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 444cabeccbdSIlpo Järvinen if (tp->rx_opt.dsack) 44540efc6faSStephen Hemminger tp->rx_opt.dsack = 0; 44640efc6faSStephen Hemminger } 44740efc6faSStephen Hemminger } 44840efc6faSStephen Hemminger 44933ad798cSAdam Langley static unsigned tcp_syn_options(struct sock *sk, struct sk_buff *skb, 45033ad798cSAdam Langley struct tcp_out_options *opts, 45133ad798cSAdam Langley struct tcp_md5sig_key **md5) { 45233ad798cSAdam Langley struct tcp_sock *tp = tcp_sk(sk); 45333ad798cSAdam Langley unsigned size = 0; 45433ad798cSAdam Langley 455cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 45633ad798cSAdam Langley *md5 = tp->af_specific->md5_lookup(sk, sk); 45733ad798cSAdam Langley if (*md5) { 45833ad798cSAdam Langley opts->options |= OPTION_MD5; 45933ad798cSAdam Langley size += TCPOLEN_MD5SIG_ALIGNED; 460cfb6eeb4SYOSHIFUJI Hideaki } 46133ad798cSAdam Langley #else 46233ad798cSAdam Langley *md5 = NULL; 463cfb6eeb4SYOSHIFUJI Hideaki #endif 46433ad798cSAdam Langley 46533ad798cSAdam Langley /* We always get an MSS option. The option bytes which will be seen in 46633ad798cSAdam Langley * normal data packets should timestamps be used, must be in the MSS 46733ad798cSAdam Langley * advertised. But we subtract them from tp->mss_cache so that 46833ad798cSAdam Langley * calculations in tcp_sendmsg are simpler etc. So account for this 46933ad798cSAdam Langley * fact here if necessary. If we don't do this correctly, as a 47033ad798cSAdam Langley * receiver we won't recognize data packets as being full sized when we 47133ad798cSAdam Langley * should, and thus we won't abide by the delayed ACK rules correctly. 47233ad798cSAdam Langley * SACKs don't matter, we never delay an ACK when we have any of those 47333ad798cSAdam Langley * going out. */ 47433ad798cSAdam Langley opts->mss = tcp_advertise_mss(sk); 47533ad798cSAdam Langley size += TCPOLEN_MSS_ALIGNED; 47633ad798cSAdam Langley 47733ad798cSAdam Langley if (likely(sysctl_tcp_timestamps && *md5 == NULL)) { 47833ad798cSAdam Langley opts->options |= OPTION_TS; 47933ad798cSAdam Langley opts->tsval = TCP_SKB_CB(skb)->when; 48033ad798cSAdam Langley opts->tsecr = tp->rx_opt.ts_recent; 48133ad798cSAdam Langley size += TCPOLEN_TSTAMP_ALIGNED; 48233ad798cSAdam Langley } 48333ad798cSAdam Langley if (likely(sysctl_tcp_window_scaling)) { 48433ad798cSAdam Langley opts->ws = tp->rx_opt.rcv_wscale; 4857982d5e1SPhilip Love if (likely(opts->ws)) 48633ad798cSAdam Langley size += TCPOLEN_WSCALE_ALIGNED; 48733ad798cSAdam Langley } 48833ad798cSAdam Langley if (likely(sysctl_tcp_sack)) { 48933ad798cSAdam Langley opts->options |= OPTION_SACK_ADVERTISE; 490b32d1310SDavid S. Miller if (unlikely(!(OPTION_TS & opts->options))) 49133ad798cSAdam Langley size += TCPOLEN_SACKPERM_ALIGNED; 49233ad798cSAdam Langley } 49333ad798cSAdam Langley 49433ad798cSAdam Langley return size; 49533ad798cSAdam Langley } 49633ad798cSAdam Langley 49733ad798cSAdam Langley static unsigned tcp_synack_options(struct sock *sk, 49833ad798cSAdam Langley struct request_sock *req, 49933ad798cSAdam Langley unsigned mss, struct sk_buff *skb, 50033ad798cSAdam Langley struct tcp_out_options *opts, 50133ad798cSAdam Langley struct tcp_md5sig_key **md5) { 50233ad798cSAdam Langley unsigned size = 0; 50333ad798cSAdam Langley struct inet_request_sock *ireq = inet_rsk(req); 50433ad798cSAdam Langley char doing_ts; 50533ad798cSAdam Langley 50633ad798cSAdam Langley #ifdef CONFIG_TCP_MD5SIG 50733ad798cSAdam Langley *md5 = tcp_rsk(req)->af_specific->md5_lookup(sk, req); 50833ad798cSAdam Langley if (*md5) { 50933ad798cSAdam Langley opts->options |= OPTION_MD5; 51033ad798cSAdam Langley size += TCPOLEN_MD5SIG_ALIGNED; 51133ad798cSAdam Langley } 51233ad798cSAdam Langley #else 51333ad798cSAdam Langley *md5 = NULL; 51433ad798cSAdam Langley #endif 51533ad798cSAdam Langley 51633ad798cSAdam Langley /* we can't fit any SACK blocks in a packet with MD5 + TS 51733ad798cSAdam Langley options. There was discussion about disabling SACK rather than TS in 51833ad798cSAdam Langley order to fit in better with old, buggy kernels, but that was deemed 51933ad798cSAdam Langley to be unnecessary. */ 52033ad798cSAdam Langley doing_ts = ireq->tstamp_ok && !(*md5 && ireq->sack_ok); 52133ad798cSAdam Langley 52233ad798cSAdam Langley opts->mss = mss; 52333ad798cSAdam Langley size += TCPOLEN_MSS_ALIGNED; 52433ad798cSAdam Langley 52533ad798cSAdam Langley if (likely(ireq->wscale_ok)) { 52633ad798cSAdam Langley opts->ws = ireq->rcv_wscale; 5277982d5e1SPhilip Love if (likely(opts->ws)) 52833ad798cSAdam Langley size += TCPOLEN_WSCALE_ALIGNED; 52933ad798cSAdam Langley } 53033ad798cSAdam Langley if (likely(doing_ts)) { 53133ad798cSAdam Langley opts->options |= OPTION_TS; 53233ad798cSAdam Langley opts->tsval = TCP_SKB_CB(skb)->when; 53333ad798cSAdam Langley opts->tsecr = req->ts_recent; 53433ad798cSAdam Langley size += TCPOLEN_TSTAMP_ALIGNED; 53533ad798cSAdam Langley } 53633ad798cSAdam Langley if (likely(ireq->sack_ok)) { 53733ad798cSAdam Langley opts->options |= OPTION_SACK_ADVERTISE; 53833ad798cSAdam Langley if (unlikely(!doing_ts)) 53933ad798cSAdam Langley size += TCPOLEN_SACKPERM_ALIGNED; 54033ad798cSAdam Langley } 54133ad798cSAdam Langley 54233ad798cSAdam Langley return size; 54333ad798cSAdam Langley } 54433ad798cSAdam Langley 54533ad798cSAdam Langley static unsigned tcp_established_options(struct sock *sk, struct sk_buff *skb, 54633ad798cSAdam Langley struct tcp_out_options *opts, 54733ad798cSAdam Langley struct tcp_md5sig_key **md5) { 54833ad798cSAdam Langley struct tcp_skb_cb *tcb = skb ? TCP_SKB_CB(skb) : NULL; 54933ad798cSAdam Langley struct tcp_sock *tp = tcp_sk(sk); 55033ad798cSAdam Langley unsigned size = 0; 551cabeccbdSIlpo Järvinen unsigned int eff_sacks; 55233ad798cSAdam Langley 55333ad798cSAdam Langley #ifdef CONFIG_TCP_MD5SIG 55433ad798cSAdam Langley *md5 = tp->af_specific->md5_lookup(sk, sk); 55533ad798cSAdam Langley if (unlikely(*md5)) { 55633ad798cSAdam Langley opts->options |= OPTION_MD5; 55733ad798cSAdam Langley size += TCPOLEN_MD5SIG_ALIGNED; 55833ad798cSAdam Langley } 55933ad798cSAdam Langley #else 56033ad798cSAdam Langley *md5 = NULL; 56133ad798cSAdam Langley #endif 56233ad798cSAdam Langley 56333ad798cSAdam Langley if (likely(tp->rx_opt.tstamp_ok)) { 56433ad798cSAdam Langley opts->options |= OPTION_TS; 56533ad798cSAdam Langley opts->tsval = tcb ? tcb->when : 0; 56633ad798cSAdam Langley opts->tsecr = tp->rx_opt.ts_recent; 56733ad798cSAdam Langley size += TCPOLEN_TSTAMP_ALIGNED; 56833ad798cSAdam Langley } 56933ad798cSAdam Langley 570cabeccbdSIlpo Järvinen eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack; 571cabeccbdSIlpo Järvinen if (unlikely(eff_sacks)) { 57233ad798cSAdam Langley const unsigned remaining = MAX_TCP_OPTION_SPACE - size; 57333ad798cSAdam Langley opts->num_sack_blocks = 574cabeccbdSIlpo Järvinen min_t(unsigned, eff_sacks, 57533ad798cSAdam Langley (remaining - TCPOLEN_SACK_BASE_ALIGNED) / 57633ad798cSAdam Langley TCPOLEN_SACK_PERBLOCK); 57733ad798cSAdam Langley size += TCPOLEN_SACK_BASE_ALIGNED + 57833ad798cSAdam Langley opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK; 57933ad798cSAdam Langley } 58033ad798cSAdam Langley 58133ad798cSAdam Langley return size; 58240efc6faSStephen Hemminger } 5831da177e4SLinus Torvalds 5841da177e4SLinus Torvalds /* This routine actually transmits TCP packets queued in by 5851da177e4SLinus Torvalds * tcp_do_sendmsg(). This is used by both the initial 5861da177e4SLinus Torvalds * transmission and possible later retransmissions. 5871da177e4SLinus Torvalds * All SKB's seen here are completely headerless. It is our 5881da177e4SLinus Torvalds * job to build the TCP header, and pass the packet down to 5891da177e4SLinus Torvalds * IP so it can do the same plus pass the packet off to the 5901da177e4SLinus Torvalds * device. 5911da177e4SLinus Torvalds * 5921da177e4SLinus Torvalds * We are working here with either a clone of the original 5931da177e4SLinus Torvalds * SKB, or a fresh unique copy made by the retransmit engine. 5941da177e4SLinus Torvalds */ 595056834d9SIlpo Järvinen static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it, 596056834d9SIlpo Järvinen gfp_t gfp_mask) 5971da177e4SLinus Torvalds { 5986687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 599dfb4b9dcSDavid S. Miller struct inet_sock *inet; 600dfb4b9dcSDavid S. Miller struct tcp_sock *tp; 601dfb4b9dcSDavid S. Miller struct tcp_skb_cb *tcb; 60233ad798cSAdam Langley struct tcp_out_options opts; 60333ad798cSAdam Langley unsigned tcp_options_size, tcp_header_size; 604cfb6eeb4SYOSHIFUJI Hideaki struct tcp_md5sig_key *md5; 605cfb6eeb4SYOSHIFUJI Hideaki __u8 *md5_hash_location; 6061da177e4SLinus Torvalds struct tcphdr *th; 6071da177e4SLinus Torvalds int err; 6081da177e4SLinus Torvalds 609dfb4b9dcSDavid S. Miller BUG_ON(!skb || !tcp_skb_pcount(skb)); 610dfb4b9dcSDavid S. Miller 611dfb4b9dcSDavid S. Miller /* If congestion control is doing timestamping, we must 612dfb4b9dcSDavid S. Miller * take such a timestamp before we potentially clone/copy. 613dfb4b9dcSDavid S. Miller */ 614164891aaSStephen Hemminger if (icsk->icsk_ca_ops->flags & TCP_CONG_RTT_STAMP) 615dfb4b9dcSDavid S. Miller __net_timestamp(skb); 616dfb4b9dcSDavid S. Miller 617dfb4b9dcSDavid S. Miller if (likely(clone_it)) { 618dfb4b9dcSDavid S. Miller if (unlikely(skb_cloned(skb))) 619dfb4b9dcSDavid S. Miller skb = pskb_copy(skb, gfp_mask); 620dfb4b9dcSDavid S. Miller else 621dfb4b9dcSDavid S. Miller skb = skb_clone(skb, gfp_mask); 622dfb4b9dcSDavid S. Miller if (unlikely(!skb)) 623dfb4b9dcSDavid S. Miller return -ENOBUFS; 624dfb4b9dcSDavid S. Miller } 625dfb4b9dcSDavid S. Miller 626dfb4b9dcSDavid S. Miller inet = inet_sk(sk); 627dfb4b9dcSDavid S. Miller tp = tcp_sk(sk); 628dfb4b9dcSDavid S. Miller tcb = TCP_SKB_CB(skb); 62933ad798cSAdam Langley memset(&opts, 0, sizeof(opts)); 6301da177e4SLinus Torvalds 63133ad798cSAdam Langley if (unlikely(tcb->flags & TCPCB_FLAG_SYN)) 63233ad798cSAdam Langley tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5); 63333ad798cSAdam Langley else 63433ad798cSAdam Langley tcp_options_size = tcp_established_options(sk, skb, &opts, 63533ad798cSAdam Langley &md5); 63633ad798cSAdam Langley tcp_header_size = tcp_options_size + sizeof(struct tcphdr); 6371da177e4SLinus Torvalds 638317a76f9SStephen Hemminger if (tcp_packets_in_flight(tp) == 0) 6396687e988SArnaldo Carvalho de Melo tcp_ca_event(sk, CA_EVENT_TX_START); 6401da177e4SLinus Torvalds 641aa8223c7SArnaldo Carvalho de Melo skb_push(skb, tcp_header_size); 642aa8223c7SArnaldo Carvalho de Melo skb_reset_transport_header(skb); 643e89862f4SDavid S. Miller skb_set_owner_w(skb, sk); 6441da177e4SLinus Torvalds 6451da177e4SLinus Torvalds /* Build TCP header and checksum it. */ 646aa8223c7SArnaldo Carvalho de Melo th = tcp_hdr(skb); 6471da177e4SLinus Torvalds th->source = inet->sport; 6481da177e4SLinus Torvalds th->dest = inet->dport; 6491da177e4SLinus Torvalds th->seq = htonl(tcb->seq); 6501da177e4SLinus Torvalds th->ack_seq = htonl(tp->rcv_nxt); 651df7a3b07SAl Viro *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) | 652dfb4b9dcSDavid S. Miller tcb->flags); 653dfb4b9dcSDavid S. Miller 654dfb4b9dcSDavid S. Miller if (unlikely(tcb->flags & TCPCB_FLAG_SYN)) { 6551da177e4SLinus Torvalds /* RFC1323: The window in SYN & SYN/ACK segments 6561da177e4SLinus Torvalds * is never scaled. 6571da177e4SLinus Torvalds */ 658600ff0c2SIlpo Järvinen th->window = htons(min(tp->rcv_wnd, 65535U)); 6591da177e4SLinus Torvalds } else { 6601da177e4SLinus Torvalds th->window = htons(tcp_select_window(sk)); 6611da177e4SLinus Torvalds } 6621da177e4SLinus Torvalds th->check = 0; 6631da177e4SLinus Torvalds th->urg_ptr = 0; 6641da177e4SLinus Torvalds 66533f5f57eSIlpo Järvinen /* The urg_mode check is necessary during a below snd_una win probe */ 6667691367dSHerbert Xu if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) { 6677691367dSHerbert Xu if (before(tp->snd_up, tcb->seq + 0x10000)) { 6681da177e4SLinus Torvalds th->urg_ptr = htons(tp->snd_up - tcb->seq); 6691da177e4SLinus Torvalds th->urg = 1; 6707691367dSHerbert Xu } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) { 6717691367dSHerbert Xu th->urg_ptr = 0xFFFF; 6727691367dSHerbert Xu th->urg = 1; 6737691367dSHerbert Xu } 6741da177e4SLinus Torvalds } 6751da177e4SLinus Torvalds 67633ad798cSAdam Langley tcp_options_write((__be32 *)(th + 1), tp, &opts, &md5_hash_location); 67733ad798cSAdam Langley if (likely((tcb->flags & TCPCB_FLAG_SYN) == 0)) 6789e412ba7SIlpo Järvinen TCP_ECN_send(sk, skb, tcp_header_size); 679dfb4b9dcSDavid S. Miller 680cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 681cfb6eeb4SYOSHIFUJI Hideaki /* Calculate the MD5 hash, as we have all we need now */ 682cfb6eeb4SYOSHIFUJI Hideaki if (md5) { 68333ad798cSAdam Langley sk->sk_route_caps &= ~NETIF_F_GSO_MASK; 684cfb6eeb4SYOSHIFUJI Hideaki tp->af_specific->calc_md5_hash(md5_hash_location, 68549a72dfbSAdam Langley md5, sk, NULL, skb); 686cfb6eeb4SYOSHIFUJI Hideaki } 687cfb6eeb4SYOSHIFUJI Hideaki #endif 688cfb6eeb4SYOSHIFUJI Hideaki 6898292a17aSArnaldo Carvalho de Melo icsk->icsk_af_ops->send_check(sk, skb->len, skb); 6901da177e4SLinus Torvalds 691dfb4b9dcSDavid S. Miller if (likely(tcb->flags & TCPCB_FLAG_ACK)) 692fc6415bcSDavid S. Miller tcp_event_ack_sent(sk, tcp_skb_pcount(skb)); 6931da177e4SLinus Torvalds 6941da177e4SLinus Torvalds if (skb->len != tcp_header_size) 6951da177e4SLinus Torvalds tcp_event_data_sent(tp, skb, sk); 6961da177e4SLinus Torvalds 697bd37a088SWei Yongjun if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq) 69881cc8a75SPavel Emelyanov TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTSEGS); 6991da177e4SLinus Torvalds 700e89862f4SDavid S. Miller err = icsk->icsk_af_ops->queue_xmit(skb, 0); 70183de47cdSHua Zhong if (likely(err <= 0)) 7021da177e4SLinus Torvalds return err; 7031da177e4SLinus Torvalds 7043cfe3baaSIlpo Järvinen tcp_enter_cwr(sk, 1); 7051da177e4SLinus Torvalds 706b9df3cb8SGerrit Renker return net_xmit_eval(err); 7071da177e4SLinus Torvalds } 7081da177e4SLinus Torvalds 7091da177e4SLinus Torvalds /* This routine just queue's the buffer 7101da177e4SLinus Torvalds * 7111da177e4SLinus Torvalds * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames, 7121da177e4SLinus Torvalds * otherwise socket can stall. 