| 1 |
/* -*- linux-c -*- |
| 2 |
* |
| 3 |
* start_cmd - cleanly launch a child process |
| 4 |
* |
| 5 |
* This file is part of systemtap, and is free software. You can |
| 6 |
* redistribute it and/or modify it under the terms of the GNU General |
| 7 |
* Public License (GPL); either version 2, or (at your option) any |
| 8 |
* later version. |
| 9 |
* |
| 10 |
* Copyright (C) 2005-2021 Red Hat Inc. |
| 11 |
*/ |
| 12 |
|
| 13 |
#include "staprun.h" |
| 14 |
#include <sys/utsname.h> |
| 15 |
#include <sys/ptrace.h> |
| 16 |
#include <sys/select.h> |
| 17 |
#include <limits.h> |
| 18 |
#include <search.h> |
| 19 |
#include <wordexp.h> |
| 20 |
|
| 21 |
|
| 22 |
#define WORKAROUND_BZ467568 1 /* PR 6964; XXX: autoconf when able; also in mainloop.c */ |
| 23 |
|
| 24 |
#ifndef OPEN_MAX |
| 25 |
#define OPEN_MAX 256 |
| 26 |
#endif |
| 27 |
|
| 28 |
void |
| 29 |
closefrom(int lowfd) |
| 30 |
{ |
| 31 |
long fd, maxfd; |
| 32 |
char *endp; |
| 33 |
struct dirent *dent; |
| 34 |
DIR *dirp; |
| 35 |
|
| 36 |
/* Check for a /proc/self/fd directory. */ |
| 37 |
if ((dirp = opendir("/proc/self/fd"))) { |
| 38 |
int dir_fd = dirfd(dirp); |
| 39 |
while ((dent = readdir(dirp)) != NULL) { |
| 40 |
fd = strtol(dent->d_name, &endp, 10); |
| 41 |
if (dent->d_name != endp && *endp == '\0' |
| 42 |
&& fd >= 0 && fd < INT_MAX && fd >= lowfd |
| 43 |
&& fd != dir_fd) |
| 44 |
(void) close((int)fd); |
| 45 |
} |
| 46 |
(void) closedir(dirp); |
| 47 |
} |
| 48 |
else { |
| 49 |
/* |
| 50 |
* Here we fall back on sysconf(). Why? It is possible |
| 51 |
* /proc isn't mounted, we're out of file descriptors, |
| 52 |
* etc., which could cause the opendir() to fail. Also |
| 53 |
* note thet it is possible to open a file descriptor |
| 54 |
* and then drop the rlimit such that it is below the |
| 55 |
* open fd. |
| 56 |
*/ |
| 57 |
maxfd = sysconf(_SC_OPEN_MAX); |
| 58 |
if (maxfd < 0) |
| 59 |
maxfd = OPEN_MAX; |
| 60 |
|
| 61 |
for (fd = lowfd; fd < maxfd; fd++) |
| 62 |
(void) close((int) fd); |
| 63 |
} |
| 64 |
} |
| 65 |
|
| 66 |
#if WORKAROUND_BZ467568 |
| 67 |
/* When a SIGUSR1 signal arrives, set this variable. */ |
| 68 |
volatile sig_atomic_t usr1_interrupt = 0; |
| 69 |
|
| 70 |
static void signal_usr1(int signum) |
| 71 |
{ |
| 72 |
(void) signum; |
| 73 |
usr1_interrupt = 1; |
| 74 |
} |
| 75 |
#else |
| 76 |
static void signal_usr1(int signum) |
| 77 |
{ |
| 78 |
(void) signum; |
| 79 |
} |
| 80 |
#endif /* WORKAROUND_BZ467568 */ |
| 81 |
|
| 82 |
/* |
| 83 |
* start_cmd forks the command given on the command line with the "-c" |
| 84 |
* option. It will wait just at the cusp of the exec until the |
| 85 |
* stap module is ready to let it run. We do it this way because we |
| 86 |
* must have the pid of the forked command so it can be set to the |
| 87 |
* module and made available internally as _stp_target. PTRACE_DETACH |
| 88 |
* is sent from resume_cmd() below when it receives STP_START from |
| 89 |
* the module. |
| 90 |
*/ |
| 91 |
void start_cmd(void) |
| 92 |
{ |
| 93 |
pid_t pid; |
| 94 |
struct sigaction a; |
| 95 |
#if WORKAROUND_BZ467568 |
| 96 |
struct sigaction usr1_action, old_action; |
| 97 |
sigset_t blockmask, oldmask; |
| 98 |
#endif /* WORKAROUND_BZ467568 */ |
| 99 |
|
| 100 |
/* if we are execing a target cmd, ignore ^C in stapio */ |
| 101 |
/* and let the target cmd get it. */ |
| 102 |
memset(&a, 0, sizeof(a)); |
| 103 |
