return 0;
4228
4229         printed += fprintf(fp, "\n");
4230
4231         printed += fprintf(fp, "   syscall            calls  errors  total       min       avg       max       stddev\n");
4232         printed += fprintf(fp, "                                     (msec)    (msec)    (msec)    (msec)        (%%)\n");
4233         printed += fprintf(fp, "   --------------- --------  ------ -------- --------- --------- ---------     ------\n");
4234
4235         resort_rb__for_each_entry(nd, syscall_stats) {
4236                 struct syscall_stats *stats = syscall_stats_entry->stats;
4237                 if (stats) {
4238                         double min = (double)(stats->stats.min) / NSEC_PER_MSEC;
4239                         double max = (double)(stats->stats.max) / NSEC_PER_MSEC;
4240                         double avg = avg_stats(&stats->stats);
4241                         double pct;
4242                         u64 n = (u64)stats->stats.n;
4243
4244                         pct = avg ? 100.0 * stddev_stats(&stats->stats) / avg : 0.0;
4245                         avg /= NSEC_PER_MSEC;
4246
4247                         sc = &trace->syscalls.table[syscall_stats_entry->syscall];
4248                         printed += fprintf(fp, "   %-15s", sc->name);
4249                         printed += fprintf(fp, " %8" PRIu64 " %6" PRIu64 " %9.3f %9.3f %9.3f",
4250                                            n, stats->nr_failures, syscall_stats_entry->msecs, min, avg);
4251                         printed += fprintf(fp, " %9.3f %9.2f%%\n", max, pct);
4252
4253                         if (trace->errno_summary && stats->nr_failures) {
4254                                 const char *arch_name = perf_env__arch(trace->host->env);
4255                                 int e;
4256
4257                                 for (e = 0; e < stats->max_errno; ++e) {
4258                                         if (stats->errnos[e] != 0)
4259                                                 fprintf(fp, "\t\t\t\t%s: %d\n", arch_syscalls__strerrno(arch_name, e + 1), stats->errnos[e]);
4260                                 }
4261                         }
4262                 }
4263         }
4264
4265         resort_rb__delete(syscall_stats);
4266         printed += fprintf(fp, "\n\n");
4267
4268         return printed;
4269 }
4270
4271 static size_t trace__fprintf_thread(FILE *fp, struct thread *thread, struct trace *trace)
4272 {
4273         size_t printed = 0;
4274         struct thread_trace *ttrace = thread__priv(thread);
4275         double ratio;
4276
4277         if (ttrace == NULL)
4278                 return 0;
4279
4280         ratio = (double)ttrace->nr_events / trace->nr_events * 100.0;
4281
4282         printed += fprintf(fp, " %s (%d), ", thread__comm_str(thread), thread__tid(thread));
4283         printed += fprintf(fp, "%lu events, ", ttrace->nr_events);
4284         printed += fprintf(fp, "%.1f%%", ratio);
4285         if (ttrace->pfmaj)
4286                 printed += fprintf(fp, ", %lu majfaults", ttrace->pfmaj);
4287         if (ttrace->pfmin)
4288                 printed += fprintf(fp, ", %lu minfaults", ttrace->pfmin);
4289         if (trace->sched)
4290                 printed += fprintf(fp, ", %.3f msec\n", ttrace->runtime_ms);
4291         else if (fputc('\n', fp) != EOF)
4292                 ++printed;
4293
4294         printed += thread__dump_stats(ttrace, trace, fp);
4295
4296         return printed;
4297 }
4298
4299 static unsigned long thread__nr_events(struct thread_trace *ttrace)
4300 {
4301         return ttrace ? ttrace->nr_events : 0;
4302 }
4303
4304 DEFINE_RESORT_RB(threads,
4305                 (thread__nr_events(thread__priv(a->thread)) <
4306                  thread__nr_events(thread__priv(b->thread))),
4307         struct thread *thread;
4308 )
4309 {
4310         entry->thread = rb_entry(nd, struct thread_rb_node, rb_node)->thread;
4311 }
4312
4313 static size_t trace__fprintf_thread_summary(struct trace *trace, FILE *fp)
4314 {
4315         size_t printed = trace__fprintf_threads_header(fp);
4316         struct rb_node *nd;
4317         int i;
4318
4319         for (i = 0; i < THREADS__TABLE_SIZE; i++) {
4320                 DECLARE_RESORT_RB_MACHINE_THREADS(threads, trace->host, i);
4321
4322                 if (threads == NULL) {
4323                         fprintf(fp, "%s", "Error sorting output by nr_events!\n");
4324                         return 0;
4325                 }
4326
4327                 resort_rb__for_each_entry(nd, threads)
4328                         printed += trace__fprintf_thread(fp, threads_entry->thread, trace);
4329
4330                 resort_rb__delete(threads);
4331         }
4332         return printed;
4333 }
4334
4335 static int trace__set_duration(const struct option *opt, const char *str,
4336                                int unset __maybe_unused)
4337 {
4338         struct trace *trace = opt->value;
4339
4340         trace->duration_filter = atof(str);
4341         return 0;
4342 }
4343
4344 static int trace__set_filter_pids_from_option(const struct option *opt, const char *str,
4345                                               int unset __maybe_unused)
4346 {
4347         int ret = -1;
4348         size_t i;
4349         struct trace *trace = opt->value;
4350         /*
4351          * FIXME: introduce a intarray class, plain parse csv and create a
4352          * { int nr, int entries[] } struct...
