11da177e4SLinus Torvalds /* 21da177e4SLinus Torvalds * INET An implementation of the TCP/IP protocol suite for the LINUX 31da177e4SLinus Torvalds * operating system. INET is implemented using the BSD Socket 41da177e4SLinus Torvalds * interface as the means of communication with the user level. 51da177e4SLinus Torvalds * 61da177e4SLinus Torvalds * Implementation of the Transmission Control Protocol(TCP). 71da177e4SLinus Torvalds * 802c30a84SJesper Juhl * Authors: Ross Biro 91da177e4SLinus Torvalds * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG> 101da177e4SLinus Torvalds * Mark Evans, <evansmp@uhura.aston.ac.uk> 111da177e4SLinus Torvalds * Corey Minyard <wf-rch!minyard@relay.EU.net> 121da177e4SLinus Torvalds * Florian La Roche, <flla@stud.uni-sb.de> 131da177e4SLinus Torvalds * Charles Hedrick, <hedrick@klinzhai.rutgers.edu> 141da177e4SLinus Torvalds * Linus Torvalds, <torvalds@cs.helsinki.fi> 151da177e4SLinus Torvalds * Alan Cox, <gw4pts@gw4pts.ampr.org> 161da177e4SLinus Torvalds * Matthew Dillon, <dillon@apollo.west.oic.com> 171da177e4SLinus Torvalds * Arnt Gulbrandsen, <agulbra@nvg.unit.no> 181da177e4SLinus Torvalds * Jorge Cwik, <jorge@laser.satlink.net> 191da177e4SLinus Torvalds */ 201da177e4SLinus Torvalds 211da177e4SLinus Torvalds /* 221da177e4SLinus Torvalds * Changes: Pedro Roque : Retransmit queue handled by TCP. 231da177e4SLinus Torvalds * : Fragmentation on mtu decrease 241da177e4SLinus Torvalds * : Segment collapse on retransmit 251da177e4SLinus Torvalds * : AF independence 261da177e4SLinus Torvalds * 271da177e4SLinus Torvalds * Linus Torvalds : send_delayed_ack 281da177e4SLinus Torvalds * David S. Miller : Charge memory using the right skb 291da177e4SLinus Torvalds * during syn/ack processing. 301da177e4SLinus Torvalds * David S. Miller : Output engine completely rewritten. 311da177e4SLinus Torvalds * Andrea Arcangeli: SYNACK carry ts_recent in tsecr. 321da177e4SLinus Torvalds * Cacophonix Gaul : draft-minshall-nagle-01 331da177e4SLinus Torvalds * J Hadi Salim : ECN support 341da177e4SLinus Torvalds * 351da177e4SLinus Torvalds */ 361da177e4SLinus Torvalds 371da177e4SLinus Torvalds #include <net/tcp.h> 381da177e4SLinus Torvalds 391da177e4SLinus Torvalds #include <linux/compiler.h> 401da177e4SLinus Torvalds #include <linux/module.h> 411da177e4SLinus Torvalds 421da177e4SLinus Torvalds /* People can turn this off for buggy TCP's found in printers etc. */ 43ab32ea5dSBrian Haley int sysctl_tcp_retrans_collapse __read_mostly = 1; 441da177e4SLinus Torvalds 4515d99e02SRick Jones /* People can turn this on to work with those rare, broken TCPs that 4615d99e02SRick Jones * interpret the window field as a signed quantity. 4715d99e02SRick Jones */ 48ab32ea5dSBrian Haley int sysctl_tcp_workaround_signed_windows __read_mostly = 0; 4915d99e02SRick Jones 501da177e4SLinus Torvalds /* This limits the percentage of the congestion window which we 511da177e4SLinus Torvalds * will allow a single TSO frame to consume. Building TSO frames 521da177e4SLinus Torvalds * which are too large can cause TCP streams to be bursty. 531da177e4SLinus Torvalds */ 54ab32ea5dSBrian Haley int sysctl_tcp_tso_win_divisor __read_mostly = 3; 551da177e4SLinus Torvalds 56ab32ea5dSBrian Haley int sysctl_tcp_mtu_probing __read_mostly = 0; 57ab32ea5dSBrian Haley int sysctl_tcp_base_mss __read_mostly = 512; 585d424d5aSJohn Heffner 5935089bb2SDavid S. Miller /* By default, RFC2861 behavior. */ 60ab32ea5dSBrian Haley int sysctl_tcp_slow_start_after_idle __read_mostly = 1; 6135089bb2SDavid S. Miller 6266f5fe62SIlpo Järvinen static void tcp_event_new_data_sent(struct sock *sk, struct sk_buff *skb) 636ff03ac3SIlpo Järvinen { 646ff03ac3SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 6566f5fe62SIlpo Järvinen unsigned int prior_packets = tp->packets_out; 669e412ba7SIlpo Järvinen 67fe067e8aSDavid S. Miller tcp_advance_send_head(sk, skb); 681da177e4SLinus Torvalds tp->snd_nxt = TCP_SKB_CB(skb)->end_seq; 698512430eSIlpo Järvinen 708512430eSIlpo Järvinen /* Don't override Nagle indefinately with F-RTO */ 718512430eSIlpo Järvinen if (tp->frto_counter == 2) 728512430eSIlpo Järvinen tp->frto_counter = 3; 7366f5fe62SIlpo Järvinen 7466f5fe62SIlpo Järvinen tp->packets_out += tcp_skb_pcount(skb); 7566f5fe62SIlpo Järvinen if (!prior_packets) 7666f5fe62SIlpo Järvinen inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 7766f5fe62SIlpo Järvinen inet_csk(sk)->icsk_rto, TCP_RTO_MAX); 781da177e4SLinus Torvalds } 791da177e4SLinus Torvalds 801da177e4SLinus Torvalds /* SND.NXT, if window was not shrunk. 811da177e4SLinus Torvalds * If window has been shrunk, what should we make? It is not clear at all. 821da177e4SLinus Torvalds * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-( 831da177e4SLinus Torvalds * Anything in between SND.UNA...SND.UNA+SND.WND also can be already 841da177e4SLinus Torvalds * invalid. OK, let's make this for now: 851da177e4SLinus Torvalds */ 869e412ba7SIlpo Järvinen static inline __u32 tcp_acceptable_seq(struct sock *sk) 871da177e4SLinus Torvalds { 889e412ba7SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 899e412ba7SIlpo Järvinen 9090840defSIlpo Järvinen if (!before(tcp_wnd_end(tp), tp->snd_nxt)) 911da177e4SLinus Torvalds return tp->snd_nxt; 921da177e4SLinus Torvalds else 9390840defSIlpo Järvinen return tcp_wnd_end(tp); 941da177e4SLinus Torvalds } 951da177e4SLinus Torvalds 961da177e4SLinus Torvalds /* Calculate mss to advertise in SYN segment. 971da177e4SLinus Torvalds * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that: 981da177e4SLinus Torvalds * 991da177e4SLinus Torvalds * 1. It is independent of path mtu. 1001da177e4SLinus Torvalds * 2. Ideally, it is maximal possible segment size i.e. 65535-40. 1011da177e4SLinus Torvalds * 3. For IPv4 it is reasonable to calculate it from maximal MTU of 1021da177e4SLinus Torvalds * attached devices, because some buggy hosts are confused by 1031da177e4SLinus Torvalds * large MSS. 1041da177e4SLinus Torvalds * 4. We do not make 3, we advertise MSS, calculated from first 1051da177e4SLinus Torvalds * hop device mtu, but allow to raise it to ip_rt_min_advmss. 1061da177e4SLinus Torvalds * This may be overridden via information stored in routing table. 1071da177e4SLinus Torvalds * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible, 1081da177e4SLinus Torvalds * probably even Jumbo". 1091da177e4SLinus Torvalds */ 1101da177e4SLinus Torvalds static __u16 tcp_advertise_mss(struct sock *sk) 1111da177e4SLinus Torvalds { 1121da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 1131da177e4SLinus Torvalds struct dst_entry *dst = __sk_dst_get(sk); 1141da177e4SLinus Torvalds int mss = tp->advmss; 1151da177e4SLinus Torvalds 1161da177e4SLinus Torvalds if (dst && dst_metric(dst, RTAX_ADVMSS) < mss) { 1171da177e4SLinus Torvalds mss = dst_metric(dst, RTAX_ADVMSS); 1181da177e4SLinus Torvalds tp->advmss = mss; 1191da177e4SLinus Torvalds } 1201da177e4SLinus Torvalds 1211da177e4SLinus Torvalds return (__u16)mss; 1221da177e4SLinus Torvalds } 1231da177e4SLinus Torvalds 1241da177e4SLinus Torvalds /* RFC2861. Reset CWND after idle period longer RTO to "restart window". 1251da177e4SLinus Torvalds * This is the first part of cwnd validation mechanism. */ 126463c84b9SArnaldo Carvalho de Melo static void tcp_cwnd_restart(struct sock *sk, struct dst_entry *dst) 1271da177e4SLinus Torvalds { 128463c84b9SArnaldo Carvalho de Melo struct tcp_sock *tp = tcp_sk(sk); 1291da177e4SLinus Torvalds s32 delta = tcp_time_stamp - tp->lsndtime; 1301da177e4SLinus Torvalds u32 restart_cwnd = tcp_init_cwnd(tp, dst); 1311da177e4SLinus Torvalds u32 cwnd = tp->snd_cwnd; 1321da177e4SLinus Torvalds 1336687e988SArnaldo Carvalho de Melo tcp_ca_event(sk, CA_EVENT_CWND_RESTART); 1341da177e4SLinus Torvalds 1356687e988SArnaldo Carvalho de Melo tp->snd_ssthresh = tcp_current_ssthresh(sk); 1361da177e4SLinus Torvalds restart_cwnd = min(restart_cwnd, cwnd); 1371da177e4SLinus Torvalds 138463c84b9SArnaldo Carvalho de Melo while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd) 1391da177e4SLinus Torvalds cwnd >>= 1; 1401da177e4SLinus Torvalds tp->snd_cwnd = max(cwnd, restart_cwnd); 1411da177e4SLinus Torvalds tp->snd_cwnd_stamp = tcp_time_stamp; 1421da177e4SLinus Torvalds tp->snd_cwnd_used = 0; 1431da177e4SLinus Torvalds } 1441da177e4SLinus Torvalds 14540efc6faSStephen Hemminger static void tcp_event_data_sent(struct tcp_sock *tp, 1461da177e4SLinus Torvalds struct sk_buff *skb, struct sock *sk) 1471da177e4SLinus Torvalds { 148463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 149463c84b9SArnaldo Carvalho de Melo const u32 now = tcp_time_stamp; 1501da177e4SLinus Torvalds 15135089bb2SDavid S. Miller if (sysctl_tcp_slow_start_after_idle && 15235089bb2SDavid S. Miller (!tp->packets_out && (s32)(now - tp->lsndtime) > icsk->icsk_rto)) 153463c84b9SArnaldo Carvalho de Melo tcp_cwnd_restart(sk, __sk_dst_get(sk)); 1541da177e4SLinus Torvalds 1551da177e4SLinus Torvalds tp->lsndtime = now; 1561da177e4SLinus Torvalds 1571da177e4SLinus Torvalds /* If it is a reply for ato after last received 1581da177e4SLinus Torvalds * packet, enter pingpong mode. 1591da177e4SLinus Torvalds */ 160463c84b9SArnaldo Carvalho de Melo if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato) 161463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.pingpong = 1; 1621da177e4SLinus Torvalds } 1631da177e4SLinus Torvalds 16440efc6faSStephen Hemminger static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts) 1651da177e4SLinus Torvalds { 166463c84b9SArnaldo Carvalho de Melo tcp_dec_quickack_mode(sk, pkts); 167463c84b9SArnaldo Carvalho de Melo inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK); 1681da177e4SLinus Torvalds } 1691da177e4SLinus Torvalds 1701da177e4SLinus Torvalds /* Determine a window scaling and initial window to offer. 1711da177e4SLinus Torvalds * Based on the assumption that the given amount of space 1721da177e4SLinus Torvalds * will be offered. Store the results in the tp structure. 1731da177e4SLinus Torvalds * NOTE: for smooth operation initial space offering should 1741da177e4SLinus Torvalds * be a multiple of mss if possible. We assume here that mss >= 1. 1751da177e4SLinus Torvalds * This MUST be enforced by all callers. 1761da177e4SLinus Torvalds */ 1771da177e4SLinus Torvalds void tcp_select_initial_window(int __space, __u32 mss, 1781da177e4SLinus Torvalds __u32 *rcv_wnd, __u32 *window_clamp, 1791da177e4SLinus Torvalds int wscale_ok, __u8 *rcv_wscale) 1801da177e4SLinus Torvalds { 1811da177e4SLinus Torvalds unsigned int space = (__space < 0 ? 0 : __space); 1821da177e4SLinus Torvalds 1831da177e4SLinus Torvalds /* If no clamp set the clamp to the max possible scaled window */ 1841da177e4SLinus Torvalds if (*window_clamp == 0) 1851da177e4SLinus Torvalds (*window_clamp) = (65535 << 14); 1861da177e4SLinus Torvalds space = min(*window_clamp, space); 1871da177e4SLinus Torvalds 1881da177e4SLinus Torvalds /* Quantize space offering to a multiple of mss if possible. */ 1891da177e4SLinus Torvalds if (space > mss) 1901da177e4SLinus Torvalds space = (space / mss) * mss; 1911da177e4SLinus Torvalds 1921da177e4SLinus Torvalds /* NOTE: offering an initial window larger than 32767 19315d99e02SRick Jones * will break some buggy TCP stacks. If the admin tells us 19415d99e02SRick Jones * it is likely we could be speaking with such a buggy stack 19515d99e02SRick Jones * we will truncate our initial window offering to 32K-1 19615d99e02SRick Jones * unless the remote has sent us a window scaling option, 19715d99e02SRick Jones * which we interpret as a sign the remote TCP is not 19815d99e02SRick Jones * misinterpreting the window field as a signed quantity. 1991da177e4SLinus Torvalds */ 20015d99e02SRick Jones if (sysctl_tcp_workaround_signed_windows) 2011da177e4SLinus Torvalds (*rcv_wnd) = min(space, MAX_TCP_WINDOW); 20215d99e02SRick Jones else 20315d99e02SRick Jones (*rcv_wnd) = space; 20415d99e02SRick Jones 2051da177e4SLinus Torvalds (*rcv_wscale) = 0; 2061da177e4SLinus Torvalds if (wscale_ok) { 2071da177e4SLinus Torvalds /* Set window scaling on max possible window 2081da177e4SLinus Torvalds * See RFC1323 for an explanation of the limit to 14 2091da177e4SLinus Torvalds */ 2101da177e4SLinus Torvalds space = max_t(u32, sysctl_tcp_rmem[2], sysctl_rmem_max); 211316c1592SStephen Hemminger space = min_t(u32, space, *window_clamp); 2121da177e4SLinus Torvalds while (space > 65535 && (*rcv_wscale) < 14) { 2131da177e4SLinus Torvalds space >>= 1; 2141da177e4SLinus Torvalds (*rcv_wscale)++; 2151da177e4SLinus Torvalds } 2161da177e4SLinus Torvalds } 2171da177e4SLinus Torvalds 2181da177e4SLinus Torvalds /* Set initial window to value enough for senders, 2196b251858SDavid S. Miller * following RFC2414. Senders, not following this RFC, 2201da177e4SLinus Torvalds * will be satisfied with 2. 2211da177e4SLinus Torvalds */ 2221da177e4SLinus Torvalds if (mss > (1 << *rcv_wscale)) { 22301ff367eSDavid S. Miller int init_cwnd = 4; 22401ff367eSDavid S. Miller if (mss > 1460 * 3) 2251da177e4SLinus Torvalds init_cwnd = 2; 22601ff367eSDavid S. Miller else if (mss > 1460) 22701ff367eSDavid S. Miller init_cwnd = 3; 2281da177e4SLinus Torvalds if (*rcv_wnd > init_cwnd * mss) 2291da177e4SLinus Torvalds *rcv_wnd = init_cwnd * mss; 2301da177e4SLinus Torvalds } 2311da177e4SLinus Torvalds 2321da177e4SLinus Torvalds /* Set the clamp no higher than max representable value */ 2331da177e4SLinus Torvalds (*window_clamp) = min(65535U << (*rcv_wscale), *window_clamp); 2341da177e4SLinus Torvalds } 2351da177e4SLinus Torvalds 2361da177e4SLinus Torvalds /* Chose a new window to advertise, update state in tcp_sock for the 2371da177e4SLinus Torvalds * socket, and return result with RFC1323 scaling applied. The return 2381da177e4SLinus Torvalds * value can be stuffed directly into th->window for an outgoing 2391da177e4SLinus Torvalds * frame. 2401da177e4SLinus Torvalds */ 24140efc6faSStephen Hemminger static u16 tcp_select_window(struct sock *sk) 2421da177e4SLinus Torvalds { 2431da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 2441da177e4SLinus Torvalds u32 cur_win = tcp_receive_window(tp); 2451da177e4SLinus Torvalds u32 new_win = __tcp_select_window(sk); 2461da177e4SLinus Torvalds 2471da177e4SLinus Torvalds /* Never shrink the offered window */ 2481da177e4SLinus Torvalds if (new_win < cur_win) { 2491da177e4SLinus Torvalds /* Danger Will Robinson! 2501da177e4SLinus Torvalds * Don't update rcv_wup/rcv_wnd here or else 2511da177e4SLinus Torvalds * we will not be able to advertise a zero 2521da177e4SLinus Torvalds * window in time. --DaveM 2531da177e4SLinus Torvalds * 2541da177e4SLinus Torvalds * Relax Will Robinson. 2551da177e4SLinus Torvalds */ 256607bfbf2SPatrick McHardy new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale); 2571da177e4SLinus Torvalds } 2581da177e4SLinus Torvalds tp->rcv_wnd = new_win; 2591da177e4SLinus Torvalds tp->rcv_wup = tp->rcv_nxt; 2601da177e4SLinus Torvalds 2611da177e4SLinus Torvalds /* Make sure we do not exceed the maximum possible 2621da177e4SLinus Torvalds * scaled window. 2631da177e4SLinus Torvalds */ 26415d99e02SRick Jones if (!tp->rx_opt.rcv_wscale && sysctl_tcp_workaround_signed_windows) 2651da177e4SLinus Torvalds new_win = min(new_win, MAX_TCP_WINDOW); 2661da177e4SLinus Torvalds else 2671da177e4SLinus Torvalds new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale)); 2681da177e4SLinus Torvalds 2691da177e4SLinus Torvalds /* RFC1323 scaling applied */ 2701da177e4SLinus Torvalds new_win >>= tp->rx_opt.rcv_wscale; 2711da177e4SLinus Torvalds 2721da177e4SLinus Torvalds /* If we advertise zero window, disable fast path. */ 2731da177e4SLinus Torvalds if (new_win == 0) 2741da177e4SLinus Torvalds tp->pred_flags = 0; 2751da177e4SLinus Torvalds 2761da177e4SLinus Torvalds return new_win; 2771da177e4SLinus Torvalds } 2781da177e4SLinus Torvalds 279056834d9SIlpo Järvinen static inline void TCP_ECN_send_synack(struct tcp_sock *tp, struct sk_buff *skb) 280bdf1ee5dSIlpo Järvinen { 281bdf1ee5dSIlpo Järvinen TCP_SKB_CB(skb)->flags &= ~TCPCB_FLAG_CWR; 282bdf1ee5dSIlpo Järvinen if (!(tp->ecn_flags & TCP_ECN_OK)) 283bdf1ee5dSIlpo Järvinen TCP_SKB_CB(skb)->flags &= ~TCPCB_FLAG_ECE; 284bdf1ee5dSIlpo Järvinen } 285bdf1ee5dSIlpo Järvinen 286bdf1ee5dSIlpo Järvinen static inline void TCP_ECN_send_syn(struct sock *sk, struct sk_buff *skb) 287bdf1ee5dSIlpo Järvinen { 288bdf1ee5dSIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 289bdf1ee5dSIlpo Järvinen 290bdf1ee5dSIlpo Järvinen tp->ecn_flags = 0; 291bdf1ee5dSIlpo Järvinen if (sysctl_tcp_ecn) { 292bdf1ee5dSIlpo Järvinen TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_ECE | TCPCB_FLAG_CWR; 293bdf1ee5dSIlpo Järvinen tp->ecn_flags = TCP_ECN_OK; 294bdf1ee5dSIlpo Järvinen } 295bdf1ee5dSIlpo Järvinen } 296bdf1ee5dSIlpo Järvinen 297bdf1ee5dSIlpo Järvinen static __inline__ void 298bdf1ee5dSIlpo Järvinen TCP_ECN_make_synack(struct request_sock *req, struct tcphdr *th) 299bdf1ee5dSIlpo Järvinen { 300bdf1ee5dSIlpo Järvinen if (inet_rsk(req)->ecn_ok) 301bdf1ee5dSIlpo Järvinen th->ece = 1; 302bdf1ee5dSIlpo Järvinen } 303bdf1ee5dSIlpo Järvinen 304bdf1ee5dSIlpo Järvinen static inline void TCP_ECN_send(struct sock *sk, struct sk_buff *skb, 305bdf1ee5dSIlpo Järvinen int tcp_header_len) 306bdf1ee5dSIlpo Järvinen { 307bdf1ee5dSIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 308bdf1ee5dSIlpo Järvinen 309bdf1ee5dSIlpo Järvinen if (tp->ecn_flags & TCP_ECN_OK) { 310bdf1ee5dSIlpo Järvinen /* Not-retransmitted data segment: set ECT and inject CWR. */ 311bdf1ee5dSIlpo Järvinen if (skb->len != tcp_header_len && 312bdf1ee5dSIlpo Järvinen !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) { 313bdf1ee5dSIlpo Järvinen INET_ECN_xmit(sk); 314bdf1ee5dSIlpo Järvinen if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) { 315bdf1ee5dSIlpo Järvinen tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR; 316bdf1ee5dSIlpo Järvinen tcp_hdr(skb)->cwr = 1; 317bdf1ee5dSIlpo Järvinen skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN; 318bdf1ee5dSIlpo Järvinen } 319bdf1ee5dSIlpo Järvinen } else { 320bdf1ee5dSIlpo Järvinen /* ACK or retransmitted segment: clear ECT|CE */ 321bdf1ee5dSIlpo Järvinen INET_ECN_dontxmit(sk); 322bdf1ee5dSIlpo Järvinen } 323bdf1ee5dSIlpo Järvinen if (tp->ecn_flags & TCP_ECN_DEMAND_CWR) 324bdf1ee5dSIlpo Järvinen tcp_hdr(skb)->ece = 1; 325bdf1ee5dSIlpo Järvinen } 326bdf1ee5dSIlpo Järvinen } 327bdf1ee5dSIlpo Järvinen 328e870a8efSIlpo Järvinen /* Constructs common control bits of non-data skb. If SYN/FIN is present, 329e870a8efSIlpo Järvinen * auto increment end seqno. 