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 * 81da177e4SLinus Torvalds * Version: $Id: tcp_output.c,v 1.146 2002/02/01 22:01:04 davem Exp $ 91da177e4SLinus Torvalds * 1002c30a84SJesper Juhl * Authors: Ross Biro 111da177e4SLinus Torvalds * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG> 121da177e4SLinus Torvalds * Mark Evans, <evansmp@uhura.aston.ac.uk> 131da177e4SLinus Torvalds * Corey Minyard <wf-rch!minyard@relay.EU.net> 141da177e4SLinus Torvalds * Florian La Roche, <flla@stud.uni-sb.de> 151da177e4SLinus Torvalds * Charles Hedrick, <hedrick@klinzhai.rutgers.edu> 161da177e4SLinus Torvalds * Linus Torvalds, <torvalds@cs.helsinki.fi> 171da177e4SLinus Torvalds * Alan Cox, <gw4pts@gw4pts.ampr.org> 181da177e4SLinus Torvalds * Matthew Dillon, <dillon@apollo.west.oic.com> 191da177e4SLinus Torvalds * Arnt Gulbrandsen, <agulbra@nvg.unit.no> 201da177e4SLinus Torvalds * Jorge Cwik, <jorge@laser.satlink.net> 211da177e4SLinus Torvalds */ 221da177e4SLinus Torvalds 231da177e4SLinus Torvalds /* 241da177e4SLinus Torvalds * Changes: Pedro Roque : Retransmit queue handled by TCP. 251da177e4SLinus Torvalds * : Fragmentation on mtu decrease 261da177e4SLinus Torvalds * : Segment collapse on retransmit 271da177e4SLinus Torvalds * : AF independence 281da177e4SLinus Torvalds * 291da177e4SLinus Torvalds * Linus Torvalds : send_delayed_ack 301da177e4SLinus Torvalds * David S. Miller : Charge memory using the right skb 311da177e4SLinus Torvalds * during syn/ack processing. 321da177e4SLinus Torvalds * David S. Miller : Output engine completely rewritten. 331da177e4SLinus Torvalds * Andrea Arcangeli: SYNACK carry ts_recent in tsecr. 341da177e4SLinus Torvalds * Cacophonix Gaul : draft-minshall-nagle-01 351da177e4SLinus Torvalds * J Hadi Salim : ECN support 361da177e4SLinus Torvalds * 371da177e4SLinus Torvalds */ 381da177e4SLinus Torvalds 391da177e4SLinus Torvalds #include <net/tcp.h> 401da177e4SLinus Torvalds 411da177e4SLinus Torvalds #include <linux/compiler.h> 421da177e4SLinus Torvalds #include <linux/module.h> 431da177e4SLinus Torvalds 441da177e4SLinus Torvalds /* People can turn this off for buggy TCP's found in printers etc. */ 45ab32ea5dSBrian Haley int sysctl_tcp_retrans_collapse __read_mostly = 1; 461da177e4SLinus Torvalds 4715d99e02SRick Jones /* People can turn this on to work with those rare, broken TCPs that 4815d99e02SRick Jones * interpret the window field as a signed quantity. 4915d99e02SRick Jones */ 50ab32ea5dSBrian Haley int sysctl_tcp_workaround_signed_windows __read_mostly = 0; 5115d99e02SRick Jones 521da177e4SLinus Torvalds /* This limits the percentage of the congestion window which we 531da177e4SLinus Torvalds * will allow a single TSO frame to consume. Building TSO frames 541da177e4SLinus Torvalds * which are too large can cause TCP streams to be bursty. 551da177e4SLinus Torvalds */ 56ab32ea5dSBrian Haley int sysctl_tcp_tso_win_divisor __read_mostly = 3; 571da177e4SLinus Torvalds 58ab32ea5dSBrian Haley int sysctl_tcp_mtu_probing __read_mostly = 0; 59ab32ea5dSBrian Haley int sysctl_tcp_base_mss __read_mostly = 512; 605d424d5aSJohn Heffner 6135089bb2SDavid S. Miller /* By default, RFC2861 behavior. */ 62ab32ea5dSBrian Haley int sysctl_tcp_slow_start_after_idle __read_mostly = 1; 6335089bb2SDavid S. Miller 6466f5fe62SIlpo Järvinen static void tcp_event_new_data_sent(struct sock *sk, struct sk_buff *skb) 656ff03ac3SIlpo Järvinen { 666ff03ac3SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 6766f5fe62SIlpo Järvinen unsigned int prior_packets = tp->packets_out; 689e412ba7SIlpo Järvinen 69fe067e8aSDavid S. Miller tcp_advance_send_head(sk, skb); 701da177e4SLinus Torvalds tp->snd_nxt = TCP_SKB_CB(skb)->end_seq; 718512430eSIlpo Järvinen 728512430eSIlpo Järvinen /* Don't override Nagle indefinately with F-RTO */ 738512430eSIlpo Järvinen if (tp->frto_counter == 2) 748512430eSIlpo Järvinen tp->frto_counter = 3; 7566f5fe62SIlpo Järvinen 7666f5fe62SIlpo Järvinen tp->packets_out += tcp_skb_pcount(skb); 7766f5fe62SIlpo Järvinen if (!prior_packets) 7866f5fe62SIlpo Järvinen inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 7966f5fe62SIlpo Järvinen inet_csk(sk)->icsk_rto, TCP_RTO_MAX); 801da177e4SLinus Torvalds } 811da177e4SLinus Torvalds 821da177e4SLinus Torvalds /* SND.NXT, if window was not shrunk. 831da177e4SLinus Torvalds * If window has been shrunk, what should we make? It is not clear at all. 841da177e4SLinus Torvalds * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-( 851da177e4SLinus Torvalds * Anything in between SND.UNA...SND.UNA+SND.WND also can be already 861da177e4SLinus Torvalds * invalid. OK, let's make this for now: 871da177e4SLinus Torvalds */ 889e412ba7SIlpo Järvinen static inline __u32 tcp_acceptable_seq(struct sock *sk) 891da177e4SLinus Torvalds { 909e412ba7SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 919e412ba7SIlpo Järvinen 9290840defSIlpo Järvinen if (!before(tcp_wnd_end(tp), tp->snd_nxt)) 931da177e4SLinus Torvalds return tp->snd_nxt; 941da177e4SLinus Torvalds else 9590840defSIlpo Järvinen return tcp_wnd_end(tp); 961da177e4SLinus Torvalds } 971da177e4SLinus Torvalds 981da177e4SLinus Torvalds /* Calculate mss to advertise in SYN segment. 991da177e4SLinus Torvalds * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that: 1001da177e4SLinus Torvalds * 1011da177e4SLinus Torvalds * 1. It is independent of path mtu. 1021da177e4SLinus Torvalds * 2. Ideally, it is maximal possible segment size i.e. 65535-40. 1031da177e4SLinus Torvalds * 3. For IPv4 it is reasonable to calculate it from maximal MTU of 1041da177e4SLinus Torvalds * attached devices, because some buggy hosts are confused by 1051da177e4SLinus Torvalds * large MSS. 1061da177e4SLinus Torvalds * 4. We do not make 3, we advertise MSS, calculated from first 1071da177e4SLinus Torvalds * hop device mtu, but allow to raise it to ip_rt_min_advmss. 1081da177e4SLinus Torvalds * This may be overridden via information stored in routing table. 1091da177e4SLinus Torvalds * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible, 1101da177e4SLinus Torvalds * probably even Jumbo". 1111da177e4SLinus Torvalds */ 1121da177e4SLinus Torvalds static __u16 tcp_advertise_mss(struct sock *sk) 1131da177e4SLinus Torvalds { 1141da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 1151da177e4SLinus Torvalds struct dst_entry *dst = __sk_dst_get(sk); 1161da177e4SLinus Torvalds int mss = tp->advmss; 1171da177e4SLinus Torvalds 1181da177e4SLinus Torvalds if (dst && dst_metric(dst, RTAX_ADVMSS) < mss) { 1191da177e4SLinus Torvalds mss = dst_metric(dst, RTAX_ADVMSS); 1201da177e4SLinus Torvalds tp->advmss = mss; 1211da177e4SLinus Torvalds } 1221da177e4SLinus Torvalds 1231da177e4SLinus Torvalds return (__u16)mss; 1241da177e4SLinus Torvalds } 1251da177e4SLinus Torvalds 1261da177e4SLinus Torvalds /* RFC2861. Reset CWND after idle period longer RTO to "restart window". 1271da177e4SLinus Torvalds * This is the first part of cwnd validation mechanism. */ 128463c84b9SArnaldo Carvalho de Melo static void tcp_cwnd_restart(struct sock *sk, struct dst_entry *dst) 1291da177e4SLinus Torvalds { 130463c84b9SArnaldo Carvalho de Melo struct tcp_sock *tp = tcp_sk(sk); 1311da177e4SLinus Torvalds s32 delta = tcp_time_stamp - tp->lsndtime; 1321da177e4SLinus Torvalds u32 restart_cwnd = tcp_init_cwnd(tp, dst); 1331da177e4SLinus Torvalds u32 cwnd = tp->snd_cwnd; 1341da177e4SLinus Torvalds 1356687e988SArnaldo Carvalho de Melo tcp_ca_event(sk, CA_EVENT_CWND_RESTART); 1361da177e4SLinus Torvalds 1376687e988SArnaldo Carvalho de Melo tp->snd_ssthresh = tcp_current_ssthresh(sk); 1381da177e4SLinus Torvalds restart_cwnd = min(restart_cwnd, cwnd); 1391da177e4SLinus Torvalds 140463c84b9SArnaldo Carvalho de Melo while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd) 1411da177e4SLinus Torvalds cwnd >>= 1; 1421da177e4SLinus Torvalds tp->snd_cwnd = max(cwnd, restart_cwnd); 1431da177e4SLinus Torvalds tp->snd_cwnd_stamp = tcp_time_stamp; 1441da177e4SLinus Torvalds tp->snd_cwnd_used = 0; 1451da177e4SLinus Torvalds } 1461da177e4SLinus Torvalds 14740efc6faSStephen Hemminger static void tcp_event_data_sent(struct tcp_sock *tp, 1481da177e4SLinus Torvalds struct sk_buff *skb, struct sock *sk) 1491da177e4SLinus Torvalds { 150463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 151463c84b9SArnaldo Carvalho de Melo const u32 now = tcp_time_stamp; 1521da177e4SLinus Torvalds 15335089bb2SDavid S. Miller if (sysctl_tcp_slow_start_after_idle && 15435089bb2SDavid S. Miller (!tp->packets_out && (s32)(now - tp->lsndtime) > icsk->icsk_rto)) 155463c84b9SArnaldo Carvalho de Melo tcp_cwnd_restart(sk, __sk_dst_get(sk)); 1561da177e4SLinus Torvalds 1571da177e4SLinus Torvalds tp->lsndtime = now; 1581da177e4SLinus Torvalds 1591da177e4SLinus Torvalds /* If it is a reply for ato after last received 1601da177e4SLinus Torvalds * packet, enter pingpong mode. 1611da177e4SLinus Torvalds */ 162463c84b9SArnaldo Carvalho de Melo if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato) 163463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.pingpong = 1; 1641da177e4SLinus Torvalds } 1651da177e4SLinus Torvalds 16640efc6faSStephen Hemminger static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts) 1671da177e4SLinus Torvalds { 168463c84b9SArnaldo Carvalho de Melo tcp_dec_quickack_mode(sk, pkts); 169463c84b9SArnaldo Carvalho de Melo inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK); 1701da177e4SLinus Torvalds } 1711da177e4SLinus Torvalds 1721da177e4SLinus Torvalds /* Determine a window scaling and initial window to offer. 1731da177e4SLinus Torvalds * Based on the assumption that the given amount of space 1741da177e4SLinus Torvalds * will be offered. Store the results in the tp structure. 1751da177e4SLinus Torvalds * NOTE: for smooth operation initial space offering should 1761da177e4SLinus Torvalds * be a multiple of mss if possible. We assume here that mss >= 1. 1771da177e4SLinus Torvalds * This MUST be enforced by all callers. 1781da177e4SLinus Torvalds */ 1791da177e4SLinus Torvalds void tcp_select_initial_window(int __space, __u32 mss, 1801da177e4SLinus Torvalds __u32 *rcv_wnd, __u32 *window_clamp, 1811da177e4SLinus Torvalds int wscale_ok, __u8 *rcv_wscale) 1821da177e4SLinus Torvalds { 1831da177e4SLinus Torvalds unsigned int space = (__space < 0 ? 0 : __space); 1841da177e4SLinus Torvalds 1851da177e4SLinus Torvalds /* If no clamp set the clamp to the max possible scaled window */ 1861da177e4SLinus Torvalds if (*window_clamp == 0) 1871da177e4SLinus Torvalds (*window_clamp) = (65535 << 14); 1881da177e4SLinus Torvalds space = min(*window_clamp, space); 1891da177e4SLinus Torvalds 1901da177e4SLinus Torvalds /* Quantize space offering to a multiple of mss if possible. */ 1911da177e4SLinus Torvalds if (space > mss) 1921da177e4SLinus Torvalds space = (space / mss) * mss; 1931da177e4SLinus Torvalds 1941da177e4SLinus Torvalds /* NOTE: offering an initial window larger than 32767 19515d99e02SRick Jones * will break some buggy TCP stacks. If the admin tells us 19615d99e02SRick Jones * it is likely we could be speaking with such a buggy stack 19715d99e02SRick Jones * we will truncate our initial window offering to 32K-1 19815d99e02SRick Jones * unless the remote has sent us a window scaling option, 19915d99e02SRick Jones * which we interpret as a sign the remote TCP is not 20015d99e02SRick Jones * misinterpreting the window field as a signed quantity. 2011da177e4SLinus Torvalds */ 20215d99e02SRick Jones if (sysctl_tcp_workaround_signed_windows) 2031da177e4SLinus Torvalds (*rcv_wnd) = min(space, MAX_TCP_WINDOW); 20415d99e02SRick Jones else 20515d99e02SRick Jones (*rcv_wnd) = space; 20615d99e02SRick Jones 2071da177e4SLinus Torvalds (*rcv_wscale) = 0; 2081da177e4SLinus Torvalds if (wscale_ok) { 2091da177e4SLinus Torvalds /* Set window scaling on max possible window 2101da177e4SLinus Torvalds * See RFC1323 for an explanation of the limit to 14 2111da177e4SLinus Torvalds */ 2121da177e4SLinus Torvalds space = max_t(u32, sysctl_tcp_rmem[2], sysctl_rmem_max); 213316c1592SStephen Hemminger space = min_t(u32, space, *window_clamp); 2141da177e4SLinus Torvalds while (space > 65535 && (*rcv_wscale) < 14) { 2151da177e4SLinus Torvalds space >>= 1; 2161da177e4SLinus Torvalds (*rcv_wscale)++; 2171da177e4SLinus Torvalds } 2181da177e4SLinus Torvalds } 2191da177e4SLinus Torvalds 2201da177e4SLinus Torvalds /* Set initial window to value enough for senders, 2216b251858SDavid S. Miller * following RFC2414. Senders, not following this RFC, 2221da177e4SLinus Torvalds * will be satisfied with 2. 2231da177e4SLinus Torvalds */ 2241da177e4SLinus Torvalds if (mss > (1<<*rcv_wscale)) { 22501ff367eSDavid S. Miller int init_cwnd = 4; 22601ff367eSDavid S. Miller if (mss > 1460*3) 2271da177e4SLinus Torvalds init_cwnd = 2; 22801ff367eSDavid S. Miller else if (mss > 1460) 22901ff367eSDavid S. Miller init_cwnd = 3; 2301da177e4SLinus Torvalds if (*rcv_wnd > init_cwnd*mss) 2311da177e4SLinus Torvalds *rcv_wnd = init_cwnd*mss; 2321da177e4SLinus Torvalds } 2331da177e4SLinus Torvalds 2341da177e4SLinus Torvalds /* Set the clamp no higher than max representable value */ 2351da177e4SLinus Torvalds (*window_clamp) = min(65535U << (*rcv_wscale), *window_clamp); 2361da177e4SLinus Torvalds } 2371da177e4SLinus Torvalds 2381da177e4SLinus Torvalds /* Chose a new window to advertise, update state in tcp_sock for the 2391da177e4SLinus Torvalds * socket, and return result with RFC1323 scaling applied. The return 2401da177e4SLinus Torvalds * value can be stuffed directly into th->window for an outgoing 2411da177e4SLinus Torvalds * frame. 2421da177e4SLinus Torvalds */ 24340efc6faSStephen Hemminger static u16 tcp_select_window(struct sock *sk) 2441da177e4SLinus Torvalds { 2451da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 2461da177e4SLinus Torvalds u32 cur_win = tcp_receive_window(tp); 2471da177e4SLinus Torvalds u32 new_win = __tcp_select_window(sk); 2481da177e4SLinus Torvalds 2491da177e4SLinus Torvalds /* Never shrink the offered window */ 2501da177e4SLinus Torvalds if (new_win < cur_win) { 2511da177e4SLinus Torvalds /* Danger Will Robinson! 2521da177e4SLinus Torvalds * Don't update rcv_wup/rcv_wnd here or else 2531da177e4SLinus Torvalds * we will not be able to advertise a zero 2541da177e4SLinus Torvalds * window in time. --DaveM 2551da177e4SLinus Torvalds * 2561da177e4SLinus Torvalds * Relax Will Robinson. 2571da177e4SLinus Torvalds */ 2581da177e4SLinus Torvalds new_win = cur_win; 2591da177e4SLinus Torvalds } 2601da177e4SLinus Torvalds tp->rcv_wnd = new_win; 2611da177e4SLinus Torvalds tp->rcv_wup = tp->rcv_nxt; 2621da177e4SLinus Torvalds 2631da177e4SLinus Torvalds /* Make sure we do not exceed the maximum possible 2641da177e4SLinus Torvalds * scaled window. 2651da177e4SLinus Torvalds */ 26615d99e02SRick Jones if (!tp->rx_opt.rcv_wscale && sysctl_tcp_workaround_signed_windows) 2671da177e4SLinus Torvalds new_win = min(new_win, MAX_TCP_WINDOW); 2681da177e4SLinus Torvalds else 2691da177e4SLinus Torvalds new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale)); 2701da177e4SLinus Torvalds 2711da177e4SLinus Torvalds /* RFC1323 scaling applied */ 2721da177e4SLinus Torvalds new_win >>= tp->rx_opt.rcv_wscale; 2731da177e4SLinus Torvalds 2741da177e4SLinus Torvalds /* If we advertise zero window, disable fast path. */ 2751da177e4SLinus Torvalds if (new_win == 0) 2761da177e4SLinus Torvalds tp->pred_flags = 0; 2771da177e4SLinus Torvalds 2781da177e4SLinus Torvalds return new_win; 2791da177e4SLinus Torvalds } 2801da177e4SLinus Torvalds 281bdf1ee5dSIlpo Järvinen static inline void TCP_ECN_send_synack(struct tcp_sock *tp, 282bdf1ee5dSIlpo Järvinen struct sk_buff *skb) 283bdf1ee5dSIlpo Järvinen { 284bdf1ee5dSIlpo Järvinen TCP_SKB_CB(skb)->flags &= ~TCPCB_FLAG_CWR; 285bdf1ee5dSIlpo Järvinen if (!(tp->ecn_flags&TCP_ECN_OK)) 286bdf1ee5dSIlpo Järvinen TCP_SKB_CB(skb)->flags &= ~TCPCB_FLAG_ECE; 287bdf1ee5dSIlpo Järvinen } 288bdf1ee5dSIlpo Järvinen 289bdf1ee5dSIlpo Järvinen static inline void TCP_ECN_send_syn(struct sock *sk, struct sk_buff *skb) 290bdf1ee5dSIlpo Järvinen { 291bdf1ee5dSIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 292bdf1ee5dSIlpo Järvinen 293bdf1ee5dSIlpo Järvinen tp->ecn_flags = 0; 294bdf1ee5dSIlpo Järvinen if (sysctl_tcp_ecn) { 295bdf1ee5dSIlpo Järvinen TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_ECE|TCPCB_FLAG_CWR; 296bdf1ee5dSIlpo Järvinen tp->ecn_flags = TCP_ECN_OK; 297bdf1ee5dSIlpo Järvinen } 298bdf1ee5dSIlpo Järvinen } 299bdf1ee5dSIlpo Järvinen 300bdf1ee5dSIlpo Järvinen static __inline__ void 301bdf1ee5dSIlpo Järvinen TCP_ECN_make_synack(struct request_sock *req, struct tcphdr *th) 302bdf1ee5dSIlpo Järvinen { 303bdf1ee5dSIlpo Järvinen if (inet_rsk(req)->ecn_ok) 304bdf1ee5dSIlpo Järvinen th->ece = 1; 305bdf1ee5dSIlpo Järvinen } 306bdf1ee5dSIlpo Järvinen 307bdf1ee5dSIlpo Järvinen static inline void TCP_ECN_send(struct sock *sk, struct sk_buff *skb, 308bdf1ee5dSIlpo Järvinen int tcp_header_len) 309bdf1ee5dSIlpo Järvinen { 310bdf1ee5dSIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 311bdf1ee5dSIlpo Järvinen 312bdf1ee5dSIlpo Järvinen if (tp->ecn_flags & TCP_ECN_OK) { 313bdf1ee5dSIlpo Järvinen /* Not-retransmitted data segment: set ECT and inject CWR. */ 314bdf1ee5dSIlpo Järvinen if (skb->len != tcp_header_len && 315bdf1ee5dSIlpo Järvinen !