4353          */
4354         struct intlist *list = intlist__new(str);
4355
4356         if (list == NULL)
4357                 return -1;
4358
4359         i = trace->filter_pids.nr = intlist__nr_entries(list) + 1;
4360         trace->filter_pids.entries = calloc(i, sizeof(pid_t));
4361
4362         if (trace->filter_pids.entries == NULL)
4363                 goto out;
4364
4365         trace->filter_pids.entries[0] = getpid();
4366
4367         for (i = 1; i < trace->filter_pids.nr; ++i)
4368                 trace->filter_pids.entries[i] = intlist__entry(list, i - 1)->i;
4369
4370         intlist__delete(list);
4371         ret = 0;
4372 out:
4373         return ret;
4374 }
4375
4376 static int trace__open_output(struct trace *trace, const char *filename)
4377 {
4378         struct stat st;
4379
4380         if (!stat(filename, &st) && st.st_size) {
4381                 char oldname[PATH_MAX];
4382
4383                 scnprintf(oldname, sizeof(oldname), "%s.old", filename);
4384                 unlink(oldname);
4385                 rename(filename, oldname);
4386         }
4387
4388         trace->output = fopen(filename, "w");
4389
4390         return trace->output == NULL ? -errno : 0;
4391 }
4392
4393 static int parse_pagefaults(const struct option *opt, const char *str,
4394                             int unset __maybe_unused)
4395 {
4396         int *trace_pgfaults = opt->value;
4397
4398         if (strcmp(str, "all") == 0)
4399                 *trace_pgfaults |= TRACE_PFMAJ | TRACE_PFMIN;
4400         else if (strcmp(str, "maj") == 0)
4401                 *trace_pgfaults |= TRACE_PFMAJ;
4402         else if (strcmp(str, "min") == 0)
4403                 *trace_pgfaults |= TRACE_PFMIN;
4404         else
4405                 return -1;
4406
4407         return 0;
4408 }
4409
4410 static void evlist__set_default_evsel_handler(struct evlist *evlist, void *handler)
4411 {
4412         struct evsel *evsel;
4413
4414         evlist__for_each_entry(evlist, evsel) {
4415                 if (evsel->handler == NULL)
4416                         evsel->handler = handler;
4417         }
4418 }
4419
4420 static void evsel__set_syscall_arg_fmt(struct evsel *evsel, const char *name)
4421 {
4422         struct syscall_arg_fmt *fmt = evsel__syscall_arg_fmt(evsel);
4423
4424         if (fmt) {
4425                 const struct syscall_fmt *scfmt = syscall_fmt__find(name);
4426
4427                 if (scfmt) {
4428                         int skip = 0;
4429
4430                         if (strcmp(evsel->tp_format->format.fields->name, "__syscall_nr") == 0 ||
4431                             strcmp(evsel->tp_format->format.fields->name, "nr") == 0)
4432                                 ++skip;
4433
4434                         memcpy(fmt + skip, scfmt->arg, (evsel->tp_format->format.nr_fields - skip) * sizeof(*fmt));
4435                 }
4436         }
4437 }
4438
4439 static int evlist__set_syscall_tp_fields(struct evlist *evlist)
4440 {
4441         struct evsel *evsel;
4442
4443         evlist__for_each_entry(evlist, evsel) {
4444                 if (evsel->priv || !evsel->tp_format)
4445                         continue;
4446
4447                 if (strcmp(evsel->tp_format->system, "syscalls")) {
4448                         evsel__init_tp_arg_scnprintf(evsel);
4449                         continue;
4450                 }
4451
4452                 if (evsel__init_syscall_tp(evsel))
4453                         return -1;
4454
4455                 if (!strncmp(evsel->tp_format->name, "sys_enter_", 10)) {
4456                         struct syscall_tp *sc = __evsel__syscall_tp(evsel);
4457
4458                         if (__tp_field__init_ptr(&sc->args, sc->id.offset + sizeof(u64)))
4459                                 return -1;
4460
4461                         evsel__set_syscall_arg_fmt(evsel, evsel->tp_format->name + sizeof("sys_enter_") - 1);
4462                 } else if (!strncmp(evsel->tp_format->name, "sys_exit_", 9)) {
4463                         struct syscall_tp *sc = __evsel__syscall_tp(evsel);
4464
4465                         if (__tp_field__init_uint(&sc->ret, sizeof(u64), sc->id.offset + sizeof(u64), evsel->needs_swap))
4466                                 return -1;
4467
4468                         evsel__set_syscall_arg_fmt(evsel, evsel->tp_format->name + sizeof("sys_exit_") - 1);
4469                 }
4470         }
4471
4472         return 0;
4473 }
4474
4475 /*
4476  * XXX: Hackish, just splitting the combined -e+--event (syscalls
4477  * (raw_syscalls:{sys_{enter,exit}} + events (tracepoints, HW, SW, etc) to use
4478  * existing facilities unchanged (trace->ev_qualifier + parse_options()).