330e870a8efSIlpo Järvinen */ 331e870a8efSIlpo Järvinen static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags) 332e870a8efSIlpo Järvinen { 333e870a8efSIlpo Järvinen skb->csum = 0; 334e870a8efSIlpo Järvinen 335e870a8efSIlpo Järvinen TCP_SKB_CB(skb)->flags = flags; 336e870a8efSIlpo Järvinen TCP_SKB_CB(skb)->sacked = 0; 337e870a8efSIlpo Järvinen 338e870a8efSIlpo Järvinen skb_shinfo(skb)->gso_segs = 1; 339e870a8efSIlpo Järvinen skb_shinfo(skb)->gso_size = 0; 340e870a8efSIlpo Järvinen skb_shinfo(skb)->gso_type = 0; 341e870a8efSIlpo Järvinen 342e870a8efSIlpo Järvinen TCP_SKB_CB(skb)->seq = seq; 343e870a8efSIlpo Järvinen if (flags & (TCPCB_FLAG_SYN | TCPCB_FLAG_FIN)) 344e870a8efSIlpo Järvinen seq++; 345e870a8efSIlpo Järvinen TCP_SKB_CB(skb)->end_seq = seq; 346e870a8efSIlpo Järvinen } 347e870a8efSIlpo Järvinen 34833f5f57eSIlpo Järvinen static inline int tcp_urg_mode(const struct tcp_sock *tp) 34933f5f57eSIlpo Järvinen { 35033f5f57eSIlpo Järvinen return tp->snd_una != tp->snd_up; 35133f5f57eSIlpo Järvinen } 35233f5f57eSIlpo Järvinen 35333ad798cSAdam Langley #define OPTION_SACK_ADVERTISE (1 << 0) 35433ad798cSAdam Langley #define OPTION_TS (1 << 1) 35533ad798cSAdam Langley #define OPTION_MD5 (1 << 2) 35633ad798cSAdam Langley 35733ad798cSAdam Langley struct tcp_out_options { 35833ad798cSAdam Langley u8 options; /* bit field of OPTION_* */ 35933ad798cSAdam Langley u8 ws; /* window scale, 0 to disable */ 36033ad798cSAdam Langley u8 num_sack_blocks; /* number of SACK blocks to include */ 36133ad798cSAdam Langley u16 mss; /* 0 to disable */ 36233ad798cSAdam Langley __u32 tsval, tsecr; /* need to include OPTION_TS */ 36333ad798cSAdam Langley }; 36433ad798cSAdam Langley 365fd6149d3SIlpo Järvinen /* Beware: Something in the Internet is very sensitive to the ordering of 366fd6149d3SIlpo Järvinen * TCP options, we learned this through the hard way, so be careful here. 367fd6149d3SIlpo Järvinen * Luckily we can at least blame others for their non-compliance but from 368fd6149d3SIlpo Järvinen * inter-operatibility perspective it seems that we're somewhat stuck with 369fd6149d3SIlpo Järvinen * the ordering which we have been using if we want to keep working with 370fd6149d3SIlpo Järvinen * those broken things (not that it currently hurts anybody as there isn't 371fd6149d3SIlpo Järvinen * particular reason why the ordering would need to be changed). 372fd6149d3SIlpo Järvinen * 373fd6149d3SIlpo Järvinen * At least SACK_PERM as the first option is known to lead to a disaster 374fd6149d3SIlpo Järvinen * (but it may well be that other scenarios fail similarly). 375fd6149d3SIlpo Järvinen */ 37633ad798cSAdam Langley static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp, 37733ad798cSAdam Langley const struct tcp_out_options *opts, 37833ad798cSAdam Langley __u8 **md5_hash) { 37933ad798cSAdam Langley if (unlikely(OPTION_MD5 & opts->options)) { 38033ad798cSAdam Langley *ptr++ = htonl((TCPOPT_NOP << 24) | 38133ad798cSAdam Langley (TCPOPT_NOP << 16) | 38233ad798cSAdam Langley (TCPOPT_MD5SIG << 8) | 38333ad798cSAdam Langley TCPOLEN_MD5SIG); 38433ad798cSAdam Langley *md5_hash = (__u8 *)ptr; 38533ad798cSAdam Langley ptr += 4; 38633ad798cSAdam Langley } else { 38733ad798cSAdam Langley *md5_hash = NULL; 38833ad798cSAdam Langley } 38933ad798cSAdam Langley 390fd6149d3SIlpo Järvinen if (unlikely(opts->mss)) { 391fd6149d3SIlpo Järvinen *ptr++ = htonl((TCPOPT_MSS << 24) | 392fd6149d3SIlpo Järvinen (TCPOLEN_MSS << 16) | 393fd6149d3SIlpo Järvinen opts->mss); 394fd6149d3SIlpo Järvinen } 395fd6149d3SIlpo Järvinen 39633ad798cSAdam Langley if (likely(OPTION_TS & opts->options)) { 39733ad798cSAdam Langley if (unlikely(OPTION_SACK_ADVERTISE & opts->options)) { 39833ad798cSAdam Langley *ptr++ = htonl((TCPOPT_SACK_PERM << 24) | 39933ad798cSAdam Langley (TCPOLEN_SACK_PERM << 16) | 40033ad798cSAdam Langley (TCPOPT_TIMESTAMP << 8) | 40133ad798cSAdam Langley TCPOLEN_TIMESTAMP); 40233ad798cSAdam Langley } else { 403496c98dfSYOSHIFUJI Hideaki *ptr++ = htonl((TCPOPT_NOP << 24) | 40440efc6faSStephen Hemminger (TCPOPT_NOP << 16) | 40540efc6faSStephen Hemminger (TCPOPT_TIMESTAMP << 8) | 40640efc6faSStephen Hemminger TCPOLEN_TIMESTAMP); 40740efc6faSStephen Hemminger } 40833ad798cSAdam Langley *ptr++ = htonl(opts->tsval); 40933ad798cSAdam Langley *ptr++ = htonl(opts->tsecr); 41033ad798cSAdam Langley } 41133ad798cSAdam Langley 41233ad798cSAdam Langley if (unlikely(OPTION_SACK_ADVERTISE & opts->options && 41333ad798cSAdam Langley !(OPTION_TS & opts->options))) { 41433ad798cSAdam Langley *ptr++ = htonl((TCPOPT_NOP << 24) | 41533ad798cSAdam Langley (TCPOPT_NOP << 16) | 41633ad798cSAdam Langley (TCPOPT_SACK_PERM << 8) | 41733ad798cSAdam Langley TCPOLEN_SACK_PERM); 41833ad798cSAdam Langley } 41933ad798cSAdam Langley 42033ad798cSAdam Langley if (unlikely(opts->ws)) { 42133ad798cSAdam Langley *ptr++ = htonl((TCPOPT_NOP << 24) | 42233ad798cSAdam Langley (TCPOPT_WINDOW << 16) | 42333ad798cSAdam Langley (TCPOLEN_WINDOW << 8) | 42433ad798cSAdam Langley opts->ws); 42533ad798cSAdam Langley } 42633ad798cSAdam Langley 42733ad798cSAdam Langley if (unlikely(opts->num_sack_blocks)) { 42833ad798cSAdam Langley struct tcp_sack_block *sp = tp->rx_opt.dsack ? 42933ad798cSAdam Langley tp->duplicate_sack : tp->selective_acks; 43040efc6faSStephen Hemminger int this_sack; 43140efc6faSStephen Hemminger 43240efc6faSStephen Hemminger *ptr++ = htonl((TCPOPT_NOP << 24) | 43340efc6faSStephen Hemminger (TCPOPT_NOP << 16) | 43440efc6faSStephen Hemminger (TCPOPT_SACK << 8) | 43533ad798cSAdam Langley (TCPOLEN_SACK_BASE + (opts->num_sack_blocks * 43640efc6faSStephen Hemminger TCPOLEN_SACK_PERBLOCK))); 4372de979bdSStephen Hemminger 43833ad798cSAdam Langley for (this_sack = 0; this_sack < opts->num_sack_blocks; 43933ad798cSAdam Langley ++this_sack) { 44040efc6faSStephen Hemminger *ptr++ = htonl(sp[this_sack].start_seq); 44140efc6faSStephen Hemminger *ptr++ = htonl(sp[this_sack].end_seq); 44240efc6faSStephen Hemminger } 4432de979bdSStephen Hemminger 44440efc6faSStephen Hemminger if (tp->rx_opt.dsack) { 44540efc6faSStephen Hemminger tp->rx_opt.dsack = 0; 44675e3d8dbSIlpo Järvinen tp->rx_opt.eff_sacks = tp->rx_opt.num_sacks; 44740efc6faSStephen Hemminger } 44840efc6faSStephen Hemminger } 44940efc6faSStephen Hemminger } 45040efc6faSStephen Hemminger 45133ad798cSAdam Langley static unsigned tcp_syn_options(struct sock *sk, struct sk_buff *skb, 45233ad798cSAdam Langley struct tcp_out_options *opts, 45333ad798cSAdam Langley struct tcp_md5sig_key **md5) { 45433ad798cSAdam Langley struct tcp_sock *tp = tcp_sk(sk); 45533ad798cSAdam Langley unsigned size = 0; 45633ad798cSAdam Langley 457cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 45833ad798cSAdam Langley *md5 = tp->af_specific->md5_lookup(sk, sk); 45933ad798cSAdam Langley if (*md5) { 46033ad798cSAdam Langley opts->options |= OPTION_MD5; 46133ad798cSAdam Langley size += TCPOLEN_MD5SIG_ALIGNED; 462cfb6eeb4SYOSHIFUJI Hideaki } 46333ad798cSAdam Langley #else 46433ad798cSAdam Langley *md5 = NULL; 465cfb6eeb4SYOSHIFUJI Hideaki #endif 46633ad798cSAdam Langley 46733ad798cSAdam Langley /* We always get an MSS option. The option bytes which will be seen in 46833ad798cSAdam Langley * normal data packets should timestamps be used, must be in the MSS 46933ad798cSAdam Langley * advertised. But we subtract them from tp->mss_cache so that 47033ad798cSAdam Langley * calculations in tcp_sendmsg are simpler etc. So account for this 47133ad798cSAdam Langley * fact here if necessary. If we don't do this correctly, as a 47233ad798cSAdam Langley * receiver we won't recognize data packets as being full sized when we 47333ad798cSAdam Langley * should, and thus we won't abide by the delayed ACK rules correctly. 47433ad798cSAdam Langley * SACKs don't matter, we never delay an ACK when we have any of those 47533ad798cSAdam Langley * going out. */ 47633ad798cSAdam Langley opts->mss = tcp_advertise_mss(sk); 47733ad798cSAdam Langley size += TCPOLEN_MSS_ALIGNED; 47833ad798cSAdam Langley 47933ad798cSAdam Langley if (likely(sysctl_tcp_timestamps && *md5 == NULL)) { 48033ad798cSAdam Langley opts->options |= OPTION_TS; 48133ad798cSAdam Langley opts->tsval = TCP_SKB_CB(skb)->when; 48233ad798cSAdam Langley opts->tsecr = tp->rx_opt.ts_recent; 48333ad798cSAdam Langley size += TCPOLEN_TSTAMP_ALIGNED; 48433ad798cSAdam Langley } 48533ad798cSAdam Langley if (likely(sysctl_tcp_window_scaling)) { 48633ad798cSAdam Langley opts->ws = tp->rx_opt.rcv_wscale; 4877982d5e1SPhilip Love if (likely(opts->ws)) 48833ad798cSAdam Langley size += TCPOLEN_WSCALE_ALIGNED; 48933ad798cSAdam Langley } 49033ad798cSAdam Langley if (likely(sysctl_tcp_sack)) { 49133ad798cSAdam Langley opts->options |= OPTION_SACK_ADVERTISE; 492b32d1310SDavid S. Miller if (unlikely(!(OPTION_TS & opts->options))) 49333ad798cSAdam Langley size += TCPOLEN_SACKPERM_ALIGNED; 49433ad798cSAdam Langley } 49533ad798cSAdam Langley 49633ad798cSAdam Langley return size; 49733ad798cSAdam Langley } 49833ad798cSAdam Langley 49933ad798cSAdam Langley static unsigned tcp_synack_options(struct sock *sk, 50033ad798cSAdam Langley struct request_sock *req, 50133ad798cSAdam Langley unsigned mss, struct sk_buff *skb, 50233ad798cSAdam Langley struct tcp_out_options *opts, 50333ad798cSAdam Langley struct tcp_md5sig_key **md5) { 50433ad798cSAdam Langley unsigned size = 0; 50533ad798cSAdam Langley struct inet_request_sock *ireq = inet_rsk(req); 50633ad798cSAdam Langley char doing_ts; 50733ad798cSAdam Langley 50833ad798cSAdam Langley #ifdef CONFIG_TCP_MD5SIG 50933ad798cSAdam Langley *md5 = tcp_rsk(req)->af_specific->md5_lookup(sk, req); 51033ad798cSAdam Langley if (*md5) { 51133ad798cSAdam Langley opts->options |= OPTION_MD5; 51233ad798cSAdam Langley size += TCPOLEN_MD5SIG_ALIGNED; 51333ad798cSAdam Langley } 51433ad798cSAdam Langley #else 51533ad798cSAdam Langley *md5 = NULL; 51633ad798cSAdam Langley #endif 51733ad798cSAdam Langley 51833ad798cSAdam Langley /* we can't fit any SACK blocks in a packet with MD5 + TS 51933ad798cSAdam Langley options. There was discussion about disabling SACK rather than TS in 52033ad798cSAdam Langley order to fit in better with old, buggy kernels, but that was deemed 52133ad798cSAdam Langley to be unnecessary. */ 52233ad798cSAdam Langley doing_ts = ireq->tstamp_ok && !(*md5 && ireq->sack_ok); 52333ad798cSAdam Langley 52433ad798cSAdam Langley opts->mss = mss; 52533ad798cSAdam Langley size += TCPOLEN_MSS_ALIGNED; 52633ad798cSAdam Langley 52733ad798cSAdam Langley if (likely(ireq->wscale_ok)) { 52833ad798cSAdam Langley opts->ws = ireq->rcv_wscale; 5297982d5e1SPhilip Love if (likely(opts->ws)) 53033ad798cSAdam Langley size += TCPOLEN_WSCALE_ALIGNED; 53133ad798cSAdam Langley } 53233ad798cSAdam Langley if (likely(doing_ts)) { 53333ad798cSAdam Langley opts->options |= OPTION_TS; 53433ad798cSAdam Langley opts->tsval = TCP_SKB_CB(skb)->when; 53533ad798cSAdam Langley opts->tsecr = req->ts_recent; 53633ad798cSAdam Langley size += TCPOLEN_TSTAMP_ALIGNED; 53733ad798cSAdam Langley } 53833ad798cSAdam Langley if (likely(ireq->sack_ok)) { 53933ad798cSAdam Langley opts->options |= OPTION_SACK_ADVERTISE; 54033ad798cSAdam Langley if (unlikely(!doing_ts)) 54133ad798cSAdam Langley size += TCPOLEN_SACKPERM_ALIGNED; 54233ad798cSAdam Langley } 54333ad798cSAdam Langley 54433ad798cSAdam Langley return size; 54533ad798cSAdam Langley } 54633ad798cSAdam Langley 54733ad798cSAdam Langley static unsigned tcp_established_options(struct sock *sk, struct sk_buff *skb, 54833ad798cSAdam Langley struct tcp_out_options *opts, 54933ad798cSAdam Langley struct tcp_md5sig_key **md5) { 55033ad798cSAdam Langley struct tcp_skb_cb *tcb = skb ? TCP_SKB_CB(skb) : NULL; 55133ad798cSAdam Langley struct tcp_sock *tp = tcp_sk(sk); 55233ad798cSAdam Langley unsigned size = 0; 55333ad798cSAdam Langley 55433ad798cSAdam Langley #ifdef CONFIG_TCP_MD5SIG 55533ad798cSAdam Langley *md5 = tp->af_specific->md5_lookup(sk, sk); 55633ad798cSAdam Langley if (unlikely(*md5)) { 55733ad798cSAdam Langley opts->options |= OPTION_MD5; 55833ad798cSAdam Langley size += TCPOLEN_MD5SIG_ALIGNED; 55933ad798cSAdam Langley } 56033ad798cSAdam Langley #else 56133ad798cSAdam Langley *md5 = NULL; 56233ad798cSAdam Langley #endif 56333ad798cSAdam Langley 56433ad798cSAdam Langley if (likely(tp->rx_opt.tstamp_ok)) { 56533ad798cSAdam Langley opts->options |= OPTION_TS; 56633ad798cSAdam Langley opts->tsval = tcb ? tcb->when : 0; 56733ad798cSAdam Langley opts->tsecr = tp->rx_opt.ts_recent; 56833ad798cSAdam Langley size += TCPOLEN_TSTAMP_ALIGNED; 56933ad798cSAdam Langley } 57033ad798cSAdam Langley 57133ad798cSAdam Langley if (unlikely(tp->rx_opt.eff_sacks)) { 57233ad798cSAdam Langley const unsigned remaining = MAX_TCP_OPTION_SPACE - size; 57333ad798cSAdam Langley opts->num_sack_blocks = 57433ad798cSAdam Langley min_t(unsigned, tp->rx_opt.eff_sacks, 57533ad798cSAdam Langley (remaining - TCPOLEN_SACK_BASE_ALIGNED) / 57633ad798cSAdam Langley TCPOLEN_SACK_PERBLOCK); 57733ad798cSAdam Langley size += TCPOLEN_SACK_BASE_ALIGNED + 57833ad798cSAdam Langley opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK; 57933ad798cSAdam Langley } 58033ad798cSAdam Langley 58133ad798cSAdam Langley return size; 58240efc6faSStephen Hemminger } 5831da177e4SLinus Torvalds 5841da177e4SLinus Torvalds /* This routine actually transmits TCP packets queued in by 5851da177e4SLinus Torvalds * tcp_do_sendmsg(). This is used by both the initial 5861da177e4SLinus Torvalds * transmission and possible later retransmissions. 5871da177e4SLinus Torvalds * All SKB's seen here are completely headerless. It is our 5881da177e4SLinus Torvalds * job to build the TCP header, and pass the packet down to 5891da177e4SLinus Torvalds * IP so it can do the same plus pass the packet off to the 5901da177e4SLinus Torvalds * device. 5911da177e4SLinus Torvalds * 5921da177e4SLinus Torvalds * We are working here with either a clone of the original 5931da177e4SLinus Torvalds * SKB, or a fresh unique copy made by the retransmit engine. 5941da177e4SLinus Torvalds */ 595056834d9SIlpo Järvinen static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it, 596056834d9SIlpo Järvinen gfp_t gfp_mask) 5971da177e4SLinus Torvalds { 5986687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 599dfb4b9dcSDavid S. Miller struct inet_sock *inet; 600dfb4b9dcSDavid S. Miller struct tcp_sock *tp; 601dfb4b9dcSDavid S. Miller struct tcp_skb_cb *tcb; 60233ad798cSAdam Langley struct tcp_out_options opts; 60333ad798cSAdam Langley unsigned tcp_options_size, tcp_header_size; 604cfb6eeb4SYOSHIFUJI Hideaki struct tcp_md5sig_key *md5; 605cfb6eeb4SYOSHIFUJI Hideaki __u8 *md5_hash_location; 6061da177e4SLinus Torvalds struct tcphdr *th; 6071da177e4SLinus Torvalds int err; 6081da177e4SLinus Torvalds 609dfb4b9dcSDavid S. Miller BUG_ON(!skb || !tcp_skb_pcount(skb)); 610dfb4b9dcSDavid S. Miller 611dfb4b9dcSDavid S. Miller /* If congestion control is doing timestamping, we must 612dfb4b9dcSDavid S. Miller * take such a timestamp before we potentially clone/copy. 613dfb4b9dcSDavid S. Miller */ 614164891aaSStephen Hemminger if (icsk->icsk_ca_ops->flags & TCP_CONG_RTT_STAMP) 615dfb4b9dcSDavid S. Miller __net_timestamp(skb); 616dfb4b9dcSDavid S. Miller 617dfb4b9dcSDavid S. Miller if (likely(clone_it)) { 618dfb4b9dcSDavid S. Miller if (unlikely(skb_cloned(skb))) 619dfb4b9dcSDavid S. Miller skb = pskb_copy(skb, gfp_mask); 620dfb4b9dcSDavid S. Miller else 621dfb4b9dcSDavid S. Miller skb = skb_clone(skb, gfp_mask); 622dfb4b9dcSDavid S. Miller if (unlikely(!skb)) 623dfb4b9dcSDavid S. Miller return -ENOBUFS; 624dfb4b9dcSDavid S. Miller } 625dfb4b9dcSDavid S. Miller 626dfb4b9dcSDavid S. Miller inet = inet_sk(sk); 627dfb4b9dcSDavid S. Miller tp = tcp_sk(sk); 628dfb4b9dcSDavid S. Miller tcb = TCP_SKB_CB(skb); 62933ad798cSAdam Langley memset(&opts, 0, sizeof(opts)); 6301da177e4SLinus Torvalds 63133ad798cSAdam Langley if (unlikely(tcb->flags & TCPCB_FLAG_SYN)) 63233ad798cSAdam Langley tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5); 63333ad798cSAdam Langley else 63433ad798cSAdam Langley tcp_options_size = tcp_established_options(sk, skb, &opts, 63533ad798cSAdam Langley &md5); 63633ad798cSAdam Langley tcp_header_size = tcp_options_size + sizeof(struct tcphdr); 6371da177e4SLinus Torvalds 638317a76f9SStephen Hemminger if (tcp_packets_in_flight(tp) == 0) 6396687e988SArnaldo Carvalho de Melo tcp_ca_event(sk, CA_EVENT_TX_START); 6401da177e4SLinus Torvalds 641aa8223c7SArnaldo Carvalho de Melo skb_push(skb, tcp_header_size); 642aa8223c7SArnaldo Carvalho de Melo skb_reset_transport_header(skb); 643e89862f4SDavid S. Miller skb_set_owner_w(skb, sk); 6441da177e4SLinus Torvalds 6451da177e4SLinus Torvalds /* Build TCP header and checksum it. */ 646aa8223c7SArnaldo Carvalho de Melo th = tcp_hdr(skb); 6471da177e4SLinus Torvalds th->source = inet->sport; 6481da177e4SLinus Torvalds th->dest = inet->dport; 6491da177e4SLinus Torvalds th->seq = htonl(tcb->seq); 6501da177e4SLinus Torvalds th->ack_seq = htonl(tp->rcv_nxt); 651df7a3b07SAl Viro *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) | 652dfb4b9dcSDavid S. Miller tcb->flags); 653dfb4b9dcSDavid S. Miller 654dfb4b9dcSDavid S. Miller if (unlikely(tcb->flags & TCPCB_FLAG_SYN)) { 6551da177e4SLinus Torvalds /* RFC1323: The window in SYN & SYN/ACK segments 6561da177e4SLinus Torvalds * is never scaled. 