11da177e4SLinus Torvalds /* 21da177e4SLinus Torvalds * INET An implementation of the TCP/IP protocol suite for the LINUX 31da177e4SLinus Torvalds * operating system. INET is implemented using the BSD Socket 41da177e4SLinus Torvalds * interface as the means of communication with the user level. 51da177e4SLinus Torvalds * 61da177e4SLinus Torvalds * Implementation of the Transmission Control Protocol(TCP). 71da177e4SLinus Torvalds * 802c30a84SJesper Juhl * Authors: Ross Biro 91da177e4SLinus Torvalds * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG> 101da177e4SLinus Torvalds * Mark Evans, <evansmp@uhura.aston.ac.uk> 111da177e4SLinus Torvalds * Corey Minyard <wf-rch!minyard@relay.EU.net> 121da177e4SLinus Torvalds * Florian La Roche, <flla@stud.uni-sb.de> 131da177e4SLinus Torvalds * Charles Hedrick, <hedrick@klinzhai.rutgers.edu> 141da177e4SLinus Torvalds * Linus Torvalds, <torvalds@cs.helsinki.fi> 151da177e4SLinus Torvalds * Alan Cox, <gw4pts@gw4pts.ampr.org> 161da177e4SLinus Torvalds * Matthew Dillon, <dillon@apollo.west.oic.com> 171da177e4SLinus Torvalds * Arnt Gulbrandsen, <agulbra@nvg.unit.no> 181da177e4SLinus Torvalds * Jorge Cwik, <jorge@laser.satlink.net> 191da177e4SLinus Torvalds */ 201da177e4SLinus Torvalds 211da177e4SLinus Torvalds /* 221da177e4SLinus Torvalds * Changes: Pedro Roque : Retransmit queue handled by TCP. 231da177e4SLinus Torvalds * : Fragmentation on mtu decrease 241da177e4SLinus Torvalds * : Segment collapse on retransmit 251da177e4SLinus Torvalds * : AF independence 261da177e4SLinus Torvalds * 271da177e4SLinus Torvalds * Linus Torvalds : send_delayed_ack 281da177e4SLinus Torvalds * David S. Miller : Charge memory using the right skb 291da177e4SLinus Torvalds * during syn/ack processing. 301da177e4SLinus Torvalds * David S. Miller : Output engine completely rewritten. 311da177e4SLinus Torvalds * Andrea Arcangeli: SYNACK carry ts_recent in tsecr. 321da177e4SLinus Torvalds * Cacophonix Gaul : draft-minshall-nagle-01 331da177e4SLinus Torvalds * J Hadi Salim : ECN support 341da177e4SLinus Torvalds * 351da177e4SLinus Torvalds */ 361da177e4SLinus Torvalds 3791df42beSJoe Perches #define pr_fmt(fmt) "TCP: " fmt 3891df42beSJoe Perches 391da177e4SLinus Torvalds #include <net/tcp.h> 401da177e4SLinus Torvalds 411da177e4SLinus Torvalds #include <linux/compiler.h> 425a0e3ad6STejun Heo #include <linux/gfp.h> 431da177e4SLinus Torvalds #include <linux/module.h> 4460e2a778SUrsula Braun #include <linux/static_key.h> 451da177e4SLinus Torvalds 46e086101bSCong Wang #include <trace/events/tcp.h> 4735089bb2SDavid S. Miller 489799ccb0SEric Dumazet /* Refresh clocks of a TCP socket, 499799ccb0SEric Dumazet * ensuring monotically increasing values. 509799ccb0SEric Dumazet */ 519799ccb0SEric Dumazet void tcp_mstamp_refresh(struct tcp_sock *tp) 529799ccb0SEric Dumazet { 539799ccb0SEric Dumazet u64 val = tcp_clock_ns(); 549799ccb0SEric Dumazet 555f6188a8SEric Dumazet if (val > tp->tcp_clock_cache) 565f6188a8SEric Dumazet tp->tcp_clock_cache = val; 579799ccb0SEric Dumazet 589799ccb0SEric Dumazet val = div_u64(val, NSEC_PER_USEC); 599799ccb0SEric Dumazet if (val > tp->tcp_mstamp) 609799ccb0SEric Dumazet tp->tcp_mstamp = val; 619799ccb0SEric Dumazet } 629799ccb0SEric Dumazet 6346d3ceabSEric Dumazet static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle, 6446d3ceabSEric Dumazet int push_one, gfp_t gfp); 65519855c5SWilliam Allen Simpson 6667edfef7SAndi Kleen /* Account for new data that has been sent to the network. */ 6775c119afSEric Dumazet static void tcp_event_new_data_sent(struct sock *sk, struct sk_buff *skb) 686ff03ac3SIlpo Järvinen { 696ba8a3b1SNandita Dukkipati struct inet_connection_sock *icsk = inet_csk(sk); 706ff03ac3SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 7166f5fe62SIlpo Järvinen unsigned int prior_packets = tp->packets_out; 729e412ba7SIlpo Järvinen 731da177e4SLinus Torvalds tp->snd_nxt = TCP_SKB_CB(skb)->end_seq; 748512430eSIlpo Järvinen 7575c119afSEric Dumazet __skb_unlink(skb, &sk->sk_write_queue); 7675c119afSEric Dumazet tcp_rbtree_insert(&sk->tcp_rtx_queue, skb); 7775c119afSEric Dumazet 7866f5fe62SIlpo Järvinen tp->packets_out += tcp_skb_pcount(skb); 79bec41a11SYuchung Cheng if (!prior_packets || icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) 80750ea2baSYuchung Cheng tcp_rearm_rto(sk); 81f19c29e3SYuchung Cheng 82f7324acdSDavid S. Miller NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT, 83f19c29e3SYuchung Cheng tcp_skb_pcount(skb)); 846a5dc9e5SEric Dumazet } 851da177e4SLinus Torvalds 86a4ecb15aSCui, Cheng /* SND.NXT, if window was not shrunk or the amount of shrunk was less than one 87a4ecb15aSCui, Cheng * window scaling factor due to loss of precision. 881da177e4SLinus Torvalds * If window has been shrunk, what should we make? It is not clear at all. 891da177e4SLinus Torvalds * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-( 901da177e4SLinus Torvalds * Anything in between SND.UNA...SND.UNA+SND.WND also can be already 911da177e4SLinus Torvalds * invalid. OK, let's make this for now: 921da177e4SLinus Torvalds */ 93cf533ea5SEric Dumazet static inline __u32 tcp_acceptable_seq(const struct sock *sk) 941da177e4SLinus Torvalds { 95cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 969e412ba7SIlpo Järvinen 97a4ecb15aSCui, Cheng if (!before(tcp_wnd_end(tp), tp->snd_nxt) || 98a4ecb15aSCui, Cheng (tp->rx_opt.wscale_ok && 99a4ecb15aSCui, Cheng ((tp->snd_nxt - tcp_wnd_end(tp)) < (1 << tp->rx_opt.rcv_wscale)))) 1001da177e4SLinus Torvalds return tp->snd_nxt; 1011da177e4SLinus Torvalds else 10290840defSIlpo Järvinen return tcp_wnd_end(tp); 1031da177e4SLinus Torvalds } 1041da177e4SLinus Torvalds 1051da177e4SLinus Torvalds /* Calculate mss to advertise in SYN segment. 1061da177e4SLinus Torvalds * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that: 1071da177e4SLinus Torvalds * 1081da177e4SLinus Torvalds * 1. It is independent of path mtu. 1091da177e4SLinus Torvalds * 2. Ideally, it is maximal possible segment size i.e. 65535-40. 1101da177e4SLinus Torvalds * 3. For IPv4 it is reasonable to calculate it from maximal MTU of 1111da177e4SLinus Torvalds * attached devices, because some buggy hosts are confused by 1121da177e4SLinus Torvalds * large MSS. 1131da177e4SLinus Torvalds * 4. We do not make 3, we advertise MSS, calculated from first 1141da177e4SLinus Torvalds * hop device mtu, but allow to raise it to ip_rt_min_advmss. 1151da177e4SLinus Torvalds * This may be overridden via information stored in routing table. 1161da177e4SLinus Torvalds * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible, 1171da177e4SLinus Torvalds * probably even Jumbo". 1181da177e4SLinus Torvalds */ 1191da177e4SLinus Torvalds static __u16 tcp_advertise_mss(struct sock *sk) 1201da177e4SLinus Torvalds { 1211da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 122cf533ea5SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 1231da177e4SLinus Torvalds int mss = tp->advmss; 1241da177e4SLinus Torvalds 1250dbaee3bSDavid S. Miller if (dst) { 1260dbaee3bSDavid S. Miller unsigned int metric = dst_metric_advmss(dst); 1270dbaee3bSDavid S. Miller 1280dbaee3bSDavid S. Miller if (metric < mss) { 1290dbaee3bSDavid S. Miller mss = metric; 1301da177e4SLinus Torvalds tp->advmss = mss; 1311da177e4SLinus Torvalds } 1320dbaee3bSDavid S. Miller } 1331da177e4SLinus Torvalds 1341da177e4SLinus Torvalds return (__u16)mss; 1351da177e4SLinus Torvalds } 1361da177e4SLinus Torvalds 1371da177e4SLinus Torvalds /* RFC2861. Reset CWND after idle period longer RTO to "restart window". 1386f021c62SEric Dumazet * This is the first part of cwnd validation mechanism. 1396f021c62SEric Dumazet */ 1406f021c62SEric Dumazet void tcp_cwnd_restart(struct sock *sk, s32 delta) 1411da177e4SLinus Torvalds { 142463c84b9SArnaldo Carvalho de Melo struct tcp_sock *tp = tcp_sk(sk); 1436f021c62SEric Dumazet u32 restart_cwnd = tcp_init_cwnd(tp, __sk_dst_get(sk)); 1441da177e4SLinus Torvalds u32 cwnd = tp->snd_cwnd; 1451da177e4SLinus Torvalds 1466687e988SArnaldo Carvalho de Melo tcp_ca_event(sk, CA_EVENT_CWND_RESTART); 1471da177e4SLinus Torvalds 1486687e988SArnaldo Carvalho de Melo tp->snd_ssthresh = tcp_current_ssthresh(sk); 1491da177e4SLinus Torvalds restart_cwnd = min(restart_cwnd, cwnd); 1501da177e4SLinus Torvalds 151463c84b9SArnaldo Carvalho de Melo while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd) 1521da177e4SLinus Torvalds cwnd >>= 1; 1531da177e4SLinus Torvalds tp->snd_cwnd = max(cwnd, restart_cwnd); 154c2203cf7SEric Dumazet tp->snd_cwnd_stamp = tcp_jiffies32; 1551da177e4SLinus Torvalds tp->snd_cwnd_used = 0; 1561da177e4SLinus Torvalds } 1571da177e4SLinus Torvalds 15867edfef7SAndi Kleen /* Congestion state accounting after a packet has been sent. */ 15940efc6faSStephen Hemminger static void tcp_event_data_sent(struct tcp_sock *tp, 160cf533ea5SEric Dumazet struct sock *sk) 1611da177e4SLinus Torvalds { 162463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 163d635fbe2SEric Dumazet const u32 now = tcp_jiffies32; 1641da177e4SLinus Torvalds 16505c5a46dSNeal Cardwell if (tcp_packets_in_flight(tp) == 0) 16605c5a46dSNeal Cardwell tcp_ca_event(sk, CA_EVENT_TX_START); 16705c5a46dSNeal Cardwell 1684a41f453SWei Wang /* If this is the first data packet sent in response to the 1694a41f453SWei Wang * previous received data, 1704a41f453SWei Wang * and it is a reply for ato after last received packet, 1714a41f453SWei Wang * increase pingpong count. 1721da177e4SLinus Torvalds */ 1734a41f453SWei Wang if (before(tp->lsndtime, icsk->icsk_ack.lrcvtime) && 1744a41f453SWei Wang (u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato) 1754a41f453SWei Wang inet_csk_inc_pingpong_cnt(sk); 1764a41f453SWei Wang 1774a41f453SWei Wang tp->lsndtime = now; 1781da177e4SLinus Torvalds } 1791da177e4SLinus Torvalds 18067edfef7SAndi Kleen /* Account for an ACK we sent. */ 18127cde44aSYuchung Cheng static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts, 18227cde44aSYuchung Cheng u32 rcv_nxt) 1831da177e4SLinus Torvalds { 1845d9f4262SEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 1855d9f4262SEric Dumazet 18686de5921SEric Dumazet if (unlikely(tp->compressed_ack > TCP_FASTRETRANS_THRESH)) { 187200d95f4SEric Dumazet NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPACKCOMPRESSED, 18886de5921SEric Dumazet tp->compressed_ack - TCP_FASTRETRANS_THRESH); 18986de5921SEric Dumazet tp->compressed_ack = TCP_FASTRETRANS_THRESH; 1905d9f4262SEric Dumazet if (hrtimer_try_to_cancel(&tp->compressed_ack_timer) == 1) 1915d9f4262SEric Dumazet __sock_put(sk); 1925d9f4262SEric Dumazet } 19327cde44aSYuchung Cheng 19427cde44aSYuchung Cheng if (unlikely(rcv_nxt != tp->rcv_nxt)) 19527cde44aSYuchung Cheng return; /* Special ACK sent by DCTCP to reflect ECN */ 196463c84b9SArnaldo Carvalho de Melo tcp_dec_quickack_mode(sk, pkts); 197463c84b9SArnaldo Carvalho de Melo inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK); 1981da177e4SLinus Torvalds } 1991da177e4SLinus Torvalds 2001da177e4SLinus Torvalds /* Determine a window scaling and initial window to offer. 2011da177e4SLinus Torvalds * Based on the assumption that the given amount of space 2021da177e4SLinus Torvalds * will be offered. Store the results in the tp structure. 2031da177e4SLinus Torvalds * NOTE: for smooth operation initial space offering should 2041da177e4SLinus Torvalds * be a multiple of mss if possible. We assume here that mss >= 1. 2051da177e4SLinus Torvalds * This MUST be enforced by all callers. 2061da177e4SLinus Torvalds */ 207ceef9ab6SEric Dumazet void tcp_select_initial_window(const struct sock *sk, int __space, __u32 mss, 2081da177e4SLinus Torvalds __u32 *rcv_wnd, __u32 *window_clamp, 20931d12926Slaurent chavey int wscale_ok, __u8 *rcv_wscale, 21031d12926Slaurent chavey __u32 init_rcv_wnd) 2111da177e4SLinus Torvalds { 2121da177e4SLinus Torvalds unsigned int space = (__space < 0 ? 0 : __space); 2131da177e4SLinus Torvalds 2141da177e4SLinus Torvalds /* If no clamp set the clamp to the max possible scaled window */ 2151da177e4SLinus Torvalds if (*window_clamp == 0) 216589c49cbSGao Feng (*window_clamp) = (U16_MAX << TCP_MAX_WSCALE); 2171da177e4SLinus Torvalds space = min(*window_clamp, space); 2181da177e4SLinus Torvalds 2191da177e4SLinus Torvalds /* Quantize space offering to a multiple of mss if possible. */ 2201da177e4SLinus Torvalds if (space > mss) 221589c49cbSGao Feng space = rounddown(space, mss); 2221da177e4SLinus Torvalds 2231da177e4SLinus Torvalds /* NOTE: offering an initial window larger than 32767 22415d99e02SRick Jones * will break some buggy TCP stacks. If the admin tells us 22515d99e02SRick Jones * it is likely we could be speaking with such a buggy stack 22615d99e02SRick Jones * we will truncate our initial window offering to 32K-1 22715d99e02SRick Jones * unless the remote has sent us a window scaling option, 22815d99e02SRick Jones * which we interpret as a sign the remote TCP is not 22915d99e02SRick Jones * misinterpreting the window field as a signed quantity. 2301da177e4SLinus Torvalds */ 231ceef9ab6SEric Dumazet if (sock_net(sk)->ipv4.sysctl_tcp_workaround_signed_windows) 2321da177e4SLinus Torvalds (*rcv_wnd) = min(space, MAX_TCP_WINDOW); 23315d99e02SRick Jones else 234a337531bSYuchung Cheng (*rcv_wnd) = min_t(u32, space, U16_MAX); 235a337531bSYuchung Cheng 236a337531bSYuchung Cheng if (init_rcv_wnd) 237a337531bSYuchung Cheng *rcv_wnd = min(*rcv_wnd, init_rcv_wnd * mss); 23815d99e02SRick Jones 23919bf6261SEric Dumazet *rcv_wscale = 0; 2401da177e4SLinus Torvalds if (wscale_ok) { 241589c49cbSGao Feng /* Set window scaling on max possible window */ 242356d1833SEric Dumazet space = max_t(u32, space, sock_net(sk)->ipv4.sysctl_tcp_rmem[2]); 243f626300aSSoheil Hassas Yeganeh space = max_t(u32, space, sysctl_rmem_max); 244316c1592SStephen Hemminger space = min_t(u32, space, *window_clamp); 24519bf6261SEric Dumazet *rcv_wscale = clamp_t(int, ilog2(space) - 15, 24619bf6261SEric Dumazet 0, TCP_MAX_WSCALE); 2471da177e4SLinus Torvalds } 2481da177e4SLinus Torvalds /* Set the clamp no higher than max representable value */ 249589c49cbSGao Feng (*window_clamp) = min_t(__u32, U16_MAX << (*rcv_wscale), *window_clamp); 2501da177e4SLinus Torvalds } 2514bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_select_initial_window); 2521da177e4SLinus Torvalds 2531da177e4SLinus Torvalds /* Chose a new window to advertise, update state in tcp_sock for the 2541da177e4SLinus Torvalds * socket, and return result with RFC1323 scaling applied. The return 2551da177e4SLinus Torvalds * value can be stuffed directly into th->window for an outgoing 2561da177e4SLinus Torvalds * frame. 2571da177e4SLinus Torvalds */ 25840efc6faSStephen Hemminger static u16 tcp_select_window(struct sock *sk) 2591da177e4SLinus Torvalds { 2601da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 2618e165e20SFlorian Westphal u32 old_win = tp->rcv_wnd; 2621da177e4SLinus Torvalds u32 cur_win = tcp_receive_window(tp); 2631da177e4SLinus Torvalds u32 new_win = __tcp_select_window(sk); 2641da177e4SLinus Torvalds 2651da177e4SLinus Torvalds /* Never shrink the offered window */ 2661da177e4SLinus Torvalds if (new_win < cur_win) { 2671da177e4SLinus Torvalds /* Danger Will Robinson! 2681da177e4SLinus Torvalds * Don't update rcv_wup/rcv_wnd here or else 2691da177e4SLinus Torvalds * we will not be able to advertise a zero 2701da177e4SLinus Torvalds * window in time. --DaveM 2711da177e4SLinus Torvalds * 2721da177e4SLinus Torvalds * Relax Will Robinson. 2731da177e4SLinus Torvalds */ 2748e165e20SFlorian Westphal if (new_win == 0) 2758e165e20SFlorian Westphal NET_INC_STATS(sock_net(sk), 2768e165e20SFlorian Westphal LINUX_MIB_TCPWANTZEROWINDOWADV); 277607bfbf2SPatrick McHardy new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale); 2781da177e4SLinus Torvalds } 2791da177e4SLinus Torvalds tp->rcv_wnd = new_win; 2801da177e4SLinus Torvalds tp->rcv_wup = tp->rcv_nxt; 2811da177e4SLinus Torvalds 2821da177e4SLinus Torvalds /* Make sure we do not exceed the maximum possible 2831da177e4SLinus Torvalds * scaled window. 2841da177e4SLinus Torvalds */ 285ceef9ab6SEric Dumazet if (!tp->rx_opt.rcv_wscale && 286ceef9ab6SEric Dumazet sock_net(sk)->ipv4.sysctl_tcp_workaround_signed_windows) 2871da177e4SLinus Torvalds new_win = min(new_win, MAX_TCP_WINDOW); 2881da177e4SLinus Torvalds else 2891da177e4SLinus Torvalds new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale)); 2901da177e4SLinus Torvalds 2911da177e4SLinus Torvalds /* RFC1323 scaling applied */ 2921da177e4SLinus Torvalds new_win >>= tp->rx_opt.rcv_wscale; 2931da177e4SLinus Torvalds 29431770e34SFlorian Westphal /* If we advertise zero window, disable fast path. */ 2958e165e20SFlorian Westphal if (new_win == 0) { 29631770e34SFlorian Westphal tp->pred_flags = 0; 2978e165e20SFlorian Westphal if (old_win) 2988e165e20SFlorian Westphal NET_INC_STATS(sock_net(sk), 2998e165e20SFlorian Westphal LINUX_MIB_TCPTOZEROWINDOWADV); 3008e165e20SFlorian Westphal } else if (old_win == 0) { 3018e165e20SFlorian Westphal NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFROMZEROWINDOWADV); 3028e165e20SFlorian Westphal } 3031da177e4SLinus Torvalds 3041da177e4SLinus Torvalds return new_win; 3051da177e4SLinus Torvalds } 3061da177e4SLinus Torvalds 30767edfef7SAndi Kleen /* Packet ECN state for a SYN-ACK */ 308735d3831SFlorian Westphal static void tcp_ecn_send_synack(struct sock *sk, struct sk_buff *skb) 309bdf1ee5dSIlpo Järvinen { 31030e502a3SDaniel Borkmann const struct tcp_sock *tp = tcp_sk(sk); 31130e502a3SDaniel Borkmann 3124de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_CWR; 313bdf1ee5dSIlpo Järvinen if (!(tp->ecn_flags & TCP_ECN_OK)) 3144de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_ECE; 31591b5b21cSLawrence Brakmo else if (tcp_ca_needs_ecn(sk) || 31691b5b21cSLawrence Brakmo tcp_bpf_ca_needs_ecn(sk)) 31730e502a3SDaniel Borkmann INET_ECN_xmit(sk); 318bdf1ee5dSIlpo Järvinen } 319bdf1ee5dSIlpo Järvinen 32067edfef7SAndi Kleen /* Packet ECN state for a SYN. */ 321735d3831SFlorian Westphal static void tcp_ecn_send_syn(struct sock *sk, struct sk_buff *skb) 322bdf1ee5dSIlpo Järvinen { 323bdf1ee5dSIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 32491b5b21cSLawrence Brakmo bool bpf_needs_ecn = tcp_bpf_ca_needs_ecn(sk); 325f7b3bec6SFlorian Westphal bool use_ecn = sock_net(sk)->ipv4.sysctl_tcp_ecn == 1 || 32691b5b21cSLawrence Brakmo tcp_ca_needs_ecn(sk) || bpf_needs_ecn; 327f7b3bec6SFlorian Westphal 328f7b3bec6SFlorian Westphal if (!use_ecn) { 329f7b3bec6SFlorian Westphal const struct dst_entry *dst = __sk_dst_get(sk); 330f7b3bec6SFlorian Westphal 331f7b3bec6SFlorian Westphal if (dst && dst_feature(dst, RTAX_FEATURE_ECN)) 332f7b3bec6SFlorian Westphal use_ecn = true; 333f7b3bec6SFlorian Westphal } 334bdf1ee5dSIlpo Järvinen 335bdf1ee5dSIlpo Järvinen tp->ecn_flags = 0; 336f7b3bec6SFlorian Westphal 337f7b3bec6SFlorian Westphal if (use_ecn) { 3384de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ECE | TCPHDR_CWR; 339bdf1ee5dSIlpo Järvinen tp->ecn_flags = TCP_ECN_OK; 34091b5b21cSLawrence Brakmo if (tcp_ca_needs_ecn(sk) || bpf_needs_ecn) 34130e502a3SDaniel Borkmann INET_ECN_xmit(sk); 342bdf1ee5dSIlpo Järvinen } 343bdf1ee5dSIlpo Järvinen } 344bdf1ee5dSIlpo Järvinen 34549213555SDaniel Borkmann static void tcp_ecn_clear_syn(struct sock *sk, struct sk_buff *skb) 34649213555SDaniel Borkmann { 34749213555SDaniel Borkmann if (sock_net(sk)->ipv4.sysctl_tcp_ecn_fallback) 34849213555SDaniel Borkmann /* tp->ecn_flags are cleared at a later point in time when 34949213555SDaniel Borkmann * SYN ACK is ultimatively being received. 35049213555SDaniel Borkmann */ 35149213555SDaniel Borkmann TCP_SKB_CB(skb)->tcp_flags &= ~(TCPHDR_ECE | TCPHDR_CWR); 35249213555SDaniel Borkmann } 35349213555SDaniel Borkmann 354735d3831SFlorian Westphal static void 3556ac705b1SEric Dumazet tcp_ecn_make_synack(const struct request_sock *req, struct tcphdr *th) 356bdf1ee5dSIlpo Järvinen { 3576ac705b1SEric Dumazet if (inet_rsk(req)->ecn_ok) 358bdf1ee5dSIlpo Järvinen th->ece = 1; 359bdf1ee5dSIlpo Järvinen } 360bdf1ee5dSIlpo Järvinen 36167edfef7SAndi Kleen /* Set up ECN state for a packet on a ESTABLISHED socket that is about to 36267edfef7SAndi Kleen * be sent. 36367edfef7SAndi Kleen */ 364735d3831SFlorian Westphal static void tcp_ecn_send(struct sock *sk, struct sk_buff *skb, 365ea1627c2SEric Dumazet struct tcphdr *th, int tcp_header_len) 366bdf1ee5dSIlpo Järvinen { 367bdf1ee5dSIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 368bdf1ee5dSIlpo Järvinen 369bdf1ee5dSIlpo Järvinen if (tp->ecn_flags & TCP_ECN_OK) { 370bdf1ee5dSIlpo Järvinen /* Not-retransmitted data segment: set ECT and inject CWR. */ 371bdf1ee5dSIlpo Järvinen if (skb->len != tcp_header_len && 372bdf1ee5dSIlpo Järvinen !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) { 373bdf1ee5dSIlpo Järvinen INET_ECN_xmit(sk); 374bdf1ee5dSIlpo Järvinen if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) { 375bdf1ee5dSIlpo Järvinen tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR; 376ea1627c2SEric Dumazet th->cwr = 1; 377bdf1ee5dSIlpo Järvinen skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN; 378bdf1ee5dSIlpo Järvinen } 37930e502a3SDaniel Borkmann } else if (!tcp_ca_needs_ecn(sk)) { 380bdf1ee5dSIlpo Järvinen /* ACK or retransmitted segment: clear ECT|CE */ 381bdf1ee5dSIlpo Järvinen INET_ECN_dontxmit(sk); 382bdf1ee5dSIlpo Järvinen } 383bdf1ee5dSIlpo Järvinen if (tp->ecn_flags & TCP_ECN_DEMAND_CWR) 384ea1627c2SEric Dumazet th->ece = 1; 385bdf1ee5dSIlpo Järvinen } 386bdf1ee5dSIlpo Järvinen } 387bdf1ee5dSIlpo Järvinen 388e870a8efSIlpo Järvinen /* Constructs common control bits of non-data skb. If SYN/FIN is present, 389e870a8efSIlpo Järvinen * auto increment end seqno. 390e870a8efSIlpo Järvinen */ 391e870a8efSIlpo Järvinen static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags) 392e870a8efSIlpo Järvinen { 3932e8e18efSDavid S. Miller skb->ip_summed = CHECKSUM_PARTIAL; 394e870a8efSIlpo Järvinen 3954de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags = flags; 396e870a8efSIlpo Järvinen TCP_SKB_CB(skb)->sacked = 0; 397e870a8efSIlpo Järvinen 398cd7d8498SEric Dumazet tcp_skb_pcount_set(skb, 1); 399e870a8efSIlpo Järvinen 400e870a8efSIlpo Järvinen TCP_SKB_CB(skb)->seq = seq; 401a3433f35SChangli Gao if (flags & (TCPHDR_SYN | TCPHDR_FIN)) 402e870a8efSIlpo Järvinen seq++; 403e870a8efSIlpo Järvinen TCP_SKB_CB(skb)->end_seq = seq; 404e870a8efSIlpo Järvinen } 405e870a8efSIlpo Järvinen 406a2a385d6SEric Dumazet static inline bool tcp_urg_mode(const struct tcp_sock *tp) 40733f5f57eSIlpo Järvinen { 40833f5f57eSIlpo Järvinen return tp->snd_una != tp->snd_up; 40933f5f57eSIlpo Järvinen } 41033f5f57eSIlpo Järvinen 41133ad798cSAdam Langley #define OPTION_SACK_ADVERTISE (1 << 0) 41233ad798cSAdam Langley #define OPTION_TS (1 << 1) 41333ad798cSAdam Langley #define OPTION_MD5 (1 << 2) 41489e95a61SOri Finkelman #define OPTION_WSCALE (1 << 3) 4152100c8d2SYuchung Cheng #define OPTION_FAST_OPEN_COOKIE (1 << 8) 41660e2a778SUrsula Braun #define OPTION_SMC (1 << 9) 41760e2a778SUrsula Braun 41860e2a778SUrsula Braun static void smc_options_write(__be32 *ptr, u16 *options) 41960e2a778SUrsula Braun { 42060e2a778SUrsula Braun #if IS_ENABLED(CONFIG_SMC) 42160e2a778SUrsula Braun if (static_branch_unlikely(&tcp_have_smc)) { 42260e2a778SUrsula Braun if (unlikely(OPTION_SMC & *options)) { 42360e2a778SUrsula Braun *ptr++ = htonl((TCPOPT_NOP << 24) | 42460e2a778SUrsula Braun (TCPOPT_NOP << 16) | 42560e2a778SUrsula Braun (TCPOPT_EXP << 8) | 42660e2a778SUrsula Braun (TCPOLEN_EXP_SMC_BASE)); 42760e2a778SUrsula Braun *ptr++ = htonl(TCPOPT_SMC_MAGIC); 42860e2a778SUrsula Braun } 42960e2a778SUrsula Braun } 43060e2a778SUrsula Braun #endif 43160e2a778SUrsula Braun } 43233ad798cSAdam Langley 43333ad798cSAdam Langley struct tcp_out_options { 4342100c8d2SYuchung Cheng u16 options; /* bit field of OPTION_* */ 4352100c8d2SYuchung Cheng u16 mss; /* 0 to disable */ 43633ad798cSAdam Langley u8 ws; /* window scale, 0 to disable */ 43733ad798cSAdam Langley u8 num_sack_blocks; /* number of SACK blocks to include */ 438bd0388aeSWilliam Allen Simpson u8 hash_size; /* bytes in hash_location */ 439bd0388aeSWilliam Allen Simpson __u8 *hash_location; /* temporary pointer, overloaded */ 4402100c8d2SYuchung Cheng __u32 tsval, tsecr; /* need to include OPTION_TS */ 4412100c8d2SYuchung Cheng struct tcp_fastopen_cookie *fastopen_cookie; /* Fast open cookie */ 44233ad798cSAdam Langley }; 44333ad798cSAdam Langley 44467edfef7SAndi Kleen /* Write previously computed TCP options to the packet. 44567edfef7SAndi Kleen * 44667edfef7SAndi Kleen * Beware: Something in the Internet is very sensitive to the ordering of 447fd6149d3SIlpo Järvinen * TCP options, we learned this through the hard way, so be careful here. 448fd6149d3SIlpo Järvinen * Luckily we can at least blame others for their non-compliance but from 4498e3bff96Sstephen hemminger * inter-operability perspective it seems that we're somewhat stuck with 450fd6149d3SIlpo Järvinen * the ordering which we have been using if we want to keep working with 451fd6149d3SIlpo Järvinen * those broken things (not that it currently hurts anybody as there isn't 452fd6149d3SIlpo Järvinen * particular reason why the ordering would need to be changed). 453fd6149d3SIlpo Järvinen * 454fd6149d3SIlpo Järvinen * At least SACK_PERM as the first option is known to lead to a disaster 455fd6149d3SIlpo Järvinen * (but it may well be that other scenarios fail similarly). 