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 <[email protected]>
  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 #endif
  22 #include "util/bpf_map.h"
  23 #include "util/rlimit.h"
  24 #include "builtin.h"
  25 #include "util/cgroup.h"
  26 #include "util/color.h"
  27 #include "util/config.h"
  28 #include "util/debug.h"
  29 #include "util/dso.h"
  30 #include "util/env.h"
  31 #include "util/event.h"
  32 #include "util/evsel.h"
  33 #include "util/evsel_fprintf.h"
  34 #include "util/synthetic-events.h"
  35 #include "util/evlist.h"
  36 #include "util/evswitch.h"
  37 #include "util/mmap.h"
  38 #include <subcmd/pager.h>
  39 #include <subcmd/exec-cmd.h>
  40 #include "util/machine.h"
  41 #include "util/map.h"
  42 #include "util/symbol.h"
  43 #include "util/path.h"
  44 #include "util/session.h"
  45 #include "util/thread.h"
  46 #include <subcmd/parse-options.h>
  47 #include "util/strlist.h"
  48 #include "util/intlist.h"
  49 #include "util/thread_map.h"
  50 #include "util/stat.h"
  51 #include "util/tool.h"
  52 #include "util/util.h"
  53 #include "trace/beauty/beauty.h"
  54 #include "trace-event.h"
  55 #include "util/parse-events.h"
  56 #include "util/bpf-loader.h"
  57 #include "util/tracepoint.h"
  58 #include "callchain.h"
  59 #include "print_binary.h"
  60 #include "string2.h"
  61 #include "syscalltbl.h"
  62 #include "rb_resort.h"
  63 #include "../perf.h"
  64
  65 #include <errno.h>
  66 #include <inttypes.h>
  67 #include <poll.h>
  68 #include <signal.h>
  69 #include <stdlib.h>
  70 #include <string.h>
  71 #include <linux/err.h>
  72 #include <linux/filter.h>
  73 #include <linux/kernel.h>
  74 #include <linux/random.h>
  75 #include <linux/stringify.h>
  76 #include <linux/time64.h>
  77 #include <linux/zalloc.h>
  78 #include <fcntl.h>
  79 #include <sys/sysmacros.h>
  80
  81 #include <linux/ctype.h>
  82 #include <perf/mmap.h>
  83
  84 #ifdef HAVE_LIBTRACEEVENT
  85 #include <traceevent/event-parse.h>
  86 #endif
  87
  88 #ifndef O_CLOEXEC
  89 # define O_CLOEXEC              02000000
  90 #endif
  91
  92 #ifndef F_LINUX_SPECIFIC_BASE
  93 # define F_LINUX_SPECIFIC_BASE  1024
  94 #endif
  95
  96 #define RAW_SYSCALL_ARGS_NUM    6
  97
  98 /*
  99  * strtoul: Go from a string to a value, i.e. for msr: MSR_FS_BASE to 0xc0000100
 100  */
 101 struct syscall_arg_fmt {
 102         size_t     (*scnprintf)(char *bf, size_t size, struct syscall_arg *arg);
 103         bool       (*strtoul)(char *bf, size_t size, struct syscall_arg *arg, u64 *val);
 104         unsigned long (*mask_val)(struct syscall_arg *arg, unsigned long val);
 105         void       *parm;
 106         const char *name;
 107         u16        nr_entries; // for arrays
 108         bool       show_zero;
 109 };
 110
 111 struct syscall_fmt {
 112         const char *name;
 113         const char *alias;
 114         struct {
 115                 const char *sys_enter,
 116                            *sys_exit;
 117         }          bpf_prog_name;
 118         struct syscall_arg_fmt arg[RAW_SYSCALL_ARGS_NUM];
 119         u8         nr_args;
 120         bool       errpid;
 121         bool       timeout;
 122         bool       hexret;
 123 };
 124
 125 struct trace {
 126         struct perf_tool        tool;
 127         struct syscalltbl       *sctbl;
 128         struct {
 129                 struct syscall  *table;
 130                 struct { // per syscall BPF_MAP_TYPE_PROG_ARRAY
 131                         struct bpf_map  *sys_enter,
 132                                         *sys_exit;
 133                 }               prog_array;
 134                 struct {
 135                         struct evsel *sys_enter,
 136                                           *sys_exit,
 137                                           *augmented;
 138                 }               events;
 139                 struct bpf_program *unaugmented_prog;
 140         } syscalls;
 141         struct {
 142                 struct bpf_map *map;
 143         } dump;
 144         struct record_opts      opts;
 145         struct evlist   *evlist;
 146         struct machine          *host;
 147         struct thread           *current;
 148         struct bpf_object       *bpf_obj;
 149         struct cgroup           *cgroup;
 150         u64                     base_time;
 151         FILE                    *output;
 152         unsigned long           nr_events;
 153         unsigned long           nr_events_printed;
 154         unsigned long           max_events;
 155         struct evswitch         evswitch;
 156         struct strlist          *ev_qualifier;
 157         struct {
 158                 size_t          nr;
 159                 int             *entries;
 160         }                       ev_qualifier_ids;
 161         struct {
 162                 size_t          nr;
 163                 pid_t           *entries;
 164                 struct bpf_map  *map;
 165         }                       filter_pids;
 166         double                  duration_filter;
 167         double                  runtime_ms;
 168         struct {
 169                 u64             vfs_getname,
 170                                 proc_getname;
 171         } stats;
 172         unsigned int            max_stack;
 173         unsigned int            min_stack;
 174         int                     raw_augmented_syscalls_args_size;
 175         bool                    raw_augmented_syscalls;
 176         bool                    fd_path_disabled;
 177         bool                    sort_events;
 178         bool                    not_ev_qualifier;
 179         bool                    live;
 180         bool                    full_time;
 181         bool                    sched;
 182         bool                    multiple_threads;
 183         bool                    summary;
 184         bool                    summary_only;
 185         bool                    errno_summary;
 186         bool                    failure_only;
 187         bool                    show_comm;
 188         bool                    print_sample;
 189         bool                    show_tool_stats;
 190         bool                    trace_syscalls;
 191         bool                    libtraceevent_print;
 192         bool                    kernel_syscallchains;
 193         s16                     args_alignment;
 194         bool                    show_tstamp;
 195         bool                    show_duration;
 196         bool                    show_zeros;
 197         bool                    show_arg_names;
 198         bool                    show_string_prefix;
 199         bool                    force;
 200         bool                    vfs_getname;
 201         int                     trace_pgfaults;
 202         char                    *perfconfig_events;
 203         struct {
 204                 struct ordered_events   data;
 205                 u64                     last;
 206         } oe;
 207 };
 208
 209 struct tp_field {
 210         int offset;
 211         union {
 212                 u64 (*integer)(struct tp_field *field, struct perf_sample *sample);
 213                 void *(*pointer)(struct tp_field *field, struct perf_sample *sample);
 214         };
 215 };
 216
 217 #define TP_UINT_FIELD(bits) \
 218 static u64 tp_field__u##bits(struct tp_field *field, struct perf_sample *sample) \
 219 { \
 220         u##bits value; \
 221         memcpy(&value, sample->raw_data + field->offset, sizeof(value)); \
 222         return value;  \
 223 }
 224
 225 TP_UINT_FIELD(8);
 226 TP_UINT_FIELD(16);
 227 TP_UINT_FIELD(32);
 228 TP_UINT_FIELD(64);
 229
 230 #define TP_UINT_FIELD__SWAPPED(bits) \
 231 static u64 tp_field__swapped_u##bits(struct tp_field *field, struct perf_sample *sample) \
 232 { \
 233         u##bits value; \
 234         memcpy(&value, sample->raw_data + field->offset, sizeof(value)); \
 235         return bswap_##bits(value);\
 236 }
 237
 238 TP_UINT_FIELD__SWAPPED(16);
 239 TP_UINT_FIELD__SWAPPED(32);
 240 TP_UINT_FIELD__SWAPPED(64);
 241
 242 static int __tp_field__init_uint(struct tp_field *field, int size, int offset, bool needs_swap)
 243 {
 244         field->offset = offset;
 245
 246         switch (size) {
 247         case 1:
 248                 field->integer = tp_field__u8;
 249                 break;
 250         case 2:
 251                 field->integer = needs_swap ? tp_field__swapped_u16 : tp_field__u16;
 252                 break;
 253         case 4:
 254                 field->integer = needs_swap ? tp_field__swapped_u32 : tp_field__u32;
 255                 break;
 256         case 8:
 257                 field->integer = needs_swap ? tp_field__swapped_u64 : tp_field__u64;
 258                 break;
 259         default:
 260                 return -1;
 261         }
 262
 263         return 0;
 264 }
 265
 266 static int tp_field__init_uint(struct tp_field *field, struct tep_format_field *format_field, bool needs_swap)
 267 {
 268         return __tp_field__init_uint(field, format_field->size, format_field->offset, needs_swap);
 269 }
 270
 271 static void *tp_field__ptr(struct tp_field *field, struct perf_sample *sample)
 272 {
 273         return sample->raw_data + field->offset;
 274 }
 275
 276 static int __tp_field__init_ptr(struct tp_field *field, int offset)
 277 {
 278         field->offset = offset;
 279         field->pointer = tp_field__ptr;
 280         return 0;
 281 }
 282
 283 static int tp_field__init_ptr(struct tp_field *field, struct tep_format_field *format_field)
 284 {
 285         return __tp_field__init_ptr(field, format_field->offset);
 286 }
 287
 288 struct syscall_tp {
 289         struct tp_field id;
 290         union {
 291                 struct tp_field args, ret;
 292         };
 293 };
 294
 295 /*
 296  * The evsel->priv as used by 'perf trace'
 297  * sc:  for raw_syscalls:sys_{enter,exit} and syscalls:sys_{enter,exit}_SYSCALLNAME
 298  * fmt: for all the other tracepoints
 299  */
 300 struct evsel_trace {
 301         struct syscall_tp       sc;
 302         struct syscall_arg_fmt  *fmt;
 303 };
 304
 305 static struct evsel_trace *evsel_trace__new(void)
 306 {
 307         return zalloc(sizeof(struct evsel_trace));
 308 }
 309
 310 static void evsel_trace__delete(struct evsel_trace *et)
 311 {
 312         if (et == NULL)
 313                 return;
 314
 315         zfree(&et->fmt);
 316         free(et);
 317 }
 318
 319 /*
 320  * Used with raw_syscalls:sys_{enter,exit} and with the
 321  * syscalls:sys_{enter,exit}_SYSCALL tracepoints
 322  */
 323 static inline struct syscall_tp *__evsel__syscall_tp(struct evsel *evsel)
 324 {
 325         struct evsel_trace *et = evsel->priv;
 326
 327         return &et->sc;
 328 }
 329
 330 static struct syscall_tp *evsel__syscall_tp(struct evsel *evsel)
 331 {
 332         if (evsel->priv == NULL) {
 333                 evsel->priv = evsel_trace__new();
 334                 if (evsel->priv == NULL)
 335                         return NULL;
 336         }
 337
 338         return __evsel__syscall_tp(evsel);
 339 }
 340
 341 /*
 342  * Used with all the other tracepoints.
 343  */
 344 static inline struct syscall_arg_fmt *__evsel__syscall_arg_fmt(struct evsel *evsel)
 345 {
 346         struct evsel_trace *et = evsel->priv;
 347
 348         return et->fmt;
 349 }
 350
 351 static struct syscall_arg_fmt *evsel__syscall_arg_fmt(struct evsel *evsel)
 352 {
 353         struct evsel_trace *et = evsel->priv;
 354
 355         if (evsel->priv == NULL) {
 356                 et = evsel->priv = evsel_trace__new();
 357
 358                 if (et == NULL)
 359                         return NULL;
 360         }
 361
 362         if (et->fmt == NULL) {
 363                 et->fmt = calloc(evsel->tp_format->format.nr_fields, sizeof(struct syscall_arg_fmt));
 364                 if (et->fmt == NULL)
 365                         goto out_delete;
 366         }
 367
 368         return __evsel__syscall_arg_fmt(evsel);
 369
 370 out_delete:
 371         evsel_trace__delete(evsel->priv);
 372         evsel->priv = NULL;
 373         return NULL;
 374 }
 375
 376 static int evsel__init_tp_uint_field(struct evsel *evsel, struct tp_field *field, const char *name)
 377 {
 378         struct tep_format_field *format_field = evsel__field(evsel, name);
 379
 380         if (format_field == NULL)
 381                 return -1;
 382
 383         return tp_field__init_uint(field, format_field, evsel->needs_swap);
 384 }
 385
 386 #define perf_evsel__init_sc_tp_uint_field(evsel, name) \
 387         ({ struct syscall_tp *sc = __evsel__syscall_tp(evsel);\
 388            evsel__init_tp_uint_field(evsel, &sc->name, #name); })
 389
 390 static int evsel__init_tp_ptr_field(struct evsel *evsel, struct tp_field *field, const char *name)
 391 {
 392         struct tep_format_field *format_field = evsel__field(evsel, name);
 393
 394         if (format_field == NULL)
 395                 return -1;
 396
 397         return tp_field__init_ptr(field, format_field);
 398 }
 399
 400 #define perf_evsel__init_sc_tp_ptr_field(evsel, name) \
 401         ({ struct syscall_tp *sc = __evsel__syscall_tp(evsel);\
 402            evsel__init_tp_ptr_field(evsel, &sc->name, #name); })
 403
 404 static void evsel__delete_priv(struct evsel *evsel)
 405 {
 406         zfree(&evsel->priv);
 407         evsel__delete(evsel);
 408 }
 409
 410 static int evsel__init_syscall_tp(struct evsel *evsel)
 411 {
 412         struct syscall_tp *sc = evsel__syscall_tp(evsel);
 413
 414         if (sc != NULL) {
 415                 if (evsel__init_tp_uint_field(evsel, &sc->id, "__syscall_nr") &&
 416                     evsel__init_tp_uint_field(evsel, &sc->id, "nr"))
 417                         return -ENOENT;
 418                 return 0;
 419         }
 420
 421         return -ENOMEM;
 422 }
 423
 424 static int evsel__init_augmented_syscall_tp(struct evsel *evsel, struct evsel *tp)
 425 {
 426         struct syscall_tp *sc = evsel__syscall_tp(evsel);
 427
 428         if (sc != NULL) {
 429                 struct tep_format_field *syscall_id = evsel__field(tp, "id");
 430                 if (syscall_id == NULL)
 431                         syscall_id = evsel__field(tp, "__syscall_nr");
 432                 if (syscall_id == NULL ||
 433                     __tp_field__init_uint(&sc->id, syscall_id->size, syscall_id->offset, evsel->needs_swap))
 434                         return -EINVAL;
 435
 436                 return 0;
 437         }
 438
 439         return -ENOMEM;
 440 }
 441
 442 static int evsel__init_augmented_syscall_tp_args(struct evsel *evsel)
 443 {
 444         struct syscall_tp *sc = __evsel__syscall_tp(evsel);
 445
 446         return __tp_field__init_ptr(&sc->args, sc->id.offset + sizeof(u64));
 447 }
 448
 449 static int evsel__init_augmented_syscall_tp_ret(struct evsel *evsel)
 450 {
 451         struct syscall_tp *sc = __evsel__syscall_tp(evsel);
 452
 453         return __tp_field__init_uint(&sc->ret, sizeof(u64), sc->id.offset + sizeof(u64), evsel->needs_swap);
 454 }
 455
 456 static int evsel__init_raw_syscall_tp(struct evsel *evsel, void *handler)
 457 {
 458         if (evsel__syscall_tp(evsel) != NULL) {
 459                 if (perf_evsel__init_sc_tp_uint_field(evsel, id))
 460                         return -ENOENT;
 461
 462                 evsel->handler = handler;
 463                 return 0;
 464         }
 465
 466         return -ENOMEM;
 467 }
 468
 469 static struct evsel *perf_evsel__raw_syscall_newtp(const char *direction, void *handler)
 470 {
 471         struct evsel *evsel = evsel__newtp("raw_syscalls", direction);
 472
 473         /* older kernel (e.g., RHEL6) use syscalls:{enter,exit} */
 474         if (IS_ERR(evsel))
 475                 evsel = evsel__newtp("syscalls", direction);
 476
 477         if (IS_ERR(evsel))
 478                 return NULL;
 479
 480         if (evsel__init_raw_syscall_tp(evsel, handler))
 481                 goto out_delete;
 482
 483         return evsel;
 484
 485 out_delete:
 486         evsel__delete_priv(evsel);
 487         return NULL;
 488 }
 489
 490 #define perf_evsel__sc_tp_uint(evsel, name, sample) \
 491         ({ struct syscall_tp *fields = __evsel__syscall_tp(evsel); \
 492            fields->name.integer(&fields->name, sample); })
 493
 494 #define perf_evsel__sc_tp_ptr(evsel, name, sample) \
 495         ({ struct syscall_tp *fields = __evsel__syscall_tp(evsel); \
 496            fields->name.pointer(&fields->name, sample); })
 497
 498 size_t strarray__scnprintf_suffix(struct strarray *sa, char *bf, size_t size, const char *intfmt, bool show_suffix, int val)
 499 {
 500         int idx = val - sa->offset;
 501
 502         if (idx < 0 || idx >= sa->nr_entries || sa->entries[idx] == NULL) {
 503                 size_t printed = scnprintf(bf, size, intfmt, val);
 504                 if (show_suffix)
 505                         printed += scnprintf(bf + printed, size - printed, " /* %s??? */", sa->prefix);
 506                 return printed;
 507         }
 508
 509         return scnprintf(bf, size, "%s%s", sa->entries[idx], show_suffix ? sa->prefix : "");
 510 }
 511
 512 size_t strarray__scnprintf(struct strarray *sa, char *bf, size_t size, const char *intfmt, bool show_prefix, int val)
 513 {
 514         int idx = val - sa->offset;
 515
 516         if (idx < 0 || idx >= sa->nr_entries || sa->entries[idx] == NULL) {
 517                 size_t printed = scnprintf(bf, size, intfmt, val);
 518                 if (show_prefix)
 519                         printed += scnprintf(bf + printed, size - printed, " /* %s??? */", sa->prefix);
 520                 return printed;
 521         }
 522
 523         return scnprintf(bf, size, "%s%s", show_prefix ? sa->prefix : "", sa->entries[idx]);
 524 }
 525
 526 static size_t __syscall_arg__scnprintf_strarray(char *bf, size_t size,
 527                                                 const char *intfmt,
 528                                                 struct syscall_arg *arg)
 529 {
 530         return strarray__scnprintf(arg->parm, bf, size, intfmt, arg->show_string_prefix, arg->val);
 531 }
 532
 533 static size_t syscall_arg__scnprintf_strarray(char *bf, size_t size,
 534                                               struct syscall_arg *arg)
 535 {
 536         return __syscall_arg__scnprintf_strarray(bf, size, "%d", arg);
 537 }
 538
 539 #define SCA_STRARRAY syscall_arg__scnprintf_strarray
 540
 541 bool syscall_arg__strtoul_strarray(char *bf, size_t size, struct syscall_arg *arg, u64 *ret)
 542 {
 543         return strarray__strtoul(arg->parm, bf, size, ret);
 544 }
 545
 546 bool syscall_arg__strtoul_strarray_flags(char *bf, size_t size, struct syscall_arg *arg, u64 *ret)
 547 {
 548         return strarray__strtoul_flags(arg->parm, bf, size, ret);
 549 }
 550
 551 bool syscall_arg__strtoul_strarrays(char *bf, size_t size, struct syscall_arg *arg, u64 *ret)
 552 {
 553         return strarrays__strtoul(arg->parm, bf, size, ret);
 554 }
 555
 556 size_t syscall_arg__scnprintf_strarray_flags(char *bf, size_t size, struct syscall_arg *arg)
 557 {
 558         return strarray__scnprintf_flags(arg->parm, bf, size, arg->show_string_prefix, arg->val);
 559 }
 560
 561 size_t strarrays__scnprintf(struct strarrays *sas, char *bf, size_t size, const char *intfmt, bool show_prefix, int val)
 562 {
 563         size_t printed;
 564         int i;
 565
 566         for (i = 0; i < sas->nr_entries; ++i) {
 567                 struct strarray *sa = sas->entries[i];
 568                 int idx = val - sa->offset;
 569
 570                 if (idx >= 0 && idx < sa->nr_entries) {
 571                         if (sa->entries[idx] == NULL)
 572                                 break;
 573                         return scnprintf(bf, size, "%s%s", show_prefix ? sa->prefix : "", sa->entries[idx]);
 574                 }
 575         }
 576
 577         printed = scnprintf(bf, size, intfmt, val);
 578         if (show_prefix)
 579                 printed += scnprintf(bf + printed, size - printed, " /* %s??? */", sas->entries[0]->prefix);
 580         return printed;
 581 }
 582
 583 bool strarray__strtoul(struct strarray *sa, char *bf, size_t size, u64 *ret)
 584 {
 585         int i;
 586
 587         for (i = 0; i < sa->nr_entries; ++i) {
 588                 if (sa->entries[i] && strncmp(sa->entries[i], bf, size) == 0 && sa->entries[i][size] == '\0') {
 589                         *ret = sa->offset + i;
 590                         return true;
 591                 }
 592         }
 593
 594         return false;
 595 }
 596
 597 bool strarray__strtoul_flags(struct strarray *sa, char *bf, size_t size, u64 *ret)
 598 {
 599         u64 val = 0;
 600         char *tok = bf, *sep, *end;
 601
 602         *ret = 0;
 603
 604         while (size != 0) {
 605                 int toklen = size;
 606
 607                 sep = memchr(tok, '|', size);
 608                 if (sep != NULL) {
 609                         size -= sep - tok + 1;
 610
 611                         end = sep - 1;
 612                         while (end > tok && isspace(*end))
 613                                 --end;
 614
 615                         toklen = end - tok + 1;
 616                 }
 617
 618                 while (isspace(*tok))
 619                         ++tok;
 620
 621                 if (isalpha(*tok) || *tok == '_') {
 622                         if (!strarray__strtoul(sa, tok, toklen, &val))
 623                                 return false;
 624                 } else
 625                         val = strtoul(tok, NULL, 0);
 626
 627                 *ret |= (1 << (val - 1));
 628
 629                 if (sep == NULL)
 630                         break;
 631                 tok = sep + 1;
 632         }
 633
 634         return true;
 635 }
 636
 637 bool strarrays__strtoul(struct strarrays *sas, char *bf, size_t size, u64 *ret)
 638 {
 639         int i;
 640
 641         for (i = 0; i < sas->nr_entries; ++i) {
 642                 struct strarray *sa = sas->entries[i];
 643
 644                 if (strarray__strtoul(sa, bf, size, ret))
 645                         return true;
 646         }
 647
 648         return false;
 649 }
 650
 651 size_t syscall_arg__scnprintf_strarrays(char *bf, size_t size,
 652                                         struct syscall_arg *arg)
 653 {
 654         return strarrays__scnprintf(arg->parm, bf, size, "%d", arg->show_string_prefix, arg->val);
 655 }
 656
 657 #ifndef AT_FDCWD
 658 #define AT_FDCWD        -100
 659 #endif
 660
 661 static size_t syscall_arg__scnprintf_fd_at(char *bf, size_t size,
 662                                            struct syscall_arg *arg)
 663 {
 664         int fd = arg->val;
 665         const char *prefix = "AT_FD";
 666
 667         if (fd == AT_FDCWD)
 668                 return scnprintf(bf, size, "%s%s", arg->show_string_prefix ? prefix : "", "CWD");
 669
 670         return syscall_arg__scnprintf_fd(bf, size, arg);
 671 }
 672
 673 #define SCA_FDAT syscall_arg__scnprintf_fd_at
 674
 675 static size_t syscall_arg__scnprintf_close_fd(char *bf, size_t size,
 676                                               struct syscall_arg *arg);
 677
 678 #define SCA_CLOSE_FD syscall_arg__scnprintf_close_fd
 679
 680 size_t syscall_arg__scnprintf_hex(char *bf, size_t size, struct syscall_arg *arg)
 681 {
 682         return scnprintf(bf, size, "%#lx", arg->val);
 683 }
 684
 685 size_t syscall_arg__scnprintf_ptr(char *bf, size_t size, struct syscall_arg *arg)
 686 {
 687         if (arg->val == 0)
 688                 return scnprintf(bf, size, "NULL");
 689         return syscall_arg__scnprintf_hex(bf, size, arg);
 690 }
 691
 692 size_t syscall_arg__scnprintf_int(char *bf, size_t size, struct syscall_arg *arg)
 693 {
 694         return scnprintf(bf, size, "%d", arg->val);
 695 }
 696
 697 size_t syscall_arg__scnprintf_long(char *bf, size_t size, struct syscall_arg *arg)
 698 {
 699         return scnprintf(bf, size, "%ld", arg->val);
 700 }
 701
 702 static size_t syscall_arg__scnprintf_char_array(char *bf, size_t size, struct syscall_arg *arg)
 703 {
 704         // XXX Hey, maybe for sched:sched_switch prev/next comm fields we can
 705         //     fill missing comms using thread__set_comm()...
 706         //     here or in a special syscall_arg__scnprintf_pid_sched_tp...
