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 * Copyright (c) 1999-2017 The strace developers.
9 * Redistribution and use in source and binary forms, with or without
10 * modification, are permitted provided that the following conditions
12 * 1. Redistributions of source code must retain the above copyright
13 * notice, this list of conditions and the following disclaimer.
14 * 2. Redistributions in binary form must reproduce the above copyright
15 * notice, this list of conditions and the following disclaimer in the
16 * documentation and/or other materials provided with the distribution.
17 * 3. The name of the author may not be used to endorse or promote products
18 * derived from this software without specific prior written permission.
20 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
21 * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
22 * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
23 * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
24 * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
25 * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
26 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
27 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
28 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
29 * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
34 #include <sys/param.h>
37 #include <sys/resource.h>
43 #include <sys/utsname.h>
45 # include <sys/prctl.h>
47 #include <asm/unistd.h>
51 #include "printsiginfo.h"
53 /* In some libc, these aren't declared. Do it ourself: */
54 extern char **environ;
59 /* if this is true do the stack trace for every system call */
60 bool stack_trace_enabled = false;
63 #define my_tkill(tid, sig) syscall(__NR_tkill, (tid), (sig))
65 /* Glue for systems without a MMU that cannot provide fork() */
66 #if !defined(HAVE_FORK)
68 # define NOMMU_SYSTEM 1
71 # define fork() vfork()
74 const unsigned int syscall_trap_sig = SIGTRAP | 0x80;
76 cflag_t cflag = CFLAG_NONE;
77 unsigned int followfork = 0;
78 unsigned int ptrace_setoptions = PTRACE_O_TRACESYSGOOD | PTRACE_O_TRACEEXEC
80 unsigned int xflag = 0;
84 bool count_wallclock = 0;
85 unsigned int qflag = 0;
86 static unsigned int tflag = 0;
87 static bool rflag = 0;
88 static bool print_pid_pfx = 0;
93 INTR_ANYWHERE = 1, /* don't block/ignore any signals */
94 INTR_WHILE_WAIT = 2, /* block fatal signals while decoding syscall. default */
95 INTR_NEVER = 3, /* block fatal signals. default if '-o FILE PROG' */
96 INTR_BLOCK_TSTP_TOO = 4, /* block fatal signals and SIGTSTP (^Z) */
100 /* We play with signal mask only if this mode is active: */
101 #define interactive (opt_intr == INTR_WHILE_WAIT)
104 * daemonized_tracer supports -D option.
105 * With this option, strace forks twice.
106 * Unlike normal case, with -D *grandparent* process exec's,
107 * becoming a traced process. Child exits (this prevents traced process
108 * from having children it doesn't expect to have), and grandchild
109 * attaches to grandparent similarly to strace -p PID.
110 * This allows for more transparent interaction in cases
111 * when process and its parent are communicating via signals,
112 * wait() etc. Without -D, strace process gets lodged in between,
113 * disrupting parent<->child link.
115 static bool daemonized_tracer = 0;
118 static int post_attach_sigstop = TCB_IGNORE_ONE_SIGSTOP;
119 # define use_seize (post_attach_sigstop == 0)
121 # define post_attach_sigstop TCB_IGNORE_ONE_SIGSTOP
125 /* Sometimes we want to print only succeeding syscalls. */
126 bool not_failing_only = 0;
128 /* Show path associated with fd arguments */
129 unsigned int show_fd_path = 0;
131 static bool detach_on_execve = 0;
133 static int exit_code;
134 static int strace_child = 0;
135 static int strace_tracer_pid = 0;
137 static char *username = NULL;
138 static uid_t run_uid;
139 static gid_t run_gid;
141 unsigned int max_strlen = DEFAULT_STRLEN;
142 static int acolumn = DEFAULT_ACOLUMN;
143 static char *acolumn_spaces;
145 static char *outfname = NULL;
146 /* If -ff, points to stderr. Else, it's our common output log */
147 static FILE *shared_log;
149 struct tcb *printing_tcp = NULL;
150 static struct tcb *current_tcp;
152 static struct tcb **tcbtab;
153 static unsigned int nprocs, tcbtabsize;
154 static const char *progname;
156 unsigned os_release; /* generated from uname()'s u.release */
158 static void detach(struct tcb *tcp);
159 static void cleanup(void);
160 static void interrupt(int sig);
161 static sigset_t start_set, blocked_set;
163 #ifdef HAVE_SIG_ATOMIC_T
164 static volatile sig_atomic_t interrupted;
166 static volatile int interrupted;
169 #ifndef HAVE_STRERROR
171 #if !HAVE_DECL_SYS_ERRLIST
173 extern char *sys_errlist[];
179 static char buf[sizeof("Unknown error %d") + sizeof(int)*3];
181 if (err_no < 1 || err_no >= sys_nerr) {
182 sprintf(buf, "Unknown error %d", err_no);
185 return sys_errlist[err_no];
188 #endif /* HAVE_STERRROR */
193 printf("%s -- version %s\n"
194 "Copyright (c) 1991-%s The strace developers <%s>.\n"
195 "This is free software; see the source for copying conditions. There is NO\n"
196 "warranty; not even for MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.\n",
197 PACKAGE_NAME, PACKAGE_VERSION, COPYRIGHT_YEAR, PACKAGE_URL);
204 usage: strace [-CdffhiqrtttTvVwxxy] [-I n] [-e expr]...\n\
205 [-a column] [-o file] [-s strsize] [-P path]...\n\
206 -p pid... / [-D] [-E var=val]... [-u username] PROG [ARGS]\n\
207 or: strace -c[dfw] [-I n] [-e expr]... [-O overhead] [-S sortby]\n\
208 -p pid... / [-D] [-E var=val]... [-u username] PROG [ARGS]\n\
211 -a column alignment COLUMN for printing syscall results (default %d)\n\
212 -i print instruction pointer at time of syscall\n\
216 -k obtain stack trace between each syscall (experimental)\n\
220 -o file send trace output to FILE instead of stderr\n\
221 -q suppress messages about attaching, detaching, etc.\n\
222 -r print relative timestamp\n\
223 -s strsize limit length of print strings to STRSIZE chars (default %d)\n\
224 -t print absolute timestamp\n\
225 -tt print absolute timestamp with usecs\n\
226 -T print time spent in each syscall\n\
227 -x print non-ascii strings in hex\n\
228 -xx print all strings in hex\n\
229 -y print paths associated with file descriptor arguments\n\
230 -yy print protocol specific information associated with socket file descriptors\n\
233 -c count time, calls, and errors for each syscall and report summary\n\
234 -C like -c but also print regular output\n\
235 -O overhead set overhead for tracing syscalls to OVERHEAD usecs\n\
236 -S sortby sort syscall counts by: time, calls, name, nothing (default %s)\n\
237 -w summarise syscall latency (default is system time)\n\
240 -e expr a qualifying expression: option=[!]all or option=[!]val1[,val2]...\n\
241 options: trace, abbrev, verbose, raw, signal, read, write, fault\n\
242 -P path trace accesses to path\n\
245 -b execve detach on execve syscall\n\
246 -D run tracer process as a detached grandchild, not as parent\n\
248 -ff follow forks with output into separate files\n\
250 1: no signals are blocked\n\
251 2: fatal signals are blocked while decoding syscall (default)\n\
252 3: fatal signals are always blocked (default if '-o FILE PROG')\n\
253 4: fatal signals and SIGTSTP (^Z) are always blocked\n\
254 (useful to make 'strace -o FILE PROG' not stop on ^Z)\n\
257 -E var remove var from the environment for command\n\
258 -E var=val put var=val in the environment for command\n\
259 -p pid trace process with process id PID, may be repeated\n\
260 -u username run command as username handling setuid and/or setgid\n\
263 -d enable debug output to stderr\n\
264 -v verbose mode: print unabbreviated argv, stat, termios, etc. args\n\
265 -h print help message\n\
268 /* ancient, no one should use it
269 -F -- attempt to follow vforks (deprecated, use -f)\n\
271 /* this is broken, so don't document it
272 -z -- print only succeeding syscalls\n\
274 , DEFAULT_ACOLUMN, DEFAULT_STRLEN, DEFAULT_SORTBY);
278 static void ATTRIBUTE_NORETURN
281 if (strace_tracer_pid == getpid()) {
288 static void verror_msg(int err_no, const char *fmt, va_list p)
294 /* We want to print entire message with single fprintf to ensure
295 * message integrity if stderr is shared with other programs.
