11da177e4SLinus Torvalds /* 21da177e4SLinus Torvalds * INET An implementation of the TCP/IP protocol suite for the LINUX 31da177e4SLinus Torvalds * operating system. INET is implemented using the BSD Socket 41da177e4SLinus Torvalds * interface as the means of communication with the user level. 51da177e4SLinus Torvalds * 61da177e4SLinus Torvalds * Implementation of the Transmission Control Protocol(TCP). 71da177e4SLinus Torvalds * 802c30a84SJesper Juhl * Authors: Ross Biro 91da177e4SLinus Torvalds * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG> 101da177e4SLinus Torvalds * Mark Evans, <evansmp@uhura.aston.ac.uk> 111da177e4SLinus Torvalds * Corey Minyard <wf-rch!minyard@relay.EU.net> 121da177e4SLinus Torvalds * Florian La Roche, <flla@stud.uni-sb.de> 131da177e4SLinus Torvalds * Charles Hedrick, <hedrick@klinzhai.rutgers.edu> 141da177e4SLinus Torvalds * Linus Torvalds, <torvalds@cs.helsinki.fi> 151da177e4SLinus Torvalds * Alan Cox, <gw4pts@gw4pts.ampr.org> 161da177e4SLinus Torvalds * Matthew Dillon, <dillon@apollo.west.oic.com> 171da177e4SLinus Torvalds * Arnt Gulbrandsen, <agulbra@nvg.unit.no> 181da177e4SLinus Torvalds * Jorge Cwik, <jorge@laser.satlink.net> 191da177e4SLinus Torvalds */ 201da177e4SLinus Torvalds 211da177e4SLinus Torvalds /* 221da177e4SLinus Torvalds * Changes: Pedro Roque : Retransmit queue handled by TCP. 231da177e4SLinus Torvalds * : Fragmentation on mtu decrease 241da177e4SLinus Torvalds * : Segment collapse on retransmit 251da177e4SLinus Torvalds * : AF independence 261da177e4SLinus Torvalds * 271da177e4SLinus Torvalds * Linus Torvalds : send_delayed_ack 281da177e4SLinus Torvalds * David S. Miller : Charge memory using the right skb 291da177e4SLinus Torvalds * during syn/ack processing. 301da177e4SLinus Torvalds * David S. Miller : Output engine completely rewritten. 311da177e4SLinus Torvalds * Andrea Arcangeli: SYNACK carry ts_recent in tsecr. 321da177e4SLinus Torvalds * Cacophonix Gaul : draft-minshall-nagle-01 331da177e4SLinus Torvalds * J Hadi Salim : ECN support 341da177e4SLinus Torvalds * 351da177e4SLinus Torvalds */ 361da177e4SLinus Torvalds 371da177e4SLinus Torvalds #include <net/tcp.h> 381da177e4SLinus Torvalds 391da177e4SLinus Torvalds #include <linux/compiler.h> 401da177e4SLinus Torvalds #include <linux/module.h> 411da177e4SLinus Torvalds 421da177e4SLinus Torvalds /* People can turn this off for buggy TCP's found in printers etc. */ 43ab32ea5dSBrian Haley int sysctl_tcp_retrans_collapse __read_mostly = 1; 441da177e4SLinus Torvalds 4515d99e02SRick Jones /* People can turn this on to work with those rare, broken TCPs that 4615d99e02SRick Jones * interpret the window field as a signed quantity. 4715d99e02SRick Jones */ 48ab32ea5dSBrian Haley int sysctl_tcp_workaround_signed_windows __read_mostly = 0; 4915d99e02SRick Jones 501da177e4SLinus Torvalds /* This limits the percentage of the congestion window which we 511da177e4SLinus Torvalds * will allow a single TSO frame to consume. Building TSO frames 521da177e4SLinus Torvalds * which are too large can cause TCP streams to be bursty. 531da177e4SLinus Torvalds */ 54ab32ea5dSBrian Haley int sysctl_tcp_tso_win_divisor __read_mostly = 3; 551da177e4SLinus Torvalds 56ab32ea5dSBrian Haley int sysctl_tcp_mtu_probing __read_mostly = 0; 57ab32ea5dSBrian Haley int sysctl_tcp_base_mss __read_mostly = 512; 585d424d5aSJohn Heffner 5935089bb2SDavid S. Miller /* By default, RFC2861 behavior. */ 60ab32ea5dSBrian Haley int sysctl_tcp_slow_start_after_idle __read_mostly = 1; 6135089bb2SDavid S. Miller 6266f5fe62SIlpo Järvinen static void tcp_event_new_data_sent(struct sock *sk, struct sk_buff *skb) 636ff03ac3SIlpo Järvinen { 646ff03ac3SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 6566f5fe62SIlpo Järvinen unsigned int prior_packets = tp->packets_out; 669e412ba7SIlpo Järvinen 67fe067e8aSDavid S. Miller tcp_advance_send_head(sk, skb); 681da177e4SLinus Torvalds tp->snd_nxt = TCP_SKB_CB(skb)->end_seq; 698512430eSIlpo Järvinen 708512430eSIlpo Järvinen /* Don't override Nagle indefinately with F-RTO */ 718512430eSIlpo Järvinen if (tp->frto_counter == 2) 728512430eSIlpo Järvinen tp->frto_counter = 3; 7366f5fe62SIlpo Järvinen 7466f5fe62SIlpo Järvinen tp->packets_out += tcp_skb_pcount(skb); 7566f5fe62SIlpo Järvinen if (!prior_packets) 7666f5fe62SIlpo Järvinen inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 7766f5fe62SIlpo Järvinen inet_csk(sk)->icsk_rto, TCP_RTO_MAX); 781da177e4SLinus Torvalds } 791da177e4SLinus Torvalds 801da177e4SLinus Torvalds /* SND.NXT, if window was not shrunk. 811da177e4SLinus Torvalds * If window has been shrunk, what should we make? It is not clear at all. 821da177e4SLinus Torvalds * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-( 831da177e4SLinus Torvalds * Anything in between SND.UNA...SND.UNA+SND.WND also can be already 841da177e4SLinus Torvalds * invalid. OK, let's make this for now: 851da177e4SLinus Torvalds */ 869e412ba7SIlpo Järvinen static inline __u32 tcp_acceptable_seq(struct sock *sk) 871da177e4SLinus Torvalds { 889e412ba7SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 899e412ba7SIlpo Järvinen 9090840defSIlpo Järvinen if (!before(tcp_wnd_end(tp), tp->snd_nxt)) 911da177e4SLinus Torvalds return tp->snd_nxt; 921da177e4SLinus Torvalds else 9390840defSIlpo Järvinen return tcp_wnd_end(tp); 941da177e4SLinus Torvalds } 951da177e4SLinus Torvalds 961da177e4SLinus Torvalds /* Calculate mss to advertise in SYN segment. 971da177e4SLinus Torvalds * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that: 981da177e4SLinus Torvalds * 991da177e4SLinus Torvalds * 1. It is independent of path mtu. 1001da177e4SLinus Torvalds * 2. Ideally, it is maximal possible segment size i.e. 65535-40. 1011da177e4SLinus Torvalds * 3. For IPv4 it is reasonable to calculate it from maximal MTU of 1021da177e4SLinus Torvalds * attached devices, because some buggy hosts are confused by 1031da177e4SLinus Torvalds * large MSS. 1041da177e4SLinus Torvalds * 4. We do not make 3, we advertise MSS, calculated from first 1051da177e4SLinus Torvalds * hop device mtu, but allow to raise it to ip_rt_min_advmss. 1061da177e4SLinus Torvalds * This may be overridden via information stored in routing table. 1071da177e4SLinus Torvalds * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible, 1081da177e4SLinus Torvalds * probably even Jumbo". 1091da177e4SLinus Torvalds */ 1101da177e4SLinus Torvalds static __u16 tcp_advertise_mss(struct sock *sk) 1111da177e4SLinus Torvalds { 1121da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 1131da177e4SLinus Torvalds struct dst_entry *dst = __sk_dst_get(sk); 1141da177e4SLinus Torvalds int mss = tp->advmss; 1151da177e4SLinus Torvalds 1161da177e4SLinus Torvalds if (dst && dst_metric(dst, RTAX_ADVMSS) < mss) { 1171da177e4SLinus Torvalds mss = dst_metric(dst, RTAX_ADVMSS); 1181da177e4SLinus Torvalds tp->advmss = mss; 1191da177e4SLinus Torvalds } 1201da177e4SLinus Torvalds 1211da177e4SLinus Torvalds return (__u16)mss; 1221da177e4SLinus Torvalds } 1231da177e4SLinus Torvalds 1241da177e4SLinus Torvalds /* RFC2861. Reset CWND after idle period longer RTO to "restart window". 1251da177e4SLinus Torvalds * This is the first part of cwnd validation mechanism. */ 126463c84b9SArnaldo Carvalho de Melo static void tcp_cwnd_restart(struct sock *sk, struct dst_entry *dst) 1271da177e4SLinus Torvalds { 128463c84b9SArnaldo Carvalho de Melo struct tcp_sock *tp = tcp_sk(sk); 1291da177e4SLinus Torvalds s32 delta = tcp_time_stamp - tp->lsndtime; 1301da177e4SLinus Torvalds u32 restart_cwnd = tcp_init_cwnd(tp, dst); 1311da177e4SLinus Torvalds u32 cwnd = tp->snd_cwnd; 1321da177e4SLinus Torvalds 1336687e988SArnaldo Carvalho de Melo tcp_ca_event(sk, CA_EVENT_CWND_RESTART); 1341da177e4SLinus Torvalds 1356687e988SArnaldo Carvalho de Melo tp->snd_ssthresh = tcp_current_ssthresh(sk); 1361da177e4SLinus Torvalds restart_cwnd = min(restart_cwnd, cwnd); 1371da177e4SLinus Torvalds 138463c84b9SArnaldo Carvalho de Melo while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd) 1391da177e4SLinus Torvalds cwnd >>= 1; 1401da177e4SLinus Torvalds tp->snd_cwnd = max(cwnd, restart_cwnd); 1411da177e4SLinus Torvalds tp->snd_cwnd_stamp = tcp_time_stamp; 1421da177e4SLinus Torvalds tp->snd_cwnd_used = 0; 1431da177e4SLinus Torvalds } 1441da177e4SLinus Torvalds 14540efc6faSStephen Hemminger static void tcp_event_data_sent(struct tcp_sock *tp, 1461da177e4SLinus Torvalds struct sk_buff *skb, struct sock *sk) 1471da177e4SLinus Torvalds { 148463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 149463c84b9SArnaldo Carvalho de Melo const u32 now = tcp_time_stamp; 1501da177e4SLinus Torvalds 15135089bb2SDavid S. Miller if (sysctl_tcp_slow_start_after_idle && 15235089bb2SDavid S. Miller (!tp->packets_out && (s32)(now - tp->lsndtime) > icsk->icsk_rto)) 153463c84b9SArnaldo Carvalho de Melo tcp_cwnd_restart(sk, __sk_dst_get(sk)); 1541da177e4SLinus Torvalds 1551da177e4SLinus Torvalds tp->lsndtime = now; 1561da177e4SLinus Torvalds 1571da177e4SLinus Torvalds /* If it is a reply for ato after last received 1581da177e4SLinus Torvalds * packet, enter pingpong mode. 1591da177e4SLinus Torvalds */ 160463c84b9SArnaldo Carvalho de Melo if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato) 161463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.pingpong = 1; 1621da177e4SLinus Torvalds } 1631da177e4SLinus Torvalds 16440efc6faSStephen Hemminger static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts) 1651da177e4SLinus Torvalds { 166463c84b9SArnaldo Carvalho de Melo tcp_dec_quickack_mode(sk, pkts); 167463c84b9SArnaldo Carvalho de Melo inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK); 1681da177e4SLinus Torvalds } 1691da177e4SLinus Torvalds 1701da177e4SLinus Torvalds /* Determine a window scaling and initial window to offer. 1711da177e4SLinus Torvalds * Based on the assumption that the given amount of space 1721da177e4SLinus Torvalds * will be offered. Store the results in the tp structure. 1731da177e4SLinus Torvalds * NOTE: for smooth operation initial space offering should 1741da177e4SLinus Torvalds * be a multiple of mss if possible. We assume here that mss >= 1. 1751da177e4SLinus Torvalds * This MUST be enforced by all callers. 1761da177e4SLinus Torvalds */ 1771da177e4SLinus Torvalds void tcp_select_initial_window(int __space, __u32 mss, 1781da177e4SLinus Torvalds __u32 *rcv_wnd, __u32 *window_clamp, 1791da177e4SLinus Torvalds int wscale_ok, __u8 *rcv_wscale) 1801da177e4SLinus Torvalds { 1811da177e4SLinus Torvalds unsigned int space = (__space < 0 ? 0 : __space); 1821da177e4SLinus Torvalds 1831da177e4SLinus Torvalds /* If no clamp set the clamp to the max possible scaled window */ 1841da177e4SLinus Torvalds if (*window_clamp == 0) 1851da177e4SLinus Torvalds (*window_clamp) = (65535 << 14); 1861da177e4SLinus Torvalds space = min(*window_clamp, space); 1871da177e4SLinus Torvalds 1881da177e4SLinus Torvalds /* Quantize space offering to a multiple of mss if possible. */ 1891da177e4SLinus Torvalds if (space > mss) 1901da177e4SLinus Torvalds space = (space / mss) * mss; 1911da177e4SLinus Torvalds 1921da177e4SLinus Torvalds /* NOTE: offering an initial window larger than 32767 19315d99e02SRick Jones * will break some buggy TCP stacks. If the admin tells us 19415d99e02SRick Jones * it is likely we could be speaking with such a buggy stack 19515d99e02SRick Jones * we will truncate our initial window offering to 32K-1 19615d99e02SRick Jones * unless the remote has sent us a window scaling option, 19715d99e02SRick Jones * which we interpret as a sign the remote TCP is not 19815d99e02SRick Jones * misinterpreting the window field as a signed quantity. 1991da177e4SLinus Torvalds */ 20015d99e02SRick Jones if (sysctl_tcp_workaround_signed_windows) 2011da177e4SLinus Torvalds (*rcv_wnd) = min(space, MAX_TCP_WINDOW); 20215d99e02SRick Jones else 20315d99e02SRick Jones (*rcv_wnd) = space; 20415d99e02SRick Jones 2051da177e4SLinus Torvalds (*rcv_wscale) = 0; 2061da177e4SLinus Torvalds if (wscale_ok) { 2071da177e4SLinus Torvalds /* Set window scaling on max possible window 2081da177e4SLinus Torvalds * See RFC1323 for an explanation of the limit to 14 2091da177e4SLinus Torvalds */ 2101da177e4SLinus Torvalds space = max_t(u32, sysctl_tcp_rmem[2], sysctl_rmem_max); 211316c1592SStephen Hemminger space = min_t(u32, space, *window_clamp); 2121da177e4SLinus Torvalds while (space > 65535 && (*rcv_wscale) < 14) { 2131da177e4SLinus Torvalds space >>= 1; 2141da177e4SLinus Torvalds (*rcv_wscale)++; 2151da177e4SLinus Torvalds } 2161da177e4SLinus Torvalds } 2171da177e4SLinus Torvalds 2181da177e4SLinus Torvalds /* Set initial window to value enough for senders, 2196b251858SDavid S. Miller * following RFC2414. Senders, not following this RFC, 2201da177e4SLinus Torvalds * will be satisfied with 2. 2211da177e4SLinus Torvalds */ 2221da177e4SLinus Torvalds if (mss > (1 << *rcv_wscale)) { 22301ff367eSDavid S. Miller int init_cwnd = 4; 22401ff367eSDavid S. Miller if (mss > 1460 * 3) 2251da177e4SLinus Torvalds init_cwnd = 2; 22601ff367eSDavid S. Miller else if (mss > 1460) 22701ff367eSDavid S. Miller init_cwnd = 3; 2281da177e4SLinus Torvalds if (*rcv_wnd > init_cwnd * mss) 2291da177e4SLinus Torvalds *rcv_wnd = init_cwnd * mss; 2301da177e4SLinus Torvalds } 2311da177e4SLinus Torvalds 2321da177e4SLinus Torvalds /* Set the clamp no higher than max representable value */ 2331da177e4SLinus Torvalds (*window_clamp) = min(65535U << (*rcv_wscale), *window_clamp); 2341da177e4SLinus Torvalds } 2351da177e4SLinus Torvalds 2361da177e4SLinus Torvalds /* Chose a new window to advertise, update state in tcp_sock for the 2371da177e4SLinus Torvalds * socket, and return result with RFC1323 scaling applied. The return 2381da177e4SLinus Torvalds * value can be stuffed directly into th->window for an outgoing 2391da177e4SLinus Torvalds * frame. 2401da177e4SLinus Torvalds */ 24140efc6faSStephen Hemminger static u16 tcp_select_window(struct sock *sk) 2421da177e4SLinus Torvalds { 2431da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 2441da177e4SLinus Torvalds u32 cur_win = tcp_receive_window(tp); 2451da177e4SLinus Torvalds u32 new_win = __tcp_select_window(sk); 2461da177e4SLinus Torvalds 2471da177e4SLinus Torvalds /* Never shrink the offered window */ 2481da177e4SLinus Torvalds if (new_win < cur_win) { 2491da177e4SLinus Torvalds /* Danger Will Robinson! 2501da177e4SLinus Torvalds * Don't update rcv_wup/rcv_wnd here or else 2511da177e4SLinus Torvalds * we will not be able to advertise a zero 2521da177e4SLinus Torvalds * window in time. --DaveM 2531da177e4SLinus Torvalds * 2541da177e4SLinus Torvalds * Relax Will Robinson. 2551da177e4SLinus Torvalds */ 256607bfbf2SPatrick McHardy new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale); 2571da177e4SLinus Torvalds } 2581da177e4SLinus Torvalds tp->rcv_wnd = new_win; 2591da177e4SLinus Torvalds tp->rcv_wup = tp->rcv_nxt; 2601da177e4SLinus Torvalds 2611da177e4SLinus Torvalds /* Make sure we do not exceed the maximum possible 2621da177e4SLinus Torvalds * scaled window. 2631da177e4SLinus Torvalds */ 26415d99e02SRick Jones if (!tp->rx_opt.rcv_wscale && sysctl_tcp_workaround_signed_windows) 2651da177e4SLinus Torvalds new_win = min(new_win, MAX_TCP_WINDOW); 2661da177e4SLinus Torvalds else 2671da177e4SLinus Torvalds new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale)); 2681da177e4SLinus Torvalds 2691da177e4SLinus Torvalds /* RFC1323 scaling applied */ 2701da177e4SLinus Torvalds new_win >>= tp->rx_opt.rcv_wscale; 2711da177e4SLinus Torvalds 2721da177e4SLinus Torvalds /* If we advertise zero window, disable fast path. */ 2731da177e4SLinus Torvalds if (new_win == 0) 2741da177e4SLinus Torvalds tp->pred_flags = 0; 2751da177e4SLinus Torvalds 2761da177e4SLinus Torvalds return new_win; 2771da177e4SLinus Torvalds } 2781da177e4SLinus Torvalds 279056834d9SIlpo Järvinen static inline void TCP_ECN_send_synack(struct tcp_sock *tp, struct sk_buff *skb) 280bdf1ee5dSIlpo Järvinen { 281bdf1ee5dSIlpo Järvinen TCP_SKB_CB(skb)->flags &= ~TCPCB_FLAG_CWR; 282bdf1ee5dSIlpo Järvinen if (!(tp->ecn_flags & TCP_ECN_OK)) 283bdf1ee5dSIlpo Järvinen TCP_SKB_CB(skb)->flags &= ~TCPCB_FLAG_ECE; 284bdf1ee5dSIlpo Järvinen } 285bdf1ee5dSIlpo Järvinen 286bdf1ee5dSIlpo Järvinen static inline void TCP_ECN_send_syn(struct sock *sk, struct sk_buff *skb) 287bdf1ee5dSIlpo Järvinen { 288bdf1ee5dSIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 289bdf1ee5dSIlpo Järvinen 290bdf1ee5dSIlpo Järvinen tp->ecn_flags = 0; 291bdf1ee5dSIlpo Järvinen if (sysctl_tcp_ecn) { 292bdf1ee5dSIlpo Järvinen TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_ECE | TCPCB_FLAG_CWR; 293bdf1ee5dSIlpo Järvinen tp->ecn_flags = TCP_ECN_OK; 294bdf1ee5dSIlpo Järvinen } 295bdf1ee5dSIlpo Järvinen } 296bdf1ee5dSIlpo Järvinen 297bdf1ee5dSIlpo Järvinen static __inline__ void 298bdf1ee5dSIlpo Järvinen TCP_ECN_make_synack(struct request_sock *req, struct tcphdr *th) 299bdf1ee5dSIlpo Järvinen { 300bdf1ee5dSIlpo Järvinen if (inet_rsk(req)->ecn_ok) 301bdf1ee5dSIlpo Järvinen th->ece = 1; 302bdf1ee5dSIlpo Järvinen } 303bdf1ee5dSIlpo Järvinen 304bdf1ee5dSIlpo Järvinen static inline void TCP_ECN_send(struct sock *sk, struct sk_buff *skb, 305bdf1ee5dSIlpo Järvinen int tcp_header_len) 306bdf1ee5dSIlpo Järvinen { 307bdf1ee5dSIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 308bdf1ee5dSIlpo Järvinen 309bdf1ee5dSIlpo Järvinen if (tp->ecn_flags & TCP_ECN_OK) { 310bdf1ee5dSIlpo Järvinen /* Not-retransmitted data segment: set ECT and inject CWR. */ 311bdf1ee5dSIlpo Järvinen if (skb->len != tcp_header_len && 312bdf1ee5dSIlpo Järvinen !