From 375fe02c91792917aa26d68a87ab110d1937f44e Mon Sep 17 00:00:00 2001 From: Yuchung Cheng Date: Mon, 22 Jul 2013 16:20:45 -0700 Subject: tcp: consolidate SYNACK RTT sampling The first patch consolidates SYNACK and other RTT measurement to use a central function tcp_ack_update_rtt(). A (small) bonus is now SYNACK RTT measurement happens after PAWS check, potentially reducing the impact of RTO seeding on bad TCP timestamps values. Signed-off-by: Yuchung Cheng Signed-off-by: David S. Miller --- net/ipv4/tcp_input.c | 14 ++++++++++++-- 1 file changed, 12 insertions(+), 2 deletions(-) (limited to 'net/ipv4/tcp_input.c') diff --git a/net/ipv4/tcp_input.c b/net/ipv4/tcp_input.c index 28af45abe062..b531710596ec 100644 --- a/net/ipv4/tcp_input.c +++ b/net/ipv4/tcp_input.c @@ -2853,6 +2853,17 @@ static inline void tcp_ack_update_rtt(struct sock *sk, const int flag, tcp_ack_no_tstamp(sk, seq_rtt, flag); } +/* Compute time elapsed between (last) SYNACK and the ACK completing 3WHS. */ +static void tcp_synack_rtt_meas(struct sock *sk, struct request_sock *req) +{ + struct tcp_sock *tp = tcp_sk(sk); + s32 seq_rtt = -1; + + if (tp->lsndtime && !tp->total_retrans) + seq_rtt = tcp_time_stamp - tp->lsndtime; + tcp_ack_update_rtt(sk, FLAG_SYN_ACKED, seq_rtt); +} + static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 in_flight) { const struct inet_connection_sock *icsk = inet_csk(sk); @@ -5624,9 +5635,7 @@ int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb, * so release it. */ if (req) { - tcp_synack_rtt_meas(sk, req); tp->total_retrans = req->num_retrans; - reqsk_fastopen_remove(sk, req, false); } else { /* Make sure socket is routed, for correct metrics. */ @@ -5651,6 +5660,7 @@ int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb, tp->snd_una = TCP_SKB_CB(skb)->ack_seq; tp->snd_wnd = ntohs(th->window) << tp->rx_opt.snd_wscale; tcp_init_wl(tp, TCP_SKB_CB(skb)->seq); + tcp_synack_rtt_meas(sk, req); if (tp->rx_opt.tstamp_ok) tp->advmss -= TCPOLEN_TSTAMP_ALIGNED; -- cgit v1.2.3-59-g8ed1b From 5b08e47caf1f2034a3a5b566bbccc8b0be3961ca Mon Sep 17 00:00:00 2001 From: Yuchung Cheng Date: Mon, 22 Jul 2013 16:20:46 -0700 Subject: tcp: prefer packet timing to TS-ECR for RTT Prefer packet timings to TS-ecr for RTT measurements when both sources are available. That's because broken middle-boxes and remote peer can return packets with corrupted TS ECR fields. Similarly most congestion controls that require RTT signals favor timing-based sources as well. Also check for bad TS ECR values to avoid RTT blow-ups. It has happened on production Web servers. Signed-off-by: Yuchung Cheng Acked-by: Neal Cardwell Signed-off-by: David S. Miller --- include/net/tcp.h | 1 - net/ipv4/tcp_input.c | 67 ++++++++++++++-------------------------------------- 2 files changed, 18 insertions(+), 50 deletions(-) (limited to 'net/ipv4/tcp_input.c') diff --git a/include/net/tcp.h b/include/net/tcp.h index f9777dbede75..c5868471abae 100644 --- a/include/net/tcp.h +++ b/include/net/tcp.h @@ -591,7 +591,6 @@ extern void tcp_initialize_rcv_mss(struct sock *sk); extern int tcp_mtu_to_mss(struct sock *sk, int pmtu); extern int tcp_mss_to_mtu(struct sock *sk, int mss); extern void tcp_mtup_init(struct sock *sk); -extern void tcp_valid_rtt_meas(struct sock *sk, u32 seq_rtt); extern void tcp_init_buffer_space(struct sock *sk); static inline void tcp_bound_rto(const struct sock *sk) diff --git a/net/ipv4/tcp_input.c b/net/ipv4/tcp_input.c index b531710596ec..c7398f05d12b 100644 --- a/net/ipv4/tcp_input.c +++ b/net/ipv4/tcp_input.c @@ -2792,65 +2792,36 @@ static void tcp_fastretrans_alert(struct sock *sk, const int acked, tcp_xmit_retransmit_queue(sk); } -void tcp_valid_rtt_meas(struct sock *sk, u32 seq_rtt) +static inline void tcp_ack_update_rtt(struct sock *sk, const int flag, + s32 seq_rtt) { - tcp_rtt_estimator(sk, seq_rtt); - tcp_set_rto(sk); - inet_csk(sk)->icsk_backoff = 0; -} -EXPORT_SYMBOL(tcp_valid_rtt_meas); + const struct tcp_sock *tp = tcp_sk(sk); + + /* Prefer RTT measured from ACK's timing to TS-ECR. This is because + * broken middle-boxes or peers may corrupt TS-ECR fields. But + * Karn's algorithm forbids taking RTT if some retransmitted data + * is acked (RFC6298). + */ + if (flag & FLAG_RETRANS_DATA_ACKED) + seq_rtt = -1; -/* Read draft-ietf-tcplw-high-performance before mucking - * with this code. (Supersedes RFC1323) - */ -static void tcp_ack_saw_tstamp(struct sock *sk, int flag) -{ /* RTTM Rule: A TSecr value received in a segment is used to * update the averaged RTT measurement only if the segment * acknowledges some new data, i.e., only if it advances the * left edge of the send window. - * * See draft-ietf-tcplw-high-performance-00, section 3.3. - * 1998/04/10 Andrey V. Savochkin - * - * Changed: reset backoff as soon as we see the first valid sample. - * If we do not, we get strongly overestimated rto. With timestamps - * samples