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>
48 #include <sys/utsname.h>
50 # include <sys/prctl.h>
52 #include <asm/unistd.h>
54 #include "largefile_wrappers.h"
55 #include "mmap_cache.h"
56 #include "number_set.h"
58 #include "printsiginfo.h"
59 #include "trace_event.h"
63 /* In some libc, these aren't declared. Do it ourself: */
64 extern char **environ;
68 #ifdef ENABLE_STACKTRACE
69 /* if this is true do the stack trace for every system call */
70 bool stack_trace_enabled;
73 #define my_tkill(tid, sig) syscall(__NR_tkill, (tid), (sig))
75 /* Glue for systems without a MMU that cannot provide fork() */
76 #if !defined(HAVE_FORK)
78 # define NOMMU_SYSTEM 1
81 # define fork() vfork()
84 const unsigned int syscall_trap_sig = SIGTRAP | 0x80;
86 cflag_t cflag = CFLAG_NONE;
87 unsigned int followfork;
88 unsigned int ptrace_setoptions = PTRACE_O_TRACESYSGOOD | PTRACE_O_TRACEEXEC
96 static unsigned int tflag;
98 static bool print_pid_pfx;
103 INTR_ANYWHERE = 1, /* don't block/ignore any signals */
104 INTR_WHILE_WAIT = 2, /* block fatal signals while decoding syscall. default */
105 INTR_NEVER = 3, /* block fatal signals. default if '-o FILE PROG' */
106 INTR_BLOCK_TSTP_TOO = 4, /* block fatal signals and SIGTSTP (^Z) */
110 /* We play with signal mask only if this mode is active: */
111 #define interactive (opt_intr == INTR_WHILE_WAIT)
114 * daemonized_tracer supports -D option.
115 * With this option, strace forks twice.
116 * Unlike normal case, with -D *grandparent* process exec's,
117 * becoming a traced process. Child exits (this prevents traced process
118 * from having children it doesn't expect to have), and grandchild
119 * attaches to grandparent similarly to strace -p PID.
120 * This allows for more transparent interaction in cases
121 * when process and its parent are communicating via signals,
122 * wait() etc. Without -D, strace process gets lodged in between,
123 * disrupting parent<->child link.
125 static bool daemonized_tracer;
127 static int post_attach_sigstop = TCB_IGNORE_ONE_SIGSTOP;
128 #define use_seize (post_attach_sigstop == 0)
130 /* Sometimes we want to print only succeeding syscalls. */
131 bool not_failing_only;
133 /* Show path associated with fd arguments */
134 unsigned int show_fd_path;
136 static bool detach_on_execve;
138 static int exit_code;
139 static int strace_child;
140 static int strace_tracer_pid;
142 static const char *username;
143 static uid_t run_uid;
144 static gid_t run_gid;
146 unsigned int max_strlen = DEFAULT_STRLEN;
147 static int acolumn = DEFAULT_ACOLUMN;
148 static char *acolumn_spaces;
150 /* Default output style for xlat entities */
151 enum xlat_style xlat_verbosity = XLAT_STYLE_ABBREV;
153 static const char *outfname;
154 /* If -ff, points to stderr. Else, it's our common output log */
155 static FILE *shared_log;
156 static bool open_append;
158 struct tcb *printing_tcp;
159 static struct tcb *current_tcp;
161 static struct tcb **tcbtab;
162 static unsigned int nprocs;
163 static size_t tcbtabsize;
165 #ifndef HAVE_PROGRAM_INVOCATION_NAME
166 char *program_invocation_name;
169 unsigned os_release; /* generated from uname()'s u.release */
171 static void detach(struct tcb *tcp);
172 static void cleanup(void);
173 static void interrupt(int sig);
175 #ifdef HAVE_SIG_ATOMIC_T
176 static volatile sig_atomic_t interrupted, restart_failed;
178 static volatile int interrupted, restart_failed;
181 static sigset_t timer_set;
182 static void timer_sighandler(int);
184 #ifndef HAVE_STRERROR
186 #if !HAVE_DECL_SYS_ERRLIST
188 extern char *sys_errlist[];
194 static char buf[sizeof("Unknown error %d") + sizeof(int)*3];
196 if (err_no < 1 || err_no >= sys_nerr) {
197 xsprintf(buf, "Unknown error %d", err_no);
200 return sys_errlist[err_no];
203 #endif /* HAVE_STERRROR */
208 static const char features[] =
209 #ifdef ENABLE_STACKTRACE
210 " stack-trace=" USE_UNWINDER
215 #if SUPPORTED_PERSONALITIES > 1
216 # if defined HAVE_M32_MPERS
221 #endif /* SUPPORTED_PERSONALITIES > 1 */
222 #if SUPPORTED_PERSONALITIES > 2
223 # if defined HAVE_MX32_MPERS
228 #endif /* SUPPORTED_PERSONALITIES > 2 */
231 printf("%s -- version %s\n"
232 "Copyright (c) 1991-%s The strace developers <%s>.\n"
233 "This is free software; see the source for copying conditions. There is NO\n"
234 "warranty; not even for MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.\n",
235 PACKAGE_NAME, PACKAGE_VERSION, COPYRIGHT_YEAR, PACKAGE_URL);
236 printf("\nOptional features enabled:%s\n",
237 features[0] ? features : " (none)");
244 usage: strace [-CdffhiqrtttTvVwxxy] [-I n] [-e expr]...\n\
245 [-a column] [-o file] [-s strsize] [-P path]...\n\
246 -p pid... / [-D] [-E var=val]... [-u username] PROG [ARGS]\n\
247 or: strace -c[dfw] [-I n] [-e expr]... [-O overhead] [-S sortby]\n\
248 -p pid... / [-D] [-E var=val]... [-u username] PROG [ARGS]\n\
251 -a column alignment COLUMN for printing syscall results (default %d)\n\
252 -i print instruction pointer at time of syscall\n\
254 #ifdef ENABLE_STACKTRACE
256 -k obtain stack trace between each syscall\n\
260 -o file send trace output to FILE instead of stderr\n\
261 -q suppress messages about attaching, detaching, etc.\n\
262 -r print relative timestamp\n\
263 -s strsize limit length of print strings to STRSIZE chars (default %d)\n\
264 -t print absolute timestamp\n\
265 -tt print absolute timestamp with usecs\n\
266 -T print time spent in each syscall\n\
267 -x print non-ascii strings in hex\n\
268 -xx print all strings in hex\n\
269 -y print paths associated with file descriptor arguments\n\
270 -yy print protocol specific information associated with socket file descriptors\n\
273 -c count time, calls, and errors for each syscall and report summary\n\
274 -C like -c but also print regular output\n\
275 -O overhead set overhead for tracing syscalls to OVERHEAD usecs\n\
276 -S sortby sort syscall counts by: time, calls, name, nothing (default %s)\n\
277 -w summarise syscall latency (default is system time)\n\
280 -e expr a qualifying expression: option=[!]all or option=[!]val1[,val2]...\n\
281 options: trace, abbrev, verbose, raw, signal, read, write, fault\n\
282 -P path trace accesses to path\n\
285 -b execve detach on execve syscall\n\
286 -D run tracer process as a detached grandchild, not as parent\n\
288 -ff follow forks with output into separate files\n\
290 1: no signals are blocked\n\
291 2: fatal signals are blocked while decoding syscall (default)\n\
292 3: fatal signals are always blocked (default if '-o FILE PROG')\n\
293 4: fatal signals and SIGTSTP (^Z) are always blocked\n\
294 (useful to make 'strace -o FILE PROG' not stop on ^Z)\n\
297 -E var remove var from the environment for command\n\
298 -E var=val put var=val in the environment for command\n\
299 -p pid trace process with process id PID, may be repeated\n\
300 -u username run command as username handling setuid and/or setgid\n\
303 -d enable debug output to stderr\n\
304 -v verbose mode: print unabbreviated argv, stat, termios, etc. args\n\
305 -h print help message\n\
308 /* ancient, no one should use it
309 -F -- attempt to follow vforks (deprecated, use -f)\n\
311 /* this is broken, so don't document it
312 -z -- print only succeeding syscalls\n\
314 , DEFAULT_ACOLUMN, DEFAULT_STRLEN, DEFAULT_SORTBY);
318 void ATTRIBUTE_NORETURN
321 if (strace_tracer_pid == getpid()) {
331 error_opt_arg(int opt, const char *arg)
333 error_msg_and_help("invalid -%c argument: '%s'", opt, arg);
336 static const char *ptrace_attach_cmd;
339 ptrace_attach_or_seize(int pid)
343 return ptrace_attach_cmd = "PTRACE_ATTACH",
344 ptrace(PTRACE_ATTACH, pid, 0L, 0L);
345 r = ptrace(PTRACE_SEIZE, pid, 0L, (unsigned long) ptrace_setoptions);
347 return ptrace_attach_cmd = "PTRACE_SEIZE", r;
348 r = ptrace(PTRACE_INTERRUPT, pid, 0L, 0L);
349 return ptrace_attach_cmd = "PTRACE_INTERRUPT", r;
353 * Used when we want to unblock stopped traced process.