 707         return scnprintf(bf, size, "\"%-.*s\"", arg->fmt->nr_entries ?: arg->len, arg->val);
 708 }
 709
 710 #define SCA_CHAR_ARRAY syscall_arg__scnprintf_char_array
 711
 712 static const char *bpf_cmd[] = {
 713         "MAP_CREATE", "MAP_LOOKUP_ELEM", "MAP_UPDATE_ELEM", "MAP_DELETE_ELEM",
 714         "MAP_GET_NEXT_KEY", "PROG_LOAD", "OBJ_PIN", "OBJ_GET", "PROG_ATTACH",
 715         "PROG_DETACH", "PROG_TEST_RUN", "PROG_GET_NEXT_ID", "MAP_GET_NEXT_ID",
 716         "PROG_GET_FD_BY_ID", "MAP_GET_FD_BY_ID", "OBJ_GET_INFO_BY_FD",
 717         "PROG_QUERY", "RAW_TRACEPOINT_OPEN", "BTF_LOAD", "BTF_GET_FD_BY_ID",
 718         "TASK_FD_QUERY", "MAP_LOOKUP_AND_DELETE_ELEM", "MAP_FREEZE",
 719         "BTF_GET_NEXT_ID", "MAP_LOOKUP_BATCH", "MAP_LOOKUP_AND_DELETE_BATCH",
 720         "MAP_UPDATE_BATCH", "MAP_DELETE_BATCH", "LINK_CREATE", "LINK_UPDATE",
 721         "LINK_GET_FD_BY_ID", "LINK_GET_NEXT_ID", "ENABLE_STATS", "ITER_CREATE",
 722         "LINK_DETACH", "PROG_BIND_MAP",
 723 };
 724 static DEFINE_STRARRAY(bpf_cmd, "BPF_");
 725
 726 static const char *fsmount_flags[] = {
 727         [1] = "CLOEXEC",
 728 };
 729 static DEFINE_STRARRAY(fsmount_flags, "FSMOUNT_");
 730
 731 #include "trace/beauty/generated/fsconfig_arrays.c"
 732
 733 static DEFINE_STRARRAY(fsconfig_cmds, "FSCONFIG_");
 734
 735 static const char *epoll_ctl_ops[] = { "ADD", "DEL", "MOD", };
 736 static DEFINE_STRARRAY_OFFSET(epoll_ctl_ops, "EPOLL_CTL_", 1);
 737
 738 static const char *itimers[] = { "REAL", "VIRTUAL", "PROF", };
 739 static DEFINE_STRARRAY(itimers, "ITIMER_");
 740
 741 static const char *keyctl_options[] = {
 742         "GET_KEYRING_ID", "JOIN_SESSION_KEYRING", "UPDATE", "REVOKE", "CHOWN",
 743         "SETPERM", "DESCRIBE", "CLEAR", "LINK", "UNLINK", "SEARCH", "READ",
 744         "INSTANTIATE", "NEGATE", "SET_REQKEY_KEYRING", "SET_TIMEOUT",
 745         "ASSUME_AUTHORITY", "GET_SECURITY", "SESSION_TO_PARENT", "REJECT",
 746         "INSTANTIATE_IOV", "INVALIDATE", "GET_PERSISTENT",
 747 };
 748 static DEFINE_STRARRAY(keyctl_options, "KEYCTL_");
 749
 750 static const char *whences[] = { "SET", "CUR", "END",
 751 #ifdef SEEK_DATA
 752 "DATA",
 753 #endif
 754 #ifdef SEEK_HOLE
 755 "HOLE",
 756 #endif
 757 };
 758 static DEFINE_STRARRAY(whences, "SEEK_");
 759
 760 static const char *fcntl_cmds[] = {
 761         "DUPFD", "GETFD", "SETFD", "GETFL", "SETFL", "GETLK", "SETLK",
 762         "SETLKW", "SETOWN", "GETOWN", "SETSIG", "GETSIG", "GETLK64",
 763         "SETLK64", "SETLKW64", "SETOWN_EX", "GETOWN_EX",
 764         "GETOWNER_UIDS",
 765 };
 766 static DEFINE_STRARRAY(fcntl_cmds, "F_");
 767
 768 static const char *fcntl_linux_specific_cmds[] = {
 769         "SETLEASE", "GETLEASE", "NOTIFY", [5] = "CANCELLK", "DUPFD_CLOEXEC",
 770         "SETPIPE_SZ", "GETPIPE_SZ", "ADD_SEALS", "GET_SEALS",
 771         "GET_RW_HINT", "SET_RW_HINT", "GET_FILE_RW_HINT", "SET_FILE_RW_HINT",
 772 };
 773
 774 static DEFINE_STRARRAY_OFFSET(fcntl_linux_specific_cmds, "F_", F_LINUX_SPECIFIC_BASE);
 775
 776 static struct strarray *fcntl_cmds_arrays[] = {
 777         &strarray__fcntl_cmds,
 778         &strarray__fcntl_linux_specific_cmds,
 779 };
 780
 781 static DEFINE_STRARRAYS(fcntl_cmds_arrays);
 782
 783 static const char *rlimit_resources[] = {
 784         "CPU", "FSIZE", "DATA", "STACK", "CORE", "RSS", "NPROC", "NOFILE",
 785         "MEMLOCK", "AS", "LOCKS", "SIGPENDING", "MSGQUEUE", "NICE", "RTPRIO",
 786         "RTTIME",
 787 };
 788 static DEFINE_STRARRAY(rlimit_resources, "RLIMIT_");
 789
 790 static const char *sighow[] = { "BLOCK", "UNBLOCK", "SETMASK", };
 791 static DEFINE_STRARRAY(sighow, "SIG_");
 792
 793 static const char *clockid[] = {
 794         "REALTIME", "MONOTONIC", "PROCESS_CPUTIME_ID", "THREAD_CPUTIME_ID",
 795         "MONOTONIC_RAW", "REALTIME_COARSE", "MONOTONIC_COARSE", "BOOTTIME",
 796         "REALTIME_ALARM", "BOOTTIME_ALARM", "SGI_CYCLE", "TAI"
 797 };
 798 static DEFINE_STRARRAY(clockid, "CLOCK_");
 799
 800 static size_t syscall_arg__scnprintf_access_mode(char *bf, size_t size,
 801                                                  struct syscall_arg *arg)
 802 {
 803         bool show_prefix = arg->show_string_prefix;
 804         const char *suffix = "_OK";
 805         size_t printed = 0;
 806         int mode = arg->val;
 807
 808         if (mode == F_OK) /* 0 */
 809                 return scnprintf(bf, size, "F%s", show_prefix ? suffix : "");
 810 #define P_MODE(n) \
 811         if (mode & n##_OK) { \
 812                 printed += scnprintf(bf + printed, size - printed, "%s%s", #n, show_prefix ? suffix : ""); \
 813                 mode &= ~n##_OK; \
 814         }
 815
 816         P_MODE(R);
 817         P_MODE(W);
 818         P_MODE(X);
 819 #undef P_MODE
 820
 821         if (mode)
 822                 printed += scnprintf(bf + printed, size - printed, "|%#x", mode);
 823
 824         return printed;
 825 }
 826
 827 #define SCA_ACCMODE syscall_arg__scnprintf_access_mode
 828
 829 static size_t syscall_arg__scnprintf_filename(char *bf, size_t size,
 830                                               struct syscall_arg *arg);
 831
 832 #define SCA_FILENAME syscall_arg__scnprintf_filename
 833
 834 static size_t syscall_arg__scnprintf_pipe_flags(char *bf, size_t size,
 835                                                 struct syscall_arg *arg)
 836 {
 837         bool show_prefix = arg->show_string_prefix;
 838         const char *prefix = "O_";
 839         int printed = 0, flags = arg->val;
 840
 841 #define P_FLAG(n) \
 842         if (flags & O_##n) { \
 843                 printed += scnprintf(bf + printed, size - printed, "%s%s%s", printed ? "|" : "", show_prefix ? prefix : "", #n); \
 844                 flags &= ~O_##n; \
 845         }
 846
 847         P_FLAG(CLOEXEC);
 848         P_FLAG(NONBLOCK);
 849 #undef P_FLAG
 850
 851         if (flags)
 852                 printed += scnprintf(bf + printed, size - printed, "%s%#x", printed ? "|" : "", flags);
 853
 854         return printed;
 855 }
 856
 857 #define SCA_PIPE_FLAGS syscall_arg__scnprintf_pipe_flags
 858
 859 #ifndef GRND_NONBLOCK
 860 #define GRND_NONBLOCK   0x0001
 861 #endif
 862 #ifndef GRND_RANDOM
 863 #define GRND_RANDOM     0x0002
 864 #endif
 865
 866 static size_t syscall_arg__scnprintf_getrandom_flags(char *bf, size_t size,
 867                                                    struct syscall_arg *arg)
 868 {
 869         bool show_prefix = arg->show_string_prefix;
 870         const char *prefix = "GRND_";
 871         int printed = 0, flags = arg->val;
 872
 873 #define P_FLAG(n) \
 874         if (flags & GRND_##n) { \
 875                 printed += scnprintf(bf + printed, size - printed, "%s%s%s", printed ? "|" : "", show_prefix ? prefix : "", #n); \
 876                 flags &= ~GRND_##n; \
 877         }
 878
 879         P_FLAG(RANDOM);
 880         P_FLAG(NONBLOCK);
 881 #undef P_FLAG
 882
 883         if (flags)
 884                 printed += scnprintf(bf + printed, size - printed, "%s%#x", printed ? "|" : "", flags);
 885
 886         return printed;
 887 }
 888
 889 #define SCA_GETRANDOM_FLAGS syscall_arg__scnprintf_getrandom_flags
 890
 891 #define STRARRAY(name, array) \
 892           { .scnprintf  = SCA_STRARRAY, \
 893             .strtoul    = STUL_STRARRAY, \
 894             .parm       = &strarray__##array, }
 895
 896 #define STRARRAY_FLAGS(name, array) \
 897           { .scnprintf  = SCA_STRARRAY_FLAGS, \
 898             .strtoul    = STUL_STRARRAY_FLAGS, \
 899             .parm       = &strarray__##array, }
 900
 901 #include "trace/beauty/arch_errno_names.c"
 902 #include "trace/beauty/eventfd.c"
 903 #include "trace/beauty/futex_op.c"
 904 #include "trace/beauty/futex_val3.c"
 905 #include "trace/beauty/mmap.c"
 906 #include "trace/beauty/mode_t.c"
 907 #include "trace/beauty/msg_flags.c"
 908 #include "trace/beauty/open_flags.c"
 909 #include "trace/beauty/perf_event_open.c"
 910 #include "trace/beauty/pid.c"
 911 #include "trace/beauty/sched_policy.c"
 912 #include "trace/beauty/seccomp.c"
 913 #include "trace/beauty/signum.c"
 914 #include "trace/beauty/socket_type.c"
 915 #include "trace/beauty/waitid_options.c"
 916
 917 static const struct syscall_fmt syscall_fmts[] = {
 918         { .name     = "access",
 919           .arg = { [1] = { .scnprintf = SCA_ACCMODE,  /* mode */ }, }, },
 920         { .name     = "arch_prctl",
 921           .arg = { [0] = { .scnprintf = SCA_X86_ARCH_PRCTL_CODE, /* code */ },
 922                    [1] = { .scnprintf = SCA_PTR, /* arg2 */ }, }, },
 923         { .name     = "bind",
 924           .arg = { [0] = { .scnprintf = SCA_INT, /* fd */ },
 925                    [1] = { .scnprintf = SCA_SOCKADDR, /* umyaddr */ },
 926                    [2] = { .scnprintf = SCA_INT, /* addrlen */ }, }, },
 927         { .name     = "bpf",
 928           .arg = { [0] = STRARRAY(cmd, bpf_cmd), }, },
 929         { .name     = "brk",        .hexret = true,
 930           .arg = { [0] = { .scnprintf = SCA_PTR, /* brk */ }, }, },
 931         { .name     = "clock_gettime",
 932           .arg = { [0] = STRARRAY(clk_id, clockid), }, },
 933         { .name     = "clock_nanosleep",
 934           .arg = { [2] = { .scnprintf = SCA_TIMESPEC,  /* rqtp */ }, }, },
 935         { .name     = "clone",      .errpid = true, .nr_args = 5,
 936           .arg = { [0] = { .name = "flags",         .scnprintf = SCA_CLONE_FLAGS, },
 937                    [1] = { .name = "child_stack",   .scnprintf = SCA_HEX, },
 938                    [2] = { .name = "parent_tidptr", .scnprintf = SCA_HEX, },
 939                    [3] = { .name = "child_tidptr",  .scnprintf = SCA_HEX, },
 940                    [4] = { .name = "tls",           .scnprintf = SCA_HEX, }, }, },
 941         { .name     = "close",
 942           .arg = { [0] = { .scnprintf = SCA_CLOSE_FD, /* fd */ }, }, },
 943         { .name     = "connect",
 944           .arg = { [0] = { .scnprintf = SCA_INT, /* fd */ },
 945                    [1] = { .scnprintf = SCA_SOCKADDR, /* servaddr */ },
 946                    [2] = { .scnprintf = SCA_INT, /* addrlen */ }, }, },
 947         { .name     = "epoll_ctl",
 948           .arg = { [1] = STRARRAY(op, epoll_ctl_ops), }, },
 949         { .name     = "eventfd2",
 950           .arg = { [1] = { .scnprintf = SCA_EFD_FLAGS, /* flags */ }, }, },
 951         { .name     = "fchmodat",
 952           .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
 953         { .name     = "fchownat",
 954           .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
 955         { .name     = "fcntl",
 956           .arg = { [1] = { .scnprintf = SCA_FCNTL_CMD,  /* cmd */
 957                            .strtoul   = STUL_STRARRAYS,
 958                            .parm      = &strarrays__fcntl_cmds_arrays,
 959                            .show_zero = true, },
 960                    [2] = { .scnprintf =  SCA_FCNTL_ARG, /* arg */ }, }, },
 961         { .name     = "flock",
 962           .arg = { [1] = { .scnprintf = SCA_FLOCK, /* cmd */ }, }, },
 963         { .name     = "fsconfig",
 964           .arg = { [1] = STRARRAY(cmd, fsconfig_cmds), }, },
 965         { .name     = "fsmount",
 966           .arg = { [1] = STRARRAY_FLAGS(flags, fsmount_flags),
 967                    [2] = { .scnprintf = SCA_FSMOUNT_ATTR_FLAGS, /* attr_flags */ }, }, },
 968         { .name     = "fspick",
 969           .arg = { [0] = { .scnprintf = SCA_FDAT,         /* dfd */ },
 970                    [1] = { .scnprintf = SCA_FILENAME,     /* path */ },
 971                    [2] = { .scnprintf = SCA_FSPICK_FLAGS, /* flags */ }, }, },
 972         { .name     = "fstat", .alias = "newfstat", },
 973         { .name     = "fstatat", .alias = "newfstatat", },
 974         { .name     = "futex",
 975           .arg = { [1] = { .scnprintf = SCA_FUTEX_OP, /* op */ },
 976                    [5] = { .scnprintf = SCA_FUTEX_VAL3, /* val3 */ }, }, },
 977         { .name     = "futimesat",
 978           .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
 979         { .name     = "getitimer",
 980           .arg = { [0] = STRARRAY(which, itimers), }, },
 981         { .name     = "getpid",     .errpid = true, },
 982         { .name     = "getpgid",    .errpid = true, },
 983         { .name     = "getppid",    .errpid = true, },
 984         { .name     = "getrandom",
 985           .arg = { [2] = { .scnprintf = SCA_GETRANDOM_FLAGS, /* flags */ }, }, },
 986         { .name     = "getrlimit",
 987           .arg = { [0] = STRARRAY(resource, rlimit_resources), }, },
 988         { .name     = "getsockopt",
 989           .arg = { [1] = STRARRAY(level, socket_level), }, },
 990         { .name     = "gettid",     .errpid = true, },
 991         { .name     = "ioctl",
 992           .arg = {
 993 #if defined(__i386__) || defined(__x86_64__)
 994 /*
 995  * FIXME: Make this available to all arches.
 996  */
 997                    [1] = { .scnprintf = SCA_IOCTL_CMD, /* cmd */ },
 998                    [2] = { .scnprintf = SCA_HEX, /* arg */ }, }, },
 999 #else
1000                    [2] = { .scnprintf = SCA_HEX, /* arg */ }, }, },
1001 #endif
1002         { .name     = "kcmp",       .nr_args = 5,
1003           .arg = { [0] = { .name = "pid1",      .scnprintf = SCA_PID, },
1004                    [1] = { .name = "pid2",      .scnprintf = SCA_PID, },
1005                    [2] = { .name = "type",      .scnprintf = SCA_KCMP_TYPE, },
1006                    [3] = { .name = "idx1",      .scnprintf = SCA_KCMP_IDX, },
1007                    [4] = { .name = "idx2",      .scnprintf = SCA_KCMP_IDX, }, }, },
1008         { .name     = "keyctl",
1009           .arg = { [0] = STRARRAY(option, keyctl_options), }, },
1010         { .name     = "kill",
1011           .arg = { [1] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
1012         { .name     = "linkat",
1013           .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
1014         { .name     = "lseek",
1015           .arg = { [2] = STRARRAY(whence, whences), }, },
1016         { .name     = "lstat", .alias = "newlstat", },
1017         { .name     = "madvise",
1018           .arg = { [0] = { .scnprintf = SCA_HEX,      /* start */ },
1019                    [2] = { .scnprintf = SCA_MADV_BHV, /* behavior */ }, }, },
1020         { .name     = "mkdirat",
1021           .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
1022         { .name     = "mknodat",
1023           .arg = { [0] = { .scnprintf = SCA_FDAT, /* fd */ }, }, },
1024         { .name     = "mmap",       .hexret = true,
1025 /* The standard mmap maps to old_mmap on s390x */
1026 #if defined(__s390x__)
1027         .alias = "old_mmap",
1028 #endif
1029           .arg = { [2] = { .scnprintf = SCA_MMAP_PROT,  /* prot */ },
1030                    [3] = { .scnprintf = SCA_MMAP_FLAGS, /* flags */
1031                            .strtoul   = STUL_STRARRAY_FLAGS,
1032                            .parm      = &strarray__mmap_flags, },
1033                    [5] = { .scnprintf = SCA_HEX,        /* offset */ }, }, },
1034         { .name     = "mount",
1035           .arg = { [0] = { .scnprintf = SCA_FILENAME, /* dev_name */ },
1036                    [3] = { .scnprintf = SCA_MOUNT_FLAGS, /* flags */
1037                            .mask_val  = SCAMV_MOUNT_FLAGS, /* flags */ }, }, },
1038         { .name     = "move_mount",
1039           .arg = { [0] = { .scnprintf = SCA_FDAT,       /* from_dfd */ },
1040                    [1] = { .scnprintf = SCA_FILENAME, /* from_pathname */ },
1041                    [2] = { .scnprintf = SCA_FDAT,       /* to_dfd */ },
1042                    [3] = { .scnprintf = SCA_FILENAME, /* to_pathname */ },
1043                    [4] = { .scnprintf = SCA_MOVE_MOUNT_FLAGS, /* flags */ }, }, },
1044         { .name     = "mprotect",
1045           .arg = { [0] = { .scnprintf = SCA_HEX,        /* start */ },
1046                    [2] = { .scnprintf = SCA_MMAP_PROT,  /* prot */ }, }, },
1047         { .name     = "mq_unlink",
1048           .arg = { [0] = { .scnprintf = SCA_FILENAME, /* u_name */ }, }, },
1049         { .name     = "mremap",     .hexret = true,
1050           .arg = { [3] = { .scnprintf = SCA_MREMAP_FLAGS, /* flags */ }, }, },
1051         { .name     = "name_to_handle_at",
1052           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
1053         { .name     = "newfstatat",
1054           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
1055         { .name     = "open",
1056           .arg = { [1] = { .scnprintf = SCA_OPEN_FLAGS, /* flags */ }, }, },
1057         { .name     = "open_by_handle_at",
1058           .arg = { [0] = { .scnprintf = SCA_FDAT,       /* dfd */ },
1059                    [2] = { .scnprintf = SCA_OPEN_FLAGS, /* flags */ }, }, },
1060         { .name     = "openat",
1061           .arg = { [0] = { .scnprintf = SCA_FDAT,       /* dfd */ },
1062                    [2] = { .scnprintf = SCA_OPEN_FLAGS, /* flags */ }, }, },
1063         { .name     = "perf_event_open",
1064           .arg = { [0] = { .scnprintf = SCA_PERF_ATTR,  /* attr */ },
1065                    [2] = { .scnprintf = SCA_INT,        /* cpu */ },
1066                    [3] = { .scnprintf = SCA_FD,         /* group_fd */ },
1067                    [4] = { .scnprintf = SCA_PERF_FLAGS, /* flags */ }, }, },
1068         { .name     = "pipe2",
1069           .arg = { [1] = { .scnprintf = SCA_PIPE_FLAGS, /* flags */ }, }, },
1070         { .name     = "pkey_alloc",
1071           .arg = { [1] = { .scnprintf = SCA_PKEY_ALLOC_ACCESS_RIGHTS,   /* access_rights */ }, }, },
1072         { .name     = "pkey_free",
1073           .arg = { [0] = { .scnprintf = SCA_INT,        /* key */ }, }, },
1074         { .name     = "pkey_mprotect",
1075           .arg = { [0] = { .scnprintf = SCA_HEX,        /* start */ },
1076                    [2] = { .scnprintf = SCA_MMAP_PROT,  /* prot */ },
1077                    [3] = { .scnprintf = SCA_INT,        /* pkey */ }, }, },
1078         { .name     = "poll", .timeout = true, },
1079         { .name     = "ppoll", .timeout = true, },
1080         { .name     = "prctl",
1081           .arg = { [0] = { .scnprintf = SCA_PRCTL_OPTION, /* option */
1082                            .strtoul   = STUL_STRARRAY,
1083                            .parm      = &strarray__prctl_options, },
1084                    [1] = { .scnprintf = SCA_PRCTL_ARG2, /* arg2 */ },
1085                    [2] = { .scnprintf = SCA_PRCTL_ARG3, /* arg3 */ }, }, },
1086         { .name     = "pread", .alias = "pread64", },
1087         { .name     = "preadv", .alias = "pread", },
1088         { .name     = "prlimit64",
1089           .arg = { [1] = STRARRAY(resource, rlimit_resources), }, },
1090         { .name     = "pwrite", .alias = "pwrite64", },
1091         { .name     = "readlinkat",
1092           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
1093         { .name     = "recvfrom",
1094           .arg = { [3] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
1095         { .name     = "recvmmsg",
1096           .arg = { [3] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
1097         { .name     = "recvmsg",
1098           .arg = { [2] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
1099         { .name     = "renameat",
1100           .arg = { [0] = { .scnprintf = SCA_FDAT, /* olddirfd */ },
1101                    [2] = { .scnprintf = SCA_FDAT, /* newdirfd */ }, }, },
1102         { .name     = "renameat2",
1103           .arg = { [0] = { .scnprintf = SCA_FDAT, /* olddirfd */ },
1104                    [2] = { .scnprintf = SCA_FDAT, /* newdirfd */ },
1105                    [4] = { .scnprintf = SCA_RENAMEAT2_FLAGS, /* flags */ }, }, },
1106         { .name     = "rt_sigaction",
1107           .arg = { [0] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
1108         { .name     = "rt_sigprocmask",
1109           .arg = { [0] = STRARRAY(how, sighow), }, },
1110         { .name     = "rt_sigqueueinfo",
1111           .arg = { [1] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
1112         { .name     = "rt_tgsigqueueinfo",
1113           .arg = { [2] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
1114         { .name     = "sched_setscheduler",
1115           .arg = { [1] = { .scnprintf = SCA_SCHED_POLICY, /* policy */ }, }, },
1116         { .name     = "seccomp",
1117           .arg = { [0] = { .scnprintf = SCA_SECCOMP_OP,    /* op */ },
1118                    [1] = { .scnprintf = SCA_SECCOMP_FLAGS, /* flags */ }, }, },
1119         { .name     = "select", .timeout = true, },
1120         { .name     = "sendfile", .alias = "sendfile64", },
1121         { .name     = "sendmmsg",
1122           .arg = { [3] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
1123         { .name     = "sendmsg",
1124           .arg = { [2] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ }, }, },
1125         { .name     = "sendto",
1126           .arg = { [3] = { .scnprintf = SCA_MSG_FLAGS, /* flags */ },
1127                    [4] = { .scnprintf = SCA_SOCKADDR, /* addr */ }, }, },
1128         { .name     = "set_tid_address", .errpid = true, },
1129         { .name     = "setitimer",
1130           .arg = { [0] = STRARRAY(which, itimers), }, },
1131         { .name     = "setrlimit",
1132           .arg = { [0] = STRARRAY(resource, rlimit_resources), }, },
1133         { .name     = "setsockopt",
1134           .arg = { [1] = STRARRAY(level, socket_level), }, },
1135         { .name     = "socket",
1136           .arg = { [0] = STRARRAY(family, socket_families),
1137                    [1] = { .scnprintf = SCA_SK_TYPE, /* type */ },
1138                    [2] = { .scnprintf = SCA_SK_PROTO, /* protocol */ }, }, },
1139         { .name     = "socketpair",
1140           .arg = { [0] = STRARRAY(family, socket_families),
1141                    [1] = { .scnprintf = SCA_SK_TYPE, /* type */ },
1142                    [2] = { .scnprintf = SCA_SK_PROTO, /* protocol */ }, }, },
1143         { .name     = "stat", .alias = "newstat", },
1144         { .name     = "statx",
1145           .arg = { [0] = { .scnprintf = SCA_FDAT,        /* fdat */ },
1146                    [2] = { .scnprintf = SCA_STATX_FLAGS, /* flags */ } ,
1147                    [3] = { .scnprintf = SCA_STATX_MASK,  /* mask */ }, }, },
1148         { .name     = "swapoff",
1149           .arg = { [0] = { .scnprintf = SCA_FILENAME, /* specialfile */ }, }, },
1150         { .name     = "swapon",
1151           .arg = { [0] = { .scnprintf = SCA_FILENAME, /* specialfile */ }, }, },
1152         { .name     = "symlinkat",
1153           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
1154         { .name     = "sync_file_range",
1155           .arg = { [3] = { .scnprintf = SCA_SYNC_FILE_RANGE_FLAGS, /* flags */ }, }, },
1156         { .name     = "tgkill",
1157           .arg = { [2] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
1158         { .name     = "tkill",
1159           .arg = { [1] = { .scnprintf = SCA_SIGNUM, /* sig */ }, }, },
1160         { .name     = "umount2", .alias = "umount",
1161           .arg = { [0] = { .scnprintf = SCA_FILENAME, /* name */ }, }, },
1162         { .name     = "uname", .alias = "newuname", },
1163         { .name     = "unlinkat",
1164           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dfd */ }, }, },
1165         { .name     = "utimensat",
1166           .arg = { [0] = { .scnprintf = SCA_FDAT, /* dirfd */ }, }, },
1167         { .name     = "wait4",      .errpid = true,
1168           .arg = { [2] = { .scnprintf = SCA_WAITID_OPTIONS, /* options */ }, }, },
1169         { .name     = "waitid",     .errpid = true,
1170           .arg = { [3] = { .scnprintf = SCA_WAITID_OPTIONS, /* options */ }, }, },
1171 };
1172
1173 static int syscall_fmt__cmp(const void *name, const void *fmtp)
1174 {
1175         const struct syscall_fmt *fmt = fmtp;
1176         return strcmp(name, fmt->name);
1177 }
1178
1179 static const struct syscall_fmt *__syscall_fmt__find(const struct syscall_fmt *fmts,
1180                                                      const int nmemb,
1181                                                      const char *name)
1182 {
1183         return bsearch(name, fmts, nmemb, sizeof(struct syscall_fmt), syscall_fmt__cmp);
1184 }
1185
1186 static const struct syscall_fmt *syscall_fmt__find(const char *name)
1187 {
1188         const int nmemb = ARRAY_SIZE(syscall_fmts);
1189         return __syscall_fmt__find(syscall_fmts, nmemb, name);
1190 }
1191
1192 static const struct syscall_fmt *__syscall_fmt__find_by_alias(const struct syscall_fmt *fmts,
1193                                                               const int nmemb, const char *alias)
1194 {
1195         int i;
1196
1197         for (i = 0; i < nmemb; ++i) {
1198                 if (fmts[i].alias && strcmp(fmts[i].alias, alias) == 0)
1199                         return &fmts[i];
1200         }
1201
1202         return NULL;
1203 }
1204
1205 static const struct syscall_fmt *syscall_fmt__find_by_alias(const char *alias)
1206 {
1207         const int nmemb = ARRAY_SIZE(syscall_fmts);
1208         return __syscall_fmt__find_by_alias(syscall_fmts, nmemb, alias);
1209 }
1210
1211 /*
1212  * is_exit: is this "exit" or "exit_group"?