4479  *
4480  * It'd be better to introduce a parse_options() variant that would return a
4481  * list with the terms it didn't match to an event...
4482  */
4483 static int trace__parse_events_option(const struct option *opt, const char *str,
4484                                       int unset __maybe_unused)
4485 {
4486         struct trace *trace = (struct trace *)opt->value;
4487         const char *s = str;
4488         char *sep = NULL, *lists[2] = { NULL, NULL, };
4489         int len = strlen(str) + 1, err = -1, list, idx;
4490         char *strace_groups_dir = system_path(STRACE_GROUPS_DIR);
4491         char group_name[PATH_MAX];
4492         const struct syscall_fmt *fmt;
4493
4494         if (strace_groups_dir == NULL)
4495                 return -1;
4496
4497         if (*s == '!') {
4498                 ++s;
4499                 trace->not_ev_qualifier = true;
4500         }
4501
4502         while (1) {
4503                 if ((sep = strchr(s, ',')) != NULL)
4504                         *sep = '\0';
4505
4506                 list = 0;
4507                 if (syscalltbl__id(trace->sctbl, s) >= 0 ||
4508                     syscalltbl__strglobmatch_first(trace->sctbl, s, &idx) >= 0) {
4509                         list = 1;
4510                         goto do_concat;
4511                 }
4512
4513                 fmt = syscall_fmt__find_by_alias(s);
4514                 if (fmt != NULL) {
4515                         list = 1;
4516                         s = fmt->name;
4517                 } else {
4518                         path__join(group_name, sizeof(group_name), strace_groups_dir, s);
4519                         if (access(group_name, R_OK) == 0)
4520                                 list = 1;
4521                 }
4522 do_concat:
4523                 if (lists[list]) {
4524                         sprintf(lists[list] + strlen(lists[list]), ",%s", s);
4525                 } else {
4526                         lists[list] = malloc(len);
4527                         if (lists[list] == NULL)
4528                                 goto out;
4529                         strcpy(lists[list], s);
4530                 }
4531
4532                 if (!sep)
4533                         break;
4534
4535                 *sep = ',';
4536                 s = sep + 1;
4537         }
4538
4539         if (lists[1] != NULL) {
4540                 struct strlist_config slist_config = {
4541                         .dirname = strace_groups_dir,
4542                 };
4543
4544                 trace->ev_qualifier = strlist__new(lists[1], &slist_config);
4545                 if (trace->ev_qualifier == NULL) {
4546                         fputs("Not enough memory to parse event qualifier", trace->output);
4547                         goto out;
4548                 }
4549
4550                 if (trace__validate_ev_qualifier(trace))
4551                         goto out;
4552                 trace->trace_syscalls = true;
4553         }
4554
4555         err = 0;
4556
4557         if (lists[0]) {
4558                 struct parse_events_option_args parse_events_option_args = {
4559                         .evlistp = &trace->evlist,
4560                 };
4561                 struct option o = {
4562                         .value = &parse_events_option_args,
4563                 };
4564                 err = parse_events_option(&o, lists[0], 0);
4565         }
4566 out:
4567         free(strace_groups_dir);
4568         free(lists[0]);
4569         free(lists[1]);
4570         if (sep)
4571                 *sep = ',';
4572
4573         return err;
4574 }
4575
4576 static int trace__parse_cgroups(const struct option *opt, const char *str, int unset)
4577 {
4578         struct trace *trace = opt->value;
4579
4580         if (!list_empty(&trace->evlist->core.entries)) {
4581                 struct option o = {
4582                         .value = &trace->evlist,
4583                 };
4584                 return parse_cgroups(&o, str, unset);
4585         }
4586         trace->cgroup = evlist__findnew_cgroup(trace->evlist, str);
4587
4588         return 0;
4589 }
4590
4591 static int trace__config(const char *var, const char *value, void *arg)
4592 {
4593         struct trace *trace = arg;
4594         int err = 0;
4595
4596         if (!strcmp(var, "trace.add_events")) {
4597                 trace->perfconfig_events = strdup(value);
4598                 if (trace->perfconfig_events == NULL) {
4599                         pr_err("Not enough memory for %s\n", "trace.add_events");
4600                         return -1;