354 * Should be only used with PTRACE_CONT, PTRACE_DETACH and PTRACE_SYSCALL.
355 * Returns 0 on success or if error was ESRCH
356 * (presumably process was killed while we talk to it).
357 * Otherwise prints error message and returns -1.
360 ptrace_restart(const unsigned int op, struct tcb *const tcp, unsigned int sig)
366 ptrace(op, tcp->pid, 0L, (unsigned long) sig);
386 * Why curcol != 0? Otherwise sometimes we get this:
388 * 10252 kill(10253, SIGKILL) = 0
389 * <ptrace(SYSCALL,10252):No such process>10253 ...next decode...
391 * 10252 died after we retrieved syscall exit data,
392 * but before we tried to restart it. Log looks ugly.
394 if (current_tcp && current_tcp->curcol != 0) {
395 tprintf(" <ptrace(%s):%s>\n", msg, strerror(err));
401 perror_msg("ptrace(PTRACE_%s,pid:%d,sig:%u)", msg, tcp->pid, sig);
406 set_cloexec_flag(int fd)
410 flags = fcntl(fd, F_GETFD);
412 /* Can happen only if fd is bad.
413 * Should never happen: if it does, we have a bug
414 * in the caller. Therefore we just abort
415 * instead of propagating the error.
417 perror_msg_and_die("fcntl(%d, F_GETFD)", fd);
420 newflags = flags | FD_CLOEXEC;
421 if (flags == newflags)
424 fcntl(fd, F_SETFD, newflags); /* never fails */
428 kill_save_errno(pid_t pid, int sig)
430 int saved_errno = errno;
432 (void) kill(pid, sig);
437 * When strace is setuid executable, we have to swap uids
438 * before and after filesystem and process management operations.
443 int euid = geteuid(), uid = getuid();
445 if (euid != uid && setreuid(euid, uid) < 0) {
446 perror_msg_and_die("setreuid");
451 strace_fopen(const char *path)
456 fp = fopen_stream(path, open_append ? "a" : "w");
458 perror_msg_and_die("Can't fopen '%s'", path);
460 set_cloexec_flag(fileno(fp));
464 static int popen_pid;
467 # define _PATH_BSHELL "/bin/sh"
471 * We cannot use standard popen(3) here because we have to distinguish
472 * popen child process from other processes we trace, and standard popen(3)
473 * does not export its child's pid.
476 strace_popen(const char *command)
484 perror_msg_and_die("pipe");
486 set_cloexec_flag(fds[1]); /* never fails */
490 perror_msg_and_die("vfork");
497 perror_msg_and_die("dup2");
500 execl(_PATH_BSHELL, "sh", "-c", command, NULL);
501 perror_msg_and_die("Can't execute '%s'", _PATH_BSHELL);
508 fp = fdopen(fds[1], "w");
510 perror_msg_and_die("fdopen");
515 outf_perror(const struct tcb * const tcp)
517 if (tcp->outf == stderr)
520 /* This is ugly, but we don't store separate file names */
522 perror_msg("%s.%u", outfname, tcp->pid);
524 perror_msg("%s", outfname);
527 ATTRIBUTE_FORMAT((printf, 1, 0))
529 tvprintf(const char *const fmt, va_list args)
532 int n = vfprintf(current_tcp->outf, fmt, args);
534 /* very unlikely due to vfprintf buffering */
535 outf_perror(current_tcp);
537 current_tcp->curcol += n;
542 tprintf(const char *fmt, ...)
550 #ifndef HAVE_FPUTS_UNLOCKED
551 # define fputs_unlocked fputs
555 tprints(const char *str)
558 int n = fputs_unlocked(str, current_tcp->outf);
560 current_tcp->curcol += strlen(str);
563 /* very unlikely due to fputs_unlocked buffering */
564 outf_perror(current_tcp);
569 tprints_comment(const char *const str)
572 tprintf(" /* %s */", str);
576 tprintf_comment(const char *fmt, ...)
590 flush_tcp_output(const struct tcb *const tcp)
592 if (fflush(tcp->outf))
600 current_tcp->curcol = 0;
601 flush_tcp_output(current_tcp);
604 printing_tcp->curcol = 0;
610 set_current_tcp(const struct tcb *tcp)
612 current_tcp = (struct tcb *) tcp;
614 /* Sync current_personality and stuff */
616 set_personality(current_tcp->currpers);
620 printleader(struct tcb *tcp)
622 /* If -ff, "previous tcb we printed" is always the same as current,
623 * because we have per-tcb output files.
629 set_current_tcp(printing_tcp);
630 if (printing_tcp->curcol != 0 && (followfork < 2 || printing_tcp == tcp)) {
632 * case 1: we have a shared log (i.e. not -ff), and last line
633 * wasn't finished (same or different tcb, doesn't matter).
634 * case 2: split log, we are the same tcb, but our last line
635 * didn't finish ("SIGKILL nuked us after syscall entry" etc).
637 tprints(" <unfinished ...>\n");
638 printing_tcp->curcol = 0;
643 set_current_tcp(tcp);
644 current_tcp->curcol = 0;
647 tprintf("%-5d ", tcp->pid);
648 else if (nprocs > 1 && !outfname)
649 tprintf("[pid %5u] ", tcp->pid);
653 clock_gettime(CLOCK_REALTIME, &ts);
656 tprintf("%lld.%06ld ",
657 (long long) ts.tv_sec, (long) ts.tv_nsec / 1000);
659 time_t local = ts.tv_sec;
660 char str[MAX(sizeof("HH:MM:SS"), sizeof(ts.tv_sec) * 3)];
661 struct tm *tm = localtime(&local);
664 strftime(str, sizeof(str), "%T", tm);
666 xsprintf(str, "%lld", (long long) local);
669 str, (long) ts.tv_nsec / 1000);
677 clock_gettime(CLOCK_MONOTONIC, &ts);
679 static struct timespec ots;
684 ts_sub(&dts, &ts, &ots);
687 tprintf("%s%6ld.%06ld%s ",
689 (long) dts.tv_sec, (long) dts.tv_nsec / 1000,
700 if (current_tcp->curcol < acolumn)
701 tprints(acolumn_spaces + current_tcp->curcol);
704 /* Should be only called directly *after successful attach* to a tracee.
705 * Otherwise, "strace -oFILE -ff -p<nonexistant_pid>"
706 * may create bogus empty FILE.<nonexistant_pid>, and then die.
709 after_successful_attach(struct tcb *tcp, const unsigned int flags)
711 tcp->flags |= TCB_ATTACHED | TCB_STARTUP | flags;
712 tcp->outf = shared_log; /* if not -ff mode, the same file is for all */
713 if (followfork >= 2) {
715 xsprintf(name, "%s.%u", outfname, tcp->pid);
716 tcp->outf = strace_fopen(name);
719 #ifdef ENABLE_STACKTRACE
720 if (stack_trace_enabled)
721 unwind_tcb_init(tcp);
728 /* Allocate some (more) TCBs (and expand the table).
729 We don't want to relocate the TCBs because our
730 callers have pointers and it would be a pain.