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) { 313bdf1ee5dSIlpo Järvinen INET_ECN_xmit(sk); 314bdf1ee5dSIlpo Järvinen if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) { 315bdf1ee5dSIlpo Järvinen tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR; 316bdf1ee5dSIlpo Järvinen tcp_hdr(skb)->cwr = 1; 317bdf1ee5dSIlpo Järvinen skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN; 318bdf1ee5dSIlpo Järvinen } 319bdf1ee5dSIlpo Järvinen } else { 320bdf1ee5dSIlpo Järvinen /* ACK or retransmitted segment: clear ECT|CE */ 321bdf1ee5dSIlpo Järvinen INET_ECN_dontxmit(sk); 322bdf1ee5dSIlpo Järvinen } 323bdf1ee5dSIlpo Järvinen if (tp->ecn_flags & TCP_ECN_DEMAND_CWR) 324bdf1ee5dSIlpo Järvinen tcp_hdr(skb)->ece = 1; 325bdf1ee5dSIlpo Järvinen } 326bdf1ee5dSIlpo Järvinen } 327bdf1ee5dSIlpo Järvinen 328e870a8efSIlpo Järvinen /* Constructs common control bits of non-data skb. If SYN/FIN is present, 329e870a8efSIlpo Järvinen * auto increment end seqno. 330e870a8efSIlpo Järvinen */ 331e870a8efSIlpo Järvinen static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags) 332e870a8efSIlpo Järvinen { 333e870a8efSIlpo Järvinen skb->csum = 0; 334e870a8efSIlpo Järvinen 335e870a8efSIlpo Järvinen TCP_SKB_CB(skb)->flags = flags; 336e870a8efSIlpo Järvinen TCP_SKB_CB(skb)->sacked = 0; 337e870a8efSIlpo Järvinen 338e870a8efSIlpo Järvinen skb_shinfo(skb)->gso_segs = 1; 339e870a8efSIlpo Järvinen skb_shinfo(skb)->gso_size = 0; 340e870a8efSIlpo Järvinen skb_shinfo(skb)->gso_type = 0; 341e870a8efSIlpo Järvinen 342e870a8efSIlpo Järvinen TCP_SKB_CB(skb)->seq = seq; 343e870a8efSIlpo Järvinen if (flags & (TCPCB_FLAG_SYN | TCPCB_FLAG_FIN)) 344e870a8efSIlpo Järvinen seq++; 345e870a8efSIlpo Järvinen TCP_SKB_CB(skb)->end_seq = seq; 346e870a8efSIlpo Järvinen } 347e870a8efSIlpo Järvinen 34833f5f57eSIlpo Järvinen static inline int tcp_urg_mode(const struct tcp_sock *tp) 34933f5f57eSIlpo Järvinen { 35033f5f57eSIlpo Järvinen return tp->snd_una != tp->snd_up; 35133f5f57eSIlpo Järvinen } 35233f5f57eSIlpo Järvinen 35333ad798cSAdam Langley #define OPTION_SACK_ADVERTISE (1 << 0) 35433ad798cSAdam Langley #define OPTION_TS (1 << 1) 35533ad798cSAdam Langley #define OPTION_MD5 (1 << 2) 35633ad798cSAdam Langley 35733ad798cSAdam Langley struct tcp_out_options { 35833ad798cSAdam Langley u8 options; /* bit field of OPTION_* */ 35933ad798cSAdam Langley u8 ws; /* window scale, 0 to disable */ 36033ad798cSAdam Langley u8 num_sack_blocks; /* number of SACK blocks to include */ 36133ad798cSAdam Langley u16 mss; /* 0 to disable */ 36233ad798cSAdam Langley __u32 tsval, tsecr; /* need to include OPTION_TS */ 36333ad798cSAdam Langley }; 36433ad798cSAdam Langley 365fd6149d3SIlpo Järvinen /* Beware: Something in the Internet is very sensitive to the ordering of 366fd6149d3SIlpo Järvinen * TCP options, we learned this through the hard way, so be careful here. 367fd6149d3SIlpo Järvinen * Luckily we can at least blame others for their non-compliance but from 368fd6149d3SIlpo Järvinen * inter-operatibility perspective it seems that we're somewhat stuck with 369fd6149d3SIlpo Järvinen * the ordering which we have been using if we want to keep working with 370fd6149d3SIlpo Järvinen * those broken things (not that it currently hurts anybody as there isn't 371fd6149d3SIlpo Järvinen * particular reason why the ordering would need to be changed). 372fd6149d3SIlpo Järvinen * 373fd6149d3SIlpo Järvinen * At least SACK_PERM as the first option is known to lead to a disaster 374fd6149d3SIlpo Järvinen * (but it may well be that other scenarios fail similarly). 375fd6149d3SIlpo Järvinen */ 37633ad798cSAdam Langley static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp, 37733ad798cSAdam Langley const struct tcp_out_options *opts, 37833ad798cSAdam Langley __u8 **md5_hash) { 37933ad798cSAdam Langley if (unlikely(OPTION_MD5 & opts->options)) { 38033ad798cSAdam Langley *ptr++ = htonl((TCPOPT_NOP << 24) | 38133ad798cSAdam Langley (TCPOPT_NOP << 16) | 38233ad798cSAdam Langley (TCPOPT_MD5SIG << 8) | 38333ad798cSAdam Langley TCPOLEN_MD5SIG); 38433ad798cSAdam Langley *md5_hash = (__u8 *)ptr; 38533ad798cSAdam Langley ptr += 4; 38633ad798cSAdam Langley } else { 38733ad798cSAdam Langley *md5_hash = NULL; 38833ad798cSAdam Langley } 38933ad798cSAdam Langley 390fd6149d3SIlpo Järvinen if (unlikely(opts->mss)) { 391fd6149d3SIlpo Järvinen *ptr++ = htonl((TCPOPT_MSS << 24) | 392fd6149d3SIlpo Järvinen (TCPOLEN_MSS << 16) | 393fd6149d3SIlpo Järvinen opts->mss); 394fd6149d3SIlpo Järvinen } 395fd6149d3SIlpo Järvinen 39633ad798cSAdam Langley if (likely(OPTION_TS & opts->options)) { 39733ad798cSAdam Langley if (unlikely(OPTION_SACK_ADVERTISE & opts->options)) { 39833ad798cSAdam Langley *ptr++ = htonl((TCPOPT_SACK_PERM << 24) | 39933ad798cSAdam Langley (TCPOLEN_SACK_PERM << 16) | 40033ad798cSAdam Langley (TCPOPT_TIMESTAMP << 8) | 40133ad798cSAdam Langley TCPOLEN_TIMESTAMP); 40233ad798cSAdam Langley } else { 403496c98dfSYOSHIFUJI Hideaki *ptr++ = htonl((TCPOPT_NOP << 24) | 40440efc6faSStephen Hemminger (TCPOPT_NOP << 16) | 40540efc6faSStephen Hemminger (TCPOPT_TIMESTAMP << 8) | 40640efc6faSStephen Hemminger TCPOLEN_TIMESTAMP); 40740efc6faSStephen Hemminger } 40833ad798cSAdam Langley *ptr++ = htonl(opts->tsval); 40933ad798cSAdam Langley *ptr++ = htonl(opts->tsecr); 41033ad798cSAdam Langley } 41133ad798cSAdam Langley 41233ad798cSAdam Langley if (unlikely(OPTION_SACK_ADVERTISE & opts->options && 41333ad798cSAdam Langley !(OPTION_TS & opts->options))) { 41433ad798cSAdam Langley *ptr++ = htonl((TCPOPT_NOP << 24) | 41533ad798cSAdam Langley (TCPOPT_NOP << 16) | 41633ad798cSAdam Langley (TCPOPT_SACK_PERM << 8) | 41733ad798cSAdam Langley TCPOLEN_SACK_PERM); 41833ad798cSAdam Langley } 41933ad798cSAdam Langley 42033ad798cSAdam Langley if (unlikely(opts->ws)) { 42133ad798cSAdam Langley *ptr++ = htonl((TCPOPT_NOP << 24) | 42233ad798cSAdam Langley (TCPOPT_WINDOW << 16) | 42333ad798cSAdam Langley (TCPOLEN_WINDOW << 8) | 42433ad798cSAdam Langley opts->ws); 42533ad798cSAdam Langley } 42633ad798cSAdam Langley 42733ad798cSAdam Langley if (unlikely(opts->num_sack_blocks)) { 42833ad798cSAdam Langley struct tcp_sack_block *sp = tp->rx_opt.dsack ? 42933ad798cSAdam Langley tp->duplicate_sack : tp->selective_acks; 43040efc6faSStephen Hemminger int this_sack; 43140efc6faSStephen Hemminger 43240efc6faSStephen Hemminger *ptr++ = htonl((TCPOPT_NOP << 24) | 43340efc6faSStephen Hemminger (TCPOPT_NOP << 16) | 43440efc6faSStephen Hemminger (TCPOPT_SACK << 8) | 43533ad798cSAdam Langley (TCPOLEN_SACK_BASE + (opts->num_sack_blocks * 43640efc6faSStephen Hemminger TCPOLEN_SACK_PERBLOCK))); 4372de979bdSStephen Hemminger 43833ad798cSAdam Langley for (this_sack = 0; this_sack < opts->num_sack_blocks; 43933ad798cSAdam Langley ++this_sack) { 44040efc6faSStephen Hemminger *ptr++ = htonl(sp[this_sack].start_seq); 44140efc6faSStephen Hemminger *ptr++ = htonl(sp[this_sack].end_seq); 44240efc6faSStephen Hemminger } 4432de979bdSStephen Hemminger 44440efc6faSStephen Hemminger tp->rx_opt.dsack = 0; 44540efc6faSStephen Hemminger } 44640efc6faSStephen Hemminger } 44740efc6faSStephen Hemminger 44833ad798cSAdam Langley static unsigned tcp_syn_options(struct sock *sk, struct sk_buff *skb, 44933ad798cSAdam Langley struct tcp_out_options *opts, 45033ad798cSAdam Langley struct tcp_md5sig_key **md5) { 45133ad798cSAdam Langley struct tcp_sock *tp = tcp_sk(sk); 45233ad798cSAdam Langley unsigned size = 0; 45333ad798cSAdam Langley 454cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 45533ad798cSAdam Langley *md5 = tp->af_specific->md5_lookup(sk, sk); 45633ad798cSAdam Langley if (*md5) { 45733ad798cSAdam Langley opts->options |= OPTION_MD5; 45833ad798cSAdam Langley size += TCPOLEN_MD5SIG_ALIGNED; 459cfb6eeb4SYOSHIFUJI Hideaki } 46033ad798cSAdam Langley #else 46133ad798cSAdam Langley *md5 = NULL; 462cfb6eeb4SYOSHIFUJI Hideaki #endif 46333ad798cSAdam Langley 46433ad798cSAdam Langley /* We always get an MSS option. The option bytes which will be seen in 46533ad798cSAdam Langley * normal data packets should timestamps be used, must be in the MSS 46633ad798cSAdam Langley * advertised. But we subtract them from tp->mss_cache so that 46733ad798cSAdam Langley * calculations in tcp_sendmsg are simpler etc. So account for this 46833ad798cSAdam Langley * fact here if necessary. If we don't do this correctly, as a 46933ad798cSAdam Langley * receiver we won't recognize data packets as being full sized when we 47033ad798cSAdam Langley * should, and thus we won't abide by the delayed ACK rules correctly. 47133ad798cSAdam Langley * SACKs don't matter, we never delay an ACK when we have any of those 47233ad798cSAdam Langley * going out. */ 47333ad798cSAdam Langley opts->mss = tcp_advertise_mss(sk); 47433ad798cSAdam Langley size += TCPOLEN_MSS_ALIGNED; 47533ad798cSAdam Langley 47633ad798cSAdam Langley if (likely(sysctl_tcp_timestamps && *md5 == NULL)) { 47733ad798cSAdam Langley opts->options |= OPTION_TS; 47833ad798cSAdam Langley opts->tsval = TCP_SKB_CB(skb)->when; 47933ad798cSAdam Langley opts->tsecr = tp->rx_opt.ts_recent; 48033ad798cSAdam Langley size += TCPOLEN_TSTAMP_ALIGNED; 48133ad798cSAdam Langley } 48233ad798cSAdam Langley if (likely(sysctl_tcp_window_scaling)) { 48333ad798cSAdam Langley opts->ws = tp->rx_opt.rcv_wscale; 4847982d5e1SPhilip Love if (likely(opts->ws)) 48533ad798cSAdam Langley size += TCPOLEN_WSCALE_ALIGNED; 48633ad798cSAdam Langley } 48733ad798cSAdam Langley if (likely(sysctl_tcp_sack)) { 48833ad798cSAdam Langley opts->options |= OPTION_SACK_ADVERTISE; 489b32d1310SDavid S. Miller if (unlikely(!(OPTION_TS & opts->options))) 49033ad798cSAdam Langley size += TCPOLEN_SACKPERM_ALIGNED; 49133ad798cSAdam Langley } 49233ad798cSAdam Langley 49333ad798cSAdam Langley return size; 49433ad798cSAdam Langley } 49533ad798cSAdam Langley 49633ad798cSAdam Langley static unsigned tcp_synack_options(struct sock *sk, 49733ad798cSAdam Langley struct request_sock *req, 49833ad798cSAdam Langley unsigned mss, struct sk_buff *skb, 49933ad798cSAdam Langley struct tcp_out_options *opts, 50033ad798cSAdam Langley struct tcp_md5sig_key **md5) { 50133ad798cSAdam Langley unsigned size = 0; 50233ad798cSAdam Langley struct inet_request_sock *ireq = inet_rsk(req); 50333ad798cSAdam Langley char doing_ts; 50433ad798cSAdam Langley 50533ad798cSAdam Langley #ifdef CONFIG_TCP_MD5SIG 50633ad798cSAdam Langley *md5 = tcp_rsk(req)->af_specific->md5_lookup(sk, req); 50733ad798cSAdam Langley if (*md5) { 50833ad798cSAdam Langley opts->options |= OPTION_MD5; 50933ad798cSAdam Langley size += TCPOLEN_MD5SIG_ALIGNED; 51033ad798cSAdam Langley } 51133ad798cSAdam Langley #else 51233ad798cSAdam Langley *md5 = NULL; 51333ad798cSAdam Langley #endif 51433ad798cSAdam Langley 51533ad798cSAdam Langley /* we can't fit any SACK blocks in a packet with MD5 + TS 51633ad798cSAdam Langley options. There was discussion about disabling SACK rather than TS in 51733ad798cSAdam Langley order to fit in better with old, buggy kernels, but that was deemed 51833ad798cSAdam Langley to be unnecessary. */ 51933ad798cSAdam Langley doing_ts = ireq->tstamp_ok && !(*md5 && ireq->sack_ok); 52033ad798cSAdam Langley 52133ad798cSAdam Langley opts->mss = mss; 52233ad798cSAdam Langley size += TCPOLEN_MSS_ALIGNED; 52333ad798cSAdam Langley 52433ad798cSAdam Langley if (likely(ireq->wscale_ok)) { 52533ad798cSAdam Langley opts->ws = ireq->rcv_wscale; 5267982d5e1SPhilip Love if (likely(opts->ws)) 52733ad798cSAdam Langley size += TCPOLEN_WSCALE_ALIGNED; 52833ad798cSAdam Langley } 52933ad798cSAdam Langley if (likely(doing_ts)) { 53033ad798cSAdam Langley opts->options |= OPTION_TS; 53133ad798cSAdam Langley opts->tsval = TCP_SKB_CB(skb)->when; 53233ad798cSAdam Langley opts->tsecr = req->ts_recent; 53333ad798cSAdam Langley size += TCPOLEN_TSTAMP_ALIGNED; 53433ad798cSAdam Langley } 53533ad798cSAdam Langley if (likely(ireq->sack_ok)) { 53633ad798cSAdam Langley opts->options |= OPTION_SACK_ADVERTISE; 53733ad798cSAdam Langley if (unlikely(!doing_ts)) 53833ad798cSAdam Langley size += TCPOLEN_SACKPERM_ALIGNED; 53933ad798cSAdam Langley } 54033ad798cSAdam Langley 54133ad798cSAdam Langley return size; 54233ad798cSAdam Langley } 54333ad798cSAdam Langley 54433ad798cSAdam Langley static unsigned tcp_established_options(struct sock *sk, struct sk_buff *skb, 54533ad798cSAdam Langley struct tcp_out_options *opts, 54633ad798cSAdam Langley struct tcp_md5sig_key **md5) { 54733ad798cSAdam Langley struct tcp_skb_cb *tcb = skb ? TCP_SKB_CB(skb) : NULL; 54833ad798cSAdam Langley struct tcp_sock *tp = tcp_sk(sk); 54933ad798cSAdam Langley unsigned size = 0; 550cabeccbdSIlpo Järvinen unsigned int eff_sacks; 55133ad798cSAdam Langley 55233ad798cSAdam Langley #ifdef CONFIG_TCP_MD5SIG 55333ad798cSAdam Langley *md5 = tp->af_specific->md5_lookup(sk, sk); 55433ad798cSAdam Langley if (unlikely(*md5)) { 55533ad798cSAdam Langley opts->options |= OPTION_MD5; 55633ad798cSAdam Langley size += TCPOLEN_MD5SIG_ALIGNED; 55733ad798cSAdam Langley } 55833ad798cSAdam Langley #else 55933ad798cSAdam Langley *md5 = NULL; 56033ad798cSAdam Langley #endif 56133ad798cSAdam Langley 56233ad798cSAdam Langley if (likely(tp->rx_opt.tstamp_ok)) { 56333ad798cSAdam Langley opts->options |= OPTION_TS; 56433ad798cSAdam Langley opts->tsval = tcb ? tcb->when : 0; 56533ad798cSAdam Langley opts->tsecr = tp->rx_opt.ts_recent; 56633ad798cSAdam Langley size += TCPOLEN_TSTAMP_ALIGNED; 56733ad798cSAdam Langley } 56833ad798cSAdam Langley 569cabeccbdSIlpo Järvinen eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack; 570cabeccbdSIlpo Järvinen if (unlikely(eff_sacks)) { 57133ad798cSAdam Langley const unsigned remaining = MAX_TCP_OPTION_SPACE - size; 57233ad798cSAdam Langley opts->num_sack_blocks = 573cabeccbdSIlpo Järvinen min_t(unsigned, eff_sacks, 57433ad798cSAdam Langley (remaining - TCPOLEN_SACK_BASE_ALIGNED) / 57533ad798cSAdam Langley TCPOLEN_SACK_PERBLOCK); 57633ad798cSAdam Langley size += TCPOLEN_SACK_BASE_ALIGNED + 57733ad798cSAdam Langley opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK; 57833ad798cSAdam Langley } 57933ad798cSAdam Langley 58033ad798cSAdam Langley return size; 58140efc6faSStephen Hemminger } 5821da177e4SLinus Torvalds 5831da177e4SLinus Torvalds /* This routine actually transmits TCP packets queued in by 5841da177e4SLinus Torvalds * tcp_do_sendmsg(). This is used by both the initial 5851da177e4SLinus Torvalds * transmission and possible later retransmissions. 5861da177e4SLinus Torvalds * All SKB's seen here are completely headerless. It is our 5871da177e4SLinus Torvalds * job to build the TCP header, and pass the packet down to 5881da177e4SLinus Torvalds * IP so it can do the same plus pass the packet off to the 5891da177e4SLinus Torvalds * device. 5901da177e4SLinus Torvalds * 5911da177e4SLinus Torvalds * We are working here with either a clone of the original 5921da177e4SLinus Torvalds * SKB, or a fresh unique copy made by the retransmit engine. 5931da177e4SLinus Torvalds */ 594056834d9SIlpo Järvinen static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it, 595056834d9SIlpo Järvinen gfp_t gfp_mask) 5961da177e4SLinus Torvalds { 5976687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 598dfb4b9dcSDavid S. Miller struct inet_sock *inet; 599dfb4b9dcSDavid S. Miller struct tcp_sock *tp; 600dfb4b9dcSDavid S. Miller struct tcp_skb_cb *tcb; 60133ad798cSAdam Langley struct tcp_out_options opts; 60233ad798cSAdam Langley unsigned tcp_options_size, tcp_header_size; 603cfb6eeb4SYOSHIFUJI Hideaki struct tcp_md5sig_key *md5; 604cfb6eeb4SYOSHIFUJI Hideaki __u8 *md5_hash_location; 6051da177e4SLinus Torvalds struct tcphdr *th; 6061da177e4SLinus Torvalds int err; 6071da177e4SLinus Torvalds 608dfb4b9dcSDavid S. Miller BUG_ON(!skb || !tcp_skb_pcount(skb)); 609dfb4b9dcSDavid S. Miller 610dfb4b9dcSDavid S. Miller /* If congestion control is doing timestamping, we must 611dfb4b9dcSDavid S. Miller * take such a timestamp before we potentially clone/copy. 612dfb4b9dcSDavid S. Miller */ 613164891aaSStephen Hemminger if (icsk->icsk_ca_ops->flags & TCP_CONG_RTT_STAMP) 614dfb4b9dcSDavid S. Miller __net_timestamp(skb); 615dfb4b9dcSDavid S. Miller 616dfb4b9dcSDavid S. Miller if (likely(clone_it)) { 617dfb4b9dcSDavid S. Miller if (unlikely(skb_cloned(skb))) 618dfb4b9dcSDavid S. Miller skb = pskb_copy(skb, gfp_mask); 619dfb4b9dcSDavid S. Miller else 620dfb4b9dcSDavid S. Miller skb = skb_clone(skb, gfp_mask); 621dfb4b9dcSDavid S. Miller if (unlikely(!skb)) 622dfb4b9dcSDavid S. Miller return -ENOBUFS; 623dfb4b9dcSDavid S. Miller } 624dfb4b9dcSDavid S. Miller 625dfb4b9dcSDavid S. Miller inet = inet_sk(sk); 626dfb4b9dcSDavid S. Miller tp = tcp_sk(sk); 627dfb4b9dcSDavid S. Miller tcb = TCP_SKB_CB(skb); 62833ad798cSAdam Langley memset(&opts, 0, sizeof(opts)); 6291da177e4SLinus Torvalds 63033ad798cSAdam Langley if (unlikely(tcb->flags & TCPCB_FLAG_SYN)) 63133ad798cSAdam Langley tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5); 63233ad798cSAdam Langley else 63333ad798cSAdam Langley tcp_options_size = tcp_established_options(sk, skb, &opts, 63433ad798cSAdam Langley &md5); 63533ad798cSAdam Langley tcp_header_size = tcp_options_size + sizeof(struct tcphdr); 6361da177e4SLinus Torvalds 637317a76f9SStephen Hemminger if (tcp_packets_in_flight(tp) == 0) 6386687e988SArnaldo Carvalho de Melo tcp_ca_event(sk, CA_EVENT_TX_START); 6391da177e4SLinus Torvalds 640aa8223c7SArnaldo Carvalho de Melo skb_push(skb, tcp_header_size); 641aa8223c7SArnaldo Carvalho de Melo skb_reset_transport_header(skb); 642e89862f4SDavid S. Miller skb_set_owner_w(skb, sk); 6431da177e4SLinus Torvalds 6441da177e4SLinus Torvalds /* Build TCP header and checksum it. */ 645aa8223c7SArnaldo Carvalho de Melo th = tcp_hdr(skb); 6461da177e4SLinus Torvalds th->source = inet->sport; 6471da177e4SLinus Torvalds th->dest = inet->dport; 6481da177e4SLinus Torvalds th->seq = htonl(tcb->seq); 6491da177e4SLinus Torvalds th->ack_seq = htonl(tp->rcv_nxt); 650df7a3b07SAl Viro *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) | 651dfb4b9dcSDavid S. Miller tcb->flags); 652dfb4b9dcSDavid S. Miller 653dfb4b9dcSDavid S. Miller if (unlikely(tcb->flags & TCPCB_FLAG_SYN)) { 6541da177e4SLinus Torvalds /* RFC1323: The window in SYN & SYN/ACK segments 6551da177e4SLinus Torvalds * is never scaled. 