strace.c

/*
 * Copyright (c) 1991, 1992 Paul Kranenburg <pk@cs.few.eur.nl>
 * Copyright (c) 1993 Branko Lankester <branko@hacktic.nl>
 * Copyright (c) 1993, 1994, 1995, 1996 Rick Sladkey <jrs@world.std.com>
 * Copyright (c) 1996-1999 Wichert Akkerman <wichert@cistron.nl>
 * All rights reserved.
 *
 * Redistribution and use in source and binary forms, with or without
 * modification, are permitted provided that the following conditions
 * are met:
 * 1. Redistributions of source code must retain the above copyright
 *    notice, this list of conditions and the following disclaimer.
 * 2. Redistributions in binary form must reproduce the above copyright
 *    notice, this list of conditions and the following disclaimer in the
 *    documentation and/or other materials provided with the distribution.
 * 3. The name of the author may not be used to endorse or promote products
 *    derived from this software without specific prior written permission.
 *
 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
 * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
 * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
 * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
 * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
 * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
 * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 */

#include "defs.h"
#include <stdarg.h>
#include <sys/param.h>
#include <fcntl.h>
#include <signal.h>
#include <sys/resource.h>
#include <sys/wait.h>
#include <sys/stat.h>
#include <pwd.h>
#include <grp.h>
#include <dirent.h>
#include <sys/utsname.h>
#ifdef HAVE_PRCTL
# include <sys/prctl.h>
#endif

#include "ptrace.h"
#include "printsiginfo.h"

/* In some libc, these aren't declared. Do it ourself: */
extern char **environ;
extern int optind;
extern char *optarg;

#ifdef USE_LIBUNWIND
/* if this is true do the stack trace for every system call */
bool stack_trace_enabled = false;
#endif

#if defined __NR_tkill
# define my_tkill(tid, sig) syscall(__NR_tkill, (tid), (sig))
#else
   /* kill() may choose arbitrarily the target task of the process group
      while we later wait on a that specific TID.  PID process waits become
      TID task specific waits for a process under ptrace(2).  */
# warning "tkill(2) not available, risk of strace hangs!"
# define my_tkill(tid, sig) kill((tid), (sig))
#endif

/* Glue for systems without a MMU that cannot provide fork() */
#if !defined(HAVE_FORK)
# undef NOMMU_SYSTEM
# define NOMMU_SYSTEM 1
#endif
#if NOMMU_SYSTEM
# define fork() vfork()
#endif

const unsigned int syscall_trap_sig = SIGTRAP | 0x80;

cflag_t cflag = CFLAG_NONE;
unsigned int followfork = 0;
unsigned int ptrace_setoptions = PTRACE_O_TRACESYSGOOD | PTRACE_O_TRACEEXEC;
unsigned int xflag = 0;
bool debug_flag = 0;
bool Tflag = 0;
bool iflag = 0;
bool count_wallclock = 0;
unsigned int qflag = 0;
static unsigned int tflag = 0;
static bool rflag = 0;
static bool print_pid_pfx = 0;

/* -I n */
enum {
    INTR_NOT_SET        = 0,
    INTR_ANYWHERE       = 1, /* don't block/ignore any signals */
    INTR_WHILE_WAIT     = 2, /* block fatal signals while decoding syscall. default */
    INTR_NEVER          = 3, /* block fatal signals. default if '-o FILE PROG' */
    INTR_BLOCK_TSTP_TOO = 4, /* block fatal signals and SIGTSTP (^Z) */
    NUM_INTR_OPTS
};
static int opt_intr;
/* We play with signal mask only if this mode is active: */
#define interactive (opt_intr == INTR_WHILE_WAIT)

/*
 * daemonized_tracer supports -D option.
 * With this option, strace forks twice.
 * Unlike normal case, with -D *grandparent* process exec's,
 * becoming a traced process. Child exits (this prevents traced process
 * from having children it doesn't expect to have), and grandchild
 * attaches to grandparent similarly to strace -p PID.
 * This allows for more transparent interaction in cases
 * when process and its parent are communicating via signals,
 * wait() etc. Without -D, strace process gets lodged in between,
 * disrupting parent<->child link.
 */
static bool daemonized_tracer = 0;

#if USE_SEIZE
static int post_attach_sigstop = TCB_IGNORE_ONE_SIGSTOP;
# define use_seize (post_attach_sigstop == 0)
#else
# define post_attach_sigstop TCB_IGNORE_ONE_SIGSTOP
# define use_seize 0
#endif

/* Sometimes we want to print only succeeding syscalls. */
bool not_failing_only = 0;

/* Show path associated with fd arguments */
unsigned int show_fd_path = 0;

static bool detach_on_execve = 0;
/* Are we "strace PROG" and need to skip detach on first execve? */
static bool skip_one_b_execve = 0;
/* Are we "strace PROG" and need to hide everything until execve? */
bool hide_log_until_execve = 0;

static int exit_code = 0;
static int strace_child = 0;
static int strace_tracer_pid = 0;

static char *username = NULL;
static uid_t run_uid;
static gid_t run_gid;

unsigned int max_strlen = DEFAULT_STRLEN;
static int acolumn = DEFAULT_ACOLUMN;
static char *acolumn_spaces;

static char *outfname = NULL;
/* If -ff, points to stderr. Else, it's our common output log */
static FILE *shared_log;

struct tcb *printing_tcp = NULL;
static struct tcb *current_tcp;

static struct tcb **tcbtab;
static unsigned int nprocs, tcbtabsize;
static const char *progname;

unsigned os_release; /* generated from uname()'s u.release */

static void detach(struct tcb *tcp);
static void cleanup(void);
static void interrupt(int sig);
static sigset_t empty_set, blocked_set;

#ifdef HAVE_SIG_ATOMIC_T
static volatile sig_atomic_t interrupted;
#else
static volatile int interrupted;
#endif

#ifndef HAVE_STRERROR

#if !HAVE_DECL_SYS_ERRLIST
extern int sys_nerr;
extern char *sys_errlist[];
#endif

const char *
strerror(int err_no)
{
	static char buf[sizeof("Unknown error %d") + sizeof(int)*3];

	if (err_no < 1 || err_no >= sys_nerr) {
		sprintf(buf, "Unknown error %d", err_no);
		return buf;
	}
	return sys_errlist[err_no];
}

#endif /* HAVE_STERRROR */

static void
usage()
{
	printf("\
usage: strace [-CdffhiqrtttTvVwxxy] [-I n] [-e expr]...\n\
              [-a column] [-o file] [-s strsize] [-P path]...\n\
              -p pid... / [-D] [-E var=val]... [-u username] PROG [ARGS]\n\
   or: strace -c[dfw] [-I n] [-e expr]... [-O overhead] [-S sortby]\n\
              -p pid... / [-D] [-E var=val]... [-u username] PROG [ARGS]\n\
\n\
Output format:\n\
  -a column      alignment COLUMN for printing syscall results (default %d)\n\
  -i             print instruction pointer at time of syscall\n\
  -o file        send trace output to FILE instead of stderr\n\
  -q             suppress messages about attaching, detaching, etc.\n\
  -r             print relative timestamp\n\
  -s strsize     limit length of print strings to STRSIZE chars (default %d)\n\
  -t             print absolute timestamp\n\
  -tt            print absolute timestamp with usecs\n\
  -T             print time spent in each syscall\n\
  -x             print non-ascii strings in hex\n\
  -xx            print all strings in hex\n\
  -y             print paths associated with file descriptor arguments\n\
  -yy            print ip:port pairs associated with socket file descriptors\n\
\n\
Statistics:\n\
  -c             count time, calls, and errors for each syscall and report summary\n\
  -C             like -c but also print regular output\n\
  -O overhead    set overhead for tracing syscalls to OVERHEAD usecs\n\
  -S sortby      sort syscall counts by: time, calls, name, nothing (default %s)\n\
  -w             summarise syscall latency (default is system time)\n\
\n\
Filtering:\n\
  -e expr        a qualifying expression: option=[!]all or option=[!]val1[,val2]...\n\
     options:    trace, abbrev, verbose, raw, signal, read, write\n\
  -P path        trace accesses to path\n\
\n\
Tracing:\n\
  -b execve      detach on execve syscall\n\
  -D             run tracer process as a detached grandchild, not as parent\n\
  -f             follow forks\n\
  -ff            follow forks with output into separate files\n\
  -I interruptible\n\
     1:          no signals are blocked\n\
     2:          fatal signals are blocked while decoding syscall (default)\n\
     3:          fatal signals are always blocked (default if '-o FILE PROG')\n\
     4:          fatal signals and SIGTSTP (^Z) are always blocked\n\
                 (useful to make 'strace -o FILE PROG' not stop on ^Z)\n\
\n\
Startup:\n\
  -E var         remove var from the environment for command\n\
  -E var=val     put var=val in the environment for command\n\
  -p pid         trace process with process id PID, may be repeated\n\
  -u username    run command as username handling setuid and/or setgid\n\
\n\
Miscellaneous:\n\
  -d             enable debug output to stderr\n\
  -v             verbose mode: print unabbreviated argv, stat, termios, etc. args\n\
  -h             print help message\n\
  -V             print version\n\
"
#ifdef USE_LIBUNWIND
"  -k             obtain stack trace between each syscall (experimental)\n\
"
#endif
/* ancient, no one should use it
-F -- attempt to follow vforks (deprecated, use -f)\n\
 */
/* this is broken, so don't document it
-z -- print only succeeding syscalls\n\
 */
, DEFAULT_ACOLUMN, DEFAULT_STRLEN, DEFAULT_SORTBY);
	exit(0);
}

static void ATTRIBUTE_NORETURN
die(void)
{
	if (strace_tracer_pid == getpid()) {
		cflag = 0;
		cleanup();
	}
	exit(1);
}

static void verror_msg(int err_no, const char *fmt, va_list p)
{
	char *msg;

	fflush(NULL);

