Re: [PATCH bpf-next 3/7] bpf: Add socket assign support

On Thu, Mar 19, 2020 at 6:55 PM Martin KaFai Lau [off-list ref] wrote:

On Wed, Mar 18, 2020 at 11:24:11PM -0700, Joe Stringer wrote:

quoted

On Wed, Mar 18, 2020 at 11:49 AM Martin KaFai Lau [off-list ref] wrote:

quoted

On Tue, Mar 17, 2020 at 05:46:58PM -0700, Joe Stringer wrote:

quoted

On Mon, Mar 16, 2020 at 11:27 PM Martin KaFai Lau [off-list ref] wrote:

quoted

On Mon, Mar 16, 2020 at 08:06:38PM -0700, Joe Stringer wrote:

quoted

On Mon, Mar 16, 2020 at 3:58 PM Martin KaFai Lau [off-list ref] wrote:

quoted

On Thu, Mar 12, 2020 at 04:36:44PM -0700, Joe Stringer wrote:

quoted

Add support for TPROXY via a new bpf helper, bpf_sk_assign().

This helper requires the BPF program to discover the socket via a call
to bpf_sk*_lookup_*(), then pass this socket to the new helper. The
helper takes its own reference to the socket in addition to any existing
reference that may or may not currently be obtained for the duration of
BPF processing. For the destination socket to receive the traffic, the
traffic must be routed towards that socket via local route, the socket

I also missed where is the local route check in the patch.
Is it implied by a sk can be found in bpf_sk*_lookup_*()?

This is a requirement for traffic redirection, it's not enforced by
the patch. If the operator does not configure routing for the relevant
traffic to ensure that the traffic is delivered locally, then after
the eBPF program terminates, it will pass up through ip_rcv() and
friends and be subject to the whims of the routing table. (or
alternatively if the BPF program redirects somewhere else then this
reference will be dropped).

Maybe there's a path to simplifying this configuration path in future
to loosen this requirement, but for now I've kept the series as
minimal as possible on that front.

quoted

[ ... ]

quoted

diff --git a/net/core/filter.c b/net/core/filter.c
index cd0a532db4e7..bae0874289d8 100644
--- a/net/core/filter.c
+++ b/net/core/filter.c

@@ -5846,6 +5846,32 @@ static const struct bpf_func_proto bpf_tcp_gen_syncookie_proto = {
      .arg5_type      = ARG_CONST_SIZE,
 };

+BPF_CALL_3(bpf_sk_assign, struct sk_buff *, skb, struct sock *, sk, u64, flags)
+{
+     if (flags != 0)
+             return -EINVAL;
+     if (!skb_at_tc_ingress(skb))
+             return -EOPNOTSUPP;
+     if (unlikely(!refcount_inc_not_zero(&sk->sk_refcnt)))
+             return -ENOENT;
+
+     skb_orphan(skb);
+     skb->sk = sk;

sk is from the bpf_sk*_lookup_*() which does not consider
the bpf_prog installed in SO_ATTACH_REUSEPORT_EBPF.
However, the use-case is currently limited to sk inspection.

It now supports selecting a particular sk to receive traffic.
Any plan in supporting that?

I think this is a general bpf_sk*_lookup_*() question, previous
discussion[0] settled on avoiding that complexity before a use case
arises, for both TC and XDP versions of these helpers; I still don't
have a specific use case in mind for such functionality. If we were to
do it, I would presume that the socket lookup caller would need to
pass a dedicated flag (supported at TC and likely not at XDP) to
communicate that SO_ATTACH_REUSEPORT_EBPF progs should be respected
and used to select the reuseport socket.

It is more about the expectation on the existing SO_ATTACH_REUSEPORT_EBPF
usecase.  It has been fine because SO_ATTACH_REUSEPORT_EBPF's bpf prog
will still be run later (e.g. from tcp_v4_rcv) to decide which sk to
recieve the skb.

If the bpf@tc assigns a TCP_LISTEN sk in bpf_sk_assign(),
will the SO_ATTACH_REUSEPORT_EBPF's bpf still be run later
to make the final sk decision?

I don't believe so, no:

ip_local_deliver()
-> ...
-> ip_protocol_deliver_rcu()
-> tcp_v4_rcv()
-> __inet_lookup_skb()
-> skb_steal_sock(skb)

But this will only affect you if you are running both the bpf@tc
program with sk_assign() and the reuseport BPF sock programs at the
same time.

I don't think it is the right answer to ask the user to be careful and
only use either bpf_sk_assign()@tc or bpf_prog@so_reuseport.