296 * Thus we use vasprintf + single fprintf.
299 if (vasprintf(&msg, fmt, p) >= 0) {
301 fprintf(stderr, "%s: %s: %s\n", progname, msg, strerror(err_no));
303 fprintf(stderr, "%s: %s\n", progname, msg);
306 /* malloc in vasprintf failed, try it without malloc */
307 fprintf(stderr, "%s: ", progname);
308 vfprintf(stderr, fmt, p);
310 fprintf(stderr, ": %s\n", strerror(err_no));
314 /* We don't switch stderr to buffered, thus fprintf(stderr)
315 * always flushes its output and this is not necessary: */
316 /* fflush(stderr); */
319 void error_msg(const char *fmt, ...)
323 verror_msg(0, fmt, p);
327 void error_msg_and_die(const char *fmt, ...)
331 verror_msg(0, fmt, p);
335 void error_msg_and_help(const char *fmt, ...)
340 verror_msg(0, fmt, p);
342 fprintf(stderr, "Try '%s -h' for more information.\n", progname);
346 void perror_msg(const char *fmt, ...)
350 verror_msg(errno, fmt, p);
354 void perror_msg_and_die(const char *fmt, ...)
358 verror_msg(errno, fmt, p);
363 error_opt_arg(int opt, const char *arg)
365 error_msg_and_help("invalid -%c argument: '%s'", opt, arg);
368 static const char *ptrace_attach_cmd;
371 ptrace_attach_or_seize(int pid)
376 return ptrace_attach_cmd = "PTRACE_ATTACH",
377 ptrace(PTRACE_ATTACH, pid, 0L, 0L);
378 r = ptrace(PTRACE_SEIZE, pid, 0L, (unsigned long) ptrace_setoptions);
380 return ptrace_attach_cmd = "PTRACE_SEIZE", r;
381 r = ptrace(PTRACE_INTERRUPT, pid, 0L, 0L);
382 return ptrace_attach_cmd = "PTRACE_INTERRUPT", r;
384 return ptrace_attach_cmd = "PTRACE_ATTACH",
385 ptrace(PTRACE_ATTACH, pid, 0L, 0L);
390 * Used when we want to unblock stopped traced process.
391 * Should be only used with PTRACE_CONT, PTRACE_DETACH and PTRACE_SYSCALL.
392 * Returns 0 on success or if error was ESRCH
393 * (presumably process was killed while we talk to it).
394 * Otherwise prints error message and returns -1.
397 ptrace_restart(const unsigned int op, struct tcb *const tcp, unsigned int sig)
403 ptrace(op, tcp->pid, 0L, (unsigned long) sig);
423 * Why curcol != 0? Otherwise sometimes we get this:
425 * 10252 kill(10253, SIGKILL) = 0
426 * <ptrace(SYSCALL,10252):No such process>10253 ...next decode...
428 * 10252 died after we retrieved syscall exit data,
429 * but before we tried to restart it. Log looks ugly.
431 if (current_tcp && current_tcp->curcol != 0) {
432 tprintf(" <ptrace(%s):%s>\n", msg, strerror(err));
438 perror_msg("ptrace(PTRACE_%s,pid:%d,sig:%u)", msg, tcp->pid, sig);
443 set_cloexec_flag(int fd)
447 flags = fcntl(fd, F_GETFD);
449 /* Can happen only if fd is bad.
450 * Should never happen: if it does, we have a bug
451 * in the caller. Therefore we just abort
452 * instead of propagating the error.
454 perror_msg_and_die("fcntl(%d, F_GETFD)", fd);
457 newflags = flags | FD_CLOEXEC;
458 if (flags == newflags)
461 fcntl(fd, F_SETFD, newflags); /* never fails */
465 kill_save_errno(pid_t pid, int sig)
467 int saved_errno = errno;
469 (void) kill(pid, sig);
474 * When strace is setuid executable, we have to swap uids
475 * before and after filesystem and process management operations.
480 int euid = geteuid(), uid = getuid();
482 if (euid != uid && setreuid(euid, uid) < 0) {
483 perror_msg_and_die("setreuid");
487 #ifdef _LARGEFILE64_SOURCE
489 # define fopen_for_output fopen64
491 # define fopen_for_output fopen
493 # define struct_stat struct stat64
494 # define stat_file stat64
495 # define struct_dirent struct dirent64
496 # define read_dir readdir64
497 # define struct_rlimit struct rlimit64
498 # define set_rlimit setrlimit64
500 # define fopen_for_output fopen
501 # define struct_stat struct stat
502 # define stat_file stat
503 # define struct_dirent struct dirent
504 # define read_dir readdir
505 # define struct_rlimit struct rlimit
506 # define set_rlimit setrlimit
510 strace_fopen(const char *path)
515 fp = fopen_for_output(path, "w");
517 perror_msg_and_die("Can't fopen '%s'", path);
519 set_cloexec_flag(fileno(fp));
523 static int popen_pid = 0;
526 # define _PATH_BSHELL "/bin/sh"
530 * We cannot use standard popen(3) here because we have to distinguish
531 * popen child process from other processes we trace, and standard popen(3)
532 * does not export its child's pid.
535 strace_popen(const char *command)
543 perror_msg_and_die("pipe");
545 set_cloexec_flag(fds[1]); /* never fails */
549 perror_msg_and_die("vfork");
556 perror_msg_and_die("dup2");
559 execl(_PATH_BSHELL, "sh", "-c", command, NULL);
560 perror_msg_and_die("Can't execute '%s'", _PATH_BSHELL);
567 fp = fdopen(fds[1], "w");
573 ATTRIBUTE_FORMAT((printf, 1, 0))
575 tvprintf(const char *const fmt, va_list args)
578 int n = vfprintf(current_tcp->outf, fmt, args);
580 if (current_tcp->outf != stderr)
581 perror_msg("%s", outfname);
583 current_tcp->curcol += n;
588 tprintf(const char *fmt, ...)
596 #ifndef HAVE_FPUTS_UNLOCKED
597 # define fputs_unlocked fputs
601 tprints(const char *str)
604 int n = fputs_unlocked(str, current_tcp->outf);
606 current_tcp->curcol += strlen(str);
609 if (current_tcp->outf != stderr)
610 perror_msg("%s", outfname);
615 tprints_comment(const char *const str)
618 tprintf(" /* %s */", str);
622 tprintf_comment(const char *fmt, ...)
639 current_tcp->curcol = 0;
640 fflush(current_tcp->outf);
643 printing_tcp->curcol = 0;
649 printleader(struct tcb *tcp)
651 /* If -ff, "previous tcb we printed" is always the same as current,
652 * because we have per-tcb output files.
658 current_tcp = printing_tcp;
659 if (printing_tcp->curcol != 0 && (followfork < 2 || printing_tcp == tcp)) {
661 * case 1: we have a shared log (i.e. not -ff), and last line
662 * wasn't finished (same or different tcb, doesn't matter).
663 * case 2: split log, we are the same tcb, but our last line
664 * didn't finish ("SIGKILL nuked us after syscall entry" etc).
666 tprints(" <unfinished ...>\n");
667 printing_tcp->curcol = 0;
673 current_tcp->curcol = 0;
676 tprintf("%-5d ", tcp->pid);
677 else if (nprocs > 1 && !outfname)
678 tprintf("[pid %5u] ", tcp->pid);
681 char str[sizeof("HH:MM:SS")];
682 struct timeval tv, dtv;
683 static struct timeval otv;
685 gettimeofday(&tv, NULL);
689 tv_sub(&dtv, &tv, &otv);
690 tprintf("%6ld.%06ld ",
691 (long) dtv.tv_sec, (long) dtv.tv_usec);
694 else if (tflag > 2) {
695 tprintf("%ld.%06ld ",
696 (long) tv.tv_sec, (long) tv.tv_usec);
699 time_t local = tv.tv_sec;
700 strftime(str, sizeof(str), "%T", localtime(&local));
702 tprintf("%s.%06ld ", str, (long) tv.tv_usec);
714 if (current_tcp->curcol < acolumn)
715 tprints(acolumn_spaces + current_tcp->curcol);
718 /* Should be only called directly *after successful attach* to a tracee.
