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 
46 #include "ptrace.h"
47 
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 #ifdef USE_LIBUNWIND
54 /* if this is true do the stack trace for every system call */
55 bool stack_trace_enabled = false;
56 #endif
57 
58 #if defined __NR_tkill
59 # define my_tkill(tid, sig) syscall(__NR_tkill, (tid), (sig))
60 #else
61    /* kill() may choose arbitrarily the target task of the process group
62       while we later wait on a that specific TID.  PID process waits become
63       TID task specific waits for a process under ptrace(2).  */
64 # warning "tkill(2) not available, risk of strace hangs!"
65 # define my_tkill(tid, sig) kill((tid), (sig))
66 #endif
67 
68 /* Glue for systems without a MMU that cannot provide fork() */
69 #if !defined(HAVE_FORK)
70 # undef NOMMU_SYSTEM
71 # define NOMMU_SYSTEM 1
72 #endif
73 #if NOMMU_SYSTEM
74 # define fork() vfork()
75 #endif
76 
77 const unsigned int syscall_trap_sig = SIGTRAP | 0x80;
78 
79 cflag_t cflag = CFLAG_NONE;
80 unsigned int followfork = 0;
81 unsigned int ptrace_setoptions = PTRACE_O_TRACESYSGOOD | PTRACE_O_TRACEEXEC;
82 unsigned int xflag = 0;
83 bool debug_flag = 0;
84 bool Tflag = 0;
85 bool iflag = 0;
86 bool count_wallclock = 0;
87 unsigned int qflag = 0;
88 static unsigned int tflag = 0;
89 static bool rflag = 0;
90 static bool print_pid_pfx = 0;
91 
92 /* -I n */
93 enum {
94     INTR_NOT_SET        = 0,
95     INTR_ANYWHERE       = 1, /* don't block/ignore any signals */
96     INTR_WHILE_WAIT     = 2, /* block fatal signals while decoding syscall. default */
97     INTR_NEVER          = 3, /* block fatal signals. default if '-o FILE PROG' */
98     INTR_BLOCK_TSTP_TOO = 4, /* block fatal signals and SIGTSTP (^Z) */
99     NUM_INTR_OPTS
100 };
101 static int opt_intr;
102 /* We play with signal mask only if this mode is active: */
103 #define interactive (opt_intr == INTR_WHILE_WAIT)
104 
105 /*
106  * daemonized_tracer supports -D option.
107  * With this option, strace forks twice.
108  * Unlike normal case, with -D *grandparent* process exec's,
109  * becoming a traced process. Child exits (this prevents traced process
110  * from having children it doesn't expect to have), and grandchild
111  * attaches to grandparent similarly to strace -p PID.
112  * This allows for more transparent interaction in cases
113  * when process and its parent are communicating via signals,
114  * wait() etc. Without -D, strace process gets lodged in between,
115  * disrupting parent<->child link.
116  */
117 static bool daemonized_tracer = 0;
118 
119 #if USE_SEIZE
120 static int post_attach_sigstop = TCB_IGNORE_ONE_SIGSTOP;
121 # define use_seize (post_attach_sigstop == 0)
122 #else
123 # define post_attach_sigstop TCB_IGNORE_ONE_SIGSTOP
124 # define use_seize 0
125 #endif
126 
127 /* Sometimes we want to print only succeeding syscalls. */
128 bool not_failing_only = 0;
129 
130 /* Show path associated with fd arguments */
131 unsigned int show_fd_path = 0;
132 
133 static bool detach_on_execve = 0;
134 /* Are we "strace PROG" and need to skip detach on first execve? */
135 static bool skip_one_b_execve = 0;
136 /* Are we "strace PROG" and need to hide everything until execve? */
137 bool hide_log_until_execve = 0;
138 
139 static int exit_code = 0;
140 static int strace_child = 0;
141 static int strace_tracer_pid = 0;
142 
143 static char *username = NULL;
144 static uid_t run_uid;
145 static gid_t run_gid;
146 
147 unsigned int max_strlen = DEFAULT_STRLEN;
148 static int acolumn = DEFAULT_ACOLUMN;
149 static char *acolumn_spaces;
150 
151 static char *outfname = NULL;
152 /* If -ff, points to stderr. Else, it's our common output log */
153 static FILE *shared_log;
154 
155 struct tcb *printing_tcp = NULL;
156 static struct tcb *current_tcp;
157 
158 static struct tcb **tcbtab;
159 static unsigned int nprocs, tcbtabsize;
160 static const char *progname;
161 
162 unsigned os_release; /* generated from uname()'s u.release */
163 
164 static void detach(struct tcb *tcp);
165 static void cleanup(void);
166 static void interrupt(int sig);
167 static sigset_t empty_set, blocked_set;
168 
169 #ifdef HAVE_SIG_ATOMIC_T
170 static volatile sig_atomic_t interrupted;
171 #else
172 static volatile int interrupted;
173 #endif
174 
175 #ifndef HAVE_STRERROR
176 
177 #if !HAVE_DECL_SYS_ERRLIST
178 extern int sys_nerr;
179 extern char *sys_errlist[];
180 #endif
181 
182 const char *
strerror(int err_no)183 strerror(int err_no)
184 {
185 	static char buf[sizeof("Unknown error %d") + sizeof(int)*3];
186 
187 	if (err_no < 1 || err_no >= sys_nerr) {
188 		sprintf(buf, "Unknown error %d", err_no);
189 		return buf;
190 	}
191 	return sys_errlist[err_no];
192 }
193 
194 #endif /* HAVE_STERRROR */
195 
196 static void
usage(FILE * ofp,int exitval)197 usage(FILE *ofp, int exitval)
198 {
199 	fprintf(ofp, "\
200 usage: strace [-CdffhiqrtttTvVxxy] [-I n] [-e expr]...\n\
201               [-a column] [-o file] [-s strsize] [-P path]...\n\
202               -p pid... / [-D] [-E var=val]... [-u username] PROG [ARGS]\n\
203    or: strace -c[df] [-I n] [-e expr]... [-O overhead] [-S sortby]\n\
204               -p pid... / [-D] [-E var=val]... [-u username] PROG [ARGS]\n\
205 -c -- count time, calls, and errors for each syscall and report summary\n\
206 -C -- like -c but also print regular output\n\
207 -w -- summarise syscall latency (default is system time)\n\
208 -d -- enable debug output to stderr\n\
209 -D -- run tracer process as a detached grandchild, not as parent\n\
210 -f -- follow forks, -ff -- with output into separate files\n\
211 -i -- print instruction pointer at time of syscall\n\
212 -q -- suppress messages about attaching, detaching, etc.\n\
213 -r -- print relative timestamp, -t -- absolute timestamp, -tt -- with usecs\n\
214 -T -- print time spent in each syscall\n\
215 -v -- verbose mode: print unabbreviated argv, stat, termios, etc. args\n\
216 -x -- print non-ascii strings in hex, -xx -- print all strings in hex\n\
217 -y -- print paths associated with file descriptor arguments\n\
218 -yy -- print ip:port pairs associated with socket file descriptors\n\
219 -h -- print help message, -V -- print version\n\
220 -a column -- alignment COLUMN for printing syscall results (default %d)\n\
221 -b execve -- detach on this syscall\n\
222 -e expr -- a qualifying expression: option=[!]all or option=[!]val1[,val2]...\n\
223    options: trace, abbrev, verbose, raw, signal, read, write\n\
224 -I interruptible --\n\
225    1: no signals are blocked\n\
226    2: fatal signals are blocked while decoding syscall (default)\n\
227    3: fatal signals are always blocked (default if '-o FILE PROG')\n\
228    4: fatal signals and SIGTSTP (^Z) are always blocked\n\
229       (useful to make 'strace -o FILE PROG' not stop on ^Z)\n\
230 -o file -- send trace output to FILE instead of stderr\n\
231 -O overhead -- set overhead for tracing syscalls to OVERHEAD usecs\n\
232 -p pid -- trace process with process id PID, may be repeated\n\
233 -s strsize -- limit length of print strings to STRSIZE chars (default %d)\n\
234 -S sortby -- sort syscall counts by: time, calls, name, nothing (default %s)\n\
235 -u username -- run command as username handling setuid and/or setgid\n\
236 -E var=val -- put var=val in the environment for command\n\
237 -E var -- remove var from the environment for command\n\
238 -P path -- trace accesses to path\n\
239 "
240 #ifdef USE_LIBUNWIND
241 "-k obtain stack trace between each syscall (experimental)\n\
242 "
243 #endif
244 /* ancient, no one should use it
245 -F -- attempt to follow vforks (deprecated, use -f)\n\
246  */
247 /* this is broken, so don't document it
248 -z -- print only succeeding syscalls\n\
249  */
250 , DEFAULT_ACOLUMN, DEFAULT_STRLEN, DEFAULT_SORTBY);
251 	exit(exitval);
252 }
253 
254 static void ATTRIBUTE_NORETURN
die(void)255 die(void)
256 {
257 	if (strace_tracer_pid == getpid()) {
258 		cflag = 0;
259 		cleanup();
260 	}
261 	exit(1);
262 }
263 
verror_msg(int err_no,const char * fmt,va_list p)264 static void verror_msg(int err_no, const char *fmt, va_list p)
265 {
266 	char *msg;
267 
268 	fflush(NULL);
269 
270 	/* We want to print entire message with single fprintf to ensure
271 	 * message integrity if stderr is shared with other programs.
272 	 * Thus we use vasprintf + single fprintf.
273 	 */
274 	msg = NULL;
275 	if (vasprintf(&msg, fmt, p) >= 0) {
276 		if (err_no)
277 			fprintf(stderr, "%s: %s: %s\n", progname, msg, strerror(err_no));
278 		else
279 			fprintf(stderr, "%s: %s\n", progname, msg);
280 		free(msg);
281 	} else {
282 		/* malloc in vasprintf failed, try it without malloc */
283 		fprintf(stderr, "%s: ", progname);
284 		vfprintf(stderr, fmt, p);
285 		if (err_no)
286 			fprintf(stderr, ": %s\n", strerror(err_no));
287 		else
288 			putc('\n', stderr);
289 	}
290 	/* We don't switch stderr to buffered, thus fprintf(stderr)
291 	 * always flushes its output and this is not necessary: */
292 	/* fflush(stderr); */
293 }
294 
error_msg(const char * fmt,...)295 void error_msg(const char *fmt, ...)
296 {
297 	va_list p;
298 	va_start(p, fmt);
299 	verror_msg(0, fmt, p);
300 	va_end(p);
301 }
302 
error_msg_and_die(const char * fmt,...)303 void error_msg_and_die(const char *fmt, ...)
304 {
305 	va_list p;
306 	va_start(p, fmt);
307 	verror_msg(0, fmt, p);
308 	die();
309 }
310 
perror_msg(const char * fmt,...)311 void perror_msg(const char *fmt, ...)
