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 371da177e4SLinus Torvalds #include <net/tcp.h> 381da177e4SLinus Torvalds 391da177e4SLinus Torvalds #include <linux/compiler.h> 405a0e3ad6STejun Heo #include <linux/gfp.h> 411da177e4SLinus Torvalds #include <linux/module.h> 421da177e4SLinus Torvalds 431da177e4SLinus Torvalds /* People can turn this off for buggy TCP's found in printers etc. */ 44ab32ea5dSBrian Haley int sysctl_tcp_retrans_collapse __read_mostly = 1; 451da177e4SLinus Torvalds 4615d99e02SRick Jones /* People can turn this on to work with those rare, broken TCPs that 4715d99e02SRick Jones * interpret the window field as a signed quantity. 4815d99e02SRick Jones */ 49ab32ea5dSBrian Haley int sysctl_tcp_workaround_signed_windows __read_mostly = 0; 5015d99e02SRick Jones 511da177e4SLinus Torvalds /* This limits the percentage of the congestion window which we 521da177e4SLinus Torvalds * will allow a single TSO frame to consume. Building TSO frames 531da177e4SLinus Torvalds * which are too large can cause TCP streams to be bursty. 541da177e4SLinus Torvalds */ 55ab32ea5dSBrian Haley int sysctl_tcp_tso_win_divisor __read_mostly = 3; 561da177e4SLinus Torvalds 57ab32ea5dSBrian Haley int sysctl_tcp_mtu_probing __read_mostly = 0; 58ab32ea5dSBrian Haley int sysctl_tcp_base_mss __read_mostly = 512; 595d424d5aSJohn Heffner 6035089bb2SDavid S. Miller /* By default, RFC2861 behavior. */ 61ab32ea5dSBrian Haley int sysctl_tcp_slow_start_after_idle __read_mostly = 1; 6235089bb2SDavid S. Miller 63519855c5SWilliam Allen Simpson int sysctl_tcp_cookie_size __read_mostly = 0; /* TCP_COOKIE_MAX */ 64e6b09ccaSDavid S. Miller EXPORT_SYMBOL_GPL(sysctl_tcp_cookie_size); 65519855c5SWilliam Allen Simpson 66519855c5SWilliam Allen Simpson 6767edfef7SAndi Kleen /* Account for new data that has been sent to the network. */ 6866f5fe62SIlpo Järvinen static void tcp_event_new_data_sent(struct sock *sk, struct sk_buff *skb) 696ff03ac3SIlpo Järvinen { 706ff03ac3SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 7166f5fe62SIlpo Järvinen unsigned int prior_packets = tp->packets_out; 729e412ba7SIlpo Järvinen 73fe067e8aSDavid S. Miller tcp_advance_send_head(sk, skb); 741da177e4SLinus Torvalds tp->snd_nxt = TCP_SKB_CB(skb)->end_seq; 758512430eSIlpo Järvinen 768512430eSIlpo Järvinen /* Don't override Nagle indefinately with F-RTO */ 778512430eSIlpo Järvinen if (tp->frto_counter == 2) 788512430eSIlpo Järvinen tp->frto_counter = 3; 7966f5fe62SIlpo Järvinen 8066f5fe62SIlpo Järvinen tp->packets_out += tcp_skb_pcount(skb); 8166f5fe62SIlpo Järvinen if (!prior_packets) 8266f5fe62SIlpo Järvinen inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 8366f5fe62SIlpo Järvinen inet_csk(sk)->icsk_rto, TCP_RTO_MAX); 841da177e4SLinus Torvalds } 851da177e4SLinus Torvalds 861da177e4SLinus Torvalds /* SND.NXT, if window was not shrunk. 871da177e4SLinus Torvalds * If window has been shrunk, what should we make? It is not clear at all. 881da177e4SLinus Torvalds * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-( 891da177e4SLinus Torvalds * Anything in between SND.UNA...SND.UNA+SND.WND also can be already 901da177e4SLinus Torvalds * invalid. OK, let's make this for now: 911da177e4SLinus Torvalds */ 929e412ba7SIlpo Järvinen static inline __u32 tcp_acceptable_seq(struct sock *sk) 931da177e4SLinus Torvalds { 949e412ba7SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 959e412ba7SIlpo Järvinen 9690840defSIlpo Järvinen if (!before(tcp_wnd_end(tp), tp->snd_nxt)) 971da177e4SLinus Torvalds return tp->snd_nxt; 981da177e4SLinus Torvalds else 9990840defSIlpo Järvinen return tcp_wnd_end(tp); 1001da177e4SLinus Torvalds } 1011da177e4SLinus Torvalds 1021da177e4SLinus Torvalds /* Calculate mss to advertise in SYN segment. 1031da177e4SLinus Torvalds * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that: 1041da177e4SLinus Torvalds * 1051da177e4SLinus Torvalds * 1. It is independent of path mtu. 1061da177e4SLinus Torvalds * 2. Ideally, it is maximal possible segment size i.e. 65535-40. 1071da177e4SLinus Torvalds * 3. For IPv4 it is reasonable to calculate it from maximal MTU of 1081da177e4SLinus Torvalds * attached devices, because some buggy hosts are confused by 1091da177e4SLinus Torvalds * large MSS. 1101da177e4SLinus Torvalds * 4. We do not make 3, we advertise MSS, calculated from first 1111da177e4SLinus Torvalds * hop device mtu, but allow to raise it to ip_rt_min_advmss. 1121da177e4SLinus Torvalds * This may be overridden via information stored in routing table. 1131da177e4SLinus Torvalds * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible, 1141da177e4SLinus Torvalds * probably even Jumbo". 1151da177e4SLinus Torvalds */ 1161da177e4SLinus Torvalds static __u16 tcp_advertise_mss(struct sock *sk) 1171da177e4SLinus Torvalds { 1181da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 1191da177e4SLinus Torvalds struct dst_entry *dst = __sk_dst_get(sk); 1201da177e4SLinus Torvalds int mss = tp->advmss; 1211da177e4SLinus Torvalds 1221da177e4SLinus Torvalds if (dst && dst_metric(dst, RTAX_ADVMSS) < mss) { 1231da177e4SLinus Torvalds mss = dst_metric(dst, RTAX_ADVMSS); 1241da177e4SLinus Torvalds tp->advmss = mss; 1251da177e4SLinus Torvalds } 1261da177e4SLinus Torvalds 1271da177e4SLinus Torvalds return (__u16)mss; 1281da177e4SLinus Torvalds } 1291da177e4SLinus Torvalds 1301da177e4SLinus Torvalds /* RFC2861. Reset CWND after idle period longer RTO to "restart window". 1311da177e4SLinus Torvalds * This is the first part of cwnd validation mechanism. */ 132463c84b9SArnaldo Carvalho de Melo static void tcp_cwnd_restart(struct sock *sk, struct dst_entry *dst) 1331da177e4SLinus Torvalds { 134463c84b9SArnaldo Carvalho de Melo struct tcp_sock *tp = tcp_sk(sk); 1351da177e4SLinus Torvalds s32 delta = tcp_time_stamp - tp->lsndtime; 1361da177e4SLinus Torvalds u32 restart_cwnd = tcp_init_cwnd(tp, dst); 1371da177e4SLinus Torvalds u32 cwnd = tp->snd_cwnd; 1381da177e4SLinus Torvalds 1396687e988SArnaldo Carvalho de Melo tcp_ca_event(sk, CA_EVENT_CWND_RESTART); 1401da177e4SLinus Torvalds 1416687e988SArnaldo Carvalho de Melo tp->snd_ssthresh = tcp_current_ssthresh(sk); 1421da177e4SLinus Torvalds restart_cwnd = min(restart_cwnd, cwnd); 1431da177e4SLinus Torvalds 144463c84b9SArnaldo Carvalho de Melo while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd) 1451da177e4SLinus Torvalds cwnd >>= 1; 1461da177e4SLinus Torvalds tp->snd_cwnd = max(cwnd, restart_cwnd); 1471da177e4SLinus Torvalds tp->snd_cwnd_stamp = tcp_time_stamp; 1481da177e4SLinus Torvalds tp->snd_cwnd_used = 0; 1491da177e4SLinus Torvalds } 1501da177e4SLinus Torvalds 15167edfef7SAndi Kleen /* Congestion state accounting after a packet has been sent. */ 15240efc6faSStephen Hemminger static void tcp_event_data_sent(struct tcp_sock *tp, 1531da177e4SLinus Torvalds struct sk_buff *skb, struct sock *sk) 1541da177e4SLinus Torvalds { 155463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 156463c84b9SArnaldo Carvalho de Melo const u32 now = tcp_time_stamp; 1571da177e4SLinus Torvalds 15835089bb2SDavid S. Miller if (sysctl_tcp_slow_start_after_idle && 15935089bb2SDavid S. Miller (!tp->packets_out && (s32)(now - tp->lsndtime) > icsk->icsk_rto)) 160463c84b9SArnaldo Carvalho de Melo tcp_cwnd_restart(sk, __sk_dst_get(sk)); 1611da177e4SLinus Torvalds 1621da177e4SLinus Torvalds tp->lsndtime = now; 1631da177e4SLinus Torvalds 1641da177e4SLinus Torvalds /* If it is a reply for ato after last received 1651da177e4SLinus Torvalds * packet, enter pingpong mode. 1661da177e4SLinus Torvalds */ 167463c84b9SArnaldo Carvalho de Melo if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato) 168463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.pingpong = 1; 1691da177e4SLinus Torvalds } 1701da177e4SLinus Torvalds 17167edfef7SAndi Kleen /* Account for an ACK we sent. */ 17240efc6faSStephen Hemminger static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts) 1731da177e4SLinus Torvalds { 174463c84b9SArnaldo Carvalho de Melo tcp_dec_quickack_mode(sk, pkts); 175463c84b9SArnaldo Carvalho de Melo inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK); 1761da177e4SLinus Torvalds } 1771da177e4SLinus Torvalds 1781da177e4SLinus Torvalds /* Determine a window scaling and initial window to offer. 1791da177e4SLinus Torvalds * Based on the assumption that the given amount of space 1801da177e4SLinus Torvalds * will be offered. Store the results in the tp structure. 1811da177e4SLinus Torvalds * NOTE: for smooth operation initial space offering should 1821da177e4SLinus Torvalds * be a multiple of mss if possible. We assume here that mss >= 1. 1831da177e4SLinus Torvalds * This MUST be enforced by all callers. 1841da177e4SLinus Torvalds */ 1851da177e4SLinus Torvalds void tcp_select_initial_window(int __space, __u32 mss, 1861da177e4SLinus Torvalds __u32 *rcv_wnd, __u32 *window_clamp, 18731d12926Slaurent chavey int wscale_ok, __u8 *rcv_wscale, 18831d12926Slaurent chavey __u32 init_rcv_wnd) 1891da177e4SLinus Torvalds { 1901da177e4SLinus Torvalds unsigned int space = (__space < 0 ? 0 : __space); 1911da177e4SLinus Torvalds 1921da177e4SLinus Torvalds /* If no clamp set the clamp to the max possible scaled window */ 1931da177e4SLinus Torvalds if (*window_clamp == 0) 1941da177e4SLinus Torvalds (*window_clamp) = (65535 << 14); 1951da177e4SLinus Torvalds space = min(*window_clamp, space); 1961da177e4SLinus Torvalds 1971da177e4SLinus Torvalds /* Quantize space offering to a multiple of mss if possible. */ 1981da177e4SLinus Torvalds if (space > mss) 1991da177e4SLinus Torvalds space = (space / mss) * mss; 2001da177e4SLinus Torvalds 2011da177e4SLinus Torvalds /* NOTE: offering an initial window larger than 32767 20215d99e02SRick Jones * will break some buggy TCP stacks. If the admin tells us 20315d99e02SRick Jones * it is likely we could be speaking with such a buggy stack 20415d99e02SRick Jones * we will truncate our initial window offering to 32K-1 20515d99e02SRick Jones * unless the remote has sent us a window scaling option, 20615d99e02SRick Jones * which we interpret as a sign the remote TCP is not 20715d99e02SRick Jones * misinterpreting the window field as a signed quantity. 2081da177e4SLinus Torvalds */ 20915d99e02SRick Jones if (sysctl_tcp_workaround_signed_windows) 2101da177e4SLinus Torvalds (*rcv_wnd) = min(space, MAX_TCP_WINDOW); 21115d99e02SRick Jones else 21215d99e02SRick Jones (*rcv_wnd) = space; 21315d99e02SRick Jones 2141da177e4SLinus Torvalds (*rcv_wscale) = 0; 2151da177e4SLinus Torvalds if (wscale_ok) { 2161da177e4SLinus Torvalds /* Set window scaling on max possible window 2171da177e4SLinus Torvalds * See RFC1323 for an explanation of the limit to 14 2181da177e4SLinus Torvalds */ 2191da177e4SLinus Torvalds space = max_t(u32, sysctl_tcp_rmem[2], sysctl_rmem_max); 220316c1592SStephen Hemminger space = min_t(u32, space, *window_clamp); 2211da177e4SLinus Torvalds while (space > 65535 && (*rcv_wscale) < 14) { 2221da177e4SLinus Torvalds space >>= 1; 2231da177e4SLinus Torvalds (*rcv_wscale)++; 2241da177e4SLinus Torvalds } 2251da177e4SLinus Torvalds } 2261da177e4SLinus Torvalds 2270705c6f0SGerrit Renker /* Set initial window to value enough for senders, following RFC5681. */ 2281da177e4SLinus Torvalds if (mss > (1 << *rcv_wscale)) { 2290705c6f0SGerrit Renker int init_cwnd = rfc3390_bytes_to_packets(mss); 2300705c6f0SGerrit Renker 23131d12926Slaurent chavey /* when initializing use the value from init_rcv_wnd 23231d12926Slaurent chavey * rather than the default from above 23331d12926Slaurent chavey */ 23431d12926Slaurent chavey if (init_rcv_wnd && 23531d12926Slaurent chavey (*rcv_wnd > init_rcv_wnd * mss)) 23631d12926Slaurent chavey *rcv_wnd = init_rcv_wnd * mss; 23731d12926Slaurent chavey else if (*rcv_wnd > init_cwnd * mss) 2381da177e4SLinus Torvalds *rcv_wnd = init_cwnd * mss; 2391da177e4SLinus Torvalds } 2401da177e4SLinus Torvalds 2411da177e4SLinus Torvalds /* Set the clamp no higher than max representable value */ 2421da177e4SLinus Torvalds (*window_clamp) = min(65535U << (*rcv_wscale), *window_clamp); 2431da177e4SLinus Torvalds } 2444bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_select_initial_window); 2451da177e4SLinus Torvalds 2461da177e4SLinus Torvalds /* Chose a new window to advertise, update state in tcp_sock for the 2471da177e4SLinus Torvalds * socket, and return result with RFC1323 scaling applied. The return 2481da177e4SLinus Torvalds * value can be stuffed directly into th->window for an outgoing 2491da177e4SLinus Torvalds * frame. 2501da177e4SLinus Torvalds */ 25140efc6faSStephen Hemminger static u16 tcp_select_window(struct sock *sk) 2521da177e4SLinus Torvalds { 2531da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 2541da177e4SLinus Torvalds u32 cur_win = tcp_receive_window(tp); 2551da177e4SLinus Torvalds u32 new_win = __tcp_select_window(sk); 2561da177e4SLinus Torvalds 2571da177e4SLinus Torvalds /* Never shrink the offered window */ 2581da177e4SLinus Torvalds if (new_win < cur_win) { 2591da177e4SLinus Torvalds /* Danger Will Robinson! 2601da177e4SLinus Torvalds * Don't update rcv_wup/rcv_wnd here or else 2611da177e4SLinus Torvalds * we will not be able to advertise a zero 2621da177e4SLinus Torvalds * window in time. --DaveM 2631da177e4SLinus Torvalds * 2641da177e4SLinus Torvalds * Relax Will Robinson. 2651da177e4SLinus Torvalds */ 266607bfbf2SPatrick McHardy new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale); 2671da177e4SLinus Torvalds } 2681da177e4SLinus Torvalds tp->rcv_wnd = new_win; 2691da177e4SLinus Torvalds tp->rcv_wup = tp->rcv_nxt; 2701da177e4SLinus Torvalds 2711da177e4SLinus Torvalds /* Make sure we do not exceed the maximum possible 2721da177e4SLinus Torvalds * scaled window. 2731da177e4SLinus Torvalds */ 27415d99e02SRick Jones if (!tp->rx_opt.rcv_wscale && sysctl_tcp_workaround_signed_windows) 2751da177e4SLinus Torvalds new_win = min(new_win, MAX_TCP_WINDOW); 2761da177e4SLinus Torvalds else 2771da177e4SLinus Torvalds new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale)); 2781da177e4SLinus Torvalds 2791da177e4SLinus Torvalds /* RFC1323 scaling applied */ 2801da177e4SLinus Torvalds new_win >>= tp->rx_opt.rcv_wscale; 2811da177e4SLinus Torvalds 2821da177e4SLinus Torvalds /* If we advertise zero window, disable fast path. */ 2831da177e4SLinus Torvalds if (new_win == 0) 2841da177e4SLinus Torvalds tp->pred_flags = 0; 2851da177e4SLinus Torvalds 2861da177e4SLinus Torvalds return new_win; 2871da177e4SLinus Torvalds } 2881da177e4SLinus Torvalds 28967edfef7SAndi Kleen /* Packet ECN state for a SYN-ACK */ 290056834d9SIlpo Järvinen static inline void TCP_ECN_send_synack(struct tcp_sock *tp, struct sk_buff *skb) 291bdf1ee5dSIlpo Järvinen { 292a3433f35SChangli Gao TCP_SKB_CB(skb)->flags &= ~TCPHDR_CWR; 293bdf1ee5dSIlpo Järvinen if (!(tp->ecn_flags & TCP_ECN_OK)) 294a3433f35SChangli Gao TCP_SKB_CB(skb)->flags &= ~TCPHDR_ECE; 295bdf1ee5dSIlpo Järvinen } 296bdf1ee5dSIlpo Järvinen 29767edfef7SAndi Kleen /* Packet ECN state for a SYN. */ 298bdf1ee5dSIlpo Järvinen static inline void TCP_ECN_send_syn(struct sock *sk, struct sk_buff *skb) 299bdf1ee5dSIlpo Järvinen { 300bdf1ee5dSIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 301bdf1ee5dSIlpo Järvinen 302bdf1ee5dSIlpo Järvinen tp->ecn_flags = 0; 303255cac91SIlpo Järvinen if (sysctl_tcp_ecn == 1) { 304a3433f35SChangli Gao TCP_SKB_CB(skb)->flags |= TCPHDR_ECE | TCPHDR_CWR; 305bdf1ee5dSIlpo Järvinen tp->ecn_flags = TCP_ECN_OK; 306bdf1ee5dSIlpo Järvinen } 307bdf1ee5dSIlpo Järvinen } 308bdf1ee5dSIlpo Järvinen 309bdf1ee5dSIlpo Järvinen static __inline__ void 310bdf1ee5dSIlpo Järvinen TCP_ECN_make_synack(struct request_sock *req, struct tcphdr *th) 311bdf1ee5dSIlpo Järvinen { 312bdf1ee5dSIlpo Järvinen if (inet_rsk(req)->ecn_ok) 313bdf1ee5dSIlpo Järvinen th->ece = 1; 314bdf1ee5dSIlpo Järvinen } 315bdf1ee5dSIlpo Järvinen 31667edfef7SAndi Kleen /* Set up ECN state for a packet on a ESTABLISHED socket that is about to 31767edfef7SAndi Kleen * be sent. 31867edfef7SAndi Kleen */ 319bdf1ee5dSIlpo Järvinen static inline void TCP_ECN_send(struct sock *sk, struct sk_buff *skb, 320bdf1ee5dSIlpo Järvinen int tcp_header_len) 321bdf1ee5dSIlpo Järvinen { 322bdf1ee5dSIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 323bdf1ee5dSIlpo Järvinen 324bdf1ee5dSIlpo Järvinen if (tp->ecn_flags & TCP_ECN_OK) { 325bdf1ee5dSIlpo Järvinen /* Not-retransmitted data segment: set ECT and inject CWR. */ 326bdf1ee5dSIlpo Järvinen if (skb->len != tcp_header_len && 327bdf1ee5dSIlpo Järvinen !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) { 328bdf1ee5dSIlpo Järvinen INET_ECN_xmit(sk); 329bdf1ee5dSIlpo Järvinen if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) { 330bdf1ee5dSIlpo Järvinen tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR; 331bdf1ee5dSIlpo Järvinen tcp_hdr(skb)->cwr = 1; 332bdf1ee5dSIlpo Järvinen skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN; 333bdf1ee5dSIlpo Järvinen } 334bdf1ee5dSIlpo Järvinen } else { 335bdf1ee5dSIlpo Järvinen /* ACK or retransmitted segment: clear ECT|CE */ 336bdf1ee5dSIlpo Järvinen INET_ECN_dontxmit(sk); 337bdf1ee5dSIlpo Järvinen } 338bdf1ee5dSIlpo Järvinen if (tp->ecn_flags & TCP_ECN_DEMAND_CWR) 339bdf1ee5dSIlpo Järvinen tcp_hdr(skb)->ece = 1; 340bdf1ee5dSIlpo Järvinen } 341bdf1ee5dSIlpo Järvinen } 342bdf1ee5dSIlpo Järvinen 343e870a8efSIlpo Järvinen /* Constructs common control bits of non-data skb. If SYN/FIN is present, 344e870a8efSIlpo Järvinen * auto increment end seqno. 345e870a8efSIlpo Järvinen */ 346e870a8efSIlpo Järvinen static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags) 347e870a8efSIlpo Järvinen { 3482e8e18efSDavid S. Miller skb->ip_summed = CHECKSUM_PARTIAL; 349e870a8efSIlpo Järvinen skb->csum = 0; 350e870a8efSIlpo Järvinen 351e870a8efSIlpo Järvinen TCP_SKB_CB(skb)->flags = flags; 352e870a8efSIlpo Järvinen TCP_SKB_CB(skb)->sacked = 0; 353e870a8efSIlpo Järvinen 354e870a8efSIlpo Järvinen skb_shinfo(skb)->gso_segs = 1; 355e870a8efSIlpo Järvinen skb_shinfo(skb)->gso_size = 0; 356e870a8efSIlpo Järvinen skb_shinfo(skb)->gso_type = 0; 357e870a8efSIlpo Järvinen 358e870a8efSIlpo Järvinen TCP_SKB_CB(skb)->seq = seq; 359a3433f35SChangli Gao if (flags & (TCPHDR_SYN | TCPHDR_FIN)) 360e870a8efSIlpo Järvinen seq++; 361e870a8efSIlpo Järvinen TCP_SKB_CB(skb)->end_seq = seq; 362e870a8efSIlpo Järvinen } 363e870a8efSIlpo Järvinen 36433f5f57eSIlpo Järvinen static inline int tcp_urg_mode(const struct tcp_sock *tp) 36533f5f57eSIlpo Järvinen { 36633f5f57eSIlpo Järvinen return tp->snd_una != tp->snd_up; 36733f5f57eSIlpo Järvinen } 36833f5f57eSIlpo Järvinen 36933ad798cSAdam Langley #define OPTION_SACK_ADVERTISE (1 << 0) 37033ad798cSAdam Langley #define OPTION_TS (1 << 1) 37133ad798cSAdam Langley #define OPTION_MD5 (1 << 2) 37289e95a61SOri Finkelman #define OPTION_WSCALE (1 << 3) 373bd0388aeSWilliam Allen Simpson #define OPTION_COOKIE_EXTENSION (1 << 4) 37433ad798cSAdam Langley 37533ad798cSAdam Langley struct tcp_out_options { 37633ad798cSAdam Langley u8 options; /* bit field of OPTION_* */ 37733ad798cSAdam Langley u8 ws; /* window scale, 0 to disable */ 37833ad798cSAdam Langley u8 num_sack_blocks; /* number of SACK blocks to include */ 379bd0388aeSWilliam Allen Simpson u8 hash_size; /* bytes in hash_location */ 38033ad798cSAdam Langley u16 mss; /* 0 to disable */ 38133ad798cSAdam Langley __u32 tsval, tsecr; /* need to include OPTION_TS */ 382bd0388aeSWilliam Allen Simpson __u8 *hash_location; /* temporary pointer, overloaded */ 38333ad798cSAdam Langley }; 38433ad798cSAdam Langley 385bd0388aeSWilliam Allen Simpson /* The sysctl int routines are generic, so check consistency here. 386bd0388aeSWilliam Allen Simpson */ 387bd0388aeSWilliam Allen Simpson static u8 tcp_cookie_size_check(u8 desired) 388bd0388aeSWilliam Allen Simpson { 389bd0388aeSWilliam Allen Simpson if (desired > 0) { 390bd0388aeSWilliam Allen Simpson /* previously specified */ 391bd0388aeSWilliam Allen Simpson return desired; 392bd0388aeSWilliam Allen Simpson } 393bd0388aeSWilliam Allen Simpson if (sysctl_tcp_cookie_size <= 0) { 394bd0388aeSWilliam Allen Simpson /* no default specified */ 395bd0388aeSWilliam Allen Simpson return 0; 396bd0388aeSWilliam Allen Simpson } 397bd0388aeSWilliam Allen Simpson if (sysctl_tcp_cookie_size <= TCP_COOKIE_MIN) { 398bd0388aeSWilliam Allen Simpson /* value too small, specify minimum */ 399bd0388aeSWilliam Allen Simpson return TCP_COOKIE_MIN; 400bd0388aeSWilliam Allen Simpson } 401bd0388aeSWilliam Allen Simpson if (sysctl_tcp_cookie_size >= TCP_COOKIE_MAX) { 402bd0388aeSWilliam Allen Simpson /* value too large, specify maximum */ 403bd0388aeSWilliam Allen Simpson return TCP_COOKIE_MAX; 404bd0388aeSWilliam Allen Simpson } 405bd0388aeSWilliam Allen Simpson if (0x1 & sysctl_tcp_cookie_size) { 406bd0388aeSWilliam Allen Simpson /* 8-bit multiple, illegal, fix it */ 407bd0388aeSWilliam Allen Simpson return (u8)(sysctl_tcp_cookie_size + 0x1); 408bd0388aeSWilliam Allen Simpson } 409bd0388aeSWilliam Allen Simpson return (u8)sysctl_tcp_cookie_size; 410bd0388aeSWilliam Allen Simpson } 411bd0388aeSWilliam Allen Simpson 41267edfef7SAndi Kleen /* Write previously computed TCP options to the packet. 