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) 1999-2018 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.
38 #include <sys/resource.h>
47 #include <sys/utsname.h>
49 # include <sys/prctl.h>
51 #include <asm/unistd.h>
53 #include "largefile_wrappers.h"
54 #include "mmap_cache.h"
55 #include "number_set.h"
57 #include "printsiginfo.h"
58 #include "trace_event.h"
61 /* In some libc, these aren't declared. Do it ourself: */
62 extern char **environ;
67 /* if this is true do the stack trace for every system call */
68 bool stack_trace_enabled;
71 #define my_tkill(tid, sig) syscall(__NR_tkill, (tid), (sig))
73 /* Glue for systems without a MMU that cannot provide fork() */
74 #if !defined(HAVE_FORK)
76 # define NOMMU_SYSTEM 1
79 # define fork() vfork()
82 const unsigned int syscall_trap_sig = SIGTRAP | 0x80;
84 cflag_t cflag = CFLAG_NONE;
85 unsigned int followfork;
86 unsigned int ptrace_setoptions = PTRACE_O_TRACESYSGOOD | PTRACE_O_TRACEEXEC
94 static unsigned int tflag;
96 static bool print_pid_pfx;
101 INTR_ANYWHERE = 1, /* don't block/ignore any signals */
102 INTR_WHILE_WAIT = 2, /* block fatal signals while decoding syscall. default */
103 INTR_NEVER = 3, /* block fatal signals. default if '-o FILE PROG' */
104 INTR_BLOCK_TSTP_TOO = 4, /* block fatal signals and SIGTSTP (^Z) */
108 /* We play with signal mask only if this mode is active: */
109 #define interactive (opt_intr == INTR_WHILE_WAIT)
112 * daemonized_tracer supports -D option.
113 * With this option, strace forks twice.
114 * Unlike normal case, with -D *grandparent* process exec's,
115 * becoming a traced process. Child exits (this prevents traced process
116 * from having children it doesn't expect to have), and grandchild
117 * attaches to grandparent similarly to strace -p PID.
118 * This allows for more transparent interaction in cases
119 * when process and its parent are communicating via signals,
120 * wait() etc. Without -D, strace process gets lodged in between,
121 * disrupting parent<->child link.
123 static bool daemonized_tracer;
126 static int post_attach_sigstop = TCB_IGNORE_ONE_SIGSTOP;
127 # define use_seize (post_attach_sigstop == 0)
129 # define post_attach_sigstop TCB_IGNORE_ONE_SIGSTOP
133 /* Sometimes we want to print only succeeding syscalls. */
134 bool not_failing_only;
136 /* Show path associated with fd arguments */
137 unsigned int show_fd_path;
139 static bool detach_on_execve;
141 static int exit_code;
142 static int strace_child;
143 static int strace_tracer_pid;
145 static const char *username;
146 static uid_t run_uid;
147 static gid_t run_gid;
149 unsigned int max_strlen = DEFAULT_STRLEN;
150 static int acolumn = DEFAULT_ACOLUMN;
151 static char *acolumn_spaces;
153 static const char *outfname;
154 /* If -ff, points to stderr. Else, it's our common output log */
155 static FILE *shared_log;
157 struct tcb *printing_tcp;
158 static struct tcb *current_tcp;
160 static struct tcb **tcbtab;
161 static unsigned int nprocs;
162 static size_t tcbtabsize;
164 #ifndef HAVE_PROGRAM_INVOCATION_NAME
165 char *program_invocation_name;
168 unsigned os_release; /* generated from uname()'s u.release */
170 static void detach(struct tcb *tcp);
171 static void cleanup(void);
172 static void interrupt(int sig);
173 static sigset_t start_set, blocked_set;
175 #ifdef HAVE_SIG_ATOMIC_T
176 static volatile sig_atomic_t interrupted;
178 static volatile int interrupted;
181 #ifndef HAVE_STRERROR
183 #if !HAVE_DECL_SYS_ERRLIST
185 extern char *sys_errlist[];
191 static char buf[sizeof("Unknown error %d") + sizeof(int)*3];
193 if (err_no < 1 || err_no >= sys_nerr) {
194 xsprintf(buf, "Unknown error %d", err_no);
197 return sys_errlist[err_no];
200 #endif /* HAVE_STERRROR */
205 static const char features[] =
208 #endif /* USE_LIBUNWIND */
211 #endif /* USE_DEMANGLE */
212 #if SUPPORTED_PERSONALITIES > 1
213 # if defined HAVE_M32_MPERS
218 #endif /* SUPPORTED_PERSONALITIES > 1 */
219 #if SUPPORTED_PERSONALITIES > 2
220 # if defined HAVE_MX32_MPERS
225 #endif /* SUPPORTED_PERSONALITIES > 2 */
228 printf("%s -- version %s\n"
229 "Copyright (c) 1991-%s The strace developers <%s>.\n"
230 "This is free software; see the source for copying conditions. There is NO\n"
231 "warranty; not even for MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.\n",
232 PACKAGE_NAME, PACKAGE_VERSION, COPYRIGHT_YEAR, PACKAGE_URL);
233 printf("\nOptional features enabled:%s\n",
234 features[0] ? features : " (none)");
241 usage: strace [-CdffhiqrtttTvVwxxy] [-I n] [-e expr]...\n\
242 [-a column] [-o file] [-s strsize] [-P path]...\n\
243 -p pid... / [-D] [-E var=val]... [-u username] PROG [ARGS]\n\
244 or: strace -c[dfw] [-I n] [-e expr]... [-O overhead] [-S sortby]\n\
245 -p pid... / [-D] [-E var=val]... [-u username] PROG [ARGS]\n\
248 -a column alignment COLUMN for printing syscall results (default %d)\n\
249 -i print instruction pointer at time of syscall\n\
253 -k obtain stack trace between each syscall (experimental)\n\
257 -o file send trace output to FILE instead of stderr\n\
258 -q suppress messages about attaching, detaching, etc.\n\
259 -r print relative timestamp\n\
260 -s strsize limit length of print strings to STRSIZE chars (default %d)\n\
261 -t print absolute timestamp\n\
262 -tt print absolute timestamp with usecs\n\
263 -T print time spent in each syscall\n\
264 -x print non-ascii strings in hex\n\
265 -xx print all strings in hex\n\
266 -y print paths associated with file descriptor arguments\n\
267 -yy print protocol specific information associated with socket file descriptors\n\
270 -c count time, calls, and errors for each syscall and report summary\n\
271 -C like -c but also print regular output\n\
272 -O overhead set overhead for tracing syscalls to OVERHEAD usecs\n\
273 -S sortby sort syscall counts by: time, calls, name, nothing (default %s)\n\
274 -w summarise syscall latency (default is system time)\n\
277 -e expr a qualifying expression: option=[!]all or option=[!]val1[,val2]...\n\
278 options: trace, abbrev, verbose, raw, signal, read, write, fault\n\
279 -P path trace accesses to path\n\
282 -b execve detach on execve syscall\n\
283 -D run tracer process as a detached grandchild, not as parent\n\
285 -ff follow forks with output into separate files\n\
287 1: no signals are blocked\n\
288 2: fatal signals are blocked while decoding syscall (default)\n\
289 3: fatal signals are always blocked (default if '-o FILE PROG')\n\
290 4: fatal signals and SIGTSTP (^Z) are always blocked\n\
291 (useful to make 'strace -o FILE PROG' not stop on ^Z)\n\
294 -E var remove var from the environment for command\n\
295 -E var=val put var=val in the environment for command\n\
296 -p pid trace process with process id PID, may be repeated\n\
297 -u username run command as username handling setuid and/or setgid\n\
300 -d enable debug output to stderr\n\
301 -v verbose mode: print unabbreviated argv, stat, termios, etc. args\n\
302 -h print help message\n\
305 /* ancient, no one should use it
306 -F -- attempt to follow vforks (deprecated, use -f)\n\
308 /* this is broken, so don't document it
309 -z -- print only succeeding syscalls\n\
311 , DEFAULT_ACOLUMN, DEFAULT_STRLEN, DEFAULT_SORTBY);
315 void ATTRIBUTE_NORETURN
318 if (strace_tracer_pid == getpid()) {
328 error_opt_arg(int opt, const char *arg)
330 error_msg_and_help("invalid -%c argument: '%s'", opt, arg);
333 static const char *ptrace_attach_cmd;
336 ptrace_attach_or_seize(int pid)
341 return ptrace_attach_cmd = "PTRACE_ATTACH",
342 ptrace(PTRACE_ATTACH, pid, 0L, 0L);
343 r = ptrace(PTRACE_SEIZE, pid, 0L, (unsigned long) ptrace_setoptions);
345 return ptrace_attach_cmd = "PTRACE_SEIZE", r;
346 r = ptrace(PTRACE_INTERRUPT, pid, 0L, 0L);
347 return ptrace_attach_cmd = "PTRACE_INTERRUPT", r;
349 return ptrace_attach_cmd = "PTRACE_ATTACH",
350 ptrace(PTRACE_ATTACH, pid, 0L, 0L);
355 * Used when we want to unblock stopped traced process.