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) { 316bdf1ee5dSIlpo Järvinen INET_ECN_xmit(sk); 317bdf1ee5dSIlpo Järvinen if (tp->ecn_flags&TCP_ECN_QUEUE_CWR) { 318bdf1ee5dSIlpo Järvinen tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR; 319bdf1ee5dSIlpo Järvinen tcp_hdr(skb)->cwr = 1; 320bdf1ee5dSIlpo Järvinen skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN; 321bdf1ee5dSIlpo Järvinen } 322bdf1ee5dSIlpo Järvinen } else { 323bdf1ee5dSIlpo Järvinen /* ACK or retransmitted segment: clear ECT|CE */ 324bdf1ee5dSIlpo Järvinen INET_ECN_dontxmit(sk); 325bdf1ee5dSIlpo Järvinen } 326bdf1ee5dSIlpo Järvinen if (tp->ecn_flags & TCP_ECN_DEMAND_CWR) 327bdf1ee5dSIlpo Järvinen tcp_hdr(skb)->ece = 1; 328bdf1ee5dSIlpo Järvinen } 329bdf1ee5dSIlpo Järvinen } 330bdf1ee5dSIlpo Järvinen 331df7a3b07SAl Viro static void tcp_build_and_update_options(__be32 *ptr, struct tcp_sock *tp, 332cfb6eeb4SYOSHIFUJI Hideaki __u32 tstamp, __u8 **md5_hash) 33340efc6faSStephen Hemminger { 33440efc6faSStephen Hemminger if (tp->rx_opt.tstamp_ok) { 335496c98dfSYOSHIFUJI Hideaki *ptr++ = htonl((TCPOPT_NOP << 24) | 33640efc6faSStephen Hemminger (TCPOPT_NOP << 16) | 33740efc6faSStephen Hemminger (TCPOPT_TIMESTAMP << 8) | 33840efc6faSStephen Hemminger TCPOLEN_TIMESTAMP); 33940efc6faSStephen Hemminger *ptr++ = htonl(tstamp); 34040efc6faSStephen Hemminger *ptr++ = htonl(tp->rx_opt.ts_recent); 34140efc6faSStephen Hemminger } 34240efc6faSStephen Hemminger if (tp->rx_opt.eff_sacks) { 34340efc6faSStephen Hemminger struct tcp_sack_block *sp = tp->rx_opt.dsack ? tp->duplicate_sack : tp->selective_acks; 34440efc6faSStephen Hemminger int this_sack; 34540efc6faSStephen Hemminger 34640efc6faSStephen Hemminger *ptr++ = htonl((TCPOPT_NOP << 24) | 34740efc6faSStephen Hemminger (TCPOPT_NOP << 16) | 34840efc6faSStephen Hemminger (TCPOPT_SACK << 8) | 34940efc6faSStephen Hemminger (TCPOLEN_SACK_BASE + (tp->rx_opt.eff_sacks * 35040efc6faSStephen Hemminger TCPOLEN_SACK_PERBLOCK))); 3512de979bdSStephen Hemminger 35240efc6faSStephen Hemminger for (this_sack = 0; this_sack < tp->rx_opt.eff_sacks; this_sack++) { 35340efc6faSStephen Hemminger *ptr++ = htonl(sp[this_sack].start_seq); 35440efc6faSStephen Hemminger *ptr++ = htonl(sp[this_sack].end_seq); 35540efc6faSStephen Hemminger } 3562de979bdSStephen Hemminger 35740efc6faSStephen Hemminger if (tp->rx_opt.dsack) { 35840efc6faSStephen Hemminger tp->rx_opt.dsack = 0; 35940efc6faSStephen Hemminger tp->rx_opt.eff_sacks--; 36040efc6faSStephen Hemminger } 36140efc6faSStephen Hemminger } 362cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 363cfb6eeb4SYOSHIFUJI Hideaki if (md5_hash) { 364cfb6eeb4SYOSHIFUJI Hideaki *ptr++ = htonl((TCPOPT_NOP << 24) | 365cfb6eeb4SYOSHIFUJI Hideaki (TCPOPT_NOP << 16) | 366cfb6eeb4SYOSHIFUJI Hideaki (TCPOPT_MD5SIG << 8) | 367cfb6eeb4SYOSHIFUJI Hideaki TCPOLEN_MD5SIG); 368cfb6eeb4SYOSHIFUJI Hideaki *md5_hash = (__u8 *)ptr; 369cfb6eeb4SYOSHIFUJI Hideaki } 370cfb6eeb4SYOSHIFUJI Hideaki #endif 37140efc6faSStephen Hemminger } 37240efc6faSStephen Hemminger 37340efc6faSStephen Hemminger /* Construct a tcp options header for a SYN or SYN_ACK packet. 37440efc6faSStephen Hemminger * If this is every changed make sure to change the definition of 37540efc6faSStephen Hemminger * MAX_SYN_SIZE to match the new maximum number of options that you 37640efc6faSStephen Hemminger * can generate. 377cfb6eeb4SYOSHIFUJI Hideaki * 378cfb6eeb4SYOSHIFUJI Hideaki * Note - that with the RFC2385 TCP option, we make room for the 379cfb6eeb4SYOSHIFUJI Hideaki * 16 byte MD5 hash. This will be filled in later, so the pointer for the 380cfb6eeb4SYOSHIFUJI Hideaki * location to be filled is passed back up. 38140efc6faSStephen Hemminger */ 382df7a3b07SAl Viro static void tcp_syn_build_options(__be32 *ptr, int mss, int ts, int sack, 38340efc6faSStephen Hemminger int offer_wscale, int wscale, __u32 tstamp, 384cfb6eeb4SYOSHIFUJI Hideaki __u32 ts_recent, __u8 **md5_hash) 38540efc6faSStephen Hemminger { 38640efc6faSStephen Hemminger /* We always get an MSS option. 38740efc6faSStephen Hemminger * The option bytes which will be seen in normal data 38840efc6faSStephen Hemminger * packets should timestamps be used, must be in the MSS 38940efc6faSStephen Hemminger * advertised. But we subtract them from tp->mss_cache so 39040efc6faSStephen Hemminger * that calculations in tcp_sendmsg are simpler etc. 39140efc6faSStephen Hemminger * So account for this fact here if necessary. If we 39240efc6faSStephen Hemminger * don't do this correctly, as a receiver we won't 39340efc6faSStephen Hemminger * recognize data packets as being full sized when we 39440efc6faSStephen Hemminger * should, and thus we won't abide by the delayed ACK 39540efc6faSStephen Hemminger * rules correctly. 39640efc6faSStephen Hemminger * SACKs don't matter, we never delay an ACK when we 39740efc6faSStephen Hemminger * have any of those going out. 39840efc6faSStephen Hemminger */ 39940efc6faSStephen Hemminger *ptr++ = htonl((TCPOPT_MSS << 24) | (TCPOLEN_MSS << 16) | mss); 40040efc6faSStephen Hemminger if (ts) { 40140efc6faSStephen Hemminger if (sack) 402496c98dfSYOSHIFUJI Hideaki *ptr++ = htonl((TCPOPT_SACK_PERM << 24) | 403496c98dfSYOSHIFUJI Hideaki (TCPOLEN_SACK_PERM << 16) | 404496c98dfSYOSHIFUJI Hideaki (TCPOPT_TIMESTAMP << 8) | 405496c98dfSYOSHIFUJI Hideaki TCPOLEN_TIMESTAMP); 40640efc6faSStephen Hemminger else 407496c98dfSYOSHIFUJI Hideaki *ptr++ = htonl((TCPOPT_NOP << 24) | 408496c98dfSYOSHIFUJI Hideaki (TCPOPT_NOP << 16) | 409496c98dfSYOSHIFUJI Hideaki (TCPOPT_TIMESTAMP << 8) | 410496c98dfSYOSHIFUJI Hideaki TCPOLEN_TIMESTAMP); 41140efc6faSStephen Hemminger *ptr++ = htonl(tstamp); /* TSVAL */ 41240efc6faSStephen Hemminger *ptr++ = htonl(ts_recent); /* TSECR */ 41340efc6faSStephen Hemminger } else if (sack) 414496c98dfSYOSHIFUJI Hideaki *ptr++ = htonl((TCPOPT_NOP << 24) | 415496c98dfSYOSHIFUJI Hideaki (TCPOPT_NOP << 16) | 416496c98dfSYOSHIFUJI Hideaki (TCPOPT_SACK_PERM << 8) | 417496c98dfSYOSHIFUJI Hideaki TCPOLEN_SACK_PERM); 41840efc6faSStephen Hemminger if (offer_wscale) 419496c98dfSYOSHIFUJI Hideaki *ptr++ = htonl((TCPOPT_NOP << 24) | 420496c98dfSYOSHIFUJI Hideaki (TCPOPT_WINDOW << 16) | 421496c98dfSYOSHIFUJI Hideaki (TCPOLEN_WINDOW << 8) | 422496c98dfSYOSHIFUJI Hideaki (wscale)); 423cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 424cfb6eeb4SYOSHIFUJI Hideaki /* 425cfb6eeb4SYOSHIFUJI Hideaki * If MD5 is enabled, then we set the option, and include the size 426cfb6eeb4SYOSHIFUJI Hideaki * (always 18). The actual MD5 hash is added just before the 427cfb6eeb4SYOSHIFUJI Hideaki * packet is sent. 428cfb6eeb4SYOSHIFUJI Hideaki */ 429cfb6eeb4SYOSHIFUJI Hideaki if (md5_hash) { 430cfb6eeb4SYOSHIFUJI Hideaki *ptr++ = htonl((TCPOPT_NOP << 24) | 431cfb6eeb4SYOSHIFUJI Hideaki (TCPOPT_NOP << 16) | 432cfb6eeb4SYOSHIFUJI Hideaki (TCPOPT_MD5SIG << 8) | 433cfb6eeb4SYOSHIFUJI Hideaki TCPOLEN_MD5SIG); 434cfb6eeb4SYOSHIFUJI Hideaki *md5_hash = (__u8 *) ptr; 435cfb6eeb4SYOSHIFUJI Hideaki } 436cfb6eeb4SYOSHIFUJI Hideaki #endif 43740efc6faSStephen Hemminger } 4381da177e4SLinus Torvalds 4391da177e4SLinus Torvalds /* This routine actually transmits TCP packets queued in by 4401da177e4SLinus Torvalds * tcp_do_sendmsg(). This is used by both the initial 4411da177e4SLinus Torvalds * transmission and possible later retransmissions. 4421da177e4SLinus Torvalds * All SKB's seen here are completely headerless. It is our 4431da177e4SLinus Torvalds * job to build the TCP header, and pass the packet down to 4441da177e4SLinus Torvalds * IP so it can do the same plus pass the packet off to the 4451da177e4SLinus Torvalds * device. 4461da177e4SLinus Torvalds * 4471da177e4SLinus Torvalds * We are working here with either a clone of the original 4481da177e4SLinus Torvalds * SKB, or a fresh unique copy made by the retransmit engine. 4491da177e4SLinus Torvalds */ 450dfb4b9dcSDavid S. Miller static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it, gfp_t gfp_mask) 4511da177e4SLinus Torvalds { 4526687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 453dfb4b9dcSDavid S. Miller struct inet_sock *inet; 454dfb4b9dcSDavid S. Miller struct tcp_sock *tp; 455dfb4b9dcSDavid S. Miller struct tcp_skb_cb *tcb; 456dfb4b9dcSDavid S. Miller int tcp_header_size; 457cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 458cfb6eeb4SYOSHIFUJI Hideaki struct tcp_md5sig_key *md5; 459cfb6eeb4SYOSHIFUJI Hideaki __u8 *md5_hash_location; 460cfb6eeb4SYOSHIFUJI Hideaki #endif 4611da177e4SLinus Torvalds struct tcphdr *th; 4621da177e4SLinus Torvalds int sysctl_flags; 4631da177e4SLinus Torvalds int err; 4641da177e4SLinus Torvalds 465dfb4b9dcSDavid S. Miller BUG_ON(!skb || !tcp_skb_pcount(skb)); 466dfb4b9dcSDavid S. Miller 467dfb4b9dcSDavid S. Miller /* If congestion control is doing timestamping, we must 468dfb4b9dcSDavid S. Miller * take such a timestamp before we potentially clone/copy. 469dfb4b9dcSDavid S. Miller */ 470164891aaSStephen Hemminger if (icsk->icsk_ca_ops->flags & TCP_CONG_RTT_STAMP) 471dfb4b9dcSDavid S. Miller __net_timestamp(skb); 472dfb4b9dcSDavid S. Miller 473dfb4b9dcSDavid S. Miller if (likely(clone_it)) { 474dfb4b9dcSDavid S. Miller if (unlikely(skb_cloned(skb))) 475dfb4b9dcSDavid S. Miller skb = pskb_copy(skb, gfp_mask); 476dfb4b9dcSDavid S. Miller else 477dfb4b9dcSDavid S. Miller skb = skb_clone(skb, gfp_mask); 478dfb4b9dcSDavid S. Miller if (unlikely(!skb)) 479dfb4b9dcSDavid S. Miller return -ENOBUFS; 480dfb4b9dcSDavid S. Miller } 481dfb4b9dcSDavid S. Miller 482dfb4b9dcSDavid S. Miller inet = inet_sk(sk); 483dfb4b9dcSDavid S. Miller tp = tcp_sk(sk); 484dfb4b9dcSDavid S. Miller tcb = TCP_SKB_CB(skb); 485dfb4b9dcSDavid S. Miller tcp_header_size = tp->tcp_header_len; 4861da177e4SLinus Torvalds 4871da177e4SLinus Torvalds #define SYSCTL_FLAG_TSTAMPS 0x1 4881da177e4SLinus Torvalds #define SYSCTL_FLAG_WSCALE 0x2 4891da177e4SLinus Torvalds #define SYSCTL_FLAG_SACK 0x4 4901da177e4SLinus Torvalds 4911da177e4SLinus Torvalds sysctl_flags = 0; 492dfb4b9dcSDavid S. Miller if (unlikely(tcb->flags & TCPCB_FLAG_SYN)) { 4931da177e4SLinus Torvalds tcp_header_size = sizeof(struct tcphdr) + TCPOLEN_MSS; 4941da177e4SLinus Torvalds if (sysctl_tcp_timestamps) { 4951da177e4SLinus Torvalds tcp_header_size += TCPOLEN_TSTAMP_ALIGNED; 4961da177e4SLinus Torvalds sysctl_flags |= SYSCTL_FLAG_TSTAMPS; 4971da177e4SLinus Torvalds } 4981da177e4SLinus Torvalds if (sysctl_tcp_window_scaling) { 4991da177e4SLinus Torvalds tcp_header_size += TCPOLEN_WSCALE_ALIGNED; 5001da177e4SLinus Torvalds sysctl_flags |= SYSCTL_FLAG_WSCALE; 5011da177e4SLinus Torvalds } 5021da177e4SLinus Torvalds if (sysctl_tcp_sack) { 5031da177e4SLinus Torvalds sysctl_flags |= SYSCTL_FLAG_SACK; 5041da177e4SLinus Torvalds if (!(sysctl_flags & SYSCTL_FLAG_TSTAMPS)) 5051da177e4SLinus Torvalds tcp_header_size += TCPOLEN_SACKPERM_ALIGNED; 5061da177e4SLinus Torvalds } 507dfb4b9dcSDavid S. Miller } else if (unlikely(tp->rx_opt.eff_sacks)) { 5081da177e4SLinus Torvalds /* A SACK is 2 pad bytes, a 2 byte header, plus 5091da177e4SLinus Torvalds * 2 32-bit sequence numbers for each SACK block. 5101da177e4SLinus Torvalds */ 5111da177e4SLinus Torvalds tcp_header_size += (TCPOLEN_SACK_BASE_ALIGNED + 512dfb4b9dcSDavid S. Miller (tp->rx_opt.eff_sacks * 513dfb4b9dcSDavid S. Miller TCPOLEN_SACK_PERBLOCK)); 5141da177e4SLinus Torvalds } 5151da177e4SLinus Torvalds 516317a76f9SStephen Hemminger if (tcp_packets_in_flight(tp) == 0) 5176687e988SArnaldo Carvalho de Melo tcp_ca_event(sk, CA_EVENT_TX_START); 5181da177e4SLinus Torvalds 519cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 520cfb6eeb4SYOSHIFUJI Hideaki /* 521cfb6eeb4SYOSHIFUJI Hideaki * Are we doing MD5 on this segment? If so - make 522cfb6eeb4SYOSHIFUJI Hideaki * room for it. 523cfb6eeb4SYOSHIFUJI Hideaki */ 524cfb6eeb4SYOSHIFUJI Hideaki md5 = tp->af_specific->md5_lookup(sk, sk); 525cfb6eeb4SYOSHIFUJI Hideaki if (md5) 526cfb6eeb4SYOSHIFUJI Hideaki tcp_header_size += TCPOLEN_MD5SIG_ALIGNED; 527cfb6eeb4SYOSHIFUJI Hideaki #endif 528cfb6eeb4SYOSHIFUJI Hideaki 529aa8223c7SArnaldo Carvalho de Melo skb_push(skb, tcp_header_size); 530aa8223c7SArnaldo Carvalho de Melo skb_reset_transport_header(skb); 531e89862f4SDavid S. Miller skb_set_owner_w(skb, sk); 5321da177e4SLinus Torvalds 5331da177e4SLinus Torvalds /* Build TCP header and checksum it. */ 534aa8223c7SArnaldo Carvalho de Melo th = tcp_hdr(skb); 5351da177e4SLinus Torvalds th->source = inet->sport; 5361da177e4SLinus Torvalds th->dest = inet->dport; 5371da177e4SLinus Torvalds th->seq = htonl(tcb->seq); 5381da177e4SLinus Torvalds th->ack_seq = htonl(tp->rcv_nxt); 539df7a3b07SAl Viro *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) | 540dfb4b9dcSDavid S. Miller tcb->flags); 541dfb4b9dcSDavid S. Miller 542dfb4b9dcSDavid S. Miller if (unlikely(tcb->flags & TCPCB_FLAG_SYN)) { 5431da177e4SLinus Torvalds /* RFC1323: The window in SYN & SYN/ACK segments 5441da177e4SLinus Torvalds * is never scaled. 5451da177e4SLinus Torvalds */ 546600ff0c2SIlpo Järvinen th->window = htons(min(tp->rcv_wnd, 65535U)); 5471da177e4SLinus Torvalds } else { 5481da177e4SLinus Torvalds th->window = htons(tcp_select_window(sk)); 5491da177e4SLinus Torvalds } 5501da177e4SLinus Torvalds th->check = 0; 5511da177e4SLinus Torvalds th->urg_ptr = 0; 5521da177e4SLinus Torvalds 553dfb4b9dcSDavid S. Miller if (unlikely(tp->urg_mode && 554dfb4b9dcSDavid S. Miller between(tp->snd_up, tcb->seq+1, tcb->seq+0xFFFF))) { 5551da177e4SLinus Torvalds th->urg_ptr = htons(tp->snd_up-tcb->seq); 5561da177e4SLinus Torvalds th->urg = 1; 5571da177e4SLinus Torvalds } 5581da177e4SLinus Torvalds 559dfb4b9dcSDavid S. Miller if (unlikely(tcb->flags & TCPCB_FLAG_SYN)) { 560df7a3b07SAl Viro tcp_syn_build_options((__be32 *)(th + 1), 5611da177e4SLinus Torvalds tcp_advertise_mss(sk), 5621da177e4SLinus Torvalds (sysctl_flags & SYSCTL_FLAG_TSTAMPS), 5631da177e4SLinus Torvalds (sysctl_flags & SYSCTL_FLAG_SACK), 5641da177e4SLinus Torvalds (sysctl_flags & SYSCTL_FLAG_WSCALE), 5651da177e4SLinus Torvalds tp->rx_opt.rcv_wscale, 5661da177e4SLinus Torvalds tcb->when, 567cfb6eeb4SYOSHIFUJI Hideaki tp->rx_opt.ts_recent, 568cfb6eeb4SYOSHIFUJI Hideaki 569cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 570cfb6eeb4SYOSHIFUJI Hideaki md5 ? &md5_hash_location : 571cfb6eeb4SYOSHIFUJI Hideaki #endif 572cfb6eeb4SYOSHIFUJI Hideaki NULL); 5731da177e4SLinus Torvalds } else { 574df7a3b07SAl Viro tcp_build_and_update_options((__be32 *)(th + 1), 575cfb6eeb4SYOSHIFUJI Hideaki tp, tcb->when, 576cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 577cfb6eeb4SYOSHIFUJI Hideaki md5 ? &md5_hash_location : 578cfb6eeb4SYOSHIFUJI Hideaki #endif 579cfb6eeb4SYOSHIFUJI Hideaki NULL); 5809e412ba7SIlpo Järvinen TCP_ECN_send(sk, skb, tcp_header_size); 5811da177e4SLinus Torvalds } 582dfb4b9dcSDavid S. Miller 583cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 584cfb6eeb4SYOSHIFUJI Hideaki /* Calculate the MD5 hash, as we have all we need now */ 585cfb6eeb4SYOSHIFUJI Hideaki if (md5) { 586cfb6eeb4SYOSHIFUJI Hideaki tp->af_specific->calc_md5_hash(md5_hash_location, 587cfb6eeb4SYOSHIFUJI Hideaki md5, 588cfb6eeb4SYOSHIFUJI Hideaki sk, NULL, NULL, 589aa8223c7SArnaldo Carvalho de Melo tcp_hdr(skb), 590cfb6eeb4SYOSHIFUJI Hideaki sk->sk_protocol, 591cfb6eeb4SYOSHIFUJI Hideaki skb->len); 592cfb6eeb4SYOSHIFUJI Hideaki } 593cfb6eeb4SYOSHIFUJI Hideaki #endif 594cfb6eeb4SYOSHIFUJI Hideaki 5958292a17aSArnaldo Carvalho de Melo icsk->icsk_af_ops->send_check(sk, skb->len, skb); 5961da177e4SLinus Torvalds 597dfb4b9dcSDavid S. Miller if (likely(tcb->flags & TCPCB_FLAG_ACK)) 598fc6415bcSDavid S. Miller tcp_event_ack_sent(sk, tcp_skb_pcount(skb)); 5991da177e4SLinus Torvalds 6001da177e4SLinus Torvalds if (skb->len != tcp_header_size) 6011da177e4SLinus Torvalds tcp_event_data_sent(tp, skb, sk); 6021da177e4SLinus Torvalds 603bd37a088SWei Yongjun if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq) 6041da177e4SLinus Torvalds TCP_INC_STATS(TCP_MIB_OUTSEGS); 6051da177e4SLinus Torvalds 606e89862f4SDavid S. Miller err = icsk->icsk_af_ops->queue_xmit(skb, 0); 60783de47cdSHua Zhong if (likely(err <= 0)) 6081da177e4SLinus Torvalds return err; 6091da177e4SLinus Torvalds 6103cfe3baaSIlpo Järvinen tcp_enter_cwr(sk, 1); 6111da177e4SLinus Torvalds 612b9df3cb8SGerrit Renker return net_xmit_eval(err); 613dfb4b9dcSDavid S. Miller 6141da177e4SLinus Torvalds #undef SYSCTL_FLAG_TSTAMPS 6151da177e4SLinus Torvalds #undef SYSCTL_FLAG_WSCALE 6161da177e4SLinus Torvalds #undef SYSCTL_FLAG_SACK 6171da177e4SLinus Torvalds } 6181da177e4SLinus Torvalds 6191da177e4SLinus Torvalds 6201da177e4SLinus Torvalds /* This routine just queue's the buffer 6211da177e4SLinus Torvalds * 6221da177e4SLinus Torvalds * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames, 6231da177e4SLinus Torvalds * otherwise socket can stall. 6241da177e4SLinus Torvalds */ 6251da177e4SLinus Torvalds static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb) 6261da177e4SLinus Torvalds { 6271da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 6281da177e4SLinus Torvalds 6291da177e4SLinus Torvalds /* Advance write_seq and place onto the write_queue. */ 6301da177e4SLinus Torvalds tp->write_seq = TCP_SKB_CB(skb)->end_seq; 6311da177e4SLinus Torvalds skb_header_release(skb); 632fe067e8aSDavid S. Miller tcp_add_write_queue_tail(sk, skb); 6333ab224beSHideo Aoki sk->sk_wmem_queued += skb->truesize; 6343ab224beSHideo Aoki sk_mem_charge(sk, skb->truesize); 6351da177e4SLinus Torvalds } 6361da177e4SLinus Torvalds 637846998aeSDavid S. Miller static void tcp_set_skb_tso_segs(struct sock *sk, struct sk_buff *skb, unsigned int mss_now) 638f6302d1dSDavid S. Miller { 639bcd76111SHerbert Xu if (skb->len <= mss_now || !sk_can_gso(sk)) { 640f6302d1dSDavid S. Miller /* Avoid the costly divide in the normal 641f6302d1dSDavid S. Miller * non-TSO case. 642f6302d1dSDavid S. Miller */ 6437967168cSHerbert Xu skb_shinfo(skb)->gso_segs = 1; 6447967168cSHerbert Xu skb_shinfo(skb)->gso_size = 0; 6457967168cSHerbert Xu skb_shinfo(skb)->gso_type = 0; 646f6302d1dSDavid S. Miller } else { 647356f89e1SIlpo Järvinen skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss_now); 6487967168cSHerbert Xu skb_shinfo(skb)->gso_size = mss_now; 649bcd76111SHerbert Xu skb_shinfo(skb)->gso_type = sk->sk_gso_type; 6501da177e4SLinus Torvalds } 6511da177e4SLinus Torvalds } 6521da177e4SLinus Torvalds 65391fed7a1SIlpo Järvinen /* When a modification to fackets out becomes necessary, we need to check 65468f8353bSIlpo Järvinen * skb is counted to fackets_out or not. 65591fed7a1SIlpo Järvinen */ 656a47e5a98SIlpo Järvinen static void tcp_adjust_fackets_out(struct sock *sk, struct sk_buff *skb, 65791fed7a1SIlpo Järvinen int decr) 65891fed7a1SIlpo Järvinen { 659a47e5a98SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 660a47e5a98SIlpo Järvinen 661dc86967bSIlpo Järvinen if (!tp->sacked_out || tcp_is_reno(tp)) 66291fed7a1SIlpo Järvinen return; 66391fed7a1SIlpo Järvinen 6646859d494SIlpo Järvinen if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq)) 66591fed7a1SIlpo Järvinen tp->fackets_out -= decr; 66691fed7a1SIlpo Järvinen } 66791fed7a1SIlpo Järvinen 6681da177e4SLinus Torvalds /* Function to create two new TCP segments. Shrinks the given segment 6691da177e4SLinus Torvalds * to the specified size and appends a new segment with the rest of the 6701da177e4SLinus Torvalds * packet to the list. This won't be called frequently, I hope. 6711da177e4SLinus Torvalds * Remember, these are still headerless SKBs at this point. 