On Wed, Jun 16, 2010 at 3:57 PM, Jamie Strandboge <jamie@canonical.com> wrote:
On Wed, 2010-06-16 at 14:04 -0600, Eric Blake wrote:
I'm with Stefan - the whole point of mkstemp is that the created file has 0600 permissions, and /tmp is restricted-deletion, so no other user can either overwrite the file contents or unlink it and replace it with an alternate file. Then again, gnulib documents that glibc 2.0.7 or older would create a file with group/other permissions if the umask wasn't set prior to the mkstemp() call, and gnulib's mkstemp() does not work around this issue; but that's a rather old version of glibc to be worrying about.
This has nothing to do with mkstemp(). As I said, libvirt's use of it is fine and there is no symlink race or security vulnerability by itself.
The issue is that use of /tmp is not required *and* it becomes difficult to properly confine libvirtd via a MAC if you must allow execution of files in /tmp. See my answer to Stefan's question for an example scenario.
-- Jamie Strandboge | http://www.canonical.com
For what it's worth, I've seen many times when people secure system that /tmp is mounted with noexec. If any of you have a shared web host and you can ssh into it and confirm. $ mount ...snip... /dev/mapper/vg-tmp on /tmp type ext3 (rw,noexec,nosuid,noatime) ...snip... -- Doug Goldstein