sigemptyset(&a.sa_mask); |
| 104 |
a.sa_flags = 0; |
| 105 |
a.sa_handler = SIG_IGN; |
| 106 |
sigaction(SIGINT, &a, NULL); |
| 107 |
|
| 108 |
#if WORKAROUND_BZ467568 |
| 109 |
/* Set up the mask of signals to temporarily block. */ |
| 110 |
sigemptyset (&blockmask); |
| 111 |
sigaddset (&blockmask, SIGUSR1); |
| 112 |
|
| 113 |
/* Establish the SIGUSR1 signal handler. */ |
| 114 |
memset(&usr1_action, 0, sizeof(usr1_action)); |
| 115 |
sigfillset (&usr1_action.sa_mask); |
| 116 |
usr1_action.sa_flags = 0; |
| 117 |
usr1_action.sa_handler = signal_usr1; |
| 118 |
sigaction (SIGUSR1, &usr1_action, &old_action); |
| 119 |
|
| 120 |
/* Block SIGUSR1 */ |
| 121 |
sigprocmask(SIG_BLOCK, &blockmask, &oldmask); |
| 122 |
#endif /* WORKAROUND_BZ467568 */ |
| 123 |
|
| 124 |
if ((pid = fork()) < 0) { |
| 125 |
_perr("fork"); |
| 126 |
exit(1); |
| 127 |
} else if (pid == 0) { |
| 128 |
/* We're in the target process. |
| 129 |
* Let's start the execve of target_cmd. */ |
| 130 |
int rc; |
| 131 |
wordexp_t words; |
| 132 |
char *sh_c_argv[4] = { NULL, NULL, NULL, NULL }; |
| 133 |
|
| 134 |
a.sa_handler = SIG_DFL; |
| 135 |
sigaction(SIGINT, &a, NULL); |
| 136 |
|
| 137 |
/* Close any FDs we still hold, similarly as though this were |
| 138 |
* a program being spawned due to an system("") tapset function. */ |
| 139 |
closefrom(3); |
| 140 |
|
| 141 |
/* We could call closefrom() here, to make sure we don't leak any |
| 142 |
* fds to the target, but it really isn't needed here since |
| 143 |
* close-on-exec should catch everything. We don't have the |
| 144 |
* synchronizations issues here we have with system_cmd(). */ |
| 145 |
|
| 146 |
/* Formerly, we just execl'd(sh,-c,$target_cmd). But this does't |
| 147 |
work well if target_cmd is a shell builtin. We really want to |
| 148 |
probe a new child process, not a mishmash of shell-interpreted |
| 149 |
stuff. */ |
| 150 |
rc = wordexp (target_cmd, & words, WRDE_NOCMD|WRDE_UNDEF); |
| 151 |
if (rc == WRDE_BADCHAR) |
| 152 |
{ |
| 153 |
/* The user must have used a shell metacharacter, thinking that |
| 154 |
we use system(3) to evaluate 'stap -c CMD'. We could generate |
| 155 |
an error message ... but let's just do what the user meant. |
| 156 |
rhbz 467652. */ |
| 157 |
sh_c_argv[0] = "sh"; |
| 158 |
sh_c_argv[1] = "-c"; |
| 159 |
sh_c_argv[2] = target_cmd; |
| 160 |
sh_c_argv[3] = NULL; |
| 161 |
|
| 162 |
if (read_stdin) |
| 163 |
{ |
| 164 |
/* close target_cmd's stdin to prevent a data race */ |
| 165 |
char *buf = malloc(sizeof(char) * (strlen(target_cmd) + 13)); |
| 166 |
if (buf == NULL) |
| 167 |
{ |
| 168 |
_err (_("Failed to allocate memory.\n")); |
| 169 |
_exit(1); |
| 170 |
} |
| 171 |
sprintf(buf, "%s < /dev/null", target_cmd); |
| 172 |
target_cmd = buf; |
| 173 |
} |
| 174 |
} |
| 175 |
else |
| 176 |
{ |
| 177 |
switch (rc) |
| 178 |
{ |
| 179 |
case 0: |
| 180 |
break; |
| 181 |
case WRDE_SYNTAX: |
| 182 |
_err (_("wordexp: syntax error (unmatched quotes?) in -c COMMAND\n")); |
| 183 |
_exit(1); |
| 184 |
default: |
| 185 |
_err (_("wordexp: parsing error (%d)\n"), rc); |
| 186 |
_exit (1); |
| 187 |
} |
| 188 |
if (words.we_wordc < 1) { _err ("empty -c COMMAND"); _exit (1); } |
| 189 |
} |
| 190 |
|
| 191 |
/* PR 6964: when tracing all the user space process including the |
| 192 |
child the signal will be messed due to uprobe module or utrace |
| 193 |
bug. The kernel sometimes crashes. So as an alternative |
