Re: Allowing space for packet headers in Wintun Tx/Rx

[Daniel Lenski <dlenski@gmail.com>]

On Sat, Apr 10, 2021 at 7:35 AM David Woodhouse <dwmw2@infradead.org> wrote:
> On Sat, 2021-04-10 at 13:38 +0000, Simon Rozman wrote:
> > Hi David,This is my proposal:
> > https://git.zx2c4.com/wintun/commit/?id=eebd6aea4f75551f6e847a1d4fff857450bac6e9
> > Awaiting review and zx2c4 approval.
> > Regards, Simon
>
>
> Looks good to me; thanks. Just need to work out how to cross-build it
> (I can muster up a Windows VM for testing, but *building* on it is
> beyond my tolerance of Windows for now).

+1 to all that.

> We'll also need to be able to WintunAllocateSendPacket() of the full
> possible MTU, then receive and decrypt into that, and send only the
> actual size of the packet we received.
>
> A per-packet tail would have let us do that, but I agree that we don't
> want to expand the TUN_PACKET header if we can avoid doing so.
>
> Perhaps a WintunShrinkAndSendPacket() — which can only *shrink*, of
> course, and which can only be used on the *last* packet allocated,
> checking that its tail *is* the Session->Receive.Tail before adjusting
> the latter accordingly.

In addition to the use case for chopping ESP trailers and
less-than-full-size packets, OpenConnect has the case of "PPP packets
in HDLC-like framing" which need to be "un-HDLC-ed" in a way that can
only cause them to shrink.
(https://gitlab.com/openconnect/openconnect/blob/master/ppp.c#L102-158)

There are two cases worth considering where the packet size could
actually *expand*:

1) Some VPN protocols support compression of the tunneled packets. It
would be bad behavior to use this to stuff a packet of >(advertised
MTU) bytes in <(advertised MTU) bytes, but it wouldn't surprise me if
it exists in the wild. We now deal with receipt of
larger-than-expected-MTU packets in OpenConnect in a relatively
uniform way: allocate MAX(mtu, 16384) bytes for packets coming from
the VPN (if using TLS transport) or MAX(mtu, 2048) if using DTLS.
2) Some VPN protocols concatenate multiple packets into a single
aggregate on the wire. On Linux we can decrypt, truncate, and send to
the tunnel interface without further copying.

Case (1) can be handled with overallocate-and-shrink. Case (2) is
pretty rare among the protocols that OpenConnect supports, so fallback
to memcpy seems fine.

Dan