1213  * is_open: is this "open" or "openat"? To associate the fd returned in sys_exit with the pathname in sys_enter.
1214  * args_size: sum of the sizes of the syscall arguments, anything after that is augmented stuff: pathname for openat, etc.
1215  * nonexistent: Just a hole in the syscall table, syscall id not allocated
1216  */
1217 struct syscall {
1218         struct tep_event    *tp_format;
1219         int                 nr_args;
1220         int                 args_size;
1221         struct {
1222                 struct bpf_program *sys_enter,
1223                                    *sys_exit;
1224         }                   bpf_prog;
1225         bool                is_exit;
1226         bool                is_open;
1227         bool                nonexistent;
1228         struct tep_format_field *args;
1229         const char          *name;
1230         const struct syscall_fmt  *fmt;
1231         struct syscall_arg_fmt *arg_fmt;
1232 };
1233
1234 /*
1235  * We need to have this 'calculated' boolean because in some cases we really
1236  * don't know what is the duration of a syscall, for instance, when we start
1237  * a session and some threads are waiting for a syscall to finish, say 'poll',
1238  * in which case all we can do is to print "( ? ) for duration and for the
1239  * start timestamp.
1240  */
1241 static size_t fprintf_duration(unsigned long t, bool calculated, FILE *fp)
1242 {
1243         double duration = (double)t / NSEC_PER_MSEC;
1244         size_t printed = fprintf(fp, "(");
1245
1246         if (!calculated)
1247                 printed += fprintf(fp, "         ");
1248         else if (duration >= 1.0)
1249                 printed += color_fprintf(fp, PERF_COLOR_RED, "%6.3f ms", duration);
1250         else if (duration >= 0.01)
1251                 printed += color_fprintf(fp, PERF_COLOR_YELLOW, "%6.3f ms", duration);
1252         else
1253                 printed += color_fprintf(fp, PERF_COLOR_NORMAL, "%6.3f ms", duration);
1254         return printed + fprintf(fp, "): ");
1255 }
1256
1257 /**
1258  * filename.ptr: The filename char pointer that will be vfs_getname'd
1259  * filename.entry_str_pos: Where to insert the string translated from
1260  *                         filename.ptr by the vfs_getname tracepoint/kprobe.
1261  * ret_scnprintf: syscall args may set this to a different syscall return
1262  *                formatter, for instance, fcntl may return fds, file flags, etc.
1263  */
1264 struct thread_trace {
1265         u64               entry_time;
1266         bool              entry_pending;
1267         unsigned long     nr_events;
1268         unsigned long     pfmaj, pfmin;
1269         char              *entry_str;
1270         double            runtime_ms;
1271         size_t            (*ret_scnprintf)(char *bf, size_t size, struct syscall_arg *arg);
1272         struct {
1273                 unsigned long ptr;
1274                 short int     entry_str_pos;
1275                 bool          pending_open;
1276                 unsigned int  namelen;
1277                 char          *name;
1278         } filename;
1279         struct {
1280                 int           max;
1281                 struct file   *table;
1282         } files;
1283
1284         struct intlist *syscall_stats;
1285 };
1286
1287 static struct thread_trace *thread_trace__new(void)
1288 {
1289         struct thread_trace *ttrace =  zalloc(sizeof(struct thread_trace));
1290
1291         if (ttrace) {
1292                 ttrace->files.max = -1;
1293                 ttrace->syscall_stats = intlist__new(NULL);
1294         }
1295
1296         return ttrace;
1297 }
1298
1299 static struct thread_trace *thread__trace(struct thread *thread, FILE *fp)
1300 {
1301         struct thread_trace *ttrace;
1302
1303         if (thread == NULL)
1304                 goto fail;
1305
1306         if (thread__priv(thread) == NULL)
1307                 thread__set_priv(thread, thread_trace__new());
1308
1309         if (thread__priv(thread) == NULL)
1310                 goto fail;
1311
1312         ttrace = thread__priv(thread);
1313         ++ttrace->nr_events;
1314
1315         return ttrace;
1316 fail:
1317         color_fprintf(fp, PERF_COLOR_RED,
1318                       "WARNING: not enough memory, dropping samples!\n");
1319         return NULL;
1320 }
1321
1322
1323 void syscall_arg__set_ret_scnprintf(struct syscall_arg *arg,
1324                                     size_t (*ret_scnprintf)(char *bf, size_t size, struct syscall_arg *arg))
1325 {
1326         struct thread_trace *ttrace = thread__priv(arg->thread);
1327
1328         ttrace->ret_scnprintf = ret_scnprintf;
1329 }
1330
1331 #define TRACE_PFMAJ             (1 << 0)
1332 #define TRACE_PFMIN             (1 << 1)
1333
1334 static const size_t trace__entry_str_size = 2048;
1335
1336 static struct file *thread_trace__files_entry(struct thread_trace *ttrace, int fd)
1337 {
1338         if (fd < 0)
1339                 return NULL;
1340
1341         if (fd > ttrace->files.max) {
1342                 struct file *nfiles = realloc(ttrace->files.table, (fd + 1) * sizeof(struct file));
1343
1344                 if (nfiles == NULL)
1345                         return NULL;
1346
1347                 if (ttrace->files.max != -1) {
1348                         memset(nfiles + ttrace->files.max + 1, 0,
1349                                (fd - ttrace->files.max) * sizeof(struct file));
1350                 } else {
1351                         memset(nfiles, 0, (fd + 1) * sizeof(struct file));
1352                 }
1353
1354                 ttrace->files.table = nfiles;
1355                 ttrace->files.max   = fd;
1356         }
1357
1358         return ttrace->files.table + fd;
1359 }
1360
1361 struct file *thread__files_entry(struct thread *thread, int fd)
1362 {
1363         return thread_trace__files_entry(thread__priv(thread), fd);
1364 }
1365
1366 static int trace__set_fd_pathname(struct thread *thread, int fd, const char *pathname)
1367 {
1368         struct thread_trace *ttrace = thread__priv(thread);
1369         struct file *file = thread_trace__files_entry(ttrace, fd);
1370
1371         if (file != NULL) {
1372                 struct stat st;
1373                 if (stat(pathname, &st) == 0)
1374                         file->dev_maj = major(st.st_rdev);
1375                 file->pathname = strdup(pathname);
1376                 if (file->pathname)
1377                         return 0;
1378         }
1379
1380         return -1;
1381 }
1382
1383 static int thread__read_fd_path(struct thread *thread, int fd)
1384 {
1385         char linkname[PATH_MAX], pathname[PATH_MAX];
1386         struct stat st;
1387         int ret;
1388
1389         if (thread__pid(thread) == thread__tid(thread)) {
1390                 scnprintf(linkname, sizeof(linkname),
1391                           "/proc/%d/fd/%d", thread__pid(thread), fd);
1392         } else {
1393                 scnprintf(linkname, sizeof(linkname),
1394                           "/proc/%d/task/%d/fd/%d",
1395                           thread__pid(thread), thread__tid(thread), fd);
1396         }
1397
1398         if (lstat(linkname, &st) < 0 || st.st_size + 1 > (off_t)sizeof(pathname))
1399                 return -1;
1400
1401         ret = readlink(linkname, pathname, sizeof(pathname));
1402
1403         if (ret < 0 || ret > st.st_size)
1404                 return -1;
1405
1406         pathname[ret] = '\0';
1407         return trace__set_fd_pathname(thread, fd, pathname);
1408 }
1409
1410 static const char *thread__fd_path(struct thread *thread, int fd,
1411                                    struct trace *trace)
1412 {
1413         struct thread_trace *ttrace = thread__priv(thread);
1414
1415         if (ttrace == NULL || trace->fd_path_disabled)
1416                 return NULL;
1417
1418         if (fd < 0)
1419                 return NULL;
1420
1421         if ((fd > ttrace->files.max || ttrace->files.table[fd].pathname == NULL)) {
1422                 if (!trace->live)
1423                         return NULL;
1424                 ++trace->stats.proc_getname;
1425                 if (thread__read_fd_path(thread, fd))
1426                         return NULL;
1427         }
1428
1429         return ttrace->files.table[fd].pathname;
1430 }
1431
1432 size_t syscall_arg__scnprintf_fd(char *bf, size_t size, struct syscall_arg *arg)
1433 {
1434         int fd = arg->val;
1435         size_t printed = scnprintf(bf, size, "%d", fd);
1436         const char *path = thread__fd_path(arg->thread, fd, arg->trace);
1437
1438         if (path)
1439                 printed += scnprintf(bf + printed, size - printed, "<%s>", path);
1440
1441         return printed;
1442 }
1443
1444 size_t pid__scnprintf_fd(struct trace *trace, pid_t pid, int fd, char *bf, size_t size)
1445 {
1446         size_t printed = scnprintf(bf, size, "%d", fd);
1447         struct thread *thread = machine__find_thread(trace->host, pid, pid);
1448
1449         if (thread) {
1450                 const char *path = thread__fd_path(thread, fd, trace);
1451
1452                 if (path)
1453                         printed += scnprintf(bf + printed, size - printed, "<%s>", path);
1454
1455                 thread__put(thread);
1456         }
1457
1458         return printed;
1459 }
1460
1461 static size_t syscall_arg__scnprintf_close_fd(char *bf, size_t size,
1462                                               struct syscall_arg *arg)
1463 {
1464         int fd = arg->val;
1465         size_t printed = syscall_arg__scnprintf_fd(bf, size, arg);
1466         struct thread_trace *ttrace = thread__priv(arg->thread);
1467
1468         if (ttrace && fd >= 0 && fd <= ttrace->files.max)
1469                 zfree(&ttrace->files.table[fd].pathname);
1470
1471         return printed;
1472 }
1473
1474 static void thread__set_filename_pos(struct thread *thread, const char *bf,
1475                                      unsigned long ptr)
1476 {
1477         struct thread_trace *ttrace = thread__priv(thread);
1478
1479         ttrace->filename.ptr = ptr;
1480         ttrace->filename.entry_str_pos = bf - ttrace->entry_str;
1481 }
1482
1483 static size_t syscall_arg__scnprintf_augmented_string(struct syscall_arg *arg, char *bf, size_t size)
1484 {
1485         struct augmented_arg *augmented_arg = arg->augmented.args;
1486         size_t printed = scnprintf(bf, size, "\"%.*s\"", augmented_arg->size, augmented_arg->value);
1487         /*
1488          * So that the next arg with a payload can consume its augmented arg, i.e. for rename* syscalls
1489          * we would have two strings, each prefixed by its size.
1490          */
1491         int consumed = sizeof(*augmented_arg) + augmented_arg->size;
1492
1493         arg->augmented.args = ((void *)arg->augmented.args) + consumed;
1494         arg->augmented.size -= consumed;
1495
1496         return printed;
1497 }
1498
1499 static size_t syscall_arg__scnprintf_filename(char *bf, size_t size,
1500                                               struct syscall_arg *arg)
1501 {
1502         unsigned long ptr = arg->val;
1503
1504         if (arg->augmented.args)
1505                 return syscall_arg__scnprintf_augmented_string(arg, bf, size);
1506
1507         if (!arg->trace->vfs_getname)
1508                 return scnprintf(bf, size, "%#x", ptr);
1509
1510         thread__set_filename_pos(arg->thread, bf, ptr);
1511         return 0;
1512 }
1513
1514 static bool trace__filter_duration(struct trace *trace, double t)
1515 {
1516         return t < (trace->duration_filter * NSEC_PER_MSEC);
1517 }
1518
1519 static size_t __trace__fprintf_tstamp(struct trace *trace, u64 tstamp, FILE *fp)
1520 {
1521         double ts = (double)(tstamp - trace->base_time) / NSEC_PER_MSEC;
1522
1523         return fprintf(fp, "%10.3f ", ts);
1524 }
1525
1526 /*
1527  * We're handling tstamp=0 as an undefined tstamp, i.e. like when we are
1528  * using ttrace->entry_time for a thread that receives a sys_exit without
1529  * first having received a sys_enter ("poll" issued before tracing session
1530  * starts, lost sys_enter exit due to ring buffer overflow).
1531  */
1532 static size_t trace__fprintf_tstamp(struct trace *trace, u64 tstamp, FILE *fp)
1533 {
1534         if (tstamp > 0)
1535                 return __trace__fprintf_tstamp(trace, tstamp, fp);
1536
1537         return fprintf(fp, "         ? ");
1538 }
1539
1540 static pid_t workload_pid = -1;
1541 static volatile sig_atomic_t done = false;
1542 static volatile sig_atomic_t interrupted = false;
1543
1544 static void sighandler_interrupt(int sig __maybe_unused)
1545 {
1546         done = interrupted = true;
1547 }
1548
1549 static void sighandler_chld(int sig __maybe_unused, siginfo_t *info,
1550                             void *context __maybe_unused)
1551 {
1552         if (info->si_pid == workload_pid)
1553                 done = true;
1554 }
1555
1556 static size_t trace__fprintf_comm_tid(struct trace *trace, struct thread *thread, FILE *fp)
1557 {
1558         size_t printed = 0;
1559
1560         if (trace->multiple_threads) {
1561                 if (trace->show_comm)
1562                         printed += fprintf(fp, "%.14s/", thread__comm_str(thread));
1563                 printed += fprintf(fp, "%d ", thread__tid(thread));
1564         }
1565
1566         return printed;
1567 }
1568
1569 static size_t trace__fprintf_entry_head(struct trace *trace, struct thread *thread,
1570                                         u64 duration, bool duration_calculated, u64 tstamp, FILE *fp)
1571 {
1572         size_t printed = 0;
1573
1574         if (trace->show_tstamp)
1575                 printed = trace__fprintf_tstamp(trace, tstamp, fp);
1576         if (trace->show_duration)
1577                 printed += fprintf_duration(duration, duration_calculated, fp);
1578         return printed + trace__fprintf_comm_tid(trace, thread, fp);
1579 }
1580
1581 static int trace__process_event(struct trace *trace, struct machine *machine,
1582                                 union perf_event *event, struct perf_sample *sample)
1583 {
1584         int ret = 0;
1585
1586         switch (event->header.type) {
1587         case PERF_RECORD_LOST:
1588                 color_fprintf(trace->output, PERF_COLOR_RED,
1589                               "LOST %" PRIu64 " events!\n", event->lost.lost);
1590                 ret = machine__process_lost_event(machine, event, sample);
1591                 break;
1592         default:
1593                 ret = machine__process_event(machine, event, sample);
1594                 break;
1595         }
1596
1597         return ret;
1598 }
1599
1600 static int trace__tool_process(struct perf_tool *tool,
1601                                union perf_event *event,
1602                                struct perf_sample *sample,
1603                                struct machine *machine)
1604 {
1605         struct trace *trace = container_of(tool, struct trace, tool);
1606         return trace__process_event(trace, machine, event, sample);
1607 }
1608
1609 static char *trace__machine__resolve_kernel_addr(void *vmachine, unsigned long long *addrp, char **modp)
1610 {
1611         struct machine *machine = vmachine;
1612
1613         if (machine->kptr_restrict_warned)
1614                 return NULL;
1615
1616         if (symbol_conf.kptr_restrict) {
1617                 pr_warning("Kernel address maps (/proc/{kallsyms,modules}) are restricted.\n\n"
1618                            "Check /proc/sys/kernel/kptr_restrict and /proc/sys/kernel/perf_event_paranoid.\n\n"
1619                            "Kernel samples will not be resolved.\n");
1620                 machine->kptr_restrict_warned = true;
1621                 return NULL;
1622         }
1623
1624         return machine__resolve_kernel_addr(vmachine, addrp, modp);
1625 }
1626
1627 static int trace__symbols_init(struct trace *trace, struct evlist *evlist)
1628 {
1629         int err = symbol__init(NULL);
1630
1631         if (err)
1632                 return err;
1633
1634         trace->host = machine__new_host();
1635         if (trace->host == NULL)
1636                 return -ENOMEM;
1637
1638         err = trace_event__register_resolver(trace->host, trace__machine__resolve_kernel_addr);
1639         if (err < 0)
1640                 goto out;
1641
1642         err = __machine__synthesize_threads(trace->host, &trace->tool, &trace->opts.target,
1643                                             evlist->core.threads, trace__tool_process,
1644                                             true, false, 1);
1645 out:
1646         if (err)
1647                 symbol__exit();
1648
1649         return err;
1650 }
1651
1652 static void trace__symbols__exit(struct trace *trace)
1653 {
1654         machine__exit(trace->host);
1655         trace->host = NULL;
1656
1657         symbol__exit();
1658 }
1659
1660 static int syscall__alloc_arg_fmts(struct syscall *sc, int nr_args)
1661 {
1662         int idx;
1663
1664         if (nr_args == RAW_SYSCALL_ARGS_NUM && sc->fmt && sc->fmt->nr_args != 0)
1665                 nr_args = sc->fmt->nr_args;
1666
1667         sc->arg_fmt = calloc(nr_args, sizeof(*sc->arg_fmt));
1668         if (sc->arg_fmt == NULL)
1669                 return -1;
1670
1671         for (idx = 0; idx < nr_args; ++idx) {
1672                 if (sc->fmt)
1673                         sc->arg_fmt[idx] = sc->fmt->arg[idx];
1674         }
1675
1676         sc->nr_args = nr_args;
1677         return 0;
1678 }
1679
1680 static const struct syscall_arg_fmt syscall_arg_fmts__by_name[] = {
1681         { .name = "msr",        .scnprintf = SCA_X86_MSR,         .strtoul = STUL_X86_MSR,         },
1682         { .name = "vector",     .scnprintf = SCA_X86_IRQ_VECTORS, .strtoul = STUL_X86_IRQ_VECTORS, },
1683 };
1684
1685 static int syscall_arg_fmt__cmp(const void *name, const void *fmtp)
1686 {
1687        const struct syscall_arg_fmt *fmt = fmtp;
1688        return strcmp(name, fmt->name);
1689 }
1690
1691 static const struct syscall_arg_fmt *
1692 __syscall_arg_fmt__find_by_name(const struct syscall_arg_fmt *fmts, const int nmemb,
1693                                 const char *name)
1694 {
1695        return bsearch(name, fmts, nmemb, sizeof(struct syscall_arg_fmt), syscall_arg_fmt__cmp);
1696 }
1697
1698 static const struct syscall_arg_fmt *syscall_arg_fmt__find_by_name(const char *name)
1699 {
1700        const int nmemb = ARRAY_SIZE(syscall_arg_fmts__by_name);
1701        return __syscall_arg_fmt__find_by_name(syscall_arg_fmts__by_name, nmemb, name);
1702 }
1703
1704 static struct tep_format_field *
1705 syscall_arg_fmt__init_array(struct syscall_arg_fmt *arg, struct tep_format_field *field)
1706 {
1707         struct tep_format_field *last_field = NULL;
1708         int len;
1709
1710         for (; field; field = field->next, ++arg) {
1711                 last_field = field;
1712
1713                 if (arg->scnprintf)
1714                         continue;
1715
1716                 len = strlen(field->name);
1717
1718                 if (strcmp(field->type, "const char *") == 0 &&
1719                     ((len >= 4 && strcmp(field->name + len - 4, "name") == 0) ||
1720                      strstr(field->name, "path") != NULL))
1721                         arg->scnprintf = SCA_FILENAME;
1722                 else if ((field->flags & TEP_FIELD_IS_POINTER) || strstr(field->name, "addr"))
1723                         arg->scnprintf = SCA_PTR;
1724                 else if (strcmp(field->type, "pid_t") == 0)
1725                         arg->scnprintf = SCA_PID;
1726                 else if (strcmp(field->type, "umode_t") == 0)
1727                         arg->scnprintf = SCA_MODE_T;
1728                 else if ((field->flags & TEP_FIELD_IS_ARRAY) && strstr(field->type, "char")) {
1729                         arg->scnprintf = SCA_CHAR_ARRAY;
1730                         arg->nr_entries = field->arraylen;
1731                 } else if ((strcmp(field->type, "int") == 0 ||
1732                           strcmp(field->type, "unsigned int") == 0 ||
1733                           strcmp(field->type, "long") == 0) &&
1734                          len >= 2 && strcmp(field->name + len - 2, "fd") == 0) {
1735                         /*
1736                          * /sys/kernel/tracing/events/syscalls/sys_enter*
1737                          * grep -E 'field:.*fd;' .../format|sed -r 's/.*field:([a-z ]+) [a-z_]*fd.+/\1/g'|sort|uniq -c
1738                          * 65 int
1739                          * 23 unsigned int
1740                          * 7 unsigned long
1741                          */
1742                         arg->scnprintf = SCA_FD;
1743                 } else {
1744                         const struct syscall_arg_fmt *fmt =
1745                                 syscall_arg_fmt__find_by_name(field->name);
1746
1747                         if (fmt) {
1748                                 arg->scnprintf = fmt->scnprintf;
1749                                 arg->strtoul   = fmt->strtoul;
1750                         }
1751                 }
1752         }
1753
1754         return last_field;
1755 }
1756
1757 static int syscall__set_arg_fmts(struct syscall *sc)
1758 {
1759         struct tep_format_field *last_field = syscall_arg_fmt__init_array(sc->arg_fmt, sc->args);
1760
1761         if (last_field)
1762                 sc->args_size = last_field->offset + last_field->size;
1763
1764         return 0;
1765 }
1766
1767 static int trace__read_syscall_info(struct trace *trace, int id)
1768 {
1769         char tp_name[128];
1770         struct syscall *sc;
1771         const char *name = syscalltbl__name(trace->sctbl, id);
1772
1773 #ifdef HAVE_SYSCALL_TABLE_SUPPORT
1774         if (trace->syscalls.table == NULL) {
1775                 trace->syscalls.table = calloc(trace->sctbl->syscalls.max_id + 1, sizeof(*sc));
1776                 if (trace->syscalls.table == NULL)
1777                         return -ENOMEM;
1778         }
1779 #else
1780         if (id > trace->sctbl->syscalls.max_id || (id == 0 && trace->syscalls.table == NULL)) {
1781                 // When using libaudit we don't know beforehand what is the max syscall id
1782                 struct syscall *table = realloc(trace->syscalls.table, (id + 1) * sizeof(*sc));
1783
1784                 if (table == NULL)
1785                         return -ENOMEM;
1786
1787                 // Need to memset from offset 0 and +1 members if brand new
1788                 if (trace->syscalls.table == NULL)
1789                         memset(table, 0, (id + 1) * sizeof(*sc));
1790                 else
1791                         memset(table + trace->sctbl->syscalls.max_id + 1, 0, (id - trace->sctbl->syscalls.max_id) * sizeof(*sc));
1792
1793                 trace->syscalls.table         = table;
1794                 trace->sctbl->syscalls.max_id = id;
1795         }
1796 #endif
1797         sc = trace->syscalls.table + id;
1798         if (sc->nonexistent)
1799                 return -EEXIST;
1800
1801         if (name == NULL) {
1802                 sc->nonexistent = true;
1803                 return -EEXIST;
1804         }
1805
1806         sc->name = name;
1807         sc->fmt  = syscall_fmt__find(sc->name);
1808
1809         snprintf(tp_name, sizeof(tp_name), "sys_enter_%s", sc->name);
1810         sc->tp_format = trace_event__tp_format("syscalls", tp_name);
1811
1812         if (IS_ERR(sc->tp_format) && sc->fmt && sc->fmt->alias) {
1813                 snprintf(tp_name, sizeof(tp_name), "sys_enter_%s", sc->fmt->alias);
1814                 sc->tp_format = trace_event__tp_format("syscalls", tp_name);
1815         }
1816
1817         /*
1818          * Fails to read trace point format via sysfs node, so the trace point
1819          * doesn't exist.  Set the 'nonexistent' flag as true.
1820          */
1821         if (IS_ERR(sc->tp_format)) {
1822                 sc->nonexistent = true;
1823                 return PTR_ERR(sc->tp_format);
1824         }
1825
1826         if (syscall__alloc_arg_fmts(sc, IS_ERR(sc->tp_format) ?
1827                                         RAW_SYSCALL_ARGS_NUM : sc->tp_format->format.nr_fields))
1828                 return -ENOMEM;
1829
1830         sc->args = sc->tp_format->format.fields;
1831         /*
1832          * We need to check and discard the first variable '__syscall_nr'
1833          * or 'nr' that mean the syscall number. It is needless here.
1834          * So drop '__syscall_nr' or 'nr' field but does not exist on older kernels.
