Re: [PATCH v4 6/7] net/tcp: Store SNEs + SEQs on ao_info

[Eric Dumazet <edumazet@google.com>]

On Wed, Nov 29, 2023 at 8:58 PM Dmitry Safonov <dima@arista.com> wrote:
>
> On 11/29/23 18:34, Eric Dumazet wrote:
> > On Wed, Nov 29, 2023 at 7:14 PM Dmitry Safonov <dima@arista.com> wrote:
> >>
> >> On 11/29/23 18:09, Eric Dumazet wrote:
> >>> On Wed, Nov 29, 2023 at 5:57 PM Dmitry Safonov <dima@arista.com> wrote:
> >>>>
> >>>> RFC 5925 (6.2):
> >>>>> TCP-AO emulates a 64-bit sequence number space by inferring when to
> >>>>> increment the high-order 32-bit portion (the SNE) based on
> >>>>> transitions in the low-order portion (the TCP sequence number).
> >>>>
> >>>> snd_sne and rcv_sne are the upper 4 bytes of extended SEQ number.
> >>>> Unfortunately, reading two 4-bytes pointers can't be performed
> >>>> atomically (without synchronization).
> >>>>
> >>>> In order to avoid locks on TCP fastpath, let's just double-account for
> >>>> SEQ changes: snd_una/rcv_nxt will be lower 4 bytes of snd_sne/rcv_sne.
> >>>>
> >>>
> >>> This will not work on 32bit kernels ?
> >>
> >> Yeah, unsure if there's someone who wants to run BGP on 32bit box, so at
> >> this moment it's already limited:
> >>
> >> config TCP_AO
> >>         bool "TCP: Authentication Option (RFC5925)"
> >>         select CRYPTO
> >>         select TCP_SIGPOOL
> >>         depends on 64BIT && IPV6 != m # seq-number extension needs WRITE_ONCE(u64)
> >>
> >
> > Oh well, this seems quite strange to have such a limitation.
>
> I guess so. On the other side, it seems that there aren't many
> non-hobbyist 32bit platforms: ia32 compatible layer will even be limited
> with a boot parameter/compile option. Maybe I'm not aware of, but it
> seems that arm64/ppc64/risc-v/x86_64 are the ones everyone interested in
> these days.
>
> >
> >> Probably, if there will be a person who is interested in this, it can
> >> get a spinlock for !CONFIG_64BIT.
> >
> >
> >>
> >>> Unless ao->snd_sne and ao->rcv_sneare only read/written under the
> >>> socket lock (and in this case no READ_ONCE()/WRITE_ONCE() should be
> >>> necessary)
> >>
> >
> > You have not commented on where these are read without the socket lock held ?
>
> Sorry for missing this, the SNEs are used with this helper
> tcp_ao_compute_sne(), so these places are (in square brackets AFAICS,
> there is a chance that I miss something obvious from your message):
>
> - tcp_v4_send_reset() => tcp_ao_prepare_reset() [rcu_read_lock()]
> - __tcp_transmit_skb() => tcp_ao_transmit_skb() [TX softirq]
> - tcp_v4_rcv() => tcp_inbound_ao_hash() [RX softirq]

All these should/must have the socket lock held !

Or reading tcp_sk(sk)->rcv_nxt would be racy anyway (note the lack of
READ_ONCE() on it)

I think you need more work to make sure this is done correctly.

ie tcp_inbound_hash() should be called from tcp_v4_do_rcv() after the
bh_lock_sock_nested() and sock_owned_by_user() checks.



>
>
> > tcp_ao_get_repair() can lock the socket.
>
> It can, sure.
>
> > In TW state, I guess these values can not be changed ?
>
> Currently, they are considered constant on TW. The incoming segments are
> not verified on twsk (so no need for SNEs). And from ACK side not
> expecting SEQ roll-over (tcp_ao_compute_sne() is not called) - this may
> change, but not quite critical it seems.
>
> If we go with this patch in question, I'll have to update this:
> :               key.sne = READ_ONCE(ao_info->snd_sne);
> (didn't adjust it for higher-bytes shift)
>
> > I think you can remove all these READ_ONCE()/WRITE_ONCE() which are not needed,
> > or please add a comment if they really are.
>
> Not sure if I answered above..
>
> > Then, you might be able to remove the 64BIT dependency ...
>
> At this moment I fail to imagine anyone running BGP + TCP-AO on 32bit
> kernel. I may be wrong, for sure.

I fail to see anyone using TCP-AO today. (up to linux-6.6), regardless
of the architecture.

Would that be a reason for not supporting it in the future ?

32bit or 64bit should not be in the picture, especially if done for
the wrong reasons.