6561da177e4SLinus Torvalds */ 657600ff0c2SIlpo Järvinen th->window = htons(min(tp->rcv_wnd, 65535U)); 6581da177e4SLinus Torvalds } else { 6591da177e4SLinus Torvalds th->window = htons(tcp_select_window(sk)); 6601da177e4SLinus Torvalds } 6611da177e4SLinus Torvalds th->check = 0; 6621da177e4SLinus Torvalds th->urg_ptr = 0; 6631da177e4SLinus Torvalds 66433f5f57eSIlpo Järvinen /* The urg_mode check is necessary during a below snd_una win probe */ 6657691367dSHerbert Xu if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) { 6667691367dSHerbert Xu if (before(tp->snd_up, tcb->seq + 0x10000)) { 6671da177e4SLinus Torvalds th->urg_ptr = htons(tp->snd_up - tcb->seq); 6681da177e4SLinus Torvalds th->urg = 1; 6697691367dSHerbert Xu } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) { 6707691367dSHerbert Xu th->urg_ptr = 0xFFFF; 6717691367dSHerbert Xu th->urg = 1; 6727691367dSHerbert Xu } 6731da177e4SLinus Torvalds } 6741da177e4SLinus Torvalds 67533ad798cSAdam Langley tcp_options_write((__be32 *)(th + 1), tp, &opts, &md5_hash_location); 67633ad798cSAdam Langley if (likely((tcb->flags & TCPCB_FLAG_SYN) == 0)) 6779e412ba7SIlpo Järvinen TCP_ECN_send(sk, skb, tcp_header_size); 678dfb4b9dcSDavid S. Miller 679cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 680cfb6eeb4SYOSHIFUJI Hideaki /* Calculate the MD5 hash, as we have all we need now */ 681cfb6eeb4SYOSHIFUJI Hideaki if (md5) { 68233ad798cSAdam Langley sk->sk_route_caps &= ~NETIF_F_GSO_MASK; 683cfb6eeb4SYOSHIFUJI Hideaki tp->af_specific->calc_md5_hash(md5_hash_location, 68449a72dfbSAdam Langley md5, sk, NULL, skb); 685cfb6eeb4SYOSHIFUJI Hideaki } 686cfb6eeb4SYOSHIFUJI Hideaki #endif 687cfb6eeb4SYOSHIFUJI Hideaki 6888292a17aSArnaldo Carvalho de Melo icsk->icsk_af_ops->send_check(sk, skb->len, skb); 6891da177e4SLinus Torvalds 690dfb4b9dcSDavid S. Miller if (likely(tcb->flags & TCPCB_FLAG_ACK)) 691fc6415bcSDavid S. Miller tcp_event_ack_sent(sk, tcp_skb_pcount(skb)); 6921da177e4SLinus Torvalds 6931da177e4SLinus Torvalds if (skb->len != tcp_header_size) 6941da177e4SLinus Torvalds tcp_event_data_sent(tp, skb, sk); 6951da177e4SLinus Torvalds 696bd37a088SWei Yongjun if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq) 69781cc8a75SPavel Emelyanov TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTSEGS); 6981da177e4SLinus Torvalds 699e89862f4SDavid S. Miller err = icsk->icsk_af_ops->queue_xmit(skb, 0); 70083de47cdSHua Zhong if (likely(err <= 0)) 7011da177e4SLinus Torvalds return err; 7021da177e4SLinus Torvalds 7033cfe3baaSIlpo Järvinen tcp_enter_cwr(sk, 1); 7041da177e4SLinus Torvalds 705b9df3cb8SGerrit Renker return net_xmit_eval(err); 7061da177e4SLinus Torvalds } 7071da177e4SLinus Torvalds 7081da177e4SLinus Torvalds /* This routine just queue's the buffer 7091da177e4SLinus Torvalds * 7101da177e4SLinus Torvalds * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames, 7111da177e4SLinus Torvalds * otherwise socket can stall. 7121da177e4SLinus Torvalds */ 7131da177e4SLinus Torvalds static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb) 7141da177e4SLinus Torvalds { 7151da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 7161da177e4SLinus Torvalds 7171da177e4SLinus Torvalds /* Advance write_seq and place onto the write_queue. */ 7181da177e4SLinus Torvalds tp->write_seq = TCP_SKB_CB(skb)->end_seq; 7191da177e4SLinus Torvalds skb_header_release(skb); 720fe067e8aSDavid S. Miller tcp_add_write_queue_tail(sk, skb); 7213ab224beSHideo Aoki sk->sk_wmem_queued += skb->truesize; 7223ab224beSHideo Aoki sk_mem_charge(sk, skb->truesize); 7231da177e4SLinus Torvalds } 7241da177e4SLinus Torvalds 725056834d9SIlpo Järvinen static void tcp_set_skb_tso_segs(struct sock *sk, struct sk_buff *skb, 726056834d9SIlpo Järvinen unsigned int mss_now) 727f6302d1dSDavid S. Miller { 728f8269a49SIlpo Järvinen if (skb->len <= mss_now || !sk_can_gso(sk)) { 729f6302d1dSDavid S. Miller /* Avoid the costly divide in the normal 730f6302d1dSDavid S. Miller * non-TSO case. 731f6302d1dSDavid S. Miller */ 7327967168cSHerbert Xu skb_shinfo(skb)->gso_segs = 1; 7337967168cSHerbert Xu skb_shinfo(skb)->gso_size = 0; 7347967168cSHerbert Xu skb_shinfo(skb)->gso_type = 0; 735f6302d1dSDavid S. Miller } else { 736356f89e1SIlpo Järvinen skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss_now); 7377967168cSHerbert Xu skb_shinfo(skb)->gso_size = mss_now; 738bcd76111SHerbert Xu skb_shinfo(skb)->gso_type = sk->sk_gso_type; 7391da177e4SLinus Torvalds } 7401da177e4SLinus Torvalds } 7411da177e4SLinus Torvalds 74291fed7a1SIlpo Järvinen /* When a modification to fackets out becomes necessary, we need to check 74368f8353bSIlpo Järvinen * skb is counted to fackets_out or not. 74491fed7a1SIlpo Järvinen */ 745a47e5a98SIlpo Järvinen static void tcp_adjust_fackets_out(struct sock *sk, struct sk_buff *skb, 74691fed7a1SIlpo Järvinen int decr) 74791fed7a1SIlpo Järvinen { 748a47e5a98SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 749a47e5a98SIlpo Järvinen 750dc86967bSIlpo Järvinen if (!tp->sacked_out || tcp_is_reno(tp)) 75191fed7a1SIlpo Järvinen return; 75291fed7a1SIlpo Järvinen 7536859d494SIlpo Järvinen if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq)) 75491fed7a1SIlpo Järvinen tp->fackets_out -= decr; 75591fed7a1SIlpo Järvinen } 75691fed7a1SIlpo Järvinen 757797108d1SIlpo Järvinen /* Pcount in the middle of the write queue got changed, we need to do various 758797108d1SIlpo Järvinen * tweaks to fix counters 759797108d1SIlpo Järvinen */ 760797108d1SIlpo Järvinen static void tcp_adjust_pcount(struct sock *sk, struct sk_buff *skb, int decr) 761797108d1SIlpo Järvinen { 762797108d1SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 763797108d1SIlpo Järvinen 764797108d1SIlpo Järvinen tp->packets_out -= decr; 765797108d1SIlpo Järvinen 766797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 767797108d1SIlpo Järvinen tp->sacked_out -= decr; 768797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) 769797108d1SIlpo Järvinen tp->retrans_out -= decr; 770797108d1SIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST) 771797108d1SIlpo Järvinen tp->lost_out -= decr; 772797108d1SIlpo Järvinen 773797108d1SIlpo Järvinen /* Reno case is special. Sigh... */ 774797108d1SIlpo Järvinen if (tcp_is_reno(tp) && decr > 0) 775797108d1SIlpo Järvinen tp->sacked_out -= min_t(u32, tp->sacked_out, decr); 776797108d1SIlpo Järvinen 777797108d1SIlpo Järvinen tcp_adjust_fackets_out(sk, skb, decr); 778797108d1SIlpo Järvinen 779797108d1SIlpo Järvinen if (tp->lost_skb_hint && 780797108d1SIlpo Järvinen before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) && 781*52cf3cc8SIlpo Järvinen (tcp_is_fack(tp) || (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))) 782797108d1SIlpo Järvinen tp->lost_cnt_hint -= decr; 783797108d1SIlpo Järvinen 784797108d1SIlpo Järvinen tcp_verify_left_out(tp); 785797108d1SIlpo Järvinen } 786797108d1SIlpo Järvinen 7871da177e4SLinus Torvalds /* Function to create two new TCP segments. Shrinks the given segment 7881da177e4SLinus Torvalds * to the specified size and appends a new segment with the rest of the 7891da177e4SLinus Torvalds * packet to the list. This won't be called frequently, I hope. 7901da177e4SLinus Torvalds * Remember, these are still headerless SKBs at this point. 7911da177e4SLinus Torvalds */ 792056834d9SIlpo Järvinen int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len, 793056834d9SIlpo Järvinen unsigned int mss_now) 7941da177e4SLinus Torvalds { 7951da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 7961da177e4SLinus Torvalds struct sk_buff *buff; 7976475be16SDavid S. Miller int nsize, old_factor; 798b60b49eaSHerbert Xu int nlen; 7999ce01461SIlpo Järvinen u8 flags; 8001da177e4SLinus Torvalds 801b2cc99f0SHerbert Xu BUG_ON(len > skb->len); 8026a438bbeSStephen Hemminger 8031da177e4SLinus Torvalds nsize = skb_headlen(skb) - len; 8041da177e4SLinus Torvalds if (nsize < 0) 8051da177e4SLinus Torvalds nsize = 0; 8061da177e4SLinus Torvalds 8071da177e4SLinus Torvalds if (skb_cloned(skb) && 8081da177e4SLinus Torvalds skb_is_nonlinear(skb) && 8091da177e4SLinus Torvalds pskb_expand_head(skb, 0, 0, GFP_ATOMIC)) 8101da177e4SLinus Torvalds return -ENOMEM; 8111da177e4SLinus Torvalds 8121da177e4SLinus Torvalds /* Get a new skb... force flag on. */ 8131da177e4SLinus Torvalds buff = sk_stream_alloc_skb(sk, nsize, GFP_ATOMIC); 8141da177e4SLinus Torvalds if (buff == NULL) 8151da177e4SLinus Torvalds return -ENOMEM; /* We'll just try again later. */ 816ef5cb973SHerbert Xu 8173ab224beSHideo Aoki sk->sk_wmem_queued += buff->truesize; 8183ab224beSHideo Aoki sk_mem_charge(sk, buff->truesize); 819b60b49eaSHerbert Xu nlen = skb->len - len - nsize; 820b60b49eaSHerbert Xu buff->truesize += nlen; 821b60b49eaSHerbert Xu skb->truesize -= nlen; 8221da177e4SLinus Torvalds 8231da177e4SLinus Torvalds /* Correct the sequence numbers. */ 8241da177e4SLinus Torvalds TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; 8251da177e4SLinus Torvalds TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; 8261da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; 8271da177e4SLinus Torvalds 8281da177e4SLinus Torvalds /* PSH and FIN should only be set in the second packet. */ 8291da177e4SLinus Torvalds flags = TCP_SKB_CB(skb)->flags; 8301da177e4SLinus Torvalds TCP_SKB_CB(skb)->flags = flags & ~(TCPCB_FLAG_FIN | TCPCB_FLAG_PSH); 8311da177e4SLinus Torvalds TCP_SKB_CB(buff)->flags = flags; 832e14c3cafSHerbert Xu TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked; 8331da177e4SLinus Torvalds 83484fa7933SPatrick McHardy if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) { 8351da177e4SLinus Torvalds /* Copy and checksum data tail into the new buffer. */ 836056834d9SIlpo Järvinen buff->csum = csum_partial_copy_nocheck(skb->data + len, 837056834d9SIlpo Järvinen skb_put(buff, nsize), 8381da177e4SLinus Torvalds nsize, 0); 8391da177e4SLinus Torvalds 8401da177e4SLinus Torvalds skb_trim(skb, len); 8411da177e4SLinus Torvalds 8421da177e4SLinus Torvalds skb->csum = csum_block_sub(skb->csum, buff->csum, len); 8431da177e4SLinus Torvalds } else { 84484fa7933SPatrick McHardy skb->ip_summed = CHECKSUM_PARTIAL; 8451da177e4SLinus Torvalds skb_split(skb, buff, len); 8461da177e4SLinus Torvalds } 8471da177e4SLinus Torvalds 8481da177e4SLinus Torvalds buff->ip_summed = skb->ip_summed; 8491da177e4SLinus Torvalds 8501da177e4SLinus Torvalds /* Looks stupid, but our code really uses when of 8511da177e4SLinus Torvalds * skbs, which it never sent before. --ANK 8521da177e4SLinus Torvalds */ 8531da177e4SLinus Torvalds TCP_SKB_CB(buff)->when = TCP_SKB_CB(skb)->when; 854a61bbcf2SPatrick McHardy buff->tstamp = skb->tstamp; 8551da177e4SLinus Torvalds 8566475be16SDavid S. Miller old_factor = tcp_skb_pcount(skb); 8576475be16SDavid S. Miller 8581da177e4SLinus Torvalds /* Fix up tso_factor for both original and new SKB. */ 859846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, skb, mss_now); 860846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, buff, mss_now); 8611da177e4SLinus Torvalds 8626475be16SDavid S. Miller /* If this packet has been sent out already, we must 8636475be16SDavid S. Miller * adjust the various packet counters. 8646475be16SDavid S. Miller */ 865cf0b450cSHerbert Xu if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) { 8666475be16SDavid S. Miller int diff = old_factor - tcp_skb_pcount(skb) - 8676475be16SDavid S. Miller tcp_skb_pcount(buff); 8681da177e4SLinus Torvalds 869797108d1SIlpo Järvinen if (diff) 870797108d1SIlpo Järvinen tcp_adjust_pcount(sk, skb, diff); 8711da177e4SLinus Torvalds } 8721da177e4SLinus Torvalds 8731da177e4SLinus Torvalds /* Link BUFF into the send queue. */ 874f44b5271SDavid S. Miller skb_header_release(buff); 875fe067e8aSDavid S. Miller tcp_insert_write_queue_after(skb, buff, sk); 8761da177e4SLinus Torvalds 8771da177e4SLinus Torvalds return 0; 8781da177e4SLinus Torvalds } 8791da177e4SLinus Torvalds 8801da177e4SLinus Torvalds /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c 8811da177e4SLinus Torvalds * eventually). The difference is that pulled data not copied, but 8821da177e4SLinus Torvalds * immediately discarded. 8831da177e4SLinus Torvalds */ 884f2911969SHerbert Xu ~{PmVHI~} static void __pskb_trim_head(struct sk_buff *skb, int len) 8851da177e4SLinus Torvalds { 8861da177e4SLinus Torvalds int i, k, eat; 8871da177e4SLinus Torvalds 8881da177e4SLinus Torvalds eat = len; 8891da177e4SLinus Torvalds k = 0; 8901da177e4SLinus Torvalds for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) { 8911da177e4SLinus Torvalds if (skb_shinfo(skb)->frags[i].size <= eat) { 8921da177e4SLinus Torvalds put_page(skb_shinfo(skb)->frags[i].page); 8931da177e4SLinus Torvalds eat -= skb_shinfo(skb)->frags[i].size; 8941da177e4SLinus Torvalds } else { 8951da177e4SLinus Torvalds skb_shinfo(skb)->frags[k] = skb_shinfo(skb)->frags[i]; 8961da177e4SLinus Torvalds if (eat) { 8971da177e4SLinus Torvalds skb_shinfo(skb)->frags[k].page_offset += eat; 8981da177e4SLinus Torvalds skb_shinfo(skb)->frags[k].size -= eat; 8991da177e4SLinus Torvalds eat = 0; 9001da177e4SLinus Torvalds } 9011da177e4SLinus Torvalds k++; 9021da177e4SLinus Torvalds } 9031da177e4SLinus Torvalds } 9041da177e4SLinus Torvalds skb_shinfo(skb)->nr_frags = k; 9051da177e4SLinus Torvalds 90627a884dcSArnaldo Carvalho de Melo skb_reset_tail_pointer(skb); 9071da177e4SLinus Torvalds skb->data_len -= len; 9081da177e4SLinus Torvalds skb->len = skb->data_len; 9091da177e4SLinus Torvalds } 9101da177e4SLinus Torvalds 9111da177e4SLinus Torvalds int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len) 9121da177e4SLinus Torvalds { 913056834d9SIlpo Järvinen if (skb_cloned(skb) && pskb_expand_head(skb, 0, 0, GFP_ATOMIC)) 9141da177e4SLinus Torvalds return -ENOMEM; 9151da177e4SLinus Torvalds 916f2911969SHerbert Xu ~{PmVHI~} /* If len == headlen, we avoid __skb_pull to preserve alignment. */ 917f2911969SHerbert Xu ~{PmVHI~} if (unlikely(len < skb_headlen(skb))) 9181da177e4SLinus Torvalds __skb_pull(skb, len); 919f2911969SHerbert Xu ~{PmVHI~} else 920f2911969SHerbert Xu ~{PmVHI~} __pskb_trim_head(skb, len - skb_headlen(skb)); 9211da177e4SLinus Torvalds 9221da177e4SLinus Torvalds TCP_SKB_CB(skb)->seq += len; 92384fa7933SPatrick McHardy skb->ip_summed = CHECKSUM_PARTIAL; 9241da177e4SLinus Torvalds 9251da177e4SLinus Torvalds skb->truesize -= len; 9261da177e4SLinus Torvalds sk->sk_wmem_queued -= len; 9273ab224beSHideo Aoki sk_mem_uncharge(sk, len); 9281da177e4SLinus Torvalds sock_set_flag(sk, SOCK_QUEUE_SHRUNK); 9291da177e4SLinus Torvalds 9301da177e4SLinus Torvalds /* Any change of skb->len requires recalculation of tso 9311da177e4SLinus Torvalds * factor and mss. 9321da177e4SLinus Torvalds */ 9331da177e4SLinus Torvalds if (tcp_skb_pcount(skb) > 1) 9340c54b85fSIlpo Järvinen tcp_set_skb_tso_segs(sk, skb, tcp_current_mss(sk)); 9351da177e4SLinus Torvalds 9361da177e4SLinus Torvalds return 0; 9371da177e4SLinus Torvalds } 9381da177e4SLinus Torvalds 9395d424d5aSJohn Heffner /* Not accounting for SACKs here. */ 9405d424d5aSJohn Heffner int tcp_mtu_to_mss(struct sock *sk, int pmtu) 9415d424d5aSJohn Heffner { 9425d424d5aSJohn Heffner struct tcp_sock *tp = tcp_sk(sk); 9435d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 9445d424d5aSJohn Heffner int mss_now; 9455d424d5aSJohn Heffner 9465d424d5aSJohn Heffner /* Calculate base mss without TCP options: 9475d424d5aSJohn Heffner It is MMS_S - sizeof(tcphdr) of rfc1122 9485d424d5aSJohn Heffner */ 9495d424d5aSJohn Heffner mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr); 9505d424d5aSJohn Heffner 9515d424d5aSJohn Heffner /* Clamp it (mss_clamp does not include tcp options) */ 9525d424d5aSJohn Heffner if (mss_now > tp->rx_opt.mss_clamp) 9535d424d5aSJohn Heffner mss_now = tp->rx_opt.mss_clamp; 9545d424d5aSJohn Heffner 9555d424d5aSJohn Heffner /* Now subtract optional transport overhead */ 9565d424d5aSJohn Heffner mss_now -= icsk->icsk_ext_hdr_len; 9575d424d5aSJohn Heffner 9585d424d5aSJohn Heffner /* Then reserve room for full set of TCP options and 8 bytes of data */ 9595d424d5aSJohn Heffner if (mss_now < 48) 9605d424d5aSJohn Heffner mss_now = 48; 9615d424d5aSJohn Heffner 9625d424d5aSJohn Heffner /* Now subtract TCP options size, not including SACKs */ 9635d424d5aSJohn Heffner mss_now -= tp->tcp_header_len - sizeof(struct tcphdr); 9645d424d5aSJohn Heffner 9655d424d5aSJohn Heffner return mss_now; 9665d424d5aSJohn Heffner } 9675d424d5aSJohn Heffner 9685d424d5aSJohn Heffner /* Inverse of above */ 9695d424d5aSJohn Heffner int tcp_mss_to_mtu(struct sock *sk, int mss) 9705d424d5aSJohn Heffner { 9715d424d5aSJohn Heffner struct tcp_sock *tp = tcp_sk(sk); 9725d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 9735d424d5aSJohn Heffner int mtu; 9745d424d5aSJohn Heffner 9755d424d5aSJohn Heffner mtu = mss + 9765d424d5aSJohn Heffner tp->tcp_header_len + 9775d424d5aSJohn Heffner icsk->icsk_ext_hdr_len + 9785d424d5aSJohn Heffner icsk->icsk_af_ops->net_header_len; 9795d424d5aSJohn Heffner 9805d424d5aSJohn Heffner return mtu; 9815d424d5aSJohn Heffner } 9825d424d5aSJohn Heffner 9835d424d5aSJohn Heffner void tcp_mtup_init(struct sock *sk) 9845d424d5aSJohn Heffner { 9855d424d5aSJohn Heffner struct tcp_sock *tp = tcp_sk(sk); 9865d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 9875d424d5aSJohn Heffner 9885d424d5aSJohn Heffner icsk->icsk_mtup.enabled = sysctl_tcp_mtu_probing > 1; 9895d424d5aSJohn Heffner icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) + 9905d424d5aSJohn Heffner icsk->icsk_af_ops->net_header_len; 9915d424d5aSJohn Heffner icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, sysctl_tcp_base_mss); 9925d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = 0; 9935d424d5aSJohn Heffner } 9945d424d5aSJohn Heffner 9951da177e4SLinus Torvalds /* This function synchronize snd mss to current pmtu/exthdr set. 9961da177e4SLinus Torvalds 9971da177e4SLinus Torvalds tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts 9981da177e4SLinus Torvalds for TCP options, but includes only bare TCP header. 