356 * Should be only used with PTRACE_CONT, PTRACE_DETACH and PTRACE_SYSCALL.
357 * Returns 0 on success or if error was ESRCH
358 * (presumably process was killed while we talk to it).
359 * Otherwise prints error message and returns -1.
362 ptrace_restart(const unsigned int op, struct tcb *const tcp, unsigned int sig)
368 ptrace(op, tcp->pid, 0L, (unsigned long) sig);
388 * Why curcol != 0? Otherwise sometimes we get this:
390 * 10252 kill(10253, SIGKILL) = 0
391 * <ptrace(SYSCALL,10252):No such process>10253 ...next decode...
393 * 10252 died after we retrieved syscall exit data,
394 * but before we tried to restart it. Log looks ugly.
396 if (current_tcp && current_tcp->curcol != 0) {
397 tprintf(" <ptrace(%s):%s>\n", msg, strerror(err));
403 perror_msg("ptrace(PTRACE_%s,pid:%d,sig:%u)", msg, tcp->pid, sig);
408 set_cloexec_flag(int fd)
412 flags = fcntl(fd, F_GETFD);
414 /* Can happen only if fd is bad.
415 * Should never happen: if it does, we have a bug
416 * in the caller. Therefore we just abort
417 * instead of propagating the error.
419 perror_msg_and_die("fcntl(%d, F_GETFD)", fd);
422 newflags = flags | FD_CLOEXEC;
423 if (flags == newflags)
426 fcntl(fd, F_SETFD, newflags); /* never fails */
430 kill_save_errno(pid_t pid, int sig)
432 int saved_errno = errno;
434 (void) kill(pid, sig);
439 * When strace is setuid executable, we have to swap uids
440 * before and after filesystem and process management operations.
445 int euid = geteuid(), uid = getuid();
447 if (euid != uid && setreuid(euid, uid) < 0) {
448 perror_msg_and_die("setreuid");
453 strace_fopen(const char *path)
458 fp = fopen_stream(path, "w");
460 perror_msg_and_die("Can't fopen '%s'", path);
462 set_cloexec_flag(fileno(fp));
466 static int popen_pid;
469 # define _PATH_BSHELL "/bin/sh"
473 * We cannot use standard popen(3) here because we have to distinguish
474 * popen child process from other processes we trace, and standard popen(3)
475 * does not export its child's pid.
478 strace_popen(const char *command)
486 perror_msg_and_die("pipe");
488 set_cloexec_flag(fds[1]); /* never fails */
492 perror_msg_and_die("vfork");
499 perror_msg_and_die("dup2");
502 execl(_PATH_BSHELL, "sh", "-c", command, NULL);
503 perror_msg_and_die("Can't execute '%s'", _PATH_BSHELL);
510 fp = fdopen(fds[1], "w");
512 perror_msg_and_die("fdopen");
517 outf_perror(const struct tcb * const tcp)
519 if (tcp->outf == stderr)
522 /* This is ugly, but we don't store separate file names */
524 perror_msg("%s.%u", outfname, tcp->pid);
526 perror_msg("%s", outfname);
529 ATTRIBUTE_FORMAT((printf, 1, 0))
531 tvprintf(const char *const fmt, va_list args)
534 int n = vfprintf(current_tcp->outf, fmt, args);
536 /* very unlikely due to vfprintf buffering */
537 outf_perror(current_tcp);
539 current_tcp->curcol += n;
544 tprintf(const char *fmt, ...)
552 #ifndef HAVE_FPUTS_UNLOCKED
553 # define fputs_unlocked fputs
557 tprints(const char *str)
560 int n = fputs_unlocked(str, current_tcp->outf);
562 current_tcp->curcol += strlen(str);
565 /* very unlikely due to fputs_unlocked buffering */
566 outf_perror(current_tcp);
571 tprints_comment(const char *const str)
574 tprintf(" /* %s */", str);
578 tprintf_comment(const char *fmt, ...)
592 flush_tcp_output(const struct tcb *const tcp)
594 if (fflush(tcp->outf))
602 current_tcp->curcol = 0;
603 flush_tcp_output(current_tcp);
606 printing_tcp->curcol = 0;
612 set_current_tcp(const struct tcb *tcp)
614 current_tcp = (struct tcb *) tcp;
616 /* Sync current_personality and stuff */
618 set_personality(current_tcp->currpers);
622 printleader(struct tcb *tcp)
624 /* If -ff, "previous tcb we printed" is always the same as current,
625 * because we have per-tcb output files.
631 set_current_tcp(printing_tcp);
632 if (printing_tcp->curcol != 0 && (followfork < 2 || printing_tcp == tcp)) {
634 * case 1: we have a shared log (i.e. not -ff), and last line
635 * wasn't finished (same or different tcb, doesn't matter).
636 * case 2: split log, we are the same tcb, but our last line
637 * didn't finish ("SIGKILL nuked us after syscall entry" etc).
639 tprints(" <unfinished ...>\n");
640 printing_tcp->curcol = 0;
645 set_current_tcp(tcp);
646 current_tcp->curcol = 0;
649 tprintf("%-5d ", tcp->pid);
650 else if (nprocs > 1 && !outfname)
651 tprintf("[pid %5u] ", tcp->pid);
654 char str[sizeof("HH:MM:SS")];
655 struct timeval tv, dtv;
656 static struct timeval otv;
658 gettimeofday(&tv, NULL);
662 tv_sub(&dtv, &tv, &otv);
663 tprintf("%6ld.%06ld ",
664 (long) dtv.tv_sec, (long) dtv.tv_usec);
666 } else if (tflag > 2) {
667 tprintf("%ld.%06ld ",
668 (long) tv.tv_sec, (long) tv.tv_usec);
670 time_t local = tv.tv_sec;
671 strftime(str, sizeof(str), "%T", localtime(&local));
673 tprintf("%s.%06ld ", str, (long) tv.tv_usec);
685 if (current_tcp->curcol < acolumn)
686 tprints(acolumn_spaces + current_tcp->curcol);
689 /* Should be only called directly *after successful attach* to a tracee.
690 * Otherwise, "strace -oFILE -ff -p<nonexistant_pid>"
691 * may create bogus empty FILE.<nonexistant_pid>, and then die.
694 newoutf(struct tcb *tcp)
696 tcp->outf = shared_log; /* if not -ff mode, the same file is for all */
697 if (followfork >= 2) {
699 xsprintf(name, "%s.%u", outfname, tcp->pid);
700 tcp->outf = strace_fopen(name);
707 /* Allocate some (more) TCBs (and expand the table).
