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>
8 * Redistribution and use in source and binary forms, with or without
9 * modification, are permitted provided that the following conditions
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.
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.
33 #include <sys/param.h>
36 #include <sys/resource.h>
42 #include <sys/utsname.h>
44 # include <sys/prctl.h>
48 #include "printsiginfo.h"
50 /* In some libc, these aren't declared. Do it ourself: */
51 extern char **environ;
56 /* if this is true do the stack trace for every system call */
57 bool stack_trace_enabled = false;
60 #if defined __NR_tkill
61 # define my_tkill(tid, sig) syscall(__NR_tkill, (tid), (sig))
63 /* kill() may choose arbitrarily the target task of the process group
64 while we later wait on a that specific TID. PID process waits become
65 TID task specific waits for a process under ptrace(2). */
66 # warning "tkill(2) not available, risk of strace hangs!"
67 # define my_tkill(tid, sig) kill((tid), (sig))
70 /* Glue for systems without a MMU that cannot provide fork() */
71 #if !defined(HAVE_FORK)
73 # define NOMMU_SYSTEM 1
76 # define fork() vfork()
79 const unsigned int syscall_trap_sig = SIGTRAP | 0x80;
81 cflag_t cflag = CFLAG_NONE;
82 unsigned int followfork = 0;
83 unsigned int ptrace_setoptions = PTRACE_O_TRACESYSGOOD | PTRACE_O_TRACEEXEC
85 unsigned int xflag = 0;
89 bool count_wallclock = 0;
90 unsigned int qflag = 0;
91 static unsigned int tflag = 0;
92 static bool rflag = 0;
93 static bool print_pid_pfx = 0;
98 INTR_ANYWHERE = 1, /* don't block/ignore any signals */
99 INTR_WHILE_WAIT = 2, /* block fatal signals while decoding syscall. default */
100 INTR_NEVER = 3, /* block fatal signals. default if '-o FILE PROG' */
101 INTR_BLOCK_TSTP_TOO = 4, /* block fatal signals and SIGTSTP (^Z) */
105 /* We play with signal mask only if this mode is active: */
106 #define interactive (opt_intr == INTR_WHILE_WAIT)
109 * daemonized_tracer supports -D option.
110 * With this option, strace forks twice.
111 * Unlike normal case, with -D *grandparent* process exec's,
112 * becoming a traced process. Child exits (this prevents traced process
113 * from having children it doesn't expect to have), and grandchild
114 * attaches to grandparent similarly to strace -p PID.
115 * This allows for more transparent interaction in cases
116 * when process and its parent are communicating via signals,
117 * wait() etc. Without -D, strace process gets lodged in between,
118 * disrupting parent<->child link.
120 static bool daemonized_tracer = 0;
123 static int post_attach_sigstop = TCB_IGNORE_ONE_SIGSTOP;
124 # define use_seize (post_attach_sigstop == 0)
126 # define post_attach_sigstop TCB_IGNORE_ONE_SIGSTOP
130 /* Sometimes we want to print only succeeding syscalls. */
131 bool not_failing_only = 0;
133 /* Show path associated with fd arguments */
134 unsigned int show_fd_path = 0;
136 static bool detach_on_execve = 0;
138 static int exit_code;
139 static int strace_child = 0;
140 static int strace_tracer_pid = 0;
142 static char *username = NULL;
143 static uid_t run_uid;
144 static gid_t run_gid;
146 unsigned int max_strlen = DEFAULT_STRLEN;
147 static int acolumn = DEFAULT_ACOLUMN;
148 static char *acolumn_spaces;
150 static char *outfname = NULL;
151 /* If -ff, points to stderr. Else, it's our common output log */
152 static FILE *shared_log;
154 struct tcb *printing_tcp = NULL;
155 static struct tcb *current_tcp;
157 static struct tcb **tcbtab;
158 static unsigned int nprocs, tcbtabsize;
159 static const char *progname;
161 unsigned os_release; /* generated from uname()'s u.release */
163 static void detach(struct tcb *tcp);
164 static void cleanup(void);
165 static void interrupt(int sig);
166 static sigset_t empty_set, blocked_set;
168 #ifdef HAVE_SIG_ATOMIC_T
169 static volatile sig_atomic_t interrupted;
171 static volatile int interrupted;
174 #ifndef HAVE_STRERROR
176 #if !HAVE_DECL_SYS_ERRLIST
178 extern char *sys_errlist[];
184 static char buf[sizeof("Unknown error %d") + sizeof(int)*3];
186 if (err_no < 1 || err_no >= sys_nerr) {
187 sprintf(buf, "Unknown error %d", err_no);
190 return sys_errlist[err_no];
193 #endif /* HAVE_STERRROR */
199 usage: strace [-CdffhiqrtttTvVwxxy] [-I n] [-e expr]...\n\
200 [-a column] [-o file] [-s strsize] [-P path]...\n\
201 -p pid... / [-D] [-E var=val]... [-u username] PROG [ARGS]\n\
202 or: strace -c[dfw] [-I n] [-e expr]... [-O overhead] [-S sortby]\n\
203 -p pid... / [-D] [-E var=val]... [-u username] PROG [ARGS]\n\
206 -a column alignment COLUMN for printing syscall results (default %d)\n\
207 -i print instruction pointer at time of syscall\n\
208 -o file send trace output to FILE instead of stderr\n\
209 -q suppress messages about attaching, detaching, etc.\n\
210 -r print relative timestamp\n\
211 -s strsize limit length of print strings to STRSIZE chars (default %d)\n\
212 -t print absolute timestamp\n\
213 -tt print absolute timestamp with usecs\n\
214 -T print time spent in each syscall\n\
215 -x print non-ascii strings in hex\n\
216 -xx print all strings in hex\n\
217 -y print paths associated with file descriptor arguments\n\
218 -yy print protocol specific information associated with socket file descriptors\n\
221 -c count time, calls, and errors for each syscall and report summary\n\
222 -C like -c but also print regular output\n\
223 -O overhead set overhead for tracing syscalls to OVERHEAD usecs\n\
224 -S sortby sort syscall counts by: time, calls, name, nothing (default %s)\n\
225 -w summarise syscall latency (default is system time)\n\
228 -e expr a qualifying expression: option=[!]all or option=[!]val1[,val2]...\n\
229 options: trace, abbrev, verbose, raw, signal, read, write, fault\n\
230 -P path trace accesses to path\n\
233 -b execve detach on execve syscall\n\
234 -D run tracer process as a detached grandchild, not as parent\n\
236 -ff follow forks with output into separate files\n\
238 1: no signals are blocked\n\
239 2: fatal signals are blocked while decoding syscall (default)\n\
240 3: fatal signals are always blocked (default if '-o FILE PROG')\n\
241 4: fatal signals and SIGTSTP (^Z) are always blocked\n\
242 (useful to make 'strace -o FILE PROG' not stop on ^Z)\n\
245 -E var remove var from the environment for command\n\
246 -E var=val put var=val in the environment for command\n\
247 -p pid trace process with process id PID, may be repeated\n\
248 -u username run command as username handling setuid and/or setgid\n\
251 -d enable debug output to stderr\n\
252 -v verbose mode: print unabbreviated argv, stat, termios, etc. args\n\
253 -h print help message\n\
257 " -k obtain stack trace between each syscall (experimental)\n\
260 /* ancient, no one should use it
261 -F -- attempt to follow vforks (deprecated, use -f)\n\
263 /* this is broken, so don't document it
264 -z -- print only succeeding syscalls\n\
266 , DEFAULT_ACOLUMN, DEFAULT_STRLEN, DEFAULT_SORTBY);
270 static void ATTRIBUTE_NORETURN
273 if (strace_tracer_pid == getpid()) {
280 static void verror_msg(int err_no, const char *fmt, va_list p)
286 /* We want to print entire message with single fprintf to ensure
287 * message integrity if stderr is shared with other programs.