	/* We want to print entire message with single fprintf to ensure
	 * message integrity if stderr is shared with other programs.
	 * Thus we use vasprintf + single fprintf.
	 */
	msg = NULL;
	if (vasprintf(&msg, fmt, p) >= 0) {
		if (err_no)
			fprintf(stderr, "%s: %s: %s\n", progname, msg, strerror(err_no));
		else
			fprintf(stderr, "%s: %s\n", progname, msg);
		free(msg);
	} else {
		/* malloc in vasprintf failed, try it without malloc */
		fprintf(stderr, "%s: ", progname);
		vfprintf(stderr, fmt, p);
		if (err_no)
			fprintf(stderr, ": %s\n", strerror(err_no));
		else
			putc('\n', stderr);
	}
	/* We don't switch stderr to buffered, thus fprintf(stderr)
	 * always flushes its output and this is not necessary: */
	/* fflush(stderr); */
}

void error_msg(const char *fmt, ...)
{
	va_list p;
	va_start(p, fmt);
	verror_msg(0, fmt, p);
	va_end(p);
}

void error_msg_and_die(const char *fmt, ...)
{
	va_list p;
	va_start(p, fmt);
	verror_msg(0, fmt, p);
	die();
}

void error_msg_and_help(const char *fmt, ...)
{
	if (fmt != NULL) {
		va_list p;
		va_start(p, fmt);
		verror_msg(0, fmt, p);
	}
	fprintf(stderr, "Try '%s -h' for more information.\n", progname);
	die();
}

void perror_msg(const char *fmt, ...)
{
	va_list p;
	va_start(p, fmt);
	verror_msg(errno, fmt, p);
	va_end(p);
}

void perror_msg_and_die(const char *fmt, ...)
{
	va_list p;
	va_start(p, fmt);
	verror_msg(errno, fmt, p);
	die();
}

static void
error_opt_arg(int opt, const char *arg)
{
	error_msg_and_help("invalid -%c argument: '%s'", opt, arg);
}

#if USE_SEIZE
static int
ptrace_attach_or_seize(int pid)
{
	int r;
	if (!use_seize)
		return ptrace(PTRACE_ATTACH, pid, 0L, 0L);
	r = ptrace(PTRACE_SEIZE, pid, 0L, (unsigned long) ptrace_setoptions);
	if (r)
		return r;
	r = ptrace(PTRACE_INTERRUPT, pid, 0L, 0L);
	return r;
}
#else
# define ptrace_attach_or_seize(pid) ptrace(PTRACE_ATTACH, (pid), 0, 0)
#endif

/*
 * Used when we want to unblock stopped traced process.
 * Should be only used with PTRACE_CONT, PTRACE_DETACH and PTRACE_SYSCALL.
 * Returns 0 on success or if error was ESRCH
 * (presumably process was killed while we talk to it).
 * Otherwise prints error message and returns -1.
 */
static int
ptrace_restart(int op, struct tcb *tcp, int sig)
{
	int err;
	const char *msg;

	errno = 0;
	ptrace(op, tcp->pid, (void *) 0, (long) sig);
	err = errno;
	if (!err)
		return 0;

	msg = "SYSCALL";
	if (op == PTRACE_CONT)
		msg = "CONT";
	if (op == PTRACE_DETACH)
		msg = "DETACH";
#ifdef PTRACE_LISTEN
	if (op == PTRACE_LISTEN)
		msg = "LISTEN";
#endif
	/*
	 * Why curcol != 0? Otherwise sometimes we get this:
	 *
	 * 10252 kill(10253, SIGKILL)              = 0
	 *  <ptrace(SYSCALL,10252):No such process>10253 ...next decode...
	 *
	 * 10252 died after we retrieved syscall exit data,
	 * but before we tried to restart it. Log looks ugly.
	 */
	if (current_tcp && current_tcp->curcol != 0) {
		tprintf(" <ptrace(%s):%s>\n", msg, strerror(err));
		line_ended();
	}
	if (err == ESRCH)
		return 0;
	errno = err;
	perror_msg("ptrace(PTRACE_%s,pid:%d,sig:%d)", msg, tcp->pid, sig);
	return -1;
}

static void
set_cloexec_flag(int fd)
{
	int flags, newflags;

	flags = fcntl(fd, F_GETFD);
	if (flags < 0) {
		/* Can happen only if fd is bad.
		 * Should never happen: if it does, we have a bug
		 * in the caller. Therefore we just abort
		 * instead of propagating the error.
		 */
		perror_msg_and_die("fcntl(%d, F_GETFD)", fd);
	}

	newflags = flags | FD_CLOEXEC;
	if (flags == newflags)
		return;

	fcntl(fd, F_SETFD, newflags); /* never fails */
}

static void kill_save_errno(pid_t pid, int sig)
{
	int saved_errno = errno;

	(void) kill(pid, sig);
	errno = saved_errno;
}

/*
 * When strace is setuid executable, we have to swap uids
 * before and after filesystem and process management operations.
 */
static void
swap_uid(void)
{
	int euid = geteuid(), uid = getuid();

	if (euid != uid && setreuid(euid, uid) < 0) {
		perror_msg_and_die("setreuid");
	}
}

#ifdef _LARGEFILE64_SOURCE
# ifdef HAVE_FOPEN64
#  define fopen_for_output fopen64
# else
#  define fopen_for_output fopen
# endif
# define struct_stat struct stat64
# define stat_file stat64
# define struct_dirent struct dirent64
# define read_dir readdir64
# define struct_rlimit struct rlimit64
# define set_rlimit setrlimit64
#else
# define fopen_for_output fopen
# define struct_stat struct stat
# define stat_file stat
# define struct_dirent struct dirent
# define read_dir readdir
# define struct_rlimit struct rlimit
# define set_rlimit setrlimit
#endif

static FILE *
strace_fopen(const char *path)
{
	FILE *fp;

	swap_uid();
	fp = fopen_for_output(path, "w");
	if (!fp)
		perror_msg_and_die("Can't fopen '%s'", path);
	swap_uid();
	set_cloexec_flag(fileno(fp));
	return fp;
}

static int popen_pid = 0;

#ifndef _PATH_BSHELL
# define _PATH_BSHELL "/bin/sh"
#endif

/*
 * We cannot use standard popen(3) here because we have to distinguish
 * popen child process from other processes we trace, and standard popen(3)
 * does not export its child's pid.
 */
static FILE *
strace_popen(const char *command)
{
	FILE *fp;
	int pid;
	int fds[2];

	swap_uid();
	if (pipe(fds) < 0)
		perror_msg_and_die("pipe");

	set_cloexec_flag(fds[1]); /* never fails */

	pid = vfork();
	if (pid < 0)
		perror_msg_and_die("vfork");

	if (pid == 0) {
		/* child */
		close(fds[1]);
		if (fds[0] != 0) {
			if (dup2(fds[0], 0))
				perror_msg_and_die("dup2");
			close(fds[0]);
		}
		execl(_PATH_BSHELL, "sh", "-c", command, NULL);
		perror_msg_and_die("Can't execute '%s'", _PATH_BSHELL);
	}

	/* parent */
	popen_pid = pid;
	close(fds[0]);
	swap_uid();
	fp = fdopen(fds[1], "w");
	if (!fp)
		die_out_of_memory();
	return fp;
}

void
tprintf(const char *fmt, ...)
{
	va_list args;

	va_start(args, fmt);
	if (current_tcp) {
		int n = strace_vfprintf(current_tcp->outf, fmt, args);
		if (n < 0) {
			if (current_tcp->outf != stderr)
				perror_msg("%s", outfname);
		} else
			current_tcp->curcol += n;
	}
	va_end(args);
}