6721da177e4SLinus Torvalds */ 6736475be16SDavid S. Miller int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len, unsigned int mss_now) 6741da177e4SLinus Torvalds { 6751da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 6761da177e4SLinus Torvalds struct sk_buff *buff; 6776475be16SDavid S. Miller int nsize, old_factor; 678b60b49eaSHerbert Xu int nlen; 6791da177e4SLinus Torvalds u16 flags; 6801da177e4SLinus Torvalds 681b2cc99f0SHerbert Xu BUG_ON(len > skb->len); 6826a438bbeSStephen Hemminger 683b7689205SIlpo Järvinen tcp_clear_retrans_hints_partial(tp); 6841da177e4SLinus Torvalds nsize = skb_headlen(skb) - len; 6851da177e4SLinus Torvalds if (nsize < 0) 6861da177e4SLinus Torvalds nsize = 0; 6871da177e4SLinus Torvalds 6881da177e4SLinus Torvalds if (skb_cloned(skb) && 6891da177e4SLinus Torvalds skb_is_nonlinear(skb) && 6901da177e4SLinus Torvalds pskb_expand_head(skb, 0, 0, GFP_ATOMIC)) 6911da177e4SLinus Torvalds return -ENOMEM; 6921da177e4SLinus Torvalds 6931da177e4SLinus Torvalds /* Get a new skb... force flag on. */ 6941da177e4SLinus Torvalds buff = sk_stream_alloc_skb(sk, nsize, GFP_ATOMIC); 6951da177e4SLinus Torvalds if (buff == NULL) 6961da177e4SLinus Torvalds return -ENOMEM; /* We'll just try again later. */ 697ef5cb973SHerbert Xu 6983ab224beSHideo Aoki sk->sk_wmem_queued += buff->truesize; 6993ab224beSHideo Aoki sk_mem_charge(sk, buff->truesize); 700b60b49eaSHerbert Xu nlen = skb->len - len - nsize; 701b60b49eaSHerbert Xu buff->truesize += nlen; 702b60b49eaSHerbert Xu skb->truesize -= nlen; 7031da177e4SLinus Torvalds 7041da177e4SLinus Torvalds /* Correct the sequence numbers. */ 7051da177e4SLinus Torvalds TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; 7061da177e4SLinus Torvalds TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; 7071da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; 7081da177e4SLinus Torvalds 7091da177e4SLinus Torvalds /* PSH and FIN should only be set in the second packet. */ 7101da177e4SLinus Torvalds flags = TCP_SKB_CB(skb)->flags; 7111da177e4SLinus Torvalds TCP_SKB_CB(skb)->flags = flags & ~(TCPCB_FLAG_FIN|TCPCB_FLAG_PSH); 7121da177e4SLinus Torvalds TCP_SKB_CB(buff)->flags = flags; 713e14c3cafSHerbert Xu TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked; 7141da177e4SLinus Torvalds 71584fa7933SPatrick McHardy if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) { 7161da177e4SLinus Torvalds /* Copy and checksum data tail into the new buffer. */ 7171da177e4SLinus Torvalds buff->csum = csum_partial_copy_nocheck(skb->data + len, skb_put(buff, nsize), 7181da177e4SLinus Torvalds nsize, 0); 7191da177e4SLinus Torvalds 7201da177e4SLinus Torvalds skb_trim(skb, len); 7211da177e4SLinus Torvalds 7221da177e4SLinus Torvalds skb->csum = csum_block_sub(skb->csum, buff->csum, len); 7231da177e4SLinus Torvalds } else { 72484fa7933SPatrick McHardy skb->ip_summed = CHECKSUM_PARTIAL; 7251da177e4SLinus Torvalds skb_split(skb, buff, len); 7261da177e4SLinus Torvalds } 7271da177e4SLinus Torvalds 7281da177e4SLinus Torvalds buff->ip_summed = skb->ip_summed; 7291da177e4SLinus Torvalds 7301da177e4SLinus Torvalds /* Looks stupid, but our code really uses when of 7311da177e4SLinus Torvalds * skbs, which it never sent before. --ANK 7321da177e4SLinus Torvalds */ 7331da177e4SLinus Torvalds TCP_SKB_CB(buff)->when = TCP_SKB_CB(skb)->when; 734a61bbcf2SPatrick McHardy buff->tstamp = skb->tstamp; 7351da177e4SLinus Torvalds 7366475be16SDavid S. Miller old_factor = tcp_skb_pcount(skb); 7376475be16SDavid S. Miller 7381da177e4SLinus Torvalds /* Fix up tso_factor for both original and new SKB. */ 739846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, skb, mss_now); 740846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, buff, mss_now); 7411da177e4SLinus Torvalds 7426475be16SDavid S. Miller /* If this packet has been sent out already, we must 7436475be16SDavid S. Miller * adjust the various packet counters. 7446475be16SDavid S. Miller */ 745cf0b450cSHerbert Xu if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) { 7466475be16SDavid S. Miller int diff = old_factor - tcp_skb_pcount(skb) - 7476475be16SDavid S. Miller tcp_skb_pcount(buff); 7481da177e4SLinus Torvalds 7496475be16SDavid S. Miller tp->packets_out -= diff; 750e14c3cafSHerbert Xu 751e14c3cafSHerbert Xu if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 752e14c3cafSHerbert Xu tp->sacked_out -= diff; 753e14c3cafSHerbert Xu if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) 754e14c3cafSHerbert Xu tp->retrans_out -= diff; 755e14c3cafSHerbert Xu 756b5860bbaSIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST) 7576475be16SDavid S. Miller tp->lost_out -= diff; 75883ca28beSHerbert Xu 75983ca28beSHerbert Xu /* Adjust Reno SACK estimate. */ 76091fed7a1SIlpo Järvinen if (tcp_is_reno(tp) && diff > 0) { 761af610b4cSIlpo Järvinen tcp_dec_pcount_approx_int(&tp->sacked_out, diff); 762005903bcSIlpo Järvinen tcp_verify_left_out(tp); 76383ca28beSHerbert Xu } 764a47e5a98SIlpo Järvinen tcp_adjust_fackets_out(sk, skb, diff); 7651da177e4SLinus Torvalds } 7661da177e4SLinus Torvalds 7671da177e4SLinus Torvalds /* Link BUFF into the send queue. */ 768f44b5271SDavid S. Miller skb_header_release(buff); 769fe067e8aSDavid S. Miller tcp_insert_write_queue_after(skb, buff, sk); 7701da177e4SLinus Torvalds 7711da177e4SLinus Torvalds return 0; 7721da177e4SLinus Torvalds } 7731da177e4SLinus Torvalds 7741da177e4SLinus Torvalds /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c 7751da177e4SLinus Torvalds * eventually). The difference is that pulled data not copied, but 7761da177e4SLinus Torvalds * immediately discarded. 7771da177e4SLinus Torvalds */ 778f2911969SHerbert Xu ~{PmVHI~} static void __pskb_trim_head(struct sk_buff *skb, int len) 7791da177e4SLinus Torvalds { 7801da177e4SLinus Torvalds int i, k, eat; 7811da177e4SLinus Torvalds 7821da177e4SLinus Torvalds eat = len; 7831da177e4SLinus Torvalds k = 0; 7841da177e4SLinus Torvalds for (i=0; i<skb_shinfo(skb)->nr_frags; i++) { 7851da177e4SLinus Torvalds if (skb_shinfo(skb)->frags[i].size <= eat) { 7861da177e4SLinus Torvalds put_page(skb_shinfo(skb)->frags[i].page); 7871da177e4SLinus Torvalds eat -= skb_shinfo(skb)->frags[i].size; 7881da177e4SLinus Torvalds } else { 7891da177e4SLinus Torvalds skb_shinfo(skb)->frags[k] = skb_shinfo(skb)->frags[i]; 7901da177e4SLinus Torvalds if (eat) { 7911da177e4SLinus Torvalds skb_shinfo(skb)->frags[k].page_offset += eat; 7921da177e4SLinus Torvalds skb_shinfo(skb)->frags[k].size -= eat; 7931da177e4SLinus Torvalds eat = 0; 7941da177e4SLinus Torvalds } 7951da177e4SLinus Torvalds k++; 7961da177e4SLinus Torvalds } 7971da177e4SLinus Torvalds } 7981da177e4SLinus Torvalds skb_shinfo(skb)->nr_frags = k; 7991da177e4SLinus Torvalds 80027a884dcSArnaldo Carvalho de Melo skb_reset_tail_pointer(skb); 8011da177e4SLinus Torvalds skb->data_len -= len; 8021da177e4SLinus Torvalds skb->len = skb->data_len; 8031da177e4SLinus Torvalds } 8041da177e4SLinus Torvalds 8051da177e4SLinus Torvalds int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len) 8061da177e4SLinus Torvalds { 8071da177e4SLinus Torvalds if (skb_cloned(skb) && 8081da177e4SLinus Torvalds pskb_expand_head(skb, 0, 0, GFP_ATOMIC)) 8091da177e4SLinus Torvalds return -ENOMEM; 8101da177e4SLinus Torvalds 811f2911969SHerbert Xu ~{PmVHI~} /* If len == headlen, we avoid __skb_pull to preserve alignment. */ 812f2911969SHerbert Xu ~{PmVHI~} if (unlikely(len < skb_headlen(skb))) 8131da177e4SLinus Torvalds __skb_pull(skb, len); 814f2911969SHerbert Xu ~{PmVHI~} else 815f2911969SHerbert Xu ~{PmVHI~} __pskb_trim_head(skb, len - skb_headlen(skb)); 8161da177e4SLinus Torvalds 8171da177e4SLinus Torvalds TCP_SKB_CB(skb)->seq += len; 81884fa7933SPatrick McHardy skb->ip_summed = CHECKSUM_PARTIAL; 8191da177e4SLinus Torvalds 8201da177e4SLinus Torvalds skb->truesize -= len; 8211da177e4SLinus Torvalds sk->sk_wmem_queued -= len; 8223ab224beSHideo Aoki sk_mem_uncharge(sk, len); 8231da177e4SLinus Torvalds sock_set_flag(sk, SOCK_QUEUE_SHRUNK); 8241da177e4SLinus Torvalds 8251da177e4SLinus Torvalds /* Any change of skb->len requires recalculation of tso 8261da177e4SLinus Torvalds * factor and mss. 8271da177e4SLinus Torvalds */ 8281da177e4SLinus Torvalds if (tcp_skb_pcount(skb) > 1) 829846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, skb, tcp_current_mss(sk, 1)); 8301da177e4SLinus Torvalds 8311da177e4SLinus Torvalds return 0; 8321da177e4SLinus Torvalds } 8331da177e4SLinus Torvalds 8345d424d5aSJohn Heffner /* Not accounting for SACKs here. */ 8355d424d5aSJohn Heffner int tcp_mtu_to_mss(struct sock *sk, int pmtu) 8365d424d5aSJohn Heffner { 8375d424d5aSJohn Heffner struct tcp_sock *tp = tcp_sk(sk); 8385d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 8395d424d5aSJohn Heffner int mss_now; 8405d424d5aSJohn Heffner 8415d424d5aSJohn Heffner /* Calculate base mss without TCP options: 8425d424d5aSJohn Heffner It is MMS_S - sizeof(tcphdr) of rfc1122 8435d424d5aSJohn Heffner */ 8445d424d5aSJohn Heffner mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr); 8455d424d5aSJohn Heffner 8465d424d5aSJohn Heffner /* Clamp it (mss_clamp does not include tcp options) */ 8475d424d5aSJohn Heffner if (mss_now > tp->rx_opt.mss_clamp) 8485d424d5aSJohn Heffner mss_now = tp->rx_opt.mss_clamp; 8495d424d5aSJohn Heffner 8505d424d5aSJohn Heffner /* Now subtract optional transport overhead */ 8515d424d5aSJohn Heffner mss_now -= icsk->icsk_ext_hdr_len; 8525d424d5aSJohn Heffner 8535d424d5aSJohn Heffner /* Then reserve room for full set of TCP options and 8 bytes of data */ 8545d424d5aSJohn Heffner if (mss_now < 48) 8555d424d5aSJohn Heffner mss_now = 48; 8565d424d5aSJohn Heffner 8575d424d5aSJohn Heffner /* Now subtract TCP options size, not including SACKs */ 8585d424d5aSJohn Heffner mss_now -= tp->tcp_header_len - sizeof(struct tcphdr); 8595d424d5aSJohn Heffner 8605d424d5aSJohn Heffner return mss_now; 8615d424d5aSJohn Heffner } 8625d424d5aSJohn Heffner 8635d424d5aSJohn Heffner /* Inverse of above */ 8645d424d5aSJohn Heffner int tcp_mss_to_mtu(struct sock *sk, int mss) 8655d424d5aSJohn Heffner { 8665d424d5aSJohn Heffner struct tcp_sock *tp = tcp_sk(sk); 8675d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 8685d424d5aSJohn Heffner int mtu; 8695d424d5aSJohn Heffner 8705d424d5aSJohn Heffner mtu = mss + 8715d424d5aSJohn Heffner tp->tcp_header_len + 8725d424d5aSJohn Heffner icsk->icsk_ext_hdr_len + 8735d424d5aSJohn Heffner icsk->icsk_af_ops->net_header_len; 8745d424d5aSJohn Heffner 8755d424d5aSJohn Heffner return mtu; 8765d424d5aSJohn Heffner } 8775d424d5aSJohn Heffner 8785d424d5aSJohn Heffner void tcp_mtup_init(struct sock *sk) 8795d424d5aSJohn Heffner { 8805d424d5aSJohn Heffner struct tcp_sock *tp = tcp_sk(sk); 8815d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 8825d424d5aSJohn Heffner 8835d424d5aSJohn Heffner icsk->icsk_mtup.enabled = sysctl_tcp_mtu_probing > 1; 8845d424d5aSJohn Heffner icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) + 8855d424d5aSJohn Heffner icsk->icsk_af_ops->net_header_len; 8865d424d5aSJohn Heffner icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, sysctl_tcp_base_mss); 8875d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = 0; 8885d424d5aSJohn Heffner } 8895d424d5aSJohn Heffner 8901da177e4SLinus Torvalds /* This function synchronize snd mss to current pmtu/exthdr set. 8911da177e4SLinus Torvalds 8921da177e4SLinus Torvalds tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts 8931da177e4SLinus Torvalds for TCP options, but includes only bare TCP header. 8941da177e4SLinus Torvalds 8951da177e4SLinus Torvalds tp->rx_opt.mss_clamp is mss negotiated at connection setup. 896caa20d9aSStephen Hemminger It is minimum of user_mss and mss received with SYN. 8971da177e4SLinus Torvalds It also does not include TCP options. 8981da177e4SLinus Torvalds 899d83d8461SArnaldo Carvalho de Melo inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function. 9001da177e4SLinus Torvalds 9011da177e4SLinus Torvalds tp->mss_cache is current effective sending mss, including 9021da177e4SLinus Torvalds all tcp options except for SACKs. It is evaluated, 9031da177e4SLinus Torvalds taking into account current pmtu, but never exceeds 9041da177e4SLinus Torvalds tp->rx_opt.mss_clamp. 9051da177e4SLinus Torvalds 9061da177e4SLinus Torvalds NOTE1. rfc1122 clearly states that advertised MSS 9071da177e4SLinus Torvalds DOES NOT include either tcp or ip options. 9081da177e4SLinus Torvalds 909d83d8461SArnaldo Carvalho de Melo NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache 910d83d8461SArnaldo Carvalho de Melo are READ ONLY outside this function. --ANK (980731) 9111da177e4SLinus Torvalds */ 9121da177e4SLinus Torvalds 9131da177e4SLinus Torvalds unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu) 9141da177e4SLinus Torvalds { 9151da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 916d83d8461SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 9175d424d5aSJohn Heffner int mss_now; 9181da177e4SLinus Torvalds 9195d424d5aSJohn Heffner if (icsk->icsk_mtup.search_high > pmtu) 9205d424d5aSJohn Heffner icsk->icsk_mtup.search_high = pmtu; 9211da177e4SLinus Torvalds 9225d424d5aSJohn Heffner mss_now = tcp_mtu_to_mss(sk, pmtu); 9231da177e4SLinus Torvalds 9241da177e4SLinus Torvalds /* Bound mss with half of window */ 9251da177e4SLinus Torvalds if (tp->max_window && mss_now > (tp->max_window>>1)) 9261da177e4SLinus Torvalds mss_now = max((tp->max_window>>1), 68U - tp->tcp_header_len); 9271da177e4SLinus Torvalds 9281da177e4SLinus Torvalds /* And store cached results */ 929d83d8461SArnaldo Carvalho de Melo icsk->icsk_pmtu_cookie = pmtu; 9305d424d5aSJohn Heffner if (icsk->icsk_mtup.enabled) 9315d424d5aSJohn Heffner mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low)); 932c1b4a7e6SDavid S. Miller tp->mss_cache = mss_now; 9331da177e4SLinus Torvalds 9341da177e4SLinus Torvalds return mss_now; 9351da177e4SLinus Torvalds } 9361da177e4SLinus Torvalds 9371da177e4SLinus Torvalds /* Compute the current effective MSS, taking SACKs and IP options, 9381da177e4SLinus Torvalds * and even PMTU discovery events into account. 9391da177e4SLinus Torvalds * 9401da177e4SLinus Torvalds * LARGESEND note: !urg_mode is overkill, only frames up to snd_up 9411da177e4SLinus Torvalds * cannot be large. However, taking into account rare use of URG, this 9421da177e4SLinus Torvalds * is not a big flaw. 9431da177e4SLinus Torvalds */ 944c1b4a7e6SDavid S. Miller unsigned int tcp_current_mss(struct sock *sk, int large_allowed) 9451da177e4SLinus Torvalds { 9461da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 9471da177e4SLinus Torvalds struct dst_entry *dst = __sk_dst_get(sk); 948c1b4a7e6SDavid S. Miller u32 mss_now; 949c1b4a7e6SDavid S. Miller u16 xmit_size_goal; 950c1b4a7e6SDavid S. Miller int doing_tso = 0; 9511da177e4SLinus Torvalds 952c1b4a7e6SDavid S. Miller mss_now = tp->mss_cache; 953c1b4a7e6SDavid S. Miller 954bcd76111SHerbert Xu if (large_allowed && sk_can_gso(sk) && !tp->urg_mode) 955c1b4a7e6SDavid S. Miller doing_tso = 1; 956c1b4a7e6SDavid S. Miller 9571da177e4SLinus Torvalds if (dst) { 9581da177e4SLinus Torvalds u32 mtu = dst_mtu(dst); 959d83d8461SArnaldo Carvalho de Melo if (mtu != inet_csk(sk)->icsk_pmtu_cookie) 9601da177e4SLinus Torvalds mss_now = tcp_sync_mss(sk, mtu); 9611da177e4SLinus Torvalds } 9621da177e4SLinus Torvalds 9631da177e4SLinus Torvalds if (tp->rx_opt.eff_sacks) 9641da177e4SLinus Torvalds mss_now -= (TCPOLEN_SACK_BASE_ALIGNED + 9651da177e4SLinus Torvalds (tp->rx_opt.eff_sacks * TCPOLEN_SACK_PERBLOCK)); 966c1b4a7e6SDavid S. Miller 967cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 968cfb6eeb4SYOSHIFUJI Hideaki if (tp->af_specific->md5_lookup(sk, sk)) 969cfb6eeb4SYOSHIFUJI Hideaki mss_now -= TCPOLEN_MD5SIG_ALIGNED; 970cfb6eeb4SYOSHIFUJI Hideaki #endif 971cfb6eeb4SYOSHIFUJI Hideaki 972c1b4a7e6SDavid S. Miller xmit_size_goal = mss_now; 973c1b4a7e6SDavid S. Miller 974c1b4a7e6SDavid S. Miller if (doing_tso) { 9758292a17aSArnaldo Carvalho de Melo xmit_size_goal = (65535 - 9768292a17aSArnaldo Carvalho de Melo inet_csk(sk)->icsk_af_ops->net_header_len - 977d83d8461SArnaldo Carvalho de Melo inet_csk(sk)->icsk_ext_hdr_len - 978d83d8461SArnaldo Carvalho de Melo tp->tcp_header_len); 979c1b4a7e6SDavid S. Miller 980c1b4a7e6SDavid S. Miller if (tp->max_window && 981c1b4a7e6SDavid S. Miller (xmit_size_goal > (tp->max_window >> 1))) 982c1b4a7e6SDavid S. Miller xmit_size_goal = max((tp->max_window >> 1), 983c1b4a7e6SDavid S. Miller 68U - tp->tcp_header_len); 984c1b4a7e6SDavid S. Miller 985c1b4a7e6SDavid S. Miller xmit_size_goal -= (xmit_size_goal % mss_now); 986c1b4a7e6SDavid S. Miller } 987c1b4a7e6SDavid S. Miller tp->xmit_size_goal = xmit_size_goal; 988c1b4a7e6SDavid S. Miller 9891da177e4SLinus Torvalds return mss_now; 9901da177e4SLinus Torvalds } 9911da177e4SLinus Torvalds 992a762a980SDavid S. Miller /* Congestion window validation. (RFC2861) */ 993a762a980SDavid S. Miller 9949e412ba7SIlpo Järvinen static void tcp_cwnd_validate(struct sock *sk) 995a762a980SDavid S. Miller { 9969e412ba7SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 997a762a980SDavid S. Miller __u32 packets_out = tp->packets_out; 998a762a980SDavid S. Miller 999a762a980SDavid S. Miller if (packets_out >= tp->snd_cwnd) { 1000a762a980SDavid S. Miller /* Network is feed fully. */ 1001a762a980SDavid S. Miller tp->snd_cwnd_used = 0; 1002a762a980SDavid S. Miller tp->snd_cwnd_stamp = tcp_time_stamp; 1003a762a980SDavid S. Miller } else { 1004a762a980SDavid S. Miller /* Network starves. */ 1005a762a980SDavid S. Miller if (tp->packets_out > tp->snd_cwnd_used) 1006a762a980SDavid S. Miller tp->snd_cwnd_used = tp->packets_out; 1007a762a980SDavid S. Miller 100815d33c07SDavid S. Miller if (sysctl_tcp_slow_start_after_idle && 100915d33c07SDavid S. Miller (s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto) 1010a762a980SDavid S. Miller tcp_cwnd_application_limited(sk); 1011a762a980SDavid S. Miller } 1012a762a980SDavid S. Miller } 1013a762a980SDavid S. Miller 10140e3a4803SIlpo Järvinen /* Returns the portion of skb which can be sent right away without 10150e3a4803SIlpo Järvinen * introducing MSS oddities to segment boundaries. In rare cases where 10160e3a4803SIlpo Järvinen * mss_now != mss_cache, we will request caller to create a small skb 10170e3a4803SIlpo Järvinen * per input skb which could be mostly avoided here (if desired). 