1457c8996SThomas Gleixner // SPDX-License-Identifier: GPL-2.0-only 21da177e4SLinus Torvalds /* 31da177e4SLinus Torvalds * INET An implementation of the TCP/IP protocol suite for the LINUX 41da177e4SLinus Torvalds * operating system. INET is implemented using the BSD Socket 51da177e4SLinus Torvalds * interface as the means of communication with the user level. 61da177e4SLinus Torvalds * 71da177e4SLinus Torvalds * Implementation of the Transmission Control Protocol(TCP). 81da177e4SLinus Torvalds * 902c30a84SJesper Juhl * Authors: Ross Biro 101da177e4SLinus Torvalds * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG> 111da177e4SLinus Torvalds * Mark Evans, <evansmp@uhura.aston.ac.uk> 121da177e4SLinus Torvalds * Corey Minyard <wf-rch!minyard@relay.EU.net> 131da177e4SLinus Torvalds * Florian La Roche, <flla@stud.uni-sb.de> 141da177e4SLinus Torvalds * Charles Hedrick, <hedrick@klinzhai.rutgers.edu> 151da177e4SLinus Torvalds * Linus Torvalds, <torvalds@cs.helsinki.fi> 161da177e4SLinus Torvalds * Alan Cox, <gw4pts@gw4pts.ampr.org> 171da177e4SLinus Torvalds * Matthew Dillon, <dillon@apollo.west.oic.com> 181da177e4SLinus Torvalds * Arnt Gulbrandsen, <agulbra@nvg.unit.no> 191da177e4SLinus Torvalds * Jorge Cwik, <jorge@laser.satlink.net> 201da177e4SLinus Torvalds */ 211da177e4SLinus Torvalds 221da177e4SLinus Torvalds /* 231da177e4SLinus Torvalds * Changes: Pedro Roque : Retransmit queue handled by TCP. 241da177e4SLinus Torvalds * : Fragmentation on mtu decrease 251da177e4SLinus Torvalds * : Segment collapse on retransmit 261da177e4SLinus Torvalds * : AF independence 271da177e4SLinus Torvalds * 281da177e4SLinus Torvalds * Linus Torvalds : send_delayed_ack 291da177e4SLinus Torvalds * David S. Miller : Charge memory using the right skb 301da177e4SLinus Torvalds * during syn/ack processing. 311da177e4SLinus Torvalds * David S. Miller : Output engine completely rewritten. 321da177e4SLinus Torvalds * Andrea Arcangeli: SYNACK carry ts_recent in tsecr. 331da177e4SLinus Torvalds * Cacophonix Gaul : draft-minshall-nagle-01 341da177e4SLinus Torvalds * J Hadi Salim : ECN support 351da177e4SLinus Torvalds * 361da177e4SLinus Torvalds */ 371da177e4SLinus Torvalds 3891df42beSJoe Perches #define pr_fmt(fmt) "TCP: " fmt 3991df42beSJoe Perches 401da177e4SLinus Torvalds #include <net/tcp.h> 411da177e4SLinus Torvalds 421da177e4SLinus Torvalds #include <linux/compiler.h> 435a0e3ad6STejun Heo #include <linux/gfp.h> 441da177e4SLinus Torvalds #include <linux/module.h> 4560e2a778SUrsula Braun #include <linux/static_key.h> 461da177e4SLinus Torvalds 47e086101bSCong Wang #include <trace/events/tcp.h> 4835089bb2SDavid S. Miller 499799ccb0SEric Dumazet /* Refresh clocks of a TCP socket, 509799ccb0SEric Dumazet * ensuring monotically increasing values. 519799ccb0SEric Dumazet */ 529799ccb0SEric Dumazet void tcp_mstamp_refresh(struct tcp_sock *tp) 539799ccb0SEric Dumazet { 549799ccb0SEric Dumazet u64 val = tcp_clock_ns(); 559799ccb0SEric Dumazet 565f6188a8SEric Dumazet tp->tcp_clock_cache = val; 57e6d14070SEric Dumazet tp->tcp_mstamp = div_u64(val, NSEC_PER_USEC); 589799ccb0SEric Dumazet } 599799ccb0SEric Dumazet 6046d3ceabSEric Dumazet static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle, 6146d3ceabSEric Dumazet int push_one, gfp_t gfp); 62519855c5SWilliam Allen Simpson 6367edfef7SAndi Kleen /* Account for new data that has been sent to the network. */ 6475c119afSEric Dumazet static void tcp_event_new_data_sent(struct sock *sk, struct sk_buff *skb) 656ff03ac3SIlpo Järvinen { 666ba8a3b1SNandita Dukkipati struct inet_connection_sock *icsk = inet_csk(sk); 676ff03ac3SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 6866f5fe62SIlpo Järvinen unsigned int prior_packets = tp->packets_out; 699e412ba7SIlpo Järvinen 701da177e4SLinus Torvalds tp->snd_nxt = TCP_SKB_CB(skb)->end_seq; 718512430eSIlpo Järvinen 7275c119afSEric Dumazet __skb_unlink(skb, &sk->sk_write_queue); 7375c119afSEric Dumazet tcp_rbtree_insert(&sk->tcp_rtx_queue, skb); 7475c119afSEric Dumazet 7566f5fe62SIlpo Järvinen tp->packets_out += tcp_skb_pcount(skb); 76bec41a11SYuchung Cheng if (!prior_packets || icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) 77750ea2baSYuchung Cheng tcp_rearm_rto(sk); 78f19c29e3SYuchung Cheng 79f7324acdSDavid S. Miller NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT, 80f19c29e3SYuchung Cheng tcp_skb_pcount(skb)); 816a5dc9e5SEric Dumazet } 821da177e4SLinus Torvalds 83a4ecb15aSCui, Cheng /* SND.NXT, if window was not shrunk or the amount of shrunk was less than one 84a4ecb15aSCui, Cheng * window scaling factor due to loss of precision. 851da177e4SLinus Torvalds * If window has been shrunk, what should we make? It is not clear at all. 861da177e4SLinus Torvalds * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-( 871da177e4SLinus Torvalds * Anything in between SND.UNA...SND.UNA+SND.WND also can be already 881da177e4SLinus Torvalds * invalid. OK, let's make this for now: 891da177e4SLinus Torvalds */ 90cf533ea5SEric Dumazet static inline __u32 tcp_acceptable_seq(const struct sock *sk) 911da177e4SLinus Torvalds { 92cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 939e412ba7SIlpo Järvinen 94a4ecb15aSCui, Cheng if (!before(tcp_wnd_end(tp), tp->snd_nxt) || 95a4ecb15aSCui, Cheng (tp->rx_opt.wscale_ok && 96a4ecb15aSCui, Cheng ((tp->snd_nxt - tcp_wnd_end(tp)) < (1 << tp->rx_opt.rcv_wscale)))) 971da177e4SLinus Torvalds return tp->snd_nxt; 981da177e4SLinus Torvalds else 9990840defSIlpo Järvinen return tcp_wnd_end(tp); 1001da177e4SLinus Torvalds } 1011da177e4SLinus Torvalds 1021da177e4SLinus Torvalds /* Calculate mss to advertise in SYN segment. 1031da177e4SLinus Torvalds * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that: 1041da177e4SLinus Torvalds * 1051da177e4SLinus Torvalds * 1. It is independent of path mtu. 1061da177e4SLinus Torvalds * 2. Ideally, it is maximal possible segment size i.e. 65535-40. 1071da177e4SLinus Torvalds * 3. For IPv4 it is reasonable to calculate it from maximal MTU of 1081da177e4SLinus Torvalds * attached devices, because some buggy hosts are confused by 1091da177e4SLinus Torvalds * large MSS. 1101da177e4SLinus Torvalds * 4. We do not make 3, we advertise MSS, calculated from first 1111da177e4SLinus Torvalds * hop device mtu, but allow to raise it to ip_rt_min_advmss. 1121da177e4SLinus Torvalds * This may be overridden via information stored in routing table. 1131da177e4SLinus Torvalds * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible, 1141da177e4SLinus Torvalds * probably even Jumbo". 1151da177e4SLinus Torvalds */ 1161da177e4SLinus Torvalds static __u16 tcp_advertise_mss(struct sock *sk) 1171da177e4SLinus Torvalds { 1181da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 119cf533ea5SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 1201da177e4SLinus Torvalds int mss = tp->advmss; 1211da177e4SLinus Torvalds 1220dbaee3bSDavid S. Miller if (dst) { 1230dbaee3bSDavid S. Miller unsigned int metric = dst_metric_advmss(dst); 1240dbaee3bSDavid S. Miller 1250dbaee3bSDavid S. Miller if (metric < mss) { 1260dbaee3bSDavid S. Miller mss = metric; 1271da177e4SLinus Torvalds tp->advmss = mss; 1281da177e4SLinus Torvalds } 1290dbaee3bSDavid S. Miller } 1301da177e4SLinus Torvalds 1311da177e4SLinus Torvalds return (__u16)mss; 1321da177e4SLinus Torvalds } 1331da177e4SLinus Torvalds 1341da177e4SLinus Torvalds /* RFC2861. Reset CWND after idle period longer RTO to "restart window". 1356f021c62SEric Dumazet * This is the first part of cwnd validation mechanism. 1366f021c62SEric Dumazet */ 1376f021c62SEric Dumazet void tcp_cwnd_restart(struct sock *sk, s32 delta) 1381da177e4SLinus Torvalds { 139463c84b9SArnaldo Carvalho de Melo struct tcp_sock *tp = tcp_sk(sk); 1406f021c62SEric Dumazet u32 restart_cwnd = tcp_init_cwnd(tp, __sk_dst_get(sk)); 1411da177e4SLinus Torvalds u32 cwnd = tp->snd_cwnd; 1421da177e4SLinus Torvalds 1436687e988SArnaldo Carvalho de Melo tcp_ca_event(sk, CA_EVENT_CWND_RESTART); 1441da177e4SLinus Torvalds 1456687e988SArnaldo Carvalho de Melo tp->snd_ssthresh = tcp_current_ssthresh(sk); 1461da177e4SLinus Torvalds restart_cwnd = min(restart_cwnd, cwnd); 1471da177e4SLinus Torvalds 148463c84b9SArnaldo Carvalho de Melo while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd) 1491da177e4SLinus Torvalds cwnd >>= 1; 1501da177e4SLinus Torvalds tp->snd_cwnd = max(cwnd, restart_cwnd); 151c2203cf7SEric Dumazet tp->snd_cwnd_stamp = tcp_jiffies32; 1521da177e4SLinus Torvalds tp->snd_cwnd_used = 0; 1531da177e4SLinus Torvalds } 1541da177e4SLinus Torvalds 15567edfef7SAndi Kleen /* Congestion state accounting after a packet has been sent. */ 15640efc6faSStephen Hemminger static void tcp_event_data_sent(struct tcp_sock *tp, 157cf533ea5SEric Dumazet struct sock *sk) 1581da177e4SLinus Torvalds { 159463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 160d635fbe2SEric Dumazet const u32 now = tcp_jiffies32; 1611da177e4SLinus Torvalds 16205c5a46dSNeal Cardwell if (tcp_packets_in_flight(tp) == 0) 16305c5a46dSNeal Cardwell tcp_ca_event(sk, CA_EVENT_TX_START); 16405c5a46dSNeal Cardwell 1654a41f453SWei Wang /* If this is the first data packet sent in response to the 1664a41f453SWei Wang * previous received data, 1674a41f453SWei Wang * and it is a reply for ato after last received packet, 1684a41f453SWei Wang * increase pingpong count. 1691da177e4SLinus Torvalds */ 1704a41f453SWei Wang if (before(tp->lsndtime, icsk->icsk_ack.lrcvtime) && 1714a41f453SWei Wang (u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato) 1724a41f453SWei Wang inet_csk_inc_pingpong_cnt(sk); 1734a41f453SWei Wang 1744a41f453SWei Wang tp->lsndtime = now; 1751da177e4SLinus Torvalds } 1761da177e4SLinus Torvalds 17767edfef7SAndi Kleen /* Account for an ACK we sent. */ 17827cde44aSYuchung Cheng static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts, 17927cde44aSYuchung Cheng u32 rcv_nxt) 1801da177e4SLinus Torvalds { 1815d9f4262SEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 1825d9f4262SEric Dumazet 18386de5921SEric Dumazet if (unlikely(tp->compressed_ack > TCP_FASTRETRANS_THRESH)) { 184200d95f4SEric Dumazet NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPACKCOMPRESSED, 18586de5921SEric Dumazet tp->compressed_ack - TCP_FASTRETRANS_THRESH); 18686de5921SEric Dumazet tp->compressed_ack = TCP_FASTRETRANS_THRESH; 1875d9f4262SEric Dumazet if (hrtimer_try_to_cancel(&tp->compressed_ack_timer) == 1) 1885d9f4262SEric Dumazet __sock_put(sk); 1895d9f4262SEric Dumazet } 19027cde44aSYuchung Cheng 19127cde44aSYuchung Cheng if (unlikely(rcv_nxt != tp->rcv_nxt)) 19227cde44aSYuchung Cheng return; /* Special ACK sent by DCTCP to reflect ECN */ 193463c84b9SArnaldo Carvalho de Melo tcp_dec_quickack_mode(sk, pkts); 194463c84b9SArnaldo Carvalho de Melo inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK); 1951da177e4SLinus Torvalds } 1961da177e4SLinus Torvalds 1971da177e4SLinus Torvalds /* Determine a window scaling and initial window to offer. 1981da177e4SLinus Torvalds * Based on the assumption that the given amount of space 1991da177e4SLinus Torvalds * will be offered. Store the results in the tp structure. 2001da177e4SLinus Torvalds * NOTE: for smooth operation initial space offering should 2011da177e4SLinus Torvalds * be a multiple of mss if possible. We assume here that mss >= 1. 2021da177e4SLinus Torvalds * This MUST be enforced by all callers. 2031da177e4SLinus Torvalds */ 204ceef9ab6SEric Dumazet void tcp_select_initial_window(const struct sock *sk, int __space, __u32 mss, 2051da177e4SLinus Torvalds __u32 *rcv_wnd, __u32 *window_clamp, 20631d12926Slaurent chavey int wscale_ok, __u8 *rcv_wscale, 20731d12926Slaurent chavey __u32 init_rcv_wnd) 2081da177e4SLinus Torvalds { 2091da177e4SLinus Torvalds unsigned int space = (__space < 0 ? 0 : __space); 2101da177e4SLinus Torvalds 2111da177e4SLinus Torvalds /* If no clamp set the clamp to the max possible scaled window */ 2121da177e4SLinus Torvalds if (*window_clamp == 0) 213589c49cbSGao Feng (*window_clamp) = (U16_MAX << TCP_MAX_WSCALE); 2141da177e4SLinus Torvalds space = min(*window_clamp, space); 2151da177e4SLinus Torvalds 2161da177e4SLinus Torvalds /* Quantize space offering to a multiple of mss if possible. */ 2171da177e4SLinus Torvalds if (space > mss) 218589c49cbSGao Feng space = rounddown(space, mss); 2191da177e4SLinus Torvalds 2201da177e4SLinus Torvalds /* NOTE: offering an initial window larger than 32767 22115d99e02SRick Jones * will break some buggy TCP stacks. If the admin tells us 22215d99e02SRick Jones * it is likely we could be speaking with such a buggy stack 22315d99e02SRick Jones * we will truncate our initial window offering to 32K-1 22415d99e02SRick Jones * unless the remote has sent us a window scaling option, 22515d99e02SRick Jones * which we interpret as a sign the remote TCP is not 22615d99e02SRick Jones * misinterpreting the window field as a signed quantity. 2271da177e4SLinus Torvalds */ 228ceef9ab6SEric Dumazet if (sock_net(sk)->ipv4.sysctl_tcp_workaround_signed_windows) 2291da177e4SLinus Torvalds (*rcv_wnd) = min(space, MAX_TCP_WINDOW); 23015d99e02SRick Jones else 231a337531bSYuchung Cheng (*rcv_wnd) = min_t(u32, space, U16_MAX); 232a337531bSYuchung Cheng 233a337531bSYuchung Cheng if (init_rcv_wnd) 234a337531bSYuchung Cheng *rcv_wnd = min(*rcv_wnd, init_rcv_wnd * mss); 23515d99e02SRick Jones 23619bf6261SEric Dumazet *rcv_wscale = 0; 2371da177e4SLinus Torvalds if (wscale_ok) { 238589c49cbSGao Feng /* Set window scaling on max possible window */ 239356d1833SEric Dumazet space = max_t(u32, space, sock_net(sk)->ipv4.sysctl_tcp_rmem[2]); 240f626300aSSoheil Hassas Yeganeh space = max_t(u32, space, sysctl_rmem_max); 241316c1592SStephen Hemminger space = min_t(u32, space, *window_clamp); 24219bf6261SEric Dumazet *rcv_wscale = clamp_t(int, ilog2(space) - 15, 24319bf6261SEric Dumazet 0, TCP_MAX_WSCALE); 2441da177e4SLinus Torvalds } 2451da177e4SLinus Torvalds /* Set the clamp no higher than max representable value */ 246589c49cbSGao Feng (*window_clamp) = min_t(__u32, U16_MAX << (*rcv_wscale), *window_clamp); 2471da177e4SLinus Torvalds } 2484bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_select_initial_window); 2491da177e4SLinus Torvalds 2501da177e4SLinus Torvalds /* Chose a new window to advertise, update state in tcp_sock for the 2511da177e4SLinus Torvalds * socket, and return result with RFC1323 scaling applied. The return 2521da177e4SLinus Torvalds * value can be stuffed directly into th->window for an outgoing 2531da177e4SLinus Torvalds * frame. 2541da177e4SLinus Torvalds */ 25540efc6faSStephen Hemminger static u16 tcp_select_window(struct sock *sk) 2561da177e4SLinus Torvalds { 2571da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 2588e165e20SFlorian Westphal u32 old_win = tp->rcv_wnd; 2591da177e4SLinus Torvalds u32 cur_win = tcp_receive_window(tp); 2601da177e4SLinus Torvalds u32 new_win = __tcp_select_window(sk); 2611da177e4SLinus Torvalds 2621da177e4SLinus Torvalds /* Never shrink the offered window */ 2631da177e4SLinus Torvalds if (new_win < cur_win) { 2641da177e4SLinus Torvalds /* Danger Will Robinson! 2651da177e4SLinus Torvalds * Don't update rcv_wup/rcv_wnd here or else 2661da177e4SLinus Torvalds * we will not be able to advertise a zero 2671da177e4SLinus Torvalds * window in time. --DaveM 2681da177e4SLinus Torvalds * 2691da177e4SLinus Torvalds * Relax Will Robinson. 2701da177e4SLinus Torvalds */ 2718e165e20SFlorian Westphal if (new_win == 0) 2728e165e20SFlorian Westphal NET_INC_STATS(sock_net(sk), 2738e165e20SFlorian Westphal LINUX_MIB_TCPWANTZEROWINDOWADV); 274607bfbf2SPatrick McHardy new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale); 2751da177e4SLinus Torvalds } 2761da177e4SLinus Torvalds tp->rcv_wnd = new_win; 2771da177e4SLinus Torvalds tp->rcv_wup = tp->rcv_nxt; 2781da177e4SLinus Torvalds 2791da177e4SLinus Torvalds /* Make sure we do not exceed the maximum possible 2801da177e4SLinus Torvalds * scaled window. 2811da177e4SLinus Torvalds */ 282ceef9ab6SEric Dumazet if (!tp->rx_opt.rcv_wscale && 283ceef9ab6SEric Dumazet sock_net(sk)->ipv4.sysctl_tcp_workaround_signed_windows) 2841da177e4SLinus Torvalds new_win = min(new_win, MAX_TCP_WINDOW); 2851da177e4SLinus Torvalds else 2861da177e4SLinus Torvalds new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale)); 2871da177e4SLinus Torvalds 2881da177e4SLinus Torvalds /* RFC1323 scaling applied */ 2891da177e4SLinus Torvalds new_win >>= tp->rx_opt.rcv_wscale; 2901da177e4SLinus Torvalds 29131770e34SFlorian Westphal /* If we advertise zero window, disable fast path. */ 2928e165e20SFlorian Westphal if (new_win == 0) { 29331770e34SFlorian Westphal tp->pred_flags = 0; 2948e165e20SFlorian Westphal if (old_win) 2958e165e20SFlorian Westphal NET_INC_STATS(sock_net(sk), 2968e165e20SFlorian Westphal LINUX_MIB_TCPTOZEROWINDOWADV); 2978e165e20SFlorian Westphal } else if (old_win == 0) { 2988e165e20SFlorian Westphal NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFROMZEROWINDOWADV); 2998e165e20SFlorian Westphal } 3001da177e4SLinus Torvalds 3011da177e4SLinus Torvalds return new_win; 3021da177e4SLinus Torvalds } 3031da177e4SLinus Torvalds 30467edfef7SAndi Kleen /* Packet ECN state for a SYN-ACK */ 305735d3831SFlorian Westphal static void tcp_ecn_send_synack(struct sock *sk, struct sk_buff *skb) 306bdf1ee5dSIlpo Järvinen { 30730e502a3SDaniel Borkmann const struct tcp_sock *tp = tcp_sk(sk); 30830e502a3SDaniel Borkmann 3094de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_CWR; 310bdf1ee5dSIlpo Järvinen if (!(tp->ecn_flags & TCP_ECN_OK)) 3114de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_ECE; 31291b5b21cSLawrence Brakmo else if (tcp_ca_needs_ecn(sk) || 31391b5b21cSLawrence Brakmo tcp_bpf_ca_needs_ecn(sk)) 31430e502a3SDaniel Borkmann INET_ECN_xmit(sk); 315bdf1ee5dSIlpo Järvinen } 316bdf1ee5dSIlpo Järvinen 31767edfef7SAndi Kleen /* Packet ECN state for a SYN. */ 318735d3831SFlorian Westphal static void tcp_ecn_send_syn(struct sock *sk, struct sk_buff *skb) 319bdf1ee5dSIlpo Järvinen { 320bdf1ee5dSIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 32191b5b21cSLawrence Brakmo bool bpf_needs_ecn = tcp_bpf_ca_needs_ecn(sk); 322f7b3bec6SFlorian Westphal bool use_ecn = sock_net(sk)->ipv4.sysctl_tcp_ecn == 1 || 32391b5b21cSLawrence Brakmo tcp_ca_needs_ecn(sk) || bpf_needs_ecn; 324f7b3bec6SFlorian Westphal 325f7b3bec6SFlorian Westphal if (!use_ecn) { 326f7b3bec6SFlorian Westphal const struct dst_entry *dst = __sk_dst_get(sk); 327f7b3bec6SFlorian Westphal 328f7b3bec6SFlorian Westphal if (dst && dst_feature(dst, RTAX_FEATURE_ECN)) 329f7b3bec6SFlorian Westphal use_ecn = true; 330f7b3bec6SFlorian Westphal } 331bdf1ee5dSIlpo Järvinen 332bdf1ee5dSIlpo Järvinen tp->ecn_flags = 0; 333f7b3bec6SFlorian Westphal 334f7b3bec6SFlorian Westphal if (use_ecn) { 3354de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ECE | TCPHDR_CWR; 336bdf1ee5dSIlpo Järvinen tp->ecn_flags = TCP_ECN_OK; 33791b5b21cSLawrence Brakmo if (tcp_ca_needs_ecn(sk) || bpf_needs_ecn) 33830e502a3SDaniel Borkmann INET_ECN_xmit(sk); 339bdf1ee5dSIlpo Järvinen } 340bdf1ee5dSIlpo Järvinen } 341bdf1ee5dSIlpo Järvinen 34249213555SDaniel Borkmann static void tcp_ecn_clear_syn(struct sock *sk, struct sk_buff *skb) 34349213555SDaniel Borkmann { 34449213555SDaniel Borkmann if (sock_net(sk)->ipv4.sysctl_tcp_ecn_fallback) 34549213555SDaniel Borkmann /* tp->ecn_flags are cleared at a later point in time when 34649213555SDaniel Borkmann * SYN ACK is ultimatively being received. 34749213555SDaniel Borkmann */ 34849213555SDaniel Borkmann TCP_SKB_CB(skb)->tcp_flags &= ~(TCPHDR_ECE | TCPHDR_CWR); 34949213555SDaniel Borkmann } 35049213555SDaniel Borkmann 351735d3831SFlorian Westphal static void 3526ac705b1SEric Dumazet tcp_ecn_make_synack(const struct request_sock *req, struct tcphdr *th) 353bdf1ee5dSIlpo Järvinen { 3546ac705b1SEric Dumazet if (inet_rsk(req)->ecn_ok) 355bdf1ee5dSIlpo Järvinen th->ece = 1; 356bdf1ee5dSIlpo Järvinen } 357bdf1ee5dSIlpo Järvinen 35867edfef7SAndi Kleen /* Set up ECN state for a packet on a ESTABLISHED socket that is about to 35967edfef7SAndi Kleen * be sent. 36067edfef7SAndi Kleen */ 361735d3831SFlorian Westphal static void tcp_ecn_send(struct sock *sk, struct sk_buff *skb, 362ea1627c2SEric Dumazet struct tcphdr *th, int tcp_header_len) 363bdf1ee5dSIlpo Järvinen { 364bdf1ee5dSIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 365bdf1ee5dSIlpo Järvinen 366bdf1ee5dSIlpo Järvinen if (tp->ecn_flags & TCP_ECN_OK) { 367bdf1ee5dSIlpo Järvinen /* Not-retransmitted data segment: set ECT and inject CWR. */ 368bdf1ee5dSIlpo Järvinen if (skb->len != tcp_header_len && 369bdf1ee5dSIlpo Järvinen !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) { 370bdf1ee5dSIlpo Järvinen INET_ECN_xmit(sk); 371bdf1ee5dSIlpo Järvinen if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) { 372bdf1ee5dSIlpo Järvinen tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR; 373ea1627c2SEric Dumazet th->cwr = 1; 374bdf1ee5dSIlpo Järvinen skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN; 375bdf1ee5dSIlpo Järvinen } 37630e502a3SDaniel Borkmann } else if (!tcp_ca_needs_ecn(sk)) { 377bdf1ee5dSIlpo Järvinen /* ACK or retransmitted segment: clear ECT|CE */ 378bdf1ee5dSIlpo Järvinen INET_ECN_dontxmit(sk); 379bdf1ee5dSIlpo Järvinen } 380bdf1ee5dSIlpo Järvinen if (tp->ecn_flags & TCP_ECN_DEMAND_CWR) 381ea1627c2SEric Dumazet th->ece = 1; 382bdf1ee5dSIlpo Järvinen } 383bdf1ee5dSIlpo Järvinen } 384bdf1ee5dSIlpo Järvinen 385e870a8efSIlpo Järvinen /* Constructs common control bits of non-data skb. If SYN/FIN is present, 386e870a8efSIlpo Järvinen * auto increment end seqno. 387e870a8efSIlpo Järvinen */ 388e870a8efSIlpo Järvinen static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags) 389e870a8efSIlpo Järvinen { 3902e8e18efSDavid S. Miller skb->ip_summed = CHECKSUM_PARTIAL; 391e870a8efSIlpo Järvinen 3924de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags = flags; 393e870a8efSIlpo Järvinen TCP_SKB_CB(skb)->sacked = 0; 394e870a8efSIlpo Järvinen 395cd7d8498SEric Dumazet tcp_skb_pcount_set(skb, 1); 396e870a8efSIlpo Järvinen 397e870a8efSIlpo Järvinen TCP_SKB_CB(skb)->seq = seq; 398a3433f35SChangli Gao if (flags & (TCPHDR_SYN | TCPHDR_FIN)) 399e870a8efSIlpo Järvinen seq++; 400e870a8efSIlpo Järvinen TCP_SKB_CB(skb)->end_seq = seq; 401e870a8efSIlpo Järvinen } 402e870a8efSIlpo Järvinen 403a2a385d6SEric Dumazet static inline bool tcp_urg_mode(const struct tcp_sock *tp) 40433f5f57eSIlpo Järvinen { 40533f5f57eSIlpo Järvinen return tp->snd_una != tp->snd_up; 40633f5f57eSIlpo Järvinen } 40733f5f57eSIlpo Järvinen 40833ad798cSAdam Langley #define OPTION_SACK_ADVERTISE (1 << 0) 40933ad798cSAdam Langley #define OPTION_TS (1 << 1) 41033ad798cSAdam Langley #define OPTION_MD5 (1 << 2) 41189e95a61SOri Finkelman #define OPTION_WSCALE (1 << 3) 4122100c8d2SYuchung Cheng #define OPTION_FAST_OPEN_COOKIE (1 << 8) 41360e2a778SUrsula Braun #define OPTION_SMC (1 << 9) 41460e2a778SUrsula Braun 41560e2a778SUrsula Braun static void smc_options_write(__be32 *ptr, u16 *options) 41660e2a778SUrsula Braun { 41760e2a778SUrsula Braun #if IS_ENABLED(CONFIG_SMC) 41860e2a778SUrsula Braun if (static_branch_unlikely(&tcp_have_smc)) { 41960e2a778SUrsula Braun if (unlikely(OPTION_SMC & *options)) { 42060e2a778SUrsula Braun *ptr++ = htonl((TCPOPT_NOP << 24) | 42160e2a778SUrsula Braun (TCPOPT_NOP << 16) | 42260e2a778SUrsula Braun (TCPOPT_EXP << 8) | 42360e2a778SUrsula Braun (TCPOLEN_EXP_SMC_BASE)); 42460e2a778SUrsula Braun *ptr++ = htonl(TCPOPT_SMC_MAGIC); 42560e2a778SUrsula Braun } 42660e2a778SUrsula Braun } 42760e2a778SUrsula Braun #endif 42860e2a778SUrsula Braun } 42933ad798cSAdam Langley 43033ad798cSAdam Langley struct tcp_out_options { 4312100c8d2SYuchung Cheng u16 options; /* bit field of OPTION_* */ 4322100c8d2SYuchung Cheng u16 mss; /* 0 to disable */ 43333ad798cSAdam Langley u8 ws; /* window scale, 0 to disable */ 43433ad798cSAdam Langley u8 num_sack_blocks; /* number of SACK blocks to include */ 435bd0388aeSWilliam Allen Simpson u8 hash_size; /* bytes in hash_location */ 436bd0388aeSWilliam Allen Simpson __u8 *hash_location; /* temporary pointer, overloaded */ 4372100c8d2SYuchung Cheng __u32 tsval, tsecr; /* need to include OPTION_TS */ 4382100c8d2SYuchung Cheng struct tcp_fastopen_cookie *fastopen_cookie; /* Fast open cookie */ 43933ad798cSAdam Langley }; 44033ad798cSAdam Langley 44167edfef7SAndi Kleen /* Write previously computed TCP options to the packet. 44267edfef7SAndi Kleen * 44367edfef7SAndi Kleen * Beware: Something in the Internet is very sensitive to the ordering of 444fd6149d3SIlpo Järvinen * TCP options, we learned this through the hard way, so be careful here. 445fd6149d3SIlpo Järvinen * Luckily we can at least blame others for their non-compliance but from 4468e3bff96Sstephen hemminger * inter-operability perspective it seems that we're somewhat stuck with 447fd6149d3SIlpo Järvinen * the ordering which we have been using if we want to keep working with 448fd6149d3SIlpo Järvinen * those broken things (not that it currently hurts anybody as there isn't 449fd6149d3SIlpo Järvinen * particular reason why the ordering would need to be changed). 450fd6149d3SIlpo Järvinen * 451fd6149d3SIlpo Järvinen * At least SACK_PERM as the first option is known to lead to a disaster 452fd6149d3SIlpo Järvinen * (but it may well be that other scenarios fail similarly). 