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;
125 static int post_attach_sigstop = TCB_IGNORE_ONE_SIGSTOP;
126 #define use_seize (post_attach_sigstop == 0)
128 /* Sometimes we want to print only succeeding syscalls. */
129 bool not_failing_only;
131 /* Show path associated with fd arguments */
132 unsigned int show_fd_path;
134 static bool detach_on_execve;
136 static int exit_code;
137 static int strace_child;
138 static int strace_tracer_pid;
140 static const char *username;
141 static uid_t run_uid;
142 static gid_t run_gid;
144 unsigned int max_strlen = DEFAULT_STRLEN;
145 static int acolumn = DEFAULT_ACOLUMN;
146 static char *acolumn_spaces;
148 static const char *outfname;
149 /* If -ff, points to stderr. Else, it's our common output log */
150 static FILE *shared_log;
152 struct tcb *printing_tcp;
153 static struct tcb *current_tcp;
155 static struct tcb **tcbtab;
156 static unsigned int nprocs;
157 static size_t tcbtabsize;
159 #ifndef HAVE_PROGRAM_INVOCATION_NAME
160 char *program_invocation_name;
163 unsigned os_release; /* generated from uname()'s u.release */
165 static void detach(struct tcb *tcp);
166 static void cleanup(void);
167 static void interrupt(int sig);
168 static sigset_t start_set, blocked_set;
170 #ifdef HAVE_SIG_ATOMIC_T
171 static volatile sig_atomic_t interrupted;
173 static volatile int interrupted;
176 #ifndef HAVE_STRERROR
178 #if !HAVE_DECL_SYS_ERRLIST
180 extern char *sys_errlist[];
186 static char buf[sizeof("Unknown error %d") + sizeof(int)*3];
188 if (err_no < 1 || err_no >= sys_nerr) {
189 xsprintf(buf, "Unknown error %d", err_no);
192 return sys_errlist[err_no];
195 #endif /* HAVE_STERRROR */
200 static const char features[] =
203 #endif /* USE_LIBUNWIND */
206 #endif /* USE_DEMANGLE */
207 #if SUPPORTED_PERSONALITIES > 1
208 # if defined HAVE_M32_MPERS
213 #endif /* SUPPORTED_PERSONALITIES > 1 */
214 #if SUPPORTED_PERSONALITIES > 2
215 # if defined HAVE_MX32_MPERS
220 #endif /* SUPPORTED_PERSONALITIES > 2 */
223 printf("%s -- version %s\n"
224 "Copyright (c) 1991-%s The strace developers <%s>.\n"
225 "This is free software; see the source for copying conditions. There is NO\n"
226 "warranty; not even for MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.\n",
227 PACKAGE_NAME, PACKAGE_VERSION, COPYRIGHT_YEAR, PACKAGE_URL);
228 printf("\nOptional features enabled:%s\n",
229 features[0] ? features : " (none)");
236 usage: strace [-CdffhiqrtttTvVwxxy] [-I n] [-e expr]...\n\
237 [-a column] [-o file] [-s strsize] [-P path]...\n\
238 -p pid... / [-D] [-E var=val]... [-u username] PROG [ARGS]\n\
239 or: strace -c[dfw] [-I n] [-e expr]... [-O overhead] [-S sortby]\n\
240 -p pid... / [-D] [-E var=val]... [-u username] PROG [ARGS]\n\
243 -a column alignment COLUMN for printing syscall results (default %d)\n\
244 -i print instruction pointer at time of syscall\n\
248 -k obtain stack trace between each syscall (experimental)\n\
252 -o file send trace output to FILE instead of stderr\n\
253 -q suppress messages about attaching, detaching, etc.\n\
254 -r print relative timestamp\n\
255 -s strsize limit length of print strings to STRSIZE chars (default %d)\n\
256 -t print absolute timestamp\n\
257 -tt print absolute timestamp with usecs\n\
258 -T print time spent in each syscall\n\
259 -x print non-ascii strings in hex\n\
260 -xx print all strings in hex\n\
261 -y print paths associated with file descriptor arguments\n\
262 -yy print protocol specific information associated with socket file descriptors\n\
265 -c count time, calls, and errors for each syscall and report summary\n\
266 -C like -c but also print regular output\n\
267 -O overhead set overhead for tracing syscalls to OVERHEAD usecs\n\
268 -S sortby sort syscall counts by: time, calls, name, nothing (default %s)\n\
269 -w summarise syscall latency (default is system time)\n\
272 -e expr a qualifying expression: option=[!]all or option=[!]val1[,val2]...\n\
273 options: trace, abbrev, verbose, raw, signal, read, write, fault\n\
274 -P path trace accesses to path\n\
277 -b execve detach on execve syscall\n\
278 -D run tracer process as a detached grandchild, not as parent\n\
280 -ff follow forks with output into separate files\n\
282 1: no signals are blocked\n\
283 2: fatal signals are blocked while decoding syscall (default)\n\
284 3: fatal signals are always blocked (default if '-o FILE PROG')\n\
285 4: fatal signals and SIGTSTP (^Z) are always blocked\n\
286 (useful to make 'strace -o FILE PROG' not stop on ^Z)\n\
289 -E var remove var from the environment for command\n\
290 -E var=val put var=val in the environment for command\n\
291 -p pid trace process with process id PID, may be repeated\n\
292 -u username run command as username handling setuid and/or setgid\n\
295 -d enable debug output to stderr\n\
296 -v verbose mode: print unabbreviated argv, stat, termios, etc. args\n\
297 -h print help message\n\
300 /* ancient, no one should use it
301 -F -- attempt to follow vforks (deprecated, use -f)\n\
303 /* this is broken, so don't document it
304 -z -- print only succeeding syscalls\n\
306 , DEFAULT_ACOLUMN, DEFAULT_STRLEN, DEFAULT_SORTBY);
310 void ATTRIBUTE_NORETURN
313 if (strace_tracer_pid == getpid()) {
323 error_opt_arg(int opt, const char *arg)
325 error_msg_and_help("invalid -%c argument: '%s'", opt, arg);
328 static const char *ptrace_attach_cmd;
331 ptrace_attach_or_seize(int pid)
335 return ptrace_attach_cmd = "PTRACE_ATTACH",
336 ptrace(PTRACE_ATTACH, pid, 0L, 0L);
337 r = ptrace(PTRACE_SEIZE, pid, 0L, (unsigned long) ptrace_setoptions);
339 return ptrace_attach_cmd = "PTRACE_SEIZE", r;
340 r = ptrace(PTRACE_INTERRUPT, pid, 0L, 0L);
341 return ptrace_attach_cmd = "PTRACE_INTERRUPT", r;
345 * Used when we want to unblock stopped traced process.
346 * Should be only used with PTRACE_CONT, PTRACE_DETACH and PTRACE_SYSCALL.
347 * Returns 0 on success or if error was ESRCH
348 * (presumably process was killed while we talk to it).
349 * Otherwise prints error message and returns -1.
352 ptrace_restart(const unsigned int op, struct tcb *const tcp, unsigned int sig)
358 ptrace(op, tcp->pid, 0L, (unsigned long) sig);
378 * Why curcol != 0? Otherwise sometimes we get this:
380 * 10252 kill(10253, SIGKILL) = 0
381 * <ptrace(SYSCALL,10252):No such process>10253 ...next decode...
383 * 10252 died after we retrieved syscall exit data,
384 * but before we tried to restart it. Log looks ugly.