719 * Otherwise, "strace -oFILE -ff -p<nonexistant_pid>"
720 * may create bogus empty FILE.<nonexistant_pid>, and then die.
723 newoutf(struct tcb *tcp)
725 tcp->outf = shared_log; /* if not -ff mode, the same file is for all */
726 if (followfork >= 2) {
727 char name[520 + sizeof(int) * 3];
728 sprintf(name, "%.512s.%u", outfname, tcp->pid);
729 tcp->outf = strace_fopen(name);
736 /* Allocate some (more) TCBs (and expand the table).
737 We don't want to relocate the TCBs because our
738 callers have pointers and it would be a pain.
739 So tcbtab is a table of pointers. Since we never
740 free the TCBs, we allocate a single chunk of many. */
741 unsigned int new_tcbtabsize, alloc_tcbtabsize;
745 alloc_tcbtabsize = tcbtabsize;
746 new_tcbtabsize = tcbtabsize * 2;
748 new_tcbtabsize = alloc_tcbtabsize = 1;
751 newtcbs = xcalloc(alloc_tcbtabsize, sizeof(newtcbs[0]));
752 tcbtab = xreallocarray(tcbtab, new_tcbtabsize, sizeof(tcbtab[0]));
753 while (tcbtabsize < new_tcbtabsize)
754 tcbtab[tcbtabsize++] = newtcbs++;
763 if (nprocs == tcbtabsize)
766 for (i = 0; i < tcbtabsize; i++) {
769 memset(tcp, 0, sizeof(*tcp));
771 #if SUPPORTED_PERSONALITIES > 1
772 tcp->currpers = current_personality;
776 if (stack_trace_enabled)
777 unwind_tcb_init(tcp);
782 error_msg("new tcb for pid %d, active tcbs:%d",
787 error_msg_and_die("bug in alloctcb");
791 get_tcb_priv_data(const struct tcb *tcp)
793 return tcp->_priv_data;
797 set_tcb_priv_data(struct tcb *tcp, void *const priv_data,
798 void (*const free_priv_data)(void *))
803 tcp->_free_priv_data = free_priv_data;
804 tcp->_priv_data = priv_data;
810 free_tcb_priv_data(struct tcb *tcp)
812 if (tcp->_priv_data) {
813 if (tcp->_free_priv_data) {
814 tcp->_free_priv_data(tcp->_priv_data);
815 tcp->_free_priv_data = NULL;
817 tcp->_priv_data = NULL;
822 droptcb(struct tcb *tcp)
828 for (p = 0; p < SUPPORTED_PERSONALITIES; ++p)
829 free(tcp->inject_vec[p]);
831 free_tcb_priv_data(tcp);
834 if (stack_trace_enabled) {
841 error_msg("dropped tcb for pid %d, %d remain",
845 if (followfork >= 2) {
846 if (tcp->curcol != 0)
847 fprintf(tcp->outf, " <detached ...>\n");
850 if (printing_tcp == tcp && tcp->curcol != 0)
851 fprintf(tcp->outf, " <detached ...>\n");
856 if (current_tcp == tcp)
858 if (printing_tcp == tcp)
861 memset(tcp, 0, sizeof(*tcp));
864 /* Detach traced process.
865 * Never call DETACH twice on the same process as both unattached and
866 * attached-unstopped processes give the same ESRCH. For unattached process we
867 * would SIGSTOP it and wait for its SIGSTOP notification forever.
870 detach(struct tcb *tcp)
876 * Linux wrongly insists the child be stopped
877 * before detaching. Arghh. We go through hoops
878 * to make a clean break of things.
881 if (!(tcp->flags & TCB_ATTACHED))
884 /* We attached but possibly didn't see the expected SIGSTOP.
885 * We must catch exactly one as otherwise the detached process
886 * would be left stopped (process state T).
888 if (tcp->flags & TCB_IGNORE_ONE_SIGSTOP)
891 error = ptrace(PTRACE_DETACH, tcp->pid, 0, 0);
893 /* On a clear day, you can see forever. */
896 if (errno != ESRCH) {
897 /* Shouldn't happen. */
898 perror_msg("detach: ptrace(PTRACE_DETACH,%u)", tcp->pid);
901 /* ESRCH: process is either not stopped or doesn't exist. */
902 if (my_tkill(tcp->pid, 0) < 0) {
904 /* Shouldn't happen. */
905 perror_msg("detach: tkill(%u,0)", tcp->pid);
906 /* else: process doesn't exist. */
909 /* Process is not stopped, need to stop it. */
912 * With SEIZE, tracee can be in group-stop already.
913 * In this state sending it another SIGSTOP does nothing.
914 * Need to use INTERRUPT.
915 * Testcase: trying to ^C a "strace -p <stopped_process>".
917 error = ptrace(PTRACE_INTERRUPT, tcp->pid, 0, 0);
921 perror_msg("detach: ptrace(PTRACE_INTERRUPT,%u)", tcp->pid);
924 error = my_tkill(tcp->pid, SIGSTOP);
928 perror_msg("detach: tkill(%u,SIGSTOP)", tcp->pid);
930 /* Either process doesn't exist, or some weird error. */
934 /* We end up here in three cases:
935 * 1. We sent PTRACE_INTERRUPT (use_seize case)
936 * 2. We sent SIGSTOP (!use_seize)
937 * 3. Attach SIGSTOP was already pending (TCB_IGNORE_ONE_SIGSTOP set)
941 if (waitpid(tcp->pid, &status, __WALL) < 0) {
945 * if (errno == ECHILD) break;
946 * ^^^ WRONG! We expect this PID to exist,
947 * and want to emit a message otherwise:
949 perror_msg("detach: waitpid(%u)", tcp->pid);
952 if (!WIFSTOPPED(status)) {
954 * Tracee exited or was killed by signal.
955 * We shouldn't normally reach this place:
956 * we don't want to consume exit status.
957 * Consider "strace -p PID" being ^C-ed:
958 * we want merely to detach from PID.
960 * However, we _can_ end up here if tracee
965 sig = WSTOPSIG(status);
967 error_msg("detach wait: event:%d sig:%d",
968 (unsigned)status >> 16, sig);
970 unsigned event = (unsigned)status >> 16;
971 if (event == PTRACE_EVENT_STOP /*&& sig == SIGTRAP*/) {
973 * sig == SIGTRAP: PTRACE_INTERRUPT stop.
974 * sig == other: process was already stopped
975 * with this stopping sig (see tests/detach-stopped).
976 * Looks like re-injecting this sig is not necessary
977 * in DETACH for the tracee to remain stopped.
982 * PTRACE_INTERRUPT is not guaranteed to produce
983 * the above event if other ptrace-stop is pending.
984 * See tests/detach-sleeping testcase:
985 * strace got SIGINT while tracee is sleeping.
986 * We sent PTRACE_INTERRUPT.
987 * We see syscall exit, not PTRACE_INTERRUPT stop.
988 * We won't get PTRACE_INTERRUPT stop
989 * if we would CONT now. Need to DETACH.
991 if (sig == syscall_trap_sig)
993 /* else: not sure in which case we can be here.
994 * Signal stop? Inject it while detaching.
996 ptrace_restart(PTRACE_DETACH, tcp, sig);
999 /* Note: this check has to be after use_seize check */
1000 /* (else, in use_seize case SIGSTOP will be mistreated) */
1001 if (sig == SIGSTOP) {
1002 /* Detach, suppressing SIGSTOP */
1003 ptrace_restart(PTRACE_DETACH, tcp, 0);
1006 if (sig == syscall_trap_sig)
1008 /* Can't detach just yet, may need to wait for SIGSTOP */
1009 error = ptrace_restart(PTRACE_CONT, tcp, sig);
1011 /* Should not happen.
1012 * Note: ptrace_restart returns 0 on ESRCH, so it's not it.
1013 * ptrace_restart already emitted error message.
1020 if (!qflag && (tcp->flags & TCB_ATTACHED))
1021 error_msg("Process %u detached", tcp->pid);
1027 process_opt_p_list(char *opt)
1031 * We accept -p PID,PID; -p "`pidof PROG`"; -p "`pgrep PROG`".