288 * Thus we use vasprintf + single fprintf.
291 if (vasprintf(&msg, fmt, p) >= 0) {
293 fprintf(stderr, "%s: %s: %s\n", progname, msg, strerror(err_no));
295 fprintf(stderr, "%s: %s\n", progname, msg);
298 /* malloc in vasprintf failed, try it without malloc */
299 fprintf(stderr, "%s: ", progname);
300 vfprintf(stderr, fmt, p);
302 fprintf(stderr, ": %s\n", strerror(err_no));
306 /* We don't switch stderr to buffered, thus fprintf(stderr)
307 * always flushes its output and this is not necessary: */
308 /* fflush(stderr); */
311 void error_msg(const char *fmt, ...)
315 verror_msg(0, fmt, p);
319 void error_msg_and_die(const char *fmt, ...)
323 verror_msg(0, fmt, p);
327 void error_msg_and_help(const char *fmt, ...)
332 verror_msg(0, fmt, p);
334 fprintf(stderr, "Try '%s -h' for more information.\n", progname);
338 void perror_msg(const char *fmt, ...)
342 verror_msg(errno, fmt, p);
346 void perror_msg_and_die(const char *fmt, ...)
350 verror_msg(errno, fmt, p);
355 error_opt_arg(int opt, const char *arg)
357 error_msg_and_help("invalid -%c argument: '%s'", opt, arg);
360 static const char *ptrace_attach_cmd;
363 ptrace_attach_or_seize(int pid)
368 return ptrace_attach_cmd = "PTRACE_ATTACH",
369 ptrace(PTRACE_ATTACH, pid, 0L, 0L);
370 r = ptrace(PTRACE_SEIZE, pid, 0L, (unsigned long) ptrace_setoptions);
372 return ptrace_attach_cmd = "PTRACE_SEIZE", r;
373 r = ptrace(PTRACE_INTERRUPT, pid, 0L, 0L);
374 return ptrace_attach_cmd = "PTRACE_INTERRUPT", r;
376 return ptrace_attach_cmd = "PTRACE_ATTACH",
377 ptrace(PTRACE_ATTACH, pid, 0L, 0L);
382 * Used when we want to unblock stopped traced process.
383 * Should be only used with PTRACE_CONT, PTRACE_DETACH and PTRACE_SYSCALL.
384 * Returns 0 on success or if error was ESRCH
385 * (presumably process was killed while we talk to it).
386 * Otherwise prints error message and returns -1.
389 ptrace_restart(const unsigned int op, struct tcb *const tcp, unsigned int sig)
395 ptrace(op, tcp->pid, 0L, (unsigned long) sig);
415 * Why curcol != 0? Otherwise sometimes we get this:
417 * 10252 kill(10253, SIGKILL) = 0
418 * <ptrace(SYSCALL,10252):No such process>10253 ...next decode...
420 * 10252 died after we retrieved syscall exit data,
421 * but before we tried to restart it. Log looks ugly.
423 if (current_tcp && current_tcp->curcol != 0) {
424 tprintf(" <ptrace(%s):%s>\n", msg, strerror(err));
430 perror_msg("ptrace(PTRACE_%s,pid:%d,sig:%u)", msg, tcp->pid, sig);
435 set_cloexec_flag(int fd)
439 flags = fcntl(fd, F_GETFD);
441 /* Can happen only if fd is bad.
442 * Should never happen: if it does, we have a bug
443 * in the caller. Therefore we just abort
444 * instead of propagating the error.
446 perror_msg_and_die("fcntl(%d, F_GETFD)", fd);
449 newflags = flags | FD_CLOEXEC;
450 if (flags == newflags)
453 fcntl(fd, F_SETFD, newflags); /* never fails */
457 kill_save_errno(pid_t pid, int sig)
459 int saved_errno = errno;
461 (void) kill(pid, sig);
466 * When strace is setuid executable, we have to swap uids
467 * before and after filesystem and process management operations.
472 int euid = geteuid(), uid = getuid();
474 if (euid != uid && setreuid(euid, uid) < 0) {
475 perror_msg_and_die("setreuid");
479 #ifdef _LARGEFILE64_SOURCE
481 # define fopen_for_output fopen64
483 # define fopen_for_output fopen
485 # define struct_stat struct stat64
486 # define stat_file stat64
487 # define struct_dirent struct dirent64
488 # define read_dir readdir64
489 # define struct_rlimit struct rlimit64
490 # define set_rlimit setrlimit64
492 # define fopen_for_output fopen
493 # define struct_stat struct stat
494 # define stat_file stat
495 # define struct_dirent struct dirent
496 # define read_dir readdir
497 # define struct_rlimit struct rlimit
498 # define set_rlimit setrlimit
502 strace_fopen(const char *path)
507 fp = fopen_for_output(path, "w");
509 perror_msg_and_die("Can't fopen '%s'", path);
511 set_cloexec_flag(fileno(fp));
515 static int popen_pid = 0;
518 # define _PATH_BSHELL "/bin/sh"
522 * We cannot use standard popen(3) here because we have to distinguish
523 * popen child process from other processes we trace, and standard popen(3)
524 * does not export its child's pid.
527 strace_popen(const char *command)
535 perror_msg_and_die("pipe");
537 set_cloexec_flag(fds[1]); /* never fails */
541 perror_msg_and_die("vfork");
548 perror_msg_and_die("dup2");
551 execl(_PATH_BSHELL, "sh", "-c", command, NULL);
552 perror_msg_and_die("Can't execute '%s'", _PATH_BSHELL);
559 fp = fdopen(fds[1], "w");
566 tprintf(const char *fmt, ...)
572 int n = strace_vfprintf(current_tcp->outf, fmt, args);
574 if (current_tcp->outf != stderr)
575 perror_msg("%s", outfname);
577 current_tcp->curcol += n;
582 #ifndef HAVE_FPUTS_UNLOCKED
583 # define fputs_unlocked fputs
587 tprints(const char *str)
590 int n = fputs_unlocked(str, current_tcp->outf);
592 current_tcp->curcol += strlen(str);
595 if (current_tcp->outf != stderr)
596 perror_msg("%s", outfname);
604 current_tcp->curcol = 0;
605 fflush(current_tcp->outf);
608 printing_tcp->curcol = 0;
614 printleader(struct tcb *tcp)
616 /* If -ff, "previous tcb we printed" is always the same as current,
617 * because we have per-tcb output files.
623 current_tcp = printing_tcp;
624 if (printing_tcp->curcol != 0 && (followfork < 2 || printing_tcp == tcp)) {
626 * case 1: we have a shared log (i.e. not -ff), and last line
627 * wasn't finished (same or different tcb, doesn't matter).
628 * case 2: split log, we are the same tcb, but our last line
629 * didn't finish ("SIGKILL nuked us after syscall entry" etc).
631 tprints(" <unfinished ...>\n");
632 printing_tcp->curcol = 0;
638 current_tcp->curcol = 0;
641 tprintf("%-5d ", tcp->pid);
642 else if (nprocs > 1 && !outfname)
643 tprintf("[pid %5u] ", tcp->pid);
646 char str[sizeof("HH:MM:SS")];
647 struct timeval tv, dtv;
648 static struct timeval otv;
650 gettimeofday(&tv, NULL);
654 tv_sub(&dtv, &tv, &otv);
655 tprintf("%6ld.%06ld ",
656 (long) dtv.tv_sec, (long) dtv.tv_usec);
659 else if (tflag > 2) {
660 tprintf("%ld.%06ld ",
661 (long) tv.tv_sec, (long) tv.tv_usec);
664 time_t local = tv.tv_sec;
665 strftime(str, sizeof(str), "%T", localtime(&local));
667 tprintf("%s.%06ld ", str, (long) tv.tv_usec);
679 if (current_tcp->curcol < acolumn)
680 tprints(acolumn_spaces + current_tcp->curcol);
683 /* Should be only called directly *after successful attach* to a tracee.