386 if (current_tcp && current_tcp->curcol != 0) {
387 tprintf(" <ptrace(%s):%s>\n", msg, strerror(err));
393 perror_msg("ptrace(PTRACE_%s,pid:%d,sig:%u)", msg, tcp->pid, sig);
398 set_cloexec_flag(int fd)
402 flags = fcntl(fd, F_GETFD);
404 /* Can happen only if fd is bad.
405 * Should never happen: if it does, we have a bug
406 * in the caller. Therefore we just abort
407 * instead of propagating the error.
409 perror_msg_and_die("fcntl(%d, F_GETFD)", fd);
412 newflags = flags | FD_CLOEXEC;
413 if (flags == newflags)
416 fcntl(fd, F_SETFD, newflags); /* never fails */
420 kill_save_errno(pid_t pid, int sig)
422 int saved_errno = errno;
424 (void) kill(pid, sig);
429 * When strace is setuid executable, we have to swap uids
430 * before and after filesystem and process management operations.
435 int euid = geteuid(), uid = getuid();
437 if (euid != uid && setreuid(euid, uid) < 0) {
438 perror_msg_and_die("setreuid");
443 strace_fopen(const char *path)
448 fp = fopen_stream(path, "w");
450 perror_msg_and_die("Can't fopen '%s'", path);
452 set_cloexec_flag(fileno(fp));
456 static int popen_pid;
459 # define _PATH_BSHELL "/bin/sh"
463 * We cannot use standard popen(3) here because we have to distinguish
464 * popen child process from other processes we trace, and standard popen(3)
465 * does not export its child's pid.
468 strace_popen(const char *command)
476 perror_msg_and_die("pipe");
478 set_cloexec_flag(fds[1]); /* never fails */
482 perror_msg_and_die("vfork");
489 perror_msg_and_die("dup2");
492 execl(_PATH_BSHELL, "sh", "-c", command, NULL);
493 perror_msg_and_die("Can't execute '%s'", _PATH_BSHELL);
500 fp = fdopen(fds[1], "w");
502 perror_msg_and_die("fdopen");
507 outf_perror(const struct tcb * const tcp)
509 if (tcp->outf == stderr)
512 /* This is ugly, but we don't store separate file names */
514 perror_msg("%s.%u", outfname, tcp->pid);
516 perror_msg("%s", outfname);
519 ATTRIBUTE_FORMAT((printf, 1, 0))
521 tvprintf(const char *const fmt, va_list args)
524 int n = vfprintf(current_tcp->outf, fmt, args);
526 /* very unlikely due to vfprintf buffering */
527 outf_perror(current_tcp);
529 current_tcp->curcol += n;
534 tprintf(const char *fmt, ...)
542 #ifndef HAVE_FPUTS_UNLOCKED
543 # define fputs_unlocked fputs
547 tprints(const char *str)
550 int n = fputs_unlocked(str, current_tcp->outf);
552 current_tcp->curcol += strlen(str);
555 /* very unlikely due to fputs_unlocked buffering */
556 outf_perror(current_tcp);
561 tprints_comment(const char *const str)
564 tprintf(" /* %s */", str);
568 tprintf_comment(const char *fmt, ...)
582 flush_tcp_output(const struct tcb *const tcp)
584 if (fflush(tcp->outf))
592 current_tcp->curcol = 0;
593 flush_tcp_output(current_tcp);
596 printing_tcp->curcol = 0;
602 set_current_tcp(const struct tcb *tcp)
604 current_tcp = (struct tcb *) tcp;
606 /* Sync current_personality and stuff */
608 set_personality(current_tcp->currpers);
612 printleader(struct tcb *tcp)
614 /* If -ff, "previous tcb we printed" is always the same as current,
615 * because we have per-tcb output files.
621 set_current_tcp(printing_tcp);
622 if (printing_tcp->curcol != 0 && (followfork < 2 || printing_tcp == tcp)) {
624 * case 1: we have a shared log (i.e. not -ff), and last line
625 * wasn't finished (same or different tcb, doesn't matter).
626 * case 2: split log, we are the same tcb, but our last line
627 * didn't finish ("SIGKILL nuked us after syscall entry" etc).
629 tprints(" <unfinished ...>\n");
630 printing_tcp->curcol = 0;
635 set_current_tcp(tcp);
636 current_tcp->curcol = 0;
639 tprintf("%-5d ", tcp->pid);
640 else if (nprocs > 1 && !outfname)
641 tprintf("[pid %5u] ", tcp->pid);
645 clock_gettime(rflag ? CLOCK_MONOTONIC : CLOCK_REALTIME, &ts);
648 static struct timespec ots;
653 ts_sub(&dts, &ts, &ots);
656 tprintf("%6ld.%06ld ",
657 (long) dts.tv_sec, (long) dts.tv_nsec / 1000);
658 } else if (tflag > 2) {
659 tprintf("%lld.%06ld ",
660 (long long) ts.tv_sec, (long) ts.tv_nsec / 1000);
662 time_t local = ts.tv_sec;
663 char str[sizeof("HH:MM:SS")];
665 strftime(str, sizeof(str), "%T", localtime(&local));
668 str, (long) ts.tv_nsec / 1000);
680 if (current_tcp->curcol < acolumn)
681 tprints(acolumn_spaces + current_tcp->curcol);
684 /* Should be only called directly *after successful attach* to a tracee.
685 * Otherwise, "strace -oFILE -ff -p<nonexistant_pid>"
686 * may create bogus empty FILE.<nonexistant_pid>, and then die.
689 newoutf(struct tcb *tcp)
691 tcp->outf = shared_log; /* if not -ff mode, the same file is for all */
692 if (followfork >= 2) {
694 xsprintf(name, "%s.%u", outfname, tcp->pid);
695 tcp->outf = strace_fopen(name);
702 /* Allocate some (more) TCBs (and expand the table).
703 We don't want to relocate the TCBs because our
704 callers have pointers and it would be a pain.
705 So tcbtab is a table of pointers. Since we never
706 free the TCBs, we allocate a single chunk of many. */
707 size_t old_tcbtabsize;
709 struct tcb **tcb_ptr;
711 old_tcbtabsize = tcbtabsize;
713 tcbtab = xgrowarray(tcbtab, &tcbtabsize, sizeof(tcbtab[0]));
714 newtcbs = xcalloc(tcbtabsize - old_tcbtabsize, sizeof(newtcbs[0]));
716 for (tcb_ptr = tcbtab + old_tcbtabsize;
717 tcb_ptr < tcbtab + tcbtabsize; tcb_ptr++, newtcbs++)
727 if (nprocs == tcbtabsize)
730 for (i = 0; i < tcbtabsize; i++) {
733 memset(tcp, 0, sizeof(*tcp));
735 #if SUPPORTED_PERSONALITIES > 1
736 tcp->currpers = current_personality;
740 if (stack_trace_enabled)
741 unwind_tcb_init(tcp);
745 debug_msg("new tcb for pid %d, active tcbs:%d",
750 error_msg_and_die("bug in alloctcb");
754 get_tcb_priv_data(const struct tcb *tcp)
756 return tcp->_priv_data;
760 set_tcb_priv_data(struct tcb *tcp, void *const priv_data,
761 void (*const free_priv_data)(void *))
766 tcp->_free_priv_data = free_priv_data;
767 tcp->_priv_data = priv_data;
773 free_tcb_priv_data(struct tcb *tcp)
775 if (tcp->_priv_data) {
776 if (tcp->_free_priv_data) {
777 tcp->_free_priv_data(tcp->_priv_data);
778 tcp->_free_priv_data = NULL;
780 tcp->_priv_data = NULL;
785 droptcb(struct tcb *tcp)
791 for (p = 0; p < SUPPORTED_PERSONALITIES; ++p)
792 free(tcp->inject_vec[p]);
794 free_tcb_priv_data(tcp);
797 if (stack_trace_enabled) {
802 mmap_cache_delete(tcp, __func__);
805 debug_msg("dropped tcb for pid %d, %d remain", tcp->pid, nprocs);
808 if (followfork >= 2) {
809 if (tcp->curcol != 0)
810 fprintf(tcp->outf, " <detached ...>\n");
813 if (printing_tcp == tcp && tcp->curcol != 0)
814 fprintf(tcp->outf, " <detached ...>\n");
815 flush_tcp_output(tcp);
819 if (current_tcp == tcp)
820 set_current_tcp(NULL);
821 if (printing_tcp == tcp)
824 memset(tcp, 0, sizeof(*tcp));
827 /* Detach traced process.
