granicus.if.org Git - strace/blob - strace.c

   1 /*
   2  * Copyright (c) 1991, 1992 Paul Kranenburg <pk@cs.few.eur.nl>
   3  * Copyright (c) 1993 Branko Lankester <branko@hacktic.nl>
   4  * Copyright (c) 1993, 1994, 1995, 1996 Rick Sladkey <jrs@world.std.com>
   5  * Copyright (c) 1996-1999 Wichert Akkerman <wichert@cistron.nl>
   6  * All rights reserved.
   7  *
   8  * Redistribution and use in source and binary forms, with or without
   9  * modification, are permitted provided that the following conditions
  10  * are met:
  11  * 1. Redistributions of source code must retain the above copyright
  12  *    notice, this list of conditions and the following disclaimer.
  13  * 2. Redistributions in binary form must reproduce the above copyright
  14  *    notice, this list of conditions and the following disclaimer in the
  15  *    documentation and/or other materials provided with the distribution.
  16  * 3. The name of the author may not be used to endorse or promote products
  17  *    derived from this software without specific prior written permission.
  18  *
  19  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
  20  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
  21  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
  22  * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
  23  * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
  24  * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
  25  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
  26  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
  27  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
  28  * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  29  */
  30
  31 #include "defs.h"
  32 #include <stdarg.h>
  33 #include <sys/param.h>
  34 #include <fcntl.h>
  35 #include <sys/resource.h>
  36 #include <sys/wait.h>
  37 #include <sys/stat.h>
  38 #include <pwd.h>
  39 #include <grp.h>
  40 #include <dirent.h>
  41 #include <sys/utsname.h>
  42 #ifdef HAVE_PRCTL
  43 # include <sys/prctl.h>
  44 #endif
  45 #if defined(IA64)
  46 # include <asm/ptrace_offsets.h>
  47 #endif
  48 /* In some libc, these aren't declared. Do it ourself: */
  49 extern char **environ;
  50 extern int optind;
  51 extern char *optarg;
  52
  53
  54 #if defined __NR_tkill
  55 # define my_tkill(tid, sig) syscall(__NR_tkill, (tid), (sig))
  56 #else
  57    /* kill() may choose arbitrarily the target task of the process group
  58       while we later wait on a that specific TID.  PID process waits become
  59       TID task specific waits for a process under ptrace(2).  */
  60 # warning "tkill(2) not available, risk of strace hangs!"
  61 # define my_tkill(tid, sig) kill((tid), (sig))
  62 #endif
  63
  64 /* Glue for systems without a MMU that cannot provide fork() */
  65 #if !defined(HAVE_FORK)
  66 # undef NOMMU_SYSTEM
  67 # define NOMMU_SYSTEM 1
  68 #endif
  69 #if NOMMU_SYSTEM
  70 # define fork() vfork()
  71 #endif
  72
  73 cflag_t cflag = CFLAG_NONE;
  74 unsigned int followfork = 0;
  75 unsigned int ptrace_setoptions = 0;
  76 unsigned int xflag = 0;
  77 bool need_fork_exec_workarounds = 0;
  78 bool debug_flag = 0;
  79 bool Tflag = 0;
  80 unsigned int qflag = 0;
  81 /* Which WSTOPSIG(status) value marks syscall traps? */
  82 static unsigned int syscall_trap_sig = SIGTRAP;
  83 static unsigned int tflag = 0;
  84 static bool iflag = 0;
  85 static bool rflag = 0;
  86 static bool print_pid_pfx = 0;
  87
  88 /* -I n */
  89 enum {
  90     INTR_NOT_SET        = 0,
  91     INTR_ANYWHERE       = 1, /* don't block/ignore any signals */
  92     INTR_WHILE_WAIT     = 2, /* block fatal signals while decoding syscall. default */
  93     INTR_NEVER          = 3, /* block fatal signals. default if '-o FILE PROG' */
  94     INTR_BLOCK_TSTP_TOO = 4, /* block fatal signals and SIGTSTP (^Z) */
  95     NUM_INTR_OPTS
  96 };
  97 static int opt_intr;
  98 /* We play with signal mask only if this mode is active: */
  99 #define interactive (opt_intr == INTR_WHILE_WAIT)
 100
 101 /*
 102  * daemonized_tracer supports -D option.
 103  * With this option, strace forks twice.
 104  * Unlike normal case, with -D *grandparent* process exec's,
 105  * becoming a traced process. Child exits (this prevents traced process
 106  * from having children it doesn't expect to have), and grandchild
 107  * attaches to grandparent similarly to strace -p PID.
 108  * This allows for more transparent interaction in cases
 109  * when process and its parent are communicating via signals,
 110  * wait() etc. Without -D, strace process gets lodged in between,
 111  * disrupting parent<->child link.
 112  */
 113 static bool daemonized_tracer = 0;
 114
 115 #if USE_SEIZE
 116 static int post_attach_sigstop = TCB_IGNORE_ONE_SIGSTOP;
 117 # define use_seize (post_attach_sigstop == 0)
 118 #else
 119 # define post_attach_sigstop TCB_IGNORE_ONE_SIGSTOP
 120 # define use_seize 0
 121 #endif
 122
 123 /* Sometimes we want to print only succeeding syscalls. */
 124 bool not_failing_only = 0;
 125
 126 /* Show path associated with fd arguments */
 127 bool show_fd_path = 0;
 128
 129 static bool detach_on_execve = 0;
 130 /* Are we "strace PROG" and need to skip detach on first execve? */
 131 static bool skip_one_b_execve = 0;
 132 /* Are we "strace PROG" and need to hide everything until execve? */
 133 bool hide_log_until_execve = 0;
 134
 135 static int exit_code = 0;
 136 static int strace_child = 0;
 137 static int strace_tracer_pid = 0;
 138
 139 static char *username = NULL;
 140 static uid_t run_uid;
 141 static gid_t run_gid;
 142
 143 unsigned int max_strlen = DEFAULT_STRLEN;
 144 static int acolumn = DEFAULT_ACOLUMN;
 145 static char *acolumn_spaces;
 146
 147 static char *outfname = NULL;
 148 /* If -ff, points to stderr. Else, it's our common output log */
 149 static FILE *shared_log;
 150
 151 struct tcb *printing_tcp = NULL;
 152 static struct tcb *current_tcp;
 153
 154 static struct tcb **tcbtab;
 155 static unsigned int nprocs, tcbtabsize;
 156 static const char *progname;
 157
 158 unsigned os_release; /* generated from uname()'s u.release */
 159
 160 static int detach(struct tcb *tcp);
 161 static int trace(void);
 162 static void cleanup(void);
 163 static void interrupt(int sig);
 164 static sigset_t empty_set, blocked_set;
 165
 166 #ifdef HAVE_SIG_ATOMIC_T
 167 static volatile sig_atomic_t interrupted;
 168 #else
 169 static volatile int interrupted;
 170 #endif
 171
 172 #ifndef HAVE_STRERROR
 173
 174 #if !HAVE_DECL_SYS_ERRLIST
 175 extern int sys_nerr;
 176 extern char *sys_errlist[];
 177 #endif
 178
 179 const char *
 180 strerror(int err_no)
 181 {
 182         static char buf[sizeof("Unknown error %d") + sizeof(int)*3];
 183
 184         if (err_no < 1 || err_no >= sys_nerr) {
 185                 sprintf(buf, "Unknown error %d", err_no);
 186                 return buf;
 187         }
 188         return sys_errlist[err_no];
 189 }
 190
 191 #endif /* HAVE_STERRROR */
 192
 193 static void
 194 usage(FILE *ofp, int exitval)
 195 {
 196         fprintf(ofp, "\
 197 usage: strace [-CdffhiqrtttTvVxxy] [-I n] [-e expr]...\n\
 198               [-a column] [-o file] [-s strsize] [-P path]...\n\
 199               -p pid... / [-D] [-E var=val]... [-u username] PROG [ARGS]\n\
 200    or: strace -c[df] [-I n] [-e expr]... [-O overhead] [-S sortby]\n\
 201               -p pid... / [-D] [-E var=val]... [-u username] PROG [ARGS]\n\
 202 -c -- count time, calls, and errors for each syscall and report summary\n\
 203 -C -- like -c but also print regular output\n\
 204 -d -- enable debug output to stderr\n\
 205 -D -- run tracer process as a detached grandchild, not as parent\n\
 206 -f -- follow forks, -ff -- with output into separate files\n\
 207 -i -- print instruction pointer at time of syscall\n\
 208 -q -- suppress messages about attaching, detaching, etc.\n\
 209 -r -- print relative timestamp, -t -- absolute timestamp, -tt -- with usecs\n\
 210 -T -- print time spent in each syscall\n\
 211 -v -- verbose mode: print unabbreviated argv, stat, termios, etc. args\n\
 212 -x -- print non-ascii strings in hex, -xx -- print all strings in hex\n\
 213 -y -- print paths associated with file descriptor arguments\n\
 214 -h -- print help message, -V -- print version\n\
 215 -a column -- alignment COLUMN for printing syscall results (default %d)\n\
 216 -b execve -- detach on this syscall\n\
 217 -e expr -- a qualifying expression: option=[!]all or option=[!]val1[,val2]...\n\
 218    options: trace, abbrev, verbose, raw, signal, read, write\n\
 219 -I interruptible --\n\
 220    1: no signals are blocked\n\
 221    2: fatal signals are blocked while decoding syscall (default)\n\
 222    3: fatal signals are always blocked (default if '-o FILE PROG')\n\
 223    4: fatal signals and SIGTSTP (^Z) are always blocked\n\
 224       (useful to make 'strace -o FILE PROG' not stop on ^Z)\n\
 225 -o file -- send trace output to FILE instead of stderr\n\
 226 -O overhead -- set overhead for tracing syscalls to OVERHEAD usecs\n\
 227 -p pid -- trace process with process id PID, may be repeated\n\
 228 -s strsize -- limit length of print strings to STRSIZE chars (default %d)\n\
 229 -S sortby -- sort syscall counts by: time, calls, name, nothing (default %s)\n\
 230 -u username -- run command as username handling setuid and/or setgid\n\
 231 -E var=val -- put var=val in the environment for command\n\
 232 -E var -- remove var from the environment for command\n\
 233 -P path -- trace accesses to path\n\
 234 "
 235 /* ancient, no one should use it
 236 -F -- attempt to follow vforks (deprecated, use -f)\n\
 237  */
 238 /* this is broken, so don't document it
 239 -z -- print only succeeding syscalls\n\
 240  */
 241 , DEFAULT_ACOLUMN, DEFAULT_STRLEN, DEFAULT_SORTBY);
 242         exit(exitval);
 243 }
 244
 245 static void die(void) __attribute__ ((noreturn));
 246 static void die(void)
 247 {
 248         if (strace_tracer_pid == getpid()) {
 249                 cflag = 0;
 250                 cleanup();
 251         }
 252         exit(1);
 253 }
 254
 255 static void verror_msg(int err_no, const char *fmt, va_list p)
 256 {
 257         char *msg;
 258
 259         fflush(NULL);
 260
 261         /* We want to print entire message with single fprintf to ensure
 262          * message integrity if stderr is shared with other programs.
 263          * Thus we use vasprintf + single fprintf.