453fd6149d3SIlpo Järvinen */ 45433ad798cSAdam Langley static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp, 455bd0388aeSWilliam Allen Simpson struct tcp_out_options *opts) 456bd0388aeSWilliam Allen Simpson { 4572100c8d2SYuchung Cheng u16 options = opts->options; /* mungable copy */ 458bd0388aeSWilliam Allen Simpson 459bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_MD5 & options)) { 4601a2c6181SChristoph Paasch *ptr++ = htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) | 4611a2c6181SChristoph Paasch (TCPOPT_MD5SIG << 8) | TCPOLEN_MD5SIG); 462bd0388aeSWilliam Allen Simpson /* overload cookie hash location */ 463bd0388aeSWilliam Allen Simpson opts->hash_location = (__u8 *)ptr; 46433ad798cSAdam Langley ptr += 4; 46533ad798cSAdam Langley } 46633ad798cSAdam Langley 467fd6149d3SIlpo Järvinen if (unlikely(opts->mss)) { 468fd6149d3SIlpo Järvinen *ptr++ = htonl((TCPOPT_MSS << 24) | 469fd6149d3SIlpo Järvinen (TCPOLEN_MSS << 16) | 470fd6149d3SIlpo Järvinen opts->mss); 471fd6149d3SIlpo Järvinen } 472fd6149d3SIlpo Järvinen 473bd0388aeSWilliam Allen Simpson if (likely(OPTION_TS & options)) { 474bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_SACK_ADVERTISE & options)) { 47533ad798cSAdam Langley *ptr++ = htonl((TCPOPT_SACK_PERM << 24) | 47633ad798cSAdam Langley (TCPOLEN_SACK_PERM << 16) | 47733ad798cSAdam Langley (TCPOPT_TIMESTAMP << 8) | 47833ad798cSAdam Langley TCPOLEN_TIMESTAMP); 479bd0388aeSWilliam Allen Simpson options &= ~OPTION_SACK_ADVERTISE; 48033ad798cSAdam Langley } else { 481496c98dfSYOSHIFUJI Hideaki *ptr++ = htonl((TCPOPT_NOP << 24) | 48240efc6faSStephen Hemminger (TCPOPT_NOP << 16) | 48340efc6faSStephen Hemminger (TCPOPT_TIMESTAMP << 8) | 48440efc6faSStephen Hemminger TCPOLEN_TIMESTAMP); 48540efc6faSStephen Hemminger } 48633ad798cSAdam Langley *ptr++ = htonl(opts->tsval); 48733ad798cSAdam Langley *ptr++ = htonl(opts->tsecr); 48833ad798cSAdam Langley } 48933ad798cSAdam Langley 490bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_SACK_ADVERTISE & options)) { 49133ad798cSAdam Langley *ptr++ = htonl((TCPOPT_NOP << 24) | 49233ad798cSAdam Langley (TCPOPT_NOP << 16) | 49333ad798cSAdam Langley (TCPOPT_SACK_PERM << 8) | 49433ad798cSAdam Langley TCPOLEN_SACK_PERM); 49533ad798cSAdam Langley } 49633ad798cSAdam Langley 497bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_WSCALE & options)) { 49833ad798cSAdam Langley *ptr++ = htonl((TCPOPT_NOP << 24) | 49933ad798cSAdam Langley (TCPOPT_WINDOW << 16) | 50033ad798cSAdam Langley (TCPOLEN_WINDOW << 8) | 50133ad798cSAdam Langley opts->ws); 50233ad798cSAdam Langley } 50333ad798cSAdam Langley 50433ad798cSAdam Langley if (unlikely(opts->num_sack_blocks)) { 50533ad798cSAdam Langley struct tcp_sack_block *sp = tp->rx_opt.dsack ? 50633ad798cSAdam Langley tp->duplicate_sack : tp->selective_acks; 50740efc6faSStephen Hemminger int this_sack; 50840efc6faSStephen Hemminger 50940efc6faSStephen Hemminger *ptr++ = htonl((TCPOPT_NOP << 24) | 51040efc6faSStephen Hemminger (TCPOPT_NOP << 16) | 51140efc6faSStephen Hemminger (TCPOPT_SACK << 8) | 51233ad798cSAdam Langley (TCPOLEN_SACK_BASE + (opts->num_sack_blocks * 51340efc6faSStephen Hemminger TCPOLEN_SACK_PERBLOCK))); 5142de979bdSStephen Hemminger 51533ad798cSAdam Langley for (this_sack = 0; this_sack < opts->num_sack_blocks; 51633ad798cSAdam Langley ++this_sack) { 51740efc6faSStephen Hemminger *ptr++ = htonl(sp[this_sack].start_seq); 51840efc6faSStephen Hemminger *ptr++ = htonl(sp[this_sack].end_seq); 51940efc6faSStephen Hemminger } 5202de979bdSStephen Hemminger 52140efc6faSStephen Hemminger tp->rx_opt.dsack = 0; 52240efc6faSStephen Hemminger } 5232100c8d2SYuchung Cheng 5242100c8d2SYuchung Cheng if (unlikely(OPTION_FAST_OPEN_COOKIE & options)) { 5252100c8d2SYuchung Cheng struct tcp_fastopen_cookie *foc = opts->fastopen_cookie; 5267f9b838bSDaniel Lee u8 *p = (u8 *)ptr; 5277f9b838bSDaniel Lee u32 len; /* Fast Open option length */ 5282100c8d2SYuchung Cheng 5297f9b838bSDaniel Lee if (foc->exp) { 5307f9b838bSDaniel Lee len = TCPOLEN_EXP_FASTOPEN_BASE + foc->len; 5317f9b838bSDaniel Lee *ptr = htonl((TCPOPT_EXP << 24) | (len << 16) | 5322100c8d2SYuchung Cheng TCPOPT_FASTOPEN_MAGIC); 5337f9b838bSDaniel Lee p += TCPOLEN_EXP_FASTOPEN_BASE; 5347f9b838bSDaniel Lee } else { 5357f9b838bSDaniel Lee len = TCPOLEN_FASTOPEN_BASE + foc->len; 5367f9b838bSDaniel Lee *p++ = TCPOPT_FASTOPEN; 5377f9b838bSDaniel Lee *p++ = len; 5382100c8d2SYuchung Cheng } 5397f9b838bSDaniel Lee 5407f9b838bSDaniel Lee memcpy(p, foc->val, foc->len); 5417f9b838bSDaniel Lee if ((len & 3) == 2) { 5427f9b838bSDaniel Lee p[foc->len] = TCPOPT_NOP; 5437f9b838bSDaniel Lee p[foc->len + 1] = TCPOPT_NOP; 5447f9b838bSDaniel Lee } 5457f9b838bSDaniel Lee ptr += (len + 3) >> 2; 5462100c8d2SYuchung Cheng } 54760e2a778SUrsula Braun 54860e2a778SUrsula Braun smc_options_write(ptr, &options); 54960e2a778SUrsula Braun } 55060e2a778SUrsula Braun 55160e2a778SUrsula Braun static void smc_set_option(const struct tcp_sock *tp, 55260e2a778SUrsula Braun struct tcp_out_options *opts, 55360e2a778SUrsula Braun unsigned int *remaining) 55460e2a778SUrsula Braun { 55560e2a778SUrsula Braun #if IS_ENABLED(CONFIG_SMC) 55660e2a778SUrsula Braun if (static_branch_unlikely(&tcp_have_smc)) { 55760e2a778SUrsula Braun if (tp->syn_smc) { 55860e2a778SUrsula Braun if (*remaining >= TCPOLEN_EXP_SMC_BASE_ALIGNED) { 55960e2a778SUrsula Braun opts->options |= OPTION_SMC; 56060e2a778SUrsula Braun *remaining -= TCPOLEN_EXP_SMC_BASE_ALIGNED; 56160e2a778SUrsula Braun } 56260e2a778SUrsula Braun } 56360e2a778SUrsula Braun } 56460e2a778SUrsula Braun #endif 56560e2a778SUrsula Braun } 56660e2a778SUrsula Braun 56760e2a778SUrsula Braun static void smc_set_option_cond(const struct tcp_sock *tp, 56860e2a778SUrsula Braun const struct inet_request_sock *ireq, 56960e2a778SUrsula Braun struct tcp_out_options *opts, 57060e2a778SUrsula Braun unsigned int *remaining) 57160e2a778SUrsula Braun { 57260e2a778SUrsula Braun #if IS_ENABLED(CONFIG_SMC) 57360e2a778SUrsula Braun if (static_branch_unlikely(&tcp_have_smc)) { 57460e2a778SUrsula Braun if (tp->syn_smc && ireq->smc_ok) { 57560e2a778SUrsula Braun if (*remaining >= TCPOLEN_EXP_SMC_BASE_ALIGNED) { 57660e2a778SUrsula Braun opts->options |= OPTION_SMC; 57760e2a778SUrsula Braun *remaining -= TCPOLEN_EXP_SMC_BASE_ALIGNED; 57860e2a778SUrsula Braun } 57960e2a778SUrsula Braun } 58060e2a778SUrsula Braun } 58160e2a778SUrsula Braun #endif 58240efc6faSStephen Hemminger } 58340efc6faSStephen Hemminger 58467edfef7SAndi Kleen /* Compute TCP options for SYN packets. This is not the final 58567edfef7SAndi Kleen * network wire format yet. 58667edfef7SAndi Kleen */ 58795c96174SEric Dumazet static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb, 58833ad798cSAdam Langley struct tcp_out_options *opts, 589cf533ea5SEric Dumazet struct tcp_md5sig_key **md5) 590cf533ea5SEric Dumazet { 59133ad798cSAdam Langley struct tcp_sock *tp = tcp_sk(sk); 59295c96174SEric Dumazet unsigned int remaining = MAX_TCP_OPTION_SPACE; 593783237e8SYuchung Cheng struct tcp_fastopen_request *fastopen = tp->fastopen_req; 59433ad798cSAdam Langley 5958c2320e8SEric Dumazet *md5 = NULL; 596cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 597921f9a0fSEric Dumazet if (static_branch_unlikely(&tcp_md5_needed) && 5986015c71eSEric Dumazet rcu_access_pointer(tp->md5sig_info)) { 59933ad798cSAdam Langley *md5 = tp->af_specific->md5_lookup(sk, sk); 60033ad798cSAdam Langley if (*md5) { 60133ad798cSAdam Langley opts->options |= OPTION_MD5; 602bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_MD5SIG_ALIGNED; 603cfb6eeb4SYOSHIFUJI Hideaki } 6048c2320e8SEric Dumazet } 605cfb6eeb4SYOSHIFUJI Hideaki #endif 60633ad798cSAdam Langley 60733ad798cSAdam Langley /* We always get an MSS option. The option bytes which will be seen in 60833ad798cSAdam Langley * normal data packets should timestamps be used, must be in the MSS 60933ad798cSAdam Langley * advertised. But we subtract them from tp->mss_cache so that 61033ad798cSAdam Langley * calculations in tcp_sendmsg are simpler etc. So account for this 61133ad798cSAdam Langley * fact here if necessary. If we don't do this correctly, as a 61233ad798cSAdam Langley * receiver we won't recognize data packets as being full sized when we 61333ad798cSAdam Langley * should, and thus we won't abide by the delayed ACK rules correctly. 61433ad798cSAdam Langley * SACKs don't matter, we never delay an ACK when we have any of those 61533ad798cSAdam Langley * going out. */ 61633ad798cSAdam Langley opts->mss = tcp_advertise_mss(sk); 617bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_MSS_ALIGNED; 61833ad798cSAdam Langley 6195d2ed052SEric Dumazet if (likely(sock_net(sk)->ipv4.sysctl_tcp_timestamps && !*md5)) { 62033ad798cSAdam Langley opts->options |= OPTION_TS; 6217faee5c0SEric Dumazet opts->tsval = tcp_skb_timestamp(skb) + tp->tsoffset; 62233ad798cSAdam Langley opts->tsecr = tp->rx_opt.ts_recent; 623bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_TSTAMP_ALIGNED; 62433ad798cSAdam Langley } 6259bb37ef0SEric Dumazet if (likely(sock_net(sk)->ipv4.sysctl_tcp_window_scaling)) { 62633ad798cSAdam Langley opts->ws = tp->rx_opt.rcv_wscale; 62789e95a61SOri Finkelman opts->options |= OPTION_WSCALE; 628bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_WSCALE_ALIGNED; 62933ad798cSAdam Langley } 630f9301034SEric Dumazet if (likely(sock_net(sk)->ipv4.sysctl_tcp_sack)) { 63133ad798cSAdam Langley opts->options |= OPTION_SACK_ADVERTISE; 632b32d1310SDavid S. Miller if (unlikely(!(OPTION_TS & opts->options))) 633bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_SACKPERM_ALIGNED; 63433ad798cSAdam Langley } 63533ad798cSAdam Langley 636783237e8SYuchung Cheng if (fastopen && fastopen->cookie.len >= 0) { 6372646c831SDaniel Lee u32 need = fastopen->cookie.len; 6382646c831SDaniel Lee 6392646c831SDaniel Lee need += fastopen->cookie.exp ? TCPOLEN_EXP_FASTOPEN_BASE : 6402646c831SDaniel Lee TCPOLEN_FASTOPEN_BASE; 641783237e8SYuchung Cheng need = (need + 3) & ~3U; /* Align to 32 bits */ 642783237e8SYuchung Cheng if (remaining >= need) { 643783237e8SYuchung Cheng opts->options |= OPTION_FAST_OPEN_COOKIE; 644783237e8SYuchung Cheng opts->fastopen_cookie = &fastopen->cookie; 645783237e8SYuchung Cheng remaining -= need; 646783237e8SYuchung Cheng tp->syn_fastopen = 1; 6472646c831SDaniel Lee tp->syn_fastopen_exp = fastopen->cookie.exp ? 1 : 0; 648783237e8SYuchung Cheng } 649783237e8SYuchung Cheng } 650bd0388aeSWilliam Allen Simpson 65160e2a778SUrsula Braun smc_set_option(tp, opts, &remaining); 65260e2a778SUrsula Braun 653bd0388aeSWilliam Allen Simpson return MAX_TCP_OPTION_SPACE - remaining; 65433ad798cSAdam Langley } 65533ad798cSAdam Langley 65667edfef7SAndi Kleen /* Set up TCP options for SYN-ACKs. */ 65760e2a778SUrsula Braun static unsigned int tcp_synack_options(const struct sock *sk, 65860e2a778SUrsula Braun struct request_sock *req, 65995c96174SEric Dumazet unsigned int mss, struct sk_buff *skb, 66033ad798cSAdam Langley struct tcp_out_options *opts, 66180f03e27SEric Dumazet const struct tcp_md5sig_key *md5, 6628336886fSJerry Chu struct tcp_fastopen_cookie *foc) 6634957faadSWilliam Allen Simpson { 66433ad798cSAdam Langley struct inet_request_sock *ireq = inet_rsk(req); 66595c96174SEric Dumazet unsigned int remaining = MAX_TCP_OPTION_SPACE; 66633ad798cSAdam Langley 66733ad798cSAdam Langley #ifdef CONFIG_TCP_MD5SIG 66880f03e27SEric Dumazet if (md5) { 66933ad798cSAdam Langley opts->options |= OPTION_MD5; 6704957faadSWilliam Allen Simpson remaining -= TCPOLEN_MD5SIG_ALIGNED; 6714957faadSWilliam Allen Simpson 6724957faadSWilliam Allen Simpson /* We can't fit any SACK blocks in a packet with MD5 + TS 6734957faadSWilliam Allen Simpson * options. There was discussion about disabling SACK 6744957faadSWilliam Allen Simpson * rather than TS in order to fit in better with old, 6754957faadSWilliam Allen Simpson * buggy kernels, but that was deemed to be unnecessary. 6764957faadSWilliam Allen Simpson */ 677de213e5eSEric Dumazet ireq->tstamp_ok &= !ireq->sack_ok; 67833ad798cSAdam Langley } 67933ad798cSAdam Langley #endif 68033ad798cSAdam Langley 6814957faadSWilliam Allen Simpson /* We always send an MSS option. */ 68233ad798cSAdam Langley opts->mss = mss; 6834957faadSWilliam Allen Simpson remaining -= TCPOLEN_MSS_ALIGNED; 68433ad798cSAdam Langley 68533ad798cSAdam Langley if (likely(ireq->wscale_ok)) { 68633ad798cSAdam Langley opts->ws = ireq->rcv_wscale; 68789e95a61SOri Finkelman opts->options |= OPTION_WSCALE; 6884957faadSWilliam Allen Simpson remaining -= TCPOLEN_WSCALE_ALIGNED; 68933ad798cSAdam Langley } 690de213e5eSEric Dumazet if (likely(ireq->tstamp_ok)) { 69133ad798cSAdam Langley opts->options |= OPTION_TS; 69295a22caeSFlorian Westphal opts->tsval = tcp_skb_timestamp(skb) + tcp_rsk(req)->ts_off; 69333ad798cSAdam Langley opts->tsecr = req->ts_recent; 6944957faadSWilliam Allen Simpson remaining -= TCPOLEN_TSTAMP_ALIGNED; 69533ad798cSAdam Langley } 69633ad798cSAdam Langley if (likely(ireq->sack_ok)) { 69733ad798cSAdam Langley opts->options |= OPTION_SACK_ADVERTISE; 698de213e5eSEric Dumazet if (unlikely(!ireq->tstamp_ok)) 6994957faadSWilliam Allen Simpson remaining -= TCPOLEN_SACKPERM_ALIGNED; 70033ad798cSAdam Langley } 7017f9b838bSDaniel Lee if (foc != NULL && foc->len >= 0) { 7027f9b838bSDaniel Lee u32 need = foc->len; 7037f9b838bSDaniel Lee 7047f9b838bSDaniel Lee need += foc->exp ? TCPOLEN_EXP_FASTOPEN_BASE : 7057f9b838bSDaniel Lee TCPOLEN_FASTOPEN_BASE; 7068336886fSJerry Chu need = (need + 3) & ~3U; /* Align to 32 bits */ 7078336886fSJerry Chu if (remaining >= need) { 7088336886fSJerry Chu opts->options |= OPTION_FAST_OPEN_COOKIE; 7098336886fSJerry Chu opts->fastopen_cookie = foc; 7108336886fSJerry Chu remaining -= need; 7118336886fSJerry Chu } 7128336886fSJerry Chu } 7134957faadSWilliam Allen Simpson 71460e2a778SUrsula Braun smc_set_option_cond(tcp_sk(sk), ireq, opts, &remaining); 71560e2a778SUrsula Braun 7164957faadSWilliam Allen Simpson return MAX_TCP_OPTION_SPACE - remaining; 71733ad798cSAdam Langley } 71833ad798cSAdam Langley 71967edfef7SAndi Kleen /* Compute TCP options for ESTABLISHED sockets. This is not the 72067edfef7SAndi Kleen * final wire format yet. 72167edfef7SAndi Kleen */ 72295c96174SEric Dumazet static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb, 72333ad798cSAdam Langley struct tcp_out_options *opts, 724cf533ea5SEric Dumazet struct tcp_md5sig_key **md5) 725cf533ea5SEric Dumazet { 72633ad798cSAdam Langley struct tcp_sock *tp = tcp_sk(sk); 72795c96174SEric Dumazet unsigned int size = 0; 728cabeccbdSIlpo Järvinen unsigned int eff_sacks; 72933ad798cSAdam Langley 7305843ef42SAndi Kleen opts->options = 0; 7315843ef42SAndi Kleen 7328c2320e8SEric Dumazet *md5 = NULL; 73333ad798cSAdam Langley #ifdef CONFIG_TCP_MD5SIG 734921f9a0fSEric Dumazet if (static_branch_unlikely(&tcp_md5_needed) && 7356015c71eSEric Dumazet rcu_access_pointer(tp->md5sig_info)) { 73633ad798cSAdam Langley *md5 = tp->af_specific->md5_lookup(sk, sk); 7378c2320e8SEric Dumazet if (*md5) { 73833ad798cSAdam Langley opts->options |= OPTION_MD5; 73933ad798cSAdam Langley size += TCPOLEN_MD5SIG_ALIGNED; 74033ad798cSAdam Langley } 7418c2320e8SEric Dumazet } 74233ad798cSAdam Langley #endif 74333ad798cSAdam Langley 74433ad798cSAdam Langley if (likely(tp->rx_opt.tstamp_ok)) { 74533ad798cSAdam Langley opts->options |= OPTION_TS; 7467faee5c0SEric Dumazet opts->tsval = skb ? tcp_skb_timestamp(skb) + tp->tsoffset : 0; 74733ad798cSAdam Langley opts->tsecr = tp->rx_opt.ts_recent; 74833ad798cSAdam Langley size += TCPOLEN_TSTAMP_ALIGNED; 74933ad798cSAdam Langley } 75033ad798cSAdam Langley 751cabeccbdSIlpo Järvinen eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack; 752cabeccbdSIlpo Järvinen if (unlikely(eff_sacks)) { 75395c96174SEric Dumazet const unsigned int remaining = MAX_TCP_OPTION_SPACE - size; 75433ad798cSAdam Langley opts->num_sack_blocks = 75595c96174SEric Dumazet min_t(unsigned int, eff_sacks, 75633ad798cSAdam Langley (remaining - TCPOLEN_SACK_BASE_ALIGNED) / 75733ad798cSAdam Langley TCPOLEN_SACK_PERBLOCK); 75833ad798cSAdam Langley size += TCPOLEN_SACK_BASE_ALIGNED + 75933ad798cSAdam Langley opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK; 76033ad798cSAdam Langley } 76133ad798cSAdam Langley 76233ad798cSAdam Langley return size; 76340efc6faSStephen Hemminger } 7641da177e4SLinus Torvalds 76546d3ceabSEric Dumazet 76646d3ceabSEric Dumazet /* TCP SMALL QUEUES (TSQ) 76746d3ceabSEric Dumazet * 76846d3ceabSEric Dumazet * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev) 76946d3ceabSEric Dumazet * to reduce RTT and bufferbloat. 77046d3ceabSEric Dumazet * We do this using a special skb destructor (tcp_wfree). 77146d3ceabSEric Dumazet * 77246d3ceabSEric Dumazet * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb 77346d3ceabSEric Dumazet * needs to be reallocated in a driver. 7748e3bff96Sstephen hemminger * The invariant being skb->truesize subtracted from sk->sk_wmem_alloc 77546d3ceabSEric Dumazet * 77646d3ceabSEric Dumazet * Since transmit from skb destructor is forbidden, we use a tasklet 77746d3ceabSEric Dumazet * to process all sockets that eventually need to send more skbs. 77846d3ceabSEric Dumazet * We use one tasklet per cpu, with its own queue of sockets. 77946d3ceabSEric Dumazet */ 78046d3ceabSEric Dumazet struct tsq_tasklet { 78146d3ceabSEric Dumazet struct tasklet_struct tasklet; 78246d3ceabSEric Dumazet struct list_head head; /* queue of tcp sockets */ 78346d3ceabSEric Dumazet }; 78446d3ceabSEric Dumazet static DEFINE_PER_CPU(struct tsq_tasklet, tsq_tasklet); 78546d3ceabSEric Dumazet 78673a6bab5SEric Dumazet static void tcp_tsq_write(struct sock *sk) 7876f458dfbSEric Dumazet { 7886f458dfbSEric Dumazet if ((1 << sk->sk_state) & 7896f458dfbSEric Dumazet (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_CLOSING | 790f9616c35SEric Dumazet TCPF_CLOSE_WAIT | TCPF_LAST_ACK)) { 791f9616c35SEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 792f9616c35SEric Dumazet 793f9616c35SEric Dumazet if (tp->lost_out > tp->retrans_out && 7943a91d29fSKoichiro Den tp->snd_cwnd > tcp_packets_in_flight(tp)) { 7953a91d29fSKoichiro Den tcp_mstamp_refresh(tp); 796f9616c35SEric Dumazet tcp_xmit_retransmit_queue(sk); 7973a91d29fSKoichiro Den } 798f9616c35SEric Dumazet 799f9616c35SEric Dumazet tcp_write_xmit(sk, tcp_current_mss(sk), tp->nonagle, 800bf06200eSJohn Ogness 0, GFP_ATOMIC); 8016f458dfbSEric Dumazet } 802f9616c35SEric Dumazet } 80373a6bab5SEric Dumazet 80473a6bab5SEric Dumazet static void tcp_tsq_handler(struct sock *sk) 80573a6bab5SEric Dumazet { 80673a6bab5SEric Dumazet bh_lock_sock(sk); 80773a6bab5SEric Dumazet if (!sock_owned_by_user(sk)) 80873a6bab5SEric Dumazet tcp_tsq_write(sk); 80973a6bab5SEric Dumazet else if (!test_and_set_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags)) 81073a6bab5SEric Dumazet sock_hold(sk); 81173a6bab5SEric Dumazet bh_unlock_sock(sk); 81273a6bab5SEric Dumazet } 81346d3ceabSEric Dumazet /* 8148e3bff96Sstephen hemminger * One tasklet per cpu tries to send more skbs. 81546d3ceabSEric Dumazet * We run in tasklet context but need to disable irqs when 8168e3bff96Sstephen hemminger * transferring tsq->head because tcp_wfree() might 81746d3ceabSEric Dumazet * interrupt us (non NAPI drivers) 81846d3ceabSEric Dumazet */ 81946d3ceabSEric Dumazet static void tcp_tasklet_func(unsigned long data) 82046d3ceabSEric Dumazet { 82146d3ceabSEric Dumazet struct tsq_tasklet *tsq = (struct tsq_tasklet *)data; 82246d3ceabSEric Dumazet LIST_HEAD(list); 82346d3ceabSEric Dumazet unsigned long flags; 82446d3ceabSEric Dumazet struct list_head *q, *n; 82546d3ceabSEric Dumazet struct tcp_sock *tp; 82646d3ceabSEric Dumazet struct sock *sk; 82746d3ceabSEric Dumazet 82846d3ceabSEric Dumazet local_irq_save(flags); 82946d3ceabSEric Dumazet list_splice_init(&tsq->head, &list); 83046d3ceabSEric Dumazet local_irq_restore(flags); 83146d3ceabSEric Dumazet 83246d3ceabSEric Dumazet list_for_each_safe(q, n, &list) { 83346d3ceabSEric Dumazet tp = list_entry(q, struct tcp_sock, tsq_node); 83446d3ceabSEric Dumazet list_del(&tp->tsq_node); 83546d3ceabSEric Dumazet 83646d3ceabSEric Dumazet sk = (struct sock *)tp; 8370a9648f1SEric Dumazet smp_mb__before_atomic(); 8387aa5470cSEric Dumazet clear_bit(TSQ_QUEUED, &sk->sk_tsq_flags); 8397aa5470cSEric Dumazet 8406f458dfbSEric Dumazet tcp_tsq_handler(sk); 84146d3ceabSEric Dumazet sk_free(sk); 84246d3ceabSEric Dumazet } 84346d3ceabSEric Dumazet } 84446d3ceabSEric Dumazet 84540fc3423SEric Dumazet #define TCP_DEFERRED_ALL (TCPF_TSQ_DEFERRED | \ 84640fc3423SEric Dumazet TCPF_WRITE_TIMER_DEFERRED | \ 84740fc3423SEric Dumazet TCPF_DELACK_TIMER_DEFERRED | \ 84840fc3423SEric Dumazet TCPF_MTU_REDUCED_DEFERRED) 84946d3ceabSEric Dumazet /** 85046d3ceabSEric Dumazet * tcp_release_cb - tcp release_sock() callback 85146d3ceabSEric Dumazet * @sk: socket 85246d3ceabSEric Dumazet * 85346d3ceabSEric Dumazet * called from release_sock() to perform protocol dependent 85446d3ceabSEric Dumazet * actions before socket release. 85546d3ceabSEric Dumazet */ 85646d3ceabSEric Dumazet void tcp_release_cb(struct sock *sk) 85746d3ceabSEric Dumazet { 8586f458dfbSEric Dumazet unsigned long flags, nflags; 85946d3ceabSEric Dumazet 8606f458dfbSEric Dumazet /* perform an atomic operation only if at least one flag is set */ 8616f458dfbSEric Dumazet do { 8627aa5470cSEric Dumazet flags = sk->sk_tsq_flags; 8636f458dfbSEric Dumazet if (!(flags & TCP_DEFERRED_ALL)) 8646f458dfbSEric Dumazet return; 8656f458dfbSEric Dumazet nflags = flags & ~TCP_DEFERRED_ALL; 8667aa5470cSEric Dumazet } while (cmpxchg(&sk->sk_tsq_flags, flags, nflags) != flags); 8676f458dfbSEric Dumazet 86873a6bab5SEric Dumazet if (flags & TCPF_TSQ_DEFERRED) { 86973a6bab5SEric Dumazet tcp_tsq_write(sk); 87073a6bab5SEric Dumazet __sock_put(sk); 87173a6bab5SEric Dumazet } 872c3f9b018SEric Dumazet /* Here begins the tricky part : 873c3f9b018SEric Dumazet * We are called from release_sock() with : 874c3f9b018SEric Dumazet * 1) BH disabled 875c3f9b018SEric Dumazet * 2) sk_lock.slock spinlock held 876c3f9b018SEric Dumazet * 3) socket owned by us (sk->sk_lock.owned == 1) 877c3f9b018SEric Dumazet * 878c3f9b018SEric Dumazet * But following code is meant to be called from BH handlers, 879c3f9b018SEric Dumazet * so we should keep BH disabled, but early release socket ownership 880c3f9b018SEric Dumazet */ 881c3f9b018SEric Dumazet sock_release_ownership(sk); 882c3f9b018SEric Dumazet 88340fc3423SEric Dumazet if (flags & TCPF_WRITE_TIMER_DEFERRED) { 8846f458dfbSEric Dumazet tcp_write_timer_handler(sk); 885144d56e9SEric Dumazet __sock_put(sk); 886144d56e9SEric Dumazet } 88740fc3423SEric Dumazet if (flags & TCPF_DELACK_TIMER_DEFERRED) { 8886f458dfbSEric Dumazet tcp_delack_timer_handler(sk); 889144d56e9SEric Dumazet __sock_put(sk); 890144d56e9SEric Dumazet } 89140fc3423SEric Dumazet if (flags & TCPF_MTU_REDUCED_DEFERRED) { 8924fab9071SNeal Cardwell inet_csk(sk)->icsk_af_ops->mtu_reduced(sk); 893144d56e9SEric Dumazet __sock_put(sk); 894144d56e9SEric Dumazet } 89546d3ceabSEric Dumazet } 89646d3ceabSEric Dumazet EXPORT_SYMBOL(tcp_release_cb); 89746d3ceabSEric Dumazet 89846d3ceabSEric Dumazet void __init tcp_tasklet_init(void) 89946d3ceabSEric Dumazet { 90046d3ceabSEric Dumazet int i; 90146d3ceabSEric Dumazet 90246d3ceabSEric Dumazet for_each_possible_cpu(i) { 90346d3ceabSEric Dumazet struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i); 90446d3ceabSEric Dumazet 90546d3ceabSEric Dumazet INIT_LIST_HEAD(&tsq->head); 90646d3ceabSEric Dumazet tasklet_init(&tsq->tasklet, 90746d3ceabSEric Dumazet tcp_tasklet_func, 90846d3ceabSEric Dumazet (unsigned long)tsq); 90946d3ceabSEric Dumazet } 91046d3ceabSEric Dumazet } 91146d3ceabSEric Dumazet 91246d3ceabSEric Dumazet /* 91346d3ceabSEric Dumazet * Write buffer destructor automatically called from kfree_skb. 9148e3bff96Sstephen hemminger * We can't xmit new skbs from this context, as we might already 91546d3ceabSEric Dumazet * hold qdisc lock. 91646d3ceabSEric Dumazet */ 917d6a4a104SEric Dumazet void tcp_wfree(struct sk_buff *skb) 91846d3ceabSEric Dumazet { 91946d3ceabSEric Dumazet struct sock *sk = skb->sk; 92046d3ceabSEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 921408f0a6cSEric Dumazet unsigned long flags, nval, oval; 9229b462d02SEric Dumazet 9239b462d02SEric Dumazet /* Keep one reference on sk_wmem_alloc. 9249b462d02SEric Dumazet * Will be released by sk_free() from here or tcp_tasklet_func() 9259b462d02SEric Dumazet */ 92614afee4bSReshetova, Elena WARN_ON(refcount_sub_and_test(skb->truesize - 1, &sk->sk_wmem_alloc)); 9279b462d02SEric Dumazet 9289b462d02SEric Dumazet /* If this softirq is serviced by ksoftirqd, we are likely under stress. 9299b462d02SEric Dumazet * Wait until our queues (qdisc + devices) are drained. 9309b462d02SEric Dumazet * This gives : 9319b462d02SEric Dumazet * - less callbacks to tcp_write_xmit(), reducing stress (batches) 9329b462d02SEric Dumazet * - chance for incoming ACK (processed by another cpu maybe) 9339b462d02SEric Dumazet * to migrate this flow (skb->ooo_okay will be eventually set) 9349b462d02SEric Dumazet */ 93514afee4bSReshetova, Elena if (refcount_read(&sk->sk_wmem_alloc) >= SKB_TRUESIZE(1) && this_cpu_ksoftirqd() == current) 9369b462d02SEric Dumazet goto out; 93746d3ceabSEric Dumazet 9387aa5470cSEric Dumazet for (oval = READ_ONCE(sk->sk_tsq_flags);; oval = nval) { 93946d3ceabSEric Dumazet struct tsq_tasklet *tsq; 940a9b204d1SEric Dumazet bool empty; 94146d3ceabSEric Dumazet 942408f0a6cSEric Dumazet if (!(oval & TSQF_THROTTLED) || (oval & TSQF_QUEUED)) 943408f0a6cSEric Dumazet goto out; 944408f0a6cSEric Dumazet 94573a6bab5SEric Dumazet nval = (oval & ~TSQF_THROTTLED) | TSQF_QUEUED; 9467aa5470cSEric Dumazet nval = cmpxchg(&sk->sk_tsq_flags, oval, nval); 947408f0a6cSEric Dumazet if (nval != oval) 948408f0a6cSEric Dumazet continue; 949408f0a6cSEric Dumazet 95046d3ceabSEric Dumazet /* queue this socket to tasklet queue */ 95146d3ceabSEric Dumazet local_irq_save(flags); 952903ceff7SChristoph Lameter tsq = this_cpu_ptr(&tsq_tasklet); 953a9b204d1SEric Dumazet empty = list_empty(&tsq->head); 95446d3ceabSEric Dumazet list_add(&tp->tsq_node, &tsq->head); 955a9b204d1SEric Dumazet if (empty) 95646d3ceabSEric Dumazet tasklet_schedule(&tsq->tasklet); 95746d3ceabSEric Dumazet local_irq_restore(flags); 9589b462d02SEric Dumazet return; 95946d3ceabSEric Dumazet } 9609b462d02SEric Dumazet out: 9619b462d02SEric Dumazet sk_free(sk); 96246d3ceabSEric Dumazet } 96346d3ceabSEric Dumazet 96473a6bab5SEric Dumazet /* Note: Called under soft irq. 96573a6bab5SEric Dumazet * We can call TCP stack right away, unless socket is owned by user. 966218af599SEric Dumazet */ 967218af599SEric Dumazet enum hrtimer_restart tcp_pace_kick(struct hrtimer *timer) 968218af599SEric Dumazet { 969218af599SEric Dumazet struct tcp_sock *tp = container_of(timer, struct tcp_sock, pacing_timer); 970218af599SEric Dumazet struct sock *sk = (struct sock *)tp; 971218af599SEric Dumazet 97273a6bab5SEric Dumazet tcp_tsq_handler(sk); 97373a6bab5SEric Dumazet sock_put(sk); 974218af599SEric Dumazet 975218af599SEric Dumazet return HRTIMER_NORESTART; 976218af599SEric Dumazet } 977218af599SEric Dumazet 978a7a25630SEric Dumazet static void tcp_update_skb_after_send(struct sock *sk, struct sk_buff *skb, 979a7a25630SEric Dumazet u64 prior_wstamp) 980e2080072SEric Dumazet { 981ab408b6dSEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 982ab408b6dSEric Dumazet 983ab408b6dSEric Dumazet if (sk->sk_pacing_status != SK_PACING_NONE) { 98476a9ebe8SEric Dumazet unsigned long rate = sk->sk_pacing_rate; 985ab408b6dSEric Dumazet 986ab408b6dSEric Dumazet /* Original sch_fq does not pace first 10 MSS 987ab408b6dSEric Dumazet * Note that tp->data_segs_out overflows after 2^32 packets, 988ab408b6dSEric Dumazet * this is a minor annoyance. 