456fd6149d3SIlpo Järvinen */ 45733ad798cSAdam Langley static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp, 458bd0388aeSWilliam Allen Simpson struct tcp_out_options *opts) 459bd0388aeSWilliam Allen Simpson { 4602100c8d2SYuchung Cheng u16 options = opts->options; /* mungable copy */ 461bd0388aeSWilliam Allen Simpson 462bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_MD5 & options)) { 4631a2c6181SChristoph Paasch *ptr++ = htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) | 4641a2c6181SChristoph Paasch (TCPOPT_MD5SIG << 8) | TCPOLEN_MD5SIG); 465bd0388aeSWilliam Allen Simpson /* overload cookie hash location */ 466bd0388aeSWilliam Allen Simpson opts->hash_location = (__u8 *)ptr; 46733ad798cSAdam Langley ptr += 4; 46833ad798cSAdam Langley } 46933ad798cSAdam Langley 470fd6149d3SIlpo Järvinen if (unlikely(opts->mss)) { 471fd6149d3SIlpo Järvinen *ptr++ = htonl((TCPOPT_MSS << 24) | 472fd6149d3SIlpo Järvinen (TCPOLEN_MSS << 16) | 473fd6149d3SIlpo Järvinen opts->mss); 474fd6149d3SIlpo Järvinen } 475fd6149d3SIlpo Järvinen 476bd0388aeSWilliam Allen Simpson if (likely(OPTION_TS & options)) { 477bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_SACK_ADVERTISE & options)) { 47833ad798cSAdam Langley *ptr++ = htonl((TCPOPT_SACK_PERM << 24) | 47933ad798cSAdam Langley (TCPOLEN_SACK_PERM << 16) | 48033ad798cSAdam Langley (TCPOPT_TIMESTAMP << 8) | 48133ad798cSAdam Langley TCPOLEN_TIMESTAMP); 482bd0388aeSWilliam Allen Simpson options &= ~OPTION_SACK_ADVERTISE; 48333ad798cSAdam Langley } else { 484496c98dfSYOSHIFUJI Hideaki *ptr++ = htonl((TCPOPT_NOP << 24) | 48540efc6faSStephen Hemminger (TCPOPT_NOP << 16) | 48640efc6faSStephen Hemminger (TCPOPT_TIMESTAMP << 8) | 48740efc6faSStephen Hemminger TCPOLEN_TIMESTAMP); 48840efc6faSStephen Hemminger } 48933ad798cSAdam Langley *ptr++ = htonl(opts->tsval); 49033ad798cSAdam Langley *ptr++ = htonl(opts->tsecr); 49133ad798cSAdam Langley } 49233ad798cSAdam Langley 493bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_SACK_ADVERTISE & options)) { 49433ad798cSAdam Langley *ptr++ = htonl((TCPOPT_NOP << 24) | 49533ad798cSAdam Langley (TCPOPT_NOP << 16) | 49633ad798cSAdam Langley (TCPOPT_SACK_PERM << 8) | 49733ad798cSAdam Langley TCPOLEN_SACK_PERM); 49833ad798cSAdam Langley } 49933ad798cSAdam Langley 500bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_WSCALE & options)) { 50133ad798cSAdam Langley *ptr++ = htonl((TCPOPT_NOP << 24) | 50233ad798cSAdam Langley (TCPOPT_WINDOW << 16) | 50333ad798cSAdam Langley (TCPOLEN_WINDOW << 8) | 50433ad798cSAdam Langley opts->ws); 50533ad798cSAdam Langley } 50633ad798cSAdam Langley 50733ad798cSAdam Langley if (unlikely(opts->num_sack_blocks)) { 50833ad798cSAdam Langley struct tcp_sack_block *sp = tp->rx_opt.dsack ? 50933ad798cSAdam Langley tp->duplicate_sack : tp->selective_acks; 51040efc6faSStephen Hemminger int this_sack; 51140efc6faSStephen Hemminger 51240efc6faSStephen Hemminger *ptr++ = htonl((TCPOPT_NOP << 24) | 51340efc6faSStephen Hemminger (TCPOPT_NOP << 16) | 51440efc6faSStephen Hemminger (TCPOPT_SACK << 8) | 51533ad798cSAdam Langley (TCPOLEN_SACK_BASE + (opts->num_sack_blocks * 51640efc6faSStephen Hemminger TCPOLEN_SACK_PERBLOCK))); 5172de979bdSStephen Hemminger 51833ad798cSAdam Langley for (this_sack = 0; this_sack < opts->num_sack_blocks; 51933ad798cSAdam Langley ++this_sack) { 52040efc6faSStephen Hemminger *ptr++ = htonl(sp[this_sack].start_seq); 52140efc6faSStephen Hemminger *ptr++ = htonl(sp[this_sack].end_seq); 52240efc6faSStephen Hemminger } 5232de979bdSStephen Hemminger 52440efc6faSStephen Hemminger tp->rx_opt.dsack = 0; 52540efc6faSStephen Hemminger } 5262100c8d2SYuchung Cheng 5272100c8d2SYuchung Cheng if (unlikely(OPTION_FAST_OPEN_COOKIE & options)) { 5282100c8d2SYuchung Cheng struct tcp_fastopen_cookie *foc = opts->fastopen_cookie; 5297f9b838bSDaniel Lee u8 *p = (u8 *)ptr; 5307f9b838bSDaniel Lee u32 len; /* Fast Open option length */ 5312100c8d2SYuchung Cheng 5327f9b838bSDaniel Lee if (foc->exp) { 5337f9b838bSDaniel Lee len = TCPOLEN_EXP_FASTOPEN_BASE + foc->len; 5347f9b838bSDaniel Lee *ptr = htonl((TCPOPT_EXP << 24) | (len << 16) | 5352100c8d2SYuchung Cheng TCPOPT_FASTOPEN_MAGIC); 5367f9b838bSDaniel Lee p += TCPOLEN_EXP_FASTOPEN_BASE; 5377f9b838bSDaniel Lee } else { 5387f9b838bSDaniel Lee len = TCPOLEN_FASTOPEN_BASE + foc->len; 5397f9b838bSDaniel Lee *p++ = TCPOPT_FASTOPEN; 5407f9b838bSDaniel Lee *p++ = len; 5412100c8d2SYuchung Cheng } 5427f9b838bSDaniel Lee 5437f9b838bSDaniel Lee memcpy(p, foc->val, foc->len); 5447f9b838bSDaniel Lee if ((len & 3) == 2) { 5457f9b838bSDaniel Lee p[foc->len] = TCPOPT_NOP; 5467f9b838bSDaniel Lee p[foc->len + 1] = TCPOPT_NOP; 5477f9b838bSDaniel Lee } 5487f9b838bSDaniel Lee ptr += (len + 3) >> 2; 5492100c8d2SYuchung Cheng } 55060e2a778SUrsula Braun 55160e2a778SUrsula Braun smc_options_write(ptr, &options); 55260e2a778SUrsula Braun } 55360e2a778SUrsula Braun 55460e2a778SUrsula Braun static void smc_set_option(const struct tcp_sock *tp, 55560e2a778SUrsula Braun struct tcp_out_options *opts, 55660e2a778SUrsula Braun unsigned int *remaining) 55760e2a778SUrsula Braun { 55860e2a778SUrsula Braun #if IS_ENABLED(CONFIG_SMC) 55960e2a778SUrsula Braun if (static_branch_unlikely(&tcp_have_smc)) { 56060e2a778SUrsula Braun if (tp->syn_smc) { 56160e2a778SUrsula Braun if (*remaining >= TCPOLEN_EXP_SMC_BASE_ALIGNED) { 56260e2a778SUrsula Braun opts->options |= OPTION_SMC; 56360e2a778SUrsula Braun *remaining -= TCPOLEN_EXP_SMC_BASE_ALIGNED; 56460e2a778SUrsula Braun } 56560e2a778SUrsula Braun } 56660e2a778SUrsula Braun } 56760e2a778SUrsula Braun #endif 56860e2a778SUrsula Braun } 56960e2a778SUrsula Braun 57060e2a778SUrsula Braun static void smc_set_option_cond(const struct tcp_sock *tp, 57160e2a778SUrsula Braun const struct inet_request_sock *ireq, 57260e2a778SUrsula Braun struct tcp_out_options *opts, 57360e2a778SUrsula Braun unsigned int *remaining) 57460e2a778SUrsula Braun { 57560e2a778SUrsula Braun #if IS_ENABLED(CONFIG_SMC) 57660e2a778SUrsula Braun if (static_branch_unlikely(&tcp_have_smc)) { 57760e2a778SUrsula Braun if (tp->syn_smc && ireq->smc_ok) { 57860e2a778SUrsula Braun if (*remaining >= TCPOLEN_EXP_SMC_BASE_ALIGNED) { 57960e2a778SUrsula Braun opts->options |= OPTION_SMC; 58060e2a778SUrsula Braun *remaining -= TCPOLEN_EXP_SMC_BASE_ALIGNED; 58160e2a778SUrsula Braun } 58260e2a778SUrsula Braun } 58360e2a778SUrsula Braun } 58460e2a778SUrsula Braun #endif 58540efc6faSStephen Hemminger } 58640efc6faSStephen Hemminger 58767edfef7SAndi Kleen /* Compute TCP options for SYN packets. This is not the final 58867edfef7SAndi Kleen * network wire format yet. 58967edfef7SAndi Kleen */ 59095c96174SEric Dumazet static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb, 59133ad798cSAdam Langley struct tcp_out_options *opts, 592cf533ea5SEric Dumazet struct tcp_md5sig_key **md5) 593cf533ea5SEric Dumazet { 59433ad798cSAdam Langley struct tcp_sock *tp = tcp_sk(sk); 59595c96174SEric Dumazet unsigned int remaining = MAX_TCP_OPTION_SPACE; 596783237e8SYuchung Cheng struct tcp_fastopen_request *fastopen = tp->fastopen_req; 59733ad798cSAdam Langley 5988c2320e8SEric Dumazet *md5 = NULL; 599cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 600921f9a0fSEric Dumazet if (static_branch_unlikely(&tcp_md5_needed) && 6016015c71eSEric Dumazet rcu_access_pointer(tp->md5sig_info)) { 60233ad798cSAdam Langley *md5 = tp->af_specific->md5_lookup(sk, sk); 60333ad798cSAdam Langley if (*md5) { 60433ad798cSAdam Langley opts->options |= OPTION_MD5; 605bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_MD5SIG_ALIGNED; 606cfb6eeb4SYOSHIFUJI Hideaki } 6078c2320e8SEric Dumazet } 608cfb6eeb4SYOSHIFUJI Hideaki #endif 60933ad798cSAdam Langley 61033ad798cSAdam Langley /* We always get an MSS option. The option bytes which will be seen in 61133ad798cSAdam Langley * normal data packets should timestamps be used, must be in the MSS 61233ad798cSAdam Langley * advertised. But we subtract them from tp->mss_cache so that 61333ad798cSAdam Langley * calculations in tcp_sendmsg are simpler etc. So account for this 61433ad798cSAdam Langley * fact here if necessary. If we don't do this correctly, as a 61533ad798cSAdam Langley * receiver we won't recognize data packets as being full sized when we 61633ad798cSAdam Langley * should, and thus we won't abide by the delayed ACK rules correctly. 61733ad798cSAdam Langley * SACKs don't matter, we never delay an ACK when we have any of those 61833ad798cSAdam Langley * going out. */ 61933ad798cSAdam Langley opts->mss = tcp_advertise_mss(sk); 620bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_MSS_ALIGNED; 62133ad798cSAdam Langley 6225d2ed052SEric Dumazet if (likely(sock_net(sk)->ipv4.sysctl_tcp_timestamps && !*md5)) { 62333ad798cSAdam Langley opts->options |= OPTION_TS; 6247faee5c0SEric Dumazet opts->tsval = tcp_skb_timestamp(skb) + tp->tsoffset; 62533ad798cSAdam Langley opts->tsecr = tp->rx_opt.ts_recent; 626bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_TSTAMP_ALIGNED; 62733ad798cSAdam Langley } 6289bb37ef0SEric Dumazet if (likely(sock_net(sk)->ipv4.sysctl_tcp_window_scaling)) { 62933ad798cSAdam Langley opts->ws = tp->rx_opt.rcv_wscale; 63089e95a61SOri Finkelman opts->options |= OPTION_WSCALE; 631bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_WSCALE_ALIGNED; 63233ad798cSAdam Langley } 633f9301034SEric Dumazet if (likely(sock_net(sk)->ipv4.sysctl_tcp_sack)) { 63433ad798cSAdam Langley opts->options |= OPTION_SACK_ADVERTISE; 635b32d1310SDavid S. Miller if (unlikely(!(OPTION_TS & opts->options))) 636bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_SACKPERM_ALIGNED; 63733ad798cSAdam Langley } 63833ad798cSAdam Langley 639783237e8SYuchung Cheng if (fastopen && fastopen->cookie.len >= 0) { 6402646c831SDaniel Lee u32 need = fastopen->cookie.len; 6412646c831SDaniel Lee 6422646c831SDaniel Lee need += fastopen->cookie.exp ? TCPOLEN_EXP_FASTOPEN_BASE : 6432646c831SDaniel Lee TCPOLEN_FASTOPEN_BASE; 644783237e8SYuchung Cheng need = (need + 3) & ~3U; /* Align to 32 bits */ 645783237e8SYuchung Cheng if (remaining >= need) { 646783237e8SYuchung Cheng opts->options |= OPTION_FAST_OPEN_COOKIE; 647783237e8SYuchung Cheng opts->fastopen_cookie = &fastopen->cookie; 648783237e8SYuchung Cheng remaining -= need; 649783237e8SYuchung Cheng tp->syn_fastopen = 1; 6502646c831SDaniel Lee tp->syn_fastopen_exp = fastopen->cookie.exp ? 1 : 0; 651783237e8SYuchung Cheng } 652783237e8SYuchung Cheng } 653bd0388aeSWilliam Allen Simpson 65460e2a778SUrsula Braun smc_set_option(tp, opts, &remaining); 65560e2a778SUrsula Braun 656bd0388aeSWilliam Allen Simpson return MAX_TCP_OPTION_SPACE - remaining; 65733ad798cSAdam Langley } 65833ad798cSAdam Langley 65967edfef7SAndi Kleen /* Set up TCP options for SYN-ACKs. */ 66060e2a778SUrsula Braun static unsigned int tcp_synack_options(const struct sock *sk, 66160e2a778SUrsula Braun struct request_sock *req, 66295c96174SEric Dumazet unsigned int mss, struct sk_buff *skb, 66333ad798cSAdam Langley struct tcp_out_options *opts, 66480f03e27SEric Dumazet const struct tcp_md5sig_key *md5, 6658336886fSJerry Chu struct tcp_fastopen_cookie *foc) 6664957faadSWilliam Allen Simpson { 66733ad798cSAdam Langley struct inet_request_sock *ireq = inet_rsk(req); 66895c96174SEric Dumazet unsigned int remaining = MAX_TCP_OPTION_SPACE; 66933ad798cSAdam Langley 67033ad798cSAdam Langley #ifdef CONFIG_TCP_MD5SIG 67180f03e27SEric Dumazet if (md5) { 67233ad798cSAdam Langley opts->options |= OPTION_MD5; 6734957faadSWilliam Allen Simpson remaining -= TCPOLEN_MD5SIG_ALIGNED; 6744957faadSWilliam Allen Simpson 6754957faadSWilliam Allen Simpson /* We can't fit any SACK blocks in a packet with MD5 + TS 6764957faadSWilliam Allen Simpson * options. There was discussion about disabling SACK 6774957faadSWilliam Allen Simpson * rather than TS in order to fit in better with old, 6784957faadSWilliam Allen Simpson * buggy kernels, but that was deemed to be unnecessary. 6794957faadSWilliam Allen Simpson */ 680de213e5eSEric Dumazet ireq->tstamp_ok &= !ireq->sack_ok; 68133ad798cSAdam Langley } 68233ad798cSAdam Langley #endif 68333ad798cSAdam Langley 6844957faadSWilliam Allen Simpson /* We always send an MSS option. */ 68533ad798cSAdam Langley opts->mss = mss; 6864957faadSWilliam Allen Simpson remaining -= TCPOLEN_MSS_ALIGNED; 68733ad798cSAdam Langley 68833ad798cSAdam Langley if (likely(ireq->wscale_ok)) { 68933ad798cSAdam Langley opts->ws = ireq->rcv_wscale; 69089e95a61SOri Finkelman opts->options |= OPTION_WSCALE; 6914957faadSWilliam Allen Simpson remaining -= TCPOLEN_WSCALE_ALIGNED; 69233ad798cSAdam Langley } 693de213e5eSEric Dumazet if (likely(ireq->tstamp_ok)) { 69433ad798cSAdam Langley opts->options |= OPTION_TS; 69595a22caeSFlorian Westphal opts->tsval = tcp_skb_timestamp(skb) + tcp_rsk(req)->ts_off; 69633ad798cSAdam Langley opts->tsecr = req->ts_recent; 6974957faadSWilliam Allen Simpson remaining -= TCPOLEN_TSTAMP_ALIGNED; 69833ad798cSAdam Langley } 69933ad798cSAdam Langley if (likely(ireq->sack_ok)) { 70033ad798cSAdam Langley opts->options |= OPTION_SACK_ADVERTISE; 701de213e5eSEric Dumazet if (unlikely(!ireq->tstamp_ok)) 7024957faadSWilliam Allen Simpson remaining -= TCPOLEN_SACKPERM_ALIGNED; 70333ad798cSAdam Langley } 7047f9b838bSDaniel Lee if (foc != NULL && foc->len >= 0) { 7057f9b838bSDaniel Lee u32 need = foc->len; 7067f9b838bSDaniel Lee 7077f9b838bSDaniel Lee need += foc->exp ? TCPOLEN_EXP_FASTOPEN_BASE : 7087f9b838bSDaniel Lee TCPOLEN_FASTOPEN_BASE; 7098336886fSJerry Chu need = (need + 3) & ~3U; /* Align to 32 bits */ 7108336886fSJerry Chu if (remaining >= need) { 7118336886fSJerry Chu opts->options |= OPTION_FAST_OPEN_COOKIE; 7128336886fSJerry Chu opts->fastopen_cookie = foc; 7138336886fSJerry Chu remaining -= need; 7148336886fSJerry Chu } 7158336886fSJerry Chu } 7164957faadSWilliam Allen Simpson 71760e2a778SUrsula Braun smc_set_option_cond(tcp_sk(sk), ireq, opts, &remaining); 71860e2a778SUrsula Braun 7194957faadSWilliam Allen Simpson return MAX_TCP_OPTION_SPACE - remaining; 72033ad798cSAdam Langley } 72133ad798cSAdam Langley 72267edfef7SAndi Kleen /* Compute TCP options for ESTABLISHED sockets. This is not the 72367edfef7SAndi Kleen * final wire format yet. 72467edfef7SAndi Kleen */ 72595c96174SEric Dumazet static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb, 72633ad798cSAdam Langley struct tcp_out_options *opts, 727cf533ea5SEric Dumazet struct tcp_md5sig_key **md5) 728cf533ea5SEric Dumazet { 72933ad798cSAdam Langley struct tcp_sock *tp = tcp_sk(sk); 73095c96174SEric Dumazet unsigned int size = 0; 731cabeccbdSIlpo Järvinen unsigned int eff_sacks; 73233ad798cSAdam Langley 7335843ef42SAndi Kleen opts->options = 0; 7345843ef42SAndi Kleen 7358c2320e8SEric Dumazet *md5 = NULL; 73633ad798cSAdam Langley #ifdef CONFIG_TCP_MD5SIG 737921f9a0fSEric Dumazet if (static_branch_unlikely(&tcp_md5_needed) && 7386015c71eSEric Dumazet rcu_access_pointer(tp->md5sig_info)) { 73933ad798cSAdam Langley *md5 = tp->af_specific->md5_lookup(sk, sk); 7408c2320e8SEric Dumazet if (*md5) { 74133ad798cSAdam Langley opts->options |= OPTION_MD5; 74233ad798cSAdam Langley size += TCPOLEN_MD5SIG_ALIGNED; 74333ad798cSAdam Langley } 7448c2320e8SEric Dumazet } 74533ad798cSAdam Langley #endif 74633ad798cSAdam Langley 74733ad798cSAdam Langley if (likely(tp->rx_opt.tstamp_ok)) { 74833ad798cSAdam Langley opts->options |= OPTION_TS; 7497faee5c0SEric Dumazet opts->tsval = skb ? tcp_skb_timestamp(skb) + tp->tsoffset : 0; 75033ad798cSAdam Langley opts->tsecr = tp->rx_opt.ts_recent; 75133ad798cSAdam Langley size += TCPOLEN_TSTAMP_ALIGNED; 75233ad798cSAdam Langley } 75333ad798cSAdam Langley 754cabeccbdSIlpo Järvinen eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack; 755cabeccbdSIlpo Järvinen if (unlikely(eff_sacks)) { 75695c96174SEric Dumazet const unsigned int remaining = MAX_TCP_OPTION_SPACE - size; 75733ad798cSAdam Langley opts->num_sack_blocks = 75895c96174SEric Dumazet min_t(unsigned int, eff_sacks, 75933ad798cSAdam Langley (remaining - TCPOLEN_SACK_BASE_ALIGNED) / 76033ad798cSAdam Langley TCPOLEN_SACK_PERBLOCK); 76133ad798cSAdam Langley size += TCPOLEN_SACK_BASE_ALIGNED + 76233ad798cSAdam Langley opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK; 76333ad798cSAdam Langley } 76433ad798cSAdam Langley 76533ad798cSAdam Langley return size; 76640efc6faSStephen Hemminger } 7671da177e4SLinus Torvalds 76846d3ceabSEric Dumazet 76946d3ceabSEric Dumazet /* TCP SMALL QUEUES (TSQ) 77046d3ceabSEric Dumazet * 77146d3ceabSEric Dumazet * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev) 77246d3ceabSEric Dumazet * to reduce RTT and bufferbloat. 77346d3ceabSEric Dumazet * We do this using a special skb destructor (tcp_wfree). 77446d3ceabSEric Dumazet * 77546d3ceabSEric Dumazet * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb 77646d3ceabSEric Dumazet * needs to be reallocated in a driver. 7778e3bff96Sstephen hemminger * The invariant being skb->truesize subtracted from sk->sk_wmem_alloc 77846d3ceabSEric Dumazet * 77946d3ceabSEric Dumazet * Since transmit from skb destructor is forbidden, we use a tasklet 78046d3ceabSEric Dumazet * to process all sockets that eventually need to send more skbs. 78146d3ceabSEric Dumazet * We use one tasklet per cpu, with its own queue of sockets. 78246d3ceabSEric Dumazet */ 78346d3ceabSEric Dumazet struct tsq_tasklet { 78446d3ceabSEric Dumazet struct tasklet_struct tasklet; 78546d3ceabSEric Dumazet struct list_head head; /* queue of tcp sockets */ 78646d3ceabSEric Dumazet }; 78746d3ceabSEric Dumazet static DEFINE_PER_CPU(struct tsq_tasklet, tsq_tasklet); 78846d3ceabSEric Dumazet 78973a6bab5SEric Dumazet static void tcp_tsq_write(struct sock *sk) 7906f458dfbSEric Dumazet { 7916f458dfbSEric Dumazet if ((1 << sk->sk_state) & 7926f458dfbSEric Dumazet (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_CLOSING | 793f9616c35SEric Dumazet TCPF_CLOSE_WAIT | TCPF_LAST_ACK)) { 794f9616c35SEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 795f9616c35SEric Dumazet 796f9616c35SEric Dumazet if (tp->lost_out > tp->retrans_out && 7973a91d29fSKoichiro Den tp->snd_cwnd > tcp_packets_in_flight(tp)) { 7983a91d29fSKoichiro Den tcp_mstamp_refresh(tp); 799f9616c35SEric Dumazet tcp_xmit_retransmit_queue(sk); 8003a91d29fSKoichiro Den } 801f9616c35SEric Dumazet 802f9616c35SEric Dumazet tcp_write_xmit(sk, tcp_current_mss(sk), tp->nonagle, 803bf06200eSJohn Ogness 0, GFP_ATOMIC); 8046f458dfbSEric Dumazet } 805f9616c35SEric Dumazet } 80673a6bab5SEric Dumazet 80773a6bab5SEric Dumazet static void tcp_tsq_handler(struct sock *sk) 80873a6bab5SEric Dumazet { 80973a6bab5SEric Dumazet bh_lock_sock(sk); 81073a6bab5SEric Dumazet if (!sock_owned_by_user(sk)) 81173a6bab5SEric Dumazet tcp_tsq_write(sk); 81273a6bab5SEric Dumazet else if (!test_and_set_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags)) 81373a6bab5SEric Dumazet sock_hold(sk); 81473a6bab5SEric Dumazet bh_unlock_sock(sk); 81573a6bab5SEric Dumazet } 81646d3ceabSEric Dumazet /* 8178e3bff96Sstephen hemminger * One tasklet per cpu tries to send more skbs. 81846d3ceabSEric Dumazet * We run in tasklet context but need to disable irqs when 8198e3bff96Sstephen hemminger * transferring tsq->head because tcp_wfree() might 82046d3ceabSEric Dumazet * interrupt us (non NAPI drivers) 82146d3ceabSEric Dumazet */ 82246d3ceabSEric Dumazet static void tcp_tasklet_func(unsigned long data) 82346d3ceabSEric Dumazet { 82446d3ceabSEric Dumazet struct tsq_tasklet *tsq = (struct tsq_tasklet *)data; 82546d3ceabSEric Dumazet LIST_HEAD(list); 82646d3ceabSEric Dumazet unsigned long flags; 82746d3ceabSEric Dumazet struct list_head *q, *n; 82846d3ceabSEric Dumazet struct tcp_sock *tp; 82946d3ceabSEric Dumazet struct sock *sk; 83046d3ceabSEric Dumazet 83146d3ceabSEric Dumazet local_irq_save(flags); 83246d3ceabSEric Dumazet list_splice_init(&tsq->head, &list); 83346d3ceabSEric Dumazet local_irq_restore(flags); 83446d3ceabSEric Dumazet 83546d3ceabSEric Dumazet list_for_each_safe(q, n, &list) { 83646d3ceabSEric Dumazet tp = list_entry(q, struct tcp_sock, tsq_node); 83746d3ceabSEric Dumazet list_del(&tp->tsq_node); 83846d3ceabSEric Dumazet 83946d3ceabSEric Dumazet sk = (struct sock *)tp; 8400a9648f1SEric Dumazet smp_mb__before_atomic(); 8417aa5470cSEric Dumazet clear_bit(TSQ_QUEUED, &sk->sk_tsq_flags); 8427aa5470cSEric Dumazet 8436f458dfbSEric Dumazet tcp_tsq_handler(sk); 84446d3ceabSEric Dumazet sk_free(sk); 84546d3ceabSEric Dumazet } 84646d3ceabSEric Dumazet } 84746d3ceabSEric Dumazet 84840fc3423SEric Dumazet #define TCP_DEFERRED_ALL (TCPF_TSQ_DEFERRED | \ 84940fc3423SEric Dumazet TCPF_WRITE_TIMER_DEFERRED | \ 85040fc3423SEric Dumazet TCPF_DELACK_TIMER_DEFERRED | \ 85140fc3423SEric Dumazet TCPF_MTU_REDUCED_DEFERRED) 85246d3ceabSEric Dumazet /** 85346d3ceabSEric Dumazet * tcp_release_cb - tcp release_sock() callback 85446d3ceabSEric Dumazet * @sk: socket 85546d3ceabSEric Dumazet * 85646d3ceabSEric Dumazet * called from release_sock() to perform protocol dependent 85746d3ceabSEric Dumazet * actions before socket release. 85846d3ceabSEric Dumazet */ 85946d3ceabSEric Dumazet void tcp_release_cb(struct sock *sk) 86046d3ceabSEric Dumazet { 8616f458dfbSEric Dumazet unsigned long flags, nflags; 86246d3ceabSEric Dumazet 8636f458dfbSEric Dumazet /* perform an atomic operation only if at least one flag is set */ 8646f458dfbSEric Dumazet do { 8657aa5470cSEric Dumazet flags = sk->sk_tsq_flags; 8666f458dfbSEric Dumazet if (!(flags & TCP_DEFERRED_ALL)) 8676f458dfbSEric Dumazet return; 8686f458dfbSEric Dumazet nflags = flags & ~TCP_DEFERRED_ALL; 8697aa5470cSEric Dumazet } while (cmpxchg(&sk->sk_tsq_flags, flags, nflags) != flags); 8706f458dfbSEric Dumazet 87173a6bab5SEric Dumazet if (flags & TCPF_TSQ_DEFERRED) { 87273a6bab5SEric Dumazet tcp_tsq_write(sk); 87373a6bab5SEric Dumazet __sock_put(sk); 87473a6bab5SEric Dumazet } 875c3f9b018SEric Dumazet /* Here begins the tricky part : 876c3f9b018SEric Dumazet * We are called from release_sock() with : 877c3f9b018SEric Dumazet * 1) BH disabled 878c3f9b018SEric Dumazet * 2) sk_lock.slock spinlock held 879c3f9b018SEric Dumazet * 3) socket owned by us (sk->sk_lock.owned == 1) 880c3f9b018SEric Dumazet * 881c3f9b018SEric Dumazet * But following code is meant to be called from BH handlers, 882c3f9b018SEric Dumazet * so we should keep BH disabled, but early release socket ownership 883c3f9b018SEric Dumazet */ 884c3f9b018SEric Dumazet sock_release_ownership(sk); 885c3f9b018SEric Dumazet 88640fc3423SEric Dumazet if (flags & TCPF_WRITE_TIMER_DEFERRED) { 8876f458dfbSEric Dumazet tcp_write_timer_handler(sk); 888144d56e9SEric Dumazet __sock_put(sk); 889144d56e9SEric Dumazet } 89040fc3423SEric Dumazet if (flags & TCPF_DELACK_TIMER_DEFERRED) { 8916f458dfbSEric Dumazet tcp_delack_timer_handler(sk); 892144d56e9SEric Dumazet __sock_put(sk); 893144d56e9SEric Dumazet } 89440fc3423SEric Dumazet if (flags & TCPF_MTU_REDUCED_DEFERRED) { 8954fab9071SNeal Cardwell inet_csk(sk)->icsk_af_ops->mtu_reduced(sk); 896144d56e9SEric Dumazet __sock_put(sk); 897144d56e9SEric Dumazet } 89846d3ceabSEric Dumazet } 89946d3ceabSEric Dumazet EXPORT_SYMBOL(tcp_release_cb); 90046d3ceabSEric Dumazet 90146d3ceabSEric Dumazet void __init tcp_tasklet_init(void) 90246d3ceabSEric Dumazet { 90346d3ceabSEric Dumazet int i; 90446d3ceabSEric Dumazet 90546d3ceabSEric Dumazet for_each_possible_cpu(i) { 90646d3ceabSEric Dumazet struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i); 90746d3ceabSEric Dumazet 90846d3ceabSEric Dumazet INIT_LIST_HEAD(&tsq->head); 90946d3ceabSEric Dumazet tasklet_init(&tsq->tasklet, 91046d3ceabSEric Dumazet tcp_tasklet_func, 91146d3ceabSEric Dumazet (unsigned long)tsq); 91246d3ceabSEric Dumazet } 91346d3ceabSEric Dumazet } 91446d3ceabSEric Dumazet 91546d3ceabSEric Dumazet /* 91646d3ceabSEric Dumazet * Write buffer destructor automatically called from kfree_skb. 9178e3bff96Sstephen hemminger * We can't xmit new skbs from this context, as we might already 91846d3ceabSEric Dumazet * hold qdisc lock. 91946d3ceabSEric Dumazet */ 920d6a4a104SEric Dumazet void tcp_wfree(struct sk_buff *skb) 92146d3ceabSEric Dumazet { 92246d3ceabSEric Dumazet struct sock *sk = skb->sk; 92346d3ceabSEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 924408f0a6cSEric Dumazet unsigned long flags, nval, oval; 9259b462d02SEric Dumazet 9269b462d02SEric Dumazet /* Keep one reference on sk_wmem_alloc. 9279b462d02SEric Dumazet * Will be released by sk_free() from here or tcp_tasklet_func() 9289b462d02SEric Dumazet */ 92914afee4bSReshetova, Elena WARN_ON(refcount_sub_and_test(skb->truesize - 1, &sk->sk_wmem_alloc)); 9309b462d02SEric Dumazet 9319b462d02SEric Dumazet /* If this softirq is serviced by ksoftirqd, we are likely under stress. 9329b462d02SEric Dumazet * Wait until our queues (qdisc + devices) are drained. 9339b462d02SEric Dumazet * This gives : 9349b462d02SEric Dumazet * - less callbacks to tcp_write_xmit(), reducing stress (batches) 9359b462d02SEric Dumazet * - chance for incoming ACK (processed by another cpu maybe) 9369b462d02SEric Dumazet * to migrate this flow (skb->ooo_okay will be eventually set) 9379b462d02SEric Dumazet */ 93814afee4bSReshetova, Elena if (refcount_read(&sk->sk_wmem_alloc) >= SKB_TRUESIZE(1) && this_cpu_ksoftirqd() == current) 9399b462d02SEric Dumazet goto out; 94046d3ceabSEric Dumazet 9417aa5470cSEric Dumazet for (oval = READ_ONCE(sk->sk_tsq_flags);; oval = nval) { 94246d3ceabSEric Dumazet struct tsq_tasklet *tsq; 943a9b204d1SEric Dumazet bool empty; 94446d3ceabSEric Dumazet 945408f0a6cSEric Dumazet if (!