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) 1999-2019 The strace developers.
9 * SPDX-License-Identifier: LGPL-2.1-or-later
18 #include <sys/resource.h>
27 #include <sys/utsname.h>
29 # include <sys/prctl.h>
32 #include "kill_save_errno.h"
33 #include "largefile_wrappers.h"
34 #include "mmap_cache.h"
35 #include "number_set.h"
36 #include "ptrace_syscall_info.h"
38 #include "printsiginfo.h"
39 #include "trace_event.h"
44 /* In some libc, these aren't declared. Do it ourself: */
45 extern char **environ;
49 #ifdef ENABLE_STACKTRACE
50 /* if this is true do the stack trace for every system call */
51 bool stack_trace_enabled;
54 #define my_tkill(tid, sig) syscall(__NR_tkill, (tid), (sig))
56 /* Glue for systems without a MMU that cannot provide fork() */
57 #if !defined(HAVE_FORK)
59 # define NOMMU_SYSTEM 1
62 # define fork() vfork()
65 const unsigned int syscall_trap_sig = SIGTRAP | 0x80;
67 cflag_t cflag = CFLAG_NONE;
68 unsigned int followfork;
69 unsigned int ptrace_setoptions = PTRACE_O_TRACESYSGOOD | PTRACE_O_TRACEEXEC
77 static unsigned int tflag;
79 static bool print_pid_pfx;
84 INTR_ANYWHERE = 1, /* don't block/ignore any signals */
85 INTR_WHILE_WAIT = 2, /* block fatal signals while decoding syscall. default */
86 INTR_NEVER = 3, /* block fatal signals. default if '-o FILE PROG' */
87 INTR_BLOCK_TSTP_TOO = 4, /* block fatal signals and SIGTSTP (^Z); default if -D */
91 /* We play with signal mask only if this mode is active: */
92 #define interactive (opt_intr == INTR_WHILE_WAIT)
95 * daemonized_tracer supports -D option.
96 * With this option, strace forks twice.
97 * Unlike normal case, with -D *grandparent* process exec's,
98 * becoming a traced process. Child exits (this prevents traced process
99 * from having children it doesn't expect to have), and grandchild
100 * attaches to grandparent similarly to strace -p PID.
101 * This allows for more transparent interaction in cases
102 * when process and its parent are communicating via signals,
103 * wait() etc. Without -D, strace process gets lodged in between,
104 * disrupting parent<->child link.
106 static bool daemonized_tracer;
108 static int post_attach_sigstop = TCB_IGNORE_ONE_SIGSTOP;
109 #define use_seize (post_attach_sigstop == 0)
111 /* Show path associated with fd arguments */
112 unsigned int show_fd_path;
114 static bool detach_on_execve;
116 static int exit_code;
117 static int strace_child;
118 static int strace_tracer_pid;
120 static const char *username;
121 static uid_t run_uid;
122 static gid_t run_gid;
124 unsigned int max_strlen = DEFAULT_STRLEN;
125 static int acolumn = DEFAULT_ACOLUMN;
126 static char *acolumn_spaces;
128 /* Default output style for xlat entities */
129 enum xlat_style xlat_verbosity = XLAT_STYLE_ABBREV;
131 static const char *outfname;
132 /* If -ff, points to stderr. Else, it's our common output log */
133 static FILE *shared_log;
134 static bool open_append;
136 struct tcb *printing_tcp;
137 static struct tcb *current_tcp;
139 struct tcb_wait_data {
140 enum trace_event te; /**< Event passed to dispatch_event() */
141 int status; /**< status, returned by wait4() */
142 unsigned long msg; /**< Value returned by PTRACE_GETEVENTMSG */
143 siginfo_t si; /**< siginfo, returned by PTRACE_GETSIGINFO */
146 static struct tcb **tcbtab;
147 static unsigned int nprocs;
148 static size_t tcbtabsize;
150 static struct tcb_wait_data *tcb_wait_tab;
151 static size_t tcb_wait_tab_size;
154 #ifndef HAVE_PROGRAM_INVOCATION_NAME
155 char *program_invocation_name;
158 unsigned os_release; /* generated from uname()'s u.release */
160 static void detach(struct tcb *tcp);
161 static void cleanup(int sig);
162 static void interrupt(int sig);
164 #ifdef HAVE_SIG_ATOMIC_T
165 static volatile sig_atomic_t interrupted, restart_failed;
167 static volatile int interrupted, restart_failed;
170 static sigset_t timer_set;
171 static void timer_sighandler(int);
173 #ifndef HAVE_STRERROR
175 # if !HAVE_DECL_SYS_ERRLIST
177 extern char *sys_errlist[];
183 static char buf[sizeof("Unknown error %d") + sizeof(int)*3];
185 if (err_no < 1 || err_no >= sys_nerr) {
186 xsprintf(buf, "Unknown error %d", err_no);
189 return sys_errlist[err_no];
192 #endif /* HAVE_STERRROR */
197 static const char features[] =
198 #ifdef ENABLE_STACKTRACE
199 " stack-trace=" USE_UNWINDER
204 #if SUPPORTED_PERSONALITIES > 1
205 # if defined HAVE_M32_MPERS
210 #endif /* SUPPORTED_PERSONALITIES > 1 */
211 #if SUPPORTED_PERSONALITIES > 2
212 # if defined HAVE_MX32_MPERS
217 #endif /* SUPPORTED_PERSONALITIES > 2 */
220 printf("%s -- version %s\n"
221 "Copyright (c) 1991-%s The strace developers <%s>.\n"
222 "This is free software; see the source for copying conditions. There is NO\n"
223 "warranty; not even for MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.\n",
224 PACKAGE_NAME, PACKAGE_VERSION, COPYRIGHT_YEAR, PACKAGE_URL);
225 printf("\nOptional features enabled:%s\n",
226 features[0] ? features : " (none)");
232 #ifdef ENABLE_STACKTRACE
239 usage: strace [-ACdffhi" K_OPT "qqrtttTvVwxxyyzZ] [-I n] [-b execve] [-e expr]...\n\
240 [-a column] [-o file] [-s strsize] [-X format] [-P path]...\n\
242 { -p pid | [-D] [-E var=val]... [-u username] PROG [ARGS] }\n\
243 or: strace -c[dfwzZ] [-I n] [-b execve] [-e expr]... [-O overhead]\n\
244 [-S sortby] [-P path]... [-p pid]...\n\
245 { -p pid | [-D] [-E var=val]... [-u username] PROG [ARGS] }\n\
248 -A open the file provided in the -o option in append mode\n\
249 -a column alignment COLUMN for printing syscall results (default %d)\n\
250 -i print instruction pointer at time of syscall\n\
252 #ifdef ENABLE_STACKTRACE
254 -k obtain stack trace between each syscall\n\
258 -o file send trace output to FILE instead of stderr\n\
259 -q suppress messages about attaching, detaching, etc.\n\
260 -qq suppress messages about process exit status as well.\n\
261 -r print relative timestamp\n\
262 -s strsize limit length of print strings to STRSIZE chars (default %d)\n\
263 -t print absolute timestamp\n\
264 -tt print absolute timestamp with usecs\n\
265 -T print time spent in each syscall\n\
266 -v verbose mode: print entities unabbreviated\n\
267 -x print non-ascii strings in hex\n\
268 -xx print all strings in hex\n\
269 -X format set the format for printing of named constants and flags\n\
270 -y print paths associated with file descriptor arguments\n\
271 -yy print protocol specific information associated with socket\n\
275 -c count time, calls, and errors for each syscall and report\n\
277 -C like -c but also print regular output\n\
278 -O overhead set overhead for tracing syscalls to OVERHEAD usecs\n\
279 -S sortby sort syscall counts by: time, calls, errors, name, nothing\n\
281 -w summarise syscall latency (default is system time)\n\
284 -e expr a qualifying expression: option=[!]all or option=[!]val1[,val2]...\n\
285 options: trace, abbrev, verbose, raw, signal, read, write, fault,\n\
286 inject, status, kvm\n\
287 -P path trace accesses to path\n\
288 -z print only syscalls that returned without an error code\n\
289 -Z print only syscalls that returned with an error code\n\
292 -b execve detach on execve syscall\n\
293 -D run tracer process as a detached grandchild, not as parent\n\
295 -ff follow forks with output into separate files\n\
297 1: no signals are blocked\n\
298 2: fatal signals are blocked while decoding syscall (default)\n\
299 3: fatal signals are always blocked (default if '-o FILE PROG')\n\
300 4: fatal signals and SIGTSTP (^Z) are always blocked\n\
301 (useful to make 'strace -o FILE PROG' not stop on ^Z)\n\
304 -E var remove var from the environment for command\n\
305 -E var=val put var=val in the environment for command\n\
306 -p pid trace process with process id PID, may be repeated\n\
307 -u username run command as username handling setuid and/or setgid\n\
310 -d enable debug output to stderr\n\
311 -h print help message\n\
314 /* ancient, no one should use it
315 -F -- attempt to follow vforks (deprecated, use -f)\n\
317 , DEFAULT_ACOLUMN, DEFAULT_STRLEN, DEFAULT_SORTBY);
323 void ATTRIBUTE_NORETURN
326 if (strace_tracer_pid == getpid()) {
335 error_opt_arg(int opt, const char *arg)
337 error_msg_and_help("invalid -%c argument: '%s'", opt, arg);
340 static const char *ptrace_attach_cmd;
343 ptrace_attach_or_seize(int pid)
347 return ptrace_attach_cmd = "PTRACE_ATTACH",
348 ptrace(PTRACE_ATTACH, pid, 0L, 0L);
349 r = ptrace(PTRACE_SEIZE, pid, 0L, (unsigned long) ptrace_setoptions);
351 return ptrace_attach_cmd = "PTRACE_SEIZE", r;
352 r = ptrace(PTRACE_INTERRUPT, pid, 0L, 0L);
353 return ptrace_attach_cmd = "PTRACE_INTERRUPT", r;
357 ptrace_op_str(unsigned int op)
359 const char *str = xlookup(ptrace_cmds, op);
363 static char buf[sizeof(op) * 3];
364 xsprintf(buf, "%u", op);
369 * Used when we want to unblock stopped traced process.
