mailing list archives
Re: CVE Request -- rpm -- Fails to remove the SUID/SGID bits on package upgrade (RH BZ#598775)
From: Jeff Johnson <n3npq () mac com>
Date: Mon, 25 Jul 2011 21:49:46 -0400
On Jul 25, 2011, at 7:22 PM, Solar Designer wrote:
On Mon, Jul 25, 2011 at 09:30:35PM +0400, Vasiliy Kulikov wrote:
On Mon, Jul 25, 2011 at 06:08 +0400, Solar Designer wrote:
- rc = Unlink(fsm->path);
+ struct stat stb;
+ int saved_errno;
+ int saved_rc = lstat(fsm->path, &stb);
+ if (!saved_rc && !S_ISLNK(stb.st_mode))
+ saved_rc = chmod(fsm->path, 0);
If the directory containing the file was owned by nonroot, then the file
could be overwritten with a symlink. So, there is a race between
lstat() and chmod(), which might lead to chmod'ing arbitrary files by
Right. The same risk is present in upstream's version of the fix.
Is there really any risk factor when the operation at the end of the
possible raciness is removing all mode related privilege?
Is it possible with these orphaned files (I'm not familiar with the code
Yes, but this problem is not limited to this specific piece of code.
rpm appears to treat the target directory tree as trusted - not only
when it removes files, but also when it creates files, etc. I did not
fully verify this, though - that's just how the code looks to me.
Here is what is implemented:
RPM opens file paths based on paths in the metadata, not based on
dynamic state in a directory whatsoever.
The paths are appended with a uniqifier called a "transaction ID"
which is essentially a time(2) stamp.
The file is written, the file is renamed into place, and chmod/chown/utimes
(only lchown for symlinks, whose type was determined from metadata)
Directories are checked to see if they exist (in which case the operation is skipped),
and are otherwise treated as above. Parent directories were checked and created
as needed before attempting to write file content, with default 755 root:root.
There's some speshul painfulness when rpm attempts to attach SELinux file contexts
to orphan directories somehow: my personal belief is that this is a fool's
errand that would be better dealt with using zero tolerance, like an assertion,
or QA to ensure that there are no "orphan" directories, but is the behavior
that was requested at the time that RPM was made responsible for attaching
SELinux file contexts on every installed path.
Devices are created by mknod(2) and otherwise renamed into place as above.
Exotica like FIFO's and sockets have no business in *.rpm packages imho:
all code I've ever seen that uses FIFO's/sockets is perfectly prepared
to create the fifo/socket where/when needed at runtime, and there is no reason to
"package" these types of files other than a fetishism for "packaging". But
RPM handles these similarly to devices: unique path created, renamed into place,
and then chmod/chown/utimes.
While one might argue that uniqifier is predictable, and hence
vulnerable to raciness, one would have to already have root in order
to modify the content being written.
SHow me the flaw in the above, and I will instantly fix. I have a harder
time attempting to guess precisely with phrases like
rpm appears to treat the target directory tree as trusted
because it uses imprecise terms like "appears" and "trusted". All content
is installed by a state machine with strict behavior. If the behavior is racy,
then the time ordered state transitions have an hole (and the state transitions
placing content onto the file system are fairly predictable).
Please: I'm not trying to be obscure or obtuse, merely trying to
get the best possible implementation in place in RPM code.
This general issue is in fact a security risk. For example, if the
directory tree contains a subdirectory writable by a pseudo-user, then a
possible compromise of this pseudo-user account might lead to worse
things via rpm. Here's an example of such directory on Owl:
# ls -la /var/lib/dhcp/dhcpd/state/
drwxrwx--T 2 root dhcp 4096 Dec 14 2010 .
drwxr-x--- 3 root dhcp 4096 Dec 8 2010 ..
-rw------- 1 dhcp dhcp 0 Dec 8 2010 dhcpd.leases
I'm not sure I follow the example, but perhaps I'm tracking too closely
with what is typically done in packaging semantics for the given example.
I would expect that dhcpd.leases either would be left purely as a side
effect (and unpackaged) or be added as %ghost so that it would be
removed when the package is erased. A path in %ghost is never touched
while installing iirc (I wrote this code years ago, can easily re-read
if necessary, but nothing much has been changed in the file state machine
code since forever, the CVE being one of the only changes I recall in years.
So if I understand the threat vector you see then I would say
that a file like dhcpd.leases isn't typically touched by rpm
and so even if the pseudo-user account is compromised, and that directory
were writable so that races might be attempted, that this is a
packaging flaw, not otherwise. Yes in the general case rpm SHOULD
handle this operation somehow, and there might well be better examples
that illustrate the attack that you are describing.
The state machine used to touch the file system is quite clear if
rpm -Uvv --fsmdebug
and examine the debugging output. If you can point me at the
specific point in the debugging output where a race exists,
I will have a fix checked in within 24 hours (I promise).
We may discuss this general issue (of rpm trusting the target tree, and
the resulting risks) separately.
Thank you for the review!
BTW, another detail I thought someone might notice is that I am applying
the chmod's not only to binary packages, like the upstream fix does, but
I think also to source packages being removed/upgraded (I did not
actually test this, though). This might be excessive, or it might not,
but I felt that it does not hurt either way.
The handling of source RPM's and the exploits that might exist on
build systems, have never really been thought through.
Here is one exploit that comes alarmingly close to tricking rpm into
removing one's home directory:
Name: foo;rm -rf ~
I will leave it as an exercise how to get the spaces past rpmbuild's
parser, not hard.
The flaws has been addressed both at @rpm5.org and @rpm.org
18 or 30 months ago (I forget), basically by restricting the character set permitted
in certain tags like Name:.
I point the flaw out to show that the risks and attacks of SRPM's and build systems
are rather different than binary RPM's on client machines.
And in the hope that someone will look more deeply at build machine security issues.
The usual argument for NOT looking at build machine security goes something like this
Build machines are usually well protected in order to avoid trojans
in source code. So there's less need of examining rpmbuild or the
installation of SRPM's.
73 de Jeff