[PATCH v2 0/6] Landlock: ioctl support

[Günther Noack]

Hi!

On Wed, Jul 12, 2023 at 07:48:29PM +0200, Mickaël Salaün wrote:
> On 12/07/2023 16:56, Günther Noack wrote:
> > On Wed, Jul 12, 2023 at 12:55:19PM +0200, Mickaël Salaün wrote:
> > > Thinking more about this, this first step is too restrictive, which
> > > might lead to dangerous situations.
> > > 
> > > My main concern is that this approach will deny innocuous or even "good"
> > > IOCTL. For instance, if FIOCLEX is denied, this could leak file
> > > descriptors and then introduce vulnerabilities.
> > 
> > This is a good point.
> > 
> > > As we discussed before, we cannot categorize all IOCTLs, but I think we
> > > need to add exceptions for a subset of them, and maintain this list.
> > > SELinux has some special handling within selinux_file_ioctl(), and we
> > > should use that as a starting point. The do_vfs_ioctl() function is
> > > another important place to look at. The main thing to keep in mind is
> > > that Landlock's goal is to restrict access to data (e.g. FS, network,
> > > IPC, bypass through other processes), not to restrict innocuous (at
> > > least in theory) kernel features.
> > > 
> > > I think, at least all IOCTLs targeting file descriptors themselves should
> > > always be allowed, similar to fcntl(2)'s F_SETFD and F_SETFL commands:
> > > - FIOCLEX
> > > - FIONCLEX
> > > - FIONBIO
> > > - FIOASYNC
> > > 
> > > Some others may not be a good idea:
> > > - FIONREAD should be OK in theory but the VFS part only target regular
> > > files and there is no access check according to the read right, which is
> > > weird.
> > > - FICLONE, FICLONERANGE, FIDEDUPRANGE: read/write actions.
> > > 
> > > We should add a built-time or run-time safeguard to be sure that future
> > > FD IOCTLs will be added to this list. I'm not sure how to efficiently
> > > implement such protection though.
> > 
> > I need to ponder it a bit.. :)  I also don't see an obvious solution yet how to
> > tie these lists of ioctls together.
> 
> I guess it should be ok to manually watch the do_vfs_ioctl() changes, but
> definitely not optimal.
> 
> > 
> > 
> > > I'm also wondering if we should not split the IOCTL access right into
> > > three: mostly read, mostly write, and misc. _IOC_READ and _IOC_WRITE are
> > > definitely not perfect, but tied to specific drivers (i.e. not a file
> > > hierarchy but a block or character device) this might help until we get
> > > a more fine-grained IOCTL access control. We should check if it's worth
> > > it according to commonly used drivers. Looking at the TTY driver, most
> > > IOCTLs are without read or write markers. Using this split could induce
> > > a false sense of security, so it should be well motivated.
> > 
> > As it was pointed out by the LWN article that Jeff Xu pointed to [1], this
> > read/write bit in the ioctl command number is only referring to whether the
> > *argument pointer* to the ioctl is being read or written, but you can not use
> > this bit to infer whether the ioctl itself performs a "reading" or "writing"
> > access to the underlying file.
> > 
> > As the LWN article explains, SELinux has fallen for the same trap in the past,
> > the post [2] has an example for an ioctl where the read/write bits for the
> > argument are not related to what the underlying operation does.
> > 
> > It might be that you could potentially use the _IOC_READ and _IOC_WRITE bits to
> > group smaller subsets of the ioctl cmd space, such as for a single device type.
> > But then, the users would have to go through the list of supported ioctls one by
> > one anyway, to ensure that this works for that subset.  If they are going
> > through them one by one anyway, they might maybe just as well list them out in
> > the filter rule...?
> > 
> > [1] https://lwn.net/Articles/428140
> > [2] https://lwn.net/Articles/428142/
> 
> Right, I fell again in this trap, !_IOC_READ cannot even guarantee non-write
> actions.
> 
> A useful split would be at least between devices and regular
> files/directories, something like this:
> - LANDLOCK_ACCESS_FS_IOCTL_DEV: allows IOCTLs on character or block devices,
> which should be targeted on specific paths.
> - LANDLOCK_ACCESS_FS_IOCTL_NODEV: allows IOCTLs on regular files,
> directories, unix sockets, pipes, and symlinks. These are targeting
> filesystems (e.g. ext4's fsverity) or common Linux file types.