731 So tcbtab is a table of pointers. Since we never
732 free the TCBs, we allocate a single chunk of many. */
733 size_t old_tcbtabsize;
735 struct tcb **tcb_ptr;
737 old_tcbtabsize = tcbtabsize;
739 tcbtab = xgrowarray(tcbtab, &tcbtabsize, sizeof(tcbtab[0]));
740 newtcbs = xcalloc(tcbtabsize - old_tcbtabsize, sizeof(newtcbs[0]));
742 for (tcb_ptr = tcbtab + old_tcbtabsize;
743 tcb_ptr < tcbtab + tcbtabsize; tcb_ptr++, newtcbs++)
753 if (nprocs == tcbtabsize)
756 for (i = 0; i < tcbtabsize; i++) {
759 memset(tcp, 0, sizeof(*tcp));
761 #if SUPPORTED_PERSONALITIES > 1
762 tcp->currpers = current_personality;
765 debug_msg("new tcb for pid %d, active tcbs:%d",
770 error_msg_and_die("bug in alloctcb");
774 get_tcb_priv_data(const struct tcb *tcp)
776 return tcp->_priv_data;
780 set_tcb_priv_data(struct tcb *tcp, void *const priv_data,
781 void (*const free_priv_data)(void *))
786 tcp->_free_priv_data = free_priv_data;
787 tcp->_priv_data = priv_data;
793 free_tcb_priv_data(struct tcb *tcp)
795 if (tcp->_priv_data) {
796 if (tcp->_free_priv_data) {
797 tcp->_free_priv_data(tcp->_priv_data);
798 tcp->_free_priv_data = NULL;
800 tcp->_priv_data = NULL;
805 droptcb(struct tcb *tcp)
811 for (p = 0; p < SUPPORTED_PERSONALITIES; ++p)
812 free(tcp->inject_vec[p]);
814 free_tcb_priv_data(tcp);
816 #ifdef ENABLE_STACKTRACE
817 if (stack_trace_enabled)
821 mmap_cache_delete(tcp, __func__);
824 debug_msg("dropped tcb for pid %d, %d remain", tcp->pid, nprocs);
827 if (followfork >= 2) {
828 if (tcp->curcol != 0)
829 fprintf(tcp->outf, " <detached ...>\n");
832 if (printing_tcp == tcp && tcp->curcol != 0)
833 fprintf(tcp->outf, " <detached ...>\n");
834 flush_tcp_output(tcp);
838 if (current_tcp == tcp)
839 set_current_tcp(NULL);
840 if (printing_tcp == tcp)
843 memset(tcp, 0, sizeof(*tcp));
846 /* Detach traced process.
847 * Never call DETACH twice on the same process as both unattached and
848 * attached-unstopped processes give the same ESRCH. For unattached process we
849 * would SIGSTOP it and wait for its SIGSTOP notification forever.
852 detach(struct tcb *tcp)
858 * Linux wrongly insists the child be stopped
859 * before detaching. Arghh. We go through hoops
860 * to make a clean break of things.
863 if (!(tcp->flags & TCB_ATTACHED))
866 /* We attached but possibly didn't see the expected SIGSTOP.
867 * We must catch exactly one as otherwise the detached process
868 * would be left stopped (process state T).
870 if (tcp->flags & TCB_IGNORE_ONE_SIGSTOP)
873 error = ptrace(PTRACE_DETACH, tcp->pid, 0, 0);
875 /* On a clear day, you can see forever. */
878 if (errno != ESRCH) {
879 /* Shouldn't happen. */
880 perror_func_msg("ptrace(PTRACE_DETACH,%u)", tcp->pid);
883 /* ESRCH: process is either not stopped or doesn't exist. */
884 if (my_tkill(tcp->pid, 0) < 0) {
886 /* Shouldn't happen. */
887 perror_func_msg("tkill(%u,0)", tcp->pid);
888 /* else: process doesn't exist. */
891 /* Process is not stopped, need to stop it. */
894 * With SEIZE, tracee can be in group-stop already.
895 * In this state sending it another SIGSTOP does nothing.
896 * Need to use INTERRUPT.
897 * Testcase: trying to ^C a "strace -p <stopped_process>".
899 error = ptrace(PTRACE_INTERRUPT, tcp->pid, 0, 0);
903 perror_func_msg("ptrace(PTRACE_INTERRUPT,%u)", tcp->pid);
905 error = my_tkill(tcp->pid, SIGSTOP);
909 perror_func_msg("tkill(%u,SIGSTOP)", tcp->pid);
911 /* Either process doesn't exist, or some weird error. */
915 /* We end up here in three cases:
916 * 1. We sent PTRACE_INTERRUPT (use_seize case)
917 * 2. We sent SIGSTOP (!use_seize)
918 * 3. Attach SIGSTOP was already pending (TCB_IGNORE_ONE_SIGSTOP set)
922 if (waitpid(tcp->pid, &status, __WALL) < 0) {
926 * if (errno == ECHILD) break;
927 * ^^^ WRONG! We expect this PID to exist,
928 * and want to emit a message otherwise:
930 perror_func_msg("waitpid(%u)", tcp->pid);
933 if (!WIFSTOPPED(status)) {
935 * Tracee exited or was killed by signal.
936 * We shouldn't normally reach this place:
937 * we don't want to consume exit status.
938 * Consider "strace -p PID" being ^C-ed:
939 * we want merely to detach from PID.
941 * However, we _can_ end up here if tracee
946 sig = WSTOPSIG(status);
947 debug_msg("detach wait: event:%d sig:%d",
948 (unsigned) status >> 16, sig);
950 unsigned event = (unsigned)status >> 16;
951 if (event == PTRACE_EVENT_STOP /*&& sig == SIGTRAP*/) {
953 * sig == SIGTRAP: PTRACE_INTERRUPT stop.
954 * sig == other: process was already stopped
955 * with this stopping sig (see tests/detach-stopped).
956 * Looks like re-injecting this sig is not necessary
957 * in DETACH for the tracee to remain stopped.
962 * PTRACE_INTERRUPT is not guaranteed to produce
963 * the above event if other ptrace-stop is pending.
964 * See tests/detach-sleeping testcase:
965 * strace got SIGINT while tracee is sleeping.
966 * We sent PTRACE_INTERRUPT.
967 * We see syscall exit, not PTRACE_INTERRUPT stop.
968 * We won't get PTRACE_INTERRUPT stop
969 * if we would CONT now. Need to DETACH.
971 if (sig == syscall_trap_sig)
973 /* else: not sure in which case we can be here.
974 * Signal stop? Inject it while detaching.
976 ptrace_restart(PTRACE_DETACH, tcp, sig);
979 /* Note: this check has to be after use_seize check */
980 /* (else, in use_seize case SIGSTOP will be mistreated) */
981 if (sig == SIGSTOP) {
982 /* Detach, suppressing SIGSTOP */
983 ptrace_restart(PTRACE_DETACH, tcp, 0);
986 if (sig == syscall_trap_sig)
988 /* Can't detach just yet, may need to wait for SIGSTOP */
989 error = ptrace_restart(PTRACE_CONT, tcp, sig);
991 /* Should not happen.
992 * Note: ptrace_restart returns 0 on ESRCH, so it's not it.
993 * ptrace_restart already emitted error message.
1000 if (!qflag && (tcp->flags & TCB_ATTACHED))
1001 error_msg("Process %u detached", tcp->pid);
1007 process_opt_p_list(char *opt)
1011 * We accept -p PID,PID; -p "`pidof PROG`"; -p "`pgrep PROG`".
1012 * pidof uses space as delim, pgrep uses newline. :(
1015 char *delim = opt + strcspn(opt, "\n\t ,");
1019 pid = string_to_uint(opt);
1021 error_msg_and_die("Invalid process id: '%s'", opt);
1023 if (pid == strace_tracer_pid) {
1024 error_msg_and_die("I'm sorry, I can't let you do that, Dave.");
1035 attach_tcb(struct tcb *const tcp)
1037 if (ptrace_attach_or_seize(tcp->pid) < 0) {
1038 perror_msg("attach: ptrace(%s, %d)",
1039 ptrace_attach_cmd, tcp->pid);
1044 after_successful_attach(tcp, TCB_GRABBED | post_attach_sigstop);
1045 debug_msg("attach to pid %d (main) succeeded", tcp->pid);
1047 static const char task_path[] = "/proc/%d/task";
1048 char procdir[sizeof(task_path) + sizeof(int) * 3];
1050 unsigned int ntid = 0, nerr = 0;
1052 if (followfork && tcp->pid != strace_child &&
1053 xsprintf(procdir, task_path, tcp->pid) > 0 &&
1054 (dir = opendir(procdir)) != NULL) {
1057 while ((de = read_dir(dir)) != NULL) {
1058 if (de->d_fileno == 0)
1061 int tid = string_to_uint(de->d_name);
1062 if (tid <= 0 || tid == tcp->pid)
1066 if (ptrace_attach_or_seize(tid) < 0) {
1068 debug_perror_msg("attach: ptrace(%s, %d)",
1069 ptrace_attach_cmd, tid);
1073 after_successful_attach(alloctcb(tid),
1074 TCB_GRABBED | post_attach_sigstop);
1075 debug_msg("attach to pid %d succeeded", tid);
1083 error_msg("Process %u attached"
1085 tcp->pid, ntid - nerr + 1);
1087 error_msg("Process %u attached",
1093 startup_attach(void)
1095 pid_t parent_pid = strace_tracer_pid;
1099 if (daemonized_tracer) {
1102 perror_func_msg_and_die("fork");
1104 if (pid) { /* parent */
1106 * Wait for grandchild to attach to straced process
1107 * (grandparent). Grandchild SIGKILLs us after it attached.