828 * Never call DETACH twice on the same process as both unattached and
829 * attached-unstopped processes give the same ESRCH. For unattached process we
830 * would SIGSTOP it and wait for its SIGSTOP notification forever.
833 detach(struct tcb *tcp)
839 * Linux wrongly insists the child be stopped
840 * before detaching. Arghh. We go through hoops
841 * to make a clean break of things.
844 if (!(tcp->flags & TCB_ATTACHED))
847 /* We attached but possibly didn't see the expected SIGSTOP.
848 * We must catch exactly one as otherwise the detached process
849 * would be left stopped (process state T).
851 if (tcp->flags & TCB_IGNORE_ONE_SIGSTOP)
854 error = ptrace(PTRACE_DETACH, tcp->pid, 0, 0);
856 /* On a clear day, you can see forever. */
859 if (errno != ESRCH) {
860 /* Shouldn't happen. */
861 perror_func_msg("ptrace(PTRACE_DETACH,%u)", tcp->pid);
864 /* ESRCH: process is either not stopped or doesn't exist. */
865 if (my_tkill(tcp->pid, 0) < 0) {
867 /* Shouldn't happen. */
868 perror_func_msg("tkill(%u,0)", tcp->pid);
869 /* else: process doesn't exist. */
872 /* Process is not stopped, need to stop it. */
875 * With SEIZE, tracee can be in group-stop already.
876 * In this state sending it another SIGSTOP does nothing.
877 * Need to use INTERRUPT.
878 * Testcase: trying to ^C a "strace -p <stopped_process>".
880 error = ptrace(PTRACE_INTERRUPT, tcp->pid, 0, 0);
884 perror_func_msg("ptrace(PTRACE_INTERRUPT,%u)", tcp->pid);
886 error = my_tkill(tcp->pid, SIGSTOP);
890 perror_func_msg("tkill(%u,SIGSTOP)", tcp->pid);
892 /* Either process doesn't exist, or some weird error. */
896 /* We end up here in three cases:
897 * 1. We sent PTRACE_INTERRUPT (use_seize case)
898 * 2. We sent SIGSTOP (!use_seize)
899 * 3. Attach SIGSTOP was already pending (TCB_IGNORE_ONE_SIGSTOP set)
903 if (waitpid(tcp->pid, &status, __WALL) < 0) {
907 * if (errno == ECHILD) break;
908 * ^^^ WRONG! We expect this PID to exist,
909 * and want to emit a message otherwise:
911 perror_func_msg("waitpid(%u)", tcp->pid);
914 if (!WIFSTOPPED(status)) {
916 * Tracee exited or was killed by signal.
917 * We shouldn't normally reach this place:
918 * we don't want to consume exit status.
919 * Consider "strace -p PID" being ^C-ed:
920 * we want merely to detach from PID.
922 * However, we _can_ end up here if tracee
927 sig = WSTOPSIG(status);
928 debug_msg("detach wait: event:%d sig:%d",
929 (unsigned) status >> 16, sig);
931 unsigned event = (unsigned)status >> 16;
932 if (event == PTRACE_EVENT_STOP /*&& sig == SIGTRAP*/) {
934 * sig == SIGTRAP: PTRACE_INTERRUPT stop.
935 * sig == other: process was already stopped
936 * with this stopping sig (see tests/detach-stopped).
937 * Looks like re-injecting this sig is not necessary
938 * in DETACH for the tracee to remain stopped.
943 * PTRACE_INTERRUPT is not guaranteed to produce
944 * the above event if other ptrace-stop is pending.
945 * See tests/detach-sleeping testcase:
946 * strace got SIGINT while tracee is sleeping.
947 * We sent PTRACE_INTERRUPT.
948 * We see syscall exit, not PTRACE_INTERRUPT stop.
949 * We won't get PTRACE_INTERRUPT stop
950 * if we would CONT now. Need to DETACH.
952 if (sig == syscall_trap_sig)
954 /* else: not sure in which case we can be here.
955 * Signal stop? Inject it while detaching.
957 ptrace_restart(PTRACE_DETACH, tcp, sig);
960 /* Note: this check has to be after use_seize check */
961 /* (else, in use_seize case SIGSTOP will be mistreated) */
962 if (sig == SIGSTOP) {
963 /* Detach, suppressing SIGSTOP */
964 ptrace_restart(PTRACE_DETACH, tcp, 0);
967 if (sig == syscall_trap_sig)
969 /* Can't detach just yet, may need to wait for SIGSTOP */
970 error = ptrace_restart(PTRACE_CONT, tcp, sig);
972 /* Should not happen.
973 * Note: ptrace_restart returns 0 on ESRCH, so it's not it.
974 * ptrace_restart already emitted error message.
981 if (!qflag && (tcp->flags & TCB_ATTACHED))
982 error_msg("Process %u detached", tcp->pid);
988 process_opt_p_list(char *opt)
992 * We accept -p PID,PID; -p "`pidof PROG`"; -p "`pgrep PROG`".
993 * pidof uses space as delim, pgrep uses newline. :(
996 char *delim = opt + strcspn(opt, "\n\t ,");
1000 pid = string_to_uint(opt);
1002 error_msg_and_die("Invalid process id: '%s'", opt);
1004 if (pid == strace_tracer_pid) {
1005 error_msg_and_die("I'm sorry, I can't let you do that, Dave.");
1016 attach_tcb(struct tcb *const tcp)
1018 if (ptrace_attach_or_seize(tcp->pid) < 0) {
1019 perror_msg("attach: ptrace(%s, %d)",
1020 ptrace_attach_cmd, tcp->pid);
1025 tcp->flags |= TCB_ATTACHED | TCB_GRABBED | TCB_STARTUP |
1026 post_attach_sigstop;
1028 debug_msg("attach to pid %d (main) succeeded", tcp->pid);
1030 static const char task_path[] = "/proc/%d/task";
1031 char procdir[sizeof(task_path) + sizeof(int) * 3];
1033 unsigned int ntid = 0, nerr = 0;
1035 if (followfork && tcp->pid != strace_child &&
1036 xsprintf(procdir, task_path, tcp->pid) > 0 &&
1037 (dir = opendir(procdir)) != NULL) {
1040 while ((de = read_dir(dir)) != NULL) {
1041 if (de->d_fileno == 0)
1044 int tid = string_to_uint(de->d_name);
1045 if (tid <= 0 || tid == tcp->pid)
1049 if (ptrace_attach_or_seize(tid) < 0) {
1051 debug_perror_msg("attach: ptrace(%s, %d)",
1052 ptrace_attach_cmd, tid);
1055 debug_msg("attach to pid %d succeeded", tid);
1057 struct tcb *tid_tcp = alloctcb(tid);
1058 tid_tcp->flags |= TCB_ATTACHED | TCB_GRABBED |
1059 TCB_STARTUP | post_attach_sigstop;
1068 error_msg("Process %u attached"
1070 tcp->pid, ntid - nerr + 1);
1072 error_msg("Process %u attached",
1078 startup_attach(void)
1080 pid_t parent_pid = strace_tracer_pid;
1085 * Block user interruptions as we would leave the traced
1086 * process stopped (process state T) if we would terminate in
1087 * between PTRACE_ATTACH and wait4() on SIGSTOP.