708 We don't want to relocate the TCBs because our
709 callers have pointers and it would be a pain.
710 So tcbtab is a table of pointers. Since we never
711 free the TCBs, we allocate a single chunk of many. */
712 size_t old_tcbtabsize;
714 struct tcb **tcb_ptr;
716 old_tcbtabsize = tcbtabsize;
718 tcbtab = xgrowarray(tcbtab, &tcbtabsize, sizeof(tcbtab[0]));
719 newtcbs = xcalloc(tcbtabsize - old_tcbtabsize, sizeof(newtcbs[0]));
721 for (tcb_ptr = tcbtab + old_tcbtabsize;
722 tcb_ptr < tcbtab + tcbtabsize; tcb_ptr++, newtcbs++)
732 if (nprocs == tcbtabsize)
735 for (i = 0; i < tcbtabsize; i++) {
738 memset(tcp, 0, sizeof(*tcp));
740 #if SUPPORTED_PERSONALITIES > 1
741 tcp->currpers = current_personality;
745 if (stack_trace_enabled)
746 unwind_tcb_init(tcp);
750 debug_msg("new tcb for pid %d, active tcbs:%d",
755 error_msg_and_die("bug in alloctcb");
759 get_tcb_priv_data(const struct tcb *tcp)
761 return tcp->_priv_data;
765 set_tcb_priv_data(struct tcb *tcp, void *const priv_data,
766 void (*const free_priv_data)(void *))
771 tcp->_free_priv_data = free_priv_data;
772 tcp->_priv_data = priv_data;
778 free_tcb_priv_data(struct tcb *tcp)
780 if (tcp->_priv_data) {
781 if (tcp->_free_priv_data) {
782 tcp->_free_priv_data(tcp->_priv_data);
783 tcp->_free_priv_data = NULL;
785 tcp->_priv_data = NULL;
790 droptcb(struct tcb *tcp)
796 for (p = 0; p < SUPPORTED_PERSONALITIES; ++p)
797 free(tcp->inject_vec[p]);
799 free_tcb_priv_data(tcp);
802 if (stack_trace_enabled) {
807 mmap_cache_delete(tcp, __func__);
810 debug_msg("dropped tcb for pid %d, %d remain", tcp->pid, nprocs);
813 if (followfork >= 2) {
814 if (tcp->curcol != 0)
815 fprintf(tcp->outf, " <detached ...>\n");
818 if (printing_tcp == tcp && tcp->curcol != 0)
819 fprintf(tcp->outf, " <detached ...>\n");
820 flush_tcp_output(tcp);
824 if (current_tcp == tcp)
825 set_current_tcp(NULL);
826 if (printing_tcp == tcp)
829 memset(tcp, 0, sizeof(*tcp));
832 /* Detach traced process.
833 * Never call DETACH twice on the same process as both unattached and
834 * attached-unstopped processes give the same ESRCH. For unattached process we
835 * would SIGSTOP it and wait for its SIGSTOP notification forever.
838 detach(struct tcb *tcp)
844 * Linux wrongly insists the child be stopped
845 * before detaching. Arghh. We go through hoops
846 * to make a clean break of things.
849 if (!(tcp->flags & TCB_ATTACHED))
852 /* We attached but possibly didn't see the expected SIGSTOP.
853 * We must catch exactly one as otherwise the detached process
854 * would be left stopped (process state T).
856 if (tcp->flags & TCB_IGNORE_ONE_SIGSTOP)
859 error = ptrace(PTRACE_DETACH, tcp->pid, 0, 0);
861 /* On a clear day, you can see forever. */
864 if (errno != ESRCH) {
865 /* Shouldn't happen. */
866 perror_func_msg("ptrace(PTRACE_DETACH,%u)", tcp->pid);
869 /* ESRCH: process is either not stopped or doesn't exist. */
870 if (my_tkill(tcp->pid, 0) < 0) {
872 /* Shouldn't happen. */
873 perror_func_msg("tkill(%u,0)", tcp->pid);
874 /* else: process doesn't exist. */
877 /* Process is not stopped, need to stop it. */
880 * With SEIZE, tracee can be in group-stop already.
881 * In this state sending it another SIGSTOP does nothing.
882 * Need to use INTERRUPT.
883 * Testcase: trying to ^C a "strace -p <stopped_process>".
885 error = ptrace(PTRACE_INTERRUPT, tcp->pid, 0, 0);
889 perror_func_msg("ptrace(PTRACE_INTERRUPT,%u)", tcp->pid);
891 error = my_tkill(tcp->pid, SIGSTOP);
895 perror_func_msg("tkill(%u,SIGSTOP)", tcp->pid);
897 /* Either process doesn't exist, or some weird error. */
901 /* We end up here in three cases:
902 * 1. We sent PTRACE_INTERRUPT (use_seize case)
903 * 2. We sent SIGSTOP (!use_seize)
904 * 3. Attach SIGSTOP was already pending (TCB_IGNORE_ONE_SIGSTOP set)
908 if (waitpid(tcp->pid, &status, __WALL) < 0) {
912 * if (errno == ECHILD) break;
913 * ^^^ WRONG! We expect this PID to exist,
914 * and want to emit a message otherwise:
916 perror_func_msg("waitpid(%u)", tcp->pid);
919 if (!WIFSTOPPED(status)) {
921 * Tracee exited or was killed by signal.
922 * We shouldn't normally reach this place:
923 * we don't want to consume exit status.
924 * Consider "strace -p PID" being ^C-ed:
925 * we want merely to detach from PID.
927 * However, we _can_ end up here if tracee
932 sig = WSTOPSIG(status);
933 debug_msg("detach wait: event:%d sig:%d",
934 (unsigned) status >> 16, sig);
936 unsigned event = (unsigned)status >> 16;
937 if (event == PTRACE_EVENT_STOP /*&& sig == SIGTRAP*/) {
939 * sig == SIGTRAP: PTRACE_INTERRUPT stop.
940 * sig == other: process was already stopped
941 * with this stopping sig (see tests/detach-stopped).
942 * Looks like re-injecting this sig is not necessary
943 * in DETACH for the tracee to remain stopped.
948 * PTRACE_INTERRUPT is not guaranteed to produce
949 * the above event if other ptrace-stop is pending.
950 * See tests/detach-sleeping testcase:
951 * strace got SIGINT while tracee is sleeping.
952 * We sent PTRACE_INTERRUPT.
953 * We see syscall exit, not PTRACE_INTERRUPT stop.
954 * We won't get PTRACE_INTERRUPT stop
955 * if we would CONT now. Need to DETACH.
957 if (sig == syscall_trap_sig)
959 /* else: not sure in which case we can be here.
960 * Signal stop? Inject it while detaching.
962 ptrace_restart(PTRACE_DETACH, tcp, sig);
965 /* Note: this check has to be after use_seize check */
966 /* (else, in use_seize case SIGSTOP will be mistreated) */
967 if (sig == SIGSTOP) {
968 /* Detach, suppressing SIGSTOP */
969 ptrace_restart(PTRACE_DETACH, tcp, 0);
972 if (sig == syscall_trap_sig)
974 /* Can't detach just yet, may need to wait for SIGSTOP */
975 error = ptrace_restart(PTRACE_CONT, tcp, sig);
977 /* Should not happen.
978 * Note: ptrace_restart returns 0 on ESRCH, so it's not it.
979 * ptrace_restart already emitted error message.
986 if (!qflag && (tcp->flags & TCB_ATTACHED))
987 error_msg("Process %u detached", tcp->pid);
993 process_opt_p_list(char *opt)
997 * We accept -p PID,PID; -p "`pidof PROG`"; -p "`pgrep PROG`".