41367edfef7SAndi Kleen * 41467edfef7SAndi Kleen * Beware: Something in the Internet is very sensitive to the ordering of 415fd6149d3SIlpo Järvinen * TCP options, we learned this through the hard way, so be careful here. 416fd6149d3SIlpo Järvinen * Luckily we can at least blame others for their non-compliance but from 417fd6149d3SIlpo Järvinen * inter-operatibility perspective it seems that we're somewhat stuck with 418fd6149d3SIlpo Järvinen * the ordering which we have been using if we want to keep working with 419fd6149d3SIlpo Järvinen * those broken things (not that it currently hurts anybody as there isn't 420fd6149d3SIlpo Järvinen * particular reason why the ordering would need to be changed). 421fd6149d3SIlpo Järvinen * 422fd6149d3SIlpo Järvinen * At least SACK_PERM as the first option is known to lead to a disaster 423fd6149d3SIlpo Järvinen * (but it may well be that other scenarios fail similarly). 424fd6149d3SIlpo Järvinen */ 42533ad798cSAdam Langley static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp, 426bd0388aeSWilliam Allen Simpson struct tcp_out_options *opts) 427bd0388aeSWilliam Allen Simpson { 428bd0388aeSWilliam Allen Simpson u8 options = opts->options; /* mungable copy */ 429bd0388aeSWilliam Allen Simpson 430bd0388aeSWilliam Allen Simpson /* Having both authentication and cookies for security is redundant, 431bd0388aeSWilliam Allen Simpson * and there's certainly not enough room. Instead, the cookie-less 432bd0388aeSWilliam Allen Simpson * extension variant is proposed. 433bd0388aeSWilliam Allen Simpson * 434bd0388aeSWilliam Allen Simpson * Consider the pessimal case with authentication. The options 435bd0388aeSWilliam Allen Simpson * could look like: 436bd0388aeSWilliam Allen Simpson * COOKIE|MD5(20) + MSS(4) + SACK|TS(12) + WSCALE(4) == 40 437bd0388aeSWilliam Allen Simpson */ 438bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_MD5 & options)) { 439bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_COOKIE_EXTENSION & options)) { 440bd0388aeSWilliam Allen Simpson *ptr++ = htonl((TCPOPT_COOKIE << 24) | 441bd0388aeSWilliam Allen Simpson (TCPOLEN_COOKIE_BASE << 16) | 442bd0388aeSWilliam Allen Simpson (TCPOPT_MD5SIG << 8) | 443bd0388aeSWilliam Allen Simpson TCPOLEN_MD5SIG); 444bd0388aeSWilliam Allen Simpson } else { 44533ad798cSAdam Langley *ptr++ = htonl((TCPOPT_NOP << 24) | 44633ad798cSAdam Langley (TCPOPT_NOP << 16) | 44733ad798cSAdam Langley (TCPOPT_MD5SIG << 8) | 44833ad798cSAdam Langley TCPOLEN_MD5SIG); 449bd0388aeSWilliam Allen Simpson } 450bd0388aeSWilliam Allen Simpson options &= ~OPTION_COOKIE_EXTENSION; 451bd0388aeSWilliam Allen Simpson /* overload cookie hash location */ 452bd0388aeSWilliam Allen Simpson opts->hash_location = (__u8 *)ptr; 45333ad798cSAdam Langley ptr += 4; 45433ad798cSAdam Langley } 45533ad798cSAdam Langley 456fd6149d3SIlpo Järvinen if (unlikely(opts->mss)) { 457fd6149d3SIlpo Järvinen *ptr++ = htonl((TCPOPT_MSS << 24) | 458fd6149d3SIlpo Järvinen (TCPOLEN_MSS << 16) | 459fd6149d3SIlpo Järvinen opts->mss); 460fd6149d3SIlpo Järvinen } 461fd6149d3SIlpo Järvinen 462bd0388aeSWilliam Allen Simpson if (likely(OPTION_TS & options)) { 463bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_SACK_ADVERTISE & options)) { 46433ad798cSAdam Langley *ptr++ = htonl((TCPOPT_SACK_PERM << 24) | 46533ad798cSAdam Langley (TCPOLEN_SACK_PERM << 16) | 46633ad798cSAdam Langley (TCPOPT_TIMESTAMP << 8) | 46733ad798cSAdam Langley TCPOLEN_TIMESTAMP); 468bd0388aeSWilliam Allen Simpson options &= ~OPTION_SACK_ADVERTISE; 46933ad798cSAdam Langley } else { 470496c98dfSYOSHIFUJI Hideaki *ptr++ = htonl((TCPOPT_NOP << 24) | 47140efc6faSStephen Hemminger (TCPOPT_NOP << 16) | 47240efc6faSStephen Hemminger (TCPOPT_TIMESTAMP << 8) | 47340efc6faSStephen Hemminger TCPOLEN_TIMESTAMP); 47440efc6faSStephen Hemminger } 47533ad798cSAdam Langley *ptr++ = htonl(opts->tsval); 47633ad798cSAdam Langley *ptr++ = htonl(opts->tsecr); 47733ad798cSAdam Langley } 47833ad798cSAdam Langley 479bd0388aeSWilliam Allen Simpson /* Specification requires after timestamp, so do it now. 480bd0388aeSWilliam Allen Simpson * 481bd0388aeSWilliam Allen Simpson * Consider the pessimal case without authentication. The options 482bd0388aeSWilliam Allen Simpson * could look like: 483bd0388aeSWilliam Allen Simpson * MSS(4) + SACK|TS(12) + COOKIE(20) + WSCALE(4) == 40 484bd0388aeSWilliam Allen Simpson */ 485bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_COOKIE_EXTENSION & options)) { 486bd0388aeSWilliam Allen Simpson __u8 *cookie_copy = opts->hash_location; 487bd0388aeSWilliam Allen Simpson u8 cookie_size = opts->hash_size; 488bd0388aeSWilliam Allen Simpson 489bd0388aeSWilliam Allen Simpson /* 8-bit multiple handled in tcp_cookie_size_check() above, 490bd0388aeSWilliam Allen Simpson * and elsewhere. 491bd0388aeSWilliam Allen Simpson */ 492bd0388aeSWilliam Allen Simpson if (0x2 & cookie_size) { 493bd0388aeSWilliam Allen Simpson __u8 *p = (__u8 *)ptr; 494bd0388aeSWilliam Allen Simpson 495bd0388aeSWilliam Allen Simpson /* 16-bit multiple */ 496bd0388aeSWilliam Allen Simpson *p++ = TCPOPT_COOKIE; 497bd0388aeSWilliam Allen Simpson *p++ = TCPOLEN_COOKIE_BASE + cookie_size; 498bd0388aeSWilliam Allen Simpson *p++ = *cookie_copy++; 499bd0388aeSWilliam Allen Simpson *p++ = *cookie_copy++; 500bd0388aeSWilliam Allen Simpson ptr++; 501bd0388aeSWilliam Allen Simpson cookie_size -= 2; 502bd0388aeSWilliam Allen Simpson } else { 503bd0388aeSWilliam Allen Simpson /* 32-bit multiple */ 504bd0388aeSWilliam Allen Simpson *ptr++ = htonl(((TCPOPT_NOP << 24) | 505bd0388aeSWilliam Allen Simpson (TCPOPT_NOP << 16) | 506bd0388aeSWilliam Allen Simpson (TCPOPT_COOKIE << 8) | 507bd0388aeSWilliam Allen Simpson TCPOLEN_COOKIE_BASE) + 508bd0388aeSWilliam Allen Simpson cookie_size); 509bd0388aeSWilliam Allen Simpson } 510bd0388aeSWilliam Allen Simpson 511bd0388aeSWilliam Allen Simpson if (cookie_size > 0) { 512bd0388aeSWilliam Allen Simpson memcpy(ptr, cookie_copy, cookie_size); 513bd0388aeSWilliam Allen Simpson ptr += (cookie_size / 4); 514bd0388aeSWilliam Allen Simpson } 515bd0388aeSWilliam Allen Simpson } 516bd0388aeSWilliam Allen Simpson 517bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_SACK_ADVERTISE & options)) { 51833ad798cSAdam Langley *ptr++ = htonl((TCPOPT_NOP << 24) | 51933ad798cSAdam Langley (TCPOPT_NOP << 16) | 52033ad798cSAdam Langley (TCPOPT_SACK_PERM << 8) | 52133ad798cSAdam Langley TCPOLEN_SACK_PERM); 52233ad798cSAdam Langley } 52333ad798cSAdam Langley 524bd0388aeSWilliam Allen Simpson if (unlikely(OPTION_WSCALE & options)) { 52533ad798cSAdam Langley *ptr++ = htonl((TCPOPT_NOP << 24) | 52633ad798cSAdam Langley (TCPOPT_WINDOW << 16) | 52733ad798cSAdam Langley (TCPOLEN_WINDOW << 8) | 52833ad798cSAdam Langley opts->ws); 52933ad798cSAdam Langley } 53033ad798cSAdam Langley 53133ad798cSAdam Langley if (unlikely(opts->num_sack_blocks)) { 53233ad798cSAdam Langley struct tcp_sack_block *sp = tp->rx_opt.dsack ? 53333ad798cSAdam Langley tp->duplicate_sack : tp->selective_acks; 53440efc6faSStephen Hemminger int this_sack; 53540efc6faSStephen Hemminger 53640efc6faSStephen Hemminger *ptr++ = htonl((TCPOPT_NOP << 24) | 53740efc6faSStephen Hemminger (TCPOPT_NOP << 16) | 53840efc6faSStephen Hemminger (TCPOPT_SACK << 8) | 53933ad798cSAdam Langley (TCPOLEN_SACK_BASE + (opts->num_sack_blocks * 54040efc6faSStephen Hemminger TCPOLEN_SACK_PERBLOCK))); 5412de979bdSStephen Hemminger 54233ad798cSAdam Langley for (this_sack = 0; this_sack < opts->num_sack_blocks; 54333ad798cSAdam Langley ++this_sack) { 54440efc6faSStephen Hemminger *ptr++ = htonl(sp[this_sack].start_seq); 54540efc6faSStephen Hemminger *ptr++ = htonl(sp[this_sack].end_seq); 54640efc6faSStephen Hemminger } 5472de979bdSStephen Hemminger 54840efc6faSStephen Hemminger tp->rx_opt.dsack = 0; 54940efc6faSStephen Hemminger } 55040efc6faSStephen Hemminger } 55140efc6faSStephen Hemminger 55267edfef7SAndi Kleen /* Compute TCP options for SYN packets. This is not the final 55367edfef7SAndi Kleen * network wire format yet. 55467edfef7SAndi Kleen */ 55533ad798cSAdam Langley static unsigned tcp_syn_options(struct sock *sk, struct sk_buff *skb, 55633ad798cSAdam Langley struct tcp_out_options *opts, 55733ad798cSAdam Langley struct tcp_md5sig_key **md5) { 55833ad798cSAdam Langley struct tcp_sock *tp = tcp_sk(sk); 559bd0388aeSWilliam Allen Simpson struct tcp_cookie_values *cvp = tp->cookie_values; 560bd0388aeSWilliam Allen Simpson unsigned remaining = MAX_TCP_OPTION_SPACE; 561bd0388aeSWilliam Allen Simpson u8 cookie_size = (!tp->rx_opt.cookie_out_never && cvp != NULL) ? 562bd0388aeSWilliam Allen Simpson tcp_cookie_size_check(cvp->cookie_desired) : 563bd0388aeSWilliam Allen Simpson 0; 56433ad798cSAdam Langley 565cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 56633ad798cSAdam Langley *md5 = tp->af_specific->md5_lookup(sk, sk); 56733ad798cSAdam Langley if (*md5) { 56833ad798cSAdam Langley opts->options |= OPTION_MD5; 569bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_MD5SIG_ALIGNED; 570cfb6eeb4SYOSHIFUJI Hideaki } 57133ad798cSAdam Langley #else 57233ad798cSAdam Langley *md5 = NULL; 573cfb6eeb4SYOSHIFUJI Hideaki #endif 57433ad798cSAdam Langley 57533ad798cSAdam Langley /* We always get an MSS option. The option bytes which will be seen in 57633ad798cSAdam Langley * normal data packets should timestamps be used, must be in the MSS 57733ad798cSAdam Langley * advertised. But we subtract them from tp->mss_cache so that 57833ad798cSAdam Langley * calculations in tcp_sendmsg are simpler etc. So account for this 57933ad798cSAdam Langley * fact here if necessary. If we don't do this correctly, as a 58033ad798cSAdam Langley * receiver we won't recognize data packets as being full sized when we 58133ad798cSAdam Langley * should, and thus we won't abide by the delayed ACK rules correctly. 58233ad798cSAdam Langley * SACKs don't matter, we never delay an ACK when we have any of those 58333ad798cSAdam Langley * going out. */ 58433ad798cSAdam Langley opts->mss = tcp_advertise_mss(sk); 585bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_MSS_ALIGNED; 58633ad798cSAdam Langley 587bb5b7c11SDavid S. Miller if (likely(sysctl_tcp_timestamps && *md5 == NULL)) { 58833ad798cSAdam Langley opts->options |= OPTION_TS; 58933ad798cSAdam Langley opts->tsval = TCP_SKB_CB(skb)->when; 59033ad798cSAdam Langley opts->tsecr = tp->rx_opt.ts_recent; 591bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_TSTAMP_ALIGNED; 59233ad798cSAdam Langley } 593bb5b7c11SDavid S. Miller if (likely(sysctl_tcp_window_scaling)) { 59433ad798cSAdam Langley opts->ws = tp->rx_opt.rcv_wscale; 59589e95a61SOri Finkelman opts->options |= OPTION_WSCALE; 596bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_WSCALE_ALIGNED; 59733ad798cSAdam Langley } 598bb5b7c11SDavid S. Miller if (likely(sysctl_tcp_sack)) { 59933ad798cSAdam Langley opts->options |= OPTION_SACK_ADVERTISE; 600b32d1310SDavid S. Miller if (unlikely(!(OPTION_TS & opts->options))) 601bd0388aeSWilliam Allen Simpson remaining -= TCPOLEN_SACKPERM_ALIGNED; 60233ad798cSAdam Langley } 60333ad798cSAdam Langley 604bd0388aeSWilliam Allen Simpson /* Note that timestamps are required by the specification. 605bd0388aeSWilliam Allen Simpson * 606bd0388aeSWilliam Allen Simpson * Odd numbers of bytes are prohibited by the specification, ensuring 607bd0388aeSWilliam Allen Simpson * that the cookie is 16-bit aligned, and the resulting cookie pair is 608bd0388aeSWilliam Allen Simpson * 32-bit aligned. 609bd0388aeSWilliam Allen Simpson */ 610bd0388aeSWilliam Allen Simpson if (*md5 == NULL && 611bd0388aeSWilliam Allen Simpson (OPTION_TS & opts->options) && 612bd0388aeSWilliam Allen Simpson cookie_size > 0) { 613bd0388aeSWilliam Allen Simpson int need = TCPOLEN_COOKIE_BASE + cookie_size; 614bd0388aeSWilliam Allen Simpson 615bd0388aeSWilliam Allen Simpson if (0x2 & need) { 616bd0388aeSWilliam Allen Simpson /* 32-bit multiple */ 617bd0388aeSWilliam Allen Simpson need += 2; /* NOPs */ 618bd0388aeSWilliam Allen Simpson 619bd0388aeSWilliam Allen Simpson if (need > remaining) { 620bd0388aeSWilliam Allen Simpson /* try shrinking cookie to fit */ 621bd0388aeSWilliam Allen Simpson cookie_size -= 2; 622bd0388aeSWilliam Allen Simpson need -= 4; 623bd0388aeSWilliam Allen Simpson } 624bd0388aeSWilliam Allen Simpson } 625bd0388aeSWilliam Allen Simpson while (need > remaining && TCP_COOKIE_MIN <= cookie_size) { 626bd0388aeSWilliam Allen Simpson cookie_size -= 4; 627bd0388aeSWilliam Allen Simpson need -= 4; 628bd0388aeSWilliam Allen Simpson } 629bd0388aeSWilliam Allen Simpson if (TCP_COOKIE_MIN <= cookie_size) { 630bd0388aeSWilliam Allen Simpson opts->options |= OPTION_COOKIE_EXTENSION; 631bd0388aeSWilliam Allen Simpson opts->hash_location = (__u8 *)&cvp->cookie_pair[0]; 632bd0388aeSWilliam Allen Simpson opts->hash_size = cookie_size; 633bd0388aeSWilliam Allen Simpson 634bd0388aeSWilliam Allen Simpson /* Remember for future incarnations. */ 635bd0388aeSWilliam Allen Simpson cvp->cookie_desired = cookie_size; 636bd0388aeSWilliam Allen Simpson 637bd0388aeSWilliam Allen Simpson if (cvp->cookie_desired != cvp->cookie_pair_size) { 638bd0388aeSWilliam Allen Simpson /* Currently use random bytes as a nonce, 639bd0388aeSWilliam Allen Simpson * assuming these are completely unpredictable 640bd0388aeSWilliam Allen Simpson * by hostile users of the same system. 641bd0388aeSWilliam Allen Simpson */ 642bd0388aeSWilliam Allen Simpson get_random_bytes(&cvp->cookie_pair[0], 643bd0388aeSWilliam Allen Simpson cookie_size); 644bd0388aeSWilliam Allen Simpson cvp->cookie_pair_size = cookie_size; 645bd0388aeSWilliam Allen Simpson } 646bd0388aeSWilliam Allen Simpson 647bd0388aeSWilliam Allen Simpson remaining -= need; 648bd0388aeSWilliam Allen Simpson } 649bd0388aeSWilliam Allen Simpson } 650bd0388aeSWilliam Allen Simpson return MAX_TCP_OPTION_SPACE - remaining; 65133ad798cSAdam Langley } 65233ad798cSAdam Langley 65367edfef7SAndi Kleen /* Set up TCP options for SYN-ACKs. */ 65433ad798cSAdam Langley static unsigned tcp_synack_options(struct sock *sk, 65533ad798cSAdam Langley struct request_sock *req, 65633ad798cSAdam Langley unsigned mss, struct sk_buff *skb, 65733ad798cSAdam Langley struct tcp_out_options *opts, 6584957faadSWilliam Allen Simpson struct tcp_md5sig_key **md5, 6594957faadSWilliam Allen Simpson struct tcp_extend_values *xvp) 6604957faadSWilliam Allen Simpson { 66133ad798cSAdam Langley struct inet_request_sock *ireq = inet_rsk(req); 6624957faadSWilliam Allen Simpson unsigned remaining = MAX_TCP_OPTION_SPACE; 6634957faadSWilliam Allen Simpson u8 cookie_plus = (xvp != NULL && !xvp->cookie_out_never) ? 6644957faadSWilliam Allen Simpson xvp->cookie_plus : 6654957faadSWilliam Allen Simpson 0; 66633ad798cSAdam Langley 66733ad798cSAdam Langley #ifdef CONFIG_TCP_MD5SIG 66833ad798cSAdam Langley *md5 = tcp_rsk(req)->af_specific->md5_lookup(sk, req); 66933ad798cSAdam Langley if (*md5) { 67033ad798cSAdam Langley opts->options |= OPTION_MD5; 6714957faadSWilliam Allen Simpson remaining -= TCPOLEN_MD5SIG_ALIGNED; 6724957faadSWilliam Allen Simpson 6734957faadSWilliam Allen Simpson /* We can't fit any SACK blocks in a packet with MD5 + TS 6744957faadSWilliam Allen Simpson * options. There was discussion about disabling SACK 6754957faadSWilliam Allen Simpson * rather than TS in order to fit in better with old, 6764957faadSWilliam Allen Simpson * buggy kernels, but that was deemed to be unnecessary. 6774957faadSWilliam Allen Simpson */ 678de213e5eSEric Dumazet ireq->tstamp_ok &= !ireq->sack_ok; 67933ad798cSAdam Langley } 68033ad798cSAdam Langley #else 68133ad798cSAdam Langley *md5 = NULL; 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; 69533ad798cSAdam Langley opts->tsval = TCP_SKB_CB(skb)->when; 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 } 70433ad798cSAdam Langley 7054957faadSWilliam Allen Simpson /* Similar rationale to tcp_syn_options() applies here, too. 7064957faadSWilliam Allen Simpson * If the <SYN> options fit, the same options should fit now! 7074957faadSWilliam Allen Simpson */ 7084957faadSWilliam Allen Simpson if (*md5 == NULL && 709de213e5eSEric Dumazet ireq->tstamp_ok && 7104957faadSWilliam Allen Simpson cookie_plus > TCPOLEN_COOKIE_BASE) { 7114957faadSWilliam Allen Simpson int need = cookie_plus; /* has TCPOLEN_COOKIE_BASE */ 7124957faadSWilliam Allen Simpson 7134957faadSWilliam Allen Simpson if (0x2 & need) { 7144957faadSWilliam Allen Simpson /* 32-bit multiple */ 7154957faadSWilliam Allen Simpson need += 2; /* NOPs */ 7164957faadSWilliam Allen Simpson } 7174957faadSWilliam Allen Simpson if (need <= remaining) { 7184957faadSWilliam Allen Simpson opts->options |= OPTION_COOKIE_EXTENSION; 7194957faadSWilliam Allen Simpson opts->hash_size = cookie_plus - TCPOLEN_COOKIE_BASE; 7204957faadSWilliam Allen Simpson remaining -= need; 7214957faadSWilliam Allen Simpson } else { 7224957faadSWilliam Allen Simpson /* There's no error return, so flag it. */ 7234957faadSWilliam Allen Simpson xvp->cookie_out_never = 1; /* true */ 7244957faadSWilliam Allen Simpson opts->hash_size = 0; 7254957faadSWilliam Allen Simpson } 7264957faadSWilliam Allen Simpson } 7274957faadSWilliam Allen Simpson return MAX_TCP_OPTION_SPACE - remaining; 72833ad798cSAdam Langley } 72933ad798cSAdam Langley 73067edfef7SAndi Kleen /* Compute TCP options for ESTABLISHED sockets. This is not the 73167edfef7SAndi Kleen * final wire format yet. 73267edfef7SAndi Kleen */ 73333ad798cSAdam Langley static unsigned tcp_established_options(struct sock *sk, struct sk_buff *skb, 73433ad798cSAdam Langley struct tcp_out_options *opts, 73533ad798cSAdam Langley struct tcp_md5sig_key **md5) { 73633ad798cSAdam Langley struct tcp_skb_cb *tcb = skb ? TCP_SKB_CB(skb) : NULL; 73733ad798cSAdam Langley struct tcp_sock *tp = tcp_sk(sk); 73833ad798cSAdam Langley unsigned size = 0; 739cabeccbdSIlpo Järvinen unsigned int eff_sacks; 74033ad798cSAdam Langley 74133ad798cSAdam Langley #ifdef CONFIG_TCP_MD5SIG 74233ad798cSAdam Langley *md5 = tp->af_specific->md5_lookup(sk, sk); 74333ad798cSAdam Langley if (unlikely(*md5)) { 74433ad798cSAdam Langley opts->options |= OPTION_MD5; 74533ad798cSAdam Langley size += TCPOLEN_MD5SIG_ALIGNED; 74633ad798cSAdam Langley } 74733ad798cSAdam Langley #else 74833ad798cSAdam Langley *md5 = NULL; 74933ad798cSAdam Langley #endif 75033ad798cSAdam Langley 75133ad798cSAdam Langley if (likely(tp->rx_opt.tstamp_ok)) { 75233ad798cSAdam Langley opts->options |= OPTION_TS; 75333ad798cSAdam Langley opts->tsval = tcb ? tcb->when : 0; 75433ad798cSAdam Langley opts->tsecr = tp->rx_opt.ts_recent; 75533ad798cSAdam Langley size += TCPOLEN_TSTAMP_ALIGNED; 75633ad798cSAdam Langley } 75733ad798cSAdam Langley 758cabeccbdSIlpo Järvinen eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack; 759cabeccbdSIlpo Järvinen if (unlikely(eff_sacks)) { 76033ad798cSAdam Langley const unsigned remaining = MAX_TCP_OPTION_SPACE - size; 76133ad798cSAdam Langley opts->num_sack_blocks = 762cabeccbdSIlpo Järvinen min_t(unsigned, eff_sacks, 76333ad798cSAdam Langley (remaining - TCPOLEN_SACK_BASE_ALIGNED) / 76433ad798cSAdam Langley TCPOLEN_SACK_PERBLOCK); 76533ad798cSAdam Langley size += TCPOLEN_SACK_BASE_ALIGNED + 76633ad798cSAdam Langley opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK; 76733ad798cSAdam Langley } 76833ad798cSAdam Langley 76933ad798cSAdam Langley return size; 77040efc6faSStephen Hemminger } 7711da177e4SLinus Torvalds 7721da177e4SLinus Torvalds /* This routine actually transmits TCP packets queued in by 7731da177e4SLinus Torvalds * tcp_do_sendmsg(). This is used by both the initial 7741da177e4SLinus Torvalds * transmission and possible later retransmissions. 