1032 * pidof uses space as delim, pgrep uses newline. :(
1035 char *delim = opt + strcspn(opt, ", \n\t");
1039 pid = string_to_uint(opt);
1041 error_msg_and_die("Invalid process id: '%s'", opt);
1043 if (pid == strace_tracer_pid) {
1044 error_msg_and_die("I'm sorry, I can't let you do that, Dave.");
1055 attach_tcb(struct tcb *const tcp)
1057 if (ptrace_attach_or_seize(tcp->pid) < 0) {
1058 perror_msg("attach: ptrace(%s, %d)",
1059 ptrace_attach_cmd, tcp->pid);
1064 tcp->flags |= TCB_ATTACHED | TCB_STARTUP | post_attach_sigstop;
1067 error_msg("attach to pid %d (main) succeeded", tcp->pid);
1069 char procdir[sizeof("/proc/%d/task") + sizeof(int) * 3];
1071 unsigned int ntid = 0, nerr = 0;
1073 if (followfork && tcp->pid != strace_child &&
1074 sprintf(procdir, "/proc/%d/task", tcp->pid) > 0 &&
1075 (dir = opendir(procdir)) != NULL) {
1078 while ((de = read_dir(dir)) != NULL) {
1079 if (de->d_fileno == 0)
1082 int tid = string_to_uint(de->d_name);
1083 if (tid <= 0 || tid == tcp->pid)
1087 if (ptrace_attach_or_seize(tid) < 0) {
1090 perror_msg("attach: ptrace(%s, %d)",
1091 ptrace_attach_cmd, tid);
1095 error_msg("attach to pid %d succeeded", tid);
1097 struct tcb *tid_tcp = alloctcb(tid);
1098 tid_tcp->flags |= TCB_ATTACHED | TCB_STARTUP |
1099 post_attach_sigstop;
1108 error_msg("Process %u attached"
1110 tcp->pid, ntid - nerr + 1);
1112 error_msg("Process %u attached",
1118 startup_attach(void)
1120 pid_t parent_pid = strace_tracer_pid;
1125 * Block user interruptions as we would leave the traced
1126 * process stopped (process state T) if we would terminate in
1127 * between PTRACE_ATTACH and wait4() on SIGSTOP.
1128 * We rely on cleanup() from this point on.
1131 sigprocmask(SIG_SETMASK, &blocked_set, NULL);
1133 if (daemonized_tracer) {
1136 perror_msg_and_die("fork");
1138 if (pid) { /* parent */
1140 * Wait for grandchild to attach to straced process
1141 * (grandparent). Grandchild SIGKILLs us after it attached.
1142 * Grandparent's wait() is unblocked by our death,
1143 * it proceeds to exec the straced program.
1146 _exit(0); /* paranoia */
1149 /* We will be the tracer process. Remember our new pid: */
1150 strace_tracer_pid = getpid();
1153 for (tcbi = 0; tcbi < tcbtabsize; tcbi++) {
1159 /* Is this a process we should attach to, but not yet attached? */
1160 if (tcp->flags & TCB_ATTACHED)
1161 continue; /* no, we already attached it */
1163 if (tcp->pid == parent_pid || tcp->pid == strace_tracer_pid) {
1165 perror_msg("attach: pid %d", tcp->pid);
1173 sigprocmask(SIG_SETMASK, &start_set, NULL);
1176 sigprocmask(SIG_SETMASK, &blocked_set, NULL);
1178 } /* for each tcbtab[] */
1180 if (daemonized_tracer) {
1182 * Make parent go away.
1183 * Also makes grandparent's wait() unblock.
1185 kill(parent_pid, SIGKILL);
1191 sigprocmask(SIG_SETMASK, &start_set, NULL);
1194 /* Stack-o-phobic exec helper, in the hope to work around
1195 * NOMMU + "daemonized tracer" difficulty.
1197 struct exec_params {
1203 struct sigaction child_sa;
1205 static struct exec_params params_for_tracee;
1207 static void ATTRIBUTE_NOINLINE ATTRIBUTE_NORETURN
1210 struct exec_params *params = ¶ms_for_tracee;
1212 if (params->fd_to_close >= 0)
1213 close(params->fd_to_close);
1214 if (!daemonized_tracer && !use_seize) {
1215 if (ptrace(PTRACE_TRACEME, 0L, 0L, 0L) < 0) {
1216 perror_msg_and_die("ptrace(PTRACE_TRACEME, ...)");
1220 if (username != NULL) {
1222 * It is important to set groups before we
1223 * lose privileges on setuid.
1225 if (initgroups(username, run_gid) < 0) {
1226 perror_msg_and_die("initgroups");
1228 if (setregid(run_gid, params->run_egid) < 0) {
1229 perror_msg_and_die("setregid");
1231 if (setreuid(run_uid, params->run_euid) < 0) {
1232 perror_msg_and_die("setreuid");
1235 else if (geteuid() != 0)
1236 if (setreuid(run_uid, run_uid) < 0) {
1237 perror_msg_and_die("setreuid");
1240 if (!daemonized_tracer) {
1242 * Induce a ptrace stop. Tracer (our parent)
1243 * will resume us with PTRACE_SYSCALL and display
1244 * the immediately following execve syscall.
1245 * Can't do this on NOMMU systems, we are after
1246 * vfork: parent is blocked, stopping would deadlock.
1249 kill(getpid(), SIGSTOP);
1252 /* we depend on SIGCHLD set to SIG_DFL by init code */
1253 /* if it happens to be SIG_IGN'ed, wait won't block */
1258 if (params_for_tracee.child_sa.sa_handler != SIG_DFL)
1259 sigaction(SIGCHLD, ¶ms_for_tracee.child_sa, NULL);
1261 execv(params->pathname, params->argv);
1262 perror_msg_and_die("exec");
1266 * Open a dummy descriptor for use as a placeholder.
1267 * The descriptor is O_RDONLY with FD_CLOEXEC flag set.
1268 * A read attempt from such descriptor ends with EOF,
1269 * a write attempt is rejected with EBADF.
1272 open_dummy_desc(void)
1277 perror_msg_and_die("pipe");
1279 set_cloexec_flag(fds[0]);
1283 /* placeholder fds status for stdin and stdout */
1284 static bool fd_is_placeholder[2];
1287 * Ensure that all standard file descriptors are open by opening placeholder
1288 * file descriptors for those standard file descriptors that are not open.
1290 * The information which descriptors have been made open is saved
1291 * in fd_is_placeholder for later use.
1294 ensure_standard_fds_opened(void)
1298 while ((fd = open_dummy_desc()) <= 2) {
1301 fd_is_placeholder[fd] = true;
1309 * Redirect stdin and stdout unless they have been opened earlier
1310 * by ensure_standard_fds_opened as placeholders.
1313 redirect_standard_fds(void)
1318 * It might be a good idea to redirect stderr as well,
1319 * but we sometimes need to print error messages.
1321 for (i = 0; i <= 1; ++i) {
1322 if (!fd_is_placeholder[i]) {
1330 startup_child(char **argv)
1332 struct_stat statbuf;
1333 const char *filename;
1334 size_t filename_len;
1335 char pathname[PATH_MAX];
1340 filename_len = strlen(filename);
1342 if (filename_len > sizeof(pathname) - 1) {
1343 errno = ENAMETOOLONG;
1344 perror_msg_and_die("exec");
1346 if (strchr(filename, '/')) {
1347 strcpy(pathname, filename);
1349 #ifdef USE_DEBUGGING_EXEC
1351 * Debuggers customarily check the current directory
1352 * first regardless of the path but doing that gives
1353 * security geeks a panic attack.
1355 else if (stat_file(filename, &statbuf) == 0)
1356 strcpy(pathname, filename);
1357 #endif /* USE_DEBUGGING_EXEC */
1362 for (path = getenv("PATH"); path && *path; path += m) {
1363 const char *colon = strchr(path, ':');
1369 m = n = strlen(path);
1371 if (!getcwd(pathname, PATH_MAX))
1373 len = strlen(pathname);
1375 else if (n > sizeof(pathname) - 1)
1378 strncpy(pathname, path, n);
1381 if (len && pathname[len - 1] != '/')
1382 pathname[len++] = '/';
1383 if (filename_len + len > sizeof(pathname) - 1)
1385 strcpy(pathname + len, filename);
1386 if (stat_file(pathname, &statbuf) == 0 &&
1387 /* Accept only regular files
1388 with some execute bits set.