6571da177e4SLinus Torvalds */ 658600ff0c2SIlpo Järvinen th->window = htons(min(tp->rcv_wnd, 65535U)); 6591da177e4SLinus Torvalds } else { 6601da177e4SLinus Torvalds th->window = htons(tcp_select_window(sk)); 6611da177e4SLinus Torvalds } 6621da177e4SLinus Torvalds th->check = 0; 6631da177e4SLinus Torvalds th->urg_ptr = 0; 6641da177e4SLinus Torvalds 66533f5f57eSIlpo Järvinen /* The urg_mode check is necessary during a below snd_una win probe */ 66633f5f57eSIlpo Järvinen if (unlikely(tcp_urg_mode(tp) && 667dfb4b9dcSDavid S. Miller between(tp->snd_up, tcb->seq + 1, tcb->seq + 0xFFFF))) { 6681da177e4SLinus Torvalds th->urg_ptr = htons(tp->snd_up - tcb->seq); 6691da177e4SLinus Torvalds th->urg = 1; 6701da177e4SLinus Torvalds } 6711da177e4SLinus Torvalds 67233ad798cSAdam Langley tcp_options_write((__be32 *)(th + 1), tp, &opts, &md5_hash_location); 67333ad798cSAdam Langley if (likely((tcb->flags & TCPCB_FLAG_SYN) == 0)) 6749e412ba7SIlpo Järvinen TCP_ECN_send(sk, skb, tcp_header_size); 675dfb4b9dcSDavid S. Miller 676cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 677cfb6eeb4SYOSHIFUJI Hideaki /* Calculate the MD5 hash, as we have all we need now */ 678cfb6eeb4SYOSHIFUJI Hideaki if (md5) { 67933ad798cSAdam Langley sk->sk_route_caps &= ~NETIF_F_GSO_MASK; 680cfb6eeb4SYOSHIFUJI Hideaki tp->af_specific->calc_md5_hash(md5_hash_location, 68149a72dfbSAdam Langley md5, sk, NULL, skb); 682cfb6eeb4SYOSHIFUJI Hideaki } 683cfb6eeb4SYOSHIFUJI Hideaki #endif 684cfb6eeb4SYOSHIFUJI Hideaki 6858292a17aSArnaldo Carvalho de Melo icsk->icsk_af_ops->send_check(sk, skb->len, skb); 6861da177e4SLinus Torvalds 687dfb4b9dcSDavid S. Miller if (likely(tcb->flags & TCPCB_FLAG_ACK)) 688fc6415bcSDavid S. Miller tcp_event_ack_sent(sk, tcp_skb_pcount(skb)); 6891da177e4SLinus Torvalds 6901da177e4SLinus Torvalds if (skb->len != tcp_header_size) 6911da177e4SLinus Torvalds tcp_event_data_sent(tp, skb, sk); 6921da177e4SLinus Torvalds 693bd37a088SWei Yongjun if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq) 69481cc8a75SPavel Emelyanov TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTSEGS); 6951da177e4SLinus Torvalds 696e89862f4SDavid S. Miller err = icsk->icsk_af_ops->queue_xmit(skb, 0); 69783de47cdSHua Zhong if (likely(err <= 0)) 6981da177e4SLinus Torvalds return err; 6991da177e4SLinus Torvalds 7003cfe3baaSIlpo Järvinen tcp_enter_cwr(sk, 1); 7011da177e4SLinus Torvalds 702b9df3cb8SGerrit Renker return net_xmit_eval(err); 7031da177e4SLinus Torvalds } 7041da177e4SLinus Torvalds 7051da177e4SLinus Torvalds /* This routine just queue's the buffer 7061da177e4SLinus Torvalds * 7071da177e4SLinus Torvalds * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames, 7081da177e4SLinus Torvalds * otherwise socket can stall. 7091da177e4SLinus Torvalds */ 7101da177e4SLinus Torvalds static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb) 7111da177e4SLinus Torvalds { 7121da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 7131da177e4SLinus Torvalds 7141da177e4SLinus Torvalds /* Advance write_seq and place onto the write_queue. */ 7151da177e4SLinus Torvalds tp->write_seq = TCP_SKB_CB(skb)->end_seq; 7161da177e4SLinus Torvalds skb_header_release(skb); 717fe067e8aSDavid S. Miller tcp_add_write_queue_tail(sk, skb); 7183ab224beSHideo Aoki sk->sk_wmem_queued += skb->truesize; 7193ab224beSHideo Aoki sk_mem_charge(sk, skb->truesize); 7201da177e4SLinus Torvalds } 7211da177e4SLinus Torvalds 722056834d9SIlpo Järvinen static void tcp_set_skb_tso_segs(struct sock *sk, struct sk_buff *skb, 723056834d9SIlpo Järvinen unsigned int mss_now) 724f6302d1dSDavid S. Miller { 725bcd76111SHerbert Xu if (skb->len <= mss_now || !sk_can_gso(sk)) { 726f6302d1dSDavid S. Miller /* Avoid the costly divide in the normal 727f6302d1dSDavid S. Miller * non-TSO case. 728f6302d1dSDavid S. Miller */ 7297967168cSHerbert Xu skb_shinfo(skb)->gso_segs = 1; 7307967168cSHerbert Xu skb_shinfo(skb)->gso_size = 0; 7317967168cSHerbert Xu skb_shinfo(skb)->gso_type = 0; 732f6302d1dSDavid S. Miller } else { 733356f89e1SIlpo Järvinen skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss_now); 7347967168cSHerbert Xu skb_shinfo(skb)->gso_size = mss_now; 735bcd76111SHerbert Xu skb_shinfo(skb)->gso_type = sk->sk_gso_type; 7361da177e4SLinus Torvalds } 7371da177e4SLinus Torvalds } 7381da177e4SLinus Torvalds 73991fed7a1SIlpo Järvinen /* When a modification to fackets out becomes necessary, we need to check 74068f8353bSIlpo Järvinen * skb is counted to fackets_out or not. 74191fed7a1SIlpo Järvinen */ 742a47e5a98SIlpo Järvinen static void tcp_adjust_fackets_out(struct sock *sk, struct sk_buff *skb, 74391fed7a1SIlpo Järvinen int decr) 74491fed7a1SIlpo Järvinen { 745a47e5a98SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 746a47e5a98SIlpo Järvinen 747dc86967bSIlpo Järvinen if (!tp->sacked_out || tcp_is_reno(tp)) 74891fed7a1SIlpo Järvinen return; 74991fed7a1SIlpo Järvinen 7506859d494SIlpo Järvinen if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq)) 75191fed7a1SIlpo Järvinen tp->fackets_out -= decr; 75291fed7a1SIlpo Järvinen } 75391fed7a1SIlpo Järvinen 7541da177e4SLinus Torvalds /* Function to create two new TCP segments. Shrinks the given segment 7551da177e4SLinus Torvalds * to the specified size and appends a new segment with the rest of the 7561da177e4SLinus Torvalds * packet to the list. This won't be called frequently, I hope. 7571da177e4SLinus Torvalds * Remember, these are still headerless SKBs at this point. 7581da177e4SLinus Torvalds */ 759056834d9SIlpo Järvinen int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len, 760056834d9SIlpo Järvinen unsigned int mss_now) 7611da177e4SLinus Torvalds { 7621da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 7631da177e4SLinus Torvalds struct sk_buff *buff; 7646475be16SDavid S. Miller int nsize, old_factor; 765b60b49eaSHerbert Xu int nlen; 7661da177e4SLinus Torvalds u16 flags; 7671da177e4SLinus Torvalds 768b2cc99f0SHerbert Xu BUG_ON(len > skb->len); 7696a438bbeSStephen Hemminger 770ef9da47cSIlpo Järvinen tcp_clear_retrans_hints_partial(tp); 7711da177e4SLinus Torvalds nsize = skb_headlen(skb) - len; 7721da177e4SLinus Torvalds if (nsize < 0) 7731da177e4SLinus Torvalds nsize = 0; 7741da177e4SLinus Torvalds 7751da177e4SLinus Torvalds if (skb_cloned(skb) && 7761da177e4SLinus Torvalds skb_is_nonlinear(skb) && 7771da177e4SLinus Torvalds pskb_expand_head(skb, 0, 0, GFP_ATOMIC)) 7781da177e4SLinus Torvalds return -ENOMEM; 7791da177e4SLinus Torvalds 7801da177e4SLinus Torvalds /* Get a new skb... force flag on. */ 7811da177e4SLinus Torvalds buff = sk_stream_alloc_skb(sk, nsize, GFP_ATOMIC); 7821da177e4SLinus Torvalds if (buff == NULL) 7831da177e4SLinus Torvalds return -ENOMEM; /* We'll just try again later. */ 784ef5cb973SHerbert Xu 7853ab224beSHideo Aoki sk->sk_wmem_queued += buff->truesize; 7863ab224beSHideo Aoki sk_mem_charge(sk, buff->truesize); 787b60b49eaSHerbert Xu nlen = skb->len - len - nsize; 788b60b49eaSHerbert Xu buff->truesize += nlen; 789b60b49eaSHerbert Xu skb->truesize -= nlen; 7901da177e4SLinus Torvalds 7911da177e4SLinus Torvalds /* Correct the sequence numbers. */ 7921da177e4SLinus Torvalds TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; 7931da177e4SLinus Torvalds TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; 7941da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; 7951da177e4SLinus Torvalds 7961da177e4SLinus Torvalds /* PSH and FIN should only be set in the second packet. */ 7971da177e4SLinus Torvalds flags = TCP_SKB_CB(skb)->flags; 7981da177e4SLinus Torvalds TCP_SKB_CB(skb)->flags = flags & ~(TCPCB_FLAG_FIN | TCPCB_FLAG_PSH); 7991da177e4SLinus Torvalds TCP_SKB_CB(buff)->flags = flags; 800e14c3cafSHerbert Xu TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked; 8011da177e4SLinus Torvalds 80284fa7933SPatrick McHardy if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) { 8031da177e4SLinus Torvalds /* Copy and checksum data tail into the new buffer. */ 804056834d9SIlpo Järvinen buff->csum = csum_partial_copy_nocheck(skb->data + len, 805056834d9SIlpo Järvinen skb_put(buff, nsize), 8061da177e4SLinus Torvalds nsize, 0); 8071da177e4SLinus Torvalds 8081da177e4SLinus Torvalds skb_trim(skb, len); 8091da177e4SLinus Torvalds 8101da177e4SLinus Torvalds skb->csum = csum_block_sub(skb->csum, buff->csum, len); 8111da177e4SLinus Torvalds } else { 81284fa7933SPatrick McHardy skb->ip_summed = CHECKSUM_PARTIAL; 8131da177e4SLinus Torvalds skb_split(skb, buff, len); 8141da177e4SLinus Torvalds } 8151da177e4SLinus Torvalds 8161da177e4SLinus Torvalds buff->ip_summed = skb->ip_summed; 8171da177e4SLinus Torvalds 8181da177e4SLinus Torvalds /* Looks stupid, but our code really uses when of 8191da177e4SLinus Torvalds * skbs, which it never sent before. --ANK 8201da177e4SLinus Torvalds */ 8211da177e4SLinus Torvalds TCP_SKB_CB(buff)->when = TCP_SKB_CB(skb)->when; 822a61bbcf2SPatrick McHardy buff->tstamp = skb->tstamp; 8231da177e4SLinus Torvalds 8246475be16SDavid S. Miller old_factor = tcp_skb_pcount(skb); 8256475be16SDavid S. Miller 8261da177e4SLinus Torvalds /* Fix up tso_factor for both original and new SKB. */ 827846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, skb, mss_now); 828846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, buff, mss_now); 8291da177e4SLinus Torvalds 8306475be16SDavid S. Miller /* If this packet has been sent out already, we must 8316475be16SDavid S. Miller * adjust the various packet counters. 8326475be16SDavid S. Miller */ 833cf0b450cSHerbert Xu if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) { 8346475be16SDavid S. Miller int diff = old_factor - tcp_skb_pcount(skb) - 8356475be16SDavid S. Miller tcp_skb_pcount(buff); 8361da177e4SLinus Torvalds 8376475be16SDavid S. Miller tp->packets_out -= diff; 838e14c3cafSHerbert Xu 839e14c3cafSHerbert Xu if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 840e14c3cafSHerbert Xu tp->sacked_out -= diff; 841e14c3cafSHerbert Xu if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) 842e14c3cafSHerbert Xu tp->retrans_out -= diff; 843e14c3cafSHerbert Xu 844b5860bbaSIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST) 8456475be16SDavid S. Miller tp->lost_out -= diff; 84683ca28beSHerbert Xu 84783ca28beSHerbert Xu /* Adjust Reno SACK estimate. */ 84891fed7a1SIlpo Järvinen if (tcp_is_reno(tp) && diff > 0) { 849af610b4cSIlpo Järvinen tcp_dec_pcount_approx_int(&tp->sacked_out, diff); 850005903bcSIlpo Järvinen tcp_verify_left_out(tp); 85183ca28beSHerbert Xu } 852a47e5a98SIlpo Järvinen tcp_adjust_fackets_out(sk, skb, diff); 8531da177e4SLinus Torvalds } 8541da177e4SLinus Torvalds 8551da177e4SLinus Torvalds /* Link BUFF into the send queue. */ 856f44b5271SDavid S. Miller skb_header_release(buff); 857fe067e8aSDavid S. Miller tcp_insert_write_queue_after(skb, buff, sk); 8581da177e4SLinus Torvalds 8591da177e4SLinus Torvalds return 0; 8601da177e4SLinus Torvalds } 8611da177e4SLinus Torvalds 8621da177e4SLinus Torvalds /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c 8631da177e4SLinus Torvalds * eventually). The difference is that pulled data not copied, but 8641da177e4SLinus Torvalds * immediately discarded. 8651da177e4SLinus Torvalds */ 866f2911969SHerbert Xu ~{PmVHI~} static void __pskb_trim_head(struct sk_buff *skb, int len) 8671da177e4SLinus Torvalds { 8681da177e4SLinus Torvalds int i, k, eat; 8691da177e4SLinus Torvalds 8701da177e4SLinus Torvalds eat = len; 8711da177e4SLinus Torvalds k = 0; 8721da177e4SLinus Torvalds for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) { 8731da177e4SLinus Torvalds if (skb_shinfo(skb)->frags[i].size <= eat) { 8741da177e4SLinus Torvalds put_page(skb_shinfo(skb)->frags[i].page); 8751da177e4SLinus Torvalds eat -= skb_shinfo(skb)->frags[i].size; 8761da177e4SLinus Torvalds } else { 8771da177e4SLinus Torvalds skb_shinfo(skb)->frags[k] = skb_shinfo(skb)->frags[i]; 8781da177e4SLinus Torvalds if (eat) { 8791da177e4SLinus Torvalds skb_shinfo(skb)->frags[k].page_offset += eat; 8801da177e4SLinus Torvalds skb_shinfo(skb)->frags[k].size -= eat; 8811da177e4SLinus Torvalds eat = 0; 8821da177e4SLinus Torvalds } 8831da177e4SLinus Torvalds k++; 8841da177e4SLinus Torvalds } 8851da177e4SLinus Torvalds } 8861da177e4SLinus Torvalds skb_shinfo(skb)->nr_frags = k; 8871da177e4SLinus Torvalds 88827a884dcSArnaldo Carvalho de Melo skb_reset_tail_pointer(skb); 8891da177e4SLinus Torvalds skb->data_len -= len; 8901da177e4SLinus Torvalds skb->len = skb->data_len; 8911da177e4SLinus Torvalds } 8921da177e4SLinus Torvalds 8931da177e4SLinus Torvalds int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len) 8941da177e4SLinus Torvalds { 895056834d9SIlpo Järvinen if (skb_cloned(skb) && pskb_expand_head(skb, 0, 0, GFP_ATOMIC)) 8961da177e4SLinus Torvalds return -ENOMEM; 8971da177e4SLinus Torvalds 898f2911969SHerbert Xu ~{PmVHI~} /* If len == headlen, we avoid __skb_pull to preserve alignment. */ 899f2911969SHerbert Xu ~{PmVHI~} if (unlikely(len < skb_headlen(skb))) 9001da177e4SLinus Torvalds __skb_pull(skb, len); 901f2911969SHerbert Xu ~{PmVHI~} else 902f2911969SHerbert Xu ~{PmVHI~} __pskb_trim_head(skb, len - skb_headlen(skb)); 9031da177e4SLinus Torvalds 9041da177e4SLinus Torvalds TCP_SKB_CB(skb)->seq += len; 90584fa7933SPatrick McHardy skb->ip_summed = CHECKSUM_PARTIAL; 9061da177e4SLinus Torvalds 9071da177e4SLinus Torvalds skb->truesize -= len; 9081da177e4SLinus Torvalds sk->sk_wmem_queued -= len; 9093ab224beSHideo Aoki sk_mem_uncharge(sk, len); 9101da177e4SLinus Torvalds sock_set_flag(sk, SOCK_QUEUE_SHRUNK); 9111da177e4SLinus Torvalds 9121da177e4SLinus Torvalds /* Any change of skb->len requires recalculation of tso 9131da177e4SLinus Torvalds * factor and mss. 9141da177e4SLinus Torvalds */ 9151da177e4SLinus Torvalds if (tcp_skb_pcount(skb) > 1) 916846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, skb, tcp_current_mss(sk, 1)); 9171da177e4SLinus Torvalds 9181da177e4SLinus Torvalds return 0; 9191da177e4SLinus Torvalds } 9201da177e4SLinus Torvalds 9215d424d5aSJohn Heffner /* Not accounting for SACKs here. */ 9225d424d5aSJohn Heffner int tcp_mtu_to_mss(struct sock *sk, int pmtu) 9235d424d5aSJohn Heffner { 9245d424d5aSJohn Heffner struct tcp_sock *tp = tcp_sk(sk); 9255d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 9265d424d5aSJohn Heffner int mss_now; 9275d424d5aSJohn Heffner 9285d424d5aSJohn Heffner /* Calculate base mss without TCP options: 9295d424d5aSJohn Heffner It is MMS_S - sizeof(tcphdr) of rfc1122 9305d424d5aSJohn Heffner */ 9315d424d5aSJohn Heffner mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr); 9325d424d5aSJohn Heffner 9335d424d5aSJohn Heffner /* Clamp it (mss_clamp does not include tcp options) */ 9345d424d5aSJohn Heffner if (mss_now > tp->rx_opt.mss_clamp) 9355d424d5aSJohn Heffner mss_now = tp->rx_opt.mss_clamp; 9365d424d5aSJohn Heffner 9375d424d5aSJohn Heffner /* Now subtract optional transport overhead */ 9385d424d5aSJohn Heffner mss_now -= icsk->icsk_ext_hdr_len; 9395d424d5aSJohn Heffner 9405d424d5aSJohn Heffner /* Then reserve room for full set of TCP options and 8 bytes of data */ 9415d424d5aSJohn Heffner if (mss_now < 48) 9425d424d5aSJohn Heffner mss_now = 48; 9435d424d5aSJohn Heffner 9445d424d5aSJohn Heffner /* Now subtract TCP options size, not including SACKs */ 9455d424d5aSJohn Heffner mss_now -= tp->tcp_header_len - sizeof(struct tcphdr); 9465d424d5aSJohn Heffner 9475d424d5aSJohn Heffner return mss_now; 9485d424d5aSJohn Heffner } 9495d424d5aSJohn Heffner 9505d424d5aSJohn Heffner /* Inverse of above */ 9515d424d5aSJohn Heffner int tcp_mss_to_mtu(struct sock *sk, int mss) 9525d424d5aSJohn Heffner { 9535d424d5aSJohn Heffner struct tcp_sock *tp = tcp_sk(sk); 9545d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 9555d424d5aSJohn Heffner int mtu; 9565d424d5aSJohn Heffner 9575d424d5aSJohn Heffner mtu = mss + 9585d424d5aSJohn Heffner tp->tcp_header_len + 9595d424d5aSJohn Heffner icsk->icsk_ext_hdr_len + 9605d424d5aSJohn Heffner icsk->icsk_af_ops->net_header_len; 9615d424d5aSJohn Heffner 9625d424d5aSJohn Heffner return mtu; 9635d424d5aSJohn Heffner } 9645d424d5aSJohn Heffner 9655d424d5aSJohn Heffner void tcp_mtup_init(struct sock *sk) 9665d424d5aSJohn Heffner { 9675d424d5aSJohn Heffner struct tcp_sock *tp = tcp_sk(sk); 9685d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 9695d424d5aSJohn Heffner 9705d424d5aSJohn Heffner icsk->icsk_mtup.enabled = sysctl_tcp_mtu_probing > 1; 9715d424d5aSJohn Heffner icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) + 9725d424d5aSJohn Heffner icsk->icsk_af_ops->net_header_len; 9735d424d5aSJohn Heffner icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, sysctl_tcp_base_mss); 9745d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = 0; 9755d424d5aSJohn Heffner } 9765d424d5aSJohn Heffner 977409d22b4SIlpo Järvinen /* Bound MSS / TSO packet size with the half of the window */ 978409d22b4SIlpo Järvinen static int tcp_bound_to_half_wnd(struct tcp_sock *tp, int pktsize) 979409d22b4SIlpo Järvinen { 980409d22b4SIlpo Järvinen if (tp->max_window && pktsize > (tp->max_window >> 1)) 981409d22b4SIlpo Järvinen return max(tp->max_window >> 1, 68U - tp->tcp_header_len); 982409d22b4SIlpo Järvinen else 983409d22b4SIlpo Järvinen return pktsize; 984409d22b4SIlpo Järvinen } 985409d22b4SIlpo Järvinen 9861da177e4SLinus Torvalds /* This function synchronize snd mss to current pmtu/exthdr set. 9871da177e4SLinus Torvalds 9881da177e4SLinus Torvalds tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts 9891da177e4SLinus Torvalds for TCP options, but includes only bare TCP header. 9901da177e4SLinus Torvalds 9911da177e4SLinus Torvalds tp->rx_opt.mss_clamp is mss negotiated at connection setup. 992caa20d9aSStephen Hemminger It is minimum of user_mss and mss received with SYN. 9931da177e4SLinus Torvalds It also does not include TCP options. 9941da177e4SLinus Torvalds 995d83d8461SArnaldo Carvalho de Melo inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function. 9961da177e4SLinus Torvalds 9971da177e4SLinus Torvalds tp->mss_cache is current effective sending mss, including 9981da177e4SLinus Torvalds all tcp options except for SACKs. It is evaluated, 9991da177e4SLinus Torvalds taking into account current pmtu, but never exceeds 10001da177e4SLinus Torvalds tp->rx_opt.mss_clamp. 10011da177e4SLinus Torvalds 10021da177e4SLinus Torvalds NOTE1. rfc1122 clearly states that advertised MSS 10031da177e4SLinus Torvalds DOES NOT include either tcp or ip options. 