So I've spent a fair bit of time the last 2 weeks trying to get livecd-creator and an selinux enforcing machine to play nicely together. It doesn't look like much, but from the point of view of the livecd creator I think the following patch is all we need. Working with rawhide as the host system I was able to build F8, F9 and rawhide livecd's with an enforcing machine.
I wouldn't suggest jumping into enfocing builds just yet as there are still some policy issues I need to work out with the selinux people but I would like comments. Basically its quite simple, if selinux is on the host we create a fake /selinux which tells the install chroot lies. I've had to make some changes to some selinux libraries to support all this, but I think we are just about there.
I'll probably backport some of the kernel changes to F9 after they are all tested and better settled but for now I'd like input on my livecd changes....
---
diff -Naupr /usr/lib/python2.5/site-packages/imgcreate/creator.py /root/imgcreate-5-28-08/creator.py --- /usr/lib/python2.5/site-packages/imgcreate/creator.py 2008-05-06 12:16:08.000000000 -0400 +++ /root/imgcreate-5-28-08/creator.py 2008-05-28 15:48:30.000000000 -0400 @@ -23,6 +23,7 @@ import sys import tempfile import shutil
+import selinux import yum import rpm
@@ -427,7 +428,7 @@ class ImageCreator(object):
self._mount_instroot(base_on)
- for d in ("/dev/pts", "/etc", "/boot", "/var/log", "/var/cache/yum"): + for d in ("/dev/pts", "/etc", "/boot", "/var/log", "/var/cache/yum", "/sys", "/proc", "/selinux"): makedirs(self._instroot + d)
cachesrc = cachedir or (self.__builddir + "/yum-cache") @@ -439,10 +440,6 @@ class ImageCreator(object): (cachesrc, "/var/cache/yum")]: self.__bindmounts.append(BindChrootMount(f, self._instroot, dest))
- # /selinux should only be mounted if selinux is enabled (enforcing or permissive) - if kickstart.selinux_enabled(self.ks): - self.__bindmounts.append(BindChrootMount("/selinux", self._instroot, None)) - # Create minimum /dev origumask = os.umask(0000) devices = [('null', 1, 3, 0666), @@ -460,6 +457,37 @@ class ImageCreator(object): os.symlink('/proc/self/fd/2', self._instroot + "/dev/stderr") os.umask(origumask)
+ # if selinux exists on the host we need to lie to the chroot + if os.path.exists("/selinux/enforce"): + selinux_dir = self._instroot + "/selinux" + + # enforce=0 tells the chroot selinux is not enforcing + # policyvers=99 tell the chroot to make the highest version of policy it can + files = [('/enforce', '0'), + ('/policyvers', '99')] + for (file, value) in files: + fd = os.open(selinux_dir + file, os.O_WRONLY | os.O_TRUNC | os.O_CREAT) + os.write(fd, value) + os.close(fd) + + # we steal mls from the host system for now, might be best to always set it to 1???? + files = ["/selinux/mls"] + for file in files: + shutil.copyfile(file, self._instroot + file) + + # make /load -> /dev/null so chroot policy loads don't hurt anything + os.mknod(selinux_dir + "/load", 0666 | stat.S_IFCHR, os.makedev(1, 3)) + + # selinux is on whoo hooo + if kickstart.selinux_enabled(self.ks): + # label the fs like it is a root before the bind mounting + cmd = "/sbin/setfiles -F -r %s %s %s" % (self._instroot, selinux.selinux_file_context_path(), self._instroot) + os.system(cmd) + # these dumb things don't get magically fixed, so make the user generic + for f in ["/proc", "/sys", "/selinux"]: + cmd = "chcon -u system_u %s" % (self._instroot + f) + os.system(cmd) + self._do_bindmounts()
os.symlink("../proc/mounts", self._instroot + "/etc/mtab") @@ -853,6 +881,18 @@ class LoopImageCreator(ImageCreator): (self._image, e))
def _unmount_instroot(self): + # if the system was running selinux clean up our lies + if os.path.exists("/selinux/enforce"): + files = ['/enforce', + '/policyvers', + '/mls', + '/load'] + for file in files: + try: + os.unlink(self._instroot + "/selinux" + file) + except OSError: + pass + if not self.__instloop is None: self.__instloop.cleanup()
diff -Naupr /usr/lib/python2.5/site-packages/imgcreate/kickstart.py /root/imgcreate-5-28-08/kickstart.py --- /usr/lib/python2.5/site-packages/imgcreate/kickstart.py 2008-05-06 12:16:08.000000000 -0400 +++ /root/imgcreate-5-28-08/kickstart.py 2008-05-19 11:22:41.000000000 -0400 @@ -372,11 +372,11 @@ class SelinuxConfig(KickstartConfig):
if ksselinux.selinux == ksconstants.SELINUX_DISABLED: return - if not os.path.exists(self.path("/sbin/restorecon")): + if os.path.exists(self.path("/sbin/restorecon")): + self.call(["/sbin/restorecon", "-l", "-v", "-r", "-F", "-e", "/proc", "-e", "/sys", "-e", "/dev", "-e", "/selinux", "/"]) + else: return
- self.call(["/sbin/restorecon", "-l", "-v", "-r", "/"]) - def apply(self, ksselinux): if os.path.exists(self.path("/usr/sbin/lokkit")): args = ["/usr/sbin/lokkit", "-f", "--quiet", "--nostart"]
Eric Paris (eparis@redhat.com) said:
So I've spent a fair bit of time the last 2 weeks trying to get livecd-creator and an selinux enforcing machine to play nicely together. It doesn't look like much, but from the point of view of the livecd creator I think the following patch is all we need. Working with rawhide as the host system I was able to build F8, F9 and rawhide livecd's with an enforcing machine.