 264          */
 265         msg = NULL;
 266         if (vasprintf(&msg, fmt, p) >= 0) {
 267                 if (err_no)
 268                         fprintf(stderr, "%s: %s: %s\n", progname, msg, strerror(err_no));
 269                 else
 270                         fprintf(stderr, "%s: %s\n", progname, msg);
 271                 free(msg);
 272         } else {
 273                 /* malloc in vasprintf failed, try it without malloc */
 274                 fprintf(stderr, "%s: ", progname);
 275                 vfprintf(stderr, fmt, p);
 276                 if (err_no)
 277                         fprintf(stderr, ": %s\n", strerror(err_no));
 278                 else
 279                         putc('\n', stderr);
 280         }
 281         /* We don't switch stderr to buffered, thus fprintf(stderr)
 282          * always flushes its output and this is not necessary: */
 283         /* fflush(stderr); */
 284 }
 285
 286 void error_msg(const char *fmt, ...)
 287 {
 288         va_list p;
 289         va_start(p, fmt);
 290         verror_msg(0, fmt, p);
 291         va_end(p);
 292 }
 293
 294 void error_msg_and_die(const char *fmt, ...)
 295 {
 296         va_list p;
 297         va_start(p, fmt);
 298         verror_msg(0, fmt, p);
 299         die();
 300 }
 301
 302 void perror_msg(const char *fmt, ...)
 303 {
 304         va_list p;
 305         va_start(p, fmt);
 306         verror_msg(errno, fmt, p);
 307         va_end(p);
 308 }
 309
 310 void perror_msg_and_die(const char *fmt, ...)
 311 {
 312         va_list p;
 313         va_start(p, fmt);
 314         verror_msg(errno, fmt, p);
 315         die();
 316 }
 317
 318 void die_out_of_memory(void)
 319 {
 320         static bool recursed = 0;
 321         if (recursed)
 322                 exit(1);
 323         recursed = 1;
 324         error_msg_and_die("Out of memory");
 325 }
 326
 327 static void
 328 error_opt_arg(int opt, const char *arg)
 329 {
 330         error_msg_and_die("Invalid -%c argument: '%s'", opt, arg);
 331 }
 332
 333 #if USE_SEIZE
 334 static int
 335 ptrace_attach_or_seize(int pid)
 336 {
 337         int r;
 338         if (!use_seize)
 339                 return ptrace(PTRACE_ATTACH, pid, 0, 0);
 340         r = ptrace(PTRACE_SEIZE, pid, 0, 0);
 341         if (r)
 342                 return r;
 343         r = ptrace(PTRACE_INTERRUPT, pid, 0, 0);
 344         return r;
 345 }
 346 #else
 347 # define ptrace_attach_or_seize(pid) ptrace(PTRACE_ATTACH, (pid), 0, 0)
 348 #endif
 349
 350 /*
 351  * Used when we want to unblock stopped traced process.
 352  * Should be only used with PTRACE_CONT, PTRACE_DETACH and PTRACE_SYSCALL.
 353  * Returns 0 on success or if error was ESRCH
 354  * (presumably process was killed while we talk to it).
 355  * Otherwise prints error message and returns -1.
 356  */
 357 static int
 358 ptrace_restart(int op, struct tcb *tcp, int sig)
 359 {
 360         int err;
 361         const char *msg;
 362
 363         errno = 0;
 364         ptrace(op, tcp->pid, (void *) 0, (long) sig);
 365         err = errno;
 366         if (!err)
 367                 return 0;
 368
 369         msg = "SYSCALL";
 370         if (op == PTRACE_CONT)
 371                 msg = "CONT";
 372         if (op == PTRACE_DETACH)
 373                 msg = "DETACH";
 374 #ifdef PTRACE_LISTEN
 375         if (op == PTRACE_LISTEN)
 376                 msg = "LISTEN";
 377 #endif
 378         /*
 379          * Why curcol != 0? Otherwise sometimes we get this:
 380          *
 381          * 10252 kill(10253, SIGKILL)              = 0
 382          *  <ptrace(SYSCALL,10252):No such process>10253 ...next decode...
 383          *
 384          * 10252 died after we retrieved syscall exit data,
 385          * but before we tried to restart it. Log looks ugly.
 386          */
 387         if (current_tcp && current_tcp->curcol != 0) {
 388                 tprintf(" <ptrace(%s):%s>\n", msg, strerror(err));
 389                 line_ended();
 390         }
 391         if (err == ESRCH)
 392                 return 0;
 393         errno = err;
 394         perror_msg("ptrace(PTRACE_%s,pid:%d,sig:%d)", msg, tcp->pid, sig);
 395         return -1;
 396 }
 397
 398 static void
 399 set_cloexec_flag(int fd)
 400 {
 401         int flags, newflags;
 402
 403         flags = fcntl(fd, F_GETFD);
 404         if (flags < 0) {
 405                 /* Can happen only if fd is bad.
 406                  * Should never happen: if it does, we have a bug
 407                  * in the caller. Therefore we just abort
 408                  * instead of propagating the error.
 409                  */
 410                 perror_msg_and_die("fcntl(%d, F_GETFD)", fd);
 411         }
 412
 413         newflags = flags | FD_CLOEXEC;
 414         if (flags == newflags)
 415                 return;
 416
 417         fcntl(fd, F_SETFD, newflags); /* never fails */
 418 }
 419
 420 static void kill_save_errno(pid_t pid, int sig)
 421 {
 422         int saved_errno = errno;
 423
 424         (void) kill(pid, sig);
 425         errno = saved_errno;
 426 }
 427
 428 /*
 429  * When strace is setuid executable, we have to swap uids
 430  * before and after filesystem and process management operations.
 431  */
 432 static void
 433 swap_uid(void)
 434 {
 435         int euid = geteuid(), uid = getuid();
 436
 437         if (euid != uid && setreuid(euid, uid) < 0) {
 438                 perror_msg_and_die("setreuid");
 439         }
 440 }
 441
 442 #if _LFS64_LARGEFILE
 443 # define fopen_for_output fopen64
 444 # define struct_stat struct stat64
 445 # define stat_file stat64
 446 # define struct_dirent struct dirent64
 447 # define read_dir readdir64
 448 # define struct_rlimit struct rlimit64
 449 # define set_rlimit setrlimit64
 450 #else
 451 # define fopen_for_output fopen
 452 # define struct_stat struct stat
 453 # define stat_file stat
 454 # define struct_dirent struct dirent
 455 # define read_dir readdir
 456 # define struct_rlimit struct rlimit
 457 # define set_rlimit setrlimit
 458 #endif
 459
 460 static FILE *
 461 strace_fopen(const char *path)
 462 {
 463         FILE *fp;
 464
 465         swap_uid();
 466         fp = fopen_for_output(path, "w");
 467         if (!fp)
 468                 perror_msg_and_die("Can't fopen '%s'", path);
 469         swap_uid();
 470         set_cloexec_flag(fileno(fp));
 471         return fp;
 472 }
 473
 474 static int popen_pid = 0;
 475
 476 #ifndef _PATH_BSHELL
 477 # define _PATH_BSHELL "/bin/sh"
 478 #endif
 479
 480 /*
 481  * We cannot use standard popen(3) here because we have to distinguish
 482  * popen child process from other processes we trace, and standard popen(3)
 483  * does not export its child's pid.
 484  */
 485 static FILE *
 486 strace_popen(const char *command)
 487 {
 488         FILE *fp;
 489         int fds[2];
 490
 491         swap_uid();
 492         if (pipe(fds) < 0)
 493                 perror_msg_and_die("pipe");
 494
 495         set_cloexec_flag(fds[1]); /* never fails */
 496
 497         popen_pid = vfork();
 498         if (popen_pid == -1)
 499                 perror_msg_and_die("vfork");
 500
 501         if (popen_pid == 0) {
 502                 /* child */
 503                 close(fds[1]);
 504                 if (fds[0] != 0) {
 505                         if (dup2(fds[0], 0))
 506                                 perror_msg_and_die("dup2");
 507                         close(fds[0]);
 508                 }
 509                 execl(_PATH_BSHELL, "sh", "-c", command, NULL);
 510                 perror_msg_and_die("Can't execute '%s'", _PATH_BSHELL);
 511         }
 512
 513         /* parent */
 514         close(fds[0]);
 515         swap_uid();
 516         fp = fdopen(fds[1], "w");
 517         if (!fp)
 518                 die_out_of_memory();
 519         return fp;
 520 }
 521
 522 void
 523 tprintf(const char *fmt, ...)
 524 {
 525         va_list args;
 526
 527         va_start(args, fmt);
 528         if (current_tcp) {
 529                 int n = strace_vfprintf(current_tcp->outf, fmt, args);
 530                 if (n < 0) {
 531                         if (current_tcp->outf != stderr)
 532                                 perror_msg("%s", outfname);
 533                 } else
 534                         current_tcp->curcol += n;
 535         }
 536         va_end(args);
 537 }
 538
 539 void
 540 tprints(const char *str)
 541 {
 542         if (current_tcp) {
 543                 int n = fputs_unlocked(str, current_tcp->outf);
 544                 if (n >= 0) {
 545                         current_tcp->curcol += strlen(str);
 546                         return;
 547                 }
 548                 if (current_tcp->outf != stderr)
 549                         perror_msg("%s", outfname);
 550         }
 551 }
 552
 553 void
 554 line_ended(void)
 555 {
 556         if (current_tcp) {
 557                 current_tcp->curcol = 0;
 558                 fflush(current_tcp->outf);
 559         }
 560         if (printing_tcp) {
 561                 printing_tcp->curcol = 0;
 562                 printing_tcp = NULL;
 563         }
 564 }
 565
 566 void
 567 printleader(struct tcb *tcp)
 568 {
 569         /* If -ff, "previous tcb we printed" is always the same as current,
 570          * because we have per-tcb output files.
 571          */
 572         if (followfork >= 2)
 573                 printing_tcp = tcp;
 574
 575         if (printing_tcp) {
 576                 current_tcp = printing_tcp;
 577                 if (printing_tcp->curcol != 0 && (followfork < 2 || printing_tcp == tcp)) {
 578                         /*
 579                          * case 1: we have a shared log (i.e. not -ff), and last line
 580                          * wasn't finished (same or different tcb, doesn't matter).
 581                          * case 2: split log, we are the same tcb, but our last line
 582                          * didn't finish ("SIGKILL nuked us after syscall entry" etc).
 583                          */
 584                         tprints(" <unfinished ...>\n");
 585                         printing_tcp->curcol = 0;
 586                 }
 587         }
 588
 589         printing_tcp = tcp;
 590         current_tcp = tcp;
 591         current_tcp->curcol = 0;
 592
 593         if (print_pid_pfx)
 594                 tprintf("%-5d ", tcp->pid);
 595         else if (nprocs > 1 && !outfname)
 596                 tprintf("[pid %5u] ", tcp->pid);
 597
 598         if (tflag) {
 599                 char str[sizeof("HH:MM:SS")];
 600                 struct timeval tv, dtv;
 601                 static struct timeval otv;
 602
 603                 gettimeofday(&tv, NULL);
 604                 if (rflag) {
 605                         if (otv.tv_sec == 0)
 606                                 otv = tv;
 607                         tv_sub(&dtv, &tv, &otv);
 608                         tprintf("%6ld.%06ld ",
 609                                 (long) dtv.tv_sec, (long) dtv.tv_usec);
 610                         otv = tv;
 611                 }
 612                 else if (tflag > 2) {
 613                         tprintf("%ld.%06ld ",
 614                                 (long) tv.tv_sec, (long) tv.tv_usec);
 615                 }
 616                 else {
 617                         time_t local = tv.tv_sec;
 618                         strftime(str, sizeof(str), "%T", localtime(&local));
 619                         if (tflag > 1)
 620                                 tprintf("%s.%06ld ", str, (long) tv.tv_usec);
 621                         else
 622                                 tprintf("%s ", str);
 623                 }
 624         }
 625         if (iflag)
 626                 printcall(tcp);
 627 }
 628
 629 void
 630 tabto(void)
 631 {
 632         if (current_tcp->curcol < acolumn)
 633                 tprints(acolumn_spaces + current_tcp->curcol);
 634 }
 635
 636 /* Should be only called directly *after successful attach* to a tracee.