1108 * Grandparent's wait() is unblocked by our death,
1109 * it proceeds to exec the straced program.
1112 _exit(0); /* paranoia */
1115 /* We will be the tracer process. Remember our new pid: */
1116 strace_tracer_pid = getpid();
1119 for (tcbi = 0; tcbi < tcbtabsize; tcbi++) {
1125 /* Is this a process we should attach to, but not yet attached? */
1126 if (tcp->flags & TCB_ATTACHED)
1127 continue; /* no, we already attached it */
1129 if (tcp->pid == parent_pid || tcp->pid == strace_tracer_pid) {
1131 perror_msg("attach: pid %d", tcp->pid);
1140 } /* for each tcbtab[] */
1142 if (daemonized_tracer) {
1144 * Make parent go away.
1145 * Also makes grandparent's wait() unblock.
1147 kill(parent_pid, SIGKILL);
1152 /* Stack-o-phobic exec helper, in the hope to work around
1153 * NOMMU + "daemonized tracer" difficulty.
1155 struct exec_params {
1161 struct sigaction child_sa;
1163 static struct exec_params params_for_tracee;
1165 static void ATTRIBUTE_NOINLINE ATTRIBUTE_NORETURN
1168 struct exec_params *params = ¶ms_for_tracee;
1170 if (params->fd_to_close >= 0)
1171 close(params->fd_to_close);
1172 if (!daemonized_tracer && !use_seize) {
1173 if (ptrace(PTRACE_TRACEME, 0L, 0L, 0L) < 0) {
1174 perror_msg_and_die("ptrace(PTRACE_TRACEME, ...)");
1178 if (username != NULL) {
1180 * It is important to set groups before we
1181 * lose privileges on setuid.
1183 if (initgroups(username, run_gid) < 0) {
1184 perror_msg_and_die("initgroups");
1186 if (setregid(run_gid, params->run_egid) < 0) {
1187 perror_msg_and_die("setregid");
1189 if (setreuid(run_uid, params->run_euid) < 0) {
1190 perror_msg_and_die("setreuid");
1192 } else if (geteuid() != 0)
1193 if (setreuid(run_uid, run_uid) < 0) {
1194 perror_msg_and_die("setreuid");
1197 if (!daemonized_tracer) {
1199 * Induce a ptrace stop. Tracer (our parent)
1200 * will resume us with PTRACE_SYSCALL and display
1201 * the immediately following execve syscall.
1202 * Can't do this on NOMMU systems, we are after
1203 * vfork: parent is blocked, stopping would deadlock.
1206 kill(getpid(), SIGSTOP);
1209 /* we depend on SIGCHLD set to SIG_DFL by init code */
1210 /* if it happens to be SIG_IGN'ed, wait won't block */
1215 if (params_for_tracee.child_sa.sa_handler != SIG_DFL)
1216 sigaction(SIGCHLD, ¶ms_for_tracee.child_sa, NULL);
1218 execv(params->pathname, params->argv);
1219 perror_msg_and_die("exec");
1223 * Open a dummy descriptor for use as a placeholder.
1224 * The descriptor is O_RDONLY with FD_CLOEXEC flag set.
1225 * A read attempt from such descriptor ends with EOF,
1226 * a write attempt is rejected with EBADF.
1229 open_dummy_desc(void)
1234 perror_func_msg_and_die("pipe");
1236 set_cloexec_flag(fds[0]);
1240 /* placeholder fds status for stdin and stdout */
1241 static bool fd_is_placeholder[2];
1244 * Ensure that all standard file descriptors are open by opening placeholder
1245 * file descriptors for those standard file descriptors that are not open.
1247 * The information which descriptors have been made open is saved
1248 * in fd_is_placeholder for later use.
1251 ensure_standard_fds_opened(void)
1255 while ((fd = open_dummy_desc()) <= 2) {
1258 fd_is_placeholder[fd] = true;
1266 * Redirect stdin and stdout unless they have been opened earlier
1267 * by ensure_standard_fds_opened as placeholders.
1270 redirect_standard_fds(void)
1275 * It might be a good idea to redirect stderr as well,
1276 * but we sometimes need to print error messages.
1278 for (i = 0; i <= 1; ++i) {
1279 if (!fd_is_placeholder[i]) {
1287 startup_child(char **argv)
1289 struct_stat statbuf;
1290 const char *filename;
1291 size_t filename_len;
1292 char pathname[PATH_MAX];
1297 filename_len = strlen(filename);
1299 if (filename_len > sizeof(pathname) - 1) {
1300 errno = ENAMETOOLONG;
1301 perror_msg_and_die("exec");
1303 if (strchr(filename, '/')) {
1304 strcpy(pathname, filename);
1306 #ifdef USE_DEBUGGING_EXEC
1308 * Debuggers customarily check the current directory
1309 * first regardless of the path but doing that gives
1310 * security geeks a panic attack.
1312 else if (stat_file(filename, &statbuf) == 0)
1313 strcpy(pathname, filename);
1314 #endif /* USE_DEBUGGING_EXEC */
1319 for (path = getenv("PATH"); path && *path; path += m) {
1320 const char *colon = strchr(path, ':');
1325 m = n = strlen(path);
1327 if (!getcwd(pathname, PATH_MAX))
1329 len = strlen(pathname);
1330 } else if (n > sizeof(pathname) - 1)
1333 strncpy(pathname, path, n);
1336 if (len && pathname[len - 1] != '/')
1337 pathname[len++] = '/';
1338 if (filename_len + len > sizeof(pathname) - 1)
1340 strcpy(pathname + len, filename);
1341 if (stat_file(pathname, &statbuf) == 0 &&
1342 /* Accept only regular files
1343 with some execute bits set.
1344 XXX not perfect, might still fail */
1345 S_ISREG(statbuf.st_mode) &&
1346 (statbuf.st_mode & 0111))
1349 if (!path || !*path)
1352 if (stat_file(pathname, &statbuf) < 0) {
1353 perror_msg_and_die("Can't stat '%s'", filename);
1356 params_for_tracee.fd_to_close = (shared_log != stderr) ? fileno(shared_log) : -1;
1357 params_for_tracee.run_euid = (statbuf.st_mode & S_ISUID) ? statbuf.st_uid : run_uid;
1358 params_for_tracee.run_egid = (statbuf.st_mode & S_ISGID) ? statbuf.st_gid : run_gid;
1359 params_for_tracee.argv = argv;
1361 * On NOMMU, can be safely freed only after execve in tracee.
1362 * It's hard to know when that happens, so we just leak it.
1364 params_for_tracee.pathname = NOMMU_SYSTEM ? xstrdup(pathname) : pathname;
1366 #if defined HAVE_PRCTL && defined PR_SET_PTRACER && defined PR_SET_PTRACER_ANY
1367 if (daemonized_tracer)
1368 prctl(PR_SET_PTRACER, PR_SET_PTRACER_ANY);
1373 perror_func_msg_and_die("fork");
1375 if ((pid != 0 && daemonized_tracer)
1376 || (pid == 0 && !daemonized_tracer)
1378 /* We are to become the tracee. Two cases:
1380 * not -D: we are child
1385 /* We are the tracer */
1387 if (!daemonized_tracer) {
1390 /* child did PTRACE_TRACEME, nothing to do in parent */
1392 if (!NOMMU_SYSTEM) {
1393 /* Wait until child stopped itself */
1395 while (waitpid(pid, &status, WSTOPPED) < 0) {
1398 perror_msg_and_die("waitpid");
1400 if (!WIFSTOPPED(status) || WSTOPSIG(status) != SIGSTOP) {
1401 kill_save_errno(pid, SIGKILL);
1402 perror_msg_and_die("Unexpected wait status %#x",
1406 /* Else: NOMMU case, we have no way to sync.