I wouldn't suggest jumping into enfocing builds just yet as there are still some policy issues I need to work out with the selinux people but I would like comments. Basically its quite simple, if selinux is on the host we create a fake /selinux which tells the install chroot lies. I've had to make some changes to some selinux libraries to support all this, but I think we are just about there.
I'll probably backport some of the kernel changes to F9 after they are all tested and better settled but for now I'd like input on my livecd changes....
My concern is this is a normal occurence (needing a chroot) that you're only patching in one place. Do we code this same logic into mock? Into pungi? Into yum --installroot? Into the documentation for admins on how to set up a chroot?
(Also, for general use, we need this in a RHEL 5 kernel. Fun!)
Bill
On Wed, 2008-05-28 at 16:04 -0400, Bill Nottingham wrote:
Eric Paris (eparis@redhat.com) said:
So I've spent a fair bit of time the last 2 weeks trying to get livecd-creator and an selinux enforcing machine to play nicely together. It doesn't look like much, but from the point of view of the livecd creator I think the following patch is all we need. Working with rawhide as the host system I was able to build F8, F9 and rawhide livecd's with an enforcing machine.
I wouldn't suggest jumping into enfocing builds just yet as there are still some policy issues I need to work out with the selinux people but I would like comments. Basically its quite simple, if selinux is on the host we create a fake /selinux which tells the install chroot lies. I've had to make some changes to some selinux libraries to support all this, but I think we are just about there.
I'll probably backport some of the kernel changes to F9 after they are all tested and better settled but for now I'd like input on my livecd changes....
My concern is this is a normal occurence (needing a chroot)
Yes and no....
that you're only patching in one place. Do we code this same logic into mock?
Mock doesn't need any special labeling knowledge, mock just has to have policy that doesn't get in the way.... But mock is the next thing I'm going to start trying to find issues with.
Into pungi?
don't know pungi...
Into yum --installroot?
possibly....
Into the documentation for admins on how to set up a chroot?
where is this documentation? usually, no, but if the chroot is supposed to be drastically different than the host policy maybe...
(Also, for general use, we need this in a RHEL 5 kernel. Fun!)
Yeah, I've heard that, and its likely to not be possible.... Once we all agree on how to make it work going forward we can talk about RHEL5.
On Wed, 2008-05-28 at 16:11 -0400, Eric Paris wrote:
My concern is this is a normal occurence (needing a chroot)
Yes and no....
sure looks like we'd need to make sure:
yum, mock and rpm all know how to set this up given how it would impact chroot creation.
We may want to drop this back to the lowest level chroot creation.
-sv
seth vidal wrote:
On Wed, 2008-05-28 at 16:11 -0400, Eric Paris wrote:
My concern is this is a normal occurence (needing a chroot)
Yes and no....
sure looks like we'd need to make sure:
yum, mock and rpm all know how to set this up given how it would impact chroot creation.
If we do so, then we should also make sure that all do things consistently wrt /dev for creating a chroot as well. And /proc and /sys. The reality is that the different applications do have a somewhat different idea of what they need/want out of their chroots and do things (or don't) accordingly.
We may want to drop this back to the lowest level chroot creation.
Which isn't to say that we might not decide down the road to push it down the stack, but I don't know that livecd-creator is a bad place in the short to medium term as Eric continues looking at SELinux and chroot interactions (Right Eric? :-)
Jeremy
Bill Nottingham wrote:
Eric Paris (eparis@redhat.com) said:
So I've spent a fair bit of time the last 2 weeks trying to get livecd-creator and an selinux enforcing machine to play nicely together. It doesn't look like much, but from the point of view of the livecd creator I think the following patch is all we need. Working with rawhide as the host system I was able to build F8, F9 and rawhide livecd's with an enforcing machine.
I wouldn't suggest jumping into enfocing builds just yet as there are still some policy issues I need to work out with the selinux people but I would like comments. Basically its quite simple, if selinux is on the host we create a fake /selinux which tells the install chroot lies. I've had to make some changes to some selinux libraries to support all this, but I think we are just about there.
I'll probably backport some of the kernel changes to F9 after they are all tested and better settled but for now I'd like input on my livecd changes....
My concern is this is a normal occurence (needing a chroot) that you're only patching in one place. Do we code this same logic into mock? Into pungi? Into yum --installroot? Into the documentation for admins on how to set up a chroot?
(Also, for general use, we need this in a RHEL 5 kernel. Fun!)