10041da177e4SLinus Torvalds 1005d83d8461SArnaldo Carvalho de Melo NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache 1006d83d8461SArnaldo Carvalho de Melo are READ ONLY outside this function. --ANK (980731) 10071da177e4SLinus Torvalds */ 10081da177e4SLinus Torvalds unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu) 10091da177e4SLinus Torvalds { 10101da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 1011d83d8461SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 10125d424d5aSJohn Heffner int mss_now; 10131da177e4SLinus Torvalds 10145d424d5aSJohn Heffner if (icsk->icsk_mtup.search_high > pmtu) 10155d424d5aSJohn Heffner icsk->icsk_mtup.search_high = pmtu; 10161da177e4SLinus Torvalds 10175d424d5aSJohn Heffner mss_now = tcp_mtu_to_mss(sk, pmtu); 1018409d22b4SIlpo Järvinen mss_now = tcp_bound_to_half_wnd(tp, mss_now); 10191da177e4SLinus Torvalds 10201da177e4SLinus Torvalds /* And store cached results */ 1021d83d8461SArnaldo Carvalho de Melo icsk->icsk_pmtu_cookie = pmtu; 10225d424d5aSJohn Heffner if (icsk->icsk_mtup.enabled) 10235d424d5aSJohn Heffner mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low)); 1024c1b4a7e6SDavid S. Miller tp->mss_cache = mss_now; 10251da177e4SLinus Torvalds 10261da177e4SLinus Torvalds return mss_now; 10271da177e4SLinus Torvalds } 10281da177e4SLinus Torvalds 10291da177e4SLinus Torvalds /* Compute the current effective MSS, taking SACKs and IP options, 10301da177e4SLinus Torvalds * and even PMTU discovery events into account. 10311da177e4SLinus Torvalds * 103233f5f57eSIlpo Järvinen * LARGESEND note: !tcp_urg_mode is overkill, only frames up to snd_up 10331da177e4SLinus Torvalds * cannot be large. However, taking into account rare use of URG, this 10341da177e4SLinus Torvalds * is not a big flaw. 10351da177e4SLinus Torvalds */ 1036c1b4a7e6SDavid S. Miller unsigned int tcp_current_mss(struct sock *sk, int large_allowed) 10371da177e4SLinus Torvalds { 10381da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 10391da177e4SLinus Torvalds struct dst_entry *dst = __sk_dst_get(sk); 1040c1b4a7e6SDavid S. Miller u32 mss_now; 1041c1b4a7e6SDavid S. Miller u16 xmit_size_goal; 1042c1b4a7e6SDavid S. Miller int doing_tso = 0; 104333ad798cSAdam Langley unsigned header_len; 104433ad798cSAdam Langley struct tcp_out_options opts; 104533ad798cSAdam Langley struct tcp_md5sig_key *md5; 10461da177e4SLinus Torvalds 1047c1b4a7e6SDavid S. Miller mss_now = tp->mss_cache; 1048c1b4a7e6SDavid S. Miller 104933f5f57eSIlpo Järvinen if (large_allowed && sk_can_gso(sk) && !tcp_urg_mode(tp)) 1050c1b4a7e6SDavid S. Miller doing_tso = 1; 1051c1b4a7e6SDavid S. Miller 10521da177e4SLinus Torvalds if (dst) { 10531da177e4SLinus Torvalds u32 mtu = dst_mtu(dst); 1054d83d8461SArnaldo Carvalho de Melo if (mtu != inet_csk(sk)->icsk_pmtu_cookie) 10551da177e4SLinus Torvalds mss_now = tcp_sync_mss(sk, mtu); 10561da177e4SLinus Torvalds } 10571da177e4SLinus Torvalds 105833ad798cSAdam Langley header_len = tcp_established_options(sk, NULL, &opts, &md5) + 105933ad798cSAdam Langley sizeof(struct tcphdr); 106033ad798cSAdam Langley /* The mss_cache is sized based on tp->tcp_header_len, which assumes 106133ad798cSAdam Langley * some common options. If this is an odd packet (because we have SACK 106233ad798cSAdam Langley * blocks etc) then our calculated header_len will be different, and 106333ad798cSAdam Langley * we have to adjust mss_now correspondingly */ 106433ad798cSAdam Langley if (header_len != tp->tcp_header_len) { 106533ad798cSAdam Langley int delta = (int) header_len - tp->tcp_header_len; 106633ad798cSAdam Langley mss_now -= delta; 106733ad798cSAdam Langley } 1068cfb6eeb4SYOSHIFUJI Hideaki 1069c1b4a7e6SDavid S. Miller xmit_size_goal = mss_now; 1070c1b4a7e6SDavid S. Miller 1071c1b4a7e6SDavid S. Miller if (doing_tso) { 107282cc1a7aSPeter P Waskiewicz Jr xmit_size_goal = ((sk->sk_gso_max_size - 1) - 10738292a17aSArnaldo Carvalho de Melo inet_csk(sk)->icsk_af_ops->net_header_len - 1074d83d8461SArnaldo Carvalho de Melo inet_csk(sk)->icsk_ext_hdr_len - 1075d83d8461SArnaldo Carvalho de Melo tp->tcp_header_len); 1076c1b4a7e6SDavid S. Miller 1077409d22b4SIlpo Järvinen xmit_size_goal = tcp_bound_to_half_wnd(tp, xmit_size_goal); 1078c1b4a7e6SDavid S. Miller xmit_size_goal -= (xmit_size_goal % mss_now); 1079c1b4a7e6SDavid S. Miller } 1080c1b4a7e6SDavid S. Miller tp->xmit_size_goal = xmit_size_goal; 1081c1b4a7e6SDavid S. Miller 10821da177e4SLinus Torvalds return mss_now; 10831da177e4SLinus Torvalds } 10841da177e4SLinus Torvalds 1085a762a980SDavid S. Miller /* Congestion window validation. (RFC2861) */ 10869e412ba7SIlpo Järvinen static void tcp_cwnd_validate(struct sock *sk) 1087a762a980SDavid S. Miller { 10889e412ba7SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 1089a762a980SDavid S. Miller 1090d436d686SIlpo Järvinen if (tp->packets_out >= tp->snd_cwnd) { 1091a762a980SDavid S. Miller /* Network is feed fully. */ 1092a762a980SDavid S. Miller tp->snd_cwnd_used = 0; 1093a762a980SDavid S. Miller tp->snd_cwnd_stamp = tcp_time_stamp; 1094a762a980SDavid S. Miller } else { 1095a762a980SDavid S. Miller /* Network starves. */ 1096a762a980SDavid S. Miller if (tp->packets_out > tp->snd_cwnd_used) 1097a762a980SDavid S. Miller tp->snd_cwnd_used = tp->packets_out; 1098a762a980SDavid S. Miller 109915d33c07SDavid S. Miller if (sysctl_tcp_slow_start_after_idle && 110015d33c07SDavid S. Miller (s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto) 1101a762a980SDavid S. Miller tcp_cwnd_application_limited(sk); 1102a762a980SDavid S. Miller } 1103a762a980SDavid S. Miller } 1104a762a980SDavid S. Miller 11050e3a4803SIlpo Järvinen /* Returns the portion of skb which can be sent right away without 11060e3a4803SIlpo Järvinen * introducing MSS oddities to segment boundaries. In rare cases where 11070e3a4803SIlpo Järvinen * mss_now != mss_cache, we will request caller to create a small skb 11080e3a4803SIlpo Järvinen * per input skb which could be mostly avoided here (if desired). 11095ea3a748SIlpo Järvinen * 11105ea3a748SIlpo Järvinen * We explicitly want to create a request for splitting write queue tail 11115ea3a748SIlpo Järvinen * to a small skb for Nagle purposes while avoiding unnecessary modulos, 11125ea3a748SIlpo Järvinen * thus all the complexity (cwnd_len is always MSS multiple which we 11135ea3a748SIlpo Järvinen * return whenever allowed by the other factors). Basically we need the 11145ea3a748SIlpo Järvinen * modulo only when the receiver window alone is the limiting factor or 11155ea3a748SIlpo Järvinen * when we would be allowed to send the split-due-to-Nagle skb fully. 11160e3a4803SIlpo Järvinen */ 11170e3a4803SIlpo Järvinen static unsigned int tcp_mss_split_point(struct sock *sk, struct sk_buff *skb, 1118056834d9SIlpo Järvinen unsigned int mss_now, unsigned int cwnd) 1119c1b4a7e6SDavid S. Miller { 11200e3a4803SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 11210e3a4803SIlpo Järvinen u32 needed, window, cwnd_len; 1122c1b4a7e6SDavid S. Miller 112390840defSIlpo Järvinen window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 1124c1b4a7e6SDavid S. Miller cwnd_len = mss_now * cwnd; 11250e3a4803SIlpo Järvinen 11260e3a4803SIlpo Järvinen if (likely(cwnd_len <= window && skb != tcp_write_queue_tail(sk))) 11270e3a4803SIlpo Järvinen return cwnd_len; 11280e3a4803SIlpo Järvinen 11295ea3a748SIlpo Järvinen needed = min(skb->len, window); 11305ea3a748SIlpo Järvinen 113117515408SIlpo Järvinen if (cwnd_len <= needed) 11320e3a4803SIlpo Järvinen return cwnd_len; 11330e3a4803SIlpo Järvinen 11340e3a4803SIlpo Järvinen return needed - needed % mss_now; 1135c1b4a7e6SDavid S. Miller } 1136c1b4a7e6SDavid S. Miller 1137c1b4a7e6SDavid S. Miller /* Can at least one segment of SKB be sent right now, according to the 1138c1b4a7e6SDavid S. Miller * congestion window rules? If so, return how many segments are allowed. 1139c1b4a7e6SDavid S. Miller */ 1140056834d9SIlpo Järvinen static inline unsigned int tcp_cwnd_test(struct tcp_sock *tp, 1141056834d9SIlpo Järvinen struct sk_buff *skb) 1142c1b4a7e6SDavid S. Miller { 1143c1b4a7e6SDavid S. Miller u32 in_flight, cwnd; 1144c1b4a7e6SDavid S. Miller 1145c1b4a7e6SDavid S. Miller /* Don't be strict about the congestion window for the final FIN. */ 1146104439a8SJohn Heffner if ((TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN) && 1147104439a8SJohn Heffner tcp_skb_pcount(skb) == 1) 1148c1b4a7e6SDavid S. Miller return 1; 1149c1b4a7e6SDavid S. Miller 1150c1b4a7e6SDavid S. Miller in_flight = tcp_packets_in_flight(tp); 1151c1b4a7e6SDavid S. Miller cwnd = tp->snd_cwnd; 1152c1b4a7e6SDavid S. Miller if (in_flight < cwnd) 1153c1b4a7e6SDavid S. Miller return (cwnd - in_flight); 1154c1b4a7e6SDavid S. Miller 1155c1b4a7e6SDavid S. Miller return 0; 1156c1b4a7e6SDavid S. Miller } 1157c1b4a7e6SDavid S. Miller 1158c1b4a7e6SDavid S. Miller /* This must be invoked the first time we consider transmitting 1159c1b4a7e6SDavid S. Miller * SKB onto the wire. 1160c1b4a7e6SDavid S. Miller */ 1161056834d9SIlpo Järvinen static int tcp_init_tso_segs(struct sock *sk, struct sk_buff *skb, 1162056834d9SIlpo Järvinen unsigned int mss_now) 1163c1b4a7e6SDavid S. Miller { 1164c1b4a7e6SDavid S. Miller int tso_segs = tcp_skb_pcount(skb); 1165c1b4a7e6SDavid S. Miller 1166056834d9SIlpo Järvinen if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) { 1167846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, skb, mss_now); 1168c1b4a7e6SDavid S. Miller tso_segs = tcp_skb_pcount(skb); 1169c1b4a7e6SDavid S. Miller } 1170c1b4a7e6SDavid S. Miller return tso_segs; 1171c1b4a7e6SDavid S. Miller } 1172c1b4a7e6SDavid S. Miller 1173c1b4a7e6SDavid S. Miller static inline int tcp_minshall_check(const struct tcp_sock *tp) 1174c1b4a7e6SDavid S. Miller { 1175c1b4a7e6SDavid S. Miller return after(tp->snd_sml, tp->snd_una) && 1176c1b4a7e6SDavid S. Miller !after(tp->snd_sml, tp->snd_nxt); 1177c1b4a7e6SDavid S. Miller } 1178c1b4a7e6SDavid S. Miller 1179c1b4a7e6SDavid S. Miller /* Return 0, if packet can be sent now without violation Nagle's rules: 1180c1b4a7e6SDavid S. Miller * 1. It is full sized. 1181c1b4a7e6SDavid S. Miller * 2. Or it contains FIN. (already checked by caller) 1182c1b4a7e6SDavid S. Miller * 3. Or TCP_NODELAY was set. 1183c1b4a7e6SDavid S. Miller * 4. Or TCP_CORK is not set, and all sent packets are ACKed. 1184c1b4a7e6SDavid S. Miller * With Minshall's modification: all sent small packets are ACKed. 1185c1b4a7e6SDavid S. Miller */ 1186c1b4a7e6SDavid S. Miller static inline int tcp_nagle_check(const struct tcp_sock *tp, 1187c1b4a7e6SDavid S. Miller const struct sk_buff *skb, 1188c1b4a7e6SDavid S. Miller unsigned mss_now, int nonagle) 1189c1b4a7e6SDavid S. Miller { 1190c1b4a7e6SDavid S. Miller return (skb->len < mss_now && 1191c1b4a7e6SDavid S. Miller ((nonagle & TCP_NAGLE_CORK) || 1192056834d9SIlpo Järvinen (!nonagle && tp->packets_out && tcp_minshall_check(tp)))); 1193c1b4a7e6SDavid S. Miller } 1194c1b4a7e6SDavid S. Miller 1195c1b4a7e6SDavid S. Miller /* Return non-zero if the Nagle test allows this packet to be 1196c1b4a7e6SDavid S. Miller * sent now. 1197c1b4a7e6SDavid S. Miller */ 1198c1b4a7e6SDavid S. Miller static inline int tcp_nagle_test(struct tcp_sock *tp, struct sk_buff *skb, 1199c1b4a7e6SDavid S. Miller unsigned int cur_mss, int nonagle) 1200c1b4a7e6SDavid S. Miller { 1201c1b4a7e6SDavid S. Miller /* Nagle rule does not apply to frames, which sit in the middle of the 1202c1b4a7e6SDavid S. Miller * write_queue (they have no chances to get new data). 1203c1b4a7e6SDavid S. Miller * 1204c1b4a7e6SDavid S. Miller * This is implemented in the callers, where they modify the 'nonagle' 1205c1b4a7e6SDavid S. Miller * argument based upon the location of SKB in the send queue. 1206c1b4a7e6SDavid S. Miller */ 1207c1b4a7e6SDavid S. Miller if (nonagle & TCP_NAGLE_PUSH) 1208c1b4a7e6SDavid S. Miller return 1; 1209c1b4a7e6SDavid S. Miller 1210d551e454SIlpo Järvinen /* Don't use the nagle rule for urgent data (or for the final FIN). 1211d551e454SIlpo Järvinen * Nagle can be ignored during F-RTO too (see RFC4138). 1212d551e454SIlpo Järvinen */ 121333f5f57eSIlpo Järvinen if (tcp_urg_mode(tp) || (tp->frto_counter == 2) || 1214c1b4a7e6SDavid S. Miller (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN)) 1215c1b4a7e6SDavid S. Miller return 1; 1216c1b4a7e6SDavid S. Miller 1217c1b4a7e6SDavid S. Miller if (!tcp_nagle_check(tp, skb, cur_mss, nonagle)) 1218c1b4a7e6SDavid S. Miller return 1; 1219c1b4a7e6SDavid S. Miller 1220c1b4a7e6SDavid S. Miller return 0; 1221c1b4a7e6SDavid S. Miller } 1222c1b4a7e6SDavid S. Miller 1223c1b4a7e6SDavid S. Miller /* Does at least the first segment of SKB fit into the send window? */ 1224056834d9SIlpo Järvinen static inline int tcp_snd_wnd_test(struct tcp_sock *tp, struct sk_buff *skb, 1225056834d9SIlpo Järvinen unsigned int cur_mss) 1226c1b4a7e6SDavid S. Miller { 1227c1b4a7e6SDavid S. Miller u32 end_seq = TCP_SKB_CB(skb)->end_seq; 1228c1b4a7e6SDavid S. Miller 1229c1b4a7e6SDavid S. Miller if (skb->len > cur_mss) 1230c1b4a7e6SDavid S. Miller end_seq = TCP_SKB_CB(skb)->seq + cur_mss; 1231c1b4a7e6SDavid S. Miller 123290840defSIlpo Järvinen return !after(end_seq, tcp_wnd_end(tp)); 1233c1b4a7e6SDavid S. Miller } 1234c1b4a7e6SDavid S. Miller 1235fe067e8aSDavid S. Miller /* This checks if the data bearing packet SKB (usually tcp_send_head(sk)) 1236c1b4a7e6SDavid S. Miller * should be put on the wire right now. If so, it returns the number of 1237c1b4a7e6SDavid S. Miller * packets allowed by the congestion window. 1238c1b4a7e6SDavid S. Miller */ 1239c1b4a7e6SDavid S. Miller static unsigned int tcp_snd_test(struct sock *sk, struct sk_buff *skb, 1240c1b4a7e6SDavid S. Miller unsigned int cur_mss, int nonagle) 1241c1b4a7e6SDavid S. Miller { 1242c1b4a7e6SDavid S. Miller struct tcp_sock *tp = tcp_sk(sk); 1243c1b4a7e6SDavid S. Miller unsigned int cwnd_quota; 1244c1b4a7e6SDavid S. Miller 1245846998aeSDavid S. Miller tcp_init_tso_segs(sk, skb, cur_mss); 1246c1b4a7e6SDavid S. Miller 1247c1b4a7e6SDavid S. Miller if (!tcp_nagle_test(tp, skb, cur_mss, nonagle)) 1248c1b4a7e6SDavid S. Miller return 0; 1249c1b4a7e6SDavid S. Miller 1250c1b4a7e6SDavid S. Miller cwnd_quota = tcp_cwnd_test(tp, skb); 1251056834d9SIlpo Järvinen if (cwnd_quota && !tcp_snd_wnd_test(tp, skb, cur_mss)) 1252c1b4a7e6SDavid S. Miller cwnd_quota = 0; 1253c1b4a7e6SDavid S. Miller 1254c1b4a7e6SDavid S. Miller return cwnd_quota; 1255c1b4a7e6SDavid S. Miller } 1256c1b4a7e6SDavid S. Miller 12579e412ba7SIlpo Järvinen int tcp_may_send_now(struct sock *sk) 1258c1b4a7e6SDavid S. Miller { 12599e412ba7SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 1260fe067e8aSDavid S. Miller struct sk_buff *skb = tcp_send_head(sk); 1261c1b4a7e6SDavid S. Miller 1262c1b4a7e6SDavid S. Miller return (skb && 1263c1b4a7e6SDavid S. Miller tcp_snd_test(sk, skb, tcp_current_mss(sk, 1), 1264c1b4a7e6SDavid S. Miller (tcp_skb_is_last(sk, skb) ? 12654e67d876SIlpo Järvinen tp->nonagle : TCP_NAGLE_PUSH))); 1266c1b4a7e6SDavid S. Miller } 1267c1b4a7e6SDavid S. Miller 1268c1b4a7e6SDavid S. Miller /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet 1269c1b4a7e6SDavid S. Miller * which is put after SKB on the list. It is very much like 1270c1b4a7e6SDavid S. Miller * tcp_fragment() except that it may make several kinds of assumptions 1271c1b4a7e6SDavid S. Miller * in order to speed up the splitting operation. In particular, we 1272c1b4a7e6SDavid S. Miller * know that all the data is in scatter-gather pages, and that the 1273c1b4a7e6SDavid S. Miller * packet has never been sent out before (and thus is not cloned). 1274c1b4a7e6SDavid S. Miller */ 1275056834d9SIlpo Järvinen static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len, 1276056834d9SIlpo Järvinen unsigned int mss_now) 1277c1b4a7e6SDavid S. Miller { 1278c1b4a7e6SDavid S. Miller struct sk_buff *buff; 1279c1b4a7e6SDavid S. Miller int nlen = skb->len - len; 1280c1b4a7e6SDavid S. Miller u16 flags; 1281c1b4a7e6SDavid S. Miller 1282c1b4a7e6SDavid S. Miller /* All of a TSO frame must be composed of paged data. */ 1283c8ac3774SHerbert Xu if (skb->len != skb->data_len) 1284c8ac3774SHerbert Xu return tcp_fragment(sk, skb, len, mss_now); 1285c1b4a7e6SDavid S. Miller 1286df97c708SPavel Emelyanov buff = sk_stream_alloc_skb(sk, 0, GFP_ATOMIC); 1287c1b4a7e6SDavid S. Miller if (unlikely(buff == NULL)) 1288c1b4a7e6SDavid S. Miller return -ENOMEM; 1289c1b4a7e6SDavid S. Miller 12903ab224beSHideo Aoki sk->sk_wmem_queued += buff->truesize; 12913ab224beSHideo Aoki sk_mem_charge(sk, buff->truesize); 1292b60b49eaSHerbert Xu buff->truesize += nlen; 1293c1b4a7e6SDavid S. Miller skb->truesize -= nlen; 1294c1b4a7e6SDavid S. Miller 1295c1b4a7e6SDavid S. Miller /* Correct the sequence numbers. */ 1296c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; 1297c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; 1298c1b4a7e6SDavid S. Miller TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; 1299c1b4a7e6SDavid S. Miller 1300c1b4a7e6SDavid S. Miller /* PSH and FIN should only be set in the second packet. */ 1301c1b4a7e6SDavid S. Miller flags = TCP_SKB_CB(skb)->flags; 1302c1b4a7e6SDavid S. Miller TCP_SKB_CB(skb)->flags = flags & ~(TCPCB_FLAG_FIN | TCPCB_FLAG_PSH); 1303c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->flags = flags; 1304c1b4a7e6SDavid S. Miller 1305c1b4a7e6SDavid S. Miller /* This packet was never sent out yet, so no SACK bits. */ 1306c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->sacked = 0; 1307c1b4a7e6SDavid S. Miller 130884fa7933SPatrick McHardy buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL; 1309c1b4a7e6SDavid S. Miller skb_split(skb, buff, len); 1310c1b4a7e6SDavid S. Miller 1311c1b4a7e6SDavid S. Miller /* Fix up tso_factor for both original and new SKB. */ 1312846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, skb, mss_now); 1313846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, buff, mss_now); 1314c1b4a7e6SDavid S. Miller 1315c1b4a7e6SDavid S. Miller /* Link BUFF into the send queue. */ 1316c1b4a7e6SDavid S. Miller skb_header_release(buff); 1317fe067e8aSDavid S. Miller tcp_insert_write_queue_after(skb, buff, sk); 1318c1b4a7e6SDavid S. Miller 1319c1b4a7e6SDavid S. Miller return 0; 1320c1b4a7e6SDavid S. Miller } 1321c1b4a7e6SDavid S. Miller 1322c1b4a7e6SDavid S. Miller /* Try to defer sending, if possible, in order to minimize the amount 1323c1b4a7e6SDavid S. Miller * of TSO splitting we do. View it as a kind of TSO Nagle test. 1324c1b4a7e6SDavid S. Miller * 1325c1b4a7e6SDavid S. Miller * This algorithm is from John Heffner. 