312 {
313 	va_list p;
314 	va_start(p, fmt);
315 	verror_msg(errno, fmt, p);
316 	va_end(p);
317 }
318 
perror_msg_and_die(const char * fmt,...)319 void perror_msg_and_die(const char *fmt, ...)
320 {
321 	va_list p;
322 	va_start(p, fmt);
323 	verror_msg(errno, fmt, p);
324 	die();
325 }
326 
die_out_of_memory(void)327 void die_out_of_memory(void)
328 {
329 	static bool recursed = 0;
330 	if (recursed)
331 		exit(1);
332 	recursed = 1;
333 	error_msg_and_die("Out of memory");
334 }
335 
336 static void
error_opt_arg(int opt,const char * arg)337 error_opt_arg(int opt, const char *arg)
338 {
339 	error_msg_and_die("Invalid -%c argument: '%s'", opt, arg);
340 }
341 
342 #if USE_SEIZE
343 static int
ptrace_attach_or_seize(int pid)344 ptrace_attach_or_seize(int pid)
345 {
346 	int r;
347 	if (!use_seize)
348 		return ptrace(PTRACE_ATTACH, pid, 0L, 0L);
349 	r = ptrace(PTRACE_SEIZE, pid, 0L, (unsigned long) ptrace_setoptions);
350 	if (r)
351 		return r;
352 	r = ptrace(PTRACE_INTERRUPT, pid, 0L, 0L);
353 	return r;
354 }
355 #else
356 # define ptrace_attach_or_seize(pid) ptrace(PTRACE_ATTACH, (pid), 0, 0)
357 #endif
358 
359 /*
360  * Used when we want to unblock stopped traced process.
361  * Should be only used with PTRACE_CONT, PTRACE_DETACH and PTRACE_SYSCALL.
362  * Returns 0 on success or if error was ESRCH
363  * (presumably process was killed while we talk to it).
364  * Otherwise prints error message and returns -1.
365  */
366 static int
ptrace_restart(int op,struct tcb * tcp,int sig)367 ptrace_restart(int op, struct tcb *tcp, int sig)
368 {
369 	int err;
370 	const char *msg;
371 
372 	errno = 0;
373 	ptrace(op, tcp->pid, (void *) 0, (long) sig);
374 	err = errno;
375 	if (!err)
376 		return 0;
377 
378 	msg = "SYSCALL";
379 	if (op == PTRACE_CONT)
380 		msg = "CONT";
381 	if (op == PTRACE_DETACH)
382 		msg = "DETACH";
383 #ifdef PTRACE_LISTEN
384 	if (op == PTRACE_LISTEN)
385 		msg = "LISTEN";
386 #endif
387 	/*
388 	 * Why curcol != 0? Otherwise sometimes we get this:
389 	 *
390 	 * 10252 kill(10253, SIGKILL)              = 0
391 	 *  <ptrace(SYSCALL,10252):No such process>10253 ...next decode...
392 	 *
393 	 * 10252 died after we retrieved syscall exit data,
394 	 * but before we tried to restart it. Log looks ugly.
395 	 */
396 	if (current_tcp && current_tcp->curcol != 0) {
397 		tprintf(" <ptrace(%s):%s>\n", msg, strerror(err));
398 		line_ended();
399 	}
400 	if (err == ESRCH)
401 		return 0;
402 	errno = err;
403 	perror_msg("ptrace(PTRACE_%s,pid:%d,sig:%d)", msg, tcp->pid, sig);
404 	return -1;
405 }
406 
407 static void
set_cloexec_flag(int fd)408 set_cloexec_flag(int fd)
409 {
410 	int flags, newflags;
411 
412 	flags = fcntl(fd, F_GETFD);
413 	if (flags < 0) {
414 		/* Can happen only if fd is bad.
415 		 * Should never happen: if it does, we have a bug
416 		 * in the caller. Therefore we just abort
417 		 * instead of propagating the error.
418 		 */
419 		perror_msg_and_die("fcntl(%d, F_GETFD)", fd);
420 	}
421 
422 	newflags = flags | FD_CLOEXEC;
423 	if (flags == newflags)
424 		return;
425 
426 	fcntl(fd, F_SETFD, newflags); /* never fails */
427 }
428 
kill_save_errno(pid_t pid,int sig)429 static void kill_save_errno(pid_t pid, int sig)
430 {
431 	int saved_errno = errno;
432 
433 	(void) kill(pid, sig);
434 	errno = saved_errno;
435 }
436 
437 /*
438  * When strace is setuid executable, we have to swap uids
439  * before and after filesystem and process management operations.
440  */
441 static void
swap_uid(void)442 swap_uid(void)
443 {
444 	int euid = geteuid(), uid = getuid();
445 
446 	if (euid != uid && setreuid(euid, uid) < 0) {
447 		perror_msg_and_die("setreuid");
448 	}
449 }
450 
451 #ifdef _LARGEFILE64_SOURCE
452 # ifdef HAVE_FOPEN64
453 #  define fopen_for_output fopen64
454 # else
455 #  define fopen_for_output fopen
456 # endif
457 # define struct_stat struct stat64
458 # define stat_file stat64
459 # define struct_dirent struct dirent64
460 # define read_dir readdir64
461 # define struct_rlimit struct rlimit64
462 # define set_rlimit setrlimit64
463 #else
464 # define fopen_for_output fopen
465 # define struct_stat struct stat
466 # define stat_file stat
467 # define struct_dirent struct dirent
468 # define read_dir readdir
469 # define struct_rlimit struct rlimit
470 # define set_rlimit setrlimit
471 #endif
472 
473 static FILE *
strace_fopen(const char * path)474 strace_fopen(const char *path)
475 {
476 	FILE *fp;
477 
478 	swap_uid();
479 	fp = fopen_for_output(path, "w");
480 	if (!fp)
481 		perror_msg_and_die("Can't fopen '%s'", path);
482 	swap_uid();
483 	set_cloexec_flag(fileno(fp));
484 	return fp;
485 }
486 
487 static int popen_pid = 0;
488 
489 #ifndef _PATH_BSHELL
490 # define _PATH_BSHELL "/bin/sh"
491 #endif
492 
493 /*
494  * We cannot use standard popen(3) here because we have to distinguish
495  * popen child process from other processes we trace, and standard popen(3)
496  * does not export its child's pid.
497  */
498 static FILE *
strace_popen(const char * command)499 strace_popen(const char *command)
500 {
501 	FILE *fp;
502 	int pid;
503 	int fds[2];
504 
505 	swap_uid();
506 	if (pipe(fds) < 0)
507 		perror_msg_and_die("pipe");
508 
509 	set_cloexec_flag(fds[1]); /* never fails */
510 
511 	pid = vfork();
512 	if (pid < 0)
513 		perror_msg_and_die("vfork");
514 
515 	if (pid == 0) {
516 		/* child */
517 		close(fds[1]);
518 		if (fds[0] != 0) {
519 			if (dup2(fds[0], 0))
520 				perror_msg_and_die("dup2");
521 			close(fds[0]);
522 		}
523 		execl(_PATH_BSHELL, "sh", "-c", command, NULL);
524 		perror_msg_and_die("Can't execute '%s'", _PATH_BSHELL);
525 	}
526 
527 	/* parent */
528 	popen_pid = pid;
529 	close(fds[0]);
530 	swap_uid();
531 	fp = fdopen(fds[1], "w");
532 	if (!fp)
533 		die_out_of_memory();
534 	return fp;
535 }
536 
537 void
tprintf(const char * fmt,...)538 tprintf(const char *fmt, ...)
539 {
540 	va_list args;
541 
542 	va_start(args, fmt);
543 	if (current_tcp) {
544 		int n = strace_vfprintf(current_tcp->outf, fmt, args);
545 		if (n < 0) {
546 			if (current_tcp->outf != stderr)
547 				perror_msg("%s", outfname);
548 		} else
549 			current_tcp->curcol += n;
550 	}
551 	va_end(args);
552 }
553 
554 #ifndef HAVE_FPUTS_UNLOCKED
555 # define fputs_unlocked fputs
556 #endif
557 
558 void
tprints(const char * str)559 tprints(const char *str)
560 {
561 	if (current_tcp) {
562 		int n = fputs_unlocked(str, current_tcp->outf);
563 		if (n >= 0) {
564 			current_tcp->curcol += strlen(str);
565 			return;
566 		}
567 		if (current_tcp->outf != stderr)
568 			perror_msg("%s", outfname);
569 	}
570 }
571 
572 void
line_ended(void)573 line_ended(void)
574 {
575 	if (current_tcp) {
576 		current_tcp->curcol = 0;
577 		fflush(current_tcp->outf);
578 	}
579 	if (printing_tcp) {
580 		printing_tcp->curcol = 0;
581 		printing_tcp = NULL;
582 	}
583 }
584 
585 void
printleader(struct tcb * tcp)586 printleader(struct tcb *tcp)
587 {
588 	/* If -ff, "previous tcb we printed" is always the same as current,
589 	 * because we have per-tcb output files.
590 	 */
591 	if (followfork >= 2)
592 		printing_tcp = tcp;
593 
594 	if (printing_tcp) {
595 		current_tcp = printing_tcp;
596 		if (printing_tcp->curcol != 0 && (followfork < 2 || printing_tcp == tcp)) {
597 			/*
598 			 * case 1: we have a shared log (i.e. not -ff), and last line
599 			 * wasn't finished (same or different tcb, doesn't matter).
600 			 * case 2: split log, we are the same tcb, but our last line
601 			 * didn't finish ("SIGKILL nuked us after syscall entry" etc).
602 			 */
603 			tprints(" <unfinished ...>\n");
604 			printing_tcp->curcol = 0;
605 		}
606 	}
607 
608 	printing_tcp = tcp;
609 	current_tcp = tcp;
610 	current_tcp->curcol = 0;
611 
612 	if (print_pid_pfx)
613 		tprintf("%-5d ", tcp->pid);
614 	else if (nprocs > 1 && !outfname)
615 		tprintf("[pid %5u] ", tcp->pid);
616 
617 	if (tflag) {
618 		char str[sizeof("HH:MM:SS")];
619 		struct timeval tv, dtv;
620 		static struct timeval otv;
621 
622 		gettimeofday(&tv, NULL);
623 		if (rflag) {
624 			if (otv.tv_sec == 0)
625 				otv = tv;
626 			tv_sub(&dtv, &tv, &otv);
627 			tprintf("%6ld.%06ld ",
628 				(long) dtv.tv_sec, (long) dtv.tv_usec);
629 			otv = tv;
630 		}
631 		else if (tflag > 2) {
632 			tprintf("%ld.%06ld ",
633 				(long) tv.tv_sec, (long) tv.tv_usec);
634 		}
635 		else {
636 			time_t local = tv.tv_sec;
637 			strftime(str, sizeof(str), "%T", localtime(&local));
638 			if (tflag > 1)
639 				tprintf("%s.%06ld ", str, (long) tv.tv_usec);
640 			else
641 				tprintf("%s ", str);
642 		}
643 	}
644 	if (iflag)
645 		print_pc(tcp);
646 }
647 
648 void
tabto(void)649 tabto(void)
650 {
651 	if (current_tcp->curcol < acolumn)
652 		tprints(acolumn_spaces + current_tcp->curcol);
653 }
654 
655 /* Should be only called directly *after successful attach* to a tracee.
