11da177e4SLinus Torvalds /* 21da177e4SLinus Torvalds * INET An implementation of the TCP/IP protocol suite for the LINUX 31da177e4SLinus Torvalds * operating system. INET is implemented using the BSD Socket 41da177e4SLinus Torvalds * interface as the means of communication with the user level. 51da177e4SLinus Torvalds * 61da177e4SLinus Torvalds * Implementation of the Transmission Control Protocol(TCP). 71da177e4SLinus Torvalds * 81da177e4SLinus Torvalds * Version: $Id: tcp_output.c,v 1.146 2002/02/01 22:01:04 davem Exp $ 91da177e4SLinus Torvalds * 1002c30a84SJesper Juhl * Authors: Ross Biro 111da177e4SLinus Torvalds * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG> 121da177e4SLinus Torvalds * Mark Evans, <evansmp@uhura.aston.ac.uk> 131da177e4SLinus Torvalds * Corey Minyard <wf-rch!minyard@relay.EU.net> 141da177e4SLinus Torvalds * Florian La Roche, <flla@stud.uni-sb.de> 151da177e4SLinus Torvalds * Charles Hedrick, <hedrick@klinzhai.rutgers.edu> 161da177e4SLinus Torvalds * Linus Torvalds, <torvalds@cs.helsinki.fi> 171da177e4SLinus Torvalds * Alan Cox, <gw4pts@gw4pts.ampr.org> 181da177e4SLinus Torvalds * Matthew Dillon, <dillon@apollo.west.oic.com> 191da177e4SLinus Torvalds * Arnt Gulbrandsen, <agulbra@nvg.unit.no> 201da177e4SLinus Torvalds * Jorge Cwik, <jorge@laser.satlink.net> 211da177e4SLinus Torvalds */ 221da177e4SLinus Torvalds 231da177e4SLinus Torvalds /* 241da177e4SLinus Torvalds * Changes: Pedro Roque : Retransmit queue handled by TCP. 251da177e4SLinus Torvalds * : Fragmentation on mtu decrease 261da177e4SLinus Torvalds * : Segment collapse on retransmit 271da177e4SLinus Torvalds * : AF independence 281da177e4SLinus Torvalds * 291da177e4SLinus Torvalds * Linus Torvalds : send_delayed_ack 301da177e4SLinus Torvalds * David S. Miller : Charge memory using the right skb 311da177e4SLinus Torvalds * during syn/ack processing. 321da177e4SLinus Torvalds * David S. Miller : Output engine completely rewritten. 331da177e4SLinus Torvalds * Andrea Arcangeli: SYNACK carry ts_recent in tsecr. 341da177e4SLinus Torvalds * Cacophonix Gaul : draft-minshall-nagle-01 351da177e4SLinus Torvalds * J Hadi Salim : ECN support 361da177e4SLinus Torvalds * 371da177e4SLinus Torvalds */ 381da177e4SLinus Torvalds 391da177e4SLinus Torvalds #include <net/tcp.h> 401da177e4SLinus Torvalds 411da177e4SLinus Torvalds #include <linux/compiler.h> 421da177e4SLinus Torvalds #include <linux/module.h> 431da177e4SLinus Torvalds 441da177e4SLinus Torvalds /* People can turn this off for buggy TCP's found in printers etc. */ 45ab32ea5dSBrian Haley int sysctl_tcp_retrans_collapse __read_mostly = 1; 461da177e4SLinus Torvalds 4715d99e02SRick Jones /* People can turn this on to work with those rare, broken TCPs that 4815d99e02SRick Jones * interpret the window field as a signed quantity. 4915d99e02SRick Jones */ 50ab32ea5dSBrian Haley int sysctl_tcp_workaround_signed_windows __read_mostly = 0; 5115d99e02SRick Jones 521da177e4SLinus Torvalds /* This limits the percentage of the congestion window which we 531da177e4SLinus Torvalds * will allow a single TSO frame to consume. Building TSO frames 541da177e4SLinus Torvalds * which are too large can cause TCP streams to be bursty. 551da177e4SLinus Torvalds */ 56ab32ea5dSBrian Haley int sysctl_tcp_tso_win_divisor __read_mostly = 3; 571da177e4SLinus Torvalds 58ab32ea5dSBrian Haley int sysctl_tcp_mtu_probing __read_mostly = 0; 59ab32ea5dSBrian Haley int sysctl_tcp_base_mss __read_mostly = 512; 605d424d5aSJohn Heffner 6135089bb2SDavid S. Miller /* By default, RFC2861 behavior. */ 62ab32ea5dSBrian Haley int sysctl_tcp_slow_start_after_idle __read_mostly = 1; 6335089bb2SDavid S. Miller 64*6ff03ac3SIlpo Järvinen static inline void tcp_packets_out_inc(struct sock *sk, 65*6ff03ac3SIlpo Järvinen const struct sk_buff *skb) 66*6ff03ac3SIlpo Järvinen { 67*6ff03ac3SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 68*6ff03ac3SIlpo Järvinen int orig = tp->packets_out; 69*6ff03ac3SIlpo Järvinen 70*6ff03ac3SIlpo Järvinen tp->packets_out += tcp_skb_pcount(skb); 71*6ff03ac3SIlpo Järvinen if (!orig) 72*6ff03ac3SIlpo Järvinen inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 73*6ff03ac3SIlpo Järvinen inet_csk(sk)->icsk_rto, TCP_RTO_MAX); 74*6ff03ac3SIlpo Järvinen } 75*6ff03ac3SIlpo Järvinen 769e412ba7SIlpo Järvinen static void update_send_head(struct sock *sk, struct sk_buff *skb) 771da177e4SLinus Torvalds { 789e412ba7SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 799e412ba7SIlpo Järvinen 80fe067e8aSDavid S. Miller tcp_advance_send_head(sk, skb); 811da177e4SLinus Torvalds tp->snd_nxt = TCP_SKB_CB(skb)->end_seq; 829e412ba7SIlpo Järvinen tcp_packets_out_inc(sk, skb); 831da177e4SLinus Torvalds } 841da177e4SLinus Torvalds 851da177e4SLinus Torvalds /* SND.NXT, if window was not shrunk. 861da177e4SLinus Torvalds * If window has been shrunk, what should we make? It is not clear at all. 871da177e4SLinus Torvalds * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-( 881da177e4SLinus Torvalds * Anything in between SND.UNA...SND.UNA+SND.WND also can be already 891da177e4SLinus Torvalds * invalid. OK, let's make this for now: 901da177e4SLinus Torvalds */ 919e412ba7SIlpo Järvinen static inline __u32 tcp_acceptable_seq(struct sock *sk) 921da177e4SLinus Torvalds { 939e412ba7SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 949e412ba7SIlpo Järvinen 951da177e4SLinus Torvalds if (!before(tp->snd_una+tp->snd_wnd, tp->snd_nxt)) 961da177e4SLinus Torvalds return tp->snd_nxt; 971da177e4SLinus Torvalds else 981da177e4SLinus Torvalds return tp->snd_una+tp->snd_wnd; 991da177e4SLinus Torvalds } 1001da177e4SLinus Torvalds 1011da177e4SLinus Torvalds /* Calculate mss to advertise in SYN segment. 1021da177e4SLinus Torvalds * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that: 1031da177e4SLinus Torvalds * 1041da177e4SLinus Torvalds * 1. It is independent of path mtu. 1051da177e4SLinus Torvalds * 2. Ideally, it is maximal possible segment size i.e. 65535-40. 1061da177e4SLinus Torvalds * 3. For IPv4 it is reasonable to calculate it from maximal MTU of 1071da177e4SLinus Torvalds * attached devices, because some buggy hosts are confused by 1081da177e4SLinus Torvalds * large MSS. 1091da177e4SLinus Torvalds * 4. We do not make 3, we advertise MSS, calculated from first 1101da177e4SLinus Torvalds * hop device mtu, but allow to raise it to ip_rt_min_advmss. 1111da177e4SLinus Torvalds * This may be overridden via information stored in routing table. 1121da177e4SLinus Torvalds * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible, 1131da177e4SLinus Torvalds * probably even Jumbo". 1141da177e4SLinus Torvalds */ 1151da177e4SLinus Torvalds static __u16 tcp_advertise_mss(struct sock *sk) 1161da177e4SLinus Torvalds { 1171da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 1181da177e4SLinus Torvalds struct dst_entry *dst = __sk_dst_get(sk); 1191da177e4SLinus Torvalds int mss = tp->advmss; 1201da177e4SLinus Torvalds 1211da177e4SLinus Torvalds if (dst && dst_metric(dst, RTAX_ADVMSS) < mss) { 1221da177e4SLinus Torvalds mss = dst_metric(dst, RTAX_ADVMSS); 1231da177e4SLinus Torvalds tp->advmss = mss; 1241da177e4SLinus Torvalds } 1251da177e4SLinus Torvalds 1261da177e4SLinus Torvalds return (__u16)mss; 1271da177e4SLinus Torvalds } 1281da177e4SLinus Torvalds 1291da177e4SLinus Torvalds /* RFC2861. Reset CWND after idle period longer RTO to "restart window". 1301da177e4SLinus Torvalds * This is the first part of cwnd validation mechanism. */ 131463c84b9SArnaldo Carvalho de Melo static void tcp_cwnd_restart(struct sock *sk, struct dst_entry *dst) 1321da177e4SLinus Torvalds { 133463c84b9SArnaldo Carvalho de Melo struct tcp_sock *tp = tcp_sk(sk); 1341da177e4SLinus Torvalds s32 delta = tcp_time_stamp - tp->lsndtime; 1351da177e4SLinus Torvalds u32 restart_cwnd = tcp_init_cwnd(tp, dst); 1361da177e4SLinus Torvalds u32 cwnd = tp->snd_cwnd; 1371da177e4SLinus Torvalds 1386687e988SArnaldo Carvalho de Melo tcp_ca_event(sk, CA_EVENT_CWND_RESTART); 1391da177e4SLinus Torvalds 1406687e988SArnaldo Carvalho de Melo tp->snd_ssthresh = tcp_current_ssthresh(sk); 1411da177e4SLinus Torvalds restart_cwnd = min(restart_cwnd, cwnd); 1421da177e4SLinus Torvalds 143463c84b9SArnaldo Carvalho de Melo while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd) 1441da177e4SLinus Torvalds cwnd >>= 1; 1451da177e4SLinus Torvalds tp->snd_cwnd = max(cwnd, restart_cwnd); 1461da177e4SLinus Torvalds tp->snd_cwnd_stamp = tcp_time_stamp; 1471da177e4SLinus Torvalds tp->snd_cwnd_used = 0; 1481da177e4SLinus Torvalds } 1491da177e4SLinus Torvalds 15040efc6faSStephen Hemminger static void tcp_event_data_sent(struct tcp_sock *tp, 1511da177e4SLinus Torvalds struct sk_buff *skb, struct sock *sk) 1521da177e4SLinus Torvalds { 153463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 154463c84b9SArnaldo Carvalho de Melo const u32 now = tcp_time_stamp; 1551da177e4SLinus Torvalds 15635089bb2SDavid S. Miller if (sysctl_tcp_slow_start_after_idle && 15735089bb2SDavid S. Miller (!tp->packets_out && (s32)(now - tp->lsndtime) > icsk->icsk_rto)) 158463c84b9SArnaldo Carvalho de Melo tcp_cwnd_restart(sk, __sk_dst_get(sk)); 1591da177e4SLinus Torvalds 1601da177e4SLinus Torvalds tp->lsndtime = now; 1611da177e4SLinus Torvalds 1621da177e4SLinus Torvalds /* If it is a reply for ato after last received 1631da177e4SLinus Torvalds * packet, enter pingpong mode. 1641da177e4SLinus Torvalds */ 165463c84b9SArnaldo Carvalho de Melo if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato) 166463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.pingpong = 1; 1671da177e4SLinus Torvalds } 1681da177e4SLinus Torvalds 16940efc6faSStephen Hemminger static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts) 1701da177e4SLinus Torvalds { 171463c84b9SArnaldo Carvalho de Melo tcp_dec_quickack_mode(sk, pkts); 172463c84b9SArnaldo Carvalho de Melo inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK); 1731da177e4SLinus Torvalds } 1741da177e4SLinus Torvalds 1751da177e4SLinus Torvalds /* Determine a window scaling and initial window to offer. 1761da177e4SLinus Torvalds * Based on the assumption that the given amount of space 1771da177e4SLinus Torvalds * will be offered. Store the results in the tp structure. 1781da177e4SLinus Torvalds * NOTE: for smooth operation initial space offering should 1791da177e4SLinus Torvalds * be a multiple of mss if possible. We assume here that mss >= 1. 1801da177e4SLinus Torvalds * This MUST be enforced by all callers. 1811da177e4SLinus Torvalds */ 1821da177e4SLinus Torvalds void tcp_select_initial_window(int __space, __u32 mss, 1831da177e4SLinus Torvalds __u32 *rcv_wnd, __u32 *window_clamp, 1841da177e4SLinus Torvalds int wscale_ok, __u8 *rcv_wscale) 1851da177e4SLinus Torvalds { 1861da177e4SLinus Torvalds unsigned int space = (__space < 0 ? 0 : __space); 1871da177e4SLinus Torvalds 1881da177e4SLinus Torvalds /* If no clamp set the clamp to the max possible scaled window */ 1891da177e4SLinus Torvalds if (*window_clamp == 0) 1901da177e4SLinus Torvalds (*window_clamp) = (65535 << 14); 1911da177e4SLinus Torvalds space = min(*window_clamp, space); 1921da177e4SLinus Torvalds 1931da177e4SLinus Torvalds /* Quantize space offering to a multiple of mss if possible. */ 1941da177e4SLinus Torvalds if (space > mss) 1951da177e4SLinus Torvalds space = (space / mss) * mss; 1961da177e4SLinus Torvalds 1971da177e4SLinus Torvalds /* NOTE: offering an initial window larger than 32767 19815d99e02SRick Jones * will break some buggy TCP stacks. If the admin tells us 19915d99e02SRick Jones * it is likely we could be speaking with such a buggy stack 20015d99e02SRick Jones * we will truncate our initial window offering to 32K-1 20115d99e02SRick Jones * unless the remote has sent us a window scaling option, 20215d99e02SRick Jones * which we interpret as a sign the remote TCP is not 20315d99e02SRick Jones * misinterpreting the window field as a signed quantity. 2041da177e4SLinus Torvalds */ 20515d99e02SRick Jones if (sysctl_tcp_workaround_signed_windows) 2061da177e4SLinus Torvalds (*rcv_wnd) = min(space, MAX_TCP_WINDOW); 20715d99e02SRick Jones else 20815d99e02SRick Jones (*rcv_wnd) = space; 20915d99e02SRick Jones 2101da177e4SLinus Torvalds (*rcv_wscale) = 0; 2111da177e4SLinus Torvalds if (wscale_ok) { 2121da177e4SLinus Torvalds /* Set window scaling on max possible window 2131da177e4SLinus Torvalds * See RFC1323 for an explanation of the limit to 14 2141da177e4SLinus Torvalds */ 2151da177e4SLinus Torvalds space = max_t(u32, sysctl_tcp_rmem[2], sysctl_rmem_max); 216316c1592SStephen Hemminger space = min_t(u32, space, *window_clamp); 2171da177e4SLinus Torvalds while (space > 65535 && (*rcv_wscale) < 14) { 2181da177e4SLinus Torvalds space >>= 1; 2191da177e4SLinus Torvalds (*rcv_wscale)++; 2201da177e4SLinus Torvalds } 2211da177e4SLinus Torvalds } 2221da177e4SLinus Torvalds 2231da177e4SLinus Torvalds /* Set initial window to value enough for senders, 2246b251858SDavid S. Miller * following RFC2414. Senders, not following this RFC, 2251da177e4SLinus Torvalds * will be satisfied with 2. 2261da177e4SLinus Torvalds */ 2271da177e4SLinus Torvalds if (mss > (1<<*rcv_wscale)) { 22801ff367eSDavid S. Miller int init_cwnd = 4; 22901ff367eSDavid S. Miller if (mss > 1460*3) 2301da177e4SLinus Torvalds init_cwnd = 2; 23101ff367eSDavid S. Miller else if (mss > 1460) 23201ff367eSDavid S. Miller init_cwnd = 3; 2331da177e4SLinus Torvalds if (*rcv_wnd > init_cwnd*mss) 2341da177e4SLinus Torvalds *rcv_wnd = init_cwnd*mss; 2351da177e4SLinus Torvalds } 2361da177e4SLinus Torvalds 2371da177e4SLinus Torvalds /* Set the clamp no higher than max representable value */ 2381da177e4SLinus Torvalds (*window_clamp) = min(65535U << (*rcv_wscale), *window_clamp); 2391da177e4SLinus Torvalds } 2401da177e4SLinus Torvalds 2411da177e4SLinus Torvalds /* Chose a new window to advertise, update state in tcp_sock for the 2421da177e4SLinus Torvalds * socket, and return result with RFC1323 scaling applied. The return 2431da177e4SLinus Torvalds * value can be stuffed directly into th->window for an outgoing 2441da177e4SLinus Torvalds * frame. 2451da177e4SLinus Torvalds */ 24640efc6faSStephen Hemminger static u16 tcp_select_window(struct sock *sk) 2471da177e4SLinus Torvalds { 2481da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 2491da177e4SLinus Torvalds u32 cur_win = tcp_receive_window(tp); 2501da177e4SLinus Torvalds u32 new_win = __tcp_select_window(sk); 2511da177e4SLinus Torvalds 2521da177e4SLinus Torvalds /* Never shrink the offered window */ 2531da177e4SLinus Torvalds if (new_win < cur_win) { 2541da177e4SLinus Torvalds /* Danger Will Robinson! 2551da177e4SLinus Torvalds * Don't update rcv_wup/rcv_wnd here or else 2561da177e4SLinus Torvalds * we will not be able to advertise a zero 2571da177e4SLinus Torvalds * window in time. --DaveM 2581da177e4SLinus Torvalds * 2591da177e4SLinus Torvalds * Relax Will Robinson. 2601da177e4SLinus Torvalds */ 2611da177e4SLinus Torvalds new_win = cur_win; 2621da177e4SLinus Torvalds } 2631da177e4SLinus Torvalds tp->rcv_wnd = new_win; 2641da177e4SLinus Torvalds tp->rcv_wup = tp->rcv_nxt; 2651da177e4SLinus Torvalds 2661da177e4SLinus Torvalds /* Make sure we do not exceed the maximum possible 2671da177e4SLinus Torvalds * scaled window. 2681da177e4SLinus Torvalds */ 26915d99e02SRick Jones if (!tp->rx_opt.rcv_wscale && sysctl_tcp_workaround_signed_windows) 2701da177e4SLinus Torvalds new_win = min(new_win, MAX_TCP_WINDOW); 2711da177e4SLinus Torvalds else 2721da177e4SLinus Torvalds new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale)); 2731da177e4SLinus Torvalds 2741da177e4SLinus Torvalds /* RFC1323 scaling applied */ 2751da177e4SLinus Torvalds new_win >>= tp->rx_opt.rcv_wscale; 2761da177e4SLinus Torvalds 2771da177e4SLinus Torvalds /* If we advertise zero window, disable fast path. */ 2781da177e4SLinus Torvalds if (new_win == 0) 2791da177e4SLinus Torvalds tp->pred_flags = 0; 2801da177e4SLinus Torvalds 2811da177e4SLinus Torvalds return new_win; 2821da177e4SLinus Torvalds } 2831da177e4SLinus Torvalds 284bdf1ee5dSIlpo Järvinen static inline void TCP_ECN_send_synack(struct tcp_sock *tp, 285bdf1ee5dSIlpo Järvinen struct sk_buff *skb) 286bdf1ee5dSIlpo Järvinen { 287bdf1ee5dSIlpo Järvinen TCP_SKB_CB(skb)->flags &= ~TCPCB_FLAG_CWR; 288bdf1ee5dSIlpo Järvinen if (!(tp->ecn_flags&TCP_ECN_OK)) 289bdf1ee5dSIlpo Järvinen TCP_SKB_CB(skb)->flags &= ~TCPCB_FLAG_ECE; 290bdf1ee5dSIlpo Järvinen } 291bdf1ee5dSIlpo Järvinen 292bdf1ee5dSIlpo Järvinen static inline void TCP_ECN_send_syn(struct sock *sk, struct sk_buff *skb) 293bdf1ee5dSIlpo Järvinen { 294bdf1ee5dSIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 295bdf1ee5dSIlpo Järvinen 296bdf1ee5dSIlpo Järvinen tp->ecn_flags = 0; 297bdf1ee5dSIlpo Järvinen if (sysctl_tcp_ecn) { 298bdf1ee5dSIlpo Järvinen TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_ECE|TCPCB_FLAG_CWR; 299bdf1ee5dSIlpo Järvinen tp->ecn_flags = TCP_ECN_OK; 300bdf1ee5dSIlpo Järvinen } 301bdf1ee5dSIlpo Järvinen } 302bdf1ee5dSIlpo Järvinen 303bdf1ee5dSIlpo Järvinen static __inline__ void 304bdf1ee5dSIlpo Järvinen TCP_ECN_make_synack(struct request_sock *req, struct tcphdr *th) 305bdf1ee5dSIlpo Järvinen { 306bdf1ee5dSIlpo Järvinen if (inet_rsk(req)->ecn_ok) 307bdf1ee5dSIlpo Järvinen th->ece = 1; 308bdf1ee5dSIlpo Järvinen } 309bdf1ee5dSIlpo Järvinen 310bdf1ee5dSIlpo Järvinen static inline void TCP_ECN_send(struct sock *sk, struct sk_buff *skb, 311bdf1ee5dSIlpo Järvinen int tcp_header_len) 312bdf1ee5dSIlpo Järvinen { 313bdf1ee5dSIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 314bdf1ee5dSIlpo Järvinen 315bdf1ee5dSIlpo Järvinen if (tp->ecn_flags & TCP_ECN_OK) { 316bdf1ee5dSIlpo Järvinen /* Not-retransmitted data segment: set ECT and inject CWR. */ 317bdf1ee5dSIlpo Järvinen if (skb->len != tcp_header_len && 318bdf1ee5dSIlpo Järvinen !