7751da177e4SLinus Torvalds * All SKB's seen here are completely headerless. It is our 7761da177e4SLinus Torvalds * job to build the TCP header, and pass the packet down to 7771da177e4SLinus Torvalds * IP so it can do the same plus pass the packet off to the 7781da177e4SLinus Torvalds * device. 7791da177e4SLinus Torvalds * 7801da177e4SLinus Torvalds * We are working here with either a clone of the original 7811da177e4SLinus Torvalds * SKB, or a fresh unique copy made by the retransmit engine. 7821da177e4SLinus Torvalds */ 783056834d9SIlpo Järvinen static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it, 784056834d9SIlpo Järvinen gfp_t gfp_mask) 7851da177e4SLinus Torvalds { 7866687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 787dfb4b9dcSDavid S. Miller struct inet_sock *inet; 788dfb4b9dcSDavid S. Miller struct tcp_sock *tp; 789dfb4b9dcSDavid S. Miller struct tcp_skb_cb *tcb; 79033ad798cSAdam Langley struct tcp_out_options opts; 79133ad798cSAdam Langley unsigned tcp_options_size, tcp_header_size; 792cfb6eeb4SYOSHIFUJI Hideaki struct tcp_md5sig_key *md5; 7931da177e4SLinus Torvalds struct tcphdr *th; 7941da177e4SLinus Torvalds int err; 7951da177e4SLinus Torvalds 796dfb4b9dcSDavid S. Miller BUG_ON(!skb || !tcp_skb_pcount(skb)); 797dfb4b9dcSDavid S. Miller 798dfb4b9dcSDavid S. Miller /* If congestion control is doing timestamping, we must 799dfb4b9dcSDavid S. Miller * take such a timestamp before we potentially clone/copy. 800dfb4b9dcSDavid S. Miller */ 801164891aaSStephen Hemminger if (icsk->icsk_ca_ops->flags & TCP_CONG_RTT_STAMP) 802dfb4b9dcSDavid S. Miller __net_timestamp(skb); 803dfb4b9dcSDavid S. Miller 804dfb4b9dcSDavid S. Miller if (likely(clone_it)) { 805dfb4b9dcSDavid S. Miller if (unlikely(skb_cloned(skb))) 806dfb4b9dcSDavid S. Miller skb = pskb_copy(skb, gfp_mask); 807dfb4b9dcSDavid S. Miller else 808dfb4b9dcSDavid S. Miller skb = skb_clone(skb, gfp_mask); 809dfb4b9dcSDavid S. Miller if (unlikely(!skb)) 810dfb4b9dcSDavid S. Miller return -ENOBUFS; 811dfb4b9dcSDavid S. Miller } 812dfb4b9dcSDavid S. Miller 813dfb4b9dcSDavid S. Miller inet = inet_sk(sk); 814dfb4b9dcSDavid S. Miller tp = tcp_sk(sk); 815dfb4b9dcSDavid S. Miller tcb = TCP_SKB_CB(skb); 81633ad798cSAdam Langley memset(&opts, 0, sizeof(opts)); 8171da177e4SLinus Torvalds 818a3433f35SChangli Gao if (unlikely(tcb->flags & TCPHDR_SYN)) 81933ad798cSAdam Langley tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5); 82033ad798cSAdam Langley else 82133ad798cSAdam Langley tcp_options_size = tcp_established_options(sk, skb, &opts, 82233ad798cSAdam Langley &md5); 82333ad798cSAdam Langley tcp_header_size = tcp_options_size + sizeof(struct tcphdr); 8241da177e4SLinus Torvalds 825317a76f9SStephen Hemminger if (tcp_packets_in_flight(tp) == 0) 8266687e988SArnaldo Carvalho de Melo tcp_ca_event(sk, CA_EVENT_TX_START); 8271da177e4SLinus Torvalds 828aa8223c7SArnaldo Carvalho de Melo skb_push(skb, tcp_header_size); 829aa8223c7SArnaldo Carvalho de Melo skb_reset_transport_header(skb); 830e89862f4SDavid S. Miller skb_set_owner_w(skb, sk); 8311da177e4SLinus Torvalds 8321da177e4SLinus Torvalds /* Build TCP header and checksum it. */ 833aa8223c7SArnaldo Carvalho de Melo th = tcp_hdr(skb); 834c720c7e8SEric Dumazet th->source = inet->inet_sport; 835c720c7e8SEric Dumazet th->dest = inet->inet_dport; 8361da177e4SLinus Torvalds th->seq = htonl(tcb->seq); 8371da177e4SLinus Torvalds th->ack_seq = htonl(tp->rcv_nxt); 838df7a3b07SAl Viro *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) | 839dfb4b9dcSDavid S. Miller tcb->flags); 840dfb4b9dcSDavid S. Miller 841a3433f35SChangli Gao if (unlikely(tcb->flags & TCPHDR_SYN)) { 8421da177e4SLinus Torvalds /* RFC1323: The window in SYN & SYN/ACK segments 8431da177e4SLinus Torvalds * is never scaled. 8441da177e4SLinus Torvalds */ 845600ff0c2SIlpo Järvinen th->window = htons(min(tp->rcv_wnd, 65535U)); 8461da177e4SLinus Torvalds } else { 8471da177e4SLinus Torvalds th->window = htons(tcp_select_window(sk)); 8481da177e4SLinus Torvalds } 8491da177e4SLinus Torvalds th->check = 0; 8501da177e4SLinus Torvalds th->urg_ptr = 0; 8511da177e4SLinus Torvalds 85233f5f57eSIlpo Järvinen /* The urg_mode check is necessary during a below snd_una win probe */ 8537691367dSHerbert Xu if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) { 8547691367dSHerbert Xu if (before(tp->snd_up, tcb->seq + 0x10000)) { 8551da177e4SLinus Torvalds th->urg_ptr = htons(tp->snd_up - tcb->seq); 8561da177e4SLinus Torvalds th->urg = 1; 8577691367dSHerbert Xu } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) { 8580eae88f3SEric Dumazet th->urg_ptr = htons(0xFFFF); 8597691367dSHerbert Xu th->urg = 1; 8607691367dSHerbert Xu } 8611da177e4SLinus Torvalds } 8621da177e4SLinus Torvalds 863bd0388aeSWilliam Allen Simpson tcp_options_write((__be32 *)(th + 1), tp, &opts); 864a3433f35SChangli Gao if (likely((tcb->flags & TCPHDR_SYN) == 0)) 8659e412ba7SIlpo Järvinen TCP_ECN_send(sk, skb, tcp_header_size); 866dfb4b9dcSDavid S. Miller 867cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 868cfb6eeb4SYOSHIFUJI Hideaki /* Calculate the MD5 hash, as we have all we need now */ 869cfb6eeb4SYOSHIFUJI Hideaki if (md5) { 870a465419bSEric Dumazet sk_nocaps_add(sk, NETIF_F_GSO_MASK); 871bd0388aeSWilliam Allen Simpson tp->af_specific->calc_md5_hash(opts.hash_location, 87249a72dfbSAdam Langley md5, sk, NULL, skb); 873cfb6eeb4SYOSHIFUJI Hideaki } 874cfb6eeb4SYOSHIFUJI Hideaki #endif 875cfb6eeb4SYOSHIFUJI Hideaki 876bb296246SHerbert Xu icsk->icsk_af_ops->send_check(sk, skb); 8771da177e4SLinus Torvalds 878a3433f35SChangli Gao if (likely(tcb->flags & TCPHDR_ACK)) 879fc6415bcSDavid S. Miller tcp_event_ack_sent(sk, tcp_skb_pcount(skb)); 8801da177e4SLinus Torvalds 8811da177e4SLinus Torvalds if (skb->len != tcp_header_size) 8821da177e4SLinus Torvalds tcp_event_data_sent(tp, skb, sk); 8831da177e4SLinus Torvalds 884bd37a088SWei Yongjun if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq) 885aa2ea058STom Herbert TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS, 886aa2ea058STom Herbert tcp_skb_pcount(skb)); 8871da177e4SLinus Torvalds 8884e15ed4dSShan Wei err = icsk->icsk_af_ops->queue_xmit(skb); 88983de47cdSHua Zhong if (likely(err <= 0)) 8901da177e4SLinus Torvalds return err; 8911da177e4SLinus Torvalds 8923cfe3baaSIlpo Järvinen tcp_enter_cwr(sk, 1); 8931da177e4SLinus Torvalds 894b9df3cb8SGerrit Renker return net_xmit_eval(err); 8951da177e4SLinus Torvalds } 8961da177e4SLinus Torvalds 89767edfef7SAndi Kleen /* This routine just queues the buffer for sending. 8981da177e4SLinus Torvalds * 8991da177e4SLinus Torvalds * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames, 9001da177e4SLinus Torvalds * otherwise socket can stall. 9011da177e4SLinus Torvalds */ 9021da177e4SLinus Torvalds static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb) 9031da177e4SLinus Torvalds { 9041da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 9051da177e4SLinus Torvalds 9061da177e4SLinus Torvalds /* Advance write_seq and place onto the write_queue. */ 9071da177e4SLinus Torvalds tp->write_seq = TCP_SKB_CB(skb)->end_seq; 9081da177e4SLinus Torvalds skb_header_release(skb); 909fe067e8aSDavid S. Miller tcp_add_write_queue_tail(sk, skb); 9103ab224beSHideo Aoki sk->sk_wmem_queued += skb->truesize; 9113ab224beSHideo Aoki sk_mem_charge(sk, skb->truesize); 9121da177e4SLinus Torvalds } 9131da177e4SLinus Torvalds 91467edfef7SAndi Kleen /* Initialize TSO segments for a packet. */ 915056834d9SIlpo Järvinen static void tcp_set_skb_tso_segs(struct sock *sk, struct sk_buff *skb, 916056834d9SIlpo Järvinen unsigned int mss_now) 917f6302d1dSDavid S. Miller { 9188e5b9ddaSHerbert Xu if (skb->len <= mss_now || !sk_can_gso(sk) || 9198e5b9ddaSHerbert Xu skb->ip_summed == CHECKSUM_NONE) { 920f6302d1dSDavid S. Miller /* Avoid the costly divide in the normal 921f6302d1dSDavid S. Miller * non-TSO case. 922f6302d1dSDavid S. Miller */ 9237967168cSHerbert Xu skb_shinfo(skb)->gso_segs = 1; 9247967168cSHerbert Xu skb_shinfo(skb)->gso_size = 0; 9257967168cSHerbert Xu skb_shinfo(skb)->gso_type = 0; 926f6302d1dSDavid S. Miller } else { 927356f89e1SIlpo Järvinen skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss_now); 9287967168cSHerbert Xu skb_shinfo(skb)->gso_size = mss_now; 929bcd76111SHerbert Xu skb_shinfo(skb)->gso_type = sk->sk_gso_type; 9301da177e4SLinus Torvalds } 9311da177e4SLinus Torvalds } 9321da177e4SLinus Torvalds 93391fed7a1SIlpo Järvinen /* When a modification to fackets out becomes necessary, we need to check 93468f8353bSIlpo Järvinen * skb is counted to fackets_out or not. 93591fed7a1SIlpo Järvinen */ 936a47e5a98SIlpo Järvinen static void tcp_adjust_fackets_out(struct sock *sk, struct sk_buff *skb, 93791fed7a1SIlpo Järvinen int decr) 93891fed7a1SIlpo Järvinen { 939a47e5a98SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 940a47e5a98SIlpo Järvinen 941dc86967bSIlpo Järvinen if (!tp->sacked_out || tcp_is_reno(tp)) 94291fed7a1SIlpo Järvinen return; 94391fed7a1SIlpo Järvinen 9446859d494SIlpo Järvinen if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq)) 94591fed7a1SIlpo Järvinen tp->fackets_out -= decr; 94691fed7a1SIlpo Järvinen } 94791fed7a1SIlpo Järvinen 948797108d1SIlpo Järvinen /* Pcount in the middle of the write queue got changed, we need to do various 949797108d1SIlpo Järvinen * tweaks to fix counters 950797108d1SIlpo Järvinen */ 951797108d1SIlpo Järvinen static void tcp_adjust_pcount(struct sock *sk, struct sk_buff *skb, int decr) 952797108d1SIlpo Järvinen { 953797108d1SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 954797108d1SIlpo Järvinen 955797108d1SIlpo Järvinen tp->packets_out -= decr; 956797108d1SIlpo Järvinen 957797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 958797108d1SIlpo Järvinen tp->sacked_out -= decr; 959797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) 960797108d1SIlpo Järvinen tp->retrans_out -= decr; 961797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST) 962797108d1SIlpo Järvinen tp->lost_out -= decr; 963797108d1SIlpo Järvinen 964797108d1SIlpo Järvinen /* Reno case is special. Sigh... */ 965797108d1SIlpo Järvinen if (tcp_is_reno(tp) && decr > 0) 966797108d1SIlpo Järvinen tp->sacked_out -= min_t(u32, tp->sacked_out, decr); 967797108d1SIlpo Järvinen 968797108d1SIlpo Järvinen tcp_adjust_fackets_out(sk, skb, decr); 969797108d1SIlpo Järvinen 970797108d1SIlpo Järvinen if (tp->lost_skb_hint && 971797108d1SIlpo Järvinen before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) && 97252cf3cc8SIlpo Järvinen (tcp_is_fack(tp) || (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))) 973797108d1SIlpo Järvinen tp->lost_cnt_hint -= decr; 974797108d1SIlpo Järvinen 975797108d1SIlpo Järvinen tcp_verify_left_out(tp); 976797108d1SIlpo Järvinen } 977797108d1SIlpo Järvinen 9781da177e4SLinus Torvalds /* Function to create two new TCP segments. Shrinks the given segment 9791da177e4SLinus Torvalds * to the specified size and appends a new segment with the rest of the 9801da177e4SLinus Torvalds * packet to the list. This won't be called frequently, I hope. 9811da177e4SLinus Torvalds * Remember, these are still headerless SKBs at this point. 9821da177e4SLinus Torvalds */ 983056834d9SIlpo Järvinen int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len, 984056834d9SIlpo Järvinen unsigned int mss_now) 9851da177e4SLinus Torvalds { 9861da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 9871da177e4SLinus Torvalds struct sk_buff *buff; 9886475be16SDavid S. Miller int nsize, old_factor; 989b60b49eaSHerbert Xu int nlen; 9909ce01461SIlpo Järvinen u8 flags; 9911da177e4SLinus Torvalds 992b2cc99f0SHerbert Xu BUG_ON(len > skb->len); 9936a438bbeSStephen Hemminger 9941da177e4SLinus Torvalds nsize = skb_headlen(skb) - len; 9951da177e4SLinus Torvalds if (nsize < 0) 9961da177e4SLinus Torvalds nsize = 0; 9971da177e4SLinus Torvalds 9981da177e4SLinus Torvalds if (skb_cloned(skb) && 9991da177e4SLinus Torvalds skb_is_nonlinear(skb) && 10001da177e4SLinus Torvalds pskb_expand_head(skb, 0, 0, GFP_ATOMIC)) 10011da177e4SLinus Torvalds return -ENOMEM; 10021da177e4SLinus Torvalds 10031da177e4SLinus Torvalds /* Get a new skb... force flag on. */ 10041da177e4SLinus Torvalds buff = sk_stream_alloc_skb(sk, nsize, GFP_ATOMIC); 10051da177e4SLinus Torvalds if (buff == NULL) 10061da177e4SLinus Torvalds return -ENOMEM; /* We'll just try again later. */ 1007ef5cb973SHerbert Xu 10083ab224beSHideo Aoki sk->sk_wmem_queued += buff->truesize; 10093ab224beSHideo Aoki sk_mem_charge(sk, buff->truesize); 1010b60b49eaSHerbert Xu nlen = skb->len - len - nsize; 1011b60b49eaSHerbert Xu buff->truesize += nlen; 1012b60b49eaSHerbert Xu skb->truesize -= nlen; 10131da177e4SLinus Torvalds 10141da177e4SLinus Torvalds /* Correct the sequence numbers. */ 10151da177e4SLinus Torvalds TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; 10161da177e4SLinus Torvalds TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; 10171da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; 10181da177e4SLinus Torvalds 10191da177e4SLinus Torvalds /* PSH and FIN should only be set in the second packet. */ 10201da177e4SLinus Torvalds flags = TCP_SKB_CB(skb)->flags; 1021a3433f35SChangli Gao TCP_SKB_CB(skb)->flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); 10221da177e4SLinus Torvalds TCP_SKB_CB(buff)->flags = flags; 1023e14c3cafSHerbert Xu TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked; 10241da177e4SLinus Torvalds 102584fa7933SPatrick McHardy if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) { 10261da177e4SLinus Torvalds /* Copy and checksum data tail into the new buffer. */ 1027056834d9SIlpo Järvinen buff->csum = csum_partial_copy_nocheck(skb->data + len, 1028056834d9SIlpo Järvinen skb_put(buff, nsize), 10291da177e4SLinus Torvalds nsize, 0); 10301da177e4SLinus Torvalds 10311da177e4SLinus Torvalds skb_trim(skb, len); 10321da177e4SLinus Torvalds 10331da177e4SLinus Torvalds skb->csum = csum_block_sub(skb->csum, buff->csum, len); 10341da177e4SLinus Torvalds } else { 103584fa7933SPatrick McHardy skb->ip_summed = CHECKSUM_PARTIAL; 10361da177e4SLinus Torvalds skb_split(skb, buff, len); 10371da177e4SLinus Torvalds } 10381da177e4SLinus Torvalds 10391da177e4SLinus Torvalds buff->ip_summed = skb->ip_summed; 10401da177e4SLinus Torvalds 10411da177e4SLinus Torvalds /* Looks stupid, but our code really uses when of 10421da177e4SLinus Torvalds * skbs, which it never sent before. --ANK 10431da177e4SLinus Torvalds */ 10441da177e4SLinus Torvalds TCP_SKB_CB(buff)->when = TCP_SKB_CB(skb)->when; 1045a61bbcf2SPatrick McHardy buff->tstamp = skb->tstamp; 10461da177e4SLinus Torvalds 10476475be16SDavid S. Miller old_factor = tcp_skb_pcount(skb); 10486475be16SDavid S. Miller 10491da177e4SLinus Torvalds /* Fix up tso_factor for both original and new SKB. */ 1050846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, skb, mss_now); 1051846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, buff, mss_now); 10521da177e4SLinus Torvalds 10536475be16SDavid S. Miller /* If this packet has been sent out already, we must 10546475be16SDavid S. Miller * adjust the various packet counters. 10556475be16SDavid S. Miller */ 1056cf0b450cSHerbert Xu if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) { 10576475be16SDavid S. Miller int diff = old_factor - tcp_skb_pcount(skb) - 10586475be16SDavid S. Miller tcp_skb_pcount(buff); 10591da177e4SLinus Torvalds 1060797108d1SIlpo Järvinen if (diff) 1061797108d1SIlpo Järvinen tcp_adjust_pcount(sk, skb, diff); 10621da177e4SLinus Torvalds } 10631da177e4SLinus Torvalds 10641da177e4SLinus Torvalds /* Link BUFF into the send queue. */ 1065f44b5271SDavid S. Miller skb_header_release(buff); 1066fe067e8aSDavid S. Miller tcp_insert_write_queue_after(skb, buff, sk); 10671da177e4SLinus Torvalds 10681da177e4SLinus Torvalds return 0; 10691da177e4SLinus Torvalds } 10701da177e4SLinus Torvalds 10711da177e4SLinus Torvalds /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c 10721da177e4SLinus Torvalds * eventually). The difference is that pulled data not copied, but 10731da177e4SLinus Torvalds * immediately discarded. 