1088 * We rely on cleanup() from this point on.
1091 sigprocmask(SIG_SETMASK, &blocked_set, NULL);
1093 if (daemonized_tracer) {
1096 perror_func_msg_and_die("fork");
1098 if (pid) { /* parent */
1100 * Wait for grandchild to attach to straced process
1101 * (grandparent). Grandchild SIGKILLs us after it attached.
1102 * Grandparent's wait() is unblocked by our death,
1103 * it proceeds to exec the straced program.
1106 _exit(0); /* paranoia */
1109 /* We will be the tracer process. Remember our new pid: */
1110 strace_tracer_pid = getpid();
1113 for (tcbi = 0; tcbi < tcbtabsize; tcbi++) {
1119 /* Is this a process we should attach to, but not yet attached? */
1120 if (tcp->flags & TCB_ATTACHED)
1121 continue; /* no, we already attached it */
1123 if (tcp->pid == parent_pid || tcp->pid == strace_tracer_pid) {
1125 perror_msg("attach: pid %d", tcp->pid);
1133 sigprocmask(SIG_SETMASK, &start_set, NULL);
1136 sigprocmask(SIG_SETMASK, &blocked_set, NULL);
1138 } /* for each tcbtab[] */
1140 if (daemonized_tracer) {
1142 * Make parent go away.
1143 * Also makes grandparent's wait() unblock.
1145 kill(parent_pid, SIGKILL);
1151 sigprocmask(SIG_SETMASK, &start_set, NULL);
1154 /* Stack-o-phobic exec helper, in the hope to work around
1155 * NOMMU + "daemonized tracer" difficulty.
1157 struct exec_params {
1163 struct sigaction child_sa;
1165 static struct exec_params params_for_tracee;
1167 static void ATTRIBUTE_NOINLINE ATTRIBUTE_NORETURN
1170 struct exec_params *params = ¶ms_for_tracee;
1172 if (params->fd_to_close >= 0)
1173 close(params->fd_to_close);
1174 if (!daemonized_tracer && !use_seize) {
1175 if (ptrace(PTRACE_TRACEME, 0L, 0L, 0L) < 0) {
1176 perror_msg_and_die("ptrace(PTRACE_TRACEME, ...)");
1180 if (username != NULL) {
1182 * It is important to set groups before we
1183 * lose privileges on setuid.
1185 if (initgroups(username, run_gid) < 0) {
1186 perror_msg_and_die("initgroups");
1188 if (setregid(run_gid, params->run_egid) < 0) {
1189 perror_msg_and_die("setregid");
1191 if (setreuid(run_uid, params->run_euid) < 0) {
1192 perror_msg_and_die("setreuid");
1194 } else if (geteuid() != 0)
1195 if (setreuid(run_uid, run_uid) < 0) {
1196 perror_msg_and_die("setreuid");
1199 if (!daemonized_tracer) {
1201 * Induce a ptrace stop. Tracer (our parent)
1202 * will resume us with PTRACE_SYSCALL and display
1203 * the immediately following execve syscall.
1204 * Can't do this on NOMMU systems, we are after
1205 * vfork: parent is blocked, stopping would deadlock.
1208 kill(getpid(), SIGSTOP);
1211 /* we depend on SIGCHLD set to SIG_DFL by init code */
1212 /* if it happens to be SIG_IGN'ed, wait won't block */
1217 if (params_for_tracee.child_sa.sa_handler != SIG_DFL)
1218 sigaction(SIGCHLD, ¶ms_for_tracee.child_sa, NULL);
1220 execv(params->pathname, params->argv);
1221 perror_msg_and_die("exec");
1225 * Open a dummy descriptor for use as a placeholder.
1226 * The descriptor is O_RDONLY with FD_CLOEXEC flag set.
1227 * A read attempt from such descriptor ends with EOF,
1228 * a write attempt is rejected with EBADF.
1231 open_dummy_desc(void)
1236 perror_func_msg_and_die("pipe");
1238 set_cloexec_flag(fds[0]);
1242 /* placeholder fds status for stdin and stdout */
1243 static bool fd_is_placeholder[2];
1246 * Ensure that all standard file descriptors are open by opening placeholder
1247 * file descriptors for those standard file descriptors that are not open.
1249 * The information which descriptors have been made open is saved
1250 * in fd_is_placeholder for later use.
1253 ensure_standard_fds_opened(void)
1257 while ((fd = open_dummy_desc()) <= 2) {
1260 fd_is_placeholder[fd] = true;
1268 * Redirect stdin and stdout unless they have been opened earlier
1269 * by ensure_standard_fds_opened as placeholders.
1272 redirect_standard_fds(void)
1277 * It might be a good idea to redirect stderr as well,
1278 * but we sometimes need to print error messages.
1280 for (i = 0; i <= 1; ++i) {
1281 if (!fd_is_placeholder[i]) {
1289 startup_child(char **argv)
1291 struct_stat statbuf;
1292 const char *filename;
1293 size_t filename_len;
1294 char pathname[PATH_MAX];
1299 filename_len = strlen(filename);
1301 if (filename_len > sizeof(pathname) - 1) {
1302 errno = ENAMETOOLONG;
1303 perror_msg_and_die("exec");
1305 if (strchr(filename, '/')) {
1306 strcpy(pathname, filename);
1308 #ifdef USE_DEBUGGING_EXEC
1310 * Debuggers customarily check the current directory
1311 * first regardless of the path but doing that gives
1312 * security geeks a panic attack.
1314 else if (stat_file(filename, &statbuf) == 0)
1315 strcpy(pathname, filename);
1316 #endif /* USE_DEBUGGING_EXEC */
1321 for (path = getenv("PATH"); path && *path; path += m) {
1322 const char *colon = strchr(path, ':');
1327 m = n = strlen(path);
1329 if (!getcwd(pathname, PATH_MAX))
1331 len = strlen(pathname);
1332 } else if (n > sizeof(pathname) - 1)
1335 strncpy(pathname, path, n);
1338 if (len && pathname[len - 1] != '/')
1339 pathname[len++] = '/';
1340 if (filename_len + len > sizeof(pathname) - 1)
1342 strcpy(pathname + len, filename);
1343 if (stat_file(pathname, &statbuf) == 0 &&
1344 /* Accept only regular files
1345 with some execute bits set.
1346 XXX not perfect, might still fail */
1347 S_ISREG(statbuf.st_mode) &&
1348 (statbuf.st_mode & 0111))
1351 if (!path || !*path)
1354 if (stat_file(pathname, &statbuf) < 0) {
1355 perror_msg_and_die("Can't stat '%s'", filename);
1358 params_for_tracee.fd_to_close = (shared_log != stderr) ? fileno(shared_log) : -1;
1359 params_for_tracee.run_euid = (statbuf.st_mode & S_ISUID) ? statbuf.st_uid : run_uid;
1360 params_for_tracee.run_egid = (statbuf.st_mode & S_ISGID) ? statbuf.st_gid : run_gid;
1361 params_for_tracee.argv = argv;
1363 * On NOMMU, can be safely freed only after execve in tracee.