9991da177e4SLinus Torvalds 10001da177e4SLinus Torvalds tp->rx_opt.mss_clamp is mss negotiated at connection setup. 1001caa20d9aSStephen Hemminger It is minimum of user_mss and mss received with SYN. 10021da177e4SLinus Torvalds It also does not include TCP options. 10031da177e4SLinus Torvalds 1004d83d8461SArnaldo Carvalho de Melo inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function. 10051da177e4SLinus Torvalds 10061da177e4SLinus Torvalds tp->mss_cache is current effective sending mss, including 10071da177e4SLinus Torvalds all tcp options except for SACKs. It is evaluated, 10081da177e4SLinus Torvalds taking into account current pmtu, but never exceeds 10091da177e4SLinus Torvalds tp->rx_opt.mss_clamp. 10101da177e4SLinus Torvalds 10111da177e4SLinus Torvalds NOTE1. rfc1122 clearly states that advertised MSS 10121da177e4SLinus Torvalds DOES NOT include either tcp or ip options. 10131da177e4SLinus Torvalds 1014d83d8461SArnaldo Carvalho de Melo NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache 1015d83d8461SArnaldo Carvalho de Melo are READ ONLY outside this function. --ANK (980731) 10161da177e4SLinus Torvalds */ 10171da177e4SLinus Torvalds unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu) 10181da177e4SLinus Torvalds { 10191da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 1020d83d8461SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 10215d424d5aSJohn Heffner int mss_now; 10221da177e4SLinus Torvalds 10235d424d5aSJohn Heffner if (icsk->icsk_mtup.search_high > pmtu) 10245d424d5aSJohn Heffner icsk->icsk_mtup.search_high = pmtu; 10251da177e4SLinus Torvalds 10265d424d5aSJohn Heffner mss_now = tcp_mtu_to_mss(sk, pmtu); 1027409d22b4SIlpo Järvinen mss_now = tcp_bound_to_half_wnd(tp, mss_now); 10281da177e4SLinus Torvalds 10291da177e4SLinus Torvalds /* And store cached results */ 1030d83d8461SArnaldo Carvalho de Melo icsk->icsk_pmtu_cookie = pmtu; 10315d424d5aSJohn Heffner if (icsk->icsk_mtup.enabled) 10325d424d5aSJohn Heffner mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low)); 1033c1b4a7e6SDavid S. Miller tp->mss_cache = mss_now; 10341da177e4SLinus Torvalds 10351da177e4SLinus Torvalds return mss_now; 10361da177e4SLinus Torvalds } 10371da177e4SLinus Torvalds 10381da177e4SLinus Torvalds /* Compute the current effective MSS, taking SACKs and IP options, 10391da177e4SLinus Torvalds * and even PMTU discovery events into account. 10401da177e4SLinus Torvalds */ 10410c54b85fSIlpo Järvinen unsigned int tcp_current_mss(struct sock *sk) 10421da177e4SLinus Torvalds { 10431da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 10441da177e4SLinus Torvalds struct dst_entry *dst = __sk_dst_get(sk); 1045c1b4a7e6SDavid S. Miller u32 mss_now; 104633ad798cSAdam Langley unsigned header_len; 104733ad798cSAdam Langley struct tcp_out_options opts; 104833ad798cSAdam Langley struct tcp_md5sig_key *md5; 10491da177e4SLinus Torvalds 1050c1b4a7e6SDavid S. Miller mss_now = tp->mss_cache; 1051c1b4a7e6SDavid S. Miller 10521da177e4SLinus Torvalds if (dst) { 10531da177e4SLinus Torvalds u32 mtu = dst_mtu(dst); 1054d83d8461SArnaldo Carvalho de Melo if (mtu != inet_csk(sk)->icsk_pmtu_cookie) 10551da177e4SLinus Torvalds mss_now = tcp_sync_mss(sk, mtu); 10561da177e4SLinus Torvalds } 10571da177e4SLinus Torvalds 105833ad798cSAdam Langley header_len = tcp_established_options(sk, NULL, &opts, &md5) + 105933ad798cSAdam Langley sizeof(struct tcphdr); 106033ad798cSAdam Langley /* The mss_cache is sized based on tp->tcp_header_len, which assumes 106133ad798cSAdam Langley * some common options. If this is an odd packet (because we have SACK 106233ad798cSAdam Langley * blocks etc) then our calculated header_len will be different, and 106333ad798cSAdam Langley * we have to adjust mss_now correspondingly */ 106433ad798cSAdam Langley if (header_len != tp->tcp_header_len) { 106533ad798cSAdam Langley int delta = (int) header_len - tp->tcp_header_len; 106633ad798cSAdam Langley mss_now -= delta; 106733ad798cSAdam Langley } 1068cfb6eeb4SYOSHIFUJI Hideaki 10691da177e4SLinus Torvalds return mss_now; 10701da177e4SLinus Torvalds } 10711da177e4SLinus Torvalds 1072a762a980SDavid S. Miller /* Congestion window validation. (RFC2861) */ 10739e412ba7SIlpo Järvinen static void tcp_cwnd_validate(struct sock *sk) 1074a762a980SDavid S. Miller { 10759e412ba7SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 1076a762a980SDavid S. Miller 1077d436d686SIlpo Järvinen if (tp->packets_out >= tp->snd_cwnd) { 1078a762a980SDavid S. Miller /* Network is feed fully. */ 1079a762a980SDavid S. Miller tp->snd_cwnd_used = 0; 1080a762a980SDavid S. Miller tp->snd_cwnd_stamp = tcp_time_stamp; 1081a762a980SDavid S. Miller } else { 1082a762a980SDavid S. Miller /* Network starves. */ 1083a762a980SDavid S. Miller if (tp->packets_out > tp->snd_cwnd_used) 1084a762a980SDavid S. Miller tp->snd_cwnd_used = tp->packets_out; 1085a762a980SDavid S. Miller 108615d33c07SDavid S. Miller if (sysctl_tcp_slow_start_after_idle && 108715d33c07SDavid S. Miller (s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto) 1088a762a980SDavid S. Miller tcp_cwnd_application_limited(sk); 1089a762a980SDavid S. Miller } 1090a762a980SDavid S. Miller } 1091a762a980SDavid S. Miller 10920e3a4803SIlpo Järvinen /* Returns the portion of skb which can be sent right away without 10930e3a4803SIlpo Järvinen * introducing MSS oddities to segment boundaries. In rare cases where 10940e3a4803SIlpo Järvinen * mss_now != mss_cache, we will request caller to create a small skb 10950e3a4803SIlpo Järvinen * per input skb which could be mostly avoided here (if desired). 10965ea3a748SIlpo Järvinen * 10975ea3a748SIlpo Järvinen * We explicitly want to create a request for splitting write queue tail 10985ea3a748SIlpo Järvinen * to a small skb for Nagle purposes while avoiding unnecessary modulos, 10995ea3a748SIlpo Järvinen * thus all the complexity (cwnd_len is always MSS multiple which we 11005ea3a748SIlpo Järvinen * return whenever allowed by the other factors). Basically we need the 11015ea3a748SIlpo Järvinen * modulo only when the receiver window alone is the limiting factor or 11025ea3a748SIlpo Järvinen * when we would be allowed to send the split-due-to-Nagle skb fully. 11030e3a4803SIlpo Järvinen */ 11040e3a4803SIlpo Järvinen static unsigned int tcp_mss_split_point(struct sock *sk, struct sk_buff *skb, 1105056834d9SIlpo Järvinen unsigned int mss_now, unsigned int cwnd) 1106c1b4a7e6SDavid S. Miller { 11070e3a4803SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 11080e3a4803SIlpo Järvinen u32 needed, window, cwnd_len; 1109c1b4a7e6SDavid S. Miller 111090840defSIlpo Järvinen window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 1111c1b4a7e6SDavid S. Miller cwnd_len = mss_now * cwnd; 11120e3a4803SIlpo Järvinen 11130e3a4803SIlpo Järvinen if (likely(cwnd_len <= window && skb != tcp_write_queue_tail(sk))) 11140e3a4803SIlpo Järvinen return cwnd_len; 11150e3a4803SIlpo Järvinen 11165ea3a748SIlpo Järvinen needed = min(skb->len, window); 11175ea3a748SIlpo Järvinen 111817515408SIlpo Järvinen if (cwnd_len <= needed) 11190e3a4803SIlpo Järvinen return cwnd_len; 11200e3a4803SIlpo Järvinen 11210e3a4803SIlpo Järvinen return needed - needed % mss_now; 1122c1b4a7e6SDavid S. Miller } 1123c1b4a7e6SDavid S. Miller 1124c1b4a7e6SDavid S. Miller /* Can at least one segment of SKB be sent right now, according to the 1125c1b4a7e6SDavid S. Miller * congestion window rules? If so, return how many segments are allowed. 1126c1b4a7e6SDavid S. Miller */ 1127056834d9SIlpo Järvinen static inline unsigned int tcp_cwnd_test(struct tcp_sock *tp, 1128056834d9SIlpo Järvinen struct sk_buff *skb) 1129c1b4a7e6SDavid S. Miller { 1130c1b4a7e6SDavid S. Miller u32 in_flight, cwnd; 1131c1b4a7e6SDavid S. Miller 1132c1b4a7e6SDavid S. Miller /* Don't be strict about the congestion window for the final FIN. */ 1133104439a8SJohn Heffner if ((TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN) && 1134104439a8SJohn Heffner tcp_skb_pcount(skb) == 1) 1135c1b4a7e6SDavid S. Miller return 1; 1136c1b4a7e6SDavid S. Miller 1137c1b4a7e6SDavid S. Miller in_flight = tcp_packets_in_flight(tp); 1138c1b4a7e6SDavid S. Miller cwnd = tp->snd_cwnd; 1139c1b4a7e6SDavid S. Miller if (in_flight < cwnd) 1140c1b4a7e6SDavid S. Miller return (cwnd - in_flight); 1141c1b4a7e6SDavid S. Miller 1142c1b4a7e6SDavid S. Miller return 0; 1143c1b4a7e6SDavid S. Miller } 1144c1b4a7e6SDavid S. Miller 1145c1b4a7e6SDavid S. Miller /* This must be invoked the first time we consider transmitting 1146c1b4a7e6SDavid S. Miller * SKB onto the wire. 1147c1b4a7e6SDavid S. Miller */ 1148056834d9SIlpo Järvinen static int tcp_init_tso_segs(struct sock *sk, struct sk_buff *skb, 1149056834d9SIlpo Järvinen unsigned int mss_now) 1150c1b4a7e6SDavid S. Miller { 1151c1b4a7e6SDavid S. Miller int tso_segs = tcp_skb_pcount(skb); 1152c1b4a7e6SDavid S. Miller 1153f8269a49SIlpo Järvinen if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) { 1154846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, skb, mss_now); 1155c1b4a7e6SDavid S. Miller tso_segs = tcp_skb_pcount(skb); 1156c1b4a7e6SDavid S. Miller } 1157c1b4a7e6SDavid S. Miller return tso_segs; 1158c1b4a7e6SDavid S. Miller } 1159c1b4a7e6SDavid S. Miller 1160c1b4a7e6SDavid S. Miller static inline int tcp_minshall_check(const struct tcp_sock *tp) 1161c1b4a7e6SDavid S. Miller { 1162c1b4a7e6SDavid S. Miller return after(tp->snd_sml, tp->snd_una) && 1163c1b4a7e6SDavid S. Miller !after(tp->snd_sml, tp->snd_nxt); 1164c1b4a7e6SDavid S. Miller } 1165c1b4a7e6SDavid S. Miller 1166c1b4a7e6SDavid S. Miller /* Return 0, if packet can be sent now without violation Nagle's rules: 1167c1b4a7e6SDavid S. Miller * 1. It is full sized. 1168c1b4a7e6SDavid S. Miller * 2. Or it contains FIN. (already checked by caller) 1169c1b4a7e6SDavid S. Miller * 3. Or TCP_NODELAY was set. 1170c1b4a7e6SDavid S. Miller * 4. Or TCP_CORK is not set, and all sent packets are ACKed. 1171c1b4a7e6SDavid S. Miller * With Minshall's modification: all sent small packets are ACKed. 1172c1b4a7e6SDavid S. Miller */ 1173c1b4a7e6SDavid S. Miller static inline int tcp_nagle_check(const struct tcp_sock *tp, 1174c1b4a7e6SDavid S. Miller const struct sk_buff *skb, 1175c1b4a7e6SDavid S. Miller unsigned mss_now, int nonagle) 1176c1b4a7e6SDavid S. Miller { 1177c1b4a7e6SDavid S. Miller return (skb->len < mss_now && 1178c1b4a7e6SDavid S. Miller ((nonagle & TCP_NAGLE_CORK) || 1179056834d9SIlpo Järvinen (!nonagle && tp->packets_out && tcp_minshall_check(tp)))); 1180c1b4a7e6SDavid S. Miller } 1181c1b4a7e6SDavid S. Miller 1182c1b4a7e6SDavid S. Miller /* Return non-zero if the Nagle test allows this packet to be 1183c1b4a7e6SDavid S. Miller * sent now. 1184c1b4a7e6SDavid S. Miller */ 1185c1b4a7e6SDavid S. Miller static inline int tcp_nagle_test(struct tcp_sock *tp, struct sk_buff *skb, 1186c1b4a7e6SDavid S. Miller unsigned int cur_mss, int nonagle) 1187c1b4a7e6SDavid S. Miller { 1188c1b4a7e6SDavid S. Miller /* Nagle rule does not apply to frames, which sit in the middle of the 1189c1b4a7e6SDavid S. Miller * write_queue (they have no chances to get new data). 1190c1b4a7e6SDavid S. Miller * 1191c1b4a7e6SDavid S. Miller * This is implemented in the callers, where they modify the 'nonagle' 1192c1b4a7e6SDavid S. Miller * argument based upon the location of SKB in the send queue. 1193c1b4a7e6SDavid S. Miller */ 1194c1b4a7e6SDavid S. Miller if (nonagle & TCP_NAGLE_PUSH) 1195c1b4a7e6SDavid S. Miller return 1; 1196c1b4a7e6SDavid S. Miller 1197d551e454SIlpo Järvinen /* Don't use the nagle rule for urgent data (or for the final FIN). 1198d551e454SIlpo Järvinen * Nagle can be ignored during F-RTO too (see RFC4138). 1199d551e454SIlpo Järvinen */ 120033f5f57eSIlpo Järvinen if (tcp_urg_mode(tp) || (tp->frto_counter == 2) || 1201c1b4a7e6SDavid S. Miller (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN)) 1202c1b4a7e6SDavid S. Miller return 1; 1203c1b4a7e6SDavid S. Miller 1204c1b4a7e6SDavid S. Miller if (!tcp_nagle_check(tp, skb, cur_mss, nonagle)) 1205c1b4a7e6SDavid S. Miller return 1; 1206c1b4a7e6SDavid S. Miller 1207c1b4a7e6SDavid S. Miller return 0; 1208c1b4a7e6SDavid S. Miller } 1209c1b4a7e6SDavid S. Miller 1210c1b4a7e6SDavid S. Miller /* Does at least the first segment of SKB fit into the send window? */ 1211056834d9SIlpo Järvinen static inline int tcp_snd_wnd_test(struct tcp_sock *tp, struct sk_buff *skb, 1212056834d9SIlpo Järvinen unsigned int cur_mss) 1213c1b4a7e6SDavid S. Miller { 1214c1b4a7e6SDavid S. Miller u32 end_seq = TCP_SKB_CB(skb)->end_seq; 1215c1b4a7e6SDavid S. Miller 1216c1b4a7e6SDavid S. Miller if (skb->len > cur_mss) 1217c1b4a7e6SDavid S. Miller end_seq = TCP_SKB_CB(skb)->seq + cur_mss; 1218c1b4a7e6SDavid S. Miller 121990840defSIlpo Järvinen return !after(end_seq, tcp_wnd_end(tp)); 1220c1b4a7e6SDavid S. Miller } 1221c1b4a7e6SDavid S. Miller 1222fe067e8aSDavid S. Miller /* This checks if the data bearing packet SKB (usually tcp_send_head(sk)) 1223c1b4a7e6SDavid S. Miller * should be put on the wire right now. If so, it returns the number of 1224c1b4a7e6SDavid S. Miller * packets allowed by the congestion window. 1225c1b4a7e6SDavid S. Miller */ 1226c1b4a7e6SDavid S. Miller static unsigned int tcp_snd_test(struct sock *sk, struct sk_buff *skb, 1227c1b4a7e6SDavid S. Miller unsigned int cur_mss, int nonagle) 1228c1b4a7e6SDavid S. Miller { 1229c1b4a7e6SDavid S. Miller struct tcp_sock *tp = tcp_sk(sk); 1230c1b4a7e6SDavid S. Miller unsigned int cwnd_quota; 1231c1b4a7e6SDavid S. Miller 1232846998aeSDavid S. Miller tcp_init_tso_segs(sk, skb, cur_mss); 1233c1b4a7e6SDavid S. Miller 1234c1b4a7e6SDavid S. Miller if (!tcp_nagle_test(tp, skb, cur_mss, nonagle)) 1235c1b4a7e6SDavid S. Miller return 0; 1236c1b4a7e6SDavid S. Miller 1237c1b4a7e6SDavid S. Miller cwnd_quota = tcp_cwnd_test(tp, skb); 1238056834d9SIlpo Järvinen if (cwnd_quota && !tcp_snd_wnd_test(tp, skb, cur_mss)) 1239c1b4a7e6SDavid S. Miller cwnd_quota = 0; 1240c1b4a7e6SDavid S. Miller 1241c1b4a7e6SDavid S. Miller return cwnd_quota; 1242c1b4a7e6SDavid S. Miller } 1243c1b4a7e6SDavid S. Miller 12449e412ba7SIlpo Järvinen int tcp_may_send_now(struct sock *sk) 1245c1b4a7e6SDavid S. Miller { 12469e412ba7SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 1247fe067e8aSDavid S. Miller struct sk_buff *skb = tcp_send_head(sk); 1248c1b4a7e6SDavid S. Miller 1249c1b4a7e6SDavid S. Miller return (skb && 12500c54b85fSIlpo Järvinen tcp_snd_test(sk, skb, tcp_current_mss(sk), 1251c1b4a7e6SDavid S. Miller (tcp_skb_is_last(sk, skb) ? 12524e67d876SIlpo Järvinen tp->nonagle : TCP_NAGLE_PUSH))); 1253c1b4a7e6SDavid S. Miller } 1254c1b4a7e6SDavid S. Miller 1255c1b4a7e6SDavid S. Miller /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet 1256c1b4a7e6SDavid S. Miller * which is put after SKB on the list. It is very much like 1257c1b4a7e6SDavid S. Miller * tcp_fragment() except that it may make several kinds of assumptions 1258c1b4a7e6SDavid S. Miller * in order to speed up the splitting operation. In particular, we 1259c1b4a7e6SDavid S. Miller * know that all the data is in scatter-gather pages, and that the 1260c1b4a7e6SDavid S. Miller * packet has never been sent out before (and thus is not cloned). 1261c1b4a7e6SDavid S. Miller */ 1262056834d9SIlpo Järvinen static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len, 1263056834d9SIlpo Järvinen unsigned int mss_now) 1264c1b4a7e6SDavid S. Miller { 1265c1b4a7e6SDavid S. Miller struct sk_buff *buff; 1266c1b4a7e6SDavid S. Miller int nlen = skb->len - len; 12679ce01461SIlpo Järvinen u8 flags; 1268c1b4a7e6SDavid S. Miller 1269c1b4a7e6SDavid S. Miller /* All of a TSO frame must be composed of paged data. */ 1270c8ac3774SHerbert Xu if (skb->len != skb->data_len) 1271c8ac3774SHerbert Xu return tcp_fragment(sk, skb, len, mss_now); 1272c1b4a7e6SDavid S. Miller 1273df97c708SPavel Emelyanov buff = sk_stream_alloc_skb(sk, 0, GFP_ATOMIC); 1274c1b4a7e6SDavid S. Miller if (unlikely(buff == NULL)) 1275c1b4a7e6SDavid S. Miller return -ENOMEM; 1276c1b4a7e6SDavid S. Miller 12773ab224beSHideo Aoki sk->sk_wmem_queued += buff->truesize; 12783ab224beSHideo Aoki sk_mem_charge(sk, buff->truesize); 1279b60b49eaSHerbert Xu buff->truesize += nlen; 1280c1b4a7e6SDavid S. Miller skb->truesize -= nlen; 1281c1b4a7e6SDavid S. Miller 1282c1b4a7e6SDavid S. Miller /* Correct the sequence numbers. */ 1283c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; 1284c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; 1285c1b4a7e6SDavid S. Miller TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; 1286c1b4a7e6SDavid S. Miller 1287c1b4a7e6SDavid S. Miller /* PSH and FIN should only be set in the second packet. */ 1288c1b4a7e6SDavid S. Miller flags = TCP_SKB_CB(skb)->flags; 1289c1b4a7e6SDavid S. Miller TCP_SKB_CB(skb)->flags = flags & ~(TCPCB_FLAG_FIN | TCPCB_FLAG_PSH); 1290c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->flags = flags; 1291c1b4a7e6SDavid S. Miller 1292c1b4a7e6SDavid S. Miller /* This packet was never sent out yet, so no SACK bits. */ 1293c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->sacked = 0; 1294c1b4a7e6SDavid S. Miller 129584fa7933SPatrick McHardy buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL; 1296c1b4a7e6SDavid S. Miller skb_split(skb, buff, len); 1297c1b4a7e6SDavid S. Miller 1298c1b4a7e6SDavid S. Miller /* Fix up tso_factor for both original and new SKB. */ 1299846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, skb, mss_now); 1300846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, buff, mss_now); 1301c1b4a7e6SDavid S. Miller 1302c1b4a7e6SDavid S. Miller /* Link BUFF into the send queue. */ 1303c1b4a7e6SDavid S. Miller skb_header_release(buff); 1304fe067e8aSDavid S. Miller tcp_insert_write_queue_after(skb, buff, sk); 1305c1b4a7e6SDavid S. Miller 1306c1b4a7e6SDavid S. Miller return 0; 1307c1b4a7e6SDavid S. Miller } 1308c1b4a7e6SDavid S. Miller 1309c1b4a7e6SDavid S. Miller /* Try to defer sending, if possible, in order to minimize the amount 1310c1b4a7e6SDavid S. Miller * of TSO splitting we do. View it as a kind of TSO Nagle test. 1311c1b4a7e6SDavid S. Miller * 1312c1b4a7e6SDavid S. Miller * This algorithm is from John Heffner. 