To make sure we are on the same page, let me paraphrase:

You are suggesting that we should split the LANDLOCK_ACCESS_FS_IOCTL
right into a LANDLOCK_ACCESS_FS_IOCTL_DEV part (for block and
character devices) and a LANDLOCK_ACCESS_FS_IOCTL_NODEV part (for
regular files, directories, named(!) unix sockets, named(!) pipes and
symlinks)?  The check would presumably be done during the open(2) call
and then store the access right on the freshly opened struct file?

If Landlock only checks the ioctl criteria during open(2), that would
mean that file descriptors created through other means would be
unaffected.

In particular, the protection would only apply to named pipes and Unix
sockets which get newly opened through the file system, but it would
not apply to pipes created through pipe(2) and Unix sockets created
through socketpair(2)?

(It is more clearly a philosophy of "protecting resources", rather
than a philosophy of limiting access to the thousands of potentially
buggy ioctl implementations. - But I think it might be reasonable to
permit unnamed pipes and socketpairs - they are useful mechanisms and
seem harmless as long as their implementations don't have bugs.)


> I think it makes sense because the underlying filesystems should already
> check for read/write access, which is not the case for block/char devices.
> Pipe and unix socket IOCTLs are quite specific but don't touch the
> underlying filesystem, and it should be allowed to properly use them. It
> should be noted that the pipe and socket IOCTL implementations don't care
> about their file mode though; I guess the rationale might be that IOCTLs may
> be required to (efficiently) either read or write.
> 

I don't understand your remark about the read/write access.

Pipes have a read end and a write end, where only one of the two
operations should work.  Unix sockets are always bidirectional, if I
remember this correctly.

> Reading your following comments, this dev/nodev classification would be like
> the *file type* item, but simpler and only for file descriptors accessible
> through the filesystem, which I guess could be everything because of procfs…
> 
> This split might also help for the landlock_inode_attr properties, but it
> would also be a bit redundant with the file type match…

I agree that this dev/nodev classification seems like a simpler
version of the *file type* item from below.

> 
> 
> > 
> > 
> > I've also pondered more about the ioctl support. I have a work-in-progress patch
> > set which filters ioctls according to various criteria, but it's not fully
> > fleshed out yet.
> > 
> > In the big picture: I think that the main ways how we can build this differently
> > are (a) the criteria on which to decide whether an ioctl is permitted, and (b)
> > the time at which we evaluate these criteria (time of open() vs. time of
> > ioctl()).  We can also evaluate the criteria at different times, depending on
> > which criterium it is.
> > 
> > So:
> > 
> > (a) The criteria on which to decide that an ioctl is permitted:
> > 
> >      We have discussed the followowing ones so far:
> > 
> >      * The *ioctl cmd* (request) itself
> >         - needs to be taken into account, obviously.
> >         - ioctl cmds do not have an obvious ordering exposed to userspace,
> >           so asking users to specify ranges is potentially difficult
> >         - asking users to list all individual ioctls they do might result in
> >           lists that are larger than I had thought. I've straced Firefox and
> >           found that it did about 20-30 direct-rendering related ioctls, and most
> >           of them were specific to my graphics card... o_O so I assume that there
> >           are more of these for other graphics cards.
> > 
> >      * The *file device ID* (major / minor)
> >         - specifying ranges is a good idea - ranges of device IDs are logically
> >           grouped and the order is also exposed and documented to user space.
> > 
> >      * The *file type*, read from filp->f_mode
> >         - includes regular files, directories, char devices, block devices,
> >           fifos and sockets
> >         - BUT this list of types in non-exhaustive:
> >           - there are valid struct file objects which have special types and are
> >             not distinguishable. They might not have a file type set in f_mode,
> >             even.  Examples include pidfds, or the Landlock ruleset FD. -- so: we
> >             do need a way to ignore file type altogether in an ioctl rule, so
> >             that such "special" file types can still be matched in the rule.
> > 
> >      * The *file path*
> >         - This can only really be checked against at open() time, imho.
> >           Doing it during the ioctl is too late, because the file might
> >           have been created in a different mount namespace, and then the
> >           current thread can't really make that file work with ioctls.
> >         - Not all open files *have* a file path (i.e. sockets, Landlock ruleset)
> 
> I think we can reach a lot through /proc/self/fd/