684 * Otherwise, "strace -oFILE -ff -p<nonexistant_pid>"
685 * may create bogus empty FILE.<nonexistant_pid>, and then die.
688 newoutf(struct tcb *tcp)
690 tcp->outf = shared_log; /* if not -ff mode, the same file is for all */
691 if (followfork >= 2) {
692 char name[520 + sizeof(int) * 3];
693 sprintf(name, "%.512s.%u", outfname, tcp->pid);
694 tcp->outf = strace_fopen(name);
701 /* Allocate some (more) TCBs (and expand the table).
702 We don't want to relocate the TCBs because our
703 callers have pointers and it would be a pain.
704 So tcbtab is a table of pointers. Since we never
705 free the TCBs, we allocate a single chunk of many. */
706 unsigned int new_tcbtabsize, alloc_tcbtabsize;
710 alloc_tcbtabsize = tcbtabsize;
711 new_tcbtabsize = tcbtabsize * 2;
713 new_tcbtabsize = alloc_tcbtabsize = 1;
716 newtcbs = xcalloc(alloc_tcbtabsize, sizeof(newtcbs[0]));
717 tcbtab = xreallocarray(tcbtab, new_tcbtabsize, sizeof(tcbtab[0]));
718 while (tcbtabsize < new_tcbtabsize)
719 tcbtab[tcbtabsize++] = newtcbs++;
728 if (nprocs == tcbtabsize)
731 for (i = 0; i < tcbtabsize; i++) {
734 memset(tcp, 0, sizeof(*tcp));
736 #if SUPPORTED_PERSONALITIES > 1
737 tcp->currpers = current_personality;
741 if (stack_trace_enabled)
742 unwind_tcb_init(tcp);
747 error_msg("new tcb for pid %d, active tcbs:%d",
752 error_msg_and_die("bug in alloctcb");
756 get_tcb_priv_data(const struct tcb *tcp)
758 return tcp->_priv_data;
762 set_tcb_priv_data(struct tcb *tcp, void *const priv_data,
763 void (*const free_priv_data)(void *))
768 tcp->_free_priv_data = free_priv_data;
769 tcp->_priv_data = priv_data;
775 free_tcb_priv_data(struct tcb *tcp)
777 if (tcp->_priv_data) {
778 if (tcp->_free_priv_data) {
779 tcp->_free_priv_data(tcp->_priv_data);
780 tcp->_free_priv_data = NULL;
782 tcp->_priv_data = NULL;
787 droptcb(struct tcb *tcp)
793 for (p = 0; p < SUPPORTED_PERSONALITIES; ++p)
794 free(tcp->fault_vec[p]);
796 free_tcb_priv_data(tcp);
799 if (stack_trace_enabled) {
806 error_msg("dropped tcb for pid %d, %d remain",
810 if (followfork >= 2) {
811 if (tcp->curcol != 0)
812 fprintf(tcp->outf, " <detached ...>\n");
815 if (printing_tcp == tcp && tcp->curcol != 0)
816 fprintf(tcp->outf, " <detached ...>\n");
821 if (current_tcp == tcp)
823 if (printing_tcp == tcp)
826 memset(tcp, 0, sizeof(*tcp));
829 /* Detach traced process.
830 * Never call DETACH twice on the same process as both unattached and
831 * attached-unstopped processes give the same ESRCH. For unattached process we
832 * would SIGSTOP it and wait for its SIGSTOP notification forever.
835 detach(struct tcb *tcp)
841 * Linux wrongly insists the child be stopped
842 * before detaching. Arghh. We go through hoops
843 * to make a clean break of things.
846 if (!(tcp->flags & TCB_ATTACHED))
849 /* We attached but possibly didn't see the expected SIGSTOP.
850 * We must catch exactly one as otherwise the detached process
851 * would be left stopped (process state T).
853 if (tcp->flags & TCB_IGNORE_ONE_SIGSTOP)
856 error = ptrace(PTRACE_DETACH, tcp->pid, 0, 0);
858 /* On a clear day, you can see forever. */
861 if (errno != ESRCH) {
862 /* Shouldn't happen. */
863 perror_msg("detach: ptrace(PTRACE_DETACH,%u)", tcp->pid);
866 /* ESRCH: process is either not stopped or doesn't exist. */
867 if (my_tkill(tcp->pid, 0) < 0) {
869 /* Shouldn't happen. */
870 perror_msg("detach: tkill(%u,0)", tcp->pid);
871 /* else: process doesn't exist. */
874 /* Process is not stopped, need to stop it. */
877 * With SEIZE, tracee can be in group-stop already.
878 * In this state sending it another SIGSTOP does nothing.
879 * Need to use INTERRUPT.
880 * Testcase: trying to ^C a "strace -p <stopped_process>".
882 error = ptrace(PTRACE_INTERRUPT, tcp->pid, 0, 0);
886 perror_msg("detach: ptrace(PTRACE_INTERRUPT,%u)", tcp->pid);
889 error = my_tkill(tcp->pid, SIGSTOP);
893 perror_msg("detach: tkill(%u,SIGSTOP)", tcp->pid);
895 /* Either process doesn't exist, or some weird error. */
899 /* We end up here in three cases:
900 * 1. We sent PTRACE_INTERRUPT (use_seize case)
901 * 2. We sent SIGSTOP (!use_seize)
902 * 3. Attach SIGSTOP was already pending (TCB_IGNORE_ONE_SIGSTOP set)
906 if (waitpid(tcp->pid, &status, __WALL) < 0) {
910 * if (errno == ECHILD) break;
911 * ^^^ WRONG! We expect this PID to exist,
912 * and want to emit a message otherwise:
914 perror_msg("detach: waitpid(%u)", tcp->pid);
917 if (!WIFSTOPPED(status)) {
919 * Tracee exited or was killed by signal.
920 * We shouldn't normally reach this place:
921 * we don't want to consume exit status.
922 * Consider "strace -p PID" being ^C-ed:
923 * we want merely to detach from PID.
925 * However, we _can_ end up here if tracee
930 sig = WSTOPSIG(status);
932 error_msg("detach wait: event:%d sig:%d",
933 (unsigned)status >> 16, sig);
935 unsigned event = (unsigned)status >> 16;
936 if (event == PTRACE_EVENT_STOP /*&& sig == SIGTRAP*/) {
938 * sig == SIGTRAP: PTRACE_INTERRUPT stop.
939 * sig == other: process was already stopped
940 * with this stopping sig (see tests/detach-stopped).
941 * Looks like re-injecting this sig is not necessary
942 * in DETACH for the tracee to remain stopped.
947 * PTRACE_INTERRUPT is not guaranteed to produce
948 * the above event if other ptrace-stop is pending.
949 * See tests/detach-sleeping testcase:
950 * strace got SIGINT while tracee is sleeping.
951 * We sent PTRACE_INTERRUPT.
952 * We see syscall exit, not PTRACE_INTERRUPT stop.
953 * We won't get PTRACE_INTERRUPT stop
954 * if we would CONT now. Need to DETACH.
956 if (sig == syscall_trap_sig)
958 /* else: not sure in which case we can be here.
959 * Signal stop? Inject it while detaching.
961 ptrace_restart(PTRACE_DETACH, tcp, sig);
964 /* Note: this check has to be after use_seize check */
965 /* (else, in use_seize case SIGSTOP will be mistreated) */
966 if (sig == SIGSTOP) {
967 /* Detach, suppressing SIGSTOP */
968 ptrace_restart(PTRACE_DETACH, tcp, 0);
971 if (sig == syscall_trap_sig)
973 /* Can't detach just yet, may need to wait for SIGSTOP */
974 error = ptrace_restart(PTRACE_CONT, tcp, sig);
976 /* Should not happen.
977 * Note: ptrace_restart returns 0 on ESRCH, so it's not it.