1313c1b4a7e6SDavid S. Miller */ 13149e412ba7SIlpo Järvinen static int tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb) 1315c1b4a7e6SDavid S. Miller { 13169e412ba7SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 13176687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 1318c1b4a7e6SDavid S. Miller u32 send_win, cong_win, limit, in_flight; 1319c1b4a7e6SDavid S. Miller 1320c1b4a7e6SDavid S. Miller if (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN) 1321ae8064acSJohn Heffner goto send_now; 1322c1b4a7e6SDavid S. Miller 13236687e988SArnaldo Carvalho de Melo if (icsk->icsk_ca_state != TCP_CA_Open) 1324ae8064acSJohn Heffner goto send_now; 1325ae8064acSJohn Heffner 1326ae8064acSJohn Heffner /* Defer for less than two clock ticks. */ 1327bd515c3eSIlpo Järvinen if (tp->tso_deferred && 1328a2acde07SIlpo Järvinen (((u32)jiffies << 1) >> 1) - (tp->tso_deferred >> 1) > 1) 1329ae8064acSJohn Heffner goto send_now; 1330908a75c1SDavid S. Miller 1331c1b4a7e6SDavid S. Miller in_flight = tcp_packets_in_flight(tp); 1332c1b4a7e6SDavid S. Miller 1333056834d9SIlpo Järvinen BUG_ON(tcp_skb_pcount(skb) <= 1 || (tp->snd_cwnd <= in_flight)); 1334c1b4a7e6SDavid S. Miller 133590840defSIlpo Järvinen send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 1336c1b4a7e6SDavid S. Miller 1337c1b4a7e6SDavid S. Miller /* From in_flight test above, we know that cwnd > in_flight. */ 1338c1b4a7e6SDavid S. Miller cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache; 1339c1b4a7e6SDavid S. Miller 1340c1b4a7e6SDavid S. Miller limit = min(send_win, cong_win); 1341c1b4a7e6SDavid S. Miller 1342ba244fe9SDavid S. Miller /* If a full-sized TSO skb can be sent, do it. */ 134382cc1a7aSPeter P Waskiewicz Jr if (limit >= sk->sk_gso_max_size) 1344ae8064acSJohn Heffner goto send_now; 1345ba244fe9SDavid S. Miller 134662ad2761SIlpo Järvinen /* Middle in queue won't get any more data, full sendable already? */ 134762ad2761SIlpo Järvinen if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len)) 134862ad2761SIlpo Järvinen goto send_now; 134962ad2761SIlpo Järvinen 1350c1b4a7e6SDavid S. Miller if (sysctl_tcp_tso_win_divisor) { 1351c1b4a7e6SDavid S. Miller u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache); 1352c1b4a7e6SDavid S. Miller 1353c1b4a7e6SDavid S. Miller /* If at least some fraction of a window is available, 1354c1b4a7e6SDavid S. Miller * just use it. 1355c1b4a7e6SDavid S. Miller */ 1356c1b4a7e6SDavid S. Miller chunk /= sysctl_tcp_tso_win_divisor; 1357c1b4a7e6SDavid S. Miller if (limit >= chunk) 1358ae8064acSJohn Heffner goto send_now; 1359c1b4a7e6SDavid S. Miller } else { 1360c1b4a7e6SDavid S. Miller /* Different approach, try not to defer past a single 1361c1b4a7e6SDavid S. Miller * ACK. Receiver should ACK every other full sized 1362c1b4a7e6SDavid S. Miller * frame, so if we have space for more than 3 frames 1363c1b4a7e6SDavid S. Miller * then send now. 1364c1b4a7e6SDavid S. Miller */ 1365c1b4a7e6SDavid S. Miller if (limit > tcp_max_burst(tp) * tp->mss_cache) 1366ae8064acSJohn Heffner goto send_now; 1367c1b4a7e6SDavid S. Miller } 1368c1b4a7e6SDavid S. Miller 1369c1b4a7e6SDavid S. Miller /* Ok, it looks like it is advisable to defer. */ 1370ae8064acSJohn Heffner tp->tso_deferred = 1 | (jiffies << 1); 1371ae8064acSJohn Heffner 1372c1b4a7e6SDavid S. Miller return 1; 1373ae8064acSJohn Heffner 1374ae8064acSJohn Heffner send_now: 1375ae8064acSJohn Heffner tp->tso_deferred = 0; 1376ae8064acSJohn Heffner return 0; 1377c1b4a7e6SDavid S. Miller } 1378c1b4a7e6SDavid S. Miller 13795d424d5aSJohn Heffner /* Create a new MTU probe if we are ready. 13805d424d5aSJohn Heffner * Returns 0 if we should wait to probe (no cwnd available), 13815d424d5aSJohn Heffner * 1 if a probe was sent, 1382056834d9SIlpo Järvinen * -1 otherwise 1383056834d9SIlpo Järvinen */ 13845d424d5aSJohn Heffner static int tcp_mtu_probe(struct sock *sk) 13855d424d5aSJohn Heffner { 13865d424d5aSJohn Heffner struct tcp_sock *tp = tcp_sk(sk); 13875d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 13885d424d5aSJohn Heffner struct sk_buff *skb, *nskb, *next; 13895d424d5aSJohn Heffner int len; 13905d424d5aSJohn Heffner int probe_size; 139191cc17c0SIlpo Järvinen int size_needed; 13925d424d5aSJohn Heffner int copy; 13935d424d5aSJohn Heffner int mss_now; 13945d424d5aSJohn Heffner 13955d424d5aSJohn Heffner /* Not currently probing/verifying, 13965d424d5aSJohn Heffner * not in recovery, 13975d424d5aSJohn Heffner * have enough cwnd, and 13985d424d5aSJohn Heffner * not SACKing (the variable headers throw things off) */ 13995d424d5aSJohn Heffner if (!icsk->icsk_mtup.enabled || 14005d424d5aSJohn Heffner icsk->icsk_mtup.probe_size || 14015d424d5aSJohn Heffner inet_csk(sk)->icsk_ca_state != TCP_CA_Open || 14025d424d5aSJohn Heffner tp->snd_cwnd < 11 || 1403cabeccbdSIlpo Järvinen tp->rx_opt.num_sacks || tp->rx_opt.dsack) 14045d424d5aSJohn Heffner return -1; 14055d424d5aSJohn Heffner 14065d424d5aSJohn Heffner /* Very simple search strategy: just double the MSS. */ 14070c54b85fSIlpo Järvinen mss_now = tcp_current_mss(sk); 14085d424d5aSJohn Heffner probe_size = 2 * tp->mss_cache; 140991cc17c0SIlpo Järvinen size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache; 14105d424d5aSJohn Heffner if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high)) { 14115d424d5aSJohn Heffner /* TODO: set timer for probe_converge_event */ 14125d424d5aSJohn Heffner return -1; 14135d424d5aSJohn Heffner } 14145d424d5aSJohn Heffner 14155d424d5aSJohn Heffner /* Have enough data in the send queue to probe? */ 14167f9c33e5SIlpo Järvinen if (tp->write_seq - tp->snd_nxt < size_needed) 14175d424d5aSJohn Heffner return -1; 14185d424d5aSJohn Heffner 141991cc17c0SIlpo Järvinen if (tp->snd_wnd < size_needed) 14205d424d5aSJohn Heffner return -1; 142190840defSIlpo Järvinen if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp))) 14225d424d5aSJohn Heffner return 0; 14235d424d5aSJohn Heffner 1424d67c58e9SIlpo Järvinen /* Do we need to wait to drain cwnd? With none in flight, don't stall */ 1425d67c58e9SIlpo Järvinen if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) { 1426d67c58e9SIlpo Järvinen if (!tcp_packets_in_flight(tp)) 14275d424d5aSJohn Heffner return -1; 14285d424d5aSJohn Heffner else 14295d424d5aSJohn Heffner return 0; 14305d424d5aSJohn Heffner } 14315d424d5aSJohn Heffner 14325d424d5aSJohn Heffner /* We're allowed to probe. Build it now. */ 14335d424d5aSJohn Heffner if ((nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC)) == NULL) 14345d424d5aSJohn Heffner return -1; 14353ab224beSHideo Aoki sk->sk_wmem_queued += nskb->truesize; 14363ab224beSHideo Aoki sk_mem_charge(sk, nskb->truesize); 14375d424d5aSJohn Heffner 1438fe067e8aSDavid S. Miller skb = tcp_send_head(sk); 14395d424d5aSJohn Heffner 14405d424d5aSJohn Heffner TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq; 14415d424d5aSJohn Heffner TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size; 14425d424d5aSJohn Heffner TCP_SKB_CB(nskb)->flags = TCPCB_FLAG_ACK; 14435d424d5aSJohn Heffner TCP_SKB_CB(nskb)->sacked = 0; 14445d424d5aSJohn Heffner nskb->csum = 0; 144584fa7933SPatrick McHardy nskb->ip_summed = skb->ip_summed; 14465d424d5aSJohn Heffner 144750c4817eSIlpo Järvinen tcp_insert_write_queue_before(nskb, skb, sk); 144850c4817eSIlpo Järvinen 14495d424d5aSJohn Heffner len = 0; 1450234b6860SIlpo Järvinen tcp_for_write_queue_from_safe(skb, next, sk) { 14515d424d5aSJohn Heffner copy = min_t(int, skb->len, probe_size - len); 14525d424d5aSJohn Heffner if (nskb->ip_summed) 14535d424d5aSJohn Heffner skb_copy_bits(skb, 0, skb_put(nskb, copy), copy); 14545d424d5aSJohn Heffner else 14555d424d5aSJohn Heffner nskb->csum = skb_copy_and_csum_bits(skb, 0, 1456056834d9SIlpo Järvinen skb_put(nskb, copy), 1457056834d9SIlpo Järvinen copy, nskb->csum); 14585d424d5aSJohn Heffner 14595d424d5aSJohn Heffner if (skb->len <= copy) { 14605d424d5aSJohn Heffner /* We've eaten all the data from this skb. 14615d424d5aSJohn Heffner * Throw it away. */ 14625d424d5aSJohn Heffner TCP_SKB_CB(nskb)->flags |= TCP_SKB_CB(skb)->flags; 1463fe067e8aSDavid S. Miller tcp_unlink_write_queue(skb, sk); 14643ab224beSHideo Aoki sk_wmem_free_skb(sk, skb); 14655d424d5aSJohn Heffner } else { 14665d424d5aSJohn Heffner TCP_SKB_CB(nskb)->flags |= TCP_SKB_CB(skb)->flags & 14675d424d5aSJohn Heffner ~(TCPCB_FLAG_FIN|TCPCB_FLAG_PSH); 14685d424d5aSJohn Heffner if (!skb_shinfo(skb)->nr_frags) { 14695d424d5aSJohn Heffner skb_pull(skb, copy); 147084fa7933SPatrick McHardy if (skb->ip_summed != CHECKSUM_PARTIAL) 1471056834d9SIlpo Järvinen skb->csum = csum_partial(skb->data, 1472056834d9SIlpo Järvinen skb->len, 0); 14735d424d5aSJohn Heffner } else { 14745d424d5aSJohn Heffner __pskb_trim_head(skb, copy); 14755d424d5aSJohn Heffner tcp_set_skb_tso_segs(sk, skb, mss_now); 14765d424d5aSJohn Heffner } 14775d424d5aSJohn Heffner TCP_SKB_CB(skb)->seq += copy; 14785d424d5aSJohn Heffner } 14795d424d5aSJohn Heffner 14805d424d5aSJohn Heffner len += copy; 1481234b6860SIlpo Järvinen 1482234b6860SIlpo Järvinen if (len >= probe_size) 1483234b6860SIlpo Järvinen break; 14845d424d5aSJohn Heffner } 14855d424d5aSJohn Heffner tcp_init_tso_segs(sk, nskb, nskb->len); 14865d424d5aSJohn Heffner 14875d424d5aSJohn Heffner /* We're ready to send. If this fails, the probe will 14885d424d5aSJohn Heffner * be resegmented into mss-sized pieces by tcp_write_xmit(). */ 14895d424d5aSJohn Heffner TCP_SKB_CB(nskb)->when = tcp_time_stamp; 14905d424d5aSJohn Heffner if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) { 14915d424d5aSJohn Heffner /* Decrement cwnd here because we are sending 14925d424d5aSJohn Heffner * effectively two packets. */ 14935d424d5aSJohn Heffner tp->snd_cwnd--; 149466f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, nskb); 14955d424d5aSJohn Heffner 14965d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len); 14970e7b1368SJohn Heffner tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq; 14980e7b1368SJohn Heffner tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq; 14995d424d5aSJohn Heffner 15005d424d5aSJohn Heffner return 1; 15015d424d5aSJohn Heffner } 15025d424d5aSJohn Heffner 15035d424d5aSJohn Heffner return -1; 15045d424d5aSJohn Heffner } 15055d424d5aSJohn Heffner 15061da177e4SLinus Torvalds /* This routine writes packets to the network. It advances the 15071da177e4SLinus Torvalds * send_head. This happens as incoming acks open up the remote 15081da177e4SLinus Torvalds * window for us. 15091da177e4SLinus Torvalds * 1510f8269a49SIlpo Järvinen * LARGESEND note: !tcp_urg_mode is overkill, only frames between 1511f8269a49SIlpo Järvinen * snd_up-64k-mss .. snd_up cannot be large. However, taking into 1512f8269a49SIlpo Järvinen * account rare use of URG, this is not a big flaw. 1513f8269a49SIlpo Järvinen * 15141da177e4SLinus Torvalds * Returns 1, if no segments are in flight and we have queued segments, but 15151da177e4SLinus Torvalds * cannot send anything now because of SWS or another problem. 15161da177e4SLinus Torvalds */ 1517d5dd9175SIlpo Järvinen static int tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle, 1518d5dd9175SIlpo Järvinen int push_one, gfp_t gfp) 15191da177e4SLinus Torvalds { 15201da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 152192df7b51SDavid S. Miller struct sk_buff *skb; 1522c1b4a7e6SDavid S. Miller unsigned int tso_segs, sent_pkts; 1523c1b4a7e6SDavid S. Miller int cwnd_quota; 15245d424d5aSJohn Heffner int result; 15251da177e4SLinus Torvalds 1526c1b4a7e6SDavid S. Miller sent_pkts = 0; 15275d424d5aSJohn Heffner 1528d5dd9175SIlpo Järvinen if (!push_one) { 15295d424d5aSJohn Heffner /* Do MTU probing. */ 1530d5dd9175SIlpo Järvinen result = tcp_mtu_probe(sk); 1531d5dd9175SIlpo Järvinen if (!result) { 15325d424d5aSJohn Heffner return 0; 15335d424d5aSJohn Heffner } else if (result > 0) { 15345d424d5aSJohn Heffner sent_pkts = 1; 15355d424d5aSJohn Heffner } 1536d5dd9175SIlpo Järvinen } 15375d424d5aSJohn Heffner 1538fe067e8aSDavid S. Miller while ((skb = tcp_send_head(sk))) { 1539c8ac3774SHerbert Xu unsigned int limit; 1540c8ac3774SHerbert Xu 1541b68e9f85SHerbert Xu tso_segs = tcp_init_tso_segs(sk, skb, mss_now); 1542c1b4a7e6SDavid S. Miller BUG_ON(!tso_segs); 1543c1b4a7e6SDavid S. Miller 1544b68e9f85SHerbert Xu cwnd_quota = tcp_cwnd_test(tp, skb); 1545b68e9f85SHerbert Xu if (!cwnd_quota) 1546b68e9f85SHerbert Xu break; 1547b68e9f85SHerbert Xu 1548b68e9f85SHerbert Xu if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now))) 1549b68e9f85SHerbert Xu break; 1550b68e9f85SHerbert Xu 1551c1b4a7e6SDavid S. Miller if (tso_segs == 1) { 1552aa93466bSDavid S. Miller if (unlikely(!tcp_nagle_test(tp, skb, mss_now, 1553aa93466bSDavid S. Miller (tcp_skb_is_last(sk, skb) ? 1554aa93466bSDavid S. Miller nonagle : TCP_NAGLE_PUSH)))) 1555aa93466bSDavid S. Miller break; 1556c1b4a7e6SDavid S. Miller } else { 1557d5dd9175SIlpo Järvinen if (!push_one && tcp_tso_should_defer(sk, skb)) 1558aa93466bSDavid S. Miller break; 1559c1b4a7e6SDavid S. Miller } 1560aa93466bSDavid S. Miller 1561c8ac3774SHerbert Xu limit = mss_now; 1562f8269a49SIlpo Järvinen if (tso_segs > 1 && !tcp_urg_mode(tp)) 15630e3a4803SIlpo Järvinen limit = tcp_mss_split_point(sk, skb, mss_now, 15640e3a4803SIlpo Järvinen cwnd_quota); 1565c8ac3774SHerbert Xu 1566c8ac3774SHerbert Xu if (skb->len > limit && 1567c8ac3774SHerbert Xu unlikely(tso_fragment(sk, skb, limit, mss_now))) 15681da177e4SLinus Torvalds break; 15691da177e4SLinus Torvalds 15701da177e4SLinus Torvalds TCP_SKB_CB(skb)->when = tcp_time_stamp; 1571c1b4a7e6SDavid S. Miller 1572d5dd9175SIlpo Järvinen if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp))) 15731da177e4SLinus Torvalds break; 15741da177e4SLinus Torvalds 15751da177e4SLinus Torvalds /* Advance the send_head. This one is sent out. 15761da177e4SLinus Torvalds * This call will increment packets_out. 15771da177e4SLinus Torvalds */ 157866f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, skb); 15791da177e4SLinus Torvalds 15801da177e4SLinus Torvalds tcp_minshall_update(tp, mss_now, skb); 1581aa93466bSDavid S. Miller sent_pkts++; 1582d5dd9175SIlpo Järvinen 1583d5dd9175SIlpo Järvinen if (push_one) 1584d5dd9175SIlpo Järvinen break; 15851da177e4SLinus Torvalds } 15861da177e4SLinus Torvalds 1587aa93466bSDavid S. Miller if (likely(sent_pkts)) { 15889e412ba7SIlpo Järvinen tcp_cwnd_validate(sk); 15891da177e4SLinus Torvalds return 0; 15901da177e4SLinus Torvalds } 1591fe067e8aSDavid S. Miller return !tp->packets_out && tcp_send_head(sk); 15921da177e4SLinus Torvalds } 15931da177e4SLinus Torvalds 1594a762a980SDavid S. Miller /* Push out any pending frames which were held back due to 1595a762a980SDavid S. Miller * TCP_CORK or attempt at coalescing tiny packets. 1596a762a980SDavid S. Miller * The socket must be locked by the caller. 1597a762a980SDavid S. Miller */ 15989e412ba7SIlpo Järvinen void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss, 15999e412ba7SIlpo Järvinen int nonagle) 1600a762a980SDavid S. Miller { 1601fe067e8aSDavid S. Miller struct sk_buff *skb = tcp_send_head(sk); 1602a762a980SDavid S. Miller 1603726e07a8SIlpo Järvinen if (!skb) 1604726e07a8SIlpo Järvinen return; 1605726e07a8SIlpo Järvinen 1606726e07a8SIlpo Järvinen /* If we are closed, the bytes will have to remain here. 1607726e07a8SIlpo Järvinen * In time closedown will finish, we empty the write queue and 1608726e07a8SIlpo Järvinen * all will be happy. 1609726e07a8SIlpo Järvinen */ 1610726e07a8SIlpo Järvinen if (unlikely(sk->sk_state == TCP_CLOSE)) 1611726e07a8SIlpo Järvinen return; 1612726e07a8SIlpo Järvinen 1613d5dd9175SIlpo Järvinen if (tcp_write_xmit(sk, cur_mss, nonagle, 0, GFP_ATOMIC)) 16149e412ba7SIlpo Järvinen tcp_check_probe_timer(sk); 1615a762a980SDavid S. Miller } 1616a762a980SDavid S. Miller 1617c1b4a7e6SDavid S. Miller /* Send _single_ skb sitting at the send head. This function requires 1618c1b4a7e6SDavid S. Miller * true push pending frames to setup probe timer etc. 1619c1b4a7e6SDavid S. Miller */ 1620c1b4a7e6SDavid S. Miller void tcp_push_one(struct sock *sk, unsigned int mss_now) 1621c1b4a7e6SDavid S. Miller { 1622fe067e8aSDavid S. Miller struct sk_buff *skb = tcp_send_head(sk); 1623c1b4a7e6SDavid S. Miller 1624c1b4a7e6SDavid S. Miller BUG_ON(!skb || skb->len < mss_now); 1625c1b4a7e6SDavid S. Miller 1626d5dd9175SIlpo Järvinen tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation); 1627c1b4a7e6SDavid S. Miller } 1628c1b4a7e6SDavid S. Miller 16291da177e4SLinus Torvalds /* This function returns the amount that we can raise the 16301da177e4SLinus Torvalds * usable window based on the following constraints 16311da177e4SLinus Torvalds * 16321da177e4SLinus Torvalds * 1. The window can never be shrunk once it is offered (RFC 793) 16331da177e4SLinus Torvalds * 2. We limit memory per socket 16341da177e4SLinus Torvalds * 16351da177e4SLinus Torvalds * RFC 1122: 16361da177e4SLinus Torvalds * "the suggested [SWS] avoidance algorithm for the receiver is to keep 16371da177e4SLinus Torvalds * RECV.NEXT + RCV.WIN fixed until: 16381da177e4SLinus Torvalds * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)" 16391da177e4SLinus Torvalds * 16401da177e4SLinus Torvalds * i.e. don't raise the right edge of the window until you can raise 16411da177e4SLinus Torvalds * it at least MSS bytes. 