#ifndef HAVE_FPUTS_UNLOCKED
# define fputs_unlocked fputs
#endif

void
tprints(const char *str)
{
	if (current_tcp) {
		int n = fputs_unlocked(str, current_tcp->outf);
		if (n >= 0) {
			current_tcp->curcol += strlen(str);
			return;
		}
		if (current_tcp->outf != stderr)
			perror_msg("%s", outfname);
	}
}

void
line_ended(void)
{
	if (current_tcp) {
		current_tcp->curcol = 0;
		fflush(current_tcp->outf);
	}
	if (printing_tcp) {
		printing_tcp->curcol = 0;
		printing_tcp = NULL;
	}
}

void
printleader(struct tcb *tcp)
{
	/* If -ff, "previous tcb we printed" is always the same as current,
	 * because we have per-tcb output files.
	 */
	if (followfork >= 2)
		printing_tcp = tcp;

	if (printing_tcp) {
		current_tcp = printing_tcp;
		if (printing_tcp->curcol != 0 && (followfork < 2 || printing_tcp == tcp)) {
			/*
			 * case 1: we have a shared log (i.e. not -ff), and last line
			 * wasn't finished (same or different tcb, doesn't matter).
			 * case 2: split log, we are the same tcb, but our last line
			 * didn't finish ("SIGKILL nuked us after syscall entry" etc).
			 */
			tprints(" <unfinished ...>\n");
			printing_tcp->curcol = 0;
		}
	}

	printing_tcp = tcp;
	current_tcp = tcp;
	current_tcp->curcol = 0;

	if (print_pid_pfx)
		tprintf("%-5d ", tcp->pid);
	else if (nprocs > 1 && !outfname)
		tprintf("[pid %5u] ", tcp->pid);

	if (tflag) {
		char str[sizeof("HH:MM:SS")];
		struct timeval tv, dtv;
		static struct timeval otv;

		gettimeofday(&tv, NULL);
		if (rflag) {
			if (otv.tv_sec == 0)
				otv = tv;
			tv_sub(&dtv, &tv, &otv);
			tprintf("%6ld.%06ld ",
				(long) dtv.tv_sec, (long) dtv.tv_usec);
			otv = tv;
		}
		else if (tflag > 2) {
			tprintf("%ld.%06ld ",
				(long) tv.tv_sec, (long) tv.tv_usec);
		}
		else {
			time_t local = tv.tv_sec;
			strftime(str, sizeof(str), "%T", localtime(&local));
			if (tflag > 1)
				tprintf("%s.%06ld ", str, (long) tv.tv_usec);
			else
				tprintf("%s ", str);
		}
	}
	if (iflag)
		print_pc(tcp);
}

void
tabto(void)
{
	if (current_tcp->curcol < acolumn)
		tprints(acolumn_spaces + current_tcp->curcol);
}

/* Should be only called directly *after successful attach* to a tracee.
 * Otherwise, "strace -oFILE -ff -p<nonexistant_pid>"
 * may create bogus empty FILE.<nonexistant_pid>, and then die.
 */
static void
newoutf(struct tcb *tcp)
{
	tcp->outf = shared_log; /* if not -ff mode, the same file is for all */
	if (followfork >= 2) {
		char name[520 + sizeof(int) * 3];
		sprintf(name, "%.512s.%u", outfname, tcp->pid);
		tcp->outf = strace_fopen(name);
	}
}

static void
expand_tcbtab(void)
{
	/* Allocate some more TCBs and expand the table.
	   We don't want to relocate the TCBs because our
	   callers have pointers and it would be a pain.
	   So tcbtab is a table of pointers.  Since we never
	   free the TCBs, we allocate a single chunk of many.  */
	unsigned int i = tcbtabsize;
	struct tcb *newtcbs = xcalloc(tcbtabsize, sizeof(newtcbs[0]));
	struct tcb **newtab = xreallocarray(tcbtab, tcbtabsize * 2,
					    sizeof(tcbtab[0]));
	tcbtabsize *= 2;
	tcbtab = newtab;
	while (i < tcbtabsize)
		tcbtab[i++] = newtcbs++;
}

static struct tcb *
alloctcb(int pid)
{
	unsigned int i;
	struct tcb *tcp;

	if (nprocs == tcbtabsize)
		expand_tcbtab();

	for (i = 0; i < tcbtabsize; i++) {
		tcp = tcbtab[i];
		if (!tcp->pid) {
			memset(tcp, 0, sizeof(*tcp));
			tcp->pid = pid;
#if SUPPORTED_PERSONALITIES > 1
			tcp->currpers = current_personality;
#endif

#ifdef USE_LIBUNWIND
			if (stack_trace_enabled)
				unwind_tcb_init(tcp);
#endif

			nprocs++;
			if (debug_flag)
				error_msg("new tcb for pid %d, active tcbs:%d",
					  tcp->pid, nprocs);
			return tcp;
		}
	}
	error_msg_and_die("bug in alloctcb");
}

static void
droptcb(struct tcb *tcp)
{
	if (tcp->pid == 0)
		return;

#ifdef USE_LIBUNWIND
	if (stack_trace_enabled) {
		unwind_tcb_fin(tcp);
	}
#endif

	nprocs--;
	if (debug_flag)
		error_msg("dropped tcb for pid %d, %d remain",
			  tcp->pid, nprocs);

	if (tcp->outf) {
		if (followfork >= 2) {
			if (tcp->curcol != 0)
				fprintf(tcp->outf, " <detached ...>\n");
			fclose(tcp->outf);
		} else {
			if (printing_tcp == tcp && tcp->curcol != 0)
				fprintf(tcp->outf, " <detached ...>\n");
			fflush(tcp->outf);
		}
	}

	if (current_tcp == tcp)
		current_tcp = NULL;
	if (printing_tcp == tcp)
		printing_tcp = NULL;

	memset(tcp, 0, sizeof(*tcp));
}

/* Detach traced process.
 * Never call DETACH twice on the same process as both unattached and
 * attached-unstopped processes give the same ESRCH.  For unattached process we
 * would SIGSTOP it and wait for its SIGSTOP notification forever.
 */
static void
detach(struct tcb *tcp)
{
	int error;
	int status;

	/*
	 * Linux wrongly insists the child be stopped
	 * before detaching.  Arghh.  We go through hoops
	 * to make a clean break of things.
	 */
#if defined(SPARC)
# undef PTRACE_DETACH
# define PTRACE_DETACH PTRACE_SUNDETACH
#endif

	if (!(tcp->flags & TCB_ATTACHED))
		goto drop;

	/* We attached but possibly didn't see the expected SIGSTOP.
	 * We must catch exactly one as otherwise the detached process
	 * would be left stopped (process state T).
	 */
	if (tcp->flags & TCB_IGNORE_ONE_SIGSTOP)
		goto wait_loop;

	error = ptrace(PTRACE_DETACH, tcp->pid, 0, 0);
	if (!error) {
		/* On a clear day, you can see forever. */
		goto drop;
	}
	if (errno != ESRCH) {
		/* Shouldn't happen. */
		perror_msg("detach: ptrace(PTRACE_DETACH,%u)", tcp->pid);
		goto drop;
	}
	/* ESRCH: process is either not stopped or doesn't exist. */
	if (my_tkill(tcp->pid, 0) < 0) {
		if (errno != ESRCH)
			/* Shouldn't happen. */
			perror_msg("detach: tkill(%u,0)", tcp->pid);
		/* else: process doesn't exist. */
		goto drop;
	}
	/* Process is not stopped, need to stop it. */
	if (use_seize) {
		/*
		 * With SEIZE, tracee can be in group-stop already.
		 * In this state sending it another SIGSTOP does nothing.
		 * Need to use INTERRUPT.
		 * Testcase: trying to ^C a "strace -p <stopped_process>".
		 */
		error = ptrace(PTRACE_INTERRUPT, tcp->pid, 0, 0);
		if (!error)
			goto wait_loop;
		if (errno != ESRCH)
			perror_msg("detach: ptrace(PTRACE_INTERRUPT,%u)", tcp->pid);
	}
	else {
		error = my_tkill(tcp->pid, SIGSTOP);
		if (!error)
			goto wait_loop;
		if (errno != ESRCH)
			perror_msg("detach: tkill(%u,SIGSTOP)", tcp->pid);
	}
	/* Either process doesn't exist, or some weird error. */
	goto drop;

 wait_loop:
	/* We end up here in three cases:
	 * 1. We sent PTRACE_INTERRUPT (use_seize case)
	 * 2. We sent SIGSTOP (!use_seize)
	 * 3. Attach SIGSTOP was already pending (TCB_IGNORE_ONE_SIGSTOP set)
	 */
	for (;;) {
		unsigned int sig;
		if (waitpid(tcp->pid, &status, __WALL) < 0) {
			if (errno == EINTR)
				continue;
			/*
			 * if (errno == ECHILD) break;
			 * ^^^  WRONG! We expect this PID to exist,
			 * and want to emit a message otherwise:
			 */
			perror_msg("detach: waitpid(%u)", tcp->pid);
			break;
		}
		if (!WIFSTOPPED(status)) {
			/*
			 * Tracee exited or was killed by signal.
			 * We shouldn't normally reach this place:
			 * we don't want to consume exit status.
			 * Consider "strace -p PID" being ^C-ed:
			 * we want merely to detach from PID.
			 *
			 * However, we _can_ end up here if tracee
			 * was SIGKILLed.
			 */
			break;
		}
		sig = WSTOPSIG(status);
		if (debug_flag)
			error_msg("detach wait: event:%d sig:%d",
				  (unsigned)status >> 16, sig);
		if (use_seize) {
			unsigned event = (unsigned)status >> 16;
			if (event == PTRACE_EVENT_STOP /*&& sig == SIGTRAP*/) {
				/*
				 * sig == SIGTRAP: PTRACE_INTERRUPT stop.
				 * sig == other: process was already stopped
				 * with this stopping sig (see tests/detach-stopped).
				 * Looks like re-injecting this sig is not necessary
				 * in DETACH for the tracee to remain stopped.
				 */
				sig = 0;
			}
			/*
			 * PTRACE_INTERRUPT is not guaranteed to produce
			 * the above event if other ptrace-stop is pending.
			 * See tests/detach-sleeping testcase:
			 * strace got SIGINT while tracee is sleeping.
			 * We sent PTRACE_INTERRUPT.
			 * We see syscall exit, not PTRACE_INTERRUPT stop.
			 * We won't get PTRACE_INTERRUPT stop
			 * if we would CONT now. Need to DETACH.
			 */
			if (sig == syscall_trap_sig)
				sig = 0;
			/* else: not sure in which case we can be here.
			 * Signal stop? Inject it while detaching.
			 */
			ptrace_restart(PTRACE_DETACH, tcp, sig);
			break;
		}
		/* Note: this check has to be after use_seize check */
		/* (else, in use_seize case SIGSTOP will be mistreated) */
		if (sig == SIGSTOP) {
			/* Detach, suppressing SIGSTOP */
			ptrace_restart(PTRACE_DETACH, tcp, 0);
			break;
		}
		if (sig == syscall_trap_sig)
			sig = 0;
		/* Can't detach just yet, may need to wait for SIGSTOP */
		error = ptrace_restart(PTRACE_CONT, tcp, sig);
		if (error < 0) {
			/* Should not happen.
			 * Note: ptrace_restart returns 0 on ESRCH, so it's not it.
			 * ptrace_restart already emitted error message.
			 */
			break;
		}
	}