4601                 }
4602         } else if (!strcmp(var, "trace.show_timestamp")) {
4603                 trace->show_tstamp = perf_config_bool(var, value);
4604         } else if (!strcmp(var, "trace.show_duration")) {
4605                 trace->show_duration = perf_config_bool(var, value);
4606         } else if (!strcmp(var, "trace.show_arg_names")) {
4607                 trace->show_arg_names = perf_config_bool(var, value);
4608                 if (!trace->show_arg_names)
4609                         trace->show_zeros = true;
4610         } else if (!strcmp(var, "trace.show_zeros")) {
4611                 bool new_show_zeros = perf_config_bool(var, value);
4612                 if (!trace->show_arg_names && !new_show_zeros) {
4613                         pr_warning("trace.show_zeros has to be set when trace.show_arg_names=no\n");
4614                         goto out;
4615                 }
4616                 trace->show_zeros = new_show_zeros;
4617         } else if (!strcmp(var, "trace.show_prefix")) {
4618                 trace->show_string_prefix = perf_config_bool(var, value);
4619         } else if (!strcmp(var, "trace.no_inherit")) {
4620                 trace->opts.no_inherit = perf_config_bool(var, value);
4621         } else if (!strcmp(var, "trace.args_alignment")) {
4622                 int args_alignment = 0;
4623                 if (perf_config_int(&args_alignment, var, value) == 0)
4624                         trace->args_alignment = args_alignment;
4625         } else if (!strcmp(var, "trace.tracepoint_beautifiers")) {
4626                 if (strcasecmp(value, "libtraceevent") == 0)
4627                         trace->libtraceevent_print = true;
4628                 else if (strcasecmp(value, "libbeauty") == 0)
4629                         trace->libtraceevent_print = false;
4630         }
4631 out:
4632         return err;
4633 }
4634
4635 static void trace__exit(struct trace *trace)
4636 {
4637         int i;
4638
4639         strlist__delete(trace->ev_qualifier);
4640         zfree(&trace->ev_qualifier_ids.entries);
4641         if (trace->syscalls.table) {
4642                 for (i = 0; i <= trace->sctbl->syscalls.max_id; i++)
4643                         syscall__exit(&trace->syscalls.table[i]);
4644                 zfree(&trace->syscalls.table);
4645         }
4646         syscalltbl__delete(trace->sctbl);
4647         zfree(&trace->perfconfig_events);
4648 }
4649
4650 #ifdef HAVE_BPF_SKEL
4651 static int bpf__setup_bpf_output(struct evlist *evlist)
4652 {
4653         int err = parse_event(evlist, "bpf-output/no-inherit=1,name=__augmented_syscalls__/");
4654
4655         if (err)
4656                 pr_debug("ERROR: failed to create the \"__augmented_syscalls__\" bpf-output event\n");
4657
4658         return err;
4659 }
4660 #endif
4661
4662 int cmd_trace(int argc, const char **argv)
4663 {
4664         const char *trace_usage[] = {
4665                 "perf trace [<options>] [<command>]",
4666                 "perf trace [<options>] -- <command> [<options>]",
4667                 "perf trace record [<options>] [<command>]",
4668                 "perf trace record [<options>] -- <command> [<options>]",
4669                 NULL
4670         };
4671         struct trace trace = {
4672                 .opts = {
4673                         .target = {
4674                                 .uid       = UINT_MAX,
4675                                 .uses_mmap = true,
4676                         },
4677                         .user_freq     = UINT_MAX,
4678                         .user_interval = ULLONG_MAX,
4679                         .no_buffering  = true,
4680                         .mmap_pages    = UINT_MAX,
4681                 },
4682                 .output = stderr,
4683                 .show_comm = true,
4684                 .show_tstamp = true,
4685                 .show_duration = true,
4686                 .show_arg_names = true,
4687                 .args_alignment = 70,
4688                 .trace_syscalls = false,
4689                 .kernel_syscallchains = false,
4690                 .max_stack = UINT_MAX,
4691                 .max_events = ULONG_MAX,
4692         };
4693         const char *output_name = NULL;
4694         const struct option trace_options[] = {
4695         OPT_CALLBACK('e', "event", &trace, "event",
4696                      "event/syscall selector. use 'perf list' to list available events",
4697                      trace__parse_events_option),
4698         OPT_CALLBACK(0, "filter", &trace.evlist, "filter",
4699                      "event filter", parse_filter),
