Skip to contentSkip to conversation

netfilter: conntrack: always store window size un-scaled

PATCH Applied Florian Westphal 5 participants 5 messages Jul 11, 2019 – Aug 13, 2019

Subsystems (2)
lore
Share & tools
lore

Florian Westphal

author maintainer
#0 lore raw

PATCH 1/1 netfilter: conntrack: always store window size un-scaled

Jakub Jankowski reported following oddity:

After 3 way handshake completes, timeout of new connection is set to max_retrans (300s) instead of established (5 days).

shortened excerpt from pcap provided:
25.070622 IP (flags [DF], proto TCP (6), length 52)
10.8.5.4.1025 > 10.8.1.2.80: Flags [S], seq 11, win 64240, [wscale 8] 26.070462 IP (flags [DF], proto TCP (6), length 48)
10.8.1.2.80 > 10.8.5.4.1025: Flags [S.], seq 82, ack 12, win 65535, [wscale 3] 27.070449 IP (flags [DF], proto TCP (6), length 40)
10.8.5.4.1025 > 10.8.1.2.80: Flags [.], ack 83, win 512, length 0

Turns out the last_win is of u16 type, but we store the scaled value:
512 << 8 (== 0x20000) becomes 0 window.

The Fixes tag is not correct, as the bug has existed forever, but without that change all that this causes might cause is to mistake a window update (to-nonzero-from-zero) for a retransmit.

Fixes fbcd253d2448b8 ("netfilter: conntrack: lower timeout to RETRANS seconds if window is 0")
---
 net/netfilter/nf_conntrack_proto_tcp.c | 8 +++++---
 1 file changed, 5 insertions(+), 3 deletions(-)
net/netfilter/nf_conntrack_proto_tcp.c +5 -3
index 7ba01d8ee165..9fe1d5e46249 100644
@@ -475,6 +475,7 @@ static bool tcp_in_window(const struct nf_conn *ct,
 	struct ip_ct_tcp_state *receiver = &state->seen[!dir];
 	const struct nf_conntrack_tuple *tuple = &ct->tuplehash[dir].tuple;
 	__u32 seq, ack, sack, end, win, swin;
+	u16 win_raw;
 	s32 receiver_offset;
 	bool res, in_recv_win;
 
@@ -483,7 +484,8 @@ static bool tcp_in_window(const struct nf_conn *ct,
 	 */
 	seq = ntohl(tcph->seq);
 	ack = sack = ntohl(tcph->ack_seq);
-	win = ntohs(tcph->window);
+	win_raw = ntohs(tcph->window);
+	win = win_raw;
 	end = segment_seq_plus_len(seq, skb->len, dataoff, tcph);
 
 	if (receiver->flags & IP_CT_TCP_FLAG_SACK_PERM)
@@ -658,14 +660,14 @@ static bool tcp_in_window(const struct nf_conn *ct,
 			    && state->last_seq == seq
 			    && state->last_ack == ack
 			    && state->last_end == end
-			    && state->last_win == win)
+			    && state->last_win == win_raw)
 				state->retrans++;
 			else {
 				state->last_dir = dir;
 				state->last_seq = seq;
 				state->last_ack = ack;
 				state->last_end = end;
-				state->last_win = win;
+				state->last_win = win_raw;
 				state->retrans = 0;
 			}
 		}
Signature · 1 line
2.21.0
Acked-by on v1 1/1
#1 lore raw
Quoted 16 lines · Florian Westphal

Jakub Jankowski reported following oddity:

After 3 way handshake completes, timeout of new connection is set to max_retrans (300s) instead of established (5 days).

shortened excerpt from pcap provided:
25.070622 IP (flags [DF], proto TCP (6), length 52)
10.8.5.4.1025 > 10.8.1.2.80: Flags [S], seq 11, win 64240, [wscale 8] 26.070462 IP (flags [DF], proto TCP (6), length 48)
10.8.1.2.80 > 10.8.5.4.1025: Flags [S.], seq 82, ack 12, win 65535, [wscale 3] 27.070449 IP (flags [DF], proto TCP (6), length 40)
10.8.5.4.1025 > 10.8.1.2.80: Flags [.], ack 83, win 512, length 0

Turns out the last_win is of u16 type, but we store the scaled value:
512 << 8 (== 0x20000) becomes 0 window.