 drop:
	if (!qflag && (tcp->flags & TCB_ATTACHED))
		error_msg("Process %u detached", tcp->pid);

	droptcb(tcp);
}

static void
process_opt_p_list(char *opt)
{
	while (*opt) {
		/*
		 * We accept -p PID,PID; -p "`pidof PROG`"; -p "`pgrep PROG`".
		 * pidof uses space as delim, pgrep uses newline. :(
		 */
		int pid;
		char *delim = opt + strcspn(opt, ", \n\t");
		char c = *delim;

		*delim = '\0';
		pid = string_to_uint(opt);
		if (pid <= 0) {
			error_msg_and_die("Invalid process id: '%s'", opt);
		}
		if (pid == strace_tracer_pid) {
			error_msg_and_die("I'm sorry, I can't let you do that, Dave.");
		}
		*delim = c;
		alloctcb(pid);
		if (c == '\0')
			break;
		opt = delim + 1;
	}
}

static void
startup_attach(void)
{
	pid_t parent_pid = strace_tracer_pid;
	unsigned int tcbi;
	struct tcb *tcp;

	/*
	 * Block user interruptions as we would leave the traced
	 * process stopped (process state T) if we would terminate in
	 * between PTRACE_ATTACH and wait4() on SIGSTOP.
	 * We rely on cleanup() from this point on.
	 */
	if (interactive)
		sigprocmask(SIG_BLOCK, &blocked_set, NULL);

	if (daemonized_tracer) {
		pid_t pid = fork();
		if (pid < 0) {
			perror_msg_and_die("fork");
		}
		if (pid) { /* parent */
			/*
			 * Wait for grandchild to attach to straced process
			 * (grandparent). Grandchild SIGKILLs us after it attached.
			 * Grandparent's wait() is unblocked by our death,
			 * it proceeds to exec the straced program.
			 */
			pause();
			_exit(0); /* paranoia */
		}
		/* grandchild */
		/* We will be the tracer process. Remember our new pid: */
		strace_tracer_pid = getpid();
	}

	for (tcbi = 0; tcbi < tcbtabsize; tcbi++) {
		tcp = tcbtab[tcbi];

		if (!tcp->pid)
			continue;

		/* Is this a process we should attach to, but not yet attached? */
		if (tcp->flags & TCB_ATTACHED)
			continue; /* no, we already attached it */

		if (tcp->pid == parent_pid || tcp->pid == strace_tracer_pid) {
			errno = EPERM;
			perror_msg("attach: %d", tcp->pid);
			droptcb(tcp);
			continue;
		}
		if (followfork && tcp->pid != strace_child) {
			char procdir[sizeof("/proc/%d/task") + sizeof(int) * 3];
			DIR *dir;

			sprintf(procdir, "/proc/%d/task", tcp->pid);
			dir = opendir(procdir);
			if (dir != NULL) {
				unsigned int ntid = 0, nerr = 0;
				struct_dirent *de;

				while ((de = read_dir(dir)) != NULL) {
					struct tcb *cur_tcp;
					int tid;

					if (de->d_fileno == 0)
						continue;
					/* we trust /proc filesystem */
					tid = atoi(de->d_name);
					if (tid <= 0)
						continue;
					++ntid;
					if (ptrace_attach_or_seize(tid) < 0) {
						++nerr;
						if (debug_flag)
							error_msg("attach to pid %d failed", tid);
						continue;
					}
					if (debug_flag)
						error_msg("attach to pid %d succeeded", tid);
					cur_tcp = tcp;
					if (tid != tcp->pid)
						cur_tcp = alloctcb(tid);
					cur_tcp->flags |= TCB_ATTACHED | TCB_STARTUP | post_attach_sigstop;
					newoutf(cur_tcp);
				}
				closedir(dir);
				if (interactive) {
					sigprocmask(SIG_SETMASK, &empty_set, NULL);
					if (interrupted)
						goto ret;
					sigprocmask(SIG_BLOCK, &blocked_set, NULL);
				}
				ntid -= nerr;
				if (ntid == 0) {
					perror_msg("attach: ptrace(PTRACE_ATTACH, ...)");
					droptcb(tcp);
					continue;
				}
				if (!qflag) {
					if (ntid > 1)
						error_msg("Process %u attached"
							  " with %u threads",
							  tcp->pid, ntid);
					else
						error_msg("Process %u attached",
							  tcp->pid);
				}
				if (!(tcp->flags & TCB_ATTACHED)) {
					/* -p PID, we failed to attach to PID itself
					 * but did attach to some of its sibling threads.
					 * Drop PID's tcp.
					 */
					droptcb(tcp);
				}
				continue;
			} /* if (opendir worked) */
		} /* if (-f) */
		if (ptrace_attach_or_seize(tcp->pid) < 0) {
			perror_msg("attach: ptrace(PTRACE_ATTACH, ...)");
			droptcb(tcp);
			continue;
		}
		tcp->flags |= TCB_ATTACHED | TCB_STARTUP | post_attach_sigstop;
		newoutf(tcp);
		if (debug_flag)
			error_msg("attach to pid %d (main) succeeded", tcp->pid);

		if (!qflag)
			error_msg("Process %u attached", tcp->pid);
	} /* for each tcbtab[] */

	if (daemonized_tracer) {
		/*
		 * Make parent go away.
		 * Also makes grandparent's wait() unblock.
		 */
		kill(parent_pid, SIGKILL);
		strace_child = 0;
	}

 ret:
	if (interactive)
		sigprocmask(SIG_SETMASK, &empty_set, NULL);
}

/* Stack-o-phobic exec helper, in the hope to work around
 * NOMMU + "daemonized tracer" difficulty.
 */
struct exec_params {
	int fd_to_close;
	uid_t run_euid;
	gid_t run_egid;
	char **argv;
	char *pathname;
};
static struct exec_params params_for_tracee;

static void ATTRIBUTE_NOINLINE ATTRIBUTE_NORETURN
exec_or_die(void)
{
	struct exec_params *params = &params_for_tracee;

	if (params->fd_to_close >= 0)
		close(params->fd_to_close);
	if (!daemonized_tracer && !use_seize) {
		if (ptrace(PTRACE_TRACEME, 0L, 0L, 0L) < 0) {
			perror_msg_and_die("ptrace(PTRACE_TRACEME, ...)");
		}
	}

	if (username != NULL) {
		/*
		 * It is important to set groups before we
		 * lose privileges on setuid.
		 */
		if (initgroups(username, run_gid) < 0) {
			perror_msg_and_die("initgroups");
		}
		if (setregid(run_gid, params->run_egid) < 0) {
			perror_msg_and_die("setregid");
		}
		if (setreuid(run_uid, params->run_euid) < 0) {
			perror_msg_and_die("setreuid");
		}
	}
	else if (geteuid() != 0)
		if (setreuid(run_uid, run_uid) < 0) {
			perror_msg_and_die("setreuid");
		}

	if (!daemonized_tracer) {
		/*
		 * Induce a ptrace stop. Tracer (our parent)
		 * will resume us with PTRACE_SYSCALL and display
		 * the immediately following execve syscall.
		 * Can't do this on NOMMU systems, we are after
		 * vfork: parent is blocked, stopping would deadlock.
		 */
		if (!NOMMU_SYSTEM)
			kill(getpid(), SIGSTOP);
	} else {
		alarm(3);
		/* we depend on SIGCHLD set to SIG_DFL by init code */
		/* if it happens to be SIG_IGN'ed, wait won't block */
		wait(NULL);
		alarm(0);
	}

	execv(params->pathname, params->argv);
	perror_msg_and_die("exec");
}

static int
open_dummy_desc(void)
{
	int fds[2];

	if (pipe(fds))
		perror_msg_and_die("pipe");
	close(fds[1]);
	return fds[0];
}

static void
startup_child(char **argv)
{
	struct_stat statbuf;
	const char *filename;
	size_t filename_len;
	char pathname[PATH_MAX];
	int pid;
	struct tcb *tcp;

	filename = argv[0];
	filename_len = strlen(filename);