 637  * Otherwise, "strace -oFILE -ff -p<nonexistant_pid>"
 638  * may create bogus empty FILE.<nonexistant_pid>, and then die.
 639  */
 640 static void
 641 newoutf(struct tcb *tcp)
 642 {
 643         tcp->outf = shared_log; /* if not -ff mode, the same file is for all */
 644         if (followfork >= 2) {
 645                 char name[520 + sizeof(int) * 3];
 646                 sprintf(name, "%.512s.%u", outfname, tcp->pid);
 647                 tcp->outf = strace_fopen(name);
 648         }
 649 }
 650
 651 static void
 652 expand_tcbtab(void)
 653 {
 654         /* Allocate some more TCBs and expand the table.
 655            We don't want to relocate the TCBs because our
 656            callers have pointers and it would be a pain.
 657            So tcbtab is a table of pointers.  Since we never
 658            free the TCBs, we allocate a single chunk of many.  */
 659         int i = tcbtabsize;
 660         struct tcb *newtcbs = calloc(tcbtabsize, sizeof(newtcbs[0]));
 661         struct tcb **newtab = realloc(tcbtab, tcbtabsize * 2 * sizeof(tcbtab[0]));
 662         if (!newtab || !newtcbs)
 663                 die_out_of_memory();
 664         tcbtabsize *= 2;
 665         tcbtab = newtab;
 666         while (i < tcbtabsize)
 667                 tcbtab[i++] = newtcbs++;
 668 }
 669
 670 static struct tcb *
 671 alloctcb(int pid)
 672 {
 673         int i;
 674         struct tcb *tcp;
 675
 676         if (nprocs == tcbtabsize)
 677                 expand_tcbtab();
 678
 679         for (i = 0; i < tcbtabsize; i++) {
 680                 tcp = tcbtab[i];
 681                 if ((tcp->flags & TCB_INUSE) == 0) {
 682                         memset(tcp, 0, sizeof(*tcp));
 683                         tcp->pid = pid;
 684                         tcp->flags = TCB_INUSE;
 685 #if SUPPORTED_PERSONALITIES > 1
 686                         tcp->currpers = current_personality;
 687 #endif
 688                         nprocs++;
 689                         if (debug_flag)
 690                                 fprintf(stderr, "new tcb for pid %d, active tcbs:%d\n", tcp->pid, nprocs);
 691                         return tcp;
 692                 }
 693         }
 694         error_msg_and_die("bug in alloctcb");
 695 }
 696
 697 static void
 698 droptcb(struct tcb *tcp)
 699 {
 700         if (tcp->pid == 0)
 701                 return;
 702
 703         nprocs--;
 704         if (debug_flag)
 705                 fprintf(stderr, "dropped tcb for pid %d, %d remain\n", tcp->pid, nprocs);
 706
 707         if (tcp->outf) {
 708                 if (followfork >= 2) {
 709                         if (tcp->curcol != 0)
 710                                 fprintf(tcp->outf, " <detached ...>\n");
 711                         fclose(tcp->outf);
 712                 } else {
 713                         if (printing_tcp == tcp && tcp->curcol != 0)
 714                                 fprintf(tcp->outf, " <detached ...>\n");
 715                         fflush(tcp->outf);
 716                 }
 717         }
 718
 719         if (current_tcp == tcp)
 720                 current_tcp = NULL;
 721         if (printing_tcp == tcp)
 722                 printing_tcp = NULL;
 723
 724         memset(tcp, 0, sizeof(*tcp));
 725 }
 726
 727 /* Detach traced process.
 728  * Never call DETACH twice on the same process as both unattached and
 729  * attached-unstopped processes give the same ESRCH.  For unattached process we
 730  * would SIGSTOP it and wait for its SIGSTOP notification forever.
 731  */
 732 static int
 733 detach(struct tcb *tcp)
 734 {
 735         int error;
 736         int status, sigstop_expected, interrupt_done;
 737
 738         if (tcp->flags & TCB_BPTSET)
 739                 clearbpt(tcp);
 740
 741         /*
 742          * Linux wrongly insists the child be stopped
 743          * before detaching.  Arghh.  We go through hoops
 744          * to make a clean break of things.
 745          */
 746 #if defined(SPARC)
 747 # undef PTRACE_DETACH
 748 # define PTRACE_DETACH PTRACE_SUNDETACH
 749 #endif
 750
 751         error = 0;
 752         sigstop_expected = 0;
 753         interrupt_done = 0;
 754         if (tcp->flags & TCB_ATTACHED) {
 755                 /*
 756                  * We attached but possibly didn't see the expected SIGSTOP.
 757                  * We must catch exactly one as otherwise the detached process
 758                  * would be left stopped (process state T).
 759                  */
 760                 sigstop_expected = (tcp->flags & TCB_IGNORE_ONE_SIGSTOP);
 761                 error = ptrace(PTRACE_DETACH, tcp->pid, 0, 0);
 762                 if (error == 0) {
 763                         /* On a clear day, you can see forever. */
 764                 }
 765                 else if (errno != ESRCH) {
 766                         /* Shouldn't happen. */
 767                         perror_msg("detach: ptrace(PTRACE_DETACH, ...)");
 768                 }
 769                 else
 770                 /* ESRCH: process is either not stopped or doesn't exist. */
 771                 if (my_tkill(tcp->pid, 0) < 0) {
 772                         if (errno != ESRCH)
 773                                 /* Shouldn't happen. */
 774                                 perror_msg("detach: checking sanity");
 775                         /* else: process doesn't exist. */
 776                 }
 777                 else
 778                 /* Process is not stopped. */
 779                 if (!sigstop_expected) {
 780                         /* We need to stop it. */
 781                         if (use_seize) {
 782                                 /*
 783                                  * With SEIZE, tracee can be in group-stop already.
 784                                  * In this state sending it another SIGSTOP does nothing.
 785                                  * Need to use INTERRUPT.
 786                                  * Testcase: trying to ^C a "strace -p <stopped_process>".
 787                                  */
 788                                 error = ptrace(PTRACE_INTERRUPT, tcp->pid, 0, 0);
 789                                 if (!error)
 790                                         interrupt_done = 1;
 791                         }
 792                         else {
 793                                 error = my_tkill(tcp->pid, SIGSTOP);
 794                                 if (!error)
 795                                         sigstop_expected = 1;
 796                         }
 797                         if (error && errno != ESRCH) {
 798                                 if (use_seize)
 799                                         perror_msg("detach: ptrace(PTRACE_INTERRUPT, ...)");
 800                                 else
 801                                         perror_msg("detach: stopping child");
 802                         }
 803                 }
 804         }
 805
 806         if (sigstop_expected || interrupt_done) {
 807                 for (;;) {
 808                         int sig;
 809                         if (waitpid(tcp->pid, &status, __WALL) < 0) {
 810                                 if (errno == EINTR)
 811                                         continue;
 812                                 if (errno == ECHILD) /* Already gone.  */
 813                                         break;
 814                                 perror_msg("detach: waiting");
 815                                 break;
 816                         }
 817                         if (!WIFSTOPPED(status)) {
 818                                 /* Au revoir, mon ami. */
 819                                 break;
 820                         }
 821                         sig = WSTOPSIG(status);
 822                         if (debug_flag)
 823                                 fprintf(stderr, "detach wait: event:%d sig:%d\n",
 824                                                 (unsigned)status >> 16, sig);
 825                         if (sigstop_expected && sig == SIGSTOP) {
 826                                 /* Detach, suppressing SIGSTOP */
 827                                 ptrace_restart(PTRACE_DETACH, tcp, 0);
 828                                 break;
 829                         }
 830                         if (interrupt_done) {
 831                                 unsigned event = (unsigned)status >> 16;
 832                                 if (event == PTRACE_EVENT_STOP /*&& sig == SIGTRAP*/) {
 833                                         /*
 834                                          * sig == SIGTRAP: PTRACE_INTERRUPT stop.
 835                                          * sig == other: process was already stopped
 836                                          * with this stopping sig (see tests/detach-stopped).
 837                                          * Looks like re-injecting this sig is not necessary
 838                                          * in DETACH for the tracee to remain stopped.
 839                                          */
 840                                         sig = 0;
 841                                 }
 842                                 /*
 843                                  * PTRACE_INTERRUPT is not guaranteed to produce
 844                                  * the above event if other ptrace-stop is pending.
 845                                  * See tests/detach-sleeping testcase:
 846                                  * strace got SIGINT while tracee is sleeping.
 847                                  * We sent PTRACE_INTERRUPT.
 848                                  * We see syscall exit, not PTRACE_INTERRUPT stop.
 849                                  * We won't get PTRACE_INTERRUPT stop
 850                                  * if we would CONT now. Need to DETACH.
 851                                  */
 852                                 if (sig == syscall_trap_sig)
 853                                         sig = 0;
 854                                 /* else: not sure in which case we can be here.
 855                                  * Signal stop? Inject it while detaching.
 856                                  */
 857                                 ptrace_restart(PTRACE_DETACH, tcp, sig);
 858                                 break;
 859                         }
 860                         if (sig == syscall_trap_sig)
 861                                 sig = 0;
 862                         /* Can't detach just yet, may need to wait for SIGSTOP */
 863                         error = ptrace_restart(PTRACE_CONT, tcp, sig);
 864                         if (error < 0)
 865                                 break;
 866                 }
 867         }
 868
 869         if (!qflag && (tcp->flags & TCB_ATTACHED))
 870                 fprintf(stderr, "Process %u detached\n", tcp->pid);
 871
 872         droptcb(tcp);
 873
 874         return error;
 875 }
 876
 877 static void
 878 process_opt_p_list(char *opt)
 879 {
 880         while (*opt) {
 881                 /*
 882                  * We accept -p PID,PID; -p "`pidof PROG`"; -p "`pgrep PROG`".
 883                  * pidof uses space as delim, pgrep uses newline. :(
 884                  */
 885                 int pid;
 886                 char *delim = opt + strcspn(opt, ", \n\t");
 887                 char c = *delim;
 888
 889                 *delim = '\0';
 890                 pid = string_to_uint(opt);
 891                 if (pid <= 0) {
 892                         error_msg_and_die("Invalid process id: '%s'", opt);
 893                 }
 894                 if (pid == strace_tracer_pid) {
 895                         error_msg_and_die("I'm sorry, I can't let you do that, Dave.");
 896                 }
 897                 *delim = c;
 898                 alloctcb(pid);
 899                 if (c == '\0')
 900                         break;
 901                 opt = delim + 1;
 902         }
 903 }
 904
 905 static void
 906 startup_attach(void)
 907 {
 908         int tcbi;
 909         struct tcb *tcp;
 910
 911         /*
 912          * Block user interruptions as we would leave the traced
 913          * process stopped (process state T) if we would terminate in
 914          * between PTRACE_ATTACH and wait4() on SIGSTOP.
