d/freebsd-skq
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

update the SACK loss recovery to RFC6675, with the following new features:

Browse Source 
- improved pipe calculation which does not degrade under heavy loss
- engaging in Loss Recovery earlier under adverse conditions
- Rescue Retransmission in case some of the trailing packets of a request got lost

All above changes are toggled with the sysctl "rfc6675_pipe" (disabled by default).

Reviewers:	#transport, tuexen, lstewart, slavash, jtl, hselasky, kib, rgrimes, chengc_netapp.com, thj, #manpages, kbowling, #netapp, rscheff
Reviewed By:	#transport
Subscribers:	imp, melifaro
MFC after:	2 weeks
Sponsored by:	NetApp, Inc.
Differential Revision:	https://reviews.freebsd.org/D18985
This commit is contained in:
Richard Scheffenegger 2021-02-16 12:18:43 +01:00
parent afcb3c4cb4
commit 3c40e1d52c
3 changed files with 73 additions and 6 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

10
share/man/man4/tcp.4
View File

@ -34,7 +34,7 @@
.\" From: @(#)tcp.4 8.1 (Berkeley) 6/5/93
.\" $FreeBSD$
.\"
.Dd February 11, 2021
.Dd February 13, 2021
.Dt TCP 4
.Os
.Sh NAME
@ -560,6 +560,14 @@ high losses leading to RTO, but reduces PRR effectiveness in more common setting
.It Va rfc6675_pipe
Calculate the bytes in flight using the algorithm described in RFC 6675, and
is also an improvement when Proportional Rate Reduction is enabled.
Also enables two other mechanisms from RFC6675.
Rescue Retransmission helps timely loss recovery, when the trailing segments
of a transmission are lost, while no additional data is ready to be sent.
In case a partial ACK without a SACK block is received during SACK loss
recovery, the trailing segment is immediately resent, rather than waiting
for a Retransmission timeout.
SACK loss recovery is also engaged, once two segments plus one byte are
SACKed - even if no traditional duplicate ACKs were seen.
.It Va rfc3042
Enable the Limited Transmit algorithm as described in RFC 3042.
It helps avoid timeouts on lossy links and also when the congestion window

34
sys/netinet/tcp_input.c
View File

@ -1501,6 +1501,7 @@ tcp_do_segment(struct mbuf *m, struct tcphdr *th, struct socket *so,
struct mbuf *mfree;
struct tcpopt to;
int tfo_syn;
u_int maxseg;
#ifdef TCPDEBUG
/*
@ -2502,8 +2503,6 @@ tcp_do_segment(struct mbuf *m, struct tcphdr *th, struct socket *so,
#endif
if (SEQ_LEQ(th->th_ack, tp->snd_una)) {
u_int maxseg;
maxseg = tcp_maxseg(tp);
if (tlen == 0 &&
(tiwin == tp->snd_wnd ||
@ -2644,7 +2643,21 @@ tcp_do_segment(struct mbuf *m, struct tcphdr *th, struct socket *so,
tp->snd_cwnd += maxseg;
(void) tp->t_fb->tfb_tcp_output(tp);
goto drop;
} else if (tp->t_dupacks == tcprexmtthresh) {
} else if (tp->t_dupacks == tcprexmtthresh ||
(tp->t_flags & TF_SACK_PERMIT &&
V_tcp_do_rfc6675_pipe &&
tp->sackhint.sacked_bytes >
(tcprexmtthresh - 1) * maxseg)) {
enter_recovery:
/*
* Above is the RFC6675 trigger condition of
* more than (dupthresh-1)*maxseg sacked data.
* If the count of holes in the
* scoreboard is >= dupthresh, we could
* also enter loss recovery, but don't
* have that value readily available.
*/
tp->t_dupacks = tcprexmtthresh;
tcp_seq onxt = tp->snd_nxt;
/*
@ -2689,6 +2702,8 @@ tcp_do_segment(struct mbuf *m, struct tcphdr *th, struct socket *so,
tp->snd_recover = tp->snd_nxt;
tp->snd_cwnd = maxseg;
(void) tp->t_fb->tfb_tcp_output(tp);
if (SEQ_GT(th->th_ack, tp->snd_una))
goto resume_partialack;
goto drop;
}
tp->snd_nxt = th->th_ack;
@ -2775,10 +2790,19 @@ tcp_do_segment(struct mbuf *m, struct tcphdr *th, struct socket *so,
*/
if ((tp->t_flags & TF_SACK_PERMIT) &&
(to.to_flags & TOF_SACK) &&
sack_changed)
sack_changed) {
tp->t_dupacks++;
/* limit overhead by setting maxseg last */
if (!IN_FASTRECOVERY(tp->t_flags) &&
(tp->sackhint.sacked_bytes >
((tcprexmtthresh - 1) *
(maxseg = tcp_maxseg(tp))))) {
goto enter_recovery;
}
}
}
resume_partialack:
KASSERT(SEQ_GT(th->th_ack, tp->snd_una),
("%s: th_ack <= snd_una", __func__));
@ -2789,7 +2813,7 @@ tcp_do_segment(struct mbuf *m, struct tcphdr *th, struct socket *so,
if (IN_FASTRECOVERY(tp->t_flags)) {
if (SEQ_LT(th->th_ack, tp->snd_recover)) {
if (tp->t_flags & TF_SACK_PERMIT)
if (V_tcp_do_prr)
if (V_tcp_do_prr && to.to_flags & TOF_SACK)
tcp_prr_partialack(tp, th);
else
tcp_sack_partialack(tp, th);