	if (filename_len > sizeof(pathname) - 1) {
		errno = ENAMETOOLONG;
		perror_msg_and_die("exec");
	}
	if (strchr(filename, '/')) {
		strcpy(pathname, filename);
	}
#ifdef USE_DEBUGGING_EXEC
	/*
	 * Debuggers customarily check the current directory
	 * first regardless of the path but doing that gives
	 * security geeks a panic attack.
	 */
	else if (stat_file(filename, &statbuf) == 0)
		strcpy(pathname, filename);
#endif /* USE_DEBUGGING_EXEC */
	else {
		const char *path;
		size_t m, n, len;

		for (path = getenv("PATH"); path && *path; path += m) {
			const char *colon = strchr(path, ':');
			if (colon) {
				n = colon - path;
				m = n + 1;
			}
			else
				m = n = strlen(path);
			if (n == 0) {
				if (!getcwd(pathname, PATH_MAX))
					continue;
				len = strlen(pathname);
			}
			else if (n > sizeof pathname - 1)
				continue;
			else {
				strncpy(pathname, path, n);
				len = n;
			}
			if (len && pathname[len - 1] != '/')
				pathname[len++] = '/';
			if (filename_len + len > sizeof(pathname) - 1)
				continue;
			strcpy(pathname + len, filename);
			if (stat_file(pathname, &statbuf) == 0 &&
			    /* Accept only regular files
			       with some execute bits set.
			       XXX not perfect, might still fail */
			    S_ISREG(statbuf.st_mode) &&
			    (statbuf.st_mode & 0111))
				break;
		}
		if (!path || !*path)
			pathname[0] = '\0';
	}
	if (stat_file(pathname, &statbuf) < 0) {
		perror_msg_and_die("Can't stat '%s'", filename);
	}

	params_for_tracee.fd_to_close = (shared_log != stderr) ? fileno(shared_log) : -1;
	params_for_tracee.run_euid = (statbuf.st_mode & S_ISUID) ? statbuf.st_uid : run_uid;
	params_for_tracee.run_egid = (statbuf.st_mode & S_ISGID) ? statbuf.st_gid : run_gid;
	params_for_tracee.argv = argv;
	/*
	 * On NOMMU, can be safely freed only after execve in tracee.
	 * It's hard to know when that happens, so we just leak it.
	 */
	params_for_tracee.pathname = NOMMU_SYSTEM ? xstrdup(pathname) : pathname;

#if defined HAVE_PRCTL && defined PR_SET_PTRACER && defined PR_SET_PTRACER_ANY
	if (daemonized_tracer)
		prctl(PR_SET_PTRACER, PR_SET_PTRACER_ANY);
#endif

	pid = fork();
	if (pid < 0) {
		perror_msg_and_die("fork");
	}
	if ((pid != 0 && daemonized_tracer)
	 || (pid == 0 && !daemonized_tracer)
	) {
		/* We are to become the tracee. Two cases:
		 * -D: we are parent
		 * not -D: we are child
		 */
		exec_or_die();
	}

	/* We are the tracer */

	if (!daemonized_tracer) {
		strace_child = pid;
		if (!use_seize) {
			/* child did PTRACE_TRACEME, nothing to do in parent */
		} else {
			if (!NOMMU_SYSTEM) {
				/* Wait until child stopped itself */
				int status;
				while (waitpid(pid, &status, WSTOPPED) < 0) {
					if (errno == EINTR)
						continue;
					perror_msg_and_die("waitpid");
				}
				if (!WIFSTOPPED(status) || WSTOPSIG(status) != SIGSTOP) {
					kill_save_errno(pid, SIGKILL);
					perror_msg_and_die("Unexpected wait status %x", status);
				}
			}
			/* Else: NOMMU case, we have no way to sync.
			 * Just attach to it as soon as possible.
			 * This means that we may miss a few first syscalls...
			 */

			if (ptrace_attach_or_seize(pid)) {
				kill_save_errno(pid, SIGKILL);
				perror_msg_and_die("Can't attach to %d", pid);
			}
			if (!NOMMU_SYSTEM)
				kill(pid, SIGCONT);
		}
		tcp = alloctcb(pid);
		if (!NOMMU_SYSTEM)
			tcp->flags |= TCB_ATTACHED | TCB_STARTUP | post_attach_sigstop;
		else
			tcp->flags |= TCB_ATTACHED | TCB_STARTUP;
		newoutf(tcp);
	}
	else {
		/* With -D, we are *child* here, the tracee is our parent. */
		strace_child = strace_tracer_pid;
		strace_tracer_pid = getpid();
		alloctcb(strace_child);
		/* attaching will be done later, by startup_attach */
		/* note: we don't do newoutf(tcp) here either! */

		/* NOMMU BUG! -D mode is active, we (child) return,
		 * and we will scribble over parent's stack!
		 * When parent later unpauses, it segfaults.
		 *
		 * We work around it
		 * (1) by declaring exec_or_die() NORETURN,
		 * hopefully compiler will just jump to it
		 * instead of call (won't push anything to stack),
		 * (2) by trying very hard in exec_or_die()
		 * to not use any stack,
		 * (3) having a really big (PATH_MAX) stack object
		 * in this function, which creates a "buffer" between
		 * child's and parent's stack pointers.
		 * This may save us if (1) and (2) failed
		 * and compiler decided to use stack in exec_or_die() anyway
		 * (happens on i386 because of stack parameter passing).
		 *
		 * A cleaner solution is to use makecontext + setcontext
		 * to create a genuine separate stack and execute on it.
		 */
	}
	/*
	 * A case where straced process is part of a pipe:
	 * { sleep 1; yes | head -n99999; } | strace -o/dev/null sh -c 'exec <&-; sleep 9'
	 * If strace won't close its fd#0, closing it in tracee is not enough:
	 * the pipe is still open, it has a reader. Thus, "head" will not get its
	 * SIGPIPE at once, on the first write.
	 *
	 * Preventing it by closing strace's stdin/out.
	 * (Don't leave fds 0 and 1 closed, this is bad practice: future opens
	 * will reuse them, unexpectedly making a newly opened object "stdin").
	 */
	close(0);
	open_dummy_desc(); /* opens to fd#0 */
	dup2(0, 1);
#if 0
	/* A good idea too, but we sometimes need to print error messages */
	if (shared_log != stderr)
		dup2(0, 2);
#endif
}

#if USE_SEIZE
static void
test_ptrace_seize(void)
{
	int pid;

	/* Need fork for test. NOMMU has no forks */
	if (NOMMU_SYSTEM) {
		post_attach_sigstop = 0; /* this sets use_seize to 1 */
		return;
	}

	pid = fork();
	if (pid < 0)
		perror_msg_and_die("fork");

	if (pid == 0) {
		pause();
		_exit(0);
	}

	/* PTRACE_SEIZE, unlike ATTACH, doesn't force tracee to trap.  After
	 * attaching tracee continues to run unless a trap condition occurs.
	 * PTRACE_SEIZE doesn't affect signal or group stop state.
	 */
	if (ptrace(PTRACE_SEIZE, pid, 0, 0) == 0) {
		post_attach_sigstop = 0; /* this sets use_seize to 1 */
	} else if (debug_flag) {
		error_msg("PTRACE_SEIZE doesn't work");
	}

	kill(pid, SIGKILL);

	while (1) {
		int status, tracee_pid;

		errno = 0;
		tracee_pid = waitpid(pid, &status, 0);
		if (tracee_pid <= 0) {
			if (errno == EINTR)
				continue;
			perror_msg_and_die("%s: unexpected wait result %d",
					 __func__, tracee_pid);
		}
		if (WIFSIGNALED(status)) {
			return;
		}
		error_msg_and_die("%s: unexpected wait status %x",
				__func__, status);
	}
}
#else /* !USE_SEIZE */
# define test_ptrace_seize() ((void)0)
#endif

static unsigned
get_os_release(void)
{
	unsigned rel;
	const char *p;
	struct utsname u;
	if (uname(&u) < 0)
		perror_msg_and_die("uname");
	/* u.release has this form: "3.2.9[-some-garbage]" */
	rel = 0;
	p = u.release;
	for (;;) {
		if (!(*p >= '0' && *p <= '9'))
			error_msg_and_die("Bad OS release string: '%s'", u.release);
		/* Note: this open-codes KERNEL_VERSION(): */
		rel = (rel << 8) | atoi(p);
		if (rel >= KERNEL_VERSION(1,0,0))
			break;
		while (*p >= '0' && *p <= '9')
			p++;
		if (*p != '.') {
			if (rel >= KERNEL_VERSION(0,1,0)) {
				/* "X.Y-something" means "X.Y.0" */
				rel <<= 8;
				break;
			}
			error_msg_and_die("Bad OS release string: '%s'", u.release);
		}
		p++;
	}
	return rel;
}

/*
 * Initialization part of main() was eating much stack (~0.5k),
 * which was unused after init.
 * We can reuse it if we move init code into a separate function.
 *
 * Don't want main() to inline us and defeat the reason
 * we have a separate function.
 */
static void ATTRIBUTE_NOINLINE
init(int argc, char *argv[])
{
	struct tcb *tcp;
	int c, i;
	int optF = 0;
	unsigned int tcbi;
	struct sigaction sa;

	progname = argv[0] ? argv[0] : "strace";

