On Wed, May 13, 2020 at 01:58 AM CEST, Alexei Starovoitov wrote:
> On Tue, May 12, 2020 at 03:52:52PM +0200, Jakub Sitnicki wrote:
>> On Mon, May 11, 2020 at 10:44 PM CEST, Alexei Starovoitov wrote:
>> > On Mon, May 11, 2020 at 08:52:06PM +0200, Jakub Sitnicki wrote:
>> >> Run a BPF program before looking up a listening socket on the receive path.
>> >> Program selects a listening socket to yield as result of socket lookup by
>> >> calling bpf_sk_assign() helper and returning BPF_REDIRECT code.
>> >>
>> >> Alternatively, program can also fail the lookup by returning with BPF_DROP,
>> >> or let the lookup continue as usual with BPF_OK on return.
>> >>
>> >> This lets the user match packets with listening sockets freely at the last
>> >> possible point on the receive path, where we know that packets are destined
>> >> for local delivery after undergoing policing, filtering, and routing.
>> >>
>> >> With BPF code selecting the socket, directing packets destined to an IP
>> >> range or to a port range to a single socket becomes possible.
>> >>
>> >> Suggested-by: Marek Majkowski <marek@xxxxxxxxxxxxxx>
>> >> Reviewed-by: Lorenz Bauer <lmb@xxxxxxxxxxxxxx>
>> >> Signed-off-by: Jakub Sitnicki <jakub@xxxxxxxxxxxxxx>
>> >> ---
>> >> include/net/inet_hashtables.h | 36 +++++++++++++++++++++++++++++++++++
>> >> net/ipv4/inet_hashtables.c | 15 ++++++++++++++-
>> >> 2 files changed, 50 insertions(+), 1 deletion(-)
>> >>
>> >> diff --git a/include/net/inet_hashtables.h b/include/net/inet_hashtables.h
>> >> index 6072dfbd1078..3fcbc8f66f88 100644
>> >> --- a/include/net/inet_hashtables.h
>> >> +++ b/include/net/inet_hashtables.h
>> >> @@ -422,4 +422,40 @@ int __inet_hash_connect(struct inet_timewait_death_row *death_row,
>> >>
>> >> int inet_hash_connect(struct inet_timewait_death_row *death_row,
>> >> struct sock *sk);
>> >> +
>> >> +static inline struct sock *bpf_sk_lookup_run(struct net *net,
>> >> + struct bpf_sk_lookup_kern *ctx)
>> >> +{
>> >> + struct bpf_prog *prog;
>> >> + int ret = BPF_OK;
>> >> +
>> >> + rcu_read_lock();
>> >> + prog = rcu_dereference(net->sk_lookup_prog);
>> >> + if (prog)
>> >> + ret = BPF_PROG_RUN(prog, ctx);
>> >> + rcu_read_unlock();
>> >> +
>> >> + if (ret == BPF_DROP)
>> >> + return ERR_PTR(-ECONNREFUSED);
>> >> + if (ret == BPF_REDIRECT)
>> >> + return ctx->selected_sk;
>> >> + return NULL;
>> >> +}
>> >> +
>> >> +static inline struct sock *inet_lookup_run_bpf(struct net *net, u8 protocol,
>> >> + __be32 saddr, __be16 sport,
>> >> + __be32 daddr, u16 dport)
>> >> +{
>> >> + struct bpf_sk_lookup_kern ctx = {
>> >> + .family = AF_INET,
>> >> + .protocol = protocol,
>> >> + .v4.saddr = saddr,
>> >> + .v4.daddr = daddr,
>> >> + .sport = sport,
>> >> + .dport = dport,
>> >> + };
>> >> +
>> >> + return bpf_sk_lookup_run(net, &ctx);
>> >> +}
>> >> +
>> >> #endif /* _INET_HASHTABLES_H */
>> >> diff --git a/net/ipv4/inet_hashtables.c b/net/ipv4/inet_hashtables.c
>> >> index ab64834837c8..f4d07285591a 100644
>> >> --- a/net/ipv4/inet_hashtables.c
>> >> +++ b/net/ipv4/inet_hashtables.c
>> >> @@ -307,9 +307,22 @@ struct sock *__inet_lookup_listener(struct net *net,
>> >> const int dif, const int sdif)
>> >> {
>> >> struct inet_listen_hashbucket *ilb2;
>> >> - struct sock *result = NULL;
>> >> + struct sock *result, *reuse_sk;
>> >> unsigned int hash2;
>> >>
>> >> + /* Lookup redirect from BPF */
>> >> + result = inet_lookup_run_bpf(net, hashinfo->protocol,
>> >> + saddr, sport, daddr, hnum);
>> >> + if (IS_ERR(result))
>> >> + return NULL;
>> >> + if (result) {
>> >> + reuse_sk = lookup_reuseport(net, result, skb, doff,
>> >> + saddr, sport, daddr, hnum);
>> >> + if (reuse_sk)
>> >> + result = reuse_sk;
>> >> + goto done;
>> >> + }
>> >> +
>> >
>> > The overhead is too high to do this all the time.