370 * Should be only used with PTRACE_CONT, PTRACE_DETACH and PTRACE_SYSCALL.
371 * Returns 0 on success or if error was ESRCH
372 * (presumably process was killed while we talk to it).
373 * Otherwise prints error message and returns -1.
376 ptrace_restart(const unsigned int op, struct tcb *const tcp, unsigned int sig)
381 ptrace(op, tcp->pid, 0L, (unsigned long) sig);
383 if (!err || err == ESRCH)
387 * Why curcol != 0? Otherwise sometimes we get this:
389 * 10252 kill(10253, SIGKILL) = 0
390 * <ptrace(SYSCALL,10252):No such process>10253 ...next decode...
392 * 10252 died after we retrieved syscall exit data,
393 * but before we tried to restart it. Log looks ugly.
395 if (current_tcp && current_tcp->curcol != 0) {
396 tprintf(" <Cannot restart pid %d with ptrace(%s): %s>\n",
397 tcp->pid, ptrace_op_str(op), strerror(err));
401 perror_msg("ptrace(%s,pid:%d,sig:%u)",
402 ptrace_op_str(op), tcp->pid, sig);
407 set_cloexec_flag(int fd)
411 flags = fcntl(fd, F_GETFD);
413 /* Can happen only if fd is bad.
414 * Should never happen: if it does, we have a bug
415 * in the caller. Therefore we just abort
416 * instead of propagating the error.
418 perror_msg_and_die("fcntl(%d, F_GETFD)", fd);
421 newflags = flags | FD_CLOEXEC;
422 if (flags == newflags)
425 if (fcntl(fd, F_SETFD, newflags)) /* never fails */
426 perror_msg_and_die("fcntl(%d, F_SETFD, %#x)", fd, newflags);
430 * When strace is setuid executable, we have to swap uids
431 * before and after filesystem and process management operations.
436 int euid = geteuid(), uid = getuid();
438 if (euid != uid && setreuid(euid, uid) < 0) {
439 perror_msg_and_die("setreuid");
444 strace_fopen(const char *path)
449 fp = fopen_stream(path, open_append ? "a" : "w");
451 perror_msg_and_die("Can't fopen '%s'", path);
453 set_cloexec_flag(fileno(fp));
457 static int popen_pid;
460 # define _PATH_BSHELL "/bin/sh"
464 * We cannot use standard popen(3) here because we have to distinguish
465 * popen child process from other processes we trace, and standard popen(3)
466 * does not export its child's pid.
469 strace_popen(const char *command)
477 perror_msg_and_die("pipe");
479 set_cloexec_flag(fds[1]); /* never fails */
483 perror_msg_and_die("vfork");
490 perror_msg_and_die("dup2");
493 execl(_PATH_BSHELL, "sh", "-c", command, NULL);
494 perror_msg_and_die("Can't execute '%s'", _PATH_BSHELL);
501 fp = fdopen(fds[1], "w");
503 perror_msg_and_die("fdopen");
508 outf_perror(const struct tcb * const tcp)
510 if (tcp->outf == stderr)
513 /* This is ugly, but we don't store separate file names */
515 perror_msg("%s.%u", outfname, tcp->pid);
517 perror_msg("%s", outfname);
520 ATTRIBUTE_FORMAT((printf, 1, 0))
522 tvprintf(const char *const fmt, va_list args)
525 int n = vfprintf(current_tcp->outf, fmt, args);
527 /* very unlikely due to vfprintf buffering */
528 outf_perror(current_tcp);
530 current_tcp->curcol += n;
535 tprintf(const char *fmt, ...)
543 #ifndef HAVE_FPUTS_UNLOCKED
544 # define fputs_unlocked fputs
548 tprints(const char *str)
551 int n = fputs_unlocked(str, current_tcp->outf);
553 current_tcp->curcol += strlen(str);
556 /* very unlikely due to fputs_unlocked buffering */
557 outf_perror(current_tcp);
562 tprints_comment(const char *const str)
565 tprintf(" /* %s */", str);
569 tprintf_comment(const char *fmt, ...)
583 flush_tcp_output(const struct tcb *const tcp)
585 if (fflush(tcp->outf))
593 current_tcp->curcol = 0;
594 flush_tcp_output(current_tcp);
597 printing_tcp->curcol = 0;
603 set_current_tcp(const struct tcb *tcp)
605 current_tcp = (struct tcb *) tcp;
607 /* Sync current_personality and stuff */
609 set_personality(current_tcp->currpers);
613 printleader(struct tcb *tcp)
615 /* If -ff, "previous tcb we printed" is always the same as current,
616 * because we have per-tcb output files.
622 set_current_tcp(printing_tcp);
623 if (!tcp->staged_output_data && printing_tcp->curcol != 0 &&
624 (followfork < 2 || printing_tcp == tcp)) {
626 * case 1: we have a shared log (i.e. not -ff), and last line
627 * wasn't finished (same or different tcb, doesn't matter).
628 * case 2: split log, we are the same tcb, but our last line
629 * didn't finish ("SIGKILL nuked us after syscall entry" etc).
631 tprints(" <unfinished ...>\n");
632 printing_tcp->curcol = 0;
637 set_current_tcp(tcp);
638 current_tcp->curcol = 0;
641 tprintf("%-5d ", tcp->pid);
642 else if (nprocs > 1 && !outfname)
643 tprintf("[pid %5u] ", tcp->pid);
647 clock_gettime(CLOCK_REALTIME, &ts);
650 tprintf("%lld.%06ld ",
651 (long long) ts.tv_sec, (long) ts.tv_nsec / 1000);
653 time_t local = ts.tv_sec;
654 char str[MAX(sizeof("HH:MM:SS"), sizeof(ts.tv_sec) * 3)];
655 struct tm *tm = localtime(&local);
658 strftime(str, sizeof(str), "%T", tm);
660 xsprintf(str, "%lld", (long long) local);
663 str, (long) ts.tv_nsec / 1000);
671 clock_gettime(CLOCK_MONOTONIC, &ts);
673 static struct timespec ots;
678 ts_sub(&dts, &ts, &ots);
681 tprintf("%s%6ld.%06ld%s ",
683 (long) dts.tv_sec, (long) dts.tv_nsec / 1000,
688 print_instruction_pointer(tcp);
694 if (current_tcp->curcol < acolumn)
695 tprints(acolumn_spaces + current_tcp->curcol);
698 /* Should be only called directly *after successful attach* to a tracee.
699 * Otherwise, "strace -oFILE -ff -p<nonexistant_pid>"
700 * may create bogus empty FILE.<nonexistant_pid>, and then die.
703 after_successful_attach(struct tcb *tcp, const unsigned int flags)
705 tcp->flags |= TCB_ATTACHED | TCB_STARTUP | flags;
706 tcp->outf = shared_log; /* if not -ff mode, the same file is for all */
707 if (followfork >= 2) {
709 xsprintf(name, "%s.%u", outfname, tcp->pid);
710 tcp->outf = strace_fopen(name);
713 #ifdef ENABLE_STACKTRACE
714 if (stack_trace_enabled)
715 unwind_tcb_init(tcp);
722 /* Allocate some (more) TCBs (and expand the table).
723 We don't want to relocate the TCBs because our
724 callers have pointers and it would be a pain.