4700         OPT_BOOLEAN(0, "comm", &trace.show_comm,
4701                     "show the thread COMM next to its id"),
4702         OPT_BOOLEAN(0, "tool_stats", &trace.show_tool_stats, "show tool stats"),
4703         OPT_CALLBACK(0, "expr", &trace, "expr", "list of syscalls/events to trace",
4704                      trace__parse_events_option),
4705         OPT_STRING('o', "output", &output_name, "file", "output file name"),
4706         OPT_STRING('i', "input", &input_name, "file", "Analyze events in file"),
4707         OPT_STRING('p', "pid", &trace.opts.target.pid, "pid",
4708                     "trace events on existing process id"),
4709         OPT_STRING('t', "tid", &trace.opts.target.tid, "tid",
4710                     "trace events on existing thread id"),
4711         OPT_CALLBACK(0, "filter-pids", &trace, "CSV list of pids",
4712                      "pids to filter (by the kernel)", trace__set_filter_pids_from_option),
4713         OPT_BOOLEAN('a', "all-cpus", &trace.opts.target.system_wide,
4714                     "system-wide collection from all CPUs"),
4715         OPT_STRING('C', "cpu", &trace.opts.target.cpu_list, "cpu",
4716                     "list of cpus to monitor"),
4717         OPT_BOOLEAN(0, "no-inherit", &trace.opts.no_inherit,
4718                     "child tasks do not inherit counters"),
4719         OPT_CALLBACK('m', "mmap-pages", &trace.opts.mmap_pages, "pages",
4720                      "number of mmap data pages", evlist__parse_mmap_pages),
4721         OPT_STRING('u', "uid", &trace.opts.target.uid_str, "user",
4722                    "user to profile"),
4723         OPT_CALLBACK(0, "duration", &trace, "float",
4724                      "show only events with duration > N.M ms",
4725                      trace__set_duration),
4726         OPT_BOOLEAN(0, "sched", &trace.sched, "show blocking scheduler events"),
4727         OPT_INCR('v', "verbose", &verbose, "be more verbose"),
4728         OPT_BOOLEAN('T', "time", &trace.full_time,
4729                     "Show full timestamp, not time relative to first start"),
4730         OPT_BOOLEAN(0, "failure", &trace.failure_only,
4731                     "Show only syscalls that failed"),
4732         OPT_BOOLEAN('s', "summary", &trace.summary_only,
4733                     "Show only syscall summary with statistics"),
4734         OPT_BOOLEAN('S', "with-summary", &trace.summary,
4735                     "Show all syscalls and summary with statistics"),
4736         OPT_BOOLEAN(0, "errno-summary", &trace.errno_summary,
4737                     "Show errno stats per syscall, use with -s or -S"),
4738         OPT_CALLBACK_DEFAULT('F', "pf", &trace.trace_pgfaults, "all|maj|min",
4739                      "Trace pagefaults", parse_pagefaults, "maj"),
4740         OPT_BOOLEAN(0, "syscalls", &trace.trace_syscalls, "Trace syscalls"),
4741         OPT_BOOLEAN('f', "force", &trace.force, "don't complain, do it"),
4742         OPT_CALLBACK(0, "call-graph", &trace.opts,
4743                      "record_mode[,record_size]", record_callchain_help,
4744                      &record_parse_callchain_opt),
4745         OPT_BOOLEAN(0, "libtraceevent_print", &trace.libtraceevent_print,
4746                     "Use libtraceevent to print the tracepoint arguments."),
4747         OPT_BOOLEAN(0, "kernel-syscall-graph", &trace.kernel_syscallchains,
4748                     "Show the kernel callchains on the syscall exit path"),
4749         OPT_ULONG(0, "max-events", &trace.max_events,
4750                 "Set the maximum number of events to print, exit after that is reached. "),
4751         OPT_UINTEGER(0, "min-stack", &trace.min_stack,
4752                      "Set the minimum stack depth when parsing the callchain, "
4753                      "anything below the specified depth will be ignored."),
4754         OPT_UINTEGER(0, "max-stack", &trace.max_stack,
4755                      "Set the maximum stack depth when parsing the callchain, "
4756                      "anything beyond the specified depth will be ignored. "
4757                      "Default: kernel.perf_event_max_stack or " __stringify(PERF_MAX_STACK_DEPTH)),
4758         OPT_BOOLEAN(0, "sort-events", &trace.sort_events,
4759                         "Sort batch of events before processing, use if getting out of order events"),
4760         OPT_BOOLEAN(0, "print-sample", &trace.print_sample,
4761                         "print the PERF_RECORD_SAMPLE PERF_SAMPLE_ info, for debugging"),
4762         OPT_UINTEGER(0, "proc-map-timeout", &proc_map_timeout,
4763                         "per thread proc mmap processing timeout in ms"),
