Hi, I'm completely unable to get ECN to work for IPv6. /proc/sys/net/ipv4/tcp_ecn is set to 1 on both sides, and ECN works just fine for IPv4 TCP connections, but when I connect over IPv6 tclass just stays at 0x0, and as far as I'd understood, there should be two bits here set to 10 (like in the diffserv field of IPv4), right? I do get ECN-echo bits in the TCP header, though (for the initial SYN and SYN/ACK packets).
I looked at the code, and although I don't understand all of it, it looks like e9df2e8fd8fbc95c57dbd1d33dada66c4627b44c (which also brought ECN for IPv6 SCTP) might be the culprit. It seems like TCP_ECN_send() calls INET_ECN_xmit(), which only sets the ECN bit in the IPv4 ToS field (inet_sk(sk)->tos), but after the patch, what's checked is inet6_sk(sk)->tclass, which is a completely different field. Is the analysis correct? Should the tclass be set as well?
OK, as a quick hack, I did this: --- a/linux-2.6.38.5/include/net/inet_ecn.h 2011-04-14 22:03:56.000000000 +0200 +++ b/linux-2.6.38.5/include/net/inet_ecn.h 2011-05-04 00:36:52.803377902 +0200 @@ -38,7 +38,7 @@ return outer; } -#define INET_ECN_xmit(sk) do { inet_sk(sk)->tos |= INET_ECN_ECT_0; } while (0) +#define INET_ECN_xmit(sk) do { inet_sk(sk)->tos |= INET_ECN_ECT_0; if (inet6_sk(sk) != NULL) inet6_sk(sk)->tclass |= INET_ECN_ECT_0; } while (0) #define INET_ECN_dontxmit(sk) \ do { inet_sk(sk)->tos &= ~INET_ECN_MASK; } while (0) and now my packets are properly marked with tclass 0x02 (ie., signalling ECN-capable transport, no congestion experienced yet). I guess this isn't the right way of doing it, but at least it confirms that the lack of setting tclass is part of the problem.
(switched to email. Please respond via emailed reply-to-all, not via the bugzilla web interface). On Tue, 3 May 2011 20:05:00 GMT bugzilla-daemon@bugzilla.kernel.org wrote: > https://bugzilla.kernel.org/show_bug.cgi?id=34322 > > Summary: No ECN marking in IPv6 > Product: Networking > Version: 2.5 > Kernel Version: 2.6.38.4 > Platform: All > OS/Version: Linux > Tree: Mainline > Status: NEW > Severity: normal > Priority: P1 > Component: IPV6 > AssignedTo: yoshfuji@linux-ipv6.org > ReportedBy: sgunderson@bigfoot.com > Regression: No > > > Hi, > > I'm completely unable to get ECN to work for IPv6. /proc/sys/net/ipv4/tcp_ecn > is set to 1 on both sides, and ECN works just fine for IPv4 TCP connections, > but when I connect over IPv6 tclass just stays at 0x0, and as far as I'd > understood, there should be two bits here set to 10 (like in the diffserv > field > of IPv4), right? > > I do get ECN-echo bits in the TCP header, though (for the initial SYN and > SYN/ACK packets). > and > I looked at the code, and although I don't understand all of it, it > looks like e9df2e8fd8fbc95c57dbd1d33dada66c4627b44c (which also brought > ECN for IPv6 SCTP) might be the culprit. It seems like TCP_ECN_send() > calls INET_ECN_xmit(), which only sets the ECN bit in the IPv4 ToS > field (inet_sk(sk)->tos), but after the patch, what's checked is > inet6_sk(sk)->tclass, which is a completely different field. > > Is the analysis correct? Should the tclass be set as well? and > OK, as a quick hack, I did this: > > --- a/linux-2.6.38.5/include/net/inet_ecn.h 2011-04-14 > 22:03:56.000000000+0200 > +++ b/linux-2.6.38.5/include/net/inet_ecn.h 2011-05-04 > 00:36:52.803377902+0200 > @@ -38,7 +38,7 @@ > return outer; > } > > -#define INET_ECN_xmit(sk) do { inet_sk(sk)->tos |= INET_ECN_ECT_0; } > while (0) > +#define INET_ECN_xmit(sk) do { inet_sk(sk)->tos |= INET_ECN_ECT_0; if > (inet6_sk(sk) != NULL) inet6_sk(sk)->tclass |= INET_ECN_ECT_0; } while (0) > #define INET_ECN_dontxmit(sk) \ > do { inet_sk(sk)->tos &= ~INET_ECN_MASK; } while (0) > > and now my packets are properly marked with tclass 0x02 (ie., > signalling ECN-capable transport, no congestion experienced yet). > > I guess this isn't the right way of doing it, but at least it confirms > that the lack of setting tclass is part of the problem.