725 So tcbtab is a table of pointers. Since we never
726 free the TCBs, we allocate a single chunk of many. */
727 size_t old_tcbtabsize;
729 struct tcb **tcb_ptr;
731 old_tcbtabsize = tcbtabsize;
733 tcbtab = xgrowarray(tcbtab, &tcbtabsize, sizeof(tcbtab[0]));
734 newtcbs = xcalloc(tcbtabsize - old_tcbtabsize, sizeof(newtcbs[0]));
736 for (tcb_ptr = tcbtab + old_tcbtabsize;
737 tcb_ptr < tcbtab + tcbtabsize; tcb_ptr++, newtcbs++)
747 if (nprocs == tcbtabsize)
750 for (i = 0; i < tcbtabsize; i++) {
753 memset(tcp, 0, sizeof(*tcp));
754 list_init(&tcp->wait_list);
756 #if SUPPORTED_PERSONALITIES > 1
757 tcp->currpers = current_personality;
760 debug_msg("new tcb for pid %d, active tcbs:%d",
765 error_msg_and_die("bug in alloctcb");
769 get_tcb_priv_data(const struct tcb *tcp)
771 return tcp->_priv_data;
775 set_tcb_priv_data(struct tcb *tcp, void *const priv_data,
776 void (*const free_priv_data)(void *))
781 tcp->_free_priv_data = free_priv_data;
782 tcp->_priv_data = priv_data;
788 free_tcb_priv_data(struct tcb *tcp)
790 if (tcp->_priv_data) {
791 if (tcp->_free_priv_data) {
792 tcp->_free_priv_data(tcp->_priv_data);
793 tcp->_free_priv_data = NULL;
795 tcp->_priv_data = NULL;
800 droptcb(struct tcb *tcp)
805 if (cflag && debug_flag) {
808 ts_sub(&dt, &tcp->stime, &tcp->atime);
809 debug_func_msg("pid %d: %.9f seconds of system time spent "
810 "since attach", tcp->pid, ts_float(&dt));
814 for (p = 0; p < SUPPORTED_PERSONALITIES; ++p)
815 free(tcp->inject_vec[p]);
817 free_tcb_priv_data(tcp);
819 #ifdef ENABLE_STACKTRACE
820 if (stack_trace_enabled)
824 #ifdef HAVE_LINUX_KVM_H
825 kvm_vcpu_info_free(tcp);
829 tcp->mmap_cache->free_fn(tcp, __func__);
832 debug_msg("dropped tcb for pid %d, %d remain", tcp->pid, nprocs);
836 if (!is_complete_set(status_set, NUMBER_OF_STATUSES)) {
837 publish = is_number_in_set(STATUS_DETACHED, status_set);
838 strace_close_memstream(tcp, publish);
841 if (followfork >= 2) {
842 if (tcp->curcol != 0 && publish)
843 fprintf(tcp->outf, " <detached ...>\n");
846 if (printing_tcp == tcp && tcp->curcol != 0 && publish)
847 fprintf(tcp->outf, " <detached ...>\n");
848 flush_tcp_output(tcp);
852 if (current_tcp == tcp)
853 set_current_tcp(NULL);
854 if (printing_tcp == tcp)
857 list_remove(&tcp->wait_list);
859 memset(tcp, 0, sizeof(*tcp));
862 /* Detach traced process.
863 * Never call DETACH twice on the same process as both unattached and
864 * attached-unstopped processes give the same ESRCH. For unattached process we
865 * would SIGSTOP it and wait for its SIGSTOP notification forever.
868 detach(struct tcb *tcp)
874 * Linux wrongly insists the child be stopped
875 * before detaching. Arghh. We go through hoops
876 * to make a clean break of things.
879 if (!(tcp->flags & TCB_ATTACHED))
882 /* We attached but possibly didn't see the expected SIGSTOP.
883 * We must catch exactly one as otherwise the detached process
884 * would be left stopped (process state T).
886 if (tcp->flags & TCB_IGNORE_ONE_SIGSTOP)
889 error = ptrace(PTRACE_DETACH, tcp->pid, 0, 0);
891 /* On a clear day, you can see forever. */
894 if (errno != ESRCH) {
895 /* Shouldn't happen. */
896 perror_func_msg("ptrace(PTRACE_DETACH,%u)", tcp->pid);
899 /* ESRCH: process is either not stopped or doesn't exist. */
900 if (my_tkill(tcp->pid, 0) < 0) {
902 /* Shouldn't happen. */
903 perror_func_msg("tkill(%u,0)", tcp->pid);
904 /* else: process doesn't exist. */
907 /* Process is not stopped, need to stop it. */
910 * With SEIZE, tracee can be in group-stop already.
911 * In this state sending it another SIGSTOP does nothing.
912 * Need to use INTERRUPT.
913 * Testcase: trying to ^C a "strace -p <stopped_process>".
915 error = ptrace(PTRACE_INTERRUPT, tcp->pid, 0, 0);
919 perror_func_msg("ptrace(PTRACE_INTERRUPT,%u)", tcp->pid);
921 error = my_tkill(tcp->pid, SIGSTOP);
925 perror_func_msg("tkill(%u,SIGSTOP)", tcp->pid);
927 /* Either process doesn't exist, or some weird error. */
931 /* We end up here in three cases:
932 * 1. We sent PTRACE_INTERRUPT (use_seize case)
933 * 2. We sent SIGSTOP (!use_seize)
934 * 3. Attach SIGSTOP was already pending (TCB_IGNORE_ONE_SIGSTOP set)
938 if (waitpid(tcp->pid, &status, __WALL) < 0) {
942 * if (errno == ECHILD) break;
943 * ^^^ WRONG! We expect this PID to exist,
944 * and want to emit a message otherwise:
946 perror_func_msg("waitpid(%u)", tcp->pid);
949 if (!WIFSTOPPED(status)) {
951 * Tracee exited or was killed by signal.
952 * We shouldn't normally reach this place:
953 * we don't want to consume exit status.
954 * Consider "strace -p PID" being ^C-ed:
955 * we want merely to detach from PID.
957 * However, we _can_ end up here if tracee
962 sig = WSTOPSIG(status);
963 debug_msg("detach wait: event:%d sig:%d",
964 (unsigned) status >> 16, sig);
966 unsigned event = (unsigned)status >> 16;
967 if (event == PTRACE_EVENT_STOP /*&& sig == SIGTRAP*/) {
969 * sig == SIGTRAP: PTRACE_INTERRUPT stop.
970 * sig == other: process was already stopped
971 * with this stopping sig (see tests/detach-stopped).
972 * Looks like re-injecting this sig is not necessary
973 * in DETACH for the tracee to remain stopped.
978 * PTRACE_INTERRUPT is not guaranteed to produce
979 * the above event if other ptrace-stop is pending.
980 * See tests/detach-sleeping testcase:
981 * strace got SIGINT while tracee is sleeping.
982 * We sent PTRACE_INTERRUPT.
983 * We see syscall exit, not PTRACE_INTERRUPT stop.
984 * We won't get PTRACE_INTERRUPT stop
985 * if we would CONT now. Need to DETACH.
987 if (sig == syscall_trap_sig)
989 /* else: not sure in which case we can be here.
990 * Signal stop? Inject it while detaching.
992 ptrace_restart(PTRACE_DETACH, tcp, sig);
995 /* Note: this check has to be after use_seize check */
996 /* (else, in use_seize case SIGSTOP will be mistreated) */
997 if (sig == SIGSTOP) {
998 /* Detach, suppressing SIGSTOP */
999 ptrace_restart(PTRACE_DETACH, tcp, 0);
1002 if (sig == syscall_trap_sig)
1004 /* Can't detach just yet, may need to wait for SIGSTOP */
1005 error = ptrace_restart(PTRACE_CONT, tcp, sig);
1007 /* Should not happen.
1008 * Note: ptrace_restart returns 0 on ESRCH, so it's not it.
1009 * ptrace_restart already emitted error message.
1016 if (!qflag && (tcp->flags & TCB_ATTACHED))
1017 error_msg("Process %u detached", tcp->pid);
1023 process_opt_p_list(char *opt)
1027 * We accept -p PID,PID; -p "`pidof PROG`"; -p "`pgrep PROG`".
1028 * pidof uses space as delim, pgrep uses newline. :(
1031 char *delim = opt + strcspn(opt, "\n\t ,");
1035 pid = string_to_uint(opt);
1037 error_msg_and_die("Invalid process id: '%s'", opt);
1039 if (pid == strace_tracer_pid) {
1040 error_msg_and_die("I'm sorry, I can't let you do that, Dave.");
1051 attach_tcb(struct tcb *const tcp)
1053 if (ptrace_attach_or_seize(tcp->pid) < 0) {
1054 perror_msg("attach: ptrace(%s, %d)",
1055 ptrace_attach_cmd, tcp->pid);
1060 after_successful_attach(tcp, TCB_GRABBED | post_attach_sigstop);
1061 debug_msg("attach to pid %d (main) succeeded", tcp->pid);
1063 static const char task_path[] = "/proc/%d/task";
1064 char procdir[sizeof(task_path) + sizeof(int) * 3];
1066 unsigned int ntid = 0, nerr = 0;
1068 if (followfork && tcp->pid != strace_child &&
1069 xsprintf(procdir, task_path, tcp->pid) > 0 &&
1070 (dir = opendir(procdir)) != NULL) {
1073 while ((de = read_dir(dir)) != NULL) {
1074 if (de->d_fileno == 0)
1077 int tid = string_to_uint(de->d_name);
1078 if (tid <= 0 || tid == tcp->pid)
1082 if (ptrace_attach_or_seize(tid) < 0) {
1084 debug_perror_msg("attach: ptrace(%s, %d)",
1085 ptrace_attach_cmd, tid);
1089 after_successful_attach(alloctcb(tid),
1090 TCB_GRABBED | post_attach_sigstop);
1091 debug_msg("attach to pid %d succeeded", tid);
1099 error_msg("Process %u attached"
1101 tcp->pid, ntid - nerr + 1);
1103 error_msg("Process %u attached",
1109 startup_attach(void)
1111 pid_t parent_pid = strace_tracer_pid;
1115 if (daemonized_tracer) {
1118 perror_func_msg_and_die("fork");
1120 if (pid) { /* parent */
1122 * Wait for grandchild to attach to straced process
1123 * (grandparent). Grandchild SIGKILLs us after it attached.
1124 * Grandparent's wait() is unblocked by our death,
1125 * it proceeds to exec the straced program.