1326c1b4a7e6SDavid S. Miller */ 13279e412ba7SIlpo Järvinen static int tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb) 1328c1b4a7e6SDavid S. Miller { 13299e412ba7SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 13306687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 1331c1b4a7e6SDavid S. Miller u32 send_win, cong_win, limit, in_flight; 1332c1b4a7e6SDavid S. Miller 1333c1b4a7e6SDavid S. Miller if (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN) 1334ae8064acSJohn Heffner goto send_now; 1335c1b4a7e6SDavid S. Miller 13366687e988SArnaldo Carvalho de Melo if (icsk->icsk_ca_state != TCP_CA_Open) 1337ae8064acSJohn Heffner goto send_now; 1338ae8064acSJohn Heffner 1339ae8064acSJohn Heffner /* Defer for less than two clock ticks. */ 1340bd515c3eSIlpo Järvinen if (tp->tso_deferred && 1341bd515c3eSIlpo Järvinen ((jiffies << 1) >> 1) - (tp->tso_deferred >> 1) > 1) 1342ae8064acSJohn Heffner goto send_now; 1343908a75c1SDavid S. Miller 1344c1b4a7e6SDavid S. Miller in_flight = tcp_packets_in_flight(tp); 1345c1b4a7e6SDavid S. Miller 1346056834d9SIlpo Järvinen BUG_ON(tcp_skb_pcount(skb) <= 1 || (tp->snd_cwnd <= in_flight)); 1347c1b4a7e6SDavid S. Miller 134890840defSIlpo Järvinen send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 1349c1b4a7e6SDavid S. Miller 1350c1b4a7e6SDavid S. Miller /* From in_flight test above, we know that cwnd > in_flight. */ 1351c1b4a7e6SDavid S. Miller cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache; 1352c1b4a7e6SDavid S. Miller 1353c1b4a7e6SDavid S. Miller limit = min(send_win, cong_win); 1354c1b4a7e6SDavid S. Miller 1355ba244fe9SDavid S. Miller /* If a full-sized TSO skb can be sent, do it. */ 135682cc1a7aSPeter P Waskiewicz Jr if (limit >= sk->sk_gso_max_size) 1357ae8064acSJohn Heffner goto send_now; 1358ba244fe9SDavid S. Miller 1359c1b4a7e6SDavid S. Miller if (sysctl_tcp_tso_win_divisor) { 1360c1b4a7e6SDavid S. Miller u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache); 1361c1b4a7e6SDavid S. Miller 1362c1b4a7e6SDavid S. Miller /* If at least some fraction of a window is available, 1363c1b4a7e6SDavid S. Miller * just use it. 1364c1b4a7e6SDavid S. Miller */ 1365c1b4a7e6SDavid S. Miller chunk /= sysctl_tcp_tso_win_divisor; 1366c1b4a7e6SDavid S. Miller if (limit >= chunk) 1367ae8064acSJohn Heffner goto send_now; 1368c1b4a7e6SDavid S. Miller } else { 1369c1b4a7e6SDavid S. Miller /* Different approach, try not to defer past a single 1370c1b4a7e6SDavid S. Miller * ACK. Receiver should ACK every other full sized 1371c1b4a7e6SDavid S. Miller * frame, so if we have space for more than 3 frames 1372c1b4a7e6SDavid S. Miller * then send now. 1373c1b4a7e6SDavid S. Miller */ 1374c1b4a7e6SDavid S. Miller if (limit > tcp_max_burst(tp) * tp->mss_cache) 1375ae8064acSJohn Heffner goto send_now; 1376c1b4a7e6SDavid S. Miller } 1377c1b4a7e6SDavid S. Miller 1378c1b4a7e6SDavid S. Miller /* Ok, it looks like it is advisable to defer. */ 1379ae8064acSJohn Heffner tp->tso_deferred = 1 | (jiffies << 1); 1380ae8064acSJohn Heffner 1381c1b4a7e6SDavid S. Miller return 1; 1382ae8064acSJohn Heffner 1383ae8064acSJohn Heffner send_now: 1384ae8064acSJohn Heffner tp->tso_deferred = 0; 1385ae8064acSJohn Heffner return 0; 1386c1b4a7e6SDavid S. Miller } 1387c1b4a7e6SDavid S. Miller 13885d424d5aSJohn Heffner /* Create a new MTU probe if we are ready. 13895d424d5aSJohn Heffner * Returns 0 if we should wait to probe (no cwnd available), 13905d424d5aSJohn Heffner * 1 if a probe was sent, 1391056834d9SIlpo Järvinen * -1 otherwise 1392056834d9SIlpo Järvinen */ 13935d424d5aSJohn Heffner static int tcp_mtu_probe(struct sock *sk) 13945d424d5aSJohn Heffner { 13955d424d5aSJohn Heffner struct tcp_sock *tp = tcp_sk(sk); 13965d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 13975d424d5aSJohn Heffner struct sk_buff *skb, *nskb, *next; 13985d424d5aSJohn Heffner int len; 13995d424d5aSJohn Heffner int probe_size; 140091cc17c0SIlpo Järvinen int size_needed; 14015d424d5aSJohn Heffner int copy; 14025d424d5aSJohn Heffner int mss_now; 14035d424d5aSJohn Heffner 14045d424d5aSJohn Heffner /* Not currently probing/verifying, 14055d424d5aSJohn Heffner * not in recovery, 14065d424d5aSJohn Heffner * have enough cwnd, and 14075d424d5aSJohn Heffner * not SACKing (the variable headers throw things off) */ 14085d424d5aSJohn Heffner if (!icsk->icsk_mtup.enabled || 14095d424d5aSJohn Heffner icsk->icsk_mtup.probe_size || 14105d424d5aSJohn Heffner inet_csk(sk)->icsk_ca_state != TCP_CA_Open || 14115d424d5aSJohn Heffner tp->snd_cwnd < 11 || 14125d424d5aSJohn Heffner tp->rx_opt.eff_sacks) 14135d424d5aSJohn Heffner return -1; 14145d424d5aSJohn Heffner 14155d424d5aSJohn Heffner /* Very simple search strategy: just double the MSS. */ 14165d424d5aSJohn Heffner mss_now = tcp_current_mss(sk, 0); 14175d424d5aSJohn Heffner probe_size = 2 * tp->mss_cache; 141891cc17c0SIlpo Järvinen size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache; 14195d424d5aSJohn Heffner if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high)) { 14205d424d5aSJohn Heffner /* TODO: set timer for probe_converge_event */ 14215d424d5aSJohn Heffner return -1; 14225d424d5aSJohn Heffner } 14235d424d5aSJohn Heffner 14245d424d5aSJohn Heffner /* Have enough data in the send queue to probe? */ 14257f9c33e5SIlpo Järvinen if (tp->write_seq - tp->snd_nxt < size_needed) 14265d424d5aSJohn Heffner return -1; 14275d424d5aSJohn Heffner 142891cc17c0SIlpo Järvinen if (tp->snd_wnd < size_needed) 14295d424d5aSJohn Heffner return -1; 143090840defSIlpo Järvinen if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp))) 14315d424d5aSJohn Heffner return 0; 14325d424d5aSJohn Heffner 1433d67c58e9SIlpo Järvinen /* Do we need to wait to drain cwnd? With none in flight, don't stall */ 1434d67c58e9SIlpo Järvinen if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) { 1435d67c58e9SIlpo Järvinen if (!tcp_packets_in_flight(tp)) 14365d424d5aSJohn Heffner return -1; 14375d424d5aSJohn Heffner else 14385d424d5aSJohn Heffner return 0; 14395d424d5aSJohn Heffner } 14405d424d5aSJohn Heffner 14415d424d5aSJohn Heffner /* We're allowed to probe. Build it now. */ 14425d424d5aSJohn Heffner if ((nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC)) == NULL) 14435d424d5aSJohn Heffner return -1; 14443ab224beSHideo Aoki sk->sk_wmem_queued += nskb->truesize; 14453ab224beSHideo Aoki sk_mem_charge(sk, nskb->truesize); 14465d424d5aSJohn Heffner 1447fe067e8aSDavid S. Miller skb = tcp_send_head(sk); 14485d424d5aSJohn Heffner 14495d424d5aSJohn Heffner TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq; 14505d424d5aSJohn Heffner TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size; 14515d424d5aSJohn Heffner TCP_SKB_CB(nskb)->flags = TCPCB_FLAG_ACK; 14525d424d5aSJohn Heffner TCP_SKB_CB(nskb)->sacked = 0; 14535d424d5aSJohn Heffner nskb->csum = 0; 145484fa7933SPatrick McHardy nskb->ip_summed = skb->ip_summed; 14555d424d5aSJohn Heffner 145650c4817eSIlpo Järvinen tcp_insert_write_queue_before(nskb, skb, sk); 145750c4817eSIlpo Järvinen 14585d424d5aSJohn Heffner len = 0; 1459234b6860SIlpo Järvinen tcp_for_write_queue_from_safe(skb, next, sk) { 14605d424d5aSJohn Heffner copy = min_t(int, skb->len, probe_size - len); 14615d424d5aSJohn Heffner if (nskb->ip_summed) 14625d424d5aSJohn Heffner skb_copy_bits(skb, 0, skb_put(nskb, copy), copy); 14635d424d5aSJohn Heffner else 14645d424d5aSJohn Heffner nskb->csum = skb_copy_and_csum_bits(skb, 0, 1465056834d9SIlpo Järvinen skb_put(nskb, copy), 1466056834d9SIlpo Järvinen copy, nskb->csum); 14675d424d5aSJohn Heffner 14685d424d5aSJohn Heffner if (skb->len <= copy) { 14695d424d5aSJohn Heffner /* We've eaten all the data from this skb. 14705d424d5aSJohn Heffner * Throw it away. */ 14715d424d5aSJohn Heffner TCP_SKB_CB(nskb)->flags |= TCP_SKB_CB(skb)->flags; 1472fe067e8aSDavid S. Miller tcp_unlink_write_queue(skb, sk); 14733ab224beSHideo Aoki sk_wmem_free_skb(sk, skb); 14745d424d5aSJohn Heffner } else { 14755d424d5aSJohn Heffner TCP_SKB_CB(nskb)->flags |= TCP_SKB_CB(skb)->flags & 14765d424d5aSJohn Heffner ~(TCPCB_FLAG_FIN|TCPCB_FLAG_PSH); 14775d424d5aSJohn Heffner if (!skb_shinfo(skb)->nr_frags) { 14785d424d5aSJohn Heffner skb_pull(skb, copy); 147984fa7933SPatrick McHardy if (skb->ip_summed != CHECKSUM_PARTIAL) 1480056834d9SIlpo Järvinen skb->csum = csum_partial(skb->data, 1481056834d9SIlpo Järvinen skb->len, 0); 14825d424d5aSJohn Heffner } else { 14835d424d5aSJohn Heffner __pskb_trim_head(skb, copy); 14845d424d5aSJohn Heffner tcp_set_skb_tso_segs(sk, skb, mss_now); 14855d424d5aSJohn Heffner } 14865d424d5aSJohn Heffner TCP_SKB_CB(skb)->seq += copy; 14875d424d5aSJohn Heffner } 14885d424d5aSJohn Heffner 14895d424d5aSJohn Heffner len += copy; 1490234b6860SIlpo Järvinen 1491234b6860SIlpo Järvinen if (len >= probe_size) 1492234b6860SIlpo Järvinen break; 14935d424d5aSJohn Heffner } 14945d424d5aSJohn Heffner tcp_init_tso_segs(sk, nskb, nskb->len); 14955d424d5aSJohn Heffner 14965d424d5aSJohn Heffner /* We're ready to send. If this fails, the probe will 14975d424d5aSJohn Heffner * be resegmented into mss-sized pieces by tcp_write_xmit(). */ 14985d424d5aSJohn Heffner TCP_SKB_CB(nskb)->when = tcp_time_stamp; 14995d424d5aSJohn Heffner if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) { 15005d424d5aSJohn Heffner /* Decrement cwnd here because we are sending 15015d424d5aSJohn Heffner * effectively two packets. */ 15025d424d5aSJohn Heffner tp->snd_cwnd--; 150366f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, nskb); 15045d424d5aSJohn Heffner 15055d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len); 15060e7b1368SJohn Heffner tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq; 15070e7b1368SJohn Heffner tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq; 15085d424d5aSJohn Heffner 15095d424d5aSJohn Heffner return 1; 15105d424d5aSJohn Heffner } 15115d424d5aSJohn Heffner 15125d424d5aSJohn Heffner return -1; 15135d424d5aSJohn Heffner } 15145d424d5aSJohn Heffner 15151da177e4SLinus Torvalds /* This routine writes packets to the network. It advances the 15161da177e4SLinus Torvalds * send_head. This happens as incoming acks open up the remote 15171da177e4SLinus Torvalds * window for us. 15181da177e4SLinus Torvalds * 15191da177e4SLinus Torvalds * Returns 1, if no segments are in flight and we have queued segments, but 15201da177e4SLinus Torvalds * cannot send anything now because of SWS or another problem. 15211da177e4SLinus Torvalds */ 1522a2e2a59cSDavid S. Miller static int tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle) 15231da177e4SLinus Torvalds { 15241da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 152592df7b51SDavid S. Miller struct sk_buff *skb; 1526c1b4a7e6SDavid S. Miller unsigned int tso_segs, sent_pkts; 1527c1b4a7e6SDavid S. Miller int cwnd_quota; 15285d424d5aSJohn Heffner int result; 15291da177e4SLinus Torvalds 15301da177e4SLinus Torvalds /* If we are closed, the bytes will have to remain here. 15311da177e4SLinus Torvalds * In time closedown will finish, we empty the write queue and all 15321da177e4SLinus Torvalds * will be happy. 15331da177e4SLinus Torvalds */ 153492df7b51SDavid S. Miller if (unlikely(sk->sk_state == TCP_CLOSE)) 153592df7b51SDavid S. Miller return 0; 153692df7b51SDavid S. Miller 1537c1b4a7e6SDavid S. Miller sent_pkts = 0; 15385d424d5aSJohn Heffner 15395d424d5aSJohn Heffner /* Do MTU probing. */ 15405d424d5aSJohn Heffner if ((result = tcp_mtu_probe(sk)) == 0) { 15415d424d5aSJohn Heffner return 0; 15425d424d5aSJohn Heffner } else if (result > 0) { 15435d424d5aSJohn Heffner sent_pkts = 1; 15445d424d5aSJohn Heffner } 15455d424d5aSJohn Heffner 1546fe067e8aSDavid S. Miller while ((skb = tcp_send_head(sk))) { 1547c8ac3774SHerbert Xu unsigned int limit; 1548c8ac3774SHerbert Xu 1549b68e9f85SHerbert Xu tso_segs = tcp_init_tso_segs(sk, skb, mss_now); 1550c1b4a7e6SDavid S. Miller BUG_ON(!tso_segs); 1551c1b4a7e6SDavid S. Miller 1552b68e9f85SHerbert Xu cwnd_quota = tcp_cwnd_test(tp, skb); 1553b68e9f85SHerbert Xu if (!cwnd_quota) 1554b68e9f85SHerbert Xu break; 1555b68e9f85SHerbert Xu 1556b68e9f85SHerbert Xu if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now))) 1557b68e9f85SHerbert Xu break; 1558b68e9f85SHerbert Xu 1559c1b4a7e6SDavid S. Miller if (tso_segs == 1) { 1560aa93466bSDavid S. Miller if (unlikely(!tcp_nagle_test(tp, skb, mss_now, 1561aa93466bSDavid S. Miller (tcp_skb_is_last(sk, skb) ? 1562aa93466bSDavid S. Miller nonagle : TCP_NAGLE_PUSH)))) 1563aa93466bSDavid S. Miller break; 1564c1b4a7e6SDavid S. Miller } else { 15659e412ba7SIlpo Järvinen if (tcp_tso_should_defer(sk, skb)) 1566aa93466bSDavid S. Miller break; 1567c1b4a7e6SDavid S. Miller } 1568aa93466bSDavid S. Miller 1569c8ac3774SHerbert Xu limit = mss_now; 15700e3a4803SIlpo Järvinen if (tso_segs > 1) 15710e3a4803SIlpo Järvinen limit = tcp_mss_split_point(sk, skb, mss_now, 15720e3a4803SIlpo Järvinen cwnd_quota); 1573c8ac3774SHerbert Xu 1574c8ac3774SHerbert Xu if (skb->len > limit && 1575c8ac3774SHerbert Xu unlikely(tso_fragment(sk, skb, limit, mss_now))) 15761da177e4SLinus Torvalds break; 15771da177e4SLinus Torvalds 15781da177e4SLinus Torvalds TCP_SKB_CB(skb)->when = tcp_time_stamp; 1579c1b4a7e6SDavid S. Miller 1580dfb4b9dcSDavid S. Miller if (unlikely(tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC))) 15811da177e4SLinus Torvalds break; 15821da177e4SLinus Torvalds 15831da177e4SLinus Torvalds /* Advance the send_head. This one is sent out. 15841da177e4SLinus Torvalds * This call will increment packets_out. 15851da177e4SLinus Torvalds */ 158666f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, skb); 15871da177e4SLinus Torvalds 15881da177e4SLinus Torvalds tcp_minshall_update(tp, mss_now, skb); 1589aa93466bSDavid S. Miller sent_pkts++; 15901da177e4SLinus Torvalds } 15911da177e4SLinus Torvalds 1592aa93466bSDavid S. Miller if (likely(sent_pkts)) { 15939e412ba7SIlpo Järvinen tcp_cwnd_validate(sk); 15941da177e4SLinus Torvalds return 0; 15951da177e4SLinus Torvalds } 1596fe067e8aSDavid S. Miller return !tp->packets_out && tcp_send_head(sk); 15971da177e4SLinus Torvalds } 15981da177e4SLinus Torvalds 1599a762a980SDavid S. Miller /* Push out any pending frames which were held back due to 1600a762a980SDavid S. Miller * TCP_CORK or attempt at coalescing tiny packets. 1601a762a980SDavid S. Miller * The socket must be locked by the caller. 1602a762a980SDavid S. Miller */ 16039e412ba7SIlpo Järvinen void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss, 16049e412ba7SIlpo Järvinen int nonagle) 1605a762a980SDavid S. Miller { 1606fe067e8aSDavid S. Miller struct sk_buff *skb = tcp_send_head(sk); 1607a762a980SDavid S. Miller 1608a762a980SDavid S. Miller if (skb) { 160955c97f3eSDavid S. Miller if (tcp_write_xmit(sk, cur_mss, nonagle)) 16109e412ba7SIlpo Järvinen tcp_check_probe_timer(sk); 1611a762a980SDavid S. Miller } 1612a762a980SDavid S. Miller } 1613a762a980SDavid S. Miller 1614c1b4a7e6SDavid S. Miller /* Send _single_ skb sitting at the send head. This function requires 1615c1b4a7e6SDavid S. Miller * true push pending frames to setup probe timer etc. 1616c1b4a7e6SDavid S. Miller */ 1617c1b4a7e6SDavid S. Miller void tcp_push_one(struct sock *sk, unsigned int mss_now) 1618c1b4a7e6SDavid S. Miller { 1619fe067e8aSDavid S. Miller struct sk_buff *skb = tcp_send_head(sk); 1620c1b4a7e6SDavid S. Miller unsigned int tso_segs, cwnd_quota; 1621c1b4a7e6SDavid S. Miller 1622c1b4a7e6SDavid S. Miller BUG_ON(!skb || skb->len < mss_now); 1623c1b4a7e6SDavid S. Miller 1624846998aeSDavid S. Miller tso_segs = tcp_init_tso_segs(sk, skb, mss_now); 1625c1b4a7e6SDavid S. Miller cwnd_quota = tcp_snd_test(sk, skb, mss_now, TCP_NAGLE_PUSH); 1626c1b4a7e6SDavid S. Miller 1627c1b4a7e6SDavid S. Miller if (likely(cwnd_quota)) { 1628c8ac3774SHerbert Xu unsigned int limit; 1629c8ac3774SHerbert Xu 1630c1b4a7e6SDavid S. Miller BUG_ON(!tso_segs); 1631c1b4a7e6SDavid S. Miller 1632c8ac3774SHerbert Xu limit = mss_now; 16330e3a4803SIlpo Järvinen if (tso_segs > 1) 16340e3a4803SIlpo Järvinen limit = tcp_mss_split_point(sk, skb, mss_now, 16350e3a4803SIlpo Järvinen cwnd_quota); 1636c8ac3774SHerbert Xu 1637c8ac3774SHerbert Xu if (skb->len > limit && 1638c8ac3774SHerbert Xu unlikely(tso_fragment(sk, skb, limit, mss_now))) 1639c1b4a7e6SDavid S. Miller return; 1640c1b4a7e6SDavid S. Miller 1641c1b4a7e6SDavid S. Miller /* Send it out now. */ 1642c1b4a7e6SDavid S. Miller TCP_SKB_CB(skb)->when = tcp_time_stamp; 1643c1b4a7e6SDavid S. Miller 1644dfb4b9dcSDavid S. Miller if (likely(!tcp_transmit_skb(sk, skb, 1, sk->sk_allocation))) { 164566f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, skb); 16469e412ba7SIlpo Järvinen tcp_cwnd_validate(sk); 1647c1b4a7e6SDavid S. Miller return; 1648c1b4a7e6SDavid S. Miller } 1649c1b4a7e6SDavid S. Miller } 1650c1b4a7e6SDavid S. Miller } 1651c1b4a7e6SDavid S. Miller 16521da177e4SLinus Torvalds /* This function returns the amount that we can raise the 16531da177e4SLinus Torvalds * usable window based on the following constraints 16541da177e4SLinus Torvalds * 16551da177e4SLinus Torvalds * 1. The window can never be shrunk once it is offered (RFC 793) 16561da177e4SLinus Torvalds * 2. We limit memory per socket 16571da177e4SLinus Torvalds * 16581da177e4SLinus Torvalds * RFC 1122: 16591da177e4SLinus Torvalds * "the suggested [SWS] avoidance algorithm for the receiver is to keep 16601da177e4SLinus Torvalds * RECV.NEXT + RCV.WIN fixed until: 16611da177e4SLinus Torvalds * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)" 16621da177e4SLinus Torvalds * 16631da177e4SLinus Torvalds * i.e. don't raise the right edge of the window until you can raise 16641da177e4SLinus Torvalds * it at least MSS bytes. 16651da177e4SLinus Torvalds * 16661da177e4SLinus Torvalds * Unfortunately, the recommended algorithm breaks header prediction, 16671da177e4SLinus Torvalds * since header prediction assumes th->window stays fixed. 16681da177e4SLinus Torvalds * 16691da177e4SLinus Torvalds * Strictly speaking, keeping th->window fixed violates the receiver 16701da177e4SLinus Torvalds * side SWS prevention criteria. The problem is that under this rule 16711da177e4SLinus Torvalds * a stream of single byte packets will cause the right side of the 16721da177e4SLinus Torvalds * window to always advance by a single byte. 16731da177e4SLinus Torvalds * 16741da177e4SLinus Torvalds * Of course, if the sender implements sender side SWS prevention 16751da177e4SLinus Torvalds * then this will not be a problem. 16761da177e4SLinus Torvalds * 16771da177e4SLinus Torvalds * BSD seems to make the following compromise: 16781da177e4SLinus Torvalds * 16791da177e4SLinus Torvalds * If the free space is less than the 1/4 of the maximum 16801da177e4SLinus Torvalds * space available and the free space is less than 1/2 mss, 16811da177e4SLinus Torvalds * then set the window to 0. 