656  * Otherwise, "strace -oFILE -ff -p<nonexistant_pid>"
657  * may create bogus empty FILE.<nonexistant_pid>, and then die.
658  */
659 static void
newoutf(struct tcb * tcp)660 newoutf(struct tcb *tcp)
661 {
662 	tcp->outf = shared_log; /* if not -ff mode, the same file is for all */
663 	if (followfork >= 2) {
664 		char name[520 + sizeof(int) * 3];
665 		sprintf(name, "%.512s.%u", outfname, tcp->pid);
666 		tcp->outf = strace_fopen(name);
667 	}
668 }
669 
670 static void
expand_tcbtab(void)671 expand_tcbtab(void)
672 {
673 	/* Allocate some more TCBs and expand the table.
674 	   We don't want to relocate the TCBs because our
675 	   callers have pointers and it would be a pain.
676 	   So tcbtab is a table of pointers.  Since we never
677 	   free the TCBs, we allocate a single chunk of many.  */
678 	unsigned int i = tcbtabsize;
679 	struct tcb *newtcbs = calloc(tcbtabsize, sizeof(newtcbs[0]));
680 	struct tcb **newtab = realloc(tcbtab, tcbtabsize * 2 * sizeof(tcbtab[0]));
681 	if (!newtab || !newtcbs)
682 		die_out_of_memory();
683 	tcbtabsize *= 2;
684 	tcbtab = newtab;
685 	while (i < tcbtabsize)
686 		tcbtab[i++] = newtcbs++;
687 }
688 
689 static struct tcb *
alloctcb(int pid)690 alloctcb(int pid)
691 {
692 	unsigned int i;
693 	struct tcb *tcp;
694 
695 	if (nprocs == tcbtabsize)
696 		expand_tcbtab();
697 
698 	for (i = 0; i < tcbtabsize; i++) {
699 		tcp = tcbtab[i];
700 		if (!tcp->pid) {
701 			memset(tcp, 0, sizeof(*tcp));
702 			tcp->pid = pid;
703 #if SUPPORTED_PERSONALITIES > 1
704 			tcp->currpers = current_personality;
705 #endif
706 
707 #ifdef USE_LIBUNWIND
708 			if (stack_trace_enabled)
709 				unwind_tcb_init(tcp);
710 #endif
711 
712 			nprocs++;
713 			if (debug_flag)
714 				fprintf(stderr, "new tcb for pid %d, active tcbs:%d\n", tcp->pid, nprocs);
715 			return tcp;
716 		}
717 	}
718 	error_msg_and_die("bug in alloctcb");
719 }
720 
721 static void
droptcb(struct tcb * tcp)722 droptcb(struct tcb *tcp)
723 {
724 	if (tcp->pid == 0)
725 		return;
726 
727 #ifdef USE_LIBUNWIND
728 	if (stack_trace_enabled) {
729 		unwind_tcb_fin(tcp);
730 	}
731 #endif
732 
733 	nprocs--;
734 	if (debug_flag)
735 		fprintf(stderr, "dropped tcb for pid %d, %d remain\n", tcp->pid, nprocs);
736 
737 	if (tcp->outf) {
738 		if (followfork >= 2) {
739 			if (tcp->curcol != 0)
740 				fprintf(tcp->outf, " <detached ...>\n");
741 			fclose(tcp->outf);
742 		} else {
743 			if (printing_tcp == tcp && tcp->curcol != 0)
744 				fprintf(tcp->outf, " <detached ...>\n");
745 			fflush(tcp->outf);
746 		}
747 	}
748 
749 	if (current_tcp == tcp)
750 		current_tcp = NULL;
751 	if (printing_tcp == tcp)
752 		printing_tcp = NULL;
753 
754 	memset(tcp, 0, sizeof(*tcp));
755 }
756 
757 /* Detach traced process.
758  * Never call DETACH twice on the same process as both unattached and
759  * attached-unstopped processes give the same ESRCH.  For unattached process we
760  * would SIGSTOP it and wait for its SIGSTOP notification forever.
761  */
762 static void
detach(struct tcb * tcp)763 detach(struct tcb *tcp)
764 {
765 	int error;
766 	int status;
767 
768 	/*
769 	 * Linux wrongly insists the child be stopped
770 	 * before detaching.  Arghh.  We go through hoops
771 	 * to make a clean break of things.
772 	 */
773 #if defined(SPARC)
774 # undef PTRACE_DETACH
775 # define PTRACE_DETACH PTRACE_SUNDETACH
776 #endif
777 
778 	if (!(tcp->flags & TCB_ATTACHED))
779 		goto drop;
780 
781 	/* We attached but possibly didn't see the expected SIGSTOP.
782 	 * We must catch exactly one as otherwise the detached process
783 	 * would be left stopped (process state T).
784 	 */
785 	if (tcp->flags & TCB_IGNORE_ONE_SIGSTOP)
786 		goto wait_loop;
787 
788 	error = ptrace(PTRACE_DETACH, tcp->pid, 0, 0);
789 	if (!error) {
790 		/* On a clear day, you can see forever. */
791 		goto drop;
792 	}
793 	if (errno != ESRCH) {
794 		/* Shouldn't happen. */
795 		perror_msg("detach: ptrace(PTRACE_DETACH,%u)", tcp->pid);
796 		goto drop;
797 	}
798 	/* ESRCH: process is either not stopped or doesn't exist. */
799 	if (my_tkill(tcp->pid, 0) < 0) {
800 		if (errno != ESRCH)
801 			/* Shouldn't happen. */
802 			perror_msg("detach: tkill(%u,0)", tcp->pid);
803 		/* else: process doesn't exist. */
804 		goto drop;
805 	}
806 	/* Process is not stopped, need to stop it. */
807 	if (use_seize) {
808 		/*
809 		 * With SEIZE, tracee can be in group-stop already.
810 		 * In this state sending it another SIGSTOP does nothing.
811 		 * Need to use INTERRUPT.
812 		 * Testcase: trying to ^C a "strace -p <stopped_process>".
813 		 */
814 		error = ptrace(PTRACE_INTERRUPT, tcp->pid, 0, 0);
815 		if (!error)
816 			goto wait_loop;
817 		if (errno != ESRCH)
818 			perror_msg("detach: ptrace(PTRACE_INTERRUPT,%u)", tcp->pid);
819 	}
820 	else {
821 		error = my_tkill(tcp->pid, SIGSTOP);
822 		if (!error)
823 			goto wait_loop;
824 		if (errno != ESRCH)
825 			perror_msg("detach: tkill(%u,SIGSTOP)", tcp->pid);
826 	}
827 	/* Either process doesn't exist, or some weird error. */
828 	goto drop;
829 
830  wait_loop:
831 	/* We end up here in three cases:
832 	 * 1. We sent PTRACE_INTERRUPT (use_seize case)
833 	 * 2. We sent SIGSTOP (!use_seize)
834 	 * 3. Attach SIGSTOP was already pending (TCB_IGNORE_ONE_SIGSTOP set)
835 	 */
836 	for (;;) {
837 		unsigned int sig;
838 		if (waitpid(tcp->pid, &status, __WALL) < 0) {
839 			if (errno == EINTR)
840 				continue;
841 			/*
842 			 * if (errno == ECHILD) break;
843 			 * ^^^  WRONG! We expect this PID to exist,
844 			 * and want to emit a message otherwise:
845 			 */
846 			perror_msg("detach: waitpid(%u)", tcp->pid);
847 			break;
848 		}
849 		if (!WIFSTOPPED(status)) {
850 			/*
851 			 * Tracee exited or was killed by signal.
852 			 * We shouldn't normally reach this place:
853 			 * we don't want to consume exit status.
854 			 * Consider "strace -p PID" being ^C-ed:
855 			 * we want merely to detach from PID.
856 			 *
857 			 * However, we _can_ end up here if tracee
858 			 * was SIGKILLed.
859 			 */
860 			break;
861 		}
862 		sig = WSTOPSIG(status);
863 		if (debug_flag)
864 			fprintf(stderr, "detach wait: event:%d sig:%d\n",
865 					(unsigned)status >> 16, sig);
866 		if (use_seize) {
867 			unsigned event = (unsigned)status >> 16;
868 			if (event == PTRACE_EVENT_STOP /*&& sig == SIGTRAP*/) {
869 				/*
870 				 * sig == SIGTRAP: PTRACE_INTERRUPT stop.
871 				 * sig == other: process was already stopped
872 				 * with this stopping sig (see tests/detach-stopped).
873 				 * Looks like re-injecting this sig is not necessary
874 				 * in DETACH for the tracee to remain stopped.
875 				 */
876 				sig = 0;
877 			}
878 			/*
879 			 * PTRACE_INTERRUPT is not guaranteed to produce
880 			 * the above event if other ptrace-stop is pending.
881 			 * See tests/detach-sleeping testcase:
882 			 * strace got SIGINT while tracee is sleeping.
883 			 * We sent PTRACE_INTERRUPT.
884 			 * We see syscall exit, not PTRACE_INTERRUPT stop.
885 			 * We won't get PTRACE_INTERRUPT stop
886 			 * if we would CONT now. Need to DETACH.
887 			 */
888 			if (sig == syscall_trap_sig)
889 				sig = 0;
890 			/* else: not sure in which case we can be here.
891 			 * Signal stop? Inject it while detaching.
892 			 */
893 			ptrace_restart(PTRACE_DETACH, tcp, sig);
894 			break;
895 		}
896 		/* Note: this check has to be after use_seize check */
897 		/* (else, in use_seize case SIGSTOP will be mistreated) */
898 		if (sig == SIGSTOP) {
899 			/* Detach, suppressing SIGSTOP */
900 			ptrace_restart(PTRACE_DETACH, tcp, 0);
901 			break;
902 		}
903 		if (sig == syscall_trap_sig)
904 			sig = 0;
905 		/* Can't detach just yet, may need to wait for SIGSTOP */
906 		error = ptrace_restart(PTRACE_CONT, tcp, sig);
907 		if (error < 0) {
908 			/* Should not happen.
909 			 * Note: ptrace_restart returns 0 on ESRCH, so it's not it.
910 			 * ptrace_restart already emitted error message.
