From: Martin KaFai Lau <martin.lau@xxxxxxxxx>
Date: Fri, 20 Oct 2023 12:59:00 -0700
> On 10/19/23 11:01 AM, Kuniyuki Iwashima wrote:
> > From: Martin KaFai Lau <martin.lau@xxxxxxxxx>
> > Date: Thu, 19 Oct 2023 00:25:00 -0700
> >> On 10/18/23 3:31 PM, Kuniyuki Iwashima wrote:
> >>> From: Kui-Feng Lee <sinquersw@xxxxxxxxx>
> >>> Date: Wed, 18 Oct 2023 14:47:43 -0700
> >>>> On 10/18/23 10:20, Kuniyuki Iwashima wrote:
> >>>>> From: Eric Dumazet <edumazet@xxxxxxxxxx>
> >>>>> Date: Wed, 18 Oct 2023 10:02:51 +0200
> >>>>>> On Wed, Oct 18, 2023 at 8:19 AM Martin KaFai Lau <martin.lau@xxxxxxxxx> wrote:
> >>>>>>>
> >>>>>>> On 10/17/23 9:48 AM, Kuniyuki Iwashima wrote:
> >>>>>>>> From: Martin KaFai Lau <martin.lau@xxxxxxxxx>
> >>>>>>>> Date: Mon, 16 Oct 2023 22:53:15 -0700
> >>>>>>>>> On 10/13/23 3:04 PM, Kuniyuki Iwashima wrote:
> >>>>>>>>>> Under SYN Flood, the TCP stack generates SYN Cookie to remain stateless
> >>>>>>>>>> After 3WHS, the proxy restores SYN and forwards it and ACK 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.
> >>>>>>>>>
> >>>>>>>>> Does the current kernel module also use the timestamp bits differently?
> >>>>>>>>> (something like patch 8 and patch 10 trying to do)
> >>>>>>>>
> >>>>>>>> Our SYN Proxy uses TS as is. The proxy nodes generate a random number
> >>>>>>>> if TS is in SYN.
> >>>>>>>>
> >>>>>>>> But I thought someone would suggest making TS available so that we can
> >>>>>>>> mock the default behaviour at least, and it would be more acceptable.
> >>>>>>>>
> >>>>>>>> The selftest uses TS just to strengthen security by validating 32-bits
> >>>>>>>> hash. Dropping a part of hash makes collision easier to happen, but
> >>>>>>>> 24-bits were sufficient for us to reduce SYN flood to the managable
> >>>>>>>> level at the backend.
> >>>>>>>
> >>>>>>> While enabling bpf to customize the syncookie (and timestamp), I want to explore
> >>>>>>> where can this also be done other than at the tcp layer.
> >>>>>>>
> >>>>>>> Have you thought about directly sending the SYNACK back at a lower layer like
> >>>>>>> tc/xdp after receiving the SYN?
> >>>>>
> >>>>> Yes. Actually, at netconf I mentioned the cookie generation hook will not
> >>>>> be necessary and should be replaced with XDP.
> >>
> >> Right, it is also what I have been thinking when seeing the
> >> BPF_SOCK_OPS_GEN_SYNCOOKIE_CB carrying the bpf generated timestamp to the
> >> tcp_make_synack. It feels like trying hard to work with the tcp want_cookie
> >> logic while there is an existing better alternative in tc/xdp to deal with synflood.
> >>
> >>>>>
> >>>>>
> >>>>>>> There are already bpf_tcp_{gen,check}_syncookie
> >>>>>>> helper that allows to do this for the performance reason to absorb synflood. It
> >>>>>>> will be natural to extend it to handle the customized syncookie also.
> >>>>>
> >>>>> Maybe we even need not extend it and can use XDP as said below.
> >>>>>
> >>>>>
> >>>>>>>
> >>>>>>> I think it should already be doable to send a SYNACK back with customized
> >>>>>>> syncookie (and timestamp) at tc/xdp today.
> >>>>>>>
> >>>>>>> When ack is received, the prog@tc/xdp can verify the cookie. It will probably
> >>>>>>> need some new kfuncs to create the ireq and queue the child socket. The bpf prog
> >>>>>>> can change the ireq->{snd_wscale, sack_ok...} if needed. The details of the
> >>>>>>> kfuncs need some more thoughts. I think most of the bpf-side infra is ready,
> >>>>>>> e.g. acquire/release/ref-tracking...etc.
> >>>>>>>
> >>>>>>
> >>>>>> I think I mostly agree with this.
> >>>>>
> >>>>> I didn't come up with kfunc to create ireq and queue it to listener, so
> >>>>> cookie_v[46]_check() were best place for me to extend easily, but now it
> >>>>> sounds like kfunc would be the way to go.
> >>>>>
> >>>>> Maybe we can move the core part of cookie_v[46]_check() except for kernel
> >>>>> cookie's validation to __cookie_v[46]_check() and expose a wrapper of it
> >>>>> as kfunc ?
