Page MenuHomeFreeBSD
Files F160958022

D28931.id88001.diff
No OneTemporary
Actions

D28931.id88001.diff
View Options

Index: share/man/man4/tcp.4
===================================================================
--- share/man/man4/tcp.4
+++ share/man/man4/tcp.4
@@ -547,6 +547,13 @@
application limited and the network bandwidth is not utilized completely.
That prevents self-inflicted packet losses once the application starts to
transmit data at a higher speed.
+.It Va do_lrd
+Enable Lost Retransmission Detection for SACK-enabled sessions, disabled by
+default.
+Under severe congestion, a retransmission can be lost which then leads to a
+mandatory Retransmission Timeout (RTO), followed by slow-start.
+LRD will try to resend the repeatedly lost packet, preventing the time-consuming
+RTO and performance reducing slow-start.
.It Va do_prr
Perform SACK loss recovery using the Proportional Rate Reduction (PRR) algorithm
described in RFC6937.
Index: sys/netinet/tcp_input.c
===================================================================
--- sys/netinet/tcp_input.c
+++ sys/netinet/tcp_input.c
@@ -164,6 +164,11 @@
&VNET_NAME(tcp_do_prr), 1,
"Enable Proportional Rate Reduction per RFC 6937");
+VNET_DEFINE(int, tcp_do_lrd) = 0;
+SYSCTL_INT(_net_inet_tcp, OID_AUTO, do_lrd, CTLFLAG_VNET | CTLFLAG_RW,
+ &VNET_NAME(tcp_do_lrd), 1,
+ "Perform Lost Retransmission Detection");
+
VNET_DEFINE(int, tcp_do_newcwv) = 0;
SYSCTL_INT(_net_inet_tcp, OID_AUTO, newcwv, CTLFLAG_VNET | CTLFLAG_RW,
&VNET_NAME(tcp_do_newcwv), 0,
@@ -2516,9 +2521,12 @@
}
if ((tp->t_flags & TF_SACK_PERMIT) &&
((to.to_flags & TOF_SACK) ||
- !TAILQ_EMPTY(&tp->snd_holes)))
- sack_changed = tcp_sack_doack(tp, &to, th->th_ack);
- else
+ !TAILQ_EMPTY(&tp->snd_holes))) {
+ if (((sack_changed = tcp_sack_doack(tp, &to, th->th_ack)) != 0) &&
+ (V_tcp_do_lrd)) {
+ tcp_lost_retransmission(tp, th);
+ }
+ } else
/*
* Reset the value so that previous (valid) value
* from the last ack with SACK doesn't get used.
@@ -4070,3 +4078,56 @@
return (4 * maxseg);
}
}
+
+/*
+ * Lost Retransmission Detection
+ * Check is FACK is beyond the rexmit of the leftmost hole.
+ * If yes, we restart sending from still existing holes,
+ * and adjust cwnd via the congestion control module.
+ */
+void
+tcp_lost_retransmission(struct tcpcb *tp, struct tcphdr *th)
+{
+ struct sackhole *temp;
+ uint32_t prev_cwnd;
+ if (IN_RECOVERY(tp->t_flags) &&
+ SEQ_GT(tp->snd_fack, tp->snd_recover) &&
+ ((temp = TAILQ_FIRST(&tp->snd_holes)) != NULL) &&
+ SEQ_GEQ(temp->rxmit, temp->end) &&
+ SEQ_GEQ(tp->snd_fack, temp->rxmit)) {
+ /*
+ * Start retransmissions from the first hole, and
+ * subsequently all other remaining holes, including
+ * those, which had been sent completely before.
+ */
+ tp->sackhint.nexthole = temp;
+ TAILQ_FOREACH(temp, &tp->snd_holes, scblink) {
+ if (SEQ_GEQ(tp->snd_fack, temp->rxmit) &&
+ SEQ_GEQ(temp->rxmit, temp->end))
+ temp->rxmit = temp->start;
+ }
+ /*
+ * Remember the old ssthresh, to deduct the beta factor used
+ * by the CC module. Finally, set cwnd to ssthresh just
+ * prior to invoking another cwnd reduction by the CC
+ * module, to not shrink it excessively.
+ */
+ prev_cwnd = tp->snd_cwnd;
+ tp->snd_cwnd = tp->snd_ssthresh;
+ /*
+ * Formally exit recovery, and let the CC module adjust
+ * ssthresh as intended.
+ */
+ EXIT_RECOVERY(tp->t_flags);
+ cc_cong_signal(tp, th, CC_NDUPACK);
+ /*
+ * For PRR, adjust recover_fs as if this new reduction
+ * initialized this variable.
+ * cwnd will be adjusted by SACK or PRR processing
+ * subsequently, only set it to a safe value here.
+ */