7131da177e4SLinus Torvalds */ 7141da177e4SLinus Torvalds static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb) 7151da177e4SLinus Torvalds { 7161da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 7171da177e4SLinus Torvalds 7181da177e4SLinus Torvalds /* Advance write_seq and place onto the write_queue. */ 7191da177e4SLinus Torvalds tp->write_seq = TCP_SKB_CB(skb)->end_seq; 7201da177e4SLinus Torvalds skb_header_release(skb); 721fe067e8aSDavid S. Miller tcp_add_write_queue_tail(sk, skb); 7223ab224beSHideo Aoki sk->sk_wmem_queued += skb->truesize; 7233ab224beSHideo Aoki sk_mem_charge(sk, skb->truesize); 7241da177e4SLinus Torvalds } 7251da177e4SLinus Torvalds 726056834d9SIlpo Järvinen static void tcp_set_skb_tso_segs(struct sock *sk, struct sk_buff *skb, 727056834d9SIlpo Järvinen unsigned int mss_now) 728f6302d1dSDavid S. Miller { 729f8269a49SIlpo Järvinen if (skb->len <= mss_now || !sk_can_gso(sk)) { 730f6302d1dSDavid S. Miller /* Avoid the costly divide in the normal 731f6302d1dSDavid S. Miller * non-TSO case. 732f6302d1dSDavid S. Miller */ 7337967168cSHerbert Xu skb_shinfo(skb)->gso_segs = 1; 7347967168cSHerbert Xu skb_shinfo(skb)->gso_size = 0; 7357967168cSHerbert Xu skb_shinfo(skb)->gso_type = 0; 736f6302d1dSDavid S. Miller } else { 737356f89e1SIlpo Järvinen skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss_now); 7387967168cSHerbert Xu skb_shinfo(skb)->gso_size = mss_now; 739bcd76111SHerbert Xu skb_shinfo(skb)->gso_type = sk->sk_gso_type; 7401da177e4SLinus Torvalds } 7411da177e4SLinus Torvalds } 7421da177e4SLinus Torvalds 74391fed7a1SIlpo Järvinen /* When a modification to fackets out becomes necessary, we need to check 74468f8353bSIlpo Järvinen * skb is counted to fackets_out or not. 74591fed7a1SIlpo Järvinen */ 746a47e5a98SIlpo Järvinen static void tcp_adjust_fackets_out(struct sock *sk, struct sk_buff *skb, 74791fed7a1SIlpo Järvinen int decr) 74891fed7a1SIlpo Järvinen { 749a47e5a98SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 750a47e5a98SIlpo Järvinen 751dc86967bSIlpo Järvinen if (!tp->sacked_out || tcp_is_reno(tp)) 75291fed7a1SIlpo Järvinen return; 75391fed7a1SIlpo Järvinen 7546859d494SIlpo Järvinen if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq)) 75591fed7a1SIlpo Järvinen tp->fackets_out -= decr; 75691fed7a1SIlpo Järvinen } 75791fed7a1SIlpo Järvinen 7581da177e4SLinus Torvalds /* Function to create two new TCP segments. Shrinks the given segment 7591da177e4SLinus Torvalds * to the specified size and appends a new segment with the rest of the 7601da177e4SLinus Torvalds * packet to the list. This won't be called frequently, I hope. 7611da177e4SLinus Torvalds * Remember, these are still headerless SKBs at this point. 7621da177e4SLinus Torvalds */ 763056834d9SIlpo Järvinen int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len, 764056834d9SIlpo Järvinen unsigned int mss_now) 7651da177e4SLinus Torvalds { 7661da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 7671da177e4SLinus Torvalds struct sk_buff *buff; 7686475be16SDavid S. Miller int nsize, old_factor; 769b60b49eaSHerbert Xu int nlen; 770*9ce01461SIlpo Järvinen u8 flags; 7711da177e4SLinus Torvalds 772b2cc99f0SHerbert Xu BUG_ON(len > skb->len); 7736a438bbeSStephen Hemminger 7741da177e4SLinus Torvalds nsize = skb_headlen(skb) - len; 7751da177e4SLinus Torvalds if (nsize < 0) 7761da177e4SLinus Torvalds nsize = 0; 7771da177e4SLinus Torvalds 7781da177e4SLinus Torvalds if (skb_cloned(skb) && 7791da177e4SLinus Torvalds skb_is_nonlinear(skb) && 7801da177e4SLinus Torvalds pskb_expand_head(skb, 0, 0, GFP_ATOMIC)) 7811da177e4SLinus Torvalds return -ENOMEM; 7821da177e4SLinus Torvalds 7831da177e4SLinus Torvalds /* Get a new skb... force flag on. */ 7841da177e4SLinus Torvalds buff = sk_stream_alloc_skb(sk, nsize, GFP_ATOMIC); 7851da177e4SLinus Torvalds if (buff == NULL) 7861da177e4SLinus Torvalds return -ENOMEM; /* We'll just try again later. */ 787ef5cb973SHerbert Xu 7883ab224beSHideo Aoki sk->sk_wmem_queued += buff->truesize; 7893ab224beSHideo Aoki sk_mem_charge(sk, buff->truesize); 790b60b49eaSHerbert Xu nlen = skb->len - len - nsize; 791b60b49eaSHerbert Xu buff->truesize += nlen; 792b60b49eaSHerbert Xu skb->truesize -= nlen; 7931da177e4SLinus Torvalds 7941da177e4SLinus Torvalds /* Correct the sequence numbers. */ 7951da177e4SLinus Torvalds TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; 7961da177e4SLinus Torvalds TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; 7971da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; 7981da177e4SLinus Torvalds 7991da177e4SLinus Torvalds /* PSH and FIN should only be set in the second packet. */ 8001da177e4SLinus Torvalds flags = TCP_SKB_CB(skb)->flags; 8011da177e4SLinus Torvalds TCP_SKB_CB(skb)->flags = flags & ~(TCPCB_FLAG_FIN | TCPCB_FLAG_PSH); 8021da177e4SLinus Torvalds TCP_SKB_CB(buff)->flags = flags; 803e14c3cafSHerbert Xu TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked; 8041da177e4SLinus Torvalds 80584fa7933SPatrick McHardy if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) { 8061da177e4SLinus Torvalds /* Copy and checksum data tail into the new buffer. */ 807056834d9SIlpo Järvinen buff->csum = csum_partial_copy_nocheck(skb->data + len, 808056834d9SIlpo Järvinen skb_put(buff, nsize), 8091da177e4SLinus Torvalds nsize, 0); 8101da177e4SLinus Torvalds 8111da177e4SLinus Torvalds skb_trim(skb, len); 8121da177e4SLinus Torvalds 8131da177e4SLinus Torvalds skb->csum = csum_block_sub(skb->csum, buff->csum, len); 8141da177e4SLinus Torvalds } else { 81584fa7933SPatrick McHardy skb->ip_summed = CHECKSUM_PARTIAL; 8161da177e4SLinus Torvalds skb_split(skb, buff, len); 8171da177e4SLinus Torvalds } 8181da177e4SLinus Torvalds 8191da177e4SLinus Torvalds buff->ip_summed = skb->ip_summed; 8201da177e4SLinus Torvalds 8211da177e4SLinus Torvalds /* Looks stupid, but our code really uses when of 8221da177e4SLinus Torvalds * skbs, which it never sent before. --ANK 8231da177e4SLinus Torvalds */ 8241da177e4SLinus Torvalds TCP_SKB_CB(buff)->when = TCP_SKB_CB(skb)->when; 825a61bbcf2SPatrick McHardy buff->tstamp = skb->tstamp; 8261da177e4SLinus Torvalds 8276475be16SDavid S. Miller old_factor = tcp_skb_pcount(skb); 8286475be16SDavid S. Miller 8291da177e4SLinus Torvalds /* Fix up tso_factor for both original and new SKB. */ 830846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, skb, mss_now); 831846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, buff, mss_now); 8321da177e4SLinus Torvalds 8336475be16SDavid S. Miller /* If this packet has been sent out already, we must 8346475be16SDavid S. Miller * adjust the various packet counters. 8356475be16SDavid S. Miller */ 836cf0b450cSHerbert Xu if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) { 8376475be16SDavid S. Miller int diff = old_factor - tcp_skb_pcount(skb) - 8386475be16SDavid S. Miller tcp_skb_pcount(buff); 8391da177e4SLinus Torvalds 8406475be16SDavid S. Miller tp->packets_out -= diff; 841e14c3cafSHerbert Xu 842e14c3cafSHerbert Xu if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 843e14c3cafSHerbert Xu tp->sacked_out -= diff; 844e14c3cafSHerbert Xu if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) 845e14c3cafSHerbert Xu tp->retrans_out -= diff; 846e14c3cafSHerbert Xu 847b5860bbaSIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST) 8486475be16SDavid S. Miller tp->lost_out -= diff; 84983ca28beSHerbert Xu 85083ca28beSHerbert Xu /* Adjust Reno SACK estimate. */ 85191fed7a1SIlpo Järvinen if (tcp_is_reno(tp) && diff > 0) { 852af610b4cSIlpo Järvinen tcp_dec_pcount_approx_int(&tp->sacked_out, diff); 853005903bcSIlpo Järvinen tcp_verify_left_out(tp); 85483ca28beSHerbert Xu } 855a47e5a98SIlpo Järvinen tcp_adjust_fackets_out(sk, skb, diff); 856d3d2ae45SIlpo Järvinen 857d3d2ae45SIlpo Järvinen if (tp->lost_skb_hint && 858d3d2ae45SIlpo Järvinen before(TCP_SKB_CB(skb)->seq, 859d3d2ae45SIlpo Järvinen TCP_SKB_CB(tp->lost_skb_hint)->seq) && 860d3d2ae45SIlpo Järvinen (tcp_is_fack(tp) || TCP_SKB_CB(skb)->sacked)) 861d3d2ae45SIlpo Järvinen tp->lost_cnt_hint -= diff; 8621da177e4SLinus Torvalds } 8631da177e4SLinus Torvalds 8641da177e4SLinus Torvalds /* Link BUFF into the send queue. */ 865f44b5271SDavid S. Miller skb_header_release(buff); 866fe067e8aSDavid S. Miller tcp_insert_write_queue_after(skb, buff, sk); 8671da177e4SLinus Torvalds 8681da177e4SLinus Torvalds return 0; 8691da177e4SLinus Torvalds } 8701da177e4SLinus Torvalds 8711da177e4SLinus Torvalds /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c 8721da177e4SLinus Torvalds * eventually). The difference is that pulled data not copied, but 8731da177e4SLinus Torvalds * immediately discarded. 8741da177e4SLinus Torvalds */ 875f2911969SHerbert Xu ~{PmVHI~} static void __pskb_trim_head(struct sk_buff *skb, int len) 8761da177e4SLinus Torvalds { 8771da177e4SLinus Torvalds int i, k, eat; 8781da177e4SLinus Torvalds 8791da177e4SLinus Torvalds eat = len; 8801da177e4SLinus Torvalds k = 0; 8811da177e4SLinus Torvalds for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) { 8821da177e4SLinus Torvalds if (skb_shinfo(skb)->frags[i].size <= eat) { 8831da177e4SLinus Torvalds put_page(skb_shinfo(skb)->frags[i].page); 8841da177e4SLinus Torvalds eat -= skb_shinfo(skb)->frags[i].size; 8851da177e4SLinus Torvalds } else { 8861da177e4SLinus Torvalds skb_shinfo(skb)->frags[k] = skb_shinfo(skb)->frags[i]; 8871da177e4SLinus Torvalds if (eat) { 8881da177e4SLinus Torvalds skb_shinfo(skb)->frags[k].page_offset += eat; 8891da177e4SLinus Torvalds skb_shinfo(skb)->frags[k].size -= eat; 8901da177e4SLinus Torvalds eat = 0; 8911da177e4SLinus Torvalds } 8921da177e4SLinus Torvalds k++; 8931da177e4SLinus Torvalds } 8941da177e4SLinus Torvalds } 8951da177e4SLinus Torvalds skb_shinfo(skb)->nr_frags = k; 8961da177e4SLinus Torvalds 89727a884dcSArnaldo Carvalho de Melo skb_reset_tail_pointer(skb); 8981da177e4SLinus Torvalds skb->data_len -= len; 8991da177e4SLinus Torvalds skb->len = skb->data_len; 9001da177e4SLinus Torvalds } 9011da177e4SLinus Torvalds 9021da177e4SLinus Torvalds int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len) 9031da177e4SLinus Torvalds { 904056834d9SIlpo Järvinen if (skb_cloned(skb) && pskb_expand_head(skb, 0, 0, GFP_ATOMIC)) 9051da177e4SLinus Torvalds return -ENOMEM; 9061da177e4SLinus Torvalds 907f2911969SHerbert Xu ~{PmVHI~} /* If len == headlen, we avoid __skb_pull to preserve alignment. */ 908f2911969SHerbert Xu ~{PmVHI~} if (unlikely(len < skb_headlen(skb))) 9091da177e4SLinus Torvalds __skb_pull(skb, len); 910f2911969SHerbert Xu ~{PmVHI~} else 911f2911969SHerbert Xu ~{PmVHI~} __pskb_trim_head(skb, len - skb_headlen(skb)); 9121da177e4SLinus Torvalds 9131da177e4SLinus Torvalds TCP_SKB_CB(skb)->seq += len; 91484fa7933SPatrick McHardy skb->ip_summed = CHECKSUM_PARTIAL; 9151da177e4SLinus Torvalds 9161da177e4SLinus Torvalds skb->truesize -= len; 9171da177e4SLinus Torvalds sk->sk_wmem_queued -= len; 9183ab224beSHideo Aoki sk_mem_uncharge(sk, len); 9191da177e4SLinus Torvalds sock_set_flag(sk, SOCK_QUEUE_SHRUNK); 9201da177e4SLinus Torvalds 9211da177e4SLinus Torvalds /* Any change of skb->len requires recalculation of tso 9221da177e4SLinus Torvalds * factor and mss. 9231da177e4SLinus Torvalds */ 9241da177e4SLinus Torvalds if (tcp_skb_pcount(skb) > 1) 925846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, skb, tcp_current_mss(sk, 1)); 9261da177e4SLinus Torvalds 9271da177e4SLinus Torvalds return 0; 9281da177e4SLinus Torvalds } 9291da177e4SLinus Torvalds 9305d424d5aSJohn Heffner /* Not accounting for SACKs here. */ 9315d424d5aSJohn Heffner int tcp_mtu_to_mss(struct sock *sk, int pmtu) 9325d424d5aSJohn Heffner { 9335d424d5aSJohn Heffner struct tcp_sock *tp = tcp_sk(sk); 9345d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 9355d424d5aSJohn Heffner int mss_now; 9365d424d5aSJohn Heffner 9375d424d5aSJohn Heffner /* Calculate base mss without TCP options: 9385d424d5aSJohn Heffner It is MMS_S - sizeof(tcphdr) of rfc1122 9395d424d5aSJohn Heffner */ 9405d424d5aSJohn Heffner mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr); 9415d424d5aSJohn Heffner 9425d424d5aSJohn Heffner /* Clamp it (mss_clamp does not include tcp options) */ 9435d424d5aSJohn Heffner if (mss_now > tp->rx_opt.mss_clamp) 9445d424d5aSJohn Heffner mss_now = tp->rx_opt.mss_clamp; 9455d424d5aSJohn Heffner 9465d424d5aSJohn Heffner /* Now subtract optional transport overhead */ 9475d424d5aSJohn Heffner mss_now -= icsk->icsk_ext_hdr_len; 9485d424d5aSJohn Heffner 9495d424d5aSJohn Heffner /* Then reserve room for full set of TCP options and 8 bytes of data */ 9505d424d5aSJohn Heffner if (mss_now < 48) 9515d424d5aSJohn Heffner mss_now = 48; 9525d424d5aSJohn Heffner 9535d424d5aSJohn Heffner /* Now subtract TCP options size, not including SACKs */ 9545d424d5aSJohn Heffner mss_now -= tp->tcp_header_len - sizeof(struct tcphdr); 9555d424d5aSJohn Heffner 9565d424d5aSJohn Heffner return mss_now; 9575d424d5aSJohn Heffner } 9585d424d5aSJohn Heffner 9595d424d5aSJohn Heffner /* Inverse of above */ 9605d424d5aSJohn Heffner int tcp_mss_to_mtu(struct sock *sk, int mss) 9615d424d5aSJohn Heffner { 9625d424d5aSJohn Heffner struct tcp_sock *tp = tcp_sk(sk); 9635d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 9645d424d5aSJohn Heffner int mtu; 9655d424d5aSJohn Heffner 9665d424d5aSJohn Heffner mtu = mss + 9675d424d5aSJohn Heffner tp->tcp_header_len + 9685d424d5aSJohn Heffner icsk->icsk_ext_hdr_len + 9695d424d5aSJohn Heffner icsk->icsk_af_ops->net_header_len; 9705d424d5aSJohn Heffner 9715d424d5aSJohn Heffner return mtu; 9725d424d5aSJohn Heffner } 9735d424d5aSJohn Heffner 9745d424d5aSJohn Heffner void tcp_mtup_init(struct sock *sk) 9755d424d5aSJohn Heffner { 9765d424d5aSJohn Heffner struct tcp_sock *tp = tcp_sk(sk); 9775d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 9785d424d5aSJohn Heffner 9795d424d5aSJohn Heffner icsk->icsk_mtup.enabled = sysctl_tcp_mtu_probing > 1; 9805d424d5aSJohn Heffner icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) + 9815d424d5aSJohn Heffner icsk->icsk_af_ops->net_header_len; 9825d424d5aSJohn Heffner icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, sysctl_tcp_base_mss); 9835d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = 0; 9845d424d5aSJohn Heffner } 9855d424d5aSJohn Heffner 986409d22b4SIlpo Järvinen /* Bound MSS / TSO packet size with the half of the window */ 987409d22b4SIlpo Järvinen static int tcp_bound_to_half_wnd(struct tcp_sock *tp, int pktsize) 988409d22b4SIlpo Järvinen { 989409d22b4SIlpo Järvinen if (tp->max_window && pktsize > (tp->max_window >> 1)) 990409d22b4SIlpo Järvinen return max(tp->max_window >> 1, 68U - tp->tcp_header_len); 991409d22b4SIlpo Järvinen else 992409d22b4SIlpo Järvinen return pktsize; 993409d22b4SIlpo Järvinen } 994409d22b4SIlpo Järvinen 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 */ 1041c1b4a7e6SDavid S. Miller unsigned int tcp_current_mss(struct sock *sk, int large_allowed) 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; 1046c1b4a7e6SDavid S. Miller u16 xmit_size_goal; 1047c1b4a7e6SDavid S. Miller int doing_tso = 0; 104833ad798cSAdam Langley unsigned header_len; 104933ad798cSAdam Langley struct tcp_out_options opts; 105033ad798cSAdam Langley struct tcp_md5sig_key *md5; 10511da177e4SLinus Torvalds 1052c1b4a7e6SDavid S. Miller mss_now = tp->mss_cache; 1053c1b4a7e6SDavid S. Miller 1054f8269a49SIlpo Järvinen if (large_allowed && sk_can_gso(sk)) 1055c1b4a7e6SDavid S. Miller doing_tso = 1; 1056c1b4a7e6SDavid S. Miller 10571da177e4SLinus Torvalds if (dst) { 10581da177e4SLinus Torvalds u32 mtu = dst_mtu(dst); 1059d83d8461SArnaldo Carvalho de Melo if (mtu != inet_csk(sk)->icsk_pmtu_cookie) 10601da177e4SLinus Torvalds mss_now = tcp_sync_mss(sk, mtu); 10611da177e4SLinus Torvalds } 10621da177e4SLinus Torvalds 106333ad798cSAdam Langley header_len = tcp_established_options(sk, NULL, &opts, &md5) + 106433ad798cSAdam Langley sizeof(struct tcphdr); 106533ad798cSAdam Langley /* The mss_cache is sized based on tp->tcp_header_len, which assumes 106633ad798cSAdam Langley * some common options. If this is an odd packet (because we have SACK 106733ad798cSAdam Langley * blocks etc) then our calculated header_len will be different, and 106833ad798cSAdam Langley * we have to adjust mss_now correspondingly */ 106933ad798cSAdam Langley if (header_len != tp->tcp_header_len) { 107033ad798cSAdam Langley int delta = (int) header_len - tp->tcp_header_len; 107133ad798cSAdam Langley mss_now -= delta; 107233ad798cSAdam Langley } 1073cfb6eeb4SYOSHIFUJI Hideaki 1074c1b4a7e6SDavid S. Miller xmit_size_goal = mss_now; 1075c1b4a7e6SDavid S. Miller 1076c1b4a7e6SDavid S. Miller if (doing_tso) { 107782cc1a7aSPeter P Waskiewicz Jr xmit_size_goal = ((sk->sk_gso_max_size - 1) - 10788292a17aSArnaldo Carvalho de Melo inet_csk(sk)->icsk_af_ops->net_header_len - 1079d83d8461SArnaldo Carvalho de Melo inet_csk(sk)->icsk_ext_hdr_len - 1080d83d8461SArnaldo Carvalho de Melo tp->tcp_header_len); 1081c1b4a7e6SDavid S. Miller 1082409d22b4SIlpo Järvinen xmit_size_goal = tcp_bound_to_half_wnd(tp, xmit_size_goal); 1083c1b4a7e6SDavid S. Miller xmit_size_goal -= (xmit_size_goal % mss_now); 1084c1b4a7e6SDavid S. Miller } 1085c1b4a7e6SDavid S. Miller tp->xmit_size_goal = xmit_size_goal; 1086c1b4a7e6SDavid S. Miller 10871da177e4SLinus Torvalds return mss_now; 10881da177e4SLinus Torvalds } 10891da177e4SLinus Torvalds 1090a762a980SDavid S. Miller /* Congestion window validation. (RFC2861) */ 10919e412ba7SIlpo Järvinen static void tcp_cwnd_validate(struct sock *sk) 1092a762a980SDavid S. Miller { 10939e412ba7SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 1094a762a980SDavid S. Miller 1095d436d686SIlpo Järvinen if (tp->packets_out >= tp->snd_cwnd) { 1096a762a980SDavid S. Miller /* Network is feed fully. */ 1097a762a980SDavid S. Miller tp->snd_cwnd_used = 0; 1098a762a980SDavid S. Miller tp->snd_cwnd_stamp = tcp_time_stamp; 1099a762a980SDavid S. Miller } else { 1100a762a980SDavid S. Miller /* Network starves. */ 1101a762a980SDavid S. Miller if (tp->packets_out > tp->snd_cwnd_used) 1102a762a980SDavid S. Miller tp->snd_cwnd_used = tp->packets_out; 1103a762a980SDavid S. Miller 110415d33c07SDavid S. Miller if (sysctl_tcp_slow_start_after_idle && 110515d33c07SDavid S. Miller (s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto) 1106a762a980SDavid S. Miller tcp_cwnd_application_limited(sk); 1107a762a980SDavid S. Miller } 1108a762a980SDavid S. Miller } 1109a762a980SDavid S. Miller 11100e3a4803SIlpo Järvinen /* Returns the portion of skb which can be sent right away without 11110e3a4803SIlpo Järvinen * introducing MSS oddities to segment boundaries. In rare cases where 11120e3a4803SIlpo Järvinen * mss_now != mss_cache, we will request caller to create a small skb 11130e3a4803SIlpo Järvinen * per input skb which could be mostly avoided here (if desired). 11145ea3a748SIlpo Järvinen * 11155ea3a748SIlpo Järvinen * We explicitly want to create a request for splitting write queue tail 11165ea3a748SIlpo Järvinen * to a small skb for Nagle purposes while avoiding unnecessary modulos, 11175ea3a748SIlpo Järvinen * thus all the complexity (cwnd_len is always MSS multiple which we 11185ea3a748SIlpo Järvinen * return whenever allowed by the other factors). Basically we need the 11195ea3a748SIlpo Järvinen * modulo only when the receiver window alone is the limiting factor or 11205ea3a748SIlpo Järvinen * when we would be allowed to send the split-due-to-Nagle skb fully. 11210e3a4803SIlpo Järvinen */ 11220e3a4803SIlpo Järvinen static unsigned int tcp_mss_split_point(struct sock *sk, struct sk_buff *skb, 1123056834d9SIlpo Järvinen unsigned int mss_now, unsigned int cwnd) 1124c1b4a7e6SDavid S. Miller { 11250e3a4803SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 11260e3a4803SIlpo Järvinen u32 needed, window, cwnd_len; 1127c1b4a7e6SDavid S. Miller 112890840defSIlpo Järvinen window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 1129c1b4a7e6SDavid S. Miller cwnd_len = mss_now * cwnd; 11300e3a4803SIlpo Järvinen 11310e3a4803SIlpo Järvinen if (likely(cwnd_len <= window && skb != tcp_write_queue_tail(sk))) 11320e3a4803SIlpo Järvinen return cwnd_len; 11330e3a4803SIlpo Järvinen 11345ea3a748SIlpo Järvinen needed = min(skb->len, window); 11355ea3a748SIlpo Järvinen 113617515408SIlpo Järvinen if (cwnd_len <= needed) 11370e3a4803SIlpo Järvinen return cwnd_len; 11380e3a4803SIlpo Järvinen 11390e3a4803SIlpo Järvinen return needed - needed % mss_now; 1140c1b4a7e6SDavid S. Miller } 1141c1b4a7e6SDavid S. Miller 1142c1b4a7e6SDavid S. Miller /* Can at least one segment of SKB be sent right now, according to the 1143c1b4a7e6SDavid S. Miller * congestion window rules? If so, return how many segments are allowed. 1144c1b4a7e6SDavid S. Miller */ 1145056834d9SIlpo Järvinen static inline unsigned int tcp_cwnd_test(struct tcp_sock *tp, 1146056834d9SIlpo Järvinen struct sk_buff *skb) 1147c1b4a7e6SDavid S. Miller { 1148c1b4a7e6SDavid S. Miller u32 in_flight, cwnd; 1149c1b4a7e6SDavid S. Miller 1150c1b4a7e6SDavid S. Miller /* Don't be strict about the congestion window for the final FIN. */ 1151104439a8SJohn Heffner if ((TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN) && 1152104439a8SJohn Heffner tcp_skb_pcount(skb) == 1) 1153c1b4a7e6SDavid S. Miller return 1; 1154c1b4a7e6SDavid S. Miller 1155c1b4a7e6SDavid S. Miller in_flight = tcp_packets_in_flight(tp); 1156c1b4a7e6SDavid S. Miller cwnd = tp->snd_cwnd; 1157c1b4a7e6SDavid S. Miller if (in_flight < cwnd) 1158c1b4a7e6SDavid S. Miller return (cwnd - in_flight); 1159c1b4a7e6SDavid S. Miller 1160c1b4a7e6SDavid S. Miller return 0; 1161c1b4a7e6SDavid S. Miller } 1162c1b4a7e6SDavid S. Miller 1163c1b4a7e6SDavid S. Miller /* This must be invoked the first time we consider transmitting 1164c1b4a7e6SDavid S. Miller * SKB onto the wire. 1165c1b4a7e6SDavid S. Miller */ 1166056834d9SIlpo Järvinen static int tcp_init_tso_segs(struct sock *sk, struct sk_buff *skb, 1167056834d9SIlpo Järvinen unsigned int mss_now) 1168c1b4a7e6SDavid S. Miller { 1169c1b4a7e6SDavid S. Miller int tso_segs = tcp_skb_pcount(skb); 1170c1b4a7e6SDavid S. Miller 1171f8269a49SIlpo Järvinen if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) { 1172846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, skb, mss_now); 1173c1b4a7e6SDavid S. Miller tso_segs = tcp_skb_pcount(skb); 1174c1b4a7e6SDavid S. Miller } 1175c1b4a7e6SDavid S. Miller return tso_segs; 1176c1b4a7e6SDavid S. Miller } 1177c1b4a7e6SDavid S. Miller 1178c1b4a7e6SDavid S. Miller static inline int tcp_minshall_check(const struct tcp_sock *tp) 1179c1b4a7e6SDavid S. Miller { 1180c1b4a7e6SDavid S. Miller return after(tp->snd_sml, tp->snd_una) && 1181c1b4a7e6SDavid S. Miller !after(tp->snd_sml, tp->snd_nxt); 1182c1b4a7e6SDavid S. Miller } 1183c1b4a7e6SDavid S. Miller 1184c1b4a7e6SDavid S. Miller /* Return 0, if packet can be sent now without violation Nagle's rules: 1185c1b4a7e6SDavid S. Miller * 1. It is full sized. 1186c1b4a7e6SDavid S. Miller * 2. Or it contains FIN. (already checked by caller) 1187c1b4a7e6SDavid S. Miller * 3. Or TCP_NODELAY was set. 1188c1b4a7e6SDavid S. Miller * 4. Or TCP_CORK is not set, and all sent packets are ACKed. 1189c1b4a7e6SDavid S. Miller * With Minshall's modification: all sent small packets are ACKed. 1190c1b4a7e6SDavid S. Miller */ 1191c1b4a7e6SDavid S. Miller static inline int tcp_nagle_check(const struct tcp_sock *tp, 1192c1b4a7e6SDavid S. Miller const struct sk_buff *skb, 1193c1b4a7e6SDavid S. Miller unsigned mss_now, int nonagle) 1194c1b4a7e6SDavid S. Miller { 1195c1b4a7e6SDavid S. Miller return (skb->len < mss_now && 1196c1b4a7e6SDavid S. Miller ((nonagle & TCP_NAGLE_CORK) || 1197056834d9SIlpo Järvinen (!nonagle && tp->packets_out && tcp_minshall_check(tp)))); 1198c1b4a7e6SDavid S. Miller } 1199c1b4a7e6SDavid S. Miller 1200c1b4a7e6SDavid S. Miller /* Return non-zero if the Nagle test allows this packet to be 1201c1b4a7e6SDavid S. Miller * sent now. 1202c1b4a7e6SDavid S. Miller */ 1203c1b4a7e6SDavid S. Miller static inline int tcp_nagle_test(struct tcp_sock *tp, struct sk_buff *skb, 1204c1b4a7e6SDavid S. Miller unsigned int cur_mss, int nonagle) 1205c1b4a7e6SDavid S. Miller { 1206c1b4a7e6SDavid S. Miller /* Nagle rule does not apply to frames, which sit in the middle of the 1207c1b4a7e6SDavid S. Miller * write_queue (they have no chances to get new data). 1208c1b4a7e6SDavid S. Miller * 1209c1b4a7e6SDavid S. Miller * This is implemented in the callers, where they modify the 'nonagle' 1210c1b4a7e6SDavid S. Miller * argument based upon the location of SKB in the send queue. 1211c1b4a7e6SDavid S. Miller */ 1212c1b4a7e6SDavid S. Miller if (nonagle & TCP_NAGLE_PUSH) 1213c1b4a7e6SDavid S. Miller return 1; 1214c1b4a7e6SDavid S. Miller 1215d551e454SIlpo Järvinen /* Don't use the nagle rule for urgent data (or for the final FIN). 1216d551e454SIlpo Järvinen * Nagle can be ignored during F-RTO too (see RFC4138). 1217d551e454SIlpo Järvinen */ 121833f5f57eSIlpo Järvinen if (tcp_urg_mode(tp) || (tp->frto_counter == 2) || 1219c1b4a7e6SDavid S. Miller (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN)) 1220c1b4a7e6SDavid S. Miller return 1; 1221c1b4a7e6SDavid S. Miller 1222c1b4a7e6SDavid S. Miller if (!tcp_nagle_check(tp, skb, cur_mss, nonagle)) 1223c1b4a7e6SDavid S. Miller return 1; 1224c1b4a7e6SDavid S. Miller 1225c1b4a7e6SDavid S. Miller return 0; 1226c1b4a7e6SDavid S. Miller } 1227c1b4a7e6SDavid S. Miller 1228c1b4a7e6SDavid S. Miller /* Does at least the first segment of SKB fit into the send window? */ 1229056834d9SIlpo Järvinen static inline int tcp_snd_wnd_test(struct tcp_sock *tp, struct sk_buff *skb, 1230056834d9SIlpo Järvinen unsigned int cur_mss) 1231c1b4a7e6SDavid S. Miller { 1232c1b4a7e6SDavid S. Miller u32 end_seq = TCP_SKB_CB(skb)->end_seq; 1233c1b4a7e6SDavid S. Miller 1234c1b4a7e6SDavid S. Miller if (skb->len > cur_mss) 1235c1b4a7e6SDavid S. Miller end_seq = TCP_SKB_CB(skb)->seq + cur_mss; 1236c1b4a7e6SDavid S. Miller 123790840defSIlpo Järvinen return !after(end_seq, tcp_wnd_end(tp)); 1238c1b4a7e6SDavid S. Miller } 1239c1b4a7e6SDavid S. Miller 1240fe067e8aSDavid S. Miller /* This checks if the data bearing packet SKB (usually tcp_send_head(sk)) 1241c1b4a7e6SDavid S. Miller * should be put on the wire right now. If so, it returns the number of 1242c1b4a7e6SDavid S. Miller * packets allowed by the congestion window. 1243c1b4a7e6SDavid S. Miller */ 1244c1b4a7e6SDavid S. Miller static unsigned int tcp_snd_test(struct sock *sk, struct sk_buff *skb, 1245c1b4a7e6SDavid S. Miller unsigned int cur_mss, int nonagle) 1246c1b4a7e6SDavid S. Miller { 1247c1b4a7e6SDavid S. Miller struct tcp_sock *tp = tcp_sk(sk); 1248c1b4a7e6SDavid S. Miller unsigned int cwnd_quota; 1249c1b4a7e6SDavid S. Miller 1250846998aeSDavid S. Miller tcp_init_tso_segs(sk, skb, cur_mss); 1251c1b4a7e6SDavid S. Miller 1252c1b4a7e6SDavid S. Miller if (!tcp_nagle_test(tp, skb, cur_mss, nonagle)) 1253c1b4a7e6SDavid S. Miller return 0; 1254c1b4a7e6SDavid S. Miller 1255c1b4a7e6SDavid S. Miller cwnd_quota = tcp_cwnd_test(tp, skb); 1256056834d9SIlpo Järvinen if (cwnd_quota && !tcp_snd_wnd_test(tp, skb, cur_mss)) 1257c1b4a7e6SDavid S. Miller cwnd_quota = 0; 1258c1b4a7e6SDavid S. Miller 1259c1b4a7e6SDavid S. Miller return cwnd_quota; 1260c1b4a7e6SDavid S. Miller } 1261c1b4a7e6SDavid S. Miller 12629e412ba7SIlpo Järvinen int tcp_may_send_now(struct sock *sk) 1263c1b4a7e6SDavid S. Miller { 12649e412ba7SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 1265fe067e8aSDavid S. Miller struct sk_buff *skb = tcp_send_head(sk); 1266c1b4a7e6SDavid S. Miller 1267c1b4a7e6SDavid S. Miller return (skb && 1268c1b4a7e6SDavid S. Miller tcp_snd_test(sk, skb, tcp_current_mss(sk, 1), 1269c1b4a7e6SDavid S. Miller (tcp_skb_is_last(sk, skb) ? 12704e67d876SIlpo Järvinen tp->nonagle : TCP_NAGLE_PUSH))); 1271c1b4a7e6SDavid S. Miller } 1272c1b4a7e6SDavid S. Miller 1273c1b4a7e6SDavid S. Miller /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet 1274c1b4a7e6SDavid S. Miller * which is put after SKB on the list. It is very much like 1275c1b4a7e6SDavid S. Miller * tcp_fragment() except that it may make several kinds of assumptions 1276c1b4a7e6SDavid S. Miller * in order to speed up the splitting operation. In particular, we 1277c1b4a7e6SDavid S. Miller * know that all the data is in scatter-gather pages, and that the 1278c1b4a7e6SDavid S. Miller * packet has never been sent out before (and thus is not cloned). 