Applying a restriction on reuseport sockets until we sort this out per
my other email should resolve this concern.

quoted

quoted

This is why I link it back to the bpf_sk*_lookup_*()
functions: If the socket lookup in the initial step respects reuseport
BPF prog logic and returns the socket using the same logic, then the
packet will be directed to the socket you expect. Just like how
non-BPF reuseport would work with this series today.

Changing bpf_sk*_lookup_*() is a way to solve it but I don't know what it
may run into when recurring bpf_prog, i.e. running bpf@so-reuseport inside
bpf@tc. That may need a closer look.

Right, that's my initial concern as well.

One alternative might be something like: in the helper implementation,
store some bit somewhere to say "we need to resolve the reuseport
later" and then when the TC BPF program returns, check this bit and if
reuseport is necessary, trigger the BPF program for it and fix up the
socket after-the-fact.

skb_dst_is_sk_prefetch() could be that bit.  One major thing
is that bpf@so_reuseport is currently run at the transport layer
and expecting skb->data pointing to udp/tcp hdr.  The ideal
place is to run it there.

My initial thought above was much simpler - just holding it long
enough to exit the current BPF program so we know that the tc@bpf
program terminates, then preparing & running the so_reuseport program.

However, the skb_dst_is_sk_prefetch() bit
is currently lost at ip[6]_rcv_core.

Yeah, I think this is tricky. Here's three paths I'm tracking right now:
* Loopback destination is already assigned to skb, so we wrap that
with the dst_sk_prefetch in sk_assign. We unwrap it again currently in
ip[6]_rcv_core, but if we retained it, we'd need to convince
ip[6]_rcv_finish_core() to respect the metadata dst, then convince
ip[6]_rcv_finish() to call the right receive destination function.
* In the connected case of patch #4 in this series, we wrap the
rx_dst. Needs similar treatment.
* In the regular initial packet case for non-loopback destination, we
don't actually have a destination yet so the routing check would need
to track this bit and ensure it's propagated through that routing
check and again back out to ip[6]_rcv_finish() and call the right
destination receive.

Even if we get through all of that and we get up to the transport
layer, one of the main points of this feature is to guide the packet
to a socket that may be associated with a different tuple, so AFAIK we
end up needing another path even up at the transport layer to check
this bit and jump straight to the reuseport_select_sock() call.
Looking at __udp4_lib_rcv() path I'm eyeing the conditional branch
after stealing the socket, in there we'd need more special casing for
the new call to select the reuseport socket.

Following that rabbit-hole, it seems less invasive to either get the
exact right reuseport socket in the first place or something closer to
the simpler approach above.

quoted

A bit uglier though, also not sure how socket
refcounting would work there; maybe we can avoid the refcount in the
socket lookup and then fix it up in the later execution.

That should not be an issue if refcnt is not taken for
SOCK_RCU_FREE (e.g. TCP_LISTEN) in the first place.

This is likely the part that I was missing---I assumed that all bets
were off when we rcu_read_unlock(). But it sounds like you're saying
it's actually the RCU grace period, and that won't happen before we
enqueue the skb at the socket?

quoted

quoted

[...]
It is another question that I have.  The TCP_LISTEN sk will suffer
from this extra refcnt, e.g. SYNFLOOD.  Can something smarter
be done in skb->destructor?

Can you elaborate a bit more on the idea you have here?

I am thinking can skb->destructor do something like bpf_sk_release()?
This patch reuses tcp sock_edemux which currently only lookups the
established sk.

I can try it out.

quoted

Looking at the BPF API, it seems like the writer of the program can
use bpf_tcp_gen_syncookie() / bpf_tcp_check_syncookie() to generate
and check syn cookies to mitigate this kind of attack. This at least
provides an option beyond what existing tproxy implementations
provide.

When the SYNACK comes back, it will still be served by a TCP_LISTEN sk.
I know refcnt sucks on synflood test.  I don't know what the effect
may be on serving those valid synack since there is no need
to measure after SOCK_RCU_FREE is in ;)

UDP is also in SOCK_RCU_FREE.  I think only early_demux, which
seems to be for connected only,  takes a refnct.
btw, it may be a good idea to add a udp test.

UDP was next on my list, in my local testing I needed a new
skc_lookup_udp() to find the right sockets. That patch series is more
straightforward than this one so I thought it'd be better to get the
feedback on this socket assign approach first then follow up with UDP
support.

I am fine to push them to optimize/support later bucket
It is still good to explore a little more such that we don't
regret later.

I'll dig around a bit.

Thanks for the feedback,
Joe