What I meant to say is: The struct file for some files does not refer
to a path on the file system that the file was opened from ==> Using
the file path as criterium does not cover all existing ioctl use
cases.

The thing that struck me about the above list of criteria is that each
of them seems to have gaps.  As an example, take timerfds
(timerfd_create(2)):

 * these do not get opened through a file system path, so the *file
   path* can not restrict them.
 * they are not character or block devices and do not have a device ID.
 * they don't match any of the file types in filp->f_mode.

So in order to permit the TFD_IOC_SET_TICKS ioctl on them, these three
criteria can't be used to describe a timerfd.

This is more important in an implementation where the criteria are
checked in security_file_ioctl, rather than in security_file_open.  In
an implementation where the criteria are only checked in
security_file_open, it would anyway not be possible to restrict ioctls
on the timerfd, and all files for which it would be possible, they
must have a path in the file system when they end up in that hook, I
suspect?

> > (b) The time at which the criteria are checked:
> > 
> >      * During open():
> >         - A check at this time is necessary to match against file paths, imho,
> >           as we already to in the ioctl patch set I've sent.
> > 
> >      * During ioctl():
> >         - A check at this time is *also* necessary, because without that, we will
> >           not be able to restrict ioctls on TTYs and other file descriptors that
> >           are obtained from other processes.

For completeness: I forgot to list here: The other reason where a
check during ioctl() is needed is the case as for the timerfd, the
pipe(2) and socketpair(2), where a file is created through a simple
syscall, but without spelling out a path.  If these kinds of files are
in scope for ioctl protection, it can't be done during the open()
check alone, I suspect?

> As I explained before, I don't think we should care about inherited or
> passed FDs. Other ways to get FDs (e.g. landlock_create_ruleset) should
> probably not be a priority for now.

I don't know what we should do about the "basic Unix tool" and
TIOCSTI/TIOCLINUX case, where it is possible to gain control over the
shell running in the tty that we get as stdout fd.

I'm in that situation with the little web application I run at home,
but the patch that you have sent for GNU tar at some point (and which
we should really revive :)) has the same problem: If an attacker
manages to do a Remote Code Execution in that tar process, they can
ioctl(1, TIOCSTI, ...) their way out into the shell which invoked tar,
and which is not restricted with tar's Landlock policy.

(I don't really see tar create a pty/tty pair either and shovel data
between them in a sidecar process or thread, just to protect against
that.)

Remark: For the specific TIOCSTI problem, I'm seeing a glimmer of
light with this patch set which has appeared in the meantime:
https://lore.kernel.org/all/20230710002645.v565c7xq5iddruse@begin/
(This will still require that distributions flip that Kconfig option
off, but the only(?) known user of TIOCSTI, BRLTTY, would continue
working.)

I would be more comfortable with doing the checks only at open(2) time
if the above patch landed in distributions so that you would need to
have CAP_SYS_ADMIN in order to use TIOCSTI.

