On Thu, Feb 17, 2022 at 10:05 AM Christian Göttsche <cgzones@xxxxxxxxxxxxxx> wrote: > On Tue, 15 Feb 2022 at 21:35, Paul Moore <paul@xxxxxxxxxxxxxx> wrote: > > On Mon, Feb 14, 2022 at 2:11 AM Jeffrey Vander Stoep <jeffv@xxxxxxxxxx> wrote: > > > On Tue, Feb 8, 2022 at 3:18 PM William Roberts <bill.c.roberts@xxxxxxxxx> wrote: > > > > > > > > <snip> > > > > > > > > This is getting too long for me. > > > > > > > > > > > > > > > > I don't have a strong opinion either way. If one were to allow this > > > > > > using a policy rule, it would result in a major policy breakage. The > > > > > > rule would turn on extended perm checks across the entire system, > > > > > > which the SELinux Reference Policy isn't written for. I can't speak > > > > > > to the Android policy, but I would imagine it would be the similar > > > > > > problem there too. > > > > > > > > > > Excuse me if I am wrong but AFAIK adding a xperm rule does not turn on > > > > > xperm checks across the entire system. > > > > > > > > It doesn't as you state below its target + class. > > > > > > > > > > > > > > If i am not mistaken it will turn on xperm checks only for the > > > > > operations that have the same source and target/target class. > > > > > > > > That's correct. > > > > > > > > > > > > > > This is also why i don't (with the exception TIOSCTI for termdev > > > > > chr_file) use xperms by default. > > > > > > > > > > 1. it is really easy to selectively filter ioctls by adding xperm rules > > > > > for end users (and since ioctls are often device/driver specific they > > > > > know best what is needed and what not) > > > > > > > > > >>> and FIONCLEX can be trivially bypassed unless fcntl(F_SETFD) > > > > > > > > > > 2. if you filter ioctls in upstream policy for example like i do with > > > > > TIOSCTI using for example (allowx foo bar (ioctl chr_file (not > > > > > (0xXXXX)))) then you cannot easily exclude additional ioctls later where source is > > > > > foo and target/tclass is bar/chr_file because there is already a rule in > > > > > place allowing the ioctl (and you cannot add rules) > > > > > > > > Currently, fcntl flag F_SETFD is never checked, it's silently allowed, but > > > > the equivalent FIONCLEX and FIOCLEX are checked. So if you wrote policy > > > > to block the FIO*CLEX flags, it would be bypassable through F_SETFD and > > > > FD_CLOEXEC. So the patch proposed makes the FIO flags behave like > > > > F_SETFD. Which means upstream policy users could drop this allow, which > > > > could then remove the target/class rule and allow all icotls. Which is easy > > > > to prevent and fix you could be a rule in to allowx 0 as documented in the > > > > wiki: https://selinuxproject.org/page/XpermRules > > > > > > > > The questions I think we have here are: > > > > 1. Do we agree that the behavior between SETFD and the FIO flags are equivalent? > > > > I think they are. > > > > 2. Do we want the interfaces to behave the same? > > > > I think they should. > > > > 3. Do upstream users of the policy construct care? > > > > The patch is backwards compat, but I don't want their to be cruft > > > > floating around with extra allowxperm rules. > > > > > > I think this proposed change is fine from Android's perspective. It > > > implements in the kernel what we've already already put in place in > > > our policy - that all domains are allowed to use these IOCLTs. > > > https://cs.android.com/android/platform/superproject/+/master:system/sepolicy/public/domain.te;l=312 > > > > > > It'll be a few years before we can clean up our policy since we need > > > to support older kernels, but that's fine. > > > > Thanks for the discussion everyone, it sounds like everybody is okay > > with the change - that's good. However, as I said earlier in this > > thread I think we need to put this behind a policy capability, how > > does POLICYDB_CAPABILITY_IOCTL_CLOEXEC/"ioctl_skip_cloexec" sound to > > everyone? > > May I ask why? > To my understanding policy capabilities exist to retain backwards > compatibility for older > policies, e.g. if a new check is introduced or a new essential class > or permission, which > would break systems running an updated kernel with a non updated policy. > In this case no check or class/permission is added, the xperm checks > against FIO(N)CLEX > are just dropped. Old policies still defining related allow rules > continue to work. Existing > polices explicitly not allowing them and relying on SELinux to block changes on > the close-on-exec flag are already broken due to the bypasses via > fnctl(2) and dup(2). Policy capabilities are a general tool that we can use when we make a change in the kernel that could potentially have an effect on the policy; it allows the policy to (typically) "opt-in" to the change. In this particular case we are talking about removing access controls, which is a Very Serious Thing, and protecting this behavior with an opt-in policy capability seems like a good way to not surprise anyone with the change. You are correct in that old policy would continue to load and work regardless, but I believe it is safer to create a new policy capability for this. -- paul-moore.com