998 * pidof uses space as delim, pgrep uses newline. :(
1001 char *delim = opt + strcspn(opt, "\n\t ,");
1005 pid = string_to_uint(opt);
1007 error_msg_and_die("Invalid process id: '%s'", opt);
1009 if (pid == strace_tracer_pid) {
1010 error_msg_and_die("I'm sorry, I can't let you do that, Dave.");
1021 attach_tcb(struct tcb *const tcp)
1023 if (ptrace_attach_or_seize(tcp->pid) < 0) {
1024 perror_msg("attach: ptrace(%s, %d)",
1025 ptrace_attach_cmd, tcp->pid);
1030 tcp->flags |= TCB_ATTACHED | TCB_GRABBED | TCB_STARTUP |
1031 post_attach_sigstop;
1033 debug_msg("attach to pid %d (main) succeeded", tcp->pid);
1035 static const char task_path[] = "/proc/%d/task";
1036 char procdir[sizeof(task_path) + sizeof(int) * 3];
1038 unsigned int ntid = 0, nerr = 0;
1040 if (followfork && tcp->pid != strace_child &&
1041 xsprintf(procdir, task_path, tcp->pid) > 0 &&
1042 (dir = opendir(procdir)) != NULL) {
1045 while ((de = read_dir(dir)) != NULL) {
1046 if (de->d_fileno == 0)
1049 int tid = string_to_uint(de->d_name);
1050 if (tid <= 0 || tid == tcp->pid)
1054 if (ptrace_attach_or_seize(tid) < 0) {
1056 debug_perror_msg("attach: ptrace(%s, %d)",
1057 ptrace_attach_cmd, tid);
1060 debug_msg("attach to pid %d succeeded", tid);
1062 struct tcb *tid_tcp = alloctcb(tid);
1063 tid_tcp->flags |= TCB_ATTACHED | TCB_GRABBED |
1064 TCB_STARTUP | 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_SETMASK, &blocked_set, NULL);
1098 if (daemonized_tracer) {
1101 perror_func_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, &start_set, NULL);
1141 sigprocmask(SIG_SETMASK, &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, &start_set, NULL);
1159 /* Stack-o-phobic exec helper, in the hope to work around
1160 * NOMMU + "daemonized tracer" difficulty.
1162 struct exec_params {
1168 struct sigaction child_sa;
1170 static struct exec_params params_for_tracee;
1172 static void ATTRIBUTE_NOINLINE ATTRIBUTE_NORETURN
1175 struct exec_params *params = ¶ms_for_tracee;
1177 if (params->fd_to_close >= 0)
1178 close(params->fd_to_close);
1179 if (!daemonized_tracer && !use_seize) {
1180 if (ptrace(PTRACE_TRACEME, 0L, 0L, 0L) < 0) {
1181 perror_msg_and_die("ptrace(PTRACE_TRACEME, ...)");
1185 if (username != NULL) {
1187 * It is important to set groups before we
1188 * lose privileges on setuid.
1190 if (initgroups(username, run_gid) < 0) {
1191 perror_msg_and_die("initgroups");
1193 if (setregid(run_gid, params->run_egid) < 0) {
1194 perror_msg_and_die("setregid");
1196 if (setreuid(run_uid, params->run_euid) < 0) {
1197 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 if (params_for_tracee.child_sa.sa_handler != SIG_DFL)
1223 sigaction(SIGCHLD, ¶ms_for_tracee.child_sa, NULL);
1225 execv(params->pathname, params->argv);
1226 perror_msg_and_die("exec");
1230 * Open a dummy descriptor for use as a placeholder.
1231 * The descriptor is O_RDONLY with FD_CLOEXEC flag set.
1232 * A read attempt from such descriptor ends with EOF,
1233 * a write attempt is rejected with EBADF.
1236 open_dummy_desc(void)
1241 perror_func_msg_and_die("pipe");
1243 set_cloexec_flag(fds[0]);
1247 /* placeholder fds status for stdin and stdout */
1248 static bool fd_is_placeholder[2];
1251 * Ensure that all standard file descriptors are open by opening placeholder
1252 * file descriptors for those standard file descriptors that are not open.
1254 * The information which descriptors have been made open is saved
1255 * in fd_is_placeholder for later use.
1258 ensure_standard_fds_opened(void)
1262 while ((fd = open_dummy_desc()) <= 2) {
1265 fd_is_placeholder[fd] = true;
1273 * Redirect stdin and stdout unless they have been opened earlier
1274 * by ensure_standard_fds_opened as placeholders.
1277 redirect_standard_fds(void)
1282 * It might be a good idea to redirect stderr as well,
1283 * but we sometimes need to print error messages.
1285 for (i = 0; i <= 1; ++i) {
1286 if (!fd_is_placeholder[i]) {
1294 startup_child(char **argv)
1296 struct_stat statbuf;
1297 const char *filename;
1298 size_t filename_len;
1299 char pathname[PATH_MAX];
1304 filename_len = strlen(filename);
1306 if (filename_len > sizeof(pathname) - 1) {
1307 errno = ENAMETOOLONG;
1308 perror_msg_and_die("exec");
1310 if (strchr(filename, '/')) {
1311 strcpy(pathname, filename);
1313 #ifdef USE_DEBUGGING_EXEC
1315 * Debuggers customarily check the current directory
1316 * first regardless of the path but doing that gives
1317 * security geeks a panic attack.
1319 else if (stat_file(filename, &statbuf) == 0)
1320 strcpy(pathname, filename);
1321 #endif /* USE_DEBUGGING_EXEC */
1326 for (path = getenv("PATH"); path && *path; path += m) {
1327 const char *colon = strchr(path, ':');
1332 m = n = strlen(path);
1334 if (!getcwd(pathname, PATH_MAX))
1336 len = strlen(pathname);
1337 } else if (n > sizeof(pathname) - 1)
1340 strncpy(pathname, path, n);
1343 if (len && pathname[len - 1] != '/')
1344 pathname[len++] = '/';
1345 if (filename_len + len > sizeof(pathname) - 1)
1347 strcpy(pathname + len, filename);
1348 if (stat_file(pathname, &statbuf) == 0 &&
1349 /* Accept only regular files
1350 with some execute bits set.
1351 XXX not perfect, might still fail */
1352 S_ISREG(statbuf.st_mode) &&
1353 (statbuf.st_mode & 0111))
1356 if (!path || !*path)
1359 if (stat_file(pathname, &statbuf) < 0) {
1360 perror_msg_and_die("Can't stat '%s'", filename);
1363 params_for_tracee.fd_to_close = (shared_log != stderr) ? fileno(shared_log) : -1;
1364 params_for_tracee.run_euid = (statbuf.st_mode & S_ISUID) ? statbuf.st_uid : run_uid;
1365 params_for_tracee.run_egid = (statbuf.st_mode & S_ISGID) ? statbuf.st_gid : run_gid;
1366 params_for_tracee.argv = argv;
1368 * On NOMMU, can be safely freed only after execve in tracee.
1369 * It's hard to know when that happens, so we just leak it.
1371 params_for_tracee.pathname = NOMMU_SYSTEM ? xstrdup(pathname) : pathname;
1373 #if defined HAVE_PRCTL && defined PR_SET_PTRACER && defined PR_SET_PTRACER_ANY
1374 if (daemonized_tracer)
1375 prctl(PR_SET_PTRACER, PR_SET_PTRACER_ANY);
1380 perror_func_msg_and_die("fork");
1382 if ((pid != 0 && daemonized_tracer)
1383 || (pid == 0 && !daemonized_tracer)
1385 /* We are to become the tracee. Two cases:
1387 * not -D: we are child
1392 /* We are the tracer */
1394 if (!daemonized_tracer) {
1397 /* child did PTRACE_TRACEME, nothing to do in parent */
1399 if (!NOMMU_SYSTEM) {
1400 /* Wait until child stopped itself */
1402 while (waitpid(pid, &status, WSTOPPED) < 0) {
1405 perror_msg_and_die("waitpid");
1407 if (!WIFSTOPPED(status) || WSTOPSIG(status) != SIGSTOP) {
1408 kill_save_errno(pid, SIGKILL);
1409 perror_msg_and_die("Unexpected wait status %#x",
1413 /* Else: NOMMU case, we have no way to sync.