911 			 */
912 			break;
913 		}
914 	}
915 
916  drop:
917 	if (!qflag && (tcp->flags & TCB_ATTACHED))
918 		fprintf(stderr, "Process %u detached\n", tcp->pid);
919 
920 	droptcb(tcp);
921 }
922 
923 static void
process_opt_p_list(char * opt)924 process_opt_p_list(char *opt)
925 {
926 	while (*opt) {
927 		/*
928 		 * We accept -p PID,PID; -p "`pidof PROG`"; -p "`pgrep PROG`".
929 		 * pidof uses space as delim, pgrep uses newline. :(
930 		 */
931 		int pid;
932 		char *delim = opt + strcspn(opt, ", \n\t");
933 		char c = *delim;
934 
935 		*delim = '\0';
936 		pid = string_to_uint(opt);
937 		if (pid <= 0) {
938 			error_msg_and_die("Invalid process id: '%s'", opt);
939 		}
940 		if (pid == strace_tracer_pid) {
941 			error_msg_and_die("I'm sorry, I can't let you do that, Dave.");
942 		}
943 		*delim = c;
944 		alloctcb(pid);
945 		if (c == '\0')
946 			break;
947 		opt = delim + 1;
948 	}
949 }
950 
951 static void
startup_attach(void)952 startup_attach(void)
953 {
954 	unsigned int tcbi;
955 	struct tcb *tcp;
956 
957 	/*
958 	 * Block user interruptions as we would leave the traced
959 	 * process stopped (process state T) if we would terminate in
960 	 * between PTRACE_ATTACH and wait4() on SIGSTOP.
961 	 * We rely on cleanup() from this point on.
962 	 */
963 	if (interactive)
964 		sigprocmask(SIG_BLOCK, &blocked_set, NULL);
965 
966 	if (daemonized_tracer) {
967 		pid_t pid = fork();
968 		if (pid < 0) {
969 			perror_msg_and_die("fork");
970 		}
971 		if (pid) { /* parent */
972 			/*
973 			 * Wait for grandchild to attach to straced process
974 			 * (grandparent). Grandchild SIGKILLs us after it attached.
975 			 * Grandparent's wait() is unblocked by our death,
976 			 * it proceeds to exec the straced program.
977 			 */
978 			pause();
979 			_exit(0); /* paranoia */
980 		}
981 		/* grandchild */
982 		/* We will be the tracer process. Remember our new pid: */
983 		strace_tracer_pid = getpid();
984 	}
985 
986 	for (tcbi = 0; tcbi < tcbtabsize; tcbi++) {
987 		tcp = tcbtab[tcbi];
988 
989 		if (!tcp->pid)
990 			continue;
991 
992 		/* Is this a process we should attach to, but not yet attached? */
993 		if (tcp->flags & TCB_ATTACHED)
994 			continue; /* no, we already attached it */
995 
996 		if (followfork && !daemonized_tracer) {
997 			char procdir[sizeof("/proc/%d/task") + sizeof(int) * 3];
998 			DIR *dir;
999 
1000 			sprintf(procdir, "/proc/%d/task", tcp->pid);
1001 			dir = opendir(procdir);
1002 			if (dir != NULL) {
1003 				unsigned int ntid = 0, nerr = 0;
1004 				struct_dirent *de;
1005 
1006 				while ((de = read_dir(dir)) != NULL) {
1007 					struct tcb *cur_tcp;
1008 					int tid;
1009 
1010 					if (de->d_fileno == 0)
1011 						continue;
1012 					/* we trust /proc filesystem */
1013 					tid = atoi(de->d_name);
1014 					if (tid <= 0)
1015 						continue;
1016 					++ntid;
1017 					if (ptrace_attach_or_seize(tid) < 0) {
1018 						++nerr;
1019 						if (debug_flag)
1020 							fprintf(stderr, "attach to pid %d failed\n", tid);
1021 						continue;
1022 					}
1023 					if (debug_flag)
1024 						fprintf(stderr, "attach to pid %d succeeded\n", tid);
1025 					cur_tcp = tcp;
1026 					if (tid != tcp->pid)
1027 						cur_tcp = alloctcb(tid);
1028 					cur_tcp->flags |= TCB_ATTACHED | TCB_STARTUP | post_attach_sigstop;
1029 					newoutf(cur_tcp);
1030 				}
1031 				closedir(dir);
1032 				if (interactive) {
1033 					sigprocmask(SIG_SETMASK, &empty_set, NULL);
1034 					if (interrupted)
1035 						goto ret;
1036 					sigprocmask(SIG_BLOCK, &blocked_set, NULL);
1037 				}
1038 				ntid -= nerr;
1039 				if (ntid == 0) {
1040 					perror_msg("attach: ptrace(PTRACE_ATTACH, ...)");
1041 					droptcb(tcp);
1042 					continue;
1043 				}
1044 				if (!qflag) {
1045 					fprintf(stderr, ntid > 1
1046 ? "Process %u attached with %u threads\n"
1047 : "Process %u attached\n",
1048 						tcp->pid, ntid);
1049 				}
1050 				if (!(tcp->flags & TCB_ATTACHED)) {
1051 					/* -p PID, we failed to attach to PID itself
1052 					 * but did attach to some of its sibling threads.
1053 					 * Drop PID's tcp.
1054 					 */
1055 					droptcb(tcp);
1056 				}
1057 				continue;
1058 			} /* if (opendir worked) */
1059 		} /* if (-f) */
1060 		if (ptrace_attach_or_seize(tcp->pid) < 0) {
1061 			perror_msg("attach: ptrace(PTRACE_ATTACH, ...)");
1062 			droptcb(tcp);
1063 			continue;
1064 		}
1065 		tcp->flags |= TCB_ATTACHED | TCB_STARTUP | post_attach_sigstop;
1066 		newoutf(tcp);
1067 		if (debug_flag)
1068 			fprintf(stderr, "attach to pid %d (main) succeeded\n", tcp->pid);
1069 
1070 		if (daemonized_tracer) {
1071 			/*
1072 			 * Make parent go away.
1073 			 * Also makes grandparent's wait() unblock.
1074 			 */
1075 			kill(getppid(), SIGKILL);
1076 		}
1077 
1078 		if (!qflag)
1079 			fprintf(stderr,
1080 				"Process %u attached\n",
1081 				tcp->pid);
1082 	} /* for each tcbtab[] */
1083 
1084  ret:
1085 	if (interactive)
1086 		sigprocmask(SIG_SETMASK, &empty_set, NULL);
1087 }
1088 
1089 /* Stack-o-phobic exec helper, in the hope to work around
1090  * NOMMU + "daemonized tracer" difficulty.
1091  */
1092 struct exec_params {
1093 	int fd_to_close;
1094 	uid_t run_euid;
1095 	gid_t run_egid;
1096 	char **argv;
1097 	char *pathname;
1098 };
1099 static struct exec_params params_for_tracee;
1100 
1101 static void ATTRIBUTE_NOINLINE ATTRIBUTE_NORETURN
exec_or_die(void)1102 exec_or_die(void)
1103 {
1104 	struct exec_params *params = &params_for_tracee;
1105 
1106 	if (params->fd_to_close >= 0)
1107 		close(params->fd_to_close);
1108 	if (!daemonized_tracer && !use_seize) {
1109 		if (ptrace(PTRACE_TRACEME, 0L, 0L, 0L) < 0) {
1110 			perror_msg_and_die("ptrace(PTRACE_TRACEME, ...)");
1111 		}
1112 	}
1113 
1114 	if (username != NULL) {
1115 		/*
1116 		 * It is important to set groups before we
1117 		 * lose privileges on setuid.
1118 		 */
1119 		if (initgroups(username, run_gid) < 0) {
1120 			perror_msg_and_die("initgroups");
1121 		}
1122 		if (setregid(run_gid, params->run_egid) < 0) {
1123 			perror_msg_and_die("setregid");
1124 		}
1125 		if (setreuid(run_uid, params->run_euid) < 0) {
1126 			perror_msg_and_die("setreuid");
1127 		}
1128 	}
1129 	else if (geteuid() != 0)
1130 		if (setreuid(run_uid, run_uid) < 0) {
1131 			perror_msg_and_die("setreuid");
1132 		}
1133 
1134 	if (!daemonized_tracer) {
1135 		/*
1136 		 * Induce a ptrace stop. Tracer (our parent)
1137 		 * will resume us with PTRACE_SYSCALL and display
1138 		 * the immediately following execve syscall.
1139 		 * Can't do this on NOMMU systems, we are after
1140 		 * vfork: parent is blocked, stopping would deadlock.
1141 		 */
1142 		if (!NOMMU_SYSTEM)
1143 			kill(getpid(), SIGSTOP);
1144 	} else {
1145 		alarm(3);
1146 		/* we depend on SIGCHLD set to SIG_DFL by init code */
1147 		/* if it happens to be SIG_IGN'ed, wait won't block */
1148 		wait(NULL);
1149 		alarm(0);
1150 	}
1151 
1152 	execv(params->pathname, params->argv);
1153 	perror_msg_and_die("exec");
1154 }
1155 
1156 static void
startup_child(char ** argv)1157 startup_child(char **argv)
1158 {
1159 	struct_stat statbuf;
1160 	const char *filename;
1161 	size_t filename_len;
1162 	char pathname[PATH_MAX];
1163 	int pid;
1164 	struct tcb *tcp;
1165 
1166 	filename = argv[0];
1167 	filename_len = strlen(filename);
1168 
1169 	if (filename_len > sizeof(pathname) - 1) {
1170 		errno = ENAMETOOLONG;
1171 		perror_msg_and_die("exec");
1172 	}
1173 	if (strchr(filename, '/')) {
1174 		strcpy(pathname, filename);
1175 	}
1176 #ifdef USE_DEBUGGING_EXEC
1177 	/*
1178 	 * Debuggers customarily check the current directory
1179 	 * first regardless of the path but doing that gives
1180 	 * security geeks a panic attack.
1181 	 */
1182 	else if (stat_file(filename, &statbuf) == 0)
1183 		strcpy(pathname, filename);
1184 #endif /* USE_DEBUGGING_EXEC */
1185 	else {
1186 		const char *path;
1187 		size_t m, n, len;
1188 
1189 		for (path = getenv("PATH"); path && *path; path += m) {
1190 			const char *colon = strchr(path, ':');
1191 			if (colon) {
1192 				n = colon - path;
1193 				m = n + 1;
1194 			}
1195 			else
1196 				m = n = strlen(path);
1197 			if (n == 0) {
1198 				if (!getcwd(pathname, PATH_MAX))
1199 					continue;
1200 				len = strlen(pathname);
1201 			}
1202 			else if (n > sizeof pathname - 1)
1203 				continue;
1204 			else {
1205 				strncpy(pathname, path, n);
1206 				len = n;
1207 			}
1208 			if (len && pathname[len - 1] != '/')
1209 				pathname[len++] = '/';
1210 			if (filename_len + len > sizeof(pathname) - 1)
1211 				continue;
1212 			strcpy(pathname + len, filename);
1213 			if (stat_file(pathname, &statbuf) == 0 &&
1214 			    /* Accept only regular files
1215 			       with some execute bits set.