Quoted 8 lines

The Fixes tag is not correct, as the bug has existed forever, but without that change all that this causes might cause is to mistake a window update (to-nonzero-from-zero) for a retransmit.

Fixes fbcd253d2448b8 ("netfilter: conntrack: lower timeout to RETRANS seconds if window is 0")

It was a nice report and catch!

Best regards,
Jozsef

Quoted 47 lines
---
 net/netfilter/nf_conntrack_proto_tcp.c | 8 +++++---
 1 file changed, 5 insertions(+), 3 deletions(-)
net/netfilter/nf_conntrack_proto_tcp.c +5 -3
index 7ba01d8ee165..9fe1d5e46249 100644
@@ -475,6 +475,7 @@ static bool tcp_in_window(const struct nf_conn *ct,
 	struct ip_ct_tcp_state *receiver = &state->seen[!dir];
 	const struct nf_conntrack_tuple *tuple = &ct->tuplehash[dir].tuple;
 	__u32 seq, ack, sack, end, win, swin;
+	u16 win_raw;
 	s32 receiver_offset;
 	bool res, in_recv_win;
 
@@ -483,7 +484,8 @@ static bool tcp_in_window(const struct nf_conn *ct,
 	 */
 	seq = ntohl(tcph->seq);
 	ack = sack = ntohl(tcph->ack_seq);
-	win = ntohs(tcph->window);
+	win_raw = ntohs(tcph->window);
+	win = win_raw;
 	end = segment_seq_plus_len(seq, skb->len, dataoff, tcph);
 
 	if (receiver->flags & IP_CT_TCP_FLAG_SACK_PERM)
@@ -658,14 +660,14 @@ static bool tcp_in_window(const struct nf_conn *ct,
 			    && state->last_seq == seq
 			    && state->last_ack == ack
 			    && state->last_end == end
-			    && state->last_win == win)
+			    && state->last_win == win_raw)
 				state->retrans++;
 			else {
 				state->last_dir = dir;
 				state->last_seq = seq;
 				state->last_ack = ack;
 				state->last_end = end;
-				state->last_win = win;
+				state->last_win = win_raw;
 				state->retrans = 0;
 			}
 		}
Signature · 1 line
2.21.0

-
E-mail : [email protected], [email protected] PGP key : http://www.kfki.hu/~kadlec/pgp_public_key.txt
Address : Wigner Research Centre for Physics, Hungarian Academy of Sciences

          H-1525 Budapest 114, POB. 49, Hungary
maintainer Applied
#2 lore raw
Quoted 22 lines · Jozsef Kadlecsik
Quoted 20 lines · Florian Westphal

Jakub Jankowski reported following oddity:

After 3 way handshake completes, timeout of new connection is set to max_retrans (300s) instead of established (5 days).

shortened excerpt from pcap provided:
25.070622 IP (flags [DF], proto TCP (6), length 52)
10.8.5.4.1025 > 10.8.1.2.80: Flags [S], seq 11, win 64240, [wscale 8] 26.070462 IP (flags [DF], proto TCP (6), length 48)
10.8.1.2.80 > 10.8.5.4.1025: Flags [S.], seq 82, ack 12, win 65535, [wscale 3] 27.070449 IP (flags [DF], proto TCP (6), length 40)
10.8.5.4.1025 > 10.8.1.2.80: Flags [.], ack 83, win 512, length 0

Turns out the last_win is of u16 type, but we store the scaled value:
512 << 8 (== 0x20000) becomes 0 window.

The Fixes tag is not correct, as the bug has existed forever, but without that change all that this causes might cause is to mistake a window update (to-nonzero-from-zero) for a retransmit.

Quoted 6 lines
Quoted 4 lines
Fixes fbcd253d2448b8 ("netfilter: conntrack: lower timeout to RETRANS seconds if window is 0")

Applied, thanks for reviewing Jozsef.