(oval & TSQF_THROTTLED) || (oval & TSQF_QUEUED)) 946408f0a6cSEric Dumazet goto out; 947408f0a6cSEric Dumazet 94873a6bab5SEric Dumazet nval = (oval & ~TSQF_THROTTLED) | TSQF_QUEUED; 9497aa5470cSEric Dumazet nval = cmpxchg(&sk->sk_tsq_flags, oval, nval); 950408f0a6cSEric Dumazet if (nval != oval) 951408f0a6cSEric Dumazet continue; 952408f0a6cSEric Dumazet 95346d3ceabSEric Dumazet /* queue this socket to tasklet queue */ 95446d3ceabSEric Dumazet local_irq_save(flags); 955903ceff7SChristoph Lameter tsq = this_cpu_ptr(&tsq_tasklet); 956a9b204d1SEric Dumazet empty = list_empty(&tsq->head); 95746d3ceabSEric Dumazet list_add(&tp->tsq_node, &tsq->head); 958a9b204d1SEric Dumazet if (empty) 95946d3ceabSEric Dumazet tasklet_schedule(&tsq->tasklet); 96046d3ceabSEric Dumazet local_irq_restore(flags); 9619b462d02SEric Dumazet return; 96246d3ceabSEric Dumazet } 9639b462d02SEric Dumazet out: 9649b462d02SEric Dumazet sk_free(sk); 96546d3ceabSEric Dumazet } 96646d3ceabSEric Dumazet 96773a6bab5SEric Dumazet /* Note: Called under soft irq. 96873a6bab5SEric Dumazet * We can call TCP stack right away, unless socket is owned by user. 969218af599SEric Dumazet */ 970218af599SEric Dumazet enum hrtimer_restart tcp_pace_kick(struct hrtimer *timer) 971218af599SEric Dumazet { 972218af599SEric Dumazet struct tcp_sock *tp = container_of(timer, struct tcp_sock, pacing_timer); 973218af599SEric Dumazet struct sock *sk = (struct sock *)tp; 974218af599SEric Dumazet 97573a6bab5SEric Dumazet tcp_tsq_handler(sk); 97673a6bab5SEric Dumazet sock_put(sk); 977218af599SEric Dumazet 978218af599SEric Dumazet return HRTIMER_NORESTART; 979218af599SEric Dumazet } 980218af599SEric Dumazet 981a7a25630SEric Dumazet static void tcp_update_skb_after_send(struct sock *sk, struct sk_buff *skb, 982a7a25630SEric Dumazet u64 prior_wstamp) 983e2080072SEric Dumazet { 984ab408b6dSEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 985ab408b6dSEric Dumazet 986ab408b6dSEric Dumazet if (sk->sk_pacing_status != SK_PACING_NONE) { 98776a9ebe8SEric Dumazet unsigned long rate = sk->sk_pacing_rate; 988ab408b6dSEric Dumazet 989ab408b6dSEric Dumazet /* Original sch_fq does not pace first 10 MSS 990ab408b6dSEric Dumazet * Note that tp->data_segs_out overflows after 2^32 packets, 991ab408b6dSEric Dumazet * this is a minor annoyance. 992ab408b6dSEric Dumazet */ 99376a9ebe8SEric Dumazet if (rate != ~0UL && rate && tp->data_segs_out >= 10) { 994a7a25630SEric Dumazet u64 len_ns = div64_ul((u64)skb->len * NSEC_PER_SEC, rate); 995a7a25630SEric Dumazet u64 credit = tp->tcp_wstamp_ns - prior_wstamp; 996a7a25630SEric Dumazet 997a7a25630SEric Dumazet /* take into account OS jitter */ 998a7a25630SEric Dumazet len_ns -= min_t(u64, len_ns / 2, credit); 999a7a25630SEric Dumazet tp->tcp_wstamp_ns += len_ns; 1000ab408b6dSEric Dumazet } 1001ab408b6dSEric Dumazet } 1002e2080072SEric Dumazet list_move_tail(&skb->tcp_tsorted_anchor, &tp->tsorted_sent_queue); 1003e2080072SEric Dumazet } 1004e2080072SEric Dumazet 10051da177e4SLinus Torvalds /* This routine actually transmits TCP packets queued in by 10061da177e4SLinus Torvalds * tcp_do_sendmsg(). This is used by both the initial 10071da177e4SLinus Torvalds * transmission and possible later retransmissions. 10081da177e4SLinus Torvalds * All SKB's seen here are completely headerless. It is our 10091da177e4SLinus Torvalds * job to build the TCP header, and pass the packet down to 10101da177e4SLinus Torvalds * IP so it can do the same plus pass the packet off to the 10111da177e4SLinus Torvalds * device. 10121da177e4SLinus Torvalds * 10131da177e4SLinus Torvalds * We are working here with either a clone of the original 10141da177e4SLinus Torvalds * SKB, or a fresh unique copy made by the retransmit engine. 10151da177e4SLinus Torvalds */ 10162987babbSYuchung Cheng static int __tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, 10172987babbSYuchung Cheng int clone_it, gfp_t gfp_mask, u32 rcv_nxt) 10181da177e4SLinus Torvalds { 10196687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 1020dfb4b9dcSDavid S. Miller struct inet_sock *inet; 1021dfb4b9dcSDavid S. Miller struct tcp_sock *tp; 1022dfb4b9dcSDavid S. Miller struct tcp_skb_cb *tcb; 102333ad798cSAdam Langley struct tcp_out_options opts; 102495c96174SEric Dumazet unsigned int tcp_options_size, tcp_header_size; 10258c72c65bSEric Dumazet struct sk_buff *oskb = NULL; 1026cfb6eeb4SYOSHIFUJI Hideaki struct tcp_md5sig_key *md5; 10271da177e4SLinus Torvalds struct tcphdr *th; 1028a7a25630SEric Dumazet u64 prior_wstamp; 10291da177e4SLinus Torvalds int err; 10301da177e4SLinus Torvalds 1031dfb4b9dcSDavid S. Miller BUG_ON(!skb || !tcp_skb_pcount(skb)); 10326f094b9eSLawrence Brakmo tp = tcp_sk(sk); 10337f12422cSYuchung Cheng prior_wstamp = tp->tcp_wstamp_ns; 10347f12422cSYuchung Cheng tp->tcp_wstamp_ns = max(tp->tcp_wstamp_ns, tp->tcp_clock_cache); 10357f12422cSYuchung Cheng skb->skb_mstamp_ns = tp->tcp_wstamp_ns; 1036ccdbb6e9SEric Dumazet if (clone_it) { 10376f094b9eSLawrence Brakmo TCP_SKB_CB(skb)->tx.in_flight = TCP_SKB_CB(skb)->end_seq 10386f094b9eSLawrence Brakmo - tp->snd_una; 10398c72c65bSEric Dumazet oskb = skb; 1040e2080072SEric Dumazet 1041e2080072SEric Dumazet tcp_skb_tsorted_save(oskb) { 1042e2080072SEric Dumazet if (unlikely(skb_cloned(oskb))) 1043e2080072SEric Dumazet skb = pskb_copy(oskb, gfp_mask); 1044dfb4b9dcSDavid S. Miller else 1045e2080072SEric Dumazet skb = skb_clone(oskb, gfp_mask); 1046e2080072SEric Dumazet } tcp_skb_tsorted_restore(oskb); 1047e2080072SEric Dumazet 1048dfb4b9dcSDavid S. Miller if (unlikely(!skb)) 1049dfb4b9dcSDavid S. Miller return -ENOBUFS; 1050dfb4b9dcSDavid S. Miller } 10515f6188a8SEric Dumazet 1052dfb4b9dcSDavid S. Miller inet = inet_sk(sk); 1053dfb4b9dcSDavid S. Miller tcb = TCP_SKB_CB(skb); 105433ad798cSAdam Langley memset(&opts, 0, sizeof(opts)); 10551da177e4SLinus Torvalds 10564de075e0SEric Dumazet if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) 105733ad798cSAdam Langley tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5); 105833ad798cSAdam Langley else 105933ad798cSAdam Langley tcp_options_size = tcp_established_options(sk, skb, &opts, 106033ad798cSAdam Langley &md5); 106133ad798cSAdam Langley tcp_header_size = tcp_options_size + sizeof(struct tcphdr); 10621da177e4SLinus Torvalds 1063547669d4SEric Dumazet /* if no packet is in qdisc/device queue, then allow XPS to select 1064b2532eb9SEric Dumazet * another queue. We can be called from tcp_tsq_handler() 106573a6bab5SEric Dumazet * which holds one reference to sk. 1066b2532eb9SEric Dumazet * 1067b2532eb9SEric Dumazet * TODO: Ideally, in-flight pure ACK packets should not matter here. 1068b2532eb9SEric Dumazet * One way to get this would be to set skb->truesize = 2 on them. 1069547669d4SEric Dumazet */ 1070b2532eb9SEric Dumazet skb->ooo_okay = sk_wmem_alloc_get(sk) < SKB_TRUESIZE(1); 10711da177e4SLinus Torvalds 107238ab52e8SEric Dumazet /* If we had to use memory reserve to allocate this skb, 107338ab52e8SEric Dumazet * this might cause drops if packet is looped back : 107438ab52e8SEric Dumazet * Other socket might not have SOCK_MEMALLOC. 107538ab52e8SEric Dumazet * Packets not looped back do not care about pfmemalloc. 107638ab52e8SEric Dumazet */ 107738ab52e8SEric Dumazet skb->pfmemalloc = 0; 107838ab52e8SEric Dumazet 1079aa8223c7SArnaldo Carvalho de Melo skb_push(skb, tcp_header_size); 1080aa8223c7SArnaldo Carvalho de Melo skb_reset_transport_header(skb); 108146d3ceabSEric Dumazet 108246d3ceabSEric Dumazet skb_orphan(skb); 108346d3ceabSEric Dumazet skb->sk = sk; 10841d2077acSEric Dumazet skb->destructor = skb_is_tcp_pure_ack(skb) ? __sock_wfree : tcp_wfree; 1085b73c3d0eSTom Herbert skb_set_hash_from_sk(skb, sk); 108614afee4bSReshetova, Elena refcount_add(skb->truesize, &sk->sk_wmem_alloc); 10871da177e4SLinus Torvalds 1088c3a2e837SJulian Anastasov skb_set_dst_pending_confirm(skb, sk->sk_dst_pending_confirm); 1089c3a2e837SJulian Anastasov 10901da177e4SLinus Torvalds /* Build TCP header and checksum it. */ 1091ea1627c2SEric Dumazet th = (struct tcphdr *)skb->data; 1092c720c7e8SEric Dumazet th->source = inet->inet_sport; 1093c720c7e8SEric Dumazet th->dest = inet->inet_dport; 10941da177e4SLinus Torvalds th->seq = htonl(tcb->seq); 10952987babbSYuchung Cheng th->ack_seq = htonl(rcv_nxt); 1096df7a3b07SAl Viro *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) | 10974de075e0SEric Dumazet tcb->tcp_flags); 1098dfb4b9dcSDavid S. Miller 10991da177e4SLinus Torvalds th->check = 0; 11001da177e4SLinus Torvalds th->urg_ptr = 0; 11011da177e4SLinus Torvalds 110233f5f57eSIlpo Järvinen /* The urg_mode check is necessary during a below snd_una win probe */ 11037691367dSHerbert Xu if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) { 11047691367dSHerbert Xu if (before(tp->snd_up, tcb->seq + 0x10000)) { 11051da177e4SLinus Torvalds th->urg_ptr = htons(tp->snd_up - tcb->seq); 11061da177e4SLinus Torvalds th->urg = 1; 11077691367dSHerbert Xu } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) { 11080eae88f3SEric Dumazet th->urg_ptr = htons(0xFFFF); 11097691367dSHerbert Xu th->urg = 1; 11107691367dSHerbert Xu } 11111da177e4SLinus Torvalds } 11121da177e4SLinus Torvalds 1113bd0388aeSWilliam Allen Simpson tcp_options_write((__be32 *)(th + 1), tp, &opts); 111451466a75SEric Dumazet skb_shinfo(skb)->gso_type = sk->sk_gso_type; 1115ea1627c2SEric Dumazet if (likely(!(tcb->tcp_flags & TCPHDR_SYN))) { 1116ea1627c2SEric Dumazet th->window = htons(tcp_select_window(sk)); 1117ea1627c2SEric Dumazet tcp_ecn_send(sk, skb, th, tcp_header_size); 1118ea1627c2SEric Dumazet } else { 1119ea1627c2SEric Dumazet /* RFC1323: The window in SYN & SYN/ACK segments 1120ea1627c2SEric Dumazet * is never scaled. 1121ea1627c2SEric Dumazet */ 1122ea1627c2SEric Dumazet th->window = htons(min(tp->rcv_wnd, 65535U)); 1123ea1627c2SEric Dumazet } 1124cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 1125cfb6eeb4SYOSHIFUJI Hideaki /* Calculate the MD5 hash, as we have all we need now */ 1126cfb6eeb4SYOSHIFUJI Hideaki if (md5) { 1127a465419bSEric Dumazet sk_nocaps_add(sk, NETIF_F_GSO_MASK); 1128bd0388aeSWilliam Allen Simpson tp->af_specific->calc_md5_hash(opts.hash_location, 112939f8e58eSEric Dumazet md5, sk, skb); 1130cfb6eeb4SYOSHIFUJI Hideaki } 1131cfb6eeb4SYOSHIFUJI Hideaki #endif 1132cfb6eeb4SYOSHIFUJI Hideaki 1133bb296246SHerbert Xu icsk->icsk_af_ops->send_check(sk, skb); 11341da177e4SLinus Torvalds 11354de075e0SEric Dumazet if (likely(tcb->tcp_flags & TCPHDR_ACK)) 113627cde44aSYuchung Cheng tcp_event_ack_sent(sk, tcp_skb_pcount(skb), rcv_nxt); 11371da177e4SLinus Torvalds 1138a44d6eacSMartin KaFai Lau if (skb->len != tcp_header_size) { 1139cf533ea5SEric Dumazet tcp_event_data_sent(tp, sk); 1140a44d6eacSMartin KaFai Lau tp->data_segs_out += tcp_skb_pcount(skb); 1141ba113c3aSWei Wang tp->bytes_sent += skb->len - tcp_header_size; 1142a44d6eacSMartin KaFai Lau } 11431da177e4SLinus Torvalds 1144bd37a088SWei Yongjun if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq) 1145aa2ea058STom Herbert TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS, 1146aa2ea058STom Herbert tcp_skb_pcount(skb)); 11471da177e4SLinus Torvalds 11482efd055cSMarcelo Ricardo Leitner tp->segs_out += tcp_skb_pcount(skb); 1149f69ad292SEric Dumazet /* OK, its time to fill skb_shinfo(skb)->gso_{segs|size} */ 1150cd7d8498SEric Dumazet skb_shinfo(skb)->gso_segs = tcp_skb_pcount(skb); 1151f69ad292SEric Dumazet skb_shinfo(skb)->gso_size = tcp_skb_mss(skb); 1152cd7d8498SEric Dumazet 1153d3edd06eSEric Dumazet /* Leave earliest departure time in skb->tstamp (skb->skb_mstamp_ns) */ 1154971f10ecSEric Dumazet 1155971f10ecSEric Dumazet /* Cleanup our debris for IP stacks */ 1156971f10ecSEric Dumazet memset(skb->cb, 0, max(sizeof(struct inet_skb_parm), 1157971f10ecSEric Dumazet sizeof(struct inet6_skb_parm))); 1158971f10ecSEric Dumazet 1159b0270e91SEric Dumazet err = icsk->icsk_af_ops->queue_xmit(sk, skb, &inet->cork.fl); 11607faee5c0SEric Dumazet 11618c72c65bSEric Dumazet if (unlikely(err > 0)) { 11625ee2c941SChristoph Paasch tcp_enter_cwr(sk); 11638c72c65bSEric Dumazet err = net_xmit_eval(err); 11648c72c65bSEric Dumazet } 1165fc225799SEric Dumazet if (!err && oskb) { 1166a7a25630SEric Dumazet tcp_update_skb_after_send(sk, oskb, prior_wstamp); 1167fc225799SEric Dumazet tcp_rate_skb_sent(sk, oskb); 1168fc225799SEric Dumazet } 11698c72c65bSEric Dumazet return err; 11701da177e4SLinus Torvalds } 11711da177e4SLinus Torvalds 11722987babbSYuchung Cheng static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it, 11732987babbSYuchung Cheng gfp_t gfp_mask) 11742987babbSYuchung Cheng { 11752987babbSYuchung Cheng return __tcp_transmit_skb(sk, skb, clone_it, gfp_mask, 11762987babbSYuchung Cheng tcp_sk(sk)->rcv_nxt); 11772987babbSYuchung Cheng } 11782987babbSYuchung Cheng 117967edfef7SAndi Kleen /* This routine just queues the buffer for sending. 11801da177e4SLinus Torvalds * 11811da177e4SLinus Torvalds * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames, 11821da177e4SLinus Torvalds * otherwise socket can stall. 11831da177e4SLinus Torvalds */ 11841da177e4SLinus Torvalds static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb) 11851da177e4SLinus Torvalds { 11861da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 11871da177e4SLinus Torvalds 11881da177e4SLinus Torvalds /* Advance write_seq and place onto the write_queue. */ 11891da177e4SLinus Torvalds tp->write_seq = TCP_SKB_CB(skb)->end_seq; 1190f4a775d1SEric Dumazet __skb_header_release(skb); 1191fe067e8aSDavid S. Miller tcp_add_write_queue_tail(sk, skb); 11923ab224beSHideo Aoki sk->sk_wmem_queued += skb->truesize; 11933ab224beSHideo Aoki sk_mem_charge(sk, skb->truesize); 11941da177e4SLinus Torvalds } 11951da177e4SLinus Torvalds 119667edfef7SAndi Kleen /* Initialize TSO segments for a packet. */ 11975bbb432cSEric Dumazet static void tcp_set_skb_tso_segs(struct sk_buff *skb, unsigned int mss_now) 1198f6302d1dSDavid S. Miller { 11994a64fd6cSEric Dumazet if (skb->len <= mss_now) { 1200f6302d1dSDavid S. Miller /* Avoid the costly divide in the normal 1201f6302d1dSDavid S. Miller * non-TSO case. 1202f6302d1dSDavid S. Miller */ 1203cd7d8498SEric Dumazet tcp_skb_pcount_set(skb, 1); 1204f69ad292SEric Dumazet TCP_SKB_CB(skb)->tcp_gso_size = 0; 1205f6302d1dSDavid S. Miller } else { 1206cd7d8498SEric Dumazet tcp_skb_pcount_set(skb, DIV_ROUND_UP(skb->len, mss_now)); 1207f69ad292SEric Dumazet TCP_SKB_CB(skb)->tcp_gso_size = mss_now; 12081da177e4SLinus Torvalds } 12091da177e4SLinus Torvalds } 12101da177e4SLinus Torvalds 1211797108d1SIlpo Järvinen /* Pcount in the middle of the write queue got changed, we need to do various 1212797108d1SIlpo Järvinen * tweaks to fix counters 1213797108d1SIlpo Järvinen */ 1214cf533ea5SEric Dumazet static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr) 1215797108d1SIlpo Järvinen { 1216797108d1SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 1217797108d1SIlpo Järvinen 1218797108d1SIlpo Järvinen tp->packets_out -= decr; 1219797108d1SIlpo Järvinen 1220797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 1221797108d1SIlpo Järvinen tp->sacked_out -= decr; 1222797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) 1223797108d1SIlpo Järvinen tp->retrans_out -= decr; 1224797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST) 1225797108d1SIlpo Järvinen tp->lost_out -= decr; 1226797108d1SIlpo Järvinen 1227797108d1SIlpo Järvinen /* Reno case is special. Sigh... */ 1228797108d1SIlpo Järvinen if (tcp_is_reno(tp) && decr > 0) 1229797108d1SIlpo Järvinen tp->sacked_out -= min_t(u32, tp->sacked_out, decr); 1230797108d1SIlpo Järvinen 1231797108d1SIlpo Järvinen if (tp->lost_skb_hint && 1232797108d1SIlpo Järvinen before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) && 1233713bafeaSYuchung Cheng (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) 1234797108d1SIlpo Järvinen tp->lost_cnt_hint -= decr; 1235797108d1SIlpo Järvinen 1236797108d1SIlpo Järvinen tcp_verify_left_out(tp); 1237797108d1SIlpo Järvinen } 1238797108d1SIlpo Järvinen 12390a2cf20cSSoheil Hassas Yeganeh static bool tcp_has_tx_tstamp(const struct sk_buff *skb) 12400a2cf20cSSoheil Hassas Yeganeh { 12410a2cf20cSSoheil Hassas Yeganeh return TCP_SKB_CB(skb)->txstamp_ack || 12420a2cf20cSSoheil Hassas Yeganeh (skb_shinfo(skb)->tx_flags & SKBTX_ANY_TSTAMP); 12430a2cf20cSSoheil Hassas Yeganeh } 12440a2cf20cSSoheil Hassas Yeganeh 1245490cc7d0SWillem de Bruijn static void tcp_fragment_tstamp(struct sk_buff *skb, struct sk_buff *skb2) 1246490cc7d0SWillem de Bruijn { 1247490cc7d0SWillem de Bruijn struct skb_shared_info *shinfo = skb_shinfo(skb); 1248490cc7d0SWillem de Bruijn 12490a2cf20cSSoheil Hassas Yeganeh if (unlikely(tcp_has_tx_tstamp(skb)) && 1250490cc7d0SWillem de Bruijn !before(shinfo->tskey, TCP_SKB_CB(skb2)->seq)) { 1251490cc7d0SWillem de Bruijn struct skb_shared_info *shinfo2 = skb_shinfo(skb2); 1252490cc7d0SWillem de Bruijn u8 tsflags = shinfo->tx_flags & SKBTX_ANY_TSTAMP; 1253490cc7d0SWillem de Bruijn 1254490cc7d0SWillem de Bruijn shinfo->tx_flags &= ~tsflags; 1255490cc7d0SWillem de Bruijn shinfo2->tx_flags |= tsflags; 1256490cc7d0SWillem de Bruijn swap(shinfo->tskey, shinfo2->tskey); 1257b51e13faSMartin KaFai Lau TCP_SKB_CB(skb2)->txstamp_ack = TCP_SKB_CB(skb)->txstamp_ack; 1258b51e13faSMartin KaFai Lau TCP_SKB_CB(skb)->txstamp_ack = 0; 1259490cc7d0SWillem de Bruijn } 1260490cc7d0SWillem de Bruijn } 1261490cc7d0SWillem de Bruijn 1262a166140eSMartin KaFai Lau static void tcp_skb_fragment_eor(struct sk_buff *skb, struct sk_buff *skb2) 1263a166140eSMartin KaFai Lau { 1264a166140eSMartin KaFai Lau TCP_SKB_CB(skb2)->eor = TCP_SKB_CB(skb)->eor; 1265a166140eSMartin KaFai Lau TCP_SKB_CB(skb)->eor = 0; 1266a166140eSMartin KaFai Lau } 1267a166140eSMartin KaFai Lau 126875c119afSEric Dumazet /* Insert buff after skb on the write or rtx queue of sk. */ 126975c119afSEric Dumazet static void tcp_insert_write_queue_after(struct sk_buff *skb, 127075c119afSEric Dumazet struct sk_buff *buff, 127175c119afSEric Dumazet struct sock *sk, 127275c119afSEric Dumazet enum tcp_queue tcp_queue) 127375c119afSEric Dumazet { 127475c119afSEric Dumazet if (tcp_queue == TCP_FRAG_IN_WRITE_QUEUE) 127575c119afSEric Dumazet __skb_queue_after(&sk->sk_write_queue, skb, buff); 127675c119afSEric Dumazet else 127775c119afSEric Dumazet tcp_rbtree_insert(&sk->tcp_rtx_queue, buff); 127875c119afSEric Dumazet } 127975c119afSEric Dumazet 12801da177e4SLinus Torvalds /* Function to create two new TCP segments. Shrinks the given segment 12811da177e4SLinus Torvalds * to the specified size and appends a new segment with the rest of the 12821da177e4SLinus Torvalds * packet to the list. This won't be called frequently, I hope. 12831da177e4SLinus Torvalds * Remember, these are still headerless SKBs at this point. 12841da177e4SLinus Torvalds */ 128575c119afSEric Dumazet int tcp_fragment(struct sock *sk, enum tcp_queue tcp_queue, 128675c119afSEric Dumazet struct sk_buff *skb, u32 len, 12876cc55e09SOctavian Purdila unsigned int mss_now, gfp_t gfp) 12881da177e4SLinus Torvalds { 12891da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 12901da177e4SLinus Torvalds struct sk_buff *buff; 12916475be16SDavid S. Miller int nsize, old_factor; 1292b60b49eaSHerbert Xu int nlen; 12939ce01461SIlpo Järvinen u8 flags; 12941da177e4SLinus Torvalds 12952fceec13SIlpo Järvinen if (WARN_ON(len > skb->len)) 12962fceec13SIlpo Järvinen return -EINVAL; 12976a438bbeSStephen Hemminger 12981da177e4SLinus Torvalds nsize = skb_headlen(skb) - len; 12991da177e4SLinus Torvalds if (nsize < 0) 13001da177e4SLinus Torvalds nsize = 0; 13011da177e4SLinus Torvalds 13026cc55e09SOctavian Purdila if (skb_unclone(skb, gfp)) 13031da177e4SLinus Torvalds return -ENOMEM; 13041da177e4SLinus Torvalds 13051da177e4SLinus Torvalds /* Get a new skb... force flag on. */ 1306eb934478SEric Dumazet buff = sk_stream_alloc_skb(sk, nsize, gfp, true); 130751456b29SIan Morris if (!buff) 13081da177e4SLinus Torvalds return -ENOMEM; /* We'll just try again later. */ 1309ef5cb973SHerbert Xu 13103ab224beSHideo Aoki sk->sk_wmem_queued += buff->truesize; 13113ab224beSHideo Aoki sk_mem_charge(sk, buff->truesize); 1312b60b49eaSHerbert Xu nlen = skb->len - len - nsize; 1313b60b49eaSHerbert Xu buff->truesize += nlen; 1314b60b49eaSHerbert Xu skb->truesize -= nlen; 13151da177e4SLinus Torvalds 13161da177e4SLinus Torvalds /* Correct the sequence numbers. */ 13171da177e4SLinus Torvalds TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; 13181da177e4SLinus Torvalds TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; 13191da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; 13201da177e4SLinus Torvalds 13211da177e4SLinus Torvalds /* PSH and FIN should only be set in the second packet. */ 13224de075e0SEric Dumazet flags = TCP_SKB_CB(skb)->tcp_flags; 13234de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); 13244de075e0SEric Dumazet TCP_SKB_CB(buff)->tcp_flags = flags; 1325e14c3cafSHerbert Xu TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked; 1326a166140eSMartin KaFai Lau tcp_skb_fragment_eor(skb, buff); 13271da177e4SLinus Torvalds 13281da177e4SLinus Torvalds skb_split(skb, buff, len); 13291da177e4SLinus Torvalds 133098be9b12SEric Dumazet buff->ip_summed = CHECKSUM_PARTIAL; 13311da177e4SLinus Torvalds 1332a61bbcf2SPatrick McHardy buff->tstamp = skb->tstamp; 1333490cc7d0SWillem de Bruijn tcp_fragment_tstamp(skb, buff); 13341da177e4SLinus Torvalds 13356475be16SDavid S. Miller old_factor = tcp_skb_pcount(skb); 13366475be16SDavid S. Miller 13371da177e4SLinus Torvalds /* Fix up tso_factor for both original and new SKB. */ 13385bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss_now); 13395bbb432cSEric Dumazet tcp_set_skb_tso_segs(buff, mss_now); 13401da177e4SLinus Torvalds 1341b9f64820SYuchung Cheng /* Update delivered info for the new segment */ 1342b9f64820SYuchung Cheng TCP_SKB_CB(buff)->tx = TCP_SKB_CB(skb)->tx; 1343b9f64820SYuchung Cheng 13446475be16SDavid S. Miller /* If this packet has been sent out already, we must 13456475be16SDavid S. Miller * adjust the various packet counters. 13466475be16SDavid S. Miller */ 1347cf0b450cSHerbert Xu if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) { 13486475be16SDavid S. Miller int diff = old_factor - tcp_skb_pcount(skb) - 13496475be16SDavid S. Miller tcp_skb_pcount(buff); 13501da177e4SLinus Torvalds 1351797108d1SIlpo Järvinen if (diff) 1352797108d1SIlpo Järvinen tcp_adjust_pcount(sk, skb, diff); 13531da177e4SLinus Torvalds } 13541da177e4SLinus Torvalds 13551da177e4SLinus Torvalds /* Link BUFF into the send queue. */ 1356f4a775d1SEric Dumazet __skb_header_release(buff); 135775c119afSEric Dumazet tcp_insert_write_queue_after(skb, buff, sk, tcp_queue); 1358f67971e6SEric Dumazet if (tcp_queue == TCP_FRAG_IN_RTX_QUEUE) 1359e2080072SEric Dumazet list_add(&buff->tcp_tsorted_anchor, &skb->tcp_tsorted_anchor); 13601da177e4SLinus Torvalds 13611da177e4SLinus Torvalds return 0; 13621da177e4SLinus Torvalds } 13631da177e4SLinus Torvalds 1364f4d01666SEric Dumazet /* This is similar to __pskb_pull_tail(). The difference is that pulled 1365f4d01666SEric Dumazet * data is not copied, but immediately discarded. 13661da177e4SLinus Torvalds */ 13677162fb24SEric Dumazet static int __pskb_trim_head(struct sk_buff *skb, int len) 13681da177e4SLinus Torvalds { 13697b7fc97aSEric Dumazet struct skb_shared_info *shinfo; 13701da177e4SLinus Torvalds int i, k, eat; 13711da177e4SLinus Torvalds 13724fa48bf3SEric Dumazet eat = min_t(int, len, skb_headlen(skb)); 13734fa48bf3SEric Dumazet if (eat) { 13744fa48bf3SEric Dumazet __skb_pull(skb, eat); 13754fa48bf3SEric Dumazet len -= eat; 13764fa48bf3SEric Dumazet if (!len) 13777162fb24SEric Dumazet return 0; 13784fa48bf3SEric Dumazet } 13791da177e4SLinus Torvalds eat = len; 13801da177e4SLinus Torvalds k = 0; 13817b7fc97aSEric Dumazet shinfo = skb_shinfo(skb); 13827b7fc97aSEric Dumazet for (i = 0; i < shinfo->nr_frags; i++) { 13837b7fc97aSEric Dumazet int size = skb_frag_size(&shinfo->frags[i]); 13849e903e08SEric Dumazet 13859e903e08SEric Dumazet if (size <= eat) { 1386aff65da0SIan Campbell skb_frag_unref(skb, i); 13879e903e08SEric Dumazet eat -= size; 13881da177e4SLinus Torvalds } else { 13897b7fc97aSEric Dumazet shinfo->frags[k] = shinfo->frags[i]; 13901da177e4SLinus Torvalds if (eat) { 13917b7fc97aSEric Dumazet shinfo->frags[k].page_offset += eat; 13927b7fc97aSEric Dumazet skb_frag_size_sub(&shinfo->frags[k], eat); 13931da177e4SLinus Torvalds eat = 0; 13941da177e4SLinus Torvalds } 13951da177e4SLinus Torvalds k++; 13961da177e4SLinus Torvalds } 13971da177e4SLinus Torvalds } 13987b7fc97aSEric Dumazet shinfo->nr_frags = k; 13991da177e4SLinus Torvalds 14001da177e4SLinus Torvalds skb->data_len -= len; 14011da177e4SLinus Torvalds skb->len = skb->data_len; 14027162fb24SEric Dumazet return len; 14031da177e4SLinus Torvalds } 14041da177e4SLinus Torvalds 140567edfef7SAndi Kleen /* Remove acked data from a packet in the transmit queue. */ 14061da177e4SLinus Torvalds int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len) 14071da177e4SLinus Torvalds { 14087162fb24SEric Dumazet u32 delta_truesize; 14097162fb24SEric Dumazet 141014bbd6a5SPravin B Shelar if (skb_unclone(skb, GFP_ATOMIC)) 14111da177e4SLinus Torvalds return -ENOMEM; 14121da177e4SLinus Torvalds 14137162fb24SEric Dumazet delta_truesize = __pskb_trim_head(skb, len); 14141da177e4SLinus Torvalds 14151da177e4SLinus Torvalds TCP_SKB_CB(skb)->seq += len; 141684fa7933SPatrick McHardy skb->ip_summed = CHECKSUM_PARTIAL; 14171da177e4SLinus Torvalds 14187162fb24SEric Dumazet if (delta_truesize) { 14197162fb24SEric Dumazet skb->truesize -= delta_truesize; 14207162fb24SEric Dumazet sk->sk_wmem_queued -= delta_truesize; 14217162fb24SEric Dumazet sk_mem_uncharge(sk, delta_truesize); 14221da177e4SLinus Torvalds sock_set_flag(sk, SOCK_QUEUE_SHRUNK); 14237162fb24SEric Dumazet } 14241da177e4SLinus Torvalds 14255b35e1e6SNeal Cardwell /* Any change of skb->len requires recalculation of tso factor. */ 14261da177e4SLinus Torvalds if (tcp_skb_pcount(skb) > 1) 14275bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, tcp_skb_mss(skb)); 14281da177e4SLinus Torvalds 14291da177e4SLinus Torvalds return 0; 14301da177e4SLinus Torvalds } 14311da177e4SLinus Torvalds 14321b63edd6SYuchung Cheng /* Calculate MSS not accounting any TCP options. */ 14331b63edd6SYuchung Cheng static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu) 14345d424d5aSJohn Heffner { 1435cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1436cf533ea5SEric Dumazet const struct inet_connection_sock *icsk = inet_csk(sk); 14375d424d5aSJohn Heffner int mss_now; 14385d424d5aSJohn Heffner 14395d424d5aSJohn Heffner /* Calculate base mss without TCP options: 14405d424d5aSJohn Heffner It is MMS_S - sizeof(tcphdr) of rfc1122 14415d424d5aSJohn Heffner */ 14425d424d5aSJohn Heffner mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr); 14435d424d5aSJohn Heffner 144467469601SEric Dumazet /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */ 144567469601SEric Dumazet if (icsk->icsk_af_ops->net_frag_header_len) { 144667469601SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 144767469601SEric Dumazet 144867469601SEric Dumazet if (dst && dst_allfrag(dst)) 144967469601SEric Dumazet mss_now -= icsk->icsk_af_ops->net_frag_header_len; 145067469601SEric Dumazet } 145167469601SEric Dumazet 14525d424d5aSJohn Heffner /* Clamp it (mss_clamp does not include tcp options) */ 14535d424d5aSJohn Heffner if (mss_now > tp->rx_opt.mss_clamp) 14545d424d5aSJohn Heffner mss_now = tp->rx_opt.mss_clamp; 14555d424d5aSJohn Heffner 14565d424d5aSJohn Heffner /* Now subtract optional transport overhead */ 14575d424d5aSJohn Heffner mss_now -= icsk->icsk_ext_hdr_len; 14585d424d5aSJohn Heffner 14595d424d5aSJohn Heffner /* Then reserve room for full set of TCP options and 8 bytes of data */ 14605d424d5aSJohn Heffner if (mss_now < 48) 14615d424d5aSJohn Heffner mss_now = 48; 14625d424d5aSJohn Heffner return mss_now; 14635d424d5aSJohn Heffner } 14645d424d5aSJohn Heffner 14651b63edd6SYuchung Cheng /* Calculate MSS. Not accounting for SACKs here. */ 14661b63edd6SYuchung Cheng int tcp_mtu_to_mss(struct sock *sk, int pmtu) 14671b63edd6SYuchung Cheng { 14681b63edd6SYuchung Cheng /* Subtract TCP options size, not including SACKs */ 14691b63edd6SYuchung Cheng return __tcp_mtu_to_mss(sk, pmtu) - 14701b63edd6SYuchung Cheng (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr)); 14711b63edd6SYuchung Cheng } 14721b63edd6SYuchung Cheng 14735d424d5aSJohn Heffner /* Inverse of above */ 147467469601SEric Dumazet int tcp_mss_to_mtu(struct sock *sk, int mss) 14755d424d5aSJohn Heffner { 1476cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1477cf533ea5SEric Dumazet const struct inet_connection_sock *icsk = inet_csk(sk); 14785d424d5aSJohn Heffner int mtu; 14795d424d5aSJohn Heffner 14805d424d5aSJohn Heffner mtu = mss + 14815d424d5aSJohn Heffner tp->tcp_header_len + 14825d424d5aSJohn Heffner icsk->icsk_ext_hdr_len + 14835d424d5aSJohn Heffner icsk->icsk_af_ops->net_header_len; 14845d424d5aSJohn Heffner 148567469601SEric Dumazet /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */ 148667469601SEric Dumazet if (icsk->icsk_af_ops->net_frag_header_len) { 148767469601SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 148867469601SEric Dumazet 148967469601SEric Dumazet if (dst && dst_allfrag(dst)) 149067469601SEric Dumazet mtu += icsk->icsk_af_ops->net_frag_header_len; 149167469601SEric Dumazet } 14925d424d5aSJohn Heffner return mtu; 14935d424d5aSJohn Heffner } 1494556c6b46SNeal Cardwell EXPORT_SYMBOL(tcp_mss_to_mtu); 14955d424d5aSJohn Heffner 149667edfef7SAndi Kleen /* MTU probing init per socket */ 14975d424d5aSJohn Heffner void tcp_mtup_init(struct sock *sk) 14985d424d5aSJohn Heffner { 14995d424d5aSJohn Heffner struct tcp_sock *tp = tcp_sk(sk); 15005d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 1501b0f9ca53SFan Du struct net *net = sock_net(sk); 15025d424d5aSJohn Heffner 1503b0f9ca53SFan Du icsk->icsk_mtup.enabled = net->ipv4.sysctl_tcp_mtu_probing > 1; 15045d424d5aSJohn Heffner icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) + 15055d424d5aSJohn Heffner icsk->icsk_af_ops->net_header_len; 1506b0f9ca53SFan Du icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, net->ipv4.sysctl_tcp_base_mss); 15075d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = 0; 150805cbc0dbSFan Du if (icsk->icsk_mtup.enabled) 1509c74df29aSEric Dumazet icsk->icsk_mtup.probe_timestamp = tcp_jiffies32; 15105d424d5aSJohn Heffner } 15114bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_mtup_init); 15125d424d5aSJohn Heffner 15131da177e4SLinus Torvalds /* This function synchronize snd mss to current pmtu/exthdr set. 15141da177e4SLinus Torvalds 15151da177e4SLinus Torvalds tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts 15161da177e4SLinus Torvalds for TCP options, but includes only bare TCP header. 