10741da177e4SLinus Torvalds */ 1075f2911969SHerbert Xu ~{PmVHI~} static void __pskb_trim_head(struct sk_buff *skb, int len) 10761da177e4SLinus Torvalds { 10771da177e4SLinus Torvalds int i, k, eat; 10781da177e4SLinus Torvalds 10791da177e4SLinus Torvalds eat = len; 10801da177e4SLinus Torvalds k = 0; 10811da177e4SLinus Torvalds for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) { 10821da177e4SLinus Torvalds if (skb_shinfo(skb)->frags[i].size <= eat) { 10831da177e4SLinus Torvalds put_page(skb_shinfo(skb)->frags[i].page); 10841da177e4SLinus Torvalds eat -= skb_shinfo(skb)->frags[i].size; 10851da177e4SLinus Torvalds } else { 10861da177e4SLinus Torvalds skb_shinfo(skb)->frags[k] = skb_shinfo(skb)->frags[i]; 10871da177e4SLinus Torvalds if (eat) { 10881da177e4SLinus Torvalds skb_shinfo(skb)->frags[k].page_offset += eat; 10891da177e4SLinus Torvalds skb_shinfo(skb)->frags[k].size -= eat; 10901da177e4SLinus Torvalds eat = 0; 10911da177e4SLinus Torvalds } 10921da177e4SLinus Torvalds k++; 10931da177e4SLinus Torvalds } 10941da177e4SLinus Torvalds } 10951da177e4SLinus Torvalds skb_shinfo(skb)->nr_frags = k; 10961da177e4SLinus Torvalds 109727a884dcSArnaldo Carvalho de Melo skb_reset_tail_pointer(skb); 10981da177e4SLinus Torvalds skb->data_len -= len; 10991da177e4SLinus Torvalds skb->len = skb->data_len; 11001da177e4SLinus Torvalds } 11011da177e4SLinus Torvalds 110267edfef7SAndi Kleen /* Remove acked data from a packet in the transmit queue. */ 11031da177e4SLinus Torvalds int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len) 11041da177e4SLinus Torvalds { 1105056834d9SIlpo Järvinen if (skb_cloned(skb) && pskb_expand_head(skb, 0, 0, GFP_ATOMIC)) 11061da177e4SLinus Torvalds return -ENOMEM; 11071da177e4SLinus Torvalds 1108f2911969SHerbert Xu ~{PmVHI~} /* If len == headlen, we avoid __skb_pull to preserve alignment. */ 1109f2911969SHerbert Xu ~{PmVHI~} if (unlikely(len < skb_headlen(skb))) 11101da177e4SLinus Torvalds __skb_pull(skb, len); 1111f2911969SHerbert Xu ~{PmVHI~} else 1112f2911969SHerbert Xu ~{PmVHI~} __pskb_trim_head(skb, len - skb_headlen(skb)); 11131da177e4SLinus Torvalds 11141da177e4SLinus Torvalds TCP_SKB_CB(skb)->seq += len; 111584fa7933SPatrick McHardy skb->ip_summed = CHECKSUM_PARTIAL; 11161da177e4SLinus Torvalds 11171da177e4SLinus Torvalds skb->truesize -= len; 11181da177e4SLinus Torvalds sk->sk_wmem_queued -= len; 11193ab224beSHideo Aoki sk_mem_uncharge(sk, len); 11201da177e4SLinus Torvalds sock_set_flag(sk, SOCK_QUEUE_SHRUNK); 11211da177e4SLinus Torvalds 11221da177e4SLinus Torvalds /* Any change of skb->len requires recalculation of tso 11231da177e4SLinus Torvalds * factor and mss. 11241da177e4SLinus Torvalds */ 11251da177e4SLinus Torvalds if (tcp_skb_pcount(skb) > 1) 11260c54b85fSIlpo Järvinen tcp_set_skb_tso_segs(sk, skb, tcp_current_mss(sk)); 11271da177e4SLinus Torvalds 11281da177e4SLinus Torvalds return 0; 11291da177e4SLinus Torvalds } 11301da177e4SLinus Torvalds 113167edfef7SAndi Kleen /* Calculate MSS. Not accounting for SACKs here. */ 11325d424d5aSJohn Heffner int tcp_mtu_to_mss(struct sock *sk, int pmtu) 11335d424d5aSJohn Heffner { 11345d424d5aSJohn Heffner struct tcp_sock *tp = tcp_sk(sk); 11355d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 11365d424d5aSJohn Heffner int mss_now; 11375d424d5aSJohn Heffner 11385d424d5aSJohn Heffner /* Calculate base mss without TCP options: 11395d424d5aSJohn Heffner It is MMS_S - sizeof(tcphdr) of rfc1122 11405d424d5aSJohn Heffner */ 11415d424d5aSJohn Heffner mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr); 11425d424d5aSJohn Heffner 11435d424d5aSJohn Heffner /* Clamp it (mss_clamp does not include tcp options) */ 11445d424d5aSJohn Heffner if (mss_now > tp->rx_opt.mss_clamp) 11455d424d5aSJohn Heffner mss_now = tp->rx_opt.mss_clamp; 11465d424d5aSJohn Heffner 11475d424d5aSJohn Heffner /* Now subtract optional transport overhead */ 11485d424d5aSJohn Heffner mss_now -= icsk->icsk_ext_hdr_len; 11495d424d5aSJohn Heffner 11505d424d5aSJohn Heffner /* Then reserve room for full set of TCP options and 8 bytes of data */ 11515d424d5aSJohn Heffner if (mss_now < 48) 11525d424d5aSJohn Heffner mss_now = 48; 11535d424d5aSJohn Heffner 11545d424d5aSJohn Heffner /* Now subtract TCP options size, not including SACKs */ 11555d424d5aSJohn Heffner mss_now -= tp->tcp_header_len - sizeof(struct tcphdr); 11565d424d5aSJohn Heffner 11575d424d5aSJohn Heffner return mss_now; 11585d424d5aSJohn Heffner } 11595d424d5aSJohn Heffner 11605d424d5aSJohn Heffner /* Inverse of above */ 11615d424d5aSJohn Heffner int tcp_mss_to_mtu(struct sock *sk, int mss) 11625d424d5aSJohn Heffner { 11635d424d5aSJohn Heffner struct tcp_sock *tp = tcp_sk(sk); 11645d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 11655d424d5aSJohn Heffner int mtu; 11665d424d5aSJohn Heffner 11675d424d5aSJohn Heffner mtu = mss + 11685d424d5aSJohn Heffner tp->tcp_header_len + 11695d424d5aSJohn Heffner icsk->icsk_ext_hdr_len + 11705d424d5aSJohn Heffner icsk->icsk_af_ops->net_header_len; 11715d424d5aSJohn Heffner 11725d424d5aSJohn Heffner return mtu; 11735d424d5aSJohn Heffner } 11745d424d5aSJohn Heffner 117567edfef7SAndi Kleen /* MTU probing init per socket */ 11765d424d5aSJohn Heffner void tcp_mtup_init(struct sock *sk) 11775d424d5aSJohn Heffner { 11785d424d5aSJohn Heffner struct tcp_sock *tp = tcp_sk(sk); 11795d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 11805d424d5aSJohn Heffner 11815d424d5aSJohn Heffner icsk->icsk_mtup.enabled = sysctl_tcp_mtu_probing > 1; 11825d424d5aSJohn Heffner icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) + 11835d424d5aSJohn Heffner icsk->icsk_af_ops->net_header_len; 11845d424d5aSJohn Heffner icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, sysctl_tcp_base_mss); 11855d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = 0; 11865d424d5aSJohn Heffner } 11874bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_mtup_init); 11885d424d5aSJohn Heffner 11891da177e4SLinus Torvalds /* This function synchronize snd mss to current pmtu/exthdr set. 11901da177e4SLinus Torvalds 11911da177e4SLinus Torvalds tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts 11921da177e4SLinus Torvalds for TCP options, but includes only bare TCP header. 11931da177e4SLinus Torvalds 11941da177e4SLinus Torvalds tp->rx_opt.mss_clamp is mss negotiated at connection setup. 1195caa20d9aSStephen Hemminger It is minimum of user_mss and mss received with SYN. 11961da177e4SLinus Torvalds It also does not include TCP options. 11971da177e4SLinus Torvalds 1198d83d8461SArnaldo Carvalho de Melo inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function. 11991da177e4SLinus Torvalds 12001da177e4SLinus Torvalds tp->mss_cache is current effective sending mss, including 12011da177e4SLinus Torvalds all tcp options except for SACKs. It is evaluated, 12021da177e4SLinus Torvalds taking into account current pmtu, but never exceeds 12031da177e4SLinus Torvalds tp->rx_opt.mss_clamp. 12041da177e4SLinus Torvalds 12051da177e4SLinus Torvalds NOTE1. rfc1122 clearly states that advertised MSS 12061da177e4SLinus Torvalds DOES NOT include either tcp or ip options. 12071da177e4SLinus Torvalds 1208d83d8461SArnaldo Carvalho de Melo NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache 1209d83d8461SArnaldo Carvalho de Melo are READ ONLY outside this function. --ANK (980731) 12101da177e4SLinus Torvalds */ 12111da177e4SLinus Torvalds unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu) 12121da177e4SLinus Torvalds { 12131da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 1214d83d8461SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 12155d424d5aSJohn Heffner int mss_now; 12161da177e4SLinus Torvalds 12175d424d5aSJohn Heffner if (icsk->icsk_mtup.search_high > pmtu) 12185d424d5aSJohn Heffner icsk->icsk_mtup.search_high = pmtu; 12191da177e4SLinus Torvalds 12205d424d5aSJohn Heffner mss_now = tcp_mtu_to_mss(sk, pmtu); 1221409d22b4SIlpo Järvinen mss_now = tcp_bound_to_half_wnd(tp, mss_now); 12221da177e4SLinus Torvalds 12231da177e4SLinus Torvalds /* And store cached results */ 1224d83d8461SArnaldo Carvalho de Melo icsk->icsk_pmtu_cookie = pmtu; 12255d424d5aSJohn Heffner if (icsk->icsk_mtup.enabled) 12265d424d5aSJohn Heffner mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low)); 1227c1b4a7e6SDavid S. Miller tp->mss_cache = mss_now; 12281da177e4SLinus Torvalds 12291da177e4SLinus Torvalds return mss_now; 12301da177e4SLinus Torvalds } 12314bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_sync_mss); 12321da177e4SLinus Torvalds 12331da177e4SLinus Torvalds /* Compute the current effective MSS, taking SACKs and IP options, 12341da177e4SLinus Torvalds * and even PMTU discovery events into account. 12351da177e4SLinus Torvalds */ 12360c54b85fSIlpo Järvinen unsigned int tcp_current_mss(struct sock *sk) 12371da177e4SLinus Torvalds { 12381da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 12391da177e4SLinus Torvalds struct dst_entry *dst = __sk_dst_get(sk); 1240c1b4a7e6SDavid S. Miller u32 mss_now; 124133ad798cSAdam Langley unsigned header_len; 124233ad798cSAdam Langley struct tcp_out_options opts; 124333ad798cSAdam Langley struct tcp_md5sig_key *md5; 12441da177e4SLinus Torvalds 1245c1b4a7e6SDavid S. Miller mss_now = tp->mss_cache; 1246c1b4a7e6SDavid S. Miller 12471da177e4SLinus Torvalds if (dst) { 12481da177e4SLinus Torvalds u32 mtu = dst_mtu(dst); 1249d83d8461SArnaldo Carvalho de Melo if (mtu != inet_csk(sk)->icsk_pmtu_cookie) 12501da177e4SLinus Torvalds mss_now = tcp_sync_mss(sk, mtu); 12511da177e4SLinus Torvalds } 12521da177e4SLinus Torvalds 125333ad798cSAdam Langley header_len = tcp_established_options(sk, NULL, &opts, &md5) + 125433ad798cSAdam Langley sizeof(struct tcphdr); 125533ad798cSAdam Langley /* The mss_cache is sized based on tp->tcp_header_len, which assumes 125633ad798cSAdam Langley * some common options. If this is an odd packet (because we have SACK 125733ad798cSAdam Langley * blocks etc) then our calculated header_len will be different, and 125833ad798cSAdam Langley * we have to adjust mss_now correspondingly */ 125933ad798cSAdam Langley if (header_len != tp->tcp_header_len) { 126033ad798cSAdam Langley int delta = (int) header_len - tp->tcp_header_len; 126133ad798cSAdam Langley mss_now -= delta; 126233ad798cSAdam Langley } 1263cfb6eeb4SYOSHIFUJI Hideaki 12641da177e4SLinus Torvalds return mss_now; 12651da177e4SLinus Torvalds } 12661da177e4SLinus Torvalds 1267a762a980SDavid S. Miller /* Congestion window validation. (RFC2861) */ 12689e412ba7SIlpo Järvinen static void tcp_cwnd_validate(struct sock *sk) 1269a762a980SDavid S. Miller { 12709e412ba7SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 1271a762a980SDavid S. Miller 1272d436d686SIlpo Järvinen if (tp->packets_out >= tp->snd_cwnd) { 1273a762a980SDavid S. Miller /* Network is feed fully. */ 1274a762a980SDavid S. Miller tp->snd_cwnd_used = 0; 1275a762a980SDavid S. Miller tp->snd_cwnd_stamp = tcp_time_stamp; 1276a762a980SDavid S. Miller } else { 1277a762a980SDavid S. Miller /* Network starves. */ 1278a762a980SDavid S. Miller if (tp->packets_out > tp->snd_cwnd_used) 1279a762a980SDavid S. Miller tp->snd_cwnd_used = tp->packets_out; 1280a762a980SDavid S. Miller 128115d33c07SDavid S. Miller if (sysctl_tcp_slow_start_after_idle && 128215d33c07SDavid S. Miller (s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto) 1283a762a980SDavid S. Miller tcp_cwnd_application_limited(sk); 1284a762a980SDavid S. Miller } 1285a762a980SDavid S. Miller } 1286a762a980SDavid S. Miller 12870e3a4803SIlpo Järvinen /* Returns the portion of skb which can be sent right away without 12880e3a4803SIlpo Järvinen * introducing MSS oddities to segment boundaries. In rare cases where 12890e3a4803SIlpo Järvinen * mss_now != mss_cache, we will request caller to create a small skb 12900e3a4803SIlpo Järvinen * per input skb which could be mostly avoided here (if desired). 12915ea3a748SIlpo Järvinen * 12925ea3a748SIlpo Järvinen * We explicitly want to create a request for splitting write queue tail 12935ea3a748SIlpo Järvinen * to a small skb for Nagle purposes while avoiding unnecessary modulos, 12945ea3a748SIlpo Järvinen * thus all the complexity (cwnd_len is always MSS multiple which we 12955ea3a748SIlpo Järvinen * return whenever allowed by the other factors). Basically we need the 12965ea3a748SIlpo Järvinen * modulo only when the receiver window alone is the limiting factor or 12975ea3a748SIlpo Järvinen * when we would be allowed to send the split-due-to-Nagle skb fully. 12980e3a4803SIlpo Järvinen */ 12990e3a4803SIlpo Järvinen static unsigned int tcp_mss_split_point(struct sock *sk, struct sk_buff *skb, 1300056834d9SIlpo Järvinen unsigned int mss_now, unsigned int cwnd) 1301c1b4a7e6SDavid S. Miller { 13020e3a4803SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 13030e3a4803SIlpo Järvinen u32 needed, window, cwnd_len; 1304c1b4a7e6SDavid S. Miller 130590840defSIlpo Järvinen window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 1306c1b4a7e6SDavid S. Miller cwnd_len = mss_now * cwnd; 13070e3a4803SIlpo Järvinen 13080e3a4803SIlpo Järvinen if (likely(cwnd_len <= window && skb != tcp_write_queue_tail(sk))) 13090e3a4803SIlpo Järvinen return cwnd_len; 13100e3a4803SIlpo Järvinen 13115ea3a748SIlpo Järvinen needed = min(skb->len, window); 13125ea3a748SIlpo Järvinen 131317515408SIlpo Järvinen if (cwnd_len <= needed) 13140e3a4803SIlpo Järvinen return cwnd_len; 13150e3a4803SIlpo Järvinen 13160e3a4803SIlpo Järvinen return needed - needed % mss_now; 1317c1b4a7e6SDavid S. Miller } 1318c1b4a7e6SDavid S. Miller 1319c1b4a7e6SDavid S. Miller /* Can at least one segment of SKB be sent right now, according to the 1320c1b4a7e6SDavid S. Miller * congestion window rules? If so, return how many segments are allowed. 1321c1b4a7e6SDavid S. Miller */ 1322056834d9SIlpo Järvinen static inline unsigned int tcp_cwnd_test(struct tcp_sock *tp, 1323056834d9SIlpo Järvinen struct sk_buff *skb) 1324c1b4a7e6SDavid S. Miller { 1325c1b4a7e6SDavid S. Miller u32 in_flight, cwnd; 1326c1b4a7e6SDavid S. Miller 1327c1b4a7e6SDavid S. Miller /* Don't be strict about the congestion window for the final FIN. */ 1328a3433f35SChangli Gao if ((TCP_SKB_CB(skb)->flags & TCPHDR_FIN) && tcp_skb_pcount(skb) == 1) 1329c1b4a7e6SDavid S. Miller return 1; 1330c1b4a7e6SDavid S. Miller 1331c1b4a7e6SDavid S. Miller in_flight = tcp_packets_in_flight(tp); 1332c1b4a7e6SDavid S. Miller cwnd = tp->snd_cwnd; 1333c1b4a7e6SDavid S. Miller if (in_flight < cwnd) 1334c1b4a7e6SDavid S. Miller return (cwnd - in_flight); 1335c1b4a7e6SDavid S. Miller 1336c1b4a7e6SDavid S. Miller return 0; 1337c1b4a7e6SDavid S. Miller } 1338c1b4a7e6SDavid S. Miller 133967edfef7SAndi Kleen /* Intialize TSO state of a skb. 134067edfef7SAndi Kleen * This must be invoked the first time we consider transmitting 1341c1b4a7e6SDavid S. Miller * SKB onto the wire. 1342c1b4a7e6SDavid S. Miller */ 1343056834d9SIlpo Järvinen static int tcp_init_tso_segs(struct sock *sk, struct sk_buff *skb, 1344056834d9SIlpo Järvinen unsigned int mss_now) 1345c1b4a7e6SDavid S. Miller { 1346c1b4a7e6SDavid S. Miller int tso_segs = tcp_skb_pcount(skb); 1347c1b4a7e6SDavid S. Miller 1348f8269a49SIlpo Järvinen if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) { 1349846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, skb, mss_now); 1350c1b4a7e6SDavid S. Miller tso_segs = tcp_skb_pcount(skb); 1351c1b4a7e6SDavid S. Miller } 1352c1b4a7e6SDavid S. Miller return tso_segs; 1353c1b4a7e6SDavid S. Miller } 1354c1b4a7e6SDavid S. Miller 135567edfef7SAndi Kleen /* Minshall's variant of the Nagle send check. */ 1356c1b4a7e6SDavid S. Miller static inline int tcp_minshall_check(const struct tcp_sock *tp) 1357c1b4a7e6SDavid S. Miller { 1358c1b4a7e6SDavid S. Miller return after(tp->snd_sml, tp->snd_una) && 1359c1b4a7e6SDavid S. Miller !after(tp->snd_sml, tp->snd_nxt); 1360c1b4a7e6SDavid S. Miller } 1361c1b4a7e6SDavid S. Miller 1362c1b4a7e6SDavid S. Miller /* Return 0, if packet can be sent now without violation Nagle's rules: 1363c1b4a7e6SDavid S. Miller * 1. It is full sized. 1364c1b4a7e6SDavid S. Miller * 2. Or it contains FIN. (already checked by caller) 1365c1b4a7e6SDavid S. Miller * 3. Or TCP_NODELAY was set. 1366c1b4a7e6SDavid S. Miller * 4. Or TCP_CORK is not set, and all sent packets are ACKed. 1367c1b4a7e6SDavid S. Miller * With Minshall's modification: all sent small packets are ACKed. 1368c1b4a7e6SDavid S. Miller */ 1369c1b4a7e6SDavid S. Miller static inline int tcp_nagle_check(const struct tcp_sock *tp, 1370c1b4a7e6SDavid S. Miller const struct sk_buff *skb, 1371c1b4a7e6SDavid S. Miller unsigned mss_now, int nonagle) 1372c1b4a7e6SDavid S. Miller { 1373*a02cec21SEric Dumazet return skb->len < mss_now && 1374c1b4a7e6SDavid S. Miller ((nonagle & TCP_NAGLE_CORK) || 1375*a02cec21SEric Dumazet (!nonagle && tp->packets_out && tcp_minshall_check(tp))); 1376c1b4a7e6SDavid S. Miller } 1377c1b4a7e6SDavid S. Miller 1378c1b4a7e6SDavid S. Miller /* Return non-zero if the Nagle test allows this packet to be 1379c1b4a7e6SDavid S. Miller * sent now. 1380c1b4a7e6SDavid S. Miller */ 1381c1b4a7e6SDavid S. Miller static inline int tcp_nagle_test(struct tcp_sock *tp, struct sk_buff *skb, 1382c1b4a7e6SDavid S. Miller unsigned int cur_mss, int nonagle) 1383c1b4a7e6SDavid S. Miller { 1384c1b4a7e6SDavid S. Miller /* Nagle rule does not apply to frames, which sit in the middle of the 1385c1b4a7e6SDavid S. Miller * write_queue (they have no chances to get new data). 1386c1b4a7e6SDavid S. Miller * 1387c1b4a7e6SDavid S. Miller * This is implemented in the callers, where they modify the 'nonagle' 1388c1b4a7e6SDavid S. Miller * argument based upon the location of SKB in the send queue. 1389c1b4a7e6SDavid S. Miller */ 1390c1b4a7e6SDavid S. Miller if (nonagle & TCP_NAGLE_PUSH) 1391c1b4a7e6SDavid S. Miller return 1; 1392c1b4a7e6SDavid S. Miller 1393d551e454SIlpo Järvinen /* Don't use the nagle rule for urgent data (or for the final FIN). 1394d551e454SIlpo Järvinen * Nagle can be ignored during F-RTO too (see RFC4138). 1395d551e454SIlpo Järvinen */ 139633f5f57eSIlpo Järvinen if (tcp_urg_mode(tp) || (tp->frto_counter == 2) || 1397a3433f35SChangli Gao (TCP_SKB_CB(skb)->flags & TCPHDR_FIN)) 1398c1b4a7e6SDavid S. Miller return 1; 1399c1b4a7e6SDavid S. Miller 1400c1b4a7e6SDavid S. Miller if (!tcp_nagle_check(tp, skb, cur_mss, nonagle)) 1401c1b4a7e6SDavid S. Miller return 1; 1402c1b4a7e6SDavid S. Miller 1403c1b4a7e6SDavid S. Miller return 0; 1404c1b4a7e6SDavid S. Miller } 1405c1b4a7e6SDavid S. Miller 1406c1b4a7e6SDavid S. Miller /* Does at least the first segment of SKB fit into the send window? */ 1407056834d9SIlpo Järvinen static inline int tcp_snd_wnd_test(struct tcp_sock *tp, struct sk_buff *skb, 1408056834d9SIlpo Järvinen unsigned int cur_mss) 1409c1b4a7e6SDavid S. Miller { 1410c1b4a7e6SDavid S. Miller u32 end_seq = TCP_SKB_CB(skb)->end_seq; 1411c1b4a7e6SDavid S. Miller 1412c1b4a7e6SDavid S. Miller if (skb->len > cur_mss) 1413c1b4a7e6SDavid S. Miller end_seq = TCP_SKB_CB(skb)->seq + cur_mss; 1414c1b4a7e6SDavid S. Miller 141590840defSIlpo Järvinen return !after(end_seq, tcp_wnd_end(tp)); 1416c1b4a7e6SDavid S. Miller } 1417c1b4a7e6SDavid S. Miller 1418fe067e8aSDavid S. Miller /* This checks if the data bearing packet SKB (usually tcp_send_head(sk)) 1419c1b4a7e6SDavid S. Miller * should be put on the wire right now. If so, it returns the number of 1420c1b4a7e6SDavid S. Miller * packets allowed by the congestion window. 1421c1b4a7e6SDavid S. Miller */ 1422c1b4a7e6SDavid S. Miller static unsigned int tcp_snd_test(struct sock *sk, struct sk_buff *skb, 1423c1b4a7e6SDavid S. Miller unsigned int cur_mss, int nonagle) 1424c1b4a7e6SDavid S. Miller { 1425c1b4a7e6SDavid S. Miller struct tcp_sock *tp = tcp_sk(sk); 1426c1b4a7e6SDavid S. Miller unsigned int cwnd_quota; 1427c1b4a7e6SDavid S. Miller 1428846998aeSDavid S. Miller tcp_init_tso_segs(sk, skb, cur_mss); 1429c1b4a7e6SDavid S. Miller 1430c1b4a7e6SDavid S. Miller if (!tcp_nagle_test(tp, skb, cur_mss, nonagle)) 1431c1b4a7e6SDavid S. Miller return 0; 1432c1b4a7e6SDavid S. Miller 1433c1b4a7e6SDavid S. Miller cwnd_quota = tcp_cwnd_test(tp, skb); 1434056834d9SIlpo Järvinen if (cwnd_quota && !tcp_snd_wnd_test(tp, skb, cur_mss)) 1435c1b4a7e6SDavid S. Miller cwnd_quota = 0; 1436c1b4a7e6SDavid S. Miller 1437c1b4a7e6SDavid S. Miller return cwnd_quota; 1438c1b4a7e6SDavid S. Miller } 1439c1b4a7e6SDavid S. Miller 144067edfef7SAndi Kleen /* Test if sending is allowed right now. */ 14419e412ba7SIlpo Järvinen int tcp_may_send_now(struct sock *sk) 1442c1b4a7e6SDavid S. Miller { 14439e412ba7SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 1444fe067e8aSDavid S. Miller struct sk_buff *skb = tcp_send_head(sk); 1445c1b4a7e6SDavid S. Miller 1446*a02cec21SEric Dumazet return skb && 14470c54b85fSIlpo Järvinen tcp_snd_test(sk, skb, tcp_current_mss(sk), 1448c1b4a7e6SDavid S. Miller (tcp_skb_is_last(sk, skb) ? 