1414 * Just attach to it as soon as possible.
1415 * This means that we may miss a few first syscalls...
1418 if (ptrace_attach_or_seize(pid)) {
1419 kill_save_errno(pid, SIGKILL);
1420 perror_msg_and_die("attach: ptrace(%s, %d)",
1421 ptrace_attach_cmd, pid);
1426 tcp = alloctcb(pid);
1427 tcp->flags |= TCB_ATTACHED | TCB_STARTUP
1428 | TCB_SKIP_DETACH_ON_FIRST_EXEC
1429 | (NOMMU_SYSTEM ? 0 : (TCB_HIDE_LOG | post_attach_sigstop));
1432 /* With -D, we are *child* here, the tracee is our parent. */
1433 strace_child = strace_tracer_pid;
1434 strace_tracer_pid = getpid();
1435 tcp = alloctcb(strace_child);
1436 tcp->flags |= TCB_SKIP_DETACH_ON_FIRST_EXEC | TCB_HIDE_LOG;
1437 /* attaching will be done later, by startup_attach */
1438 /* note: we don't do newoutf(tcp) here either! */
1440 /* NOMMU BUG! -D mode is active, we (child) return,
1441 * and we will scribble over parent's stack!
1442 * When parent later unpauses, it segfaults.
1445 * (1) by declaring exec_or_die() NORETURN,
1446 * hopefully compiler will just jump to it
1447 * instead of call (won't push anything to stack),
1448 * (2) by trying very hard in exec_or_die()
1449 * to not use any stack,
1450 * (3) having a really big (PATH_MAX) stack object
1451 * in this function, which creates a "buffer" between
1452 * child's and parent's stack pointers.
1453 * This may save us if (1) and (2) failed
1454 * and compiler decided to use stack in exec_or_die() anyway
1455 * (happens on i386 because of stack parameter passing).
1457 * A cleaner solution is to use makecontext + setcontext
1458 * to create a genuine separate stack and execute on it.
1462 * A case where straced process is part of a pipe:
1463 * { sleep 1; yes | head -n99999; } | strace -o/dev/null sh -c 'exec <&-; sleep 9'
1464 * If strace won't close its fd#0, closing it in tracee is not enough:
1465 * the pipe is still open, it has a reader. Thus, "head" will not get its
1466 * SIGPIPE at once, on the first write.
1468 * Preventing it by redirecting strace's stdin/out.
1469 * (Don't leave fds 0 and 1 closed, this is bad practice: future opens
1470 * will reuse them, unexpectedly making a newly opened object "stdin").
1472 redirect_standard_fds();
1477 test_ptrace_seize(void)
1481 /* Need fork for test. NOMMU has no forks */
1483 post_attach_sigstop = 0; /* this sets use_seize to 1 */
1489 perror_func_msg_and_die("fork");
1496 /* PTRACE_SEIZE, unlike ATTACH, doesn't force tracee to trap. After
1497 * attaching tracee continues to run unless a trap condition occurs.
1498 * PTRACE_SEIZE doesn't affect signal or group stop state.
1500 if (ptrace(PTRACE_SEIZE, pid, 0, 0) == 0) {
1501 post_attach_sigstop = 0; /* this sets use_seize to 1 */
1503 debug_msg("PTRACE_SEIZE doesn't work");
1509 int status, tracee_pid;
1512 tracee_pid = waitpid(pid, &status, 0);
1513 if (tracee_pid <= 0) {
1516 perror_func_msg_and_die("unexpected wait result %d",
1519 if (WIFSIGNALED(status))
1522 error_func_msg_and_die("unexpected wait status %#x", status);
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 set_sighandler(int signo, void (*sighandler)(int), struct sigaction *oldact)
1565 /* if signal handler is a function, add the signal to blocked_set */
1566 if (sighandler != SIG_IGN && sighandler != SIG_DFL)
1567 sigaddset(&blocked_set, signo);
1569 const struct sigaction sa = { .sa_handler = sighandler };
1570 sigaction(signo, &sa, oldact);
1574 * Initialization part of main() was eating much stack (~0.5k),
1575 * which was unused after init.
1576 * We can reuse it if we move init code into a separate function.
1578 * Don't want main() to inline us and defeat the reason
1579 * we have a separate function.
1581 static void ATTRIBUTE_NOINLINE
1582 init(int argc, char *argv[])
1587 if (!program_invocation_name || !*program_invocation_name) {
1588 static char name[] = "strace";
1589 program_invocation_name =
1590 (argc > 0 && argv[0] && *argv[0]) ? argv[0] : name;
1593 strace_tracer_pid = getpid();
1595 os_release = get_os_release();
1597 shared_log = stderr;
1598 set_sortby(DEFAULT_SORTBY);
1599 set_personality(DEFAULT_PERSONALITY);
1600 qualify("trace=all");
1601 qualify("abbrev=all");
1602 qualify("verbose=all");
1603 #if DEFAULT_QUAL_FLAGS != (QUAL_TRACE | QUAL_ABBREV | QUAL_VERBOSE)
1604 # error Bug in DEFAULT_QUAL_FLAGS
1606 qualify("signal=all");
1607 while ((c = getopt(argc, argv, "+"
1608 #ifdef USE_LIBUNWIND
1611 "a:b:cCdDe:E:fFhiI:o:O:p:P:qrs:S:tTu:vVwxyz")) != EOF) {
1614 acolumn = string_to_uint(optarg);
1616 error_opt_arg(c, optarg);
1619 if (strcmp(optarg, "execve") != 0)
1620 error_msg_and_die("Syscall '%s' for -b isn't supported",
1622 detach_on_execve = 1;
1625 if (cflag == CFLAG_BOTH) {
1626 error_msg_and_help("-c and -C are mutually exclusive");
1628 cflag = CFLAG_ONLY_STATS;
1631 if (cflag == CFLAG_ONLY_STATS) {
1632 error_msg_and_help("-c and -C are mutually exclusive");
1640 daemonized_tracer = 1;
1646 if (putenv(optarg) < 0)
1647 perror_msg_and_die("putenv");
1662 opt_intr = string_to_uint_upto(optarg, NUM_INTR_OPTS - 1);
1664 error_opt_arg(c, optarg);
1666 #ifdef USE_LIBUNWIND
1668 stack_trace_enabled = true;
1675 i = string_to_uint(optarg);
1677 error_opt_arg(c, optarg);
1681 process_opt_p_list(optarg);
1684 pathtrace_select(optarg);
1693 i = string_to_uint(optarg);
1694 if (i < 0 || (unsigned int) i > -1U / 4)
1695 error_opt_arg(c, optarg);
1711 qualify("abbrev=none");
1718 count_wallclock = 1;
1727 not_failing_only = 1;
1730 error_msg_and_help(NULL);
1738 if (argc < 0 || (!nprocs && !argc)) {
1739 error_msg_and_help("must have PROG [ARGS] or -p PID");
1742 if (!argc && daemonized_tracer) {
1743 error_msg_and_help("PROG [ARGS] must be specified with -D");
1748 error_msg("deprecated option -F ignored");
1750 error_msg("option -F is deprecated, "
1751 "please use -f instead");
1756 if (followfork >= 2 && cflag) {
1757 error_msg_and_help("(-c or -C) and -ff are mutually exclusive");
1760 if (count_wallclock && !cflag) {
1761 error_msg_and_help("-w must be given with (-c or -C)");
1764 if (cflag == CFLAG_ONLY_STATS) {
1766 error_msg("-%c has no effect with -c", 'i');
1767 #ifdef USE_LIBUNWIND
1768 if (stack_trace_enabled)
1769 error_msg("-%c has no effect with -c", 'k');
1772 error_msg("-%c has no effect with -c", 'r');
1774 error_msg("-%c has no effect with -c", 't');
1776 error_msg("-%c has no effect with -c", 'T');
1778 error_msg("-%c has no effect with -c", 'y');
1783 error_msg("-tt has no effect with -r");
1787 acolumn_spaces = xmalloc(acolumn + 1);
1788 memset(acolumn_spaces, ' ', acolumn);
1789 acolumn_spaces[acolumn] = '\0';
1791 sigprocmask(SIG_SETMASK, NULL, &start_set);
1792 memcpy(&blocked_set, &start_set, sizeof(blocked_set));
1794 set_sighandler(SIGCHLD, SIG_DFL, ¶ms_for_tracee.child_sa);
1796 #ifdef USE_LIBUNWIND
1797 if (stack_trace_enabled) {
1801 for (tcbi = 0; tcbi < tcbtabsize; ++tcbi) {
1802 unwind_tcb_init(tcbtab[tcbi]);
1807 /* See if they want to run as another user. */
1808 if (username != NULL) {
1809 struct passwd *pent;
1811 if (getuid() != 0 || geteuid() != 0) {
1812 error_msg_and_die("You must be root to use the -u option");
1814 pent = getpwnam(username);
1816 error_msg_and_die("Cannot find user '%s'", username);
1818 run_uid = pent->pw_uid;
1819 run_gid = pent->pw_gid;
1826 ptrace_setoptions |= PTRACE_O_TRACECLONE |
1827 PTRACE_O_TRACEFORK |
1828 PTRACE_O_TRACEVFORK;
1829 debug_msg("ptrace_setoptions = %#x", ptrace_setoptions);
1830 test_ptrace_seize();
1833 * Is something weird with our stdin and/or stdout -
1834 * for example, may they be not open? In this case,
1835 * ensure that none of the future opens uses them.