Do you think this is realistic?  If this does not get flipped by
distributions, Landlock would continue to have these TIOCSTI problems
on these platforms (unless its users do the pty/tty pair thing, but
that seems like an unrealistic demand).


> > The tentative approach I've taken in my own patch set and the WIP part so far is:
> > 
> >   (1) Do file path checks at open() time (modeled as a access_fs right)
> >   (2) Do cmds, device ID and file type checks at ioctl() time.
> >       This is modeled independently of the file path check. -- both checks need to
> >       pass independently for an ioctl invocation to be permitted.
> 
> This looks good! However, see below an alternative approach for the rules
> combination.
> 
> > 
> > The API of that approach is:
> >   * The ruleset attribute gets a new handled_misc field,
> >     and when setting the first bit in it, it'll deny all ioctls
> >     unless there is a special ioctl rule added for them
> >     (and the path of the file was OK for ioctl at open time).
> >   * A new rule type with an associated struct landlock_ioctl_attr -
> >     that struct lets users define:
> >       - the desired mask of file types (or 0 for "all")
> >       - the designed device ID range
> >       - the list of ioctl cmds to be permitted for such files
> > 
> > An open question is whether the ruleset attr's "handled_misc" field should
> > rather be a "handled_ioctl_cmds" field, a set of restricted ioctl cmds
> > (potentially [0, 2^32)).  I think that would be more consistent conceptually
> > with how it was done for file system access rights, but obviously we can't model
> > it as a bit field any more - it would have to be some other suitable
> > representation of a set of integers, which also lets people say "all ioctls".
> 
> We might not need another ruleset's field because we can reuse the existing
> FS access rights, including the new IOCTL one(s). The trick is just to
> define a new way to match files (and optionally specific IOCTL commands),
> hence the landlock_inode_attr proposal. As you explained, this type of rule
> could match device IDs and file types.

I have to think about it and maybe try it out in code.  This might be
a better option if we go for doing the checks only at open(2) time.

I do think that device IDs are often a better way to specify device
files than their paths are.  Device IDs are a stable numbering scheme
that won't change, whereas the structure of /dev can be defined by
user space and is also often dynamically adding and removing devices.

> An alternative way to identify such
> properties would be to pass an FD and specify a subset of these properties
> to match on. This would avoid some side channel issues, and could be used to
> check for directory or file (as done for path_beneath) to avoid irrelevant
> access rights.

I don't fully understand what you mean here.  Do you mean to use an
open device file as example for what to match?  I don't see how
specifying the file type and device ID range as plain numbers could
lead to a race condition.


> I suggest to first handle path_beneath and inode rules as a binary OR set of
> rights (i.e. the sandboxed processes only needs one rule to match for the
> defined action to be allowed). Then, with a way to identify inodes rules, we
> could treat them as synthetic access rights and add a new allowed_inode
> field to path_beneath rules and remove the IOCTL access rights from their
> allowed_access field. This way, sandboxed processes would need both rules to
> match.

I'm not sure what the inode rule is.  Do you mean "ioctl rule"?

If yes, I do agree that a list of permitted ioctls is similar to the
access rights flags that we already have, and it would have to get
passed around in a similar fashion (as "synthetic access rights"),
albeit using a different data structure.

I'm still skeptical of the API approach where we tie previously
unrelated rules together, if that is what you mean here.  I find this
difficult to explain and reason about.  But in doubt we'll see in the
implementation how unwieldy it actually gets.


> > The upside of that approach would be that it could also be used to selectively
> > restrict specific known-evil ioctls, and letting all others continue to work.
> > For example, sandboxing or sudo-like programs could filter out TIOCSTI and
> > TIOCLINUX.

By the way, selectively restricting known-bad ioctls is still not
possible with the approach we discussed now, I think.  Maybe TIOCSTI
is the only bad one... I hope.

> > 
> > I'd be interested in hearing your opinion about this (also from the Chromium
> > side).
> > 
> > Thanks,
> > —Günther
> > 

Thanks,
–Günther