 915          * We rely on cleanup() from this point on.
 916          */
 917         if (interactive)
 918                 sigprocmask(SIG_BLOCK, &blocked_set, NULL);
 919
 920         if (daemonized_tracer) {
 921                 pid_t pid = fork();
 922                 if (pid < 0) {
 923                         perror_msg_and_die("fork");
 924                 }
 925                 if (pid) { /* parent */
 926                         /*
 927                          * Wait for grandchild to attach to straced process
 928                          * (grandparent). Grandchild SIGKILLs us after it attached.
 929                          * Grandparent's wait() is unblocked by our death,
 930                          * it proceeds to exec the straced program.
 931                          */
 932                         pause();
 933                         _exit(0); /* paranoia */
 934                 }
 935                 /* grandchild */
 936                 /* We will be the tracer process. Remember our new pid: */
 937                 strace_tracer_pid = getpid();
 938         }
 939
 940         for (tcbi = 0; tcbi < tcbtabsize; tcbi++) {
 941                 tcp = tcbtab[tcbi];
 942
 943                 if (!(tcp->flags & TCB_INUSE))
 944                         continue;
 945
 946                 /* Is this a process we should attach to, but not yet attached? */
 947                 if (tcp->flags & TCB_ATTACHED)
 948                         continue; /* no, we already attached it */
 949
 950                 if (followfork && !daemonized_tracer) {
 951                         char procdir[sizeof("/proc/%d/task") + sizeof(int) * 3];
 952                         DIR *dir;
 953
 954                         sprintf(procdir, "/proc/%d/task", tcp->pid);
 955                         dir = opendir(procdir);
 956                         if (dir != NULL) {
 957                                 unsigned int ntid = 0, nerr = 0;
 958                                 struct_dirent *de;
 959
 960                                 while ((de = read_dir(dir)) != NULL) {
 961                                         struct tcb *cur_tcp;
 962                                         int tid;
 963
 964                                         if (de->d_fileno == 0)
 965                                                 continue;
 966                                         /* we trust /proc filesystem */
 967                                         tid = atoi(de->d_name);
 968                                         if (tid <= 0)
 969                                                 continue;
 970                                         ++ntid;
 971                                         if (ptrace_attach_or_seize(tid) < 0) {
 972                                                 ++nerr;
 973                                                 if (debug_flag)
 974                                                         fprintf(stderr, "attach to pid %d failed\n", tid);
 975                                                 continue;
 976                                         }
 977                                         if (debug_flag)
 978                                                 fprintf(stderr, "attach to pid %d succeeded\n", tid);
 979                                         cur_tcp = tcp;
 980                                         if (tid != tcp->pid)
 981                                                 cur_tcp = alloctcb(tid);
 982                                         cur_tcp->flags |= TCB_ATTACHED | TCB_STARTUP | post_attach_sigstop;
 983                                         newoutf(cur_tcp);
 984                                 }
 985                                 closedir(dir);
 986                                 if (interactive) {
 987                                         sigprocmask(SIG_SETMASK, &empty_set, NULL);
 988                                         if (interrupted)
 989                                                 goto ret;
 990                                         sigprocmask(SIG_BLOCK, &blocked_set, NULL);
 991                                 }
 992                                 ntid -= nerr;
 993                                 if (ntid == 0) {
 994                                         perror_msg("attach: ptrace(PTRACE_ATTACH, ...)");
 995                                         droptcb(tcp);
 996                                         continue;
 997                                 }
 998                                 if (!qflag) {
 999                                         fprintf(stderr, ntid > 1
1000 ? "Process %u attached with %u threads\n"
1001 : "Process %u attached\n",
1002                                                 tcp->pid, ntid);
1003                                 }
1004                                 if (!(tcp->flags & TCB_ATTACHED)) {
1005                                         /* -p PID, we failed to attach to PID itself
1006                                          * but did attach to some of its sibling threads.
1007                                          * Drop PID's tcp.
1008                                          */
1009                                         droptcb(tcp);
1010                                 }
1011                                 continue;
1012                         } /* if (opendir worked) */
1013                 } /* if (-f) */
1014                 if (ptrace_attach_or_seize(tcp->pid) < 0) {
1015                         perror_msg("attach: ptrace(PTRACE_ATTACH, ...)");
1016                         droptcb(tcp);
1017                         continue;
1018                 }
1019                 tcp->flags |= TCB_ATTACHED | TCB_STARTUP | post_attach_sigstop;
1020                 newoutf(tcp);
1021                 if (debug_flag)
1022                         fprintf(stderr, "attach to pid %d (main) succeeded\n", tcp->pid);
1023
1024                 if (daemonized_tracer) {
1025                         /*
1026                          * Make parent go away.
1027                          * Also makes grandparent's wait() unblock.
1028                          */
1029                         kill(getppid(), SIGKILL);
1030                 }
1031
1032                 if (!qflag)
1033                         fprintf(stderr,
1034                                 "Process %u attached\n",
1035                                 tcp->pid);
1036         } /* for each tcbtab[] */
1037
1038  ret:
1039         if (interactive)
1040                 sigprocmask(SIG_SETMASK, &empty_set, NULL);
1041 }
1042
1043 /* Stack-o-phobic exec helper, in the hope to work around
1044  * NOMMU + "daemonized tracer" difficulty.
1045  */
1046 struct exec_params {
1047         int fd_to_close;
1048         uid_t run_euid;
1049         gid_t run_egid;
1050         char **argv;
1051         char *pathname;
1052 };
1053 static struct exec_params params_for_tracee;
1054 static void __attribute__ ((noinline, noreturn))
1055 exec_or_die(void)
1056 {
1057         struct exec_params *params = &params_for_tracee;
1058
1059         if (params->fd_to_close >= 0)
1060                 close(params->fd_to_close);
1061         if (!daemonized_tracer && !use_seize) {
1062                 if (ptrace(PTRACE_TRACEME, 0L, 0L, 0L) < 0) {
1063                         perror_msg_and_die("ptrace(PTRACE_TRACEME, ...)");
1064                 }
1065         }
1066
1067         if (username != NULL) {
1068                 /*
1069                  * It is important to set groups before we
1070                  * lose privileges on setuid.
1071                  */
1072                 if (initgroups(username, run_gid) < 0) {
1073                         perror_msg_and_die("initgroups");
1074                 }
1075                 if (setregid(run_gid, params->run_egid) < 0) {
1076                         perror_msg_and_die("setregid");
1077                 }
1078                 if (setreuid(run_uid, params->run_euid) < 0) {
1079                         perror_msg_and_die("setreuid");
1080                 }
1081         }
1082         else if (geteuid() != 0)
1083                 if (setreuid(run_uid, run_uid) < 0) {
1084                         perror_msg_and_die("setreuid");
1085                 }
1086
1087         if (!daemonized_tracer) {
1088                 /*
1089                  * Induce a ptrace stop. Tracer (our parent)
1090                  * will resume us with PTRACE_SYSCALL and display
1091                  * the immediately following execve syscall.
1092                  * Can't do this on NOMMU systems, we are after
1093                  * vfork: parent is blocked, stopping would deadlock.
1094                  */
1095                 if (!NOMMU_SYSTEM)
1096                         kill(getpid(), SIGSTOP);
1097         } else {
1098                 alarm(3);
1099                 /* we depend on SIGCHLD set to SIG_DFL by init code */
1100                 /* if it happens to be SIG_IGN'ed, wait won't block */
1101                 wait(NULL);
1102                 alarm(0);
1103         }
1104
1105         execv(params->pathname, params->argv);
1106         perror_msg_and_die("exec");
1107 }
1108
1109 static void
1110 startup_child(char **argv)
1111 {
1112         struct_stat statbuf;
1113         const char *filename;
1114         char pathname[MAXPATHLEN];
1115         int pid;
1116         struct tcb *tcp;
1117
1118         filename = argv[0];
1119         if (strchr(filename, '/')) {
1120                 if (strlen(filename) > sizeof pathname - 1) {
1121                         errno = ENAMETOOLONG;
1122                         perror_msg_and_die("exec");
1123                 }
1124                 strcpy(pathname, filename);
1125         }
1126 #ifdef USE_DEBUGGING_EXEC
1127         /*
1128          * Debuggers customarily check the current directory
1129          * first regardless of the path but doing that gives
1130          * security geeks a panic attack.
1131          */
1132         else if (stat_file(filename, &statbuf) == 0)
1133                 strcpy(pathname, filename);
1134 #endif /* USE_DEBUGGING_EXEC */
1135         else {
1136                 const char *path;
1137                 int m, n, len;
1138
1139                 for (path = getenv("PATH"); path && *path; path += m) {
1140                         const char *colon = strchr(path, ':');
1141                         if (colon) {
1142                                 n = colon - path;
1143                                 m = n + 1;
1144                         }
1145                         else
1146                                 m = n = strlen(path);
1147                         if (n == 0) {
1148                                 if (!getcwd(pathname, MAXPATHLEN))
1149                                         continue;
1150                                 len = strlen(pathname);
1151                         }
1152                         else if (n > sizeof pathname - 1)
1153                                 continue;
1154                         else {
1155                                 strncpy(pathname, path, n);
1156                                 len = n;
1157                         }
1158                         if (len && pathname[len - 1] != '/')
1159                                 pathname[len++] = '/';
1160                         strcpy(pathname + len, filename);
1161                         if (stat_file(pathname, &statbuf) == 0 &&
1162                             /* Accept only regular files
1163                                with some execute bits set.
1164                                XXX not perfect, might still fail */
1165                             S_ISREG(statbuf.st_mode) &&
1166                             (statbuf.st_mode & 0111))
1167                                 break;
1168                 }
1169         }
1170         if (stat_file(pathname, &statbuf) < 0) {
1171                 perror_msg_and_die("Can't stat '%s'", filename);
1172         }
1173
1174         params_for_tracee.fd_to_close = (shared_log != stderr) ? fileno(shared_log) : -1;
1175         params_for_tracee.run_euid = (statbuf.st_mode & S_ISUID) ? statbuf.st_uid : run_uid;
1176         params_for_tracee.run_egid = (statbuf.st_mode & S_ISGID) ? statbuf.st_gid : run_gid;
1177         params_for_tracee.argv = argv;
1178         /*
1179          * On NOMMU, can be safely freed only after execve in tracee.
1180          * It's hard to know when that happens, so we just leak it.
1181          */
1182         params_for_tracee.pathname = NOMMU_SYSTEM ? strdup(pathname) : pathname;
1183
1184 #if defined HAVE_PRCTL && defined PR_SET_PTRACER && defined PR_SET_PTRACER_ANY
1185         if (daemonized_tracer)
1186                 prctl(PR_SET_PTRACER, PR_SET_PTRACER_ANY);
1187 #endif
1188
1189         strace_child = pid = fork();
1190         if (pid < 0) {
1191                 perror_msg_and_die("fork");
1192         }
1193         if ((pid != 0 && daemonized_tracer)
1194          || (pid == 0 && !daemonized_tracer)
1195         ) {
1196                 /* We are to become the tracee. Two cases:
1197                  * -D: we are parent
1198                  * not -D: we are child
1199                  */
1200                 exec_or_die();
1201         }
1202
1203         /* We are the tracer */
1204
1205         if (!daemonized_tracer) {
1206                 if (!use_seize) {
1207                         /* child did PTRACE_TRACEME, nothing to do in parent */
1208                 } else {
1209                         if (!NOMMU_SYSTEM) {
1210                                 /* Wait until child stopped itself */
1211                                 int status;
1212                                 while (waitpid(pid, &status, WSTOPPED) < 0) {
1213                                         if (errno == EINTR)
1214                                                 continue;
1215                                         perror_msg_and_die("waitpid");
1216                                 }
1217                                 if (!WIFSTOPPED(status) || WSTOPSIG(status) != SIGSTOP) {
1218                                         kill_save_errno(pid, SIGKILL);
1219                                         perror_msg_and_die("Unexpected wait status %x", status);
1220                                 }
1221                         }
1222                         /* Else: NOMMU case, we have no way to sync.