989ab408b6dSEric Dumazet */ 99076a9ebe8SEric Dumazet if (rate != ~0UL && rate && tp->data_segs_out >= 10) { 991a7a25630SEric Dumazet u64 len_ns = div64_ul((u64)skb->len * NSEC_PER_SEC, rate); 992a7a25630SEric Dumazet u64 credit = tp->tcp_wstamp_ns - prior_wstamp; 993a7a25630SEric Dumazet 994a7a25630SEric Dumazet /* take into account OS jitter */ 995a7a25630SEric Dumazet len_ns -= min_t(u64, len_ns / 2, credit); 996a7a25630SEric Dumazet tp->tcp_wstamp_ns += len_ns; 997ab408b6dSEric Dumazet } 998ab408b6dSEric Dumazet } 999e2080072SEric Dumazet list_move_tail(&skb->tcp_tsorted_anchor, &tp->tsorted_sent_queue); 1000e2080072SEric Dumazet } 1001e2080072SEric Dumazet 10021da177e4SLinus Torvalds /* This routine actually transmits TCP packets queued in by 10031da177e4SLinus Torvalds * tcp_do_sendmsg(). This is used by both the initial 10041da177e4SLinus Torvalds * transmission and possible later retransmissions. 10051da177e4SLinus Torvalds * All SKB's seen here are completely headerless. It is our 10061da177e4SLinus Torvalds * job to build the TCP header, and pass the packet down to 10071da177e4SLinus Torvalds * IP so it can do the same plus pass the packet off to the 10081da177e4SLinus Torvalds * device. 10091da177e4SLinus Torvalds * 10101da177e4SLinus Torvalds * We are working here with either a clone of the original 10111da177e4SLinus Torvalds * SKB, or a fresh unique copy made by the retransmit engine. 10121da177e4SLinus Torvalds */ 10132987babbSYuchung Cheng static int __tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, 10142987babbSYuchung Cheng int clone_it, gfp_t gfp_mask, u32 rcv_nxt) 10151da177e4SLinus Torvalds { 10166687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 1017dfb4b9dcSDavid S. Miller struct inet_sock *inet; 1018dfb4b9dcSDavid S. Miller struct tcp_sock *tp; 1019dfb4b9dcSDavid S. Miller struct tcp_skb_cb *tcb; 102033ad798cSAdam Langley struct tcp_out_options opts; 102195c96174SEric Dumazet unsigned int tcp_options_size, tcp_header_size; 10228c72c65bSEric Dumazet struct sk_buff *oskb = NULL; 1023cfb6eeb4SYOSHIFUJI Hideaki struct tcp_md5sig_key *md5; 10241da177e4SLinus Torvalds struct tcphdr *th; 1025a7a25630SEric Dumazet u64 prior_wstamp; 10261da177e4SLinus Torvalds int err; 10271da177e4SLinus Torvalds 1028dfb4b9dcSDavid S. Miller BUG_ON(!skb || !tcp_skb_pcount(skb)); 10296f094b9eSLawrence Brakmo tp = tcp_sk(sk); 10307f12422cSYuchung Cheng prior_wstamp = tp->tcp_wstamp_ns; 10317f12422cSYuchung Cheng tp->tcp_wstamp_ns = max(tp->tcp_wstamp_ns, tp->tcp_clock_cache); 10327f12422cSYuchung Cheng skb->skb_mstamp_ns = tp->tcp_wstamp_ns; 1033ccdbb6e9SEric Dumazet if (clone_it) { 10346f094b9eSLawrence Brakmo TCP_SKB_CB(skb)->tx.in_flight = TCP_SKB_CB(skb)->end_seq 10356f094b9eSLawrence Brakmo - tp->snd_una; 10368c72c65bSEric Dumazet oskb = skb; 1037e2080072SEric Dumazet 1038e2080072SEric Dumazet tcp_skb_tsorted_save(oskb) { 1039e2080072SEric Dumazet if (unlikely(skb_cloned(oskb))) 1040e2080072SEric Dumazet skb = pskb_copy(oskb, gfp_mask); 1041dfb4b9dcSDavid S. Miller else 1042e2080072SEric Dumazet skb = skb_clone(oskb, gfp_mask); 1043e2080072SEric Dumazet } tcp_skb_tsorted_restore(oskb); 1044e2080072SEric Dumazet 1045dfb4b9dcSDavid S. Miller if (unlikely(!skb)) 1046dfb4b9dcSDavid S. Miller return -ENOBUFS; 1047dfb4b9dcSDavid S. Miller } 10485f6188a8SEric Dumazet 1049dfb4b9dcSDavid S. Miller inet = inet_sk(sk); 1050dfb4b9dcSDavid S. Miller tcb = TCP_SKB_CB(skb); 105133ad798cSAdam Langley memset(&opts, 0, sizeof(opts)); 10521da177e4SLinus Torvalds 10534de075e0SEric Dumazet if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) 105433ad798cSAdam Langley tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5); 105533ad798cSAdam Langley else 105633ad798cSAdam Langley tcp_options_size = tcp_established_options(sk, skb, &opts, 105733ad798cSAdam Langley &md5); 105833ad798cSAdam Langley tcp_header_size = tcp_options_size + sizeof(struct tcphdr); 10591da177e4SLinus Torvalds 1060547669d4SEric Dumazet /* if no packet is in qdisc/device queue, then allow XPS to select 1061b2532eb9SEric Dumazet * another queue. We can be called from tcp_tsq_handler() 106273a6bab5SEric Dumazet * which holds one reference to sk. 1063b2532eb9SEric Dumazet * 1064b2532eb9SEric Dumazet * TODO: Ideally, in-flight pure ACK packets should not matter here. 1065b2532eb9SEric Dumazet * One way to get this would be to set skb->truesize = 2 on them. 1066547669d4SEric Dumazet */ 1067b2532eb9SEric Dumazet skb->ooo_okay = sk_wmem_alloc_get(sk) < SKB_TRUESIZE(1); 10681da177e4SLinus Torvalds 106938ab52e8SEric Dumazet /* If we had to use memory reserve to allocate this skb, 107038ab52e8SEric Dumazet * this might cause drops if packet is looped back : 107138ab52e8SEric Dumazet * Other socket might not have SOCK_MEMALLOC. 107238ab52e8SEric Dumazet * Packets not looped back do not care about pfmemalloc. 107338ab52e8SEric Dumazet */ 107438ab52e8SEric Dumazet skb->pfmemalloc = 0; 107538ab52e8SEric Dumazet 1076aa8223c7SArnaldo Carvalho de Melo skb_push(skb, tcp_header_size); 1077aa8223c7SArnaldo Carvalho de Melo skb_reset_transport_header(skb); 107846d3ceabSEric Dumazet 107946d3ceabSEric Dumazet skb_orphan(skb); 108046d3ceabSEric Dumazet skb->sk = sk; 10811d2077acSEric Dumazet skb->destructor = skb_is_tcp_pure_ack(skb) ? __sock_wfree : tcp_wfree; 1082b73c3d0eSTom Herbert skb_set_hash_from_sk(skb, sk); 108314afee4bSReshetova, Elena refcount_add(skb->truesize, &sk->sk_wmem_alloc); 10841da177e4SLinus Torvalds 1085c3a2e837SJulian Anastasov skb_set_dst_pending_confirm(skb, sk->sk_dst_pending_confirm); 1086c3a2e837SJulian Anastasov 10871da177e4SLinus Torvalds /* Build TCP header and checksum it. */ 1088ea1627c2SEric Dumazet th = (struct tcphdr *)skb->data; 1089c720c7e8SEric Dumazet th->source = inet->inet_sport; 1090c720c7e8SEric Dumazet th->dest = inet->inet_dport; 10911da177e4SLinus Torvalds th->seq = htonl(tcb->seq); 10922987babbSYuchung Cheng th->ack_seq = htonl(rcv_nxt); 1093df7a3b07SAl Viro *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) | 10944de075e0SEric Dumazet tcb->tcp_flags); 1095dfb4b9dcSDavid S. Miller 10961da177e4SLinus Torvalds th->check = 0; 10971da177e4SLinus Torvalds th->urg_ptr = 0; 10981da177e4SLinus Torvalds 109933f5f57eSIlpo Järvinen /* The urg_mode check is necessary during a below snd_una win probe */ 11007691367dSHerbert Xu if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) { 11017691367dSHerbert Xu if (before(tp->snd_up, tcb->seq + 0x10000)) { 11021da177e4SLinus Torvalds th->urg_ptr = htons(tp->snd_up - tcb->seq); 11031da177e4SLinus Torvalds th->urg = 1; 11047691367dSHerbert Xu } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) { 11050eae88f3SEric Dumazet th->urg_ptr = htons(0xFFFF); 11067691367dSHerbert Xu th->urg = 1; 11077691367dSHerbert Xu } 11081da177e4SLinus Torvalds } 11091da177e4SLinus Torvalds 1110bd0388aeSWilliam Allen Simpson tcp_options_write((__be32 *)(th + 1), tp, &opts); 111151466a75SEric Dumazet skb_shinfo(skb)->gso_type = sk->sk_gso_type; 1112ea1627c2SEric Dumazet if (likely(!(tcb->tcp_flags & TCPHDR_SYN))) { 1113ea1627c2SEric Dumazet th->window = htons(tcp_select_window(sk)); 1114ea1627c2SEric Dumazet tcp_ecn_send(sk, skb, th, tcp_header_size); 1115ea1627c2SEric Dumazet } else { 1116ea1627c2SEric Dumazet /* RFC1323: The window in SYN & SYN/ACK segments 1117ea1627c2SEric Dumazet * is never scaled. 1118ea1627c2SEric Dumazet */ 1119ea1627c2SEric Dumazet th->window = htons(min(tp->rcv_wnd, 65535U)); 1120ea1627c2SEric Dumazet } 1121cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 1122cfb6eeb4SYOSHIFUJI Hideaki /* Calculate the MD5 hash, as we have all we need now */ 1123cfb6eeb4SYOSHIFUJI Hideaki if (md5) { 1124a465419bSEric Dumazet sk_nocaps_add(sk, NETIF_F_GSO_MASK); 1125bd0388aeSWilliam Allen Simpson tp->af_specific->calc_md5_hash(opts.hash_location, 112639f8e58eSEric Dumazet md5, sk, skb); 1127cfb6eeb4SYOSHIFUJI Hideaki } 1128cfb6eeb4SYOSHIFUJI Hideaki #endif 1129cfb6eeb4SYOSHIFUJI Hideaki 1130bb296246SHerbert Xu icsk->icsk_af_ops->send_check(sk, skb); 11311da177e4SLinus Torvalds 11324de075e0SEric Dumazet if (likely(tcb->tcp_flags & TCPHDR_ACK)) 113327cde44aSYuchung Cheng tcp_event_ack_sent(sk, tcp_skb_pcount(skb), rcv_nxt); 11341da177e4SLinus Torvalds 1135a44d6eacSMartin KaFai Lau if (skb->len != tcp_header_size) { 1136cf533ea5SEric Dumazet tcp_event_data_sent(tp, sk); 1137a44d6eacSMartin KaFai Lau tp->data_segs_out += tcp_skb_pcount(skb); 1138ba113c3aSWei Wang tp->bytes_sent += skb->len - tcp_header_size; 1139a44d6eacSMartin KaFai Lau } 11401da177e4SLinus Torvalds 1141bd37a088SWei Yongjun if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq) 1142aa2ea058STom Herbert TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS, 1143aa2ea058STom Herbert tcp_skb_pcount(skb)); 11441da177e4SLinus Torvalds 11452efd055cSMarcelo Ricardo Leitner tp->segs_out += tcp_skb_pcount(skb); 1146f69ad292SEric Dumazet /* OK, its time to fill skb_shinfo(skb)->gso_{segs|size} */ 1147cd7d8498SEric Dumazet skb_shinfo(skb)->gso_segs = tcp_skb_pcount(skb); 1148f69ad292SEric Dumazet skb_shinfo(skb)->gso_size = tcp_skb_mss(skb); 1149cd7d8498SEric Dumazet 1150d3edd06eSEric Dumazet /* Leave earliest departure time in skb->tstamp (skb->skb_mstamp_ns) */ 1151971f10ecSEric Dumazet 1152971f10ecSEric Dumazet /* Cleanup our debris for IP stacks */ 1153971f10ecSEric Dumazet memset(skb->cb, 0, max(sizeof(struct inet_skb_parm), 1154971f10ecSEric Dumazet sizeof(struct inet6_skb_parm))); 1155971f10ecSEric Dumazet 1156b0270e91SEric Dumazet err = icsk->icsk_af_ops->queue_xmit(sk, skb, &inet->cork.fl); 11577faee5c0SEric Dumazet 11588c72c65bSEric Dumazet if (unlikely(err > 0)) { 11595ee2c941SChristoph Paasch tcp_enter_cwr(sk); 11608c72c65bSEric Dumazet err = net_xmit_eval(err); 11618c72c65bSEric Dumazet } 1162fc225799SEric Dumazet if (!err && oskb) { 1163a7a25630SEric Dumazet tcp_update_skb_after_send(sk, oskb, prior_wstamp); 1164fc225799SEric Dumazet tcp_rate_skb_sent(sk, oskb); 1165fc225799SEric Dumazet } 11668c72c65bSEric Dumazet return err; 11671da177e4SLinus Torvalds } 11681da177e4SLinus Torvalds 11692987babbSYuchung Cheng static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it, 11702987babbSYuchung Cheng gfp_t gfp_mask) 11712987babbSYuchung Cheng { 11722987babbSYuchung Cheng return __tcp_transmit_skb(sk, skb, clone_it, gfp_mask, 11732987babbSYuchung Cheng tcp_sk(sk)->rcv_nxt); 11742987babbSYuchung Cheng } 11752987babbSYuchung Cheng 117667edfef7SAndi Kleen /* This routine just queues the buffer for sending. 11771da177e4SLinus Torvalds * 11781da177e4SLinus Torvalds * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames, 11791da177e4SLinus Torvalds * otherwise socket can stall. 11801da177e4SLinus Torvalds */ 11811da177e4SLinus Torvalds static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb) 11821da177e4SLinus Torvalds { 11831da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 11841da177e4SLinus Torvalds 11851da177e4SLinus Torvalds /* Advance write_seq and place onto the write_queue. */ 11861da177e4SLinus Torvalds tp->write_seq = TCP_SKB_CB(skb)->end_seq; 1187f4a775d1SEric Dumazet __skb_header_release(skb); 1188fe067e8aSDavid S. Miller tcp_add_write_queue_tail(sk, skb); 11893ab224beSHideo Aoki sk->sk_wmem_queued += skb->truesize; 11903ab224beSHideo Aoki sk_mem_charge(sk, skb->truesize); 11911da177e4SLinus Torvalds } 11921da177e4SLinus Torvalds 119367edfef7SAndi Kleen /* Initialize TSO segments for a packet. */ 11945bbb432cSEric Dumazet static void tcp_set_skb_tso_segs(struct sk_buff *skb, unsigned int mss_now) 1195f6302d1dSDavid S. Miller { 11964a64fd6cSEric Dumazet if (skb->len <= mss_now) { 1197f6302d1dSDavid S. Miller /* Avoid the costly divide in the normal 1198f6302d1dSDavid S. Miller * non-TSO case. 1199f6302d1dSDavid S. Miller */ 1200cd7d8498SEric Dumazet tcp_skb_pcount_set(skb, 1); 1201f69ad292SEric Dumazet TCP_SKB_CB(skb)->tcp_gso_size = 0; 1202f6302d1dSDavid S. Miller } else { 1203cd7d8498SEric Dumazet tcp_skb_pcount_set(skb, DIV_ROUND_UP(skb->len, mss_now)); 1204f69ad292SEric Dumazet TCP_SKB_CB(skb)->tcp_gso_size = mss_now; 12051da177e4SLinus Torvalds } 12061da177e4SLinus Torvalds } 12071da177e4SLinus Torvalds 1208797108d1SIlpo Järvinen /* Pcount in the middle of the write queue got changed, we need to do various 1209797108d1SIlpo Järvinen * tweaks to fix counters 1210797108d1SIlpo Järvinen */ 1211cf533ea5SEric Dumazet static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr) 1212797108d1SIlpo Järvinen { 1213797108d1SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 1214797108d1SIlpo Järvinen 1215797108d1SIlpo Järvinen tp->packets_out -= decr; 1216797108d1SIlpo Järvinen 1217797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 1218797108d1SIlpo Järvinen tp->sacked_out -= decr; 1219797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) 1220797108d1SIlpo Järvinen tp->retrans_out -= decr; 1221797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST) 1222797108d1SIlpo Järvinen tp->lost_out -= decr; 1223797108d1SIlpo Järvinen 1224797108d1SIlpo Järvinen /* Reno case is special. Sigh... */ 1225797108d1SIlpo Järvinen if (tcp_is_reno(tp) && decr > 0) 1226797108d1SIlpo Järvinen tp->sacked_out -= min_t(u32, tp->sacked_out, decr); 1227797108d1SIlpo Järvinen 1228797108d1SIlpo Järvinen if (tp->lost_skb_hint && 1229797108d1SIlpo Järvinen before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) && 1230713bafeaSYuchung Cheng (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) 1231797108d1SIlpo Järvinen tp->lost_cnt_hint -= decr; 1232797108d1SIlpo Järvinen 1233797108d1SIlpo Järvinen tcp_verify_left_out(tp); 1234797108d1SIlpo Järvinen } 1235797108d1SIlpo Järvinen 12360a2cf20cSSoheil Hassas Yeganeh static bool tcp_has_tx_tstamp(const struct sk_buff *skb) 12370a2cf20cSSoheil Hassas Yeganeh { 12380a2cf20cSSoheil Hassas Yeganeh return TCP_SKB_CB(skb)->txstamp_ack || 12390a2cf20cSSoheil Hassas Yeganeh (skb_shinfo(skb)->tx_flags & SKBTX_ANY_TSTAMP); 12400a2cf20cSSoheil Hassas Yeganeh } 12410a2cf20cSSoheil Hassas Yeganeh 1242490cc7d0SWillem de Bruijn static void tcp_fragment_tstamp(struct sk_buff *skb, struct sk_buff *skb2) 1243490cc7d0SWillem de Bruijn { 1244490cc7d0SWillem de Bruijn struct skb_shared_info *shinfo = skb_shinfo(skb); 1245490cc7d0SWillem de Bruijn 12460a2cf20cSSoheil Hassas Yeganeh if (unlikely(tcp_has_tx_tstamp(skb)) && 1247490cc7d0SWillem de Bruijn !before(shinfo->tskey, TCP_SKB_CB(skb2)->seq)) { 1248490cc7d0SWillem de Bruijn struct skb_shared_info *shinfo2 = skb_shinfo(skb2); 1249490cc7d0SWillem de Bruijn u8 tsflags = shinfo->tx_flags & SKBTX_ANY_TSTAMP; 1250490cc7d0SWillem de Bruijn 1251490cc7d0SWillem de Bruijn shinfo->tx_flags &= ~tsflags; 1252490cc7d0SWillem de Bruijn shinfo2->tx_flags |= tsflags; 1253490cc7d0SWillem de Bruijn swap(shinfo->tskey, shinfo2->tskey); 1254b51e13faSMartin KaFai Lau TCP_SKB_CB(skb2)->txstamp_ack = TCP_SKB_CB(skb)->txstamp_ack; 1255b51e13faSMartin KaFai Lau TCP_SKB_CB(skb)->txstamp_ack = 0; 1256490cc7d0SWillem de Bruijn } 1257490cc7d0SWillem de Bruijn } 1258490cc7d0SWillem de Bruijn 1259a166140eSMartin KaFai Lau static void tcp_skb_fragment_eor(struct sk_buff *skb, struct sk_buff *skb2) 1260a166140eSMartin KaFai Lau { 1261a166140eSMartin KaFai Lau TCP_SKB_CB(skb2)->eor = TCP_SKB_CB(skb)->eor; 1262a166140eSMartin KaFai Lau TCP_SKB_CB(skb)->eor = 0; 1263a166140eSMartin KaFai Lau } 1264a166140eSMartin KaFai Lau 126575c119afSEric Dumazet /* Insert buff after skb on the write or rtx queue of sk. */ 126675c119afSEric Dumazet static void tcp_insert_write_queue_after(struct sk_buff *skb, 126775c119afSEric Dumazet struct sk_buff *buff, 126875c119afSEric Dumazet struct sock *sk, 126975c119afSEric Dumazet enum tcp_queue tcp_queue) 127075c119afSEric Dumazet { 127175c119afSEric Dumazet if (tcp_queue == TCP_FRAG_IN_WRITE_QUEUE) 127275c119afSEric Dumazet __skb_queue_after(&sk->sk_write_queue, skb, buff); 127375c119afSEric Dumazet else 127475c119afSEric Dumazet tcp_rbtree_insert(&sk->tcp_rtx_queue, buff); 127575c119afSEric Dumazet } 127675c119afSEric Dumazet 12771da177e4SLinus Torvalds /* Function to create two new TCP segments. Shrinks the given segment 12781da177e4SLinus Torvalds * to the specified size and appends a new segment with the rest of the 12791da177e4SLinus Torvalds * packet to the list. This won't be called frequently, I hope. 12801da177e4SLinus Torvalds * Remember, these are still headerless SKBs at this point. 12811da177e4SLinus Torvalds */ 128275c119afSEric Dumazet int tcp_fragment(struct sock *sk, enum tcp_queue tcp_queue, 128375c119afSEric Dumazet struct sk_buff *skb, u32 len, 12846cc55e09SOctavian Purdila unsigned int mss_now, gfp_t gfp) 12851da177e4SLinus Torvalds { 12861da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 12871da177e4SLinus Torvalds struct sk_buff *buff; 12886475be16SDavid S. Miller int nsize, old_factor; 1289b60b49eaSHerbert Xu int nlen; 12909ce01461SIlpo Järvinen u8 flags; 12911da177e4SLinus Torvalds 12922fceec13SIlpo Järvinen if (WARN_ON(len > skb->len)) 12932fceec13SIlpo Järvinen return -EINVAL; 12946a438bbeSStephen Hemminger 12951da177e4SLinus Torvalds nsize = skb_headlen(skb) - len; 12961da177e4SLinus Torvalds if (nsize < 0) 12971da177e4SLinus Torvalds nsize = 0; 12981da177e4SLinus Torvalds 1299*f070ef2aSEric Dumazet if (unlikely((sk->sk_wmem_queued >> 1) > sk->sk_sndbuf)) { 1300*f070ef2aSEric Dumazet NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPWQUEUETOOBIG); 1301*f070ef2aSEric Dumazet return -ENOMEM; 1302*f070ef2aSEric Dumazet } 1303*f070ef2aSEric Dumazet 13046cc55e09SOctavian Purdila if (skb_unclone(skb, gfp)) 13051da177e4SLinus Torvalds return -ENOMEM; 13061da177e4SLinus Torvalds 13071da177e4SLinus Torvalds /* Get a new skb... force flag on. */ 1308eb934478SEric Dumazet buff = sk_stream_alloc_skb(sk, nsize, gfp, true); 130951456b29SIan Morris if (!buff) 13101da177e4SLinus Torvalds return -ENOMEM; /* We'll just try again later. */ 1311ef5cb973SHerbert Xu 13123ab224beSHideo Aoki sk->sk_wmem_queued += buff->truesize; 13133ab224beSHideo Aoki sk_mem_charge(sk, buff->truesize); 1314b60b49eaSHerbert Xu nlen = skb->len - len - nsize; 1315b60b49eaSHerbert Xu buff->truesize += nlen; 1316b60b49eaSHerbert Xu skb->truesize -= nlen; 13171da177e4SLinus Torvalds 13181da177e4SLinus Torvalds /* Correct the sequence numbers. */ 13191da177e4SLinus Torvalds TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; 13201da177e4SLinus Torvalds TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; 13211da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; 13221da177e4SLinus Torvalds 13231da177e4SLinus Torvalds /* PSH and FIN should only be set in the second packet. */ 13244de075e0SEric Dumazet flags = TCP_SKB_CB(skb)->tcp_flags; 13254de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); 13264de075e0SEric Dumazet TCP_SKB_CB(buff)->tcp_flags = flags; 1327e14c3cafSHerbert Xu TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked; 1328a166140eSMartin KaFai Lau tcp_skb_fragment_eor(skb, buff); 13291da177e4SLinus Torvalds 13301da177e4SLinus Torvalds skb_split(skb, buff, len); 13311da177e4SLinus Torvalds 133298be9b12SEric Dumazet buff->ip_summed = CHECKSUM_PARTIAL; 13331da177e4SLinus Torvalds 1334a61bbcf2SPatrick McHardy buff->tstamp = skb->tstamp; 1335490cc7d0SWillem de Bruijn tcp_fragment_tstamp(skb, buff); 13361da177e4SLinus Torvalds 13376475be16SDavid S. Miller old_factor = tcp_skb_pcount(skb); 13386475be16SDavid S. Miller 13391da177e4SLinus Torvalds /* Fix up tso_factor for both original and new SKB. */ 13405bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss_now); 13415bbb432cSEric Dumazet tcp_set_skb_tso_segs(buff, mss_now); 13421da177e4SLinus Torvalds 1343b9f64820SYuchung Cheng /* Update delivered info for the new segment */ 1344b9f64820SYuchung Cheng TCP_SKB_CB(buff)->tx = TCP_SKB_CB(skb)->tx; 1345b9f64820SYuchung Cheng 13466475be16SDavid S. Miller /* If this packet has been sent out already, we must 13476475be16SDavid S. Miller * adjust the various packet counters. 13486475be16SDavid S. Miller */ 1349cf0b450cSHerbert Xu if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) { 13506475be16SDavid S. Miller int diff = old_factor - tcp_skb_pcount(skb) - 13516475be16SDavid S. Miller tcp_skb_pcount(buff); 13521da177e4SLinus Torvalds 1353797108d1SIlpo Järvinen if (diff) 1354797108d1SIlpo Järvinen tcp_adjust_pcount(sk, skb, diff); 13551da177e4SLinus Torvalds } 13561da177e4SLinus Torvalds 13571da177e4SLinus Torvalds /* Link BUFF into the send queue. */ 1358f4a775d1SEric Dumazet __skb_header_release(buff); 135975c119afSEric Dumazet tcp_insert_write_queue_after(skb, buff, sk, tcp_queue); 1360f67971e6SEric Dumazet if (tcp_queue == TCP_FRAG_IN_RTX_QUEUE) 1361e2080072SEric Dumazet list_add(&buff->tcp_tsorted_anchor, &skb->tcp_tsorted_anchor); 13621da177e4SLinus Torvalds 13631da177e4SLinus Torvalds return 0; 13641da177e4SLinus Torvalds } 13651da177e4SLinus Torvalds 1366f4d01666SEric Dumazet /* This is similar to __pskb_pull_tail(). The difference is that pulled 1367f4d01666SEric Dumazet * data is not copied, but immediately discarded. 13681da177e4SLinus Torvalds */ 13697162fb24SEric Dumazet static int __pskb_trim_head(struct sk_buff *skb, int len) 13701da177e4SLinus Torvalds { 13717b7fc97aSEric Dumazet struct skb_shared_info *shinfo; 13721da177e4SLinus Torvalds int i, k, eat; 13731da177e4SLinus Torvalds 13744fa48bf3SEric Dumazet eat = min_t(int, len, skb_headlen(skb)); 13754fa48bf3SEric Dumazet if (eat) { 13764fa48bf3SEric Dumazet __skb_pull(skb, eat); 13774fa48bf3SEric Dumazet len -= eat; 13784fa48bf3SEric Dumazet if (!len) 13797162fb24SEric Dumazet return 0; 13804fa48bf3SEric Dumazet } 13811da177e4SLinus Torvalds eat = len; 13821da177e4SLinus Torvalds k = 0; 13837b7fc97aSEric Dumazet shinfo = skb_shinfo(skb); 13847b7fc97aSEric Dumazet for (i = 0; i < shinfo->nr_frags; i++) { 13857b7fc97aSEric Dumazet int size = skb_frag_size(&shinfo->frags[i]); 13869e903e08SEric Dumazet 13879e903e08SEric Dumazet if (size <= eat) { 1388aff65da0SIan Campbell skb_frag_unref(skb, i); 13899e903e08SEric Dumazet eat -= size; 13901da177e4SLinus Torvalds } else { 13917b7fc97aSEric Dumazet shinfo->frags[k] = shinfo->frags[i]; 13921da177e4SLinus Torvalds if (eat) { 13937b7fc97aSEric Dumazet shinfo->frags[k].page_offset += eat; 13947b7fc97aSEric Dumazet skb_frag_size_sub(&shinfo->frags[k], eat); 13951da177e4SLinus Torvalds eat = 0; 13961da177e4SLinus Torvalds } 13971da177e4SLinus Torvalds k++; 13981da177e4SLinus Torvalds } 13991da177e4SLinus Torvalds } 14007b7fc97aSEric Dumazet shinfo->nr_frags = k; 14011da177e4SLinus Torvalds 14021da177e4SLinus Torvalds skb->data_len -= len; 14031da177e4SLinus Torvalds skb->len = skb->data_len; 14047162fb24SEric Dumazet return len; 14051da177e4SLinus Torvalds } 14061da177e4SLinus Torvalds 140767edfef7SAndi Kleen /* Remove acked data from a packet in the transmit queue. */ 14081da177e4SLinus Torvalds int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len) 14091da177e4SLinus Torvalds { 14107162fb24SEric Dumazet u32 delta_truesize; 14117162fb24SEric Dumazet 141214bbd6a5SPravin B Shelar if (skb_unclone(skb, GFP_ATOMIC)) 14131da177e4SLinus Torvalds return -ENOMEM; 14141da177e4SLinus Torvalds 14157162fb24SEric Dumazet delta_truesize = __pskb_trim_head(skb, len); 14161da177e4SLinus Torvalds 14171da177e4SLinus Torvalds TCP_SKB_CB(skb)->seq += len; 141884fa7933SPatrick McHardy skb->ip_summed = CHECKSUM_PARTIAL; 14191da177e4SLinus Torvalds 14207162fb24SEric Dumazet if (delta_truesize) { 14217162fb24SEric Dumazet skb->truesize -= delta_truesize; 14227162fb24SEric Dumazet sk->sk_wmem_queued -= delta_truesize; 14237162fb24SEric Dumazet sk_mem_uncharge(sk, delta_truesize); 14241da177e4SLinus Torvalds sock_set_flag(sk, SOCK_QUEUE_SHRUNK); 14257162fb24SEric Dumazet } 14261da177e4SLinus Torvalds 14275b35e1e6SNeal Cardwell /* Any change of skb->len requires recalculation of tso factor. */ 14281da177e4SLinus Torvalds if (tcp_skb_pcount(skb) > 1) 14295bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, tcp_skb_mss(skb)); 14301da177e4SLinus Torvalds 14311da177e4SLinus Torvalds return 0; 14321da177e4SLinus Torvalds } 14331da177e4SLinus Torvalds 14341b63edd6SYuchung Cheng /* Calculate MSS not accounting any TCP options. */ 14351b63edd6SYuchung Cheng static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu) 14365d424d5aSJohn Heffner { 1437cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1438cf533ea5SEric Dumazet const struct inet_connection_sock *icsk = inet_csk(sk); 14395d424d5aSJohn Heffner int mss_now; 14405d424d5aSJohn Heffner 14415d424d5aSJohn Heffner /* Calculate base mss without TCP options: 14425d424d5aSJohn Heffner It is MMS_S - sizeof(tcphdr) of rfc1122 14435d424d5aSJohn Heffner */ 14445d424d5aSJohn Heffner mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr); 14455d424d5aSJohn Heffner 144667469601SEric Dumazet /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */ 144767469601SEric Dumazet if (icsk->icsk_af_ops->net_frag_header_len) { 144867469601SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 144967469601SEric Dumazet 145067469601SEric Dumazet if (dst && dst_allfrag(dst)) 145167469601SEric Dumazet mss_now -= icsk->icsk_af_ops->net_frag_header_len; 145267469601SEric Dumazet } 145367469601SEric Dumazet 14545d424d5aSJohn Heffner /* Clamp it (mss_clamp does not include tcp options) */ 14555d424d5aSJohn Heffner if (mss_now > tp->rx_opt.mss_clamp) 14565d424d5aSJohn Heffner mss_now = tp->rx_opt.mss_clamp; 14575d424d5aSJohn Heffner 14585d424d5aSJohn Heffner /* Now subtract optional transport overhead */ 14595d424d5aSJohn Heffner mss_now -= icsk->icsk_ext_hdr_len; 14605d424d5aSJohn Heffner 14615d424d5aSJohn Heffner /* Then reserve room for full set of TCP options and 8 bytes of data */ 14623b4929f6SEric Dumazet if (mss_now < TCP_MIN_SND_MSS) 14633b4929f6SEric Dumazet mss_now = TCP_MIN_SND_MSS; 14645d424d5aSJohn Heffner return mss_now; 14655d424d5aSJohn Heffner } 14665d424d5aSJohn Heffner 14671b63edd6SYuchung Cheng /* Calculate MSS. Not accounting for SACKs here. */ 14681b63edd6SYuchung Cheng int tcp_mtu_to_mss(struct sock *sk, int pmtu) 14691b63edd6SYuchung Cheng { 14701b63edd6SYuchung Cheng /* Subtract TCP options size, not including SACKs */ 14711b63edd6SYuchung Cheng return __tcp_mtu_to_mss(sk, pmtu) - 14721b63edd6SYuchung Cheng (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr)); 14731b63edd6SYuchung Cheng } 14741b63edd6SYuchung Cheng 14755d424d5aSJohn Heffner /* Inverse of above */ 147667469601SEric Dumazet int tcp_mss_to_mtu(struct sock *sk, int mss) 14775d424d5aSJohn Heffner { 1478cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1479cf533ea5SEric Dumazet const struct inet_connection_sock *icsk = inet_csk(sk); 14805d424d5aSJohn Heffner int mtu; 14815d424d5aSJohn Heffner 14825d424d5aSJohn Heffner mtu = mss + 14835d424d5aSJohn Heffner tp->tcp_header_len + 14845d424d5aSJohn Heffner icsk->icsk_ext_hdr_len + 14855d424d5aSJohn Heffner icsk->icsk_af_ops->net_header_len; 14865d424d5aSJohn Heffner 148767469601SEric Dumazet /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */ 148867469601SEric Dumazet if (icsk->icsk_af_ops->net_frag_header_len) { 148967469601SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 149067469601SEric Dumazet 149167469601SEric Dumazet if (dst && dst_allfrag(dst)) 149267469601SEric Dumazet mtu += icsk->icsk_af_ops->net_frag_header_len; 149367469601SEric Dumazet } 14945d424d5aSJohn Heffner return mtu; 14955d424d5aSJohn Heffner } 1496556c6b46SNeal Cardwell EXPORT_SYMBOL(tcp_mss_to_mtu); 14975d424d5aSJohn Heffner 149867edfef7SAndi Kleen /* MTU probing init per socket */ 14995d424d5aSJohn Heffner void tcp_mtup_init(struct sock *sk) 15005d424d5aSJohn Heffner { 15015d424d5aSJohn Heffner struct tcp_sock *tp = tcp_sk(sk); 15025d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 1503b0f9ca53SFan Du struct net *net = sock_net(sk); 15045d424d5aSJohn Heffner 1505b0f9ca53SFan Du icsk->icsk_mtup.enabled = net->ipv4.sysctl_tcp_mtu_probing > 1; 15065d424d5aSJohn Heffner icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) + 15075d424d5aSJohn Heffner icsk->icsk_af_ops->net_header_len; 1508b0f9ca53SFan Du icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, net->ipv4.sysctl_tcp_base_mss); 15095d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = 0; 151005cbc0dbSFan Du if (icsk->icsk_mtup.enabled) 1511c74df29aSEric Dumazet icsk->icsk_mtup.probe_timestamp = tcp_jiffies32; 15125d424d5aSJohn Heffner } 15134bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_mtup_init); 15145d424d5aSJohn Heffner 15151da177e4SLinus Torvalds /* This function synchronize snd mss to current pmtu/exthdr set. 15161da177e4SLinus Torvalds 15171da177e4SLinus Torvalds tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts 15181da177e4SLinus Torvalds for TCP options, but includes only bare TCP header. 