1837 * This was seen in the wild when /proc/sys/kernel/core_pattern
1838 * was set to "|/bin/strace -o/tmp/LOG PROG":
1839 * kernel runs coredump helper with fd#0 open but fd#1 closed (!),
1840 * therefore LOG gets opened to fd#1, and fd#1 is closed by
1841 * "don't hold up stdin/out open" code soon after.
1843 ensure_standard_fds_opened();
1845 /* Check if they want to redirect the output. */
1847 /* See if they want to pipe the output. */
1848 if (outfname[0] == '|' || outfname[0] == '!') {
1850 * We can't do the <outfname>.PID funny business
1851 * when using popen, so prohibit it.
1853 if (followfork >= 2)
1854 error_msg_and_help("piping the output and -ff "
1855 "are mutually exclusive");
1856 shared_log = strace_popen(outfname + 1);
1857 } else if (followfork < 2) {
1858 shared_log = strace_fopen(outfname);
1859 } else if (strlen(outfname) >= PATH_MAX - sizeof(int) * 3) {
1860 errno = ENAMETOOLONG;
1861 perror_msg_and_die("%s", outfname);
1864 /* -ff without -o FILE is the same as single -f */
1865 if (followfork >= 2)
1869 if (!outfname || outfname[0] == '|' || outfname[0] == '!') {
1870 setvbuf(shared_log, NULL, _IOLBF, 0);
1874 * argv[0] -pPID -oFILE Default interactive setting
1875 * yes * 0 INTR_WHILE_WAIT
1876 * no 1 0 INTR_WHILE_WAIT
1877 * yes * 1 INTR_NEVER
1878 * no 1 1 INTR_WHILE_WAIT
1881 if (outfname && argc) {
1883 opt_intr = INTR_NEVER;
1888 opt_intr = INTR_WHILE_WAIT;
1891 * startup_child() must be called before the signal handlers get
1892 * installed below as they are inherited into the spawned process.
1893 * Also we do not need to be protected by them as during interruption
1894 * in the startup_child() mode we kill the spawned process anyway.
1897 startup_child(argv);
1900 set_sighandler(SIGTTOU, SIG_IGN, NULL);
1901 set_sighandler(SIGTTIN, SIG_IGN, NULL);
1902 if (opt_intr != INTR_ANYWHERE) {
1903 if (opt_intr == INTR_BLOCK_TSTP_TOO)
1904 set_sighandler(SIGTSTP, SIG_IGN, NULL);
1906 * In interactive mode (if no -o OUTFILE, or -p PID is used),
1907 * fatal signals are blocked while syscall stop is processed,
1908 * and acted on in between, when waiting for new syscall stops.
1909 * In non-interactive mode, signals are ignored.
1911 set_sighandler(SIGHUP, interactive ? interrupt : SIG_IGN, NULL);
1912 set_sighandler(SIGINT, interactive ? interrupt : SIG_IGN, NULL);
1913 set_sighandler(SIGQUIT, interactive ? interrupt : SIG_IGN, NULL);
1914 set_sighandler(SIGPIPE, interactive ? interrupt : SIG_IGN, NULL);
1915 set_sighandler(SIGTERM, interactive ? interrupt : SIG_IGN, NULL);
1918 if (nprocs != 0 || daemonized_tracer)
1921 /* Do we want pids printed in our -o OUTFILE?
1922 * -ff: no (every pid has its own file); or
1923 * -f: yes (there can be more pids in the future); or
1924 * -p PID1,PID2: yes (there are already more than one pid)
1926 print_pid_pfx = (outfname && followfork < 2 && (followfork == 1 || nprocs > 1));
1937 for (i = 0; i < tcbtabsize; i++) {
1938 struct tcb *tcp = tcbtab[i];
1939 if (tcp->pid == pid)
1953 /* 'interrupted' is a volatile object, fetch it only once */
1954 fatal_sig = interrupted;
1956 fatal_sig = SIGTERM;
1958 for (i = 0; i < tcbtabsize; i++) {
1962 debug_func_msg("looking at pid %u", tcp->pid);
1963 if (tcp->pid == strace_child) {
1964 kill(tcp->pid, SIGCONT);
1965 kill(tcp->pid, fatal_sig);
1970 call_summary(shared_log);
1980 print_debug_info(const int pid, int status)
1982 const unsigned int event = (unsigned int) status >> 16;
1983 char buf[sizeof("WIFEXITED,exitcode=%u") + sizeof(int)*3 /*paranoia:*/ + 16];
1984 char evbuf[sizeof(",EVENT_VFORK_DONE (%u)") + sizeof(int)*3 /*paranoia:*/ + 16];
1987 if (WIFSIGNALED(status))
1989 xsprintf(buf, "WIFSIGNALED,%ssig=%s",
1990 WCOREDUMP(status) ? "core," : "",
1991 signame(WTERMSIG(status)));
1993 xsprintf(buf, "WIFSIGNALED,sig=%s",
1994 signame(WTERMSIG(status)));
1996 if (WIFEXITED(status))
1997 xsprintf(buf, "WIFEXITED,exitcode=%u", WEXITSTATUS(status));
1998 if (WIFSTOPPED(status))
1999 xsprintf(buf, "WIFSTOPPED,sig=%s", signame(WSTOPSIG(status)));
2002 static const char *const event_names[] = {
2003 [PTRACE_EVENT_CLONE] = "CLONE",
2004 [PTRACE_EVENT_FORK] = "FORK",
2005 [PTRACE_EVENT_VFORK] = "VFORK",
2006 [PTRACE_EVENT_VFORK_DONE] = "VFORK_DONE",
2007 [PTRACE_EVENT_EXEC] = "EXEC",
2008 [PTRACE_EVENT_EXIT] = "EXIT",
2009 /* [PTRACE_EVENT_STOP (=128)] would make biggish array */
2011 const char *e = "??";
2012 if (event < ARRAY_SIZE(event_names))
2013 e = event_names[event];
2014 else if (event == PTRACE_EVENT_STOP)
2016 xsprintf(evbuf, ",EVENT_%s (%u)", e, event);
2018 error_msg("[wait(0x%06x) = %u] %s%s", status, pid, buf, evbuf);
2022 maybe_allocate_tcb(const int pid, int status)
2024 if (!WIFSTOPPED(status)) {
2025 if (detach_on_execve && pid == strace_child) {
2026 /* example: strace -bexecve sh -c 'exec true' */
2031 * This can happen if we inherited an unknown child.
