X-Git-Url: http://mj.ucw.cz/gitweb/?a=blobdiff_plain;f=src%2Fbox.c;h=a2c5f683ce16af8dd2dd8d314917d7c3255054f2;hb=e663f0222d8299a6b802cc35983858cfeca13a70;hp=cc780cd2c174cbc04c622830dc84160b9099bccc;hpb=04cca68f101b8071504da6b304b9ffd83bc2d238;p=moe.git

diff --git a/src/box.c b/src/box.c
index cc780cd..a2c5f68 100644
--- a/src/box.c
+++ b/src/box.c
@@ -1,7 +1,7 @@
 /*
- *	A Simple Testing Sandbox
+ *	A Simple Sandbox for MO-Eval
  *
- *	(c) 2001--2007 Martin Mares <mj@ucw.cz>
+ *	(c) 2001--2008 Martin Mares <mj@ucw.cz>
  */
 
 #define _LARGEFILE64_SOURCE
@@ -27,6 +27,7 @@
 
 #define NONRET __attribute__((noreturn))
 #define UNUSED __attribute__((unused))
+#define ARRAY_SIZE(a) (int)(sizeof(a)/sizeof(a[0]))
 
 static int filter_syscalls;		/* 0=off, 1=liberal, 2=totalitarian */
 static int timeout;			/* milliseconds */
@@ -35,7 +36,6 @@ static int pass_environ;
 static int file_access;
 static int verbose;
 static int memory_limit;
-static int allow_times;
 static char *redir_stdin, *redir_stdout;
 static char *set_cwd;
 
@@ -44,6 +44,7 @@ static int is_ptraced;
 static volatile int timer_tick;
 static struct timeval start_time;
 static int ticks_per_sec;
+static int exec_seen;
 
 #if defined(__GLIBC__) && __GLIBC__ == 2 && __GLIBC_MINOR__ > 0
 /* glibc 2.1 or newer -> has lseek64 */
@@ -90,6 +91,215 @@ msg(char *msg, ...)
   va_end(args);
 }
 