1216 			       XXX not perfect, might still fail */
1217 			    S_ISREG(statbuf.st_mode) &&
1218 			    (statbuf.st_mode & 0111))
1219 				break;
1220 		}
1221 		if (!path || !*path)
1222 			pathname[0] = '\0';
1223 	}
1224 	if (stat_file(pathname, &statbuf) < 0) {
1225 		perror_msg_and_die("Can't stat '%s'", filename);
1226 	}
1227 
1228 	params_for_tracee.fd_to_close = (shared_log != stderr) ? fileno(shared_log) : -1;
1229 	params_for_tracee.run_euid = (statbuf.st_mode & S_ISUID) ? statbuf.st_uid : run_uid;
1230 	params_for_tracee.run_egid = (statbuf.st_mode & S_ISGID) ? statbuf.st_gid : run_gid;
1231 	params_for_tracee.argv = argv;
1232 	/*
1233 	 * On NOMMU, can be safely freed only after execve in tracee.
1234 	 * It's hard to know when that happens, so we just leak it.
1235 	 */
1236 	params_for_tracee.pathname = NOMMU_SYSTEM ? strdup(pathname) : pathname;
1237 
1238 #if defined HAVE_PRCTL && defined PR_SET_PTRACER && defined PR_SET_PTRACER_ANY
1239 	if (daemonized_tracer)
1240 		prctl(PR_SET_PTRACER, PR_SET_PTRACER_ANY);
1241 #endif
1242 
1243 	pid = fork();
1244 	if (pid < 0) {
1245 		perror_msg_and_die("fork");
1246 	}
1247 	if ((pid != 0 && daemonized_tracer)
1248 	 || (pid == 0 && !daemonized_tracer)
1249 	) {
1250 		/* We are to become the tracee. Two cases:
1251 		 * -D: we are parent
1252 		 * not -D: we are child
1253 		 */
1254 		exec_or_die();
1255 	}
1256 
1257 	/* We are the tracer */
1258 
1259 	if (!daemonized_tracer) {
1260 		strace_child = pid;
1261 		if (!use_seize) {
1262 			/* child did PTRACE_TRACEME, nothing to do in parent */
1263 		} else {
1264 			if (!NOMMU_SYSTEM) {
1265 				/* Wait until child stopped itself */
1266 				int status;
1267 				while (waitpid(pid, &status, WSTOPPED) < 0) {
1268 					if (errno == EINTR)
1269 						continue;
1270 					perror_msg_and_die("waitpid");
1271 				}
1272 				if (!WIFSTOPPED(status) || WSTOPSIG(status) != SIGSTOP) {
1273 					kill_save_errno(pid, SIGKILL);
1274 					perror_msg_and_die("Unexpected wait status %x", status);
1275 				}
1276 			}
1277 			/* Else: NOMMU case, we have no way to sync.
1278 			 * Just attach to it as soon as possible.
1279 			 * This means that we may miss a few first syscalls...
1280 			 */
1281 
1282 			if (ptrace_attach_or_seize(pid)) {
1283 				kill_save_errno(pid, SIGKILL);
1284 				perror_msg_and_die("Can't attach to %d", pid);
1285 			}
1286 			if (!NOMMU_SYSTEM)
1287 				kill(pid, SIGCONT);
1288 		}
1289 		tcp = alloctcb(pid);
1290 		if (!NOMMU_SYSTEM)
1291 			tcp->flags |= TCB_ATTACHED | TCB_STARTUP | post_attach_sigstop;
1292 		else
1293 			tcp->flags |= TCB_ATTACHED | TCB_STARTUP;
1294 		newoutf(tcp);
1295 	}
1296 	else {
1297 		/* With -D, we are *child* here, IOW: different pid. Fetch it: */
1298 		strace_tracer_pid = getpid();
1299 		/* The tracee is our parent: */
1300 		pid = getppid();
1301 		alloctcb(pid);
1302 		/* attaching will be done later, by startup_attach */
1303 		/* note: we don't do newoutf(tcp) here either! */
1304 
1305 		/* NOMMU BUG! -D mode is active, we (child) return,
1306 		 * and we will scribble over parent's stack!
1307 		 * When parent later unpauses, it segfaults.
1308 		 *
1309 		 * We work around it
1310 		 * (1) by declaring exec_or_die() NORETURN,
1311 		 * hopefully compiler will just jump to it
1312 		 * instead of call (won't push anything to stack),
1313 		 * (2) by trying very hard in exec_or_die()
1314 		 * to not use any stack,
1315 		 * (3) having a really big (PATH_MAX) stack object
1316 		 * in this function, which creates a "buffer" between
1317 		 * child's and parent's stack pointers.
1318 		 * This may save us if (1) and (2) failed
1319 		 * and compiler decided to use stack in exec_or_die() anyway
1320 		 * (happens on i386 because of stack parameter passing).
1321 		 *
1322 		 * A cleaner solution is to use makecontext + setcontext
1323 		 * to create a genuine separate stack and execute on it.
1324 		 */
1325 	}
1326 }
1327 
1328 #if USE_SEIZE
1329 static void
test_ptrace_seize(void)1330 test_ptrace_seize(void)
1331 {
1332 	int pid;
1333 
1334 	/* Need fork for test. NOMMU has no forks */
1335 	if (NOMMU_SYSTEM) {
1336 		post_attach_sigstop = 0; /* this sets use_seize to 1 */
1337 		return;
1338 	}
1339 
1340 	pid = fork();
1341 	if (pid < 0)
1342 		perror_msg_and_die("fork");
1343 
1344 	if (pid == 0) {
1345 		pause();
1346 		_exit(0);
1347 	}
1348 
1349 	/* PTRACE_SEIZE, unlike ATTACH, doesn't force tracee to trap.  After
1350 	 * attaching tracee continues to run unless a trap condition occurs.
1351 	 * PTRACE_SEIZE doesn't affect signal or group stop state.
1352 	 */
1353 	if (ptrace(PTRACE_SEIZE, pid, 0, 0) == 0) {
1354 		post_attach_sigstop = 0; /* this sets use_seize to 1 */
1355 	} else if (debug_flag) {
1356 		fprintf(stderr, "PTRACE_SEIZE doesn't work\n");
1357 	}
1358 
1359 	kill(pid, SIGKILL);
1360 
1361 	while (1) {
1362 		int status, tracee_pid;
1363 
1364 		errno = 0;
1365 		tracee_pid = waitpid(pid, &status, 0);
1366 		if (tracee_pid <= 0) {
1367 			if (errno == EINTR)
1368 				continue;
1369 			perror_msg_and_die("%s: unexpected wait result %d",
1370 					 __func__, tracee_pid);
1371 		}
1372 		if (WIFSIGNALED(status)) {
1373 			return;
1374 		}
1375 		error_msg_and_die("%s: unexpected wait status %x",
1376 				__func__, status);
1377 	}
1378 }
1379 #else /* !USE_SEIZE */
1380 # define test_ptrace_seize() ((void)0)
1381 #endif
1382 
1383 static unsigned
get_os_release(void)1384 get_os_release(void)
1385 {
1386 	unsigned rel;
1387 	const char *p;
1388 	struct utsname u;
1389 	if (uname(&u) < 0)
1390 		perror_msg_and_die("uname");
1391 	/* u.release has this form: "3.2.9[-some-garbage]" */
1392 	rel = 0;
1393 	p = u.release;
1394 	for (;;) {
1395 		if (!(*p >= '0' && *p <= '9'))
1396 			error_msg_and_die("Bad OS release string: '%s'", u.release);
1397 		/* Note: this open-codes KERNEL_VERSION(): */
1398 		rel = (rel << 8) | atoi(p);
1399 		if (rel >= KERNEL_VERSION(1,0,0))
1400 			break;
1401 		while (*p >= '0' && *p <= '9')
1402 			p++;
1403 		if (*p != '.') {
1404 			if (rel >= KERNEL_VERSION(0,1,0)) {
1405 				/* "X.Y-something" means "X.Y.0" */
1406 				rel <<= 8;
1407 				break;
1408 			}
1409 			error_msg_and_die("Bad OS release string: '%s'", u.release);
1410 		}
1411 		p++;
1412 	}
1413 	return rel;
1414 }
1415 
1416 /*
1417  * Initialization part of main() was eating much stack (~0.5k),
1418  * which was unused after init.
1419  * We can reuse it if we move init code into a separate function.
1420  *
1421  * Don't want main() to inline us and defeat the reason
1422  * we have a separate function.