1364 * It's hard to know when that happens, so we just leak it.
1366 params_for_tracee.pathname = NOMMU_SYSTEM ? xstrdup(pathname) : pathname;
1368 #if defined HAVE_PRCTL && defined PR_SET_PTRACER && defined PR_SET_PTRACER_ANY
1369 if (daemonized_tracer)
1370 prctl(PR_SET_PTRACER, PR_SET_PTRACER_ANY);
1375 perror_func_msg_and_die("fork");
1377 if ((pid != 0 && daemonized_tracer)
1378 || (pid == 0 && !daemonized_tracer)
1380 /* We are to become the tracee. Two cases:
1382 * not -D: we are child
1387 /* We are the tracer */
1389 if (!daemonized_tracer) {
1392 /* child did PTRACE_TRACEME, nothing to do in parent */
1394 if (!NOMMU_SYSTEM) {
1395 /* Wait until child stopped itself */
1397 while (waitpid(pid, &status, WSTOPPED) < 0) {
1400 perror_msg_and_die("waitpid");
1402 if (!WIFSTOPPED(status) || WSTOPSIG(status) != SIGSTOP) {
1403 kill_save_errno(pid, SIGKILL);
1404 perror_msg_and_die("Unexpected wait status %#x",
1408 /* Else: NOMMU case, we have no way to sync.
1409 * Just attach to it as soon as possible.
1410 * This means that we may miss a few first syscalls...
1413 if (ptrace_attach_or_seize(pid)) {
1414 kill_save_errno(pid, SIGKILL);
1415 perror_msg_and_die("attach: ptrace(%s, %d)",
1416 ptrace_attach_cmd, pid);
1421 tcp = alloctcb(pid);
1422 tcp->flags |= TCB_ATTACHED | TCB_STARTUP
1423 | TCB_SKIP_DETACH_ON_FIRST_EXEC
1424 | (NOMMU_SYSTEM ? 0 : (TCB_HIDE_LOG | post_attach_sigstop));
1427 /* With -D, we are *child* here, the tracee is our parent. */
1428 strace_child = strace_tracer_pid;
1429 strace_tracer_pid = getpid();
1430 tcp = alloctcb(strace_child);
1431 tcp->flags |= TCB_SKIP_DETACH_ON_FIRST_EXEC | TCB_HIDE_LOG;
1432 /* attaching will be done later, by startup_attach */
1433 /* note: we don't do newoutf(tcp) 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 /* if signal handler is a function, add the signal to blocked_set */
1557 if (sighandler != SIG_IGN && sighandler != SIG_DFL)
1558 sigaddset(&blocked_set, signo);
1560 const struct sigaction sa = { .sa_handler = sighandler };
1561 sigaction(signo, &sa, oldact);
1565 * Initialization part of main() was eating much stack (~0.5k),
1566 * which was unused after init.
1567 * We can reuse it if we move init code into a separate function.
1569 * Don't want main() to inline us and defeat the reason
1570 * we have a separate function.
1572 static void ATTRIBUTE_NOINLINE
1573 init(int argc, char *argv[])
1578 if (!program_invocation_name || !*program_invocation_name) {
1579 static char name[] = "strace";
1580 program_invocation_name =
1581 (argc > 0 && argv[0] && *argv[0]) ? argv[0] : name;
1584 strace_tracer_pid = getpid();
1586 os_release = get_os_release();
1588 shared_log = stderr;
1589 set_sortby(DEFAULT_SORTBY);
1590 set_personality(DEFAULT_PERSONALITY);
1591 qualify("trace=all");
1592 qualify("abbrev=all");
1593 qualify("verbose=all");
1594 #if DEFAULT_QUAL_FLAGS != (QUAL_TRACE | QUAL_ABBREV | QUAL_VERBOSE)
1595 # error Bug in DEFAULT_QUAL_FLAGS
1597 qualify("signal=all");
1598 while ((c = getopt(argc, argv, "+"
1599 #ifdef USE_LIBUNWIND
1602 "a:b:cCdDe:E:fFhiI:o:O:p:P:qrs:S:tTu:vVwxyz")) != EOF) {
1605 acolumn = string_to_uint(optarg);
1607 error_opt_arg(c, optarg);
1610 if (strcmp(optarg, "execve") != 0)
1611 error_msg_and_die("Syscall '%s' for -b isn't supported",
1613 detach_on_execve = 1;
1616 if (cflag == CFLAG_BOTH) {
1617 error_msg_and_help("-c and -C are mutually exclusive");
1619 cflag = CFLAG_ONLY_STATS;
1622 if (cflag == CFLAG_ONLY_STATS) {
1623 error_msg_and_help("-c and -C are mutually exclusive");
1631 daemonized_tracer = 1;
1637 if (putenv(optarg) < 0)
1638 perror_msg_and_die("putenv");
1653 opt_intr = string_to_uint_upto(optarg, NUM_INTR_OPTS - 1);
1655 error_opt_arg(c, optarg);
1657 #ifdef USE_LIBUNWIND
1659 stack_trace_enabled = true;
1666 i = string_to_uint(optarg);
1668 error_opt_arg(c, optarg);
1672 process_opt_p_list(optarg);
1675 pathtrace_select(optarg);
1684 i = string_to_uint(optarg);
1685 if (i < 0 || (unsigned int) i > -1U / 4)
1686 error_opt_arg(c, optarg);
1702 qualify("abbrev=none");
1709 count_wallclock = 1;
1718 not_failing_only = 1;
1721 error_msg_and_help(NULL);
1729 if (argc < 0 || (!nprocs && !argc)) {
1730 error_msg_and_help("must have PROG [ARGS] or -p PID");
1733 if (!argc && daemonized_tracer) {
1734 error_msg_and_help("PROG [ARGS] must be specified with -D");
1739 error_msg("deprecated option -F ignored");
1741 error_msg("option -F is deprecated, "
1742 "please use -f instead");
1747 if (followfork >= 2 && cflag) {
1748 error_msg_and_help("(-c or -C) and -ff are mutually exclusive");
1751 if (count_wallclock && !cflag) {
1752 error_msg_and_help("-w must be given with (-c or -C)");
1755 if (cflag == CFLAG_ONLY_STATS) {
1757 error_msg("-%c has no effect with -c", 'i');
1758 #ifdef USE_LIBUNWIND
1759 if (stack_trace_enabled)
1760 error_msg("-%c has no effect with -c", 'k');
1763 error_msg("-%c has no effect with -c", 'r');
1765 error_msg("-%c has no effect with -c", 't');
1767 error_msg("-%c has no effect with -c", 'T');
1769 error_msg("-%c has no effect with -c", 'y');
1774 error_msg("-tt has no effect with -r");
1778 acolumn_spaces = xmalloc(acolumn + 1);
1779 memset(acolumn_spaces, ' ', acolumn);
1780 acolumn_spaces[acolumn] = '\0';
1782 sigprocmask(SIG_SETMASK, NULL, &start_set);
1783 memcpy(&blocked_set, &start_set, sizeof(blocked_set));
1785 set_sighandler(SIGCHLD, SIG_DFL, ¶ms_for_tracee.child_sa);
1787 #ifdef USE_LIBUNWIND
1788 if (stack_trace_enabled) {
1792 for (tcbi = 0; tcbi < tcbtabsize; ++tcbi) {
1793 unwind_tcb_init(tcbtab[tcbi]);
1798 /* See if they want to run as another user. */
1799 if (username != NULL) {
1800 struct passwd *pent;
1802 if (getuid() != 0 || geteuid() != 0) {
1803 error_msg_and_die("You must be root to use the -u option");
1805 pent = getpwnam(username);
1807 error_msg_and_die("Cannot find user '%s'", username);
1809 run_uid = pent->pw_uid;
1810 run_gid = pent->pw_gid;
1817 ptrace_setoptions |= PTRACE_O_TRACECLONE |
1818 PTRACE_O_TRACEFORK |
1819 PTRACE_O_TRACEVFORK;
1820 debug_msg("ptrace_setoptions = %#x", ptrace_setoptions);
1821 test_ptrace_seize();
1824 * Is something weird with our stdin and/or stdout -
1825 * for example, may they be not open? In this case,
1826 * ensure that none of the future opens uses them.