1389 XXX not perfect, might still fail */
1390 S_ISREG(statbuf.st_mode) &&
1391 (statbuf.st_mode & 0111))
1394 if (!path || !*path)
1397 if (stat_file(pathname, &statbuf) < 0) {
1398 perror_msg_and_die("Can't stat '%s'", filename);
1401 params_for_tracee.fd_to_close = (shared_log != stderr) ? fileno(shared_log) : -1;
1402 params_for_tracee.run_euid = (statbuf.st_mode & S_ISUID) ? statbuf.st_uid : run_uid;
1403 params_for_tracee.run_egid = (statbuf.st_mode & S_ISGID) ? statbuf.st_gid : run_gid;
1404 params_for_tracee.argv = argv;
1406 * On NOMMU, can be safely freed only after execve in tracee.
1407 * It's hard to know when that happens, so we just leak it.
1409 params_for_tracee.pathname = NOMMU_SYSTEM ? xstrdup(pathname) : pathname;
1411 #if defined HAVE_PRCTL && defined PR_SET_PTRACER && defined PR_SET_PTRACER_ANY
1412 if (daemonized_tracer)
1413 prctl(PR_SET_PTRACER, PR_SET_PTRACER_ANY);
1418 perror_msg_and_die("fork");
1420 if ((pid != 0 && daemonized_tracer)
1421 || (pid == 0 && !daemonized_tracer)
1423 /* We are to become the tracee. Two cases:
1425 * not -D: we are child
1430 /* We are the tracer */
1432 if (!daemonized_tracer) {
1435 /* child did PTRACE_TRACEME, nothing to do in parent */
1437 if (!NOMMU_SYSTEM) {
1438 /* Wait until child stopped itself */
1440 while (waitpid(pid, &status, WSTOPPED) < 0) {
1443 perror_msg_and_die("waitpid");
1445 if (!WIFSTOPPED(status) || WSTOPSIG(status) != SIGSTOP) {
1446 kill_save_errno(pid, SIGKILL);
1447 perror_msg_and_die("Unexpected wait status %#x",
1451 /* Else: NOMMU case, we have no way to sync.
1452 * Just attach to it as soon as possible.
1453 * This means that we may miss a few first syscalls...
1456 if (ptrace_attach_or_seize(pid)) {
1457 kill_save_errno(pid, SIGKILL);
1458 perror_msg_and_die("attach: ptrace(%s, %d)",
1459 ptrace_attach_cmd, pid);
1464 tcp = alloctcb(pid);
1465 tcp->flags |= TCB_ATTACHED | TCB_STARTUP
1466 | TCB_SKIP_DETACH_ON_FIRST_EXEC
1467 | (NOMMU_SYSTEM ? 0 : (TCB_HIDE_LOG | post_attach_sigstop));
1471 /* With -D, we are *child* here, the tracee is our parent. */
1472 strace_child = strace_tracer_pid;
1473 strace_tracer_pid = getpid();
1474 tcp = alloctcb(strace_child);
1475 tcp->flags |= TCB_SKIP_DETACH_ON_FIRST_EXEC | TCB_HIDE_LOG;
1476 /* attaching will be done later, by startup_attach */
1477 /* note: we don't do newoutf(tcp) here either! */
1479 /* NOMMU BUG! -D mode is active, we (child) return,
1480 * and we will scribble over parent's stack!
1481 * When parent later unpauses, it segfaults.
1484 * (1) by declaring exec_or_die() NORETURN,
1485 * hopefully compiler will just jump to it
1486 * instead of call (won't push anything to stack),
1487 * (2) by trying very hard in exec_or_die()
1488 * to not use any stack,
1489 * (3) having a really big (PATH_MAX) stack object
1490 * in this function, which creates a "buffer" between
1491 * child's and parent's stack pointers.
1492 * This may save us if (1) and (2) failed
1493 * and compiler decided to use stack in exec_or_die() anyway
1494 * (happens on i386 because of stack parameter passing).
1496 * A cleaner solution is to use makecontext + setcontext
1497 * to create a genuine separate stack and execute on it.
1501 * A case where straced process is part of a pipe:
1502 * { sleep 1; yes | head -n99999; } | strace -o/dev/null sh -c 'exec <&-; sleep 9'
1503 * If strace won't close its fd#0, closing it in tracee is not enough:
1504 * the pipe is still open, it has a reader. Thus, "head" will not get its
1505 * SIGPIPE at once, on the first write.
1507 * Preventing it by redirecting strace's stdin/out.
1508 * (Don't leave fds 0 and 1 closed, this is bad practice: future opens
1509 * will reuse them, unexpectedly making a newly opened object "stdin").
1511 redirect_standard_fds();
1516 test_ptrace_seize(void)
1520 /* Need fork for test. NOMMU has no forks */
1522 post_attach_sigstop = 0; /* this sets use_seize to 1 */
1528 perror_msg_and_die("fork");
1535 /* PTRACE_SEIZE, unlike ATTACH, doesn't force tracee to trap. After
1536 * attaching tracee continues to run unless a trap condition occurs.
1537 * PTRACE_SEIZE doesn't affect signal or group stop state.
1539 if (ptrace(PTRACE_SEIZE, pid, 0, 0) == 0) {
1540 post_attach_sigstop = 0; /* this sets use_seize to 1 */
1541 } else if (debug_flag) {
1542 error_msg("PTRACE_SEIZE doesn't work");
1548 int status, tracee_pid;
1551 tracee_pid = waitpid(pid, &status, 0);
1552 if (tracee_pid <= 0) {
1555 perror_msg_and_die("%s: unexpected wait result %d",
1556 __func__, tracee_pid);
1558 if (WIFSIGNALED(status)) {
1561 error_msg_and_die("%s: unexpected wait status %#x",
1565 #else /* !USE_SEIZE */
1566 # define test_ptrace_seize() ((void)0)
1570 get_os_release(void)
1576 perror_msg_and_die("uname");
1577 /* u.release has this form: "3.2.9[-some-garbage]" */
1581 if (!(*p >= '0' && *p <= '9'))
1582 error_msg_and_die("Bad OS release string: '%s'", u.release);
1583 /* Note: this open-codes KERNEL_VERSION(): */
1584 rel = (rel << 8) | atoi(p);
1585 if (rel >= KERNEL_VERSION(1,0,0))
1587 while (*p >= '0' && *p <= '9')
1590 if (rel >= KERNEL_VERSION(0,1,0)) {
1591 /* "X.Y-something" means "X.Y.0" */
1595 error_msg_and_die("Bad OS release string: '%s'", u.release);
1603 set_sigaction(int signo, void (*sighandler)(int), struct sigaction *oldact)
1605 /* if signal handler is a function, add the signal to blocked_set */
1606 if (sighandler != SIG_IGN && sighandler != SIG_DFL)
1607 sigaddset(&blocked_set, signo);
1609 const struct sigaction sa = { .sa_handler = sighandler };
1610 sigaction(signo, &sa, oldact);
1614 * Initialization part of main() was eating much stack (~0.5k),
1615 * which was unused after init.
1616 * We can reuse it if we move init code into a separate function.
1618 * Don't want main() to inline us and defeat the reason
1619 * we have a separate function.