	/* Make sure SIGCHLD has the default action so that waitpid
	   definitely works without losing track of children.  The user
	   should not have given us a bogus state to inherit, but he might
	   have.  Arguably we should detect SIG_IGN here and pass it on
	   to children, but probably noone really needs that.  */
	signal(SIGCHLD, SIG_DFL);

	strace_tracer_pid = getpid();

	os_release = get_os_release();

	/* Allocate the initial tcbtab.  */
	tcbtabsize = argc;	/* Surely enough for all -p args.  */
	tcbtab = xcalloc(tcbtabsize, sizeof(tcbtab[0]));
	tcp = xcalloc(tcbtabsize, sizeof(*tcp));
	for (tcbi = 0; tcbi < tcbtabsize; tcbi++)
		tcbtab[tcbi] = tcp++;

	shared_log = stderr;
	set_sortby(DEFAULT_SORTBY);
	set_personality(DEFAULT_PERSONALITY);
	qualify("trace=all");
	qualify("abbrev=all");
	qualify("verbose=all");
#if DEFAULT_QUAL_FLAGS != (QUAL_TRACE | QUAL_ABBREV | QUAL_VERBOSE)
# error Bug in DEFAULT_QUAL_FLAGS
#endif
	qualify("signal=all");
	while ((c = getopt(argc, argv,
		"+b:cCdfFhiqrtTvVwxyz"
#ifdef USE_LIBUNWIND
		"k"
#endif
		"D"
		"a:e:o:O:p:s:S:u:E:P:I:")) != EOF) {
		switch (c) {
		case 'b':
			if (strcmp(optarg, "execve") != 0)
				error_msg_and_die("Syscall '%s' for -b isn't supported",
					optarg);
			detach_on_execve = 1;
			break;
		case 'c':
			if (cflag == CFLAG_BOTH) {
				error_msg_and_help("-c and -C are mutually exclusive");
			}
			cflag = CFLAG_ONLY_STATS;
			break;
		case 'C':
			if (cflag == CFLAG_ONLY_STATS) {
				error_msg_and_help("-c and -C are mutually exclusive");
			}
			cflag = CFLAG_BOTH;
			break;
		case 'd':
			debug_flag = 1;
			break;
		case 'D':
			daemonized_tracer = 1;
			break;
		case 'F':
			optF = 1;
			break;
		case 'f':
			followfork++;
			break;
		case 'h':
			usage();
			break;
		case 'i':
			iflag = 1;
			break;
		case 'q':
			qflag++;
			break;
		case 'r':
			rflag = 1;
			/* fall through to tflag++ */
		case 't':
			tflag++;
			break;
		case 'T':
			Tflag = 1;
			break;
		case 'w':
			count_wallclock = 1;
			break;
		case 'x':
			xflag++;
			break;
		case 'y':
			show_fd_path++;
			break;
		case 'v':
			qualify("abbrev=none");
			break;
		case 'V':
			printf("%s -- version %s\n", PACKAGE_NAME, VERSION);
			exit(0);
			break;
		case 'z':
			not_failing_only = 1;
			break;
		case 'a':
			acolumn = string_to_uint(optarg);
			if (acolumn < 0)
				error_opt_arg(c, optarg);
			break;
		case 'e':
			qualify(optarg);
			break;
		case 'o':
			outfname = xstrdup(optarg);
			break;
		case 'O':
			i = string_to_uint(optarg);
			if (i < 0)
				error_opt_arg(c, optarg);
			set_overhead(i);
			break;
		case 'p':
			process_opt_p_list(optarg);
			break;
		case 'P':
			pathtrace_select(optarg);
			break;
		case 's':
			i = string_to_uint(optarg);
			if (i < 0)
				error_opt_arg(c, optarg);
			max_strlen = i;
			break;
		case 'S':
			set_sortby(optarg);
			break;
		case 'u':
			username = xstrdup(optarg);
			break;
#ifdef USE_LIBUNWIND
		case 'k':
			stack_trace_enabled = true;
			break;
#endif
		case 'E':
			if (putenv(optarg) < 0)
				die_out_of_memory();
			break;
		case 'I':
			opt_intr = string_to_uint(optarg);
			if (opt_intr <= 0 || opt_intr >= NUM_INTR_OPTS)
				error_opt_arg(c, optarg);
			break;
		default:
			error_msg_and_help(NULL);
			break;
		}
	}
	argv += optind;
	/* argc -= optind; - no need, argc is not used below */

	acolumn_spaces = xmalloc(acolumn + 1);
	memset(acolumn_spaces, ' ', acolumn);
	acolumn_spaces[acolumn] = '\0';

	if (!argv[0] && !nprocs) {
		error_msg_and_help("must have PROG [ARGS] or -p PID");
	}

	if (!argv[0] && daemonized_tracer) {
		error_msg_and_help("PROG [ARGS] must be specified with -D");
	}

	if (!followfork)
		followfork = optF;

	if (followfork >= 2 && cflag) {
		error_msg_and_help("(-c or -C) and -ff are mutually exclusive");
	}

	if (count_wallclock && !cflag) {
		error_msg_and_help("-w must be given with (-c or -C)");
	}

	if (cflag == CFLAG_ONLY_STATS) {
		if (iflag)
			error_msg("-%c has no effect with -c", 'i');
#ifdef USE_LIBUNWIND
		if (stack_trace_enabled)
			error_msg("-%c has no effect with -c", 'k');
#endif
		if (rflag)
			error_msg("-%c has no effect with -c", 'r');
		if (tflag)
			error_msg("-%c has no effect with -c", 't');
		if (Tflag)
			error_msg("-%c has no effect with -c", 'T');
		if (show_fd_path)
			error_msg("-%c has no effect with -c", 'y');
	}

#ifdef USE_LIBUNWIND
	if (stack_trace_enabled)
		unwind_init();
#endif

	/* See if they want to run as another user. */
	if (username != NULL) {
		struct passwd *pent;

		if (getuid() != 0 || geteuid() != 0) {
			error_msg_and_die("You must be root to use the -u option");
		}
		pent = getpwnam(username);
		if (pent == NULL) {
			error_msg_and_die("Cannot find user '%s'", username);
		}
		run_uid = pent->pw_uid;
		run_gid = pent->pw_gid;
	}
	else {
		run_uid = getuid();
		run_gid = getgid();
	}

	if (followfork)
		ptrace_setoptions |= PTRACE_O_TRACECLONE |
				     PTRACE_O_TRACEFORK |
				     PTRACE_O_TRACEVFORK;
	if (debug_flag)
		error_msg("ptrace_setoptions = %#x", ptrace_setoptions);
	test_ptrace_seize();

	if (fcntl(0, F_GETFD) == -1 || fcntl(1, F_GETFD) == -1) {
		/*
		 * Something weird with our stdin and/or stdout -
		 * for example, may be not open? In this case,
		 * ensure that none of the future opens uses them.
		 *
		 * This was seen in the wild when /proc/sys/kernel/core_pattern
		 * was set to "|/bin/strace -o/tmp/LOG PROG":
		 * kernel runs coredump helper with fd#0 open but fd#1 closed (!),
		 * therefore LOG gets opened to fd#1, and fd#1 is closed by
		 * "don't hold up stdin/out open" code soon after.
		 */
		int fd = open_dummy_desc();
		while (fd >= 0 && fd < 2)
			fd = dup(fd);
		if (fd > 2)
			close(fd);
	}

	/* Check if they want to redirect the output. */
	if (outfname) {
		/* See if they want to pipe the output. */
		if (outfname[0] == '|' || outfname[0] == '!') {
			/*
			 * We can't do the <outfname>.PID funny business
			 * when using popen, so prohibit it.
			 */
			if (followfork >= 2)
				error_msg_and_help("piping the output and -ff are mutually exclusive");
			shared_log = strace_popen(outfname + 1);
		}
		else if (followfork < 2)
			shared_log = strace_fopen(outfname);
	} else {
		/* -ff without -o FILE is the same as single -f */
		if (followfork >= 2)
			followfork = 1;
	}

	if (!outfname || outfname[0] == '|' || outfname[0] == '!') {
		char *buf = xmalloc(BUFSIZ);
		setvbuf(shared_log, buf, _IOLBF, BUFSIZ);
	}
	if (outfname && argv[0]) {
		if (!opt_intr)
			opt_intr = INTR_NEVER;
		if (!qflag)
			qflag = 1;
	}
	if (!opt_intr)
		opt_intr = INTR_WHILE_WAIT;

	/* argv[0]	-pPID	-oFILE	Default interactive setting
	 * yes		*	0	INTR_WHILE_WAIT
	 * no		1	0	INTR_WHILE_WAIT
	 * yes		*	1	INTR_NEVER
	 * no		1	1	INTR_WHILE_WAIT
	 */

	sigemptyset(&empty_set);
	sigemptyset(&blocked_set);