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) { 319bdf1ee5dSIlpo Järvinen INET_ECN_xmit(sk); 320bdf1ee5dSIlpo Järvinen if (tp->ecn_flags&TCP_ECN_QUEUE_CWR) { 321bdf1ee5dSIlpo Järvinen tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR; 322bdf1ee5dSIlpo Järvinen tcp_hdr(skb)->cwr = 1; 323bdf1ee5dSIlpo Järvinen skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN; 324bdf1ee5dSIlpo Järvinen } 325bdf1ee5dSIlpo Järvinen } else { 326bdf1ee5dSIlpo Järvinen /* ACK or retransmitted segment: clear ECT|CE */ 327bdf1ee5dSIlpo Järvinen INET_ECN_dontxmit(sk); 328bdf1ee5dSIlpo Järvinen } 329bdf1ee5dSIlpo Järvinen if (tp->ecn_flags & TCP_ECN_DEMAND_CWR) 330bdf1ee5dSIlpo Järvinen tcp_hdr(skb)->ece = 1; 331bdf1ee5dSIlpo Järvinen } 332bdf1ee5dSIlpo Järvinen } 333bdf1ee5dSIlpo Järvinen 334df7a3b07SAl Viro static void tcp_build_and_update_options(__be32 *ptr, struct tcp_sock *tp, 335cfb6eeb4SYOSHIFUJI Hideaki __u32 tstamp, __u8 **md5_hash) 33640efc6faSStephen Hemminger { 33740efc6faSStephen Hemminger if (tp->rx_opt.tstamp_ok) { 338496c98dfSYOSHIFUJI Hideaki *ptr++ = htonl((TCPOPT_NOP << 24) | 33940efc6faSStephen Hemminger (TCPOPT_NOP << 16) | 34040efc6faSStephen Hemminger (TCPOPT_TIMESTAMP << 8) | 34140efc6faSStephen Hemminger TCPOLEN_TIMESTAMP); 34240efc6faSStephen Hemminger *ptr++ = htonl(tstamp); 34340efc6faSStephen Hemminger *ptr++ = htonl(tp->rx_opt.ts_recent); 34440efc6faSStephen Hemminger } 34540efc6faSStephen Hemminger if (tp->rx_opt.eff_sacks) { 34640efc6faSStephen Hemminger struct tcp_sack_block *sp = tp->rx_opt.dsack ? tp->duplicate_sack : tp->selective_acks; 34740efc6faSStephen Hemminger int this_sack; 34840efc6faSStephen Hemminger 34940efc6faSStephen Hemminger *ptr++ = htonl((TCPOPT_NOP << 24) | 35040efc6faSStephen Hemminger (TCPOPT_NOP << 16) | 35140efc6faSStephen Hemminger (TCPOPT_SACK << 8) | 35240efc6faSStephen Hemminger (TCPOLEN_SACK_BASE + (tp->rx_opt.eff_sacks * 35340efc6faSStephen Hemminger TCPOLEN_SACK_PERBLOCK))); 3542de979bdSStephen Hemminger 35540efc6faSStephen Hemminger for (this_sack = 0; this_sack < tp->rx_opt.eff_sacks; this_sack++) { 35640efc6faSStephen Hemminger *ptr++ = htonl(sp[this_sack].start_seq); 35740efc6faSStephen Hemminger *ptr++ = htonl(sp[this_sack].end_seq); 35840efc6faSStephen Hemminger } 3592de979bdSStephen Hemminger 36040efc6faSStephen Hemminger if (tp->rx_opt.dsack) { 36140efc6faSStephen Hemminger tp->rx_opt.dsack = 0; 36240efc6faSStephen Hemminger tp->rx_opt.eff_sacks--; 36340efc6faSStephen Hemminger } 36440efc6faSStephen Hemminger } 365cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 366cfb6eeb4SYOSHIFUJI Hideaki if (md5_hash) { 367cfb6eeb4SYOSHIFUJI Hideaki *ptr++ = htonl((TCPOPT_NOP << 24) | 368cfb6eeb4SYOSHIFUJI Hideaki (TCPOPT_NOP << 16) | 369cfb6eeb4SYOSHIFUJI Hideaki (TCPOPT_MD5SIG << 8) | 370cfb6eeb4SYOSHIFUJI Hideaki TCPOLEN_MD5SIG); 371cfb6eeb4SYOSHIFUJI Hideaki *md5_hash = (__u8 *)ptr; 372cfb6eeb4SYOSHIFUJI Hideaki } 373cfb6eeb4SYOSHIFUJI Hideaki #endif 37440efc6faSStephen Hemminger } 37540efc6faSStephen Hemminger 37640efc6faSStephen Hemminger /* Construct a tcp options header for a SYN or SYN_ACK packet. 37740efc6faSStephen Hemminger * If this is every changed make sure to change the definition of 37840efc6faSStephen Hemminger * MAX_SYN_SIZE to match the new maximum number of options that you 37940efc6faSStephen Hemminger * can generate. 380cfb6eeb4SYOSHIFUJI Hideaki * 381cfb6eeb4SYOSHIFUJI Hideaki * Note - that with the RFC2385 TCP option, we make room for the 382cfb6eeb4SYOSHIFUJI Hideaki * 16 byte MD5 hash. This will be filled in later, so the pointer for the 383cfb6eeb4SYOSHIFUJI Hideaki * location to be filled is passed back up. 38440efc6faSStephen Hemminger */ 385df7a3b07SAl Viro static void tcp_syn_build_options(__be32 *ptr, int mss, int ts, int sack, 38640efc6faSStephen Hemminger int offer_wscale, int wscale, __u32 tstamp, 387cfb6eeb4SYOSHIFUJI Hideaki __u32 ts_recent, __u8 **md5_hash) 38840efc6faSStephen Hemminger { 38940efc6faSStephen Hemminger /* We always get an MSS option. 39040efc6faSStephen Hemminger * The option bytes which will be seen in normal data 39140efc6faSStephen Hemminger * packets should timestamps be used, must be in the MSS 39240efc6faSStephen Hemminger * advertised. But we subtract them from tp->mss_cache so 39340efc6faSStephen Hemminger * that calculations in tcp_sendmsg are simpler etc. 39440efc6faSStephen Hemminger * So account for this fact here if necessary. If we 39540efc6faSStephen Hemminger * don't do this correctly, as a receiver we won't 39640efc6faSStephen Hemminger * recognize data packets as being full sized when we 39740efc6faSStephen Hemminger * should, and thus we won't abide by the delayed ACK 39840efc6faSStephen Hemminger * rules correctly. 39940efc6faSStephen Hemminger * SACKs don't matter, we never delay an ACK when we 40040efc6faSStephen Hemminger * have any of those going out. 40140efc6faSStephen Hemminger */ 40240efc6faSStephen Hemminger *ptr++ = htonl((TCPOPT_MSS << 24) | (TCPOLEN_MSS << 16) | mss); 40340efc6faSStephen Hemminger if (ts) { 40440efc6faSStephen Hemminger if (sack) 405496c98dfSYOSHIFUJI Hideaki *ptr++ = htonl((TCPOPT_SACK_PERM << 24) | 406496c98dfSYOSHIFUJI Hideaki (TCPOLEN_SACK_PERM << 16) | 407496c98dfSYOSHIFUJI Hideaki (TCPOPT_TIMESTAMP << 8) | 408496c98dfSYOSHIFUJI Hideaki TCPOLEN_TIMESTAMP); 40940efc6faSStephen Hemminger else 410496c98dfSYOSHIFUJI Hideaki *ptr++ = htonl((TCPOPT_NOP << 24) | 411496c98dfSYOSHIFUJI Hideaki (TCPOPT_NOP << 16) | 412496c98dfSYOSHIFUJI Hideaki (TCPOPT_TIMESTAMP << 8) | 413496c98dfSYOSHIFUJI Hideaki TCPOLEN_TIMESTAMP); 41440efc6faSStephen Hemminger *ptr++ = htonl(tstamp); /* TSVAL */ 41540efc6faSStephen Hemminger *ptr++ = htonl(ts_recent); /* TSECR */ 41640efc6faSStephen Hemminger } else if (sack) 417496c98dfSYOSHIFUJI Hideaki *ptr++ = htonl((TCPOPT_NOP << 24) | 418496c98dfSYOSHIFUJI Hideaki (TCPOPT_NOP << 16) | 419496c98dfSYOSHIFUJI Hideaki (TCPOPT_SACK_PERM << 8) | 420496c98dfSYOSHIFUJI Hideaki TCPOLEN_SACK_PERM); 42140efc6faSStephen Hemminger if (offer_wscale) 422496c98dfSYOSHIFUJI Hideaki *ptr++ = htonl((TCPOPT_NOP << 24) | 423496c98dfSYOSHIFUJI Hideaki (TCPOPT_WINDOW << 16) | 424496c98dfSYOSHIFUJI Hideaki (TCPOLEN_WINDOW << 8) | 425496c98dfSYOSHIFUJI Hideaki (wscale)); 426cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 427cfb6eeb4SYOSHIFUJI Hideaki /* 428cfb6eeb4SYOSHIFUJI Hideaki * If MD5 is enabled, then we set the option, and include the size 429cfb6eeb4SYOSHIFUJI Hideaki * (always 18). The actual MD5 hash is added just before the 430cfb6eeb4SYOSHIFUJI Hideaki * packet is sent. 431cfb6eeb4SYOSHIFUJI Hideaki */ 432cfb6eeb4SYOSHIFUJI Hideaki if (md5_hash) { 433cfb6eeb4SYOSHIFUJI Hideaki *ptr++ = htonl((TCPOPT_NOP << 24) | 434cfb6eeb4SYOSHIFUJI Hideaki (TCPOPT_NOP << 16) | 435cfb6eeb4SYOSHIFUJI Hideaki (TCPOPT_MD5SIG << 8) | 436cfb6eeb4SYOSHIFUJI Hideaki TCPOLEN_MD5SIG); 437cfb6eeb4SYOSHIFUJI Hideaki *md5_hash = (__u8 *) ptr; 438cfb6eeb4SYOSHIFUJI Hideaki } 439cfb6eeb4SYOSHIFUJI Hideaki #endif 44040efc6faSStephen Hemminger } 4411da177e4SLinus Torvalds 4421da177e4SLinus Torvalds /* This routine actually transmits TCP packets queued in by 4431da177e4SLinus Torvalds * tcp_do_sendmsg(). This is used by both the initial 4441da177e4SLinus Torvalds * transmission and possible later retransmissions. 4451da177e4SLinus Torvalds * All SKB's seen here are completely headerless. It is our 4461da177e4SLinus Torvalds * job to build the TCP header, and pass the packet down to 4471da177e4SLinus Torvalds * IP so it can do the same plus pass the packet off to the 4481da177e4SLinus Torvalds * device. 4491da177e4SLinus Torvalds * 4501da177e4SLinus Torvalds * We are working here with either a clone of the original 4511da177e4SLinus Torvalds * SKB, or a fresh unique copy made by the retransmit engine. 4521da177e4SLinus Torvalds */ 453dfb4b9dcSDavid S. Miller static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it, gfp_t gfp_mask) 4541da177e4SLinus Torvalds { 4556687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 456dfb4b9dcSDavid S. Miller struct inet_sock *inet; 457dfb4b9dcSDavid S. Miller struct tcp_sock *tp; 458dfb4b9dcSDavid S. Miller struct tcp_skb_cb *tcb; 459dfb4b9dcSDavid S. Miller int tcp_header_size; 460cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 461cfb6eeb4SYOSHIFUJI Hideaki struct tcp_md5sig_key *md5; 462cfb6eeb4SYOSHIFUJI Hideaki __u8 *md5_hash_location; 463cfb6eeb4SYOSHIFUJI Hideaki #endif 4641da177e4SLinus Torvalds struct tcphdr *th; 4651da177e4SLinus Torvalds int sysctl_flags; 4661da177e4SLinus Torvalds int err; 4671da177e4SLinus Torvalds 468dfb4b9dcSDavid S. Miller BUG_ON(!skb || !tcp_skb_pcount(skb)); 469dfb4b9dcSDavid S. Miller 470dfb4b9dcSDavid S. Miller /* If congestion control is doing timestamping, we must 471dfb4b9dcSDavid S. Miller * take such a timestamp before we potentially clone/copy. 472dfb4b9dcSDavid S. Miller */ 473164891aaSStephen Hemminger if (icsk->icsk_ca_ops->flags & TCP_CONG_RTT_STAMP) 474dfb4b9dcSDavid S. Miller __net_timestamp(skb); 475dfb4b9dcSDavid S. Miller 476dfb4b9dcSDavid S. Miller if (likely(clone_it)) { 477dfb4b9dcSDavid S. Miller if (unlikely(skb_cloned(skb))) 478dfb4b9dcSDavid S. Miller skb = pskb_copy(skb, gfp_mask); 479dfb4b9dcSDavid S. Miller else 480dfb4b9dcSDavid S. Miller skb = skb_clone(skb, gfp_mask); 481dfb4b9dcSDavid S. Miller if (unlikely(!skb)) 482dfb4b9dcSDavid S. Miller return -ENOBUFS; 483dfb4b9dcSDavid S. Miller } 484dfb4b9dcSDavid S. Miller 485dfb4b9dcSDavid S. Miller inet = inet_sk(sk); 486dfb4b9dcSDavid S. Miller tp = tcp_sk(sk); 487dfb4b9dcSDavid S. Miller tcb = TCP_SKB_CB(skb); 488dfb4b9dcSDavid S. Miller tcp_header_size = tp->tcp_header_len; 4891da177e4SLinus Torvalds 4901da177e4SLinus Torvalds #define SYSCTL_FLAG_TSTAMPS 0x1 4911da177e4SLinus Torvalds #define SYSCTL_FLAG_WSCALE 0x2 4921da177e4SLinus Torvalds #define SYSCTL_FLAG_SACK 0x4 4931da177e4SLinus Torvalds 4941da177e4SLinus Torvalds sysctl_flags = 0; 495dfb4b9dcSDavid S. Miller if (unlikely(tcb->flags & TCPCB_FLAG_SYN)) { 4961da177e4SLinus Torvalds tcp_header_size = sizeof(struct tcphdr) + TCPOLEN_MSS; 4971da177e4SLinus Torvalds if (sysctl_tcp_timestamps) { 4981da177e4SLinus Torvalds tcp_header_size += TCPOLEN_TSTAMP_ALIGNED; 4991da177e4SLinus Torvalds sysctl_flags |= SYSCTL_FLAG_TSTAMPS; 5001da177e4SLinus Torvalds } 5011da177e4SLinus Torvalds if (sysctl_tcp_window_scaling) { 5021da177e4SLinus Torvalds tcp_header_size += TCPOLEN_WSCALE_ALIGNED; 5031da177e4SLinus Torvalds sysctl_flags |= SYSCTL_FLAG_WSCALE; 5041da177e4SLinus Torvalds } 5051da177e4SLinus Torvalds if (sysctl_tcp_sack) { 5061da177e4SLinus Torvalds sysctl_flags |= SYSCTL_FLAG_SACK; 5071da177e4SLinus Torvalds if (!(sysctl_flags & SYSCTL_FLAG_TSTAMPS)) 5081da177e4SLinus Torvalds tcp_header_size += TCPOLEN_SACKPERM_ALIGNED; 5091da177e4SLinus Torvalds } 510dfb4b9dcSDavid S. Miller } else if (unlikely(tp->rx_opt.eff_sacks)) { 5111da177e4SLinus Torvalds /* A SACK is 2 pad bytes, a 2 byte header, plus 5121da177e4SLinus Torvalds * 2 32-bit sequence numbers for each SACK block. 5131da177e4SLinus Torvalds */ 5141da177e4SLinus Torvalds tcp_header_size += (TCPOLEN_SACK_BASE_ALIGNED + 515dfb4b9dcSDavid S. Miller (tp->rx_opt.eff_sacks * 516dfb4b9dcSDavid S. Miller TCPOLEN_SACK_PERBLOCK)); 5171da177e4SLinus Torvalds } 5181da177e4SLinus Torvalds 519317a76f9SStephen Hemminger if (tcp_packets_in_flight(tp) == 0) 5206687e988SArnaldo Carvalho de Melo tcp_ca_event(sk, CA_EVENT_TX_START); 5211da177e4SLinus Torvalds 522cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 523cfb6eeb4SYOSHIFUJI Hideaki /* 524cfb6eeb4SYOSHIFUJI Hideaki * Are we doing MD5 on this segment? If so - make 525cfb6eeb4SYOSHIFUJI Hideaki * room for it. 526cfb6eeb4SYOSHIFUJI Hideaki */ 527cfb6eeb4SYOSHIFUJI Hideaki md5 = tp->af_specific->md5_lookup(sk, sk); 528cfb6eeb4SYOSHIFUJI Hideaki if (md5) 529cfb6eeb4SYOSHIFUJI Hideaki tcp_header_size += TCPOLEN_MD5SIG_ALIGNED; 530cfb6eeb4SYOSHIFUJI Hideaki #endif 531cfb6eeb4SYOSHIFUJI Hideaki 532aa8223c7SArnaldo Carvalho de Melo skb_push(skb, tcp_header_size); 533aa8223c7SArnaldo Carvalho de Melo skb_reset_transport_header(skb); 534e89862f4SDavid S. Miller skb_set_owner_w(skb, sk); 5351da177e4SLinus Torvalds 5361da177e4SLinus Torvalds /* Build TCP header and checksum it. */ 537aa8223c7SArnaldo Carvalho de Melo th = tcp_hdr(skb); 5381da177e4SLinus Torvalds th->source = inet->sport; 5391da177e4SLinus Torvalds th->dest = inet->dport; 5401da177e4SLinus Torvalds th->seq = htonl(tcb->seq); 5411da177e4SLinus Torvalds th->ack_seq = htonl(tp->rcv_nxt); 542df7a3b07SAl Viro *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) | 543dfb4b9dcSDavid S. Miller tcb->flags); 544dfb4b9dcSDavid S. Miller 545dfb4b9dcSDavid S. Miller if (unlikely(tcb->flags & TCPCB_FLAG_SYN)) { 5461da177e4SLinus Torvalds /* RFC1323: The window in SYN & SYN/ACK segments 5471da177e4SLinus Torvalds * is never scaled. 5481da177e4SLinus Torvalds */ 549600ff0c2SIlpo Järvinen th->window = htons(min(tp->rcv_wnd, 65535U)); 5501da177e4SLinus Torvalds } else { 5511da177e4SLinus Torvalds th->window = htons(tcp_select_window(sk)); 5521da177e4SLinus Torvalds } 5531da177e4SLinus Torvalds th->check = 0; 5541da177e4SLinus Torvalds th->urg_ptr = 0; 5551da177e4SLinus Torvalds 556dfb4b9dcSDavid S. Miller if (unlikely(tp->urg_mode && 557dfb4b9dcSDavid S. Miller between(tp->snd_up, tcb->seq+1, tcb->seq+0xFFFF))) { 5581da177e4SLinus Torvalds th->urg_ptr = htons(tp->snd_up-tcb->seq); 5591da177e4SLinus Torvalds th->urg = 1; 5601da177e4SLinus Torvalds } 5611da177e4SLinus Torvalds 562dfb4b9dcSDavid S. Miller if (unlikely(tcb->flags & TCPCB_FLAG_SYN)) { 563df7a3b07SAl Viro tcp_syn_build_options((__be32 *)(th + 1), 5641da177e4SLinus Torvalds tcp_advertise_mss(sk), 5651da177e4SLinus Torvalds (sysctl_flags & SYSCTL_FLAG_TSTAMPS), 5661da177e4SLinus Torvalds (sysctl_flags & SYSCTL_FLAG_SACK), 5671da177e4SLinus Torvalds (sysctl_flags & SYSCTL_FLAG_WSCALE), 5681da177e4SLinus Torvalds tp->rx_opt.rcv_wscale, 5691da177e4SLinus Torvalds tcb->when, 570cfb6eeb4SYOSHIFUJI Hideaki tp->rx_opt.ts_recent, 571cfb6eeb4SYOSHIFUJI Hideaki 572cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 573cfb6eeb4SYOSHIFUJI Hideaki md5 ? &md5_hash_location : 574cfb6eeb4SYOSHIFUJI Hideaki #endif 575cfb6eeb4SYOSHIFUJI Hideaki NULL); 5761da177e4SLinus Torvalds } else { 577df7a3b07SAl Viro tcp_build_and_update_options((__be32 *)(th + 1), 578cfb6eeb4SYOSHIFUJI Hideaki tp, tcb->when, 579cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 580cfb6eeb4SYOSHIFUJI Hideaki md5 ? &md5_hash_location : 581cfb6eeb4SYOSHIFUJI Hideaki #endif 582cfb6eeb4SYOSHIFUJI Hideaki NULL); 5839e412ba7SIlpo Järvinen TCP_ECN_send(sk, skb, tcp_header_size); 5841da177e4SLinus Torvalds } 585dfb4b9dcSDavid S. Miller 586cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 587cfb6eeb4SYOSHIFUJI Hideaki /* Calculate the MD5 hash, as we have all we need now */ 588cfb6eeb4SYOSHIFUJI Hideaki if (md5) { 589cfb6eeb4SYOSHIFUJI Hideaki tp->af_specific->calc_md5_hash(md5_hash_location, 590cfb6eeb4SYOSHIFUJI Hideaki md5, 591cfb6eeb4SYOSHIFUJI Hideaki sk, NULL, NULL, 592aa8223c7SArnaldo Carvalho de Melo tcp_hdr(skb), 593cfb6eeb4SYOSHIFUJI Hideaki sk->sk_protocol, 594cfb6eeb4SYOSHIFUJI Hideaki skb->len); 595cfb6eeb4SYOSHIFUJI Hideaki } 596cfb6eeb4SYOSHIFUJI Hideaki #endif 597cfb6eeb4SYOSHIFUJI Hideaki 5988292a17aSArnaldo Carvalho de Melo icsk->icsk_af_ops->send_check(sk, skb->len, skb); 5991da177e4SLinus Torvalds 600dfb4b9dcSDavid S. Miller if (likely(tcb->flags & TCPCB_FLAG_ACK)) 601fc6415bcSDavid S. Miller tcp_event_ack_sent(sk, tcp_skb_pcount(skb)); 6021da177e4SLinus Torvalds 6031da177e4SLinus Torvalds if (skb->len != tcp_header_size) 6041da177e4SLinus Torvalds tcp_event_data_sent(tp, skb, sk); 6051da177e4SLinus Torvalds 606bd37a088SWei Yongjun if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq) 6071da177e4SLinus Torvalds TCP_INC_STATS(TCP_MIB_OUTSEGS); 6081da177e4SLinus Torvalds 609e89862f4SDavid S. Miller err = icsk->icsk_af_ops->queue_xmit(skb, 0); 61083de47cdSHua Zhong if (likely(err <= 0)) 6111da177e4SLinus Torvalds return err; 6121da177e4SLinus Torvalds 6133cfe3baaSIlpo Järvinen tcp_enter_cwr(sk, 1); 6141da177e4SLinus Torvalds 615b9df3cb8SGerrit Renker return net_xmit_eval(err); 616dfb4b9dcSDavid S. Miller 6171da177e4SLinus Torvalds #undef SYSCTL_FLAG_TSTAMPS 6181da177e4SLinus Torvalds #undef SYSCTL_FLAG_WSCALE 6191da177e4SLinus Torvalds #undef SYSCTL_FLAG_SACK 6201da177e4SLinus Torvalds } 6211da177e4SLinus Torvalds 6221da177e4SLinus Torvalds 6231da177e4SLinus Torvalds /* This routine just queue's the buffer 6241da177e4SLinus Torvalds * 6251da177e4SLinus Torvalds * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames, 6261da177e4SLinus Torvalds * otherwise socket can stall. 6271da177e4SLinus Torvalds */ 6281da177e4SLinus Torvalds static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb) 6291da177e4SLinus Torvalds { 6301da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 6311da177e4SLinus Torvalds 6321da177e4SLinus Torvalds /* Advance write_seq and place onto the write_queue. */ 6331da177e4SLinus Torvalds tp->write_seq = TCP_SKB_CB(skb)->end_seq; 6341da177e4SLinus Torvalds skb_header_release(skb); 635fe067e8aSDavid S. Miller tcp_add_write_queue_tail(sk, skb); 6361da177e4SLinus Torvalds sk_charge_skb(sk, skb); 6371da177e4SLinus Torvalds } 6381da177e4SLinus Torvalds 639846998aeSDavid S. Miller static void tcp_set_skb_tso_segs(struct sock *sk, struct sk_buff *skb, unsigned int mss_now) 640f6302d1dSDavid S. Miller { 641bcd76111SHerbert Xu if (skb->len <= mss_now || !sk_can_gso(sk)) { 642f6302d1dSDavid S. Miller /* Avoid the costly divide in the normal 643f6302d1dSDavid S. Miller * non-TSO case. 644f6302d1dSDavid S. Miller */ 6457967168cSHerbert Xu skb_shinfo(skb)->gso_segs = 1; 6467967168cSHerbert Xu skb_shinfo(skb)->gso_size = 0; 6477967168cSHerbert Xu skb_shinfo(skb)->gso_type = 0; 648f6302d1dSDavid S. Miller } else { 649f6302d1dSDavid S. Miller unsigned int factor; 650f6302d1dSDavid S. Miller 651846998aeSDavid S. Miller factor = skb->len + (mss_now - 1); 652846998aeSDavid S. Miller factor /= mss_now; 6537967168cSHerbert Xu skb_shinfo(skb)->gso_segs = factor; 6547967168cSHerbert Xu skb_shinfo(skb)->gso_size = mss_now; 655bcd76111SHerbert Xu skb_shinfo(skb)->gso_type = sk->sk_gso_type; 6561da177e4SLinus Torvalds } 6571da177e4SLinus Torvalds } 6581da177e4SLinus Torvalds 6591da177e4SLinus Torvalds /* Function to create two new TCP segments. Shrinks the given segment 6601da177e4SLinus Torvalds * to the specified size and appends a new segment with the rest of the 6611da177e4SLinus Torvalds * packet to the list. This won't be called frequently, I hope. 6621da177e4SLinus Torvalds * Remember, these are still headerless SKBs at this point. 