Re: [PATCH v19 08/12] landlock: Add syscall implementation

On Wed, Jul 8, 2020 at 3:04 PM Mickaël Salaün [off-list ref] wrote:

On 08/07/2020 10:57, Arnd Bergmann wrote:

quoted

On Tue, Jul 7, 2020 at 8:10 PM Mickaël Salaün [off-list ref] wrote:

It looks like all you need here today is a single argument bit, plus
possibly some room for extensibility. I would suggest removing all
the extra bits and using a syscall like

SYSCALL_DEFINE1(landlock_create_ruleset, u32, flags);

I don't really see how this needs any variable-length arguments,
it really doesn't do much.

We need the attr_ptr/attr_size pattern because the number of ruleset
properties will increase (e.g. network access mask).

But how many bits do you think you will *actually* need in total that
this needs to be a two-dimensional set of flags? At the moment you
only have a single bit that you interpret.

quoted

To be on the safe side, you might split up the flags into either the
upper/lower 16 bits or two u32 arguments, to allow both compatible
(ignored by older kernels if flag is set) and incompatible (return error
when an unknown flag is set) bits.

This may be a good idea in general, but in the case of Landlock, because
this kind of (discretionary) sandboxing should be a best-effort security
feature, we should avoid incompatible behavior. In practice, every
unknown bit returns an error because userland can probe for available
bits thanks to the get_features command. This kind of (in)compatibility
can then be handled by userland.

If there are not going to be incompatible extensions, then just ignore
all unknown bits and never return an error but get rid of the user
space probing that just complicates the interface.

In general, it's hard to rely on user space to first ask the kernel
what it can do, the way this normally works is that user space
asks the kernel for something and it either does it or not, but gives
an indication of whether it worked.

I suggest this syscall signature:
SYSCALL_DEFINE3(landlock_create_ruleset, __u32, options, const struct
landlock_attr_ruleset __user *, ruleset_ptr, size_t, ruleset_size);

The other problem here is that indirect variable-size structured arguments
are a pain to instrument with things like strace or seccomp, so you
should first try to use a fixed argument list, and fall back to a fixed
structure if that fails.

quoted

quoted

+static int syscall_add_rule_path_beneath(const void __user *const attr_ptr,
+               const size_t attr_size)
+{
+       struct landlock_attr_path_beneath attr_path_beneath;
+       struct path path;
+       struct landlock_ruleset *ruleset;
+       int err;

Similarly, it looks like this wants to be

SYSCALL_DEFINE3(landlock_add_rule_path_beneath, int, ruleset, int,
path, __u32, flags)

I don't see any need to extend this in a way that wouldn't already
be served better by adding another system call. You might argue
that 'flags' and 'allowed_access' could be separate, with the latter
being an indirect in/out argument here, like

SYSCALL_DEFINE4(landlock_add_rule_path_beneath, int, ruleset, int, path,
                           __u64 *, allowed_acces, __u32, flags)

To avoid adding a new syscall for each new rule type (e.g. path_beneath,
path_range, net_ipv4_range, etc.), I think it would be better to keep
the attr_ptr/attr_size pattern and to explicitely set a dedicated option
flag to specify the attr type.

This would look like this:
SYSCALL_DEFINE4(landlock_add_rule, __u32, options, int, ruleset, const
void __user *, rule_ptr, size_t, rule_size);

The rule_ptr could then point to multiple types like struct
landlock_attr_path_beneath (without the current ruleset_fd field).

This again introduces variable-sized structured data. How many different
kinds of rule types do you think there will be (most likely, and maybe an
upper bound)?

Could (some of) these be generalized to use the same data structure?

quoted

quoted

+static int syscall_enforce_ruleset(const void __user *const attr_ptr,
+               const size_t attr_size)

Here it seems like you just need to pass the file descriptor, or maybe

SYSCALL_DEFINE2(landlock_enforce, int, ruleset, __u32 flags);

if you need flags for extensibility.

Right, but for consistency I prefer to change the arguments like this:
SYSCALL_DEFINE2(landlock_enforce, __u32 options, int, ruleset);

Most system calls pass the object they work on as the first argument,
in this case this would be the ruleset file descriptor.

     Arnd