commit a93f581705787105ac28bccdc179778fd4400298
author Victor Boivie <boivie@webrtc.org> Mon Sep 25 07:53:22 2023
committer WebRTC LUCI CQ <webrtc-scoped@luci-project-accounts.iam.gserviceaccount.com> Thu Sep 28 11:31:02 2023
tree cb9ae729f7514741eeba42cf94dc06ee8803a9fd
parent d8633868b34dc1d841f0a9fd1afe2bc22aa8bde6

dcsctp: Don't generate FORWARD-TSN across stream resets

This was a fun bug which proved to be challenging to find a good
solution for. The issue comes from the combination of partial
reliability and stream resetting, which are covered in different RFCs,
and where they don't refer to each other...

Stream resetting (RFC 6525) is used in WebRTC for closing a Data
Channel, and is done by signaling to the receiver that the stream
sequence number (SSN) should be set to zero (0) at some time. Partial
reliability (RFC 3758) - and expiring messages that will not be
retransmitted - is done by signaling that the SSN should be set to a
certain value at a certain TSN, as the messages up until the provided
SSN are not to be expected to be sent again.

As these two functionalities both work by signaling to the receiver
what the next expected SSN should be, they need to do it correctly not
to overwrite each others' intent. And here was the bug. An example
scenario where this caused issues, where we are Z (the receiver),
getting packets from the sender (A):

 5  A->Z          DATA (TSN=30, B, SID=2, SSN=0)
 6          Z->A  SACK (Ack=30)
 7  A->Z          DATA (TSN=31, E, SID=2, SSN=0)
 8  A->Z          RE_CONFIG (REQ=30, TSN=31, SID=2)
 9          Z->A  RE_CONFIG (RESP=30, Performed)
10          Z->A  SACK (Ack=31)
11  A->Z          DATA (TSN=32, SID=1)
12  A->Z          FORWARD_TSN (TSN=32, SID=2, SSN=0)

Let's assume that the path Z->A had packet loss and A never really
received our responses (#6, #9, #10) in time.

At #5, Z receives a DATA fragment, which it acks, and at #7 the end of
that message. The stream is then reset (#8) which it signals that it
was performed (#9) and acked (#10), and data on another stream (2) was
received (#11). Since A hasn't received any ACKS yet, and those chunks
on SID=2 all expired, A sends a FORWARD-TSN saying that "Skip to TSN=32,
and don't expect SID=2, SSN=0". That makes the receiver expect the SSN
on SID=2 to be SSN=1 next time at TSN=32.

But that's not good at all - A reset the stream at #8 and will want to
send the next message on SID=2 using SSN=0 - not 1. The FORWARD-TSN
clearly can't have a TSN that is beyond the stream reset TSN for that
stream.

This is just one example - combining stream resetting and partial
reliability, together with a lossy network, and different variants of
this can occur, which results in the receiver possibly not delivering
packets because it expects a different SSN than the one the sender is
later using.

So this CL adds "breakpoints" to how far a FORWARD-TSN can stretch. It
will simply not cross any Stream Reset last assigned TSNs, and only when
a receiver has acked that all TSNs up till the Stream Reset last
assigned TSN has been received, it will proceed expiring chunks after
that.

Bug: webrtc:14600
Change-Id: Ibae8c9308f5dfe8d734377d42cce653e69e95731
Reviewed-on: https://webrtc-review.googlesource.com/c/src/+/321600
Commit-Queue: Victor Boivie <boivie@webrtc.org>
Reviewed-by: Harald Alvestrand <hta@webrtc.org>
Cr-Commit-Position: refs/heads/main@{#40829}