are accepted even from very old segments: f.e., when rtt=1 - * increases to 8, we retransmit 5 times and after 8 seconds delayed - * answer arrives rto becomes 120 seconds! If at least one of segments - * in window is lost... Voila. --ANK (010210) */ - struct tcp_sock *tp = tcp_sk(sk); + if (seq_rtt < 0 && tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr) + seq_rtt = tcp_time_stamp - tp->rx_opt.rcv_tsecr; - tcp_valid_rtt_meas(sk, tcp_time_stamp - tp->rx_opt.rcv_tsecr); -} - -static void tcp_ack_no_tstamp(struct sock *sk, u32 seq_rtt, int flag) -{ - /* We don't have a timestamp. Can only use - * packets that are not retransmitted to determine - * rtt estimates. Also, we must not reset the - * backoff for rto until we get a non-retransmitted - * packet. This allows us to deal with a situation - * where the network delay has increased suddenly. - * I.e. Karn's algorithm. (SIGCOMM '87, p5.) - */ - - if (flag & FLAG_RETRANS_DATA_ACKED) + if (seq_rtt < 0) return; - tcp_valid_rtt_meas(sk, seq_rtt); -} + tcp_rtt_estimator(sk, seq_rtt); + tcp_set_rto(sk); -static inline void tcp_ack_update_rtt(struct sock *sk, const int flag, - const s32 seq_rtt) -{ - const struct tcp_sock *tp = tcp_sk(sk); - /* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */ - if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr) - tcp_ack_saw_tstamp(sk, flag); - else if (seq_rtt >= 0) - tcp_ack_no_tstamp(sk, seq_rtt, flag); + /* RFC6298: only reset backoff on valid RTT measurement. */ + inet_csk(sk)->icsk_backoff = 0; } /* Compute time elapsed between (last) SYNACK and the ACK completing 3WHS. */ @@ -2989,8 +2960,6 @@ static int tcp_clean_rtx_queue(struct sock *sk, int prior_fackets, if (sacked & TCPCB_SACKED_RETRANS) tp->retrans_out -= acked_pcount; flag |= FLAG_RETRANS_DATA_ACKED; - ca_seq_rtt = -1; - seq_rtt = -1; } else { ca_seq_rtt = now - scb->when; last_ackt = skb->tstamp; -- cgit v1.2.3-59-g8ed1b From 59c9af4234b0c21a1ed05cf65bf014d0c1a67bfd Mon Sep 17 00:00:00 2001 From: Yuchung Cheng Date: Mon, 22 Jul 2013 16:20:47 -0700 Subject: tcp: measure RTT from new SACK Take RTT sample if an ACK selectively acks some sequences that have never been retransmitted. The Karn's algorithm does not apply even if that ACK (s)acks other retransmitted sequences, because it must been generated by an original but perhaps out-of-order packet. There is no ambiguity. In case when multiple blocks are newly sacked because of ACK losses the earliest block is used to measure RTT, similar to cummulative ACKs. Such RTT samples allow the sender to estimate the RTO during loss recovery and packet reordering events. It is still useful even with TCP timestamps. That's because during these events the SND.UNA may not advance preventing RTT samples from TS ECR (thus the FLAG_ACKED check before calling tcp_ack_update_rtt()). Therefore this new RTT source is complementary to existing ACK and TS RTT mechanisms. This patch does not update the RTO. It is done in the next patch. Signed-off-by: Yuchung Cheng Signed-off-by: David S. Miller --- net/ipv4/tcp_input.c | 23 +++++++++++++++++------ 1 file changed, 17 insertions(+), 6 deletions(-) (limited to 'net/ipv4/tcp_input.c') diff --git a/net/ipv4/tcp_input.c b/net/ipv4/tcp_input.c index c7398f05d12b..b85bc7c3736a 100644 --- a/net/ipv4/tcp_input.c +++ b/net/ipv4/tcp_input.c @@ -1048,6 +1048,7 @@ struct tcp_sacktag_state { int reord; int fack_count; int flag; + s32 rtt; /* RTT measured by SACKing never-retransmitted data */ }; /* Check if skb is fully within the SACK block. In presence of GSO skbs, @@ -1108,7 +1109,7 @@ static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb, static u8 tcp_sacktag_one(struct sock *sk, struct tcp_sacktag_state *state, u8 sacked, u32 start_seq, u32 end_seq, - bool dup_sack, int pcount) + int dup_sack, int pcount, u32 xmit_time) { struct tcp_sock *tp = tcp_sk(sk); int fack_count = state->fack_count; @@ -1148,6 +1149,9 @@ static u8 tcp_sacktag_one(struct sock *sk, state->reord); if (!after(end_seq, tp->high_seq)) state->flag |= FLAG_ORIG_SACK_ACKED; + /* Pick the earliest sequence sacked for RTT */ + if (state->rtt < 0) + state->rtt = tcp_time_stamp - xmit_time; } if (sacked & TCPCB_LOST) { @@ -1205,7 +1209,8 @@ static bool tcp_shifted_skb(struct sock *sk, struct sk_buff *skb, * tcp_highest_sack_seq() when skb is highest_sack. */ tcp_sacktag_one(sk, state, TCP_SKB_CB(skb)->sacked, - start_seq, end_seq, dup_sack, pcount); + start_seq, end_seq, dup_sack, pcount, + TCP_SKB_CB(skb)->when); if (skb == tp->lost_skb_hint) tp->lost_cnt_hint += pcount; @@ -1479,7 +1484,8 @@ static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq, dup_sack, - tcp_skb_pcount(skb)); + tcp_skb_pcount(skb), + TCP_SKB_CB(skb)->when); if (!before(TCP_SKB_CB(skb)->seq, tcp_highest_sack_seq(tp))) @@ -1536,7 +1542,7 @@ static int tcp_sack_cache_ok(const struct tcp_sock *tp, const struct tcp_sack_bl