15191da177e4SLinus Torvalds 15201da177e4SLinus Torvalds tp->rx_opt.mss_clamp is mss negotiated at connection setup. 1521caa20d9aSStephen Hemminger It is minimum of user_mss and mss received with SYN. 15221da177e4SLinus Torvalds It also does not include TCP options. 15231da177e4SLinus Torvalds 1524d83d8461SArnaldo Carvalho de Melo inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function. 15251da177e4SLinus Torvalds 15261da177e4SLinus Torvalds tp->mss_cache is current effective sending mss, including 15271da177e4SLinus Torvalds all tcp options except for SACKs. It is evaluated, 15281da177e4SLinus Torvalds taking into account current pmtu, but never exceeds 15291da177e4SLinus Torvalds tp->rx_opt.mss_clamp. 15301da177e4SLinus Torvalds 15311da177e4SLinus Torvalds NOTE1. rfc1122 clearly states that advertised MSS 15321da177e4SLinus Torvalds DOES NOT include either tcp or ip options. 15331da177e4SLinus Torvalds 1534d83d8461SArnaldo Carvalho de Melo NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache 1535d83d8461SArnaldo Carvalho de Melo are READ ONLY outside this function. --ANK (980731) 15361da177e4SLinus Torvalds */ 15371da177e4SLinus Torvalds unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu) 15381da177e4SLinus Torvalds { 15391da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 1540d83d8461SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 15415d424d5aSJohn Heffner int mss_now; 15421da177e4SLinus Torvalds 15435d424d5aSJohn Heffner if (icsk->icsk_mtup.search_high > pmtu) 15445d424d5aSJohn Heffner icsk->icsk_mtup.search_high = pmtu; 15451da177e4SLinus Torvalds 15465d424d5aSJohn Heffner mss_now = tcp_mtu_to_mss(sk, pmtu); 1547409d22b4SIlpo Järvinen mss_now = tcp_bound_to_half_wnd(tp, mss_now); 15481da177e4SLinus Torvalds 15491da177e4SLinus Torvalds /* And store cached results */ 1550d83d8461SArnaldo Carvalho de Melo icsk->icsk_pmtu_cookie = pmtu; 15515d424d5aSJohn Heffner if (icsk->icsk_mtup.enabled) 15525d424d5aSJohn Heffner mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low)); 1553c1b4a7e6SDavid S. Miller tp->mss_cache = mss_now; 15541da177e4SLinus Torvalds 15551da177e4SLinus Torvalds return mss_now; 15561da177e4SLinus Torvalds } 15574bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_sync_mss); 15581da177e4SLinus Torvalds 15591da177e4SLinus Torvalds /* Compute the current effective MSS, taking SACKs and IP options, 15601da177e4SLinus Torvalds * and even PMTU discovery events into account. 15611da177e4SLinus Torvalds */ 15620c54b85fSIlpo Järvinen unsigned int tcp_current_mss(struct sock *sk) 15631da177e4SLinus Torvalds { 1564cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1565cf533ea5SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 1566c1b4a7e6SDavid S. Miller u32 mss_now; 156795c96174SEric Dumazet unsigned int header_len; 156833ad798cSAdam Langley struct tcp_out_options opts; 156933ad798cSAdam Langley struct tcp_md5sig_key *md5; 15701da177e4SLinus Torvalds 1571c1b4a7e6SDavid S. Miller mss_now = tp->mss_cache; 1572c1b4a7e6SDavid S. Miller 15731da177e4SLinus Torvalds if (dst) { 15741da177e4SLinus Torvalds u32 mtu = dst_mtu(dst); 1575d83d8461SArnaldo Carvalho de Melo if (mtu != inet_csk(sk)->icsk_pmtu_cookie) 15761da177e4SLinus Torvalds mss_now = tcp_sync_mss(sk, mtu); 15771da177e4SLinus Torvalds } 15781da177e4SLinus Torvalds 157933ad798cSAdam Langley header_len = tcp_established_options(sk, NULL, &opts, &md5) + 158033ad798cSAdam Langley sizeof(struct tcphdr); 158133ad798cSAdam Langley /* The mss_cache is sized based on tp->tcp_header_len, which assumes 158233ad798cSAdam Langley * some common options. If this is an odd packet (because we have SACK 158333ad798cSAdam Langley * blocks etc) then our calculated header_len will be different, and 158433ad798cSAdam Langley * we have to adjust mss_now correspondingly */ 158533ad798cSAdam Langley if (header_len != tp->tcp_header_len) { 158633ad798cSAdam Langley int delta = (int) header_len - tp->tcp_header_len; 158733ad798cSAdam Langley mss_now -= delta; 158833ad798cSAdam Langley } 1589cfb6eeb4SYOSHIFUJI Hideaki 15901da177e4SLinus Torvalds return mss_now; 15911da177e4SLinus Torvalds } 15921da177e4SLinus Torvalds 159386fd14adSWeiping Pan /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto. 159486fd14adSWeiping Pan * As additional protections, we do not touch cwnd in retransmission phases, 159586fd14adSWeiping Pan * and if application hit its sndbuf limit recently. 159686fd14adSWeiping Pan */ 159786fd14adSWeiping Pan static void tcp_cwnd_application_limited(struct sock *sk) 1598a762a980SDavid S. Miller { 15999e412ba7SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 1600a762a980SDavid S. Miller 160186fd14adSWeiping Pan if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open && 160286fd14adSWeiping Pan sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) { 160386fd14adSWeiping Pan /* Limited by application or receiver window. */ 160486fd14adSWeiping Pan u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk)); 160586fd14adSWeiping Pan u32 win_used = max(tp->snd_cwnd_used, init_win); 160686fd14adSWeiping Pan if (win_used < tp->snd_cwnd) { 160786fd14adSWeiping Pan tp->snd_ssthresh = tcp_current_ssthresh(sk); 160886fd14adSWeiping Pan tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1; 160986fd14adSWeiping Pan } 161086fd14adSWeiping Pan tp->snd_cwnd_used = 0; 161186fd14adSWeiping Pan } 1612c2203cf7SEric Dumazet tp->snd_cwnd_stamp = tcp_jiffies32; 161386fd14adSWeiping Pan } 161486fd14adSWeiping Pan 1615ca8a2263SNeal Cardwell static void tcp_cwnd_validate(struct sock *sk, bool is_cwnd_limited) 1616a762a980SDavid S. Miller { 16171b1fc3fdSWei Wang const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops; 1618a762a980SDavid S. Miller struct tcp_sock *tp = tcp_sk(sk); 1619a762a980SDavid S. Miller 1620ca8a2263SNeal Cardwell /* Track the maximum number of outstanding packets in each 1621ca8a2263SNeal Cardwell * window, and remember whether we were cwnd-limited then. 1622ca8a2263SNeal Cardwell */ 1623ca8a2263SNeal Cardwell if (!before(tp->snd_una, tp->max_packets_seq) || 1624ca8a2263SNeal Cardwell tp->packets_out > tp->max_packets_out) { 1625ca8a2263SNeal Cardwell tp->max_packets_out = tp->packets_out; 1626ca8a2263SNeal Cardwell tp->max_packets_seq = tp->snd_nxt; 1627ca8a2263SNeal Cardwell tp->is_cwnd_limited = is_cwnd_limited; 1628ca8a2263SNeal Cardwell } 1629e114a710SEric Dumazet 163024901551SEric Dumazet if (tcp_is_cwnd_limited(sk)) { 1631a762a980SDavid S. Miller /* Network is feed fully. */ 1632a762a980SDavid S. Miller tp->snd_cwnd_used = 0; 1633c2203cf7SEric Dumazet tp->snd_cwnd_stamp = tcp_jiffies32; 1634a762a980SDavid S. Miller } else { 1635a762a980SDavid S. Miller /* Network starves. */ 1636a762a980SDavid S. Miller if (tp->packets_out > tp->snd_cwnd_used) 1637a762a980SDavid S. Miller tp->snd_cwnd_used = tp->packets_out; 1638a762a980SDavid S. Miller 1639b510f0d2SEric Dumazet if (sock_net(sk)->ipv4.sysctl_tcp_slow_start_after_idle && 1640c2203cf7SEric Dumazet (s32)(tcp_jiffies32 - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto && 16411b1fc3fdSWei Wang !ca_ops->cong_control) 1642a762a980SDavid S. Miller tcp_cwnd_application_limited(sk); 1643b0f71bd3SFrancis Yan 1644b0f71bd3SFrancis Yan /* The following conditions together indicate the starvation 1645b0f71bd3SFrancis Yan * is caused by insufficient sender buffer: 1646b0f71bd3SFrancis Yan * 1) just sent some data (see tcp_write_xmit) 1647b0f71bd3SFrancis Yan * 2) not cwnd limited (this else condition) 164875c119afSEric Dumazet * 3) no more data to send (tcp_write_queue_empty()) 1649b0f71bd3SFrancis Yan * 4) application is hitting buffer limit (SOCK_NOSPACE) 1650b0f71bd3SFrancis Yan */ 165175c119afSEric Dumazet if (tcp_write_queue_empty(sk) && sk->sk_socket && 1652b0f71bd3SFrancis Yan test_bit(SOCK_NOSPACE, &sk->sk_socket->flags) && 1653b0f71bd3SFrancis Yan (1 << sk->sk_state) & (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) 1654b0f71bd3SFrancis Yan tcp_chrono_start(sk, TCP_CHRONO_SNDBUF_LIMITED); 1655a762a980SDavid S. Miller } 1656a762a980SDavid S. Miller } 1657a762a980SDavid S. Miller 1658d4589926SEric Dumazet /* Minshall's variant of the Nagle send check. */ 1659d4589926SEric Dumazet static bool tcp_minshall_check(const struct tcp_sock *tp) 1660d4589926SEric Dumazet { 1661d4589926SEric Dumazet return after(tp->snd_sml, tp->snd_una) && 1662d4589926SEric Dumazet !after(tp->snd_sml, tp->snd_nxt); 1663d4589926SEric Dumazet } 1664d4589926SEric Dumazet 1665d4589926SEric Dumazet /* Update snd_sml if this skb is under mss 1666d4589926SEric Dumazet * Note that a TSO packet might end with a sub-mss segment 1667d4589926SEric Dumazet * The test is really : 1668d4589926SEric Dumazet * if ((skb->len % mss) != 0) 1669d4589926SEric Dumazet * tp->snd_sml = TCP_SKB_CB(skb)->end_seq; 1670d4589926SEric Dumazet * But we can avoid doing the divide again given we already have 1671d4589926SEric Dumazet * skb_pcount = skb->len / mss_now 16720e3a4803SIlpo Järvinen */ 1673d4589926SEric Dumazet static void tcp_minshall_update(struct tcp_sock *tp, unsigned int mss_now, 1674d4589926SEric Dumazet const struct sk_buff *skb) 1675d4589926SEric Dumazet { 1676d4589926SEric Dumazet if (skb->len < tcp_skb_pcount(skb) * mss_now) 1677d4589926SEric Dumazet tp->snd_sml = TCP_SKB_CB(skb)->end_seq; 1678d4589926SEric Dumazet } 1679d4589926SEric Dumazet 1680d4589926SEric Dumazet /* Return false, if packet can be sent now without violation Nagle's rules: 1681d4589926SEric Dumazet * 1. It is full sized. (provided by caller in %partial bool) 1682d4589926SEric Dumazet * 2. Or it contains FIN. (already checked by caller) 1683d4589926SEric Dumazet * 3. Or TCP_CORK is not set, and TCP_NODELAY is set. 1684d4589926SEric Dumazet * 4. Or TCP_CORK is not set, and all sent packets are ACKed. 1685d4589926SEric Dumazet * With Minshall's modification: all sent small packets are ACKed. 1686d4589926SEric Dumazet */ 1687d4589926SEric Dumazet static bool tcp_nagle_check(bool partial, const struct tcp_sock *tp, 1688cc93fc51SPeter Pan(潘卫平) int nonagle) 1689d4589926SEric Dumazet { 1690d4589926SEric Dumazet return partial && 1691d4589926SEric Dumazet ((nonagle & TCP_NAGLE_CORK) || 1692d4589926SEric Dumazet (!nonagle && tp->packets_out && tcp_minshall_check(tp))); 1693d4589926SEric Dumazet } 1694605ad7f1SEric Dumazet 1695605ad7f1SEric Dumazet /* Return how many segs we'd like on a TSO packet, 1696605ad7f1SEric Dumazet * to send one TSO packet per ms 1697605ad7f1SEric Dumazet */ 1698dcb8c9b4SEric Dumazet static u32 tcp_tso_autosize(const struct sock *sk, unsigned int mss_now, 16991b3878caSNeal Cardwell int min_tso_segs) 1700605ad7f1SEric Dumazet { 1701605ad7f1SEric Dumazet u32 bytes, segs; 1702605ad7f1SEric Dumazet 170376a9ebe8SEric Dumazet bytes = min_t(unsigned long, 170476a9ebe8SEric Dumazet sk->sk_pacing_rate >> sk->sk_pacing_shift, 1705605ad7f1SEric Dumazet sk->sk_gso_max_size - 1 - MAX_TCP_HEADER); 1706605ad7f1SEric Dumazet 1707605ad7f1SEric Dumazet /* Goal is to send at least one packet per ms, 1708605ad7f1SEric Dumazet * not one big TSO packet every 100 ms. 1709605ad7f1SEric Dumazet * This preserves ACK clocking and is consistent 1710605ad7f1SEric Dumazet * with tcp_tso_should_defer() heuristic. 1711605ad7f1SEric Dumazet */ 17121b3878caSNeal Cardwell segs = max_t(u32, bytes / mss_now, min_tso_segs); 1713605ad7f1SEric Dumazet 1714350c9f48SEric Dumazet return segs; 1715605ad7f1SEric Dumazet } 1716605ad7f1SEric Dumazet 1717ed6e7268SNeal Cardwell /* Return the number of segments we want in the skb we are transmitting. 1718ed6e7268SNeal Cardwell * See if congestion control module wants to decide; otherwise, autosize. 1719ed6e7268SNeal Cardwell */ 1720ed6e7268SNeal Cardwell static u32 tcp_tso_segs(struct sock *sk, unsigned int mss_now) 1721ed6e7268SNeal Cardwell { 1722ed6e7268SNeal Cardwell const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops; 1723dcb8c9b4SEric Dumazet u32 min_tso, tso_segs; 1724ed6e7268SNeal Cardwell 1725dcb8c9b4SEric Dumazet min_tso = ca_ops->min_tso_segs ? 1726dcb8c9b4SEric Dumazet ca_ops->min_tso_segs(sk) : 1727dcb8c9b4SEric Dumazet sock_net(sk)->ipv4.sysctl_tcp_min_tso_segs; 1728dcb8c9b4SEric Dumazet 1729dcb8c9b4SEric Dumazet tso_segs = tcp_tso_autosize(sk, mss_now, min_tso); 1730350c9f48SEric Dumazet return min_t(u32, tso_segs, sk->sk_gso_max_segs); 1731ed6e7268SNeal Cardwell } 1732ed6e7268SNeal Cardwell 1733d4589926SEric Dumazet /* Returns the portion of skb which can be sent right away */ 1734d4589926SEric Dumazet static unsigned int tcp_mss_split_point(const struct sock *sk, 1735d4589926SEric Dumazet const struct sk_buff *skb, 1736d4589926SEric Dumazet unsigned int mss_now, 1737d4589926SEric Dumazet unsigned int max_segs, 1738d4589926SEric Dumazet int nonagle) 1739c1b4a7e6SDavid S. Miller { 1740cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1741d4589926SEric Dumazet u32 partial, needed, window, max_len; 1742c1b4a7e6SDavid S. Miller 174390840defSIlpo Järvinen window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 17441485348dSBen Hutchings max_len = mss_now * max_segs; 17450e3a4803SIlpo Järvinen 17461485348dSBen Hutchings if (likely(max_len <= window && skb != tcp_write_queue_tail(sk))) 17471485348dSBen Hutchings return max_len; 17480e3a4803SIlpo Järvinen 17495ea3a748SIlpo Järvinen needed = min(skb->len, window); 17505ea3a748SIlpo Järvinen 17511485348dSBen Hutchings if (max_len <= needed) 17521485348dSBen Hutchings return max_len; 17530e3a4803SIlpo Järvinen 1754d4589926SEric Dumazet partial = needed % mss_now; 1755d4589926SEric Dumazet /* If last segment is not a full MSS, check if Nagle rules allow us 1756d4589926SEric Dumazet * to include this last segment in this skb. 1757d4589926SEric Dumazet * Otherwise, we'll split the skb at last MSS boundary 1758d4589926SEric Dumazet */ 1759cc93fc51SPeter Pan(潘卫平) if (tcp_nagle_check(partial != 0, tp, nonagle)) 1760d4589926SEric Dumazet return needed - partial; 1761d4589926SEric Dumazet 1762d4589926SEric Dumazet return needed; 1763c1b4a7e6SDavid S. Miller } 1764c1b4a7e6SDavid S. Miller 1765c1b4a7e6SDavid S. Miller /* Can at least one segment of SKB be sent right now, according to the 1766c1b4a7e6SDavid S. Miller * congestion window rules? If so, return how many segments are allowed. 1767c1b4a7e6SDavid S. Miller */ 1768cf533ea5SEric Dumazet static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp, 1769cf533ea5SEric Dumazet const struct sk_buff *skb) 1770c1b4a7e6SDavid S. Miller { 1771d649a7a8SEric Dumazet u32 in_flight, cwnd, halfcwnd; 1772c1b4a7e6SDavid S. Miller 1773c1b4a7e6SDavid S. Miller /* Don't be strict about the congestion window for the final FIN. */ 17744de075e0SEric Dumazet if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) && 17754de075e0SEric Dumazet tcp_skb_pcount(skb) == 1) 1776c1b4a7e6SDavid S. Miller return 1; 1777c1b4a7e6SDavid S. Miller 1778c1b4a7e6SDavid S. Miller in_flight = tcp_packets_in_flight(tp); 1779c1b4a7e6SDavid S. Miller cwnd = tp->snd_cwnd; 1780d649a7a8SEric Dumazet if (in_flight >= cwnd) 1781c1b4a7e6SDavid S. Miller return 0; 1782d649a7a8SEric Dumazet 1783d649a7a8SEric Dumazet /* For better scheduling, ensure we have at least 1784d649a7a8SEric Dumazet * 2 GSO packets in flight. 1785d649a7a8SEric Dumazet */ 1786d649a7a8SEric Dumazet halfcwnd = max(cwnd >> 1, 1U); 1787d649a7a8SEric Dumazet return min(halfcwnd, cwnd - in_flight); 1788c1b4a7e6SDavid S. Miller } 1789c1b4a7e6SDavid S. Miller 1790b595076aSUwe Kleine-König /* Initialize TSO state of a skb. 179167edfef7SAndi Kleen * This must be invoked the first time we consider transmitting 1792c1b4a7e6SDavid S. Miller * SKB onto the wire. 1793c1b4a7e6SDavid S. Miller */ 17945bbb432cSEric Dumazet static int tcp_init_tso_segs(struct sk_buff *skb, unsigned int mss_now) 1795c1b4a7e6SDavid S. Miller { 1796c1b4a7e6SDavid S. Miller int tso_segs = tcp_skb_pcount(skb); 1797c1b4a7e6SDavid S. Miller 1798f8269a49SIlpo Järvinen if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) { 17995bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss_now); 1800c1b4a7e6SDavid S. Miller tso_segs = tcp_skb_pcount(skb); 1801c1b4a7e6SDavid S. Miller } 1802c1b4a7e6SDavid S. Miller return tso_segs; 1803c1b4a7e6SDavid S. Miller } 1804c1b4a7e6SDavid S. Miller 1805c1b4a7e6SDavid S. Miller 1806a2a385d6SEric Dumazet /* Return true if the Nagle test allows this packet to be 1807c1b4a7e6SDavid S. Miller * sent now. 1808c1b4a7e6SDavid S. Miller */ 1809a2a385d6SEric Dumazet static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb, 1810c1b4a7e6SDavid S. Miller unsigned int cur_mss, int nonagle) 1811c1b4a7e6SDavid S. Miller { 1812c1b4a7e6SDavid S. Miller /* Nagle rule does not apply to frames, which sit in the middle of the 1813c1b4a7e6SDavid S. Miller * write_queue (they have no chances to get new data). 1814c1b4a7e6SDavid S. Miller * 1815c1b4a7e6SDavid S. Miller * This is implemented in the callers, where they modify the 'nonagle' 1816c1b4a7e6SDavid S. Miller * argument based upon the location of SKB in the send queue. 1817c1b4a7e6SDavid S. Miller */ 1818c1b4a7e6SDavid S. Miller if (nonagle & TCP_NAGLE_PUSH) 1819a2a385d6SEric Dumazet return true; 1820c1b4a7e6SDavid S. Miller 18219b44190dSYuchung Cheng /* Don't use the nagle rule for urgent data (or for the final FIN). */ 18229b44190dSYuchung Cheng if (tcp_urg_mode(tp) || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) 1823a2a385d6SEric Dumazet return true; 1824c1b4a7e6SDavid S. Miller 1825cc93fc51SPeter Pan(潘卫平) if (!tcp_nagle_check(skb->len < cur_mss, tp, nonagle)) 1826a2a385d6SEric Dumazet return true; 1827c1b4a7e6SDavid S. Miller 1828a2a385d6SEric Dumazet return false; 1829c1b4a7e6SDavid S. Miller } 1830c1b4a7e6SDavid S. Miller 1831c1b4a7e6SDavid S. Miller /* Does at least the first segment of SKB fit into the send window? */ 1832a2a385d6SEric Dumazet static bool tcp_snd_wnd_test(const struct tcp_sock *tp, 1833a2a385d6SEric Dumazet const struct sk_buff *skb, 1834056834d9SIlpo Järvinen unsigned int cur_mss) 1835c1b4a7e6SDavid S. Miller { 1836c1b4a7e6SDavid S. Miller u32 end_seq = TCP_SKB_CB(skb)->end_seq; 1837c1b4a7e6SDavid S. Miller 1838c1b4a7e6SDavid S. Miller if (skb->len > cur_mss) 1839c1b4a7e6SDavid S. Miller end_seq = TCP_SKB_CB(skb)->seq + cur_mss; 1840c1b4a7e6SDavid S. Miller 184190840defSIlpo Järvinen return !after(end_seq, tcp_wnd_end(tp)); 1842c1b4a7e6SDavid S. Miller } 1843c1b4a7e6SDavid S. Miller 1844c1b4a7e6SDavid S. Miller /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet 1845c1b4a7e6SDavid S. Miller * which is put after SKB on the list. It is very much like 1846c1b4a7e6SDavid S. Miller * tcp_fragment() except that it may make several kinds of assumptions 1847c1b4a7e6SDavid S. Miller * in order to speed up the splitting operation. In particular, we 1848c1b4a7e6SDavid S. Miller * know that all the data is in scatter-gather pages, and that the 1849c1b4a7e6SDavid S. Miller * packet has never been sent out before (and thus is not cloned). 1850c1b4a7e6SDavid S. Miller */ 185156483341SEric Dumazet static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len, 1852c4ead4c5SEric Dumazet unsigned int mss_now, gfp_t gfp) 1853c1b4a7e6SDavid S. Miller { 1854c1b4a7e6SDavid S. Miller int nlen = skb->len - len; 185556483341SEric Dumazet struct sk_buff *buff; 18569ce01461SIlpo Järvinen u8 flags; 1857c1b4a7e6SDavid S. Miller 1858c1b4a7e6SDavid S. Miller /* All of a TSO frame must be composed of paged data. */ 1859c8ac3774SHerbert Xu if (skb->len != skb->data_len) 186056483341SEric Dumazet return tcp_fragment(sk, TCP_FRAG_IN_WRITE_QUEUE, 186156483341SEric Dumazet skb, len, mss_now, gfp); 1862c1b4a7e6SDavid S. Miller 1863eb934478SEric Dumazet buff = sk_stream_alloc_skb(sk, 0, gfp, true); 186451456b29SIan Morris if (unlikely(!buff)) 1865c1b4a7e6SDavid S. Miller return -ENOMEM; 1866c1b4a7e6SDavid S. Miller 18673ab224beSHideo Aoki sk->sk_wmem_queued += buff->truesize; 18683ab224beSHideo Aoki sk_mem_charge(sk, buff->truesize); 1869b60b49eaSHerbert Xu buff->truesize += nlen; 1870c1b4a7e6SDavid S. Miller skb->truesize -= nlen; 1871c1b4a7e6SDavid S. Miller 1872c1b4a7e6SDavid S. Miller /* Correct the sequence numbers. */ 1873c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; 1874c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; 1875c1b4a7e6SDavid S. Miller TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; 1876c1b4a7e6SDavid S. Miller 1877c1b4a7e6SDavid S. Miller /* PSH and FIN should only be set in the second packet. */ 18784de075e0SEric Dumazet flags = TCP_SKB_CB(skb)->tcp_flags; 18794de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); 18804de075e0SEric Dumazet TCP_SKB_CB(buff)->tcp_flags = flags; 1881c1b4a7e6SDavid S. Miller 1882c1b4a7e6SDavid S. Miller /* This packet was never sent out yet, so no SACK bits. */ 1883c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->sacked = 0; 1884c1b4a7e6SDavid S. Miller 1885a166140eSMartin KaFai Lau tcp_skb_fragment_eor(skb, buff); 1886a166140eSMartin KaFai Lau 188798be9b12SEric Dumazet buff->ip_summed = CHECKSUM_PARTIAL; 1888c1b4a7e6SDavid S. Miller skb_split(skb, buff, len); 1889490cc7d0SWillem de Bruijn tcp_fragment_tstamp(skb, buff); 1890c1b4a7e6SDavid S. Miller 1891c1b4a7e6SDavid S. Miller /* Fix up tso_factor for both original and new SKB. */ 18925bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss_now); 18935bbb432cSEric Dumazet tcp_set_skb_tso_segs(buff, mss_now); 1894c1b4a7e6SDavid S. Miller 1895c1b4a7e6SDavid S. Miller /* Link BUFF into the send queue. */ 1896f4a775d1SEric Dumazet __skb_header_release(buff); 189756483341SEric Dumazet tcp_insert_write_queue_after(skb, buff, sk, TCP_FRAG_IN_WRITE_QUEUE); 1898c1b4a7e6SDavid S. Miller 1899c1b4a7e6SDavid S. Miller return 0; 1900c1b4a7e6SDavid S. Miller } 1901c1b4a7e6SDavid S. Miller 1902c1b4a7e6SDavid S. Miller /* Try to defer sending, if possible, in order to minimize the amount 1903c1b4a7e6SDavid S. Miller * of TSO splitting we do. View it as a kind of TSO Nagle test. 1904c1b4a7e6SDavid S. Miller * 1905c1b4a7e6SDavid S. Miller * This algorithm is from John Heffner. 