1835          */
1836         if (sc->args && (!strcmp(sc->args->name, "__syscall_nr") || !strcmp(sc->args->name, "nr"))) {
1837                 sc->args = sc->args->next;
1838                 --sc->nr_args;
1839         }
1840
1841         sc->is_exit = !strcmp(name, "exit_group") || !strcmp(name, "exit");
1842         sc->is_open = !strcmp(name, "open") || !strcmp(name, "openat");
1843
1844         return syscall__set_arg_fmts(sc);
1845 }
1846
1847 static int evsel__init_tp_arg_scnprintf(struct evsel *evsel)
1848 {
1849         struct syscall_arg_fmt *fmt = evsel__syscall_arg_fmt(evsel);
1850
1851         if (fmt != NULL) {
1852                 syscall_arg_fmt__init_array(fmt, evsel->tp_format->format.fields);
1853                 return 0;
1854         }
1855
1856         return -ENOMEM;
1857 }
1858
1859 static int intcmp(const void *a, const void *b)
1860 {
1861         const int *one = a, *another = b;
1862
1863         return *one - *another;
1864 }
1865
1866 static int trace__validate_ev_qualifier(struct trace *trace)
1867 {
1868         int err = 0;
1869         bool printed_invalid_prefix = false;
1870         struct str_node *pos;
1871         size_t nr_used = 0, nr_allocated = strlist__nr_entries(trace->ev_qualifier);
1872
1873         trace->ev_qualifier_ids.entries = malloc(nr_allocated *
1874                                                  sizeof(trace->ev_qualifier_ids.entries[0]));
1875
1876         if (trace->ev_qualifier_ids.entries == NULL) {
1877                 fputs("Error:\tNot enough memory for allocating events qualifier ids\n",
1878                        trace->output);
1879                 err = -EINVAL;
1880                 goto out;
1881         }
1882
1883         strlist__for_each_entry(pos, trace->ev_qualifier) {
1884                 const char *sc = pos->s;
1885                 int id = syscalltbl__id(trace->sctbl, sc), match_next = -1;
1886
1887                 if (id < 0) {
1888                         id = syscalltbl__strglobmatch_first(trace->sctbl, sc, &match_next);
1889                         if (id >= 0)
1890                                 goto matches;
1891
1892                         if (!printed_invalid_prefix) {
1893                                 pr_debug("Skipping unknown syscalls: ");
1894                                 printed_invalid_prefix = true;
1895                         } else {
1896                                 pr_debug(", ");
1897                         }
1898
1899                         pr_debug("%s", sc);
1900                         continue;
1901                 }
1902 matches:
1903                 trace->ev_qualifier_ids.entries[nr_used++] = id;
1904                 if (match_next == -1)
1905                         continue;
1906
1907                 while (1) {
1908                         id = syscalltbl__strglobmatch_next(trace->sctbl, sc, &match_next);
1909                         if (id < 0)
1910                                 break;
1911                         if (nr_allocated == nr_used) {
1912                                 void *entries;
1913
1914                                 nr_allocated += 8;
1915                                 entries = realloc(trace->ev_qualifier_ids.entries,
1916                                                   nr_allocated * sizeof(trace->ev_qualifier_ids.entries[0]));
1917                                 if (entries == NULL) {
1918                                         err = -ENOMEM;
1919                                         fputs("\nError:\t Not enough memory for parsing\n", trace->output);
1920                                         goto out_free;
1921                                 }
1922                                 trace->ev_qualifier_ids.entries = entries;
1923                         }
1924                         trace->ev_qualifier_ids.entries[nr_used++] = id;
1925                 }
1926         }
1927
1928         trace->ev_qualifier_ids.nr = nr_used;
1929         qsort(trace->ev_qualifier_ids.entries, nr_used, sizeof(int), intcmp);
1930 out:
1931         if (printed_invalid_prefix)
1932                 pr_debug("\n");
1933         return err;
1934 out_free:
1935         zfree(&trace->ev_qualifier_ids.entries);
1936         trace->ev_qualifier_ids.nr = 0;
1937         goto out;
1938 }
1939
1940 static __maybe_unused bool trace__syscall_enabled(struct trace *trace, int id)
1941 {
1942         bool in_ev_qualifier;
1943
1944         if (trace->ev_qualifier_ids.nr == 0)
1945                 return true;
1946
1947         in_ev_qualifier = bsearch(&id, trace->ev_qualifier_ids.entries,
1948                                   trace->ev_qualifier_ids.nr, sizeof(int), intcmp) != NULL;
1949
1950         if (in_ev_qualifier)
1951                return !trace->not_ev_qualifier;
1952
1953         return trace->not_ev_qualifier;
1954 }
1955
1956 /*
1957  * args is to be interpreted as a series of longs but we need to handle
1958  * 8-byte unaligned accesses. args points to raw_data within the event
1959  * and raw_data is guaranteed to be 8-byte unaligned because it is
1960  * preceded by raw_size which is a u32. So we need to copy args to a temp
1961  * variable to read it. Most notably this avoids extended load instructions
1962  * on unaligned addresses
1963  */
1964 unsigned long syscall_arg__val(struct syscall_arg *arg, u8 idx)
1965 {
1966         unsigned long val;
1967         unsigned char *p = arg->args + sizeof(unsigned long) * idx;
1968
1969         memcpy(&val, p, sizeof(val));
1970         return val;
1971 }
1972
1973 static size_t syscall__scnprintf_name(struct syscall *sc, char *bf, size_t size,
1974                                       struct syscall_arg *arg)
1975 {
1976         if (sc->arg_fmt && sc->arg_fmt[arg->idx].name)
1977                 return scnprintf(bf, size, "%s: ", sc->arg_fmt[arg->idx].name);
1978
1979         return scnprintf(bf, size, "arg%d: ", arg->idx);
1980 }
1981
1982 /*
1983  * Check if the value is in fact zero, i.e. mask whatever needs masking, such
1984  * as mount 'flags' argument that needs ignoring some magic flag, see comment
1985  * in tools/perf/trace/beauty/mount_flags.c
1986  */
1987 static unsigned long syscall_arg_fmt__mask_val(struct syscall_arg_fmt *fmt, struct syscall_arg *arg, unsigned long val)
1988 {
1989         if (fmt && fmt->mask_val)
1990                 return fmt->mask_val(arg, val);
1991
1992         return val;
1993 }
1994
1995 static size_t syscall_arg_fmt__scnprintf_val(struct syscall_arg_fmt *fmt, char *bf, size_t size,
1996                                              struct syscall_arg *arg, unsigned long val)
1997 {
1998         if (fmt && fmt->scnprintf) {
1999                 arg->val = val;
2000                 if (fmt->parm)
2001                         arg->parm = fmt->parm;
2002                 return fmt->scnprintf(bf, size, arg);
2003         }
2004         return scnprintf(bf, size, "%ld", val);
2005 }
2006
2007 static size_t syscall__scnprintf_args(struct syscall *sc, char *bf, size_t size,
2008                                       unsigned char *args, void *augmented_args, int augmented_args_size,
2009                                       struct trace *trace, struct thread *thread)
2010 {
2011         size_t printed = 0;
2012         unsigned long val;
2013         u8 bit = 1;
2014         struct syscall_arg arg = {
2015                 .args   = args,
2016                 .augmented = {
2017                         .size = augmented_args_size,
2018                         .args = augmented_args,
2019                 },
2020                 .idx    = 0,
2021                 .mask   = 0,
2022                 .trace  = trace,
2023                 .thread = thread,
2024                 .show_string_prefix = trace->show_string_prefix,
2025         };
2026         struct thread_trace *ttrace = thread__priv(thread);
2027
2028         /*
2029          * Things like fcntl will set this in its 'cmd' formatter to pick the
2030          * right formatter for the return value (an fd? file flags?), which is
2031          * not needed for syscalls that always return a given type, say an fd.
2032          */
2033         ttrace->ret_scnprintf = NULL;
2034
2035         if (sc->args != NULL) {
2036                 struct tep_format_field *field;
2037
2038                 for (field = sc->args; field;
2039                      field = field->next, ++arg.idx, bit <<= 1) {
2040                         if (arg.mask & bit)
2041                                 continue;
2042
2043                         arg.fmt = &sc->arg_fmt[arg.idx];
2044                         val = syscall_arg__val(&arg, arg.idx);
2045                         /*
2046                          * Some syscall args need some mask, most don't and
2047                          * return val untouched.
2048                          */
2049                         val = syscall_arg_fmt__mask_val(&sc->arg_fmt[arg.idx], &arg, val);
2050
2051                         /*
2052                          * Suppress this argument if its value is zero and
2053                          * and we don't have a string associated in an
2054                          * strarray for it.
2055                          */
2056                         if (val == 0 &&
2057                             !trace->show_zeros &&
2058                             !(sc->arg_fmt &&
2059                               (sc->arg_fmt[arg.idx].show_zero ||
2060                                sc->arg_fmt[arg.idx].scnprintf == SCA_STRARRAY ||
2061                                sc->arg_fmt[arg.idx].scnprintf == SCA_STRARRAYS) &&
2062                               sc->arg_fmt[arg.idx].parm))
2063                                 continue;
2064
2065                         printed += scnprintf(bf + printed, size - printed, "%s", printed ? ", " : "");
2066
2067                         if (trace->show_arg_names)
2068                                 printed += scnprintf(bf + printed, size - printed, "%s: ", field->name);
2069
2070                         printed += syscall_arg_fmt__scnprintf_val(&sc->arg_fmt[arg.idx],
2071                                                                   bf + printed, size - printed, &arg, val);
2072                 }
2073         } else if (IS_ERR(sc->tp_format)) {
2074                 /*
2075                  * If we managed to read the tracepoint /format file, then we
2076                  * may end up not having any args, like with gettid(), so only
2077                  * print the raw args when we didn't manage to read it.
2078                  */
2079                 while (arg.idx < sc->nr_args) {
2080                         if (arg.mask & bit)
2081                                 goto next_arg;
2082                         val = syscall_arg__val(&arg, arg.idx);
2083                         if (printed)
2084                                 printed += scnprintf(bf + printed, size - printed, ", ");
2085                         printed += syscall__scnprintf_name(sc, bf + printed, size - printed, &arg);
2086                         printed += syscall_arg_fmt__scnprintf_val(&sc->arg_fmt[arg.idx], bf + printed, size - printed, &arg, val);
2087 next_arg:
2088                         ++arg.idx;
2089                         bit <<= 1;
2090                 }
2091         }
2092
2093         return printed;
2094 }
2095
2096 typedef int (*tracepoint_handler)(struct trace *trace, struct evsel *evsel,
2097                                   union perf_event *event,
2098                                   struct perf_sample *sample);
2099
2100 static struct syscall *trace__syscall_info(struct trace *trace,
2101                                            struct evsel *evsel, int id)
2102 {
2103         int err = 0;
2104
2105         if (id < 0) {
2106
2107                 /*
2108                  * XXX: Noticed on x86_64, reproduced as far back as 3.0.36, haven't tried
2109                  * before that, leaving at a higher verbosity level till that is
2110                  * explained. Reproduced with plain ftrace with:
2111                  *
2112                  * echo 1 > /t/events/raw_syscalls/sys_exit/enable
2113                  * grep "NR -1 " /t/trace_pipe
2114                  *
2115                  * After generating some load on the machine.
2116                  */
2117                 if (verbose > 1) {
2118                         static u64 n;
2119                         fprintf(trace->output, "Invalid syscall %d id, skipping (%s, %" PRIu64 ") ...\n",
2120                                 id, evsel__name(evsel), ++n);
2121                 }
2122                 return NULL;
2123         }
2124
2125         err = -EINVAL;
2126
2127 #ifdef HAVE_SYSCALL_TABLE_SUPPORT
2128         if (id > trace->sctbl->syscalls.max_id) {
2129 #else
2130         if (id >= trace->sctbl->syscalls.max_id) {
2131                 /*
2132                  * With libaudit we don't know beforehand what is the max_id,
2133                  * so we let trace__read_syscall_info() figure that out as we
2134                  * go on reading syscalls.
2135                  */
2136                 err = trace__read_syscall_info(trace, id);
2137                 if (err)
2138 #endif
2139                 goto out_cant_read;
2140         }
2141
2142         if ((trace->syscalls.table == NULL || trace->syscalls.table[id].name == NULL) &&
2143             (err = trace__read_syscall_info(trace, id)) != 0)
2144                 goto out_cant_read;
2145
2146         if (trace->syscalls.table && trace->syscalls.table[id].nonexistent)
2147                 goto out_cant_read;
2148
2149         return &trace->syscalls.table[id];
2150
2151 out_cant_read:
2152         if (verbose > 0) {
2153                 char sbuf[STRERR_BUFSIZE];
2154                 fprintf(trace->output, "Problems reading syscall %d: %d (%s)", id, -err, str_error_r(-err, sbuf, sizeof(sbuf)));
2155                 if (id <= trace->sctbl->syscalls.max_id && trace->syscalls.table[id].name != NULL)
2156                         fprintf(trace->output, "(%s)", trace->syscalls.table[id].name);
2157                 fputs(" information\n", trace->output);
2158         }
2159         return NULL;
2160 }
2161
2162 struct syscall_stats {
2163         struct stats stats;
2164         u64          nr_failures;
2165         int          max_errno;
2166         u32          *errnos;
2167 };
2168
2169 static void thread__update_stats(struct thread *thread, struct thread_trace *ttrace,
2170                                  int id, struct perf_sample *sample, long err, bool errno_summary)
2171 {
2172         struct int_node *inode;
2173         struct syscall_stats *stats;
2174         u64 duration = 0;
2175
2176         inode = intlist__findnew(ttrace->syscall_stats, id);
2177         if (inode == NULL)
2178                 return;
2179
2180         stats = inode->priv;
2181         if (stats == NULL) {
2182                 stats = zalloc(sizeof(*stats));
2183                 if (stats == NULL)
2184                         return;
2185
2186                 init_stats(&stats->stats);
2187                 inode->priv = stats;
2188         }
2189
2190         if (ttrace->entry_time && sample->time > ttrace->entry_time)
2191                 duration = sample->time - ttrace->entry_time;
2192
2193         update_stats(&stats->stats, duration);
2194
2195         if (err < 0) {
2196                 ++stats->nr_failures;
2197
2198                 if (!errno_summary)
2199                         return;
2200
2201                 err = -err;
2202                 if (err > stats->max_errno) {
2203                         u32 *new_errnos = realloc(stats->errnos, err * sizeof(u32));
2204
2205                         if (new_errnos) {
2206                                 memset(new_errnos + stats->max_errno, 0, (err - stats->max_errno) * sizeof(u32));
2207                         } else {
2208                                 pr_debug("Not enough memory for errno stats for thread \"%s\"(%d/%d), results will be incomplete\n",
2209                                          thread__comm_str(thread), thread__pid(thread),
2210                                          thread__tid(thread));
2211                                 return;
2212                         }
2213
2214                         stats->errnos = new_errnos;
2215                         stats->max_errno = err;
2216                 }
2217
2218                 ++stats->errnos[err - 1];
2219         }
2220 }
2221
2222 static int trace__printf_interrupted_entry(struct trace *trace)
2223 {
2224         struct thread_trace *ttrace;
2225         size_t printed;
2226         int len;
2227
2228         if (trace->failure_only || trace->current == NULL)
2229                 return 0;
2230
2231         ttrace = thread__priv(trace->current);
2232
2233         if (!ttrace->entry_pending)
2234                 return 0;
2235
2236         printed  = trace__fprintf_entry_head(trace, trace->current, 0, false, ttrace->entry_time, trace->output);
2237         printed += len = fprintf(trace->output, "%s)", ttrace->entry_str);
2238
2239         if (len < trace->args_alignment - 4)
2240                 printed += fprintf(trace->output, "%-*s", trace->args_alignment - 4 - len, " ");
2241
2242         printed += fprintf(trace->output, " ...\n");
2243
2244         ttrace->entry_pending = false;
2245         ++trace->nr_events_printed;
2246
2247         return printed;
2248 }
2249
2250 static int trace__fprintf_sample(struct trace *trace, struct evsel *evsel,
2251                                  struct perf_sample *sample, struct thread *thread)
2252 {
2253         int printed = 0;
2254
2255         if (trace->print_sample) {
2256                 double ts = (double)sample->time / NSEC_PER_MSEC;
2257
2258                 printed += fprintf(trace->output, "%22s %10.3f %s %d/%d [%d]\n",
2259                                    evsel__name(evsel), ts,
2260                                    thread__comm_str(thread),
2261                                    sample->pid, sample->tid, sample->cpu);
2262         }
2263
2264         return printed;
2265 }
2266
2267 static void *syscall__augmented_args(struct syscall *sc, struct perf_sample *sample, int *augmented_args_size, int raw_augmented_args_size)
2268 {
2269         void *augmented_args = NULL;
2270         /*
2271          * For now with BPF raw_augmented we hook into raw_syscalls:sys_enter
2272          * and there we get all 6 syscall args plus the tracepoint common fields
2273          * that gets calculated at the start and the syscall_nr (another long).
2274          * So we check if that is the case and if so don't look after the
2275          * sc->args_size but always after the full raw_syscalls:sys_enter payload,
2276          * which is fixed.
2277          *
2278          * We'll revisit this later to pass s->args_size to the BPF augmenter
2279          * (now tools/perf/examples/bpf/augmented_raw_syscalls.c, so that it
2280          * copies only what we need for each syscall, like what happens when we
2281          * use syscalls:sys_enter_NAME, so that we reduce the kernel/userspace
2282          * traffic to just what is needed for each syscall.
2283          */
2284         int args_size = raw_augmented_args_size ?: sc->args_size;
2285
2286         *augmented_args_size = sample->raw_size - args_size;
2287         if (*augmented_args_size > 0)
2288                 augmented_args = sample->raw_data + args_size;
2289
2290         return augmented_args;
2291 }
2292
2293 static void syscall__exit(struct syscall *sc)
2294 {
2295         if (!sc)
2296                 return;
2297
2298         zfree(&sc->arg_fmt);
2299 }
2300
2301 static int trace__sys_enter(struct trace *trace, struct evsel *evsel,
2302                             union perf_event *event __maybe_unused,
2303                             struct perf_sample *sample)
2304 {
2305         char *msg;
2306         void *args;
2307         int printed = 0;
2308         struct thread *thread;
2309         int id = perf_evsel__sc_tp_uint(evsel, id, sample), err = -1;
2310         int augmented_args_size = 0;
2311         void *augmented_args = NULL;
2312         struct syscall *sc = trace__syscall_info(trace, evsel, id);
2313         struct thread_trace *ttrace;
2314
2315         if (sc == NULL)
2316                 return -1;
2317
2318         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2319         ttrace = thread__trace(thread, trace->output);
2320         if (ttrace == NULL)
2321                 goto out_put;
2322
2323         trace__fprintf_sample(trace, evsel, sample, thread);
2324
2325         args = perf_evsel__sc_tp_ptr(evsel, args, sample);
2326
2327         if (ttrace->entry_str == NULL) {
2328                 ttrace->entry_str = malloc(trace__entry_str_size);
2329                 if (!ttrace->entry_str)
2330                         goto out_put;
2331         }
2332
2333         if (!(trace->duration_filter || trace->summary_only || trace->min_stack))
2334                 trace__printf_interrupted_entry(trace);
2335         /*
2336          * If this is raw_syscalls.sys_enter, then it always comes with the 6 possible
2337          * arguments, even if the syscall being handled, say "openat", uses only 4 arguments
2338          * this breaks syscall__augmented_args() check for augmented args, as we calculate
2339          * syscall->args_size using each syscalls:sys_enter_NAME tracefs format file,
2340          * so when handling, say the openat syscall, we end up getting 6 args for the
2341          * raw_syscalls:sys_enter event, when we expected just 4, we end up mistakenly
2342          * thinking that the extra 2 u64 args are the augmented filename, so just check
2343          * here and avoid using augmented syscalls when the evsel is the raw_syscalls one.
2344          */
2345         if (evsel != trace->syscalls.events.sys_enter)
2346                 augmented_args = syscall__augmented_args(sc, sample, &augmented_args_size, trace->raw_augmented_syscalls_args_size);
2347         ttrace->entry_time = sample->time;
2348         msg = ttrace->entry_str;
2349         printed += scnprintf(msg + printed, trace__entry_str_size - printed, "%s(", sc->name);
2350
2351         printed += syscall__scnprintf_args(sc, msg + printed, trace__entry_str_size - printed,
2352                                            args, augmented_args, augmented_args_size, trace, thread);
2353
2354         if (sc->is_exit) {
2355                 if (!(trace->duration_filter || trace->summary_only || trace->failure_only || trace->min_stack)) {
2356                         int alignment = 0;
2357
2358                         trace__fprintf_entry_head(trace, thread, 0, false, ttrace->entry_time, trace->output);
2359                         printed = fprintf(trace->output, "%s)", ttrace->entry_str);
2360                         if (trace->args_alignment > printed)
2361                                 alignment = trace->args_alignment - printed;
2362                         fprintf(trace->output, "%*s= ?\n", alignment, " ");
2363                 }
2364         } else {
2365                 ttrace->entry_pending = true;
2366                 /* See trace__vfs_getname & trace__sys_exit */
2367                 ttrace->filename.pending_open = false;
2368         }
2369
2370         if (trace->current != thread) {
2371                 thread__put(trace->current);
2372                 trace->current = thread__get(thread);
2373         }
2374         err = 0;
2375 out_put:
2376         thread__put(thread);
2377         return err;
2378 }
2379
2380 static int trace__fprintf_sys_enter(struct trace *trace, struct evsel *evsel,
2381                                     struct perf_sample *sample)
2382 {
2383         struct thread_trace *ttrace;
2384         struct thread *thread;
2385         int id = perf_evsel__sc_tp_uint(evsel, id, sample), err = -1;
2386         struct syscall *sc = trace__syscall_info(trace, evsel, id);
2387         char msg[1024];
2388         void *args, *augmented_args = NULL;
2389         int augmented_args_size;
2390
2391         if (sc == NULL)
2392                 return -1;
2393
2394         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2395         ttrace = thread__trace(thread, trace->output);
2396         /*
2397          * We need to get ttrace just to make sure it is there when syscall__scnprintf_args()
2398          * and the rest of the beautifiers accessing it via struct syscall_arg touches it.
2399          */
2400         if (ttrace == NULL)
2401                 goto out_put;
2402
2403         args = perf_evsel__sc_tp_ptr(evsel, args, sample);
2404         augmented_args = syscall__augmented_args(sc, sample, &augmented_args_size, trace->raw_augmented_syscalls_args_size);
2405         syscall__scnprintf_args(sc, msg, sizeof(msg), args, augmented_args, augmented_args_size, trace, thread);
2406         fprintf(trace->output, "%s", msg);
2407         err = 0;
2408 out_put:
2409         thread__put(thread);
2410         return err;
2411 }
2412
2413 static int trace__resolve_callchain(struct trace *trace, struct evsel *evsel,
2414                                     struct perf_sample *sample,
2415                                     struct callchain_cursor *cursor)
2416 {
2417         struct addr_location al;
2418         int max_stack = evsel->core.attr.sample_max_stack ?
2419                         evsel->core.attr.sample_max_stack :
2420                         trace->max_stack;
2421         int err = -1;
2422
2423         addr_location__init(&al);
2424         if (machine__resolve(trace->host, &al, sample) < 0)
2425                 goto out;
2426
2427         err = thread__resolve_callchain(al.thread, cursor, evsel, sample, NULL, NULL, max_stack);
2428 out:
2429         addr_location__exit(&al);
2430         return err;
2431 }
2432
2433 static int trace__fprintf_callchain(struct trace *trace, struct perf_sample *sample)
2434 {
2435         /* TODO: user-configurable print_opts */
2436         const unsigned int print_opts = EVSEL__PRINT_SYM |
2437                                         EVSEL__PRINT_DSO |
2438                                         EVSEL__PRINT_UNKNOWN_AS_ADDR;
2439
2440         return sample__fprintf_callchain(sample, 38, print_opts, get_tls_callchain_cursor(), symbol_conf.bt_stop_list, trace->output);
2441 }
2442
2443 static const char *errno_to_name(struct evsel *evsel, int err)
2444 {
2445         struct perf_env *env = evsel__env(evsel);
2446         const char *arch_name = perf_env__arch(env);
2447
2448         return arch_syscalls__strerrno(arch_name, err);
2449 }
2450
2451 static int trace__sys_exit(struct trace *trace, struct evsel *evsel,
2452                            union perf_event *event __maybe_unused,
2453                            struct perf_sample *sample)
2454 {
2455         long ret;
2456         u64 duration = 0;
2457         bool duration_calculated = false;
2458         struct thread *thread;
2459         int id = perf_evsel__sc_tp_uint(evsel, id, sample), err = -1, callchain_ret = 0, printed = 0;
2460         int alignment = trace->args_alignment;
2461         struct syscall *sc = trace__syscall_info(trace, evsel, id);
2462         struct thread_trace *ttrace;
2463
2464         if (sc == NULL)
2465                 return -1;
2466
2467         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2468         ttrace = thread__trace(thread, trace->output);
2469         if (ttrace == NULL)
2470                 goto out_put;
2471
2472         trace__fprintf_sample(trace, evsel, sample, thread);
2473
2474         ret = perf_evsel__sc_tp_uint(evsel, ret, sample);
2475
2476         if (trace->summary)
2477                 thread__update_stats(thread, ttrace, id, sample, ret, trace->errno_summary);
2478
2479         if (!trace->fd_path_disabled && sc->is_open && ret >= 0 && ttrace->filename.pending_open) {
2480                 trace__set_fd_pathname(thread, ret, ttrace->filename.name);
2481                 ttrace->filename.pending_open = false;
2482                 ++trace->stats.vfs_getname;
2483         }
2484
2485         if (ttrace->entry_time) {
2486                 duration = sample->time - ttrace->entry_time;
2487                 if (trace__filter_duration(trace, duration))
2488                         goto out;
2489                 duration_calculated = true;
2490         } else if (trace->duration_filter)
2491                 goto out;
2492
2493         if (sample->callchain) {
2494                 struct callchain_cursor *cursor = get_tls_callchain_cursor();
2495
2496                 callchain_ret = trace__resolve_callchain(trace, evsel, sample, cursor);
2497                 if (callchain_ret == 0) {
2498                         if (cursor->nr < trace->min_stack)
2499                                 goto out;
2500                         callchain_ret = 1;
2501                 }
2502         }
2503
2504         if (trace->summary_only || (ret >= 0 && trace->failure_only))
2505                 goto out;
2506
2507         trace__fprintf_entry_head(trace, thread, duration, duration_calculated, ttrace->entry_time, trace->output);
2508
2509         if (ttrace->entry_pending) {
2510                 printed = fprintf(trace->output, "%s", ttrace->entry_str);
2511         } else {
2512                 printed += fprintf(trace->output, " ... [");
2513                 color_fprintf(trace->output, PERF_COLOR_YELLOW, "continued");
2514                 printed += 9;
2515                 printed += fprintf(trace->output, "]: %s()", sc->name);
2516         }
2517
2518         printed++; /* the closing ')' */
2519
2520         if (alignment > printed)
2521                 alignment -= printed;
2522         else
2523                 alignment = 0;
2524
2525         fprintf(trace->output, ")%*s= ", alignment, " ");
2526
2527         if (sc->fmt == NULL) {
2528                 if (ret < 0)
2529                         goto errno_print;
2530 signed_print:
2531                 fprintf(trace->output, "%ld", ret);
2532         } else if (ret < 0) {
2533 errno_print: {
2534                 char bf[STRERR_BUFSIZE];
2535                 const char *emsg = str_error_r(-ret, bf, sizeof(bf)),
2536                            *e = errno_to_name(evsel, -ret);
2537
2538                 fprintf(trace->output, "-1 %s (%s)", e, emsg);
2539         }
2540         } else if (ret == 0 && sc->fmt->timeout)
2541                 fprintf(trace->output, "0 (Timeout)");
2542         else if (ttrace->ret_scnprintf) {
2543                 char bf[1024];
2544                 struct syscall_arg arg = {
2545                         .val    = ret,
2546                         .thread = thread,
2547                         .trace  = trace,
2548                 };
2549                 ttrace->ret_scnprintf(bf, sizeof(bf), &arg);
2550                 ttrace->ret_scnprintf = NULL;
2551                 fprintf(trace->output, "%s", bf);
2552         } else if (sc->fmt->hexret)
2553                 fprintf(trace->output, "%#lx", ret);
2554         else if (sc->fmt->errpid) {
2555                 struct thread *child = machine__find_thread(trace->host, ret, ret);
2556
2557                 if (child != NULL) {
2558                         fprintf(trace->output, "%ld", ret);
2559                         if (thread__comm_set(child))
2560                                 fprintf(trace->output, " (%s)", thread__comm_str(child));
2561                         thread__put(child);
2562                 }
2563         } else
2564                 goto signed_print;
2565
2566         fputc('\n', trace->output);
2567
2568         /*
2569          * We only consider an 'event' for the sake of --max-events a non-filtered
2570          * sys_enter + sys_exit and other tracepoint events.