1423  */
1424 static void ATTRIBUTE_NOINLINE
init(int argc,char * argv[])1425 init(int argc, char *argv[])
1426 {
1427 	struct tcb *tcp;
1428 	int c, i;
1429 	int optF = 0;
1430 	unsigned int tcbi;
1431 	struct sigaction sa;
1432 
1433 	progname = argv[0] ? argv[0] : "strace";
1434 
1435 	/* Make sure SIGCHLD has the default action so that waitpid
1436 	   definitely works without losing track of children.  The user
1437 	   should not have given us a bogus state to inherit, but he might
1438 	   have.  Arguably we should detect SIG_IGN here and pass it on
1439 	   to children, but probably noone really needs that.  */
1440 	signal(SIGCHLD, SIG_DFL);
1441 
1442 	strace_tracer_pid = getpid();
1443 
1444 	os_release = get_os_release();
1445 
1446 	/* Allocate the initial tcbtab.  */
1447 	tcbtabsize = argc;	/* Surely enough for all -p args.  */
1448 	tcbtab = calloc(tcbtabsize, sizeof(tcbtab[0]));
1449 	if (!tcbtab)
1450 		die_out_of_memory();
1451 	tcp = calloc(tcbtabsize, sizeof(*tcp));
1452 	if (!tcp)
1453 		die_out_of_memory();
1454 	for (tcbi = 0; tcbi < tcbtabsize; tcbi++)
1455 		tcbtab[tcbi] = tcp++;
1456 
1457 	shared_log = stderr;
1458 	set_sortby(DEFAULT_SORTBY);
1459 	set_personality(DEFAULT_PERSONALITY);
1460 	qualify("trace=all");
1461 	qualify("abbrev=all");
1462 	qualify("verbose=all");
1463 #if DEFAULT_QUAL_FLAGS != (QUAL_TRACE | QUAL_ABBREV | QUAL_VERBOSE)
1464 # error Bug in DEFAULT_QUAL_FLAGS
1465 #endif
1466 	qualify("signal=all");
1467 	while ((c = getopt(argc, argv,
1468 		"+b:cCdfFhiqrtTvVwxyz"
1469 #ifdef USE_LIBUNWIND
1470 		"k"
1471 #endif
1472 		"D"
1473 		"a:e:o:O:p:s:S:u:E:P:I:")) != EOF) {
1474 		switch (c) {
1475 		case 'b':
1476 			if (strcmp(optarg, "execve") != 0)
1477 				error_msg_and_die("Syscall '%s' for -b isn't supported",
1478 					optarg);
1479 			detach_on_execve = 1;
1480 			break;
1481 		case 'c':
1482 			if (cflag == CFLAG_BOTH) {
1483 				error_msg_and_die("-c and -C are mutually exclusive");
1484 			}
1485 			cflag = CFLAG_ONLY_STATS;
1486 			break;
1487 		case 'C':
1488 			if (cflag == CFLAG_ONLY_STATS) {
1489 				error_msg_and_die("-c and -C are mutually exclusive");
1490 			}
1491 			cflag = CFLAG_BOTH;
1492 			break;
1493 		case 'd':
1494 			debug_flag = 1;
1495 			break;
1496 		case 'D':
1497 			daemonized_tracer = 1;
1498 			break;
1499 		case 'F':
1500 			optF = 1;
1501 			break;
1502 		case 'f':
1503 			followfork++;
1504 			break;
1505 		case 'h':
1506 			usage(stdout, 0);
1507 			break;
1508 		case 'i':
1509 			iflag = 1;
1510 			break;
1511 		case 'q':
1512 			qflag++;
1513 			break;
1514 		case 'r':
1515 			rflag = 1;
1516 			/* fall through to tflag++ */
1517 		case 't':
1518 			tflag++;
1519 			break;
1520 		case 'T':
1521 			Tflag = 1;
1522 			break;
1523 		case 'w':
1524 			count_wallclock = 1;
1525 			break;
1526 		case 'x':
1527 			xflag++;
1528 			break;
1529 		case 'y':
1530 			show_fd_path++;
1531 			break;
1532 		case 'v':
1533 			qualify("abbrev=none");
1534 			break;
1535 		case 'V':
1536 			printf("%s -- version %s\n", PACKAGE_NAME, VERSION);
1537 			exit(0);
1538 			break;
1539 		case 'z':
1540 			not_failing_only = 1;
1541 			break;
1542 		case 'a':
1543 			acolumn = string_to_uint(optarg);
1544 			if (acolumn < 0)
1545 				error_opt_arg(c, optarg);
1546 			break;
1547 		case 'e':
1548 			qualify(optarg);
1549 			break;
1550 		case 'o':
1551 			outfname = strdup(optarg);
1552 			break;
1553 		case 'O':
1554 			i = string_to_uint(optarg);
1555 			if (i < 0)
1556 				error_opt_arg(c, optarg);
1557 			set_overhead(i);
1558 			break;
1559 		case 'p':
1560 			process_opt_p_list(optarg);
1561 			break;
1562 		case 'P':
1563 			pathtrace_select(optarg);
1564 			break;
1565 		case 's':
1566 			i = string_to_uint(optarg);
1567 			if (i < 0)
1568 				error_opt_arg(c, optarg);
1569 			max_strlen = i;
1570 			break;
1571 		case 'S':
1572 			set_sortby(optarg);
1573 			break;
1574 		case 'u':
1575 			username = strdup(optarg);
1576 			break;
1577 #ifdef USE_LIBUNWIND
1578 		case 'k':
1579 			stack_trace_enabled = true;
1580 			break;
1581 #endif
1582 		case 'E':
1583 			if (putenv(optarg) < 0)
1584 				die_out_of_memory();
1585 			break;
1586 		case 'I':
1587 			opt_intr = string_to_uint(optarg);
1588 			if (opt_intr <= 0 || opt_intr >= NUM_INTR_OPTS)
1589 				error_opt_arg(c, optarg);
1590 			break;
1591 		default:
1592 			usage(stderr, 1);
1593 			break;
1594 		}
1595 	}
1596 	argv += optind;
1597 	/* argc -= optind; - no need, argc is not used below */
1598 
1599 	acolumn_spaces = malloc(acolumn + 1);
1600 	if (!acolumn_spaces)
1601 		die_out_of_memory();
1602 	memset(acolumn_spaces, ' ', acolumn);
1603 	acolumn_spaces[acolumn] = '\0';
1604 
1605 	/* Must have PROG [ARGS], or -p PID. Not both. */
1606 	if (!argv[0] == !nprocs)
1607 		usage(stderr, 1);
1608 
1609 	if (nprocs != 0 && daemonized_tracer) {
1610 		error_msg_and_die("-D and -p are mutually exclusive");
1611 	}
1612 
1613 	if (!followfork)
1614 		followfork = optF;
1615 
1616 	if (followfork >= 2 && cflag) {
1617 		error_msg_and_die("(-c or -C) and -ff are mutually exclusive");
1618 	}
1619 
1620 	if (count_wallclock && !cflag) {
1621 		error_msg_and_die("-w must be given with (-c or -C)");
1622 	}
1623 
1624 	if (cflag == CFLAG_ONLY_STATS) {
1625 		if (iflag)
1626 			error_msg("-%c has no effect with -c", 'i');
1627 #ifdef USE_LIBUNWIND
1628 		if (stack_trace_enabled)
1629 			error_msg("-%c has no effect with -c", 'k');
1630 #endif
1631 		if (rflag)
1632 			error_msg("-%c has no effect with -c", 'r');
1633 		if (tflag)
1634 			error_msg("-%c has no effect with -c", 't');
1635 		if (Tflag)
1636 			error_msg("-%c has no effect with -c", 'T');
1637 		if (show_fd_path)
1638 			error_msg("-%c has no effect with -c", 'y');
1639 	}
1640 
1641 #ifdef USE_LIBUNWIND
1642 	if (stack_trace_enabled)
1643 		unwind_init();
1644 #endif
1645 
1646 	/* See if they want to run as another user. */
1647 	if (username != NULL) {
1648 		struct passwd *pent;
1649 
1650 		if (getuid() != 0 || geteuid() != 0) {
1651 			error_msg_and_die("You must be root to use the -u option");
1652 		}
1653 		pent = getpwnam(username);
1654 		if (pent == NULL) {
1655 			error_msg_and_die("Cannot find user '%s'", username);
1656 		}
1657 		run_uid = pent->pw_uid;
1658 		run_gid = pent->pw_gid;
1659 	}
1660 	else {
1661 		run_uid = getuid();
1662 		run_gid = getgid();
1663 	}
1664 
1665 	if (followfork)
1666 		ptrace_setoptions |= PTRACE_O_TRACECLONE |
1667 				     PTRACE_O_TRACEFORK |
1668 				     PTRACE_O_TRACEVFORK;
1669 	if (debug_flag)
1670 		fprintf(stderr, "ptrace_setoptions = %#x\n", ptrace_setoptions);
1671 	test_ptrace_seize();
1672 
1673 	/* Check if they want to redirect the output. */
1674 	if (outfname) {
1675 		/* See if they want to pipe the output. */
1676 		if (outfname[0] == '|' || outfname[0] == '!') {
1677 			/*
1678 			 * We can't do the <outfname>.PID funny business
1679 			 * when using popen, so prohibit it.
1680 			 */
1681 			if (followfork >= 2)
1682 				error_msg_and_die("Piping the output and -ff are mutually exclusive");
1683 			shared_log = strace_popen(outfname + 1);
1684 		}
1685 		else if (followfork < 2)
1686 			shared_log = strace_fopen(outfname);
1687 	} else {
1688 		/* -ff without -o FILE is the same as single -f */
1689 		if (followfork >= 2)
1690 			followfork = 1;
1691 	}
1692 
1693 	if (!outfname || outfname[0] == '|' || outfname[0] == '!') {
1694 		char *buf = malloc(BUFSIZ);
1695 		if (!buf)
1696 			die_out_of_memory();
1697 		setvbuf(shared_log, buf, _IOLBF, BUFSIZ);
1698 	}
1699 	if (outfname && argv[0]) {
1700 		if (!opt_intr)
1701 			opt_intr = INTR_NEVER;
1702 		qflag = 1;
1703 	}
1704 	if (!opt_intr)
1705 		opt_intr = INTR_WHILE_WAIT;
1706 
1707 	/* argv[0]	-pPID	-oFILE	Default interactive setting
1708 	 * yes		0	0	INTR_WHILE_WAIT
1709 	 * no		1	0	INTR_WHILE_WAIT
1710 	 * yes		0	1	INTR_NEVER
1711 	 * no		1	1	INTR_WHILE_WAIT
1712 	 */
1713 
1714 	sigemptyset(&empty_set);
1715 	sigemptyset(&blocked_set);
1716 
1717 	/* startup_child() must be called before the signal handlers get
1718 	 * installed below as they are inherited into the spawned process.
1719 	 * Also we do not need to be protected by them as during interruption
1720 	 * in the startup_child() mode we kill the spawned process anyway.
1721 	 */
1722 	if (argv[0]) {
1723 		if (!NOMMU_SYSTEM || daemonized_tracer)
1724 			hide_log_until_execve = 1;
1725 		skip_one_b_execve = 1;
1726 		startup_child(argv);
1727 	}
1728 
1729 	sa.sa_handler = SIG_IGN;
1730 	sigemptyset(&sa.sa_mask);
1731 	sa.sa_flags = 0;
1732 	sigaction(SIGTTOU, &sa, NULL); /* SIG_IGN */
1733 	sigaction(SIGTTIN, &sa, NULL); /* SIG_IGN */
1734 	if (opt_intr != INTR_ANYWHERE) {
1735 		if (opt_intr == INTR_BLOCK_TSTP_TOO)
1736 			sigaction(SIGTSTP, &sa, NULL); /* SIG_IGN */
1737 		/*
1738 		 * In interactive mode (if no -o OUTFILE, or -p PID is used),
1739 		 * fatal signals are blocked while syscall stop is processed,
1740 		 * and acted on in between, when waiting for new syscall stops.
1741 		 * In non-interactive mode, signals are ignored.
1742 		 */
1743 		if (opt_intr == INTR_WHILE_WAIT) {
1744 			sigaddset(&blocked_set, SIGHUP);
1745 			sigaddset(&blocked_set, SIGINT);
1746 			sigaddset(&blocked_set, SIGQUIT);
1747 			sigaddset(&blocked_set, SIGPIPE);
1748 			sigaddset(&blocked_set, SIGTERM);
1749 			sa.sa_handler = interrupt;
1750 		}
1751 		/* SIG_IGN, or set handler for these */
1752 		sigaction(SIGHUP, &sa, NULL);
1753 		sigaction(SIGINT, &sa, NULL);
1754 		sigaction(SIGQUIT, &sa, NULL);
1755 		sigaction(SIGPIPE, &sa, NULL);
1756 		sigaction(SIGTERM, &sa, NULL);
1757 	}
1758 	if (nprocs != 0 || daemonized_tracer)
1759 		startup_attach();
1760 
1761 	/* Do we want pids printed in our -o OUTFILE?