4764         OPT_CALLBACK('G', "cgroup", &trace, "name", "monitor event in cgroup name only",
4765                      trace__parse_cgroups),
4766         OPT_INTEGER('D', "delay", &trace.opts.target.initial_delay,
4767                      "ms to wait before starting measurement after program "
4768                      "start"),
4769         OPTS_EVSWITCH(&trace.evswitch),
4770         OPT_END()
4771         };
4772         bool __maybe_unused max_stack_user_set = true;
4773         bool mmap_pages_user_set = true;
4774         struct evsel *evsel;
4775         const char * const trace_subcommands[] = { "record", NULL };
4776         int err = -1;
4777         char bf[BUFSIZ];
4778         struct sigaction sigchld_act;
4779
4780         signal(SIGSEGV, sighandler_dump_stack);
4781         signal(SIGFPE, sighandler_dump_stack);
4782         signal(SIGINT, sighandler_interrupt);
4783
4784         memset(&sigchld_act, 0, sizeof(sigchld_act));
4785         sigchld_act.sa_flags = SA_SIGINFO;
4786         sigchld_act.sa_sigaction = sighandler_chld;
4787         sigaction(SIGCHLD, &sigchld_act, NULL);
4788
4789         trace.evlist = evlist__new();
4790         trace.sctbl = syscalltbl__new();
4791
4792         if (trace.evlist == NULL || trace.sctbl == NULL) {
4793                 pr_err("Not enough memory to run!\n");
4794                 err = -ENOMEM;
4795                 goto out;
4796         }
4797
4798         /*
4799          * Parsing .perfconfig may entail creating a BPF event, that may need
4800          * to create BPF maps, so bump RLIM_MEMLOCK as the default 64K setting
4801          * is too small. This affects just this process, not touching the
4802          * global setting. If it fails we'll get something in 'perf trace -v'
4803          * to help diagnose the problem.
4804          */
4805         rlimit__bump_memlock();
4806
4807         err = perf_config(trace__config, &trace);
4808         if (err)
4809                 goto out;
4810
4811         argc = parse_options_subcommand(argc, argv, trace_options, trace_subcommands,
4812                                  trace_usage, PARSE_OPT_STOP_AT_NON_OPTION);
4813
4814         /*
4815          * Here we already passed thru trace__parse_events_option() and it has
4816          * already figured out if -e syscall_name, if not but if --event
4817          * foo:bar was used, the user is interested _just_ in those, say,
4818          * tracepoint events, not in the strace-like syscall-name-based mode.
4819          *
4820          * This is important because we need to check if strace-like mode is
4821          * needed to decided if we should filter out the eBPF
4822          * __augmented_syscalls__ code, if it is in the mix, say, via
4823          * .perfconfig trace.add_events, and filter those out.
4824          */
4825         if (!trace.trace_syscalls && !trace.trace_pgfaults &&
4826             trace.evlist->core.nr_entries == 0 /* Was --events used? */) {
4827                 trace.trace_syscalls = true;
4828         }
4829         /*
4830          * Now that we have --verbose figured out, lets see if we need to parse
4831          * events from .perfconfig, so that if those events fail parsing, say some
4832          * BPF program fails, then we'll be able to use --verbose to see what went
4833          * wrong in more detail.
4834          */
4835         if (trace.perfconfig_events != NULL) {
4836                 struct parse_events_error parse_err;
4837
4838                 parse_events_error__init(&parse_err);
4839                 err = parse_events(trace.evlist, trace.perfconfig_events, &parse_err);
4840                 if (err)
4841                         parse_events_error__print(&parse_err, trace.perfconfig_events);
4842                 parse_events_error__exit(&parse_err);
4843                 if (err)
4844                         goto out;
4845         }
4846
4847         if ((nr_cgroups || trace.cgroup) && !trace.opts.target.system_wide) {
4848                 usage_with_options_msg(trace_usage, trace_options,
4849                                        "cgroup monitoring only available in system-wide mode");
4850         }
4851
4852 #ifdef HAVE_BPF_SKEL
4853         trace.skel = augmented_raw_syscalls_bpf__open();
4854         if (!trace.skel) {
4855                 pr_debug("Failed to open augmented syscalls BPF skeleton");
4856         } else {
4857                 /*
4858                  * Disable attaching the BPF programs except for sys_enter and
4859                  * sys_exit that tail call into this as necessary.
