Re: [libvirt] [PATCH 2/4] libvirt-guest.init: quoting variables

On 03/09/2011 01:54 AM, Philipp Hahn wrote: Hmm. Valid URIs can contain '?', which _is_ a glob character, so I agree that we need to quote "$uri" everywhere that we have already correctly obtatined it in order to avoid inadvertent globbing. Meanwhile, $URIS is indeed a user-settable variable, via /etc/sysconfig/libvirt-guests, so we should want to protect ourselves from malicious input. (On the other hand, we already blindly source /etc/sysconfig/libvirt-guests without validating that it contains _only_ shell assignments, so a malicious user could already put arbitrary shell code in there. The real way to be safe would be to parse out _just_ the assignments we care about, using sed or read, and eval those assignments rather than sourcing the entire file. I guess we have to draw a line somewhere at how much paranoia we want to express. If we can trust that /etc/sysconfig/libvirt-guests can't be subverted by a malicious user, then we can trust that it won't contain a malicious $URIS and instead chalk up bugs in that file as system administration shortcomings.) However, that points out a shortcoming in your patch - we already do constructs like for uri in $URIS; do which _already_ corrupted any globbing and performed IFS splitting, at which point $uri would not contain any further globbing (except in the weird case of a glob that expanded to an existing file name which happens to also be a valid glob), so quoting just "$uri" as in your patch is a lost cause unless we address the problem from the point of the user input to begin with. Meanwhile, we already know that a valid URI does not contain whitespace (a valid URI can represent whitespace via encoding, where needed), so it's already acceptable to use IFS splitting on $URIS to break it into individual URIs, it's just that we need to suppress globbing in the process. 'set -f' can be used for this, if needed. If we could use bash shell arrays, it would be a matter of creating an array variable ${URI[]} that contains the individual URIs, doing the work of safely splitting things only once. But libvirt-guests has to run under /bin/sh == dash for Debian, which lacks shell arrays. I'm still thinking about how to handle this one... Meanwhile, I'm not a fan of blindly quoting everything; there are documented cases where you can trigger shell bugs by doing too much quoting. For example: foo=`some_command "with quotes"` is portable, but foo="`some_command "with quotes"`" is not. So I prefer to quote variables that come from external source, but not internal variables that are obviously safe (that said, quoting generally doesn't hurt, so even if something is overkill does not meant that I am rejecting the patch). This part's good. Quoting "$id" is overkill; we know it's numeric. But it probably doesn't hurt. Likewise, if we know $uuid is valid, then it contains neither $IFS nor globbing characters. Should we instead be tightening up the list_guests function to guarantee that we never encounter garbage data? To be really robust (for example, $confuri could be '(', which is known to trip up some buggy test(1) implementations), this should be: [ "x$confuri" = "x$uri" ] This one's overkill - we explicitly set $configured to either 'true' or 'false'; no external data possible. Overkill - we know $virsh_pid is numeric. Reasonable, since SHUTDOWN_TIMEOUT might come from external sources. Overkill, since once you are inside the while loop, you are guaranteed that $timeout is now numeric. Overkill, since $guest_running is under our control. Overkill, since $empty is under our control. Likewise. Overkill, since we know that $1 is one of three safe values. I think I'll apply your patch as-is, in spite of the overkill, then worry about how to safely split $URIS as a separate patch. -- Eric Blake eblake(a)redhat.com +1-801-349-2682 Libvirt virtualization library http://libvirt.org