6631da177e4SLinus Torvalds */ 6646475be16SDavid S. Miller int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len, unsigned int mss_now) 6651da177e4SLinus Torvalds { 6661da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 6671da177e4SLinus Torvalds struct sk_buff *buff; 6686475be16SDavid S. Miller int nsize, old_factor; 669b60b49eaSHerbert Xu int nlen; 6701da177e4SLinus Torvalds u16 flags; 6711da177e4SLinus Torvalds 672b2cc99f0SHerbert Xu BUG_ON(len > skb->len); 6736a438bbeSStephen Hemminger 6746a438bbeSStephen Hemminger clear_all_retrans_hints(tp); 6751da177e4SLinus Torvalds nsize = skb_headlen(skb) - len; 6761da177e4SLinus Torvalds if (nsize < 0) 6771da177e4SLinus Torvalds nsize = 0; 6781da177e4SLinus Torvalds 6791da177e4SLinus Torvalds if (skb_cloned(skb) && 6801da177e4SLinus Torvalds skb_is_nonlinear(skb) && 6811da177e4SLinus Torvalds pskb_expand_head(skb, 0, 0, GFP_ATOMIC)) 6821da177e4SLinus Torvalds return -ENOMEM; 6831da177e4SLinus Torvalds 6841da177e4SLinus Torvalds /* Get a new skb... force flag on. */ 6851da177e4SLinus Torvalds buff = sk_stream_alloc_skb(sk, nsize, GFP_ATOMIC); 6861da177e4SLinus Torvalds if (buff == NULL) 6871da177e4SLinus Torvalds return -ENOMEM; /* We'll just try again later. */ 688ef5cb973SHerbert Xu 689b60b49eaSHerbert Xu sk_charge_skb(sk, buff); 690b60b49eaSHerbert Xu nlen = skb->len - len - nsize; 691b60b49eaSHerbert Xu buff->truesize += nlen; 692b60b49eaSHerbert Xu skb->truesize -= nlen; 6931da177e4SLinus Torvalds 6941da177e4SLinus Torvalds /* Correct the sequence numbers. */ 6951da177e4SLinus Torvalds TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; 6961da177e4SLinus Torvalds TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; 6971da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; 6981da177e4SLinus Torvalds 6991da177e4SLinus Torvalds /* PSH and FIN should only be set in the second packet. */ 7001da177e4SLinus Torvalds flags = TCP_SKB_CB(skb)->flags; 7011da177e4SLinus Torvalds TCP_SKB_CB(skb)->flags = flags & ~(TCPCB_FLAG_FIN|TCPCB_FLAG_PSH); 7021da177e4SLinus Torvalds TCP_SKB_CB(buff)->flags = flags; 703e14c3cafSHerbert Xu TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked; 7041da177e4SLinus Torvalds TCP_SKB_CB(skb)->sacked &= ~TCPCB_AT_TAIL; 7051da177e4SLinus Torvalds 70684fa7933SPatrick McHardy if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) { 7071da177e4SLinus Torvalds /* Copy and checksum data tail into the new buffer. */ 7081da177e4SLinus Torvalds buff->csum = csum_partial_copy_nocheck(skb->data + len, skb_put(buff, nsize), 7091da177e4SLinus Torvalds nsize, 0); 7101da177e4SLinus Torvalds 7111da177e4SLinus Torvalds skb_trim(skb, len); 7121da177e4SLinus Torvalds 7131da177e4SLinus Torvalds skb->csum = csum_block_sub(skb->csum, buff->csum, len); 7141da177e4SLinus Torvalds } else { 71584fa7933SPatrick McHardy skb->ip_summed = CHECKSUM_PARTIAL; 7161da177e4SLinus Torvalds skb_split(skb, buff, len); 7171da177e4SLinus Torvalds } 7181da177e4SLinus Torvalds 7191da177e4SLinus Torvalds buff->ip_summed = skb->ip_summed; 7201da177e4SLinus Torvalds 7211da177e4SLinus Torvalds /* Looks stupid, but our code really uses when of 7221da177e4SLinus Torvalds * skbs, which it never sent before. --ANK 7231da177e4SLinus Torvalds */ 7241da177e4SLinus Torvalds TCP_SKB_CB(buff)->when = TCP_SKB_CB(skb)->when; 725a61bbcf2SPatrick McHardy buff->tstamp = skb->tstamp; 7261da177e4SLinus Torvalds 7276475be16SDavid S. Miller old_factor = tcp_skb_pcount(skb); 7286475be16SDavid S. Miller 7291da177e4SLinus Torvalds /* Fix up tso_factor for both original and new SKB. */ 730846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, skb, mss_now); 731846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, buff, mss_now); 7321da177e4SLinus Torvalds 7336475be16SDavid S. Miller /* If this packet has been sent out already, we must 7346475be16SDavid S. Miller * adjust the various packet counters. 7356475be16SDavid S. Miller */ 736cf0b450cSHerbert Xu if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) { 7376475be16SDavid S. Miller int diff = old_factor - tcp_skb_pcount(skb) - 7386475be16SDavid S. Miller tcp_skb_pcount(buff); 7391da177e4SLinus Torvalds 7406475be16SDavid S. Miller tp->packets_out -= diff; 741e14c3cafSHerbert Xu 742e14c3cafSHerbert Xu if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 743e14c3cafSHerbert Xu tp->sacked_out -= diff; 744e14c3cafSHerbert Xu if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) 745e14c3cafSHerbert Xu tp->retrans_out -= diff; 746e14c3cafSHerbert Xu 747b5860bbaSIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST) 7486475be16SDavid S. Miller tp->lost_out -= diff; 74983ca28beSHerbert Xu 7506475be16SDavid S. Miller if (diff > 0) { 75183ca28beSHerbert Xu /* Adjust Reno SACK estimate. */ 752e60402d0SIlpo Järvinen if (tcp_is_reno(tp)) { 753af610b4cSIlpo Järvinen tcp_dec_pcount_approx_int(&tp->sacked_out, diff); 754005903bcSIlpo Järvinen tcp_verify_left_out(tp); 75583ca28beSHerbert Xu } 75683ca28beSHerbert Xu 757af610b4cSIlpo Järvinen tcp_dec_pcount_approx_int(&tp->fackets_out, diff); 758d0410051SIlpo Järvinen /* SACK fastpath might overwrite it unless dealt with */ 759d0410051SIlpo Järvinen if (tp->fastpath_skb_hint != NULL && 760d0410051SIlpo Järvinen after(TCP_SKB_CB(tp->fastpath_skb_hint)->seq, 761d0410051SIlpo Järvinen TCP_SKB_CB(skb)->seq)) { 762af610b4cSIlpo Järvinen tcp_dec_pcount_approx_int(&tp->fastpath_cnt_hint, diff); 763d0410051SIlpo Järvinen } 7646475be16SDavid S. Miller } 7651da177e4SLinus Torvalds } 7661da177e4SLinus Torvalds 7671da177e4SLinus Torvalds /* Link BUFF into the send queue. */ 768f44b5271SDavid S. Miller skb_header_release(buff); 769fe067e8aSDavid S. Miller tcp_insert_write_queue_after(skb, buff, sk); 7701da177e4SLinus Torvalds 7711da177e4SLinus Torvalds return 0; 7721da177e4SLinus Torvalds } 7731da177e4SLinus Torvalds 7741da177e4SLinus Torvalds /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c 7751da177e4SLinus Torvalds * eventually). The difference is that pulled data not copied, but 7761da177e4SLinus Torvalds * immediately discarded. 7771da177e4SLinus Torvalds */ 778f2911969SHerbert Xu ~{PmVHI~} static void __pskb_trim_head(struct sk_buff *skb, int len) 7791da177e4SLinus Torvalds { 7801da177e4SLinus Torvalds int i, k, eat; 7811da177e4SLinus Torvalds 7821da177e4SLinus Torvalds eat = len; 7831da177e4SLinus Torvalds k = 0; 7841da177e4SLinus Torvalds for (i=0; i<skb_shinfo(skb)->nr_frags; i++) { 7851da177e4SLinus Torvalds if (skb_shinfo(skb)->frags[i].size <= eat) { 7861da177e4SLinus Torvalds put_page(skb_shinfo(skb)->frags[i].page); 7871da177e4SLinus Torvalds eat -= skb_shinfo(skb)->frags[i].size; 7881da177e4SLinus Torvalds } else { 7891da177e4SLinus Torvalds skb_shinfo(skb)->frags[k] = skb_shinfo(skb)->frags[i]; 7901da177e4SLinus Torvalds if (eat) { 7911da177e4SLinus Torvalds skb_shinfo(skb)->frags[k].page_offset += eat; 7921da177e4SLinus Torvalds skb_shinfo(skb)->frags[k].size -= eat; 7931da177e4SLinus Torvalds eat = 0; 7941da177e4SLinus Torvalds } 7951da177e4SLinus Torvalds k++; 7961da177e4SLinus Torvalds } 7971da177e4SLinus Torvalds } 7981da177e4SLinus Torvalds skb_shinfo(skb)->nr_frags = k; 7991da177e4SLinus Torvalds 80027a884dcSArnaldo Carvalho de Melo skb_reset_tail_pointer(skb); 8011da177e4SLinus Torvalds skb->data_len -= len; 8021da177e4SLinus Torvalds skb->len = skb->data_len; 8031da177e4SLinus Torvalds } 8041da177e4SLinus Torvalds 8051da177e4SLinus Torvalds int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len) 8061da177e4SLinus Torvalds { 8071da177e4SLinus Torvalds if (skb_cloned(skb) && 8081da177e4SLinus Torvalds pskb_expand_head(skb, 0, 0, GFP_ATOMIC)) 8091da177e4SLinus Torvalds return -ENOMEM; 8101da177e4SLinus Torvalds 811f2911969SHerbert Xu ~{PmVHI~} /* If len == headlen, we avoid __skb_pull to preserve alignment. */ 812f2911969SHerbert Xu ~{PmVHI~} if (unlikely(len < skb_headlen(skb))) 8131da177e4SLinus Torvalds __skb_pull(skb, len); 814f2911969SHerbert Xu ~{PmVHI~} else 815f2911969SHerbert Xu ~{PmVHI~} __pskb_trim_head(skb, len - skb_headlen(skb)); 8161da177e4SLinus Torvalds 8171da177e4SLinus Torvalds TCP_SKB_CB(skb)->seq += len; 81884fa7933SPatrick McHardy skb->ip_summed = CHECKSUM_PARTIAL; 8191da177e4SLinus Torvalds 8201da177e4SLinus Torvalds skb->truesize -= len; 8211da177e4SLinus Torvalds sk->sk_wmem_queued -= len; 8221da177e4SLinus Torvalds sk->sk_forward_alloc += len; 8231da177e4SLinus Torvalds sock_set_flag(sk, SOCK_QUEUE_SHRUNK); 8241da177e4SLinus Torvalds 8251da177e4SLinus Torvalds /* Any change of skb->len requires recalculation of tso 8261da177e4SLinus Torvalds * factor and mss. 8271da177e4SLinus Torvalds */ 8281da177e4SLinus Torvalds if (tcp_skb_pcount(skb) > 1) 829846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, skb, tcp_current_mss(sk, 1)); 8301da177e4SLinus Torvalds 8311da177e4SLinus Torvalds return 0; 8321da177e4SLinus Torvalds } 8331da177e4SLinus Torvalds 8345d424d5aSJohn Heffner /* Not accounting for SACKs here. */ 8355d424d5aSJohn Heffner int tcp_mtu_to_mss(struct sock *sk, int pmtu) 8365d424d5aSJohn Heffner { 8375d424d5aSJohn Heffner struct tcp_sock *tp = tcp_sk(sk); 8385d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 8395d424d5aSJohn Heffner int mss_now; 8405d424d5aSJohn Heffner 8415d424d5aSJohn Heffner /* Calculate base mss without TCP options: 8425d424d5aSJohn Heffner It is MMS_S - sizeof(tcphdr) of rfc1122 8435d424d5aSJohn Heffner */ 8445d424d5aSJohn Heffner mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr); 8455d424d5aSJohn Heffner 8465d424d5aSJohn Heffner /* Clamp it (mss_clamp does not include tcp options) */ 8475d424d5aSJohn Heffner if (mss_now > tp->rx_opt.mss_clamp) 8485d424d5aSJohn Heffner mss_now = tp->rx_opt.mss_clamp; 8495d424d5aSJohn Heffner 8505d424d5aSJohn Heffner /* Now subtract optional transport overhead */ 8515d424d5aSJohn Heffner mss_now -= icsk->icsk_ext_hdr_len; 8525d424d5aSJohn Heffner 8535d424d5aSJohn Heffner /* Then reserve room for full set of TCP options and 8 bytes of data */ 8545d424d5aSJohn Heffner if (mss_now < 48) 8555d424d5aSJohn Heffner mss_now = 48; 8565d424d5aSJohn Heffner 8575d424d5aSJohn Heffner /* Now subtract TCP options size, not including SACKs */ 8585d424d5aSJohn Heffner mss_now -= tp->tcp_header_len - sizeof(struct tcphdr); 8595d424d5aSJohn Heffner 8605d424d5aSJohn Heffner return mss_now; 8615d424d5aSJohn Heffner } 8625d424d5aSJohn Heffner 8635d424d5aSJohn Heffner /* Inverse of above */ 8645d424d5aSJohn Heffner int tcp_mss_to_mtu(struct sock *sk, int mss) 8655d424d5aSJohn Heffner { 8665d424d5aSJohn Heffner struct tcp_sock *tp = tcp_sk(sk); 8675d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 8685d424d5aSJohn Heffner int mtu; 8695d424d5aSJohn Heffner 8705d424d5aSJohn Heffner mtu = mss + 8715d424d5aSJohn Heffner tp->tcp_header_len + 8725d424d5aSJohn Heffner icsk->icsk_ext_hdr_len + 8735d424d5aSJohn Heffner icsk->icsk_af_ops->net_header_len; 8745d424d5aSJohn Heffner 8755d424d5aSJohn Heffner return mtu; 8765d424d5aSJohn Heffner } 8775d424d5aSJohn Heffner 8785d424d5aSJohn Heffner void tcp_mtup_init(struct sock *sk) 8795d424d5aSJohn Heffner { 8805d424d5aSJohn Heffner struct tcp_sock *tp = tcp_sk(sk); 8815d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 8825d424d5aSJohn Heffner 8835d424d5aSJohn Heffner icsk->icsk_mtup.enabled = sysctl_tcp_mtu_probing > 1; 8845d424d5aSJohn Heffner icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) + 8855d424d5aSJohn Heffner icsk->icsk_af_ops->net_header_len; 8865d424d5aSJohn Heffner icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, sysctl_tcp_base_mss); 8875d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = 0; 8885d424d5aSJohn Heffner } 8895d424d5aSJohn Heffner 8901da177e4SLinus Torvalds /* This function synchronize snd mss to current pmtu/exthdr set. 8911da177e4SLinus Torvalds 8921da177e4SLinus Torvalds tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts 8931da177e4SLinus Torvalds for TCP options, but includes only bare TCP header. 8941da177e4SLinus Torvalds 8951da177e4SLinus Torvalds tp->rx_opt.mss_clamp is mss negotiated at connection setup. 896caa20d9aSStephen Hemminger It is minimum of user_mss and mss received with SYN. 8971da177e4SLinus Torvalds It also does not include TCP options. 8981da177e4SLinus Torvalds 899d83d8461SArnaldo Carvalho de Melo inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function. 9001da177e4SLinus Torvalds 9011da177e4SLinus Torvalds tp->mss_cache is current effective sending mss, including 9021da177e4SLinus Torvalds all tcp options except for SACKs. It is evaluated, 9031da177e4SLinus Torvalds taking into account current pmtu, but never exceeds 9041da177e4SLinus Torvalds tp->rx_opt.mss_clamp. 9051da177e4SLinus Torvalds 9061da177e4SLinus Torvalds NOTE1. rfc1122 clearly states that advertised MSS 9071da177e4SLinus Torvalds DOES NOT include either tcp or ip options. 9081da177e4SLinus Torvalds 909d83d8461SArnaldo Carvalho de Melo NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache 910d83d8461SArnaldo Carvalho de Melo are READ ONLY outside this function. --ANK (980731) 9111da177e4SLinus Torvalds */ 9121da177e4SLinus Torvalds 9131da177e4SLinus Torvalds unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu) 9141da177e4SLinus Torvalds { 9151da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 916d83d8461SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 9175d424d5aSJohn Heffner int mss_now; 9181da177e4SLinus Torvalds 9195d424d5aSJohn Heffner if (icsk->icsk_mtup.search_high > pmtu) 9205d424d5aSJohn Heffner icsk->icsk_mtup.search_high = pmtu; 9211da177e4SLinus Torvalds 9225d424d5aSJohn Heffner mss_now = tcp_mtu_to_mss(sk, pmtu); 9231da177e4SLinus Torvalds 9241da177e4SLinus Torvalds /* Bound mss with half of window */ 9251da177e4SLinus Torvalds if (tp->max_window && mss_now > (tp->max_window>>1)) 9261da177e4SLinus Torvalds mss_now = max((tp->max_window>>1), 68U - tp->tcp_header_len); 9271da177e4SLinus Torvalds 9281da177e4SLinus Torvalds /* And store cached results */ 929d83d8461SArnaldo Carvalho de Melo icsk->icsk_pmtu_cookie = pmtu; 9305d424d5aSJohn Heffner if (icsk->icsk_mtup.enabled) 9315d424d5aSJohn Heffner mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low)); 932c1b4a7e6SDavid S. Miller tp->mss_cache = mss_now; 9331da177e4SLinus Torvalds 9341da177e4SLinus Torvalds return mss_now; 9351da177e4SLinus Torvalds } 9361da177e4SLinus Torvalds 9371da177e4SLinus Torvalds /* Compute the current effective MSS, taking SACKs and IP options, 9381da177e4SLinus Torvalds * and even PMTU discovery events into account. 9391da177e4SLinus Torvalds * 9401da177e4SLinus Torvalds * LARGESEND note: !urg_mode is overkill, only frames up to snd_up 9411da177e4SLinus Torvalds * cannot be large. However, taking into account rare use of URG, this 9421da177e4SLinus Torvalds * is not a big flaw. 9431da177e4SLinus Torvalds */ 944c1b4a7e6SDavid S. Miller unsigned int tcp_current_mss(struct sock *sk, int large_allowed) 9451da177e4SLinus Torvalds { 9461da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 9471da177e4SLinus Torvalds struct dst_entry *dst = __sk_dst_get(sk); 948c1b4a7e6SDavid S. Miller u32 mss_now; 949c1b4a7e6SDavid S. Miller u16 xmit_size_goal; 950c1b4a7e6SDavid S. Miller int doing_tso = 0; 9511da177e4SLinus Torvalds 952c1b4a7e6SDavid S. Miller mss_now = tp->mss_cache; 953c1b4a7e6SDavid S. Miller 954bcd76111SHerbert Xu if (large_allowed && sk_can_gso(sk) && !tp->urg_mode) 955c1b4a7e6SDavid S. Miller doing_tso = 1; 956c1b4a7e6SDavid S. Miller 9571da177e4SLinus Torvalds if (dst) { 9581da177e4SLinus Torvalds u32 mtu = dst_mtu(dst); 959d83d8461SArnaldo Carvalho de Melo if (mtu != inet_csk(sk)->icsk_pmtu_cookie) 9601da177e4SLinus Torvalds mss_now = tcp_sync_mss(sk, mtu); 9611da177e4SLinus Torvalds } 9621da177e4SLinus Torvalds 9631da177e4SLinus Torvalds if (tp->rx_opt.eff_sacks) 9641da177e4SLinus Torvalds mss_now -= (TCPOLEN_SACK_BASE_ALIGNED + 9651da177e4SLinus Torvalds (tp->rx_opt.eff_sacks * TCPOLEN_SACK_PERBLOCK)); 966c1b4a7e6SDavid S. Miller 967cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 968cfb6eeb4SYOSHIFUJI Hideaki if (tp->af_specific->md5_lookup(sk, sk)) 969cfb6eeb4SYOSHIFUJI Hideaki mss_now -= TCPOLEN_MD5SIG_ALIGNED; 970cfb6eeb4SYOSHIFUJI Hideaki #endif 971cfb6eeb4SYOSHIFUJI Hideaki 972c1b4a7e6SDavid S. Miller xmit_size_goal = mss_now; 973c1b4a7e6SDavid S. Miller 974c1b4a7e6SDavid S. Miller if (doing_tso) { 9758292a17aSArnaldo Carvalho de Melo xmit_size_goal = (65535 - 9768292a17aSArnaldo Carvalho de Melo inet_csk(sk)->icsk_af_ops->net_header_len - 977d83d8461SArnaldo Carvalho de Melo inet_csk(sk)->icsk_ext_hdr_len - 978d83d8461SArnaldo Carvalho de Melo tp->tcp_header_len); 979c1b4a7e6SDavid S. Miller 980c1b4a7e6SDavid S. Miller if (tp->max_window && 981c1b4a7e6SDavid S. Miller (xmit_size_goal > (tp->max_window >> 1))) 982c1b4a7e6SDavid S. Miller xmit_size_goal = max((tp->max_window >> 1), 983c1b4a7e6SDavid S. Miller 68U - tp->tcp_header_len); 984c1b4a7e6SDavid S. Miller 985c1b4a7e6SDavid S. Miller xmit_size_goal -= (xmit_size_goal % mss_now); 986c1b4a7e6SDavid S. Miller } 987c1b4a7e6SDavid S. Miller tp->xmit_size_goal = xmit_size_goal; 988c1b4a7e6SDavid S. Miller 9891da177e4SLinus Torvalds return mss_now; 9901da177e4SLinus Torvalds } 9911da177e4SLinus Torvalds 992a762a980SDavid S. Miller /* Congestion window validation. (RFC2861) */ 993a762a980SDavid S. Miller 9949e412ba7SIlpo Järvinen static void tcp_cwnd_validate(struct sock *sk) 995a762a980SDavid S. Miller { 9969e412ba7SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 997a762a980SDavid S. Miller __u32 packets_out = tp->packets_out; 998a762a980SDavid S. Miller 999a762a980SDavid S. Miller if (packets_out >= tp->snd_cwnd) { 1000a762a980SDavid S. Miller /* Network is feed fully. */ 1001a762a980SDavid S. Miller tp->snd_cwnd_used = 0; 1002a762a980SDavid S. Miller tp->snd_cwnd_stamp = tcp_time_stamp; 1003a762a980SDavid S. Miller } else { 1004a762a980SDavid S. Miller /* Network starves. */ 1005a762a980SDavid S. Miller if (tp->packets_out > tp->snd_cwnd_used) 1006a762a980SDavid S. Miller tp->snd_cwnd_used = tp->packets_out; 1007a762a980SDavid S. Miller 100815d33c07SDavid S. Miller if (sysctl_tcp_slow_start_after_idle && 100915d33c07SDavid S. Miller (s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto) 1010a762a980SDavid S. Miller tcp_cwnd_application_limited(sk); 1011a762a980SDavid S. Miller } 1012a762a980SDavid S. Miller } 1013a762a980SDavid S. Miller 1014c1b4a7e6SDavid S. Miller static unsigned int tcp_window_allows(struct tcp_sock *tp, struct sk_buff *skb, unsigned int mss_now, unsigned int cwnd) 1015c1b4a7e6SDavid S. Miller { 1016c1b4a7e6SDavid S. Miller u32 window, cwnd_len; 1017c1b4a7e6SDavid S. Miller 1018c1b4a7e6SDavid S. Miller window = (tp->snd_una + tp->snd_wnd - TCP_SKB_CB(skb)->seq); 1019c1b4a7e6SDavid S. Miller cwnd_len = mss_now * cwnd; 1020c1b4a7e6SDavid S. Miller return min(window, cwnd_len); 1021c1b4a7e6SDavid S. Miller } 1022c1b4a7e6SDavid S. Miller 1023c1b4a7e6SDavid S. Miller /* Can at least one segment of SKB be sent right now, according to the 1024c1b4a7e6SDavid S. Miller * congestion window rules? If so, return how many segments are allowed. 