1223                          * Just attach to it as soon as possible.
1224                          * This means that we may miss a few first syscalls...
1225                          */
1226
1227                         if (ptrace_attach_or_seize(pid)) {
1228                                 kill_save_errno(pid, SIGKILL);
1229                                 perror_msg_and_die("Can't attach to %d", pid);
1230                         }
1231                         if (!NOMMU_SYSTEM)
1232                                 kill(pid, SIGCONT);
1233                 }
1234                 tcp = alloctcb(pid);
1235                 if (!NOMMU_SYSTEM)
1236                         tcp->flags |= TCB_ATTACHED | TCB_STRACE_CHILD | TCB_STARTUP | post_attach_sigstop;
1237                 else
1238                         tcp->flags |= TCB_ATTACHED | TCB_STRACE_CHILD | TCB_STARTUP;
1239                 newoutf(tcp);
1240         }
1241         else {
1242                 /* With -D, we are *child* here, IOW: different pid. Fetch it: */
1243                 strace_tracer_pid = getpid();
1244                 /* The tracee is our parent: */
1245                 pid = getppid();
1246                 alloctcb(pid);
1247                 /* attaching will be done later, by startup_attach */
1248                 /* note: we don't do newoutf(tcp) here either! */
1249
1250                 /* NOMMU BUG! -D mode is active, we (child) return,
1251                  * and we will scribble over parent's stack!
1252                  * When parent later unpauses, it segfaults.
1253                  *
1254                  * We work around it
1255                  * (1) by declaring exec_or_die() NORETURN,
1256                  * hopefully compiler will just jump to it
1257                  * instead of call (won't push anything to stack),
1258                  * (2) by trying very hard in exec_or_die()
1259                  * to not use any stack,
1260                  * (3) having a really big (MAXPATHLEN) stack object
1261                  * in this function, which creates a "buffer" between
1262                  * child's and parent's stack pointers.
1263                  * This may save us if (1) and (2) failed
1264                  * and compiler decided to use stack in exec_or_die() anyway
1265                  * (happens on i386 because of stack parameter passing).
1266                  *
1267                  * A cleaner solution is to use makecontext + setcontext
1268                  * to create a genuine separate stack and execute on it.
1269                  */
1270         }
1271 }
1272
1273 /*
1274  * Test whether the kernel support PTRACE_O_TRACECLONE et al options.
1275  * First fork a new child, call ptrace with PTRACE_SETOPTIONS on it,
1276  * and then see which options are supported by the kernel.
1277  */
1278 static int
1279 test_ptrace_setoptions_followfork(void)
1280 {
1281         int pid, expected_grandchild = 0, found_grandchild = 0;
1282         const unsigned int test_options = PTRACE_O_TRACECLONE |
1283                                           PTRACE_O_TRACEFORK |
1284                                           PTRACE_O_TRACEVFORK;
1285
1286         /* Need fork for test. NOMMU has no forks */
1287         if (NOMMU_SYSTEM)
1288                 goto worked; /* be bold, and pretend that test succeeded */
1289
1290         pid = fork();
1291         if (pid < 0)
1292                 perror_msg_and_die("fork");
1293         if (pid == 0) {
1294                 pid = getpid();
1295                 if (ptrace(PTRACE_TRACEME, 0L, 0L, 0L) < 0)
1296                         perror_msg_and_die("%s: PTRACE_TRACEME doesn't work",
1297                                            __func__);
1298                 kill_save_errno(pid, SIGSTOP);
1299                 if (fork() < 0)
1300                         perror_msg_and_die("fork");
1301                 _exit(0);
1302         }
1303
1304         while (1) {
1305                 int status, tracee_pid;
1306
1307                 errno = 0;
1308                 tracee_pid = wait(&status);
1309                 if (tracee_pid <= 0) {
1310                         if (errno == EINTR)
1311                                 continue;
1312                         if (errno == ECHILD)
1313                                 break;
1314                         kill_save_errno(pid, SIGKILL);
1315                         perror_msg_and_die("%s: unexpected wait result %d",
1316                                            __func__, tracee_pid);
1317                 }
1318                 if (WIFEXITED(status)) {
1319                         if (WEXITSTATUS(status)) {
1320                                 if (tracee_pid != pid)
1321                                         kill_save_errno(pid, SIGKILL);
1322                                 error_msg_and_die("%s: unexpected exit status %u",
1323                                                   __func__, WEXITSTATUS(status));
1324                         }
1325                         continue;
1326                 }
1327                 if (WIFSIGNALED(status)) {
1328                         if (tracee_pid != pid)
1329                                 kill_save_errno(pid, SIGKILL);
1330                         error_msg_and_die("%s: unexpected signal %u",
1331                                           __func__, WTERMSIG(status));
1332                 }
1333                 if (!WIFSTOPPED(status)) {
1334                         if (tracee_pid != pid)
1335                                 kill_save_errno(tracee_pid, SIGKILL);
1336                         kill_save_errno(pid, SIGKILL);
1337                         error_msg_and_die("%s: unexpected wait status %x",
1338                                           __func__, status);
1339                 }
1340                 if (tracee_pid != pid) {
1341                         found_grandchild = tracee_pid;
1342                         if (ptrace(PTRACE_CONT, tracee_pid, 0, 0) < 0) {
1343                                 kill_save_errno(tracee_pid, SIGKILL);
1344                                 kill_save_errno(pid, SIGKILL);
1345                                 perror_msg_and_die("PTRACE_CONT doesn't work");
1346                         }
1347                         continue;
1348                 }
1349                 switch (WSTOPSIG(status)) {
1350                 case SIGSTOP:
1351                         if (ptrace(PTRACE_SETOPTIONS, pid, 0, test_options) < 0
1352                             && errno != EINVAL && errno != EIO)
1353                                 perror_msg("PTRACE_SETOPTIONS");
1354                         break;
1355                 case SIGTRAP:
1356                         if (status >> 16 == PTRACE_EVENT_FORK) {
1357                                 long msg = 0;
1358
1359                                 if (ptrace(PTRACE_GETEVENTMSG, pid,
1360                                            NULL, (long) &msg) == 0)
1361                                         expected_grandchild = msg;
1362                         }
1363                         break;
1364                 }
1365                 if (ptrace(PTRACE_SYSCALL, pid, 0, 0) < 0) {
1366                         kill_save_errno(pid, SIGKILL);
1367                         perror_msg_and_die("PTRACE_SYSCALL doesn't work");
1368                 }
1369         }
1370         if (expected_grandchild && expected_grandchild == found_grandchild) {
1371  worked:
1372                 ptrace_setoptions |= test_options;
1373                 if (debug_flag)
1374                         fprintf(stderr, "ptrace_setoptions = %#x\n",
1375                                 ptrace_setoptions);
1376                 return 0;
1377         }
1378         error_msg("Test for PTRACE_O_TRACECLONE failed, "
1379                   "giving up using this feature.");
1380         return 1;
1381 }
1382
1383 /*
1384  * Test whether the kernel support PTRACE_O_TRACESYSGOOD.
1385  * First fork a new child, call ptrace(PTRACE_SETOPTIONS) on it,
1386  * and then see whether it will stop with (SIGTRAP | 0x80).
1387  *
1388  * Use of this option enables correct handling of user-generated SIGTRAPs,
1389  * and SIGTRAPs generated by special instructions such as int3 on x86:
1390  * _start:      .globl  _start
1391  *              int3
1392  *              movl    $42, %ebx
1393  *              movl    $1, %eax
1394  *              int     $0x80
1395  * (compile with: "gcc -nostartfiles -nostdlib -o int3 int3.S")
1396  */
1397 static int
1398 test_ptrace_setoptions_for_all(void)
1399 {
1400         const unsigned int test_options = PTRACE_O_TRACESYSGOOD |
1401                                           PTRACE_O_TRACEEXEC;
1402         int pid;
1403         int it_worked = 0;
1404
1405         /* Need fork for test. NOMMU has no forks */
1406         if (NOMMU_SYSTEM)
1407                 goto worked; /* be bold, and pretend that test succeeded */
1408
1409         pid = fork();
1410         if (pid < 0)
1411                 perror_msg_and_die("fork");
1412
1413         if (pid == 0) {
1414                 pid = getpid();
1415                 if (ptrace(PTRACE_TRACEME, 0L, 0L, 0L) < 0)
1416                         /* Note: exits with exitcode 1 */
1417                         perror_msg_and_die("%s: PTRACE_TRACEME doesn't work",
1418                                            __func__);
1419                 kill(pid, SIGSTOP);
1420                 _exit(0); /* parent should see entry into this syscall */
1421         }
1422
1423         while (1) {
1424                 int status, tracee_pid;
1425
1426                 errno = 0;
1427                 tracee_pid = wait(&status);
1428                 if (tracee_pid <= 0) {
1429                         if (errno == EINTR)
1430                                 continue;
1431                         kill_save_errno(pid, SIGKILL);
1432                         perror_msg_and_die("%s: unexpected wait result %d",
1433                                            __func__, tracee_pid);
1434                 }
1435                 if (WIFEXITED(status)) {
1436                         if (WEXITSTATUS(status) == 0)
1437                                 break;
1438                         error_msg_and_die("%s: unexpected exit status %u",
1439                                           __func__, WEXITSTATUS(status));
1440                 }
1441                 if (WIFSIGNALED(status)) {
1442                         error_msg_and_die("%s: unexpected signal %u",
1443                                           __func__, WTERMSIG(status));
1444                 }
1445                 if (!WIFSTOPPED(status)) {
1446                         kill(pid, SIGKILL);
1447                         error_msg_and_die("%s: unexpected wait status %x",
1448                                           __func__, status);
1449                 }
1450                 if (WSTOPSIG(status) == SIGSTOP) {
1451                         /*
1452                          * We don't check "options aren't accepted" error.
1453                          * If it happens, we'll never get (SIGTRAP | 0x80),
1454                          * and thus will decide to not use the option.
1455                          * IOW: the outcome of the test will be correct.