16821da177e4SLinus Torvalds * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ] 16831da177e4SLinus Torvalds * Otherwise, just prevent the window from shrinking 16841da177e4SLinus Torvalds * and from being larger than the largest representable value. 16851da177e4SLinus Torvalds * 16861da177e4SLinus Torvalds * This prevents incremental opening of the window in the regime 16871da177e4SLinus Torvalds * where TCP is limited by the speed of the reader side taking 16881da177e4SLinus Torvalds * data out of the TCP receive queue. It does nothing about 16891da177e4SLinus Torvalds * those cases where the window is constrained on the sender side 16901da177e4SLinus Torvalds * because the pipeline is full. 16911da177e4SLinus Torvalds * 16921da177e4SLinus Torvalds * BSD also seems to "accidentally" limit itself to windows that are a 16931da177e4SLinus Torvalds * multiple of MSS, at least until the free space gets quite small. 16941da177e4SLinus Torvalds * This would appear to be a side effect of the mbuf implementation. 16951da177e4SLinus Torvalds * Combining these two algorithms results in the observed behavior 16961da177e4SLinus Torvalds * of having a fixed window size at almost all times. 16971da177e4SLinus Torvalds * 16981da177e4SLinus Torvalds * Below we obtain similar behavior by forcing the offered window to 16991da177e4SLinus Torvalds * a multiple of the mss when it is feasible to do so. 17001da177e4SLinus Torvalds * 17011da177e4SLinus Torvalds * Note, we don't "adjust" for TIMESTAMP or SACK option bytes. 17021da177e4SLinus Torvalds * Regular options like TIMESTAMP are taken into account. 17031da177e4SLinus Torvalds */ 17041da177e4SLinus Torvalds u32 __tcp_select_window(struct sock *sk) 17051da177e4SLinus Torvalds { 1706463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 17071da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 1708caa20d9aSStephen Hemminger /* MSS for the peer's data. Previous versions used mss_clamp 17091da177e4SLinus Torvalds * here. I don't know if the value based on our guesses 17101da177e4SLinus Torvalds * of peer's MSS is better for the performance. It's more correct 17111da177e4SLinus Torvalds * but may be worse for the performance because of rcv_mss 17121da177e4SLinus Torvalds * fluctuations. --SAW 1998/11/1 17131da177e4SLinus Torvalds */ 1714463c84b9SArnaldo Carvalho de Melo int mss = icsk->icsk_ack.rcv_mss; 17151da177e4SLinus Torvalds int free_space = tcp_space(sk); 17161da177e4SLinus Torvalds int full_space = min_t(int, tp->window_clamp, tcp_full_space(sk)); 17171da177e4SLinus Torvalds int window; 17181da177e4SLinus Torvalds 17191da177e4SLinus Torvalds if (mss > full_space) 17201da177e4SLinus Torvalds mss = full_space; 17211da177e4SLinus Torvalds 1722b92edbe0SEric Dumazet if (free_space < (full_space >> 1)) { 1723463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.quick = 0; 17241da177e4SLinus Torvalds 17251da177e4SLinus Torvalds if (tcp_memory_pressure) 1726056834d9SIlpo Järvinen tp->rcv_ssthresh = min(tp->rcv_ssthresh, 1727056834d9SIlpo Järvinen 4U * tp->advmss); 17281da177e4SLinus Torvalds 17291da177e4SLinus Torvalds if (free_space < mss) 17301da177e4SLinus Torvalds return 0; 17311da177e4SLinus Torvalds } 17321da177e4SLinus Torvalds 17331da177e4SLinus Torvalds if (free_space > tp->rcv_ssthresh) 17341da177e4SLinus Torvalds free_space = tp->rcv_ssthresh; 17351da177e4SLinus Torvalds 17361da177e4SLinus Torvalds /* Don't do rounding if we are using window scaling, since the 17371da177e4SLinus Torvalds * scaled window will not line up with the MSS boundary anyway. 17381da177e4SLinus Torvalds */ 17391da177e4SLinus Torvalds window = tp->rcv_wnd; 17401da177e4SLinus Torvalds if (tp->rx_opt.rcv_wscale) { 17411da177e4SLinus Torvalds window = free_space; 17421da177e4SLinus Torvalds 17431da177e4SLinus Torvalds /* Advertise enough space so that it won't get scaled away. 17441da177e4SLinus Torvalds * Import case: prevent zero window announcement if 17451da177e4SLinus Torvalds * 1<<rcv_wscale > mss. 17461da177e4SLinus Torvalds */ 17471da177e4SLinus Torvalds if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window) 17481da177e4SLinus Torvalds window = (((window >> tp->rx_opt.rcv_wscale) + 1) 17491da177e4SLinus Torvalds << tp->rx_opt.rcv_wscale); 17501da177e4SLinus Torvalds } else { 17511da177e4SLinus Torvalds /* Get the largest window that is a nice multiple of mss. 17521da177e4SLinus Torvalds * Window clamp already applied above. 17531da177e4SLinus Torvalds * If our current window offering is within 1 mss of the 17541da177e4SLinus Torvalds * free space we just keep it. This prevents the divide 17551da177e4SLinus Torvalds * and multiply from happening most of the time. 17561da177e4SLinus Torvalds * We also don't do any window rounding when the free space 17571da177e4SLinus Torvalds * is too small. 17581da177e4SLinus Torvalds */ 17591da177e4SLinus Torvalds if (window <= free_space - mss || window > free_space) 17601da177e4SLinus Torvalds window = (free_space / mss) * mss; 176184565070SJohn Heffner else if (mss == full_space && 1762b92edbe0SEric Dumazet free_space > window + (full_space >> 1)) 176384565070SJohn Heffner window = free_space; 17641da177e4SLinus Torvalds } 17651da177e4SLinus Torvalds 17661da177e4SLinus Torvalds return window; 17671da177e4SLinus Torvalds } 17681da177e4SLinus Torvalds 1769*4a17fc3aSIlpo Järvinen /* Collapses two adjacent SKB's during retransmission. */ 1770*4a17fc3aSIlpo Järvinen static void tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb) 17711da177e4SLinus Torvalds { 17721da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 1773fe067e8aSDavid S. Miller struct sk_buff *next_skb = tcp_write_queue_next(sk, skb); 1774058dc334SIlpo Järvinen int skb_size, next_skb_size; 1775058dc334SIlpo Järvinen u16 flags; 17761da177e4SLinus Torvalds 1777058dc334SIlpo Järvinen skb_size = skb->len; 1778058dc334SIlpo Järvinen next_skb_size = next_skb->len; 1779058dc334SIlpo Järvinen flags = TCP_SKB_CB(skb)->flags; 17801da177e4SLinus Torvalds 1781058dc334SIlpo Järvinen BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1); 17821da177e4SLinus Torvalds 17836859d494SIlpo Järvinen tcp_highest_sack_combine(sk, next_skb, skb); 1784a6963a6bSIlpo Järvinen 1785fe067e8aSDavid S. Miller tcp_unlink_write_queue(next_skb, sk); 17861da177e4SLinus Torvalds 1787058dc334SIlpo Järvinen skb_copy_from_linear_data(next_skb, skb_put(skb, next_skb_size), 17881a4e2d09SArnaldo Carvalho de Melo next_skb_size); 17891da177e4SLinus Torvalds 179052d570aaSJarek Poplawski if (next_skb->ip_summed == CHECKSUM_PARTIAL) 179152d570aaSJarek Poplawski skb->ip_summed = CHECKSUM_PARTIAL; 17921da177e4SLinus Torvalds 179384fa7933SPatrick McHardy if (skb->ip_summed != CHECKSUM_PARTIAL) 17941da177e4SLinus Torvalds skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size); 17951da177e4SLinus Torvalds 17961da177e4SLinus Torvalds /* Update sequence range on original skb. */ 17971da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq; 17981da177e4SLinus Torvalds 17991da177e4SLinus Torvalds /* Merge over control information. */ 18001da177e4SLinus Torvalds flags |= TCP_SKB_CB(next_skb)->flags; /* This moves PSH/FIN etc. over */ 18011da177e4SLinus Torvalds TCP_SKB_CB(skb)->flags = flags; 18021da177e4SLinus Torvalds 18031da177e4SLinus Torvalds /* All done, get rid of second SKB and account for it so 18041da177e4SLinus Torvalds * packet counting does not break. 18051da177e4SLinus Torvalds */ 18064828e7f4SIlpo Järvinen TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS; 18071da177e4SLinus Torvalds if (TCP_SKB_CB(next_skb)->sacked & TCPCB_SACKED_RETRANS) 18081da177e4SLinus Torvalds tp->retrans_out -= tcp_skb_pcount(next_skb); 1809b5860bbaSIlpo Järvinen if (TCP_SKB_CB(next_skb)->sacked & TCPCB_LOST) 18101da177e4SLinus Torvalds tp->lost_out -= tcp_skb_pcount(next_skb); 18111da177e4SLinus Torvalds /* Reno case is special. Sigh... */ 1812e60402d0SIlpo Järvinen if (tcp_is_reno(tp) && tp->sacked_out) 18131da177e4SLinus Torvalds tcp_dec_pcount_approx(&tp->sacked_out, next_skb); 18141da177e4SLinus Torvalds 1815a47e5a98SIlpo Järvinen tcp_adjust_fackets_out(sk, next_skb, tcp_skb_pcount(next_skb)); 1816e9144bd8SIlpo Järvinen tp->packets_out -= tcp_skb_pcount(next_skb); 1817b7689205SIlpo Järvinen 1818b7689205SIlpo Järvinen /* changed transmit queue under us so clear hints */ 1819ef9da47cSIlpo Järvinen tcp_clear_retrans_hints_partial(tp); 1820ef9da47cSIlpo Järvinen if (next_skb == tp->retransmit_skb_hint) 1821ef9da47cSIlpo Järvinen tp->retransmit_skb_hint = skb; 1822b7689205SIlpo Järvinen 18233ab224beSHideo Aoki sk_wmem_free_skb(sk, next_skb); 18241da177e4SLinus Torvalds } 18251da177e4SLinus Torvalds 1826*4a17fc3aSIlpo Järvinen static int tcp_can_collapse(struct sock *sk, struct sk_buff *skb) 1827*4a17fc3aSIlpo Järvinen { 1828*4a17fc3aSIlpo Järvinen if (tcp_skb_pcount(skb) > 1) 1829*4a17fc3aSIlpo Järvinen return 0; 1830*4a17fc3aSIlpo Järvinen /* TODO: SACK collapsing could be used to remove this condition */ 1831*4a17fc3aSIlpo Järvinen if (skb_shinfo(skb)->nr_frags != 0) 1832*4a17fc3aSIlpo Järvinen return 0; 1833*4a17fc3aSIlpo Järvinen if (skb_cloned(skb)) 1834*4a17fc3aSIlpo Järvinen return 0; 1835*4a17fc3aSIlpo Järvinen if (skb == tcp_send_head(sk)) 1836*4a17fc3aSIlpo Järvinen return 0; 1837*4a17fc3aSIlpo Järvinen /* Some heurestics for collapsing over SACK'd could be invented */ 1838*4a17fc3aSIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 1839*4a17fc3aSIlpo Järvinen return 0; 1840*4a17fc3aSIlpo Järvinen 1841*4a17fc3aSIlpo Järvinen return 1; 1842*4a17fc3aSIlpo Järvinen } 1843*4a17fc3aSIlpo Järvinen 1844*4a17fc3aSIlpo Järvinen static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to, 1845*4a17fc3aSIlpo Järvinen int space) 1846*4a17fc3aSIlpo Järvinen { 1847*4a17fc3aSIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 1848*4a17fc3aSIlpo Järvinen struct sk_buff *skb = to, *tmp; 1849*4a17fc3aSIlpo Järvinen int first = 1; 1850*4a17fc3aSIlpo Järvinen 1851*4a17fc3aSIlpo Järvinen if (!sysctl_tcp_retrans_collapse) 1852*4a17fc3aSIlpo Järvinen return; 1853*4a17fc3aSIlpo Järvinen if (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_SYN) 1854*4a17fc3aSIlpo Järvinen return; 1855*4a17fc3aSIlpo Järvinen 1856*4a17fc3aSIlpo Järvinen tcp_for_write_queue_from_safe(skb, tmp, sk) { 1857*4a17fc3aSIlpo Järvinen if (!tcp_can_collapse(sk, skb)) 1858*4a17fc3aSIlpo Järvinen break; 1859*4a17fc3aSIlpo Järvinen 1860*4a17fc3aSIlpo Järvinen space -= skb->len; 1861*4a17fc3aSIlpo Järvinen 1862*4a17fc3aSIlpo Järvinen if (first) { 1863*4a17fc3aSIlpo Järvinen first = 0; 1864*4a17fc3aSIlpo Järvinen continue; 1865*4a17fc3aSIlpo Järvinen } 1866*4a17fc3aSIlpo Järvinen 1867*4a17fc3aSIlpo Järvinen if (space < 0) 1868*4a17fc3aSIlpo Järvinen break; 1869*4a17fc3aSIlpo Järvinen /* Punt if not enough space exists in the first SKB for 1870*4a17fc3aSIlpo Järvinen * the data in the second 1871*4a17fc3aSIlpo Järvinen */ 1872*4a17fc3aSIlpo Järvinen if (skb->len > skb_tailroom(to)) 1873*4a17fc3aSIlpo Järvinen break; 1874*4a17fc3aSIlpo Järvinen 1875*4a17fc3aSIlpo Järvinen if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp))) 1876*4a17fc3aSIlpo Järvinen break; 1877*4a17fc3aSIlpo Järvinen 1878*4a17fc3aSIlpo Järvinen tcp_collapse_retrans(sk, to); 1879*4a17fc3aSIlpo Järvinen } 1880*4a17fc3aSIlpo Järvinen } 1881*4a17fc3aSIlpo Järvinen 18821da177e4SLinus Torvalds /* Do a simple retransmit without using the backoff mechanisms in 18831da177e4SLinus Torvalds * tcp_timer. This is used for path mtu discovery. 18841da177e4SLinus Torvalds * The socket is already locked here. 18851da177e4SLinus Torvalds */ 18861da177e4SLinus Torvalds void tcp_simple_retransmit(struct sock *sk) 18871da177e4SLinus Torvalds { 18886687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 18891da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 18901da177e4SLinus Torvalds struct sk_buff *skb; 18911da177e4SLinus Torvalds unsigned int mss = tcp_current_mss(sk, 0); 1892006f582cSIlpo Järvinen u32 prior_lost = tp->lost_out; 18931da177e4SLinus Torvalds 1894fe067e8aSDavid S. Miller tcp_for_write_queue(skb, sk) { 1895fe067e8aSDavid S. Miller if (skb == tcp_send_head(sk)) 1896fe067e8aSDavid S. Miller break; 18971da177e4SLinus Torvalds if (skb->len > mss && 18981da177e4SLinus Torvalds !(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) { 18991da177e4SLinus Torvalds if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) { 19001da177e4SLinus Torvalds TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS; 19011da177e4SLinus Torvalds tp->retrans_out -= tcp_skb_pcount(skb); 19021da177e4SLinus Torvalds } 1903006f582cSIlpo Järvinen tcp_skb_mark_lost_uncond_verify(tp, skb); 19041da177e4SLinus Torvalds } 19051da177e4SLinus Torvalds } 19061da177e4SLinus Torvalds 1907ef9da47cSIlpo Järvinen tcp_clear_retrans_hints_partial(tp); 19086a438bbeSStephen Hemminger 1909006f582cSIlpo Järvinen if (prior_lost == tp->lost_out) 19101da177e4SLinus Torvalds return; 19111da177e4SLinus Torvalds 1912882bebaaSIlpo Järvinen if (tcp_is_reno(tp)) 1913882bebaaSIlpo Järvinen tcp_limit_reno_sacked(tp); 1914882bebaaSIlpo Järvinen 1915005903bcSIlpo Järvinen tcp_verify_left_out(tp); 19161da177e4SLinus Torvalds 19171da177e4SLinus Torvalds /* Don't muck with the congestion window here. 19181da177e4SLinus Torvalds * Reason is that we do not increase amount of _data_ 19191da177e4SLinus Torvalds * in network, but units changed and effective 19201da177e4SLinus Torvalds * cwnd/ssthresh really reduced now. 19211da177e4SLinus Torvalds */ 19226687e988SArnaldo Carvalho de Melo if (icsk->icsk_ca_state != TCP_CA_Loss) { 19231da177e4SLinus Torvalds tp->high_seq = tp->snd_nxt; 19246687e988SArnaldo Carvalho de Melo tp->snd_ssthresh = tcp_current_ssthresh(sk); 19251da177e4SLinus Torvalds tp->prior_ssthresh = 0; 19261da177e4SLinus Torvalds tp->undo_marker = 0; 19276687e988SArnaldo Carvalho de Melo tcp_set_ca_state(sk, TCP_CA_Loss); 19281da177e4SLinus Torvalds } 19291da177e4SLinus Torvalds tcp_xmit_retransmit_queue(sk); 19301da177e4SLinus Torvalds } 19311da177e4SLinus Torvalds 19321da177e4SLinus Torvalds /* This retransmits one SKB. Policy decisions and retransmit queue 19331da177e4SLinus Torvalds * state updates are done by the caller. Returns non-zero if an 19341da177e4SLinus Torvalds * error occurred which prevented the send. 19351da177e4SLinus Torvalds */ 19361da177e4SLinus Torvalds int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb) 19371da177e4SLinus Torvalds { 19381da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 19395d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 19407d227cd2SSridhar Samudrala unsigned int cur_mss; 19411da177e4SLinus Torvalds int err; 19421da177e4SLinus Torvalds 19435d424d5aSJohn Heffner /* Inconslusive MTU probe */ 19445d424d5aSJohn Heffner if (icsk->icsk_mtup.probe_size) { 19455d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = 0; 19465d424d5aSJohn Heffner } 19475d424d5aSJohn Heffner 19481da177e4SLinus Torvalds /* Do not sent more than we queued. 1/4 is reserved for possible 1949caa20d9aSStephen Hemminger * copying overhead: fragmentation, tunneling, mangling etc. 19501da177e4SLinus Torvalds */ 19511da177e4SLinus Torvalds if (atomic_read(&sk->sk_wmem_alloc) > 19521da177e4SLinus Torvalds min(sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), sk->sk_sndbuf)) 19531da177e4SLinus Torvalds return -EAGAIN; 19541da177e4SLinus Torvalds 19551da177e4SLinus Torvalds if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) { 19561da177e4SLinus Torvalds if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) 19571da177e4SLinus Torvalds BUG(); 19581da177e4SLinus Torvalds if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq)) 19591da177e4SLinus Torvalds return -ENOMEM; 19601da177e4SLinus Torvalds } 19611da177e4SLinus Torvalds 19627d227cd2SSridhar Samudrala if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk)) 19637d227cd2SSridhar Samudrala return -EHOSTUNREACH; /* Routing failure or similar. */ 19647d227cd2SSridhar Samudrala 19657d227cd2SSridhar Samudrala cur_mss = tcp_current_mss(sk, 0); 19667d227cd2SSridhar Samudrala 19671da177e4SLinus Torvalds /* If receiver has shrunk his window, and skb is out of 19681da177e4SLinus Torvalds * new window, do not retransmit it. The exception is the 19691da177e4SLinus Torvalds * case, when window is shrunk to zero. In this case 19701da177e4SLinus Torvalds * our retransmit serves as a zero window probe. 19711da177e4SLinus Torvalds */ 197290840defSIlpo Järvinen if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) 19731da177e4SLinus Torvalds && TCP_SKB_CB(skb)->seq != tp->snd_una) 19741da177e4SLinus Torvalds return -EAGAIN; 19751da177e4SLinus Torvalds 19761da177e4SLinus Torvalds if (skb->len > cur_mss) { 1977846998aeSDavid S. Miller if (tcp_fragment(sk, skb, cur_mss, cur_mss)) 19781da177e4SLinus Torvalds return -ENOMEM; /* We'll try again later. */ 19791da177e4SLinus Torvalds } 19801da177e4SLinus Torvalds 19811da177e4SLinus Torvalds tcp_retrans_try_collapse(sk, skb, cur_mss); 19821da177e4SLinus Torvalds 19831da177e4SLinus Torvalds /* Some Solaris stacks overoptimize and ignore the FIN on a 19841da177e4SLinus Torvalds * retransmit when old data is attached. So strip it off 19851da177e4SLinus Torvalds * since it is cheap to do so and saves bytes on the network. 19861da177e4SLinus Torvalds */ 19871da177e4SLinus Torvalds if (skb->len > 0 && 19881da177e4SLinus Torvalds (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN) && 19891da177e4SLinus Torvalds tp->snd_una == (TCP_SKB_CB(skb)->end_seq - 1)) { 19901da177e4SLinus Torvalds if (!pskb_trim(skb, 0)) { 1991e870a8efSIlpo Järvinen /* Reuse, even though it does some unnecessary work */ 1992e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, TCP_SKB_CB(skb)->end_seq - 1, 1993e870a8efSIlpo Järvinen TCP_SKB_CB(skb)->flags); 19941da177e4SLinus Torvalds skb->ip_summed = CHECKSUM_NONE; 19951da177e4SLinus Torvalds } 19961da177e4SLinus Torvalds } 19971da177e4SLinus Torvalds 19981da177e4SLinus Torvalds /* Make a copy, if the first transmission SKB clone we made 19991da177e4SLinus Torvalds * is still in somebody's hands, else make a clone. 