1407 * Just attach to it as soon as possible.
1408 * This means that we may miss a few first syscalls...
1411 if (ptrace_attach_or_seize(pid)) {
1412 kill_save_errno(pid, SIGKILL);
1413 perror_msg_and_die("attach: ptrace(%s, %d)",
1414 ptrace_attach_cmd, pid);
1419 tcp = alloctcb(pid);
1420 after_successful_attach(tcp, TCB_SKIP_DETACH_ON_FIRST_EXEC
1423 | post_attach_sigstop)));
1425 /* With -D, we are *child* here, the tracee is our parent. */
1426 strace_child = strace_tracer_pid;
1427 strace_tracer_pid = getpid();
1428 tcp = alloctcb(strace_child);
1429 tcp->flags |= TCB_SKIP_DETACH_ON_FIRST_EXEC | TCB_HIDE_LOG;
1431 * Attaching will be done later, by startup_attach.
1432 * Note: we don't do after_successful_attach() here either!
1435 /* NOMMU BUG! -D mode is active, we (child) return,
1436 * and we will scribble over parent's stack!
1437 * When parent later unpauses, it segfaults.
1440 * (1) by declaring exec_or_die() NORETURN,
1441 * hopefully compiler will just jump to it
1442 * instead of call (won't push anything to stack),
1443 * (2) by trying very hard in exec_or_die()
1444 * to not use any stack,
1445 * (3) having a really big (PATH_MAX) stack object
1446 * in this function, which creates a "buffer" between
1447 * child's and parent's stack pointers.
1448 * This may save us if (1) and (2) failed
1449 * and compiler decided to use stack in exec_or_die() anyway
1450 * (happens on i386 because of stack parameter passing).
1452 * A cleaner solution is to use makecontext + setcontext
1453 * to create a genuine separate stack and execute on it.
1457 * A case where straced process is part of a pipe:
1458 * { sleep 1; yes | head -n99999; } | strace -o/dev/null sh -c 'exec <&-; sleep 9'
1459 * If strace won't close its fd#0, closing it in tracee is not enough:
1460 * the pipe is still open, it has a reader. Thus, "head" will not get its
1461 * SIGPIPE at once, on the first write.
1463 * Preventing it by redirecting strace's stdin/out.
1464 * (Don't leave fds 0 and 1 closed, this is bad practice: future opens
1465 * will reuse them, unexpectedly making a newly opened object "stdin").
1467 redirect_standard_fds();
1471 test_ptrace_seize(void)
1475 /* Need fork for test. NOMMU has no forks */
1477 post_attach_sigstop = 0; /* this sets use_seize to 1 */
1483 perror_func_msg_and_die("fork");
1490 /* PTRACE_SEIZE, unlike ATTACH, doesn't force tracee to trap. After
1491 * attaching tracee continues to run unless a trap condition occurs.
1492 * PTRACE_SEIZE doesn't affect signal or group stop state.
1494 if (ptrace(PTRACE_SEIZE, pid, 0, 0) == 0) {
1495 post_attach_sigstop = 0; /* this sets use_seize to 1 */
1497 debug_msg("PTRACE_SEIZE doesn't work");
1503 int status, tracee_pid;
1506 tracee_pid = waitpid(pid, &status, 0);
1507 if (tracee_pid <= 0) {
1510 perror_func_msg_and_die("unexpected wait result %d",
1513 if (WIFSIGNALED(status))
1516 error_func_msg_and_die("unexpected wait status %#x", status);
1521 get_os_release(void)
1527 perror_msg_and_die("uname");
1528 /* u.release has this form: "3.2.9[-some-garbage]" */
1532 if (!(*p >= '0' && *p <= '9'))
1533 error_msg_and_die("Bad OS release string: '%s'", u.release);
1534 /* Note: this open-codes KERNEL_VERSION(): */
1535 rel = (rel << 8) | atoi(p);
1536 if (rel >= KERNEL_VERSION(1, 0, 0))
1538 while (*p >= '0' && *p <= '9')
1541 if (rel >= KERNEL_VERSION(0, 1, 0)) {
1542 /* "X.Y-something" means "X.Y.0" */
1546 error_msg_and_die("Bad OS release string: '%s'", u.release);
1554 set_sighandler(int signo, void (*sighandler)(int), struct sigaction *oldact)
1556 const struct sigaction sa = { .sa_handler = sighandler };
1557 sigaction(signo, &sa, oldact);
1561 * Initialization part of main() was eating much stack (~0.5k),
1562 * which was unused after init.
1563 * We can reuse it if we move init code into a separate function.
1565 * Don't want main() to inline us and defeat the reason
1566 * we have a separate function.
1568 static void ATTRIBUTE_NOINLINE
1569 init(int argc, char *argv[])
1574 if (!program_invocation_name || !*program_invocation_name) {
1575 static char name[] = "strace";
1576 program_invocation_name =
1577 (argc > 0 && argv[0] && *argv[0]) ? argv[0] : name;
1580 strace_tracer_pid = getpid();
1582 os_release = get_os_release();
1584 shared_log = stderr;
1585 set_sortby(DEFAULT_SORTBY);
1586 set_personality(DEFAULT_PERSONALITY);
1587 qualify("trace=all");
1588 qualify("abbrev=all");
1589 qualify("verbose=all");
1590 #if DEFAULT_QUAL_FLAGS != (QUAL_TRACE | QUAL_ABBREV | QUAL_VERBOSE)
1591 # error Bug in DEFAULT_QUAL_FLAGS
1593 qualify("signal=all");
1594 while ((c = getopt(argc, argv, "+"
1595 #ifdef ENABLE_STACKTRACE
1598 "a:Ab:cCdDe:E:fFhiI:o:O:p:P:qrs:S:tTu:vVwxX:yz")) != EOF) {
1601 acolumn = string_to_uint(optarg);
1603 error_opt_arg(c, optarg);
1609 if (strcmp(optarg, "execve") != 0)
1610 error_msg_and_die("Syscall '%s' for -b isn't supported",
1612 detach_on_execve = 1;
1615 if (cflag == CFLAG_BOTH) {
1616 error_msg_and_help("-c and -C are mutually exclusive");
1618 cflag = CFLAG_ONLY_STATS;
1621 if (cflag == CFLAG_ONLY_STATS) {
1622 error_msg_and_help("-c and -C are mutually exclusive");
1630 daemonized_tracer = 1;
1636 if (putenv(optarg) < 0)
1637 perror_msg_and_die("putenv");
1652 opt_intr = string_to_uint_upto(optarg, NUM_INTR_OPTS - 1);
1654 error_opt_arg(c, optarg);
1656 #ifdef ENABLE_STACKTRACE
1658 stack_trace_enabled = true;
1665 i = string_to_uint(optarg);
1667 error_opt_arg(c, optarg);
1671 process_opt_p_list(optarg);
1674 pathtrace_select(optarg);
1683 i = string_to_uint(optarg);
1684 if (i < 0 || (unsigned int) i > -1U / 4)
1685 error_opt_arg(c, optarg);
1701 qualify("abbrev=none");
1708 count_wallclock = 1;
1714 if (!strcmp(optarg, "raw"))
1715 xlat_verbosity = XLAT_STYLE_RAW;
1716 else if (!strcmp(optarg, "abbrev"))
1717 xlat_verbosity = XLAT_STYLE_ABBREV;
1718 else if (!strcmp(optarg, "verbose"))
1719 xlat_verbosity = XLAT_STYLE_VERBOSE;
1721 error_opt_arg(c, optarg);
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 ENABLE_STACKTRACE
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');
1781 acolumn_spaces = xmalloc(acolumn + 1);
1782 memset(acolumn_spaces, ' ', acolumn);
1783 acolumn_spaces[acolumn] = '\0';
1785 set_sighandler(SIGCHLD, SIG_DFL, ¶ms_for_tracee.child_sa);
1787 #ifdef ENABLE_STACKTRACE
1788 if (stack_trace_enabled)
1792 /* See if they want to run as another user. */
1793 if (username != NULL) {
1794 struct passwd *pent;
1796 if (getuid() != 0 || geteuid() != 0) {
1797 error_msg_and_die("You must be root to use the -u option");
1799 pent = getpwnam(username);
1801 error_msg_and_die("Cannot find user '%s'", username);
1803 run_uid = pent->pw_uid;
1804 run_gid = pent->pw_gid;
1811 ptrace_setoptions |= PTRACE_O_TRACECLONE |
1812 PTRACE_O_TRACEFORK |
1813 PTRACE_O_TRACEVFORK;
1814 debug_msg("ptrace_setoptions = %#x", ptrace_setoptions);
1815 test_ptrace_seize();
1818 * Is something weird with our stdin and/or stdout -
1819 * for example, may they be not open? In this case,
1820 * ensure that none of the future opens uses them.