10180e3a4803SIlpo Järvinen */ 10190e3a4803SIlpo Järvinen static unsigned int tcp_mss_split_point(struct sock *sk, struct sk_buff *skb, 10200e3a4803SIlpo Järvinen unsigned int mss_now, 10210e3a4803SIlpo Järvinen unsigned int cwnd) 1022c1b4a7e6SDavid S. Miller { 10230e3a4803SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 10240e3a4803SIlpo Järvinen u32 needed, window, cwnd_len; 1025c1b4a7e6SDavid S. Miller 102690840defSIlpo Järvinen window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 1027c1b4a7e6SDavid S. Miller cwnd_len = mss_now * cwnd; 10280e3a4803SIlpo Järvinen 10290e3a4803SIlpo Järvinen if (likely(cwnd_len <= window && skb != tcp_write_queue_tail(sk))) 10300e3a4803SIlpo Järvinen return cwnd_len; 10310e3a4803SIlpo Järvinen 10320e3a4803SIlpo Järvinen if (skb == tcp_write_queue_tail(sk) && cwnd_len <= skb->len) 10330e3a4803SIlpo Järvinen return cwnd_len; 10340e3a4803SIlpo Järvinen 10350e3a4803SIlpo Järvinen needed = min(skb->len, window); 10360e3a4803SIlpo Järvinen return needed - needed % mss_now; 1037c1b4a7e6SDavid S. Miller } 1038c1b4a7e6SDavid S. Miller 1039c1b4a7e6SDavid S. Miller /* Can at least one segment of SKB be sent right now, according to the 1040c1b4a7e6SDavid S. Miller * congestion window rules? If so, return how many segments are allowed. 1041c1b4a7e6SDavid S. Miller */ 1042c1b4a7e6SDavid S. Miller static inline unsigned int tcp_cwnd_test(struct tcp_sock *tp, struct sk_buff *skb) 1043c1b4a7e6SDavid S. Miller { 1044c1b4a7e6SDavid S. Miller u32 in_flight, cwnd; 1045c1b4a7e6SDavid S. Miller 1046c1b4a7e6SDavid S. Miller /* Don't be strict about the congestion window for the final FIN. */ 1047104439a8SJohn Heffner if ((TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN) && 1048104439a8SJohn Heffner tcp_skb_pcount(skb) == 1) 1049c1b4a7e6SDavid S. Miller return 1; 1050c1b4a7e6SDavid S. Miller 1051c1b4a7e6SDavid S. Miller in_flight = tcp_packets_in_flight(tp); 1052c1b4a7e6SDavid S. Miller cwnd = tp->snd_cwnd; 1053c1b4a7e6SDavid S. Miller if (in_flight < cwnd) 1054c1b4a7e6SDavid S. Miller return (cwnd - in_flight); 1055c1b4a7e6SDavid S. Miller 1056c1b4a7e6SDavid S. Miller return 0; 1057c1b4a7e6SDavid S. Miller } 1058c1b4a7e6SDavid S. Miller 1059c1b4a7e6SDavid S. Miller /* This must be invoked the first time we consider transmitting 1060c1b4a7e6SDavid S. Miller * SKB onto the wire. 1061c1b4a7e6SDavid S. Miller */ 106240efc6faSStephen Hemminger static int tcp_init_tso_segs(struct sock *sk, struct sk_buff *skb, unsigned int mss_now) 1063c1b4a7e6SDavid S. Miller { 1064c1b4a7e6SDavid S. Miller int tso_segs = tcp_skb_pcount(skb); 1065c1b4a7e6SDavid S. Miller 1066846998aeSDavid S. Miller if (!tso_segs || 1067846998aeSDavid S. Miller (tso_segs > 1 && 10687967168cSHerbert Xu tcp_skb_mss(skb) != mss_now)) { 1069846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, skb, mss_now); 1070c1b4a7e6SDavid S. Miller tso_segs = tcp_skb_pcount(skb); 1071c1b4a7e6SDavid S. Miller } 1072c1b4a7e6SDavid S. Miller return tso_segs; 1073c1b4a7e6SDavid S. Miller } 1074c1b4a7e6SDavid S. Miller 1075c1b4a7e6SDavid S. Miller static inline int tcp_minshall_check(const struct tcp_sock *tp) 1076c1b4a7e6SDavid S. Miller { 1077c1b4a7e6SDavid S. Miller return after(tp->snd_sml,tp->snd_una) && 1078c1b4a7e6SDavid S. Miller !after(tp->snd_sml, tp->snd_nxt); 1079c1b4a7e6SDavid S. Miller } 1080c1b4a7e6SDavid S. Miller 1081c1b4a7e6SDavid S. Miller /* Return 0, if packet can be sent now without violation Nagle's rules: 1082c1b4a7e6SDavid S. Miller * 1. It is full sized. 1083c1b4a7e6SDavid S. Miller * 2. Or it contains FIN. (already checked by caller) 1084c1b4a7e6SDavid S. Miller * 3. Or TCP_NODELAY was set. 1085c1b4a7e6SDavid S. Miller * 4. Or TCP_CORK is not set, and all sent packets are ACKed. 1086c1b4a7e6SDavid S. Miller * With Minshall's modification: all sent small packets are ACKed. 1087c1b4a7e6SDavid S. Miller */ 1088c1b4a7e6SDavid S. Miller 1089c1b4a7e6SDavid S. Miller static inline int tcp_nagle_check(const struct tcp_sock *tp, 1090c1b4a7e6SDavid S. Miller const struct sk_buff *skb, 1091c1b4a7e6SDavid S. Miller unsigned mss_now, int nonagle) 1092c1b4a7e6SDavid S. Miller { 1093c1b4a7e6SDavid S. Miller return (skb->len < mss_now && 1094c1b4a7e6SDavid S. Miller ((nonagle&TCP_NAGLE_CORK) || 1095c1b4a7e6SDavid S. Miller (!nonagle && 1096c1b4a7e6SDavid S. Miller tp->packets_out && 1097c1b4a7e6SDavid S. Miller tcp_minshall_check(tp)))); 1098c1b4a7e6SDavid S. Miller } 1099c1b4a7e6SDavid S. Miller 1100c1b4a7e6SDavid S. Miller /* Return non-zero if the Nagle test allows this packet to be 1101c1b4a7e6SDavid S. Miller * sent now. 1102c1b4a7e6SDavid S. Miller */ 1103c1b4a7e6SDavid S. Miller static inline int tcp_nagle_test(struct tcp_sock *tp, struct sk_buff *skb, 1104c1b4a7e6SDavid S. Miller unsigned int cur_mss, int nonagle) 1105c1b4a7e6SDavid S. Miller { 1106c1b4a7e6SDavid S. Miller /* Nagle rule does not apply to frames, which sit in the middle of the 1107c1b4a7e6SDavid S. Miller * write_queue (they have no chances to get new data). 1108c1b4a7e6SDavid S. Miller * 1109c1b4a7e6SDavid S. Miller * This is implemented in the callers, where they modify the 'nonagle' 1110c1b4a7e6SDavid S. Miller * argument based upon the location of SKB in the send queue. 1111c1b4a7e6SDavid S. Miller */ 1112c1b4a7e6SDavid S. Miller if (nonagle & TCP_NAGLE_PUSH) 1113c1b4a7e6SDavid S. Miller return 1; 1114c1b4a7e6SDavid S. Miller 1115d551e454SIlpo Järvinen /* Don't use the nagle rule for urgent data (or for the final FIN). 1116d551e454SIlpo Järvinen * Nagle can be ignored during F-RTO too (see RFC4138). 1117d551e454SIlpo Järvinen */ 1118d551e454SIlpo Järvinen if (tp->urg_mode || (tp->frto_counter == 2) || 1119c1b4a7e6SDavid S. Miller (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN)) 1120c1b4a7e6SDavid S. Miller return 1; 1121c1b4a7e6SDavid S. Miller 1122c1b4a7e6SDavid S. Miller if (!tcp_nagle_check(tp, skb, cur_mss, nonagle)) 1123c1b4a7e6SDavid S. Miller return 1; 1124c1b4a7e6SDavid S. Miller 1125c1b4a7e6SDavid S. Miller return 0; 1126c1b4a7e6SDavid S. Miller } 1127c1b4a7e6SDavid S. Miller 1128c1b4a7e6SDavid S. Miller /* Does at least the first segment of SKB fit into the send window? */ 1129c1b4a7e6SDavid S. Miller static inline int tcp_snd_wnd_test(struct tcp_sock *tp, struct sk_buff *skb, unsigned int cur_mss) 1130c1b4a7e6SDavid S. Miller { 1131c1b4a7e6SDavid S. Miller u32 end_seq = TCP_SKB_CB(skb)->end_seq; 1132c1b4a7e6SDavid S. Miller 1133c1b4a7e6SDavid S. Miller if (skb->len > cur_mss) 1134c1b4a7e6SDavid S. Miller end_seq = TCP_SKB_CB(skb)->seq + cur_mss; 1135c1b4a7e6SDavid S. Miller 113690840defSIlpo Järvinen return !after(end_seq, tcp_wnd_end(tp)); 1137c1b4a7e6SDavid S. Miller } 1138c1b4a7e6SDavid S. Miller 1139fe067e8aSDavid S. Miller /* This checks if the data bearing packet SKB (usually tcp_send_head(sk)) 1140c1b4a7e6SDavid S. Miller * should be put on the wire right now. If so, it returns the number of 1141c1b4a7e6SDavid S. Miller * packets allowed by the congestion window. 1142c1b4a7e6SDavid S. Miller */ 1143c1b4a7e6SDavid S. Miller static unsigned int tcp_snd_test(struct sock *sk, struct sk_buff *skb, 1144c1b4a7e6SDavid S. Miller unsigned int cur_mss, int nonagle) 1145c1b4a7e6SDavid S. Miller { 1146c1b4a7e6SDavid S. Miller struct tcp_sock *tp = tcp_sk(sk); 1147c1b4a7e6SDavid S. Miller unsigned int cwnd_quota; 1148c1b4a7e6SDavid S. Miller 1149846998aeSDavid S. Miller tcp_init_tso_segs(sk, skb, cur_mss); 1150c1b4a7e6SDavid S. Miller 1151c1b4a7e6SDavid S. Miller if (!tcp_nagle_test(tp, skb, cur_mss, nonagle)) 1152c1b4a7e6SDavid S. Miller return 0; 1153c1b4a7e6SDavid S. Miller 1154c1b4a7e6SDavid S. Miller cwnd_quota = tcp_cwnd_test(tp, skb); 1155c1b4a7e6SDavid S. Miller if (cwnd_quota && 1156c1b4a7e6SDavid S. Miller !tcp_snd_wnd_test(tp, skb, cur_mss)) 1157c1b4a7e6SDavid S. Miller cwnd_quota = 0; 1158c1b4a7e6SDavid S. Miller 1159c1b4a7e6SDavid S. Miller return cwnd_quota; 1160c1b4a7e6SDavid S. Miller } 1161c1b4a7e6SDavid S. Miller 11629e412ba7SIlpo Järvinen int tcp_may_send_now(struct sock *sk) 1163c1b4a7e6SDavid S. Miller { 11649e412ba7SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 1165fe067e8aSDavid S. Miller struct sk_buff *skb = tcp_send_head(sk); 1166c1b4a7e6SDavid S. Miller 1167c1b4a7e6SDavid S. Miller return (skb && 1168c1b4a7e6SDavid S. Miller tcp_snd_test(sk, skb, tcp_current_mss(sk, 1), 1169c1b4a7e6SDavid S. Miller (tcp_skb_is_last(sk, skb) ? 11704e67d876SIlpo Järvinen tp->nonagle : TCP_NAGLE_PUSH))); 1171c1b4a7e6SDavid S. Miller } 1172c1b4a7e6SDavid S. Miller 1173c1b4a7e6SDavid S. Miller /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet 1174c1b4a7e6SDavid S. Miller * which is put after SKB on the list. It is very much like 1175c1b4a7e6SDavid S. Miller * tcp_fragment() except that it may make several kinds of assumptions 1176c1b4a7e6SDavid S. Miller * in order to speed up the splitting operation. In particular, we 1177c1b4a7e6SDavid S. Miller * know that all the data is in scatter-gather pages, and that the 1178c1b4a7e6SDavid S. Miller * packet has never been sent out before (and thus is not cloned). 1179c1b4a7e6SDavid S. Miller */ 1180846998aeSDavid S. Miller static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len, unsigned int mss_now) 1181c1b4a7e6SDavid S. Miller { 1182c1b4a7e6SDavid S. Miller struct sk_buff *buff; 1183c1b4a7e6SDavid S. Miller int nlen = skb->len - len; 1184c1b4a7e6SDavid S. Miller u16 flags; 1185c1b4a7e6SDavid S. Miller 1186c1b4a7e6SDavid S. Miller /* All of a TSO frame must be composed of paged data. */ 1187c8ac3774SHerbert Xu if (skb->len != skb->data_len) 1188c8ac3774SHerbert Xu return tcp_fragment(sk, skb, len, mss_now); 1189c1b4a7e6SDavid S. Miller 1190df97c708SPavel Emelyanov buff = sk_stream_alloc_skb(sk, 0, GFP_ATOMIC); 1191c1b4a7e6SDavid S. Miller if (unlikely(buff == NULL)) 1192c1b4a7e6SDavid S. Miller return -ENOMEM; 1193c1b4a7e6SDavid S. Miller 11943ab224beSHideo Aoki sk->sk_wmem_queued += buff->truesize; 11953ab224beSHideo Aoki sk_mem_charge(sk, buff->truesize); 1196b60b49eaSHerbert Xu buff->truesize += nlen; 1197c1b4a7e6SDavid S. Miller skb->truesize -= nlen; 1198c1b4a7e6SDavid S. Miller 1199c1b4a7e6SDavid S. Miller /* Correct the sequence numbers. */ 1200c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; 1201c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; 1202c1b4a7e6SDavid S. Miller TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; 1203c1b4a7e6SDavid S. Miller 1204c1b4a7e6SDavid S. Miller /* PSH and FIN should only be set in the second packet. */ 1205c1b4a7e6SDavid S. Miller flags = TCP_SKB_CB(skb)->flags; 1206c1b4a7e6SDavid S. Miller TCP_SKB_CB(skb)->flags = flags & ~(TCPCB_FLAG_FIN|TCPCB_FLAG_PSH); 1207c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->flags = flags; 1208c1b4a7e6SDavid S. Miller 1209c1b4a7e6SDavid S. Miller /* This packet was never sent out yet, so no SACK bits. */ 1210c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->sacked = 0; 1211c1b4a7e6SDavid S. Miller 121284fa7933SPatrick McHardy buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL; 1213c1b4a7e6SDavid S. Miller skb_split(skb, buff, len); 1214c1b4a7e6SDavid S. Miller 1215c1b4a7e6SDavid S. Miller /* Fix up tso_factor for both original and new SKB. */ 1216846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, skb, mss_now); 1217846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, buff, mss_now); 1218c1b4a7e6SDavid S. Miller 1219c1b4a7e6SDavid S. Miller /* Link BUFF into the send queue. */ 1220c1b4a7e6SDavid S. Miller skb_header_release(buff); 1221fe067e8aSDavid S. Miller tcp_insert_write_queue_after(skb, buff, sk); 1222c1b4a7e6SDavid S. Miller 1223c1b4a7e6SDavid S. Miller return 0; 1224c1b4a7e6SDavid S. Miller } 1225c1b4a7e6SDavid S. Miller 1226c1b4a7e6SDavid S. Miller /* Try to defer sending, if possible, in order to minimize the amount 1227c1b4a7e6SDavid S. Miller * of TSO splitting we do. View it as a kind of TSO Nagle test. 1228c1b4a7e6SDavid S. Miller * 1229c1b4a7e6SDavid S. Miller * This algorithm is from John Heffner. 1230c1b4a7e6SDavid S. Miller */ 12319e412ba7SIlpo Järvinen static int tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb) 1232c1b4a7e6SDavid S. Miller { 12339e412ba7SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 12346687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 1235c1b4a7e6SDavid S. Miller u32 send_win, cong_win, limit, in_flight; 1236c1b4a7e6SDavid S. Miller 1237c1b4a7e6SDavid S. Miller if (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN) 1238ae8064acSJohn Heffner goto send_now; 1239c1b4a7e6SDavid S. Miller 12406687e988SArnaldo Carvalho de Melo if (icsk->icsk_ca_state != TCP_CA_Open) 1241ae8064acSJohn Heffner goto send_now; 1242ae8064acSJohn Heffner 1243ae8064acSJohn Heffner /* Defer for less than two clock ticks. */ 1244bd515c3eSIlpo Järvinen if (tp->tso_deferred && 1245bd515c3eSIlpo Järvinen ((jiffies << 1) >> 1) - (tp->tso_deferred >> 1) > 1) 1246ae8064acSJohn Heffner goto send_now; 1247908a75c1SDavid S. Miller 1248c1b4a7e6SDavid S. Miller in_flight = tcp_packets_in_flight(tp); 1249c1b4a7e6SDavid S. Miller 1250c1b4a7e6SDavid S. Miller BUG_ON(tcp_skb_pcount(skb) <= 1 || 1251c1b4a7e6SDavid S. Miller (tp->snd_cwnd <= in_flight)); 1252c1b4a7e6SDavid S. Miller 125390840defSIlpo Järvinen send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 1254c1b4a7e6SDavid S. Miller 1255c1b4a7e6SDavid S. Miller /* From in_flight test above, we know that cwnd > in_flight. */ 1256c1b4a7e6SDavid S. Miller cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache; 1257c1b4a7e6SDavid S. Miller 1258c1b4a7e6SDavid S. Miller limit = min(send_win, cong_win); 1259c1b4a7e6SDavid S. Miller 1260ba244fe9SDavid S. Miller /* If a full-sized TSO skb can be sent, do it. */ 1261ba244fe9SDavid S. Miller if (limit >= 65536) 1262ae8064acSJohn Heffner goto send_now; 1263ba244fe9SDavid S. Miller 1264c1b4a7e6SDavid S. Miller if (sysctl_tcp_tso_win_divisor) { 1265c1b4a7e6SDavid S. Miller u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache); 1266c1b4a7e6SDavid S. Miller 1267c1b4a7e6SDavid S. Miller /* If at least some fraction of a window is available, 1268c1b4a7e6SDavid S. Miller * just use it. 1269c1b4a7e6SDavid S. Miller */ 1270c1b4a7e6SDavid S. Miller chunk /= sysctl_tcp_tso_win_divisor; 1271c1b4a7e6SDavid S. Miller if (limit >= chunk) 1272ae8064acSJohn Heffner goto send_now; 1273c1b4a7e6SDavid S. Miller } else { 1274c1b4a7e6SDavid S. Miller /* Different approach, try not to defer past a single 1275c1b4a7e6SDavid S. Miller * ACK. Receiver should ACK every other full sized 1276c1b4a7e6SDavid S. Miller * frame, so if we have space for more than 3 frames 1277c1b4a7e6SDavid S. Miller * then send now. 1278c1b4a7e6SDavid S. Miller */ 1279c1b4a7e6SDavid S. Miller if (limit > tcp_max_burst(tp) * tp->mss_cache) 1280ae8064acSJohn Heffner goto send_now; 1281c1b4a7e6SDavid S. Miller } 1282c1b4a7e6SDavid S. Miller 1283c1b4a7e6SDavid S. Miller /* Ok, it looks like it is advisable to defer. */ 1284ae8064acSJohn Heffner tp->tso_deferred = 1 | (jiffies<<1); 1285ae8064acSJohn Heffner 1286c1b4a7e6SDavid S. Miller return 1; 1287ae8064acSJohn Heffner 1288ae8064acSJohn Heffner send_now: 1289ae8064acSJohn Heffner tp->tso_deferred = 0; 1290ae8064acSJohn Heffner return 0; 1291c1b4a7e6SDavid S. Miller } 1292c1b4a7e6SDavid S. Miller 12935d424d5aSJohn Heffner /* Create a new MTU probe if we are ready. 