1621 static void ATTRIBUTE_NOINLINE
1622 init(int argc, char *argv[])
1627 progname = argv[0] ? argv[0] : "strace";
1629 strace_tracer_pid = getpid();
1631 os_release = get_os_release();
1633 shared_log = stderr;
1634 set_sortby(DEFAULT_SORTBY);
1635 set_personality(DEFAULT_PERSONALITY);
1636 qualify("trace=all");
1637 qualify("abbrev=all");
1638 qualify("verbose=all");
1639 #if DEFAULT_QUAL_FLAGS != (QUAL_TRACE | QUAL_ABBREV | QUAL_VERBOSE)
1640 # error Bug in DEFAULT_QUAL_FLAGS
1642 qualify("signal=all");
1643 while ((c = getopt(argc, argv,
1644 "+b:cCdfFhiqrtTvVwxyz"
1645 #ifdef USE_LIBUNWIND
1649 "a:e:o:O:p:s:S:u:E:P:I:")) != EOF) {
1652 if (strcmp(optarg, "execve") != 0)
1653 error_msg_and_die("Syscall '%s' for -b isn't supported",
1655 detach_on_execve = 1;
1658 if (cflag == CFLAG_BOTH) {
1659 error_msg_and_help("-c and -C are mutually exclusive");
1661 cflag = CFLAG_ONLY_STATS;
1664 if (cflag == CFLAG_ONLY_STATS) {
1665 error_msg_and_help("-c and -C are mutually exclusive");
1673 daemonized_tracer = 1;
1700 count_wallclock = 1;
1709 qualify("abbrev=none");
1716 not_failing_only = 1;
1719 acolumn = string_to_uint(optarg);
1721 error_opt_arg(c, optarg);
1727 outfname = xstrdup(optarg);
1730 i = string_to_uint(optarg);
1732 error_opt_arg(c, optarg);
1736 process_opt_p_list(optarg);
1739 pathtrace_select(optarg);
1742 i = string_to_uint(optarg);
1744 error_opt_arg(c, optarg);
1751 username = xstrdup(optarg);
1753 #ifdef USE_LIBUNWIND
1755 stack_trace_enabled = true;
1759 if (putenv(optarg) < 0)
1760 die_out_of_memory();
1763 opt_intr = string_to_uint_upto(optarg, NUM_INTR_OPTS - 1);
1765 error_opt_arg(c, optarg);
1768 error_msg_and_help(NULL);
1773 /* argc -= optind; - no need, argc is not used below */
1775 acolumn_spaces = xmalloc(acolumn + 1);
1776 memset(acolumn_spaces, ' ', acolumn);
1777 acolumn_spaces[acolumn] = '\0';
1779 if (!argv[0] && !nprocs) {
1780 error_msg_and_help("must have PROG [ARGS] or -p PID");
1783 if (!argv[0] && daemonized_tracer) {
1784 error_msg_and_help("PROG [ARGS] must be specified with -D");
1790 if (followfork >= 2 && cflag) {
1791 error_msg_and_help("(-c or -C) and -ff are mutually exclusive");
1794 if (count_wallclock && !cflag) {
1795 error_msg_and_help("-w must be given with (-c or -C)");
1798 if (cflag == CFLAG_ONLY_STATS) {
1800 error_msg("-%c has no effect with -c", 'i');
1801 #ifdef USE_LIBUNWIND
1802 if (stack_trace_enabled)
1803 error_msg("-%c has no effect with -c", 'k');
1806 error_msg("-%c has no effect with -c", 'r');
1808 error_msg("-%c has no effect with -c", 't');
1810 error_msg("-%c has no effect with -c", 'T');
1812 error_msg("-%c has no effect with -c", 'y');
1817 error_msg("-tt has no effect with -r");
1821 sigprocmask(SIG_SETMASK, NULL, &start_set);
1822 memcpy(&blocked_set, &start_set, sizeof(blocked_set));
1824 set_sigaction(SIGCHLD, SIG_DFL, ¶ms_for_tracee.child_sa);
1826 #ifdef USE_LIBUNWIND
1827 if (stack_trace_enabled) {
1831 for (tcbi = 0; tcbi < tcbtabsize; ++tcbi) {
1832 unwind_tcb_init(tcbtab[tcbi]);
1837 /* See if they want to run as another user. */
1838 if (username != NULL) {
1839 struct passwd *pent;
1841 if (getuid() != 0 || geteuid() != 0) {
1842 error_msg_and_die("You must be root to use the -u option");
1844 pent = getpwnam(username);
1846 error_msg_and_die("Cannot find user '%s'", username);
1848 run_uid = pent->pw_uid;
1849 run_gid = pent->pw_gid;
1857 ptrace_setoptions |= PTRACE_O_TRACECLONE |
1858 PTRACE_O_TRACEFORK |
1859 PTRACE_O_TRACEVFORK;
1861 error_msg("ptrace_setoptions = %#x", ptrace_setoptions);
1862 test_ptrace_seize();
1865 * Is something weird with our stdin and/or stdout -
1866 * for example, may they be not open? In this case,
1867 * ensure that none of the future opens uses them.
1869 * This was seen in the wild when /proc/sys/kernel/core_pattern
1870 * was set to "|/bin/strace -o/tmp/LOG PROG":
1871 * kernel runs coredump helper with fd#0 open but fd#1 closed (!),
1872 * therefore LOG gets opened to fd#1, and fd#1 is closed by
1873 * "don't hold up stdin/out open" code soon after.
1875 ensure_standard_fds_opened();
1877 /* Check if they want to redirect the output. */
1879 /* See if they want to pipe the output. */
1880 if (outfname[0] == '|' || outfname[0] == '!') {
1882 * We can't do the <outfname>.PID funny business
1883 * when using popen, so prohibit it.
1885 if (followfork >= 2)
1886 error_msg_and_help("piping the output and -ff are mutually exclusive");
1887 shared_log = strace_popen(outfname + 1);
1889 else if (followfork < 2)
1890 shared_log = strace_fopen(outfname);
1892 /* -ff without -o FILE is the same as single -f */
1893 if (followfork >= 2)
1897 if (!outfname || outfname[0] == '|' || outfname[0] == '!') {
1898 setvbuf(shared_log, NULL, _IOLBF, 0);
1902 * argv[0] -pPID -oFILE Default interactive setting
1903 * yes * 0 INTR_WHILE_WAIT
1904 * no 1 0 INTR_WHILE_WAIT
1905 * yes * 1 INTR_NEVER
1906 * no 1 1 INTR_WHILE_WAIT
1909 if (outfname && argv[0]) {
1911 opt_intr = INTR_NEVER;
1916 opt_intr = INTR_WHILE_WAIT;
1919 * startup_child() must be called before the signal handlers get
1920 * installed below as they are inherited into the spawned process.
1921 * Also we do not need to be protected by them as during interruption
1922 * in the startup_child() mode we kill the spawned process anyway.
1925 startup_child(argv);
1928 set_sigaction(SIGTTOU, SIG_IGN, NULL);
1929 set_sigaction(SIGTTIN, SIG_IGN, NULL);
1930 if (opt_intr != INTR_ANYWHERE) {
1931 if (opt_intr == INTR_BLOCK_TSTP_TOO)
1932 set_sigaction(SIGTSTP, SIG_IGN, NULL);
1934 * In interactive mode (if no -o OUTFILE, or -p PID is used),
1935 * fatal signals are blocked while syscall stop is processed,
1936 * and acted on in between, when waiting for new syscall stops.
1937 * In non-interactive mode, signals are ignored.
1939 set_sigaction(SIGHUP, interactive ? interrupt : SIG_IGN, NULL);
1940 set_sigaction(SIGINT, interactive ? interrupt : SIG_IGN, NULL);
1941 set_sigaction(SIGQUIT, interactive ? interrupt : SIG_IGN, NULL);
1942 set_sigaction(SIGPIPE, interactive ? interrupt : SIG_IGN, NULL);
1943 set_sigaction(SIGTERM, interactive ? interrupt : SIG_IGN, NULL);
1946 if (nprocs != 0 || daemonized_tracer)
1949 /* Do we want pids printed in our -o OUTFILE?