1906c1b4a7e6SDavid S. Miller */ 1907ca8a2263SNeal Cardwell static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb, 1908f9bfe4e6SEric Dumazet bool *is_cwnd_limited, 1909f9bfe4e6SEric Dumazet bool *is_rwnd_limited, 1910f9bfe4e6SEric Dumazet u32 max_segs) 1911c1b4a7e6SDavid S. Miller { 19126687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 1913f1c6ea38SEric Dumazet u32 send_win, cong_win, limit, in_flight; 191450c8339eSEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 191550c8339eSEric Dumazet struct sk_buff *head; 1916ad9f4f50SEric Dumazet int win_divisor; 1917f1c6ea38SEric Dumazet s64 delta; 1918c1b4a7e6SDavid S. Miller 191999d7662aSEric Dumazet if (icsk->icsk_ca_state >= TCP_CA_Recovery) 1920ae8064acSJohn Heffner goto send_now; 1921ae8064acSJohn Heffner 19225f852eb5SEric Dumazet /* Avoid bursty behavior by allowing defer 1923a682850aSEric Dumazet * only if the last write was recent (1 ms). 1924a682850aSEric Dumazet * Note that tp->tcp_wstamp_ns can be in the future if we have 1925a682850aSEric Dumazet * packets waiting in a qdisc or device for EDT delivery. 19265f852eb5SEric Dumazet */ 1927a682850aSEric Dumazet delta = tp->tcp_clock_cache - tp->tcp_wstamp_ns - NSEC_PER_MSEC; 1928a682850aSEric Dumazet if (delta > 0) 1929ae8064acSJohn Heffner goto send_now; 1930908a75c1SDavid S. Miller 1931c1b4a7e6SDavid S. Miller in_flight = tcp_packets_in_flight(tp); 1932c1b4a7e6SDavid S. Miller 1933c8c9aeb5SStefano Brivio BUG_ON(tcp_skb_pcount(skb) <= 1); 1934c8c9aeb5SStefano Brivio BUG_ON(tp->snd_cwnd <= in_flight); 1935c1b4a7e6SDavid S. Miller 193690840defSIlpo Järvinen send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 1937c1b4a7e6SDavid S. Miller 1938c1b4a7e6SDavid S. Miller /* From in_flight test above, we know that cwnd > in_flight. */ 1939c1b4a7e6SDavid S. Miller cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache; 1940c1b4a7e6SDavid S. Miller 1941c1b4a7e6SDavid S. Miller limit = min(send_win, cong_win); 1942c1b4a7e6SDavid S. Miller 1943ba244fe9SDavid S. Miller /* If a full-sized TSO skb can be sent, do it. */ 1944605ad7f1SEric Dumazet if (limit >= max_segs * tp->mss_cache) 1945ae8064acSJohn Heffner goto send_now; 1946ba244fe9SDavid S. Miller 194762ad2761SIlpo Järvinen /* Middle in queue won't get any more data, full sendable already? */ 194862ad2761SIlpo Järvinen if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len)) 194962ad2761SIlpo Järvinen goto send_now; 195062ad2761SIlpo Järvinen 19515bbcc0f5SLinus Torvalds win_divisor = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_tso_win_divisor); 1952ad9f4f50SEric Dumazet if (win_divisor) { 1953c1b4a7e6SDavid S. Miller u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache); 1954c1b4a7e6SDavid S. Miller 1955c1b4a7e6SDavid S. Miller /* If at least some fraction of a window is available, 1956c1b4a7e6SDavid S. Miller * just use it. 1957c1b4a7e6SDavid S. Miller */ 1958ad9f4f50SEric Dumazet chunk /= win_divisor; 1959c1b4a7e6SDavid S. Miller if (limit >= chunk) 1960ae8064acSJohn Heffner goto send_now; 1961c1b4a7e6SDavid S. Miller } else { 1962c1b4a7e6SDavid S. Miller /* Different approach, try not to defer past a single 1963c1b4a7e6SDavid S. Miller * ACK. Receiver should ACK every other full sized 1964c1b4a7e6SDavid S. Miller * frame, so if we have space for more than 3 frames 1965c1b4a7e6SDavid S. Miller * then send now. 1966c1b4a7e6SDavid S. Miller */ 19676b5a5c0dSNeal Cardwell if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache) 1968ae8064acSJohn Heffner goto send_now; 1969c1b4a7e6SDavid S. Miller } 1970c1b4a7e6SDavid S. Miller 197175c119afSEric Dumazet /* TODO : use tsorted_sent_queue ? */ 197275c119afSEric Dumazet head = tcp_rtx_queue_head(sk); 197375c119afSEric Dumazet if (!head) 197475c119afSEric Dumazet goto send_now; 1975f1c6ea38SEric Dumazet delta = tp->tcp_clock_cache - head->tstamp; 197650c8339eSEric Dumazet /* If next ACK is likely to come too late (half srtt), do not defer */ 1977f1c6ea38SEric Dumazet if ((s64)(delta - (u64)NSEC_PER_USEC * (tp->srtt_us >> 4)) < 0) 197850c8339eSEric Dumazet goto send_now; 197950c8339eSEric Dumazet 1980f9bfe4e6SEric Dumazet /* Ok, it looks like it is advisable to defer. 1981f9bfe4e6SEric Dumazet * Three cases are tracked : 1982f9bfe4e6SEric Dumazet * 1) We are cwnd-limited 1983f9bfe4e6SEric Dumazet * 2) We are rwnd-limited 1984f9bfe4e6SEric Dumazet * 3) We are application limited. 1985f9bfe4e6SEric Dumazet */ 1986f9bfe4e6SEric Dumazet if (cong_win < send_win) { 1987f9bfe4e6SEric Dumazet if (cong_win <= skb->len) { 1988ca8a2263SNeal Cardwell *is_cwnd_limited = true; 1989f9bfe4e6SEric Dumazet return true; 1990f9bfe4e6SEric Dumazet } 1991f9bfe4e6SEric Dumazet } else { 1992f9bfe4e6SEric Dumazet if (send_win <= skb->len) { 1993f9bfe4e6SEric Dumazet *is_rwnd_limited = true; 1994f9bfe4e6SEric Dumazet return true; 1995f9bfe4e6SEric Dumazet } 1996f9bfe4e6SEric Dumazet } 1997f9bfe4e6SEric Dumazet 1998f9bfe4e6SEric Dumazet /* If this packet won't get more data, do not wait. */ 1999d8ed257fSEric Dumazet if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) || 2000d8ed257fSEric Dumazet TCP_SKB_CB(skb)->eor) 2001f9bfe4e6SEric Dumazet goto send_now; 2002ca8a2263SNeal Cardwell 2003a2a385d6SEric Dumazet return true; 2004ae8064acSJohn Heffner 2005ae8064acSJohn Heffner send_now: 2006a2a385d6SEric Dumazet return false; 2007c1b4a7e6SDavid S. Miller } 2008c1b4a7e6SDavid S. Miller 200905cbc0dbSFan Du static inline void tcp_mtu_check_reprobe(struct sock *sk) 201005cbc0dbSFan Du { 201105cbc0dbSFan Du struct inet_connection_sock *icsk = inet_csk(sk); 201205cbc0dbSFan Du struct tcp_sock *tp = tcp_sk(sk); 201305cbc0dbSFan Du struct net *net = sock_net(sk); 201405cbc0dbSFan Du u32 interval; 201505cbc0dbSFan Du s32 delta; 201605cbc0dbSFan Du 201705cbc0dbSFan Du interval = net->ipv4.sysctl_tcp_probe_interval; 2018c74df29aSEric Dumazet delta = tcp_jiffies32 - icsk->icsk_mtup.probe_timestamp; 201905cbc0dbSFan Du if (unlikely(delta >= interval * HZ)) { 202005cbc0dbSFan Du int mss = tcp_current_mss(sk); 202105cbc0dbSFan Du 202205cbc0dbSFan Du /* Update current search range */ 202305cbc0dbSFan Du icsk->icsk_mtup.probe_size = 0; 202405cbc0dbSFan Du icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + 202505cbc0dbSFan Du sizeof(struct tcphdr) + 202605cbc0dbSFan Du icsk->icsk_af_ops->net_header_len; 202705cbc0dbSFan Du icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss); 202805cbc0dbSFan Du 202905cbc0dbSFan Du /* Update probe time stamp */ 2030c74df29aSEric Dumazet icsk->icsk_mtup.probe_timestamp = tcp_jiffies32; 203105cbc0dbSFan Du } 203205cbc0dbSFan Du } 203305cbc0dbSFan Du 2034808cf9e3SIlya Lesokhin static bool tcp_can_coalesce_send_queue_head(struct sock *sk, int len) 2035808cf9e3SIlya Lesokhin { 2036808cf9e3SIlya Lesokhin struct sk_buff *skb, *next; 2037808cf9e3SIlya Lesokhin 2038808cf9e3SIlya Lesokhin skb = tcp_send_head(sk); 2039808cf9e3SIlya Lesokhin tcp_for_write_queue_from_safe(skb, next, sk) { 2040808cf9e3SIlya Lesokhin if (len <= skb->len) 2041808cf9e3SIlya Lesokhin break; 2042808cf9e3SIlya Lesokhin 2043808cf9e3SIlya Lesokhin if (unlikely(TCP_SKB_CB(skb)->eor)) 2044808cf9e3SIlya Lesokhin return false; 2045808cf9e3SIlya Lesokhin 2046808cf9e3SIlya Lesokhin len -= skb->len; 2047808cf9e3SIlya Lesokhin } 2048808cf9e3SIlya Lesokhin 2049808cf9e3SIlya Lesokhin return true; 2050808cf9e3SIlya Lesokhin } 2051808cf9e3SIlya Lesokhin 20525d424d5aSJohn Heffner /* Create a new MTU probe if we are ready. 205367edfef7SAndi Kleen * MTU probe is regularly attempting to increase the path MTU by 205467edfef7SAndi Kleen * deliberately sending larger packets. This discovers routing 205567edfef7SAndi Kleen * changes resulting in larger path MTUs. 205667edfef7SAndi Kleen * 20575d424d5aSJohn Heffner * Returns 0 if we should wait to probe (no cwnd available), 20585d424d5aSJohn Heffner * 1 if a probe was sent, 2059056834d9SIlpo Järvinen * -1 otherwise 2060056834d9SIlpo Järvinen */ 20615d424d5aSJohn Heffner static int tcp_mtu_probe(struct sock *sk) 20625d424d5aSJohn Heffner { 20635d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 206412a59abcSEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 20655d424d5aSJohn Heffner struct sk_buff *skb, *nskb, *next; 20666b58e0a5SFan Du struct net *net = sock_net(sk); 20675d424d5aSJohn Heffner int probe_size; 206891cc17c0SIlpo Järvinen int size_needed; 206912a59abcSEric Dumazet int copy, len; 20705d424d5aSJohn Heffner int mss_now; 20716b58e0a5SFan Du int interval; 20725d424d5aSJohn Heffner 20735d424d5aSJohn Heffner /* Not currently probing/verifying, 20745d424d5aSJohn Heffner * not in recovery, 20755d424d5aSJohn Heffner * have enough cwnd, and 207612a59abcSEric Dumazet * not SACKing (the variable headers throw things off) 207712a59abcSEric Dumazet */ 207812a59abcSEric Dumazet if (likely(!icsk->icsk_mtup.enabled || 20795d424d5aSJohn Heffner icsk->icsk_mtup.probe_size || 20805d424d5aSJohn Heffner inet_csk(sk)->icsk_ca_state != TCP_CA_Open || 20815d424d5aSJohn Heffner tp->snd_cwnd < 11 || 208212a59abcSEric Dumazet tp->rx_opt.num_sacks || tp->rx_opt.dsack)) 20835d424d5aSJohn Heffner return -1; 20845d424d5aSJohn Heffner 20856b58e0a5SFan Du /* Use binary search for probe_size between tcp_mss_base, 20866b58e0a5SFan Du * and current mss_clamp. if (search_high - search_low) 20876b58e0a5SFan Du * smaller than a threshold, backoff from probing. 20886b58e0a5SFan Du */ 20890c54b85fSIlpo Järvinen mss_now = tcp_current_mss(sk); 20906b58e0a5SFan Du probe_size = tcp_mtu_to_mss(sk, (icsk->icsk_mtup.search_high + 20916b58e0a5SFan Du icsk->icsk_mtup.search_low) >> 1); 209291cc17c0SIlpo Järvinen size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache; 20936b58e0a5SFan Du interval = icsk->icsk_mtup.search_high - icsk->icsk_mtup.search_low; 209405cbc0dbSFan Du /* When misfortune happens, we are reprobing actively, 209505cbc0dbSFan Du * and then reprobe timer has expired. We stick with current 209605cbc0dbSFan Du * probing process by not resetting search range to its orignal. 209705cbc0dbSFan Du */ 20986b58e0a5SFan Du if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high) || 209905cbc0dbSFan Du interval < net->ipv4.sysctl_tcp_probe_threshold) { 210005cbc0dbSFan Du /* Check whether enough time has elaplased for 210105cbc0dbSFan Du * another round of probing. 210205cbc0dbSFan Du */ 210305cbc0dbSFan Du tcp_mtu_check_reprobe(sk); 21045d424d5aSJohn Heffner return -1; 21055d424d5aSJohn Heffner } 21065d424d5aSJohn Heffner 21075d424d5aSJohn Heffner /* Have enough data in the send queue to probe? */ 21087f9c33e5SIlpo Järvinen if (tp->write_seq - tp->snd_nxt < size_needed) 21095d424d5aSJohn Heffner return -1; 21105d424d5aSJohn Heffner 211191cc17c0SIlpo Järvinen if (tp->snd_wnd < size_needed) 21125d424d5aSJohn Heffner return -1; 211390840defSIlpo Järvinen if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp))) 21145d424d5aSJohn Heffner return 0; 21155d424d5aSJohn Heffner 2116d67c58e9SIlpo Järvinen /* Do we need to wait to drain cwnd? With none in flight, don't stall */ 2117d67c58e9SIlpo Järvinen if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) { 2118d67c58e9SIlpo Järvinen if (!tcp_packets_in_flight(tp)) 21195d424d5aSJohn Heffner return -1; 21205d424d5aSJohn Heffner else 21215d424d5aSJohn Heffner return 0; 21225d424d5aSJohn Heffner } 21235d424d5aSJohn Heffner 2124808cf9e3SIlya Lesokhin if (!tcp_can_coalesce_send_queue_head(sk, probe_size)) 2125808cf9e3SIlya Lesokhin return -1; 2126808cf9e3SIlya Lesokhin 21275d424d5aSJohn Heffner /* We're allowed to probe. Build it now. */ 2128eb934478SEric Dumazet nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC, false); 212951456b29SIan Morris if (!nskb) 21305d424d5aSJohn Heffner return -1; 21313ab224beSHideo Aoki sk->sk_wmem_queued += nskb->truesize; 21323ab224beSHideo Aoki sk_mem_charge(sk, nskb->truesize); 21335d424d5aSJohn Heffner 2134fe067e8aSDavid S. Miller skb = tcp_send_head(sk); 21355d424d5aSJohn Heffner 21365d424d5aSJohn Heffner TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq; 21375d424d5aSJohn Heffner TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size; 21384de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK; 21395d424d5aSJohn Heffner TCP_SKB_CB(nskb)->sacked = 0; 21405d424d5aSJohn Heffner nskb->csum = 0; 214198be9b12SEric Dumazet nskb->ip_summed = CHECKSUM_PARTIAL; 21425d424d5aSJohn Heffner 214350c4817eSIlpo Järvinen tcp_insert_write_queue_before(nskb, skb, sk); 21442b7cda9cSEric Dumazet tcp_highest_sack_replace(sk, skb, nskb); 214550c4817eSIlpo Järvinen 21465d424d5aSJohn Heffner len = 0; 2147234b6860SIlpo Järvinen tcp_for_write_queue_from_safe(skb, next, sk) { 21485d424d5aSJohn Heffner copy = min_t(int, skb->len, probe_size - len); 21495d424d5aSJohn Heffner skb_copy_bits(skb, 0, skb_put(nskb, copy), copy); 21505d424d5aSJohn Heffner 21515d424d5aSJohn Heffner if (skb->len <= copy) { 21525d424d5aSJohn Heffner /* We've eaten all the data from this skb. 21535d424d5aSJohn Heffner * Throw it away. */ 21544de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags; 2155808cf9e3SIlya Lesokhin /* If this is the last SKB we copy and eor is set 2156808cf9e3SIlya Lesokhin * we need to propagate it to the new skb. 2157808cf9e3SIlya Lesokhin */ 2158808cf9e3SIlya Lesokhin TCP_SKB_CB(nskb)->eor = TCP_SKB_CB(skb)->eor; 2159fe067e8aSDavid S. Miller tcp_unlink_write_queue(skb, sk); 21603ab224beSHideo Aoki sk_wmem_free_skb(sk, skb); 21615d424d5aSJohn Heffner } else { 21624de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags & 2163a3433f35SChangli Gao ~(TCPHDR_FIN|TCPHDR_PSH); 21645d424d5aSJohn Heffner if (!skb_shinfo(skb)->nr_frags) { 21655d424d5aSJohn Heffner skb_pull(skb, copy); 21665d424d5aSJohn Heffner } else { 21675d424d5aSJohn Heffner __pskb_trim_head(skb, copy); 21685bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss_now); 21695d424d5aSJohn Heffner } 21705d424d5aSJohn Heffner TCP_SKB_CB(skb)->seq += copy; 21715d424d5aSJohn Heffner } 21725d424d5aSJohn Heffner 21735d424d5aSJohn Heffner len += copy; 2174234b6860SIlpo Järvinen 2175234b6860SIlpo Järvinen if (len >= probe_size) 2176234b6860SIlpo Järvinen break; 21775d424d5aSJohn Heffner } 21785bbb432cSEric Dumazet tcp_init_tso_segs(nskb, nskb->len); 21795d424d5aSJohn Heffner 21805d424d5aSJohn Heffner /* We're ready to send. If this fails, the probe will 21817faee5c0SEric Dumazet * be resegmented into mss-sized pieces by tcp_write_xmit(). 21827faee5c0SEric Dumazet */ 21835d424d5aSJohn Heffner if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) { 21845d424d5aSJohn Heffner /* Decrement cwnd here because we are sending 21855d424d5aSJohn Heffner * effectively two packets. */ 21865d424d5aSJohn Heffner tp->snd_cwnd--; 218766f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, nskb); 21885d424d5aSJohn Heffner 21895d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len); 21900e7b1368SJohn Heffner tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq; 21910e7b1368SJohn Heffner tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq; 21925d424d5aSJohn Heffner 21935d424d5aSJohn Heffner return 1; 21945d424d5aSJohn Heffner } 21955d424d5aSJohn Heffner 21965d424d5aSJohn Heffner return -1; 21975d424d5aSJohn Heffner } 21985d424d5aSJohn Heffner 2199864e5c09SEric Dumazet static bool tcp_pacing_check(struct sock *sk) 2200218af599SEric Dumazet { 2201864e5c09SEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 2202864e5c09SEric Dumazet 2203864e5c09SEric Dumazet if (!tcp_needs_internal_pacing(sk)) 2204864e5c09SEric Dumazet return false; 2205864e5c09SEric Dumazet 2206864e5c09SEric Dumazet if (tp->tcp_wstamp_ns <= tp->tcp_clock_cache) 2207864e5c09SEric Dumazet return false; 2208864e5c09SEric Dumazet 2209864e5c09SEric Dumazet if (!hrtimer_is_queued(&tp->pacing_timer)) { 2210864e5c09SEric Dumazet hrtimer_start(&tp->pacing_timer, 2211864e5c09SEric Dumazet ns_to_ktime(tp->tcp_wstamp_ns), 2212864e5c09SEric Dumazet HRTIMER_MODE_ABS_PINNED_SOFT); 2213864e5c09SEric Dumazet sock_hold(sk); 2214864e5c09SEric Dumazet } 2215864e5c09SEric Dumazet return true; 2216218af599SEric Dumazet } 2217218af599SEric Dumazet 2218f9616c35SEric Dumazet /* TCP Small Queues : 2219f9616c35SEric Dumazet * Control number of packets in qdisc/devices to two packets / or ~1 ms. 2220f9616c35SEric Dumazet * (These limits are doubled for retransmits) 2221f9616c35SEric Dumazet * This allows for : 2222f9616c35SEric Dumazet * - better RTT estimation and ACK scheduling 2223f9616c35SEric Dumazet * - faster recovery 2224f9616c35SEric Dumazet * - high rates 2225f9616c35SEric Dumazet * Alas, some drivers / subsystems require a fair amount 2226f9616c35SEric Dumazet * of queued bytes to ensure line rate. 2227f9616c35SEric Dumazet * One example is wifi aggregation (802.11 AMPDU) 2228f9616c35SEric Dumazet */ 2229f9616c35SEric Dumazet static bool tcp_small_queue_check(struct sock *sk, const struct sk_buff *skb, 2230f9616c35SEric Dumazet unsigned int factor) 2231f9616c35SEric Dumazet { 223276a9ebe8SEric Dumazet unsigned long limit; 2233f9616c35SEric Dumazet 223476a9ebe8SEric Dumazet limit = max_t(unsigned long, 223576a9ebe8SEric Dumazet 2 * skb->truesize, 223676a9ebe8SEric Dumazet sk->sk_pacing_rate >> sk->sk_pacing_shift); 2237c73e5807SEric Dumazet if (sk->sk_pacing_status == SK_PACING_NONE) 223876a9ebe8SEric Dumazet limit = min_t(unsigned long, limit, 22399184d8bbSEric Dumazet sock_net(sk)->ipv4.sysctl_tcp_limit_output_bytes); 2240f9616c35SEric Dumazet limit <<= factor; 2241f9616c35SEric Dumazet 224214afee4bSReshetova, Elena if (refcount_read(&sk->sk_wmem_alloc) > limit) { 224375c119afSEric Dumazet /* Always send skb if rtx queue is empty. 224475eefc6cSEric Dumazet * No need to wait for TX completion to call us back, 224575eefc6cSEric Dumazet * after softirq/tasklet schedule. 224675eefc6cSEric Dumazet * This helps when TX completions are delayed too much. 224775eefc6cSEric Dumazet */ 224875c119afSEric Dumazet if (tcp_rtx_queue_empty(sk)) 224975eefc6cSEric Dumazet return false; 225075eefc6cSEric Dumazet 22517aa5470cSEric Dumazet set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags); 2252f9616c35SEric Dumazet /* It is possible TX completion already happened 2253f9616c35SEric Dumazet * before we set TSQ_THROTTLED, so we must 2254f9616c35SEric Dumazet * test again the condition. 2255f9616c35SEric Dumazet */ 2256f9616c35SEric Dumazet smp_mb__after_atomic(); 225714afee4bSReshetova, Elena if (refcount_read(&sk->sk_wmem_alloc) > limit) 2258f9616c35SEric Dumazet return true; 2259f9616c35SEric Dumazet } 2260f9616c35SEric Dumazet return false; 2261f9616c35SEric Dumazet } 2262f9616c35SEric Dumazet 226305b055e8SFrancis Yan static void tcp_chrono_set(struct tcp_sock *tp, const enum tcp_chrono new) 226405b055e8SFrancis Yan { 2265628174ccSEric Dumazet const u32 now = tcp_jiffies32; 2266efe967cdSArnd Bergmann enum tcp_chrono old = tp->chrono_type; 226705b055e8SFrancis Yan 2268efe967cdSArnd Bergmann if (old > TCP_CHRONO_UNSPEC) 2269efe967cdSArnd Bergmann tp->chrono_stat[old - 1] += now - tp->chrono_start; 227005b055e8SFrancis Yan tp->chrono_start = now; 227105b055e8SFrancis Yan tp->chrono_type = new; 227205b055e8SFrancis Yan } 227305b055e8SFrancis Yan 227405b055e8SFrancis Yan void tcp_chrono_start(struct sock *sk, const enum tcp_chrono type) 227505b055e8SFrancis Yan { 227605b055e8SFrancis Yan struct tcp_sock *tp = tcp_sk(sk); 227705b055e8SFrancis Yan 227805b055e8SFrancis Yan /* If there are multiple conditions worthy of tracking in a 22790f87230dSFrancis Yan * chronograph then the highest priority enum takes precedence 22800f87230dSFrancis Yan * over the other conditions. So that if something "more interesting" 228105b055e8SFrancis Yan * starts happening, stop the previous chrono and start a new one. 228205b055e8SFrancis Yan */ 228305b055e8SFrancis Yan if (type > tp->chrono_type) 228405b055e8SFrancis Yan tcp_chrono_set(tp, type); 228505b055e8SFrancis Yan } 228605b055e8SFrancis Yan 228705b055e8SFrancis Yan void tcp_chrono_stop(struct sock *sk, const enum tcp_chrono type) 228805b055e8SFrancis Yan { 228905b055e8SFrancis Yan struct tcp_sock *tp = tcp_sk(sk); 229005b055e8SFrancis Yan 22910f87230dSFrancis Yan 22920f87230dSFrancis Yan /* There are multiple conditions worthy of tracking in a 22930f87230dSFrancis Yan * chronograph, so that the highest priority enum takes 22940f87230dSFrancis Yan * precedence over the other conditions (see tcp_chrono_start). 22950f87230dSFrancis Yan * If a condition stops, we only stop chrono tracking if 22960f87230dSFrancis Yan * it's the "most interesting" or current chrono we are 22970f87230dSFrancis Yan * tracking and starts busy chrono if we have pending data. 22980f87230dSFrancis Yan */ 229975c119afSEric Dumazet if (tcp_rtx_and_write_queues_empty(sk)) 230005b055e8SFrancis Yan tcp_chrono_set(tp, TCP_CHRONO_UNSPEC); 23010f87230dSFrancis Yan else if (type == tp->chrono_type) 23020f87230dSFrancis Yan tcp_chrono_set(tp, TCP_CHRONO_BUSY); 230305b055e8SFrancis Yan } 230405b055e8SFrancis Yan 23051da177e4SLinus Torvalds /* This routine writes packets to the network. It advances the 23061da177e4SLinus Torvalds * send_head. This happens as incoming acks open up the remote 23071da177e4SLinus Torvalds * window for us. 23081da177e4SLinus Torvalds * 2309f8269a49SIlpo Järvinen * LARGESEND note: !tcp_urg_mode is overkill, only frames between 2310f8269a49SIlpo Järvinen * snd_up-64k-mss .. snd_up cannot be large. However, taking into 2311f8269a49SIlpo Järvinen * account rare use of URG, this is not a big flaw. 2312f8269a49SIlpo Järvinen * 23136ba8a3b1SNandita Dukkipati * Send at most one packet when push_one > 0. Temporarily ignore 23146ba8a3b1SNandita Dukkipati * cwnd limit to force at most one packet out when push_one == 2. 23156ba8a3b1SNandita Dukkipati 2316a2a385d6SEric Dumazet * Returns true, if no segments are in flight and we have queued segments, 2317a2a385d6SEric Dumazet * but cannot send anything now because of SWS or another problem. 23181da177e4SLinus Torvalds */ 2319a2a385d6SEric Dumazet static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle, 2320d5dd9175SIlpo Järvinen int push_one, gfp_t gfp) 23211da177e4SLinus Torvalds { 23221da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 232392df7b51SDavid S. Miller struct sk_buff *skb; 2324c1b4a7e6SDavid S. Miller unsigned int tso_segs, sent_pkts; 2325c1b4a7e6SDavid S. Miller int cwnd_quota; 23265d424d5aSJohn Heffner int result; 23275615f886SFrancis Yan bool is_cwnd_limited = false, is_rwnd_limited = false; 2328605ad7f1SEric Dumazet u32 max_segs; 23291da177e4SLinus Torvalds 2330c1b4a7e6SDavid S. Miller sent_pkts = 0; 23315d424d5aSJohn Heffner 2332ee1836aeSEric Dumazet tcp_mstamp_refresh(tp); 2333d5dd9175SIlpo Järvinen if (!push_one) { 23345d424d5aSJohn Heffner /* Do MTU probing. */ 2335d5dd9175SIlpo Järvinen result = tcp_mtu_probe(sk); 2336d5dd9175SIlpo Järvinen if (!result) { 2337a2a385d6SEric Dumazet return false; 23385d424d5aSJohn Heffner } else if (result > 0) { 23395d424d5aSJohn Heffner sent_pkts = 1; 23405d424d5aSJohn Heffner } 2341d5dd9175SIlpo Järvinen } 23425d424d5aSJohn Heffner 2343ed6e7268SNeal Cardwell max_segs = tcp_tso_segs(sk, mss_now); 2344fe067e8aSDavid S. Miller while ((skb = tcp_send_head(sk))) { 2345c8ac3774SHerbert Xu unsigned int limit; 2346c8ac3774SHerbert Xu 234779861919SEric Dumazet if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE) { 234879861919SEric Dumazet /* "skb_mstamp_ns" is used as a start point for the retransmit timer */ 234979861919SEric Dumazet skb->skb_mstamp_ns = tp->tcp_wstamp_ns = tp->tcp_clock_cache; 235079861919SEric Dumazet list_move_tail(&skb->tcp_tsorted_anchor, &tp->tsorted_sent_queue); 2351bf50b606SEric Dumazet tcp_init_tso_segs(skb, mss_now); 235279861919SEric Dumazet goto repair; /* Skip network transmission */ 235379861919SEric Dumazet } 235479861919SEric Dumazet 2355218af599SEric Dumazet if (tcp_pacing_check(sk)) 2356218af599SEric Dumazet break; 2357218af599SEric Dumazet 23585bbb432cSEric Dumazet tso_segs = tcp_init_tso_segs(skb, mss_now); 2359c1b4a7e6SDavid S. Miller BUG_ON(!tso_segs); 2360c1b4a7e6SDavid S. Miller 2361b68e9f85SHerbert Xu cwnd_quota = tcp_cwnd_test(tp, skb); 23626ba8a3b1SNandita Dukkipati if (!cwnd_quota) { 23636ba8a3b1SNandita Dukkipati if (push_one == 2) 23646ba8a3b1SNandita Dukkipati /* Force out a loss probe pkt. */ 23656ba8a3b1SNandita Dukkipati cwnd_quota = 1; 23666ba8a3b1SNandita Dukkipati else 2367b68e9f85SHerbert Xu break; 23686ba8a3b1SNandita Dukkipati } 2369b68e9f85SHerbert Xu 23705615f886SFrancis Yan if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now))) { 23715615f886SFrancis Yan is_rwnd_limited = true; 2372b68e9f85SHerbert Xu break; 23735615f886SFrancis Yan } 2374b68e9f85SHerbert Xu 2375d6a4e26aSEric Dumazet if (tso_segs == 1) { 2376aa93466bSDavid S. Miller if (unlikely(!tcp_nagle_test(tp, skb, mss_now, 2377aa93466bSDavid S. Miller (tcp_skb_is_last(sk, skb) ? 