15171da177e4SLinus Torvalds 15181da177e4SLinus Torvalds tp->rx_opt.mss_clamp is mss negotiated at connection setup. 1519caa20d9aSStephen Hemminger It is minimum of user_mss and mss received with SYN. 15201da177e4SLinus Torvalds It also does not include TCP options. 15211da177e4SLinus Torvalds 1522d83d8461SArnaldo Carvalho de Melo inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function. 15231da177e4SLinus Torvalds 15241da177e4SLinus Torvalds tp->mss_cache is current effective sending mss, including 15251da177e4SLinus Torvalds all tcp options except for SACKs. It is evaluated, 15261da177e4SLinus Torvalds taking into account current pmtu, but never exceeds 15271da177e4SLinus Torvalds tp->rx_opt.mss_clamp. 15281da177e4SLinus Torvalds 15291da177e4SLinus Torvalds NOTE1. rfc1122 clearly states that advertised MSS 15301da177e4SLinus Torvalds DOES NOT include either tcp or ip options. 15311da177e4SLinus Torvalds 1532d83d8461SArnaldo Carvalho de Melo NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache 1533d83d8461SArnaldo Carvalho de Melo are READ ONLY outside this function. --ANK (980731) 15341da177e4SLinus Torvalds */ 15351da177e4SLinus Torvalds unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu) 15361da177e4SLinus Torvalds { 15371da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 1538d83d8461SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 15395d424d5aSJohn Heffner int mss_now; 15401da177e4SLinus Torvalds 15415d424d5aSJohn Heffner if (icsk->icsk_mtup.search_high > pmtu) 15425d424d5aSJohn Heffner icsk->icsk_mtup.search_high = pmtu; 15431da177e4SLinus Torvalds 15445d424d5aSJohn Heffner mss_now = tcp_mtu_to_mss(sk, pmtu); 1545409d22b4SIlpo Järvinen mss_now = tcp_bound_to_half_wnd(tp, mss_now); 15461da177e4SLinus Torvalds 15471da177e4SLinus Torvalds /* And store cached results */ 1548d83d8461SArnaldo Carvalho de Melo icsk->icsk_pmtu_cookie = pmtu; 15495d424d5aSJohn Heffner if (icsk->icsk_mtup.enabled) 15505d424d5aSJohn Heffner mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low)); 1551c1b4a7e6SDavid S. Miller tp->mss_cache = mss_now; 15521da177e4SLinus Torvalds 15531da177e4SLinus Torvalds return mss_now; 15541da177e4SLinus Torvalds } 15554bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_sync_mss); 15561da177e4SLinus Torvalds 15571da177e4SLinus Torvalds /* Compute the current effective MSS, taking SACKs and IP options, 15581da177e4SLinus Torvalds * and even PMTU discovery events into account. 15591da177e4SLinus Torvalds */ 15600c54b85fSIlpo Järvinen unsigned int tcp_current_mss(struct sock *sk) 15611da177e4SLinus Torvalds { 1562cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1563cf533ea5SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 1564c1b4a7e6SDavid S. Miller u32 mss_now; 156595c96174SEric Dumazet unsigned int header_len; 156633ad798cSAdam Langley struct tcp_out_options opts; 156733ad798cSAdam Langley struct tcp_md5sig_key *md5; 15681da177e4SLinus Torvalds 1569c1b4a7e6SDavid S. Miller mss_now = tp->mss_cache; 1570c1b4a7e6SDavid S. Miller 15711da177e4SLinus Torvalds if (dst) { 15721da177e4SLinus Torvalds u32 mtu = dst_mtu(dst); 1573d83d8461SArnaldo Carvalho de Melo if (mtu != inet_csk(sk)->icsk_pmtu_cookie) 15741da177e4SLinus Torvalds mss_now = tcp_sync_mss(sk, mtu); 15751da177e4SLinus Torvalds } 15761da177e4SLinus Torvalds 157733ad798cSAdam Langley header_len = tcp_established_options(sk, NULL, &opts, &md5) + 157833ad798cSAdam Langley sizeof(struct tcphdr); 157933ad798cSAdam Langley /* The mss_cache is sized based on tp->tcp_header_len, which assumes 158033ad798cSAdam Langley * some common options. If this is an odd packet (because we have SACK 158133ad798cSAdam Langley * blocks etc) then our calculated header_len will be different, and 158233ad798cSAdam Langley * we have to adjust mss_now correspondingly */ 158333ad798cSAdam Langley if (header_len != tp->tcp_header_len) { 158433ad798cSAdam Langley int delta = (int) header_len - tp->tcp_header_len; 158533ad798cSAdam Langley mss_now -= delta; 158633ad798cSAdam Langley } 1587cfb6eeb4SYOSHIFUJI Hideaki 15881da177e4SLinus Torvalds return mss_now; 15891da177e4SLinus Torvalds } 15901da177e4SLinus Torvalds 159186fd14adSWeiping Pan /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto. 159286fd14adSWeiping Pan * As additional protections, we do not touch cwnd in retransmission phases, 159386fd14adSWeiping Pan * and if application hit its sndbuf limit recently. 159486fd14adSWeiping Pan */ 159586fd14adSWeiping Pan static void tcp_cwnd_application_limited(struct sock *sk) 1596a762a980SDavid S. Miller { 15979e412ba7SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 1598a762a980SDavid S. Miller 159986fd14adSWeiping Pan if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open && 160086fd14adSWeiping Pan sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) { 160186fd14adSWeiping Pan /* Limited by application or receiver window. */ 160286fd14adSWeiping Pan u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk)); 160386fd14adSWeiping Pan u32 win_used = max(tp->snd_cwnd_used, init_win); 160486fd14adSWeiping Pan if (win_used < tp->snd_cwnd) { 160586fd14adSWeiping Pan tp->snd_ssthresh = tcp_current_ssthresh(sk); 160686fd14adSWeiping Pan tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1; 160786fd14adSWeiping Pan } 160886fd14adSWeiping Pan tp->snd_cwnd_used = 0; 160986fd14adSWeiping Pan } 1610c2203cf7SEric Dumazet tp->snd_cwnd_stamp = tcp_jiffies32; 161186fd14adSWeiping Pan } 161286fd14adSWeiping Pan 1613ca8a2263SNeal Cardwell static void tcp_cwnd_validate(struct sock *sk, bool is_cwnd_limited) 1614a762a980SDavid S. Miller { 16151b1fc3fdSWei Wang const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops; 1616a762a980SDavid S. Miller struct tcp_sock *tp = tcp_sk(sk); 1617a762a980SDavid S. Miller 1618ca8a2263SNeal Cardwell /* Track the maximum number of outstanding packets in each 1619ca8a2263SNeal Cardwell * window, and remember whether we were cwnd-limited then. 1620ca8a2263SNeal Cardwell */ 1621ca8a2263SNeal Cardwell if (!before(tp->snd_una, tp->max_packets_seq) || 1622ca8a2263SNeal Cardwell tp->packets_out > tp->max_packets_out) { 1623ca8a2263SNeal Cardwell tp->max_packets_out = tp->packets_out; 1624ca8a2263SNeal Cardwell tp->max_packets_seq = tp->snd_nxt; 1625ca8a2263SNeal Cardwell tp->is_cwnd_limited = is_cwnd_limited; 1626ca8a2263SNeal Cardwell } 1627e114a710SEric Dumazet 162824901551SEric Dumazet if (tcp_is_cwnd_limited(sk)) { 1629a762a980SDavid S. Miller /* Network is feed fully. */ 1630a762a980SDavid S. Miller tp->snd_cwnd_used = 0; 1631c2203cf7SEric Dumazet tp->snd_cwnd_stamp = tcp_jiffies32; 1632a762a980SDavid S. Miller } else { 1633a762a980SDavid S. Miller /* Network starves. */ 1634a762a980SDavid S. Miller if (tp->packets_out > tp->snd_cwnd_used) 1635a762a980SDavid S. Miller tp->snd_cwnd_used = tp->packets_out; 1636a762a980SDavid S. Miller 1637b510f0d2SEric Dumazet if (sock_net(sk)->ipv4.sysctl_tcp_slow_start_after_idle && 1638c2203cf7SEric Dumazet (s32)(tcp_jiffies32 - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto && 16391b1fc3fdSWei Wang !ca_ops->cong_control) 1640a762a980SDavid S. Miller tcp_cwnd_application_limited(sk); 1641b0f71bd3SFrancis Yan 1642b0f71bd3SFrancis Yan /* The following conditions together indicate the starvation 1643b0f71bd3SFrancis Yan * is caused by insufficient sender buffer: 1644b0f71bd3SFrancis Yan * 1) just sent some data (see tcp_write_xmit) 1645b0f71bd3SFrancis Yan * 2) not cwnd limited (this else condition) 164675c119afSEric Dumazet * 3) no more data to send (tcp_write_queue_empty()) 1647b0f71bd3SFrancis Yan * 4) application is hitting buffer limit (SOCK_NOSPACE) 1648b0f71bd3SFrancis Yan */ 164975c119afSEric Dumazet if (tcp_write_queue_empty(sk) && sk->sk_socket && 1650b0f71bd3SFrancis Yan test_bit(SOCK_NOSPACE, &sk->sk_socket->flags) && 1651b0f71bd3SFrancis Yan (1 << sk->sk_state) & (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) 1652b0f71bd3SFrancis Yan tcp_chrono_start(sk, TCP_CHRONO_SNDBUF_LIMITED); 1653a762a980SDavid S. Miller } 1654a762a980SDavid S. Miller } 1655a762a980SDavid S. Miller 1656d4589926SEric Dumazet /* Minshall's variant of the Nagle send check. */ 1657d4589926SEric Dumazet static bool tcp_minshall_check(const struct tcp_sock *tp) 1658d4589926SEric Dumazet { 1659d4589926SEric Dumazet return after(tp->snd_sml, tp->snd_una) && 1660d4589926SEric Dumazet !after(tp->snd_sml, tp->snd_nxt); 1661d4589926SEric Dumazet } 1662d4589926SEric Dumazet 1663d4589926SEric Dumazet /* Update snd_sml if this skb is under mss 1664d4589926SEric Dumazet * Note that a TSO packet might end with a sub-mss segment 1665d4589926SEric Dumazet * The test is really : 1666d4589926SEric Dumazet * if ((skb->len % mss) != 0) 1667d4589926SEric Dumazet * tp->snd_sml = TCP_SKB_CB(skb)->end_seq; 1668d4589926SEric Dumazet * But we can avoid doing the divide again given we already have 1669d4589926SEric Dumazet * skb_pcount = skb->len / mss_now 16700e3a4803SIlpo Järvinen */ 1671d4589926SEric Dumazet static void tcp_minshall_update(struct tcp_sock *tp, unsigned int mss_now, 1672d4589926SEric Dumazet const struct sk_buff *skb) 1673d4589926SEric Dumazet { 1674d4589926SEric Dumazet if (skb->len < tcp_skb_pcount(skb) * mss_now) 1675d4589926SEric Dumazet tp->snd_sml = TCP_SKB_CB(skb)->end_seq; 1676d4589926SEric Dumazet } 1677d4589926SEric Dumazet 1678d4589926SEric Dumazet /* Return false, if packet can be sent now without violation Nagle's rules: 1679d4589926SEric Dumazet * 1. It is full sized. (provided by caller in %partial bool) 1680d4589926SEric Dumazet * 2. Or it contains FIN. (already checked by caller) 1681d4589926SEric Dumazet * 3. Or TCP_CORK is not set, and TCP_NODELAY is set. 1682d4589926SEric Dumazet * 4. Or TCP_CORK is not set, and all sent packets are ACKed. 1683d4589926SEric Dumazet * With Minshall's modification: all sent small packets are ACKed. 1684d4589926SEric Dumazet */ 1685d4589926SEric Dumazet static bool tcp_nagle_check(bool partial, const struct tcp_sock *tp, 1686cc93fc51SPeter Pan(潘卫平) int nonagle) 1687d4589926SEric Dumazet { 1688d4589926SEric Dumazet return partial && 1689d4589926SEric Dumazet ((nonagle & TCP_NAGLE_CORK) || 1690d4589926SEric Dumazet (!nonagle && tp->packets_out && tcp_minshall_check(tp))); 1691d4589926SEric Dumazet } 1692605ad7f1SEric Dumazet 1693605ad7f1SEric Dumazet /* Return how many segs we'd like on a TSO packet, 1694605ad7f1SEric Dumazet * to send one TSO packet per ms 1695605ad7f1SEric Dumazet */ 1696dcb8c9b4SEric Dumazet static u32 tcp_tso_autosize(const struct sock *sk, unsigned int mss_now, 16971b3878caSNeal Cardwell int min_tso_segs) 1698605ad7f1SEric Dumazet { 1699605ad7f1SEric Dumazet u32 bytes, segs; 1700605ad7f1SEric Dumazet 170176a9ebe8SEric Dumazet bytes = min_t(unsigned long, 170276a9ebe8SEric Dumazet sk->sk_pacing_rate >> sk->sk_pacing_shift, 1703605ad7f1SEric Dumazet sk->sk_gso_max_size - 1 - MAX_TCP_HEADER); 1704605ad7f1SEric Dumazet 1705605ad7f1SEric Dumazet /* Goal is to send at least one packet per ms, 1706605ad7f1SEric Dumazet * not one big TSO packet every 100 ms. 1707605ad7f1SEric Dumazet * This preserves ACK clocking and is consistent 1708605ad7f1SEric Dumazet * with tcp_tso_should_defer() heuristic. 1709605ad7f1SEric Dumazet */ 17101b3878caSNeal Cardwell segs = max_t(u32, bytes / mss_now, min_tso_segs); 1711605ad7f1SEric Dumazet 1712350c9f48SEric Dumazet return segs; 1713605ad7f1SEric Dumazet } 1714605ad7f1SEric Dumazet 1715ed6e7268SNeal Cardwell /* Return the number of segments we want in the skb we are transmitting. 1716ed6e7268SNeal Cardwell * See if congestion control module wants to decide; otherwise, autosize. 1717ed6e7268SNeal Cardwell */ 1718ed6e7268SNeal Cardwell static u32 tcp_tso_segs(struct sock *sk, unsigned int mss_now) 1719ed6e7268SNeal Cardwell { 1720ed6e7268SNeal Cardwell const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops; 1721dcb8c9b4SEric Dumazet u32 min_tso, tso_segs; 1722ed6e7268SNeal Cardwell 1723dcb8c9b4SEric Dumazet min_tso = ca_ops->min_tso_segs ? 1724dcb8c9b4SEric Dumazet ca_ops->min_tso_segs(sk) : 1725dcb8c9b4SEric Dumazet sock_net(sk)->ipv4.sysctl_tcp_min_tso_segs; 1726dcb8c9b4SEric Dumazet 1727dcb8c9b4SEric Dumazet tso_segs = tcp_tso_autosize(sk, mss_now, min_tso); 1728350c9f48SEric Dumazet return min_t(u32, tso_segs, sk->sk_gso_max_segs); 1729ed6e7268SNeal Cardwell } 1730ed6e7268SNeal Cardwell 1731d4589926SEric Dumazet /* Returns the portion of skb which can be sent right away */ 1732d4589926SEric Dumazet static unsigned int tcp_mss_split_point(const struct sock *sk, 1733d4589926SEric Dumazet const struct sk_buff *skb, 1734d4589926SEric Dumazet unsigned int mss_now, 1735d4589926SEric Dumazet unsigned int max_segs, 1736d4589926SEric Dumazet int nonagle) 1737c1b4a7e6SDavid S. Miller { 1738cf533ea5SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 1739d4589926SEric Dumazet u32 partial, needed, window, max_len; 1740c1b4a7e6SDavid S. Miller 174190840defSIlpo Järvinen window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 17421485348dSBen Hutchings max_len = mss_now * max_segs; 17430e3a4803SIlpo Järvinen 17441485348dSBen Hutchings if (likely(max_len <= window && skb != tcp_write_queue_tail(sk))) 17451485348dSBen Hutchings return max_len; 17460e3a4803SIlpo Järvinen 17475ea3a748SIlpo Järvinen needed = min(skb->len, window); 17485ea3a748SIlpo Järvinen 17491485348dSBen Hutchings if (max_len <= needed) 17501485348dSBen Hutchings return max_len; 17510e3a4803SIlpo Järvinen 1752d4589926SEric Dumazet partial = needed % mss_now; 1753d4589926SEric Dumazet /* If last segment is not a full MSS, check if Nagle rules allow us 1754d4589926SEric Dumazet * to include this last segment in this skb. 1755d4589926SEric Dumazet * Otherwise, we'll split the skb at last MSS boundary 1756d4589926SEric Dumazet */ 1757cc93fc51SPeter Pan(潘卫平) if (tcp_nagle_check(partial != 0, tp, nonagle)) 1758d4589926SEric Dumazet return needed - partial; 1759d4589926SEric Dumazet 1760d4589926SEric Dumazet return needed; 1761c1b4a7e6SDavid S. Miller } 1762c1b4a7e6SDavid S. Miller 1763c1b4a7e6SDavid S. Miller /* Can at least one segment of SKB be sent right now, according to the 1764c1b4a7e6SDavid S. Miller * congestion window rules? If so, return how many segments are allowed. 1765c1b4a7e6SDavid S. Miller */ 1766cf533ea5SEric Dumazet static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp, 1767cf533ea5SEric Dumazet const struct sk_buff *skb) 1768c1b4a7e6SDavid S. Miller { 1769d649a7a8SEric Dumazet u32 in_flight, cwnd, halfcwnd; 1770c1b4a7e6SDavid S. Miller 1771c1b4a7e6SDavid S. Miller /* Don't be strict about the congestion window for the final FIN. */ 17724de075e0SEric Dumazet if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) && 17734de075e0SEric Dumazet tcp_skb_pcount(skb) == 1) 1774c1b4a7e6SDavid S. Miller return 1; 1775c1b4a7e6SDavid S. Miller 1776c1b4a7e6SDavid S. Miller in_flight = tcp_packets_in_flight(tp); 1777c1b4a7e6SDavid S. Miller cwnd = tp->snd_cwnd; 1778d649a7a8SEric Dumazet if (in_flight >= cwnd) 1779c1b4a7e6SDavid S. Miller return 0; 1780d649a7a8SEric Dumazet 1781d649a7a8SEric Dumazet /* For better scheduling, ensure we have at least 1782d649a7a8SEric Dumazet * 2 GSO packets in flight. 1783d649a7a8SEric Dumazet */ 1784d649a7a8SEric Dumazet halfcwnd = max(cwnd >> 1, 1U); 1785d649a7a8SEric Dumazet return min(halfcwnd, cwnd - in_flight); 1786c1b4a7e6SDavid S. Miller } 1787c1b4a7e6SDavid S. Miller 1788b595076aSUwe Kleine-König /* Initialize TSO state of a skb. 178967edfef7SAndi Kleen * This must be invoked the first time we consider transmitting 1790c1b4a7e6SDavid S. Miller * SKB onto the wire. 1791c1b4a7e6SDavid S. Miller */ 17925bbb432cSEric Dumazet static int tcp_init_tso_segs(struct sk_buff *skb, unsigned int mss_now) 1793c1b4a7e6SDavid S. Miller { 1794c1b4a7e6SDavid S. Miller int tso_segs = tcp_skb_pcount(skb); 1795c1b4a7e6SDavid S. Miller 1796f8269a49SIlpo Järvinen if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) { 17975bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss_now); 1798c1b4a7e6SDavid S. Miller tso_segs = tcp_skb_pcount(skb); 1799c1b4a7e6SDavid S. Miller } 1800c1b4a7e6SDavid S. Miller return tso_segs; 1801c1b4a7e6SDavid S. Miller } 1802c1b4a7e6SDavid S. Miller 1803c1b4a7e6SDavid S. Miller 1804a2a385d6SEric Dumazet /* Return true if the Nagle test allows this packet to be 1805c1b4a7e6SDavid S. Miller * sent now. 1806c1b4a7e6SDavid S. Miller */ 1807a2a385d6SEric Dumazet static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb, 1808c1b4a7e6SDavid S. Miller unsigned int cur_mss, int nonagle) 1809c1b4a7e6SDavid S. Miller { 1810c1b4a7e6SDavid S. Miller /* Nagle rule does not apply to frames, which sit in the middle of the 1811c1b4a7e6SDavid S. Miller * write_queue (they have no chances to get new data). 1812c1b4a7e6SDavid S. Miller * 1813c1b4a7e6SDavid S. Miller * This is implemented in the callers, where they modify the 'nonagle' 1814c1b4a7e6SDavid S. Miller * argument based upon the location of SKB in the send queue. 1815c1b4a7e6SDavid S. Miller */ 1816c1b4a7e6SDavid S. Miller if (nonagle & TCP_NAGLE_PUSH) 1817a2a385d6SEric Dumazet return true; 1818c1b4a7e6SDavid S. Miller 18199b44190dSYuchung Cheng /* Don't use the nagle rule for urgent data (or for the final FIN). */ 18209b44190dSYuchung Cheng if (tcp_urg_mode(tp) || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) 1821a2a385d6SEric Dumazet return true; 1822c1b4a7e6SDavid S. Miller 1823cc93fc51SPeter Pan(潘卫平) if (!tcp_nagle_check(skb->len < cur_mss, tp, nonagle)) 1824a2a385d6SEric Dumazet return true; 1825c1b4a7e6SDavid S. Miller 1826a2a385d6SEric Dumazet return false; 1827c1b4a7e6SDavid S. Miller } 1828c1b4a7e6SDavid S. Miller 1829c1b4a7e6SDavid S. Miller /* Does at least the first segment of SKB fit into the send window? */ 1830a2a385d6SEric Dumazet static bool tcp_snd_wnd_test(const struct tcp_sock *tp, 1831a2a385d6SEric Dumazet const struct sk_buff *skb, 1832056834d9SIlpo Järvinen unsigned int cur_mss) 1833c1b4a7e6SDavid S. Miller { 1834c1b4a7e6SDavid S. Miller u32 end_seq = TCP_SKB_CB(skb)->end_seq; 1835c1b4a7e6SDavid S. Miller 1836c1b4a7e6SDavid S. Miller if (skb->len > cur_mss) 1837c1b4a7e6SDavid S. Miller end_seq = TCP_SKB_CB(skb)->seq + cur_mss; 1838c1b4a7e6SDavid S. Miller 183990840defSIlpo Järvinen return !after(end_seq, tcp_wnd_end(tp)); 1840c1b4a7e6SDavid S. Miller } 1841c1b4a7e6SDavid S. Miller 1842c1b4a7e6SDavid S. Miller /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet 1843c1b4a7e6SDavid S. Miller * which is put after SKB on the list. It is very much like 1844c1b4a7e6SDavid S. Miller * tcp_fragment() except that it may make several kinds of assumptions 1845c1b4a7e6SDavid S. Miller * in order to speed up the splitting operation. In particular, we 1846c1b4a7e6SDavid S. Miller * know that all the data is in scatter-gather pages, and that the 1847c1b4a7e6SDavid S. Miller * packet has never been sent out before (and thus is not cloned). 1848c1b4a7e6SDavid S. Miller */ 1849*56483341SEric Dumazet static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len, 1850c4ead4c5SEric Dumazet unsigned int mss_now, gfp_t gfp) 1851c1b4a7e6SDavid S. Miller { 1852c1b4a7e6SDavid S. Miller int nlen = skb->len - len; 1853*56483341SEric Dumazet struct sk_buff *buff; 18549ce01461SIlpo Järvinen u8 flags; 1855c1b4a7e6SDavid S. Miller 1856c1b4a7e6SDavid S. Miller /* All of a TSO frame must be composed of paged data. */ 1857c8ac3774SHerbert Xu if (skb->len != skb->data_len) 1858*56483341SEric Dumazet return tcp_fragment(sk, TCP_FRAG_IN_WRITE_QUEUE, 1859*56483341SEric Dumazet skb, len, mss_now, gfp); 1860c1b4a7e6SDavid S. Miller 1861eb934478SEric Dumazet buff = sk_stream_alloc_skb(sk, 0, gfp, true); 186251456b29SIan Morris if (unlikely(!buff)) 1863c1b4a7e6SDavid S. Miller return -ENOMEM; 1864c1b4a7e6SDavid S. Miller 18653ab224beSHideo Aoki sk->sk_wmem_queued += buff->truesize; 18663ab224beSHideo Aoki sk_mem_charge(sk, buff->truesize); 1867b60b49eaSHerbert Xu buff->truesize += nlen; 1868c1b4a7e6SDavid S. Miller skb->truesize -= nlen; 1869c1b4a7e6SDavid S. Miller 1870c1b4a7e6SDavid S. Miller /* Correct the sequence numbers. */ 1871c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; 1872c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; 1873c1b4a7e6SDavid S. Miller TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; 1874c1b4a7e6SDavid S. Miller 1875c1b4a7e6SDavid S. Miller /* PSH and FIN should only be set in the second packet. */ 18764de075e0SEric Dumazet flags = TCP_SKB_CB(skb)->tcp_flags; 18774de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); 18784de075e0SEric Dumazet TCP_SKB_CB(buff)->tcp_flags = flags; 1879c1b4a7e6SDavid S. Miller 1880c1b4a7e6SDavid S. Miller /* This packet was never sent out yet, so no SACK bits. */ 1881c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->sacked = 0; 1882c1b4a7e6SDavid S. Miller 1883a166140eSMartin KaFai Lau tcp_skb_fragment_eor(skb, buff); 1884a166140eSMartin KaFai Lau 188598be9b12SEric Dumazet buff->ip_summed = CHECKSUM_PARTIAL; 1886c1b4a7e6SDavid S. Miller skb_split(skb, buff, len); 1887490cc7d0SWillem de Bruijn tcp_fragment_tstamp(skb, buff); 1888c1b4a7e6SDavid S. Miller 1889c1b4a7e6SDavid S. Miller /* Fix up tso_factor for both original and new SKB. */ 18905bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss_now); 18915bbb432cSEric Dumazet tcp_set_skb_tso_segs(buff, mss_now); 1892c1b4a7e6SDavid S. Miller 1893c1b4a7e6SDavid S. Miller /* Link BUFF into the send queue. */ 1894f4a775d1SEric Dumazet __skb_header_release(buff); 1895*56483341SEric Dumazet tcp_insert_write_queue_after(skb, buff, sk, TCP_FRAG_IN_WRITE_QUEUE); 1896c1b4a7e6SDavid S. Miller 1897c1b4a7e6SDavid S. Miller return 0; 1898c1b4a7e6SDavid S. Miller } 1899c1b4a7e6SDavid S. Miller 1900c1b4a7e6SDavid S. Miller /* Try to defer sending, if possible, in order to minimize the amount 1901c1b4a7e6SDavid S. Miller * of TSO splitting we do. View it as a kind of TSO Nagle test. 1902c1b4a7e6SDavid S. Miller * 1903c1b4a7e6SDavid S. Miller * This algorithm is from John Heffner. 