1950 * -ff: no (every pid has its own file); or
1951 * -f: yes (there can be more pids in the future); or
1952 * -p PID1,PID2: yes (there are already more than one pid)
1954 print_pid_pfx = (outfname && followfork < 2 && (followfork == 1 || nprocs > 1));
1965 for (i = 0; i < tcbtabsize; i++) {
1966 struct tcb *tcp = tcbtab[i];
1967 if (tcp->pid == pid)
1981 /* 'interrupted' is a volatile object, fetch it only once */
1982 fatal_sig = interrupted;
1984 fatal_sig = SIGTERM;
1986 for (i = 0; i < tcbtabsize; i++) {
1991 error_msg("cleanup: looking at pid %u", tcp->pid);
1992 if (tcp->pid == strace_child) {
1993 kill(tcp->pid, SIGCONT);
1994 kill(tcp->pid, fatal_sig);
1999 call_summary(shared_log);
2009 print_debug_info(const int pid, int status)
2011 const unsigned int event = (unsigned int) status >> 16;
2012 char buf[sizeof("WIFEXITED,exitcode=%u") + sizeof(int)*3 /*paranoia:*/ + 16];
2013 char evbuf[sizeof(",EVENT_VFORK_DONE (%u)") + sizeof(int)*3 /*paranoia:*/ + 16];
2016 if (WIFSIGNALED(status))
2018 sprintf(buf, "WIFSIGNALED,%ssig=%s",
2019 WCOREDUMP(status) ? "core," : "",
2020 signame(WTERMSIG(status)));
2022 sprintf(buf, "WIFSIGNALED,sig=%s",
2023 signame(WTERMSIG(status)));
2025 if (WIFEXITED(status))
2026 sprintf(buf, "WIFEXITED,exitcode=%u", WEXITSTATUS(status));
2027 if (WIFSTOPPED(status))
2028 sprintf(buf, "WIFSTOPPED,sig=%s", signame(WSTOPSIG(status)));
2031 static const char *const event_names[] = {
2032 [PTRACE_EVENT_CLONE] = "CLONE",
2033 [PTRACE_EVENT_FORK] = "FORK",
2034 [PTRACE_EVENT_VFORK] = "VFORK",
2035 [PTRACE_EVENT_VFORK_DONE] = "VFORK_DONE",
2036 [PTRACE_EVENT_EXEC] = "EXEC",
2037 [PTRACE_EVENT_EXIT] = "EXIT",
2038 /* [PTRACE_EVENT_STOP (=128)] would make biggish array */
2040 const char *e = "??";
2041 if (event < ARRAY_SIZE(event_names))
2042 e = event_names[event];
2043 else if (event == PTRACE_EVENT_STOP)
2045 sprintf(evbuf, ",EVENT_%s (%u)", e, event);
2047 error_msg("[wait(0x%06x) = %u] %s%s", status, pid, buf, evbuf);
2051 maybe_allocate_tcb(const int pid, int status)
2053 if (!WIFSTOPPED(status)) {
2054 if (detach_on_execve && pid == strace_child) {
2055 /* example: strace -bexecve sh -c 'exec true' */
2060 * This can happen if we inherited an unknown child.
2061 * Example: (sleep 1 & exec strace true)
2063 error_msg("Exit of unknown pid %u ignored", pid);
2067 /* We assume it's a fork/vfork/clone child */
2068 struct tcb *tcp = alloctcb(pid);
2069 tcp->flags |= TCB_ATTACHED | TCB_STARTUP | post_attach_sigstop;
2072 error_msg("Process %d attached", pid);
2075 /* This can happen if a clone call used
2076 * CLONE_PTRACE itself.
2078 ptrace(PTRACE_CONT, pid, NULL, 0);
2079 error_msg("Stop of unknown pid %u seen, PTRACE_CONTed it", pid);
2085 maybe_switch_tcbs(struct tcb *tcp, const int pid)
2088 struct tcb *execve_thread;
2091 if (ptrace(PTRACE_GETEVENTMSG, pid, NULL, &old_pid) < 0)
2093 /* Avoid truncation in pid2tcb() param passing */
2094 if (old_pid <= 0 || old_pid == pid)
2096 if ((unsigned long) old_pid > UINT_MAX)
2098 execve_thread = pid2tcb(old_pid);
2099 /* It should be !NULL, but I feel paranoid */
2103 if (execve_thread->curcol != 0) {
2105 * One case we are here is -ff:
2106 * try "strace -oLOG -ff test/threaded_execve"
2108 fprintf(execve_thread->outf, " <pid changed to %d ...>\n", pid);
2109 /*execve_thread->curcol = 0; - no need, see code below */
2111 /* Swap output FILEs (needed for -ff) */
2112 fp = execve_thread->outf;
2113 execve_thread->outf = tcp->outf;
2115 /* And their column positions */
2116 execve_thread->curcol = tcp->curcol;
2118 /* Drop leader, but close execve'd thread outfile (if -ff) */
2120 /* Switch to the thread, reusing leader's outfile and pid */
2121 tcp = execve_thread;
2123 if (cflag != CFLAG_ONLY_STATS) {
2125 tprintf("+++ superseded by execve in pid %lu +++\n", old_pid);
2127 tcp->flags |= TCB_REPRINT;
2134 print_signalled(struct tcb *tcp, const int pid, int status)
2136 if (pid == strace_child) {
2137 exit_code = 0x100 | WTERMSIG(status);
2141 if (cflag != CFLAG_ONLY_STATS
2142 && is_number_in_set(WTERMSIG(status), &signal_set)) {
2145 tprintf("+++ killed by %s %s+++\n",
2146 signame(WTERMSIG(status)),
2147 WCOREDUMP(status) ? "(core dumped) " : "");
2149 tprintf("+++ killed by %s +++\n",
2150 signame(WTERMSIG(status)));
2157 print_exited(struct tcb *tcp, const int pid, int status)
2159 if (pid == strace_child) {
2160 exit_code = WEXITSTATUS(status);
2164 if (cflag != CFLAG_ONLY_STATS &&
2167 tprintf("+++ exited with %d +++\n", WEXITSTATUS(status));
2173 print_stopped(struct tcb *tcp, const siginfo_t *si, const unsigned int sig)
2175 if (cflag != CFLAG_ONLY_STATS
2177 && is_number_in_set(sig, &signal_set)) {
2180 tprintf("--- %s ", signame(sig));
2184 tprintf("--- stopped by %s ---\n", signame(sig));
2190 startup_tcb(struct tcb *tcp)
2193 error_msg("pid %d has TCB_STARTUP, initializing it", tcp->pid);
2195 tcp->flags &= ~TCB_STARTUP;
2199 error_msg("setting opts 0x%x on pid %d",
2200 ptrace_setoptions, tcp->pid);
2201 if (ptrace(PTRACE_SETOPTIONS, tcp->pid, NULL, ptrace_setoptions) < 0) {
2202 if (errno != ESRCH) {
2203 /* Should never happen, really */
2204 perror_msg_and_die("PTRACE_SETOPTIONS");
2209 if (get_scno(tcp) == 1)
2210 tcp->s_prev_ent = tcp->s_ent;
2214 print_event_exit(struct tcb *tcp)
2216 if (entering(tcp) || filtered(tcp) || hide_log(tcp)
2217 || cflag == CFLAG_ONLY_STATS) {
2221 if (followfork < 2 && printing_tcp && printing_tcp != tcp
2222 && printing_tcp->curcol != 0) {
2223 current_tcp = printing_tcp;
2224 tprints(" <unfinished ...>\n");
2225 fflush(printing_tcp->outf);
2226 printing_tcp->curcol = 0;
2230 if ((followfork < 2 && printing_tcp != tcp)
2231 || (tcp->flags & TCB_REPRINT)) {
2232 tcp->flags &= ~TCB_REPRINT;
2234 tprintf("<... %s resumed>", tcp->s_ent->sys_name);
2237 if (!(tcp->sys_func_rval & RVAL_DECODED)) {
2239 * The decoder has probably decided to print something
2240 * on exiting syscall which is not going to happen.
2242 tprints(" <unfinished ...>");
2251 /* Break the main loop. */
2254 /* Call next_event() again. */
2257 /* Restart the tracee with signal 0 and call next_event() again. */
2261 * For all the events below, current_tcp is set to current tracee's
2262 * tcb. All the suggested actions imply that you want to continue
2263 * tracing of the current tracee; alternatively, you can detach it.
2267 * Syscall entry or exit.
2268 * Restart the tracee with signal 0, or with an injected signal number.
2273 * Tracee received signal with number WSTOPSIG(*pstatus); signal info
2274 * is written to *si. Restart the tracee (with that signal number
2275 * if you want to deliver it).
2277 TE_SIGNAL_DELIVERY_STOP,
2280 * Tracee was killed by a signal with number WTERMSIG(*pstatus).
2285 * Tracee was stopped by a signal with number WSTOPSIG(*pstatus).
2286 * Restart the tracee with that signal number.
2291 * Tracee exited with status WEXITSTATUS(*pstatus).
2296 * Tracee is going to perform execve().
2297 * Restart the tracee with signal 0.
2299 TE_STOP_BEFORE_EXECVE,
2302 * Tracee is going to terminate.
2303 * Restart the tracee with signal 0.
2305 TE_STOP_BEFORE_EXIT,
2308 static enum trace_event
2309 next_event(int *pstatus, siginfo_t *si)
2321 * Used to exit simply when nprocs hits zero, but in this testcase:
2322 * int main() { _exit(!!fork()); }
2323 * under strace -f, parent sometimes (rarely) manages
2324 * to exit before we see the first stop of the child,
2325 * and we are losing track of it:
2326 * 19923 clone(...) = 19924
2327 * 19923 exit_group(1) = ?
