Skip to contentSkip to conversation

tcp_cubic: fix spurious HYSTART_DELAY on RTT decrease

PATCH In review Neal Cardwell 2 participants 4 messages Jun 24, 2020 – Jun 25, 2020

Subsystems (6)
lore
Share & tools
lore

Neal Cardwell

author maintainer
#0 lore raw

cover tcp_cubic: fix spurious HYSTART_DELAY on RTT decrease

This series fixes a long-standing bug in the TCP CUBIC HYSTART_DELAY mechanim recently reported by Mirja Kuehlewind. The code can cause a spurious exit of slow start in some particular
cases: upon an RTT decrease that happens on the 9th or later ACK in a round trip. This series fixes the original Hystart code and also the recent BPF implementation.

Neal Cardwell (2):
tcp_cubic: fix spurious HYSTART_DELAY exit upon drop in min RTT
bpf: tcp: bpf_cubic: fix spurious HYSTART_DELAY exit upon drop in min
RTT

 net/ipv4/tcp_cubic.c                          | 5 ++---
 tools/testing/selftests/bpf/progs/bpf_cubic.c | 5 ++---
 2 files changed, 4 insertions(+), 6 deletions(-)
Signature · 1 line
2.27.0.111.gc72c7da667-goog
author maintainer
#1 lore raw

PATCH 1/2 tcp_cubic: fix spurious HYSTART_DELAY exit upon drop in min RTT

Mirja Kuehlewind reported a bug in Linux TCP CUBIC Hystart, where Hystart HYSTART_DELAY mechanism can exit Slow Start spuriously on an ACK when the minimum rtt of a connection goes down. From inspection it is clear from the existing code that this could happen in an example like the following:

o The first 8 RTT samples in a round trip are 150ms, resulting in a curr_rtt of 150ms and a delay_min of 150ms.

o The 9th RTT sample is 100ms. The curr_rtt does not change after the first 8 samples, so curr_rtt remains 150ms. But delay_min can be lowered at any time, so delay_min falls to 100ms. The code executes the HYSTART_DELAY comparison between curr_rtt of 150ms and delay_min of 100ms, and the curr_rtt is declared far enough above delay_min to force a (spurious) exit of Slow start.

The fix here is simple: allow every RTT sample in a round trip to lower the curr_rtt.

---
 net/ipv4/tcp_cubic.c | 5 ++---
 1 file changed, 2 insertions(+), 3 deletions(-)
net/ipv4/tcp_cubic.c +2 -3
index 8f8eefd3a3ce..c7bf5b26bf0c 100644
@@ -432,10 +432,9 @@ static void hystart_update(struct sock *sk, u32 delay)
 
 	if (hystart_detect & HYSTART_DELAY) {
 		/* obtain the minimum delay of more than sampling packets */
+		if (ca->curr_rtt > delay)
+			ca->curr_rtt = delay;
 		if (ca->sample_cnt < HYSTART_MIN_SAMPLES) {
-			if (ca->curr_rtt > delay)
-				ca->curr_rtt = delay;
-
 			ca->sample_cnt++;
 		} else {
 			if (ca->curr_rtt > ca->delay_min +
Signature · 1 line
2.27.0.111.gc72c7da667-goog
author maintainer
#2 lore raw

PATCH 2/2 bpf: tcp: bpf_cubic: fix spurious HYSTART_DELAY exit upon drop in min RTT

In reply to Neal Cardwell

Apply the fix from:
"tcp_cubic: fix spurious HYSTART_DELAY exit upon drop in min RTT"
to the BPF implementation of TCP CUBIC congestion control.

Repeating the commit description here for completeness:

Mirja Kuehlewind reported a bug in Linux TCP CUBIC Hystart, where Hystart HYSTART_DELAY mechanism can exit Slow Start spuriously on an ACK when the minimum rtt of a connection goes down. From inspection it is clear from the existing code that this could happen in an example like the following:

o The first 8 RTT samples in a round trip are 150ms, resulting in a curr_rtt of 150ms and a delay_min of 150ms.

o The 9th RTT sample is 100ms. The curr_rtt does not change after the first 8 samples, so curr_rtt remains 150ms. But delay_min can be lowered at any time, so delay_min falls to 100ms. The code executes the HYSTART_DELAY comparison between curr_rtt of 150ms and delay_min of 100ms, and the curr_rtt is declared far enough above delay_min to force a (spurious) exit of Slow start.

The fix here is simple: allow every RTT sample in a round trip to lower the curr_rtt.

Fixes 6de4a9c430b5 ("bpf: tcp: Add bpf_cubic example")
---
 tools/testing/selftests/bpf/progs/bpf_cubic.c | 5 ++---
 1 file changed, 2 insertions(+), 3 deletions(-)
tools/testing/selftests/bpf/progs/bpf_cubic.c +2 -3
index 7897c8f4d363..ef574087f1e1 100644
@@ -480,10 +480,9 @@ static __always_inline void hystart_update(struct sock *sk, __u32 delay)
 
 	if (hystart_detect & HYSTART_DELAY) {
 		/* obtain the minimum delay of more than sampling packets */
+		if (ca->curr_rtt > delay)
+			ca->curr_rtt = delay;
 		if (ca->sample_cnt < HYSTART_MIN_SAMPLES) {
-			if (ca->curr_rtt > delay)
-				ca->curr_rtt = delay;
-
 			ca->sample_cnt++;
 		} else {
 			if (ca->curr_rtt > ca->delay_min +
Signature · 1 line
2.27.0.111.gc72c7da667-goog
maintainer
#3 lore raw
In reply to Neal Cardwell

From: Neal Cardwell <[email protected]>
Date: Wed, 24 Jun 2020 12:42:01 -0400

Quoted 6 lines

This series fixes a long-standing bug in the TCP CUBIC HYSTART_DELAY mechanim recently reported by Mirja Kuehlewind. The code can cause a spurious exit of slow start in some particular
cases: upon an RTT decrease that happens on the 9th or later ACK in a round trip. This series fixes the original Hystart code and also the recent BPF implementation.

Series applied and queued up for -stable, thank you.