1128 _exit(0); /* paranoia */
1131 /* We will be the tracer process. Remember our new pid: */
1132 strace_tracer_pid = getpid();
1135 for (tcbi = 0; tcbi < tcbtabsize; tcbi++) {
1141 /* Is this a process we should attach to, but not yet attached? */
1142 if (tcp->flags & TCB_ATTACHED)
1143 continue; /* no, we already attached it */
1145 if (tcp->pid == parent_pid || tcp->pid == strace_tracer_pid) {
1147 perror_msg("attach: pid %d", tcp->pid);
1156 } /* for each tcbtab[] */
1158 if (daemonized_tracer) {
1160 * Make parent go away.
1161 * Also makes grandparent's wait() unblock.
1163 kill(parent_pid, SIGKILL);
1168 /* Stack-o-phobic exec helper, in the hope to work around
1169 * NOMMU + "daemonized tracer" difficulty.
1171 struct exec_params {
1177 struct sigaction child_sa;
1179 static struct exec_params params_for_tracee;
1181 static void ATTRIBUTE_NOINLINE ATTRIBUTE_NORETURN
1184 struct exec_params *params = ¶ms_for_tracee;
1186 if (params->fd_to_close >= 0)
1187 close(params->fd_to_close);
1188 if (!daemonized_tracer && !use_seize) {
1189 if (ptrace(PTRACE_TRACEME, 0L, 0L, 0L) < 0) {
1190 perror_msg_and_die("ptrace(PTRACE_TRACEME, ...)");
1194 if (username != NULL) {
1196 * It is important to set groups before we
1197 * lose privileges on setuid.
1199 if (initgroups(username, run_gid) < 0) {
1200 perror_msg_and_die("initgroups");
1202 if (setregid(run_gid, params->run_egid) < 0) {
1203 perror_msg_and_die("setregid");
1205 if (setreuid(run_uid, params->run_euid) < 0) {
1206 perror_msg_and_die("setreuid");
1208 } else if (geteuid() != 0)
1209 if (setreuid(run_uid, run_uid) < 0) {
1210 perror_msg_and_die("setreuid");
1213 if (!daemonized_tracer) {
1215 * Induce a ptrace stop. Tracer (our parent)
1216 * will resume us with PTRACE_SYSCALL and display
1217 * the immediately following execve syscall.
1218 * Can't do this on NOMMU systems, we are after
1219 * vfork: parent is blocked, stopping would deadlock.
1222 kill(getpid(), SIGSTOP);
1225 /* we depend on SIGCHLD set to SIG_DFL by init code */
1226 /* if it happens to be SIG_IGN'ed, wait won't block */
1231 if (params_for_tracee.child_sa.sa_handler != SIG_DFL)
1232 sigaction(SIGCHLD, ¶ms_for_tracee.child_sa, NULL);
1234 execv(params->pathname, params->argv);
1235 perror_msg_and_die("exec");
1239 * Open a dummy descriptor for use as a placeholder.
1240 * The descriptor is O_RDONLY with FD_CLOEXEC flag set.
1241 * A read attempt from such descriptor ends with EOF,
1242 * a write attempt is rejected with EBADF.
1245 open_dummy_desc(void)
1250 perror_func_msg_and_die("pipe");
1252 set_cloexec_flag(fds[0]);
1256 /* placeholder fds status for stdin and stdout */
1257 static bool fd_is_placeholder[2];
1260 * Ensure that all standard file descriptors are open by opening placeholder
1261 * file descriptors for those standard file descriptors that are not open.
1263 * The information which descriptors have been made open is saved
1264 * in fd_is_placeholder for later use.
1267 ensure_standard_fds_opened(void)
1271 while ((fd = open_dummy_desc()) <= 2) {
1274 fd_is_placeholder[fd] = true;
1282 * Redirect stdin and stdout unless they have been opened earlier
1283 * by ensure_standard_fds_opened as placeholders.
1286 redirect_standard_fds(void)
1291 * It might be a good idea to redirect stderr as well,
1292 * but we sometimes need to print error messages.
1294 for (i = 0; i <= 1; ++i) {
1295 if (!fd_is_placeholder[i]) {
1303 startup_child(char **argv)
1305 strace_stat_t statbuf;
1306 const char *filename;
1307 size_t filename_len;
1308 char pathname[PATH_MAX];
1313 filename_len = strlen(filename);
1315 if (filename_len > sizeof(pathname) - 1) {
1316 errno = ENAMETOOLONG;
1317 perror_msg_and_die("exec");
1319 if (strchr(filename, '/')) {
1320 strcpy(pathname, filename);
1322 #ifdef USE_DEBUGGING_EXEC
1324 * Debuggers customarily check the current directory
1325 * first regardless of the path but doing that gives
1326 * security geeks a panic attack.
1328 else if (stat_file(filename, &statbuf) == 0)
1329 strcpy(pathname, filename);
1330 #endif /* USE_DEBUGGING_EXEC */
1335 for (path = getenv("PATH"); path && *path; path += m) {
1336 const char *colon = strchr(path, ':');
1341 m = n = strlen(path);
1343 if (!getcwd(pathname, PATH_MAX))
1345 len = strlen(pathname);
1346 } else if (n > sizeof(pathname) - 1)
1349 strncpy(pathname, path, n);
1352 if (len && pathname[len - 1] != '/')
1353 pathname[len++] = '/';
1354 if (filename_len + len > sizeof(pathname) - 1)
1356 strcpy(pathname + len, filename);
1357 if (stat_file(pathname, &statbuf) == 0 &&
1358 /* Accept only regular files
1359 with some execute bits set.
1360 XXX not perfect, might still fail */
1361 S_ISREG(statbuf.st_mode) &&
1362 (statbuf.st_mode & 0111))
1365 if (!path || !*path)
1368 if (stat_file(pathname, &statbuf) < 0) {
1369 perror_msg_and_die("Can't stat '%s'", filename);
1372 params_for_tracee.fd_to_close = (shared_log != stderr) ? fileno(shared_log) : -1;
1373 params_for_tracee.run_euid = (statbuf.st_mode & S_ISUID) ? statbuf.st_uid : run_uid;
1374 params_for_tracee.run_egid = (statbuf.st_mode & S_ISGID) ? statbuf.st_gid : run_gid;
1375 params_for_tracee.argv = argv;
1377 * On NOMMU, can be safely freed only after execve in tracee.
1378 * It's hard to know when that happens, so we just leak it.
1380 params_for_tracee.pathname = NOMMU_SYSTEM ? xstrdup(pathname) : pathname;
1382 #if defined HAVE_PRCTL && defined PR_SET_PTRACER && defined PR_SET_PTRACER_ANY
1383 if (daemonized_tracer)
1384 prctl(PR_SET_PTRACER, PR_SET_PTRACER_ANY);
1389 perror_func_msg_and_die("fork");
1391 if ((pid != 0 && daemonized_tracer)
1392 || (pid == 0 && !daemonized_tracer)
1394 /* We are to become the tracee. Two cases:
1396 * not -D: we are child
1401 /* We are the tracer */
1403 if (!daemonized_tracer) {
1406 /* child did PTRACE_TRACEME, nothing to do in parent */
1408 if (!NOMMU_SYSTEM) {
1409 /* Wait until child stopped itself */
1411 while (waitpid(pid, &status, WSTOPPED) < 0) {
1414 perror_msg_and_die("waitpid");
1416 if (!WIFSTOPPED(status) || WSTOPSIG(status) != SIGSTOP) {
1417 kill_save_errno(pid, SIGKILL);
1418 perror_msg_and_die("Unexpected wait status %#x",
1422 /* Else: NOMMU case, we have no way to sync.
1423 * Just attach to it as soon as possible.
1424 * This means that we may miss a few first syscalls...
1427 if (ptrace_attach_or_seize(pid)) {
1428 kill_save_errno(pid, SIGKILL);
1429 perror_msg_and_die("attach: ptrace(%s, %d)",
1430 ptrace_attach_cmd, pid);
1435 tcp = alloctcb(pid);
1436 after_successful_attach(tcp, TCB_SKIP_DETACH_ON_FIRST_EXEC
1439 | post_attach_sigstop)));
1441 /* With -D, we are *child* here, the tracee is our parent. */
1442 strace_child = strace_tracer_pid;
1443 strace_tracer_pid = getpid();
1444 tcp = alloctcb(strace_child);
1445 tcp->flags |= TCB_SKIP_DETACH_ON_FIRST_EXEC | TCB_HIDE_LOG;
1447 * Attaching will be done later, by startup_attach.
1448 * Note: we don't do after_successful_attach() here either!
1451 /* NOMMU BUG! -D mode is active, we (child) return,
1452 * and we will scribble over parent's stack!
1453 * When parent later unpauses, it segfaults.
1456 * (1) by declaring exec_or_die() NORETURN,
1457 * hopefully compiler will just jump to it
1458 * instead of call (won't push anything to stack),
1459 * (2) by trying very hard in exec_or_die()
1460 * to not use any stack,
1461 * (3) having a really big (PATH_MAX) stack object
1462 * in this function, which creates a "buffer" between
1463 * child's and parent's stack pointers.
1464 * This may save us if (1) and (2) failed
1465 * and compiler decided to use stack in exec_or_die() anyway
1466 * (happens on i386 because of stack parameter passing).
1468 * A cleaner solution is to use makecontext + setcontext
1469 * to create a genuine separate stack and execute on it.
1473 * A case where straced process is part of a pipe:
1474 * { sleep 1; yes | head -n99999; } | strace -o/dev/null sh -c 'exec <&-; sleep 9'
1475 * If strace won't close its fd#0, closing it in tracee is not enough:
1476 * the pipe is still open, it has a reader. Thus, "head" will not get its
1477 * SIGPIPE at once, on the first write.