+static const char * const syscall_tab[] = {
+#include "syscall-table.h"
+};
+#define NUM_SYSCALLS (sizeof(syscall_tab)/sizeof(syscall_tab[0]))
+#define NUM_ACTIONS (NUM_SYSCALLS+64)
+
+enum action {
+  SC_DEFAULT,		// Use the default action
+  SC_NO,		// Always forbid
+  SC_YES,		// Always permit
+  SC_FILENAME,		// Permit if arg1 is a known filename
+  SC_LIBERAL = 128,	// Valid only in liberal mode
+};
+
+static unsigned char syscall_action[NUM_ACTIONS] = {
+#define S(x) [__NR_##x]
+
+    // Syscalls permitted for specific file names
+    S(open) = SC_FILENAME,
+    S(creat) = SC_FILENAME,
+    S(unlink) = SC_FILENAME,
+    S(oldstat) = SC_FILENAME,
+    S(access) = SC_FILENAME,			
+    S(oldlstat) = SC_FILENAME,			
+    S(truncate) = SC_FILENAME,
+    S(stat) = SC_FILENAME,
+    S(lstat) = SC_FILENAME,
+    S(truncate64) = SC_FILENAME,
+    S(stat64) = SC_FILENAME,
+    S(lstat64) = SC_FILENAME,
+    S(readlink) = SC_FILENAME,
+
+    // Syscalls permitted always
+    S(exit) = SC_YES,
+    S(read) = SC_YES,
+    S(write) = SC_YES,
+    S(close) = SC_YES,
+    S(lseek) = SC_YES,
+    S(getpid) = SC_YES,
+    S(getuid) = SC_YES,
+    S(oldfstat) = SC_YES,
+    S(dup) = SC_YES,
+    S(brk) = SC_YES,
+    S(getgid) = SC_YES,
+    S(geteuid) = SC_YES,
+    S(getegid) = SC_YES,
+    S(dup2) = SC_YES,
+    S(ftruncate) = SC_YES,
+    S(fstat) = SC_YES,
+    S(personality) = SC_YES,
+    S(_llseek) = SC_YES,
+    S(readv) = SC_YES,
+    S(writev) = SC_YES,
+    S(getresuid) = SC_YES,
+#ifdef __NR_pread64
+    S(pread64) = SC_YES,
+    S(pwrite64) = SC_YES,
+#else
+    S(pread) = SC_YES,
+    S(pwrite) = SC_YES,
+#endif
+    S(ftruncate64) = SC_YES,
+    S(fstat64) = SC_YES,
+    S(fcntl) = SC_YES,
+    S(fcntl64) = SC_YES,
+    S(mmap) = SC_YES,
+    S(munmap) = SC_YES,
+    S(ioctl) = SC_YES,
+    S(uname) = SC_YES,
+    S(gettid) = SC_YES,
+    S(set_thread_area) = SC_YES,
+    S(get_thread_area) = SC_YES,
+    S(exit_group) = SC_YES,
+
+    // Syscalls permitted only in liberal mode
+    S(time) = SC_YES | SC_LIBERAL,
+    S(alarm) = SC_YES | SC_LIBERAL,
+    S(pause) = SC_YES | SC_LIBERAL,
+    S(signal) = SC_YES | SC_LIBERAL,
+    S(fchmod) = SC_YES | SC_LIBERAL,
+    S(sigaction) = SC_YES | SC_LIBERAL,
+    S(sgetmask) = SC_YES | SC_LIBERAL,
+    S(ssetmask) = SC_YES | SC_LIBERAL,
+    S(sigsuspend) = SC_YES | SC_LIBERAL,
+    S(sigpending) = SC_YES | SC_LIBERAL,
+    S(getrlimit) = SC_YES | SC_LIBERAL,
+    S(getrusage) = SC_YES | SC_LIBERAL,
+    S(ugetrlimit) = SC_YES | SC_LIBERAL,
+    S(gettimeofday) = SC_YES | SC_LIBERAL,
+    S(select) = SC_YES | SC_LIBERAL,
+    S(readdir) = SC_YES | SC_LIBERAL,
+    S(setitimer) = SC_YES | SC_LIBERAL,
+    S(getitimer) = SC_YES | SC_LIBERAL,
+    S(sigreturn) = SC_YES | SC_LIBERAL,
+    S(mprotect) = SC_YES | SC_LIBERAL,
+    S(sigprocmask) = SC_YES | SC_LIBERAL,
+    S(getdents) = SC_YES | SC_LIBERAL,
+    S(getdents64) = SC_YES | SC_LIBERAL,
+    S(_newselect) = SC_YES | SC_LIBERAL,
+    S(fdatasync) = SC_YES | SC_LIBERAL,
+    S(mremap) = SC_YES | SC_LIBERAL,
+    S(poll) = SC_YES | SC_LIBERAL,
+    S(getcwd) = SC_YES | SC_LIBERAL,
+    S(nanosleep) = SC_YES | SC_LIBERAL,
+    S(rt_sigreturn) = SC_YES | SC_LIBERAL,
+    S(rt_sigaction) = SC_YES | SC_LIBERAL,
+    S(rt_sigprocmask) = SC_YES | SC_LIBERAL,
+    S(rt_sigpending) = SC_YES | SC_LIBERAL,
+    S(rt_sigtimedwait) = SC_YES | SC_LIBERAL,