978 * ptrace_restart already emitted error message.
985 if (!qflag && (tcp->flags & TCB_ATTACHED))
986 error_msg("Process %u detached", tcp->pid);
992 process_opt_p_list(char *opt)
996 * We accept -p PID,PID; -p "`pidof PROG`"; -p "`pgrep PROG`".
997 * pidof uses space as delim, pgrep uses newline. :(
1000 char *delim = opt + strcspn(opt, ", \n\t");
1004 pid = string_to_uint(opt);
1006 error_msg_and_die("Invalid process id: '%s'", opt);
1008 if (pid == strace_tracer_pid) {
1009 error_msg_and_die("I'm sorry, I can't let you do that, Dave.");
1020 attach_tcb(struct tcb *const tcp)
1022 if (ptrace_attach_or_seize(tcp->pid) < 0) {
1023 perror_msg("attach: ptrace(%s, %d)",
1024 ptrace_attach_cmd, tcp->pid);
1029 tcp->flags |= TCB_ATTACHED | TCB_STARTUP | post_attach_sigstop;
1032 error_msg("attach to pid %d (main) succeeded", tcp->pid);
1034 char procdir[sizeof("/proc/%d/task") + sizeof(int) * 3];
1036 unsigned int ntid = 0, nerr = 0;
1038 if (followfork && tcp->pid != strace_child &&
1039 sprintf(procdir, "/proc/%d/task", tcp->pid) > 0 &&
1040 (dir = opendir(procdir)) != NULL) {
1043 while ((de = read_dir(dir)) != NULL) {
1044 if (de->d_fileno == 0)
1047 int tid = string_to_uint(de->d_name);
1048 if (tid <= 0 || tid == tcp->pid)
1052 if (ptrace_attach_or_seize(tid) < 0) {
1055 perror_msg("attach: ptrace(%s, %d)",
1056 ptrace_attach_cmd, tid);
1060 error_msg("attach to pid %d succeeded", tid);
1062 struct tcb *tid_tcp = alloctcb(tid);
1063 tid_tcp->flags |= TCB_ATTACHED | TCB_STARTUP |
1064 post_attach_sigstop;
1073 error_msg("Process %u attached"
1075 tcp->pid, ntid - nerr + 1);
1077 error_msg("Process %u attached",
1083 startup_attach(void)
1085 pid_t parent_pid = strace_tracer_pid;
1090 * Block user interruptions as we would leave the traced
1091 * process stopped (process state T) if we would terminate in
1092 * between PTRACE_ATTACH and wait4() on SIGSTOP.
1093 * We rely on cleanup() from this point on.
1096 sigprocmask(SIG_BLOCK, &blocked_set, NULL);
1098 if (daemonized_tracer) {
1101 perror_msg_and_die("fork");
1103 if (pid) { /* parent */
1105 * Wait for grandchild to attach to straced process
1106 * (grandparent). Grandchild SIGKILLs us after it attached.
1107 * Grandparent's wait() is unblocked by our death,
1108 * it proceeds to exec the straced program.
1111 _exit(0); /* paranoia */
1114 /* We will be the tracer process. Remember our new pid: */
1115 strace_tracer_pid = getpid();
1118 for (tcbi = 0; tcbi < tcbtabsize; tcbi++) {
1124 /* Is this a process we should attach to, but not yet attached? */
1125 if (tcp->flags & TCB_ATTACHED)
1126 continue; /* no, we already attached it */
1128 if (tcp->pid == parent_pid || tcp->pid == strace_tracer_pid) {
1130 perror_msg("attach: pid %d", tcp->pid);
1138 sigprocmask(SIG_SETMASK, &empty_set, NULL);
1141 sigprocmask(SIG_BLOCK, &blocked_set, NULL);
1143 } /* for each tcbtab[] */
1145 if (daemonized_tracer) {
1147 * Make parent go away.
1148 * Also makes grandparent's wait() unblock.
1150 kill(parent_pid, SIGKILL);
1156 sigprocmask(SIG_SETMASK, &empty_set, NULL);
1159 /* Stack-o-phobic exec helper, in the hope to work around
1160 * NOMMU + "daemonized tracer" difficulty.
1162 struct exec_params {
1169 static struct exec_params params_for_tracee;
1171 static void ATTRIBUTE_NOINLINE ATTRIBUTE_NORETURN
1174 struct exec_params *params = ¶ms_for_tracee;
1176 if (params->fd_to_close >= 0)
1177 close(params->fd_to_close);
1178 if (!daemonized_tracer && !use_seize) {
1179 if (ptrace(PTRACE_TRACEME, 0L, 0L, 0L) < 0) {
1180 perror_msg_and_die("ptrace(PTRACE_TRACEME, ...)");
1184 if (username != NULL) {
1186 * It is important to set groups before we
1187 * lose privileges on setuid.
1189 if (initgroups(username, run_gid) < 0) {
1190 perror_msg_and_die("initgroups");
1192 if (setregid(run_gid, params->run_egid) < 0) {
1193 perror_msg_and_die("setregid");
1195 if (setreuid(run_uid, params->run_euid) < 0) {
1196 perror_msg_and_die("setreuid");
1199 else if (geteuid() != 0)
1200 if (setreuid(run_uid, run_uid) < 0) {
1201 perror_msg_and_die("setreuid");
1204 if (!daemonized_tracer) {
1206 * Induce a ptrace stop. Tracer (our parent)
1207 * will resume us with PTRACE_SYSCALL and display
1208 * the immediately following execve syscall.
1209 * Can't do this on NOMMU systems, we are after
1210 * vfork: parent is blocked, stopping would deadlock.
1213 kill(getpid(), SIGSTOP);
1216 /* we depend on SIGCHLD set to SIG_DFL by init code */
1217 /* if it happens to be SIG_IGN'ed, wait won't block */
1222 execv(params->pathname, params->argv);
1223 perror_msg_and_die("exec");
1227 * Open a dummy descriptor for use as a placeholder.
1228 * The descriptor is O_RDONLY with FD_CLOEXEC flag set.
1229 * A read attempt from such descriptor ends with EOF,
1230 * a write attempt is rejected with EBADF.
1233 open_dummy_desc(void)
1238 perror_msg_and_die("pipe");
1240 set_cloexec_flag(fds[0]);
1244 /* placeholder fds status for stdin and stdout */
1245 static bool fd_is_placeholder[2];
1248 * Ensure that all standard file descriptors are open by opening placeholder
1249 * file descriptors for those standard file descriptors that are not open.
1251 * The information which descriptors have been made open is saved
1252 * in fd_is_placeholder for later use.
1255 ensure_standard_fds_opened(void)
1259 while ((fd = open_dummy_desc()) <= 2) {
1262 fd_is_placeholder[fd] = true;
1270 * Redirect stdin and stdout unless they have been opened earlier
1271 * by ensure_standard_fds_opened as placeholders.
1274 redirect_standard_fds(void)
1279 * It might be a good idea to redirect stderr as well,
1280 * but we sometimes need to print error messages.
1282 for (i = 0; i <= 1; ++i) {
1283 if (!fd_is_placeholder[i]) {
1291 startup_child(char **argv)
1293 struct_stat statbuf;
1294 const char *filename;
1295 size_t filename_len;
1296 char pathname[PATH_MAX];
1301 filename_len = strlen(filename);
1303 if (filename_len > sizeof(pathname) - 1) {
1304 errno = ENAMETOOLONG;
1305 perror_msg_and_die("exec");
1307 if (strchr(filename, '/')) {
1308 strcpy(pathname, filename);
1310 #ifdef USE_DEBUGGING_EXEC
1312 * Debuggers customarily check the current directory
1313 * first regardless of the path but doing that gives
1314 * security geeks a panic attack.