1479 * Preventing it by redirecting strace's stdin/out.
1480 * (Don't leave fds 0 and 1 closed, this is bad practice: future opens
1481 * will reuse them, unexpectedly making a newly opened object "stdin").
1483 redirect_standard_fds();
1487 test_ptrace_seize(void)
1491 /* Need fork for test. NOMMU has no forks */
1493 post_attach_sigstop = 0; /* this sets use_seize to 1 */
1499 perror_func_msg_and_die("fork");
1506 /* PTRACE_SEIZE, unlike ATTACH, doesn't force tracee to trap. After
1507 * attaching tracee continues to run unless a trap condition occurs.
1508 * PTRACE_SEIZE doesn't affect signal or group stop state.
1510 if (ptrace(PTRACE_SEIZE, pid, 0, 0) == 0) {
1511 post_attach_sigstop = 0; /* this sets use_seize to 1 */
1513 debug_msg("PTRACE_SEIZE doesn't work");
1519 int status, tracee_pid;
1522 tracee_pid = waitpid(pid, &status, 0);
1523 if (tracee_pid <= 0) {
1526 perror_func_msg_and_die("unexpected wait result %d",
1529 if (WIFSIGNALED(status))
1532 error_func_msg_and_die("unexpected wait status %#x", status);
1537 get_os_release(void)
1541 perror_msg_and_die("uname");
1543 * u.release string consists of at most three parts
1544 * and normally has this form: "3.2.9[-some-garbage]",
1545 * "X.Y-something" means "X.Y.0".
1547 const char *p = u.release;
1548 unsigned int rel = 0;
1549 for (unsigned int parts = 0; parts < 3; ++parts) {
1551 for (; (*p >= '0') && (*p <= '9'); ++p) {
1564 set_sighandler(int signo, void (*sighandler)(int), struct sigaction *oldact)
1566 const struct sigaction sa = { .sa_handler = sighandler };
1567 sigaction(signo, &sa, oldact);
1571 * Initialization part of main() was eating much stack (~0.5k),
1572 * which was unused after init.
1573 * We can reuse it if we move init code into a separate function.
1575 * Don't want main() to inline us and defeat the reason
1576 * we have a separate function.
1578 static void ATTRIBUTE_NOINLINE
1579 init(int argc, char *argv[])
1582 int optF = 0, zflags = 0;
1584 if (!program_invocation_name || !*program_invocation_name) {
1585 static char name[] = "strace";
1586 program_invocation_name =
1587 (argc > 0 && argv[0] && *argv[0]) ? argv[0] : name;
1590 strace_tracer_pid = getpid();
1592 os_release = get_os_release();
1594 shared_log = stderr;
1595 set_sortby(DEFAULT_SORTBY);
1596 set_personality(DEFAULT_PERSONALITY);
1597 qualify("trace=all");
1598 qualify("abbrev=all");
1599 qualify("verbose=all");
1600 qualify("status=all");
1601 #if DEFAULT_QUAL_FLAGS != (QUAL_TRACE | QUAL_ABBREV | QUAL_VERBOSE)
1602 # error Bug in DEFAULT_QUAL_FLAGS
1604 qualify("signal=all");
1605 while ((c = getopt(argc, argv, "+"
1606 #ifdef ENABLE_STACKTRACE
1609 "a:Ab:cCdDe:E:fFhiI:o:O:p:P:qrs:S:tTu:vVwxX:yzZ")) != EOF) {
1612 acolumn = string_to_uint(optarg);
1614 error_opt_arg(c, optarg);
1620 if (strcmp(optarg, "execve") != 0)
1621 error_msg_and_die("Syscall '%s' for -b isn't supported",
1623 detach_on_execve = 1;
1626 if (cflag == CFLAG_BOTH) {
1627 error_msg_and_help("-c and -C are mutually exclusive");
1629 cflag = CFLAG_ONLY_STATS;
1632 if (cflag == CFLAG_ONLY_STATS) {
1633 error_msg_and_help("-c and -C are mutually exclusive");
1641 daemonized_tracer = 1;
1647 if (putenv(optarg) < 0)
1648 perror_msg_and_die("putenv");
1663 opt_intr = string_to_uint_upto(optarg, NUM_INTR_OPTS - 1);
1665 error_opt_arg(c, optarg);
1667 #ifdef ENABLE_STACKTRACE
1669 stack_trace_enabled = true;
1676 if (set_overhead(optarg) < 0)
1677 error_opt_arg(c, optarg);
1680 process_opt_p_list(optarg);
1683 pathtrace_select(optarg);
1692 i = string_to_uint(optarg);
1693 if (i < 0 || (unsigned int) i > -1U / 4)
1694 error_opt_arg(c, optarg);
1710 qualify("abbrev=none");
1717 count_wallclock = 1;
1723 if (!strcmp(optarg, "raw"))
1724 xlat_verbosity = XLAT_STYLE_RAW;
1725 else if (!strcmp(optarg, "abbrev"))
1726 xlat_verbosity = XLAT_STYLE_ABBREV;
1727 else if (!strcmp(optarg, "verbose"))
1728 xlat_verbosity = XLAT_STYLE_VERBOSE;
1730 error_opt_arg(c, optarg);
1736 clear_number_set_array(status_set, 1);
1737 add_number_to_set(STATUS_SUCCESSFUL, status_set);
1741 clear_number_set_array(status_set, 1);
1742 add_number_to_set(STATUS_FAILED, status_set);
1746 error_msg_and_help(NULL);
1754 if (argc < 0 || (!nprocs && !argc)) {
1755 error_msg_and_help("must have PROG [ARGS] or -p PID");
1758 if (!argc && daemonized_tracer) {
1759 error_msg_and_help("PROG [ARGS] must be specified with -D");
1764 error_msg("deprecated option -F ignored");
1766 error_msg("option -F is deprecated, "
1767 "please use -f instead");
1772 if (followfork >= 2 && cflag) {
1773 error_msg_and_help("(-c or -C) and -ff are mutually exclusive");
1776 if (count_wallclock && !cflag) {
1777 error_msg_and_help("-w must be given with (-c or -C)");
1780 if (cflag == CFLAG_ONLY_STATS) {
1782 error_msg("-%c has no effect with -c", 'i');
1783 if (stack_trace_enabled)
1784 error_msg("-%c has no effect with -c", 'k');
1786 error_msg("-%c has no effect with -c", 'r');
1788 error_msg("-%c has no effect with -c", 't');
1790 error_msg("-%c has no effect with -c", 'T');
1792 error_msg("-%c has no effect with -c", 'y');
1795 #ifndef HAVE_OPEN_MEMSTREAM
1796 if (!is_complete_set(status_set, NUMBER_OF_STATUSES))
1797 error_msg_and_help("open_memstream is required to use -z, -Z, or -e status");
1801 error_msg("Only the last of -z/-Z options will take effect. "
1802 "See status qualifier for more complex filters.");
1804 acolumn_spaces = xmalloc(acolumn + 1);
1805 memset(acolumn_spaces, ' ', acolumn);
1806 acolumn_spaces[acolumn] = '\0';
1808 set_sighandler(SIGCHLD, SIG_DFL, ¶ms_for_tracee.child_sa);
1810 #ifdef ENABLE_STACKTRACE
1811 if (stack_trace_enabled)
1815 /* See if they want to run as another user. */
1816 if (username != NULL) {
1817 struct passwd *pent;
1819 if (getuid() != 0 || geteuid() != 0) {
1820 error_msg_and_die("You must be root to use the -u option");
1822 pent = getpwnam(username);
1824 error_msg_and_die("Cannot find user '%s'", username);
1826 run_uid = pent->pw_uid;
1827 run_gid = pent->pw_gid;
1834 ptrace_setoptions |= PTRACE_O_TRACECLONE |
1835 PTRACE_O_TRACEFORK |
1836 PTRACE_O_TRACEVFORK;
1837 debug_msg("ptrace_setoptions = %#x", ptrace_setoptions);
1838 test_ptrace_seize();
1839 test_ptrace_get_syscall_info();
1842 * Is something weird with our stdin and/or stdout -
1843 * for example, may they be not open? In this case,
1844 * ensure that none of the future opens uses them.
1846 * This was seen in the wild when /proc/sys/kernel/core_pattern
1847 * was set to "|/bin/strace -o/tmp/LOG PROG":
1848 * kernel runs coredump helper with fd#0 open but fd#1 closed (!),
1849 * therefore LOG gets opened to fd#1, and fd#1 is closed by
1850 * "don't hold up stdin/out open" code soon after.
1852 ensure_standard_fds_opened();
1854 /* Check if they want to redirect the output. */
1856 /* See if they want to pipe the output. */
1857 if (outfname[0] == '|' || outfname[0] == '!') {
1859 * We can't do the <outfname>.PID funny business
1860 * when using popen, so prohibit it.
