From 31daf0393fbb17cf6efe613fb538a3ea4b5202e4 Mon Sep 17 00:00:00 2001 From: Gerrit Renker Date: Sun, 24 Jul 2011 21:18:25 -0600 Subject: dccp ccid-2: use feature-negotiation to report Ack Ratio changes This uses the new feature-negotiation framework to signal Ack Ratio changes, as required by RFC 4341, sec. 6.1.2. That raises some problems with CCID-2, which at the moment can not cope gracefully with Ack Ratios > 1. Since these issues are not directly related to feature negotiation, they are marked by a FIXME. Signed-off-by: Gerrit Renker Signed-off-by: Samuel Jero Acked-by: Ian McDonald --- net/dccp/ccids/ccid2.c | 10 +++++++++- 1 file changed, 9 insertions(+), 1 deletion(-) (limited to 'net/dccp/ccids') diff --git a/net/dccp/ccids/ccid2.c b/net/dccp/ccids/ccid2.c index 0462040..b51cc92 100644 --- a/net/dccp/ccids/ccid2.c +++ b/net/dccp/ccids/ccid2.c @@ -494,8 +494,16 @@ static void ccid2_hc_tx_packet_recv(struct sock *sk, struct sk_buff *skb) if (hc->tx_rpdupack >= NUMDUPACK) { hc->tx_rpdupack = -1; /* XXX lame */ hc->tx_rpseq = 0; - +#ifdef __CCID2_COPES_GRACEFULLY_WITH_ACK_CONGESTION_CONTROL__ + /* + * FIXME: Ack Congestion Control is broken; in + * the current state instabilities occurred with + * Ack Ratios greater than 1; causing hang-ups + * and long RTO timeouts. This needs to be fixed + * before opening up dynamic changes. -- gerrit + */ ccid2_change_l_ack_ratio(sk, 2 * dp->dccps_l_ack_ratio); +#endif } } } -- cgit v1.1 From d346d886a4c7f771c184e73833133f23a18de884 Mon Sep 17 00:00:00 2001 From: Samuel Jero Date: Sun, 24 Jul 2011 20:49:19 -0600 Subject: dccp ccid-2: prevent cwnd > Sequence Window Add a check to prevent CCID-2 from increasing the cwnd greater than the Sequence Window. When the congestion window becomes bigger than the Sequence Window, CCID-2 will attempt to keep more data in the network than the DCCP Sequence Window code considers possible. This results in the Sequence Window code issuing a Sync, thereby inducing needless overhead. Further, if this occurs at the sender, CCID-2 will never detect the problem because the Acks it receives will indicate no losses. I have seen this cause a drop of 1/3rd in throughput for a connection. Also add code to adjust the Sequence Window to be about 5 times the number of packets in the network (RFC 4340, 7.5.2) and to adjust the Ack Ratio so that the remote Sequence Window will hold about 5 times the number of packets in the network. This allows the congestion window to increase correctly without being limited by the Sequence Window. Signed-off-by: Samuel Jero Acked-by: Gerrit Renker --- net/dccp/ccids/ccid2.c | 50 +++++++++++++++++++++++++++++++++++--------------- net/dccp/ccids/ccid2.h | 6 ++++++ 2 files changed, 41 insertions(+), 15 deletions(-) (limited to 'net/dccp/ccids') diff --git a/net/dccp/ccids/ccid2.c b/net/dccp/ccids/ccid2.c index b51cc92..9dbc4d8 100644 --- a/net/dccp/ccids/ccid2.c +++ b/net/dccp/ccids/ccid2.c @@ -85,7 +85,6 @@ static int ccid2_hc_tx_send_packet(struct sock *sk, struct sk_buff *skb) static void ccid2_change_l_ack_ratio(struct sock *sk, u32 val) { - struct dccp_sock *dp = dccp_sk(sk); u32 max_ratio = DIV_ROUND_UP(ccid2_hc_tx_sk(sk)->tx_cwnd, 2); /* @@ -98,14 +97,15 @@ static void ccid2_change_l_ack_ratio(struct sock *sk, u32 val) DCCP_WARN("Limiting Ack Ratio (%u) to %u\n", val, max_ratio); val = max_ratio; } - if (val > DCCPF_ACK_RATIO_MAX) - val = DCCPF_ACK_RATIO_MAX; - - if (val == dp->dccps_l_ack_ratio) - return; + dccp_feat_signal_nn_change(sk, DCCPF_ACK_RATIO, + min_t(u32, val, DCCPF_ACK_RATIO_MAX)); +} - ccid2_pr_debug("changing local ack