2571          */
2572         if (++trace->nr_events_printed == trace->max_events && trace->max_events != ULONG_MAX)
2573                 interrupted = true;
2574
2575         if (callchain_ret > 0)
2576                 trace__fprintf_callchain(trace, sample);
2577         else if (callchain_ret < 0)
2578                 pr_err("Problem processing %s callchain, skipping...\n", evsel__name(evsel));
2579 out:
2580         ttrace->entry_pending = false;
2581         err = 0;
2582 out_put:
2583         thread__put(thread);
2584         return err;
2585 }
2586
2587 static int trace__vfs_getname(struct trace *trace, struct evsel *evsel,
2588                               union perf_event *event __maybe_unused,
2589                               struct perf_sample *sample)
2590 {
2591         struct thread *thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2592         struct thread_trace *ttrace;
2593         size_t filename_len, entry_str_len, to_move;
2594         ssize_t remaining_space;
2595         char *pos;
2596         const char *filename = evsel__rawptr(evsel, sample, "pathname");
2597
2598         if (!thread)
2599                 goto out;
2600
2601         ttrace = thread__priv(thread);
2602         if (!ttrace)
2603                 goto out_put;
2604
2605         filename_len = strlen(filename);
2606         if (filename_len == 0)
2607                 goto out_put;
2608
2609         if (ttrace->filename.namelen < filename_len) {
2610                 char *f = realloc(ttrace->filename.name, filename_len + 1);
2611
2612                 if (f == NULL)
2613                         goto out_put;
2614
2615                 ttrace->filename.namelen = filename_len;
2616                 ttrace->filename.name = f;
2617         }
2618
2619         strcpy(ttrace->filename.name, filename);
2620         ttrace->filename.pending_open = true;
2621
2622         if (!ttrace->filename.ptr)
2623                 goto out_put;
2624
2625         entry_str_len = strlen(ttrace->entry_str);
2626         remaining_space = trace__entry_str_size - entry_str_len - 1; /* \0 */
2627         if (remaining_space <= 0)
2628                 goto out_put;
2629
2630         if (filename_len > (size_t)remaining_space) {
2631                 filename += filename_len - remaining_space;
2632                 filename_len = remaining_space;
2633         }
2634
2635         to_move = entry_str_len - ttrace->filename.entry_str_pos + 1; /* \0 */
2636         pos = ttrace->entry_str + ttrace->filename.entry_str_pos;
2637         memmove(pos + filename_len, pos, to_move);
2638         memcpy(pos, filename, filename_len);
2639
2640         ttrace->filename.ptr = 0;
2641         ttrace->filename.entry_str_pos = 0;
2642 out_put:
2643         thread__put(thread);
2644 out:
2645         return 0;
2646 }
2647
2648 static int trace__sched_stat_runtime(struct trace *trace, struct evsel *evsel,
2649                                      union perf_event *event __maybe_unused,
2650                                      struct perf_sample *sample)
2651 {
2652         u64 runtime = evsel__intval(evsel, sample, "runtime");
2653         double runtime_ms = (double)runtime / NSEC_PER_MSEC;
2654         struct thread *thread = machine__findnew_thread(trace->host,
2655                                                         sample->pid,
2656                                                         sample->tid);
2657         struct thread_trace *ttrace = thread__trace(thread, trace->output);
2658
2659         if (ttrace == NULL)
2660                 goto out_dump;
2661
2662         ttrace->runtime_ms += runtime_ms;
2663         trace->runtime_ms += runtime_ms;
2664 out_put:
2665         thread__put(thread);
2666         return 0;
2667
2668 out_dump:
2669         fprintf(trace->output, "%s: comm=%s,pid=%u,runtime=%" PRIu64 ",vruntime=%" PRIu64 ")\n",
2670                evsel->name,
2671                evsel__strval(evsel, sample, "comm"),
2672                (pid_t)evsel__intval(evsel, sample, "pid"),
2673                runtime,
2674                evsel__intval(evsel, sample, "vruntime"));
2675         goto out_put;
2676 }
2677
2678 static int bpf_output__printer(enum binary_printer_ops op,
2679                                unsigned int val, void *extra __maybe_unused, FILE *fp)
2680 {
2681         unsigned char ch = (unsigned char)val;
2682
2683         switch (op) {
2684         case BINARY_PRINT_CHAR_DATA:
2685                 return fprintf(fp, "%c", isprint(ch) ? ch : '.');
2686         case BINARY_PRINT_DATA_BEGIN:
2687         case BINARY_PRINT_LINE_BEGIN:
2688         case BINARY_PRINT_ADDR:
2689         case BINARY_PRINT_NUM_DATA:
2690         case BINARY_PRINT_NUM_PAD:
2691         case BINARY_PRINT_SEP:
2692         case BINARY_PRINT_CHAR_PAD:
2693         case BINARY_PRINT_LINE_END:
2694         case BINARY_PRINT_DATA_END:
2695         default:
2696                 break;
2697         }
2698
2699         return 0;
2700 }
2701
2702 static void bpf_output__fprintf(struct trace *trace,
2703                                 struct perf_sample *sample)
2704 {
2705         binary__fprintf(sample->raw_data, sample->raw_size, 8,
2706                         bpf_output__printer, NULL, trace->output);
2707         ++trace->nr_events_printed;
2708 }
2709
2710 static size_t trace__fprintf_tp_fields(struct trace *trace, struct evsel *evsel, struct perf_sample *sample,
2711                                        struct thread *thread, void *augmented_args, int augmented_args_size)
2712 {
2713         char bf[2048];
2714         size_t size = sizeof(bf);
2715         struct tep_format_field *field = evsel->tp_format->format.fields;
2716         struct syscall_arg_fmt *arg = __evsel__syscall_arg_fmt(evsel);
2717         size_t printed = 0;
2718         unsigned long val;
2719         u8 bit = 1;
2720         struct syscall_arg syscall_arg = {
2721                 .augmented = {
2722                         .size = augmented_args_size,
2723                         .args = augmented_args,
2724                 },
2725                 .idx    = 0,
2726                 .mask   = 0,
2727                 .trace  = trace,
2728                 .thread = thread,
2729                 .show_string_prefix = trace->show_string_prefix,
2730         };
2731
2732         for (; field && arg; field = field->next, ++syscall_arg.idx, bit <<= 1, ++arg) {
2733                 if (syscall_arg.mask & bit)
2734                         continue;
2735
2736                 syscall_arg.len = 0;
2737                 syscall_arg.fmt = arg;
2738                 if (field->flags & TEP_FIELD_IS_ARRAY) {
2739                         int offset = field->offset;
2740
2741                         if (field->flags & TEP_FIELD_IS_DYNAMIC) {
2742                                 offset = format_field__intval(field, sample, evsel->needs_swap);
2743                                 syscall_arg.len = offset >> 16;
2744                                 offset &= 0xffff;
2745                                 if (tep_field_is_relative(field->flags))
2746                                         offset += field->offset + field->size;
2747                         }
2748
2749                         val = (uintptr_t)(sample->raw_data + offset);
2750                 } else
2751                         val = format_field__intval(field, sample, evsel->needs_swap);
2752                 /*
2753                  * Some syscall args need some mask, most don't and
2754                  * return val untouched.
2755                  */
2756                 val = syscall_arg_fmt__mask_val(arg, &syscall_arg, val);
2757
2758                 /*
2759                  * Suppress this argument if its value is zero and
2760                  * we don't have a string associated in an
2761                  * strarray for it.
2762                  */
2763                 if (val == 0 &&
2764                     !trace->show_zeros &&
2765                     !((arg->show_zero ||
2766                        arg->scnprintf == SCA_STRARRAY ||
2767                        arg->scnprintf == SCA_STRARRAYS) &&
2768                       arg->parm))
2769                         continue;
2770
2771                 printed += scnprintf(bf + printed, size - printed, "%s", printed ? ", " : "");
2772
2773                 if (trace->show_arg_names)
2774                         printed += scnprintf(bf + printed, size - printed, "%s: ", field->name);
2775
2776                 printed += syscall_arg_fmt__scnprintf_val(arg, bf + printed, size - printed, &syscall_arg, val);
2777         }
2778
2779         return printed + fprintf(trace->output, "%s", bf);
2780 }
2781
2782 static int trace__event_handler(struct trace *trace, struct evsel *evsel,
2783                                 union perf_event *event __maybe_unused,
2784                                 struct perf_sample *sample)
2785 {
2786         struct thread *thread;
2787         int callchain_ret = 0;
2788         /*
2789          * Check if we called perf_evsel__disable(evsel) due to, for instance,
2790          * this event's max_events having been hit and this is an entry coming
2791          * from the ring buffer that we should discard, since the max events
2792          * have already been considered/printed.
2793          */
2794         if (evsel->disabled)
2795                 return 0;
2796
2797         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2798
2799         if (sample->callchain) {
2800                 struct callchain_cursor *cursor = get_tls_callchain_cursor();
2801
2802                 callchain_ret = trace__resolve_callchain(trace, evsel, sample, cursor);
2803                 if (callchain_ret == 0) {
2804                         if (cursor->nr < trace->min_stack)
2805                                 goto out;
2806                         callchain_ret = 1;
2807                 }
2808         }
2809
2810         trace__printf_interrupted_entry(trace);
2811         trace__fprintf_tstamp(trace, sample->time, trace->output);
2812
2813         if (trace->trace_syscalls && trace->show_duration)
2814                 fprintf(trace->output, "(         ): ");
2815
2816         if (thread)
2817                 trace__fprintf_comm_tid(trace, thread, trace->output);
2818
2819         if (evsel == trace->syscalls.events.augmented) {
2820                 int id = perf_evsel__sc_tp_uint(evsel, id, sample);
2821                 struct syscall *sc = trace__syscall_info(trace, evsel, id);
2822
2823                 if (sc) {
2824                         fprintf(trace->output, "%s(", sc->name);
2825                         trace__fprintf_sys_enter(trace, evsel, sample);
2826                         fputc(')', trace->output);
2827                         goto newline;
2828                 }
2829
2830                 /*
2831                  * XXX: Not having the associated syscall info or not finding/adding
2832                  *      the thread should never happen, but if it does...
2833                  *      fall thru and print it as a bpf_output event.
2834                  */
2835         }
2836
2837         fprintf(trace->output, "%s(", evsel->name);
2838
2839         if (evsel__is_bpf_output(evsel)) {
2840                 bpf_output__fprintf(trace, sample);
2841         } else if (evsel->tp_format) {
2842                 if (strncmp(evsel->tp_format->name, "sys_enter_", 10) ||
2843                     trace__fprintf_sys_enter(trace, evsel, sample)) {
2844                         if (trace->libtraceevent_print) {
2845                                 event_format__fprintf(evsel->tp_format, sample->cpu,
2846                                                       sample->raw_data, sample->raw_size,
2847                                                       trace->output);
2848                         } else {
2849                                 trace__fprintf_tp_fields(trace, evsel, sample, thread, NULL, 0);
2850                         }
2851                 }
2852         }
2853
2854 newline:
2855         fprintf(trace->output, ")\n");
2856
2857         if (callchain_ret > 0)
2858                 trace__fprintf_callchain(trace, sample);
2859         else if (callchain_ret < 0)
2860                 pr_err("Problem processing %s callchain, skipping...\n", evsel__name(evsel));
2861
2862         ++trace->nr_events_printed;
2863
2864         if (evsel->max_events != ULONG_MAX && ++evsel->nr_events_printed == evsel->max_events) {
2865                 evsel__disable(evsel);
2866                 evsel__close(evsel);
2867         }
2868 out:
2869         thread__put(thread);
2870         return 0;
2871 }
2872
2873 static void print_location(FILE *f, struct perf_sample *sample,
2874                            struct addr_location *al,
2875                            bool print_dso, bool print_sym)
2876 {
2877
2878         if ((verbose > 0 || print_dso) && al->map)
2879                 fprintf(f, "%s@", map__dso(al->map)->long_name);
2880
2881         if ((verbose > 0 || print_sym) && al->sym)
2882                 fprintf(f, "%s+0x%" PRIx64, al->sym->name,
2883                         al->addr - al->sym->start);
2884         else if (al->map)
2885                 fprintf(f, "0x%" PRIx64, al->addr);
2886         else
2887                 fprintf(f, "0x%" PRIx64, sample->addr);
2888 }
2889
2890 static int trace__pgfault(struct trace *trace,
2891                           struct evsel *evsel,
2892                           union perf_event *event __maybe_unused,
2893                           struct perf_sample *sample)
2894 {
2895         struct thread *thread;
2896         struct addr_location al;
2897         char map_type = 'd';
2898         struct thread_trace *ttrace;
2899         int err = -1;
2900         int callchain_ret = 0;
2901
2902         addr_location__init(&al);
2903         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2904
2905         if (sample->callchain) {
2906                 struct callchain_cursor *cursor = get_tls_callchain_cursor();
2907
2908                 callchain_ret = trace__resolve_callchain(trace, evsel, sample, cursor);
2909                 if (callchain_ret == 0) {
2910                         if (cursor->nr < trace->min_stack)
2911                                 goto out_put;
2912                         callchain_ret = 1;
2913                 }
2914         }
2915
2916         ttrace = thread__trace(thread, trace->output);
2917         if (ttrace == NULL)
2918                 goto out_put;
2919
2920         if (evsel->core.attr.config == PERF_COUNT_SW_PAGE_FAULTS_MAJ)
2921                 ttrace->pfmaj++;
2922         else
2923                 ttrace->pfmin++;
2924
2925         if (trace->summary_only)
2926                 goto out;
2927
2928         thread__find_symbol(thread, sample->cpumode, sample->ip, &al);
2929
2930         trace__fprintf_entry_head(trace, thread, 0, true, sample->time, trace->output);
2931
2932         fprintf(trace->output, "%sfault [",
2933                 evsel->core.attr.config == PERF_COUNT_SW_PAGE_FAULTS_MAJ ?
2934                 "maj" : "min");
2935
2936         print_location(trace->output, sample, &al, false, true);
2937
2938         fprintf(trace->output, "] => ");
2939
2940         thread__find_symbol(thread, sample->cpumode, sample->addr, &al);
2941
2942         if (!al.map) {
2943                 thread__find_symbol(thread, sample->cpumode, sample->addr, &al);
2944
2945                 if (al.map)
2946                         map_type = 'x';
2947                 else
2948                         map_type = '?';
2949         }
2950
2951         print_location(trace->output, sample, &al, true, false);
2952
2953         fprintf(trace->output, " (%c%c)\n", map_type, al.level);
2954
2955         if (callchain_ret > 0)
2956                 trace__fprintf_callchain(trace, sample);
2957         else if (callchain_ret < 0)
2958                 pr_err("Problem processing %s callchain, skipping...\n", evsel__name(evsel));
2959
2960         ++trace->nr_events_printed;
2961 out:
2962         err = 0;
2963 out_put:
2964         thread__put(thread);
2965         addr_location__exit(&al);
2966         return err;
2967 }
2968
2969 static void trace__set_base_time(struct trace *trace,
2970                                  struct evsel *evsel,
2971                                  struct perf_sample *sample)
2972 {
2973         /*
2974          * BPF events were not setting PERF_SAMPLE_TIME, so be more robust
2975          * and don't use sample->time unconditionally, we may end up having
2976          * some other event in the future without PERF_SAMPLE_TIME for good
2977          * reason, i.e. we may not be interested in its timestamps, just in
2978          * it taking place, picking some piece of information when it
2979          * appears in our event stream (vfs_getname comes to mind).
2980          */
2981         if (trace->base_time == 0 && !trace->full_time &&
2982             (evsel->core.attr.sample_type & PERF_SAMPLE_TIME))
2983                 trace->base_time = sample->time;
2984 }
2985
2986 static int trace__process_sample(struct perf_tool *tool,
2987                                  union perf_event *event,
2988                                  struct perf_sample *sample,
2989                                  struct evsel *evsel,
2990                                  struct machine *machine __maybe_unused)
2991 {
2992         struct trace *trace = container_of(tool, struct trace, tool);
2993         struct thread *thread;
2994         int err = 0;
2995
2996         tracepoint_handler handler = evsel->handler;
2997
2998         thread = machine__findnew_thread(trace->host, sample->pid, sample->tid);
2999         if (thread && thread__is_filtered(thread))
3000                 goto out;
3001
3002         trace__set_base_time(trace, evsel, sample);
3003
3004         if (handler) {
3005                 ++trace->nr_events;
3006                 handler(trace, evsel, event, sample);
3007         }
3008 out:
3009         thread__put(thread);
3010         return err;
3011 }
3012
3013 static int trace__record(struct trace *trace, int argc, const char **argv)
3014 {
3015         unsigned int rec_argc, i, j;
3016         const char **rec_argv;
3017         const char * const record_args[] = {
3018                 "record",
3019                 "-R",
3020                 "-m", "1024",
3021                 "-c", "1",
3022         };
3023         pid_t pid = getpid();
3024         char *filter = asprintf__tp_filter_pids(1, &pid);
3025         const char * const sc_args[] = { "-e", };
3026         unsigned int sc_args_nr = ARRAY_SIZE(sc_args);
3027         const char * const majpf_args[] = { "-e", "major-faults" };
3028         unsigned int majpf_args_nr = ARRAY_SIZE(majpf_args);
3029         const char * const minpf_args[] = { "-e", "minor-faults" };
3030         unsigned int minpf_args_nr = ARRAY_SIZE(minpf_args);
3031         int err = -1;
3032
3033         /* +3 is for the event string below and the pid filter */
3034         rec_argc = ARRAY_SIZE(record_args) + sc_args_nr + 3 +
3035                 majpf_args_nr + minpf_args_nr + argc;
3036         rec_argv = calloc(rec_argc + 1, sizeof(char *));
3037
3038         if (rec_argv == NULL || filter == NULL)
3039                 goto out_free;
3040
3041         j = 0;
3042         for (i = 0; i < ARRAY_SIZE(record_args); i++)
3043                 rec_argv[j++] = record_args[i];
3044
3045         if (trace->trace_syscalls) {
3046                 for (i = 0; i < sc_args_nr; i++)
3047                         rec_argv[j++] = sc_args[i];
3048
3049                 /* event string may be different for older kernels - e.g., RHEL6 */
3050                 if (is_valid_tracepoint("raw_syscalls:sys_enter"))
3051                         rec_argv[j++] = "raw_syscalls:sys_enter,raw_syscalls:sys_exit";
3052                 else if (is_valid_tracepoint("syscalls:sys_enter"))
3053                         rec_argv[j++] = "syscalls:sys_enter,syscalls:sys_exit";
3054                 else {
3055                         pr_err("Neither raw_syscalls nor syscalls events exist.\n");
3056                         goto out_free;
3057                 }
3058         }
3059
3060         rec_argv[j++] = "--filter";
3061         rec_argv[j++] = filter;
3062
3063         if (trace->trace_pgfaults & TRACE_PFMAJ)
3064                 for (i = 0; i < majpf_args_nr; i++)
3065                         rec_argv[j++] = majpf_args[i];
3066
3067         if (trace->trace_pgfaults & TRACE_PFMIN)
3068                 for (i = 0; i < minpf_args_nr; i++)
3069                         rec_argv[j++] = minpf_args[i];
3070
3071         for (i = 0; i < (unsigned int)argc; i++)
3072                 rec_argv[j++] = argv[i];
3073
3074         err = cmd_record(j, rec_argv);
3075 out_free:
3076         free(filter);
3077         free(rec_argv);
3078         return err;
3079 }
3080
3081 static size_t trace__fprintf_thread_summary(struct trace *trace, FILE *fp);
3082
3083 static bool evlist__add_vfs_getname(struct evlist *evlist)
3084 {
3085         bool found = false;
3086         struct evsel *evsel, *tmp;
3087         struct parse_events_error err;
3088         int ret;
3089
3090         parse_events_error__init(&err);
3091         ret = parse_events(evlist, "probe:vfs_getname*", &err);
3092         parse_events_error__exit(&err);
3093         if (ret)
3094                 return false;
3095
3096         evlist__for_each_entry_safe(evlist, evsel, tmp) {
3097                 if (!strstarts(evsel__name(evsel), "probe:vfs_getname"))
3098                         continue;
3099
3100                 if (evsel__field(evsel, "pathname")) {
3101                         evsel->handler = trace__vfs_getname;
3102                         found = true;
3103                         continue;
3104                 }
3105
3106                 list_del_init(&evsel->core.node);
3107                 evsel->evlist = NULL;
3108                 evsel__delete(evsel);
3109         }
3110
3111         return found;
3112 }
3113
3114 static struct evsel *evsel__new_pgfault(u64 config)
3115 {
3116         struct evsel *evsel;
3117         struct perf_event_attr attr = {
3118                 .type = PERF_TYPE_SOFTWARE,
3119                 .mmap_data = 1,
3120         };
3121
3122         attr.config = config;
3123         attr.sample_period = 1;
3124
3125         event_attr_init(&attr);
3126
3127         evsel = evsel__new(&attr);
3128         if (evsel)
3129                 evsel->handler = trace__pgfault;
3130
3131         return evsel;
3132 }
3133
3134 static void evlist__free_syscall_tp_fields(struct evlist *evlist)
3135 {
3136         struct evsel *evsel;
3137
3138         evlist__for_each_entry(evlist, evsel) {
3139                 struct evsel_trace *et = evsel->priv;
3140
3141                 if (!et || !evsel->tp_format || strcmp(evsel->tp_format->system, "syscalls"))
3142                         continue;
3143
3144                 zfree(&et->fmt);
3145                 free(et);
3146         }
3147 }
3148
3149 static void trace__handle_event(struct trace *trace, union perf_event *event, struct perf_sample *sample)
3150 {
3151         const u32 type = event->header.type;
3152         struct evsel *evsel;
3153
3154         if (type != PERF_RECORD_SAMPLE) {
3155                 trace__process_event(trace, trace->host, event, sample);
3156                 return;
3157         }
3158
3159         evsel = evlist__id2evsel(trace->evlist, sample->id);
3160         if (evsel == NULL) {
3161                 fprintf(trace->output, "Unknown tp ID %" PRIu64 ", skipping...\n", sample->id);
3162                 return;
3163         }
3164
3165         if (evswitch__discard(&trace->evswitch, evsel))
3166                 return;
3167
3168         trace__set_base_time(trace, evsel, sample);
3169
3170         if (evsel->core.attr.type == PERF_TYPE_TRACEPOINT &&
3171             sample->raw_data == NULL) {
3172                 fprintf(trace->output, "%s sample with no payload for tid: %d, cpu %d, raw_size=%d, skipping...\n",
3173                        evsel__name(evsel), sample->tid,
3174                        sample->cpu, sample->raw_size);
3175         } else {
3176                 tracepoint_handler handler = evsel->handler;
3177                 handler(trace, evsel, event, sample);
3178         }
3179
3180         if (trace->nr_events_printed >= trace->max_events && trace->max_events != ULONG_MAX)
3181                 interrupted = true;
3182 }
3183
3184 static int trace__add_syscall_newtp(struct trace *trace)
3185 {
3186         int ret = -1;
3187         struct evlist *evlist = trace->evlist;
3188         struct evsel *sys_enter, *sys_exit;
3189
3190         sys_enter = perf_evsel__raw_syscall_newtp("sys_enter", trace__sys_enter);
3191         if (sys_enter == NULL)
3192                 goto out;
3193
3194         if (perf_evsel__init_sc_tp_ptr_field(sys_enter, args))
3195                 goto out_delete_sys_enter;
3196
3197         sys_exit = perf_evsel__raw_syscall_newtp("sys_exit", trace__sys_exit);
3198         if (sys_exit == NULL)
3199                 goto out_delete_sys_enter;
3200
3201         if (perf_evsel__init_sc_tp_uint_field(sys_exit, ret))
3202                 goto out_delete_sys_exit;
3203
3204         evsel__config_callchain(sys_enter, &trace->opts, &callchain_param);
3205         evsel__config_callchain(sys_exit, &trace->opts, &callchain_param);
3206
3207         evlist__add(evlist, sys_enter);
3208         evlist__add(evlist, sys_exit);
3209
3210         if (callchain_param.enabled && !trace->kernel_syscallchains) {
3211                 /*
3212                  * We're interested only in the user space callchain
3213                  * leading to the syscall, allow overriding that for
3214                  * debugging reasons using --kernel_syscall_callchains
3215                  */
3216                 sys_exit->core.attr.exclude_callchain_kernel = 1;
3217         }
3218
3219         trace->syscalls.events.sys_enter = sys_enter;
3220         trace->syscalls.events.sys_exit  = sys_exit;
3221
3222         ret = 0;
3223 out:
3224         return ret;
3225
3226 out_delete_sys_exit:
3227         evsel__delete_priv(sys_exit);
3228 out_delete_sys_enter:
3229         evsel__delete_priv(sys_enter);
3230         goto out;
3231 }
3232
3233 static int trace__set_ev_qualifier_tp_filter(struct trace *trace)
3234 {
3235         int err = -1;
3236         struct evsel *sys_exit;
3237         char *filter = asprintf_expr_inout_ints("id", !trace->not_ev_qualifier,
3238                                                 trace->ev_qualifier_ids.nr,
3239                                                 trace->ev_qualifier_ids.entries);
3240
3241         if (filter == NULL)
3242                 goto out_enomem;
3243
3244         if (!evsel__append_tp_filter(trace->syscalls.events.sys_enter, filter)) {
3245                 sys_exit = trace->syscalls.events.sys_exit;
3246                 err = evsel__append_tp_filter(sys_exit, filter);
3247         }
3248
3249         free(filter);
3250 out:
3251         return err;
3252 out_enomem:
3253         errno = ENOMEM;
3254         goto out;
3255 }
3256
3257 #ifdef HAVE_LIBBPF_SUPPORT
3258 static struct bpf_map *trace__find_bpf_map_by_name(struct trace *trace, const char *name)
3259 {
3260         if (trace->bpf_obj == NULL)
3261                 return NULL;
3262
3263         return bpf_object__find_map_by_name(trace->bpf_obj, name);
3264 }
3265
3266 static void trace__set_bpf_map_filtered_pids(struct trace *trace)
3267 {
3268         trace->filter_pids.map = trace__find_bpf_map_by_name(trace, "pids_filtered");
3269 }
3270
3271 static void trace__set_bpf_map_syscalls(struct trace *trace)
3272 {
3273         trace->syscalls.prog_array.sys_enter = trace__find_bpf_map_by_name(trace, "syscalls_sys_enter");
3274         trace->syscalls.prog_array.sys_exit  = trace__find_bpf_map_by_name(trace, "syscalls_sys_exit");
3275 }
3276
3277 static struct bpf_program *trace__find_bpf_program_by_title(struct trace *trace, const char *name)
3278 {
3279         struct bpf_program *pos, *prog = NULL;
3280         const char *sec_name;
3281
3282         if (trace->bpf_obj == NULL)
3283                 return NULL;
3284
3285         bpf_object__for_each_program(pos, trace->bpf_obj) {
3286                 sec_name = bpf_program__section_name(pos);
3287                 if (sec_name && !strcmp(sec_name, name)) {
3288                         prog = pos;
3289                         break;
3290                 }
3291         }
3292
3293         return prog;
3294 }
3295
3296 static struct bpf_program *trace__find_syscall_bpf_prog(struct trace *trace, struct syscall *sc,
3297                                                         const char *prog_name, const char *type)
3298 {
3299         struct bpf_program *prog;
3300
3301         if (prog_name == NULL) {
3302                 char default_prog_name[256];
3303                 scnprintf(default_prog_name, sizeof(default_prog_name), "!syscalls:sys_%s_%s", type, sc->name);
3304                 prog = trace__find_bpf_program_by_title(trace, default_prog_name);
3305                 if (prog != NULL)
3306                         goto out_found;
3307                 if (sc->fmt && sc->fmt->alias) {
3308                         scnprintf(default_prog_name, sizeof(default_prog_name), "!syscalls:sys_%s_%s", type, sc->fmt->alias);
3309                         prog = trace__find_bpf_program_by_title(trace, default_prog_name);
3310                         if (prog != NULL)
3311                                 goto out_found;
3312                 }
3313                 goto out_unaugmented;
3314         }
3315
3316         prog = trace__find_bpf_program_by_title(trace, prog_name);
3317
3318         if (prog != NULL) {
3319 out_found:
3320                 return prog;
3321         }
3322
3323         pr_debug("Couldn't find BPF prog \"%s\" to associate with syscalls:sys_%s_%s, not augmenting it\n",
3324                  prog_name, type, sc->name);
3325 out_unaugmented:
3326         return trace->syscalls.unaugmented_prog;
3327 }
3328
3329 static void trace__init_syscall_bpf_progs(struct trace *trace, int id)
3330 {
3331         struct syscall *sc = trace__syscall_info(trace, NULL, id);
3332
3333         if (sc == NULL)
3334                 return;
3335
3336         sc->bpf_prog.sys_enter = trace__find_syscall_bpf_prog(trace, sc, sc->fmt ? sc->fmt->bpf_prog_name.sys_enter : NULL, "enter");
3337         sc->bpf_prog.sys_exit  = trace__find_syscall_bpf_prog(trace, sc, sc->fmt ? sc->fmt->bpf_prog_name.sys_exit  : NULL,  "exit");
3338 }
3339
3340 static int trace__bpf_prog_sys_enter_fd(struct trace *trace, int id)
3341 {
3342         struct syscall *sc = trace__syscall_info(trace, NULL, id);
3343         return sc ? bpf_program__fd(sc->bpf_prog.sys_enter) : bpf_program__fd(trace->syscalls.unaugmented_prog);
3344 }
3345
3346 static int trace__bpf_prog_sys_exit_fd(struct trace *trace, int id)
3347 {
3348         struct syscall *sc = trace__syscall_info(trace, NULL, id);
3349         return sc ? bpf_program__fd(sc->bpf_prog.sys_exit) : bpf_program__fd(trace->syscalls.unaugmented_prog);
3350 }
3351
3352 static struct bpf_program *trace__find_usable_bpf_prog_entry(struct trace *trace, struct syscall *sc)
3353 {
3354         struct tep_format_field *field, *candidate_field;
3355         int id;
3356
3357         /*
3358          * We're only interested in syscalls that have a pointer:
3359          */
3360         for (field = sc->args; field; field = field->next) {
3361                 if (field->flags & TEP_FIELD_IS_POINTER)
3362                         goto try_to_find_pair;
3363         }
3364
3365         return NULL;
3366
3367 try_to_find_pair:
3368         for (id = 0; id < trace->sctbl->syscalls.nr_entries; ++id) {
3369                 struct syscall *pair = trace__syscall_info(trace, NULL, id);
3370                 struct bpf_program *pair_prog;
3371                 bool is_candidate = false;
3372
3373                 if (pair == NULL || pair == sc ||
3374                     pair->bpf_prog.sys_enter == trace->syscalls.unaugmented_prog)
3375                         continue;
3376
3377                 for (field = sc->args, candidate_field = pair->args;
3378                      field && candidate_field; field = field->next, candidate_field = candidate_field->next) {
3379                         bool is_pointer = field->flags & TEP_FIELD_IS_POINTER,
3380                              candidate_is_pointer = candidate_field->flags & TEP_FIELD_IS_POINTER;
3381
3382                         if (is_pointer) {
3383                                if (!candidate_is_pointer) {
3384                                         // The candidate just doesn't copies our pointer arg, might copy other pointers we want.