1904c1b4a7e6SDavid S. Miller */ 1905ca8a2263SNeal Cardwell static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb, 1906f9bfe4e6SEric Dumazet bool *is_cwnd_limited, 1907f9bfe4e6SEric Dumazet bool *is_rwnd_limited, 1908f9bfe4e6SEric Dumazet u32 max_segs) 1909c1b4a7e6SDavid S. Miller { 19106687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 1911f1c6ea38SEric Dumazet u32 send_win, cong_win, limit, in_flight; 191250c8339eSEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 191350c8339eSEric Dumazet struct sk_buff *head; 1914ad9f4f50SEric Dumazet int win_divisor; 1915f1c6ea38SEric Dumazet s64 delta; 1916c1b4a7e6SDavid S. Miller 191799d7662aSEric Dumazet if (icsk->icsk_ca_state >= TCP_CA_Recovery) 1918ae8064acSJohn Heffner goto send_now; 1919ae8064acSJohn Heffner 19205f852eb5SEric Dumazet /* Avoid bursty behavior by allowing defer 1921a682850aSEric Dumazet * only if the last write was recent (1 ms). 1922a682850aSEric Dumazet * Note that tp->tcp_wstamp_ns can be in the future if we have 1923a682850aSEric Dumazet * packets waiting in a qdisc or device for EDT delivery. 19245f852eb5SEric Dumazet */ 1925a682850aSEric Dumazet delta = tp->tcp_clock_cache - tp->tcp_wstamp_ns - NSEC_PER_MSEC; 1926a682850aSEric Dumazet if (delta > 0) 1927ae8064acSJohn Heffner goto send_now; 1928908a75c1SDavid S. Miller 1929c1b4a7e6SDavid S. Miller in_flight = tcp_packets_in_flight(tp); 1930c1b4a7e6SDavid S. Miller 1931c8c9aeb5SStefano Brivio BUG_ON(tcp_skb_pcount(skb) <= 1); 1932c8c9aeb5SStefano Brivio BUG_ON(tp->snd_cwnd <= in_flight); 1933c1b4a7e6SDavid S. Miller 193490840defSIlpo Järvinen send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 1935c1b4a7e6SDavid S. Miller 1936c1b4a7e6SDavid S. Miller /* From in_flight test above, we know that cwnd > in_flight. */ 1937c1b4a7e6SDavid S. Miller cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache; 1938c1b4a7e6SDavid S. Miller 1939c1b4a7e6SDavid S. Miller limit = min(send_win, cong_win); 1940c1b4a7e6SDavid S. Miller 1941ba244fe9SDavid S. Miller /* If a full-sized TSO skb can be sent, do it. */ 1942605ad7f1SEric Dumazet if (limit >= max_segs * tp->mss_cache) 1943ae8064acSJohn Heffner goto send_now; 1944ba244fe9SDavid S. Miller 194562ad2761SIlpo Järvinen /* Middle in queue won't get any more data, full sendable already? */ 194662ad2761SIlpo Järvinen if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len)) 194762ad2761SIlpo Järvinen goto send_now; 194862ad2761SIlpo Järvinen 19495bbcc0f5SLinus Torvalds win_divisor = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_tso_win_divisor); 1950ad9f4f50SEric Dumazet if (win_divisor) { 1951c1b4a7e6SDavid S. Miller u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache); 1952c1b4a7e6SDavid S. Miller 1953c1b4a7e6SDavid S. Miller /* If at least some fraction of a window is available, 1954c1b4a7e6SDavid S. Miller * just use it. 1955c1b4a7e6SDavid S. Miller */ 1956ad9f4f50SEric Dumazet chunk /= win_divisor; 1957c1b4a7e6SDavid S. Miller if (limit >= chunk) 1958ae8064acSJohn Heffner goto send_now; 1959c1b4a7e6SDavid S. Miller } else { 1960c1b4a7e6SDavid S. Miller /* Different approach, try not to defer past a single 1961c1b4a7e6SDavid S. Miller * ACK. Receiver should ACK every other full sized 1962c1b4a7e6SDavid S. Miller * frame, so if we have space for more than 3 frames 1963c1b4a7e6SDavid S. Miller * then send now. 1964c1b4a7e6SDavid S. Miller */ 19656b5a5c0dSNeal Cardwell if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache) 1966ae8064acSJohn Heffner goto send_now; 1967c1b4a7e6SDavid S. Miller } 1968c1b4a7e6SDavid S. Miller 196975c119afSEric Dumazet /* TODO : use tsorted_sent_queue ? */ 197075c119afSEric Dumazet head = tcp_rtx_queue_head(sk); 197175c119afSEric Dumazet if (!head) 197275c119afSEric Dumazet goto send_now; 1973f1c6ea38SEric Dumazet delta = tp->tcp_clock_cache - head->tstamp; 197450c8339eSEric Dumazet /* If next ACK is likely to come too late (half srtt), do not defer */ 1975f1c6ea38SEric Dumazet if ((s64)(delta - (u64)NSEC_PER_USEC * (tp->srtt_us >> 4)) < 0) 197650c8339eSEric Dumazet goto send_now; 197750c8339eSEric Dumazet 1978f9bfe4e6SEric Dumazet /* Ok, it looks like it is advisable to defer. 1979f9bfe4e6SEric Dumazet * Three cases are tracked : 1980f9bfe4e6SEric Dumazet * 1) We are cwnd-limited 1981f9bfe4e6SEric Dumazet * 2) We are rwnd-limited 1982f9bfe4e6SEric Dumazet * 3) We are application limited. 1983f9bfe4e6SEric Dumazet */ 1984f9bfe4e6SEric Dumazet if (cong_win < send_win) { 1985f9bfe4e6SEric Dumazet if (cong_win <= skb->len) { 1986ca8a2263SNeal Cardwell *is_cwnd_limited = true; 1987f9bfe4e6SEric Dumazet return true; 1988f9bfe4e6SEric Dumazet } 1989f9bfe4e6SEric Dumazet } else { 1990f9bfe4e6SEric Dumazet if (send_win <= skb->len) { 1991f9bfe4e6SEric Dumazet *is_rwnd_limited = true; 1992f9bfe4e6SEric Dumazet return true; 1993f9bfe4e6SEric Dumazet } 1994f9bfe4e6SEric Dumazet } 1995f9bfe4e6SEric Dumazet 1996f9bfe4e6SEric Dumazet /* If this packet won't get more data, do not wait. */ 1997d8ed257fSEric Dumazet if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) || 1998d8ed257fSEric Dumazet TCP_SKB_CB(skb)->eor) 1999f9bfe4e6SEric Dumazet goto send_now; 2000ca8a2263SNeal Cardwell 2001a2a385d6SEric Dumazet return true; 2002ae8064acSJohn Heffner 2003ae8064acSJohn Heffner send_now: 2004a2a385d6SEric Dumazet return false; 2005c1b4a7e6SDavid S. Miller } 2006c1b4a7e6SDavid S. Miller 200705cbc0dbSFan Du static inline void tcp_mtu_check_reprobe(struct sock *sk) 200805cbc0dbSFan Du { 200905cbc0dbSFan Du struct inet_connection_sock *icsk = inet_csk(sk); 201005cbc0dbSFan Du struct tcp_sock *tp = tcp_sk(sk); 201105cbc0dbSFan Du struct net *net = sock_net(sk); 201205cbc0dbSFan Du u32 interval; 201305cbc0dbSFan Du s32 delta; 201405cbc0dbSFan Du 201505cbc0dbSFan Du interval = net->ipv4.sysctl_tcp_probe_interval; 2016c74df29aSEric Dumazet delta = tcp_jiffies32 - icsk->icsk_mtup.probe_timestamp; 201705cbc0dbSFan Du if (unlikely(delta >= interval * HZ)) { 201805cbc0dbSFan Du int mss = tcp_current_mss(sk); 201905cbc0dbSFan Du 202005cbc0dbSFan Du /* Update current search range */ 202105cbc0dbSFan Du icsk->icsk_mtup.probe_size = 0; 202205cbc0dbSFan Du icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + 202305cbc0dbSFan Du sizeof(struct tcphdr) + 202405cbc0dbSFan Du icsk->icsk_af_ops->net_header_len; 202505cbc0dbSFan Du icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss); 202605cbc0dbSFan Du 202705cbc0dbSFan Du /* Update probe time stamp */ 2028c74df29aSEric Dumazet icsk->icsk_mtup.probe_timestamp = tcp_jiffies32; 202905cbc0dbSFan Du } 203005cbc0dbSFan Du } 203105cbc0dbSFan Du 2032808cf9e3SIlya Lesokhin static bool tcp_can_coalesce_send_queue_head(struct sock *sk, int len) 2033808cf9e3SIlya Lesokhin { 2034808cf9e3SIlya Lesokhin struct sk_buff *skb, *next; 2035808cf9e3SIlya Lesokhin 2036808cf9e3SIlya Lesokhin skb = tcp_send_head(sk); 2037808cf9e3SIlya Lesokhin tcp_for_write_queue_from_safe(skb, next, sk) { 2038808cf9e3SIlya Lesokhin if (len <= skb->len) 2039808cf9e3SIlya Lesokhin break; 2040808cf9e3SIlya Lesokhin 2041808cf9e3SIlya Lesokhin if (unlikely(TCP_SKB_CB(skb)->eor)) 2042808cf9e3SIlya Lesokhin return false; 2043808cf9e3SIlya Lesokhin 2044808cf9e3SIlya Lesokhin len -= skb->len; 2045808cf9e3SIlya Lesokhin } 2046808cf9e3SIlya Lesokhin 2047808cf9e3SIlya Lesokhin return true; 2048808cf9e3SIlya Lesokhin } 2049808cf9e3SIlya Lesokhin 20505d424d5aSJohn Heffner /* Create a new MTU probe if we are ready. 205167edfef7SAndi Kleen * MTU probe is regularly attempting to increase the path MTU by 205267edfef7SAndi Kleen * deliberately sending larger packets. This discovers routing 205367edfef7SAndi Kleen * changes resulting in larger path MTUs. 205467edfef7SAndi Kleen * 20555d424d5aSJohn Heffner * Returns 0 if we should wait to probe (no cwnd available), 20565d424d5aSJohn Heffner * 1 if a probe was sent, 2057056834d9SIlpo Järvinen * -1 otherwise 2058056834d9SIlpo Järvinen */ 20595d424d5aSJohn Heffner static int tcp_mtu_probe(struct sock *sk) 20605d424d5aSJohn Heffner { 20615d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 206212a59abcSEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 20635d424d5aSJohn Heffner struct sk_buff *skb, *nskb, *next; 20646b58e0a5SFan Du struct net *net = sock_net(sk); 20655d424d5aSJohn Heffner int probe_size; 206691cc17c0SIlpo Järvinen int size_needed; 206712a59abcSEric Dumazet int copy, len; 20685d424d5aSJohn Heffner int mss_now; 20696b58e0a5SFan Du int interval; 20705d424d5aSJohn Heffner 20715d424d5aSJohn Heffner /* Not currently probing/verifying, 20725d424d5aSJohn Heffner * not in recovery, 20735d424d5aSJohn Heffner * have enough cwnd, and 207412a59abcSEric Dumazet * not SACKing (the variable headers throw things off) 207512a59abcSEric Dumazet */ 207612a59abcSEric Dumazet if (likely(!icsk->icsk_mtup.enabled || 20775d424d5aSJohn Heffner icsk->icsk_mtup.probe_size || 20785d424d5aSJohn Heffner inet_csk(sk)->icsk_ca_state != TCP_CA_Open || 20795d424d5aSJohn Heffner tp->snd_cwnd < 11 || 208012a59abcSEric Dumazet tp->rx_opt.num_sacks || tp->rx_opt.dsack)) 20815d424d5aSJohn Heffner return -1; 20825d424d5aSJohn Heffner 20836b58e0a5SFan Du /* Use binary search for probe_size between tcp_mss_base, 20846b58e0a5SFan Du * and current mss_clamp. if (search_high - search_low) 20856b58e0a5SFan Du * smaller than a threshold, backoff from probing. 20866b58e0a5SFan Du */ 20870c54b85fSIlpo Järvinen mss_now = tcp_current_mss(sk); 20886b58e0a5SFan Du probe_size = tcp_mtu_to_mss(sk, (icsk->icsk_mtup.search_high + 20896b58e0a5SFan Du icsk->icsk_mtup.search_low) >> 1); 209091cc17c0SIlpo Järvinen size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache; 20916b58e0a5SFan Du interval = icsk->icsk_mtup.search_high - icsk->icsk_mtup.search_low; 209205cbc0dbSFan Du /* When misfortune happens, we are reprobing actively, 209305cbc0dbSFan Du * and then reprobe timer has expired. We stick with current 209405cbc0dbSFan Du * probing process by not resetting search range to its orignal. 209505cbc0dbSFan Du */ 20966b58e0a5SFan Du if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high) || 209705cbc0dbSFan Du interval < net->ipv4.sysctl_tcp_probe_threshold) { 209805cbc0dbSFan Du /* Check whether enough time has elaplased for 209905cbc0dbSFan Du * another round of probing. 210005cbc0dbSFan Du */ 210105cbc0dbSFan Du tcp_mtu_check_reprobe(sk); 21025d424d5aSJohn Heffner return -1; 21035d424d5aSJohn Heffner } 21045d424d5aSJohn Heffner 21055d424d5aSJohn Heffner /* Have enough data in the send queue to probe? */ 21067f9c33e5SIlpo Järvinen if (tp->write_seq - tp->snd_nxt < size_needed) 21075d424d5aSJohn Heffner return -1; 21085d424d5aSJohn Heffner 210991cc17c0SIlpo Järvinen if (tp->snd_wnd < size_needed) 21105d424d5aSJohn Heffner return -1; 211190840defSIlpo Järvinen if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp))) 21125d424d5aSJohn Heffner return 0; 21135d424d5aSJohn Heffner 2114d67c58e9SIlpo Järvinen /* Do we need to wait to drain cwnd? With none in flight, don't stall */ 2115d67c58e9SIlpo Järvinen if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) { 2116d67c58e9SIlpo Järvinen if (!tcp_packets_in_flight(tp)) 21175d424d5aSJohn Heffner return -1; 21185d424d5aSJohn Heffner else 21195d424d5aSJohn Heffner return 0; 21205d424d5aSJohn Heffner } 21215d424d5aSJohn Heffner 2122808cf9e3SIlya Lesokhin if (!tcp_can_coalesce_send_queue_head(sk, probe_size)) 2123808cf9e3SIlya Lesokhin return -1; 2124808cf9e3SIlya Lesokhin 21255d424d5aSJohn Heffner /* We're allowed to probe. Build it now. */ 2126eb934478SEric Dumazet nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC, false); 212751456b29SIan Morris if (!nskb) 21285d424d5aSJohn Heffner return -1; 21293ab224beSHideo Aoki sk->sk_wmem_queued += nskb->truesize; 21303ab224beSHideo Aoki sk_mem_charge(sk, nskb->truesize); 21315d424d5aSJohn Heffner 2132fe067e8aSDavid S. Miller skb = tcp_send_head(sk); 21335d424d5aSJohn Heffner 21345d424d5aSJohn Heffner TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq; 21355d424d5aSJohn Heffner TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size; 21364de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK; 21375d424d5aSJohn Heffner TCP_SKB_CB(nskb)->sacked = 0; 21385d424d5aSJohn Heffner nskb->csum = 0; 213998be9b12SEric Dumazet nskb->ip_summed = CHECKSUM_PARTIAL; 21405d424d5aSJohn Heffner 214150c4817eSIlpo Järvinen tcp_insert_write_queue_before(nskb, skb, sk); 21422b7cda9cSEric Dumazet tcp_highest_sack_replace(sk, skb, nskb); 214350c4817eSIlpo Järvinen 21445d424d5aSJohn Heffner len = 0; 2145234b6860SIlpo Järvinen tcp_for_write_queue_from_safe(skb, next, sk) { 21465d424d5aSJohn Heffner copy = min_t(int, skb->len, probe_size - len); 21475d424d5aSJohn Heffner skb_copy_bits(skb, 0, skb_put(nskb, copy), copy); 21485d424d5aSJohn Heffner 21495d424d5aSJohn Heffner if (skb->len <= copy) { 21505d424d5aSJohn Heffner /* We've eaten all the data from this skb. 21515d424d5aSJohn Heffner * Throw it away. */ 21524de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags; 2153808cf9e3SIlya Lesokhin /* If this is the last SKB we copy and eor is set 2154808cf9e3SIlya Lesokhin * we need to propagate it to the new skb. 2155808cf9e3SIlya Lesokhin */ 2156808cf9e3SIlya Lesokhin TCP_SKB_CB(nskb)->eor = TCP_SKB_CB(skb)->eor; 2157fe067e8aSDavid S. Miller tcp_unlink_write_queue(skb, sk); 21583ab224beSHideo Aoki sk_wmem_free_skb(sk, skb); 21595d424d5aSJohn Heffner } else { 21604de075e0SEric Dumazet TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags & 2161a3433f35SChangli Gao ~(TCPHDR_FIN|TCPHDR_PSH); 21625d424d5aSJohn Heffner if (!skb_shinfo(skb)->nr_frags) { 21635d424d5aSJohn Heffner skb_pull(skb, copy); 21645d424d5aSJohn Heffner } else { 21655d424d5aSJohn Heffner __pskb_trim_head(skb, copy); 21665bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss_now); 21675d424d5aSJohn Heffner } 21685d424d5aSJohn Heffner TCP_SKB_CB(skb)->seq += copy; 21695d424d5aSJohn Heffner } 21705d424d5aSJohn Heffner 21715d424d5aSJohn Heffner len += copy; 2172234b6860SIlpo Järvinen 2173234b6860SIlpo Järvinen if (len >= probe_size) 2174234b6860SIlpo Järvinen break; 21755d424d5aSJohn Heffner } 21765bbb432cSEric Dumazet tcp_init_tso_segs(nskb, nskb->len); 21775d424d5aSJohn Heffner 21785d424d5aSJohn Heffner /* We're ready to send. If this fails, the probe will 21797faee5c0SEric Dumazet * be resegmented into mss-sized pieces by tcp_write_xmit(). 21807faee5c0SEric Dumazet */ 21815d424d5aSJohn Heffner if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) { 21825d424d5aSJohn Heffner /* Decrement cwnd here because we are sending 21835d424d5aSJohn Heffner * effectively two packets. */ 21845d424d5aSJohn Heffner tp->snd_cwnd--; 218566f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, nskb); 21865d424d5aSJohn Heffner 21875d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len); 21880e7b1368SJohn Heffner tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq; 21890e7b1368SJohn Heffner tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq; 21905d424d5aSJohn Heffner 21915d424d5aSJohn Heffner return 1; 21925d424d5aSJohn Heffner } 21935d424d5aSJohn Heffner 21945d424d5aSJohn Heffner return -1; 21955d424d5aSJohn Heffner } 21965d424d5aSJohn Heffner 2197864e5c09SEric Dumazet static bool tcp_pacing_check(struct sock *sk) 2198218af599SEric Dumazet { 2199864e5c09SEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 2200864e5c09SEric Dumazet 2201864e5c09SEric Dumazet if (!tcp_needs_internal_pacing(sk)) 2202864e5c09SEric Dumazet return false; 2203864e5c09SEric Dumazet 2204864e5c09SEric Dumazet if (tp->tcp_wstamp_ns <= tp->tcp_clock_cache) 2205864e5c09SEric Dumazet return false; 2206864e5c09SEric Dumazet 2207864e5c09SEric Dumazet if (!hrtimer_is_queued(&tp->pacing_timer)) { 2208864e5c09SEric Dumazet hrtimer_start(&tp->pacing_timer, 2209864e5c09SEric Dumazet ns_to_ktime(tp->tcp_wstamp_ns), 2210864e5c09SEric Dumazet HRTIMER_MODE_ABS_PINNED_SOFT); 2211864e5c09SEric Dumazet sock_hold(sk); 2212864e5c09SEric Dumazet } 2213864e5c09SEric Dumazet return true; 2214218af599SEric Dumazet } 2215218af599SEric Dumazet 2216f9616c35SEric Dumazet /* TCP Small Queues : 2217f9616c35SEric Dumazet * Control number of packets in qdisc/devices to two packets / or ~1 ms. 2218f9616c35SEric Dumazet * (These limits are doubled for retransmits) 2219f9616c35SEric Dumazet * This allows for : 2220f9616c35SEric Dumazet * - better RTT estimation and ACK scheduling 2221f9616c35SEric Dumazet * - faster recovery 2222f9616c35SEric Dumazet * - high rates 2223f9616c35SEric Dumazet * Alas, some drivers / subsystems require a fair amount 2224f9616c35SEric Dumazet * of queued bytes to ensure line rate. 2225f9616c35SEric Dumazet * One example is wifi aggregation (802.11 AMPDU) 2226f9616c35SEric Dumazet */ 2227f9616c35SEric Dumazet static bool tcp_small_queue_check(struct sock *sk, const struct sk_buff *skb, 2228f9616c35SEric Dumazet unsigned int factor) 2229f9616c35SEric Dumazet { 223076a9ebe8SEric Dumazet unsigned long limit; 2231f9616c35SEric Dumazet 223276a9ebe8SEric Dumazet limit = max_t(unsigned long, 223376a9ebe8SEric Dumazet 2 * skb->truesize, 223476a9ebe8SEric Dumazet sk->sk_pacing_rate >> sk->sk_pacing_shift); 2235c73e5807SEric Dumazet if (sk->sk_pacing_status == SK_PACING_NONE) 223676a9ebe8SEric Dumazet limit = min_t(unsigned long, limit, 22379184d8bbSEric Dumazet sock_net(sk)->ipv4.sysctl_tcp_limit_output_bytes); 2238f9616c35SEric Dumazet limit <<= factor; 2239f9616c35SEric Dumazet 224014afee4bSReshetova, Elena if (refcount_read(&sk->sk_wmem_alloc) > limit) { 224175c119afSEric Dumazet /* Always send skb if rtx queue is empty. 224275eefc6cSEric Dumazet * No need to wait for TX completion to call us back, 224375eefc6cSEric Dumazet * after softirq/tasklet schedule. 224475eefc6cSEric Dumazet * This helps when TX completions are delayed too much. 224575eefc6cSEric Dumazet */ 224675c119afSEric Dumazet if (tcp_rtx_queue_empty(sk)) 224775eefc6cSEric Dumazet return false; 224875eefc6cSEric Dumazet 22497aa5470cSEric Dumazet set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags); 2250f9616c35SEric Dumazet /* It is possible TX completion already happened 2251f9616c35SEric Dumazet * before we set TSQ_THROTTLED, so we must 2252f9616c35SEric Dumazet * test again the condition. 2253f9616c35SEric Dumazet */ 2254f9616c35SEric Dumazet smp_mb__after_atomic(); 225514afee4bSReshetova, Elena if (refcount_read(&sk->sk_wmem_alloc) > limit) 2256f9616c35SEric Dumazet return true; 2257f9616c35SEric Dumazet } 2258f9616c35SEric Dumazet return false; 2259f9616c35SEric Dumazet } 2260f9616c35SEric Dumazet 226105b055e8SFrancis Yan static void tcp_chrono_set(struct tcp_sock *tp, const enum tcp_chrono new) 226205b055e8SFrancis Yan { 2263628174ccSEric Dumazet const u32 now = tcp_jiffies32; 2264efe967cdSArnd Bergmann enum tcp_chrono old = tp->chrono_type; 226505b055e8SFrancis Yan 2266efe967cdSArnd Bergmann if (old > TCP_CHRONO_UNSPEC) 2267efe967cdSArnd Bergmann tp->chrono_stat[old - 1] += now - tp->chrono_start; 226805b055e8SFrancis Yan tp->chrono_start = now; 226905b055e8SFrancis Yan tp->chrono_type = new; 227005b055e8SFrancis Yan } 227105b055e8SFrancis Yan 227205b055e8SFrancis Yan void tcp_chrono_start(struct sock *sk, const enum tcp_chrono type) 227305b055e8SFrancis Yan { 227405b055e8SFrancis Yan struct tcp_sock *tp = tcp_sk(sk); 227505b055e8SFrancis Yan 227605b055e8SFrancis Yan /* If there are multiple conditions worthy of tracking in a 22770f87230dSFrancis Yan * chronograph then the highest priority enum takes precedence 22780f87230dSFrancis Yan * over the other conditions. So that if something "more interesting" 227905b055e8SFrancis Yan * starts happening, stop the previous chrono and start a new one. 228005b055e8SFrancis Yan */ 228105b055e8SFrancis Yan if (type > tp->chrono_type) 228205b055e8SFrancis Yan tcp_chrono_set(tp, type); 228305b055e8SFrancis Yan } 228405b055e8SFrancis Yan 228505b055e8SFrancis Yan void tcp_chrono_stop(struct sock *sk, const enum tcp_chrono type) 228605b055e8SFrancis Yan { 228705b055e8SFrancis Yan struct tcp_sock *tp = tcp_sk(sk); 228805b055e8SFrancis Yan 22890f87230dSFrancis Yan 22900f87230dSFrancis Yan /* There are multiple conditions worthy of tracking in a 22910f87230dSFrancis Yan * chronograph, so that the highest priority enum takes 22920f87230dSFrancis Yan * precedence over the other conditions (see tcp_chrono_start). 22930f87230dSFrancis Yan * If a condition stops, we only stop chrono tracking if 22940f87230dSFrancis Yan * it's the "most interesting" or current chrono we are 22950f87230dSFrancis Yan * tracking and starts busy chrono if we have pending data. 22960f87230dSFrancis Yan */ 229775c119afSEric Dumazet if (tcp_rtx_and_write_queues_empty(sk)) 229805b055e8SFrancis Yan tcp_chrono_set(tp, TCP_CHRONO_UNSPEC); 22990f87230dSFrancis Yan else if (type == tp->chrono_type) 23000f87230dSFrancis Yan tcp_chrono_set(tp, TCP_CHRONO_BUSY); 230105b055e8SFrancis Yan } 230205b055e8SFrancis Yan 23031da177e4SLinus Torvalds /* This routine writes packets to the network. It advances the 23041da177e4SLinus Torvalds * send_head. This happens as incoming acks open up the remote 23051da177e4SLinus Torvalds * window for us. 23061da177e4SLinus Torvalds * 2307f8269a49SIlpo Järvinen * LARGESEND note: !tcp_urg_mode is overkill, only frames between 2308f8269a49SIlpo Järvinen * snd_up-64k-mss .. snd_up cannot be large. However, taking into 2309f8269a49SIlpo Järvinen * account rare use of URG, this is not a big flaw. 2310f8269a49SIlpo Järvinen * 23116ba8a3b1SNandita Dukkipati * Send at most one packet when push_one > 0. Temporarily ignore 23126ba8a3b1SNandita Dukkipati * cwnd limit to force at most one packet out when push_one == 2. 23136ba8a3b1SNandita Dukkipati 2314a2a385d6SEric Dumazet * Returns true, if no segments are in flight and we have queued segments, 2315a2a385d6SEric Dumazet * but cannot send anything now because of SWS or another problem. 23161da177e4SLinus Torvalds */ 2317a2a385d6SEric Dumazet static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle, 2318d5dd9175SIlpo Järvinen int push_one, gfp_t gfp) 23191da177e4SLinus Torvalds { 23201da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 232192df7b51SDavid S. Miller struct sk_buff *skb; 2322c1b4a7e6SDavid S. Miller unsigned int tso_segs, sent_pkts; 2323c1b4a7e6SDavid S. Miller int cwnd_quota; 23245d424d5aSJohn Heffner int result; 23255615f886SFrancis Yan bool is_cwnd_limited = false, is_rwnd_limited = false; 2326605ad7f1SEric Dumazet u32 max_segs; 23271da177e4SLinus Torvalds 2328c1b4a7e6SDavid S. Miller sent_pkts = 0; 23295d424d5aSJohn Heffner 2330ee1836aeSEric Dumazet tcp_mstamp_refresh(tp); 2331d5dd9175SIlpo Järvinen if (!push_one) { 23325d424d5aSJohn Heffner /* Do MTU probing. */ 2333d5dd9175SIlpo Järvinen result = tcp_mtu_probe(sk); 2334d5dd9175SIlpo Järvinen if (!result) { 2335a2a385d6SEric Dumazet return false; 23365d424d5aSJohn Heffner } else if (result > 0) { 23375d424d5aSJohn Heffner sent_pkts = 1; 23385d424d5aSJohn Heffner } 2339d5dd9175SIlpo Järvinen } 23405d424d5aSJohn Heffner 2341ed6e7268SNeal Cardwell max_segs = tcp_tso_segs(sk, mss_now); 2342fe067e8aSDavid S. Miller while ((skb = tcp_send_head(sk))) { 2343c8ac3774SHerbert Xu unsigned int limit; 2344c8ac3774SHerbert Xu 234579861919SEric Dumazet if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE) { 234679861919SEric Dumazet /* "skb_mstamp_ns" is used as a start point for the retransmit timer */ 234779861919SEric Dumazet skb->skb_mstamp_ns = tp->tcp_wstamp_ns = tp->tcp_clock_cache; 234879861919SEric Dumazet list_move_tail(&skb->tcp_tsorted_anchor, &tp->tsorted_sent_queue); 2349bf50b606SEric Dumazet tcp_init_tso_segs(skb, mss_now); 235079861919SEric Dumazet goto repair; /* Skip network transmission */ 235179861919SEric Dumazet } 235279861919SEric Dumazet 2353218af599SEric Dumazet if (tcp_pacing_check(sk)) 2354218af599SEric Dumazet break; 2355218af599SEric Dumazet 23565bbb432cSEric Dumazet tso_segs = tcp_init_tso_segs(skb, mss_now); 2357c1b4a7e6SDavid S. Miller BUG_ON(!tso_segs); 2358c1b4a7e6SDavid S. Miller 2359b68e9f85SHerbert Xu cwnd_quota = tcp_cwnd_test(tp, skb); 23606ba8a3b1SNandita Dukkipati if (!cwnd_quota) { 23616ba8a3b1SNandita Dukkipati if (push_one == 2) 23626ba8a3b1SNandita Dukkipati /* Force out a loss probe pkt. */ 23636ba8a3b1SNandita Dukkipati cwnd_quota = 1; 23646ba8a3b1SNandita Dukkipati else 2365b68e9f85SHerbert Xu break; 23666ba8a3b1SNandita Dukkipati } 2367b68e9f85SHerbert Xu 23685615f886SFrancis Yan if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now))) { 23695615f886SFrancis Yan is_rwnd_limited = true; 2370b68e9f85SHerbert Xu break; 23715615f886SFrancis Yan } 2372b68e9f85SHerbert Xu 2373d6a4e26aSEric Dumazet if (tso_segs == 1) { 2374aa93466bSDavid S. Miller if (unlikely(!tcp_nagle_test(tp, skb, mss_now, 2375aa93466bSDavid S. Miller (tcp_skb_is_last(sk, skb) ? 