Le jeudi 05 mai 2011 à 14:41 -0700, Andrew Morton a écrit : > (switched to email. Please respond via emailed reply-to-all, not via the > bugzilla web interface). > > On Tue, 3 May 2011 20:05:00 GMT > bugzilla-daemon@bugzilla.kernel.org wrote: > > > https://bugzilla.kernel.org/show_bug.cgi?id=34322 > > > > Summary: No ECN marking in IPv6 > > Product: Networking > > Version: 2.5 > > Kernel Version: 2.6.38.4 > > Platform: All > > OS/Version: Linux > > Tree: Mainline > > Status: NEW > > Severity: normal > > Priority: P1 > > Component: IPV6 > > AssignedTo: yoshfuji@linux-ipv6.org > > ReportedBy: sgunderson@bigfoot.com > > Regression: No > > > > > > Hi, > > > > I'm completely unable to get ECN to work for IPv6. > /proc/sys/net/ipv4/tcp_ecn > > is set to 1 on both sides, and ECN works just fine for IPv4 TCP > connections, > > but when I connect over IPv6 tclass just stays at 0x0, and as far as I'd > > understood, there should be two bits here set to 10 (like in the diffserv > field > > of IPv4), right? > > > > I do get ECN-echo bits in the TCP header, though (for the initial SYN and > > SYN/ACK packets). > > > > and > > > I looked at the code, and although I don't understand all of it, it > > looks like e9df2e8fd8fbc95c57dbd1d33dada66c4627b44c (which also brought > > ECN for IPv6 SCTP) might be the culprit. It seems like TCP_ECN_send() > > calls INET_ECN_xmit(), which only sets the ECN bit in the IPv4 ToS > > field (inet_sk(sk)->tos), but after the patch, what's checked is > > inet6_sk(sk)->tclass, which is a completely different field. > > > > Is the analysis correct? Should the tclass be set as well? > > and > > > OK, as a quick hack, I did this: > > > > --- a/linux-2.6.38.5/include/net/inet_ecn.h 2011-04-14 > 22:03:56.000000000+0200 > > +++ b/linux-2.6.38.5/include/net/inet_ecn.h 2011-05-04 > 00:36:52.803377902+0200 > > @@ -38,7 +38,7 @@ > > return outer; > > } > > > > -#define INET_ECN_xmit(sk) do { inet_sk(sk)->tos |= INET_ECN_ECT_0; } > while (0) > > +#define INET_ECN_xmit(sk) do { inet_sk(sk)->tos |= INET_ECN_ECT_0; if > (inet6_sk(sk) != NULL) inet6_sk(sk)->tclass |= INET_ECN_ECT_0; } while (0) > > #define INET_ECN_dontxmit(sk) \ > > do { inet_sk(sk)->tos &= ~INET_ECN_MASK; } while (0) > > > > and now my packets are properly marked with tclass 0x02 (ie., > > signalling ECN-capable transport, no congestion experienced yet). > > > > I guess this isn't the right way of doing it, but at least it confirms > > that the lack of setting tclass is part of the problem. > Cc YOSHIFUJI Hideaki <yoshfuji@linux-ipv6.org> Hello Steinar Analysis seems fine, but you also need to change INET_ECN_dontxmit() for retransmitted packets. Any chance you can refine your patch ? Thanks !
On Fri, May 06, 2011 at 05:04:52PM +0200, Eric Dumazet wrote: > Analysis seems fine, but you also need to change INET_ECN_dontxmit() for > retransmitted packets. > > Any chance you can refine your patch ? Sure, but is really checking against NULL the right way of checking for IPv6 sockets? I'd imagined I should have checked address family or something instead... /* Steinar */
From: Steinar H. Gunderson <sgunderson@bigfoot.com> Le vendredi 06 mai 2011 à 19:12 +0200, Steinar H. Gunderson a écrit : > Sure, but is really checking against NULL the right way of checking for IPv6 > sockets? I'd imagined I should have checked address family or something > instead... > It should be fine. I cooked for you the official patch and made sure it worked with a RED ECN setup, and one ipv6 tcp xmit. # tc -s -d qdisc show dev eth1 ... qdisc red 11: parent 1:11 limit 120Kb min 8Kb max 80Kb ecn ewma 2 Plog 21 Scell_log 11 Sent 114694826 bytes 76446 pkt (dropped 15, overlimits 485 requeues 0) rate 12126Kbit 1011pps backlog 0b 0p requeues 0 marked 470 early 15 pdrop 0 other 0 Thanks again ! [PATCH] ipv6: restore correct ECN handling on TCP xmit Since commit e9df2e8fd8fbc9 (Use appropriate sock tclass setting for routing lookup) we lost ability to properly add ECN codemarks to ipv6 TCP frames. It seems like TCP_ECN_send() calls INET_ECN_xmit(), which only sets the ECN bit in the IPv4 ToS field (inet_sk(sk)->tos), but after the patch, what's checked is inet6_sk(sk)->tclass, which is a completely different field. Close bug https://bugzilla.kernel.org/show_bug.cgi?id=34322 [Eric Dumazet] : added the INET_ECN_dontxmit() fix and replace macros by inline functions for clarity. Signed-off-by: Steinar H. Gunderson <sgunderson@bigfoot.com> Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com> Cc: YOSHIFUJI Hideaki <yoshfuji@linux-ipv6.org> Cc: Andrew Morton <akpm@linux-foundation.org> --- include/net/inet_ecn.h | 16 +++++++++++++--- 1 files changed, 13 insertions(+), 3 deletions(-) diff --git a/include/net/inet_ecn.h b/include/net/inet_ecn.h index 88bdd01..2fa8d13 100644 --- a/include/net/inet_ecn.h +++ b/include/net/inet_ecn.h @@ -38,9 +38,19 @@ static inline __u8 INET_ECN_encapsulate(__u8 outer, __u8 inner) return outer; } -#define INET_ECN_xmit(sk) do { inet_sk(sk)->tos |= INET_ECN_ECT_0; } while (0) -#define INET_ECN_dontxmit(sk) \ - do { inet_sk(sk)->tos &= ~INET_ECN_MASK; } while (0) +static inline void INET_ECN_xmit(struct sock *sk) +{ + inet_sk(sk)->tos |= INET_ECN_ECT_0; + if (inet6_sk(sk) != NULL) + inet6_sk(sk)->tclass |= INET_ECN_ECT_0; +} + +static inline void INET_ECN_dontxmit(struct sock *sk) +{ + inet_sk(sk)->tos &= ~INET_ECN_MASK; + if (inet6_sk(sk) != NULL) + inet6_sk(sk)->tclass &= ~INET_ECN_MASK; +} #define IP6_ECN_flow_init(label) do { \ (label) &= ~htonl(INET_ECN_MASK << 20); \
On Sat, May 07, 2011 at 11:44:46AM +0200, Eric Dumazet wrote: > I cooked for you the official patch and made sure it worked with a RED > ECN setup, and one ipv6 tcp xmit. Great, thanks :-) This looks good to me. /* Steinar */
From: "Steinar H. Gunderson" <sgunderson@bigfoot.com> Date: Sat, 7 May 2011 11:59:10 +0200 > On Sat, May 07, 2011 at 11:44:46AM +0200, Eric Dumazet wrote: >> I cooked for you the official patch and made sure it worked with a RED >> ECN setup, and one ipv6 tcp xmit. > > Great, thanks :-) This looks good to me. Applied, thanks everyone.
On Thu, May 12, 2011 at 03:52:21PM -0700, David Miller wrote: >>> I cooked for you the official patch and made sure it worked with a RED >>> ECN setup, and one ipv6 tcp xmit. >> Great, thanks :-) This looks good to me. > Applied, thanks everyone. Do you think it would be possible to get it backported to stable kernels? (I've been told getting it into the 2.6.32.x series is the easiest way to get it into Debian stable, and I guess it should be a candidate.) /* Steinar */
From: "Steinar H. Gunderson" <sgunderson@bigfoot.com> Date: Fri, 13 May 2011 00:54:58 +0200 > On Thu, May 12, 2011 at 03:52:21PM -0700, David Miller wrote: >>>> I cooked for you the official patch and made sure it worked with a RED >>>> ECN setup, and one ipv6 tcp xmit. >>> Great, thanks :-) This looks good to me. >> Applied, thanks everyone. > > Do you think it would be possible to get it backported to stable kernels? > (I've been told getting it into the 2.6.32.x series is the easiest way to get > it into Debian stable, and I guess it should be a candidate.) I only submit patches for what stable@kernel.org is actively maintaining which is 2.6.38 right now.
On Fri, 2011-05-13 at 00:15 +0100, Steinar H. Gunderson wrote: > On Thu, May 12, 2011 at 03:52:21PM -0700, David Miller wrote: > >>> I cooked for you the official patch and made sure it worked with a RED > >>> ECN setup, and one ipv6 tcp xmit. > >> Great, thanks :-) This looks good to me. > > Applied, thanks everyone. > > Do you think it would be possible to get it backported to stable kernels? > (I've been told getting it into the 2.6.32.x series is the easiest way to get > it into Debian stable, and I guess it should be a candidate.) If this doesn't apply directly to 2.6.32 then you'll need to prepare the backport and send it to stable@kernel.org with a reference to the commit in mainline. Ben. -- Ben Hutchings Once a job is fouled up, anything done to improve it makes it worse.
On Fri, May 13, 2011 at 04:58:25AM +0000, bugzilla-daemon@bugzilla.kernel.org wrote: > If this doesn't apply directly to 2.6.32 then you'll need to prepare the > backport and send it to stable@kernel.org with a reference to the commit > in mainline. It applies directly (no rejects, no fuzz). /* Steinar */
A patch referencing this bug report has been merged in v2.6.39: commit ca06707022d6ba4744198a8ebbe4994786b0c613 Author: Steinar H. Gunderson <sgunderson@bigfoot.com> Date: Fri May 6 23:44:46 2011 +0000 ipv6: restore correct ECN handling on TCP xmit
On Fri, May 13, 2011 at 11:45:04AM +0200, Steinar H. Gunderson wrote: >> If this doesn't apply directly to 2.6.32 then you'll need to prepare the >> backport and send it to stable@kernel.org with a reference to the commit >> in mainline. > It applies directly (no rejects, no fuzz). Just to clear up: If it applies directly, will it automatically get picked up in 2.6.32? Or should I sent it to stable@kernel.org myself? /* Steinar */