+    S(rt_sigqueueinfo) = SC_YES | SC_LIBERAL,
+    S(rt_sigsuspend) = SC_YES | SC_LIBERAL,
+    S(mmap2) = SC_YES | SC_LIBERAL,
+    S(_sysctl) = SC_YES | SC_LIBERAL,
+#undef S
+};
+
+static const char *
+syscall_name(unsigned int id, char *buf)
+{
+  if (id < NUM_SYSCALLS && syscall_tab[id])
+    return syscall_tab[id];
+  else
+    {
+      sprintf(buf, "#%d", id);
+      return buf;
+    }
+}
+
+static int
+syscall_by_name(char *name)
+{
+  for (unsigned int i=0; i<sizeof(syscall_tab)/sizeof(syscall_tab[0]); i++)
+    if (syscall_tab[i] && !strcmp(syscall_tab[i], name))
+      return i;
+  if (name[0] == '#')
+    name++;
+  if (!*name)
+    return -1;
+  char *ep;
+  unsigned long l = strtoul(name, &ep, 0);
+  if (*ep)
+    return -1;
+  if (l >= NUM_ACTIONS)
+    return NUM_ACTIONS;
+  return l;
+}
+
+static int
+set_action(char *a)
+{
+  char *sep = strchr(a, '=');
+  enum action act = SC_YES;
+  if (sep)
+    {
+      *sep++ = 0;
+      if (!strcmp(sep, "yes"))
+	act = SC_YES;
+      else if (!strcmp(sep, "no"))
+	act = SC_NO;
+      else if (!strcmp(sep, "file"))
+	act = SC_FILENAME;
+      else
+	return 0;
+    }
+
+  int sys = syscall_by_name(a);
+  if (sys < 0)
+    die("Unknown syscall `%s'", a);
+  if (sys >= (int)NUM_ACTIONS)
+    die("Syscall `%s' out of range", a);
+  syscall_action[sys] = act;
+  return 1;
+}
+
+struct path_rule {
+  char *path;
+  enum action action;
+  struct path_rule *next;
+};
+
+static struct path_rule default_path_rules[] = {
+  { "/etc/", SC_YES },
+  { "/lib/", SC_YES },
+  { "/usr/lib/", SC_YES },
+  { "/opt/lib/", SC_YES },
+  { "/usr/share/zoneinfo/", SC_YES },
+  { "/dev/null", SC_YES },
+  { "/dev/zero", SC_YES },
+  { "/proc/meminfo", SC_YES },
+  { "/proc/self/stat", SC_YES },
+  { "/proc/self/exe", SC_YES },			// Needed by FPC 2.0.x runtime
+};
+
+static enum action
+match_path_rule(struct path_rule *r, char *path)
+{
+  char *rr = r->path;
+  while (*rr)
+    if (*rr++ != *path++)
+      {
+	if (rr[-1] == '/' && !path[-1])
+	  break;
+	return SC_DEFAULT;
+      }
+  if (rr > r->path && rr[-1] != '/' && *path)
+    return SC_DEFAULT;
+  return r->action;
+}
+
 static void
 valid_filename(unsigned long addr)
 {
@@ -135,140 +345,59 @@ valid_filename(unsigned long addr)
   msg("[%s] ", namebuf);
   if (file_access >= 3)
     return;
+
+  // Everything in current directory is permitted
   if (!strchr(namebuf, '/') && strcmp(namebuf, ".."))
     return;
+
+  // ".." anywhere in the path is forbidden
+  enum action act = SC_DEFAULT;
+  if (strstr(namebuf, ".."))
+    act = SC_NO;
+
+  // Scan built-in rules
   if (file_access >= 2)
-    {
-      if ((!strncmp(namebuf, "/etc/", 5) ||
-	   !strncmp(namebuf, "/lib/", 5) ||
-	   !strncmp(namebuf, "/usr/lib/", 9) ||
-	   !strncmp(namebuf, "/opt/lib/", 9))
-	  && !strstr(namebuf, ".."))
-	return;
-      if (!strcmp(namebuf, "/dev/null") ||
-	  !strcmp(namebuf, "/dev/zero") ||
-	  !strcmp(namebuf, "/proc/meminfo") ||
-	  !strcmp(namebuf, "/proc/self/stat") ||
-	  !strcmp(namebuf, "/proc/self/exe") ||			/* Needed by FPC 2.0.x runtime */
-	  !strncmp(namebuf, "/usr/share/zoneinfo/", 20))
-	return;
-    }
-  die("Forbidden access to file `%s'", namebuf);
+    for (int i=0; i<ARRAY_SIZE(default_path_rules) && !act; i++)
+      act = match_path_rule(&default_path_rules[i], namebuf);
+
+  if (act != SC_YES)
+    die("Forbidden access to file `%s'", namebuf);
 }
 