2376aa93466bSDavid S. Miller nonagle : TCP_NAGLE_PUSH)))) 2377aa93466bSDavid S. Miller break; 2378c1b4a7e6SDavid S. Miller } else { 2379ca8a2263SNeal Cardwell if (!push_one && 2380605ad7f1SEric Dumazet tcp_tso_should_defer(sk, skb, &is_cwnd_limited, 2381f9bfe4e6SEric Dumazet &is_rwnd_limited, max_segs)) 2382aa93466bSDavid S. Miller break; 2383c1b4a7e6SDavid S. Miller } 2384aa93466bSDavid S. Miller 2385605ad7f1SEric Dumazet limit = mss_now; 2386d6a4e26aSEric Dumazet if (tso_segs > 1 && !tcp_urg_mode(tp)) 2387605ad7f1SEric Dumazet limit = tcp_mss_split_point(sk, skb, mss_now, 2388605ad7f1SEric Dumazet min_t(unsigned int, 2389605ad7f1SEric Dumazet cwnd_quota, 2390605ad7f1SEric Dumazet max_segs), 2391605ad7f1SEric Dumazet nonagle); 2392605ad7f1SEric Dumazet 2393605ad7f1SEric Dumazet if (skb->len > limit && 2394*56483341SEric Dumazet unlikely(tso_fragment(sk, skb, limit, mss_now, gfp))) 2395605ad7f1SEric Dumazet break; 2396605ad7f1SEric Dumazet 2397f9616c35SEric Dumazet if (tcp_small_queue_check(sk, skb, 0)) 239846d3ceabSEric Dumazet break; 2399c9eeec26SEric Dumazet 2400d5dd9175SIlpo Järvinen if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp))) 24011da177e4SLinus Torvalds break; 24021da177e4SLinus Torvalds 2403ec342325SAndrew Vagin repair: 24041da177e4SLinus Torvalds /* Advance the send_head. This one is sent out. 24051da177e4SLinus Torvalds * This call will increment packets_out. 24061da177e4SLinus Torvalds */ 240766f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, skb); 24081da177e4SLinus Torvalds 24091da177e4SLinus Torvalds tcp_minshall_update(tp, mss_now, skb); 2410a262f0cdSNandita Dukkipati sent_pkts += tcp_skb_pcount(skb); 2411d5dd9175SIlpo Järvinen 2412d5dd9175SIlpo Järvinen if (push_one) 2413d5dd9175SIlpo Järvinen break; 24141da177e4SLinus Torvalds } 24151da177e4SLinus Torvalds 24165615f886SFrancis Yan if (is_rwnd_limited) 24175615f886SFrancis Yan tcp_chrono_start(sk, TCP_CHRONO_RWND_LIMITED); 24185615f886SFrancis Yan else 24195615f886SFrancis Yan tcp_chrono_stop(sk, TCP_CHRONO_RWND_LIMITED); 24205615f886SFrancis Yan 2421aa93466bSDavid S. Miller if (likely(sent_pkts)) { 2422684bad11SYuchung Cheng if (tcp_in_cwnd_reduction(sk)) 2423684bad11SYuchung Cheng tp->prr_out += sent_pkts; 24246ba8a3b1SNandita Dukkipati 24256ba8a3b1SNandita Dukkipati /* Send one loss probe per tail loss episode. */ 24266ba8a3b1SNandita Dukkipati if (push_one != 2) 2427ed66dfafSNeal Cardwell tcp_schedule_loss_probe(sk, false); 2428d2e1339fSBendik Rønning Opstad is_cwnd_limited |= (tcp_packets_in_flight(tp) >= tp->snd_cwnd); 2429ca8a2263SNeal Cardwell tcp_cwnd_validate(sk, is_cwnd_limited); 2430a2a385d6SEric Dumazet return false; 24311da177e4SLinus Torvalds } 243275c119afSEric Dumazet return !tp->packets_out && !tcp_write_queue_empty(sk); 24336ba8a3b1SNandita Dukkipati } 24346ba8a3b1SNandita Dukkipati 2435ed66dfafSNeal Cardwell bool tcp_schedule_loss_probe(struct sock *sk, bool advancing_rto) 24366ba8a3b1SNandita Dukkipati { 24376ba8a3b1SNandita Dukkipati struct inet_connection_sock *icsk = inet_csk(sk); 24386ba8a3b1SNandita Dukkipati struct tcp_sock *tp = tcp_sk(sk); 2439a2815817SNeal Cardwell u32 timeout, rto_delta_us; 24402ae21cf5SEric Dumazet int early_retrans; 24416ba8a3b1SNandita Dukkipati 24426ba8a3b1SNandita Dukkipati /* Don't do any loss probe on a Fast Open connection before 3WHS 24436ba8a3b1SNandita Dukkipati * finishes. 24446ba8a3b1SNandita Dukkipati */ 2445f9b99582SYuchung Cheng if (tp->fastopen_rsk) 24466ba8a3b1SNandita Dukkipati return false; 24476ba8a3b1SNandita Dukkipati 24482ae21cf5SEric Dumazet early_retrans = sock_net(sk)->ipv4.sysctl_tcp_early_retrans; 24496ba8a3b1SNandita Dukkipati /* Schedule a loss probe in 2*RTT for SACK capable connections 2450b4f70c3dSNeal Cardwell * not in loss recovery, that are either limited by cwnd or application. 24516ba8a3b1SNandita Dukkipati */ 24522ae21cf5SEric Dumazet if ((early_retrans != 3 && early_retrans != 4) || 2453bec41a11SYuchung Cheng !tp->packets_out || !tcp_is_sack(tp) || 2454b4f70c3dSNeal Cardwell (icsk->icsk_ca_state != TCP_CA_Open && 2455b4f70c3dSNeal Cardwell icsk->icsk_ca_state != TCP_CA_CWR)) 24566ba8a3b1SNandita Dukkipati return false; 24576ba8a3b1SNandita Dukkipati 2458bb4d991aSYuchung Cheng /* Probe timeout is 2*rtt. Add minimum RTO to account 2459f9b99582SYuchung Cheng * for delayed ack when there's one outstanding packet. If no RTT 2460f9b99582SYuchung Cheng * sample is available then probe after TCP_TIMEOUT_INIT. 24616ba8a3b1SNandita Dukkipati */ 2462bb4d991aSYuchung Cheng if (tp->srtt_us) { 2463bb4d991aSYuchung Cheng timeout = usecs_to_jiffies(tp->srtt_us >> 2); 24646ba8a3b1SNandita Dukkipati if (tp->packets_out == 1) 2465bb4d991aSYuchung Cheng timeout += TCP_RTO_MIN; 2466bb4d991aSYuchung Cheng else 2467bb4d991aSYuchung Cheng timeout += TCP_TIMEOUT_MIN; 2468bb4d991aSYuchung Cheng } else { 2469bb4d991aSYuchung Cheng timeout = TCP_TIMEOUT_INIT; 2470bb4d991aSYuchung Cheng } 24716ba8a3b1SNandita Dukkipati 2472a2815817SNeal Cardwell /* If the RTO formula yields an earlier time, then use that time. */ 2473ed66dfafSNeal Cardwell rto_delta_us = advancing_rto ? 2474ed66dfafSNeal Cardwell jiffies_to_usecs(inet_csk(sk)->icsk_rto) : 2475ed66dfafSNeal Cardwell tcp_rto_delta_us(sk); /* How far in future is RTO? */ 2476a2815817SNeal Cardwell if (rto_delta_us > 0) 2477a2815817SNeal Cardwell timeout = min_t(u32, timeout, usecs_to_jiffies(rto_delta_us)); 24786ba8a3b1SNandita Dukkipati 24793f80e08fSEric Dumazet tcp_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout, 24803f80e08fSEric Dumazet TCP_RTO_MAX, NULL); 24816ba8a3b1SNandita Dukkipati return true; 24826ba8a3b1SNandita Dukkipati } 24836ba8a3b1SNandita Dukkipati 24841f3279aeSEric Dumazet /* Thanks to skb fast clones, we can detect if a prior transmit of 24851f3279aeSEric Dumazet * a packet is still in a qdisc or driver queue. 24861f3279aeSEric Dumazet * In this case, there is very little point doing a retransmit ! 24871f3279aeSEric Dumazet */ 24881f3279aeSEric Dumazet static bool skb_still_in_host_queue(const struct sock *sk, 24891f3279aeSEric Dumazet const struct sk_buff *skb) 24901f3279aeSEric Dumazet { 249139bb5e62SEric Dumazet if (unlikely(skb_fclone_busy(sk, skb))) { 2492c10d9310SEric Dumazet NET_INC_STATS(sock_net(sk), 24931f3279aeSEric Dumazet LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES); 24941f3279aeSEric Dumazet return true; 24951f3279aeSEric Dumazet } 24961f3279aeSEric Dumazet return false; 24971f3279aeSEric Dumazet } 24981f3279aeSEric Dumazet 2499b340b264SYuchung Cheng /* When probe timeout (PTO) fires, try send a new segment if possible, else 25006ba8a3b1SNandita Dukkipati * retransmit the last segment. 25016ba8a3b1SNandita Dukkipati */ 25026ba8a3b1SNandita Dukkipati void tcp_send_loss_probe(struct sock *sk) 25036ba8a3b1SNandita Dukkipati { 25049b717a8dSNandita Dukkipati struct tcp_sock *tp = tcp_sk(sk); 25056ba8a3b1SNandita Dukkipati struct sk_buff *skb; 25066ba8a3b1SNandita Dukkipati int pcount; 25076ba8a3b1SNandita Dukkipati int mss = tcp_current_mss(sk); 25086ba8a3b1SNandita Dukkipati 2509b340b264SYuchung Cheng skb = tcp_send_head(sk); 251075c119afSEric Dumazet if (skb && tcp_snd_wnd_test(tp, skb, mss)) { 2511b340b264SYuchung Cheng pcount = tp->packets_out; 2512b340b264SYuchung Cheng tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC); 2513b340b264SYuchung Cheng if (tp->packets_out > pcount) 2514b340b264SYuchung Cheng goto probe_sent; 25156ba8a3b1SNandita Dukkipati goto rearm_timer; 25166ba8a3b1SNandita Dukkipati } 251775c119afSEric Dumazet skb = skb_rb_last(&sk->tcp_rtx_queue); 2518b2b7af86SYuchung Cheng if (unlikely(!skb)) { 2519b2b7af86SYuchung Cheng WARN_ONCE(tp->packets_out, 2520b2b7af86SYuchung Cheng "invalid inflight: %u state %u cwnd %u mss %d\n", 2521b2b7af86SYuchung Cheng tp->packets_out, sk->sk_state, tp->snd_cwnd, mss); 2522b2b7af86SYuchung Cheng inet_csk(sk)->icsk_pending = 0; 2523b2b7af86SYuchung Cheng return; 2524b2b7af86SYuchung Cheng } 25256ba8a3b1SNandita Dukkipati 25269b717a8dSNandita Dukkipati /* At most one outstanding TLP retransmission. */ 25279b717a8dSNandita Dukkipati if (tp->tlp_high_seq) 25289b717a8dSNandita Dukkipati goto rearm_timer; 25299b717a8dSNandita Dukkipati 25301f3279aeSEric Dumazet if (skb_still_in_host_queue(sk, skb)) 25311f3279aeSEric Dumazet goto rearm_timer; 25321f3279aeSEric Dumazet 25336ba8a3b1SNandita Dukkipati pcount = tcp_skb_pcount(skb); 25346ba8a3b1SNandita Dukkipati if (WARN_ON(!pcount)) 25356ba8a3b1SNandita Dukkipati goto rearm_timer; 25366ba8a3b1SNandita Dukkipati 25376ba8a3b1SNandita Dukkipati if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) { 253875c119afSEric Dumazet if (unlikely(tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb, 253975c119afSEric Dumazet (pcount - 1) * mss, mss, 25406cc55e09SOctavian Purdila GFP_ATOMIC))) 25416ba8a3b1SNandita Dukkipati goto rearm_timer; 254275c119afSEric Dumazet skb = skb_rb_next(skb); 25436ba8a3b1SNandita Dukkipati } 25446ba8a3b1SNandita Dukkipati 25456ba8a3b1SNandita Dukkipati if (WARN_ON(!skb || !tcp_skb_pcount(skb))) 25466ba8a3b1SNandita Dukkipati goto rearm_timer; 25476ba8a3b1SNandita Dukkipati 254810d3be56SEric Dumazet if (__tcp_retransmit_skb(sk, skb, 1)) 2549b340b264SYuchung Cheng goto rearm_timer; 25506ba8a3b1SNandita Dukkipati 25519b717a8dSNandita Dukkipati /* Record snd_nxt for loss detection. */ 25529b717a8dSNandita Dukkipati tp->tlp_high_seq = tp->snd_nxt; 25539b717a8dSNandita Dukkipati 2554b340b264SYuchung Cheng probe_sent: 2555c10d9310SEric Dumazet NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSPROBES); 2556fcd16c0aSYuchung Cheng /* Reset s.t. tcp_rearm_rto will restart timer from now */ 2557fcd16c0aSYuchung Cheng inet_csk(sk)->icsk_pending = 0; 2558b340b264SYuchung Cheng rearm_timer: 2559fcd16c0aSYuchung Cheng tcp_rearm_rto(sk); 25601da177e4SLinus Torvalds } 25611da177e4SLinus Torvalds 2562a762a980SDavid S. Miller /* Push out any pending frames which were held back due to 2563a762a980SDavid S. Miller * TCP_CORK or attempt at coalescing tiny packets. 2564a762a980SDavid S. Miller * The socket must be locked by the caller. 2565a762a980SDavid S. Miller */ 25669e412ba7SIlpo Järvinen void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss, 25679e412ba7SIlpo Järvinen int nonagle) 2568a762a980SDavid S. Miller { 2569726e07a8SIlpo Järvinen /* If we are closed, the bytes will have to remain here. 2570726e07a8SIlpo Järvinen * In time closedown will finish, we empty the write queue and 2571726e07a8SIlpo Järvinen * all will be happy. 2572726e07a8SIlpo Järvinen */ 2573726e07a8SIlpo Järvinen if (unlikely(sk->sk_state == TCP_CLOSE)) 2574726e07a8SIlpo Järvinen return; 2575726e07a8SIlpo Järvinen 257699a1dec7SMel Gorman if (tcp_write_xmit(sk, cur_mss, nonagle, 0, 25777450aaf6SEric Dumazet sk_gfp_mask(sk, GFP_ATOMIC))) 25789e412ba7SIlpo Järvinen tcp_check_probe_timer(sk); 2579a762a980SDavid S. Miller } 2580a762a980SDavid S. Miller 2581c1b4a7e6SDavid S. Miller /* Send _single_ skb sitting at the send head. This function requires 2582c1b4a7e6SDavid S. Miller * true push pending frames to setup probe timer etc. 2583c1b4a7e6SDavid S. Miller */ 2584c1b4a7e6SDavid S. Miller void tcp_push_one(struct sock *sk, unsigned int mss_now) 2585c1b4a7e6SDavid S. Miller { 2586fe067e8aSDavid S. Miller struct sk_buff *skb = tcp_send_head(sk); 2587c1b4a7e6SDavid S. Miller 2588c1b4a7e6SDavid S. Miller BUG_ON(!skb || skb->len < mss_now); 2589c1b4a7e6SDavid S. Miller 2590d5dd9175SIlpo Järvinen tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation); 2591c1b4a7e6SDavid S. Miller } 2592c1b4a7e6SDavid S. Miller 25931da177e4SLinus Torvalds /* This function returns the amount that we can raise the 25941da177e4SLinus Torvalds * usable window based on the following constraints 25951da177e4SLinus Torvalds * 25961da177e4SLinus Torvalds * 1. The window can never be shrunk once it is offered (RFC 793) 25971da177e4SLinus Torvalds * 2. We limit memory per socket 25981da177e4SLinus Torvalds * 25991da177e4SLinus Torvalds * RFC 1122: 26001da177e4SLinus Torvalds * "the suggested [SWS] avoidance algorithm for the receiver is to keep 26011da177e4SLinus Torvalds * RECV.NEXT + RCV.WIN fixed until: 26021da177e4SLinus Torvalds * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)" 26031da177e4SLinus Torvalds * 26041da177e4SLinus Torvalds * i.e. don't raise the right edge of the window until you can raise 26051da177e4SLinus Torvalds * it at least MSS bytes. 26061da177e4SLinus Torvalds * 26071da177e4SLinus Torvalds * Unfortunately, the recommended algorithm breaks header prediction, 26081da177e4SLinus Torvalds * since header prediction assumes th->window stays fixed. 26091da177e4SLinus Torvalds * 26101da177e4SLinus Torvalds * Strictly speaking, keeping th->window fixed violates the receiver 26111da177e4SLinus Torvalds * side SWS prevention criteria. The problem is that under this rule 26121da177e4SLinus Torvalds * a stream of single byte packets will cause the right side of the 26131da177e4SLinus Torvalds * window to always advance by a single byte. 26141da177e4SLinus Torvalds * 26151da177e4SLinus Torvalds * Of course, if the sender implements sender side SWS prevention 26161da177e4SLinus Torvalds * then this will not be a problem. 26171da177e4SLinus Torvalds * 26181da177e4SLinus Torvalds * BSD seems to make the following compromise: 26191da177e4SLinus Torvalds * 26201da177e4SLinus Torvalds * If the free space is less than the 1/4 of the maximum 26211da177e4SLinus Torvalds * space available and the free space is less than 1/2 mss, 26221da177e4SLinus Torvalds * then set the window to 0. 26231da177e4SLinus Torvalds * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ] 26241da177e4SLinus Torvalds * Otherwise, just prevent the window from shrinking 26251da177e4SLinus Torvalds * and from being larger than the largest representable value. 26261da177e4SLinus Torvalds * 26271da177e4SLinus Torvalds * This prevents incremental opening of the window in the regime 26281da177e4SLinus Torvalds * where TCP is limited by the speed of the reader side taking 26291da177e4SLinus Torvalds * data out of the TCP receive queue. It does nothing about 26301da177e4SLinus Torvalds * those cases where the window is constrained on the sender side 26311da177e4SLinus Torvalds * because the pipeline is full. 26321da177e4SLinus Torvalds * 26331da177e4SLinus Torvalds * BSD also seems to "accidentally" limit itself to windows that are a 26341da177e4SLinus Torvalds * multiple of MSS, at least until the free space gets quite small. 26351da177e4SLinus Torvalds * This would appear to be a side effect of the mbuf implementation. 26361da177e4SLinus Torvalds * Combining these two algorithms results in the observed behavior 26371da177e4SLinus Torvalds * of having a fixed window size at almost all times. 26381da177e4SLinus Torvalds * 26391da177e4SLinus Torvalds * Below we obtain similar behavior by forcing the offered window to 26401da177e4SLinus Torvalds * a multiple of the mss when it is feasible to do so. 26411da177e4SLinus Torvalds * 26421da177e4SLinus Torvalds * Note, we don't "adjust" for TIMESTAMP or SACK option bytes. 26431da177e4SLinus Torvalds * Regular options like TIMESTAMP are taken into account. 26441da177e4SLinus Torvalds */ 26451da177e4SLinus Torvalds u32 __tcp_select_window(struct sock *sk) 26461da177e4SLinus Torvalds { 2647463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 26481da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 2649caa20d9aSStephen Hemminger /* MSS for the peer's data. Previous versions used mss_clamp 26501da177e4SLinus Torvalds * here. I don't know if the value based on our guesses 26511da177e4SLinus Torvalds * of peer's MSS is better for the performance. It's more correct 26521da177e4SLinus Torvalds * but may be worse for the performance because of rcv_mss 26531da177e4SLinus Torvalds * fluctuations. --SAW 1998/11/1 26541da177e4SLinus Torvalds */ 2655463c84b9SArnaldo Carvalho de Melo int mss = icsk->icsk_ack.rcv_mss; 26561da177e4SLinus Torvalds int free_space = tcp_space(sk); 265786c1a045SFlorian Westphal int allowed_space = tcp_full_space(sk); 265886c1a045SFlorian Westphal int full_space = min_t(int, tp->window_clamp, allowed_space); 26591da177e4SLinus Torvalds int window; 26601da177e4SLinus Torvalds 266106425c30SEric Dumazet if (unlikely(mss > full_space)) { 26621da177e4SLinus Torvalds mss = full_space; 266306425c30SEric Dumazet if (mss <= 0) 266406425c30SEric Dumazet return 0; 266506425c30SEric Dumazet } 2666b92edbe0SEric Dumazet if (free_space < (full_space >> 1)) { 2667463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.quick = 0; 26681da177e4SLinus Torvalds 2669b8da51ebSEric Dumazet if (tcp_under_memory_pressure(sk)) 2670056834d9SIlpo Järvinen tp->rcv_ssthresh = min(tp->rcv_ssthresh, 2671056834d9SIlpo Järvinen 4U * tp->advmss); 26721da177e4SLinus Torvalds 267386c1a045SFlorian Westphal /* free_space might become our new window, make sure we don't 267486c1a045SFlorian Westphal * increase it due to wscale. 267586c1a045SFlorian Westphal */ 267686c1a045SFlorian Westphal free_space = round_down(free_space, 1 << tp->rx_opt.rcv_wscale); 267786c1a045SFlorian Westphal 267886c1a045SFlorian Westphal /* if free space is less than mss estimate, or is below 1/16th 267986c1a045SFlorian Westphal * of the maximum allowed, try to move to zero-window, else 268086c1a045SFlorian Westphal * tcp_clamp_window() will grow rcv buf up to tcp_rmem[2], and 268186c1a045SFlorian Westphal * new incoming data is dropped due to memory limits. 268286c1a045SFlorian Westphal * With large window, mss test triggers way too late in order 268386c1a045SFlorian Westphal * to announce zero window in time before rmem limit kicks in. 268486c1a045SFlorian Westphal */ 268586c1a045SFlorian Westphal if (free_space < (allowed_space >> 4) || free_space < mss) 26861da177e4SLinus Torvalds return 0; 26871da177e4SLinus Torvalds } 26881da177e4SLinus Torvalds 26891da177e4SLinus Torvalds if (free_space > tp->rcv_ssthresh) 26901da177e4SLinus Torvalds free_space = tp->rcv_ssthresh; 26911da177e4SLinus Torvalds 26921da177e4SLinus Torvalds /* Don't do rounding if we are using window scaling, since the 26931da177e4SLinus Torvalds * scaled window will not line up with the MSS boundary anyway. 26941da177e4SLinus Torvalds */ 26951da177e4SLinus Torvalds if (tp->rx_opt.rcv_wscale) { 26961da177e4SLinus Torvalds window = free_space; 26971da177e4SLinus Torvalds 26981da177e4SLinus Torvalds /* Advertise enough space so that it won't get scaled away. 26991da177e4SLinus Torvalds * Import case: prevent zero window announcement if 27001da177e4SLinus Torvalds * 1<<rcv_wscale > mss. 27011da177e4SLinus Torvalds */ 27021935299dSGao Feng window = ALIGN(window, (1 << tp->rx_opt.rcv_wscale)); 27031da177e4SLinus Torvalds } else { 27041935299dSGao Feng window = tp->rcv_wnd; 27051da177e4SLinus Torvalds /* Get the largest window that is a nice multiple of mss. 27061da177e4SLinus Torvalds * Window clamp already applied above. 27071da177e4SLinus Torvalds * If our current window offering is within 1 mss of the 27081da177e4SLinus Torvalds * free space we just keep it. This prevents the divide 27091da177e4SLinus Torvalds * and multiply from happening most of the time. 27101da177e4SLinus Torvalds * We also don't do any window rounding when the free space 27111da177e4SLinus Torvalds * is too small. 27121da177e4SLinus Torvalds */ 27131da177e4SLinus Torvalds if (window <= free_space - mss || window > free_space) 27141935299dSGao Feng window = rounddown(free_space, mss); 271584565070SJohn Heffner else if (mss == full_space && 2716b92edbe0SEric Dumazet free_space > window + (full_space >> 1)) 271784565070SJohn Heffner window = free_space; 27181da177e4SLinus Torvalds } 27191da177e4SLinus Torvalds 27201da177e4SLinus Torvalds return window; 27211da177e4SLinus Torvalds } 27221da177e4SLinus Torvalds 2723cfea5a68SMartin KaFai Lau void tcp_skb_collapse_tstamp(struct sk_buff *skb, 2724082ac2d5SMartin KaFai Lau const struct sk_buff *next_skb) 2725082ac2d5SMartin KaFai Lau { 27260a2cf20cSSoheil Hassas Yeganeh if (unlikely(tcp_has_tx_tstamp(next_skb))) { 27270a2cf20cSSoheil Hassas Yeganeh const struct skb_shared_info *next_shinfo = 27280a2cf20cSSoheil Hassas Yeganeh skb_shinfo(next_skb); 2729082ac2d5SMartin KaFai Lau struct skb_shared_info *shinfo = skb_shinfo(skb); 2730082ac2d5SMartin KaFai Lau 27310a2cf20cSSoheil Hassas Yeganeh shinfo->tx_flags |= next_shinfo->tx_flags & SKBTX_ANY_TSTAMP; 2732082ac2d5SMartin KaFai Lau shinfo->tskey = next_shinfo->tskey; 27332de8023eSMartin KaFai Lau TCP_SKB_CB(skb)->txstamp_ack |= 27342de8023eSMartin KaFai Lau TCP_SKB_CB(next_skb)->txstamp_ack; 2735082ac2d5SMartin KaFai Lau } 2736082ac2d5SMartin KaFai Lau } 2737082ac2d5SMartin KaFai Lau 27384a17fc3aSIlpo Järvinen /* Collapses two adjacent SKB's during retransmission. */ 2739f8071cdeSEric Dumazet static bool tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb) 27401da177e4SLinus Torvalds { 27411da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 274275c119afSEric Dumazet struct sk_buff *next_skb = skb_rb_next(skb); 274313dde04fSWei Yongjun int next_skb_size; 27441da177e4SLinus Torvalds 2745058dc334SIlpo Järvinen next_skb_size = next_skb->len; 27461da177e4SLinus Torvalds 2747058dc334SIlpo Järvinen BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1); 27481da177e4SLinus Torvalds 2749f8071cdeSEric Dumazet if (next_skb_size) { 2750f8071cdeSEric Dumazet if (next_skb_size <= skb_availroom(skb)) 2751f8071cdeSEric Dumazet skb_copy_bits(next_skb, 0, skb_put(skb, next_skb_size), 2752f8071cdeSEric Dumazet next_skb_size); 2753f8071cdeSEric Dumazet else if (!skb_shift(skb, next_skb, next_skb_size)) 2754f8071cdeSEric Dumazet return false; 2755f8071cdeSEric Dumazet } 27562b7cda9cSEric Dumazet tcp_highest_sack_replace(sk, next_skb, skb); 2757a6963a6bSIlpo Järvinen 27581da177e4SLinus Torvalds /* Update sequence range on original skb. */ 27591da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq; 27601da177e4SLinus Torvalds 2761e6c7d085SIlpo Järvinen /* Merge over control information. This moves PSH/FIN etc. over */ 27624de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags; 27631da177e4SLinus Torvalds 27641da177e4SLinus Torvalds /* All done, get rid of second SKB and account for it so 27651da177e4SLinus Torvalds * packet counting does not break. 27661da177e4SLinus Torvalds */ 27674828e7f4SIlpo Järvinen TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS; 2768a643b5d4SMartin KaFai Lau TCP_SKB_CB(skb)->eor = TCP_SKB_CB(next_skb)->eor; 2769b7689205SIlpo Järvinen 2770b7689205SIlpo Järvinen /* changed transmit queue under us so clear hints */ 2771ef9da47cSIlpo Järvinen tcp_clear_retrans_hints_partial(tp); 2772ef9da47cSIlpo Järvinen if (next_skb == tp->retransmit_skb_hint) 2773ef9da47cSIlpo Järvinen tp->retransmit_skb_hint = skb; 2774b7689205SIlpo Järvinen 2775797108d1SIlpo Järvinen tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb)); 2776797108d1SIlpo Järvinen 2777082ac2d5SMartin KaFai Lau tcp_skb_collapse_tstamp(skb, next_skb); 2778082ac2d5SMartin KaFai Lau 277975c119afSEric Dumazet tcp_rtx_queue_unlink_and_free(next_skb, sk); 2780f8071cdeSEric Dumazet return true; 27811da177e4SLinus Torvalds } 27821da177e4SLinus Torvalds 278367edfef7SAndi Kleen /* Check if coalescing SKBs is legal. */ 2784a2a385d6SEric Dumazet static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb) 27854a17fc3aSIlpo Järvinen { 27864a17fc3aSIlpo Järvinen if (tcp_skb_pcount(skb) > 1) 2787a2a385d6SEric Dumazet return false; 27884a17fc3aSIlpo Järvinen if (skb_cloned(skb)) 2789a2a385d6SEric Dumazet return false; 27902331ccc5SEric Dumazet /* Some heuristics for collapsing over SACK'd could be invented */ 27914a17fc3aSIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 2792a2a385d6SEric Dumazet return false; 27934a17fc3aSIlpo Järvinen 2794a2a385d6SEric Dumazet return true; 27954a17fc3aSIlpo Järvinen } 27964a17fc3aSIlpo Järvinen 279767edfef7SAndi Kleen /* Collapse packets in the retransmit queue to make to create 279867edfef7SAndi Kleen * less packets on the wire. This is only done on retransmission. 279967edfef7SAndi Kleen */ 28004a17fc3aSIlpo Järvinen static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to, 28014a17fc3aSIlpo Järvinen int space) 28024a17fc3aSIlpo Järvinen { 28034a17fc3aSIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 28044a17fc3aSIlpo Järvinen struct sk_buff *skb = to, *tmp; 2805a2a385d6SEric Dumazet bool first = true; 28064a17fc3aSIlpo Järvinen 2807e0a1e5b5SEric Dumazet if (!sock_net(sk)->ipv4.sysctl_tcp_retrans_collapse) 28084a17fc3aSIlpo Järvinen return; 28094de075e0SEric Dumazet if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN) 28104a17fc3aSIlpo Järvinen return; 28114a17fc3aSIlpo Järvinen 281275c119afSEric Dumazet skb_rbtree_walk_from_safe(skb, tmp) { 28134a17fc3aSIlpo Järvinen if (!tcp_can_collapse(sk, skb)) 28144a17fc3aSIlpo Järvinen break; 28154a17fc3aSIlpo Järvinen 2816a643b5d4SMartin KaFai Lau if (!tcp_skb_can_collapse_to(to)) 2817a643b5d4SMartin KaFai Lau break; 2818a643b5d4SMartin KaFai Lau 28194a17fc3aSIlpo Järvinen space -= skb->len; 28204a17fc3aSIlpo Järvinen 28214a17fc3aSIlpo Järvinen if (first) { 2822a2a385d6SEric Dumazet first = false; 28234a17fc3aSIlpo Järvinen continue; 28244a17fc3aSIlpo Järvinen } 28254a17fc3aSIlpo Järvinen 28264a17fc3aSIlpo Järvinen if (space < 0) 28274a17fc3aSIlpo Järvinen break; 28284a17fc3aSIlpo Järvinen 28294a17fc3aSIlpo Järvinen if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp))) 28304a17fc3aSIlpo Järvinen break; 28314a17fc3aSIlpo Järvinen 2832f8071cdeSEric Dumazet if (!tcp_collapse_retrans(sk, to)) 2833f8071cdeSEric Dumazet break; 28344a17fc3aSIlpo Järvinen } 28354a17fc3aSIlpo Järvinen } 28364a17fc3aSIlpo Järvinen 28371da177e4SLinus Torvalds /* This retransmits one SKB. Policy decisions and retransmit queue 28381da177e4SLinus Torvalds * state updates are done by the caller. Returns non-zero if an 28391da177e4SLinus Torvalds * error occurred which prevented the send. 28401da177e4SLinus Torvalds */ 284110d3be56SEric Dumazet int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs) 28421da177e4SLinus Torvalds { 28435d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 284410d3be56SEric Dumazet struct tcp_sock *tp = tcp_sk(sk); 28457d227cd2SSridhar Samudrala unsigned int cur_mss; 284610d3be56SEric Dumazet int diff, len, err; 28471da177e4SLinus Torvalds 284810d3be56SEric Dumazet 284910d3be56SEric Dumazet /* Inconclusive MTU probe */ 285010d3be56SEric Dumazet if (icsk->icsk_mtup.probe_size) 28515d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = 0; 28525d424d5aSJohn Heffner 28531da177e4SLinus Torvalds /* Do not sent more than we queued. 