1279c1b4a7e6SDavid S. Miller */ 1280056834d9SIlpo Järvinen static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len, 1281056834d9SIlpo Järvinen unsigned int mss_now) 1282c1b4a7e6SDavid S. Miller { 1283c1b4a7e6SDavid S. Miller struct sk_buff *buff; 1284c1b4a7e6SDavid S. Miller int nlen = skb->len - len; 1285*9ce01461SIlpo Järvinen u8 flags; 1286c1b4a7e6SDavid S. Miller 1287c1b4a7e6SDavid S. Miller /* All of a TSO frame must be composed of paged data. */ 1288c8ac3774SHerbert Xu if (skb->len != skb->data_len) 1289c8ac3774SHerbert Xu return tcp_fragment(sk, skb, len, mss_now); 1290c1b4a7e6SDavid S. Miller 1291df97c708SPavel Emelyanov buff = sk_stream_alloc_skb(sk, 0, GFP_ATOMIC); 1292c1b4a7e6SDavid S. Miller if (unlikely(buff == NULL)) 1293c1b4a7e6SDavid S. Miller return -ENOMEM; 1294c1b4a7e6SDavid S. Miller 12953ab224beSHideo Aoki sk->sk_wmem_queued += buff->truesize; 12963ab224beSHideo Aoki sk_mem_charge(sk, buff->truesize); 1297b60b49eaSHerbert Xu buff->truesize += nlen; 1298c1b4a7e6SDavid S. Miller skb->truesize -= nlen; 1299c1b4a7e6SDavid S. Miller 1300c1b4a7e6SDavid S. Miller /* Correct the sequence numbers. */ 1301c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; 1302c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; 1303c1b4a7e6SDavid S. Miller TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; 1304c1b4a7e6SDavid S. Miller 1305c1b4a7e6SDavid S. Miller /* PSH and FIN should only be set in the second packet. */ 1306c1b4a7e6SDavid S. Miller flags = TCP_SKB_CB(skb)->flags; 1307c1b4a7e6SDavid S. Miller TCP_SKB_CB(skb)->flags = flags & ~(TCPCB_FLAG_FIN | TCPCB_FLAG_PSH); 1308c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->flags = flags; 1309c1b4a7e6SDavid S. Miller 1310c1b4a7e6SDavid S. Miller /* This packet was never sent out yet, so no SACK bits. */ 1311c1b4a7e6SDavid S. Miller TCP_SKB_CB(buff)->sacked = 0; 1312c1b4a7e6SDavid S. Miller 131384fa7933SPatrick McHardy buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL; 1314c1b4a7e6SDavid S. Miller skb_split(skb, buff, len); 1315c1b4a7e6SDavid S. Miller 1316c1b4a7e6SDavid S. Miller /* Fix up tso_factor for both original and new SKB. */ 1317846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, skb, mss_now); 1318846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, buff, mss_now); 1319c1b4a7e6SDavid S. Miller 1320c1b4a7e6SDavid S. Miller /* Link BUFF into the send queue. */ 1321c1b4a7e6SDavid S. Miller skb_header_release(buff); 1322fe067e8aSDavid S. Miller tcp_insert_write_queue_after(skb, buff, sk); 1323c1b4a7e6SDavid S. Miller 1324c1b4a7e6SDavid S. Miller return 0; 1325c1b4a7e6SDavid S. Miller } 1326c1b4a7e6SDavid S. Miller 1327c1b4a7e6SDavid S. Miller /* Try to defer sending, if possible, in order to minimize the amount 1328c1b4a7e6SDavid S. Miller * of TSO splitting we do. View it as a kind of TSO Nagle test. 1329c1b4a7e6SDavid S. Miller * 1330c1b4a7e6SDavid S. Miller * This algorithm is from John Heffner. 1331c1b4a7e6SDavid S. Miller */ 13329e412ba7SIlpo Järvinen static int tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb) 1333c1b4a7e6SDavid S. Miller { 13349e412ba7SIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 13356687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 1336c1b4a7e6SDavid S. Miller u32 send_win, cong_win, limit, in_flight; 1337c1b4a7e6SDavid S. Miller 1338c1b4a7e6SDavid S. Miller if (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN) 1339ae8064acSJohn Heffner goto send_now; 1340c1b4a7e6SDavid S. Miller 13416687e988SArnaldo Carvalho de Melo if (icsk->icsk_ca_state != TCP_CA_Open) 1342ae8064acSJohn Heffner goto send_now; 1343ae8064acSJohn Heffner 1344ae8064acSJohn Heffner /* Defer for less than two clock ticks. */ 1345bd515c3eSIlpo Järvinen if (tp->tso_deferred && 1346a2acde07SIlpo Järvinen (((u32)jiffies << 1) >> 1) - (tp->tso_deferred >> 1) > 1) 1347ae8064acSJohn Heffner goto send_now; 1348908a75c1SDavid S. Miller 1349c1b4a7e6SDavid S. Miller in_flight = tcp_packets_in_flight(tp); 1350c1b4a7e6SDavid S. Miller 1351056834d9SIlpo Järvinen BUG_ON(tcp_skb_pcount(skb) <= 1 || (tp->snd_cwnd <= in_flight)); 1352c1b4a7e6SDavid S. Miller 135390840defSIlpo Järvinen send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 1354c1b4a7e6SDavid S. Miller 1355c1b4a7e6SDavid S. Miller /* From in_flight test above, we know that cwnd > in_flight. */ 1356c1b4a7e6SDavid S. Miller cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache; 1357c1b4a7e6SDavid S. Miller 1358c1b4a7e6SDavid S. Miller limit = min(send_win, cong_win); 1359c1b4a7e6SDavid S. Miller 1360ba244fe9SDavid S. Miller /* If a full-sized TSO skb can be sent, do it. */ 136182cc1a7aSPeter P Waskiewicz Jr if (limit >= sk->sk_gso_max_size) 1362ae8064acSJohn Heffner goto send_now; 1363ba244fe9SDavid S. Miller 136462ad2761SIlpo Järvinen /* Middle in queue won't get any more data, full sendable already? */ 136562ad2761SIlpo Järvinen if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len)) 136662ad2761SIlpo Järvinen goto send_now; 136762ad2761SIlpo Järvinen 1368c1b4a7e6SDavid S. Miller if (sysctl_tcp_tso_win_divisor) { 1369c1b4a7e6SDavid S. Miller u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache); 1370c1b4a7e6SDavid S. Miller 1371c1b4a7e6SDavid S. Miller /* If at least some fraction of a window is available, 1372c1b4a7e6SDavid S. Miller * just use it. 1373c1b4a7e6SDavid S. Miller */ 1374c1b4a7e6SDavid S. Miller chunk /= sysctl_tcp_tso_win_divisor; 1375c1b4a7e6SDavid S. Miller if (limit >= chunk) 1376ae8064acSJohn Heffner goto send_now; 1377c1b4a7e6SDavid S. Miller } else { 1378c1b4a7e6SDavid S. Miller /* Different approach, try not to defer past a single 1379c1b4a7e6SDavid S. Miller * ACK. Receiver should ACK every other full sized 1380c1b4a7e6SDavid S. Miller * frame, so if we have space for more than 3 frames 1381c1b4a7e6SDavid S. Miller * then send now. 1382c1b4a7e6SDavid S. Miller */ 1383c1b4a7e6SDavid S. Miller if (limit > tcp_max_burst(tp) * tp->mss_cache) 1384ae8064acSJohn Heffner goto send_now; 1385c1b4a7e6SDavid S. Miller } 1386c1b4a7e6SDavid S. Miller 1387c1b4a7e6SDavid S. Miller /* Ok, it looks like it is advisable to defer. */ 1388ae8064acSJohn Heffner tp->tso_deferred = 1 | (jiffies << 1); 1389ae8064acSJohn Heffner 1390c1b4a7e6SDavid S. Miller return 1; 1391ae8064acSJohn Heffner 1392ae8064acSJohn Heffner send_now: 1393ae8064acSJohn Heffner tp->tso_deferred = 0; 1394ae8064acSJohn Heffner return 0; 1395c1b4a7e6SDavid S. Miller } 1396c1b4a7e6SDavid S. Miller 13975d424d5aSJohn Heffner /* Create a new MTU probe if we are ready. 13985d424d5aSJohn Heffner * Returns 0 if we should wait to probe (no cwnd available), 13995d424d5aSJohn Heffner * 1 if a probe was sent, 1400056834d9SIlpo Järvinen * -1 otherwise 1401056834d9SIlpo Järvinen */ 14025d424d5aSJohn Heffner static int tcp_mtu_probe(struct sock *sk) 14035d424d5aSJohn Heffner { 14045d424d5aSJohn Heffner struct tcp_sock *tp = tcp_sk(sk); 14055d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 14065d424d5aSJohn Heffner struct sk_buff *skb, *nskb, *next; 14075d424d5aSJohn Heffner int len; 14085d424d5aSJohn Heffner int probe_size; 140991cc17c0SIlpo Järvinen int size_needed; 14105d424d5aSJohn Heffner int copy; 14115d424d5aSJohn Heffner int mss_now; 14125d424d5aSJohn Heffner 14135d424d5aSJohn Heffner /* Not currently probing/verifying, 14145d424d5aSJohn Heffner * not in recovery, 14155d424d5aSJohn Heffner * have enough cwnd, and 14165d424d5aSJohn Heffner * not SACKing (the variable headers throw things off) */ 14175d424d5aSJohn Heffner if (!icsk->icsk_mtup.enabled || 14185d424d5aSJohn Heffner icsk->icsk_mtup.probe_size || 14195d424d5aSJohn Heffner inet_csk(sk)->icsk_ca_state != TCP_CA_Open || 14205d424d5aSJohn Heffner tp->snd_cwnd < 11 || 1421cabeccbdSIlpo Järvinen tp->rx_opt.num_sacks || tp->rx_opt.dsack) 14225d424d5aSJohn Heffner return -1; 14235d424d5aSJohn Heffner 14245d424d5aSJohn Heffner /* Very simple search strategy: just double the MSS. */ 14255d424d5aSJohn Heffner mss_now = tcp_current_mss(sk, 0); 14265d424d5aSJohn Heffner probe_size = 2 * tp->mss_cache; 142791cc17c0SIlpo Järvinen size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache; 14285d424d5aSJohn Heffner if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high)) { 14295d424d5aSJohn Heffner /* TODO: set timer for probe_converge_event */ 14305d424d5aSJohn Heffner return -1; 14315d424d5aSJohn Heffner } 14325d424d5aSJohn Heffner 14335d424d5aSJohn Heffner /* Have enough data in the send queue to probe? */ 14347f9c33e5SIlpo Järvinen if (tp->write_seq - tp->snd_nxt < size_needed) 14355d424d5aSJohn Heffner return -1; 14365d424d5aSJohn Heffner 143791cc17c0SIlpo Järvinen if (tp->snd_wnd < size_needed) 14385d424d5aSJohn Heffner return -1; 143990840defSIlpo Järvinen if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp))) 14405d424d5aSJohn Heffner return 0; 14415d424d5aSJohn Heffner 1442d67c58e9SIlpo Järvinen /* Do we need to wait to drain cwnd? With none in flight, don't stall */ 1443d67c58e9SIlpo Järvinen if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) { 1444d67c58e9SIlpo Järvinen if (!tcp_packets_in_flight(tp)) 14455d424d5aSJohn Heffner return -1; 14465d424d5aSJohn Heffner else 14475d424d5aSJohn Heffner return 0; 14485d424d5aSJohn Heffner } 14495d424d5aSJohn Heffner 14505d424d5aSJohn Heffner /* We're allowed to probe. Build it now. */ 14515d424d5aSJohn Heffner if ((nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC)) == NULL) 14525d424d5aSJohn Heffner return -1; 14533ab224beSHideo Aoki sk->sk_wmem_queued += nskb->truesize; 14543ab224beSHideo Aoki sk_mem_charge(sk, nskb->truesize); 14555d424d5aSJohn Heffner 1456fe067e8aSDavid S. Miller skb = tcp_send_head(sk); 14575d424d5aSJohn Heffner 14585d424d5aSJohn Heffner TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq; 14595d424d5aSJohn Heffner TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size; 14605d424d5aSJohn Heffner TCP_SKB_CB(nskb)->flags = TCPCB_FLAG_ACK; 14615d424d5aSJohn Heffner TCP_SKB_CB(nskb)->sacked = 0; 14625d424d5aSJohn Heffner nskb->csum = 0; 146384fa7933SPatrick McHardy nskb->ip_summed = skb->ip_summed; 14645d424d5aSJohn Heffner 146550c4817eSIlpo Järvinen tcp_insert_write_queue_before(nskb, skb, sk); 146650c4817eSIlpo Järvinen 14675d424d5aSJohn Heffner len = 0; 1468234b6860SIlpo Järvinen tcp_for_write_queue_from_safe(skb, next, sk) { 14695d424d5aSJohn Heffner copy = min_t(int, skb->len, probe_size - len); 14705d424d5aSJohn Heffner if (nskb->ip_summed) 14715d424d5aSJohn Heffner skb_copy_bits(skb, 0, skb_put(nskb, copy), copy); 14725d424d5aSJohn Heffner else 14735d424d5aSJohn Heffner nskb->csum = skb_copy_and_csum_bits(skb, 0, 1474056834d9SIlpo Järvinen skb_put(nskb, copy), 1475056834d9SIlpo Järvinen copy, nskb->csum); 14765d424d5aSJohn Heffner 14775d424d5aSJohn Heffner if (skb->len <= copy) { 14785d424d5aSJohn Heffner /* We've eaten all the data from this skb. 14795d424d5aSJohn Heffner * Throw it away. */ 14805d424d5aSJohn Heffner TCP_SKB_CB(nskb)->flags |= TCP_SKB_CB(skb)->flags; 1481fe067e8aSDavid S. Miller tcp_unlink_write_queue(skb, sk); 14823ab224beSHideo Aoki sk_wmem_free_skb(sk, skb); 14835d424d5aSJohn Heffner } else { 14845d424d5aSJohn Heffner TCP_SKB_CB(nskb)->flags |= TCP_SKB_CB(skb)->flags & 14855d424d5aSJohn Heffner ~(TCPCB_FLAG_FIN|TCPCB_FLAG_PSH); 14865d424d5aSJohn Heffner if (!skb_shinfo(skb)->nr_frags) { 14875d424d5aSJohn Heffner skb_pull(skb, copy); 148884fa7933SPatrick McHardy if (skb->ip_summed != CHECKSUM_PARTIAL) 1489056834d9SIlpo Järvinen skb->csum = csum_partial(skb->data, 1490056834d9SIlpo Järvinen skb->len, 0); 14915d424d5aSJohn Heffner } else { 14925d424d5aSJohn Heffner __pskb_trim_head(skb, copy); 14935d424d5aSJohn Heffner tcp_set_skb_tso_segs(sk, skb, mss_now); 14945d424d5aSJohn Heffner } 14955d424d5aSJohn Heffner TCP_SKB_CB(skb)->seq += copy; 14965d424d5aSJohn Heffner } 14975d424d5aSJohn Heffner 14985d424d5aSJohn Heffner len += copy; 1499234b6860SIlpo Järvinen 1500234b6860SIlpo Järvinen if (len >= probe_size) 1501234b6860SIlpo Järvinen break; 15025d424d5aSJohn Heffner } 15035d424d5aSJohn Heffner tcp_init_tso_segs(sk, nskb, nskb->len); 15045d424d5aSJohn Heffner 15055d424d5aSJohn Heffner /* We're ready to send. If this fails, the probe will 15065d424d5aSJohn Heffner * be resegmented into mss-sized pieces by tcp_write_xmit(). */ 15075d424d5aSJohn Heffner TCP_SKB_CB(nskb)->when = tcp_time_stamp; 15085d424d5aSJohn Heffner if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) { 15095d424d5aSJohn Heffner /* Decrement cwnd here because we are sending 15105d424d5aSJohn Heffner * effectively two packets. */ 15115d424d5aSJohn Heffner tp->snd_cwnd--; 151266f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, nskb); 15135d424d5aSJohn Heffner 15145d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len); 15150e7b1368SJohn Heffner tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq; 15160e7b1368SJohn Heffner tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq; 15175d424d5aSJohn Heffner 15185d424d5aSJohn Heffner return 1; 15195d424d5aSJohn Heffner } 15205d424d5aSJohn Heffner 15215d424d5aSJohn Heffner return -1; 15225d424d5aSJohn Heffner } 15235d424d5aSJohn Heffner 15241da177e4SLinus Torvalds /* This routine writes packets to the network. It advances the 15251da177e4SLinus Torvalds * send_head. This happens as incoming acks open up the remote 15261da177e4SLinus Torvalds * window for us. 15271da177e4SLinus Torvalds * 1528f8269a49SIlpo Järvinen * LARGESEND note: !tcp_urg_mode is overkill, only frames between 1529f8269a49SIlpo Järvinen * snd_up-64k-mss .. snd_up cannot be large. However, taking into 1530f8269a49SIlpo Järvinen * account rare use of URG, this is not a big flaw. 1531f8269a49SIlpo Järvinen * 15321da177e4SLinus Torvalds * Returns 1, if no segments are in flight and we have queued segments, but 15331da177e4SLinus Torvalds * cannot send anything now because of SWS or another problem. 15341da177e4SLinus Torvalds */ 1535d5dd9175SIlpo Järvinen static int tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle, 1536d5dd9175SIlpo Järvinen int push_one, gfp_t gfp) 15371da177e4SLinus Torvalds { 15381da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 153992df7b51SDavid S. Miller struct sk_buff *skb; 1540c1b4a7e6SDavid S. Miller unsigned int tso_segs, sent_pkts; 1541c1b4a7e6SDavid S. Miller int cwnd_quota; 15425d424d5aSJohn Heffner int result; 15431da177e4SLinus Torvalds 1544c1b4a7e6SDavid S. Miller sent_pkts = 0; 15455d424d5aSJohn Heffner 1546d5dd9175SIlpo Järvinen if (!push_one) { 15475d424d5aSJohn Heffner /* Do MTU probing. */ 1548d5dd9175SIlpo Järvinen result = tcp_mtu_probe(sk); 1549d5dd9175SIlpo Järvinen if (!result) { 15505d424d5aSJohn Heffner return 0; 15515d424d5aSJohn Heffner } else if (result > 0) { 15525d424d5aSJohn Heffner sent_pkts = 1; 15535d424d5aSJohn Heffner } 1554d5dd9175SIlpo Järvinen } 15555d424d5aSJohn Heffner 1556fe067e8aSDavid S. Miller while ((skb = tcp_send_head(sk))) { 1557c8ac3774SHerbert Xu unsigned int limit; 1558c8ac3774SHerbert Xu 1559b68e9f85SHerbert Xu tso_segs = tcp_init_tso_segs(sk, skb, mss_now); 1560c1b4a7e6SDavid S. Miller BUG_ON(!tso_segs); 1561c1b4a7e6SDavid S. Miller 1562b68e9f85SHerbert Xu cwnd_quota = tcp_cwnd_test(tp, skb); 1563b68e9f85SHerbert Xu if (!cwnd_quota) 1564b68e9f85SHerbert Xu break; 1565b68e9f85SHerbert Xu 1566b68e9f85SHerbert Xu if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now))) 1567b68e9f85SHerbert Xu break; 1568b68e9f85SHerbert Xu 1569c1b4a7e6SDavid S. Miller if (tso_segs == 1) { 1570aa93466bSDavid S. Miller if (unlikely(!tcp_nagle_test(tp, skb, mss_now, 1571aa93466bSDavid S. Miller (tcp_skb_is_last(sk, skb) ? 