1025c1b4a7e6SDavid S. Miller */ 1026c1b4a7e6SDavid S. Miller static inline unsigned int tcp_cwnd_test(struct tcp_sock *tp, struct sk_buff *skb) 1027c1b4a7e6SDavid S. Miller { 1028c1b4a7e6SDavid S. Miller u32 in_flight, cwnd; 1029c1b4a7e6SDavid S. Miller 1030c1b4a7e6SDavid S. Miller /* Don't be strict about the congestion window for the final FIN. */ 1031104439a8SJohn Heffner if ((TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN) && 1032104439a8SJohn Heffner tcp_skb_pcount(skb) == 1) 1033c1b4a7e6SDavid S. Miller return 1; 1034c1b4a7e6SDavid S. Miller 1035c1b4a7e6SDavid S. Miller in_flight = tcp_packets_in_flight(tp); 1036c1b4a7e6SDavid S. Miller cwnd = tp->snd_cwnd; 1037c1b4a7e6SDavid S. Miller if (in_flight < cwnd) 1038c1b4a7e6SDavid S. Miller return (cwnd - in_flight); 1039c1b4a7e6SDavid S. Miller 1040c1b4a7e6SDavid S. Miller return 0; 1041c1b4a7e6SDavid S. Miller } 1042c1b4a7e6SDavid S. Miller 1043c1b4a7e6SDavid S. Miller /* This must be invoked the first time we consider transmitting 1044c1b4a7e6SDavid S. Miller * SKB onto the wire. 1045c1b4a7e6SDavid S. Miller */ 104640efc6faSStephen Hemminger static int tcp_init_tso_segs(struct sock *sk, struct sk_buff *skb, unsigned int mss_now) 1047c1b4a7e6SDavid S. Miller { 1048c1b4a7e6SDavid S. Miller int tso_segs = tcp_skb_pcount(skb); 1049c1b4a7e6SDavid S. Miller 1050846998aeSDavid S. Miller if (!tso_segs || 1051846998aeSDavid S. Miller (tso_segs > 1 && 10527967168cSHerbert Xu tcp_skb_mss(skb) != mss_now)) { 1053846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, skb, mss_now); 1054c1b4a7e6SDavid S. Miller tso_segs = tcp_skb_pcount(skb); 1055c1b4a7e6SDavid S. Miller } 1056c1b4a7e6SDavid S. Miller return tso_segs; 1057c1b4a7e6SDavid S. Miller } 1058c1b4a7e6SDavid S. Miller 1059c1b4a7e6SDavid S. Miller static inline int tcp_minshall_check(const struct tcp_sock *tp) 1060c1b4a7e6SDavid S. Miller { 1061c1b4a7e6SDavid S. Miller return after(tp->snd_sml,tp->snd_una) && 1062c1b4a7e6SDavid S. Miller !after(tp->snd_sml, tp->snd_nxt); 1063c1b4a7e6SDavid S. Miller } 1064c1b4a7e6SDavid S. Miller 1065c1b4a7e6SDavid S. Miller /* Return 0, if packet can be sent now without violation Nagle's rules: 1066c1b4a7e6SDavid S. Miller * 1. It is full sized. 1067c1b4a7e6SDavid S. Miller * 2. Or it contains FIN. (already checked by caller) 1068c1b4a7e6SDavid S. Miller * 3. Or TCP_NODELAY was set. 1069c1b4a7e6SDavid S. Miller * 4. Or TCP_CORK is not set, and all sent packets are ACKed. 1070c1b4a7e6SDavid S. Miller * With Minshall's modification: all sent small packets are ACKed. 1071c1b4a7e6SDavid S. Miller */ 1072c1b4a7e6SDavid S. Miller 1073c1b4a7e6SDavid S. Miller static inline int tcp_nagle_check(const struct tcp_sock *tp, 1074c1b4a7e6SDavid S. Miller const struct sk_buff *skb, 1075c1b4a7e6SDavid S. Miller unsigned mss_now, int nonagle) 1076c1b4a7e6SDavid S. Miller { 1077c1b4a7e6SDavid S. Miller return (skb->len < mss_now && 1078c1b4a7e6SDavid S. Miller ((nonagle&TCP_NAGLE_CORK) || 1079c1b4a7e6SDavid S. Miller (!nonagle && 1080c1b4a7e6SDavid S. Miller tp->packets_out && 1081c1b4a7e6SDavid S. Miller tcp_minshall_check(tp)))); 1082c1b4a7e6SDavid S. Miller } 1083c1b4a7e6SDavid S. Miller 1084c1b4a7e6SDavid S. Miller /* Return non-zero if the Nagle test allows this packet to be 1085c1b4a7e6SDavid S. Miller * sent now. 1086c1b4a7e6SDavid S. Miller */ 1087c1b4a7e6SDavid S. Miller static inline int tcp_nagle_test(struct tcp_sock *tp, struct sk_buff *skb, 1088c1b4a7e6SDavid S. Miller unsigned int cur_mss, int nonagle) 1089c1b4a7e6SDavid S. Miller { 1090c1b4a7e6SDavid S. Miller /* Nagle rule does not apply to frames, which sit in the middle of the 1091c1b4a7e6SDavid S. Miller * write_queue (they have no chances to get new data). 1092c1b4a7e6SDavid S. Miller * 1093c1b4a7e6SDavid S. Miller * This is implemented in the callers, where they modify the 'nonagle' 1094c1b4a7e6SDavid S. Miller * argument based upon the location of SKB in the send queue. 1095c1b4a7e6SDavid S. Miller */ 1096c1b4a7e6SDavid S. Miller if (nonagle & TCP_NAGLE_PUSH) 1097c1b4a7e6SDavid S. Miller return 1; 1098c1b4a7e6SDavid S. Miller 1099d551e454SIlpo Järvinen /* Don't use the nagle rule for urgent data (or for the final FIN). 1100d551e454SIlpo Järvinen * Nagle can be ignored during F-RTO too (see RFC4138). 1101d551e454SIlpo Järvinen */ 1102d551e454SIlpo Järvinen if (tp->urg_mode || (tp->frto_counter == 2) || 1103c1b4a7e6SDavid S. Miller (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN)) 1104c1b4a7e6SDavid S. Miller return 1; 1105c1b4a7e6SDavid S. Miller 1106c1b4a7e6SDavid S. Miller if (!tcp_nagle_check(tp, skb, cur_mss, nonagle)) 1107c1b4a7e6SDavid S. Miller return 1; 1108c1b4a7e6SDavid S. Miller 1109c1b4a7e6SDavid S. Miller return 0; 1110c1b4a7e6SDavid S. Miller } 1111c1b4a7e6SDavid S. Miller 1112c1b4a7e6SDavid S. Miller /* Does at least the first segment of SKB fit into the send window? */ 1113c1b4a7e6SDavid S. Miller static inline int tcp_snd_wnd_test(struct tcp_sock *tp, struct sk_buff *skb, unsigned int cur_mss) 1114c1b4a7e6SDavid S. Miller { 1115c1b4a7e6SDavid S. Miller u32 end_seq = TCP_SKB_CB(skb)->end_seq; 1116c1b4a7e6SDavid S. Miller 1117c1b4a7e6SDavid S. Miller if (skb->len > cur_mss) 1118c1b4a7e6SDavid S. Miller end_seq = TCP_SKB_CB(skb)->seq + cur_mss; 1119c1b4a7e6SDavid S. Miller 1120c1b4a7e6SDavid S. Miller return !after(end_seq, tp->snd_una + tp->snd_wnd); 1121c1b4a7e6SDavid S. Miller } 1122c1b4a7e6SDavid S. Miller 1123fe067e8aSDavid S. Miller /* This checks if the data bearing packet SKB (usually tcp_send_head(sk)) 1124c1b4a7e6SDavid S. Miller * should be put on the wire right now. If so, it returns the number of 1125c1b4a7e6SDavid S. Miller * packets allowed by the congestion window. 1126c1b4a7e6SDavid S. Miller */ 1127c1b4a7e6SDavid S. Miller static unsigned int tcp_snd_test(struct sock *sk, struct sk_buff *skb, 1128c1b4a7e6SDavid S. Miller unsigned int cur_mss, int nonagle) 1129c1b4a7e6SDavid S. Miller { 1130c1b4a7e6SDavid S. Miller struct tcp_sock *tp = tcp_sk(sk); 1131c1b4a7e6SDavid S. Miller unsigned int cwnd_quota; 1132c1b4a7e6SDavid S. Miller 1133846998aeSDavid S. Miller tcp_init_tso_segs(sk, skb, cur_mss); 1134c1b4a7e6SDavid S. Miller 1135c1b4a7e6SDavid S. Miller if (!tcp_nagle_test(tp, skb, cur_mss, nonagle)) 1136c1b4a7e6SDavid S. Miller return 0; 1137c1b4a7e6SDavid S. Miller 1138c1b4a7e6SDavid S. Miller cwnd_quota = tcp_cwnd_test(tp, skb); 1139c1b4a7e6SDavid S. Miller if (cwnd_quota && 1140c1b4a7e6SDavid S. Miller !tcp_snd_wnd_test(tp, skb, cur_mss)) 1141c1b4a7e6SDavid S. Miller cwnd_quota = 0; 1142c1b4a7e6SDavid S. Miller 1143c1b4a7e6SDavid S. Miller return cwnd_quota; 1144c1b4a7e6SDavid S. Miller } 1145c1b4a7e6SDavid S. Miller 11469e412ba7SIlpo Järvinen int tcp_may_send_now(struct sock *sk) 1147c1b4a7e6SDavid S. Miller { 11489e412ba7SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 1149fe067e8aSDavid S. Miller struct sk_buff *skb = tcp_send_head(sk); 1150c1b4a7e6SDavid S. Miller 1151c1b4a7e6SDavid S. Miller return (skb && 1152c1b4a7e6SDavid S. Miller tcp_snd_test(sk, skb, tcp_current_mss(sk, 1), 1153c1b4a7e6SDavid S. Miller (tcp_skb_is_last(sk, skb) ? 1154c1b4a7e6SDavid S. Miller TCP_NAGLE_PUSH : 1155c1b4a7e6SDavid S. Miller tp->nonagle))); 1156c1b4a7e6SDavid S. Miller } 1157c1b4a7e6SDavid S. Miller 1158c1b4a7e6SDavid S. Miller /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet 1159c1b4a7e6SDavid S. Miller * which is put after SKB on the list. It is very much like 1160c1b4a7e6SDavid S. Miller * tcp_fragment() except that it may make several kinds of assumptions 1161c1b4a7e6SDavid S. Miller * in order to speed up the splitting operation. In particular, we 1162c1b4a7e6SDavid S. Miller * know that all the data is in scatter-gather pages, and that the 1163c1b4a7e6SDavid S. Miller * packet has never been sent out before (and thus is not cloned). 1164c1b4a7e6SDavid S. Miller */ 1165846998aeSDavid S. Miller static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len, unsigned int mss_now) 1166c1b4a7e6SDavid S. Miller { 1167c1b4a7e6SDavid S. Miller struct sk_buff *buff; 1168c1b4a7e6SDavid S. Miller int nlen = skb->len - len; 1169c1b4a7e6SDavid S. Miller u16 flags; 1170c1b4a7e6SDavid S. Miller 1171c1b4a7e6SDavid S. Miller /* All of a TSO frame must be composed of paged data. */ 1172c8ac3774SHerbert Xu if (skb->len != skb->data_len) 1173c8ac3774SHerbert Xu return tcp_fragment(sk, skb, len, mss_now); 1174c1b4a7e6SDavid S. Miller 1175c1b4a7e6SDavid S. Miller buff = sk_stream_alloc_pskb(sk, 0, 0, GFP_ATOMIC); 1176c1b4a7e6SDavid S. Miller if (unlikely(buff == NULL)) 1177c1b4a7e6SDavid S. Miller return -ENOMEM; 1178c1b4a7e6SDavid S. Miller 1179b60b49eaSHerbert Xu sk_charge_skb(sk, buff); 1180b60b49eaSHerbert Xu buff->truesize += nlen; 1181c1b4a7e6SDavid S. Miller skb->truesize -= nlen; 1182c1b4a7e6SDavid S. Miller 1183c1b4a7e6SDavid S. Miller /* Correct the sequence numbers. */ 1184c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; 1185c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; 1186c1b4a7e6SDavid S. Miller TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; 1187c1b4a7e6SDavid S. Miller 1188c1b4a7e6SDavid S. Miller /* PSH and FIN should only be set in the second packet. */ 1189c1b4a7e6SDavid S. Miller flags = TCP_SKB_CB(skb)->flags; 1190c1b4a7e6SDavid S. Miller TCP_SKB_CB(skb)->flags = flags & ~(TCPCB_FLAG_FIN|TCPCB_FLAG_PSH); 1191c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->flags = flags; 1192c1b4a7e6SDavid S. Miller 1193c1b4a7e6SDavid S. Miller /* This packet was never sent out yet, so no SACK bits. */ 1194c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->sacked = 0; 1195c1b4a7e6SDavid S. Miller 119684fa7933SPatrick McHardy buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL; 1197c1b4a7e6SDavid S. Miller skb_split(skb, buff, len); 1198c1b4a7e6SDavid S. Miller 1199c1b4a7e6SDavid S. Miller /* Fix up tso_factor for both original and new SKB. */ 1200846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, skb, mss_now); 1201846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, buff, mss_now); 1202c1b4a7e6SDavid S. Miller 1203c1b4a7e6SDavid S. Miller /* Link BUFF into the send queue. */ 1204c1b4a7e6SDavid S. Miller skb_header_release(buff); 1205fe067e8aSDavid S. Miller tcp_insert_write_queue_after(skb, buff, sk); 1206c1b4a7e6SDavid S. Miller 1207c1b4a7e6SDavid S. Miller return 0; 1208c1b4a7e6SDavid S. Miller } 1209c1b4a7e6SDavid S. Miller 1210c1b4a7e6SDavid S. Miller /* Try to defer sending, if possible, in order to minimize the amount 1211c1b4a7e6SDavid S. Miller * of TSO splitting we do. View it as a kind of TSO Nagle test. 1212c1b4a7e6SDavid S. Miller * 1213c1b4a7e6SDavid S. Miller * This algorithm is from John Heffner. 1214c1b4a7e6SDavid S. Miller */ 12159e412ba7SIlpo Järvinen static int tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb) 1216c1b4a7e6SDavid S. Miller { 12179e412ba7SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 12186687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 1219c1b4a7e6SDavid S. Miller u32 send_win, cong_win, limit, in_flight; 1220c1b4a7e6SDavid S. Miller 1221c1b4a7e6SDavid S. Miller if (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN) 1222ae8064acSJohn Heffner goto send_now; 1223c1b4a7e6SDavid S. Miller 12246687e988SArnaldo Carvalho de Melo if (icsk->icsk_ca_state != TCP_CA_Open) 1225ae8064acSJohn Heffner goto send_now; 1226ae8064acSJohn Heffner 1227ae8064acSJohn Heffner /* Defer for less than two clock ticks. */ 1228ae8064acSJohn Heffner if (!tp->tso_deferred && ((jiffies<<1)>>1) - (tp->tso_deferred>>1) > 1) 1229ae8064acSJohn Heffner goto send_now; 1230908a75c1SDavid S. Miller 1231c1b4a7e6SDavid S. Miller in_flight = tcp_packets_in_flight(tp); 1232c1b4a7e6SDavid S. Miller 1233c1b4a7e6SDavid S. Miller BUG_ON(tcp_skb_pcount(skb) <= 1 || 1234c1b4a7e6SDavid S. Miller (tp->snd_cwnd <= in_flight)); 1235c1b4a7e6SDavid S. Miller 1236c1b4a7e6SDavid S. Miller send_win = (tp->snd_una + tp->snd_wnd) - TCP_SKB_CB(skb)->seq; 1237c1b4a7e6SDavid S. Miller 1238c1b4a7e6SDavid S. Miller /* From in_flight test above, we know that cwnd > in_flight. */ 1239c1b4a7e6SDavid S. Miller cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache; 1240c1b4a7e6SDavid S. Miller 1241c1b4a7e6SDavid S. Miller limit = min(send_win, cong_win); 1242c1b4a7e6SDavid S. Miller 1243ba244fe9SDavid S. Miller /* If a full-sized TSO skb can be sent, do it. */ 1244ba244fe9SDavid S. Miller if (limit >= 65536) 1245ae8064acSJohn Heffner goto send_now; 1246ba244fe9SDavid S. Miller 1247c1b4a7e6SDavid S. Miller if (sysctl_tcp_tso_win_divisor) { 1248c1b4a7e6SDavid S. Miller u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache); 1249c1b4a7e6SDavid S. Miller 1250c1b4a7e6SDavid S. Miller /* If at least some fraction of a window is available, 1251c1b4a7e6SDavid S. Miller * just use it. 1252c1b4a7e6SDavid S. Miller */ 1253c1b4a7e6SDavid S. Miller chunk /= sysctl_tcp_tso_win_divisor; 1254c1b4a7e6SDavid S. Miller if (limit >= chunk) 1255ae8064acSJohn Heffner goto send_now; 1256c1b4a7e6SDavid S. Miller } else { 1257c1b4a7e6SDavid S. Miller /* Different approach, try not to defer past a single 1258c1b4a7e6SDavid S. Miller * ACK. Receiver should ACK every other full sized 1259c1b4a7e6SDavid S. Miller * frame, so if we have space for more than 3 frames 1260c1b4a7e6SDavid S. Miller * then send now. 1261c1b4a7e6SDavid S. Miller */ 1262c1b4a7e6SDavid S. Miller if (limit > tcp_max_burst(tp) * tp->mss_cache) 1263ae8064acSJohn Heffner goto send_now; 1264c1b4a7e6SDavid S. Miller } 1265c1b4a7e6SDavid S. Miller 1266c1b4a7e6SDavid S. Miller /* Ok, it looks like it is advisable to defer. */ 1267ae8064acSJohn Heffner tp->tso_deferred = 1 | (jiffies<<1); 1268ae8064acSJohn Heffner 1269c1b4a7e6SDavid S. Miller return 1; 1270ae8064acSJohn Heffner 1271ae8064acSJohn Heffner send_now: 1272ae8064acSJohn Heffner tp->tso_deferred = 0; 1273ae8064acSJohn Heffner return 0; 1274c1b4a7e6SDavid S. Miller } 1275c1b4a7e6SDavid S. Miller 12765d424d5aSJohn Heffner /* Create a new MTU probe if we are ready. 