1316 else if (stat_file(filename, &statbuf) == 0)
1317 strcpy(pathname, filename);
1318 #endif /* USE_DEBUGGING_EXEC */
1323 for (path = getenv("PATH"); path && *path; path += m) {
1324 const char *colon = strchr(path, ':');
1330 m = n = strlen(path);
1332 if (!getcwd(pathname, PATH_MAX))
1334 len = strlen(pathname);
1336 else if (n > sizeof pathname - 1)
1339 strncpy(pathname, path, n);
1342 if (len && pathname[len - 1] != '/')
1343 pathname[len++] = '/';
1344 if (filename_len + len > sizeof(pathname) - 1)
1346 strcpy(pathname + len, filename);
1347 if (stat_file(pathname, &statbuf) == 0 &&
1348 /* Accept only regular files
1349 with some execute bits set.
1350 XXX not perfect, might still fail */
1351 S_ISREG(statbuf.st_mode) &&
1352 (statbuf.st_mode & 0111))
1355 if (!path || !*path)
1358 if (stat_file(pathname, &statbuf) < 0) {
1359 perror_msg_and_die("Can't stat '%s'", filename);
1362 params_for_tracee.fd_to_close = (shared_log != stderr) ? fileno(shared_log) : -1;
1363 params_for_tracee.run_euid = (statbuf.st_mode & S_ISUID) ? statbuf.st_uid : run_uid;
1364 params_for_tracee.run_egid = (statbuf.st_mode & S_ISGID) ? statbuf.st_gid : run_gid;
1365 params_for_tracee.argv = argv;
1367 * On NOMMU, can be safely freed only after execve in tracee.
1368 * It's hard to know when that happens, so we just leak it.
1370 params_for_tracee.pathname = NOMMU_SYSTEM ? xstrdup(pathname) : pathname;
1372 #if defined HAVE_PRCTL && defined PR_SET_PTRACER && defined PR_SET_PTRACER_ANY
1373 if (daemonized_tracer)
1374 prctl(PR_SET_PTRACER, PR_SET_PTRACER_ANY);
1379 perror_msg_and_die("fork");
1381 if ((pid != 0 && daemonized_tracer)
1382 || (pid == 0 && !daemonized_tracer)
1384 /* We are to become the tracee. Two cases:
1386 * not -D: we are child
1391 /* We are the tracer */
1393 if (!daemonized_tracer) {
1396 /* child did PTRACE_TRACEME, nothing to do in parent */
1398 if (!NOMMU_SYSTEM) {
1399 /* Wait until child stopped itself */
1401 while (waitpid(pid, &status, WSTOPPED) < 0) {
1404 perror_msg_and_die("waitpid");
1406 if (!WIFSTOPPED(status) || WSTOPSIG(status) != SIGSTOP) {
1407 kill_save_errno(pid, SIGKILL);
1408 perror_msg_and_die("Unexpected wait status %x", status);
1411 /* Else: NOMMU case, we have no way to sync.
1412 * Just attach to it as soon as possible.
1413 * This means that we may miss a few first syscalls...
1416 if (ptrace_attach_or_seize(pid)) {
1417 kill_save_errno(pid, SIGKILL);
1418 perror_msg_and_die("attach: ptrace(%s, %d)",
1419 ptrace_attach_cmd, pid);
1424 tcp = alloctcb(pid);
1425 tcp->flags |= TCB_ATTACHED | TCB_STARTUP
1426 | TCB_SKIP_DETACH_ON_FIRST_EXEC
1427 | (NOMMU_SYSTEM ? 0 : (TCB_HIDE_LOG | post_attach_sigstop));
1431 /* With -D, we are *child* here, the tracee is our parent. */
1432 strace_child = strace_tracer_pid;
1433 strace_tracer_pid = getpid();
1434 tcp = alloctcb(strace_child);
1435 tcp->flags |= TCB_SKIP_DETACH_ON_FIRST_EXEC | TCB_HIDE_LOG;
1436 /* attaching will be done later, by startup_attach */
1437 /* note: we don't do newoutf(tcp) here either! */
1439 /* NOMMU BUG! -D mode is active, we (child) return,
1440 * and we will scribble over parent's stack!
1441 * When parent later unpauses, it segfaults.
1444 * (1) by declaring exec_or_die() NORETURN,
1445 * hopefully compiler will just jump to it
1446 * instead of call (won't push anything to stack),
1447 * (2) by trying very hard in exec_or_die()
1448 * to not use any stack,
1449 * (3) having a really big (PATH_MAX) stack object
1450 * in this function, which creates a "buffer" between
1451 * child's and parent's stack pointers.
1452 * This may save us if (1) and (2) failed
1453 * and compiler decided to use stack in exec_or_die() anyway
1454 * (happens on i386 because of stack parameter passing).
1456 * A cleaner solution is to use makecontext + setcontext
1457 * to create a genuine separate stack and execute on it.
1461 * A case where straced process is part of a pipe:
1462 * { sleep 1; yes | head -n99999; } | strace -o/dev/null sh -c 'exec <&-; sleep 9'
1463 * If strace won't close its fd#0, closing it in tracee is not enough:
1464 * the pipe is still open, it has a reader. Thus, "head" will not get its
1465 * SIGPIPE at once, on the first write.
1467 * Preventing it by redirecting strace's stdin/out.
1468 * (Don't leave fds 0 and 1 closed, this is bad practice: future opens
1469 * will reuse them, unexpectedly making a newly opened object "stdin").
1471 redirect_standard_fds();
1476 test_ptrace_seize(void)
1480 /* Need fork for test. NOMMU has no forks */
1482 post_attach_sigstop = 0; /* this sets use_seize to 1 */
1488 perror_msg_and_die("fork");
1495 /* PTRACE_SEIZE, unlike ATTACH, doesn't force tracee to trap. After
1496 * attaching tracee continues to run unless a trap condition occurs.
1497 * PTRACE_SEIZE doesn't affect signal or group stop state.
1499 if (ptrace(PTRACE_SEIZE, pid, 0, 0) == 0) {
1500 post_attach_sigstop = 0; /* this sets use_seize to 1 */
1501 } else if (debug_flag) {
1502 error_msg("PTRACE_SEIZE doesn't work");
1508 int status, tracee_pid;
1511 tracee_pid = waitpid(pid, &status, 0);
1512 if (tracee_pid <= 0) {
1515 perror_msg_and_die("%s: unexpected wait result %d",
1516 __func__, tracee_pid);
1518 if (WIFSIGNALED(status)) {
1521 error_msg_and_die("%s: unexpected wait status %x",
1525 #else /* !USE_SEIZE */
1526 # define test_ptrace_seize() ((void)0)
1530 get_os_release(void)
1536 perror_msg_and_die("uname");
1537 /* u.release has this form: "3.2.9[-some-garbage]" */
1541 if (!(*p >= '0' && *p <= '9'))
1542 error_msg_and_die("Bad OS release string: '%s'", u.release);
1543 /* Note: this open-codes KERNEL_VERSION(): */
1544 rel = (rel << 8) | atoi(p);
1545 if (rel >= KERNEL_VERSION(1,0,0))
1547 while (*p >= '0' && *p <= '9')
1550 if (rel >= KERNEL_VERSION(0,1,0)) {
1551 /* "X.Y-something" means "X.Y.0" */
1555 error_msg_and_die("Bad OS release string: '%s'", u.release);
1563 * Initialization part of main() was eating much stack (~0.5k),
1564 * which was unused after init.
1565 * We can reuse it if we move init code into a separate function.
1567 * Don't want main() to inline us and defeat the reason
1568 * we have a separate function.