1828 * This was seen in the wild when /proc/sys/kernel/core_pattern
1829 * was set to "|/bin/strace -o/tmp/LOG PROG":
1830 * kernel runs coredump helper with fd#0 open but fd#1 closed (!),
1831 * therefore LOG gets opened to fd#1, and fd#1 is closed by
1832 * "don't hold up stdin/out open" code soon after.
1834 ensure_standard_fds_opened();
1836 /* Check if they want to redirect the output. */
1838 /* See if they want to pipe the output. */
1839 if (outfname[0] == '|' || outfname[0] == '!') {
1841 * We can't do the <outfname>.PID funny business
1842 * when using popen, so prohibit it.
1844 if (followfork >= 2)
1845 error_msg_and_help("piping the output and -ff "
1846 "are mutually exclusive");
1847 shared_log = strace_popen(outfname + 1);
1848 } else if (followfork < 2) {
1849 shared_log = strace_fopen(outfname);
1850 } else if (strlen(outfname) >= PATH_MAX - sizeof(int) * 3) {
1851 errno = ENAMETOOLONG;
1852 perror_msg_and_die("%s", outfname);
1855 /* -ff without -o FILE is the same as single -f */
1856 if (followfork >= 2)
1860 if (!outfname || outfname[0] == '|' || outfname[0] == '!') {
1861 setvbuf(shared_log, NULL, _IOLBF, 0);
1865 * argv[0] -pPID -oFILE Default interactive setting
1866 * yes * 0 INTR_WHILE_WAIT
1867 * no 1 0 INTR_WHILE_WAIT
1868 * yes * 1 INTR_NEVER
1869 * no 1 1 INTR_WHILE_WAIT
1872 if (outfname && argc) {
1874 opt_intr = INTR_NEVER;
1879 opt_intr = INTR_WHILE_WAIT;
1882 * startup_child() must be called before the signal handlers get
1883 * installed below as they are inherited into the spawned process.
1884 * Also we do not need to be protected by them as during interruption
1885 * in the startup_child() mode we kill the spawned process anyway.
1888 startup_child(argv);
1891 set_sighandler(SIGTTOU, SIG_IGN, NULL);
1892 set_sighandler(SIGTTIN, SIG_IGN, NULL);
1893 if (opt_intr != INTR_ANYWHERE) {
1894 if (opt_intr == INTR_BLOCK_TSTP_TOO)
1895 set_sighandler(SIGTSTP, SIG_IGN, NULL);
1897 * In interactive mode (if no -o OUTFILE, or -p PID is used),
1898 * fatal signals are blocked while syscall stop is processed,
1899 * and acted on in between, when waiting for new syscall stops.
1900 * In non-interactive mode, signals are ignored.
1902 set_sighandler(SIGHUP, interactive ? interrupt : SIG_IGN, NULL);
1903 set_sighandler(SIGINT, interactive ? interrupt : SIG_IGN, NULL);
1904 set_sighandler(SIGQUIT, interactive ? interrupt : SIG_IGN, NULL);
1905 set_sighandler(SIGPIPE, interactive ? interrupt : SIG_IGN, NULL);
1906 set_sighandler(SIGTERM, interactive ? interrupt : SIG_IGN, NULL);
1909 if (nprocs != 0 || daemonized_tracer)
1912 /* Do we want pids printed in our -o OUTFILE?
1913 * -ff: no (every pid has its own file); or
1914 * -f: yes (there can be more pids in the future); or
1915 * -p PID1,PID2: yes (there are already more than one pid)
1917 print_pid_pfx = (outfname && followfork < 2 && (followfork == 1 || nprocs > 1));
1921 pid2tcb(const int pid)
1926 #define PID2TCB_CACHE_SIZE 1024U
1927 #define PID2TCB_CACHE_MASK (PID2TCB_CACHE_SIZE - 1)
1929 static struct tcb *pid2tcb_cache[PID2TCB_CACHE_SIZE];
1930 struct tcb **const ptcp = &pid2tcb_cache[pid & PID2TCB_CACHE_MASK];
1931 struct tcb *tcp = *ptcp;
1933 if (tcp && tcp->pid == pid)
1936 for (unsigned int i = 0; i < tcbtabsize; ++i) {
1938 if (tcp->pid == pid)
1952 /* 'interrupted' is a volatile object, fetch it only once */
1953 fatal_sig = interrupted;
1955 fatal_sig = SIGTERM;
1957 for (i = 0; i < tcbtabsize; i++) {
1961 debug_func_msg("looking at pid %u", tcp->pid);
1962 if (tcp->pid == strace_child) {
1963 kill(tcp->pid, SIGCONT);
1964 kill(tcp->pid, fatal_sig);
1969 call_summary(shared_log);
1979 print_debug_info(const int pid, int status)
1981 const unsigned int event = (unsigned int) status >> 16;
1982 char buf[sizeof("WIFEXITED,exitcode=%u") + sizeof(int)*3 /*paranoia:*/ + 16];
1983 char evbuf[sizeof(",EVENT_VFORK_DONE (%u)") + sizeof(int)*3 /*paranoia:*/ + 16];
1986 if (WIFSIGNALED(status))
1988 xsprintf(buf, "WIFSIGNALED,%ssig=%s",
1989 WCOREDUMP(status) ? "core," : "",
1990 signame(WTERMSIG(status)));
1992 xsprintf(buf, "WIFSIGNALED,sig=%s",
1993 signame(WTERMSIG(status)));
1995 if (WIFEXITED(status))
1996 xsprintf(buf, "WIFEXITED,exitcode=%u", WEXITSTATUS(status));
1997 if (WIFSTOPPED(status))
1998 xsprintf(buf, "WIFSTOPPED,sig=%s", signame(WSTOPSIG(status)));
2001 static const char *const event_names[] = {
2002 [PTRACE_EVENT_CLONE] = "CLONE",
2003 [PTRACE_EVENT_FORK] = "FORK",
2004 [PTRACE_EVENT_VFORK] = "VFORK",
2005 [PTRACE_EVENT_VFORK_DONE] = "VFORK_DONE",
2006 [PTRACE_EVENT_EXEC] = "EXEC",
2007 [PTRACE_EVENT_EXIT] = "EXIT",
2008 /* [PTRACE_EVENT_STOP (=128)] would make biggish array */
2010 const char *e = "??";
2011 if (event < ARRAY_SIZE(event_names))
2012 e = event_names[event];
2013 else if (event == PTRACE_EVENT_STOP)
2015 xsprintf(evbuf, ",EVENT_%s (%u)", e, event);
2017 error_msg("[wait(0x%06x) = %u] %s%s", status, pid, buf, evbuf);
2021 maybe_allocate_tcb(const int pid, int status)
2023 if (!WIFSTOPPED(status)) {
2024 if (detach_on_execve && pid == strace_child) {
2025 /* example: strace -bexecve sh -c 'exec true' */
2030 * This can happen if we inherited an unknown child.