 static int
 valid_syscall(struct user *u)
 {
-  switch (u->regs.orig_eax)
+  unsigned int sys = u->regs.orig_eax;
+  enum action act = (sys < NUM_ACTIONS) ? syscall_action[sys] : SC_DEFAULT;
+
+  if (act & SC_LIBERAL)
     {
-    case __NR_execve:
-      {
-	static int exec_counter;
-	return !exec_counter++;
-      }
-    case __NR_open:
-    case __NR_creat:
-    case __NR_unlink:
-    case __NR_oldstat:
-    case __NR_access:			
-    case __NR_oldlstat:			
-    case __NR_truncate:
-    case __NR_stat:
-    case __NR_lstat:
-    case __NR_truncate64:
-    case __NR_stat64:
-    case __NR_lstat64:
-    case __NR_readlink:
-      valid_filename(u->regs.ebx);
+      if (filter_syscalls == 1)
+        act &= ~SC_LIBERAL;
+      else
+        act = SC_DEFAULT;
+    }
+  switch (act)
+    {
+    case SC_YES:
       return 1;
-    case __NR_exit:
-    case __NR_read:
-    case __NR_write:
-    case __NR_close:
-    case __NR_lseek:
-    case __NR_getpid:
-    case __NR_getuid:
-    case __NR_oldfstat:
-    case __NR_dup:
-    case __NR_brk:
-    case __NR_getgid:
-    case __NR_geteuid:
-    case __NR_getegid:
-    case __NR_dup2:
-    case __NR_ftruncate:
-    case __NR_fstat:
-    case __NR_personality:
-    case __NR__llseek:
-    case __NR_readv:
-    case __NR_writev:
-    case __NR_getresuid:
-#ifdef __NR_pread64
-    case __NR_pread64:
-    case __NR_pwrite64:
-#else
-    case __NR_pread:
-    case __NR_pwrite:
-#endif
-    case __NR_ftruncate64:
-    case __NR_fstat64:
-    case __NR_fcntl:
-    case __NR_fcntl64:
-    case __NR_mmap:
-    case __NR_munmap:
-    case __NR_ioctl:
-    case __NR_uname:
-    case __NR_gettid:
-    case __NR_set_thread_area:
-    case __NR_get_thread_area:
-    case __NR_exit_group:
+    case SC_NO:
+      return 0;
+    case SC_FILENAME:
+      valid_filename(u->regs.ebx);
       return 1;
-    case __NR_time:
-    case __NR_alarm:
-    case __NR_pause:
-    case __NR_signal:
-    case __NR_fchmod:
-    case __NR_sigaction:
-    case __NR_sgetmask:
-    case __NR_ssetmask:
-    case __NR_sigsuspend:
-    case __NR_sigpending:
-    case __NR_getrlimit:
-    case __NR_getrusage:
-    case __NR_gettimeofday:
-    case __NR_select:
-    case __NR_readdir:
-    case __NR_setitimer:
-    case __NR_getitimer:
-    case __NR_sigreturn:
-    case __NR_mprotect:
-    case __NR_sigprocmask:
-    case __NR_getdents:
-    case __NR_getdents64:
-    case __NR__newselect:
-    case __NR_fdatasync:
-    case __NR_mremap:
-    case __NR_poll:
-    case __NR_getcwd:
-    case __NR_nanosleep:
-    case __NR_rt_sigreturn:
-    case __NR_rt_sigaction:
-    case __NR_rt_sigprocmask:
-    case __NR_rt_sigpending:
-    case __NR_rt_sigtimedwait:
-    case __NR_rt_sigqueueinfo:
-    case __NR_rt_sigsuspend:
-    case __NR_mmap2:
-    case __NR__sysctl:
-      return (filter_syscalls == 1);
-    case __NR_times:
-      return allow_times;
+    default: ;
+    }
+
+  switch (sys)
+    {
     case __NR_kill:
       if (u->regs.ebx == box_pid)
-	die("Commited suicide by signal %d", (int)u->regs.ecx);
+	die("Committed suicide by signal %d", (int)u->regs.ecx);
       return 0;
     case __NR_tgkill:
       if (u->regs.ebx == box_pid && u->regs.ecx == box_pid)
-	die("Commited suicide by signal %d", (int)u->regs.edx);
+	die("Committed suicide by signal %d", (int)u->regs.edx);
       return 0;
     default:
       return 0;
@@ -408,7 +537,7 @@ boxkeeper(void)
       if (WIFSIGNALED(stat))
 	{
 	  box_pid = 0;
-	  die("Caught fatal signal %d", WTERMSIG(stat));
+	  die("Caught fatal signal %d%s", WTERMSIG(stat), (syscall_count ? "" : " during startup"));
 	}
       if (WIFSTOPPED(stat))
 	{
@@ -424,9 +553,17 @@ boxkeeper(void)
 		msg(">> Traceme request caught\n");
 	      else if (stop_count & 1)		/* Syscall entry */
 		{
-		  msg(">> Syscall %3ld (%08lx,%08lx,%08lx) ", u.regs.orig_eax, u.regs.ebx, u.regs.ecx, u.regs.edx);
-		  syscall_count++;
-		  if (!valid_syscall(&u))
+		  char namebuf[32];
+		  msg(">> Syscall %-12s (%08lx,%08lx,%08lx) ", syscall_name(u.regs.orig_eax, namebuf), u.regs.ebx, u.regs.ecx, u.regs.edx);
+		  if (!exec_seen)
+		    {
+		      msg("[master] ");
+		      if (u.regs.orig_eax == __NR_execve)
+			exec_seen = 1;
+		    }
+		  else if (valid_syscall(&u))
+		    syscall_count++;
+		  else
 		    {
 		      /*
 		       * Unfortunately, PTRACE_KILL kills _after_ the syscall completes,
@@ -437,7 +574,7 @@ boxkeeper(void)
 		      u.regs.orig_eax = 0xffffffff;
 		      if (ptrace(PTRACE_SETREGS, box_pid, NULL, &u) < 0)
 			die("ptrace(PTRACE_SETREGS): %m");
-		      die("Forbidden syscall %d", sys);
+		      die("Forbidden syscall %s", syscall_name(sys, namebuf));
 		    }
 		}
 	      else					/* Syscall return */
@@ -462,7 +599,7 @@ box_inside(int argc, char **argv)
 {
   struct rlimit rl;
   char *args[argc+1];
-  char *env[1] = { NULL };
+  char *env[] = { "LIBC_FATAL_STDERR_=1", NULL };
 