Bill
-- fedora-selinux-list mailing list fedora-selinux-list@redhat.com https://www.redhat.com/mailman/listinfo/fedora-selinux-list
Well I think we need to do a couple of these to figure out the common requirements.
I envision mock to be quite different then livecd. I think we need to full the mock chroot to think SELinux is disabled and to do no labeling in the chroot. This would allow us to confine the mock process to be able to write to the chroot and label the chroot mock_rw_t. We could then use SELinux to prevent mock environments from breaking out of the chroot, and stop mock environments from doing evil network things within the chroot.
In livecd we need to be able to put down labels that the host machine does not understand.
Daniel J Walsh (dwalsh@redhat.com) said:
Well I think we need to do a couple of these to figure out the common requirements.
I envision mock to be quite different then livecd. I think we need to full the mock chroot to think SELinux is disabled and to do no labeling in the chroot. This would allow us to confine the mock process to be able to write to the chroot and label the chroot mock_rw_t. We could then use SELinux to prevent mock environments from breaking out of the chroot, and stop mock environments from doing evil network things within the chroot.
In livecd we need to be able to put down labels that the host machine does not understand.
The problem is that mock can be used to do non-build things. (For example, creating the anaconda install images.)
Bill
Eric Paris wrote:
So I've spent a fair bit of time the last 2 weeks trying to get livecd-creator and an selinux enforcing machine to play nicely together. It doesn't look like much, but from the point of view of the livecd creator I think the following patch is all we need. Working with rawhide as the host system I was able to build F8, F9 and rawhide livecd's with an enforcing machine.
I wouldn't suggest jumping into enfocing builds just yet as there are still some policy issues I need to work out with the selinux people but I would like comments. Basically its quite simple, if selinux is on the host we create a fake /selinux which tells the install chroot lies. I've had to make some changes to some selinux libraries to support all this, but I think we are just about there.
Very cool and definitely long needed. Thanks for taking the time to really dive into this. And this is far simpler than the approach I had started looking at once upon a time (... which involved fuse)
A few comments on the patch
diff -Naupr /usr/lib/python2.5/site-packages/imgcreate/creator.py /root/imgcreate-5-28-08/creator.py --- /usr/lib/python2.5/site-packages/imgcreate/creator.py 2008-05-06 12:16:08.000000000 -0400 +++ /root/imgcreate-5-28-08/creator.py 2008-05-28 15:48:30.000000000 -0400 @@ -460,6 +457,37 @@ class ImageCreator(object): os.symlink('/proc/self/fd/2', self._instroot + "/dev/stderr") os.umask(origumask)
# if selinux exists on the host we need to lie to the chrootif os.path.exists("/selinux/enforce"):selinux_dir = self._instroot + "/selinux"# enforce=0 tells the chroot selinux is not enforcing# policyvers=99 tell the chroot to make the highest version of policy it canfiles = [('/enforce', '0'),('/policyvers', '99')]
Does the kernel guarantee that 99 will be the highest version of policy? Not that it likely matters much. Also, having this as a tuple rather than a list makes it marginally faster since we're never going to modify it
for (file, value) in files:fd = os.open(selinux_dir + file, os.O_WRONLY | os.O_TRUNC | os.O_CREAT)os.write(fd, value)os.close(fd)# we steal mls from the host system for now, might be best to always set it to 1????files = ["/selinux/mls"]for file in files:shutil.copyfile(file, self._instroot + file)# make /load -> /dev/null so chroot policy loads don't hurt anythingos.mknod(selinux_dir + "/load", 0666 | stat.S_IFCHR, os.makedev(1, 3))
This being the big win :)
# selinux is on whoo hoooif kickstart.selinux_enabled(self.ks):# label the fs like it is a root before the bind mountingcmd = "/sbin/setfiles -F -r %s %s %s" % (self._instroot, selinux.selinux_file_context_path(), self._instroot)os.system(cmd)# these dumb things don't get magically fixed, so make the user genericfor f in ["/proc", "/sys", "/selinux"]:cmd = "chcon -u system_u %s" % (self._instroot + f)os.system(cmd)
os.system is generally not preferred -- using the subprocess module is a lot safer.
Also, overall it might be nice to encapsulate the /selinux creation here into its own __create_selinuxfs() method that gets called. /me makes a note to do that to the /dev creation too.
@@ -853,6 +881,18 @@ class LoopImageCreator(ImageCreator): (self._image, e))
def _unmount_instroot(self):
# if the system was running selinux clean up our liesif os.path.exists("/selinux/enforce"):files = ['/enforce','/policyvers','/mls','/load']
Again a tuple versus a list
for file in files:try:os.unlink(self._instroot + "/selinux" + file)except OSError:pass
And again having it in a method is probably the nice thing to do. And I know I said to stick it into _unmount_instroot, but seeing where you've put the mount side, it probably makes more sense in unmount() instead
Jeremy
selinux@lists.fedoraproject.org
-
Bill Nottingham -
Daniel J Walsh -
Eric Paris -
Jeremy Katz -
Seth Vidal