>> > The feature has to be static_key-ed.
>>
>> Static keys is something that Lorenz has also suggested internally, but
>> we wanted to keep it simple at first.
>>
>> Introduction of static keys forces us to decide when non-init_net netns
>> are allowed to attach to SK_LOOKUP, as attaching enabling SK_LOOKUP in
>> isolated netns will affect the rx path in init_net.
>>
>> I see two options, which seem sensible:
>>
>> 1) limit SK_LOOKUP to init_net, which makes testing setup harder, or
>>
>> 2) allow non-init_net netns to attach to SK_LOOKUP only if static key
>> has been already enabled (via sysctl?).
>
> I think both are overkill.
> Just enable that static_key if any netns has progs.
> Loading this prog type will be privileged operation even after cap_bpf.
>
OK, right. In the new model caps are checked at load time. And
CAP_BPF+CAP_NET_ADMIN check on load is done against init_user_ns.
[...]
>> I'm curious what downside do you see here?
>> Is overriding an earlier DROP/REDIRECT verdict useful?
>>
>> > Another option could be to execute all attached progs regardless
>> > of return code, but don't let second prog override selected_sk blindly.
>> > bpf_sk_assign() could get smarter.
>>
>> So if IIUC the rough idea here would be like below?
>>
>> - 1st program calls
>>
>> bpf_sk_assign(ctx, sk1, 0 /*flags*/) -> 0 (OK)
>>
>> - 2nd program calls
>>
>> bpf_sk_assign(ctx, sk2, 0) -> -EBUSY (already selected)
>> bpf_sk_assign(ctx, sk2, BPF_EXIST) -> 0 (OK, replace existing)
>>
>> In this case the last program to run has the final say, as opposed to
>> the semantics where DROP/REDIRECT terminates.
>>
>> Also, 2nd and subsequent programs would probably need to know if and
>> which socket has been already selected. I think the selection could be
>> exposed in context as bpf_sock pointer.
>
> I think running all is better.
> The main down side of terminating early is predictability.
> Imagine first prog is doing the sock selection based on some map configuration.
> Then second prog gets loaded and doing its own selection.
> These two progs are managed by different user space processes.
> Now first map got changed and second prog stopped seeing the packets.
> No warning. Nothing. With "bpf_sk_assign(ctx, sk2, 0) -> -EBUSY"
> the second prog at least will see errors and will be able to log
> and alert humans to do something about it.
> The question of ordering come up, of course. But that ordering concerns
> we had for some time with cgroup-bpf run array and it wasn't horrible.
> We're still trying to solve it on cgroup-bpf side in a generic way,
> but simple first-to-attach -> first-to-run was good enough there
> and I think will be here as well. The whole dispatcher project
> and managing policy, priority, ordering in user space better to solve
> it generically for all cases. But the kernel should do simple basics.
That makes sense. Thanks for guidance.
-Jakub