static int tcp_sacktag_write_queue(struct sock *sk, const struct sk_buff *ack_skb, - u32 prior_snd_una) + u32 prior_snd_una, s32 *sack_rtt) { struct tcp_sock *tp = tcp_sk(sk); const unsigned char *ptr = (skb_transport_header(ack_skb) + @@ -1554,6 +1560,7 @@ tcp_sacktag_write_queue(struct sock *sk, const struct sk_buff *ack_skb, state.flag = 0; state.reord = tp->packets_out; + state.rtt = -1; if (!tp->sacked_out) { if (WARN_ON(tp->fackets_out)) @@ -1737,6 +1744,7 @@ out: WARN_ON((int)tp->retrans_out < 0); WARN_ON((int)tcp_packets_in_flight(tp) < 0); #endif + *sack_rtt = state.rtt; return state.flag; } @@ -3254,6 +3262,7 @@ static int tcp_ack(struct sock *sk, const struct sk_buff *skb, int flag) int prior_packets = tp->packets_out; const int prior_unsacked = tp->packets_out - tp->sacked_out; int acked = 0; /* Number of packets newly acked */ + s32 sack_rtt = -1; /* If the ack is older than previous acks * then we can probably ignore it. @@ -3310,7 +3319,8 @@ static int tcp_ack(struct sock *sk, const struct sk_buff *skb, int flag) flag |= tcp_ack_update_window(sk, skb, ack, ack_seq); if (TCP_SKB_CB(skb)->sacked) - flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una); + flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una, + &sack_rtt); if (TCP_ECN_rcv_ecn_echo(tp, tcp_hdr(skb))) flag |= FLAG_ECE; @@ -3382,7 +3392,8 @@ old_ack: * If data was DSACKed, see if we can undo a cwnd reduction. */ if (TCP_SKB_CB(skb)->sacked) { - flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una); + flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una, + &sack_rtt); tcp_fastretrans_alert(sk, acked, prior_unsacked, is_dupack, flag); } -- cgit v1.2.3-59-g8ed1b From ed08495c31bb991de636d2488abaa50b39f2ff4a Mon Sep 17 00:00:00 2001 From: Yuchung Cheng Date: Mon, 22 Jul 2013 16:20:48 -0700 Subject: tcp: use RTT from SACK for RTO If RTT is not available because Karn's check has failed or no new packet is acked, use the RTT measured from SACK to estimate the RTO. The sender can continue to estimate the RTO during loss recovery or reordering event upon receiving non-partial ACKs. This also changes when the RTO is re-armed. Previously it is only re-armed when some data is cummulatively acknowledged (i.e., SND.UNA advances), but now it is re-armed whenever RTT estimator is updated. This feature is particularly useful to reduce spurious timeout for buffer bloat including cellular carriers [1], and RTT estimation on reordering events. [1] "An In-depth Study of LTE: Effect of Network Protocol and Application Behavior on Performance", In Proc. of SIGCOMM 2013 Signed-off-by: Yuchung Cheng Acked-by: Neal Cardwell Signed-off-by: David S. Miller --- net/ipv4/tcp_input.c | 23 ++++++++++++++--------- 1 file changed, 14 insertions(+), 9 deletions(-) (limited to 'net/ipv4/tcp_input.c') diff --git a/net/ipv4/tcp_input.c b/net/ipv4/tcp_input.c index b85bc7c3736a..b61274b666f6 100644 --- a/net/ipv4/tcp_input.c +++ b/net/ipv4/tcp_input.c @@ -2800,8 +2800,8 @@ static void tcp_fastretrans_alert(struct sock *sk, const int acked, tcp_xmit_retransmit_queue(sk); } -static inline void tcp_ack_update_rtt(struct sock *sk, const int flag, - s32 seq_rtt) +static inline bool tcp_ack_update_rtt(struct sock *sk, const int flag, + s32 seq_rtt, s32 sack_rtt) { const struct tcp_sock *tp = tcp_sk(sk); @@ -2813,6 +2813,9 @@ static inline void tcp_ack_update_rtt(struct sock *sk, const int flag, if (flag & FLAG_RETRANS_DATA_ACKED) seq_rtt = -1; + if (seq_rtt < 0) + seq_rtt = sack_rtt; + /* RTTM Rule: A TSecr value received in a segment is used to * update the averaged RTT measurement only if the segment * acknowledges some new data, i.e., only if it advances the @@ -2823,13 +2826,14 @@ static inline void tcp_ack_update_rtt(struct sock *sk, const int flag, seq_rtt = tcp_time_stamp - tp->rx_opt.rcv_tsecr; if (seq_rtt < 0) - return; + return false; tcp_rtt_estimator(sk, seq_rtt); tcp_set_rto(sk); /* RFC6298: only reset backoff on valid RTT measurement. */ inet_csk(sk)->icsk_backoff = 0; + return true; } /* Compute time elapsed between (last) SYNACK and the ACK completing 3WHS. */ @@ -2840,7 +2844,7 @@ static void tcp_synack_rtt_meas(struct sock *sk, struct request_sock *req) if (tp->lsndtime && !tp->total_retrans) seq_rtt = tcp_time_stamp - tp->lsndtime; - tcp_ack_update_rtt(sk, FLAG_SYN_ACKED, seq_rtt); + tcp_ack_update_rtt(sk, FLAG_SYN_ACKED, seq_rtt, -1); } static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 in_flight) @@ -2929,7 +2933,7 @@ static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb) * arrived at the other end. */ static int tcp_clean_rtx_queue(struct sock *sk, int prior_fackets, - u32 prior_snd_una) + u32 prior_snd_una, s32 sack_rtt) { struct tcp_sock *tp = tcp_sk(sk); const struct inet_connection_sock *icsk = inet_csk(sk); @@ -3019,6 +3023,10 @@ static int tcp_clean_rtx_queue(struct sock *sk, int prior_fackets, if (skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) flag |= FLAG_SACK_RENEGING; + if (tcp_ack_update_rtt(sk, flag, seq_rtt, sack_rtt) || + (flag & FLAG_ACKED)) + tcp_rearm_rto(sk); + if (flag & FLAG_ACKED) { const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops; @@ -3028,9 +3036,6 @@ static int tcp_clean_rtx_queue(struct sock *sk, int prior_fackets, tcp_mtup_probe_success(sk); } - tcp_ack_update_rtt(sk, flag, seq_rtt); - tcp_rearm_rto(sk); - if (tcp_is_reno(tp)) { tcp_remove_reno_sacks(sk, pkts_acked); } else { @@ -3339,7 +3344,7 @@ static int tcp_ack(struct sock *sk, const struct sk_buff *skb, int flag) /* See if we can take anything off of the retransmit queue. */ acked = tp->packets_out; - flag |= tcp_clean_rtx_queue(sk, prior_fackets, prior_snd_una); + flag |= tcp_clean_rtx_queue(sk, prior_fackets, prior_snd_una, sack_rtt); acked -= tp->packets_out; if (tcp_ack_is_dubious(sk, flag)) { -- cgit v1.2.3-59-g8ed1b From 74c181d528bd8b5989f424a489262d0742ca31ae Mon Sep 17 00:00:00 2001 From: Yuchung Cheng Date: Mon, 12 Aug 2013 16:41:25 -0700 Subject: tcp: reset reordering est. selectively on timeout On timeout the TCP sender unconditionally resets the estimated degree of network reordering (tp->reordering). The idea behind this is that the estimate is too large to trigger fast recovery (e.g., due to a IP path change). But for example if the sender only had 2 packets outstanding, then a timeout doesn't tell much about reordering. A sender that learns about reordering on big writes and loses packets on small writes will end up falsely retransmitting again and again, especially when reordering is more likely on big writes. Therefore the sender should only suspect that tp->reordering is too high if it could have gone into fast recovery with the (lower) default estimate. Signed-off-by: Yuchung Cheng Acked-by: Neal Cardwell Signed-off-by: David S. Miller --- net/ipv4/tcp_input.c | 9 +++++++-- 1 file changed, 7 insertions(+), 2 deletions(-) (limited to 'net/ipv4/tcp_input.c') diff --git a/net/ipv4/tcp_input.c b/net/ipv4/tcp_input.c index b61274b666f6..e965cc7b87ff 100644 --- a/net/ipv4/tcp_input.c +++ b/net/ipv4/tcp_input.c @@ -1877,8 +1877,13 @@ void tcp_enter_loss(struct sock *sk, int how) } tcp_verify_left_out(tp); - tp->reordering = min_t(unsigned int, tp->reordering, - sysctl_tcp_reordering); + /* Timeout in disordered state after receiving substantial DUPACKs + * suggests that the degree of reordering is over-estimated. + */ + if (icsk->icsk_ca_state <= TCP_CA_Disorder && + tp->sacked_out >= sysctl_tcp_reordering) + tp->reordering = min_t(unsigned int, tp->reordering, + sysctl_tcp_reordering); tcp_set_ca_state(sk, TCP_CA_Loss); tp->high_seq = tp->snd_nxt; TCP_ECN_queue_cwr(tp); -- cgit v1.2.3-59-g8ed1b From 0f7cc9a3c2bd89b15720dbf358e9b9e62af27126 Mon Sep 17 00:00:00 2001 From: Yuchung Cheng Date: Wed, 21 Aug 2013 17:29:23 -0700 Subject: tcp: increase throughput when reordering is high The stack currently detects reordering and avoid spurious retransmission very well. However the throughput is sub-optimal under high reordering because cwnd is increased only if the data is deliverd in order. I.e., FLAG_DATA_ACKED check in tcp_ack(). The more packet are reordered the worse the throughput is. Therefore when reordering is proven high, cwnd should advance whenever the data is delivered regardless of its ordering. If reordering is low, conservatively advance cwnd only on ordered deliveries in Open state, and retain cwnd in Disordered state (RFC5681). Using netperf on a qdisc setup of 20Mbps BW and random RTT from 45ms to 55ms (for reordering effect). This change increases TCP throughput by 20 - 25% to near bottleneck BW. A special case is the stretched ACK with new SACK and/or ECE mark. For example, a receiver may receive an out of order or ECN packet with unacked data buffered because of LRO or delayed ACK. The principle on such an ACK is to advance cwnd on the cummulative acked part first, then reduce cwnd in tcp_fastretrans_alert(). Signed-off-by: Yuchung Cheng Acked-by: Neal Cardwell Acked-by: Eric Dumazet Signed-off-by: David S. Miller --- net/ipv4/tcp_input.c | 32 ++++++++++++++++++++------------ 1 file changed, 20 insertions(+), 12 deletions(-) (limited to 'net/ipv4/tcp_input.c') diff --git a/net/ipv4/tcp_input.c b/net/ipv4/tcp_input.c index e965cc7b87ff..ec492eae0cd7 100644 --- a/net/ipv4/tcp_input.c +++ b/net/ipv4/tcp_input.c @@ -2485,8 +2485,6 @@ static void tcp_try_to_open(struct sock *sk, int flag, const int prior_unsacked) if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) { tcp_try_keep_open(sk); - if (inet_csk(sk)->icsk_ca_state != TCP_CA_Open) - tcp_moderate_cwnd(tp); } else { tcp_cwnd_reduction(sk, prior_unsacked, 0); } @@ -3128,11 +3126,24 @@ static inline bool tcp_ack_is_dubious(const struct sock *sk, const int flag) inet_csk(sk)->icsk_ca_state != TCP_CA_Open; } +/* Decide wheather to run the increase function of congestion control. */ static inline bool tcp_may_raise_cwnd(const struct sock *sk, const int flag) { - const struct tcp_sock *tp = tcp_sk(sk); - return (!