1822 * This was seen in the wild when /proc/sys/kernel/core_pattern
1823 * was set to "|/bin/strace -o/tmp/LOG PROG":
1824 * kernel runs coredump helper with fd#0 open but fd#1 closed (!),
1825 * therefore LOG gets opened to fd#1, and fd#1 is closed by
1826 * "don't hold up stdin/out open" code soon after.
1828 ensure_standard_fds_opened();
1830 /* Check if they want to redirect the output. */
1832 /* See if they want to pipe the output. */
1833 if (outfname[0] == '|' || outfname[0] == '!') {
1835 * We can't do the <outfname>.PID funny business
1836 * when using popen, so prohibit it.
1838 if (followfork >= 2)
1839 error_msg_and_help("piping the output and -ff "
1840 "are mutually exclusive");
1841 shared_log = strace_popen(outfname + 1);
1842 } else if (followfork < 2) {
1843 shared_log = strace_fopen(outfname);
1844 } else if (strlen(outfname) >= PATH_MAX - sizeof(int) * 3) {
1845 errno = ENAMETOOLONG;
1846 perror_msg_and_die("%s", outfname);
1849 /* -ff without -o FILE is the same as single -f */
1850 if (followfork >= 2)
1854 if (!outfname || outfname[0] == '|' || outfname[0] == '!') {
1855 setvbuf(shared_log, NULL, _IOLBF, 0);
1859 * argv[0] -pPID -oFILE Default interactive setting
1860 * yes * 0 INTR_WHILE_WAIT
1861 * no 1 0 INTR_WHILE_WAIT
1862 * yes * 1 INTR_NEVER
1863 * no 1 1 INTR_WHILE_WAIT
1866 if (outfname && argc) {
1868 opt_intr = INTR_NEVER;
1873 opt_intr = INTR_WHILE_WAIT;
1876 * startup_child() must be called before the signal handlers get
1877 * installed below as they are inherited into the spawned process.
1878 * Also we do not need to be protected by them as during interruption
1879 * in the startup_child() mode we kill the spawned process anyway.
1882 startup_child(argv);
1885 set_sighandler(SIGTTOU, SIG_IGN, NULL);
1886 set_sighandler(SIGTTIN, SIG_IGN, NULL);
1887 if (opt_intr != INTR_ANYWHERE) {
1888 if (opt_intr == INTR_BLOCK_TSTP_TOO)
1889 set_sighandler(SIGTSTP, SIG_IGN, NULL);
1891 * In interactive mode (if no -o OUTFILE, or -p PID is used),
1892 * fatal signals are handled asynchronously and acted
1893 * when waiting for process state changes.
1894 * In non-interactive mode these signals are ignored.
1896 set_sighandler(SIGHUP, interactive ? interrupt : SIG_IGN, NULL);
1897 set_sighandler(SIGINT, interactive ? interrupt : SIG_IGN, NULL);
1898 set_sighandler(SIGQUIT, interactive ? interrupt : SIG_IGN, NULL);
1899 set_sighandler(SIGPIPE, interactive ? interrupt : SIG_IGN, NULL);
1900 set_sighandler(SIGTERM, interactive ? interrupt : SIG_IGN, NULL);
1903 sigemptyset(&timer_set);
1904 sigaddset(&timer_set, SIGALRM);
1905 sigprocmask(SIG_BLOCK, &timer_set, NULL);
1906 set_sighandler(SIGALRM, timer_sighandler, NULL);
1908 if (nprocs != 0 || daemonized_tracer)
1911 /* Do we want pids printed in our -o OUTFILE?
1912 * -ff: no (every pid has its own file); or
1913 * -f: yes (there can be more pids in the future); or
1914 * -p PID1,PID2: yes (there are already more than one pid)
1916 print_pid_pfx = (outfname && followfork < 2 && (followfork == 1 || nprocs > 1));
1920 pid2tcb(const int pid)
1925 #define PID2TCB_CACHE_SIZE 1024U
1926 #define PID2TCB_CACHE_MASK (PID2TCB_CACHE_SIZE - 1)
1928 static struct tcb *pid2tcb_cache[PID2TCB_CACHE_SIZE];
1929 struct tcb **const ptcp = &pid2tcb_cache[pid & PID2TCB_CACHE_MASK];
1930 struct tcb *tcp = *ptcp;
1932 if (tcp && tcp->pid == pid)
1935 for (unsigned int i = 0; i < tcbtabsize; ++i) {
1937 if (tcp->pid == pid)
1951 /* 'interrupted' is a volatile object, fetch it only once */
1952 fatal_sig = interrupted;
1954 fatal_sig = SIGTERM;
1956 for (i = 0; i < tcbtabsize; i++) {
1960 debug_func_msg("looking at pid %u", tcp->pid);
1961 if (tcp->pid == strace_child) {
1962 kill(tcp->pid, SIGCONT);
1963 kill(tcp->pid, fatal_sig);
1968 call_summary(shared_log);
1978 print_debug_info(const int pid, int status)
1980 const unsigned int event = (unsigned int) status >> 16;
1981 char buf[sizeof("WIFEXITED,exitcode=%u") + sizeof(int)*3 /*paranoia:*/ + 16];
1982 char evbuf[sizeof(",EVENT_VFORK_DONE (%u)") + sizeof(int)*3 /*paranoia:*/ + 16];
1985 if (WIFSIGNALED(status))
1987 xsprintf(buf, "WIFSIGNALED,%ssig=%s",
1988 WCOREDUMP(status) ? "core," : "",
1989 signame(WTERMSIG(status)));
1991 xsprintf(buf, "WIFSIGNALED,sig=%s",
1992 signame(WTERMSIG(status)));
1994 if (WIFEXITED(status))
1995 xsprintf(buf, "WIFEXITED,exitcode=%u", WEXITSTATUS(status));
1996 if (WIFSTOPPED(status))
1997 xsprintf(buf, "WIFSTOPPED,sig=%s", signame(WSTOPSIG(status)));
2000 static const char *const event_names[] = {
2001 [PTRACE_EVENT_CLONE] = "CLONE",
2002 [PTRACE_EVENT_FORK] = "FORK",
2003 [PTRACE_EVENT_VFORK] = "VFORK",
2004 [PTRACE_EVENT_VFORK_DONE] = "VFORK_DONE",
2005 [PTRACE_EVENT_EXEC] = "EXEC",
2006 [PTRACE_EVENT_EXIT] = "EXIT",
2007 /* [PTRACE_EVENT_STOP (=128)] would make biggish array */
2009 const char *e = "??";
2010 if (event < ARRAY_SIZE(event_names))
2011 e = event_names[event];
2012 else if (event == PTRACE_EVENT_STOP)
2014 xsprintf(evbuf, ",EVENT_%s (%u)", e, event);
2016 error_msg("[wait(0x%06x) = %u] %s%s", status, pid, buf, evbuf);
2020 maybe_allocate_tcb(const int pid, int status)
2022 if (!WIFSTOPPED(status)) {
2023 if (detach_on_execve && pid == strace_child) {
2024 /* example: strace -bexecve sh -c 'exec true' */
2029 * This can happen if we inherited an unknown child.
2030 * Example: (sleep 1 & exec strace true)
2032 error_msg("Exit of unknown pid %u ignored", pid);
2036 /* We assume it's a fork/vfork/clone child */
2037 struct tcb *tcp = alloctcb(pid);
2038 after_successful_attach(tcp, post_attach_sigstop);
2040 error_msg("Process %d attached", pid);
2044 * This can happen if a clone call misused CLONE_PTRACE itself.