12775d424d5aSJohn Heffner * Returns 0 if we should wait to probe (no cwnd available), 12785d424d5aSJohn Heffner * 1 if a probe was sent, 12795d424d5aSJohn Heffner * -1 otherwise */ 12805d424d5aSJohn Heffner static int tcp_mtu_probe(struct sock *sk) 12815d424d5aSJohn Heffner { 12825d424d5aSJohn Heffner struct tcp_sock *tp = tcp_sk(sk); 12835d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 12845d424d5aSJohn Heffner struct sk_buff *skb, *nskb, *next; 12855d424d5aSJohn Heffner int len; 12865d424d5aSJohn Heffner int probe_size; 12875d424d5aSJohn Heffner unsigned int pif; 12885d424d5aSJohn Heffner int copy; 12895d424d5aSJohn Heffner int mss_now; 12905d424d5aSJohn Heffner 12915d424d5aSJohn Heffner /* Not currently probing/verifying, 12925d424d5aSJohn Heffner * not in recovery, 12935d424d5aSJohn Heffner * have enough cwnd, and 12945d424d5aSJohn Heffner * not SACKing (the variable headers throw things off) */ 12955d424d5aSJohn Heffner if (!icsk->icsk_mtup.enabled || 12965d424d5aSJohn Heffner icsk->icsk_mtup.probe_size || 12975d424d5aSJohn Heffner inet_csk(sk)->icsk_ca_state != TCP_CA_Open || 12985d424d5aSJohn Heffner tp->snd_cwnd < 11 || 12995d424d5aSJohn Heffner tp->rx_opt.eff_sacks) 13005d424d5aSJohn Heffner return -1; 13015d424d5aSJohn Heffner 13025d424d5aSJohn Heffner /* Very simple search strategy: just double the MSS. */ 13035d424d5aSJohn Heffner mss_now = tcp_current_mss(sk, 0); 13045d424d5aSJohn Heffner probe_size = 2*tp->mss_cache; 13055d424d5aSJohn Heffner if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high)) { 13065d424d5aSJohn Heffner /* TODO: set timer for probe_converge_event */ 13075d424d5aSJohn Heffner return -1; 13085d424d5aSJohn Heffner } 13095d424d5aSJohn Heffner 13105d424d5aSJohn Heffner /* Have enough data in the send queue to probe? */ 13115d424d5aSJohn Heffner len = 0; 1312fe067e8aSDavid S. Miller if ((skb = tcp_send_head(sk)) == NULL) 13135d424d5aSJohn Heffner return -1; 13145d424d5aSJohn Heffner while ((len += skb->len) < probe_size && !tcp_skb_is_last(sk, skb)) 1315fe067e8aSDavid S. Miller skb = tcp_write_queue_next(sk, skb); 13165d424d5aSJohn Heffner if (len < probe_size) 13175d424d5aSJohn Heffner return -1; 13185d424d5aSJohn Heffner 13195d424d5aSJohn Heffner /* Receive window check. */ 13205d424d5aSJohn Heffner if (after(TCP_SKB_CB(skb)->seq + probe_size, tp->snd_una + tp->snd_wnd)) { 13215d424d5aSJohn Heffner if (tp->snd_wnd < probe_size) 13225d424d5aSJohn Heffner return -1; 13235d424d5aSJohn Heffner else 13245d424d5aSJohn Heffner return 0; 13255d424d5aSJohn Heffner } 13265d424d5aSJohn Heffner 13275d424d5aSJohn Heffner /* Do we need to wait to drain cwnd? */ 13285d424d5aSJohn Heffner pif = tcp_packets_in_flight(tp); 13295d424d5aSJohn Heffner if (pif + 2 > tp->snd_cwnd) { 13305d424d5aSJohn Heffner /* With no packets in flight, don't stall. */ 13315d424d5aSJohn Heffner if (pif == 0) 13325d424d5aSJohn Heffner return -1; 13335d424d5aSJohn Heffner else 13345d424d5aSJohn Heffner return 0; 13355d424d5aSJohn Heffner } 13365d424d5aSJohn Heffner 13375d424d5aSJohn Heffner /* We're allowed to probe. Build it now. */ 13385d424d5aSJohn Heffner if ((nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC)) == NULL) 13395d424d5aSJohn Heffner return -1; 13405d424d5aSJohn Heffner sk_charge_skb(sk, nskb); 13415d424d5aSJohn Heffner 1342fe067e8aSDavid S. Miller skb = tcp_send_head(sk); 1343fe067e8aSDavid S. Miller tcp_insert_write_queue_before(nskb, skb, sk); 1344fe067e8aSDavid S. Miller tcp_advance_send_head(sk, skb); 13455d424d5aSJohn Heffner 13465d424d5aSJohn Heffner TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq; 13475d424d5aSJohn Heffner TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size; 13485d424d5aSJohn Heffner TCP_SKB_CB(nskb)->flags = TCPCB_FLAG_ACK; 13495d424d5aSJohn Heffner TCP_SKB_CB(nskb)->sacked = 0; 13505d424d5aSJohn Heffner nskb->csum = 0; 135184fa7933SPatrick McHardy nskb->ip_summed = skb->ip_summed; 13525d424d5aSJohn Heffner 13535d424d5aSJohn Heffner len = 0; 13545d424d5aSJohn Heffner while (len < probe_size) { 1355fe067e8aSDavid S. Miller next = tcp_write_queue_next(sk, skb); 13565d424d5aSJohn Heffner 13575d424d5aSJohn Heffner copy = min_t(int, skb->len, probe_size - len); 13585d424d5aSJohn Heffner if (nskb->ip_summed) 13595d424d5aSJohn Heffner skb_copy_bits(skb, 0, skb_put(nskb, copy), copy); 13605d424d5aSJohn Heffner else 13615d424d5aSJohn Heffner nskb->csum = skb_copy_and_csum_bits(skb, 0, 13625d424d5aSJohn Heffner skb_put(nskb, copy), copy, nskb->csum); 13635d424d5aSJohn Heffner 13645d424d5aSJohn Heffner if (skb->len <= copy) { 13655d424d5aSJohn Heffner /* We've eaten all the data from this skb. 13665d424d5aSJohn Heffner * Throw it away. */ 13675d424d5aSJohn Heffner TCP_SKB_CB(nskb)->flags |= TCP_SKB_CB(skb)->flags; 1368fe067e8aSDavid S. Miller tcp_unlink_write_queue(skb, sk); 13695d424d5aSJohn Heffner sk_stream_free_skb(sk, skb); 13705d424d5aSJohn Heffner } else { 13715d424d5aSJohn Heffner TCP_SKB_CB(nskb)->flags |= TCP_SKB_CB(skb)->flags & 13725d424d5aSJohn Heffner ~(TCPCB_FLAG_FIN|TCPCB_FLAG_PSH); 13735d424d5aSJohn Heffner if (!skb_shinfo(skb)->nr_frags) { 13745d424d5aSJohn Heffner skb_pull(skb, copy); 137584fa7933SPatrick McHardy if (skb->ip_summed != CHECKSUM_PARTIAL) 13765d424d5aSJohn Heffner skb->csum = csum_partial(skb->data, skb->len, 0); 13775d424d5aSJohn Heffner } else { 13785d424d5aSJohn Heffner __pskb_trim_head(skb, copy); 13795d424d5aSJohn Heffner tcp_set_skb_tso_segs(sk, skb, mss_now); 13805d424d5aSJohn Heffner } 13815d424d5aSJohn Heffner TCP_SKB_CB(skb)->seq += copy; 13825d424d5aSJohn Heffner } 13835d424d5aSJohn Heffner 13845d424d5aSJohn Heffner len += copy; 13855d424d5aSJohn Heffner skb = next; 13865d424d5aSJohn Heffner } 13875d424d5aSJohn Heffner tcp_init_tso_segs(sk, nskb, nskb->len); 13885d424d5aSJohn Heffner 13895d424d5aSJohn Heffner /* We're ready to send. If this fails, the probe will 13905d424d5aSJohn Heffner * be resegmented into mss-sized pieces by tcp_write_xmit(). */ 13915d424d5aSJohn Heffner TCP_SKB_CB(nskb)->when = tcp_time_stamp; 13925d424d5aSJohn Heffner if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) { 13935d424d5aSJohn Heffner /* Decrement cwnd here because we are sending 13945d424d5aSJohn Heffner * effectively two packets. */ 13955d424d5aSJohn Heffner tp->snd_cwnd--; 13969e412ba7SIlpo Järvinen update_send_head(sk, nskb); 13975d424d5aSJohn Heffner 13985d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len); 13990e7b1368SJohn Heffner tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq; 14000e7b1368SJohn Heffner tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq; 14015d424d5aSJohn Heffner 14025d424d5aSJohn Heffner return 1; 14035d424d5aSJohn Heffner } 14045d424d5aSJohn Heffner 14055d424d5aSJohn Heffner return -1; 14065d424d5aSJohn Heffner } 14075d424d5aSJohn Heffner 14085d424d5aSJohn Heffner 14091da177e4SLinus Torvalds /* This routine writes packets to the network. It advances the 14101da177e4SLinus Torvalds * send_head. This happens as incoming acks open up the remote 14111da177e4SLinus Torvalds * window for us. 14121da177e4SLinus Torvalds * 14131da177e4SLinus Torvalds * Returns 1, if no segments are in flight and we have queued segments, but 14141da177e4SLinus Torvalds * cannot send anything now because of SWS or another problem. 14151da177e4SLinus Torvalds */ 1416a2e2a59cSDavid S. Miller static int tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle) 14171da177e4SLinus Torvalds { 14181da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 141992df7b51SDavid S. Miller struct sk_buff *skb; 1420c1b4a7e6SDavid S. Miller unsigned int tso_segs, sent_pkts; 1421c1b4a7e6SDavid S. Miller int cwnd_quota; 14225d424d5aSJohn Heffner int result; 14231da177e4SLinus Torvalds 14241da177e4SLinus Torvalds /* If we are closed, the bytes will have to remain here. 14251da177e4SLinus Torvalds * In time closedown will finish, we empty the write queue and all 14261da177e4SLinus Torvalds * will be happy. 14271da177e4SLinus Torvalds */ 142892df7b51SDavid S. Miller if (unlikely(sk->sk_state == TCP_CLOSE)) 142992df7b51SDavid S. Miller return 0; 143092df7b51SDavid S. Miller 1431c1b4a7e6SDavid S. Miller sent_pkts = 0; 14325d424d5aSJohn Heffner 14335d424d5aSJohn Heffner /* Do MTU probing. */ 14345d424d5aSJohn Heffner if ((result = tcp_mtu_probe(sk)) == 0) { 14355d424d5aSJohn Heffner return 0; 14365d424d5aSJohn Heffner } else if (result > 0) { 14375d424d5aSJohn Heffner sent_pkts = 1; 14385d424d5aSJohn Heffner } 14395d424d5aSJohn Heffner 1440fe067e8aSDavid S. Miller while ((skb = tcp_send_head(sk))) { 1441c8ac3774SHerbert Xu unsigned int limit; 1442c8ac3774SHerbert Xu 1443b68e9f85SHerbert Xu tso_segs = tcp_init_tso_segs(sk, skb, mss_now); 1444c1b4a7e6SDavid S. Miller BUG_ON(!tso_segs); 1445c1b4a7e6SDavid S. Miller 1446b68e9f85SHerbert Xu cwnd_quota = tcp_cwnd_test(tp, skb); 1447b68e9f85SHerbert Xu if (!cwnd_quota) 1448b68e9f85SHerbert Xu break; 1449b68e9f85SHerbert Xu 1450b68e9f85SHerbert Xu if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now))) 1451b68e9f85SHerbert Xu break; 1452b68e9f85SHerbert Xu 1453c1b4a7e6SDavid S. Miller if (tso_segs == 1) { 1454aa93466bSDavid S. Miller if (unlikely(!tcp_nagle_test(tp, skb, mss_now, 1455aa93466bSDavid S. Miller (tcp_skb_is_last(sk, skb) ? 1456aa93466bSDavid S. Miller nonagle : TCP_NAGLE_PUSH)))) 1457aa93466bSDavid S. Miller break; 1458c1b4a7e6SDavid S. Miller } else { 14599e412ba7SIlpo Järvinen if (tcp_tso_should_defer(sk, skb)) 1460aa93466bSDavid S. Miller break; 1461c1b4a7e6SDavid S. Miller } 1462aa93466bSDavid S. Miller 1463c8ac3774SHerbert Xu limit = mss_now; 1464c1b4a7e6SDavid S. Miller if (tso_segs > 1) { 1465c8ac3774SHerbert Xu limit = tcp_window_allows(tp, skb, 1466c1b4a7e6SDavid S. Miller mss_now, cwnd_quota); 1467c1b4a7e6SDavid S. Miller 1468c1b4a7e6SDavid S. Miller if (skb->len < limit) { 1469c1b4a7e6SDavid S. Miller unsigned int trim = skb->len % mss_now; 1470c1b4a7e6SDavid S. Miller 1471c1b4a7e6SDavid S. Miller if (trim) 1472c1b4a7e6SDavid S. Miller limit = skb->len - trim; 1473c1b4a7e6SDavid S. Miller } 1474c1b4a7e6SDavid S. Miller } 1475c8ac3774SHerbert Xu 1476c8ac3774SHerbert Xu if (skb->len > limit && 1477c8ac3774SHerbert Xu unlikely(tso_fragment(sk, skb, limit, mss_now))) 14781da177e4SLinus Torvalds break; 14791da177e4SLinus Torvalds 14801da177e4SLinus Torvalds TCP_SKB_CB(skb)->when = tcp_time_stamp; 1481c1b4a7e6SDavid S. Miller 1482dfb4b9dcSDavid S. Miller if (unlikely(tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC))) 14831da177e4SLinus Torvalds break; 14841da177e4SLinus Torvalds 14851da177e4SLinus Torvalds /* Advance the send_head. This one is sent out. 14861da177e4SLinus Torvalds * This call will increment packets_out. 14871da177e4SLinus Torvalds */ 14889e412ba7SIlpo Järvinen update_send_head(sk, skb); 14891da177e4SLinus Torvalds 14901da177e4SLinus Torvalds tcp_minshall_update(tp, mss_now, skb); 1491aa93466bSDavid S. Miller sent_pkts++; 14921da177e4SLinus Torvalds } 14931da177e4SLinus Torvalds 1494aa93466bSDavid S. Miller if (likely(sent_pkts)) { 14959e412ba7SIlpo Järvinen tcp_cwnd_validate(sk); 14961da177e4SLinus Torvalds return 0; 14971da177e4SLinus Torvalds } 1498fe067e8aSDavid S. Miller return !tp->packets_out && tcp_send_head(sk); 14991da177e4SLinus Torvalds } 15001da177e4SLinus Torvalds 1501a762a980SDavid S. Miller /* Push out any pending frames which were held back due to 1502a762a980SDavid S. Miller * TCP_CORK or attempt at coalescing tiny packets. 1503a762a980SDavid S. Miller * The socket must be locked by the caller. 1504a762a980SDavid S. Miller */ 15059e412ba7SIlpo Järvinen void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss, 15069e412ba7SIlpo Järvinen int nonagle) 1507a762a980SDavid S. Miller { 1508fe067e8aSDavid S. Miller struct sk_buff *skb = tcp_send_head(sk); 1509a762a980SDavid S. Miller 1510a762a980SDavid S. Miller if (skb) { 151155c97f3eSDavid S. Miller if (tcp_write_xmit(sk, cur_mss, nonagle)) 15129e412ba7SIlpo Järvinen tcp_check_probe_timer(sk); 1513a762a980SDavid S. Miller } 1514a762a980SDavid S. Miller } 1515a762a980SDavid S. Miller 1516c1b4a7e6SDavid S. Miller /* Send _single_ skb sitting at the send head. This function requires 1517c1b4a7e6SDavid S. Miller * true push pending frames to setup probe timer etc. 1518c1b4a7e6SDavid S. Miller */ 1519c1b4a7e6SDavid S. Miller void tcp_push_one(struct sock *sk, unsigned int mss_now) 1520c1b4a7e6SDavid S. Miller { 1521c1b4a7e6SDavid S. Miller struct tcp_sock *tp = tcp_sk(sk); 1522fe067e8aSDavid S. Miller struct sk_buff *skb = tcp_send_head(sk); 1523c1b4a7e6SDavid S. Miller unsigned int tso_segs, cwnd_quota; 1524c1b4a7e6SDavid S. Miller 1525c1b4a7e6SDavid S. Miller BUG_ON(!skb || skb->len < mss_now); 1526c1b4a7e6SDavid S. Miller 1527846998aeSDavid S. Miller tso_segs = tcp_init_tso_segs(sk, skb, mss_now); 1528c1b4a7e6SDavid S. Miller cwnd_quota = tcp_snd_test(sk, skb, mss_now, TCP_NAGLE_PUSH); 1529c1b4a7e6SDavid S. Miller 1530c1b4a7e6SDavid S. Miller if (likely(cwnd_quota)) { 1531c8ac3774SHerbert Xu unsigned int limit; 1532c8ac3774SHerbert Xu 1533c1b4a7e6SDavid S. Miller BUG_ON(!tso_segs); 1534c1b4a7e6SDavid S. Miller 1535c8ac3774SHerbert Xu limit = mss_now; 1536c1b4a7e6SDavid S. Miller if (tso_segs > 1) { 1537c8ac3774SHerbert Xu limit = tcp_window_allows(tp, skb, 1538c1b4a7e6SDavid S. Miller mss_now, cwnd_quota); 1539c1b4a7e6SDavid S. Miller 1540c1b4a7e6SDavid S. Miller if (skb->len < limit) { 1541c1b4a7e6SDavid S. Miller unsigned int trim = skb->len % mss_now; 1542c1b4a7e6SDavid S. Miller 1543c1b4a7e6SDavid S. Miller if (trim) 1544c1b4a7e6SDavid S. Miller limit = skb->len - trim; 1545c1b4a7e6SDavid S. Miller } 1546c1b4a7e6SDavid S. Miller } 1547c8ac3774SHerbert Xu 1548c8ac3774SHerbert Xu if (skb->len > limit && 1549c8ac3774SHerbert Xu unlikely(tso_fragment(sk, skb, limit, mss_now))) 1550c1b4a7e6SDavid S. Miller return; 1551c1b4a7e6SDavid S. Miller 1552c1b4a7e6SDavid S. Miller /* Send it out now. */ 1553c1b4a7e6SDavid S. Miller TCP_SKB_CB(skb)->when = tcp_time_stamp; 1554c1b4a7e6SDavid S. Miller 1555dfb4b9dcSDavid S. Miller if (likely(!tcp_transmit_skb(sk, skb, 1, sk->sk_allocation))) { 15569e412ba7SIlpo Järvinen update_send_head(sk, skb); 15579e412ba7SIlpo Järvinen tcp_cwnd_validate(sk); 1558c1b4a7e6SDavid S. Miller return; 1559c1b4a7e6SDavid S. Miller } 1560c1b4a7e6SDavid S. Miller } 1561c1b4a7e6SDavid S. Miller } 1562c1b4a7e6SDavid S. Miller 15631da177e4SLinus Torvalds /* This function returns the amount that we can raise the 15641da177e4SLinus Torvalds * usable window based on the following constraints 15651da177e4SLinus Torvalds * 15661da177e4SLinus Torvalds * 1. The window can never be shrunk once it is offered (RFC 793) 15671da177e4SLinus Torvalds * 2. We limit memory per socket 15681da177e4SLinus Torvalds * 15691da177e4SLinus Torvalds * RFC 1122: 15701da177e4SLinus Torvalds * "the suggested [SWS] avoidance algorithm for the receiver is to keep 15711da177e4SLinus Torvalds * RECV.NEXT + RCV.WIN fixed until: 15721da177e4SLinus Torvalds * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)" 15731da177e4SLinus Torvalds * 15741da177e4SLinus Torvalds * i.e. don't raise the right edge of the window until you can raise 15751da177e4SLinus Torvalds * it at least MSS bytes. 15761da177e4SLinus Torvalds * 15771da177e4SLinus Torvalds * Unfortunately, the recommended algorithm breaks header prediction, 15781da177e4SLinus Torvalds * since header prediction assumes th->window stays fixed. 15791da177e4SLinus Torvalds * 15801da177e4SLinus Torvalds * Strictly speaking, keeping th->window fixed violates the receiver 15811da177e4SLinus Torvalds * side SWS prevention criteria. The problem is that under this rule 15821da177e4SLinus Torvalds * a stream of single byte packets will cause the right side of the 15831da177e4SLinus Torvalds * window to always advance by a single byte. 15841da177e4SLinus Torvalds * 15851da177e4SLinus Torvalds * Of course, if the sender implements sender side SWS prevention 15861da177e4SLinus Torvalds * then this will not be a problem. 15871da177e4SLinus Torvalds * 15881da177e4SLinus Torvalds * BSD seems to make the following compromise: 15891da177e4SLinus Torvalds * 15901da177e4SLinus Torvalds * If the free space is less than the 1/4 of the maximum 15911da177e4SLinus Torvalds * space available and the free space is less than 1/2 mss, 15921da177e4SLinus Torvalds * then set the window to 0. 15931da177e4SLinus Torvalds * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ] 15941da177e4SLinus Torvalds * Otherwise, just prevent the window from shrinking 15951da177e4SLinus Torvalds * and from being larger than the largest representable value. 15961da177e4SLinus Torvalds * 15971da177e4SLinus Torvalds * This prevents incremental opening of the window in the regime 15981da177e4SLinus Torvalds * where TCP is limited by the speed of the reader side taking 15991da177e4SLinus Torvalds * data out of the TCP receive queue. It does nothing about 16001da177e4SLinus Torvalds * those cases where the window is constrained on the sender side 16011da177e4SLinus Torvalds * because the pipeline is full. 16021da177e4SLinus Torvalds * 16031da177e4SLinus Torvalds * BSD also seems to "accidentally" limit itself to windows that are a 16041da177e4SLinus Torvalds * multiple of MSS, at least until the free space gets quite small. 16051da177e4SLinus Torvalds * This would appear to be a side effect of the mbuf implementation. 16061da177e4SLinus Torvalds * Combining these two algorithms results in the observed behavior 16071da177e4SLinus Torvalds * of having a fixed window size at almost all times. 16081da177e4SLinus Torvalds * 16091da177e4SLinus Torvalds * Below we obtain similar behavior by forcing the offered window to 16101da177e4SLinus Torvalds * a multiple of the mss when it is feasible to do so. 16111da177e4SLinus Torvalds * 16121da177e4SLinus Torvalds * Note, we don't "adjust" for TIMESTAMP or SACK option bytes. 16131da177e4SLinus Torvalds * Regular options like TIMESTAMP are taken into account. 16141da177e4SLinus Torvalds */ 16151da177e4SLinus Torvalds u32 __tcp_select_window(struct sock *sk) 16161da177e4SLinus Torvalds { 1617463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 16181da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 1619caa20d9aSStephen Hemminger /* MSS for the peer's data. Previous versions used mss_clamp 16201da177e4SLinus Torvalds * here. I don't know if the value based on our guesses 16211da177e4SLinus Torvalds * of peer's MSS is better for the performance. It's more correct 16221da177e4SLinus Torvalds * but may be worse for the performance because of rcv_mss 16231da177e4SLinus Torvalds * fluctuations. --SAW 1998/11/1 16241da177e4SLinus Torvalds */ 1625463c84b9SArnaldo Carvalho de Melo int mss = icsk->icsk_ack.rcv_mss; 16261da177e4SLinus Torvalds int free_space = tcp_space(sk); 16271da177e4SLinus Torvalds int full_space = min_t(int, tp->window_clamp, tcp_full_space(sk)); 16281da177e4SLinus Torvalds int window; 16291da177e4SLinus Torvalds 16301da177e4SLinus Torvalds if (mss > full_space) 16311da177e4SLinus Torvalds mss = full_space; 16321da177e4SLinus Torvalds 16331da177e4SLinus Torvalds if (free_space < full_space/2) { 1634463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.quick = 0; 16351da177e4SLinus Torvalds 16361da177e4SLinus Torvalds if (tcp_memory_pressure) 16371da177e4SLinus Torvalds tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U*tp->advmss); 16381da177e4SLinus Torvalds 16391da177e4SLinus Torvalds if (free_space < mss) 16401da177e4SLinus Torvalds return 0; 16411da177e4SLinus Torvalds } 16421da177e4SLinus Torvalds 16431da177e4SLinus Torvalds if (free_space > tp->rcv_ssthresh) 16441da177e4SLinus Torvalds free_space = tp->rcv_ssthresh; 16451da177e4SLinus Torvalds 16461da177e4SLinus Torvalds /* Don't do rounding if we are using window scaling, since the 16471da177e4SLinus Torvalds * scaled window will not line up with the MSS boundary anyway. 