1570 static void ATTRIBUTE_NOINLINE
1571 init(int argc, char *argv[])
1575 struct sigaction sa;
1577 progname = argv[0] ? argv[0] : "strace";
1579 /* Make sure SIGCHLD has the default action so that waitpid
1580 definitely works without losing track of children. The user
1581 should not have given us a bogus state to inherit, but he might
1582 have. Arguably we should detect SIG_IGN here and pass it on
1583 to children, but probably noone really needs that. */
1584 signal(SIGCHLD, SIG_DFL);
1586 strace_tracer_pid = getpid();
1588 os_release = get_os_release();
1590 shared_log = stderr;
1591 set_sortby(DEFAULT_SORTBY);
1592 set_personality(DEFAULT_PERSONALITY);
1593 qualify("trace=all");
1594 qualify("abbrev=all");
1595 qualify("verbose=all");
1596 #if DEFAULT_QUAL_FLAGS != (QUAL_TRACE | QUAL_ABBREV | QUAL_VERBOSE)
1597 # error Bug in DEFAULT_QUAL_FLAGS
1599 qualify("signal=all");
1600 while ((c = getopt(argc, argv,
1601 "+b:cCdfFhiqrtTvVwxyz"
1602 #ifdef USE_LIBUNWIND
1606 "a:e:o:O:p:s:S:u:E:P:I:")) != EOF) {
1609 if (strcmp(optarg, "execve") != 0)
1610 error_msg_and_die("Syscall '%s' for -b isn't supported",
1612 detach_on_execve = 1;
1615 if (cflag == CFLAG_BOTH) {
1616 error_msg_and_help("-c and -C are mutually exclusive");
1618 cflag = CFLAG_ONLY_STATS;
1621 if (cflag == CFLAG_ONLY_STATS) {
1622 error_msg_and_help("-c and -C are mutually exclusive");
1630 daemonized_tracer = 1;
1657 count_wallclock = 1;
1666 qualify("abbrev=none");
1669 printf("%s -- version %s\n", PACKAGE_NAME, VERSION);
1673 not_failing_only = 1;
1676 acolumn = string_to_uint(optarg);
1678 error_opt_arg(c, optarg);
1684 outfname = xstrdup(optarg);
1687 i = string_to_uint(optarg);
1689 error_opt_arg(c, optarg);
1693 process_opt_p_list(optarg);
1696 pathtrace_select(optarg);
1699 i = string_to_uint(optarg);
1701 error_opt_arg(c, optarg);
1708 username = xstrdup(optarg);
1710 #ifdef USE_LIBUNWIND
1712 stack_trace_enabled = true;
1716 if (putenv(optarg) < 0)
1717 die_out_of_memory();
1720 opt_intr = string_to_uint_upto(optarg, NUM_INTR_OPTS - 1);
1722 error_opt_arg(c, optarg);
1725 error_msg_and_help(NULL);
1730 /* argc -= optind; - no need, argc is not used below */
1732 acolumn_spaces = xmalloc(acolumn + 1);
1733 memset(acolumn_spaces, ' ', acolumn);
1734 acolumn_spaces[acolumn] = '\0';
1736 if (!argv[0] && !nprocs) {
1737 error_msg_and_help("must have PROG [ARGS] or -p PID");
1740 if (!argv[0] && daemonized_tracer) {
1741 error_msg_and_help("PROG [ARGS] must be specified with -D");
1747 if (followfork >= 2 && cflag) {
1748 error_msg_and_help("(-c or -C) and -ff are mutually exclusive");
1751 if (count_wallclock && !cflag) {
1752 error_msg_and_help("-w must be given with (-c or -C)");
1755 if (cflag == CFLAG_ONLY_STATS) {
1757 error_msg("-%c has no effect with -c", 'i');
1758 #ifdef USE_LIBUNWIND
1759 if (stack_trace_enabled)
1760 error_msg("-%c has no effect with -c", 'k');
1763 error_msg("-%c has no effect with -c", 'r');
1765 error_msg("-%c has no effect with -c", 't');
1767 error_msg("-%c has no effect with -c", 'T');
1769 error_msg("-%c has no effect with -c", 'y');
1774 error_msg("-tt has no effect with -r");
1778 #ifdef USE_LIBUNWIND
1779 if (stack_trace_enabled) {
1783 for (tcbi = 0; tcbi < tcbtabsize; ++tcbi) {
1784 unwind_tcb_init(tcbtab[tcbi]);
1789 /* See if they want to run as another user. */
1790 if (username != NULL) {
1791 struct passwd *pent;
1793 if (getuid() != 0 || geteuid() != 0) {
1794 error_msg_and_die("You must be root to use the -u option");
1796 pent = getpwnam(username);
1798 error_msg_and_die("Cannot find user '%s'", username);
1800 run_uid = pent->pw_uid;
1801 run_gid = pent->pw_gid;
1809 ptrace_setoptions |= PTRACE_O_TRACECLONE |
1810 PTRACE_O_TRACEFORK |
1811 PTRACE_O_TRACEVFORK;
1813 error_msg("ptrace_setoptions = %#x", ptrace_setoptions);
1814 test_ptrace_seize();
1817 * Is something weird with our stdin and/or stdout -
1818 * for example, may they be not open? In this case,
1819 * ensure that none of the future opens uses them.
1821 * This was seen in the wild when /proc/sys/kernel/core_pattern
1822 * was set to "|/bin/strace -o/tmp/LOG PROG":
1823 * kernel runs coredump helper with fd#0 open but fd#1 closed (!),
1824 * therefore LOG gets opened to fd#1, and fd#1 is closed by
1825 * "don't hold up stdin/out open" code soon after.
1827 ensure_standard_fds_opened();
1829 /* Check if they want to redirect the output. */
1831 /* See if they want to pipe the output. */
1832 if (outfname[0] == '|' || outfname[0] == '!') {
1834 * We can't do the <outfname>.PID funny business
1835 * when using popen, so prohibit it.
1837 if (followfork >= 2)
1838 error_msg_and_help("piping the output and -ff are mutually exclusive");
1839 shared_log = strace_popen(outfname + 1);
1841 else if (followfork < 2)
1842 shared_log = strace_fopen(outfname);
1844 /* -ff without -o FILE is the same as single -f */
1845 if (followfork >= 2)
1849 if (!outfname || outfname[0] == '|' || outfname[0] == '!') {
1850 setvbuf(shared_log, NULL, _IOLBF, 0);
1852 if (outfname && argv[0]) {
1854 opt_intr = INTR_NEVER;
1859 opt_intr = INTR_WHILE_WAIT;
1861 /* argv[0] -pPID -oFILE Default interactive setting
1862 * yes * 0 INTR_WHILE_WAIT
1863 * no 1 0 INTR_WHILE_WAIT
1864 * yes * 1 INTR_NEVER
1865 * no 1 1 INTR_WHILE_WAIT
1868 sigemptyset(&empty_set);
1869 sigemptyset(&blocked_set);
1871 /* startup_child() must be called before the signal handlers get
1872 * installed below as they are inherited into the spawned process.
1873 * Also we do not need to be protected by them as during interruption
1874 * in the startup_child() mode we kill the spawned process anyway.
1877 startup_child(argv);
1880 sa.sa_handler = SIG_IGN;
1881 sigemptyset(&sa.sa_mask);
1883 sigaction(SIGTTOU, &sa, NULL); /* SIG_IGN */
1884 sigaction(SIGTTIN, &sa, NULL); /* SIG_IGN */
1885 if (opt_intr != INTR_ANYWHERE) {
1886 if (opt_intr == INTR_BLOCK_TSTP_TOO)
1887 sigaction(SIGTSTP, &sa, NULL); /* SIG_IGN */
1889 * In interactive mode (if no -o OUTFILE, or -p PID is used),
1890 * fatal signals are blocked while syscall stop is processed,
1891 * and acted on in between, when waiting for new syscall stops.
1892 * In non-interactive mode, signals are ignored.