(flag & FLAG_ECE) || tp->snd_cwnd < tp->snd_ssthresh) && - !tcp_in_cwnd_reduction(sk); + if (tcp_in_cwnd_reduction(sk)) + return false; + + /* If reordering is high then always grow cwnd whenever data is + * delivered regardless of its ordering. Otherwise stay conservative + * and only grow cwnd on in-order delivery in Open state, and retain + * cwnd in Disordered state (RFC5681). A stretched ACK with + * new SACK or ECE mark may first advance cwnd here and later reduce + * cwnd in tcp_fastretrans_alert() based on more states. + */ + if (tcp_sk(sk)->reordering > sysctl_tcp_reordering) + return flag & FLAG_FORWARD_PROGRESS; + + return inet_csk(sk)->icsk_ca_state == TCP_CA_Open && + flag & FLAG_DATA_ACKED; } /* Check that window update is acceptable. @@ -3352,18 +3363,15 @@ static int tcp_ack(struct sock *sk, const struct sk_buff *skb, int flag) flag |= tcp_clean_rtx_queue(sk, prior_fackets, prior_snd_una, sack_rtt); acked -= tp->packets_out; + /* Advance cwnd if state allows */ + if (tcp_may_raise_cwnd(sk, flag)) + tcp_cong_avoid(sk, ack, prior_in_flight); + if (tcp_ack_is_dubious(sk, flag)) { - /* Advance CWND, if state allows this. */ - if ((flag & FLAG_DATA_ACKED) && tcp_may_raise_cwnd(sk, flag)) - tcp_cong_avoid(sk, ack, prior_in_flight); is_dupack = !(flag & (FLAG_SND_UNA_ADVANCED | FLAG_NOT_DUP)); tcp_fastretrans_alert(sk, acked, prior_unsacked, is_dupack, flag); - } else { - if (flag & FLAG_DATA_ACKED) - tcp_cong_avoid(sk, ack, prior_in_flight); } - if (tp->tlp_high_seq) tcp_process_tlp_ack(sk, ack, flag); -- cgit v1.2.3-59-g8ed1b From 95bd09eb27507691520d39ee1044d6ad831c1168 Mon Sep 17 00:00:00 2001 From: Eric Dumazet Date: Tue, 27 Aug 2013 05:46:32 -0700 Subject: tcp: TSO packets automatic sizing After hearing many people over past years complaining against TSO being bursty or even buggy, we are proud to present automatic sizing of TSO packets. One part of the problem is that tcp_tso_should_defer() uses an heuristic relying on upcoming ACKS instead of a timer, but more generally, having big TSO packets makes little sense for low rates, as it tends to create micro bursts on the network, and general consensus is to reduce the buffering amount. This patch introduces a per socket sk_pacing_rate, that approximates the current sending rate, and allows us to size the TSO packets so that we try to send one packet every ms. This field could be set by other transports. Patch has no impact for high speed flows, where having large TSO packets makes sense to reach line rate. For other flows, this helps better packet scheduling and ACK clocking. This patch increases performance of TCP flows in lossy environments. A new sysctl (tcp_min_tso_segs) is added, to specify the minimal size of a TSO packet (default being 2). A follow-up patch will provide a new packet scheduler (FQ), using sk_pacing_rate as an input to perform optional per flow pacing. This explains why we chose to set sk_pacing_rate to twice the current rate, allowing 'slow start' ramp up. sk_pacing_rate = 2 * cwnd * mss / srtt v2: Neal Cardwell reported a suspect deferring of last two segments on initial write of 10 MSS, I had to change tcp_tso_should_defer() to take into account tp->xmit_size_goal_segs Signed-off-by: Eric Dumazet Cc: Neal Cardwell Cc: Yuchung Cheng Cc: Van Jacobson Cc: Tom Herbert Acked-by: Yuchung Cheng Acked-by: Neal Cardwell Signed-off-by: David S. Miller --- Documentation/networking/ip-sysctl.txt | 9 +++++++++ include/net/sock.h | 2 ++ include/net/tcp.h | 1 + net/ipv4/sysctl_net_ipv4.c | 10 ++++++++++ net/ipv4/tcp.c | 28 +++++++++++++++++++++++----- net/ipv4/tcp_input.c | 32 +++++++++++++++++++++++++++++++- net/ipv4/tcp_output.c | 2 +- 7 files changed, 77 insertions(+), 7 deletions(-) (limited to 'net/ipv4/tcp_input.c') diff --git a/Documentation/networking/ip-sysctl.txt b/Documentation/networking/ip-sysctl.txt index a2be556032c9..1cb3aeb4baff 100644 --- a/Documentation/networking/ip-sysctl.txt +++ b/Documentation/networking/ip-sysctl.txt @@ -482,6 +482,15 @@ tcp_syn_retries - INTEGER tcp_timestamps - BOOLEAN Enable timestamps as defined in RFC1323. +tcp_min_tso_segs - INTEGER + Minimal number of segments per TSO frame. + Since linux-3.12, TCP does an automatic sizing of TSO frames, + depending on flow rate, instead of filling 64Kbytes packets. + For specific usages, it's possible to force TCP to build big + TSO frames. Note that TCP stack might split too big TSO packets + if available window is too small. + Default: 2 + tcp_tso_win_divisor - INTEGER This allows control over what percentage of the congestion window can be consumed by a single TSO frame. diff --git a/include/net/sock.h b/include/net/sock.h index e4bbcbfd07ea..6ba2e7b0e2b1 