+ tp->snd_cwnd = tcp_maxseg(tp);
+ tp->sackhint.recover_fs = (tp->snd_max - tp->snd_una) -
+ tp->sackhint.recover_fs;
+ }
+}
Index: sys/netinet/tcp_output.c
===================================================================
--- sys/netinet/tcp_output.c
+++ sys/netinet/tcp_output.c
@@ -1264,6 +1264,14 @@
} else {
th->th_seq = htonl(p->rxmit);
p->rxmit += len;
+ /*
+ * Lost Retransmission Detection
+ * trigger resending of a (then
+ * still existing) hole, when
+ * fack acks recoverypoint.
+ */
+ if (V_tcp_do_lrd && SEQ_GEQ(p->rxmit, p->end))
+ p->rxmit = tp->snd_recover;
tp->sackhint.sack_bytes_rexmit += len;
}
if (IN_RECOVERY(tp->t_flags)) {
Index: sys/netinet/tcp_sack.c
===================================================================
--- sys/netinet/tcp_sack.c
+++ sys/netinet/tcp_sack.c
@@ -719,7 +719,8 @@
cur = TAILQ_PREV(cur, sackhole_head, scblink);
continue;
}
- tp->sackhint.sack_bytes_rexmit -= (cur->rxmit - cur->start);
+ tp->sackhint.sack_bytes_rexmit -=
+ (SEQ_MIN(cur->rxmit, cur->end) - cur->start);
KASSERT(tp->sackhint.sack_bytes_rexmit >= 0,
("sackhint bytes rtx >= 0"));
sack_changed = 1;
@@ -750,6 +751,8 @@
delivered_data += (cur->end - sblkp->start);
cur->end = sblkp->start;
cur->rxmit = SEQ_MIN(cur->rxmit, cur->end);
+ if (V_tcp_do_lrd && SEQ_GEQ(cur->rxmit, cur->end))
+ cur->rxmit = tp->snd_recover;
} else {
/*
* ACKs some data in middle of a hole; need
@@ -761,17 +764,20 @@
if (SEQ_GT(cur->rxmit, temp->rxmit)) {
temp->rxmit = cur->rxmit;
tp->sackhint.sack_bytes_rexmit
- += (temp->rxmit
- - temp->start);
+ += (SEQ_MIN(temp->rxmit,
+ temp->end) - temp->start);
}
cur->end = sblkp->start;
cur->rxmit = SEQ_MIN(cur->rxmit,
cur->end);
+ if (V_tcp_do_lrd && SEQ_GEQ(cur->rxmit, cur->end))
+ cur->rxmit = tp->snd_recover;
delivered_data += (sblkp->end - sblkp->start);
}
}
}
- tp->sackhint.sack_bytes_rexmit += (cur->rxmit - cur->start);
+ tp->sackhint.sack_bytes_rexmit +=
+ (SEQ_MIN(cur->rxmit, cur->end) - cur->start);
/*
* Testing sblkp->start against cur->start tells us whether
* we're done with the sack block or the sack hole.
@@ -899,7 +905,7 @@
*sack_bytes_rexmt += (p->rxmit - p->start);
break;
}
- *sack_bytes_rexmt += (p->rxmit - p->start);
+ *sack_bytes_rexmt += (SEQ_MIN(p->rxmit, p->end) - p->start);
}
return (p);
}
Index: sys/netinet/tcp_var.h
===================================================================
--- sys/netinet/tcp_var.h
+++ sys/netinet/tcp_var.h
@@ -855,6 +855,7 @@
VNET_DECLARE(int, tcp_do_autorcvbuf);
VNET_DECLARE(int, tcp_do_autosndbuf);
VNET_DECLARE(int, tcp_do_ecn);
+VNET_DECLARE(int, tcp_do_lrd);
VNET_DECLARE(int, tcp_do_prr);
VNET_DECLARE(int, tcp_do_prr_conservative);
VNET_DECLARE(int, tcp_do_newcwv);
@@ -889,6 +890,7 @@
VNET_DECLARE(struct inpcbhead, tcb);
VNET_DECLARE(struct inpcbinfo, tcbinfo);
+#define V_tcp_do_lrd VNET(tcp_do_lrd)
#define V_tcp_do_prr VNET(tcp_do_prr)
#define V_tcp_do_prr_conservative VNET(tcp_do_prr_conservative)
#define V_tcp_do_newcwv VNET(tcp_do_newcwv)
@@ -1082,6 +1084,7 @@
void tcp_sack_adjust(struct tcpcb *tp);
struct sackhole *tcp_sack_output(struct tcpcb *tp, int *sack_bytes_rexmt);
void tcp_do_prr_ack(struct tcpcb *, struct tcphdr *, struct tcpopt *);
+void tcp_lost_retransmission(struct tcpcb *, struct tcphdr *);
void tcp_sack_partialack(struct tcpcb *, struct tcphdr *);
void tcp_free_sackholes(struct tcpcb *tp);
int tcp_newreno(struct tcpcb *, struct tcphdr *);

File Metadata

Mime Type
text/plain
Expires
Tue, Jun 30, 10:19 AM (5 h, 16 m)
Storage Engine
blob
Storage Format
Raw Data
Storage Handle
34496517
Default Alt Text
D28931.id88001.diff (7 KB)

Event Timeline