James Yonan wrote:
Matthias Andree wrote:
On Thu, 21 Aug 2008, Alberto Gonzalez Iniesta wrote:
It seems that tightening the security on OpenVPN brought some surprises
[1] to users and broke some features [2].
As for [1], I included a note in the Debian NEWS file on the new
--script-security option. But those updating a VPN using the very same
VPN (and without previous knowledge of this option) may find themselves
without access to the remote system (if the VPN/system is restarted, and
a script is to be executed). Also, those using NetworkManager [3] aren't
able to specify the '--script-security' option, and since NetworkManager
may/will call external scripts, this new security feature will break
their VPNs. The option is really useful, but 2 would be a more sensible
default IMHO.
One might argue that NetworkManager doesn't support all *RC versions of
OpenVPN's. At the end of the day, it will be a NetworkManager issue.
Interfaces change...
Regarding [2] an strace shows that calls to external commands with
arguments include the arguments as part of the command filename:
For:
--up "/tmp/foo up"
The call is:
[pid 3519] execve("/tmp/foo up", ["/tmp/foo up", "tun0", "1500", "1542",
"10.XXX.XXX.X", "10.XXX.XXX.X", "init"], [/* 30 vars */]) = -1 ENOENT
(No such file or directory)
Apparently argument splitting - as documented in the --up section of the
manpage - no longer works; it was probably formerly done by the implicit
/bin/sh -c that is now gone with the switch to exec*().
(I didn't check, and didn't check the two Debian BTS reports either.)
So either the code needs argument splitting or you need a two-line shell
wrapper similar to:
#! /bin/sh -e
exec /tmp/foo up "$@"
Not my call to make.
I agree that argument splitting in --up "/tmp/foo up" should be
supported, as it doesn't detract from the sought-after security goals in
migrating from system() to execve(), and preserves backward compatibility.
Okay, argument splitting should be working again with this patch.
One of the above posters asked if --script-security could be defaulted
to 2 to preserve backward compatibility. I would tend to disagree on
this as it would defeat, in my view, the primary purpose of
--script-security.
I think the guiding principle here is that security software should
refrain from doing potentially dangerous things (and I don't think
anyone can deny that running user-defined scripts from a VPN daemon is
potentially dangerous) unless explicit permission is given by the user.
------------------------------------------------------------------------
r3311 | james | 2008-09-06 03:42:17 -0600 (Sat, 06 Sep 2008) | 20 lines
Changed paths:
M /branches/BETA21/openvpn/buffer.c
M /branches/BETA21/openvpn/buffer.h
M /branches/BETA21/openvpn/errlevel.h
M /branches/BETA21/openvpn/init.c
M /branches/BETA21/openvpn/misc.c
M /branches/BETA21/openvpn/misc.h
M /branches/BETA21/openvpn/multi.c
M /branches/BETA21/openvpn/socket.c
M /branches/BETA21/openvpn/ssl.c
2.1_rc8 and earlier did implicit shell expansion on script
arguments since all scripts were called by system().
The security hardening changes made to 2.1_rc9 no longer
use system(), but rather use the safer execve or CreateProcess
system calls. The security hardening also introduced a
backward incompatibility with 2.1_rc8 and earlier in that
script parameters were no longer shell-expanded, so
for example:
client-connect "docc CLIENT-CONNECT"
would fail to work because execve would try to execute
a script called "docc CLIENT-CONNECT" instead of "docc"
with "CLIENT-CONNECT" as the first argument.
This patch fixes the issue, bringing the script argument
semantics back to pre 2.1_rc9 behavior in order to preserve
backward compatibility while still using execve or CreateProcess
to execute the script/executable.
------------------------------------------------------------------------