100644 --- a/include/net/sock.h +++ b/include/net/sock.h @@ -232,6 +232,7 @@ struct cg_proto; * @sk_napi_id: id of the last napi context to receive data for sk * @sk_ll_usec: usecs to busypoll when there is no data * @sk_allocation: allocation mode + * @sk_pacing_rate: Pacing rate (if supported by transport/packet scheduler) * @sk_sndbuf: size of send buffer in bytes * @sk_flags: %SO_LINGER (l_onoff), %SO_BROADCAST, %SO_KEEPALIVE, * %SO_OOBINLINE settings, %SO_TIMESTAMPING settings @@ -361,6 +362,7 @@ struct sock { kmemcheck_bitfield_end(flags); int sk_wmem_queued; gfp_t sk_allocation; + u32 sk_pacing_rate; /* bytes per second */ netdev_features_t sk_route_caps; netdev_features_t sk_route_nocaps; int sk_gso_type; diff --git a/include/net/tcp.h b/include/net/tcp.h index dd5e16f66f84..6a6a88db462d 100644 --- a/include/net/tcp.h +++ b/include/net/tcp.h @@ -281,6 +281,7 @@ extern int sysctl_tcp_early_retrans; extern int sysctl_tcp_limit_output_bytes; extern int sysctl_tcp_challenge_ack_limit; extern unsigned int sysctl_tcp_notsent_lowat; +extern int sysctl_tcp_min_tso_segs; extern atomic_long_t tcp_memory_allocated; extern struct percpu_counter tcp_sockets_allocated; diff --git a/net/ipv4/sysctl_net_ipv4.c b/net/ipv4/sysctl_net_ipv4.c index 8ed7c32ae28e..540279f4c531 100644 --- a/net/ipv4/sysctl_net_ipv4.c +++ b/net/ipv4/sysctl_net_ipv4.c @@ -29,6 +29,7 @@ static int zero; static int one = 1; static int four = 4; +static int gso_max_segs = GSO_MAX_SEGS; static int tcp_retr1_max = 255; static int ip_local_port_range_min[] = { 1, 1 }; static int ip_local_port_range_max[] = { 65535, 65535 }; @@ -760,6 +761,15 @@ static struct ctl_table ipv4_table[] = { .extra1 = &zero, .extra2 = &four, }, + { + .procname = "tcp_min_tso_segs", + .data = &sysctl_tcp_min_tso_segs, + .maxlen = sizeof(int), + .mode = 0644, + .proc_handler = proc_dointvec_minmax, + .extra1 = &zero, + .extra2 = &gso_max_segs, + }, { .procname = "udp_mem", .data = &sysctl_udp_mem, diff --git a/net/ipv4/tcp.c b/net/ipv4/tcp.c index 4e42c03859f4..fdf74090a001 100644 --- a/net/ipv4/tcp.c +++ b/net/ipv4/tcp.c @@ -283,6 +283,8 @@ int sysctl_tcp_fin_timeout __read_mostly = TCP_FIN_TIMEOUT; +int sysctl_tcp_min_tso_segs __read_mostly = 2; + struct percpu_counter tcp_orphan_count; EXPORT_SYMBOL_GPL(tcp_orphan_count); @@ -785,12 +787,28 @@ static unsigned int tcp_xmit_size_goal(struct sock *sk, u32 mss_now, xmit_size_goal = mss_now; if (large_allowed && sk_can_gso(sk)) { - xmit_size_goal = ((sk->sk_gso_max_size - 1) - - inet_csk(sk)->icsk_af_ops->net_header_len - - inet_csk(sk)->icsk_ext_hdr_len - - tp->tcp_header_len); + u32 gso_size, hlen; + + /* Maybe we should/could use sk->sk_prot->max_header here ? */ + hlen = inet_csk(sk)->icsk_af_ops->net_header_len + + inet_csk(sk)->icsk_ext_hdr_len + + tp->tcp_header_len; + + /* Goal is to send at least one packet per ms, + * not one big TSO packet every 100 ms. + * This preserves ACK clocking and is consistent + * with tcp_tso_should_defer() heuristic. + */ + gso_size = sk->sk_pacing_rate / (2 * MSEC_PER_SEC); + gso_size = max_t(u32, gso_size, + sysctl_tcp_min_tso_segs * mss_now); + + xmit_size_goal = min_t(u32, gso_size, + sk->sk_gso_max_size - 1 - hlen); - /* TSQ : try to have two TSO segments in flight */ + /* TSQ : try to have at least two segments in flight + * (one in NIC TX ring, another in Qdisc) + */ xmit_size_goal = min_t(u32, xmit_size_goal, sysctl_tcp_limit_output_bytes >> 1); diff --git a/net/ipv4/tcp_input.c b/net/ipv4/tcp_input.c index ec492eae0cd7..1a84fffe6993 100644 --- a/net/ipv4/tcp_input.c +++ b/net/ipv4/tcp_input.c @@ -688,6 +688,34 @@ static void tcp_rtt_estimator(struct sock *sk, const __u32 mrtt) } } +/* Set the sk_pacing_rate to allow proper sizing of TSO packets. + * Note: TCP stack does not yet implement pacing. + * FQ packet scheduler can be used to implement cheap but effective + * TCP pacing, to smooth the burst on large writes when packets + * in flight is significantly lower than cwnd (or rwin) + */ +static void tcp_update_pacing_rate(struct sock *sk) +{ + const struct tcp_sock *tp = tcp_sk(sk); + u64 rate; + + /* set sk_pacing_rate to 200 % of current rate (mss * cwnd / srtt) */ + rate = (u64)tp->mss_cache * 2 * (HZ << 3); + + rate *= max(tp->snd_cwnd, tp->packets_out); + + /* Correction for small srtt : minimum srtt being 8 (1 jiffy << 3), + * be conservative and assume srtt = 1 (125 us instead of 1.25 ms) + * We probably need usec resolution in the future. + * Note: This also takes care of possible srtt=0 case, + * when tcp_rtt_estimator() was not yet called. + */ + if (tp->srtt > 8 + 2) + do_div(rate, tp->srtt); + + sk->sk_pacing_rate = min_t(u64, rate, ~0U); +} + /* Calculate rto without backoff. This is the second half of Van Jacobson's * routine referred to above. */ @@ -3278,7 +3306,7 @@ static int tcp_ack(struct sock *sk, const struct sk_buff *skb, int flag) u32 ack_seq = TCP_SKB_CB(skb)->seq; u32 ack = TCP_SKB_CB(skb)->ack_seq; bool is_dupack = false; - u32 prior_in_flight; + u32 prior_in_flight, prior_cwnd = tp->snd_cwnd, prior_rtt = tp->srtt; u32 prior_fackets; int prior_packets = tp->packets_out; const int prior_unsacked = tp->packets_out - tp->sacked_out; @@ -3383,6 +3411,8 @@ static int tcp_ack(struct sock *sk, const struct sk_buff *skb, int flag) if (icsk->icsk_pending == ICSK_TIME_RETRANS) tcp_schedule_loss_probe(sk); + if (tp->srtt != prior_rtt || tp->snd_cwnd != prior_cwnd) + tcp_update_pacing_rate(sk); return 1; no_queue: diff --git a/net/ipv4/tcp_output.c b/net/ipv4/tcp_output.c index 884efff5b531..e63ae4c9691d 100644 --- a/net/ipv4/tcp_output.c +++ b/net/ipv4/tcp_output.c @@ -1631,7 +1631,7 @@ static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb) /* If a full-sized TSO skb can be sent, do it. */ if (limit >= min_t(unsigned int, sk->sk_gso_max_size, - sk->sk_gso_max_segs * tp->mss_cache)) + tp->xmit_size_goal_segs * tp->mss_cache)) goto send_now; /* Middle in queue won't get any more data, full sendable already? */ -- cgit v1.2.3-59-g8ed1b From c995ae2259ee36caf48bbfacf40111998dacd4af Mon Sep 17 00:00:00 2001 From: Vijay Subramanian Date: Tue, 3 Sep 2013 12:23:22 -0700 Subject: tcp: Change return value of tcp_rcv_established() tcp_rcv_established() returns only one value namely 0. We change the return value to void (as suggested by David Miller). After commit 0c24604b (tcp: implement RFC 5961 4.2), we no longer send RSTs in response to SYNs. We can remove the check and processing on the return value of tcp_rcv_established(). We also fix jtcp_rcv_established() in tcp_probe.c to match that of tcp_rcv_established(). Signed-off-by: Vijay Subramanian Signed-off-by: David S. Miller --- include/net/tcp.h | 4 ++-- net/ipv4/tcp_input.c | 13 ++++++------- net/ipv4/tcp_ipv4.c | 5 +---- net/ipv4/tcp_probe.c | 5 ++--- net/ipv6/tcp_ipv6.c | 3 +-- 5 files changed, 12 insertions(+), 18 deletions(-) (limited to 'net/ipv4/tcp_input.c') diff --git a/include/net/tcp.h b/include/net/tcp.h index 6a6a88db462d..b1aa324c5e65 100644 --- a/include/net/tcp.h +++ b/include/net/tcp.h @@ -371,8 +371,8 @@ extern void tcp_delack_timer_handler(struct sock *sk); extern int tcp_ioctl(struct sock *sk, int cmd, unsigned long arg); extern int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb, const struct tcphdr *th, unsigned int len); -extern int tcp_rcv_established(struct sock *sk, struct sk_buff *skb, - const struct tcphdr *th, unsigned int len); +extern void tcp_rcv_established(struct sock *sk, struct sk_buff *skb, + const struct tcphdr *th, unsigned int len); extern void tcp_rcv_space_adjust(struct sock *sk); extern void tcp_cleanup_rbuf(struct sock *sk, int copied); extern int tcp_twsk_unique(struct sock *sk, struct sock *sktw, void *twp); diff --git a/net/ipv4/tcp_input.c b/net/ipv4/tcp_input.c index 1a84fffe6993..93d7e9de4143 100644 --- a/net/ipv4/tcp_input.c +++ b/net/ipv4/tcp_input.c @@ -5049,8 +5049,8 @@ discard: * the rest is checked inline. Fast processing is turned on in * tcp_data_queue when everything is OK. */ -int tcp_rcv_established(struct sock *sk, struct sk_buff *skb, - const struct tcphdr *th, unsigned int len) +void tcp_rcv_established(struct sock *sk, struct sk_buff *skb, + const struct tcphdr *th, unsigned int len) { struct tcp_sock *tp = tcp_sk(sk); @@ -5127,7 +5127,7 @@ int tcp_rcv_established(struct sock *sk, struct sk_buff *skb, tcp_ack(sk, skb, 0); __kfree_skb(skb); tcp_data_snd_check(sk); - return 0; + return; } else { /* Header too small */ TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS); goto discard; @@ -5220,7 +5220,7 @@ no_ack: if (eaten) kfree_skb_partial(skb, fragstolen); sk->sk_data_ready(sk, 0); - return 0; + return; } } @@ -5236,7 +5236,7 @@ slow_path: */ if (!tcp_validate_incoming(sk, skb, th, 1)) - return 0; + return; step5: if (tcp_ack(sk, skb, FLAG_SLOWPATH | FLAG_UPDATE_TS_RECENT) < 0) @@ -5252,7 +5252,7 @@ step5: tcp_data_snd_check(sk); tcp_ack_snd_check(sk); - return 0; + return; csum_error: TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_CSUMERRORS); @@ -5260,7 +5260,6 @@ csum_error: discard: __kfree_skb(skb); - return 0; } EXPORT_SYMBOL(tcp_rcv_established); diff --git a/net/ipv4/tcp_ipv4.c b/net/ipv4/tcp_ipv4.c index 09d45d718973..b14266bb91eb 100644 --- a/net/ipv4/tcp_ipv4.c +++ b/net/ipv4/tcp_ipv4.c @@ -1799,10 +1799,7 @@ int tcp_v4_do_rcv(struct sock *sk, struct sk_buff *skb) sk->sk_rx_dst = NULL; } } - if (tcp_rcv_established(sk, skb, tcp_hdr(skb), skb->len)) { - rsk = sk; - goto reset; - } + tcp_rcv_established(sk, skb, tcp_hdr(skb), skb->len); return 0; } diff --git a/net/ipv4/tcp_probe.c b/net/ipv4/tcp_probe.c index 1f6aa543b64e..611beab38a00 100644 --- a/net/ipv4/tcp_probe.c +++ b/net/ipv4/tcp_probe.c @@ -122,8 +122,8 @@ static inline int tcp_probe_avail(void) * Hook inserted to be called before each receive