Page MenuHomeFreeBSD
Differential D23364 Diff 67297 sys/netinet/tcp_stacks/rack.c

Changeset View

Standalone View

View Options

sys/netinet/tcp_stacks/rack.c

  • This file is larger than 256 KB, so syntax highlighting is disabled by default.
Show First 20 Lines • Show All 1,793 Lines • ▼ Show 20 Lines if (!IN_FASTRECOVERY(tp->t_flags)) {
rack_log_to_prr(rack, 2); rack_log_to_prr(rack, 2);
rack->r_ctl.rc_prr_recovery_fs = tp->snd_max - tp->snd_una; rack->r_ctl.rc_prr_recovery_fs = tp->snd_max - tp->snd_una;
tp->snd_recover = tp->snd_max; tp->snd_recover = tp->snd_max;
if (tp->t_flags2 & TF2_ECN_PERMIT) if (tp->t_flags2 & TF2_ECN_PERMIT)
tp->t_flags2 |= TF2_ECN_SND_CWR; tp->t_flags2 |= TF2_ECN_SND_CWR;
} }
break; break;
case CC_ECN: case CC_ECN:
if (!IN_CONGRECOVERY(tp->t_flags)) { if (!IN_CONGRECOVERY(tp->t_flags) ||
/*
* allow ECN reaction on ACK to CWR, if
* that data segment was also CE marked
*/
SEQ_GEQ(th->th_ack, tp->snd_recover)) {
rscheffAuthorUnsubmitted
Done
Inline Actions

At this stage, the congestion episode was not yet finished on an ACK covering snd_recover (both loss or ECN cc-reaction).

As per RFC guidance, only a single CC reaction per window is required; Since ECN and loss are tracked by two flags, we may enter (and reduce cwnd) independently.
However, exit is done only when th_ack >= snd_recover - and snd_recover is pulled forward by whatever happened last.

Most problematic may be non-SACK, NewReno recovery, where an ECN after loss could prolong the fast recovery episode (a subsequent loss episode may not result in another cc reaction, if overlapped by an ECN - however, there already were two cc reactions earlier).

rscheff: At this stage, the congestion episode was not yet finished on an ACK covering snd_recover (both…
EXIT_CONGRECOVERY(tp->t_flags);
TCPSTAT_INC(tcps_ecn_rcwnd); TCPSTAT_INC(tcps_ecn_rcwnd);
tp->snd_recover = tp->snd_max; tp->snd_recover = tp->snd_max+1;
rscheffAuthorUnsubmitted
Done
Inline Actions

As CWR will be sent only on the next new data segment, the congestion epoch is over at snd_max+1 with ECN, rather than when snd_max is cumulative acked as in loss recovery.

rscheff: As CWR will be sent only on the next new data segment, the congestion epoch is over at…
if (tp->t_flags2 & TF2_ECN_PERMIT) if (tp->t_flags2 & TF2_ECN_PERMIT)
tp->t_flags2 |= TF2_ECN_SND_CWR; tp->t_flags2 |= TF2_ECN_SND_CWR;
} }
break; break;
case CC_RTO: case CC_RTO:
tp->t_dupacks = 0; tp->t_dupacks = 0;
tp->t_bytes_acked = 0; tp->t_bytes_acked = 0;
EXIT_RECOVERY(tp->t_flags); EXIT_RECOVERY(tp->t_flags);
▲ Show 20 Lines • Show All 7,661 Lines • ▼ Show 20 Lines#endif
if (tp->t_state == TCPS_ESTABLISHED && if (tp->t_state == TCPS_ESTABLISHED &&
(tp->t_flags2 & TF2_ECN_PERMIT)) { (tp->t_flags2 & TF2_ECN_PERMIT)) {
/* /*
* If the peer has ECN, mark data packets with ECN capable * If the peer has ECN, mark data packets with ECN capable
* transmission (ECT). Ignore pure ack packets, * transmission (ECT). Ignore pure ack packets,
* retransmissions and window probes. * retransmissions and window probes.
*/ */
if (len > 0 && SEQ_GEQ(tp->snd_nxt, tp->snd_max) && if (len > 0 && SEQ_GEQ(tp->snd_nxt, tp->snd_max) &&
(sack_rxmit == 0) && (sack_rxmit == 0) &&
!((tp->t_flags & TF_FORCEDATA) && len == 1)) { !((tp->t_flags & TF_FORCEDATA) && len == 1)) {
#ifdef INET6 #ifdef INET6
rscheffAuthorUnsubmitted
Done
Inline Actions

this is to address the case, where persist window probes are done using new 1-octet data segments.

rscheff: this is to address the case, where persist window probes are done using new 1-octet data…
rscheffAuthorUnsubmitted
Done
Inline Actions

rS360639 had removed the simple check for window probe packets (snd_nxt == snd_max, len==1, but no forward progress).

@rrs what was the motivation for that change, how do you exclude setting ECT0 on window probes in RACK?

rscheff: rS360639 had removed the simple check for window probe packets (snd_nxt == snd_max, len==1, but…
if (isipv6) if (isipv6)
ip6->ip6_flow |= htonl(IPTOS_ECN_ECT0 << 20); ip6->ip6_flow |= htonl(IPTOS_ECN_ECT0 << 20);
else else
#endif #endif
ip->ip_tos |= IPTOS_ECN_ECT0; ip->ip_tos |= IPTOS_ECN_ECT0;
TCPSTAT_INC(tcps_ecn_ect0); TCPSTAT_INC(tcps_ecn_ect0);
}
/* /*
* Reply with proper ECN notifications. * Reply with proper ECN notifications.
*/ */
if (tp->t_flags2 & TF2_ECN_SND_CWR) { if (tp->t_flags2 & TF2_ECN_SND_CWR) {
flags |= TH_CWR; flags |= TH_CWR;
tp->t_flags2 &= ~TF2_ECN_SND_CWR; tp->t_flags2 &= ~TF2_ECN_SND_CWR;
}
} }
if (tp->t_flags2 & TF2_ECN_SND_ECE) if (tp->t_flags2 & TF2_ECN_SND_ECE)
flags |= TH_ECE; flags |= TH_ECE;
} }
/* /*
* If we are doing retransmissions, then snd_nxt will not reflect * If we are doing retransmissions, then snd_nxt will not reflect
* the first unsent octet. For ACK only packets, we do not want the * the first unsent octet. For ACK only packets, we do not want the
* sequence number of the retransmitted packet, we want the sequence * sequence number of the retransmitted packet, we want the sequence
▲ Show 20 Lines • Show All 1,066 Lines • Show Last 20 Lines