1449*a02cec21SEric Dumazet tp->nonagle : TCP_NAGLE_PUSH)); 1450c1b4a7e6SDavid S. Miller } 1451c1b4a7e6SDavid S. Miller 1452c1b4a7e6SDavid S. Miller /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet 1453c1b4a7e6SDavid S. Miller * which is put after SKB on the list. It is very much like 1454c1b4a7e6SDavid S. Miller * tcp_fragment() except that it may make several kinds of assumptions 1455c1b4a7e6SDavid S. Miller * in order to speed up the splitting operation. In particular, we 1456c1b4a7e6SDavid S. Miller * know that all the data is in scatter-gather pages, and that the 1457c1b4a7e6SDavid S. Miller * packet has never been sent out before (and thus is not cloned). 1458c1b4a7e6SDavid S. Miller */ 1459056834d9SIlpo Järvinen static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len, 1460c4ead4c5SEric Dumazet unsigned int mss_now, gfp_t gfp) 1461c1b4a7e6SDavid S. Miller { 1462c1b4a7e6SDavid S. Miller struct sk_buff *buff; 1463c1b4a7e6SDavid S. Miller int nlen = skb->len - len; 14649ce01461SIlpo Järvinen u8 flags; 1465c1b4a7e6SDavid S. Miller 1466c1b4a7e6SDavid S. Miller /* All of a TSO frame must be composed of paged data. */ 1467c8ac3774SHerbert Xu if (skb->len != skb->data_len) 1468c8ac3774SHerbert Xu return tcp_fragment(sk, skb, len, mss_now); 1469c1b4a7e6SDavid S. Miller 1470c4ead4c5SEric Dumazet buff = sk_stream_alloc_skb(sk, 0, gfp); 1471c1b4a7e6SDavid S. Miller if (unlikely(buff == NULL)) 1472c1b4a7e6SDavid S. Miller return -ENOMEM; 1473c1b4a7e6SDavid S. Miller 14743ab224beSHideo Aoki sk->sk_wmem_queued += buff->truesize; 14753ab224beSHideo Aoki sk_mem_charge(sk, buff->truesize); 1476b60b49eaSHerbert Xu buff->truesize += nlen; 1477c1b4a7e6SDavid S. Miller skb->truesize -= nlen; 1478c1b4a7e6SDavid S. Miller 1479c1b4a7e6SDavid S. Miller /* Correct the sequence numbers. */ 1480c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; 1481c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; 1482c1b4a7e6SDavid S. Miller TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; 1483c1b4a7e6SDavid S. Miller 1484c1b4a7e6SDavid S. Miller /* PSH and FIN should only be set in the second packet. */ 1485c1b4a7e6SDavid S. Miller flags = TCP_SKB_CB(skb)->flags; 1486a3433f35SChangli Gao TCP_SKB_CB(skb)->flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); 1487c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->flags = flags; 1488c1b4a7e6SDavid S. Miller 1489c1b4a7e6SDavid S. Miller /* This packet was never sent out yet, so no SACK bits. */ 1490c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->sacked = 0; 1491c1b4a7e6SDavid S. Miller 149284fa7933SPatrick McHardy buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL; 1493c1b4a7e6SDavid S. Miller skb_split(skb, buff, len); 1494c1b4a7e6SDavid S. Miller 1495c1b4a7e6SDavid S. Miller /* Fix up tso_factor for both original and new SKB. */ 1496846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, skb, mss_now); 1497846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, buff, mss_now); 1498c1b4a7e6SDavid S. Miller 1499c1b4a7e6SDavid S. Miller /* Link BUFF into the send queue. */ 1500c1b4a7e6SDavid S. Miller skb_header_release(buff); 1501fe067e8aSDavid S. Miller tcp_insert_write_queue_after(skb, buff, sk); 1502c1b4a7e6SDavid S. Miller 1503c1b4a7e6SDavid S. Miller return 0; 1504c1b4a7e6SDavid S. Miller } 1505c1b4a7e6SDavid S. Miller 1506c1b4a7e6SDavid S. Miller /* Try to defer sending, if possible, in order to minimize the amount 1507c1b4a7e6SDavid S. Miller * of TSO splitting we do. View it as a kind of TSO Nagle test. 1508c1b4a7e6SDavid S. Miller * 1509c1b4a7e6SDavid S. Miller * This algorithm is from John Heffner. 1510c1b4a7e6SDavid S. Miller */ 15119e412ba7SIlpo Järvinen static int tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb) 1512c1b4a7e6SDavid S. Miller { 15139e412ba7SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 15146687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 1515c1b4a7e6SDavid S. Miller u32 send_win, cong_win, limit, in_flight; 1516c1b4a7e6SDavid S. Miller 1517a3433f35SChangli Gao if (TCP_SKB_CB(skb)->flags & TCPHDR_FIN) 1518ae8064acSJohn Heffner goto send_now; 1519c1b4a7e6SDavid S. Miller 15206687e988SArnaldo Carvalho de Melo if (icsk->icsk_ca_state != TCP_CA_Open) 1521ae8064acSJohn Heffner goto send_now; 1522ae8064acSJohn Heffner 1523ae8064acSJohn Heffner /* Defer for less than two clock ticks. */ 1524bd515c3eSIlpo Järvinen if (tp->tso_deferred && 1525a2acde07SIlpo Järvinen (((u32)jiffies << 1) >> 1) - (tp->tso_deferred >> 1) > 1) 1526ae8064acSJohn Heffner goto send_now; 1527908a75c1SDavid S. Miller 1528c1b4a7e6SDavid S. Miller in_flight = tcp_packets_in_flight(tp); 1529c1b4a7e6SDavid S. Miller 1530056834d9SIlpo Järvinen BUG_ON(tcp_skb_pcount(skb) <= 1 || (tp->snd_cwnd <= in_flight)); 1531c1b4a7e6SDavid S. Miller 153290840defSIlpo Järvinen send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 1533c1b4a7e6SDavid S. Miller 1534c1b4a7e6SDavid S. Miller /* From in_flight test above, we know that cwnd > in_flight. */ 1535c1b4a7e6SDavid S. Miller cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache; 1536c1b4a7e6SDavid S. Miller 1537c1b4a7e6SDavid S. Miller limit = min(send_win, cong_win); 1538c1b4a7e6SDavid S. Miller 1539ba244fe9SDavid S. Miller /* If a full-sized TSO skb can be sent, do it. */ 154082cc1a7aSPeter P Waskiewicz Jr if (limit >= sk->sk_gso_max_size) 1541ae8064acSJohn Heffner goto send_now; 1542ba244fe9SDavid S. Miller 154362ad2761SIlpo Järvinen /* Middle in queue won't get any more data, full sendable already? */ 154462ad2761SIlpo Järvinen if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len)) 154562ad2761SIlpo Järvinen goto send_now; 154662ad2761SIlpo Järvinen 1547c1b4a7e6SDavid S. Miller if (sysctl_tcp_tso_win_divisor) { 1548c1b4a7e6SDavid S. Miller u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache); 1549c1b4a7e6SDavid S. Miller 1550c1b4a7e6SDavid S. Miller /* If at least some fraction of a window is available, 1551c1b4a7e6SDavid S. Miller * just use it. 1552c1b4a7e6SDavid S. Miller */ 1553c1b4a7e6SDavid S. Miller chunk /= sysctl_tcp_tso_win_divisor; 1554c1b4a7e6SDavid S. Miller if (limit >= chunk) 1555ae8064acSJohn Heffner goto send_now; 1556c1b4a7e6SDavid S. Miller } else { 1557c1b4a7e6SDavid S. Miller /* Different approach, try not to defer past a single 1558c1b4a7e6SDavid S. Miller * ACK. Receiver should ACK every other full sized 1559c1b4a7e6SDavid S. Miller * frame, so if we have space for more than 3 frames 1560c1b4a7e6SDavid S. Miller * then send now. 1561c1b4a7e6SDavid S. Miller */ 1562c1b4a7e6SDavid S. Miller if (limit > tcp_max_burst(tp) * tp->mss_cache) 1563ae8064acSJohn Heffner goto send_now; 1564c1b4a7e6SDavid S. Miller } 1565c1b4a7e6SDavid S. Miller 1566c1b4a7e6SDavid S. Miller /* Ok, it looks like it is advisable to defer. */ 1567ae8064acSJohn Heffner tp->tso_deferred = 1 | (jiffies << 1); 1568ae8064acSJohn Heffner 1569c1b4a7e6SDavid S. Miller return 1; 1570ae8064acSJohn Heffner 1571ae8064acSJohn Heffner send_now: 1572ae8064acSJohn Heffner tp->tso_deferred = 0; 1573ae8064acSJohn Heffner return 0; 1574c1b4a7e6SDavid S. Miller } 1575c1b4a7e6SDavid S. Miller 15765d424d5aSJohn Heffner /* Create a new MTU probe if we are ready. 157767edfef7SAndi Kleen * MTU probe is regularly attempting to increase the path MTU by 157867edfef7SAndi Kleen * deliberately sending larger packets. This discovers routing 157967edfef7SAndi Kleen * changes resulting in larger path MTUs. 158067edfef7SAndi Kleen * 15815d424d5aSJohn Heffner * Returns 0 if we should wait to probe (no cwnd available), 15825d424d5aSJohn Heffner * 1 if a probe was sent, 1583056834d9SIlpo Järvinen * -1 otherwise 1584056834d9SIlpo Järvinen */ 15855d424d5aSJohn Heffner static int tcp_mtu_probe(struct sock *sk) 15865d424d5aSJohn Heffner { 15875d424d5aSJohn Heffner struct tcp_sock *tp = tcp_sk(sk); 15885d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 15895d424d5aSJohn Heffner struct sk_buff *skb, *nskb, *next; 15905d424d5aSJohn Heffner int len; 15915d424d5aSJohn Heffner int probe_size; 159291cc17c0SIlpo Järvinen int size_needed; 15935d424d5aSJohn Heffner int copy; 15945d424d5aSJohn Heffner int mss_now; 15955d424d5aSJohn Heffner 15965d424d5aSJohn Heffner /* Not currently probing/verifying, 15975d424d5aSJohn Heffner * not in recovery, 15985d424d5aSJohn Heffner * have enough cwnd, and 15995d424d5aSJohn Heffner * not SACKing (the variable headers throw things off) */ 16005d424d5aSJohn Heffner if (!icsk->icsk_mtup.enabled || 16015d424d5aSJohn Heffner icsk->icsk_mtup.probe_size || 16025d424d5aSJohn Heffner inet_csk(sk)->icsk_ca_state != TCP_CA_Open || 16035d424d5aSJohn Heffner tp->snd_cwnd < 11 || 1604cabeccbdSIlpo Järvinen tp->rx_opt.num_sacks || tp->rx_opt.dsack) 16055d424d5aSJohn Heffner return -1; 16065d424d5aSJohn Heffner 16075d424d5aSJohn Heffner /* Very simple search strategy: just double the MSS. */ 16080c54b85fSIlpo Järvinen mss_now = tcp_current_mss(sk); 16095d424d5aSJohn Heffner probe_size = 2 * tp->mss_cache; 161091cc17c0SIlpo Järvinen size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache; 16115d424d5aSJohn Heffner if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high)) { 16125d424d5aSJohn Heffner /* TODO: set timer for probe_converge_event */ 16135d424d5aSJohn Heffner return -1; 16145d424d5aSJohn Heffner } 16155d424d5aSJohn Heffner 16165d424d5aSJohn Heffner /* Have enough data in the send queue to probe? */ 16177f9c33e5SIlpo Järvinen if (tp->write_seq - tp->snd_nxt < size_needed) 16185d424d5aSJohn Heffner return -1; 16195d424d5aSJohn Heffner 162091cc17c0SIlpo Järvinen if (tp->snd_wnd < size_needed) 16215d424d5aSJohn Heffner return -1; 162290840defSIlpo Järvinen if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp))) 16235d424d5aSJohn Heffner return 0; 16245d424d5aSJohn Heffner 1625d67c58e9SIlpo Järvinen /* Do we need to wait to drain cwnd? With none in flight, don't stall */ 1626d67c58e9SIlpo Järvinen if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) { 1627d67c58e9SIlpo Järvinen if (!tcp_packets_in_flight(tp)) 16285d424d5aSJohn Heffner return -1; 16295d424d5aSJohn Heffner else 16305d424d5aSJohn Heffner return 0; 16315d424d5aSJohn Heffner } 16325d424d5aSJohn Heffner 16335d424d5aSJohn Heffner /* We're allowed to probe. Build it now. */ 16345d424d5aSJohn Heffner if ((nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC)) == NULL) 16355d424d5aSJohn Heffner return -1; 16363ab224beSHideo Aoki sk->sk_wmem_queued += nskb->truesize; 16373ab224beSHideo Aoki sk_mem_charge(sk, nskb->truesize); 16385d424d5aSJohn Heffner 1639fe067e8aSDavid S. Miller skb = tcp_send_head(sk); 16405d424d5aSJohn Heffner 16415d424d5aSJohn Heffner TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq; 16425d424d5aSJohn Heffner TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size; 1643a3433f35SChangli Gao TCP_SKB_CB(nskb)->flags = TCPHDR_ACK; 16445d424d5aSJohn Heffner TCP_SKB_CB(nskb)->sacked = 0; 16455d424d5aSJohn Heffner nskb->csum = 0; 164684fa7933SPatrick McHardy nskb->ip_summed = skb->ip_summed; 16475d424d5aSJohn Heffner 164850c4817eSIlpo Järvinen tcp_insert_write_queue_before(nskb, skb, sk); 164950c4817eSIlpo Järvinen 16505d424d5aSJohn Heffner len = 0; 1651234b6860SIlpo Järvinen tcp_for_write_queue_from_safe(skb, next, sk) { 16525d424d5aSJohn Heffner copy = min_t(int, skb->len, probe_size - len); 16535d424d5aSJohn Heffner if (nskb->ip_summed) 16545d424d5aSJohn Heffner skb_copy_bits(skb, 0, skb_put(nskb, copy), copy); 16555d424d5aSJohn Heffner else 16565d424d5aSJohn Heffner nskb->csum = skb_copy_and_csum_bits(skb, 0, 1657056834d9SIlpo Järvinen skb_put(nskb, copy), 1658056834d9SIlpo Järvinen copy, nskb->csum); 16595d424d5aSJohn Heffner 16605d424d5aSJohn Heffner if (skb->len <= copy) { 16615d424d5aSJohn Heffner /* We've eaten all the data from this skb. 16625d424d5aSJohn Heffner * Throw it away. */ 16635d424d5aSJohn Heffner TCP_SKB_CB(nskb)->flags |= TCP_SKB_CB(skb)->flags; 1664fe067e8aSDavid S. Miller tcp_unlink_write_queue(skb, sk); 16653ab224beSHideo Aoki sk_wmem_free_skb(sk, skb); 16665d424d5aSJohn Heffner } else { 16675d424d5aSJohn Heffner TCP_SKB_CB(nskb)->flags |= TCP_SKB_CB(skb)->flags & 1668a3433f35SChangli Gao ~(TCPHDR_FIN|TCPHDR_PSH); 16695d424d5aSJohn Heffner if (!skb_shinfo(skb)->nr_frags) { 16705d424d5aSJohn Heffner skb_pull(skb, copy); 167184fa7933SPatrick McHardy if (skb->ip_summed != CHECKSUM_PARTIAL) 1672056834d9SIlpo Järvinen skb->csum = csum_partial(skb->data, 1673056834d9SIlpo Järvinen skb->len, 0); 16745d424d5aSJohn Heffner } else { 16755d424d5aSJohn Heffner __pskb_trim_head(skb, copy); 16765d424d5aSJohn Heffner tcp_set_skb_tso_segs(sk, skb, mss_now); 16775d424d5aSJohn Heffner } 16785d424d5aSJohn Heffner TCP_SKB_CB(skb)->seq += copy; 16795d424d5aSJohn Heffner } 16805d424d5aSJohn Heffner 16815d424d5aSJohn Heffner len += copy; 1682234b6860SIlpo Järvinen 1683234b6860SIlpo Järvinen if (len >= probe_size) 1684234b6860SIlpo Järvinen break; 16855d424d5aSJohn Heffner } 16865d424d5aSJohn Heffner tcp_init_tso_segs(sk, nskb, nskb->len); 16875d424d5aSJohn Heffner 16885d424d5aSJohn Heffner /* We're ready to send. If this fails, the probe will 16895d424d5aSJohn Heffner * be resegmented into mss-sized pieces by tcp_write_xmit(). */ 16905d424d5aSJohn Heffner TCP_SKB_CB(nskb)->when = tcp_time_stamp; 16915d424d5aSJohn Heffner if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) { 16925d424d5aSJohn Heffner /* Decrement cwnd here because we are sending 16935d424d5aSJohn Heffner * effectively two packets. */ 16945d424d5aSJohn Heffner tp->snd_cwnd--; 169566f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, nskb); 16965d424d5aSJohn Heffner 16975d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len); 16980e7b1368SJohn Heffner tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq; 16990e7b1368SJohn Heffner tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq; 17005d424d5aSJohn Heffner 17015d424d5aSJohn Heffner return 1; 17025d424d5aSJohn Heffner } 17035d424d5aSJohn Heffner 17045d424d5aSJohn Heffner return -1; 17055d424d5aSJohn Heffner } 17065d424d5aSJohn Heffner 17071da177e4SLinus Torvalds /* This routine writes packets to the network. It advances the 17081da177e4SLinus Torvalds * send_head. This happens as incoming acks open up the remote 17091da177e4SLinus Torvalds * window for us. 17101da177e4SLinus Torvalds * 1711f8269a49SIlpo Järvinen * LARGESEND note: !tcp_urg_mode is overkill, only frames between 1712f8269a49SIlpo Järvinen * snd_up-64k-mss .. snd_up cannot be large. However, taking into 1713f8269a49SIlpo Järvinen * account rare use of URG, this is not a big flaw. 1714f8269a49SIlpo Järvinen * 17151da177e4SLinus Torvalds * Returns 1, if no segments are in flight and we have queued segments, but 17161da177e4SLinus Torvalds * cannot send anything now because of SWS or another problem. 17171da177e4SLinus Torvalds */ 1718d5dd9175SIlpo Järvinen static int tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle, 1719d5dd9175SIlpo Järvinen int push_one, gfp_t gfp) 17201da177e4SLinus Torvalds { 17211da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 172292df7b51SDavid S. Miller struct sk_buff *skb; 1723c1b4a7e6SDavid S. Miller unsigned int tso_segs, sent_pkts; 1724c1b4a7e6SDavid S. Miller int cwnd_quota; 17255d424d5aSJohn Heffner int result; 17261da177e4SLinus Torvalds 1727c1b4a7e6SDavid S. Miller sent_pkts = 0; 17285d424d5aSJohn Heffner 1729d5dd9175SIlpo Järvinen if (!push_one) { 17305d424d5aSJohn Heffner /* Do MTU probing. */ 1731d5dd9175SIlpo Järvinen result = tcp_mtu_probe(sk); 1732d5dd9175SIlpo Järvinen if (!result) { 17335d424d5aSJohn Heffner return 0; 17345d424d5aSJohn Heffner } else if (result > 0) { 17355d424d5aSJohn Heffner sent_pkts = 1; 17365d424d5aSJohn Heffner } 1737d5dd9175SIlpo Järvinen } 17385d424d5aSJohn Heffner 1739fe067e8aSDavid S. Miller while ((skb = tcp_send_head(sk))) { 1740c8ac3774SHerbert Xu unsigned int limit; 1741c8ac3774SHerbert Xu 1742b68e9f85SHerbert Xu tso_segs = tcp_init_tso_segs(sk, skb, mss_now); 1743c1b4a7e6SDavid S. Miller BUG_ON(!tso_segs); 1744c1b4a7e6SDavid S. Miller 1745b68e9f85SHerbert Xu cwnd_quota = tcp_cwnd_test(tp, skb); 1746b68e9f85SHerbert Xu if (!cwnd_quota) 1747b68e9f85SHerbert Xu break; 1748b68e9f85SHerbert Xu 1749b68e9f85SHerbert Xu if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now))) 1750b68e9f85SHerbert Xu break; 1751b68e9f85SHerbert Xu 1752c1b4a7e6SDavid S. Miller if (tso_segs == 1) { 1753aa93466bSDavid S. Miller if (unlikely(!tcp_nagle_test(tp, skb, mss_now, 1754aa93466bSDavid S. Miller (tcp_skb_is_last(sk, skb) ? 1755aa93466bSDavid S. Miller nonagle : TCP_NAGLE_PUSH)))) 1756aa93466bSDavid S. Miller break; 1757c1b4a7e6SDavid S. Miller } else { 1758d5dd9175SIlpo Järvinen if (!push_one && tcp_tso_should_defer(sk, skb)) 1759aa93466bSDavid S. Miller break; 1760c1b4a7e6SDavid S. Miller } 1761aa93466bSDavid S. Miller 1762c8ac3774SHerbert Xu limit = mss_now; 1763f8269a49SIlpo Järvinen if (tso_segs > 1 && !tcp_urg_mode(tp)) 17640e3a4803SIlpo Järvinen limit = tcp_mss_split_point(sk, skb, mss_now, 17650e3a4803SIlpo Järvinen cwnd_quota); 1766c8ac3774SHerbert Xu 1767c8ac3774SHerbert Xu if (skb->len > limit && 1768c4ead4c5SEric Dumazet unlikely(tso_fragment(sk, skb, limit, mss_now, gfp))) 17691da177e4SLinus Torvalds break; 17701da177e4SLinus Torvalds 17711da177e4SLinus Torvalds TCP_SKB_CB(skb)->when = tcp_time_stamp; 1772c1b4a7e6SDavid S. Miller 1773d5dd9175SIlpo Järvinen if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp))) 17741da177e4SLinus Torvalds break; 17751da177e4SLinus Torvalds 17761da177e4SLinus Torvalds /* Advance the send_head. This one is sent out. 17771da177e4SLinus Torvalds * This call will increment packets_out. 17781da177e4SLinus Torvalds */ 177966f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, skb); 17801da177e4SLinus Torvalds 17811da177e4SLinus Torvalds tcp_minshall_update(tp, mss_now, skb); 1782aa93466bSDavid S. Miller sent_pkts++; 1783d5dd9175SIlpo Järvinen 1784d5dd9175SIlpo Järvinen if (push_one) 1785d5dd9175SIlpo Järvinen break; 17861da177e4SLinus Torvalds } 17871da177e4SLinus Torvalds 1788aa93466bSDavid S. Miller if (likely(sent_pkts)) { 17899e412ba7SIlpo Järvinen tcp_cwnd_validate(sk); 17901da177e4SLinus Torvalds return 0; 17911da177e4SLinus Torvalds } 1792fe067e8aSDavid S. Miller return !tp->packets_out && tcp_send_head(sk); 17931da177e4SLinus Torvalds } 17941da177e4SLinus Torvalds 1795a762a980SDavid S. Miller /* Push out any pending frames which were held back due to 1796a762a980SDavid S. Miller * TCP_CORK or attempt at coalescing tiny packets. 1797a762a980SDavid S. Miller * The socket must be locked by the caller. 1798a762a980SDavid S. Miller */ 17999e412ba7SIlpo Järvinen void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss, 18009e412ba7SIlpo Järvinen int nonagle) 1801a762a980SDavid S. Miller { 1802726e07a8SIlpo Järvinen /* If we are closed, the bytes will have to remain here. 1803726e07a8SIlpo Järvinen * In time closedown will finish, we empty the write queue and 1804726e07a8SIlpo Järvinen * all will be happy. 