	/* startup_child() must be called before the signal handlers get
	 * installed below as they are inherited into the spawned process.
	 * Also we do not need to be protected by them as during interruption
	 * in the startup_child() mode we kill the spawned process anyway.
	 */
	if (argv[0]) {
		if (!NOMMU_SYSTEM || daemonized_tracer)
			hide_log_until_execve = 1;
		skip_one_b_execve = 1;
		startup_child(argv);
	}

	sa.sa_handler = SIG_IGN;
	sigemptyset(&sa.sa_mask);
	sa.sa_flags = 0;
	sigaction(SIGTTOU, &sa, NULL); /* SIG_IGN */
	sigaction(SIGTTIN, &sa, NULL); /* SIG_IGN */
	if (opt_intr != INTR_ANYWHERE) {
		if (opt_intr == INTR_BLOCK_TSTP_TOO)
			sigaction(SIGTSTP, &sa, NULL); /* SIG_IGN */
		/*
		 * In interactive mode (if no -o OUTFILE, or -p PID is used),
		 * fatal signals are blocked while syscall stop is processed,
		 * and acted on in between, when waiting for new syscall stops.
		 * In non-interactive mode, signals are ignored.
		 */
		if (opt_intr == INTR_WHILE_WAIT) {
			sigaddset(&blocked_set, SIGHUP);
			sigaddset(&blocked_set, SIGINT);
			sigaddset(&blocked_set, SIGQUIT);
			sigaddset(&blocked_set, SIGPIPE);
			sigaddset(&blocked_set, SIGTERM);
			sa.sa_handler = interrupt;
		}
		/* SIG_IGN, or set handler for these */
		sigaction(SIGHUP, &sa, NULL);
		sigaction(SIGINT, &sa, NULL);
		sigaction(SIGQUIT, &sa, NULL);
		sigaction(SIGPIPE, &sa, NULL);
		sigaction(SIGTERM, &sa, NULL);
	}
	if (nprocs != 0 || daemonized_tracer)
		startup_attach();

	/* Do we want pids printed in our -o OUTFILE?
	 * -ff: no (every pid has its own file); or
	 * -f: yes (there can be more pids in the future); or
	 * -p PID1,PID2: yes (there are already more than one pid)
	 */
	print_pid_pfx = (outfname && followfork < 2 && (followfork == 1 || nprocs > 1));
}

static struct tcb *
pid2tcb(int pid)
{
	unsigned int i;

	if (pid <= 0)
		return NULL;

	for (i = 0; i < tcbtabsize; i++) {
		struct tcb *tcp = tcbtab[i];
		if (tcp->pid == pid)
			return tcp;
	}

	return NULL;
}

static void
cleanup(void)
{
	unsigned int i;
	struct tcb *tcp;
	int fatal_sig;

	/* 'interrupted' is a volatile object, fetch it only once */
	fatal_sig = interrupted;
	if (!fatal_sig)
		fatal_sig = SIGTERM;

	for (i = 0; i < tcbtabsize; i++) {
		tcp = tcbtab[i];
		if (!tcp->pid)
			continue;
		if (debug_flag)
			error_msg("cleanup: looking at pid %u", tcp->pid);
		if (tcp->pid == strace_child) {
			kill(tcp->pid, SIGCONT);
			kill(tcp->pid, fatal_sig);
		}
		detach(tcp);
	}
	if (cflag)
		call_summary(shared_log);
}

static void
interrupt(int sig)
{
	interrupted = sig;
}

static void
print_debug_info(const int pid, int status)
{
	const unsigned int event = (unsigned int) status >> 16;
	char buf[sizeof("WIFEXITED,exitcode=%u") + sizeof(int)*3 /*paranoia:*/ + 16];
	char evbuf[sizeof(",EVENT_VFORK_DONE (%u)") + sizeof(int)*3 /*paranoia:*/ + 16];

	strcpy(buf, "???");
	if (WIFSIGNALED(status))
#ifdef WCOREDUMP
		sprintf(buf, "WIFSIGNALED,%ssig=%s",
				WCOREDUMP(status) ? "core," : "",
				signame(WTERMSIG(status)));
#else
		sprintf(buf, "WIFSIGNALED,sig=%s",
				signame(WTERMSIG(status)));
#endif
	if (WIFEXITED(status))
		sprintf(buf, "WIFEXITED,exitcode=%u", WEXITSTATUS(status));
	if (WIFSTOPPED(status))
		sprintf(buf, "WIFSTOPPED,sig=%s", signame(WSTOPSIG(status)));
#ifdef WIFCONTINUED
	/* Should never be seen */
	if (WIFCONTINUED(status))
		strcpy(buf, "WIFCONTINUED");
#endif
	evbuf[0] = '\0';
	if (event != 0) {
		static const char *const event_names[] = {
			[PTRACE_EVENT_CLONE] = "CLONE",
			[PTRACE_EVENT_FORK]  = "FORK",
			[PTRACE_EVENT_VFORK] = "VFORK",
			[PTRACE_EVENT_VFORK_DONE] = "VFORK_DONE",
			[PTRACE_EVENT_EXEC]  = "EXEC",
			[PTRACE_EVENT_EXIT]  = "EXIT",
			/* [PTRACE_EVENT_STOP (=128)] would make biggish array */
		};
		const char *e = "??";
		if (event < ARRAY_SIZE(event_names))
			e = event_names[event];
		else if (event == PTRACE_EVENT_STOP)
			e = "STOP";
		sprintf(evbuf, ",EVENT_%s (%u)", e, event);
	}
	error_msg("[wait(0x%06x) = %u] %s%s", status, pid, buf, evbuf);
}

static struct tcb *
maybe_allocate_tcb(const int pid, int status)
{
	if (!WIFSTOPPED(status)) {
		if (detach_on_execve && pid == strace_child) {
			/* example: strace -bexecve sh -c 'exec true' */
			strace_child = 0;
			return NULL;
		}
		/*
		 * This can happen if we inherited an unknown child.
		 * Example: (sleep 1 & exec strace true)
		 */
		error_msg("Exit of unknown pid %u ignored", pid);
		return NULL;
	}
	if (followfork) {
		/* We assume it's a fork/vfork/clone child */
		struct tcb *tcp = alloctcb(pid);
		tcp->flags |= TCB_ATTACHED | TCB_STARTUP | post_attach_sigstop;
		newoutf(tcp);
		if (!qflag)
			error_msg("Process %d attached", pid);
		return tcp;
	} else {
		/* This can happen if a clone call used
		 * CLONE_PTRACE itself.
		 */
		ptrace(PTRACE_CONT, pid, (char *) 0, 0);
		error_msg("Stop of unknown pid %u seen, PTRACE_CONTed it", pid);
		return NULL;
	}
}

static struct tcb *
maybe_switch_tcbs(struct tcb *tcp, const int pid)
{
	FILE *fp;
	struct tcb *execve_thread;
	long old_pid = 0;

	if (ptrace(PTRACE_GETEVENTMSG, pid, NULL, (long) &old_pid) < 0)
		return tcp;
	/* Avoid truncation in pid2tcb() param passing */
	if (old_pid <= 0 || old_pid == pid)
		return tcp;
	if ((unsigned long) old_pid > UINT_MAX)
		return tcp;
	execve_thread = pid2tcb(old_pid);
	/* It should be !NULL, but I feel paranoid */
	if (!execve_thread)
		return tcp;

	if (execve_thread->curcol != 0) {
		/*
		 * One case we are here is -ff:
		 * try "strace -oLOG -ff test/threaded_execve"
		 */
		fprintf(execve_thread->outf, " <pid changed to %d ...>\n", pid);
		/*execve_thread->curcol = 0; - no need, see code below */
	}
	/* Swap output FILEs (needed for -ff) */
	fp = execve_thread->outf;
	execve_thread->outf = tcp->outf;
	tcp->outf = fp;
	/* And their column positions */
	execve_thread->curcol = tcp->curcol;
	tcp->curcol = 0;
	/* Drop leader, but close execve'd thread outfile (if -ff) */
	droptcb(tcp);
	/* Switch to the thread, reusing leader's outfile and pid */
	tcp = execve_thread;
	tcp->pid = pid;
	if (cflag != CFLAG_ONLY_STATS) {
		printleader(tcp);
		tprintf("+++ superseded by execve in pid %lu +++\n", old_pid);
		line_ended();
		tcp->flags |= TCB_REPRINT;
	}

	return tcp;
}

static void
print_signalled(struct tcb *tcp, const int pid, int status)
{
	if (pid == strace_child) {
		exit_code = 0x100 | WTERMSIG(status);
		strace_child = 0;
	}

	if (cflag != CFLAG_ONLY_STATS
	 && (qual_flags[WTERMSIG(status)] & QUAL_SIGNAL)
	) {
		printleader(tcp);
#ifdef WCOREDUMP
		tprintf("+++ killed by %s %s+++\n",
			signame(WTERMSIG(status)),
			WCOREDUMP(status) ? "(core dumped) " : "");
#else
		tprintf("+++ killed by %s +++\n",
			signame(WTERMSIG(status)));
#endif
		line_ended();
	}
}

static void
print_exited(struct tcb *tcp, const int pid, int status)
{
	if (pid == strace_child) {
		exit_code = WEXITSTATUS(status);
		strace_child = 0;
	}

	if (cflag != CFLAG_ONLY_STATS &&
	    qflag < 2) {
		printleader(tcp);
		tprintf("+++ exited with %d +++\n", WEXITSTATUS(status));
		line_ended();
	}
}

static void
print_stopped(struct tcb *tcp, const siginfo_t *si, const unsigned int sig)
{
	if (cflag != CFLAG_ONLY_STATS
	    && !hide_log_until_execve
	    && (qual_flags[sig] & QUAL_SIGNAL)
	   ) {
		printleader(tcp);
		if (si) {
			tprintf("--- %s ", signame(sig));
			printsiginfo(si, verbose(tcp));
			tprints(" ---\n");
		} else
			tprintf("--- stopped by %s ---\n", signame(sig));
		line_ended();
	}
}

static void
startup_tcb(struct tcb *tcp)
{
	if (debug_flag)
		error_msg("pid %d has TCB_STARTUP, initializing it", tcp->pid);

	tcp->flags &= ~TCB_STARTUP;

	if (!use_seize) {
		if (debug_flag)
			error_msg("setting opts 0x%x on pid %d",
				  ptrace_setoptions, tcp->pid);
		if (ptrace(PTRACE_SETOPTIONS, tcp->pid, NULL, ptrace_setoptions) < 0) {
			if (errno != ESRCH) {
				/* Should never happen, really */
				perror_msg_and_die("PTRACE_SETOPTIONS");
			}
		}
	}
}