16421da177e4SLinus Torvalds * 16431da177e4SLinus Torvalds * Unfortunately, the recommended algorithm breaks header prediction, 16441da177e4SLinus Torvalds * since header prediction assumes th->window stays fixed. 16451da177e4SLinus Torvalds * 16461da177e4SLinus Torvalds * Strictly speaking, keeping th->window fixed violates the receiver 16471da177e4SLinus Torvalds * side SWS prevention criteria. The problem is that under this rule 16481da177e4SLinus Torvalds * a stream of single byte packets will cause the right side of the 16491da177e4SLinus Torvalds * window to always advance by a single byte. 16501da177e4SLinus Torvalds * 16511da177e4SLinus Torvalds * Of course, if the sender implements sender side SWS prevention 16521da177e4SLinus Torvalds * then this will not be a problem. 16531da177e4SLinus Torvalds * 16541da177e4SLinus Torvalds * BSD seems to make the following compromise: 16551da177e4SLinus Torvalds * 16561da177e4SLinus Torvalds * If the free space is less than the 1/4 of the maximum 16571da177e4SLinus Torvalds * space available and the free space is less than 1/2 mss, 16581da177e4SLinus Torvalds * then set the window to 0. 16591da177e4SLinus Torvalds * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ] 16601da177e4SLinus Torvalds * Otherwise, just prevent the window from shrinking 16611da177e4SLinus Torvalds * and from being larger than the largest representable value. 16621da177e4SLinus Torvalds * 16631da177e4SLinus Torvalds * This prevents incremental opening of the window in the regime 16641da177e4SLinus Torvalds * where TCP is limited by the speed of the reader side taking 16651da177e4SLinus Torvalds * data out of the TCP receive queue. It does nothing about 16661da177e4SLinus Torvalds * those cases where the window is constrained on the sender side 16671da177e4SLinus Torvalds * because the pipeline is full. 16681da177e4SLinus Torvalds * 16691da177e4SLinus Torvalds * BSD also seems to "accidentally" limit itself to windows that are a 16701da177e4SLinus Torvalds * multiple of MSS, at least until the free space gets quite small. 16711da177e4SLinus Torvalds * This would appear to be a side effect of the mbuf implementation. 16721da177e4SLinus Torvalds * Combining these two algorithms results in the observed behavior 16731da177e4SLinus Torvalds * of having a fixed window size at almost all times. 16741da177e4SLinus Torvalds * 16751da177e4SLinus Torvalds * Below we obtain similar behavior by forcing the offered window to 16761da177e4SLinus Torvalds * a multiple of the mss when it is feasible to do so. 16771da177e4SLinus Torvalds * 16781da177e4SLinus Torvalds * Note, we don't "adjust" for TIMESTAMP or SACK option bytes. 16791da177e4SLinus Torvalds * Regular options like TIMESTAMP are taken into account. 16801da177e4SLinus Torvalds */ 16811da177e4SLinus Torvalds u32 __tcp_select_window(struct sock *sk) 16821da177e4SLinus Torvalds { 1683463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 16841da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 1685caa20d9aSStephen Hemminger /* MSS for the peer's data. Previous versions used mss_clamp 16861da177e4SLinus Torvalds * here. I don't know if the value based on our guesses 16871da177e4SLinus Torvalds * of peer's MSS is better for the performance. It's more correct 16881da177e4SLinus Torvalds * but may be worse for the performance because of rcv_mss 16891da177e4SLinus Torvalds * fluctuations. --SAW 1998/11/1 16901da177e4SLinus Torvalds */ 1691463c84b9SArnaldo Carvalho de Melo int mss = icsk->icsk_ack.rcv_mss; 16921da177e4SLinus Torvalds int free_space = tcp_space(sk); 16931da177e4SLinus Torvalds int full_space = min_t(int, tp->window_clamp, tcp_full_space(sk)); 16941da177e4SLinus Torvalds int window; 16951da177e4SLinus Torvalds 16961da177e4SLinus Torvalds if (mss > full_space) 16971da177e4SLinus Torvalds mss = full_space; 16981da177e4SLinus Torvalds 1699b92edbe0SEric Dumazet if (free_space < (full_space >> 1)) { 1700463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.quick = 0; 17011da177e4SLinus Torvalds 17021da177e4SLinus Torvalds if (tcp_memory_pressure) 1703056834d9SIlpo Järvinen tp->rcv_ssthresh = min(tp->rcv_ssthresh, 1704056834d9SIlpo Järvinen 4U * tp->advmss); 17051da177e4SLinus Torvalds 17061da177e4SLinus Torvalds if (free_space < mss) 17071da177e4SLinus Torvalds return 0; 17081da177e4SLinus Torvalds } 17091da177e4SLinus Torvalds 17101da177e4SLinus Torvalds if (free_space > tp->rcv_ssthresh) 17111da177e4SLinus Torvalds free_space = tp->rcv_ssthresh; 17121da177e4SLinus Torvalds 17131da177e4SLinus Torvalds /* Don't do rounding if we are using window scaling, since the 17141da177e4SLinus Torvalds * scaled window will not line up with the MSS boundary anyway. 17151da177e4SLinus Torvalds */ 17161da177e4SLinus Torvalds window = tp->rcv_wnd; 17171da177e4SLinus Torvalds if (tp->rx_opt.rcv_wscale) { 17181da177e4SLinus Torvalds window = free_space; 17191da177e4SLinus Torvalds 17201da177e4SLinus Torvalds /* Advertise enough space so that it won't get scaled away. 17211da177e4SLinus Torvalds * Import case: prevent zero window announcement if 17221da177e4SLinus Torvalds * 1<<rcv_wscale > mss. 17231da177e4SLinus Torvalds */ 17241da177e4SLinus Torvalds if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window) 17251da177e4SLinus Torvalds window = (((window >> tp->rx_opt.rcv_wscale) + 1) 17261da177e4SLinus Torvalds << tp->rx_opt.rcv_wscale); 17271da177e4SLinus Torvalds } else { 17281da177e4SLinus Torvalds /* Get the largest window that is a nice multiple of mss. 17291da177e4SLinus Torvalds * Window clamp already applied above. 17301da177e4SLinus Torvalds * If our current window offering is within 1 mss of the 17311da177e4SLinus Torvalds * free space we just keep it. This prevents the divide 17321da177e4SLinus Torvalds * and multiply from happening most of the time. 17331da177e4SLinus Torvalds * We also don't do any window rounding when the free space 17341da177e4SLinus Torvalds * is too small. 17351da177e4SLinus Torvalds */ 17361da177e4SLinus Torvalds if (window <= free_space - mss || window > free_space) 17371da177e4SLinus Torvalds window = (free_space / mss) * mss; 173884565070SJohn Heffner else if (mss == full_space && 1739b92edbe0SEric Dumazet free_space > window + (full_space >> 1)) 174084565070SJohn Heffner window = free_space; 17411da177e4SLinus Torvalds } 17421da177e4SLinus Torvalds 17431da177e4SLinus Torvalds return window; 17441da177e4SLinus Torvalds } 17451da177e4SLinus Torvalds 17464a17fc3aSIlpo Järvinen /* Collapses two adjacent SKB's during retransmission. */ 17474a17fc3aSIlpo Järvinen static void tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb) 17481da177e4SLinus Torvalds { 17491da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 1750fe067e8aSDavid S. Miller struct sk_buff *next_skb = tcp_write_queue_next(sk, skb); 1751058dc334SIlpo Järvinen int skb_size, next_skb_size; 17521da177e4SLinus Torvalds 1753058dc334SIlpo Järvinen skb_size = skb->len; 1754058dc334SIlpo Järvinen next_skb_size = next_skb->len; 17551da177e4SLinus Torvalds 1756058dc334SIlpo Järvinen BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1); 17571da177e4SLinus Torvalds 17586859d494SIlpo Järvinen tcp_highest_sack_combine(sk, next_skb, skb); 1759a6963a6bSIlpo Järvinen 1760fe067e8aSDavid S. Miller tcp_unlink_write_queue(next_skb, sk); 17611da177e4SLinus Torvalds 1762058dc334SIlpo Järvinen skb_copy_from_linear_data(next_skb, skb_put(skb, next_skb_size), 17631a4e2d09SArnaldo Carvalho de Melo next_skb_size); 17641da177e4SLinus Torvalds 176552d570aaSJarek Poplawski if (next_skb->ip_summed == CHECKSUM_PARTIAL) 176652d570aaSJarek Poplawski skb->ip_summed = CHECKSUM_PARTIAL; 17671da177e4SLinus Torvalds 176884fa7933SPatrick McHardy if (skb->ip_summed != CHECKSUM_PARTIAL) 17691da177e4SLinus Torvalds skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size); 17701da177e4SLinus Torvalds 17711da177e4SLinus Torvalds /* Update sequence range on original skb. */ 17721da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq; 17731da177e4SLinus Torvalds 1774e6c7d085SIlpo Järvinen /* Merge over control information. This moves PSH/FIN etc. over */ 1775e6c7d085SIlpo Järvinen TCP_SKB_CB(skb)->flags |= TCP_SKB_CB(next_skb)->flags; 17761da177e4SLinus Torvalds 17771da177e4SLinus Torvalds /* All done, get rid of second SKB and account for it so 17781da177e4SLinus Torvalds * packet counting does not break. 17791da177e4SLinus Torvalds */ 17804828e7f4SIlpo Järvinen TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS; 1781b7689205SIlpo Järvinen 1782b7689205SIlpo Järvinen /* changed transmit queue under us so clear hints */ 1783ef9da47cSIlpo Järvinen tcp_clear_retrans_hints_partial(tp); 1784ef9da47cSIlpo Järvinen if (next_skb == tp->retransmit_skb_hint) 1785ef9da47cSIlpo Järvinen tp->retransmit_skb_hint = skb; 1786b7689205SIlpo Järvinen 1787797108d1SIlpo Järvinen tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb)); 1788797108d1SIlpo Järvinen 17893ab224beSHideo Aoki sk_wmem_free_skb(sk, next_skb); 17901da177e4SLinus Torvalds } 17911da177e4SLinus Torvalds 17924a17fc3aSIlpo Järvinen static int tcp_can_collapse(struct sock *sk, struct sk_buff *skb) 17934a17fc3aSIlpo Järvinen { 17944a17fc3aSIlpo Järvinen if (tcp_skb_pcount(skb) > 1) 17954a17fc3aSIlpo Järvinen return 0; 17964a17fc3aSIlpo Järvinen /* TODO: SACK collapsing could be used to remove this condition */ 17974a17fc3aSIlpo Järvinen if (skb_shinfo(skb)->nr_frags != 0) 17984a17fc3aSIlpo Järvinen return 0; 17994a17fc3aSIlpo Järvinen if (skb_cloned(skb)) 18004a17fc3aSIlpo Järvinen return 0; 18014a17fc3aSIlpo Järvinen if (skb == tcp_send_head(sk)) 18024a17fc3aSIlpo Järvinen return 0; 18034a17fc3aSIlpo Järvinen /* Some heurestics for collapsing over SACK'd could be invented */ 18044a17fc3aSIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 18054a17fc3aSIlpo Järvinen return 0; 18064a17fc3aSIlpo Järvinen 18074a17fc3aSIlpo Järvinen return 1; 18084a17fc3aSIlpo Järvinen } 18094a17fc3aSIlpo Järvinen 18104a17fc3aSIlpo Järvinen static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to, 18114a17fc3aSIlpo Järvinen int space) 18124a17fc3aSIlpo Järvinen { 18134a17fc3aSIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 18144a17fc3aSIlpo Järvinen struct sk_buff *skb = to, *tmp; 18154a17fc3aSIlpo Järvinen int first = 1; 18164a17fc3aSIlpo Järvinen 18174a17fc3aSIlpo Järvinen if (!sysctl_tcp_retrans_collapse) 18184a17fc3aSIlpo Järvinen return; 18194a17fc3aSIlpo Järvinen if (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_SYN) 18204a17fc3aSIlpo Järvinen return; 18214a17fc3aSIlpo Järvinen 18224a17fc3aSIlpo Järvinen tcp_for_write_queue_from_safe(skb, tmp, sk) { 18234a17fc3aSIlpo Järvinen if (!tcp_can_collapse(sk, skb)) 18244a17fc3aSIlpo Järvinen break; 18254a17fc3aSIlpo Järvinen 18264a17fc3aSIlpo Järvinen space -= skb->len; 18274a17fc3aSIlpo Järvinen 18284a17fc3aSIlpo Järvinen if (first) { 18294a17fc3aSIlpo Järvinen first = 0; 18304a17fc3aSIlpo Järvinen continue; 18314a17fc3aSIlpo Järvinen } 18324a17fc3aSIlpo Järvinen 18334a17fc3aSIlpo Järvinen if (space < 0) 18344a17fc3aSIlpo Järvinen break; 18354a17fc3aSIlpo Järvinen /* Punt if not enough space exists in the first SKB for 18364a17fc3aSIlpo Järvinen * the data in the second 18374a17fc3aSIlpo Järvinen */ 18384a17fc3aSIlpo Järvinen if (skb->len > skb_tailroom(to)) 18394a17fc3aSIlpo Järvinen break; 18404a17fc3aSIlpo Järvinen 18414a17fc3aSIlpo Järvinen if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp))) 18424a17fc3aSIlpo Järvinen break; 18434a17fc3aSIlpo Järvinen 18444a17fc3aSIlpo Järvinen tcp_collapse_retrans(sk, to); 18454a17fc3aSIlpo Järvinen } 18464a17fc3aSIlpo Järvinen } 18474a17fc3aSIlpo Järvinen 18481da177e4SLinus Torvalds /* This retransmits one SKB. Policy decisions and retransmit queue 18491da177e4SLinus Torvalds * state updates are done by the caller. Returns non-zero if an 18501da177e4SLinus Torvalds * error occurred which prevented the send. 18511da177e4SLinus Torvalds */ 18521da177e4SLinus Torvalds int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb) 18531da177e4SLinus Torvalds { 18541da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 18555d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 18567d227cd2SSridhar Samudrala unsigned int cur_mss; 18571da177e4SLinus Torvalds int err; 18581da177e4SLinus Torvalds 18595d424d5aSJohn Heffner /* Inconslusive MTU probe */ 18605d424d5aSJohn Heffner if (icsk->icsk_mtup.probe_size) { 18615d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = 0; 18625d424d5aSJohn Heffner } 18635d424d5aSJohn Heffner 18641da177e4SLinus Torvalds /* Do not sent more than we queued. 1/4 is reserved for possible 1865caa20d9aSStephen Hemminger * copying overhead: fragmentation, tunneling, mangling etc. 18661da177e4SLinus Torvalds */ 18671da177e4SLinus Torvalds if (atomic_read(&sk->sk_wmem_alloc) > 18681da177e4SLinus Torvalds min(sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), sk->sk_sndbuf)) 18691da177e4SLinus Torvalds return -EAGAIN; 18701da177e4SLinus Torvalds 18711da177e4SLinus Torvalds if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) { 18721da177e4SLinus Torvalds if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) 18731da177e4SLinus Torvalds BUG(); 18741da177e4SLinus Torvalds if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq)) 18751da177e4SLinus Torvalds return -ENOMEM; 18761da177e4SLinus Torvalds } 18771da177e4SLinus Torvalds 18787d227cd2SSridhar Samudrala if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk)) 18797d227cd2SSridhar Samudrala return -EHOSTUNREACH; /* Routing failure or similar. */ 18807d227cd2SSridhar Samudrala 18810c54b85fSIlpo Järvinen cur_mss = tcp_current_mss(sk); 18827d227cd2SSridhar Samudrala 18831da177e4SLinus Torvalds /* If receiver has shrunk his window, and skb is out of 18841da177e4SLinus Torvalds * new window, do not retransmit it. The exception is the 18851da177e4SLinus Torvalds * case, when window is shrunk to zero. In this case 18861da177e4SLinus Torvalds * our retransmit serves as a zero window probe. 18871da177e4SLinus Torvalds */ 188890840defSIlpo Järvinen if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) 18891da177e4SLinus Torvalds && TCP_SKB_CB(skb)->seq != tp->snd_una) 18901da177e4SLinus Torvalds return -EAGAIN; 18911da177e4SLinus Torvalds 18921da177e4SLinus Torvalds if (skb->len > cur_mss) { 1893846998aeSDavid S. Miller if (tcp_fragment(sk, skb, cur_mss, cur_mss)) 18941da177e4SLinus Torvalds return -ENOMEM; /* We'll try again later. */ 189502276f3cSIlpo Järvinen } else { 18969eb9362eSIlpo Järvinen int oldpcount = tcp_skb_pcount(skb); 18979eb9362eSIlpo Järvinen 18989eb9362eSIlpo Järvinen if (unlikely(oldpcount > 1)) { 189902276f3cSIlpo Järvinen tcp_init_tso_segs(sk, skb, cur_mss); 19009eb9362eSIlpo Järvinen tcp_adjust_pcount(sk, skb, oldpcount - tcp_skb_pcount(skb)); 19019eb9362eSIlpo Järvinen } 19021da177e4SLinus Torvalds } 19031da177e4SLinus Torvalds 19041da177e4SLinus Torvalds tcp_retrans_try_collapse(sk, skb, cur_mss); 19051da177e4SLinus Torvalds 19061da177e4SLinus Torvalds /* Some Solaris stacks overoptimize and ignore the FIN on a 19071da177e4SLinus Torvalds * retransmit when old data is attached. So strip it off 19081da177e4SLinus Torvalds * since it is cheap to do so and saves bytes on the network. 19091da177e4SLinus Torvalds */ 19101da177e4SLinus Torvalds if (skb->len > 0 && 19111da177e4SLinus Torvalds (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN) && 19121da177e4SLinus Torvalds tp->snd_una == (TCP_SKB_CB(skb)->end_seq - 1)) { 19131da177e4SLinus Torvalds if (!pskb_trim(skb, 0)) { 1914e870a8efSIlpo Järvinen /* Reuse, even though it does some unnecessary work */ 1915e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, TCP_SKB_CB(skb)->end_seq - 1, 1916e870a8efSIlpo Järvinen TCP_SKB_CB(skb)->flags); 19171da177e4SLinus Torvalds skb->ip_summed = CHECKSUM_NONE; 19181da177e4SLinus Torvalds } 19191da177e4SLinus Torvalds } 19201da177e4SLinus Torvalds 19211da177e4SLinus Torvalds /* Make a copy, if the first transmission SKB clone we made 19221da177e4SLinus Torvalds * is still in somebody's hands, else make a clone. 19231da177e4SLinus Torvalds */ 19241da177e4SLinus Torvalds TCP_SKB_CB(skb)->when = tcp_time_stamp; 19251da177e4SLinus Torvalds 1926dfb4b9dcSDavid S. Miller err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 19271da177e4SLinus Torvalds 19281da177e4SLinus Torvalds if (err == 0) { 19291da177e4SLinus Torvalds /* Update global TCP statistics. */ 193081cc8a75SPavel Emelyanov TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS); 19311da177e4SLinus Torvalds 19321da177e4SLinus Torvalds tp->total_retrans++; 19331da177e4SLinus Torvalds 19341da177e4SLinus Torvalds #if FASTRETRANS_DEBUG > 0 19351da177e4SLinus Torvalds if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) { 19361da177e4SLinus Torvalds if (net_ratelimit()) 19371da177e4SLinus Torvalds printk(KERN_DEBUG "retrans_out leaked.