2378aa93466bSDavid S. Miller nonagle : TCP_NAGLE_PUSH)))) 2379aa93466bSDavid S. Miller break; 2380c1b4a7e6SDavid S. Miller } else { 2381ca8a2263SNeal Cardwell if (!push_one && 2382605ad7f1SEric Dumazet tcp_tso_should_defer(sk, skb, &is_cwnd_limited, 2383f9bfe4e6SEric Dumazet &is_rwnd_limited, max_segs)) 2384aa93466bSDavid S. Miller break; 2385c1b4a7e6SDavid S. Miller } 2386aa93466bSDavid S. Miller 2387605ad7f1SEric Dumazet limit = mss_now; 2388d6a4e26aSEric Dumazet if (tso_segs > 1 && !tcp_urg_mode(tp)) 2389605ad7f1SEric Dumazet limit = tcp_mss_split_point(sk, skb, mss_now, 2390605ad7f1SEric Dumazet min_t(unsigned int, 2391605ad7f1SEric Dumazet cwnd_quota, 2392605ad7f1SEric Dumazet max_segs), 2393605ad7f1SEric Dumazet nonagle); 2394605ad7f1SEric Dumazet 2395605ad7f1SEric Dumazet if (skb->len > limit && 239656483341SEric Dumazet unlikely(tso_fragment(sk, skb, limit, mss_now, gfp))) 2397605ad7f1SEric Dumazet break; 2398605ad7f1SEric Dumazet 2399f9616c35SEric Dumazet if (tcp_small_queue_check(sk, skb, 0)) 240046d3ceabSEric Dumazet break; 2401c9eeec26SEric Dumazet 2402d5dd9175SIlpo Järvinen if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp))) 24031da177e4SLinus Torvalds break; 24041da177e4SLinus Torvalds 2405ec342325SAndrew Vagin repair: 24061da177e4SLinus Torvalds /* Advance the send_head. This one is sent out. 24071da177e4SLinus Torvalds * This call will increment packets_out. 24081da177e4SLinus Torvalds */ 240966f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, skb); 24101da177e4SLinus Torvalds 24111da177e4SLinus Torvalds tcp_minshall_update(tp, mss_now, skb); 2412a262f0cdSNandita Dukkipati sent_pkts += tcp_skb_pcount(skb); 2413d5dd9175SIlpo Järvinen 2414d5dd9175SIlpo Järvinen if (push_one) 2415d5dd9175SIlpo Järvinen break; 24161da177e4SLinus Torvalds } 24171da177e4SLinus Torvalds 24185615f886SFrancis Yan if (is_rwnd_limited) 24195615f886SFrancis Yan tcp_chrono_start(sk, TCP_CHRONO_RWND_LIMITED); 24205615f886SFrancis Yan else 24215615f886SFrancis Yan tcp_chrono_stop(sk, TCP_CHRONO_RWND_LIMITED); 24225615f886SFrancis Yan 2423aa93466bSDavid S. Miller if (likely(sent_pkts)) { 2424684bad11SYuchung Cheng if (tcp_in_cwnd_reduction(sk)) 2425684bad11SYuchung Cheng tp->prr_out += sent_pkts; 24266ba8a3b1SNandita Dukkipati 24276ba8a3b1SNandita Dukkipati /* Send one loss probe per tail loss episode. */ 24286ba8a3b1SNandita Dukkipati if (push_one != 2) 2429ed66dfafSNeal Cardwell tcp_schedule_loss_probe(sk, false); 2430d2e1339fSBendik Rønning Opstad is_cwnd_limited |= (tcp_packets_in_flight(tp) >= tp->snd_cwnd); 2431ca8a2263SNeal Cardwell tcp_cwnd_validate(sk, is_cwnd_limited); 2432a2a385d6SEric Dumazet return false; 24331da177e4SLinus Torvalds } 243475c119afSEric Dumazet return !tp->packets_out && !tcp_write_queue_empty(sk); 24356ba8a3b1SNandita Dukkipati } 24366ba8a3b1SNandita Dukkipati 2437ed66dfafSNeal Cardwell bool tcp_schedule_loss_probe(struct sock *sk, bool advancing_rto) 24386ba8a3b1SNandita Dukkipati { 24396ba8a3b1SNandita Dukkipati struct inet_connection_sock *icsk = inet_csk(sk); 24406ba8a3b1SNandita Dukkipati struct tcp_sock *tp = tcp_sk(sk); 2441a2815817SNeal Cardwell u32 timeout, rto_delta_us; 24422ae21cf5SEric Dumazet int early_retrans; 24436ba8a3b1SNandita Dukkipati 24446ba8a3b1SNandita Dukkipati /* Don't do any loss probe on a Fast Open connection before 3WHS 24456ba8a3b1SNandita Dukkipati * finishes. 24466ba8a3b1SNandita Dukkipati */ 2447f9b99582SYuchung Cheng if (tp->fastopen_rsk) 24486ba8a3b1SNandita Dukkipati return false; 24496ba8a3b1SNandita Dukkipati 24502ae21cf5SEric Dumazet early_retrans = sock_net(sk)->ipv4.sysctl_tcp_early_retrans; 24516ba8a3b1SNandita Dukkipati /* Schedule a loss probe in 2*RTT for SACK capable connections 2452b4f70c3dSNeal Cardwell * not in loss recovery, that are either limited by cwnd or application. 24536ba8a3b1SNandita Dukkipati */ 24542ae21cf5SEric Dumazet if ((early_retrans != 3 && early_retrans != 4) || 2455bec41a11SYuchung Cheng !tp->packets_out || !tcp_is_sack(tp) || 2456b4f70c3dSNeal Cardwell (icsk->icsk_ca_state != TCP_CA_Open && 2457b4f70c3dSNeal Cardwell icsk->icsk_ca_state != TCP_CA_CWR)) 24586ba8a3b1SNandita Dukkipati return false; 24596ba8a3b1SNandita Dukkipati 2460bb4d991aSYuchung Cheng /* Probe timeout is 2*rtt. Add minimum RTO to account 2461f9b99582SYuchung Cheng * for delayed ack when there's one outstanding packet. If no RTT 2462f9b99582SYuchung Cheng * sample is available then probe after TCP_TIMEOUT_INIT. 24636ba8a3b1SNandita Dukkipati */ 2464bb4d991aSYuchung Cheng if (tp->srtt_us) { 2465bb4d991aSYuchung Cheng timeout = usecs_to_jiffies(tp->srtt_us >> 2); 24666ba8a3b1SNandita Dukkipati if (tp->packets_out == 1) 2467bb4d991aSYuchung Cheng timeout += TCP_RTO_MIN; 2468bb4d991aSYuchung Cheng else 2469bb4d991aSYuchung Cheng timeout += TCP_TIMEOUT_MIN; 2470bb4d991aSYuchung Cheng } else { 2471bb4d991aSYuchung Cheng timeout = TCP_TIMEOUT_INIT; 2472bb4d991aSYuchung Cheng } 24736ba8a3b1SNandita Dukkipati 2474a2815817SNeal Cardwell /* If the RTO formula yields an earlier time, then use that time. */ 2475ed66dfafSNeal Cardwell rto_delta_us = advancing_rto ? 2476ed66dfafSNeal Cardwell jiffies_to_usecs(inet_csk(sk)->icsk_rto) : 2477ed66dfafSNeal Cardwell tcp_rto_delta_us(sk); /* How far in future is RTO? */ 2478a2815817SNeal Cardwell if (rto_delta_us > 0) 2479a2815817SNeal Cardwell timeout = min_t(u32, timeout, usecs_to_jiffies(rto_delta_us)); 24806ba8a3b1SNandita Dukkipati 24813f80e08fSEric Dumazet tcp_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout, 24823f80e08fSEric Dumazet TCP_RTO_MAX, NULL); 24836ba8a3b1SNandita Dukkipati return true; 24846ba8a3b1SNandita Dukkipati } 24856ba8a3b1SNandita Dukkipati 24861f3279aeSEric Dumazet /* Thanks to skb fast clones, we can detect if a prior transmit of 24871f3279aeSEric Dumazet * a packet is still in a qdisc or driver queue. 24881f3279aeSEric Dumazet * In this case, there is very little point doing a retransmit ! 24891f3279aeSEric Dumazet */ 24901f3279aeSEric Dumazet static bool skb_still_in_host_queue(const struct sock *sk, 24911f3279aeSEric Dumazet const struct sk_buff *skb) 24921f3279aeSEric Dumazet { 249339bb5e62SEric Dumazet if (unlikely(skb_fclone_busy(sk, skb))) { 2494c10d9310SEric Dumazet NET_INC_STATS(sock_net(sk), 24951f3279aeSEric Dumazet LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES); 24961f3279aeSEric Dumazet return true; 24971f3279aeSEric Dumazet } 24981f3279aeSEric Dumazet return false; 24991f3279aeSEric Dumazet } 25001f3279aeSEric Dumazet 2501b340b264SYuchung Cheng /* When probe timeout (PTO) fires, try send a new segment if possible, else 25026ba8a3b1SNandita Dukkipati * retransmit the last segment. 25036ba8a3b1SNandita Dukkipati */ 25046ba8a3b1SNandita Dukkipati void tcp_send_loss_probe(struct sock *sk) 25056ba8a3b1SNandita Dukkipati { 25069b717a8dSNandita Dukkipati struct tcp_sock *tp = tcp_sk(sk); 25076ba8a3b1SNandita Dukkipati struct sk_buff *skb; 25086ba8a3b1SNandita Dukkipati int pcount; 25096ba8a3b1SNandita Dukkipati int mss = tcp_current_mss(sk); 25106ba8a3b1SNandita Dukkipati 2511b340b264SYuchung Cheng skb = tcp_send_head(sk); 251275c119afSEric Dumazet if (skb && tcp_snd_wnd_test(tp, skb, mss)) { 2513b340b264SYuchung Cheng pcount = tp->packets_out; 2514b340b264SYuchung Cheng tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC); 2515b340b264SYuchung Cheng if (tp->packets_out > pcount) 2516b340b264SYuchung Cheng goto probe_sent; 25176ba8a3b1SNandita Dukkipati goto rearm_timer; 25186ba8a3b1SNandita Dukkipati } 251975c119afSEric Dumazet skb = skb_rb_last(&sk->tcp_rtx_queue); 2520b2b7af86SYuchung Cheng if (unlikely(!skb)) { 2521b2b7af86SYuchung Cheng WARN_ONCE(tp->packets_out, 2522b2b7af86SYuchung Cheng "invalid inflight: %u state %u cwnd %u mss %d\n", 2523b2b7af86SYuchung Cheng tp->packets_out, sk->sk_state, tp->snd_cwnd, mss); 2524b2b7af86SYuchung Cheng inet_csk(sk)->icsk_pending = 0; 2525b2b7af86SYuchung Cheng return; 2526b2b7af86SYuchung Cheng } 25276ba8a3b1SNandita Dukkipati 25289b717a8dSNandita Dukkipati /* At most one outstanding TLP retransmission. */ 25299b717a8dSNandita Dukkipati if (tp->tlp_high_seq) 25309b717a8dSNandita Dukkipati goto rearm_timer; 25319b717a8dSNandita Dukkipati 25321f3279aeSEric Dumazet if (skb_still_in_host_queue(sk, skb)) 25331f3279aeSEric Dumazet goto rearm_timer; 25341f3279aeSEric Dumazet 25356ba8a3b1SNandita Dukkipati pcount = tcp_skb_pcount(skb); 25366ba8a3b1SNandita Dukkipati if (WARN_ON(!pcount)) 25376ba8a3b1SNandita Dukkipati goto rearm_timer; 25386ba8a3b1SNandita Dukkipati 25396ba8a3b1SNandita Dukkipati if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) { 254075c119afSEric Dumazet if (unlikely(tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb, 254175c119afSEric Dumazet (pcount - 1) * mss, mss, 25426cc55e09SOctavian Purdila GFP_ATOMIC))) 25436ba8a3b1SNandita Dukkipati goto rearm_timer; 254475c119afSEric Dumazet skb = skb_rb_next(skb); 25456ba8a3b1SNandita Dukkipati } 25466ba8a3b1SNandita Dukkipati 25476ba8a3b1SNandita Dukkipati if (WARN_ON(!skb || !tcp_skb_pcount(skb))) 25486ba8a3b1SNandita Dukkipati goto rearm_timer; 25496ba8a3b1SNandita Dukkipati 255010d3be56SEric Dumazet if (__tcp_retransmit_skb(sk, skb, 1)) 2551b340b264SYuchung Cheng goto rearm_timer; 25526ba8a3b1SNandita Dukkipati 25539b717a8dSNandita Dukkipati /* Record snd_nxt for loss detection. */ 25549b717a8dSNandita Dukkipati tp->tlp_high_seq = tp->snd_nxt; 25559b717a8dSNandita Dukkipati 2556b340b264SYuchung Cheng probe_sent: 2557c10d9310SEric Dumazet NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSPROBES); 2558fcd16c0aSYuchung Cheng /* Reset s.t. tcp_rearm_rto will restart timer from now */ 2559fcd16c0aSYuchung Cheng inet_csk(sk)->icsk_pending = 0; 2560b340b264SYuchung Cheng rearm_timer: 2561fcd16c0aSYuchung Cheng tcp_rearm_rto(sk); 25621da177e4SLinus Torvalds } 25631da177e4SLinus Torvalds 2564a762a980SDavid S. Miller /* Push out any pending frames which were held back due to 2565a762a980SDavid S. Miller * TCP_CORK or attempt at coalescing tiny packets. 2566a762a980SDavid S. Miller * The socket must be locked by the caller. 2567a762a980SDavid S. Miller */ 25689e412ba7SIlpo Järvinen void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss, 25699e412ba7SIlpo Järvinen int nonagle) 2570a762a980SDavid S. Miller { 2571726e07a8SIlpo Järvinen /* If we are closed, the bytes will have to remain here. 2572726e07a8SIlpo Järvinen * In time closedown will finish, we empty the write queue and 2573726e07a8SIlpo Järvinen * all will be happy. 2574726e07a8SIlpo Järvinen */ 2575726e07a8SIlpo Järvinen if (unlikely(sk->sk_state == TCP_CLOSE)) 2576726e07a8SIlpo Järvinen return; 2577726e07a8SIlpo Järvinen 257899a1dec7SMel Gorman if (tcp_write_xmit(sk, cur_mss, nonagle, 0, 25797450aaf6SEric Dumazet sk_gfp_mask(sk, GFP_ATOMIC))) 25809e412ba7SIlpo Järvinen tcp_check_probe_timer(sk); 2581a762a980SDavid S. Miller } 2582a762a980SDavid S. Miller 2583c1b4a7e6SDavid S. Miller /* Send _single_ skb sitting at the send head. This function requires 2584c1b4a7e6SDavid S. Miller * true push pending frames to setup probe timer etc. 2585c1b4a7e6SDavid S. Miller */ 2586c1b4a7e6SDavid S. Miller void tcp_push_one(struct sock *sk, unsigned int mss_now) 2587c1b4a7e6SDavid S. Miller { 2588fe067e8aSDavid S. Miller struct sk_buff *skb = tcp_send_head(sk); 2589c1b4a7e6SDavid S. Miller 2590c1b4a7e6SDavid S. Miller BUG_ON(!skb || skb->len < mss_now); 2591c1b4a7e6SDavid S. Miller 2592d5dd9175SIlpo Järvinen tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation); 2593c1b4a7e6SDavid S. Miller } 2594c1b4a7e6SDavid S. Miller 25951da177e4SLinus Torvalds /* This function returns the amount that we can raise the 25961da177e4SLinus Torvalds * usable window based on the following constraints 25971da177e4SLinus Torvalds * 25981da177e4SLinus Torvalds * 1. The window can never be shrunk once it is offered (RFC 793) 25991da177e4SLinus Torvalds * 2. We limit memory per socket 26001da177e4SLinus Torvalds * 26011da177e4SLinus Torvalds * RFC 1122: 26021da177e4SLinus Torvalds * "the suggested [SWS] avoidance algorithm for the receiver is to keep 26031da177e4SLinus Torvalds * RECV.NEXT + RCV.WIN fixed until: 26041da177e4SLinus Torvalds * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)" 26051da177e4SLinus Torvalds * 26061da177e4SLinus Torvalds * i.e. don't raise the right edge of the window until you can raise 26071da177e4SLinus Torvalds * it at least MSS bytes. 26081da177e4SLinus Torvalds * 26091da177e4SLinus Torvalds * Unfortunately, the recommended algorithm breaks header prediction, 26101da177e4SLinus Torvalds * since header prediction assumes th->window stays fixed. 26111da177e4SLinus Torvalds * 26121da177e4SLinus Torvalds * Strictly speaking, keeping th->window fixed violates the receiver 26131da177e4SLinus Torvalds * side SWS prevention criteria. The problem is that under this rule 26141da177e4SLinus Torvalds * a stream of single byte packets will cause the right side of the 26151da177e4SLinus Torvalds * window to always advance by a single byte. 26161da177e4SLinus Torvalds * 26171da177e4SLinus Torvalds * Of course, if the sender implements sender side SWS prevention 26181da177e4SLinus Torvalds * then this will not be a problem. 26191da177e4SLinus Torvalds * 26201da177e4SLinus Torvalds * BSD seems to make the following compromise: 26211da177e4SLinus Torvalds * 26221da177e4SLinus Torvalds * If the free space is less than the 1/4 of the maximum 26231da177e4SLinus Torvalds * space available and the free space is less than 1/2 mss, 26241da177e4SLinus Torvalds * then set the window to 0. 26251da177e4SLinus Torvalds * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ] 26261da177e4SLinus Torvalds * Otherwise, just prevent the window from shrinking 26271da177e4SLinus Torvalds * and from being larger than the largest representable value. 26281da177e4SLinus Torvalds * 26291da177e4SLinus Torvalds * This prevents incremental opening of the window in the regime 26301da177e4SLinus Torvalds * where TCP is limited by the speed of the reader side taking 26311da177e4SLinus Torvalds * data out of the TCP receive queue. It does nothing about 26321da177e4SLinus Torvalds * those cases where the window is constrained on the sender side 26331da177e4SLinus Torvalds * because the pipeline is full. 26341da177e4SLinus Torvalds * 26351da177e4SLinus Torvalds * BSD also seems to "accidentally" limit itself to windows that are a 26361da177e4SLinus Torvalds * multiple of MSS, at least until the free space gets quite small. 26371da177e4SLinus Torvalds * This would appear to be a side effect of the mbuf implementation. 26381da177e4SLinus Torvalds * Combining these two algorithms results in the observed behavior 26391da177e4SLinus Torvalds * of having a fixed window size at almost all times. 26401da177e4SLinus Torvalds * 26411da177e4SLinus Torvalds * Below we obtain similar behavior by forcing the offered window to 26421da177e4SLinus Torvalds * a multiple of the mss when it is feasible to do so. 26431da177e4SLinus Torvalds * 26441da177e4SLinus Torvalds * Note, we don't "adjust" for TIMESTAMP or SACK option bytes. 26451da177e4SLinus Torvalds * Regular options like TIMESTAMP are taken into account. 26461da177e4SLinus Torvalds */ 26471da177e4SLinus Torvalds u32 __tcp_select_window(struct sock *sk) 26481da177e4SLinus Torvalds { 2649463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 26501da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 2651caa20d9aSStephen Hemminger /* MSS for the peer's data. Previous versions used mss_clamp 26521da177e4SLinus Torvalds * here. I don't know if the value based on our guesses 26531da177e4SLinus Torvalds * of peer's MSS is better for the performance. It's more correct 26541da177e4SLinus Torvalds * but may be worse for the performance because of rcv_mss 26551da177e4SLinus Torvalds * fluctuations. --SAW 1998/11/1 26561da177e4SLinus Torvalds */ 2657463c84b9SArnaldo Carvalho de Melo int mss = icsk->icsk_ack.rcv_mss; 26581da177e4SLinus Torvalds int free_space = tcp_space(sk); 265986c1a045SFlorian Westphal int allowed_space = tcp_full_space(sk); 266086c1a045SFlorian Westphal int full_space = min_t(int, tp->window_clamp, allowed_space); 26611da177e4SLinus Torvalds int window; 26621da177e4SLinus Torvalds 266306425c30SEric Dumazet if (unlikely(mss > full_space)) { 26641da177e4SLinus Torvalds mss = full_space; 266506425c30SEric Dumazet if (mss <= 0) 266606425c30SEric Dumazet return 0; 266706425c30SEric Dumazet } 2668b92edbe0SEric Dumazet if (free_space < (full_space >> 1)) { 2669463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.quick = 0; 26701da177e4SLinus Torvalds 2671b8da51ebSEric Dumazet if (tcp_under_memory_pressure(sk)) 2672056834d9SIlpo Järvinen tp->rcv_ssthresh = min(tp->rcv_ssthresh, 2673056834d9SIlpo Järvinen 4U * tp->advmss); 26741da177e4SLinus Torvalds 267586c1a045SFlorian Westphal /* free_space might become our new window, make sure we don't 267686c1a045SFlorian Westphal * increase it due to wscale. 267786c1a045SFlorian Westphal */ 267886c1a045SFlorian Westphal free_space = round_down(free_space, 1 << tp->rx_opt.rcv_wscale); 267986c1a045SFlorian Westphal 268086c1a045SFlorian Westphal /* if free space is less than mss estimate, or is below 1/16th 268186c1a045SFlorian Westphal * of the maximum allowed, try to move to zero-window, else 268286c1a045SFlorian Westphal * tcp_clamp_window() will grow rcv buf up to tcp_rmem[2], and 268386c1a045SFlorian Westphal * new incoming data is dropped due to memory limits. 268486c1a045SFlorian Westphal * With large window, mss test triggers way too late in order 268586c1a045SFlorian Westphal * to announce zero window in time before rmem limit kicks in. 268686c1a045SFlorian Westphal */ 268786c1a045SFlorian Westphal if (free_space < (allowed_space >> 4) || free_space < mss) 26881da177e4SLinus Torvalds return 0; 26891da177e4SLinus Torvalds } 26901da177e4SLinus Torvalds 26911da177e4SLinus Torvalds if (free_space > tp->rcv_ssthresh) 26921da177e4SLinus Torvalds free_space = tp->rcv_ssthresh; 26931da177e4SLinus Torvalds 26941da177e4SLinus Torvalds /* Don't do rounding if we are using window scaling, since the 26951da177e4SLinus Torvalds * scaled window will not line up with the MSS boundary anyway. 26961da177e4SLinus Torvalds */ 26971da177e4SLinus Torvalds if (tp->rx_opt.rcv_wscale) { 26981da177e4SLinus Torvalds window = free_space; 26991da177e4SLinus Torvalds 27001da177e4SLinus Torvalds /* Advertise enough space so that it won't get scaled away. 27011da177e4SLinus Torvalds * Import case: prevent zero window announcement if 27021da177e4SLinus Torvalds * 1<<rcv_wscale > mss. 27031da177e4SLinus Torvalds */ 27041935299dSGao Feng window = ALIGN(window, (1 << tp->rx_opt.rcv_wscale)); 27051da177e4SLinus Torvalds } else { 27061935299dSGao Feng window = tp->rcv_wnd; 27071da177e4SLinus Torvalds /* Get the largest window that is a nice multiple of mss. 27081da177e4SLinus Torvalds * Window clamp already applied above. 27091da177e4SLinus Torvalds * If our current window offering is within 1 mss of the 27101da177e4SLinus Torvalds * free space we just keep it. This prevents the divide 27111da177e4SLinus Torvalds * and multiply from happening most of the time. 27121da177e4SLinus Torvalds * We also don't do any window rounding when the free space 27131da177e4SLinus Torvalds * is too small. 27141da177e4SLinus Torvalds */ 27151da177e4SLinus Torvalds if (window <= free_space - mss || window > free_space) 27161935299dSGao Feng window = rounddown(free_space, mss); 271784565070SJohn Heffner else if (mss == full_space && 2718b92edbe0SEric Dumazet free_space > window + (full_space >> 1)) 271984565070SJohn Heffner window = free_space; 27201da177e4SLinus Torvalds } 27211da177e4SLinus Torvalds 27221da177e4SLinus Torvalds return window; 27231da177e4SLinus Torvalds } 27241da177e4SLinus Torvalds 2725cfea5a68SMartin KaFai Lau void tcp_skb_collapse_tstamp(struct sk_buff *skb, 2726082ac2d5SMartin KaFai Lau const struct sk_buff *next_skb) 2727082ac2d5SMartin KaFai Lau { 27280a2cf20cSSoheil Hassas Yeganeh if (unlikely(tcp_has_tx_tstamp(next_skb))) { 27290a2cf20cSSoheil Hassas Yeganeh const struct skb_shared_info *next_shinfo = 27300a2cf20cSSoheil Hassas Yeganeh skb_shinfo(next_skb); 2731082ac2d5SMartin KaFai Lau struct skb_shared_info *shinfo = skb_shinfo(skb); 2732082ac2d5SMartin KaFai Lau 27330a2cf20cSSoheil Hassas Yeganeh shinfo->tx_flags |= next_shinfo->tx_flags & SKBTX_ANY_TSTAMP; 2734082ac2d5SMartin KaFai Lau shinfo->tskey = next_shinfo->tskey; 27352de8023eSMartin KaFai Lau TCP_SKB_CB(skb)->txstamp_ack |= 27362de8023eSMartin KaFai Lau TCP_SKB_CB(next_skb)->txstamp_ack; 2737082ac2d5SMartin KaFai Lau } 2738082ac2d5SMartin KaFai Lau } 2739082ac2d5SMartin KaFai Lau 27404a17fc3aSIlpo Järvinen /* Collapses two adjacent SKB's during retransmission. */ 2741f8071cdeSEric Dumazet static bool tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb) 27421da177e4SLinus Torvalds { 27431da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 274475c119afSEric Dumazet struct sk_buff *next_skb = skb_rb_next(skb); 274513dde04fSWei Yongjun int next_skb_size; 27461da177e4SLinus Torvalds 2747058dc334SIlpo Järvinen next_skb_size = next_skb->len; 27481da177e4SLinus Torvalds 2749058dc334SIlpo Järvinen BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1); 27501da177e4SLinus Torvalds 2751f8071cdeSEric Dumazet if (next_skb_size) { 2752f8071cdeSEric Dumazet if (next_skb_size <= skb_availroom(skb)) 2753f8071cdeSEric Dumazet skb_copy_bits(next_skb, 0, skb_put(skb, next_skb_size), 2754f8071cdeSEric Dumazet next_skb_size); 27553b4929f6SEric Dumazet else if (!tcp_skb_shift(skb, next_skb, 1, next_skb_size)) 2756f8071cdeSEric Dumazet return false; 2757f8071cdeSEric Dumazet } 27582b7cda9cSEric Dumazet tcp_highest_sack_replace(sk, next_skb, skb); 2759a6963a6bSIlpo Järvinen 27601da177e4SLinus Torvalds /* Update sequence range on original skb. */ 27611da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq; 27621da177e4SLinus Torvalds 2763e6c7d085SIlpo Järvinen /* Merge over control information. This moves PSH/FIN etc. over */ 27644de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags; 27651da177e4SLinus Torvalds 27661da177e4SLinus Torvalds /* All done, get rid of second SKB and account for it so 27671da177e4SLinus Torvalds * packet counting does not break. 27681da177e4SLinus Torvalds */ 27694828e7f4SIlpo Järvinen TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS; 2770a643b5d4SMartin KaFai Lau TCP_SKB_CB(skb)->eor = TCP_SKB_CB(next_skb)->eor; 2771b7689205SIlpo Järvinen 2772b7689205SIlpo Järvinen /* changed transmit queue under us so clear hints */ 2773ef9da47cSIlpo Järvinen tcp_clear_retrans_hints_partial(tp); 2774ef9da47cSIlpo Järvinen if (next_skb == tp->retransmit_skb_hint) 2775ef9da47cSIlpo Järvinen tp->retransmit_skb_hint = skb; 2776b7689205SIlpo Järvinen 2777797108d1SIlpo Järvinen tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb)); 2778797108d1SIlpo Järvinen 2779082ac2d5SMartin KaFai Lau tcp_skb_collapse_tstamp(skb, next_skb); 2780082ac2d5SMartin KaFai Lau 278175c119afSEric Dumazet tcp_rtx_queue_unlink_and_free(next_skb, sk); 2782f8071cdeSEric Dumazet return true; 27831da177e4SLinus Torvalds } 27841da177e4SLinus Torvalds 278567edfef7SAndi Kleen /* Check if coalescing SKBs is legal. */ 2786a2a385d6SEric Dumazet static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb) 27874a17fc3aSIlpo Järvinen { 27884a17fc3aSIlpo Järvinen if (tcp_skb_pcount(skb) > 1) 2789a2a385d6SEric Dumazet return false; 27904a17fc3aSIlpo Järvinen if (skb_cloned(skb)) 2791a2a385d6SEric Dumazet return false; 27922331ccc5SEric Dumazet /* Some heuristics for collapsing over SACK'd could be invented */ 27934a17fc3aSIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 2794a2a385d6SEric Dumazet return false; 27954a17fc3aSIlpo Järvinen 2796a2a385d6SEric Dumazet return true; 27974a17fc3aSIlpo Järvinen } 27984a17fc3aSIlpo Järvinen 279967edfef7SAndi Kleen /* Collapse packets in the retransmit queue to make to create 280067edfef7SAndi Kleen * less packets on the wire. This is only done on retransmission. 280167edfef7SAndi Kleen */ 28024a17fc3aSIlpo Järvinen static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to, 28034a17fc3aSIlpo Järvinen int space) 28044a17fc3aSIlpo Järvinen { 28054a17fc3aSIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 28064a17fc3aSIlpo Järvinen struct sk_buff *skb = to, *tmp; 2807a2a385d6SEric Dumazet bool first = true; 28084a17fc3aSIlpo Järvinen 2809e0a1e5b5SEric Dumazet if (!sock_net(sk)->ipv4.sysctl_tcp_retrans_collapse) 28104a17fc3aSIlpo Järvinen return; 28114de075e0SEric Dumazet if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN) 28124a17fc3aSIlpo Järvinen return; 28134a17fc3aSIlpo Järvinen 281475c119afSEric Dumazet skb_rbtree_walk_from_safe(skb, tmp) { 28154a17fc3aSIlpo Järvinen if (!tcp_can_collapse(sk, skb)) 28164a17fc3aSIlpo Järvinen break; 28174a17fc3aSIlpo Järvinen 2818a643b5d4SMartin KaFai Lau if (!tcp_skb_can_collapse_to(to)) 2819a643b5d4SMartin KaFai Lau break; 2820a643b5d4SMartin KaFai Lau 28214a17fc3aSIlpo Järvinen space -= skb->len; 28224a17fc3aSIlpo Järvinen 28234a17fc3aSIlpo Järvinen if (first) { 2824a2a385d6SEric Dumazet first = false; 28254a17fc3aSIlpo Järvinen continue; 28264a17fc3aSIlpo Järvinen } 28274a17fc3aSIlpo Järvinen 28284a17fc3aSIlpo Järvinen if (space < 0) 28294a17fc3aSIlpo Järvinen break; 28304a17fc3aSIlpo Järvinen 28314a17fc3aSIlpo Järvinen if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp))) 28324a17fc3aSIlpo Järvinen break; 28334a17fc3aSIlpo Järvinen 2834f8071cdeSEric Dumazet if (!tcp_collapse_retrans(sk, to)) 2835f8071cdeSEric Dumazet break; 28364a17fc3aSIlpo Järvinen } 28374a17fc3aSIlpo Järvinen } 28384a17fc3aSIlpo Järvinen 28391da177e4SLinus Torvalds /* This retransmits one SKB. Policy decisions and retransmit queue 28401da177e4SLinus Torvalds * state updates are done by the caller. Returns non-zero if an 28411da177e4SLinus Torvalds * error occurred which prevented the send. 28421da177e4SLinus Torvalds */ 284310d3be56SEric Dumazet int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs) 28441da177e4SLinus Torvalds { 28455d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 284610d3be56SEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 28477d227cd2SSridhar Samudrala unsigned int cur_mss; 284810d3be56SEric Dumazet int diff, len, err; 28491da177e4SLinus Torvalds 285010d3be56SEric Dumazet 285110d3be56SEric Dumazet /* Inconclusive MTU probe */ 285210d3be56SEric Dumazet if (icsk->icsk_mtup.probe_size) 28535d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = 0; 28545d424d5aSJohn Heffner 28551da177e4SLinus Torvalds /* Do not sent more than we queued. 