1762 	 * -ff: no (every pid has its own file); or
1763 	 * -f: yes (there can be more pids in the future); or
1764 	 * -p PID1,PID2: yes (there are already more than one pid)
1765 	 */
1766 	print_pid_pfx = (outfname && followfork < 2 && (followfork == 1 || nprocs > 1));
1767 }
1768 
1769 static struct tcb *
pid2tcb(int pid)1770 pid2tcb(int pid)
1771 {
1772 	unsigned int i;
1773 
1774 	if (pid <= 0)
1775 		return NULL;
1776 
1777 	for (i = 0; i < tcbtabsize; i++) {
1778 		struct tcb *tcp = tcbtab[i];
1779 		if (tcp->pid == pid)
1780 			return tcp;
1781 	}
1782 
1783 	return NULL;
1784 }
1785 
1786 static void
cleanup(void)1787 cleanup(void)
1788 {
1789 	unsigned int i;
1790 	struct tcb *tcp;
1791 	int fatal_sig;
1792 
1793 	/* 'interrupted' is a volatile object, fetch it only once */
1794 	fatal_sig = interrupted;
1795 	if (!fatal_sig)
1796 		fatal_sig = SIGTERM;
1797 
1798 	for (i = 0; i < tcbtabsize; i++) {
1799 		tcp = tcbtab[i];
1800 		if (!tcp->pid)
1801 			continue;
1802 		if (debug_flag)
1803 			fprintf(stderr,
1804 				"cleanup: looking at pid %u\n", tcp->pid);
1805 		if (tcp->pid == strace_child) {
1806 			kill(tcp->pid, SIGCONT);
1807 			kill(tcp->pid, fatal_sig);
1808 		}
1809 		detach(tcp);
1810 	}
1811 	if (cflag)
1812 		call_summary(shared_log);
1813 }
1814 
1815 static void
interrupt(int sig)1816 interrupt(int sig)
1817 {
1818 	interrupted = sig;
1819 }
1820 
1821 static void
print_debug_info(const int pid,int status)1822 print_debug_info(const int pid, int status)
1823 {
1824 	const unsigned int event = (unsigned int) status >> 16;
1825 	char buf[sizeof("WIFEXITED,exitcode=%u") + sizeof(int)*3 /*paranoia:*/ + 16];
1826 	char evbuf[sizeof(",EVENT_VFORK_DONE (%u)") + sizeof(int)*3 /*paranoia:*/ + 16];
1827 
1828 	strcpy(buf, "???");
1829 	if (WIFSIGNALED(status))
1830 #ifdef WCOREDUMP
1831 		sprintf(buf, "WIFSIGNALED,%ssig=%s",
1832 				WCOREDUMP(status) ? "core," : "",
1833 				signame(WTERMSIG(status)));
1834 #else
1835 		sprintf(buf, "WIFSIGNALED,sig=%s",
1836 				signame(WTERMSIG(status)));
1837 #endif
1838 	if (WIFEXITED(status))
1839 		sprintf(buf, "WIFEXITED,exitcode=%u", WEXITSTATUS(status));
1840 	if (WIFSTOPPED(status))
1841 		sprintf(buf, "WIFSTOPPED,sig=%s", signame(WSTOPSIG(status)));
1842 #ifdef WIFCONTINUED
1843 	/* Should never be seen */
1844 	if (WIFCONTINUED(status))
1845 		strcpy(buf, "WIFCONTINUED");
1846 #endif
1847 	evbuf[0] = '\0';
1848 	if (event != 0) {
1849 		static const char *const event_names[] = {
1850 			[PTRACE_EVENT_CLONE] = "CLONE",
1851 			[PTRACE_EVENT_FORK]  = "FORK",
1852 			[PTRACE_EVENT_VFORK] = "VFORK",
1853 			[PTRACE_EVENT_VFORK_DONE] = "VFORK_DONE",
1854 			[PTRACE_EVENT_EXEC]  = "EXEC",
1855 			[PTRACE_EVENT_EXIT]  = "EXIT",
1856 			/* [PTRACE_EVENT_STOP (=128)] would make biggish array */
1857 		};
1858 		const char *e = "??";
1859 		if (event < ARRAY_SIZE(event_names))
1860 			e = event_names[event];
1861 		else if (event == PTRACE_EVENT_STOP)
1862 			e = "STOP";
1863 		sprintf(evbuf, ",EVENT_%s (%u)", e, event);
1864 	}
1865 	fprintf(stderr, " [wait(0x%06x) = %u] %s%s\n", status, pid, buf, evbuf);
1866 }
1867 
1868 static struct tcb *
maybe_allocate_tcb(const int pid,int status)1869 maybe_allocate_tcb(const int pid, int status)
1870 {
1871 	if (!WIFSTOPPED(status)) {
1872 		if (detach_on_execve && pid == strace_child) {
1873 			/* example: strace -bexecve sh -c 'exec true' */
1874 			strace_child = 0;
1875 			return NULL;
1876 		}
1877 		/*
1878 		 * This can happen if we inherited an unknown child.
1879 		 * Example: (sleep 1 & exec strace true)
1880 		 */
1881 		error_msg("Exit of unknown pid %u ignored", pid);
1882 		return NULL;
1883 	}
1884 	if (followfork) {
1885 		/* We assume it's a fork/vfork/clone child */
1886 		struct tcb *tcp = alloctcb(pid);
1887 		tcp->flags |= TCB_ATTACHED | TCB_STARTUP | post_attach_sigstop;
1888 		newoutf(tcp);
1889 		if (!qflag)
1890 			fprintf(stderr, "Process %d attached\n", pid);
1891 		return tcp;
1892 	} else {
1893 		/* This can happen if a clone call used
1894 		 * CLONE_PTRACE itself.
1895 		 */
1896 		ptrace(PTRACE_CONT, pid, (char *) 0, 0);
1897 		error_msg("Stop of unknown pid %u seen, PTRACE_CONTed it", pid);
1898 		return NULL;
1899 	}
1900 }
1901 
1902 static struct tcb *
maybe_switch_tcbs(struct tcb * tcp,const int pid)1903 maybe_switch_tcbs(struct tcb *tcp, const int pid)
1904 {
1905 	FILE *fp;
1906 	struct tcb *execve_thread;
1907 	long old_pid = 0;
1908 
1909 	if (ptrace(PTRACE_GETEVENTMSG, pid, NULL, (long) &old_pid) < 0)
1910 		return tcp;
1911 	/* Avoid truncation in pid2tcb() param passing */
1912 	if (old_pid <= 0 || old_pid == pid)
1913 		return tcp;
1914 	if ((unsigned long) old_pid > UINT_MAX)
1915 		return tcp;
1916 	execve_thread = pid2tcb(old_pid);
1917 	/* It should be !NULL, but I feel paranoid */
1918 	if (!execve_thread)
1919 		return tcp;
1920 
1921 	if (execve_thread->curcol != 0) {
1922 		/*
1923 		 * One case we are here is -ff:
1924 		 * try "strace -oLOG -ff test/threaded_execve"
1925 		 */
1926 		fprintf(execve_thread->outf, " <pid changed to %d ...>\n", pid);
1927 		/*execve_thread->curcol = 0; - no need, see code below */
1928 	}
1929 	/* Swap output FILEs (needed for -ff) */
1930 	fp = execve_thread->outf;
1931 	execve_thread->outf = tcp->outf;
1932 	tcp->outf = fp;
1933 	/* And their column positions */
1934 	execve_thread->curcol = tcp->curcol;
1935 	tcp->curcol = 0;
1936 	/* Drop leader, but close execve'd thread outfile (if -ff) */
1937 	droptcb(tcp);
1938 	/* Switch to the thread, reusing leader's outfile and pid */
1939 	tcp = execve_thread;
1940 	tcp->pid = pid;
1941 	if (cflag != CFLAG_ONLY_STATS) {
1942 		printleader(tcp);
1943 		tprintf("+++ superseded by execve in pid %lu +++\n", old_pid);
1944 		line_ended();
1945 		tcp->flags |= TCB_REPRINT;
1946 	}
1947 
1948 	return tcp;
1949 }
1950 
1951 static void
print_signalled(struct tcb * tcp,const int pid,int status)1952 print_signalled(struct tcb *tcp, const int pid, int status)
1953 {
1954 	if (pid == strace_child) {
1955 		exit_code = 0x100 | WTERMSIG(status);
1956 		strace_child = 0;
1957 	}
1958 
1959 	if (cflag != CFLAG_ONLY_STATS
1960 	 && (qual_flags[WTERMSIG(status)] & QUAL_SIGNAL)
1961 	) {
1962 		printleader(tcp);
1963 #ifdef WCOREDUMP
1964 		tprintf("+++ killed by %s %s+++\n",
1965 			signame(WTERMSIG(status)),
1966 			WCOREDUMP(status) ? "(core dumped) " : "");
1967 #else
1968 		tprintf("+++ killed by %s +++\n",
1969 			signame(WTERMSIG(status)));
1970 #endif
1971 		line_ended();
1972 	}
1973 }
1974 
1975 static void
print_exited(struct tcb * tcp,const int pid,int status)1976 print_exited(struct tcb *tcp, const int pid, int status)
1977 {
1978 	if (pid == strace_child) {
1979 		exit_code = WEXITSTATUS(status);
1980 		strace_child = 0;
1981 	}
1982 
1983 	if (cflag != CFLAG_ONLY_STATS &&
1984 	    qflag < 2) {
1985 		printleader(tcp);
1986 		tprintf("+++ exited with %d +++\n", WEXITSTATUS(status));
1987 		line_ended();
1988 	}
1989 }
1990 
1991 static void
print_stopped(struct tcb * tcp,const siginfo_t * si,const unsigned int sig)1992 print_stopped(struct tcb *tcp, const siginfo_t *si, const unsigned int sig)
1993 {
1994 	if (cflag != CFLAG_ONLY_STATS
1995 	    && !hide_log_until_execve
1996 	    && (qual_flags[sig] & QUAL_SIGNAL)
1997 	   ) {
1998 		printleader(tcp);
1999 		if (si) {
2000 			tprintf("--- %s ", signame(sig));
2001 			printsiginfo(si, verbose(tcp));
2002 			tprints(" ---\n");
2003 		} else
2004 			tprintf("--- stopped by %s ---\n", signame(sig));
2005 		line_ended();
2006 	}
2007 }
2008 
2009 static void
startup_tcb(struct tcb * tcp)2010 startup_tcb(struct tcb *tcp)
2011 {
2012 	if (debug_flag)
2013 		fprintf(stderr, "pid %d has TCB_STARTUP, initializing it\n",
2014 			tcp->pid);
2015 
2016 	tcp->flags &= ~TCB_STARTUP;
2017 
2018 	if (!use_seize) {
2019 		if (debug_flag)
2020 			fprintf(stderr, "setting opts 0x%x on pid %d\n",
2021 				ptrace_setoptions, tcp->pid);
2022 		if (ptrace(PTRACE_SETOPTIONS, tcp->pid, NULL, ptrace_setoptions) < 0) {
2023 			if (errno != ESRCH) {
2024 				/* Should never happen, really */
2025 				perror_msg_and_die("PTRACE_SETOPTIONS");
2026 			}
2027 		}
2028 	}
2029 }
2030 
2031 /* Returns true iff the main trace loop has to continue. */
2032 static bool
trace(void)2033 trace(void)
2034 {
2035 	int pid;
2036 	int wait_errno;
2037 	int status;
2038 	bool stopped;
2039 	unsigned int sig;
2040 	unsigned int event;
2041 	struct tcb *tcp;
2042 	struct rusage ru;
2043 
2044 	if (interrupted)
2045 		return false;
2046 
2047 	/*
2048 	 * Used to exit simply when nprocs hits zero, but in this testcase:
2049 	 *  int main() { _exit(!!fork()); }
2050 	 * under strace -f, parent sometimes (rarely) manages
2051 	 * to exit before we see the first stop of the child,
2052 	 * and we are losing track of it:
2053 	 *  19923 clone(...) = 19924
2054 	 *  19923 exit_group(1)     = ?