1805726e07a8SIlpo Järvinen */ 1806726e07a8SIlpo Järvinen if (unlikely(sk->sk_state == TCP_CLOSE)) 1807726e07a8SIlpo Järvinen return; 1808726e07a8SIlpo Järvinen 1809d5dd9175SIlpo Järvinen if (tcp_write_xmit(sk, cur_mss, nonagle, 0, GFP_ATOMIC)) 18109e412ba7SIlpo Järvinen tcp_check_probe_timer(sk); 1811a762a980SDavid S. Miller } 1812a762a980SDavid S. Miller 1813c1b4a7e6SDavid S. Miller /* Send _single_ skb sitting at the send head. This function requires 1814c1b4a7e6SDavid S. Miller * true push pending frames to setup probe timer etc. 1815c1b4a7e6SDavid S. Miller */ 1816c1b4a7e6SDavid S. Miller void tcp_push_one(struct sock *sk, unsigned int mss_now) 1817c1b4a7e6SDavid S. Miller { 1818fe067e8aSDavid S. Miller struct sk_buff *skb = tcp_send_head(sk); 1819c1b4a7e6SDavid S. Miller 1820c1b4a7e6SDavid S. Miller BUG_ON(!skb || skb->len < mss_now); 1821c1b4a7e6SDavid S. Miller 1822d5dd9175SIlpo Järvinen tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation); 1823c1b4a7e6SDavid S. Miller } 1824c1b4a7e6SDavid S. Miller 18251da177e4SLinus Torvalds /* This function returns the amount that we can raise the 18261da177e4SLinus Torvalds * usable window based on the following constraints 18271da177e4SLinus Torvalds * 18281da177e4SLinus Torvalds * 1. The window can never be shrunk once it is offered (RFC 793) 18291da177e4SLinus Torvalds * 2. We limit memory per socket 18301da177e4SLinus Torvalds * 18311da177e4SLinus Torvalds * RFC 1122: 18321da177e4SLinus Torvalds * "the suggested [SWS] avoidance algorithm for the receiver is to keep 18331da177e4SLinus Torvalds * RECV.NEXT + RCV.WIN fixed until: 18341da177e4SLinus Torvalds * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)" 18351da177e4SLinus Torvalds * 18361da177e4SLinus Torvalds * i.e. don't raise the right edge of the window until you can raise 18371da177e4SLinus Torvalds * it at least MSS bytes. 18381da177e4SLinus Torvalds * 18391da177e4SLinus Torvalds * Unfortunately, the recommended algorithm breaks header prediction, 18401da177e4SLinus Torvalds * since header prediction assumes th->window stays fixed. 18411da177e4SLinus Torvalds * 18421da177e4SLinus Torvalds * Strictly speaking, keeping th->window fixed violates the receiver 18431da177e4SLinus Torvalds * side SWS prevention criteria. The problem is that under this rule 18441da177e4SLinus Torvalds * a stream of single byte packets will cause the right side of the 18451da177e4SLinus Torvalds * window to always advance by a single byte. 18461da177e4SLinus Torvalds * 18471da177e4SLinus Torvalds * Of course, if the sender implements sender side SWS prevention 18481da177e4SLinus Torvalds * then this will not be a problem. 18491da177e4SLinus Torvalds * 18501da177e4SLinus Torvalds * BSD seems to make the following compromise: 18511da177e4SLinus Torvalds * 18521da177e4SLinus Torvalds * If the free space is less than the 1/4 of the maximum 18531da177e4SLinus Torvalds * space available and the free space is less than 1/2 mss, 18541da177e4SLinus Torvalds * then set the window to 0. 18551da177e4SLinus Torvalds * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ] 18561da177e4SLinus Torvalds * Otherwise, just prevent the window from shrinking 18571da177e4SLinus Torvalds * and from being larger than the largest representable value. 18581da177e4SLinus Torvalds * 18591da177e4SLinus Torvalds * This prevents incremental opening of the window in the regime 18601da177e4SLinus Torvalds * where TCP is limited by the speed of the reader side taking 18611da177e4SLinus Torvalds * data out of the TCP receive queue. It does nothing about 18621da177e4SLinus Torvalds * those cases where the window is constrained on the sender side 18631da177e4SLinus Torvalds * because the pipeline is full. 18641da177e4SLinus Torvalds * 18651da177e4SLinus Torvalds * BSD also seems to "accidentally" limit itself to windows that are a 18661da177e4SLinus Torvalds * multiple of MSS, at least until the free space gets quite small. 18671da177e4SLinus Torvalds * This would appear to be a side effect of the mbuf implementation. 18681da177e4SLinus Torvalds * Combining these two algorithms results in the observed behavior 18691da177e4SLinus Torvalds * of having a fixed window size at almost all times. 18701da177e4SLinus Torvalds * 18711da177e4SLinus Torvalds * Below we obtain similar behavior by forcing the offered window to 18721da177e4SLinus Torvalds * a multiple of the mss when it is feasible to do so. 18731da177e4SLinus Torvalds * 18741da177e4SLinus Torvalds * Note, we don't "adjust" for TIMESTAMP or SACK option bytes. 18751da177e4SLinus Torvalds * Regular options like TIMESTAMP are taken into account. 18761da177e4SLinus Torvalds */ 18771da177e4SLinus Torvalds u32 __tcp_select_window(struct sock *sk) 18781da177e4SLinus Torvalds { 1879463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 18801da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 1881caa20d9aSStephen Hemminger /* MSS for the peer's data. Previous versions used mss_clamp 18821da177e4SLinus Torvalds * here. I don't know if the value based on our guesses 18831da177e4SLinus Torvalds * of peer's MSS is better for the performance. It's more correct 18841da177e4SLinus Torvalds * but may be worse for the performance because of rcv_mss 18851da177e4SLinus Torvalds * fluctuations. --SAW 1998/11/1 18861da177e4SLinus Torvalds */ 1887463c84b9SArnaldo Carvalho de Melo int mss = icsk->icsk_ack.rcv_mss; 18881da177e4SLinus Torvalds int free_space = tcp_space(sk); 18891da177e4SLinus Torvalds int full_space = min_t(int, tp->window_clamp, tcp_full_space(sk)); 18901da177e4SLinus Torvalds int window; 18911da177e4SLinus Torvalds 18921da177e4SLinus Torvalds if (mss > full_space) 18931da177e4SLinus Torvalds mss = full_space; 18941da177e4SLinus Torvalds 1895b92edbe0SEric Dumazet if (free_space < (full_space >> 1)) { 1896463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.quick = 0; 18971da177e4SLinus Torvalds 18981da177e4SLinus Torvalds if (tcp_memory_pressure) 1899056834d9SIlpo Järvinen tp->rcv_ssthresh = min(tp->rcv_ssthresh, 1900056834d9SIlpo Järvinen 4U * tp->advmss); 19011da177e4SLinus Torvalds 19021da177e4SLinus Torvalds if (free_space < mss) 19031da177e4SLinus Torvalds return 0; 19041da177e4SLinus Torvalds } 19051da177e4SLinus Torvalds 19061da177e4SLinus Torvalds if (free_space > tp->rcv_ssthresh) 19071da177e4SLinus Torvalds free_space = tp->rcv_ssthresh; 19081da177e4SLinus Torvalds 19091da177e4SLinus Torvalds /* Don't do rounding if we are using window scaling, since the 19101da177e4SLinus Torvalds * scaled window will not line up with the MSS boundary anyway. 19111da177e4SLinus Torvalds */ 19121da177e4SLinus Torvalds window = tp->rcv_wnd; 19131da177e4SLinus Torvalds if (tp->rx_opt.rcv_wscale) { 19141da177e4SLinus Torvalds window = free_space; 19151da177e4SLinus Torvalds 19161da177e4SLinus Torvalds /* Advertise enough space so that it won't get scaled away. 19171da177e4SLinus Torvalds * Import case: prevent zero window announcement if 19181da177e4SLinus Torvalds * 1<<rcv_wscale > mss. 19191da177e4SLinus Torvalds */ 19201da177e4SLinus Torvalds if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window) 19211da177e4SLinus Torvalds window = (((window >> tp->rx_opt.rcv_wscale) + 1) 19221da177e4SLinus Torvalds << tp->rx_opt.rcv_wscale); 19231da177e4SLinus Torvalds } else { 19241da177e4SLinus Torvalds /* Get the largest window that is a nice multiple of mss. 19251da177e4SLinus Torvalds * Window clamp already applied above. 19261da177e4SLinus Torvalds * If our current window offering is within 1 mss of the 19271da177e4SLinus Torvalds * free space we just keep it. This prevents the divide 19281da177e4SLinus Torvalds * and multiply from happening most of the time. 19291da177e4SLinus Torvalds * We also don't do any window rounding when the free space 19301da177e4SLinus Torvalds * is too small. 19311da177e4SLinus Torvalds */ 19321da177e4SLinus Torvalds if (window <= free_space - mss || window > free_space) 19331da177e4SLinus Torvalds window = (free_space / mss) * mss; 193484565070SJohn Heffner else if (mss == full_space && 1935b92edbe0SEric Dumazet free_space > window + (full_space >> 1)) 193684565070SJohn Heffner window = free_space; 19371da177e4SLinus Torvalds } 19381da177e4SLinus Torvalds 19391da177e4SLinus Torvalds return window; 19401da177e4SLinus Torvalds } 19411da177e4SLinus Torvalds 19424a17fc3aSIlpo Järvinen /* Collapses two adjacent SKB's during retransmission. */ 19434a17fc3aSIlpo Järvinen static void tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb) 19441da177e4SLinus Torvalds { 19451da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 1946fe067e8aSDavid S. Miller struct sk_buff *next_skb = tcp_write_queue_next(sk, skb); 1947058dc334SIlpo Järvinen int skb_size, next_skb_size; 19481da177e4SLinus Torvalds 1949058dc334SIlpo Järvinen skb_size = skb->len; 1950058dc334SIlpo Järvinen next_skb_size = next_skb->len; 19511da177e4SLinus Torvalds 1952058dc334SIlpo Järvinen BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1); 19531da177e4SLinus Torvalds 19546859d494SIlpo Järvinen tcp_highest_sack_combine(sk, next_skb, skb); 1955a6963a6bSIlpo Järvinen 1956fe067e8aSDavid S. Miller tcp_unlink_write_queue(next_skb, sk); 19571da177e4SLinus Torvalds 1958058dc334SIlpo Järvinen skb_copy_from_linear_data(next_skb, skb_put(skb, next_skb_size), 19591a4e2d09SArnaldo Carvalho de Melo next_skb_size); 19601da177e4SLinus Torvalds 196152d570aaSJarek Poplawski if (next_skb->ip_summed == CHECKSUM_PARTIAL) 196252d570aaSJarek Poplawski skb->ip_summed = CHECKSUM_PARTIAL; 19631da177e4SLinus Torvalds 196484fa7933SPatrick McHardy if (skb->ip_summed != CHECKSUM_PARTIAL) 19651da177e4SLinus Torvalds skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size); 19661da177e4SLinus Torvalds 19671da177e4SLinus Torvalds /* Update sequence range on original skb. */ 19681da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq; 19691da177e4SLinus Torvalds 1970e6c7d085SIlpo Järvinen /* Merge over control information. This moves PSH/FIN etc. over */ 1971e6c7d085SIlpo Järvinen TCP_SKB_CB(skb)->flags |= TCP_SKB_CB(next_skb)->flags; 19721da177e4SLinus Torvalds 19731da177e4SLinus Torvalds /* All done, get rid of second SKB and account for it so 19741da177e4SLinus Torvalds * packet counting does not break. 19751da177e4SLinus Torvalds */ 19764828e7f4SIlpo Järvinen TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS; 1977b7689205SIlpo Järvinen 1978b7689205SIlpo Järvinen /* changed transmit queue under us so clear hints */ 1979ef9da47cSIlpo Järvinen tcp_clear_retrans_hints_partial(tp); 1980ef9da47cSIlpo Järvinen if (next_skb == tp->retransmit_skb_hint) 1981ef9da47cSIlpo Järvinen tp->retransmit_skb_hint = skb; 1982b7689205SIlpo Järvinen 1983797108d1SIlpo Järvinen tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb)); 1984797108d1SIlpo Järvinen 19853ab224beSHideo Aoki sk_wmem_free_skb(sk, next_skb); 19861da177e4SLinus Torvalds } 19871da177e4SLinus Torvalds 198867edfef7SAndi Kleen /* Check if coalescing SKBs is legal. */ 19894a17fc3aSIlpo Järvinen static int tcp_can_collapse(struct sock *sk, struct sk_buff *skb) 19904a17fc3aSIlpo Järvinen { 19914a17fc3aSIlpo Järvinen if (tcp_skb_pcount(skb) > 1) 19924a17fc3aSIlpo Järvinen return 0; 19934a17fc3aSIlpo Järvinen /* TODO: SACK collapsing could be used to remove this condition */ 19944a17fc3aSIlpo Järvinen if (skb_shinfo(skb)->nr_frags != 0) 19954a17fc3aSIlpo Järvinen return 0; 19964a17fc3aSIlpo Järvinen if (skb_cloned(skb)) 19974a17fc3aSIlpo Järvinen return 0; 19984a17fc3aSIlpo Järvinen if (skb == tcp_send_head(sk)) 19994a17fc3aSIlpo Järvinen return 0; 20004a17fc3aSIlpo Järvinen /* Some heurestics for collapsing over SACK'd could be invented */ 20014a17fc3aSIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 20024a17fc3aSIlpo Järvinen return 0; 20034a17fc3aSIlpo Järvinen 20044a17fc3aSIlpo Järvinen return 1; 20054a17fc3aSIlpo Järvinen } 20064a17fc3aSIlpo Järvinen 200767edfef7SAndi Kleen /* Collapse packets in the retransmit queue to make to create 200867edfef7SAndi Kleen * less packets on the wire. This is only done on retransmission. 200967edfef7SAndi Kleen */ 20104a17fc3aSIlpo Järvinen static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to, 20114a17fc3aSIlpo Järvinen int space) 20124a17fc3aSIlpo Järvinen { 20134a17fc3aSIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 20144a17fc3aSIlpo Järvinen struct sk_buff *skb = to, *tmp; 20154a17fc3aSIlpo Järvinen int first = 1; 20164a17fc3aSIlpo Järvinen 20174a17fc3aSIlpo Järvinen if (!sysctl_tcp_retrans_collapse) 20184a17fc3aSIlpo Järvinen return; 2019a3433f35SChangli Gao if (TCP_SKB_CB(skb)->flags & TCPHDR_SYN) 20204a17fc3aSIlpo Järvinen return; 20214a17fc3aSIlpo Järvinen 20224a17fc3aSIlpo Järvinen tcp_for_write_queue_from_safe(skb, tmp, sk) { 20234a17fc3aSIlpo Järvinen if (!tcp_can_collapse(sk, skb)) 20244a17fc3aSIlpo Järvinen break; 20254a17fc3aSIlpo Järvinen 20264a17fc3aSIlpo Järvinen space -= skb->len; 20274a17fc3aSIlpo Järvinen 20284a17fc3aSIlpo Järvinen if (first) { 20294a17fc3aSIlpo Järvinen first = 0; 20304a17fc3aSIlpo Järvinen continue; 20314a17fc3aSIlpo Järvinen } 20324a17fc3aSIlpo Järvinen 20334a17fc3aSIlpo Järvinen if (space < 0) 20344a17fc3aSIlpo Järvinen break; 20354a17fc3aSIlpo Järvinen /* Punt if not enough space exists in the first SKB for 20364a17fc3aSIlpo Järvinen * the data in the second 20374a17fc3aSIlpo Järvinen */ 20384a17fc3aSIlpo Järvinen if (skb->len > skb_tailroom(to)) 20394a17fc3aSIlpo Järvinen break; 20404a17fc3aSIlpo Järvinen 20414a17fc3aSIlpo Järvinen if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp))) 20424a17fc3aSIlpo Järvinen break; 20434a17fc3aSIlpo Järvinen 20444a17fc3aSIlpo Järvinen tcp_collapse_retrans(sk, to); 20454a17fc3aSIlpo Järvinen } 20464a17fc3aSIlpo Järvinen } 20474a17fc3aSIlpo Järvinen 20481da177e4SLinus Torvalds /* This retransmits one SKB. Policy decisions and retransmit queue 20491da177e4SLinus Torvalds * state updates are done by the caller. Returns non-zero if an 20501da177e4SLinus Torvalds * error occurred which prevented the send. 20511da177e4SLinus Torvalds */ 20521da177e4SLinus Torvalds int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb) 20531da177e4SLinus Torvalds { 20541da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 20555d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 20567d227cd2SSridhar Samudrala unsigned int cur_mss; 20571da177e4SLinus Torvalds int err; 20581da177e4SLinus Torvalds 20595d424d5aSJohn Heffner /* Inconslusive MTU probe */ 20605d424d5aSJohn Heffner if (icsk->icsk_mtup.probe_size) { 20615d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = 0; 20625d424d5aSJohn Heffner } 20635d424d5aSJohn Heffner 20641da177e4SLinus Torvalds /* Do not sent more than we queued. 1/4 is reserved for possible 2065caa20d9aSStephen Hemminger * copying overhead: fragmentation, tunneling, mangling etc. 20661da177e4SLinus Torvalds */ 20671da177e4SLinus Torvalds if (atomic_read(&sk->sk_wmem_alloc) > 20681da177e4SLinus Torvalds min(sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), sk->sk_sndbuf)) 20691da177e4SLinus Torvalds return -EAGAIN; 20701da177e4SLinus Torvalds 20711da177e4SLinus Torvalds if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) { 20721da177e4SLinus Torvalds if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) 20731da177e4SLinus Torvalds BUG(); 20741da177e4SLinus Torvalds if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq)) 20751da177e4SLinus Torvalds return -ENOMEM; 20761da177e4SLinus Torvalds } 20771da177e4SLinus Torvalds 20787d227cd2SSridhar Samudrala if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk)) 20797d227cd2SSridhar Samudrala return -EHOSTUNREACH; /* Routing failure or similar. */ 20807d227cd2SSridhar Samudrala 20810c54b85fSIlpo Järvinen cur_mss = tcp_current_mss(sk); 20827d227cd2SSridhar Samudrala 20831da177e4SLinus Torvalds /* If receiver has shrunk his window, and skb is out of 20841da177e4SLinus Torvalds * new window, do not retransmit it. The exception is the 20851da177e4SLinus Torvalds * case, when window is shrunk to zero. In this case 20861da177e4SLinus Torvalds * our retransmit serves as a zero window probe. 20871da177e4SLinus Torvalds */ 20889d4fb27dSJoe Perches if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) && 20899d4fb27dSJoe Perches TCP_SKB_CB(skb)->seq != tp->snd_una) 20901da177e4SLinus Torvalds return -EAGAIN; 20911da177e4SLinus Torvalds 20921da177e4SLinus Torvalds if (skb->len > cur_mss) { 2093846998aeSDavid S. Miller if (tcp_fragment(sk, skb, cur_mss, cur_mss)) 20941da177e4SLinus Torvalds return -ENOMEM; /* We'll try again later. */ 209502276f3cSIlpo Järvinen } else { 20969eb9362eSIlpo Järvinen int oldpcount = tcp_skb_pcount(skb); 20979eb9362eSIlpo Järvinen 20989eb9362eSIlpo Järvinen if (unlikely(oldpcount > 1)) { 209902276f3cSIlpo Järvinen tcp_init_tso_segs(sk, skb, cur_mss); 21009eb9362eSIlpo Järvinen tcp_adjust_pcount(sk, skb, oldpcount - tcp_skb_pcount(skb)); 21019eb9362eSIlpo Järvinen } 21021da177e4SLinus Torvalds } 21031da177e4SLinus Torvalds 21041da177e4SLinus Torvalds tcp_retrans_try_collapse(sk, skb, cur_mss); 21051da177e4SLinus Torvalds 21061da177e4SLinus Torvalds /* Some Solaris stacks overoptimize and ignore the FIN on a 21071da177e4SLinus Torvalds * retransmit when old data is attached. So strip it off 21081da177e4SLinus Torvalds * since it is cheap to do so and saves bytes on the network. 21091da177e4SLinus Torvalds */ 21101da177e4SLinus Torvalds if (skb->len > 0 && 2111a3433f35SChangli Gao (TCP_SKB_CB(skb)->flags & TCPHDR_FIN) && 21121da177e4SLinus Torvalds tp->snd_una == (TCP_SKB_CB(skb)->end_seq - 1)) { 21131da177e4SLinus Torvalds if (!pskb_trim(skb, 0)) { 2114e870a8efSIlpo Järvinen /* Reuse, even though it does some unnecessary work */ 2115e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, TCP_SKB_CB(skb)->end_seq - 1, 2116e870a8efSIlpo Järvinen TCP_SKB_CB(skb)->flags); 21171da177e4SLinus Torvalds skb->ip_summed = CHECKSUM_NONE; 21181da177e4SLinus Torvalds } 21191da177e4SLinus Torvalds } 21201da177e4SLinus Torvalds 21211da177e4SLinus Torvalds /* Make a copy, if the first transmission SKB clone we made 21221da177e4SLinus Torvalds * is still in somebody's hands, else make a clone. 21231da177e4SLinus Torvalds */ 21241da177e4SLinus Torvalds TCP_SKB_CB(skb)->when = tcp_time_stamp; 21251da177e4SLinus Torvalds 2126dfb4b9dcSDavid S. Miller err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 21271da177e4SLinus Torvalds 21281da177e4SLinus Torvalds if (err == 0) { 21291da177e4SLinus Torvalds /* Update global TCP statistics. */ 213081cc8a75SPavel Emelyanov TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS); 21311da177e4SLinus Torvalds 21321da177e4SLinus Torvalds tp->total_retrans++; 21331da177e4SLinus Torvalds 21341da177e4SLinus Torvalds #if FASTRETRANS_DEBUG > 0 21351da177e4SLinus Torvalds if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) { 21361da177e4SLinus Torvalds if (net_ratelimit()) 21371da177e4SLinus Torvalds printk(KERN_DEBUG "retrans_out leaked.\n"); 21381da177e4SLinus Torvalds } 21391da177e4SLinus Torvalds #endif 2140b08d6cb2SIlpo Järvinen if (!tp->retrans_out) 2141b08d6cb2SIlpo Järvinen tp->lost_retrans_low = tp->snd_nxt; 21421da177e4SLinus Torvalds TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS; 21431da177e4SLinus Torvalds tp->retrans_out += tcp_skb_pcount(skb); 21441da177e4SLinus Torvalds 21451da177e4SLinus Torvalds /* Save stamp of the first retransmit. */ 21461da177e4SLinus Torvalds if (!tp->retrans_stamp) 21471da177e4SLinus Torvalds tp->retrans_stamp = TCP_SKB_CB(skb)->when; 21481da177e4SLinus Torvalds 21491da177e4SLinus Torvalds tp->undo_retrans++; 21501da177e4SLinus Torvalds 21511da177e4SLinus Torvalds /* snd_nxt is stored to detect loss of retransmitted segment, 21521da177e4SLinus Torvalds * see tcp_input.c tcp_sacktag_write_queue(). 