3385                                         continue;
3386                                }
3387                         } else {
3388                                 if (candidate_is_pointer) {
3389                                         // The candidate might copy a pointer we don't have, skip it.
3390                                         goto next_candidate;
3391                                 }
3392                                 continue;
3393                         }
3394
3395                         if (strcmp(field->type, candidate_field->type))
3396                                 goto next_candidate;
3397
3398                         is_candidate = true;
3399                 }
3400
3401                 if (!is_candidate)
3402                         goto next_candidate;
3403
3404                 /*
3405                  * Check if the tentative pair syscall augmenter has more pointers, if it has,
3406                  * then it may be collecting that and we then can't use it, as it would collect
3407                  * more than what is common to the two syscalls.
3408                  */
3409                 if (candidate_field) {
3410                         for (candidate_field = candidate_field->next; candidate_field; candidate_field = candidate_field->next)
3411                                 if (candidate_field->flags & TEP_FIELD_IS_POINTER)
3412                                         goto next_candidate;
3413                 }
3414
3415                 pair_prog = pair->bpf_prog.sys_enter;
3416                 /*
3417                  * If the pair isn't enabled, then its bpf_prog.sys_enter will not
3418                  * have been searched for, so search it here and if it returns the
3419                  * unaugmented one, then ignore it, otherwise we'll reuse that BPF
3420                  * program for a filtered syscall on a non-filtered one.
3421                  *
3422                  * For instance, we have "!syscalls:sys_enter_renameat" and that is
3423                  * useful for "renameat2".
3424                  */
3425                 if (pair_prog == NULL) {
3426                         pair_prog = trace__find_syscall_bpf_prog(trace, pair, pair->fmt ? pair->fmt->bpf_prog_name.sys_enter : NULL, "enter");
3427                         if (pair_prog == trace->syscalls.unaugmented_prog)
3428                                 goto next_candidate;
3429                 }
3430
3431                 pr_debug("Reusing \"%s\" BPF sys_enter augmenter for \"%s\"\n", pair->name, sc->name);
3432                 return pair_prog;
3433         next_candidate:
3434                 continue;
3435         }
3436
3437         return NULL;
3438 }
3439
3440 static int trace__init_syscalls_bpf_prog_array_maps(struct trace *trace)
3441 {
3442         int map_enter_fd = bpf_map__fd(trace->syscalls.prog_array.sys_enter),
3443             map_exit_fd  = bpf_map__fd(trace->syscalls.prog_array.sys_exit);
3444         int err = 0, key;
3445
3446         for (key = 0; key < trace->sctbl->syscalls.nr_entries; ++key) {
3447                 int prog_fd;
3448
3449                 if (!trace__syscall_enabled(trace, key))
3450                         continue;
3451
3452                 trace__init_syscall_bpf_progs(trace, key);
3453
3454                 // It'll get at least the "!raw_syscalls:unaugmented"
3455                 prog_fd = trace__bpf_prog_sys_enter_fd(trace, key);
3456                 err = bpf_map_update_elem(map_enter_fd, &key, &prog_fd, BPF_ANY);
3457                 if (err)
3458                         break;
3459                 prog_fd = trace__bpf_prog_sys_exit_fd(trace, key);
3460                 err = bpf_map_update_elem(map_exit_fd, &key, &prog_fd, BPF_ANY);
3461                 if (err)
3462                         break;
3463         }
3464
3465         /*
3466          * Now lets do a second pass looking for enabled syscalls without
3467          * an augmenter that have a signature that is a superset of another
3468          * syscall with an augmenter so that we can auto-reuse it.
3469          *
3470          * I.e. if we have an augmenter for the "open" syscall that has
3471          * this signature:
3472          *
3473          *   int open(const char *pathname, int flags, mode_t mode);
3474          *
3475          * I.e. that will collect just the first string argument, then we
3476          * can reuse it for the 'creat' syscall, that has this signature:
3477          *
3478          *   int creat(const char *pathname, mode_t mode);
3479          *
3480          * and for:
3481          *
3482          *   int stat(const char *pathname, struct stat *statbuf);
3483          *   int lstat(const char *pathname, struct stat *statbuf);
3484          *
3485          * Because the 'open' augmenter will collect the first arg as a string,
3486          * and leave alone all the other args, which already helps with
3487          * beautifying 'stat' and 'lstat''s pathname arg.
3488          *
3489          * Then, in time, when 'stat' gets an augmenter that collects both
3490          * first and second arg (this one on the raw_syscalls:sys_exit prog
3491          * array tail call, then that one will be used.
3492          */
3493         for (key = 0; key < trace->sctbl->syscalls.nr_entries; ++key) {
3494                 struct syscall *sc = trace__syscall_info(trace, NULL, key);
3495                 struct bpf_program *pair_prog;
3496                 int prog_fd;
3497
3498                 if (sc == NULL || sc->bpf_prog.sys_enter == NULL)
3499                         continue;
3500
3501                 /*
3502                  * For now we're just reusing the sys_enter prog, and if it
3503                  * already has an augmenter, we don't need to find one.
3504                  */
3505                 if (sc->bpf_prog.sys_enter != trace->syscalls.unaugmented_prog)
3506                         continue;
3507
3508                 /*
3509                  * Look at all the other syscalls for one that has a signature
3510                  * that is close enough that we can share:
3511                  */
3512                 pair_prog = trace__find_usable_bpf_prog_entry(trace, sc);
3513                 if (pair_prog == NULL)
3514                         continue;
3515
3516                 sc->bpf_prog.sys_enter = pair_prog;
3517
3518                 /*
3519                  * Update the BPF_MAP_TYPE_PROG_SHARED for raw_syscalls:sys_enter
3520                  * with the fd for the program we're reusing:
3521                  */
3522                 prog_fd = bpf_program__fd(sc->bpf_prog.sys_enter);
3523                 err = bpf_map_update_elem(map_enter_fd, &key, &prog_fd, BPF_ANY);
3524                 if (err)
3525                         break;
3526         }
3527
3528
3529         return err;
3530 }
3531
3532 static void trace__delete_augmented_syscalls(struct trace *trace)
3533 {
3534         struct evsel *evsel, *tmp;
3535
3536         evlist__remove(trace->evlist, trace->syscalls.events.augmented);
3537         evsel__delete(trace->syscalls.events.augmented);
3538         trace->syscalls.events.augmented = NULL;
3539
3540         evlist__for_each_entry_safe(trace->evlist, tmp, evsel) {
3541                 if (evsel->bpf_obj == trace->bpf_obj) {
3542                         evlist__remove(trace->evlist, evsel);
3543                         evsel__delete(evsel);
3544                 }
3545
3546         }
3547
3548         bpf_object__close(trace->bpf_obj);
3549         trace->bpf_obj = NULL;
3550 }
3551 #else // HAVE_LIBBPF_SUPPORT
3552 static struct bpf_map *trace__find_bpf_map_by_name(struct trace *trace __maybe_unused,
3553                                                    const char *name __maybe_unused)
3554 {
3555         return NULL;
3556 }
3557
3558 static void trace__set_bpf_map_filtered_pids(struct trace *trace __maybe_unused)
3559 {
3560 }
3561
3562 static void trace__set_bpf_map_syscalls(struct trace *trace __maybe_unused)
3563 {
3564 }
3565
3566 static struct bpf_program *trace__find_bpf_program_by_title(struct trace *trace __maybe_unused,
3567                                                             const char *name __maybe_unused)
3568 {
3569         return NULL;
3570 }
3571
3572 static int trace__init_syscalls_bpf_prog_array_maps(struct trace *trace __maybe_unused)
3573 {
3574         return 0;
3575 }
3576
3577 static void trace__delete_augmented_syscalls(struct trace *trace __maybe_unused)
3578 {
3579 }
3580 #endif // HAVE_LIBBPF_SUPPORT
3581
3582 static bool trace__only_augmented_syscalls_evsels(struct trace *trace)
3583 {
3584         struct evsel *evsel;
3585
3586         evlist__for_each_entry(trace->evlist, evsel) {
3587                 if (evsel == trace->syscalls.events.augmented ||
3588                     evsel->bpf_obj == trace->bpf_obj)
3589                         continue;
3590
3591                 return false;
3592         }
3593
3594         return true;
3595 }
3596
3597 static int trace__set_ev_qualifier_filter(struct trace *trace)
3598 {
3599         if (trace->syscalls.events.sys_enter)
3600                 return trace__set_ev_qualifier_tp_filter(trace);
3601         return 0;
3602 }
3603
3604 static int bpf_map__set_filter_pids(struct bpf_map *map __maybe_unused,
3605                                     size_t npids __maybe_unused, pid_t *pids __maybe_unused)
3606 {
3607         int err = 0;
3608 #ifdef HAVE_LIBBPF_SUPPORT
3609         bool value = true;
3610         int map_fd = bpf_map__fd(map);
3611         size_t i;
3612
3613         for (i = 0; i < npids; ++i) {
3614                 err = bpf_map_update_elem(map_fd, &pids[i], &value, BPF_ANY);
3615                 if (err)
3616                         break;
3617         }
3618 #endif
3619         return err;
3620 }
3621
3622 static int trace__set_filter_loop_pids(struct trace *trace)
3623 {
3624         unsigned int nr = 1, err;
3625         pid_t pids[32] = {
3626                 getpid(),
3627         };
3628         struct thread *thread = machine__find_thread(trace->host, pids[0], pids[0]);
3629
3630         while (thread && nr < ARRAY_SIZE(pids)) {
3631                 struct thread *parent = machine__find_thread(trace->host,
3632                                                              thread__ppid(thread),
3633                                                              thread__ppid(thread));
3634
3635                 if (parent == NULL)
3636                         break;
3637
3638                 if (!strcmp(thread__comm_str(parent), "sshd") ||
3639                     strstarts(thread__comm_str(parent), "gnome-terminal")) {
3640                         pids[nr++] = thread__tid(parent);
3641                         break;
3642                 }
3643                 thread = parent;
3644         }
3645
3646         err = evlist__append_tp_filter_pids(trace->evlist, nr, pids);
3647         if (!err && trace->filter_pids.map)
3648                 err = bpf_map__set_filter_pids(trace->filter_pids.map, nr, pids);
3649
3650         return err;
3651 }
3652
3653 static int trace__set_filter_pids(struct trace *trace)
3654 {
3655         int err = 0;
3656         /*
3657          * Better not use !target__has_task() here because we need to cover the
3658          * case where no threads were specified in the command line, but a
3659          * workload was, and in that case we will fill in the thread_map when
3660          * we fork the workload in evlist__prepare_workload.
3661          */
3662         if (trace->filter_pids.nr > 0) {
3663                 err = evlist__append_tp_filter_pids(trace->evlist, trace->filter_pids.nr,
3664                                                     trace->filter_pids.entries);
3665                 if (!err && trace->filter_pids.map) {
3666                         err = bpf_map__set_filter_pids(trace->filter_pids.map, trace->filter_pids.nr,
3667                                                        trace->filter_pids.entries);
3668                 }
3669         } else if (perf_thread_map__pid(trace->evlist->core.threads, 0) == -1) {
3670                 err = trace__set_filter_loop_pids(trace);
3671         }
3672
3673         return err;
3674 }
3675
3676 static int __trace__deliver_event(struct trace *trace, union perf_event *event)
3677 {
3678         struct evlist *evlist = trace->evlist;
3679         struct perf_sample sample;
3680         int err = evlist__parse_sample(evlist, event, &sample);
3681
3682         if (err)
3683                 fprintf(trace->output, "Can't parse sample, err = %d, skipping...\n", err);
3684         else
3685                 trace__handle_event(trace, event, &sample);
3686
3687         return 0;
3688 }
3689
3690 static int __trace__flush_events(struct trace *trace)
3691 {
3692         u64 first = ordered_events__first_time(&trace->oe.data);
3693         u64 flush = trace->oe.last - NSEC_PER_SEC;
3694
3695         /* Is there some thing to flush.. */
3696         if (first && first < flush)
3697                 return ordered_events__flush_time(&trace->oe.data, flush);
3698
3699         return 0;
3700 }
3701
3702 static int trace__flush_events(struct trace *trace)
3703 {
3704         return !trace->sort_events ? 0 : __trace__flush_events(trace);
3705 }
3706
3707 static int trace__deliver_event(struct trace *trace, union perf_event *event)
3708 {
3709         int err;
3710
3711         if (!trace->sort_events)
3712                 return __trace__deliver_event(trace, event);
3713
3714         err = evlist__parse_sample_timestamp(trace->evlist, event, &trace->oe.last);
3715         if (err && err != -1)
3716                 return err;
3717
3718         err = ordered_events__queue(&trace->oe.data, event, trace->oe.last, 0, NULL);
3719         if (err)
3720                 return err;
3721
3722         return trace__flush_events(trace);
3723 }
3724
3725 static int ordered_events__deliver_event(struct ordered_events *oe,
3726                                          struct ordered_event *event)
3727 {
3728         struct trace *trace = container_of(oe, struct trace, oe.data);
3729
3730         return __trace__deliver_event(trace, event->event);
3731 }
3732
3733 static struct syscall_arg_fmt *evsel__find_syscall_arg_fmt_by_name(struct evsel *evsel, char *arg)
3734 {
3735         struct tep_format_field *field;
3736         struct syscall_arg_fmt *fmt = __evsel__syscall_arg_fmt(evsel);
3737
3738         if (evsel->tp_format == NULL || fmt == NULL)
3739                 return NULL;
3740
3741         for (field = evsel->tp_format->format.fields; field; field = field->next, ++fmt)
3742                 if (strcmp(field->name, arg) == 0)
3743                         return fmt;
3744
3745         return NULL;
3746 }
3747
3748 static int trace__expand_filter(struct trace *trace __maybe_unused, struct evsel *evsel)
3749 {
3750         char *tok, *left = evsel->filter, *new_filter = evsel->filter;
3751
3752         while ((tok = strpbrk(left, "=<>!")) != NULL) {
3753                 char *right = tok + 1, *right_end;
3754
3755                 if (*right == '=')
3756                         ++right;
3757
3758                 while (isspace(*right))
3759                         ++right;
3760
3761                 if (*right == '\0')
3762                         break;
3763
3764                 while (!isalpha(*left))
3765                         if (++left == tok) {
3766                                 /*
3767                                  * Bail out, can't find the name of the argument that is being
3768                                  * used in the filter, let it try to set this filter, will fail later.
3769                                  */
3770                                 return 0;
3771                         }
3772
3773                 right_end = right + 1;
3774                 while (isalnum(*right_end) || *right_end == '_' || *right_end == '|')
3775                         ++right_end;
3776
3777                 if (isalpha(*right)) {
3778                         struct syscall_arg_fmt *fmt;
3779                         int left_size = tok - left,
3780                             right_size = right_end - right;
3781                         char arg[128];
3782
3783                         while (isspace(left[left_size - 1]))
3784                                 --left_size;
3785
3786                         scnprintf(arg, sizeof(arg), "%.*s", left_size, left);
3787
3788                         fmt = evsel__find_syscall_arg_fmt_by_name(evsel, arg);
3789                         if (fmt == NULL) {
3790                                 pr_err("\"%s\" not found in \"%s\", can't set filter \"%s\"\n",
3791                                        arg, evsel->name, evsel->filter);
3792                                 return -1;
3793                         }
3794
3795                         pr_debug2("trying to expand \"%s\" \"%.*s\" \"%.*s\" -> ",
3796                                  arg, (int)(right - tok), tok, right_size, right);
3797
3798                         if (fmt->strtoul) {
3799                                 u64 val;
3800                                 struct syscall_arg syscall_arg = {
3801                                         .parm = fmt->parm,
3802                                 };
3803
3804                                 if (fmt->strtoul(right, right_size, &syscall_arg, &val)) {
3805                                         char *n, expansion[19];
3806                                         int expansion_lenght = scnprintf(expansion, sizeof(expansion), "%#" PRIx64, val);
3807                                         int expansion_offset = right - new_filter;
3808
3809                                         pr_debug("%s", expansion);
3810
3811                                         if (asprintf(&n, "%.*s%s%s", expansion_offset, new_filter, expansion, right_end) < 0) {
3812                                                 pr_debug(" out of memory!\n");
3813                                                 free(new_filter);
3814                                                 return -1;
3815                                         }
3816                                         if (new_filter != evsel->filter)
3817                                                 free(new_filter);
3818                                         left = n + expansion_offset + expansion_lenght;
3819                                         new_filter = n;
3820                                 } else {
3821                                         pr_err("\"%.*s\" not found for \"%s\" in \"%s\", can't set filter \"%s\"\n",
3822                                                right_size, right, arg, evsel->name, evsel->filter);
3823                                         return -1;
3824                                 }
3825                         } else {
3826                                 pr_err("No resolver (strtoul) for \"%s\" in \"%s\", can't set filter \"%s\"\n",
3827                                        arg, evsel->name, evsel->filter);
3828                                 return -1;
3829                         }
3830
3831                         pr_debug("\n");
3832                 } else {
3833                         left = right_end;
3834                 }
3835         }
3836
3837         if (new_filter != evsel->filter) {
3838                 pr_debug("New filter for %s: %s\n", evsel->name, new_filter);
3839                 evsel__set_filter(evsel, new_filter);
3840                 free(new_filter);
3841         }
3842
3843         return 0;
3844 }
3845
3846 static int trace__expand_filters(struct trace *trace, struct evsel **err_evsel)
3847 {
3848         struct evlist *evlist = trace->evlist;
3849         struct evsel *evsel;
3850
3851         evlist__for_each_entry(evlist, evsel) {
3852                 if (evsel->filter == NULL)
3853                         continue;
3854
3855                 if (trace__expand_filter(trace, evsel)) {
3856                         *err_evsel = evsel;
3857                         return -1;
3858                 }
3859         }
3860
3861         return 0;
3862 }
3863
3864 static int trace__run(struct trace *trace, int argc, const char **argv)
3865 {
3866         struct evlist *evlist = trace->evlist;
3867         struct evsel *evsel, *pgfault_maj = NULL, *pgfault_min = NULL;
3868         int err = -1, i;
3869         unsigned long before;
3870         const bool forks = argc > 0;
3871         bool draining = false;
3872
3873         trace->live = true;
3874
3875         if (!trace->raw_augmented_syscalls) {
3876                 if (trace->trace_syscalls && trace__add_syscall_newtp(trace))
3877                         goto out_error_raw_syscalls;
3878
3879                 if (trace->trace_syscalls)
3880                         trace->vfs_getname = evlist__add_vfs_getname(evlist);
3881         }
3882
3883         if ((trace->trace_pgfaults & TRACE_PFMAJ)) {
3884                 pgfault_maj = evsel__new_pgfault(PERF_COUNT_SW_PAGE_FAULTS_MAJ);
3885                 if (pgfault_maj == NULL)
3886                         goto out_error_mem;
3887                 evsel__config_callchain(pgfault_maj, &trace->opts, &callchain_param);
3888                 evlist__add(evlist, pgfault_maj);
3889         }
3890
3891         if ((trace->trace_pgfaults & TRACE_PFMIN)) {
3892                 pgfault_min = evsel__new_pgfault(PERF_COUNT_SW_PAGE_FAULTS_MIN);
3893                 if (pgfault_min == NULL)
3894                         goto out_error_mem;
3895                 evsel__config_callchain(pgfault_min, &trace->opts, &callchain_param);
3896                 evlist__add(evlist, pgfault_min);
3897         }
3898
3899         /* Enable ignoring missing threads when -u/-p option is defined. */
3900         trace->opts.ignore_missing_thread = trace->opts.target.uid != UINT_MAX || trace->opts.target.pid;
3901
3902         if (trace->sched &&
3903             evlist__add_newtp(evlist, "sched", "sched_stat_runtime", trace__sched_stat_runtime))
3904                 goto out_error_sched_stat_runtime;
3905         /*
3906          * If a global cgroup was set, apply it to all the events without an
3907          * explicit cgroup. I.e.:
3908          *
3909          *      trace -G A -e sched:*switch
3910          *
3911          * Will set all raw_syscalls:sys_{enter,exit}, pgfault, vfs_getname, etc
3912          * _and_ sched:sched_switch to the 'A' cgroup, while:
3913          *
3914          * trace -e sched:*switch -G A
3915          *
3916          * will only set the sched:sched_switch event to the 'A' cgroup, all the
3917          * other events (raw_syscalls:sys_{enter,exit}, etc are left "without"
3918          * a cgroup (on the root cgroup, sys wide, etc).