/* Returns true iff the main trace loop has to continue. */
static bool
trace(void)
{
	int pid;
	int wait_errno;
	int status;
	bool stopped;
	unsigned int sig;
	unsigned int event;
	struct tcb *tcp;
	struct rusage ru;

	if (interrupted)
		return false;

	/*
	 * Used to exit simply when nprocs hits zero, but in this testcase:
	 *  int main() { _exit(!!fork()); }
	 * under strace -f, parent sometimes (rarely) manages
	 * to exit before we see the first stop of the child,
	 * and we are losing track of it:
	 *  19923 clone(...) = 19924
	 *  19923 exit_group(1)     = ?
	 *  19923 +++ exited with 1 +++
	 * Exiting only when wait() returns ECHILD works better.
	 */
	if (popen_pid != 0) {
		/* However, if -o|logger is in use, we can't do that.
		 * Can work around that by double-forking the logger,
		 * but that loses the ability to wait for its completion
		 * on exit. Oh well...
		 */
		if (nprocs == 0)
			return false;
	}

	if (interactive)
		sigprocmask(SIG_SETMASK, &empty_set, NULL);
	pid = wait4(-1, &status, __WALL, (cflag ? &ru : NULL));
	wait_errno = errno;
	if (interactive)
		sigprocmask(SIG_BLOCK, &blocked_set, NULL);

	if (pid < 0) {
		if (wait_errno == EINTR)
			return true;
		if (nprocs == 0 && wait_errno == ECHILD)
			return false;
		/*
		 * If nprocs > 0, ECHILD is not expected,
		 * treat it as any other error here:
		 */
		errno = wait_errno;
		perror_msg_and_die("wait4(__WALL)");
	}

	if (pid == popen_pid) {
		if (!WIFSTOPPED(status))
			popen_pid = 0;
		return true;
	}

	if (debug_flag)
		print_debug_info(pid, status);

	/* Look up 'pid' in our table. */
	tcp = pid2tcb(pid);

	if (!tcp) {
		tcp = maybe_allocate_tcb(pid, status);
		if (!tcp)
			return true;
	}

	if (WIFSTOPPED(status))
		get_regs(pid);
	else
		clear_regs();

	event = (unsigned int) status >> 16;

	if (event == PTRACE_EVENT_EXEC) {
		/*
		 * Under Linux, execve changes pid to thread leader's pid,
		 * and we see this changed pid on EVENT_EXEC and later,
		 * execve sysexit. Leader "disappears" without exit
		 * notification. Let user know that, drop leader's tcb,
		 * and fix up pid in execve thread's tcb.
		 * Effectively, execve thread's tcb replaces leader's tcb.
		 *
		 * BTW, leader is 'stuck undead' (doesn't report WIFEXITED
		 * on exit syscall) in multithreaded programs exactly
		 * in order to handle this case.
		 *
		 * PTRACE_GETEVENTMSG returns old pid starting from Linux 3.0.
		 * On 2.6 and earlier, it can return garbage.
		 */
		if (os_release >= KERNEL_VERSION(3,0,0))
			tcp = maybe_switch_tcbs(tcp, pid);

		if (detach_on_execve && !skip_one_b_execve) {
			detach(tcp); /* do "-b execve" thingy */
			return true;
		}
		skip_one_b_execve = 0;
	}

	/* Set current output file */
	current_tcp = tcp;

	if (cflag) {
		tv_sub(&tcp->dtime, &ru.ru_stime, &tcp->stime);
		tcp->stime = ru.ru_stime;
	}

	if (WIFSIGNALED(status)) {
		print_signalled(tcp, pid, status);
		droptcb(tcp);
		return true;
	}

	if (WIFEXITED(status)) {
		print_exited(tcp, pid, status);
		droptcb(tcp);
		return true;
	}

	if (!WIFSTOPPED(status)) {
		/*
		 * Neither signalled, exited or stopped.
		 * How could that be?
		 */
		error_msg("pid %u not stopped!", pid);
		droptcb(tcp);
		return true;
	}

	/* Is this the very first time we see this tracee stopped? */
	if (tcp->flags & TCB_STARTUP) {
		startup_tcb(tcp);
		if (get_scno(tcp) == 1)
			tcp->s_prev_ent = tcp->s_ent;
	}

	sig = WSTOPSIG(status);

	if (event != 0) {
		/* Ptrace event */
#if USE_SEIZE
		if (event == PTRACE_EVENT_STOP) {
			/*
			 * PTRACE_INTERRUPT-stop or group-stop.
			 * PTRACE_INTERRUPT-stop has sig == SIGTRAP here.
			 */
			switch (sig) {
				case SIGSTOP:
				case SIGTSTP:
				case SIGTTIN:
				case SIGTTOU:
					stopped = true;
					goto show_stopsig;
			}
		}
#endif
		goto restart_tracee_with_sig_0;
	}

	/*
	 * Is this post-attach SIGSTOP?
	 * Interestingly, the process may stop
	 * with STOPSIG equal to some other signal
	 * than SIGSTOP if we happend to attach
	 * just before the process takes a signal.
	 */
	if (sig == SIGSTOP && (tcp->flags & TCB_IGNORE_ONE_SIGSTOP)) {
		if (debug_flag)
			error_msg("ignored SIGSTOP on pid %d", tcp->pid);
		tcp->flags &= ~TCB_IGNORE_ONE_SIGSTOP;
		goto restart_tracee_with_sig_0;
	}

	if (sig != syscall_trap_sig) {
		siginfo_t si = {};

		/*
		 * True if tracee is stopped by signal
		 * (as opposed to "tracee received signal").
		 * TODO: shouldn't we check for errno == EINVAL too?
		 * We can get ESRCH instead, you know...
		 */
		stopped = ptrace(PTRACE_GETSIGINFO, pid, 0, (long) &si) < 0;
#if USE_SEIZE
show_stopsig:
#endif
		print_stopped(tcp, stopped ? NULL : &si, sig);

		if (!stopped)
			/* It's signal-delivery-stop. Inject the signal */
			goto restart_tracee;

		/* It's group-stop */
		if (use_seize) {
			/*
			 * This ends ptrace-stop, but does *not* end group-stop.
			 * This makes stopping signals work properly on straced process
			 * (that is, process really stops. It used to continue to run).
			 */
			if (ptrace_restart(PTRACE_LISTEN, tcp, 0) < 0) {
				/* Note: ptrace_restart emitted error message */
				exit_code = 1;
				return false;
			}
			return true;
		}
		/* We don't have PTRACE_LISTEN support... */
		goto restart_tracee;
	}

	/* We handled quick cases, we are permitted to interrupt now. */
	if (interrupted)
		return false;

	/*
	 * This should be syscall entry or exit.
	 * Handle it.
	 */
	if (trace_syscall(tcp) < 0) {
		/*
		 * ptrace() failed in trace_syscall().
		 * Likely a result of process disappearing mid-flight.
		 * Observed case: exit_group() or SIGKILL terminating
		 * all processes in thread group.
		 * We assume that ptrace error was caused by process death.
		 * We used to detach(tcp) here, but since we no longer
		 * implement "detach before death" policy/hack,
		 * we can let this process to report its death to us
		 * normally, via WIFEXITED or WIFSIGNALED wait status.
		 */
		return true;
	}

restart_tracee_with_sig_0:
	sig = 0;

restart_tracee:
	if (ptrace_restart(PTRACE_SYSCALL, tcp, sig) < 0) {
		/* Note: ptrace_restart emitted error message */
		exit_code = 1;
		return false;
	}

	return true;
}

int
main(int argc, char *argv[])
{
	init(argc, argv);

	while (trace())
		;

	cleanup();
	fflush(NULL);
	if (shared_log != stderr)
		fclose(shared_log);
	if (popen_pid) {
		while (waitpid(popen_pid, NULL, 0) < 0 && errno == EINTR)
			;
	}
	if (exit_code > 0xff) {
		/* Avoid potential core file clobbering.  */
		struct_rlimit rlim = {0, 0};
		set_rlimit(RLIMIT_CORE, &rlim);

		/* Child was killed by a signal, mimic that.  */
		exit_code &= 0xff;
		signal(exit_code, SIG_DFL);
		raise(exit_code);
		/* Paranoia - what if this signal is not fatal?
		   Exit with 128 + signo then.  */
		exit_code += 128;
	}

	return exit_code;
}