20001da177e4SLinus Torvalds */ 20011da177e4SLinus Torvalds TCP_SKB_CB(skb)->when = tcp_time_stamp; 20021da177e4SLinus Torvalds 2003dfb4b9dcSDavid S. Miller err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 20041da177e4SLinus Torvalds 20051da177e4SLinus Torvalds if (err == 0) { 20061da177e4SLinus Torvalds /* Update global TCP statistics. */ 200781cc8a75SPavel Emelyanov TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS); 20081da177e4SLinus Torvalds 20091da177e4SLinus Torvalds tp->total_retrans++; 20101da177e4SLinus Torvalds 20111da177e4SLinus Torvalds #if FASTRETRANS_DEBUG > 0 20121da177e4SLinus Torvalds if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) { 20131da177e4SLinus Torvalds if (net_ratelimit()) 20141da177e4SLinus Torvalds printk(KERN_DEBUG "retrans_out leaked.\n"); 20151da177e4SLinus Torvalds } 20161da177e4SLinus Torvalds #endif 2017b08d6cb2SIlpo Järvinen if (!tp->retrans_out) 2018b08d6cb2SIlpo Järvinen tp->lost_retrans_low = tp->snd_nxt; 20191da177e4SLinus Torvalds TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS; 20201da177e4SLinus Torvalds tp->retrans_out += tcp_skb_pcount(skb); 20211da177e4SLinus Torvalds 20221da177e4SLinus Torvalds /* Save stamp of the first retransmit. */ 20231da177e4SLinus Torvalds if (!tp->retrans_stamp) 20241da177e4SLinus Torvalds tp->retrans_stamp = TCP_SKB_CB(skb)->when; 20251da177e4SLinus Torvalds 20261da177e4SLinus Torvalds tp->undo_retrans++; 20271da177e4SLinus Torvalds 20281da177e4SLinus Torvalds /* snd_nxt is stored to detect loss of retransmitted segment, 20291da177e4SLinus Torvalds * see tcp_input.c tcp_sacktag_write_queue(). 20301da177e4SLinus Torvalds */ 20311da177e4SLinus Torvalds TCP_SKB_CB(skb)->ack_seq = tp->snd_nxt; 20321da177e4SLinus Torvalds } 20331da177e4SLinus Torvalds return err; 20341da177e4SLinus Torvalds } 20351da177e4SLinus Torvalds 2036b5afe7bcSIlpo Järvinen static int tcp_can_forward_retransmit(struct sock *sk) 2037b5afe7bcSIlpo Järvinen { 2038b5afe7bcSIlpo Järvinen const struct inet_connection_sock *icsk = inet_csk(sk); 2039b5afe7bcSIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 2040b5afe7bcSIlpo Järvinen 2041b5afe7bcSIlpo Järvinen /* Forward retransmissions are possible only during Recovery. */ 2042b5afe7bcSIlpo Järvinen if (icsk->icsk_ca_state != TCP_CA_Recovery) 2043b5afe7bcSIlpo Järvinen return 0; 2044b5afe7bcSIlpo Järvinen 2045b5afe7bcSIlpo Järvinen /* No forward retransmissions in Reno are possible. */ 2046b5afe7bcSIlpo Järvinen if (tcp_is_reno(tp)) 2047b5afe7bcSIlpo Järvinen return 0; 2048b5afe7bcSIlpo Järvinen 2049b5afe7bcSIlpo Järvinen /* Yeah, we have to make difficult choice between forward transmission 2050b5afe7bcSIlpo Järvinen * and retransmission... Both ways have their merits... 2051b5afe7bcSIlpo Järvinen * 2052b5afe7bcSIlpo Järvinen * For now we do not retransmit anything, while we have some new 2053b5afe7bcSIlpo Järvinen * segments to send. In the other cases, follow rule 3 for 2054b5afe7bcSIlpo Järvinen * NextSeg() specified in RFC3517. 2055b5afe7bcSIlpo Järvinen */ 2056b5afe7bcSIlpo Järvinen 2057b5afe7bcSIlpo Järvinen if (tcp_may_send_now(sk)) 2058b5afe7bcSIlpo Järvinen return 0; 2059b5afe7bcSIlpo Järvinen 2060b5afe7bcSIlpo Järvinen return 1; 2061b5afe7bcSIlpo Järvinen } 2062b5afe7bcSIlpo Järvinen 20631da177e4SLinus Torvalds /* This gets called after a retransmit timeout, and the initially 20641da177e4SLinus Torvalds * retransmitted data is acknowledged. It tries to continue 20651da177e4SLinus Torvalds * resending the rest of the retransmit queue, until either 20661da177e4SLinus Torvalds * we've sent it all or the congestion window limit is reached. 20671da177e4SLinus Torvalds * If doing SACK, the first ACK which comes back for a timeout 20681da177e4SLinus Torvalds * based retransmit packet might feed us FACK information again. 20691da177e4SLinus Torvalds * If so, we use it to avoid unnecessarily retransmissions. 20701da177e4SLinus Torvalds */ 20711da177e4SLinus Torvalds void tcp_xmit_retransmit_queue(struct sock *sk) 20721da177e4SLinus Torvalds { 20736687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 20741da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 20751da177e4SLinus Torvalds struct sk_buff *skb; 20760e1c54c2SIlpo Järvinen struct sk_buff *hole = NULL; 2077618d9f25SIlpo Järvinen u32 last_lost; 207861eb55f4SIlpo Järvinen int mib_idx; 20790e1c54c2SIlpo Järvinen int fwd_rexmitting = 0; 20806a438bbeSStephen Hemminger 208108ebd172SIlpo Järvinen if (!tp->lost_out) 208208ebd172SIlpo Järvinen tp->retransmit_high = tp->snd_una; 208308ebd172SIlpo Järvinen 2084618d9f25SIlpo Järvinen if (tp->retransmit_skb_hint) { 20856a438bbeSStephen Hemminger skb = tp->retransmit_skb_hint; 2086618d9f25SIlpo Järvinen last_lost = TCP_SKB_CB(skb)->end_seq; 2087618d9f25SIlpo Järvinen if (after(last_lost, tp->retransmit_high)) 2088618d9f25SIlpo Järvinen last_lost = tp->retransmit_high; 2089618d9f25SIlpo Järvinen } else { 2090fe067e8aSDavid S. Miller skb = tcp_write_queue_head(sk); 2091618d9f25SIlpo Järvinen last_lost = tp->snd_una; 2092618d9f25SIlpo Järvinen } 20931da177e4SLinus Torvalds 20941da177e4SLinus Torvalds /* First pass: retransmit lost packets. */ 2095fe067e8aSDavid S. Miller tcp_for_write_queue_from(skb, sk) { 20961da177e4SLinus Torvalds __u8 sacked = TCP_SKB_CB(skb)->sacked; 20971da177e4SLinus Torvalds 2098fe067e8aSDavid S. Miller if (skb == tcp_send_head(sk)) 2099fe067e8aSDavid S. Miller break; 21006a438bbeSStephen Hemminger /* we could do better than to assign each time */ 21010e1c54c2SIlpo Järvinen if (hole == NULL) 21026a438bbeSStephen Hemminger tp->retransmit_skb_hint = skb; 21036a438bbeSStephen Hemminger 21041da177e4SLinus Torvalds /* Assume this retransmit will generate 21051da177e4SLinus Torvalds * only one packet for congestion window 21061da177e4SLinus Torvalds * calculation purposes. This works because 21071da177e4SLinus Torvalds * tcp_retransmit_skb() will chop up the 21081da177e4SLinus Torvalds * packet to be MSS sized and all the 21091da177e4SLinus Torvalds * packet counting works out. 21101da177e4SLinus Torvalds */ 21111da177e4SLinus Torvalds if (tcp_packets_in_flight(tp) >= tp->snd_cwnd) 21121da177e4SLinus Torvalds return; 21130e1c54c2SIlpo Järvinen 21140e1c54c2SIlpo Järvinen if (fwd_rexmitting) { 21150e1c54c2SIlpo Järvinen begin_fwd: 21160e1c54c2SIlpo Järvinen if (!before(TCP_SKB_CB(skb)->seq, tcp_highest_sack_seq(tp))) 2117006f582cSIlpo Järvinen break; 21180e1c54c2SIlpo Järvinen mib_idx = LINUX_MIB_TCPFORWARDRETRANS; 21190e1c54c2SIlpo Järvinen 21200e1c54c2SIlpo Järvinen } else if (!before(TCP_SKB_CB(skb)->seq, tp->retransmit_high)) { 2121618d9f25SIlpo Järvinen tp->retransmit_high = last_lost; 21220e1c54c2SIlpo Järvinen if (!tcp_can_forward_retransmit(sk)) 21230e1c54c2SIlpo Järvinen break; 21240e1c54c2SIlpo Järvinen /* Backtrack if necessary to non-L'ed skb */ 21250e1c54c2SIlpo Järvinen if (hole != NULL) { 21260e1c54c2SIlpo Järvinen skb = hole; 21270e1c54c2SIlpo Järvinen hole = NULL; 21280e1c54c2SIlpo Järvinen } 21290e1c54c2SIlpo Järvinen fwd_rexmitting = 1; 21300e1c54c2SIlpo Järvinen goto begin_fwd; 21310e1c54c2SIlpo Järvinen 21320e1c54c2SIlpo Järvinen } else if (!(sacked & TCPCB_LOST)) { 21330e1c54c2SIlpo Järvinen if (hole == NULL && !(sacked & TCPCB_SACKED_RETRANS)) 21340e1c54c2SIlpo Järvinen hole = skb; 213561eb55f4SIlpo Järvinen continue; 21361da177e4SLinus Torvalds 21370e1c54c2SIlpo Järvinen } else { 2138618d9f25SIlpo Järvinen last_lost = TCP_SKB_CB(skb)->end_seq; 21390e1c54c2SIlpo Järvinen if (icsk->icsk_ca_state != TCP_CA_Loss) 21400e1c54c2SIlpo Järvinen mib_idx = LINUX_MIB_TCPFASTRETRANS; 21410e1c54c2SIlpo Järvinen else 21420e1c54c2SIlpo Järvinen mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS; 21430e1c54c2SIlpo Järvinen } 21440e1c54c2SIlpo Järvinen 21450e1c54c2SIlpo Järvinen if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS)) 214661eb55f4SIlpo Järvinen continue; 214740b215e5SPavel Emelyanov 2148f0ceb0edSIlpo Järvinen if (tcp_retransmit_skb(sk, skb)) 21491da177e4SLinus Torvalds return; 2150de0744afSPavel Emelyanov NET_INC_STATS_BH(sock_net(sk), mib_idx); 21511da177e4SLinus Torvalds 2152fe067e8aSDavid S. Miller if (skb == tcp_write_queue_head(sk)) 2153463c84b9SArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 21543f421baaSArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto, 21553f421baaSArnaldo Carvalho de Melo TCP_RTO_MAX); 21561da177e4SLinus Torvalds } 21571da177e4SLinus Torvalds } 21581da177e4SLinus Torvalds 21591da177e4SLinus Torvalds /* Send a fin. The caller locks the socket for us. This cannot be 21601da177e4SLinus Torvalds * allowed to fail queueing a FIN frame under any circumstances. 21611da177e4SLinus Torvalds */ 21621da177e4SLinus Torvalds void tcp_send_fin(struct sock *sk) 21631da177e4SLinus Torvalds { 21641da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 2165fe067e8aSDavid S. Miller struct sk_buff *skb = tcp_write_queue_tail(sk); 21661da177e4SLinus Torvalds int mss_now; 21671da177e4SLinus Torvalds 21681da177e4SLinus Torvalds /* Optimization, tack on the FIN if we have a queue of 21691da177e4SLinus Torvalds * unsent frames. But be careful about outgoing SACKS 21701da177e4SLinus Torvalds * and IP options. 21711da177e4SLinus Torvalds */ 21721da177e4SLinus Torvalds mss_now = tcp_current_mss(sk, 1); 21731da177e4SLinus Torvalds 2174fe067e8aSDavid S. Miller if (tcp_send_head(sk) != NULL) { 21751da177e4SLinus Torvalds TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_FIN; 21761da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq++; 21771da177e4SLinus Torvalds tp->write_seq++; 21781da177e4SLinus Torvalds } else { 21791da177e4SLinus Torvalds /* Socket is locked, keep trying until memory is available. */ 21801da177e4SLinus Torvalds for (;;) { 2181d179cd12SDavid S. Miller skb = alloc_skb_fclone(MAX_TCP_HEADER, GFP_KERNEL); 21821da177e4SLinus Torvalds if (skb) 21831da177e4SLinus Torvalds break; 21841da177e4SLinus Torvalds yield(); 21851da177e4SLinus Torvalds } 21861da177e4SLinus Torvalds 21871da177e4SLinus Torvalds /* Reserve space for headers and prepare control bits. */ 21881da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 21891da177e4SLinus Torvalds /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */ 2190e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tp->write_seq, 2191e870a8efSIlpo Järvinen TCPCB_FLAG_ACK | TCPCB_FLAG_FIN); 21921da177e4SLinus Torvalds tcp_queue_skb(sk, skb); 21931da177e4SLinus Torvalds } 21949e412ba7SIlpo Järvinen __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_OFF); 21951da177e4SLinus Torvalds } 21961da177e4SLinus Torvalds 21971da177e4SLinus Torvalds /* We get here when a process closes a file descriptor (either due to 21981da177e4SLinus Torvalds * an explicit close() or as a byproduct of exit()'ing) and there 21991da177e4SLinus Torvalds * was unread data in the receive queue. This behavior is recommended 220065bb723cSGerrit Renker * by RFC 2525, section 2.17. -DaveM 22011da177e4SLinus Torvalds */ 2202dd0fc66fSAl Viro void tcp_send_active_reset(struct sock *sk, gfp_t priority) 22031da177e4SLinus Torvalds { 22041da177e4SLinus Torvalds struct sk_buff *skb; 22051da177e4SLinus Torvalds 22061da177e4SLinus Torvalds /* NOTE: No TCP options attached and we never retransmit this. */ 22071da177e4SLinus Torvalds skb = alloc_skb(MAX_TCP_HEADER, priority); 22081da177e4SLinus Torvalds if (!skb) { 22094e673444SPavel Emelyanov NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); 22101da177e4SLinus Torvalds return; 22111da177e4SLinus Torvalds } 22121da177e4SLinus Torvalds 22131da177e4SLinus Torvalds /* Reserve space for headers and prepare control bits. */ 22141da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 2215e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk), 2216e870a8efSIlpo Järvinen TCPCB_FLAG_ACK | TCPCB_FLAG_RST); 22171da177e4SLinus Torvalds /* Send it off. */ 22181da177e4SLinus Torvalds TCP_SKB_CB(skb)->when = tcp_time_stamp; 2219dfb4b9dcSDavid S. Miller if (tcp_transmit_skb(sk, skb, 0, priority)) 22204e673444SPavel Emelyanov NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); 222126af65cbSSridhar Samudrala 222281cc8a75SPavel Emelyanov TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS); 22231da177e4SLinus Torvalds } 22241da177e4SLinus Torvalds 22251da177e4SLinus Torvalds /* WARNING: This routine must only be called when we have already sent 22261da177e4SLinus Torvalds * a SYN packet that crossed the incoming SYN that caused this routine 22271da177e4SLinus Torvalds * to get called. If this assumption fails then the initial rcv_wnd 22281da177e4SLinus Torvalds * and rcv_wscale values will not be correct. 22291da177e4SLinus Torvalds */ 22301da177e4SLinus Torvalds int tcp_send_synack(struct sock *sk) 22311da177e4SLinus Torvalds { 22321da177e4SLinus Torvalds struct sk_buff *skb; 22331da177e4SLinus Torvalds 2234fe067e8aSDavid S. Miller skb = tcp_write_queue_head(sk); 22351da177e4SLinus Torvalds if (skb == NULL || !(TCP_SKB_CB(skb)->flags & TCPCB_FLAG_SYN)) { 22361da177e4SLinus Torvalds printk(KERN_DEBUG "tcp_send_synack: wrong queue state\n"); 22371da177e4SLinus Torvalds return -EFAULT; 22381da177e4SLinus Torvalds } 22391da177e4SLinus Torvalds if (!(TCP_SKB_CB(skb)->flags & TCPCB_FLAG_ACK)) { 22401da177e4SLinus Torvalds if (skb_cloned(skb)) { 22411da177e4SLinus Torvalds struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC); 22421da177e4SLinus Torvalds if (nskb == NULL) 22431da177e4SLinus Torvalds return -ENOMEM; 2244fe067e8aSDavid S. Miller tcp_unlink_write_queue(skb, sk); 22451da177e4SLinus Torvalds skb_header_release(nskb); 2246fe067e8aSDavid S. Miller __tcp_add_write_queue_head(sk, nskb); 22473ab224beSHideo Aoki sk_wmem_free_skb(sk, skb); 22483ab224beSHideo Aoki sk->sk_wmem_queued += nskb->truesize; 22493ab224beSHideo Aoki sk_mem_charge(sk, nskb->truesize); 22501da177e4SLinus Torvalds skb = nskb; 22511da177e4SLinus Torvalds } 22521da177e4SLinus Torvalds 22531da177e4SLinus Torvalds TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_ACK; 22541da177e4SLinus Torvalds TCP_ECN_send_synack(tcp_sk(sk), skb); 22551da177e4SLinus Torvalds } 22561da177e4SLinus Torvalds TCP_SKB_CB(skb)->when = tcp_time_stamp; 2257dfb4b9dcSDavid S. Miller return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 22581da177e4SLinus Torvalds } 22591da177e4SLinus Torvalds 22601da177e4SLinus Torvalds /* 22611da177e4SLinus Torvalds * Prepare a SYN-ACK. 22621da177e4SLinus Torvalds */ 22631da177e4SLinus Torvalds struct sk_buff *tcp_make_synack(struct sock *sk, struct dst_entry *dst, 226460236fddSArnaldo Carvalho de Melo struct request_sock *req) 22651da177e4SLinus Torvalds { 22662e6599cbSArnaldo Carvalho de Melo struct inet_request_sock *ireq = inet_rsk(req); 22671da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 22681da177e4SLinus Torvalds struct tcphdr *th; 22691da177e4SLinus Torvalds int tcp_header_size; 227033ad798cSAdam Langley struct tcp_out_options opts; 22711da177e4SLinus Torvalds struct sk_buff *skb; 2272cfb6eeb4SYOSHIFUJI Hideaki struct tcp_md5sig_key *md5; 2273cfb6eeb4SYOSHIFUJI Hideaki __u8 *md5_hash_location; 2274f5fff5dcSTom Quetchenbach int mss; 22751da177e4SLinus Torvalds 22761da177e4SLinus Torvalds skb = sock_wmalloc(sk, MAX_TCP_HEADER + 15, 1, GFP_ATOMIC); 22771da177e4SLinus Torvalds if (skb == NULL) 22781da177e4SLinus Torvalds return NULL; 22791da177e4SLinus Torvalds 22801da177e4SLinus Torvalds /* Reserve space for headers. */ 22811da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 22821da177e4SLinus Torvalds 22831da177e4SLinus Torvalds skb->dst = dst_clone(dst); 22841da177e4SLinus Torvalds 2285f5fff5dcSTom Quetchenbach mss = dst_metric(dst, RTAX_ADVMSS); 2286f5fff5dcSTom Quetchenbach if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < mss) 2287f5fff5dcSTom Quetchenbach mss = tp->rx_opt.user_mss; 2288f5fff5dcSTom Quetchenbach 228933ad798cSAdam Langley if (req->rcv_wnd == 0) { /* ignored for retransmitted syns */ 229033ad798cSAdam Langley __u8 rcv_wscale; 229133ad798cSAdam Langley /* Set this up on the first call only */ 229233ad798cSAdam Langley req->window_clamp = tp->window_clamp ? : dst_metric(dst, RTAX_WINDOW); 229333ad798cSAdam Langley /* tcp_full_space because it is guaranteed to be the first packet */ 229433ad798cSAdam Langley tcp_select_initial_window(tcp_full_space(sk), 2295f5fff5dcSTom Quetchenbach mss - (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0), 229633ad798cSAdam Langley &req->rcv_wnd, 229733ad798cSAdam Langley &req->window_clamp, 229833ad798cSAdam Langley ireq->wscale_ok, 229933ad798cSAdam Langley &rcv_wscale); 230033ad798cSAdam Langley ireq->rcv_wscale = rcv_wscale; 230133ad798cSAdam Langley } 2302cfb6eeb4SYOSHIFUJI Hideaki 230333ad798cSAdam Langley memset(&opts, 0, sizeof(opts)); 23048b5f12d0SFlorian Westphal #ifdef CONFIG_SYN_COOKIES 23058b5f12d0SFlorian Westphal if (unlikely(req->cookie_ts)) 23068b5f12d0SFlorian Westphal TCP_SKB_CB(skb)->when = cookie_init_timestamp(req); 23078b5f12d0SFlorian Westphal else 23088b5f12d0SFlorian Westphal #endif 230933ad798cSAdam Langley TCP_SKB_CB(skb)->when = tcp_time_stamp; 2310f5fff5dcSTom Quetchenbach tcp_header_size = tcp_synack_options(sk, req, mss, 231133ad798cSAdam Langley skb, &opts, &md5) + 231233ad798cSAdam Langley sizeof(struct tcphdr); 231333ad798cSAdam Langley 2314aa8223c7SArnaldo Carvalho de Melo skb_push(skb, tcp_header_size); 2315aa8223c7SArnaldo Carvalho de Melo skb_reset_transport_header(skb); 23161da177e4SLinus Torvalds 2317aa8223c7SArnaldo Carvalho de Melo th = tcp_hdr(skb); 23181da177e4SLinus Torvalds memset(th, 0, sizeof(struct tcphdr)); 23191da177e4SLinus Torvalds th->syn = 1; 23201da177e4SLinus Torvalds th->ack = 1; 23211da177e4SLinus Torvalds TCP_ECN_make_synack(req, th); 2322a3116ac5SKOVACS Krisztian th->source = ireq->loc_port; 23232e6599cbSArnaldo Carvalho de Melo th->dest = ireq->rmt_port; 2324e870a8efSIlpo Järvinen /* Setting of flags are superfluous here for callers (and ECE is 2325e870a8efSIlpo Järvinen * not even correctly set) 2326e870a8efSIlpo Järvinen */ 2327e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tcp_rsk(req)->snt_isn, 2328e870a8efSIlpo Järvinen TCPCB_FLAG_SYN | TCPCB_FLAG_ACK); 23291da177e4SLinus Torvalds th->seq = htonl(TCP_SKB_CB(skb)->seq); 23302e6599cbSArnaldo Carvalho de Melo th->ack_seq = htonl(tcp_rsk(req)->rcv_isn + 1); 23311da177e4SLinus Torvalds 23321da177e4SLinus Torvalds /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */ 2333600ff0c2SIlpo Järvinen th->window = htons(min(req->rcv_wnd, 65535U)); 233433ad798cSAdam Langley tcp_options_write((__be32 *)(th + 1), tp, &opts, &md5_hash_location); 23351da177e4SLinus Torvalds th->doff = (tcp_header_size >> 2); 233681cc8a75SPavel Emelyanov TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTSEGS); 2337cfb6eeb4SYOSHIFUJI Hideaki 2338cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 2339cfb6eeb4SYOSHIFUJI Hideaki /* Okay, we have all we need - do the md5 hash if needed */ 2340cfb6eeb4SYOSHIFUJI Hideaki if (md5) { 2341cfb6eeb4SYOSHIFUJI Hideaki tp->af_specific->calc_md5_hash(md5_hash_location, 234249a72dfbSAdam Langley md5, NULL, req, skb); 2343cfb6eeb4SYOSHIFUJI Hideaki } 2344cfb6eeb4SYOSHIFUJI Hideaki #endif 2345cfb6eeb4SYOSHIFUJI Hideaki 23461da177e4SLinus Torvalds return skb; 23471da177e4SLinus Torvalds } 23481da177e4SLinus Torvalds 23491da177e4SLinus Torvalds /* 23501da177e4SLinus Torvalds * Do all connect socket setups that can be done AF independent. 