\n"); 19381da177e4SLinus Torvalds } 19391da177e4SLinus Torvalds #endif 1940b08d6cb2SIlpo Järvinen if (!tp->retrans_out) 1941b08d6cb2SIlpo Järvinen tp->lost_retrans_low = tp->snd_nxt; 19421da177e4SLinus Torvalds TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS; 19431da177e4SLinus Torvalds tp->retrans_out += tcp_skb_pcount(skb); 19441da177e4SLinus Torvalds 19451da177e4SLinus Torvalds /* Save stamp of the first retransmit. */ 19461da177e4SLinus Torvalds if (!tp->retrans_stamp) 19471da177e4SLinus Torvalds tp->retrans_stamp = TCP_SKB_CB(skb)->when; 19481da177e4SLinus Torvalds 19491da177e4SLinus Torvalds tp->undo_retrans++; 19501da177e4SLinus Torvalds 19511da177e4SLinus Torvalds /* snd_nxt is stored to detect loss of retransmitted segment, 19521da177e4SLinus Torvalds * see tcp_input.c tcp_sacktag_write_queue(). 19531da177e4SLinus Torvalds */ 19541da177e4SLinus Torvalds TCP_SKB_CB(skb)->ack_seq = tp->snd_nxt; 19551da177e4SLinus Torvalds } 19561da177e4SLinus Torvalds return err; 19571da177e4SLinus Torvalds } 19581da177e4SLinus Torvalds 1959b5afe7bcSIlpo Järvinen static int tcp_can_forward_retransmit(struct sock *sk) 1960b5afe7bcSIlpo Järvinen { 1961b5afe7bcSIlpo Järvinen const struct inet_connection_sock *icsk = inet_csk(sk); 1962b5afe7bcSIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 1963b5afe7bcSIlpo Järvinen 1964b5afe7bcSIlpo Järvinen /* Forward retransmissions are possible only during Recovery. */ 1965b5afe7bcSIlpo Järvinen if (icsk->icsk_ca_state != TCP_CA_Recovery) 1966b5afe7bcSIlpo Järvinen return 0; 1967b5afe7bcSIlpo Järvinen 1968b5afe7bcSIlpo Järvinen /* No forward retransmissions in Reno are possible. */ 1969b5afe7bcSIlpo Järvinen if (tcp_is_reno(tp)) 1970b5afe7bcSIlpo Järvinen return 0; 1971b5afe7bcSIlpo Järvinen 1972b5afe7bcSIlpo Järvinen /* Yeah, we have to make difficult choice between forward transmission 1973b5afe7bcSIlpo Järvinen * and retransmission... Both ways have their merits... 1974b5afe7bcSIlpo Järvinen * 1975b5afe7bcSIlpo Järvinen * For now we do not retransmit anything, while we have some new 1976b5afe7bcSIlpo Järvinen * segments to send. In the other cases, follow rule 3 for 1977b5afe7bcSIlpo Järvinen * NextSeg() specified in RFC3517. 1978b5afe7bcSIlpo Järvinen */ 1979b5afe7bcSIlpo Järvinen 1980b5afe7bcSIlpo Järvinen if (tcp_may_send_now(sk)) 1981b5afe7bcSIlpo Järvinen return 0; 1982b5afe7bcSIlpo Järvinen 1983b5afe7bcSIlpo Järvinen return 1; 1984b5afe7bcSIlpo Järvinen } 1985b5afe7bcSIlpo Järvinen 19861da177e4SLinus Torvalds /* This gets called after a retransmit timeout, and the initially 19871da177e4SLinus Torvalds * retransmitted data is acknowledged. It tries to continue 19881da177e4SLinus Torvalds * resending the rest of the retransmit queue, until either 19891da177e4SLinus Torvalds * we've sent it all or the congestion window limit is reached. 19901da177e4SLinus Torvalds * If doing SACK, the first ACK which comes back for a timeout 19911da177e4SLinus Torvalds * based retransmit packet might feed us FACK information again. 19921da177e4SLinus Torvalds * If so, we use it to avoid unnecessarily retransmissions. 19931da177e4SLinus Torvalds */ 19941da177e4SLinus Torvalds void tcp_xmit_retransmit_queue(struct sock *sk) 19951da177e4SLinus Torvalds { 19966687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 19971da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 19981da177e4SLinus Torvalds struct sk_buff *skb; 19990e1c54c2SIlpo Järvinen struct sk_buff *hole = NULL; 2000618d9f25SIlpo Järvinen u32 last_lost; 200161eb55f4SIlpo Järvinen int mib_idx; 20020e1c54c2SIlpo Järvinen int fwd_rexmitting = 0; 20036a438bbeSStephen Hemminger 200408ebd172SIlpo Järvinen if (!tp->lost_out) 200508ebd172SIlpo Järvinen tp->retransmit_high = tp->snd_una; 200608ebd172SIlpo Järvinen 2007618d9f25SIlpo Järvinen if (tp->retransmit_skb_hint) { 20086a438bbeSStephen Hemminger skb = tp->retransmit_skb_hint; 2009618d9f25SIlpo Järvinen last_lost = TCP_SKB_CB(skb)->end_seq; 2010618d9f25SIlpo Järvinen if (after(last_lost, tp->retransmit_high)) 2011618d9f25SIlpo Järvinen last_lost = tp->retransmit_high; 2012618d9f25SIlpo Järvinen } else { 2013fe067e8aSDavid S. Miller skb = tcp_write_queue_head(sk); 2014618d9f25SIlpo Järvinen last_lost = tp->snd_una; 2015618d9f25SIlpo Järvinen } 20161da177e4SLinus Torvalds 2017fe067e8aSDavid S. Miller tcp_for_write_queue_from(skb, sk) { 20181da177e4SLinus Torvalds __u8 sacked = TCP_SKB_CB(skb)->sacked; 20191da177e4SLinus Torvalds 2020fe067e8aSDavid S. Miller if (skb == tcp_send_head(sk)) 2021fe067e8aSDavid S. Miller break; 20226a438bbeSStephen Hemminger /* we could do better than to assign each time */ 20230e1c54c2SIlpo Järvinen if (hole == NULL) 20246a438bbeSStephen Hemminger tp->retransmit_skb_hint = skb; 20256a438bbeSStephen Hemminger 20261da177e4SLinus Torvalds /* Assume this retransmit will generate 20271da177e4SLinus Torvalds * only one packet for congestion window 20281da177e4SLinus Torvalds * calculation purposes. This works because 20291da177e4SLinus Torvalds * tcp_retransmit_skb() will chop up the 20301da177e4SLinus Torvalds * packet to be MSS sized and all the 20311da177e4SLinus Torvalds * packet counting works out. 20321da177e4SLinus Torvalds */ 20331da177e4SLinus Torvalds if (tcp_packets_in_flight(tp) >= tp->snd_cwnd) 20341da177e4SLinus Torvalds return; 20350e1c54c2SIlpo Järvinen 20360e1c54c2SIlpo Järvinen if (fwd_rexmitting) { 20370e1c54c2SIlpo Järvinen begin_fwd: 20380e1c54c2SIlpo Järvinen if (!before(TCP_SKB_CB(skb)->seq, tcp_highest_sack_seq(tp))) 2039006f582cSIlpo Järvinen break; 20400e1c54c2SIlpo Järvinen mib_idx = LINUX_MIB_TCPFORWARDRETRANS; 20410e1c54c2SIlpo Järvinen 20420e1c54c2SIlpo Järvinen } else if (!before(TCP_SKB_CB(skb)->seq, tp->retransmit_high)) { 2043618d9f25SIlpo Järvinen tp->retransmit_high = last_lost; 20440e1c54c2SIlpo Järvinen if (!tcp_can_forward_retransmit(sk)) 20450e1c54c2SIlpo Järvinen break; 20460e1c54c2SIlpo Järvinen /* Backtrack if necessary to non-L'ed skb */ 20470e1c54c2SIlpo Järvinen if (hole != NULL) { 20480e1c54c2SIlpo Järvinen skb = hole; 20490e1c54c2SIlpo Järvinen hole = NULL; 20500e1c54c2SIlpo Järvinen } 20510e1c54c2SIlpo Järvinen fwd_rexmitting = 1; 20520e1c54c2SIlpo Järvinen goto begin_fwd; 20530e1c54c2SIlpo Järvinen 20540e1c54c2SIlpo Järvinen } else if (!(sacked & TCPCB_LOST)) { 2055ac11ba75SIlpo Järvinen if (hole == NULL && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED))) 20560e1c54c2SIlpo Järvinen hole = skb; 205761eb55f4SIlpo Järvinen continue; 20581da177e4SLinus Torvalds 20590e1c54c2SIlpo Järvinen } else { 2060618d9f25SIlpo Järvinen last_lost = TCP_SKB_CB(skb)->end_seq; 20610e1c54c2SIlpo Järvinen if (icsk->icsk_ca_state != TCP_CA_Loss) 20620e1c54c2SIlpo Järvinen mib_idx = LINUX_MIB_TCPFASTRETRANS; 20630e1c54c2SIlpo Järvinen else 20640e1c54c2SIlpo Järvinen mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS; 20650e1c54c2SIlpo Järvinen } 20660e1c54c2SIlpo Järvinen 20670e1c54c2SIlpo Järvinen if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS)) 206861eb55f4SIlpo Järvinen continue; 206940b215e5SPavel Emelyanov 2070f0ceb0edSIlpo Järvinen if (tcp_retransmit_skb(sk, skb)) 20711da177e4SLinus Torvalds return; 2072de0744afSPavel Emelyanov NET_INC_STATS_BH(sock_net(sk), mib_idx); 20731da177e4SLinus Torvalds 2074fe067e8aSDavid S. Miller if (skb == tcp_write_queue_head(sk)) 2075463c84b9SArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 20763f421baaSArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto, 20773f421baaSArnaldo Carvalho de Melo TCP_RTO_MAX); 20781da177e4SLinus Torvalds } 20791da177e4SLinus Torvalds } 20801da177e4SLinus Torvalds 20811da177e4SLinus Torvalds /* Send a fin. The caller locks the socket for us. This cannot be 20821da177e4SLinus Torvalds * allowed to fail queueing a FIN frame under any circumstances. 20831da177e4SLinus Torvalds */ 20841da177e4SLinus Torvalds void tcp_send_fin(struct sock *sk) 20851da177e4SLinus Torvalds { 20861da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 2087fe067e8aSDavid S. Miller struct sk_buff *skb = tcp_write_queue_tail(sk); 20881da177e4SLinus Torvalds int mss_now; 20891da177e4SLinus Torvalds 20901da177e4SLinus Torvalds /* Optimization, tack on the FIN if we have a queue of 20911da177e4SLinus Torvalds * unsent frames. But be careful about outgoing SACKS 20921da177e4SLinus Torvalds * and IP options. 20931da177e4SLinus Torvalds */ 20940c54b85fSIlpo Järvinen mss_now = tcp_current_mss(sk); 20951da177e4SLinus Torvalds 2096fe067e8aSDavid S. Miller if (tcp_send_head(sk) != NULL) { 20971da177e4SLinus Torvalds TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_FIN; 20981da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq++; 20991da177e4SLinus Torvalds tp->write_seq++; 21001da177e4SLinus Torvalds } else { 21011da177e4SLinus Torvalds /* Socket is locked, keep trying until memory is available. */ 21021da177e4SLinus Torvalds for (;;) { 2103d179cd12SDavid S. Miller skb = alloc_skb_fclone(MAX_TCP_HEADER, GFP_KERNEL); 21041da177e4SLinus Torvalds if (skb) 21051da177e4SLinus Torvalds break; 21061da177e4SLinus Torvalds yield(); 21071da177e4SLinus Torvalds } 21081da177e4SLinus Torvalds 21091da177e4SLinus Torvalds /* Reserve space for headers and prepare control bits. */ 21101da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 21111da177e4SLinus Torvalds /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */ 2112e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tp->write_seq, 2113e870a8efSIlpo Järvinen TCPCB_FLAG_ACK | TCPCB_FLAG_FIN); 21141da177e4SLinus Torvalds tcp_queue_skb(sk, skb); 21151da177e4SLinus Torvalds } 21169e412ba7SIlpo Järvinen __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_OFF); 21171da177e4SLinus Torvalds } 21181da177e4SLinus Torvalds 21191da177e4SLinus Torvalds /* We get here when a process closes a file descriptor (either due to 21201da177e4SLinus Torvalds * an explicit close() or as a byproduct of exit()'ing) and there 21211da177e4SLinus Torvalds * was unread data in the receive queue. This behavior is recommended 212265bb723cSGerrit Renker * by RFC 2525, section 2.17. -DaveM 21231da177e4SLinus Torvalds */ 2124dd0fc66fSAl Viro void tcp_send_active_reset(struct sock *sk, gfp_t priority) 21251da177e4SLinus Torvalds { 21261da177e4SLinus Torvalds struct sk_buff *skb; 21271da177e4SLinus Torvalds 21281da177e4SLinus Torvalds /* NOTE: No TCP options attached and we never retransmit this. */ 21291da177e4SLinus Torvalds skb = alloc_skb(MAX_TCP_HEADER, priority); 21301da177e4SLinus Torvalds if (!skb) { 21314e673444SPavel Emelyanov NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); 21321da177e4SLinus Torvalds return; 21331da177e4SLinus Torvalds } 21341da177e4SLinus Torvalds 21351da177e4SLinus Torvalds /* Reserve space for headers and prepare control bits. */ 21361da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 2137e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk), 2138e870a8efSIlpo Järvinen TCPCB_FLAG_ACK | TCPCB_FLAG_RST); 21391da177e4SLinus Torvalds /* Send it off. */ 21401da177e4SLinus Torvalds TCP_SKB_CB(skb)->when = tcp_time_stamp; 2141dfb4b9dcSDavid S. Miller if (tcp_transmit_skb(sk, skb, 0, priority)) 21424e673444SPavel Emelyanov NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); 214326af65cbSSridhar Samudrala 214481cc8a75SPavel Emelyanov TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS); 21451da177e4SLinus Torvalds } 21461da177e4SLinus Torvalds 21471da177e4SLinus Torvalds /* WARNING: This routine must only be called when we have already sent 21481da177e4SLinus Torvalds * a SYN packet that crossed the incoming SYN that caused this routine 21491da177e4SLinus Torvalds * to get called. If this assumption fails then the initial rcv_wnd 21501da177e4SLinus Torvalds * and rcv_wscale values will not be correct. 21511da177e4SLinus Torvalds */ 21521da177e4SLinus Torvalds int tcp_send_synack(struct sock *sk) 21531da177e4SLinus Torvalds { 21541da177e4SLinus Torvalds struct sk_buff *skb; 21551da177e4SLinus Torvalds 2156fe067e8aSDavid S. Miller skb = tcp_write_queue_head(sk); 21571da177e4SLinus Torvalds if (skb == NULL || !(TCP_SKB_CB(skb)->flags & TCPCB_FLAG_SYN)) { 21581da177e4SLinus Torvalds printk(KERN_DEBUG "tcp_send_synack: wrong queue state\n"); 21591da177e4SLinus Torvalds return -EFAULT; 21601da177e4SLinus Torvalds } 21611da177e4SLinus Torvalds if (!(TCP_SKB_CB(skb)->flags & TCPCB_FLAG_ACK)) { 21621da177e4SLinus Torvalds if (skb_cloned(skb)) { 21631da177e4SLinus Torvalds struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC); 21641da177e4SLinus Torvalds if (nskb == NULL) 21651da177e4SLinus Torvalds return -ENOMEM; 2166fe067e8aSDavid S. Miller tcp_unlink_write_queue(skb, sk); 21671da177e4SLinus Torvalds skb_header_release(nskb); 2168fe067e8aSDavid S. Miller __tcp_add_write_queue_head(sk, nskb); 21693ab224beSHideo Aoki sk_wmem_free_skb(sk, skb); 21703ab224beSHideo Aoki sk->sk_wmem_queued += nskb->truesize; 21713ab224beSHideo Aoki sk_mem_charge(sk, nskb->truesize); 21721da177e4SLinus Torvalds skb = nskb; 21731da177e4SLinus Torvalds } 21741da177e4SLinus Torvalds 21751da177e4SLinus Torvalds TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_ACK; 21761da177e4SLinus Torvalds TCP_ECN_send_synack(tcp_sk(sk), skb); 21771da177e4SLinus Torvalds } 21781da177e4SLinus Torvalds TCP_SKB_CB(skb)->when = tcp_time_stamp; 2179dfb4b9dcSDavid S. Miller return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 21801da177e4SLinus Torvalds } 21811da177e4SLinus Torvalds 21821da177e4SLinus Torvalds /* 21831da177e4SLinus Torvalds * Prepare a SYN-ACK. 21841da177e4SLinus Torvalds */ 21851da177e4SLinus Torvalds struct sk_buff *tcp_make_synack(struct sock *sk, struct dst_entry *dst, 218660236fddSArnaldo Carvalho de Melo struct request_sock *req) 21871da177e4SLinus Torvalds { 21882e6599cbSArnaldo Carvalho de Melo struct inet_request_sock *ireq = inet_rsk(req); 21891da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 21901da177e4SLinus Torvalds struct tcphdr *th; 21911da177e4SLinus Torvalds int tcp_header_size; 219233ad798cSAdam Langley struct tcp_out_options opts; 21931da177e4SLinus Torvalds struct sk_buff *skb; 2194cfb6eeb4SYOSHIFUJI Hideaki struct tcp_md5sig_key *md5; 2195cfb6eeb4SYOSHIFUJI Hideaki __u8 *md5_hash_location; 2196f5fff5dcSTom Quetchenbach int mss; 21971da177e4SLinus Torvalds 21981da177e4SLinus Torvalds skb = sock_wmalloc(sk, MAX_TCP_HEADER + 15, 1, GFP_ATOMIC); 21991da177e4SLinus Torvalds if (skb == NULL) 22001da177e4SLinus Torvalds return NULL; 22011da177e4SLinus Torvalds 22021da177e4SLinus Torvalds /* Reserve space for headers. */ 22031da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 22041da177e4SLinus Torvalds 22051da177e4SLinus Torvalds skb->dst = dst_clone(dst); 22061da177e4SLinus Torvalds 2207f5fff5dcSTom Quetchenbach mss = dst_metric(dst, RTAX_ADVMSS); 2208f5fff5dcSTom Quetchenbach if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < mss) 2209f5fff5dcSTom Quetchenbach mss = tp->rx_opt.user_mss; 2210f5fff5dcSTom Quetchenbach 221133ad798cSAdam Langley if (req->rcv_wnd == 0) { /* ignored for retransmitted syns */ 221233ad798cSAdam Langley __u8 rcv_wscale; 221333ad798cSAdam Langley /* Set this up on the first call only */ 221433ad798cSAdam Langley req->window_clamp = tp->window_clamp ? : dst_metric(dst, RTAX_WINDOW); 221533ad798cSAdam Langley /* tcp_full_space because it is guaranteed to be the first packet */ 221633ad798cSAdam Langley tcp_select_initial_window(tcp_full_space(sk), 2217f5fff5dcSTom Quetchenbach mss - (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0), 221833ad798cSAdam Langley &req->rcv_wnd, 221933ad798cSAdam Langley &req->window_clamp, 222033ad798cSAdam Langley ireq->wscale_ok, 222133ad798cSAdam Langley &rcv_wscale); 222233ad798cSAdam Langley ireq->rcv_wscale = rcv_wscale; 222333ad798cSAdam Langley } 2224cfb6eeb4SYOSHIFUJI Hideaki 222533ad798cSAdam Langley memset(&opts, 0, sizeof(opts)); 22268b5f12d0SFlorian Westphal #ifdef CONFIG_SYN_COOKIES 22278b5f12d0SFlorian Westphal if (unlikely(req->cookie_ts)) 22288b5f12d0SFlorian Westphal TCP_SKB_CB(skb)->when = cookie_init_timestamp(req); 22298b5f12d0SFlorian Westphal else 22308b5f12d0SFlorian Westphal #endif 223133ad798cSAdam Langley TCP_SKB_CB(skb)->when = tcp_time_stamp; 2232f5fff5dcSTom Quetchenbach tcp_header_size = tcp_synack_options(sk, req, mss, 223333ad798cSAdam Langley skb, &opts, &md5) + 223433ad798cSAdam Langley sizeof(struct tcphdr); 223533ad798cSAdam Langley 2236aa8223c7SArnaldo Carvalho de Melo skb_push(skb, tcp_header_size); 2237aa8223c7SArnaldo Carvalho de Melo skb_reset_transport_header(skb); 22381da177e4SLinus Torvalds 2239aa8223c7SArnaldo Carvalho de Melo th = tcp_hdr(skb); 22401da177e4SLinus Torvalds memset(th, 0, sizeof(struct tcphdr)); 22411da177e4SLinus Torvalds th->syn = 1; 22421da177e4SLinus Torvalds th->ack = 1; 22431da177e4SLinus Torvalds TCP_ECN_make_synack(req, th); 2244a3116ac5SKOVACS Krisztian th->source = ireq->loc_port; 22452e6599cbSArnaldo Carvalho de Melo th->dest = ireq->rmt_port; 2246e870a8efSIlpo Järvinen /* Setting of flags are superfluous here for callers (and ECE is 2247e870a8efSIlpo Järvinen * not even correctly set) 2248e870a8efSIlpo Järvinen */ 2249e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tcp_rsk(req)->snt_isn, 2250e870a8efSIlpo Järvinen TCPCB_FLAG_SYN | TCPCB_FLAG_ACK); 22511da177e4SLinus Torvalds th->seq = htonl(TCP_SKB_CB(skb)->seq); 22522e6599cbSArnaldo Carvalho de Melo th->ack_seq = htonl(tcp_rsk(req)->rcv_isn + 1); 22531da177e4SLinus Torvalds 22541da177e4SLinus Torvalds /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */ 2255600ff0c2SIlpo Järvinen th->window = htons(min(req->rcv_wnd, 65535U)); 225633ad798cSAdam Langley tcp_options_write((__be32 *)(th + 1), tp, &opts, &md5_hash_location); 22571da177e4SLinus Torvalds th->doff = (tcp_header_size >> 2); 225881cc8a75SPavel Emelyanov TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTSEGS); 2259cfb6eeb4SYOSHIFUJI Hideaki 2260cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 2261cfb6eeb4SYOSHIFUJI Hideaki /* Okay, we have all we need - do the md5 hash if needed */ 2262cfb6eeb4SYOSHIFUJI Hideaki if (md5) { 2263cfb6eeb4SYOSHIFUJI Hideaki tp->af_specific->calc_md5_hash(md5_hash_location, 226449a72dfbSAdam Langley md5, NULL, req, skb); 2265cfb6eeb4SYOSHIFUJI Hideaki } 2266cfb6eeb4SYOSHIFUJI Hideaki #endif 2267cfb6eeb4SYOSHIFUJI Hideaki 22681da177e4SLinus Torvalds return skb; 22691da177e4SLinus Torvalds } 22701da177e4SLinus Torvalds 22711da177e4SLinus Torvalds /* 22721da177e4SLinus Torvalds * Do all connect socket setups that can be done AF independent. 