On 12/20/23 5:28 PM, Kuniyuki Iwashima wrote:
Under SYN Flood, the TCP stack generates SYN Cookie to remain stateless
for the connection request until a valid ACK is responded to the SYN+ACK.
The cookie contains two kinds of host-specific bits, a timestamp and
secrets, so only can it be validated by the generator. It means SYN
Cookie consumes network resources between the client and the server;
intermediate nodes must remember which nodes to route ACK for the cookie.
SYN Proxy reduces such unwanted resource allocation by handling 3WHS at
the edge network. After SYN Proxy completes 3WHS, it forwards SYN to the
backend server and completes another 3WHS. However, since the server's
ISN differs from the cookie, the proxy must manage the ISN mappings and
fix up SEQ/ACK numbers in every packet for each connection. If a proxy
node goes down, all the connections through it are terminated. Keeping
a state at proxy is painful from that perspective.
At AWS, we use a dirty hack to build truly stateless SYN Proxy at scale.
Our SYN Proxy consists of the front proxy layer and the backend kernel
module. (See slides of LPC2023 [0], p37 - p48)
The cookie that SYN Proxy generates differs from the kernel's cookie in
that it contains a secret (called rolling salt) (i) shared by all the proxy
nodes so that any node can validate ACK and (ii) updated periodically so
that old cookies cannot be validated and we need not encode a timestamp for
the cookie. Also, ISN contains WScale, SACK, and ECN, not in TS val. This
is not to sacrifice any connection quality, where some customers turn off
TCP timestamps option due to retro CVE.
After 3WHS, the proxy restores SYN, encapsulates ACK into SYN, and forward
the TCP-in-TCP packet to the backend server. Our kernel module works at
Netfilter input/output hooks and first feeds SYN to the TCP stack to
initiate 3WHS. When the module is triggered for SYN+ACK, it looks up the
corresponding request socket and overwrites tcp_rsk(req)->snt_isn with the
proxy's cookie. Then, the module can complete 3WHS with the original ACK
as is.
This way, our SYN Proxy does not manage the ISN mappings nor wait for
SYN+ACK from the backend thus can remain stateless. It's working very
well for high-bandwidth services like multiple Tbps, but we are looking
for a way to drop the dirty hack and further optimise the sequences.
If we could validate an arbitrary SYN Cookie on the backend server with
BPF, the proxy would need not restore SYN nor pass it. After validating
ACK, the proxy node just needs to forward it, and then the server can do
the lightweight validation (e.g. check if ACK came from proxy nodes, etc)
and create a connection from the ACK.
This series allows us to create a full sk from an arbitrary SYN Cookie,
which is done in 3 steps.
1) At tc, BPF prog calls a new kfunc to create a reqsk and configure
it based on the argument populated from SYN Cookie. The reqsk has
its listener as req->rsk_listener and is passed to the TCP stack as
skb->sk.
2) During TCP socket lookup for the skb, skb_steal_sock() returns a
listener in the reuseport group that inet_reqsk(skb->sk)->rsk_listener
belongs to.
3) In cookie_v[46]_check(), the reqsk (skb->sk) is fully initialised and
a full sk is created.
The kfunc usage is as follows:
struct bpf_tcp_req_attrs attrs = {
.mss = mss,
.wscale_ok = wscale_ok,
.rcv_wscale = rcv_wscale, /* Server's WScale < 15 */
.snd_wscale = snd_wscale, /* Client's WScale < 15 */
.tstamp_ok = tstamp_ok,
.rcv_tsval = tsval,
.rcv_tsecr = tsecr, /* Server's Initial TSval */
.usec_ts_ok = usec_ts_ok,
.sack_ok = sack_ok,
.ecn_ok = ecn_ok,
}
skc = bpf_skc_lookup_tcp(...);
sk = (struct sock *)bpf_skc_to_tcp_sock(skc);
bpf_sk_assign_tcp_reqsk(skb, sk, attrs, sizeof(attrs));
bpf_sk_release(skc);
[0]: https://lpc.events/event/17/contributions/1645/attachments/1350/2701/SYN_Proxy_at_Scale_with_BPF.pdf
Changes:
v7:
* Patch 5 & 6
* Drop MPTCP support
I think Yonghong's (thanks!) cpuv4 patch
(https://lore.kernel.org/bpf/20240110051348.2737007-1-yonghong.song@xxxxxxxxx/)
has addressed the issue that the selftest in patch 6 has encountered.
There are some minor comments in v7. Please respin v8 when the cpuv4 patch has
concluded so that it can kick off the CI also.