2055 	 *  19923 +++ exited with 1 +++
2056 	 * Exiting only when wait() returns ECHILD works better.
2057 	 */
2058 	if (popen_pid != 0) {
2059 		/* However, if -o|logger is in use, we can't do that.
2060 		 * Can work around that by double-forking the logger,
2061 		 * but that loses the ability to wait for its completion
2062 		 * on exit. Oh well...
2063 		 */
2064 		if (nprocs == 0)
2065 			return false;
2066 	}
2067 
2068 	if (interactive)
2069 		sigprocmask(SIG_SETMASK, &empty_set, NULL);
2070 	pid = wait4(-1, &status, __WALL, (cflag ? &ru : NULL));
2071 	wait_errno = errno;
2072 	if (interactive)
2073 		sigprocmask(SIG_BLOCK, &blocked_set, NULL);
2074 
2075 	if (pid < 0) {
2076 		if (wait_errno == EINTR)
2077 			return true;
2078 		if (nprocs == 0 && wait_errno == ECHILD)
2079 			return false;
2080 		/*
2081 		 * If nprocs > 0, ECHILD is not expected,
2082 		 * treat it as any other error here:
2083 		 */
2084 		errno = wait_errno;
2085 		perror_msg_and_die("wait4(__WALL)");
2086 	}
2087 
2088 	if (pid == popen_pid) {
2089 		if (!WIFSTOPPED(status))
2090 			popen_pid = 0;
2091 		return true;
2092 	}
2093 
2094 	if (debug_flag)
2095 		print_debug_info(pid, status);
2096 
2097 	/* Look up 'pid' in our table. */
2098 	tcp = pid2tcb(pid);
2099 
2100 	if (!tcp) {
2101 		tcp = maybe_allocate_tcb(pid, status);
2102 		if (!tcp)
2103 			return true;
2104 	}
2105 
2106 	if (WIFSTOPPED(status))
2107 		get_regs(pid);
2108 	else
2109 		clear_regs();
2110 
2111 	event = (unsigned int) status >> 16;
2112 
2113 	if (event == PTRACE_EVENT_EXEC) {
2114 		/*
2115 		 * Under Linux, execve changes pid to thread leader's pid,
2116 		 * and we see this changed pid on EVENT_EXEC and later,
2117 		 * execve sysexit. Leader "disappears" without exit
2118 		 * notification. Let user know that, drop leader's tcb,
2119 		 * and fix up pid in execve thread's tcb.
2120 		 * Effectively, execve thread's tcb replaces leader's tcb.
2121 		 *
2122 		 * BTW, leader is 'stuck undead' (doesn't report WIFEXITED
2123 		 * on exit syscall) in multithreaded programs exactly
2124 		 * in order to handle this case.
2125 		 *
2126 		 * PTRACE_GETEVENTMSG returns old pid starting from Linux 3.0.
2127 		 * On 2.6 and earlier, it can return garbage.
2128 		 */
2129 		if (os_release >= KERNEL_VERSION(3,0,0))
2130 			tcp = maybe_switch_tcbs(tcp, pid);
2131 
2132 		if (detach_on_execve && !skip_one_b_execve) {
2133 			detach(tcp); /* do "-b execve" thingy */
2134 			return true;
2135 		}
2136 		skip_one_b_execve = 0;
2137 	}
2138 
2139 	/* Set current output file */
2140 	current_tcp = tcp;
2141 
2142 	if (cflag) {
2143 		tv_sub(&tcp->dtime, &ru.ru_stime, &tcp->stime);
2144 		tcp->stime = ru.ru_stime;
2145 	}
2146 
2147 	if (WIFSIGNALED(status)) {
2148 		print_signalled(tcp, pid, status);
2149 		droptcb(tcp);
2150 		return true;
2151 	}
2152 
2153 	if (WIFEXITED(status)) {
2154 		print_exited(tcp, pid, status);
2155 		droptcb(tcp);
2156 		return true;
2157 	}
2158 
2159 	if (!WIFSTOPPED(status)) {
2160 		/*
2161 		 * Neither signalled, exited or stopped.
2162 		 * How could that be?
2163 		 */
2164 		error_msg("pid %u not stopped!", pid);
2165 		droptcb(tcp);
2166 		return true;
2167 	}
2168 
2169 	/* Is this the very first time we see this tracee stopped? */
2170 	if (tcp->flags & TCB_STARTUP) {
2171 		startup_tcb(tcp);
2172 		if (get_scno(tcp) == 1)
2173 			tcp->s_prev_ent = tcp->s_ent;
2174 	}
2175 
2176 	sig = WSTOPSIG(status);
2177 
2178 	if (event != 0) {
2179 		/* Ptrace event */
2180 #if USE_SEIZE
2181 		if (event == PTRACE_EVENT_STOP) {
2182 			/*
2183 			 * PTRACE_INTERRUPT-stop or group-stop.
2184 			 * PTRACE_INTERRUPT-stop has sig == SIGTRAP here.
2185 			 */
2186 			switch (sig) {
2187 				case SIGSTOP:
2188 				case SIGTSTP:
2189 				case SIGTTIN:
2190 				case SIGTTOU:
2191 					stopped = true;
2192 					goto show_stopsig;
2193 			}
2194 		}
2195 #endif
2196 		goto restart_tracee_with_sig_0;
2197 	}
2198 
2199 	/*
2200 	 * Is this post-attach SIGSTOP?
2201 	 * Interestingly, the process may stop
2202 	 * with STOPSIG equal to some other signal
2203 	 * than SIGSTOP if we happend to attach
2204 	 * just before the process takes a signal.
2205 	 */
2206 	if (sig == SIGSTOP && (tcp->flags & TCB_IGNORE_ONE_SIGSTOP)) {
2207 		if (debug_flag)
2208 			fprintf(stderr, "ignored SIGSTOP on pid %d\n", tcp->pid);
2209 		tcp->flags &= ~TCB_IGNORE_ONE_SIGSTOP;
2210 		goto restart_tracee_with_sig_0;
2211 	}
2212 
2213 	if (sig != syscall_trap_sig) {
2214 		siginfo_t si;
2215 
2216 		/*
2217 		 * True if tracee is stopped by signal
2218 		 * (as opposed to "tracee received signal").
2219 		 * TODO: shouldn't we check for errno == EINVAL too?
2220 		 * We can get ESRCH instead, you know...
2221 		 */
2222 		stopped = ptrace(PTRACE_GETSIGINFO, pid, 0, (long) &si) < 0;
2223 #if USE_SEIZE
2224 show_stopsig:
2225 #endif
2226 		print_stopped(tcp, stopped ? NULL : &si, sig);
2227 
2228 		if (!stopped)
2229 			/* It's signal-delivery-stop. Inject the signal */
2230 			goto restart_tracee;
2231 
2232 		/* It's group-stop */
2233 		if (use_seize) {
2234 			/*
2235 			 * This ends ptrace-stop, but does *not* end group-stop.
2236 			 * This makes stopping signals work properly on straced process
2237 			 * (that is, process really stops. It used to continue to run).
2238 			 */
2239 			if (ptrace_restart(PTRACE_LISTEN, tcp, 0) < 0) {
2240 				/* Note: ptrace_restart emitted error message */
2241 				exit_code = 1;
2242 				return false;
2243 			}
2244 			return true;
2245 		}
2246 		/* We don't have PTRACE_LISTEN support... */
2247 		goto restart_tracee;
2248 	}
2249 
2250 	/* We handled quick cases, we are permitted to interrupt now. */
2251 	if (interrupted)
2252 		return false;
2253 
2254 	/*
2255 	 * This should be syscall entry or exit.
2256 	 * Handle it.
2257 	 */
2258 	if (trace_syscall(tcp) < 0) {
2259 		/*
2260 		 * ptrace() failed in trace_syscall().
2261 		 * Likely a result of process disappearing mid-flight.
2262 		 * Observed case: exit_group() or SIGKILL terminating
2263 		 * all processes in thread group.
2264 		 * We assume that ptrace error was caused by process death.
2265 		 * We used to detach(tcp) here, but since we no longer
2266 		 * implement "detach before death" policy/hack,
2267 		 * we can let this process to report its death to us
2268 		 * normally, via WIFEXITED or WIFSIGNALED wait status.
2269 		 */
2270 		return true;
2271 	}
2272 
2273 restart_tracee_with_sig_0:
2274 	sig = 0;
2275 
2276 restart_tracee:
2277 	if (ptrace_restart(PTRACE_SYSCALL, tcp, sig) < 0) {
2278 		/* Note: ptrace_restart emitted error message */
2279 		exit_code = 1;
2280 		return false;
2281 	}
2282 
2283 	return true;
2284 }
2285 
2286 int
main(int argc,char * argv[])2287 main(int argc, char *argv[])
2288 {
2289 	init(argc, argv);
2290 
2291 	while (trace())
2292 		;
2293 
2294 	cleanup();
2295 	fflush(NULL);
2296 	if (shared_log != stderr)
2297 		fclose(shared_log);
2298 	if (popen_pid) {
2299 		while (waitpid(popen_pid, NULL, 0) < 0 && errno == EINTR)
2300 			;
2301 	}
2302 	if (exit_code > 0xff) {
2303 		/* Avoid potential core file clobbering.  */
2304 		struct_rlimit rlim = {0, 0};
2305 		set_rlimit(RLIMIT_CORE, &rlim);
2306 
2307 		/* Child was killed by a signal, mimic that.  */
2308 		exit_code &= 0xff;
2309 		signal(exit_code, SIG_DFL);
2310 		raise(exit_code);
2311 		/* Paranoia - what if this signal is not fatal?
2312 		   Exit with 128 + signo then.  */
2313 		exit_code += 128;
2314 	}
2315 
2316 	return exit_code;
2317 }
2318