2031 * Example: (sleep 1 & exec strace true)
2033 error_msg("Exit of unknown pid %u ignored", pid);
2037 /* We assume it's a fork/vfork/clone child */
2038 struct tcb *tcp = alloctcb(pid);
2039 tcp->flags |= TCB_ATTACHED | TCB_STARTUP | post_attach_sigstop;
2042 error_msg("Process %d attached", pid);
2046 * This can happen if a clone call misused CLONE_PTRACE itself.
2048 * There used to be a dance around possible re-injection of
2049 * WSTOPSIG(status), but it was later removed as the only
2050 * observable stop here is the initial ptrace-stop.
2052 ptrace(PTRACE_DETACH, pid, NULL, 0L);
2053 error_msg("Detached unknown pid %d", pid);
2059 maybe_switch_tcbs(struct tcb *tcp, const int pid)
2062 struct tcb *execve_thread;
2065 if (ptrace(PTRACE_GETEVENTMSG, pid, NULL, &old_pid) < 0)
2067 /* Avoid truncation in pid2tcb() param passing */
2068 if (old_pid <= 0 || old_pid == pid)
2070 if ((unsigned long) old_pid > UINT_MAX)
2072 execve_thread = pid2tcb(old_pid);
2073 /* It should be !NULL, but I feel paranoid */
2077 if (execve_thread->curcol != 0) {
2079 * One case we are here is -ff:
2080 * try "strace -oLOG -ff test/threaded_execve"
2082 fprintf(execve_thread->outf, " <pid changed to %d ...>\n", pid);
2083 /*execve_thread->curcol = 0; - no need, see code below */
2085 /* Swap output FILEs (needed for -ff) */
2086 fp = execve_thread->outf;
2087 execve_thread->outf = tcp->outf;
2089 /* And their column positions */
2090 execve_thread->curcol = tcp->curcol;
2092 /* Drop leader, but close execve'd thread outfile (if -ff) */
2094 /* Switch to the thread, reusing leader's outfile and pid */
2095 tcp = execve_thread;
2097 if (cflag != CFLAG_ONLY_STATS) {
2099 tprintf("+++ superseded by execve in pid %lu +++\n", old_pid);
2101 tcp->flags |= TCB_REPRINT;
2108 print_signalled(struct tcb *tcp, const int pid, int status)
2110 if (pid == strace_child) {
2111 exit_code = 0x100 | WTERMSIG(status);
2115 if (cflag != CFLAG_ONLY_STATS
2116 && is_number_in_set(WTERMSIG(status), signal_set)) {
2119 tprintf("+++ killed by %s %s+++\n",
2120 signame(WTERMSIG(status)),
2121 WCOREDUMP(status) ? "(core dumped) " : "");
2123 tprintf("+++ killed by %s +++\n",
2124 signame(WTERMSIG(status)));
2131 print_exited(struct tcb *tcp, const int pid, int status)
2133 if (pid == strace_child) {
2134 exit_code = WEXITSTATUS(status);
2138 if (cflag != CFLAG_ONLY_STATS &&
2141 tprintf("+++ exited with %d +++\n", WEXITSTATUS(status));
2147 print_stopped(struct tcb *tcp, const siginfo_t *si, const unsigned int sig)
2149 if (cflag != CFLAG_ONLY_STATS
2151 && is_number_in_set(sig, signal_set)) {
2154 tprintf("--- %s ", signame(sig));
2158 tprintf("--- stopped by %s ---\n", signame(sig));
2164 startup_tcb(struct tcb *tcp)
2166 debug_msg("pid %d has TCB_STARTUP, initializing it", tcp->pid);
2168 tcp->flags &= ~TCB_STARTUP;
2171 debug_msg("setting opts 0x%x on pid %d",
2172 ptrace_setoptions, tcp->pid);
2173 if (ptrace(PTRACE_SETOPTIONS, tcp->pid, NULL, ptrace_setoptions) < 0) {
2174 if (errno != ESRCH) {
2175 /* Should never happen, really */
2176 perror_msg_and_die("PTRACE_SETOPTIONS");
2181 if ((tcp->flags & TCB_GRABBED) && (get_scno(tcp) == 1))
2182 tcp->s_prev_ent = tcp->s_ent;
2186 print_event_exit(struct tcb *tcp)
2188 if (entering(tcp) || filtered(tcp) || hide_log(tcp)
2189 || cflag == CFLAG_ONLY_STATS) {
2193 if (followfork < 2 && printing_tcp && printing_tcp != tcp
2194 && printing_tcp->curcol != 0) {
2195 set_current_tcp(printing_tcp);
2196 tprints(" <unfinished ...>\n");
2197 flush_tcp_output(printing_tcp);
2198 printing_tcp->curcol = 0;
2199 set_current_tcp(tcp);
2202 if ((followfork < 2 && printing_tcp != tcp)
2203 || (tcp->flags & TCB_REPRINT)) {
2204 tcp->flags &= ~TCB_REPRINT;
2206 tprintf("<... %s resumed>", tcp->s_ent->sys_name);
2209 if (!(tcp->sys_func_rval & RVAL_DECODED)) {
2211 * The decoder has probably decided to print something
2212 * on exiting syscall which is not going to happen.
2214 tprints(" <unfinished ...>");
2224 static enum trace_event
2225 next_event(int *pstatus, siginfo_t *si)
2237 * Used to exit simply when nprocs hits zero, but in this testcase:
2238 * int main(void) { _exit(!!fork()); }
2239 * under strace -f, parent sometimes (rarely) manages
2240 * to exit before we see the first stop of the child,
2241 * and we are losing track of it:
2242 * 19923 clone(...) = 19924
2243 * 19923 exit_group(1) = ?
2244 * 19923 +++ exited with 1 +++
2245 * Exiting only when wait() returns ECHILD works better.
2247 if (popen_pid != 0) {
2248 /* However, if -o|logger is in use, we can't do that.
2249 * Can work around that by double-forking the logger,
2250 * but that loses the ability to wait for its completion
2251 * on exit. Oh well...
2258 sigprocmask(SIG_SETMASK, &start_set, NULL);
2259 pid = wait4(-1, pstatus, __WALL, (cflag ? &ru : NULL));
2262 sigprocmask(SIG_SETMASK, &blocked_set, NULL);
2265 if (wait_errno == EINTR)
2267 if (nprocs == 0 && wait_errno == ECHILD)
2270 * If nprocs > 0, ECHILD is not expected,
2271 * treat it as any other error here:
2274 perror_msg_and_die("wait4(__WALL)");
2279 if (pid == popen_pid) {
2280 if (!WIFSTOPPED(status))
2286 print_debug_info(pid, status);
2288 /* Look up 'pid' in our table. */
2292 tcp = maybe_allocate_tcb(pid, status);
2299 /* Set current output file */
2300 set_current_tcp(tcp);
2303 struct timespec stime = {
2304 .tv_sec = ru.ru_stime.tv_sec,
2305 .tv_nsec = ru.ru_stime.tv_usec * 1000
2307 ts_sub(&tcp->dtime, &stime, &tcp->stime);
2311 if (WIFSIGNALED(status))
2312 return TE_SIGNALLED;
2314 if (WIFEXITED(status))
2318 * As WCONTINUED flag has not been specified to wait4,
2319 * it cannot be WIFCONTINUED(status), so the only case
2320 * that remains is WIFSTOPPED(status).