4860                  */
4861                 struct bpf_program *prog;
4862
4863                 bpf_object__for_each_program(prog, trace.skel->obj) {
4864                         if (prog != trace.skel->progs.sys_enter && prog != trace.skel->progs.sys_exit)
4865                                 bpf_program__set_autoattach(prog, /*autoattach=*/false);
4866                 }
4867
4868                 err = augmented_raw_syscalls_bpf__load(trace.skel);
4869
4870                 if (err < 0) {
4871                         libbpf_strerror(err, bf, sizeof(bf));
4872                         pr_debug("Failed to load augmented syscalls BPF skeleton: %s\n", bf);
4873                 } else {
4874                         augmented_raw_syscalls_bpf__attach(trace.skel);
4875                         trace__add_syscall_newtp(&trace);
4876                 }
4877         }
4878
4879         err = bpf__setup_bpf_output(trace.evlist);
4880         if (err) {
4881                 libbpf_strerror(err, bf, sizeof(bf));
4882                 pr_err("ERROR: Setup BPF output event failed: %s\n", bf);
4883                 goto out;
4884         }
4885         trace.syscalls.events.bpf_output = evlist__last(trace.evlist);
4886         assert(!strcmp(evsel__name(trace.syscalls.events.bpf_output), "__augmented_syscalls__"));
4887 #endif
4888         err = -1;
4889
4890         if (trace.trace_pgfaults) {
4891                 trace.opts.sample_address = true;
4892                 trace.opts.sample_time = true;
4893         }
4894
4895         if (trace.opts.mmap_pages == UINT_MAX)
4896                 mmap_pages_user_set = false;
4897
4898         if (trace.max_stack == UINT_MAX) {
4899                 trace.max_stack = input_name ? PERF_MAX_STACK_DEPTH : sysctl__max_stack();
4900                 max_stack_user_set = false;
4901         }
4902
4903 #ifdef HAVE_DWARF_UNWIND_SUPPORT
4904         if ((trace.min_stack || max_stack_user_set) && !callchain_param.enabled) {
4905                 record_opts__parse_callchain(&trace.opts, &callchain_param, "dwarf", false);
4906         }
4907 #endif
4908
4909         if (callchain_param.enabled) {
4910                 if (!mmap_pages_user_set && geteuid() == 0)
4911                         trace.opts.mmap_pages = perf_event_mlock_kb_in_pages() * 4;
4912
4913                 symbol_conf.use_callchain = true;
4914         }
4915
4916         if (trace.evlist->core.nr_entries > 0) {
4917                 evlist__set_default_evsel_handler(trace.evlist, trace__event_handler);
4918                 if (evlist__set_syscall_tp_fields(trace.evlist)) {
4919                         perror("failed to set syscalls:* tracepoint fields");
4920                         goto out;
4921                 }
4922         }
4923
4924         if (trace.sort_events) {
4925                 ordered_events__init(&trace.oe.data, ordered_events__deliver_event, &trace);
4926                 ordered_events__set_copy_on_queue(&trace.oe.data, true);
4927         }
4928
4929         /*
4930          * If we are augmenting syscalls, then combine what we put in the
4931          * __augmented_syscalls__ BPF map with what is in the
4932          * syscalls:sys_exit_FOO tracepoints, i.e. just like we do without BPF,
4933          * combining raw_syscalls:sys_enter with raw_syscalls:sys_exit.
4934          *
4935          * We'll switch to look at two BPF maps, one for sys_enter and the
4936          * other for sys_exit when we start augmenting the sys_exit paths with
4937          * buffers that are being copied from kernel to userspace, think 'read'
4938          * syscall.
4939          */
4940         if (trace.syscalls.events.bpf_output) {
4941                 evlist__for_each_entry(trace.evlist, evsel) {
4942                         bool raw_syscalls_sys_exit = strcmp(evsel__name(evsel), "raw_syscalls:sys_exit") == 0;
4943
4944                         if (raw_syscalls_sys_exit) {
4945                                 trace.raw_augmented_syscalls = true;
4946                                 goto init_augmented_syscall_tp;
4947                         }
4948
4949                         if (trace.syscalls.events.bpf_output->priv == NULL &&
4950                             strstr(evsel__name(evsel), "syscalls:sys_enter")) {
4951                                 struct evsel *augmented = trace.syscalls.events.bpf_output;
4952                                 if (evsel__init_augmented_syscall_tp(augmented, evsel) ||
4953                                     evsel__init_augmented_syscall_tp_args(augmented))
4954                                         goto out;
4955                                 /*
4956                                  * Augmented is __augmented_syscalls__ BPF_OUTPUT event
4957                                  * Above we made sure we can get from the payload the tp fields
4958                                  * that we get from syscalls:sys_enter tracefs format file.