21531da177e4SLinus Torvalds */ 21541da177e4SLinus Torvalds TCP_SKB_CB(skb)->ack_seq = tp->snd_nxt; 21551da177e4SLinus Torvalds } 21561da177e4SLinus Torvalds return err; 21571da177e4SLinus Torvalds } 21581da177e4SLinus Torvalds 215967edfef7SAndi Kleen /* Check if we forward retransmits are possible in the current 216067edfef7SAndi Kleen * window/congestion state. 216167edfef7SAndi Kleen */ 2162b5afe7bcSIlpo Järvinen static int tcp_can_forward_retransmit(struct sock *sk) 2163b5afe7bcSIlpo Järvinen { 2164b5afe7bcSIlpo Järvinen const struct inet_connection_sock *icsk = inet_csk(sk); 2165b5afe7bcSIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 2166b5afe7bcSIlpo Järvinen 2167b5afe7bcSIlpo Järvinen /* Forward retransmissions are possible only during Recovery. */ 2168b5afe7bcSIlpo Järvinen if (icsk->icsk_ca_state != TCP_CA_Recovery) 2169b5afe7bcSIlpo Järvinen return 0; 2170b5afe7bcSIlpo Järvinen 2171b5afe7bcSIlpo Järvinen /* No forward retransmissions in Reno are possible. */ 2172b5afe7bcSIlpo Järvinen if (tcp_is_reno(tp)) 2173b5afe7bcSIlpo Järvinen return 0; 2174b5afe7bcSIlpo Järvinen 2175b5afe7bcSIlpo Järvinen /* Yeah, we have to make difficult choice between forward transmission 2176b5afe7bcSIlpo Järvinen * and retransmission... Both ways have their merits... 2177b5afe7bcSIlpo Järvinen * 2178b5afe7bcSIlpo Järvinen * For now we do not retransmit anything, while we have some new 2179b5afe7bcSIlpo Järvinen * segments to send. In the other cases, follow rule 3 for 2180b5afe7bcSIlpo Järvinen * NextSeg() specified in RFC3517. 2181b5afe7bcSIlpo Järvinen */ 2182b5afe7bcSIlpo Järvinen 2183b5afe7bcSIlpo Järvinen if (tcp_may_send_now(sk)) 2184b5afe7bcSIlpo Järvinen return 0; 2185b5afe7bcSIlpo Järvinen 2186b5afe7bcSIlpo Järvinen return 1; 2187b5afe7bcSIlpo Järvinen } 2188b5afe7bcSIlpo Järvinen 21891da177e4SLinus Torvalds /* This gets called after a retransmit timeout, and the initially 21901da177e4SLinus Torvalds * retransmitted data is acknowledged. It tries to continue 21911da177e4SLinus Torvalds * resending the rest of the retransmit queue, until either 21921da177e4SLinus Torvalds * we've sent it all or the congestion window limit is reached. 21931da177e4SLinus Torvalds * If doing SACK, the first ACK which comes back for a timeout 21941da177e4SLinus Torvalds * based retransmit packet might feed us FACK information again. 21951da177e4SLinus Torvalds * If so, we use it to avoid unnecessarily retransmissions. 21961da177e4SLinus Torvalds */ 21971da177e4SLinus Torvalds void tcp_xmit_retransmit_queue(struct sock *sk) 21981da177e4SLinus Torvalds { 21996687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 22001da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 22011da177e4SLinus Torvalds struct sk_buff *skb; 22020e1c54c2SIlpo Järvinen struct sk_buff *hole = NULL; 2203618d9f25SIlpo Järvinen u32 last_lost; 220461eb55f4SIlpo Järvinen int mib_idx; 22050e1c54c2SIlpo Järvinen int fwd_rexmitting = 0; 22066a438bbeSStephen Hemminger 220745e77d31SIlpo Järvinen if (!tp->packets_out) 220845e77d31SIlpo Järvinen return; 220945e77d31SIlpo Järvinen 221008ebd172SIlpo Järvinen if (!tp->lost_out) 221108ebd172SIlpo Järvinen tp->retransmit_high = tp->snd_una; 221208ebd172SIlpo Järvinen 2213618d9f25SIlpo Järvinen if (tp->retransmit_skb_hint) { 22146a438bbeSStephen Hemminger skb = tp->retransmit_skb_hint; 2215618d9f25SIlpo Järvinen last_lost = TCP_SKB_CB(skb)->end_seq; 2216618d9f25SIlpo Järvinen if (after(last_lost, tp->retransmit_high)) 2217618d9f25SIlpo Järvinen last_lost = tp->retransmit_high; 2218618d9f25SIlpo Järvinen } else { 2219fe067e8aSDavid S. Miller skb = tcp_write_queue_head(sk); 2220618d9f25SIlpo Järvinen last_lost = tp->snd_una; 2221618d9f25SIlpo Järvinen } 22221da177e4SLinus Torvalds 2223fe067e8aSDavid S. Miller tcp_for_write_queue_from(skb, sk) { 22241da177e4SLinus Torvalds __u8 sacked = TCP_SKB_CB(skb)->sacked; 22251da177e4SLinus Torvalds 2226fe067e8aSDavid S. Miller if (skb == tcp_send_head(sk)) 2227fe067e8aSDavid S. Miller break; 22286a438bbeSStephen Hemminger /* we could do better than to assign each time */ 22290e1c54c2SIlpo Järvinen if (hole == NULL) 22306a438bbeSStephen Hemminger tp->retransmit_skb_hint = skb; 22316a438bbeSStephen Hemminger 22321da177e4SLinus Torvalds /* Assume this retransmit will generate 22331da177e4SLinus Torvalds * only one packet for congestion window 22341da177e4SLinus Torvalds * calculation purposes. This works because 22351da177e4SLinus Torvalds * tcp_retransmit_skb() will chop up the 22361da177e4SLinus Torvalds * packet to be MSS sized and all the 22371da177e4SLinus Torvalds * packet counting works out. 22381da177e4SLinus Torvalds */ 22391da177e4SLinus Torvalds if (tcp_packets_in_flight(tp) >= tp->snd_cwnd) 22401da177e4SLinus Torvalds return; 22410e1c54c2SIlpo Järvinen 22420e1c54c2SIlpo Järvinen if (fwd_rexmitting) { 22430e1c54c2SIlpo Järvinen begin_fwd: 22440e1c54c2SIlpo Järvinen if (!before(TCP_SKB_CB(skb)->seq, tcp_highest_sack_seq(tp))) 2245006f582cSIlpo Järvinen break; 22460e1c54c2SIlpo Järvinen mib_idx = LINUX_MIB_TCPFORWARDRETRANS; 22470e1c54c2SIlpo Järvinen 22480e1c54c2SIlpo Järvinen } else if (!before(TCP_SKB_CB(skb)->seq, tp->retransmit_high)) { 2249618d9f25SIlpo Järvinen tp->retransmit_high = last_lost; 22500e1c54c2SIlpo Järvinen if (!tcp_can_forward_retransmit(sk)) 22510e1c54c2SIlpo Järvinen break; 22520e1c54c2SIlpo Järvinen /* Backtrack if necessary to non-L'ed skb */ 22530e1c54c2SIlpo Järvinen if (hole != NULL) { 22540e1c54c2SIlpo Järvinen skb = hole; 22550e1c54c2SIlpo Järvinen hole = NULL; 22560e1c54c2SIlpo Järvinen } 22570e1c54c2SIlpo Järvinen fwd_rexmitting = 1; 22580e1c54c2SIlpo Järvinen goto begin_fwd; 22590e1c54c2SIlpo Järvinen 22600e1c54c2SIlpo Järvinen } else if (!(sacked & TCPCB_LOST)) { 2261ac11ba75SIlpo Järvinen if (hole == NULL && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED))) 22620e1c54c2SIlpo Järvinen hole = skb; 226361eb55f4SIlpo Järvinen continue; 22641da177e4SLinus Torvalds 22650e1c54c2SIlpo Järvinen } else { 2266618d9f25SIlpo Järvinen last_lost = TCP_SKB_CB(skb)->end_seq; 22670e1c54c2SIlpo Järvinen if (icsk->icsk_ca_state != TCP_CA_Loss) 22680e1c54c2SIlpo Järvinen mib_idx = LINUX_MIB_TCPFASTRETRANS; 22690e1c54c2SIlpo Järvinen else 22700e1c54c2SIlpo Järvinen mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS; 22710e1c54c2SIlpo Järvinen } 22720e1c54c2SIlpo Järvinen 22730e1c54c2SIlpo Järvinen if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS)) 227461eb55f4SIlpo Järvinen continue; 227540b215e5SPavel Emelyanov 2276f0ceb0edSIlpo Järvinen if (tcp_retransmit_skb(sk, skb)) 22771da177e4SLinus Torvalds return; 2278de0744afSPavel Emelyanov NET_INC_STATS_BH(sock_net(sk), mib_idx); 22791da177e4SLinus Torvalds 2280fe067e8aSDavid S. Miller if (skb == tcp_write_queue_head(sk)) 2281463c84b9SArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 22823f421baaSArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto, 22833f421baaSArnaldo Carvalho de Melo TCP_RTO_MAX); 22841da177e4SLinus Torvalds } 22851da177e4SLinus Torvalds } 22861da177e4SLinus Torvalds 22871da177e4SLinus Torvalds /* Send a fin. The caller locks the socket for us. This cannot be 22881da177e4SLinus Torvalds * allowed to fail queueing a FIN frame under any circumstances. 22891da177e4SLinus Torvalds */ 22901da177e4SLinus Torvalds void tcp_send_fin(struct sock *sk) 22911da177e4SLinus Torvalds { 22921da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 2293fe067e8aSDavid S. Miller struct sk_buff *skb = tcp_write_queue_tail(sk); 22941da177e4SLinus Torvalds int mss_now; 22951da177e4SLinus Torvalds 22961da177e4SLinus Torvalds /* Optimization, tack on the FIN if we have a queue of 22971da177e4SLinus Torvalds * unsent frames. But be careful about outgoing SACKS 22981da177e4SLinus Torvalds * and IP options. 22991da177e4SLinus Torvalds */ 23000c54b85fSIlpo Järvinen mss_now = tcp_current_mss(sk); 23011da177e4SLinus Torvalds 2302fe067e8aSDavid S. Miller if (tcp_send_head(sk) != NULL) { 2303a3433f35SChangli Gao TCP_SKB_CB(skb)->flags |= TCPHDR_FIN; 23041da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq++; 23051da177e4SLinus Torvalds tp->write_seq++; 23061da177e4SLinus Torvalds } else { 23071da177e4SLinus Torvalds /* Socket is locked, keep trying until memory is available. */ 23081da177e4SLinus Torvalds for (;;) { 2309aa133076SWu Fengguang skb = alloc_skb_fclone(MAX_TCP_HEADER, 2310aa133076SWu Fengguang sk->sk_allocation); 23111da177e4SLinus Torvalds if (skb) 23121da177e4SLinus Torvalds break; 23131da177e4SLinus Torvalds yield(); 23141da177e4SLinus Torvalds } 23151da177e4SLinus Torvalds 23161da177e4SLinus Torvalds /* Reserve space for headers and prepare control bits. */ 23171da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 23181da177e4SLinus Torvalds /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */ 2319e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tp->write_seq, 2320a3433f35SChangli Gao TCPHDR_ACK | TCPHDR_FIN); 23211da177e4SLinus Torvalds tcp_queue_skb(sk, skb); 23221da177e4SLinus Torvalds } 23239e412ba7SIlpo Järvinen __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_OFF); 23241da177e4SLinus Torvalds } 23251da177e4SLinus Torvalds 23261da177e4SLinus Torvalds /* We get here when a process closes a file descriptor (either due to 23271da177e4SLinus Torvalds * an explicit close() or as a byproduct of exit()'ing) and there 23281da177e4SLinus Torvalds * was unread data in the receive queue. This behavior is recommended 232965bb723cSGerrit Renker * by RFC 2525, section 2.17. -DaveM 23301da177e4SLinus Torvalds */ 2331dd0fc66fSAl Viro void tcp_send_active_reset(struct sock *sk, gfp_t priority) 23321da177e4SLinus Torvalds { 23331da177e4SLinus Torvalds struct sk_buff *skb; 23341da177e4SLinus Torvalds 23351da177e4SLinus Torvalds /* NOTE: No TCP options attached and we never retransmit this. */ 23361da177e4SLinus Torvalds skb = alloc_skb(MAX_TCP_HEADER, priority); 23371da177e4SLinus Torvalds if (!skb) { 23384e673444SPavel Emelyanov NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); 23391da177e4SLinus Torvalds return; 23401da177e4SLinus Torvalds } 23411da177e4SLinus Torvalds 23421da177e4SLinus Torvalds /* Reserve space for headers and prepare control bits. */ 23431da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 2344e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk), 2345a3433f35SChangli Gao TCPHDR_ACK | TCPHDR_RST); 23461da177e4SLinus Torvalds /* Send it off. */ 23471da177e4SLinus Torvalds TCP_SKB_CB(skb)->when = tcp_time_stamp; 2348dfb4b9dcSDavid S. Miller if (tcp_transmit_skb(sk, skb, 0, priority)) 23494e673444SPavel Emelyanov NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); 235026af65cbSSridhar Samudrala 235181cc8a75SPavel Emelyanov TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS); 23521da177e4SLinus Torvalds } 23531da177e4SLinus Torvalds 235467edfef7SAndi Kleen /* Send a crossed SYN-ACK during socket establishment. 235567edfef7SAndi Kleen * WARNING: This routine must only be called when we have already sent 23561da177e4SLinus Torvalds * a SYN packet that crossed the incoming SYN that caused this routine 23571da177e4SLinus Torvalds * to get called. If this assumption fails then the initial rcv_wnd 23581da177e4SLinus Torvalds * and rcv_wscale values will not be correct. 23591da177e4SLinus Torvalds */ 23601da177e4SLinus Torvalds int tcp_send_synack(struct sock *sk) 23611da177e4SLinus Torvalds { 23621da177e4SLinus Torvalds struct sk_buff *skb; 23631da177e4SLinus Torvalds 2364fe067e8aSDavid S. Miller skb = tcp_write_queue_head(sk); 2365a3433f35SChangli Gao if (skb == NULL || !(TCP_SKB_CB(skb)->flags & TCPHDR_SYN)) { 23661da177e4SLinus Torvalds printk(KERN_DEBUG "tcp_send_synack: wrong queue state\n"); 23671da177e4SLinus Torvalds return -EFAULT; 23681da177e4SLinus Torvalds } 2369a3433f35SChangli Gao if (!(TCP_SKB_CB(skb)->flags & TCPHDR_ACK)) { 23701da177e4SLinus Torvalds if (skb_cloned(skb)) { 23711da177e4SLinus Torvalds struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC); 23721da177e4SLinus Torvalds if (nskb == NULL) 23731da177e4SLinus Torvalds return -ENOMEM; 2374fe067e8aSDavid S. Miller tcp_unlink_write_queue(skb, sk); 23751da177e4SLinus Torvalds skb_header_release(nskb); 2376fe067e8aSDavid S. Miller __tcp_add_write_queue_head(sk, nskb); 23773ab224beSHideo Aoki sk_wmem_free_skb(sk, skb); 23783ab224beSHideo Aoki sk->sk_wmem_queued += nskb->truesize; 23793ab224beSHideo Aoki sk_mem_charge(sk, nskb->truesize); 23801da177e4SLinus Torvalds skb = nskb; 23811da177e4SLinus Torvalds } 23821da177e4SLinus Torvalds 2383a3433f35SChangli Gao TCP_SKB_CB(skb)->flags |= TCPHDR_ACK; 23841da177e4SLinus Torvalds TCP_ECN_send_synack(tcp_sk(sk), skb); 23851da177e4SLinus Torvalds } 23861da177e4SLinus Torvalds TCP_SKB_CB(skb)->when = tcp_time_stamp; 2387dfb4b9dcSDavid S. Miller return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 23881da177e4SLinus Torvalds } 23891da177e4SLinus Torvalds 239067edfef7SAndi Kleen /* Prepare a SYN-ACK. */ 23911da177e4SLinus Torvalds struct sk_buff *tcp_make_synack(struct sock *sk, struct dst_entry *dst, 2392e6b4d113SWilliam Allen Simpson struct request_sock *req, 2393e6b4d113SWilliam Allen Simpson struct request_values *rvp) 23941da177e4SLinus Torvalds { 2395bd0388aeSWilliam Allen Simpson struct tcp_out_options opts; 23964957faadSWilliam Allen Simpson struct tcp_extend_values *xvp = tcp_xv(rvp); 23972e6599cbSArnaldo Carvalho de Melo struct inet_request_sock *ireq = inet_rsk(req); 23981da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 239928b2774aSEric Dumazet const struct tcp_cookie_values *cvp = tp->cookie_values; 24001da177e4SLinus Torvalds struct tcphdr *th; 24011da177e4SLinus Torvalds struct sk_buff *skb; 2402cfb6eeb4SYOSHIFUJI Hideaki struct tcp_md5sig_key *md5; 2403bd0388aeSWilliam Allen Simpson int tcp_header_size; 2404f5fff5dcSTom Quetchenbach int mss; 240528b2774aSEric Dumazet int s_data_desired = 0; 24061da177e4SLinus Torvalds 240728b2774aSEric Dumazet if (cvp != NULL && cvp->s_data_constant && cvp->s_data_desired) 240828b2774aSEric Dumazet s_data_desired = cvp->s_data_desired; 240928b2774aSEric Dumazet skb = sock_wmalloc(sk, MAX_TCP_HEADER + 15 + s_data_desired, 1, GFP_ATOMIC); 24101da177e4SLinus Torvalds if (skb == NULL) 24111da177e4SLinus Torvalds return NULL; 24121da177e4SLinus Torvalds 24131da177e4SLinus Torvalds /* Reserve space for headers. */ 24141da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 24151da177e4SLinus Torvalds 2416adf30907SEric Dumazet skb_dst_set(skb, dst_clone(dst)); 24171da177e4SLinus Torvalds 2418f5fff5dcSTom Quetchenbach mss = dst_metric(dst, RTAX_ADVMSS); 2419f5fff5dcSTom Quetchenbach if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < mss) 2420f5fff5dcSTom Quetchenbach mss = tp->rx_opt.user_mss; 2421f5fff5dcSTom Quetchenbach 242233ad798cSAdam Langley if (req->rcv_wnd == 0) { /* ignored for retransmitted syns */ 242333ad798cSAdam Langley __u8 rcv_wscale; 242433ad798cSAdam Langley /* Set this up on the first call only */ 242533ad798cSAdam Langley req->window_clamp = tp->window_clamp ? : dst_metric(dst, RTAX_WINDOW); 2426e88c64f0SHagen Paul Pfeifer 2427e88c64f0SHagen Paul Pfeifer /* limit the window selection if the user enforce a smaller rx buffer */ 2428e88c64f0SHagen Paul Pfeifer if (sk->sk_userlocks & SOCK_RCVBUF_LOCK && 2429e88c64f0SHagen Paul Pfeifer (req->window_clamp > tcp_full_space(sk) || req->window_clamp == 0)) 2430e88c64f0SHagen Paul Pfeifer req->window_clamp = tcp_full_space(sk); 2431e88c64f0SHagen Paul Pfeifer 243233ad798cSAdam Langley /* tcp_full_space because it is guaranteed to be the first packet */ 243333ad798cSAdam Langley tcp_select_initial_window(tcp_full_space(sk), 2434f5fff5dcSTom Quetchenbach mss - (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0), 243533ad798cSAdam Langley &req->rcv_wnd, 243633ad798cSAdam Langley &req->window_clamp, 243733ad798cSAdam Langley ireq->wscale_ok, 243831d12926Slaurent chavey &rcv_wscale, 243931d12926Slaurent chavey dst_metric(dst, RTAX_INITRWND)); 244033ad798cSAdam Langley ireq->rcv_wscale = rcv_wscale; 244133ad798cSAdam Langley } 2442cfb6eeb4SYOSHIFUJI Hideaki 244333ad798cSAdam Langley memset(&opts, 0, sizeof(opts)); 24448b5f12d0SFlorian Westphal #ifdef CONFIG_SYN_COOKIES 24458b5f12d0SFlorian Westphal if (unlikely(req->cookie_ts)) 24468b5f12d0SFlorian Westphal TCP_SKB_CB(skb)->when = cookie_init_timestamp(req); 24478b5f12d0SFlorian Westphal else 24488b5f12d0SFlorian Westphal #endif 244933ad798cSAdam Langley TCP_SKB_CB(skb)->when = tcp_time_stamp; 2450f5fff5dcSTom Quetchenbach tcp_header_size = tcp_synack_options(sk, req, mss, 24514957faadSWilliam Allen Simpson skb, &opts, &md5, xvp) 24524957faadSWilliam Allen Simpson + sizeof(*th); 245333ad798cSAdam Langley 2454aa8223c7SArnaldo Carvalho de Melo skb_push(skb, tcp_header_size); 2455aa8223c7SArnaldo Carvalho de Melo skb_reset_transport_header(skb); 24561da177e4SLinus Torvalds 2457aa8223c7SArnaldo Carvalho de Melo th = tcp_hdr(skb); 24581da177e4SLinus Torvalds memset(th, 0, sizeof(struct tcphdr)); 24591da177e4SLinus Torvalds th->syn = 1; 24601da177e4SLinus Torvalds th->ack = 1; 24611da177e4SLinus Torvalds TCP_ECN_make_synack(req, th); 2462a3116ac5SKOVACS Krisztian th->source = ireq->loc_port; 24632e6599cbSArnaldo Carvalho de Melo th->dest = ireq->rmt_port; 2464e870a8efSIlpo Järvinen /* Setting of flags are superfluous here for callers (and ECE is 2465e870a8efSIlpo Järvinen * not even correctly set) 2466e870a8efSIlpo Järvinen */ 2467e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tcp_rsk(req)->snt_isn, 2468a3433f35SChangli Gao TCPHDR_SYN | TCPHDR_ACK); 24694957faadSWilliam Allen Simpson 24704957faadSWilliam Allen Simpson if (OPTION_COOKIE_EXTENSION & opts.options) { 247128b2774aSEric Dumazet if (s_data_desired) { 247228b2774aSEric Dumazet u8 *buf = skb_put(skb, s_data_desired); 24734957faadSWilliam Allen Simpson 24744957faadSWilliam Allen Simpson /* copy data directly from the listening socket. */ 247528b2774aSEric Dumazet memcpy(buf, cvp->s_data_payload, s_data_desired); 247628b2774aSEric Dumazet TCP_SKB_CB(skb)->end_seq += s_data_desired; 24774957faadSWilliam Allen Simpson } 24784957faadSWilliam Allen Simpson 24794957faadSWilliam Allen Simpson if (opts.hash_size > 0) { 24804957faadSWilliam Allen Simpson __u32 workspace[SHA_WORKSPACE_WORDS]; 24814957faadSWilliam Allen Simpson u32 *mess = &xvp->cookie_bakery[COOKIE_DIGEST_WORDS]; 24824957faadSWilliam Allen Simpson u32 *tail = &mess[COOKIE_MESSAGE_WORDS-1]; 24834957faadSWilliam Allen Simpson 24844957faadSWilliam Allen Simpson /* Secret recipe depends on the Timestamp, (future) 24854957faadSWilliam Allen Simpson * Sequence and Acknowledgment Numbers, Initiator 24864957faadSWilliam Allen Simpson * Cookie, and others handled by IP variant caller. 