1456                          */
1457                         if (ptrace(PTRACE_SETOPTIONS, pid, 0L, test_options) < 0
1458                             && errno != EINVAL && errno != EIO)
1459                                 perror_msg("PTRACE_SETOPTIONS");
1460                 }
1461                 if (WSTOPSIG(status) == (SIGTRAP | 0x80)) {
1462                         it_worked = 1;
1463                 }
1464                 if (ptrace(PTRACE_SYSCALL, pid, 0L, 0L) < 0) {
1465                         kill_save_errno(pid, SIGKILL);
1466                         perror_msg_and_die("PTRACE_SYSCALL doesn't work");
1467                 }
1468         }
1469
1470         if (it_worked) {
1471  worked:
1472                 syscall_trap_sig = (SIGTRAP | 0x80);
1473                 ptrace_setoptions |= test_options;
1474                 if (debug_flag)
1475                         fprintf(stderr, "ptrace_setoptions = %#x\n",
1476                                 ptrace_setoptions);
1477                 return 0;
1478         }
1479
1480         error_msg("Test for PTRACE_O_TRACESYSGOOD failed, "
1481                   "giving up using this feature.");
1482         return 1;
1483 }
1484
1485 #if USE_SEIZE
1486 static void
1487 test_ptrace_seize(void)
1488 {
1489         int pid;
1490
1491         /* Need fork for test. NOMMU has no forks */
1492         if (NOMMU_SYSTEM) {
1493                 post_attach_sigstop = 0; /* this sets use_seize to 1 */
1494                 return;
1495         }
1496
1497         pid = fork();
1498         if (pid < 0)
1499                 perror_msg_and_die("fork");
1500
1501         if (pid == 0) {
1502                 pause();
1503                 _exit(0);
1504         }
1505
1506         /* PTRACE_SEIZE, unlike ATTACH, doesn't force tracee to trap.  After
1507          * attaching tracee continues to run unless a trap condition occurs.
1508          * PTRACE_SEIZE doesn't affect signal or group stop state.
1509          */
1510         if (ptrace(PTRACE_SEIZE, pid, 0, 0) == 0) {
1511                 post_attach_sigstop = 0; /* this sets use_seize to 1 */
1512         } else if (debug_flag) {
1513                 fprintf(stderr, "PTRACE_SEIZE doesn't work\n");
1514         }
1515
1516         kill(pid, SIGKILL);
1517
1518         while (1) {
1519                 int status, tracee_pid;
1520
1521                 errno = 0;
1522                 tracee_pid = waitpid(pid, &status, 0);
1523                 if (tracee_pid <= 0) {
1524                         if (errno == EINTR)
1525                                 continue;
1526                         perror_msg_and_die("%s: unexpected wait result %d",
1527                                          __func__, tracee_pid);
1528                 }
1529                 if (WIFSIGNALED(status)) {
1530                         return;
1531                 }
1532                 error_msg_and_die("%s: unexpected wait status %x",
1533                                 __func__, status);
1534         }
1535 }
1536 #else /* !USE_SEIZE */
1537 # define test_ptrace_seize() ((void)0)
1538 #endif
1539
1540 static unsigned
1541 get_os_release(void)
1542 {
1543         unsigned rel;
1544         const char *p;
1545         struct utsname u;
1546         if (uname(&u) < 0)
1547                 perror_msg_and_die("uname");
1548         /* u.release has this form: "3.2.9[-some-garbage]" */
1549         rel = 0;
1550         p = u.release;
1551         for (;;) {
1552                 if (!(*p >= '0' && *p <= '9'))
1553                         error_msg_and_die("Bad OS release string: '%s'", u.release);
1554                 /* Note: this open-codes KERNEL_VERSION(): */
1555                 rel = (rel << 8) | atoi(p);
1556                 if (rel >= KERNEL_VERSION(1,0,0))
1557                         break;
1558                 while (*p >= '0' && *p <= '9')
1559                         p++;
1560                 if (*p != '.') {
1561                         if (rel >= KERNEL_VERSION(0,1,0)) {
1562                                 /* "X.Y-something" means "X.Y.0" */
1563                                 rel <<= 8;
1564                                 break;
1565                         }
1566                         error_msg_and_die("Bad OS release string: '%s'", u.release);
1567                 }
1568                 p++;
1569         }
1570         return rel;
1571 }
1572
1573 /*
1574  * Initialization part of main() was eating much stack (~0.5k),
1575  * which was unused after init.
1576  * We can reuse it if we move init code into a separate function.
1577  *
1578  * Don't want main() to inline us and defeat the reason
1579  * we have a separate function.
1580  */
1581 static void __attribute__ ((noinline))
1582 init(int argc, char *argv[])
1583 {
1584         struct tcb *tcp;
1585         int c, i;
1586         int optF = 0;
1587         struct sigaction sa;
1588
1589         progname = argv[0] ? argv[0] : "strace";
1590
1591         /* Make sure SIGCHLD has the default action so that waitpid
1592            definitely works without losing track of children.  The user
1593            should not have given us a bogus state to inherit, but he might
1594            have.  Arguably we should detect SIG_IGN here and pass it on
1595            to children, but probably noone really needs that.  */
1596         signal(SIGCHLD, SIG_DFL);
1597
1598         strace_tracer_pid = getpid();
1599
1600         os_release = get_os_release();
1601
1602         /* Allocate the initial tcbtab.  */
1603         tcbtabsize = argc;      /* Surely enough for all -p args.  */
1604         tcbtab = calloc(tcbtabsize, sizeof(tcbtab[0]));
1605         if (!tcbtab)
1606                 die_out_of_memory();
1607         tcp = calloc(tcbtabsize, sizeof(*tcp));
1608         if (!tcp)
1609                 die_out_of_memory();
1610         for (c = 0; c < tcbtabsize; c++)
1611                 tcbtab[c] = tcp++;
1612
1613         shared_log = stderr;
1614         set_sortby(DEFAULT_SORTBY);
1615         set_personality(DEFAULT_PERSONALITY);
1616         qualify("trace=all");
1617         qualify("abbrev=all");
1618         qualify("verbose=all");
1619 #if DEFAULT_QUAL_FLAGS != (QUAL_TRACE | QUAL_ABBREV | QUAL_VERBOSE)
1620 # error Bug in DEFAULT_QUAL_FLAGS
1621 #endif
1622         qualify("signal=all");
1623         while ((c = getopt(argc, argv,
1624                 "+b:cCdfFhiqrtTvVxyz"
1625                 "D"
1626                 "a:e:o:O:p:s:S:u:E:P:I:")) != EOF) {
1627                 switch (c) {
1628                 case 'b':
1629                         if (strcmp(optarg, "execve") != 0)
1630                                 error_msg_and_die("Syscall '%s' for -b isn't supported",
1631                                         optarg);
1632                         detach_on_execve = 1;
1633                         break;
1634                 case 'c':
1635                         if (cflag == CFLAG_BOTH) {
1636                                 error_msg_and_die("-c and -C are mutually exclusive");
1637                         }
1638                         cflag = CFLAG_ONLY_STATS;
1639                         break;
1640                 case 'C':
1641                         if (cflag == CFLAG_ONLY_STATS) {
1642                                 error_msg_and_die("-c and -C are mutually exclusive");
1643                         }
1644                         cflag = CFLAG_BOTH;
1645                         break;
1646                 case 'd':
1647                         debug_flag = 1;
1648                         break;
1649                 case 'D':
1650                         daemonized_tracer = 1;
1651                         break;
1652                 case 'F':
1653                         optF = 1;
1654                         break;
1655                 case 'f':
1656                         followfork++;
1657                         break;
1658                 case 'h':
1659                         usage(stdout, 0);
1660                         break;
1661                 case 'i':
1662                         iflag = 1;
1663                         break;
1664                 case 'q':
1665                         qflag++;
1666                         break;
1667                 case 'r':
1668                         rflag = 1;
1669                         /* fall through to tflag++ */
1670                 case 't':
1671                         tflag++;
1672                         break;
1673                 case 'T':
1674                         Tflag = 1;
1675                         break;
1676                 case 'x':
1677                         xflag++;
1678                         break;
1679                 case 'y':
1680                         show_fd_path = 1;
1681                         break;
1682                 case 'v':
1683                         qualify("abbrev=none");
1684                         break;
1685                 case 'V':
1686                         printf("%s -- version %s\n", PACKAGE_NAME, VERSION);
1687                         exit(0);
1688                         break;
1689                 case 'z':
1690                         not_failing_only = 1;
1691                         break;
1692                 case 'a':
1693                         acolumn = string_to_uint(optarg);
1694                         if (acolumn < 0)
1695                                 error_opt_arg(c, optarg);
1696                         break;
1697                 case 'e':
1698                         qualify(optarg);
1699                         break;
1700                 case 'o':
1701                         outfname = strdup(optarg);
1702                         break;
1703                 case 'O':
1704                         i = string_to_uint(optarg);
1705                         if (i < 0)
1706                                 error_opt_arg(c, optarg);
1707                         set_overhead(i);
1708                         break;
1709                 case 'p':
1710                         process_opt_p_list(optarg);
1711                         break;
1712                 case 'P':
1713                         pathtrace_select(optarg);
1714                         break;
1715                 case 's':
1716                         i = string_to_uint(optarg);
1717                         if (i < 0)
1718                                 error_opt_arg(c, optarg);
1719                         max_strlen = i;
1720                         break;
1721                 case 'S':
1722                         set_sortby(optarg);
1723                         break;
1724                 case 'u':
1725                         username = strdup(optarg);
1726                         break;
1727                 case 'E':
1728                         if (putenv(optarg) < 0)
1729                                 die_out_of_memory();
1730                         break;
1731                 case 'I':
1732                         opt_intr = string_to_uint(optarg);
1733                         if (opt_intr <= 0 || opt_intr >= NUM_INTR_OPTS)
1734                                 error_opt_arg(c, optarg);
1735                         break;
1736                 default:
1737                         usage(stderr, 1);
1738                         break;
1739                 }
1740         }
1741         argv += optind;
1742         /* argc -= optind; - no need, argc is not used below */
1743
1744         acolumn_spaces = malloc(acolumn + 1);
1745         if (!acolumn_spaces)
1746                 die_out_of_memory();
1747         memset(acolumn_spaces, ' ', acolumn);
1748         acolumn_spaces[acolumn] = '\0';
1749
1750         /* Must have PROG [ARGS], or -p PID. Not both. */
1751         if (!argv[0] == !nprocs)
1752                 usage(stderr, 1);
1753
1754         if (nprocs != 0 && daemonized_tracer) {
1755                 error_msg_and_die("-D and -p are mutually exclusive");
1756         }
1757
1758         if (!followfork)
1759                 followfork = optF;
1760
1761         if (followfork >= 2 && cflag) {
1762                 error_msg_and_die("(-c or -C) and -ff are mutually exclusive");
1763         }
1764
1765         /* See if they want to run as another user. */
1766         if (username != NULL) {
1767                 struct passwd *pent;
1768
1769                 if (getuid() != 0 || geteuid() != 0) {
1770                         error_msg_and_die("You must be root to use the -u option");
1771                 }
1772                 pent = getpwnam(username);
1773                 if (pent == NULL) {
1774                         error_msg_and_die("Cannot find user '%s'", username);
1775                 }
1776                 run_uid = pent->pw_uid;
1777                 run_gid = pent->pw_gid;
1778         }
1779         else {
1780                 run_uid = getuid();
1781                 run_gid = getgid();
1782         }
1783
1784         /*
1785          * On any reasonably recent Linux kernel (circa about 2.5.46)
1786          * need_fork_exec_workarounds should stay 0 after these tests:
1787          */
1788         /*need_fork_exec_workarounds = 0; - already is */
1789         if (followfork)
1790                 need_fork_exec_workarounds = test_ptrace_setoptions_followfork();
1791         need_fork_exec_workarounds |= test_ptrace_setoptions_for_all();
1792         test_ptrace_seize();
1793
1794         /* Check if they want to redirect the output. */
1795         if (outfname) {
1796                 /* See if they want to pipe the output. */
1797                 if (outfname[0] == '|' || outfname[0] == '!') {
1798                         /*
1799                          * We can't do the <outfname>.PID funny business
1800                          * when using popen, so prohibit it.