22731da177e4SLinus Torvalds */ 227440efc6faSStephen Hemminger static void tcp_connect_init(struct sock *sk) 22751da177e4SLinus Torvalds { 22761da177e4SLinus Torvalds struct dst_entry *dst = __sk_dst_get(sk); 22771da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 22781da177e4SLinus Torvalds __u8 rcv_wscale; 22791da177e4SLinus Torvalds 22801da177e4SLinus Torvalds /* We'll fix this up when we get a response from the other end. 22811da177e4SLinus Torvalds * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT. 22821da177e4SLinus Torvalds */ 22831da177e4SLinus Torvalds tp->tcp_header_len = sizeof(struct tcphdr) + 22841da177e4SLinus Torvalds (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0); 22851da177e4SLinus Torvalds 2286cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 2287cfb6eeb4SYOSHIFUJI Hideaki if (tp->af_specific->md5_lookup(sk, sk) != NULL) 2288cfb6eeb4SYOSHIFUJI Hideaki tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED; 2289cfb6eeb4SYOSHIFUJI Hideaki #endif 2290cfb6eeb4SYOSHIFUJI Hideaki 22911da177e4SLinus Torvalds /* If user gave his TCP_MAXSEG, record it to clamp */ 22921da177e4SLinus Torvalds if (tp->rx_opt.user_mss) 22931da177e4SLinus Torvalds tp->rx_opt.mss_clamp = tp->rx_opt.user_mss; 22941da177e4SLinus Torvalds tp->max_window = 0; 22955d424d5aSJohn Heffner tcp_mtup_init(sk); 22961da177e4SLinus Torvalds tcp_sync_mss(sk, dst_mtu(dst)); 22971da177e4SLinus Torvalds 22981da177e4SLinus Torvalds if (!tp->window_clamp) 22991da177e4SLinus Torvalds tp->window_clamp = dst_metric(dst, RTAX_WINDOW); 23001da177e4SLinus Torvalds tp->advmss = dst_metric(dst, RTAX_ADVMSS); 2301f5fff5dcSTom Quetchenbach if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->advmss) 2302f5fff5dcSTom Quetchenbach tp->advmss = tp->rx_opt.user_mss; 2303f5fff5dcSTom Quetchenbach 23041da177e4SLinus Torvalds tcp_initialize_rcv_mss(sk); 23051da177e4SLinus Torvalds 23061da177e4SLinus Torvalds tcp_select_initial_window(tcp_full_space(sk), 23071da177e4SLinus Torvalds tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0), 23081da177e4SLinus Torvalds &tp->rcv_wnd, 23091da177e4SLinus Torvalds &tp->window_clamp, 23101da177e4SLinus Torvalds sysctl_tcp_window_scaling, 23111da177e4SLinus Torvalds &rcv_wscale); 23121da177e4SLinus Torvalds 23131da177e4SLinus Torvalds tp->rx_opt.rcv_wscale = rcv_wscale; 23141da177e4SLinus Torvalds tp->rcv_ssthresh = tp->rcv_wnd; 23151da177e4SLinus Torvalds 23161da177e4SLinus Torvalds sk->sk_err = 0; 23171da177e4SLinus Torvalds sock_reset_flag(sk, SOCK_DONE); 23181da177e4SLinus Torvalds tp->snd_wnd = 0; 2319ee7537b6SHantzis Fotis tcp_init_wl(tp, 0); 23201da177e4SLinus Torvalds tp->snd_una = tp->write_seq; 23211da177e4SLinus Torvalds tp->snd_sml = tp->write_seq; 232233f5f57eSIlpo Järvinen tp->snd_up = tp->write_seq; 23231da177e4SLinus Torvalds tp->rcv_nxt = 0; 23241da177e4SLinus Torvalds tp->rcv_wup = 0; 23251da177e4SLinus Torvalds tp->copied_seq = 0; 23261da177e4SLinus Torvalds 2327463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT; 2328463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_retransmits = 0; 23291da177e4SLinus Torvalds tcp_clear_retrans(tp); 23301da177e4SLinus Torvalds } 23311da177e4SLinus Torvalds 23321da177e4SLinus Torvalds /* 23331da177e4SLinus Torvalds * Build a SYN and send it off. 23341da177e4SLinus Torvalds */ 23351da177e4SLinus Torvalds int tcp_connect(struct sock *sk) 23361da177e4SLinus Torvalds { 23371da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 23381da177e4SLinus Torvalds struct sk_buff *buff; 23391da177e4SLinus Torvalds 23401da177e4SLinus Torvalds tcp_connect_init(sk); 23411da177e4SLinus Torvalds 2342d179cd12SDavid S. Miller buff = alloc_skb_fclone(MAX_TCP_HEADER + 15, sk->sk_allocation); 23431da177e4SLinus Torvalds if (unlikely(buff == NULL)) 23441da177e4SLinus Torvalds return -ENOBUFS; 23451da177e4SLinus Torvalds 23461da177e4SLinus Torvalds /* Reserve space for headers. */ 23471da177e4SLinus Torvalds skb_reserve(buff, MAX_TCP_HEADER); 23481da177e4SLinus Torvalds 2349bd37a088SWei Yongjun tp->snd_nxt = tp->write_seq; 2350e870a8efSIlpo Järvinen tcp_init_nondata_skb(buff, tp->write_seq++, TCPCB_FLAG_SYN); 2351e870a8efSIlpo Järvinen TCP_ECN_send_syn(sk, buff); 23521da177e4SLinus Torvalds 23531da177e4SLinus Torvalds /* Send it off. */ 23541da177e4SLinus Torvalds TCP_SKB_CB(buff)->when = tcp_time_stamp; 23551da177e4SLinus Torvalds tp->retrans_stamp = TCP_SKB_CB(buff)->when; 23561da177e4SLinus Torvalds skb_header_release(buff); 2357fe067e8aSDavid S. Miller __tcp_add_write_queue_tail(sk, buff); 23583ab224beSHideo Aoki sk->sk_wmem_queued += buff->truesize; 23593ab224beSHideo Aoki sk_mem_charge(sk, buff->truesize); 23601da177e4SLinus Torvalds tp->packets_out += tcp_skb_pcount(buff); 2361dfb4b9dcSDavid S. Miller tcp_transmit_skb(sk, buff, 1, GFP_KERNEL); 2362bd37a088SWei Yongjun 2363bd37a088SWei Yongjun /* We change tp->snd_nxt after the tcp_transmit_skb() call 2364bd37a088SWei Yongjun * in order to make this packet get counted in tcpOutSegs. 2365bd37a088SWei Yongjun */ 2366bd37a088SWei Yongjun tp->snd_nxt = tp->write_seq; 2367bd37a088SWei Yongjun tp->pushed_seq = tp->write_seq; 236881cc8a75SPavel Emelyanov TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS); 23691da177e4SLinus Torvalds 23701da177e4SLinus Torvalds /* Timer for repeating the SYN until an answer. */ 23713f421baaSArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 23723f421baaSArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto, TCP_RTO_MAX); 23731da177e4SLinus Torvalds return 0; 23741da177e4SLinus Torvalds } 23751da177e4SLinus Torvalds 23761da177e4SLinus Torvalds /* Send out a delayed ack, the caller does the policy checking 23771da177e4SLinus Torvalds * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check() 23781da177e4SLinus Torvalds * for details. 23791da177e4SLinus Torvalds */ 23801da177e4SLinus Torvalds void tcp_send_delayed_ack(struct sock *sk) 23811da177e4SLinus Torvalds { 2382463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 2383463c84b9SArnaldo Carvalho de Melo int ato = icsk->icsk_ack.ato; 23841da177e4SLinus Torvalds unsigned long timeout; 23851da177e4SLinus Torvalds 23861da177e4SLinus Torvalds if (ato > TCP_DELACK_MIN) { 2387463c84b9SArnaldo Carvalho de Melo const struct tcp_sock *tp = tcp_sk(sk); 23881da177e4SLinus Torvalds int max_ato = HZ / 2; 23891da177e4SLinus Torvalds 2390056834d9SIlpo Järvinen if (icsk->icsk_ack.pingpong || 2391056834d9SIlpo Järvinen (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)) 23921da177e4SLinus Torvalds max_ato = TCP_DELACK_MAX; 23931da177e4SLinus Torvalds 23941da177e4SLinus Torvalds /* Slow path, intersegment interval is "high". */ 23951da177e4SLinus Torvalds 23961da177e4SLinus Torvalds /* If some rtt estimate is known, use it to bound delayed ack. 2397463c84b9SArnaldo Carvalho de Melo * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements 23981da177e4SLinus Torvalds * directly. 23991da177e4SLinus Torvalds */ 24001da177e4SLinus Torvalds if (tp->srtt) { 24011da177e4SLinus Torvalds int rtt = max(tp->srtt >> 3, TCP_DELACK_MIN); 24021da177e4SLinus Torvalds 24031da177e4SLinus Torvalds if (rtt < max_ato) 24041da177e4SLinus Torvalds max_ato = rtt; 24051da177e4SLinus Torvalds } 24061da177e4SLinus Torvalds 24071da177e4SLinus Torvalds ato = min(ato, max_ato); 24081da177e4SLinus Torvalds } 24091da177e4SLinus Torvalds 24101da177e4SLinus Torvalds /* Stay within the limit we were given */ 24111da177e4SLinus Torvalds timeout = jiffies + ato; 24121da177e4SLinus Torvalds 24131da177e4SLinus Torvalds /* Use new timeout only if there wasn't a older one earlier. */ 2414463c84b9SArnaldo Carvalho de Melo if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) { 24151da177e4SLinus Torvalds /* If delack timer was blocked or is about to expire, 24161da177e4SLinus Torvalds * send ACK now. 24171da177e4SLinus Torvalds */ 2418463c84b9SArnaldo Carvalho de Melo if (icsk->icsk_ack.blocked || 2419463c84b9SArnaldo Carvalho de Melo time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) { 24201da177e4SLinus Torvalds tcp_send_ack(sk); 24211da177e4SLinus Torvalds return; 24221da177e4SLinus Torvalds } 24231da177e4SLinus Torvalds 2424463c84b9SArnaldo Carvalho de Melo if (!time_before(timeout, icsk->icsk_ack.timeout)) 2425463c84b9SArnaldo Carvalho de Melo timeout = icsk->icsk_ack.timeout; 24261da177e4SLinus Torvalds } 2427463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER; 2428463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.timeout = timeout; 2429463c84b9SArnaldo Carvalho de Melo sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout); 24301da177e4SLinus Torvalds } 24311da177e4SLinus Torvalds 24321da177e4SLinus Torvalds /* This routine sends an ack and also updates the window. */ 24331da177e4SLinus Torvalds void tcp_send_ack(struct sock *sk) 24341da177e4SLinus Torvalds { 24351da177e4SLinus Torvalds struct sk_buff *buff; 24361da177e4SLinus Torvalds 2437058dc334SIlpo Järvinen /* If we have been reset, we may not send again. */ 2438058dc334SIlpo Järvinen if (sk->sk_state == TCP_CLOSE) 2439058dc334SIlpo Järvinen return; 2440058dc334SIlpo Järvinen 24411da177e4SLinus Torvalds /* We are not putting this on the write queue, so 24421da177e4SLinus Torvalds * tcp_transmit_skb() will set the ownership to this 24431da177e4SLinus Torvalds * sock. 24441da177e4SLinus Torvalds */ 24451da177e4SLinus Torvalds buff = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC); 24461da177e4SLinus Torvalds if (buff == NULL) { 2447463c84b9SArnaldo Carvalho de Melo inet_csk_schedule_ack(sk); 2448463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN; 24493f421baaSArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK, 24503f421baaSArnaldo Carvalho de Melo TCP_DELACK_MAX, TCP_RTO_MAX); 24511da177e4SLinus Torvalds return; 24521da177e4SLinus Torvalds } 24531da177e4SLinus Torvalds 24541da177e4SLinus Torvalds /* Reserve space for headers and prepare control bits. */ 24551da177e4SLinus Torvalds skb_reserve(buff, MAX_TCP_HEADER); 2456e870a8efSIlpo Järvinen tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPCB_FLAG_ACK); 24571da177e4SLinus Torvalds 24581da177e4SLinus Torvalds /* Send it off, this clears delayed acks for us. */ 24591da177e4SLinus Torvalds TCP_SKB_CB(buff)->when = tcp_time_stamp; 2460dfb4b9dcSDavid S. Miller tcp_transmit_skb(sk, buff, 0, GFP_ATOMIC); 24611da177e4SLinus Torvalds } 24621da177e4SLinus Torvalds 24631da177e4SLinus Torvalds /* This routine sends a packet with an out of date sequence 24641da177e4SLinus Torvalds * number. It assumes the other end will try to ack it. 24651da177e4SLinus Torvalds * 24661da177e4SLinus Torvalds * Question: what should we make while urgent mode? 24671da177e4SLinus Torvalds * 4.4BSD forces sending single byte of data. We cannot send 24681da177e4SLinus Torvalds * out of window data, because we have SND.NXT==SND.MAX... 24691da177e4SLinus Torvalds * 24701da177e4SLinus Torvalds * Current solution: to send TWO zero-length segments in urgent mode: 24711da177e4SLinus Torvalds * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is 24721da177e4SLinus Torvalds * out-of-date with SND.UNA-1 to probe window. 24731da177e4SLinus Torvalds */ 24741da177e4SLinus Torvalds static int tcp_xmit_probe_skb(struct sock *sk, int urgent) 24751da177e4SLinus Torvalds { 24761da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 24771da177e4SLinus Torvalds struct sk_buff *skb; 24781da177e4SLinus Torvalds 24791da177e4SLinus Torvalds /* We don't queue it, tcp_transmit_skb() sets ownership. */ 24801da177e4SLinus Torvalds skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC); 24811da177e4SLinus Torvalds if (skb == NULL) 24821da177e4SLinus Torvalds return -1; 24831da177e4SLinus Torvalds 24841da177e4SLinus Torvalds /* Reserve space for headers and set control bits. */ 24851da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 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 */ 2490e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPCB_FLAG_ACK); 24911da177e4SLinus Torvalds TCP_SKB_CB(skb)->when = tcp_time_stamp; 2492dfb4b9dcSDavid S. Miller return tcp_transmit_skb(sk, skb, 0, GFP_ATOMIC); 24931da177e4SLinus Torvalds } 24941da177e4SLinus Torvalds 24951da177e4SLinus Torvalds int tcp_write_wakeup(struct sock *sk) 24961da177e4SLinus Torvalds { 24971da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 24981da177e4SLinus Torvalds struct sk_buff *skb; 24991da177e4SLinus Torvalds 2500058dc334SIlpo Järvinen if (sk->sk_state == TCP_CLOSE) 2501058dc334SIlpo Järvinen return -1; 2502058dc334SIlpo Järvinen 2503fe067e8aSDavid S. Miller if ((skb = tcp_send_head(sk)) != NULL && 250490840defSIlpo Järvinen before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) { 25051da177e4SLinus Torvalds int err; 25060c54b85fSIlpo Järvinen unsigned int mss = tcp_current_mss(sk); 250790840defSIlpo Järvinen unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 25081da177e4SLinus Torvalds 25091da177e4SLinus Torvalds if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq)) 25101da177e4SLinus Torvalds tp->pushed_seq = TCP_SKB_CB(skb)->end_seq; 25111da177e4SLinus Torvalds 25121da177e4SLinus Torvalds /* We are probing the opening of a window 25131da177e4SLinus Torvalds * but the window size is != 0 25141da177e4SLinus Torvalds * must have been a result SWS avoidance ( sender ) 25151da177e4SLinus Torvalds */ 25161da177e4SLinus Torvalds if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq || 25171da177e4SLinus Torvalds skb->len > mss) { 25181da177e4SLinus Torvalds seg_size = min(seg_size, mss); 25191da177e4SLinus Torvalds TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH; 2520846998aeSDavid S. Miller if (tcp_fragment(sk, skb, seg_size, mss)) 25211da177e4SLinus Torvalds return -1; 25221da177e4SLinus Torvalds } else if (!tcp_skb_pcount(skb)) 2523846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, skb, mss); 25241da177e4SLinus Torvalds 25251da177e4SLinus Torvalds TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH; 25261da177e4SLinus Torvalds TCP_SKB_CB(skb)->when = tcp_time_stamp; 2527dfb4b9dcSDavid S. Miller err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 252866f5fe62SIlpo Järvinen if (!err) 252966f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, skb); 25301da177e4SLinus Torvalds return err; 25311da177e4SLinus Torvalds } else { 253233f5f57eSIlpo Järvinen if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF)) 25334828e7f4SIlpo Järvinen tcp_xmit_probe_skb(sk, 1); 25341da177e4SLinus Torvalds return tcp_xmit_probe_skb(sk, 0); 25351da177e4SLinus Torvalds } 25361da177e4SLinus Torvalds } 25371da177e4SLinus Torvalds 25381da177e4SLinus Torvalds /* A window probe timeout has occurred. If window is not closed send 25391da177e4SLinus Torvalds * a partial packet else a zero probe. 25401da177e4SLinus Torvalds */ 25411da177e4SLinus Torvalds void tcp_send_probe0(struct sock *sk) 25421da177e4SLinus Torvalds { 2543463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 25441da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 25451da177e4SLinus Torvalds int err; 25461da177e4SLinus Torvalds 25471da177e4SLinus Torvalds err = tcp_write_wakeup(sk); 25481da177e4SLinus Torvalds 2549fe067e8aSDavid S. Miller if (tp->packets_out || !tcp_send_head(sk)) { 25501da177e4SLinus Torvalds /* Cancel probe timer, if it is not required. */ 25516687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out = 0; 2552463c84b9SArnaldo Carvalho de Melo icsk->icsk_backoff = 0; 25531da177e4SLinus Torvalds return; 25541da177e4SLinus Torvalds } 25551da177e4SLinus Torvalds 25561da177e4SLinus Torvalds if (err <= 0) { 2557463c84b9SArnaldo Carvalho de Melo if (icsk->icsk_backoff < sysctl_tcp_retries2) 2558463c84b9SArnaldo Carvalho de Melo icsk->icsk_backoff++; 25596687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out++; 2560463c84b9SArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0, 25613f421baaSArnaldo Carvalho de Melo min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX), 25623f421baaSArnaldo Carvalho de Melo TCP_RTO_MAX); 25631da177e4SLinus Torvalds } else { 25641da177e4SLinus Torvalds /* If packet was not sent due to local congestion, 25656687e988SArnaldo Carvalho de Melo * do not backoff and do not remember icsk_probes_out. 25661da177e4SLinus Torvalds * Let local senders to fight for local resources. 25671da177e4SLinus Torvalds * 25681da177e4SLinus Torvalds * Use accumulated backoff yet. 25691da177e4SLinus Torvalds */ 25706687e988SArnaldo Carvalho de Melo if (!icsk->icsk_probes_out) 25716687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out = 1; 2572463c84b9SArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0, 2573463c84b9SArnaldo Carvalho de Melo min(icsk->icsk_rto << icsk->icsk_backoff, 25743f421baaSArnaldo Carvalho de Melo TCP_RESOURCE_PROBE_INTERVAL), 25753f421baaSArnaldo Carvalho de Melo TCP_RTO_MAX); 25761da177e4SLinus Torvalds } 25771da177e4SLinus Torvalds } 25781da177e4SLinus Torvalds 2579c6aefafbSGlenn Griffin EXPORT_SYMBOL(tcp_select_initial_window); 25801da177e4SLinus Torvalds EXPORT_SYMBOL(tcp_connect); 25811da177e4SLinus Torvalds EXPORT_SYMBOL(tcp_make_synack); 25821da177e4SLinus Torvalds EXPORT_SYMBOL(tcp_simple_retransmit); 25831da177e4SLinus Torvalds EXPORT_SYMBOL(tcp_sync_mss); 25845d424d5aSJohn Heffner EXPORT_SYMBOL(tcp_mtup_init); 2585