ratio to %u\n", val); - dp->dccps_l_ack_ratio = val; +static void ccid2_change_l_seq_window(struct sock *sk, u64 val) +{ + dccp_feat_signal_nn_change(sk, DCCPF_SEQUENCE_WINDOW, + clamp_val(val, DCCPF_SEQ_WMIN, + DCCPF_SEQ_WMAX)); } static void ccid2_hc_tx_rto_expire(unsigned long data) @@ -405,17 +405,37 @@ static void ccid2_new_ack(struct sock *sk, struct ccid2_seq *seqp, unsigned int *maxincr) { struct ccid2_hc_tx_sock *hc = ccid2_hc_tx_sk(sk); - - if (hc->tx_cwnd < hc->tx_ssthresh) { - if (*maxincr > 0 && ++hc->tx_packets_acked == 2) { + struct dccp_sock *dp = dccp_sk(sk); + int r_seq_used = hc->tx_cwnd / dp->dccps_l_ack_ratio; + + if (hc->tx_cwnd < dp->dccps_l_seq_win && + r_seq_used < dp->dccps_r_seq_win) { + if (hc->tx_cwnd < hc->tx_ssthresh) { + if (*maxincr > 0 && ++hc->tx_packets_acked == 2) { + hc->tx_cwnd += 1; + *maxincr -= 1; + hc->tx_packets_acked = 0; + } + } else if (++hc->tx_packets_acked >= hc->tx_cwnd) { hc->tx_cwnd += 1; - *maxincr -= 1; hc->tx_packets_acked = 0; } - } else if (++hc->tx_packets_acked >= hc->tx_cwnd) { - hc->tx_cwnd += 1; - hc->tx_packets_acked = 0; } + + /* + * Adjust the local sequence window and the ack ratio to allow about + * 5 times the number of packets in the network (RFC 4340 7.5.2) + */ + if (r_seq_used * CCID2_WIN_CHANGE_FACTOR >= dp->dccps_r_seq_win) + ccid2_change_l_ack_ratio(sk, dp->dccps_l_ack_ratio * 2); + else if (r_seq_used * CCID2_WIN_CHANGE_FACTOR < dp->dccps_r_seq_win/2) + ccid2_change_l_ack_ratio(sk, dp->dccps_l_ack_ratio / 2 ? : 1U); + + if (hc->tx_cwnd * CCID2_WIN_CHANGE_FACTOR >= dp->dccps_l_seq_win) + ccid2_change_l_seq_window(sk, dp->dccps_l_seq_win * 2); + else if (hc->tx_cwnd * CCID2_WIN_CHANGE_FACTOR < dp->dccps_l_seq_win/2) + ccid2_change_l_seq_window(sk, dp->dccps_l_seq_win / 2); + /* * FIXME: RTT is sampled several times per acknowledgment (for each * entry in the Ack Vector), instead of once per Ack (as in TCP SACK). diff --git a/net/dccp/ccids/ccid2.h b/net/dccp/ccids/ccid2.h index f585d33..18c9754 100644 --- a/net/dccp/ccids/ccid2.h +++ b/net/dccp/ccids/ccid2.h @@ -43,6 +43,12 @@ struct ccid2_seq { #define CCID2_SEQBUF_LEN 1024 #define CCID2_SEQBUF_MAX 128 +/* + * Multiple of congestion window to keep the sequence window at + * (RFC 4340 7.5.2) + */ +#define CCID2_WIN_CHANGE_FACTOR 5 + /** * struct ccid2_hc_tx_sock - CCID2 TX half connection * @tx_{cwnd,ssthresh,pipe}: as per RFC 4341, section 5 -- cgit v1.1 From 0ce95dc792549e0cf704e74aa8acb15a401f8cca Mon Sep 17 00:00:00 2001 From: Samuel Jero Date: Sun, 24 Jul 2011 21:05:16 -0600 Subject: dccp ccid-2: increment cwnd correctly This patch fixes an issue where CCID-2 will not increase the congestion window for numerous RTTs after an idle period, application-limited period, or a loss once the algorithm is in Congestion Avoidance. What happens is that, when CCID-2 is in Congestion Avoidance mode, it will increase hc->tx_packets_acked by one for every packet and will increment cwnd every cwnd packets. However, if there is now an idle period in the connection, cwnd will be reduced, possibly below the slow start threshold. This will cause the connection to go into Slow Start. However, in Slow Start CCID-2 performs this test to increment cwnd every second ack: ++hc->tx_packets_acked == 2 Unfortunately, this will be incorrect, if cwnd previous to the idle period was larger than 2 and if tx_packets_acked was close to cwnd. For example: cwnd=50 and tx_packets_acked=45. In this case, the current code, will increment tx_packets_acked until it equals two, which will only be once tx_packets_acked (an unsigned 32-bit integer) overflows. My fix is simply to change that test for tx_packets_acked greater than or equal to two in slow start. Signed-off-by: Samuel Jero Acked-by: Gerrit Renker --- net/dccp/ccids/ccid2.