1/4 is reserved for possible 2854caa20d9aSStephen Hemminger * copying overhead: fragmentation, tunneling, mangling etc. 28551da177e4SLinus Torvalds */ 285614afee4bSReshetova, Elena if (refcount_read(&sk->sk_wmem_alloc) > 2857ffb4d6c8SEric Dumazet min_t(u32, sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), 2858ffb4d6c8SEric Dumazet sk->sk_sndbuf)) 28591da177e4SLinus Torvalds return -EAGAIN; 28601da177e4SLinus Torvalds 28611f3279aeSEric Dumazet if (skb_still_in_host_queue(sk, skb)) 28621f3279aeSEric Dumazet return -EBUSY; 28631f3279aeSEric Dumazet 28641da177e4SLinus Torvalds if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) { 28657f582b24SEric Dumazet if (unlikely(before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))) { 28667f582b24SEric Dumazet WARN_ON_ONCE(1); 28677f582b24SEric Dumazet return -EINVAL; 28687f582b24SEric Dumazet } 28691da177e4SLinus Torvalds if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq)) 28701da177e4SLinus Torvalds return -ENOMEM; 28711da177e4SLinus Torvalds } 28721da177e4SLinus Torvalds 28737d227cd2SSridhar Samudrala if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk)) 28747d227cd2SSridhar Samudrala return -EHOSTUNREACH; /* Routing failure or similar. */ 28757d227cd2SSridhar Samudrala 28760c54b85fSIlpo Järvinen cur_mss = tcp_current_mss(sk); 28777d227cd2SSridhar Samudrala 28781da177e4SLinus Torvalds /* If receiver has shrunk his window, and skb is out of 28791da177e4SLinus Torvalds * new window, do not retransmit it. The exception is the 28801da177e4SLinus Torvalds * case, when window is shrunk to zero. In this case 28811da177e4SLinus Torvalds * our retransmit serves as a zero window probe. 28821da177e4SLinus Torvalds */ 28839d4fb27dSJoe Perches if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) && 28849d4fb27dSJoe Perches TCP_SKB_CB(skb)->seq != tp->snd_una) 28851da177e4SLinus Torvalds return -EAGAIN; 28861da177e4SLinus Torvalds 288710d3be56SEric Dumazet len = cur_mss * segs; 288810d3be56SEric Dumazet if (skb->len > len) { 288975c119afSEric Dumazet if (tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb, len, 289075c119afSEric Dumazet cur_mss, GFP_ATOMIC)) 28911da177e4SLinus Torvalds return -ENOMEM; /* We'll try again later. */ 289202276f3cSIlpo Järvinen } else { 2893c52e2421SEric Dumazet if (skb_unclone(skb, GFP_ATOMIC)) 2894c52e2421SEric Dumazet return -ENOMEM; 289510d3be56SEric Dumazet 289610d3be56SEric Dumazet diff = tcp_skb_pcount(skb); 289710d3be56SEric Dumazet tcp_set_skb_tso_segs(skb, cur_mss); 289810d3be56SEric Dumazet diff -= tcp_skb_pcount(skb); 289910d3be56SEric Dumazet if (diff) 290010d3be56SEric Dumazet tcp_adjust_pcount(sk, skb, diff); 290110d3be56SEric Dumazet if (skb->len < cur_mss) 290210d3be56SEric Dumazet tcp_retrans_try_collapse(sk, skb, cur_mss); 29031da177e4SLinus Torvalds } 29041da177e4SLinus Torvalds 290549213555SDaniel Borkmann /* RFC3168, section 6.1.1.1. ECN fallback */ 290649213555SDaniel Borkmann if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN_ECN) == TCPHDR_SYN_ECN) 290749213555SDaniel Borkmann tcp_ecn_clear_syn(sk, skb); 290849213555SDaniel Borkmann 2909678550c6SYuchung Cheng /* Update global and local TCP statistics. */ 2910678550c6SYuchung Cheng segs = tcp_skb_pcount(skb); 2911678550c6SYuchung Cheng TCP_ADD_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS, segs); 2912678550c6SYuchung Cheng if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN) 2913678550c6SYuchung Cheng __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS); 2914678550c6SYuchung Cheng tp->total_retrans += segs; 2915fb31c9b9SWei Wang tp->bytes_retrans += skb->len; 2916678550c6SYuchung Cheng 291750bceae9SThomas Graf /* make sure skb->data is aligned on arches that require it 291850bceae9SThomas Graf * and check if ack-trimming & collapsing extended the headroom 291950bceae9SThomas Graf * beyond what csum_start can cover. 292050bceae9SThomas Graf */ 292150bceae9SThomas Graf if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) || 292250bceae9SThomas Graf skb_headroom(skb) >= 0xFFFF)) { 292310a81980SEric Dumazet struct sk_buff *nskb; 292410a81980SEric Dumazet 2925e2080072SEric Dumazet tcp_skb_tsorted_save(skb) { 292610a81980SEric Dumazet nskb = __pskb_copy(skb, MAX_TCP_HEADER, GFP_ATOMIC); 2927c84a5711SYuchung Cheng err = nskb ? tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC) : 2928117632e6SEric Dumazet -ENOBUFS; 2929e2080072SEric Dumazet } tcp_skb_tsorted_restore(skb); 2930e2080072SEric Dumazet 29315889e2c0SYousuk Seung if (!err) { 2932a7a25630SEric Dumazet tcp_update_skb_after_send(sk, skb, tp->tcp_wstamp_ns); 29335889e2c0SYousuk Seung tcp_rate_skb_sent(sk, skb); 29345889e2c0SYousuk Seung } 2935117632e6SEric Dumazet } else { 2936c84a5711SYuchung Cheng err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 2937117632e6SEric Dumazet } 2938c84a5711SYuchung Cheng 29397f12422cSYuchung Cheng /* To avoid taking spuriously low RTT samples based on a timestamp 29407f12422cSYuchung Cheng * for a transmit that never happened, always mark EVER_RETRANS 29417f12422cSYuchung Cheng */ 29427f12422cSYuchung Cheng TCP_SKB_CB(skb)->sacked |= TCPCB_EVER_RETRANS; 29437f12422cSYuchung Cheng 2944a31ad29eSLawrence Brakmo if (BPF_SOCK_OPS_TEST_FLAG(tp, BPF_SOCK_OPS_RETRANS_CB_FLAG)) 2945a31ad29eSLawrence Brakmo tcp_call_bpf_3arg(sk, BPF_SOCK_OPS_RETRANS_CB, 2946a31ad29eSLawrence Brakmo TCP_SKB_CB(skb)->seq, segs, err); 2947a31ad29eSLawrence Brakmo 2948fc9f3501SEric Dumazet if (likely(!err)) { 2949e086101bSCong Wang trace_tcp_retransmit_skb(sk, skb); 2950678550c6SYuchung Cheng } else if (err != -EBUSY) { 2951ec641b39SYuchung Cheng NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL, segs); 2952fc9f3501SEric Dumazet } 2953c84a5711SYuchung Cheng return err; 295493b174adSYuchung Cheng } 295593b174adSYuchung Cheng 295610d3be56SEric Dumazet int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs) 295793b174adSYuchung Cheng { 295893b174adSYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 295910d3be56SEric Dumazet int err = __tcp_retransmit_skb(sk, skb, segs); 29601da177e4SLinus Torvalds 29611da177e4SLinus Torvalds if (err == 0) { 29621da177e4SLinus Torvalds #if FASTRETRANS_DEBUG > 0 29631da177e4SLinus Torvalds if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) { 2964e87cc472SJoe Perches net_dbg_ratelimited("retrans_out leaked\n"); 29651da177e4SLinus Torvalds } 29661da177e4SLinus Torvalds #endif 29671da177e4SLinus Torvalds TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS; 29681da177e4SLinus Torvalds tp->retrans_out += tcp_skb_pcount(skb); 29697ae18975SYuchung Cheng } 29701da177e4SLinus Torvalds 29717ae18975SYuchung Cheng /* Save stamp of the first (attempted) retransmit. */ 29721da177e4SLinus Torvalds if (!tp->retrans_stamp) 29737faee5c0SEric Dumazet tp->retrans_stamp = tcp_skb_timestamp(skb); 29741da177e4SLinus Torvalds 29756e08d5e3SYuchung Cheng if (tp->undo_retrans < 0) 29766e08d5e3SYuchung Cheng tp->undo_retrans = 0; 29776e08d5e3SYuchung Cheng tp->undo_retrans += tcp_skb_pcount(skb); 29781da177e4SLinus Torvalds return err; 29791da177e4SLinus Torvalds } 29801da177e4SLinus Torvalds 29811da177e4SLinus Torvalds /* This gets called after a retransmit timeout, and the initially 29821da177e4SLinus Torvalds * retransmitted data is acknowledged. It tries to continue 29831da177e4SLinus Torvalds * resending the rest of the retransmit queue, until either 29841da177e4SLinus Torvalds * we've sent it all or the congestion window limit is reached. 29851da177e4SLinus Torvalds */ 29861da177e4SLinus Torvalds void tcp_xmit_retransmit_queue(struct sock *sk) 29871da177e4SLinus Torvalds { 29886687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 2989b9f1f1ceSEric Dumazet struct sk_buff *skb, *rtx_head, *hole = NULL; 29901da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 2991840a3cbeSYuchung Cheng u32 max_segs; 299261eb55f4SIlpo Järvinen int mib_idx; 29936a438bbeSStephen Hemminger 299445e77d31SIlpo Järvinen if (!tp->packets_out) 299545e77d31SIlpo Järvinen return; 299645e77d31SIlpo Järvinen 299775c119afSEric Dumazet rtx_head = tcp_rtx_queue_head(sk); 2998b9f1f1ceSEric Dumazet skb = tp->retransmit_skb_hint ?: rtx_head; 2999ed6e7268SNeal Cardwell max_segs = tcp_tso_segs(sk, tcp_current_mss(sk)); 300075c119afSEric Dumazet skb_rbtree_walk_from(skb) { 3001dca0aaf8SEric Dumazet __u8 sacked; 300210d3be56SEric Dumazet int segs; 30031da177e4SLinus Torvalds 3004218af599SEric Dumazet if (tcp_pacing_check(sk)) 3005218af599SEric Dumazet break; 3006218af599SEric Dumazet 30076a438bbeSStephen Hemminger /* we could do better than to assign each time */ 300851456b29SIan Morris if (!hole) 30096a438bbeSStephen Hemminger tp->retransmit_skb_hint = skb; 30106a438bbeSStephen Hemminger 301110d3be56SEric Dumazet segs = tp->snd_cwnd - tcp_packets_in_flight(tp); 301210d3be56SEric Dumazet if (segs <= 0) 30131da177e4SLinus Torvalds return; 3014dca0aaf8SEric Dumazet sacked = TCP_SKB_CB(skb)->sacked; 3015a3d2e9f8SEric Dumazet /* In case tcp_shift_skb_data() have aggregated large skbs, 3016a3d2e9f8SEric Dumazet * we need to make sure not sending too bigs TSO packets 3017a3d2e9f8SEric Dumazet */ 3018a3d2e9f8SEric Dumazet segs = min_t(int, segs, max_segs); 30190e1c54c2SIlpo Järvinen 3020840a3cbeSYuchung Cheng if (tp->retrans_out >= tp->lost_out) { 3021006f582cSIlpo Järvinen break; 30220e1c54c2SIlpo Järvinen } else if (!(sacked & TCPCB_LOST)) { 302351456b29SIan Morris if (!hole && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED))) 30240e1c54c2SIlpo Järvinen hole = skb; 302561eb55f4SIlpo Järvinen continue; 30261da177e4SLinus Torvalds 30270e1c54c2SIlpo Järvinen } else { 30280e1c54c2SIlpo Järvinen if (icsk->icsk_ca_state != TCP_CA_Loss) 30290e1c54c2SIlpo Järvinen mib_idx = LINUX_MIB_TCPFASTRETRANS; 30300e1c54c2SIlpo Järvinen else 30310e1c54c2SIlpo Järvinen mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS; 30320e1c54c2SIlpo Järvinen } 30330e1c54c2SIlpo Järvinen 30340e1c54c2SIlpo Järvinen if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS)) 303561eb55f4SIlpo Järvinen continue; 303640b215e5SPavel Emelyanov 3037f9616c35SEric Dumazet if (tcp_small_queue_check(sk, skb, 1)) 3038f9616c35SEric Dumazet return; 3039f9616c35SEric Dumazet 304010d3be56SEric Dumazet if (tcp_retransmit_skb(sk, skb, segs)) 30411da177e4SLinus Torvalds return; 304224ab6becSYuchung Cheng 3043de1d6578SYuchung Cheng NET_ADD_STATS(sock_net(sk), mib_idx, tcp_skb_pcount(skb)); 30441da177e4SLinus Torvalds 3045684bad11SYuchung Cheng if (tcp_in_cwnd_reduction(sk)) 3046a262f0cdSNandita Dukkipati tp->prr_out += tcp_skb_pcount(skb); 3047a262f0cdSNandita Dukkipati 304875c119afSEric Dumazet if (skb == rtx_head && 304957dde7f7SYuchung Cheng icsk->icsk_pending != ICSK_TIME_REO_TIMEOUT) 30503f80e08fSEric Dumazet tcp_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 30513f421baaSArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto, 30523f80e08fSEric Dumazet TCP_RTO_MAX, 30533f80e08fSEric Dumazet skb); 30541da177e4SLinus Torvalds } 30551da177e4SLinus Torvalds } 30561da177e4SLinus Torvalds 3057d83769a5SEric Dumazet /* We allow to exceed memory limits for FIN packets to expedite 3058d83769a5SEric Dumazet * connection tear down and (memory) recovery. 3059845704a5SEric Dumazet * Otherwise tcp_send_fin() could be tempted to either delay FIN 3060845704a5SEric Dumazet * or even be forced to close flow without any FIN. 3061a6c5ea4cSEric Dumazet * In general, we want to allow one skb per socket to avoid hangs 3062a6c5ea4cSEric Dumazet * with edge trigger epoll() 3063d83769a5SEric Dumazet */ 3064a6c5ea4cSEric Dumazet void sk_forced_mem_schedule(struct sock *sk, int size) 3065d83769a5SEric Dumazet { 3066e805605cSJohannes Weiner int amt; 3067d83769a5SEric Dumazet 3068d83769a5SEric Dumazet if (size <= sk->sk_forward_alloc) 3069d83769a5SEric Dumazet return; 3070d83769a5SEric Dumazet amt = sk_mem_pages(size); 3071d83769a5SEric Dumazet sk->sk_forward_alloc += amt * SK_MEM_QUANTUM; 3072e805605cSJohannes Weiner sk_memory_allocated_add(sk, amt); 3073e805605cSJohannes Weiner 3074baac50bbSJohannes Weiner if (mem_cgroup_sockets_enabled && sk->sk_memcg) 3075baac50bbSJohannes Weiner mem_cgroup_charge_skmem(sk->sk_memcg, amt); 3076d83769a5SEric Dumazet } 3077d83769a5SEric Dumazet 3078845704a5SEric Dumazet /* Send a FIN. The caller locks the socket for us. 3079845704a5SEric Dumazet * We should try to send a FIN packet really hard, but eventually give up. 30801da177e4SLinus Torvalds */ 30811da177e4SLinus Torvalds void tcp_send_fin(struct sock *sk) 30821da177e4SLinus Torvalds { 3083845704a5SEric Dumazet struct sk_buff *skb, *tskb = tcp_write_queue_tail(sk); 30841da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 30851da177e4SLinus Torvalds 3086845704a5SEric Dumazet /* Optimization, tack on the FIN if we have one skb in write queue and 3087845704a5SEric Dumazet * this skb was not yet sent, or we are under memory pressure. 3088845704a5SEric Dumazet * Note: in the latter case, FIN packet will be sent after a timeout, 3089845704a5SEric Dumazet * as TCP stack thinks it has already been transmitted. 30901da177e4SLinus Torvalds */ 309175c119afSEric Dumazet if (!tskb && tcp_under_memory_pressure(sk)) 309275c119afSEric Dumazet tskb = skb_rb_last(&sk->tcp_rtx_queue); 309375c119afSEric Dumazet 309475c119afSEric Dumazet if (tskb) { 3095845704a5SEric Dumazet coalesce: 3096845704a5SEric Dumazet TCP_SKB_CB(tskb)->tcp_flags |= TCPHDR_FIN; 3097845704a5SEric Dumazet TCP_SKB_CB(tskb)->end_seq++; 30981da177e4SLinus Torvalds tp->write_seq++; 309975c119afSEric Dumazet if (tcp_write_queue_empty(sk)) { 3100845704a5SEric Dumazet /* This means tskb was already sent. 3101845704a5SEric Dumazet * Pretend we included the FIN on previous transmit. 3102845704a5SEric Dumazet * We need to set tp->snd_nxt to the value it would have 3103845704a5SEric Dumazet * if FIN had been sent. This is because retransmit path 3104845704a5SEric Dumazet * does not change tp->snd_nxt. 3105845704a5SEric Dumazet */ 3106845704a5SEric Dumazet tp->snd_nxt++; 3107845704a5SEric Dumazet return; 3108845704a5SEric Dumazet } 31091da177e4SLinus Torvalds } else { 3110845704a5SEric Dumazet skb = alloc_skb_fclone(MAX_TCP_HEADER, sk->sk_allocation); 3111845704a5SEric Dumazet if (unlikely(!skb)) { 3112845704a5SEric Dumazet if (tskb) 3113845704a5SEric Dumazet goto coalesce; 3114845704a5SEric Dumazet return; 31151da177e4SLinus Torvalds } 3116e2080072SEric Dumazet INIT_LIST_HEAD(&skb->tcp_tsorted_anchor); 3117d83769a5SEric Dumazet skb_reserve(skb, MAX_TCP_HEADER); 3118a6c5ea4cSEric Dumazet sk_forced_mem_schedule(sk, skb->truesize); 31191da177e4SLinus Torvalds /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */ 3120e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tp->write_seq, 3121a3433f35SChangli Gao TCPHDR_ACK | TCPHDR_FIN); 31221da177e4SLinus Torvalds tcp_queue_skb(sk, skb); 31231da177e4SLinus Torvalds } 3124845704a5SEric Dumazet __tcp_push_pending_frames(sk, tcp_current_mss(sk), TCP_NAGLE_OFF); 31251da177e4SLinus Torvalds } 31261da177e4SLinus Torvalds 31271da177e4SLinus Torvalds /* We get here when a process closes a file descriptor (either due to 31281da177e4SLinus Torvalds * an explicit close() or as a byproduct of exit()'ing) and there 31291da177e4SLinus Torvalds * was unread data in the receive queue. This behavior is recommended 313065bb723cSGerrit Renker * by RFC 2525, section 2.17. -DaveM 31311da177e4SLinus Torvalds */ 3132dd0fc66fSAl Viro void tcp_send_active_reset(struct sock *sk, gfp_t priority) 31331da177e4SLinus Torvalds { 31341da177e4SLinus Torvalds struct sk_buff *skb; 31351da177e4SLinus Torvalds 31367cc2b043SGao Feng TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS); 31377cc2b043SGao Feng 31381da177e4SLinus Torvalds /* NOTE: No TCP options attached and we never retransmit this. */ 31391da177e4SLinus Torvalds skb = alloc_skb(MAX_TCP_HEADER, priority); 31401da177e4SLinus Torvalds if (!skb) { 31414e673444SPavel Emelyanov NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); 31421da177e4SLinus Torvalds return; 31431da177e4SLinus Torvalds } 31441da177e4SLinus Torvalds 31451da177e4SLinus Torvalds /* Reserve space for headers and prepare control bits. */ 31461da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 3147e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk), 3148a3433f35SChangli Gao TCPHDR_ACK | TCPHDR_RST); 31499a568de4SEric Dumazet tcp_mstamp_refresh(tcp_sk(sk)); 31501da177e4SLinus Torvalds /* Send it off. */ 3151dfb4b9dcSDavid S. Miller if (tcp_transmit_skb(sk, skb, 0, priority)) 31524e673444SPavel Emelyanov NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); 3153c24b14c4SSong Liu 3154c24b14c4SSong Liu /* skb of trace_tcp_send_reset() keeps the skb that caused RST, 3155c24b14c4SSong Liu * skb here is different to the troublesome skb, so use NULL 3156c24b14c4SSong Liu */ 3157c24b14c4SSong Liu trace_tcp_send_reset(sk, NULL); 31581da177e4SLinus Torvalds } 31591da177e4SLinus Torvalds 316067edfef7SAndi Kleen /* Send a crossed SYN-ACK during socket establishment. 316167edfef7SAndi Kleen * WARNING: This routine must only be called when we have already sent 31621da177e4SLinus Torvalds * a SYN packet that crossed the incoming SYN that caused this routine 31631da177e4SLinus Torvalds * to get called. If this assumption fails then the initial rcv_wnd 31641da177e4SLinus Torvalds * and rcv_wscale values will not be correct. 31651da177e4SLinus Torvalds */ 31661da177e4SLinus Torvalds int tcp_send_synack(struct sock *sk) 31671da177e4SLinus Torvalds { 31681da177e4SLinus Torvalds struct sk_buff *skb; 31691da177e4SLinus Torvalds 317075c119afSEric Dumazet skb = tcp_rtx_queue_head(sk); 317151456b29SIan Morris if (!skb || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) { 317275c119afSEric Dumazet pr_err("%s: wrong queue state\n", __func__); 31731da177e4SLinus Torvalds return -EFAULT; 31741da177e4SLinus Torvalds } 31754de075e0SEric Dumazet if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) { 31761da177e4SLinus Torvalds if (skb_cloned(skb)) { 3177e2080072SEric Dumazet struct sk_buff *nskb; 3178e2080072SEric Dumazet 3179e2080072SEric Dumazet tcp_skb_tsorted_save(skb) { 3180e2080072SEric Dumazet nskb = skb_copy(skb, GFP_ATOMIC); 3181e2080072SEric Dumazet } tcp_skb_tsorted_restore(skb); 318251456b29SIan Morris if (!nskb) 31831da177e4SLinus Torvalds return -ENOMEM; 3184e2080072SEric Dumazet INIT_LIST_HEAD(&nskb->tcp_tsorted_anchor); 318575c119afSEric Dumazet tcp_rtx_queue_unlink_and_free(skb, sk); 3186f4a775d1SEric Dumazet __skb_header_release(nskb); 318775c119afSEric Dumazet tcp_rbtree_insert(&sk->tcp_rtx_queue, nskb); 31883ab224beSHideo Aoki sk->sk_wmem_queued += nskb->truesize; 31893ab224beSHideo Aoki sk_mem_charge(sk, nskb->truesize); 31901da177e4SLinus Torvalds skb = nskb; 31911da177e4SLinus Torvalds } 31921da177e4SLinus Torvalds 31934de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK; 3194735d3831SFlorian Westphal tcp_ecn_send_synack(sk, skb); 31951da177e4SLinus Torvalds } 3196dfb4b9dcSDavid S. Miller return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 31971da177e4SLinus Torvalds } 31981da177e4SLinus Torvalds 31994aea39c1SEric Dumazet /** 32004aea39c1SEric Dumazet * tcp_make_synack - Prepare a SYN-ACK. 32014aea39c1SEric Dumazet * sk: listener socket 32024aea39c1SEric Dumazet * dst: dst entry attached to the SYNACK 32034aea39c1SEric Dumazet * req: request_sock pointer 32044aea39c1SEric Dumazet * 32054aea39c1SEric Dumazet * Allocate one skb and build a SYNACK packet. 32064aea39c1SEric Dumazet * @dst is consumed : Caller should not use it again. 32074aea39c1SEric Dumazet */ 32085d062de7SEric Dumazet struct sk_buff *tcp_make_synack(const struct sock *sk, struct dst_entry *dst, 3209e6b4d113SWilliam Allen Simpson struct request_sock *req, 3210ca6fb065SEric Dumazet struct tcp_fastopen_cookie *foc, 3211b3d05147SEric Dumazet enum tcp_synack_type synack_type) 32121da177e4SLinus Torvalds { 32132e6599cbSArnaldo Carvalho de Melo struct inet_request_sock *ireq = inet_rsk(req); 32145d062de7SEric Dumazet const struct tcp_sock *tp = tcp_sk(sk); 321580f03e27SEric Dumazet struct tcp_md5sig_key *md5 = NULL; 32165d062de7SEric Dumazet struct tcp_out_options opts; 32175d062de7SEric Dumazet struct sk_buff *skb; 3218bd0388aeSWilliam Allen Simpson int tcp_header_size; 32195d062de7SEric Dumazet struct tcphdr *th; 3220f5fff5dcSTom Quetchenbach int mss; 32211da177e4SLinus Torvalds 3222ca6fb065SEric Dumazet skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC); 32234aea39c1SEric Dumazet if (unlikely(!skb)) { 32244aea39c1SEric Dumazet dst_release(dst); 32251da177e4SLinus Torvalds return NULL; 32264aea39c1SEric Dumazet } 32271da177e4SLinus Torvalds /* Reserve space for headers. */ 32281da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 32291da177e4SLinus Torvalds 3230b3d05147SEric Dumazet switch (synack_type) { 3231b3d05147SEric Dumazet case TCP_SYNACK_NORMAL: 32329e17f8a4SEric Dumazet skb_set_owner_w(skb, req_to_sk(req)); 3233b3d05147SEric Dumazet break; 3234b3d05147SEric Dumazet case TCP_SYNACK_COOKIE: 3235b3d05147SEric Dumazet /* Under synflood, we do not attach skb to a socket, 3236b3d05147SEric Dumazet * to avoid false sharing. 3237b3d05147SEric Dumazet */ 3238b3d05147SEric Dumazet break; 3239b3d05147SEric Dumazet case TCP_SYNACK_FASTOPEN: 3240ca6fb065SEric Dumazet /* sk is a const pointer, because we want to express multiple 3241ca6fb065SEric Dumazet * cpu might call us concurrently. 3242ca6fb065SEric Dumazet * sk->sk_wmem_alloc in an atomic, we can promote to rw. 3243ca6fb065SEric Dumazet */ 3244ca6fb065SEric Dumazet skb_set_owner_w(skb, (struct sock *)sk); 3245b3d05147SEric Dumazet break; 3246ca6fb065SEric Dumazet } 32474aea39c1SEric Dumazet skb_dst_set(skb, dst); 32481da177e4SLinus Torvalds 32493541f9e8SEric Dumazet mss = tcp_mss_clamp(tp, dst_metric_advmss(dst)); 3250f5fff5dcSTom Quetchenbach 325133ad798cSAdam Langley memset(&opts, 0, sizeof(opts)); 32528b5f12d0SFlorian Westphal #ifdef CONFIG_SYN_COOKIES 32538b5f12d0SFlorian Westphal if (unlikely(req->cookie_ts)) 3254d3edd06eSEric Dumazet skb->skb_mstamp_ns = cookie_init_timestamp(req); 32558b5f12d0SFlorian Westphal else 32568b5f12d0SFlorian Westphal #endif 3257d3edd06eSEric Dumazet skb->skb_mstamp_ns = tcp_clock_ns(); 325880f03e27SEric Dumazet 325980f03e27SEric Dumazet #ifdef CONFIG_TCP_MD5SIG 326080f03e27SEric Dumazet rcu_read_lock(); 3261fd3a154aSEric Dumazet md5 = tcp_rsk(req)->af_specific->req_md5_lookup(sk, req_to_sk(req)); 326280f03e27SEric Dumazet #endif 326358d607d3SEric Dumazet skb_set_hash(skb, tcp_rsk(req)->txhash, PKT_HASH_TYPE_L4); 326460e2a778SUrsula Braun tcp_header_size = tcp_synack_options(sk, req, mss, skb, &opts, md5, 326560e2a778SUrsula Braun foc) + sizeof(*th); 326633ad798cSAdam Langley 3267aa8223c7SArnaldo Carvalho de Melo skb_push(skb, tcp_header_size); 3268aa8223c7SArnaldo Carvalho de Melo skb_reset_transport_header(skb); 32691da177e4SLinus Torvalds 3270ea1627c2SEric Dumazet th = (struct tcphdr *)skb->data; 32711da177e4SLinus Torvalds memset(th, 0, sizeof(struct tcphdr)); 32721da177e4SLinus Torvalds th->syn = 1; 32731da177e4SLinus Torvalds th->ack = 1; 32746ac705b1SEric Dumazet tcp_ecn_make_synack(req, th); 3275b44084c2SEric Dumazet th->source = htons(ireq->ir_num); 3276634fb979SEric Dumazet th->dest = ireq->ir_rmt_port; 3277e05a90ecSJamal Hadi Salim skb->mark = ireq->ir_mark; 32783b117750SEric Dumazet skb->ip_summed = CHECKSUM_PARTIAL; 32793b117750SEric Dumazet th->seq = htonl(tcp_rsk(req)->snt_isn); 32808336886fSJerry Chu /* XXX data is queued and acked as is. No buffer/window check */ 32818336886fSJerry Chu th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt); 32821da177e4SLinus Torvalds 32831da177e4SLinus Torvalds /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */ 3284ed53d0abSEric Dumazet th->window = htons(min(req->rsk_rcv_wnd, 65535U)); 32855d062de7SEric Dumazet tcp_options_write((__be32 *)(th + 1), NULL, &opts); 32861da177e4SLinus Torvalds th->doff = (tcp_header_size >> 2); 328790bbcc60SEric Dumazet __TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTSEGS); 3288cfb6eeb4SYOSHIFUJI Hideaki 3289cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 3290cfb6eeb4SYOSHIFUJI Hideaki /* Okay, we have all we need - do the md5 hash if needed */ 329180f03e27SEric Dumazet if (md5) 3292bd0388aeSWilliam Allen Simpson tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location, 329339f8e58eSEric Dumazet md5, req_to_sk(req), skb); 329480f03e27SEric Dumazet rcu_read_unlock(); 3295cfb6eeb4SYOSHIFUJI Hideaki #endif 3296cfb6eeb4SYOSHIFUJI Hideaki 3297b50edd78SEric Dumazet /* Do not fool tcpdump (if any), clean our debris */ 32982456e855SThomas Gleixner skb->tstamp = 0; 32991da177e4SLinus Torvalds return skb; 33001da177e4SLinus Torvalds } 33014bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_make_synack); 33021da177e4SLinus Torvalds 330381164413SDaniel Borkmann static void tcp_ca_dst_init(struct sock *sk, const struct dst_entry *dst) 330481164413SDaniel Borkmann { 330581164413SDaniel Borkmann struct inet_connection_sock *icsk = inet_csk(sk); 330681164413SDaniel Borkmann const struct tcp_congestion_ops *ca; 330781164413SDaniel Borkmann u32 ca_key = dst_metric(dst, RTAX_CC_ALGO); 330881164413SDaniel Borkmann 330981164413SDaniel Borkmann if (ca_key == TCP_CA_UNSPEC) 331081164413SDaniel Borkmann return; 331181164413SDaniel Borkmann 331281164413SDaniel Borkmann rcu_read_lock(); 331381164413SDaniel Borkmann ca = tcp_ca_find_key(ca_key); 331481164413SDaniel Borkmann if (likely(ca && try_module_get(ca->owner))) { 331581164413SDaniel Borkmann module_put(icsk->icsk_ca_ops->owner); 331681164413SDaniel Borkmann icsk->icsk_ca_dst_locked = tcp_ca_dst_locked(dst); 331781164413SDaniel Borkmann icsk->icsk_ca_ops = ca; 331881164413SDaniel Borkmann } 331981164413SDaniel Borkmann rcu_read_unlock(); 332081164413SDaniel Borkmann } 332181164413SDaniel Borkmann 332267edfef7SAndi Kleen /* Do all connect socket setups that can be done AF independent. */ 3323f7e56a76Sstephen hemminger static void tcp_connect_init(struct sock *sk) 33241da177e4SLinus Torvalds { 3325cf533ea5SEric Dumazet const struct dst_entry *dst = __sk_dst_get(sk); 33261da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 33271da177e4SLinus Torvalds __u8 rcv_wscale; 332813d3b1ebSLawrence Brakmo u32 rcv_wnd; 33291da177e4SLinus Torvalds 33301da177e4SLinus Torvalds /* We'll fix this up when we get a response from the other end. 33311da177e4SLinus Torvalds * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT. 33321da177e4SLinus Torvalds */ 33335d2ed052SEric Dumazet tp->tcp_header_len = sizeof(struct tcphdr); 33345d2ed052SEric Dumazet if (sock_net(sk)->ipv4.sysctl_tcp_timestamps) 