____________________________________________________________ RUS-CERT Advisory 200004-01: GNU Emacs 20 RUS-CERT, University of Stuttgart 2000-04-18 Summary Several vulnerabilities were discovered in all Emacs versions up to 20.6, namely: Under certain circumstances, unprivileged local users can eavesdrop the communication between Emacs and its subprocesses. It is impossible to safely create temporary files in a public directory from Emacs Lisp. The history of recently typed keys may expose passwords. Especially the first two vulnerabilities seriously impact the use of tools like Mailcrypt in a multi-user environment. 1. Improper permissions on slave PTYs 1.1. Scope Affected systems: Linux (both libc and glibc) FreeBSD (and probably other *BSD variants) HP-UX 10.x, 11.00 AIX 4 Unaffected systems: Solaris (The Solaris runtime system automatically adjusts the PTY permissions.) Data General's DG/UX seems to be unaffected, according to the source code. Other systems have not been examined. 1.2. Severity High in multi-user environments, low otherwise. 1.3. Problem On the systems listed above, when a new subprocess is created using the builtin Lisp function start-process, Emacs doesn't set proper permissions for the slave PTY device. 1.4. Impact Unprivileged local users can eavesdrop the data which Emacs sends to its subprocess and fake responses from the subprocess. This impacts Emacs packages such as Mailcrypt, which transmit (among other things) PGP passphrases over this data channel. 1.5. Solution At Emacs Lisp level, the only workaround is to use call-process instead of start-process. Of course, this is not always an option because the functionality provided by these functions is not the same (synchronous vs. asynchronous subprocesses). The real solution requires modification of the Emacs C source code. A patch for Emacs 20.6 is included below which enables Emacs to Unix98 PTYs. The patch is known to work on the following systems: Linux with glibc 2.1 AIX 4.2 HP-UX 11.00 It is expected to work on HP-UX 10.x as well. (Under some versions of HP-UX, grantpt() does not behave as specified. The patch contains a suitable workaround.) Unfortunately, systems lacking Unix98 support (such as Linux with libc5 and glibc 2.0, FreeBSD and AIX 3) require a completely different fix and a setuid root binary to change the PTY permissions (in other words: some kind of userspace Unix 98 PTY emulation). There are no plans to provide this emulation; Unix 98 PTYs are already widely adopted and most Unix derivatives provide them (with the notable exception of several *BSD variants). For FreeBSD, an enhancement to openpty() has been proposed which sets proper permissions on the slave TTY device (see problem report bin/9770). The proposal has yet to be adopted, though. Future Emacs releases will contain a similar fix. 2. Unsafe creation of temporary files 2.1. Scope All Unix-like Emacs platforms on which public directories are used to store temporary files. 2.2. Severity High in multi-user environments, low otherwise. 2.3. Problem Emacs Lisp does not provide any functionality to create a file in a publicly writable directory in a safe way. 2.4. Impact Many Emacs packages use the make-temp-name Lisp function to create names for temporary files. These names are not very hard to guess. Because it is impossible to create the actual temporary file in a safe manner, the usual symlink attacks are likely successful. 2.5. Solution Emacs 21 will provide a new make-temp-file function (which creates the file in question in safe way) and the functionality to safely create temporary files. In the meantime, until Emacs 21 is released and package maintainers adopt the new function, private directories for temporary files should be used. Most packages provide variables for that. For example, for Mailcrypt, the variable mc-temp-directory has to be set, and for Python Mode, it's py-temp-directory. 3. Passwords are stored in the key history 3.1. Scope All platforms. 