3919          *
3920          * Multiple cgroups:
3921          *
3922          * trace -G A -e sched:*switch -G B
3923          *
3924          * the syscall ones go to the 'A' cgroup, the sched:sched_switch goes
3925          * to the 'B' cgroup.
3926          *
3927          * evlist__set_default_cgroup() grabs a reference of the passed cgroup
3928          * only for the evsels still without a cgroup, i.e. evsel->cgroup == NULL.
3929          */
3930         if (trace->cgroup)
3931                 evlist__set_default_cgroup(trace->evlist, trace->cgroup);
3932
3933         err = evlist__create_maps(evlist, &trace->opts.target);
3934         if (err < 0) {
3935                 fprintf(trace->output, "Problems parsing the target to trace, check your options!\n");
3936                 goto out_delete_evlist;
3937         }
3938
3939         err = trace__symbols_init(trace, evlist);
3940         if (err < 0) {
3941                 fprintf(trace->output, "Problems initializing symbol libraries!\n");
3942                 goto out_delete_evlist;
3943         }
3944
3945         evlist__config(evlist, &trace->opts, &callchain_param);
3946
3947         if (forks) {
3948                 err = evlist__prepare_workload(evlist, &trace->opts.target, argv, false, NULL);
3949                 if (err < 0) {
3950                         fprintf(trace->output, "Couldn't run the workload!\n");
3951                         goto out_delete_evlist;
3952                 }
3953                 workload_pid = evlist->workload.pid;
3954         }
3955
3956         err = evlist__open(evlist);
3957         if (err < 0)
3958                 goto out_error_open;
3959
3960         err = bpf__apply_obj_config();
3961         if (err) {
3962                 char errbuf[BUFSIZ];
3963
3964                 bpf__strerror_apply_obj_config(err, errbuf, sizeof(errbuf));
3965                 pr_err("ERROR: Apply config to BPF failed: %s\n",
3966                          errbuf);
3967                 goto out_error_open;
3968         }
3969
3970         err = trace__set_filter_pids(trace);
3971         if (err < 0)
3972                 goto out_error_mem;
3973
3974         if (trace->syscalls.prog_array.sys_enter)
3975                 trace__init_syscalls_bpf_prog_array_maps(trace);
3976
3977         if (trace->ev_qualifier_ids.nr > 0) {
3978                 err = trace__set_ev_qualifier_filter(trace);
3979                 if (err < 0)
3980                         goto out_errno;
3981
3982                 if (trace->syscalls.events.sys_exit) {
3983                         pr_debug("event qualifier tracepoint filter: %s\n",
3984                                  trace->syscalls.events.sys_exit->filter);
3985                 }
3986         }
3987
3988         /*
3989          * If the "close" syscall is not traced, then we will not have the
3990          * opportunity to, in syscall_arg__scnprintf_close_fd() invalidate the
3991          * fd->pathname table and were ending up showing the last value set by
3992          * syscalls opening a pathname and associating it with a descriptor or
3993          * reading it from /proc/pid/fd/ in cases where that doesn't make
3994          * sense.
3995          *
3996          *  So just disable this beautifier (SCA_FD, SCA_FDAT) when 'close' is
3997          *  not in use.
3998          */
3999         trace->fd_path_disabled = !trace__syscall_enabled(trace, syscalltbl__id(trace->sctbl, "close"));
4000
4001         err = trace__expand_filters(trace, &evsel);
4002         if (err)
4003                 goto out_delete_evlist;
4004         err = evlist__apply_filters(evlist, &evsel);
4005         if (err < 0)
4006                 goto out_error_apply_filters;
4007
4008         if (trace->dump.map)
4009                 bpf_map__fprintf(trace->dump.map, trace->output);
4010
4011         err = evlist__mmap(evlist, trace->opts.mmap_pages);
4012         if (err < 0)
4013                 goto out_error_mmap;
4014
4015         if (!target__none(&trace->opts.target) && !trace->opts.target.initial_delay)
4016                 evlist__enable(evlist);
4017
4018         if (forks)
4019                 evlist__start_workload(evlist);
4020
4021         if (trace->opts.target.initial_delay) {
4022                 usleep(trace->opts.target.initial_delay * 1000);
4023                 evlist__enable(evlist);
4024         }
4025
4026         trace->multiple_threads = perf_thread_map__pid(evlist->core.threads, 0) == -1 ||
4027                 perf_thread_map__nr(evlist->core.threads) > 1 ||
4028                 evlist__first(evlist)->core.attr.inherit;
4029
4030         /*
4031          * Now that we already used evsel->core.attr to ask the kernel to setup the
4032          * events, lets reuse evsel->core.attr.sample_max_stack as the limit in
4033          * trace__resolve_callchain(), allowing per-event max-stack settings
4034          * to override an explicitly set --max-stack global setting.
4035          */
4036         evlist__for_each_entry(evlist, evsel) {
4037                 if (evsel__has_callchain(evsel) &&
4038                     evsel->core.attr.sample_max_stack == 0)
4039                         evsel->core.attr.sample_max_stack = trace->max_stack;
4040         }
4041 again:
4042         before = trace->nr_events;
4043
4044         for (i = 0; i < evlist->core.nr_mmaps; i++) {
4045                 union perf_event *event;
4046                 struct mmap *md;
4047
4048                 md = &evlist->mmap[i];
4049                 if (perf_mmap__read_init(&md->core) < 0)
4050                         continue;
4051
4052                 while ((event = perf_mmap__read_event(&md->core)) != NULL) {
4053                         ++trace->nr_events;
4054
4055                         err = trace__deliver_event(trace, event);
4056                         if (err)
4057                                 goto out_disable;
4058
4059                         perf_mmap__consume(&md->core);
4060
4061                         if (interrupted)
4062                                 goto out_disable;
4063
4064                         if (done && !draining) {
4065                                 evlist__disable(evlist);
4066                                 draining = true;
4067                         }
4068                 }
4069                 perf_mmap__read_done(&md->core);
4070         }
4071
4072         if (trace->nr_events == before) {
4073                 int timeout = done ? 100 : -1;
4074
4075                 if (!draining && evlist__poll(evlist, timeout) > 0) {
4076                         if (evlist__filter_pollfd(evlist, POLLERR | POLLHUP | POLLNVAL) == 0)
4077                                 draining = true;
4078
4079                         goto again;
4080                 } else {
4081                         if (trace__flush_events(trace))
4082                                 goto out_disable;
4083                 }
4084         } else {
4085                 goto again;
4086         }
4087
4088 out_disable:
4089         thread__zput(trace->current);
4090
4091         evlist__disable(evlist);
4092
4093         if (trace->sort_events)
4094                 ordered_events__flush(&trace->oe.data, OE_FLUSH__FINAL);
4095
4096         if (!err) {
4097                 if (trace->summary)
4098                         trace__fprintf_thread_summary(trace, trace->output);
4099
4100                 if (trace->show_tool_stats) {
4101                         fprintf(trace->output, "Stats:\n "
4102                                                " vfs_getname : %" PRIu64 "\n"
4103                                                " proc_getname: %" PRIu64 "\n",
4104                                 trace->stats.vfs_getname,
4105                                 trace->stats.proc_getname);
4106                 }
4107         }
4108
4109 out_delete_evlist:
4110         trace__symbols__exit(trace);
4111         evlist__free_syscall_tp_fields(evlist);
4112         evlist__delete(evlist);
4113         cgroup__put(trace->cgroup);
4114         trace->evlist = NULL;
4115         trace->live = false;
4116         return err;
4117 {
4118         char errbuf[BUFSIZ];
4119
4120 out_error_sched_stat_runtime:
4121         tracing_path__strerror_open_tp(errno, errbuf, sizeof(errbuf), "sched", "sched_stat_runtime");
4122         goto out_error;
4123
4124 out_error_raw_syscalls:
4125         tracing_path__strerror_open_tp(errno, errbuf, sizeof(errbuf), "raw_syscalls", "sys_(enter|exit)");
4126         goto out_error;
4127
4128 out_error_mmap:
4129         evlist__strerror_mmap(evlist, errno, errbuf, sizeof(errbuf));
4130         goto out_error;
4131
4132 out_error_open:
4133         evlist__strerror_open(evlist, errno, errbuf, sizeof(errbuf));
4134
4135 out_error:
4136         fprintf(trace->output, "%s\n", errbuf);
4137         goto out_delete_evlist;
4138
4139 out_error_apply_filters:
4140         fprintf(trace->output,
4141                 "Failed to set filter \"%s\" on event %s with %d (%s)\n",
4142                 evsel->filter, evsel__name(evsel), errno,
4143                 str_error_r(errno, errbuf, sizeof(errbuf)));
4144         goto out_delete_evlist;
4145 }
4146 out_error_mem:
4147         fprintf(trace->output, "Not enough memory to run!\n");
4148         goto out_delete_evlist;
4149
4150 out_errno:
4151         fprintf(trace->output, "errno=%d,%s\n", errno, strerror(errno));
4152         goto out_delete_evlist;
4153 }
4154
4155 static int trace__replay(struct trace *trace)
4156 {
4157         const struct evsel_str_handler handlers[] = {
4158                 { "probe:vfs_getname",       trace__vfs_getname, },
4159         };
4160         struct perf_data data = {
4161                 .path  = input_name,
4162                 .mode  = PERF_DATA_MODE_READ,
4163                 .force = trace->force,
4164         };
4165         struct perf_session *session;
4166         struct evsel *evsel;
4167         int err = -1;
4168
4169         trace->tool.sample        = trace__process_sample;
4170         trace->tool.mmap          = perf_event__process_mmap;
4171         trace->tool.mmap2         = perf_event__process_mmap2;
4172         trace->tool.comm          = perf_event__process_comm;
4173         trace->tool.exit          = perf_event__process_exit;
4174         trace->tool.fork          = perf_event__process_fork;
4175         trace->tool.attr          = perf_event__process_attr;
4176         trace->tool.tracing_data  = perf_event__process_tracing_data;
4177         trace->tool.build_id      = perf_event__process_build_id;
4178         trace->tool.namespaces    = perf_event__process_namespaces;
4179
4180         trace->tool.ordered_events = true;
4181         trace->tool.ordering_requires_timestamps = true;
4182
4183         /* add tid to output */
4184         trace->multiple_threads = true;
4185
4186         session = perf_session__new(&data, &trace->tool);
4187         if (IS_ERR(session))
4188                 return PTR_ERR(session);
4189
4190         if (trace->opts.target.pid)
4191                 symbol_conf.pid_list_str = strdup(trace->opts.target.pid);
4192
4193         if (trace->opts.target.tid)
4194                 symbol_conf.tid_list_str = strdup(trace->opts.target.tid);
4195
4196         if (symbol__init(&session->header.env) < 0)
4197                 goto out;
4198
4199         trace->host = &session->machines.host;
4200
4201         err = perf_session__set_tracepoints_handlers(session, handlers);
4202         if (err)
4203                 goto out;
4204
4205         evsel = evlist__find_tracepoint_by_name(session->evlist, "raw_syscalls:sys_enter");
4206         trace->syscalls.events.sys_enter = evsel;
4207         /* older kernels have syscalls tp versus raw_syscalls */
4208         if (evsel == NULL)
4209                 evsel = evlist__find_tracepoint_by_name(session->evlist, "syscalls:sys_enter");
4210
4211         if (evsel &&
4212             (evsel__init_raw_syscall_tp(evsel, trace__sys_enter) < 0 ||
4213             perf_evsel__init_sc_tp_ptr_field(evsel, args))) {
4214                 pr_err("Error during initialize raw_syscalls:sys_enter event\n");
4215                 goto out;
4216         }
4217
4218         evsel = evlist__find_tracepoint_by_name(session->evlist, "raw_syscalls:sys_exit");
4219         trace->syscalls.events.sys_exit = evsel;
4220         if (evsel == NULL)
4221                 evsel = evlist__find_tracepoint_by_name(session->evlist, "syscalls:sys_exit");
4222         if (evsel &&
4223             (evsel__init_raw_syscall_tp(evsel, trace__sys_exit) < 0 ||
4224             perf_evsel__init_sc_tp_uint_field(evsel, ret))) {
4225                 pr_err("Error during initialize raw_syscalls:sys_exit event\n");
4226                 goto out;
4227         }
4228
4229         evlist__for_each_entry(session->evlist, evsel) {
4230                 if (evsel->core.attr.type == PERF_TYPE_SOFTWARE &&
4231                     (evsel->core.attr.config == PERF_COUNT_SW_PAGE_FAULTS_MAJ ||
4232                      evsel->core.attr.config == PERF_COUNT_SW_PAGE_FAULTS_MIN ||
4233                      evsel->core.attr.config == PERF_COUNT_SW_PAGE_FAULTS))
4234                         evsel->handler = trace__pgfault;
4235         }
4236
4237         setup_pager();
4238
4239         err = perf_session__process_events(session);
4240         if (err)
4241                 pr_err("Failed to process events, error %d", err);
4242
4243         else if (trace->summary)
4244                 trace__fprintf_thread_summary(trace, trace->output);
4245
4246 out:
4247         perf_session__delete(session);
4248
4249         return err;
4250 }
4251
4252 static size_t trace__fprintf_threads_header(FILE *fp)
4253 {
4254         size_t printed;
4255
4256         printed  = fprintf(fp, "\n Summary of events:\n\n");
4257
4258         return printed;
4259 }
4260
4261 DEFINE_RESORT_RB(syscall_stats, a->msecs > b->msecs,
4262         struct syscall_stats *stats;
4263         double               msecs;
4264         int                  syscall;
4265 )
4266 {
4267         struct int_node *source = rb_entry(nd, struct int_node, rb_node);
4268         struct syscall_stats *stats = source->priv;
4269
4270         entry->syscall = source->i;
4271         entry->stats   = stats;
4272         entry->msecs   = stats ? (u64)stats->stats.n * (avg_stats(&stats->stats) / NSEC_PER_MSEC) : 0;
4273 }
4274
4275 static size_t thread__dump_stats(struct thread_trace *ttrace,
4276                                  struct trace *trace, FILE *fp)
4277 {
4278         size_t printed = 0;
4279         struct syscall *sc;
4280         struct rb_node *nd;
4281         DECLARE_RESORT_RB_INTLIST(syscall_stats, ttrace->syscall_stats);
4282
4283         if (syscall_stats == NULL)
4284                 return 0;
4285
4286         printed += fprintf(fp, "\n");
4287
4288         printed += fprintf(fp, "   syscall            calls  errors  total       min       avg       max       stddev\n");
4289         printed += fprintf(fp, "                                     (msec)    (msec)    (msec)    (msec)        (%%)\n");
4290         printed += fprintf(fp, "   --------------- --------  ------ -------- --------- --------- ---------     ------\n");
4291
4292         resort_rb__for_each_entry(nd, syscall_stats) {
4293                 struct syscall_stats *stats = syscall_stats_entry->stats;
4294                 if (stats) {
4295                         double min = (double)(stats->stats.min) / NSEC_PER_MSEC;
4296                         double max = (double)(stats->stats.max) / NSEC_PER_MSEC;
4297                         double avg = avg_stats(&stats->stats);
4298                         double pct;
4299                         u64 n = (u64)stats->stats.n;
4300
4301                         pct = avg ? 100.0 * stddev_stats(&stats->stats) / avg : 0.0;
4302                         avg /= NSEC_PER_MSEC;
4303
4304                         sc = &trace->syscalls.table[syscall_stats_entry->syscall];
4305                         printed += fprintf(fp, "   %-15s", sc->name);
4306                         printed += fprintf(fp, " %8" PRIu64 " %6" PRIu64 " %9.3f %9.3f %9.3f",
4307                                            n, stats->nr_failures, syscall_stats_entry->msecs, min, avg);
4308                         printed += fprintf(fp, " %9.3f %9.2f%%\n", max, pct);
4309
4310                         if (trace->errno_summary && stats->nr_failures) {
4311                                 const char *arch_name = perf_env__arch(trace->host->env);
4312                                 int e;
4313
4314                                 for (e = 0; e < stats->max_errno; ++e) {
4315                                         if (stats->errnos[e] != 0)
4316                                                 fprintf(fp, "\t\t\t\t%s: %d\n", arch_syscalls__strerrno(arch_name, e + 1), stats->errnos[e]);
4317                                 }
4318                         }
4319                 }
4320         }
4321
4322         resort_rb__delete(syscall_stats);
4323         printed += fprintf(fp, "\n\n");
4324
4325         return printed;
4326 }
4327
4328 static size_t trace__fprintf_thread(FILE *fp, struct thread *thread, struct trace *trace)
4329 {
4330         size_t printed = 0;
4331         struct thread_trace *ttrace = thread__priv(thread);
4332         double ratio;
4333
4334         if (ttrace == NULL)
4335                 return 0;
4336
4337         ratio = (double)ttrace->nr_events / trace->nr_events * 100.0;
4338
4339         printed += fprintf(fp, " %s (%d), ", thread__comm_str(thread), thread__tid(thread));
4340         printed += fprintf(fp, "%lu events, ", ttrace->nr_events);
4341         printed += fprintf(fp, "%.1f%%", ratio);
4342         if (ttrace->pfmaj)
4343                 printed += fprintf(fp, ", %lu majfaults", ttrace->pfmaj);
4344         if (ttrace->pfmin)
4345                 printed += fprintf(fp, ", %lu minfaults", ttrace->pfmin);
4346         if (trace->sched)
4347                 printed += fprintf(fp, ", %.3f msec\n", ttrace->runtime_ms);
4348         else if (fputc('\n', fp) != EOF)
4349                 ++printed;
4350
4351         printed += thread__dump_stats(ttrace, trace, fp);
4352
4353         return printed;
4354 }
4355
4356 static unsigned long thread__nr_events(struct thread_trace *ttrace)
4357 {
4358         return ttrace ? ttrace->nr_events : 0;
4359 }
4360
4361 DEFINE_RESORT_RB(threads,
4362                 (thread__nr_events(thread__priv(a->thread)) <
4363                  thread__nr_events(thread__priv(b->thread))),
4364         struct thread *thread;
4365 )
4366 {
4367         entry->thread = rb_entry(nd, struct thread_rb_node, rb_node)->thread;
4368 }
4369
4370 static size_t trace__fprintf_thread_summary(struct trace *trace, FILE *fp)
4371 {
4372         size_t printed = trace__fprintf_threads_header(fp);
4373         struct rb_node *nd;
4374         int i;
4375
4376         for (i = 0; i < THREADS__TABLE_SIZE; i++) {
4377                 DECLARE_RESORT_RB_MACHINE_THREADS(threads, trace->host, i);
4378
4379                 if (threads == NULL) {
4380                         fprintf(fp, "%s", "Error sorting output by nr_events!\n");
4381                         return 0;
4382                 }
4383
4384                 resort_rb__for_each_entry(nd, threads)
4385                         printed += trace__fprintf_thread(fp, threads_entry->thread, trace);
4386
4387                 resort_rb__delete(threads);
4388         }
4389         return printed;
4390 }
4391
4392 static int trace__set_duration(const struct option *opt, const char *str,
4393                                int unset __maybe_unused)
4394 {
4395         struct trace *trace = opt->value;
4396
4397         trace->duration_filter = atof(str);
4398         return 0;
4399 }
4400
4401 static int trace__set_filter_pids_from_option(const struct option *opt, const char *str,
4402                                               int unset __maybe_unused)
4403 {
4404         int ret = -1;
4405         size_t i;
4406         struct trace *trace = opt->value;
4407         /*
4408          * FIXME: introduce a intarray class, plain parse csv and create a
4409          * { int nr, int entries[] } struct...
4410          */
4411         struct intlist *list = intlist__new(str);
4412
4413         if (list == NULL)
4414                 return -1;
4415
4416         i = trace->filter_pids.nr = intlist__nr_entries(list) + 1;
4417         trace->filter_pids.entries = calloc(i, sizeof(pid_t));
4418
4419         if (trace->filter_pids.entries == NULL)
4420                 goto out;
4421
4422         trace->filter_pids.entries[0] = getpid();
4423
4424         for (i = 1; i < trace->filter_pids.nr; ++i)
4425                 trace->filter_pids.entries[i] = intlist__entry(list, i - 1)->i;
4426
4427         intlist__delete(list);
4428         ret = 0;
4429 out:
4430         return ret;
4431 }
4432
4433 static int trace__open_output(struct trace *trace, const char *filename)
4434 {
4435         struct stat st;
4436
4437         if (!stat(filename, &st) && st.st_size) {
4438                 char oldname[PATH_MAX];
4439
4440                 scnprintf(oldname, sizeof(oldname), "%s.old", filename);
4441                 unlink(oldname);
4442                 rename(filename, oldname);
4443         }
4444
4445         trace->output = fopen(filename, "w");
4446
4447         return trace->output == NULL ? -errno : 0;
4448 }
4449
4450 static int parse_pagefaults(const struct option *opt, const char *str,
4451                             int unset __maybe_unused)
4452 {
4453         int *trace_pgfaults = opt->value;
4454
4455         if (strcmp(str, "all") == 0)
4456                 *trace_pgfaults |= TRACE_PFMAJ | TRACE_PFMIN;
4457         else if (strcmp(str, "maj") == 0)
4458                 *trace_pgfaults |= TRACE_PFMAJ;
4459         else if (strcmp(str, "min") == 0)
4460                 *trace_pgfaults |= TRACE_PFMIN;
4461         else
4462                 return -1;
4463
4464         return 0;
4465 }
4466
4467 static void evlist__set_default_evsel_handler(struct evlist *evlist, void *handler)
4468 {
4469         struct evsel *evsel;
4470
4471         evlist__for_each_entry(evlist, evsel) {
4472                 if (evsel->handler == NULL)
4473                         evsel->handler = handler;
4474         }
4475 }
4476
4477 static void evsel__set_syscall_arg_fmt(struct evsel *evsel, const char *name)
4478 {
4479         struct syscall_arg_fmt *fmt = evsel__syscall_arg_fmt(evsel);
4480
4481         if (fmt) {
4482                 const struct syscall_fmt *scfmt = syscall_fmt__find(name);
4483
4484                 if (scfmt) {
4485                         int skip = 0;
4486
4487                         if (strcmp(evsel->tp_format->format.fields->name, "__syscall_nr") == 0 ||
4488                             strcmp(evsel->tp_format->format.fields->name, "nr") == 0)
4489                                 ++skip;
4490
4491                         memcpy(fmt + skip, scfmt->arg, (evsel->tp_format->format.nr_fields - skip) * sizeof(*fmt));
4492                 }
4493         }
4494 }
4495
4496 static int evlist__set_syscall_tp_fields(struct evlist *evlist)
4497 {
4498         struct evsel *evsel;
4499
4500         evlist__for_each_entry(evlist, evsel) {
4501                 if (evsel->priv || !evsel->tp_format)
4502                         continue;
4503
4504                 if (strcmp(evsel->tp_format->system, "syscalls")) {
4505                         evsel__init_tp_arg_scnprintf(evsel);
4506                         continue;
4507                 }
4508
4509                 if (evsel__init_syscall_tp(evsel))
4510                         return -1;
4511
4512                 if (!strncmp(evsel->tp_format->name, "sys_enter_", 10)) {
4513                         struct syscall_tp *sc = __evsel__syscall_tp(evsel);
4514
4515                         if (__tp_field__init_ptr(&sc->args, sc->id.offset + sizeof(u64)))
4516                                 return -1;
4517
4518                         evsel__set_syscall_arg_fmt(evsel, evsel->tp_format->name + sizeof("sys_enter_") - 1);
4519                 } else if (!strncmp(evsel->tp_format->name, "sys_exit_", 9)) {
4520                         struct syscall_tp *sc = __evsel__syscall_tp(evsel);
4521
4522                         if (__tp_field__init_uint(&sc->ret, sizeof(u64), sc->id.offset + sizeof(u64), evsel->needs_swap))
4523                                 return -1;
4524
4525                         evsel__set_syscall_arg_fmt(evsel, evsel->tp_format->name + sizeof("sys_exit_") - 1);
4526                 }
4527         }
4528
4529         return 0;
4530 }
4531
4532 /*
4533  * XXX: Hackish, just splitting the combined -e+--event (syscalls
4534  * (raw_syscalls:{sys_{enter,exit}} + events (tracepoints, HW, SW, etc) to use
4535  * existing facilities unchanged (trace->ev_qualifier + parse_options()).
4536  *
4537  * It'd be better to introduce a parse_options() variant that would return a
4538  * list with the terms it didn't match to an event...