2046 * There used to be a dance around possible re-injection of
2047 * WSTOPSIG(status), but it was later removed as the only
2048 * observable stop here is the initial ptrace-stop.
2050 ptrace(PTRACE_DETACH, pid, NULL, 0L);
2051 error_msg("Detached unknown pid %d", pid);
2057 maybe_switch_tcbs(struct tcb *tcp, const int pid)
2060 struct tcb *execve_thread;
2063 if (ptrace(PTRACE_GETEVENTMSG, pid, NULL, &old_pid) < 0)
2065 /* Avoid truncation in pid2tcb() param passing */
2066 if (old_pid <= 0 || old_pid == pid)
2068 if ((unsigned long) old_pid > UINT_MAX)
2070 execve_thread = pid2tcb(old_pid);
2071 /* It should be !NULL, but I feel paranoid */
2075 if (execve_thread->curcol != 0) {
2077 * One case we are here is -ff:
2078 * try "strace -oLOG -ff test/threaded_execve"
2080 fprintf(execve_thread->outf, " <pid changed to %d ...>\n", pid);
2081 /*execve_thread->curcol = 0; - no need, see code below */
2083 /* Swap output FILEs (needed for -ff) */
2084 fp = execve_thread->outf;
2085 execve_thread->outf = tcp->outf;
2087 /* And their column positions */
2088 execve_thread->curcol = tcp->curcol;
2090 /* Drop leader, but close execve'd thread outfile (if -ff) */
2092 /* Switch to the thread, reusing leader's outfile and pid */
2093 tcp = execve_thread;
2095 if (cflag != CFLAG_ONLY_STATS) {
2097 tprintf("+++ superseded by execve in pid %lu +++\n", old_pid);
2099 tcp->flags |= TCB_REPRINT;
2106 print_signalled(struct tcb *tcp, const int pid, int status)
2108 if (pid == strace_child) {
2109 exit_code = 0x100 | WTERMSIG(status);
2113 if (cflag != CFLAG_ONLY_STATS
2114 && is_number_in_set(WTERMSIG(status), signal_set)) {
2117 tprintf("+++ killed by %s %s+++\n",
2118 signame(WTERMSIG(status)),
2119 WCOREDUMP(status) ? "(core dumped) " : "");
2121 tprintf("+++ killed by %s +++\n",
2122 signame(WTERMSIG(status)));
2129 print_exited(struct tcb *tcp, const int pid, int status)
2131 if (pid == strace_child) {
2132 exit_code = WEXITSTATUS(status);
2136 if (cflag != CFLAG_ONLY_STATS &&
2139 tprintf("+++ exited with %d +++\n", WEXITSTATUS(status));
2145 print_stopped(struct tcb *tcp, const siginfo_t *si, const unsigned int sig)
2147 if (cflag != CFLAG_ONLY_STATS
2149 && is_number_in_set(sig, signal_set)) {
2152 tprintf("--- %s ", signame(sig));
2156 tprintf("--- stopped by %s ---\n", signame(sig));
2162 startup_tcb(struct tcb *tcp)
2164 debug_msg("pid %d has TCB_STARTUP, initializing it", tcp->pid);
2166 tcp->flags &= ~TCB_STARTUP;
2169 debug_msg("setting opts 0x%x on pid %d",
2170 ptrace_setoptions, tcp->pid);
2171 if (ptrace(PTRACE_SETOPTIONS, tcp->pid, NULL, ptrace_setoptions) < 0) {
2172 if (errno != ESRCH) {
2173 /* Should never happen, really */
2174 perror_msg_and_die("PTRACE_SETOPTIONS");
2179 if ((tcp->flags & TCB_GRABBED) && (get_scno(tcp) == 1))
2180 tcp->s_prev_ent = tcp->s_ent;
2184 print_event_exit(struct tcb *tcp)
2186 if (entering(tcp) || filtered(tcp) || hide_log(tcp)
2187 || cflag == CFLAG_ONLY_STATS) {
2191 if (followfork < 2 && printing_tcp && printing_tcp != tcp
2192 && printing_tcp->curcol != 0) {
2193 set_current_tcp(printing_tcp);
2194 tprints(" <unfinished ...>\n");
2195 flush_tcp_output(printing_tcp);
2196 printing_tcp->curcol = 0;
2197 set_current_tcp(tcp);
2200 if ((followfork < 2 && printing_tcp != tcp)
2201 || (tcp->flags & TCB_REPRINT)) {
2202 tcp->flags &= ~TCB_REPRINT;
2204 tprintf("<... %s resumed>", tcp->s_ent->sys_name);
2207 if (!(tcp->sys_func_rval & RVAL_DECODED)) {
2209 * The decoder has probably decided to print something
2210 * on exiting syscall which is not going to happen.
2212 tprints(" <unfinished ...>");
2222 static enum trace_event
2223 next_event(int *pstatus, siginfo_t *si)
2234 * Used to exit simply when nprocs hits zero, but in this testcase:
2235 * int main(void) { _exit(!!fork()); }
2236 * under strace -f, parent sometimes (rarely) manages
2237 * to exit before we see the first stop of the child,
2238 * and we are losing track of it:
2239 * 19923 clone(...) = 19924
2240 * 19923 exit_group(1) = ?
2241 * 19923 +++ exited with 1 +++
2242 * Exiting only when wait() returns ECHILD works better.
2244 if (popen_pid != 0) {
2245 /* However, if -o|logger is in use, we can't do that.
2246 * Can work around that by double-forking the logger,
2247 * but that loses the ability to wait for its completion
2248 * on exit. Oh well...
2254 const bool unblock_delay_timer = is_delay_timer_armed();
2257 * The window of opportunity to handle expirations
2258 * of the delay timer opens here.
2260 * Unblock the signal handler for the delay timer
2261 * iff the delay timer is already created.
2263 if (unblock_delay_timer)
2264 sigprocmask(SIG_UNBLOCK, &timer_set, NULL);
2267 * If the delay timer has expired, then its expiration
2268 * has been handled already by the signal handler.
2270 * If the delay timer expires during wait4(),
2271 * then the system call will be interrupted and
2272 * the expiration will be handled by the signal handler.
2274 pid = wait4(-1, pstatus, __WALL, (cflag ? &ru : NULL));
2275 const int wait_errno = errno;
2278 * The window of opportunity to handle expirations
2279 * of the delay timer closes here.
2281 * Block the signal handler for the delay timer
2282 * iff it was unblocked earlier.
2284 if (unblock_delay_timer) {
2285 sigprocmask(SIG_BLOCK, &timer_set, NULL);
2292 if (wait_errno == EINTR)
2294 if (nprocs == 0 && wait_errno == ECHILD)
2297 * If nprocs > 0, ECHILD is not expected,
2298 * treat it as any other error here:
2301 perror_msg_and_die("wait4(__WALL)");
2306 if (pid == popen_pid) {
2307 if (!WIFSTOPPED(status))
2313 print_debug_info(pid, status);
2315 /* Look up 'pid' in our table. */
2319 tcp = maybe_allocate_tcb(pid, status);
2326 /* Set current output file */
2327 set_current_tcp(tcp);
2330 struct timespec stime = {
2331 .tv_sec = ru.ru_stime.tv_sec,
2332 .tv_nsec = ru.ru_stime.tv_usec * 1000
2334 ts_sub(&tcp->dtime, &stime, &tcp->stime);
2338 if (WIFSIGNALED(status))
2339 return TE_SIGNALLED;
2341 if (WIFEXITED(status))
2345 * As WCONTINUED flag has not been specified to wait4,
2346 * it cannot be WIFCONTINUED(status), so the only case
2347 * that remains is WIFSTOPPED(status).
2350 /* Is this the very first time we see this tracee stopped? */
2351 if (tcp->flags & TCB_STARTUP)
2354 const unsigned int sig = WSTOPSIG(status);
2355 const unsigned int event = (unsigned int) status >> 16;
2360 * Is this post-attach SIGSTOP?
2361 * Interestingly, the process may stop
2362 * with STOPSIG equal to some other signal
2363 * than SIGSTOP if we happened to attach
2364 * just before the process takes a signal.