3.2. Severity Low. 3.3. Problem Functions like read-passwd do not clear the the history of recently typed keys. In fact, there is no way to do that from Emacs Lisp. 3.4. Impact Passwords might be recovered by someone who has got access to the console on which Emacs is running, subverting password expiring as, for example, provided by Mailcrypt. (Usually, there are many other ways to obtain passwords if you can type C-h l inside a foreign Emacs, though.) 3.5. Solution The patch below adds code to clear-this-command-keys which will erase the vector containing the last 100 events. In the past, this function was already used as if it behaved that way. 4. Acknowledgements Helmut Waitzmann for rediscovering the PTY permissions problem and testing the HP-UX patch. Gerd Moellmann of the Emacs development team for the patch to clear-this-command-keys and helpful comments. 5. Patch against Emacs 20.6 The patch below is against GNU Emacs 20.6, as available at GNU FTP mirrors. Note that you have to run autoconf to recreate the configure script (including it would have enormously increased the size of the patch). diff --unified --recursive emacs-20.6-orig/configure.in emacs-20.6/configure.in --- emacs-20.6-orig/configure.in Sat Feb 26 13:07:02 2000 +++ emacs-20.6/configure.in Fri Mar 10 19:13:05 2000 @@ -1636,6 +1636,11 @@ strerror fpathconf select mktime euidaccess getpagesize tzset setlocale \ utimes setrlimit setpgid getcwd shutdown strftime) +# Check for UNIX98 PTYs. +# getpt is a glibc addition which emulates the master device on +# systems without kernel support. +AC_CHECK_FUNCS(grantpt unlockpt getpt ptsname) + # Check this now, so that we will NOT find the above functions in ncurses. # That is because we have not set up to link ncurses in lib-src. # It's better to believe a function is not available diff --unified --recursive emacs-20.6-orig/src/config.in emacs-20.6/src/config.in --- emacs-20.6-orig/src/config.in Mon Apr 26 07:19:44 1999 +++ emacs-20.6/src/config.in Fri Mar 10 19:13:05 2000 @@ -235,6 +235,14 @@ #undef HAVE_SHUTDOWN #undef HAVE_STRFTIME +/* UNIX98 PTY support functions + getpt is a glibc addition which emulates the master device on + systems without kernel support. */ +#undef HAVE_GRANTPT +#undef HAVE_UNLOCKPT +#undef HAVE_GETPT +#undef HAVE_PTSNAME + #undef LOCALTIME_CACHE #undef HAVE_INET_SOCKETS diff --unified --recursive emacs-20.6-orig/src/keyboard.c emacs-20.6/src/keyboard.c --- emacs-20.6-orig/src/keyboard.c Thu Nov 18 05:57:32 1999 +++ emacs-20.6/src/keyboard.c Fri Mar 10 19:13:05 2000 @@ -8318,10 +8318,18 @@ DEFUN ("clear-this-command-keys", Fclear_this_command_keys, Sclear_this_command_keys, 0, 0, 0, - "Clear out the vector that `this-command-keys' returns.") + "Clear out the vector that `this-command-keys' returns.\n\ +Clear vector containing last 100 events.") () { + int i; + this_command_key_count = 0; + + for (i = 0; i < XVECTOR (recent_keys)->size; ++i) + XVECTOR (recent_keys)->contents[i] = Qnil; + total_keys = 0; + recent_keys_index = 0; return Qnil; } diff --unified --recursive emacs-20.6-orig/src/s/aix4.h emacs-20.6/src/s/aix4.h --- emacs-20.6-orig/src/s/aix4.h Sat Jul 25 08:45:27 1998 +++ emacs-20.6/src/s/aix4.h Fri Mar 17 20:44:08 2000 @@ -12,3 +12,33 @@ /* Specify the type that the 3rd arg of `accept' points to. It is just a guess which versions of AIX need this definition. */ #define SOCKLEN_TYPE int + +#if defined(HAVE_GRANTPT) && defined(HAVE_UNLOCKPT) && defined(HAVE_PTSNAME) +/* UNIX98 PTYs are available. + Added by Florian Weimer , + RUS-CERT, University of Stuttgart. Based on Emacs code for DGUX. */ + +/* Most of the #defines are already provided by aix3-1.h. */ + +/* This sets the name of the slave side of the PTY. grantpt(3) and + unlockpt(3) may fork a subprocess, so keep sigchld_handler() from + intercepting that death. */ + +#undef PTY_TTY_NAME_SPRINTF +#define PTY_TTY_NAME_SPRINTF \ + { \ + char *ptsname(), *ptyname; \ + \ + sigblock(sigmask(SIGCHLD)); \ + if (grantpt(fd) == -1) \ + fatal("could not grant slave pty"); \ + if (unlockpt(fd) == -1) \ + fatal("could not unlock slave pty"); \ + sigunblock(sigmask(SIGCHLD)); \ + if (!