   memcpy(args, argv, argc * sizeof(char *));
   args[argc] = NULL;
@@ -491,8 +628,14 @@ box_inside(int argc, char **argv)
   rl.rlim_cur = rl.rlim_max = 64;
   if (setrlimit(RLIMIT_NOFILE, &rl) < 0)
     die("setrlimit: %m");
-  if (filter_syscalls && ptrace(PTRACE_TRACEME) < 0)
-    die("ptrace(PTRACE_TRACEME): %m");
+  if (filter_syscalls)
+    {
+      if (ptrace(PTRACE_TRACEME) < 0)
+	die("ptrace(PTRACE_TRACEME): %m");
+      /* Trick: Make sure that we are stopped until the boxkeeper wakes up. */
+      signal(SIGCHLD, SIG_IGN);
+      raise(SIGCHLD);
+    }
   execve(args[0], args, (pass_environ ? environ : env));
   die("execve(\"%s\"): %m", args[0]);
 }
@@ -512,6 +655,8 @@ Options:\n\
 -i <file>\tRedirect stdin from <file>\n\
 -m <size>\tLimit address space to <size> KB\n\
 -o <file>\tRedirect stdout to <file>\n\
+-s <sys>\tPermit the specified syscall (be careful)\n\
+-s <sys>=<act>\tDefine action for the specified syscall (<act>=yes/no/file)\n\
 -t <time>\tSet run time limit (seconds, fractions allowed)\n\
 -T\t\tAllow syscalls for measuring run time\n\
 -v\t\tBe verbose\n\
@@ -526,7 +671,7 @@ main(int argc, char **argv)
   int c;
   uid_t uid;
 
-  while ((c = getopt(argc, argv, "a:c:efi:m:o:t:Tvw:")) >= 0)
+  while ((c = getopt(argc, argv, "a:c:efi:m:o:s:t:Tvw:")) >= 0)
     switch (c)
       {
       case 'a':
@@ -550,11 +695,15 @@ main(int argc, char **argv)
       case 'o':
 	redir_stdout = optarg;
 	break;
+      case 's':
+	if (!set_action(optarg))
+	  usage();
+	break;
       case 't':
 	timeout = 1000*atof(optarg);
 	break;
       case 'T':
-	allow_times++;
+	syscall_action[__NR_times] = SC_YES;
 	break;
       case 'v':
 	verbose++;