Re: Unrecoverable loss pattern

[Martin Thomson <[e-mail]>]

Two timers solves the problem because it leads to the request being
sent on TLP (or RTO), which unsticks the whole thing.  That is, if you
consider the problem to be less that the Finished isn't read, but more
that the handshake timers disable TLP/RTO.

On Fri, Mar 2, 2018 at 6:13 AM, Ian Swett <[e-mail]> wrote:
> I'm a bit confused about which case we're concerned by, so I'll try to talk
> through my understanding.
>
> Two timers won't solve the problem of the ack of the Finished being lost(and
> then the acknowledgement never being sent again) that Marten was discussing,
> because as Christian said, we won't process re-transmissions of the Finished
> message because they're not encrypted and it will be stuck in an
> undeliverable/unackable state.
>
> The one byte frame could be an explicit signal to move on, but we still need
> a way to stop re-transmitting the Finished.  Does that frame serve as an
> acknowledgement of the Finished as well?  Doesn't the receipt of encrypted
> packets also serve as an acknowledgement of the Finished?  In either case,
> we're discussing an implicit ack of sorts.
>
> One totally different(and possibly poor) alternative would be to only send
> unencrypted handshake data on stream 0, and when you receive an encrypted
> packet and your handshake has completed, if stream 0 isn't already closed,
> reset it, since you'd know the handshake has completed.  We'd have to
> allocate a different stream for post-handshake encrypted data, but that may
> be a win anyway, since it would clearly define which streams were encrypted
> and which weren't?
>
> Related Q: At some point, weren't we discussing not sending finished?
>
>
>
> On Mon, Feb 26, 2018 at 1:45 PM Christian Huitema <[e-mail]>
> wrote:
>>
>>
>>
>> On 2/26/2018 12:51 AM, Martin Thomson wrote:
>> > The issue in #1018 (as I understood it) was the fact that Finished
>> > could be protected (and thereby authenticated), but it was not.  Not
>> > having it protected makes it vulnerable to all sorts of attack.
>> >
>> > This is different, it's not an attack, it's just an unfortunate
>> > interaction between the loss recovery algorithms and the TLS mapping.
>>
>> Actually, the issue was that allowing Finished to be repeated in clear
>> text allows for a DOS attack on the server. The spec changed to ignore
>> handshake packets after Finished is received once. That closes the DOS
>> issue, but it opens the "ACK of Finished" issue.
>>
>> >
>> > Other than running multiple timers as I suggested, you could also
>> > force the server to respond to cleartext packets with a repeat of its
>> > ACK.  It doesn't even have to look at what it received, it just plays
>> > out the last ACK.  That's what we do in our DTLS 1.3 implementation.
>> > It works.  The trick there is that once you have an ACK for the ACKs
>> > that acknowledge the handshake you stop the special behaviour.
>> >
>> That was actually the first solution I thought off. That works too, but
>> there are two downsides. First, you get a "twitch trigger" -- send an
>> handshake packet, get the server to send an ACK. Not a very big deal,
>> but slightly annoying. Second, it relies not only on implementing ACK of
>> ACK, but also on triggering a special behavior for the ACK of the ACK of
>> the last handshake packet. Many implementations do ACK of ACK
>> management, but grafting a special case in that path is a bit tricky.
>> That was my thought process before proposing a simple one byte
>> "handshake complete" frame. The special frame is easier to integrate in
>> the receive processing.
>>
>> -- Christian Huitema
>>
>