1894 if (opt_intr == INTR_WHILE_WAIT) {
1895 sigaddset(&blocked_set, SIGHUP);
1896 sigaddset(&blocked_set, SIGINT);
1897 sigaddset(&blocked_set, SIGQUIT);
1898 sigaddset(&blocked_set, SIGPIPE);
1899 sigaddset(&blocked_set, SIGTERM);
1900 sa.sa_handler = interrupt;
1902 /* SIG_IGN, or set handler for these */
1903 sigaction(SIGHUP, &sa, NULL);
1904 sigaction(SIGINT, &sa, NULL);
1905 sigaction(SIGQUIT, &sa, NULL);
1906 sigaction(SIGPIPE, &sa, NULL);
1907 sigaction(SIGTERM, &sa, NULL);
1909 if (nprocs != 0 || daemonized_tracer)
1912 /* Do we want pids printed in our -o OUTFILE?
1913 * -ff: no (every pid has its own file); or
1914 * -f: yes (there can be more pids in the future); or
1915 * -p PID1,PID2: yes (there are already more than one pid)
1917 print_pid_pfx = (outfname && followfork < 2 && (followfork == 1 || nprocs > 1));
1928 for (i = 0; i < tcbtabsize; i++) {
1929 struct tcb *tcp = tcbtab[i];
1930 if (tcp->pid == pid)
1944 /* 'interrupted' is a volatile object, fetch it only once */
1945 fatal_sig = interrupted;
1947 fatal_sig = SIGTERM;
1949 for (i = 0; i < tcbtabsize; i++) {
1954 error_msg("cleanup: looking at pid %u", tcp->pid);
1955 if (tcp->pid == strace_child) {
1956 kill(tcp->pid, SIGCONT);
1957 kill(tcp->pid, fatal_sig);
1962 call_summary(shared_log);
1972 print_debug_info(const int pid, int status)
1974 const unsigned int event = (unsigned int) status >> 16;
1975 char buf[sizeof("WIFEXITED,exitcode=%u") + sizeof(int)*3 /*paranoia:*/ + 16];
1976 char evbuf[sizeof(",EVENT_VFORK_DONE (%u)") + sizeof(int)*3 /*paranoia:*/ + 16];
1979 if (WIFSIGNALED(status))
1981 sprintf(buf, "WIFSIGNALED,%ssig=%s",
1982 WCOREDUMP(status) ? "core," : "",
1983 signame(WTERMSIG(status)));
1985 sprintf(buf, "WIFSIGNALED,sig=%s",
1986 signame(WTERMSIG(status)));
1988 if (WIFEXITED(status))
1989 sprintf(buf, "WIFEXITED,exitcode=%u", WEXITSTATUS(status));
1990 if (WIFSTOPPED(status))
1991 sprintf(buf, "WIFSTOPPED,sig=%s", signame(WSTOPSIG(status)));
1993 /* Should never be seen */
1994 if (WIFCONTINUED(status))
1995 strcpy(buf, "WIFCONTINUED");
1999 static const char *const event_names[] = {
2000 [PTRACE_EVENT_CLONE] = "CLONE",
2001 [PTRACE_EVENT_FORK] = "FORK",
2002 [PTRACE_EVENT_VFORK] = "VFORK",
2003 [PTRACE_EVENT_VFORK_DONE] = "VFORK_DONE",
2004 [PTRACE_EVENT_EXEC] = "EXEC",
2005 [PTRACE_EVENT_EXIT] = "EXIT",
2006 /* [PTRACE_EVENT_STOP (=128)] would make biggish array */
2008 const char *e = "??";
2009 if (event < ARRAY_SIZE(event_names))
2010 e = event_names[event];
2011 else if (event == PTRACE_EVENT_STOP)
2013 sprintf(evbuf, ",EVENT_%s (%u)", e, event);
2015 error_msg("[wait(0x%06x) = %u] %s%s", status, pid, buf, evbuf);
2019 maybe_allocate_tcb(const int pid, int status)
2021 if (!WIFSTOPPED(status)) {
2022 if (detach_on_execve && pid == strace_child) {
2023 /* example: strace -bexecve sh -c 'exec true' */
2028 * This can happen if we inherited an unknown child.
2029 * Example: (sleep 1 & exec strace true)
2031 error_msg("Exit of unknown pid %u ignored", pid);
2035 /* We assume it's a fork/vfork/clone child */
2036 struct tcb *tcp = alloctcb(pid);
2037 tcp->flags |= TCB_ATTACHED | TCB_STARTUP | post_attach_sigstop;
2040 error_msg("Process %d attached", pid);
2043 /* This can happen if a clone call used
2044 * CLONE_PTRACE itself.
2046 ptrace(PTRACE_CONT, pid, NULL, 0);
2047 error_msg("Stop of unknown pid %u seen, PTRACE_CONTed it", pid);
2053 maybe_switch_tcbs(struct tcb *tcp, const int pid)
2056 struct tcb *execve_thread;
2059 if (ptrace(PTRACE_GETEVENTMSG, pid, NULL, &old_pid) < 0)
2061 /* Avoid truncation in pid2tcb() param passing */
2062 if (old_pid <= 0 || old_pid == pid)
2064 if ((unsigned long) old_pid > UINT_MAX)
2066 execve_thread = pid2tcb(old_pid);
2067 /* It should be !NULL, but I feel paranoid */
2071 if (execve_thread->curcol != 0) {
2073 * One case we are here is -ff:
2074 * try "strace -oLOG -ff test/threaded_execve"
2076 fprintf(execve_thread->outf, " <pid changed to %d ...>\n", pid);
2077 /*execve_thread->curcol = 0; - no need, see code below */
2079 /* Swap output FILEs (needed for -ff) */
2080 fp = execve_thread->outf;
2081 execve_thread->outf = tcp->outf;
2083 /* And their column positions */
2084 execve_thread->curcol = tcp->curcol;
2086 /* Drop leader, but close execve'd thread outfile (if -ff) */
2088 /* Switch to the thread, reusing leader's outfile and pid */
2089 tcp = execve_thread;
2091 if (cflag != CFLAG_ONLY_STATS) {
2093 tprintf("+++ superseded by execve in pid %lu +++\n", old_pid);
2095 tcp->flags |= TCB_REPRINT;
2102 print_signalled(struct tcb *tcp, const int pid, int status)
2104 if (pid == strace_child) {
2105 exit_code = 0x100 | WTERMSIG(status);
2109 if (cflag != CFLAG_ONLY_STATS
2110 && is_number_in_set(WTERMSIG(status), &signal_set)) {
2113 tprintf("+++ killed by %s %s+++\n",
2114 signame(WTERMSIG(status)),
2115 WCOREDUMP(status) ? "(core dumped) " : "");
2117 tprintf("+++ killed by %s +++\n",
2118 signame(WTERMSIG(status)));
2125 print_exited(struct tcb *tcp, const int pid, int status)
2127 if (pid == strace_child) {
2128 exit_code = WEXITSTATUS(status);
2132 if (cflag != CFLAG_ONLY_STATS &&
2135 tprintf("+++ exited with %d +++\n", WEXITSTATUS(status));
2141 print_stopped(struct tcb *tcp, const siginfo_t *si, const unsigned int sig)
2143 if (cflag != CFLAG_ONLY_STATS
2145 && is_number_in_set(sig, &signal_set)) {
2148 tprintf("--- %s ", signame(sig));
2152 tprintf("--- stopped by %s ---\n", signame(sig));
2158 startup_tcb(struct tcb *tcp)
2161 error_msg("pid %d has TCB_STARTUP, initializing it", tcp->pid);
2163 tcp->flags &= ~TCB_STARTUP;
2167 error_msg("setting opts 0x%x on pid %d",
2168 ptrace_setoptions, tcp->pid);
2169 if (ptrace(PTRACE_SETOPTIONS, tcp->pid, NULL, ptrace_setoptions) < 0) {
2170 if (errno != ESRCH) {
2171 /* Should never happen, really */
2172 perror_msg_and_die("PTRACE_SETOPTIONS");
2179 print_event_exit(struct tcb *tcp)
2181 if (entering(tcp) || filtered(tcp) || hide_log(tcp)
2182 || cflag == CFLAG_ONLY_STATS) {
2186 if (followfork < 2 && printing_tcp && printing_tcp != tcp
2187 && printing_tcp->curcol != 0) {
2188 current_tcp = printing_tcp;
2189 tprints(" <unfinished ...>\n");
2190 fflush(printing_tcp->outf);
2191 printing_tcp->curcol = 0;
2195 if ((followfork < 2 && printing_tcp != tcp)
2196 || (tcp->flags & TCB_REPRINT)) {
2197 tcp->flags &= ~TCB_REPRINT;
2199 tprintf("<... %s resumed>", tcp->s_ent->sys_name);
2202 if (!(tcp->sys_func_rval & RVAL_DECODED)) {
2204 * The decoder has probably decided to print something
2205 * on exiting syscall which is not going to happen.