12945d424d5aSJohn Heffner * Returns 0 if we should wait to probe (no cwnd available), 12955d424d5aSJohn Heffner * 1 if a probe was sent, 12965d424d5aSJohn Heffner * -1 otherwise */ 12975d424d5aSJohn Heffner static int tcp_mtu_probe(struct sock *sk) 12985d424d5aSJohn Heffner { 12995d424d5aSJohn Heffner struct tcp_sock *tp = tcp_sk(sk); 13005d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 13015d424d5aSJohn Heffner struct sk_buff *skb, *nskb, *next; 13025d424d5aSJohn Heffner int len; 13035d424d5aSJohn Heffner int probe_size; 130491cc17c0SIlpo Järvinen int size_needed; 13055d424d5aSJohn Heffner int copy; 13065d424d5aSJohn Heffner int mss_now; 13075d424d5aSJohn Heffner 13085d424d5aSJohn Heffner /* Not currently probing/verifying, 13095d424d5aSJohn Heffner * not in recovery, 13105d424d5aSJohn Heffner * have enough cwnd, and 13115d424d5aSJohn Heffner * not SACKing (the variable headers throw things off) */ 13125d424d5aSJohn Heffner if (!icsk->icsk_mtup.enabled || 13135d424d5aSJohn Heffner icsk->icsk_mtup.probe_size || 13145d424d5aSJohn Heffner inet_csk(sk)->icsk_ca_state != TCP_CA_Open || 13155d424d5aSJohn Heffner tp->snd_cwnd < 11 || 13165d424d5aSJohn Heffner tp->rx_opt.eff_sacks) 13175d424d5aSJohn Heffner return -1; 13185d424d5aSJohn Heffner 13195d424d5aSJohn Heffner /* Very simple search strategy: just double the MSS. */ 13205d424d5aSJohn Heffner mss_now = tcp_current_mss(sk, 0); 13215d424d5aSJohn Heffner probe_size = 2*tp->mss_cache; 132291cc17c0SIlpo Järvinen size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache; 13235d424d5aSJohn Heffner if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high)) { 13245d424d5aSJohn Heffner /* TODO: set timer for probe_converge_event */ 13255d424d5aSJohn Heffner return -1; 13265d424d5aSJohn Heffner } 13275d424d5aSJohn Heffner 13285d424d5aSJohn Heffner /* Have enough data in the send queue to probe? */ 13297f9c33e5SIlpo Järvinen if (tp->write_seq - tp->snd_nxt < size_needed) 13305d424d5aSJohn Heffner return -1; 13315d424d5aSJohn Heffner 133291cc17c0SIlpo Järvinen if (tp->snd_wnd < size_needed) 13335d424d5aSJohn Heffner return -1; 133490840defSIlpo Järvinen if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp))) 13355d424d5aSJohn Heffner return 0; 13365d424d5aSJohn Heffner 1337d67c58e9SIlpo Järvinen /* Do we need to wait to drain cwnd? With none in flight, don't stall */ 1338d67c58e9SIlpo Järvinen if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) { 1339d67c58e9SIlpo Järvinen if (!tcp_packets_in_flight(tp)) 13405d424d5aSJohn Heffner return -1; 13415d424d5aSJohn Heffner else 13425d424d5aSJohn Heffner return 0; 13435d424d5aSJohn Heffner } 13445d424d5aSJohn Heffner 13455d424d5aSJohn Heffner /* We're allowed to probe. Build it now. */ 13465d424d5aSJohn Heffner if ((nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC)) == NULL) 13475d424d5aSJohn Heffner return -1; 13483ab224beSHideo Aoki sk->sk_wmem_queued += nskb->truesize; 13493ab224beSHideo Aoki sk_mem_charge(sk, nskb->truesize); 13505d424d5aSJohn Heffner 1351fe067e8aSDavid S. Miller skb = tcp_send_head(sk); 13525d424d5aSJohn Heffner 13535d424d5aSJohn Heffner TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq; 13545d424d5aSJohn Heffner TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size; 13555d424d5aSJohn Heffner TCP_SKB_CB(nskb)->flags = TCPCB_FLAG_ACK; 13565d424d5aSJohn Heffner TCP_SKB_CB(nskb)->sacked = 0; 13575d424d5aSJohn Heffner nskb->csum = 0; 135884fa7933SPatrick McHardy nskb->ip_summed = skb->ip_summed; 13595d424d5aSJohn Heffner 136050c4817eSIlpo Järvinen tcp_insert_write_queue_before(nskb, skb, sk); 136150c4817eSIlpo Järvinen 13625d424d5aSJohn Heffner len = 0; 1363234b6860SIlpo Järvinen tcp_for_write_queue_from_safe(skb, next, sk) { 13645d424d5aSJohn Heffner copy = min_t(int, skb->len, probe_size - len); 13655d424d5aSJohn Heffner if (nskb->ip_summed) 13665d424d5aSJohn Heffner skb_copy_bits(skb, 0, skb_put(nskb, copy), copy); 13675d424d5aSJohn Heffner else 13685d424d5aSJohn Heffner nskb->csum = skb_copy_and_csum_bits(skb, 0, 13695d424d5aSJohn Heffner skb_put(nskb, copy), copy, nskb->csum); 13705d424d5aSJohn Heffner 13715d424d5aSJohn Heffner if (skb->len <= copy) { 13725d424d5aSJohn Heffner /* We've eaten all the data from this skb. 13735d424d5aSJohn Heffner * Throw it away. */ 13745d424d5aSJohn Heffner TCP_SKB_CB(nskb)->flags |= TCP_SKB_CB(skb)->flags; 1375fe067e8aSDavid S. Miller tcp_unlink_write_queue(skb, sk); 13763ab224beSHideo Aoki sk_wmem_free_skb(sk, skb); 13775d424d5aSJohn Heffner } else { 13785d424d5aSJohn Heffner TCP_SKB_CB(nskb)->flags |= TCP_SKB_CB(skb)->flags & 13795d424d5aSJohn Heffner ~(TCPCB_FLAG_FIN|TCPCB_FLAG_PSH); 13805d424d5aSJohn Heffner if (!skb_shinfo(skb)->nr_frags) { 13815d424d5aSJohn Heffner skb_pull(skb, copy); 138284fa7933SPatrick McHardy if (skb->ip_summed != CHECKSUM_PARTIAL) 13835d424d5aSJohn Heffner skb->csum = csum_partial(skb->data, skb->len, 0); 13845d424d5aSJohn Heffner } else { 13855d424d5aSJohn Heffner __pskb_trim_head(skb, copy); 13865d424d5aSJohn Heffner tcp_set_skb_tso_segs(sk, skb, mss_now); 13875d424d5aSJohn Heffner } 13885d424d5aSJohn Heffner TCP_SKB_CB(skb)->seq += copy; 13895d424d5aSJohn Heffner } 13905d424d5aSJohn Heffner 13915d424d5aSJohn Heffner len += copy; 1392234b6860SIlpo Järvinen 1393234b6860SIlpo Järvinen if (len >= probe_size) 1394234b6860SIlpo Järvinen break; 13955d424d5aSJohn Heffner } 13965d424d5aSJohn Heffner tcp_init_tso_segs(sk, nskb, nskb->len); 13975d424d5aSJohn Heffner 13985d424d5aSJohn Heffner /* We're ready to send. If this fails, the probe will 13995d424d5aSJohn Heffner * be resegmented into mss-sized pieces by tcp_write_xmit(). */ 14005d424d5aSJohn Heffner TCP_SKB_CB(nskb)->when = tcp_time_stamp; 14015d424d5aSJohn Heffner if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) { 14025d424d5aSJohn Heffner /* Decrement cwnd here because we are sending 14035d424d5aSJohn Heffner * effectively two packets. */ 14045d424d5aSJohn Heffner tp->snd_cwnd--; 140566f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, nskb); 14065d424d5aSJohn Heffner 14075d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len); 14080e7b1368SJohn Heffner tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq; 14090e7b1368SJohn Heffner tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq; 14105d424d5aSJohn Heffner 14115d424d5aSJohn Heffner return 1; 14125d424d5aSJohn Heffner } 14135d424d5aSJohn Heffner 14145d424d5aSJohn Heffner return -1; 14155d424d5aSJohn Heffner } 14165d424d5aSJohn Heffner 14175d424d5aSJohn Heffner 14181da177e4SLinus Torvalds /* This routine writes packets to the network. It advances the 14191da177e4SLinus Torvalds * send_head. This happens as incoming acks open up the remote 14201da177e4SLinus Torvalds * window for us. 14211da177e4SLinus Torvalds * 14221da177e4SLinus Torvalds * Returns 1, if no segments are in flight and we have queued segments, but 14231da177e4SLinus Torvalds * cannot send anything now because of SWS or another problem. 14241da177e4SLinus Torvalds */ 1425a2e2a59cSDavid S. Miller static int tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle) 14261da177e4SLinus Torvalds { 14271da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 142892df7b51SDavid S. Miller struct sk_buff *skb; 1429c1b4a7e6SDavid S. Miller unsigned int tso_segs, sent_pkts; 1430c1b4a7e6SDavid S. Miller int cwnd_quota; 14315d424d5aSJohn Heffner int result; 14321da177e4SLinus Torvalds 14331da177e4SLinus Torvalds /* If we are closed, the bytes will have to remain here. 14341da177e4SLinus Torvalds * In time closedown will finish, we empty the write queue and all 14351da177e4SLinus Torvalds * will be happy. 14361da177e4SLinus Torvalds */ 143792df7b51SDavid S. Miller if (unlikely(sk->sk_state == TCP_CLOSE)) 143892df7b51SDavid S. Miller return 0; 143992df7b51SDavid S. Miller 1440c1b4a7e6SDavid S. Miller sent_pkts = 0; 14415d424d5aSJohn Heffner 14425d424d5aSJohn Heffner /* Do MTU probing. */ 14435d424d5aSJohn Heffner if ((result = tcp_mtu_probe(sk)) == 0) { 14445d424d5aSJohn Heffner return 0; 14455d424d5aSJohn Heffner } else if (result > 0) { 14465d424d5aSJohn Heffner sent_pkts = 1; 14475d424d5aSJohn Heffner } 14485d424d5aSJohn Heffner 1449fe067e8aSDavid S. Miller while ((skb = tcp_send_head(sk))) { 1450c8ac3774SHerbert Xu unsigned int limit; 1451c8ac3774SHerbert Xu 1452b68e9f85SHerbert Xu tso_segs = tcp_init_tso_segs(sk, skb, mss_now); 1453c1b4a7e6SDavid S. Miller BUG_ON(!tso_segs); 1454c1b4a7e6SDavid S. Miller 1455b68e9f85SHerbert Xu cwnd_quota = tcp_cwnd_test(tp, skb); 1456b68e9f85SHerbert Xu if (!cwnd_quota) 1457b68e9f85SHerbert Xu break; 1458b68e9f85SHerbert Xu 1459b68e9f85SHerbert Xu if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now))) 1460b68e9f85SHerbert Xu break; 1461b68e9f85SHerbert Xu 1462c1b4a7e6SDavid S. Miller if (tso_segs == 1) { 1463aa93466bSDavid S. Miller if (unlikely(!tcp_nagle_test(tp, skb, mss_now, 1464aa93466bSDavid S. Miller (tcp_skb_is_last(sk, skb) ? 1465aa93466bSDavid S. Miller nonagle : TCP_NAGLE_PUSH)))) 1466aa93466bSDavid S. Miller break; 1467c1b4a7e6SDavid S. Miller } else { 14689e412ba7SIlpo Järvinen if (tcp_tso_should_defer(sk, skb)) 1469aa93466bSDavid S. Miller break; 1470c1b4a7e6SDavid S. Miller } 1471aa93466bSDavid S. Miller 1472c8ac3774SHerbert Xu limit = mss_now; 14730e3a4803SIlpo Järvinen if (tso_segs > 1) 14740e3a4803SIlpo Järvinen limit = tcp_mss_split_point(sk, skb, mss_now, 14750e3a4803SIlpo Järvinen cwnd_quota); 1476c8ac3774SHerbert Xu 1477c8ac3774SHerbert Xu if (skb->len > limit && 1478c8ac3774SHerbert Xu unlikely(tso_fragment(sk, skb, limit, mss_now))) 14791da177e4SLinus Torvalds break; 14801da177e4SLinus Torvalds 14811da177e4SLinus Torvalds TCP_SKB_CB(skb)->when = tcp_time_stamp; 1482c1b4a7e6SDavid S. Miller 1483dfb4b9dcSDavid S. Miller if (unlikely(tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC))) 14841da177e4SLinus Torvalds break; 14851da177e4SLinus Torvalds 14861da177e4SLinus Torvalds /* Advance the send_head. This one is sent out. 14871da177e4SLinus Torvalds * This call will increment packets_out. 14881da177e4SLinus Torvalds */ 148966f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, skb); 14901da177e4SLinus Torvalds 14911da177e4SLinus Torvalds tcp_minshall_update(tp, mss_now, skb); 1492aa93466bSDavid S. Miller sent_pkts++; 14931da177e4SLinus Torvalds } 14941da177e4SLinus Torvalds 1495aa93466bSDavid S. Miller if (likely(sent_pkts)) { 14969e412ba7SIlpo Järvinen tcp_cwnd_validate(sk); 14971da177e4SLinus Torvalds return 0; 14981da177e4SLinus Torvalds } 1499fe067e8aSDavid S. Miller return !tp->packets_out && tcp_send_head(sk); 15001da177e4SLinus Torvalds } 15011da177e4SLinus Torvalds 1502a762a980SDavid S. Miller /* Push out any pending frames which were held back due to 1503a762a980SDavid S. Miller * TCP_CORK or attempt at coalescing tiny packets. 1504a762a980SDavid S. Miller * The socket must be locked by the caller. 1505a762a980SDavid S. Miller */ 15069e412ba7SIlpo Järvinen void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss, 15079e412ba7SIlpo Järvinen int nonagle) 1508a762a980SDavid S. Miller { 1509fe067e8aSDavid S. Miller struct sk_buff *skb = tcp_send_head(sk); 1510a762a980SDavid S. Miller 1511a762a980SDavid S. Miller if (skb) { 151255c97f3eSDavid S. Miller if (tcp_write_xmit(sk, cur_mss, nonagle)) 15139e412ba7SIlpo Järvinen tcp_check_probe_timer(sk); 1514a762a980SDavid S. Miller } 1515a762a980SDavid S. Miller } 1516a762a980SDavid S. Miller 1517c1b4a7e6SDavid S. Miller /* Send _single_ skb sitting at the send head. This function requires 1518c1b4a7e6SDavid S. Miller * true push pending frames to setup probe timer etc. 1519c1b4a7e6SDavid S. Miller */ 1520c1b4a7e6SDavid S. Miller void tcp_push_one(struct sock *sk, unsigned int mss_now) 1521c1b4a7e6SDavid S. Miller { 1522fe067e8aSDavid S. Miller struct sk_buff *skb = tcp_send_head(sk); 1523c1b4a7e6SDavid S. Miller unsigned int tso_segs, cwnd_quota; 1524c1b4a7e6SDavid S. Miller 1525c1b4a7e6SDavid S. Miller BUG_ON(!skb || skb->len < mss_now); 1526c1b4a7e6SDavid S. Miller 1527846998aeSDavid S. Miller tso_segs = tcp_init_tso_segs(sk, skb, mss_now); 1528c1b4a7e6SDavid S. Miller cwnd_quota = tcp_snd_test(sk, skb, mss_now, TCP_NAGLE_PUSH); 1529c1b4a7e6SDavid S. Miller 1530c1b4a7e6SDavid S. Miller if (likely(cwnd_quota)) { 1531c8ac3774SHerbert Xu unsigned int limit; 1532c8ac3774SHerbert Xu 1533c1b4a7e6SDavid S. Miller BUG_ON(!tso_segs); 1534c1b4a7e6SDavid S. Miller 1535c8ac3774SHerbert Xu limit = mss_now; 15360e3a4803SIlpo Järvinen if (tso_segs > 1) 15370e3a4803SIlpo Järvinen limit = tcp_mss_split_point(sk, skb, mss_now, 15380e3a4803SIlpo Järvinen cwnd_quota); 1539c8ac3774SHerbert Xu 1540c8ac3774SHerbert Xu if (skb->len > limit && 1541c8ac3774SHerbert Xu unlikely(tso_fragment(sk, skb, limit, mss_now))) 1542c1b4a7e6SDavid S. Miller return; 1543c1b4a7e6SDavid S. Miller 1544c1b4a7e6SDavid S. Miller /* Send it out now. */ 1545c1b4a7e6SDavid S. Miller TCP_SKB_CB(skb)->when = tcp_time_stamp; 1546c1b4a7e6SDavid S. Miller 1547dfb4b9dcSDavid S. Miller if (likely(!tcp_transmit_skb(sk, skb, 1, sk->sk_allocation))) { 154866f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, skb); 15499e412ba7SIlpo Järvinen tcp_cwnd_validate(sk); 1550c1b4a7e6SDavid S. Miller return; 1551c1b4a7e6SDavid S. Miller } 1552c1b4a7e6SDavid S. Miller } 1553c1b4a7e6SDavid S. Miller } 1554c1b4a7e6SDavid S. Miller 15551da177e4SLinus Torvalds /* This function returns the amount that we can raise the 15561da177e4SLinus Torvalds * usable window based on the following constraints 15571da177e4SLinus Torvalds * 15581da177e4SLinus Torvalds * 1. The window can never be shrunk once it is offered (RFC 793) 15591da177e4SLinus Torvalds * 2. We limit memory per socket 15601da177e4SLinus Torvalds * 15611da177e4SLinus Torvalds * RFC 1122: 15621da177e4SLinus Torvalds * "the suggested [SWS] avoidance algorithm for the receiver is to keep 15631da177e4SLinus Torvalds * RECV.NEXT + RCV.WIN fixed until: 15641da177e4SLinus Torvalds * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)" 15651da177e4SLinus Torvalds * 15661da177e4SLinus Torvalds * i.e. don't raise the right edge of the window until you can raise 15671da177e4SLinus Torvalds * it at least MSS bytes. 15681da177e4SLinus Torvalds * 15691da177e4SLinus Torvalds * Unfortunately, the recommended algorithm breaks header prediction, 15701da177e4SLinus Torvalds * since header prediction assumes th->window stays fixed. 15711da177e4SLinus Torvalds * 15721da177e4SLinus Torvalds * Strictly speaking, keeping th->window fixed violates the receiver 15731da177e4SLinus Torvalds * side SWS prevention criteria. The problem is that under this rule 15741da177e4SLinus Torvalds * a stream of single byte packets will cause the right side of the 15751da177e4SLinus Torvalds * window to always advance by a single byte. 15761da177e4SLinus Torvalds * 15771da177e4SLinus Torvalds * Of course, if the sender implements sender side SWS prevention 15781da177e4SLinus Torvalds * then this will not be a problem. 15791da177e4SLinus Torvalds * 15801da177e4SLinus Torvalds * BSD seems to make the following compromise: 15811da177e4SLinus Torvalds * 15821da177e4SLinus Torvalds * If the free space is less than the 1/4 of the maximum 15831da177e4SLinus Torvalds * space available and the free space is less than 1/2 mss, 15841da177e4SLinus Torvalds * then set the window to 0. 15851da177e4SLinus Torvalds * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ] 15861da177e4SLinus Torvalds * Otherwise, just prevent the window from shrinking 15871da177e4SLinus Torvalds * and from being larger than the largest representable value. 15881da177e4SLinus Torvalds * 15891da177e4SLinus Torvalds * This prevents incremental opening of the window in the regime 15901da177e4SLinus Torvalds * where TCP is limited by the speed of the reader side taking 15911da177e4SLinus Torvalds * data out of the TCP receive queue. It does nothing about 15921da177e4SLinus Torvalds * those cases where the window is constrained on the sender side 15931da177e4SLinus Torvalds * because the pipeline is full. 15941da177e4SLinus Torvalds * 15951da177e4SLinus Torvalds * BSD also seems to "accidentally" limit itself to windows that are a 15961da177e4SLinus Torvalds * multiple of MSS, at least until the free space gets quite small. 15971da177e4SLinus Torvalds * This would appear to be a side effect of the mbuf implementation. 15981da177e4SLinus Torvalds * Combining these two algorithms results in the observed behavior 15991da177e4SLinus Torvalds * of having a fixed window size at almost all times. 16001da177e4SLinus Torvalds * 16011da177e4SLinus Torvalds * Below we obtain similar behavior by forcing the offered window to 16021da177e4SLinus Torvalds * a multiple of the mss when it is feasible to do so. 16031da177e4SLinus Torvalds * 16041da177e4SLinus Torvalds * Note, we don't "adjust" for TIMESTAMP or SACK option bytes. 16051da177e4SLinus Torvalds * Regular options like TIMESTAMP are taken into account. 16061da177e4SLinus Torvalds */ 16071da177e4SLinus Torvalds u32 __tcp_select_window(struct sock *sk) 16081da177e4SLinus Torvalds { 1609463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 16101da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 1611caa20d9aSStephen Hemminger /* MSS for the peer's data. Previous versions used mss_clamp 16121da177e4SLinus Torvalds * here. I don't know if the value based on our guesses 16131da177e4SLinus Torvalds * of peer's MSS is better for the performance. It's more correct 16141da177e4SLinus Torvalds * but may be worse for the performance because of rcv_mss 16151da177e4SLinus Torvalds * fluctuations. --SAW 1998/11/1 16161da177e4SLinus Torvalds */ 1617463c84b9SArnaldo Carvalho de Melo int mss = icsk->icsk_ack.rcv_mss; 16181da177e4SLinus Torvalds int free_space = tcp_space(sk); 16191da177e4SLinus Torvalds int full_space = min_t(int, tp->window_clamp, tcp_full_space(sk)); 16201da177e4SLinus Torvalds int window; 16211da177e4SLinus Torvalds 16221da177e4SLinus Torvalds if (mss > full_space) 16231da177e4SLinus Torvalds mss = full_space; 16241da177e4SLinus Torvalds 1625b92edbe0SEric Dumazet if (free_space < (full_space >> 1)) { 1626463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.quick = 0; 16271da177e4SLinus Torvalds 16281da177e4SLinus Torvalds if (tcp_memory_pressure) 16291da177e4SLinus Torvalds tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U*tp->advmss); 16301da177e4SLinus Torvalds 16311da177e4SLinus Torvalds if (free_space < mss) 16321da177e4SLinus Torvalds return 0; 16331da177e4SLinus Torvalds } 16341da177e4SLinus Torvalds 16351da177e4SLinus Torvalds if (free_space > tp->rcv_ssthresh) 16361da177e4SLinus Torvalds free_space = tp->rcv_ssthresh; 16371da177e4SLinus Torvalds 16381da177e4SLinus Torvalds /* Don't do rounding if we are using window scaling, since the 16391da177e4SLinus Torvalds * scaled window will not line up with the MSS boundary anyway. 