packet. * Note: arguments must match tcp_rcv_established()! */ -static int jtcp_rcv_established(struct sock *sk, struct sk_buff *skb, - const struct tcphdr *th, unsigned int len) +static void jtcp_rcv_established(struct sock *sk, struct sk_buff *skb, + const struct tcphdr *th, unsigned int len) { const struct tcp_sock *tp = tcp_sk(sk); const struct inet_sock *inet = inet_sk(sk); @@ -172,7 +172,6 @@ static int jtcp_rcv_established(struct sock *sk, struct sk_buff *skb, } jprobe_return(); - return 0; } static struct jprobe tcp_jprobe = { diff --git a/net/ipv6/tcp_ipv6.c b/net/ipv6/tcp_ipv6.c index 5bcfadf09e95..9acdcedf9a14 100644 --- a/net/ipv6/tcp_ipv6.c +++ b/net/ipv6/tcp_ipv6.c @@ -1360,8 +1360,7 @@ static int tcp_v6_do_rcv(struct sock *sk, struct sk_buff *skb) } } - if (tcp_rcv_established(sk, skb, tcp_hdr(skb), skb->len)) - goto reset; + tcp_rcv_established(sk, skb, tcp_hdr(skb), skb->len); if (opt_skb) goto ipv6_pktoptions; return 0; -- cgit v1.2.3-59-g8ed1b From 16edfe7ee02dd7fcc13855ea19158333529533b2 Mon Sep 17 00:00:00 2001 From: Yuchung Cheng Date: Thu, 5 Sep 2013 15:36:09 -0700 Subject: tcp: fix no cwnd growth after timeout In commit 0f7cc9a3 "tcp: increase throughput when reordering is high", it only allows cwnd to increase in Open state. This mistakenly disables slow start after timeout (CA_Loss). Moreover cwnd won't grow if the state moves from Disorder to Open later in tcp_fastretrans_alert(). Therefore the correct logic should be to allow cwnd to grow as long as the data is received in order in Open, Loss, or even Disorder state. Signed-off-by: Yuchung Cheng Acked-by: Neal Cardwell Signed-off-by: David S. Miller --- net/ipv4/tcp_input.c | 6 ++---- 1 file changed, 2 insertions(+), 4 deletions(-) (limited to 'net/ipv4/tcp_input.c') diff --git a/net/ipv4/tcp_input.c b/net/ipv4/tcp_input.c index 1969e16d936d..894bc174f472 100644 --- a/net/ipv4/tcp_input.c +++ b/net/ipv4/tcp_input.c @@ -3162,16 +3162,14 @@ static inline bool tcp_may_raise_cwnd(const struct sock *sk, const int flag) /* If reordering is high then always grow cwnd whenever data is * delivered regardless of its ordering. Otherwise stay conservative - * and only grow cwnd on in-order delivery in Open state, and retain - * cwnd in Disordered state (RFC5681). A stretched ACK with + * and only grow cwnd on in-order delivery (RFC5681). A stretched ACK w/ * new SACK or ECE mark may first advance cwnd here and later reduce * cwnd in tcp_fastretrans_alert() based on more states. */ if (tcp_sk(sk)->reordering > sysctl_tcp_reordering) return flag & FLAG_FORWARD_PROGRESS; - return inet_csk(sk)->icsk_ca_state == TCP_CA_Open && - flag & FLAG_DATA_ACKED; + return flag & FLAG_DATA_ACKED; } /* Check that window update is acceptable. -- cgit v1.2.3-59-g8ed1b From 4e4f1fc226816905c937f9b29dabe351075dfe0f Mon Sep 17 00:00:00 2001 From: Eric Dumazet Date: Fri, 6 Sep 2013 10:35:58 -0700 Subject: tcp: properly increase rcv_ssthresh for ofo packets TCP receive window handling is multi staged. A socket has a memory budget, static or dynamic, in sk_rcvbuf. Because we do not really know how this memory budget translates to a TCP window (payload), TCP announces a small initial window (about 20 MSS). When a packet is received, we increase TCP rcv_win depending on the payload/truesize ratio of this packet. Good citizen packets give a hint that it's reasonable to have rcv_win = sk_rcvbuf/2 This heuristic takes place in tcp_grow_window() Problem is : We currently call tcp_grow_window() only for in-order packets. This means that reorders or packet losses stop proper grow of rcv_win, and senders are unable to benefit from fast recovery, or proper reordering level detection. Really, a packet being stored in OFO queue is not a bad citizen. It should be part of the game as in-order packets. In our traces, we very often see sender is limited by linux small receive windows, even if linux hosts use autotuning (DRS) and should allow rcv_win to grow to ~3MB. Signed-off-by: Eric Dumazet Acked-by: Neal Cardwell Signed-off-by: David S. Miller --- net/ipv4/tcp_input.c | 5 ++++- 1 file changed, 4 insertions(+), 1 deletion(-) (limited to 'net/ipv4/tcp_input.c') diff --git a/net/ipv4/tcp_input.c b/net/ipv4/tcp_input.c index 894bc174f472..25a89eaa669d 100644 --- a/net/ipv4/tcp_input.c +++ b/net/ipv4/tcp_input.c @@ -4139,6 +4139,7 @@ static void tcp_data_queue_ofo(struct sock *sk, struct sk_buff *skb) if (!tcp_try_coalesce(sk, skb1, skb, &fragstolen)) { __skb_queue_after(&tp->out_of_order_queue, skb1, skb); } else { + tcp_grow_window(sk, skb); kfree_skb_partial(skb, fragstolen); skb = NULL; } @@ -4214,8 +4215,10 @@ add_sack: if (tcp_is_sack(tp)) tcp_sack_new_ofo_skb(sk, seq, end_seq); end: - if (skb) + if (skb) { + tcp_grow_window(sk, skb); skb_set_owner_r(skb, sk); + } } static int __must_check tcp_queue_rcv(struct sock *sk, struct sk_buff *skb, int hdrlen, -- cgit v1.2.3-59-g8ed1b