33355d2ed052SEric Dumazet tp->tcp_header_len += TCPOLEN_TSTAMP_ALIGNED; 33361da177e4SLinus Torvalds 3337cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 333800db4124SIan Morris if (tp->af_specific->md5_lookup(sk, sk)) 3339cfb6eeb4SYOSHIFUJI Hideaki tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED; 3340cfb6eeb4SYOSHIFUJI Hideaki #endif 3341cfb6eeb4SYOSHIFUJI Hideaki 33421da177e4SLinus Torvalds /* If user gave his TCP_MAXSEG, record it to clamp */ 33431da177e4SLinus Torvalds if (tp->rx_opt.user_mss) 33441da177e4SLinus Torvalds tp->rx_opt.mss_clamp = tp->rx_opt.user_mss; 33451da177e4SLinus Torvalds tp->max_window = 0; 33465d424d5aSJohn Heffner tcp_mtup_init(sk); 33471da177e4SLinus Torvalds tcp_sync_mss(sk, dst_mtu(dst)); 33481da177e4SLinus Torvalds 334981164413SDaniel Borkmann tcp_ca_dst_init(sk, dst); 335081164413SDaniel Borkmann 33511da177e4SLinus Torvalds if (!tp->window_clamp) 33521da177e4SLinus Torvalds tp->window_clamp = dst_metric(dst, RTAX_WINDOW); 33533541f9e8SEric Dumazet tp->advmss = tcp_mss_clamp(tp, dst_metric_advmss(dst)); 3354f5fff5dcSTom Quetchenbach 33551da177e4SLinus Torvalds tcp_initialize_rcv_mss(sk); 33561da177e4SLinus Torvalds 3357e88c64f0SHagen Paul Pfeifer /* limit the window selection if the user enforce a smaller rx buffer */ 3358e88c64f0SHagen Paul Pfeifer if (sk->sk_userlocks & SOCK_RCVBUF_LOCK && 3359e88c64f0SHagen Paul Pfeifer (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0)) 3360e88c64f0SHagen Paul Pfeifer tp->window_clamp = tcp_full_space(sk); 3361e88c64f0SHagen Paul Pfeifer 336213d3b1ebSLawrence Brakmo rcv_wnd = tcp_rwnd_init_bpf(sk); 336313d3b1ebSLawrence Brakmo if (rcv_wnd == 0) 336413d3b1ebSLawrence Brakmo rcv_wnd = dst_metric(dst, RTAX_INITRWND); 336513d3b1ebSLawrence Brakmo 3366ceef9ab6SEric Dumazet tcp_select_initial_window(sk, tcp_full_space(sk), 33671da177e4SLinus Torvalds tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0), 33681da177e4SLinus Torvalds &tp->rcv_wnd, 33691da177e4SLinus Torvalds &tp->window_clamp, 33709bb37ef0SEric Dumazet sock_net(sk)->ipv4.sysctl_tcp_window_scaling, 337131d12926Slaurent chavey &rcv_wscale, 337213d3b1ebSLawrence Brakmo rcv_wnd); 33731da177e4SLinus Torvalds 33741da177e4SLinus Torvalds tp->rx_opt.rcv_wscale = rcv_wscale; 33751da177e4SLinus Torvalds tp->rcv_ssthresh = tp->rcv_wnd; 33761da177e4SLinus Torvalds 33771da177e4SLinus Torvalds sk->sk_err = 0; 33781da177e4SLinus Torvalds sock_reset_flag(sk, SOCK_DONE); 33791da177e4SLinus Torvalds tp->snd_wnd = 0; 3380ee7537b6SHantzis Fotis tcp_init_wl(tp, 0); 33817f582b24SEric Dumazet tcp_write_queue_purge(sk); 33821da177e4SLinus Torvalds tp->snd_una = tp->write_seq; 33831da177e4SLinus Torvalds tp->snd_sml = tp->write_seq; 338433f5f57eSIlpo Järvinen tp->snd_up = tp->write_seq; 3385370816aeSPavel Emelyanov tp->snd_nxt = tp->write_seq; 3386ee995283SPavel Emelyanov 3387ee995283SPavel Emelyanov if (likely(!tp->repair)) 33881da177e4SLinus Torvalds tp->rcv_nxt = 0; 3389c7781a6eSAndrew Vagin else 339070eabf0eSEric Dumazet tp->rcv_tstamp = tcp_jiffies32; 3391ee995283SPavel Emelyanov tp->rcv_wup = tp->rcv_nxt; 3392ee995283SPavel Emelyanov tp->copied_seq = tp->rcv_nxt; 33931da177e4SLinus Torvalds 33948550f328SLawrence Brakmo inet_csk(sk)->icsk_rto = tcp_timeout_init(sk); 3395463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_retransmits = 0; 33961da177e4SLinus Torvalds tcp_clear_retrans(tp); 33971da177e4SLinus Torvalds } 33981da177e4SLinus Torvalds 3399783237e8SYuchung Cheng static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb) 3400783237e8SYuchung Cheng { 3401783237e8SYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 3402783237e8SYuchung Cheng struct tcp_skb_cb *tcb = TCP_SKB_CB(skb); 3403783237e8SYuchung Cheng 3404783237e8SYuchung Cheng tcb->end_seq += skb->len; 3405f4a775d1SEric Dumazet __skb_header_release(skb); 3406783237e8SYuchung Cheng sk->sk_wmem_queued += skb->truesize; 3407783237e8SYuchung Cheng sk_mem_charge(sk, skb->truesize); 3408783237e8SYuchung Cheng tp->write_seq = tcb->end_seq; 3409783237e8SYuchung Cheng tp->packets_out += tcp_skb_pcount(skb); 3410783237e8SYuchung Cheng } 3411783237e8SYuchung Cheng 3412783237e8SYuchung Cheng /* Build and send a SYN with data and (cached) Fast Open cookie. However, 3413783237e8SYuchung Cheng * queue a data-only packet after the regular SYN, such that regular SYNs 3414783237e8SYuchung Cheng * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges 3415783237e8SYuchung Cheng * only the SYN sequence, the data are retransmitted in the first ACK. 3416783237e8SYuchung Cheng * If cookie is not cached or other error occurs, falls back to send a 3417783237e8SYuchung Cheng * regular SYN with Fast Open cookie request option. 3418783237e8SYuchung Cheng */ 3419783237e8SYuchung Cheng static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn) 3420783237e8SYuchung Cheng { 3421783237e8SYuchung Cheng struct tcp_sock *tp = tcp_sk(sk); 3422783237e8SYuchung Cheng struct tcp_fastopen_request *fo = tp->fastopen_req; 3423065263f4SWei Wang int space, err = 0; 3424355a901eSEric Dumazet struct sk_buff *syn_data; 3425783237e8SYuchung Cheng 342667da22d2SYuchung Cheng tp->rx_opt.mss_clamp = tp->advmss; /* If MSS is not cached */ 3427065263f4SWei Wang if (!tcp_fastopen_cookie_check(sk, &tp->rx_opt.mss_clamp, &fo->cookie)) 3428783237e8SYuchung Cheng goto fallback; 3429783237e8SYuchung Cheng 3430783237e8SYuchung Cheng /* MSS for SYN-data is based on cached MSS and bounded by PMTU and 3431783237e8SYuchung Cheng * user-MSS. Reserve maximum option space for middleboxes that add 3432783237e8SYuchung Cheng * private TCP options. The cost is reduced data space in SYN :( 3433783237e8SYuchung Cheng */ 34343541f9e8SEric Dumazet tp->rx_opt.mss_clamp = tcp_mss_clamp(tp, tp->rx_opt.mss_clamp); 34353541f9e8SEric Dumazet 34361b63edd6SYuchung Cheng space = __tcp_mtu_to_mss(sk, inet_csk(sk)->icsk_pmtu_cookie) - 3437783237e8SYuchung Cheng MAX_TCP_OPTION_SPACE; 3438783237e8SYuchung Cheng 3439f5ddcbbbSEric Dumazet space = min_t(size_t, space, fo->size); 3440f5ddcbbbSEric Dumazet 3441f5ddcbbbSEric Dumazet /* limit to order-0 allocations */ 3442f5ddcbbbSEric Dumazet space = min_t(size_t, space, SKB_MAX_HEAD(MAX_TCP_HEADER)); 3443f5ddcbbbSEric Dumazet 3444eb934478SEric Dumazet syn_data = sk_stream_alloc_skb(sk, space, sk->sk_allocation, false); 3445355a901eSEric Dumazet if (!syn_data) 3446783237e8SYuchung Cheng goto fallback; 3447355a901eSEric Dumazet syn_data->ip_summed = CHECKSUM_PARTIAL; 3448355a901eSEric Dumazet memcpy(syn_data->cb, syn->cb, sizeof(syn->cb)); 344907e100f9SEric Dumazet if (space) { 345007e100f9SEric Dumazet int copied = copy_from_iter(skb_put(syn_data, space), space, 345157be5bdaSAl Viro &fo->data->msg_iter); 345257be5bdaSAl Viro if (unlikely(!copied)) { 3453ba233b34SEric Dumazet tcp_skb_tsorted_anchor_cleanup(syn_data); 3454355a901eSEric Dumazet kfree_skb(syn_data); 3455783237e8SYuchung Cheng goto fallback; 3456783237e8SYuchung Cheng } 345757be5bdaSAl Viro if (copied != space) { 345857be5bdaSAl Viro skb_trim(syn_data, copied); 345957be5bdaSAl Viro space = copied; 346057be5bdaSAl Viro } 3461f859a448SWillem de Bruijn skb_zcopy_set(syn_data, fo->uarg, NULL); 346207e100f9SEric Dumazet } 3463355a901eSEric Dumazet /* No more data pending in inet_wait_for_connect() */ 3464355a901eSEric Dumazet if (space == fo->size) 3465355a901eSEric Dumazet fo->data = NULL; 3466355a901eSEric Dumazet fo->copied = space; 3467783237e8SYuchung Cheng 3468355a901eSEric Dumazet tcp_connect_queue_skb(sk, syn_data); 34690f87230dSFrancis Yan if (syn_data->len) 34700f87230dSFrancis Yan tcp_chrono_start(sk, TCP_CHRONO_BUSY); 3471355a901eSEric Dumazet 3472355a901eSEric Dumazet err = tcp_transmit_skb(sk, syn_data, 1, sk->sk_allocation); 3473355a901eSEric Dumazet 3474d3edd06eSEric Dumazet syn->skb_mstamp_ns = syn_data->skb_mstamp_ns; 3475355a901eSEric Dumazet 3476355a901eSEric Dumazet /* Now full SYN+DATA was cloned and sent (or not), 3477355a901eSEric Dumazet * remove the SYN from the original skb (syn_data) 3478355a901eSEric Dumazet * we keep in write queue in case of a retransmit, as we 3479355a901eSEric Dumazet * also have the SYN packet (with no data) in the same queue. 3480431a9124SEric Dumazet */ 3481355a901eSEric Dumazet TCP_SKB_CB(syn_data)->seq++; 3482355a901eSEric Dumazet TCP_SKB_CB(syn_data)->tcp_flags = TCPHDR_ACK | TCPHDR_PSH; 3483355a901eSEric Dumazet if (!err) { 348467da22d2SYuchung Cheng tp->syn_data = (fo->copied > 0); 348575c119afSEric Dumazet tcp_rbtree_insert(&sk->tcp_rtx_queue, syn_data); 3486f19c29e3SYuchung Cheng NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT); 3487783237e8SYuchung Cheng goto done; 3488783237e8SYuchung Cheng } 3489783237e8SYuchung Cheng 349075c119afSEric Dumazet /* data was not sent, put it in write_queue */ 349175c119afSEric Dumazet __skb_queue_tail(&sk->sk_write_queue, syn_data); 3492b5b7db8dSEric Dumazet tp->packets_out -= tcp_skb_pcount(syn_data); 3493b5b7db8dSEric Dumazet 3494783237e8SYuchung Cheng fallback: 3495783237e8SYuchung Cheng /* Send a regular SYN with Fast Open cookie request option */ 3496783237e8SYuchung Cheng if (fo->cookie.len > 0) 3497783237e8SYuchung Cheng fo->cookie.len = 0; 3498783237e8SYuchung Cheng err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation); 3499783237e8SYuchung Cheng if (err) 3500783237e8SYuchung Cheng tp->syn_fastopen = 0; 3501783237e8SYuchung Cheng done: 3502783237e8SYuchung Cheng fo->cookie.len = -1; /* Exclude Fast Open option for SYN retries */ 3503783237e8SYuchung Cheng return err; 3504783237e8SYuchung Cheng } 3505783237e8SYuchung Cheng 350667edfef7SAndi Kleen /* Build a SYN and send it off. */ 35071da177e4SLinus Torvalds int tcp_connect(struct sock *sk) 35081da177e4SLinus Torvalds { 35091da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 35101da177e4SLinus Torvalds struct sk_buff *buff; 3511ee586811SEric Paris int err; 35121da177e4SLinus Torvalds 3513de525be2SLawrence Brakmo tcp_call_bpf(sk, BPF_SOCK_OPS_TCP_CONNECT_CB, 0, NULL); 35148ba60924SEric Dumazet 35158ba60924SEric Dumazet if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk)) 35168ba60924SEric Dumazet return -EHOSTUNREACH; /* Routing failure or similar. */ 35178ba60924SEric Dumazet 35181da177e4SLinus Torvalds tcp_connect_init(sk); 35191da177e4SLinus Torvalds 35202b916477SAndrey Vagin if (unlikely(tp->repair)) { 35212b916477SAndrey Vagin tcp_finish_connect(sk, NULL); 35222b916477SAndrey Vagin return 0; 35232b916477SAndrey Vagin } 35242b916477SAndrey Vagin 3525eb934478SEric Dumazet buff = sk_stream_alloc_skb(sk, 0, sk->sk_allocation, true); 3526355a901eSEric Dumazet if (unlikely(!buff)) 35271da177e4SLinus Torvalds return -ENOBUFS; 35281da177e4SLinus Torvalds 3529a3433f35SChangli Gao tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN); 35309a568de4SEric Dumazet tcp_mstamp_refresh(tp); 35319a568de4SEric Dumazet tp->retrans_stamp = tcp_time_stamp(tp); 3532783237e8SYuchung Cheng tcp_connect_queue_skb(sk, buff); 3533735d3831SFlorian Westphal tcp_ecn_send_syn(sk, buff); 353475c119afSEric Dumazet tcp_rbtree_insert(&sk->tcp_rtx_queue, buff); 35351da177e4SLinus Torvalds 3536783237e8SYuchung Cheng /* Send off SYN; include data in Fast Open. */ 3537783237e8SYuchung Cheng err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) : 3538783237e8SYuchung Cheng tcp_transmit_skb(sk, buff, 1, sk->sk_allocation); 3539ee586811SEric Paris if (err == -ECONNREFUSED) 3540ee586811SEric Paris return err; 3541bd37a088SWei Yongjun 3542bd37a088SWei Yongjun /* We change tp->snd_nxt after the tcp_transmit_skb() call 3543bd37a088SWei Yongjun * in order to make this packet get counted in tcpOutSegs. 3544bd37a088SWei Yongjun */ 3545bd37a088SWei Yongjun tp->snd_nxt = tp->write_seq; 3546bd37a088SWei Yongjun tp->pushed_seq = tp->write_seq; 3547b5b7db8dSEric Dumazet buff = tcp_send_head(sk); 3548b5b7db8dSEric Dumazet if (unlikely(buff)) { 3549b5b7db8dSEric Dumazet tp->snd_nxt = TCP_SKB_CB(buff)->seq; 3550b5b7db8dSEric Dumazet tp->pushed_seq = TCP_SKB_CB(buff)->seq; 3551b5b7db8dSEric Dumazet } 355281cc8a75SPavel Emelyanov TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS); 35531da177e4SLinus Torvalds 35541da177e4SLinus Torvalds /* Timer for repeating the SYN until an answer. */ 35553f421baaSArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 35563f421baaSArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto, TCP_RTO_MAX); 35571da177e4SLinus Torvalds return 0; 35581da177e4SLinus Torvalds } 35594bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_connect); 35601da177e4SLinus Torvalds 35611da177e4SLinus Torvalds /* Send out a delayed ack, the caller does the policy checking 35621da177e4SLinus Torvalds * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check() 35631da177e4SLinus Torvalds * for details. 35641da177e4SLinus Torvalds */ 35651da177e4SLinus Torvalds void tcp_send_delayed_ack(struct sock *sk) 35661da177e4SLinus Torvalds { 3567463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 3568463c84b9SArnaldo Carvalho de Melo int ato = icsk->icsk_ack.ato; 35691da177e4SLinus Torvalds unsigned long timeout; 35701da177e4SLinus Torvalds 35711da177e4SLinus Torvalds if (ato > TCP_DELACK_MIN) { 3572463c84b9SArnaldo Carvalho de Melo const struct tcp_sock *tp = tcp_sk(sk); 35731da177e4SLinus Torvalds int max_ato = HZ / 2; 35741da177e4SLinus Torvalds 357531954cd8SWei Wang if (inet_csk_in_pingpong_mode(sk) || 3576056834d9SIlpo Järvinen (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)) 35771da177e4SLinus Torvalds max_ato = TCP_DELACK_MAX; 35781da177e4SLinus Torvalds 35791da177e4SLinus Torvalds /* Slow path, intersegment interval is "high". */ 35801da177e4SLinus Torvalds 35811da177e4SLinus Torvalds /* If some rtt estimate is known, use it to bound delayed ack. 3582463c84b9SArnaldo Carvalho de Melo * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements 35831da177e4SLinus Torvalds * directly. 35841da177e4SLinus Torvalds */ 3585740b0f18SEric Dumazet if (tp->srtt_us) { 3586740b0f18SEric Dumazet int rtt = max_t(int, usecs_to_jiffies(tp->srtt_us >> 3), 3587740b0f18SEric Dumazet TCP_DELACK_MIN); 35881da177e4SLinus Torvalds 35891da177e4SLinus Torvalds if (rtt < max_ato) 35901da177e4SLinus Torvalds max_ato = rtt; 35911da177e4SLinus Torvalds } 35921da177e4SLinus Torvalds 35931da177e4SLinus Torvalds ato = min(ato, max_ato); 35941da177e4SLinus Torvalds } 35951da177e4SLinus Torvalds 35961da177e4SLinus Torvalds /* Stay within the limit we were given */ 35971da177e4SLinus Torvalds timeout = jiffies + ato; 35981da177e4SLinus Torvalds 35991da177e4SLinus Torvalds /* Use new timeout only if there wasn't a older one earlier. */ 3600463c84b9SArnaldo Carvalho de Melo if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) { 36011da177e4SLinus Torvalds /* If delack timer was blocked or is about to expire, 36021da177e4SLinus Torvalds * send ACK now. 36031da177e4SLinus Torvalds */ 3604463c84b9SArnaldo Carvalho de Melo if (icsk->icsk_ack.blocked || 3605463c84b9SArnaldo Carvalho de Melo time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) { 36061da177e4SLinus Torvalds tcp_send_ack(sk); 36071da177e4SLinus Torvalds return; 36081da177e4SLinus Torvalds } 36091da177e4SLinus Torvalds 3610463c84b9SArnaldo Carvalho de Melo if (!time_before(timeout, icsk->icsk_ack.timeout)) 3611463c84b9SArnaldo Carvalho de Melo timeout = icsk->icsk_ack.timeout; 36121da177e4SLinus Torvalds } 3613463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER; 3614463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.timeout = timeout; 3615463c84b9SArnaldo Carvalho de Melo sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout); 36161da177e4SLinus Torvalds } 36171da177e4SLinus Torvalds 36181da177e4SLinus Torvalds /* This routine sends an ack and also updates the window. */ 36192987babbSYuchung Cheng void __tcp_send_ack(struct sock *sk, u32 rcv_nxt) 36201da177e4SLinus Torvalds { 36211da177e4SLinus Torvalds struct sk_buff *buff; 36221da177e4SLinus Torvalds 3623058dc334SIlpo Järvinen /* If we have been reset, we may not send again. */ 3624058dc334SIlpo Järvinen if (sk->sk_state == TCP_CLOSE) 3625058dc334SIlpo Järvinen return; 3626058dc334SIlpo Järvinen 36271da177e4SLinus Torvalds /* We are not putting this on the write queue, so 36281da177e4SLinus Torvalds * tcp_transmit_skb() will set the ownership to this 36291da177e4SLinus Torvalds * sock. 36301da177e4SLinus Torvalds */ 36317450aaf6SEric Dumazet buff = alloc_skb(MAX_TCP_HEADER, 36327450aaf6SEric Dumazet sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN)); 36337450aaf6SEric Dumazet if (unlikely(!buff)) { 3634463c84b9SArnaldo Carvalho de Melo inet_csk_schedule_ack(sk); 3635463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN; 36363f421baaSArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK, 36373f421baaSArnaldo Carvalho de Melo TCP_DELACK_MAX, TCP_RTO_MAX); 36381da177e4SLinus Torvalds return; 36391da177e4SLinus Torvalds } 36401da177e4SLinus Torvalds 36411da177e4SLinus Torvalds /* Reserve space for headers and prepare control bits. */ 36421da177e4SLinus Torvalds skb_reserve(buff, MAX_TCP_HEADER); 3643a3433f35SChangli Gao tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK); 36441da177e4SLinus Torvalds 364598781965SEric Dumazet /* We do not want pure acks influencing TCP Small Queues or fq/pacing 364698781965SEric Dumazet * too much. 364798781965SEric Dumazet * SKB_TRUESIZE(max(1 .. 66, MAX_TCP_HEADER)) is unfortunately ~784 364898781965SEric Dumazet */ 364998781965SEric Dumazet skb_set_tcp_pure_ack(buff); 365098781965SEric Dumazet 36511da177e4SLinus Torvalds /* Send it off, this clears delayed acks for us. */ 36522987babbSYuchung Cheng __tcp_transmit_skb(sk, buff, 0, (__force gfp_t)0, rcv_nxt); 36531da177e4SLinus Torvalds } 365427cde44aSYuchung Cheng EXPORT_SYMBOL_GPL(__tcp_send_ack); 36552987babbSYuchung Cheng 36562987babbSYuchung Cheng void tcp_send_ack(struct sock *sk) 36572987babbSYuchung Cheng { 36582987babbSYuchung Cheng __tcp_send_ack(sk, tcp_sk(sk)->rcv_nxt); 36591da177e4SLinus Torvalds } 36601da177e4SLinus Torvalds 36611da177e4SLinus Torvalds /* This routine sends a packet with an out of date sequence 36621da177e4SLinus Torvalds * number. It assumes the other end will try to ack it. 36631da177e4SLinus Torvalds * 36641da177e4SLinus Torvalds * Question: what should we make while urgent mode? 36651da177e4SLinus Torvalds * 4.4BSD forces sending single byte of data. We cannot send 36661da177e4SLinus Torvalds * out of window data, because we have SND.NXT==SND.MAX... 36671da177e4SLinus Torvalds * 36681da177e4SLinus Torvalds * Current solution: to send TWO zero-length segments in urgent mode: 36691da177e4SLinus Torvalds * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is 36701da177e4SLinus Torvalds * out-of-date with SND.UNA-1 to probe window. 36711da177e4SLinus Torvalds */ 3672e520af48SEric Dumazet static int tcp_xmit_probe_skb(struct sock *sk, int urgent, int mib) 36731da177e4SLinus Torvalds { 36741da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 36751da177e4SLinus Torvalds struct sk_buff *skb; 36761da177e4SLinus Torvalds 36771da177e4SLinus Torvalds /* We don't queue it, tcp_transmit_skb() sets ownership. */ 36787450aaf6SEric Dumazet skb = alloc_skb(MAX_TCP_HEADER, 36797450aaf6SEric Dumazet sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN)); 368051456b29SIan Morris if (!skb) 36811da177e4SLinus Torvalds return -1; 36821da177e4SLinus Torvalds 36831da177e4SLinus Torvalds /* Reserve space for headers and set control bits. */ 36841da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 36851da177e4SLinus Torvalds /* Use a previous sequence. This should cause the other 36861da177e4SLinus Torvalds * end to send an ack. Don't queue or clone SKB, just 36871da177e4SLinus Torvalds * send it. 36881da177e4SLinus Torvalds */ 3689a3433f35SChangli Gao tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK); 3690e2e8009fSRenato Westphal NET_INC_STATS(sock_net(sk), mib); 36917450aaf6SEric Dumazet return tcp_transmit_skb(sk, skb, 0, (__force gfp_t)0); 36921da177e4SLinus Torvalds } 36931da177e4SLinus Torvalds 3694385e2070SEric Dumazet /* Called from setsockopt( ... TCP_REPAIR ) */ 3695ee995283SPavel Emelyanov void tcp_send_window_probe(struct sock *sk) 3696ee995283SPavel Emelyanov { 3697ee995283SPavel Emelyanov if (sk->sk_state == TCP_ESTABLISHED) { 3698ee995283SPavel Emelyanov tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1; 36999a568de4SEric Dumazet tcp_mstamp_refresh(tcp_sk(sk)); 3700e520af48SEric Dumazet tcp_xmit_probe_skb(sk, 0, LINUX_MIB_TCPWINPROBE); 3701ee995283SPavel Emelyanov } 3702ee995283SPavel Emelyanov } 3703ee995283SPavel Emelyanov 370467edfef7SAndi Kleen /* Initiate keepalive or window probe from timer. */ 3705e520af48SEric Dumazet int tcp_write_wakeup(struct sock *sk, int mib) 37061da177e4SLinus Torvalds { 37071da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 37081da177e4SLinus Torvalds struct sk_buff *skb; 37091da177e4SLinus Torvalds 3710058dc334SIlpo Järvinen if (sk->sk_state == TCP_CLOSE) 3711058dc334SIlpo Järvinen return -1; 3712058dc334SIlpo Järvinen 371300db4124SIan Morris skb = tcp_send_head(sk); 371400db4124SIan Morris if (skb && before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) { 37151da177e4SLinus Torvalds int err; 37160c54b85fSIlpo Järvinen unsigned int mss = tcp_current_mss(sk); 371790840defSIlpo Järvinen unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 37181da177e4SLinus Torvalds 37191da177e4SLinus Torvalds if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq)) 37201da177e4SLinus Torvalds tp->pushed_seq = TCP_SKB_CB(skb)->end_seq; 37211da177e4SLinus Torvalds 37221da177e4SLinus Torvalds /* We are probing the opening of a window 37231da177e4SLinus Torvalds * but the window size is != 0 37241da177e4SLinus Torvalds * must have been a result SWS avoidance ( sender ) 37251da177e4SLinus Torvalds */ 37261da177e4SLinus Torvalds if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq || 37271da177e4SLinus Torvalds skb->len > mss) { 37281da177e4SLinus Torvalds seg_size = min(seg_size, mss); 37294de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; 373075c119afSEric Dumazet if (tcp_fragment(sk, TCP_FRAG_IN_WRITE_QUEUE, 373175c119afSEric Dumazet skb, seg_size, mss, GFP_ATOMIC)) 37321da177e4SLinus Torvalds return -1; 37331da177e4SLinus Torvalds } else if (!tcp_skb_pcount(skb)) 37345bbb432cSEric Dumazet tcp_set_skb_tso_segs(skb, mss); 37351da177e4SLinus Torvalds 37364de075e0SEric Dumazet TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; 3737dfb4b9dcSDavid S. Miller err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 373866f5fe62SIlpo Järvinen if (!err) 373966f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, skb); 37401da177e4SLinus Torvalds return err; 37411da177e4SLinus Torvalds } else { 374233f5f57eSIlpo Järvinen if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF)) 3743e520af48SEric Dumazet tcp_xmit_probe_skb(sk, 1, mib); 3744e520af48SEric Dumazet return tcp_xmit_probe_skb(sk, 0, mib); 37451da177e4SLinus Torvalds } 37461da177e4SLinus Torvalds } 37471da177e4SLinus Torvalds 37481da177e4SLinus Torvalds /* A window probe timeout has occurred. If window is not closed send 37491da177e4SLinus Torvalds * a partial packet else a zero probe. 37501da177e4SLinus Torvalds */ 37511da177e4SLinus Torvalds void tcp_send_probe0(struct sock *sk) 37521da177e4SLinus Torvalds { 3753463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 37541da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 3755c6214a97SNikolay Borisov struct net *net = sock_net(sk); 3756c1d5674fSYuchung Cheng unsigned long timeout; 37571da177e4SLinus Torvalds int err; 37581da177e4SLinus Torvalds 3759e520af48SEric Dumazet err = tcp_write_wakeup(sk, LINUX_MIB_TCPWINPROBE); 37601da177e4SLinus Torvalds 376175c119afSEric Dumazet if (tp->packets_out || tcp_write_queue_empty(sk)) { 37621da177e4SLinus Torvalds /* Cancel probe timer, if it is not required. */ 37636687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out = 0; 3764463c84b9SArnaldo Carvalho de Melo icsk->icsk_backoff = 0; 37651da177e4SLinus Torvalds return; 37661da177e4SLinus Torvalds } 37671da177e4SLinus Torvalds 3768c1d5674fSYuchung Cheng icsk->icsk_probes_out++; 37691da177e4SLinus Torvalds if (err <= 0) { 3770c6214a97SNikolay Borisov if (icsk->icsk_backoff < net->ipv4.sysctl_tcp_retries2) 3771463c84b9SArnaldo Carvalho de Melo icsk->icsk_backoff++; 3772c1d5674fSYuchung Cheng timeout = tcp_probe0_when(sk, TCP_RTO_MAX); 37731da177e4SLinus Torvalds } else { 37741da177e4SLinus Torvalds /* If packet was not sent due to local congestion, 3775c1d5674fSYuchung Cheng * Let senders fight for local resources conservatively. 37761da177e4SLinus Torvalds */ 3777c1d5674fSYuchung Cheng timeout = TCP_RESOURCE_PROBE_INTERVAL; 37781da177e4SLinus Torvalds } 3779c1d5674fSYuchung Cheng tcp_reset_xmit_timer(sk, ICSK_TIME_PROBE0, timeout, TCP_RTO_MAX, NULL); 37801da177e4SLinus Torvalds } 37815db92c99SOctavian Purdila 3782ea3bea3aSEric Dumazet int tcp_rtx_synack(const struct sock *sk, struct request_sock *req) 37835db92c99SOctavian Purdila { 37845db92c99SOctavian Purdila const struct tcp_request_sock_ops *af_ops = tcp_rsk(req)->af_specific; 37855db92c99SOctavian Purdila struct flowi fl; 37865db92c99SOctavian Purdila int res; 37875db92c99SOctavian Purdila 378858d607d3SEric Dumazet tcp_rsk(req)->txhash = net_tx_rndhash(); 3789b3d05147SEric Dumazet res = af_ops->send_synack(sk, NULL, &fl, req, NULL, TCP_SYNACK_NORMAL); 37905db92c99SOctavian Purdila if (!res) { 379190bbcc60SEric Dumazet __TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS); 379202a1d6e7SEric Dumazet __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS); 37937e32b443SYuchung Cheng if (unlikely(tcp_passive_fastopen(sk))) 37947e32b443SYuchung Cheng tcp_sk(sk)->total_retrans++; 3795cf34ce3dSSong Liu trace_tcp_retransmit_synack(sk, req); 37965db92c99SOctavian Purdila } 37975db92c99SOctavian Purdila return res; 37985db92c99SOctavian Purdila } 37995db92c99SOctavian Purdila EXPORT_SYMBOL(tcp_rtx_synack); 3800