16401da177e4SLinus Torvalds */ 16411da177e4SLinus Torvalds window = tp->rcv_wnd; 16421da177e4SLinus Torvalds if (tp->rx_opt.rcv_wscale) { 16431da177e4SLinus Torvalds window = free_space; 16441da177e4SLinus Torvalds 16451da177e4SLinus Torvalds /* Advertise enough space so that it won't get scaled away. 16461da177e4SLinus Torvalds * Import case: prevent zero window announcement if 16471da177e4SLinus Torvalds * 1<<rcv_wscale > mss. 16481da177e4SLinus Torvalds */ 16491da177e4SLinus Torvalds if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window) 16501da177e4SLinus Torvalds window = (((window >> tp->rx_opt.rcv_wscale) + 1) 16511da177e4SLinus Torvalds << tp->rx_opt.rcv_wscale); 16521da177e4SLinus Torvalds } else { 16531da177e4SLinus Torvalds /* Get the largest window that is a nice multiple of mss. 16541da177e4SLinus Torvalds * Window clamp already applied above. 16551da177e4SLinus Torvalds * If our current window offering is within 1 mss of the 16561da177e4SLinus Torvalds * free space we just keep it. This prevents the divide 16571da177e4SLinus Torvalds * and multiply from happening most of the time. 16581da177e4SLinus Torvalds * We also don't do any window rounding when the free space 16591da177e4SLinus Torvalds * is too small. 16601da177e4SLinus Torvalds */ 16611da177e4SLinus Torvalds if (window <= free_space - mss || window > free_space) 16621da177e4SLinus Torvalds window = (free_space/mss)*mss; 166384565070SJohn Heffner else if (mss == full_space && 1664b92edbe0SEric Dumazet free_space > window + (full_space >> 1)) 166584565070SJohn Heffner window = free_space; 16661da177e4SLinus Torvalds } 16671da177e4SLinus Torvalds 16681da177e4SLinus Torvalds return window; 16691da177e4SLinus Torvalds } 16701da177e4SLinus Torvalds 16711da177e4SLinus Torvalds /* Attempt to collapse two adjacent SKB's during retransmission. */ 16721da177e4SLinus Torvalds static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *skb, int mss_now) 16731da177e4SLinus Torvalds { 16741da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 1675fe067e8aSDavid S. Miller struct sk_buff *next_skb = tcp_write_queue_next(sk, skb); 16761da177e4SLinus Torvalds 16771da177e4SLinus Torvalds /* The first test we must make is that neither of these two 16781da177e4SLinus Torvalds * SKB's are still referenced by someone else. 16791da177e4SLinus Torvalds */ 16801da177e4SLinus Torvalds if (!skb_cloned(skb) && !skb_cloned(next_skb)) { 16811da177e4SLinus Torvalds int skb_size = skb->len, next_skb_size = next_skb->len; 16821da177e4SLinus Torvalds u16 flags = TCP_SKB_CB(skb)->flags; 16831da177e4SLinus Torvalds 16841da177e4SLinus Torvalds /* Also punt if next skb has been SACK'd. */ 16851da177e4SLinus Torvalds if (TCP_SKB_CB(next_skb)->sacked & TCPCB_SACKED_ACKED) 16861da177e4SLinus Torvalds return; 16871da177e4SLinus Torvalds 16881da177e4SLinus Torvalds /* Next skb is out of window. */ 168990840defSIlpo Järvinen if (after(TCP_SKB_CB(next_skb)->end_seq, tcp_wnd_end(tp))) 16901da177e4SLinus Torvalds return; 16911da177e4SLinus Torvalds 16921da177e4SLinus Torvalds /* Punt if not enough space exists in the first SKB for 16931da177e4SLinus Torvalds * the data in the second, or the total combined payload 16941da177e4SLinus Torvalds * would exceed the MSS. 16951da177e4SLinus Torvalds */ 16961da177e4SLinus Torvalds if ((next_skb_size > skb_tailroom(skb)) || 16971da177e4SLinus Torvalds ((skb_size + next_skb_size) > mss_now)) 16981da177e4SLinus Torvalds return; 16991da177e4SLinus Torvalds 17001da177e4SLinus Torvalds BUG_ON(tcp_skb_pcount(skb) != 1 || 17011da177e4SLinus Torvalds tcp_skb_pcount(next_skb) != 1); 17021da177e4SLinus Torvalds 17036859d494SIlpo Järvinen tcp_highest_sack_combine(sk, next_skb, skb); 1704a6963a6bSIlpo Järvinen 17051da177e4SLinus Torvalds /* Ok. We will be able to collapse the packet. */ 1706fe067e8aSDavid S. Miller tcp_unlink_write_queue(next_skb, sk); 17071da177e4SLinus Torvalds 17081a4e2d09SArnaldo Carvalho de Melo skb_copy_from_linear_data(next_skb, 17091a4e2d09SArnaldo Carvalho de Melo skb_put(skb, next_skb_size), 17101a4e2d09SArnaldo Carvalho de Melo next_skb_size); 17111da177e4SLinus Torvalds 171252d570aaSJarek Poplawski if (next_skb->ip_summed == CHECKSUM_PARTIAL) 171352d570aaSJarek Poplawski skb->ip_summed = CHECKSUM_PARTIAL; 17141da177e4SLinus Torvalds 171584fa7933SPatrick McHardy if (skb->ip_summed != CHECKSUM_PARTIAL) 17161da177e4SLinus Torvalds skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size); 17171da177e4SLinus Torvalds 17181da177e4SLinus Torvalds /* Update sequence range on original skb. */ 17191da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq; 17201da177e4SLinus Torvalds 17211da177e4SLinus Torvalds /* Merge over control information. */ 17221da177e4SLinus Torvalds flags |= TCP_SKB_CB(next_skb)->flags; /* This moves PSH/FIN etc. over */ 17231da177e4SLinus Torvalds TCP_SKB_CB(skb)->flags = flags; 17241da177e4SLinus Torvalds 17251da177e4SLinus Torvalds /* All done, get rid of second SKB and account for it so 17261da177e4SLinus Torvalds * packet counting does not break. 17271da177e4SLinus Torvalds */ 1728*4828e7f4SIlpo Järvinen TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS; 17291da177e4SLinus Torvalds if (TCP_SKB_CB(next_skb)->sacked&TCPCB_SACKED_RETRANS) 17301da177e4SLinus Torvalds tp->retrans_out -= tcp_skb_pcount(next_skb); 1731b5860bbaSIlpo Järvinen if (TCP_SKB_CB(next_skb)->sacked&TCPCB_LOST) 17321da177e4SLinus Torvalds tp->lost_out -= tcp_skb_pcount(next_skb); 17331da177e4SLinus Torvalds /* Reno case is special. Sigh... */ 1734e60402d0SIlpo Järvinen if (tcp_is_reno(tp) && tp->sacked_out) 17351da177e4SLinus Torvalds tcp_dec_pcount_approx(&tp->sacked_out, next_skb); 17361da177e4SLinus Torvalds 1737a47e5a98SIlpo Järvinen tcp_adjust_fackets_out(sk, next_skb, tcp_skb_pcount(next_skb)); 1738e9144bd8SIlpo Järvinen tp->packets_out -= tcp_skb_pcount(next_skb); 1739b7689205SIlpo Järvinen 1740b7689205SIlpo Järvinen /* changed transmit queue under us so clear hints */ 1741b7689205SIlpo Järvinen tcp_clear_retrans_hints_partial(tp); 1742b7689205SIlpo Järvinen 17433ab224beSHideo Aoki sk_wmem_free_skb(sk, next_skb); 17441da177e4SLinus Torvalds } 17451da177e4SLinus Torvalds } 17461da177e4SLinus Torvalds 17471da177e4SLinus Torvalds /* Do a simple retransmit without using the backoff mechanisms in 17481da177e4SLinus Torvalds * tcp_timer. This is used for path mtu discovery. 17491da177e4SLinus Torvalds * The socket is already locked here. 17501da177e4SLinus Torvalds */ 17511da177e4SLinus Torvalds void tcp_simple_retransmit(struct sock *sk) 17521da177e4SLinus Torvalds { 17536687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 17541da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 17551da177e4SLinus Torvalds struct sk_buff *skb; 17561da177e4SLinus Torvalds unsigned int mss = tcp_current_mss(sk, 0); 17571da177e4SLinus Torvalds int lost = 0; 17581da177e4SLinus Torvalds 1759fe067e8aSDavid S. Miller tcp_for_write_queue(skb, sk) { 1760fe067e8aSDavid S. Miller if (skb == tcp_send_head(sk)) 1761fe067e8aSDavid S. Miller break; 17621da177e4SLinus Torvalds if (skb->len > mss && 17631da177e4SLinus Torvalds !(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED)) { 17641da177e4SLinus Torvalds if (TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS) { 17651da177e4SLinus Torvalds TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS; 17661da177e4SLinus Torvalds tp->retrans_out -= tcp_skb_pcount(skb); 17671da177e4SLinus Torvalds } 17681da177e4SLinus Torvalds if (!(TCP_SKB_CB(skb)->sacked&TCPCB_LOST)) { 17691da177e4SLinus Torvalds TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; 17701da177e4SLinus Torvalds tp->lost_out += tcp_skb_pcount(skb); 17711da177e4SLinus Torvalds lost = 1; 17721da177e4SLinus Torvalds } 17731da177e4SLinus Torvalds } 17741da177e4SLinus Torvalds } 17751da177e4SLinus Torvalds 17765af4ec23SIlpo Järvinen tcp_clear_all_retrans_hints(tp); 17776a438bbeSStephen Hemminger 17781da177e4SLinus Torvalds if (!lost) 17791da177e4SLinus Torvalds return; 17801da177e4SLinus Torvalds 1781005903bcSIlpo Järvinen tcp_verify_left_out(tp); 17821da177e4SLinus Torvalds 17831da177e4SLinus Torvalds /* Don't muck with the congestion window here. 17841da177e4SLinus Torvalds * Reason is that we do not increase amount of _data_ 17851da177e4SLinus Torvalds * in network, but units changed and effective 17861da177e4SLinus Torvalds * cwnd/ssthresh really reduced now. 17871da177e4SLinus Torvalds */ 17886687e988SArnaldo Carvalho de Melo if (icsk->icsk_ca_state != TCP_CA_Loss) { 17891da177e4SLinus Torvalds tp->high_seq = tp->snd_nxt; 17906687e988SArnaldo Carvalho de Melo tp->snd_ssthresh = tcp_current_ssthresh(sk); 17911da177e4SLinus Torvalds tp->prior_ssthresh = 0; 17921da177e4SLinus Torvalds tp->undo_marker = 0; 17936687e988SArnaldo Carvalho de Melo tcp_set_ca_state(sk, TCP_CA_Loss); 17941da177e4SLinus Torvalds } 17951da177e4SLinus Torvalds tcp_xmit_retransmit_queue(sk); 17961da177e4SLinus Torvalds } 17971da177e4SLinus Torvalds 17981da177e4SLinus Torvalds /* This retransmits one SKB. Policy decisions and retransmit queue 17991da177e4SLinus Torvalds * state updates are done by the caller. Returns non-zero if an 18001da177e4SLinus Torvalds * error occurred which prevented the send. 18011da177e4SLinus Torvalds */ 18021da177e4SLinus Torvalds int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb) 18031da177e4SLinus Torvalds { 18041da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 18055d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 18061da177e4SLinus Torvalds unsigned int cur_mss = tcp_current_mss(sk, 0); 18071da177e4SLinus Torvalds int err; 18081da177e4SLinus Torvalds 18095d424d5aSJohn Heffner /* Inconslusive MTU probe */ 18105d424d5aSJohn Heffner if (icsk->icsk_mtup.probe_size) { 18115d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = 0; 18125d424d5aSJohn Heffner } 18135d424d5aSJohn Heffner 18141da177e4SLinus Torvalds /* Do not sent more than we queued. 1/4 is reserved for possible 1815caa20d9aSStephen Hemminger * copying overhead: fragmentation, tunneling, mangling etc. 18161da177e4SLinus Torvalds */ 18171da177e4SLinus Torvalds if (atomic_read(&sk->sk_wmem_alloc) > 18181da177e4SLinus Torvalds min(sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), sk->sk_sndbuf)) 18191da177e4SLinus Torvalds return -EAGAIN; 18201da177e4SLinus Torvalds 18211da177e4SLinus Torvalds if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) { 18221da177e4SLinus Torvalds if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) 18231da177e4SLinus Torvalds BUG(); 18241da177e4SLinus Torvalds if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq)) 18251da177e4SLinus Torvalds return -ENOMEM; 18261da177e4SLinus Torvalds } 18271da177e4SLinus Torvalds 18281da177e4SLinus Torvalds /* If receiver has shrunk his window, and skb is out of 18291da177e4SLinus Torvalds * new window, do not retransmit it. The exception is the 18301da177e4SLinus Torvalds * case, when window is shrunk to zero. In this case 18311da177e4SLinus Torvalds * our retransmit serves as a zero window probe. 18321da177e4SLinus Torvalds */ 183390840defSIlpo Järvinen if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) 18341da177e4SLinus Torvalds && TCP_SKB_CB(skb)->seq != tp->snd_una) 18351da177e4SLinus Torvalds return -EAGAIN; 18361da177e4SLinus Torvalds 18371da177e4SLinus Torvalds if (skb->len > cur_mss) { 1838846998aeSDavid S. Miller if (tcp_fragment(sk, skb, cur_mss, cur_mss)) 18391da177e4SLinus Torvalds return -ENOMEM; /* We'll try again later. */ 18401da177e4SLinus Torvalds } 18411da177e4SLinus Torvalds 18421da177e4SLinus Torvalds /* Collapse two adjacent packets if worthwhile and we can. */ 18431da177e4SLinus Torvalds if (!(TCP_SKB_CB(skb)->flags & TCPCB_FLAG_SYN) && 18441da177e4SLinus Torvalds (skb->len < (cur_mss >> 1)) && 1845fe067e8aSDavid S. Miller (tcp_write_queue_next(sk, skb) != tcp_send_head(sk)) && 1846fe067e8aSDavid S. Miller (!tcp_skb_is_last(sk, skb)) && 1847fe067e8aSDavid S. Miller (skb_shinfo(skb)->nr_frags == 0 && skb_shinfo(tcp_write_queue_next(sk, skb))->nr_frags == 0) && 1848fe067e8aSDavid S. Miller (tcp_skb_pcount(skb) == 1 && tcp_skb_pcount(tcp_write_queue_next(sk, skb)) == 1) && 18491da177e4SLinus Torvalds (sysctl_tcp_retrans_collapse != 0)) 18501da177e4SLinus Torvalds tcp_retrans_try_collapse(sk, skb, cur_mss); 18511da177e4SLinus Torvalds 18528292a17aSArnaldo Carvalho de Melo if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk)) 18531da177e4SLinus Torvalds return -EHOSTUNREACH; /* Routing failure or similar. */ 18541da177e4SLinus Torvalds 18551da177e4SLinus Torvalds /* Some Solaris stacks overoptimize and ignore the FIN on a 18561da177e4SLinus Torvalds * retransmit when old data is attached. So strip it off 18571da177e4SLinus Torvalds * since it is cheap to do so and saves bytes on the network. 18581da177e4SLinus Torvalds */ 18591da177e4SLinus Torvalds if (skb->len > 0 && 18601da177e4SLinus Torvalds (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN) && 18611da177e4SLinus Torvalds tp->snd_una == (TCP_SKB_CB(skb)->end_seq - 1)) { 18621da177e4SLinus Torvalds if (!pskb_trim(skb, 0)) { 18631da177e4SLinus Torvalds TCP_SKB_CB(skb)->seq = TCP_SKB_CB(skb)->end_seq - 1; 18647967168cSHerbert Xu skb_shinfo(skb)->gso_segs = 1; 18657967168cSHerbert Xu skb_shinfo(skb)->gso_size = 0; 18667967168cSHerbert Xu skb_shinfo(skb)->gso_type = 0; 18671da177e4SLinus Torvalds skb->ip_summed = CHECKSUM_NONE; 18681da177e4SLinus Torvalds skb->csum = 0; 18691da177e4SLinus Torvalds } 18701da177e4SLinus Torvalds } 18711da177e4SLinus Torvalds 18721da177e4SLinus Torvalds /* Make a copy, if the first transmission SKB clone we made 18731da177e4SLinus Torvalds * is still in somebody's hands, else make a clone. 18741da177e4SLinus Torvalds */ 18751da177e4SLinus Torvalds TCP_SKB_CB(skb)->when = tcp_time_stamp; 18761da177e4SLinus Torvalds 1877dfb4b9dcSDavid S. Miller err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 18781da177e4SLinus Torvalds 18791da177e4SLinus Torvalds if (err == 0) { 18801da177e4SLinus Torvalds /* Update global TCP statistics. */ 18811da177e4SLinus Torvalds TCP_INC_STATS(TCP_MIB_RETRANSSEGS); 18821da177e4SLinus Torvalds 18831da177e4SLinus Torvalds tp->total_retrans++; 18841da177e4SLinus Torvalds 18851da177e4SLinus Torvalds #if FASTRETRANS_DEBUG > 0 18861da177e4SLinus Torvalds if (TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS) { 18871da177e4SLinus Torvalds if (net_ratelimit()) 18881da177e4SLinus Torvalds printk(KERN_DEBUG "retrans_out leaked.\n"); 18891da177e4SLinus Torvalds } 18901da177e4SLinus Torvalds #endif 1891b08d6cb2SIlpo Järvinen if (!tp->retrans_out) 1892b08d6cb2SIlpo Järvinen tp->lost_retrans_low = tp->snd_nxt; 18931da177e4SLinus Torvalds TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS; 18941da177e4SLinus Torvalds tp->retrans_out += tcp_skb_pcount(skb); 18951da177e4SLinus Torvalds 18961da177e4SLinus Torvalds /* Save stamp of the first retransmit. */ 18971da177e4SLinus Torvalds if (!tp->retrans_stamp) 18981da177e4SLinus Torvalds tp->retrans_stamp = TCP_SKB_CB(skb)->when; 18991da177e4SLinus Torvalds 19001da177e4SLinus Torvalds tp->undo_retrans++; 19011da177e4SLinus Torvalds 19021da177e4SLinus Torvalds /* snd_nxt is stored to detect loss of retransmitted segment, 19031da177e4SLinus Torvalds * see tcp_input.c tcp_sacktag_write_queue(). 19041da177e4SLinus Torvalds */ 19051da177e4SLinus Torvalds TCP_SKB_CB(skb)->ack_seq = tp->snd_nxt; 19061da177e4SLinus Torvalds } 19071da177e4SLinus Torvalds return err; 19081da177e4SLinus Torvalds } 19091da177e4SLinus Torvalds 19101da177e4SLinus Torvalds /* This gets called after a retransmit timeout, and the initially 19111da177e4SLinus Torvalds * retransmitted data is acknowledged. It tries to continue 19121da177e4SLinus Torvalds * resending the rest of the retransmit queue, until either 19131da177e4SLinus Torvalds * we've sent it all or the congestion window limit is reached. 19141da177e4SLinus Torvalds * If doing SACK, the first ACK which comes back for a timeout 19151da177e4SLinus Torvalds * based retransmit packet might feed us FACK information again. 19161da177e4SLinus Torvalds * If so, we use it to avoid unnecessarily retransmissions. 19171da177e4SLinus Torvalds */ 19181da177e4SLinus Torvalds void tcp_xmit_retransmit_queue(struct sock *sk) 19191da177e4SLinus Torvalds { 19206687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 19211da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 19221da177e4SLinus Torvalds struct sk_buff *skb; 19236a438bbeSStephen Hemminger int packet_cnt; 19246a438bbeSStephen Hemminger 19256a438bbeSStephen Hemminger if (tp->retransmit_skb_hint) { 19266a438bbeSStephen Hemminger skb = tp->retransmit_skb_hint; 19276a438bbeSStephen Hemminger packet_cnt = tp->retransmit_cnt_hint; 19286a438bbeSStephen Hemminger }else{ 1929fe067e8aSDavid S. Miller skb = tcp_write_queue_head(sk); 19306a438bbeSStephen Hemminger packet_cnt = 0; 19316a438bbeSStephen Hemminger } 19321da177e4SLinus Torvalds 19331da177e4SLinus Torvalds /* First pass: retransmit lost packets. */ 19346a438bbeSStephen Hemminger if (tp->lost_out) { 1935fe067e8aSDavid S. Miller tcp_for_write_queue_from(skb, sk) { 19361da177e4SLinus Torvalds __u8 sacked = TCP_SKB_CB(skb)->sacked; 19371da177e4SLinus Torvalds 1938fe067e8aSDavid S. Miller if (skb == tcp_send_head(sk)) 1939fe067e8aSDavid S. Miller break; 19406a438bbeSStephen Hemminger /* we could do better than to assign each time */ 19416a438bbeSStephen Hemminger tp->retransmit_skb_hint = skb; 19426a438bbeSStephen Hemminger tp->retransmit_cnt_hint = packet_cnt; 19436a438bbeSStephen Hemminger 19441da177e4SLinus Torvalds /* Assume this retransmit will generate 19451da177e4SLinus Torvalds * only one packet for congestion window 19461da177e4SLinus Torvalds * calculation purposes. This works because 19471da177e4SLinus Torvalds * tcp_retransmit_skb() will chop up the 19481da177e4SLinus Torvalds * packet to be MSS sized and all the 19491da177e4SLinus Torvalds * packet counting works out. 19501da177e4SLinus Torvalds */ 19511da177e4SLinus Torvalds if (tcp_packets_in_flight(tp) >= tp->snd_cwnd) 19521da177e4SLinus Torvalds return; 19531da177e4SLinus Torvalds 19541da177e4SLinus Torvalds if (sacked & TCPCB_LOST) { 19551da177e4SLinus Torvalds if (!