16481da177e4SLinus Torvalds */ 16491da177e4SLinus Torvalds window = tp->rcv_wnd; 16501da177e4SLinus Torvalds if (tp->rx_opt.rcv_wscale) { 16511da177e4SLinus Torvalds window = free_space; 16521da177e4SLinus Torvalds 16531da177e4SLinus Torvalds /* Advertise enough space so that it won't get scaled away. 16541da177e4SLinus Torvalds * Import case: prevent zero window announcement if 16551da177e4SLinus Torvalds * 1<<rcv_wscale > mss. 16561da177e4SLinus Torvalds */ 16571da177e4SLinus Torvalds if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window) 16581da177e4SLinus Torvalds window = (((window >> tp->rx_opt.rcv_wscale) + 1) 16591da177e4SLinus Torvalds << tp->rx_opt.rcv_wscale); 16601da177e4SLinus Torvalds } else { 16611da177e4SLinus Torvalds /* Get the largest window that is a nice multiple of mss. 16621da177e4SLinus Torvalds * Window clamp already applied above. 16631da177e4SLinus Torvalds * If our current window offering is within 1 mss of the 16641da177e4SLinus Torvalds * free space we just keep it. This prevents the divide 16651da177e4SLinus Torvalds * and multiply from happening most of the time. 16661da177e4SLinus Torvalds * We also don't do any window rounding when the free space 16671da177e4SLinus Torvalds * is too small. 16681da177e4SLinus Torvalds */ 16691da177e4SLinus Torvalds if (window <= free_space - mss || window > free_space) 16701da177e4SLinus Torvalds window = (free_space/mss)*mss; 167184565070SJohn Heffner else if (mss == full_space && 167284565070SJohn Heffner free_space > window + full_space/2) 167384565070SJohn Heffner window = free_space; 16741da177e4SLinus Torvalds } 16751da177e4SLinus Torvalds 16761da177e4SLinus Torvalds return window; 16771da177e4SLinus Torvalds } 16781da177e4SLinus Torvalds 16791da177e4SLinus Torvalds /* Attempt to collapse two adjacent SKB's during retransmission. */ 16801da177e4SLinus Torvalds static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *skb, int mss_now) 16811da177e4SLinus Torvalds { 16821da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 1683fe067e8aSDavid S. Miller struct sk_buff *next_skb = tcp_write_queue_next(sk, skb); 16841da177e4SLinus Torvalds 16851da177e4SLinus Torvalds /* The first test we must make is that neither of these two 16861da177e4SLinus Torvalds * SKB's are still referenced by someone else. 16871da177e4SLinus Torvalds */ 16881da177e4SLinus Torvalds if (!skb_cloned(skb) && !skb_cloned(next_skb)) { 16891da177e4SLinus Torvalds int skb_size = skb->len, next_skb_size = next_skb->len; 16901da177e4SLinus Torvalds u16 flags = TCP_SKB_CB(skb)->flags; 16911da177e4SLinus Torvalds 16921da177e4SLinus Torvalds /* Also punt if next skb has been SACK'd. */ 16931da177e4SLinus Torvalds if (TCP_SKB_CB(next_skb)->sacked & TCPCB_SACKED_ACKED) 16941da177e4SLinus Torvalds return; 16951da177e4SLinus Torvalds 16961da177e4SLinus Torvalds /* Next skb is out of window. */ 16971da177e4SLinus Torvalds if (after(TCP_SKB_CB(next_skb)->end_seq, tp->snd_una+tp->snd_wnd)) 16981da177e4SLinus Torvalds return; 16991da177e4SLinus Torvalds 17001da177e4SLinus Torvalds /* Punt if not enough space exists in the first SKB for 17011da177e4SLinus Torvalds * the data in the second, or the total combined payload 17021da177e4SLinus Torvalds * would exceed the MSS. 17031da177e4SLinus Torvalds */ 17041da177e4SLinus Torvalds if ((next_skb_size > skb_tailroom(skb)) || 17051da177e4SLinus Torvalds ((skb_size + next_skb_size) > mss_now)) 17061da177e4SLinus Torvalds return; 17071da177e4SLinus Torvalds 17081da177e4SLinus Torvalds BUG_ON(tcp_skb_pcount(skb) != 1 || 17091da177e4SLinus Torvalds tcp_skb_pcount(next_skb) != 1); 17101da177e4SLinus Torvalds 17116a438bbeSStephen Hemminger /* changing transmit queue under us so clear hints */ 17126a438bbeSStephen Hemminger clear_all_retrans_hints(tp); 17136a438bbeSStephen Hemminger 17141da177e4SLinus Torvalds /* Ok. We will be able to collapse the packet. */ 1715fe067e8aSDavid S. Miller tcp_unlink_write_queue(next_skb, sk); 17161da177e4SLinus Torvalds 17171a4e2d09SArnaldo Carvalho de Melo skb_copy_from_linear_data(next_skb, 17181a4e2d09SArnaldo Carvalho de Melo skb_put(skb, next_skb_size), 17191a4e2d09SArnaldo Carvalho de Melo next_skb_size); 17201da177e4SLinus Torvalds 172152d570aaSJarek Poplawski if (next_skb->ip_summed == CHECKSUM_PARTIAL) 172252d570aaSJarek Poplawski skb->ip_summed = CHECKSUM_PARTIAL; 17231da177e4SLinus Torvalds 172484fa7933SPatrick McHardy if (skb->ip_summed != CHECKSUM_PARTIAL) 17251da177e4SLinus Torvalds skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size); 17261da177e4SLinus Torvalds 17271da177e4SLinus Torvalds /* Update sequence range on original skb. */ 17281da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq; 17291da177e4SLinus Torvalds 17301da177e4SLinus Torvalds /* Merge over control information. */ 17311da177e4SLinus Torvalds flags |= TCP_SKB_CB(next_skb)->flags; /* This moves PSH/FIN etc. over */ 17321da177e4SLinus Torvalds TCP_SKB_CB(skb)->flags = flags; 17331da177e4SLinus Torvalds 17341da177e4SLinus Torvalds /* All done, get rid of second SKB and account for it so 17351da177e4SLinus Torvalds * packet counting does not break. 17361da177e4SLinus Torvalds */ 17371da177e4SLinus Torvalds TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked&(TCPCB_EVER_RETRANS|TCPCB_AT_TAIL); 17381da177e4SLinus Torvalds if (TCP_SKB_CB(next_skb)->sacked&TCPCB_SACKED_RETRANS) 17391da177e4SLinus Torvalds tp->retrans_out -= tcp_skb_pcount(next_skb); 1740b5860bbaSIlpo Järvinen if (TCP_SKB_CB(next_skb)->sacked&TCPCB_LOST) 17411da177e4SLinus Torvalds tp->lost_out -= tcp_skb_pcount(next_skb); 17421da177e4SLinus Torvalds /* Reno case is special. Sigh... */ 1743e60402d0SIlpo Järvinen if (tcp_is_reno(tp) && tp->sacked_out) 17441da177e4SLinus Torvalds tcp_dec_pcount_approx(&tp->sacked_out, next_skb); 17451da177e4SLinus Torvalds 17461da177e4SLinus Torvalds /* Not quite right: it can be > snd.fack, but 17471da177e4SLinus Torvalds * it is better to underestimate fackets. 17481da177e4SLinus Torvalds */ 17491da177e4SLinus Torvalds tcp_dec_pcount_approx(&tp->fackets_out, next_skb); 1750e9144bd8SIlpo Järvinen tp->packets_out -= tcp_skb_pcount(next_skb); 17511da177e4SLinus Torvalds sk_stream_free_skb(sk, next_skb); 17521da177e4SLinus Torvalds } 17531da177e4SLinus Torvalds } 17541da177e4SLinus Torvalds 17551da177e4SLinus Torvalds /* Do a simple retransmit without using the backoff mechanisms in 17561da177e4SLinus Torvalds * tcp_timer. This is used for path mtu discovery. 17571da177e4SLinus Torvalds * The socket is already locked here. 17581da177e4SLinus Torvalds */ 17591da177e4SLinus Torvalds void tcp_simple_retransmit(struct sock *sk) 17601da177e4SLinus Torvalds { 17616687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 17621da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 17631da177e4SLinus Torvalds struct sk_buff *skb; 17641da177e4SLinus Torvalds unsigned int mss = tcp_current_mss(sk, 0); 17651da177e4SLinus Torvalds int lost = 0; 17661da177e4SLinus Torvalds 1767fe067e8aSDavid S. Miller tcp_for_write_queue(skb, sk) { 1768fe067e8aSDavid S. Miller if (skb == tcp_send_head(sk)) 1769fe067e8aSDavid S. Miller break; 17701da177e4SLinus Torvalds if (skb->len > mss && 17711da177e4SLinus Torvalds !(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED)) { 17721da177e4SLinus Torvalds if (TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS) { 17731da177e4SLinus Torvalds TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS; 17741da177e4SLinus Torvalds tp->retrans_out -= tcp_skb_pcount(skb); 17751da177e4SLinus Torvalds } 17761da177e4SLinus Torvalds if (!(TCP_SKB_CB(skb)->sacked&TCPCB_LOST)) { 17771da177e4SLinus Torvalds TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; 17781da177e4SLinus Torvalds tp->lost_out += tcp_skb_pcount(skb); 17791da177e4SLinus Torvalds lost = 1; 17801da177e4SLinus Torvalds } 17811da177e4SLinus Torvalds } 17821da177e4SLinus Torvalds } 17831da177e4SLinus Torvalds 17846a438bbeSStephen Hemminger clear_all_retrans_hints(tp); 17856a438bbeSStephen Hemminger 17861da177e4SLinus Torvalds if (!lost) 17871da177e4SLinus Torvalds return; 17881da177e4SLinus Torvalds 1789005903bcSIlpo Järvinen tcp_verify_left_out(tp); 17901da177e4SLinus Torvalds 17911da177e4SLinus Torvalds /* Don't muck with the congestion window here. 17921da177e4SLinus Torvalds * Reason is that we do not increase amount of _data_ 17931da177e4SLinus Torvalds * in network, but units changed and effective 17941da177e4SLinus Torvalds * cwnd/ssthresh really reduced now. 17951da177e4SLinus Torvalds */ 17966687e988SArnaldo Carvalho de Melo if (icsk->icsk_ca_state != TCP_CA_Loss) { 17971da177e4SLinus Torvalds tp->high_seq = tp->snd_nxt; 17986687e988SArnaldo Carvalho de Melo tp->snd_ssthresh = tcp_current_ssthresh(sk); 17991da177e4SLinus Torvalds tp->prior_ssthresh = 0; 18001da177e4SLinus Torvalds tp->undo_marker = 0; 18016687e988SArnaldo Carvalho de Melo tcp_set_ca_state(sk, TCP_CA_Loss); 18021da177e4SLinus Torvalds } 18031da177e4SLinus Torvalds tcp_xmit_retransmit_queue(sk); 18041da177e4SLinus Torvalds } 18051da177e4SLinus Torvalds 18061da177e4SLinus Torvalds /* This retransmits one SKB. Policy decisions and retransmit queue 18071da177e4SLinus Torvalds * state updates are done by the caller. Returns non-zero if an 18081da177e4SLinus Torvalds * error occurred which prevented the send. 18091da177e4SLinus Torvalds */ 18101da177e4SLinus Torvalds int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb) 18111da177e4SLinus Torvalds { 18121da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 18135d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 18141da177e4SLinus Torvalds unsigned int cur_mss = tcp_current_mss(sk, 0); 18151da177e4SLinus Torvalds int err; 18161da177e4SLinus Torvalds 18175d424d5aSJohn Heffner /* Inconslusive MTU probe */ 18185d424d5aSJohn Heffner if (icsk->icsk_mtup.probe_size) { 18195d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = 0; 18205d424d5aSJohn Heffner } 18215d424d5aSJohn Heffner 18221da177e4SLinus Torvalds /* Do not sent more than we queued. 1/4 is reserved for possible 1823caa20d9aSStephen Hemminger * copying overhead: fragmentation, tunneling, mangling etc. 18241da177e4SLinus Torvalds */ 18251da177e4SLinus Torvalds if (atomic_read(&sk->sk_wmem_alloc) > 18261da177e4SLinus Torvalds min(sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), sk->sk_sndbuf)) 18271da177e4SLinus Torvalds return -EAGAIN; 18281da177e4SLinus Torvalds 18291da177e4SLinus Torvalds if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) { 18301da177e4SLinus Torvalds if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) 18311da177e4SLinus Torvalds BUG(); 18321da177e4SLinus Torvalds if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq)) 18331da177e4SLinus Torvalds return -ENOMEM; 18341da177e4SLinus Torvalds } 18351da177e4SLinus Torvalds 18361da177e4SLinus Torvalds /* If receiver has shrunk his window, and skb is out of 18371da177e4SLinus Torvalds * new window, do not retransmit it. The exception is the 18381da177e4SLinus Torvalds * case, when window is shrunk to zero. In this case 18391da177e4SLinus Torvalds * our retransmit serves as a zero window probe. 18401da177e4SLinus Torvalds */ 18411da177e4SLinus Torvalds if (!before(TCP_SKB_CB(skb)->seq, tp->snd_una+tp->snd_wnd) 18421da177e4SLinus Torvalds && TCP_SKB_CB(skb)->seq != tp->snd_una) 18431da177e4SLinus Torvalds return -EAGAIN; 18441da177e4SLinus Torvalds 18451da177e4SLinus Torvalds if (skb->len > cur_mss) { 1846846998aeSDavid S. Miller if (tcp_fragment(sk, skb, cur_mss, cur_mss)) 18471da177e4SLinus Torvalds return -ENOMEM; /* We'll try again later. */ 18481da177e4SLinus Torvalds } 18491da177e4SLinus Torvalds 18501da177e4SLinus Torvalds /* Collapse two adjacent packets if worthwhile and we can. */ 18511da177e4SLinus Torvalds if (!(TCP_SKB_CB(skb)->flags & TCPCB_FLAG_SYN) && 18521da177e4SLinus Torvalds (skb->len < (cur_mss >> 1)) && 1853fe067e8aSDavid S. Miller (tcp_write_queue_next(sk, skb) != tcp_send_head(sk)) && 1854fe067e8aSDavid S. Miller (!tcp_skb_is_last(sk, skb)) && 1855fe067e8aSDavid S. Miller (skb_shinfo(skb)->nr_frags == 0 && skb_shinfo(tcp_write_queue_next(sk, skb))->nr_frags == 0) && 1856fe067e8aSDavid S. Miller (tcp_skb_pcount(skb) == 1 && tcp_skb_pcount(tcp_write_queue_next(sk, skb)) == 1) && 18571da177e4SLinus Torvalds (sysctl_tcp_retrans_collapse != 0)) 18581da177e4SLinus Torvalds tcp_retrans_try_collapse(sk, skb, cur_mss); 18591da177e4SLinus Torvalds 18608292a17aSArnaldo Carvalho de Melo if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk)) 18611da177e4SLinus Torvalds return -EHOSTUNREACH; /* Routing failure or similar. */ 18621da177e4SLinus Torvalds 18631da177e4SLinus Torvalds /* Some Solaris stacks overoptimize and ignore the FIN on a 18641da177e4SLinus Torvalds * retransmit when old data is attached. So strip it off 18651da177e4SLinus Torvalds * since it is cheap to do so and saves bytes on the network. 18661da177e4SLinus Torvalds */ 18671da177e4SLinus Torvalds if (skb->len > 0 && 18681da177e4SLinus Torvalds (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN) && 18691da177e4SLinus Torvalds tp->snd_una == (TCP_SKB_CB(skb)->end_seq - 1)) { 18701da177e4SLinus Torvalds if (!pskb_trim(skb, 0)) { 18711da177e4SLinus Torvalds TCP_SKB_CB(skb)->seq = TCP_SKB_CB(skb)->end_seq - 1; 18727967168cSHerbert Xu skb_shinfo(skb)->gso_segs = 1; 18737967168cSHerbert Xu skb_shinfo(skb)->gso_size = 0; 18747967168cSHerbert Xu skb_shinfo(skb)->gso_type = 0; 18751da177e4SLinus Torvalds skb->ip_summed = CHECKSUM_NONE; 18761da177e4SLinus Torvalds skb->csum = 0; 18771da177e4SLinus Torvalds } 18781da177e4SLinus Torvalds } 18791da177e4SLinus Torvalds 18801da177e4SLinus Torvalds /* Make a copy, if the first transmission SKB clone we made 18811da177e4SLinus Torvalds * is still in somebody's hands, else make a clone. 18821da177e4SLinus Torvalds */ 18831da177e4SLinus Torvalds TCP_SKB_CB(skb)->when = tcp_time_stamp; 18841da177e4SLinus Torvalds 1885dfb4b9dcSDavid S. Miller err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 18861da177e4SLinus Torvalds 18871da177e4SLinus Torvalds if (err == 0) { 18881da177e4SLinus Torvalds /* Update global TCP statistics. */ 18891da177e4SLinus Torvalds TCP_INC_STATS(TCP_MIB_RETRANSSEGS); 18901da177e4SLinus Torvalds 18911da177e4SLinus Torvalds tp->total_retrans++; 18921da177e4SLinus Torvalds 18931da177e4SLinus Torvalds #if FASTRETRANS_DEBUG > 0 18941da177e4SLinus Torvalds if (TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS) { 18951da177e4SLinus Torvalds if (net_ratelimit()) 18961da177e4SLinus Torvalds printk(KERN_DEBUG "retrans_out leaked.\n"); 18971da177e4SLinus Torvalds } 18981da177e4SLinus Torvalds #endif 18991da177e4SLinus Torvalds TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS; 19001da177e4SLinus Torvalds tp->retrans_out += tcp_skb_pcount(skb); 19011da177e4SLinus Torvalds 19021da177e4SLinus Torvalds /* Save stamp of the first retransmit. */ 19031da177e4SLinus Torvalds if (!tp->retrans_stamp) 19041da177e4SLinus Torvalds tp->retrans_stamp = TCP_SKB_CB(skb)->when; 19051da177e4SLinus Torvalds 19061da177e4SLinus Torvalds tp->undo_retrans++; 19071da177e4SLinus Torvalds 19081da177e4SLinus Torvalds /* snd_nxt is stored to detect loss of retransmitted segment, 19091da177e4SLinus Torvalds * see tcp_input.c tcp_sacktag_write_queue(). 19101da177e4SLinus Torvalds */ 19111da177e4SLinus Torvalds TCP_SKB_CB(skb)->ack_seq = tp->snd_nxt; 19121da177e4SLinus Torvalds } 19131da177e4SLinus Torvalds return err; 19141da177e4SLinus Torvalds } 19151da177e4SLinus Torvalds 19161da177e4SLinus Torvalds /* This gets called after a retransmit timeout, and the initially 19171da177e4SLinus Torvalds * retransmitted data is acknowledged. It tries to continue 19181da177e4SLinus Torvalds * resending the rest of the retransmit queue, until either 19191da177e4SLinus Torvalds * we've sent it all or the congestion window limit is reached. 19201da177e4SLinus Torvalds * If doing SACK, the first ACK which comes back for a timeout 19211da177e4SLinus Torvalds * based retransmit packet might feed us FACK information again. 19221da177e4SLinus Torvalds * If so, we use it to avoid unnecessarily retransmissions. 19231da177e4SLinus Torvalds */ 19241da177e4SLinus Torvalds void tcp_xmit_retransmit_queue(struct sock *sk) 19251da177e4SLinus Torvalds { 19266687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 19271da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 19281da177e4SLinus Torvalds struct sk_buff *skb; 19296a438bbeSStephen Hemminger int packet_cnt; 19306a438bbeSStephen Hemminger 19316a438bbeSStephen Hemminger if (tp->retransmit_skb_hint) { 19326a438bbeSStephen Hemminger skb = tp->retransmit_skb_hint; 19336a438bbeSStephen Hemminger packet_cnt = tp->retransmit_cnt_hint; 19346a438bbeSStephen Hemminger }else{ 1935fe067e8aSDavid S. Miller skb = tcp_write_queue_head(sk); 19366a438bbeSStephen Hemminger packet_cnt = 0; 19376a438bbeSStephen Hemminger } 19381da177e4SLinus Torvalds 19391da177e4SLinus Torvalds /* First pass: retransmit lost packets. */ 19406a438bbeSStephen Hemminger if (tp->lost_out) { 1941fe067e8aSDavid S. Miller tcp_for_write_queue_from(skb, sk) { 19421da177e4SLinus Torvalds __u8 sacked = TCP_SKB_CB(skb)->sacked; 19431da177e4SLinus Torvalds 1944fe067e8aSDavid S. Miller if (skb == tcp_send_head(sk)) 1945fe067e8aSDavid S. Miller break; 19466a438bbeSStephen Hemminger /* we could do better than to assign each time */ 19476a438bbeSStephen Hemminger tp->retransmit_skb_hint = skb; 19486a438bbeSStephen Hemminger tp->retransmit_cnt_hint = packet_cnt; 19496a438bbeSStephen Hemminger 19501da177e4SLinus Torvalds /* Assume this retransmit will generate 19511da177e4SLinus Torvalds * only one packet for congestion window 19521da177e4SLinus Torvalds * calculation purposes. This works because 19531da177e4SLinus Torvalds * tcp_retransmit_skb() will chop up the 19541da177e4SLinus Torvalds * packet to be MSS sized and all the 19551da177e4SLinus Torvalds * packet counting works out. 19561da177e4SLinus Torvalds */ 19571da177e4SLinus Torvalds if (tcp_packets_in_flight(tp) >= tp->snd_cwnd) 19581da177e4SLinus Torvalds return; 19591da177e4SLinus Torvalds 19601da177e4SLinus Torvalds if (sacked & TCPCB_LOST) { 19611da177e4SLinus Torvalds if (!