2207 tprints(" <unfinished ...>");
2215 /* Returns true iff the main trace loop has to continue. */
2232 * Used to exit simply when nprocs hits zero, but in this testcase:
2233 * int main() { _exit(!!fork()); }
2234 * under strace -f, parent sometimes (rarely) manages
2235 * to exit before we see the first stop of the child,
2236 * and we are losing track of it:
2237 * 19923 clone(...) = 19924
2238 * 19923 exit_group(1) = ?
2239 * 19923 +++ exited with 1 +++
2240 * Exiting only when wait() returns ECHILD works better.
2242 if (popen_pid != 0) {
2243 /* However, if -o|logger is in use, we can't do that.
2244 * Can work around that by double-forking the logger,
2245 * but that loses the ability to wait for its completion
2246 * on exit. Oh well...
2253 sigprocmask(SIG_SETMASK, &empty_set, NULL);
2254 pid = wait4(-1, &status, __WALL, (cflag ? &ru : NULL));
2257 sigprocmask(SIG_BLOCK, &blocked_set, NULL);
2260 if (wait_errno == EINTR)
2262 if (nprocs == 0 && wait_errno == ECHILD)
2265 * If nprocs > 0, ECHILD is not expected,
2266 * treat it as any other error here:
2269 perror_msg_and_die("wait4(__WALL)");
2272 if (pid == popen_pid) {
2273 if (!WIFSTOPPED(status))
2279 print_debug_info(pid, status);
2281 /* Look up 'pid' in our table. */
2285 tcp = maybe_allocate_tcb(pid, status);
2290 if (WIFSTOPPED(status))
2295 event = (unsigned int) status >> 16;
2297 if (event == PTRACE_EVENT_EXEC) {
2299 * Under Linux, execve changes pid to thread leader's pid,
2300 * and we see this changed pid on EVENT_EXEC and later,
2301 * execve sysexit. Leader "disappears" without exit
2302 * notification. Let user know that, drop leader's tcb,
2303 * and fix up pid in execve thread's tcb.
2304 * Effectively, execve thread's tcb replaces leader's tcb.
2306 * BTW, leader is 'stuck undead' (doesn't report WIFEXITED
2307 * on exit syscall) in multithreaded programs exactly
2308 * in order to handle this case.
2310 * PTRACE_GETEVENTMSG returns old pid starting from Linux 3.0.
2311 * On 2.6 and earlier, it can return garbage.
2313 if (os_release >= KERNEL_VERSION(3,0,0))
2314 tcp = maybe_switch_tcbs(tcp, pid);
2316 if (detach_on_execve) {
2317 if (tcp->flags & TCB_SKIP_DETACH_ON_FIRST_EXEC) {
2318 tcp->flags &= ~TCB_SKIP_DETACH_ON_FIRST_EXEC;
2320 detach(tcp); /* do "-b execve" thingy */
2326 /* Set current output file */
2330 tv_sub(&tcp->dtime, &ru.ru_stime, &tcp->stime);
2331 tcp->stime = ru.ru_stime;
2334 if (WIFSIGNALED(status)) {
2335 print_signalled(tcp, pid, status);
2340 if (WIFEXITED(status)) {
2341 print_exited(tcp, pid, status);
2346 if (!WIFSTOPPED(status)) {
2348 * Neither signalled, exited or stopped.
2349 * How could that be?
2351 error_msg("pid %u not stopped!", pid);
2356 /* Is this the very first time we see this tracee stopped? */
2357 if (tcp->flags & TCB_STARTUP) {
2359 if (get_scno(tcp) == 1)
2360 tcp->s_prev_ent = tcp->s_ent;
2363 sig = WSTOPSIG(status);
2368 case PTRACE_EVENT_EXIT:
2369 print_event_exit(tcp);
2370 goto restart_tracee_with_sig_0;
2372 case PTRACE_EVENT_STOP:
2374 * PTRACE_INTERRUPT-stop or group-stop.
2375 * PTRACE_INTERRUPT-stop has sig == SIGTRAP here.
2388 goto restart_tracee_with_sig_0;
2392 * Is this post-attach SIGSTOP?
2393 * Interestingly, the process may stop
2394 * with STOPSIG equal to some other signal
2395 * than SIGSTOP if we happend to attach
2396 * just before the process takes a signal.
2398 if (sig == SIGSTOP && (tcp->flags & TCB_IGNORE_ONE_SIGSTOP)) {
2400 error_msg("ignored SIGSTOP on pid %d", tcp->pid);
2401 tcp->flags &= ~TCB_IGNORE_ONE_SIGSTOP;
2402 goto restart_tracee_with_sig_0;
2405 if (sig != syscall_trap_sig) {
2409 * True if tracee is stopped by signal
2410 * (as opposed to "tracee received signal").
2411 * TODO: shouldn't we check for errno == EINVAL too?
2412 * We can get ESRCH instead, you know...
2414 stopped = ptrace(PTRACE_GETSIGINFO, pid, 0, &si) < 0;
2418 print_stopped(tcp, stopped ? NULL : &si, sig);
2421 /* It's signal-delivery-stop. Inject the signal */
2422 goto restart_tracee;
2424 /* It's group-stop */
2427 * This ends ptrace-stop, but does *not* end group-stop.
2428 * This makes stopping signals work properly on straced process
2429 * (that is, process really stops. It used to continue to run).
2431 if (ptrace_restart(PTRACE_LISTEN, tcp, 0) < 0) {
2432 /* Note: ptrace_restart emitted error message */
2438 /* We don't have PTRACE_LISTEN support... */
2439 goto restart_tracee;
2442 /* We handled quick cases, we are permitted to interrupt now. */
2447 * This should be syscall entry or exit.
2450 if (trace_syscall(tcp) < 0) {
2452 * ptrace() failed in trace_syscall().
2453 * Likely a result of process disappearing mid-flight.
2454 * Observed case: exit_group() or SIGKILL terminating
2455 * all processes in thread group.
2456 * We assume that ptrace error was caused by process death.
2457 * We used to detach(tcp) here, but since we no longer
2458 * implement "detach before death" policy/hack,
2459 * we can let this process to report its death to us
2460 * normally, via WIFEXITED or WIFSIGNALED wait status.
2465 restart_tracee_with_sig_0:
2469 if (ptrace_restart(PTRACE_SYSCALL, tcp, sig) < 0) {
2470 /* Note: ptrace_restart emitted error message */
2479 main(int argc, char *argv[])
2483 exit_code = !nprocs;
2490 if (shared_log != stderr)
2493 while (waitpid(popen_pid, NULL, 0) < 0 && errno == EINTR)
2496 if (exit_code > 0xff) {
2497 /* Avoid potential core file clobbering. */
2498 struct_rlimit rlim = {0, 0};
2499 set_rlimit(RLIMIT_CORE, &rlim);
2501 /* Child was killed by a signal, mimic that. */
2503 signal(exit_code, SIG_DFL);
2505 /* Paranoia - what if this signal is not fatal?
2506 Exit with 128 + signo then. */