(sacked&(TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))) { 19566a438bbeSStephen Hemminger if (tcp_retransmit_skb(sk, skb)) { 19576a438bbeSStephen Hemminger tp->retransmit_skb_hint = NULL; 19581da177e4SLinus Torvalds return; 19596a438bbeSStephen Hemminger } 19606687e988SArnaldo Carvalho de Melo if (icsk->icsk_ca_state != TCP_CA_Loss) 19611da177e4SLinus Torvalds NET_INC_STATS_BH(LINUX_MIB_TCPFASTRETRANS); 19621da177e4SLinus Torvalds else 19631da177e4SLinus Torvalds NET_INC_STATS_BH(LINUX_MIB_TCPSLOWSTARTRETRANS); 19641da177e4SLinus Torvalds 1965fe067e8aSDavid S. Miller if (skb == tcp_write_queue_head(sk)) 1966463c84b9SArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 19673f421baaSArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto, 19683f421baaSArnaldo Carvalho de Melo TCP_RTO_MAX); 19691da177e4SLinus Torvalds } 19701da177e4SLinus Torvalds 19716a438bbeSStephen Hemminger packet_cnt += tcp_skb_pcount(skb); 19726a438bbeSStephen Hemminger if (packet_cnt >= tp->lost_out) 19731da177e4SLinus Torvalds break; 19741da177e4SLinus Torvalds } 19751da177e4SLinus Torvalds } 19761da177e4SLinus Torvalds } 19771da177e4SLinus Torvalds 19781da177e4SLinus Torvalds /* OK, demanded retransmission is finished. */ 19791da177e4SLinus Torvalds 19801da177e4SLinus Torvalds /* Forward retransmissions are possible only during Recovery. */ 19816687e988SArnaldo Carvalho de Melo if (icsk->icsk_ca_state != TCP_CA_Recovery) 19821da177e4SLinus Torvalds return; 19831da177e4SLinus Torvalds 19841da177e4SLinus Torvalds /* No forward retransmissions in Reno are possible. */ 1985e60402d0SIlpo Järvinen if (tcp_is_reno(tp)) 19861da177e4SLinus Torvalds return; 19871da177e4SLinus Torvalds 19881da177e4SLinus Torvalds /* Yeah, we have to make difficult choice between forward transmission 19891da177e4SLinus Torvalds * and retransmission... Both ways have their merits... 19901da177e4SLinus Torvalds * 19911da177e4SLinus Torvalds * For now we do not retransmit anything, while we have some new 1992539d243fSIlpo Järvinen * segments to send. In the other cases, follow rule 3 for 1993539d243fSIlpo Järvinen * NextSeg() specified in RFC3517. 19941da177e4SLinus Torvalds */ 19951da177e4SLinus Torvalds 19969e412ba7SIlpo Järvinen if (tcp_may_send_now(sk)) 19971da177e4SLinus Torvalds return; 19981da177e4SLinus Torvalds 1999539d243fSIlpo Järvinen /* If nothing is SACKed, highest_sack in the loop won't be valid */ 2000539d243fSIlpo Järvinen if (!tp->sacked_out) 2001539d243fSIlpo Järvinen return; 2002539d243fSIlpo Järvinen 2003539d243fSIlpo Järvinen if (tp->forward_skb_hint) 20046a438bbeSStephen Hemminger skb = tp->forward_skb_hint; 2005539d243fSIlpo Järvinen else 2006fe067e8aSDavid S. Miller skb = tcp_write_queue_head(sk); 20071da177e4SLinus Torvalds 2008fe067e8aSDavid S. Miller tcp_for_write_queue_from(skb, sk) { 2009fe067e8aSDavid S. Miller if (skb == tcp_send_head(sk)) 2010fe067e8aSDavid S. Miller break; 20116a438bbeSStephen Hemminger tp->forward_skb_hint = skb; 20126a438bbeSStephen Hemminger 20136859d494SIlpo Järvinen if (!before(TCP_SKB_CB(skb)->seq, tcp_highest_sack_seq(tp))) 20141da177e4SLinus Torvalds break; 20151da177e4SLinus Torvalds 20161da177e4SLinus Torvalds if (tcp_packets_in_flight(tp) >= tp->snd_cwnd) 20171da177e4SLinus Torvalds break; 20181da177e4SLinus Torvalds 20191da177e4SLinus Torvalds if (TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) 20201da177e4SLinus Torvalds continue; 20211da177e4SLinus Torvalds 20221da177e4SLinus Torvalds /* Ok, retransmit it. */ 20236a438bbeSStephen Hemminger if (tcp_retransmit_skb(sk, skb)) { 20246a438bbeSStephen Hemminger tp->forward_skb_hint = NULL; 20251da177e4SLinus Torvalds break; 20266a438bbeSStephen Hemminger } 20271da177e4SLinus Torvalds 2028fe067e8aSDavid S. Miller if (skb == tcp_write_queue_head(sk)) 20293f421baaSArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 20303f421baaSArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto, 20313f421baaSArnaldo Carvalho de Melo TCP_RTO_MAX); 20321da177e4SLinus Torvalds 20331da177e4SLinus Torvalds NET_INC_STATS_BH(LINUX_MIB_TCPFORWARDRETRANS); 20341da177e4SLinus Torvalds } 20351da177e4SLinus Torvalds } 20361da177e4SLinus Torvalds 20371da177e4SLinus Torvalds 20381da177e4SLinus Torvalds /* Send a fin. The caller locks the socket for us. This cannot be 20391da177e4SLinus Torvalds * allowed to fail queueing a FIN frame under any circumstances. 20401da177e4SLinus Torvalds */ 20411da177e4SLinus Torvalds void tcp_send_fin(struct sock *sk) 20421da177e4SLinus Torvalds { 20431da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 2044fe067e8aSDavid S. Miller struct sk_buff *skb = tcp_write_queue_tail(sk); 20451da177e4SLinus Torvalds int mss_now; 20461da177e4SLinus Torvalds 20471da177e4SLinus Torvalds /* Optimization, tack on the FIN if we have a queue of 20481da177e4SLinus Torvalds * unsent frames. But be careful about outgoing SACKS 20491da177e4SLinus Torvalds * and IP options. 20501da177e4SLinus Torvalds */ 20511da177e4SLinus Torvalds mss_now = tcp_current_mss(sk, 1); 20521da177e4SLinus Torvalds 2053fe067e8aSDavid S. Miller if (tcp_send_head(sk) != NULL) { 20541da177e4SLinus Torvalds TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_FIN; 20551da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq++; 20561da177e4SLinus Torvalds tp->write_seq++; 20571da177e4SLinus Torvalds } else { 20581da177e4SLinus Torvalds /* Socket is locked, keep trying until memory is available. */ 20591da177e4SLinus Torvalds for (;;) { 2060d179cd12SDavid S. Miller skb = alloc_skb_fclone(MAX_TCP_HEADER, GFP_KERNEL); 20611da177e4SLinus Torvalds if (skb) 20621da177e4SLinus Torvalds break; 20631da177e4SLinus Torvalds yield(); 20641da177e4SLinus Torvalds } 20651da177e4SLinus Torvalds 20661da177e4SLinus Torvalds /* Reserve space for headers and prepare control bits. */ 20671da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 20681da177e4SLinus Torvalds skb->csum = 0; 20691da177e4SLinus Torvalds TCP_SKB_CB(skb)->flags = (TCPCB_FLAG_ACK | TCPCB_FLAG_FIN); 20701da177e4SLinus Torvalds TCP_SKB_CB(skb)->sacked = 0; 20717967168cSHerbert Xu skb_shinfo(skb)->gso_segs = 1; 20727967168cSHerbert Xu skb_shinfo(skb)->gso_size = 0; 20737967168cSHerbert Xu skb_shinfo(skb)->gso_type = 0; 20741da177e4SLinus Torvalds 20751da177e4SLinus Torvalds /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */ 20761da177e4SLinus Torvalds TCP_SKB_CB(skb)->seq = tp->write_seq; 20771da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + 1; 20781da177e4SLinus Torvalds tcp_queue_skb(sk, skb); 20791da177e4SLinus Torvalds } 20809e412ba7SIlpo Järvinen __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_OFF); 20811da177e4SLinus Torvalds } 20821da177e4SLinus Torvalds 20831da177e4SLinus Torvalds /* We get here when a process closes a file descriptor (either due to 20841da177e4SLinus Torvalds * an explicit close() or as a byproduct of exit()'ing) and there 20851da177e4SLinus Torvalds * was unread data in the receive queue. This behavior is recommended 208665bb723cSGerrit Renker * by RFC 2525, section 2.17. -DaveM 20871da177e4SLinus Torvalds */ 2088dd0fc66fSAl Viro void tcp_send_active_reset(struct sock *sk, gfp_t priority) 20891da177e4SLinus Torvalds { 20901da177e4SLinus Torvalds struct sk_buff *skb; 20911da177e4SLinus Torvalds 20921da177e4SLinus Torvalds /* NOTE: No TCP options attached and we never retransmit this. */ 20931da177e4SLinus Torvalds skb = alloc_skb(MAX_TCP_HEADER, priority); 20941da177e4SLinus Torvalds if (!skb) { 20951da177e4SLinus Torvalds NET_INC_STATS(LINUX_MIB_TCPABORTFAILED); 20961da177e4SLinus Torvalds return; 20971da177e4SLinus Torvalds } 20981da177e4SLinus Torvalds 20991da177e4SLinus Torvalds /* Reserve space for headers and prepare control bits. */ 21001da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 21011da177e4SLinus Torvalds skb->csum = 0; 21021da177e4SLinus Torvalds TCP_SKB_CB(skb)->flags = (TCPCB_FLAG_ACK | TCPCB_FLAG_RST); 21031da177e4SLinus Torvalds TCP_SKB_CB(skb)->sacked = 0; 21047967168cSHerbert Xu skb_shinfo(skb)->gso_segs = 1; 21057967168cSHerbert Xu skb_shinfo(skb)->gso_size = 0; 21067967168cSHerbert Xu skb_shinfo(skb)->gso_type = 0; 21071da177e4SLinus Torvalds 21081da177e4SLinus Torvalds /* Send it off. */ 21099e412ba7SIlpo Järvinen TCP_SKB_CB(skb)->seq = tcp_acceptable_seq(sk); 21101da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq; 21111da177e4SLinus Torvalds TCP_SKB_CB(skb)->when = tcp_time_stamp; 2112dfb4b9dcSDavid S. Miller if (tcp_transmit_skb(sk, skb, 0, priority)) 21131da177e4SLinus Torvalds NET_INC_STATS(LINUX_MIB_TCPABORTFAILED); 21141da177e4SLinus Torvalds } 21151da177e4SLinus Torvalds 21161da177e4SLinus Torvalds /* WARNING: This routine must only be called when we have already sent 21171da177e4SLinus Torvalds * a SYN packet that crossed the incoming SYN that caused this routine 21181da177e4SLinus Torvalds * to get called. If this assumption fails then the initial rcv_wnd 21191da177e4SLinus Torvalds * and rcv_wscale values will not be correct. 21201da177e4SLinus Torvalds */ 21211da177e4SLinus Torvalds int tcp_send_synack(struct sock *sk) 21221da177e4SLinus Torvalds { 21231da177e4SLinus Torvalds struct sk_buff* skb; 21241da177e4SLinus Torvalds 2125fe067e8aSDavid S. Miller skb = tcp_write_queue_head(sk); 21261da177e4SLinus Torvalds if (skb == NULL || !(TCP_SKB_CB(skb)->flags&TCPCB_FLAG_SYN)) { 21271da177e4SLinus Torvalds printk(KERN_DEBUG "tcp_send_synack: wrong queue state\n"); 21281da177e4SLinus Torvalds return -EFAULT; 21291da177e4SLinus Torvalds } 21301da177e4SLinus Torvalds if (!(TCP_SKB_CB(skb)->flags&TCPCB_FLAG_ACK)) { 21311da177e4SLinus Torvalds if (skb_cloned(skb)) { 21321da177e4SLinus Torvalds struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC); 21331da177e4SLinus Torvalds if (nskb == NULL) 21341da177e4SLinus Torvalds return -ENOMEM; 2135fe067e8aSDavid S. Miller tcp_unlink_write_queue(skb, sk); 21361da177e4SLinus Torvalds skb_header_release(nskb); 2137fe067e8aSDavid S. Miller __tcp_add_write_queue_head(sk, nskb); 21383ab224beSHideo Aoki sk_wmem_free_skb(sk, skb); 21393ab224beSHideo Aoki sk->sk_wmem_queued += nskb->truesize; 21403ab224beSHideo Aoki sk_mem_charge(sk, nskb->truesize); 21411da177e4SLinus Torvalds skb = nskb; 21421da177e4SLinus Torvalds } 21431da177e4SLinus Torvalds 21441da177e4SLinus Torvalds TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_ACK; 21451da177e4SLinus Torvalds TCP_ECN_send_synack(tcp_sk(sk), skb); 21461da177e4SLinus Torvalds } 21471da177e4SLinus Torvalds TCP_SKB_CB(skb)->when = tcp_time_stamp; 2148dfb4b9dcSDavid S. Miller return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 21491da177e4SLinus Torvalds } 21501da177e4SLinus Torvalds 21511da177e4SLinus Torvalds /* 21521da177e4SLinus Torvalds * Prepare a SYN-ACK. 21531da177e4SLinus Torvalds */ 21541da177e4SLinus Torvalds struct sk_buff * tcp_make_synack(struct sock *sk, struct dst_entry *dst, 215560236fddSArnaldo Carvalho de Melo struct request_sock *req) 21561da177e4SLinus Torvalds { 21572e6599cbSArnaldo Carvalho de Melo struct inet_request_sock *ireq = inet_rsk(req); 21581da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 21591da177e4SLinus Torvalds struct tcphdr *th; 21601da177e4SLinus Torvalds int tcp_header_size; 21611da177e4SLinus Torvalds struct sk_buff *skb; 2162cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 2163cfb6eeb4SYOSHIFUJI Hideaki struct tcp_md5sig_key *md5; 2164cfb6eeb4SYOSHIFUJI Hideaki __u8 *md5_hash_location; 2165cfb6eeb4SYOSHIFUJI Hideaki #endif 21661da177e4SLinus Torvalds 21671da177e4SLinus Torvalds skb = sock_wmalloc(sk, MAX_TCP_HEADER + 15, 1, GFP_ATOMIC); 21681da177e4SLinus Torvalds if (skb == NULL) 21691da177e4SLinus Torvalds return NULL; 21701da177e4SLinus Torvalds 21711da177e4SLinus Torvalds /* Reserve space for headers. */ 21721da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 21731da177e4SLinus Torvalds 21741da177e4SLinus Torvalds skb->dst = dst_clone(dst); 21751da177e4SLinus Torvalds 21761da177e4SLinus Torvalds tcp_header_size = (sizeof(struct tcphdr) + TCPOLEN_MSS + 21772e6599cbSArnaldo Carvalho de Melo (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0) + 21782e6599cbSArnaldo Carvalho de Melo (ireq->wscale_ok ? TCPOLEN_WSCALE_ALIGNED : 0) + 21791da177e4SLinus Torvalds /* SACK_PERM is in the place of NOP NOP of TS */ 21802e6599cbSArnaldo Carvalho de Melo ((ireq->sack_ok && !ireq->tstamp_ok) ? TCPOLEN_SACKPERM_ALIGNED : 0)); 2181cfb6eeb4SYOSHIFUJI Hideaki 2182cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 2183cfb6eeb4SYOSHIFUJI Hideaki /* Are we doing MD5 on this segment? If so - make room for it */ 2184cfb6eeb4SYOSHIFUJI Hideaki md5 = tcp_rsk(req)->af_specific->md5_lookup(sk, req); 2185cfb6eeb4SYOSHIFUJI Hideaki if (md5) 2186cfb6eeb4SYOSHIFUJI Hideaki tcp_header_size += TCPOLEN_MD5SIG_ALIGNED; 2187cfb6eeb4SYOSHIFUJI Hideaki #endif 2188aa8223c7SArnaldo Carvalho de Melo skb_push(skb, tcp_header_size); 2189aa8223c7SArnaldo Carvalho de Melo skb_reset_transport_header(skb); 21901da177e4SLinus Torvalds 2191aa8223c7SArnaldo Carvalho de Melo th = tcp_hdr(skb); 21921da177e4SLinus Torvalds memset(th, 0, sizeof(struct tcphdr)); 21931da177e4SLinus Torvalds th->syn = 1; 21941da177e4SLinus Torvalds th->ack = 1; 21951da177e4SLinus Torvalds TCP_ECN_make_synack(req, th); 21961da177e4SLinus Torvalds th->source = inet_sk(sk)->sport; 21972e6599cbSArnaldo Carvalho de Melo th->dest = ireq->rmt_port; 21982e6599cbSArnaldo Carvalho de Melo TCP_SKB_CB(skb)->seq = tcp_rsk(req)->snt_isn; 21991da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + 1; 22001da177e4SLinus Torvalds TCP_SKB_CB(skb)->sacked = 0; 22017967168cSHerbert Xu skb_shinfo(skb)->gso_segs = 1; 22027967168cSHerbert Xu skb_shinfo(skb)->gso_size = 0; 22037967168cSHerbert Xu skb_shinfo(skb)->gso_type = 0; 22041da177e4SLinus Torvalds th->seq = htonl(TCP_SKB_CB(skb)->seq); 22052e6599cbSArnaldo Carvalho de Melo th->ack_seq = htonl(tcp_rsk(req)->rcv_isn + 1); 22061da177e4SLinus Torvalds if (req->rcv_wnd == 0) { /* ignored for retransmitted syns */ 22071da177e4SLinus Torvalds __u8 rcv_wscale; 22081da177e4SLinus Torvalds /* Set this up on the first call only */ 22091da177e4SLinus Torvalds req->window_clamp = tp->window_clamp ? : dst_metric(dst, RTAX_WINDOW); 22101da177e4SLinus Torvalds /* tcp_full_space because it is guaranteed to be the first packet */ 22111da177e4SLinus Torvalds tcp_select_initial_window(tcp_full_space(sk), 22122e6599cbSArnaldo Carvalho de Melo dst_metric(dst, RTAX_ADVMSS) - (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0), 22131da177e4SLinus Torvalds &req->rcv_wnd, 22141da177e4SLinus Torvalds &req->window_clamp, 22152e6599cbSArnaldo Carvalho de Melo ireq->wscale_ok, 22161da177e4SLinus Torvalds &rcv_wscale); 22172e6599cbSArnaldo Carvalho de Melo ireq->rcv_wscale = rcv_wscale; 22181da177e4SLinus Torvalds } 22191da177e4SLinus Torvalds 22201da177e4SLinus Torvalds /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */ 2221600ff0c2SIlpo Järvinen th->window = htons(min(req->rcv_wnd, 65535U)); 22221da177e4SLinus Torvalds 22231da177e4SLinus Torvalds TCP_SKB_CB(skb)->when = tcp_time_stamp; 2224df7a3b07SAl Viro tcp_syn_build_options((__be32 *)(th + 1), dst_metric(dst, RTAX_ADVMSS), ireq->tstamp_ok, 22252e6599cbSArnaldo Carvalho de Melo ireq->sack_ok, ireq->wscale_ok, ireq->rcv_wscale, 22261da177e4SLinus Torvalds TCP_SKB_CB(skb)->when, 2227cfb6eeb4SYOSHIFUJI Hideaki req->ts_recent, 2228cfb6eeb4SYOSHIFUJI Hideaki ( 2229cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 2230cfb6eeb4SYOSHIFUJI Hideaki md5 ? &md5_hash_location : 2231cfb6eeb4SYOSHIFUJI Hideaki #endif 2232cfb6eeb4SYOSHIFUJI Hideaki NULL) 2233cfb6eeb4SYOSHIFUJI Hideaki ); 22341da177e4SLinus Torvalds 22351da177e4SLinus Torvalds skb->csum = 0; 22361da177e4SLinus Torvalds th->doff = (tcp_header_size >> 2); 22371da177e4SLinus Torvalds TCP_INC_STATS(TCP_MIB_OUTSEGS); 2238cfb6eeb4SYOSHIFUJI Hideaki 2239cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 2240cfb6eeb4SYOSHIFUJI Hideaki /* Okay, we have all we need - do the md5 hash if needed */ 2241cfb6eeb4SYOSHIFUJI Hideaki if (md5) { 2242cfb6eeb4SYOSHIFUJI Hideaki tp->af_specific->calc_md5_hash(md5_hash_location, 2243cfb6eeb4SYOSHIFUJI Hideaki md5, 2244cfb6eeb4SYOSHIFUJI Hideaki NULL, dst, req, 2245aa8223c7SArnaldo Carvalho de Melo tcp_hdr(skb), sk->sk_protocol, 2246cfb6eeb4SYOSHIFUJI Hideaki skb->len); 2247cfb6eeb4SYOSHIFUJI Hideaki } 2248cfb6eeb4SYOSHIFUJI Hideaki #endif 2249cfb6eeb4SYOSHIFUJI Hideaki 22501da177e4SLinus Torvalds return skb; 22511da177e4SLinus Torvalds } 22521da177e4SLinus Torvalds 22531da177e4SLinus Torvalds /* 22541da177e4SLinus Torvalds * Do all connect socket setups that can be done AF independent. 