1801                          */
1802                         if (followfork >= 2)
1803                                 error_msg_and_die("Piping the output and -ff are mutually exclusive");
1804                         shared_log = strace_popen(outfname + 1);
1805                 }
1806                 else if (followfork < 2)
1807                         shared_log = strace_fopen(outfname);
1808         } else {
1809                 /* -ff without -o FILE is the same as single -f */
1810                 if (followfork >= 2)
1811                         followfork = 1;
1812         }
1813
1814         if (!outfname || outfname[0] == '|' || outfname[0] == '!') {
1815                 char *buf = malloc(BUFSIZ);
1816                 if (!buf)
1817                         die_out_of_memory();
1818                 setvbuf(shared_log, buf, _IOLBF, BUFSIZ);
1819         }
1820         if (outfname && argv[0]) {
1821                 if (!opt_intr)
1822                         opt_intr = INTR_NEVER;
1823                 qflag = 1;
1824         }
1825         if (!opt_intr)
1826                 opt_intr = INTR_WHILE_WAIT;
1827
1828         /* argv[0]      -pPID   -oFILE  Default interactive setting
1829          * yes          0       0       INTR_WHILE_WAIT
1830          * no           1       0       INTR_WHILE_WAIT
1831          * yes          0       1       INTR_NEVER
1832          * no           1       1       INTR_WHILE_WAIT
1833          */
1834
1835         sigemptyset(&empty_set);
1836         sigemptyset(&blocked_set);
1837
1838         /* startup_child() must be called before the signal handlers get
1839          * installed below as they are inherited into the spawned process.
1840          * Also we do not need to be protected by them as during interruption
1841          * in the startup_child() mode we kill the spawned process anyway.
1842          */
1843         if (argv[0]) {
1844                 if (!NOMMU_SYSTEM || daemonized_tracer)
1845                         hide_log_until_execve = 1;
1846                 skip_one_b_execve = 1;
1847                 startup_child(argv);
1848         }
1849
1850         sa.sa_handler = SIG_IGN;
1851         sigemptyset(&sa.sa_mask);
1852         sa.sa_flags = 0;
1853         sigaction(SIGTTOU, &sa, NULL); /* SIG_IGN */
1854         sigaction(SIGTTIN, &sa, NULL); /* SIG_IGN */
1855         if (opt_intr != INTR_ANYWHERE) {
1856                 if (opt_intr == INTR_BLOCK_TSTP_TOO)
1857                         sigaction(SIGTSTP, &sa, NULL); /* SIG_IGN */
1858                 /*
1859                  * In interactive mode (if no -o OUTFILE, or -p PID is used),
1860                  * fatal signals are blocked while syscall stop is processed,
1861                  * and acted on in between, when waiting for new syscall stops.
1862                  * In non-interactive mode, signals are ignored.
1863                  */
1864                 if (opt_intr == INTR_WHILE_WAIT) {
1865                         sigaddset(&blocked_set, SIGHUP);
1866                         sigaddset(&blocked_set, SIGINT);
1867                         sigaddset(&blocked_set, SIGQUIT);
1868                         sigaddset(&blocked_set, SIGPIPE);
1869                         sigaddset(&blocked_set, SIGTERM);
1870                         sa.sa_handler = interrupt;
1871                 }
1872                 /* SIG_IGN, or set handler for these */
1873                 sigaction(SIGHUP, &sa, NULL);
1874                 sigaction(SIGINT, &sa, NULL);
1875                 sigaction(SIGQUIT, &sa, NULL);
1876                 sigaction(SIGPIPE, &sa, NULL);
1877                 sigaction(SIGTERM, &sa, NULL);
1878         }
1879         if (nprocs != 0 || daemonized_tracer)
1880                 startup_attach();
1881
1882         /* Do we want pids printed in our -o OUTFILE?
1883          * -ff: no (every pid has its own file); or
1884          * -f: yes (there can be more pids in the future); or
1885          * -p PID1,PID2: yes (there are already more than one pid)
1886          */
1887         print_pid_pfx = (outfname && followfork < 2 && (followfork == 1 || nprocs > 1));
1888 }
1889
1890 static struct tcb *
1891 pid2tcb(int pid)
1892 {
1893         int i;
1894
1895         if (pid <= 0)
1896                 return NULL;
1897
1898         for (i = 0; i < tcbtabsize; i++) {
1899                 struct tcb *tcp = tcbtab[i];
1900                 if (tcp->pid == pid && (tcp->flags & TCB_INUSE))
1901                         return tcp;
1902         }
1903
1904         return NULL;
1905 }
1906
1907 static void
1908 cleanup(void)
1909 {
1910         int i;
1911         struct tcb *tcp;
1912         int fatal_sig;
1913
1914         /* 'interrupted' is a volatile object, fetch it only once */
1915         fatal_sig = interrupted;
1916         if (!fatal_sig)
1917                 fatal_sig = SIGTERM;
1918
1919         for (i = 0; i < tcbtabsize; i++) {
1920                 tcp = tcbtab[i];
1921                 if (!(tcp->flags & TCB_INUSE))
1922                         continue;
1923                 if (debug_flag)
1924                         fprintf(stderr,
1925                                 "cleanup: looking at pid %u\n", tcp->pid);
1926                 if (tcp->flags & TCB_STRACE_CHILD) {
1927                         kill(tcp->pid, SIGCONT);
1928                         kill(tcp->pid, fatal_sig);
1929                 }
1930                 detach(tcp);
1931         }
1932         if (cflag)
1933                 call_summary(shared_log);
1934 }
1935
1936 static void
1937 interrupt(int sig)
1938 {
1939         interrupted = sig;
1940 }
1941
1942 static int
1943 trace(void)
1944 {
1945         struct rusage ru;
1946
1947         while (nprocs != 0) {
1948                 int pid;
1949                 int wait_errno;
1950                 int status, sig;
1951                 int stopped;
1952                 struct tcb *tcp;
1953                 unsigned event;
1954
1955                 if (interrupted)
1956                         return 0;
1957
1958                 if (interactive)
1959                         sigprocmask(SIG_SETMASK, &empty_set, NULL);
1960                 pid = wait4(-1, &status, __WALL, (cflag ? &ru : NULL));
1961                 wait_errno = errno;
1962                 if (interactive)
1963                         sigprocmask(SIG_BLOCK, &blocked_set, NULL);
1964
1965                 if (pid < 0) {
1966                         if (wait_errno == EINTR)
1967                                 continue;
1968                         if (wait_errno == ECHILD)
1969                                 /* Should not happen since nprocs > 0 */
1970                                 return 0;
1971                         errno = wait_errno;
1972                         perror_msg("wait4(__WALL)");
1973                         return -1;
1974                 }
1975
1976                 if (pid == popen_pid) {
1977                         if (WIFEXITED(status) || WIFSIGNALED(status))
1978                                 popen_pid = 0;
1979                         continue;
1980                 }
1981
1982                 event = ((unsigned)status >> 16);
1983                 if (debug_flag) {
1984                         char buf[sizeof("WIFEXITED,exitcode=%u") + sizeof(int)*3 /*paranoia:*/ + 16];
1985                         char evbuf[sizeof(",PTRACE_EVENT_?? (%u)") + sizeof(int)*3 /*paranoia:*/ + 16];
1986                         strcpy(buf, "???");
1987                         if (WIFSIGNALED(status))
1988 #ifdef WCOREDUMP
1989                                 sprintf(buf, "WIFSIGNALED,%ssig=%s",
1990                                                 WCOREDUMP(status) ? "core," : "",
1991                                                 signame(WTERMSIG(status)));
1992 #else
1993                                 sprintf(buf, "WIFSIGNALED,sig=%s",
1994                                                 signame(WTERMSIG(status)));
1995 #endif
1996                         if (WIFEXITED(status))
1997                                 sprintf(buf, "WIFEXITED,exitcode=%u", WEXITSTATUS(status));
1998                         if (WIFSTOPPED(status))
1999                                 sprintf(buf, "WIFSTOPPED,sig=%s", signame(WSTOPSIG(status)));
2000 #ifdef WIFCONTINUED
2001                         if (WIFCONTINUED(status))
2002                                 strcpy(buf, "WIFCONTINUED");
2003 #endif
2004                         evbuf[0] = '\0';
2005                         if (event != 0) {
2006                                 static const char *const event_names[] = {
2007                                         [PTRACE_EVENT_CLONE] = "CLONE",
2008                                         [PTRACE_EVENT_FORK]  = "FORK",
2009                                         [PTRACE_EVENT_VFORK] = "VFORK",
2010                                         [PTRACE_EVENT_VFORK_DONE] = "VFORK_DONE",
2011                                         [PTRACE_EVENT_EXEC]  = "EXEC",
2012                                         [PTRACE_EVENT_EXIT]  = "EXIT",
2013                                 };
2014                                 const char *e;
2015                                 if (event < ARRAY_SIZE(event_names))
2016                                         e = event_names[event];
2017                                 else {
2018                                         sprintf(buf, "?? (%u)", event);
2019                                         e = buf;
2020                                 }
2021                                 sprintf(evbuf, ",PTRACE_EVENT_%s", e);
2022                         }
2023                         fprintf(stderr, " [wait(0x%04x) = %u] %s%s\n", status, pid, buf, evbuf);
2024                 }
2025
2026                 /* Look up 'pid' in our table. */
2027                 tcp = pid2tcb(pid);
2028
2029                 if (!tcp) {
2030                         if (followfork) {
2031                                 tcp = alloctcb(pid);
2032                                 tcp->flags |= TCB_ATTACHED | TCB_STARTUP | post_attach_sigstop;
2033                                 newoutf(tcp);
2034                                 if (!qflag)
2035                                         fprintf(stderr, "Process %d attached\n",
2036                                                 pid);
2037                         } else {
2038                                 /* This can happen if a clone call used
2039                                    CLONE_PTRACE itself.  */
2040                                 if (WIFSTOPPED(status))
2041                                         ptrace(PTRACE_CONT, pid, (char *) 0, 0);
2042                                 error_msg_and_die("Unknown pid: %u", pid);
2043                         }
2044                 }
2045
2046                 clear_regs();
2047                 if (WIFSTOPPED(status))
2048                         get_regs(pid);
2049
2050                 /* Under Linux, execve changes pid to thread leader's pid,
2051                  * and we see this changed pid on EVENT_EXEC and later,
2052                  * execve sysexit. Leader "disappears" without exit
2053                  * notification. Let user know that, drop leader's tcb,
2054                  * and fix up pid in execve thread's tcb.
2055                  * Effectively, execve thread's tcb replaces leader's tcb.
2056                  *
2057                  * BTW, leader is 'stuck undead' (doesn't report WIFEXITED
2058                  * on exit syscall) in multithreaded programs exactly
2059                  * in order to handle this case.
2060                  *
2061                  * PTRACE_GETEVENTMSG returns old pid starting from Linux 3.0.
2062                  * On 2.6 and earlier, it can return garbage.