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) (limited to 'net/dccp/ccids') diff --git a/net/dccp/ccids/ccid2.c b/net/dccp/ccids/ccid2.c index 9dbc4d8..d9dbd9f 100644 --- a/net/dccp/ccids/ccid2.c +++ b/net/dccp/ccids/ccid2.c @@ -411,7 +411,7 @@ static void ccid2_new_ack(struct sock *sk, struct ccid2_seq *seqp, if (hc->tx_cwnd < dp->dccps_l_seq_win && r_seq_used < dp->dccps_r_seq_win) { if (hc->tx_cwnd < hc->tx_ssthresh) { - if (*maxincr > 0 && ++hc->tx_packets_acked == 2) { + if (*maxincr > 0 && ++hc->tx_packets_acked >= 2) { hc->tx_cwnd += 1; *maxincr -= 1; hc->tx_packets_acked = 0; -- cgit v1.1 From d96a9e8dd04cf5ab2782ca6192e395c5ca373f7d Mon Sep 17 00:00:00 2001 From: Samuel Jero Date: Sun, 24 Jul 2011 20:57:49 -0600 Subject: dccp ccid-2: check Ack Ratio when reducing cwnd This patch causes CCID-2 to check the Ack Ratio after reducing the congestion window. If the Ack Ratio is greater than the congestion window, it is reduced. This prevents timeouts caused by an Ack Ratio larger than the congestion window. In this situation, we choose to set the Ack Ratio to half the congestion window (or one if that's zero) so that if we loose one ack we don't trigger a timeout. Signed-off-by: Samuel Jero Acked-by: Gerrit Renker --- net/dccp/ccids/ccid2.c | 26 +++++++++++++++++++++++--- 1 file changed, 23 insertions(+), 3 deletions(-) (limited to 'net/dccp/ccids') diff --git a/net/dccp/ccids/ccid2.c b/net/dccp/ccids/ccid2.c index d9dbd9f..67164bb 100644 --- a/net/dccp/ccids/ccid2.c +++ b/net/dccp/ccids/ccid2.c @@ -101,6 +101,24 @@ static void ccid2_change_l_ack_ratio(struct sock *sk, u32 val) min_t(u32, val, DCCPF_ACK_RATIO_MAX)); } +static void ccid2_check_l_ack_ratio(struct sock *sk) +{ + struct ccid2_hc_tx_sock *hc = ccid2_hc_tx_sk(sk); + + /* + * After a loss, idle period, application limited period, or RTO we + * need to check that the ack ratio is still less than the congestion + * window. Otherwise, we will send an entire congestion window of + * packets and got no response because we haven't sent ack ratio + * packets yet. + * If the ack ratio does need to be reduced, we reduce it to half of + * the congestion window (or 1 if that's zero) instead of to the + * congestion window. This prevents problems if one ack is lost. + */ + if (dccp_feat_nn_get(sk, DCCPF_ACK_RATIO) > hc->tx_cwnd) + ccid2_change_l_ack_ratio(sk, hc->tx_cwnd/2 ? : 1U); +} + static void ccid2_change_l_seq_window(struct sock *sk, u64 val) { dccp_feat_signal_nn_change(sk, DCCPF_SEQUENCE_WINDOW, @@ -187,6 +205,8 @@ static void ccid2_cwnd_application_limited(struct sock *sk, const u32 now) } hc->tx_cwnd_used = 0; hc->tx_cwnd_stamp = now; + + ccid2_check_l_ack_ratio(sk); } /* This borrows the code of tcp_cwnd_restart() */ @@ -205,6 +225,8 @@ static void ccid2_cwnd_restart(struct sock *sk, const u32 now) hc->tx_cwnd_stamp = now; hc->tx_cwnd_used = 0; + + ccid2_check_l_ack_ratio(sk); } static void ccid2_hc_tx_packet_sent(struct sock *sk, unsigned int len) @@ -461,9 +483,7 @@ static void ccid2_congestion_event(struct sock *sk, struct ccid2_seq *seqp) hc->tx_cwnd = hc->tx_cwnd / 2 ? : 1U; hc->tx_ssthresh = max(hc->tx_cwnd, 2U); - /* Avoid spurious timeouts resulting from Ack Ratio > cwnd */ - if (dccp_sk(sk)->dccps_l_ack_ratio > hc->tx_cwnd) - ccid2_change_l_ack_ratio(sk, hc->tx_cwnd); + ccid2_check_l_ack_ratio(sk); } static int ccid2_hc_tx_parse_options(struct sock *sk, u8 packet_type, -- cgit v1.1