4959                                  */
4960                                 augmented->handler = trace__sys_enter;
4961                                 /*
4962                                  * Now we do the same for the *syscalls:sys_enter event so that
4963                                  * if we handle it directly, i.e. if the BPF prog returns 0 so
4964                                  * as not to filter it, then we'll handle it just like we would
4965                                  * for the BPF_OUTPUT one:
4966                                  */
4967                                 if (evsel__init_augmented_syscall_tp(evsel, evsel) ||
4968                                     evsel__init_augmented_syscall_tp_args(evsel))
4969                                         goto out;
4970                                 evsel->handler = trace__sys_enter;
4971                         }
4972
4973                         if (strstarts(evsel__name(evsel), "syscalls:sys_exit_")) {
4974                                 struct syscall_tp *sc;
4975 init_augmented_syscall_tp:
4976                                 if (evsel__init_augmented_syscall_tp(evsel, evsel))
4977                                         goto out;
4978                                 sc = __evsel__syscall_tp(evsel);
4979                                 /*
4980                                  * For now with BPF raw_augmented we hook into
4981                                  * raw_syscalls:sys_enter and there we get all
4982                                  * 6 syscall args plus the tracepoint common
4983                                  * fields and the syscall_nr (another long).
4984                                  * So we check if that is the case and if so
4985                                  * don't look after the sc->args_size but
4986                                  * always after the full raw_syscalls:sys_enter
4987                                  * payload, which is fixed.
4988                                  *
4989                                  * We'll revisit this later to pass
4990                                  * s->args_size to the BPF augmenter (now
4991                                  * tools/perf/examples/bpf/augmented_raw_syscalls.c,
4992                                  * so that it copies only what we need for each
4993                                  * syscall, like what happens when we use
4994                                  * syscalls:sys_enter_NAME, so that we reduce
4995                                  * the kernel/userspace traffic to just what is
4996                                  * needed for each syscall.
4997                                  */
4998                                 if (trace.raw_augmented_syscalls)
4999                                         trace.raw_augmented_syscalls_args_size = (6 + 1) * sizeof(long) + sc->id.offset;
5000                                 evsel__init_augmented_syscall_tp_ret(evsel);
5001                                 evsel->handler = trace__sys_exit;
5002                         }
5003                 }
5004         }
5005
5006         if ((argc >= 1) && (strcmp(argv[0], "record") == 0))
5007                 return trace__record(&trace, argc-1, &argv[1]);
5008
5009         /* Using just --errno-summary will trigger --summary */
5010         if (trace.errno_summary && !trace.summary && !trace.summary_only)
5011                 trace.summary_only = true;
5012
5013         /* summary_only implies summary option, but don't overwrite summary if set */
5014         if (trace.summary_only)
5015                 trace.summary = trace.summary_only;
5016
5017         if (output_name != NULL) {
5018                 err = trace__open_output(&trace, output_name);
5019                 if (err < 0) {
5020                         perror("failed to create output file");
5021                         goto out;
5022                 }
5023         }
5024
5025         err = evswitch__init(&trace.evswitch, trace.evlist, stderr);
5026         if (err)
5027                 goto out_close;
5028
5029         err = target__validate(&trace.opts.target);
5030         if (err) {
5031                 target__strerror(&trace.opts.target, err, bf, sizeof(bf));
5032                 fprintf(trace.output, "%s", bf);
5033                 goto out_close;
5034         }
5035
5036         err = target__parse_uid(&trace.opts.target);
5037         if (err) {
5038                 target__strerror(&trace.opts.target, err, bf, sizeof(bf));
5039                 fprintf(trace.output, "%s", bf);
5040                 goto out_close;
5041         }
5042
5043         if (!argc && target__none(&trace.opts.target))
5044                 trace.opts.target.system_wide = true;
5045
5046         if (input_name)
5047                 err = trace__replay(&trace);
5048         else
5049                 err = trace__run(&trace, argc, argv);
5050
5051 out_close:
5052         if (output_name != NULL)
5053                 fclose(trace.output);
5054 out:
5055         trace__exit(&trace);
5056 #ifdef HAVE_BPF_SKEL
5057         augmented_raw_syscalls_bpf__destroy(trace.skel);
5058 #endif
5059         return err;
5060 }