24874957faadSWilliam Allen Simpson */ 24884957faadSWilliam Allen Simpson *tail-- ^= opts.tsval; 24894957faadSWilliam Allen Simpson *tail-- ^= tcp_rsk(req)->rcv_isn + 1; 24904957faadSWilliam Allen Simpson *tail-- ^= TCP_SKB_CB(skb)->seq + 1; 24914957faadSWilliam Allen Simpson 24924957faadSWilliam Allen Simpson /* recommended */ 24930eae88f3SEric Dumazet *tail-- ^= (((__force u32)th->dest << 16) | (__force u32)th->source); 2494f9a2e69eSDavid S. Miller *tail-- ^= (u32)(unsigned long)cvp; /* per sockopt */ 24954957faadSWilliam Allen Simpson 24964957faadSWilliam Allen Simpson sha_transform((__u32 *)&xvp->cookie_bakery[0], 24974957faadSWilliam Allen Simpson (char *)mess, 24984957faadSWilliam Allen Simpson &workspace[0]); 24994957faadSWilliam Allen Simpson opts.hash_location = 25004957faadSWilliam Allen Simpson (__u8 *)&xvp->cookie_bakery[0]; 25014957faadSWilliam Allen Simpson } 25024957faadSWilliam Allen Simpson } 25034957faadSWilliam Allen Simpson 25041da177e4SLinus Torvalds th->seq = htonl(TCP_SKB_CB(skb)->seq); 25052e6599cbSArnaldo Carvalho de Melo th->ack_seq = htonl(tcp_rsk(req)->rcv_isn + 1); 25061da177e4SLinus Torvalds 25071da177e4SLinus Torvalds /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */ 2508600ff0c2SIlpo Järvinen th->window = htons(min(req->rcv_wnd, 65535U)); 2509bd0388aeSWilliam Allen Simpson tcp_options_write((__be32 *)(th + 1), tp, &opts); 25101da177e4SLinus Torvalds th->doff = (tcp_header_size >> 2); 2511aa2ea058STom Herbert TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS, tcp_skb_pcount(skb)); 2512cfb6eeb4SYOSHIFUJI Hideaki 2513cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 2514cfb6eeb4SYOSHIFUJI Hideaki /* Okay, we have all we need - do the md5 hash if needed */ 2515cfb6eeb4SYOSHIFUJI Hideaki if (md5) { 2516bd0388aeSWilliam Allen Simpson tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location, 251749a72dfbSAdam Langley md5, NULL, req, skb); 2518cfb6eeb4SYOSHIFUJI Hideaki } 2519cfb6eeb4SYOSHIFUJI Hideaki #endif 2520cfb6eeb4SYOSHIFUJI Hideaki 25211da177e4SLinus Torvalds return skb; 25221da177e4SLinus Torvalds } 25234bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_make_synack); 25241da177e4SLinus Torvalds 252567edfef7SAndi Kleen /* Do all connect socket setups that can be done AF independent. */ 252640efc6faSStephen Hemminger static void tcp_connect_init(struct sock *sk) 25271da177e4SLinus Torvalds { 25281da177e4SLinus Torvalds struct dst_entry *dst = __sk_dst_get(sk); 25291da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 25301da177e4SLinus Torvalds __u8 rcv_wscale; 25311da177e4SLinus Torvalds 25321da177e4SLinus Torvalds /* We'll fix this up when we get a response from the other end. 25331da177e4SLinus Torvalds * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT. 25341da177e4SLinus Torvalds */ 25351da177e4SLinus Torvalds tp->tcp_header_len = sizeof(struct tcphdr) + 2536bb5b7c11SDavid S. Miller (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0); 25371da177e4SLinus Torvalds 2538cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 2539cfb6eeb4SYOSHIFUJI Hideaki if (tp->af_specific->md5_lookup(sk, sk) != NULL) 2540cfb6eeb4SYOSHIFUJI Hideaki tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED; 2541cfb6eeb4SYOSHIFUJI Hideaki #endif 2542cfb6eeb4SYOSHIFUJI Hideaki 25431da177e4SLinus Torvalds /* If user gave his TCP_MAXSEG, record it to clamp */ 25441da177e4SLinus Torvalds if (tp->rx_opt.user_mss) 25451da177e4SLinus Torvalds tp->rx_opt.mss_clamp = tp->rx_opt.user_mss; 25461da177e4SLinus Torvalds tp->max_window = 0; 25475d424d5aSJohn Heffner tcp_mtup_init(sk); 25481da177e4SLinus Torvalds tcp_sync_mss(sk, dst_mtu(dst)); 25491da177e4SLinus Torvalds 25501da177e4SLinus Torvalds if (!tp->window_clamp) 25511da177e4SLinus Torvalds tp->window_clamp = dst_metric(dst, RTAX_WINDOW); 25521da177e4SLinus Torvalds tp->advmss = dst_metric(dst, RTAX_ADVMSS); 2553f5fff5dcSTom Quetchenbach if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->advmss) 2554f5fff5dcSTom Quetchenbach tp->advmss = tp->rx_opt.user_mss; 2555f5fff5dcSTom Quetchenbach 25561da177e4SLinus Torvalds tcp_initialize_rcv_mss(sk); 25571da177e4SLinus Torvalds 2558e88c64f0SHagen Paul Pfeifer /* limit the window selection if the user enforce a smaller rx buffer */ 2559e88c64f0SHagen Paul Pfeifer if (sk->sk_userlocks & SOCK_RCVBUF_LOCK && 2560e88c64f0SHagen Paul Pfeifer (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0)) 2561e88c64f0SHagen Paul Pfeifer tp->window_clamp = tcp_full_space(sk); 2562e88c64f0SHagen Paul Pfeifer 25631da177e4SLinus Torvalds tcp_select_initial_window(tcp_full_space(sk), 25641da177e4SLinus Torvalds tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0), 25651da177e4SLinus Torvalds &tp->rcv_wnd, 25661da177e4SLinus Torvalds &tp->window_clamp, 2567bb5b7c11SDavid S. Miller sysctl_tcp_window_scaling, 256831d12926Slaurent chavey &rcv_wscale, 256931d12926Slaurent chavey dst_metric(dst, RTAX_INITRWND)); 25701da177e4SLinus Torvalds 25711da177e4SLinus Torvalds tp->rx_opt.rcv_wscale = rcv_wscale; 25721da177e4SLinus Torvalds tp->rcv_ssthresh = tp->rcv_wnd; 25731da177e4SLinus Torvalds 25741da177e4SLinus Torvalds sk->sk_err = 0; 25751da177e4SLinus Torvalds sock_reset_flag(sk, SOCK_DONE); 25761da177e4SLinus Torvalds tp->snd_wnd = 0; 2577ee7537b6SHantzis Fotis tcp_init_wl(tp, 0); 25781da177e4SLinus Torvalds tp->snd_una = tp->write_seq; 25791da177e4SLinus Torvalds tp->snd_sml = tp->write_seq; 258033f5f57eSIlpo Järvinen tp->snd_up = tp->write_seq; 25811da177e4SLinus Torvalds tp->rcv_nxt = 0; 25821da177e4SLinus Torvalds tp->rcv_wup = 0; 25831da177e4SLinus Torvalds tp->copied_seq = 0; 25841da177e4SLinus Torvalds 2585463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT; 2586463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_retransmits = 0; 25871da177e4SLinus Torvalds tcp_clear_retrans(tp); 25881da177e4SLinus Torvalds } 25891da177e4SLinus Torvalds 259067edfef7SAndi Kleen /* Build a SYN and send it off. */ 25911da177e4SLinus Torvalds int tcp_connect(struct sock *sk) 25921da177e4SLinus Torvalds { 25931da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 25941da177e4SLinus Torvalds struct sk_buff *buff; 25951da177e4SLinus Torvalds 25961da177e4SLinus Torvalds tcp_connect_init(sk); 25971da177e4SLinus Torvalds 2598d179cd12SDavid S. Miller buff = alloc_skb_fclone(MAX_TCP_HEADER + 15, sk->sk_allocation); 25991da177e4SLinus Torvalds if (unlikely(buff == NULL)) 26001da177e4SLinus Torvalds return -ENOBUFS; 26011da177e4SLinus Torvalds 26021da177e4SLinus Torvalds /* Reserve space for headers. */ 26031da177e4SLinus Torvalds skb_reserve(buff, MAX_TCP_HEADER); 26041da177e4SLinus Torvalds 2605bd37a088SWei Yongjun tp->snd_nxt = tp->write_seq; 2606a3433f35SChangli Gao tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN); 2607e870a8efSIlpo Järvinen TCP_ECN_send_syn(sk, buff); 26081da177e4SLinus Torvalds 26091da177e4SLinus Torvalds /* Send it off. */ 26101da177e4SLinus Torvalds TCP_SKB_CB(buff)->when = tcp_time_stamp; 26111da177e4SLinus Torvalds tp->retrans_stamp = TCP_SKB_CB(buff)->when; 26121da177e4SLinus Torvalds skb_header_release(buff); 2613fe067e8aSDavid S. Miller __tcp_add_write_queue_tail(sk, buff); 26143ab224beSHideo Aoki sk->sk_wmem_queued += buff->truesize; 26153ab224beSHideo Aoki sk_mem_charge(sk, buff->truesize); 26161da177e4SLinus Torvalds tp->packets_out += tcp_skb_pcount(buff); 2617aa133076SWu Fengguang tcp_transmit_skb(sk, buff, 1, sk->sk_allocation); 2618bd37a088SWei Yongjun 2619bd37a088SWei Yongjun /* We change tp->snd_nxt after the tcp_transmit_skb() call 2620bd37a088SWei Yongjun * in order to make this packet get counted in tcpOutSegs. 2621bd37a088SWei Yongjun */ 2622bd37a088SWei Yongjun tp->snd_nxt = tp->write_seq; 2623bd37a088SWei Yongjun tp->pushed_seq = tp->write_seq; 262481cc8a75SPavel Emelyanov TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS); 26251da177e4SLinus Torvalds 26261da177e4SLinus Torvalds /* Timer for repeating the SYN until an answer. */ 26273f421baaSArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 26283f421baaSArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto, TCP_RTO_MAX); 26291da177e4SLinus Torvalds return 0; 26301da177e4SLinus Torvalds } 26314bc2f18bSEric Dumazet EXPORT_SYMBOL(tcp_connect); 26321da177e4SLinus Torvalds 26331da177e4SLinus Torvalds /* Send out a delayed ack, the caller does the policy checking 26341da177e4SLinus Torvalds * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check() 26351da177e4SLinus Torvalds * for details. 26361da177e4SLinus Torvalds */ 26371da177e4SLinus Torvalds void tcp_send_delayed_ack(struct sock *sk) 26381da177e4SLinus Torvalds { 2639463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 2640463c84b9SArnaldo Carvalho de Melo int ato = icsk->icsk_ack.ato; 26411da177e4SLinus Torvalds unsigned long timeout; 26421da177e4SLinus Torvalds 26431da177e4SLinus Torvalds if (ato > TCP_DELACK_MIN) { 2644463c84b9SArnaldo Carvalho de Melo const struct tcp_sock *tp = tcp_sk(sk); 26451da177e4SLinus Torvalds int max_ato = HZ / 2; 26461da177e4SLinus Torvalds 2647056834d9SIlpo Järvinen if (icsk->icsk_ack.pingpong || 2648056834d9SIlpo Järvinen (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)) 26491da177e4SLinus Torvalds max_ato = TCP_DELACK_MAX; 26501da177e4SLinus Torvalds 26511da177e4SLinus Torvalds /* Slow path, intersegment interval is "high". */ 26521da177e4SLinus Torvalds 26531da177e4SLinus Torvalds /* If some rtt estimate is known, use it to bound delayed ack. 2654463c84b9SArnaldo Carvalho de Melo * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements 26551da177e4SLinus Torvalds * directly. 26561da177e4SLinus Torvalds */ 26571da177e4SLinus Torvalds if (tp->srtt) { 26581da177e4SLinus Torvalds int rtt = max(tp->srtt >> 3, TCP_DELACK_MIN); 26591da177e4SLinus Torvalds 26601da177e4SLinus Torvalds if (rtt < max_ato) 26611da177e4SLinus Torvalds max_ato = rtt; 26621da177e4SLinus Torvalds } 26631da177e4SLinus Torvalds 26641da177e4SLinus Torvalds ato = min(ato, max_ato); 26651da177e4SLinus Torvalds } 26661da177e4SLinus Torvalds 26671da177e4SLinus Torvalds /* Stay within the limit we were given */ 26681da177e4SLinus Torvalds timeout = jiffies + ato; 26691da177e4SLinus Torvalds 26701da177e4SLinus Torvalds /* Use new timeout only if there wasn't a older one earlier. */ 2671463c84b9SArnaldo Carvalho de Melo if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) { 26721da177e4SLinus Torvalds /* If delack timer was blocked or is about to expire, 26731da177e4SLinus Torvalds * send ACK now. 26741da177e4SLinus Torvalds */ 2675463c84b9SArnaldo Carvalho de Melo if (icsk->icsk_ack.blocked || 2676463c84b9SArnaldo Carvalho de Melo time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) { 26771da177e4SLinus Torvalds tcp_send_ack(sk); 26781da177e4SLinus Torvalds return; 26791da177e4SLinus Torvalds } 26801da177e4SLinus Torvalds 2681463c84b9SArnaldo Carvalho de Melo if (!time_before(timeout, icsk->icsk_ack.timeout)) 2682463c84b9SArnaldo Carvalho de Melo timeout = icsk->icsk_ack.timeout; 26831da177e4SLinus Torvalds } 2684463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER; 2685463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.timeout = timeout; 2686463c84b9SArnaldo Carvalho de Melo sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout); 26871da177e4SLinus Torvalds } 26881da177e4SLinus Torvalds 26891da177e4SLinus Torvalds /* This routine sends an ack and also updates the window. */ 26901da177e4SLinus Torvalds void tcp_send_ack(struct sock *sk) 26911da177e4SLinus Torvalds { 26921da177e4SLinus Torvalds struct sk_buff *buff; 26931da177e4SLinus Torvalds 2694058dc334SIlpo Järvinen /* If we have been reset, we may not send again. */ 2695058dc334SIlpo Järvinen if (sk->sk_state == TCP_CLOSE) 2696058dc334SIlpo Järvinen return; 2697058dc334SIlpo Järvinen 26981da177e4SLinus Torvalds /* We are not putting this on the write queue, so 26991da177e4SLinus Torvalds * tcp_transmit_skb() will set the ownership to this 27001da177e4SLinus Torvalds * sock. 27011da177e4SLinus Torvalds */ 27021da177e4SLinus Torvalds buff = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC); 27031da177e4SLinus Torvalds if (buff == NULL) { 2704463c84b9SArnaldo Carvalho de Melo inet_csk_schedule_ack(sk); 2705463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN; 27063f421baaSArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK, 27073f421baaSArnaldo Carvalho de Melo TCP_DELACK_MAX, TCP_RTO_MAX); 27081da177e4SLinus Torvalds return; 27091da177e4SLinus Torvalds } 27101da177e4SLinus Torvalds 27111da177e4SLinus Torvalds /* Reserve space for headers and prepare control bits. */ 27121da177e4SLinus Torvalds skb_reserve(buff, MAX_TCP_HEADER); 2713a3433f35SChangli Gao tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK); 27141da177e4SLinus Torvalds 27151da177e4SLinus Torvalds /* Send it off, this clears delayed acks for us. */ 27161da177e4SLinus Torvalds TCP_SKB_CB(buff)->when = tcp_time_stamp; 2717dfb4b9dcSDavid S. Miller tcp_transmit_skb(sk, buff, 0, GFP_ATOMIC); 27181da177e4SLinus Torvalds } 27191da177e4SLinus Torvalds 27201da177e4SLinus Torvalds /* This routine sends a packet with an out of date sequence 27211da177e4SLinus Torvalds * number. It assumes the other end will try to ack it. 27221da177e4SLinus Torvalds * 27231da177e4SLinus Torvalds * Question: what should we make while urgent mode? 27241da177e4SLinus Torvalds * 4.4BSD forces sending single byte of data. We cannot send 27251da177e4SLinus Torvalds * out of window data, because we have SND.NXT==SND.MAX... 27261da177e4SLinus Torvalds * 27271da177e4SLinus Torvalds * Current solution: to send TWO zero-length segments in urgent mode: 27281da177e4SLinus Torvalds * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is 27291da177e4SLinus Torvalds * out-of-date with SND.UNA-1 to probe window. 27301da177e4SLinus Torvalds */ 27311da177e4SLinus Torvalds static int tcp_xmit_probe_skb(struct sock *sk, int urgent) 27321da177e4SLinus Torvalds { 27331da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 27341da177e4SLinus Torvalds struct sk_buff *skb; 27351da177e4SLinus Torvalds 27361da177e4SLinus Torvalds /* We don't queue it, tcp_transmit_skb() sets ownership. */ 27371da177e4SLinus Torvalds skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC); 27381da177e4SLinus Torvalds if (skb == NULL) 27391da177e4SLinus Torvalds return -1; 27401da177e4SLinus Torvalds 27411da177e4SLinus Torvalds /* Reserve space for headers and set control bits. */ 27421da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 27431da177e4SLinus Torvalds /* Use a previous sequence. This should cause the other 27441da177e4SLinus Torvalds * end to send an ack. Don't queue or clone SKB, just 27451da177e4SLinus Torvalds * send it. 27461da177e4SLinus Torvalds */ 2747a3433f35SChangli Gao tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK); 27481da177e4SLinus Torvalds TCP_SKB_CB(skb)->when = tcp_time_stamp; 2749dfb4b9dcSDavid S. Miller return tcp_transmit_skb(sk, skb, 0, GFP_ATOMIC); 27501da177e4SLinus Torvalds } 27511da177e4SLinus Torvalds 275267edfef7SAndi Kleen /* Initiate keepalive or window probe from timer. */ 27531da177e4SLinus Torvalds int tcp_write_wakeup(struct sock *sk) 27541da177e4SLinus Torvalds { 27551da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 27561da177e4SLinus Torvalds struct sk_buff *skb; 27571da177e4SLinus Torvalds 2758058dc334SIlpo Järvinen if (sk->sk_state == TCP_CLOSE) 2759058dc334SIlpo Järvinen return -1; 2760058dc334SIlpo Järvinen 2761fe067e8aSDavid S. Miller if ((skb = tcp_send_head(sk)) != NULL && 276290840defSIlpo Järvinen before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) { 27631da177e4SLinus Torvalds int err; 27640c54b85fSIlpo Järvinen unsigned int mss = tcp_current_mss(sk); 276590840defSIlpo Järvinen unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 27661da177e4SLinus Torvalds 27671da177e4SLinus Torvalds if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq)) 27681da177e4SLinus Torvalds tp->pushed_seq = TCP_SKB_CB(skb)->end_seq; 27691da177e4SLinus Torvalds 27701da177e4SLinus Torvalds /* We are probing the opening of a window 27711da177e4SLinus Torvalds * but the window size is != 0 27721da177e4SLinus Torvalds * must have been a result SWS avoidance ( sender ) 27731da177e4SLinus Torvalds */ 27741da177e4SLinus Torvalds if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq || 27751da177e4SLinus Torvalds skb->len > mss) { 27761da177e4SLinus Torvalds seg_size = min(seg_size, mss); 2777a3433f35SChangli Gao TCP_SKB_CB(skb)->flags |= TCPHDR_PSH; 2778846998aeSDavid S. Miller if (tcp_fragment(sk, skb, seg_size, mss)) 27791da177e4SLinus Torvalds return -1; 27801da177e4SLinus Torvalds } else if (!tcp_skb_pcount(skb)) 2781846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, skb, mss); 27821da177e4SLinus Torvalds 2783a3433f35SChangli Gao TCP_SKB_CB(skb)->flags |= TCPHDR_PSH; 27841da177e4SLinus Torvalds TCP_SKB_CB(skb)->when = tcp_time_stamp; 2785dfb4b9dcSDavid S. Miller err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 278666f5fe62SIlpo Järvinen if (!err) 278766f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, skb); 27881da177e4SLinus Torvalds return err; 27891da177e4SLinus Torvalds } else { 279033f5f57eSIlpo Järvinen if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF)) 27914828e7f4SIlpo Järvinen tcp_xmit_probe_skb(sk, 1); 27921da177e4SLinus Torvalds return tcp_xmit_probe_skb(sk, 0); 27931da177e4SLinus Torvalds } 27941da177e4SLinus Torvalds } 27951da177e4SLinus Torvalds 27961da177e4SLinus Torvalds /* A window probe timeout has occurred. If window is not closed send 27971da177e4SLinus Torvalds * a partial packet else a zero probe. 27981da177e4SLinus Torvalds */ 27991da177e4SLinus Torvalds void tcp_send_probe0(struct sock *sk) 28001da177e4SLinus Torvalds { 2801463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 28021da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 28031da177e4SLinus Torvalds int err; 28041da177e4SLinus Torvalds 28051da177e4SLinus Torvalds err = tcp_write_wakeup(sk); 28061da177e4SLinus Torvalds 2807fe067e8aSDavid S. Miller if (tp->packets_out || !tcp_send_head(sk)) { 28081da177e4SLinus Torvalds /* Cancel probe timer, if it is not required. */ 28096687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out = 0; 2810463c84b9SArnaldo Carvalho de Melo icsk->icsk_backoff = 0; 28111da177e4SLinus Torvalds return; 28121da177e4SLinus Torvalds } 28131da177e4SLinus Torvalds 28141da177e4SLinus Torvalds if (err <= 0) { 2815463c84b9SArnaldo Carvalho de Melo if (icsk->icsk_backoff < sysctl_tcp_retries2) 2816463c84b9SArnaldo Carvalho de Melo icsk->icsk_backoff++; 28176687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out++; 2818463c84b9SArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0, 28193f421baaSArnaldo Carvalho de Melo min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX), 28203f421baaSArnaldo Carvalho de Melo TCP_RTO_MAX); 28211da177e4SLinus Torvalds } else { 28221da177e4SLinus Torvalds /* If packet was not sent due to local congestion, 28236687e988SArnaldo Carvalho de Melo * do not backoff and do not remember icsk_probes_out. 28241da177e4SLinus Torvalds * Let local senders to fight for local resources. 28251da177e4SLinus Torvalds * 28261da177e4SLinus Torvalds * Use accumulated backoff yet. 28271da177e4SLinus Torvalds */ 28286687e988SArnaldo Carvalho de Melo if (!icsk->icsk_probes_out) 28296687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out = 1; 2830463c84b9SArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0, 2831463c84b9SArnaldo Carvalho de Melo min(icsk->icsk_rto << icsk->icsk_backoff, 28323f421baaSArnaldo Carvalho de Melo TCP_RESOURCE_PROBE_INTERVAL), 28333f421baaSArnaldo Carvalho de Melo TCP_RTO_MAX); 28341da177e4SLinus Torvalds } 28351da177e4SLinus Torvalds } 2836