2063                  */
2064                 if (event == PTRACE_EVENT_EXEC && os_release >= KERNEL_VERSION(3,0,0)) {
2065                         FILE *fp;
2066                         struct tcb *execve_thread;
2067                         long old_pid = 0;
2068
2069                         if (ptrace(PTRACE_GETEVENTMSG, pid, NULL, (long) &old_pid) < 0)
2070                                 goto dont_switch_tcbs;
2071                         if (old_pid <= 0 || old_pid == pid)
2072                                 goto dont_switch_tcbs;
2073                         execve_thread = pid2tcb(old_pid);
2074                         /* It should be !NULL, but I feel paranoid */
2075                         if (!execve_thread)
2076                                 goto dont_switch_tcbs;
2077
2078                         if (execve_thread->curcol != 0) {
2079                                 /*
2080                                  * One case we are here is -ff:
2081                                  * try "strace -oLOG -ff test/threaded_execve"
2082                                  */
2083                                 fprintf(execve_thread->outf, " <pid changed to %d ...>\n", pid);
2084                                 /*execve_thread->curcol = 0; - no need, see code below */
2085                         }
2086                         /* Swap output FILEs (needed for -ff) */
2087                         fp = execve_thread->outf;
2088                         execve_thread->outf = tcp->outf;
2089                         tcp->outf = fp;
2090                         /* And their column positions */
2091                         execve_thread->curcol = tcp->curcol;
2092                         tcp->curcol = 0;
2093                         /* Drop leader, but close execve'd thread outfile (if -ff) */
2094                         droptcb(tcp);
2095                         /* Switch to the thread, reusing leader's outfile and pid */
2096                         tcp = execve_thread;
2097                         tcp->pid = pid;
2098                         if (cflag != CFLAG_ONLY_STATS) {
2099                                 printleader(tcp);
2100                                 tprintf("+++ superseded by execve in pid %lu +++\n", old_pid);
2101                                 line_ended();
2102                                 tcp->flags |= TCB_REPRINT;
2103                         }
2104                 }
2105  dont_switch_tcbs:
2106
2107                 if (event == PTRACE_EVENT_EXEC) {
2108                         if (detach_on_execve && !skip_one_b_execve)
2109                                 detach(tcp); /* do "-b execve" thingy */
2110                         skip_one_b_execve = 0;
2111                 }
2112
2113                 /* Set current output file */
2114                 current_tcp = tcp;
2115
2116                 if (cflag) {
2117                         tv_sub(&tcp->dtime, &ru.ru_stime, &tcp->stime);
2118                         tcp->stime = ru.ru_stime;
2119                 }
2120
2121                 if (WIFSIGNALED(status)) {
2122                         if (pid == strace_child)
2123                                 exit_code = 0x100 | WTERMSIG(status);
2124                         if (cflag != CFLAG_ONLY_STATS
2125                          && (qual_flags[WTERMSIG(status)] & QUAL_SIGNAL)
2126                         ) {
2127                                 printleader(tcp);
2128 #ifdef WCOREDUMP
2129                                 tprintf("+++ killed by %s %s+++\n",
2130                                         signame(WTERMSIG(status)),
2131                                         WCOREDUMP(status) ? "(core dumped) " : "");
2132 #else
2133                                 tprintf("+++ killed by %s +++\n",
2134                                         signame(WTERMSIG(status)));
2135 #endif
2136                                 line_ended();
2137                         }
2138                         droptcb(tcp);
2139                         continue;
2140                 }
2141                 if (WIFEXITED(status)) {
2142                         if (pid == strace_child)
2143                                 exit_code = WEXITSTATUS(status);
2144                         if (cflag != CFLAG_ONLY_STATS &&
2145                             qflag < 2) {
2146                                 printleader(tcp);
2147                                 tprintf("+++ exited with %d +++\n", WEXITSTATUS(status));
2148                                 line_ended();
2149                         }
2150                         droptcb(tcp);
2151                         continue;
2152                 }
2153                 if (!WIFSTOPPED(status)) {
2154                         fprintf(stderr, "PANIC: pid %u not stopped\n", pid);
2155                         droptcb(tcp);
2156                         continue;
2157                 }
2158
2159                 /* Is this the very first time we see this tracee stopped? */
2160                 if (tcp->flags & TCB_STARTUP) {
2161                         if (debug_flag)
2162                                 fprintf(stderr, "pid %d has TCB_STARTUP, initializing it\n", tcp->pid);
2163                         tcp->flags &= ~TCB_STARTUP;
2164                         if (tcp->flags & TCB_BPTSET) {
2165                                 /*
2166                                  * One example is a breakpoint inherited from
2167                                  * parent through fork().
2168                                  */
2169                                 if (clearbpt(tcp) < 0) {
2170                                         /* Pretty fatal */
2171                                         droptcb(tcp);
2172                                         cleanup();
2173                                         return -1;
2174                                 }
2175                         }
2176                         if (ptrace_setoptions) {
2177                                 if (debug_flag)
2178                                         fprintf(stderr, "setting opts %x on pid %d\n", ptrace_setoptions, tcp->pid);
2179                                 if (ptrace(PTRACE_SETOPTIONS, tcp->pid, NULL, ptrace_setoptions) < 0) {
2180                                         if (errno != ESRCH) {
2181                                                 /* Should never happen, really */
2182                                                 perror_msg_and_die("PTRACE_SETOPTIONS");
2183                                         }
2184                                 }
2185                         }
2186                 }
2187
2188                 sig = WSTOPSIG(status);
2189
2190                 if (event != 0) {
2191                         /* Ptrace event */
2192 #if USE_SEIZE
2193                         if (event == PTRACE_EVENT_STOP) {
2194                                 /*
2195                                  * PTRACE_INTERRUPT-stop or group-stop.
2196                                  * PTRACE_INTERRUPT-stop has sig == SIGTRAP here.
2197                                  */
2198                                 if (sig == SIGSTOP
2199                                  || sig == SIGTSTP
2200                                  || sig == SIGTTIN
2201                                  || sig == SIGTTOU
2202                                 ) {
2203                                         stopped = 1;
2204                                         goto show_stopsig;
2205                                 }
2206                         }
2207 #endif
2208                         goto restart_tracee_with_sig_0;
2209                 }
2210
2211                 /* Is this post-attach SIGSTOP?
2212                  * Interestingly, the process may stop
2213                  * with STOPSIG equal to some other signal
2214                  * than SIGSTOP if we happend to attach
2215                  * just before the process takes a signal.
2216                  */
2217                 if (sig == SIGSTOP && (tcp->flags & TCB_IGNORE_ONE_SIGSTOP)) {
2218                         if (debug_flag)
2219                                 fprintf(stderr, "ignored SIGSTOP on pid %d\n", tcp->pid);
2220                         tcp->flags &= ~TCB_IGNORE_ONE_SIGSTOP;
2221                         goto restart_tracee_with_sig_0;
2222                 }
2223
2224                 if (sig != syscall_trap_sig) {
2225                         siginfo_t si;
2226
2227                         /* Nonzero (true) if tracee is stopped by signal
2228                          * (as opposed to "tracee received signal").
2229                          * TODO: shouldn't we check for errno == EINVAL too?
2230                          * We can get ESRCH instead, you know...
2231                          */
2232                         stopped = (ptrace(PTRACE_GETSIGINFO, pid, 0, (long) &si) < 0);
2233 #if USE_SEIZE
2234  show_stopsig:
2235 #endif
2236                         if (cflag != CFLAG_ONLY_STATS
2237                             && !hide_log_until_execve
2238                             && (qual_flags[sig] & QUAL_SIGNAL)
2239                            ) {
2240 #if defined(PT_CR_IPSR) && defined(PT_CR_IIP)
2241                                 long pc = 0;
2242                                 long psr = 0;
2243
2244                                 upeek(tcp, PT_CR_IPSR, &psr);
2245                                 upeek(tcp, PT_CR_IIP, &pc);
2246
2247 # define PSR_RI 41
2248                                 pc += (psr >> PSR_RI) & 0x3;
2249 # define PC_FORMAT_STR  " @ %lx"
2250 # define PC_FORMAT_ARG  , pc
2251 #else
2252 # define PC_FORMAT_STR  ""
2253 # define PC_FORMAT_ARG  /* nothing */
2254 #endif
2255                                 printleader(tcp);
2256                                 if (!stopped) {
2257                                         tprintf("--- %s ", signame(sig));
2258                                         printsiginfo(&si, verbose(tcp));
2259                                         tprintf(PC_FORMAT_STR " ---\n"
2260                                                 PC_FORMAT_ARG);
2261                                 } else
2262                                         tprintf("--- stopped by %s" PC_FORMAT_STR " ---\n",
2263                                                 signame(sig)
2264                                                 PC_FORMAT_ARG);
2265                                 line_ended();
2266                         }
2267
2268                         if (!stopped)
2269                                 /* It's signal-delivery-stop. Inject the signal */
2270                                 goto restart_tracee;
2271
2272                         /* It's group-stop */
2273                         if (use_seize) {
2274                                 /*
2275                                  * This ends ptrace-stop, but does *not* end group-stop.
2276                                  * This makes stopping signals work properly on straced process
2277                                  * (that is, process really stops. It used to continue to run).
2278                                  */
2279                                 if (ptrace_restart(PTRACE_LISTEN, tcp, 0) < 0) {
2280                                         cleanup();
2281                                         return -1;
2282                                 }
2283                                 continue;
2284                         }
2285                         /* We don't have PTRACE_LISTEN support... */
2286                         goto restart_tracee;
2287                 }
2288
2289                 /* We handled quick cases, we are permitted to interrupt now. */
2290                 if (interrupted)
2291                         return 0;
2292
2293                 /* This should be syscall entry or exit.
2294                  * (Or it still can be that pesky post-execve SIGTRAP!)
2295                  * Handle it.
2296                  */
2297                 if (trace_syscall(tcp) < 0) {
2298                         /* ptrace() failed in trace_syscall().
2299                          * Likely a result of process disappearing mid-flight.
2300                          * Observed case: exit_group() or SIGKILL terminating
2301                          * all processes in thread group.
2302                          * We assume that ptrace error was caused by process death.
2303                          * We used to detach(tcp) here, but since we no longer
2304                          * implement "detach before death" policy/hack,
2305                          * we can let this process to report its death to us
2306                          * normally, via WIFEXITED or WIFSIGNALED wait status.
2307                          */
2308                         continue;
2309                 }
2310  restart_tracee_with_sig_0:
2311                 sig = 0;
2312  restart_tracee:
2313                 if (ptrace_restart(PTRACE_SYSCALL, tcp, sig) < 0) {
2314                         cleanup();
2315                         return -1;
2316                 }
2317         }
2318         return 0;
2319 }
2320
2321 int
2322 main(int argc, char *argv[])
2323 {
2324         init(argc, argv);
2325
2326         /* Run main tracing loop */
2327         if (trace() < 0)
2328                 return 1;
2329
2330         cleanup();
2331         fflush(NULL);
2332         if (shared_log != stderr)
2333                 fclose(shared_log);
2334         if (popen_pid) {
2335                 while (waitpid(popen_pid, NULL, 0) < 0 && errno == EINTR)
2336                         ;
2337         }
2338         if (exit_code > 0xff) {
2339                 /* Avoid potential core file clobbering.  */
2340                 struct_rlimit rlim = {0, 0};
2341                 set_rlimit(RLIMIT_CORE, &rlim);
2342
2343                 /* Child was killed by a signal, mimic that.  */
2344                 exit_code &= 0xff;
2345                 signal(exit_code, SIG_DFL);
2346                 raise(exit_code);
2347                 /* Paranoia - what if this signal is not fatal?
2348                    Exit with 128 + signo then.  */
2349                 exit_code += 128;
2350         }
2351
2352         return exit_code;
2353 }