2366 if (sig == SIGSTOP && (tcp->flags & TCB_IGNORE_ONE_SIGSTOP)) {
2367 debug_func_msg("ignored SIGSTOP on pid %d", tcp->pid);
2368 tcp->flags &= ~TCB_IGNORE_ONE_SIGSTOP;
2370 } else if (sig == syscall_trap_sig) {
2371 return TE_SYSCALL_STOP;
2373 *si = (siginfo_t) {};
2375 * True if tracee is stopped by signal
2376 * (as opposed to "tracee received signal").
2377 * TODO: shouldn't we check for errno == EINVAL too?
2378 * We can get ESRCH instead, you know...
2380 bool stopped = ptrace(PTRACE_GETSIGINFO, pid, 0, si) < 0;
2381 return stopped ? TE_GROUP_STOP : TE_SIGNAL_DELIVERY_STOP;
2384 case PTRACE_EVENT_STOP:
2386 * PTRACE_INTERRUPT-stop or group-stop.
2387 * PTRACE_INTERRUPT-stop has sig == SIGTRAP here.
2394 return TE_GROUP_STOP;
2397 case PTRACE_EVENT_EXEC:
2398 return TE_STOP_BEFORE_EXECVE;
2399 case PTRACE_EVENT_EXIT:
2400 return TE_STOP_BEFORE_EXIT;
2407 trace_syscall(struct tcb *tcp, unsigned int *sig)
2409 if (entering(tcp)) {
2410 int res = syscall_entering_decode(tcp);
2415 res = syscall_entering_trace(tcp, sig);
2417 syscall_entering_finish(tcp, res);
2420 struct timespec ts = {};
2421 int res = syscall_exiting_decode(tcp, &ts);
2423 res = syscall_exiting_trace(tcp, &ts, res);
2425 syscall_exiting_finish(tcp);
2430 /* Returns true iff the main trace loop has to continue. */
2432 dispatch_event(enum trace_event ret, int *pstatus, siginfo_t *si)
2434 unsigned int restart_op = PTRACE_SYSCALL;
2435 unsigned int restart_sig = 0;
2447 case TE_SYSCALL_STOP:
2448 if (trace_syscall(current_tcp, &restart_sig) < 0) {
2450 * ptrace() failed in trace_syscall().
2451 * Likely a result of process disappearing mid-flight.
2452 * Observed case: exit_group() or SIGKILL terminating
2453 * all processes in thread group.
2454 * We assume that ptrace error was caused by process death.
2455 * We used to detach(current_tcp) here, but since we no
2456 * longer implement "detach before death" policy/hack,
2457 * we can let this process to report its death to us
2458 * normally, via WIFEXITED or WIFSIGNALED wait status.
2464 case TE_SIGNAL_DELIVERY_STOP:
2465 restart_sig = WSTOPSIG(*pstatus);
2466 print_stopped(current_tcp, si, restart_sig);
2470 print_signalled(current_tcp, current_tcp->pid, *pstatus);
2471 droptcb(current_tcp);
2475 restart_sig = WSTOPSIG(*pstatus);
2476 print_stopped(current_tcp, NULL, restart_sig);
2479 * This ends ptrace-stop, but does *not* end group-stop.
2480 * This makes stopping signals work properly on straced
2481 * process (that is, process really stops. It used to
2484 restart_op = PTRACE_LISTEN;
2490 print_exited(current_tcp, current_tcp->pid, *pstatus);
2491 droptcb(current_tcp);
2494 case TE_STOP_BEFORE_EXECVE:
2496 * Check that we are inside syscall now (next event after
2497 * PTRACE_EVENT_EXEC should be for syscall exiting). If it is
2498 * not the case, we might have a situation when we attach to a
2499 * process and the first thing we see is a PTRACE_EVENT_EXEC
2500 * and all the following syscall state tracking is screwed up
2503 if (entering(current_tcp)) {
2506 error_msg("Stray PTRACE_EVENT_EXEC from pid %d"
2507 ", trying to recover...",
2510 current_tcp->flags |= TCB_RECOVERING;
2511 ret = trace_syscall(current_tcp, &restart_sig);
2512 current_tcp->flags &= ~TCB_RECOVERING;
2515 /* The reason is described in TE_SYSCALL_STOP */
2521 * Under Linux, execve changes pid to thread leader's pid,
2522 * and we see this changed pid on EVENT_EXEC and later,
2523 * execve sysexit. Leader "disappears" without exit
2524 * notification. Let user know that, drop leader's tcb,
2525 * and fix up pid in execve thread's tcb.
2526 * Effectively, execve thread's tcb replaces leader's tcb.
2528 * BTW, leader is 'stuck undead' (doesn't report WIFEXITED
2529 * on exit syscall) in multithreaded programs exactly
2530 * in order to handle this case.
2532 * PTRACE_GETEVENTMSG returns old pid starting from Linux 3.0.
2533 * On 2.6 and earlier, it can return garbage.
2535 if (os_release >= KERNEL_VERSION(3, 0, 0))
2536 set_current_tcp(maybe_switch_tcbs(current_tcp,
2539 if (detach_on_execve) {
2540 if (current_tcp->flags & TCB_SKIP_DETACH_ON_FIRST_EXEC) {
2541 current_tcp->flags &= ~TCB_SKIP_DETACH_ON_FIRST_EXEC;
2543 detach(current_tcp); /* do "-b execve" thingy */
2549 case TE_STOP_BEFORE_EXIT:
2550 print_event_exit(current_tcp);
2554 /* We handled quick cases, we are permitted to interrupt now. */
2558 /* If the process is being delayed, do not ptrace_restart just yet */
2559 if (syscall_delayed(current_tcp))
2562 if (ptrace_restart(restart_op, current_tcp, restart_sig) < 0) {
2563 /* Note: ptrace_restart emitted error message */
2571 restart_delayed_tcb(struct tcb *const tcp)
2573 debug_func_msg("pid %d", tcp->pid);
2575 tcp->flags &= ~TCB_DELAYED;
2577 struct tcb *const prev_tcp = current_tcp;
2579 bool ret = dispatch_event(TE_RESTART, NULL, NULL);
2580 current_tcp = prev_tcp;
2586 restart_delayed_tcbs(void)
2588 struct tcb *tcp_next = NULL;
2589 struct timespec ts_now;
2591 clock_gettime(CLOCK_MONOTONIC, &ts_now);
2593 for (size_t i = 0; i < tcbtabsize; i++) {
2594 struct tcb *tcp = tcbtab[i];
2596 if (tcp->pid && syscall_delayed(tcp)) {
2597 if (ts_cmp(&ts_now, &tcp->delay_expiration_time) > 0) {
2598 if (!restart_delayed_tcb(tcp))
2601 /* Check whether this tcb is the next. */
2603 ts_cmp(&tcp_next->delay_expiration_time,
2604 &tcp->delay_expiration_time) > 0) {
2612 arm_delay_timer(tcp_next);
2618 * As this signal handler does a lot of work that is not suitable
2619 * for signal handlers, extra care must be taken to ensure that
2620 * it is enabled only in those places where it's safe.
2623 timer_sighandler(int sig)
2625 delay_timer_expired();
2630 int saved_errno = errno;
2632 if (!restart_delayed_tcbs())
2635 errno = saved_errno;
2638 #ifdef ENABLE_COVERAGE_GCOV
2639 extern void __gcov_flush(void);
2642 static void ATTRIBUTE_NORETURN
2647 if (shared_log != stderr)
2650 while (waitpid(popen_pid, NULL, 0) < 0 && errno == EINTR)
2653 if (exit_code > 0xff) {
2654 /* Avoid potential core file clobbering. */
2655 struct_rlimit rlim = {0, 0};
2656 set_rlimit(RLIMIT_CORE, &rlim);
2658 /* Child was killed by a signal, mimic that. */
2660 signal(exit_code, SIG_DFL);
2661 #ifdef ENABLE_COVERAGE_GCOV
2666 /* Unblock the signal. */
2669 sigaddset(&mask, exit_code);
2670 #ifdef ENABLE_COVERAGE_GCOV
2673 sigprocmask(SIG_UNBLOCK, &mask, NULL);
2675 /* Paranoia - what if this signal is not fatal?
2676 Exit with 128 + signo then. */
2683 main(int argc, char *argv[])
2685 setlocale(LC_ALL, "");
2688 exit_code = !nprocs;
2692 while (dispatch_event(next_event(&status, &si), &status, &si))