2328 * 19923 +++ exited with 1 +++
2329 * Exiting only when wait() returns ECHILD works better.
2331 if (popen_pid != 0) {
2332 /* However, if -o|logger is in use, we can't do that.
2333 * Can work around that by double-forking the logger,
2334 * but that loses the ability to wait for its completion
2335 * on exit. Oh well...
2342 sigprocmask(SIG_SETMASK, &start_set, NULL);
2343 pid = wait4(-1, pstatus, __WALL, (cflag ? &ru : NULL));
2346 sigprocmask(SIG_SETMASK, &blocked_set, NULL);
2349 if (wait_errno == EINTR)
2351 if (nprocs == 0 && wait_errno == ECHILD)
2354 * If nprocs > 0, ECHILD is not expected,
2355 * treat it as any other error here:
2358 perror_msg_and_die("wait4(__WALL)");
2363 if (pid == popen_pid) {
2364 if (!WIFSTOPPED(status))
2370 print_debug_info(pid, status);
2372 /* Look up 'pid' in our table. */
2376 tcp = maybe_allocate_tcb(pid, status);
2383 /* Set current output file */
2387 tv_sub(&tcp->dtime, &ru.ru_stime, &tcp->stime);
2388 tcp->stime = ru.ru_stime;
2391 if (WIFSIGNALED(status))
2392 return TE_SIGNALLED;
2394 if (WIFEXITED(status))
2398 * As WCONTINUED flag has not been specified to wait4,
2399 * it cannot be WIFCONTINUED(status), so the only case
2400 * that remains is WIFSTOPPED(status).
2403 /* Is this the very first time we see this tracee stopped? */
2404 if (tcp->flags & TCB_STARTUP)
2407 const unsigned int sig = WSTOPSIG(status);
2408 const unsigned int event = (unsigned int) status >> 16;
2413 * Is this post-attach SIGSTOP?
2414 * Interestingly, the process may stop
2415 * with STOPSIG equal to some other signal
2416 * than SIGSTOP if we happened to attach
2417 * just before the process takes a signal.
2419 if (sig == SIGSTOP && (tcp->flags & TCB_IGNORE_ONE_SIGSTOP)) {
2421 error_msg("ignored SIGSTOP on pid %d", tcp->pid);
2422 tcp->flags &= ~TCB_IGNORE_ONE_SIGSTOP;
2424 } else if (sig == syscall_trap_sig) {
2425 return TE_SYSCALL_STOP;
2427 *si = (siginfo_t) {};
2429 * True if tracee is stopped by signal
2430 * (as opposed to "tracee received signal").
2431 * TODO: shouldn't we check for errno == EINVAL too?
2432 * We can get ESRCH instead, you know...
2434 bool stopped = ptrace(PTRACE_GETSIGINFO, pid, 0, si) < 0;
2435 return stopped ? TE_GROUP_STOP : TE_SIGNAL_DELIVERY_STOP;
2439 case PTRACE_EVENT_STOP:
2441 * PTRACE_INTERRUPT-stop or group-stop.
2442 * PTRACE_INTERRUPT-stop has sig == SIGTRAP here.
2449 return TE_GROUP_STOP;
2453 case PTRACE_EVENT_EXEC:
2454 return TE_STOP_BEFORE_EXECVE;
2455 case PTRACE_EVENT_EXIT:
2456 return TE_STOP_BEFORE_EXIT;
2463 trace_syscall(struct tcb *tcp, unsigned int *sig)
2465 if (entering(tcp)) {
2466 int res = syscall_entering_decode(tcp);
2471 res = syscall_entering_trace(tcp, sig);
2473 syscall_entering_finish(tcp, res);
2476 struct timeval tv = {};
2477 int res = syscall_exiting_decode(tcp, &tv);
2479 res = syscall_exiting_trace(tcp, tv, res);
2481 syscall_exiting_finish(tcp);
2486 /* Returns true iff the main trace loop has to continue. */
2488 dispatch_event(enum trace_event ret, int *pstatus, siginfo_t *si)
2490 unsigned int restart_op = PTRACE_SYSCALL;
2491 unsigned int restart_sig = 0;
2503 case TE_SYSCALL_STOP:
2504 if (trace_syscall(current_tcp, &restart_sig) < 0) {
2506 * ptrace() failed in trace_syscall().
2507 * Likely a result of process disappearing mid-flight.
2508 * Observed case: exit_group() or SIGKILL terminating
2509 * all processes in thread group.
2510 * We assume that ptrace error was caused by process death.
2511 * We used to detach(current_tcp) here, but since we no
2512 * longer implement "detach before death" policy/hack,
2513 * we can let this process to report its death to us
2514 * normally, via WIFEXITED or WIFSIGNALED wait status.
2520 case TE_SIGNAL_DELIVERY_STOP:
2521 restart_sig = WSTOPSIG(*pstatus);
2522 print_stopped(current_tcp, si, restart_sig);
2526 print_signalled(current_tcp, current_tcp->pid, *pstatus);
2527 droptcb(current_tcp);
2531 restart_sig = WSTOPSIG(*pstatus);
2532 print_stopped(current_tcp, NULL, restart_sig);
2535 * This ends ptrace-stop, but does *not* end group-stop.
2536 * This makes stopping signals work properly on straced
2537 * process (that is, process really stops. It used to
2540 restart_op = PTRACE_LISTEN;
2546 print_exited(current_tcp, current_tcp->pid, *pstatus);
2547 droptcb(current_tcp);
2550 case TE_STOP_BEFORE_EXECVE:
2552 * Under Linux, execve changes pid to thread leader's pid,
2553 * and we see this changed pid on EVENT_EXEC and later,
2554 * execve sysexit. Leader "disappears" without exit
2555 * notification. Let user know that, drop leader's tcb,
2556 * and fix up pid in execve thread's tcb.
2557 * Effectively, execve thread's tcb replaces leader's tcb.
2559 * BTW, leader is 'stuck undead' (doesn't report WIFEXITED
2560 * on exit syscall) in multithreaded programs exactly
2561 * in order to handle this case.
2563 * PTRACE_GETEVENTMSG returns old pid starting from Linux 3.0.
2564 * On 2.6 and earlier, it can return garbage.
2566 if (os_release >= KERNEL_VERSION(3,0,0))
2567 current_tcp = maybe_switch_tcbs(current_tcp, current_tcp->pid);
2569 if (detach_on_execve) {
2570 if (current_tcp->flags & TCB_SKIP_DETACH_ON_FIRST_EXEC) {
2571 current_tcp->flags &= ~TCB_SKIP_DETACH_ON_FIRST_EXEC;
2573 detach(current_tcp); /* do "-b execve" thingy */
2579 case TE_STOP_BEFORE_EXIT:
2580 print_event_exit(current_tcp);
2584 /* We handled quick cases, we are permitted to interrupt now. */
2588 if (ptrace_restart(restart_op, current_tcp, restart_sig) < 0) {
2589 /* Note: ptrace_restart emitted error message */
2596 #ifdef ENABLE_COVERAGE_GCOV
2597 extern void __gcov_flush();
2600 static void ATTRIBUTE_NORETURN
2605 if (shared_log != stderr)
2608 while (waitpid(popen_pid, NULL, 0) < 0 && errno == EINTR)
2611 if (exit_code > 0xff) {
2612 /* Avoid potential core file clobbering. */
2613 struct_rlimit rlim = {0, 0};
2614 set_rlimit(RLIMIT_CORE, &rlim);
2616 /* Child was killed by a signal, mimic that. */
2618 signal(exit_code, SIG_DFL);
2619 #ifdef ENABLE_COVERAGE_GCOV
2624 /* Unblock the signal. */
2627 sigaddset(&mask, exit_code);
2628 #ifdef ENABLE_COVERAGE_GCOV
2631 sigprocmask(SIG_UNBLOCK, &mask, NULL);
2633 /* Paranoia - what if this signal is not fatal?
2634 Exit with 128 + signo then. */
2641 main(int argc, char *argv[])
2645 exit_code = !nprocs;
2649 while (dispatch_event(next_event(&status, &si), &status, &si))