#3 lore raw
In reply to Florian Westphal

this seemed to be fixed in 5.1.19 but not announced in
https://cdn.kernel.org/pub/linux/kernel/v5.x/ChangeLog-5.1.19 and after update to 5.1.20 last night the freezes of a ssh-tunneled vnc session are starting again without exclude "lo" while the session was stable 3 days

see topic "Connection timeouts due to INVALID state rule"

Am 12.07.19 um 00:29 schrieb Florian Westphal:

Quoted 68 lines

Jakub Jankowski reported following oddity:

After 3 way handshake completes, timeout of new connection is set to max_retrans (300s) instead of established (5 days).

shortened excerpt from pcap provided:
25.070622 IP (flags [DF], proto TCP (6), length 52)
10.8.5.4.1025 > 10.8.1.2.80: Flags [S], seq 11, win 64240, [wscale 8] 26.070462 IP (flags [DF], proto TCP (6), length 48)
10.8.1.2.80 > 10.8.5.4.1025: Flags [S.], seq 82, ack 12, win 65535, [wscale 3] 27.070449 IP (flags [DF], proto TCP (6), length 40)
10.8.5.4.1025 > 10.8.1.2.80: Flags [.], ack 83, win 512, length 0

Turns out the last_win is of u16 type, but we store the scaled value:
512 << 8 (== 0x20000) becomes 0 window.

The Fixes tag is not correct, as the bug has existed forever, but without that change all that this causes might cause is to mistake a window update (to-nonzero-from-zero) for a retransmit.

Fixes fbcd253d2448b8 ("netfilter: conntrack: lower timeout to RETRANS seconds if window is 0")
---
 net/netfilter/nf_conntrack_proto_tcp.c | 8 +++++---
 1 file changed, 5 insertions(+), 3 deletions(-)
net/netfilter/nf_conntrack_proto_tcp.c +5 -3
index 7ba01d8ee165..9fe1d5e46249 100644
@@ -475,6 +475,7 @@ static bool tcp_in_window(const struct nf_conn *ct,
 	struct ip_ct_tcp_state *receiver = &state->seen[!dir];
 	const struct nf_conntrack_tuple *tuple = &ct->tuplehash[dir].tuple;
 	__u32 seq, ack, sack, end, win, swin;
+	u16 win_raw;
 	s32 receiver_offset;
 	bool res, in_recv_win;
 
@@ -483,7 +484,8 @@ static bool tcp_in_window(const struct nf_conn *ct,
 	 */
 	seq = ntohl(tcph->seq);
 	ack = sack = ntohl(tcph->ack_seq);
-	win = ntohs(tcph->window);
+	win_raw = ntohs(tcph->window);
+	win = win_raw;
 	end = segment_seq_plus_len(seq, skb->len, dataoff, tcph);
 
 	if (receiver->flags & IP_CT_TCP_FLAG_SACK_PERM)
@@ -658,14 +660,14 @@ static bool tcp_in_window(const struct nf_conn *ct,
 			    && state->last_seq == seq
 			    && state->last_ack == ack
 			    && state->last_end == end
-			    && state->last_win == win)
+			    && state->last_win == win_raw)
 				state->retrans++;
 			else {
 				state->last_dir = dir;
 				state->last_seq = seq;
 				state->last_ack = ack;
 				state->last_end = end;
-				state->last_win = win;
+				state->last_win = win_raw;
 				state->retrans = 0;
 			}
 		}
#4 lore raw
In reply to Florian Westphal

Hi Florian,

You wrote on Fri, Jul 12, 2019 at 12:29:05AM +0200:

Quoted 16 lines

Jakub Jankowski reported following oddity:

After 3 way handshake completes, timeout of new connection is set to max_retrans (300s) instead of established (5 days).

shortened excerpt from pcap provided:
25.070622 IP (flags [DF], proto TCP (6), length 52)
10.8.5.4.1025 > 10.8.1.2.80: Flags [S], seq 11, win 64240, [wscale 8] 26.070462 IP (flags [DF], proto TCP (6), length 48)
10.8.1.2.80 > 10.8.5.4.1025: Flags [S.], seq 82, ack 12, win 65535, [wscale 3] 27.070449 IP (flags [DF], proto TCP (6), length 40)
10.8.5.4.1025 > 10.8.1.2.80: Flags [.], ack 83, win 512, length 0

Turns out the last_win is of u16 type, but we store the scaled value:
512 << 8 (== 0x20000) becomes 0 window.

Quoted 8 lines

The Fixes tag is not correct, as the bug has existed forever, but without that change all that this causes might cause is to mistake a window update (to-nonzero-from-zero) for a retransmit.

Fixes fbcd253d2448b8 ("netfilter: conntrack: lower timeout to RETRANS seconds if window is 0")

it seems the patch fixed a kernel bugzilla entry, too:
https://bugzilla.kernel.org/show_bug.cgi?id=202287

I had a feeling it could be related since the reporter bisected it down to changes in the TCP window scaling defaults.

Cheers,
Thomas