1/4 is reserved for possible 2856caa20d9aSStephen Hemminger * copying overhead: fragmentation, tunneling, mangling etc. 28571da177e4SLinus Torvalds */ 285814afee4bSReshetova, Elena if (refcount_read(&sk->sk_wmem_alloc) > 2859ffb4d6c8SEric Dumazet min_t(u32, sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), 2860ffb4d6c8SEric Dumazet sk->sk_sndbuf)) 28611da177e4SLinus Torvalds return -EAGAIN; 28621da177e4SLinus Torvalds 28631f3279aeSEric Dumazet if (skb_still_in_host_queue(sk, skb)) 28641f3279aeSEric Dumazet return -EBUSY; 28651f3279aeSEric Dumazet 28661da177e4SLinus Torvalds if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) { 28677f582b24SEric Dumazet if (unlikely(before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))) { 28687f582b24SEric Dumazet WARN_ON_ONCE(1); 28697f582b24SEric Dumazet return -EINVAL; 28707f582b24SEric Dumazet } 28711da177e4SLinus Torvalds if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq)) 28721da177e4SLinus Torvalds return -ENOMEM; 28731da177e4SLinus Torvalds } 28741da177e4SLinus Torvalds 28757d227cd2SSridhar Samudrala if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk)) 28767d227cd2SSridhar Samudrala return -EHOSTUNREACH; /* Routing failure or similar. */ 28777d227cd2SSridhar Samudrala 28780c54b85fSIlpo Järvinen cur_mss = tcp_current_mss(sk); 28797d227cd2SSridhar Samudrala 28801da177e4SLinus Torvalds /* If receiver has shrunk his window, and skb is out of 28811da177e4SLinus Torvalds * new window, do not retransmit it. The exception is the 28821da177e4SLinus Torvalds * case, when window is shrunk to zero. In this case 28831da177e4SLinus Torvalds * our retransmit serves as a zero window probe. 28841da177e4SLinus Torvalds */ 28859d4fb27dSJoe Perches if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) && 28869d4fb27dSJoe Perches TCP_SKB_CB(skb)->seq != tp->snd_una) 28871da177e4SLinus Torvalds return -EAGAIN; 28881da177e4SLinus Torvalds 288910d3be56SEric Dumazet len = cur_mss * segs; 289010d3be56SEric Dumazet if (skb->len > len) { 289175c119afSEric Dumazet if (tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb, len, 289275c119afSEric Dumazet cur_mss, GFP_ATOMIC)) 28931da177e4SLinus Torvalds return -ENOMEM; /* We'll try again later. */ 289402276f3cSIlpo Järvinen } else { 2895c52e2421SEric Dumazet if (skb_unclone(skb, GFP_ATOMIC)) 2896c52e2421SEric Dumazet return -ENOMEM; 289710d3be56SEric Dumazet 289810d3be56SEric Dumazet diff = tcp_skb_pcount(skb); 289910d3be56SEric Dumazet tcp_set_skb_tso_segs(skb, cur_mss); 290010d3be56SEric Dumazet diff -= tcp_skb_pcount(skb); 290110d3be56SEric Dumazet if (diff) 290210d3be56SEric Dumazet tcp_adjust_pcount(sk, skb, diff); 290310d3be56SEric Dumazet if (skb->len < cur_mss) 290410d3be56SEric Dumazet tcp_retrans_try_collapse(sk, skb, cur_mss); 29051da177e4SLinus Torvalds } 29061da177e4SLinus Torvalds 290749213555SDaniel Borkmann /* RFC3168, section 6.1.1.1. ECN fallback */ 290849213555SDaniel Borkmann if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN_ECN) == TCPHDR_SYN_ECN) 290949213555SDaniel Borkmann tcp_ecn_clear_syn(sk, skb); 291049213555SDaniel Borkmann 2911678550c6SYuchung Cheng /* Update global and local TCP statistics. */ 2912678550c6SYuchung Cheng segs = tcp_skb_pcount(skb); 2913678550c6SYuchung Cheng TCP_ADD_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS, segs); 2914678550c6SYuchung Cheng if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN) 2915678550c6SYuchung Cheng __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS); 2916678550c6SYuchung Cheng tp->total_retrans += segs; 2917fb31c9b9SWei Wang tp->bytes_retrans += skb->len; 2918678550c6SYuchung Cheng 291950bceae9SThomas Graf /* make sure skb->data is aligned on arches that require it 292050bceae9SThomas Graf * and check if ack-trimming & collapsing extended the headroom 292150bceae9SThomas Graf * beyond what csum_start can cover. 292250bceae9SThomas Graf */ 292350bceae9SThomas Graf if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) || 292450bceae9SThomas Graf skb_headroom(skb) >= 0xFFFF)) { 292510a81980SEric Dumazet struct sk_buff *nskb; 292610a81980SEric Dumazet 2927e2080072SEric Dumazet tcp_skb_tsorted_save(skb) { 292810a81980SEric Dumazet nskb = __pskb_copy(skb, MAX_TCP_HEADER, GFP_ATOMIC); 2929c84a5711SYuchung Cheng err = nskb ? tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC) : 2930117632e6SEric Dumazet -ENOBUFS; 2931e2080072SEric Dumazet } tcp_skb_tsorted_restore(skb); 2932e2080072SEric Dumazet 29335889e2c0SYousuk Seung if (!err) { 2934a7a25630SEric Dumazet tcp_update_skb_after_send(sk, skb, tp->tcp_wstamp_ns); 29355889e2c0SYousuk Seung tcp_rate_skb_sent(sk, skb); 29365889e2c0SYousuk Seung } 2937117632e6SEric Dumazet } else { 2938c84a5711SYuchung Cheng err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 2939117632e6SEric Dumazet } 2940c84a5711SYuchung Cheng 29417f12422cSYuchung Cheng /* To avoid taking spuriously low RTT samples based on a timestamp 29427f12422cSYuchung Cheng * for a transmit that never happened, always mark EVER_RETRANS 29437f12422cSYuchung Cheng */ 29447f12422cSYuchung Cheng TCP_SKB_CB(skb)->sacked |= TCPCB_EVER_RETRANS; 29457f12422cSYuchung Cheng 2946a31ad29eSLawrence Brakmo if (BPF_SOCK_OPS_TEST_FLAG(tp, BPF_SOCK_OPS_RETRANS_CB_FLAG)) 2947a31ad29eSLawrence Brakmo tcp_call_bpf_3arg(sk, BPF_SOCK_OPS_RETRANS_CB, 2948a31ad29eSLawrence Brakmo TCP_SKB_CB(skb)->seq, segs, err); 2949a31ad29eSLawrence Brakmo 2950fc9f3501SEric Dumazet if (likely(!err)) { 2951e086101bSCong Wang trace_tcp_retransmit_skb(sk, skb); 2952678550c6SYuchung Cheng } else if (err != -EBUSY) { 2953ec641b39SYuchung Cheng NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL, segs); 2954fc9f3501SEric Dumazet } 2955c84a5711SYuchung Cheng return err; 295693b174adSYuchung Cheng } 295793b174adSYuchung Cheng 295810d3be56SEric Dumazet int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs) 295993b174adSYuchung Cheng { 296093b174adSYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 296110d3be56SEric Dumazet int err = __tcp_retransmit_skb(sk, skb, segs); 29621da177e4SLinus Torvalds 29631da177e4SLinus Torvalds if (err == 0) { 29641da177e4SLinus Torvalds #if FASTRETRANS_DEBUG > 0 29651da177e4SLinus Torvalds if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) { 2966e87cc472SJoe Perches net_dbg_ratelimited("retrans_out leaked\n"); 29671da177e4SLinus Torvalds } 29681da177e4SLinus Torvalds #endif 29691da177e4SLinus Torvalds TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS; 29701da177e4SLinus Torvalds tp->retrans_out += tcp_skb_pcount(skb); 29717ae18975SYuchung Cheng } 29721da177e4SLinus Torvalds 29737ae18975SYuchung Cheng /* Save stamp of the first (attempted) retransmit. */ 29741da177e4SLinus Torvalds if (!tp->retrans_stamp) 29757faee5c0SEric Dumazet tp->retrans_stamp = tcp_skb_timestamp(skb); 29761da177e4SLinus Torvalds 29776e08d5e3SYuchung Cheng if (tp->undo_retrans < 0) 29786e08d5e3SYuchung Cheng tp->undo_retrans = 0; 29796e08d5e3SYuchung Cheng tp->undo_retrans += tcp_skb_pcount(skb); 29801da177e4SLinus Torvalds return err; 29811da177e4SLinus Torvalds } 29821da177e4SLinus Torvalds 29831da177e4SLinus Torvalds /* This gets called after a retransmit timeout, and the initially 29841da177e4SLinus Torvalds * retransmitted data is acknowledged. It tries to continue 29851da177e4SLinus Torvalds * resending the rest of the retransmit queue, until either 29861da177e4SLinus Torvalds * we've sent it all or the congestion window limit is reached. 29871da177e4SLinus Torvalds */ 29881da177e4SLinus Torvalds void tcp_xmit_retransmit_queue(struct sock *sk) 29891da177e4SLinus Torvalds { 29906687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 2991b9f1f1ceSEric Dumazet struct sk_buff *skb, *rtx_head, *hole = NULL; 29921da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 2993840a3cbeSYuchung Cheng u32 max_segs; 299461eb55f4SIlpo Järvinen int mib_idx; 29956a438bbeSStephen Hemminger 299645e77d31SIlpo Järvinen if (!tp->packets_out) 299745e77d31SIlpo Järvinen return; 299845e77d31SIlpo Järvinen 299975c119afSEric Dumazet rtx_head = tcp_rtx_queue_head(sk); 3000b9f1f1ceSEric Dumazet skb = tp->retransmit_skb_hint ?: rtx_head; 3001ed6e7268SNeal Cardwell max_segs = tcp_tso_segs(sk, tcp_current_mss(sk)); 300275c119afSEric Dumazet skb_rbtree_walk_from(skb) { 3003dca0aaf8SEric Dumazet __u8 sacked; 300410d3be56SEric Dumazet int segs; 30051da177e4SLinus Torvalds 3006218af599SEric Dumazet if (tcp_pacing_check(sk)) 3007218af599SEric Dumazet break; 3008218af599SEric Dumazet 30096a438bbeSStephen Hemminger /* we could do better than to assign each time */ 301051456b29SIan Morris if (!hole) 30116a438bbeSStephen Hemminger tp->retransmit_skb_hint = skb; 30126a438bbeSStephen Hemminger 301310d3be56SEric Dumazet segs = tp->snd_cwnd - tcp_packets_in_flight(tp); 301410d3be56SEric Dumazet if (segs <= 0) 30151da177e4SLinus Torvalds return; 3016dca0aaf8SEric Dumazet sacked = TCP_SKB_CB(skb)->sacked; 3017a3d2e9f8SEric Dumazet /* In case tcp_shift_skb_data() have aggregated large skbs, 3018a3d2e9f8SEric Dumazet * we need to make sure not sending too bigs TSO packets 3019a3d2e9f8SEric Dumazet */ 3020a3d2e9f8SEric Dumazet segs = min_t(int, segs, max_segs); 30210e1c54c2SIlpo Järvinen 3022840a3cbeSYuchung Cheng if (tp->retrans_out >= tp->lost_out) { 3023006f582cSIlpo Järvinen break; 30240e1c54c2SIlpo Järvinen } else if (!(sacked & TCPCB_LOST)) { 302551456b29SIan Morris if (!hole && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED))) 30260e1c54c2SIlpo Järvinen hole = skb; 302761eb55f4SIlpo Järvinen continue; 30281da177e4SLinus Torvalds 30290e1c54c2SIlpo Järvinen } else { 30300e1c54c2SIlpo Järvinen if (icsk->icsk_ca_state != TCP_CA_Loss) 30310e1c54c2SIlpo Järvinen mib_idx = LINUX_MIB_TCPFASTRETRANS; 30320e1c54c2SIlpo Järvinen else 30330e1c54c2SIlpo Järvinen mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS; 30340e1c54c2SIlpo Järvinen } 30350e1c54c2SIlpo Järvinen 30360e1c54c2SIlpo Järvinen if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS)) 303761eb55f4SIlpo Järvinen continue; 303840b215e5SPavel Emelyanov 3039f9616c35SEric Dumazet if (tcp_small_queue_check(sk, skb, 1)) 3040f9616c35SEric Dumazet return; 3041f9616c35SEric Dumazet 304210d3be56SEric Dumazet if (tcp_retransmit_skb(sk, skb, segs)) 30431da177e4SLinus Torvalds return; 304424ab6becSYuchung Cheng 3045de1d6578SYuchung Cheng NET_ADD_STATS(sock_net(sk), mib_idx, tcp_skb_pcount(skb)); 30461da177e4SLinus Torvalds 3047684bad11SYuchung Cheng if (tcp_in_cwnd_reduction(sk)) 3048a262f0cdSNandita Dukkipati tp->prr_out += tcp_skb_pcount(skb); 3049a262f0cdSNandita Dukkipati 305075c119afSEric Dumazet if (skb == rtx_head && 305157dde7f7SYuchung Cheng icsk->icsk_pending != ICSK_TIME_REO_TIMEOUT) 30523f80e08fSEric Dumazet tcp_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 30533f421baaSArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto, 30543f80e08fSEric Dumazet TCP_RTO_MAX, 30553f80e08fSEric Dumazet skb); 30561da177e4SLinus Torvalds } 30571da177e4SLinus Torvalds } 30581da177e4SLinus Torvalds 3059d83769a5SEric Dumazet /* We allow to exceed memory limits for FIN packets to expedite 3060d83769a5SEric Dumazet * connection tear down and (memory) recovery. 3061845704a5SEric Dumazet * Otherwise tcp_send_fin() could be tempted to either delay FIN 3062845704a5SEric Dumazet * or even be forced to close flow without any FIN. 3063a6c5ea4cSEric Dumazet * In general, we want to allow one skb per socket to avoid hangs 3064a6c5ea4cSEric Dumazet * with edge trigger epoll() 3065d83769a5SEric Dumazet */ 3066a6c5ea4cSEric Dumazet void sk_forced_mem_schedule(struct sock *sk, int size) 3067d83769a5SEric Dumazet { 3068e805605cSJohannes Weiner int amt; 3069d83769a5SEric Dumazet 3070d83769a5SEric Dumazet if (size <= sk->sk_forward_alloc) 3071d83769a5SEric Dumazet return; 3072d83769a5SEric Dumazet amt = sk_mem_pages(size); 3073d83769a5SEric Dumazet sk->sk_forward_alloc += amt * SK_MEM_QUANTUM; 3074e805605cSJohannes Weiner sk_memory_allocated_add(sk, amt); 3075e805605cSJohannes Weiner 3076baac50bbSJohannes Weiner if (mem_cgroup_sockets_enabled && sk->sk_memcg) 3077baac50bbSJohannes Weiner mem_cgroup_charge_skmem(sk->sk_memcg, amt); 3078d83769a5SEric Dumazet } 3079d83769a5SEric Dumazet 3080845704a5SEric Dumazet /* Send a FIN. The caller locks the socket for us. 3081845704a5SEric Dumazet * We should try to send a FIN packet really hard, but eventually give up. 30821da177e4SLinus Torvalds */ 30831da177e4SLinus Torvalds void tcp_send_fin(struct sock *sk) 30841da177e4SLinus Torvalds { 3085845704a5SEric Dumazet struct sk_buff *skb, *tskb = tcp_write_queue_tail(sk); 30861da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 30871da177e4SLinus Torvalds 3088845704a5SEric Dumazet /* Optimization, tack on the FIN if we have one skb in write queue and 3089845704a5SEric Dumazet * this skb was not yet sent, or we are under memory pressure. 3090845704a5SEric Dumazet * Note: in the latter case, FIN packet will be sent after a timeout, 3091845704a5SEric Dumazet * as TCP stack thinks it has already been transmitted. 30921da177e4SLinus Torvalds */ 309375c119afSEric Dumazet if (!tskb && tcp_under_memory_pressure(sk)) 309475c119afSEric Dumazet tskb = skb_rb_last(&sk->tcp_rtx_queue); 309575c119afSEric Dumazet 309675c119afSEric Dumazet if (tskb) { 3097845704a5SEric Dumazet TCP_SKB_CB(tskb)->tcp_flags |= TCPHDR_FIN; 3098845704a5SEric Dumazet TCP_SKB_CB(tskb)->end_seq++; 30991da177e4SLinus Torvalds tp->write_seq++; 310075c119afSEric Dumazet if (tcp_write_queue_empty(sk)) { 3101845704a5SEric Dumazet /* This means tskb was already sent. 3102845704a5SEric Dumazet * Pretend we included the FIN on previous transmit. 3103845704a5SEric Dumazet * We need to set tp->snd_nxt to the value it would have 3104845704a5SEric Dumazet * if FIN had been sent. This is because retransmit path 3105845704a5SEric Dumazet * does not change tp->snd_nxt. 3106845704a5SEric Dumazet */ 3107845704a5SEric Dumazet tp->snd_nxt++; 3108845704a5SEric Dumazet return; 3109845704a5SEric Dumazet } 31101da177e4SLinus Torvalds } else { 3111845704a5SEric Dumazet skb = alloc_skb_fclone(MAX_TCP_HEADER, sk->sk_allocation); 3112d1edc085SColin Ian King if (unlikely(!skb)) 3113845704a5SEric Dumazet return; 3114d1edc085SColin Ian King 3115e2080072SEric Dumazet INIT_LIST_HEAD(&skb->tcp_tsorted_anchor); 3116d83769a5SEric Dumazet skb_reserve(skb, MAX_TCP_HEADER); 3117a6c5ea4cSEric Dumazet sk_forced_mem_schedule(sk, skb->truesize); 31181da177e4SLinus Torvalds /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */ 3119e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tp->write_seq, 3120a3433f35SChangli Gao TCPHDR_ACK | TCPHDR_FIN); 31211da177e4SLinus Torvalds tcp_queue_skb(sk, skb); 31221da177e4SLinus Torvalds } 3123845704a5SEric Dumazet __tcp_push_pending_frames(sk, tcp_current_mss(sk), TCP_NAGLE_OFF); 31241da177e4SLinus Torvalds } 31251da177e4SLinus Torvalds 31261da177e4SLinus Torvalds /* We get here when a process closes a file descriptor (either due to 31271da177e4SLinus Torvalds * an explicit close() or as a byproduct of exit()'ing) and there 31281da177e4SLinus Torvalds * was unread data in the receive queue. This behavior is recommended 312965bb723cSGerrit Renker * by RFC 2525, section 2.17. -DaveM 31301da177e4SLinus Torvalds */ 3131dd0fc66fSAl Viro void tcp_send_active_reset(struct sock *sk, gfp_t priority) 31321da177e4SLinus Torvalds { 31331da177e4SLinus Torvalds struct sk_buff *skb; 31341da177e4SLinus Torvalds 31357cc2b043SGao Feng TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS); 31367cc2b043SGao Feng 31371da177e4SLinus Torvalds /* NOTE: No TCP options attached and we never retransmit this. */ 31381da177e4SLinus Torvalds skb = alloc_skb(MAX_TCP_HEADER, priority); 31391da177e4SLinus Torvalds if (!skb) { 31404e673444SPavel Emelyanov NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); 31411da177e4SLinus Torvalds return; 31421da177e4SLinus Torvalds } 31431da177e4SLinus Torvalds 31441da177e4SLinus Torvalds /* Reserve space for headers and prepare control bits. */ 31451da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 3146e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk), 3147a3433f35SChangli Gao TCPHDR_ACK | TCPHDR_RST); 31489a568de4SEric Dumazet tcp_mstamp_refresh(tcp_sk(sk)); 31491da177e4SLinus Torvalds /* Send it off. */ 3150dfb4b9dcSDavid S. Miller if (tcp_transmit_skb(sk, skb, 0, priority)) 31514e673444SPavel Emelyanov NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); 3152c24b14c4SSong Liu 3153c24b14c4SSong Liu /* skb of trace_tcp_send_reset() keeps the skb that caused RST, 3154c24b14c4SSong Liu * skb here is different to the troublesome skb, so use NULL 3155c24b14c4SSong Liu */ 3156c24b14c4SSong Liu trace_tcp_send_reset(sk, NULL); 31571da177e4SLinus Torvalds } 31581da177e4SLinus Torvalds 315967edfef7SAndi Kleen /* Send a crossed SYN-ACK during socket establishment. 316067edfef7SAndi Kleen * WARNING: This routine must only be called when we have already sent 31611da177e4SLinus Torvalds * a SYN packet that crossed the incoming SYN that caused this routine 31621da177e4SLinus Torvalds * to get called. If this assumption fails then the initial rcv_wnd 31631da177e4SLinus Torvalds * and rcv_wscale values will not be correct. 31641da177e4SLinus Torvalds */ 31651da177e4SLinus Torvalds int tcp_send_synack(struct sock *sk) 31661da177e4SLinus Torvalds { 31671da177e4SLinus Torvalds struct sk_buff *skb; 31681da177e4SLinus Torvalds 316975c119afSEric Dumazet skb = tcp_rtx_queue_head(sk); 317051456b29SIan Morris if (!skb || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) { 317175c119afSEric Dumazet pr_err("%s: wrong queue state\n", __func__); 31721da177e4SLinus Torvalds return -EFAULT; 31731da177e4SLinus Torvalds } 31744de075e0SEric Dumazet if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) { 31751da177e4SLinus Torvalds if (skb_cloned(skb)) { 3176e2080072SEric Dumazet struct sk_buff *nskb; 3177e2080072SEric Dumazet 3178e2080072SEric Dumazet tcp_skb_tsorted_save(skb) { 3179e2080072SEric Dumazet nskb = skb_copy(skb, GFP_ATOMIC); 3180e2080072SEric Dumazet } tcp_skb_tsorted_restore(skb); 318151456b29SIan Morris if (!nskb) 31821da177e4SLinus Torvalds return -ENOMEM; 3183e2080072SEric Dumazet INIT_LIST_HEAD(&nskb->tcp_tsorted_anchor); 318475c119afSEric Dumazet tcp_rtx_queue_unlink_and_free(skb, sk); 3185f4a775d1SEric Dumazet __skb_header_release(nskb); 318675c119afSEric Dumazet tcp_rbtree_insert(&sk->tcp_rtx_queue, nskb); 31873ab224beSHideo Aoki sk->sk_wmem_queued += nskb->truesize; 31883ab224beSHideo Aoki sk_mem_charge(sk, nskb->truesize); 31891da177e4SLinus Torvalds skb = nskb; 31901da177e4SLinus Torvalds } 31911da177e4SLinus Torvalds 31924de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK; 3193735d3831SFlorian Westphal tcp_ecn_send_synack(sk, skb); 31941da177e4SLinus Torvalds } 3195dfb4b9dcSDavid S. Miller return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 31961da177e4SLinus Torvalds } 31971da177e4SLinus Torvalds 31984aea39c1SEric Dumazet /** 31994aea39c1SEric Dumazet * tcp_make_synack - Prepare a SYN-ACK. 32004aea39c1SEric Dumazet * sk: listener socket 32014aea39c1SEric Dumazet * dst: dst entry attached to the SYNACK 32024aea39c1SEric Dumazet * req: request_sock pointer 32034aea39c1SEric Dumazet * 32044aea39c1SEric Dumazet * Allocate one skb and build a SYNACK packet. 32054aea39c1SEric Dumazet * @dst is consumed : Caller should not use it again. 32064aea39c1SEric Dumazet */ 32075d062de7SEric Dumazet struct sk_buff *tcp_make_synack(const struct sock *sk, struct dst_entry *dst, 3208e6b4d113SWilliam Allen Simpson struct request_sock *req, 3209ca6fb065SEric Dumazet struct tcp_fastopen_cookie *foc, 3210b3d05147SEric Dumazet enum tcp_synack_type synack_type) 32111da177e4SLinus Torvalds { 32122e6599cbSArnaldo Carvalho de Melo struct inet_request_sock *ireq = inet_rsk(req); 32135d062de7SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 321480f03e27SEric Dumazet struct tcp_md5sig_key *md5 = NULL; 32155d062de7SEric Dumazet struct tcp_out_options opts; 32165d062de7SEric Dumazet struct sk_buff *skb; 3217bd0388aeSWilliam Allen Simpson int tcp_header_size; 32185d062de7SEric Dumazet struct tcphdr *th; 3219f5fff5dcSTom Quetchenbach int mss; 32201da177e4SLinus Torvalds 3221ca6fb065SEric Dumazet skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC); 32224aea39c1SEric Dumazet if (unlikely(!skb)) { 32234aea39c1SEric Dumazet dst_release(dst); 32241da177e4SLinus Torvalds return NULL; 32254aea39c1SEric Dumazet } 32261da177e4SLinus Torvalds /* Reserve space for headers. */ 32271da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 32281da177e4SLinus Torvalds 3229b3d05147SEric Dumazet switch (synack_type) { 3230b3d05147SEric Dumazet case TCP_SYNACK_NORMAL: 32319e17f8a4SEric Dumazet skb_set_owner_w(skb, req_to_sk(req)); 3232b3d05147SEric Dumazet break; 3233b3d05147SEric Dumazet case TCP_SYNACK_COOKIE: 3234b3d05147SEric Dumazet /* Under synflood, we do not attach skb to a socket, 3235b3d05147SEric Dumazet * to avoid false sharing. 3236b3d05147SEric Dumazet */ 3237b3d05147SEric Dumazet break; 3238b3d05147SEric Dumazet case TCP_SYNACK_FASTOPEN: 3239ca6fb065SEric Dumazet /* sk is a const pointer, because we want to express multiple 3240ca6fb065SEric Dumazet * cpu might call us concurrently. 3241ca6fb065SEric Dumazet * sk->sk_wmem_alloc in an atomic, we can promote to rw. 3242ca6fb065SEric Dumazet */ 3243ca6fb065SEric Dumazet skb_set_owner_w(skb, (struct sock *)sk); 3244b3d05147SEric Dumazet break; 3245ca6fb065SEric Dumazet } 32464aea39c1SEric Dumazet skb_dst_set(skb, dst); 32471da177e4SLinus Torvalds 32483541f9e8SEric Dumazet mss = tcp_mss_clamp(tp, dst_metric_advmss(dst)); 3249f5fff5dcSTom Quetchenbach 325033ad798cSAdam Langley memset(&opts, 0, sizeof(opts)); 32518b5f12d0SFlorian Westphal #ifdef CONFIG_SYN_COOKIES 32528b5f12d0SFlorian Westphal if (unlikely(req->cookie_ts)) 3253d3edd06eSEric Dumazet skb->skb_mstamp_ns = cookie_init_timestamp(req); 32548b5f12d0SFlorian Westphal else 32558b5f12d0SFlorian Westphal #endif 32569e450c1eSYuchung Cheng { 3257d3edd06eSEric Dumazet skb->skb_mstamp_ns = tcp_clock_ns(); 32589e450c1eSYuchung Cheng if (!tcp_rsk(req)->snt_synack) /* Timestamp first SYNACK */ 32599e450c1eSYuchung Cheng tcp_rsk(req)->snt_synack = tcp_skb_timestamp_us(skb); 32609e450c1eSYuchung Cheng } 326180f03e27SEric Dumazet 326280f03e27SEric Dumazet #ifdef CONFIG_TCP_MD5SIG 326380f03e27SEric Dumazet rcu_read_lock(); 3264fd3a154aSEric Dumazet md5 = tcp_rsk(req)->af_specific->req_md5_lookup(sk, req_to_sk(req)); 326580f03e27SEric Dumazet #endif 326658d607d3SEric Dumazet skb_set_hash(skb, tcp_rsk(req)->txhash, PKT_HASH_TYPE_L4); 326760e2a778SUrsula Braun tcp_header_size = tcp_synack_options(sk, req, mss, skb, &opts, md5, 326860e2a778SUrsula Braun foc) + sizeof(*th); 326933ad798cSAdam Langley 3270aa8223c7SArnaldo Carvalho de Melo skb_push(skb, tcp_header_size); 3271aa8223c7SArnaldo Carvalho de Melo skb_reset_transport_header(skb); 32721da177e4SLinus Torvalds 3273ea1627c2SEric Dumazet th = (struct tcphdr *)skb->data; 32741da177e4SLinus Torvalds memset(th, 0, sizeof(struct tcphdr)); 32751da177e4SLinus Torvalds th->syn = 1; 32761da177e4SLinus Torvalds th->ack = 1; 32776ac705b1SEric Dumazet tcp_ecn_make_synack(req, th); 3278b44084c2SEric Dumazet th->source = htons(ireq->ir_num); 3279634fb979SEric Dumazet th->dest = ireq->ir_rmt_port; 3280e05a90ecSJamal Hadi Salim skb->mark = ireq->ir_mark; 32813b117750SEric Dumazet skb->ip_summed = CHECKSUM_PARTIAL; 32823b117750SEric Dumazet th->seq = htonl(tcp_rsk(req)->snt_isn); 32838336886fSJerry Chu /* XXX data is queued and acked as is. No buffer/window check */ 32848336886fSJerry Chu th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt); 32851da177e4SLinus Torvalds 32861da177e4SLinus Torvalds /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */ 3287ed53d0abSEric Dumazet th->window = htons(min(req->rsk_rcv_wnd, 65535U)); 32885d062de7SEric Dumazet tcp_options_write((__be32 *)(th + 1), NULL, &opts); 32891da177e4SLinus Torvalds th->doff = (tcp_header_size >> 2); 329090bbcc60SEric Dumazet __TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTSEGS); 3291cfb6eeb4SYOSHIFUJI Hideaki 3292cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 3293cfb6eeb4SYOSHIFUJI Hideaki /* Okay, we have all we need - do the md5 hash if needed */ 329480f03e27SEric Dumazet if (md5) 3295bd0388aeSWilliam Allen Simpson tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location, 329639f8e58eSEric Dumazet md5, req_to_sk(req), skb); 329780f03e27SEric Dumazet rcu_read_unlock(); 3298cfb6eeb4SYOSHIFUJI Hideaki #endif 3299cfb6eeb4SYOSHIFUJI Hideaki 3300b50edd78SEric Dumazet /* Do not fool tcpdump (if any), clean our debris */ 33012456e855SThomas Gleixner skb->tstamp = 0; 33021da177e4SLinus Torvalds return skb; 33031da177e4SLinus Torvalds } 33044bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_make_synack); 33051da177e4SLinus Torvalds 330681164413SDaniel Borkmann static void tcp_ca_dst_init(struct sock *sk, const struct dst_entry *dst) 330781164413SDaniel Borkmann { 330881164413SDaniel Borkmann struct inet_connection_sock *icsk = inet_csk(sk); 330981164413SDaniel Borkmann const struct tcp_congestion_ops *ca; 331081164413SDaniel Borkmann u32 ca_key = dst_metric(dst, RTAX_CC_ALGO); 331181164413SDaniel Borkmann 331281164413SDaniel Borkmann if (ca_key == TCP_CA_UNSPEC) 331381164413SDaniel Borkmann return; 331481164413SDaniel Borkmann 331581164413SDaniel Borkmann rcu_read_lock(); 331681164413SDaniel Borkmann ca = tcp_ca_find_key(ca_key); 331781164413SDaniel Borkmann if (likely(ca && try_module_get(ca->owner))) { 331881164413SDaniel Borkmann module_put(icsk->icsk_ca_ops->owner); 331981164413SDaniel Borkmann icsk->icsk_ca_dst_locked = tcp_ca_dst_locked(dst); 332081164413SDaniel Borkmann icsk->icsk_ca_ops = ca; 332181164413SDaniel Borkmann } 332281164413SDaniel Borkmann rcu_read_unlock(); 332381164413SDaniel Borkmann } 332481164413SDaniel Borkmann 332567edfef7SAndi Kleen /* Do all connect socket setups that can be done AF independent. */ 3326f7e56a76Sstephen hemminger static void tcp_connect_init(struct sock *sk) 33271da177e4SLinus Torvalds { 3328cf533ea5SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 33291da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 33301da177e4SLinus Torvalds __u8 rcv_wscale; 333113d3b1ebSLawrence Brakmo u32 rcv_wnd; 33321da177e4SLinus Torvalds 33331da177e4SLinus Torvalds /* We'll fix this up when we get a response from the other end. 33341da177e4SLinus Torvalds * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT. 33351da177e4SLinus Torvalds */ 33365d2ed052SEric Dumazet tp->tcp_header_len = sizeof(struct tcphdr); 33375d2ed052SEric Dumazet if (sock_net(sk)->ipv4.sysctl_tcp_timestamps) 33385d2ed052SEric Dumazet tp->tcp_header_len += TCPOLEN_TSTAMP_ALIGNED; 33391da177e4SLinus Torvalds 3340cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 334100db4124SIan Morris if (tp->af_specific->md5_lookup(sk, sk)) 3342cfb6eeb4SYOSHIFUJI Hideaki tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED; 3343cfb6eeb4SYOSHIFUJI Hideaki #endif 3344cfb6eeb4SYOSHIFUJI Hideaki 33451da177e4SLinus Torvalds /* If user gave his TCP_MAXSEG, record it to clamp */ 33461da177e4SLinus Torvalds if (tp->rx_opt.user_mss) 33471da177e4SLinus Torvalds tp->rx_opt.mss_clamp = tp->rx_opt.user_mss; 33481da177e4SLinus Torvalds tp->max_window = 0; 33495d424d5aSJohn Heffner tcp_mtup_init(sk); 33501da177e4SLinus Torvalds tcp_sync_mss(sk, dst_mtu(dst)); 33511da177e4SLinus Torvalds 335281164413SDaniel Borkmann tcp_ca_dst_init(sk, dst); 335381164413SDaniel Borkmann 33541da177e4SLinus Torvalds if (!tp->window_clamp) 33551da177e4SLinus Torvalds tp->window_clamp = dst_metric(dst, RTAX_WINDOW); 33563541f9e8SEric Dumazet tp->advmss = tcp_mss_clamp(tp, dst_metric_advmss(dst)); 3357f5fff5dcSTom Quetchenbach 33581da177e4SLinus Torvalds tcp_initialize_rcv_mss(sk); 33591da177e4SLinus Torvalds 3360e88c64f0SHagen Paul Pfeifer /* limit the window selection if the user enforce a smaller rx buffer */ 3361e88c64f0SHagen Paul Pfeifer if (sk->sk_userlocks & SOCK_RCVBUF_LOCK && 3362e88c64f0SHagen Paul Pfeifer (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0)) 3363e88c64f0SHagen Paul Pfeifer tp->window_clamp = tcp_full_space(sk); 3364e88c64f0SHagen Paul Pfeifer 336513d3b1ebSLawrence Brakmo rcv_wnd = tcp_rwnd_init_bpf(sk); 336613d3b1ebSLawrence Brakmo if (rcv_wnd == 0) 336713d3b1ebSLawrence Brakmo rcv_wnd = dst_metric(dst, RTAX_INITRWND); 336813d3b1ebSLawrence Brakmo 3369ceef9ab6SEric Dumazet tcp_select_initial_window(sk, tcp_full_space(sk), 33701da177e4SLinus Torvalds tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0), 33711da177e4SLinus Torvalds &tp->rcv_wnd, 33721da177e4SLinus Torvalds &tp->window_clamp, 33739bb37ef0SEric Dumazet sock_net(sk)->ipv4.sysctl_tcp_window_scaling, 337431d12926Slaurent chavey &rcv_wscale, 337513d3b1ebSLawrence Brakmo rcv_wnd); 33761da177e4SLinus Torvalds 33771da177e4SLinus Torvalds tp->rx_opt.rcv_wscale = rcv_wscale; 33781da177e4SLinus Torvalds tp->rcv_ssthresh = tp->rcv_wnd; 33791da177e4SLinus Torvalds 33801da177e4SLinus Torvalds sk->sk_err = 0; 33811da177e4SLinus Torvalds sock_reset_flag(sk, SOCK_DONE); 33821da177e4SLinus Torvalds tp->snd_wnd = 0; 3383ee7537b6SHantzis Fotis tcp_init_wl(tp, 0); 33847f582b24SEric Dumazet tcp_write_queue_purge(sk); 33851da177e4SLinus Torvalds tp->snd_una = tp->write_seq; 33861da177e4SLinus Torvalds tp->snd_sml = tp->write_seq; 338733f5f57eSIlpo Järvinen tp->snd_up = tp->write_seq; 3388370816aeSPavel Emelyanov tp->snd_nxt = tp->write_seq; 3389ee995283SPavel Emelyanov 3390ee995283SPavel Emelyanov if (likely(!tp->repair)) 33911da177e4SLinus Torvalds tp->rcv_nxt = 0; 3392c7781a6eSAndrew Vagin else 339370eabf0eSEric Dumazet tp->rcv_tstamp = tcp_jiffies32; 3394ee995283SPavel Emelyanov tp->rcv_wup = tp->rcv_nxt; 3395ee995283SPavel Emelyanov tp->copied_seq = tp->rcv_nxt; 33961da177e4SLinus Torvalds 33978550f328SLawrence Brakmo inet_csk(sk)->icsk_rto = tcp_timeout_init(sk); 3398463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_retransmits = 0; 33991da177e4SLinus Torvalds tcp_clear_retrans(tp); 34001da177e4SLinus Torvalds } 34011da177e4SLinus Torvalds 3402783237e8SYuchung Cheng static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb) 3403783237e8SYuchung Cheng { 3404783237e8SYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 3405783237e8SYuchung Cheng struct tcp_skb_cb *tcb = TCP_SKB_CB(skb); 3406783237e8SYuchung Cheng 3407783237e8SYuchung Cheng tcb->end_seq += skb->len; 3408f4a775d1SEric Dumazet __skb_header_release(skb); 3409783237e8SYuchung Cheng sk->sk_wmem_queued += skb->truesize; 3410783237e8SYuchung Cheng sk_mem_charge(sk, skb->truesize); 3411783237e8SYuchung Cheng tp->write_seq = tcb->end_seq; 3412783237e8SYuchung Cheng tp->packets_out += tcp_skb_pcount(skb); 3413783237e8SYuchung Cheng } 3414783237e8SYuchung Cheng 3415783237e8SYuchung Cheng /* Build and send a SYN with data and (cached) Fast Open cookie. However, 3416783237e8SYuchung Cheng * queue a data-only packet after the regular SYN, such that regular SYNs 3417783237e8SYuchung Cheng * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges 3418783237e8SYuchung Cheng * only the SYN sequence, the data are retransmitted in the first ACK. 3419783237e8SYuchung Cheng * If cookie is not cached or other error occurs, falls back to send a 3420783237e8SYuchung Cheng * regular SYN with Fast Open cookie request option. 3421783237e8SYuchung Cheng */ 3422783237e8SYuchung Cheng static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn) 3423783237e8SYuchung Cheng { 3424783237e8SYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 3425783237e8SYuchung Cheng struct tcp_fastopen_request *fo = tp->fastopen_req; 3426065263f4SWei Wang int space, err = 0; 3427355a901eSEric Dumazet struct sk_buff *syn_data; 3428783237e8SYuchung Cheng 342967da22d2SYuchung Cheng tp->rx_opt.mss_clamp = tp->advmss; /* If MSS is not cached */ 3430065263f4SWei Wang if (!tcp_fastopen_cookie_check(sk, &tp->rx_opt.mss_clamp, &fo->cookie)) 3431783237e8SYuchung Cheng goto fallback; 3432783237e8SYuchung Cheng 3433783237e8SYuchung Cheng /* MSS for SYN-data is based on cached MSS and bounded by PMTU and 3434783237e8SYuchung Cheng * user-MSS. Reserve maximum option space for middleboxes that add 3435783237e8SYuchung Cheng * private TCP options. The cost is reduced data space in SYN :( 3436783237e8SYuchung Cheng */ 34373541f9e8SEric Dumazet tp->rx_opt.mss_clamp = tcp_mss_clamp(tp, tp->rx_opt.mss_clamp); 34383541f9e8SEric Dumazet 34391b63edd6SYuchung Cheng space = __tcp_mtu_to_mss(sk, inet_csk(sk)->icsk_pmtu_cookie) - 3440783237e8SYuchung Cheng MAX_TCP_OPTION_SPACE; 3441783237e8SYuchung Cheng 3442f5ddcbbbSEric Dumazet space = min_t(size_t, space, fo->size); 3443f5ddcbbbSEric Dumazet 3444f5ddcbbbSEric Dumazet /* limit to order-0 allocations */ 3445f5ddcbbbSEric Dumazet space = min_t(size_t, space, SKB_MAX_HEAD(MAX_TCP_HEADER)); 3446f5ddcbbbSEric Dumazet 3447eb934478SEric Dumazet syn_data = sk_stream_alloc_skb(sk, space, sk->sk_allocation, false); 3448355a901eSEric Dumazet if (!syn_data) 3449783237e8SYuchung Cheng goto fallback; 3450355a901eSEric Dumazet syn_data->ip_summed = CHECKSUM_PARTIAL; 3451355a901eSEric Dumazet memcpy(syn_data->cb, syn->cb, sizeof(syn->cb)); 345207e100f9SEric Dumazet if (space) { 345307e100f9SEric Dumazet int copied = copy_from_iter(skb_put(syn_data, space), space, 345457be5bdaSAl Viro &fo->data->msg_iter); 345557be5bdaSAl Viro if (unlikely(!copied)) { 3456ba233b34SEric Dumazet tcp_skb_tsorted_anchor_cleanup(syn_data); 3457355a901eSEric Dumazet kfree_skb(syn_data); 3458783237e8SYuchung Cheng goto fallback; 3459783237e8SYuchung Cheng } 346057be5bdaSAl Viro if (copied != space) { 346157be5bdaSAl Viro skb_trim(syn_data, copied); 346257be5bdaSAl Viro space = copied; 346357be5bdaSAl Viro } 3464f859a448SWillem de Bruijn skb_zcopy_set(syn_data, fo->uarg, NULL); 346507e100f9SEric Dumazet } 3466355a901eSEric Dumazet /* No more data pending in inet_wait_for_connect() */ 3467355a901eSEric Dumazet if (space == fo->size) 3468355a901eSEric Dumazet fo->data = NULL; 3469355a901eSEric Dumazet fo->copied = space; 3470783237e8SYuchung Cheng 3471355a901eSEric Dumazet tcp_connect_queue_skb(sk, syn_data); 34720f87230dSFrancis Yan if (syn_data->len) 34730f87230dSFrancis Yan tcp_chrono_start(sk, TCP_CHRONO_BUSY); 3474355a901eSEric Dumazet 3475355a901eSEric Dumazet err = tcp_transmit_skb(sk, syn_data, 1, sk->sk_allocation); 3476355a901eSEric Dumazet 3477d3edd06eSEric Dumazet syn->skb_mstamp_ns = syn_data->skb_mstamp_ns; 3478355a901eSEric Dumazet 3479355a901eSEric Dumazet /* Now full SYN+DATA was cloned and sent (or not), 3480355a901eSEric Dumazet * remove the SYN from the original skb (syn_data) 3481355a901eSEric Dumazet * we keep in write queue in case of a retransmit, as we 3482355a901eSEric Dumazet * also have the SYN packet (with no data) in the same queue. 3483431a9124SEric Dumazet */ 3484355a901eSEric Dumazet TCP_SKB_CB(syn_data)->seq++; 3485355a901eSEric Dumazet TCP_SKB_CB(syn_data)->tcp_flags = TCPHDR_ACK | TCPHDR_PSH; 3486355a901eSEric Dumazet if (!err) { 348767da22d2SYuchung Cheng tp->syn_data = (fo->copied > 0); 348875c119afSEric Dumazet tcp_rbtree_insert(&sk->tcp_rtx_queue, syn_data); 3489f19c29e3SYuchung Cheng NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT); 3490783237e8SYuchung Cheng goto done; 3491783237e8SYuchung Cheng } 3492783237e8SYuchung Cheng 349375c119afSEric Dumazet /* data was not sent, put it in write_queue */ 349475c119afSEric Dumazet __skb_queue_tail(&sk->sk_write_queue, syn_data); 3495b5b7db8dSEric Dumazet tp->packets_out -= tcp_skb_pcount(syn_data); 3496b5b7db8dSEric Dumazet 3497783237e8SYuchung Cheng fallback: 3498783237e8SYuchung Cheng /* Send a regular SYN with Fast Open cookie request option */ 3499783237e8SYuchung Cheng if (fo->cookie.len > 0) 3500783237e8SYuchung Cheng fo->cookie.len = 0; 3501783237e8SYuchung Cheng err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation); 3502783237e8SYuchung Cheng if (err) 3503783237e8SYuchung Cheng tp->syn_fastopen = 0; 3504783237e8SYuchung Cheng done: 3505783237e8SYuchung Cheng fo->cookie.len = -1; /* Exclude Fast Open option for SYN retries */ 3506783237e8SYuchung Cheng return err; 3507783237e8SYuchung Cheng } 3508783237e8SYuchung Cheng 350967edfef7SAndi Kleen /* Build a SYN and send it off. */ 35101da177e4SLinus Torvalds int tcp_connect(struct sock *sk) 35111da177e4SLinus Torvalds { 35121da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 35131da177e4SLinus Torvalds struct sk_buff *buff; 3514ee586811SEric Paris int err; 35151da177e4SLinus Torvalds 3516de525be2SLawrence Brakmo tcp_call_bpf(sk, BPF_SOCK_OPS_TCP_CONNECT_CB, 0, NULL); 35178ba60924SEric Dumazet 35188ba60924SEric Dumazet if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk)) 35198ba60924SEric Dumazet return -EHOSTUNREACH; /* Routing failure or similar. */ 35208ba60924SEric Dumazet 35211da177e4SLinus Torvalds tcp_connect_init(sk); 35221da177e4SLinus Torvalds 35232b916477SAndrey Vagin if (unlikely(tp->repair)) { 35242b916477SAndrey Vagin tcp_finish_connect(sk, NULL); 35252b916477SAndrey Vagin return 0; 35262b916477SAndrey Vagin } 35272b916477SAndrey Vagin 3528eb934478SEric Dumazet buff = sk_stream_alloc_skb(sk, 0, sk->sk_allocation, true); 3529355a901eSEric Dumazet if (unlikely(!buff)) 35301da177e4SLinus Torvalds return -ENOBUFS; 35311da177e4SLinus Torvalds 3532a3433f35SChangli Gao tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN); 35339a568de4SEric Dumazet tcp_mstamp_refresh(tp); 35349a568de4SEric Dumazet tp->retrans_stamp = tcp_time_stamp(tp); 3535783237e8SYuchung Cheng tcp_connect_queue_skb(sk, buff); 3536735d3831SFlorian Westphal tcp_ecn_send_syn(sk, buff); 353775c119afSEric Dumazet tcp_rbtree_insert(&sk->tcp_rtx_queue, buff); 35381da177e4SLinus Torvalds 3539783237e8SYuchung Cheng /* Send off SYN; include data in Fast Open. */ 3540783237e8SYuchung Cheng err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) : 3541783237e8SYuchung Cheng tcp_transmit_skb(sk, buff, 1, sk->sk_allocation); 3542ee586811SEric Paris if (err == -ECONNREFUSED) 3543ee586811SEric Paris return err; 3544bd37a088SWei Yongjun 3545bd37a088SWei Yongjun /* We change tp->snd_nxt after the tcp_transmit_skb() call 3546bd37a088SWei Yongjun * in order to make this packet get counted in tcpOutSegs. 3547bd37a088SWei Yongjun */ 3548bd37a088SWei Yongjun tp->snd_nxt = tp->write_seq; 3549bd37a088SWei Yongjun tp->pushed_seq = tp->write_seq; 3550b5b7db8dSEric Dumazet buff = tcp_send_head(sk); 3551b5b7db8dSEric Dumazet if (unlikely(buff)) { 3552b5b7db8dSEric Dumazet tp->snd_nxt = TCP_SKB_CB(buff)->seq; 3553b5b7db8dSEric Dumazet tp->pushed_seq = TCP_SKB_CB(buff)->seq; 3554b5b7db8dSEric Dumazet } 355581cc8a75SPavel Emelyanov TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS); 35561da177e4SLinus Torvalds 35571da177e4SLinus Torvalds /* Timer for repeating the SYN until an answer. */ 35583f421baaSArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 35593f421baaSArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto, TCP_RTO_MAX); 35601da177e4SLinus Torvalds return 0; 35611da177e4SLinus Torvalds } 35624bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_connect); 35631da177e4SLinus Torvalds 35641da177e4SLinus Torvalds /* Send out a delayed ack, the caller does the policy checking 35651da177e4SLinus Torvalds * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check() 35661da177e4SLinus Torvalds * for details. 35671da177e4SLinus Torvalds */ 35681da177e4SLinus Torvalds void tcp_send_delayed_ack(struct sock *sk) 35691da177e4SLinus Torvalds { 3570463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 3571463c84b9SArnaldo Carvalho de Melo int ato = icsk->icsk_ack.ato; 35721da177e4SLinus Torvalds unsigned long timeout; 35731da177e4SLinus Torvalds 35741da177e4SLinus Torvalds if (ato > TCP_DELACK_MIN) { 3575463c84b9SArnaldo Carvalho de Melo const struct tcp_sock *tp = tcp_sk(sk); 35761da177e4SLinus Torvalds int max_ato = HZ / 2; 35771da177e4SLinus Torvalds 357831954cd8SWei Wang if (inet_csk_in_pingpong_mode(sk) || 3579056834d9SIlpo Järvinen (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)) 35801da177e4SLinus Torvalds max_ato = TCP_DELACK_MAX; 35811da177e4SLinus Torvalds 35821da177e4SLinus Torvalds /* Slow path, intersegment interval is "high". */ 35831da177e4SLinus Torvalds 35841da177e4SLinus Torvalds /* If some rtt estimate is known, use it to bound delayed ack. 3585463c84b9SArnaldo Carvalho de Melo * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements 35861da177e4SLinus Torvalds * directly. 35871da177e4SLinus Torvalds */ 3588740b0f18SEric Dumazet if (tp->srtt_us) { 3589740b0f18SEric Dumazet int rtt = max_t(int, usecs_to_jiffies(tp->srtt_us >> 3), 3590740b0f18SEric Dumazet TCP_DELACK_MIN); 35911da177e4SLinus Torvalds 35921da177e4SLinus Torvalds if (rtt < max_ato) 35931da177e4SLinus Torvalds max_ato = rtt; 35941da177e4SLinus Torvalds } 35951da177e4SLinus Torvalds 35961da177e4SLinus Torvalds ato = min(ato, max_ato); 35971da177e4SLinus Torvalds } 35981da177e4SLinus Torvalds 35991da177e4SLinus Torvalds /* Stay within the limit we were given */ 36001da177e4SLinus Torvalds timeout = jiffies + ato; 36011da177e4SLinus Torvalds 36021da177e4SLinus Torvalds /* Use new timeout only if there wasn't a older one earlier. */ 3603463c84b9SArnaldo Carvalho de Melo if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) { 36041da177e4SLinus Torvalds /* If delack timer was blocked or is about to expire, 36051da177e4SLinus Torvalds * send ACK now. 36061da177e4SLinus Torvalds */ 3607463c84b9SArnaldo Carvalho de Melo if (icsk->icsk_ack.blocked || 3608463c84b9SArnaldo Carvalho de Melo time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) { 36091da177e4SLinus Torvalds tcp_send_ack(sk); 36101da177e4SLinus Torvalds return; 36111da177e4SLinus Torvalds } 36121da177e4SLinus Torvalds 3613463c84b9SArnaldo Carvalho de Melo if (!time_before(timeout, icsk->icsk_ack.timeout)) 3614463c84b9SArnaldo Carvalho de Melo timeout = icsk->icsk_ack.timeout; 36151da177e4SLinus Torvalds } 3616463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER; 3617463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.timeout = timeout; 3618463c84b9SArnaldo Carvalho de Melo sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout); 36191da177e4SLinus Torvalds } 36201da177e4SLinus Torvalds 36211da177e4SLinus Torvalds /* This routine sends an ack and also updates the window. */ 36222987babbSYuchung Cheng void __tcp_send_ack(struct sock *sk, u32 rcv_nxt) 36231da177e4SLinus Torvalds { 36241da177e4SLinus Torvalds struct sk_buff *buff; 36251da177e4SLinus Torvalds 3626058dc334SIlpo Järvinen /* If we have been reset, we may not send again. */ 3627058dc334SIlpo Järvinen if (sk->sk_state == TCP_CLOSE) 3628058dc334SIlpo Järvinen return; 3629058dc334SIlpo Järvinen 36301da177e4SLinus Torvalds /* We are not putting this on the write queue, so 36311da177e4SLinus Torvalds * tcp_transmit_skb() will set the ownership to this 36321da177e4SLinus Torvalds * sock. 36331da177e4SLinus Torvalds */ 36347450aaf6SEric Dumazet buff = alloc_skb(MAX_TCP_HEADER, 36357450aaf6SEric Dumazet sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN)); 36367450aaf6SEric Dumazet if (unlikely(!buff)) { 3637463c84b9SArnaldo Carvalho de Melo inet_csk_schedule_ack(sk); 3638463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN; 36393f421baaSArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK, 36403f421baaSArnaldo Carvalho de Melo TCP_DELACK_MAX, TCP_RTO_MAX); 36411da177e4SLinus Torvalds return; 36421da177e4SLinus Torvalds } 36431da177e4SLinus Torvalds 36441da177e4SLinus Torvalds /* Reserve space for headers and prepare control bits. */ 36451da177e4SLinus Torvalds skb_reserve(buff, MAX_TCP_HEADER); 3646a3433f35SChangli Gao tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK); 36471da177e4SLinus Torvalds 364898781965SEric Dumazet /* We do not want pure acks influencing TCP Small Queues or fq/pacing 364998781965SEric Dumazet * too much. 365098781965SEric Dumazet * SKB_TRUESIZE(max(1 .. 66, MAX_TCP_HEADER)) is unfortunately ~784 365198781965SEric Dumazet */ 365298781965SEric Dumazet skb_set_tcp_pure_ack(buff); 365398781965SEric Dumazet 36541da177e4SLinus Torvalds /* Send it off, this clears delayed acks for us. */ 36552987babbSYuchung Cheng __tcp_transmit_skb(sk, buff, 0, (__force gfp_t)0, rcv_nxt); 36561da177e4SLinus Torvalds } 365727cde44aSYuchung Cheng EXPORT_SYMBOL_GPL(__tcp_send_ack); 36582987babbSYuchung Cheng 36592987babbSYuchung Cheng void tcp_send_ack(struct sock *sk) 36602987babbSYuchung Cheng { 36612987babbSYuchung Cheng __tcp_send_ack(sk, tcp_sk(sk)->rcv_nxt); 36621da177e4SLinus Torvalds } 36631da177e4SLinus Torvalds 36641da177e4SLinus Torvalds /* This routine sends a packet with an out of date sequence 36651da177e4SLinus Torvalds * number. It assumes the other end will try to ack it. 36661da177e4SLinus Torvalds * 36671da177e4SLinus Torvalds * Question: what should we make while urgent mode? 36681da177e4SLinus Torvalds * 4.4BSD forces sending single byte of data. We cannot send 36691da177e4SLinus Torvalds * out of window data, because we have SND.NXT==SND.MAX... 36701da177e4SLinus Torvalds * 36711da177e4SLinus Torvalds * Current solution: to send TWO zero-length segments in urgent mode: 36721da177e4SLinus Torvalds * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is 36731da177e4SLinus Torvalds * out-of-date with SND.UNA-1 to probe window. 36741da177e4SLinus Torvalds */ 3675e520af48SEric Dumazet static int tcp_xmit_probe_skb(struct sock *sk, int urgent, int mib) 36761da177e4SLinus Torvalds { 36771da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 36781da177e4SLinus Torvalds struct sk_buff *skb; 36791da177e4SLinus Torvalds 36801da177e4SLinus Torvalds /* We don't queue it, tcp_transmit_skb() sets ownership. */ 36817450aaf6SEric Dumazet skb = alloc_skb(MAX_TCP_HEADER, 36827450aaf6SEric Dumazet sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN)); 368351456b29SIan Morris if (!skb) 36841da177e4SLinus Torvalds return -1; 36851da177e4SLinus Torvalds 36861da177e4SLinus Torvalds /* Reserve space for headers and set control bits. */ 36871da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 36881da177e4SLinus Torvalds /* Use a previous sequence. This should cause the other 36891da177e4SLinus Torvalds * end to send an ack. Don't queue or clone SKB, just 36901da177e4SLinus Torvalds * send it. 36911da177e4SLinus Torvalds */ 3692a3433f35SChangli Gao tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK); 3693e2e8009fSRenato Westphal NET_INC_STATS(sock_net(sk), mib); 36947450aaf6SEric Dumazet return tcp_transmit_skb(sk, skb, 0, (__force gfp_t)0); 36951da177e4SLinus Torvalds } 36961da177e4SLinus Torvalds 3697385e2070SEric Dumazet /* Called from setsockopt( ... TCP_REPAIR ) */ 3698ee995283SPavel Emelyanov void tcp_send_window_probe(struct sock *sk) 3699ee995283SPavel Emelyanov { 3700ee995283SPavel Emelyanov if (sk->sk_state == TCP_ESTABLISHED) { 3701ee995283SPavel Emelyanov tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1; 37029a568de4SEric Dumazet tcp_mstamp_refresh(tcp_sk(sk)); 3703e520af48SEric Dumazet tcp_xmit_probe_skb(sk, 0, LINUX_MIB_TCPWINPROBE); 3704ee995283SPavel Emelyanov } 3705ee995283SPavel Emelyanov } 3706ee995283SPavel Emelyanov 370767edfef7SAndi Kleen /* Initiate keepalive or window probe from timer. */ 3708e520af48SEric Dumazet int tcp_write_wakeup(struct sock *sk, int mib) 37091da177e4SLinus Torvalds { 37101da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 37111da177e4SLinus Torvalds struct sk_buff *skb; 37121da177e4SLinus Torvalds 3713058dc334SIlpo Järvinen if (sk->sk_state == TCP_CLOSE) 3714058dc334SIlpo Järvinen return -1; 3715058dc334SIlpo Järvinen 371600db4124SIan Morris skb = tcp_send_head(sk); 371700db4124SIan Morris if (skb && before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) { 37181da177e4SLinus Torvalds int err; 37190c54b85fSIlpo Järvinen unsigned int mss = tcp_current_mss(sk); 372090840defSIlpo Järvinen unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 37211da177e4SLinus Torvalds 37221da177e4SLinus Torvalds if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq)) 37231da177e4SLinus Torvalds tp->pushed_seq = TCP_SKB_CB(skb)->end_seq; 37241da177e4SLinus Torvalds 37251da177e4SLinus Torvalds /* We are probing the opening of a window 37261da177e4SLinus Torvalds * but the window size is != 0 37271da177e4SLinus Torvalds * must have been a result SWS avoidance ( sender ) 37281da177e4SLinus Torvalds */ 37291da177e4SLinus Torvalds if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq || 37301da177e4SLinus Torvalds skb->len > mss) { 37311da177e4SLinus Torvalds seg_size = min(seg_size, mss); 37324de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; 373375c119afSEric Dumazet if (tcp_fragment(sk, TCP_FRAG_IN_WRITE_QUEUE, 373475c119afSEric Dumazet skb, seg_size, mss, GFP_ATOMIC)) 37351da177e4SLinus Torvalds return -1; 37361da177e4SLinus Torvalds } else if (!tcp_skb_pcount(skb)) 37375bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss); 37381da177e4SLinus Torvalds 37394de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; 3740dfb4b9dcSDavid S. Miller err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 374166f5fe62SIlpo Järvinen if (!err) 374266f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, skb); 37431da177e4SLinus Torvalds return err; 37441da177e4SLinus Torvalds } else { 374533f5f57eSIlpo Järvinen if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF)) 3746e520af48SEric Dumazet tcp_xmit_probe_skb(sk, 1, mib); 3747e520af48SEric Dumazet return tcp_xmit_probe_skb(sk, 0, mib); 37481da177e4SLinus Torvalds } 37491da177e4SLinus Torvalds } 37501da177e4SLinus Torvalds 37511da177e4SLinus Torvalds /* A window probe timeout has occurred. If window is not closed send 37521da177e4SLinus Torvalds * a partial packet else a zero probe. 37531da177e4SLinus Torvalds */ 37541da177e4SLinus Torvalds void tcp_send_probe0(struct sock *sk) 37551da177e4SLinus Torvalds { 3756463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 37571da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 3758c6214a97SNikolay Borisov struct net *net = sock_net(sk); 3759c1d5674fSYuchung Cheng unsigned long timeout; 37601da177e4SLinus Torvalds int err; 37611da177e4SLinus Torvalds 3762e520af48SEric Dumazet err = tcp_write_wakeup(sk, LINUX_MIB_TCPWINPROBE); 37631da177e4SLinus Torvalds 376475c119afSEric Dumazet if (tp->packets_out || tcp_write_queue_empty(sk)) { 37651da177e4SLinus Torvalds /* Cancel probe timer, if it is not required. */ 37666687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out = 0; 3767463c84b9SArnaldo Carvalho de Melo icsk->icsk_backoff = 0; 37681da177e4SLinus Torvalds return; 37691da177e4SLinus Torvalds } 37701da177e4SLinus Torvalds 3771c1d5674fSYuchung Cheng icsk->icsk_probes_out++; 37721da177e4SLinus Torvalds if (err <= 0) { 3773c6214a97SNikolay Borisov if (icsk->icsk_backoff < net->ipv4.sysctl_tcp_retries2) 3774463c84b9SArnaldo Carvalho de Melo icsk->icsk_backoff++; 3775c1d5674fSYuchung Cheng timeout = tcp_probe0_when(sk, TCP_RTO_MAX); 37761da177e4SLinus Torvalds } else { 37771da177e4SLinus Torvalds /* If packet was not sent due to local congestion, 3778c1d5674fSYuchung Cheng * Let senders fight for local resources conservatively. 37791da177e4SLinus Torvalds */ 3780c1d5674fSYuchung Cheng timeout = TCP_RESOURCE_PROBE_INTERVAL; 37811da177e4SLinus Torvalds } 3782c1d5674fSYuchung Cheng tcp_reset_xmit_timer(sk, ICSK_TIME_PROBE0, timeout, TCP_RTO_MAX, NULL); 37831da177e4SLinus Torvalds } 37845db92c99SOctavian Purdila 3785ea3bea3aSEric Dumazet int tcp_rtx_synack(const struct sock *sk, struct request_sock *req) 37865db92c99SOctavian Purdila { 37875db92c99SOctavian Purdila const struct tcp_request_sock_ops *af_ops = tcp_rsk(req)->af_specific; 37885db92c99SOctavian Purdila struct flowi fl; 37895db92c99SOctavian Purdila int res; 37905db92c99SOctavian Purdila 379158d607d3SEric Dumazet tcp_rsk(req)->txhash = net_tx_rndhash(); 3792b3d05147SEric Dumazet res = af_ops->send_synack(sk, NULL, &fl, req, NULL, TCP_SYNACK_NORMAL); 37935db92c99SOctavian Purdila if (!res) { 379490bbcc60SEric Dumazet __TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS); 379502a1d6e7SEric Dumazet __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS); 37967e32b443SYuchung Cheng if (unlikely(tcp_passive_fastopen(sk))) 37977e32b443SYuchung Cheng tcp_sk(sk)->total_retrans++; 3798cf34ce3dSSong Liu trace_tcp_retransmit_synack(sk, req); 37995db92c99SOctavian Purdila } 38005db92c99SOctavian Purdila return res; 38015db92c99SOctavian Purdila } 38025db92c99SOctavian Purdila EXPORT_SYMBOL(tcp_rtx_synack); 3803