1572aa93466bSDavid S. Miller nonagle : TCP_NAGLE_PUSH)))) 1573aa93466bSDavid S. Miller break; 1574c1b4a7e6SDavid S. Miller } else { 1575d5dd9175SIlpo Järvinen if (!push_one && tcp_tso_should_defer(sk, skb)) 1576aa93466bSDavid S. Miller break; 1577c1b4a7e6SDavid S. Miller } 1578aa93466bSDavid S. Miller 1579c8ac3774SHerbert Xu limit = mss_now; 1580f8269a49SIlpo Järvinen if (tso_segs > 1 && !tcp_urg_mode(tp)) 15810e3a4803SIlpo Järvinen limit = tcp_mss_split_point(sk, skb, mss_now, 15820e3a4803SIlpo Järvinen cwnd_quota); 1583c8ac3774SHerbert Xu 1584c8ac3774SHerbert Xu if (skb->len > limit && 1585c8ac3774SHerbert Xu unlikely(tso_fragment(sk, skb, limit, mss_now))) 15861da177e4SLinus Torvalds break; 15871da177e4SLinus Torvalds 15881da177e4SLinus Torvalds TCP_SKB_CB(skb)->when = tcp_time_stamp; 1589c1b4a7e6SDavid S. Miller 1590d5dd9175SIlpo Järvinen if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp))) 15911da177e4SLinus Torvalds break; 15921da177e4SLinus Torvalds 15931da177e4SLinus Torvalds /* Advance the send_head. This one is sent out. 15941da177e4SLinus Torvalds * This call will increment packets_out. 15951da177e4SLinus Torvalds */ 159666f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, skb); 15971da177e4SLinus Torvalds 15981da177e4SLinus Torvalds tcp_minshall_update(tp, mss_now, skb); 1599aa93466bSDavid S. Miller sent_pkts++; 1600d5dd9175SIlpo Järvinen 1601d5dd9175SIlpo Järvinen if (push_one) 1602d5dd9175SIlpo Järvinen break; 16031da177e4SLinus Torvalds } 16041da177e4SLinus Torvalds 1605aa93466bSDavid S. Miller if (likely(sent_pkts)) { 16069e412ba7SIlpo Järvinen tcp_cwnd_validate(sk); 16071da177e4SLinus Torvalds return 0; 16081da177e4SLinus Torvalds } 1609fe067e8aSDavid S. Miller return !tp->packets_out && tcp_send_head(sk); 16101da177e4SLinus Torvalds } 16111da177e4SLinus Torvalds 1612a762a980SDavid S. Miller /* Push out any pending frames which were held back due to 1613a762a980SDavid S. Miller * TCP_CORK or attempt at coalescing tiny packets. 1614a762a980SDavid S. Miller * The socket must be locked by the caller. 1615a762a980SDavid S. Miller */ 16169e412ba7SIlpo Järvinen void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss, 16179e412ba7SIlpo Järvinen int nonagle) 1618a762a980SDavid S. Miller { 1619fe067e8aSDavid S. Miller struct sk_buff *skb = tcp_send_head(sk); 1620a762a980SDavid S. Miller 1621726e07a8SIlpo Järvinen if (!skb) 1622726e07a8SIlpo Järvinen return; 1623726e07a8SIlpo Järvinen 1624726e07a8SIlpo Järvinen /* If we are closed, the bytes will have to remain here. 1625726e07a8SIlpo Järvinen * In time closedown will finish, we empty the write queue and 1626726e07a8SIlpo Järvinen * all will be happy. 1627726e07a8SIlpo Järvinen */ 1628726e07a8SIlpo Järvinen if (unlikely(sk->sk_state == TCP_CLOSE)) 1629726e07a8SIlpo Järvinen return; 1630726e07a8SIlpo Järvinen 1631d5dd9175SIlpo Järvinen if (tcp_write_xmit(sk, cur_mss, nonagle, 0, GFP_ATOMIC)) 16329e412ba7SIlpo Järvinen tcp_check_probe_timer(sk); 1633a762a980SDavid S. Miller } 1634a762a980SDavid S. Miller 1635c1b4a7e6SDavid S. Miller /* Send _single_ skb sitting at the send head. This function requires 1636c1b4a7e6SDavid S. Miller * true push pending frames to setup probe timer etc. 1637c1b4a7e6SDavid S. Miller */ 1638c1b4a7e6SDavid S. Miller void tcp_push_one(struct sock *sk, unsigned int mss_now) 1639c1b4a7e6SDavid S. Miller { 1640fe067e8aSDavid S. Miller struct sk_buff *skb = tcp_send_head(sk); 1641c1b4a7e6SDavid S. Miller 1642c1b4a7e6SDavid S. Miller BUG_ON(!skb || skb->len < mss_now); 1643c1b4a7e6SDavid S. Miller 1644d5dd9175SIlpo Järvinen tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation); 1645c1b4a7e6SDavid S. Miller } 1646c1b4a7e6SDavid S. Miller 16471da177e4SLinus Torvalds /* This function returns the amount that we can raise the 16481da177e4SLinus Torvalds * usable window based on the following constraints 16491da177e4SLinus Torvalds * 16501da177e4SLinus Torvalds * 1. The window can never be shrunk once it is offered (RFC 793) 16511da177e4SLinus Torvalds * 2. We limit memory per socket 16521da177e4SLinus Torvalds * 16531da177e4SLinus Torvalds * RFC 1122: 16541da177e4SLinus Torvalds * "the suggested [SWS] avoidance algorithm for the receiver is to keep 16551da177e4SLinus Torvalds * RECV.NEXT + RCV.WIN fixed until: 16561da177e4SLinus Torvalds * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)" 16571da177e4SLinus Torvalds * 16581da177e4SLinus Torvalds * i.e. don't raise the right edge of the window until you can raise 16591da177e4SLinus Torvalds * it at least MSS bytes. 16601da177e4SLinus Torvalds * 16611da177e4SLinus Torvalds * Unfortunately, the recommended algorithm breaks header prediction, 16621da177e4SLinus Torvalds * since header prediction assumes th->window stays fixed. 16631da177e4SLinus Torvalds * 16641da177e4SLinus Torvalds * Strictly speaking, keeping th->window fixed violates the receiver 16651da177e4SLinus Torvalds * side SWS prevention criteria. The problem is that under this rule 16661da177e4SLinus Torvalds * a stream of single byte packets will cause the right side of the 16671da177e4SLinus Torvalds * window to always advance by a single byte. 16681da177e4SLinus Torvalds * 16691da177e4SLinus Torvalds * Of course, if the sender implements sender side SWS prevention 16701da177e4SLinus Torvalds * then this will not be a problem. 16711da177e4SLinus Torvalds * 16721da177e4SLinus Torvalds * BSD seems to make the following compromise: 16731da177e4SLinus Torvalds * 16741da177e4SLinus Torvalds * If the free space is less than the 1/4 of the maximum 16751da177e4SLinus Torvalds * space available and the free space is less than 1/2 mss, 16761da177e4SLinus Torvalds * then set the window to 0. 16771da177e4SLinus Torvalds * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ] 16781da177e4SLinus Torvalds * Otherwise, just prevent the window from shrinking 16791da177e4SLinus Torvalds * and from being larger than the largest representable value. 16801da177e4SLinus Torvalds * 16811da177e4SLinus Torvalds * This prevents incremental opening of the window in the regime 16821da177e4SLinus Torvalds * where TCP is limited by the speed of the reader side taking 16831da177e4SLinus Torvalds * data out of the TCP receive queue. It does nothing about 16841da177e4SLinus Torvalds * those cases where the window is constrained on the sender side 16851da177e4SLinus Torvalds * because the pipeline is full. 16861da177e4SLinus Torvalds * 16871da177e4SLinus Torvalds * BSD also seems to "accidentally" limit itself to windows that are a 16881da177e4SLinus Torvalds * multiple of MSS, at least until the free space gets quite small. 16891da177e4SLinus Torvalds * This would appear to be a side effect of the mbuf implementation. 16901da177e4SLinus Torvalds * Combining these two algorithms results in the observed behavior 16911da177e4SLinus Torvalds * of having a fixed window size at almost all times. 16921da177e4SLinus Torvalds * 16931da177e4SLinus Torvalds * Below we obtain similar behavior by forcing the offered window to 16941da177e4SLinus Torvalds * a multiple of the mss when it is feasible to do so. 16951da177e4SLinus Torvalds * 16961da177e4SLinus Torvalds * Note, we don't "adjust" for TIMESTAMP or SACK option bytes. 16971da177e4SLinus Torvalds * Regular options like TIMESTAMP are taken into account. 16981da177e4SLinus Torvalds */ 16991da177e4SLinus Torvalds u32 __tcp_select_window(struct sock *sk) 17001da177e4SLinus Torvalds { 1701463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 17021da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 1703caa20d9aSStephen Hemminger /* MSS for the peer's data. Previous versions used mss_clamp 17041da177e4SLinus Torvalds * here. I don't know if the value based on our guesses 17051da177e4SLinus Torvalds * of peer's MSS is better for the performance. It's more correct 17061da177e4SLinus Torvalds * but may be worse for the performance because of rcv_mss 17071da177e4SLinus Torvalds * fluctuations. --SAW 1998/11/1 17081da177e4SLinus Torvalds */ 1709463c84b9SArnaldo Carvalho de Melo int mss = icsk->icsk_ack.rcv_mss; 17101da177e4SLinus Torvalds int free_space = tcp_space(sk); 17111da177e4SLinus Torvalds int full_space = min_t(int, tp->window_clamp, tcp_full_space(sk)); 17121da177e4SLinus Torvalds int window; 17131da177e4SLinus Torvalds 17141da177e4SLinus Torvalds if (mss > full_space) 17151da177e4SLinus Torvalds mss = full_space; 17161da177e4SLinus Torvalds 1717b92edbe0SEric Dumazet if (free_space < (full_space >> 1)) { 1718463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.quick = 0; 17191da177e4SLinus Torvalds 17201da177e4SLinus Torvalds if (tcp_memory_pressure) 1721056834d9SIlpo Järvinen tp->rcv_ssthresh = min(tp->rcv_ssthresh, 1722056834d9SIlpo Järvinen 4U * tp->advmss); 17231da177e4SLinus Torvalds 17241da177e4SLinus Torvalds if (free_space < mss) 17251da177e4SLinus Torvalds return 0; 17261da177e4SLinus Torvalds } 17271da177e4SLinus Torvalds 17281da177e4SLinus Torvalds if (free_space > tp->rcv_ssthresh) 17291da177e4SLinus Torvalds free_space = tp->rcv_ssthresh; 17301da177e4SLinus Torvalds 17311da177e4SLinus Torvalds /* Don't do rounding if we are using window scaling, since the 17321da177e4SLinus Torvalds * scaled window will not line up with the MSS boundary anyway. 17331da177e4SLinus Torvalds */ 17341da177e4SLinus Torvalds window = tp->rcv_wnd; 17351da177e4SLinus Torvalds if (tp->rx_opt.rcv_wscale) { 17361da177e4SLinus Torvalds window = free_space; 17371da177e4SLinus Torvalds 17381da177e4SLinus Torvalds /* Advertise enough space so that it won't get scaled away. 17391da177e4SLinus Torvalds * Import case: prevent zero window announcement if 17401da177e4SLinus Torvalds * 1<<rcv_wscale > mss. 17411da177e4SLinus Torvalds */ 17421da177e4SLinus Torvalds if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window) 17431da177e4SLinus Torvalds window = (((window >> tp->rx_opt.rcv_wscale) + 1) 17441da177e4SLinus Torvalds << tp->rx_opt.rcv_wscale); 17451da177e4SLinus Torvalds } else { 17461da177e4SLinus Torvalds /* Get the largest window that is a nice multiple of mss. 17471da177e4SLinus Torvalds * Window clamp already applied above. 17481da177e4SLinus Torvalds * If our current window offering is within 1 mss of the 17491da177e4SLinus Torvalds * free space we just keep it. This prevents the divide 17501da177e4SLinus Torvalds * and multiply from happening most of the time. 17511da177e4SLinus Torvalds * We also don't do any window rounding when the free space 17521da177e4SLinus Torvalds * is too small. 17531da177e4SLinus Torvalds */ 17541da177e4SLinus Torvalds if (window <= free_space - mss || window > free_space) 17551da177e4SLinus Torvalds window = (free_space / mss) * mss; 175684565070SJohn Heffner else if (mss == full_space && 1757b92edbe0SEric Dumazet free_space > window + (full_space >> 1)) 175884565070SJohn Heffner window = free_space; 17591da177e4SLinus Torvalds } 17601da177e4SLinus Torvalds 17611da177e4SLinus Torvalds return window; 17621da177e4SLinus Torvalds } 17631da177e4SLinus Torvalds 17644a17fc3aSIlpo Järvinen /* Collapses two adjacent SKB's during retransmission. */ 17654a17fc3aSIlpo Järvinen static void tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb) 17661da177e4SLinus Torvalds { 17671da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 1768fe067e8aSDavid S. Miller struct sk_buff *next_skb = tcp_write_queue_next(sk, skb); 1769058dc334SIlpo Järvinen int skb_size, next_skb_size; 17701da177e4SLinus Torvalds 1771058dc334SIlpo Järvinen skb_size = skb->len; 1772058dc334SIlpo Järvinen next_skb_size = next_skb->len; 17731da177e4SLinus Torvalds 1774058dc334SIlpo Järvinen BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1); 17751da177e4SLinus Torvalds 17766859d494SIlpo Järvinen tcp_highest_sack_combine(sk, next_skb, skb); 1777a6963a6bSIlpo Järvinen 1778fe067e8aSDavid S. Miller tcp_unlink_write_queue(next_skb, sk); 17791da177e4SLinus Torvalds 1780058dc334SIlpo Järvinen skb_copy_from_linear_data(next_skb, skb_put(skb, next_skb_size), 17811a4e2d09SArnaldo Carvalho de Melo next_skb_size); 17821da177e4SLinus Torvalds 178352d570aaSJarek Poplawski if (next_skb->ip_summed == CHECKSUM_PARTIAL) 178452d570aaSJarek Poplawski skb->ip_summed = CHECKSUM_PARTIAL; 17851da177e4SLinus Torvalds 178684fa7933SPatrick McHardy if (skb->ip_summed != CHECKSUM_PARTIAL) 17871da177e4SLinus Torvalds skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size); 17881da177e4SLinus Torvalds 17891da177e4SLinus Torvalds /* Update sequence range on original skb. */ 17901da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq; 17911da177e4SLinus Torvalds 1792e6c7d085SIlpo Järvinen /* Merge over control information. This moves PSH/FIN etc. over */ 1793e6c7d085SIlpo Järvinen TCP_SKB_CB(skb)->flags |= TCP_SKB_CB(next_skb)->flags; 17941da177e4SLinus Torvalds 17951da177e4SLinus Torvalds /* All done, get rid of second SKB and account for it so 17961da177e4SLinus Torvalds * packet counting does not break. 