Re: [RFC bpf-next 0/5] Extend SOCKMAP to store listening sockets

On Mon, Oct 28, 2019 at 01:35:26PM +0100, Jakub Sitnicki wrote:

On Mon, Oct 28, 2019 at 06:52 AM CET, Martin Lau wrote:

quoted

On Tue, Oct 22, 2019 at 01:37:25PM +0200, Jakub Sitnicki wrote:

quoted

This patch set is a follow up on a suggestion from LPC '19 discussions to
make SOCKMAP (or a new map type derived from it) a generic type for storing
established as well as listening sockets.

We found ourselves in need of a map type that keeps references to listening
sockets when working on making the socket lookup programmable, aka BPF
inet_lookup [1].  Initially we repurposed REUSEPORT_SOCKARRAY but found it
problematic to extend due to being tightly coupled with reuseport
logic (see slides [2]).
So we've turned our attention to SOCKMAP instead.

As it turns out the changes needed to make SOCKMAP suitable for storing
listening sockets are self-contained and have use outside of programming
the socket lookup. Hence this patch set.

With these patches SOCKMAP can be used in SK_REUSEPORT BPF programs as a
drop-in replacement for REUSEPORT_SOCKARRAY for TCP. This can hopefully
lead to code consolidation between the two map types in the future.

What is the plan for UDP support in sockmap?

It's on our road-map because without SOCKMAP support for UDP we won't be
able to move away from TPROXY [1] and custom SO_BINDTOPREFIX extension
[2] for steering new UDP flows to receiving sockets. Also we would like
to look into using SOCKMAP for connected UDP socket splicing in the
future [3].

I was planning to split work as follows:

1. SOCKMAP support for listening sockets (this series)
2. programmable socket lookup for TCP (cut-down version of [4])
3. SOCKMAP support for UDP (work not started)

hmm...It is hard to comment how the full UDP sockmap may
work out without a code attempt because I am not fluent in
sock_map ;)

From a quick look, it seems there are quite a few things to do.
For example, the TCP_SKB_CB(skb) usage and how that may look
like in UDP.  "struct udp_skb_cb" is 28 bytes while "struct napi_gro_cb"
seems to be 48 bytes already which may need a closer look.

4. programmable socket lookup for UDP (rest of [4])

I'm open to suggestions on how to organize it.

quoted

quoted

Having said that, the main intention here is to lay groundwork for using
SOCKMAP in the next iteration of programmable socket lookup patches.

What may be the minimal to get only lookup work for UDP sockmap?
.close() and .unhash()?

John would know better. I haven't tried doing it yet.

From just reading the code - override the two proto ops you mentioned,
close and unhash, and adapt the socket checks in SOCKMAP.

Do your use cases need bpf prog attached to sock_map?

If not, would it be cleaner to delicate another map_type
for lookup-only use case to have both TCP and UDP support.

-Jakub

[1] https://urldefense.proofpoint.com/v2/url?u=https-3A__blog.cloudflare.com_how-2Dwe-2Dbuilt-2Dspectrum_&d=DwIBAg&c=5VD0RTtNlTh3ycd41b3MUw&r=VQnoQ7LvghIj0gVEaiQSUw&m=lSo-FsOeNl_8znZZ07H8I6ZYAinPKTR5C3Cn_Ol3QYQ&s=DZgW8-2Xl1P8NU59ji4ieQLzwWpx4t3gGq_tqB0l3Bo&e= 
[2] https://lore.kernel.org/netdev/1458699966-3752-1-git-send-email-gilberto.bertin@gmail.com/ (local)
[3] https://lore.kernel.org/bpf/20190828072250.29828-1-jakub@cloudflare.com/ (local)
[4] https://urldefense.proofpoint.com/v2/url?u=https-3A__blog.cloudflare.com_sockmap-2Dtcp-2Dsplicing-2Dof-2Dthe-2Dfuture_&d=DwIBAg&c=5VD0RTtNlTh3ycd41b3MUw&r=VQnoQ7LvghIj0gVEaiQSUw&m=lSo-FsOeNl_8znZZ07H8I6ZYAinPKTR5C3Cn_Ol3QYQ&s=NerUqb4j7IsGBTcni6Yxk40wf6kTkckHXn3Nx5i4mCU&e=