4539  */
4540 static int trace__parse_events_option(const struct option *opt, const char *str,
4541                                       int unset __maybe_unused)
4542 {
4543         struct trace *trace = (struct trace *)opt->value;
4544         const char *s = str;
4545         char *sep = NULL, *lists[2] = { NULL, NULL, };
4546         int len = strlen(str) + 1, err = -1, list, idx;
4547         char *strace_groups_dir = system_path(STRACE_GROUPS_DIR);
4548         char group_name[PATH_MAX];
4549         const struct syscall_fmt *fmt;
4550
4551         if (strace_groups_dir == NULL)
4552                 return -1;
4553
4554         if (*s == '!') {
4555                 ++s;
4556                 trace->not_ev_qualifier = true;
4557         }
4558
4559         while (1) {
4560                 if ((sep = strchr(s, ',')) != NULL)
4561                         *sep = '\0';
4562
4563                 list = 0;
4564                 if (syscalltbl__id(trace->sctbl, s) >= 0 ||
4565                     syscalltbl__strglobmatch_first(trace->sctbl, s, &idx) >= 0) {
4566                         list = 1;
4567                         goto do_concat;
4568                 }
4569
4570                 fmt = syscall_fmt__find_by_alias(s);
4571                 if (fmt != NULL) {
4572                         list = 1;
4573                         s = fmt->name;
4574                 } else {
4575                         path__join(group_name, sizeof(group_name), strace_groups_dir, s);
4576                         if (access(group_name, R_OK) == 0)
4577                                 list = 1;
4578                 }
4579 do_concat:
4580                 if (lists[list]) {
4581                         sprintf(lists[list] + strlen(lists[list]), ",%s", s);
4582                 } else {
4583                         lists[list] = malloc(len);
4584                         if (lists[list] == NULL)
4585                                 goto out;
4586                         strcpy(lists[list], s);
4587                 }
4588
4589                 if (!sep)
4590                         break;
4591
4592                 *sep = ',';
4593                 s = sep + 1;
4594         }
4595
4596         if (lists[1] != NULL) {
4597                 struct strlist_config slist_config = {
4598                         .dirname = strace_groups_dir,
4599                 };
4600
4601                 trace->ev_qualifier = strlist__new(lists[1], &slist_config);
4602                 if (trace->ev_qualifier == NULL) {
4603                         fputs("Not enough memory to parse event qualifier", trace->output);
4604                         goto out;
4605                 }
4606
4607                 if (trace__validate_ev_qualifier(trace))
4608                         goto out;
4609                 trace->trace_syscalls = true;
4610         }
4611
4612         err = 0;
4613
4614         if (lists[0]) {
4615                 struct parse_events_option_args parse_events_option_args = {
4616                         .evlistp = &trace->evlist,
4617                 };
4618                 struct option o = {
4619                         .value = &parse_events_option_args,
4620                 };
4621                 err = parse_events_option(&o, lists[0], 0);
4622         }
4623 out:
4624         free(strace_groups_dir);
4625         free(lists[0]);
4626         free(lists[1]);
4627         if (sep)
4628                 *sep = ',';
4629
4630         return err;
4631 }
4632
4633 static int trace__parse_cgroups(const struct option *opt, const char *str, int unset)
4634 {
4635         struct trace *trace = opt->value;
4636
4637         if (!list_empty(&trace->evlist->core.entries)) {
4638                 struct option o = {
4639                         .value = &trace->evlist,
4640                 };
4641                 return parse_cgroups(&o, str, unset);
4642         }
4643         trace->cgroup = evlist__findnew_cgroup(trace->evlist, str);
4644
4645         return 0;
4646 }
4647
4648 static int trace__config(const char *var, const char *value, void *arg)
4649 {
4650         struct trace *trace = arg;
4651         int err = 0;
4652
4653         if (!strcmp(var, "trace.add_events")) {
4654                 trace->perfconfig_events = strdup(value);
4655                 if (trace->perfconfig_events == NULL) {
4656                         pr_err("Not enough memory for %s\n", "trace.add_events");
4657                         return -1;
4658                 }
4659         } else if (!strcmp(var, "trace.show_timestamp")) {
4660                 trace->show_tstamp = perf_config_bool(var, value);
4661         } else if (!strcmp(var, "trace.show_duration")) {
4662                 trace->show_duration = perf_config_bool(var, value);
4663         } else if (!strcmp(var, "trace.show_arg_names")) {
4664                 trace->show_arg_names = perf_config_bool(var, value);
4665                 if (!trace->show_arg_names)
4666                         trace->show_zeros = true;
4667         } else if (!strcmp(var, "trace.show_zeros")) {
4668                 bool new_show_zeros = perf_config_bool(var, value);
4669                 if (!trace->show_arg_names && !new_show_zeros) {
4670                         pr_warning("trace.show_zeros has to be set when trace.show_arg_names=no\n");
4671                         goto out;
4672                 }
4673                 trace->show_zeros = new_show_zeros;
4674         } else if (!strcmp(var, "trace.show_prefix")) {
4675                 trace->show_string_prefix = perf_config_bool(var, value);
4676         } else if (!strcmp(var, "trace.no_inherit")) {
4677                 trace->opts.no_inherit = perf_config_bool(var, value);
4678         } else if (!strcmp(var, "trace.args_alignment")) {
4679                 int args_alignment = 0;
4680                 if (perf_config_int(&args_alignment, var, value) == 0)
4681                         trace->args_alignment = args_alignment;
4682         } else if (!strcmp(var, "trace.tracepoint_beautifiers")) {
4683                 if (strcasecmp(value, "libtraceevent") == 0)
4684                         trace->libtraceevent_print = true;
4685                 else if (strcasecmp(value, "libbeauty") == 0)
4686                         trace->libtraceevent_print = false;
4687         }
4688 out:
4689         return err;
4690 }
4691
4692 static void trace__exit(struct trace *trace)
4693 {
4694         int i;
4695
4696         strlist__delete(trace->ev_qualifier);
4697         zfree(&trace->ev_qualifier_ids.entries);
4698         if (trace->syscalls.table) {
4699                 for (i = 0; i <= trace->sctbl->syscalls.max_id; i++)
4700                         syscall__exit(&trace->syscalls.table[i]);
4701                 zfree(&trace->syscalls.table);
4702         }
4703         syscalltbl__delete(trace->sctbl);
4704         zfree(&trace->perfconfig_events);
4705 }
4706
4707 int cmd_trace(int argc, const char **argv)
4708 {
4709         const char *trace_usage[] = {
4710                 "perf trace [<options>] [<command>]",
4711                 "perf trace [<options>] -- <command> [<options>]",
4712                 "perf trace record [<options>] [<command>]",
4713                 "perf trace record [<options>] -- <command> [<options>]",
4714                 NULL
4715         };
4716         struct trace trace = {
4717                 .opts = {
4718                         .target = {
4719                                 .uid       = UINT_MAX,
4720                                 .uses_mmap = true,
4721                         },
4722                         .user_freq     = UINT_MAX,
4723                         .user_interval = ULLONG_MAX,
4724                         .no_buffering  = true,
4725                         .mmap_pages    = UINT_MAX,
4726                 },
4727                 .output = stderr,
4728                 .show_comm = true,
4729                 .show_tstamp = true,
4730                 .show_duration = true,
4731                 .show_arg_names = true,
4732                 .args_alignment = 70,
4733                 .trace_syscalls = false,
4734                 .kernel_syscallchains = false,
4735                 .max_stack = UINT_MAX,
4736                 .max_events = ULONG_MAX,
4737         };
4738         const char *map_dump_str = NULL;
4739         const char *output_name = NULL;
4740         const struct option trace_options[] = {
4741         OPT_CALLBACK('e', "event", &trace, "event",
4742                      "event/syscall selector. use 'perf list' to list available events",
4743                      trace__parse_events_option),
4744         OPT_CALLBACK(0, "filter", &trace.evlist, "filter",
4745                      "event filter", parse_filter),
4746         OPT_BOOLEAN(0, "comm", &trace.show_comm,
4747                     "show the thread COMM next to its id"),
4748         OPT_BOOLEAN(0, "tool_stats", &trace.show_tool_stats, "show tool stats"),
4749         OPT_CALLBACK(0, "expr", &trace, "expr", "list of syscalls/events to trace",
4750                      trace__parse_events_option),
4751         OPT_STRING('o', "output", &output_name, "file", "output file name"),
4752         OPT_STRING('i', "input", &input_name, "file", "Analyze events in file"),
4753         OPT_STRING('p', "pid", &trace.opts.target.pid, "pid",
4754                     "trace events on existing process id"),
4755         OPT_STRING('t', "tid", &trace.opts.target.tid, "tid",
4756                     "trace events on existing thread id"),
4757         OPT_CALLBACK(0, "filter-pids", &trace, "CSV list of pids",
4758                      "pids to filter (by the kernel)", trace__set_filter_pids_from_option),
4759         OPT_BOOLEAN('a', "all-cpus", &trace.opts.target.system_wide,
4760                     "system-wide collection from all CPUs"),
4761         OPT_STRING('C', "cpu", &trace.opts.target.cpu_list, "cpu",
4762                     "list of cpus to monitor"),
4763         OPT_BOOLEAN(0, "no-inherit", &trace.opts.no_inherit,
4764                     "child tasks do not inherit counters"),
4765         OPT_CALLBACK('m', "mmap-pages", &trace.opts.mmap_pages, "pages",
4766                      "number of mmap data pages", evlist__parse_mmap_pages),
4767         OPT_STRING('u', "uid", &trace.opts.target.uid_str, "user",
4768                    "user to profile"),
4769         OPT_CALLBACK(0, "duration", &trace, "float",
4770                      "show only events with duration > N.M ms",
4771                      trace__set_duration),
4772 #ifdef HAVE_LIBBPF_SUPPORT
4773         OPT_STRING(0, "map-dump", &map_dump_str, "BPF map", "BPF map to periodically dump"),
4774 #endif
4775         OPT_BOOLEAN(0, "sched", &trace.sched, "show blocking scheduler events"),
4776         OPT_INCR('v', "verbose", &verbose, "be more verbose"),
4777         OPT_BOOLEAN('T', "time", &trace.full_time,
4778                     "Show full timestamp, not time relative to first start"),
4779         OPT_BOOLEAN(0, "failure", &trace.failure_only,
4780                     "Show only syscalls that failed"),
4781         OPT_BOOLEAN('s', "summary", &trace.summary_only,
4782                     "Show only syscall summary with statistics"),
4783         OPT_BOOLEAN('S', "with-summary", &trace.summary,
4784                     "Show all syscalls and summary with statistics"),
4785         OPT_BOOLEAN(0, "errno-summary", &trace.errno_summary,
4786                     "Show errno stats per syscall, use with -s or -S"),
4787         OPT_CALLBACK_DEFAULT('F', "pf", &trace.trace_pgfaults, "all|maj|min",
4788                      "Trace pagefaults", parse_pagefaults, "maj"),
4789         OPT_BOOLEAN(0, "syscalls", &trace.trace_syscalls, "Trace syscalls"),
4790         OPT_BOOLEAN('f', "force", &trace.force, "don't complain, do it"),
4791         OPT_CALLBACK(0, "call-graph", &trace.opts,
4792                      "record_mode[,record_size]", record_callchain_help,
4793                      &record_parse_callchain_opt),
4794         OPT_BOOLEAN(0, "libtraceevent_print", &trace.libtraceevent_print,
4795                     "Use libtraceevent to print the tracepoint arguments."),
4796         OPT_BOOLEAN(0, "kernel-syscall-graph", &trace.kernel_syscallchains,
4797                     "Show the kernel callchains on the syscall exit path"),
4798         OPT_ULONG(0, "max-events", &trace.max_events,
4799                 "Set the maximum number of events to print, exit after that is reached. "),
4800         OPT_UINTEGER(0, "min-stack", &trace.min_stack,
4801                      "Set the minimum stack depth when parsing the callchain, "
4802                      "anything below the specified depth will be ignored."),
4803         OPT_UINTEGER(0, "max-stack", &trace.max_stack,
4804                      "Set the maximum stack depth when parsing the callchain, "
4805                      "anything beyond the specified depth will be ignored. "
4806                      "Default: kernel.perf_event_max_stack or " __stringify(PERF_MAX_STACK_DEPTH)),
4807         OPT_BOOLEAN(0, "sort-events", &trace.sort_events,
4808                         "Sort batch of events before processing, use if getting out of order events"),
4809         OPT_BOOLEAN(0, "print-sample", &trace.print_sample,
4810                         "print the PERF_RECORD_SAMPLE PERF_SAMPLE_ info, for debugging"),
4811         OPT_UINTEGER(0, "proc-map-timeout", &proc_map_timeout,
4812                         "per thread proc mmap processing timeout in ms"),
4813         OPT_CALLBACK('G', "cgroup", &trace, "name", "monitor event in cgroup name only",
4814                      trace__parse_cgroups),
4815         OPT_INTEGER('D', "delay", &trace.opts.target.initial_delay,
4816                      "ms to wait before starting measurement after program "
4817                      "start"),
4818         OPTS_EVSWITCH(&trace.evswitch),
4819         OPT_END()
4820         };
4821         bool __maybe_unused max_stack_user_set = true;
4822         bool mmap_pages_user_set = true;
4823         struct evsel *evsel;
4824         const char * const trace_subcommands[] = { "record", NULL };
4825         int err = -1;
4826         char bf[BUFSIZ];
4827         struct sigaction sigchld_act;
4828
4829         signal(SIGSEGV, sighandler_dump_stack);
4830         signal(SIGFPE, sighandler_dump_stack);
4831         signal(SIGINT, sighandler_interrupt);
4832
4833         memset(&sigchld_act, 0, sizeof(sigchld_act));
4834         sigchld_act.sa_flags = SA_SIGINFO;
4835         sigchld_act.sa_sigaction = sighandler_chld;
4836         sigaction(SIGCHLD, &sigchld_act, NULL);
4837
4838         trace.evlist = evlist__new();
4839         trace.sctbl = syscalltbl__new();
4840
4841         if (trace.evlist == NULL || trace.sctbl == NULL) {
4842                 pr_err("Not enough memory to run!\n");
4843                 err = -ENOMEM;
4844                 goto out;
4845         }
4846
4847         /*
4848          * Parsing .perfconfig may entail creating a BPF event, that may need
4849          * to create BPF maps, so bump RLIM_MEMLOCK as the default 64K setting
4850          * is too small. This affects just this process, not touching the
4851          * global setting. If it fails we'll get something in 'perf trace -v'
4852          * to help diagnose the problem.
4853          */
4854         rlimit__bump_memlock();
4855
4856         err = perf_config(trace__config, &trace);
4857         if (err)
4858                 goto out;
4859
4860         argc = parse_options_subcommand(argc, argv, trace_options, trace_subcommands,
4861                                  trace_usage, PARSE_OPT_STOP_AT_NON_OPTION);
4862
4863         /*
4864          * Here we already passed thru trace__parse_events_option() and it has
4865          * already figured out if -e syscall_name, if not but if --event
4866          * foo:bar was used, the user is interested _just_ in those, say,
4867          * tracepoint events, not in the strace-like syscall-name-based mode.
4868          *
4869          * This is important because we need to check if strace-like mode is
4870          * needed to decided if we should filter out the eBPF
4871          * __augmented_syscalls__ code, if it is in the mix, say, via
4872          * .perfconfig trace.add_events, and filter those out.
4873          */
4874         if (!trace.trace_syscalls && !trace.trace_pgfaults &&
4875             trace.evlist->core.nr_entries == 0 /* Was --events used? */) {
4876                 trace.trace_syscalls = true;
4877         }
4878         /*
4879          * Now that we have --verbose figured out, lets see if we need to parse
4880          * events from .perfconfig, so that if those events fail parsing, say some
4881          * BPF program fails, then we'll be able to use --verbose to see what went
4882          * wrong in more detail.
4883          */
4884         if (trace.perfconfig_events != NULL) {
4885                 struct parse_events_error parse_err;
4886
4887                 parse_events_error__init(&parse_err);
4888                 err = parse_events(trace.evlist, trace.perfconfig_events, &parse_err);
4889                 if (err)
4890                         parse_events_error__print(&parse_err, trace.perfconfig_events);
4891                 parse_events_error__exit(&parse_err);
4892                 if (err)
4893                         goto out;
4894         }
4895
4896         if ((nr_cgroups || trace.cgroup) && !trace.opts.target.system_wide) {
4897                 usage_with_options_msg(trace_usage, trace_options,
4898                                        "cgroup monitoring only available in system-wide mode");
4899         }
4900
4901         evsel = bpf__setup_output_event(trace.evlist, "__augmented_syscalls__");
4902         if (IS_ERR(evsel)) {
4903                 bpf__strerror_setup_output_event(trace.evlist, PTR_ERR(evsel), bf, sizeof(bf));
4904                 pr_err("ERROR: Setup trace syscalls enter failed: %s\n", bf);
4905                 goto out;
4906         }
4907
4908         if (evsel) {
4909                 trace.syscalls.events.augmented = evsel;
4910
4911                 evsel = evlist__find_tracepoint_by_name(trace.evlist, "raw_syscalls:sys_enter");
4912                 if (evsel == NULL) {
4913                         pr_err("ERROR: raw_syscalls:sys_enter not found in the augmented BPF object\n");
4914                         goto out;
4915                 }
4916
4917                 if (evsel->bpf_obj == NULL) {
4918                         pr_err("ERROR: raw_syscalls:sys_enter not associated to a BPF object\n");
4919                         goto out;
4920                 }
4921
4922                 trace.bpf_obj = evsel->bpf_obj;
4923
4924                 /*
4925                  * If we have _just_ the augmenter event but don't have a
4926                  * explicit --syscalls, then assume we want all strace-like
4927                  * syscalls:
4928                  */
4929                 if (!trace.trace_syscalls && trace__only_augmented_syscalls_evsels(&trace))
4930                         trace.trace_syscalls = true;
4931                 /*
4932                  * So, if we have a syscall augmenter, but trace_syscalls, aka
4933                  * strace-like syscall tracing is not set, then we need to trow
4934                  * away the augmenter, i.e. all the events that were created
4935                  * from that BPF object file.
4936                  *
4937                  * This is more to fix the current .perfconfig trace.add_events
4938                  * style of setting up the strace-like eBPF based syscall point
4939                  * payload augmenter.
4940                  *
4941                  * All this complexity will be avoided by adding an alternative
4942                  * to trace.add_events in the form of
4943                  * trace.bpf_augmented_syscalls, that will be only parsed if we
4944                  * need it.
4945                  *
4946                  * .perfconfig trace.add_events is still useful if we want, for
4947                  * instance, have msr_write.msr in some .perfconfig profile based
4948                  * 'perf trace --config determinism.profile' mode, where for some
4949                  * particular goal/workload type we want a set of events and
4950                  * output mode (with timings, etc) instead of having to add
4951                  * all via the command line.
4952                  *
4953                  * Also --config to specify an alternate .perfconfig file needs
4954                  * to be implemented.
4955                  */
4956                 if (!trace.trace_syscalls) {
4957                         trace__delete_augmented_syscalls(&trace);
4958                 } else {
4959                         trace__set_bpf_map_filtered_pids(&trace);
4960                         trace__set_bpf_map_syscalls(&trace);
4961                         trace.syscalls.unaugmented_prog = trace__find_bpf_program_by_title(&trace, "!raw_syscalls:unaugmented");
4962                 }
4963         }
4964
4965         err = bpf__setup_stdout(trace.evlist);
4966         if (err) {
4967                 bpf__strerror_setup_stdout(trace.evlist, err, bf, sizeof(bf));
4968                 pr_err("ERROR: Setup BPF stdout failed: %s\n", bf);
4969                 goto out;
4970         }
4971
4972         err = -1;
4973
4974         if (map_dump_str) {
4975                 trace.dump.map = trace__find_bpf_map_by_name(&trace, map_dump_str);
4976                 if (trace.dump.map == NULL) {
4977                         pr_err("ERROR: BPF map \"%s\" not found\n", map_dump_str);
4978                         goto out;
4979                 }
4980         }
4981
4982         if (trace.trace_pgfaults) {
4983                 trace.opts.sample_address = true;
4984                 trace.opts.sample_time = true;
4985         }
4986
4987         if (trace.opts.mmap_pages == UINT_MAX)
4988                 mmap_pages_user_set = false;
4989
4990         if (trace.max_stack == UINT_MAX) {
4991                 trace.max_stack = input_name ? PERF_MAX_STACK_DEPTH : sysctl__max_stack();
4992                 max_stack_user_set = false;
4993         }
4994
4995 #ifdef HAVE_DWARF_UNWIND_SUPPORT
4996         if ((trace.min_stack || max_stack_user_set) && !callchain_param.enabled) {
4997                 record_opts__parse_callchain(&trace.opts, &callchain_param, "dwarf", false);
4998         }
4999 #endif
5000
5001         if (callchain_param.enabled) {
5002                 if (!mmap_pages_user_set && geteuid() == 0)
5003                         trace.opts.mmap_pages = perf_event_mlock_kb_in_pages() * 4;
5004
5005                 symbol_conf.use_callchain = true;
5006         }
5007
5008         if (trace.evlist->core.nr_entries > 0) {
5009                 evlist__set_default_evsel_handler(trace.evlist, trace__event_handler);
5010                 if (evlist__set_syscall_tp_fields(trace.evlist)) {
5011                         perror("failed to set syscalls:* tracepoint fields");
5012                         goto out;
5013                 }
5014         }
5015
5016         if (trace.sort_events) {
5017                 ordered_events__init(&trace.oe.data, ordered_events__deliver_event, &trace);
5018                 ordered_events__set_copy_on_queue(&trace.oe.data, true);
5019         }
5020
5021         /*
5022          * If we are augmenting syscalls, then combine what we put in the
5023          * __augmented_syscalls__ BPF map with what is in the
5024          * syscalls:sys_exit_FOO tracepoints, i.e. just like we do without BPF,
5025          * combining raw_syscalls:sys_enter with raw_syscalls:sys_exit.
5026          *
5027          * We'll switch to look at two BPF maps, one for sys_enter and the
5028          * other for sys_exit when we start augmenting the sys_exit paths with
5029          * buffers that are being copied from kernel to userspace, think 'read'
5030          * syscall.
5031          */
5032         if (trace.syscalls.events.augmented) {
5033                 evlist__for_each_entry(trace.evlist, evsel) {
5034                         bool raw_syscalls_sys_exit = strcmp(evsel__name(evsel), "raw_syscalls:sys_exit") == 0;
5035
5036                         if (raw_syscalls_sys_exit) {
5037                                 trace.raw_augmented_syscalls = true;
5038                                 goto init_augmented_syscall_tp;
5039                         }
5040
5041                         if (trace.syscalls.events.augmented->priv == NULL &&
5042                             strstr(evsel__name(evsel), "syscalls:sys_enter")) {
5043                                 struct evsel *augmented = trace.syscalls.events.augmented;
5044                                 if (evsel__init_augmented_syscall_tp(augmented, evsel) ||
5045                                     evsel__init_augmented_syscall_tp_args(augmented))
5046                                         goto out;
5047                                 /*
5048                                  * Augmented is __augmented_syscalls__ BPF_OUTPUT event
5049                                  * Above we made sure we can get from the payload the tp fields
5050                                  * that we get from syscalls:sys_enter tracefs format file.
5051                                  */
5052                                 augmented->handler = trace__sys_enter;
5053                                 /*
5054                                  * Now we do the same for the *syscalls:sys_enter event so that
5055                                  * if we handle it directly, i.e. if the BPF prog returns 0 so
5056                                  * as not to filter it, then we'll handle it just like we would
5057                                  * for the BPF_OUTPUT one:
5058                                  */
5059                                 if (evsel__init_augmented_syscall_tp(evsel, evsel) ||
5060                                     evsel__init_augmented_syscall_tp_args(evsel))
5061                                         goto out;
5062                                 evsel->handler = trace__sys_enter;
5063                         }
5064
5065                         if (strstarts(evsel__name(evsel), "syscalls:sys_exit_")) {
5066                                 struct syscall_tp *sc;
5067 init_augmented_syscall_tp:
5068                                 if (evsel__init_augmented_syscall_tp(evsel, evsel))
5069                                         goto out;
5070                                 sc = __evsel__syscall_tp(evsel);
5071                                 /*
5072                                  * For now with BPF raw_augmented we hook into
5073                                  * raw_syscalls:sys_enter and there we get all
5074                                  * 6 syscall args plus the tracepoint common
5075                                  * fields and the syscall_nr (another long).
5076                                  * So we check if that is the case and if so
5077                                  * don't look after the sc->args_size but
5078                                  * always after the full raw_syscalls:sys_enter
5079                                  * payload, which is fixed.
5080                                  *
5081                                  * We'll revisit this later to pass
5082                                  * s->args_size to the BPF augmenter (now
5083                                  * tools/perf/examples/bpf/augmented_raw_syscalls.c,
5084                                  * so that it copies only what we need for each
5085                                  * syscall, like what happens when we use
5086                                  * syscalls:sys_enter_NAME, so that we reduce
5087                                  * the kernel/userspace traffic to just what is
5088                                  * needed for each syscall.
5089                                  */
5090                                 if (trace.raw_augmented_syscalls)
5091                                         trace.raw_augmented_syscalls_args_size = (6 + 1) * sizeof(long) + sc->id.offset;
5092                                 evsel__init_augmented_syscall_tp_ret(evsel);
5093                                 evsel->handler = trace__sys_exit;
5094                         }
5095                 }
5096         }
5097
5098         if ((argc >= 1) && (strcmp(argv[0], "record") == 0))
5099                 return trace__record(&trace, argc-1, &argv[1]);
5100
5101         /* Using just --errno-summary will trigger --summary */
5102         if (trace.errno_summary && !trace.summary && !trace.summary_only)
5103                 trace.summary_only = true;
5104
5105         /* summary_only implies summary option, but don't overwrite summary if set */
5106         if (trace.summary_only)
5107                 trace.summary = trace.summary_only;
5108
5109         if (output_name != NULL) {
5110                 err = trace__open_output(&trace, output_name);
5111                 if (err < 0) {
5112                         perror("failed to create output file");
5113                         goto out;
5114                 }
5115         }
5116
5117         err = evswitch__init(&trace.evswitch, trace.evlist, stderr);
5118         if (err)
5119                 goto out_close;
5120
5121         err = target__validate(&trace.opts.target);
5122         if (err) {
5123                 target__strerror(&trace.opts.target, err, bf, sizeof(bf));
5124                 fprintf(trace.output, "%s", bf);
5125                 goto out_close;
5126         }
5127
5128         err = target__parse_uid(&trace.opts.target);
5129         if (err) {
5130                 target__strerror(&trace.opts.target, err, bf, sizeof(bf));
5131                 fprintf(trace.output, "%s", bf);
5132                 goto out_close;
5133         }
5134
5135         if (!argc && target__none(&trace.opts.target))
5136                 trace.opts.target.system_wide = true;
5137
5138         if (input_name)
5139                 err = trace__replay(&trace);
5140         else
5141                 err = trace__run(&trace, argc, argv);
5142
5143 out_close:
5144         if (output_name != NULL)
5145                 fclose(trace.output);
5146 out:
5147         trace__exit(&trace);
5148         return err;
5149 }