(ptyname = ptsname(fd))) \ + fatal ("could not enable slave pty"); \ + strncpy(pty_name, ptyname, sizeof(pty_name)); \ + pty_name[sizeof(pty_name) - 1] = 0; \ + } + +#endif diff --unified --recursive emacs-20.6-orig/src/s/gnu-linux.h emacs-20.6/src/s/gnu-linux.h --- emacs-20.6-orig/src/s/gnu-linux.h Wed Jan 26 14:28:40 2000 +++ emacs-20.6/src/s/gnu-linux.h Fri Mar 17 20:44:31 2000 @@ -307,3 +307,49 @@ #ifdef DOUG_LEA_MALLOC #undef REL_ALLOC #endif + +#if defined(HAVE_GRANTPT) && defined(HAVE_UNLOCKPT) && defined(HAVE_PTSNAME) +/* UNIX98 PTYs are available. + Added by Florian Weimer , + RUS-CERT, University of Stuttgart. Based on Emacs code for DGUX. */ + +#define PTY_ITERATION for (i = 0; i < 1; i++) +/* no iteration at all */ + +/* Use getpt() if it's available, because it provides Unix98 PTY + emulation for kernels which doesn't support it natively. */ + +#ifdef HAVE_GETPT +#define PTY_OPEN \ + do { \ + fd = getpt(); \ + if (fcntl (fd, F_SETFL, O_NDELAY) == -1) \ + fatal ("could not set master PTY to non-block mode"); \ + } while (0) + +#else +/* the master PTY device */ +#define PTY_NAME_SPRINTF strcpy (pty_name, "/dev/ptmx"); +#endif + +/* This sets the name of the slave side of the PTY. grantpt(3) and + unlockpt(3) may fork a subprocess, so keep sigchld_handler() from + intercepting that death. */ + +#define PTY_TTY_NAME_SPRINTF \ + { \ + char *ptsname(), *ptyname; \ + \ + sigblock(sigmask(SIGCHLD)); \ + if (grantpt(fd) == -1) \ + fatal("could not grant slave pty"); \ + if (unlockpt(fd) == -1) \ + fatal("could not unlock slave pty"); \ + if (!(ptyname = ptsname(fd))) \ + fatal ("could not enable slave pty"); \ + strncpy(pty_name, ptyname, sizeof(pty_name)); \ + pty_name[sizeof(pty_name) - 1] = 0; \ + sigunblock(sigmask(SIGCHLD)); \ + } + +#endif diff --unified --recursive emacs-20.6-orig/src/s/hpux.h emacs-20.6/src/s/hpux.h --- emacs-20.6-orig/src/s/hpux.h Mon Jan 15 10:16:40 1996 +++ emacs-20.6/src/s/hpux.h Wed Mar 29 08:40:52 2000 @@ -228,6 +228,59 @@ /* This is needed for HPUX version 6.2; it may not be needed for 6.2.1. */ #define SHORT_CAST_BUG +#if defined(HAVE_GRANTPT) && defined(HAVE_UNLOCKPT) && defined(HAVE_PTSNAME) +/* UNIX98 PTYs are available. + Added by Florian Weimer , + RUS-CERT, University of Stuttgart. Based on Emacs code for DGUX. */ + +#ifdef emacs +#include +#include +#endif + +#define PTY_ITERATION for (i = 0; i < 1; i++) +/* no iteration at all */ + +/* the master PTY device */ +#define PTY_NAME_SPRINTF strcpy (pty_name, "/dev/ptmx"); + +/* This sets the name of the slave side of the PTY. grantpt(3) and + unlockpt(3) may fork a subprocess, so keep sigchld_handler() from + intercepting that death. grantpt() behavior on HP-UX differs from + what's specified in the man page: the group of the slave PTY is set + to the user's primary group, and we fix that. */ + +#define PTY_TTY_NAME_SPRINTF \ + { \ + char *ptsname(), *ptyname; \ + struct group *getgrnam (), *tty_group = getgrnam ("tty"); \ + if (tty_group == NULL) \ + fatal ("group tty not found"); \ + \ + sigblock(sigmask(SIGCHLD)); \ + if (grantpt(fd) == -1) \ + fatal("could not grant slave pty"); \ + if (!(ptyname = ptsname(fd))) \ + fatal ("could not enable slave pty"); \ + strncpy(pty_name, ptyname, sizeof(pty_name)); \ + pty_name[sizeof(pty_name) - 1] = 0; \ + if (chown (pty_name, (uid_t) -1, tty_group->gr_gid) == -1) \ + fatal ("could not chown slave pty"); \ + if (unlockpt(fd) == -1) \ + fatal("could not unlock slave pty"); \ + sigunblock(sigmask(SIGCHLD)); \ + } + +/* Push various streams modules onto a PTY channel. */ + +#define SETUP_SLAVE_PTY \ + if (ioctl (xforkin, I_PUSH, "ptem") == -1) \ + fatal ("ioctl I_PUSH ptem", errno); \ + if (ioctl (xforkin, I_PUSH, "ldterm") == -1) \ + fatal ("ioctl I_PUSH ldterm", errno); + +#else /* no UNIX98 PTYs */ + /* This is how to get the device name of the tty end of a pty. */ #define PTY_TTY_NAME_SPRINTF \ sprintf (pty_name, "/dev/pty/tty%c%x", c, i); @@ -235,6 +288,8 @@ /* This is how to get the device name of the control end of a pty. */ #define PTY_NAME_SPRINTF \ sprintf (pty_name, "/dev/ptym/pty%c%x", c, i); + +#endif /* UNIX 98 PTYs */ /* This triggers a conditional in xfaces.c. */ #define XOS_NEEDS_TIME_H