17971da177e4SLinus Torvalds */ 17984828e7f4SIlpo Järvinen TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS; 17991da177e4SLinus Torvalds if (TCP_SKB_CB(next_skb)->sacked & TCPCB_SACKED_RETRANS) 18001da177e4SLinus Torvalds tp->retrans_out -= tcp_skb_pcount(next_skb); 1801b5860bbaSIlpo Järvinen if (TCP_SKB_CB(next_skb)->sacked & TCPCB_LOST) 18021da177e4SLinus Torvalds tp->lost_out -= tcp_skb_pcount(next_skb); 18031da177e4SLinus Torvalds /* Reno case is special. Sigh... */ 1804e60402d0SIlpo Järvinen if (tcp_is_reno(tp) && tp->sacked_out) 18051da177e4SLinus Torvalds tcp_dec_pcount_approx(&tp->sacked_out, next_skb); 18061da177e4SLinus Torvalds 1807a47e5a98SIlpo Järvinen tcp_adjust_fackets_out(sk, next_skb, tcp_skb_pcount(next_skb)); 1808e9144bd8SIlpo Järvinen tp->packets_out -= tcp_skb_pcount(next_skb); 1809b7689205SIlpo Järvinen 1810b7689205SIlpo Järvinen /* changed transmit queue under us so clear hints */ 1811ef9da47cSIlpo Järvinen tcp_clear_retrans_hints_partial(tp); 1812ef9da47cSIlpo Järvinen if (next_skb == tp->retransmit_skb_hint) 1813ef9da47cSIlpo Järvinen tp->retransmit_skb_hint = skb; 1814b7689205SIlpo Järvinen 18153ab224beSHideo Aoki sk_wmem_free_skb(sk, next_skb); 18161da177e4SLinus Torvalds } 18171da177e4SLinus Torvalds 18184a17fc3aSIlpo Järvinen static int tcp_can_collapse(struct sock *sk, struct sk_buff *skb) 18194a17fc3aSIlpo Järvinen { 18204a17fc3aSIlpo Järvinen if (tcp_skb_pcount(skb) > 1) 18214a17fc3aSIlpo Järvinen return 0; 18224a17fc3aSIlpo Järvinen /* TODO: SACK collapsing could be used to remove this condition */ 18234a17fc3aSIlpo Järvinen if (skb_shinfo(skb)->nr_frags != 0) 18244a17fc3aSIlpo Järvinen return 0; 18254a17fc3aSIlpo Järvinen if (skb_cloned(skb)) 18264a17fc3aSIlpo Järvinen return 0; 18274a17fc3aSIlpo Järvinen if (skb == tcp_send_head(sk)) 18284a17fc3aSIlpo Järvinen return 0; 18294a17fc3aSIlpo Järvinen /* Some heurestics for collapsing over SACK'd could be invented */ 18304a17fc3aSIlpo Järvinen if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) 18314a17fc3aSIlpo Järvinen return 0; 18324a17fc3aSIlpo Järvinen 18334a17fc3aSIlpo Järvinen return 1; 18344a17fc3aSIlpo Järvinen } 18354a17fc3aSIlpo Järvinen 18364a17fc3aSIlpo Järvinen static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to, 18374a17fc3aSIlpo Järvinen int space) 18384a17fc3aSIlpo Järvinen { 18394a17fc3aSIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 18404a17fc3aSIlpo Järvinen struct sk_buff *skb = to, *tmp; 18414a17fc3aSIlpo Järvinen int first = 1; 18424a17fc3aSIlpo Järvinen 18434a17fc3aSIlpo Järvinen if (!sysctl_tcp_retrans_collapse) 18444a17fc3aSIlpo Järvinen return; 18454a17fc3aSIlpo Järvinen if (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_SYN) 18464a17fc3aSIlpo Järvinen return; 18474a17fc3aSIlpo Järvinen 18484a17fc3aSIlpo Järvinen tcp_for_write_queue_from_safe(skb, tmp, sk) { 18494a17fc3aSIlpo Järvinen if (!tcp_can_collapse(sk, skb)) 18504a17fc3aSIlpo Järvinen break; 18514a17fc3aSIlpo Järvinen 18524a17fc3aSIlpo Järvinen space -= skb->len; 18534a17fc3aSIlpo Järvinen 18544a17fc3aSIlpo Järvinen if (first) { 18554a17fc3aSIlpo Järvinen first = 0; 18564a17fc3aSIlpo Järvinen continue; 18574a17fc3aSIlpo Järvinen } 18584a17fc3aSIlpo Järvinen 18594a17fc3aSIlpo Järvinen if (space < 0) 18604a17fc3aSIlpo Järvinen break; 18614a17fc3aSIlpo Järvinen /* Punt if not enough space exists in the first SKB for 18624a17fc3aSIlpo Järvinen * the data in the second 18634a17fc3aSIlpo Järvinen */ 18644a17fc3aSIlpo Järvinen if (skb->len > skb_tailroom(to)) 18654a17fc3aSIlpo Järvinen break; 18664a17fc3aSIlpo Järvinen 18674a17fc3aSIlpo Järvinen if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp))) 18684a17fc3aSIlpo Järvinen break; 18694a17fc3aSIlpo Järvinen 18704a17fc3aSIlpo Järvinen tcp_collapse_retrans(sk, to); 18714a17fc3aSIlpo Järvinen } 18724a17fc3aSIlpo Järvinen } 18734a17fc3aSIlpo Järvinen 18741da177e4SLinus Torvalds /* This retransmits one SKB. Policy decisions and retransmit queue 18751da177e4SLinus Torvalds * state updates are done by the caller. Returns non-zero if an 18761da177e4SLinus Torvalds * error occurred which prevented the send. 18771da177e4SLinus Torvalds */ 18781da177e4SLinus Torvalds int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb) 18791da177e4SLinus Torvalds { 18801da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 18815d424d5aSJohn Heffner struct inet_connection_sock *icsk = inet_csk(sk); 18827d227cd2SSridhar Samudrala unsigned int cur_mss; 18831da177e4SLinus Torvalds int err; 18841da177e4SLinus Torvalds 18855d424d5aSJohn Heffner /* Inconslusive MTU probe */ 18865d424d5aSJohn Heffner if (icsk->icsk_mtup.probe_size) { 18875d424d5aSJohn Heffner icsk->icsk_mtup.probe_size = 0; 18885d424d5aSJohn Heffner } 18895d424d5aSJohn Heffner 18901da177e4SLinus Torvalds /* Do not sent more than we queued. 1/4 is reserved for possible 1891caa20d9aSStephen Hemminger * copying overhead: fragmentation, tunneling, mangling etc. 18921da177e4SLinus Torvalds */ 18931da177e4SLinus Torvalds if (atomic_read(&sk->sk_wmem_alloc) > 18941da177e4SLinus Torvalds min(sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), sk->sk_sndbuf)) 18951da177e4SLinus Torvalds return -EAGAIN; 18961da177e4SLinus Torvalds 18971da177e4SLinus Torvalds if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) { 18981da177e4SLinus Torvalds if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) 18991da177e4SLinus Torvalds BUG(); 19001da177e4SLinus Torvalds if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq)) 19011da177e4SLinus Torvalds return -ENOMEM; 19021da177e4SLinus Torvalds } 19031da177e4SLinus Torvalds 19047d227cd2SSridhar Samudrala if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk)) 19057d227cd2SSridhar Samudrala return -EHOSTUNREACH; /* Routing failure or similar. */ 19067d227cd2SSridhar Samudrala 19077d227cd2SSridhar Samudrala cur_mss = tcp_current_mss(sk, 0); 19087d227cd2SSridhar Samudrala 19091da177e4SLinus Torvalds /* If receiver has shrunk his window, and skb is out of 19101da177e4SLinus Torvalds * new window, do not retransmit it. The exception is the 19111da177e4SLinus Torvalds * case, when window is shrunk to zero. In this case 19121da177e4SLinus Torvalds * our retransmit serves as a zero window probe. 19131da177e4SLinus Torvalds */ 191490840defSIlpo Järvinen if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) 19151da177e4SLinus Torvalds && TCP_SKB_CB(skb)->seq != tp->snd_una) 19161da177e4SLinus Torvalds return -EAGAIN; 19171da177e4SLinus Torvalds 19181da177e4SLinus Torvalds if (skb->len > cur_mss) { 1919846998aeSDavid S. Miller if (tcp_fragment(sk, skb, cur_mss, cur_mss)) 19201da177e4SLinus Torvalds return -ENOMEM; /* We'll try again later. */ 192102276f3cSIlpo Järvinen } else { 192202276f3cSIlpo Järvinen tcp_init_tso_segs(sk, skb, cur_mss); 19231da177e4SLinus Torvalds } 19241da177e4SLinus Torvalds 19251da177e4SLinus Torvalds tcp_retrans_try_collapse(sk, skb, cur_mss); 19261da177e4SLinus Torvalds 19271da177e4SLinus Torvalds /* Some Solaris stacks overoptimize and ignore the FIN on a 19281da177e4SLinus Torvalds * retransmit when old data is attached. So strip it off 19291da177e4SLinus Torvalds * since it is cheap to do so and saves bytes on the network. 19301da177e4SLinus Torvalds */ 19311da177e4SLinus Torvalds if (skb->len > 0 && 19321da177e4SLinus Torvalds (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN) && 19331da177e4SLinus Torvalds tp->snd_una == (TCP_SKB_CB(skb)->end_seq - 1)) { 19341da177e4SLinus Torvalds if (!pskb_trim(skb, 0)) { 1935e870a8efSIlpo Järvinen /* Reuse, even though it does some unnecessary work */ 1936e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, TCP_SKB_CB(skb)->end_seq - 1, 1937e870a8efSIlpo Järvinen TCP_SKB_CB(skb)->flags); 19381da177e4SLinus Torvalds skb->ip_summed = CHECKSUM_NONE; 19391da177e4SLinus Torvalds } 19401da177e4SLinus Torvalds } 19411da177e4SLinus Torvalds 19421da177e4SLinus Torvalds /* Make a copy, if the first transmission SKB clone we made 19431da177e4SLinus Torvalds * is still in somebody's hands, else make a clone. 19441da177e4SLinus Torvalds */ 19451da177e4SLinus Torvalds TCP_SKB_CB(skb)->when = tcp_time_stamp; 19461da177e4SLinus Torvalds 1947dfb4b9dcSDavid S. Miller err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 19481da177e4SLinus Torvalds 19491da177e4SLinus Torvalds if (err == 0) { 19501da177e4SLinus Torvalds /* Update global TCP statistics. */ 195181cc8a75SPavel Emelyanov TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS); 19521da177e4SLinus Torvalds 19531da177e4SLinus Torvalds tp->total_retrans++; 19541da177e4SLinus Torvalds 19551da177e4SLinus Torvalds #if FASTRETRANS_DEBUG > 0 19561da177e4SLinus Torvalds if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) { 19571da177e4SLinus Torvalds if (net_ratelimit()) 19581da177e4SLinus Torvalds printk(KERN_DEBUG "retrans_out leaked.\n"); 19591da177e4SLinus Torvalds } 19601da177e4SLinus Torvalds #endif 1961b08d6cb2SIlpo Järvinen if (!tp->retrans_out) 1962b08d6cb2SIlpo Järvinen tp->lost_retrans_low = tp->snd_nxt; 19631da177e4SLinus Torvalds TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS; 19641da177e4SLinus Torvalds tp->retrans_out += tcp_skb_pcount(skb); 19651da177e4SLinus Torvalds 19661da177e4SLinus Torvalds /* Save stamp of the first retransmit. */ 19671da177e4SLinus Torvalds if (!tp->retrans_stamp) 19681da177e4SLinus Torvalds tp->retrans_stamp = TCP_SKB_CB(skb)->when; 19691da177e4SLinus Torvalds 19701da177e4SLinus Torvalds tp->undo_retrans++; 19711da177e4SLinus Torvalds 19721da177e4SLinus Torvalds /* snd_nxt is stored to detect loss of retransmitted segment, 19731da177e4SLinus Torvalds * see tcp_input.c tcp_sacktag_write_queue(). 19741da177e4SLinus Torvalds */ 19751da177e4SLinus Torvalds TCP_SKB_CB(skb)->ack_seq = tp->snd_nxt; 19761da177e4SLinus Torvalds } 19771da177e4SLinus Torvalds return err; 19781da177e4SLinus Torvalds } 19791da177e4SLinus Torvalds 1980b5afe7bcSIlpo Järvinen static int tcp_can_forward_retransmit(struct sock *sk) 1981b5afe7bcSIlpo Järvinen { 1982b5afe7bcSIlpo Järvinen const struct inet_connection_sock *icsk = inet_csk(sk); 1983b5afe7bcSIlpo Järvinen struct tcp_sock *tp = tcp_sk(sk); 1984b5afe7bcSIlpo Järvinen 1985b5afe7bcSIlpo Järvinen /* Forward retransmissions are possible only during Recovery. */ 1986b5afe7bcSIlpo Järvinen if (icsk->icsk_ca_state != TCP_CA_Recovery) 1987b5afe7bcSIlpo Järvinen return 0; 1988b5afe7bcSIlpo Järvinen 1989b5afe7bcSIlpo Järvinen /* No forward retransmissions in Reno are possible. */ 1990b5afe7bcSIlpo Järvinen if (tcp_is_reno(tp)) 1991b5afe7bcSIlpo Järvinen return 0; 1992b5afe7bcSIlpo Järvinen 1993b5afe7bcSIlpo Järvinen /* Yeah, we have to make difficult choice between forward transmission 1994b5afe7bcSIlpo Järvinen * and retransmission... Both ways have their merits... 1995b5afe7bcSIlpo Järvinen * 1996b5afe7bcSIlpo Järvinen * For now we do not retransmit anything, while we have some new 1997b5afe7bcSIlpo Järvinen * segments to send. In the other cases, follow rule 3 for 1998b5afe7bcSIlpo Järvinen * NextSeg() specified in RFC3517. 1999b5afe7bcSIlpo Järvinen */ 2000b5afe7bcSIlpo Järvinen 2001b5afe7bcSIlpo Järvinen if (tcp_may_send_now(sk)) 2002b5afe7bcSIlpo Järvinen return 0; 2003b5afe7bcSIlpo Järvinen 2004b5afe7bcSIlpo Järvinen return 1; 2005b5afe7bcSIlpo Järvinen } 2006b5afe7bcSIlpo Järvinen 20071da177e4SLinus Torvalds /* This gets called after a retransmit timeout, and the initially 20081da177e4SLinus Torvalds * retransmitted data is acknowledged. It tries to continue 20091da177e4SLinus Torvalds * resending the rest of the retransmit queue, until either 20101da177e4SLinus Torvalds * we've sent it all or the congestion window limit is reached. 20111da177e4SLinus Torvalds * If doing SACK, the first ACK which comes back for a timeout 20121da177e4SLinus Torvalds * based retransmit packet might feed us FACK information again. 20131da177e4SLinus Torvalds * If so, we use it to avoid unnecessarily retransmissions. 20141da177e4SLinus Torvalds */ 20151da177e4SLinus Torvalds void tcp_xmit_retransmit_queue(struct sock *sk) 20161da177e4SLinus Torvalds { 20176687e988SArnaldo Carvalho de Melo const struct inet_connection_sock *icsk = inet_csk(sk); 20181da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 20191da177e4SLinus Torvalds struct sk_buff *skb; 20200e1c54c2SIlpo Järvinen struct sk_buff *hole = NULL; 2021618d9f25SIlpo Järvinen u32 last_lost; 202261eb55f4SIlpo Järvinen int mib_idx; 20230e1c54c2SIlpo Järvinen int fwd_rexmitting = 0; 20246a438bbeSStephen Hemminger 202508ebd172SIlpo Järvinen if (!tp->lost_out) 202608ebd172SIlpo Järvinen tp->retransmit_high = tp->snd_una; 202708ebd172SIlpo Järvinen 2028618d9f25SIlpo Järvinen if (tp->retransmit_skb_hint) { 20296a438bbeSStephen Hemminger skb = tp->retransmit_skb_hint; 2030618d9f25SIlpo Järvinen last_lost = TCP_SKB_CB(skb)->end_seq; 2031618d9f25SIlpo Järvinen if (after(last_lost, tp->retransmit_high)) 2032618d9f25SIlpo Järvinen last_lost = tp->retransmit_high; 2033618d9f25SIlpo Järvinen } else { 2034fe067e8aSDavid S. Miller skb = tcp_write_queue_head(sk); 2035618d9f25SIlpo Järvinen last_lost = tp->snd_una; 2036618d9f25SIlpo Järvinen } 20371da177e4SLinus Torvalds 2038fe067e8aSDavid S. Miller tcp_for_write_queue_from(skb, sk) { 20391da177e4SLinus Torvalds __u8 sacked = TCP_SKB_CB(skb)->sacked; 20401da177e4SLinus Torvalds 2041fe067e8aSDavid S. Miller if (skb == tcp_send_head(sk)) 2042fe067e8aSDavid S. Miller break; 20436a438bbeSStephen Hemminger /* we could do better than to assign each time */ 20440e1c54c2SIlpo Järvinen if (hole == NULL) 20456a438bbeSStephen Hemminger tp->retransmit_skb_hint = skb; 20466a438bbeSStephen Hemminger 20471da177e4SLinus Torvalds /* Assume this retransmit will generate 20481da177e4SLinus Torvalds * only one packet for congestion window 20491da177e4SLinus Torvalds * calculation purposes. This works because 20501da177e4SLinus Torvalds * tcp_retransmit_skb() will chop up the 20511da177e4SLinus Torvalds * packet to be MSS sized and all the 20521da177e4SLinus Torvalds * packet counting works out. 20531da177e4SLinus Torvalds */ 20541da177e4SLinus Torvalds if (tcp_packets_in_flight(tp) >= tp->snd_cwnd) 20551da177e4SLinus Torvalds return; 20560e1c54c2SIlpo Järvinen 20570e1c54c2SIlpo Järvinen if (fwd_rexmitting) { 20580e1c54c2SIlpo Järvinen begin_fwd: 20590e1c54c2SIlpo Järvinen if (!before(TCP_SKB_CB(skb)->seq, tcp_highest_sack_seq(tp))) 2060006f582cSIlpo Järvinen break; 20610e1c54c2SIlpo Järvinen mib_idx = LINUX_MIB_TCPFORWARDRETRANS; 20620e1c54c2SIlpo Järvinen 20630e1c54c2SIlpo Järvinen } else if (!before(TCP_SKB_CB(skb)->seq, tp->retransmit_high)) { 2064618d9f25SIlpo Järvinen tp->retransmit_high = last_lost; 20650e1c54c2SIlpo Järvinen if (!tcp_can_forward_retransmit(sk)) 20660e1c54c2SIlpo Järvinen break; 20670e1c54c2SIlpo Järvinen /* Backtrack if necessary to non-L'ed skb */ 20680e1c54c2SIlpo Järvinen if (hole != NULL) { 20690e1c54c2SIlpo Järvinen skb = hole; 20700e1c54c2SIlpo Järvinen hole = NULL; 20710e1c54c2SIlpo Järvinen } 20720e1c54c2SIlpo Järvinen fwd_rexmitting = 1; 20730e1c54c2SIlpo Järvinen goto begin_fwd; 20740e1c54c2SIlpo Järvinen 20750e1c54c2SIlpo Järvinen } else if (!(sacked & TCPCB_LOST)) { 2076ac11ba75SIlpo Järvinen if (hole == NULL && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED))) 20770e1c54c2SIlpo Järvinen hole = skb; 207861eb55f4SIlpo Järvinen continue; 20791da177e4SLinus Torvalds 20800e1c54c2SIlpo Järvinen } else { 2081618d9f25SIlpo Järvinen last_lost = TCP_SKB_CB(skb)->end_seq; 20820e1c54c2SIlpo Järvinen if (icsk->icsk_ca_state != TCP_CA_Loss) 20830e1c54c2SIlpo Järvinen mib_idx = LINUX_MIB_TCPFASTRETRANS; 20840e1c54c2SIlpo Järvinen else 20850e1c54c2SIlpo Järvinen mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS; 20860e1c54c2SIlpo Järvinen } 20870e1c54c2SIlpo Järvinen 20880e1c54c2SIlpo Järvinen if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS)) 208961eb55f4SIlpo Järvinen continue; 209040b215e5SPavel Emelyanov 2091f0ceb0edSIlpo Järvinen if (tcp_retransmit_skb(sk, skb)) 20921da177e4SLinus Torvalds return; 2093de0744afSPavel Emelyanov NET_INC_STATS_BH(sock_net(sk), mib_idx); 20941da177e4SLinus Torvalds 2095fe067e8aSDavid S. Miller if (skb == tcp_write_queue_head(sk)) 2096463c84b9SArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 20973f421baaSArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto, 20983f421baaSArnaldo Carvalho de Melo TCP_RTO_MAX); 20991da177e4SLinus Torvalds } 21001da177e4SLinus Torvalds } 21011da177e4SLinus Torvalds 21021da177e4SLinus Torvalds /* Send a fin. The caller locks the socket for us. This cannot be 21031da177e4SLinus Torvalds * allowed to fail queueing a FIN frame under any circumstances. 