2032 * Example: (sleep 1 & exec strace true)
2034 error_msg("Exit of unknown pid %u ignored", pid);
2038 /* We assume it's a fork/vfork/clone child */
2039 struct tcb *tcp = alloctcb(pid);
2040 tcp->flags |= TCB_ATTACHED | TCB_STARTUP | post_attach_sigstop;
2043 error_msg("Process %d attached", pid);
2047 * This can happen if a clone call misused CLONE_PTRACE itself.
2049 * There used to be a dance around possible re-injection of
2050 * WSTOPSIG(status), but it was later removed as the only
2051 * observable stop here is the initial ptrace-stop.
2053 ptrace(PTRACE_DETACH, pid, NULL, 0L);
2054 error_msg("Detached unknown pid %d", pid);
2060 maybe_switch_tcbs(struct tcb *tcp, const int pid)
2063 struct tcb *execve_thread;
2066 if (ptrace(PTRACE_GETEVENTMSG, pid, NULL, &old_pid) < 0)
2068 /* Avoid truncation in pid2tcb() param passing */
2069 if (old_pid <= 0 || old_pid == pid)
2071 if ((unsigned long) old_pid > UINT_MAX)
2073 execve_thread = pid2tcb(old_pid);
2074 /* It should be !NULL, but I feel paranoid */
2078 if (execve_thread->curcol != 0) {
2080 * One case we are here is -ff:
2081 * try "strace -oLOG -ff test/threaded_execve"
2083 fprintf(execve_thread->outf, " <pid changed to %d ...>\n", pid);
2084 /*execve_thread->curcol = 0; - no need, see code below */
2086 /* Swap output FILEs (needed for -ff) */
2087 fp = execve_thread->outf;
2088 execve_thread->outf = tcp->outf;
2090 /* And their column positions */
2091 execve_thread->curcol = tcp->curcol;
2093 /* Drop leader, but close execve'd thread outfile (if -ff) */
2095 /* Switch to the thread, reusing leader's outfile and pid */
2096 tcp = execve_thread;
2098 if (cflag != CFLAG_ONLY_STATS) {
2100 tprintf("+++ superseded by execve in pid %lu +++\n", old_pid);
2102 tcp->flags |= TCB_REPRINT;
2109 print_signalled(struct tcb *tcp, const int pid, int status)
2111 if (pid == strace_child) {
2112 exit_code = 0x100 | WTERMSIG(status);
2116 if (cflag != CFLAG_ONLY_STATS
2117 && is_number_in_set(WTERMSIG(status), signal_set)) {
2120 tprintf("+++ killed by %s %s+++\n",
2121 signame(WTERMSIG(status)),
2122 WCOREDUMP(status) ? "(core dumped) " : "");
2124 tprintf("+++ killed by %s +++\n",
2125 signame(WTERMSIG(status)));
2132 print_exited(struct tcb *tcp, const int pid, int status)
2134 if (pid == strace_child) {
2135 exit_code = WEXITSTATUS(status);
2139 if (cflag != CFLAG_ONLY_STATS &&
2142 tprintf("+++ exited with %d +++\n", WEXITSTATUS(status));
2148 print_stopped(struct tcb *tcp, const siginfo_t *si, const unsigned int sig)
2150 if (cflag != CFLAG_ONLY_STATS
2152 && is_number_in_set(sig, signal_set)) {
2155 tprintf("--- %s ", signame(sig));
2159 tprintf("--- stopped by %s ---\n", signame(sig));
2165 startup_tcb(struct tcb *tcp)
2167 debug_msg("pid %d has TCB_STARTUP, initializing it", tcp->pid);
2169 tcp->flags &= ~TCB_STARTUP;
2172 debug_msg("setting opts 0x%x on pid %d",
2173 ptrace_setoptions, tcp->pid);
2174 if (ptrace(PTRACE_SETOPTIONS, tcp->pid, NULL, ptrace_setoptions) < 0) {
2175 if (errno != ESRCH) {
2176 /* Should never happen, really */
2177 perror_msg_and_die("PTRACE_SETOPTIONS");
2182 if ((tcp->flags & TCB_GRABBED) && (get_scno(tcp) == 1))
2183 tcp->s_prev_ent = tcp->s_ent;
2187 print_event_exit(struct tcb *tcp)
2189 if (entering(tcp) || filtered(tcp) || hide_log(tcp)
2190 || cflag == CFLAG_ONLY_STATS) {
2194 if (followfork < 2 && printing_tcp && printing_tcp != tcp
2195 && printing_tcp->curcol != 0) {
2196 set_current_tcp(printing_tcp);
2197 tprints(" <unfinished ...>\n");
2198 flush_tcp_output(printing_tcp);
2199 printing_tcp->curcol = 0;
2200 set_current_tcp(tcp);
2203 if ((followfork < 2 && printing_tcp != tcp)
2204 || (tcp->flags & TCB_REPRINT)) {
2205 tcp->flags &= ~TCB_REPRINT;
2207 tprintf("<... %s resumed>", tcp->s_ent->sys_name);
2210 if (!(tcp->sys_func_rval & RVAL_DECODED)) {
2212 * The decoder has probably decided to print something
2213 * on exiting syscall which is not going to happen.
2215 tprints(" <unfinished ...>");
2225 static enum trace_event
2226 next_event(int *pstatus, siginfo_t *si)
2238 * Used to exit simply when nprocs hits zero, but in this testcase:
2239 * int main(void) { _exit(!!fork()); }
2240 * under strace -f, parent sometimes (rarely) manages
2241 * to exit before we see the first stop of the child,
2242 * and we are losing track of it:
2243 * 19923 clone(...) = 19924
2244 * 19923 exit_group(1) = ?
2245 * 19923 +++ exited with 1 +++
2246 * Exiting only when wait() returns ECHILD works better.
2248 if (popen_pid != 0) {
2249 /* However, if -o|logger is in use, we can't do that.
2250 * Can work around that by double-forking the logger,
2251 * but that loses the ability to wait for its completion
2252 * on exit. Oh well...
2259 sigprocmask(SIG_SETMASK, &start_set, NULL);
2260 pid = wait4(-1, pstatus, __WALL, (cflag ? &ru : NULL));
2263 sigprocmask(SIG_SETMASK, &blocked_set, NULL);
2266 if (wait_errno == EINTR)
2268 if (nprocs == 0 && wait_errno == ECHILD)
2271 * If nprocs > 0, ECHILD is not expected,
2272 * treat it as any other error here:
2275 perror_msg_and_die("wait4(__WALL)");
2280 if (pid == popen_pid) {
2281 if (!WIFSTOPPED(status))
2287 print_debug_info(pid, status);
2289 /* Look up 'pid' in our table. */
2293 tcp = maybe_allocate_tcb(pid, status);
2300 /* Set current output file */
2301 set_current_tcp(tcp);
2304 tv_sub(&tcp->dtime, &ru.ru_stime, &tcp->stime);
2305 tcp->stime = ru.ru_stime;
2308 if (WIFSIGNALED(status))
2309 return TE_SIGNALLED;
2311 if (WIFEXITED(status))
2315 * As WCONTINUED flag has not been specified to wait4,
2316 * it cannot be WIFCONTINUED(status), so the only case
2317 * that remains is WIFSTOPPED(status).