22551da177e4SLinus Torvalds */ 225640efc6faSStephen Hemminger static void tcp_connect_init(struct sock *sk) 22571da177e4SLinus Torvalds { 22581da177e4SLinus Torvalds struct dst_entry *dst = __sk_dst_get(sk); 22591da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 22601da177e4SLinus Torvalds __u8 rcv_wscale; 22611da177e4SLinus Torvalds 22621da177e4SLinus Torvalds /* We'll fix this up when we get a response from the other end. 22631da177e4SLinus Torvalds * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT. 22641da177e4SLinus Torvalds */ 22651da177e4SLinus Torvalds tp->tcp_header_len = sizeof(struct tcphdr) + 22661da177e4SLinus Torvalds (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0); 22671da177e4SLinus Torvalds 2268cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 2269cfb6eeb4SYOSHIFUJI Hideaki if (tp->af_specific->md5_lookup(sk, sk) != NULL) 2270cfb6eeb4SYOSHIFUJI Hideaki tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED; 2271cfb6eeb4SYOSHIFUJI Hideaki #endif 2272cfb6eeb4SYOSHIFUJI Hideaki 22731da177e4SLinus Torvalds /* If user gave his TCP_MAXSEG, record it to clamp */ 22741da177e4SLinus Torvalds if (tp->rx_opt.user_mss) 22751da177e4SLinus Torvalds tp->rx_opt.mss_clamp = tp->rx_opt.user_mss; 22761da177e4SLinus Torvalds tp->max_window = 0; 22775d424d5aSJohn Heffner tcp_mtup_init(sk); 22781da177e4SLinus Torvalds tcp_sync_mss(sk, dst_mtu(dst)); 22791da177e4SLinus Torvalds 22801da177e4SLinus Torvalds if (!tp->window_clamp) 22811da177e4SLinus Torvalds tp->window_clamp = dst_metric(dst, RTAX_WINDOW); 22821da177e4SLinus Torvalds tp->advmss = dst_metric(dst, RTAX_ADVMSS); 22831da177e4SLinus Torvalds tcp_initialize_rcv_mss(sk); 22841da177e4SLinus Torvalds 22851da177e4SLinus Torvalds tcp_select_initial_window(tcp_full_space(sk), 22861da177e4SLinus Torvalds tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0), 22871da177e4SLinus Torvalds &tp->rcv_wnd, 22881da177e4SLinus Torvalds &tp->window_clamp, 22891da177e4SLinus Torvalds sysctl_tcp_window_scaling, 22901da177e4SLinus Torvalds &rcv_wscale); 22911da177e4SLinus Torvalds 22921da177e4SLinus Torvalds tp->rx_opt.rcv_wscale = rcv_wscale; 22931da177e4SLinus Torvalds tp->rcv_ssthresh = tp->rcv_wnd; 22941da177e4SLinus Torvalds 22951da177e4SLinus Torvalds sk->sk_err = 0; 22961da177e4SLinus Torvalds sock_reset_flag(sk, SOCK_DONE); 22971da177e4SLinus Torvalds tp->snd_wnd = 0; 22981da177e4SLinus Torvalds tcp_init_wl(tp, tp->write_seq, 0); 22991da177e4SLinus Torvalds tp->snd_una = tp->write_seq; 23001da177e4SLinus Torvalds tp->snd_sml = tp->write_seq; 23011da177e4SLinus Torvalds tp->rcv_nxt = 0; 23021da177e4SLinus Torvalds tp->rcv_wup = 0; 23031da177e4SLinus Torvalds tp->copied_seq = 0; 23041da177e4SLinus Torvalds 2305463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT; 2306463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_retransmits = 0; 23071da177e4SLinus Torvalds tcp_clear_retrans(tp); 23081da177e4SLinus Torvalds } 23091da177e4SLinus Torvalds 23101da177e4SLinus Torvalds /* 23111da177e4SLinus Torvalds * Build a SYN and send it off. 23121da177e4SLinus Torvalds */ 23131da177e4SLinus Torvalds int tcp_connect(struct sock *sk) 23141da177e4SLinus Torvalds { 23151da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 23161da177e4SLinus Torvalds struct sk_buff *buff; 23171da177e4SLinus Torvalds 23181da177e4SLinus Torvalds tcp_connect_init(sk); 23191da177e4SLinus Torvalds 2320d179cd12SDavid S. Miller buff = alloc_skb_fclone(MAX_TCP_HEADER + 15, sk->sk_allocation); 23211da177e4SLinus Torvalds if (unlikely(buff == NULL)) 23221da177e4SLinus Torvalds return -ENOBUFS; 23231da177e4SLinus Torvalds 23241da177e4SLinus Torvalds /* Reserve space for headers. */ 23251da177e4SLinus Torvalds skb_reserve(buff, MAX_TCP_HEADER); 23261da177e4SLinus Torvalds 23271da177e4SLinus Torvalds TCP_SKB_CB(buff)->flags = TCPCB_FLAG_SYN; 23289e412ba7SIlpo Järvinen TCP_ECN_send_syn(sk, buff); 23291da177e4SLinus Torvalds TCP_SKB_CB(buff)->sacked = 0; 23307967168cSHerbert Xu skb_shinfo(buff)->gso_segs = 1; 23317967168cSHerbert Xu skb_shinfo(buff)->gso_size = 0; 23327967168cSHerbert Xu skb_shinfo(buff)->gso_type = 0; 23331da177e4SLinus Torvalds buff->csum = 0; 2334bd37a088SWei Yongjun tp->snd_nxt = tp->write_seq; 23351da177e4SLinus Torvalds TCP_SKB_CB(buff)->seq = tp->write_seq++; 23361da177e4SLinus Torvalds TCP_SKB_CB(buff)->end_seq = tp->write_seq; 23371da177e4SLinus Torvalds 23381da177e4SLinus Torvalds /* Send it off. */ 23391da177e4SLinus Torvalds TCP_SKB_CB(buff)->when = tcp_time_stamp; 23401da177e4SLinus Torvalds tp->retrans_stamp = TCP_SKB_CB(buff)->when; 23411da177e4SLinus Torvalds skb_header_release(buff); 2342fe067e8aSDavid S. Miller __tcp_add_write_queue_tail(sk, buff); 23433ab224beSHideo Aoki sk->sk_wmem_queued += buff->truesize; 23443ab224beSHideo Aoki sk_mem_charge(sk, buff->truesize); 23451da177e4SLinus Torvalds tp->packets_out += tcp_skb_pcount(buff); 2346dfb4b9dcSDavid S. Miller tcp_transmit_skb(sk, buff, 1, GFP_KERNEL); 2347bd37a088SWei Yongjun 2348bd37a088SWei Yongjun /* We change tp->snd_nxt after the tcp_transmit_skb() call 2349bd37a088SWei Yongjun * in order to make this packet get counted in tcpOutSegs. 2350bd37a088SWei Yongjun */ 2351bd37a088SWei Yongjun tp->snd_nxt = tp->write_seq; 2352bd37a088SWei Yongjun tp->pushed_seq = tp->write_seq; 23531da177e4SLinus Torvalds TCP_INC_STATS(TCP_MIB_ACTIVEOPENS); 23541da177e4SLinus Torvalds 23551da177e4SLinus Torvalds /* Timer for repeating the SYN until an answer. */ 23563f421baaSArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 23573f421baaSArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto, TCP_RTO_MAX); 23581da177e4SLinus Torvalds return 0; 23591da177e4SLinus Torvalds } 23601da177e4SLinus Torvalds 23611da177e4SLinus Torvalds /* Send out a delayed ack, the caller does the policy checking 23621da177e4SLinus Torvalds * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check() 23631da177e4SLinus Torvalds * for details. 23641da177e4SLinus Torvalds */ 23651da177e4SLinus Torvalds void tcp_send_delayed_ack(struct sock *sk) 23661da177e4SLinus Torvalds { 2367463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 2368463c84b9SArnaldo Carvalho de Melo int ato = icsk->icsk_ack.ato; 23691da177e4SLinus Torvalds unsigned long timeout; 23701da177e4SLinus Torvalds 23711da177e4SLinus Torvalds if (ato > TCP_DELACK_MIN) { 2372463c84b9SArnaldo Carvalho de Melo const struct tcp_sock *tp = tcp_sk(sk); 23731da177e4SLinus Torvalds int max_ato = HZ/2; 23741da177e4SLinus Torvalds 2375463c84b9SArnaldo Carvalho de Melo if (icsk->icsk_ack.pingpong || (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)) 23761da177e4SLinus Torvalds max_ato = TCP_DELACK_MAX; 23771da177e4SLinus Torvalds 23781da177e4SLinus Torvalds /* Slow path, intersegment interval is "high". */ 23791da177e4SLinus Torvalds 23801da177e4SLinus Torvalds /* If some rtt estimate is known, use it to bound delayed ack. 2381463c84b9SArnaldo Carvalho de Melo * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements 23821da177e4SLinus Torvalds * directly. 23831da177e4SLinus Torvalds */ 23841da177e4SLinus Torvalds if (tp->srtt) { 23851da177e4SLinus Torvalds int rtt = max(tp->srtt>>3, TCP_DELACK_MIN); 23861da177e4SLinus Torvalds 23871da177e4SLinus Torvalds if (rtt < max_ato) 23881da177e4SLinus Torvalds max_ato = rtt; 23891da177e4SLinus Torvalds } 23901da177e4SLinus Torvalds 23911da177e4SLinus Torvalds ato = min(ato, max_ato); 23921da177e4SLinus Torvalds } 23931da177e4SLinus Torvalds 23941da177e4SLinus Torvalds /* Stay within the limit we were given */ 23951da177e4SLinus Torvalds timeout = jiffies + ato; 23961da177e4SLinus Torvalds 23971da177e4SLinus Torvalds /* Use new timeout only if there wasn't a older one earlier. */ 2398463c84b9SArnaldo Carvalho de Melo if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) { 23991da177e4SLinus Torvalds /* If delack timer was blocked or is about to expire, 24001da177e4SLinus Torvalds * send ACK now. 24011da177e4SLinus Torvalds */ 2402463c84b9SArnaldo Carvalho de Melo if (icsk->icsk_ack.blocked || 2403463c84b9SArnaldo Carvalho de Melo time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) { 24041da177e4SLinus Torvalds tcp_send_ack(sk); 24051da177e4SLinus Torvalds return; 24061da177e4SLinus Torvalds } 24071da177e4SLinus Torvalds 2408463c84b9SArnaldo Carvalho de Melo if (!time_before(timeout, icsk->icsk_ack.timeout)) 2409463c84b9SArnaldo Carvalho de Melo timeout = icsk->icsk_ack.timeout; 24101da177e4SLinus Torvalds } 2411463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER; 2412463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.timeout = timeout; 2413463c84b9SArnaldo Carvalho de Melo sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout); 24141da177e4SLinus Torvalds } 24151da177e4SLinus Torvalds 24161da177e4SLinus Torvalds /* This routine sends an ack and also updates the window. */ 24171da177e4SLinus Torvalds void tcp_send_ack(struct sock *sk) 24181da177e4SLinus Torvalds { 24191da177e4SLinus Torvalds /* If we have been reset, we may not send again. */ 24201da177e4SLinus Torvalds if (sk->sk_state != TCP_CLOSE) { 24211da177e4SLinus Torvalds struct sk_buff *buff; 24221da177e4SLinus Torvalds 24231da177e4SLinus Torvalds /* We are not putting this on the write queue, so 24241da177e4SLinus Torvalds * tcp_transmit_skb() will set the ownership to this 24251da177e4SLinus Torvalds * sock. 24261da177e4SLinus Torvalds */ 24271da177e4SLinus Torvalds buff = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC); 24281da177e4SLinus Torvalds if (buff == NULL) { 2429463c84b9SArnaldo Carvalho de Melo inet_csk_schedule_ack(sk); 2430463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN; 24313f421baaSArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK, 24323f421baaSArnaldo Carvalho de Melo TCP_DELACK_MAX, TCP_RTO_MAX); 24331da177e4SLinus Torvalds return; 24341da177e4SLinus Torvalds } 24351da177e4SLinus Torvalds 24361da177e4SLinus Torvalds /* Reserve space for headers and prepare control bits. */ 24371da177e4SLinus Torvalds skb_reserve(buff, MAX_TCP_HEADER); 24381da177e4SLinus Torvalds buff->csum = 0; 24391da177e4SLinus Torvalds TCP_SKB_CB(buff)->flags = TCPCB_FLAG_ACK; 24401da177e4SLinus Torvalds TCP_SKB_CB(buff)->sacked = 0; 24417967168cSHerbert Xu skb_shinfo(buff)->gso_segs = 1; 24427967168cSHerbert Xu skb_shinfo(buff)->gso_size = 0; 24437967168cSHerbert Xu skb_shinfo(buff)->gso_type = 0; 24441da177e4SLinus Torvalds 24451da177e4SLinus Torvalds /* Send it off, this clears delayed acks for us. */ 24469e412ba7SIlpo Järvinen TCP_SKB_CB(buff)->seq = TCP_SKB_CB(buff)->end_seq = tcp_acceptable_seq(sk); 24471da177e4SLinus Torvalds TCP_SKB_CB(buff)->when = tcp_time_stamp; 2448dfb4b9dcSDavid S. Miller tcp_transmit_skb(sk, buff, 0, GFP_ATOMIC); 24491da177e4SLinus Torvalds } 24501da177e4SLinus Torvalds } 24511da177e4SLinus Torvalds 24521da177e4SLinus Torvalds /* This routine sends a packet with an out of date sequence 24531da177e4SLinus Torvalds * number. It assumes the other end will try to ack it. 24541da177e4SLinus Torvalds * 24551da177e4SLinus Torvalds * Question: what should we make while urgent mode? 24561da177e4SLinus Torvalds * 4.4BSD forces sending single byte of data. We cannot send 24571da177e4SLinus Torvalds * out of window data, because we have SND.NXT==SND.MAX... 24581da177e4SLinus Torvalds * 24591da177e4SLinus Torvalds * Current solution: to send TWO zero-length segments in urgent mode: 24601da177e4SLinus Torvalds * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is 24611da177e4SLinus Torvalds * out-of-date with SND.UNA-1 to probe window. 24621da177e4SLinus Torvalds */ 24631da177e4SLinus Torvalds static int tcp_xmit_probe_skb(struct sock *sk, int urgent) 24641da177e4SLinus Torvalds { 24651da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 24661da177e4SLinus Torvalds struct sk_buff *skb; 24671da177e4SLinus Torvalds 24681da177e4SLinus Torvalds /* We don't queue it, tcp_transmit_skb() sets ownership. */ 24691da177e4SLinus Torvalds skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC); 24701da177e4SLinus Torvalds if (skb == NULL) 24711da177e4SLinus Torvalds return -1; 24721da177e4SLinus Torvalds 24731da177e4SLinus Torvalds /* Reserve space for headers and set control bits. */ 24741da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 24751da177e4SLinus Torvalds skb->csum = 0; 24761da177e4SLinus Torvalds TCP_SKB_CB(skb)->flags = TCPCB_FLAG_ACK; 2477*4828e7f4SIlpo Järvinen TCP_SKB_CB(skb)->sacked = 0; 24787967168cSHerbert Xu skb_shinfo(skb)->gso_segs = 1; 24797967168cSHerbert Xu skb_shinfo(skb)->gso_size = 0; 24807967168cSHerbert Xu skb_shinfo(skb)->gso_type = 0; 24811da177e4SLinus Torvalds 24821da177e4SLinus Torvalds /* Use a previous sequence. This should cause the other 24831da177e4SLinus Torvalds * end to send an ack. Don't queue or clone SKB, just 24841da177e4SLinus Torvalds * send it. 24851da177e4SLinus Torvalds */ 24861da177e4SLinus Torvalds TCP_SKB_CB(skb)->seq = urgent ? tp->snd_una : tp->snd_una - 1; 24871da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq; 24881da177e4SLinus Torvalds TCP_SKB_CB(skb)->when = tcp_time_stamp; 2489dfb4b9dcSDavid S. Miller return tcp_transmit_skb(sk, skb, 0, GFP_ATOMIC); 24901da177e4SLinus Torvalds } 24911da177e4SLinus Torvalds 24921da177e4SLinus Torvalds int tcp_write_wakeup(struct sock *sk) 24931da177e4SLinus Torvalds { 24941da177e4SLinus Torvalds if (sk->sk_state != TCP_CLOSE) { 24951da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 24961da177e4SLinus Torvalds struct sk_buff *skb; 24971da177e4SLinus Torvalds 2498fe067e8aSDavid S. Miller if ((skb = tcp_send_head(sk)) != NULL && 249990840defSIlpo Järvinen before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) { 25001da177e4SLinus Torvalds int err; 25011da177e4SLinus Torvalds unsigned int mss = tcp_current_mss(sk, 0); 250290840defSIlpo Järvinen unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 25031da177e4SLinus Torvalds 25041da177e4SLinus Torvalds if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq)) 25051da177e4SLinus Torvalds tp->pushed_seq = TCP_SKB_CB(skb)->end_seq; 25061da177e4SLinus Torvalds 25071da177e4SLinus Torvalds /* We are probing the opening of a window 25081da177e4SLinus Torvalds * but the window size is != 0 25091da177e4SLinus Torvalds * must have been a result SWS avoidance ( sender ) 25101da177e4SLinus Torvalds */ 25111da177e4SLinus Torvalds if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq || 25121da177e4SLinus Torvalds skb->len > mss) { 25131da177e4SLinus Torvalds seg_size = min(seg_size, mss); 25141da177e4SLinus Torvalds TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH; 2515846998aeSDavid S. Miller if (tcp_fragment(sk, skb, seg_size, mss)) 25161da177e4SLinus Torvalds return -1; 25171da177e4SLinus Torvalds } else if (!tcp_skb_pcount(skb)) 2518846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, skb, mss); 25191da177e4SLinus Torvalds 25201da177e4SLinus Torvalds TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH; 25211da177e4SLinus Torvalds TCP_SKB_CB(skb)->when = tcp_time_stamp; 2522dfb4b9dcSDavid S. Miller err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 252366f5fe62SIlpo Järvinen if (!err) 252466f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, skb); 25251da177e4SLinus Torvalds return err; 25261da177e4SLinus Torvalds } else { 25271da177e4SLinus Torvalds if (tp->urg_mode && 25281da177e4SLinus Torvalds between(tp->snd_up, tp->snd_una+1, tp->snd_una+0xFFFF)) 2529*4828e7f4SIlpo Järvinen tcp_xmit_probe_skb(sk, 1); 25301da177e4SLinus Torvalds return tcp_xmit_probe_skb(sk, 0); 25311da177e4SLinus Torvalds } 25321da177e4SLinus Torvalds } 25331da177e4SLinus Torvalds return -1; 25341da177e4SLinus Torvalds } 25351da177e4SLinus Torvalds 25361da177e4SLinus Torvalds /* A window probe timeout has occurred. If window is not closed send 25371da177e4SLinus Torvalds * a partial packet else a zero probe. 25381da177e4SLinus Torvalds */ 25391da177e4SLinus Torvalds void tcp_send_probe0(struct sock *sk) 25401da177e4SLinus Torvalds { 2541463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 25421da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 25431da177e4SLinus Torvalds int err; 25441da177e4SLinus Torvalds 25451da177e4SLinus Torvalds err = tcp_write_wakeup(sk); 25461da177e4SLinus Torvalds 2547fe067e8aSDavid S. Miller if (tp->packets_out || !tcp_send_head(sk)) { 25481da177e4SLinus Torvalds /* Cancel probe timer, if it is not required. */ 25496687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out = 0; 2550463c84b9SArnaldo Carvalho de Melo icsk->icsk_backoff = 0; 25511da177e4SLinus Torvalds return; 25521da177e4SLinus Torvalds } 25531da177e4SLinus Torvalds 25541da177e4SLinus Torvalds if (err <= 0) { 2555463c84b9SArnaldo Carvalho de Melo if (icsk->icsk_backoff < sysctl_tcp_retries2) 2556463c84b9SArnaldo Carvalho de Melo icsk->icsk_backoff++; 25576687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out++; 2558463c84b9SArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0, 25593f421baaSArnaldo Carvalho de Melo min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX), 25603f421baaSArnaldo Carvalho de Melo TCP_RTO_MAX); 25611da177e4SLinus Torvalds } else { 25621da177e4SLinus Torvalds /* If packet was not sent due to local congestion, 25636687e988SArnaldo Carvalho de Melo * do not backoff and do not remember icsk_probes_out. 25641da177e4SLinus Torvalds * Let local senders to fight for local resources. 25651da177e4SLinus Torvalds * 25661da177e4SLinus Torvalds * Use accumulated backoff yet. 25671da177e4SLinus Torvalds */ 25686687e988SArnaldo Carvalho de Melo if (!icsk->icsk_probes_out) 25696687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out = 1; 2570463c84b9SArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0, 2571463c84b9SArnaldo Carvalho de Melo min(icsk->icsk_rto << icsk->icsk_backoff, 25723f421baaSArnaldo Carvalho de Melo TCP_RESOURCE_PROBE_INTERVAL), 25733f421baaSArnaldo Carvalho de Melo TCP_RTO_MAX); 25741da177e4SLinus Torvalds } 25751da177e4SLinus Torvalds } 25761da177e4SLinus Torvalds 25771da177e4SLinus Torvalds EXPORT_SYMBOL(tcp_connect); 25781da177e4SLinus Torvalds EXPORT_SYMBOL(tcp_make_synack); 25791da177e4SLinus Torvalds EXPORT_SYMBOL(tcp_simple_retransmit); 25801da177e4SLinus Torvalds EXPORT_SYMBOL(tcp_sync_mss); 2581f4805edeSStephen Hemminger EXPORT_SYMBOL(sysctl_tcp_tso_win_divisor); 25825d424d5aSJohn Heffner EXPORT_SYMBOL(tcp_mtup_init); 2583