(sacked&(TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))) { 19626a438bbeSStephen Hemminger if (tcp_retransmit_skb(sk, skb)) { 19636a438bbeSStephen Hemminger tp->retransmit_skb_hint = NULL; 19641da177e4SLinus Torvalds return; 19656a438bbeSStephen Hemminger } 19666687e988SArnaldo Carvalho de Melo if (icsk->icsk_ca_state != TCP_CA_Loss) 19671da177e4SLinus Torvalds NET_INC_STATS_BH(LINUX_MIB_TCPFASTRETRANS); 19681da177e4SLinus Torvalds else 19691da177e4SLinus Torvalds NET_INC_STATS_BH(LINUX_MIB_TCPSLOWSTARTRETRANS); 19701da177e4SLinus Torvalds 1971fe067e8aSDavid S. Miller if (skb == tcp_write_queue_head(sk)) 1972463c84b9SArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 19733f421baaSArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto, 19743f421baaSArnaldo Carvalho de Melo TCP_RTO_MAX); 19751da177e4SLinus Torvalds } 19761da177e4SLinus Torvalds 19776a438bbeSStephen Hemminger packet_cnt += tcp_skb_pcount(skb); 19786a438bbeSStephen Hemminger if (packet_cnt >= tp->lost_out) 19791da177e4SLinus Torvalds break; 19801da177e4SLinus Torvalds } 19811da177e4SLinus Torvalds } 19821da177e4SLinus Torvalds } 19831da177e4SLinus Torvalds 19841da177e4SLinus Torvalds /* OK, demanded retransmission is finished. */ 19851da177e4SLinus Torvalds 19861da177e4SLinus Torvalds /* Forward retransmissions are possible only during Recovery. */ 19876687e988SArnaldo Carvalho de Melo if (icsk->icsk_ca_state != TCP_CA_Recovery) 19881da177e4SLinus Torvalds return; 19891da177e4SLinus Torvalds 19901da177e4SLinus Torvalds /* No forward retransmissions in Reno are possible. */ 1991e60402d0SIlpo Järvinen if (tcp_is_reno(tp)) 19921da177e4SLinus Torvalds return; 19931da177e4SLinus Torvalds 19941da177e4SLinus Torvalds /* Yeah, we have to make difficult choice between forward transmission 19951da177e4SLinus Torvalds * and retransmission... Both ways have their merits... 19961da177e4SLinus Torvalds * 19971da177e4SLinus Torvalds * For now we do not retransmit anything, while we have some new 1998539d243fSIlpo Järvinen * segments to send. In the other cases, follow rule 3 for 1999539d243fSIlpo Järvinen * NextSeg() specified in RFC3517. 20001da177e4SLinus Torvalds */ 20011da177e4SLinus Torvalds 20029e412ba7SIlpo Järvinen if (tcp_may_send_now(sk)) 20031da177e4SLinus Torvalds return; 20041da177e4SLinus Torvalds 2005539d243fSIlpo Järvinen /* If nothing is SACKed, highest_sack in the loop won't be valid */ 2006539d243fSIlpo Järvinen if (!tp->sacked_out) 2007539d243fSIlpo Järvinen return; 2008539d243fSIlpo Järvinen 2009539d243fSIlpo Järvinen if (tp->forward_skb_hint) 20106a438bbeSStephen Hemminger skb = tp->forward_skb_hint; 2011539d243fSIlpo Järvinen else 2012fe067e8aSDavid S. Miller skb = tcp_write_queue_head(sk); 20131da177e4SLinus Torvalds 2014fe067e8aSDavid S. Miller tcp_for_write_queue_from(skb, sk) { 2015fe067e8aSDavid S. Miller if (skb == tcp_send_head(sk)) 2016fe067e8aSDavid S. Miller break; 20176a438bbeSStephen Hemminger tp->forward_skb_hint = skb; 20186a438bbeSStephen Hemminger 2019539d243fSIlpo Järvinen if (after(TCP_SKB_CB(skb)->seq, tp->highest_sack)) 20201da177e4SLinus Torvalds break; 20211da177e4SLinus Torvalds 20221da177e4SLinus Torvalds if (tcp_packets_in_flight(tp) >= tp->snd_cwnd) 20231da177e4SLinus Torvalds break; 20241da177e4SLinus Torvalds 20251da177e4SLinus Torvalds if (TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) 20261da177e4SLinus Torvalds continue; 20271da177e4SLinus Torvalds 20281da177e4SLinus Torvalds /* Ok, retransmit it. */ 20296a438bbeSStephen Hemminger if (tcp_retransmit_skb(sk, skb)) { 20306a438bbeSStephen Hemminger tp->forward_skb_hint = NULL; 20311da177e4SLinus Torvalds break; 20326a438bbeSStephen Hemminger } 20331da177e4SLinus Torvalds 2034fe067e8aSDavid S. Miller if (skb == tcp_write_queue_head(sk)) 20353f421baaSArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 20363f421baaSArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto, 20373f421baaSArnaldo Carvalho de Melo TCP_RTO_MAX); 20381da177e4SLinus Torvalds 20391da177e4SLinus Torvalds NET_INC_STATS_BH(LINUX_MIB_TCPFORWARDRETRANS); 20401da177e4SLinus Torvalds } 20411da177e4SLinus Torvalds } 20421da177e4SLinus Torvalds 20431da177e4SLinus Torvalds 20441da177e4SLinus Torvalds /* Send a fin. The caller locks the socket for us. This cannot be 20451da177e4SLinus Torvalds * allowed to fail queueing a FIN frame under any circumstances. 20461da177e4SLinus Torvalds */ 20471da177e4SLinus Torvalds void tcp_send_fin(struct sock *sk) 20481da177e4SLinus Torvalds { 20491da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 2050fe067e8aSDavid S. Miller struct sk_buff *skb = tcp_write_queue_tail(sk); 20511da177e4SLinus Torvalds int mss_now; 20521da177e4SLinus Torvalds 20531da177e4SLinus Torvalds /* Optimization, tack on the FIN if we have a queue of 20541da177e4SLinus Torvalds * unsent frames. But be careful about outgoing SACKS 20551da177e4SLinus Torvalds * and IP options. 20561da177e4SLinus Torvalds */ 20571da177e4SLinus Torvalds mss_now = tcp_current_mss(sk, 1); 20581da177e4SLinus Torvalds 2059fe067e8aSDavid S. Miller if (tcp_send_head(sk) != NULL) { 20601da177e4SLinus Torvalds TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_FIN; 20611da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq++; 20621da177e4SLinus Torvalds tp->write_seq++; 20631da177e4SLinus Torvalds } else { 20641da177e4SLinus Torvalds /* Socket is locked, keep trying until memory is available. */ 20651da177e4SLinus Torvalds for (;;) { 2066d179cd12SDavid S. Miller skb = alloc_skb_fclone(MAX_TCP_HEADER, GFP_KERNEL); 20671da177e4SLinus Torvalds if (skb) 20681da177e4SLinus Torvalds break; 20691da177e4SLinus Torvalds yield(); 20701da177e4SLinus Torvalds } 20711da177e4SLinus Torvalds 20721da177e4SLinus Torvalds /* Reserve space for headers and prepare control bits. */ 20731da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 20741da177e4SLinus Torvalds skb->csum = 0; 20751da177e4SLinus Torvalds TCP_SKB_CB(skb)->flags = (TCPCB_FLAG_ACK | TCPCB_FLAG_FIN); 20761da177e4SLinus Torvalds TCP_SKB_CB(skb)->sacked = 0; 20777967168cSHerbert Xu skb_shinfo(skb)->gso_segs = 1; 20787967168cSHerbert Xu skb_shinfo(skb)->gso_size = 0; 20797967168cSHerbert Xu skb_shinfo(skb)->gso_type = 0; 20801da177e4SLinus Torvalds 20811da177e4SLinus Torvalds /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */ 20821da177e4SLinus Torvalds TCP_SKB_CB(skb)->seq = tp->write_seq; 20831da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + 1; 20841da177e4SLinus Torvalds tcp_queue_skb(sk, skb); 20851da177e4SLinus Torvalds } 20869e412ba7SIlpo Järvinen __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_OFF); 20871da177e4SLinus Torvalds } 20881da177e4SLinus Torvalds 20891da177e4SLinus Torvalds /* We get here when a process closes a file descriptor (either due to 20901da177e4SLinus Torvalds * an explicit close() or as a byproduct of exit()'ing) and there 20911da177e4SLinus Torvalds * was unread data in the receive queue. This behavior is recommended 209265bb723cSGerrit Renker * by RFC 2525, section 2.17. -DaveM 20931da177e4SLinus Torvalds */ 2094dd0fc66fSAl Viro void tcp_send_active_reset(struct sock *sk, gfp_t priority) 20951da177e4SLinus Torvalds { 20961da177e4SLinus Torvalds struct sk_buff *skb; 20971da177e4SLinus Torvalds 20981da177e4SLinus Torvalds /* NOTE: No TCP options attached and we never retransmit this. */ 20991da177e4SLinus Torvalds skb = alloc_skb(MAX_TCP_HEADER, priority); 21001da177e4SLinus Torvalds if (!skb) { 21011da177e4SLinus Torvalds NET_INC_STATS(LINUX_MIB_TCPABORTFAILED); 21021da177e4SLinus Torvalds return; 21031da177e4SLinus Torvalds } 21041da177e4SLinus Torvalds 21051da177e4SLinus Torvalds /* Reserve space for headers and prepare control bits. */ 21061da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 21071da177e4SLinus Torvalds skb->csum = 0; 21081da177e4SLinus Torvalds TCP_SKB_CB(skb)->flags = (TCPCB_FLAG_ACK | TCPCB_FLAG_RST); 21091da177e4SLinus Torvalds TCP_SKB_CB(skb)->sacked = 0; 21107967168cSHerbert Xu skb_shinfo(skb)->gso_segs = 1; 21117967168cSHerbert Xu skb_shinfo(skb)->gso_size = 0; 21127967168cSHerbert Xu skb_shinfo(skb)->gso_type = 0; 21131da177e4SLinus Torvalds 21141da177e4SLinus Torvalds /* Send it off. */ 21159e412ba7SIlpo Järvinen TCP_SKB_CB(skb)->seq = tcp_acceptable_seq(sk); 21161da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq; 21171da177e4SLinus Torvalds TCP_SKB_CB(skb)->when = tcp_time_stamp; 2118dfb4b9dcSDavid S. Miller if (tcp_transmit_skb(sk, skb, 0, priority)) 21191da177e4SLinus Torvalds NET_INC_STATS(LINUX_MIB_TCPABORTFAILED); 21201da177e4SLinus Torvalds } 21211da177e4SLinus Torvalds 21221da177e4SLinus Torvalds /* WARNING: This routine must only be called when we have already sent 21231da177e4SLinus Torvalds * a SYN packet that crossed the incoming SYN that caused this routine 21241da177e4SLinus Torvalds * to get called. If this assumption fails then the initial rcv_wnd 21251da177e4SLinus Torvalds * and rcv_wscale values will not be correct. 21261da177e4SLinus Torvalds */ 21271da177e4SLinus Torvalds int tcp_send_synack(struct sock *sk) 21281da177e4SLinus Torvalds { 21291da177e4SLinus Torvalds struct sk_buff* skb; 21301da177e4SLinus Torvalds 2131fe067e8aSDavid S. Miller skb = tcp_write_queue_head(sk); 21321da177e4SLinus Torvalds if (skb == NULL || !(TCP_SKB_CB(skb)->flags&TCPCB_FLAG_SYN)) { 21331da177e4SLinus Torvalds printk(KERN_DEBUG "tcp_send_synack: wrong queue state\n"); 21341da177e4SLinus Torvalds return -EFAULT; 21351da177e4SLinus Torvalds } 21361da177e4SLinus Torvalds if (!(TCP_SKB_CB(skb)->flags&TCPCB_FLAG_ACK)) { 21371da177e4SLinus Torvalds if (skb_cloned(skb)) { 21381da177e4SLinus Torvalds struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC); 21391da177e4SLinus Torvalds if (nskb == NULL) 21401da177e4SLinus Torvalds return -ENOMEM; 2141fe067e8aSDavid S. Miller tcp_unlink_write_queue(skb, sk); 21421da177e4SLinus Torvalds skb_header_release(nskb); 2143fe067e8aSDavid S. Miller __tcp_add_write_queue_head(sk, nskb); 21441da177e4SLinus Torvalds sk_stream_free_skb(sk, skb); 21451da177e4SLinus Torvalds sk_charge_skb(sk, nskb); 21461da177e4SLinus Torvalds skb = nskb; 21471da177e4SLinus Torvalds } 21481da177e4SLinus Torvalds 21491da177e4SLinus Torvalds TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_ACK; 21501da177e4SLinus Torvalds TCP_ECN_send_synack(tcp_sk(sk), skb); 21511da177e4SLinus Torvalds } 21521da177e4SLinus Torvalds TCP_SKB_CB(skb)->when = tcp_time_stamp; 2153dfb4b9dcSDavid S. Miller return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 21541da177e4SLinus Torvalds } 21551da177e4SLinus Torvalds 21561da177e4SLinus Torvalds /* 21571da177e4SLinus Torvalds * Prepare a SYN-ACK. 21581da177e4SLinus Torvalds */ 21591da177e4SLinus Torvalds struct sk_buff * tcp_make_synack(struct sock *sk, struct dst_entry *dst, 216060236fddSArnaldo Carvalho de Melo struct request_sock *req) 21611da177e4SLinus Torvalds { 21622e6599cbSArnaldo Carvalho de Melo struct inet_request_sock *ireq = inet_rsk(req); 21631da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 21641da177e4SLinus Torvalds struct tcphdr *th; 21651da177e4SLinus Torvalds int tcp_header_size; 21661da177e4SLinus Torvalds struct sk_buff *skb; 2167cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 2168cfb6eeb4SYOSHIFUJI Hideaki struct tcp_md5sig_key *md5; 2169cfb6eeb4SYOSHIFUJI Hideaki __u8 *md5_hash_location; 2170cfb6eeb4SYOSHIFUJI Hideaki #endif 21711da177e4SLinus Torvalds 21721da177e4SLinus Torvalds skb = sock_wmalloc(sk, MAX_TCP_HEADER + 15, 1, GFP_ATOMIC); 21731da177e4SLinus Torvalds if (skb == NULL) 21741da177e4SLinus Torvalds return NULL; 21751da177e4SLinus Torvalds 21761da177e4SLinus Torvalds /* Reserve space for headers. */ 21771da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 21781da177e4SLinus Torvalds 21791da177e4SLinus Torvalds skb->dst = dst_clone(dst); 21801da177e4SLinus Torvalds 21811da177e4SLinus Torvalds tcp_header_size = (sizeof(struct tcphdr) + TCPOLEN_MSS + 21822e6599cbSArnaldo Carvalho de Melo (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0) + 21832e6599cbSArnaldo Carvalho de Melo (ireq->wscale_ok ? TCPOLEN_WSCALE_ALIGNED : 0) + 21841da177e4SLinus Torvalds /* SACK_PERM is in the place of NOP NOP of TS */ 21852e6599cbSArnaldo Carvalho de Melo ((ireq->sack_ok && !ireq->tstamp_ok) ? TCPOLEN_SACKPERM_ALIGNED : 0)); 2186cfb6eeb4SYOSHIFUJI Hideaki 2187cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 2188cfb6eeb4SYOSHIFUJI Hideaki /* Are we doing MD5 on this segment? If so - make room for it */ 2189cfb6eeb4SYOSHIFUJI Hideaki md5 = tcp_rsk(req)->af_specific->md5_lookup(sk, req); 2190cfb6eeb4SYOSHIFUJI Hideaki if (md5) 2191cfb6eeb4SYOSHIFUJI Hideaki tcp_header_size += TCPOLEN_MD5SIG_ALIGNED; 2192cfb6eeb4SYOSHIFUJI Hideaki #endif 2193aa8223c7SArnaldo Carvalho de Melo skb_push(skb, tcp_header_size); 2194aa8223c7SArnaldo Carvalho de Melo skb_reset_transport_header(skb); 21951da177e4SLinus Torvalds 2196aa8223c7SArnaldo Carvalho de Melo th = tcp_hdr(skb); 21971da177e4SLinus Torvalds memset(th, 0, sizeof(struct tcphdr)); 21981da177e4SLinus Torvalds th->syn = 1; 21991da177e4SLinus Torvalds th->ack = 1; 22001da177e4SLinus Torvalds TCP_ECN_make_synack(req, th); 22011da177e4SLinus Torvalds th->source = inet_sk(sk)->sport; 22022e6599cbSArnaldo Carvalho de Melo th->dest = ireq->rmt_port; 22032e6599cbSArnaldo Carvalho de Melo TCP_SKB_CB(skb)->seq = tcp_rsk(req)->snt_isn; 22041da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + 1; 22051da177e4SLinus Torvalds TCP_SKB_CB(skb)->sacked = 0; 22067967168cSHerbert Xu skb_shinfo(skb)->gso_segs = 1; 22077967168cSHerbert Xu skb_shinfo(skb)->gso_size = 0; 22087967168cSHerbert Xu skb_shinfo(skb)->gso_type = 0; 22091da177e4SLinus Torvalds th->seq = htonl(TCP_SKB_CB(skb)->seq); 22102e6599cbSArnaldo Carvalho de Melo th->ack_seq = htonl(tcp_rsk(req)->rcv_isn + 1); 22111da177e4SLinus Torvalds if (req->rcv_wnd == 0) { /* ignored for retransmitted syns */ 22121da177e4SLinus Torvalds __u8 rcv_wscale; 22131da177e4SLinus Torvalds /* Set this up on the first call only */ 22141da177e4SLinus Torvalds req->window_clamp = tp->window_clamp ? : dst_metric(dst, RTAX_WINDOW); 22151da177e4SLinus Torvalds /* tcp_full_space because it is guaranteed to be the first packet */ 22161da177e4SLinus Torvalds tcp_select_initial_window(tcp_full_space(sk), 22172e6599cbSArnaldo Carvalho de Melo dst_metric(dst, RTAX_ADVMSS) - (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0), 22181da177e4SLinus Torvalds &req->rcv_wnd, 22191da177e4SLinus Torvalds &req->window_clamp, 22202e6599cbSArnaldo Carvalho de Melo ireq->wscale_ok, 22211da177e4SLinus Torvalds &rcv_wscale); 22222e6599cbSArnaldo Carvalho de Melo ireq->rcv_wscale = rcv_wscale; 22231da177e4SLinus Torvalds } 22241da177e4SLinus Torvalds 22251da177e4SLinus Torvalds /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */ 2226600ff0c2SIlpo Järvinen th->window = htons(min(req->rcv_wnd, 65535U)); 22271da177e4SLinus Torvalds 22281da177e4SLinus Torvalds TCP_SKB_CB(skb)->when = tcp_time_stamp; 2229df7a3b07SAl Viro tcp_syn_build_options((__be32 *)(th + 1), dst_metric(dst, RTAX_ADVMSS), ireq->tstamp_ok, 22302e6599cbSArnaldo Carvalho de Melo ireq->sack_ok, ireq->wscale_ok, ireq->rcv_wscale, 22311da177e4SLinus Torvalds TCP_SKB_CB(skb)->when, 2232cfb6eeb4SYOSHIFUJI Hideaki req->ts_recent, 2233cfb6eeb4SYOSHIFUJI Hideaki ( 2234cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 2235cfb6eeb4SYOSHIFUJI Hideaki md5 ? &md5_hash_location : 2236cfb6eeb4SYOSHIFUJI Hideaki #endif 2237cfb6eeb4SYOSHIFUJI Hideaki NULL) 2238cfb6eeb4SYOSHIFUJI Hideaki ); 22391da177e4SLinus Torvalds 22401da177e4SLinus Torvalds skb->csum = 0; 22411da177e4SLinus Torvalds th->doff = (tcp_header_size >> 2); 22421da177e4SLinus Torvalds TCP_INC_STATS(TCP_MIB_OUTSEGS); 2243cfb6eeb4SYOSHIFUJI Hideaki 2244cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 2245cfb6eeb4SYOSHIFUJI Hideaki /* Okay, we have all we need - do the md5 hash if needed */ 2246cfb6eeb4SYOSHIFUJI Hideaki if (md5) { 2247cfb6eeb4SYOSHIFUJI Hideaki tp->af_specific->calc_md5_hash(md5_hash_location, 2248cfb6eeb4SYOSHIFUJI Hideaki md5, 2249cfb6eeb4SYOSHIFUJI Hideaki NULL, dst, req, 2250aa8223c7SArnaldo Carvalho de Melo tcp_hdr(skb), sk->sk_protocol, 2251cfb6eeb4SYOSHIFUJI Hideaki skb->len); 2252cfb6eeb4SYOSHIFUJI Hideaki } 2253cfb6eeb4SYOSHIFUJI Hideaki #endif 2254cfb6eeb4SYOSHIFUJI Hideaki 22551da177e4SLinus Torvalds return skb; 22561da177e4SLinus Torvalds } 22571da177e4SLinus Torvalds 22581da177e4SLinus Torvalds /* 22591da177e4SLinus Torvalds * Do all connect socket setups that can be done AF independent. 