2320 /* Is this the very first time we see this tracee stopped? */
2321 if (tcp->flags & TCB_STARTUP)
2324 const unsigned int sig = WSTOPSIG(status);
2325 const unsigned int event = (unsigned int) status >> 16;
2330 * Is this post-attach SIGSTOP?
2331 * Interestingly, the process may stop
2332 * with STOPSIG equal to some other signal
2333 * than SIGSTOP if we happened to attach
2334 * just before the process takes a signal.
2336 if (sig == SIGSTOP && (tcp->flags & TCB_IGNORE_ONE_SIGSTOP)) {
2337 debug_func_msg("ignored SIGSTOP on pid %d", tcp->pid);
2338 tcp->flags &= ~TCB_IGNORE_ONE_SIGSTOP;
2340 } else if (sig == syscall_trap_sig) {
2341 return TE_SYSCALL_STOP;
2343 *si = (siginfo_t) {};
2345 * True if tracee is stopped by signal
2346 * (as opposed to "tracee received signal").
2347 * TODO: shouldn't we check for errno == EINVAL too?
2348 * We can get ESRCH instead, you know...
2350 bool stopped = ptrace(PTRACE_GETSIGINFO, pid, 0, si) < 0;
2351 return stopped ? TE_GROUP_STOP : TE_SIGNAL_DELIVERY_STOP;
2355 case PTRACE_EVENT_STOP:
2357 * PTRACE_INTERRUPT-stop or group-stop.
2358 * PTRACE_INTERRUPT-stop has sig == SIGTRAP here.
2365 return TE_GROUP_STOP;
2369 case PTRACE_EVENT_EXEC:
2370 return TE_STOP_BEFORE_EXECVE;
2371 case PTRACE_EVENT_EXIT:
2372 return TE_STOP_BEFORE_EXIT;
2379 trace_syscall(struct tcb *tcp, unsigned int *sig)
2381 if (entering(tcp)) {
2382 int res = syscall_entering_decode(tcp);
2387 res = syscall_entering_trace(tcp, sig);
2389 syscall_entering_finish(tcp, res);
2392 struct timeval tv = {};
2393 int res = syscall_exiting_decode(tcp, &tv);
2395 res = syscall_exiting_trace(tcp, tv, res);
2397 syscall_exiting_finish(tcp);
2402 /* Returns true iff the main trace loop has to continue. */
2404 dispatch_event(enum trace_event ret, int *pstatus, siginfo_t *si)
2406 unsigned int restart_op = PTRACE_SYSCALL;
2407 unsigned int restart_sig = 0;
2419 case TE_SYSCALL_STOP:
2420 if (trace_syscall(current_tcp, &restart_sig) < 0) {
2422 * ptrace() failed in trace_syscall().
2423 * Likely a result of process disappearing mid-flight.
2424 * Observed case: exit_group() or SIGKILL terminating
2425 * all processes in thread group.
2426 * We assume that ptrace error was caused by process death.
2427 * We used to detach(current_tcp) here, but since we no
2428 * longer implement "detach before death" policy/hack,
2429 * we can let this process to report its death to us
2430 * normally, via WIFEXITED or WIFSIGNALED wait status.
2436 case TE_SIGNAL_DELIVERY_STOP:
2437 restart_sig = WSTOPSIG(*pstatus);
2438 print_stopped(current_tcp, si, restart_sig);
2442 print_signalled(current_tcp, current_tcp->pid, *pstatus);
2443 droptcb(current_tcp);
2447 restart_sig = WSTOPSIG(*pstatus);
2448 print_stopped(current_tcp, NULL, restart_sig);
2451 * This ends ptrace-stop, but does *not* end group-stop.
2452 * This makes stopping signals work properly on straced
2453 * process (that is, process really stops. It used to
2456 restart_op = PTRACE_LISTEN;
2462 print_exited(current_tcp, current_tcp->pid, *pstatus);
2463 droptcb(current_tcp);
2466 case TE_STOP_BEFORE_EXECVE:
2468 * Check that we are inside syscall now (next event after
2469 * PTRACE_EVENT_EXEC should be for syscall exiting). If it is
2470 * not the case, we might have a situation when we attach to a
2471 * process and the first thing we see is a PTRACE_EVENT_EXEC
2472 * and all the following syscall state tracking is screwed up
2475 if (entering(current_tcp)) {
2478 error_msg("Stray PTRACE_EVENT_EXEC from pid %d"
2479 ", trying to recover...",
2482 current_tcp->flags |= TCB_RECOVERING;
2483 ret = trace_syscall(current_tcp, &restart_sig);
2484 current_tcp->flags &= ~TCB_RECOVERING;
2487 /* The reason is described in TE_SYSCALL_STOP */
2493 * Under Linux, execve changes pid to thread leader's pid,
2494 * and we see this changed pid on EVENT_EXEC and later,
2495 * execve sysexit. Leader "disappears" without exit
2496 * notification. Let user know that, drop leader's tcb,
2497 * and fix up pid in execve thread's tcb.
2498 * Effectively, execve thread's tcb replaces leader's tcb.
2500 * BTW, leader is 'stuck undead' (doesn't report WIFEXITED
2501 * on exit syscall) in multithreaded programs exactly
2502 * in order to handle this case.
2504 * PTRACE_GETEVENTMSG returns old pid starting from Linux 3.0.
2505 * On 2.6 and earlier, it can return garbage.
2507 if (os_release >= KERNEL_VERSION(3, 0, 0))
2508 set_current_tcp(maybe_switch_tcbs(current_tcp,
2511 if (detach_on_execve) {
2512 if (current_tcp->flags & TCB_SKIP_DETACH_ON_FIRST_EXEC) {
2513 current_tcp->flags &= ~TCB_SKIP_DETACH_ON_FIRST_EXEC;
2515 detach(current_tcp); /* do "-b execve" thingy */
2521 case TE_STOP_BEFORE_EXIT:
2522 print_event_exit(current_tcp);
2526 /* We handled quick cases, we are permitted to interrupt now. */
2530 if (ptrace_restart(restart_op, current_tcp, restart_sig) < 0) {
2531 /* Note: ptrace_restart emitted error message */
2538 #ifdef ENABLE_COVERAGE_GCOV
2539 extern void __gcov_flush(void);
2542 static void ATTRIBUTE_NORETURN
2547 if (shared_log != stderr)
2550 while (waitpid(popen_pid, NULL, 0) < 0 && errno == EINTR)
2553 if (exit_code > 0xff) {
2554 /* Avoid potential core file clobbering. */
2555 struct_rlimit rlim = {0, 0};
2556 set_rlimit(RLIMIT_CORE, &rlim);
2558 /* Child was killed by a signal, mimic that. */
2560 signal(exit_code, SIG_DFL);
2561 #ifdef ENABLE_COVERAGE_GCOV
2566 /* Unblock the signal. */
2569 sigaddset(&mask, exit_code);
2570 #ifdef ENABLE_COVERAGE_GCOV
2573 sigprocmask(SIG_UNBLOCK, &mask, NULL);
2575 /* Paranoia - what if this signal is not fatal?
2576 Exit with 128 + signo then. */
2583 main(int argc, char *argv[])
2587 exit_code = !nprocs;
2591 while (dispatch_event(next_event(&status, &si), &status, &si))