2323 /* Is this the very first time we see this tracee stopped? */
2324 if (tcp->flags & TCB_STARTUP)
2327 const unsigned int sig = WSTOPSIG(status);
2328 const unsigned int event = (unsigned int) status >> 16;
2333 * Is this post-attach SIGSTOP?
2334 * Interestingly, the process may stop
2335 * with STOPSIG equal to some other signal
2336 * than SIGSTOP if we happened to attach
2337 * just before the process takes a signal.
2339 if (sig == SIGSTOP && (tcp->flags & TCB_IGNORE_ONE_SIGSTOP)) {
2340 debug_func_msg("ignored SIGSTOP on pid %d", tcp->pid);
2341 tcp->flags &= ~TCB_IGNORE_ONE_SIGSTOP;
2343 } else if (sig == syscall_trap_sig) {
2344 return TE_SYSCALL_STOP;
2346 *si = (siginfo_t) {};
2348 * True if tracee is stopped by signal
2349 * (as opposed to "tracee received signal").
2350 * TODO: shouldn't we check for errno == EINVAL too?
2351 * We can get ESRCH instead, you know...
2353 bool stopped = ptrace(PTRACE_GETSIGINFO, pid, 0, si) < 0;
2354 return stopped ? TE_GROUP_STOP : TE_SIGNAL_DELIVERY_STOP;
2357 case PTRACE_EVENT_STOP:
2359 * PTRACE_INTERRUPT-stop or group-stop.
2360 * PTRACE_INTERRUPT-stop has sig == SIGTRAP here.
2367 return TE_GROUP_STOP;
2370 case PTRACE_EVENT_EXEC:
2371 return TE_STOP_BEFORE_EXECVE;
2372 case PTRACE_EVENT_EXIT:
2373 return TE_STOP_BEFORE_EXIT;
2380 trace_syscall(struct tcb *tcp, unsigned int *sig)
2382 if (entering(tcp)) {
2383 int res = syscall_entering_decode(tcp);
2388 res = syscall_entering_trace(tcp, sig);
2390 syscall_entering_finish(tcp, res);
2393 struct timespec ts = {};
2394 int res = syscall_exiting_decode(tcp, &ts);
2396 res = syscall_exiting_trace(tcp, &ts, res);
2398 syscall_exiting_finish(tcp);
2403 /* Returns true iff the main trace loop has to continue. */
2405 dispatch_event(enum trace_event ret, int *pstatus, siginfo_t *si)
2407 unsigned int restart_op = PTRACE_SYSCALL;
2408 unsigned int restart_sig = 0;
2420 case TE_SYSCALL_STOP:
2421 if (trace_syscall(current_tcp, &restart_sig) < 0) {
2423 * ptrace() failed in trace_syscall().
2424 * Likely a result of process disappearing mid-flight.
2425 * Observed case: exit_group() or SIGKILL terminating
2426 * all processes in thread group.
2427 * We assume that ptrace error was caused by process death.
2428 * We used to detach(current_tcp) here, but since we no
2429 * longer implement "detach before death" policy/hack,
2430 * we can let this process to report its death to us
2431 * normally, via WIFEXITED or WIFSIGNALED wait status.
2437 case TE_SIGNAL_DELIVERY_STOP:
2438 restart_sig = WSTOPSIG(*pstatus);
2439 print_stopped(current_tcp, si, restart_sig);
2443 print_signalled(current_tcp, current_tcp->pid, *pstatus);
2444 droptcb(current_tcp);
2448 restart_sig = WSTOPSIG(*pstatus);
2449 print_stopped(current_tcp, NULL, restart_sig);
2452 * This ends ptrace-stop, but does *not* end group-stop.
2453 * This makes stopping signals work properly on straced
2454 * process (that is, process really stops. It used to
2457 restart_op = PTRACE_LISTEN;
2463 print_exited(current_tcp, current_tcp->pid, *pstatus);
2464 droptcb(current_tcp);
2467 case TE_STOP_BEFORE_EXECVE:
2469 * Check that we are inside syscall now (next event after
2470 * PTRACE_EVENT_EXEC should be for syscall exiting). If it is
2471 * not the case, we might have a situation when we attach to a
2472 * process and the first thing we see is a PTRACE_EVENT_EXEC
2473 * and all the following syscall state tracking is screwed up
2476 if (entering(current_tcp)) {
2479 error_msg("Stray PTRACE_EVENT_EXEC from pid %d"
2480 ", trying to recover...",
2483 current_tcp->flags |= TCB_RECOVERING;
2484 ret = trace_syscall(current_tcp, &restart_sig);
2485 current_tcp->flags &= ~TCB_RECOVERING;
2488 /* The reason is described in TE_SYSCALL_STOP */
2494 * Under Linux, execve changes pid to thread leader's pid,
2495 * and we see this changed pid on EVENT_EXEC and later,
2496 * execve sysexit. Leader "disappears" without exit
2497 * notification. Let user know that, drop leader's tcb,
2498 * and fix up pid in execve thread's tcb.
2499 * Effectively, execve thread's tcb replaces leader's tcb.
2501 * BTW, leader is 'stuck undead' (doesn't report WIFEXITED
2502 * on exit syscall) in multithreaded programs exactly
2503 * in order to handle this case.
2505 * PTRACE_GETEVENTMSG returns old pid starting from Linux 3.0.
2506 * On 2.6 and earlier, it can return garbage.
2508 if (os_release >= KERNEL_VERSION(3, 0, 0))
2509 set_current_tcp(maybe_switch_tcbs(current_tcp,
2512 if (detach_on_execve) {
2513 if (current_tcp->flags & TCB_SKIP_DETACH_ON_FIRST_EXEC) {
2514 current_tcp->flags &= ~TCB_SKIP_DETACH_ON_FIRST_EXEC;
2516 detach(current_tcp); /* do "-b execve" thingy */
2522 case TE_STOP_BEFORE_EXIT:
2523 print_event_exit(current_tcp);
2527 /* We handled quick cases, we are permitted to interrupt now. */
2531 if (ptrace_restart(restart_op, current_tcp, restart_sig) < 0) {
2532 /* Note: ptrace_restart emitted error message */
2539 #ifdef ENABLE_COVERAGE_GCOV
2540 extern void __gcov_flush(void);
2543 static void ATTRIBUTE_NORETURN
2548 if (shared_log != stderr)
2551 while (waitpid(popen_pid, NULL, 0) < 0 && errno == EINTR)
2554 if (exit_code > 0xff) {
2555 /* Avoid potential core file clobbering. */
2556 struct_rlimit rlim = {0, 0};
2557 set_rlimit(RLIMIT_CORE, &rlim);
2559 /* Child was killed by a signal, mimic that. */
2561 signal(exit_code, SIG_DFL);
2562 #ifdef ENABLE_COVERAGE_GCOV
2567 /* Unblock the signal. */
2570 sigaddset(&mask, exit_code);
2571 #ifdef ENABLE_COVERAGE_GCOV
2574 sigprocmask(SIG_UNBLOCK, &mask, NULL);
2576 /* Paranoia - what if this signal is not fatal?
2577 Exit with 128 + signo then. */
2584 main(int argc, char *argv[])
2588 exit_code = !nprocs;
2592 while (dispatch_event(next_event(&status, &si), &status, &si))