22601da177e4SLinus Torvalds */ 226140efc6faSStephen Hemminger static void tcp_connect_init(struct sock *sk) 22621da177e4SLinus Torvalds { 22631da177e4SLinus Torvalds struct dst_entry *dst = __sk_dst_get(sk); 22641da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 22651da177e4SLinus Torvalds __u8 rcv_wscale; 22661da177e4SLinus Torvalds 22671da177e4SLinus Torvalds /* We'll fix this up when we get a response from the other end. 22681da177e4SLinus Torvalds * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT. 22691da177e4SLinus Torvalds */ 22701da177e4SLinus Torvalds tp->tcp_header_len = sizeof(struct tcphdr) + 22711da177e4SLinus Torvalds (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0); 22721da177e4SLinus Torvalds 2273cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 2274cfb6eeb4SYOSHIFUJI Hideaki if (tp->af_specific->md5_lookup(sk, sk) != NULL) 2275cfb6eeb4SYOSHIFUJI Hideaki tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED; 2276cfb6eeb4SYOSHIFUJI Hideaki #endif 2277cfb6eeb4SYOSHIFUJI Hideaki 22781da177e4SLinus Torvalds /* If user gave his TCP_MAXSEG, record it to clamp */ 22791da177e4SLinus Torvalds if (tp->rx_opt.user_mss) 22801da177e4SLinus Torvalds tp->rx_opt.mss_clamp = tp->rx_opt.user_mss; 22811da177e4SLinus Torvalds tp->max_window = 0; 22825d424d5aSJohn Heffner tcp_mtup_init(sk); 22831da177e4SLinus Torvalds tcp_sync_mss(sk, dst_mtu(dst)); 22841da177e4SLinus Torvalds 22851da177e4SLinus Torvalds if (!tp->window_clamp) 22861da177e4SLinus Torvalds tp->window_clamp = dst_metric(dst, RTAX_WINDOW); 22871da177e4SLinus Torvalds tp->advmss = dst_metric(dst, RTAX_ADVMSS); 22881da177e4SLinus Torvalds tcp_initialize_rcv_mss(sk); 22891da177e4SLinus Torvalds 22901da177e4SLinus Torvalds tcp_select_initial_window(tcp_full_space(sk), 22911da177e4SLinus Torvalds tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0), 22921da177e4SLinus Torvalds &tp->rcv_wnd, 22931da177e4SLinus Torvalds &tp->window_clamp, 22941da177e4SLinus Torvalds sysctl_tcp_window_scaling, 22951da177e4SLinus Torvalds &rcv_wscale); 22961da177e4SLinus Torvalds 22971da177e4SLinus Torvalds tp->rx_opt.rcv_wscale = rcv_wscale; 22981da177e4SLinus Torvalds tp->rcv_ssthresh = tp->rcv_wnd; 22991da177e4SLinus Torvalds 23001da177e4SLinus Torvalds sk->sk_err = 0; 23011da177e4SLinus Torvalds sock_reset_flag(sk, SOCK_DONE); 23021da177e4SLinus Torvalds tp->snd_wnd = 0; 23031da177e4SLinus Torvalds tcp_init_wl(tp, tp->write_seq, 0); 23041da177e4SLinus Torvalds tp->snd_una = tp->write_seq; 23051da177e4SLinus Torvalds tp->snd_sml = tp->write_seq; 23061da177e4SLinus Torvalds tp->rcv_nxt = 0; 23071da177e4SLinus Torvalds tp->rcv_wup = 0; 23081da177e4SLinus Torvalds tp->copied_seq = 0; 23091da177e4SLinus Torvalds 2310463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT; 2311463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_retransmits = 0; 23121da177e4SLinus Torvalds tcp_clear_retrans(tp); 23131da177e4SLinus Torvalds } 23141da177e4SLinus Torvalds 23151da177e4SLinus Torvalds /* 23161da177e4SLinus Torvalds * Build a SYN and send it off. 23171da177e4SLinus Torvalds */ 23181da177e4SLinus Torvalds int tcp_connect(struct sock *sk) 23191da177e4SLinus Torvalds { 23201da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 23211da177e4SLinus Torvalds struct sk_buff *buff; 23221da177e4SLinus Torvalds 23231da177e4SLinus Torvalds tcp_connect_init(sk); 23241da177e4SLinus Torvalds 2325d179cd12SDavid S. Miller buff = alloc_skb_fclone(MAX_TCP_HEADER + 15, sk->sk_allocation); 23261da177e4SLinus Torvalds if (unlikely(buff == NULL)) 23271da177e4SLinus Torvalds return -ENOBUFS; 23281da177e4SLinus Torvalds 23291da177e4SLinus Torvalds /* Reserve space for headers. */ 23301da177e4SLinus Torvalds skb_reserve(buff, MAX_TCP_HEADER); 23311da177e4SLinus Torvalds 23321da177e4SLinus Torvalds TCP_SKB_CB(buff)->flags = TCPCB_FLAG_SYN; 23339e412ba7SIlpo Järvinen TCP_ECN_send_syn(sk, buff); 23341da177e4SLinus Torvalds TCP_SKB_CB(buff)->sacked = 0; 23357967168cSHerbert Xu skb_shinfo(buff)->gso_segs = 1; 23367967168cSHerbert Xu skb_shinfo(buff)->gso_size = 0; 23377967168cSHerbert Xu skb_shinfo(buff)->gso_type = 0; 23381da177e4SLinus Torvalds buff->csum = 0; 2339bd37a088SWei Yongjun tp->snd_nxt = tp->write_seq; 23401da177e4SLinus Torvalds TCP_SKB_CB(buff)->seq = tp->write_seq++; 23411da177e4SLinus Torvalds TCP_SKB_CB(buff)->end_seq = tp->write_seq; 23421da177e4SLinus Torvalds 23431da177e4SLinus Torvalds /* Send it off. */ 23441da177e4SLinus Torvalds TCP_SKB_CB(buff)->when = tcp_time_stamp; 23451da177e4SLinus Torvalds tp->retrans_stamp = TCP_SKB_CB(buff)->when; 23461da177e4SLinus Torvalds skb_header_release(buff); 2347fe067e8aSDavid S. Miller __tcp_add_write_queue_tail(sk, buff); 23481da177e4SLinus Torvalds sk_charge_skb(sk, buff); 23491da177e4SLinus Torvalds tp->packets_out += tcp_skb_pcount(buff); 2350dfb4b9dcSDavid S. Miller tcp_transmit_skb(sk, buff, 1, GFP_KERNEL); 2351bd37a088SWei Yongjun 2352bd37a088SWei Yongjun /* We change tp->snd_nxt after the tcp_transmit_skb() call 2353bd37a088SWei Yongjun * in order to make this packet get counted in tcpOutSegs. 2354bd37a088SWei Yongjun */ 2355bd37a088SWei Yongjun tp->snd_nxt = tp->write_seq; 2356bd37a088SWei Yongjun tp->pushed_seq = tp->write_seq; 23571da177e4SLinus Torvalds TCP_INC_STATS(TCP_MIB_ACTIVEOPENS); 23581da177e4SLinus Torvalds 23591da177e4SLinus Torvalds /* Timer for repeating the SYN until an answer. */ 23603f421baaSArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 23613f421baaSArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto, TCP_RTO_MAX); 23621da177e4SLinus Torvalds return 0; 23631da177e4SLinus Torvalds } 23641da177e4SLinus Torvalds 23651da177e4SLinus Torvalds /* Send out a delayed ack, the caller does the policy checking 23661da177e4SLinus Torvalds * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check() 23671da177e4SLinus Torvalds * for details. 23681da177e4SLinus Torvalds */ 23691da177e4SLinus Torvalds void tcp_send_delayed_ack(struct sock *sk) 23701da177e4SLinus Torvalds { 2371463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 2372463c84b9SArnaldo Carvalho de Melo int ato = icsk->icsk_ack.ato; 23731da177e4SLinus Torvalds unsigned long timeout; 23741da177e4SLinus Torvalds 23751da177e4SLinus Torvalds if (ato > TCP_DELACK_MIN) { 2376463c84b9SArnaldo Carvalho de Melo const struct tcp_sock *tp = tcp_sk(sk); 23771da177e4SLinus Torvalds int max_ato = HZ/2; 23781da177e4SLinus Torvalds 2379463c84b9SArnaldo Carvalho de Melo if (icsk->icsk_ack.pingpong || (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)) 23801da177e4SLinus Torvalds max_ato = TCP_DELACK_MAX; 23811da177e4SLinus Torvalds 23821da177e4SLinus Torvalds /* Slow path, intersegment interval is "high". */ 23831da177e4SLinus Torvalds 23841da177e4SLinus Torvalds /* If some rtt estimate is known, use it to bound delayed ack. 2385463c84b9SArnaldo Carvalho de Melo * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements 23861da177e4SLinus Torvalds * directly. 23871da177e4SLinus Torvalds */ 23881da177e4SLinus Torvalds if (tp->srtt) { 23891da177e4SLinus Torvalds int rtt = max(tp->srtt>>3, TCP_DELACK_MIN); 23901da177e4SLinus Torvalds 23911da177e4SLinus Torvalds if (rtt < max_ato) 23921da177e4SLinus Torvalds max_ato = rtt; 23931da177e4SLinus Torvalds } 23941da177e4SLinus Torvalds 23951da177e4SLinus Torvalds ato = min(ato, max_ato); 23961da177e4SLinus Torvalds } 23971da177e4SLinus Torvalds 23981da177e4SLinus Torvalds /* Stay within the limit we were given */ 23991da177e4SLinus Torvalds timeout = jiffies + ato; 24001da177e4SLinus Torvalds 24011da177e4SLinus Torvalds /* Use new timeout only if there wasn't a older one earlier. */ 2402463c84b9SArnaldo Carvalho de Melo if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) { 24031da177e4SLinus Torvalds /* If delack timer was blocked or is about to expire, 24041da177e4SLinus Torvalds * send ACK now. 24051da177e4SLinus Torvalds */ 2406463c84b9SArnaldo Carvalho de Melo if (icsk->icsk_ack.blocked || 2407463c84b9SArnaldo Carvalho de Melo time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) { 24081da177e4SLinus Torvalds tcp_send_ack(sk); 24091da177e4SLinus Torvalds return; 24101da177e4SLinus Torvalds } 24111da177e4SLinus Torvalds 2412463c84b9SArnaldo Carvalho de Melo if (!time_before(timeout, icsk->icsk_ack.timeout)) 2413463c84b9SArnaldo Carvalho de Melo timeout = icsk->icsk_ack.timeout; 24141da177e4SLinus Torvalds } 2415463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER; 2416463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.timeout = timeout; 2417463c84b9SArnaldo Carvalho de Melo sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout); 24181da177e4SLinus Torvalds } 24191da177e4SLinus Torvalds 24201da177e4SLinus Torvalds /* This routine sends an ack and also updates the window. */ 24211da177e4SLinus Torvalds void tcp_send_ack(struct sock *sk) 24221da177e4SLinus Torvalds { 24231da177e4SLinus Torvalds /* If we have been reset, we may not send again. */ 24241da177e4SLinus Torvalds if (sk->sk_state != TCP_CLOSE) { 24251da177e4SLinus Torvalds struct sk_buff *buff; 24261da177e4SLinus Torvalds 24271da177e4SLinus Torvalds /* We are not putting this on the write queue, so 24281da177e4SLinus Torvalds * tcp_transmit_skb() will set the ownership to this 24291da177e4SLinus Torvalds * sock. 24301da177e4SLinus Torvalds */ 24311da177e4SLinus Torvalds buff = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC); 24321da177e4SLinus Torvalds if (buff == NULL) { 2433463c84b9SArnaldo Carvalho de Melo inet_csk_schedule_ack(sk); 2434463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN; 24353f421baaSArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK, 24363f421baaSArnaldo Carvalho de Melo TCP_DELACK_MAX, TCP_RTO_MAX); 24371da177e4SLinus Torvalds return; 24381da177e4SLinus Torvalds } 24391da177e4SLinus Torvalds 24401da177e4SLinus Torvalds /* Reserve space for headers and prepare control bits. */ 24411da177e4SLinus Torvalds skb_reserve(buff, MAX_TCP_HEADER); 24421da177e4SLinus Torvalds buff->csum = 0; 24431da177e4SLinus Torvalds TCP_SKB_CB(buff)->flags = TCPCB_FLAG_ACK; 24441da177e4SLinus Torvalds TCP_SKB_CB(buff)->sacked = 0; 24457967168cSHerbert Xu skb_shinfo(buff)->gso_segs = 1; 24467967168cSHerbert Xu skb_shinfo(buff)->gso_size = 0; 24477967168cSHerbert Xu skb_shinfo(buff)->gso_type = 0; 24481da177e4SLinus Torvalds 24491da177e4SLinus Torvalds /* Send it off, this clears delayed acks for us. */ 24509e412ba7SIlpo Järvinen TCP_SKB_CB(buff)->seq = TCP_SKB_CB(buff)->end_seq = tcp_acceptable_seq(sk); 24511da177e4SLinus Torvalds TCP_SKB_CB(buff)->when = tcp_time_stamp; 2452dfb4b9dcSDavid S. Miller tcp_transmit_skb(sk, buff, 0, GFP_ATOMIC); 24531da177e4SLinus Torvalds } 24541da177e4SLinus Torvalds } 24551da177e4SLinus Torvalds 24561da177e4SLinus Torvalds /* This routine sends a packet with an out of date sequence 24571da177e4SLinus Torvalds * number. It assumes the other end will try to ack it. 24581da177e4SLinus Torvalds * 24591da177e4SLinus Torvalds * Question: what should we make while urgent mode? 24601da177e4SLinus Torvalds * 4.4BSD forces sending single byte of data. We cannot send 24611da177e4SLinus Torvalds * out of window data, because we have SND.NXT==SND.MAX... 24621da177e4SLinus Torvalds * 24631da177e4SLinus Torvalds * Current solution: to send TWO zero-length segments in urgent mode: 24641da177e4SLinus Torvalds * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is 24651da177e4SLinus Torvalds * out-of-date with SND.UNA-1 to probe window. 24661da177e4SLinus Torvalds */ 24671da177e4SLinus Torvalds static int tcp_xmit_probe_skb(struct sock *sk, int urgent) 24681da177e4SLinus Torvalds { 24691da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 24701da177e4SLinus Torvalds struct sk_buff *skb; 24711da177e4SLinus Torvalds 24721da177e4SLinus Torvalds /* We don't queue it, tcp_transmit_skb() sets ownership. */ 24731da177e4SLinus Torvalds skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC); 24741da177e4SLinus Torvalds if (skb == NULL) 24751da177e4SLinus Torvalds return -1; 24761da177e4SLinus Torvalds 24771da177e4SLinus Torvalds /* Reserve space for headers and set control bits. */ 24781da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 24791da177e4SLinus Torvalds skb->csum = 0; 24801da177e4SLinus Torvalds TCP_SKB_CB(skb)->flags = TCPCB_FLAG_ACK; 24811da177e4SLinus Torvalds TCP_SKB_CB(skb)->sacked = urgent; 24827967168cSHerbert Xu skb_shinfo(skb)->gso_segs = 1; 24837967168cSHerbert Xu skb_shinfo(skb)->gso_size = 0; 24847967168cSHerbert Xu skb_shinfo(skb)->gso_type = 0; 24851da177e4SLinus Torvalds 24861da177e4SLinus Torvalds /* Use a previous sequence. This should cause the other 24871da177e4SLinus Torvalds * end to send an ack. Don't queue or clone SKB, just 24881da177e4SLinus Torvalds * send it. 24891da177e4SLinus Torvalds */ 24901da177e4SLinus Torvalds TCP_SKB_CB(skb)->seq = urgent ? tp->snd_una : tp->snd_una - 1; 24911da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq; 24921da177e4SLinus Torvalds TCP_SKB_CB(skb)->when = tcp_time_stamp; 2493dfb4b9dcSDavid S. Miller return tcp_transmit_skb(sk, skb, 0, GFP_ATOMIC); 24941da177e4SLinus Torvalds } 24951da177e4SLinus Torvalds 24961da177e4SLinus Torvalds int tcp_write_wakeup(struct sock *sk) 24971da177e4SLinus Torvalds { 24981da177e4SLinus Torvalds if (sk->sk_state != TCP_CLOSE) { 24991da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 25001da177e4SLinus Torvalds struct sk_buff *skb; 25011da177e4SLinus Torvalds 2502fe067e8aSDavid S. Miller if ((skb = tcp_send_head(sk)) != NULL && 25031da177e4SLinus Torvalds before(TCP_SKB_CB(skb)->seq, tp->snd_una+tp->snd_wnd)) { 25041da177e4SLinus Torvalds int err; 25051da177e4SLinus Torvalds unsigned int mss = tcp_current_mss(sk, 0); 25061da177e4SLinus Torvalds unsigned int seg_size = tp->snd_una+tp->snd_wnd-TCP_SKB_CB(skb)->seq; 25071da177e4SLinus Torvalds 25081da177e4SLinus Torvalds if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq)) 25091da177e4SLinus Torvalds tp->pushed_seq = TCP_SKB_CB(skb)->end_seq; 25101da177e4SLinus Torvalds 25111da177e4SLinus Torvalds /* We are probing the opening of a window 25121da177e4SLinus Torvalds * but the window size is != 0 25131da177e4SLinus Torvalds * must have been a result SWS avoidance ( sender ) 25141da177e4SLinus Torvalds */ 25151da177e4SLinus Torvalds if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq || 25161da177e4SLinus Torvalds skb->len > mss) { 25171da177e4SLinus Torvalds seg_size = min(seg_size, mss); 25181da177e4SLinus Torvalds TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH; 2519846998aeSDavid S. Miller if (tcp_fragment(sk, skb, seg_size, mss)) 25201da177e4SLinus Torvalds return -1; 25211da177e4SLinus Torvalds } else if (!tcp_skb_pcount(skb)) 2522846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, skb, mss); 25231da177e4SLinus Torvalds 25241da177e4SLinus Torvalds TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH; 25251da177e4SLinus Torvalds TCP_SKB_CB(skb)->when = tcp_time_stamp; 2526dfb4b9dcSDavid S. Miller err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 25271da177e4SLinus Torvalds if (!err) { 25289e412ba7SIlpo Järvinen update_send_head(sk, skb); 25291da177e4SLinus Torvalds } 25301da177e4SLinus Torvalds return err; 25311da177e4SLinus Torvalds } else { 25321da177e4SLinus Torvalds if (tp->urg_mode && 25331da177e4SLinus Torvalds between(tp->snd_up, tp->snd_una+1, tp->snd_una+0xFFFF)) 25341da177e4SLinus Torvalds tcp_xmit_probe_skb(sk, TCPCB_URG); 25351da177e4SLinus Torvalds return tcp_xmit_probe_skb(sk, 0); 25361da177e4SLinus Torvalds } 25371da177e4SLinus Torvalds } 25381da177e4SLinus Torvalds return -1; 25391da177e4SLinus Torvalds } 25401da177e4SLinus Torvalds 25411da177e4SLinus Torvalds /* A window probe timeout has occurred. If window is not closed send 25421da177e4SLinus Torvalds * a partial packet else a zero probe. 25431da177e4SLinus Torvalds */ 25441da177e4SLinus Torvalds void tcp_send_probe0(struct sock *sk) 25451da177e4SLinus Torvalds { 2546463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 25471da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 25481da177e4SLinus Torvalds int err; 25491da177e4SLinus Torvalds 25501da177e4SLinus Torvalds err = tcp_write_wakeup(sk); 25511da177e4SLinus Torvalds 2552fe067e8aSDavid S. Miller if (tp->packets_out || !tcp_send_head(sk)) { 25531da177e4SLinus Torvalds /* Cancel probe timer, if it is not required. */ 25546687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out = 0; 2555463c84b9SArnaldo Carvalho de Melo icsk->icsk_backoff = 0; 25561da177e4SLinus Torvalds return; 25571da177e4SLinus Torvalds } 25581da177e4SLinus Torvalds 25591da177e4SLinus Torvalds if (err <= 0) { 2560463c84b9SArnaldo Carvalho de Melo if (icsk->icsk_backoff < sysctl_tcp_retries2) 2561463c84b9SArnaldo Carvalho de Melo icsk->icsk_backoff++; 25626687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out++; 2563463c84b9SArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0, 25643f421baaSArnaldo Carvalho de Melo min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX), 25653f421baaSArnaldo Carvalho de Melo TCP_RTO_MAX); 25661da177e4SLinus Torvalds } else { 25671da177e4SLinus Torvalds /* If packet was not sent due to local congestion, 25686687e988SArnaldo Carvalho de Melo * do not backoff and do not remember icsk_probes_out. 25691da177e4SLinus Torvalds * Let local senders to fight for local resources. 25701da177e4SLinus Torvalds * 25711da177e4SLinus Torvalds * Use accumulated backoff yet. 25721da177e4SLinus Torvalds */ 25736687e988SArnaldo Carvalho de Melo if (!icsk->icsk_probes_out) 25746687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out = 1; 2575463c84b9SArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0, 2576463c84b9SArnaldo Carvalho de Melo min(icsk->icsk_rto << icsk->icsk_backoff, 25773f421baaSArnaldo Carvalho de Melo TCP_RESOURCE_PROBE_INTERVAL), 25783f421baaSArnaldo Carvalho de Melo TCP_RTO_MAX); 25791da177e4SLinus Torvalds } 25801da177e4SLinus Torvalds } 25811da177e4SLinus Torvalds 25821da177e4SLinus Torvalds EXPORT_SYMBOL(tcp_connect); 25831da177e4SLinus Torvalds EXPORT_SYMBOL(tcp_make_synack); 25841da177e4SLinus Torvalds EXPORT_SYMBOL(tcp_simple_retransmit); 25851da177e4SLinus Torvalds EXPORT_SYMBOL(tcp_sync_mss); 2586f4805edeSStephen Hemminger EXPORT_SYMBOL(sysctl_tcp_tso_win_divisor); 25875d424d5aSJohn Heffner EXPORT_SYMBOL(tcp_mtup_init); 2588