1862 if (followfork >= 2)
1863 error_msg_and_help("piping the output and -ff "
1864 "are mutually exclusive");
1865 shared_log = strace_popen(outfname + 1);
1866 } else if (followfork < 2) {
1867 shared_log = strace_fopen(outfname);
1868 } else if (strlen(outfname) >= PATH_MAX - sizeof(int) * 3) {
1869 errno = ENAMETOOLONG;
1870 perror_msg_and_die("%s", outfname);
1873 /* -ff without -o FILE is the same as single -f */
1874 if (followfork >= 2)
1878 if (!outfname || outfname[0] == '|' || outfname[0] == '!') {
1879 setvbuf(shared_log, NULL, _IOLBF, 0);
1883 * argv[0] -pPID -oFILE Default interactive setting
1884 * yes * 0 INTR_WHILE_WAIT
1885 * no 1 0 INTR_WHILE_WAIT
1886 * yes * 1 INTR_NEVER
1887 * no 1 1 INTR_WHILE_WAIT
1890 if (daemonized_tracer && !opt_intr)
1891 opt_intr = INTR_BLOCK_TSTP_TOO;
1892 if (outfname && argc) {
1894 opt_intr = INTR_NEVER;
1899 opt_intr = INTR_WHILE_WAIT;
1902 * startup_child() must be called before the signal handlers get
1903 * installed below as they are inherited into the spawned process.
1904 * Also we do not need to be protected by them as during interruption
1905 * in the startup_child() mode we kill the spawned process anyway.
1908 startup_child(argv);
1911 set_sighandler(SIGTTOU, SIG_IGN, NULL);
1912 set_sighandler(SIGTTIN, SIG_IGN, NULL);
1913 if (opt_intr != INTR_ANYWHERE) {
1914 if (opt_intr == INTR_BLOCK_TSTP_TOO)
1915 set_sighandler(SIGTSTP, SIG_IGN, NULL);
1917 * In interactive mode (if no -o OUTFILE, or -p PID is used),
1918 * fatal signals are handled asynchronously and acted
1919 * when waiting for process state changes.
1920 * In non-interactive mode these signals are ignored.
1922 set_sighandler(SIGHUP, interactive ? interrupt : SIG_IGN, NULL);
1923 set_sighandler(SIGINT, interactive ? interrupt : SIG_IGN, NULL);
1924 set_sighandler(SIGQUIT, interactive ? interrupt : SIG_IGN, NULL);
1925 set_sighandler(SIGPIPE, interactive ? interrupt : SIG_IGN, NULL);
1926 set_sighandler(SIGTERM, interactive ? interrupt : SIG_IGN, NULL);
1929 sigemptyset(&timer_set);
1930 sigaddset(&timer_set, SIGALRM);
1931 sigprocmask(SIG_BLOCK, &timer_set, NULL);
1932 set_sighandler(SIGALRM, timer_sighandler, NULL);
1934 if (nprocs != 0 || daemonized_tracer)
1937 /* Do we want pids printed in our -o OUTFILE?
1938 * -ff: no (every pid has its own file); or
1939 * -f: yes (there can be more pids in the future); or
1940 * -p PID1,PID2: yes (there are already more than one pid)
1942 print_pid_pfx = (outfname && followfork < 2 && (followfork == 1 || nprocs > 1));
1946 pid2tcb(const int pid)
1951 #define PID2TCB_CACHE_SIZE 1024U
1952 #define PID2TCB_CACHE_MASK (PID2TCB_CACHE_SIZE - 1)
1954 static struct tcb *pid2tcb_cache[PID2TCB_CACHE_SIZE];
1955 struct tcb **const ptcp = &pid2tcb_cache[pid & PID2TCB_CACHE_MASK];
1956 struct tcb *tcp = *ptcp;
1958 if (tcp && tcp->pid == pid)
1961 for (unsigned int i = 0; i < tcbtabsize; ++i) {
1963 if (tcp->pid == pid)
1971 cleanup(int fatal_sig)
1977 fatal_sig = SIGTERM;
1979 for (i = 0; i < tcbtabsize; i++) {
1983 debug_func_msg("looking at pid %u", tcp->pid);
1984 if (tcp->pid == strace_child) {
1985 kill(tcp->pid, SIGCONT);
1986 kill(tcp->pid, fatal_sig);
1999 print_debug_info(const int pid, int status)
2001 const unsigned int event = (unsigned int) status >> 16;
2002 char buf[sizeof("WIFEXITED,exitcode=%u") + sizeof(int)*3 /*paranoia:*/ + 16];
2003 char evbuf[sizeof(",EVENT_VFORK_DONE (%u)") + sizeof(int)*3 /*paranoia:*/ + 16];
2006 if (WIFSIGNALED(status))
2007 xsprintf(buf, "WIFSIGNALED,%ssig=%s",
2008 WCOREDUMP(status) ? "core," : "",
2009 sprintsigname(WTERMSIG(status)));
2010 if (WIFEXITED(status))
2011 xsprintf(buf, "WIFEXITED,exitcode=%u", WEXITSTATUS(status));
2012 if (WIFSTOPPED(status))
2013 xsprintf(buf, "WIFSTOPPED,sig=%s",
2014 sprintsigname(WSTOPSIG(status)));
2017 static const char *const event_names[] = {
2018 [PTRACE_EVENT_CLONE] = "CLONE",
2019 [PTRACE_EVENT_FORK] = "FORK",
2020 [PTRACE_EVENT_VFORK] = "VFORK",
2021 [PTRACE_EVENT_VFORK_DONE] = "VFORK_DONE",
2022 [PTRACE_EVENT_EXEC] = "EXEC",
2023 [PTRACE_EVENT_EXIT] = "EXIT",
2024 /* [PTRACE_EVENT_STOP (=128)] would make biggish array */
2026 const char *e = "??";
2027 if (event < ARRAY_SIZE(event_names))
2028 e = event_names[event];
2029 else if (event == PTRACE_EVENT_STOP)
2031 xsprintf(evbuf, ",EVENT_%s (%u)", e, event);
2033 error_msg("[wait(0x%06x) = %u] %s%s", status, pid, buf, evbuf);
2037 maybe_allocate_tcb(const int pid, int status)
2039 if (!WIFSTOPPED(status)) {
2040 if (detach_on_execve && pid == strace_child) {
2041 /* example: strace -bexecve sh -c 'exec true' */
2046 * This can happen if we inherited an unknown child.
2047 * Example: (sleep 1 & exec strace true)
2049 error_msg("Exit of unknown pid %u ignored", pid);
2053 /* We assume it's a fork/vfork/clone child */
2054 struct tcb *tcp = alloctcb(pid);
2055 after_successful_attach(tcp, post_attach_sigstop);
2057 error_msg("Process %d attached", pid);
2061 * This can happen if a clone call misused CLONE_PTRACE itself.
2063 * There used to be a dance around possible re-injection of
2064 * WSTOPSIG(status), but it was later removed as the only
2065 * observable stop here is the initial ptrace-stop.
2067 ptrace(PTRACE_DETACH, pid, NULL, 0L);
2068 error_msg("Detached unknown pid %d", pid);
2074 * Under Linux, execve changes pid to thread leader's pid, and we see this
2075 * changed pid on EVENT_EXEC and later, execve sysexit. Leader "disappears"
2076 * without exit notification. Let user know that, drop leader's tcb, and fix
2077 * up pid in execve thread's tcb. Effectively, execve thread's tcb replaces
2080 * BTW, leader is 'stuck undead' (doesn't report WIFEXITED on exit syscall)
2081 * in multi-threaded programs exactly in order to handle this case.
2084 maybe_switch_tcbs(struct tcb *tcp, const int pid)
2087 * PTRACE_GETEVENTMSG returns old pid starting from Linux 3.0.
2088 * On 2.6 and earlier it can return garbage.
2090 if (os_release < KERNEL_VERSION(3, 0, 0))
2093 const long old_pid = tcb_wait_tab[tcp->wait_data_idx].msg;
2095 /* Avoid truncation in pid2tcb() param passing */
2096 if (old_pid <= 0 || old_pid == pid)
2098 if ((unsigned long) old_pid > UINT_MAX)
2100 struct tcb *execve_thread = pid2tcb(old_pid);
2101 /* It should be !NULL, but I feel paranoid */
2105 if (execve_thread->curcol != 0) {
2107 * One case we are here is -ff, try
2108 * "strace -oLOG -ff test/threaded_execve".
2109 * Another case is demonstrated by
2110 * tests/maybe_switch_current_tcp.c
2112 fprintf(execve_thread->outf, " <pid changed to %d ...>\n", pid);
2113 /*execve_thread->curcol = 0; - no need, see code below */
2115 /* Swap output FILEs and memstream (needed for -ff) */
2116 FILE *fp = execve_thread->outf;
2117 execve_thread->outf = tcp->outf;
2119 if (execve_thread->staged_output_data || tcp->staged_output_data) {
2120 struct staged_output_data *staged_output_data;
2122 staged_output_data = execve_thread->staged_output_data;
2123 execve_thread->staged_output_data = tcp->staged_output_data;
2124 tcp->staged_output_data = staged_output_data;
2127 /* And their column positions */
2128 execve_thread->curcol = tcp->curcol;
2130 /* Drop leader, but close execve'd thread outfile (if -ff) */
2132 /* Switch to the thread, reusing leader's outfile and pid */
2133 tcp = execve_thread;
2135 if (cflag != CFLAG_ONLY_STATS) {
2137 tprintf("+++ superseded by execve in pid %lu +++\n", old_pid);
2140 * Need to reopen memstream for thread
2141 * as we closed it in droptcb.
