Re: [PATCH bpf-next v1 00/15] Introduce typed pointer support in BPF maps

On Tue, Feb 22, 2022 at 11:35:14AM IST, Song Liu wrote:

On Sun, Feb 20, 2022 at 5:48 AM Kumar Kartikeya Dwivedi
[off-list ref] wrote:

quoted

Introduction
------------

This set enables storing pointers of a certain type in BPF map, and extends the
verifier to enforce type safety and lifetime correctness properties.

The infrastructure being added is generic enough for allowing storing any kind
of pointers whose type is available using BTF (user or kernel) in the future
(e.g. strongly typed memory allocation in BPF program), which are internally
tracked in the verifier as PTR_TO_BTF_ID, but for now the series limits them to
four kinds of pointers obtained from the kernel.

Obviously, use of this feature depends on map BTF.

1. Unreferenced kernel pointer

In this case, there are very few restrictions. The pointer type being stored
must match the type declared in the map value. However, such a pointer when
loaded from the map can only be dereferenced, but not passed to any in-kernel
helpers or kernel functions available to the program. This is because while the
verifier's exception handling mechanism coverts BPF_LDX to PROBE_MEM loads,
which are then handled specially by the JIT implementation, the same liberty is
not available to accesses inside the kernel. The pointer by the time it is
passed into a helper has no lifetime related guarantees about the object it is
pointing to, and may well be referencing invalid memory.

2. Referenced kernel pointer

This case imposes a lot of restrictions on the programmer, to ensure safety. To
transfer the ownership of a reference in the BPF program to the map, the user
must use the BPF_XCHG instruction, which returns the old pointer contained in
the map, as an acquired reference, and releases verifier state for the
referenced pointer being exchanged, as it moves into the map.

This a normal PTR_TO_BTF_ID that can be used with in-kernel helpers and kernel
functions callable by the program.

However, if BPF_LDX is used to load a referenced pointer from the map, it is
still not permitted to pass it to in-kernel helpers or kernel functions. To
obtain a reference usable with helpers, the user must invoke a kfunc helper
which returns a usable reference (which also must be eventually released before
BPF_EXIT, or moved into a map).

Since the load of the pointer (preserving data dependency ordering) must happen
inside the RCU read section, the kfunc helper will take a pointer to the map
value, which must point to the actual pointer of the object whose reference is
to be raised. The type will be verified from the BTF information of the kfunc,
as the prototype must be:

        T *func(T **, ... /* other arguments */);

Then, the verifier checks whether pointer at offset of the map value points to
the type T, and permits the call.

This convention is followed so that such helpers may also be called from
sleepable BPF programs, where RCU read lock is not necessarily held in the BPF
program context, hence necessiating the need to pass in a pointer to the actual
pointer to perform the load inside the RCU read section.

3. per-CPU kernel pointer

These have very little restrictions. The user can store a PTR_TO_PERCPU_BTF_ID
into the map, and when loading from the map, they must NULL check it before use,
because while a non-zero value stored into the map should always be valid, it can
still be reset to zero on updates. After checking it to be non-NULL, it can be
passed to bpf_per_cpu_ptr and bpf_this_cpu_ptr helpers to obtain a PTR_TO_BTF_ID
to underlying per-CPU object.

It is also permitted to write 0 and reset the value.

4. Userspace pointer

The verifier recently gained support for annotating BTF with __user type tag.
This indicates pointers pointing to memory which must be read using the
bpf_probe_read_user helper to ensure correct results. The set also permits
storing them into the BPF map, and ensures user pointer cannot be stored
into other kinds of pointers mentioned above.

When loaded from the map, the only thing that can be done is to pass this
pointer to bpf_probe_read_user. No dereference is allowed.

I guess I missed some context here. Could you please provide some reference
to the use cases of these features?

The common usecase is caching references to objects inside BPF maps, to avoid
costly lookups, and being able to raise it once for the duration of program
invocation when passing it to multiple helpers (to avoid further re-lookups).
Storing references also allows you to control object lifetime.

One other use case is enabling xdp_frame queueing in XDP using this, but that
still needs some integration work after this lands, so it's a bit early to
comment on the specifics.

Other than that, I think Alexei already mentioned this could be easily extended
to do memory allocation returning a PTR_TO_BTF_ID in a BPF program [0] in the
future.

  [0]: https://lore.kernel.org/bpf/20220216230615.po6huyrgkswk7u67@ast-mbp.dhcp.thefacebook.com (local)

For Unreferenced kernel pointer and userspace pointer, it seems that there is
no guarantee the pointer will still be valid during access (we only know it is
valid when it is stored in the map). Is this correct?

That is correct. In the case of unreferenced and referenced kernel pointers,
when you do a BPF_LDX, both are marked as PTR_UNTRUSTED, and it is not allowed
to pass them into helpers or kfuncs, because from that point onwards we cannot
claim that the object is still alive when pointer is used later. Still,
dereference is permitted because verifier handles faults for bad accesses using
PROBE_MEM conversion for PTR_TO_BTF_ID loads in convert_ctx_accesses (which is
then later detected by JIT to build exception table used by exception handler).

In case of reading unreferenced pointer, in some cases you know that the pointer
will stay valid, so you can just store it in the map and load and directly
access it, it imposes very little restrictions.

For the referenced case, and BPF_LDX marking it as PTR_UNTRUSTED, you could say
that this makes it a lot less useful, because if BPF program already holds
reference, just to make sure I _read valid data_, I still have to use the
kptr_get style helper to raise and put reference to ensure the object is alive
when it is accessed.

So in that case, for RCU protected objects, it should still wait for BPF program
to hit BPF_EXIT before the actual release, but for other cases like the case of
sleepable programs, or objects where refcount alone manages lifetime, you can
also detect writer presence of the other BPF program (to detect if pointer
during our access was xchg'd out) using a seqlock style scheme:

	v = bpf_map_lookup_elem(&map, ...);
	if (!v)
		return 0;
	seq_begin = v->seq;
	atomic_thread_fence(memory_order_acquire); // A
	<do access>
	atomic_thread_fence(memory_order_acquire); // B
	seq_end = v->seq;
	if (seq_begin & 1 || seq_begin != seq_end)
		goto bad_read;
	<use data>

Ofcourse, barriers are not yet in BPF, but you get the idea (it should work on
x86). The updater BPF program will increment v->seq before and after xchg,
ensuring proper ordering. v->seq starts as 0, so odd seq indicates writer update
is in progress.

This would allow you to not raise refcount, while still ensuring that as long as
object was accessed, it was still valid between A and B. Even if raising
uncontended refcount is cheap, this is much cheaper.

The case of userspace pointer is different, it sets the MEM_USER flag, so the
only useful thing to do is calling bpf_probe_read_user, you can't even
dereference it. You are right that in most cases that userspace pointer won't be
useful, but for some cooperative cases between BPF program and userspace thread,
it can act as a way to share certain thread local areas/userspace memory that
the BPF program can then store keyed by the task_struct *, where using a BPF map
to share memory is not always possible.

Thanks,
Song

[...]

--
Kartikeya