> >>>>>
> >>>>> Then, we can look up sk and pass the listener, skb, and flags (for sack_ok,
> >>>>> etc) to the kfunc. (It could still introduce some conflicts with Eric's
> >>>>> patch though...)
> >>>>
> >>>> Does that mean the packets handled in this way (in XDP) will skip all
> >>>> netfilter at all?
> >>>
> >>> Good point.
> >>>
> >>> If we want not to skip other layers, maybe we can use tc ?
> >>>
> >>> 1) allocate ireq and set sack_ok etc with kfunc
> >>> 2) bpf_sk_assign() to set ireq to skb (this could be done in kfunc above)
> >>> 3) let inet_steal_sock() return req->sk_listener if not sk_fullsock(sk)
> >>> 4) if skb->sk is reqsk in cookie_v[46]_check(), skip validation and
> >>> req allocation and create full sk
> >>
> >> Haven't looked at the details. The above feels reasonable and would be nice if
> >> it works out. don't know if the skb at tc can be used in cookie_v[46]_check() as
> >> is. It probably needs more thoughts. [ note, xdp does not have skb. ]
> >>
> >> Regarding the "allocate ireq and set sack_ok etc with kfunc", do you think it
> >> will be useful (and potentially cleaner) even for the
> >> BPF_SOCK_OPS_CHECK_SYNCOOKIE_CB if it needed to go back to consider skops? Then
> >> only do the BPF_SOCK_OPS_CHECK_SYNCOOKIE_CB and the xdp/tc can generate SYNACK.
> >> The xdp/tc can still do the check and drop the bad ACK earlier in the stack.
> >
> > kfunc would be useful if we want to fall back to the default
> > validation, but I think we should not allocate ireq in kfunc.
> >
> > The SOCK_OPS prog only returns a binary value. If we decide whether
> > we skip validation or not based on kfunc call (ireq allocation), the
> > flow would be like :
> >
> > 1. CG_OK & ireq is allocated -> skip validation and req allocation
> > 2. CG_OK & no ireq -> default validation
> > 3. CG_ERR -> RST
> >
> > The problem here is that if kfunc fails with -ENOMEM and cookie
> > is valid, we need a way to tell the kernel to drop the ACK instead
> > of sending RST. (I hope the prog could return CG_DROP...)
>
> bpf_set_retval() helper allows the cgrp bpf prog to return -ENOMEM. Take a look
> at how __cgroup_bpf_run_filter_getsockopt is using the return value of
> bpf_prog_run_array_cg() and an example in progs/cgroup_getset_retval_getsockopt.c.
Oh, this is nice, I assumed -EPERM was always returned.
> > If we allocate ireq first, it would be cleaner as bpf need not care
> > about the drop path.
> >
> > 1. CG_OK & mss is set -> skip validation
> > 2. CG_OK & no mss set -> default validation
> > 3. CG_ERR -> RST
>
> Even if it uses the mss set/not-set like above to decide drop/rst. Does it
> really need to pre-allocate ireq? Looking at the test, the bpf prog is not using
> the skops->sk either.
It uses skops->remote_ip4 etc, maybe this was another reason why
I chose pre-alloc, but yes, it's not needed. The same value can
be extraced from skb with bpf_skb_load_bytes_relative(BPF_HDR_START_NET).
> It would be nice to allow bpf prog to check the cookie first before creating
> ireq. The kernel also checks the cookie first before tcp_parse_option and ireq
> creation. Beside, I suspect the multiple "if ([!]bpf_cookie)" checks in
> cookie_v[46]_check() is due to the pre-alloc ireq requirement.
>
> What does it take to create an ireq? sk, skb, tcp_opt, and mss? Potentially, it
> could have a "bpf_skops_parse_tcp_options(struct bpf_sock_ops_kern *skops,
> struct tcp_options_received *opt_rx, u32 opt_rx__sz)" to initialize the tcp_opt.
> I think the bpf prog should be able to parse the tcp options by itself also and
> directly initialize the tcp_opt.
Yes, also the prog will not need to parse all the options unless
the validation algorithm needs to becaues SACK_PERMITTED, WSCALE,
MSS (and ECN bits) are only available in SYN.
So, the prog will just need to parse timestamps option with
bpf_load_hdr_opt() and can initialise tcp_opt based on ISN
(and/or TS).
> The "bpf_skops_alloc_tcp_req(struct bpf_sock_ops_kern *skops, struct
> tcp_options_received *opt_rx, u32 opt_rx__size, int mss,...)" could directly
> save the "ireq" in skops->ireq (new member). If skops->ireq is available, the
> kernel could then skip most of the ireq initialization and directly continue the
> remaining processing (e.g. directly to security_inet_conn_request() ?). would
> that work?
Yes, that will work.