2143 if (!is_complete_set(status_set, NUMBER_OF_STATUSES))
2144 strace_open_memstream(tcp);
2145 tcp->flags |= TCB_REPRINT;
2152 maybe_switch_current_tcp(void)
2154 struct tcb *tcp = maybe_switch_tcbs(current_tcp, current_tcp->pid);
2157 set_current_tcp(tcp);
2163 print_signalled(struct tcb *tcp, const int pid, int status)
2165 if (pid == strace_child) {
2166 exit_code = 0x100 | WTERMSIG(status);
2170 if (cflag != CFLAG_ONLY_STATS
2171 && is_number_in_set(WTERMSIG(status), signal_set)) {
2173 tprintf("+++ killed by %s %s+++\n",
2174 sprintsigname(WTERMSIG(status)),
2175 WCOREDUMP(status) ? "(core dumped) " : "");
2181 print_exited(struct tcb *tcp, const int pid, int status)
2183 if (pid == strace_child) {
2184 exit_code = WEXITSTATUS(status);
2188 if (cflag != CFLAG_ONLY_STATS &&
2191 tprintf("+++ exited with %d +++\n", WEXITSTATUS(status));
2197 print_stopped(struct tcb *tcp, const siginfo_t *si, const unsigned int sig)
2199 if (cflag != CFLAG_ONLY_STATS
2201 && is_number_in_set(sig, signal_set)) {
2204 tprintf("--- %s ", sprintsigname(sig));
2208 tprintf("--- stopped by %s ---\n", sprintsigname(sig));
2211 #ifdef ENABLE_STACKTRACE
2212 if (stack_trace_enabled)
2213 unwind_tcb_print(tcp);
2219 startup_tcb(struct tcb *tcp)
2221 debug_msg("pid %d has TCB_STARTUP, initializing it", tcp->pid);
2223 tcp->flags &= ~TCB_STARTUP;
2226 debug_msg("setting opts 0x%x on pid %d",
2227 ptrace_setoptions, tcp->pid);
2228 if (ptrace(PTRACE_SETOPTIONS, tcp->pid, NULL, ptrace_setoptions) < 0) {
2229 if (errno != ESRCH) {
2230 /* Should never happen, really */
2231 perror_msg_and_die("PTRACE_SETOPTIONS");
2236 if ((tcp->flags & TCB_GRABBED) && (get_scno(tcp) == 1))
2237 tcp->s_prev_ent = tcp->s_ent;
2240 tcp->atime = tcp->stime;
2245 print_event_exit(struct tcb *tcp)
2247 if (entering(tcp) || filtered(tcp) || hide_log(tcp)
2248 || cflag == CFLAG_ONLY_STATS) {
2252 if (followfork < 2 && printing_tcp && printing_tcp != tcp
2253 && printing_tcp->curcol != 0) {
2254 set_current_tcp(printing_tcp);
2255 tprints(" <unfinished ...>\n");
2256 flush_tcp_output(printing_tcp);
2257 printing_tcp->curcol = 0;
2258 set_current_tcp(tcp);
2261 print_syscall_resume(tcp);
2263 if (!(tcp->sys_func_rval & RVAL_DECODED)) {
2265 * The decoder has probably decided to print something
2266 * on exiting syscall which is not going to happen.
2268 tprints(" <unfinished ...>");
2275 if (!is_complete_set(status_set, NUMBER_OF_STATUSES)) {
2276 bool publish = is_number_in_set(STATUS_UNFINISHED, status_set);
2277 strace_close_memstream(tcp, publish);
2283 trace_wait_data_size(struct tcb *tcp)
2285 return sizeof(struct tcb_wait_data);
2288 static struct tcb_wait_data *
2289 init_trace_wait_data(void *p)
2291 struct tcb_wait_data *wd = p;
2293 memset(wd, 0, sizeof(*wd));
2298 static struct tcb_wait_data *
2299 copy_trace_wait_data(const struct tcb_wait_data *wd)
2301 struct tcb_wait_data *new_wd = xmalloc(sizeof(*new_wd));
2303 memcpy(new_wd, wd, sizeof(*wd));
2309 free_trace_wait_data(struct tcb_wait_data *wd)
2315 tcb_wait_tab_check_size(const size_t size)
2317 while (size >= tcb_wait_tab_size) {
2318 tcb_wait_tab = xgrowarray(tcb_wait_tab,
2320 sizeof(tcb_wait_tab[0]));
2324 static const struct tcb_wait_data *
2330 invalidate_umove_cache();
2332 struct tcb *tcp = NULL;
2333 struct list_item *elem;
2335 static EMPTY_LIST(pending_tcps);
2336 /* Handle the queued tcbs before waiting for new events. */
2337 if (!list_is_empty(&pending_tcps))
2338 goto next_event_get_tcp;
2340 static struct tcb *extra_tcp;
2341 static size_t wait_extra_data_idx;
2342 /* Handle the extra tcb event. */
2346 tcp->wait_data_idx = wait_extra_data_idx;
2348 debug_msg("dequeued extra event for pid %u", tcp->pid);
2349 goto next_event_exit;
2353 * Used to exit simply when nprocs hits zero, but in this testcase:
2354 * int main(void) { _exit(!!fork()); }
2355 * under strace -f, parent sometimes (rarely) manages
2356 * to exit before we see the first stop of the child,
2357 * and we are losing track of it:
2358 * 19923 clone(...) = 19924
2359 * 19923 exit_group(1) = ?
2360 * 19923 +++ exited with 1 +++
2361 * Exiting only when wait() returns ECHILD works better.
2363 if (popen_pid != 0) {
2364 /* However, if -o|logger is in use, we can't do that.
2365 * Can work around that by double-forking the logger,
2366 * but that loses the ability to wait for its completion
2367 * on exit. Oh well...
2373 const bool unblock_delay_timer = is_delay_timer_armed();
2376 * The window of opportunity to handle expirations
2377 * of the delay timer opens here.
2379 * Unblock the signal handler for the delay timer
2380 * iff the delay timer is already created.
2382 if (unblock_delay_timer)
2383 sigprocmask(SIG_UNBLOCK, &timer_set, NULL);
2386 * If the delay timer has expired, then its expiration
2387 * has been handled already by the signal handler.
2389 * If the delay timer expires during wait4(),
2390 * then the system call will be interrupted and
2391 * the expiration will be handled by the signal handler.
2395 int pid = wait4(-1, &status, __WALL, (cflag ? &ru : NULL));
2396 int wait_errno = errno;
2399 * The window of opportunity to handle expirations
2400 * of the delay timer closes here.
2402 * Block the signal handler for the delay timer
2403 * iff it was unblocked earlier.
2405 if (unblock_delay_timer) {
2406 sigprocmask(SIG_BLOCK, &timer_set, NULL);
2412 size_t wait_tab_pos = 0;
2413 bool wait_nohang = false;
2416 * Wait for new events until wait4() returns 0 (meaning that there's
2417 * nothing more to wait for for now), or a second event for some tcb
2418 * appears (which may happen if a tracee was SIGKILL'ed, for example).
2421 struct tcb_wait_data *wd;
2424 if (wait_errno == EINTR)
2428 if (nprocs == 0 && wait_errno == ECHILD)
2431 * If nprocs > 0, ECHILD is not expected,
2432 * treat it as any other error here:
2435 perror_msg_and_die("wait4(__WALL)");
2441 if (pid == popen_pid) {
2442 if (!WIFSTOPPED(status))
2448 print_debug_info(pid, status);
2450 /* Look up 'pid' in our table. */
2454 tcp = maybe_allocate_tcb(pid, status);
2456 goto next_event_wait_next;
2460 tcp->stime.tv_sec = ru.ru_stime.tv_sec;
2461 tcp->stime.tv_nsec = ru.ru_stime.tv_usec * 1000;
2464 tcb_wait_tab_check_size(wait_tab_pos);
2466 /* Initialise a new wait data structure. */
2467 wd = tcb_wait_tab + wait_tab_pos;
2468 init_trace_wait_data(wd);
2469 wd->status = status;
2471 if (WIFSIGNALED(status)) {
2472 wd->te = TE_SIGNALLED;
2473 } else if (WIFEXITED(status)) {
2477 * As WCONTINUED flag has not been specified to wait4,
2478 * it cannot be WIFCONTINUED(status), so the only case
2479 * that remains is WIFSTOPPED(status).
2482 const unsigned int sig = WSTOPSIG(status);
2483 const unsigned int event = (unsigned int) status >> 16;
2488 * Is this post-attach SIGSTOP?
2489 * Interestingly, the process may stop
2490 * with STOPSIG equal to some other signal
2491 * than SIGSTOP if we happened to attach
2492 * just before the process takes a signal.
2494 if (sig == SIGSTOP &&
2495 (tcp->flags & TCB_IGNORE_ONE_SIGSTOP)) {
2496 debug_func_msg("ignored SIGSTOP on "
2497 "pid %d", tcp->pid);
2498 tcp->flags &= ~TCB_IGNORE_ONE_SIGSTOP;
2499 wd->te = TE_RESTART;
2500 } else if (sig == syscall_trap_sig) {
2501 wd->te = TE_SYSCALL_STOP;
2504 * True if tracee is stopped by signal
2505 * (as opposed to "tracee received
2507 * TODO: shouldn't we check for
2508 * errno == EINVAL too?
2509 * We can get ESRCH instead, you know...
2511 bool stopped = ptrace(PTRACE_GETSIGINFO,
2512 pid, 0, &wd->si) < 0;
2514 wd->te = stopped ? TE_GROUP_STOP
2515 : TE_SIGNAL_DELIVERY_STOP;
2518 case PTRACE_EVENT_STOP:
2520 * PTRACE_INTERRUPT-stop or group-stop.