35
sys/netinet/tcp_sack.c
View File

@ -750,6 +750,16 @@ tcp_sack_doack(struct tcpcb *tp, struct tcpopt *to, tcp_seq th_ack)
else
sblkp--;
}
if (!(to->to_flags & TOF_SACK))
/*
* If this ACK did not contain any
* SACK blocks, any only moved the
* left edge right, it is a pure
* cumulative ACK. Do not count
* DupAck for this. Also required
* for RFC6675 rescue retransmission.
*/
sack_changed = 0;
tp->sackhint.delivered_data = delivered_data;
tp->sackhint.sacked_bytes += delivered_data - left_edge_delta;
KASSERT((delivered_data >= 0), ("delivered_data < 0"));
@ -800,6 +810,31 @@ tcp_sack_partialack(struct tcpcb *tp, struct tcphdr *th)
if (tp->snd_cwnd > tp->snd_ssthresh)
tp->snd_cwnd = tp->snd_ssthresh;
tp->t_flags |= TF_ACKNOW;
/*
* RFC6675 rescue retransmission
* Add a hole between th_ack (snd_una is not yet set) and snd_max,
* if this was a pure cumulative ACK and no data was send beyond
* recovery point. Since the data in the socket has not been freed
* at this point, we check if the scoreboard is empty, and the ACK
* delivered some new data, indicating a full ACK. Also, if the
* recovery point is still at snd_max, we are probably application
* limited. However, this inference might not always be true. The
* rescue retransmission may rarely be slightly premature
* compared to RFC6675.
* The corresponding ACK+SACK will cause any further outstanding
* segments to be retransmitted. This addresses a corner case, when
* the trailing packets of a window are lost and no further data
* is available for sending.
*/
if ((V_tcp_do_rfc6675_pipe) &&
SEQ_LT(th->th_ack, tp->snd_recover) &&
(tp->snd_recover == tp->snd_max) &&
TAILQ_EMPTY(&tp->snd_holes) &&
(tp->sackhint.delivered_data > 0)) {
struct sackhole *hole;
int maxseg = tcp_maxseg(tp);
hole = tcp_sackhole_insert(tp, SEQ_MAX(th->th_ack, tp->snd_max - maxseg), tp->snd_max, NULL);
}
(void) tp->t_fb->tfb_tcp_output(tp);
}