On 8/29/16 12:24 PM, Tejun Heo wrote:
Hello, Sargun.
On Mon, Aug 29, 2016 at 11:49:07AM -0700, Sargun Dhillon wrote:
It would be a separate hook per LSM hook. Why wouldn't we want a separate bpf
hook per lsm hook? I think if one program has to handle them all, the first
program would be looking up the hook program in a bpf prog array. If you think
it's better to have this logic in the BPF program, that makes sense.
I had a version of this patch that allowed you to attach a prog array instead,
but I think that it's cleaner attaching a program per lsm hook. In addition,
there's a performance impact that comes from these hooks, so I wouldn't want to
execute unneccessary code if it's avoidable.
Hmm... it doesn't really matter how the backend part looks like and if
we need to implement per-call hooks to lower runtime overhead, sure.
I was mostly worried about the approach propagating through the
userland visible interface.
The prog array approach also makes stacking filters difficult. If people want
multiple filters per hook, the orchestrator would have to rewrite the existing
filters to be cooperative.
I'm not really sure "stacking" in the kernel side is a good idea.
Please see below.
I'm not convinced about the approach. It's an approach which pretty
much requires future extensions while being rigid. Not a good
combination.
Do you have an alternative recommendation? Maybe just a set of 5 u64s
as the context object along with the hook ID?
cgroup fs doesn't seem like the right interface for this but if it
were I'd go for named hook IDs instead of opaque numbers.
Unless this is properly delegatable, IOW, it's safe to fully delegate
to a lesser security domain for all operations including program
loading and assignment (I can't see how that'd be the case), making it
an explicit controller doens't work in terms of userland interface.
It's fine for bpf / lsm / whatever to attach to cgroups by extending
struct cgroup itself or implementing an implicit controller but to be
visible as an explicit controller it must be able to follow cgroup
interface rules including delegation. If not, it's best to come
through the interface which enforces the required permission checks
and then talk to cgroup from there. This was also an issue with
network cgroup bpf programs that Daniel Mack is working on. Please
chat with him.
Program assignment is possible by lesser security domains. Program loading is
limited to CAP_SYS_ADMIN in init_user_ns. We could make it accessible to
CAP_SYS_ADMIN in any userns, but it the reasoning behind this is that Checmate
BPF programs can leak kernel pointers.
That doesn't make much sense to me. Delegation doesn't mean much if a
delegatee can't load its own program (and I don't see how one can
delegate kernel pointer access to !root). Also, unless there's
per-program fine control on who can load it, it seems pretty dangerous
to let anyone load any program.
Could we potentially restrict it to only CAP_MAC_OVERRIDE, while still meeting
cgroup delegation requirements?
Wouldn't it make far more sense to pass cgroup fd to bpf syscall so
that "load this program" and "attach this program to the cgroup
identified by this fd" through the same interface and permission
checks? cgroup participating in bpf operations is all fine but
splitting the userland interface across two domains seems like a bad
idea.
Filters which are higher up in the heirarchy will still be enforced during
delegation. This was an explicit design, as the "Orchestrator in Orchestrator"
use case needs to be supported.
Given that program loading is restricted to root, wouldn't it be an a
lot more efficient approach to let userland multiplex multiple
programs? Walking the tree executing bpf programs each time one of
these operations runs can be pretty expensive. Imagine a tree like
the following.
A - B - C
\ D
Let's say program is currently loaded on D. If someone wants to add a
program on B, userland can load the program on B, combine B's and D's
program and compile them into a single program and load it on D. The
only thing kernel would need to do in terms of hierarchy is finding
what's the closest program to execute. In the above example, C would
need to use B's program and that can be determined on program
assignment time rather than on each operation.
I think that's exactly what Daniel's patches are doing and imo
it makes sense to keep this style for lsm as well
and also apply the concept of hook_id.
Daniel adds two commands to bpf syscall to attach/detach from cgroup
with hook_id.
Initially two hooks will be for socket rx and tx.
Then all interesting lsm hooks can be added one by one.
Daniel's prog type will be bpf_prog_type_cgroup_socket_filter.
LSM's prog type will be bpf_prog_type_lsm.
And verifier can check type safety since the lsm hook_id will be
passed at the program load time.
See another thread we had with Mickael.
landlock and checmate are very similar and should really be
single lsm as long as we agree that both are cgroup based.
The main difference between the two:
- landlock is proposing unpriveleged mode
- checmate is proposing writing into arguments from the program
These differences can be flags/options to one lsm.
Implementations of course are different so far, but
instead of arguing landlock vs checmate, I'd like us
to focus on how we can make one lsm that solves all use cases.