2521 * PTRACE_INTERRUPT-stop has sig == SIGTRAP here.
2528 wd->te = TE_GROUP_STOP;
2531 wd->te = TE_RESTART;
2534 case PTRACE_EVENT_EXEC:
2536 * TODO: shouldn't we check for
2537 * errno == EINVAL here, too?
2538 * We can get ESRCH instead, you know...
2540 if (ptrace(PTRACE_GETEVENTMSG, pid, NULL,
2544 wd->te = TE_STOP_BEFORE_EXECVE;
2546 case PTRACE_EVENT_EXIT:
2547 wd->te = TE_STOP_BEFORE_EXIT;
2550 wd->te = TE_RESTART;
2555 error_func_msg("Tracing event hasn't been determined "
2556 "for pid %d, status %0#x", pid, status);
2558 if (!list_is_empty(&tcp->wait_list)) {
2559 wait_extra_data_idx = wait_tab_pos;
2561 debug_func_msg("queued extra pid %d", tcp->pid);
2563 tcp->wait_data_idx = wait_tab_pos;
2564 list_append(&pending_tcps, &tcp->wait_list);
2565 debug_func_msg("queued pid %d", tcp->pid);
2573 next_event_wait_next:
2574 pid = wait4(-1, &status, __WALL | WNOHANG, (cflag ? &ru : NULL));
2580 elem = list_remove_head(&pending_tcps);
2583 tcb_wait_tab_check_size(0);
2584 memset(tcb_wait_tab, 0, sizeof(*tcb_wait_tab));
2585 tcb_wait_tab->te = TE_NEXT;
2587 return tcb_wait_tab;
2589 tcp = list_elem(elem, struct tcb, wait_list);
2590 debug_func_msg("dequeued pid %d", tcp->pid);
2594 /* Is this the very first time we see this tracee stopped? */
2595 if (tcp->flags & TCB_STARTUP)
2600 /* Set current output file */
2601 set_current_tcp(tcp);
2603 return tcb_wait_tab + tcp->wait_data_idx;
2607 trace_syscall(struct tcb *tcp, unsigned int *sig)
2609 if (entering(tcp)) {
2610 int res = syscall_entering_decode(tcp);
2615 res = syscall_entering_trace(tcp, sig);
2617 syscall_entering_finish(tcp, res);
2620 struct timespec ts = {};
2621 int res = syscall_exiting_decode(tcp, &ts);
2623 res = syscall_exiting_trace(tcp, &ts, res);
2625 syscall_exiting_finish(tcp);
2630 /* Returns true iff the main trace loop has to continue. */
2632 dispatch_event(const struct tcb_wait_data *wd)
2634 unsigned int restart_op = PTRACE_SYSCALL;
2635 unsigned int restart_sig = 0;
2636 enum trace_event te = wd ? wd->te : TE_BREAK;
2638 * Copy wd->status to a non-const variable to workaround glibc bugs
2639 * around union wait fixed by glibc commit glibc-2.24~391
2641 int status = wd ? wd->status : 0;
2653 case TE_SYSCALL_STOP:
2654 if (trace_syscall(current_tcp, &restart_sig) < 0) {
2656 * ptrace() failed in trace_syscall().
2657 * Likely a result of process disappearing mid-flight.
2658 * Observed case: exit_group() or SIGKILL terminating
2659 * all processes in thread group.
2660 * We assume that ptrace error was caused by process death.
2661 * We used to detach(current_tcp) here, but since we no
2662 * longer implement "detach before death" policy/hack,
2663 * we can let this process to report its death to us
2664 * normally, via WIFEXITED or WIFSIGNALED wait status.
2670 case TE_SIGNAL_DELIVERY_STOP:
2671 restart_sig = WSTOPSIG(status);
2672 print_stopped(current_tcp, &wd->si, restart_sig);
2676 print_signalled(current_tcp, current_tcp->pid, status);
2677 droptcb(current_tcp);
2681 restart_sig = WSTOPSIG(status);
2682 print_stopped(current_tcp, NULL, restart_sig);
2685 * This ends ptrace-stop, but does *not* end group-stop.
2686 * This makes stopping signals work properly on straced
2687 * process (that is, process really stops. It used to
2690 restart_op = PTRACE_LISTEN;
2696 print_exited(current_tcp, current_tcp->pid, status);
2697 droptcb(current_tcp);
2700 case TE_STOP_BEFORE_EXECVE:
2701 /* The syscall succeeded, clear the flag. */
2702 current_tcp->flags &= ~TCB_CHECK_EXEC_SYSCALL;
2704 * Check that we are inside syscall now (next event after
2705 * PTRACE_EVENT_EXEC should be for syscall exiting). If it is
2706 * not the case, we might have a situation when we attach to a
2707 * process and the first thing we see is a PTRACE_EVENT_EXEC
2708 * and all the following syscall state tracking is screwed up
2711 if (!maybe_switch_current_tcp() && entering(current_tcp)) {
2714 error_msg("Stray PTRACE_EVENT_EXEC from pid %d"
2715 ", trying to recover...",
2718 current_tcp->flags |= TCB_RECOVERING;
2719 ret = trace_syscall(current_tcp, &restart_sig);
2720 current_tcp->flags &= ~TCB_RECOVERING;
2723 /* The reason is described in TE_SYSCALL_STOP */
2728 if (detach_on_execve) {
2729 if (current_tcp->flags & TCB_SKIP_DETACH_ON_FIRST_EXEC) {
2730 current_tcp->flags &= ~TCB_SKIP_DETACH_ON_FIRST_EXEC;
2732 detach(current_tcp); /* do "-b execve" thingy */
2738 case TE_STOP_BEFORE_EXIT:
2739 print_event_exit(current_tcp);
2743 /* We handled quick cases, we are permitted to interrupt now. */
2747 /* If the process is being delayed, do not ptrace_restart just yet */
2748 if (syscall_delayed(current_tcp)) {
2749 if (current_tcp->delayed_wait_data)
2750 error_func_msg("pid %d has delayed wait data set"
2751 " already", current_tcp->pid);
2753 current_tcp->delayed_wait_data = copy_trace_wait_data(wd);
2758 if (ptrace_restart(restart_op, current_tcp, restart_sig) < 0) {
2759 /* Note: ptrace_restart emitted error message */
2767 restart_delayed_tcb(struct tcb *const tcp)
2769 struct tcb_wait_data *wd = tcp->delayed_wait_data;
2772 error_func_msg("No delayed wait data found for pid %d",
2774 wd = init_trace_wait_data(alloca(trace_wait_data_size(tcp)));
2777 wd->te = TE_RESTART;
2779 debug_func_msg("pid %d", tcp->pid);
2781 tcp->flags &= ~TCB_DELAYED;
2783 struct tcb *const prev_tcp = current_tcp;
2785 bool ret = dispatch_event(wd);
2786 current_tcp = prev_tcp;
2788 free_trace_wait_data(tcp->delayed_wait_data);
2789 tcp->delayed_wait_data = NULL;
2795 restart_delayed_tcbs(void)
2797 struct tcb *tcp_next = NULL;
2798 struct timespec ts_now;
2800 clock_gettime(CLOCK_MONOTONIC, &ts_now);
2802 for (size_t i = 0; i < tcbtabsize; i++) {
2803 struct tcb *tcp = tcbtab[i];
2805 if (tcp->pid && syscall_delayed(tcp)) {
2806 if (ts_cmp(&ts_now, &tcp->delay_expiration_time) > 0) {
2807 if (!restart_delayed_tcb(tcp))
2810 /* Check whether this tcb is the next. */
2812 ts_cmp(&tcp_next->delay_expiration_time,
2813 &tcp->delay_expiration_time) > 0) {
2821 arm_delay_timer(tcp_next);
2827 * As this signal handler does a lot of work that is not suitable
2828 * for signal handlers, extra care must be taken to ensure that
2829 * it is enabled only in those places where it's safe.
2832 timer_sighandler(int sig)
2834 delay_timer_expired();
2839 int saved_errno = errno;
2841 if (!restart_delayed_tcbs())
2844 errno = saved_errno;
2847 #ifdef ENABLE_COVERAGE_GCOV
2848 extern void __gcov_flush(void);
2851 static void ATTRIBUTE_NORETURN
2854 int sig = interrupted;
2858 call_summary(shared_log);
2860 if (shared_log != stderr)
2863 while (waitpid(popen_pid, NULL, 0) < 0 && errno == EINTR)
2867 exit_code = 0x100 | sig;
2869 if (exit_code > 0xff) {
2870 /* Avoid potential core file clobbering. */
2871 struct_rlimit rlim = {0, 0};
2872 set_rlimit(RLIMIT_CORE, &rlim);
2874 /* Child was killed by a signal, mimic that. */
2876 signal(exit_code, SIG_DFL);
2877 #ifdef ENABLE_COVERAGE_GCOV
2882 /* Unblock the signal. */
2885 sigaddset(&mask, exit_code);
2886 #ifdef ENABLE_COVERAGE_GCOV
2889 sigprocmask(SIG_UNBLOCK, &mask, NULL);
2891 /* Paranoia - what if this signal is not fatal?
2892 Exit with 128 + signo then. */
2899 main(int argc, char *argv[])
2901 setlocale(LC_ALL, "");
2904 exit_code = !nprocs;
2906 while (dispatch_event(next_event()))