tcp: do not purge SACK scoreboard on first RTO

[rscheff]

Summary

The errate to RFC2018 have an important change - on
RTO, the SACK data SHOULD be purged, not MUST be purged.

Keeping the SACK scoreboard intact after the first RTO,
and discarding all data only on subsequent RTOs allows
a more timely and efficient loss recovery under many
cirumstances, where the RTO was due to an unrecoverable
lost retransmissions (LRD requires additional data to be
sent and SACKed, and a sufficiently large cwnd for this
to happen).

MFC after: 10 weeks

Test Plan

after some pathological losses, induce an RTO and check that
previously SACKed data is not retransmitted again.

[rscheff]

• perform clean SACK retransmissions after RTO

[rscheff]

In order for clean SACK LR after the RTO, we don't want to exit recovery, and we also need to clear the number of retransmitted sack bytes, for the cwnd calculation to work. Without these, the stack would oscillate between SACK and non-SACK retransmissions (snd_una), resulting in many duplicate sent packets.

There are two incidental edits here (missed one magic number and replaced with the proper enum, and a whitespace change).

Note that this approach is a bit brittle in my testing (with a pathological loss pattern of 3 lost - 1 delivered - 3 lost ..., as after the RTO, every packet is clocked out by a preceeding ACK (cwnd remains at 1, since I'm not resetting any of the prr variables, which could allow some cwnd growth during loss recovery).

[tuexen]

Maybe split the editorial changes in a separate commit...

sys/netinet/tcp_sack.c

924

Nit: no double space before p->start.

[cc]

sys/netinet/siftr.c

521 ↗

(On Diff #131159)

Strongly suggest not adding this file in this change for tracking purposes. It is so benign that can wait.

[rscheff]

• remove editorial change in siftr.c; keep change of magic number to enum for SACK

[rscheff]

• split off editorial changes to MFC independently