21041da177e4SLinus Torvalds */ 21051da177e4SLinus Torvalds void tcp_send_fin(struct sock *sk) 21061da177e4SLinus Torvalds { 21071da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 2108fe067e8aSDavid S. Miller struct sk_buff *skb = tcp_write_queue_tail(sk); 21091da177e4SLinus Torvalds int mss_now; 21101da177e4SLinus Torvalds 21111da177e4SLinus Torvalds /* Optimization, tack on the FIN if we have a queue of 21121da177e4SLinus Torvalds * unsent frames. But be careful about outgoing SACKS 21131da177e4SLinus Torvalds * and IP options. 21141da177e4SLinus Torvalds */ 21151da177e4SLinus Torvalds mss_now = tcp_current_mss(sk, 1); 21161da177e4SLinus Torvalds 2117fe067e8aSDavid S. Miller if (tcp_send_head(sk) != NULL) { 21181da177e4SLinus Torvalds TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_FIN; 21191da177e4SLinus Torvalds TCP_SKB_CB(skb)->end_seq++; 21201da177e4SLinus Torvalds tp->write_seq++; 21211da177e4SLinus Torvalds } else { 21221da177e4SLinus Torvalds /* Socket is locked, keep trying until memory is available. */ 21231da177e4SLinus Torvalds for (;;) { 2124d179cd12SDavid S. Miller skb = alloc_skb_fclone(MAX_TCP_HEADER, GFP_KERNEL); 21251da177e4SLinus Torvalds if (skb) 21261da177e4SLinus Torvalds break; 21271da177e4SLinus Torvalds yield(); 21281da177e4SLinus Torvalds } 21291da177e4SLinus Torvalds 21301da177e4SLinus Torvalds /* Reserve space for headers and prepare control bits. */ 21311da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 21321da177e4SLinus Torvalds /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */ 2133e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tp->write_seq, 2134e870a8efSIlpo Järvinen TCPCB_FLAG_ACK | TCPCB_FLAG_FIN); 21351da177e4SLinus Torvalds tcp_queue_skb(sk, skb); 21361da177e4SLinus Torvalds } 21379e412ba7SIlpo Järvinen __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_OFF); 21381da177e4SLinus Torvalds } 21391da177e4SLinus Torvalds 21401da177e4SLinus Torvalds /* We get here when a process closes a file descriptor (either due to 21411da177e4SLinus Torvalds * an explicit close() or as a byproduct of exit()'ing) and there 21421da177e4SLinus Torvalds * was unread data in the receive queue. This behavior is recommended 214365bb723cSGerrit Renker * by RFC 2525, section 2.17. -DaveM 21441da177e4SLinus Torvalds */ 2145dd0fc66fSAl Viro void tcp_send_active_reset(struct sock *sk, gfp_t priority) 21461da177e4SLinus Torvalds { 21471da177e4SLinus Torvalds struct sk_buff *skb; 21481da177e4SLinus Torvalds 21491da177e4SLinus Torvalds /* NOTE: No TCP options attached and we never retransmit this. */ 21501da177e4SLinus Torvalds skb = alloc_skb(MAX_TCP_HEADER, priority); 21511da177e4SLinus Torvalds if (!skb) { 21524e673444SPavel Emelyanov NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); 21531da177e4SLinus Torvalds return; 21541da177e4SLinus Torvalds } 21551da177e4SLinus Torvalds 21561da177e4SLinus Torvalds /* Reserve space for headers and prepare control bits. */ 21571da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 2158e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk), 2159e870a8efSIlpo Järvinen TCPCB_FLAG_ACK | TCPCB_FLAG_RST); 21601da177e4SLinus Torvalds /* Send it off. */ 21611da177e4SLinus Torvalds TCP_SKB_CB(skb)->when = tcp_time_stamp; 2162dfb4b9dcSDavid S. Miller if (tcp_transmit_skb(sk, skb, 0, priority)) 21634e673444SPavel Emelyanov NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); 216426af65cbSSridhar Samudrala 216581cc8a75SPavel Emelyanov TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS); 21661da177e4SLinus Torvalds } 21671da177e4SLinus Torvalds 21681da177e4SLinus Torvalds /* WARNING: This routine must only be called when we have already sent 21691da177e4SLinus Torvalds * a SYN packet that crossed the incoming SYN that caused this routine 21701da177e4SLinus Torvalds * to get called. If this assumption fails then the initial rcv_wnd 21711da177e4SLinus Torvalds * and rcv_wscale values will not be correct. 21721da177e4SLinus Torvalds */ 21731da177e4SLinus Torvalds int tcp_send_synack(struct sock *sk) 21741da177e4SLinus Torvalds { 21751da177e4SLinus Torvalds struct sk_buff *skb; 21761da177e4SLinus Torvalds 2177fe067e8aSDavid S. Miller skb = tcp_write_queue_head(sk); 21781da177e4SLinus Torvalds if (skb == NULL || !(TCP_SKB_CB(skb)->flags & TCPCB_FLAG_SYN)) { 21791da177e4SLinus Torvalds printk(KERN_DEBUG "tcp_send_synack: wrong queue state\n"); 21801da177e4SLinus Torvalds return -EFAULT; 21811da177e4SLinus Torvalds } 21821da177e4SLinus Torvalds if (!(TCP_SKB_CB(skb)->flags & TCPCB_FLAG_ACK)) { 21831da177e4SLinus Torvalds if (skb_cloned(skb)) { 21841da177e4SLinus Torvalds struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC); 21851da177e4SLinus Torvalds if (nskb == NULL) 21861da177e4SLinus Torvalds return -ENOMEM; 2187fe067e8aSDavid S. Miller tcp_unlink_write_queue(skb, sk); 21881da177e4SLinus Torvalds skb_header_release(nskb); 2189fe067e8aSDavid S. Miller __tcp_add_write_queue_head(sk, nskb); 21903ab224beSHideo Aoki sk_wmem_free_skb(sk, skb); 21913ab224beSHideo Aoki sk->sk_wmem_queued += nskb->truesize; 21923ab224beSHideo Aoki sk_mem_charge(sk, nskb->truesize); 21931da177e4SLinus Torvalds skb = nskb; 21941da177e4SLinus Torvalds } 21951da177e4SLinus Torvalds 21961da177e4SLinus Torvalds TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_ACK; 21971da177e4SLinus Torvalds TCP_ECN_send_synack(tcp_sk(sk), skb); 21981da177e4SLinus Torvalds } 21991da177e4SLinus Torvalds TCP_SKB_CB(skb)->when = tcp_time_stamp; 2200dfb4b9dcSDavid S. Miller return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 22011da177e4SLinus Torvalds } 22021da177e4SLinus Torvalds 22031da177e4SLinus Torvalds /* 22041da177e4SLinus Torvalds * Prepare a SYN-ACK. 22051da177e4SLinus Torvalds */ 22061da177e4SLinus Torvalds struct sk_buff *tcp_make_synack(struct sock *sk, struct dst_entry *dst, 220760236fddSArnaldo Carvalho de Melo struct request_sock *req) 22081da177e4SLinus Torvalds { 22092e6599cbSArnaldo Carvalho de Melo struct inet_request_sock *ireq = inet_rsk(req); 22101da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 22111da177e4SLinus Torvalds struct tcphdr *th; 22121da177e4SLinus Torvalds int tcp_header_size; 221333ad798cSAdam Langley struct tcp_out_options opts; 22141da177e4SLinus Torvalds struct sk_buff *skb; 2215cfb6eeb4SYOSHIFUJI Hideaki struct tcp_md5sig_key *md5; 2216cfb6eeb4SYOSHIFUJI Hideaki __u8 *md5_hash_location; 2217f5fff5dcSTom Quetchenbach int mss; 22181da177e4SLinus Torvalds 22191da177e4SLinus Torvalds skb = sock_wmalloc(sk, MAX_TCP_HEADER + 15, 1, GFP_ATOMIC); 22201da177e4SLinus Torvalds if (skb == NULL) 22211da177e4SLinus Torvalds return NULL; 22221da177e4SLinus Torvalds 22231da177e4SLinus Torvalds /* Reserve space for headers. */ 22241da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 22251da177e4SLinus Torvalds 22261da177e4SLinus Torvalds skb->dst = dst_clone(dst); 22271da177e4SLinus Torvalds 2228f5fff5dcSTom Quetchenbach mss = dst_metric(dst, RTAX_ADVMSS); 2229f5fff5dcSTom Quetchenbach if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < mss) 2230f5fff5dcSTom Quetchenbach mss = tp->rx_opt.user_mss; 2231f5fff5dcSTom Quetchenbach 223233ad798cSAdam Langley if (req->rcv_wnd == 0) { /* ignored for retransmitted syns */ 223333ad798cSAdam Langley __u8 rcv_wscale; 223433ad798cSAdam Langley /* Set this up on the first call only */ 223533ad798cSAdam Langley req->window_clamp = tp->window_clamp ? : dst_metric(dst, RTAX_WINDOW); 223633ad798cSAdam Langley /* tcp_full_space because it is guaranteed to be the first packet */ 223733ad798cSAdam Langley tcp_select_initial_window(tcp_full_space(sk), 2238f5fff5dcSTom Quetchenbach mss - (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0), 223933ad798cSAdam Langley &req->rcv_wnd, 224033ad798cSAdam Langley &req->window_clamp, 224133ad798cSAdam Langley ireq->wscale_ok, 224233ad798cSAdam Langley &rcv_wscale); 224333ad798cSAdam Langley ireq->rcv_wscale = rcv_wscale; 224433ad798cSAdam Langley } 2245cfb6eeb4SYOSHIFUJI Hideaki 224633ad798cSAdam Langley memset(&opts, 0, sizeof(opts)); 22478b5f12d0SFlorian Westphal #ifdef CONFIG_SYN_COOKIES 22488b5f12d0SFlorian Westphal if (unlikely(req->cookie_ts)) 22498b5f12d0SFlorian Westphal TCP_SKB_CB(skb)->when = cookie_init_timestamp(req); 22508b5f12d0SFlorian Westphal else 22518b5f12d0SFlorian Westphal #endif 225233ad798cSAdam Langley TCP_SKB_CB(skb)->when = tcp_time_stamp; 2253f5fff5dcSTom Quetchenbach tcp_header_size = tcp_synack_options(sk, req, mss, 225433ad798cSAdam Langley skb, &opts, &md5) + 225533ad798cSAdam Langley sizeof(struct tcphdr); 225633ad798cSAdam Langley 2257aa8223c7SArnaldo Carvalho de Melo skb_push(skb, tcp_header_size); 2258aa8223c7SArnaldo Carvalho de Melo skb_reset_transport_header(skb); 22591da177e4SLinus Torvalds 2260aa8223c7SArnaldo Carvalho de Melo th = tcp_hdr(skb); 22611da177e4SLinus Torvalds memset(th, 0, sizeof(struct tcphdr)); 22621da177e4SLinus Torvalds th->syn = 1; 22631da177e4SLinus Torvalds th->ack = 1; 22641da177e4SLinus Torvalds TCP_ECN_make_synack(req, th); 2265a3116ac5SKOVACS Krisztian th->source = ireq->loc_port; 22662e6599cbSArnaldo Carvalho de Melo th->dest = ireq->rmt_port; 2267e870a8efSIlpo Järvinen /* Setting of flags are superfluous here for callers (and ECE is 2268e870a8efSIlpo Järvinen * not even correctly set) 2269e870a8efSIlpo Järvinen */ 2270e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tcp_rsk(req)->snt_isn, 2271e870a8efSIlpo Järvinen TCPCB_FLAG_SYN | TCPCB_FLAG_ACK); 22721da177e4SLinus Torvalds th->seq = htonl(TCP_SKB_CB(skb)->seq); 22732e6599cbSArnaldo Carvalho de Melo th->ack_seq = htonl(tcp_rsk(req)->rcv_isn + 1); 22741da177e4SLinus Torvalds 22751da177e4SLinus Torvalds /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */ 2276600ff0c2SIlpo Järvinen th->window = htons(min(req->rcv_wnd, 65535U)); 227733ad798cSAdam Langley tcp_options_write((__be32 *)(th + 1), tp, &opts, &md5_hash_location); 22781da177e4SLinus Torvalds th->doff = (tcp_header_size >> 2); 227981cc8a75SPavel Emelyanov TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTSEGS); 2280cfb6eeb4SYOSHIFUJI Hideaki 2281cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 2282cfb6eeb4SYOSHIFUJI Hideaki /* Okay, we have all we need - do the md5 hash if needed */ 2283cfb6eeb4SYOSHIFUJI Hideaki if (md5) { 2284cfb6eeb4SYOSHIFUJI Hideaki tp->af_specific->calc_md5_hash(md5_hash_location, 228549a72dfbSAdam Langley md5, NULL, req, skb); 2286cfb6eeb4SYOSHIFUJI Hideaki } 2287cfb6eeb4SYOSHIFUJI Hideaki #endif 2288cfb6eeb4SYOSHIFUJI Hideaki 22891da177e4SLinus Torvalds return skb; 22901da177e4SLinus Torvalds } 22911da177e4SLinus Torvalds 22921da177e4SLinus Torvalds /* 22931da177e4SLinus Torvalds * Do all connect socket setups that can be done AF independent. 22941da177e4SLinus Torvalds */ 229540efc6faSStephen Hemminger static void tcp_connect_init(struct sock *sk) 22961da177e4SLinus Torvalds { 22971da177e4SLinus Torvalds struct dst_entry *dst = __sk_dst_get(sk); 22981da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 22991da177e4SLinus Torvalds __u8 rcv_wscale; 23001da177e4SLinus Torvalds 23011da177e4SLinus Torvalds /* We'll fix this up when we get a response from the other end. 23021da177e4SLinus Torvalds * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT. 23031da177e4SLinus Torvalds */ 23041da177e4SLinus Torvalds tp->tcp_header_len = sizeof(struct tcphdr) + 23051da177e4SLinus Torvalds (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0); 23061da177e4SLinus Torvalds 2307cfb6eeb4SYOSHIFUJI Hideaki #ifdef CONFIG_TCP_MD5SIG 2308cfb6eeb4SYOSHIFUJI Hideaki if (tp->af_specific->md5_lookup(sk, sk) != NULL) 2309cfb6eeb4SYOSHIFUJI Hideaki tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED; 2310cfb6eeb4SYOSHIFUJI Hideaki #endif 2311cfb6eeb4SYOSHIFUJI Hideaki 23121da177e4SLinus Torvalds /* If user gave his TCP_MAXSEG, record it to clamp */ 23131da177e4SLinus Torvalds if (tp->rx_opt.user_mss) 23141da177e4SLinus Torvalds tp->rx_opt.mss_clamp = tp->rx_opt.user_mss; 23151da177e4SLinus Torvalds tp->max_window = 0; 23165d424d5aSJohn Heffner tcp_mtup_init(sk); 23171da177e4SLinus Torvalds tcp_sync_mss(sk, dst_mtu(dst)); 23181da177e4SLinus Torvalds 23191da177e4SLinus Torvalds if (!tp->window_clamp) 23201da177e4SLinus Torvalds tp->window_clamp = dst_metric(dst, RTAX_WINDOW); 23211da177e4SLinus Torvalds tp->advmss = dst_metric(dst, RTAX_ADVMSS); 2322f5fff5dcSTom Quetchenbach if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->advmss) 2323f5fff5dcSTom Quetchenbach tp->advmss = tp->rx_opt.user_mss; 2324f5fff5dcSTom Quetchenbach 23251da177e4SLinus Torvalds tcp_initialize_rcv_mss(sk); 23261da177e4SLinus Torvalds 23271da177e4SLinus Torvalds tcp_select_initial_window(tcp_full_space(sk), 23281da177e4SLinus Torvalds tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0), 23291da177e4SLinus Torvalds &tp->rcv_wnd, 23301da177e4SLinus Torvalds &tp->window_clamp, 23311da177e4SLinus Torvalds sysctl_tcp_window_scaling, 23321da177e4SLinus Torvalds &rcv_wscale); 23331da177e4SLinus Torvalds 23341da177e4SLinus Torvalds tp->rx_opt.rcv_wscale = rcv_wscale; 23351da177e4SLinus Torvalds tp->rcv_ssthresh = tp->rcv_wnd; 23361da177e4SLinus Torvalds 23371da177e4SLinus Torvalds sk->sk_err = 0; 23381da177e4SLinus Torvalds sock_reset_flag(sk, SOCK_DONE); 23391da177e4SLinus Torvalds tp->snd_wnd = 0; 23401da177e4SLinus Torvalds tcp_init_wl(tp, tp->write_seq, 0); 23411da177e4SLinus Torvalds tp->snd_una = tp->write_seq; 23421da177e4SLinus Torvalds tp->snd_sml = tp->write_seq; 234333f5f57eSIlpo Järvinen tp->snd_up = tp->write_seq; 23441da177e4SLinus Torvalds tp->rcv_nxt = 0; 23451da177e4SLinus Torvalds tp->rcv_wup = 0; 23461da177e4SLinus Torvalds tp->copied_seq = 0; 23471da177e4SLinus Torvalds 2348463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT; 2349463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_retransmits = 0; 23501da177e4SLinus Torvalds tcp_clear_retrans(tp); 23511da177e4SLinus Torvalds } 23521da177e4SLinus Torvalds 23531da177e4SLinus Torvalds /* 23541da177e4SLinus Torvalds * Build a SYN and send it off. 23551da177e4SLinus Torvalds */ 23561da177e4SLinus Torvalds int tcp_connect(struct sock *sk) 23571da177e4SLinus Torvalds { 23581da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 23591da177e4SLinus Torvalds struct sk_buff *buff; 23601da177e4SLinus Torvalds 23611da177e4SLinus Torvalds tcp_connect_init(sk); 23621da177e4SLinus Torvalds 2363d179cd12SDavid S. Miller buff = alloc_skb_fclone(MAX_TCP_HEADER + 15, sk->sk_allocation); 23641da177e4SLinus Torvalds if (unlikely(buff == NULL)) 23651da177e4SLinus Torvalds return -ENOBUFS; 23661da177e4SLinus Torvalds 23671da177e4SLinus Torvalds /* Reserve space for headers. */ 23681da177e4SLinus Torvalds skb_reserve(buff, MAX_TCP_HEADER); 23691da177e4SLinus Torvalds 2370bd37a088SWei Yongjun tp->snd_nxt = tp->write_seq; 2371e870a8efSIlpo Järvinen tcp_init_nondata_skb(buff, tp->write_seq++, TCPCB_FLAG_SYN); 2372e870a8efSIlpo Järvinen TCP_ECN_send_syn(sk, buff); 23731da177e4SLinus Torvalds 23741da177e4SLinus Torvalds /* Send it off. */ 23751da177e4SLinus Torvalds TCP_SKB_CB(buff)->when = tcp_time_stamp; 23761da177e4SLinus Torvalds tp->retrans_stamp = TCP_SKB_CB(buff)->when; 23771da177e4SLinus Torvalds skb_header_release(buff); 2378fe067e8aSDavid S. Miller __tcp_add_write_queue_tail(sk, buff); 23793ab224beSHideo Aoki sk->sk_wmem_queued += buff->truesize; 23803ab224beSHideo Aoki sk_mem_charge(sk, buff->truesize); 23811da177e4SLinus Torvalds tp->packets_out += tcp_skb_pcount(buff); 2382dfb4b9dcSDavid S. Miller tcp_transmit_skb(sk, buff, 1, GFP_KERNEL); 2383bd37a088SWei Yongjun 2384bd37a088SWei Yongjun /* We change tp->snd_nxt after the tcp_transmit_skb() call 2385bd37a088SWei Yongjun * in order to make this packet get counted in tcpOutSegs. 