23511da177e4SLinus Torvalds */ 235240efc6faSStephen Hemminger static void tcp_connect_init(struct sock *sk) 23531da177e4SLinus Torvalds { 23541da177e4SLinus Torvalds struct dst_entry *dst = __sk_dst_get(sk); 23551da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 23561da177e4SLinus Torvalds __u8 rcv_wscale; 23571da177e4SLinus Torvalds 23581da177e4SLinus Torvalds /* We'll fix this up when we get a response from the other end. 23591da177e4SLinus Torvalds * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT. 23601da177e4SLinus Torvalds */ 23611da177e4SLinus Torvalds tp->tcp_header_len = sizeof(struct tcphdr) + 23621da177e4SLinus Torvalds (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0); 23631da177e4SLinus Torvalds 2364cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 2365cfb6eeb4SYOSHIFUJI Hideaki if (tp->af_specific->md5_lookup(sk, sk) != NULL) 2366cfb6eeb4SYOSHIFUJI Hideaki tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED; 2367cfb6eeb4SYOSHIFUJI Hideaki #endif 2368cfb6eeb4SYOSHIFUJI Hideaki 23691da177e4SLinus Torvalds /* If user gave his TCP_MAXSEG, record it to clamp */ 23701da177e4SLinus Torvalds if (tp->rx_opt.user_mss) 23711da177e4SLinus Torvalds tp->rx_opt.mss_clamp = tp->rx_opt.user_mss; 23721da177e4SLinus Torvalds tp->max_window = 0; 23735d424d5aSJohn Heffner tcp_mtup_init(sk); 23741da177e4SLinus Torvalds tcp_sync_mss(sk, dst_mtu(dst)); 23751da177e4SLinus Torvalds 23761da177e4SLinus Torvalds if (!tp->window_clamp) 23771da177e4SLinus Torvalds tp->window_clamp = dst_metric(dst, RTAX_WINDOW); 23781da177e4SLinus Torvalds tp->advmss = dst_metric(dst, RTAX_ADVMSS); 2379f5fff5dcSTom Quetchenbach if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->advmss) 2380f5fff5dcSTom Quetchenbach tp->advmss = tp->rx_opt.user_mss; 2381f5fff5dcSTom Quetchenbach 23821da177e4SLinus Torvalds tcp_initialize_rcv_mss(sk); 23831da177e4SLinus Torvalds 23841da177e4SLinus Torvalds tcp_select_initial_window(tcp_full_space(sk), 23851da177e4SLinus Torvalds tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0), 23861da177e4SLinus Torvalds &tp->rcv_wnd, 23871da177e4SLinus Torvalds &tp->window_clamp, 23881da177e4SLinus Torvalds sysctl_tcp_window_scaling, 23891da177e4SLinus Torvalds &rcv_wscale); 23901da177e4SLinus Torvalds 23911da177e4SLinus Torvalds tp->rx_opt.rcv_wscale = rcv_wscale; 23921da177e4SLinus Torvalds tp->rcv_ssthresh = tp->rcv_wnd; 23931da177e4SLinus Torvalds 23941da177e4SLinus Torvalds sk->sk_err = 0; 23951da177e4SLinus Torvalds sock_reset_flag(sk, SOCK_DONE); 23961da177e4SLinus Torvalds tp->snd_wnd = 0; 23971da177e4SLinus Torvalds tcp_init_wl(tp, tp->write_seq, 0); 23981da177e4SLinus Torvalds tp->snd_una = tp->write_seq; 23991da177e4SLinus Torvalds tp->snd_sml = tp->write_seq; 240033f5f57eSIlpo Järvinen tp->snd_up = tp->write_seq; 24011da177e4SLinus Torvalds tp->rcv_nxt = 0; 24021da177e4SLinus Torvalds tp->rcv_wup = 0; 24031da177e4SLinus Torvalds tp->copied_seq = 0; 24041da177e4SLinus Torvalds 2405463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT; 2406463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_retransmits = 0; 24071da177e4SLinus Torvalds tcp_clear_retrans(tp); 24081da177e4SLinus Torvalds } 24091da177e4SLinus Torvalds 24101da177e4SLinus Torvalds /* 24111da177e4SLinus Torvalds * Build a SYN and send it off. 24121da177e4SLinus Torvalds */ 24131da177e4SLinus Torvalds int tcp_connect(struct sock *sk) 24141da177e4SLinus Torvalds { 24151da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 24161da177e4SLinus Torvalds struct sk_buff *buff; 24171da177e4SLinus Torvalds 24181da177e4SLinus Torvalds tcp_connect_init(sk); 24191da177e4SLinus Torvalds 2420d179cd12SDavid S. Miller buff = alloc_skb_fclone(MAX_TCP_HEADER + 15, sk->sk_allocation); 24211da177e4SLinus Torvalds if (unlikely(buff == NULL)) 24221da177e4SLinus Torvalds return -ENOBUFS; 24231da177e4SLinus Torvalds 24241da177e4SLinus Torvalds /* Reserve space for headers. */ 24251da177e4SLinus Torvalds skb_reserve(buff, MAX_TCP_HEADER); 24261da177e4SLinus Torvalds 2427bd37a088SWei Yongjun tp->snd_nxt = tp->write_seq; 2428e870a8efSIlpo Järvinen tcp_init_nondata_skb(buff, tp->write_seq++, TCPCB_FLAG_SYN); 2429e870a8efSIlpo Järvinen TCP_ECN_send_syn(sk, buff); 24301da177e4SLinus Torvalds 24311da177e4SLinus Torvalds /* Send it off. */ 24321da177e4SLinus Torvalds TCP_SKB_CB(buff)->when = tcp_time_stamp; 24331da177e4SLinus Torvalds tp->retrans_stamp = TCP_SKB_CB(buff)->when; 24341da177e4SLinus Torvalds skb_header_release(buff); 2435fe067e8aSDavid S. Miller __tcp_add_write_queue_tail(sk, buff); 24363ab224beSHideo Aoki sk->sk_wmem_queued += buff->truesize; 24373ab224beSHideo Aoki sk_mem_charge(sk, buff->truesize); 24381da177e4SLinus Torvalds tp->packets_out += tcp_skb_pcount(buff); 2439dfb4b9dcSDavid S. Miller tcp_transmit_skb(sk, buff, 1, GFP_KERNEL); 2440bd37a088SWei Yongjun 2441bd37a088SWei Yongjun /* We change tp->snd_nxt after the tcp_transmit_skb() call 2442bd37a088SWei Yongjun * in order to make this packet get counted in tcpOutSegs. 2443bd37a088SWei Yongjun */ 2444bd37a088SWei Yongjun tp->snd_nxt = tp->write_seq; 2445bd37a088SWei Yongjun tp->pushed_seq = tp->write_seq; 244681cc8a75SPavel Emelyanov TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS); 24471da177e4SLinus Torvalds 24481da177e4SLinus Torvalds /* Timer for repeating the SYN until an answer. */ 24493f421baaSArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 24503f421baaSArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto, TCP_RTO_MAX); 24511da177e4SLinus Torvalds return 0; 24521da177e4SLinus Torvalds } 24531da177e4SLinus Torvalds 24541da177e4SLinus Torvalds /* Send out a delayed ack, the caller does the policy checking 24551da177e4SLinus Torvalds * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check() 24561da177e4SLinus Torvalds * for details. 24571da177e4SLinus Torvalds */ 24581da177e4SLinus Torvalds void tcp_send_delayed_ack(struct sock *sk) 24591da177e4SLinus Torvalds { 2460463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 2461463c84b9SArnaldo Carvalho de Melo int ato = icsk->icsk_ack.ato; 24621da177e4SLinus Torvalds unsigned long timeout; 24631da177e4SLinus Torvalds 24641da177e4SLinus Torvalds if (ato > TCP_DELACK_MIN) { 2465463c84b9SArnaldo Carvalho de Melo const struct tcp_sock *tp = tcp_sk(sk); 24661da177e4SLinus Torvalds int max_ato = HZ / 2; 24671da177e4SLinus Torvalds 2468056834d9SIlpo Järvinen if (icsk->icsk_ack.pingpong || 2469056834d9SIlpo Järvinen (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)) 24701da177e4SLinus Torvalds max_ato = TCP_DELACK_MAX; 24711da177e4SLinus Torvalds 24721da177e4SLinus Torvalds /* Slow path, intersegment interval is "high". */ 24731da177e4SLinus Torvalds 24741da177e4SLinus Torvalds /* If some rtt estimate is known, use it to bound delayed ack. 2475463c84b9SArnaldo Carvalho de Melo * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements 24761da177e4SLinus Torvalds * directly. 24771da177e4SLinus Torvalds */ 24781da177e4SLinus Torvalds if (tp->srtt) { 24791da177e4SLinus Torvalds int rtt = max(tp->srtt >> 3, TCP_DELACK_MIN); 24801da177e4SLinus Torvalds 24811da177e4SLinus Torvalds if (rtt < max_ato) 24821da177e4SLinus Torvalds max_ato = rtt; 24831da177e4SLinus Torvalds } 24841da177e4SLinus Torvalds 24851da177e4SLinus Torvalds ato = min(ato, max_ato); 24861da177e4SLinus Torvalds } 24871da177e4SLinus Torvalds 24881da177e4SLinus Torvalds /* Stay within the limit we were given */ 24891da177e4SLinus Torvalds timeout = jiffies + ato; 24901da177e4SLinus Torvalds 24911da177e4SLinus Torvalds /* Use new timeout only if there wasn't a older one earlier. */ 2492463c84b9SArnaldo Carvalho de Melo if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) { 24931da177e4SLinus Torvalds /* If delack timer was blocked or is about to expire, 24941da177e4SLinus Torvalds * send ACK now. 24951da177e4SLinus Torvalds */ 2496463c84b9SArnaldo Carvalho de Melo if (icsk->icsk_ack.blocked || 2497463c84b9SArnaldo Carvalho de Melo time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) { 24981da177e4SLinus Torvalds tcp_send_ack(sk); 24991da177e4SLinus Torvalds return; 25001da177e4SLinus Torvalds } 25011da177e4SLinus Torvalds 2502463c84b9SArnaldo Carvalho de Melo if (!time_before(timeout, icsk->icsk_ack.timeout)) 2503463c84b9SArnaldo Carvalho de Melo timeout = icsk->icsk_ack.timeout; 25041da177e4SLinus Torvalds } 2505463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER; 2506463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.timeout = timeout; 2507463c84b9SArnaldo Carvalho de Melo sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout); 25081da177e4SLinus Torvalds } 25091da177e4SLinus Torvalds 25101da177e4SLinus Torvalds /* This routine sends an ack and also updates the window. */ 25111da177e4SLinus Torvalds void tcp_send_ack(struct sock *sk) 25121da177e4SLinus Torvalds { 25131da177e4SLinus Torvalds struct sk_buff *buff; 25141da177e4SLinus Torvalds 2515058dc334SIlpo Järvinen /* If we have been reset, we may not send again. */ 2516058dc334SIlpo Järvinen if (sk->sk_state == TCP_CLOSE) 2517058dc334SIlpo Järvinen return; 2518058dc334SIlpo Järvinen 25191da177e4SLinus Torvalds /* We are not putting this on the write queue, so 25201da177e4SLinus Torvalds * tcp_transmit_skb() will set the ownership to this 25211da177e4SLinus Torvalds * sock. 25221da177e4SLinus Torvalds */ 25231da177e4SLinus Torvalds buff = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC); 25241da177e4SLinus Torvalds if (buff == NULL) { 2525463c84b9SArnaldo Carvalho de Melo inet_csk_schedule_ack(sk); 2526463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN; 25273f421baaSArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK, 25283f421baaSArnaldo Carvalho de Melo TCP_DELACK_MAX, TCP_RTO_MAX); 25291da177e4SLinus Torvalds return; 25301da177e4SLinus Torvalds } 25311da177e4SLinus Torvalds 25321da177e4SLinus Torvalds /* Reserve space for headers and prepare control bits. */ 25331da177e4SLinus Torvalds skb_reserve(buff, MAX_TCP_HEADER); 2534e870a8efSIlpo Järvinen tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPCB_FLAG_ACK); 25351da177e4SLinus Torvalds 25361da177e4SLinus Torvalds /* Send it off, this clears delayed acks for us. */ 25371da177e4SLinus Torvalds TCP_SKB_CB(buff)->when = tcp_time_stamp; 2538dfb4b9dcSDavid S. Miller tcp_transmit_skb(sk, buff, 0, GFP_ATOMIC); 25391da177e4SLinus Torvalds } 25401da177e4SLinus Torvalds 25411da177e4SLinus Torvalds /* This routine sends a packet with an out of date sequence 25421da177e4SLinus Torvalds * number. It assumes the other end will try to ack it. 25431da177e4SLinus Torvalds * 25441da177e4SLinus Torvalds * Question: what should we make while urgent mode? 25451da177e4SLinus Torvalds * 4.4BSD forces sending single byte of data. We cannot send 25461da177e4SLinus Torvalds * out of window data, because we have SND.NXT==SND.MAX... 25471da177e4SLinus Torvalds * 25481da177e4SLinus Torvalds * Current solution: to send TWO zero-length segments in urgent mode: 25491da177e4SLinus Torvalds * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is 25501da177e4SLinus Torvalds * out-of-date with SND.UNA-1 to probe window. 25511da177e4SLinus Torvalds */ 25521da177e4SLinus Torvalds static int tcp_xmit_probe_skb(struct sock *sk, int urgent) 25531da177e4SLinus Torvalds { 25541da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 25551da177e4SLinus Torvalds struct sk_buff *skb; 25561da177e4SLinus Torvalds 25571da177e4SLinus Torvalds /* We don't queue it, tcp_transmit_skb() sets ownership. */ 25581da177e4SLinus Torvalds skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC); 25591da177e4SLinus Torvalds if (skb == NULL) 25601da177e4SLinus Torvalds return -1; 25611da177e4SLinus Torvalds 25621da177e4SLinus Torvalds /* Reserve space for headers and set control bits. */ 25631da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 25641da177e4SLinus Torvalds /* Use a previous sequence. This should cause the other 25651da177e4SLinus Torvalds * end to send an ack. Don't queue or clone SKB, just 25661da177e4SLinus Torvalds * send it. 25671da177e4SLinus Torvalds */ 2568e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPCB_FLAG_ACK); 25691da177e4SLinus Torvalds TCP_SKB_CB(skb)->when = tcp_time_stamp; 2570dfb4b9dcSDavid S. Miller return tcp_transmit_skb(sk, skb, 0, GFP_ATOMIC); 25711da177e4SLinus Torvalds } 25721da177e4SLinus Torvalds 25731da177e4SLinus Torvalds int tcp_write_wakeup(struct sock *sk) 25741da177e4SLinus Torvalds { 25751da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 25761da177e4SLinus Torvalds struct sk_buff *skb; 25771da177e4SLinus Torvalds 2578058dc334SIlpo Järvinen if (sk->sk_state == TCP_CLOSE) 2579058dc334SIlpo Järvinen return -1; 2580058dc334SIlpo Järvinen 2581fe067e8aSDavid S. Miller if ((skb = tcp_send_head(sk)) != NULL && 258290840defSIlpo Järvinen before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) { 25831da177e4SLinus Torvalds int err; 25841da177e4SLinus Torvalds unsigned int mss = tcp_current_mss(sk, 0); 258590840defSIlpo Järvinen unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 25861da177e4SLinus Torvalds 25871da177e4SLinus Torvalds if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq)) 25881da177e4SLinus Torvalds tp->pushed_seq = TCP_SKB_CB(skb)->end_seq; 25891da177e4SLinus Torvalds 25901da177e4SLinus Torvalds /* We are probing the opening of a window 25911da177e4SLinus Torvalds * but the window size is != 0 25921da177e4SLinus Torvalds * must have been a result SWS avoidance ( sender ) 25931da177e4SLinus Torvalds */ 25941da177e4SLinus Torvalds if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq || 25951da177e4SLinus Torvalds skb->len > mss) { 25961da177e4SLinus Torvalds seg_size = min(seg_size, mss); 25971da177e4SLinus Torvalds TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH; 2598846998aeSDavid S. Miller if (tcp_fragment(sk, skb, seg_size, mss)) 25991da177e4SLinus Torvalds return -1; 26001da177e4SLinus Torvalds } else if (!tcp_skb_pcount(skb)) 2601846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, skb, mss); 26021da177e4SLinus Torvalds 26031da177e4SLinus Torvalds TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH; 26041da177e4SLinus Torvalds TCP_SKB_CB(skb)->when = tcp_time_stamp; 2605dfb4b9dcSDavid S. Miller err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 260666f5fe62SIlpo Järvinen if (!err) 260766f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, skb); 26081da177e4SLinus Torvalds return err; 26091da177e4SLinus Torvalds } else { 261033f5f57eSIlpo Järvinen if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF)) 26114828e7f4SIlpo Järvinen tcp_xmit_probe_skb(sk, 1); 26121da177e4SLinus Torvalds return tcp_xmit_probe_skb(sk, 0); 26131da177e4SLinus Torvalds } 26141da177e4SLinus Torvalds } 26151da177e4SLinus Torvalds 26161da177e4SLinus Torvalds /* A window probe timeout has occurred. If window is not closed send 26171da177e4SLinus Torvalds * a partial packet else a zero probe. 26181da177e4SLinus Torvalds */ 26191da177e4SLinus Torvalds void tcp_send_probe0(struct sock *sk) 26201da177e4SLinus Torvalds { 2621463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 26221da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 26231da177e4SLinus Torvalds int err; 26241da177e4SLinus Torvalds 26251da177e4SLinus Torvalds err = tcp_write_wakeup(sk); 26261da177e4SLinus Torvalds 2627fe067e8aSDavid S. Miller if (tp->packets_out || !tcp_send_head(sk)) { 26281da177e4SLinus Torvalds /* Cancel probe timer, if it is not required. */ 26296687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out = 0; 2630463c84b9SArnaldo Carvalho de Melo icsk->icsk_backoff = 0; 26311da177e4SLinus Torvalds return; 26321da177e4SLinus Torvalds } 26331da177e4SLinus Torvalds 26341da177e4SLinus Torvalds if (err <= 0) { 2635463c84b9SArnaldo Carvalho de Melo if (icsk->icsk_backoff < sysctl_tcp_retries2) 2636463c84b9SArnaldo Carvalho de Melo icsk->icsk_backoff++; 26376687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out++; 2638463c84b9SArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0, 26393f421baaSArnaldo Carvalho de Melo min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX), 26403f421baaSArnaldo Carvalho de Melo TCP_RTO_MAX); 26411da177e4SLinus Torvalds } else { 26421da177e4SLinus Torvalds /* If packet was not sent due to local congestion, 26436687e988SArnaldo Carvalho de Melo * do not backoff and do not remember icsk_probes_out. 26441da177e4SLinus Torvalds * Let local senders to fight for local resources. 26451da177e4SLinus Torvalds * 26461da177e4SLinus Torvalds * Use accumulated backoff yet. 26471da177e4SLinus Torvalds */ 26486687e988SArnaldo Carvalho de Melo if (!icsk->icsk_probes_out) 26496687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out = 1; 2650463c84b9SArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0, 2651463c84b9SArnaldo Carvalho de Melo min(icsk->icsk_rto << icsk->icsk_backoff, 26523f421baaSArnaldo Carvalho de Melo TCP_RESOURCE_PROBE_INTERVAL), 26533f421baaSArnaldo Carvalho de Melo TCP_RTO_MAX); 26541da177e4SLinus Torvalds } 26551da177e4SLinus Torvalds } 26561da177e4SLinus Torvalds 2657c6aefafbSGlenn Griffin EXPORT_SYMBOL(tcp_select_initial_window); 26581da177e4SLinus Torvalds EXPORT_SYMBOL(tcp_connect); 26591da177e4SLinus Torvalds EXPORT_SYMBOL(tcp_make_synack); 26601da177e4SLinus Torvalds EXPORT_SYMBOL(tcp_simple_retransmit); 26611da177e4SLinus Torvalds EXPORT_SYMBOL(tcp_sync_mss); 26625d424d5aSJohn Heffner EXPORT_SYMBOL(tcp_mtup_init); 2663