| 194 |
approximation, we just wait here for a signal from the parent. */ |
| 195 |
|
| 196 |
dbug(1, "blocking briefly\n"); |
| 197 |
alarm(60); /* but not indefinitely */ |
| 198 |
|
| 199 |
#if WORKAROUND_BZ467568 |
| 200 |
{ |
| 201 |
/* Wait for the SIGUSR1 */ |
| 202 |
while (!usr1_interrupt) |
| 203 |
sigsuspend(&oldmask); |
| 204 |
|
| 205 |
/* Restore the old SIGUSR1 signal handler. */ |
| 206 |
sigaction (SIGUSR1, &old_action, NULL); |
| 207 |
|
| 208 |
/* Restore the original signal mask */ |
| 209 |
sigprocmask(SIG_SETMASK, &oldmask, NULL); |
| 210 |
} |
| 211 |
#else /* !WORKAROUND_BZ467568 */ |
| 212 |
rc = ptrace (PTRACE_TRACEME, 0, 0, 0); |
| 213 |
if (rc < 0) perror ("ptrace me"); |
| 214 |
raise (SIGCONT); /* Harmless; just passes control to parent. */ |
| 215 |
#endif /* !WORKAROUND_BZ467568 */ |
| 216 |
|
| 217 |
alarm(0); /* clear alarms */ |
| 218 |
dbug(1, "execing target_cmd %s\n", target_cmd); |
| 219 |
|
| 220 |
/* Note that execvp() is not a direct system call; it does a $PATH |
| 221 |
search in glibc. We would like to filter out these dummy syscalls |
| 222 |
from the utrace events seen by scripts. |
| 223 |
|
| 224 |
This filtering would be done for us for free, if we used ptrace |
| 225 |
... but see PR6964. XXX: Instead, we could open-code the |
| 226 |
$PATH search here; put the pause() afterward; and run a direct |
| 227 |
execve instead of execvp(). */ |
| 228 |
|
| 229 |
if (execvp ((sh_c_argv[0] == NULL ? words.we_wordv[0] : sh_c_argv[0]), |
| 230 |
(sh_c_argv[0] == NULL ? words.we_wordv : sh_c_argv)) < 0) |
| 231 |
perror(target_cmd); |
| 232 |
|
| 233 |
/* (There is no need to wordfree() words; they are or will be gone.) */ |
| 234 |
|
| 235 |
_exit(1); |
| 236 |
} else { |
| 237 |
/* We're in the parent. The child will parse target_cmd and |
| 238 |
execv() the result. It will be stopped thereabouts and send us |
| 239 |
a SIGTRAP. Or rather, due to PR 6964, it will stop itself and wait for |
| 240 |
us to release it. */ |
| 241 |
target_pid = pid; |
| 242 |
#if WORKAROUND_BZ467568 |
| 243 |
/* Restore the old SIGUSR1 signal handler. */ |
| 244 |
sigaction (SIGUSR1, &old_action, NULL); |
| 245 |
|
| 246 |
/* Restore the original signal mask */ |
| 247 |
sigprocmask(SIG_SETMASK, &oldmask, NULL); |
| 248 |
#else /* !WORKAROUND_BZ467568 */ |
| 249 |
int status; |
| 250 |
waitpid (target_pid, &status, 0); |
| 251 |
dbug(1, "waited for target_cmd %s pid %d status %x\n", target_cmd, target_pid, (unsigned) status); |
| 252 |
#endif /* !WORKAROUND_BZ467568 */ |
| 253 |
} |
| 254 |
} |
| 255 |
|
| 256 |
/* |
| 257 |
* resume_cmd signals the command started by start_cmd() |
| 258 |
* and identified by target_pid to continue running. |
| 259 |
* Returns 0 on success, negative PTRACE_DETACH error code on failure. */ |
| 260 |
int resume_cmd(void) |
| 261 |
{ |
| 262 |
#if WORKAROUND_BZ467568 |
| 263 |
/* Let's just send our pet signal to the child |
| 264 |
process that should be waiting for us, mid-pause(). */ |
| 265 |
kill (target_pid, SIGUSR1); |
| 266 |
#else |
| 267 |
/* Were it not for PR6964, we'd like to do it this way: */ |
| 268 |
int rc = ptrace (PTRACE_DETACH, target_pid, 0, 0); |
| 269 |
if (rc < 0) |
| 270 |
{ |
| 271 |
perror (_("ptrace detach")); |
| 272 |
if (target_cmd) |
| 273 |
kill(target_pid, SIGKILL); |
| 274 |
return rc; |
| 275 |
} |
| 276 |
#endif |
| 277 |
return 0; |
| 278 |
} |