2386bd37a088SWei Yongjun */ 2387bd37a088SWei Yongjun tp->snd_nxt = tp->write_seq; 2388bd37a088SWei Yongjun tp->pushed_seq = tp->write_seq; 238981cc8a75SPavel Emelyanov TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS); 23901da177e4SLinus Torvalds 23911da177e4SLinus Torvalds /* Timer for repeating the SYN until an answer. */ 23923f421baaSArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, 23933f421baaSArnaldo Carvalho de Melo inet_csk(sk)->icsk_rto, TCP_RTO_MAX); 23941da177e4SLinus Torvalds return 0; 23951da177e4SLinus Torvalds } 23961da177e4SLinus Torvalds 23971da177e4SLinus Torvalds /* Send out a delayed ack, the caller does the policy checking 23981da177e4SLinus Torvalds * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check() 23991da177e4SLinus Torvalds * for details. 24001da177e4SLinus Torvalds */ 24011da177e4SLinus Torvalds void tcp_send_delayed_ack(struct sock *sk) 24021da177e4SLinus Torvalds { 2403463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 2404463c84b9SArnaldo Carvalho de Melo int ato = icsk->icsk_ack.ato; 24051da177e4SLinus Torvalds unsigned long timeout; 24061da177e4SLinus Torvalds 24071da177e4SLinus Torvalds if (ato > TCP_DELACK_MIN) { 2408463c84b9SArnaldo Carvalho de Melo const struct tcp_sock *tp = tcp_sk(sk); 24091da177e4SLinus Torvalds int max_ato = HZ / 2; 24101da177e4SLinus Torvalds 2411056834d9SIlpo Järvinen if (icsk->icsk_ack.pingpong || 2412056834d9SIlpo Järvinen (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)) 24131da177e4SLinus Torvalds max_ato = TCP_DELACK_MAX; 24141da177e4SLinus Torvalds 24151da177e4SLinus Torvalds /* Slow path, intersegment interval is "high". */ 24161da177e4SLinus Torvalds 24171da177e4SLinus Torvalds /* If some rtt estimate is known, use it to bound delayed ack. 2418463c84b9SArnaldo Carvalho de Melo * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements 24191da177e4SLinus Torvalds * directly. 24201da177e4SLinus Torvalds */ 24211da177e4SLinus Torvalds if (tp->srtt) { 24221da177e4SLinus Torvalds int rtt = max(tp->srtt >> 3, TCP_DELACK_MIN); 24231da177e4SLinus Torvalds 24241da177e4SLinus Torvalds if (rtt < max_ato) 24251da177e4SLinus Torvalds max_ato = rtt; 24261da177e4SLinus Torvalds } 24271da177e4SLinus Torvalds 24281da177e4SLinus Torvalds ato = min(ato, max_ato); 24291da177e4SLinus Torvalds } 24301da177e4SLinus Torvalds 24311da177e4SLinus Torvalds /* Stay within the limit we were given */ 24321da177e4SLinus Torvalds timeout = jiffies + ato; 24331da177e4SLinus Torvalds 24341da177e4SLinus Torvalds /* Use new timeout only if there wasn't a older one earlier. */ 2435463c84b9SArnaldo Carvalho de Melo if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) { 24361da177e4SLinus Torvalds /* If delack timer was blocked or is about to expire, 24371da177e4SLinus Torvalds * send ACK now. 24381da177e4SLinus Torvalds */ 2439463c84b9SArnaldo Carvalho de Melo if (icsk->icsk_ack.blocked || 2440463c84b9SArnaldo Carvalho de Melo time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) { 24411da177e4SLinus Torvalds tcp_send_ack(sk); 24421da177e4SLinus Torvalds return; 24431da177e4SLinus Torvalds } 24441da177e4SLinus Torvalds 2445463c84b9SArnaldo Carvalho de Melo if (!time_before(timeout, icsk->icsk_ack.timeout)) 2446463c84b9SArnaldo Carvalho de Melo timeout = icsk->icsk_ack.timeout; 24471da177e4SLinus Torvalds } 2448463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER; 2449463c84b9SArnaldo Carvalho de Melo icsk->icsk_ack.timeout = timeout; 2450463c84b9SArnaldo Carvalho de Melo sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout); 24511da177e4SLinus Torvalds } 24521da177e4SLinus Torvalds 24531da177e4SLinus Torvalds /* This routine sends an ack and also updates the window. */ 24541da177e4SLinus Torvalds void tcp_send_ack(struct sock *sk) 24551da177e4SLinus Torvalds { 24561da177e4SLinus Torvalds struct sk_buff *buff; 24571da177e4SLinus Torvalds 2458058dc334SIlpo Järvinen /* If we have been reset, we may not send again. */ 2459058dc334SIlpo Järvinen if (sk->sk_state == TCP_CLOSE) 2460058dc334SIlpo Järvinen return; 2461058dc334SIlpo Järvinen 24621da177e4SLinus Torvalds /* We are not putting this on the write queue, so 24631da177e4SLinus Torvalds * tcp_transmit_skb() will set the ownership to this 24641da177e4SLinus Torvalds * sock. 24651da177e4SLinus Torvalds */ 24661da177e4SLinus Torvalds buff = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC); 24671da177e4SLinus Torvalds if (buff == NULL) { 2468463c84b9SArnaldo Carvalho de Melo inet_csk_schedule_ack(sk); 2469463c84b9SArnaldo Carvalho de Melo inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN; 24703f421baaSArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK, 24713f421baaSArnaldo Carvalho de Melo TCP_DELACK_MAX, TCP_RTO_MAX); 24721da177e4SLinus Torvalds return; 24731da177e4SLinus Torvalds } 24741da177e4SLinus Torvalds 24751da177e4SLinus Torvalds /* Reserve space for headers and prepare control bits. */ 24761da177e4SLinus Torvalds skb_reserve(buff, MAX_TCP_HEADER); 2477e870a8efSIlpo Järvinen tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPCB_FLAG_ACK); 24781da177e4SLinus Torvalds 24791da177e4SLinus Torvalds /* Send it off, this clears delayed acks for us. */ 24801da177e4SLinus Torvalds TCP_SKB_CB(buff)->when = tcp_time_stamp; 2481dfb4b9dcSDavid S. Miller tcp_transmit_skb(sk, buff, 0, GFP_ATOMIC); 24821da177e4SLinus Torvalds } 24831da177e4SLinus Torvalds 24841da177e4SLinus Torvalds /* This routine sends a packet with an out of date sequence 24851da177e4SLinus Torvalds * number. It assumes the other end will try to ack it. 24861da177e4SLinus Torvalds * 24871da177e4SLinus Torvalds * Question: what should we make while urgent mode? 24881da177e4SLinus Torvalds * 4.4BSD forces sending single byte of data. We cannot send 24891da177e4SLinus Torvalds * out of window data, because we have SND.NXT==SND.MAX... 24901da177e4SLinus Torvalds * 24911da177e4SLinus Torvalds * Current solution: to send TWO zero-length segments in urgent mode: 24921da177e4SLinus Torvalds * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is 24931da177e4SLinus Torvalds * out-of-date with SND.UNA-1 to probe window. 24941da177e4SLinus Torvalds */ 24951da177e4SLinus Torvalds static int tcp_xmit_probe_skb(struct sock *sk, int urgent) 24961da177e4SLinus Torvalds { 24971da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 24981da177e4SLinus Torvalds struct sk_buff *skb; 24991da177e4SLinus Torvalds 25001da177e4SLinus Torvalds /* We don't queue it, tcp_transmit_skb() sets ownership. */ 25011da177e4SLinus Torvalds skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC); 25021da177e4SLinus Torvalds if (skb == NULL) 25031da177e4SLinus Torvalds return -1; 25041da177e4SLinus Torvalds 25051da177e4SLinus Torvalds /* Reserve space for headers and set control bits. */ 25061da177e4SLinus Torvalds skb_reserve(skb, MAX_TCP_HEADER); 25071da177e4SLinus Torvalds /* Use a previous sequence. This should cause the other 25081da177e4SLinus Torvalds * end to send an ack. Don't queue or clone SKB, just 25091da177e4SLinus Torvalds * send it. 25101da177e4SLinus Torvalds */ 2511e870a8efSIlpo Järvinen tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPCB_FLAG_ACK); 25121da177e4SLinus Torvalds TCP_SKB_CB(skb)->when = tcp_time_stamp; 2513dfb4b9dcSDavid S. Miller return tcp_transmit_skb(sk, skb, 0, GFP_ATOMIC); 25141da177e4SLinus Torvalds } 25151da177e4SLinus Torvalds 25161da177e4SLinus Torvalds int tcp_write_wakeup(struct sock *sk) 25171da177e4SLinus Torvalds { 25181da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 25191da177e4SLinus Torvalds struct sk_buff *skb; 25201da177e4SLinus Torvalds 2521058dc334SIlpo Järvinen if (sk->sk_state == TCP_CLOSE) 2522058dc334SIlpo Järvinen return -1; 2523058dc334SIlpo Järvinen 2524fe067e8aSDavid S. Miller if ((skb = tcp_send_head(sk)) != NULL && 252590840defSIlpo Järvinen before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) { 25261da177e4SLinus Torvalds int err; 25271da177e4SLinus Torvalds unsigned int mss = tcp_current_mss(sk, 0); 252890840defSIlpo Järvinen unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; 25291da177e4SLinus Torvalds 25301da177e4SLinus Torvalds if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq)) 25311da177e4SLinus Torvalds tp->pushed_seq = TCP_SKB_CB(skb)->end_seq; 25321da177e4SLinus Torvalds 25331da177e4SLinus Torvalds /* We are probing the opening of a window 25341da177e4SLinus Torvalds * but the window size is != 0 25351da177e4SLinus Torvalds * must have been a result SWS avoidance ( sender ) 25361da177e4SLinus Torvalds */ 25371da177e4SLinus Torvalds if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq || 25381da177e4SLinus Torvalds skb->len > mss) { 25391da177e4SLinus Torvalds seg_size = min(seg_size, mss); 25401da177e4SLinus Torvalds TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH; 2541846998aeSDavid S. Miller if (tcp_fragment(sk, skb, seg_size, mss)) 25421da177e4SLinus Torvalds return -1; 25431da177e4SLinus Torvalds } else if (!tcp_skb_pcount(skb)) 2544846998aeSDavid S. Miller tcp_set_skb_tso_segs(sk, skb, mss); 25451da177e4SLinus Torvalds 25461da177e4SLinus Torvalds TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH; 25471da177e4SLinus Torvalds TCP_SKB_CB(skb)->when = tcp_time_stamp; 2548dfb4b9dcSDavid S. Miller err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); 254966f5fe62SIlpo Järvinen if (!err) 255066f5fe62SIlpo Järvinen tcp_event_new_data_sent(sk, skb); 25511da177e4SLinus Torvalds return err; 25521da177e4SLinus Torvalds } else { 255333f5f57eSIlpo Järvinen if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF)) 25544828e7f4SIlpo Järvinen tcp_xmit_probe_skb(sk, 1); 25551da177e4SLinus Torvalds return tcp_xmit_probe_skb(sk, 0); 25561da177e4SLinus Torvalds } 25571da177e4SLinus Torvalds } 25581da177e4SLinus Torvalds 25591da177e4SLinus Torvalds /* A window probe timeout has occurred. If window is not closed send 25601da177e4SLinus Torvalds * a partial packet else a zero probe. 25611da177e4SLinus Torvalds */ 25621da177e4SLinus Torvalds void tcp_send_probe0(struct sock *sk) 25631da177e4SLinus Torvalds { 2564463c84b9SArnaldo Carvalho de Melo struct inet_connection_sock *icsk = inet_csk(sk); 25651da177e4SLinus Torvalds struct tcp_sock *tp = tcp_sk(sk); 25661da177e4SLinus Torvalds int err; 25671da177e4SLinus Torvalds 25681da177e4SLinus Torvalds err = tcp_write_wakeup(sk); 25691da177e4SLinus Torvalds 2570fe067e8aSDavid S. Miller if (tp->packets_out || !tcp_send_head(sk)) { 25711da177e4SLinus Torvalds /* Cancel probe timer, if it is not required. */ 25726687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out = 0; 2573463c84b9SArnaldo Carvalho de Melo icsk->icsk_backoff = 0; 25741da177e4SLinus Torvalds return; 25751da177e4SLinus Torvalds } 25761da177e4SLinus Torvalds 25771da177e4SLinus Torvalds if (err <= 0) { 2578463c84b9SArnaldo Carvalho de Melo if (icsk->icsk_backoff < sysctl_tcp_retries2) 2579463c84b9SArnaldo Carvalho de Melo icsk->icsk_backoff++; 25806687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out++; 2581463c84b9SArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0, 25823f421baaSArnaldo Carvalho de Melo min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX), 25833f421baaSArnaldo Carvalho de Melo TCP_RTO_MAX); 25841da177e4SLinus Torvalds } else { 25851da177e4SLinus Torvalds /* If packet was not sent due to local congestion, 25866687e988SArnaldo Carvalho de Melo * do not backoff and do not remember icsk_probes_out. 25871da177e4SLinus Torvalds * Let local senders to fight for local resources. 25881da177e4SLinus Torvalds * 25891da177e4SLinus Torvalds * Use accumulated backoff yet. 25901da177e4SLinus Torvalds */ 25916687e988SArnaldo Carvalho de Melo if (!icsk->icsk_probes_out) 25926687e988SArnaldo Carvalho de Melo icsk->icsk_probes_out = 1; 2593463c84b9SArnaldo Carvalho de Melo inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0, 2594463c84b9SArnaldo Carvalho de Melo min(icsk->icsk_rto << icsk->icsk_backoff, 25953f421baaSArnaldo Carvalho de Melo TCP_RESOURCE_PROBE_INTERVAL), 25963f421baaSArnaldo Carvalho de Melo TCP_RTO_MAX); 25971da177e4SLinus Torvalds } 25981da177e4SLinus Torvalds } 25991da177e4SLinus Torvalds 2600c6aefafbSGlenn Griffin EXPORT_SYMBOL(tcp_select_initial_window); 26011da177e4SLinus Torvalds EXPORT_SYMBOL(tcp_connect); 26021da177e4SLinus Torvalds EXPORT_SYMBOL(tcp_make_synack); 26031da177e4SLinus Torvalds EXPORT_SYMBOL(tcp_simple_retransmit); 26041da177e4SLinus Torvalds EXPORT_SYMBOL(tcp_sync_mss); 26055d424d5aSJohn Heffner EXPORT_SYMBOL(tcp_mtup_init); 2606